On 5/10/21 3:10 AM, Davidlohr Bueso wrote:
On 2021-05-08 12:23, Manfred Spraul wrote:
Hi Varad,
On 5/7/21 3:38 PM, Varad Gautam wrote:
@@ -1005,11 +1022,9 @@ static inline void __pipelined_op(struct wake_q_head *wake_q, struct ext_wait_queue *this) { list_del(&this->list); - get_task_struct(this->task);
+ wake_q_add(wake_q, this->task); /* see MQ_BARRIER for purpose/pairing */ smp_store_release(&this->state, STATE_READY); - wake_q_add_safe(wake_q, this->task); } /* pipelined_send() - send a message directly to the task waiting in
First, I was too fast: I had assumed that wake_q_add() before smp_store_release() would be a potential lost wakeup.
Yeah you need wake_up_q() to actually wake anything up.
As __pipelined_op() is called within spin_lock(&info->lock), and as wq_sleep() will reread this->state after acquiring spin_lock(&info->lock), I do not see a bug anymore.
Right, and when I proposed this version of the fix I was mostly focusing on STATE_READY being set as the last operation, but the fact of the matter is we had moved to the wake_q_add_safe() version for two reasons:
(1) Ensuring the ->state = STATE_READY is done after the reference count and avoid racing with exit. In mqueue's original use of wake_q we were relying on the call's implied barrier from wake_q_add() in order to avoid reordering of setting the state. But this turned out to be insufficient hence the explicit smp_store_release().
(2) In order to prevent a potential lost wakeup when the blocked task is already queued for wakeup by another task (the failed cmpxchg case in wake_q_add), and therefore we need to set the return condition (->state = STATE_READY) before adding the task to the wake_q.
But I'm not seeing how race (2) can happen in mqueue. The race was always theoretical to begin with, with the exception of rwsems[1] in which actually the wakee task could end up in the waker's wake_q without actually blocking.
So all in all I now agree that we should keep the order of how we currently have things, just to be on the safer side, if nothing else.
Considering that moving the wake_q_add call in v2 / v3 has the potential to cause lost wakeups which has shown up in other cases, I would argue for merging the approach from v1 as the path of least surprises in favor of first factoring out the race [1]. I will resurrect it for a v4.
[1] https://lore.kernel.org/lkml/20210504155534.17270-1-varad.gautam@suse.com/
Thanks, Varad
[1] https://lore.kernel.org/lkml/1543495830-2644-1-git-send-email-xieyongji@baid...
Thanks, Davidlohr