Re: [PATCH v3] btrfs: Don't remove block group still has pinned down bytes

[Nikolay Borisov <nborisov@suse.com>]

On 20.06.2018 08:34, Qu Wenruo wrote:
> 
> 
> On 2018年06月20日 13:25, Nikolay Borisov wrote:
>>
>>
>> On 15.06.2018 04:35, Qu Wenruo wrote:
>>> [BUG]
>>> Under certain KVM load and LTP tests, we are possible to hit the
>>> following calltrace if quota is enabled:
>>> ------
>>> BTRFS critical (device vda2): unable to find logical 8820195328 length 4096
>>> BTRFS critical (device vda2): unable to find logical 8820195328 length 4096
>>> ------------[ cut here ]------------
>>> WARNING: CPU: 0 PID: 49 at ../block/blk-core.c:172 blk_status_to_errno+0x1a/0x30
>>> CPU: 0 PID: 49 Comm: kworker/u2:1 Not tainted 4.12.14-15-default #1 SLE15 (unreleased)
>>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.0.0-prebuilt.qemu-project.org 04/01/2014
>>> Workqueue: btrfs-endio-write btrfs_endio_write_helper [btrfs]
>>> task: ffff9f827b340bc0 task.stack: ffffb4f8c0304000
>>> RIP: 0010:blk_status_to_errno+0x1a/0x30
>>> Call Trace:
>>>  submit_extent_page+0x191/0x270 [btrfs]
>>>  ? btrfs_create_repair_bio+0x130/0x130 [btrfs]
>>>  __do_readpage+0x2d2/0x810 [btrfs]
>>>  ? btrfs_create_repair_bio+0x130/0x130 [btrfs]
>>>  ? run_one_async_done+0xc0/0xc0 [btrfs]
>>>  __extent_read_full_page+0xe7/0x100 [btrfs]
>>>  ? run_one_async_done+0xc0/0xc0 [btrfs]
>>>  read_extent_buffer_pages+0x1ab/0x2d0 [btrfs]
>>>  ? run_one_async_done+0xc0/0xc0 [btrfs]
>>>  btree_read_extent_buffer_pages+0x94/0xf0 [btrfs]
>>>  read_tree_block+0x31/0x60 [btrfs]
>>>  read_block_for_search.isra.35+0xf0/0x2e0 [btrfs]
>>>  btrfs_search_slot+0x46b/0xa00 [btrfs]
>>>  ? kmem_cache_alloc+0x1a8/0x510
>>>  ? btrfs_get_token_32+0x5b/0x120 [btrfs]
>>>  find_parent_nodes+0x11d/0xeb0 [btrfs]
>>>  ? leaf_space_used+0xb8/0xd0 [btrfs]
>>>  ? btrfs_leaf_free_space+0x49/0x90 [btrfs]
>>>  ? btrfs_find_all_roots_safe+0x93/0x100 [btrfs]
>>>  btrfs_find_all_roots_safe+0x93/0x100 [btrfs]
>>>  btrfs_find_all_roots+0x45/0x60 [btrfs]
>>>  btrfs_qgroup_trace_extent_post+0x20/0x40 [btrfs]
>>>  btrfs_add_delayed_data_ref+0x1a3/0x1d0 [btrfs]
>>>  btrfs_alloc_reserved_file_extent+0x38/0x40 [btrfs]
>>>  insert_reserved_file_extent.constprop.71+0x289/0x2e0 [btrfs]
>>>  btrfs_finish_ordered_io+0x2f4/0x7f0 [btrfs]
>>>  ? pick_next_task_fair+0x2cd/0x530
>>>  ? __switch_to+0x92/0x4b0
>>>  btrfs_worker_helper+0x81/0x300 [btrfs]
>>>  process_one_work+0x1da/0x3f0
>>>  worker_thread+0x2b/0x3f0
>>>  ? process_one_work+0x3f0/0x3f0
>>>  kthread+0x11a/0x130
>>>  ? kthread_create_on_node+0x40/0x40
>>>  ret_from_fork+0x35/0x40
>>> Code: 00 00 5b c3 0f 1f 00 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 40 80 ff 0c 40 0f b6 c7 77 0b 48 c1 e0 04 8b 80 00 bf c8 bd c3 <0f> 0b b8 fb ff ff ff c3 0f 1f 40 00 66 2e 0f 1f 84 00 00 00 00
>>> ---[ end trace f079fb809e7a862b ]---
>>> BTRFS critical (device vda2): unable to find logical 8820195328 length 16384
>>> BTRFS: error (device vda2) in btrfs_finish_ordered_io:3023: errno=-5 IO failure
>>> BTRFS info (device vda2): forced readonly
>>> BTRFS error (device vda2): pending csums is 2887680
>>> ------
>>>
>>> [CAUSE]
>>> It's caused by race with block group auto removal like the following
>>> case:
>>> - There is a meta block group X, which has only one tree block
>>>   The tree block belongs to fs tree 257.
>>> - In current transaction, some operation modified fs tree 257
>>>   The tree block get CoWed, so the block group X is empty, and marked as
>>>   unused, queued to be deleted.
>>> - Some workload (like fsync) wakes up cleaner_kthread()
>>>   Which will call btrfs_deleted_unused_bgs() to remove unused block
>>>   groups.
>>>   So block group X along its chunk map get removed.
>>> - Some delalloc work finished for fs tree 257
>>>   Quota needs to get the original reference of the extent, which will
>>>   reads tree blocks of commit root of 257.
>>>   Then since the chunk map get removed, above warning get triggered.
>>>
>>> [FIX]
>>> Just teach btrfs_delete_unused_bgs() to skip block group who still has
>>> pinned bytes.
>>>
>>> However there is a minor side effect, since currently we only queue
>>> empty blocks at update_block_group(), and such empty block group with
>>> pinned bytes won't go through update_block_group() again, such block
>>> group won't be removed, until it get new extent allocated and removed.
>>>
>>> But please note that, there are more problems related to extent
>>> allocator with block group auto removal.
>>>
>>> Even a block group is marked unused, extent allocator can still allocate
>>> new extents from unused block group.
>>> Thus delaying block group to next transaction won't work.
>>> (Extents get allocated in current transaction, and removed again in next
>>> transaction).
>>
>> Isn't this easily solvable by making the allocator check whether the
>> block group it has chosen is part of the unused_bgs_list ?
> 
> That's also my initial idea, however bg_cache->bg_list can also be used
> in trans->new_bgs.
> 
> Another factor is, even we check the block group of an extent in
> find_free_extent() and skip the allocation, we can lead to more frequent
> false ENOSPC.
> 
> The correct way to do it is to allow find_free_extent() to skip block
> group which is in unused_bgs_list (maybe needs extra bit to indicate
> this), if we have other block group to use.
> Or we still need to use that unused block group, and remove it from
> unused_bgs_list to avoid ENOSPC and unneeded new block group creation.
> 
> (BTW, there is a bug that find_free_extent() checked block_group->list
> wrongly)
> 
> That's the main reason sending out such small fix, as the perfect fix
> may be much complex than my expectation.

Well, if we have somewhat major deficiencies in the allocator then I'd
say in the long run those need to be fixed - properly. AFAIR we've
discussed the fact that the bg_list  is somewhat abused since it could
be linked to one of 2 lists, AFAIR it was in the context of some locking
you did around block groups. I guess the take away is that the block
group management + allocator are in need of some overhauling rather than
piling hacks on top of hacks just to make it work another day.


> 
> Thanks,
> Qu
> 
>>
>>>
>>> So the root fix need to co-operate with extent allocator.
>>> But current fix should be enough for this particular bug.
>>>
>>> Signed-off-by: Qu Wenruo <wqu@suse.com>
>>
>> LGTM,
>>
>> Reviewed-by: Nikolay Borisov <nborisov@suse.com>
>>
>>> ---
>>> changelog:
>>> v2:
>>>   Commit message update, to better indicate how pinned byte is used in
>>>   btrfs and why it's related to quota.
>>> v3:
>>>   Commit message update, further explaining the bug with an example.
>>>   And added the side effect of the fix, and possible further fix.
>>> ---
>>>  fs/btrfs/extent-tree.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
>>> index f190023386a9..7d14c4ca8232 100644
>>> --- a/fs/btrfs/extent-tree.c
>>> +++ b/fs/btrfs/extent-tree.c
>>> @@ -10675,7 +10675,7 @@ void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
>>>  		/* Don't want to race with allocators so take the groups_sem */
>>>  		down_write(&space_info->groups_sem);
>>>  		spin_lock(&block_group->lock);
>>> -		if (block_group->reserved ||
>>> +		if (block_group->reserved || block_group->pinned ||
>>>  		    btrfs_block_group_used(&block_group->item) ||
>>>  		    block_group->ro ||
>>>  		    list_is_singular(&block_group->list)) {
>>>