[PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Rik van Riel]

CPUs with nohz_full do not want disruption from timer interrupts,
or other random system things.  This includes block mq work.

There is another issue with block mq vs. realtime tasks that run
100% of the time, which is not uncommon on systems that have CPUs
dedicated to real time use with isolcpus= and nohz_full=

Specifically, on systems like that, a block work item may never
get to run, which could lead to filesystems getting stuck forever.

We can avoid both issues by not scheduling blk-mq workqueues on
cpus in nohz_full mode.

Question for Jens: should we try to spread out the load for
currently offline and nohz CPUs across the remaining CPUs in
the system, to get the full benefit of blk-mq in these situations?

If so, do you have any preference on how I should implement that?

Cc: Frederic Weisbecker <fweisbec@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jens Axboe <axboe@kernel.org>
Signed-off-by: Rik van Riel <riel@redhat.com>
---
 block/blk-mq.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 4f4bea21052e..1004d6817fa4 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -21,6 +21,7 @@
 #include <linux/sched/sysctl.h>
 #include <linux/delay.h>
 #include <linux/crash_dump.h>
+#include <linux/tick.h>
 
 #include <trace/events/block.h>
 
@@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
 		if (!cpu_online(i))
 			continue;
 
+		/* Do not schedule work on nohz full dedicated CPUs. */
+		if (tick_nohz_full_cpu(i))
+			continue;
+
 		hctx = q->mq_ops->map_queue(q, i);
 		cpumask_set_cpu(i, hctx->cpumask);
 		hctx->nr_ctx++;

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Frederic Weisbecker]

On Tue, Mar 31, 2015 at 10:27:26AM -0400, Rik van Riel wrote:
> CPUs with nohz_full do not want disruption from timer interrupts,
> or other random system things.  This includes block mq work.
> 
> There is another issue with block mq vs. realtime tasks that run
> 100% of the time, which is not uncommon on systems that have CPUs
> dedicated to real time use with isolcpus= and nohz_full=
> 
> Specifically, on systems like that, a block work item may never
> get to run, which could lead to filesystems getting stuck forever.
> 
> We can avoid both issues by not scheduling blk-mq workqueues on
> cpus in nohz_full mode.
> 
> Question for Jens: should we try to spread out the load for
> currently offline and nohz CPUs across the remaining CPUs in
> the system, to get the full benefit of blk-mq in these situations?
> 
> If so, do you have any preference on how I should implement that?
> 
> Cc: Frederic Weisbecker <fweisbec@redhat.com>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Jens Axboe <axboe@kernel.org>
> Signed-off-by: Rik van Riel <riel@redhat.com>
> ---
>  block/blk-mq.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 4f4bea21052e..1004d6817fa4 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -21,6 +21,7 @@
>  #include <linux/sched/sysctl.h>
>  #include <linux/delay.h>
>  #include <linux/crash_dump.h>
> +#include <linux/tick.h>
>  
>  #include <trace/events/block.h>
>  
> @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
>  		if (!cpu_online(i))
>  			continue;
>  
> +		/* Do not schedule work on nohz full dedicated CPUs. */
> +		if (tick_nohz_full_cpu(i))
> +			continue;

I guess in this case, the queue for this CPU will be handled by another CPU?
Is this an unbound workqueue? I guess it's not but if it is, we should wait for
the workqueue affinity patchset.

Also, since we are doing a lot of kernel pre-setting behind nohz full, it would
be nice to warn the user about each of them in dmesg.

> +
>  		hctx = q->mq_ops->map_queue(q, i);
>  		cpumask_set_cpu(i, hctx->cpumask);
>  		hctx->nr_ctx++;
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Marcelo Tosatti]

On Tue, Mar 31, 2015 at 05:02:38PM +0200, Frederic Weisbecker wrote:
> On Tue, Mar 31, 2015 at 10:27:26AM -0400, Rik van Riel wrote:
> > CPUs with nohz_full do not want disruption from timer interrupts,
> > or other random system things.  This includes block mq work.
> > 
> > There is another issue with block mq vs. realtime tasks that run
> > 100% of the time, which is not uncommon on systems that have CPUs
> > dedicated to real time use with isolcpus= and nohz_full=
> > 
> > Specifically, on systems like that, a block work item may never
> > get to run, which could lead to filesystems getting stuck forever.
> > 
> > We can avoid both issues by not scheduling blk-mq workqueues on
> > cpus in nohz_full mode.
> > 
> > Question for Jens: should we try to spread out the load for
> > currently offline and nohz CPUs across the remaining CPUs in
> > the system, to get the full benefit of blk-mq in these situations?
> > 
> > If so, do you have any preference on how I should implement that?
> > 
> > Cc: Frederic Weisbecker <fweisbec@redhat.com>
> > Cc: Ingo Molnar <mingo@kernel.org>
> > Cc: Jens Axboe <axboe@kernel.org>
> > Signed-off-by: Rik van Riel <riel@redhat.com>
> > ---
> >  block/blk-mq.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> > 
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index 4f4bea21052e..1004d6817fa4 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -21,6 +21,7 @@
> >  #include <linux/sched/sysctl.h>
> >  #include <linux/delay.h>
> >  #include <linux/crash_dump.h>
> > +#include <linux/tick.h>
> >  
> >  #include <trace/events/block.h>
> >  
> > @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
> >  		if (!cpu_online(i))
> >  			continue;
> >  
> > +		/* Do not schedule work on nohz full dedicated CPUs. */
> > +		if (tick_nohz_full_cpu(i))
> > +			continue;
> 
> I guess in this case, the queue for this CPU will be handled by another CPU?
> Is this an unbound workqueue? I guess it's not but if it is, we should wait for
> the workqueue affinity patchset.

Where is the latest version of that patchset again ?

> 
> Also, since we are doing a lot of kernel pre-setting behind nohz full, it would
> be nice to warn the user about each of them in dmesg.
> 
> > +
> >  		hctx = q->mq_ops->map_queue(q, i);
> >  		cpumask_set_cpu(i, hctx->cpumask);
> >  		hctx->nr_ctx++;
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at  http://www.tux.org/lkml/

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Luiz Capitulino]

On Tue, 31 Mar 2015 20:18:36 -0300
Marcelo Tosatti <mtosatti@redhat.com> wrote:

> On Tue, Mar 31, 2015 at 05:02:38PM +0200, Frederic Weisbecker wrote:
> > On Tue, Mar 31, 2015 at 10:27:26AM -0400, Rik van Riel wrote:
> > > CPUs with nohz_full do not want disruption from timer interrupts,
> > > or other random system things.  This includes block mq work.
> > > 
> > > There is another issue with block mq vs. realtime tasks that run
> > > 100% of the time, which is not uncommon on systems that have CPUs
> > > dedicated to real time use with isolcpus= and nohz_full=
> > > 
> > > Specifically, on systems like that, a block work item may never
> > > get to run, which could lead to filesystems getting stuck forever.
> > > 
> > > We can avoid both issues by not scheduling blk-mq workqueues on
> > > cpus in nohz_full mode.
> > > 
> > > Question for Jens: should we try to spread out the load for
> > > currently offline and nohz CPUs across the remaining CPUs in
> > > the system, to get the full benefit of blk-mq in these situations?
> > > 
> > > If so, do you have any preference on how I should implement that?
> > > 
> > > Cc: Frederic Weisbecker <fweisbec@redhat.com>
> > > Cc: Ingo Molnar <mingo@kernel.org>
> > > Cc: Jens Axboe <axboe@kernel.org>
> > > Signed-off-by: Rik van Riel <riel@redhat.com>
> > > ---
> > >  block/blk-mq.c | 5 +++++
> > >  1 file changed, 5 insertions(+)
> > > 
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index 4f4bea21052e..1004d6817fa4 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -21,6 +21,7 @@
> > >  #include <linux/sched/sysctl.h>
> > >  #include <linux/delay.h>
> > >  #include <linux/crash_dump.h>
> > > +#include <linux/tick.h>
> > >  
> > >  #include <trace/events/block.h>
> > >  
> > > @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
> > >  		if (!cpu_online(i))
> > >  			continue;
> > >  
> > > +		/* Do not schedule work on nohz full dedicated CPUs. */
> > > +		if (tick_nohz_full_cpu(i))
> > > +			continue;
> > 
> > I guess in this case, the queue for this CPU will be handled by another CPU?
> > Is this an unbound workqueue? I guess it's not but if it is, we should wait for
> > the workqueue affinity patchset.
> 
> Where is the latest version of that patchset again ?

I think you're referring to this series:

https://marc.info/?l=linux-kernel&m=142613630606404&w=2

> 
> > 
> > Also, since we are doing a lot of kernel pre-setting behind nohz full, it would
> > be nice to warn the user about each of them in dmesg.
> > 
> > > +
> > >  		hctx = q->mq_ops->map_queue(q, i);
> > >  		cpumask_set_cpu(i, hctx->cpumask);
> > >  		hctx->nr_ctx++;
> > > --
> > > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > > the body of a message to majordomo@vger.kernel.org
> > > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > > Please read the FAQ at  http://www.tux.org/lkml/
> 

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Jens Axboe]

On 03/31/2015 08:27 AM, Rik van Riel wrote:
> CPUs with nohz_full do not want disruption from timer interrupts,
> or other random system things.  This includes block mq work.
>
> There is another issue with block mq vs. realtime tasks that run
> 100% of the time, which is not uncommon on systems that have CPUs
> dedicated to real time use with isolcpus= and nohz_full=
>
> Specifically, on systems like that, a block work item may never
> get to run, which could lead to filesystems getting stuck forever.
>
> We can avoid both issues by not scheduling blk-mq workqueues on
> cpus in nohz_full mode.
>
> Question for Jens: should we try to spread out the load for
> currently offline and nohz CPUs across the remaining CPUs in
> the system, to get the full benefit of blk-mq in these situations?
>
> If so, do you have any preference on how I should implement that?
>
> Cc: Frederic Weisbecker <fweisbec@redhat.com>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Jens Axboe <axboe@kernel.org>
> Signed-off-by: Rik van Riel <riel@redhat.com>
> ---
>   block/blk-mq.c | 5 +++++
>   1 file changed, 5 insertions(+)
>
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 4f4bea21052e..1004d6817fa4 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -21,6 +21,7 @@
>   #include <linux/sched/sysctl.h>
>   #include <linux/delay.h>
>   #include <linux/crash_dump.h>
> +#include <linux/tick.h>
>
>   #include <trace/events/block.h>
>
> @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
>   		if (!cpu_online(i))
>   			continue;
>
> +		/* Do not schedule work on nohz full dedicated CPUs. */
> +		if (tick_nohz_full_cpu(i))
> +			continue;

Is this CPU ever going to queue IO? If yes, then it needs to be mapped. 
If userspace never runs on it and submits IO, then we'll never run 
completions on it nor schedule the associated workqueue. So I really 
don't see how it doesn't already work, as-is.

-- 
Jens Axboe

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Frederic Weisbecker]

On Tue, Mar 31, 2015 at 09:07:11AM -0600, Jens Axboe wrote:
> On 03/31/2015 08:27 AM, Rik van Riel wrote:
> >CPUs with nohz_full do not want disruption from timer interrupts,
> >or other random system things.  This includes block mq work.
> >
> >There is another issue with block mq vs. realtime tasks that run
> >100% of the time, which is not uncommon on systems that have CPUs
> >dedicated to real time use with isolcpus= and nohz_full=
> >
> >Specifically, on systems like that, a block work item may never
> >get to run, which could lead to filesystems getting stuck forever.
> >
> >We can avoid both issues by not scheduling blk-mq workqueues on
> >cpus in nohz_full mode.
> >
> >Question for Jens: should we try to spread out the load for
> >currently offline and nohz CPUs across the remaining CPUs in
> >the system, to get the full benefit of blk-mq in these situations?
> >
> >If so, do you have any preference on how I should implement that?
> >
> >Cc: Frederic Weisbecker <fweisbec@redhat.com>
> >Cc: Ingo Molnar <mingo@kernel.org>
> >Cc: Jens Axboe <axboe@kernel.org>
> >Signed-off-by: Rik van Riel <riel@redhat.com>
> >---
> >  block/blk-mq.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> >
> >diff --git a/block/blk-mq.c b/block/blk-mq.c
> >index 4f4bea21052e..1004d6817fa4 100644
> >--- a/block/blk-mq.c
> >+++ b/block/blk-mq.c
> >@@ -21,6 +21,7 @@
> >  #include <linux/sched/sysctl.h>
> >  #include <linux/delay.h>
> >  #include <linux/crash_dump.h>
> >+#include <linux/tick.h>
> >
> >  #include <trace/events/block.h>
> >
> >@@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
> >  		if (!cpu_online(i))
> >  			continue;
> >
> >+		/* Do not schedule work on nohz full dedicated CPUs. */
> >+		if (tick_nohz_full_cpu(i))
> >+			continue;
> 
> Is this CPU ever going to queue IO? If yes, then it needs to be mapped. If
> userspace never runs on it and submits IO, then we'll never run completions
> on it nor schedule the associated workqueue. So I really don't see how it
> doesn't already work, as-is.

Well, it's fairly possible that full dynticks CPUs do IO of any sort. Is it possible
to affine these asynchronous works to specific CPU? The usual scheme of full dynticks
is to have CPU 0 handling any kind of housekeeping and other CPUs doing latency or performance
sensitive works that don't want to be disturbed.

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Jens Axboe]

[-- Attachment #1: Type: text/plain, Size: 2785 bytes --]

On 03/31/2015 09:33 AM, Frederic Weisbecker wrote:
> On Tue, Mar 31, 2015 at 09:07:11AM -0600, Jens Axboe wrote:
>> On 03/31/2015 08:27 AM, Rik van Riel wrote:
>>> CPUs with nohz_full do not want disruption from timer interrupts,
>>> or other random system things.  This includes block mq work.
>>>
>>> There is another issue with block mq vs. realtime tasks that run
>>> 100% of the time, which is not uncommon on systems that have CPUs
>>> dedicated to real time use with isolcpus= and nohz_full=
>>>
>>> Specifically, on systems like that, a block work item may never
>>> get to run, which could lead to filesystems getting stuck forever.
>>>
>>> We can avoid both issues by not scheduling blk-mq workqueues on
>>> cpus in nohz_full mode.
>>>
>>> Question for Jens: should we try to spread out the load for
>>> currently offline and nohz CPUs across the remaining CPUs in
>>> the system, to get the full benefit of blk-mq in these situations?
>>>
>>> If so, do you have any preference on how I should implement that?
>>>
>>> Cc: Frederic Weisbecker <fweisbec@redhat.com>
>>> Cc: Ingo Molnar <mingo@kernel.org>
>>> Cc: Jens Axboe <axboe@kernel.org>
>>> Signed-off-by: Rik van Riel <riel@redhat.com>
>>> ---
>>>   block/blk-mq.c | 5 +++++
>>>   1 file changed, 5 insertions(+)
>>>
>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>> index 4f4bea21052e..1004d6817fa4 100644
>>> --- a/block/blk-mq.c
>>> +++ b/block/blk-mq.c
>>> @@ -21,6 +21,7 @@
>>>   #include <linux/sched/sysctl.h>
>>>   #include <linux/delay.h>
>>>   #include <linux/crash_dump.h>
>>> +#include <linux/tick.h>
>>>
>>>   #include <trace/events/block.h>
>>>
>>> @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
>>>   		if (!cpu_online(i))
>>>   			continue;
>>>
>>> +		/* Do not schedule work on nohz full dedicated CPUs. */
>>> +		if (tick_nohz_full_cpu(i))
>>> +			continue;
>>
>> Is this CPU ever going to queue IO? If yes, then it needs to be mapped. If
>> userspace never runs on it and submits IO, then we'll never run completions
>> on it nor schedule the associated workqueue. So I really don't see how it
>> doesn't already work, as-is.
>
> Well, it's fairly possible that full dynticks CPUs do IO of any sort. Is it possible
> to affine these asynchronous works to specific CPU? The usual scheme of full dynticks
> is to have CPU 0 handling any kind of housekeeping and other CPUs doing latency or performance
> sensitive works that don't want to be disturbed.

That'd be easy enough to do, that's how blk-mq handles offline CPUs as 
well. The attached patch is completely untested, but will handle offline 
or nohz CPUs in the same fashion - they will punt to hardware queue 0, 
which is mapped to CPU0 (and others, depending on the queue vs CPU ratio).

-- 
Jens Axboe


[-- Attachment #2: blk-mq-offline-nohz.patch --]
[-- Type: text/x-patch, Size: 2790 bytes --]

diff --git a/block/blk-mq-cpumap.c b/block/blk-mq-cpumap.c
index 5f13f4d0bcce..9cb20d14c6b9 100644
--- a/block/blk-mq-cpumap.c
+++ b/block/blk-mq-cpumap.c
@@ -51,7 +51,10 @@ int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues)
 
 	queue = 0;
 	for_each_possible_cpu(i) {
-		if (!cpu_online(i)) {
+		/*
+		 * Offline or full nohz CPUs get mapped to CPU0
+		 */
+		if (blk_mq_cpu_offline(i)) {
 			map[i] = 0;
 			continue;
 		}
diff --git a/block/blk-mq.c b/block/blk-mq.c
index b7b8933ec241..ec0de2871950 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -366,7 +366,7 @@ static void blk_mq_ipi_complete_request(struct request *rq)
 	if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 		shared = cpus_share_cache(cpu, ctx->cpu);
 
-	if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
+	if (cpu != ctx->cpu && !shared && !blk_mq_cpu_offline(ctx->cpu)) {
 		rq->csd.func = __blk_mq_complete_request_remote;
 		rq->csd.info = rq;
 		rq->csd.flags = 0;
@@ -1022,7 +1022,7 @@ void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 	struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 
 	current_ctx = blk_mq_get_ctx(q);
-	if (!cpu_online(ctx->cpu))
+	if (blk_mq_cpu_offline(ctx->cpu))
 		rq->mq_ctx = ctx = current_ctx;
 
 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
@@ -1051,7 +1051,7 @@ static void blk_mq_insert_requests(struct request_queue *q,
 
 	current_ctx = blk_mq_get_ctx(q);
 
-	if (!cpu_online(ctx->cpu))
+	if (blk_mq_cpu_offline(ctx->cpu))
 		ctx = current_ctx;
 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
 
@@ -1757,7 +1757,7 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
 		__ctx->queue = q;
 
 		/* If the cpu isn't online, the cpu is mapped to first hctx */
-		if (!cpu_online(i))
+		if (blk_mq_cpu_offline(i))
 			continue;
 
 		hctx = q->mq_ops->map_queue(q, i);
@@ -1789,7 +1789,7 @@ static void blk_mq_map_swqueue(struct request_queue *q)
 	 */
 	queue_for_each_ctx(q, ctx, i) {
 		/* If the cpu isn't online, the cpu is mapped to first hctx */
-		if (!cpu_online(i))
+		if (blk_mq_cpu_offline(i))
 			continue;
 
 		hctx = q->mq_ops->map_queue(q, i);
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 6a48c4c0d8a2..443dc8e0ea24 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -1,6 +1,8 @@
 #ifndef INT_BLK_MQ_H
 #define INT_BLK_MQ_H
 
+#include <linux/tick.h>
+
 struct blk_mq_tag_set;
 
 struct blk_mq_ctx {
@@ -123,4 +125,13 @@ static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
 	return hctx->nr_ctx && hctx->tags;
 }
 
+/*
+ * If the CPU is offline or is a nohz CPU, we will remap any IO processing
+ * to the first hardware queue.
+ */
+static inline bool blk_mq_cpu_offline(const unsigned int cpu)
+{
+	return !cpu_online(cpu) || tick_nohz_full_cpu(cpu);
+}
+
 #endif

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Rik van Riel]

On 03/31/2015 11:43 AM, Jens Axboe wrote:

> That'd be easy enough to do, that's how blk-mq handles offline CPUs as
> well. The attached patch is completely untested, but will handle offline
> or nohz CPUs in the same fashion - they will punt to hardware queue 0,
> which is mapped to CPU0 (and others, depending on the queue vs CPU ratio).

I have done some sanity testing with your patch,
starting a KVM guest with vcpus and emulator threads
all pinned to nohz_full cpus.

The guest is still able to do disk IO, so things
appear to work...

Thanks for looking into this, Jens.

Tested-by: Rik van Riel <riel@redhat.com>

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Jens Axboe]

On 04/01/2015 10:12 AM, Rik van Riel wrote:
> On 03/31/2015 11:43 AM, Jens Axboe wrote:
>
>> That'd be easy enough to do, that's how blk-mq handles offline CPUs as
>> well. The attached patch is completely untested, but will handle offline
>> or nohz CPUs in the same fashion - they will punt to hardware queue 0,
>> which is mapped to CPU0 (and others, depending on the queue vs CPU ratio).
>
> I have done some sanity testing with your patch,
> starting a KVM guest with vcpus and emulator threads
> all pinned to nohz_full cpus.
>
> The guest is still able to do disk IO, so things
> appear to work...
>
> Thanks for looking into this, Jens.
>
> Tested-by: Rik van Riel <riel@redhat.com>

Great thanks, I'll do some sanity testing here too and get it applied 
for 4.1.

-- 
Jens Axboe

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Marcelo Tosatti]

On Tue, Mar 31, 2015 at 10:27:26AM -0400, Rik van Riel wrote:
> CPUs with nohz_full do not want disruption from timer interrupts,
> or other random system things.  This includes block mq work.
> 
> There is another issue with block mq vs. realtime tasks that run
> 100% of the time, which is not uncommon on systems that have CPUs
> dedicated to real time use with isolcpus= and nohz_full=
> 
> Specifically, on systems like that, a block work item may never
> get to run, which could lead to filesystems getting stuck forever.
> 
> We can avoid both issues by not scheduling blk-mq workqueues on
> cpus in nohz_full mode.
> 
> Question for Jens: should we try to spread out the load for
> currently offline and nohz CPUs across the remaining CPUs in
> the system, to get the full benefit of blk-mq in these situations?
> 
> If so, do you have any preference on how I should implement that?
> 
> Cc: Frederic Weisbecker <fweisbec@redhat.com>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Jens Axboe <axboe@kernel.org>
> Signed-off-by: Rik van Riel <riel@redhat.com>
> ---
>  block/blk-mq.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 4f4bea21052e..1004d6817fa4 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -21,6 +21,7 @@
>  #include <linux/sched/sysctl.h>
>  #include <linux/delay.h>
>  #include <linux/crash_dump.h>
> +#include <linux/tick.h>
>  
>  #include <trace/events/block.h>
>  
> @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
>  		if (!cpu_online(i))
>  			continue;
>  
> +		/* Do not schedule work on nohz full dedicated CPUs. */
> +		if (tick_nohz_full_cpu(i))
> +			continue;
> +
>  		hctx = q->mq_ops->map_queue(q, i);
>  		cpumask_set_cpu(i, hctx->cpumask);
>  		hctx->nr_ctx++;

Rik,

I suppose any bound workqueue queued on isolated CPUs should be moved at 
queue time to other CPUs (sacrifficing performance).

So that by doing "queue_work" on an isolated CPU would move that 
work somewhere else.

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Jens Axboe]

On 03/31/2015 05:17 PM, Marcelo Tosatti wrote:
> On Tue, Mar 31, 2015 at 10:27:26AM -0400, Rik van Riel wrote:
>> CPUs with nohz_full do not want disruption from timer interrupts,
>> or other random system things.  This includes block mq work.
>>
>> There is another issue with block mq vs. realtime tasks that run
>> 100% of the time, which is not uncommon on systems that have CPUs
>> dedicated to real time use with isolcpus= and nohz_full=
>>
>> Specifically, on systems like that, a block work item may never
>> get to run, which could lead to filesystems getting stuck forever.
>>
>> We can avoid both issues by not scheduling blk-mq workqueues on
>> cpus in nohz_full mode.
>>
>> Question for Jens: should we try to spread out the load for
>> currently offline and nohz CPUs across the remaining CPUs in
>> the system, to get the full benefit of blk-mq in these situations?
>>
>> If so, do you have any preference on how I should implement that?
>>
>> Cc: Frederic Weisbecker <fweisbec@redhat.com>
>> Cc: Ingo Molnar <mingo@kernel.org>
>> Cc: Jens Axboe <axboe@kernel.org>
>> Signed-off-by: Rik van Riel <riel@redhat.com>
>> ---
>>   block/blk-mq.c | 5 +++++
>>   1 file changed, 5 insertions(+)
>>
>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>> index 4f4bea21052e..1004d6817fa4 100644
>> --- a/block/blk-mq.c
>> +++ b/block/blk-mq.c
>> @@ -21,6 +21,7 @@
>>   #include <linux/sched/sysctl.h>
>>   #include <linux/delay.h>
>>   #include <linux/crash_dump.h>
>> +#include <linux/tick.h>
>>
>>   #include <trace/events/block.h>
>>
>> @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
>>   		if (!cpu_online(i))
>>   			continue;
>>
>> +		/* Do not schedule work on nohz full dedicated CPUs. */
>> +		if (tick_nohz_full_cpu(i))
>> +			continue;
>> +
>>   		hctx = q->mq_ops->map_queue(q, i);
>>   		cpumask_set_cpu(i, hctx->cpumask);
>>   		hctx->nr_ctx++;
>
> Rik,
>
> I suppose any bound workqueue queued on isolated CPUs should be moved at
> queue time to other CPUs (sacrifficing performance).
>
> So that by doing "queue_work" on an isolated CPU would move that
> work somewhere else.

That wont work for blk-mq, we rely on the characteristics of bound 
workqueues. So it would have to be handled up front, like in the patch I 
sent out.

-- 
Jens Axboe

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Rik van Riel]

On 04/01/2015 10:36 AM, Jens Axboe wrote:

> That wont work for blk-mq, we rely on the characteristics of bound
> workqueues. So it would have to be handled up front, like in the patch I
> sent out.

Your patch from yesterday looks good.  I'll take it for
a spin today (had some other stuff on my plate yesterday).

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Jens Axboe]

On 04/01/2015 08:45 AM, Rik van Riel wrote:
> On 04/01/2015 10:36 AM, Jens Axboe wrote:
>
>> That wont work for blk-mq, we rely on the characteristics of bound
>> workqueues. So it would have to be handled up front, like in the patch I
>> sent out.
>
> Your patch from yesterday looks good.  I'll take it for
> a spin today (had some other stuff on my plate yesterday).

Great, thanks. If we can get it tested, then I can get it queued up for 4.1.


-- 
Jens Axboe

Re: [PATCH RFC] nohz,blk-mq: do not create blk-mq workqueues on nohz dedicated CPUs

[Mike Galbraith]

On Tue, 2015-03-31 at 10:27 -0400, Rik van Riel wrote:
> CPUs with nohz_full do not want disruption from timer interrupts,
> or other random system things.  This includes block mq work.
> 
> There is another issue with block mq vs. realtime tasks that run
> 100% of the time, which is not uncommon on systems that have CPUs
> dedicated to real time use with isolcpus= and nohz_full=
> 
> Specifically, on systems like that, a block work item may never
> get to run, which could lead to filesystems getting stuck forever.

What reason is there to run a compute hog as RT if it's the only task 
on an isolated core?  In an RT kernel, PI may be a reason, but if a 
task is so critical that it needs bare metal, it had better not be.

        -Mike