Re: [PATCH] mm: move idle swap cache pages to the tail of LRU after COW

[Huang, Ying]

Johannes Weiner <hannes@xxxxxxxxxxx> writes:

> On Thu, May 20, 2021 at 09:59:15AM +0800, Huang, Ying wrote:
>> Johannes Weiner <hannes@xxxxxxxxxxx> writes:
>> 
>> > On Thu, May 20, 2021 at 09:22:45AM +0800, Huang, Ying wrote:
>> >> Johannes Weiner <hannes@xxxxxxxxxxx> writes:
>> >> 
>> >> > On Wed, May 19, 2021 at 09:33:13AM +0800, Huang Ying wrote:
>> >> >> diff --git a/mm/memory.c b/mm/memory.c
>> >> >> index b83f734c4e1d..2b6847f4c03e 100644
>> >> >> --- a/mm/memory.c
>> >> >> +++ b/mm/memory.c
>> >> >> @@ -3012,6 +3012,11 @@ static vm_fault_t wp_page_copy(struct vm_fault *vmf)
>> >> >>  				munlock_vma_page(old_page);
>> >> >>  			unlock_page(old_page);
>> >> >>  		}
>> >> >> +		if (page_copied && PageSwapCache(old_page) &&
>> >> >> +		    !page_mapped(old_page) && trylock_page(old_page)) {
>> >> >> +			try_to_free_idle_swapcache(old_page);
>> >> >> +			unlock_page(old_page);
>> >> >
>> >> > If there are no more swap or pte references, can we just attempt to
>> >> > free the page right away, like we do during regular unmap?
>> >> >
>> >> > 		if (page_copied)
>> >> > 			free_swap_cache(old_page);
>> >> > 		put_page(old_page);
>> >> 
>> >> A previous version of the patch does roughly this.
>> >> 
>> >> https://lore.kernel.org/lkml/20210113024241.179113-1-ying.huang@xxxxxxxxx/
>> >> 
>> >> But Linus has concerns with the overhead introduced in the hot COW path.
>> >
>> > Sorry, I had missed that thread.
>> >
>> > It sounds like there were the same concerns about the LRU shuffling
>> > overhead in the COW page. Now we have numbers for that, but not the
>> > free_swap_cache version. Would you be able to run the numbers for that
>> > as well? It would be interesting to see how much the additional code
>> > complexity buys us.
>> 
>> The number for which workload?  The workload that is used to evaluate
>> this patch?
>
> Yeah, the pmbench one from the changelog.

Sure.  I have rebased the original patch that frees the idle swap cache
directly and done the test.  The results show that the pmbench score of
freeing directly is a little better than that of moving to the tail of
LRU.  The pmbench score increases about 3.6%.  I think this is expected,
because we need to free the page finally even if we move the idle swap
cache to the tail of LRU.

Best Regards,
Huang, Ying