Re: [PATCH 1/2] mm/mmu_notifier: Mark up direct reclaim paths with MAYFAIL

[Chris Wilson]

Quoting Jason Gunthorpe (2020-06-24 17:50:57)
> On Wed, Jun 24, 2020 at 03:37:32PM +0100, Chris Wilson wrote:
> > Quoting Jason Gunthorpe (2020-06-24 15:25:44)
> > > On Wed, Jun 24, 2020 at 03:21:49PM +0100, Chris Wilson wrote:
> > > > Quoting Jason Gunthorpe (2020-06-24 15:16:04)
> > > > > On Wed, Jun 24, 2020 at 03:12:42PM +0100, Chris Wilson wrote:
> > > > > > Quoting Jason Gunthorpe (2020-06-24 13:39:10)
> > > > > > > On Wed, Jun 24, 2020 at 01:21:03PM +0100, Chris Wilson wrote:
> > > > > > > > Quoting Jason Gunthorpe (2020-06-24 13:10:53)
> > > > > > > > > On Wed, Jun 24, 2020 at 09:02:47AM +0100, Chris Wilson wrote:
> > > > > > > > > > When direct reclaim enters the shrinker and tries to reclaim pages, it
> > > > > > > > > > has to opportunitically unmap them [try_to_unmap_one]. For direct
> > > > > > > > > > reclaim, the calling context is unknown and may include attempts to
> > > > > > > > > > unmap one page of a dma object while attempting to allocate more pages
> > > > > > > > > > for that object. Pass the information along that we are inside an
> > > > > > > > > > opportunistic unmap that can allow that page to remain referenced and
> > > > > > > > > > mapped, and let the callback opt in to avoiding a recursive wait.
> > > > > > > > > 
> > > > > > > > > i915 should already not be holding locks shared with the notifiers
> > > > > > > > > across allocations that can trigger reclaim. This is already required
> > > > > > > > > to use notifiers correctly anyhow - why do we need something in the
> > > > > > > > > notifiers?
> > > > > > > > 
> > > > > > > > for (n = 0; n < num_pages; n++)
> > > > > > > >       pin_user_page()
> > > > > > > > 
> > > > > > > > may call try_to_unmap_page from the lru shrinker for [0, n-1].
> > > > > > > 
> > > > > > > Yes, of course you can't hold any locks that intersect with notifiers
> > > > > > > across pin_user_page()/get_user_page()
> > > > > > 
> > > > > > What lock though? It's just the page refcount, shrinker asks us to drop
> > > > > > it [via mmu], we reply we would like to keep using that page as freeing
> > > > > > it for the current allocation is "robbing Peter to pay Paul".
> > > > > 
> > > > > Maybe I'm unclear what this series is actually trying to fix? 
> > > > > 
> > > > > You said "avoiding a recursive wait" which sounds like some locking
> > > > > deadlock to me.
> > > > 
> > > > It's the shrinker being called while we are allocating for/on behalf of
> > > > the object. As we are actively using the object, we don't want to free
> > > > it -- the partial object allocation being the clearest, if the object
> > > > consists of 2 pages, trying to free page 0 in order to allocate page 1
> > > > has to fail (and the shrinker should find another candidate to reclaim,
> > > > or fail the allocation).
> > > 
> > > mmu notifiers are not for influencing policy of the mm.
> > 
> > It's policy is "this may fail" regardless of the mmu notifier at this
> > point. That is not changed.
> 
> MMU notifiers are for tracking updates, they are not allowed to fail.
> The one slightly weird case of non-blocking is the only exception.
> 
> > Your suggestion is that we move the pages to the unevictable mapping so
> > that the shrinker LRU is never invoked on pages we have grabbed with
> > pin_user_page. Does that work with the rest of the mmu notifiers?
> 
> That is beyond what I'm familiar with - but generally - if you want to
> influence decisions the MM is making then it needs to be at the
> front of the process and not inside notifiers. 
> 
> So what you describe seems broadly appropriate to me.

Sadly, it's a mlock_vma_page problem all over again.
 
> I'm still a little unclear on what you are trying to fix - pinned
> pages are definitely not freed, do you have some case where pages
> which are pinned are being cleaned out from the MM despite being
> pinned? Sounds a bit strange, maybe that is worth adressing directly?

It suffices to say that pin_user_pages does not prevent try_to_unmap_one
from trying to revoke the page. But we could perhaps slip a
page_maybe_dma_pinned() in around there and see what happens.
-Chris