Re: 2.6.22-rc4-mm1

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 6 Jun 2007 09:30:53 -0700 William Lee Irwin III <wli@holomorphy.com> wrote:

> On Wed, Jun 06, 2007 at 02:07:37AM -0700, Andrew Morton wrote:
> > ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.22-rc4/2.6.22-rc4-mm1/
> > - Somebody broke it on my powerpc G5, but I didn't have time to do yet
> >   another bisection yet.
> > - There's a lengthy patch series here from Nick which attempts to address
> >   the longstanding pagefault-vs-buffered-write deadlock.
> >   A great shower of filesystems were broken and have been disabled with
> >   CONFIG_BROKEN.  This includes reiser4.
> > - Complex patches which eliminate the kernel's fixed size limit on the
> >   command-line length.  These break nommu builds.
> 
> Someone remind me what the pagefault vs. buffered write deadlock is.

generic_file_write() does lock_page(), then copies the user's data into
pagecache.  If that copy_from_user() encounters a major fault and the page
is not uptodate, the pagefault handler does lock_page() and deadlocks.

It requires that the user be writ()ing from a mmap of the page back into
the same page, which is weird.

The kernel tries to prefault the page to avoid the copy_from_user() fault,
but there are ways in whcih that can be defeated (super memory pressure,
malicious fadvise() from a second thread, etc).

> Something brings down i386/qemu before even earlyprintk can handle.
> 
> Bisection has narrowed it down to patch 1140 after everything got
> renumbered by peterz' fix for mm-variable-length-argument-support.patch,
> namely containersv10-make-cpusets-a-client-of-containers.patch

erk.  A step-by-step how-to-make-this-happen might help if poss, please.