[PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Sebastian Andrzej Siewior]

blk_mq_complete_request_remote() will dispatch request completion to
another CPU via IPI if the CPU belongs to a different cache domain.

This breaks on PREEMPT_RT because the IPI function will complete the
request in IRQ context which includes acquiring spinlock_t typed locks.
Completing the IPI request in softirq on the remote CPU is probably less
efficient because it would require to wake ksoftirqd for this task
(which runs at SCHED_OTHER).

Ignoring the IPI request and completing the request locally is probably
the best option. It be completed either in the IRQ-thread or at the end
of the routine in softirq context.

Let blk_mq_complete_need_ipi() return that there is no need for IPI on
PREEMPT_RT.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 block/blk-mq.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index e37aa31332b70..99d2fb51e0e84 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -647,7 +647,7 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 {
 	int cpu = raw_smp_processor_id();
 
-	if (!IS_ENABLED(CONFIG_SMP) ||
+	if (!IS_ENABLED(CONFIG_SMP) || IS_ENABLED(CONFIG_PREEMPT_RT) ||
 	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
 		return false;
 
-- 
2.28.0

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Christoph Hellwig]

> -	if (!IS_ENABLED(CONFIG_SMP) ||
> +	if (!IS_ENABLED(CONFIG_SMP) || IS_ENABLED(CONFIG_PREEMPT_RT) ||
>  	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))

This needs a big fat comment explaining your rationale.  And probably
a separate if statement to make it obvious as well.

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Sebastian Andrzej Siewior]

On 2020-10-23 12:21:30 [+0100], Christoph Hellwig wrote:
> > -	if (!IS_ENABLED(CONFIG_SMP) ||
> > +	if (!IS_ENABLED(CONFIG_SMP) || IS_ENABLED(CONFIG_PREEMPT_RT) ||
> >  	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
> 
> This needs a big fat comment explaining your rationale.  And probably
> a separate if statement to make it obvious as well.

Okay.
How much difference does it make between completing in-softirq vs
in-IPI? I'm asking because acquiring a spinlock_t in an IPI shouldn't be
done (as per Documentation/locking/locktypes.rst). We don't have
anything in lockdep that will complain here on !RT and we the above we
avoid the case on RT.

Sebastian

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Christoph Hellwig]

On Fri, Oct 23, 2020 at 03:52:19PM +0200, Sebastian Andrzej Siewior wrote:
> On 2020-10-23 12:21:30 [+0100], Christoph Hellwig wrote:
> > > -	if (!IS_ENABLED(CONFIG_SMP) ||
> > > +	if (!IS_ENABLED(CONFIG_SMP) || IS_ENABLED(CONFIG_PREEMPT_RT) ||
> > >  	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
> > 
> > This needs a big fat comment explaining your rationale.  And probably
> > a separate if statement to make it obvious as well.
> 
> Okay.
> How much difference does it make between completing in-softirq vs
> in-IPI?

For normal non-RT builds?  This introduces another context switch, which
for the latencies we are aiming for is noticable.

> I'm asking because acquiring a spinlock_t in an IPI shouldn't be
> done (as per Documentation/locking/locktypes.rst). We don't have
> anything in lockdep that will complain here on !RT and we the above we
> avoid the case on RT.

At least for NVMe we aren't taking locks, but with the number of drivers

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Sebastian Andrzej Siewior]

On 2020-10-27 09:26:06 [+0000], Christoph Hellwig wrote:
> On Fri, Oct 23, 2020 at 03:52:19PM +0200, Sebastian Andrzej Siewior wrote:
> > On 2020-10-23 12:21:30 [+0100], Christoph Hellwig wrote:
> > > > -	if (!IS_ENABLED(CONFIG_SMP) ||
> > > > +	if (!IS_ENABLED(CONFIG_SMP) || IS_ENABLED(CONFIG_PREEMPT_RT) ||
> > > >  	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
> > > 
> > > This needs a big fat comment explaining your rationale.  And probably
> > > a separate if statement to make it obvious as well.
> > 
> > Okay.
> > How much difference does it make between completing in-softirq vs
> > in-IPI?
> 
> For normal non-RT builds?  This introduces another context switch, which
> for the latencies we are aiming for is noticable.

There should be no context switch. The pending softirq should be
executed on irq_exit() from that IPI, that is
  irq_exit()
  -> __irq_exit_rcu()
    -> invoke_softirq()
      -> __do_softirq() || do_softirq_own_stack() 

unlike with the command line switch `threadirqs' enabled,
invoke_softirq() woukd wakeup the `ksoftirqd' thread which would involve
a context switch.

> > I'm asking because acquiring a spinlock_t in an IPI shouldn't be
> > done (as per Documentation/locking/locktypes.rst). We don't have
> > anything in lockdep that will complain here on !RT and we the above we
> > avoid the case on RT.
> 
> At least for NVMe we aren't taking locks, but with the number of drivers

Right. I found this David Runge's log:

|BUG: scheduling while atomic: swapper/19/0/0x00010002
|CPU: 19 PID: 0 Comm: swapper/19 Not tainted 5.9.1-rt18-1-rt #1
|Hardware name: System manufacturer System Product Name/Pro WS X570-ACE, BIOS 1302 01/20/2020
|Call Trace:
| <IRQ>
| dump_stack+0x6b/0x88
| __schedule_bug.cold+0x89/0x97
| __schedule+0x6a4/0xa10
| preempt_schedule_lock+0x23/0x40
| rt_spin_lock_slowlock_locked+0x117/0x2c0
| rt_spin_lock_slowlock+0x58/0x80
| rt_spin_lock+0x2a/0x40
| test_clear_page_writeback+0xcd/0x310
| end_page_writeback+0x43/0x70
| end_bio_extent_buffer_writepage+0xb2/0x100 [btrfs]
| btrfs_end_bio+0x83/0x140 [btrfs]
| clone_endio+0x84/0x1f0 [dm_mod]
| blk_update_request+0x254/0x470
| blk_mq_end_request+0x1c/0x130
| flush_smp_call_function_queue+0xd5/0x1a0
| __sysvec_call_function_single+0x36/0x150
| asm_call_irq_on_stack+0x12/0x20
| </IRQ>

so the NVME driver isn't taking any locks but lock_page_memcg() (and
xa_lock_irqsave()) in test_clear_page_writeback() is.

Sebastian

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Christoph Hellwig]

On Tue, Oct 27, 2020 at 11:11:02AM +0100, Sebastian Andrzej Siewior wrote:
> Right. I found this David Runge's log:

True, ->bi_end_io instances can do a lot of things as long as they
are hardirq safe.

And in the end the IPI case isn't the super fast path anyway, as it
means we don't use a queue per CPU.

Is there a way to raise a softirq and preferably place it on a given
CPU without our IPI dance?  That should be a win-win situation for
everyone.

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Thomas Gleixner]

On Tue, Oct 27 2020 at 16:07, Christoph Hellwig wrote:
> On Tue, Oct 27, 2020 at 11:11:02AM +0100, Sebastian Andrzej Siewior wrote:
>> Right. I found this David Runge's log:
>
> True, ->bi_end_io instances can do a lot of things as long as they
> are hardirq safe.
>
> And in the end the IPI case isn't the super fast path anyway, as it
> means we don't use a queue per CPU.
>
> Is there a way to raise a softirq and preferably place it on a given
> CPU without our IPI dance?  That should be a win-win situation for
> everyone.

Not really. Softirq pending bits are strictly per cpu and we don't have
locking or atomics to set them remotely. Even if we had that, then you'd
still need a mechanism to make sure that the remote CPU actually
processes them. So you'd still need an IPI of some sorts.

Thanks,

        tglx

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Christoph Hellwig]

On Tue, Oct 27, 2020 at 06:05:15PM +0100, Thomas Gleixner wrote:
> > Is there a way to raise a softirq and preferably place it on a given
> > CPU without our IPI dance?  That should be a win-win situation for
> > everyone.
> 
> Not really. Softirq pending bits are strictly per cpu and we don't have
> locking or atomics to set them remotely. Even if we had that, then you'd
> still need a mechanism to make sure that the remote CPU actually
> processes them. So you'd still need an IPI of some sorts.

Ok.  I was hoping we could hide this in core code somehow, especially
a peterz didn't like the use of smp_call_function_single_async in the
blk-mq completion code very much.

Sebastian, would this solve your preempt-rt and lockdep issues?


diff --git a/block/blk-mq.c b/block/blk-mq.c
index cdced4aca2e812..5c125fb11b5691 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -626,19 +626,7 @@ static void __blk_mq_complete_request_remote(void *data)
 {
 	struct request *rq = data;
 
-	/*
-	 * For most of single queue controllers, there is only one irq vector
-	 * for handling I/O completion, and the only irq's affinity is set
-	 * to all possible CPUs.  On most of ARCHs, this affinity means the irq
-	 * is handled on one specific CPU.
-	 *
-	 * So complete I/O requests in softirq context in case of single queue
-	 * devices to avoid degrading I/O performance due to irqsoff latency.
-	 */
-	if (rq->q->nr_hw_queues == 1)
-		blk_mq_trigger_softirq(rq);
-	else
-		rq->q->mq_ops->complete(rq);
+	blk_mq_trigger_softirq(rq);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Sebastian Andrzej Siewior]

On 2020-10-27 17:23:09 [+0000], Christoph Hellwig wrote:
> On Tue, Oct 27, 2020 at 06:05:15PM +0100, Thomas Gleixner wrote:
> > > Is there a way to raise a softirq and preferably place it on a given
> > > CPU without our IPI dance?  That should be a win-win situation for
> > > everyone.
> > 
> > Not really. Softirq pending bits are strictly per cpu and we don't have
> > locking or atomics to set them remotely. Even if we had that, then you'd
> > still need a mechanism to make sure that the remote CPU actually
> > processes them. So you'd still need an IPI of some sorts.
> 
> Ok.  I was hoping we could hide this in core code somehow, especially
> a peterz didn't like the use of smp_call_function_single_async in the
> blk-mq completion code very much.
> 
> Sebastian, would this solve your preempt-rt and lockdep issues?

second. I'm cooking something.

Sebastian

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Sebastian Andrzej Siewior]

On 2020-10-27 17:23:09 [+0000], Christoph Hellwig wrote:
> Ok.  I was hoping we could hide this in core code somehow, especially
> a peterz didn't like the use of smp_call_function_single_async in the
> blk-mq completion code very much.

No idea how you could efficiently avoid smp_call_function_single_async():
- workqueue (okay)
- a timer_list timer which expires now. More code plus it may delay up
  to HZ.

> Sebastian, would this solve your preempt-rt and lockdep issues?

the problem with that is that on RT/force-threaded it will always wake
ksoftirqd thread and complete there. That are extra steps which should
be probably avoided.

Now.
The hunk in blk_mq_complete_need_ipi() will avoid the waking a thread
with force-threaded enabled.

The remaining part is a switch to llist which avoids locking (IRQ
off/on) and it allows invoke the IPI/raise softirq only if something was
added. The entries are now processed in the reverse order but this
shouldn't matter right?

I would split this into two patches (the blk_mq_complete_need_ipi() hunk
and the llist part) unless there are objections.

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 55bcee5dc0320..d2452ee9b0e2c 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -41,7 +41,7 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"
 
-static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
+static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
 
 static void blk_mq_poll_stats_start(struct request_queue *q);
 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
@@ -565,80 +565,31 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
 }
 EXPORT_SYMBOL(blk_mq_end_request);
 
-/*
- * Softirq action handler - move entries to local list and loop over them
- * while passing them to the queue registered handler.
- */
-static __latent_entropy void blk_done_softirq(struct softirq_action *h)
+static void blk_complete_reqs(struct llist_head *cpu_list)
 {
-	struct list_head *cpu_list, local_list;
+	struct llist_node *entry;
+	struct request *rq, *rq_next;
 
-	local_irq_disable();
-	cpu_list = this_cpu_ptr(&blk_cpu_done);
-	list_replace_init(cpu_list, &local_list);
-	local_irq_enable();
+	entry = llist_del_all(cpu_list);
 
-	while (!list_empty(&local_list)) {
-		struct request *rq;
-
-		rq = list_entry(local_list.next, struct request, ipi_list);
-		list_del_init(&rq->ipi_list);
+	llist_for_each_entry_safe(rq, rq_next, entry, ipi_list)
 		rq->q->mq_ops->complete(rq);
-	}
 }
 
-static void blk_mq_trigger_softirq(struct request *rq)
+static __latent_entropy void blk_done_softirq(struct softirq_action *h)
 {
-	struct list_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-	list = this_cpu_ptr(&blk_cpu_done);
-	list_add_tail(&rq->ipi_list, list);
-
-	/*
-	 * If the list only contains our just added request, signal a raise of
-	 * the softirq.  If there are already entries there, someone already
-	 * raised the irq but it hasn't run yet.
-	 */
-	if (list->next == &rq->ipi_list)
-		raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_restore(flags);
+	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
 }
 
 static int blk_softirq_cpu_dead(unsigned int cpu)
 {
-	/*
-	 * If a CPU goes away, splice its entries to the current CPU
-	 * and trigger a run of the softirq
-	 */
-	local_irq_disable();
-	list_splice_init(&per_cpu(blk_cpu_done, cpu),
-			 this_cpu_ptr(&blk_cpu_done));
-	raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_enable();
-
+	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
 	return 0;
 }
 
-
 static void __blk_mq_complete_request_remote(void *data)
 {
-	struct request *rq = data;
-
-	/*
-	 * For most of single queue controllers, there is only one irq vector
-	 * for handling I/O completion, and the only irq's affinity is set
-	 * to all possible CPUs.  On most of ARCHs, this affinity means the irq
-	 * is handled on one specific CPU.
-	 *
-	 * So complete I/O requests in softirq context in case of single queue
-	 * devices to avoid degrading I/O performance due to irqsoff latency.
-	 */
-	if (rq->q->nr_hw_queues == 1)
-		blk_mq_trigger_softirq(rq);
-	else
-		rq->q->mq_ops->complete(rq);
+	__raise_softirq_irqoff(BLOCK_SOFTIRQ);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)
@@ -648,6 +599,14 @@ 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 ||
@@ -661,6 +620,7 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 
 bool blk_mq_complete_request_remote(struct request *rq)
 {
+	struct llist_head *cpu_list;
 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 
 	/*
@@ -671,14 +631,21 @@ bool blk_mq_complete_request_remote(struct request *rq)
 		return false;
 
 	if (blk_mq_complete_need_ipi(rq)) {
-		rq->csd.func = __blk_mq_complete_request_remote;
-		rq->csd.info = rq;
-		rq->csd.flags = 0;
-		smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
+		unsigned int cpu;
+
+		cpu = rq->mq_ctx->cpu;
+		cpu_list = &per_cpu(blk_cpu_done, cpu);
+		if (llist_add(&rq->ipi_list, cpu_list)) {
+			rq->csd.func = __blk_mq_complete_request_remote;
+			rq->csd.flags = 0;
+			smp_call_function_single_async(cpu, &rq->csd);
+		}
 	} else {
 		if (rq->q->nr_hw_queues > 1)
 			return false;
-		blk_mq_trigger_softirq(rq);
+		cpu_list = this_cpu_ptr(&blk_cpu_done);
+		if (llist_add(&rq->ipi_list, cpu_list))
+			raise_softirq(BLOCK_SOFTIRQ);
 	}
 
 	return true;
@@ -3909,7 +3876,7 @@ static int __init blk_mq_init(void)
 	int i;
 
 	for_each_possible_cpu(i)
-		INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
+		init_llist_head(&per_cpu(blk_cpu_done, i));
 	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
 
 	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 639cae2c158b5..331b2b675b417 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -156,7 +156,7 @@ struct request {
 	 */
 	union {
 		struct hlist_node hash;	/* merge hash */
-		struct list_head ipi_list;
+		struct llist_node ipi_list;
 	};
 
 	/*

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Christoph Hellwig]

> The remaining part is a switch to llist which avoids locking (IRQ
> off/on) and it allows invoke the IPI/raise softirq only if something was
> added. The entries are now processed in the reverse order but this
> shouldn't matter right?

For correctness it should not matter, but I think it could have
performance implications.  I think you'll have to throw in a
llist_reverse_order.

> I would split this into two patches (the blk_mq_complete_need_ipi() hunk
> and the llist part) unless there are objections.

Yes, please do.

[PATCH 1/3] blk-mq: Don't complete on a remote CPU in force threaded mode

[Sebastian Andrzej Siewior]

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>
---
 block/blk-mq.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 55bcee5dc0320..421a40968c9ff 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -648,6 +648,14 @@ 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 ||
-- 
2.28.0

[PATCH 2/3] blk-mq: Always complete remote completions requests in softirq

[Sebastian Andrzej Siewior]

Controllers with multiple queues have their IRQ-handelers pinned to a
CPU. The core shouldn't need to complete the request on a remote CPU.

Remove this case and always raise the softirq to complete the request.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 block/blk-mq.c | 14 +-------------
 1 file changed, 1 insertion(+), 13 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 421a40968c9ff..769d2d532a825 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -626,19 +626,7 @@ static void __blk_mq_complete_request_remote(void *data)
 {
 	struct request *rq = data;
 
-	/*
-	 * For most of single queue controllers, there is only one irq vector
-	 * for handling I/O completion, and the only irq's affinity is set
-	 * to all possible CPUs.  On most of ARCHs, this affinity means the irq
-	 * is handled on one specific CPU.
-	 *
-	 * So complete I/O requests in softirq context in case of single queue
-	 * devices to avoid degrading I/O performance due to irqsoff latency.
-	 */
-	if (rq->q->nr_hw_queues == 1)
-		blk_mq_trigger_softirq(rq);
-	else
-		rq->q->mq_ops->complete(rq);
+	blk_mq_trigger_softirq(rq);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)
-- 
2.28.0

[PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

With llist_head it is possible to avoid the locking (the irq-off region)
when items are added. This makes it possible to add items on a remote
CPU.
llist_add() returns true if the list was previously empty. This can be
used to invoke the SMP function call / raise sofirq only if the first
item was added (otherwise it is already pending).
This simplifies the code a little and reduces the IRQ-off regions. With
this change it possible to reduce the SMP-function call a simple
__raise_softirq_irqoff().

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 block/blk-mq.c         | 78 ++++++++++++++----------------------------
 include/linux/blkdev.h |  2 +-
 2 files changed, 26 insertions(+), 54 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 769d2d532a825..4f53de48e5038 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -41,7 +41,7 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"
 
-static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
+static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
 
 static void blk_mq_poll_stats_start(struct request_queue *q);
 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
@@ -565,68 +565,32 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
 }
 EXPORT_SYMBOL(blk_mq_end_request);
 
-/*
- * Softirq action handler - move entries to local list and loop over them
- * while passing them to the queue registered handler.
- */
-static __latent_entropy void blk_done_softirq(struct softirq_action *h)
+static void blk_complete_reqs(struct llist_head *cpu_list)
 {
-	struct list_head *cpu_list, local_list;
+	struct llist_node *entry;
+	struct request *rq, *rq_next;
 
-	local_irq_disable();
-	cpu_list = this_cpu_ptr(&blk_cpu_done);
-	list_replace_init(cpu_list, &local_list);
-	local_irq_enable();
+	entry = llist_del_all(cpu_list);
+	entry = llist_reverse_order(entry);
 
-	while (!list_empty(&local_list)) {
-		struct request *rq;
-
-		rq = list_entry(local_list.next, struct request, ipi_list);
-		list_del_init(&rq->ipi_list);
+	llist_for_each_entry_safe(rq, rq_next, entry, ipi_list)
 		rq->q->mq_ops->complete(rq);
-	}
 }
 
-static void blk_mq_trigger_softirq(struct request *rq)
+static __latent_entropy void blk_done_softirq(struct softirq_action *h)
 {
-	struct list_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-	list = this_cpu_ptr(&blk_cpu_done);
-	list_add_tail(&rq->ipi_list, list);
-
-	/*
-	 * If the list only contains our just added request, signal a raise of
-	 * the softirq.  If there are already entries there, someone already
-	 * raised the irq but it hasn't run yet.
-	 */
-	if (list->next == &rq->ipi_list)
-		raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_restore(flags);
+	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
 }
 
 static int blk_softirq_cpu_dead(unsigned int cpu)
 {
-	/*
-	 * If a CPU goes away, splice its entries to the current CPU
-	 * and trigger a run of the softirq
-	 */
-	local_irq_disable();
-	list_splice_init(&per_cpu(blk_cpu_done, cpu),
-			 this_cpu_ptr(&blk_cpu_done));
-	raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_enable();
-
+	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
 	return 0;
 }
 
-
 static void __blk_mq_complete_request_remote(void *data)
 {
-	struct request *rq = data;
-
-	blk_mq_trigger_softirq(rq);
+	__raise_softirq_irqoff(BLOCK_SOFTIRQ);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)
@@ -657,6 +621,7 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 
 bool blk_mq_complete_request_remote(struct request *rq)
 {
+	struct llist_head *cpu_list;
 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 
 	/*
@@ -667,14 +632,21 @@ bool blk_mq_complete_request_remote(struct request *rq)
 		return false;
 
 	if (blk_mq_complete_need_ipi(rq)) {
-		rq->csd.func = __blk_mq_complete_request_remote;
-		rq->csd.info = rq;
-		rq->csd.flags = 0;
-		smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
+		unsigned int cpu;
+
+		cpu = rq->mq_ctx->cpu;
+		cpu_list = &per_cpu(blk_cpu_done, cpu);
+		if (llist_add(&rq->ipi_list, cpu_list)) {
+			rq->csd.func = __blk_mq_complete_request_remote;
+			rq->csd.flags = 0;
+			smp_call_function_single_async(cpu, &rq->csd);
+		}
 	} else {
 		if (rq->q->nr_hw_queues > 1)
 			return false;
-		blk_mq_trigger_softirq(rq);
+		cpu_list = this_cpu_ptr(&blk_cpu_done);
+		if (llist_add(&rq->ipi_list, cpu_list))
+			raise_softirq(BLOCK_SOFTIRQ);
 	}
 
 	return true;
@@ -3905,7 +3877,7 @@ static int __init blk_mq_init(void)
 	int i;
 
 	for_each_possible_cpu(i)
-		INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
+		init_llist_head(&per_cpu(blk_cpu_done, i));
 	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
 
 	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 639cae2c158b5..331b2b675b417 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -156,7 +156,7 @@ struct request {
 	 */
 	union {
 		struct hlist_node hash;	/* merge hash */
-		struct list_head ipi_list;
+		struct llist_node ipi_list;
 	};
 
 	/*
-- 
2.28.0

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

On Wed, Oct 28, 2020 at 03:12:51PM +0100, Sebastian Andrzej Siewior wrote:
>  static int blk_softirq_cpu_dead(unsigned int cpu)
>  {
> -	/*
> -	 * If a CPU goes away, splice its entries to the current CPU
> -	 * and trigger a run of the softirq
> -	 */
> -	local_irq_disable();
> -	list_splice_init(&per_cpu(blk_cpu_done, cpu),
> -			 this_cpu_ptr(&blk_cpu_done));
> -	raise_softirq_irqoff(BLOCK_SOFTIRQ);
> -	local_irq_enable();
> -
> +	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
>  	return 0;

How can this be preempted?  Can't we keep using this_cpu_ptr here?

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-10-28 14:44:53 [+0000], Christoph Hellwig wrote:
> On Wed, Oct 28, 2020 at 03:12:51PM +0100, Sebastian Andrzej Siewior wrote:
> >  static int blk_softirq_cpu_dead(unsigned int cpu)
> >  {
> > -	/*
> > -	 * If a CPU goes away, splice its entries to the current CPU
> > -	 * and trigger a run of the softirq
> > -	 */
> > -	local_irq_disable();
> > -	list_splice_init(&per_cpu(blk_cpu_done, cpu),
> > -			 this_cpu_ptr(&blk_cpu_done));
> > -	raise_softirq_irqoff(BLOCK_SOFTIRQ);
> > -	local_irq_enable();
> > -
> > +	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
> >  	return 0;
> 
> How can this be preempted?  Can't we keep using this_cpu_ptr here?

cpu of the dead CPU != this CPU.

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-10-28 15:12:51 [+0100], To linux-block@vger.kernel.org wrote:
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -667,14 +632,21 @@ bool blk_mq_complete_request_remote(struct request *rq)
>  		return false;
>  
>  	if (blk_mq_complete_need_ipi(rq)) {
…
>  	} else {
>  		if (rq->q->nr_hw_queues > 1)
>  			return false;
> -		blk_mq_trigger_softirq(rq);
> +		cpu_list = this_cpu_ptr(&blk_cpu_done);
> +		if (llist_add(&rq->ipi_list, cpu_list))
> +			raise_softirq(BLOCK_SOFTIRQ);
>  	}
>  
>  	return true;

So Mike posted this:
| BUG: using smp_processor_id() in preemptible [00000000] code: usb-storage/841
| caller is blk_mq_complete_request_remote.part.0+0xa2/0x120
| CPU: 0 PID: 841 Comm: usb-storage Not tainted 5.10.0-rc1+ #61
| Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-1 04/01/2014
| Call Trace:
|  dump_stack+0x77/0x97
|  check_preemption_disabled+0xbe/0xc0
|  blk_mq_complete_request_remote.part.0+0xa2/0x120
|  blk_mq_complete_request+0x2e/0x40
|  usb_stor_control_thread+0x29a/0x300
|  kthread+0x14b/0x170
|  ret_from_fork+0x22/0x30

This comes from this_cpu_ptr() because usb_stor_control_thread() runs
with enabled preemption.

Adding preempt_disable() around it will make the warning go away but
will wake the ksoftirqd (this happens now, too).
Adding local_bh_disable() around it would perform the completion
immediately (instead of waking kssoftirqd) but local_bh_enable() feels
slightly more expensive.

Are there many drivers completing the SCSI requests in preemtible
context? In this case it would be more efficient to complete the request
directly (usb_stor_control_thread() goes to sleep after that anyway and
there is only one request at a time).

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

On Thu, Oct 29, 2020 at 02:12:12PM +0100, Sebastian Andrzej Siewior wrote:
> Are there many drivers completing the SCSI requests in preemtible
> context? In this case it would be more efficient to complete the request
> directly (usb_stor_control_thread() goes to sleep after that anyway and
> there is only one request at a time).

Well, usb-storage obviously seems to do it, and the block layer
does not prohibit it.

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-10-29 14:05:36 [+0000], Christoph Hellwig wrote:
> Well, usb-storage obviously seems to do it, and the block layer
> does not prohibit it.

Also loop, nvme-tcp and then I stopped looking.
Any objections about adding local_bh_disable() around it?

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

On Thu, Oct 29, 2020 at 03:56:23PM +0100, Sebastian Andrzej Siewior wrote:
> On 2020-10-29 14:05:36 [+0000], Christoph Hellwig wrote:
> > Well, usb-storage obviously seems to do it, and the block layer
> > does not prohibit it.
> 
> Also loop, nvme-tcp and then I stopped looking.
> Any objections about adding local_bh_disable() around it?

To me it seems like the whole IPI plus potentially softirq dance is
a little pointless when completing from process context.

Sagi, any opinion on that from the nvme-tcp POV?

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sagi Grimberg]

>>> Well, usb-storage obviously seems to do it, and the block layer
>>> does not prohibit it.
>>
>> Also loop, nvme-tcp and then I stopped looking.
>> Any objections about adding local_bh_disable() around it?
> 
> To me it seems like the whole IPI plus potentially softirq dance is
> a little pointless when completing from process context.

I agree.

> Sagi, any opinion on that from the nvme-tcp POV?

nvme-tcp should (almost) always complete from the context that matches
the rq->mq_ctx->cpu as the thread that processes incoming
completions (per hctx) should be affinitized to match it (unless cpus
come and go).

So for nvme-tcp I don't expect blk_mq_complete_need_ipi to return true
in normal operation. That leaves the teardowns+aborts, which aren't very
interesting here.

I would note that nvme-tcp does not go to sleep after completing every
I/O like how sebastian indicated usb does.

Having said that, today the network stack is calling nvme_tcp_data_ready
in napi context (softirq) which in turn triggers the queue thread to
handle network rx (and complete the I/O). It's been measured recently
that running the rx context directly in softirq will save some
latency (possible because nvme-tcp rx context is non-blocking).

So I'd think that patch #2 is unnecessary and just add overhead for
nvme-tcp.. do note that the napi softirq cpu mapping depends on the RSS
steering, which is unlikely to match rq->mq_ctx->cpu, hence if completed
from napi context, nvme-tcp will probably always go to the IPI path.

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-10-29 13:03:26 [-0700], Sagi Grimberg wrote:
> 
> > > > Well, usb-storage obviously seems to do it, and the block layer
> > > > does not prohibit it.
> > > 
> > > Also loop, nvme-tcp and then I stopped looking.
> > > Any objections about adding local_bh_disable() around it?
> > 
> > To me it seems like the whole IPI plus potentially softirq dance is
> > a little pointless when completing from process context.
> 
> I agree.
> 
> > Sagi, any opinion on that from the nvme-tcp POV?
> 
> nvme-tcp should (almost) always complete from the context that matches
> the rq->mq_ctx->cpu as the thread that processes incoming
> completions (per hctx) should be affinitized to match it (unless cpus
> come and go).

in which context? But this is probably nr_hw_queues > 1?

> So for nvme-tcp I don't expect blk_mq_complete_need_ipi to return true
> in normal operation. That leaves the teardowns+aborts, which aren't very
> interesting here.

The process context invocation is nvme_tcp_complete_timed_out().

> I would note that nvme-tcp does not go to sleep after completing every
> I/O like how sebastian indicated usb does.
> 
> Having said that, today the network stack is calling nvme_tcp_data_ready
> in napi context (softirq) which in turn triggers the queue thread to
> handle network rx (and complete the I/O). It's been measured recently
> that running the rx context directly in softirq will save some
> latency (possible because nvme-tcp rx context is non-blocking).
> 
> So I'd think that patch #2 is unnecessary and just add overhead for
> nvme-tcp.. do note that the napi softirq cpu mapping depends on the RSS
> steering, which is unlikely to match rq->mq_ctx->cpu, hence if completed
> from napi context, nvme-tcp will probably always go to the IPI path.

but running it in softirq on the remote CPU would still allow of other
packets to come on the remote CPU (which would block BLOCK sofirq if
NET_RX is already running).

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sagi Grimberg]

>>>>> Well, usb-storage obviously seems to do it, and the block layer
>>>>> does not prohibit it.
>>>>
>>>> Also loop, nvme-tcp and then I stopped looking.
>>>> Any objections about adding local_bh_disable() around it?
>>>
>>> To me it seems like the whole IPI plus potentially softirq dance is
>>> a little pointless when completing from process context.
>>
>> I agree.
>>
>>> Sagi, any opinion on that from the nvme-tcp POV?
>>
>> nvme-tcp should (almost) always complete from the context that matches
>> the rq->mq_ctx->cpu as the thread that processes incoming
>> completions (per hctx) should be affinitized to match it (unless cpus
>> come and go).
> 
> in which context?

Not sure what is the question.

> But this is probably nr_hw_queues > 1?

Yes.

>> So for nvme-tcp I don't expect blk_mq_complete_need_ipi to return true
>> in normal operation. That leaves the teardowns+aborts, which aren't very
>> interesting here.
> 
> The process context invocation is nvme_tcp_complete_timed_out().

Yes.

>> I would note that nvme-tcp does not go to sleep after completing every
>> I/O like how sebastian indicated usb does.
>>
>> Having said that, today the network stack is calling nvme_tcp_data_ready
>> in napi context (softirq) which in turn triggers the queue thread to
>> handle network rx (and complete the I/O). It's been measured recently
>> that running the rx context directly in softirq will save some
>> latency (possible because nvme-tcp rx context is non-blocking).
>>
>> So I'd think that patch #2 is unnecessary and just add overhead for
>> nvme-tcp.. do note that the napi softirq cpu mapping depends on the RSS
>> steering, which is unlikely to match rq->mq_ctx->cpu, hence if completed
>> from napi context, nvme-tcp will probably always go to the IPI path.
> 
> but running it in softirq on the remote CPU would still allow of other
> packets to come on the remote CPU (which would block BLOCK sofirq if
> NET_RX is already running).

Not sure I understand your comment, if napi triggers on core X and we
complete from that, it will trigger IPI to core Y, and there with patch 
#2 is will trigger softirq instead of calling ->complete directly no?

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-10-29 14:07:59 [-0700], Sagi Grimberg wrote:
> > in which context?
> 
> Not sure what is the question.

The question is in which context do you complete your requests. My guess
by now is "usually softirq/NAPI and context in rare error case".

> > But this is probably nr_hw_queues > 1?
> 
> Yes.

So this means it will either complete directly or issue an IPI.

> > but running it in softirq on the remote CPU would still allow of other
> > packets to come on the remote CPU (which would block BLOCK sofirq if
> > NET_RX is already running).
> 
> Not sure I understand your comment, if napi triggers on core X and we
> complete from that, it will trigger IPI to core Y, and there with patch #2
> is will trigger softirq instead of calling ->complete directly no?

This is correct. But trigger softirq does not mean that it will wake
`ksoftirqd' as it is the case for the usb-storage right now. In your
case (completing from NAPI/sofitrq (or for most other driver which
complete in their IRQ handler)) it means:
- trigger IPI
- IPI will OR the BLOCK-softirq bit.
- on exit from IPI it will invoke do_softirq() (unless softirq is
  already pending and got interrupted by the IPI) and complete the
  Block request.

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Jens Axboe]

On 10/31/20 4:41 AM, Sebastian Andrzej Siewior wrote:
> On 2020-10-29 14:07:59 [-0700], Sagi Grimberg wrote:
>>> in which context?
>>
>> Not sure what is the question.
> 
> The question is in which context do you complete your requests. My guess
> by now is "usually softirq/NAPI and context in rare error case".

There really aren't any rules for this, and it's perfectly legit to
complete from process context. Maybe you're a kthread driven driver and
that's how you handle completions. The block completion path has always
been hard IRQ safe, but possible to call from anywhere.

-- 
Jens Axboe

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Jens Axboe]

On 10/31/20 9:00 AM, Jens Axboe wrote:
> On 10/31/20 4:41 AM, Sebastian Andrzej Siewior wrote:
>> On 2020-10-29 14:07:59 [-0700], Sagi Grimberg wrote:
>>>> in which context?
>>>
>>> Not sure what is the question.
>>
>> The question is in which context do you complete your requests. My guess
>> by now is "usually softirq/NAPI and context in rare error case".
> 
> There really aren't any rules for this, and it's perfectly legit to
> complete from process context. Maybe you're a kthread driven driver and
> that's how you handle completions. The block completion path has always
> been hard IRQ safe, but possible to call from anywhere.

A more recent example is polled IO, which will always complete from
process/task context and very much is fast path.

-- 
Jens Axboe

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

On Sat, Oct 31, 2020 at 09:01:45AM -0600, Jens Axboe wrote:
> On 10/31/20 9:00 AM, Jens Axboe wrote:
> > On 10/31/20 4:41 AM, Sebastian Andrzej Siewior wrote:
> >> On 2020-10-29 14:07:59 [-0700], Sagi Grimberg wrote:
> >>>> in which context?
> >>>
> >>> Not sure what is the question.
> >>
> >> The question is in which context do you complete your requests. My guess
> >> by now is "usually softirq/NAPI and context in rare error case".
> > 
> > There really aren't any rules for this, and it's perfectly legit to
> > complete from process context. Maybe you're a kthread driven driver and
> > that's how you handle completions. The block completion path has always
> > been hard IRQ safe, but possible to call from anywhere.
> 
> A more recent example is polled IO, which will always complete from
> process/task context and very much is fast path.

But we never IPI for that anyway, so it is the easy case.

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-10-31 09:00:49 [-0600], Jens Axboe wrote:
> There really aren't any rules for this, and it's perfectly legit to
> complete from process context. Maybe you're a kthread driven driver and
> that's how you handle completions. The block completion path has always
> been hard IRQ safe, but possible to call from anywhere.

I'm not trying to put restrictions and forbidding completions from a
kthread. I'm trying to avoid the pointless softirq dance for no added
value. We could:

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 4f53de48e5038..c4693b3750878 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -644,9 +644,11 @@ bool blk_mq_complete_request_remote(struct request *rq)
 	} else {
 		if (rq->q->nr_hw_queues > 1)
 			return false;
+		preempt_disable();
 		cpu_list = this_cpu_ptr(&blk_cpu_done);
 		if (llist_add(&rq->ipi_list, cpu_list))
 			raise_softirq(BLOCK_SOFTIRQ);
+		preempt_enable();
 	}
 
 	return true;

to not break that assumption you just mentioned and provide 
|static inline void blk_mq_complete_request_local(struct request *rq)
|{
|                 rq->q->mq_ops->complete(rq);
|}

so that completion issued from from process context (like those from
usb-storage) don't end up waking `ksoftird' (running at SCHED_OTHER)
completing the requests but rather performing it right away. The softirq
dance makes no sense here.

As mentioned earlier, the alternative _could_ be to
	s/preempt_/local_bh_/

in the above patch. This would ensure that any invocation outside of
IRQ/Softirq context would invoke the softirq _directly_ at
local_bh_enable() time rather than waking the daemon for that purpose.
It would also avoid another completion function for the direct case
which could be abused if used from outside the thread context.
The last one is currently my favorite.

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

On Mon, Nov 02, 2020 at 10:55:33AM +0100, Sebastian Andrzej Siewior wrote:
> On 2020-10-31 09:00:49 [-0600], Jens Axboe wrote:
> > There really aren't any rules for this, and it's perfectly legit to
> > complete from process context. Maybe you're a kthread driven driver and
> > that's how you handle completions. The block completion path has always
> > been hard IRQ safe, but possible to call from anywhere.
> 
> I'm not trying to put restrictions and forbidding completions from a
> kthread. I'm trying to avoid the pointless softirq dance for no added
> value. We could:

> to not break that assumption you just mentioned and provide 
> |static inline void blk_mq_complete_request_local(struct request *rq)
> |{
> |                 rq->q->mq_ops->complete(rq);
> |}
> 
> so that completion issued from from process context (like those from
> usb-storage) don't end up waking `ksoftird' (running at SCHED_OTHER)
> completing the requests but rather performing it right away. The softirq
> dance makes no sense here.

Agreed.  But I don't think your above blk_mq_complete_request_local
is all that useful either as ->complete is defined by the caller,
so we could just do a direct call.  Basically we should just
return false from blk_mq_complete_request_remote after updating
the state when called from process context.  But given that IIRC
we are not supposed to check what state we are called from
we'll need a helper just for updating the state instead and
ensure the driver uses the right helper.  Now of course we might
have process context callers that still want to bounce to the
submitting CPU, but in that case we should go directly to a
workqueue or similar.

Either way doing this properly will probabl involve an audit of all
drivers, but I think that is worth it.

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sagi Grimberg]

>>> There really aren't any rules for this, and it's perfectly legit to
>>> complete from process context. Maybe you're a kthread driven driver and
>>> that's how you handle completions. The block completion path has always
>>> been hard IRQ safe, but possible to call from anywhere.
>>
>> I'm not trying to put restrictions and forbidding completions from a
>> kthread. I'm trying to avoid the pointless softirq dance for no added
>> value. We could:
> 
>> to not break that assumption you just mentioned and provide
>> |static inline void blk_mq_complete_request_local(struct request *rq)
>> |{
>> |                 rq->q->mq_ops->complete(rq);
>> |}
>>
>> so that completion issued from from process context (like those from
>> usb-storage) don't end up waking `ksoftird' (running at SCHED_OTHER)
>> completing the requests but rather performing it right away. The softirq
>> dance makes no sense here.
> 
> Agreed.  But I don't think your above blk_mq_complete_request_local
> is all that useful either as ->complete is defined by the caller,
> so we could just do a direct call.  Basically we should just
> return false from blk_mq_complete_request_remote after updating
> the state when called from process context.

Agreed.

> But given that IIRC
> we are not supposed to check what state we are called from
> we'll need a helper just for updating the state instead and
> ensure the driver uses the right helper.  Now of course we might
> have process context callers that still want to bounce to the
> submitting CPU, but in that case we should go directly to a
> workqueue or similar.

This would mean that it may be suboptimal for nvme-tcp to complete
requests in softirq context from the network context (determined by
NIC steering). Because in this case, this would trigger workqueue
schedule on a per-request basis rather than once per .data_ready
call like we do today. Is that correct?

It has been observed that completing commands in softirq context
(network determined cpu) because basically the completion does
IPI + local complete, not IPI + softirq or IPI + workqueue.

> Either way doing this properly will probabl involve an audit of all
> drivers, but I think that is worth it.

Agree.

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2020-11-02 18:12:38 [+0000], Christoph Hellwig wrote:
> > to not break that assumption you just mentioned and provide 
> > |static inline void blk_mq_complete_request_local(struct request *rq)
> > |{
> > |                 rq->q->mq_ops->complete(rq);
> > |}
> > 
> > so that completion issued from from process context (like those from
> > usb-storage) don't end up waking `ksoftird' (running at SCHED_OTHER)
> > completing the requests but rather performing it right away. The softirq
> > dance makes no sense here.
> 
> Agreed.  But I don't think your above blk_mq_complete_request_local
> is all that useful either as ->complete is defined by the caller,
> so we could just do a direct call.
In usb-storage case it is hidden somewhere in the SCSI stack but this
can probably be changed later on.

>                                     Basically we should just
> return false from blk_mq_complete_request_remote after updating
> the state when called from process context.  But given that IIRC
> we are not supposed to check what state we are called from
> we'll need a helper just for updating the state instead and
> ensure the driver uses the right helper.  Now of course we might
> have process context callers that still want to bounce to the
> submitting CPU, but in that case we should go directly to a
> workqueue or similar.

So instead blk_mq_complete_request_local() you want a helper to set the
state in which the completion function is invoked. Sounds more like an
argument :)

> Either way doing this properly will probabl involve an audit of all
> drivers, but I think that is worth it.

I'm lost. Should I repost the three patches with a preempt_disable()
section (as suggested) to not break preemptible callers? And then move
from there to provide callers from preemtible context an alternative?

Sebastian

Re: [PATCH RFC] blk-mq: Don't IPI requests on PREEMPT_RT

[Peter Zijlstra]

On Tue, Oct 27, 2020 at 05:23:09PM +0000, Christoph Hellwig wrote:
> Ok.  I was hoping we could hide this in core code somehow, especially
> a peterz didn't like the use of smp_call_function_single_async in the
> blk-mq completion code very much.

It's smp_call_function_single_async() in general that I don't much like.
But Linus seemed unconvinced, so we'll keep it for a while I suppose.

[PATCH 0/3 v2] blk-mq: Don't complete in IRQ, use llist_head

[Sebastian Andrzej Siewior]

This a repost of the patches in the old thread [0] which died, rebase against
-next.

The series avoids completing the requests on a remote CPU if booted with
threadirqs. It avoids completing requests in hard-IRQ context on remote
CPU by deferring it to the the softirq context.

One change since the last post: preempt-disable() around llist_add() +
raise_softirq() to ensure that request is added on the same CPU where
the softirq is raised. Some callers (like usb-storage) invoke this
function from preemtible context and this preserves the current "call me
from any context" semantic.
My understanding is that in a later attempt we may change such callers
to complete directly and avoid the softirq ping-pong.

[0] https://lkml.kernel.org/r/20201028141251.3608598-1-bigeasy@linutronix.de

Sebastian

[PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

With llist_head it is possible to avoid the locking (the irq-off region)
when items are added. This makes it possible to add items on a remote
CPU.
llist_add() returns true if the list was previously empty. This can be
used to invoke the SMP function call / raise sofirq only if the first
item was added (otherwise it is already pending).
This simplifies the code a little and reduces the IRQ-off regions. With
this change it possible to reduce the SMP-function call a simple
__raise_softirq_irqoff().
blk_mq_complete_request_remote() needs a preempt-disable section if the
request needs to complete on the local CPU. Some callers (USB-storage)
invoke this preemptible context and the request needs to be enqueued on
the same CPU as the softirq is raised.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 block/blk-mq.c         | 77 ++++++++++++++----------------------------
 include/linux/blkdev.h |  2 +-
 2 files changed, 27 insertions(+), 52 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 3c0e94913d874..b5138327952a4 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -41,7 +41,7 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"
 
-static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
+static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
 
 static void blk_mq_poll_stats_start(struct request_queue *q);
 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
@@ -567,68 +567,32 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
 }
 EXPORT_SYMBOL(blk_mq_end_request);
 
-/*
- * Softirq action handler - move entries to local list and loop over them
- * while passing them to the queue registered handler.
- */
-static __latent_entropy void blk_done_softirq(struct softirq_action *h)
+static void blk_complete_reqs(struct llist_head *cpu_list)
 {
-	struct list_head *cpu_list, local_list;
+	struct llist_node *entry;
+	struct request *rq, *rq_next;
 
-	local_irq_disable();
-	cpu_list = this_cpu_ptr(&blk_cpu_done);
-	list_replace_init(cpu_list, &local_list);
-	local_irq_enable();
+	entry = llist_del_all(cpu_list);
+	entry = llist_reverse_order(entry);
 
-	while (!list_empty(&local_list)) {
-		struct request *rq;
-
-		rq = list_entry(local_list.next, struct request, ipi_list);
-		list_del_init(&rq->ipi_list);
+	llist_for_each_entry_safe(rq, rq_next, entry, ipi_list)
 		rq->q->mq_ops->complete(rq);
-	}
 }
 
-static void blk_mq_trigger_softirq(struct request *rq)
+static __latent_entropy void blk_done_softirq(struct softirq_action *h)
 {
-	struct list_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-	list = this_cpu_ptr(&blk_cpu_done);
-	list_add_tail(&rq->ipi_list, list);
-
-	/*
-	 * If the list only contains our just added request, signal a raise of
-	 * the softirq.  If there are already entries there, someone already
-	 * raised the irq but it hasn't run yet.
-	 */
-	if (list->next == &rq->ipi_list)
-		raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_restore(flags);
+	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
 }
 
 static int blk_softirq_cpu_dead(unsigned int cpu)
 {
-	/*
-	 * If a CPU goes away, splice its entries to the current CPU
-	 * and trigger a run of the softirq
-	 */
-	local_irq_disable();
-	list_splice_init(&per_cpu(blk_cpu_done, cpu),
-			 this_cpu_ptr(&blk_cpu_done));
-	raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_enable();
-
+	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
 	return 0;
 }
 
-
 static void __blk_mq_complete_request_remote(void *data)
 {
-	struct request *rq = data;
-
-	blk_mq_trigger_softirq(rq);
+	__raise_softirq_irqoff(BLOCK_SOFTIRQ);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)
@@ -659,6 +623,7 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 
 bool blk_mq_complete_request_remote(struct request *rq)
 {
+	struct llist_head *cpu_list;
 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 
 	/*
@@ -669,12 +634,22 @@ bool blk_mq_complete_request_remote(struct request *rq)
 		return false;
 
 	if (blk_mq_complete_need_ipi(rq)) {
-		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
-		smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
+		unsigned int cpu;
+
+		cpu = rq->mq_ctx->cpu;
+		cpu_list = &per_cpu(blk_cpu_done, cpu);
+		if (llist_add(&rq->ipi_list, cpu_list)) {
+			INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
+			smp_call_function_single_async(cpu, &rq->csd);
+		}
 	} else {
 		if (rq->q->nr_hw_queues > 1)
 			return false;
-		blk_mq_trigger_softirq(rq);
+		preempt_disable();
+		cpu_list = this_cpu_ptr(&blk_cpu_done);
+		if (llist_add(&rq->ipi_list, cpu_list))
+			raise_softirq(BLOCK_SOFTIRQ);
+		preempt_enable();
 	}
 
 	return true;
@@ -3905,7 +3880,7 @@ static int __init blk_mq_init(void)
 	int i;
 
 	for_each_possible_cpu(i)
-		INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
+		init_llist_head(&per_cpu(blk_cpu_done, i));
 	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
 
 	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 17cedf0dc83db..7b05390766eec 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -156,7 +156,7 @@ struct request {
 	 */
 	union {
 		struct hlist_node hash;	/* merge hash */
-		struct list_head ipi_list;
+		struct llist_node ipi_list;
 	};
 
 	/*
-- 
2.29.2

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset=unknown-8bit, Size: 3467 bytes --]

On Fri, Dec 04, 2020 at 08:13:56PM +0100, Sebastian Andrzej Siewior wrote:
> With llist_head it is possible to avoid the locking (the irq-off region)
> when items are added. This makes it possible to add items on a remote
> CPU.
> llist_add() returns true if the list was previously empty. This can be
> used to invoke the SMP function call / raise sofirq only if the first
> item was added (otherwise it is already pending).
> This simplifies the code a little and reduces the IRQ-off regions. With
> this change it possible to reduce the SMP-function call a simple
> __raise_softirq_irqoff().
> blk_mq_complete_request_remote() needs a preempt-disable section if the
> request needs to complete on the local CPU. Some callers (USB-storage)
> invoke this preemptible context and the request needs to be enqueued on
> the same CPU as the softirq is raised.
> 
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> ---
>  block/blk-mq.c         | 77 ++++++++++++++----------------------------
>  include/linux/blkdev.h |  2 +-
>  2 files changed, 27 insertions(+), 52 deletions(-)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 3c0e94913d874..b5138327952a4 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -41,7 +41,7 @@
>  #include "blk-mq-sched.h"
>  #include "blk-rq-qos.h"
>  
> +static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
>  
>  static void blk_mq_poll_stats_start(struct request_queue *q);
>  static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
> @@ -567,68 +567,32 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
>  }
>  EXPORT_SYMBOL(blk_mq_end_request);
>  
> +static void blk_complete_reqs(struct llist_head *cpu_list)
>  {
> +	struct llist_node *entry;
> +	struct request *rq, *rq_next;
>  
> +	entry = llist_del_all(cpu_list);
> +	entry = llist_reverse_order(entry);

I find the variable naming and split of the assignments a little
strange.  What about:

static void blk_complete_reqs(struct llist_head *list)
{
	struct llist_node *first = llist_reverse_order(llist_del_all(list));
	struct request *rq, *next;

?

> +	llist_for_each_entry_safe(rq, rq_next, entry, ipi_list)
>  		rq->q->mq_ops->complete(rq);
>  }

Aren't some sanitizers going to be unhappy if we never delete the
request from the list?

>  bool blk_mq_complete_request_remote(struct request *rq)
>  {
> +	struct llist_head *cpu_list;
>  	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
>  
>  	/*
> @@ -669,12 +634,22 @@ bool blk_mq_complete_request_remote(struct request *rq)
>  		return false;
>  
>  	if (blk_mq_complete_need_ipi(rq)) {
> +		unsigned int cpu;
> +
> +		cpu = rq->mq_ctx->cpu;
> +		cpu_list = &per_cpu(blk_cpu_done, cpu);
> +		if (llist_add(&rq->ipi_list, cpu_list)) {
> +			INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
> +			smp_call_function_single_async(cpu, &rq->csd);
> +		}

I think the above code section inside the conditional should go into a
little helper instead of being open coded here in the fast path routine.
I laso don't really see the ¶oint of the cpu and cpulist locl variables.

>  	} else {
>  		if (rq->q->nr_hw_queues > 1)
>  			return false;
> +		preempt_disable();
> +		cpu_list = this_cpu_ptr(&blk_cpu_done);
> +		if (llist_add(&rq->ipi_list, cpu_list))
> +			raise_softirq(BLOCK_SOFTIRQ);
> +		preempt_enable();

I think the section after the return false here also would benefit from
a little helper with a descriptive name.

Otherwise this looks good to me.

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

Just to clarify what I mean, I think the flow in
blk_mq_complete_request_remote should turn into something like:


	...

	if (rq->cmd_flags & REQ_HIPRI)
		return false;

	if (blk_mq_complete_need_ipi(rq))
		blk_mq_complete_send_ipi(rq);
		return true;
	}

	if (rq->q->nr_hw_queues == 1) {
		blk_mq_raise_softirq(rq);
		return true;
	}

	return false;
}

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset=utf-8, Size: 8624 bytes --]

On 2020-12-08 13:20:04 [+0000], Christoph Hellwig wrote:
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -41,7 +41,7 @@
> >  #include "blk-mq-sched.h"
> >  #include "blk-rq-qos.h"
> >  
> > +static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
> >  
> >  static void blk_mq_poll_stats_start(struct request_queue *q);
> >  static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
> > @@ -567,68 +567,32 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
> >  }
> >  EXPORT_SYMBOL(blk_mq_end_request);
> >  
> > +static void blk_complete_reqs(struct llist_head *cpu_list)
> >  {
> > +	struct llist_node *entry;
> > +	struct request *rq, *rq_next;
> >  
> > +	entry = llist_del_all(cpu_list);
> > +	entry = llist_reverse_order(entry);
> 
> I find the variable naming and split of the assignments a little
> strange.  What about:
> 
> static void blk_complete_reqs(struct llist_head *list)
> {
> 	struct llist_node *first = llist_reverse_order(llist_del_all(list));
> 	struct request *rq, *next;
> 
> ?

Sure.

> > +	llist_for_each_entry_safe(rq, rq_next, entry, ipi_list)
> >  		rq->q->mq_ops->complete(rq);
> >  }
> 
> Aren't some sanitizers going to be unhappy if we never delete the
> request from the list?

I don't think so. If so there is more to complain about like,
flush_smp_call_function_queue(), delayed_mntput(),
irq_work_run_list(), ...


> >  bool blk_mq_complete_request_remote(struct request *rq)
> >  {
> > +	struct llist_head *cpu_list;
> >  	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
> >  
> >  	/*
> > @@ -669,12 +634,22 @@ bool blk_mq_complete_request_remote(struct request *rq)
> >  		return false;
> >  
> >  	if (blk_mq_complete_need_ipi(rq)) {
> > +		unsigned int cpu;
> > +
> > +		cpu = rq->mq_ctx->cpu;
> > +		cpu_list = &per_cpu(blk_cpu_done, cpu);
> > +		if (llist_add(&rq->ipi_list, cpu_list)) {
> > +			INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
> > +			smp_call_function_single_async(cpu, &rq->csd);
> > +		}
> 
> I think the above code section inside the conditional should go into a
> little helper instead of being open coded here in the fast path routine.
> I laso don't really see the ¶oint of the cpu and cpulist locl variables.
> 
> >  	} else {
> >  		if (rq->q->nr_hw_queues > 1)
> >  			return false;
> > +		preempt_disable();
> > +		cpu_list = this_cpu_ptr(&blk_cpu_done);
> > +		if (llist_add(&rq->ipi_list, cpu_list))
> > +			raise_softirq(BLOCK_SOFTIRQ);
> > +		preempt_enable();
> 
> I think the section after the return false here also would benefit from
> a little helper with a descriptive name.
> 
> Otherwise this looks good to me.

Please see below.

----->8-------

From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date: Wed, 28 Oct 2020 11:08:21 +0100
Subject: [PATCH] blk-mq: Use llist_head for blk_cpu_done

With llist_head it is possible to avoid the locking (the irq-off region)
when items are added. This makes it possible to add items on a remote
CPU without additional locking.
llist_add() returns true if the list was previously empty. This can be
used to invoke the SMP function call / raise sofirq only if the first
item was added (otherwise it is already pending).
This simplifies the code a little and reduces the IRQ-off regions.

blk_mq_raise_softirq() needs a preempt-disable section to ensure the
request is enqueued on the same CPU as the softirq is raised.
Some callers (USB-storage) invoke this path in preemptible context.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 block/blk-mq.c         | 97 ++++++++++++++++++------------------------
 include/linux/blkdev.h |  2 +-
 2 files changed, 42 insertions(+), 57 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 9baa681f6ee67..959b45fd41882 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -41,7 +41,7 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"
 
-static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
+static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
 
 static void blk_mq_poll_stats_start(struct request_queue *q);
 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
@@ -567,68 +567,29 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
 }
 EXPORT_SYMBOL(blk_mq_end_request);
 
-/*
- * Softirq action handler - move entries to local list and loop over them
- * while passing them to the queue registered handler.
- */
-static __latent_entropy void blk_done_softirq(struct softirq_action *h)
+static void blk_complete_reqs(struct llist_head *list)
 {
-	struct list_head *cpu_list, local_list;
+	struct llist_node *entry = llist_reverse_order(llist_del_all(list));
+	struct request *rq, *next;
 
-	local_irq_disable();
-	cpu_list = this_cpu_ptr(&blk_cpu_done);
-	list_replace_init(cpu_list, &local_list);
-	local_irq_enable();
-
-	while (!list_empty(&local_list)) {
-		struct request *rq;
-
-		rq = list_entry(local_list.next, struct request, ipi_list);
-		list_del_init(&rq->ipi_list);
+	llist_for_each_entry_safe(rq, next, entry, ipi_list)
 		rq->q->mq_ops->complete(rq);
-	}
 }
 
-static void blk_mq_trigger_softirq(struct request *rq)
+static __latent_entropy void blk_done_softirq(struct softirq_action *h)
 {
-	struct list_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-	list = this_cpu_ptr(&blk_cpu_done);
-	list_add_tail(&rq->ipi_list, list);
-
-	/*
-	 * If the list only contains our just added request, signal a raise of
-	 * the softirq.  If there are already entries there, someone already
-	 * raised the irq but it hasn't run yet.
-	 */
-	if (list->next == &rq->ipi_list)
-		raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_restore(flags);
+	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
 }
 
 static int blk_softirq_cpu_dead(unsigned int cpu)
 {
-	/*
-	 * If a CPU goes away, splice its entries to the current CPU
-	 * and trigger a run of the softirq
-	 */
-	local_irq_disable();
-	list_splice_init(&per_cpu(blk_cpu_done, cpu),
-			 this_cpu_ptr(&blk_cpu_done));
-	raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_enable();
-
+	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
 	return 0;
 }
 
-
 static void __blk_mq_complete_request_remote(void *data)
 {
-	struct request *rq = data;
-
-	blk_mq_trigger_softirq(rq);
+	__raise_softirq_irqoff(BLOCK_SOFTIRQ);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)
@@ -657,6 +618,30 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 	return cpu_online(rq->mq_ctx->cpu);
 }
 
+static void blk_mq_complete_send_ipi(struct request *rq)
+{
+	struct llist_head *list;
+	unsigned int cpu;
+
+	cpu = rq->mq_ctx->cpu;
+	list = &per_cpu(blk_cpu_done, cpu);
+	if (llist_add(&rq->ipi_list, list)) {
+		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
+		smp_call_function_single_async(cpu, &rq->csd);
+	}
+}
+
+static void blk_mq_raise_softirq(struct request *rq)
+{
+	struct llist_head *list;
+
+	preempt_disable();
+	list = this_cpu_ptr(&blk_cpu_done);
+	if (llist_add(&rq->ipi_list, list))
+		raise_softirq(BLOCK_SOFTIRQ);
+	preempt_enable();
+}
+
 bool blk_mq_complete_request_remote(struct request *rq)
 {
 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
@@ -669,15 +654,15 @@ bool blk_mq_complete_request_remote(struct request *rq)
 		return false;
 
 	if (blk_mq_complete_need_ipi(rq)) {
-		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
-		smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
-	} else {
-		if (rq->q->nr_hw_queues > 1)
-			return false;
-		blk_mq_trigger_softirq(rq);
+		blk_mq_complete_send_ipi(rq);
+		return true;
 	}
 
-	return true;
+	if (rq->q->nr_hw_queues == 1) {
+		blk_mq_raise_softirq(rq);
+		return true;
+	}
+	return false;
 }
 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
 
@@ -3917,7 +3902,7 @@ static int __init blk_mq_init(void)
 	int i;
 
 	for_each_possible_cpu(i)
-		INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
+		init_llist_head(&per_cpu(blk_cpu_done, i));
 	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
 
 	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f94ee3089e015..89a444c5a5833 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -153,7 +153,7 @@ struct request {
 	 */
 	union {
 		struct hlist_node hash;	/* merge hash */
-		struct list_head ipi_list;
+		struct llist_node ipi_list;
 	};
 
 	/*
-- 
2.29.2

[PATCH v3 0/3] blk-mq: Don't complete in IRQ, use llist_head

[Sebastian Andrzej Siewior]

Patch 2+3 were applied and then dropped by Jens due to a NOHZ+softirq
related warning [0]. Turns out a successful wakeup via
set_nr_if_polling() will not process any softirqs and the CPU may go
back to idle. This is addressed by patch #1.

smpcfd_dying_cpu() will also invoke SMP-functions calls via
flush_smp_call_function_queue() but the block layer shouldn't queue
anything because the CPU isn't online anymore.
The two caller of flush_smp_call_function_from_idle() look fine with
opening interrupts from within do_softirq().

[0] https://lkml.kernel.org/r/1ee4b31b-350e-a9f5-4349-cfb34b89829a@kernel.dk

Sebastian

[PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

With llist_head it is possible to avoid the locking (the irq-off region)
when items are added. This makes it possible to add items on a remote
CPU without additional locking.
llist_add() returns true if the list was previously empty. This can be
used to invoke the SMP function call / raise sofirq only if the first
item was added (otherwise it is already pending).
This simplifies the code a little and reduces the IRQ-off regions.

blk_mq_raise_softirq() needs a preempt-disable section to ensure the
request is enqueued on the same CPU as the softirq is raised.
Some callers (USB-storage) invoke this path in preemptible context.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 block/blk-mq.c         | 97 ++++++++++++++++++------------------------
 include/linux/blkdev.h |  2 +-
 2 files changed, 42 insertions(+), 57 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 90348ae518461..463de2981df8a 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -41,7 +41,7 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"
 
-static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
+static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
 
 static void blk_mq_poll_stats_start(struct request_queue *q);
 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
@@ -567,68 +567,29 @@ void blk_mq_end_request(struct request *rq, blk_status_t error)
 }
 EXPORT_SYMBOL(blk_mq_end_request);
 
-/*
- * Softirq action handler - move entries to local list and loop over them
- * while passing them to the queue registered handler.
- */
-static __latent_entropy void blk_done_softirq(struct softirq_action *h)
+static void blk_complete_reqs(struct llist_head *list)
 {
-	struct list_head *cpu_list, local_list;
+	struct llist_node *entry = llist_reverse_order(llist_del_all(list));
+	struct request *rq, *next;
 
-	local_irq_disable();
-	cpu_list = this_cpu_ptr(&blk_cpu_done);
-	list_replace_init(cpu_list, &local_list);
-	local_irq_enable();
-
-	while (!list_empty(&local_list)) {
-		struct request *rq;
-
-		rq = list_entry(local_list.next, struct request, ipi_list);
-		list_del_init(&rq->ipi_list);
+	llist_for_each_entry_safe(rq, next, entry, ipi_list)
 		rq->q->mq_ops->complete(rq);
-	}
 }
 
-static void blk_mq_trigger_softirq(struct request *rq)
+static __latent_entropy void blk_done_softirq(struct softirq_action *h)
 {
-	struct list_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-	list = this_cpu_ptr(&blk_cpu_done);
-	list_add_tail(&rq->ipi_list, list);
-
-	/*
-	 * If the list only contains our just added request, signal a raise of
-	 * the softirq.  If there are already entries there, someone already
-	 * raised the irq but it hasn't run yet.
-	 */
-	if (list->next == &rq->ipi_list)
-		raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_restore(flags);
+	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
 }
 
 static int blk_softirq_cpu_dead(unsigned int cpu)
 {
-	/*
-	 * If a CPU goes away, splice its entries to the current CPU
-	 * and trigger a run of the softirq
-	 */
-	local_irq_disable();
-	list_splice_init(&per_cpu(blk_cpu_done, cpu),
-			 this_cpu_ptr(&blk_cpu_done));
-	raise_softirq_irqoff(BLOCK_SOFTIRQ);
-	local_irq_enable();
-
+	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
 	return 0;
 }
 
-
 static void __blk_mq_complete_request_remote(void *data)
 {
-	struct request *rq = data;
-
-	blk_mq_trigger_softirq(rq);
+	__raise_softirq_irqoff(BLOCK_SOFTIRQ);
 }
 
 static inline bool blk_mq_complete_need_ipi(struct request *rq)
@@ -657,6 +618,30 @@ static inline bool blk_mq_complete_need_ipi(struct request *rq)
 	return cpu_online(rq->mq_ctx->cpu);
 }
 
+static void blk_mq_complete_send_ipi(struct request *rq)
+{
+	struct llist_head *list;
+	unsigned int cpu;
+
+	cpu = rq->mq_ctx->cpu;
+	list = &per_cpu(blk_cpu_done, cpu);
+	if (llist_add(&rq->ipi_list, list)) {
+		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
+		smp_call_function_single_async(cpu, &rq->csd);
+	}
+}
+
+static void blk_mq_raise_softirq(struct request *rq)
+{
+	struct llist_head *list;
+
+	preempt_disable();
+	list = this_cpu_ptr(&blk_cpu_done);
+	if (llist_add(&rq->ipi_list, list))
+		raise_softirq(BLOCK_SOFTIRQ);
+	preempt_enable();
+}
+
 bool blk_mq_complete_request_remote(struct request *rq)
 {
 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
@@ -669,15 +654,15 @@ bool blk_mq_complete_request_remote(struct request *rq)
 		return false;
 
 	if (blk_mq_complete_need_ipi(rq)) {
-		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
-		smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
-	} else {
-		if (rq->q->nr_hw_queues > 1)
-			return false;
-		blk_mq_trigger_softirq(rq);
+		blk_mq_complete_send_ipi(rq);
+		return true;
 	}
 
-	return true;
+	if (rq->q->nr_hw_queues == 1) {
+		blk_mq_raise_softirq(rq);
+		return true;
+	}
+	return false;
 }
 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
 
@@ -3892,7 +3877,7 @@ static int __init blk_mq_init(void)
 	int i;
 
 	for_each_possible_cpu(i)
-		INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
+		init_llist_head(&per_cpu(blk_cpu_done, i));
 	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
 
 	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f94ee3089e015..89a444c5a5833 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -153,7 +153,7 @@ struct request {
 	 */
 	union {
 		struct hlist_node hash;	/* merge hash */
-		struct list_head ipi_list;
+		struct llist_node ipi_list;
 	};
 
 	/*
-- 
2.30.0

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

> +static void blk_mq_complete_send_ipi(struct request *rq)
> +{
> +	struct llist_head *list;
> +	unsigned int cpu;
> +
> +	cpu = rq->mq_ctx->cpu;
> +	list = &per_cpu(blk_cpu_done, cpu);
> +	if (llist_add(&rq->ipi_list, list)) {
> +		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
> +		smp_call_function_single_async(cpu, &rq->csd);
> +	}
> +}

Nit: it would be nice to initialize cpu and list in the declaration
lines.

Otherwise looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Sebastian Andrzej Siewior]

On 2021-01-25 08:30:12 [+0000], Christoph Hellwig wrote:
> > +static void blk_mq_complete_send_ipi(struct request *rq)
> > +{
> > +	struct llist_head *list;
> > +	unsigned int cpu;
> > +
> > +	cpu = rq->mq_ctx->cpu;
> > +	list = &per_cpu(blk_cpu_done, cpu);
> > +	if (llist_add(&rq->ipi_list, list)) {
> > +		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
> > +		smp_call_function_single_async(cpu, &rq->csd);
> > +	}
> > +}
> 
> Nit: it would be nice to initialize cpu and list in the declaration
> lines.

Why? They get initialized later.

> Otherwise looks good:
> 
> Reviewed-by: Christoph Hellwig <hch@lst.de>

Sebastian

Re: [PATCH 3/3] blk-mq: Use llist_head for blk_cpu_done

[Christoph Hellwig]

On Mon, Jan 25, 2021 at 09:32:04AM +0100, Sebastian Andrzej Siewior wrote:
> On 2021-01-25 08:30:12 [+0000], Christoph Hellwig wrote:
> > > +static void blk_mq_complete_send_ipi(struct request *rq)
> > > +{
> > > +	struct llist_head *list;
> > > +	unsigned int cpu;
> > > +
> > > +	cpu = rq->mq_ctx->cpu;
> > > +	list = &per_cpu(blk_cpu_done, cpu);
> > > +	if (llist_add(&rq->ipi_list, list)) {
> > > +		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
> > > +		smp_call_function_single_async(cpu, &rq->csd);
> > > +	}
> > > +}
> > 
> > Nit: it would be nice to initialize cpu and list in the declaration
> > lines.
> 
> Why? They get initialized later.

Because:

	unsigned int cpu = rq->mq_ctx->cpu;
	struct llist_head *list = &per_cpu(blk_cpu_done, cpu);

is a lot easier to follow than:

	struct llist_head *list;
	unsigned int cpu;

	cpu = rq->mq_ctx->cpu;
	list = &per_cpu(blk_cpu_done, cpu);