On 5/23/18 4:09 PM, Ming Lei wrote:

On Thu, May 24, 2018 at 1:48 AM, Omar Sandoval [off-list ref] wrote:

quoted

On Wed, May 23, 2018 at 05:32:31PM +0800, Ming Lei wrote:

quoted

On Tue, May 22, 2018 at 09:59:17PM -0600, Jens Axboe wrote:

quoted

On 5/19/18 1:44 AM, Ming Lei wrote:

quoted

When the allocation process is scheduled back and the mapped hw queue is
changed, do one extra wake up on orignal queue for compensating wake up
miss, so other allocations on the orignal queue won't be starved.

This patch fixes one request allocation hang issue, which can be
triggered easily in case of very low nr_request.

Trying to think of better ways we can fix this, but I don't see
any right now. Getting rid of the wake_up_nr() kills us on tons
of tasks waiting.

I am not sure if I understand your point, but this issue isn't related
with wake_up_nr() actually, and it can be reproduced after reverting
4e5dff41be7b5201c1c47c (blk-mq: improve heavily contended tag case).

All tasks in current sbq_wait_state may be scheduled to other CPUs, and
there may still be tasks waiting for allocation from this sbitmap_queue,
and the root cause is about cross-queue allocation, as you said,
there are too many queues, :-)

I don't follow. Your description of the problem was that we have two
waiters and only wake up one, which doesn't in turn allocate and free a
tag and wake up the second waiter. Changing it back to wake_up_nr()
eliminates that problem. And if waking up everything doesn't fix it, how
does your fix of waking up a few extra tasks fix it?

What matters is that this patch wakes up the previous sbq, let's see if
from another view:

1) still 2 hw queues, nr_requests are 2, and wake_batch is one

2) there are 3 waiters on hw queue 0

3) two in-flight requests in hw queue 0 are completed, and only two waiters
of 3 are waken up because of wake_batch, but both the two waiters can be
scheduled to another CPU and cause to switch to hw queue 1

4) then the 3rd waiter will wait for ever, since no in-flight request
is in hw queue
0 any more.

5) this patch fixes it by the fake wakeup when waiter is scheduled to another
hw queue

The issue can be understood a bit easier if we just forget sbq_wait_state and
focus on sbq, :-)

It makes sense to me, and also explains why wake_up() vs wake_up_nr() doesn't
matter. Which is actually a relief. And the condition of moving AND having
a waiter should be rare enough that it'll work out fine in practice, I don't
see any performance implications from this. You're right that we already
abort early if we don't have pending waiters, so it's all good.

Can you respin with the comments from Omar and myself covered?

-- 
Jens Axboe