Re: [PATCH] mm: try to free swap only for reading swap fault

[zhouxianrong]

i mean for reading swap fault try_to_free_swap in do_swap_page
could hurt clean swap cache pages and make them dirty. it affects
reclaim procedure in shrink_page_list and let this function write
out much more these dirty anonymous pages. in fact these dirty
anonymous pages might keep clean originally.

On 2017/11/2 21:22, Michal Hocko wrote:
> On Thu 02-11-17 20:35:19, zhouxianrong@xxxxxxxxxx wrote:
> > From: zhouxianrong <zhouxianrong@xxxxxxxxxx>
> >
> > the purpose of this patch is that when a reading swap fault
> > happens on a clean swap cache page whose swap count is equal
> > to one, then try_to_free_swap could remove this page from
> > swap cache and mark this page dirty. so if later we reclaimed
> > this page then we could pageout this page due to this dirty.
> > so i want to allow this action only for writing swap fault.
> >
> > i sampled the data of non-dirty anonymous pages which is no
> > need to pageout and total anonymous pages in shrink_page_list.
> >
> > the results are:
> >
> >         non-dirty anonymous pages     total anonymous pages
> > before  26343                         635218
> > after   36907                         634312
>
> This data is absolutely pointless without describing the workload.
> You patch also stil fails to explain which workloads are going to
> benefit/suffer from the change and why it is a good thing to do in
> general.
>
> > Signed-off-by: zhouxianrong <zhouxianrong@xxxxxxxxxx>
> > ---
> >  mm/memory.c |    2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index a728bed..5a944fe 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -2999,7 +2999,7 @@ int do_swap_page(struct vm_fault *vmf)
> >  	}
> >
> >  	swap_free(entry);
> > -	if (mem_cgroup_swap_full(page) ||
> > +	if (((vmf->flags & FAULT_FLAG_WRITE) && mem_cgroup_swap_full(page)) ||
> >  	    (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
> >  		try_to_free_swap(page);
> >  	unlock_page(page);
> > --
> > 1.7.9.5