Re: 2.6.18 ext3 panic.

[Eric Sandeen <esandeen@redhat.com>]

Jan Kara wrote:

>> This is exactly  the solution I proposed earlier (to check 
>> buffer_mapped() before calling submit_bh()).
>> But at that time, Jan pointed out that the whole handling is wrong.
>   Yes, and it was. However it turned out that there are more problems
> than I thought ;).
> 
>> But if this is the only case we need to handle, I am okay with this band 
>> aid :)
>   I think Eric's patch may be a part of it. But we still need to check whether
> the buffer is not after EOF before submitting it (or better said just
> after we manage to lock the buffer). Because while we are waiting for
> the buffer lock, journal_unmap_buffer() can still come and steal the
> buffer - at least the write-out in journal_dirty_data() definitely needs
> the check if I haven't overlooked something.

Ok, let me think on that today.  My first reaction is that if we have
the bh state lock and pay attention to mapped in journal_dirty_data(),
then any blocks past EOF which have gotten unmapped by
journal_unmap_buffer will be recognized as such (because they are now
unmapped... without needing to check for past EOF...) and we'll be fine.

As a datapoint, davej's stresstest (several fsx's and fsstresses)
survived an overnight run on his box, which used to panic in < 2 hrs.
Survived about 6 hours on my box until I intentionally stopped it; my
box had added a write/truncate test in a loop, with a bunch of periodic
syncs as well....

-Eric