Re: [PATCH] bcache: allow more bucket reclaims during garbage collection

[Robert Pang]

On Mon, Nov 2, 2026 at 11:02 PM Robert Pang <robertpang@xxxxxxxxxx> wrote:
>
> In patch a14a68b76954 ("bcache: allow allocator to invalidate bucket in gc"),
> reclaimable buckets are added to the free list for the bucket allocator
> at the start of a GC cycle.
>
> However, once these buckets are allocated, they are no longer marked as
> reclaimable until the GC cycle finishes. Under heavy loads and frequent cache
> updates, all reclaimable buckets can be exhausted before the GC ends, causing
> the allocator to wait and stalling incoming I/O.
>
> Address this by allowing reclaimable buckets to be reclaimed during GC up to
> the number of generations left before BUCKET_GC_GEN_MAX.
>
> Signed-off-by: Robert Pang <robertpang@xxxxxxxxxx>
> ---
>  drivers/md/bcache/alloc.c  | 3 ++-
>  drivers/md/bcache/bcache.h | 2 +-
>  drivers/md/bcache/btree.c  | 5 +++--
>  3 files changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/md/bcache/alloc.c b/drivers/md/bcache/alloc.c
> index 3ec240ec685d..959d753e7cf6 100644
> --- a/drivers/md/bcache/alloc.c
> +++ b/drivers/md/bcache/alloc.c
> @@ -141,7 +141,8 @@ void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
>         bch_inc_gen(ca, b);
>         b->prio = INITIAL_PRIO;
>         atomic_inc(&b->pin);
> -       b->reclaimable_in_gc = 0;
> +       if (b->reclaimable_in_gc)
> +               b->reclaimable_in_gc--;
>  }
>
>  static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
> diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
> index ec9ff9715081..daae5ac5ffbc 100644
> --- a/drivers/md/bcache/bcache.h
> +++ b/drivers/md/bcache/bcache.h
> @@ -200,7 +200,7 @@ struct bucket {
>         uint8_t         gen;
>         uint8_t         last_gc; /* Most out of date gen in the btree */
>         uint16_t        gc_mark; /* Bitfield used by GC. See below for field */
> -       uint16_t        reclaimable_in_gc:1;
> +       uint8_t         reclaimable_in_gc; /* no. of times reclaimable in GC */
>  };
>
>  /*
> diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
> index 27a129d47d0a..fc167868c814 100644
> --- a/drivers/md/bcache/btree.c
> +++ b/drivers/md/bcache/btree.c
> @@ -1742,9 +1742,10 @@ static void btree_gc_start(struct cache_set *c)
>
>         ca = c->cache;
>         for_each_bucket(b, ca) {
> -               b->last_gc = b->gen;
>                 if (bch_can_invalidate_bucket(ca, b))
> -                       b->reclaimable_in_gc = 1;
> +                       b->reclaimable_in_gc =
> +                               BUCKET_GC_GEN_MAX - bucket_gc_gen(b);
> +               b->last_gc = b->gen;
>                 if (!atomic_read(&b->pin)) {
>                         SET_GC_MARK(b, 0);
>                         SET_GC_SECTORS_USED(b, 0);
> --
> 2.56.0.398.g882cbad275-goog
>

Hi all,

To provide some additional context on this patch:

  - This commit fixes a recurrence of the I/O stall issue [1]
originally addressed by commit a14a68b76954 ("bcache: allow allocator
to invalidate bucket in gc").
  - The recurrence stems from commit 009073806fbc ("bcache: reduce gc
latency by processing less nodes and sleep less time"), which
introduces longer GC pauses (100 ms) to reduce front IO tail latency
during GC.
  - Due to the longer GC runtime, commit a14a68b76954 becomes
ineffective under heavy cache updates because reclaimable buckets are
rapidly exhausted when each bucket is limited to a single reclaim.
  - This patch extends commit a14a68b76954 to allow buckets to be
reclaimed multiple times, bounded by the BUCKET_GC_GEN_MAX safety
range.

Best regards
Robert Pang

[1] https://lore.kernel.org/linux-bcache/1599736884-5479-1-git-send-email-dongsheng.yang@xxxxxxxxxxxx/