xfs fails dbench in 2.6.18-rc1-mm1

xfs fails dbench in 2.6.18-rc1-mm1

[Martin Bligh]

http://test.kernel.org/abat/40891/debug/test.log.1

Filesystem type for /mnt/tmp is xfs
write failed on handle 13786
4 clients started
Child failed with status 1
write failed on handle 13786
write failed on handle 13786
write failed on handle 13786

Works fine in -git4
All other fs's seemed to run OK.

Machine is a 4x Opteron.

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Eric Dumazet]

On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
> http://test.kernel.org/abat/40891/debug/test.log.1
>
> Filesystem type for /mnt/tmp is xfs
> write failed on handle 13786
> 4 clients started
> Child failed with status 1
> write failed on handle 13786
> write failed on handle 13786
> write failed on handle 13786
>
> Works fine in -git4
> All other fs's seemed to run OK.
>
> Machine is a 4x Opteron.

You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf

http://www.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blobdiff;h=3f356086061d7076a379b76e265295a5fe3750fe;hp=55f4e70225631b275f85215ee543b104507caacc;hb=92eb7a2f28d551acedeb5752263267a64b1f5ddf;f=fs/file.c

or change in alloc_fdtable()

nfds = max_t(int, 8 * L1_CACHE_BYTES, roundup_pow_of_two(nfds));

into

nfds = max_t(int, 8 * L1_CACHE_BYTES, roundup_pow_of_two(nr+1));

Eric

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Andy Whitcroft]

Eric Dumazet wrote:
> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
>> http://test.kernel.org/abat/40891/debug/test.log.1
>>
>> Filesystem type for /mnt/tmp is xfs
>> write failed on handle 13786
>> 4 clients started
>> Child failed with status 1
>> write failed on handle 13786
>> write failed on handle 13786
>> write failed on handle 13786
>>
>> Works fine in -git4
>> All other fs's seemed to run OK.
>>
>> Machine is a 4x Opteron.
> 
> You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf

Will give reverting that a spin.

-apw

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Martin Bligh]

Eric Dumazet wrote:
> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
> 
>>http://test.kernel.org/abat/40891/debug/test.log.1
>>
>>Filesystem type for /mnt/tmp is xfs
>>write failed on handle 13786
>>4 clients started
>>Child failed with status 1
>>write failed on handle 13786
>>write failed on handle 13786
>>write failed on handle 13786
>>
>>Works fine in -git4
>>All other fs's seemed to run OK.
>>
>>Machine is a 4x Opteron.
> 
> 
> You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf

Still fails (thanks Andy).

M.

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Badari Pulavarty]

Martin Bligh wrote:
> Eric Dumazet wrote:
>> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
>>
>>> http://test.kernel.org/abat/40891/debug/test.log.1
>>>
>>> Filesystem type for /mnt/tmp is xfs
>>> write failed on handle 13786
>>> 4 clients started
>>> Child failed with status 1
>>> write failed on handle 13786
>>> write failed on handle 13786
>>> write failed on handle 13786
>>>
>>> Works fine in -git4
>>> All other fs's seemed to run OK.
>>>
>>> Machine is a 4x Opteron.
>>
>>
>> You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf
>
> Still fails (thanks Andy).
>
Wondering if its my changes :(
Can you back out these and try ?

Please, Please tell me that, its not me :)

Thanks,
Badari

#
vectorize-aio_read-aio_write-fileop-methods.patch
remove-readv-writev-methods-and-use-aio_read-aio_write.patch
streamline-generic_file_-interfaces-and-filemap.patch
streamline-generic_file_-interfaces-and-filemap-ecryptfs.patch

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Martin Bligh]

Badari Pulavarty wrote:
> Martin Bligh wrote:
> 
>> Eric Dumazet wrote:
>>
>>> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
>>>
>>>> http://test.kernel.org/abat/40891/debug/test.log.1
>>>>
>>>> Filesystem type for /mnt/tmp is xfs
>>>> write failed on handle 13786
>>>> 4 clients started
>>>> Child failed with status 1
>>>> write failed on handle 13786
>>>> write failed on handle 13786
>>>> write failed on handle 13786
>>>>
>>>> Works fine in -git4
>>>> All other fs's seemed to run OK.
>>>>
>>>> Machine is a 4x Opteron.
>>>
>>>
>>>
>>> You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf
>>
>>
>> Still fails (thanks Andy).
>>
> Wondering if its my changes :(
> Can you back out these and try ?
> 
> Please, Please tell me that, its not me :)
> 
> Thanks,
> Badari
> 
> #
> vectorize-aio_read-aio_write-fileop-methods.patch
> remove-readv-writev-methods-and-use-aio_read-aio_write.patch
> streamline-generic_file_-interfaces-and-filemap.patch
> streamline-generic_file_-interfaces-and-filemap-ecryptfs.patch

You could submit a job to elm3b6 to run dbench on xfs ;-)

M.

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Martin Bligh]

> I am not able to "insmod xfs.ko" on my x86-64 machine :(
> 
> elm3b29:~ # modprobe xfs
> FATAL: Error inserting xfs (/lib/modules/2.6.18-rc1-
> mm1/kernel/fs/xfs/xfs.ko): Cannot allocate memory
> 
> #dmesg shows ..
> 
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 328 bytes percpu data
> Could not allocate 328 bytes percpu data
> Could not allocate 328 bytes percpu data
> 
> 
> Whats happening here ?



Dunno, but I built it statically, and didn't have that problem ;-)
The default config on elm3b6 should work.
http://ftp.kernel.org/pub/linux/kernel/people/mbligh/config/abat/amd64

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Badari Pulavarty]

On Wed, 2006-07-12 at 13:26 -0700, Martin Bligh wrote:
> Badari Pulavarty wrote:
> > Martin Bligh wrote:
> > 
> >> Eric Dumazet wrote:
> >>
> >>> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
> >>>
> >>>> http://test.kernel.org/abat/40891/debug/test.log.1
> >>>>
> >>>> Filesystem type for /mnt/tmp is xfs
> >>>> write failed on handle 13786
> >>>> 4 clients started
> >>>> Child failed with status 1
> >>>> write failed on handle 13786
> >>>> write failed on handle 13786
> >>>> write failed on handle 13786
> >>>>
> >>>> Works fine in -git4
> >>>> All other fs's seemed to run OK.
> >>>>
> >>>> Machine is a 4x Opteron.
> >>>
> >>>
> >>>
> >>> You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf
> >>
> >>
> >> Still fails (thanks Andy).
> >>
> > Wondering if its my changes :(
> > Can you back out these and try ?
> > 
> > Please, Please tell me that, its not me :)
> > 
> > Thanks,
> > Badari
> > 
> > #
> > vectorize-aio_read-aio_write-fileop-methods.patch
> > remove-readv-writev-methods-and-use-aio_read-aio_write.patch
> > streamline-generic_file_-interfaces-and-filemap.patch
> > streamline-generic_file_-interfaces-and-filemap-ecryptfs.patch
> 
> You could submit a job to elm3b6 to run dbench on xfs ;-)
> 
> M.


I am not able to "insmod xfs.ko" on my x86-64 machine :(

elm3b29:~ # modprobe xfs
FATAL: Error inserting xfs (/lib/modules/2.6.18-rc1-
mm1/kernel/fs/xfs/xfs.ko): Cannot allocate memory

#dmesg shows ..

Could not allocate 8 bytes percpu data
Could not allocate 8 bytes percpu data
Could not allocate 8 bytes percpu data
Could not allocate 8 bytes percpu data
Could not allocate 8 bytes percpu data
Could not allocate 8 bytes percpu data
Could not allocate 328 bytes percpu data
Could not allocate 328 bytes percpu data
Could not allocate 328 bytes percpu data


Whats happening here ?

Thanks,
Badari

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Keith Mannthey]

On 7/12/06, Badari Pulavarty <pbadari@us.ibm.com> wrote:
> On Wed, 2006-07-12 at 13:26 -0700, Martin Bligh wrote:
> > Badari Pulavarty wrote:
> > > Martin Bligh wrote:
> > >
> > >> Eric Dumazet wrote:
> > >>
> > >>> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
> > >>>
> > >>>> http://test.kernel.org/abat/40891/debug/test.log.1
> > >>>>
> > >>>> Filesystem type for /mnt/tmp is xfs
> > >>>> write failed on handle 13786
> > >>>> 4 clients started
> > >>>> Child failed with status 1
> > >>>> write failed on handle 13786
> > >>>> write failed on handle 13786
> > >>>> write failed on handle 13786
> > >>>>
> > >>>> Works fine in -git4
> > >>>> All other fs's seemed to run OK.
> > >>>>
> > >>>> Machine is a 4x Opteron.
> > >>>
> > >>>
> > >>>
> > >>> You need to revert 92eb7a2f28d551acedeb5752263267a64b1f5ddf
> > >>
> > >>
> > >> Still fails (thanks Andy).
> > >>
> > > Wondering if its my changes :(
> > > Can you back out these and try ?
> > >
> > > Please, Please tell me that, its not me :)
> > >
> > > Thanks,
> > > Badari
> > >
> > > #
> > > vectorize-aio_read-aio_write-fileop-methods.patch
> > > remove-readv-writev-methods-and-use-aio_read-aio_write.patch
> > > streamline-generic_file_-interfaces-and-filemap.patch
> > > streamline-generic_file_-interfaces-and-filemap-ecryptfs.patch
> >
> > You could submit a job to elm3b6 to run dbench on xfs ;-)
> >
> > M.
>
>
> I am not able to "insmod xfs.ko" on my x86-64 machine :(
>
> elm3b29:~ # modprobe xfs
> FATAL: Error inserting xfs (/lib/modules/2.6.18-rc1-
> mm1/kernel/fs/xfs/xfs.ko): Cannot allocate memory
>
> #dmesg shows ..
>
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 8 bytes percpu data
> Could not allocate 328 bytes percpu data
> Could not allocate 328 bytes percpu data
> Could not allocate 328 bytes percpu data
>
>
> Whats happening here ?

The per-cpu area is exhausted in -mm x68_64 (If you pump up the percpu
area or cut down NR_CPUS you can work around it).  I ran into this a
few -mm release ago. There are some details in the 2.6.17-mm6 thread
from July 5th of so.

Thanks,
  Keith

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Badari Pulavarty]

On Wed, 2006-07-12 at 13:26 -0700, Martin Bligh wrote:
> Badari Pulavarty wrote:
> > Martin Bligh wrote:
> > 
> >> Eric Dumazet wrote:
> >>
> >>> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
> >>>
> >>>> http://test.kernel.org/abat/40891/debug/test.log.1
> >>>>
> >>>> Filesystem type for /mnt/tmp is xfs
> >>>> write failed on handle 13786
> >>>> 4 clients started
> >>>> Child failed with status 1
> >>>> write failed on handle 13786
> >>>> write failed on handle 13786
> >>>> write failed on handle 13786
> >>>>
> >>>> Works fine in -git4
> >>>> All other fs's seemed to run OK.
> >>>>
> >>>> Machine is a 4x Opteron.

Sorry !! its my screw up again :(
Here is the patch to fix it.

Thanks,
Badari

Fix a bug in __xfs_file_write() which is causing writes to fail
with -EINVAL.

Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>

Index: linux-2.6.18-rc1/fs/xfs/linux-2.6/xfs_file.c
===================================================================
--- linux-2.6.18-rc1.orig/fs/xfs/linux-2.6/xfs_file.c	2006-07-11 21:28:08.000000000 -0700
+++ linux-2.6.18-rc1/fs/xfs/linux-2.6/xfs_file.c	2006-07-12 17:43:13.000000000 -0700
@@ -99,7 +99,7 @@ __xfs_file_write(
 	BUG_ON(iocb->ki_pos != pos);
 	if (unlikely(file->f_flags & O_DIRECT))
 		ioflags |= IO_ISDIRECT;
-	return bhv_vop_write(vp, iocb, &iov, nr_segs, &iocb->ki_pos,
+	return bhv_vop_write(vp, iocb, iov, nr_segs, &iocb->ki_pos,
 				ioflags, NULL);
 }
 

Re: xfs fails dbench in 2.6.18-rc1-mm1

[Andy Whitcroft]

Badari Pulavarty wrote:
> On Wed, 2006-07-12 at 13:26 -0700, Martin Bligh wrote:
>> Badari Pulavarty wrote:
>>> Martin Bligh wrote:
>>>
>>>> Eric Dumazet wrote:
>>>>
>>>>> On Wednesday 12 July 2006 18:58, Martin Bligh wrote:
>>>>>
>>>>>> http://test.kernel.org/abat/40891/debug/test.log.1
>>>>>>
>>>>>> Filesystem type for /mnt/tmp is xfs
>>>>>> write failed on handle 13786
>>>>>> 4 clients started
>>>>>> Child failed with status 1
>>>>>> write failed on handle 13786
>>>>>> write failed on handle 13786
>>>>>> write failed on handle 13786
>>>>>>
>>>>>> Works fine in -git4
>>>>>> All other fs's seemed to run OK.
>>>>>>
>>>>>> Machine is a 4x Opteron.
> 
> Sorry !! its my screw up again :(
> Here is the patch to fix it.
> 
> Thanks,
> Badari
> 
> Fix a bug in __xfs_file_write() which is causing writes to fail
> with -EINVAL.
> 
> Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>

Ok, I've thrown this one + apkm's latest fix into the tester.  Only one 
result in so far, but that is our first good on that machine.

-apw