cfq: fix IOPRIO_CLASS_IDLE delays
PATCH 1/1 cfq: fix IOPRIO_CLASS_IDLE delays
After the fresh boot:
ionice -c3 -p $$ echo cfq >> /sys/block/XXX/queue/scheduler dd if=/dev/XXX of=/dev/null bs=512 count=1
Now dd hangs in D state and the queue is completely stalled for approximately INITIAL_JIFFIES + CFQ_IDLE_GRACE jiffies. This is because cfq_init_queue() forgets to initialize cfq_data->last_end_request.
(I guess this patch is not complete, overflow is still possible)
block/cfq-iosched.c renamed from block/cfq-iosched.c~ +1 -0
@@ -2122,6 +2122,7 @@ static void *cfq_init_queue(struct requeINIT_WORK (&cfqd->unplug_work, cfq_kick_queue); + cfqd->last_end_request = jiffies; cfqd->cfq_quantum = cfq_quantum; cfqd->cfq_fifo_expire[0 ] = cfq_fifo_expire[0 ]; cfqd->cfq_fifo_expire[1 ] = cfq_fifo_expire[1 ];
Quoted 9 lines · Oleg Nesterov
After the fresh boot:
ionice -c3 -p $$ echo cfq >> /sys/block/XXX/queue/scheduler dd if=/dev/XXX of=/dev/null bs=512 count=1Now dd hangs in D state and the queue is completely stalled for approximately
INITIAL_JIFFIES+CFQ_IDLE_GRACEjiffies. This is becausecfq_init_queue()forgets to initializecfq_data->last_end_request.
Another good catch, thanks Oleg!
Quoted 1 line
(I guess this patch is not complete, overflow is still possible)
Hmm, where would it overflow?
Signature · 1 line
Jens Axboe
Quoted 6 lines · Jens Axboe
Quoted 2 lines · Oleg Nesterov
(I guess this patch is not complete, overflow is still possible)
Hmm, where would it overflow?
What if the queue was idle long enough (no !CLASS_IDLE requests) so that jiffies wraps into the past wrt ->last_end_request?
IOW, perhaps something like the patch below makes some sense. Of course, this is only a theoretical problem, I'm not sure we really need a fix.
Oleg.
[PATCH] cfq_idle_class_timer: add paranoid checks for jiffies overflow
In theory, if the queue was idle long enough, cfq_idle_class_timer may have a false (and very long) timeout because jiffies can wrap into the past wrt ->last_end_request.
block/cfq-iosched.c renamed from block/cfq-iosched.c~ +17 -11
@@ -789,6 +789,20 @@ static inline void cfq_slice_expired(str__cfq_slice_expired (cfqd, cfqq, timed_out); } +static int start_idle_class_timer (struct cfq_data *cfqd) +{ +unsigned long end = cfqd->last_end_request +CFQ_IDLE_GRACE ; +unsigned long now = jiffies; + +if (time_before (now, end) && +time_after_eq (now, cfqd->last_end_request)) { +mod_timer (&cfqd->idle_class_timer, end); +return 1 ; + } + +return 0 ; +} +/* * Get next queue for service. Unless we have a queue preemption, * we'll simply select the first cfqq in the service tree.
@@ -805,19 +819,14 @@ static struct cfq_queue *cfq_get_next_qu cfqq = rb_entry(n, struct cfq_queue, rb_node);if (cfq_class_idle (cfqq)) { -unsigned long end; -/* * if we have idle queues and no rt or be queues had * pending requests, either allow immediate service if * the grace period has passed or arm the idle grace * timer */ - end = cfqd->last_end_request +CFQ_IDLE_GRACE ; -if (time_before (jiffies, end)) { -mod_timer (&cfqd->idle_class_timer, end); +if (start_idle_class_timer (cfqd)) cfqq =NULL ; - } }return cfqq;
@@ -2033,17 +2042,14 @@ out_cont:static void cfq_idle_class_timer (unsigned long data) {struct cfq_data *cfqd = (struct cfq_data *) data; -unsigned long flags, end; +unsigned long flags;spin_lock_irqsave (cfqd->queue->queue_lock, flags);/* * race with a non-idle queue, reset timer */ - end = cfqd->last_end_request +CFQ_IDLE_GRACE ; -if (!time_after_eq (jiffies, end)) -mod_timer (&cfqd->idle_class_timer, end); -else +if (!start_idle_class_timer (cfqd))cfq_schedule_dispatch (cfqd);spin_unlock_irqrestore (cfqd->queue->queue_lock, flags);
Quoted 6 lines · Oleg Nesterov
Quoted 5 lines · Jens Axboe
Quoted 2 lines · Oleg Nesterov
(I guess this patch is not complete, overflow is still possible)
Quoted 7 lines
Quoted 1 line
Hmm, where would it overflow?
What if the queue was idle long enough (no !CLASS_IDLE requests) so that jiffies wraps into the past wrt ->last_end_request?
IOW, perhaps something like the patch below makes some sense. Of course, this is only a theoretical problem, I'm not sure we really need a fix.
OK I see what you mean, probably pretty much a theoretical problem only. But that doesn't mean it should be fixed, I've applied your patch. It also collects the two paths enabling the idle class timer, so I consider it a cleanup as well as it improves readability.
Signature · 1 line
Jens Axboe