Re: Kernel 2.6.19.2 New RAID 5 Bug (oops when writing Samba -> RAID5)

From
Neil Brown <hidden>
Date
2007-01-23 02:07:16
Also in
linux-xfs, lkml
On Monday January 22, dan.j.williams@gmail.com wrote:
On 1/22/07, Neil Brown [off-list ref] wrote:
quoted
On Monday January 22, cebbert@redhat.com wrote:
quoted
Justin Piszcz wrote:
quoted
My .config is attached, please let me know if any other information is
needed and please CC (lkml) as I am not on the list, thanks!

Running Kernel 2.6.19.2 on a MD RAID5 volume.  Copying files over Samba to
the RAID5 running XFS.

Any idea what happened here?
....
quoted
quoted
Without digging too deeply, I'd say you've hit the same bug Sami Farin
and others
have reported starting with 2.6.19: pages mapped with kmap_atomic()
become unmapped
during memcpy() or similar operations.  Try disabling preempt -- that
seems to be the
common factor.
That is exactly the conclusion I had just come to (a kmap_atomic page
must be being unmapped during memcpy).  I wasn't aware that others had
reported it - thanks for that.

Turning off CONFIG_PREEMPT certainly seems like a good idea.
Coming from an ARM background I am not yet versed in the inner
workings of kmap_atomic, but if you have time for a question I am
curious as to why spin_lock(&sh->lock)  is not sufficient pre-emption
protection for copy_data() in this case?
Presumably there is a bug somewhere.
kmap_atomic itself calls inc_preempt_count so that preemption should
be disabled at least until the kunmap_atomic is called.

But apparently not.  The symptoms point exactly to the page getting
unmapped when it shouldn't.  Until that bug is found and fixed, the
work around of turning of CONFIG_PREEMPT seems to make sense.

Of course it would be great if someone who can easily reproduce this
bug could do the 'git bisect' thing to find out where the bug crept
in.....

NeilBrown
Thread (1 message) 1 message, 1 author, last reply 2007-01-23
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help