Re: [PATCH 3/3] migrate_pages: try migrate in batch asynchronously firstly

["Huang, Ying" <ying.huang@intel.com>]

Hugh Dickins <hughd@google.com> writes:

> On Wed, 1 Mar 2023, Huang, Ying wrote:
>> Hugh Dickins <hughd@google.com> writes:
>> > On Tue, 28 Feb 2023, Huang, Ying wrote:
>> >> Hugh Dickins <hughd@google.com> writes:
>> >> > On Fri, 24 Feb 2023, Huang Ying wrote:
>> >> >> 
>> >> >> diff --git a/mm/migrate.c b/mm/migrate.c
>> >> >> index 91198b487e49..c17ce5ee8d92 100644
>> >> >> --- a/mm/migrate.c
>> >> >> +++ b/mm/migrate.c
>> >> >> @@ -1843,6 +1843,51 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
>> >> >>  	return rc;
>> >> >>  }
>> >> >>  
>> >> >> +static int migrate_pages_sync(struct list_head *from, new_page_t get_new_page,
>> >> >> +		free_page_t put_new_page, unsigned long private,
>> >> >> +		enum migrate_mode mode, int reason, struct list_head *ret_folios,
>> >> >> +		struct list_head *split_folios, struct migrate_pages_stats *stats)
>> >> >> +{
>> >> >> +	int rc, nr_failed = 0;
>> >> >> +	LIST_HEAD(folios);
>> >> >> +	struct migrate_pages_stats astats;
>> >> >> +
>> >> >> +	memset(&astats, 0, sizeof(astats));
>> >> >> +	/* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
>> >> >> +	rc = migrate_pages_batch(from, get_new_page, put_new_page, private, MIGRATE_ASYNC,
>> >> >> +				 reason, &folios, split_folios, &astats,
>> >> >> +				 NR_MAX_MIGRATE_PAGES_RETRY);
>> >> >
>> >> > I wonder if that and below would better be NR_MAX_MIGRATE_PAGES_RETRY / 2.
>> >> >
>> >> > Though I've never got down to adjusting that number (and it's not a job
>> >> > to be done in this set of patches), those 10 retries sometimes terrify
>> >> > me, from a latency point of view.  They can have such different weights:
>> >> > in the unmapped case, 10 retries is okay; but when a pinned page is mapped
>> >> > into 1000 processes, the thought of all that unmapping and TLB flushing
>> >> > and remapping is terrifying.
>> >> >
>> >> > Since you're retrying below, halve both numbers of retries for now?
>> >> 
>> >> Yes.  These are reasonable concerns.
>> >> 
>> >> And in the original implementation, we only wait to lock page and wait
>> >> the writeback to complete if pass > 2.  This is kind of trying to
>> >> migrate asynchronously for 3 times before the real synchronous
>> >> migration.  So, should we delete the "force" logic (in
>> >> migrate_folio_unmap()), and try to migrate asynchronously for 3 times in
>> >> batch before migrating synchronously for 7 times one by one?
>> >
>> > Oh, that's a good idea (but please don't imagine I've thought it through):
>> > I hadn't realized the way in which your migrate_pages_sync() addition is
>> > kind of duplicating the way that the "force" argument conditions behaviour,
>> > It would be very appealing to delete the "force" argument now if you can.
>> 
>> Sure.  Will do that in the next version.
>> 
>> > But aside from that, you've also made me wonder (again, please remember I
>> > don't have a good picture of the new migrate_pages() sequence in my head)
>> > whether you have already made a *great* strike against my 10 retries
>> > terror.  Am I reading it right, that the unmapping is now done on the
>> > first try, and the remove_migration_ptes after the last try (all the
>> > pages involved having remained locked throughout)?
>> 
>> Yes.  You are right.  Now, unmapping and moving are two separate steps,
>> and they are retried separately.  After a folio has been unmapped
>> successfully, we will not remap/unmap it 10 times if the folio is pinned
>> so that failed to move (migrate_folio_move()).  So the latency caused by
>> retrying is much better now.  But I still tend to keep the total retry
>> number as before.  Do you agree?
>
> Yes, I agree, keep the total retry number 10 as before: maybe someone in
> future will show that more than 5 is a waste of time, but there's little
> need to get into that now: if you've put an end to that 10 times unmapping
> and remapping, that's a great step forward, quite apart from the TLB flush
> batching itself.
>
> (I did change "no need" to "little need" above: I do have some some
> anxiety about the increased latencies from keeping folios locked and
> migration entries in place for significantly longer than before your
> batching: I won't be surprised if the maximum batch size has to be
> lowered, if reports of latency spikes come in; and that might extend
> to the retry count too.)

Yes.  Latency are always concerns for batching.  We may revisit this
when needed.  Something good now is that we will never wait the lock or
bit in batched mode.  Latency tolerance depends on caller too, for
example, when we migrate some cold pages from DRAM to CXL MEM, we can
tolerate relatively long latency.  If so, we can add a parameter to
migrate_pages() to restrict the batch number and retry number when
necessary too.

Best Regards,
Huang, Ying