On Fri, Nov 18, 2022 at 3:33 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
>
> On Fri, Nov 18, 2022 at 02:51:01PM -0700, Yu Zhao wrote:
> > On Fri, Nov 18, 2022 at 2:25 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Nov 17, 2022 at 06:40:05PM -0700, Yu Zhao wrote:
> > > > On Thu, Nov 17, 2022 at 6:26 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, Nov 17, 2022 at 03:22:42PM -0700, Yu Zhao wrote:
> > > > > > On Thu, Nov 17, 2022 at 12:47 AM Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Tue, Nov 15, 2022 at 06:38:07PM -0700, Yu Zhao wrote:
> > > > > > > > The page reclaim isolates a batch of folios from the tail of one of
> > > > > > > > the LRU lists and works on those folios one by one. For a suitable
> > > > > > > > swap-backed folio, if the swap device is async, it queues that folio
> > > > > > > > for writeback. After the page reclaim finishes an entire batch, it
> > > > > > > > puts back the folios it queued for writeback to the head of the
> > > > > > > > original LRU list.
> > > > > > > >
> > > > > > > > In the meantime, the page writeback flushes the queued folios also by
> > > > > > > > batches. Its batching logic is independent from that of the page
> > > > > > > > reclaim. For each of the folios it writes back, the page writeback
> > > > > > > > calls folio_rotate_reclaimable() which tries to rotate a folio to the
> > > > > > > > tail.
> > > > > > > >
> > > > > > > > folio_rotate_reclaimable() only works for a folio after the page
> > > > > > > > reclaim has put it back. If an async swap device is fast enough, the
> > > > > > > > page writeback can finish with that folio while the page reclaim is
> > > > > > > > still working on the rest of the batch containing it. In this case,
> > > > > > > > that folio will remain at the head and the page reclaim will not retry
> > > > > > > > it before reaching there.
> > > > > > > >
> > > > > > > > This patch adds a retry to evict_folios(). After evict_folios() has
> > > > > > > > finished an entire batch and before it puts back folios it cannot free
> > > > > > > > immediately, it retries those that may have missed the rotation.
> > > > > > >
> > > > > > > Can we make something like this?
> > > > > >
> > > > > > This works for both the active/inactive LRU and MGLRU.
> > > > >
> > > > > I hope we fix both altogether.
> > > > >
> > > > > >
> > > > > > But it's not my prefered way because of these two subtle differences:
> > > > > > 1. Folios eligible for retry take an unnecessary round trip below --
> > > > > > they are first added to the LRU list and then removed from there for
> > > > > > retry. For high speed swap devices, the LRU lock contention is already
> > > > > > quite high (>10% in CPU profile under heavy memory pressure). So I'm
> > > > > > hoping we can avoid this round trip.
> > > > > > 2. The number of retries of a folio on folio_wb_list is unlimited,
> > > > > > whereas this patch limits the retry to one. So in theory, we can spin
> > > > > > on a bunch of folios that keep failing.
> > > > > >
> > > > > > The most ideal solution would be to have the one-off retry logic in
> > > > > > shrink_folio_list(). But right now, that function is very cluttered. I
> > > > > > plan to refactor it (low priority at the moment), and probably after
> > > > > > that, we can add a generic retry for both the active/inactive LRU and
> > > > > > MGLRU. I'll raise its priority if you strongly prefer this. Please
> > > > > > feel free to let me know.
> > > > >
> > > > > Well, my preference for *ideal solution* is writeback completion drops
> > > > > page immediately without LRU rotating. IIRC, concern was softirq latency
> > > > > and locking relevant in the context at that time when I tried it.
> > > >
> > > > Are we good for now or are there other ideas we want to try while we are at it?
> > > >
> > >
> > > good for now with what solution you are thinking? The retry logic you
> > > suggested? I personally don't like the solution relies on the timing.
> > >
> > > If you are concerning about unnecessary round trip, it shouldn't
> > > happen frequency since your assumption is swap device is so fast
> > > so second loop would see their wb done?
> >
> > No, the round trip that hits the LRU lock in the process.
>
> I see what you meant.
>
> >
> > For folios written and ready to be freed, they'll have to go from
> > being isolated to the tail of LRU list and then to getting isolated
> > again. This requires an extra hit on the LRU lock, which is highly
> > contended for fast swap devices under heavy memory pressure.
> >
> > > Anyway, I am strongly push my preference. Feel free to go with way
>
> Oh, sorry for the typo: "not strongly push my preference"
>
> > > you want if the solution can fix both LRU schemes.
> >
> > There is another concern I listed previously:
> >
> > > > > > 2. The number of retries of a folio on folio_wb_list is unlimited,
> > > > > > whereas this patch limits the retry to one. So in theory, we can spin
> > > > > > on a bunch of folios that keep failing.
> >
> > If this can happen, it'd be really hard to track it down. Any thoughts on this?
>
> Could you elaborate why folio_wb_list can keep spinning?
Currently page reclaim never puts a folio that it failed to evict back
to the tail because that folio can fail a second, a third time, etc.
If we want to add a retry logic, we need to make sure it's a limited
retry for each failed folio. Once its limit is reached, we'll have to
put that folio back to the head so that we won't risk spinning on it
indefinitely.
The below introduces an *unlimited* retry logic, and to make it work,
we would need to exclude all conditions that could cause wb_folio to
fail when it goes through shirnk_folio_list() again. Otherwise, it can
repeat the same process and deadlock. It's difficult to enumerate
those conditions exhaustively, because they depend on other layers of
the kernel.
/*
* If writeback is already done, move the page into tail.
* Otherwise, put the page into head and folio_rotate_reclaimable
* will move it to the tail when the writeback is done
*/
if (!folio_test_writeback(wb_folio)) &&
folio_test_reclaim(wb_folio))
lruvec_add_folio_tail(lruvec, folio);
> My concern is how we can make sure the timing bet is good for most
> workloads on heterogeneous/dvfs frequency core control env.
I agree. It's not an easy problem to solve.
