Re: [PATCH] - Fix unmap_vma() bug related to mmu_notifiers

[Andrea Arcangeli <aarcange@redhat.com>]

On Mon, Jan 25, 2010 at 03:10:33PM -0600, Jack Steiner wrote:
> On Mon, Jan 25, 2010 at 08:00:52PM +0100, Andrea Arcangeli wrote:
> > On Mon, Jan 25, 2010 at 11:45:56AM -0600, Jack Steiner wrote:
> > > unmap_vmas() can fail to correctly flush the TLB if a
> > > callout to mmu_notifier_invalidate_range_start() sleeps.
> > 
> > Not sure I understand: the callbacks invoked by
> > mmu_notifier_invalidate_range_start can't sleep, or rcu locking inside
> > mmu notifier will break too (first thing that should be replaced with
> > srcu if they were allowed to sleep).
> > 
> > In short there's no schedule that could be added because of those
> > callbacks so if this code isn't ok and schedules and screw on the
> > mmu_gather tlb it's probably not mmu notifier related.
> 
> Hmmmm. I was under the impression that the range callbacks _were_ allowed to
> sleep. That was certainly in the original patches posted by christoph. I was
> not aware that it had changed.

There's rcu, srcu not... srcu is the first thing to add, and then
anon_vma refcounting, then we can go sleepable...

> I'll let Robin provide details but the way mmu_notifiers are used by
> xpmem, I believe sleeping is essential. The 'start" callout sends messages to
> another OS instances to teardown cross-SSI mappings. That is difficult
> to do w/o a sleep.
> 
> We'll dig deeper tomorrow & see what we can do....

The old patches are in my ftp area, they should still apply, you
should concentrate testing with those additional ones applied, then it
will work for xpmem too ;)

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>