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
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