Re: [PATCH] mm/mglru: check need_resched() in fill/drain_evictable()

From: Kairui Song

Date: Thu Nov 12 2026 - 11:08:28 EST


On Thu, Nov 12, 2026 at 5:01 PM Hui Zhu <hui.zhu@xxxxxxxxx> wrote:
>
> Hi Michal and Kairui,
>
> 2026/11/12 16:09, Michal Hocko:
> > On Thu 12-11-26 16:03:54, Kairui Song wrote:
> >> On Thu, Nov 12, 2026 at 03:58:27PM +0800, Hui Zhu wrote:
> >>> From: Hui Zhu <zhuhui@xxxxxxxxxx>
> >>>
> >>> lru_gen_change_state() migrates folios in batches of up to
> >>> MAX_LRU_BATCH (4096) folios under lruvec_lock_irq(). A full batch
> >>> keeps local interrupts off for up to about a millisecond, and CPUs
> >>> contending for the same lru_lock spin with interrupts disabled for
> >>> that long.
> >>>
> >>> The caller already reschedules between batches, but a batch itself
> >>> only ends after 4096 folios. Stop it early when a resched is
> >>> pending; the caller restarts the migration where it left off. A
> >>> remote resched_curr() sets TIF_NEED_RESCHED directly in the target's
> >>> thread_info, so the flag can still be set with interrupts off.
> >>>
> >>> This follows walk_pud_range(), which already pairs need_resched()
> >>> with a MAX_LRU_BATCH cap, and commit 6e9be217a3ce ("mm/mglru: use a
> >>> smaller batch for reclaim").
> >>>
> >>> Signed-off-by: Hui Zhu <zhuhui@xxxxxxxxxx>
> >>> ---
> >>> mm/vmscan.c | 6 ++++--
> >>> 1 file changed, 4 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/mm/vmscan.c b/mm/vmscan.c
> >>> index 6a1af9b92349..47f2cb25eb46 100644
> >>> --- a/mm/vmscan.c
> >>> +++ b/mm/vmscan.c
> >>> @@ -5533,7 +5533,8 @@ static bool fill_evictable(struct lruvec *lruvec)
> >>> success = lru_gen_add_folio(lruvec, folio, !active);
> >>> VM_WARN_ON_ONCE(!success);
> >>>
> >>> - if (!--remaining)
> >>> + /* stop early to shorten the irq-off section */
> >>> + if (!--remaining || need_resched())
> >> Hi Hui,
> >>
> >> I've been thinking about a different design, what if we just
> >> delete the `--remaining` check completely, and check
> >> spin_is_contended on the lruvec and need_resched()?
> >>
> >> In the optimal case we don't need to keep lock/unlock the lruvec
> >> at all so this is faster. And it still avoids lock contention or
> >> blocked schedule, and it might be even more responsive as we are
> >> release the lock earlier on contention.
> I like the spin_is_contended() idea and will add it. This file
> already uses it (lru_gen_look_around() bails out if the PTL is
> contended), and security/keys/gc.c uses exactly this combination in a
> long list walk under a spinlock.
> > Wouldn't this result in too much lock bouncing in the end? Would it make
> > sense to reduce the batch instead?
> >
> This worry is fair. Under sustained contention each yield costs an
> unlock, a requeue and a restart (microseconds), while one folio move
> costs a few hundred nanoseconds, so the migration could slow down by
> tens of times, with cgroup_lock and cpus_read_lock held for all of it.
> Requiring a minimum of work per hold bounds this: only honor
> spin_is_contended() after MIN_LRU_BATCH folios in the current hold.
> Each hold then moves at least ~64 folios (~15us), and a waiter still
> waits only ~15us instead of the ~1ms of a full MAX_LRU_BATCH batch.
>
> I'd still keep the batch cap itself as a backstop, though. Both
> triggers are event driven: with interrupts off the local tick cannot
> set TIF_NEED_RESCHED, so it takes a remote wakeup targeting this CPU,
> and spin_is_contended() takes somebody grabbing this lru_lock. On a
> quiet system neither may ever fire, and one batch becomes the whole
> lruvec: 100GB of 4K folios is ~25M of them, i.e. seconds of irq-off
> inside a single rcu_read_lock() section, enough for an RCU stall. And
> toggling MGLRU on a quiet machine is a common case.
>
> Reducing the batch alone would also work and is simpler, but it pays
> the unlock/relock round trip even when nobody is waiting, and on a
> loaded machine the unconditional cond_resched() every small batch
> stretches the whole switch more than yielding only when somebody
> actually wants the lock or the CPU.
>
> So how about all three combined, like jbd2_log_do_checkpoint() does
> (batch cap || need_resched() || spin_needbreak()):
>
> if (!--remaining || need_resched() ||
> (moved >= MIN_LRU_BATCH && spin_is_contended(&lruvec->lru_lock)))
> return false;

Do we really need the `moved` counter? Other than that it looks good
to me. Thanks!