commit a79e53d8 upstream.

On Wednesday 16 February 2011 15:49:47 Andrea Arcangeli wrote:
> Subject: fix pgd_lock deadlock
>
> From: Andrea Arcangeli <aarcange@redhat.com>
>
> It's forbidden to take the page_table_lock with the irq disabled or if
> there's contention the IPIs (for tlb flushes) sent with the page_table_lock
> held will never run leading to a deadlock.
>
> Apparently nobody takes the pgd_lock from irq so the _irqsave can be
> removed.
>
> Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>

This patch (original commit Id for 2.6.38 a79e53d8)
needs to be back-ported to 2.6.32.x as well.
I observed a dead-lock problem when running a PAE enabled Debian 2.6.32.46+
kernel with 6 VCPUs as a KVM on (2.6.32, 3.2, 3.3) kernel, which showed the
following behaviour:

1 VCPU is stuck in
  pgd_alloc() =E2=86=92 pgd_prepopulate_pmb() =E2=86=92... =E2=86=92  flush_tlb_others_ipi()
while (!cpumask_empty(to_cpumask(f->flush_cpumask)))
    cpu_relax();
(gdb) print f->flush_cpumask
$5 = {1}

while all other VCPUs are stuck in
  pgd_alloc() =E2=86=92 spin_lock_irqsave(pgd_lock)

I tracked it down to the commit
 2.6.39-rc1: 4981d01e
 2.6.32.34: ba456fd7
 x86: Flush TLB if PGD entry is changed in i386 PAE mode
which when reverted made the bug disappear.

Comparing 3.2 to 2.6.32.34 showed that the 'pgd-deadlock'-patch went into
2.6.38, that is before the 'PAE correctness'-patch, so the problem was
probably never observed in the main development branch.
But for 2.6.32 the 'pgd-deadlock' patch is still missing, so the 'PAE
corretness'-patch made the problem worse with 2.6.32.

The Patch was also back-ported to the OpenSUSE Kernel
<http://kernel.opensuse.org/cgit/kernel-source/commit/?id=ac27c01aa880c65d17043ab87249c613ac4c3635>,
Since the patch didn't apply cleanly on the current Debian kernel, I had to
backport it for us and Debian. The patch is also available from our (German)
Bugzilla <https://forge.univention.org/bugzilla/show_bug.cgi?id=26661> or
from the Debian BTS at <http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=669335>.

I have no easy test case, but running multiple parallel builds inside the VM
normally triggers the bug within seconds to minutes. With the patch applied
the VM survived a night building packages without any problem.
Signed-off-by: Philipp Hahn <hahn@univention.de>

Sincerely
Philipp
-
Philipp Hahn           Open Source Software Engineer      hahn@univention.de
Univention GmbH        be open.                       fon: +49 421 22 232- 0
Mary-Somerville-Str.1  D-28359 Bremen                 fax: +49 421 22 232-99
                                                   http://www.univention.de/

It's forbidden to take the page_table_lock with the irq disabled
or if there's contention the IPIs (for tlb flushes) sent with
the page_table_lock held will never run leading to a deadlock.

Nobody takes the pgd_lock from irq context so the _irqsave can be
removed.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: <stable@kernel.org>
LKML-Reference: <201102162345.p1GNjMjm021738@imap1.linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Git-commit: a79e53d8Signed-off-by: Willy Tarreau <w@1wt.eu>