Please contact me privately or (preferably the list) for questions,
comments, discussions, flames, names, or anything. I'll do complete
rewrites of the patches if someone tells me how they don't meet their
needs or how they can be done better. I'm here to try to bridge the
needs (and wants) of the anti-malware vendors with the technical
realities of the kernel. So everyone feel free to throw in your two
cents and I'll try to reconcile it all. These 5 patches are part 1.
They give us a working able solution.
>From my point of view patches forthcoming and mentioned below should
help with performance for those who actually have userspace scanners but
also could presents be implemented using this framework.
Background
++++++++++
There is a consensus in the security industry that protecting against
malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
of so-called on-access scanning is usable and reasonable approach.
Currently the Linux kernel does not offer a completely suitable
interface to implement such security solutions. Present solutions
involve overwriting function pointers in the LSM, in filesystem
operations, in the sycall table, and other fragile hacks. The purpose
of this project is to create a fast, clean interface for userspace
programs to look for malware when files are accessed. This malware may
be ultimately intended for this or some other Linux machine or may be
malware intended to attack a host running a different operating system
and is merely in transit across the Linux server. Since there are
almost an infinite number of ways in which information can enter and
exit a server it is not seen as reasonable to move these checks to all
the applications at the boundary (MTA, NFS, CIFS, SSH, rsync, et al.) to
look for such malware on at the border.
For this Linux kernel interface speed is of particular interest for
those who have it compiled into the kernel but have no userspace client.
There must be no measurable performance hit to just compiling this into
the kernel.
Security vendors, Linux distributors and other interested parties have
come together on the malware-list mailing list to discuss this problem
and see if they can work together to propose a solution. During these
talks couple of requirement sets were posted with the aim of fleshing
out common needs as a prerequisite of creating an interface prototype.
Collated requirements
+++++++++++++++++++++
1. Intercept file opens (exec also) for vetting (block until decision is made) and allow some userspace black magic to make decisions.
2. Intercept file closes for scanning post access
3. Cache scan results so the same file is not scanned on each and every access
4. Ability to flush the cache and cause all files to be re-scanned when accessed
5. Define which filesystems are cacheable and which are not
6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
7. Report other relevant file, process and user information associated with each interception
8. Report file pathnames to userspace (relative to process root, current working directory)
9. Mark a processes as exempt from on access scanning
10. Exclude sub-trees from scanning based on filesystem (exclude procfs, sysfs, devfs)
11. Exclude sub-trees from scanning based on filesystem path
12. Include only certain sub-trees from scanning based on filesystem path
13. Register more than one userspace client in which case behavior is restrictive
Discussion of requirements
++++++++++++++++++++++++++
The initial patch set with NOT meet all of these 'requirements.' Some
will be implemented at a later time and some will never be implemented.
Specifics are detailed below. There is no intention to (abu)use the LSM
for this purpose. The LSM provides complete internal kernel mandatory
access controls. It is not intended for userspace scanning and
detection. Users should not be forced to choose between an in kernel
mandatory access control policy and this additional userspace file
access. LSM stacking is NOT as option as has been demonstrated
repeatedly.
1., 2. Basic interception
-------------------------
Core requirement is to intercept access to files and prevent it if
malicious content is detected. This is done on open, not on read. It
may be possible to do read time checking with minimal performance impact
although not currently implemented. This means that the following race
is possible
Process1 Process2
- open file RD
- open file WR
- write virus data (1)
- read virus data
*note that any open after (1) will get properly vetted. At this time
the likely hood of this being a problem vs the performance impact of
scanning on read and the increased complexity of the code means this is
left out. This should not be a problem for local executables as writes
to files opened to be run typically return ETXTBSY.
To accomplish that two hooks were inserted, on file open in
__dentry_open and in filp_close on file close. In both cases the file
object in question is passed as a parameter for further processing. In
case of an open the operation can actually be blocked, while closes are
always immediately successful and will not cause additional blocking.
Results of a close are returned to the kernel asynchronously and may be
used to cache answers to speed up a future open.
Interception processing is done by way of three chains of filters.
Access requests are first send to the "evaluation" chain. Depending on
the results of the evaluation the decision is then send to either the
allow chain or the deny chain.
There are three basic responses each filter can make - to be indifferent
or either allow or deny access to the file. The filter may also allow
or deny access to a file while not caching that result.
One of the most important filters in the evaluation chain implements an
interface through which an userspace process can register and receive
vetting requests. Userspace process opens a misc character device to
express its interest and then receives binary structures from that
device describing basic interception information. After file contents
have been scanned a vetting response is sent by writing a different
binary structure back to the device and the intercepted process
continues its execution. These are not done over network sockets and no
endian conversions are done. The client and the kernel must have the
same endian configuration.
3., 4. Caching
---------------
To avoid scanning unchanged files on every access which would be very
bad for performance some sort of caching is needed. Although possible
to implement a cache in userspace having two context switches required
for every open is clearly not fast. We implemented it per inode object
as a serial number compared with a single global monotonically
increasing system serial number.
The cache filter is inserted into the evaluation chain before the
userspace client filter and if the inode serial number is equal to the
system one it allows access to the file.
If the file is seen for the first time, has been modified, or for any
other reason has a serial number less than the system one the cache
filter will be 'indifferent' and processing of the given vetting request
will continue down the evaluation chain. When some filter (only
Userspace in the first patch set) allows access to a file its inode
serial number is set to the system global which effectively makes it
cached. Also, when a write access is gained for a file the serial number
will automatically be reset as well as when any process actually writes
to that file.
Cache flushing is possible by simply increasing the global system serial
number.
Both positive and negative vetting results are cached by the means of
positive and negative serial numbers.
This method of caching has minimal impact on system resources while
providing maximal effectiveness and simple implementation.
5. Fine-grained caching
-----------------------
It is necessary to select which filesystems can be safely cached and
which must not be. For example it is not a good idea to allow caching of
network filesystems because their content can be changed invisibly. Disk
based and some virtual filesystems can be cached safely on the other
hand.
This first proposal only partially implements this requirement. Only
block device backed filesystems will be cached while there is no way to
enable caching for things like tmpfs. Improving this is left out of the
initial prototype. Although there may be additional work to implement
caching for certain FS types there is no plan to greatly increase the
scope of the cache granularity. There is no plan to cache based on the
operation or things of that nature. Caching of this nature can be
implemented in userspace if the vendor so chooses. We include only a
minimal safe cache for performance reasons.
6. Direct access to file content
--------------------------------
When an userspace daemon receives a vetting request, it also receives a
new RO file descriptor which provides direct access to the inode in
question. This is to enable access to the file regardless of it
accessibility from the scanner environment (consider process namespaces,
chroot's, NFS). The userspace client is responsible for closing this
file when it is finished scanning.
7. Other reporting
------------------
Along with the fd being installed in the scanning process the process
gets a binary structure of data including:
+ uint32_t version;
+ uint32_t type;
+ int32_t fd;
+ uint32_t operation;
+ uint32_t flags;
+ uint32_t mode;
+ uint32_t uid;
+ uint32_t gid;
+ uint32_t tgid;
+ uint32_t pid;
8. Path name reporting
----------------------
When a malicious content is detected in a file it is important to be
able to report its location so the user or system administrator can take
appropriate actions.
This is implemented in a amazingly simple way which will hopefully avoid
the controversy of some other solutions. Path name is only needed for
reporting purposes and it is obtained by reading the symlink of the
given file descriptor in /proc. Its as simple as userspace calling:
snprintf(link, sizeof(link), "/proc/self/fd/%d", details.fd);
ret = readlink(link, buf, sizeof(buf)-1);
9. Process exclusion
--------------------
Sometimes it is necessary to exclude certain processes from being
intercepted. For example it might be a userspace root kit scanner which
would not be able to find root kits if access to them was blocked by the
on-access scanner.
To facilitate that we have created a special file a process can open and
register itself as excluded. A flag is then put into its kernel
structure (task_struct) which makes it excluded from scanning.
This implementation is very simple and provides greatest performance. In
the proposed implementation access to the exclusion device is controlled
though permissions on the device node which are not sufficient. An LSM
call will need to be made for this type or access in a later patch.
10. Filesystem exclusions
-------------------------
One pretty important optimization is not to scan things like /proc, /sys
or similar. Basically all filesystems where user can not store
arbitrary, potentially malicious, content could and should be excluded
from scanning.
This interface prototype implements it as a run-time configurable list
of filesystem names. Again it is a filter in the evaluation chain which
can allow access before the request gets routed to the userspace client.
This will not be implemented in the first patch set but should be soon
to follow. It is done by simply comparing strings between those
supplied and the s_type->name field in an associated superblock.
11. Path exclusions
-------------------
The need for exclusions can be demonstrated with an example of a MySQL
server. It's data files are frequently modified which means they would
need to be constantly rescanned which is very bad for performance. Also,
it is most often not even possible to reasonably scan them. Therefore
the best solution is not to scan its database store which can simply be
implemented by excluding the store subdirectory.
It is a relatively simple implementation which allows run-time
configuration of a list of sub directories or files to exclude.
Exclusion paths are relative to each process root. So for example if we
want to exclude /var/lib/mysql/ and we have a mysql running in a chroot
where from the outside that directory actually lives
in /chroot/mysql/var/lib/mysql, /var/lib/mysql should actually be added
to the exclusion list.
This is also not included in the initial patch set but will be coming
shortly after.
12. Path Inclusions
-------------------
Path-based inclusions are not implemented due to concerns with
hard-linked files both inside and outside the included directories. It
is too easy to fall into a sense of false security with path inclusions
since the pathname is almost meaningless. If a vendor feels this is
particularly important for them they will have to implement it in
userspace by use of a judicious list of exclusion filters.
13. Multiple client registration with restrictive behavior
-----------------------------------------------------------
This is currently not implemented. Multiple clients can register but
they will be used for (crappy) load balancing only. Not all will be
called for a single interception. Only one of the registered clients
will process a single interception. Desire here is to enable multiple
clients servicing interceptions in parallel for performance and
reliability reasons.
Requirement for serial and restrictive behavior would be slightly more
complicated to implement because we would want to keep the current
behavior as well. Or in other words we would need to have groups of
multiple clients, where each interception would go through one client
from each group with the desired restrictive behavior.
This may be left for a future implementation for simplicity reasons but
I find it unlikely. If a vendor needs to send requests to multiple
scanners they should be able to implement that serialization in
userspace. I see no need for an in kernel event dispatcher. Note that
the audit system had this same need and has done it as a userspace event
dispatcher. We have also seen in the LSM that restrictive access
stacking is not as easy as it sounds and has been abandoned.
Closing remarks
---------------
Although some may argue some of the filters are not necessary or may
better be implemented in userspace, we think it is better to have them
in kernel primarily for performance reasons. Secondly, it is all simple
code not introducing much baggage or risk into the kernel itself. The
most complex filter and the only one with locking ramifications is the
userspace client vetting which calls into dentry_open() on both open and
close operations. There is no locking around caching or process
exclusions or other work.
**************************
The patches can be found in a git tree located:
http://git.infradead.org/users/eparis/talpa.git
since:
commit 2b12a4c524812fb3f6ee590a02e65b95c8c32229
Author: Linus Torvalds <[email protected]>
Date: Fri Aug 1 14:59:11 2008 -0700
Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6
This tree will be rebased regularly, so please do not just start pulling
and hoping it will continue to always merge. My current plan is to
commit changes and comments on the end of this tree and eventually
reroll those changes into these 5 patches for finally submission to
upstream. Likely this will be an iterative process.
The 5 patches in the following e-mails can also be found at
http://people.redhat.com/~eparis/talpa
Documentation/talpa/allow_most.c | 138 ++++++++
Documentation/talpa/cache | 17 +
Documentation/talpa/client | 85 +++++
Documentation/talpa/design.txt | 266 +++++++++++++++
Documentation/talpa/tecat.c | 50 ++
Documentation/talpa/test_deny.c | 356 ++++++++++++++++++++
Documentation/talpa/thread_exclude | 6
fs/inode.c | 6
fs/namei.c | 2
fs/open.c | 10
include/linux/fs.h | 5
include/linux/sched.h | 1
include/linux/talpa.h | 188 +++++++++++
security/Kconfig | 1
security/Makefile | 2
security/talpa/Kconfig | 51 +++
security/talpa/Makefile | 17 -
security/talpa/talpa.h | 115 ++++++
security/talpa/talpa_allow_calls.h | 12
security/talpa/talpa_cache.c | 207 ++++++++++++
security/talpa/talpa_cache.h | 22 +
security/talpa/talpa_client.c | 543 ++++++++++++++++++++++++++++++++
security/talpa/talpa_common.c | 56 +++
security/talpa/talpa_configuration.c | 156 +++++++++
security/talpa/talpa_deny_calls.h | 11
security/talpa/talpa_evaluation_calls.h | 42 ++
security/talpa/talpa_interceptor.c | 121 +++++++
security/talpa/talpa_thread_exclude.c | 67 +++
28 files changed, 2546 insertions(+), 7 deletions(-)
On Mon, Aug 04, 2008 at 05:00:16PM -0400, Eric Paris wrote:
> Please contact me privately or (preferably the list) for questions,
> comments, discussions, flames, names, or anything. I'll do complete
> rewrites of the patches if someone tells me how they don't meet their
> needs or how they can be done better. I'm here to try to bridge the
> needs (and wants) of the anti-malware vendors with the technical
> realities of the kernel. So everyone feel free to throw in your two
> cents and I'll try to reconcile it all. These 5 patches are part 1.
> They give us a working able solution.
>
> >From my point of view patches forthcoming and mentioned below should
> help with performance for those who actually have userspace scanners but
> also could presents be implemented using this framework.
>
>
> Background
> ++++++++++
> There is a consensus in the security industry that protecting against
> malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
> of so-called on-access scanning is usable and reasonable approach.
> Currently the Linux kernel does not offer a completely suitable
> interface to implement such security solutions. Present solutions
> involve overwriting function pointers in the LSM, in filesystem
> operations, in the sycall table, and other fragile hacks. The purpose
> of this project is to create a fast, clean interface for userspace
> programs to look for malware when files are accessed. This malware may
> be ultimately intended for this or some other Linux machine or may be
> malware intended to attack a host running a different operating system
> and is merely in transit across the Linux server. Since there are
> almost an infinite number of ways in which information can enter and
> exit a server it is not seen as reasonable to move these checks to all
> the applications at the boundary (MTA, NFS, CIFS, SSH, rsync, et al.) to
> look for such malware on at the border.
>
> For this Linux kernel interface speed is of particular interest for
> those who have it compiled into the kernel but have no userspace client.
> There must be no measurable performance hit to just compiling this into
> the kernel.
>
> Security vendors, Linux distributors and other interested parties have
> come together on the malware-list mailing list to discuss this problem
> and see if they can work together to propose a solution. During these
> talks couple of requirement sets were posted with the aim of fleshing
> out common needs as a prerequisite of creating an interface prototype.
These requirements were posted? Where? I don't recall seeing them.
> Collated requirements
> +++++++++++++++++++++
> 1. Intercept file opens (exec also) for vetting (block until
> decision is made) and allow some userspace black magic to make
> decisions.
> 2. Intercept file closes for scanning post access
> 3. Cache scan results so the same file is not scanned on each and every access
> 4. Ability to flush the cache and cause all files to be re-scanned when accessed
> 5. Define which filesystems are cacheable and which are not
> 6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
> 7. Report other relevant file, process and user information associated with each interception
> 8. Report file pathnames to userspace (relative to process root, current working directory)
> 9. Mark a processes as exempt from on access scanning
> 10. Exclude sub-trees from scanning based on filesystem (exclude procfs, sysfs, devfs)
> 11. Exclude sub-trees from scanning based on filesystem path
> 12. Include only certain sub-trees from scanning based on filesystem path
> 13. Register more than one userspace client in which case behavior is restrictive
I don't see anything in the list above that make this a requirement that
the code to do this be placed within the kernel.
What is wrong with doing it in glibc or some other system-wide library
(LD_PRELOAD hooks, etc.)?
> 1., 2. Basic interception
> -------------------------
> Core requirement is to intercept access to files and prevent it if
> malicious content is detected. This is done on open, not on read. It
> may be possible to do read time checking with minimal performance impact
> although not currently implemented. This means that the following race
> is possible
>
> Process1 Process2
> - open file RD
> - open file WR
> - write virus data (1)
> - read virus data
Wonderful, we are going to implement a solution that is known to not
work, with a trivial way around it?
Sorry, that's not going to fly.
> *note that any open after (1) will get properly vetted. At this time
> the likely hood of this being a problem vs the performance impact of
> scanning on read and the increased complexity of the code means this is
> left out. This should not be a problem for local executables as writes
> to files opened to be run typically return ETXTBSY.
Are you sure about this?
> One of the most important filters in the evaluation chain implements an
> interface through which an userspace process can register and receive
> vetting requests. Userspace process opens a misc character device to
> express its interest and then receives binary structures from that
> device describing basic interception information. After file contents
> have been scanned a vetting response is sent by writing a different
> binary structure back to the device and the intercepted process
> continues its execution. These are not done over network sockets and no
> endian conversions are done. The client and the kernel must have the
> same endian configuration.
How about the same 64/32bit requirement? Your implementation is
incorrect otherwise.
(hint, your current patch is also wrong in this area, you should fix
that up...)
And a binary structure? Ick, are you trying to make it hard for future
expansions and such?
And why not netlink/network socket? Why a character device? You are
already using securityfs, why not use a file node in there?
> 6. Direct access to file content
> --------------------------------
> When an userspace daemon receives a vetting request, it also receives a
> new RO file descriptor which provides direct access to the inode in
> question. This is to enable access to the file regardless of it
> accessibility from the scanner environment (consider process namespaces,
> chroot's, NFS). The userspace client is responsible for closing this
> file when it is finished scanning.
Is this secondary file handle properly checked for the security issues
involved with such a thing? What happens if the userspace client does
not close the file handle?
> 7. Other reporting
> ------------------
> Along with the fd being installed in the scanning process the process
> gets a binary structure of data including:
What's with the love of binary structures? :)
> + uint32_t version;
> + uint32_t type;
> + int32_t fd;
> + uint32_t operation;
> + uint32_t flags;
> + uint32_t mode;
> + uint32_t uid;
> + uint32_t gid;
> + uint32_t tgid;
> + uint32_t pid;
What happens when the world moves to 128bit or 64bit uids? (yes, I've
seen proposals for such a thing...)
Why would userspace care about these meta-file things, what does it want
with them?
> 8. Path name reporting
> ----------------------
> When a malicious content is detected in a file it is important to be
> able to report its location so the user or system administrator can take
> appropriate actions.
>
> This is implemented in a amazingly simple way which will hopefully avoid
> the controversy of some other solutions. Path name is only needed for
> reporting purposes and it is obtained by reading the symlink of the
> given file descriptor in /proc. Its as simple as userspace calling:
>
> snprintf(link, sizeof(link), "/proc/self/fd/%d", details.fd);
> ret = readlink(link, buf, sizeof(buf)-1);
Cute hack. What's to keep it from racing with the fd changing from the
original program?
> 9. Process exclusion
> --------------------
> Sometimes it is necessary to exclude certain processes from being
> intercepted. For example it might be a userspace root kit scanner which
> would not be able to find root kits if access to them was blocked by the
> on-access scanner.
>
> To facilitate that we have created a special file a process can open and
> register itself as excluded. A flag is then put into its kernel
> structure (task_struct) which makes it excluded from scanning.
>
> This implementation is very simple and provides greatest performance. In
> the proposed implementation access to the exclusion device is controlled
> though permissions on the device node which are not sufficient. An LSM
> call will need to be made for this type or access in a later patch.
Heh, so if you want to write a "virus" for Linux, just implement this
flag. What's to keep a "rogue" program from telling the kernel that all
programs on the system are to be excluded?
> 10. Filesystem exclusions
> -------------------------
> One pretty important optimization is not to scan things like /proc, /sys
> or similar. Basically all filesystems where user can not store
> arbitrary, potentially malicious, content could and should be excluded
> from scanning.
Why, does scanning these files take extra time? Just curious.
> 11. Path exclusions
> -------------------
> The need for exclusions can be demonstrated with an example of a MySQL
> server. It's data files are frequently modified which means they would
> need to be constantly rescanned which is very bad for performance. Also,
> it is most often not even possible to reasonably scan them. Therefore
> the best solution is not to scan its database store which can simply be
> implemented by excluding the store subdirectory.
>
> It is a relatively simple implementation which allows run-time
> configuration of a list of sub directories or files to exclude.
> Exclusion paths are relative to each process root. So for example if we
> want to exclude /var/lib/mysql/ and we have a mysql running in a chroot
> where from the outside that directory actually lives
> in /chroot/mysql/var/lib/mysql, /var/lib/mysql should actually be added
> to the exclusion list.
>
> This is also not included in the initial patch set but will be coming
> shortly after.
Again, what's to keep all files to be marked as excluded?
> Closing remarks
> ---------------
> Although some may argue some of the filters are not necessary or may
> better be implemented in userspace, we think it is better to have them
> in kernel primarily for performance reasons.
Why? What numbers do you have that say the kernel is faster in
implementing this? This is the first mention of such a requirement, we
need to see real data to back it up please.
> Secondly, it is all simple code not introducing much baggage or risk
> into the kernel itself.
I disagree, see above.
thanks,
greg k-h
On Mon, Aug 04, 2008 at 03:32:49PM -0700, Greg KH wrote:
> > 1. Intercept file opens (exec also) for vetting (block until
> > decision is made) and allow some userspace black magic to make
> > decisions.
NACK, this kind of policy should be done in kernelspace.
> > 2. Intercept file closes for scanning post access
Not even possible for mmap, and dumb otherwise, NACK.
> > 6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
Explain?
> > 9. Mark a processes as exempt from on access scanning
Nack, this completely defeats the purpose.
> > 11. Exclude sub-trees from scanning based on filesystem path
> > 12. Include only certain sub-trees from scanning based on filesystem path
Nack, please no pathname based idiocies.
On Mon, 2008-08-04 at 15:32 -0700, Greg KH wrote:
> On Mon, Aug 04, 2008 at 05:00:16PM -0400, Eric Paris wrote:
> > Security vendors, Linux distributors and other interested parties have
> > come together on the malware-list mailing list to discuss this problem
> > and see if they can work together to propose a solution. During these
> > talks couple of requirement sets were posted with the aim of fleshing
> > out common needs as a prerequisite of creating an interface prototype.
>
> These requirements were posted? Where? I don't recall seeing them.
they were collected from the comments of Sophos, CA, and McAfee on
[email protected] back in January 2008. I can't find the
lists archived on the net so I will post the raw messages tomorrow from
my local mail store and send a link.
>
> > Collated requirements
> > +++++++++++++++++++++
> > 1. Intercept file opens (exec also) for vetting (block until
> > decision is made) and allow some userspace black magic to make
> > decisions.
> > 2. Intercept file closes for scanning post access
> > 3. Cache scan results so the same file is not scanned on each and every access
> > 4. Ability to flush the cache and cause all files to be re-scanned when accessed
> > 5. Define which filesystems are cacheable and which are not
> > 6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
> > 7. Report other relevant file, process and user information associated with each interception
> > 8. Report file pathnames to userspace (relative to process root, current working directory)
> > 9. Mark a processes as exempt from on access scanning
> > 10. Exclude sub-trees from scanning based on filesystem (exclude procfs, sysfs, devfs)
> > 11. Exclude sub-trees from scanning based on filesystem path
> > 12. Include only certain sub-trees from scanning based on filesystem path
> > 13. Register more than one userspace client in which case behavior is restrictive
>
> I don't see anything in the list above that make this a requirement that
> the code to do this be placed within the kernel.
>
> What is wrong with doing it in glibc or some other system-wide library
> (LD_PRELOAD hooks, etc.)?
It may be possible to do in glibc, LD_PRELOAD doesn't exactly work for
suid binaries
>
> > 1., 2. Basic interception
> > -------------------------
> > Core requirement is to intercept access to files and prevent it if
> > malicious content is detected. This is done on open, not on read. It
> > may be possible to do read time checking with minimal performance impact
> > although not currently implemented. This means that the following race
> > is possible
> >
> > Process1 Process2
> > - open file RD
> > - open file WR
> > - write virus data (1)
> > - read virus data
>
> Wonderful, we are going to implement a solution that is known to not
> work, with a trivial way around it?
>
> Sorry, that's not going to fly.
The model only makes claims about open and I want to be forthright with
its shortcomings. It sounds rather unreasonable to think that every
time I want to read one bite from a file which is being concurrently
written by another process some virus scanner should have to reread and
validate the entire file. I think as some point we have to accept the
fact that there is no feasible perfect solution (no you can't do write
time checking since circumventing that is as simple as splitting your
bad bits into two writes...)
>
> > *note that any open after (1) will get properly vetted. At this time
> > the likely hood of this being a problem vs the performance impact of
> > scanning on read and the increased complexity of the code means this is
> > left out. This should not be a problem for local executables as writes
> > to files opened to be run typically return ETXTBSY.
>
> Are you sure about this?
I'm willing to say that opens after (1) are going to be validated. I am
not certain that all executables opened for write while they are being
executed return ETXTBSY (I do know it happens at least sometimes) so I'm
willing to drop that idea.
> > One of the most important filters in the evaluation chain implements an
> > interface through which an userspace process can register and receive
> > vetting requests. Userspace process opens a misc character device to
> > express its interest and then receives binary structures from that
> > device describing basic interception information. After file contents
> > have been scanned a vetting response is sent by writing a different
> > binary structure back to the device and the intercepted process
> > continues its execution. These are not done over network sockets and no
> > endian conversions are done. The client and the kernel must have the
> > same endian configuration.
>
> How about the same 64/32bit requirement? Your implementation is
> incorrect otherwise.
I'll definitely go back and look, but I think I use bit lengths for
everything in the communication channel so its only endian issues to
worry about.
> (hint, your current patch is also wrong in this area, you should fix
> that up...)
> And a binary structure? Ick, are you trying to make it hard for future
> expansions and such?
As long as the requirement that the first 32 bits be a version it might
make ugly code but any future expansions are easy to deal with. Read
from userspace, get the first 32 bits, cast the read from userspace to
the correct structure. What would you suggest?
>
> And why not netlink/network socket? Why a character device? You are
> already using securityfs, why not use a file node in there?
Opps, old description. I do just use an inode in securityfs, not a misc
character device. I'm not clear what netlink would buy here. I might
be able to make my async close vetting code a little cleaner, but it
would make other things more complex (like registration and actually
having to track userspace clients)
>
> > 6. Direct access to file content
> > --------------------------------
> > When an userspace daemon receives a vetting request, it also receives a
> > new RO file descriptor which provides direct access to the inode in
> > question. This is to enable access to the file regardless of it
> > accessibility from the scanner environment (consider process namespaces,
> > chroot's, NFS). The userspace client is responsible for closing this
> > file when it is finished scanning.
>
> Is this secondary file handle properly checked for the security issues
> involved with such a thing? What happens if the userspace client does
> not close the file handle?
I'm not sure the security issues that you are refering too, do you mean
do we make LSM checks and regular DAC checks for the userspace client on
the file in question? yes.
The userspace client is forced to respond to all fd's it is handed. If
userspace decided to respond to a request but never close the file the
client will eventually run out of fds and I really should make sure I
have decent error handling for that case. No real damage done aside
from extra references outstanding until the client program dies. Much
the same way as any program that calls open on files it doesn't ever
close....
>
> > 7. Other reporting
> > ------------------
> > Along with the fd being installed in the scanning process the process
> > gets a binary structure of data including:
>
> What's with the love of binary structures? :)
Its only the one structure (ok and the response)
include/linux/talpa.h
struct talpa_packet_client
struct talpa_packet_kernel
> > + uint32_t version;
> > + uint32_t type;
> > + int32_t fd;
> > + uint32_t operation;
> > + uint32_t flags;
> > + uint32_t mode;
> > + uint32_t uid;
> > + uint32_t gid;
> > + uint32_t tgid;
> > + uint32_t pid;
>
> What happens when the world moves to 128bit or 64bit uids? (yes, I've
> seen proposals for such a thing...)
The same things that happens to every other subsystem that uses uint32_t
to describe uid (like audit?) It either gets truncated massive main
ensues...
> Why would userspace care about these meta-file things, what does it want
> with them?
Honstely? I don't know. Maybe someone with access to the black magic
source code will stand up and say if most of this metadata is important
and if so how.
>
> > 8. Path name reporting
> > ----------------------
> > When a malicious content is detected in a file it is important to be
> > able to report its location so the user or system administrator can take
> > appropriate actions.
> >
> > This is implemented in a amazingly simple way which will hopefully avoid
> > the controversy of some other solutions. Path name is only needed for
> > reporting purposes and it is obtained by reading the symlink of the
> > given file descriptor in /proc. Its as simple as userspace calling:
> >
> > snprintf(link, sizeof(link), "/proc/self/fd/%d", details.fd);
> > ret = readlink(link, buf, sizeof(buf)-1);
>
> Cute hack. What's to keep it from racing with the fd changing from the
> original program?
Not sure what you mean here. On sys_open the original program is
blocking until the userspace client answers allow or deny. Both the
original program fd and the fd that magically appeared in the client
point to the same dentry. Names may move around but its going to be the
same 'name' for both of them. I don't see a race here....
>
> > 9. Process exclusion
> > --------------------
> > Sometimes it is necessary to exclude certain processes from being
> > intercepted. For example it might be a userspace root kit scanner which
> > would not be able to find root kits if access to them was blocked by the
> > on-access scanner.
> >
> > To facilitate that we have created a special file a process can open and
> > register itself as excluded. A flag is then put into its kernel
> > structure (task_struct) which makes it excluded from scanning.
> >
> > This implementation is very simple and provides greatest performance. In
> > the proposed implementation access to the exclusion device is controlled
> > though permissions on the device node which are not sufficient. An LSM
> > call will need to be made for this type or access in a later patch.
>
> Heh, so if you want to write a "virus" for Linux, just implement this
> flag. What's to keep a "rogue" program from telling the kernel that all
> programs on the system are to be excluded?
Processes can only get this flag one of 2 ways.
1) register as a client to make access decisions
2) echo 1 into the magic file to enable the flag for themselves
A process can only set this flag on itself and having this flag only
means that your opens and closes will not be scanned. And exculded
program could write a virus and it would not be caught on close, but it
would be caught on the next open.
> > 10. Filesystem exclusions
> > -------------------------
> > One pretty important optimization is not to scan things like /proc, /sys
> > or similar. Basically all filesystems where user can not store
> > arbitrary, potentially malicious, content could and should be excluded
> > from scanning.
>
> Why, does scanning these files take extra time? Just curious.
Perf win, why bothering looking for malware in /proc when it can't
exist? It doesn't take longer it just takes time having to do
userspace -> kernel -> userspace -> kernel -> userspace
just to cat /proc/mounts, all of this could probably be alliviated if we
cached access on non block backed files but then we have to come up with
a way to exclude only nfs/cifs. I'd rather list the FSs that don't need
scanning every time than those that do....
>
> > 11. Path exclusions
> > -------------------
> > The need for exclusions can be demonstrated with an example of a MySQL
> > server. It's data files are frequently modified which means they would
> > need to be constantly rescanned which is very bad for performance. Also,
> > it is most often not even possible to reasonably scan them. Therefore
> > the best solution is not to scan its database store which can simply be
> > implemented by excluding the store subdirectory.
> >
> > It is a relatively simple implementation which allows run-time
> > configuration of a list of sub directories or files to exclude.
> > Exclusion paths are relative to each process root. So for example if we
> > want to exclude /var/lib/mysql/ and we have a mysql running in a chroot
> > where from the outside that directory actually lives
> > in /chroot/mysql/var/lib/mysql, /var/lib/mysql should actually be added
> > to the exclusion list.
> >
> > This is also not included in the initial patch set but will be coming
> > shortly after.
>
> Again, what's to keep all files to be marked as excluded?
You have to be root and I'll probably add an LSM hook?
> > Closing remarks
> > ---------------
> > Although some may argue some of the filters are not necessary or may
> > better be implemented in userspace, we think it is better to have them
> > in kernel primarily for performance reasons.
>
> Why? What numbers do you have that say the kernel is faster in
> implementing this? This is the first mention of such a requirement, we
> need to see real data to back it up please.
In kernel caching is clearly a huge perf win. I couldn't even measure a
change in kernel build time when I didn't run a userspace client. If
anyone can explain a way to get race free in kernel caching and out of
kernel redirection and scanning I'd love it :)
I'll post numbers on perf in the next day or 2.
> > Secondly, it is all simple code not introducing much baggage or risk
> > into the kernel itself.
>
> I disagree, see above.
On Mon, 2008-08-04 at 20:32 -0400, Eric Paris wrote:
> The model only makes claims about open and I want to be forthright with
> its shortcomings. It sounds rather unreasonable to think that every
> time I want to read one bite from a file which is being concurrently
> written by another process some virus scanner should have to reread and
> validate the entire file. I think as some point we have to accept the
Sorry, can't help but laugh at myself for typing "bite".
-Eric
On Mon, 2008-08-04 at 20:26 -0400, Christoph Hellwig wrote:
> On Mon, Aug 04, 2008 at 03:32:49PM -0700, Greg KH wrote:
> > > 1. Intercept file opens (exec also) for vetting (block until
> > > decision is made) and allow some userspace black magic to make
> > > decisions.
>
> NACK, this kind of policy should be done in kernelspace.
What? You want to write and in kernel scanner for Window viruses?
>
> > > 2. Intercept file closes for scanning post access
>
> Not even possible for mmap, and dumb otherwise, NACK.
I don't know when files get closed and can't preemptively scan to make
sure it is clean for the next open? Any writes are going to invalidate
the allow/deny cache....
>
> > > 6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
>
> Explain?
The data connected with the file being opened must as reasonably as
possible be the data the 'scanner' looks at. Some foolish early
discussion wanted to do simplistic things like pass a pathname to a
scanner and have it call open on that path name. I'm willing to
entertain any other method of making the scanner look at the data the
process is about to get.
>
> > > 9. Mark a processes as exempt from on access scanning
>
> Nack, this completely defeats the purpose.
What? it allows a process to open a file that contains malware, how is
that horrible. If a process says "I want to see malware" it can then
see malware. Doesn't in any way affect other processes or the system
security as a whole. If 'bad' data gets into a file its going to get
blocked from everything that doesn't actively choose to see it.
>
> > > 11. Exclude sub-trees from scanning based on filesystem path
> > > 12. Include only certain sub-trees from scanning based on filesystem path
>
> Nack, please no pathname based idiocies.
Go read the long explainations, I already rules out path based
inclusions. I'm leaving exclusions up for grabs since I don't see it
weakening the security model.
On Mon, Aug 04, 2008 at 08:32:54PM -0400, Eric Paris wrote:
> On Mon, 2008-08-04 at 15:32 -0700, Greg KH wrote:
> > On Mon, Aug 04, 2008 at 05:00:16PM -0400, Eric Paris wrote:
> > > Collated requirements
> > > +++++++++++++++++++++
> > > 1. Intercept file opens (exec also) for vetting (block until
> > > decision is made) and allow some userspace black magic to make
> > > decisions.
> > > 2. Intercept file closes for scanning post access
> > > 3. Cache scan results so the same file is not scanned on each and every access
> > > 4. Ability to flush the cache and cause all files to be re-scanned when accessed
> > > 5. Define which filesystems are cacheable and which are not
> > > 6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
> > > 7. Report other relevant file, process and user information associated with each interception
> > > 8. Report file pathnames to userspace (relative to process root, current working directory)
> > > 9. Mark a processes as exempt from on access scanning
> > > 10. Exclude sub-trees from scanning based on filesystem (exclude procfs, sysfs, devfs)
> > > 11. Exclude sub-trees from scanning based on filesystem path
> > > 12. Include only certain sub-trees from scanning based on filesystem path
> > > 13. Register more than one userspace client in which case behavior is restrictive
> >
> > I don't see anything in the list above that make this a requirement that
> > the code to do this be placed within the kernel.
> >
> > What is wrong with doing it in glibc or some other system-wide library
> > (LD_PRELOAD hooks, etc.)?
>
> It may be possible to do in glibc, LD_PRELOAD doesn't exactly work for
> suid binaries
Are suid binaries something that you feel is necessary to scan from?
I don't see it on the list above :)
> > > 1., 2. Basic interception
> > > -------------------------
> > > Core requirement is to intercept access to files and prevent it if
> > > malicious content is detected. This is done on open, not on read. It
> > > may be possible to do read time checking with minimal performance impact
> > > although not currently implemented. This means that the following race
> > > is possible
> > >
> > > Process1 Process2
> > > - open file RD
> > > - open file WR
> > > - write virus data (1)
> > > - read virus data
> >
> > Wonderful, we are going to implement a solution that is known to not
> > work, with a trivial way around it?
> >
> > Sorry, that's not going to fly.
>
> The model only makes claims about open and I want to be forthright with
> its shortcomings. It sounds rather unreasonable to think that every
> time I want to read one bite from a file which is being concurrently
> written by another process some virus scanner should have to reread and
> validate the entire file. I think as some point we have to accept the
> fact that there is no feasible perfect solution (no you can't do write
> time checking since circumventing that is as simple as splitting your
> bad bits into two writes...)
So, if this isn't really going to protect anything, how can anyone
justify adding it to the kernel? I sure would not allow that.
> > > One of the most important filters in the evaluation chain implements an
> > > interface through which an userspace process can register and receive
> > > vetting requests. Userspace process opens a misc character device to
> > > express its interest and then receives binary structures from that
> > > device describing basic interception information. After file contents
> > > have been scanned a vetting response is sent by writing a different
> > > binary structure back to the device and the intercepted process
> > > continues its execution. These are not done over network sockets and no
> > > endian conversions are done. The client and the kernel must have the
> > > same endian configuration.
> >
> > How about the same 64/32bit requirement? Your implementation is
> > incorrect otherwise.
>
> I'll definitely go back and look, but I think I use bit lengths for
> everything in the communication channel so its only endian issues to
> worry about.
No, your field definitions are incorrect.
You must use __u8 and friends for variables that cross the
userspace/kernel boundry. None of the uint_* crap :)
> > (hint, your current patch is also wrong in this area, you should fix
> > that up...)
>
> > And a binary structure? Ick, are you trying to make it hard for future
> > expansions and such?
>
> As long as the requirement that the first 32 bits be a version it might
> make ugly code but any future expansions are easy to deal with. Read
> from userspace, get the first 32 bits, cast the read from userspace to
> the correct structure. What would you suggest?
Not doing this in the kernel at all.
Seriously.
I mean it.
Oh, and after that, not using a binary interface, have we not learned
from the ioctl mess? I sure thought we had...
> > And why not netlink/network socket? Why a character device? You are
> > already using securityfs, why not use a file node in there?
>
> Opps, old description. I do just use an inode in securityfs, not a misc
> character device. I'm not clear what netlink would buy here. I might
> be able to make my async close vetting code a little cleaner, but it
> would make other things more complex (like registration and actually
> having to track userspace clients)
Why would the kernel have to worry about that?
> > > 6. Direct access to file content
> > > --------------------------------
> > > When an userspace daemon receives a vetting request, it also receives a
> > > new RO file descriptor which provides direct access to the inode in
> > > question. This is to enable access to the file regardless of it
> > > accessibility from the scanner environment (consider process namespaces,
> > > chroot's, NFS). The userspace client is responsible for closing this
> > > file when it is finished scanning.
> >
> > Is this secondary file handle properly checked for the security issues
> > involved with such a thing? What happens if the userspace client does
> > not close the file handle?
>
> I'm not sure the security issues that you are refering too, do you mean
> do we make LSM checks and regular DAC checks for the userspace client on
> the file in question? yes.
Yes, that is what I was referring to.
> > > + uint32_t version;
> > > + uint32_t type;
> > > + int32_t fd;
> > > + uint32_t operation;
> > > + uint32_t flags;
> > > + uint32_t mode;
> > > + uint32_t uid;
> > > + uint32_t gid;
> > > + uint32_t tgid;
> > > + uint32_t pid;
> >
> > What happens when the world moves to 128bit or 64bit uids? (yes, I've
> > seen proposals for such a thing...)
>
> The same things that happens to every other subsystem that uses uint32_t
> to describe uid (like audit?) It either gets truncated massive main
> ensues...
audit passed the value in a binary structure from the kernel to
userspace? Really? Ick.
> > Why would userspace care about these meta-file things, what does it want
> > with them?
>
> Honstely? I don't know. Maybe someone with access to the black magic
> source code will stand up and say if most of this metadata is important
> and if so how.
Don't add things that are not needed, _everything_ must be justified.
> > > 8. Path name reporting
> > > ----------------------
> > > When a malicious content is detected in a file it is important to be
> > > able to report its location so the user or system administrator can take
> > > appropriate actions.
> > >
> > > This is implemented in a amazingly simple way which will hopefully avoid
> > > the controversy of some other solutions. Path name is only needed for
> > > reporting purposes and it is obtained by reading the symlink of the
> > > given file descriptor in /proc. Its as simple as userspace calling:
> > >
> > > snprintf(link, sizeof(link), "/proc/self/fd/%d", details.fd);
> > > ret = readlink(link, buf, sizeof(buf)-1);
> >
> > Cute hack. What's to keep it from racing with the fd changing from the
> > original program?
>
> Not sure what you mean here. On sys_open the original program is
> blocking until the userspace client answers allow or deny. Both the
> original program fd and the fd that magically appeared in the client
> point to the same dentry. Names may move around but its going to be the
> same 'name' for both of them. I don't see a race here....
Oh, forgot about the fact that the code blocks. That's probably a race
in itself :)
> > Heh, so if you want to write a "virus" for Linux, just implement this
> > flag. What's to keep a "rogue" program from telling the kernel that all
> > programs on the system are to be excluded?
>
> Processes can only get this flag one of 2 ways.
>
> 1) register as a client to make access decisions
How do you do that?
> 2) echo 1 into the magic file to enable the flag for themselves
Simple enough :)
> > > 10. Filesystem exclusions
> > > -------------------------
> > > One pretty important optimization is not to scan things like /proc, /sys
> > > or similar. Basically all filesystems where user can not store
> > > arbitrary, potentially malicious, content could and should be excluded
> > > from scanning.
> >
> > Why, does scanning these files take extra time? Just curious.
>
> Perf win, why bothering looking for malware in /proc when it can't
> exist? It doesn't take longer it just takes time having to do
>
> userspace -> kernel -> userspace -> kernel -> userspace
>
> just to cat /proc/mounts, all of this could probably be alliviated if we
> cached access on non block backed files but then we have to come up with
> a way to exclude only nfs/cifs. I'd rather list the FSs that don't need
> scanning every time than those that do....
How long does this whole process take? Seriously is it worth the added
kernel code for something that is not measurable?
> > > 11. Path exclusions
> > > -------------------
> > > The need for exclusions can be demonstrated with an example of a MySQL
> > > server. It's data files are frequently modified which means they would
> > > need to be constantly rescanned which is very bad for performance. Also,
> > > it is most often not even possible to reasonably scan them. Therefore
> > > the best solution is not to scan its database store which can simply be
> > > implemented by excluding the store subdirectory.
> > >
> > > It is a relatively simple implementation which allows run-time
> > > configuration of a list of sub directories or files to exclude.
> > > Exclusion paths are relative to each process root. So for example if we
> > > want to exclude /var/lib/mysql/ and we have a mysql running in a chroot
> > > where from the outside that directory actually lives
> > > in /chroot/mysql/var/lib/mysql, /var/lib/mysql should actually be added
> > > to the exclusion list.
> > >
> > > This is also not included in the initial patch set but will be coming
> > > shortly after.
> >
> > Again, what's to keep all files to be marked as excluded?
>
> You have to be root and I'll probably add an LSM hook?
Why an LSM hook? You aren't an LSM.
> > > Closing remarks
> > > ---------------
> > > Although some may argue some of the filters are not necessary or may
> > > better be implemented in userspace, we think it is better to have them
> > > in kernel primarily for performance reasons.
> >
> > Why? What numbers do you have that say the kernel is faster in
> > implementing this? This is the first mention of such a requirement, we
> > need to see real data to back it up please.
>
> In kernel caching is clearly a huge perf win.
Why? If the cache is also in userspace, it should be the same, right?
> I couldn't even measure a change in kernel build time when I didn't
> run a userspace client.
And when you did run a foolish userspace client?
If you did it all in userspace, if the userspace code isn't being
called, the kernel build time should be the same as well :)
> If anyone can explain a way to get race free in kernel caching and out
> of kernel redirection and scanning I'd love it :)
Again, do it all in userspace (caching, and scanning). I still really
don't see the need to do this in the kernel becides it being "the way
people have always done it."
thanks,
greg k-h
On Mon, Aug 04, 2008 at 08:47:04PM -0400, Eric Paris wrote:
> On Mon, 2008-08-04 at 20:26 -0400, Christoph Hellwig wrote:
> > On Mon, Aug 04, 2008 at 03:32:49PM -0700, Greg KH wrote:
> > > > 1. Intercept file opens (exec also) for vetting (block until
> > > > decision is made) and allow some userspace black magic to make
> > > > decisions.
> >
> > NACK, this kind of policy should be done in kernelspace.
>
> What? You want to write and in kernel scanner for Window viruses?
No, I want a sane security policy in kernelsapce that doesn't look
at the content because doing security by content properly is equivalent
to solving the halting problem. I couldn't give a rats a** about
windows viruses as they can't actually cause any harm on a Linux
machine.
> >
> > > > 6. Scan files directly not relying on path. Avoid races and problems with namespaces, chroot, containers, etc.
> >
> > Explain?
>
> The data connected with the file being opened must as reasonably as
> possible be the data the 'scanner' looks at. Some foolish early
> discussion wanted to do simplistic things like pass a pathname to a
> scanner and have it call open on that path name. I'm willing to
> entertain any other method of making the scanner look at the data the
> process is about to get.
Well, data can change all the time, as can the path name. This whole
content scanning thing doesn't make any sense at all.
> > > > 9. Mark a processes as exempt from on access scanning
> >
> > Nack, this completely defeats the purpose.
>
> What? it allows a process to open a file that contains malware, how is
> that horrible. If a process says "I want to see malware" it can then
> see malware. Doesn't in any way affect other processes or the system
> security as a whole. If 'bad' data gets into a file its going to get
> blocked from everything that doesn't actively choose to see it.
So make this opt-in and in userspace. Just LD_PRELOAD some monster lib
doing all the horrible things you propose and use it wherever you want.
Eric Paris wrote:
> Please contact me privately or (preferably the list) for questions,
> comments, discussions, flames, names, or anything. I'll do complete
> rewrites of the patches if someone tells me how they don't meet their
> needs or how they can be done better. I'm here to try to bridge the
> needs (and wants) of the anti-malware vendors with the technical
> realities of the kernel. So everyone feel free to throw in your two
> cents and I'll try to reconcile it all. These 5 patches are part 1.
> They give us a working able solution.
>
> >From my point of view patches forthcoming and mentioned below should
> help with performance for those who actually have userspace scanners but
> also could presents be implemented using this framework.
>
>
The LSM list (CCed) should be included in this discussion.
> Background
> ...
If we had stackable LSMs then the required functionality could simply be
built into the LSM interface. Then the anti-malware would simply stack
itself with other LSMs. In my opinion this is a perfect example for the
argument of stackable LSMs. So far we mainly have LSMs which provide an
extra access control mechanism (in addition to DAC). IMHO, Ideally DAC
could be another stackable LSM (enabled by default). Other security
schemes such as intrusion detection, firewalls/netfilter, anti-malware,
and application restrictions (sandboxes such as jails or finer grained
restrictions such as AppArmor) could all register LSMs onto the stack.
Additional infrastructure would be necessary. Permissible security
remains a item of contention. Perhaps I am naive but I think most LSMs
could work based on accept/reject. Where every LSM must accept an action
in order for it to be carried out.
MHO,
Cliffe.
Casey Schaufler wrote:
> Eric Paris wrote:
>> Please contact me privately or (preferably the list) for questions,
>> comments, discussions, flames, names, or anything. I'll do complete
>> rewrites of the patches if someone tells me how they don't meet their
>> needs or how they can be done better. I'm here to try to bridge the
>> needs (and wants) of the anti-malware vendors with the technical
>> realities of the kernel. So everyone feel free to throw in your two
>> cents and I'll try to reconcile it all. These 5 patches are part 1.
>> They give us a working able solution.
>>
>> >From my point of view patches forthcoming and mentioned below should
>> help with performance for those who actually have userspace scanners but
>> also could presents be implemented using this framework.
>>
>>
>
> The LSM list (CCed) should be included in this discussion.
>> Background
>> ...
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe
> linux-security-module" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
>
Cliffe wrote:
> If we had stackable LSMs then the required functionality could simply
> be built into the LSM interface. Then the anti-malware would simply
> stack itself with other LSMs. In my opinion this is a perfect example
> for the argument of stackable LSMs.
No argument from me.
> So far we mainly have LSMs which provide an extra access control
> mechanism (in addition to DAC).
Yes. This is the design center for the LSM.
> IMHO, Ideally DAC could be another stackable LSM (enabled by default).
Yup. Search the archives for "authoritative hooks".
> Other security schemes such as intrusion detection,
> firewalls/netfilter, anti-malware, and application restrictions
> (sandboxes such as jails or finer grained restrictions such as
> AppArmor) could all register LSMs onto the stack.
Stacking is easy for files. It's a real pain in the backside for UDP
packets.
> Additional infrastructure would be necessary. Permissible security
> remains a item of contention. Perhaps I am naive but I think most LSMs
> could work based on accept/reject. Where every LSM must accept an
> action in order for it to be carried out.
Please propose patches.
> MHO,
Oh, humility isn't all it's cracked up to be. Show us all up and
write the code. I'm serious, I don't think there's anyone here who
would object to a really good stacking scheme.
On Tue, Aug 05, 2008 at 11:01:59AM +0800, Cliffe wrote:
> If we had stackable LSMs then the required functionality could simply be
> built into the LSM interface. Then the anti-malware would simply stack
> itself with other LSMs.
Could it really? How would such an interface work?
Remember, the big issue here isn't the kernel "hooks", but the fact that
a lot of people are yet to be convinced that something like this needs
to be within the kernel itself.
Perhaps we should dig up the proposals for the filesystem-notify type
patches, something like that might be all the majority of the virus
people need, as they want to just scan things for Windows viruses, not
Linux ones, and to do so "lazily" might be sufficient.
thanks,
greg k-h
Casey Schaufler wrote:
> Cliffe wrote:
>> Additional infrastructure would be necessary. Permissible security
>> remains a item of contention. Perhaps I am naive but I think most
>> LSMs could work based on accept/reject. Where every LSM must accept
>> an action in order for it to be carried out.
>
> Please propose patches.
>
>> MHO,
>
> Oh, humility isn't all it's cracked up to be. Show us all up and
> write the code. I'm serious, I don't think there's anyone here who
> would object to a really good stacking scheme.
Ah, I'd love to give it a shot but I have this little PhD thesis and
related LSM and tools I need to work on. I just thought I'd put my 5c in.
Cliffe.
Greg KH wrote:
> On Tue, Aug 05, 2008 at 11:01:59AM +0800, Cliffe wrote:
>
>> If we had stackable LSMs then the required functionality could simply be
>> built into the LSM interface. Then the anti-malware would simply stack
>> itself with other LSMs.
>>
>
> Could it really? How would such an interface work?
>
> Remember, the big issue here isn't the kernel "hooks", but the fact that
> a lot of people are yet to be convinced that something like this needs
> to be within the kernel itself.
>
> Perhaps we should dig up the proposals for the filesystem-notify type
> patches, something like that might be all the majority of the virus
> people need, as they want to just scan things for Windows viruses, not
> Linux ones, and to do so "lazily" might be sufficient.
>
From memory Dazuko is a LSM designed for this purpose, although it
cannot be used on systems running other LSMs due to the lack of stacking
support. It is used by a bunch of anti-malware programs including ClamAV
and AVG.
On Mon, Aug 4, 2008 at 8:54 PM, Christoph Hellwig <[email protected]> wrote:
> On Mon, Aug 04, 2008 at 08:47:04PM -0400, Eric Paris wrote:
>> On Mon, 2008-08-04 at 20:26 -0400, Christoph Hellwig wrote:
>> > NACK, this kind of policy should be done in kernelspace.
>>
>> What? You want to write and in kernel scanner for Window viruses?
>
> No, I want a sane security policy in kernelsapce that doesn't look
> at the content because doing security by content properly is equivalent
> to solving the halting problem. I couldn't give a rats a** about
> windows viruses as they can't actually cause any harm on a Linux
> machine.
Much better solution:
Use SELinux or another similar Mandatory Access Control labeling
system. Mark some things as "trusted" or "privileged" or whatever
your particular labeling methodology requires. Mark other things as
"untrusted", "unprivileged", "internet file", etc. Disallow most
interaction between "trusted" and "untrusted" things. Configure your
userspace virus-scanner which is allowed to read "untrusted" files and
create "semi-trusted" files in a particular directory, where they can
then be picked up by "trusted" programs.
Problem solved. Untrusted and possibly-compromised files can't be
executed, or even if they could be they can't do anything
interesting/harmful. In order to execute some junk you just
downloaded from the internet you have to click "Yes I accept the
security risk" and run it through whatever virus-scanner you want.
Then you copy it from the virus-scanner output directory into
somewhere else and run it.
You can do the same thing with software updates downloaded from the
internet, just replace "trusted" with "installed package",
"semi-trusted" with "valid package", "untrusted" with "just-downloaded
package file", and "virus scanner" with "package signature
verification tool".
I could easily come up with a bunch more examples if you really care that much.
Cheers,
Kyle Moffett
Greg KH wrote:
>
>
> I don't see anything in the list above that make this a requirement that
> the code to do this be placed within the kernel.
>
> What is wrong with doing it in glibc or some other system-wide library
> (LD_PRELOAD hooks, etc.)?
>
A linux virus would trivially get around that by doing its own syscalls
instead of using glibc. (It might still link dynamically to glibc so
you don't get suspicious, but won' actually use it when doing bad stuff.)
Helge Hafting
> > It may be possible to do in glibc, LD_PRELOAD doesn't exactly work for
> > suid binaries
>
> Are suid binaries something that you feel is necessary to scan from?
>
> I don't see it on the list above :)
Doesn't work very well really does it - ld.so loads files too and can be
attacked.
> Again, do it all in userspace (caching, and scanning). I still really
> don't see the need to do this in the kernel becides it being "the way
> people have always done it."
We don't have notifiers for file segment changes that are scalable that
far. We don't have mechanisms for delaying an open for a scan.
There are various bits of this you can't just do in user space. This
doesn't just hurt virus scanners btw its also bad news for indexing disk
content.
> No, I want a sane security policy in kernelsapce that doesn't look
> at the content because doing security by content properly is equivalent
> to solving the halting problem. I couldn't give a rats a** about
> windows viruses as they can't actually cause any harm on a Linux
> machine.
Go on then.. post patches.
I think your are being incredibly naïve. Our memory debugging is not 100%
solid but work by heuristic. Our lock analysis doesn't solve the halting
problem but is extremely useful and so on.
> Well, data can change all the time, as can the path name. This whole
> content scanning thing doesn't make any sense at all.
Phone numbers change all the time, shall we burn all the phone books ?
> So make this opt-in and in userspace. Just LD_PRELOAD some monster lib
> doing all the horrible things you propose and use it wherever you want.
Rather tricky as the needed hooks don't exist and you need to get ahead
of even ld.so as well as protect suid apps. You've clearly not even
thought about the problem space before sounding off because there are
more elegant ways of tackling it even if you want to push it at
userspace, and ones that aren't implausible like LD_PRELOAD.
If you'd applied even 30 seconds thought you would at least be pointing
people at FUSE or a stacking fs.
Alan
On Tue, Aug 5, 2008 at 1:58 PM, Cliffe <[email protected]> wrote:
> Greg KH wrote:
>>
>> On Tue, Aug 05, 2008 at 11:01:59AM +0800, Cliffe wrote:
>>
>>>
>>> If we had stackable LSMs then the required functionality could simply be
>>> built into the LSM interface. Then the anti-malware would simply stack
>>> itself with other LSMs.
>>>
>>
>> Could it really? How would such an interface work?
>>
>> Remember, the big issue here isn't the kernel "hooks", but the fact that
>> a lot of people are yet to be convinced that something like this needs
>> to be within the kernel itself.
>>
>> Perhaps we should dig up the proposals for the filesystem-notify type
>> patches, something like that might be all the majority of the virus
>> people need, as they want to just scan things for Windows viruses, not
>> Linux ones, and to do so "lazily" might be sufficient.
>>
>
> From memory Dazuko is a LSM designed for this purpose, although it cannot be
> used on systems running other LSMs due to the lack of stacking support. It
> is used by a bunch of anti-malware programs including ClamAV and AVG.
Sorry to say stacking support really does need to be taken with a grain of salt.
Deeper the stack worse the issue of creating system lag.
Anti-malware stacking into LSM adding there own hooks need to be
looked at as more of a problem than a solution.
Lets take a simple example of Dazuko. It catches all filesystem
accesses and can alter them. Side effect all protected documents by
the LSM could end up going threw Dazuko so if a flaw is in the anti
virus that can be exploited straight up complete system exposed.
So really LSM stacking is not the correct solution. More correct
solution would be list out what features malware and virus scanners
need and extend LSM interfaces so the main LSM could control where the
antivirus/malware scanner was accessing. Yes there are times you
don't want virus scanners or malware scanners seeing everything.
Containers is another issue around stacking. Containers being added
to Linux are providing more and more controls. Currently LSM's are
independent to this. People have responded that it is light weight
virtualisation so dones not need the need to run a different LSM
secuirty construct inside a container. Its that it is
virtualistation is why it kinda need to.
Now any stacking design of LSM's have to take into account the needs
of containers.
Of course prefered is no stacking for containers and some form of
cross LSM mapping. Like prime kernel LSM being selinux, OS contained
being smack so smack secuirty files coveted to selinux with host OS's
limitations applied so container maintains its secuirty.
Same could be applied to chroot's. This is a common over site of the
current LSM model. chroot or containers running different
distrobution inside there is no preformated way of picking up the
secuirty. Note could even be the same distrubtion different version
just at one point of time they changed secuirty systems. So user is
running a chroot or a container to run a old application that is
needed. But now its running without secuirty. This is could become
a major issue in time.
So yes needs of stacking have some major headaches. Its not just a
malware issue.
Peter Dolding
> Remember, the big issue here isn't the kernel "hooks", but the fact that
> a lot of people are yet to be convinced that something like this needs
> to be within the kernel itself.
Mostly the same people who said that about LSM I note 8)
> Perhaps we should dig up the proposals for the filesystem-notify type
> patches, something like that might be all the majority of the virus
> people need, as they want to just scan things for Windows viruses, not
> Linux ones, and to do so "lazily" might be sufficient.
The key difference between a lazy scan and an active intervention is tiny
- the ability to block in the security decision to open a file.
Once you have that bit you have the ability to hand the file handle up to
a daemon to chew on and return a status. The same co-incidentally gives
you the hooks for doing various kinds of HSM as you can block an open
while you retrieve the archived content from wherever it was warehoused.
Alan
> Much better solution:
And one which was found lacking about 1950...
> Problem solved. Untrusted and possibly-compromised files can't be
> executed, or even if they could be they can't do anything
Two things
- Scripts
- Attacks based on compromising a live binary
You can use SELinux to control what is executed and it is a very
effective management control technique. However it doesn't control
javascript in web pages, exploits that popen perl and chat to it and so
on...
On Tue, 2008-08-05 at 12:31 +0100, Alan Cox wrote:
> > No, I want a sane security policy in kernelsapce that doesn't look
> > at the content because doing security by content properly is equivalent
> > to solving the halting problem. I couldn't give a rats a** about
> > windows viruses as they can't actually cause any harm on a Linux
> > machine.
>
> Go on then.. post patches.
>
> I think your are being incredibly naïve. Our memory debugging is not 100%
> solid but work by heuristic. Our lock analysis doesn't solve the halting
> problem but is extremely useful and so on.
Sure, but what's the point of partial security? It seems to me you're
not secure until you're fully secure, so why bother with almost?
Every bug caught with the debuggers is one caught, yay :-)
Every virus let through is a pawned system, aww :-(
> Sure, but what's the point of partial security? It seems to me you're
> not secure until you're fully secure, so why bother with almost?
>
> Every bug caught with the debuggers is one caught, yay :-)
>
> Every virus let through is a pawned system, aww :-(
Every bug missed is a crash. Security is not an absolute. If I burst into
your office with a machine gun and ask you for the password your security
system probably fails ...
Security is a risk/reward model.
Alan
I share the concern here. The idea that a piece of malware can exclude
itself seems nasty to me. I am not an expert on writing malware, but it
intuitively seems to me to be a huge opportunity for creativity. The
argument that it's ok because anything that the malware writes will
eventually be scanned anyway does not reassure me.
Also... I was one of the people who brought up the idea of a process
exclusion when the requirements list was being developed. I intended it
as a way that an AV application could exclude specific OTHER processes
by name (as selected by the AV user) -- not as a way that a process
would exclude itself. I don't think that the implementation here
reflects this goal, which still seems to me to be a requirement.
Jon Press
CA/HCL Internet Security Business Unit
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Eric Paris
Sent: Monday, August 04, 2008 8:33 PM
To: Greg KH
Cc: [email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a linux interface
foron access scanning
>
> > 9. Process exclusion
> > --------------------
> > Sometimes it is necessary to exclude certain processes from being
> > intercepted. For example it might be a userspace root kit scanner
which
> > would not be able to find root kits if access to them was blocked by
the
> > on-access scanner.
> >
> > To facilitate that we have created a special file a process can open
and
> > register itself as excluded. A flag is then put into its kernel
> > structure (task_struct) which makes it excluded from scanning.
> >
> > This implementation is very simple and provides greatest
performance. In
> > the proposed implementation access to the exclusion device is
controlled
> > though permissions on the device node which are not sufficient. An
LSM
> > call will need to be made for this type or access in a later patch.
>
> Heh, so if you want to write a "virus" for Linux, just implement this
> flag. What's to keep a "rogue" program from telling the kernel that
all
> programs on the system are to be excluded?
Processes can only get this flag one of 2 ways.
1) register as a client to make access decisions
2) echo 1 into the magic file to enable the flag for themselves
A process can only set this flag on itself and having this flag only
means that your opens and closes will not be scanned. And excluded
program could write a virus and it would not be caught on close, but it
would be caught on the next open.
On Tue, 2008-08-05 at 10:41 -0400, Press, Jonathan wrote:
> I share the concern here. The idea that a piece of malware can exclude
> itself seems nasty to me. I am not an expert on writing malware, but it
> intuitively seems to me to be a huge opportunity for creativity. The
> argument that it's ok because anything that the malware writes will
> eventually be scanned anyway does not reassure me.
You aren't doing write time scanning anyway. This exclusion means that
an 'excluded' process can OPEN things that would normally be called
malware. The model here doesn't talk about adding files with bad
information to the system it talks about stopping that bad information
from being opened and propagated further. Thread exclusions as they are
written in the patch only weaken security to those processes which
actively choose to read malware, it in no way weakens the security of
the system as a whole...
>
> Also... I was one of the people who brought up the idea of a process
> exclusion when the requirements list was being developed. I intended it
> as a way that an AV application could exclude specific OTHER processes
> by name (as selected by the AV user) -- not as a way that a process
> would exclude itself. I don't think that the implementation here
> reflects this goal, which still seems to me to be a requirement.
Wait wit, you'd rather have a 'privileged' process be allowed to exclude
every other process on a system than have a it only be allowed to
exclude itself? and somehow that is safer?
"by name" is right out the window. You are never going to win 'by name'
on anything to do with the kernel :) Maybe you can get me to
eventually buy into 'by pid' or something like that, but setting flags
on other running processes is always going to be racy and scary for me.
Can you show me some code on how to do this cleanly? And why it needs
to be done in kernel?
What is the goal you are trying to achieve? A performance win for the
application in question or is this a security aware application that
needs to be able to access 'sensitive' data?
-Eric
Regarding self-exclusion... We are doing write-time scanning, or more
accurately close-time scanning. We don't block the close, but we do
scan the file afterwards, and if it is infected, we perform any one of
several actions -- such as report to user, rename, quarantine, delete,
etc. I would not want a self-excluded executable to be able to write
something that I don't get to scan.
Regarding exclusion by name, etc... You are right. I am accustomed to
thinking about our current architecture, which does not involve a
published interface. I have to adjust my thinking. Putting such a
feature into the kernel would not be a good idea. We can handle it in
user space.
Jon
-----Original Message-----
From: Eric Paris [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 10:57 AM
To: Press, Jonathan
Cc: Greg KH; [email protected]; [email protected]
Subject: RE: [malware-list] [RFC 0/5] [TALPA] Intro to a linux
interfaceforon access scanning
On Tue, 2008-08-05 at 10:41 -0400, Press, Jonathan wrote:
> I share the concern here. The idea that a piece of malware can
exclude
> itself seems nasty to me. I am not an expert on writing malware, but
it
> intuitively seems to me to be a huge opportunity for creativity. The
> argument that it's ok because anything that the malware writes will
> eventually be scanned anyway does not reassure me.
You aren't doing write time scanning anyway. This exclusion means that
an 'excluded' process can OPEN things that would normally be called
malware. The model here doesn't talk about adding files with bad
information to the system it talks about stopping that bad information
from being opened and propagated further. Thread exclusions as they are
written in the patch only weaken security to those processes which
actively choose to read malware, it in no way weakens the security of
the system as a whole...
>
> Also... I was one of the people who brought up the idea of a process
> exclusion when the requirements list was being developed. I intended
it
> as a way that an AV application could exclude specific OTHER processes
> by name (as selected by the AV user) -- not as a way that a process
> would exclude itself. I don't think that the implementation here
> reflects this goal, which still seems to me to be a requirement.
Wait wit, you'd rather have a 'privileged' process be allowed to exclude
every other process on a system than have a it only be allowed to
exclude itself? and somehow that is safer?
"by name" is right out the window. You are never going to win 'by name'
on anything to do with the kernel :) Maybe you can get me to
eventually buy into 'by pid' or something like that, but setting flags
on other running processes is always going to be racy and scary for me.
Can you show me some code on how to do this cleanly? And why it needs
to be done in kernel?
What is the goal you are trying to achieve? A performance win for the
application in question or is this a security aware application that
needs to be able to access 'sensitive' data?
-Eric
On Tue, Aug 05, 2008 at 12:23:28PM +0100, Alan Cox wrote:
> > > It may be possible to do in glibc, LD_PRELOAD doesn't exactly work for
> > > suid binaries
> >
> > Are suid binaries something that you feel is necessary to scan from?
> >
> > I don't see it on the list above :)
>
> Doesn't work very well really does it - ld.so loads files too and can be
> attacked.
But that's an "attack" vector, which virus scanners are not addressing.
They are going for the "is this file corrupted" type issue. The
"normal" LSM interface is for malware itself, taking control of a
system.
thanks,
greg k-h
On Tue, Aug 05, 2008 at 12:25:03PM +0100, Alan Cox wrote:
> > Again, do it all in userspace (caching, and scanning). I still really
> > don't see the need to do this in the kernel becides it being "the way
> > people have always done it."
>
> We don't have notifiers for file segment changes that are scalable that
> far.
I agree, but if we did, would that help out a lot here? Lots of other
groups of people are needing/asking for something like this and if
someone can finally get it together to post something useful, that might
be a very good thing.
thanks,
greg k-h
On Tue, Aug 05, 2008 at 01:21:01PM +0200, Helge Hafting wrote:
> Greg KH wrote:
>>
>>
>> I don't see anything in the list above that make this a requirement that
>> the code to do this be placed within the kernel.
>>
>> What is wrong with doing it in glibc or some other system-wide library
>> (LD_PRELOAD hooks, etc.)?
>>
>
> A linux virus would trivially get around that by doing its own syscalls
> instead of using glibc. (It might still link dynamically to glibc so
> you don't get suspicious, but won' actually use it when doing bad stuff.)
That's fine, then the file is corrupted. It is when the "normal"
program goes to load the file that we want to block and determine if we
have a problem or not in the data.
virus scanners are not a security model in the aspect of SELinux or
SMACK. If they were, they would just use the LSM interface. virus
scanners are interested in blocking "normal" programs from reading
invalid data from disk before acting on it or executing it.
thanks,
greg k-h
On Tue, 2008-08-05 at 12:37 -0400, Press, Jonathan wrote:
> Regarding self-exclusion... We are doing write-time scanning, or more
> accurately close-time scanning. We don't block the close, but we do
> scan the file afterwards, and if it is infected, we perform any one of
> several actions -- such as report to user, rename, quarantine, delete,
> etc. I would not want a self-excluded executable to be able to write
> something that I don't get to scan.
With the current implementation there are 2 ways to be excluded. Both
require root and I plan on both requiring the access to pass a
newfangled LSM hook or maybe just require CAP_RAWIO. LSM people have
thoughts?
Method #1) Become a client listening for access decisions, basically
just open /security/talpa/client/talpa-client and you are free of
open/close scans. We have to make the scanner itself not cause its own
opens and closes to need scanned, think infinite recursion.
Method #2) Exclude yourself. This involves
opening /security/talpa/exclude/talpa-exclude and writing "1" into it.
this file is owned by root and is 600. Regular user processes cannot
exclude themselves willy nilly nor can any configuration exclude them.
It might be possible to do exclusions in userspace using the pid and
non-caching results for things other than the scanning clients
themselves.
If you can outline the design of a better method that meets your needs
I'd be glad to try to code it. In your mind how do you see programs
being able to exclude others while not being a security risk? I assume
your answer is that the program "giving out the exclusions" must be
root, which we already satisfy. There is (or could be, I don't remember
offhand) the option to disable thread exclusions in kernel (except for
those threads that act as userspace clients, they MUST be excluded
somehow). But really as it stands any root process could just enable
them again. In the non-LSM case root processes already won so, they can
just disable the whole infrastructure send kill -9 to all your clients
and have at it.....
-Eric
On Tue, 05 Aug 2008 13:19:56 -0400
Eric Paris <[email protected]> wrote:
> If you can outline the design of a better method that meets your needs
> I'd be glad to try to code it. In your mind how do you see programs
> being able to exclude others while not being a security risk?
ok so lets be specific.
You are trying to prevent an application from opening a "damaged" file,
or from someone starting a "damaged" file.
You are not trying to prevent anything once you have executed a damaged
file; once you execute one of these for this part it's game over (to
limit the damage other tools like selinux exist, but are outside the
scope of talpa).
So... as long as /sbin/init isn't compromised... intercepting exec and
open (in all variants) is all you need.
And this can be done from userland with the preload: the "workaround"
from the preload assumes you've already executed malicious code, which
is outside of your protection scope.
What am I missing?
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
> And this can be done from userland with the preload: the "workaround"
> from the preload assumes you've already executed malicious code, which
> is outside of your protection scope.
>
> What am I missing?
Scripts
Attempts to screen content
Exec occuring after ld.so is compromised
Is there anything however that cannot be done with SELinux if you added
the ability to block an open and kick it upwards (including the open of
an exec binary)
It seems you would then get a transition from a label of 'trusted' to
'untrusted_unverified' and an open of untrusted_unverified can (depending
on the SELinux rule) then block, trap upwards and continue according to a
userspace response.
At that point all the questions like 'what do I want to scan for' become
SELinux questions and we already have all the technology to do stuff like
'only scan for samba' or 'only scan for httpd and cgi' and do it
efficiently.
The cache then becomes the labels which are already part of the fs and
our existing labelling and context management.
Alan
On Tuesday 05 August 2008 21:31, Alan Cox wrote:
> > No, I want a sane security policy in kernelsapce that doesn't look
> > at the content because doing security by content properly is equivalent
> > to solving the halting problem. I couldn't give a rats a** about
> > windows viruses as they can't actually cause any harm on a Linux
> > machine.
>
> Go on then.. post patches.
Onus is not (or actually often is, but should not) be on reviewers
to do this. That effectively further reduces review bandwidth.
I can see why "feature" writers like the idea though. Either they
get someone to do their work for them, or negative reviews
disappear :)
On Tue, 5 Aug 2008 18:29:44 +0100
Alan Cox <[email protected]> wrote:
> > And this can be done from userland with the preload: the
> > "workaround" from the preload assumes you've already executed
> > malicious code, which is outside of your protection scope.
> >
> > What am I missing?
>
> Scripts
get either opened or exec'd...
(ignoring stdin-fed scripts right now.. but that's a problem regardless)
> Attempts to screen content
can you explain?
> Exec occuring after ld.so is compromised
this is post-compromise scenario; if you have enough root rights to do
that then it's game over.
I'm not sure if this is off the direct idea of this thread, or if I am
possibly missing the whole point.
However, I want to point out that scanning on close is still an integral
part of AV protection, even if intercepting opens and execs
theoretically catches everything.
You can say that there are four parts to the malware life cycle --
getting on a machine, residing there, causing local damage, and
propagating elsewhere. It is part of the philosophy of AV protection
that you do everything you can to prevent all of them. That's why there
are scans on close, scheduled scans, and scans on open. Most of our
users employ all three and do not rely on one or two. If an infection
arrives on a machine and finds a home because it is assumed that it will
be caught when it is opened for use, then it is just one more compromise
away from doing damage and/or spreading.
Jon Press
-----Original Message-----
From: Arjan van de Ven [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 1:39 PM
To: Eric Paris
Cc: Press, Jonathan; Greg KH; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a linux
interfaceforon access scanning
On Tue, 05 Aug 2008 13:19:56 -0400
Eric Paris <[email protected]> wrote:
> If you can outline the design of a better method that meets your needs
> I'd be glad to try to code it. In your mind how do you see programs
> being able to exclude others while not being a security risk?
ok so lets be specific.
You are trying to prevent an application from opening a "damaged" file,
or from someone starting a "damaged" file.
You are not trying to prevent anything once you have executed a damaged
file; once you execute one of these for this part it's game over (to
limit the damage other tools like selinux exist, but are outside the
scope of talpa).
So... as long as /sbin/init isn't compromised... intercepting exec and
open (in all variants) is all you need.
And this can be done from userland with the preload: the "workaround"
from the preload assumes you've already executed malicious code, which
is outside of your protection scope.
What am I missing?
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Tuesday 05 August 2008 07:00, Eric Paris wrote:
> Please contact me privately or (preferably the list) for questions,
> comments, discussions, flames, names, or anything. I'll do complete
> rewrites of the patches if someone tells me how they don't meet their
> needs or how they can be done better. I'm here to try to bridge the
> needs (and wants) of the anti-malware vendors with the technical
> realities of the kernel. So everyone feel free to throw in your two
> cents and I'll try to reconcile it all. These 5 patches are part 1.
> They give us a working able solution.
>
> >From my point of view patches forthcoming and mentioned below should
>
> help with performance for those who actually have userspace scanners but
> also could presents be implemented using this framework.
>
>
> Background
> ++++++++++
> There is a consensus in the security industry that protecting against
> malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
> of so-called on-access scanning is usable and reasonable approach.
> Currently the Linux kernel does not offer a completely suitable
> interface to implement such security solutions. Present solutions
> involve overwriting function pointers in the LSM, in filesystem
> operations, in the sycall table, and other fragile hacks. The purpose
> of this project is to create a fast, clean interface for userspace
> programs to look for malware when files are accessed. This malware may
> be ultimately intended for this or some other Linux machine or may be
> malware intended to attack a host running a different operating system
> and is merely in transit across the Linux server. Since there are
> almost an infinite number of ways in which information can enter and
> exit a server it is not seen as reasonable to move these checks to all
> the applications at the boundary (MTA, NFS, CIFS, SSH, rsync, et al.) to
> look for such malware on at the border.
>
> For this Linux kernel interface speed is of particular interest for
> those who have it compiled into the kernel but have no userspace client.
> There must be no measurable performance hit to just compiling this into
> the kernel.
>
> Security vendors, Linux distributors and other interested parties have
> come together on the malware-list mailing list to discuss this problem
> and see if they can work together to propose a solution. During these
> talks couple of requirement sets were posted with the aim of fleshing
> out common needs as a prerequisite of creating an interface prototype.
>
> Collated requirements
> +++++++++++++++++++++
I suspect that what people actually want is a requirement of what it is
they are trying to do. Not a list of demands on the kernel that they think
are needed to implement what they think is the right way to solve some
problem that they define.
> 5. Define which filesystems are cacheable and which are not
This is practically impossible to do completely without rewriting a lot
of code (which will never be accepted). I don't see why it is needed though
as the filesystem cache is supposed to be kept coherent with disk.
A: No.
Q: Should I include quotations after my reply?
On Tue, Aug 05, 2008 at 02:04:26PM -0400, Press, Jonathan wrote:
> I'm not sure if this is off the direct idea of this thread, or if I am
> possibly missing the whole point.
I think you might be missing the point a bit here, as the traditional
Unix model that Linux has prevents much of what the "traditional AV"
products need to do, right?
> However, I want to point out that scanning on close is still an integral
> part of AV protection, even if intercepting opens and execs
> theoretically catches everything.
Great, then put a hook in glibc and catch all closes and then kick off
your scanning.
> You can say that there are four parts to the malware life cycle --
> getting on a machine, residing there, causing local damage, and
> propagating elsewhere. It is part of the philosophy of AV protection
> that you do everything you can to prevent all of them.
But this proposed patchset does not do much to prevent all of these,
right?
> That's why there are scans on close, scheduled scans, and scans on
> open. Most of our users employ all three and do not rely on one or
> two. If an infection arrives on a machine and finds a home because it
> is assumed that it will be caught when it is opened for use, then it
> is just one more compromise away from doing damage and/or spreading.
So how are you going about preventing the "infection from arriving"
with this proposed patchset?
Isn't that something that SELinux or another LSM can prevent better?
thanks,
greg k-h
On Tue, 5 Aug 2008 14:04:26 -0400
"Press, Jonathan" <[email protected]> wrote:
>
> However, I want to point out that scanning on close is still an
> integral part of AV protection, even if intercepting opens and execs
> theoretically catches everything.
but close is... very limited in value. Open is a discrete event
traditionally associated withh permission checks.
Close... not so. (And if you mmap memory, you can then close the file
and still write to it via the mmap)
Lets ask it differently: what will you do if you find something nasty?
You can't fail the close... so why block for it?
And if you don't block for it... all you would need is an asynchronous
notification... something like... inotify
You're right...I am not talking about blocking at all -- which may be a
further indication that I am missing the specific point of this thread.
But be that as it may... I don't want to have to use more than one
interface to get all the events I am interested in. I want to register
as a client and listen, and get everything I need from the same place.
Also, it seems to me that for my purposes, close is discrete enough. It
tells me that there is now something out there that should be looked at.
Jon
-----Original Message-----
From: Arjan van de Ven [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 2:28 PM
To: Press, Jonathan
Cc: Eric Paris; Greg KH; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a linux
interfaceforon access scanning
On Tue, 5 Aug 2008 14:04:26 -0400
"Press, Jonathan" <[email protected]> wrote:
>
> However, I want to point out that scanning on close is still an
> integral part of AV protection, even if intercepting opens and execs
> theoretically catches everything.
but close is... very limited in value. Open is a discrete event
traditionally associated withh permission checks.
Close... not so. (And if you mmap memory, you can then close the file
and still write to it via the mmap)
Lets ask it differently: what will you do if you find something nasty?
You can't fail the close... so why block for it?
And if you don't block for it... all you would need is an asynchronous
notification... something like... inotify
>> I think you might be missing the point a bit here, as the traditional
Unix model that
>> Linux has prevents much of what the "traditional AV" products need to
do, right?
Is your point that Linux and Unix machines are less vulnerable to
viruses? If so, that's not relevant to my point at all. A Unix machine
can be a carrier, passing infections on to other vulnerable platforms
(guess which one). An enterprise security system sees the entire
enterprise as an integrated whole -- not just individual machines with
their own separate attributes and no impact on each other at all.
>>So how are you going about preventing the "infection from arriving"
>> with this proposed patchset?
I'm not endorsing or opposing the proposal until I digest it further.
However, I will say that while preventing infections from arriving is
not foolproof, doing a scan-on-close with the option to delete or
quarantine an infected file goes a long way.
Jon
-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 2:12 PM
To: Press, Jonathan
Cc: Arjan van de Ven; Eric Paris; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
linuxinterfaceforon access scanning
A: No.
Q: Should I include quotations after my reply?
On Tue, Aug 05, 2008 at 02:04:26PM -0400, Press, Jonathan wrote:
> I'm not sure if this is off the direct idea of this thread, or if I am
> possibly missing the whole point.
I think you might be missing the point a bit here, as the traditional
Unix model that Linux has prevents much of what the "traditional AV"
products need to do, right?
> However, I want to point out that scanning on close is still an
integral
> part of AV protection, even if intercepting opens and execs
> theoretically catches everything.
Great, then put a hook in glibc and catch all closes and then kick off
your scanning.
> You can say that there are four parts to the malware life cycle --
> getting on a machine, residing there, causing local damage, and
> propagating elsewhere. It is part of the philosophy of AV protection
> that you do everything you can to prevent all of them.
But this proposed patchset does not do much to prevent all of these,
right?
> That's why there are scans on close, scheduled scans, and scans on
> open. Most of our users employ all three and do not rely on one or
> two. If an infection arrives on a machine and finds a home because it
> is assumed that it will be caught when it is opened for use, then it
> is just one more compromise away from doing damage and/or spreading.
So how are you going about preventing the "infection from arriving" with
this proposed patchset?
Isn't that something that SELinux or another LSM can prevent better?
thanks,
greg k-h
you're still top-posting!
Please do not do that.. post your comments below the mail like anyone
else does.
On Tue, 5 Aug 2008 14:34:26 -0400
"Press, Jonathan" <[email protected]> wrote:
> But be that as it may... I don't want to have to use more than one
> interface to get all the events I am interested in. I want to
> register as a client and listen, and get everything I need from the
> same place.
that's your problem to be honest. Worst case you have a small wrapper
to funnel various sources into one.
> Also, it seems to me that for my purposes, close is discrete enough.
> It tells me that there is now something out there that should be
> looked at.
inotify tells you that. close does not.
On Tue, 2008-08-05 at 11:27 -0700, Arjan van de Ven wrote:
> On Tue, 5 Aug 2008 14:04:26 -0400
> "Press, Jonathan" <[email protected]> wrote:
>
> >
> > However, I want to point out that scanning on close is still an
> > integral part of AV protection, even if intercepting opens and execs
> > theoretically catches everything.
>
>
> but close is... very limited in value. Open is a discrete event
> traditionally associated withh permission checks.
> Close... not so. (And if you mmap memory, you can then close the file
> and still write to it via the mmap)
Thankfully my implementation will invalidate that close time check and
caching result. It does the invalidating the same place we update mtime
and my understanding is that mmap has been updating mtime for quite a
while now. So again, it might not be perfect, but that situation should
get caught the next time around. I see close time checking as more of a
performance win and as a way to help close some of the length bad
information could exist than anything since its done in a non-blocking
path.
I think we all agree that open is the most interesting time for scanning
operations, but as Jonathan points out there is some value (even if not
perfect value) in looking at closes as well.
> Lets ask it differently: what will you do if you find something nasty?
> You can't fail the close... so why block for it?
> And if you don't block for it... all you would need is an asynchronous
> notification... something like... inotify
I actually already have an async non-blocking close notification built
in. Instead of waiting for a userspace response we just queue the close
notification and move along. When userspace gets around to it scanning
and allow/deny caching can then take place at a later time.
-Eric
A: http://en.wikipedia.org/wiki/Top_post
Q: Were do I find info about this thing called top-posting?
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?
On Tue, Aug 05, 2008 at 02:34:26PM -0400, Press, Jonathan wrote:
> You're right...I am not talking about blocking at all -- which may be a
> further indication that I am missing the specific point of this thread.
>
> But be that as it may... I don't want to have to use more than one
> interface to get all the events I am interested in. I want to register
> as a client and listen, and get everything I need from the same place.
That's an implementation issue, not a requirement. If it's a
requirement, it sure is a lazy one :)
> Also, it seems to me that for my purposes, close is discrete enough. It
> tells me that there is now something out there that should be looked at.
So, if you hook glibc to catch all calls to close, is that sufficient?
thanks,
greg k-h
On Tue, Aug 05, 2008 at 02:38:23PM -0400, Press, Jonathan wrote:
> Is your point that Linux and Unix machines are less vulnerable to
> viruses? If so, that's not relevant to my point at all. A Unix machine
> can be a carrier, passing infections on to other vulnerable platforms
> (guess which one). An enterprise security system sees the entire
> enterprise as an integrated whole -- not just individual machines with
> their own separate attributes and no impact on each other at all.
Sure, but if that's the case, you don't need to have a blocking open()
interface. Having inotify tell your application that a file
descriptor that had been opened for writing has been closed
(IN_CLOSE_WRITE) should be quite sufficient.
- Ted
On Tue, 2008-08-05 at 10:03 -0700, Greg KH wrote:
> On Tue, Aug 05, 2008 at 12:23:28PM +0100, Alan Cox wrote:
> > > > It may be possible to do in glibc, LD_PRELOAD doesn't exactly work for
> > > > suid binaries
> > >
> > > Are suid binaries something that you feel is necessary to scan from?
> > >
> > > I don't see it on the list above :)
> >
> > Doesn't work very well really does it - ld.so loads files too and can be
> > attacked.
>
> But that's an "attack" vector, which virus scanners are not addressing.
> They are going for the "is this file corrupted" type issue. The
> "normal" LSM interface is for malware itself, taking control of a
> system.
So you are arguing against the defense in depth theory? LSM should
solve it all so why bother?
I don't think that anyone is claiming that they don't address that
vector. They may not be perfect but they are infinitely better than the
zero protection we offer today. Clearly in the enterprise world the
most useful purpose of these scanners is looking at inflight data
crossing valid safe non-hacked linux processes, but the implementation I
have given is such that we can start doing some validation on our own
executables. Remember the requirement wasn't just to scan data, it was
to scan everything that gets opened.
As an example of the usefulness of this approach as opposed to the
LD_PRELOAD/glibc approach is that I could be used to minimize the impact
of things like the recently much touted discussion about malicious
repository mirrors. Although clearly there were some flaws in the
common conception of the problem, the ability to trick users into
downloading and installing trojaned libraries is not something we can
presently protect against with any mechanism.
Lets assume that the black magic in userspace was able to spot a
trojaned library running programs which would still be linked to the old
files on disk would continue to work but you would very quickly find out
there was trouble when everything that tried to open its shared library
was getting EPERM. Yes, yes, even I can think of ways around it (bad
repo sends bad kernel first and waits until the machine is running the
bad kernel to disable scanning) but by that time they already won. And
its possible that the userspace scanning could/would spot the trojaned
kernel and report on it during install time.
I'm going to run perf test this afternoon but I'm going to have to look
for a test that is a good mix of reads,writes, and opens. I strongly
suspect that there will be a noticable perf lose in any other method
that doesn't include in kernel caching. (how can there not be with an
extra context switch for every open?)
-Eric
On Mon, 2008-08-04 at 17:51 -0700, Greg KH wrote:
> On Mon, Aug 04, 2008 at 08:32:54PM -0400, Eric Paris wrote:
> Oh, and after that, not using a binary interface, have we not learned
> from the ioctl mess? I sure thought we had...
I don't see a reason why we can't use strings and key=value pairs for
any metadata being sent back and forth. That seem more reasonable?
> > > Heh, so if you want to write a "virus" for Linux, just implement this
> > > flag. What's to keep a "rogue" program from telling the kernel that all
> > > programs on the system are to be excluded?
> >
> > Processes can only get this flag one of 2 ways.
> >
> > 1) register as a client to make access decisions
>
> How do you do that?
open the magic "vetting" file RW and you are a client who can answer
access decisions.
> > 2) echo 1 into the magic file to enable the flag for themselves
>
> Simple enough :)
>
> > > > 10. Filesystem exclusions
> > > > -------------------------
> > > > One pretty important optimization is not to scan things like /proc, /sys
> > > > or similar. Basically all filesystems where user can not store
> > > > arbitrary, potentially malicious, content could and should be excluded
> > > > from scanning.
> > >
> > > Why, does scanning these files take extra time? Just curious.
> >
> > Perf win, why bothering looking for malware in /proc when it can't
> > exist? It doesn't take longer it just takes time having to do
> >
> > userspace -> kernel -> userspace -> kernel -> userspace
> >
> > just to cat /proc/mounts, all of this could probably be alliviated if we
> > cached access on non block backed files but then we have to come up with
> > a way to exclude only nfs/cifs. I'd rather list the FSs that don't need
> > scanning every time than those that do....
>
> How long does this whole process take? Seriously is it worth the added
> kernel code for something that is not measurable?
Is it worth having 2 context switches for every open when none are
needed? I plan to get numbers on that.
>
> > > > 11. Path exclusions
> > > > -------------------
> > > > The need for exclusions can be demonstrated with an example of a MySQL
> > > > server. It's data files are frequently modified which means they would
> > > > need to be constantly rescanned which is very bad for performance. Also,
> > > > it is most often not even possible to reasonably scan them. Therefore
> > > > the best solution is not to scan its database store which can simply be
> > > > implemented by excluding the store subdirectory.
> > > >
> > > > It is a relatively simple implementation which allows run-time
> > > > configuration of a list of sub directories or files to exclude.
> > > > Exclusion paths are relative to each process root. So for example if we
> > > > want to exclude /var/lib/mysql/ and we have a mysql running in a chroot
> > > > where from the outside that directory actually lives
> > > > in /chroot/mysql/var/lib/mysql, /var/lib/mysql should actually be added
> > > > to the exclusion list.
> > > >
> > > > This is also not included in the initial patch set but will be coming
> > > > shortly after.
> > >
> > > Again, what's to keep all files to be marked as excluded?
> >
> > You have to be root and I'll probably add an LSM hook?
>
> Why an LSM hook? You aren't an LSM.
Maybe a new hook, maybe just a capability call, something to make root
not the
> > > > Closing remarks
> > > > ---------------
> > > > Although some may argue some of the filters are not necessary or may
> > > > better be implemented in userspace, we think it is better to have them
> > > > in kernel primarily for performance reasons.
> > >
> > > Why? What numbers do you have that say the kernel is faster in
> > > implementing this? This is the first mention of such a requirement, we
> > > need to see real data to back it up please.
> >
> > In kernel caching is clearly a huge perf win.
>
> Why? If the cache is also in userspace, it should be the same, right?
In kernel cache has 0 context switches for every open. Userspace
caching has 2. Every open has to block, switch to the context of the
userspace client/cache, get that decisions, and then switch back to the
original process.
-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 2:39 PM
To: Press, Jonathan
Cc: Arjan van de Ven; Eric Paris; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
linuxinterfaceforon access scanning
A: http://en.wikipedia.org/wiki/Top_post
Q: Were do I find info about this thing called top-posting?
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?
On Tue, Aug 05, 2008 at 02:34:26PM -0400, Press, Jonathan wrote:
> You're right...I am not talking about blocking at all -- which may be
a
> further indication that I am missing the specific point of this
thread.
>
> But be that as it may... I don't want to have to use more than one
> interface to get all the events I am interested in. I want to
register
> as a client and listen, and get everything I need from the same place.
That's an implementation issue, not a requirement. If it's a
requirement, it sure is a lazy one :)
[JON PRESS] I wouldn't call it lazy, actually. It's more like
"economical" or "ergonomic" -- or, dare I say it -- "user-friendly." In
this case, the users are the AV vendors who will have to write to the
API that will come out of this spec. We will be more inclined to
appreciate the SDK (for want of a better term) if it covers all the
bases, rather than force us to go elsewhere for some of our
requirements. When we write SDKs, we try to make sure that our users
will find whatever they need.
> Also, it seems to me that for my purposes, close is discrete enough.
It
> tells me that there is now something out there that should be looked
at.
So, if you hook glibc to catch all calls to close, is that sufficient?
[JON PRESS] Let's see...I'm going to use inotify for some events, glibc
for others, and this API for the rest. Would you really want to write
an application like that?
On Tue, Aug 05, 2008 at 03:46:03PM -0400, Eric Paris wrote:
> On Mon, 2008-08-04 at 17:51 -0700, Greg KH wrote:
> > On Mon, Aug 04, 2008 at 08:32:54PM -0400, Eric Paris wrote:
>
> > Oh, and after that, not using a binary interface, have we not learned
> > from the ioctl mess? I sure thought we had...
>
> I don't see a reason why we can't use strings and key=value pairs for
> any metadata being sent back and forth. That seem more reasonable?
Sure, but do you really want to put a parser in the kernel (well, make
that, yet-another-parser-in-the-kernel...)?
> > > > Heh, so if you want to write a "virus" for Linux, just implement this
> > > > flag. What's to keep a "rogue" program from telling the kernel that all
> > > > programs on the system are to be excluded?
> > >
> > > Processes can only get this flag one of 2 ways.
> > >
> > > 1) register as a client to make access decisions
> >
> > How do you do that?
>
> open the magic "vetting" file RW and you are a client who can answer
> access decisions.
What's to keep anyone from doing this?
> > > Perf win, why bothering looking for malware in /proc when it can't
> > > exist? It doesn't take longer it just takes time having to do
> > >
> > > userspace -> kernel -> userspace -> kernel -> userspace
> > >
> > > just to cat /proc/mounts, all of this could probably be alliviated if we
> > > cached access on non block backed files but then we have to come up with
> > > a way to exclude only nfs/cifs. I'd rather list the FSs that don't need
> > > scanning every time than those that do....
> >
> > How long does this whole process take? Seriously is it worth the added
> > kernel code for something that is not measurable?
>
> Is it worth having 2 context switches for every open when none are
> needed? I plan to get numbers on that.
Compared to the real time it takes in the "virus engine"? I bet it's
totally lost in the noise. Those things are huge beasts with thousands
to hundreds of thousands of context switches.
> > > In kernel caching is clearly a huge perf win.
> >
> > Why? If the cache is also in userspace, it should be the same, right?
>
> In kernel cache has 0 context switches for every open. Userspace
> caching has 2. Every open has to block, switch to the context of the
> userspace client/cache, get that decisions, and then switch back to the
> original process.
Again, compared to what? If you in userspace are doing big complex
things, such an overhead is trivial.
And again, realize that Linux has the fastest context switches _by far_
of any other operating system. It is ok to do things in userspace, we
are used to that :)
thanks,
greg k-h
On Tue, Aug 05, 2008 at 02:38:23PM -0400, Press, Jonathan wrote:
> >> I think you might be missing the point a bit here, as the traditional
> Unix model that
> >> Linux has prevents much of what the "traditional AV" products need to
> do, right?
>
> Is your point that Linux and Unix machines are less vulnerable to
> viruses? If so, that's not relevant to my point at all. A Unix machine
> can be a carrier, passing infections on to other vulnerable platforms
> (guess which one).
So you are going to try to force us to take something into the Linux
kernel due to the security inadiquacies of a totally different operating
system? You might want to rethink that argument :)
> An enterprise security system sees the entire enterprise as an
> integrated whole -- not just individual machines with their own
> separate attributes and no impact on each other at all.
I agree, but as others have pointed out, you don't need to do this in
the kernel, you can do it from userspace today (samba has hooks for this
for that "other" operating system already).
thanks,
greg k-h
On Tue, Aug 05, 2008 at 04:15:32PM -0400, Press, Jonathan wrote:
> On Tue, Aug 05, 2008 at 02:34:26PM -0400, Press, Jonathan wrote:
> > You're right...I am not talking about blocking at all -- which may be
> a
> > further indication that I am missing the specific point of this
> thread.
> >
> > But be that as it may... I don't want to have to use more than one
> > interface to get all the events I am interested in. I want to
> register
> > as a client and listen, and get everything I need from the same place.
>
> That's an implementation issue, not a requirement. If it's a
> requirement, it sure is a lazy one :)
>
>
> [JON PRESS] I wouldn't call it lazy, actually. It's more like
> "economical" or "ergonomic" -- or, dare I say it -- "user-friendly." In
> this case, the users are the AV vendors who will have to write to the
> API that will come out of this spec. We will be more inclined to
> appreciate the SDK (for want of a better term) if it covers all the
> bases, rather than force us to go elsewhere for some of our
> requirements. When we write SDKs, we try to make sure that our users
> will find whatever they need.
But realize that you are adding an overhead on us, the kernel community,
to make your life easier. We are the ones that are taking our time to
review and comment on this code. We are the ones who will have to live
with this code for forever, and maintain it over the lifetime of linux.
So far, you all have shown no willingness to give anything back to us at
all.
In fact, I'd go so far as to say you have been openly hostile, violating
our copyright and license by shipping closed source kernel modules,
making our users have huge problems when we can not support them if they
happen to have the misfortune of using your product, and creating code
that pokes directly into the kernel in ways we explicily do not want to
have happen (syscall hooking, walking symbol tables, etc.)
So please remember, that you should be the ones going out of your way to
be nice to us, as you are coming from a huge deficit here that you all
need to make up from.
> > Also, it seems to me that for my purposes, close is discrete enough.
> It
> > tells me that there is now something out there that should be looked
> at.
>
> So, if you hook glibc to catch all calls to close, is that sufficient?
>
> [JON PRESS] Let's see...I'm going to use inotify for some events, glibc
> for others, and this API for the rest. Would you really want to write
> an application like that?
Think of it as a way to justify your huge prices :)
thanks,
greg k-h
-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 4:18 PM
To: Press, Jonathan
Cc: Arjan van de Ven; Eric Paris; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
alinuxinterfaceforon access scanning
On Tue, Aug 05, 2008 at 02:38:23PM -0400, Press, Jonathan wrote:
> >> I think you might be missing the point a bit here, as the
traditional
> Unix model that
> >> Linux has prevents much of what the "traditional AV" products need
to
> do, right?
>
> Is your point that Linux and Unix machines are less vulnerable to
> viruses? If so, that's not relevant to my point at all. A Unix
machine
> can be a carrier, passing infections on to other vulnerable platforms
> (guess which one).
So you are going to try to force us to take something into the Linux
kernel due to the security inadiquacies of a totally different operating
system? You might want to rethink that argument :)
[JON PRESS] On the contrary...you might want to rethink your reaction.
The security inadequacies of that other operating system that happens to
have a 90+% market sure are exactly why Linux and other OS's that
coexist with it should be more conscious of their own interactions with
it. Enterprises that see Linux as a potential breeding ground for
infestations are less likely to tolerate Linux in their environment.
Why do you think we have so many customers who have a corporate mandate
to have AV software on all machines, no matter what platform type?
thanks,
greg k-h
> > Scripts
>
> get either opened or exec'd...
>
> (ignoring stdin-fed scripts right now.. but that's a problem regardless)
Don't forget pty/tty pairs, ioctls to type chars etc
> > Attempts to screen content
>
> can you explain?
One common use for scanning is to scan content not looking for executable
stuff but for things like malformed image files and the like. It's also
useful for indexing and in those cases you don't really want ld.so in the
way.
> > Exec occuring after ld.so is compromised
>
> this is post-compromise scenario; if you have enough root rights to do
> that then it's game over.
You still want to know if you spot this. And the root rights thing
assumes no selinux. For a large case of uses I would agree however.
Alan
On Tue, Aug 05, 2008 at 02:56:42PM -0400, Eric Paris wrote:
> On Tue, 2008-08-05 at 10:03 -0700, Greg KH wrote:
> > On Tue, Aug 05, 2008 at 12:23:28PM +0100, Alan Cox wrote:
> > > > > It may be possible to do in glibc, LD_PRELOAD doesn't exactly work for
> > > > > suid binaries
> > > >
> > > > Are suid binaries something that you feel is necessary to scan from?
> > > >
> > > > I don't see it on the list above :)
> > >
> > > Doesn't work very well really does it - ld.so loads files too and can be
> > > attacked.
> >
> > But that's an "attack" vector, which virus scanners are not addressing.
> > They are going for the "is this file corrupted" type issue. The
> > "normal" LSM interface is for malware itself, taking control of a
> > system.
>
> So you are arguing against the defense in depth theory? LSM should
> solve it all so why bother?
No, I am saying I have yet to see a real requirement to justify that
this code goes into the kernel.
Your list shows no such requirement, in fact it overlaps with some
things that LSM already provides. So why not just make it an LSM and be
done with it?
Oh wait, this is a commercial add-on for vendor kernels that already
have an LSM built in, that will not work.
Is that really the community's issue? :)
> I don't think that anyone is claiming that they don't address that
> vector. They may not be perfect but they are infinitely better than the
> zero protection we offer today. Clearly in the enterprise world the
> most useful purpose of these scanners is looking at inflight data
> crossing valid safe non-hacked linux processes, but the implementation I
> have given is such that we can start doing some validation on our own
> executables. Remember the requirement wasn't just to scan data, it was
> to scan everything that gets opened.
/me points at a fixed inotify.
> As an example of the usefulness of this approach as opposed to the
> LD_PRELOAD/glibc approach is that I could be used to minimize the impact
> of things like the recently much touted discussion about malicious
> repository mirrors. Although clearly there were some flaws in the
> common conception of the problem, the ability to trick users into
> downloading and installing trojaned libraries is not something we can
> presently protect against with any mechanism.
Exactly, so don't try to bring it up when it does not even pertain here.
> Lets assume that the black magic in userspace was able to spot a
> trojaned library running programs which would still be linked to the old
> files on disk would continue to work but you would very quickly find out
> there was trouble when everything that tried to open its shared library
> was getting EPERM.
But that's not what virus scanners on Linux do! So again, don't bring
up things that are outside the threat model!
> I'm going to run perf test this afternoon but I'm going to have to look
> for a test that is a good mix of reads,writes, and opens. I strongly
> suspect that there will be a noticable perf lose in any other method
> that doesn't include in kernel caching. (how can there not be with an
> extra context switch for every open?)
See my earlier comments about this.
thanks,
greg k-h
> > However, I want to point out that scanning on close is still an integral
> > part of AV protection, even if intercepting opens and execs
> > theoretically catches everything.
>
> Great, then put a hook in glibc and catch all closes and then kick off
> your scanning.
kill -9
deferred close via mmap
etc etc etc
You can't just armwave it into glibc, that doesn't hold water. You also
have shared state between processes (index on last close of a handle
shared by several threads or processes).
Same problem you have in the indexing business (which also wants the
close hook) - aside from all the practical issues that LD_PRELOAD tends
to turn up.
I'm not actually interested in the AV stuff, but content indexing I do
care about and we do need a way to get notification up to user space.
Alan
-----Original Message-----
From: Theodore Tso [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 2:55 PM
To: Press, Jonathan
Cc: Greg KH; Arjan van de Ven; Eric Paris; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
alinuxinterfaceforon access scanning
On Tue, Aug 05, 2008 at 02:38:23PM -0400, Press, Jonathan wrote:
> Is your point that Linux and Unix machines are less vulnerable to
> viruses? If so, that's not relevant to my point at all. A Unix
machine
> can be a carrier, passing infections on to other vulnerable platforms
> (guess which one). An enterprise security system sees the entire
> enterprise as an integrated whole -- not just individual machines with
> their own separate attributes and no impact on each other at all.
Sure, but if that's the case, you don't need to have a blocking open()
interface. Having inotify tell your application that a file
descriptor that had been opened for writing has been closed
(IN_CLOSE_WRITE) should be quite sufficient.
[JON PRESS] I don't get the connection between what I said and your
point about not needing blocking open() interface. If I ftp into a
Linux machine and GET an infected file, you want FTP to go right ahead
and read it and send it to me over the wire?
On Tue, 5 Aug 2008 16:15:32 -0400
"Press, Jonathan" <[email protected]> wrote:
>
> > Also, it seems to me that for my purposes, close is discrete enough.
> It
> > tells me that there is now something out there that should be looked
> at.
>
> So, if you hook glibc to catch all calls to close, is that sufficient?
>
> [JON PRESS] Let's see...I'm going to use inotify for some events,
> glibc for others, and this API for the rest. Would you really want
> to write an application like that?
so you have to do 2 cases:
1) inotify to notice files changing
(no need to hook glibc for that, and no need to hook close() since you
already get a notify for the change)
This is to catch the *creation* of "bad" content (say a browser saving a
download or somesuch)
By nature this is asynchronous for both performance and "what could you
do if" reasons.
(but so would the close() scan be, and again please explain how you
deal with write-to-mmap-after-close)
2) A synchronous check on open() or exec()
This is to prevent *use* of "bad" content, either by an application
opening a bad file, or by executing a "bad" program.
For neither do you need to hook the kernel; ld preload works great for
this.
This does assume that at some point you have a transition from "ok"
program to the first time you run a "bad" one (via exec or open); and
that you catch it at that point.
I don't yet buy the argument "but what if the virus corrupted your ld
preload", because if it can do that your own virus scanner is also
corrupted.
Can you explain what gap is left after you do these two things?
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Tue, 5 Aug 2008 21:12:43 +0100
Alan Cox <[email protected]> wrote:
> > this is post-compromise scenario; if you have enough root rights to
> > do that then it's game over.
>
> You still want to know if you spot this. And the root rights thing
> assumes no selinux. For a large case of uses I would agree however.
if selinux wouldn't stop a "not really root" root from writing to ld.so
stuff you have a different problem ;-)
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Tue, 2008-08-05 at 16:28 -0400, Press, Jonathan wrote:
> -----Original Message-----
> From: Greg KH [mailto:[email protected]]
> Sent: Tuesday, August 05, 2008 4:18 PM
> To: Press, Jonathan
> Cc: Arjan van de Ven; Eric Paris; [email protected];
> [email protected]; [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
> alinuxinterfaceforon access scanning
>
> On Tue, Aug 05, 2008 at 02:38:23PM -0400, Press, Jonathan wrote:
> > >> I think you might be missing the point a bit here, as the
> traditional
> > Unix model that
> > >> Linux has prevents much of what the "traditional AV" products need
> to
> > do, right?
> >
> > Is your point that Linux and Unix machines are less vulnerable to
> > viruses? If so, that's not relevant to my point at all. A Unix
> machine
> > can be a carrier, passing infections on to other vulnerable platforms
> > (guess which one).
>
> So you are going to try to force us to take something into the Linux
> kernel due to the security inadiquacies of a totally different operating
> system? You might want to rethink that argument :)
>
> [JON PRESS] On the contrary...you might want to rethink your reaction.
> The security inadequacies of that other operating system that happens to
> have a 90+% market sure are exactly why Linux and other OS's that
> coexist with it should be more conscious of their own interactions with
> it. Enterprises that see Linux as a potential breeding ground for
> infestations are less likely to tolerate Linux in their environment.
> Why do you think we have so many customers who have a corporate mandate
> to have AV software on all machines, no matter what platform type?
I don't think the pissing contest is going to get us anywhere. I think
Jon might want to realize that the linux kernel is not driven by
business needs, we are driven by technical correctness and technical
necessity. lkml isn't a place where you wave a bag of money and say "do
you want to be in the data center, do as I say." Techies always win,
not business. I think Greg needs to realize that not all of the AV
vendors are being or want to be difficult, thick, or stubborn. I would
like to point out for the community's enjoyment that much of the heavy
lifting here has been done by one of these vendors who is currently
using the above mentioned horrible hacks to make their product work
(although at least I believe GPL horrible hacks). Most all of the black
magic vendors agree they want to work towards a real upstream solution
so lets try to find it, not just build walls and get defensive.
I personally agree with Greg that I don't care if its 'hard' to get all
the information you need to do your job as long as it is reasonable and
sustainable. I think this interface is both and I'm going to be looking
for numbers to show it over the next couple of days.
I think Alan and I have both described how greater linux security can be
gained through this interface compared to glibc or LD_PRELOAD even if it
isn't perfect security. I certainly don't make the claim that all
malware (for any OS) is going to get stopped dead in its tracks. But
then again I also haven't heard any vendor say "we don't look for any
linux malware." Even if the majority of their business is driven by
"that other OS" it doesn't mean that software on linux is without flaws
and we don't have attackable programs. Would any vendor who does this
type of work stand up and say how your product may have stopped or been
able to stop a vulnerability that would have been impossible in
userspace.
-Eric
On Tue, 5 Aug 2008 10:01:43 -0700
Greg KH <[email protected]> wrote:
> On Tue, Aug 05, 2008 at 12:25:03PM +0100, Alan Cox wrote:
> > > Again, do it all in userspace (caching, and scanning). I still really
> > > don't see the need to do this in the kernel becides it being "the way
> > > people have always done it."
> >
> > We don't have notifiers for file segment changes that are scalable that
> > far.
>
> I agree, but if we did, would that help out a lot here? Lots of other
> groups of people are needing/asking for something like this and if
> someone can finally get it together to post something useful, that might
> be a very good thing.
A scalable notification scheme would certainly sort out content indexing
systems more nicely. Open notifiers that scale let you do path indexing
("People who opened file X also opened file Y" - both for optimising disk
layouts and for application level 'what do I prompt the user with' stuff)
The only thing I can see that is actually needed to get the whole thing
working sweetly even for the virus and HSM cases is an LSM willing to
bounce some opens via a user space helper. Even without that you could
label content 'dubious' on close and relabel it accordingly in the
asynchronous scanning app
Alan
On Tue, 2008-08-05 at 13:38 -0700, Arjan van de Ven wrote:
> and again please explain how you
> deal with write-to-mmap-after-close)
http://marc.info/?l=linux-security-module&m=121796172212429&w=2
Write after close will invalidate the close time caching decision, but
will still get caught on the next open....
-Eric
On Monday 04 August 2008 11:44:28 pm Casey Schaufler wrote:
> Cliffe wrote:
> > Other security schemes such as intrusion detection,
> > firewalls/netfilter, anti-malware, and application restrictions
> > (sandboxes such as jails or finer grained restrictions such as
> > AppArmor) could all register LSMs onto the stack.
>
> Stacking is easy for files. It's a real pain in the backside for UDP
> packets.
How is it any better/worse for UDP packets than files?
--
paul moore
linux @ hp
On Tue, Aug 05, 2008 at 01:38:32PM -0700, Arjan van de Ven wrote:
> This does assume that at some point you have a transition from "ok"
> program to the first time you run a "bad" one (via exec or open); and
> that you catch it at that point.
>
> I don't yet buy the argument "but what if the virus corrupted your ld
> preload", because if it can do that your own virus scanner is also
> corrupted.
>
>
> Can you explain what gap is left after you do these two things?
Actually, the real question (and the reason why I question the personal
integrity of the people in "AV community" pushing that kind of trash)
is very simple:
Where Is Your Threat Profile?
Various people had been asking for _years_ to define what the hell are you
trying to prevent. Not only there'd been no coherent answer (and no, this
list of requirements is _not_ that - it's "what kind of hooks do we want"),
you guys seem to be unable to decide whether you expect the malware in
question to be passive or to be actively evading detection with infected
processes running on the host that does scanning.
Moreover, the answer seems to be changing back and forth to suit the needs
of the moment in the argument. Slightly exaggregated it goes like this:
-- Why don't you do $FOO?
-- Running virus would be able to evade $FOO, of course!
-- No shit, Sherlock; it would also be able to evade much more intrusive $BAR
you are proposing; here's how <obvious evasion method>
-- Oh, but that's not a problem; think of Linux server with Windows clients
and Windows viruses...
On Tue, 05 Aug 2008 16:51:07 -0400
Eric Paris <[email protected]> wrote:
> I think Alan and I have both described how greater linux security can
> be gained through this interface compared to glibc or LD_PRELOAD even
> if it isn't perfect security.
I guess I missed that.
I will fully subscribe to the idea that "the LD_PRELOAD way assumes you
have a good->bad transition" might not be fully bullet proof (after
all, if your init is compromised you won't have this transition)
but what else is the actual gap?
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Tue, Aug 05, 2008 at 04:37:42PM -0400, Press, Jonathan wrote:
>
> [JON PRESS] I don't get the connection between what I said and your
> point about not needing blocking open() interface. If I ftp into a
> Linux machine and GET an infected file, you want FTP to go right ahead
> and read it and send it to me over the wire?
Shouldn't that be the issue of the FTP server itself not serving up
"invalid" files, and not the kernel? Why not just hook in it, I'm
pretty sure they already provide this kind of interface, right?
thanks,
greg k-h
-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 5:15 PM
To: Press, Jonathan
Cc: Theodore Tso; Arjan van de Ven; Eric Paris;
[email protected]; [email protected];
[email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
alinuxinterfaceforonaccess scanning
On Tue, Aug 05, 2008 at 04:37:42PM -0400, Press, Jonathan wrote:
>
> [JON PRESS] I don't get the connection between what I said and your
> point about not needing blocking open() interface. If I ftp into a
> Linux machine and GET an infected file, you want FTP to go right ahead
> and read it and send it to me over the wire?
Shouldn't that be the issue of the FTP server itself not serving up
"invalid" files, and not the kernel? Why not just hook in it, I'm
pretty sure they already provide this kind of interface, right?
[JON PRESS] So how would that work? The FTP server would have code
that called into someone's AV SDK (maybe CA's, maybe not) and scanned
the file before sending. OK. What about all the thousands of other
applications that might access a file and send it somewhere, or copy it
somewhere. They should all do the same thing, right? How do we make
that happen? That's the whole point of centralizing the control (the
notification, blocking and waiting -- not the actual scanning, of
course) in the kernel. The scan becomes unavoidable -- and that is the
definition (OK, a definition) of true security.
On Tue, 5 Aug 2008 17:23:39 -0400
"Press, Jonathan" <[email protected]> wrote:
> -----Original Message-----
> From: Greg KH [mailto:[email protected]]
> Sent: Tuesday, August 05, 2008 5:15 PM
> To: Press, Jonathan
> Cc: Theodore Tso; Arjan van de Ven; Eric Paris;
> [email protected]; [email protected];
> [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
> alinuxinterfaceforonaccess scanning
>
> On Tue, Aug 05, 2008 at 04:37:42PM -0400, Press, Jonathan wrote:
> >
> > [JON PRESS] I don't get the connection between what I said and your
> > point about not needing blocking open() interface. If I ftp into a
> > Linux machine and GET an infected file, you want FTP to go right
> > ahead and read it and send it to me over the wire?
>
> Shouldn't that be the issue of the FTP server itself not serving up
> "invalid" files, and not the kernel? Why not just hook in it, I'm
> pretty sure they already provide this kind of interface, right?
>
>
> [JON PRESS] So how would that work?
the admin (or distro) decides he wants all files opened by the FTP
server scanned, so he starts the FTP server with the LD_PRELOAD set.
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Tue, Aug 05, 2008 at 05:23:39PM -0400, Press, Jonathan wrote:
> [JON PRESS] So how would that work? The FTP server would have code
> that called into someone's AV SDK (maybe CA's, maybe not) and scanned
> the file before sending. OK. What about all the thousands of other
> applications that might access a file and send it somewhere, or copy it
> somewhere. They should all do the same thing, right? How do we make
> that happen? That's the whole point of centralizing the control (the
> notification, blocking and waiting -- not the actual scanning, of
> course) in the kernel. The scan becomes unavoidable -- and that is the
> definition (OK, a definition) of true security.
Give me a break... There's nothing to stop you from doing that as
a stackable fs, cachefs-based filesystem, fuse-based filesystem,
CODA with trimmed-down server that does scan-on-commit, et sodding cetera.
Again, do you or do you not expect the malware to be active on scanning
host? Yes or no, please.
These two variants are so fundamentally different that discussion without
clearly indicating which variant you are currently talking about is absolutely
pointless.
On Tue, Aug 05, 2008 at 05:23:39PM -0400, Press, Jonathan wrote:
> -----Original Message-----
> From: Greg KH [mailto:[email protected]]
> Sent: Tuesday, August 05, 2008 5:15 PM
> To: Press, Jonathan
> Cc: Theodore Tso; Arjan van de Ven; Eric Paris;
> [email protected]; [email protected];
> [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
> alinuxinterfaceforonaccess scanning
>
> On Tue, Aug 05, 2008 at 04:37:42PM -0400, Press, Jonathan wrote:
> >
> > [JON PRESS] I don't get the connection between what I said and your
> > point about not needing blocking open() interface. If I ftp into a
> > Linux machine and GET an infected file, you want FTP to go right ahead
> > and read it and send it to me over the wire?
>
> Shouldn't that be the issue of the FTP server itself not serving up
> "invalid" files, and not the kernel? Why not just hook in it, I'm
> pretty sure they already provide this kind of interface, right?
>
>
> [JON PRESS] So how would that work? The FTP server would have code
> that called into someone's AV SDK (maybe CA's, maybe not) and scanned
> the file before sending. OK. What about all the thousands of other
> applications that might access a file and send it somewhere, or copy it
> somewhere.
Ok, let's stop right here.
Please answer Al's question:
What is the threat model this proposal is attempting to address?
Unless this is well defined, we will go down the "hook this here", "but
what about..." argument back and forth for a very long time getting no
where.
So please let's end this thread right now and can someone please post
the answer to this question.
thanks,
greg k-h
On Mon, Aug 04, 2008 at 05:00:16PM -0400, Eric Paris wrote:
> Please contact me privately or (preferably the list) for questions,
> comments, discussions, flames, names, or anything. I'll do complete
> rewrites of the patches if someone tells me how they don't meet their
> needs or how they can be done better. I'm here to try to bridge the
> needs (and wants) of the anti-malware vendors with the technical
> realities of the kernel. So everyone feel free to throw in your two
> cents and I'll try to reconcile it all. These 5 patches are part 1.
> They give us a working able solution.
>
> >From my point of view patches forthcoming and mentioned below should
> help with performance for those who actually have userspace scanners but
> also could presents be implemented using this framework.
>
>
> Background
> ++++++++++
> There is a consensus in the security industry that protecting against
> malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
> of so-called on-access scanning is usable and reasonable approach.
Can you point to any helpful explanations of that concensus?
Off-hand it's surprising. (A defense that depends on cataloging every
possible individual attack sounds difficult!)
--b.
On Tue, 5 Aug 2008 14:04:26 -0400
"Press, Jonathan" <[email protected]> wrote:
> You can say that there are four parts to the malware life cycle --
> getting on a machine, residing there, causing local damage, and
> propagating elsewhere.
We're getting close.
What exactly are the threat models you want to protect against?
Once we have the answer to that question, we can figure out
whether the current patch set addresses the threat model or
whether alternate approaches are needed.
--
All rights reversed.
On Tue, 05 Aug 2008 14:39:34 -0400
Eric Paris <[email protected]> wrote:
> On Tue, 2008-08-05 at 11:27 -0700, Arjan van de Ven wrote:
> > but close is... very limited in value. Open is a discrete event
> > traditionally associated withh permission checks.
> > Close... not so. (And if you mmap memory, you can then close the file
> > and still write to it via the mmap)
>
> Thankfully my implementation will invalidate that close time check and
> caching result. It does the invalidating the same place we update mtime
> and my understanding is that mmap has been updating mtime for quite a
> while now.
Then isn't the close time check superfluous, since you do the
checks at change time already?
--
All rights reversed.
On Tue, 5 Aug 2008 14:38:23 -0400
"Press, Jonathan" <[email protected]> wrote:
> However, I will say that while preventing infections from arriving is
> not foolproof, doing a scan-on-close with the option to delete or
> quarantine an infected file goes a long way.
How can you, as a security professional, argue for a security measure
that you know is flawed, when there is a mailing list full of people
willing to help figure out what a working security measure would look
like?
Yes, abandoning some of the old DOS anti-virus security model may
cause work, but surely that will be worth it if it leads to a more
secure system?
--
All rights reversed.
On Tue, 5 Aug 2008, Rik van Riel wrote:
> What exactly are the threat models you want to protect against?
>
> Once we have the answer to that question, we can figure out
> whether the current patch set addresses the threat model or
> whether alternate approaches are needed.
Alas, we've been here before.
http://article.gmane.org/gmane.linux.kernel/608634
On Thu, 29 Nov 2007, Al Viro wrote:
> Incidentally, I would really love to see the threat profile we are
> talking about.
Exactly.
Please come up with a set of requirements that can be reviewed by the
core kernel folk, and perhaps then focus on how to meet those requirements
once they have been accepted.
To be very clear, so we don't waste any _more_ time and effort on this:
The anti-malware folk need to first provide a clearly understandable and
complete description of the problem including a characterization of the
threat.
The next step is to propose a design which addresses the problem, and to
clearly and completely demonstrate how it does so.
Only then will it be possible to conduct an informed discussion on the
underlying case for malware scanning and its possible implementation.
- James
--
James Morris
<[email protected]>
On Tue, 2008-08-05 at 20:30 -0400, Rik van Riel wrote:
> On Tue, 05 Aug 2008 14:39:34 -0400
> Eric Paris <[email protected]> wrote:
> > On Tue, 2008-08-05 at 11:27 -0700, Arjan van de Ven wrote:
>
> > > but close is... very limited in value. Open is a discrete event
> > > traditionally associated withh permission checks.
> > > Close... not so. (And if you mmap memory, you can then close the file
> > > and still write to it via the mmap)
> >
> > Thankfully my implementation will invalidate that close time check and
> > caching result. It does the invalidating the same place we update mtime
> > and my understanding is that mmap has been updating mtime for quite a
> > while now.
>
> Then isn't the close time check superfluous, since you do the
> checks at change time already?
In the patches I posted, "checks" are done at open and close if the
result is not already in the cache. Every write invalidates the cache
and thus the next open/close will do a "check."
So the longer a process keeps a file open the longer it is susceptible
to "unclean" data existing in that file.
Eric Paris <[email protected]> writes:
> 5. Fine-grained caching
> -----------------------
> It is necessary to select which filesystems can be safely cached and
> which must not be. For example it is not a good idea to allow caching of
> network filesystems because their content can be changed invisibly. Disk
> based and some virtual filesystems can be cached safely on the other
> hand.
Actually local disk file systems can be changed invisibly to the VFS too by
directly writing to the block device. This does not change the
page cache, but the on disk copy and when a page is pruned from
RAM and reloaded VFS will see the new contents without knowing
about any change. How would you stop that in your
proposal? I assume you could always require a special LKM that
forbids block writes for anything mounted, but that has other problems
too and one wuld need to be extremly careful of holes in
such a protection scheme (e.g. overlapping partitions)
[haven't read the rest of the proposal]
-Andi
Christoph Hellwig <[email protected]> writes:
> I couldn't give a rats a** about
> windows viruses as they can't actually cause any harm on a Linux
> machine.
My guess is that wine is good enough these days to run many Windows
worms actually.
-Andi
"Press, Jonathan" <[email protected]> writes:
>
> Also... I was one of the people who brought up the idea of a process
> exclusion when the requirements list was being developed. I intended it
> as a way that an AV application could exclude specific OTHER processes
> by name (as selected by the AV user)
There's no fixed process name in Linux that cannot be easily faked:
Use process name -- every process can change that by writing to its own
environment.
Use comm name -- there's a prctl to change that and there can be collisions
Use path name of binary -- breaks with chroot and name spaces. Also existing
binaries can be subverted.
Use inode of binary -- can be faked with fuse and breaks when the binaries
is copied ...
Use dev, inode -- breaks when copying binary and when running on network
file systems without a device node
-Andi
Paul Moore wrote:
> On Monday 04 August 2008 11:44:28 pm Casey Schaufler wrote:
>
>> Cliffe wrote:
>>
>>> Other security schemes such as intrusion detection,
>>> firewalls/netfilter, anti-malware, and application restrictions
>>> (sandboxes such as jails or finer grained restrictions such as
>>> AppArmor) could all register LSMs onto the stack.
>>>
>> Stacking is easy for files. It's a real pain in the backside for UDP
>> packets.
>>
>
> How is it any better/worse for UDP packets than files?
>
On delivery you'd need to decide what security scheme is actually
available on the packet and in what order to interpret any inbound
security data. If you had an MLS scheme that uses CIPSO, an integrity
mechanism using IPSEC and a DAC scheme that assigns user ids by
host address getting the ordering right and every domain registered
properly in the networking stack would be a trick. Plus, making sure
that any state the security scheme requires is tricky. Maybe it's not
actually worse if the schemes agree on what qualifies as a security
element, but if one scheme does access control outbound while another
does inbound it will get hairy.
On Wed, 2008-08-06 at 04:35 +0200, Andi Kleen wrote:
> Eric Paris <[email protected]> writes:
>
> > 5. Fine-grained caching
> > -----------------------
> > It is necessary to select which filesystems can be safely cached and
> > which must not be. For example it is not a good idea to allow caching of
> > network filesystems because their content can be changed invisibly. Disk
> > based and some virtual filesystems can be cached safely on the other
> > hand.
>
> Actually local disk file systems can be changed invisibly to the VFS too by
> directly writing to the block device. This does not change the
> page cache, but the on disk copy and when a page is pruned from
> RAM and reloaded VFS will see the new contents without knowing
> about any change. How would you stop that in your
> proposal? I assume you could always require a special LKM that
> forbids block writes for anything mounted, but that has other problems
> too and one wuld need to be extremly careful of holes in
> such a protection scheme (e.g. overlapping partitions)
I didn't consider it. Most likely at the end of the day the finding
will be, "if you can write directly to the block device you already won
since there as so many other things you can do to subvert the system."
Admittedly its the first technical point brought up on list that I
didn't consider at all (lots of other things brought up on list that
need more thought that weren't exactly technical details, don't let me
seem like I'm downplaying those)
-Eric
> I didn't consider it. Most likely at the end of the day the finding
> will be, "if you can write directly to the block device you already won
> since there as so many other things you can do to subvert the system."
This means your scheme is not generally supposed to protect against root?
I assume yes (since I can think of lots of other holes for
root), but you should state that explicitely in the spec since it
is a major limitation.
On the other hand it will also allow you to optimize significantly:
In particularly it also means that you can trust the permissions
and don't need to check any files which cannot be written by users
you don't control.
-Andi
Eric Paris wrote on 05/08/2008 01:32:54:
> On Mon, 2008-08-04 at 15:32 -0700, Greg KH wrote:
> > Why would userspace care about these meta-file things, what does it
want
> > with them?
>
> Honstely? I don't know. Maybe someone with access to the black magic
> source code will stand up and say if most of this metadata is important
> and if so how.
In general this metadata provides more context to the event that happened.
For example reporting - log message/UI popup/centralised something can be
displayed saying which user running which application was involved with
bad stuff. Also we can find out where the user is logged in and send him a
message there.
It is more descriptive than just failing the access with -EACCESS which
becomes ambigious.
> > > 10. Filesystem exclusions
> > > -------------------------
> > > One pretty important optimization is not to scan things like /proc,
/sys
> > > or similar. Basically all filesystems where user can not store
> > > arbitrary, potentially malicious, content could and should be
excluded
> > > from scanning.
> >
> > Why, does scanning these files take extra time? Just curious.
>
> Perf win, why bothering looking for malware in /proc when it can't
> exist? It doesn't take longer it just takes time having to do
>
> userspace -> kernel -> userspace -> kernel -> userspace
>
> just to cat /proc/mounts, all of this could probably be alliviated if we
> cached access on non block backed files but then we have to come up with
> a way to exclude only nfs/cifs. I'd rather list the FSs that don't need
> scanning every time than those that do....
Agreed.
> > > Closing remarks
> > > ---------------
> > > Although some may argue some of the filters are not necessary or may
> > > better be implemented in userspace, we think it is better to have
them
> > > in kernel primarily for performance reasons.
> >
> > Why? What numbers do you have that say the kernel is faster in
> > implementing this? This is the first mention of such a requirement,
we
> > need to see real data to back it up please.
>
> In kernel caching is clearly a huge perf win. I couldn't even measure a
> change in kernel build time when I didn't run a userspace client. If
> anyone can explain a way to get race free in kernel caching and out of
> kernel redirection and scanning I'd love it :)
When you don't run an userspace client cache should not come into play
because nothing will be cached (in this iteration at least). So I guess
you meant something different here? Like not running an userspace client
and having the filter disabled (or even not) will produce very little
overhead, probably not observable without micro-benchmarking. Having an
userspace client which just replies with "allow" should have even less
performance impact because most inodes will get cached which means filter
chain will be shorter on subsequent accesses to the same inode.
In either case it will become obvious how huge performance win is to have
in kernel caching once you get the numbers. Let me know if I can help you
with that somehow.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Greg KH wrote on 05/08/2008 19:11:41:
> On Tue, Aug 05, 2008 at 02:04:26PM -0400, Press, Jonathan wrote:
> > I'm not sure if this is off the direct idea of this thread, or if I am
> > possibly missing the whole point.
>
> I think you might be missing the point a bit here, as the traditional
> Unix model that Linux has prevents much of what the "traditional AV"
> products need to do, right?
Could you please explain some more what and how do you think Unix model
prevents?
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Eric Paris wrote on 05/08/2008 20:46:03:
> On Mon, 2008-08-04 at 17:51 -0700, Greg KH wrote:
> > On Mon, Aug 04, 2008 at 08:32:54PM -0400, Eric Paris wrote:
>
> > Oh, and after that, not using a binary interface, have we not learned
> > from the ioctl mess? I sure thought we had...
>
> I don't see a reason why we can't use strings and key=value pairs for
> any metadata being sent back and forth. That seem more reasonable?
Should be OK from my point of view assuming we keep cache and basic
filesystem exclusions in kernel. Otherwise it would be too much work (I am
talking about CPU time) to do with each and every interception.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Greg KH wrote on 05/08/2008 21:15:35:
> > > > Perf win, why bothering looking for malware in /proc when it can't
> > > > exist? It doesn't take longer it just takes time having to do
> > > >
> > > > userspace -> kernel -> userspace -> kernel -> userspace
> > > >
> > > > just to cat /proc/mounts, all of this could probably be alliviated
if we
> > > > cached access on non block backed files but then we have to come
up with
> > > > a way to exclude only nfs/cifs. I'd rather list the FSs that
don't need
> > > > scanning every time than those that do....
> > >
> > > How long does this whole process take? Seriously is it worth the
added
> > > kernel code for something that is not measurable?
> >
> > Is it worth having 2 context switches for every open when none are
> > needed? I plan to get numbers on that.
>
> Compared to the real time it takes in the "virus engine"? I bet it's
> totally lost in the noise. Those things are huge beasts with thousands
> to hundreds of thousands of context switches.
No, because we are talking about a case here where we don't want to do any
scanning. We want to detect if it is procfs (for example) as quickly as
possible and don't do anything. Same goes for any other filesystem where
it is not possible to store arbitrary user data.
> > > > In kernel caching is clearly a huge perf win.
> > >
> > > Why? If the cache is also in userspace, it should be the same,
right?
> >
> > In kernel cache has 0 context switches for every open. Userspace
> > caching has 2. Every open has to block, switch to the context of the
> > userspace client/cache, get that decisions, and then switch back to
the
> > original process.
>
> Again, compared to what? If you in userspace are doing big complex
> things, such an overhead is trivial.
Again similar thing as above - In case of a cache we are not doing complex
things.
So I think you can't argue that because scanning is slow everything else
has to go to userspace. On a typical running system scanning is
exceptional and everything else benefits from being in the fast path.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Nick Piggin wrote on 05/08/2008 19:08:05:
> On Tuesday 05 August 2008 07:00, Eric Paris wrote:
> > 5. Define which filesystems are cacheable and which are not
>
> This is practically impossible to do completely without rewriting a lot
> of code (which will never be accepted). I don't see why it is needed
though
> as the filesystem cache is supposed to be kept coherent with disk.
Problem is with network filesystems. So could it be a flag somewhere per
filesystem which would say something like "this filesystem guarantees
content of a file cannot change without get_write_access or
file_update_time being called locally"? That doesn't sound like a lot of
code so what am I missing?
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Greg KH wrote on 05/08/2008 21:26:21:
> > [JON PRESS] I wouldn't call it lazy, actually. It's more like
> > "economical" or "ergonomic" -- or, dare I say it -- "user-friendly."
In
> > this case, the users are the AV vendors who will have to write to the
> > API that will come out of this spec. We will be more inclined to
> > appreciate the SDK (for want of a better term) if it covers all the
> > bases, rather than force us to go elsewhere for some of our
> > requirements. When we write SDKs, we try to make sure that our users
> > will find whatever they need.
>
> But realize that you are adding an overhead on us, the kernel community,
> to make your life easier. We are the ones that are taking our time to
> review and comment on this code. We are the ones who will have to live
> with this code for forever, and maintain it over the lifetime of linux.
> So far, you all have shown no willingness to give anything back to us at
> all.
We all? How is that true? I for example wrote some code and am willing to
help maintain it if it gets accepted. And as you describe it, it would be
true for any submission because not all things are usefull for all people,
while everything is baggage for the community. And who is the community? I
thought all who take place in discussions, bug reporting, submitting code,
fixing bugs etc are the community.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
J. Bruce Fields wrote on 05/08/2008 23:55:24:
> On Mon, Aug 04, 2008 at 05:00:16PM -0400, Eric Paris wrote:
> > Please contact me privately or (preferably the list) for questions,
> > comments, discussions, flames, names, or anything. I'll do complete
> > rewrites of the patches if someone tells me how they don't meet their
> > needs or how they can be done better. I'm here to try to bridge the
> > needs (and wants) of the anti-malware vendors with the technical
> > realities of the kernel. So everyone feel free to throw in your two
> > cents and I'll try to reconcile it all. These 5 patches are part 1.
> > They give us a working able solution.
> >
> > >From my point of view patches forthcoming and mentioned below should
> > help with performance for those who actually have userspace scanners
but
> > also could presents be implemented using this framework.
> >
> >
> > Background
> > ++++++++++
> > There is a consensus in the security industry that protecting against
> > malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
> > of so-called on-access scanning is usable and reasonable approach.
>
> Can you point to any helpful explanations of that concensus?
I can't, but everyone is doing it so that is at least an implied
consensus.
> Off-hand it's surprising. (A defense that depends on cataloging every
> possible individual attack sounds difficult!)
Maybe it is not how you imagine it? It is not a database of every possible
individual attack but there are more intelligent methods. But I am not an
expert in this field to explain it better..
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Wed, Aug 06, 2008 at 11:05:43AM +0100, [email protected] wrote:
> Greg KH wrote on 05/08/2008 21:26:21:
>
> > > [JON PRESS] I wouldn't call it lazy, actually. It's more like
> > > "economical" or "ergonomic" -- or, dare I say it -- "user-friendly."
> In
> > > this case, the users are the AV vendors who will have to write to the
> > > API that will come out of this spec. We will be more inclined to
> > > appreciate the SDK (for want of a better term) if it covers all the
> > > bases, rather than force us to go elsewhere for some of our
> > > requirements. When we write SDKs, we try to make sure that our users
> > > will find whatever they need.
> >
> > But realize that you are adding an overhead on us, the kernel community,
> > to make your life easier. We are the ones that are taking our time to
> > review and comment on this code. We are the ones who will have to live
> > with this code for forever, and maintain it over the lifetime of linux.
> > So far, you all have shown no willingness to give anything back to us at
> > all.
>
> We all? How is that true? I for example wrote some code and am willing to
> help maintain it if it gets accepted. And as you describe it, it would be
> true for any submission because not all things are usefull for all people,
> while everything is baggage for the community. And who is the community? I
> thought all who take place in discussions, bug reporting, submitting code,
> fixing bugs etc are the community.
As an observer of this thread:
- Some set of requirements suddenly appears out of the void on
linux-kernel.
- Noone is able and/or willing to exactly describe the problem(s) they
are trying to solve.
With this status quo the discussion is going nowhere - Linux kernel
development does not work this way.
The aim is not to include this code, but to find the best technical
solution for your problem(s) - no matter whether this will have anything
in common with the list of requirements and the code posted or not.
> Tvrtko
cu
Adrian
--
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed
Adrian Bunk <[email protected]> wrote on 06/08/2008 11:50:08:
> On Wed, Aug 06, 2008 at 11:05:43AM +0100, [email protected]
wrote:
> > Greg KH wrote on 05/08/2008 21:26:21:
> >
> > > > [JON PRESS] I wouldn't call it lazy, actually. It's more like
> > > > "economical" or "ergonomic" -- or, dare I say it --
"user-friendly."
> > In
> > > > this case, the users are the AV vendors who will have to write to
the
> > > > API that will come out of this spec. We will be more inclined to
> > > > appreciate the SDK (for want of a better term) if it covers all
the
> > > > bases, rather than force us to go elsewhere for some of our
> > > > requirements. When we write SDKs, we try to make sure that our
users
> > > > will find whatever they need.
> > >
> > > But realize that you are adding an overhead on us, the kernel
community,
> > > to make your life easier. We are the ones that are taking our time
to
> > > review and comment on this code. We are the ones who will have to
live
> > > with this code for forever, and maintain it over the lifetime of
linux.
> > > So far, you all have shown no willingness to give anything back to
us at
> > > all.
> >
> > We all? How is that true? I for example wrote some code and am willing
to
> > help maintain it if it gets accepted. And as you describe it, it would
be
> > true for any submission because not all things are usefull for all
people,
> > while everything is baggage for the community. And who is the
community? I
> > thought all who take place in discussions, bug reporting, submitting
code,
> > fixing bugs etc are the community.
>
> As an observer of this thread:
>
> - Some set of requirements suddenly appears out of the void on
> linux-kernel.
Because previously it was said to go away and come back with a clear list
of requirements. And here you make it sound like a negative thing. See
what I am talking about?
> - Noone is able and/or willing to exactly describe the problem(s) they
> are trying to solve.
Hopefully we will get there. Very little time has passed since the
discussion has started, even less considering the time zone difference for
some.
> With this status quo the discussion is going nowhere - Linux kernel
> development does not work this way.
>
> The aim is not to include this code, but to find the best technical
> solution for your problem(s) - no matter whether this will have anything
> in common with the list of requirements and the code posted or not.
I completely agree with that. Here I was just pointing out that what Greg
wrote was untrue and exaggerated so not helping the discussion at all.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Wednesday 06 August 2008 19:44, [email protected] wrote:
> Nick Piggin wrote on 05/08/2008 19:08:05:
> > On Tuesday 05 August 2008 07:00, Eric Paris wrote:
> > > 5. Define which filesystems are cacheable and which are not
> >
> > This is practically impossible to do completely without rewriting a lot
> > of code (which will never be accepted). I don't see why it is needed
>
> though
>
> > as the filesystem cache is supposed to be kept coherent with disk.
>
> Problem is with network filesystems. So could it be a flag somewhere per
> filesystem which would say something like "this filesystem guarantees
> content of a file cannot change without get_write_access or
> file_update_time being called locally"? That doesn't sound like a lot of
> code so what am I missing?
Maybe... but that's not the same as what requirement 5 calls for.
But depending on exactly what semantics you really call for, it can get
tricky to account for all of pagecache. Writes can happen through page
tables or get_user_pages. True that a process has to at some point have
write permission to the file, but the cache itself could be modified
even after the file is closed and all mmaps disappear.
On Wed, Aug 06, 2008 at 12:07:57PM +0100, [email protected] wrote:
> Adrian Bunk <[email protected]> wrote on 06/08/2008 11:50:08:
> >
> > As an observer of this thread:
> >
> > - Some set of requirements suddenly appears out of the void on
> > linux-kernel.
>
> Because previously it was said to go away and come back with a clear list
> of requirements. And here you make it sound like a negative thing. See
> what I am talking about?
Both of my points belong together.
> > - Noone is able and/or willing to exactly describe the problem(s) they
> > are trying to solve.
>
> Hopefully we will get there. Very little time has passed since the
> discussion has started, even less considering the time zone difference for
> some.
>
> > With this status quo the discussion is going nowhere - Linux kernel
> > development does not work this way.
> >
> > The aim is not to include this code, but to find the best technical
> > solution for your problem(s) - no matter whether this will have anything
> > in common with the list of requirements and the code posted or not.
>
> I completely agree with that. Here I was just pointing out that what Greg
> wrote was untrue and exaggerated so not helping the discussion at all.
Until now the main discussion participant from the AV side is
Jonathan Press.
But the real discussion hasn't even started since the information
required is not available.
And as soon as the information for the real discussion is available all
these initial discussions become irrelevant.
> Tvrtko
cu
Adrian
--
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed
Nick Piggin <[email protected]> wrote on 06/08/2008 12:10:58:
> On Wednesday 06 August 2008 19:44, [email protected] wrote:
> > Nick Piggin wrote on 05/08/2008 19:08:05:
> > > On Tuesday 05 August 2008 07:00, Eric Paris wrote:
> > > > 5. Define which filesystems are cacheable and which are not
> > >
> > > This is practically impossible to do completely without rewriting a
lot
> > > of code (which will never be accepted). I don't see why it is needed
> >
> > though
> >
> > > as the filesystem cache is supposed to be kept coherent with disk.
> >
> > Problem is with network filesystems. So could it be a flag somewhere
per
> > filesystem which would say something like "this filesystem guarantees
> > content of a file cannot change without get_write_access or
> > file_update_time being called locally"? That doesn't sound like a lot
of
> > code so what am I missing?
>
> Maybe... but that's not the same as what requirement 5 calls for.
I see what you mean, it should have been worded better. Nevertheless that
is what was intended by it - to enable caching only on filesystems where
it is safe to do so.
> But depending on exactly what semantics you really call for, it can get
> tricky to account for all of pagecache. Writes can happen through page
> tables or get_user_pages. True that a process has to at some point have
> write permission to the file, but the cache itself could be modified
> even after the file is closed and all mmaps disappear.
I don't have a very good understanding of the VM subsystem I must admit.
So in other words with the current kernel file modification time is not
necessarily correct - it represents when the file was last opened for
modification, not when it was actually modified? (While those two points
in time can be arbitrarily separated)
How would I use those methods for file modification? I am curious to make
a test case..
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
-----Original Message-----
From: Rik van Riel [mailto:[email protected]]
Sent: Tuesday, August 05, 2008 8:51 PM
To: Press, Jonathan
Cc: Greg KH; Arjan van de Ven; Eric Paris; [email protected];
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
linuxinterfaceforon access scanning
On Tue, 5 Aug 2008 14:38:23 -0400
"Press, Jonathan" <[email protected]> wrote:
> However, I will say that while preventing infections from arriving is
> not foolproof, doing a scan-on-close with the option to delete or
> quarantine an infected file goes a long way.
How can you, as a security professional, argue for a security measure
that you know is flawed, when there is a mailing list full of people
willing to help figure out what a working security measure would look
like?
Yes, abandoning some of the old DOS anti-virus security model may cause
work, but surely that will be worth it if it leads to a more secure
system?
[JON PRESS] I hesitate to join back in the discussion, because so far I
haven't been very successful. I know that my level of technical depth
is not at all the equal of just about everybody else in this
conversation. That's not my job and I don't claim to have the answers
to all the questions that are being raised.
However... This question about arguing for a flawed security technique
is a good one, and in a way it gets to the heart of the philosophy of
security. Scan-on-close is admittedly incomplete as an anti-virus tool.
But I don't agree that that make it flawed. It is part of a repertoire
of techniques for preventing malware from residing on a machine.
Let's take a simplistic and unrealistic example. Let's suppose that you
knew that there were 20 applications on your machine that had buffer
overflow vulnerabilities that could be exploited, and you had a fix for
18 of them. Would you decide not to apply the fixes, because that meant
that there would still be 2 left -- and because theoretically there is a
way to prevent all buffer overflows from being exploited and there is a
mailing list full of people trying to figure out how to do it.
I think if it as being like the Sieve of Eratosthenes. The further down
you go, the more numbers drop out. In AV scanning, each step of the way
removes some percentage of the harmful files, and closes the window of
time that they have to operate or migrate. Or maybe it's like spraying
insecticide when there is an outbreak of some deadly mosquito-borne
illness. Without getting into the political issues about spraying,
because this is JUST AN EXAMPLE -- would you not spray because 5% of the
bugs would still be left behind? Wouldn't you then spray again, because
you wipe out another 95%?
If this were not the case, why would so many of our customers run
scheduled system scans on a regular basis in addition to scanning on use
and scanning on close? There are several reasons, including the
constant evolution of malware, the possibility that a machine's
protection was temporarily disabled for some reason, etc. The point is
that catching malware as close to the time it is written as possible is
a net gain that adds to the overall level of protection.
Hopefully, in addition to discussing the particular question that was
raised in Rik's email, this also sheds some light on how I look at the
overall problem -- i.e., from a higher more conceptual level of
functionality and value, rather than the low-level technical specifics.
On Wed, Aug 6, 2008 at 10:10 PM, Press, Jonathan <[email protected]> wrote:
> -----Original Message-----
> From: Rik van Riel [mailto:[email protected]]
> Sent: Tuesday, August 05, 2008 8:51 PM
> To: Press, Jonathan
> Cc: Greg KH; Arjan van de Ven; Eric Paris; [email protected];
> [email protected]; [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
> linuxinterfaceforon access scanning
>
> On Tue, 5 Aug 2008 14:38:23 -0400
> "Press, Jonathan" <[email protected]> wrote:
>
>> However, I will say that while preventing infections from arriving is
>> not foolproof, doing a scan-on-close with the option to delete or
>> quarantine an infected file goes a long way.
>
> How can you, as a security professional, argue for a security measure
> that you know is flawed, when there is a mailing list full of people
> willing to help figure out what a working security measure would look
> like?
>
> Yes, abandoning some of the old DOS anti-virus security model may cause
> work, but surely that will be worth it if it leads to a more secure
> system?
>
>
> [JON PRESS] I hesitate to join back in the discussion, because so far I
> haven't been very successful. I know that my level of technical depth
> is not at all the equal of just about everybody else in this
> conversation. That's not my job and I don't claim to have the answers
> to all the questions that are being raised.
>
> However... This question about arguing for a flawed security technique
> is a good one, and in a way it gets to the heart of the philosophy of
> security. Scan-on-close is admittedly incomplete as an anti-virus tool.
> But I don't agree that that make it flawed. It is part of a repertoire
> of techniques for preventing malware from residing on a machine.
>
> Let's take a simplistic and unrealistic example. Let's suppose that you
> knew that there were 20 applications on your machine that had buffer
> overflow vulnerabilities that could be exploited, and you had a fix for
> 18 of them. Would you decide not to apply the fixes, because that meant
> that there would still be 2 left -- and because theoretically there is a
> way to prevent all buffer overflows from being exploited and there is a
> mailing list full of people trying to figure out how to do it.
>
This buffer overflow risk and other equals are why LSM's exist. It
put jails around applications so they cannot do any more than they are
meant to. Its called risk reduction something that is most likely a
new idea to people coming from a windows background. It also makes
exploiting a flawed applicaiton tricky. Do something that application
should not normally do it will be blocked and trip the LSM alarm that
could be set to straight up terminate the offending application. Yes
a true shot on sign of trouble system. This is what you anti-virus
guys call behaviour monitoring same system some anti-virus software
uses to detect unknown viruses.
So 2 left should never happen since we have at least a part fix for
all of them. This is how it has to work. Failure is not a option
in our eyes. If you have a percent missed that is a failure of there
is not something else to prevent damage. We are not Windows users
with weak setup systems. We don't want weak distributions out there.
Nice if some anti-virus products started demanding a min standard LSM
to run side by side with them.
LSM's are already embedded core system exploitation prevention.
LSM currently don't extend deep enough into users to really tighten
completely down on the Users account.
So far I have not found a exact list of what is needed by AV or
Malware companies that say LSM stacking is needed. That says stacked
LSM's are needed.
So I will bring a few things to the table. Number one LSB is working
on a common packaging API using DBUS based off policy kit. So
malware application installers run in users own account and to install
into the system have to go threw a scan able interface. So far I
have not seen AV companies there working in improving the design.
Prevention beats cure.
This reduces a issue of malware to the Users own account without heavy
handed scanning.
So scanning becomes reduced to user editable files.
Linux firewall supports user mode modules so antivirus can scan
network traffic and use built in firewall to monitor for malware.
File scanning look no deeper than fusefs.
http://clamfs.sourceforge.net/. Alter the automount system to wrap
this over the top of user mounted file systems and locations of user
editable and your are done.
Now credentials patch that has not got mainline yet also provides user
mode daemon support to override filesystem permissions. Could also be
integrated into a Anti Virus Scanner. credentials is not a LSM really
its centralisation of permission information so its no longer speed
all over the kernel.
There are sections in containers as well that could cover bits..
TPM segments appear in 2.6.27 as well that will also make a core
system breach harder.
Now please layout what you need Anti-Virus Companies. Don't use
clueless desktop users as a reason. Linux could already provide the
needed interfaces just not LSM.
Now please provide a detailed list of exactly what you need Anti-virus
companies. Most likely everything you need already exists mainline
or in a development side line. Extra coders to get some of those
features mainline would be good.
Peter Dolding
PS how to I get my email on the [email protected] So it
does not bounce things to me.
-----Original Message-----
From: Peter Dolding [mailto:[email protected]]
Sent: Wednesday, August 06, 2008 8:39 AM
To: Press, Jonathan
Cc: Rik van Riel; Greg KH; Arjan van de Ven; Eric Paris;
[email protected]; [email protected]
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
linuxinterfaceforon access scanning
On Wed, Aug 6, 2008 at 10:10 PM, Press, Jonathan <[email protected]>
wrote:
> -----Original Message-----
> From: Rik van Riel [mailto:[email protected]]
> Sent: Tuesday, August 05, 2008 8:51 PM
> To: Press, Jonathan
> Cc: Greg KH; Arjan van de Ven; Eric Paris;
[email protected];
> [email protected]; [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
> linuxinterfaceforon access scanning
>
> On Tue, 5 Aug 2008 14:38:23 -0400
> "Press, Jonathan" <[email protected]> wrote:
>
>> However, I will say that while preventing infections from arriving is
>> not foolproof, doing a scan-on-close with the option to delete or
>> quarantine an infected file goes a long way.
>
> How can you, as a security professional, argue for a security measure
> that you know is flawed, when there is a mailing list full of people
> willing to help figure out what a working security measure would look
> like?
>
> Yes, abandoning some of the old DOS anti-virus security model may
cause
> work, but surely that will be worth it if it leads to a more secure
> system?
>
>
> [JON PRESS] I hesitate to join back in the discussion, because so far
I
> haven't been very successful. I know that my level of technical depth
> is not at all the equal of just about everybody else in this
> conversation. That's not my job and I don't claim to have the answers
> to all the questions that are being raised.
>
> However... This question about arguing for a flawed security
technique
> is a good one, and in a way it gets to the heart of the philosophy of
> security. Scan-on-close is admittedly incomplete as an anti-virus
tool.
> But I don't agree that that make it flawed. It is part of a
repertoire
> of techniques for preventing malware from residing on a machine.
>
> Let's take a simplistic and unrealistic example. Let's suppose that
you
> knew that there were 20 applications on your machine that had buffer
> overflow vulnerabilities that could be exploited, and you had a fix
for
> 18 of them. Would you decide not to apply the fixes, because that
meant
> that there would still be 2 left -- and because theoretically there is
a
> way to prevent all buffer overflows from being exploited and there is
a
> mailing list full of people trying to figure out how to do it.
>
This buffer overflow risk and other equals are why LSM's exist. It put
jails around applications so they cannot do any more than they are meant
to. Its called risk reduction something that is most likely a new idea
to people coming from a windows background. It also makes exploiting a
flawed applicaiton tricky. Do something that application should not
normally do it will be blocked and trip the LSM alarm that
could be set to straight up terminate the offending application. Yes a
true shot on sign of trouble system. This is what you anti-virus guys
call behaviour monitoring same system some anti-virus software uses to
detect unknown viruses.
So 2 left should never happen since we have at least a part fix for all
of them. This is how it has to work. Failure is not a option in our
eyes. If you have a percent missed that is a failure of there is not
something else to prevent damage. We are not Windows users with weak
setup systems. We don't want weak distributions out there. Nice if
some anti-virus products started demanding a min standard LSM to run
side by side with them.
LSM's are already embedded core system exploitation prevention.
LSM currently don't extend deep enough into users to really tighten
completely down on the Users account.
So far I have not found a exact list of what is needed by AV or Malware
companies that say LSM stacking is needed. That says stacked LSM's are
needed.
So I will bring a few things to the table. Number one LSB is working on
a common packaging API using DBUS based off policy kit. So malware
application installers run in users own account and to install into the
system have to go threw a scan able interface. So far I have not seen
AV companies there working in improving the design. Prevention beats
cure.
This reduces a issue of malware to the Users own account without heavy
handed scanning.
So scanning becomes reduced to user editable files.
Linux firewall supports user mode modules so antivirus can scan network
traffic and use built in firewall to monitor for malware.
File scanning look no deeper than fusefs.
http://clamfs.sourceforge.net/. Alter the automount system to wrap
this over the top of user mounted file systems and locations of user
editable and your are done.
Now credentials patch that has not got mainline yet also provides user
mode daemon support to override filesystem permissions. Could also be
integrated into a Anti Virus Scanner. credentials is not a LSM really
its centralisation of permission information so its no longer speed all
over the kernel.
There are sections in containers as well that could cover bits..
TPM segments appear in 2.6.27 as well that will also make a core system
breach harder.
Now please layout what you need Anti-Virus Companies. Don't use
clueless desktop users as a reason. Linux could already provide the
needed interfaces just not LSM.
Now please provide a detailed list of exactly what you need Anti-virus
companies. Most likely everything you need already exists mainline or
in a development side line. Extra coders to get some of those features
mainline would be good.
[JON PRESS] First of all, my examples were just hypothetical examples
designed to set a context. If they don't work, then don't take them too
literally.
Second, I'll try to put another way. The disease analogy for computer
viruses, though perhaps overused, actually goes a surprisingly long way.
The kernel guys are like medical researchers, and the AV guys (or at
least speaking for myself) are in Public Health. You guys are looking
at the human body (the computer) and looking for its low-level
vulnerabilities and ways to protect against them. We are working in the
context of imperfect human behavior that often causes people to get sick
in spite of the current state of medical knowledge, and occasionally
causes things to get out of control. Neither point of view is incorrect
-- but each attacks a very different aspect of the problem.
And this goes right to the heart of the statement "prevention beats
cure". That's true in the long run, but it's not true if you're already
sick -- or your neighbor is, and prevention is still a few years away.
Third, regarding stating what we need, we actually already had a
discussion about this several months ago, where we described this. That
discussion was the basis for Eric's proposal. To give it some
perspective, here is how I understand the flow of events. As Eric
described in his original posting this week, in the past many products
(AV, along with others that have needed access to similar events --
representatives of which have been participating in this discussion)
have done their own tricks to get at i/o events that they need to
trigger some action. Many do this by modifying the syscall table. We
were told that this was going to become an invalid operation in a coming
release of Red Hat Linux, and Red Hat offered to have a discussion with
us on what could be done to offer us another way into the same
information. This seemed like a very positive direction to go in, and
there was a discussion about what this alternative ought to look like.
On the basis of this discussion, Eric then went off and worked on the
design that he has presented.
There was probably an implicit assumption on everyone's part, including
Red Hat's, that what ought to be done was to replace the existing
syscall-based event trapping with some other interface that more or less
does the same thing in a cleaner way -- NOT to have all of the AV and
other product vendors go out and completely rethink their models. And
that's not because we inherently object to rethinking. It's really an
issue of what kind of time frame we have before a new OS goes out that
completely breaks our products. Rethinking the entire model and
developing all new products along different paradigms is not a bad thing
to do, but it takes a lot of time. Frankly we're always doing that in
some way. However, there is a more pressing problem at hand, and that's
what Eric's spec is meant to solve.
Jon
On Wed, 6 Aug 2008 08:10:53 -0400
"Press, Jonathan" <[email protected]> wrote:
> However... This question about arguing for a flawed security
> technique is a good one, and in a way it gets to the heart of the
> philosophy of security. Scan-on-close is admittedly incomplete as an
> anti-virus tool. But I don't agree that that make it flawed. It is
> part of a repertoire of techniques for preventing malware from
> residing on a machine.
this indeed gets to the heart of the problem.
several of us here are trying to argue that scan-on-close isn't very
good, BUT that if you do scan-on-change (say with inotify) you do reach
the same goal but with much better results.
notice the "but" in there. What I hope will happen is that one or more
people from the AV side (eg you and tvrtko) will actually read the "but"
part rather than just dismissing it outright because of not liking your
solution in the first part of it. Part of the answer could be "nice
however our goal is <THIS> so it won't work because of <THAT>".
At least as long as <THIS> isn't "scan on close" because that's not a
goal, that's a means.
this kind of thing seems to be indicative of the entire discussion.
For lkml proposals, both sides need to be willing to accept alternative
solutions for a problem (I know I am, just need to see why ;-).. and
explain the WHY and the goal if it's not clear.
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Wed, 6 Aug 2008 09:11:14 -0400
> There was probably an implicit assumption on everyone's part,
> including Red Hat's, that what ought to be done was to replace the
> existing syscall-based event trapping with some other interface that
> more or less does the same thing in a cleaner way -- NOT to have all
> of the AV and other product vendors go out and completely rethink
> their models. And that's not because we inherently object to
> rethinking. It's really an issue of what kind of time frame we have
> before a new OS goes out that completely breaks our products.
not writing to the syscall table hasn't been possible/allowed for..
about 5 years now. (yes I know there were still bad hacks possible
until 2 years ago). So I'm sorry, but the timeline argument doesn't
hold, you've had 5+ years of warning.
All existing RHEL products already don't allow this (I know it for the
earlier ones since I was part of the design team)... unless your
software acts entirely like a rootkit (but even then)
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Wed, 2008-08-06 at 06:49 -0700, Arjan van de Ven wrote:
> On Wed, 6 Aug 2008 09:11:14 -0400
> > There was probably an implicit assumption on everyone's part,
> > including Red Hat's, that what ought to be done was to replace the
> > existing syscall-based event trapping with some other interface that
> > more or less does the same thing in a cleaner way -- NOT to have all
> > of the AV and other product vendors go out and completely rethink
> > their models. And that's not because we inherently object to
> > rethinking. It's really an issue of what kind of time frame we have
> > before a new OS goes out that completely breaks our products.
>
> not writing to the syscall table hasn't been possible/allowed for..
> about 5 years now. (yes I know there were still bad hacks possible
> until 2 years ago). So I'm sorry, but the timeline argument doesn't
> hold, you've had 5+ years of warning.
>
> All existing RHEL products already don't allow this (I know it for the
> earlier ones since I was part of the design team)... unless your
> software acts entirely like a rootkit (but even then)
Other options involved overwriting LSM function pointers. I was told
that recently moving SELinux to be statically compiled in apparently
messed them up on that method, at least for RH products. The other
method I've heard is hunting down all of the filesystem_operations
structs and overwriting those functions. I was also told that until
recently pages marked RO could just be marked RW and then remarked RO,
although it was recently fixed to RO pages stayed RO. So yeah, I'd have
to call them all ugly rootkit like hacks.
they just keep finding uglier and uglier ways to infiltrate the kernel
which is why I was ask to try to help get a clean solution.
-Eric
On Wed, Aug 6, 2008 at 11:11 PM, Press, Jonathan <[email protected]> wrote:
> -----Original Message-----
> From: Peter Dolding [mailto:[email protected]]
> Sent: Wednesday, August 06, 2008 8:39 AM
> To: Press, Jonathan
> Cc: Rik van Riel; Greg KH; Arjan van de Ven; Eric Paris;
> [email protected]; [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
> linuxinterfaceforon access scanning
>
> On Wed, Aug 6, 2008 at 10:10 PM, Press, Jonathan <[email protected]>
> wrote:
>> -----Original Message-----
>> From: Rik van Riel [mailto:[email protected]]
>> Sent: Tuesday, August 05, 2008 8:51 PM
>> To: Press, Jonathan
>> Cc: Greg KH; Arjan van de Ven; Eric Paris;
> [email protected];
>> [email protected]; [email protected]
>> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
>> linuxinterfaceforon access scanning
>>
>> On Tue, 5 Aug 2008 14:38:23 -0400
>> "Press, Jonathan" <[email protected]> wrote:
>>
>>> However, I will say that while preventing infections from arriving is
>>> not foolproof, doing a scan-on-close with the option to delete or
>>> quarantine an infected file goes a long way.
>>
>> How can you, as a security professional, argue for a security measure
>> that you know is flawed, when there is a mailing list full of people
>> willing to help figure out what a working security measure would look
>> like?
>>
>> Yes, abandoning some of the old DOS anti-virus security model may
> cause
>> work, but surely that will be worth it if it leads to a more secure
>> system?
>>
>>
>> [JON PRESS] I hesitate to join back in the discussion, because so far
> I
>> haven't been very successful. I know that my level of technical depth
>> is not at all the equal of just about everybody else in this
>> conversation. That's not my job and I don't claim to have the answers
>> to all the questions that are being raised.
>>
>> However... This question about arguing for a flawed security
> technique
>> is a good one, and in a way it gets to the heart of the philosophy of
>> security. Scan-on-close is admittedly incomplete as an anti-virus
> tool.
>> But I don't agree that that make it flawed. It is part of a
> repertoire
>> of techniques for preventing malware from residing on a machine.
>>
>> Let's take a simplistic and unrealistic example. Let's suppose that
> you
>> knew that there were 20 applications on your machine that had buffer
>> overflow vulnerabilities that could be exploited, and you had a fix
> for
>> 18 of them. Would you decide not to apply the fixes, because that
> meant
>> that there would still be 2 left -- and because theoretically there is
> a
>> way to prevent all buffer overflows from being exploited and there is
> a
>> mailing list full of people trying to figure out how to do it.
>>
> This buffer overflow risk and other equals are why LSM's exist. It put
> jails around applications so they cannot do any more than they are meant
> to. Its called risk reduction something that is most likely a new idea
> to people coming from a windows background. It also makes exploiting a
> flawed applicaiton tricky. Do something that application should not
> normally do it will be blocked and trip the LSM alarm that
> could be set to straight up terminate the offending application. Yes a
> true shot on sign of trouble system. This is what you anti-virus guys
> call behaviour monitoring same system some anti-virus software uses to
> detect unknown viruses.
>
> So 2 left should never happen since we have at least a part fix for all
> of them. This is how it has to work. Failure is not a option in our
> eyes. If you have a percent missed that is a failure of there is not
> something else to prevent damage. We are not Windows users with weak
> setup systems. We don't want weak distributions out there. Nice if
> some anti-virus products started demanding a min standard LSM to run
> side by side with them.
>
> LSM's are already embedded core system exploitation prevention.
>
> LSM currently don't extend deep enough into users to really tighten
> completely down on the Users account.
>
> So far I have not found a exact list of what is needed by AV or Malware
> companies that say LSM stacking is needed. That says stacked LSM's are
> needed.
>
> So I will bring a few things to the table. Number one LSB is working on
> a common packaging API using DBUS based off policy kit. So malware
> application installers run in users own account and to install into the
> system have to go threw a scan able interface. So far I have not seen
> AV companies there working in improving the design. Prevention beats
> cure.
>
> This reduces a issue of malware to the Users own account without heavy
> handed scanning.
>
> So scanning becomes reduced to user editable files.
>
> Linux firewall supports user mode modules so antivirus can scan network
> traffic and use built in firewall to monitor for malware.
>
> File scanning look no deeper than fusefs.
> http://clamfs.sourceforge.net/. Alter the automount system to wrap
> this over the top of user mounted file systems and locations of user
> editable and your are done.
>
> Now credentials patch that has not got mainline yet also provides user
> mode daemon support to override filesystem permissions. Could also be
> integrated into a Anti Virus Scanner. credentials is not a LSM really
> its centralisation of permission information so its no longer speed all
> over the kernel.
>
> There are sections in containers as well that could cover bits..
>
> TPM segments appear in 2.6.27 as well that will also make a core system
> breach harder.
>
> Now please layout what you need Anti-Virus Companies. Don't use
> clueless desktop users as a reason. Linux could already provide the
> needed interfaces just not LSM.
>
> Now please provide a detailed list of exactly what you need Anti-virus
> companies. Most likely everything you need already exists mainline or
> in a development side line. Extra coders to get some of those features
> mainline would be good.
>
>
> [JON PRESS] First of all, my examples were just hypothetical examples
> designed to set a context. If they don't work, then don't take them too
> literally.
>
> Second, I'll try to put another way. The disease analogy for computer
> viruses, though perhaps overused, actually goes a surprisingly long way.
> The kernel guys are like medical researchers, and the AV guys (or at
> least speaking for myself) are in Public Health. You guys are looking
> at the human body (the computer) and looking for its low-level
> vulnerabilities and ways to protect against them. We are working in the
> context of imperfect human behavior that often causes people to get sick
> in spite of the current state of medical knowledge, and occasionally
> causes things to get out of control. Neither point of view is incorrect
> -- but each attacks a very different aspect of the problem.
We are not just kernel. http://hal.freedesktop.org/docs/PolicyKit/
Is a good read.
Long term Linux design is basically Nuke the root user and protect the
core os providing restricted access.
So anti system core damage integrates outside kernel space. We are
designing to limit how damaging that imperfect behavior is.
> And this goes right to the heart of the statement "prevention beats
> cure". That's true in the long run, but it's not true if you're already
> sick -- or your neighbor is, and prevention is still a few years away.
>
> Third, regarding stating what we need, we actually already had a
> discussion about this several months ago, where we described this. That
> discussion was the basis for Eric's proposal. To give it some
> perspective, here is how I understand the flow of events. As Eric
> described in his original posting this week, in the past many products
> (AV, along with others that have needed access to similar events --
> representatives of which have been participating in this discussion)
> have done their own tricks to get at i/o events that they need to
> trigger some action. Many do this by modifying the syscall table. We
> were told that this was going to become an invalid operation in a coming
> release of Red Hat Linux, and Red Hat offered to have a discussion with
> us on what could be done to offer us another way into the same
> information. This seemed like a very positive direction to go in, and
> there was a discussion about what this alternative ought to look like.
> On the basis of this discussion, Eric then went off and worked on the
> design that he has presented.
> There was probably an implicit assumption on everyone's part, including
> Red Hat's, that what ought to be done was to replace the existing
> syscall-based event trapping with some other interface that more or less
> does the same thing in a cleaner way -- NOT to have all of the AV and
> other product vendors go out and completely rethink their models. And
> that's not because we inherently object to rethinking. It's really an
> issue of what kind of time frame we have before a new OS goes out that
> completely breaks our products. Rethinking the entire model and
> developing all new products along different paradigms is not a bad thing
> to do, but it takes a lot of time. Frankly we're always doing that in
> some way. However, there is a more pressing problem at hand, and that's
> what Eric's spec is meant to solve.
>
You got to remember syscall trapping can be done for good and evil.
You use it to monitor actions another use it to do harm like steal data.
We need syscall trapping for debugging. ptrace and the like.
Something important you are forgetting and its critical. Windows gets
hell loads of overhead from Anti-Virus products. Our LSM's already
provide overhead. Duplication is not a way forwards.
Credentials IBM patch provides a way to override filesystem
permissions so monitor and scan can be done there.
Tracing improvements lined 2.6.27 could be used to monitor
applications a AV is suspect of.
Capabilities restrict syscalls by blocks.
Question exactly why needing to monitor at the syscall level. Really
why. There are ways to monitor and restrict accesses without
dropping to that level.
Most likely the issue is that your models are broken. Under windows
most anti-viruses copy the methods you would use to root kit a OS to
get there data. Not working with the existing secuirty model.
Learning to work side by side with the LSM and other systems already
provided anti-virus companies are going to have to learn to do. No
more of this hooking into the OS to get what you want.
Critical note on Linux. Rootkits out number viruses and malware 20
to 1. This is way different to your windows numbers. So any method
that is Rootkit style will fail in time because that hole will have to
be closed.
My main issue is TALPA, dazuko and so on of Anti-Virus Filesystem
monitoring are all going to break anyhow when
http://lwn.net/Articles/251224/ Credentials get added and common
filesystem caching gets added.
You want to change a permissions on a file/object before its opened.
So does the Credential user space daemon on file systems that cannot
store secuirty information. We only really need 1 location in the
source base for this. Expand Credentials slightly to allow anti
viruses to operate by by problem. Even better when FS-Cache can sit
on top of Credentials correctly no need for anti virus software to
have independent caching of blocked and allowed files. FS-Cache picks
a large amount of this up.
Basically TALPA, dazuko and so on of Anti-Virus Filesystem monitoring
don't fit in the future design of Linux. All they will be is
duplication of a existing interface. A interface that complete avoids
the stacking issue.
Peter Dolding
On Wed, Aug 6, 2008 at 11:55 PM, Eric Paris <[email protected]> wrote:
> On Wed, 2008-08-06 at 06:49 -0700, Arjan van de Ven wrote:
>> On Wed, 6 Aug 2008 09:11:14 -0400
>> > There was probably an implicit assumption on everyone's part,
>> > including Red Hat's, that what ought to be done was to replace the
>> > existing syscall-based event trapping with some other interface that
>> > more or less does the same thing in a cleaner way -- NOT to have all
>> > of the AV and other product vendors go out and completely rethink
>> > their models. And that's not because we inherently object to
>> > rethinking. It's really an issue of what kind of time frame we have
>> > before a new OS goes out that completely breaks our products.
>>
>> not writing to the syscall table hasn't been possible/allowed for..
>> about 5 years now. (yes I know there were still bad hacks possible
>> until 2 years ago). So I'm sorry, but the timeline argument doesn't
>> hold, you've had 5+ years of warning.
>>
>> All existing RHEL products already don't allow this (I know it for the
>> earlier ones since I was part of the design team)... unless your
>> software acts entirely like a rootkit (but even then)
>
> Other options involved overwriting LSM function pointers. I was told
> that recently moving SELinux to be statically compiled in apparently
> messed them up on that method, at least for RH products. The other
> method I've heard is hunting down all of the filesystem_operations
> structs and overwriting those functions. I was also told that until
> recently pages marked RO could just be marked RW and then remarked RO,
> although it was recently fixed to RO pages stayed RO. So yeah, I'd have
> to call them all ugly rootkit like hacks.
>
> they just keep finding uglier and uglier ways to infiltrate the kernel
> which is why I was ask to try to help get a clean solution.
>
Simplely they are following the windows way of doing things. Rootkit
the OS no one will stop us. Sorry that RootKitting is not going to
work here long term because we will fix Root Kit weaknesses. TPM from
IBM will make it even harder. Nice bits of that are in 2.6.27.
Allowing LSM stacking solves nothing. IBM's developers credentials
is a far better place for file system monitoring to hook in. Its pre
caching allows all the alterations they could want. Even better
credentials is a required patch to clean up Linux internal permission
splating all over the place. Then generic filesystem caching sits on
top of that. So 1 less cache needed. Ie the cache of passes and
failures they keep so they don't slow the system down too much.
About time they wake up Linux Different OS. We have zero tolerance of
root kits. There is a section with the design to provide the file
system monitoring and the network monitoring both without being in LSM
space. So why do they need LSM's. Cannot they just use system LSM's
to protect there anti-virus like everyone else. No need to treat
them special. If they have a issue with a LSM not being good enough
they need to speak up.
They really need to provide what they want to do to us. There issue
is they keep on looking in the wrong places and doing things the
windows way. This is not WINDOWS. In time mac and other OS's could
come tighter on root kits as well causing the programs to fail as
well. Basically get use to working with a OS that does not tolerate
secuirty flaws.
Peter Dolding
Arjan van de Ven wrote on 06/08/2008 14:44:18:
> On Wed, 6 Aug 2008 08:10:53 -0400
> "Press, Jonathan" <[email protected]> wrote:
>
> > However... This question about arguing for a flawed security
> > technique is a good one, and in a way it gets to the heart of the
> > philosophy of security. Scan-on-close is admittedly incomplete as an
> > anti-virus tool. But I don't agree that that make it flawed. It is
> > part of a repertoire of techniques for preventing malware from
> > residing on a machine.
>
> this indeed gets to the heart of the problem.
> several of us here are trying to argue that scan-on-close isn't very
> good, BUT that if you do scan-on-change (say with inotify) you do reach
> the same goal but with much better results.
>
> notice the "but" in there. What I hope will happen is that one or more
> people from the AV side (eg you and tvrtko) will actually read the "but"
> part rather than just dismissing it outright because of not liking your
> solution in the first part of it. Part of the answer could be "nice
> however our goal is <THIS> so it won't work because of <THAT>".
> At least as long as <THIS> isn't "scan on close" because that's not a
> goal, that's a means.
>
> this kind of thing seems to be indicative of the entire discussion.
> For lkml proposals, both sides need to be willing to accept alternative
> solutions for a problem (I know I am, just need to see why ;-).. and
> explain the WHY and the goal if it's not clear.
Problems with inotify as far as I know:
You can't do something like inotify("/") (made up API) but you have to set
up a watch for every directory you wan't to watch. That seems like a waste
of resources.
Then you get back a file name, if you wan't to report it or attempt* to
scan it you have to build a pathname yourself, which means you have to
maintain the whole tree of names in memory. Even bigger waste.
When I say attempt to scan it above I mean that we are back into the
pathanme teritorry. It is not guaranteed we will be able to open and scan
using that pathname. I don't know what inotify reports with chroots and
private namespaces, but it can certainly fail with NFS and root_squash. So
it is less effective as well as being resource intensive.
I think this is a good amount of flaws which shows inotify isn't really
ideal.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Tuesday 05 August 2008 11:00:31 pm Casey Schaufler wrote:
> Paul Moore wrote:
> > On Monday 04 August 2008 11:44:28 pm Casey Schaufler wrote:
> >> Cliffe wrote:
> >>> Other security schemes such as intrusion detection,
> >>> firewalls/netfilter, anti-malware, and application restrictions
> >>> (sandboxes such as jails or finer grained restrictions such as
> >>> AppArmor) could all register LSMs onto the stack.
> >>
> >> Stacking is easy for files. It's a real pain in the backside for
> >> UDP packets.
> >
> > How is it any better/worse for UDP packets than files?
>
> On delivery you'd need to decide what security scheme is actually
> available on the packet and in what order to interpret any inbound
> security data. If you had an MLS scheme that uses CIPSO, an integrity
> mechanism using IPSEC and a DAC scheme that assigns user ids by
> host address getting the ordering right and every domain registered
> properly in the networking stack would be a trick.
I still don't understand how that is any different from a file or some
other resource, local or remote. Assuming a single security label
(tag, mark, mode, etc.) on an entity on which you wish to apply an
access control decision the problem boils down to how do you
internalize the security label in such a way that it can be useful for
the security mechanism(s). In the case of a single LSM you do this
once, in the case of multiple, stacked LSMs you do this multiple times.
With multiple security markings on an entity then you have to decide if
you want to consider every marking at each LSM instance, or a subset.
The complexity in this case does go up dramatically, but I think the
key point for our discussion is that it doesn't matter if the entity is
a file or a packet.
> Plus, making sure
> that any state the security scheme requires is tricky. Maybe it's not
> actually worse if the schemes agree on what qualifies as a security
> element, but if one scheme does access control outbound while another
> does inbound it will get hairy.
Once again, these points apply equally to files as they do to packets.
--
paul moore
linux @ hp
Quoting Peter Dolding ([email protected]):
> On Wed, Aug 6, 2008 at 11:55 PM, Eric Paris <[email protected]> wrote:
> > On Wed, 2008-08-06 at 06:49 -0700, Arjan van de Ven wrote:
> >> On Wed, 6 Aug 2008 09:11:14 -0400
> >> > There was probably an implicit assumption on everyone's part,
> >> > including Red Hat's, that what ought to be done was to replace the
> >> > existing syscall-based event trapping with some other interface that
> >> > more or less does the same thing in a cleaner way -- NOT to have all
> >> > of the AV and other product vendors go out and completely rethink
> >> > their models. And that's not because we inherently object to
> >> > rethinking. It's really an issue of what kind of time frame we have
> >> > before a new OS goes out that completely breaks our products.
> >>
> >> not writing to the syscall table hasn't been possible/allowed for..
> >> about 5 years now. (yes I know there were still bad hacks possible
> >> until 2 years ago). So I'm sorry, but the timeline argument doesn't
> >> hold, you've had 5+ years of warning.
> >>
> >> All existing RHEL products already don't allow this (I know it for the
> >> earlier ones since I was part of the design team)... unless your
> >> software acts entirely like a rootkit (but even then)
> >
> > Other options involved overwriting LSM function pointers. I was told
> > that recently moving SELinux to be statically compiled in apparently
> > messed them up on that method, at least for RH products. The other
> > method I've heard is hunting down all of the filesystem_operations
> > structs and overwriting those functions. I was also told that until
> > recently pages marked RO could just be marked RW and then remarked RO,
> > although it was recently fixed to RO pages stayed RO. So yeah, I'd have
> > to call them all ugly rootkit like hacks.
> >
> > they just keep finding uglier and uglier ways to infiltrate the kernel
> > which is why I was ask to try to help get a clean solution.
> >
> Simplely they are following the windows way of doing things. Rootkit
> the OS no one will stop us. Sorry that RootKitting is not going to
> work here long term because we will fix Root Kit weaknesses. TPM from
> IBM will make it even harder. Nice bits of that are in 2.6.27.
>
> Allowing LSM stacking solves nothing. IBM's developers credentials
Seriously, it's David Howells at Redhat who's doing that. In your last
email i figured it was a typo, but credit where it's due :)
-serge
On Wed, 6 Aug 2008 15:16:02 +0100
[email protected] wrote:
> Arjan van de Ven wrote on 06/08/2008 14:44:18:
>
> > On Wed, 6 Aug 2008 08:10:53 -0400
> > "Press, Jonathan" <[email protected]> wrote:
> >
> > > However... This question about arguing for a flawed security
> > > technique is a good one, and in a way it gets to the heart of the
> > > philosophy of security. Scan-on-close is admittedly incomplete
> > > as an anti-virus tool. But I don't agree that that make it
> > > flawed. It is part of a repertoire of techniques for preventing
> > > malware from residing on a machine.
> >
> > this indeed gets to the heart of the problem.
> > several of us here are trying to argue that scan-on-close isn't very
> > good, BUT that if you do scan-on-change (say with inotify) you do
> > reach the same goal but with much better results.
> >
> > notice the "but" in there. What I hope will happen is that one or
> > more people from the AV side (eg you and tvrtko) will actually read
> > the "but" part rather than just dismissing it outright because of
> > not liking your solution in the first part of it. Part of the
> > answer could be "nice however our goal is <THIS> so it won't work
> > because of <THAT>". At least as long as <THIS> isn't "scan on
> > close" because that's not a goal, that's a means.
> >
> > this kind of thing seems to be indicative of the entire discussion.
> > For lkml proposals, both sides need to be willing to accept
> > alternative solutions for a problem (I know I am, just need to see
> > why ;-).. and explain the WHY and the goal if it's not clear.
>
> Problems with inotify as far as I know:
finally some useful discussion ;-)
>
> You can't do something like inotify("/") (made up API) but you have
> to set up a watch for every directory you wan't to watch. That seems
> like a waste of resources.
>
> Then you get back a file name, if you wan't to report it or attempt*
> to scan it you have to build a pathname yourself, which means you
> have to maintain the whole tree of names in memory. Even bigger waste.
>
> When I say attempt to scan it above I mean that we are back into the
> pathanme teritorry. It is not guaranteed we will be able to open and
> scan using that pathname. I don't know what inotify reports with
> chroots and private namespaces, but it can certainly fail with NFS
> and root_squash. So it is less effective as well as being resource
> intensive.
one argument against the namespaces one is that of security; on the one
hand I can see the argument of running the virus scanner in its own
container, on the other hand I can see the argument of not liking
cross-container access of data as a security issue by itself.
>
> I think this is a good amount of flaws which shows inotify isn't
> really ideal.
that is a fair list of complaints.. the question is can inotify be
fixed....
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Wed, Aug 06, 2008 at 12:07:57PM +0100, [email protected] wrote:
> > - Some set of requirements suddenly appears out of the void on
> > linux-kernel.
>
> Because previously it was said to go away and come back with a clear list
> of requirements. And here you make it sound like a negative thing. See
> what I am talking about?
The list of requirements you came up with was a very low-level set of
requirements. This is why Al Viro called it not much better than we
want a bunch of hooks here, here, and here.
What is needed is the high-level set of requirements --- which in the
case of security fixes, really needs to start with a threat model (or
threat models). See my previous message, where I tried to help you
guys out on this. There are scenarios such as "The Linux Desktop",
where the Clueless User may be tricked to run random binaries. Then
there is the "The Linux Fileserver", where users may upload malware
via CIFS, NFS, et. al, but there aren't any Clueless Users to start
the malware running on said Linux Fileserver, etc. When you do threat
analysis you need to know whether the malware is likely to have
compromised root (superuser) access or not. Etc.
Low-level requirements are things like "this code must take the
number, multiply by it 7, and add 42". High-level requirements answer
the question, why the heck are you trying to do this in the first
place?!?
- Ted
On Wed, Aug 06, 2008 at 08:10:53AM -0400, Press, Jonathan wrote:
> I think if it as being like the Sieve of Eratosthenes. The further down
> you go, the more numbers drop out. In AV scanning, each step of the way
> removes some percentage of the harmful files, and closes the window of
> time that they have to operate or migrate. Or maybe it's like spraying
> insecticide when there is an outbreak of some deadly mosquito-borne
> illness. Without getting into the political issues about spraying,
> because this is JUST AN EXAMPLE -- would you not spray because 5% of the
> bugs would still be left behind? Wouldn't you then spray again, because
> you wipe out another 95%?
The problem with your example is that it ignores the cost; the cost in
code maintenance; the cost in performance, etc. That's the problem an
absolutist view towards security. Going back to the your sparying
analogy, if the illness is considered *so* utterly deadly that you
don't consider the costs of beneficial insects dieing, children
getting exposed so badly that they get cancer five years later,
etc. --- then the argument would be heck, let's spray every day!
Let's spray every hour! Let's have a insectside misters going 24
hours a day in the parks and in the schools!!!
In the TSA example, let's force every single traveller to strip naked
publically and be submitted to body cavity searches! Since
**obviously** stopping terrorist bombs is so important that no other
considerations need to be taken into account. Oh, and we should
obviously also give all of our financial information to the security
agencies so they can do futher screens to look for terrorists; who
cares about the risks that laptops with all of that unencrypted data
will be stolen out of a locked office in the San Francisco airport?
Similarly there are costs to doing all of this extra scanning. You're
getting carried away here way you say that it never hurts to do extra
scanning, and that we don't need to decide whether or not it makes
sense to do it all. That's just stupid. The whole defense in depth,
taken to extremes, leads to completely nonsensical thinking. Security
is *defintiely* a cost/benefit tradeoff, and to do something
meaningful here we need to think rationally about the threat
environment --- and part of that threat environment is the existing
security systems in Linux, which are definitely far more powerful than
what DOS/Windows have.
- Ted
On Wed, Aug 06, 2008 at 11:05:43AM +0100, [email protected] wrote:
> Greg KH wrote on 05/08/2008 21:26:21:
>
> > > [JON PRESS] I wouldn't call it lazy, actually. It's more like
> > > "economical" or "ergonomic" -- or, dare I say it -- "user-friendly."
> In
> > > this case, the users are the AV vendors who will have to write to the
> > > API that will come out of this spec. We will be more inclined to
> > > appreciate the SDK (for want of a better term) if it covers all the
> > > bases, rather than force us to go elsewhere for some of our
> > > requirements. When we write SDKs, we try to make sure that our users
> > > will find whatever they need.
> >
> > But realize that you are adding an overhead on us, the kernel community,
> > to make your life easier. We are the ones that are taking our time to
> > review and comment on this code. We are the ones who will have to live
> > with this code for forever, and maintain it over the lifetime of linux.
> > So far, you all have shown no willingness to give anything back to us at
> > all.
>
> We all? How is that true? I for example wrote some code and am willing to
> help maintain it if it gets accepted.
You did? I didn't see any patches from you here, did I just happen to
miss them?
If Eric's code was based on yours, that's great, but it didn't imply
maintenance, especially as it doesn't look like your original patches
much :)
thanks,
greg k-h
On Wed, 2008-08-06 at 08:17 -0700, Greg KH wrote:
> On Wed, Aug 06, 2008 at 11:05:43AM +0100, [email protected] wrote:
> > Greg KH wrote on 05/08/2008 21:26:21:
> >
> > > > [JON PRESS] I wouldn't call it lazy, actually. It's more like
> > > > "economical" or "ergonomic" -- or, dare I say it -- "user-friendly."
> > In
> > > > this case, the users are the AV vendors who will have to write to the
> > > > API that will come out of this spec. We will be more inclined to
> > > > appreciate the SDK (for want of a better term) if it covers all the
> > > > bases, rather than force us to go elsewhere for some of our
> > > > requirements. When we write SDKs, we try to make sure that our users
> > > > will find whatever they need.
> > >
> > > But realize that you are adding an overhead on us, the kernel community,
> > > to make your life easier. We are the ones that are taking our time to
> > > review and comment on this code. We are the ones who will have to live
> > > with this code for forever, and maintain it over the lifetime of linux.
> > > So far, you all have shown no willingness to give anything back to us at
> > > all.
> >
> > We all? How is that true? I for example wrote some code and am willing to
> > help maintain it if it gets accepted.
>
> You did? I didn't see any patches from you here, did I just happen to
> miss them?
>
> If Eric's code was based on yours, that's great, but it didn't imply
> maintenance, especially as it doesn't look like your original patches
> much :)
First copyrights in all those files are Sophos. Tvrtko is where the
basis for most of these code came from. Although ongoing maintenance
has not been discussed, he and Sophos deffinitely get the credit for
working on code.
-Eric
On Wed, Aug 06, 2008 at 03:16:02PM +0100, [email protected] wrote:
>
> You can't do something like inotify("/") (made up API) but you have to set
> up a watch for every directory you wan't to watch. That seems like a waste
> of resources.
>
> Then you get back a file name, if you wan't to report it or attempt* to
> scan it you have to build a pathname yourself, which means you have to
> maintain the whole tree of names in memory. Even bigger waste.
Yes, it would be nice if inotify gave you back a full pathname and
where a single watch would return all changes anywhere in the
filesystem tree. I'd recommend that folks try to create such a patch.
> When I say attempt to scan it above I mean that we are back into the
> pathanme teritorry. It is not guaranteed we will be able to open and scan
> using that pathname. I don't know what inotify reports with chroots and
> private namespaces, but it can certainly fail with NFS and root_squash. So
> it is less effective as well as being resource intensive.
Linux's namespace support does break a lot of traditional paradigm.
I'll note the TALPA "requirements" are broken themselves since they
refer to pathnames.
Furthermore, I assume you'll always want to do the scanning in
userspace; the virus signature files for Windows are ***huge***. And
if you are going to be scanning for Windows virii on the argument that
you want to stop malware on fileservers, I don't think you want to put
all of that code into the kernel. (I'll note that all that code
complexity leads to bugs, which will in kernel code cause system
crashes. One company's Linux AV code --- I won't say which --- almost
lead to a rather big and public customer abandoning an Linux
deployment because said proprietary, badly/disastrously written,
kernel code was leaking a small amount of memory on every file open,
and no one could figure out why their file server was crashing every
five days or so. I was called in to rescue said customer before they
cancelled the contract in disgust, and I traced it back to a
proprietary AV kernel module. What fun...)
So if we are going to have to deal with namespaces, I suspect the best
we can do for any interface (whether it is inotify based or not) is to
have it return pathnames that are valid in the namespace that the
program calling said interface happens to be running in. If necessary
the AV program can be given access to a highly privileged namespace
where all mounts are visible. And if you want to restrict namespaces
from being created at all, that's a security policy decision that
should be made via the LSM hooks.
As far as blocking opens are concerned, my suggestion there would be
changes would probably be much more likely accepted if they solved
more problems than just what the AV folks need. For example, think
about hierarchical storage management, and DMAPI. DMAPI is a total
disaster because it doesn't know about namespaces and so is completely
pathname based (which doesn't work well when you have namespaces).
But a solution which is general enough that it can also be used to
support HSM would probably be a good thing.
Also, it may very well be that instead of one, purpose-specific
interface such as what you suggested in TALPA, it might be much better
if it was a series of different interfaces; and in some cases, some of
the changes might be extensions and improvments to existing
facilities, such inotify.
Regards,
- Ted
Peter Dolding wrote:
> My main issue is TALPA, dazuko and so on of Anti-Virus Filesystem
> monitoring are all going to break anyhow when
> http://lwn.net/Articles/251224/ Credentials get added and common
> filesystem caching gets added.
>
> You want to change a permissions on a file/object before its opened.
> So does the Credential user space daemon on file systems that cannot
> store secuirty information. We only really need 1 location in the
> source base for this. Expand Credentials slightly to allow anti
> viruses to operate by by problem. Even better when FS-Cache can sit
> on top of Credentials correctly no need for anti virus software to
> have independent caching of blocked and allowed files. FS-Cache picks
> a large amount of this up.
>
> Basically TALPA, dazuko and so on of Anti-Virus Filesystem monitoring
> don't fit in the future design of Linux. All they will be is
> duplication of a existing interface. A interface that complete avoids
> the stacking issue.
Then in a real sense you've solved much of their problem for them (;-))
After this comes engineering, so that they can re-use the scanning
mechanisms they already have, but from a different caller.
The requirements are probably that they know
- is this an open for read or write (somewhat less time-sensitive)?
- is the data present, or do we have to wait?
- if so, for what?
as of the time they start looking at the file. Having a race-free
mechanism using credentials and RCU is, IMHO, A Really Good Thing.
Another thing they and we will likely need is a way to discover
if a file is inacessable due to an AV operation, and if the time
it has been inacessable is less than or equal to a scanning
upper bound by file size or beyond it. The latter is for repair
of broken state introduced by the AV process failing.
--dave
>
> Peter Dolding
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
--
David Collier-Brown | Always do right. This will gratify
Sun Microsystems, Toronto | some people and astonish the rest
[email protected] | -- Mark Twain
cell: (647) 833-9377, bridge: (877) 385-4099 code: 506 9191#
-----Original Message-----
From: Theodore Tso [mailto:[email protected]]
Sent: Wednesday, August 06, 2008 11:09 AM
To: Press, Jonathan
Cc: Rik van Riel; [email protected];
[email protected]; [email protected];
Arjan van de Ven
Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
alinuxinterfaceforon access scanning
On Wed, Aug 06, 2008 at 08:10:53AM -0400, Press, Jonathan wrote:
> I think if it as being like the Sieve of Eratosthenes. The further
down
> you go, the more numbers drop out. In AV scanning, each step of the
way
> removes some percentage of the harmful files, and closes the window of
> time that they have to operate or migrate. Or maybe it's like
spraying
> insecticide when there is an outbreak of some deadly mosquito-borne
> illness. Without getting into the political issues about spraying,
> because this is JUST AN EXAMPLE -- would you not spray because 5% of
the
> bugs would still be left behind? Wouldn't you then spray again,
because
> you wipe out another 95%?
The problem with your example is that it ignores the cost; the cost in
code maintenance; the cost in performance, etc. That's the problem an
absolutist view towards security. Going back to the your sparying
analogy, if the illness is considered *so* utterly deadly that you don't
consider the costs of beneficial insects dieing, children getting
exposed so badly that they get cancer five years later, etc. --- then
the argument would be heck, let's spray every day! Let's spray every
hour! Let's have a insectside misters going 24 hours a day in the parks
and in the schools!!!
In the TSA example, let's force every single traveller to strip naked
publically and be submitted to body cavity searches! Since
**obviously** stopping terrorist bombs is so important that no other
considerations need to be taken into account. Oh, and we should
obviously also give all of our financial information to the security
agencies so they can do futher screens to look for terrorists; who cares
about the risks that laptops with all of that unencrypted data will be
stolen out of a locked office in the San Francisco airport?
Similarly there are costs to doing all of this extra scanning. You're
getting carried away here way you say that it never hurts to do extra
scanning, and that we don't need to decide whether or not it makes sense
to do it all. That's just stupid. The whole defense in depth, taken to
extremes, leads to completely nonsensical thinking. Security is
*defintiely* a cost/benefit tradeoff, and to do something meaningful
here we need to think rationally about the threat environment --- and
part of that threat environment is the existing security systems in
Linux, which are definitely far more powerful than what DOS/Windows
have.
[JON PRESS] You're absolutely right, there is a cost-benefit tradeoff,
and I am not ignoring it at all. Reasonable people can have a
reasonable discussion about what constitutes enough or too much
scanning. From my experience with customers, many if not most
enterprises consider any outbreak with any serious damage to be too
much, and don't mind redundant scanning if it will help to prevent it.
Even so, I don't think your extreme examples are really parallel to what
we do. Personally, I think that scanning on open, exec and close is not
excessive.
And in fact, we do go out of our way to avoid scanning when it really
isn't necessary. For example, that's the reason that we want a cache --
and again reasonable people can have a reasonable discussion about how
to handle that.
On Wed, Aug 06, 2008 at 10:24:21AM +0100, [email protected] wrote:
> Greg KH wrote on 05/08/2008 19:11:41:
>
> > On Tue, Aug 05, 2008 at 02:04:26PM -0400, Press, Jonathan wrote:
> > > I'm not sure if this is off the direct idea of this thread, or if I am
> > > possibly missing the whole point.
> >
> > I think you might be missing the point a bit here, as the traditional
> > Unix model that Linux has prevents much of what the "traditional AV"
> > products need to do, right?
>
> Could you please explain some more what and how do you think Unix model
> prevents?
It prevents any user from overwriting an existing executable, and it
prevents any user from adding an executable file to a common directory
(/usr/bin).
It also prevents any user from overwriting a different user's data file.
What specific threat model are you feeling is still present on Linux
today that this proposal is supposed to address?
thanks,
greg k-h
On Wed, Aug 06, 2008 at 10:37:06AM +0100, [email protected] wrote:
> Greg KH wrote on 05/08/2008 21:15:35:
>
> > > > > Perf win, why bothering looking for malware in /proc when it can't
> > > > > exist? It doesn't take longer it just takes time having to do
> > > > >
> > > > > userspace -> kernel -> userspace -> kernel -> userspace
> > > > >
> > > > > just to cat /proc/mounts, all of this could probably be alliviated
> if we
> > > > > cached access on non block backed files but then we have to come
> up with
> > > > > a way to exclude only nfs/cifs. I'd rather list the FSs that
> don't need
> > > > > scanning every time than those that do....
> > > >
> > > > How long does this whole process take? Seriously is it worth the
> added
> > > > kernel code for something that is not measurable?
> > >
> > > Is it worth having 2 context switches for every open when none are
> > > needed? I plan to get numbers on that.
> >
> > Compared to the real time it takes in the "virus engine"? I bet it's
> > totally lost in the noise. Those things are huge beasts with thousands
> > to hundreds of thousands of context switches.
>
> No, because we are talking about a case here where we don't want to do any
> scanning. We want to detect if it is procfs (for example) as quickly as
> possible and don't do anything. Same goes for any other filesystem where
> it is not possible to store arbitrary user data.
See previous messages about namespaces and paths for trying to figure this
kind of information out in a sane way within the kernel.
> > > > > In kernel caching is clearly a huge perf win.
> > > >
> > > > Why? If the cache is also in userspace, it should be the same,
> right?
> > >
> > > In kernel cache has 0 context switches for every open. Userspace
> > > caching has 2. Every open has to block, switch to the context of the
> > > userspace client/cache, get that decisions, and then switch back to
> the
> > > original process.
> >
> > Again, compared to what? If you in userspace are doing big complex
> > things, such an overhead is trivial.
>
> Again similar thing as above - In case of a cache we are not doing complex
> things.
Except for the overhead of keeping a cache :)
> So I think you can't argue that because scanning is slow everything
> else has to go to userspace. On a typical running system scanning is
> exceptional and everything else benefits from being in the fast path.
I really can not judge as we have not seen an implementation yet.
thanks,
greg k-h
"Press, Jonathan" <[email protected]> wrote:
>>but close is... very limited in value. Open is a discrete event
>>traditionally associated withh permission checks.
>>Close... not so. (And if you mmap memory, you can then close the file
>>and still write to it via the mmap)
Eric Paris wrote:
> I think we all agree that open is the most interesting time for scanning
> operations, but as Jonathan points out there is some value (even if not
> perfect value) in looking at closes as well.
Open for read is the "traditional" time for scanning, but the
sequence (open for write) -> change -> (time passes or close happens)
is specifically a good time to do content checking, so as to have the
answer to the check available for the open for read.
I'd suggest "read" and "write" are the two cases that are interesting,
and that we've been using 'open" an "close" for a not very good
approximation to them (;-))
--dave
--
David Collier-Brown | Always do right. This will gratify
Sun Microsystems, Toronto | some people and astonish the rest
[email protected] | -- Mark Twain
cell: (647) 833-9377, bridge: (877) 385-4099 code: 506 9191#
On Wed, 2008-08-06 at 08:25 -0700, Greg KH wrote:
> On Wed, Aug 06, 2008 at 10:37:06AM +0100, [email protected] wrote:
> > Greg KH wrote on 05/08/2008 21:15:35:
> >
> > > > > > Perf win, why bothering looking for malware in /proc when it can't
> > > > > > exist? It doesn't take longer it just takes time having to do
> > > > > >
> > > > > > userspace -> kernel -> userspace -> kernel -> userspace
> > > > > >
> > > > > > just to cat /proc/mounts, all of this could probably be alliviated
> > if we
> > > > > > cached access on non block backed files but then we have to come
> > up with
> > > > > > a way to exclude only nfs/cifs. I'd rather list the FSs that
> > don't need
> > > > > > scanning every time than those that do....
> > > > >
> > > > > How long does this whole process take? Seriously is it worth the
> > added
> > > > > kernel code for something that is not measurable?
> > > >
> > > > Is it worth having 2 context switches for every open when none are
> > > > needed? I plan to get numbers on that.
> > >
> > > Compared to the real time it takes in the "virus engine"? I bet it's
> > > totally lost in the noise. Those things are huge beasts with thousands
> > > to hundreds of thousands of context switches.
> >
> > No, because we are talking about a case here where we don't want to do any
> > scanning. We want to detect if it is procfs (for example) as quickly as
> > possible and don't do anything. Same goes for any other filesystem where
> > it is not possible to store arbitrary user data.
>
> See previous messages about namespaces and paths for trying to figure this
> kind of information out in a sane way within the kernel.
Didn't I already go over this? The patch for FS exclusions would not be
namespace based, rather dentry->d_inode->i_sb->fstype->name matching.
Lets not start name based discussions at this point in time. Those
patches weren't proposed on this go and reading the write up of both of
the name based items (I think number 11 and 12) one I outright reject
and the other will require future discussion.
-Eric
On Wed, 6 Aug 2008 11:33:23 -0400
"Press, Jonathan" <[email protected]> wrote:
> Even so, I don't think your extreme examples are really parallel to what
> we do. Personally, I think that scanning on open, exec and close is not
> excessive.
>
> And in fact, we do go out of our way to avoid scanning when it really
> isn't necessary. For example, that's the reason that we want a cache --
Disks are slow and files are getting larger by the day.
We can do a lot better than scanning a whole file. A mechanism
that can notify programs about what file changed and what byte
range in the file changed can reduce scanning overhead by only
needing to scan the part of the file that changed.
More importantly, getting info on which bytes in a file changed
will also help backup programs and disk indexing programs.
Blocking on open can be useful for more than just anti-virus
scanning. It can also be useful for hierarchical storage
management or simply for using a filesystem while it is being
restored from backup - an important consideration because
restoring a filesystem from backup can take days with modern
data sizes.
Blocking on exec is similar to blocking on open and useful
in the same scenarios.
What we need to work on is making sure that the interfaces
that go into the kernel are useful not just for anti-virus
programs, but also for other software.
--
All Rights Reversed
Theodore Tso <[email protected]> wrote on 06/08/2008 16:22:36:
> On Wed, Aug 06, 2008 at 03:16:02PM +0100, [email protected]
wrote:
> >
> > You can't do something like inotify("/") (made up API) but you have to
set
> > up a watch for every directory you wan't to watch. That seems like a
waste
> > of resources.
> >
> > Then you get back a file name, if you wan't to report it or attempt*
to
> > scan it you have to build a pathname yourself, which means you have to
> > maintain the whole tree of names in memory. Even bigger waste.
>
> Yes, it would be nice if inotify gave you back a full pathname and
> where a single watch would return all changes anywhere in the
> filesystem tree. I'd recommend that folks try to create such a patch.
I will have a brief look to familiarise myself with inotify
implementation. So maybe, but even if that solves only one
sub-requirement.
> > When I say attempt to scan it above I mean that we are back into the
> > pathanme teritorry. It is not guaranteed we will be able to open and
scan
> > using that pathname. I don't know what inotify reports with chroots
and
> > private namespaces, but it can certainly fail with NFS and
root_squash. So
> > it is less effective as well as being resource intensive.
>
> Linux's namespace support does break a lot of traditional paradigm.
> I'll note the TALPA "requirements" are broken themselves since they
> refer to pathnames.
Core functionality does not depend on pathnames. I thought that was
sufficiently clear from the design description. But read below...
> Furthermore, I assume you'll always want to do the scanning in
> userspace; the virus signature files for Windows are ***huge***. And
You assume correctly so the rest of the paragraph was not needed. :)
> if you are going to be scanning for Windows virii on the argument that
> you want to stop malware on fileservers, I don't think you want to put
> all of that code into the kernel. (I'll note that all that code
> complexity leads to bugs, which will in kernel code cause system
> crashes. One company's Linux AV code --- I won't say which --- almost
> lead to a rather big and public customer abandoning an Linux
> deployment because said proprietary, badly/disastrously written,
> kernel code was leaking a small amount of memory on every file open,
> and no one could figure out why their file server was crashing every
> five days or so. I was called in to rescue said customer before they
> cancelled the contract in disgust, and I traced it back to a
> proprietary AV kernel module. What fun...)
Once upon a time I worked for a different company and we used embedded
linux to drive some custom hardware, rather complex things. On some
customer sites, every week or so the OS would hang. Some free, public and
open kernel GPL code was leaking kernel memory on each USB transaction and
depending on the use it would use up all memory sooner or later. We lost
the customer, but didn't abandon Linux. Instead we helped fix the leak and
if you don't believe me it should be ChangeLogs under my name, something
like five or six years ago.
So what do ours anecdotes prove? Only lack of testing I would say.
Also, that companies from your example bad code would probably be better
if a proper interface did exist and they didn't have to hack around.
> So if we are going to have to deal with namespaces, I suspect the best
> we can do for any interface (whether it is inotify based or not) is to
> have it return pathnames that are valid in the namespace that the
> program calling said interface happens to be running in. If necessary
> the AV program can be given access to a highly privileged namespace
> where all mounts are visible. And if you want to restrict namespaces
> from being created at all, that's a security policy decision that
> should be made via the LSM hooks.
I agree wih the first part and that is how it works.
Pathnames are used for reporting purposes and for possible filesystem
hiearchy exclusions. For reporting it is obviously not critical from
security point of view and the design document showed how we get them.
There is no new code added to do it and it happens from userspace. Such
pathnames are relative to the userspace daemon since they are obtained via
/proc/self/fd/ and readlink. Exclusions use relative paths and that is
also explained in the document.
Therefore access to all namespaces should not be needed except in a way of
getting a file descriptor from another process, possibly from another
namespace, do to the scan.
> As far as blocking opens are concerned, my suggestion there would be
> changes would probably be much more likely accepted if they solved
> more problems than just what the AV folks need. For example, think
> about hierarchical storage management, and DMAPI. DMAPI is a total
> disaster because it doesn't know about namespaces and so is completely
> pathname based (which doesn't work well when you have namespaces).
> But a solution which is general enough that it can also be used to
> support HSM would probably be a good thing.
I agree and think we are completely open to this. However the story
reverses in a way that we now don't know what are the requirements for
those so couldn't possibly address them initially.
> Also, it may very well be that instead of one, purpose-specific
> interface such as what you suggested in TALPA, it might be much better
> if it was a series of different interfaces; and in some cases, some of
> the changes might be extensions and improvments to existing
> facilities, such inotify.
Could be, but I think we won't know for sure until everything is fleshed
out.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Greg KH <[email protected]> wrote on 06/08/2008 16:25:55:
> On Wed, Aug 06, 2008 at 10:37:06AM +0100, [email protected]
wrote:
> > Greg KH wrote on 05/08/2008 21:15:35:
> >
> > > > > > Perf win, why bothering looking for malware in /proc when it
can't
> > > > > > exist? It doesn't take longer it just takes time having to do
> > > > > >
> > > > > > userspace -> kernel -> userspace -> kernel -> userspace
> > > > > >
> > > > > > just to cat /proc/mounts, all of this could probably be
alliviated
> > if we
> > > > > > cached access on non block backed files but then we have to
come
> > up with
> > > > > > a way to exclude only nfs/cifs. I'd rather list the FSs that
> > don't need
> > > > > > scanning every time than those that do....
> > > > >
> > > > > How long does this whole process take? Seriously is it worth
the
> > added
> > > > > kernel code for something that is not measurable?
> > > >
> > > > Is it worth having 2 context switches for every open when none are
> > > > needed? I plan to get numbers on that.
> > >
> > > Compared to the real time it takes in the "virus engine"? I bet
it's
> > > totally lost in the noise. Those things are huge beasts with
thousands
> > > to hundreds of thousands of context switches.
> >
> > No, because we are talking about a case here where we don't want to do
any
> > scanning. We want to detect if it is procfs (for example) as quickly
as
> > possible and don't do anything. Same goes for any other filesystem
where
> > it is not possible to store arbitrary user data.
>
> See previous messages about namespaces and paths for trying to figure
this
> kind of information out in a sane way within the kernel.
How are namespaces and pathnames relevant to exclusions based on
filesystem *type*? It is not about checking for /proc but checking if the
filesystem name is proc, sysfs, ..
> > > > > > In kernel caching is clearly a huge perf win.
> > > > >
> > > > > Why? If the cache is also in userspace, it should be the same,
> > right?
> > > >
> > > > In kernel cache has 0 context switches for every open. Userspace
> > > > caching has 2. Every open has to block, switch to the context of
the
> > > > userspace client/cache, get that decisions, and then switch back
to
> > the
> > > > original process.
> > >
> > > Again, compared to what? If you in userspace are doing big complex
> > > things, such an overhead is trivial.
> >
> > Again similar thing as above - In case of a cache we are not doing
complex
> > things.
>
> Except for the overhead of keeping a cache :)
As small as possible - just one extra field in the inode struct.
And really you are talking about a different thing again and moving the
argument around. I answered you why your argument about putting a cache in
userspace since we do so much there already is wrong, because we don't do
a lot (anything really) in userspace for vast majority of opens if we have
this totally simple in-kernel caching scheme.
> > So I think you can't argue that because scanning is slow everything
> > else has to go to userspace. On a typical running system scanning is
> > exceptional and everything else benefits from being in the fast path.
>
> I really can not judge as we have not seen an implementation yet.
Did you mean "_I_ have not seen an implementation yet"? Because Eric
posted it so you can have a look at your leisure.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Rik van Riel wrote on 06/08/2008 16:46:04:
> On Wed, 6 Aug 2008 11:33:23 -0400
> "Press, Jonathan" <[email protected]> wrote:
>
> > Even so, I don't think your extreme examples are really parallel to
what
> > we do. Personally, I think that scanning on open, exec and close is
not
> > excessive.
> >
> > And in fact, we do go out of our way to avoid scanning when it really
> > isn't necessary. For example, that's the reason that we want a cache
--
>
> Disks are slow and files are getting larger by the day.
>
> We can do a lot better than scanning a whole file. A mechanism
> that can notify programs about what file changed and what byte
> range in the file changed can reduce scanning overhead by only
> needing to scan the part of the file that changed.
It is much more advanced than that, really. I don't know if ever a whole
file is read and in 99% it is just a tiny part of it. I don't know what I
am allowed to disclose and also it is not my area of expertise, but if you
are interested in how detection actually works maybe we can talk off list
and put you in touch with some other people here.
It is also wrong to think that you can scan only what has changes because
that bit may be harmless itself but present a final part of a malware
puzzle.
> More importantly, getting info on which bytes in a file changed
> will also help backup programs and disk indexing programs.
True, but Nick mentioned some huge issues with access after close and
munmap in one of your previous postings. It sounds to me that would be a
huge VM/filesystem work to actually enable things like this.
> What we need to work on is making sure that the interfaces
> that go into the kernel are useful not just for anti-virus
> programs, but also for other software.
I definitely agree with that.
Tvrtko
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Wed, 6 Aug 2008 17:12:56 +0100
[email protected] wrote:
> Rik van Riel wrote on 06/08/2008 16:46:04:
> > More importantly, getting info on which bytes in a file changed
> > will also help backup programs and disk indexing programs.
>
> True, but Nick mentioned some huge issues with access after close and
> munmap in one of your previous postings. It sounds to me that would be a
> huge VM/filesystem work to actually enable things like this.
A universally useful large change is often easier to get into
the Linux kernel than a single-purpose hack.
--
All Rights Reversed
On Wednesday 06 August 2008 21:29, [email protected] wrote:
> Nick Piggin <[email protected]> wrote on 06/08/2008 12:10:58:
> > > code so what am I missing?
> >
> > Maybe... but that's not the same as what requirement 5 calls for.
>
> I see what you mean, it should have been worded better. Nevertheless that
> is what was intended by it - to enable caching only on filesystems where
> it is safe to do so.
That is exactly what you cannot do. Otherwise you have to rewrite
filesystems not to use the pagecache, and probably have to do disallow
or do some horrible hacks to support mmap.
The requirement you actually intend is more like what you said before,
along the lines of "file can't be changed in given set of circumstances".
> > But depending on exactly what semantics you really call for, it can get
> > tricky to account for all of pagecache. Writes can happen through page
> > tables or get_user_pages. True that a process has to at some point have
> > write permission to the file, but the cache itself could be modified
> > even after the file is closed and all mmaps disappear.
>
> I don't have a very good understanding of the VM subsystem I must admit.
> So in other words with the current kernel file modification time is not
> necessarily correct - it represents when the file was last opened for
> modification, not when it was actually modified? (While those two points
> in time can be arbitrarily separated)
It is a best effort thing. When you start mmapping, that allows
get_user_pages, which in turn may be able to take pages and modify
them at fairly arbitrary points in time.
> How would I use those methods for file modification? I am curious to make
> a test case..
Even just with mmap (we do a much better job of it now, but) you
can probably write to a file after mtime if you write to already
dirty pages.
On Wed, 06 Aug 2008 04:41:04 +0200
Andi Kleen <[email protected]> wrote:
> Christoph Hellwig <[email protected]> writes:
>
> > I couldn't give a rats a** about
> > windows viruses as they can't actually cause any harm on a Linux
> > machine.
>
> My guess is that wine is good enough these days to run many Windows
> worms actually.
Firefox, Open Office and other applications are probably large and
feature rich enough that making worms for them might be possible.
Not being able to harm the operating system is not the same as not
being able to harm the user. We still want some protection from
the latter...
--
All Rights Reversed
On Wed, 2008-08-06 at 12:25 -0400, Rik van Riel wrote:
> On Wed, 6 Aug 2008 17:12:56 +0100
> [email protected] wrote:
> > Rik van Riel wrote on 06/08/2008 16:46:04:
>
> > > More importantly, getting info on which bytes in a file changed
> > > will also help backup programs and disk indexing programs.
> >
> > True, but Nick mentioned some huge issues with access after close and
> > munmap in one of your previous postings. It sounds to me that would be a
> > huge VM/filesystem work to actually enable things like this.
>
> A universally useful large change is often easier to get into
> the Linux kernel than a single-purpose hack.
Over engineering in a void for some supposed someday user isn't a good
idea either. If there are other people interested in working together
to come up with a solution to multiple needs I'm here to help!
-Eric
On Tue, 2008-08-05 at 13:30 -0700, Greg KH wrote:
> On Tue, Aug 05, 2008 at 02:56:42PM -0400, Eric Paris wrote:
> > So you are arguing against the defense in depth theory? LSM should
> > solve it all so why bother?
>
> No, I am saying I have yet to see a real requirement to justify that
> this code goes into the kernel.
While I try to get someone to talk to me, lets consider even the most
simplistic threat model. I'm not stating that this is the only threat
that vendors wish to protect against, but I think we can all agree it is
a meaningful threat. Trying to get the AV vendors to talk about their
products actual uses is like pulling teeth.
This simple thread shows what I believe to be clear and compelling
evidence of the need for an in kernel solution. Lets just consider that
we are a high input, high output, NFS file server with other OS's
mounting this NFS share RW.
Our goal is to stop, or at least reduce the throughput (I clearly
document and accept the open to read race, and until we get a working
revoke I don't see that changing) of malware across the NFS server.
This data will not be attacking the NFS server. We wish to slow and
hopefully halt the spread of this data with minimal impact to the NFS
server.
Since the NFS server interacts directly with the VFS no purely userspace
solution is possible. I certainly hope noone tells me we should hook
directly into nfsd and do everything else in userspace :)
The purpose of this e-mail is not to ask others to find 'the perfect
solution for this one example' but merely to holp move away from the
ideas that this can be done in GLIBC or in LD_PRELOAD.
Maybe my code is all crap, I'm more than willing to accept that, it
wouldn't be the first time, but I do see the open on the NFS server and
I could do "something" with that file on the server. I can continue to
define lots of other threats that my interface is useful for (even if
Greg dismissed one of them out of hand) but right now I'm trying to get
all of the AV vendors to define what it is they do (and I think we all
know they will claim to do exactly this)
-Eric
On Wed, Aug 06, 2008 at 02:49:57PM -0400, Eric Paris wrote:
>
> This simple thread shows what I believe to be clear and compelling
> evidence of the need for an in kernel solution. Lets just consider that
> we are a high input, high output, NFS file server with other OS's
> mounting this NFS share RW.
>
> Our goal is to stop, or at least reduce the throughput (I clearly
> document and accept the open to read race, and until we get a working
> revoke I don't see that changing) of malware across the NFS server.
> This data will not be attacking the NFS server. We wish to slow and
> hopefully halt the spread of this data with minimal impact to the NFS
> server.
In this scenario, are you positing that you are worried about Windows
malware, or Linux malware? What OS are the clients running? I will
note that Windows has such a sucky NFS implementation that nearly all
Widows clients will be running CIFS/SMB, not NFS --- so the right
solution there is to integrate the virus checking with Samba ---
especially since the one AV vendor has already admitted the actual
virus signature checking has to be done in userspace.
For Linux clients, one question that immediately rises is the
end-to-end argument. Wouldn't be far better to run whatever security
solution on the client? After all, a Virus checking on an NFS server
isn't going to help the user if they accidentally track in the virus
on a USB stick. (Especially if it is an infected Macro virus in an
office document.)
- Ted
On Wed, 2008-08-06 at 17:02 -0400, Theodore Tso wrote:
> On Wed, Aug 06, 2008 at 02:49:57PM -0400, Eric Paris wrote:
> >
> > This simple thread shows what I believe to be clear and compelling
> > evidence of the need for an in kernel solution. Lets just consider that
> > we are a high input, high output, NFS file server with other OS's
> > mounting this NFS share RW.
> >
> > Our goal is to stop, or at least reduce the throughput (I clearly
> > document and accept the open to read race, and until we get a working
> > revoke I don't see that changing) of malware across the NFS server.
> > This data will not be attacking the NFS server. We wish to slow and
> > hopefully halt the spread of this data with minimal impact to the NFS
> > server.
>
> In this scenario, are you positing that you are worried about Windows
> malware, or Linux malware? What OS are the clients running? I will
> note that Windows has such a sucky NFS implementation that nearly all
> Widows clients will be running CIFS/SMB, not NFS
I believe I specifically did not make any such claims at all about the
client OS and merely claimed the intended target was not the linux NFS
server. I didn't make those claims because they are irrelevant and so
that people could not jump on those details and try to offload the
solution to the wrong place. Maybe the client is not Windows but
another large desktop OS who actually has a reasonable NFS client? How
do you turn this into a straw man argument then? Remember, I'm not
claiming that my solution for the entirety of the threads that AV
vendors claim to want to protect again, I simply claim that a
GLIBC/LD_PRELOAD solution is easily shown to be infeasible for even the
most elementary of threats.
> --- so the right
> solution there is to integrate the virus checking with Samba ---
> especially since the one AV vendor has already admitted the actual
> virus signature checking has to be done in userspace.
<snide> I believe they all are going to claim it has to be in some
userspace proprietary application for them to keep making money </snide>
> For Linux clients, one question that immediately rises is the
> end-to-end argument. Wouldn't be far better to run whatever security
> solution on the client? After all, a Virus checking on an NFS server
> isn't going to help the user if they accidentally track in the virus
> on a USB stick. (Especially if it is an infected Macro virus in an
> office document.)
Your argument is irrelevant for the threat given and you seem to have
contorted the actual point of the statements to fit something else. But
I'm sure you a fan of multiple layers of security that you don't
actually believe that "just check on the clients" is the right thing to
do. Linux client side checking is most likely going to be something
that vendors claim to want to do but has no bearing on if out of kernel
scanning is feasible for NFS servers. Nor if you want to look at the
"end-to-end argument" as you claim can excluding server side scanning be
a reasonable choice.
How many clients machines at your location are controlled by some IT
organization? How many servers? I think it's quite obvious that unless
the answer to the first question is "all" then we would want scanning on
both the client and the server. I think there are many organizations in
which many, or even most, machines with access to an NFS server can be
controlled so as to enforce scanning, but its not reasonable, at least
in my mind, to throw out server side NFS scanning unless you can control
ALL of the clients.
-Eric
-Eric
On Wed, Aug 06, 2008 at 05:28:01PM -0400, Eric Paris wrote:
> > In this scenario, are you positing that you are worried about Windows
> > malware, or Linux malware? What OS are the clients running? I will
> > note that Windows has such a sucky NFS implementation that nearly all
> > Widows clients will be running CIFS/SMB, not NFS
>
> I believe I specifically did not make any such claims at all about the
> client OS and merely claimed the intended target was not the linux NFS
> server.
I know you didn't say; that's why I asked. :-)
I dispute your assertion that this quesiton is irrelevant. It's
highly relevant, because if it's Windows clients, they ***won't*** be
using NFS.
As for other large desktop OS's, that would be MacOS and Linux;
anything else? And the big, huge, vast difference between Windows and
MacOS/Linux is that with Windows, in practice people ran with
Administrator privileges because most applications (including at one
point Microsoft Visual Studio :-) died and/or completely refused to
install if you didn't have Administrator privileges. So people very
regularly ran with Root privs. With Vista, you no longer run with
root privileges by default --- instead, applications still assume they
have Administrator privileges, causing the Really Annoying Popup boxes
to pop up each time the application needs to do something that require
privileges --- which has trained users to mindlessly click "OK" each
time the Annoying Popup Box comes up.
Given that MacOS and Linux don't have these flaws with respect to
applications regularly expecting root privileges, will you admit that
perhaps some of the extreme scanning tactics that were required by
AntiVirus vendors might be not as necessary for "other desktops"?
Asking the question is important because if they are spending all of
their time on Windows virii, then your "elementary threat" is really
an "elementary strawman". Or, at the very least, it's a low priority
effort, since the number of virii out in the field for Linux and MacOS
desktops is in the noise compared to Windows. I know that it's
convenient for AV vendors to claim in their marketing literature that
this is only because Windows is more popular, but while that might be
part of it, it is also true that there are significant, structural
differences between Windows and those other large desktop candidates.
> Your argument is irrelevant for the threat given and you seem to have
> contorted the actual point of the statements to fit something else. But
> I'm sure you a fan of multiple layers of security that you don't
> actually believe that "just check on the clients" is the right thing to
> do.
Giving up my water bottles and having to take off my shoes at airport
security has been justified in the name of "multiple layers of
security". No, I'm NOT a fan of mindlessly using "defense in depth"
as an excuse for arbitrary amounts of security and giving up arbitrary
amounts of my private data. You need to prove to me that from a cost
benefit tradeoff it's really worth it.
- Ted
Eric Paris wrote:
>There is a consensus in the security industry that protecting against
>malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
>of so-called on-access scanning is usable and reasonable approach.
This is at odds with my experience. Are you sure you've been talking to
the right people? Is it possible you've only been talking to A/V vendors?
I find it entirely plausible that there is such a consensus among A/V
vendors, but I'm pretty skeptical that the rest of the security community
would make this kind of claim. What I hear, instead, is quite a bit of
skepticism about the future of A/V.
Here's an experiment for you. Walk up to a random security expert and
ask them what they think of blacklisting as a foundation for building
secure systems. Ask them what they think of the future of A/V in security
and whether they think A/V will be of increasing or decreasing relevance
to security in the future. The answers might be educational. Actually,
I suspect it's even possible you might find that many knowledgeable A/V
insiders privately share some of these same concerns about the future
of A/V -- look at how pretty much every major A/V vendor out there is
looking to diversify, to expand into other areas of computer security
and compliance, and to move beyond signature-based file scanners.
If you picked a bunch of computer security experts who don't work for an
A/V vendor and asked them what they thought about all this, I suspect
they'd be more likely to line up behind the kinds of comments that Ted
Tso has been posting. Personally, I think Ted's comments have been
highly constructive, thoughtful, and well worth re-reading.
On 8/6/08, David Collier-Brown <[email protected]> wrote:
>
>
> Peter Dolding wrote:
> > My main issue is TALPA, dazuko and so on of Anti-Virus Filesystem
>> monitoring are all going to break anyhow when
>> http://lwn.net/Articles/251224/ Credentials get added and common
>> filesystem caching gets added.
>>
>> You want to change a permissions on a file/object before its opened.
>> So does the Credential user space daemon on file systems that cannot
>> store secuirty information. We only really need 1 location in the
>> source base for this. Expand Credentials slightly to allow anti
>> viruses to operate by by problem. Even better when FS-Cache can sit
>> on top of Credentials correctly no need for anti virus software to
>> have independent caching of blocked and allowed files. FS-Cache picks
>> a large amount of this up.
>>
>> Basically TALPA, dazuko and so on of Anti-Virus Filesystem monitoring
>> don't fit in the future design of Linux. All they will be is
>> duplication of a existing interface. A interface that complete avoids
>> the stacking issue.
>
> Then in a real sense you've solved much of their problem for them (;-))
> After this comes engineering, so that they can re-use the scanning
> mechanisms they already have, but from a different caller.
>
> The requirements are probably that they know
> - is this an open for read or write (somewhat less time-sensitive)?
> - is the data present, or do we have to wait?
> - if so, for what?
> as of the time they start looking at the file. Having a race-free
> mechanism using credentials and RCU is, IMHO, A Really Good Thing.
>
> Another thing they and we will likely need is a way to discover
> if a file is inacessable due to an AV operation, and if the time
> it has been inacessable is less than or equal to a scanning
> upper bound by file size or beyond it. The latter is for repair
> of broken state introduced by the AV process failing.
>
> --dave
>
Failed daemon support has to exist in Credentials for filesystems that
cannot store all the security data and needing third party support.
This is the issue same features basically required to do the same
things. Different reasons.
One to virus scan. One to fix up security lacking file systems. From
what I see exactly the same set of features are needed.
At most only 1 feature is missing a form of copy on write. Adding
this at credentials level over the complete system would not be a bad
thing. A option to cache writes for scanning even write to like a
journal for latter applying once scanned. This could even come as
part of the generic file system cache.
This is why I am kinda scratching my head. Creating new files also
have to go past Credentials. Even creating and editing files have to
go past Credentials.
It seams like the best place to look to hook in correctly even better
most of the hook is already there. Hooking into Credentials is far
better than hooking into syscalls. Since if a new special feature
filesystem access syscall is added it will still enter Credentials.
On top Credentials interface is under the VFS. Its between the raw
filesystem and the VFS. So no more path based errors.
Yes deeper into the OS to the correct location. Anti-virus methods
are too shallow. AppArmor and some other secuirty systems like it
have the same path based issue because LSM is too shallow.
Anti-Virus company answer has been to root kit out of the LSM API
wrong answer. Correct answer is get to the correct location in the
OS.
Note clamfs is on the right path using fuse since this goes under the
VFS and not effected by path issues and is not root kitting the OS.
Of course credentials do the job better.
Really Anti-Virus guys give as a shopping list of exactly what you
need to do. Key word need. Hooking syscall's most likely is not a
need but a want. Even worse a want in the wrong location to give 100
percent coverage. Ie when a new syscall is added it creates a
backdoor around you scanners.
Peter Dolding
PS I really wound not mind a better paid job. The developers who
have to created all these incorrect ways have to been paid more than
me.
Paul Moore wrote:
> With multiple security markings on an entity then you have to decide if
> you want to consider every marking at each LSM instance, or a subset.
> The complexity in this case does go up dramatically, but I think the
> key point for our discussion is that it doesn't matter if the entity is
> a file or a packet.
>
>
Perhaps you're right. I'm thinking in terms of the notion that each
LSM can have an independent file system attribute, but if they all
want to use IP options we're talking about a very limited resource.
No real biggie, I mostly wanted to point out that there's more to
stacking than fetching multiple xattrs.
On Wednesday 06 August 2008, Adrian Bunk wrote:
> On Wed, Aug 06, 2008 at 12:07:57PM +0100, [email protected] wrote:
> > Adrian Bunk <[email protected]> wrote on 06/08/2008 11:50:08:
> > > As an observer of this thread:
> > >
> > > - Some set of requirements suddenly appears out of the void on
> > > linux-kernel.
> >
> > Because previously it was said to go away and come back with a clear list
> > of requirements. And here you make it sound like a negative thing. See
> > what I am talking about?
>
> Both of my points belong together.
>
> > > - Noone is able and/or willing to exactly describe the problem(s) they
> > > are trying to solve.
> >
> > Hopefully we will get there. Very little time has passed since the
> > discussion has started, even less considering the time zone difference
> > for some.
> >
> > > With this status quo the discussion is going nowhere - Linux kernel
> > > development does not work this way.
> > >
> > > The aim is not to include this code, but to find the best technical
> > > solution for your problem(s) - no matter whether this will have
> > > anything in common with the list of requirements and the code posted or
> > > not.
> >
> > I completely agree with that. Here I was just pointing out that what Greg
> > wrote was untrue and exaggerated so not helping the discussion at all.
>
> Until now the main discussion participant from the AV side is
> Jonathan Press.
Well, if you insist, but I must state that this mail represents my own opinion
and not my employer's (that's because all the people I could consult with are
sleeping :) ).
> But the real discussion hasn't even started since the information
> required is not available.
>
> And as soon as the information for the real discussion is available all
> these initial discussions become irrelevant.
> - Noone is able and/or willing to exactly describe the problem(s) they
> are trying to solve.
Well, here is one attempt.
A good percentage of an AV product's job is to prevent exploitation of a
security hole in a product before the vendor (assuming the vendor admits it's
bug and not a misuse of the product's features).
Most distribution makers go through a lot of work before releasing an update,
which might take days. Add to this the fact that some users refuse to update
periodically (because one operating system out there shattered the belief in
this practice) and that some of them are willing to pay to not care. This is
reason enough for most AV vendors.
In the present, on the Linux Desktop, this is [still] hypothetical talk and
God help it will remain so. However, if there is one incredibly small chance
that one (new?) type of malware can spread to a large number of users, then
AV vendors will race for creating a solution because there will _definitely_
be people needing help with this (please notice that the IQ scale starts from
zero and not from 130 :) ).
I think this patch is trying to do what dazuko hasn't managed to do (yet): get
into mainline. :)
--
Mihai Donțu
On Thu, 7 Aug 2008 03:49:55 +0300
Mihai Donțu <[email protected]> wrote:
> Well, here is one attempt.
>
> A good percentage of an AV product's job is to prevent exploitation
> of a security hole in a product before the vendor (assuming the
> vendor admits it's bug and not a misuse of the product's features).
just to get things clear;
you're not talking about preventing the actual exploitation per se
(that would be the job of the various protection technologies) or the
containment (that would be SELinux), but more about detecting the
presence and preventing to (accidental) use of pre-canned, widely used
exploit binaries/files ?
Hi!
> Problems with inotify as far as I know:
>
> You can't do something like inotify("/") (made up API) but you have to set
> up a watch for every directory you wan't to watch. That seems like a waste
> of resources.
>
> Then you get back a file name, if you wan't to report it or attempt* to
> scan it you have to build a pathname yourself, which means you have to
> maintain the whole tree of names in memory. Even bigger waste.
>
> When I say attempt to scan it above I mean that we are back into the
> pathanme teritorry. It is not guaranteed we will be able to open and scan
> using that pathname. I don't know what inotify reports with chroots and
> private namespaces, but it can certainly fail with NFS and root_squash. So
> it is less effective as well as being resource intensive.
>
> I think this is a good amount of flaws which shows inotify isn't really
> ideal.
Not ideal, but looks like good enough, and could certainly be
improved. If it is secure-enough for you (I think it is), that looks
like a way to go.
(Plus, such improvements would actually be very welcome).
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Wed, 2008-08-06 at 17:52 -0400, Theodore Tso wrote:
> On Wed, Aug 06, 2008 at 05:28:01PM -0400, Eric Paris wrote:
> > > In this scenario, are you positing that you are worried about Windows
> > > malware, or Linux malware? What OS are the clients running? I will
> > > note that Windows has such a sucky NFS implementation that nearly all
> > > Widows clients will be running CIFS/SMB, not NFS
> >
> > I believe I specifically did not make any such claims at all about the
> > client OS and merely claimed the intended target was not the linux NFS
> > server.
>
> I know you didn't say; that's why I asked. :-)
We are sort of talking past each other here, but I really am listening.
[snip]
> Given that MacOS and Linux don't have these flaws with respect to
> applications regularly expecting root privileges, will you admit that
> perhaps some of the extreme scanning tactics that were required by
> AntiVirus vendors might be not as necessary for "other desktops"?
Absolutely, I think that our solution needs worry much less about an
actively attacking root process than Windows. I think everyone on list
would tend to agree that our security model greatly helps in this
regard. I think the problem quickly becomes intractable as soon as we
talk about a locally running actively attacking root process, which I
believe is what you are discussing. I think that leaves 4 things
1) fileserver with no active attack on system
2) actively attacking non-root processes (run some process that tries to
screw people)
3) accidentally attacking, non-root process. (open a malicious
openoffice file)
4) accidentally attacking, properly functioning, root processes (my
thought on this would be a properly functioning non-exploited yum
program getting data from a bad repo)
For the purposes of this e-mail discussion can we only discuss #1? I
think its the easiest problem to look at and is at the heart of stopping
#2, 3 and 4. If we can keep the files off the disk we greatly reduce
(although clearly don't eliminate) the ability for all of the other
attacks against our system.
> Asking the question is important because if they are spending all of
> their time on Windows virii, then your "elementary threat" is really
> an "elementary strawman". Or, at the very least, it's a low priority
> effort, since the number of virii out in the field for Linux and MacOS
> desktops is in the noise compared to Windows. I know that it's
> convenient for AV vendors to claim in their marketing literature that
> this is only because Windows is more popular, but while that might be
> part of it, it is also true that there are significant, structural
> differences between Windows and those other large desktop candidates.
Remember my (contrived to prove a single point, but I think interesting)
goal was not to stop an active attack against any OS, which you appear
to be focusing on. My goal was to make a Linux fileserver an
inhospitable place for data which may someday attack a system to live.
Our systems have almost infinitely many ingress and egress mechanisms
and I don't think its tractable to attempt to do this at the edge. At a
minimum while I sit here I can think of smbd, nfs, ftp, http, smtp, ssh,
and rsync which are very likely ingress and egress points of data on a
Linux system. Trying to move the solution to all of those places gets a
bit large and certainly buggy. And I think we both agree those aren't
the only possible ingress and egress points for information.
> > Your argument is irrelevant for the threat given and you seem to have
> > contorted the actual point of the statements to fit something else. But
> > I'm sure you a fan of multiple layers of security that you don't
> > actually believe that "just check on the clients" is the right thing to
> > do.
>
> Giving up my water bottles and having to take off my shoes at airport
> security has been justified in the name of "multiple layers of
> security". No, I'm NOT a fan of mindlessly using "defense in depth"
> as an excuse for arbitrary amounts of security and giving up arbitrary
> amounts of my private data. You need to prove to me that from a cost
> benefit tradeoff it's really worth it.
I guess step one is agreeing that the goal "harden linux to be an
inhospitable host for 'bad' data which may someday be used to attack
another system" is a useful goal. Maybe you don't agree and if so
please help me understand how that goal is unreasonable. We aren't
talking about the solution just yet (I'm the first to say that the
patches I posted might be complete shiet for the final solution), just
the goal for a moment. Maybe the solution will be so unwieldy that we
later find the goal to not be worth it, but this seems the most
simplistic of goals that all AV vendors are going to claim to want to
do.
I believe this to be a reasonable goal as it reduces the attack area for
a number of attacks against linux programs (pdf rendering flaws, jpg
rendering flaws, etc). Browser downloads bad pdf, evince tries to open
that which the browser just downloaded, BLOCKED. Obviously the 'right'
thing to do there was update evince so the attack would not work, but
now what stops me, the unsuspecting use who just looked at this pdf and
didn't get hacked from passing it on? Same thing for openoffice macros?
Sure you update openoffice and you down get p0wnt but that doesn't keep
you from putting it on the NFS server for everyone else in the office
who doesn't know how to keep their system up to date from getting
screwed.
While waiting for AV vendors to see if any of them can make reasonable
claims about goals 2, 3, and 4 I think we can consider #1. (for all I
know #1 is going to be the ONLY thing that any vendor can make a
reasonable claim about) Even if they come up with other things I think
that just considering #1 is valuable since it can eliminate some of the
potential solutions. I don't see myself wasting time going down the
GLIBC path since I believe that making systems inhospitable to 'bad'
data is something that should be done (if reasonably possible) both
client and server side.
-Eric
On Thu, Aug 7, 2008 at 7:28 PM, Pavel Machek <[email protected]> wrote:
> Hi!
>
>> Problems with inotify as far as I know:
>>
>> You can't do something like inotify("/") (made up API) but you have to set
>> up a watch for every directory you wan't to watch. That seems like a waste
>> of resources.
>>
>> Then you get back a file name, if you wan't to report it or attempt* to
>> scan it you have to build a pathname yourself, which means you have to
>> maintain the whole tree of names in memory. Even bigger waste.
>>
>> When I say attempt to scan it above I mean that we are back into the
>> pathanme teritorry. It is not guaranteed we will be able to open and scan
>> using that pathname. I don't know what inotify reports with chroots and
>> private namespaces, but it can certainly fail with NFS and root_squash. So
>> it is less effective as well as being resource intensive.
>>
>> I think this is a good amount of flaws which shows inotify isn't really
>> ideal.
>
> Not ideal, but looks like good enough, and could certainly be
> improved. If it is secure-enough for you (I think it is), that looks
> like a way to go.
>
> (Plus, such improvements would actually be very welcome).
> Pavel
>
My Issue API does not have to be made up. The API kinda exists in
Credentials for a completely different use.
Issue with inotify its only monitoring. No permission alteration to
block access.
Also its the wrong level. I will explain why.
inotify is sitting on top of the vfs. Ok fine if you don't want scan
effectiveness. With bind mounts this becomes trouble quickly.
/tmp
/usr/tmp
/usr/local/tmp
All could be exactly the same physical drive. Now in the anti-virus
would have to scan each of thoose directories if it does not find out
they are bind mounted. Also could not look up and say ok that is bind
mounted its fine because since the last lookup a new bind mount could
have been put over the top letting something slip past.
Also TALPA is wrong because its a LSM. LSM are the most powerful
bits of the OS. LSM are the enforcers of Linux. They are the last
line of defence for the OS. You have your normal permission systems
when they fail then its the job of a LSM to pick up the mess.
Virus scanning is way better inside the permission system. This way
if you have a issue there is still a back stop of the LSM to prevent a
breached virus scanner doing major system harm. TALPA risks killing
the backstop.
Basically path based + Linux equals doomed. VFS is not path based friendly.
Credentials patch on the other hand. Is under the VFS sees real
partitions so can process divided by partition avoiding rescanning.
Also its in the permission system to has to get approve from the LSM
to do alterations. So its back stoped.
Also is sitting at exactly the right point to override permissions.
Filesystem cache that exists to sit on top of Credentials provides the
other needed bit scan on write function.
Also filesystem driver level was also correct. Same reason not messed
up by VFS multi mounting and over mounting.
Don't worry lots of things have been made as LSM's that should not
been. Basically avoiding having to correct the right things.
Peter Dolding
On Fri, 2008-08-08 at 00:21 +1000, Peter Dolding wrote:
> Also TALPA is wrong because its a LSM. LSM are the most powerful
> bits of the OS. LSM are the enforcers of Linux. They are the last
> line of defence for the OS. You have your normal permission systems
> when they fail then its the job of a LSM to pick up the mess.
I've been trying to ignore you, but I just can't any more. Please dear
god look at my work before spewing this crap. My work is not an LSM.
My work doesn't care about bind mounts, it cares about inodes. My work
is not path name based. I'm not sure who you are trying to convince of
something here but obviously talking about something you don't know
about is not working for you.....
-Eric
On Fri, Aug 8, 2008 at 12:31 AM, Eric Paris <[email protected]> wrote:
> On Fri, 2008-08-08 at 00:21 +1000, Peter Dolding wrote:
>
>> Also TALPA is wrong because its a LSM. LSM are the most powerful
>> bits of the OS. LSM are the enforcers of Linux. They are the last
>> line of defence for the OS. You have your normal permission systems
>> when they fail then its the job of a LSM to pick up the mess.
>
> I've been trying to ignore you, but I just can't any more. Please dear
> god look at my work before spewing this crap. My work is not an LSM.
> My work doesn't care about bind mounts, it cares about inodes. My work
> is not path name based. I'm not sure who you are trying to convince of
> something here but obviously talking about something you don't know
> about is not working for you.....
>
Path based was to using inotify that basically cannot do the job. I
guess you agree with that.
Ok you corrected that bit in the TALPA design? Past form of the
beast was hooking down threw the LSM structs to block files so if
something went wrong it could override all the permissions on the
system without any means of anything else blocking it. Effectively
killing the system. Basically the Userspace daemon answering block
everything worked. Ie connected in too high. So key operational
files of the OS could not be protected against the scanner turning
into malware itself. Yes users are tricked to install fake
anti-viruses and malware scanners on windows. So yes something
hostile to the system can and most likely will connect to the TALPA
API. Is over site to the LSM provided? Just last time I looked
still missing.
Credentials alterations to permissions must go back threw the LSM.
Also Credentials permission change will not operate without LSM built
into kernel. Running a Anti-virus without a wrapper around it is
partly foolish anyhow.
There is still the duplication issue. Two filesystem caches one in
TALPA and One the FS-cache that is designed to sit on top of
Credentials. TALPA is only a part filesystem cache but caching
approved and not approved is still a form of filesystem cache. Its
caching virtual permissions of the filesystem.
Also duplication of permission overriding Credentials provides means
to override permissions from userspace as well including creating
virtual permissions like TALPA is doing. Reason for virtual
permissions simple not all file system support the all the permissions
to run Linux on with LSM's fully active.
Simple fact you have two API doing almost exactly the same thing.
Just different ways. Credentials spreed all the way threw the OS as
well.
So if in the Credential stuct was added a flag saying what version
virus scan had been done on a running executable a search of running
programs that had not been scanned against particular virus signatures
would also be simple. Same with disconnected files. Ie files that
have been deleted but since a process is still running are still not
deleted yet in the /proc. TALPA cache current design cache is
nothing more than duplication beside what Credentials and FS-cache is
doing.
Also Credentials nice screw up you inode monitoring. FS-cache on top
of Credentials can impersonate a inode. So TALPA will think its
getting a inode from like NFS but its really getting what was stored
in the FS-cache. So now TALPA is failed it has to scan everything to
avoid Credentials and its cache is rendered useless.
Caching of scanned and not scanned is more than doable threw the
credentials. Scanning at the Filesystem cache itself also makes
sence. Only 1 cache system to run also equals not exploited by
another cache running on the system.
OS hardening will required doing away with all per filesystem caches
and come back to generic cache only. The caches currently back door
TALPA and have minor down side to Credentials. Credentials with
FS-cache does have the option to scan everything as its loaded into
the Cache so can defend against the other caches harming it at the
cost of CPU time and ram. Double full caching. But since its caching
the files themselves its can not be tricked by impersonation.
TALPA is a failed design. It has too many holes. They need to be
closed. Fastest way to close them is just merge its ideas into
Credentials.
Peter Dolding
On 07-08-08 16:16, Eric Paris wrote:
> Absolutely, I think that our solution needs worry much less about an
> actively attacking root process than Windows.
Note by the way that this assumption is being actively undermined by all
those companies releasing software released as big executable blobs that
you just have to chmod +x and run; as root if you want them in /opt, or
/usr/local/bin, or ...
Even when OpenOffice.org, IBM Lotus Symphony, RealPlayer, Flash, what
have you, might generally by themselves not be considered virusses both
the precedent they set and the opportunity for infecting systems by
offering those for re-download from elsewhere is worrying.
One of the more useful things the security crowd could do to keep the
above assumption valid would be working on a general "external software
installer" and getting all distributions to settle on it (which probably
needs a common divisor package and distro backends to feed it to the
native package manager).
Rene.
On Fri, 2008-08-08 at 04:06 +0200, Rene Herman wrote:
> On 07-08-08 16:16, Eric Paris wrote:
>
> > Absolutely, I think that our solution needs worry much less about an
> > actively attacking root process than Windows.
>
> Note by the way that this assumption is being actively undermined by all
> those companies releasing software released as big executable blobs that
> you just have to chmod +x and run; as root if you want them in /opt, or
> /usr/local/bin, or ...
but you already say that said blob exists on disk? Therefore by my most
basic of models it won't ever actually get to run since it will get
scanned right as you try to execute it and you will get EPERM instead of
a running evil process. (all of that is assuming the userspace black
magic is useful, but I don't think that's really up for debate since we
have no way of knowing exactly what these closed source AV vendors
actually are doing....)
It looks in my mind that more and more the only real model that can even
attempt to be addressed is to make disks inhospitable to data which
might be intended to do ill to another machine.
Once the process is running we are talking about an IDS right?
>
> Even when OpenOffice.org, IBM Lotus Symphony, RealPlayer, Flash, what
> have you, might generally by themselves not be considered virusses both
> the precedent they set and the opportunity for infecting systems by
> offering those for re-download from elsewhere is worrying.
>
> One of the more useful things the security crowd could do to keep the
> above assumption valid would be working on a general "external software
> installer" and getting all distributions to settle on it (which probably
> needs a common divisor package and distro backends to feed it to the
> native package manager).
maybe a good idea, but beyond my expertise or ability to push forward...
-Eric
On 08-08-08 04:15, Eric Paris wrote:
> On Fri, 2008-08-08 at 04:06 +0200, Rene Herman wrote:
>> On 07-08-08 16:16, Eric Paris wrote:
>>
>>> Absolutely, I think that our solution needs worry much less about an
>>> actively attacking root process than Windows.
>> Note by the way that this assumption is being actively undermined by all
>> those companies releasing software released as big executable blobs that
>> you just have to chmod +x and run; as root if you want them in /opt, or
>> /usr/local/bin, or ...
>
> but you already say that said blob exists on disk? Therefore by my most
> basic of models it won't ever actually get to run since it will get
> scanned right as you try to execute it and you will get EPERM instead of
> a running evil process. (all of that is assuming the userspace black
> magic is useful, but I don't think that's really up for debate since we
> have no way of knowing exactly what these closed source AV vendors
> actually are doing....)
>
> It looks in my mind that more and more the only real model that can even
> attempt to be addressed is to make disks inhospitable to data which
> might be intended to do ill to another machine.
>
> Once the process is running we are talking about an IDS right?
Once the process is running we are talking about having lost.
Note, I did not follow this thread (closely) ...
Thing is just -- just wait for how long it takes said clueless user you
are so vigorously protecting to disable that which brings about your
model the first time if interferes with him running a cracked copy of an
expensive yet popular program, playing a game, installing a wiggling
nudies screensaver, ...
The difference here is just that Linux systems are particularly bad at
those tasks to begin with which in nice circular motions keeps those
clueless users away, obviating the need to protect ourselves from them.
If this stuff is to be discussed in a context as if Linux were relevant
on the desktop though, I believe it's rather unproductive to expect a
fundamental difference with Windows in this respect.
Fine though if you wouldn't insist on that context. As said, wasn't
following along really.
>> Even when OpenOffice.org, IBM Lotus Symphony, RealPlayer, Flash, what
>> have you, might generally by themselves not be considered virusses both
>> the precedent they set and the opportunity for infecting systems by
>> offering those for re-download from elsewhere is worrying.
>>
>> One of the more useful things the security crowd could do to keep the
>> above assumption valid would be working on a general "external software
>> installer" and getting all distributions to settle on it (which probably
>> needs a common divisor package and distro backends to feed it to the
>> native package manager).
>
> maybe a good idea, but beyond my expertise or ability to push forward...
I'll leave it in just in case it spurs someone :)
Rene.
On Fri, 8 Aug 2008, Peter Dolding wrote:
> TALPA is a failed design. It has too many holes. They need to be
> closed. Fastest way to close them is just merge its ideas into
> Credentials.
Sound great, looking forward to your patches!
I suggest developing them against:
git://git.kernel.org/pub/scm/linux/kernel/git/jmorris/security-testing-2.6#next-creds
Thanks.
- James
--
James Morris
<[email protected]>
> -----Original Message-----
> From: [email protected] [mailto:malware-list-
> [email protected]] On Behalf Of Rene Herman
> Sent: Thursday, August 07, 2008 10:56 PM
> To: Eric Paris
> Cc: [email protected]; Alan Cox;
[email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a linux
interface for on
> access scanning
>
> The difference here is just that Linux systems are particularly bad at
> those tasks to begin with which in nice circular motions keeps those
> clueless users away, obviating the need to protect ourselves from
them.
> If this stuff is to be discussed in a context as if Linux were
relevant
> on the desktop though, I believe it's rather unproductive to expect a
> fundamental difference with Windows in this respect.
Just an observation about this...
Since I earn my living on the basis of users, clueless or not, I've
gotten into the habit of just taking them as they come and trying to go
out of my way to not refer to them as clueless -- except in a few
specific and particularly annoying cases.
But that's not my point. My point is that Linux has become a
commercially viable environment with a lot of enterprise users, with a
significant number of enterprises are standardizing on it, or at least
officially supporting/allowing/encouraging its use. Because of that,
for example, we have a significant number of user issues coming in that
indicate that there are actually plenty of clueless Linux users, whether
the OS was intended for them or not.
The fact that they are there is the main reason that Red Hat and Novell,
for example (at least by my observation from the outside -- I can't
speak at all about how they see if from the inside) seem to be putting
the bulk of their efforts into their enterprise editions, as opposed to
their traditional technologist editions.
The bottom line, then, is that there ARE way more clueless Linux users
out there than there used to be, which makes them a) vulnerable to
losses by virtue of their own mistakes, and b) vectors for the spread of
malware. Which is kind of why we're here.
Jon Press
On 08-08-08 13:58, Press, Jonathan wrote:
> Just an observation about this...
>
> Since I earn my living on the basis of users, clueless or not, I've
> gotten into the habit of just taking them as they come and trying to go
> out of my way to not refer to them as clueless -- except in a few
> specific and particularly annoying cases.
You say that as though you feel that calling someone clueless were a bad
thing. There are tons of subjects I'm completely and utterly clueless
about and very happily so. But, as you say, that's not the point...
> But that's not my point. My point is that Linux has become a
> commercially viable environment with a lot of enterprise users, with a
> significant number of enterprises are standardizing on it, or at least
> officially supporting/allowing/encouraging its use. Because of that,
> for example, we have a significant number of user issues coming in that
> indicate that there are actually plenty of clueless Linux users, whether
> the OS was intended for them or not.
But not users with root access, which is the context in which my own
remark was. Enterprise users in corporations are not what I call the
desktop; I'd generally call those workstations, with the desktop being
your average home PC with the enormous amounts of cheap and buggy
hardware and the definite lack of central IT management.
It's also dependent on country. Over here in the Netherlands, corporate
adoption "on the workstation" is very low (and seemingly dropping again
after some initial attempts in local government) and adoption on the
desktop is for all intents and purposes 0. It's different especially in
eastern-europe.
Funny that really, how all that Free as in Speech stuff mostly works for
people without money...
> The fact that they are there is the main reason that Red Hat and Novell,
> for example (at least by my observation from the outside -- I can't
> speak at all about how they see if from the inside) seem to be putting
> the bulk of their efforts into their enterprise editions, as opposed to
> their traditional technologist editions.
>
> The bottom line, then, is that there ARE way more clueless Linux users
> out there than there used to be, which makes them a) vulnerable to
> losses by virtue of their own mistakes, and b) vectors for the spread of
> malware. Which is kind of why we're here.
Right, so that, then, is a threat model. I myself believe you are here
mostly to guard against 11-year old girls installing infected
screensavers of horses which given the fairly low adoption of Linux by
11-year old girls says something about my view of things.
But, yes, as I myself said as well, it might be sensible to discuss this
issue simply _as if_ lots of users were brushing their My Little Pony's
while waiting for their kernels to finish compiling if you're designing
something that _should_ protect them if they were.
Goes back really to the threat model question you were asked I guess.
Rene.
> -----Original Message-----
> From: Rene Herman [mailto:[email protected]]
> Sent: Friday, August 08, 2008 8:35 AM
> To: Press, Jonathan
> Cc: Eric Paris; [email protected]; Alan Cox; malware-
> [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a linux
interface for on
> access scanning
> > Since I earn my living on the basis of users, clueless or not, I've
> > gotten into the habit of just taking them as they come and trying to
go
> > out of my way to not refer to them as clueless -- except in a few
> > specific and particularly annoying cases.
>
> You say that as though you feel that calling someone clueless were a
bad
> thing. There are tons of subjects I'm completely and utterly clueless
> about and very happily so.
Well, in all honesty, sometimes when I think of someone as clueless,
there is a certain negative emotion that goes along with it. However,
the point is more that THEY don't like to be called clueless -- both
because, for some reason they take it as an insult, and also because
there are many out there who "know enough to be dangerous" and think
they know more than they really do. So I have to be careful not to say
anything to them that sounds like I think that, and to make sure that
people in my group have the same discipline.
> > But that's not my point. My point is that Linux has become a
> > commercially viable environment with a lot of enterprise users, with
a
> > significant number of enterprises are standardizing on it, or at
least
> > officially supporting/allowing/encouraging its use. Because of
that,
> > for example, we have a significant number of user issues coming in
that
> > indicate that there are actually plenty of clueless Linux users,
whether
> > the OS was intended for them or not.
>
> But not users with root access, which is the context in which my own
> remark was. Enterprise users in corporations are not what I call the
> desktop; I'd generally call those workstations, with the desktop being
> your average home PC with the enormous amounts of cheap and buggy
> hardware and the definite lack of central IT management.
You would be surprised how many users with root access are also
clueless. I used the word "enterprise" very broadly. It applies to
large companies/institutions with sophisticated IT departments, but it
also applies to small to medium organizations who said to themselves,
"Gee what's this Linux thing I keep hearing about -- maybe we ought to
get one of those." And they end up with administrators who also don't
really know all the ins and outs they need to in order to create a
secure environment.
In a way, it's a price that Linux has paid for its growing acceptance.
The user community is not what it used to be.
> It's also dependent on country. Over here in the Netherlands,
corporate
> adoption "on the workstation" is very low (and seemingly dropping
again
> after some initial attempts in local government) and adoption on the
> desktop is for all intents and purposes 0. It's different especially
in
> eastern-europe.
I can't speak about every different country, but I would say that you
are right about Eastern Europe. But also in the US and Canada we have
plenty of adoption on the workstation.
> > The fact that they are there is the main reason that Red Hat and
Novell,
> > for example (at least by my observation from the outside -- I can't
> > speak at all about how they see if from the inside) seem to be
putting
> > the bulk of their efforts into their enterprise editions, as opposed
to
> > their traditional technologist editions.
> >
> > The bottom line, then, is that there ARE way more clueless Linux
users
> > out there than there used to be, which makes them a) vulnerable to
> > losses by virtue of their own mistakes, and b) vectors for the
spread of
> > malware. Which is kind of why we're here.
>
> Right, so that, then, is a threat model. I myself believe you are here
> mostly to guard against 11-year old girls installing infected
> screensavers of horses which given the fairly low adoption of Linux by
> 11-year old girls says something about my view of things.
>
> But, yes, as I myself said as well, it might be sensible to discuss
this
> issue simply _as if_ lots of users were brushing their My Little
Pony's
> while waiting for their kernels to finish compiling if you're
designing
> something that _should_ protect them if they were.
There are many gradations of users between the kernel pro and the
11-year-old girl.
One is the guy right out of college who really likes computers and he
gets a job at a place where they use Linux, and he does a web search and
finds this really cool application that might help him do his job
better, so he installs it on the spot. Uh-oh.
One is the guy who's been working there for a while and the head of IT,
tired of spending so much on Windows, tells him to install Linux on a
few machines, so he follows the instructions and he's now the
administrator.
The fact is that there are MILLIONS out there, and when you have
millions of anything, you are going to have lots of different kinds of
behavior that are hard to predict ahead of time.
On 08-08-08 15:11, Press, Jonathan wrote:
> One is the guy right out of college who really likes computers and he
> gets a job at a place where they use Linux, and he does a web search
> and finds this really cool application that might help him do his job
> better, so he installs it on the spot. Uh-oh.
See, that never happens here. Our guys use Linux while _in_ college (in
between the periods where they need to run Windows to get their college
work done ofcourse) and after graduation take a job collecting microsoft
certifications that their boss pays them to collect.
Then I come along a couple of years later and buy the old soundcards
that they put up for sale on our national auction site, mentioning I
want that stuff for Linux testing and they get all teary-eyed and
melancholic about the good old days when they had the time to play
around with linux. Those fun childhood days.
Not making it up.
> One is the guy who's been working there for a while and the head of
> IT, tired of spending so much on Windows, tells him to install Linux
> on a few machines, so he follows the instructions and he's now the
> administrator.
I saw that happening a few years ago here. By now, the small time shops
that I'm aware of have all gone back to Windows. Linux simply cost them
too much time which, if IT is not their business, they had no intention
whatsoever of spending on IT.
Do let me assure you though that I'm definitely aware that my personal
frame of refence might not be all that gobally applicable and bow out of
the discussion.
Rene.
On Thursday 07 August 2008, Arjan van de Ven wrote:
> On Thu, 7 Aug 2008 03:49:55 +0300
>
> Mihai Donțu <[email protected]> wrote:
> > Well, here is one attempt.
> >
> > A good percentage of an AV product's job is to prevent exploitation
> > of a security hole in a product before the vendor (assuming the
> > vendor admits it's bug and not a misuse of the product's features).
>
> just to get things clear;
> you're not talking about preventing the actual exploitation per se
> (that would be the job of the various protection technologies) or the
> containment (that would be SELinux), but more about detecting the
> presence and preventing to (accidental) use of pre-canned, widely used
> exploit binaries/files ?
I apologize for the late reply. The answer to your question is: yes. I was
planning to write some more on this subject but this is unnecessary now,
because I see [almost] everyone accepted that some kind of antimalware
scanning is needed and are looking for alternative (better) solutions to the
patch that started all this.
Can't wait to see the end result. :)
Thanks,
--
Mihai Donțu
On Mon, 11 Aug 2008 16:45:47 +0300
Mihai Donțu <[email protected]> wrote:
> On Thursday 07 August 2008, Arjan van de Ven wrote:
> > On Thu, 7 Aug 2008 03:49:55 +0300
> >
> > Mihai Donțu <[email protected]> wrote:
> > > Well, here is one attempt.
> > >
> > > A good percentage of an AV product's job is to prevent
> > > exploitation of a security hole in a product before the vendor
> > > (assuming the vendor admits it's bug and not a misuse of the
> > > product's features).
> >
> > just to get things clear;
> > you're not talking about preventing the actual exploitation per se
> > (that would be the job of the various protection technologies) or
> > the containment (that would be SELinux), but more about detecting
> > the presence and preventing to (accidental) use of pre-canned,
> > widely used exploit binaries/files ?
>
> I apologize for the late reply. The answer to your question is: yes.
> I was planning to write some more on this subject but this is
> unnecessary now, because I see [almost] everyone accepted that some
> kind of antimalware scanning is needed and are looking for
> alternative (better) solutions to the patch that started all this.
we do still appreciate your description, since I don't think there's a
clear "here's what we really try to protect against" statement yet.
Answering Ted's questions would be a really good start...
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
Arjan van de Ven wrote:
>>>just to get things clear; yyou're [ talking about ...] detecting
>>>the presence and preventing to (accidental) use of pre-canned,
>>>widely used exploit binaries/files ?
Mihai Donțu <[email protected]> wrote:
>>I apologize for the late reply. The answer to your question is: yes.
>>I was planning to write some more on this subject
Arjan van de Ven wrote:
> we do still appreciate your description, since I don't think there's a
> clear "here's what we really try to protect against" statement yet.
Perhaps I could try: the AV folks are trying to prevent the
execution of either modified normal binaries/files or
specifically exploit binaries/files, by machines for which the
files are executable or interpretable.
The experience of those communities is predominantly
with DOS/Windows executables and interpretable files, which
they have difficulty generalizing from.
In principle, they could be targeted at any machine, so any
mechanisms should be applicable to native executables and
interpretables as well as foreign ones.
--dave (who (in)famously wrote a UUCP virus, which warned sysadmins
if they had bad enough security settings to have run it as root) c-b
--
David Collier-Brown | Always do right. This will gratify
Sun Microsystems, Toronto | some people and astonish the rest
[email protected] | -- Mark Twain
cell: (647) 833-9377, bridge: (877) 385-4099 code: 506 9191#
> -----Original Message-----
> From: [email protected] [mailto:[email protected]] On
> Behalf Of David Collier-Brown
> Sent: Monday, August 11, 2008 12:11 PM
> To: Arjan van de Ven
> Cc: Mihai Donțu; Adrian Bunk; [email protected]; Greg KH; Press,
> Jonathan; [email protected]; [email protected];
> [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a linuxinterfaceforon access
> scanning
>
> Perhaps I could try: the AV folks are trying to prevent the
> execution of either modified normal binaries/files or
> specifically exploit binaries/files, by machines for which the
> files are executable or interpretable.
>
> The experience of those communities is predominantly
> with DOS/Windows executables and interpretable files, which
> they have difficulty generalizing from.
>
> In principle, they could be targeted at any machine, so any
> mechanisms should be applicable to native executables and
> interpretables as well as foreign ones.
You know, that's actually a very good statement of the model.
I think everyone understands one side of the threat model, that is Linux machines being carriers of infections aimed at other platforms. There are many ways that such infections can be stored, and many ways in which they can be communicated to the target machines. There are so many that it would not be effective or efficient for each such transfer application to be able to handle its own malware scanning, which is the short statement of why centralized AV protection with notification assistance from the kernel is appropriate.
So, putting that aside, David's statement is a reasonable summary of the nature of the other side of the model, the native attacks we are trying to protect against. That is, on any platform, not just Linux, we don't always know ahead of time what kind of attack can be launched and how it would be done. Experience has shown that no one knows enough about every piece of software on every machine to be able to know ahead of time what vulnerabilities there are that can be exploited. Therefore, we adopt a methodology that allows us to be fairly general. The signature matching approach allows us to find the signatures of either specific exploit binaries or corrupted normal binaries that our collection and research activities have identified as malware.
It is true that most of our experience is with DOS and Windows, and that the types of attacks that can be launched there are not easy to generalize to other platforms. However, that does not mean that non-Windows platforms like Linux are therefore immune, or that we are tilting at windmills to say that we address non-Windows infections. There are many specific forms of malware on non-Windows platforms that we identify in the same way that we identify them on Windows.
In a sense the reason I have found the question about "threat model" to be so difficult to answer is that the basic unpredictability of the attack makes the answer in essence: "Anything". That is, it essentially doesn't matter what the threat is or how malware is implemented, except that we know that it exists, both in theory and in practice, and we have an effective way of detecting and removing it.
Jon Press
????{.n?+???????+%?????ݶ??w??{.n?+????{??G?????{ay? ʇڙ?,j??f???h?????????z_??(?階?ݢj"???m??????G????????????&???~???iO???z??v?^?m???? ????????I?
Hi!
> > Perhaps I could try: the AV folks are trying to prevent the
> > execution of either modified normal binaries/files or
> > specifically exploit binaries/files, by machines for which the
> > files are executable or interpretable.
> >
> > The experience of those communities is predominantly
> > with DOS/Windows executables and interpretable files, which
> > they have difficulty generalizing from.
> >
> > In principle, they could be targeted at any machine, so any
> > mechanisms should be applicable to native executables and
> > interpretables as well as foreign ones.
>
>
> You know, that's actually a very good statement of the model.
>
> I think everyone understands one side of the threat model, that is Linux machines being carriers of infections aimed at other platforms. There are many ways that such infections can be stored, and many ways in which they can be communicated to the target machines. There are so many that it would not be effective or efficient for each such transfer application to be able to handle its own malware scanning, which is the short statement of why centralized AV protection with notification assistance from the kernel is appropriate.
>
No.
Proposed kernel solution did not work -- there still was write
vs. read race. You are right that it is not ok for each application to
do its own malware scanning, but libmalware.so that handles the
scanning seems very reasonable.
And as applications _need_ to be modified for the write vs. read race
to be solved, libmalware.so looks like a way forward.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
> -----Original Message-----
> From: Pavel Machek [mailto:[email protected]]
> Sent: Wednesday, August 13, 2008 6:28 AM
> To: Press, Jonathan
> Cc: [email protected]; Arjan van de Ven; Mihai Don??u; Adrian Bunk;
> [email protected]; Greg KH; [email protected];
linux-security-
> [email protected]; [email protected]
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
linuxinterfaceforon access
> scanning
>
> > I think everyone understands one side of the threat model, that is
Linux machines
> being carriers of infections aimed at other platforms. There are many
ways that
> such infections can be stored, and many ways in which they can be
communicated
> to the target machines. There are so many that it would not be
effective or efficient
> for each such transfer application to be able to handle its own
malware scanning,
> which is the short statement of why centralized AV protection with
notification
> assistance from the kernel is appropriate.
> >
>
> No.
>
> Proposed kernel solution did not work -- there still was write
> vs. read race. You are right that it is not ok for each application to
> do its own malware scanning, but libmalware.so that handles the
> scanning seems very reasonable.
>
> And as applications _need_ to be modified for the write vs. read race
> to be solved, libmalware.so looks like a way forward.
>
Pavel
I am not sure what you are suggesting, and I may have missed the
libmalware proposal (I don't see any mention of that specific idea in
any other message). However, just to be clear... At no point did we
suggest that the kernel would do any scanning. What we have been
interested in is a mechanism that can allow a scanning application to be
notified by the kernel of specific i/o events, for those events to be
blocked by the kernel until a user-space scan is done, and then the
user-space scan sends back allow or deny, at which point the i/o event
returns to the caller -- either success or error. This is the only way
that malware can be guaranteed of being detected when it is used (for
local application purposes or for transmission to another platform) or
created.
Also, a solution that requires applications to be modified will not
work, because there is no way that we would be able to get ALL
applications on the machines to be modified in the required ways. If
ANY applications are not so modified, then you have an unacceptable
malware hole. The only solution that really works is one that
guarantees that all applications are involved, which is why the kernel
has to be involved in some way. It's the only centralized authority
that can stick its nose into all of the required activities.
Whether the specific proposal currently on the table handles all the
issues or not is to me a separate point.
Jon Press
On Wed, Aug 13, 2008 at 8:46 PM, Press, Jonathan <[email protected]> wrote:
>> -----Original Message-----
>> From: Pavel Machek [mailto:[email protected]]
>> Sent: Wednesday, August 13, 2008 6:28 AM
>> To: Press, Jonathan
>> Cc: [email protected]; Arjan van de Ven; Mihai Don??u; Adrian Bunk;
>> [email protected]; Greg KH; [email protected];
> linux-security-
>> [email protected]; [email protected]
>> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
> linuxinterfaceforon access
>> scanning
>>
>> > I think everyone understands one side of the threat model, that is
> Linux machines
>> being carriers of infections aimed at other platforms. There are many
> ways that
>> such infections can be stored, and many ways in which they can be
> communicated
>> to the target machines. There are so many that it would not be
> effective or efficient
>> for each such transfer application to be able to handle its own
> malware scanning,
>> which is the short statement of why centralized AV protection with
> notification
>> assistance from the kernel is appropriate.
>> >
>>
>> No.
>>
>> Proposed kernel solution did not work -- there still was write
>> vs. read race. You are right that it is not ok for each application to
>> do its own malware scanning, but libmalware.so that handles the
>> scanning seems very reasonable.
>>
>> And as applications _need_ to be modified for the write vs. read race
>> to be solved, libmalware.so looks like a way forward.
>>
> Pavel
>
> I am not sure what you are suggesting, and I may have missed the
> libmalware proposal (I don't see any mention of that specific idea in
> any other message). However, just to be clear... At no point did we
> suggest that the kernel would do any scanning. What we have been
> interested in is a mechanism that can allow a scanning application to be
> notified by the kernel of specific i/o events, for those events to be
> blocked by the kernel until a user-space scan is done, and then the
> user-space scan sends back allow or deny, at which point the i/o event
> returns to the caller -- either success or error. This is the only way
> that malware can be guaranteed of being detected when it is used (for
> local application purposes or for transmission to another platform) or
> created.
>
> Also, a solution that requires applications to be modified will not
> work, because there is no way that we would be able to get ALL
> applications on the machines to be modified in the required ways. If
> ANY applications are not so modified, then you have an unacceptable
> malware hole. The only solution that really works is one that
> guarantees that all applications are involved, which is why the kernel
> has to be involved in some way. It's the only centralized authority
> that can stick its nose into all of the required activities.
>
> Whether the specific proposal currently on the table handles all the
> issues or not is to me a separate point.
Being in the correct place is key. Wrong place more issues more holes.
Lets get this right this time please. Windows is a mess. You cannot
run many av products side by side. Due to failures of signatures at
times to pickup every threat out there we have to move forward with a
system that works. Supporting many scanners at once if needed.
LIM guys are going after the same kind of defect detection.
Lot of designs have been created over all the years but they have all
failed to address the basics.
Scanning without file caching leaves a hole and costs cpu time. So
integration into the file system cache system is kinda key.
Designing hooks forcing users to use only 1 product. Yes good for
market lock in but kicks lots people in teeth.
Getting patch mainline is basically going to hit walls if you don't
sit down and say lets fix the kernel secuirty correctly. This is not
windows you have the power to alter the complete kernel if you can
provide valid justification for the change.
Valid justification is that someone like me looks at it and cannot
find a flaw in design and it provides needed functionality. Issue you
keep on addressing only 1 this is the reason why the patches have
stayed out side mainline when other patches have got in. In theory
the complete core of the Linux kernel could be replaced if Valid
justification could be made.
This is where Linux is different to Windows. You are free to
redesign linux to be the hardest OS on earth for a virus to operate.
Where windows you have to work with provided model. Lots here need
to drop the idea that anything in the Linux kernel is set in stone.
Only thing set in stone flawed designed get shot down.
Lets start with what would the features you would wish for as
Anti-virus designers if you could design a OS from scratch around your
Anti-virus so it was almost flawless. From that list we have what we
need to work on to give you exactly what you want in time. Even
better post that on a long term website as a wishlist for all OS's.
Basically give us direction.
Peter Dolding
Hi!
> > Proposed kernel solution did not work -- there still was write
> > vs. read race. You are right that it is not ok for each application to
> > do its own malware scanning, but libmalware.so that handles the
> > scanning seems very reasonable.
> >
> > And as applications _need_ to be modified for the write vs. read race
> > to be solved, libmalware.so looks like a way forward.
>
> I am not sure what you are suggesting, and I may have missed the
> libmalware proposal (I don't see any mention of that specific idea in
> any other message). However, just to be clear... At no point did we
> suggest that the kernel would do any scanning. What we have been
> interested in is a mechanism that can allow a scanning application to be
> notified by the kernel of specific i/o events, for those events to be
> blocked by the kernel until a user-space scan is done, and then the
> user-space scan sends back allow or deny, at which point the i/o event
> returns to the caller -- either success or error. This is the only way
> that malware can be guaranteed of being detected when it is used (for
> local application purposes or for transmission to another platform) or
> created.
As I said, that does not work. Proposed solution blocked open, while
you'd have to block read() and you still could not handle mmap.
So sorry, no kernel solution can work.
> Also, a solution that requires applications to be modified will not
> work, because there is no way that we would be able to get ALL
> applications on the machines to be modified in the required ways. If
> ANY applications are not so modified, then you have an unacceptable
> malware hole. The only solution that really works is one that
So you make sure all apps are modified. Distros are good at that, and
modifications are not that hard.
Plus, proposed solution already has three unacceptable holes:
1) it only catches known signatures
2) write vs. read race mentioned above
3) mmap problem
. Making sure all apps use libmalware.so is trivial compared to
solving 3).
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
Pavel Machek wrote on 13/08/2008 13:56:38:
Big snip since I am really only curious about libmalware.so.
> Plus, proposed solution already has three unacceptable holes:
>
> 1) it only catches known signatures
> 2) write vs. read race mentioned above
Discussions about perfect, better or no security are in danger of becoming
boring.
> 3) mmap problem
>
> . Making sure all apps use libmalware.so is trivial compared to
> solving 3).
You haven't answered what exactly is this libmalware.so, since you are the
only one mentioning it? It would be interesting to learn how it solves the
mmap problem, provides perfect security so it is acceptable, handles the
kernel NFS server serving malicious files, caters for applications which
do not use it, is better (more secure) than the kernel solution, provides
reasonalbe performance and is easier to maintain for the community? To
list only some of the requirements which have been mentioned so far.
--
Tvrtko A. Ursulin
Senior Software Engineer, Sophos
"Views and opinions expressed in this email are strictly those of the
author.
The contents has not been reviewed or approved by Sophos."
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Wed, 13 Aug 2008 06:46:46 -0400
"Press, Jonathan" <[email protected]> wrote:
> > -----Original Message-----
> > From: Pavel Machek [mailto:[email protected]]
> > Sent: Wednesday, August 13, 2008 6:28 AM
> > To: Press, Jonathan
> > Cc: [email protected]; Arjan van de Ven; Mihai Don??u; Adrian Bunk;
> > [email protected]; Greg KH; [email protected];
> linux-security-
> > [email protected]; [email protected]
> > Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to a
> linuxinterfaceforon access
> > scanning
> >
> > > I think everyone understands one side of the threat model, that is
> Linux machines
> > being carriers of infections aimed at other platforms. There are
> > many
> ways that
> > such infections can be stored, and many ways in which they can be
> communicated
> > to the target machines. There are so many that it would not be
> effective or efficient
> > for each such transfer application to be able to handle its own
> malware scanning,
> > which is the short statement of why centralized AV protection with
> notification
> > assistance from the kernel is appropriate.
> > >
> >
> > No.
> >
> > Proposed kernel solution did not work -- there still was write
> > vs. read race. You are right that it is not ok for each application
> > to do its own malware scanning, but libmalware.so that handles the
> > scanning seems very reasonable.
> >
> > And as applications _need_ to be modified for the write vs. read
> > race to be solved, libmalware.so looks like a way forward.
> >
> Pavel
>
> I am not sure what you are suggesting, and I may have missed the
> libmalware proposal (I don't see any mention of that specific idea in
> any other message). However, just to be clear... At no point did we
> suggest that the kernel would do any scanning. What we have been
> interested in is a mechanism that can allow a scanning application to
> be notified by the kernel of specific i/o events, for those events to
> be blocked by the kernel until a user-space scan is done, and then the
> user-space scan sends back allow or deny, at which point the i/o event
> returns to the caller -- either success or error. This is the only
> way that malware can be guaranteed of being detected when it is used
> (for local application purposes or for transmission to another
> platform) or created.
this is a very broad statement that ignores the LD_PRELOAD approach,
and thus not true.
>
> Also, a solution that requires applications to be modified will not
> work, because there is no way that we would be able to get ALL
> applications on the machines to be modified in the required ways. If
> ANY applications are not so modified, then you have an unacceptable
you don't need to modify applications to make them use a library...
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
On Wed, 13 Aug 2008 14:56:38 +0200
Pavel Machek <[email protected]> wrote:
> So you make sure all apps are modified. Distros are good at that, and
> modifications are not that hard.
>
> Plus, proposed solution already has three unacceptable holes:
>
> 1) it only catches known signatures
>
> 2) write vs. read race mentioned above
>
> 3) mmap problem
>
> . Making sure all apps use libmalware.so is trivial compared to
> solving 3).
the other thing is.. all applications ALREADY use such a library. It's
called "glibc".
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
Arjan van de Ven <[email protected]> wrote on 13/08/2008 14:54:01:
> > I am not sure what you are suggesting, and I may have missed the
> > libmalware proposal (I don't see any mention of that specific idea in
> > any other message). However, just to be clear... At no point did we
> > suggest that the kernel would do any scanning. What we have been
> > interested in is a mechanism that can allow a scanning application to
> > be notified by the kernel of specific i/o events, for those events to
> > be blocked by the kernel until a user-space scan is done, and then the
> > user-space scan sends back allow or deny, at which point the i/o event
> > returns to the caller -- either success or error. This is the only
> > way that malware can be guaranteed of being detected when it is used
> > (for local application purposes or for transmission to another
> > platform) or created.
>
> this is a very broad statement that ignores the LD_PRELOAD approach,
> and thus not true.
LD_PRELOAD does not solve at least knfsd and suid binaries. But we are
going in circles. :)
> > Also, a solution that requires applications to be modified will not
> > work, because there is no way that we would be able to get ALL
> > applications on the machines to be modified in the required ways. If
> > ANY applications are not so modified, then you have an unacceptable
>
> you don't need to modify applications to make them use a library...
Same is true for a kernel solution. Plus, it also works for those who make
system calls directly, knfsd and suid binaries, and we can have cheap and
ultra-efficient caching. Not much kernel code, even less complex kernel
code and unmeasurable impact when not used and compiled in. What are the
big technical objections to that?
--
Tvrtko A. Ursulin
Senior Software Engineer, Sophos
"Views and opinions expressed in this email are strictly those of the
author.
The contents has not been reviewed or approved by Sophos."
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Wed, 13 Aug 2008 15:16:12 +0100
[email protected] wrote:
> Arjan van de Ven <[email protected]> wrote on 13/08/2008 14:54:01:
>
> > > I am not sure what you are suggesting, and I may have missed the
> > > libmalware proposal (I don't see any mention of that specific
> > > idea in any other message). However, just to be clear... At no
> > > point did we suggest that the kernel would do any scanning. What
> > > we have been interested in is a mechanism that can allow a
> > > scanning application to be notified by the kernel of specific i/o
> > > events, for those events to be blocked by the kernel until a
> > > user-space scan is done, and then the user-space scan sends back
> > > allow or deny, at which point the i/o event returns to the caller
> > > -- either success or error. This is the only way that malware
> > > can be guaranteed of being detected when it is used (for local
> > > application purposes or for transmission to another platform) or
> > > created.
> >
> > this is a very broad statement that ignores the LD_PRELOAD approach,
> > and thus not true.
>
> LD_PRELOAD does not solve at least knfsd and suid binaries. But we
> are going in circles. :)
>
> > > Also, a solution that requires applications to be modified will
> > > not work, because there is no way that we would be able to get ALL
> > > applications on the machines to be modified in the required
> > > ways. If ANY applications are not so modified, then you have an
> > > unacceptable
> >
> > you don't need to modify applications to make them use a library...
>
> Same is true for a kernel solution. Plus, it also works for those who
> make system calls directly, knfsd and suid binaries, and we can have
> cheap and ultra-efficient caching. Not much kernel code, even less
> complex kernel code and unmeasurable impact when not used and
> compiled in. What are the big technical objections to that?
>
the biggest objection is the lack of security model description.
STILL nobody has answered Ted's questions.
And still the AV side of the argument keeps making circular arguments.
I'm not saying the kernel shouldn't be involved at all.
I can totally see a solution where we have a
sys_virus_scan(int fd)
that glibc calls at appropriate places (say every read() and mmap())
and that on the kernel side uses a cache to store which virus signature
version it was last scanned with, and if not new enough.. punts to some
userspace scanner for vetting.
but first someone needs to answer Ted's very basic questions or the
TALPA side really does look like a donkey in this argument.
--
If you want to reach me at my work email, use [email protected]
For development, discussion and tips for power savings,
visit http://www.lesswatts.org
Arjan van de Ven wrote on 13/08/2008 15:28:59:
> On Wed, 13 Aug 2008 15:16:12 +0100
> [email protected] wrote:
>
> > Arjan van de Ven <[email protected]> wrote on 13/08/2008 14:54:01:
> >
> > > > I am not sure what you are suggesting, and I may have missed the
> > > > libmalware proposal (I don't see any mention of that specific
> > > > idea in any other message). However, just to be clear... At no
> > > > point did we suggest that the kernel would do any scanning. What
> > > > we have been interested in is a mechanism that can allow a
> > > > scanning application to be notified by the kernel of specific i/o
> > > > events, for those events to be blocked by the kernel until a
> > > > user-space scan is done, and then the user-space scan sends back
> > > > allow or deny, at which point the i/o event returns to the caller
> > > > -- either success or error. This is the only way that malware
> > > > can be guaranteed of being detected when it is used (for local
> > > > application purposes or for transmission to another platform) or
> > > > created.
> > >
> > > this is a very broad statement that ignores the LD_PRELOAD approach,
> > > and thus not true.
> >
> > LD_PRELOAD does not solve at least knfsd and suid binaries. But we
> > are going in circles. :)
> >
> > > > Also, a solution that requires applications to be modified will
> > > > not work, because there is no way that we would be able to get ALL
> > > > applications on the machines to be modified in the required
> > > > ways. If ANY applications are not so modified, then you have an
> > > > unacceptable
> > >
> > > you don't need to modify applications to make them use a library...
> >
> > Same is true for a kernel solution. Plus, it also works for those who
> > make system calls directly, knfsd and suid binaries, and we can have
> > cheap and ultra-efficient caching. Not much kernel code, even less
> > complex kernel code and unmeasurable impact when not used and
> > compiled in. What are the big technical objections to that?
> >
>
> the biggest objection is the lack of security model description.
> STILL nobody has answered Ted's questions.
>
> And still the AV side of the argument keeps making circular arguments.
>
>
> I'm not saying the kernel shouldn't be involved at all.
> I can totally see a solution where we have a
> sys_virus_scan(int fd)
> that glibc calls at appropriate places (say every read() and mmap())
> and that on the kernel side uses a cache to store which virus signature
> version it was last scanned with, and if not new enough.. punts to some
> userspace scanner for vetting.
>
> but first someone needs to answer Ted's very basic questions or the
> TALPA side really does look like a donkey in this argument.
I think that the english saying is "It takes two to tangle". As much as I
am "guilty" for stepping out on Pavel's dejavu comments while at this
stage I agree we should be making a more coherent and systematic threat
model, you also haven't really contributed anything new and constructive
by jumping into my (or was it Jonathans?) reply to him. So I disagree that
"AV side" is the only side making circular argument and looking as a
donkey (thanks), especially in todays exchange.
So lets finish this thread now and hopefully we will have something real
to discuss soon.
--
Tvrtko A. Ursulin
Senior Software Engineer, Sophos
"Views and opinions expressed in this email are strictly those of the
author.
The contents has not been reviewed or approved by Sophos."
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
Hi!
Okay, so goal of libmalware.so is to "not allow data in the black list
to pass through Linux server". Threat model is windows machines trying
to copy infected files through the server. Viruses are not expected to
have shell access to either root or normal users on the server.
> Big snip since I am really only curious about libmalware.so.
>
> > Plus, proposed solution already has three unacceptable holes:
> >
> > 1) it only catches known signatures
> > 2) write vs. read race mentioned above
>
> Discussions about perfect, better or no security are in danger of becoming
> boring.
>
> > 3) mmap problem
> >
> > . Making sure all apps use libmalware.so is trivial compared to
> > solving 3).
>
> You haven't answered what exactly is this libmalware.so, since you are the
> only one mentioning it? It would be interesting to learn how it solves the
> mmap problem, provides perfect security so it is acceptable, handles the
> kernel NFS server serving malicious files, caters for applications which
> do not use it, is better (more secure) than the kernel solution, provides
> reasonalbe performance and is easier to maintain for the community? To
> list only some of the requirements which have been mentioned so far.
mmap problem: libmalware.so would not offer mmap() to applications (or
maybe it would do copy of the file, then allow mmap on the copy).
kernel NFS server is not handled; don't use it for serving Windows
clients. Not that you need performance for that, anyway.
Obviously libmalware.so will not help applications not using it. With
distributions, that's not a problem.
Unlike kernel solution, it does not contain races with read/write/mmap
-- untrusted files access is made through methods that can be safe.
You can query helper daemon for cache info; that should provide good
enough performance.
I never claimed it is easier to maintain than kernel solution; but
unlike kernel solution it actually _works_, 100% of time, for apps
using it.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Wed 2008-08-13 15:16:12, [email protected] wrote:
> Arjan van de Ven <[email protected]> wrote on 13/08/2008 14:54:01:
>
> > > I am not sure what you are suggesting, and I may have missed the
> > > libmalware proposal (I don't see any mention of that specific idea in
> > > any other message). However, just to be clear... At no point did we
> > > suggest that the kernel would do any scanning. What we have been
> > > interested in is a mechanism that can allow a scanning application to
> > > be notified by the kernel of specific i/o events, for those events to
> > > be blocked by the kernel until a user-space scan is done, and then the
> > > user-space scan sends back allow or deny, at which point the i/o event
> > > returns to the caller -- either success or error. This is the only
> > > way that malware can be guaranteed of being detected when it is used
> > > (for local application purposes or for transmission to another
> > > platform) or created.
> >
> > this is a very broad statement that ignores the LD_PRELOAD approach,
> > and thus not true.
>
> LD_PRELOAD does not solve at least knfsd and suid binaries. But we are
> going in circles. :)
Yes, there are about 5 suid binaries on typical linux system. Link
them to libmalware by hand.
> > > Also, a solution that requires applications to be modified will not
> > > work, because there is no way that we would be able to get ALL
> > > applications on the machines to be modified in the required ways. If
> > > ANY applications are not so modified, then you have an unacceptable
> >
> > you don't need to modify applications to make them use a library...
>
> Same is true for a kernel solution. Plus, it also works for those who make
> system calls directly, knfsd and suid binaries, and we can have cheap and
> ultra-efficient caching. Not much kernel code, even less complex kernel
> code and unmeasurable impact when not used and compiled in. What are the
> big technical objections to that?
That is does not work?
(Neither does LD_PRELOAD; it still has the old mmap problem. Too bad,
but at least you get 99.9% coverage of all the apps).
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
> -----Original Message-----
> From: Pavel Machek [mailto:[email protected]]
> Sent: Thursday, August 14, 2008 8:54 AM
> To: [email protected]
> Cc: Arjan van de Ven; Adrian Bunk; [email protected]; Greg KH; Press,
Jonathan;
> [email protected]; [email protected];
malware-
> [email protected]; Mihai Don??u
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
alinuxinterfaceforon access
> scanning
>
> Hi!
>
> Okay, so goal of libmalware.so is to "not allow data in the black list
> to pass through Linux server". Threat model is windows machines trying
> to copy infected files through the server.
That's only part of the threat model.
> Viruses are not expected to have shell access to either root or normal
> users on the server.
That's a big exception.
> it actually _works_, 100% of time, for apps using it.
Again that's a big condition.
Jon Press
> > LD_PRELOAD does not solve at least knfsd and suid binaries. But we are
> > going in circles. :)
>
> Yes, there are about 5 suid binaries on typical linux system. Link
> them to libmalware by hand
And knfsd ?
Oh yes you don't seem to have an answer just manure to throw
On Thu 2008-08-14 21:06:04, Alan Cox wrote:
> > > LD_PRELOAD does not solve at least knfsd and suid binaries. But we are
> > > going in circles. :)
> >
> > Yes, there are about 5 suid binaries on typical linux system. Link
> > them to libmalware by hand
>
> And knfsd ?
>
> Oh yes you don't seem to have an answer just manure to throw
I think I acknowledged inability to handle knfsd... just don't use
knfsd, then. Its not like nfs is critical for communicating with
Windows, right?
That is still better than proposed solution here -- TALPA breaks with
mmap, or when someone does read too soon after write.
Marketing "antivirus" system that is racy be design seems to deserve
some manure.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
Hi!
> > Okay, so goal of libmalware.so is to "not allow data in the black list
> > to pass through Linux server". Threat model is windows machines trying
> > to copy infected files through the server.
>
> That's only part of the threat model.
Yes, that's the part libmalware.so proposal solves. Given scary number
of 0 Linux viruses in wild, it seems to solve the problem pretty well.
> > it actually _works_, 100% of time, for apps using it.
>
> Again that's a big condition.
Yep, so... why don't you propose something better? I'm pretty sure
100% reliable scanning is impossible without modifying applications,
but hey, you can prove me wrong.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Fri, 15 Aug 2008 00:39:18 +0200
Pavel Machek <[email protected]> wrote:
> > > Okay, so goal of libmalware.so is to "not allow data in the black list
> > > to pass through Linux server". Threat model is windows machines trying
> > > to copy infected files through the server.
> >
> > That's only part of the threat model.
>
> Yes, that's the part libmalware.so proposal solves. Given scary number
> of 0 Linux viruses in wild, it seems to solve the problem pretty well.
If you're trolling, you're not being very good at it.
Just because you cannot easily infect a Linux system from a
user application does not mean malware cannot do all kinds
of damage with user privileges. Think of a key sniffer (using
the same interface that the X screensavers use) or a spam bot
running with user privileges.
Firefox, OpenOffice.org and other (mostly desktop) programs are
extremely large and complex, deal with untrusted data on a daily
basis and could be used to spread worms and get malware onto systems.
The old DOS model of "you need to infect system binaries" is not
a good description of how today's malware works. Malware is not
there to infect a system "as much as possible", but to accomplish
actual malice.
Consequently, the number of acceptable attack vectors on a system
is pretty large and we should protect against these kinds of
programs.
It would be good to get this additional layer of protection against
malware in place, before people start developing Linux malware.
--
All rights reversed.
On Thu, Aug 14, 2008 at 08:00:05PM -0400, Rik van Riel wrote:
> > Yes, that's the part libmalware.so proposal solves. Given scary number
> > of 0 Linux viruses in wild, it seems to solve the problem pretty well.
>
> If you're trolling, you're not being very good at it.
>
> Just because you cannot easily infect a Linux system from a
> user application does not mean malware cannot do all kinds
> of damage with user privileges. Think of a key sniffer (using
> the same interface that the X screensavers use) or a spam bot
> running with user privileges.
But Pavel is raising a good question. In Eric's proposed threat
model, he claimed the only thing that he was trying to solve was
"scanning". Just file scanning. That implies no root privileges, but
it also implied that he wasn't worried about malware running with user
privileges, either. Presumbly, that would be caught and stopped by
the file scanner before the malware had a chance to run; that is the
execve(2) system call would also be blocked until the executable was
scanned.
So if that is the threat model, then the only thing libmalware.so
doesn't solve is knfsd access, and it should be evaluated on that
basis. If the threat model *does* include malware which is **not**
caught by the AV scanner, and is running with user privileges, then
there are a whole host of other attacks that we have to worry about.
So let's be real clear, up front, what the threat model is, and avoid
changing the model around to rule out solutions that don't fit the
initially preconceived one. That's how you get to the TSA
confiscating water bottles in airport security lines.
- Ted
On Thu, 14 Aug 2008 20:43:35 -0400
Theodore Tso <[email protected]> wrote:
> But Pavel is raising a good question. In Eric's proposed threat
> model, he claimed the only thing that he was trying to solve was
> "scanning". Just file scanning. That implies no root privileges, but
> it also implied that he wasn't worried about malware running with user
> privileges, either. Presumbly, that would be caught and stopped by
> the file scanner before the malware had a chance to run;
You bring up a very good point - malware does not need to
be stored on a filesystem in order to run or cause damage.
--
All rights reversed.
On Thu, 2008-08-14 at 20:43 -0400, Theodore Tso wrote:
> On Thu, Aug 14, 2008 at 08:00:05PM -0400, Rik van Riel wrote:
> > > Yes, that's the part libmalware.so proposal solves. Given scary number
> > > of 0 Linux viruses in wild, it seems to solve the problem pretty well.
> >
> > If you're trolling, you're not being very good at it.
> >
> > Just because you cannot easily infect a Linux system from a
> > user application does not mean malware cannot do all kinds
> > of damage with user privileges. Think of a key sniffer (using
> > the same interface that the X screensavers use) or a spam bot
> > running with user privileges.
>
> But Pavel is raising a good question. In Eric's proposed threat
> model, he claimed the only thing that he was trying to solve was
> "scanning". Just file scanning. That implies no root privileges, but
> it also implied that he wasn't worried about malware running with user
> privileges, either. Presumbly, that would be caught and stopped by
> the file scanner before the malware had a chance to run; that is the
> execve(2) system call would also be blocked until the executable was
> scanned.
>
> So if that is the threat model, then the only thing libmalware.so
> doesn't solve is knfsd access, and it should be evaluated on that
> basis. If the threat model *does* include malware which is **not**
> caught by the AV scanner, and is running with user privileges, then
> there are a whole host of other attacks that we have to worry about.
> So let's be real clear, up front, what the threat model is, and avoid
> changing the model around to rule out solutions that don't fit the
> initially preconceived one. That's how you get to the TSA
> confiscating water bottles in airport security lines.
No, I'm not claiming to protect against running processes. I'll leave
that for SELinux.
I haven't seen this supposed libmalware.so so take anything I say with a
grain of sand. But I take it that the solutions to the problems are
'don't do that.'
aka malware is allowed to flow freely across linux nfs servers. Great,
I'm sure corperate IT organizations are going to love knowing there
isn't a darn thing they can do to protect their nfs server from being
storage grounds other than hope they can control all of the border.
And I still don't get this 'mmap problem' that I don't solve that
libmalware magically solves. What? don't use mmap? I certainly hope
not.
Are we seriously considering that the right thing to do is to try to
push malware scanning to every project on sourceforge? At least putting
a solution inside glibc wasn't completely insane, I just think for
numerous reasons we've seen on list for the last 2 weeks not a better
idea. In any case, having an application have to make special calls to
handle 'untrusted' data is basically like turning the keys to the castle
over on every exploit. No, I might not make promises about subverted
applications, but that doesn't mean I have to just open all the doors.
And anything that requires explicit application help is just that. Talk
about theater.
-Eric
On Thu, 14 Aug 2008, Eric Paris wrote:
>> But Pavel is raising a good question. In Eric's proposed threat
>> model, he claimed the only thing that he was trying to solve was
>> "scanning". Just file scanning. That implies no root privileges, but
>> it also implied that he wasn't worried about malware running with user
>> privileges, either. Presumbly, that would be caught and stopped by
>> the file scanner before the malware had a chance to run; that is the
>> execve(2) system call would also be blocked until the executable was
>> scanned.
>>
>> So if that is the threat model, then the only thing libmalware.so
>> doesn't solve is knfsd access, and it should be evaluated on that
>> basis. If the threat model *does* include malware which is **not**
>> caught by the AV scanner, and is running with user privileges, then
>> there are a whole host of other attacks that we have to worry about.
>> So let's be real clear, up front, what the threat model is, and avoid
>> changing the model around to rule out solutions that don't fit the
>> initially preconceived one. That's how you get to the TSA
>> confiscating water bottles in airport security lines.
>
> No, I'm not claiming to protect against running processes. I'll leave
> that for SELinux.
>
> I haven't seen this supposed libmalware.so so take anything I say with a
> grain of sand. But I take it that the solutions to the problems are
> 'don't do that.'
libmalware.so is shorthand for 'have a userspace library do the scanning
and handle the open'
> aka malware is allowed to flow freely across linux nfs servers. Great,
> I'm sure corperate IT organizations are going to love knowing there
> isn't a darn thing they can do to protect their nfs server from being
> storage grounds other than hope they can control all of the border.
nfs isn't secure anyway so you better control the border.
it's only the kernel nfs daemon that won't use the library, so the answer
is don't use that daemon. I beleive that there is a FUSE NFS option, or if
nothing else, mount a filesystem with FUSE (linked against libmalware.so)
and then export the result vi knfsd.
> And I still don't get this 'mmap problem' that I don't solve that
> libmalware magically solves. What? don't use mmap? I certainly hope
> not.
libmalware can intercept the mmap call and decide to either prohibit it,
copy the file so that the program is operating from it's own copy, or do
something else (all dependant on policy, as defined in userspace for this
library)
> Are we seriously considering that the right thing to do is to try to
> push malware scanning to every project on sourceforge? At least putting
> a solution inside glibc wasn't completely insane, I just think for
> numerous reasons we've seen on list for the last 2 weeks not a better
> idea. In any case, having an application have to make special calls to
> handle 'untrusted' data is basically like turning the keys to the castle
> over on every exploit. No, I might not make promises about subverted
> applications, but that doesn't mean I have to just open all the doors.
> And anything that requires explicit application help is just that. Talk
> about theater.
libmalware.so and a modified glibc are not mutually exclusive. you distros
could offer versions of glibc that include libmalware.
again, libmalware.so is not referring to any specific body of code, it's
referring to the concept that the handling of open/mmap/read/etc and
scanning is done via a userspace library rather then by the kernel. if
everyone can agree on that concept then hashing out the details of _which_
library it gets put in is a smaller detail.
this isn't designed to require every app that wants this protection to
have to change it's code.
David lang
On Thu, 14 Aug 2008, [email protected] wrote:
> again, libmalware.so is not referring to any specific body of code, it's
> referring to the concept that the handling of open/mmap/read/etc and scanning
> is done via a userspace library rather then by the kernel. if everyone can
> agree on that concept then hashing out the details of _which_ library it gets
> put in is a smaller detail.
one reason to layer scanners is that you could have one that just checks
to see if the file was deployed from a OS package, if it was (and still
has the same hash as the package manager thinks it should have) it sets a
flag that other scanners could look for (if they see it they can skip
scanning the file)
David Lang
On Thu, 14 Aug 2008, [email protected] wrote:
> On Thu, 14 Aug 2008, [email protected] wrote:
>
>> again, libmalware.so is not referring to any specific body of code, it's
>> referring to the concept that the handling of open/mmap/read/etc and
>> scanning is done via a userspace library rather then by the kernel. if
>> everyone can agree on that concept then hashing out the details of _which_
>> library it gets put in is a smaller detail.
>
> one reason to layer scanners is that you could have one that just checks to
> see if the file was deployed from a OS package, if it was (and still has the
> same hash as the package manager thinks it should have) it sets a flag that
> other scanners could look for (if they see it they can skip scanning the
> file)
actually, instead of trying to have one scanner trust the results of
another, a better approach would be for libmalware to look for the package
manager approval and if it finds that it could skip asking the other
scanners to look at it.
this sort of thing can easily be done in userspace libraries, but you
definantly would _not_ want to do in a kernel-level enforcement mechanism.
David Lang
> So if that is the threat model, then the only thing libmalware.so
> doesn't solve is knfsd access, and it should be evaluated on that
And static binaries, and other kernel modular file I/O done on behalf of
applications, and making a library work well which is *hard*, and
labelling and scalability, and sharing a cache between users, and
aggregating events across processes .. and a few other things.
You have a lot of shared state here.
> So let's be real clear, up front, what the threat model is, and avoid
> changing the model around to rule out solutions that don't fit the
> initially preconceived one. That's how you get to the TSA
> confiscating water bottles in airport security lines.
I don't think this is useful. I don't actually care what the virus folks
want to implement. I would rather see useful general purpose interfaces
that are practical and make sense. I cdon't really care whether people
use the write() system call to do sensible or dumb things. What we should
care about is the write syscall.
There seem so far to be two independent requests
* Noticing a file has recently changed, possibly with information about
changes and possibly being able to aggregate/time that for scalability.
This has a need to be able to accurately reference which file. This is
needed for good content indexing, and virus people want it for certain
kinds of scanning (post write, background etc). Doing this solely as a
final close notifier seems to be insufficient (as it may never close).
* Being able to pause an open pending some action. Examples include HSM
and dubiously some kinds of virus check. Problems here include the fact
it can only meaningfully be done for first open (which is fine for HSM)
and that the notion of an exclusive open, or of repeated clearly defined
open/read-write/close/done sequences are actually quite foreign to Unix
like systems.
We shouldn't need to care what people do with good interface. What
matters is in your airport example is that at the infrastructure level
there is a point you can choose to do scanning and we agree where.
Whether people use this to provide a Starbucks or goons with rubber
gloves who take away babies milk is an application layer problem.
Alan
On Fri, Aug 15, 2008 at 09:35:13AM +0100, Alan Cox wrote:
> We shouldn't need to care what people do with good interface. What
> matters is in your airport example is that at the infrastructure level
> there is a point you can choose to do scanning and we agree where.
> Whether people use this to provide a Starbucks or goons with rubber
> gloves who take away babies milk is an application layer problem.
If it's a good interface that also happens to address HSM/DMAPI
functionality, as well as a more efficient way for trackerd to work, I
agree completely. I think you will agree the proposed TALPA interface
is a bit too virus-scanner specific, though? Especially with explicit
talk of specialized (persistent or not) "clean/dirty/infected" bits
that the kernel would store in the inode for the benefit of the AV
scanner? That's rather optimized for the goons-with-rubber-gloves
that-make-mothers-drink-their-own-breast-to-prove-it's-not-explosives
crowd, I think...
- Ted
> -----Original Message-----
> From: Theodore Tso [mailto:[email protected]]
> Sent: Friday, August 15, 2008 7:35 AM
> To: Alan Cox
> Cc: Rik van Riel; Pavel Machek; Press, Jonathan; [email protected];
Adrian Bunk;
> Mihai Don??u; [email protected];
[email protected]; linux-
> [email protected]; Arjan van de Ven
> Subject: Re: [malware-list] [RFC 0/5] [TALPA] Intro to
alinuxinterfaceforonaccess
> scanning
>
> On Fri, Aug 15, 2008 at 09:35:13AM +0100, Alan Cox wrote:
> > We shouldn't need to care what people do with good interface. What
> > matters is in your airport example is that at the infrastructure
level
> > there is a point you can choose to do scanning and we agree where.
> > Whether people use this to provide a Starbucks or goons with rubber
> > gloves who take away babies milk is an application layer problem.
An interesting example is (please don't scream everyone -- it's just an
illustration of one approach to the problem) NetWare. (In fact, we
still have a lot of active NetWare customers, so the fact that it is an
archaic OS is not really the issue.) NetWare takes the kitchen-sink
approach. It has an interface that allows notification on a whole host
of i/o events, way more than we have ever found useful, and the
application can register for the ones it wants and go from there. The
kernel does not care what the application's inherent logic is, as long
as the application passes control back with some appropriate return
information (in other words, allow or deny in the context of malware)
within a reasonable amount of time. I have no idea what purpose it was
written for originally, and there are flaws to be sure, but I know that
it is used successfully for HSMs and anti-malware products.
> If it's a good interface that also happens to address HSM/DMAPI
> functionality, as well as a more efficient way for trackerd to work, I
> agree completely. I think you will agree the proposed TALPA interface
> is a bit too virus-scanner specific, though? Especially with explicit
> talk of specialized (persistent or not) "clean/dirty/infected" bits
> that the kernel would store in the inode for the benefit of the AV
> scanner? That's rather optimized for the goons-with-rubber-gloves
> that-make-mothers-drink-their-own-breast-to-prove-it's-not-explosives
> crowd, I think...
That may just be a question of terminology. If the bits are construed
not as clean/dirty/infected, but as "I care about this file" vs. "I
don't care about this file" then the rubber gloves come off.
Jon Press
On Fri, Aug 15, 2008 at 08:57:48AM -0400, Press, Jonathan wrote:
> That may just be a question of terminology. If the bits are construed
> not as clean/dirty/infected, but as "I care about this file" vs. "I
> don't care about this file" then the rubber gloves come off.
Sure, as long as we're very clear about the semantics of the bits. If
the bits are not persistent, but which get dropped if the inode is
every evicted from memory, and it's considered OK, or even desirable,
to rescan the file when it is brought back into memory, that may be
acceptable to the rubber gloves folks (make people go through lots
superflous of security scans, even when they are transfering betewen
flights --- security is always more important than passengers'
convenience!), but perhaps not to other applications such as file
indexers, who would view rescanning files that have already been
scanned, and not have been modified, as a waste of time, battery, CPU
and disk bandwidth, etc.
As I understand it, the TALPA proposal had non-persistent
clean/dirty/infected bits.
- Ted
[email protected] wrote on 2008-08-15 14:16:21:
> On Fri, Aug 15, 2008 at 08:57:48AM -0400, Press, Jonathan wrote:
> > That may just be a question of terminology. If the bits are construed
> > not as clean/dirty/infected, but as "I care about this file" vs. "I
> > don't care about this file" then the rubber gloves come off.
>
> Sure, as long as we're very clear about the semantics of the bits. If
> the bits are not persistent, but which get dropped if the inode is
> every evicted from memory, and it's considered OK, or even desirable,
> to rescan the file when it is brought back into memory, that may be
> acceptable to the rubber gloves folks (make people go through lots
> superflous of security scans, even when they are transfering betewen
> flights --- security is always more important than passengers'
> convenience!), but perhaps not to other applications such as file
> indexers, who would view rescanning files that have already been
> scanned, and not have been modified, as a waste of time, battery, CPU
> and disk bandwidth, etc.
>
> As I understand it, the TALPA proposal had non-persistent
> clean/dirty/infected bits.
>
> - Ted
Yes the current proposal has temporary markers in the in-memory
representation if inodes.
This is a problem for current anti-malware scanning, as virus data updates
come every few hours
(at which point the entire clean/infected state has to be cleared), so the
loss after a reboot is
limited.
--
Douglas Leeder
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
[email protected] wrote on 2008-08-15 14:22:27:
> [email protected] wrote on 2008-08-15 14:16:21:
>
> > On Fri, Aug 15, 2008 at 08:57:48AM -0400, Press, Jonathan wrote:
> > > That may just be a question of terminology. If the bits are
construed
> > > not as clean/dirty/infected, but as "I care about this file" vs. "I
> > > don't care about this file" then the rubber gloves come off.
> >
> > Sure, as long as we're very clear about the semantics of the bits. If
> > the bits are not persistent, but which get dropped if the inode is
> > every evicted from memory, and it's considered OK, or even desirable,
> > to rescan the file when it is brought back into memory, that may be
> > acceptable to the rubber gloves folks (make people go through lots
> > superflous of security scans, even when they are transfering betewen
> > flights --- security is always more important than passengers'
> > convenience!), but perhaps not to other applications such as file
> > indexers, who would view rescanning files that have already been
> > scanned, and not have been modified, as a waste of time, battery, CPU
> > and disk bandwidth, etc.
> >
> > As I understand it, the TALPA proposal had non-persistent
> > clean/dirty/infected bits.
> >
> > - Ted
>
> Yes the current proposal has temporary markers in the in-memory
> representation if inodes.
>
> This is a problem for current anti-malware scanning, as virus data
updates
> come every few hours
> (at which point the entire clean/infected state has to be cleared), so
the
> loss after a reboot is
> limited.
Of course I mean't NOT a problem...
--
Douglas Leeder
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Fri, Aug 15, 2008 at 02:22:27PM +0100, [email protected] wrote:
>
> This is a problem for current anti-malware scanning, as virus data updates
> come every few hours
Every few hours?!? I hadn't noticed Windows AV programs getting
updates that frequently, at least not the ones that I've been familiar
with. (Semantec, Norton, McAfee)
- Ted
Theodore Tso <[email protected]> wrote on 2008-08-15 14:55:37:
> On Fri, Aug 15, 2008 at 02:22:27PM +0100, [email protected]
wrote:
> >
> > This is a problem for current anti-malware scanning, as virus data
updates
> > come every few hours
>
> Every few hours?!? I hadn't noticed Windows AV programs getting
> updates that frequently, at least not the ones that I've been familiar
> with. (Semantec, Norton, McAfee)
>From one of our Linux machines that configured to update normally:
Fri 15 Aug 2008 15:02:23 BST
Fri 15 Aug 2008 12:52:19 BST
Fri 15 Aug 2008 08:32:24 BST
Fri 15 Aug 2008 04:12:27 BST
Fri 15 Aug 2008 02:02:23 BST
That was the sort of time period I remember being told -
but the AV products probably don't make a song and
dance about a small virus data updates so you probably wouldn't notice
them.
--
Douglas Leeder
Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon,
OX14 3YP, United Kingdom.
Company Reg No 2096520. VAT Reg No GB 348 3873 20.
On Fri, 15 Aug 2008 09:55:37 EDT, Theodore Tso said:
> On Fri, Aug 15, 2008 at 02:22:27PM +0100, [email protected] wrote:
> >
> > This is a problem for current anti-malware scanning, as virus data updates
> > come every few hours
>
> Every few hours?!? I hadn't noticed Windows AV programs getting
> updates that frequently, at least not the ones that I've been familiar
> with. (Semantec, Norton, McAfee)
Try running a mail server that provides virus scanning for a large population
of 100K or so mailboxes. You end up pulling from your vendor on an hourly
basis, just because a virus on a burn through your userbase can toast you that
quickly.
Since 9AM Sunday (is now 11:30AM Friday as I write this), we've pulled new
signatures 33 times (one new signature each time in this case) from our vendor.
So yeah, about once every 3-4 hours we get a new updated one for a new variant
of whatever. I've seen times when we've pulled a new signature file 3 hours in
a row, and each time there were 10-12 new variants, so averaging 12/hour...
> > So if that is the threat model, then the only thing libmalware.so
> > doesn't solve is knfsd access, and it should be evaluated on that
>
> And static binaries, and other kernel modular file I/O done on behalf of
> applications, and making a library work well which is *hard*, and
> labelling and scalability, and sharing a cache between users, and
> aggregating events across processes .. and a few other things.
Well, I believe libmalware.so works with threat model I described; of
course it will not protect statically linked sambad (unless you
statically link it with libmalware.so, which you should do). I'm not
actually advocating LD_PRELOAD. There are enough userspace nfsds
around, but yes, you can't use libmalware.so with knfsd.
[You could do something like fuse filesystem linked with libmalware,
and make knfsd export that, and put applications you can't link with
libmalware to use that. But that's a _hack_.]
> There seem so far to be two independent requests
>
> * Noticing a file has recently changed, possibly with information about
> changes and possibly being able to aggregate/time that for scalability.
> This has a need to be able to accurately reference which file. This is
> needed for good content indexing, and virus people want it for certain
> kinds of scanning (post write, background etc). Doing this solely as a
> final close notifier seems to be insufficient (as it may never
> close).
Agreed, we need this.
> * Being able to pause an open pending some action. Examples include HSM
> and dubiously some kinds of virus check. Problems here include the fact
> it can only meaningfully be done for first open (which is fine for HSM)
> and that the notion of an exclusive open, or of repeated clearly defined
> open/read-write/close/done sequences are actually quite foreign to Unix
> like systems.
Do HSMs really get implemented like that? I'd expect HSM to be
something like fuse or unionfs... and when it is confined to one
filesystem, you can use existing solutions.
I don't like blocking at open at all, and I don't think blocking at
open is enough for antivirus scanner.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
Hi!
> And I still don't get this 'mmap problem' that I don't solve that
> libmalware magically solves. What? don't use mmap? I certainly hope
> not.
Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
which is basically shared memory -- is fundamentally incompatible with
reliable virus scanning.
...or do you have a reasonable solution for mmap?
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
> On Thu, 14 Aug 2008, Eric Paris wrote:
>
> >>But Pavel is raising a good question. In Eric's proposed threat
> >>model, he claimed the only thing that he was trying to solve was
> >>"scanning". Just file scanning. That implies no root privileges, but
> >>it also implied that he wasn't worried about malware running with user
> >>privileges, either. Presumbly, that would be caught and stopped by
> >>the file scanner before the malware had a chance to run; that is the
> >>execve(2) system call would also be blocked until the executable was
> >>scanned.
> >>
> >>So if that is the threat model, then the only thing libmalware.so
> >>doesn't solve is knfsd access, and it should be evaluated on that
> >>basis. If the threat model *does* include malware which is **not**
> >>caught by the AV scanner, and is running with user privileges, then
> >>there are a whole host of other attacks that we have to worry about.
> >>So let's be real clear, up front, what the threat model is, and avoid
> >>changing the model around to rule out solutions that don't fit the
> >>initially preconceived one. That's how you get to the TSA
> >>confiscating water bottles in airport security lines.
> >
> >No, I'm not claiming to protect against running processes. I'll leave
> >that for SELinux.
> >
> >I haven't seen this supposed libmalware.so so take anything I say with a
> >grain of sand. But I take it that the solutions to the problems are
> >'don't do that.'
>
> libmalware.so is shorthand for 'have a userspace library do the scanning
> and handle the open'
(snip). Agreed, you explained it better than I would.
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Mon, 18 Aug 2008, Pavel Machek wrote:
> Hi!
>
>> And I still don't get this 'mmap problem' that I don't solve that
>> libmalware magically solves. What? don't use mmap? I certainly hope
>> not.
>
> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
> which is basically shared memory -- is fundamentally incompatible with
> reliable virus scanning.
>
> ...or do you have a reasonable solution for mmap?
mmap has a few different problems
1. intercepting reads and writes to take action at that time
2. the fact that two programs can use it as an inter-process communication
mechanism.
if you are worried about the IPC aspects, all you can do is forbid it,
along with shared memory, pipes, network connections, etc. none of these
provide you with a 'final result' that can be scanned, and as others have
pointed out there are too many ways to do things that get assembled by the
far side to try and catch all malware in them.
as for intercepting reads and writes. the approach I outlined addresses
this by having the kernel mark thefile as dirty if any writes happen, and
checking the file status at the time of doing the mmap instead of trying
to do it when the file is accessed after the mmap.
David Lang
Hi!
> >>And I still don't get this 'mmap problem' that I don't solve that
> >>libmalware magically solves. What? don't use mmap? I certainly hope
> >>not.
> >
> >Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
> >which is basically shared memory -- is fundamentally incompatible with
> >reliable virus scanning.
> >
> >...or do you have a reasonable solution for mmap?
>
>
> mmap has a few different problems
>
> 1. intercepting reads and writes to take action at that time
>
> 2. the fact that two programs can use it as an inter-process communication
> mechanism.
...can and will use it as an IPC. So we need to modify some
applications.
Rather than modify all the applications using mmap (you can't tell if
the other side is going to use it for shared memory... right?), we
could simply modify all the Windows-facing applications using mmap.
> if you are worried about the IPC aspects, all you can do is forbid it,
Can you automatically tell if applications are using mmap for IPC?
BTW in another mail you wanted to include /var/log/syslog from
scanning. You should not be doing that if syslog is exported to
Windows systems. Of course, you can get away with scanning syslog when
Windows client tries to read it, which should be acceptable...
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Mon, 18 Aug 2008, Pavel Machek wrote:
>>>> And I still don't get this 'mmap problem' that I don't solve that
>>>> libmalware magically solves. What? don't use mmap? I certainly hope
>>>> not.
>>>
>>> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
>>> which is basically shared memory -- is fundamentally incompatible with
>>> reliable virus scanning.
>>>
>>> ...or do you have a reasonable solution for mmap?
>>
>>
>> mmap has a few different problems
>>
>> 1. intercepting reads and writes to take action at that time
>>
>> 2. the fact that two programs can use it as an inter-process communication
>> mechanism.
>
> ...can and will use it as an IPC. So we need to modify some
> applications.
>
> Rather than modify all the applications using mmap (you can't tell if
> the other side is going to use it for shared memory... right?), we
> could simply modify all the Windows-facing applications using mmap.
>
>> if you are worried about the IPC aspects, all you can do is forbid it,
>
> Can you automatically tell if applications are using mmap for IPC?
no, but can you tell at the time of the mmap command if anyone has it
opened for writing? if you can then you can just not allow the mmap in
thid case (policy decision by userspace, as such it can try to look at
what other programs are accessing it via mmap to decide if it should allow
it or not)
> BTW in another mail you wanted to include /var/log/syslog from
> scanning. You should not be doing that if syslog is exported to
> Windows systems. Of course, you can get away with scanning syslog when
> Windows client tries to read it, which should be acceptable...
I listed that as an example of what I would consider a sane policy. by
doing the checking is a userspace library different binaries can be linked
against different libraries by the sysadmin/distro to decide which ones
need to do what checking. there's nothing inherent in the mechanism that
foces the policy in any direction.
David Lang
Pavel Machek wrote:
> Hi!
>
>
>>>> And I still don't get this 'mmap problem' that I don't solve that
>>>> libmalware magically solves. What? don't use mmap? I certainly hope
>>>> not.
>>>>
>>> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
>>> which is basically shared memory -- is fundamentally incompatible with
>>> reliable virus scanning.
>>>
>>> ...or do you have a reasonable solution for mmap?
>>>
>> mmap has a few different problems
>>
>> 1. intercepting reads and writes to take action at that time
>>
>> 2. the fact that two programs can use it as an inter-process communication
>> mechanism.
>>
>
> ...can and will use it as an IPC. So we need to modify some
> applications.
>
> Rather than modify all the applications using mmap (you can't tell if
> the other side is going to use it for shared memory... right?), we
> could simply modify all the Windows-facing applications using mmap.
>
>
>> if you are worried about the IPC aspects, all you can do is forbid it,
>>
>
> Can you automatically tell if applications are using mmap for IPC?
>
> BTW in another mail you wanted to include /var/log/syslog from
> scanning. You should not be doing that if syslog is exported to
> Windows systems. Of course, you can get away with scanning syslog when
> Windows client tries to read it, which should be acceptable...
> Pavel
>
There is a solution to this whole scanning thing, but I've been
reluctant to suggest it, and it will be pretty obvious why y'all
probably don't want to try it. Just to be sure the options are
out on the table, here goes.
Define an xattr, let's call it "UNSCANNED", which has as its value
a timestamp. A regular file with this attribute cannot be executed
or opened,(exec or open hangs or fails, either behavior has merit
and downsides) and it requires privilege (perhaps CAP_MAC_ADMIN)
to remove the attribute. File creation attaches the attribute. Any
open for write attaches the attribute.
Your scanner runs with privilege (say CAP_MAC_OVERRIDE) and passes
judgment on files with this attribute, removing either the file, if
it is Evil, or the attribute, if it is Good. The scanner is invoked
when a file that was open with write access is closed. This can be
done using mechanisms already discussed on this thread.
If you like, you could use a "SCANNED" attribute instead of an
"UNSCANNED" attribute, and reverse the sense of the test. The
major difference will show up on filesystems that don't support
xattrs. The implications should be obvious.
Now at the beginning I said that you wouldn't like this scheme,
and it shouldn't take a security expert to see the usability problems.
This is how an old school trusted systems junkie (like me) would do
it, and I don't see a better way that would actually achieve the
stated goals. If you wanted you could implement an LSM to do the
labeling bit in a day or two, the notification in about the same
time, which would leave the scanner as the long pole in your
development schedule.
P.S. - Library based security doesn't work.
On Mon, 18 Aug 2008 00:58:44 +0200
Pavel Machek <[email protected]> wrote:
> Rather than modify all the applications using mmap (you can't tell if
> the other side is going to use it for shared memory... right?), we
> could simply modify all the Windows-facing applications using mmap.
If web browsers, office suites and mail clients on Windows
have certain kinds of vulnerabilities, it is safe to assume
that the same programs on Linux will have similar problems.
Can we please get rid of the idea that "Windows facing" is
where the whole malware problem is?
As for how to solve it - lets try to come up with a solution
that is reasonably high performance and can be used for more
than just malware scanning.
Using the same code for things like HSM would be nice.
--
All rights reversed.
On Sun, 17 Aug 2008, Casey Schaufler wrote:
> Pavel Machek wrote:
>>>>> And I still don't get this 'mmap problem' that I don't solve that
>>>>> libmalware magically solves. What? don't use mmap? I certainly hope
>>>>> not.
>>>>>
>>>> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
>>>> which is basically shared memory -- is fundamentally incompatible with
>>>> reliable virus scanning.
>>>>
>>>> ...or do you have a reasonable solution for mmap?
>>>>
>>> mmap has a few different problems
>>>
>>> 1. intercepting reads and writes to take action at that time
>>>
>>> 2. the fact that two programs can use it as an inter-process communication
>>> mechanism.
>>>
>>
>> ...can and will use it as an IPC. So we need to modify some
>> applications.
>>
>> Rather than modify all the applications using mmap (you can't tell if
>> the other side is going to use it for shared memory... right?), we
>> could simply modify all the Windows-facing applications using mmap.
>>
>>
>>> if you are worried about the IPC aspects, all you can do is forbid it,
>>
>> Can you automatically tell if applications are using mmap for IPC?
>>
>> BTW in another mail you wanted to include /var/log/syslog from
>> scanning. You should not be doing that if syslog is exported to
>> Windows systems. Of course, you can get away with scanning syslog when
>> Windows client tries to read it, which should be acceptable...
>> Pavel
>>
>
> There is a solution to this whole scanning thing, but I've been
> reluctant to suggest it, and it will be pretty obvious why y'all
> probably don't want to try it. Just to be sure the options are
> out on the table, here goes.
>
> Define an xattr, let's call it "UNSCANNED", which has as its value
> a timestamp. A regular file with this attribute cannot be executed
> or opened,(exec or open hangs or fails, either behavior has merit
> and downsides) and it requires privilege (perhaps CAP_MAC_ADMIN)
> to remove the attribute. File creation attaches the attribute. Any
> open for write attaches the attribute.
>
> Your scanner runs with privilege (say CAP_MAC_OVERRIDE) and passes
> judgment on files with this attribute, removing either the file, if
> it is Evil, or the attribute, if it is Good. The scanner is invoked
> when a file that was open with write access is closed. This can be
> done using mechanisms already discussed on this thread.
>
> If you like, you could use a "SCANNED" attribute instead of an
> "UNSCANNED" attribute, and reverse the sense of the test. The
> major difference will show up on filesystems that don't support
> xattrs. The implications should be obvious.
>
> Now at the beginning I said that you wouldn't like this scheme,
> and it shouldn't take a security expert to see the usability problems.
> This is how an old school trusted systems junkie (like me) would do
> it, and I don't see a better way that would actually achieve the
> stated goals. If you wanted you could implement an LSM to do the
> labeling bit in a day or two, the notification in about the same
> time, which would leave the scanner as the long pole in your
> development schedule.
did you read the proposal I wrote up? it's similar to (but more flexible
than) what you are saying
it would allow for multiple 'scanned' tags (to allow for multiple scanning
programs). the kernel would clear the tags when the file is dirtied (not
when it is closed, the file may not be closed for weeks after all), there
is a mechanism to tell the scanner(s) that a file was modified (rather
then having to scan the entire filesystem), and if the scan was not done
yet when a file is opened the scanner(s) can be invoked at that time.
> P.S. - Library based security doesn't work.
why not? (with the appropriate kernel support)
it actually wouldn't be hard to have the kernel check the xattr, what
would be hard is hving the kernel know when it should then invoke the
scanner(s) and when it shouldn't. This is a policy decision that doesn't
belong in the kernel.
David Lang
On Mon, Aug 18, 2008 at 10:17 AM, <[email protected]> wrote:
> On Sun, 17 Aug 2008, Casey Schaufler wrote:
>
>> Pavel Machek wrote:
>>>>>>
>>>>>> And I still don't get this 'mmap problem' that I don't solve that
>>>>>> libmalware magically solves. What? don't use mmap? I certainly hope
>>>>>> not.
>>>>>>
>>>>> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
>>>>> which is basically shared memory -- is fundamentally incompatible with
>>>>> reliable virus scanning.
>>>>>
>>>>> ...or do you have a reasonable solution for mmap?
>>>>>
>>>> mmap has a few different problems
>>>>
>>>> 1. intercepting reads and writes to take action at that time
>>>>
>>>> 2. the fact that two programs can use it as an inter-process
>>>> communication mechanism.
>>>>
>>>
>>> ...can and will use it as an IPC. So we need to modify some
>>> applications.
>>>
>>> Rather than modify all the applications using mmap (you can't tell if
>>> the other side is going to use it for shared memory... right?), we
>>> could simply modify all the Windows-facing applications using mmap.
>>>
>>>
>>>> if you are worried about the IPC aspects, all you can do is forbid it,
>>>
>>> Can you automatically tell if applications are using mmap for IPC?
>>>
>>> BTW in another mail you wanted to include /var/log/syslog from
>>> scanning. You should not be doing that if syslog is exported to
>>> Windows systems. Of course, you can get away with scanning syslog when
>>> Windows client tries to read it, which should be acceptable...
>>>
>>> Pavel
>>>
>>
>> There is a solution to this whole scanning thing, but I've been
>> reluctant to suggest it, and it will be pretty obvious why y'all
>> probably don't want to try it. Just to be sure the options are
>> out on the table, here goes.
>>
>> Define an xattr, let's call it "UNSCANNED", which has as its value
>> a timestamp. A regular file with this attribute cannot be executed
>> or opened,(exec or open hangs or fails, either behavior has merit
>> and downsides) and it requires privilege (perhaps CAP_MAC_ADMIN)
>> to remove the attribute. File creation attaches the attribute. Any
>> open for write attaches the attribute.
>>
>> Your scanner runs with privilege (say CAP_MAC_OVERRIDE) and passes
>> judgment on files with this attribute, removing either the file, if
>> it is Evil, or the attribute, if it is Good. The scanner is invoked
>> when a file that was open with write access is closed. This can be
>> done using mechanisms already discussed on this thread.
>>
>> If you like, you could use a "SCANNED" attribute instead of an
>> "UNSCANNED" attribute, and reverse the sense of the test. The
>> major difference will show up on filesystems that don't support
>> xattrs. The implications should be obvious.
>>
>> Now at the beginning I said that you wouldn't like this scheme,
>> and it shouldn't take a security expert to see the usability problems.
>> This is how an old school trusted systems junkie (like me) would do
>> it, and I don't see a better way that would actually achieve the
>> stated goals. If you wanted you could implement an LSM to do the
>> labeling bit in a day or two, the notification in about the same
>> time, which would leave the scanner as the long pole in your
>> development schedule.
>
> did you read the proposal I wrote up? it's similar to (but more flexible
> than) what you are saying
>
> it would allow for multiple 'scanned' tags (to allow for multiple scanning
> programs). the kernel would clear the tags when the file is dirtied (not
> when it is closed, the file may not be closed for weeks after all), there is
> a mechanism to tell the scanner(s) that a file was modified (rather then
> having to scan the entire filesystem), and if the scan was not done yet when
> a file is opened the scanner(s) can be invoked at that time.
>
>> P.S. - Library based security doesn't work.
>
> why not? (with the appropriate kernel support)
>
> it actually wouldn't be hard to have the kernel check the xattr, what would
> be hard is hving the kernel know when it should then invoke the scanner(s)
> and when it shouldn't. This is a policy decision that doesn't belong in the
> kernel.
>
Because of syscalls and you don't need to use LD.so to run a program.
Programs can embed there own LD.so and avoid using the system core
one. This kind of function is being developed for LSB applications
for platforms that don't support LSB.
syscalls you can run programs by passing all .so files and directly
talk to kernel.
Two aways around LD Preloads there are a few more.
Peter Dolding
On Mon, 18 Aug 2008, Peter Dolding wrote:
> On Mon, Aug 18, 2008 at 10:17 AM, <[email protected]> wrote:
>> On Sun, 17 Aug 2008, Casey Schaufler wrote:
>>
>>> Pavel Machek wrote:
>>>>>>>
>>>>>>> And I still don't get this 'mmap problem' that I don't solve that
>>>>>>> libmalware magically solves. What? don't use mmap? I certainly hope
>>>>>>> not.
>>>>>>>
>>>>>> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
>>>>>> which is basically shared memory -- is fundamentally incompatible with
>>>>>> reliable virus scanning.
>>>>>>
>>>>>> ...or do you have a reasonable solution for mmap?
>>>>>>
>>>>> mmap has a few different problems
>>>>>
>>>>> 1. intercepting reads and writes to take action at that time
>>>>>
>>>>> 2. the fact that two programs can use it as an inter-process
>>>>> communication mechanism.
>>>>>
>>>>
>>>> ...can and will use it as an IPC. So we need to modify some
>>>> applications.
>>>>
>>>> Rather than modify all the applications using mmap (you can't tell if
>>>> the other side is going to use it for shared memory... right?), we
>>>> could simply modify all the Windows-facing applications using mmap.
>>>>
>>>>
>>>>> if you are worried about the IPC aspects, all you can do is forbid it,
>>>>
>>>> Can you automatically tell if applications are using mmap for IPC?
>>>>
>>>> BTW in another mail you wanted to include /var/log/syslog from
>>>> scanning. You should not be doing that if syslog is exported to
>>>> Windows systems. Of course, you can get away with scanning syslog when
>>>> Windows client tries to read it, which should be acceptable...
>>>>
>>>> Pavel
>>>>
>>>
>>> There is a solution to this whole scanning thing, but I've been
>>> reluctant to suggest it, and it will be pretty obvious why y'all
>>> probably don't want to try it. Just to be sure the options are
>>> out on the table, here goes.
>>>
>>> Define an xattr, let's call it "UNSCANNED", which has as its value
>>> a timestamp. A regular file with this attribute cannot be executed
>>> or opened,(exec or open hangs or fails, either behavior has merit
>>> and downsides) and it requires privilege (perhaps CAP_MAC_ADMIN)
>>> to remove the attribute. File creation attaches the attribute. Any
>>> open for write attaches the attribute.
>>>
>>> Your scanner runs with privilege (say CAP_MAC_OVERRIDE) and passes
>>> judgment on files with this attribute, removing either the file, if
>>> it is Evil, or the attribute, if it is Good. The scanner is invoked
>>> when a file that was open with write access is closed. This can be
>>> done using mechanisms already discussed on this thread.
>>>
>>> If you like, you could use a "SCANNED" attribute instead of an
>>> "UNSCANNED" attribute, and reverse the sense of the test. The
>>> major difference will show up on filesystems that don't support
>>> xattrs. The implications should be obvious.
>>>
>>> Now at the beginning I said that you wouldn't like this scheme,
>>> and it shouldn't take a security expert to see the usability problems.
>>> This is how an old school trusted systems junkie (like me) would do
>>> it, and I don't see a better way that would actually achieve the
>>> stated goals. If you wanted you could implement an LSM to do the
>>> labeling bit in a day or two, the notification in about the same
>>> time, which would leave the scanner as the long pole in your
>>> development schedule.
>>
>> did you read the proposal I wrote up? it's similar to (but more flexible
>> than) what you are saying
>>
>> it would allow for multiple 'scanned' tags (to allow for multiple scanning
>> programs). the kernel would clear the tags when the file is dirtied (not
>> when it is closed, the file may not be closed for weeks after all), there is
>> a mechanism to tell the scanner(s) that a file was modified (rather then
>> having to scan the entire filesystem), and if the scan was not done yet when
>> a file is opened the scanner(s) can be invoked at that time.
>>
>>> P.S. - Library based security doesn't work.
>>
>> why not? (with the appropriate kernel support)
>>
>> it actually wouldn't be hard to have the kernel check the xattr, what would
>> be hard is hving the kernel know when it should then invoke the scanner(s)
>> and when it shouldn't. This is a policy decision that doesn't belong in the
>> kernel.
>>
> Because of syscalls and you don't need to use LD.so to run a program.
>
> Programs can embed there own LD.so and avoid using the system core
> one. This kind of function is being developed for LSB applications
> for platforms that don't support LSB.
>
> syscalls you can run programs by passing all .so files and directly
> talk to kernel.
>
> Two aways around LD Preloads there are a few more.
you seem to be disagreeing with the threat model. If so, rather then
arguing agains this implementation of a defense, you need to be arguing
against the threat model.
you are arguing that this solution doesn't work in some conditions. you
are correct, but the threat model that was proposed (and nobody has
proposed any different ones) is not trying to defend against code running
on the system, even if that code is trying to bypass the protection. I
even think that this is a good thing. it allows the distros to use the
protection where appropriate and disable it where it's not (which is a
surprisingly large number of places).
David Lang
[email protected] wrote:
> On Sun, 17 Aug 2008, Casey Schaufler wrote:
>
>> Pavel Machek wrote:
>>>>>> And I still don't get this 'mmap problem' that I don't solve that
>>>>>> libmalware magically solves. What? don't use mmap? I certainly
>>>>>> hope
>>>>>> not.
>>>>>>
>>>>> Don't use mmap, it is as simple as that. AFAICS mmap(MAP_SHARED) --
>>>>> which is basically shared memory -- is fundamentally incompatible
>>>>> with
>>>>> reliable virus scanning.
>>>>>
>>>>> ...or do you have a reasonable solution for mmap?
>>>>>
>>>> mmap has a few different problems
>>>>
>>>> 1. intercepting reads and writes to take action at that time
>>>>
>>>> 2. the fact that two programs can use it as an inter-process
>>>> communication mechanism.
>>>>
>>>
>>> ...can and will use it as an IPC. So we need to modify some
>>> applications.
>>>
>>> Rather than modify all the applications using mmap (you can't tell if
>>> the other side is going to use it for shared memory... right?), we
>>> could simply modify all the Windows-facing applications using mmap.
>>>
>>>
>>>> if you are worried about the IPC aspects, all you can do is forbid it,
>>>
>>> Can you automatically tell if applications are using mmap for IPC?
>>>
>>> BTW in another mail you wanted to include /var/log/syslog from
>>> scanning. You should not be doing that if syslog is exported to
>>> Windows systems. Of course, you can get away with scanning syslog when
>>> Windows client tries to read it, which should be acceptable...
>>> Pavel
>>>
>>
>> There is a solution to this whole scanning thing, but I've been
>> reluctant to suggest it, and it will be pretty obvious why y'all
>> probably don't want to try it. Just to be sure the options are
>> out on the table, here goes.
>>
>> Define an xattr, let's call it "UNSCANNED", which has as its value
>> a timestamp. A regular file with this attribute cannot be executed
>> or opened,(exec or open hangs or fails, either behavior has merit
>> and downsides) and it requires privilege (perhaps CAP_MAC_ADMIN)
>> to remove the attribute. File creation attaches the attribute. Any
>> open for write attaches the attribute.
>>
>> Your scanner runs with privilege (say CAP_MAC_OVERRIDE) and passes
>> judgment on files with this attribute, removing either the file, if
>> it is Evil, or the attribute, if it is Good. The scanner is invoked
>> when a file that was open with write access is closed. This can be
>> done using mechanisms already discussed on this thread.
>>
>> If you like, you could use a "SCANNED" attribute instead of an
>> "UNSCANNED" attribute, and reverse the sense of the test. The
>> major difference will show up on filesystems that don't support
>> xattrs. The implications should be obvious.
>>
>> Now at the beginning I said that you wouldn't like this scheme,
>> and it shouldn't take a security expert to see the usability problems.
>> This is how an old school trusted systems junkie (like me) would do
>> it, and I don't see a better way that would actually achieve the
>> stated goals. If you wanted you could implement an LSM to do the
>> labeling bit in a day or two, the notification in about the same
>> time, which would leave the scanner as the long pole in your
>> development schedule.
>
> did you read the proposal I wrote up? it's similar to (but more
> flexible than) what you are saying
Sorry, I must have missed it. My bad.
> it would allow for multiple 'scanned' tags (to allow for multiple
> scanning programs). the kernel would clear the tags when the file is
> dirtied (not when it is closed, the file may not be closed for weeks
> after all), there is a mechanism to tell the scanner(s) that a file
> was modified (rather then having to scan the entire filesystem), and
> if the scan was not done yet when a file is opened the scanner(s) can
> be invoked at that time.
>
One xattr, multiple xattrs, all would work. It's simpler to clear the
"scanned" attribute on open than on dirty, a your scheme means looking
at writes, seeks, ioctls, fcntls, and so on.
>> P.S. - Library based security doesn't work.
>
> why not? (with the appropriate kernel support)
If you have kernel support it aint library based.
> it actually wouldn't be hard to have the kernel check the xattr, what
> would be hard is hving the kernel know when it should then invoke the
> scanner(s) and when it shouldn't. This is a policy decision that
> doesn't belong in the kernel.
I'm old school. I say that all policy decisions regarding objects that
the kernel maintains belong in the kernel. That's why I say that any time
a file is closed with scanned unset (or unscanned set) the scanner should
be invoked or notified. The scanner can choose to ignore or queue the
notification, but the kernel will be done at that point. The kernel have
done all that it can and it's up to the scanner from then on.
On Mon, 2008-08-04 at 17:00 -0400, Eric Paris wrote:
> Background
> ++++++++++
> There is a consensus in the security industry that protecting against
> malicious files (viruses, root kits, spyware, ad-ware, ...) by the way
> of so-called on-access scanning is usable and reasonable approach.
> Currently the Linux kernel does not offer a completely suitable
> interface to implement such security solutions. Present solutions
A long time ago the FUSE developers said something about implementing
write-through stacking for FUSE (i.e. 'sudo fusermount -o allow_other
encfs / /' would allow mounting on /, and encfs could read/write under
its own mount point).
Wouldn't that make more sense? You could i.e. edit /etc/avfs.conf to
say "scanner=clamscan" "clamscan=/usr/bin/clamscan" and mount avfs
on /...?
On Sun 2008-08-17 20:07:39, Rik van Riel wrote:
> On Mon, 18 Aug 2008 00:58:44 +0200
> Pavel Machek <[email protected]> wrote:
>
> > Rather than modify all the applications using mmap (you can't tell if
> > the other side is going to use it for shared memory... right?), we
> > could simply modify all the Windows-facing applications using mmap.
>
> If web browsers, office suites and mail clients on Windows
> have certain kinds of vulnerabilities, it is safe to assume
> that the same programs on Linux will have similar problems.
>
> Can we please get rid of the idea that "Windows facing" is
> where the whole malware problem is?
>
> As for how to solve it - lets try to come up with a solution
> that is reasonably high performance and can be used for more
> than just malware scanning.
Don't mix exploits with viruses -- they are different.
Exploit is where application does something very unexpected due to a
bug.
Virus is where machine works correctly, but user does something
stupid.
For exploits, randomization + patching + compartments seem like a
solution. We should be working on "how to confine openoffice.org so
that it can't do much damage" instead of "how to detect .doc documents
that makes openoffice.org do something unexpected".
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html