Re: [PATCH] mm/hmm: replace hmm_update with mmu_notifier_range

[Jason Gunthorpe <jgg@ziepe.ca>]

On Wed, Jul 24, 2019 at 07:58:58PM +0200, Michal Hocko wrote:
> On Wed 24-07-19 12:28:58, Jason Gunthorpe wrote:
> > On Wed, Jul 24, 2019 at 09:05:53AM +0200, Christoph Hellwig wrote:
> > > Looks good:
> > > 
> > > Reviewed-by: Christoph Hellwig <hch@lst.de>
> > > 
> > > One comment on a related cleanup:
> > > 
> > > >  	list_for_each_entry(mirror, &hmm->mirrors, list) {
> > > >  		int rc;
> > > >  
> > > > -		rc = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
> > > > +		rc = mirror->ops->sync_cpu_device_pagetables(mirror, nrange);
> > > >  		if (rc) {
> > > > -			if (WARN_ON(update.blockable || rc != -EAGAIN))
> > > > +			if (WARN_ON(mmu_notifier_range_blockable(nrange) ||
> > > > +			    rc != -EAGAIN))
> > > >  				continue;
> > > >  			ret = -EAGAIN;
> > > >  			break;
> > > 
> > > This magic handling of error seems odd.  I think we should merge rc and
> > > ret into one variable and just break out if any error happens instead
> > > or claiming in the comments -EAGAIN is the only valid error and then
> > > ignoring all others here.
> > 
> > The WARN_ON is enforcing the rules already commented near
> > mmuu_notifier_ops.invalidate_start - we could break or continue, it
> > doesn't much matter how to recover from a broken driver, but since we
> > did the WARN_ON this should sanitize the ret to EAGAIN or 0
> > 
> > Humm. Actually having looked this some more, I wonder if this is a
> > problem:
> > 
> > I see in __oom_reap_task_mm():
> > 
> > 			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
> > 				tlb_finish_mmu(&tlb, range.start, range.end);
> > 				ret = false;
> > 				continue;
> > 			}
> > 			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
> > 			mmu_notifier_invalidate_range_end(&range);
> > 
> > Which looks like it creates an unbalanced start/end pairing if any
> > start returns EAGAIN?
> > 
> > This does not seem OK.. Many users require start/end to be paired to
> > keep track of their internal locking. Ie for instance hmm breaks
> > because the hmm->notifiers counter becomes unable to get to 0.
> > 
> > Below is the best idea I've had so far..
> > 
> > Michal, what do you think?
> 
> IIRC we have discussed this with Jerome back then when I've introduced
> this code and unless I misremember he said the current code was OK.

Nope, it has always been broken.

> Maybe new users have started relying on a new semantic in the meantime,
> back then, none of the notifier has even started any action in blocking
> mode on a EAGAIN bailout. Most of them simply did trylock early in the
> process and bailed out so there was nothing to do for the range_end
> callback.

Single notifiers are not the problem. I tried to make this clear in
the commit message, but lets be more explicit.

We have *two* notifiers registered to the mm, A and B:

A invalidate_range_start: (has no blocking)
    spin_lock()
    counter++
    spin_unlock()

A invalidate_range_end:
    spin_lock()
    counter--
    spin_unlock()

And this one:

B invalidate_range_start: (has blocking)
    if (!try_mutex_lock())
        return -EAGAIN;
    counter++
    mutex_unlock()

B invalidate_range_end:
    spin_lock()
    counter--
    spin_unlock()

So now the oom path does:

invalidate_range_start_non_blocking:
 for each mn:
   a->invalidate_range_start
   b->invalidate_range_start
   rc = EAGAIN

Now we SKIP A's invalidate_range_end even though A had no idea this
would happen has state that needs to be unwound. A is broken.

B survived just fine.

A and B *alone* work fine, combined they fail.

When the commit was landed you can use KVM as an example of A and RDMA
ODP as an example of B

Jason