On 06/25/18 00:25, Ming Lei wrote:
On Sun, Jun 24, 2018 at 04:33:21PM +0000, Bart Van Assche wrote:
On Sun, 2018-06-24 at 18:16 +0800, Ming Lei wrote:
Now I am revisiting 'TAG_SHARED in restart' again for the long delay issue
of SCSI LUN probe. And found there is one bug in blk_mq_mark_tag_wait():
- hctx->dispatch_wait is added to wait queue by holding hctx->lock and
the wait queue's lock
- no hctx->lock is held when removing hctx->dispatch_wait from wait
queue.
- so the two actions(add, remove) may happen meantime since
hctx->dispatch_wait can be added to different wait queues.
Turns out this issue can be observed easily by applying the patches[2],
which is for removing 'TAG_SHARED in restart', then run simple shared-tag
null_blk test[4].
But if the hctx->lock is held in blk_mq_dispatch_wake(), as done in
patch [3], there isn't such issue at all, so it shows this issue is
related with hctx->lock, and adding/removing hctx->dispatch_wait to
wait queue. But the way of holding hctx->lock in irq context may not
be one accepted solution, since it has been avoided from the beginning
of blk-mq.
So does anyone have better ideas for this issue?
So far, follows what I thought of:
1) fix the mechanism of blk_mq_mark_tag_wait(), and removing
'TAG_SHARED in restart', then we can fix the long delay issue of
SCSI LUN probe, meantime performance can got improved, as I observed,
this way may improve IOPS by 20~30% in multi-LUN scsi_debug test.
But the issue is how to fix?
2) keep 'TAG_SHARED in restart' and let it cover the issue of
blk_mq_mark_tag_wait() as now, then try to improve 'TAG_SHARED in restart'
in another way, so that performance can be better, and synchronize_rcu()
can be removed from blk_mq_del_queue_tag_set(), then SCSI LUN probe long
delay can be fixed. I had wrote patches to do that last year. If anyone
is interested, I may post it out.
Or other ideas, any comments & ideas are welcome!
Please have a look at [PATCH] blk-mq: Fix a race condition in blk_mq_mark_tag_wait(),
16 Jan 2018 (https://www.mail-archive.com/linux-block@xxxxxxxxxxxxxxx/msg17474.html).
Thanks for sharing it, looks I miss your findings.
Your commit log describes the issue exactly, but unfortunately the patch
isn't correct, because hctx->lock isn't held in the removing path of
blk_mq_dispatch_wake(). Given 'hctx->dispatch_wait' may be added to
different wait queues, it isn't enough to hold wait queue lock and
hctx->lock in add path only. Otherwise, removing path can be seen as
'lockless' from the view point of add path.
I disagree. My patch is such that the waitqueue lock is held both around
the code that adds hctx->dispatch_wait to a waitqueue and around the
code that removes hctx->dispatch_wait from a waitqueue. No locking has
been added in blk_mq_dispatch_wake() because its caller holds the
appropriate wait queue lock.
Bart.
