Re: [PATCH 1/3] migrate_pages: fix deadlock in batched migration

[Hugh Dickins]

On Tue, 28 Feb 2023, Huang, Ying wrote:
> Hugh Dickins <hughd@xxxxxxxxxx> writes:
> > On Fri, 24 Feb 2023, Huang Ying wrote:
> >> @@ -1247,7 +1236,7 @@ static int migrate_folio_unmap(new_page_t get_new_page, free_page_t put_new_page
> >>  		/* Establish migration ptes */
> >>  		VM_BUG_ON_FOLIO(folio_test_anon(src) &&
> >>  			       !folio_test_ksm(src) && !anon_vma, src);
> >> -		try_to_migrate(src, TTU_BATCH_FLUSH);
> >> +		try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
> >
> > Why that change, I wonder? The TTU_BATCH_FLUSH can still be useful for
> > gathering multiple cross-CPU TLB flushes into one, even when it's only
> > a single page in the batch.
> 
> Firstly, I would have thought that we have no opportunities to batch the
> TLB flushing now.  But as you pointed out, it is still possible to batch
> if mapcount > 1.  Secondly, without TTU_BATCH_FLUSH, we may flush the
> TLB for a single page (with invlpg instruction), otherwise, we will
> flush the TLB for all pages.  The former is faster and will not
> influence other TLB entries of the process.
> 
> Or we use TTU_BATCH_FLUSH only if mapcount > 1?

I had not thought at all of the "invlpg" advantage (which I imagine
some other architectures than x86 share) to not delaying the TLB flush
of a single PTE.

Frankly, I just don't have any feeling for the tradeoff between
multiple remote invlpgs versus one remote batched TLB flush of all.
Which presumably depends on number of CPUs, size of TLBs, etc etc.

Your "mapcount > 1" idea might be good, but I cannot tell: I'd say
now that there's no reason to change your "mode == MIGRATE_ASYNC ?
TTU_BATCH_FLUSH : 0" without much more thought, or a quick insight
from someone else.  Some other time maybe.

Hugh