On Wed, 2012-03-07 at 07:30 +1100, Dave Chinner wrote:
> On Mon, Mar 05, 2012 at 10:40:12PM +0100, Rafał Kupka wrote:
> > is good and first bad commit is:
> > 0e6e847f "xfs: stop using the page cache to back the buffer cache"
> Which changed internals of the buffer cache, but not the external
> > I double tested it - 704b2907 can do 60+ loops of xfsdump without any
> > trouble and 0e6e847f breaks in <10 iterations.
> What is your test case?
I run xfsdump (to /dev/null) in a loop and execute postmark benchmark at
the same time. It's a minimal Debian/Squeeze installation in KVM guest
but it happened on another system too. An Xen server with 512MB memory
xfsdump -f /dev/null -l0 -p1 -J -L "test" -M "test" /
and postmark with file (also in loop):
set size 500 1000000
set number 1000
set transactions 1000
Typically after few iterations xfsdump is stuck with log messages
mentioned in previous e-mails.
Important thing I recently notice - I'm able to reproduce this bug only
on small memory KVM guests (384M). This very same system started with 1G
memory runs fine for 100+ xfsdump executions.
> Also, it might be worthwhile to capture an event trace (e.g. all the
> xfs_buf events) to see what occurs just before the error is
> triggered. That might tell us exactl what sequence has occurred
> leading up to the error.
I will try to provide this information in a few days.