Re: [PATCH] mm: fix page_mkclean_one (was: 2.6.19 file content corruption on ext3)

[Martin Michlmayr <tbm@cyrius.com>]

* Linus Torvalds <torvalds@osdl.org> [2006-12-20 09:35]:
> Can you remind us:
>  - your ARM is UP, right? Do you have PREEMPT on?

It's UP and PREEMPT is not set.  I used 2.6.19 plus the patch that has
been posted.

>  - This is probably a stupid question, but you did make sure that the
>    database was ok (with some rebuild command) and that you didn't have
>    preexisting corruption?

Yes, my test case is to install Debian on the ARM machine so the
database is created fresh.  While the corruption always triggers
during a fresh installation, it's much harder to see in a running
system.  Some people see it on their system but I haven't found a 100%
working recipe to reproduce it yet given a working system; doing a new
installation seems to trigger it all the time though.

> Anyway, the page_mkclean_one() fixes (along with _most_ things we've
> looked at) shouldn't matter on UP, at least certainly not without
> PREEMPT.

Hmm.  So what about UP without PREEMPT then...

Maybe the following information is helpful in some way: remember how I
said that we have applied 6 mm patches to 2.6.18 in Debian?  According
to Gordon Farquharson, who's helping me a great deal with testing
installation on this ARM machine (Linksys NSLU2), the corruption
doesn't always show up when you only apply
mm-tracking-shared-dirty-pages.patch to 2.6.18 but it shows up all the
time with all six patches applied.  As a reminder, the 6 patches we
apply are:

mm-tracking-shared-dirty-pages.patch
mm-balance-dirty-pages.patch
mm-optimize-mprotect.patch
mm-install_page-cleanup.patch
mm-do_wp_page-fixup.patch
mm-msync-cleanup.patch

-- 
Martin Michlmayr
http://www.cyrius.com/