Re: [PATCH] fix pgd_lock deadlock

["Jan Beulich" <JBeulich@novell.com>]

>>> On 21.02.11 at 15:53, Johannes Weiner <jweiner@redhat.com> wrote:
> On Mon, Feb 21, 2011 at 03:30:23PM +0100, Andrea Arcangeli wrote:
>> On Thu, Feb 17, 2011 at 11:19:41AM +0100, Johannes Weiner wrote:
>> > So Xen needs all page tables protected when pinning/unpinning and
>> > extended page_table_lock to cover kernel range, which it does nowhere
>> > else AFAICS.  But the places it extended are also taking the pgd_lock,
>> > so I wonder if Xen could just take the pgd_lock itself in these paths
>> > and we could revert page_table_lock back to cover user va only?
>> > Jeremy, could this work?  Untested.
>> 
>> If this works for Xen, I definitely prefer this.
> 
> Below is real submission, with changelog and sign-off and all (except
> testing on Xen itself, sorry).  I moved pgd_lock acquisition in this
> version to make the lock ordering perfectly clear.  Xen people, could
> you have a look at this?

While I think that it would be correct, it doesn't look like a
reasonable fix to me: It effectively serializes process (address
space) construction and destruction.

A possible alternative would be to acquire the page table lock
in vmalloc_sync_all() only in the Xen case (perhaps by storing
NULL into page->index in pgd_set_mm() when not running on
Xen). This is utilizing the fact that there aren't (supposed to
be - for non-pvops this is definitely the case) any TLB flush IPIs
under Xen, and hence the race you're trying to fix doesn't
exist there (while non-Xen doesn't need the extra locking).

Jan