After enabling CONFIG_IRQ_TIME_ACCOUNTING to monitor IRQ pressure in our
container environment, we observed several noticeable behavioral changes.

One of our IRQ-heavy services, such as Redis, reported a significant
reduction in CPU usage after upgrading to the new kernel with
CONFIG_IRQ_TIME_ACCOUNTING enabled. However, despite adding more threads
to handle an increased workload, the CPU usage could not be raised. In
other words, even though the container’s CPU usage appeared low, it was
unable to process more workloads to utilize additional CPU resources, which
caused issues.

This behavior can be demonstrated using netperf:

  function start_server() {
      for j in `seq 1 3`; do
          netserver -p $[12345+j] > /dev/null &
      done
  }

  server_ip=$1
  function start_client() {
    # That applies to cgroup2 as well.
    mkdir -p /sys/fs/cgroup/cpuacct/test
    echo $$ > /sys/fs/cgroup/cpuacct/test/cgroup.procs
    for j in `seq 1 3`; do
        port=$[12345+j]
        taskset -c 0 netperf -H ${server_ip} -l ${run_time:-30000}   \
                -t TCP_STREAM -p $port -- -D -m 1k -M 1K -s 8k -S 8k \
                > /dev/null &
    done
  }

  start_server
  start_client

We can verify the CPU usage of the test cgroup using cpuacct.stat. The
output shows:

  system: 53
  user: 2

The CPU usage of the cgroup is relatively low at around 55%, but this usage
doesn't increase, even with more netperf tasks. The reason is that CPU0 is
at 100% utilization, as confirmed by mpstat:

  02:56:22 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
  02:56:23 PM    0    0.99    0.00   55.45    0.00    0.99   42.57    0.00    0.00    0.00    0.00

  02:56:23 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
  02:56:24 PM    0    2.00    0.00   55.00    0.00    0.00   43.00    0.00    0.00    0.00    0.00

It is clear that the %soft is not accounted into the cgroup of the
interrupted task. This behavior is unexpected. We should account for IRQ
time to the cgroup to reflect the pressure the group is under.

After a thorough analysis, I discovered that this change in behavior is due
to commit 305e6835e055 ("sched: Do not account irq time to current task"),
which altered whether IRQ time should be charged to the interrupted task.
While I agree that a task should not be penalized by random interrupts, the
task itself cannot progress while interrupted. Therefore, the interrupted
time should be reported to the user.

The system metric in cpuacct.stat is crucial in indicating whether a
container is under heavy system pressure, including IRQ/softirq activity.
Hence, IRQ/softirq time should be accounted for in the cpuacct system
usage, which also applies to cgroup2’s rstat.

This patch reintroduces IRQ/softirq accounting to cgroups.

Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
---
 kernel/sched/core.c  | 33 +++++++++++++++++++++++++++++++--
 kernel/sched/psi.c   | 14 +++-----------
 kernel/sched/stats.h |  7 ++++---
 3 files changed, 38 insertions(+), 16 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 06a06f0897c3..5ed2c5c8c911 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -5579,6 +5579,35 @@ __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
 #endif /* CONFIG_SCHED_DEBUG */
 
+#ifdef CONFIG_IRQ_TIME_ACCOUNTING
+static void account_irqtime(struct rq *rq, struct task_struct *curr,
+			    struct task_struct *prev)
+{
+	int cpu = smp_processor_id();
+	s64 delta;
+	u64 irq;
+
+	if (!static_branch_likely(&sched_clock_irqtime))
+		return;
+
+	irq = irq_time_read(cpu);
+	delta = (s64)(irq - rq->psi_irq_time);
+	if (delta < 0)
+		return;
+
+	rq->psi_irq_time = irq;
+	psi_account_irqtime(rq, curr, prev, delta);
+	cgroup_account_cputime(curr, delta);
+	/* We account both softirq and irq into softirq */
+	cgroup_account_cputime_field(curr, CPUTIME_SOFTIRQ, delta);
+}
+#else
+static inline void account_irqtime(struct rq *rq, struct task_struct *curr,
+				   struct task_struct *prev)
+{
+}
+#endif
+
 /*
  * This function gets called by the timer code, with HZ frequency.
  * We call it with interrupts disabled.
@@ -5600,7 +5629,7 @@ void sched_tick(void)
 	rq_lock(rq, &rf);
 
 	curr = rq->curr;
-	psi_account_irqtime(rq, curr, NULL);
+	account_irqtime(rq, curr, NULL);
 
 	update_rq_clock(rq);
 	hw_pressure = arch_scale_hw_pressure(cpu_of(rq));
@@ -6683,7 +6712,7 @@ static void __sched notrace __schedule(int sched_mode)
 		++*switch_count;
 
 		migrate_disable_switch(rq, prev);
-		psi_account_irqtime(rq, prev, next);
+		account_irqtime(rq, prev, next);
 		psi_sched_switch(prev, next, block);
 
 		trace_sched_switch(preempt, prev, next, prev_state);
diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
index 4d26a106f03b..1adb41b2ae1d 100644
--- a/kernel/sched/psi.c
+++ b/kernel/sched/psi.c
@@ -990,16 +990,14 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 }
 
 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
-void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev)
+void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev,
+			 s64 delta)
 {
 	int cpu = task_cpu(curr);
 	struct psi_group *group;
 	struct psi_group_cpu *groupc;
-	s64 delta;
-	u64 irq;
 
-	if (static_branch_likely(&psi_disabled) ||
-	    !static_branch_likely(&sched_clock_irqtime))
+	if (static_branch_likely(&psi_disabled))
 		return;
 
 	if (!curr->pid)
@@ -1010,12 +1008,6 @@ void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_st
 	if (prev && task_psi_group(prev) == group)
 		return;
 
-	irq = irq_time_read(cpu);
-	delta = (s64)(irq - rq->psi_irq_time);
-	if (delta < 0)
-		return;
-	rq->psi_irq_time = irq;
-
 	do {
 		u64 now;
 
diff --git a/kernel/sched/stats.h b/kernel/sched/stats.h
index 767e098a3bd1..17eefe5876a5 100644
--- a/kernel/sched/stats.h
+++ b/kernel/sched/stats.h
@@ -111,10 +111,11 @@ void psi_task_change(struct task_struct *task, int clear, int set);
 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 		     bool sleep);
 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
-void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev);
+void psi_account_irqtime(struct rq *rq, struct task_struct *curr,
+			 struct task_struct *prev, s64 delta);
 #else
 static inline void psi_account_irqtime(struct rq *rq, struct task_struct *curr,
-				       struct task_struct *prev) {}
+				       struct task_struct *prev, s64 delta) {}
 #endif /*CONFIG_IRQ_TIME_ACCOUNTING */
 /*
  * PSI tracks state that persists across sleeps, such as iowaits and
@@ -215,7 +216,7 @@ static inline void psi_sched_switch(struct task_struct *prev,
 				    struct task_struct *next,
 				    bool sleep) {}
 static inline void psi_account_irqtime(struct rq *rq, struct task_struct *curr,
-				       struct task_struct *prev) {}
+				       struct task_struct *prev, s64 delta) {}
 #endif /* CONFIG_PSI */
 
 #ifdef CONFIG_SCHED_INFO
-- 
2.43.5

On Fri, Nov 01, 2024 at 11:17:50AM +0800, Yafang Shao wrote:
> After enabling CONFIG_IRQ_TIME_ACCOUNTING to monitor IRQ pressure in our
> container environment, we observed several noticeable behavioral changes.
> 
> One of our IRQ-heavy services, such as Redis, reported a significant
> reduction in CPU usage after upgrading to the new kernel with
> CONFIG_IRQ_TIME_ACCOUNTING enabled. However, despite adding more threads
> to handle an increased workload, the CPU usage could not be raised. In
> other words, even though the container’s CPU usage appeared low, it was
> unable to process more workloads to utilize additional CPU resources, which
> caused issues.

> We can verify the CPU usage of the test cgroup using cpuacct.stat. The
> output shows:
> 
>   system: 53
>   user: 2
> 
> The CPU usage of the cgroup is relatively low at around 55%, but this usage
> doesn't increase, even with more netperf tasks. The reason is that CPU0 is
> at 100% utilization, as confirmed by mpstat:
> 
>   02:56:22 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
>   02:56:23 PM    0    0.99    0.00   55.45    0.00    0.99   42.57    0.00    0.00    0.00    0.00
> 
>   02:56:23 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
>   02:56:24 PM    0    2.00    0.00   55.00    0.00    0.00   43.00    0.00    0.00    0.00    0.00
> 
> It is clear that the %soft is not accounted into the cgroup of the
> interrupted task. This behavior is unexpected. We should account for IRQ
> time to the cgroup to reflect the pressure the group is under.
> 
> After a thorough analysis, I discovered that this change in behavior is due
> to commit 305e6835e055 ("sched: Do not account irq time to current task"),
> which altered whether IRQ time should be charged to the interrupted task.
> While I agree that a task should not be penalized by random interrupts, the
> task itself cannot progress while interrupted. Therefore, the interrupted
> time should be reported to the user.
> 
> The system metric in cpuacct.stat is crucial in indicating whether a
> container is under heavy system pressure, including IRQ/softirq activity.
> Hence, IRQ/softirq time should be accounted for in the cpuacct system
> usage, which also applies to cgroup2’s rstat.
> 
> This patch reintroduces IRQ/softirq accounting to cgroups.

How !? what does it actually do?

> Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
> Cc: Johannes Weiner <hannes@cmpxchg.org>
> ---
>  kernel/sched/core.c  | 33 +++++++++++++++++++++++++++++++--
>  kernel/sched/psi.c   | 14 +++-----------
>  kernel/sched/stats.h |  7 ++++---
>  3 files changed, 38 insertions(+), 16 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 06a06f0897c3..5ed2c5c8c911 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -5579,6 +5579,35 @@ __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
>  static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
>  #endif /* CONFIG_SCHED_DEBUG */
>  
> +#ifdef CONFIG_IRQ_TIME_ACCOUNTING
> +static void account_irqtime(struct rq *rq, struct task_struct *curr,
> +			    struct task_struct *prev)
> +{
> +	int cpu = smp_processor_id();
> +	s64 delta;
> +	u64 irq;
> +
> +	if (!static_branch_likely(&sched_clock_irqtime))
> +		return;
> +
> +	irq = irq_time_read(cpu);
> +	delta = (s64)(irq - rq->psi_irq_time);

At this point the variable is no longer exclusive to PSI and should
probably be renamed.

> +	if (delta < 0)
> +		return;
> +
> +	rq->psi_irq_time = irq;
> +	psi_account_irqtime(rq, curr, prev, delta);
> +	cgroup_account_cputime(curr, delta);
> +	/* We account both softirq and irq into softirq */
> +	cgroup_account_cputime_field(curr, CPUTIME_SOFTIRQ, delta);

This seems wrong.. we have CPUTIME_IRQ.

> +}

In fact, much of this seems like it's going about things sideways.

Why can't you just add the cgroup_account_*() garbage to
irqtime_account_irq()? That is were it's still split out into softirq
and irq.

But the much bigger question is -- how can you be sure that this
interrupt is in fact for the cgroup you're attributing it to? Could be
for an entirely different cgroup.

On Fri, Nov 1, 2024 at 6:28 PM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Fri, Nov 01, 2024 at 11:17:50AM +0800, Yafang Shao wrote:
> > After enabling CONFIG_IRQ_TIME_ACCOUNTING to monitor IRQ pressure in our
> > container environment, we observed several noticeable behavioral changes.
> >
> > One of our IRQ-heavy services, such as Redis, reported a significant
> > reduction in CPU usage after upgrading to the new kernel with
> > CONFIG_IRQ_TIME_ACCOUNTING enabled. However, despite adding more threads
> > to handle an increased workload, the CPU usage could not be raised. In
> > other words, even though the container’s CPU usage appeared low, it was
> > unable to process more workloads to utilize additional CPU resources, which
> > caused issues.
>
> > We can verify the CPU usage of the test cgroup using cpuacct.stat. The
> > output shows:
> >
> >   system: 53
> >   user: 2
> >
> > The CPU usage of the cgroup is relatively low at around 55%, but this usage
> > doesn't increase, even with more netperf tasks. The reason is that CPU0 is
> > at 100% utilization, as confirmed by mpstat:
> >
> >   02:56:22 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
> >   02:56:23 PM    0    0.99    0.00   55.45    0.00    0.99   42.57    0.00    0.00    0.00    0.00
> >
> >   02:56:23 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
> >   02:56:24 PM    0    2.00    0.00   55.00    0.00    0.00   43.00    0.00    0.00    0.00    0.00
> >
> > It is clear that the %soft is not accounted into the cgroup of the
> > interrupted task. This behavior is unexpected. We should account for IRQ
> > time to the cgroup to reflect the pressure the group is under.
> >
> > After a thorough analysis, I discovered that this change in behavior is due
> > to commit 305e6835e055 ("sched: Do not account irq time to current task"),
> > which altered whether IRQ time should be charged to the interrupted task.
> > While I agree that a task should not be penalized by random interrupts, the
> > task itself cannot progress while interrupted. Therefore, the interrupted
> > time should be reported to the user.
> >
> > The system metric in cpuacct.stat is crucial in indicating whether a
> > container is under heavy system pressure, including IRQ/softirq activity.
> > Hence, IRQ/softirq time should be accounted for in the cpuacct system
> > usage, which also applies to cgroup2’s rstat.
> >
> > This patch reintroduces IRQ/softirq accounting to cgroups.
>
> How !? what does it actually do?

It seems there's some misunderstanding due to the term *accounting*
here. What it actually does is track the interrupted time within a
cgroup.

>
> > Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
> > Cc: Johannes Weiner <hannes@cmpxchg.org>
> > ---
> >  kernel/sched/core.c  | 33 +++++++++++++++++++++++++++++++--
> >  kernel/sched/psi.c   | 14 +++-----------
> >  kernel/sched/stats.h |  7 ++++---
> >  3 files changed, 38 insertions(+), 16 deletions(-)
> >
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 06a06f0897c3..5ed2c5c8c911 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -5579,6 +5579,35 @@ __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
> >  static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
> >  #endif /* CONFIG_SCHED_DEBUG */
> >
> > +#ifdef CONFIG_IRQ_TIME_ACCOUNTING
> > +static void account_irqtime(struct rq *rq, struct task_struct *curr,
> > +                         struct task_struct *prev)
> > +{
> > +     int cpu = smp_processor_id();
> > +     s64 delta;
> > +     u64 irq;
> > +
> > +     if (!static_branch_likely(&sched_clock_irqtime))
> > +             return;
> > +
> > +     irq = irq_time_read(cpu);
> > +     delta = (s64)(irq - rq->psi_irq_time);
>
> At this point the variable is no longer exclusive to PSI and should
> probably be renamed.

OK.

>
> > +     if (delta < 0)
> > +             return;
> > +
> > +     rq->psi_irq_time = irq;
> > +     psi_account_irqtime(rq, curr, prev, delta);
> > +     cgroup_account_cputime(curr, delta);
> > +     /* We account both softirq and irq into softirq */
> > +     cgroup_account_cputime_field(curr, CPUTIME_SOFTIRQ, delta);
>
> This seems wrong.. we have CPUTIME_IRQ.

OK.

>
> > +}
>
> In fact, much of this seems like it's going about things sideways.
>
> Why can't you just add the cgroup_account_*() garbage to
> irqtime_account_irq()? That is were it's still split out into softirq
> and irq.

I previously implemented this in v1: link. However, in that version,
we had to hold the irq_lock within the critical path, which could
impact performance. Taking inspiration from commit ddae0ca2a8fe
("sched: Move psi_account_irqtime() out of update_rq_clock_task()
hotpath"), I've now adapted the approach to handle it in a
non-critical path, reducing the performance impact.

>
> But the much bigger question is -- how can you be sure that this
> interrupt is in fact for the cgroup you're attributing it to? Could be
> for an entirely different cgroup.

As I explained in another thread, identifying the exact culprit can be
challenging, but identifying the victim is straightforward. That’s
precisely what this patch set accomplishes.

-- 
Regards
Yafang