blk-mq: Don't complete on a remote CPU in force threaded mode
authorSebastian Andrzej Siewior <bigeasy@linutronix.de>
Fri, 4 Dec 2020 19:13:54 +0000 (20:13 +0100)
committerJens Axboe <axboe@kernel.dk>
Thu, 17 Dec 2020 20:41:30 +0000 (13:41 -0700)
With force threaded interrupts enabled, raising softirq from an SMP
function call will always result in waking the ksoftirqd thread. This is
not optimal given that the thread runs at SCHED_OTHER priority.

Completing the request in hard IRQ-context on PREEMPT_RT (which enforces
the force threaded mode) is bad because the completion handler may
acquire sleeping locks which violate the locking context.

Disable request completing on a remote CPU in force threaded mode.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-mq.c

index a428798e1c5c6f64de056691ff57d5ca29950bbc..c338c9bc5a2c53d22331f45801a3b5dfbc299dda 100644 (file)
@@ -650,6 +650,14 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
        if (!IS_ENABLED(CONFIG_SMP) ||
            !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
                return false;
+       /*
+        * With force threaded interrupts enabled, raising softirq from an SMP
+        * function call will always result in waking the ksoftirqd thread.
+        * This is probably worse than completing the request on a different
+        * cache domain.
+        */
+       if (force_irqthreads)
+               return false;
 
        /* same CPU or cache domain?  Complete locally */
        if (cpu == rq->mq_ctx->cpu ||