On Sat, Apr 29, 2023 at 04:32:34PM +0800, Yin, Fengwei wrote:
> Hi Kirill,
>
> On 4/28/2023 10:02 PM, Kirill A. Shutemov wrote:
> > On Fri, Apr 28, 2023 at 02:28:07PM +0800, Yin, Fengwei wrote:
> >> Hi Kirill,
> >>
> >> On 4/25/2023 8:38 PM, Kirill A. Shutemov wrote:
> >>> On Tue, Apr 25, 2023 at 04:46:26PM +0800, Yin Fengwei wrote:
> >>>> free_transhuge_page() acquires split queue lock then check
> >>>> whether the THP was added to deferred list or not.
> >>>>
> >>>> It's safe to check whether the THP is in deferred list or not.
> >>>> When code hit free_transhuge_page(), there is no one tries
> >>>> to update the folio's _deferred_list.
> >>>>
> >>>> If folio is not in deferred_list, it's safe to check without
> >>>> acquiring lock.
> >>>>
> >>>> If folio is in deferred_list, the other node in deferred_list
> >>>> adding/deleteing doesn't impact the return value of
> >>>> list_epmty(@folio->_deferred_list).
> >>>
> >>> Typo.
> >>>
> >>>>
> >>>> Running page_fault1 of will-it-scale + order 2 folio for anonymous
> >>>> mapping with 96 processes on an Ice Lake 48C/96T test box, we could
> >>>> see the 61% split_queue_lock contention:
> >>>> - 71.28% 0.35% page_fault1_pro [kernel.kallsyms] [k]
> >>>> release_pages
> >>>> - 70.93% release_pages
> >>>> - 61.42% free_transhuge_page
> >>>> + 60.77% _raw_spin_lock_irqsave
> >>>>
> >>>> With this patch applied, the split_queue_lock contention is less
> >>>> than 1%.
> >>>>
> >>>> Signed-off-by: Yin Fengwei <fengwei.yin@xxxxxxxxx>
> >>>> Tested-by: Ryan Roberts <ryan.roberts@xxxxxxx>
> >>>> ---
> >>>> mm/huge_memory.c | 19 ++++++++++++++++---
> >>>> 1 file changed, 16 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >>>> index 032fb0ef9cd1..c620f1f12247 100644
> >>>> --- a/mm/huge_memory.c
> >>>> +++ b/mm/huge_memory.c
> >>>> @@ -2799,12 +2799,25 @@ void free_transhuge_page(struct page *page)
> >>>> struct deferred_split *ds_queue = get_deferred_split_queue(folio);
> >>>> unsigned long flags;
> >>>>
> >>>> - spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
> >>>> - if (!list_empty(&folio->_deferred_list)) {
> >>>> + /*
> >>>> + * At this point, there is no one trying to queue the folio
> >>>> + * to deferred_list. folio->_deferred_list is not possible
> >>>> + * being updated.
> >>>> + *
> >>>> + * If folio is already added to deferred_list, add/delete to/from
> >>>> + * deferred_list will not impact list_empty(&folio->_deferred_list).
> >>>> + * It's safe to check list_empty(&folio->_deferred_list) without
> >>>> + * acquiring the lock.
> >>>> + *
> >>>> + * If folio is not in deferred_list, it's safe to check without
> >>>> + * acquiring the lock.
> >>>> + */
> >>>> + if (data_race(!list_empty(&folio->_deferred_list))) {
> >>>> + spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
> >>>
> >>> Recheck under lock?
> >> In function deferred_split_scan(), there is following code block:
> >> if (folio_try_get(folio)) {
> >> list_move(&folio->_deferred_list, &list);
> >> } else {
> >> /* We lost race with folio_put() */
> >> list_del_init(&folio->_deferred_list);
> >> ds_queue->split_queue_len--;
> >> }
> >>
> >> I am wondering what kind of "lost race with folio_put()" can be.
> >>
> >> My understanding is that it's not necessary to handle this case here
> >> because free_transhuge_page() will handle it once folio get zero ref.
> >> But I must miss something here. Thanks.
> >
> > free_transhuge_page() got when refcount is already zero. Both
> > deferred_split_scan() and free_transhuge_page() can see the page with zero
> > refcount. The check makes deferred_split_scan() to leave the page to the
> > free_transhuge_page().
> >
> If deferred_split_scan() leaves the page to free_transhuge_page(), is it
> necessary to do
> list_del_init(&folio->_deferred_list);
> ds_queue->split_queue_len--;
>
> Can these two line be left to free_transhuge_page() either? Thanks.
I *think* (my cache is cold on deferred split) we can. But since we
already hold the lock, why not take care of it? It makes your change more
efficient.
--
Kiryl Shutsemau / Kirill A. Shutemov
