Trond, with the latest code for issuing LAYOUTGET with the right credentials
we still seem to have a problem with the objects and blocks layout where
the security enforcement over out-of-band I/O differs from than the one
over in-band I/O.
Consider the following scenario:
file is owned by <uid1, gid1>, mode 660
process p1 successfully opens the file for RW with <uid1, gid1> (client sent OPEN)
process p2 successfully opens the file for RW with <uid2, gid1> (client sent ACCESS)
client gets a layout using LAYOUTGET for IOMODE_RW
the file is chmod'ed to 600
now, empirically, in-band I/O would succeed for p1 and fail for p2 (as seen on linux
and some commercial servers)
for out-of-band I/O, an object-based server will fence-off the object and recall the layout
to enforce the client to refresh its layout, send a LAYOUTGET, reauthorize, and get
a new capability. BUT, that's not enough as the new layout and capability, would allow both
p1 and p2 access to the object (as the layout is global to the client), yet we want only p1
to have access now.
How about sending ACCESS for any principal before using a newly retrieved layout
at OPEN time or after the layout was revoked/reacquired to simulate the in-band behavior in
a practical manner?
Note that I expect some inaccuracies in behavior even with sending ACCESS as
the linux nfs server and other commercial servers bypass permission checking for the file owner
at I/O time but not for ACCESS. I believe this was done to simulate (sort of) Posix behavior
that allows I/O to an open file even after changing its security attributes.
Also, do we deal correctly with LAYOUTGET failing on NFS4ERR_ACCESS?
In the example above, if the open order was reversed, LAYOUTGET would have failed for p2's
creds as it doesn't have RW access to the file. That would result to reverting to the MDS
and the I/O would fail on NFS4ERR_ACCESS as well, yet we'll keep trying (and failing)
LAYOUTGET. Optionally, the client could try other creds that opened the file.
If the first process to open the file closes it, should we use different creds for LAYOUTGET?
With the latest implementation we keep the first opener creds referenced until we return the
whole layout, right?
Regards,
Benny
On Tue, 2013-05-28 at 15:50 +0300, Benny Halevy wrote:
> Trond, with the latest code for issuing LAYOUTGET with the right credentials
> we still seem to have a problem with the objects and blocks layout where
> the security enforcement over out-of-band I/O differs from than the one
> over in-band I/O.
>
> Consider the following scenario:
>
> file is owned by <uid1, gid1>, mode 660
> process p1 successfully opens the file for RW with <uid1, gid1> (client sent OPEN)
> process p2 successfully opens the file for RW with <uid2, gid1> (client sent ACCESS)
> client gets a layout using LAYOUTGET for IOMODE_RW
> the file is chmod'ed to 600
>
> now, empirically, in-band I/O would succeed for p1 and fail for p2 (as seen on linux
> and some commercial servers)
>
> for out-of-band I/O, an object-based server will fence-off the object and recall the layout
> to enforce the client to refresh its layout, send a LAYOUTGET, reauthorize, and get
> a new capability. BUT, that's not enough as the new layout and capability, would allow both
> p1 and p2 access to the object (as the layout is global to the client), yet we want only p1
> to have access now.
I don't understand why you think this is related to the LAYOUTGET
credential change. The only difference that the credential change brings
is to the case where the client doesn't hold a layout segment prior to
initiating the read/writeback.
IOW: If p1 had already grabbed a layout segment covering the area being
accessed by p2, then under the old code, we would still have forged
ahead and performed the read/write on the DS without calling LAYOUTGET
at all.
> How about sending ACCESS for any principal before using a newly retrieved layout
> at OPEN time or after the layout was revoked/reacquired to simulate the in-band behavior in
> a practical manner?
If you want to do that for the objects and blocks layout types, then
fine, but I see no reason to do it for files layouts: the files DSes
will do access checking using the cred passed with the READ/WRITE
regardless of what happened with LAYOUTGET.
> Note that I expect some inaccuracies in behavior even with sending ACCESS as
> the linux nfs server and other commercial servers bypass permission checking for the file owner
> at I/O time but not for ACCESS. I believe this was done to simulate (sort of) Posix behavior
> that allows I/O to an open file even after changing its security attributes.
>
> Also, do we deal correctly with LAYOUTGET failing on NFS4ERR_ACCESS?
> In the example above, if the open order was reversed, LAYOUTGET would have failed for p2's
> creds as it doesn't have RW access to the file. That would result to reverting to the MDS
> and the I/O would fail on NFS4ERR_ACCESS as well, yet we'll keep trying (and failing)
> LAYOUTGET. Optionally, the client could try other creds that opened the file.
> If the first process to open the file closes it, should we use different creds for LAYOUTGET?
> With the latest implementation we keep the first opener creds referenced until we return the
> whole layout, right?
I'm open to the idea of having an NFS4ERR_ACCESS reply to LAYOUTGET fail
the entire I/O attempt without an attempt to fail back to MDS.
As for switching creds on close, I believe that is still forbidden under
the rules guiding the EXCHGID4_FLAG_BIND_PRINC_STATEID flag (RFC5661,
section 18.35.3). Under those rules, a server that sets the
EXCHGID4_FLAG_BIND_PRINC_STATEID in the EXCHANGE_ID reply MUST return
NFS4ERR_WRONG_CRED in response to the LAYOUTGET call if it tries to
authenticate a layout or open stateid that was created by p1, using the
principal of p2.
--
Trond Myklebust
Linux NFS client maintainer
NetApp
[email protected]
http://www.netapp.com
On 2013-05-28 18:08, Myklebust, Trond wrote:
> On Tue, 2013-05-28 at 15:50 +0300, Benny Halevy wrote:
>> Trond, with the latest code for issuing LAYOUTGET with the right credentials
>> we still seem to have a problem with the objects and blocks layout where
>> the security enforcement over out-of-band I/O differs from than the one
>> over in-band I/O.
>>
>> Consider the following scenario:
>>
>> file is owned by <uid1, gid1>, mode 660
>> process p1 successfully opens the file for RW with <uid1, gid1> (client sent OPEN)
>> process p2 successfully opens the file for RW with <uid2, gid1> (client sent ACCESS)
>> client gets a layout using LAYOUTGET for IOMODE_RW
>> the file is chmod'ed to 600
>>
>> now, empirically, in-band I/O would succeed for p1 and fail for p2 (as seen on linux
>> and some commercial servers)
>>
>> for out-of-band I/O, an object-based server will fence-off the object and recall the layout
>> to enforce the client to refresh its layout, send a LAYOUTGET, reauthorize, and get
>> a new capability. BUT, that's not enough as the new layout and capability, would allow both
>> p1 and p2 access to the object (as the layout is global to the client), yet we want only p1
>> to have access now.
>
> I don't understand why you think this is related to the LAYOUTGET
> credential change. The only difference that the credential change brings
> is to the case where the client doesn't hold a layout segment prior to
> initiating the read/writeback.
Correct, this is not caused by that change, just that testing the change
raised this issue.
>
> IOW: If p1 had already grabbed a layout segment covering the area being
> accessed by p2, then under the old code, we would still have forged
> ahead and performed the read/write on the DS without calling LAYOUTGET
> at all.
>
>> How about sending ACCESS for any principal before using a newly retrieved layout
>> at OPEN time or after the layout was revoked/reacquired to simulate the in-band behavior in
>> a practical manner?
>
> If you want to do that for the objects and blocks layout types, then
> fine, but I see no reason to do it for files layouts: the files DSes
> will do access checking using the cred passed with the READ/WRITE
> regardless of what happened with LAYOUTGET.
>
True.
>> Note that I expect some inaccuracies in behavior even with sending ACCESS as
>> the linux nfs server and other commercial servers bypass permission checking for the file owner
>> at I/O time but not for ACCESS. I believe this was done to simulate (sort of) Posix behavior
>> that allows I/O to an open file even after changing its security attributes.
>>
>> Also, do we deal correctly with LAYOUTGET failing on NFS4ERR_ACCESS?
>> In the example above, if the open order was reversed, LAYOUTGET would have failed for p2's
>> creds as it doesn't have RW access to the file. That would result to reverting to the MDS
>> and the I/O would fail on NFS4ERR_ACCESS as well, yet we'll keep trying (and failing)
>> LAYOUTGET. Optionally, the client could try other creds that opened the file.
>> If the first process to open the file closes it, should we use different creds for LAYOUTGET?
>> With the latest implementation we keep the first opener creds referenced until we return the
>> whole layout, right?
>
> I'm open to the idea of having an NFS4ERR_ACCESS reply to LAYOUTGET fail
> the entire I/O attempt without an attempt to fail back to MDS.
>
> As for switching creds on close, I believe that is still forbidden under
> the rules guiding the EXCHGID4_FLAG_BIND_PRINC_STATEID flag (RFC5661,
> section 18.35.3). Under those rules, a server that sets the
> EXCHGID4_FLAG_BIND_PRINC_STATEID in the EXCHANGE_ID reply MUST return
> NFS4ERR_WRONG_CRED in response to the LAYOUTGET call if it tries to
> authenticate a layout or open stateid that was created by p1, using the
> principal of p2.
>
>
Yeah, that's a bummer.
Even if the client returns the whole layout and tried to retrieve a new layout
using the open stateid, it still bound to the original open creds.
Benny