Hi Brandon,

On Mon, Aug 26, 2024 at 12:49:34AM +0000, bugzilla-daemon@kernel.org wrote:
> https://bugzilla.kernel.org/show_bug.cgi?id=219194
> 
>             Bug ID: 219194
>            Summary: PSI full-stall duration reports nonsensically large
>                     value when reading pressure file too frequently
>            Product: Other
>            Version: 2.5
>           Hardware: AMD
>                 OS: Linux
>             Status: NEW
>           Severity: normal
>           Priority: P3
>          Component: Other
>           Assignee: other_other@kernel-bugs.osdl.org
>           Reporter: brandon@buildbuddy.io
>                 CC: hannes@cmpxchg.org
>         Regression: No
> 
> Created attachment 306775 [details]
>   --> https://bugzilla.kernel.org/attachment.cgi?id=306775&action=edit
> Minimal reproducible example
> 
> The cumulative full-stall duration ("total" field) reported in
> /sys/fs/cgroup/.../cpu.pressure is occasionally nonsensically large - a value
> much longer than the total execution duration of the only process that was
> run
> in the cgroup.
> 
> Interestingly, the nonsensical value that it reports is often exactly
> 4.294967
> seconds. This value looks suspiciously like UINT32_MAX (4294967295) which
> made
> me think that this is a bug with PSI code rather than something I was doing
> wrong.

Thanks for reporting this. Your reproducer was immensely helpful in
tracking this down, much appreciated. With the below patch, I can no
longer trigger it. If at all possible, could you try it on your end?

It's a race condition between aggregator (reader) and a task changing
state. Specifically, it's an issue with the live state snooping: the
aggregator will add cpu_clock() - state_start to the sample of any
currently active states that have not been recorded yet. This makes
the aggregator get ahead of recorded task state, but we assume this is
okay because, as the state concludes, that delta we snooped ahead will
be incorporated back into the recorded times.

But there is a bug. When the state change concludes it reads
cpu_clock() before acquiring the seq write lock. This means an
aggregator can race and snoop ahead to a later clock value than the
one at which the state will conclude. The next aggregation will see a
negative delta and record a UINT_MAX - <race time> sample:

psi_group_change()               get_recent_times()
  now = cpu_clock() (X)

  <a few cycles pass>

                                 read_seqcount_begin()
                                 now = cpu_clock() (X+42)
                                 memcpy(times, groupc->times)
                                 !read_seqcount_retry()

                                 /* Snoops live state up to X+42 */
                                 if (state_mask & (1 << s))
                                   times[s] += now - state_start

				 delta = times[s] - groupc->times_prev[s]
                                 groupc->times_prev[s] = times[s]

                                 /* times_prev up to X+42 */

  write_seqcount_begin()
  /* concludes state ending at X */
  groupc->times[s] += now - state_start
  groupc->state_mask &= 1 << s
  write_seqcount_end()

                                 read_seqcount_begin()
                                 memcpy(times, groupc->times)
                                 !read_seqcount_retry()

                                 if (state_mask & (1 << s))
                                   ; /* nope, state concluded */

				 delta = times[s] - groupc->times_prev[s]
                                 /* times_prev ahead, sample is UINT_MAX-42 */

The solution is to move the clock read of the state change into the
locked section. This way, live snooping cannot get ahead of the time
at which the state will conclude:

---

diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
index 020d58967d4e..84dad1511d1e 100644
--- a/kernel/sched/psi.c
+++ b/kernel/sched/psi.c
@@ -769,12 +769,13 @@ static void record_times(struct psi_group_cpu *groupc, u64 now)
 }
 
 static void psi_group_change(struct psi_group *group, int cpu,
-			     unsigned int clear, unsigned int set, u64 now,
+			     unsigned int clear, unsigned int set,
 			     bool wake_clock)
 {
 	struct psi_group_cpu *groupc;
 	unsigned int t, m;
 	u32 state_mask;
+	u64 now;
 
 	lockdep_assert_rq_held(cpu_rq(cpu));
 	groupc = per_cpu_ptr(group->pcpu, cpu);
@@ -789,6 +790,7 @@ static void psi_group_change(struct psi_group *group, int cpu,
 	 * SOME and FULL time these may have resulted in.
 	 */
 	write_seqcount_begin(&groupc->seq);
+	now = cpu_clock(cpu);
 
 	/*
 	 * Start with TSK_ONCPU, which doesn't have a corresponding
@@ -899,18 +901,15 @@ void psi_task_change(struct task_struct *task, int clear, int set)
 {
 	int cpu = task_cpu(task);
 	struct psi_group *group;
-	u64 now;
 
 	if (!task->pid)
 		return;
 
 	psi_flags_change(task, clear, set);
 
-	now = cpu_clock(cpu);
-
 	group = task_psi_group(task);
 	do {
-		psi_group_change(group, cpu, clear, set, now, true);
+		psi_group_change(group, cpu, clear, set, true);
 	} while ((group = group->parent));
 }
 
@@ -919,7 +918,6 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 {
 	struct psi_group *group, *common = NULL;
 	int cpu = task_cpu(prev);
-	u64 now = cpu_clock(cpu);
 
 	if (next->pid) {
 		psi_flags_change(next, 0, TSK_ONCPU);
@@ -936,7 +934,7 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 				break;
 			}
 
-			psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
+			psi_group_change(group, cpu, 0, TSK_ONCPU, true);
 		} while ((group = group->parent));
 	}
 
@@ -974,7 +972,7 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 		do {
 			if (group == common)
 				break;
-			psi_group_change(group, cpu, clear, set, now, wake_clock);
+			psi_group_change(group, cpu, clear, set, wake_clock);
 		} while ((group = group->parent));
 
 		/*
@@ -986,7 +984,7 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 		if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
 			clear &= ~TSK_ONCPU;
 			for (; group; group = group->parent)
-				psi_group_change(group, cpu, clear, set, now, wake_clock);
+				psi_group_change(group, cpu, clear, set, wake_clock);
 		}
 	}
 }
@@ -997,8 +995,8 @@ void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_st
 	int cpu = task_cpu(curr);
 	struct psi_group *group;
 	struct psi_group_cpu *groupc;
-	u64 now, irq;
 	s64 delta;
+	u64 irq;
 
 	if (static_branch_likely(&psi_disabled))
 		return;
@@ -1011,7 +1009,6 @@ void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_st
 	if (prev && task_psi_group(prev) == group)
 		return;
 
-	now = cpu_clock(cpu);
 	irq = irq_time_read(cpu);
 	delta = (s64)(irq - rq->psi_irq_time);
 	if (delta < 0)
@@ -1019,12 +1016,15 @@ void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_st
 	rq->psi_irq_time = irq;
 
 	do {
+		u64 now;
+
 		if (!group->enabled)
 			continue;
 
 		groupc = per_cpu_ptr(group->pcpu, cpu);
 
 		write_seqcount_begin(&groupc->seq);
+		now = cpu_clock(cpu);
 
 		record_times(groupc, now);
 		groupc->times[PSI_IRQ_FULL] += delta;
@@ -1223,11 +1223,9 @@ void psi_cgroup_restart(struct psi_group *group)
 	for_each_possible_cpu(cpu) {
 		struct rq *rq = cpu_rq(cpu);
 		struct rq_flags rf;
-		u64 now;
 
 		rq_lock_irq(rq, &rf);
-		now = cpu_clock(cpu);
-		psi_group_change(group, cpu, 0, 0, now, true);
+		psi_group_change(group, cpu, 0, 0, true);
 		rq_unlock_irq(rq, &rf);
 	}
 }

Hi Johannes - thank you for looking into this so quickly!

For testing kernel changes it's easiest for me to use kernel v5.15 and run it under Firecracker MicroVM, so I had to adjust the patch slightly to remove the irq PSI changes.

I checked that the issue still reproduces on v5.15 before applying the patch, and that after applying the patch, the issue no longer reproduces.

Thank you!