This patch extends functionality of e{2|4}fsprogs to add
EXT2_FLAG_DIRECT flag to be passed to ext2fs_open2() function. This
internally calls open() function with O_DIRECT and handles the memory
alignment for both read and write operations.
In some cases direct access to devices is necessary and that was the
main reason for this patch to be done.
The main reason why this was done is that pygrub (used by xen
virtualization user-space package, it's a python version of grub for
paravirtualized guests) sometimes uses outdated version of grub.conf
file. Modifications to xen package were *not* enough because e2fsprogs
doesn't open the files directly. That's why I added EXT2_FLAG_DIRECT
support to make read/write operations work directly when passed. It's
been tested with pygrub like mentioned above for read operation and it's
working fine.
Signed-off-by: Michal Novotny <[email protected]>
On 01/08/2010 04:36 AM, Michal Novotny wrote:
> This patch extends functionality of e{2|4}fsprogs to add
> EXT2_FLAG_DIRECT flag to be passed to ext2fs_open2() function. This
> internally calls open() function with O_DIRECT and handles the memory
> alignment for both read and write operations.
> In some cases direct access to devices is necessary and that was the
> main reason for this patch to be done.
>
> The main reason why this was done is that pygrub (used by xen
> virtualization user-space package, it's a python version of grub for
> paravirtualized guests) sometimes uses outdated version of grub.conf
> file. Modifications to xen package were *not* enough because e2fsprogs
> doesn't open the files directly. That's why I added EXT2_FLAG_DIRECT
> support to make read/write operations work directly when passed. It's
> been tested with pygrub like mentioned above for read operation and
> it's working fine.
>
> Signed-off-by: Michal Novotny <[email protected]>
Can you add to this some kind of data flow overview? Seems like a
really odd way to update this file system...
ric
On 01/11/2010 09:06 PM, Ric Wheeler wrote:
> On 01/08/2010 04:36 AM, Michal Novotny wrote:
>> This patch extends functionality of e{2|4}fsprogs to add
>> EXT2_FLAG_DIRECT flag to be passed to ext2fs_open2() function. This
>> internally calls open() function with O_DIRECT and handles the memory
>> alignment for both read and write operations.
>> In some cases direct access to devices is necessary and that was the
>> main reason for this patch to be done.
>>
>> The main reason why this was done is that pygrub (used by xen
>> virtualization user-space package, it's a python version of grub for
>> paravirtualized guests) sometimes uses outdated version of grub.conf
>> file. Modifications to xen package were *not* enough because
>> e2fsprogs doesn't open the files directly. That's why I added
>> EXT2_FLAG_DIRECT support to make read/write operations work directly
>> when passed. It's been tested with pygrub like mentioned above for
>> read operation and it's working fine.
>>
>> Signed-off-by: Michal Novotny <[email protected]>
>
> Can you add to this some kind of data flow overview? Seems like a
> really odd way to update this file system...
>
> ric
>
What do you mean by adding some kind of data flow overview? I am not
working on file systems but I needed option open files directly to be
added to e2fsprogs to make pygrub working right with not using
cached/outdated data so this was the first patch I did for file system
so I don't realy know what do you mean by data flow overview ...
Thanks,
Michal
On 01/12/2010 05:54 AM, Michal Novotny wrote:
> On 01/11/2010 09:06 PM, Ric Wheeler wrote:
>> On 01/08/2010 04:36 AM, Michal Novotny wrote:
>>> This patch extends functionality of e{2|4}fsprogs to add
>>> EXT2_FLAG_DIRECT flag to be passed to ext2fs_open2() function. This
>>> internally calls open() function with O_DIRECT and handles the memory
>>> alignment for both read and write operations.
>>> In some cases direct access to devices is necessary and that was the
>>> main reason for this patch to be done.
>>>
>>> The main reason why this was done is that pygrub (used by xen
>>> virtualization user-space package, it's a python version of grub for
>>> paravirtualized guests) sometimes uses outdated version of grub.conf
>>> file. Modifications to xen package were *not* enough because
>>> e2fsprogs doesn't open the files directly. That's why I added
>>> EXT2_FLAG_DIRECT support to make read/write operations work directly
>>> when passed. It's been tested with pygrub like mentioned above for
>>> read operation and it's working fine.
>>>
>>> Signed-off-by: Michal Novotny <[email protected]>
>>
>> Can you add to this some kind of data flow overview? Seems like a
>> really odd way to update this file system...
>>
>> ric
>>
> What do you mean by adding some kind of data flow overview? I am not
> working on file systems but I needed option open files directly to be
> added to e2fsprogs to make pygrub working right with not using
> cached/outdated data so this was the first patch I did for file system
> so I don't realy know what do you mean by data flow overview ...
>
> Thanks,
> Michal
Hi Michael,
What would be useful is where the data is written (guest/host) and the state of
the file system (mounted/unmounted). Which app it is that writes the data that
gets cached, who needs to read it and misses that cached data.
From the above, it is really not clear that you should not be able to use
existing interfaces to flush any cached data to disk....
ric
On 01/12/2010 12:59 PM, Ric Wheeler wrote:
> On 01/12/2010 05:54 AM, Michal Novotny wrote:
>> On 01/11/2010 09:06 PM, Ric Wheeler wrote:
>>> On 01/08/2010 04:36 AM, Michal Novotny wrote:
>>>> This patch extends functionality of e{2|4}fsprogs to add
>>>> EXT2_FLAG_DIRECT flag to be passed to ext2fs_open2() function. This
>>>> internally calls open() function with O_DIRECT and handles the memory
>>>> alignment for both read and write operations.
>>>> In some cases direct access to devices is necessary and that was the
>>>> main reason for this patch to be done.
>>>>
>>>> The main reason why this was done is that pygrub (used by xen
>>>> virtualization user-space package, it's a python version of grub for
>>>> paravirtualized guests) sometimes uses outdated version of grub.conf
>>>> file. Modifications to xen package were *not* enough because
>>>> e2fsprogs doesn't open the files directly. That's why I added
>>>> EXT2_FLAG_DIRECT support to make read/write operations work directly
>>>> when passed. It's been tested with pygrub like mentioned above for
>>>> read operation and it's working fine.
>>>>
>>>> Signed-off-by: Michal Novotny <[email protected]>
>>>
>>> Can you add to this some kind of data flow overview? Seems like a
>>> really odd way to update this file system...
>>>
>>> ric
>>>
>> What do you mean by adding some kind of data flow overview? I am not
>> working on file systems but I needed option open files directly to be
>> added to e2fsprogs to make pygrub working right with not using
>> cached/outdated data so this was the first patch I did for file system
>> so I don't realy know what do you mean by data flow overview ...
>>
>> Thanks,
>> Michal
>
> Hi Michael,
>
> What would be useful is where the data is written (guest/host) and the
> state of the file system (mounted/unmounted). Which app it is that
> writes the data that gets cached, who needs to read it and misses that
> cached data.
>
> From the above, it is really not clear that you should not be able to
> use existing interfaces to flush any cached data to disk....
>
> ric
>
Hi Ric,
I don't really know if I see your point but the thing here is that there
was no way to open a file directly (ie. using O_DIRECT). The direct
write support has been added only to make it possible to use both read
and write directly. The main reason to create this patch was to add
direct read support and flush capability won't help me at all. I am
working in Red Hat, Virtualization team on Xen so I am really not that
much familiar with file systems but what I needed was an option to read
the data directly (using O_DIRECT) in e2fsprogs. One bug was about
pygrub (Python version of GRUB of Xen PV guests that is internally using
e2fsprogs functionality to access data on ext2/3/4 partition to boot the
PV guests) uses outdated/cached data so some modifications were
necessary to open everything directly... It was done to Xen by a
collegue but there was some open() call with no O_DIRECT set so I
investigated this further and it turned out that this call was coming
from e2fsprogs library. That was the reason why I needed to add O_DIRECT
read support and O_DIRECT write support was added just to make it for
both read and write. For write, yes, there may be the option to flush
the disk but not for read. To workaround this but dropping caches is a
really ugly idea so finally patching e2fsprogs too (along with Xen)
turned out to the best idea. Once again, this was done to support
O_DIRECT for read function only - it is *not* used for write function
and it was added just to make it complete for both read and write
functions. Regarding the functionality of read function, this is done in
the host machine before the PV guest boots. It reads grub.conf, then it
reads initrd and kernel and boots from it... Some incidents of booting
outdated grub.conf/initrd and kernel has been reported to us and this
patch is to make it don't use caches so outdated data can't be used...
Is the scenario clear or do you need some other information about that?
Thanks,
Michal
On Tue, Jan 12, 2010 at 01:15:49PM +0100, Michal Novotny wrote:
> I don't really know if I see your point but the thing here is that there
> was no way to open a file directly (ie. using O_DIRECT). The direct
> write support has been added only to make it possible to use both read
> and write directly. The main reason to create this patch was to add
> direct read support and flush capability won't help me at all. I am
> working in Red Hat, Virtualization team on Xen so I am really not that
> much familiar with file systems but what I needed was an option to read
> the data directly (using O_DIRECT) in e2fsprogs. One bug was about
> pygrub (Python version of GRUB of Xen PV guests that is internally using
> e2fsprogs functionality to access data on ext2/3/4 partition to boot the
> PV guests) uses outdated/cached data so some modifications were
> necessary to open everything directly...
So to get things staigt: you're using e2fsprogs to manipulate a life
filesystem and thing using O_DIRECT saves your ass? I think you need to
rething your model of operation fundamentally in that case.
On 01/12/2010 01:23 PM, Christoph Hellwig wrote:
> On Tue, Jan 12, 2010 at 01:15:49PM +0100, Michal Novotny wrote:
>
>> I don't really know if I see your point but the thing here is that there
>> was no way to open a file directly (ie. using O_DIRECT). The direct
>> write support has been added only to make it possible to use both read
>> and write directly. The main reason to create this patch was to add
>> direct read support and flush capability won't help me at all. I am
>> working in Red Hat, Virtualization team on Xen so I am really not that
>> much familiar with file systems but what I needed was an option to read
>> the data directly (using O_DIRECT) in e2fsprogs. One bug was about
>> pygrub (Python version of GRUB of Xen PV guests that is internally using
>> e2fsprogs functionality to access data on ext2/3/4 partition to boot the
>> PV guests) uses outdated/cached data so some modifications were
>> necessary to open everything directly...
>>
> So to get things staigt: you're using e2fsprogs to manipulate a life
> filesystem and thing using O_DIRECT saves your ass? I think you need to
> rething your model of operation fundamentally in that case.
>
>
Not really, pygrub doesn't do any manipulation with file system and
also, it's not working on a life file system. It's called before the
guest boots up to read information about grub.conf/initrd and kernel for
PV guest and after this is read and selected in pygrub then the guest is
booted using the kernel and initrd extracted from the image (after which
the file is closed). Once again, nothing uses write support and it was
added just to make it use O_DIRECT for both read and write operations
but only pygrub uses only read support and O_DIRECT passed here is the
only way to make it use non-cached data.
On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
> Not really, pygrub doesn't do any manipulation with file system and
> also, it's not working on a life file system. It's called before the
> guest boots up to read information about grub.conf/initrd and kernel for
> PV guest and after this is read and selected in pygrub then the guest is
> booted using the kernel and initrd extracted from the image (after which
> the file is closed). Once again, nothing uses write support and it was
> added just to make it use O_DIRECT for both read and write operations
> but only pygrub uses only read support and O_DIRECT passed here is the
> only way to make it use non-cached data.
So what caches get in the way? From the above it seems the situation
is the following:
- filesystem N is a guest filesystem. It's not usually mounted on the
host, except for initial setup long time ago
- before booting a guest your "pygrub" tools needs to read files on
it, and it's doing so using e2fsprogs
- once the guest is life it uses the extN kernel driver to access the
filesystem
nowhere in this cycle you should have any stale cached data. The kernel
always makes sure to write back data on umount/reboot, as does e2fsprogs
if actually used to write data (which you said is not the case anyway).
The only data that may be in the cache are unmodified data from reads
on the block device from either e2fsprogs or a suboptimal virtual block
device implementation, but these can't cause any problems.
On 2010-01-12, at 08:30, Michal Novotny wrote:
> On 01/12/2010 01:23 PM, Christoph Hellwig wrote:
>> So to get things staigt: you're using e2fsprogs to manipulate a life
>> filesystem and thing using O_DIRECT saves your ass? I think you
>> need to
>> rething your model of operation fundamentally in that case.
>>
>
> Not really, pygrub doesn't do any manipulation with file system and
> also, it's not working on a life file system. It's called before the
> guest boots up to read information about grub.conf/initrd and kernel
> for PV guest and after this is read and selected in pygrub then the
> guest is booted using the kernel and initrd extracted from the image
> (after which the file is closed). Once again, nothing uses write
> support and it was added just to make it use O_DIRECT for both read
> and write operations but only pygrub uses only read support and
> O_DIRECT passed here is the only way to make it use non-cached data.
Michal, I think the thing that is confusing everyone is that if you
are not accessing a live filesystem, and you are not doing the writes
yourself, then why is it bad to read cached data? How is it that the
cached data becomes stale if pygrub isn't modifying it, and there is
nothing else mounting the filesystem?
Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.
On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>
>> Not really, pygrub doesn't do any manipulation with file system and
>> also, it's not working on a life file system. It's called before the
>> guest boots up to read information about grub.conf/initrd and kernel for
>> PV guest and after this is read and selected in pygrub then the guest is
>> booted using the kernel and initrd extracted from the image (after which
>> the file is closed). Once again, nothing uses write support and it was
>> added just to make it use O_DIRECT for both read and write operations
>> but only pygrub uses only read support and O_DIRECT passed here is the
>> only way to make it use non-cached data.
>>
> So what caches get in the way? From the above it seems the situation
> is the following:
>
> - filesystem N is a guest filesystem. It's not usually mounted on the
> host, except for initial setup long time ago
>
Yes, it is really a guest file system. This is not mounted in the host
and the reason is to get actual version of grub.conf, initrd and kernel
to be booted...
> - before booting a guest your "pygrub" tools needs to read files on
> it, and it's doing so using e2fsprogs
>
Correct.
> - once the guest is life it uses the extN kernel driver to access the
> filesystem
>
That's right. So this is no longer pygrub responsibility...
> nowhere in this cycle you should have any stale cached data. The kernel
> always makes sure to write back data on umount/reboot, as does e2fsprogs
> if actually used to write data (which you said is not the case anyway).
>
In fact I was unable to run into those problems myself but
reporter/customer did.
> The only data that may be in the cache are unmodified data from reads
> on the block device from either e2fsprogs or a suboptimal virtual block
> device implementation, but these can't cause any problems.
>
Michal
On 01/12/2010 01:47 PM, Andreas Dilger wrote:
> On 2010-01-12, at 08:30, Michal Novotny wrote:
>> On 01/12/2010 01:23 PM, Christoph Hellwig wrote:
>>> So to get things staigt: you're using e2fsprogs to manipulate a life
>>> filesystem and thing using O_DIRECT saves your ass? I think you
>>> need to
>>> rething your model of operation fundamentally in that case.
>>>
>>
>> Not really, pygrub doesn't do any manipulation with file system and
>> also, it's not working on a life file system. It's called before the
>> guest boots up to read information about grub.conf/initrd and kernel
>> for PV guest and after this is read and selected in pygrub then the
>> guest is booted using the kernel and initrd extracted from the image
>> (after which the file is closed). Once again, nothing uses write
>> support and it was added just to make it use O_DIRECT for both read
>> and write operations but only pygrub uses only read support and
>> O_DIRECT passed here is the only way to make it use non-cached data.
>
>
> Michal, I think the thing that is confusing everyone is that if you
> are not accessing a live filesystem, and you are not doing the writes
> yourself, then why is it bad to read cached data? How is it that the
> cached data becomes stale if pygrub isn't modifying it, and there is
> nothing else mounting the filesystem?
>
Hi Andreas,
it really is bad because the host have data about grub.conf, kernel and
initrd cached but if you run the guest, update kernel or grub.conf etc.,
you need to see the changes here to boot the new/updated kernel now. Not
the old, cached one. This is why it is bad to read cached data. Pygrub
isn't modifying it but the guest can modify the data itself...
Thanks,
Michal
On 01/12/2010 08:01 AM, Michal Novotny wrote:
> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>> Not really, pygrub doesn't do any manipulation with file system and
>>> also, it's not working on a life file system. It's called before the
>>> guest boots up to read information about grub.conf/initrd and kernel for
>>> PV guest and after this is read and selected in pygrub then the guest is
>>> booted using the kernel and initrd extracted from the image (after which
>>> the file is closed). Once again, nothing uses write support and it was
>>> added just to make it use O_DIRECT for both read and write operations
>>> but only pygrub uses only read support and O_DIRECT passed here is the
>>> only way to make it use non-cached data.
>> So what caches get in the way? From the above it seems the situation
>> is the following:
>>
>> - filesystem N is a guest filesystem. It's not usually mounted on the
>> host, except for initial setup long time ago
>
> Yes, it is really a guest file system. This is not mounted in the host
> and the reason is to get actual version of grub.conf, initrd and kernel
> to be booted...
>
>> - before booting a guest your "pygrub" tools needs to read files on
>> it, and it's doing so using e2fsprogs
>
> Correct.
>
>> - once the guest is life it uses the extN kernel driver to access the
>> filesystem
>
> That's right. So this is no longer pygrub responsibility...
>
>> nowhere in this cycle you should have any stale cached data. The kernel
>> always makes sure to write back data on umount/reboot, as does e2fsprogs
>> if actually used to write data (which you said is not the case anyway).
>
> In fact I was unable to run into those problems myself but
> reporter/customer did.
>
>> The only data that may be in the cache are unmodified data from reads
>> on the block device from either e2fsprogs or a suboptimal virtual block
>> device implementation, but these can't cause any problems.
> Michal
If the guest is the only one (when running) that installs a new grub.conf file
and kernel and it shuts down properly, you should be good. It if does not shut
down cleanly, it could have a stale grub.conf file (or worse, a partially
written one), but using O_DIRECT to bypass the file system cache should not help.
If we cannot reproduce this failure, sounds like we need to go back and get a
better understanding of what the customer saw?
ric
On 01/12/2010 02:04 PM, Ric Wheeler wrote:
> On 01/12/2010 08:01 AM, Michal Novotny wrote:
>> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>>> Not really, pygrub doesn't do any manipulation with file system and
>>>> also, it's not working on a life file system. It's called before the
>>>> guest boots up to read information about grub.conf/initrd and
>>>> kernel for
>>>> PV guest and after this is read and selected in pygrub then the
>>>> guest is
>>>> booted using the kernel and initrd extracted from the image (after
>>>> which
>>>> the file is closed). Once again, nothing uses write support and it was
>>>> added just to make it use O_DIRECT for both read and write operations
>>>> but only pygrub uses only read support and O_DIRECT passed here is the
>>>> only way to make it use non-cached data.
>>> So what caches get in the way? From the above it seems the situation
>>> is the following:
>>>
>>> - filesystem N is a guest filesystem. It's not usually mounted on the
>>> host, except for initial setup long time ago
>>
>> Yes, it is really a guest file system. This is not mounted in the host
>> and the reason is to get actual version of grub.conf, initrd and kernel
>> to be booted...
>>
>>> - before booting a guest your "pygrub" tools needs to read files on
>>> it, and it's doing so using e2fsprogs
>>
>> Correct.
>>
>>> - once the guest is life it uses the extN kernel driver to access the
>>> filesystem
>>
>> That's right. So this is no longer pygrub responsibility...
>>
>>> nowhere in this cycle you should have any stale cached data. The kernel
>>> always makes sure to write back data on umount/reboot, as does
>>> e2fsprogs
>>> if actually used to write data (which you said is not the case anyway).
>>
>> In fact I was unable to run into those problems myself but
>> reporter/customer did.
>>
>>> The only data that may be in the cache are unmodified data from reads
>>> on the block device from either e2fsprogs or a suboptimal virtual block
>>> device implementation, but these can't cause any problems.
>> Michal
>
> If the guest is the only one (when running) that installs a new
> grub.conf file and kernel and it shuts down properly, you should be
> good. It if does not shut down cleanly, it could have a stale
> grub.conf file (or worse, a partially written one), but using O_DIRECT
> to bypass the file system cache should not help.
>
> If we cannot reproduce this failure, sounds like we need to go back
> and get a better understanding of what the customer saw?
>
> ric
>
That's right. I am going write an e-mail regarding this information to
the reproducer if this bug and tell him that I need more information
about what's happening at the customer side.
Michal
On 01/12/2010 02:12 PM, Michal Novotny wrote:
> On 01/12/2010 02:04 PM, Ric Wheeler wrote:
>> On 01/12/2010 08:01 AM, Michal Novotny wrote:
>>> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>>>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>>>> Not really, pygrub doesn't do any manipulation with file system and
>>>>> also, it's not working on a life file system. It's called before the
>>>>> guest boots up to read information about grub.conf/initrd and
>>>>> kernel for
>>>>> PV guest and after this is read and selected in pygrub then the
>>>>> guest is
>>>>> booted using the kernel and initrd extracted from the image (after
>>>>> which
>>>>> the file is closed). Once again, nothing uses write support and it
>>>>> was
>>>>> added just to make it use O_DIRECT for both read and write operations
>>>>> but only pygrub uses only read support and O_DIRECT passed here is
>>>>> the
>>>>> only way to make it use non-cached data.
>>>> So what caches get in the way? From the above it seems the situation
>>>> is the following:
>>>>
>>>> - filesystem N is a guest filesystem. It's not usually mounted on the
>>>> host, except for initial setup long time ago
>>>
>>> Yes, it is really a guest file system. This is not mounted in the host
>>> and the reason is to get actual version of grub.conf, initrd and kernel
>>> to be booted...
>>>
>>>> - before booting a guest your "pygrub" tools needs to read files on
>>>> it, and it's doing so using e2fsprogs
>>>
>>> Correct.
>>>
>>>> - once the guest is life it uses the extN kernel driver to access the
>>>> filesystem
>>>
>>> That's right. So this is no longer pygrub responsibility...
>>>
>>>> nowhere in this cycle you should have any stale cached data. The
>>>> kernel
>>>> always makes sure to write back data on umount/reboot, as does
>>>> e2fsprogs
>>>> if actually used to write data (which you said is not the case
>>>> anyway).
>>>
>>> In fact I was unable to run into those problems myself but
>>> reporter/customer did.
>>>
>>>> The only data that may be in the cache are unmodified data from reads
>>>> on the block device from either e2fsprogs or a suboptimal virtual
>>>> block
>>>> device implementation, but these can't cause any problems.
>>> Michal
>>
>> If the guest is the only one (when running) that installs a new
>> grub.conf file and kernel and it shuts down properly, you should be
>> good. It if does not shut down cleanly, it could have a stale
>> grub.conf file (or worse, a partially written one), but using
>> O_DIRECT to bypass the file system cache should not help.
>>
>> If we cannot reproduce this failure, sounds like we need to go back
>> and get a better understanding of what the customer saw?
>>
>> ric
>>
> That's right. I am going write an e-mail regarding this information to
> the reproducer if this bug and tell him that I need more information
> about what's happening at the customer side.
>
One more thing to point out, let's have a look at:
https://bugzilla.redhat.com/show_bug.cgi?id=466681#c15 .This is about
workaround to drop caches to be added to pygrub in the host machine
using this command:
echo 1> /proc/sys/vm/drop_caches
So this really looks like the caching issue if it's working fine after
dropping the caches. That may be the reason why this could be fine with
this patch present in e2fsprogs.
Michal
On 01/12/2010 08:23 AM, Michal Novotny wrote:
> On 01/12/2010 02:12 PM, Michal Novotny wrote:
>> On 01/12/2010 02:04 PM, Ric Wheeler wrote:
>>> On 01/12/2010 08:01 AM, Michal Novotny wrote:
>>>> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>>>>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>>>>> Not really, pygrub doesn't do any manipulation with file system and
>>>>>> also, it's not working on a life file system. It's called before the
>>>>>> guest boots up to read information about grub.conf/initrd and
>>>>>> kernel for
>>>>>> PV guest and after this is read and selected in pygrub then the
>>>>>> guest is
>>>>>> booted using the kernel and initrd extracted from the image (after
>>>>>> which
>>>>>> the file is closed). Once again, nothing uses write support and it
>>>>>> was
>>>>>> added just to make it use O_DIRECT for both read and write operations
>>>>>> but only pygrub uses only read support and O_DIRECT passed here is
>>>>>> the
>>>>>> only way to make it use non-cached data.
>>>>> So what caches get in the way? From the above it seems the situation
>>>>> is the following:
>>>>>
>>>>> - filesystem N is a guest filesystem. It's not usually mounted on the
>>>>> host, except for initial setup long time ago
>>>>
>>>> Yes, it is really a guest file system. This is not mounted in the host
>>>> and the reason is to get actual version of grub.conf, initrd and kernel
>>>> to be booted...
>>>>
>>>>> - before booting a guest your "pygrub" tools needs to read files on
>>>>> it, and it's doing so using e2fsprogs
>>>>
>>>> Correct.
>>>>
>>>>> - once the guest is life it uses the extN kernel driver to access the
>>>>> filesystem
>>>>
>>>> That's right. So this is no longer pygrub responsibility...
>>>>
>>>>> nowhere in this cycle you should have any stale cached data. The
>>>>> kernel
>>>>> always makes sure to write back data on umount/reboot, as does
>>>>> e2fsprogs
>>>>> if actually used to write data (which you said is not the case
>>>>> anyway).
>>>>
>>>> In fact I was unable to run into those problems myself but
>>>> reporter/customer did.
>>>>
>>>>> The only data that may be in the cache are unmodified data from reads
>>>>> on the block device from either e2fsprogs or a suboptimal virtual
>>>>> block
>>>>> device implementation, but these can't cause any problems.
>>>> Michal
>>>
>>> If the guest is the only one (when running) that installs a new
>>> grub.conf file and kernel and it shuts down properly, you should be
>>> good. It if does not shut down cleanly, it could have a stale
>>> grub.conf file (or worse, a partially written one), but using
>>> O_DIRECT to bypass the file system cache should not help.
>>>
>>> If we cannot reproduce this failure, sounds like we need to go back
>>> and get a better understanding of what the customer saw?
>>>
>>> ric
>>>
>> That's right. I am going write an e-mail regarding this information to
>> the reproducer if this bug and tell him that I need more information
>> about what's happening at the customer side.
>>
> One more thing to point out, let's have a look at:
> https://bugzilla.redhat.com/show_bug.cgi?id=466681#c15 .This is about
> workaround to drop caches to be added to pygrub in the host machine
> using this command:
>
> echo 1> /proc/sys/vm/drop_caches
>
> So this really looks like the caching issue if it's working fine after
> dropping the caches. That may be the reason why this could be fine with
> this patch present in e2fsprogs.
>
> Michal
That BZ has a pretty long and twisted history, but after a quick read, I still
don't see why a cleanly shutdown guest would have issues with caching that using
O_DIRECT on read would help.
We will need to dig into a bit more...
ric
On 01/12/2010 02:29 PM, Ric Wheeler wrote:
> On 01/12/2010 08:23 AM, Michal Novotny wrote:
>> On 01/12/2010 02:12 PM, Michal Novotny wrote:
>>> On 01/12/2010 02:04 PM, Ric Wheeler wrote:
>>>> On 01/12/2010 08:01 AM, Michal Novotny wrote:
>>>>> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>>>>>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>>>>>> Not really, pygrub doesn't do any manipulation with file system and
>>>>>>> also, it's not working on a life file system. It's called before
>>>>>>> the
>>>>>>> guest boots up to read information about grub.conf/initrd and
>>>>>>> kernel for
>>>>>>> PV guest and after this is read and selected in pygrub then the
>>>>>>> guest is
>>>>>>> booted using the kernel and initrd extracted from the image (after
>>>>>>> which
>>>>>>> the file is closed). Once again, nothing uses write support and it
>>>>>>> was
>>>>>>> added just to make it use O_DIRECT for both read and write
>>>>>>> operations
>>>>>>> but only pygrub uses only read support and O_DIRECT passed here is
>>>>>>> the
>>>>>>> only way to make it use non-cached data.
>>>>>> So what caches get in the way? From the above it seems the situation
>>>>>> is the following:
>>>>>>
>>>>>> - filesystem N is a guest filesystem. It's not usually mounted on
>>>>>> the
>>>>>> host, except for initial setup long time ago
>>>>>
>>>>> Yes, it is really a guest file system. This is not mounted in the
>>>>> host
>>>>> and the reason is to get actual version of grub.conf, initrd and
>>>>> kernel
>>>>> to be booted...
>>>>>
>>>>>> - before booting a guest your "pygrub" tools needs to read files on
>>>>>> it, and it's doing so using e2fsprogs
>>>>>
>>>>> Correct.
>>>>>
>>>>>> - once the guest is life it uses the extN kernel driver to access
>>>>>> the
>>>>>> filesystem
>>>>>
>>>>> That's right. So this is no longer pygrub responsibility...
>>>>>
>>>>>> nowhere in this cycle you should have any stale cached data. The
>>>>>> kernel
>>>>>> always makes sure to write back data on umount/reboot, as does
>>>>>> e2fsprogs
>>>>>> if actually used to write data (which you said is not the case
>>>>>> anyway).
>>>>>
>>>>> In fact I was unable to run into those problems myself but
>>>>> reporter/customer did.
>>>>>
>>>>>> The only data that may be in the cache are unmodified data from
>>>>>> reads
>>>>>> on the block device from either e2fsprogs or a suboptimal virtual
>>>>>> block
>>>>>> device implementation, but these can't cause any problems.
>>>>> Michal
>>>>
>>>> If the guest is the only one (when running) that installs a new
>>>> grub.conf file and kernel and it shuts down properly, you should be
>>>> good. It if does not shut down cleanly, it could have a stale
>>>> grub.conf file (or worse, a partially written one), but using
>>>> O_DIRECT to bypass the file system cache should not help.
>>>>
>>>> If we cannot reproduce this failure, sounds like we need to go back
>>>> and get a better understanding of what the customer saw?
>>>>
>>>> ric
>>>>
>>> That's right. I am going write an e-mail regarding this information to
>>> the reproducer if this bug and tell him that I need more information
>>> about what's happening at the customer side.
>>>
>> One more thing to point out, let's have a look at:
>> https://bugzilla.redhat.com/show_bug.cgi?id=466681#c15 .This is about
>> workaround to drop caches to be added to pygrub in the host machine
>> using this command:
>>
>> echo 1> /proc/sys/vm/drop_caches
>>
>> So this really looks like the caching issue if it's working fine after
>> dropping the caches. That may be the reason why this could be fine with
>> this patch present in e2fsprogs.
>>
>> Michal
>
> That BZ has a pretty long and twisted history, but after a quick read,
> I still don't see why a cleanly shutdown guest would have issues with
> caching that using O_DIRECT on read would help.
>
> We will need to dig into a bit more...
>
> ric
>
I am not saying we don't need to dig a little bit more, we surely do
but unfortunately I am waiting for information from reporter. But I am
also thinking that this O_DIRECT functionality support to bypass caches
could be useful...
Thanks,
Michal
On 01/12/2010 03:33 PM, Chris Lee wrote:
>
>
> Michal Novotny wrote:
>> On 01/12/2010 02:29 PM, Ric Wheeler wrote:
>>> On 01/12/2010 08:23 AM, Michal Novotny wrote:
>>>> On 01/12/2010 02:12 PM, Michal Novotny wrote:
>>>>> On 01/12/2010 02:04 PM, Ric Wheeler wrote:
>>>>>> On 01/12/2010 08:01 AM, Michal Novotny wrote:
>>>>>>> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>>>>>>>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>>>>>>>> Not really, pygrub doesn't do any manipulation with file
>>>>>>>>> system and
>>>>>>>>> also, it's not working on a life file system. It's called
>>>>>>>>> before the
>>>>>>>>> guest boots up to read information about grub.conf/initrd and
>>>>>>>>> kernel for
>>>>>>>>> PV guest and after this is read and selected in pygrub then the
>>>>>>>>> guest is
>>>>>>>>> booted using the kernel and initrd extracted from the image
>>>>>>>>> (after
>>>>>>>>> which
>>>>>>>>> the file is closed). Once again, nothing uses write support
>>>>>>>>> and it
>>>>>>>>> was
>>>>>>>>> added just to make it use O_DIRECT for both read and write
>>>>>>>>> operations
>>>>>>>>> but only pygrub uses only read support and O_DIRECT passed
>>>>>>>>> here is
>>>>>>>>> the
>>>>>>>>> only way to make it use non-cached data.
>>>>>>>> So what caches get in the way? From the above it seems the
>>>>>>>> situation
>>>>>>>> is the following:
>>>>>>>>
>>>>>>>> - filesystem N is a guest filesystem. It's not usually mounted
>>>>>>>> on the
>>>>>>>> host, except for initial setup long time ago
>>>>>>>
>>>>>>> Yes, it is really a guest file system. This is not mounted in
>>>>>>> the host
>>>>>>> and the reason is to get actual version of grub.conf, initrd and
>>>>>>> kernel
>>>>>>> to be booted...
>>>>>>>
>>>>>>>> - before booting a guest your "pygrub" tools needs to read
>>>>>>>> files on
>>>>>>>> it, and it's doing so using e2fsprogs
>>>>>>>
>>>>>>> Correct.
>>>>>>>
>>>>>>>> - once the guest is life it uses the extN kernel driver to
>>>>>>>> access the
>>>>>>>> filesystem
>>>>>>>
>>>>>>> That's right. So this is no longer pygrub responsibility...
>>>>>>>
>>>>>>>> nowhere in this cycle you should have any stale cached data. The
>>>>>>>> kernel
>>>>>>>> always makes sure to write back data on umount/reboot, as does
>>>>>>>> e2fsprogs
>>>>>>>> if actually used to write data (which you said is not the case
>>>>>>>> anyway).
>>>>>>>
>>>>>>> In fact I was unable to run into those problems myself but
>>>>>>> reporter/customer did.
>>>>>>>
>>>>>>>> The only data that may be in the cache are unmodified data from
>>>>>>>> reads
>>>>>>>> on the block device from either e2fsprogs or a suboptimal virtual
>>>>>>>> block
>>>>>>>> device implementation, but these can't cause any problems.
>>>>>>> Michal
>>>>>>
>>>>>> If the guest is the only one (when running) that installs a new
>>>>>> grub.conf file and kernel and it shuts down properly, you should be
>>>>>> good. It if does not shut down cleanly, it could have a stale
>>>>>> grub.conf file (or worse, a partially written one), but using
>>>>>> O_DIRECT to bypass the file system cache should not help.
>>>>>>
>>>>>> If we cannot reproduce this failure, sounds like we need to go back
>>>>>> and get a better understanding of what the customer saw?
>>>>>>
>>>>>> ric
>>>>>>
>>>>> That's right. I am going write an e-mail regarding this
>>>>> information to
>>>>> the reproducer if this bug and tell him that I need more information
>>>>> about what's happening at the customer side.
>>>>>
>>>> One more thing to point out, let's have a look at:
>>>> https://bugzilla.redhat.com/show_bug.cgi?id=466681#c15 .This is about
>>>> workaround to drop caches to be added to pygrub in the host machine
>>>> using this command:
>>>>
>>>> echo 1> /proc/sys/vm/drop_caches
>>>>
>>>> So this really looks like the caching issue if it's working fine after
>>>> dropping the caches. That may be the reason why this could be fine
>>>> with
>>>> this patch present in e2fsprogs.
>>>>
>>>> Michal
>>>
>>> That BZ has a pretty long and twisted history, but after a quick
>>> read, I still don't see why a cleanly shutdown guest would have
>>> issues with caching that using O_DIRECT on read would help.
>>>
>>> We will need to dig into a bit more...
>>>
>>> ric
>>>
>> I am not saying we don't need to dig a little bit more, we surely do
>> but unfortunately I am waiting for information from reporter. But I
>> am also thinking that this O_DIRECT functionality support to bypass
>> caches could be useful...
>>
>> Thanks,
>> Michal
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>
> I can not see where the cache could cause this problem but is it
> possible that it is in the Host file system rather than than the guest
> where it is causing a problem;
This may be right because drop caches in the host is a working
workaround. Also, I am having some information about it. Scott wrote
that he was able to reproduce it but with my patches applied it is
working fine. I am waiting for more information about that and customer
test results...
Michal
Michal Novotny wrote:
> On 01/12/2010 02:29 PM, Ric Wheeler wrote:
>> On 01/12/2010 08:23 AM, Michal Novotny wrote:
>>> On 01/12/2010 02:12 PM, Michal Novotny wrote:
>>>> On 01/12/2010 02:04 PM, Ric Wheeler wrote:
>>>>> On 01/12/2010 08:01 AM, Michal Novotny wrote:
>>>>>> On 01/12/2010 01:46 PM, Christoph Hellwig wrote:
>>>>>>> On Tue, Jan 12, 2010 at 01:30:40PM +0100, Michal Novotny wrote:
>>>>>>>> Not really, pygrub doesn't do any manipulation with file system
>>>>>>>> and
>>>>>>>> also, it's not working on a life file system. It's called
>>>>>>>> before the
>>>>>>>> guest boots up to read information about grub.conf/initrd and
>>>>>>>> kernel for
>>>>>>>> PV guest and after this is read and selected in pygrub then the
>>>>>>>> guest is
>>>>>>>> booted using the kernel and initrd extracted from the image (after
>>>>>>>> which
>>>>>>>> the file is closed). Once again, nothing uses write support and it
>>>>>>>> was
>>>>>>>> added just to make it use O_DIRECT for both read and write
>>>>>>>> operations
>>>>>>>> but only pygrub uses only read support and O_DIRECT passed here is
>>>>>>>> the
>>>>>>>> only way to make it use non-cached data.
>>>>>>> So what caches get in the way? From the above it seems the
>>>>>>> situation
>>>>>>> is the following:
>>>>>>>
>>>>>>> - filesystem N is a guest filesystem. It's not usually mounted
>>>>>>> on the
>>>>>>> host, except for initial setup long time ago
>>>>>>
>>>>>> Yes, it is really a guest file system. This is not mounted in the
>>>>>> host
>>>>>> and the reason is to get actual version of grub.conf, initrd and
>>>>>> kernel
>>>>>> to be booted...
>>>>>>
>>>>>>> - before booting a guest your "pygrub" tools needs to read files on
>>>>>>> it, and it's doing so using e2fsprogs
>>>>>>
>>>>>> Correct.
>>>>>>
>>>>>>> - once the guest is life it uses the extN kernel driver to
>>>>>>> access the
>>>>>>> filesystem
>>>>>>
>>>>>> That's right. So this is no longer pygrub responsibility...
>>>>>>
>>>>>>> nowhere in this cycle you should have any stale cached data. The
>>>>>>> kernel
>>>>>>> always makes sure to write back data on umount/reboot, as does
>>>>>>> e2fsprogs
>>>>>>> if actually used to write data (which you said is not the case
>>>>>>> anyway).
>>>>>>
>>>>>> In fact I was unable to run into those problems myself but
>>>>>> reporter/customer did.
>>>>>>
>>>>>>> The only data that may be in the cache are unmodified data from
>>>>>>> reads
>>>>>>> on the block device from either e2fsprogs or a suboptimal virtual
>>>>>>> block
>>>>>>> device implementation, but these can't cause any problems.
>>>>>> Michal
>>>>>
>>>>> If the guest is the only one (when running) that installs a new
>>>>> grub.conf file and kernel and it shuts down properly, you should be
>>>>> good. It if does not shut down cleanly, it could have a stale
>>>>> grub.conf file (or worse, a partially written one), but using
>>>>> O_DIRECT to bypass the file system cache should not help.
>>>>>
>>>>> If we cannot reproduce this failure, sounds like we need to go back
>>>>> and get a better understanding of what the customer saw?
>>>>>
>>>>> ric
>>>>>
>>>> That's right. I am going write an e-mail regarding this information to
>>>> the reproducer if this bug and tell him that I need more information
>>>> about what's happening at the customer side.
>>>>
>>> One more thing to point out, let's have a look at:
>>> https://bugzilla.redhat.com/show_bug.cgi?id=466681#c15 .This is about
>>> workaround to drop caches to be added to pygrub in the host machine
>>> using this command:
>>>
>>> echo 1> /proc/sys/vm/drop_caches
>>>
>>> So this really looks like the caching issue if it's working fine after
>>> dropping the caches. That may be the reason why this could be fine with
>>> this patch present in e2fsprogs.
>>>
>>> Michal
>>
>> That BZ has a pretty long and twisted history, but after a quick
>> read, I still don't see why a cleanly shutdown guest would have
>> issues with caching that using O_DIRECT on read would help.
>>
>> We will need to dig into a bit more...
>>
>> ric
>>
> I am not saying we don't need to dig a little bit more, we surely do
> but unfortunately I am waiting for information from reporter. But I am
> also thinking that this O_DIRECT functionality support to bypass
> caches could be useful...
>
> Thanks,
> Michal
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
I can not see where the cache could cause this problem but is it
possible that it is in the Host file system rather than than the guest
where it is causing a problem; If a guest shuts down it writes to it's
file system and all is good only it's file system is a file on another
file system. So the cache looking after those writes as managed by the
hyper visor or whatever could hold that data un-flushed for whatever reason.
But when the guest starts up it should get the cached or most recent
version, this should not be stale data unless more than one guest is
using this file system and they are overwriting each others files, then
a cached version might be newer and different from what is expected and
the on disk version might be the old and expected version.
End user needs to provide more information.
Chris.
Christoph Hellwig wrote:
> On Tue, Jan 12, 2010 at 01:15:49PM +0100, Michal Novotny wrote:
>> I don't really know if I see your point but the thing here is that there
>> was no way to open a file directly (ie. using O_DIRECT). The direct
>> write support has been added only to make it possible to use both read
>> and write directly. The main reason to create this patch was to add
>> direct read support and flush capability won't help me at all. I am
>> working in Red Hat, Virtualization team on Xen so I am really not that
>> much familiar with file systems but what I needed was an option to read
>> the data directly (using O_DIRECT) in e2fsprogs. One bug was about
>> pygrub (Python version of GRUB of Xen PV guests that is internally using
>> e2fsprogs functionality to access data on ext2/3/4 partition to boot the
>> PV guests) uses outdated/cached data so some modifications were
>> necessary to open everything directly...
>
> So to get things staigt: you're using e2fsprogs to manipulate a life
> filesystem and thing using O_DIRECT saves your ass? I think you need to
> rething your model of operation fundamentally in that case.
>
Christoph -
It's my understanding that nobody is doing concurrent access.
If the host reads the block device via pygrub to boot the guest,
the guest while running updates the same device when installing
a new kernel (either through the fs, or by writing to the bdev;
probably the former...), and then the guest shuts down - is there
something which will sync the host's cache (cached from the
prior read) with the updates from the guest?
I don't know nearly enough about how the virt IO goes ...
But if access is sequential between host & guest, I'd hope that
this would all just work.
If that is all supposed to work, then maybe the description of the
data flow (as Ric asked for) is not correct.
-Eric
On 01/12/2010 04:16 PM, Eric Sandeen wrote:
> Christoph Hellwig wrote:
>
>> On Tue, Jan 12, 2010 at 01:15:49PM +0100, Michal Novotny wrote:
>>
>>> I don't really know if I see your point but the thing here is that there
>>> was no way to open a file directly (ie. using O_DIRECT). The direct
>>> write support has been added only to make it possible to use both read
>>> and write directly. The main reason to create this patch was to add
>>> direct read support and flush capability won't help me at all. I am
>>> working in Red Hat, Virtualization team on Xen so I am really not that
>>> much familiar with file systems but what I needed was an option to read
>>> the data directly (using O_DIRECT) in e2fsprogs. One bug was about
>>> pygrub (Python version of GRUB of Xen PV guests that is internally using
>>> e2fsprogs functionality to access data on ext2/3/4 partition to boot the
>>> PV guests) uses outdated/cached data so some modifications were
>>> necessary to open everything directly...
>>>
>> So to get things staigt: you're using e2fsprogs to manipulate a life
>> filesystem and thing using O_DIRECT saves your ass? I think you need to
>> rething your model of operation fundamentally in that case.
>>
>>
> Christoph -
>
> It's my understanding that nobody is doing concurrent access.
>
>
Eric,
indeed, no concurrent accesses are happening that time. The guests file
system seems to be cached since after dropping caches in host (dom0) it
is reported to work correctly.
> If the host reads the block device via pygrub to boot the guest,
> the guest while running updates the same device when installing
> a new kernel (either through the fs, or by writing to the bdev;
> probably the former...), and then the guest shuts down - is there
> something which will sync the host's cache (cached from the
> prior read) with the updates from the guest?
>
It appears that the data are cached in host (dom0) and another access
(after guest update/shutdown sequence) makes dom0 still provide old
(cached) data and not the current (newest) data.
Michal
Ok, I looked at the issue. The problem is that the Xen backend drivers
are (as expected) utterly braindead and submit bios directly from the
virtualization backed without using proper abstractions and thus
bypassing all the cache coherency features in the fileystems (the block
device nodes are just another mini-filesystem in that respect). So
when you first have buffered access in the host pages may stay in cache
and get overwritten directly on disk by a Xen guest, and once the guest
is down the host may still use the now stale cached data.
I would recommend to migrate your cutomers to KVM which uses the proper
abtractions and thus doesn't have this problem. There's a reason after
all why all the Xen dom0 mess never got merged to mainline.
On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
> Ok, I looked at the issue. The problem is that the Xen backend drivers
> are (as expected) utterly braindead and submit bios directly from the
> virtualization backed without using proper abstractions and thus
> bypassing all the cache coherency features in the fileystems (the block
> device nodes are just another mini-filesystem in that respect). So
> when you first have buffered access in the host pages may stay in cache
> and get overwritten directly on disk by a Xen guest, and once the guest
> is down the host may still use the now stale cached data.
>
> I would recommend to migrate your cutomers to KVM which uses the proper
> abtractions and thus doesn't have this problem. There's a reason after
> all why all the Xen dom0 mess never got merged to mainline.
>
So, do you think the problem is in the Xen backend drivers and to make
it working right in Xen the driver fix is needed?
On Tue, Jan 12, 2010 at 05:43:21PM +0100, Michal Novotny wrote:
> So, do you think the problem is in the Xen backend drivers and to make
> it working right in Xen the driver fix is needed?
Yes, the Xen blkback driver just submits I/O directly without using
the right interfaces to force cache coherency. It might be relatively
easy to hack a call in to flush all caches when it starts up, but given
how it bypasses all abstractions it's almost impossible to give full
coherency as if using the normal block device interfaces.
On 01/12/2010 11:43 AM, Michal Novotny wrote:
> On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
>> Ok, I looked at the issue. The problem is that the Xen backend drivers
>> are (as expected) utterly braindead and submit bios directly from the
>> virtualization backed without using proper abstractions and thus
>> bypassing all the cache coherency features in the fileystems (the block
>> device nodes are just another mini-filesystem in that respect). So
>> when you first have buffered access in the host pages may stay in cache
>> and get overwritten directly on disk by a Xen guest, and once the guest
>> is down the host may still use the now stale cached data.
>>
>> I would recommend to migrate your cutomers to KVM which uses the proper
>> abtractions and thus doesn't have this problem. There's a reason after
>> all why all the Xen dom0 mess never got merged to mainline.
> So, do you think the problem is in the Xen backend drivers and to make
> it working right in Xen the driver fix is needed?
If XEN drivers by pass the normal IO and FS stack on the host, then I can
understand why the hack to e2fsprogs works but it does not seem like a good fix.
Specifically, the data will continue to be cached (and if dirty, might be
written back to the storage eventually).
If we need a work around, you need to drop VM caches for that device before you
update the guest's files and possibly again afterwards (and make sure that
nothing pulls the data into cache during the operation).
Basically, this sounds like the backend drivers are doing something really,
really dangerous....
ric
On 01/12/2010 05:47 PM, Christoph Hellwig wrote:
> On Tue, Jan 12, 2010 at 05:43:21PM +0100, Michal Novotny wrote:
>
>> So, do you think the problem is in the Xen backend drivers and to make
>> it working right in Xen the driver fix is needed?
>>
> Yes, the Xen blkback driver just submits I/O directly without using
> the right interfaces to force cache coherency. It might be relatively
> easy to hack a call in to flush all caches when it starts up, but given
> how it bypasses all abstractions it's almost impossible to give full
> coherency as if using the normal block device interfaces.
>
>
Ok, this way the fix to the drivers should be done. Thanks for your
investigation. I just forwarded your reply to other members of my team.
Thanks,
Michal
On 01/12/2010 05:50 PM, Ric Wheeler wrote:
> On 01/12/2010 11:43 AM, Michal Novotny wrote:
>> On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
>>> Ok, I looked at the issue. The problem is that the Xen backend drivers
>>> are (as expected) utterly braindead and submit bios directly from the
>>> virtualization backed without using proper abstractions and thus
>>> bypassing all the cache coherency features in the fileystems (the block
>>> device nodes are just another mini-filesystem in that respect). So
>>> when you first have buffered access in the host pages may stay in cache
>>> and get overwritten directly on disk by a Xen guest, and once the guest
>>> is down the host may still use the now stale cached data.
>>>
>>> I would recommend to migrate your cutomers to KVM which uses the proper
>>> abtractions and thus doesn't have this problem. There's a reason after
>>> all why all the Xen dom0 mess never got merged to mainline.
>> So, do you think the problem is in the Xen backend drivers and to make
>> it working right in Xen the driver fix is needed?
>
> If XEN drivers by pass the normal IO and FS stack on the host, then I
> can understand why the hack to e2fsprogs works but it does not seem
> like a good fix.
>
> Specifically, the data will continue to be cached (and if dirty, might
> be written back to the storage eventually).
>
> If we need a work around, you need to drop VM caches for that device
> before you update the guest's files and possibly again afterwards (and
> make sure that nothing pulls the data into cache during the operation).
>
> Basically, this sounds like the backend drivers are doing something
> really, really dangerous....
>
> ric
>
Ok, so you think this is not good to do this patch for e2fsprogs for
direct access support? The only things we could do now is to fix the
backend drivers or create a workaround to drop caches? I need to discuss
this further with guys in my team...
Thanks,
Michal
Michal Novotny wrote:
> On 01/12/2010 05:50 PM, Ric Wheeler wrote:
>> On 01/12/2010 11:43 AM, Michal Novotny wrote:
>>> On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
>>>> Ok, I looked at the issue. The problem is that the Xen backend drivers
>>>> are (as expected) utterly braindead and submit bios directly from the
>>>> virtualization backed without using proper abstractions and thus
>>>> bypassing all the cache coherency features in the fileystems (the block
>>>> device nodes are just another mini-filesystem in that respect). So
>>>> when you first have buffered access in the host pages may stay in cache
>>>> and get overwritten directly on disk by a Xen guest, and once the guest
>>>> is down the host may still use the now stale cached data.
>>>>
>>>> I would recommend to migrate your cutomers to KVM which uses the proper
>>>> abtractions and thus doesn't have this problem. There's a reason after
>>>> all why all the Xen dom0 mess never got merged to mainline.
>>> So, do you think the problem is in the Xen backend drivers and to make
>>> it working right in Xen the driver fix is needed?
>>
>> If XEN drivers by pass the normal IO and FS stack on the host, then I
>> can understand why the hack to e2fsprogs works but it does not seem
>> like a good fix.
>>
>> Specifically, the data will continue to be cached (and if dirty, might
>> be written back to the storage eventually).
>>
>> If we need a work around, you need to drop VM caches for that device
>> before you update the guest's files and possibly again afterwards (and
>> make sure that nothing pulls the data into cache during the operation).
>>
>> Basically, this sounds like the backend drivers are doing something
>> really, really dangerous....
>>
>> ric
>>
> Ok, so you think this is not good to do this patch for e2fsprogs for
> direct access support? The only things we could do now is to fix the
> backend drivers or create a workaround to drop caches? I need to discuss
> this further with guys in my team...
I do think that patching it up in e2fsprogs is unnecessarily invasive;
it's fixing it at the wrong spot.
Any block dev IO from the host is dangerous; fixing it only in e2fsprogs
for this one case doesn't seem like the right course of action.
-Eric
> Thanks,
> Michal
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
On 01/12/2010 11:56 AM, Eric Sandeen wrote:
> Michal Novotny wrote:
>> On 01/12/2010 05:50 PM, Ric Wheeler wrote:
>>> On 01/12/2010 11:43 AM, Michal Novotny wrote:
>>>> On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
>>>>> Ok, I looked at the issue. The problem is that the Xen backend drivers
>>>>> are (as expected) utterly braindead and submit bios directly from the
>>>>> virtualization backed without using proper abstractions and thus
>>>>> bypassing all the cache coherency features in the fileystems (the block
>>>>> device nodes are just another mini-filesystem in that respect). So
>>>>> when you first have buffered access in the host pages may stay in cache
>>>>> and get overwritten directly on disk by a Xen guest, and once the guest
>>>>> is down the host may still use the now stale cached data.
>>>>>
>>>>> I would recommend to migrate your cutomers to KVM which uses the proper
>>>>> abtractions and thus doesn't have this problem. There's a reason after
>>>>> all why all the Xen dom0 mess never got merged to mainline.
>>>> So, do you think the problem is in the Xen backend drivers and to make
>>>> it working right in Xen the driver fix is needed?
>>>
>>> If XEN drivers by pass the normal IO and FS stack on the host, then I
>>> can understand why the hack to e2fsprogs works but it does not seem
>>> like a good fix.
>>>
>>> Specifically, the data will continue to be cached (and if dirty, might
>>> be written back to the storage eventually).
>>>
>>> If we need a work around, you need to drop VM caches for that device
>>> before you update the guest's files and possibly again afterwards (and
>>> make sure that nothing pulls the data into cache during the operation).
>>>
>>> Basically, this sounds like the backend drivers are doing something
>>> really, really dangerous....
>>>
>>> ric
>>>
>> Ok, so you think this is not good to do this patch for e2fsprogs for
>> direct access support? The only things we could do now is to fix the
>> backend drivers or create a workaround to drop caches? I need to discuss
>> this further with guys in my team...
>
> I do think that patching it up in e2fsprogs is unnecessarily invasive;
> it's fixing it at the wrong spot.
>
> Any block dev IO from the host is dangerous; fixing it only in e2fsprogs
> for this one case doesn't seem like the right course of action.
>
> -Eric
It actually could produce some nastier issues where it would work a bit, the bad
data gets flushed back to the backing store and then your O_DIRECT read would be
broken.
Also, for normal users of e2fsprogs, they should never bypass the cache...
ric
On 01/12/2010 05:59 PM, Ric Wheeler wrote:
> On 01/12/2010 11:56 AM, Eric Sandeen wrote:
>> Michal Novotny wrote:
>>> On 01/12/2010 05:50 PM, Ric Wheeler wrote:
>>>> On 01/12/2010 11:43 AM, Michal Novotny wrote:
>>>>> On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
>>>>>> Ok, I looked at the issue. The problem is that the Xen backend
>>>>>> drivers
>>>>>> are (as expected) utterly braindead and submit bios directly from
>>>>>> the
>>>>>> virtualization backed without using proper abstractions and thus
>>>>>> bypassing all the cache coherency features in the fileystems (the
>>>>>> block
>>>>>> device nodes are just another mini-filesystem in that respect). So
>>>>>> when you first have buffered access in the host pages may stay in
>>>>>> cache
>>>>>> and get overwritten directly on disk by a Xen guest, and once the
>>>>>> guest
>>>>>> is down the host may still use the now stale cached data.
>>>>>>
>>>>>> I would recommend to migrate your cutomers to KVM which uses the
>>>>>> proper
>>>>>> abtractions and thus doesn't have this problem. There's a reason
>>>>>> after
>>>>>> all why all the Xen dom0 mess never got merged to mainline.
>>>>> So, do you think the problem is in the Xen backend drivers and to
>>>>> make
>>>>> it working right in Xen the driver fix is needed?
>>>>
>>>> If XEN drivers by pass the normal IO and FS stack on the host, then I
>>>> can understand why the hack to e2fsprogs works but it does not seem
>>>> like a good fix.
>>>>
>>>> Specifically, the data will continue to be cached (and if dirty, might
>>>> be written back to the storage eventually).
>>>>
>>>> If we need a work around, you need to drop VM caches for that device
>>>> before you update the guest's files and possibly again afterwards (and
>>>> make sure that nothing pulls the data into cache during the
>>>> operation).
>>>>
>>>> Basically, this sounds like the backend drivers are doing something
>>>> really, really dangerous....
>>>>
>>>> ric
>>>>
>>> Ok, so you think this is not good to do this patch for e2fsprogs for
>>> direct access support? The only things we could do now is to fix the
>>> backend drivers or create a workaround to drop caches? I need to
>>> discuss
>>> this further with guys in my team...
>>
>> I do think that patching it up in e2fsprogs is unnecessarily invasive;
>> it's fixing it at the wrong spot.
>>
>> Any block dev IO from the host is dangerous; fixing it only in e2fsprogs
>> for this one case doesn't seem like the right course of action.
>>
>> -Eric
>
> It actually could produce some nastier issues where it would work a
> bit, the bad data gets flushed back to the backing store and then your
> O_DIRECT read would be broken.
>
> Also, for normal users of e2fsprogs, they should never bypass the
> cache...
>
> ric
Ok, good to know the normal users should never bypass the cache. That
way this seems to be some kind of kernel bug...
Thank you all for your input,
Michal
Just for completeness, this is what we think we have going on:
1) First boot: Running on the host, pygrub reads the device (unmounted)
to bootstrap the guest with image A (kernel + grub.conf)
(2) guest image updates the kernel/grub.conf using weird Xen IO path
(bypassing the host page cache, creating BIO's directly in the host
memory).
Note that at this point in time, the image from (1) is still possibly in
the page cache of the host
(3) reboot of guest - host pygrub uses page cache (stale pages) when
bootstrapping the guest who occasionally boots into the stale image.
This doesn't happen all of the time - if the pages are dropped before
(3) happens, it won't happen.
Uses O_DIRECT has a side effect of invalidating the page cache while
reading.
Dropping VM caches just for the devices in question would fix this or
(as Christoph mentioned) use kvm which does this more sanely :-)
ric
On 01/12/2010 06:00 PM, Michal Novotny wrote:
> On 01/12/2010 05:59 PM, Ric Wheeler wrote:
>> On 01/12/2010 11:56 AM, Eric Sandeen wrote:
>>> Michal Novotny wrote:
>>>> On 01/12/2010 05:50 PM, Ric Wheeler wrote:
>>>>> On 01/12/2010 11:43 AM, Michal Novotny wrote:
>>>>>> On 01/12/2010 05:38 PM, Christoph Hellwig wrote:
>>>>>>> Ok, I looked at the issue. The problem is that the Xen backend
>>>>>>> drivers
>>>>>>> are (as expected) utterly braindead and submit bios directly
>>>>>>> from the
>>>>>>> virtualization backed without using proper abstractions and thus
>>>>>>> bypassing all the cache coherency features in the fileystems
>>>>>>> (the block
>>>>>>> device nodes are just another mini-filesystem in that respect). So
>>>>>>> when you first have buffered access in the host pages may stay
>>>>>>> in cache
>>>>>>> and get overwritten directly on disk by a Xen guest, and once
>>>>>>> the guest
>>>>>>> is down the host may still use the now stale cached data.
>>>>>>>
>>>>>>> I would recommend to migrate your cutomers to KVM which uses the
>>>>>>> proper
>>>>>>> abtractions and thus doesn't have this problem. There's a reason
>>>>>>> after
>>>>>>> all why all the Xen dom0 mess never got merged to mainline.
>>>>>> So, do you think the problem is in the Xen backend drivers and to
>>>>>> make
>>>>>> it working right in Xen the driver fix is needed?
>>>>>
>>>>> If XEN drivers by pass the normal IO and FS stack on the host, then I
>>>>> can understand why the hack to e2fsprogs works but it does not seem
>>>>> like a good fix.
>>>>>
>>>>> Specifically, the data will continue to be cached (and if dirty,
>>>>> might
>>>>> be written back to the storage eventually).
>>>>>
>>>>> If we need a work around, you need to drop VM caches for that device
>>>>> before you update the guest's files and possibly again afterwards
>>>>> (and
>>>>> make sure that nothing pulls the data into cache during the
>>>>> operation).
>>>>>
>>>>> Basically, this sounds like the backend drivers are doing something
>>>>> really, really dangerous....
>>>>>
>>>>> ric
>>>>>
>>>> Ok, so you think this is not good to do this patch for e2fsprogs for
>>>> direct access support? The only things we could do now is to fix the
>>>> backend drivers or create a workaround to drop caches? I need to
>>>> discuss
>>>> this further with guys in my team...
>>>
>>> I do think that patching it up in e2fsprogs is unnecessarily invasive;
>>> it's fixing it at the wrong spot.
>>>
>>> Any block dev IO from the host is dangerous; fixing it only in
>>> e2fsprogs
>>> for this one case doesn't seem like the right course of action.
>>>
>>> -Eric
>>
>> It actually could produce some nastier issues where it would work a
>> bit, the bad data gets flushed back to the backing store and then
>> your O_DIRECT read would be broken.
>>
>> Also, for normal users of e2fsprogs, they should never bypass the
>> cache...
>>
>> ric
> Ok, good to know the normal users should never bypass the cache. That
> way this seems to be some kind of kernel bug...
>
Hi,
finally, after discussing this with several people we decided it's
better to solve it in kernel path since Xen is using some weird IO path
like mentioned. Since this patch is about using O_DIRECT in e2fsprogs
which I've been told that normal e2fsprogs user should never do I'd like
to thanks you all for your help and tell you to ignore this patch since
we're going to solve it some other way, ie. fix this in the kernel itself.
Thanks for your help!
Michal