Hi,
On Wed, Aug 10, 2016 at 5:55 AM, Tejun Heo <t...@kernel.org> wrote:
> Hello,
>
> On Mon, Aug 08, 2016 at 01:39:08PM +0200, Roman Pen wrote:
>> Long time ago there was a similar fix proposed by Akinobu Mita[1],
>> but it seems that time everyone decided to fix this subtle race in
>> percpu-refcount and Tejun Heo[2] did an attempt (as I can see that
>> patchset was not applied).
>
> So, I probably forgot about it while waiting for confirmation of fix.
> Can you please verify that the patchset fixes the issue? I can apply
> the patchset right away.
I have not checked your patchset but according to my understanding
it should not fix *this* issue. What happens here is a wrong order
of invocation of percpu_ref_reinit() and percpu_ref_kill(). So what
was observed is the following:
CPU#0 CPU#1
---------------- -----------------
percpu_ref_kill()
percpu_ref_kill() << atomic reference does
percpu_ref_reinit() << not guarantee the order
blk_mq_freeze_queue_wait() !! HANG HERE
percpu_ref_reinit()
blk_mq_freeze_queue_wait() on CPU#1 expects percpu-refcount to be
switched to ATOMIC mode (killed), but that does not happen, because
CPU#2 was faster and has been switched percpu-refcount to PERCPU
mode.
This race happens inside blk-mq, because invocation of kill/reinit
is controlled by the reference counter, which does not guarantee the
order of the following functions calls (kill/reinit).
So the fix is the same as originally proposed by Akinobu Mita, but
the issue is different.
But of course I can run tests on top of your series, just to verify
that everything goes smoothly and internally percpu-refcount members
are consistent.
--
Roman
