Re: [f2fs-dev] [PATCH Vx 1/1] f2fs: Avoid deadlock between writeback and checkpoint

[Chao Yu <chao@kernel.org>]

On 2021/12/18 0:59, Jaegeuk Kim wrote:
> On 12/17, Chao Yu wrote:
>> On 2021/11/20 9:50, Chao Yu wrote:
>>> On 2021/11/18 14:46, Chao Yu wrote:
>>>> On 2021/11/18 0:56, Jaegeuk Kim wrote:
>>>>> On 11/09, niuzhiguo84@gmail.com wrote:
>>>>>> From: Zhiguo Niu <zhiguo.niu@unisoc.com>
>>>>>>
>>>>>> There could be a scenario as following:
>>>>>> The inodeA and inodeB are in b_io queue of writeback
>>>>>> inodeA : f2fs's node inode
>>>>>> inodeB : a dir inode with only one dirty pages, and the node page
>>>>>> of inodeB cached into inodeA
>>>>>>
>>>>>> writeback:
>>>>>>
>>>>>> wb_workfn
>>>>>> wb_writeback
>>>>>> blk_start_plug
>>>>>>             loop {
>>>>>>             queue_io
>>>>>>             progress=__writeback_inodes_wb
>>>>>>                     __writeback_single_inode
>>>>>>                             do_writepages
>>>>>>                                     f2fs_write_data_pages
>>>>>>                                     wbc->pages_skipped +=get_dirty_pages
>>>>>>                             inode->i_state &= ~dirty
>>>>>>                     wrote++
>>>>>>                     requeue_inode
>>>>>>             }
>>>>>> blk_finish_plug
>>>>>>
>>>>>> checkpoint:
>>>>>>
>>>>>> f2fs_write_checkpoint
>>>>>> f2fs_sync_dirty_inodes
>>>>>> filemap_fdatawrite
>>>>>> do_writepages
>>>>>> f2fs_write_data_pages
>>>>>>             f2fs_write_single_data_page
>>>>>>                     f2fs_do_write_data_page
>>>>>>                             set_page_writeback
>>>>>>                             f2fs_outplace_write_data
>>>>>>                                     f2fs_update_data_blkaddr
>>>>>>                                             f2fs_wait_on_page_writeback
>>>>>>                     inode_dec_dirty_pages
>>>>>>
>>>>>> 1. Writeback thread flush inodeA, and push it's bio request in task's plug;
>>>>>> 2. Checkpoint thread writes inodeB's dirty page, and then wait its node
>>>>>>         page writeback cached into inodeA which is in writeback task's plug
>>>>>> 3. Writeback thread flush inodeB and skip writing the dirty page as
>>>>>>         wb_sync_req[DATA] > 0.
>>>>>> 4. As none of the inodeB's page is marked as PAGECACHE_TAG_DIRTY, writeback
>>>>>>         thread clear inodeB's dirty state.
>>>>>> 5. Then inodeB is moved from b_io to b_dirty because of pages_skipped > 0
>>>>>>         as checkpoint thread is stuck before dec dirty_pages.
>>>>>>
>>>>>> This patch collect correct pages_skipped according to the tag state in
>>>>>> page tree of inode
>>>>>>
>>>>>> Signed-off-by: Zhiguo Niu <zhiguo.niu@unisoc.com>
>>>>>> Signed-off-by: Jing Xia <jing.xia@unisoc.com>
>>>>>> ---
>>>>>>      fs/f2fs/data.c | 4 +++-
>>>>>>      1 file changed, 3 insertions(+), 1 deletion(-)
>>>>>>
>>>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>>>> index f4fd6c246c9a..e98628e3868c 100644
>>>>>> --- a/fs/f2fs/data.c
>>>>>> +++ b/fs/f2fs/data.c
>>>>>> @@ -3237,7 +3237,9 @@ static int __f2fs_write_data_pages(struct address_space *mapping,
>>>>>>      	return ret;
>>>>>>      skip_write:
>>>>>> -	wbc->pages_skipped += get_dirty_pages(inode);
>>>>>> +	wbc->pages_skipped +=
>>>>>> +		mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY) ?
>>>>>
>>>>> Is there any race condition to get 0, if there's any dirty page? IOWs, it
>>>>
>>>> Quoted from Jing Xia's explanation:
>>>>
>>>> [T:writeback]				[T:checkpoint]
>>>
>>> My bad, [1] should be here:
>>>
>>> bio contains NodeA was plugged in writeback threads
>>>
>>> Thanks,
>>>
>>>> 					- do_writepages  -- sync write inodeB, inc wb_sync_req[DATA]
>>>> 					 - f2fs_write_data_pages
>>>> 					  - f2fs_write_single_data_page -- write last dirty page
>>>> 					   - f2fs_do_write_data_page
>>>> 					    - set_page_writeback  -- clear page dirty flag and
>>>> 					    PAGECACHE_TAG_DIRTY tag in radix tree
>>>> 					    - f2fs_outplace_write_data
>>>> 					     - f2fs_update_data_blkaddr
>>>> 					      - f2fs_wait_on_page_writeback -- wait NodeA to writeback here
>>
>> Jaegeuk,
>>
>> checkpoint() will stuck on waiting NodeA to writeback.
>>
>>>> 					   - inode_dec_dirty_pages
>>>
>>>> bio contains NodeA was plugged in writeback threads
>>>
>>> [1]
>>>
>>> Thanks,
>>>
>>>> - writeback_sb_inodes
>>>>      - writeback_single_inode
>>>>       - do_writepages
>>>>        - f2fs_write_data_pages -- skip writepages due to wb_sync_req[DATA]
>>>>         - wbc->pages_skipped += get_dirty_pages() -- PAGECACHE_TAG_DIRTY is not set but get_dirty_pages() returns one
>>>>      - requeue_inode -- requeue inode to wb->b_dirty queue due to non-zero.pages_skipped
>>
>> Meanwhile, kworker plugged NodeA, and then loop writebacking inode due to
>> inode will be requeued all the time.
> 
> I'm thinking that this will be released after finishing node writes in a
> different context. Then, this loop can happen in that time period which
> is NOT forever. But, if we want to avoid this temporary loop, we may be

I doubt it may break due to below condition in writeback_sb_inodes():

		if (need_resched()) {
			/*
			 * We're trying to balance between building up a nice
			 * long list of IOs to improve our merge rate, and
			 * getting those IOs out quickly for anyone throttling
			 * in balance_dirty_pages().  cond_resched() doesn't
			 * unplug, so get our IOs out the door before we
			 * give up the CPU.
			 */
			if (current->plug)
				blk_flush_plug(current->plug, false);
			cond_resched();
		}

But as Zhiguo Niu and Jing Xia reported, blk_flush_plug was not executed...

> able to decrease the dirty count before f2fs_update_data_blkaddr().

I guess there is one simple way, we can force flushing plugged bio via
blk_flush_plug(current->plug, false) once we skip writeback in writepages()
due to sbi->wb_sync_req[DATA/NODE] is valid.

Thoughts?

Thanks,

> 
>>
>> Thanks,
>>
>>>>
>>>>> seems the current condition is just requeuing the inode as dirty, but next
>>>>> flushing time will remove it from dirty list. Is this giving too much overheads?
>>>>
>>>> I prefer to let writeback thread call blk_flush_plug() after skipping
>>>> writepages() due to wb_sync_req[DATA/NODE] check condition, thoughts?
>>>>
>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>> index 9f754aaef558..b6e1ed73f8f5 100644
>>>> --- a/fs/f2fs/data.c
>>>> +++ b/fs/f2fs/data.c
>>>> @@ -3087,6 +3087,8 @@ static int f2fs_write_cache_pages(struct address_space *mapping,
>>>>      			/* give a priority to WB_SYNC threads */
>>>>      			if (atomic_read(&sbi->wb_sync_req[DATA]) &&
>>>>      					wbc->sync_mode == WB_SYNC_NONE) {
>>>> +				if (current->plug)
>>>> +					blk_flush_plug(current->plug, false);
>>>>      				done = 1;
>>>>      				break;
>>>>      			}
>>>> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
>>>> index 556fcd8457f3..dd9a817d8dab 100644
>>>> --- a/fs/f2fs/node.c
>>>> +++ b/fs/f2fs/node.c
>>>> @@ -1946,6 +1946,8 @@ int f2fs_sync_node_pages(struct f2fs_sb_info *sbi,
>>>>      			if (atomic_read(&sbi->wb_sync_req[NODE]) &&
>>>>      					wbc->sync_mode == WB_SYNC_NONE) {
>>>>      				done = 1;
>>>> +				if (current->plug)
>>>> +					blk_flush_plug(current->plug, false);
>>>>      				break;
>>>>      			}
>>>>
>>>>
>>>>
>>>> Thanks,
>>>>
>>>>>
>>>>>> +		get_dirty_pages(inode) : 0;
>>>>>>      	trace_f2fs_writepages(mapping->host, wbc, DATA);
>>>>>>      	return 0;
>>>>>>      }
>>>>>> -- 
>>>>>> 2.28.0
>>>>
>>>>
>>>> _______________________________________________
>>>> Linux-f2fs-devel mailing list
>>>> Linux-f2fs-devel@lists.sourceforge.net
>>>> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
>>>>
>>>
>>>
>>> _______________________________________________
>>> Linux-f2fs-devel mailing list
>>> Linux-f2fs-devel@lists.sourceforge.net
>>> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel