We have seen some cases where XFS deadlocks under heavy memory pressure,
we end up trying to free memory from a code path which has xfs locks,
the freeing of memory requires other xfs locks and sometimes a dependency
between the two can cause a hang.
There are changes in the current development cvs tree which may fix this,
we know that it is not bug free right now, and are working on it. In the
meantime, could you possibly try a kernel built from the development tree
and let us know if you can hang it in the same way.
Steve
> I have run into a strange problem testing xfs. My test is quite simple,
> take 2 drives make them both xfs, create dummy data (eg tar /usr into
> usr.tar) create more dummy data so that some of the individual tar files
> are greater than 2G, when I have about 10G total data do the following:
>
> loop forever
> tar dummy data (both large and small files) from one xfs drive to the other
> delete all the data on the 2nd xfs drive
> continue loop
>
> After a while, sometimes in the middle of the first loop, other times
> after 6 loops, or 15 loops, or 1.5 days worth of loops the whole machine
> freezes. (one test machine actually turns off).
>
> I have had this happen on pentiums, pentium II, and Athlons. I test with
> maxtor drives mostly though I can get the same thing happening with
> western digital drives. The ide controlers include VIA vt82xxxx and
> Intel PIIx and Promise PDC202xxx. I have been squeezing the most speed
> out of the IDE stuff for a while, meaning compiling into the kernel the
> right drivers, and turning on multicount (multi_mode) and DMA. These
> kernel patches have worked in the past (meaning same test, just no >2G
> files) using ext2.
>
> The xfs file systems are corrupted after the freeze. After the reboot it
> is necessary to umount the files systems and xfs_repair them.
>
> I am currently using the .9 pre-release.
>
> I have tried with multi_mode off, with dma off. I have tried with
> building the file system using
>
> mkfs -t xfs -l internal,size=16000b -f /dev/hdg1
> and
> mkfs -t xfs -l internal,size=16000b -d unwritten=0 -f /dev/hdg1
> and
> mkfs -t xfs -l size=8000b -d unwritten=0 -f /dev/hdg1
> and
> mkfs -t xfs -d unwritten=0 -f /dev/hdg1
>
> I am hoping someone recognizes this, and can point me in the right
> direction. I have a hunch it is something in my making of the kernel,
> because prior to sending this email, I downloaded the install RPM, and
> installed the SGI-REDHAT7-XFS default system on yet another test
> machine--without making
> any changes at all it seems to be running my little stress test the
> longest (this test machine also has brand new drives/mother board/memory/etc)
>
> This email is long enough as it is, but if seeing the .config or my
> turning on any xfs debugging would help anyone help me...please let me know.
>
> Thank you
>
> Sean Dougherty
> Manager of way too much
> TTUHSC At Amarillo
> sean@xxxxxxxxxxxxxx
|