Re: [RFC] RDMA/umem: pin_user_pages*() can temporarily fail due to migration glitches

[Jason Gunthorpe <jgg@nvidia.com>]

On Tue, Apr 30, 2024 at 10:10:43PM -0700, Christoph Hellwig wrote:
> > +		pinned = -ENOMEM;
> > +		int attempts = 0;
> > +		/*
> > +		 * pin_user_pages_fast() can return -EAGAIN, due to falling back
> > +		 * to gup-slow and then failing to migrate pages out of
> > +		 * ZONE_MOVABLE due to a transient elevated page refcount.
> > +		 *
> > +		 * One retry is enough to avoid this problem, so far, but let's
> > +		 * use a slightly higher retry count just in case even larger
> > +		 * systems have a longer-lasting transient refcount problem.
> > +		 *
> > +		 */
> > +		static const int MAX_ATTEMPTS = 3;
> > +
> > +		while (pinned == -EAGAIN && attempts < MAX_ATTEMPTS) {
> > +			pinned = pin_user_pages_fast(cur_base,
> > +						     min_t(unsigned long,
> > +							npages, PAGE_SIZE /
> > +							sizeof(struct page *)),
> > +						     gup_flags, page_list);
> >  			ret = pinned;
> > -			goto umem_release;
> > +			attempts++;
> > +
> > +			if (pinned == -EAGAIN)
> > +				continue;
> >  		}
> > +		if (pinned < 0)
> > +			goto umem_release;
> 
> This doesn't make sense.  IFF a blind retry is all that is needed it
> should be done in the core functionality.  I fear it's not that easy,
> though.

+1

This migration retry weirdness is a GUP issue, it needs to be solved
in the mm not exposed to every pin_user_pages caller.

If it turns out ZONE_MOVEABLE pages can't actually be reliably moved
then it is pretty broken..

Jason