When running RT workloads in isolated CPUs, many cases of deadline misses
are caused by remote CPU requests such as smp_call_function*().
For those cases, having the names of those functions running around the
deadline miss moment could help (a lot) finding a target for the next
improvements.
Add tracepoints for acquiring the function name & csd before entry and
after returning from the remote-cpu requested function.
Also, add tracepoints on the remote cpus requesting them.
Signed-off-by: Leonardo Bras <[email protected]>
---
Changes since RFCv2:
- Fixed some spacing issues and trace calls
Changes since RFCv1:
- Implemented trace_csd_queue_cpu() as suggested by Valentin Schneider
- Using EVENT_CLASS in order to avoid duplication
- Introduced new helper: csd_do_func()
- Name change from smp_call_function_* to csd_function_*
- Rebased on top of torvalds/master
include/trace/events/smp.h | 72 ++++++++++++++++++++++++++++++++++++++
kernel/smp.c | 41 +++++++++++++---------
2 files changed, 96 insertions(+), 17 deletions(-)
create mode 100644 include/trace/events/smp.h
diff --git a/include/trace/events/smp.h b/include/trace/events/smp.h
new file mode 100644
index 000000000000..c304318a0203
--- /dev/null
+++ b/include/trace/events/smp.h
@@ -0,0 +1,72 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM smp
+
+#if !defined(_TRACE_SMP_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_SMP_H
+
+#include <linux/tracepoint.h>
+
+TRACE_EVENT(csd_queue_cpu,
+
+ TP_PROTO(const unsigned int cpu,
+ unsigned long callsite,
+ smp_call_func_t func,
+ call_single_data_t *csd),
+
+ TP_ARGS(cpu, callsite, func, csd),
+
+ TP_STRUCT__entry(
+ __field(unsigned int, cpu)
+ __field(void *, callsite)
+ __field(void *, func)
+ __field(void *, csd)
+ ),
+
+ TP_fast_assign(
+ __entry->cpu = cpu;
+ __entry->callsite = (void *)callsite;
+ __entry->func = func;
+ __entry->csd = csd;
+ ),
+
+ TP_printk("cpu=%u callsite=%pS func=%pS csd=%p",
+ __entry->cpu, __entry->callsite, __entry->func, __entry->csd)
+);
+
+/*
+ * Tracepoints for a function which is called as an effect of smp_call_function.*
+ */
+DECLARE_EVENT_CLASS(csd_function,
+
+ TP_PROTO(smp_call_func_t func, call_single_data_t *csd),
+
+ TP_ARGS(func, csd),
+
+ TP_STRUCT__entry(
+ __field(void *, func)
+ __field(void *, csd)
+ ),
+
+ TP_fast_assign(
+ __entry->func = func;
+ __entry->csd = csd;
+ ),
+
+ TP_printk("function %ps, csd = %p", __entry->func, __entry->csd)
+);
+
+DEFINE_EVENT(csd_function, csd_function_entry,
+ TP_PROTO(smp_call_func_t func, call_single_data_t *csd),
+ TP_ARGS(func, csd)
+);
+
+DEFINE_EVENT(csd_function, csd_function_exit,
+ TP_PROTO(smp_call_func_t func, call_single_data_t *csd),
+ TP_ARGS(func, csd)
+);
+
+#endif /* _TRACE_SMP_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>
diff --git a/kernel/smp.c b/kernel/smp.c
index ab3e5dad6cfe..cada433c5c1f 100644
--- a/kernel/smp.c
+++ b/kernel/smp.c
@@ -27,6 +27,9 @@
#include <linux/jump_label.h>
#include <trace/events/ipi.h>
+#define CREATE_TRACE_POINTS
+#include <trace/events/smp.h>
+#undef CREATE_TRACE_POINTS
#include "smpboot.h"
#include "sched/smp.h"
@@ -121,6 +124,14 @@ send_call_function_ipi_mask(struct cpumask *mask)
arch_send_call_function_ipi_mask(mask);
}
+static __always_inline void
+csd_do_func(smp_call_func_t func, void *info, call_single_data_t *csd)
+{
+ trace_csd_function_entry(func, csd);
+ func(info);
+ trace_csd_function_exit(func, csd);
+}
+
#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
static DEFINE_STATIC_KEY_MAYBE(CONFIG_CSD_LOCK_WAIT_DEBUG_DEFAULT, csdlock_debug_enabled);
@@ -329,7 +340,7 @@ void __smp_call_single_queue(int cpu, struct llist_node *node)
* even if we haven't sent the smp_call IPI yet (e.g. the stopper
* executes migration_cpu_stop() on the remote CPU).
*/
- if (trace_ipi_send_cpu_enabled()) {
+ if (trace_csd_queue_cpu_enabled()) {
call_single_data_t *csd;
smp_call_func_t func;
@@ -337,7 +348,7 @@ void __smp_call_single_queue(int cpu, struct llist_node *node)
func = CSD_TYPE(csd) == CSD_TYPE_TTWU ?
sched_ttwu_pending : csd->func;
- trace_ipi_send_cpu(cpu, _RET_IP_, func);
+ trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
}
/*
@@ -375,7 +386,7 @@ static int generic_exec_single(int cpu, struct __call_single_data *csd)
csd_lock_record(csd);
csd_unlock(csd);
local_irq_save(flags);
- func(info);
+ csd_do_func(func, info, csd);
csd_lock_record(NULL);
local_irq_restore(flags);
return 0;
@@ -477,7 +488,7 @@ static void __flush_smp_call_function_queue(bool warn_cpu_offline)
}
csd_lock_record(csd);
- func(info);
+ csd_do_func(func, info, csd);
csd_unlock(csd);
csd_lock_record(NULL);
} else {
@@ -508,7 +519,7 @@ static void __flush_smp_call_function_queue(bool warn_cpu_offline)
csd_lock_record(csd);
csd_unlock(csd);
- func(info);
+ csd_do_func(func, info, csd);
csd_lock_record(NULL);
} else if (type == CSD_TYPE_IRQ_WORK) {
irq_work_single(csd);
@@ -522,8 +533,10 @@ static void __flush_smp_call_function_queue(bool warn_cpu_offline)
/*
* Third; only CSD_TYPE_TTWU is left, issue those.
*/
- if (entry)
- sched_ttwu_pending(entry);
+ if (entry) {
+ csd = llist_entry(entry, typeof(*csd), node.llist);
+ csd_do_func(sched_ttwu_pending, entry, csd);
+ }
}
@@ -728,7 +741,7 @@ static void smp_call_function_many_cond(const struct cpumask *mask,
int cpu, last_cpu, this_cpu = smp_processor_id();
struct call_function_data *cfd;
bool wait = scf_flags & SCF_WAIT;
- int nr_cpus = 0, nr_queued = 0;
+ int nr_cpus = 0;
bool run_remote = false;
bool run_local = false;
@@ -786,21 +799,15 @@ static void smp_call_function_many_cond(const struct cpumask *mask,
csd->node.src = smp_processor_id();
csd->node.dst = cpu;
#endif
+ trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
+
if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
nr_cpus++;
last_cpu = cpu;
}
- nr_queued++;
}
- /*
- * Trace each smp_function_call_*() as an IPI, actual IPIs
- * will be traced with func==generic_smp_call_function_single_ipi().
- */
- if (nr_queued)
- trace_ipi_send_cpumask(cfd->cpumask, _RET_IP_, func);
-
/*
* Choose the most efficient way to send an IPI. Note that the
* number of CPUs might be zero due to concurrent changes to the
@@ -816,7 +823,7 @@ static void smp_call_function_many_cond(const struct cpumask *mask,
unsigned long flags;
local_irq_save(flags);
- func(info);
+ csd_do_func(func, info, NULL);
local_irq_restore(flags);
}
--
2.40.1
On 10/05/23 20:01, Leonardo Bras wrote:
> When running RT workloads in isolated CPUs, many cases of deadline misses
> are caused by remote CPU requests such as smp_call_function*().
>
> For those cases, having the names of those functions running around the
> deadline miss moment could help (a lot) finding a target for the next
> improvements.
>
> Add tracepoints for acquiring the function name & csd before entry and
> after returning from the remote-cpu requested function.
>
> Also, add tracepoints on the remote cpus requesting them.
Per the v1 thread, I got some context as to why we want this, but this
changelog is quite sparse on the subject. How about:
"""
The recently added ipi_send_{cpu,cpumask} tracepoints allow finding sources
of IPIs targeting CPUs running latency-sensitive applications.
For NOHZ_FULL CPUs, all IPIs are interference, and those tracepoints are
sufficient to find them and work on getting rid of them. In some setups
however, not *all* IPIs are to be suppressed, but long-running IPI
callbacks can still be problematic.
Add a pair of tracepoints to mark the start and end of processing a CSD IPI
callback, similar to what exists for softirq, workqueue or timer callbacks.
"""
And you can probably split the csd_queue_cpu TP into a separate commit,
with something along the lines of:
"""
Add a tracepoint for when a CSD is queued to a remote CPU's
call_single_queue. This allows finding exactly which CPU queued a given CSD
when looking at a csd_function_{entry,exit} event, and also enables us to
accurately measure IPI delivery time with e.g. a synthetic event:
$ echo 'hist:keys=cpu,csd.hex:ts=common_timestamp.usecs' >\
/sys/kernel/tracing/events/smp/csd_queue_cpu/trigger
$ echo 'csd_latency unsigned int dst_cpu; unsigned long csd; u64 time' >\
/sys/kernel/tracing/synthetic_events
$ echo \
'hist:keys=common_cpu,csd.hex:'\
'time=common_timestamp.usecs-$ts:'\
'onmatch(smp.csd_queue_cpu).trace(csd_latency,common_cpu,csd,$time)' >\
/sys/kernel/tracing/events/smp/csd_function_entry/trigger
$ trace-cmd record -e 'synthetic:csd_latency' hackbench
$ trace-cmd report
<...>-467 [001] 21.824263: csd_queue_cpu: cpu=0 callsite=try_to_wake_up+0x2ea func=sched_ttwu_pending csd=0xffff8880076148b8
<...>-467 [001] 21.824280: ipi_send_cpu: cpu=0 callsite=try_to_wake_up+0x2ea callback=generic_smp_call_function_single_interrupt+0x0
<...>-489 [000] 21.824299: csd_function_entry: func=sched_ttwu_pending csd=0xffff8880076148b8
<...>-489 [000] 21.824320: csd_latency: dst_cpu=0, csd=18446612682193848504, time=36
"""
>
> Signed-off-by: Leonardo Bras <[email protected]>
Overall I like the CSD vs IPI split, it's a saner approach than logging it
all as IPIs, even if it does generate more events as we have to emit an
event for every csd queued (i.e. no _cpumask() variant is possible).
Some nitpicks below, and one other thing: are we happy with the smp event
namespace, and with the new TP names? Should the namespace be csd instead,
to match the trace_<namespace>_* nomenclature?
> ---
>
> Changes since RFCv2:
> - Fixed some spacing issues and trace calls
>
> Changes since RFCv1:
> - Implemented trace_csd_queue_cpu() as suggested by Valentin Schneider
> - Using EVENT_CLASS in order to avoid duplication
> - Introduced new helper: csd_do_func()
> - Name change from smp_call_function_* to csd_function_*
> - Rebased on top of torvalds/master
>
> include/trace/events/smp.h | 72 ++++++++++++++++++++++++++++++++++++++
> kernel/smp.c | 41 +++++++++++++---------
> 2 files changed, 96 insertions(+), 17 deletions(-)
> create mode 100644 include/trace/events/smp.h
>
> diff --git a/include/trace/events/smp.h b/include/trace/events/smp.h
> new file mode 100644
> index 000000000000..c304318a0203
> --- /dev/null
> +++ b/include/trace/events/smp.h
> @@ -0,0 +1,72 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#undef TRACE_SYSTEM
> +#define TRACE_SYSTEM smp
> +
> +#if !defined(_TRACE_SMP_H) || defined(TRACE_HEADER_MULTI_READ)
> +#define _TRACE_SMP_H
> +
> +#include <linux/tracepoint.h>
> +
> +TRACE_EVENT(csd_queue_cpu,
> +
> + TP_PROTO(const unsigned int cpu,
> + unsigned long callsite,
> + smp_call_func_t func,
> + call_single_data_t *csd),
> +
> + TP_ARGS(cpu, callsite, func, csd),
> +
> + TP_STRUCT__entry(
> + __field(unsigned int, cpu)
> + __field(void *, callsite)
> + __field(void *, func)
> + __field(void *, csd)
> + ),
> +
> + TP_fast_assign(
> + __entry->cpu = cpu;
> + __entry->callsite = (void *)callsite;
> + __entry->func = func;
> + __entry->csd = csd;
> + ),
> +
> + TP_printk("cpu=%u callsite=%pS func=%pS csd=%p",
^^^^^^^^
This can be func=%ps, offsets should always be 0 so not very useful (I know
I put %pS for the IPI TPs, that's a force of habit...)
> + __entry->cpu, __entry->callsite, __entry->func, __entry->csd)
> +);
> +
> +/*
> + * Tracepoints for a function which is called as an effect of smp_call_function.*
> + */
> +DECLARE_EVENT_CLASS(csd_function,
> +
> + TP_PROTO(smp_call_func_t func, call_single_data_t *csd),
> +
> + TP_ARGS(func, csd),
> +
> + TP_STRUCT__entry(
> + __field(void *, func)
> + __field(void *, csd)
> + ),
> +
> + TP_fast_assign(
> + __entry->func = func;
> + __entry->csd = csd;
> + ),
> +
> + TP_printk("function %ps, csd = %p", __entry->func, __entry->csd)
To match the style of the other TPs, that should be:
TP_printk("func=%ps csd=%p", __entry->func, __entry->csd)
On Thu, 2023-05-11 at 11:56 +0100, Valentin Schneider wrote:
> On 10/05/23 20:01, Leonardo Bras wrote:
> > When running RT workloads in isolated CPUs, many cases of deadline misses
> > are caused by remote CPU requests such as smp_call_function*().
> >
> > For those cases, having the names of those functions running around the
> > deadline miss moment could help (a lot) finding a target for the next
> > improvements.
> >
> > Add tracepoints for acquiring the function name & csd before entry and
> > after returning from the remote-cpu requested function.
> >
> > Also, add tracepoints on the remote cpus requesting them.
>
> Per the v1 thread, I got some context as to why we want this, but this
> changelog is quite sparse on the subject. How about:
>
> """
> The recently added ipi_send_{cpu,cpumask} tracepoints allow finding sources
> of IPIs targeting CPUs running latency-sensitive applications.
>
> For NOHZ_FULL CPUs, all IPIs are interference, and those tracepoints are
> sufficient to find them and work on getting rid of them. In some setups
> however, not *all* IPIs are to be suppressed, but long-running IPI
> callbacks can still be problematic.
>
> Add a pair of tracepoints to mark the start and end of processing a CSD IPI
> callback, similar to what exists for softirq, workqueue or timer callbacks.
> """
>
> And you can probably split the csd_queue_cpu TP into a separate commit,
> with something along the lines of:
>
> """
> Add a tracepoint for when a CSD is queued to a remote CPU's
> call_single_queue. This allows finding exactly which CPU queued a given CSD
> when looking at a csd_function_{entry,exit} event, and also enables us to
> accurately measure IPI delivery time with e.g. a synthetic event:
>
> $ echo 'hist:keys=cpu,csd.hex:ts=common_timestamp.usecs' >\
> /sys/kernel/tracing/events/smp/csd_queue_cpu/trigger
> $ echo 'csd_latency unsigned int dst_cpu; unsigned long csd; u64 time' >\
> /sys/kernel/tracing/synthetic_events
> $ echo \
> 'hist:keys=common_cpu,csd.hex:'\
> 'time=common_timestamp.usecs-$ts:'\
> 'onmatch(smp.csd_queue_cpu).trace(csd_latency,common_cpu,csd,$time)' >\
> /sys/kernel/tracing/events/smp/csd_function_entry/trigger
>
> $ trace-cmd record -e 'synthetic:csd_latency' hackbench
> $ trace-cmd report
> <...>-467 [001] 21.824263: csd_queue_cpu: cpu=0 callsite=try_to_wake_up+0x2ea func=sched_ttwu_pending csd=0xffff8880076148b8
> <...>-467 [001] 21.824280: ipi_send_cpu: cpu=0 callsite=try_to_wake_up+0x2ea callback=generic_smp_call_function_single_interrupt+0x0
> <...>-489 [000] 21.824299: csd_function_entry: func=sched_ttwu_pending csd=0xffff8880076148b8
> <...>-489 [000] 21.824320: csd_latency: dst_cpu=0, csd=18446612682193848504, time=36
> """
Sure, I will split the patch and add the commit msg as suggested.
>
> >
> > Signed-off-by: Leonardo Bras <[email protected]>
>
> Overall I like the CSD vs IPI split, it's a saner approach than logging it
> all as IPIs, even if it does generate more events as we have to emit an
> event for every csd queued (i.e. no _cpumask() variant is possible).
>
> Some nitpicks below, and one other thing: are we happy with the smp event
> namespace, and with the new TP names? Should the namespace be csd instead,
> to match the trace_<namespace>_* nomenclature?
I selected "smp.h" filename because those tracepoints are getting used on smp.c.
I think keeping a generic name is good for any future tracepoints people end up
designing for smp.c, but I am open to discussion.
>
> > ---
> >
> > Changes since RFCv2:
> > - Fixed some spacing issues and trace calls
> >
> > Changes since RFCv1:
> > - Implemented trace_csd_queue_cpu() as suggested by Valentin Schneider
> > - Using EVENT_CLASS in order to avoid duplication
> > - Introduced new helper: csd_do_func()
> > - Name change from smp_call_function_* to csd_function_*
> > - Rebased on top of torvalds/master
> >
> > include/trace/events/smp.h | 72 ++++++++++++++++++++++++++++++++++++++
> > kernel/smp.c | 41 +++++++++++++---------
> > 2 files changed, 96 insertions(+), 17 deletions(-)
> > create mode 100644 include/trace/events/smp.h
> >
> > diff --git a/include/trace/events/smp.h b/include/trace/events/smp.h
> > new file mode 100644
> > index 000000000000..c304318a0203
> > --- /dev/null
> > +++ b/include/trace/events/smp.h
> > @@ -0,0 +1,72 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +#undef TRACE_SYSTEM
> > +#define TRACE_SYSTEM smp
> > +
> > +#if !defined(_TRACE_SMP_H) || defined(TRACE_HEADER_MULTI_READ)
> > +#define _TRACE_SMP_H
> > +
> > +#include <linux/tracepoint.h>
> > +
> > +TRACE_EVENT(csd_queue_cpu,
> > +
> > + TP_PROTO(const unsigned int cpu,
> > + unsigned long callsite,
> > + smp_call_func_t func,
> > + call_single_data_t *csd),
> > +
> > + TP_ARGS(cpu, callsite, func, csd),
> > +
> > + TP_STRUCT__entry(
> > + __field(unsigned int, cpu)
> > + __field(void *, callsite)
> > + __field(void *, func)
> > + __field(void *, csd)
> > + ),
> > +
> > + TP_fast_assign(
> > + __entry->cpu = cpu;
> > + __entry->callsite = (void *)callsite;
> > + __entry->func = func;
> > + __entry->csd = csd;
> > + ),
> > +
> > + TP_printk("cpu=%u callsite=%pS func=%pS csd=%p",
> ^^^^^^^^
> This can be func=%ps, offsets should always be 0 so not very useful (I know
> I put %pS for the IPI TPs, that's a force of habit...)
>
I agree, let's do as suggested.
> > + __entry->cpu, __entry->callsite, __entry->func, __entry->csd)
> > +);
>
> > +
> > +/*
> > + * Tracepoints for a function which is called as an effect of smp_call_function.*
> > + */
> > +DECLARE_EVENT_CLASS(csd_function,
> > +
> > + TP_PROTO(smp_call_func_t func, call_single_data_t *csd),
> > +
> > + TP_ARGS(func, csd),
> > +
> > + TP_STRUCT__entry(
> > + __field(void *, func)
> > + __field(void *, csd)
> > + ),
> > +
> > + TP_fast_assign(
> > + __entry->func = func;
> > + __entry->csd = csd;
> > + ),
> > +
> > + TP_printk("function %ps, csd = %p", __entry->func, __entry->csd)
>
> To match the style of the other TPs, that should be:
>
> TP_printk("func=%ps csd=%p", __entry->func, __entry->csd)
>
Done!
Thanks for reviewing!
Leo