Re: [PATCH v4 2/2] mm/swap: scan by cluster in find_next_to_unuse()

[Nhat Pham]

On Wed, Sep 9, 2026 at 9:16 AM Youngjun Park <youngjun.park@xxxxxxx> wrote:
>
> find_next_to_unuse() walks every offset from 0 to si->max, and swapoff
> restarts that walk on each retry, so the cost scales with the size of
> the device rather than with the few slots the shmem and mmlist passes
> could not free.  It has caused stalls before.
>
> The flat walk predates the swap table.  Slot state now lives in a per
> cluster table, and wait_for_allocation() stops all allocation before
> try_to_unuse() runs, so a cluster that holds no slot in use stays that
> way.  Skip such a cluster instead of reading all of its entries.
>
> Fill a 1 TiB swap up to some amount, then swapoff.
> What is left sits at the top of what was filled, so every slot below it
> is free.  Medians over 11 pairs at 32 and 128 GiB, 3 pairs at 256 and
> 512.
>
>         filled          swapoff
>                         old     new
>          32 GiB          92.4ms  66.3ms
>         128 GiB         157.6ms  94.4ms
>         256 GiB         209.7ms  63.8ms
>         512 GiB         391.8ms  73.4ms
>
> old grows with how much was filled, new does not.
>
> In the ordinary case swap still holds real data and swapoff spends its
> time reading it back. it is tested 4 GiB on an 8 GiB device, where the scan
> is 1.4% of try_to_unuse(), and there is no difference either way.
>
> Commit dc644a073769 ("mm: add three more cond_resched() in swapoff")
> answered those stalls with a cond_resched() every 256 offsets.  A walk
> bounded by one cluster no longer needs that counter.  The loop now runs
> at most SWAPFILE_CLUSTER times before it returns or reschedules, the
> same bound swap_reclaim_full_clusters() already scans between
> cond_resched() calls.
>
> The scan end is clamped to si->max, so the walk stops there rather than
> running into the masked tail of the last cluster.
>
> ci->count is read without ci->lock, so READ_ONCE() marks the read for
> KCSAN.  Allocation is already stopped, so the count can only drop, and a
> slot stops being counted only after its folio has left the swap cache.
> An empty cluster therefore holds nothing for try_to_unuse() to act on.
>
> Signed-off-by: Youngjun Park <youngjun.park@xxxxxxx>
> Reviewed-by: Barry Song <baohua@xxxxxxxxxx>
> Acked-by: Kairui Song <kasong@xxxxxxxxxxx>
> Reviewed-by: Baoquan He <baoquan.he@xxxxxxxxx>

Reviewed-by: Nhat Pham <nphamcs@xxxxxxxxx>