Currently, do_async_reclaim_metadata_space cycles through all flush
states even if the filesystem has been aborted. In contrast, the data
reclaim path (do_async_reclaim_data_space) explicitly checks for
BTRFS_FS_ERROR and fails all pending tickets immediately.

This inconsistency causes the metadata reclaimer to waste CPU cycles
performing useless flush operations (like attempting to commit a
transaction or allocate chunks) on a broken filesystem.

Fix this by adding a BTRFS_FS_ERROR check inside the metadata reclaim
loop, ensuring that we fail all tickets and exit as soon as a
filesystem error is detected, matching the behavior of data reclaim.

Signed-off-by: Jiasheng Jiang <jiashengjiangcool@gmail.com>
---
 fs/btrfs/space-info.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/fs/btrfs/space-info.c b/fs/btrfs/space-info.c
index 6babbe333741..b3aae44a1436 100644
--- a/fs/btrfs/space-info.c
+++ b/fs/btrfs/space-info.c
@@ -1189,6 +1189,14 @@ static void do_async_reclaim_metadata_space(struct btrfs_space_info *space_info)
 			spin_unlock(&space_info->lock);
 			return;
 		}
+
+		if (unlikely(BTRFS_FS_ERROR(fs_info))) {
+			maybe_fail_all_tickets(space_info);
+			space_info->flush = false;
+			spin_unlock(&space_info->lock);
+			return;
+		}
+
 		to_reclaim = btrfs_calc_reclaim_metadata_size(space_info);
 		if (last_tickets_id == space_info->tickets_id) {
 			flush_state++;
-- 
2.25.1

On Sat, Jan 10, 2026 at 8:07 PM Jiasheng Jiang
<jiashengjiangcool@gmail.com> wrote:
>
> Currently, do_async_reclaim_metadata_space cycles through all flush
> states even if the filesystem has been aborted. In contrast, the data
> reclaim path (do_async_reclaim_data_space) explicitly checks for
> BTRFS_FS_ERROR and fails all pending tickets immediately.
>
> This inconsistency causes the metadata reclaimer to waste CPU cycles
> performing useless flush operations (like attempting to commit a
> transaction or allocate chunks) on a broken filesystem.

No it doesn't.
If the fs is aborted we can't commit a transaction, etc - all flush
operations return early.

>
> Fix this by adding a BTRFS_FS_ERROR check inside the metadata reclaim
> loop, ensuring that we fail all tickets and exit as soon as a
> filesystem error is detected, matching the behavior of data reclaim.

Honestly I don't think it's worth adding this code.
Not only is flushing a no-op when the fs was aborted, having a
transaction abort is an excepcional and rare event.

Thanks.

>
> Signed-off-by: Jiasheng Jiang <jiashengjiangcool@gmail.com>
> ---
>  fs/btrfs/space-info.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
>
> diff --git a/fs/btrfs/space-info.c b/fs/btrfs/space-info.c
> index 6babbe333741..b3aae44a1436 100644
> --- a/fs/btrfs/space-info.c
> +++ b/fs/btrfs/space-info.c
> @@ -1189,6 +1189,14 @@ static void do_async_reclaim_metadata_space(struct btrfs_space_info *space_info)
>                         spin_unlock(&space_info->lock);
>                         return;
>                 }
> +
> +               if (unlikely(BTRFS_FS_ERROR(fs_info))) {
> +                       maybe_fail_all_tickets(space_info);
> +                       space_info->flush = false;
> +                       spin_unlock(&space_info->lock);
> +                       return;
> +               }
> +
>                 to_reclaim = btrfs_calc_reclaim_metadata_size(space_info);
>                 if (last_tickets_id == space_info->tickets_id) {
>                         flush_state++;
> --
> 2.25.1
>
>