Re: [PATCH 4/4] btrfs: zoned: serialize zone finishing per block group

[Johannes Thumshirn <johannes.thumshirn@wdc.com>]

On Mon, Aug 24, 2026 at 02:49:56PM +0800, Dongjiang Zhu wrote:
> [BUG]
> On an HC620 with max_active_zones=128, buffered writes occupied all 128
> active-zone slots.  A metadata-heavy create/delete workload followed by
> sync then forced a metadata block group switch.

Can you explain that to me. In my mental model, if all 128 active-zones are
used and you call sync, we need to add a treelog-bg. But there is no more
active zone resource available so we cannot create the treelog bg. This is why
we fail, so we need to finish a zone first to then create a treelog bg.

> 
> A kernel with additional zone finish tracing reported:
> 
>   zoned-debug: finish clear-runtime bg=140928614400 caller=btrfs_zone_finish_endio_workfn comm=kworker/u8:7
>   zoned-debug: finish early-inactive bg=140928614400 caller=btrfs_check_meta_write_pointer comm=sync
>   zoned-debug: pivot finish ret=0 old_bg=140928614400 new_bg=339839287296
>   zoned-debug: activate fail slot bg=339839287296 type=metadata left=0 max_active=128 comm=sync
>   zoned-debug: finish zone-mgmt bg=140928614400 ret=0 left=0 caller=btrfs_zone_finish_endio_workfn comm=kworker/u8:7
>   BTRFS error (device sdb): error while writing out transaction: -EAGAIN
>   BTRFS error (device sdb state A): Transaction 67 aborted (-EAGAIN)
>   BTRFS info (device sdb state EA): forced readonly
>   zoned-debug: clear-active dev=/dev/sdb pos=140928614400 left=1 max_active=128
> 
> The transaction was aborted before the old block group returned its
> active-zone slot, forcing the filesystem read-only.

Ah now I see we're on the same page.
> 
> [CAUSE]
> do_zone_finish() clears BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE before issuing
> the zone finish operation.  This allows the following interleaving:
> 
>   Thread A, endio worker              Thread B, transaction writeback
> 
>   clear ZONE_IS_ACTIVE
>   call_zone_finish(old)
>                                         observe old as inactive
>                                         return from do_zone_finish(old)
>                                         activate(new)
>                                         fail with left == 0
>   clear the device active bit
>   increment active_zones_left
> 
> Thread B treats the cleared active bit as a completed finish while thread
> A is still performing the device operation.
> 
> [FIX]
> Add BLOCK_GROUP_FLAG_ZONE_FINISHING and keep ZONE_IS_ACTIVE set until the
> zone finish operation and active-list cleanup are complete.
> 
> Concurrent finishers wait for ZONE_FINISHING to be cleared instead of
> returning early.  Wake them after clearing ZONE_IS_ACTIVE and removing
> the block group from the active list.  If a waiter acquired a temporary
> read-only reference, release it after the wait.
> 

Stupid question, wouldn't it be enough to re-order the call to
do_zone_finish() vs the clearing of  BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE?