Re: [PATCH] blk-mq: Avoid that I/O hangs in bt_get()

[Jens Axboe <axboe@kernel.dk>]

On 10/06/2014 06:27 AM, Bart Van Assche wrote:
> Ensure that bt_clear_tag() increments bs->wait_cnt if one or more
> threads are waiting for a tag. Remove a superfluous
> waitqueue_active() call from blk_mq_tag_wakeup_all(). This patch
> avoids that bt_get() hangs as follows if the number of hardware
> contexts is below the number of CPU threads:
> 
> INFO: task fio:6739 blocked for more than 120 seconds.
>       Not tainted 3.17.0-rc7+ #1
> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> fio             D ffff88085fcd2740     0  6739   6688 0x00000000
>  ffff8807fb96f830 0000000000000086 ffff8808350eb000 ffff8807fb96ffd8
>  0000000000012740 0000000000012740 ffff88085b885000 ffff8808350eb000
>  ffff88085fcd2fe0 ffff8807fb96f920 ffff88080061ddc8 ffffe8ffffcc4600
> Call Trace:
>  [<ffffffff8142f52d>] io_schedule+0x9d/0x130
>  [<ffffffff812016bf>] bt_get+0xef/0x180
>  [<ffffffff8107f440>] ? prepare_to_wait_event+0x110/0x110
>  [<ffffffff81201a0f>] blk_mq_get_tag+0x9f/0xd0
>  [<ffffffff811fdedb>] __blk_mq_alloc_request+0x1b/0x200
>  [<ffffffff811ff3bc>] blk_mq_map_request+0x15c/0x1b0
>  [<ffffffff8120079e>] blk_mq_make_request+0x6e/0x270
>  [<ffffffff8110a99f>] ? mempool_alloc+0x4f/0x130
>  [<ffffffff811f3af0>] generic_make_request+0xc0/0x110
>  [<ffffffff811f3bab>] submit_bio+0x6b/0x140
>  [<ffffffff81110e4b>] ? set_page_dirty_lock+0x2b/0x40
>  [<ffffffff811eea57>] ? bio_set_pages_dirty+0x47/0x60
>  [<ffffffff811907d0>] do_blockdev_direct_IO+0x2080/0x3220
>  [<ffffffff8118bac0>] ? I_BDEV+0x10/0x10
>  [<ffffffff811919bc>] __blockdev_direct_IO+0x4c/0x50
>  [<ffffffff8118bac0>] ? I_BDEV+0x10/0x10
>  [<ffffffff8118c1de>] blkdev_direct_IO+0x4e/0x50
>  [<ffffffff8118bac0>] ? I_BDEV+0x10/0x10
>  [<ffffffff81109a13>] generic_file_read_iter+0x513/0x5e0
>  [<ffffffff8119d8a7>] ? kiocb_free+0x37/0x40
>  [<ffffffff8118c660>] ? ioctl_by_bdev+0x40/0x40
>  [<ffffffff8118c697>] blkdev_read_iter+0x37/0x40
>  [<ffffffff8119e6b4>] aio_run_iocb+0x1e4/0x3c0
>  [<ffffffff8114dfa6>] ? kmem_cache_alloc+0xd6/0x3d0
>  [<ffffffff8114e045>] ? kmem_cache_alloc+0x175/0x3d0
>  [<ffffffff8119f5cc>] do_io_submit+0x11c/0x490
>  [<ffffffff8119f950>] SyS_io_submit+0x10/0x20
>  [<ffffffff81432fd2>] system_call_fastpath+0x16/0x1b
> 
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> Cc: Christoph Hellwig <hch@lst.de>
> Cc: Ming Lei <ming.lei@canonical.com>
> Cc: Robert Elliott <Elliott@hp.com>
> Cc: <stable@vger.kernel.org>
> ---
>  block/blk-mq-tag.c | 31 +++++++++++++++----------------
>  1 file changed, 15 insertions(+), 16 deletions(-)
> 
> diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
> index b5088f0..08d3b1c 100644
> --- a/block/blk-mq-tag.c
> +++ b/block/blk-mq-tag.c
> @@ -80,8 +80,7 @@ static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
>  	for (i = 0; i < BT_WAIT_QUEUES; i++) {
>  		struct bt_wait_state *bs = &bt->bs[wake_index];
>  
> -		if (waitqueue_active(&bs->wait))
> -			wake_up(&bs->wait);
> +		wake_up(&bs->wait);
>  
>  		wake_index = bt_index_inc(wake_index);
>  	}
> @@ -346,20 +345,20 @@ static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
>  	 */
>  	clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
>  
> -	bs = bt_wake_ptr(bt);
> -	if (!bs)
> -		return;
> -
> -	wait_cnt = atomic_dec_return(&bs->wait_cnt);
> -	if (wait_cnt == 0) {
> -wake:
> -		atomic_add(bt->wake_cnt, &bs->wait_cnt);
> -		bt_index_atomic_inc(&bt->wake_index);
> -		wake_up(&bs->wait);
> -	} else if (wait_cnt < 0) {
> -		wait_cnt = atomic_inc_return(&bs->wait_cnt);
> -		if (!wait_cnt)
> -			goto wake;
> +	for (;;) {
> +		bs = bt_wake_ptr(bt);
> +		if (!bs)
> +			return;
> +
> +		wait_cnt = atomic_dec_return(&bs->wait_cnt);
> +		if (unlikely(wait_cnt < 0))
> +			wait_cnt = atomic_inc_return(&bs->wait_cnt);
> +		if (wait_cnt == 0) {
> +			atomic_add(bt->wake_cnt, &bs->wait_cnt);
> +			bt_index_atomic_inc(&bt->wake_index);
> +			wake_up(&bs->wait);
> +			return;
> +		}
>  	}
>  }

I've been able to reproduce this this morning, and your patch does seem
to fix it. The inc/add logic is making my head spin a bit. And we now
end up banging a lot more on the waitqueue lock through
prepare_to_wait(), so there's a substantial performance regression to go
with the change.

I'll fiddle with this a bit and see if we can't retain existing
performance properties under tag contention, while still fixing the hang.

-- 
Jens Axboe