Tom Lane wrote:
> It's fairly hard to see how that could happen inside Postgres. One can
> readily imagine bugs that might replace one whole page with another,
> but there aren't any operations that manipulate half-a-page. On the
> other hand, if your kernel uses 4K blocksize, this could be explained
> as substitution of one buffer for another at the kernel level. So my
> money is on a kernel bug. As I mentioned, we've once or twice before
> seen reports that looked like similar substitutions of single pages by
> the kernel.
Yeah, I agree that's what's likely to have happened here. The
kernel's page size is 4k, as is the default block size used by XFS.
Tracking this one down any further is going to be nigh unto
impossible, I think.
There have been a pile of fixes to XFS in 2.6.16, so I'll start using
that, at least...
Thanks for looking at this.
--
Kevin Brown kevin@sysexperts.com