Re: [PATCH] btrfs: zoned: don't submit orphaned extent buffers

On 7/9/26 12:31 PM, Qu Wenruo wrote:
>
>
> 在 2026/7/3 15:24, Johannes Thumshirn 写道:
>> On a zoned filesystem btree_writepages() can encounter a dirty metadata
>> extent buffer whose block group no longer exists. Submitting a write for
>> such a buffer maps it to a stale/removed block-group and leaves the 
>> folio
>> under writeback forever, hanging later in filemap_fdatawait_range(), for
>> example the iput(btree_inode) in close_ctree(), which then hangs 
>> unmount.
>>
>> This is caused by btrfs_clear_buffer_dirty() not clearing the dirty 
>> bit of
>> a freed tree block but it sets EXTENT_BUFFER_ZONED_ZEROOUT and keeps the
>> buffer dirty so that it is still written out to keep the zone's
>> meta_write_pointer advancing sequentially. So a freed metadata block
>> legitimately stays dirty until that zero-write completes.
>>
>> Dropping these buffers is safe: the block group is empty, so they are
>> stale, unreferenced, already-freed blocks. Once the zone is reset their
>> zero-write is unneeded. Instead of submitting a such a write, finish the
>> writeback immediately.
>>
>> Reported-by: Shin'ichiro Kawasaki <[email protected]>
>> Fixes: 7db94301a980 ("btrfs: zoned: introduce block group context to 
>> btrfs_eb_write_context")
>> Cc: [email protected]
>> Signed-off-by: Johannes Thumshirn <[email protected]>
>> ---
>>   fs/btrfs/extent_io.c | 16 ++++++++++++++--
>>   1 file changed, 14 insertions(+), 2 deletions(-)
>>
>> diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
>> index 0edd532174fa..4a029ae719e9 100644
>> --- a/fs/btrfs/extent_io.c
>> +++ b/fs/btrfs/extent_io.c
>> @@ -2280,7 +2280,8 @@ static void prepare_eb_write(struct 
>> extent_buffer *eb)
>>   }
>>     static noinline_for_stack void write_one_eb(struct extent_buffer 
>> *eb,
>> -                        struct writeback_control *wbc)
>> +                        struct writeback_control *wbc,
>> +                        bool submit)
>>   {
>>       struct btrfs_fs_info *fs_info = eb->fs_info;
>>       struct btrfs_bio *bbio;
>> @@ -2310,6 +2311,12 @@ static noinline_for_stack void 
>> write_one_eb(struct extent_buffer *eb,
>>           wbc_account_cgroup_owner(wbc, folio, range_len);
>>           folio_unlock(folio);
>>       }
>> +
>> +    if (!submit) {
>> +        btrfs_bio_end_io(bbio, BLK_STS_OK);
>> +        return;
>> +    }
>> +
>>       /*
>>        * If the fs is already in error status, do not submit any 
>> writeback
>>        * but immediately finish it.
>> @@ -2397,6 +2404,8 @@ int btree_writepages(struct address_space 
>> *mapping, struct writeback_control *wb
>>           struct extent_buffer *eb;
>>             while ((eb = eb_batch_next(&batch)) != NULL) {
>> +            bool submit = true;
>> +
>>               ctx.eb = eb;
>>                 ret = btrfs_check_meta_write_pointer(eb->fs_info, &ctx);
>> @@ -2411,6 +2420,9 @@ int btree_writepages(struct address_space 
>> *mapping, struct writeback_control *wb
>>                   continue;
>>               }
>>   +            if (btrfs_is_zoned(fs_info) && !ctx.zoned_bg)
>> +                submit = false;
>> +
>>               if (!lock_extent_buffer_for_io(eb, wbc))
>>                   continue;
>>   @@ -2420,7 +2432,7 @@ int btree_writepages(struct address_space 
>> *mapping, struct writeback_control *wb
>>                   btrfs_schedule_zone_finish_bg(ctx.zoned_bg, eb);
>>                   ctx.zoned_bg->meta_write_pointer += eb->len;
>>               }
>> -            write_one_eb(eb, wbc);
>> +            write_one_eb(eb, wbc, submit);
>
> I understand this is the minimal fix, but I can't help but wondering, 
> would it be more instinctual to release all ebs inside a zoned 
> metadata bg when freeing the bg?

Yes but quite frankly I'm still hunting the root cause.

One thing I think I've found so far is, we're re-dirtying a eb, then 
relocation GC's the block-group, then writeback comes along trying to 
submit these ebs but the bg is gone.

This analysis could be totally bogus of cause because I missed something.