hrtimer_cancel() busy-waits for the hrtimer callback to stop, pretty much like del_timer_sync(). This creates a possible deadlock scenario where we hold a spinlock before calling hrtimer_cancel() while in trying to acquire the same spinlock in the callback.
This kind of deadlock is already known and is catchable by lockdep, like for del_timer_sync(), we can add lockdep annotations. However, it is still missing for hrtimer_cancel(). (I have a WIP patch to make it complete for hrtimer_cancel() but it breaks booting.)
And there is such a deadlock scenario in kernel/events/core.c too, well actually, it is a simpler version: the hrtimer callback waits for itself to finish on the same CPU! It sounds stupid but it is not obvious at all, it hides very deeply in the perf event code:
cpu_clock_event_init(): perf_swevent_init_hrtimer(): hwc->hrtimer.function = perf_swevent_hrtimer;
perf_swevent_hrtimer(): __perf_event_overflow(): __perf_event_account_interrupt(): perf_adjust_period(): pmu->stop(): cpu_clock_event_stop(): perf_swevent_cancel(): hrtimer_cancel()
Getting stuck in an hrtimer is a disaster:
Interrupts are disabled for hrtimer_interrupt(), this means other interrupt handling is blocked too, notably the IPI handler, especially when smp_call_function_*() waits for their callbacks synchronously. This is probably why we saw hundreds of soft lockup's inside smp_call_function_single(), given how widely they are used in kernel. Ironically, perf event code uses synchronous smp_call_function_single() heavily too.
The fix is not easy. To minimize the impact, ideally we should just avoid busy waiting when it is called within the hrtimer callback on the same CPU, there is no reason to wait for itself to finish anyway. Probably it doesn't even need to cancel itself either since it will restart by pmu->start() later. There are two possible fixes here:
1. Modify hrtimer API to detect if a hrtimer callback is running on the same CPU now. It does not look pretty to adjust it only for this special case though.
2. Passing some information from perf_swevent_hrtimer() down to perf_swevent_cancel().
So I pick the latter approach, it is nicer and straightforward.
Note, Jiri pointed out that the swevent pmu does not triger NMI and this timer will never be touched in NMI context, so we don't race with NMI handler here.
Fixes: abd50713944c ("perf: Reimplement frequency driven sampling") Cc: Peter Zijlstra peterz@infradead.org Cc: Ingo Molnar mingo@redhat.com Cc: Linus Torvalds torvalds@linux-foundation.org Cc: Arnaldo Carvalho de Melo acme@kernel.org Cc: Alexander Shishkin alexander.shishkin@linux.intel.com Cc: Jiri Olsa jolsa@redhat.com Cc: Namhyung Kim namhyung@kernel.org Cc: stable@vger.kernel.org Signed-off-by: Cong Wang xiyou.wangcong@gmail.com --- include/linux/perf_event.h | 3 +++ kernel/events/core.c | 43 +++++++++++++++++++++++++++---------------- 2 files changed, 30 insertions(+), 16 deletions(-)
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h index 1fa12887ec02..aab39b8aa720 100644 --- a/include/linux/perf_event.h +++ b/include/linux/perf_event.h @@ -310,6 +310,9 @@ struct pmu { #define PERF_EF_START 0x01 /* start the counter when adding */ #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */ #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */ +#define PERF_EF_NO_WAIT 0x08 /* do not wait when stopping, for + * example, waiting for a timer + */
/* * Adds/Removes a counter to/from the PMU, can be done inside a diff --git a/kernel/events/core.c b/kernel/events/core.c index 8f0434a9951a..f15832346b35 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -3555,7 +3555,8 @@ do { \ static DEFINE_PER_CPU(int, perf_throttled_count); static DEFINE_PER_CPU(u64, perf_throttled_seq);
-static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) +static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, + bool disable, bool nowait) { struct hw_perf_event *hwc = &event->hw; s64 period, sample_period; @@ -3574,8 +3575,13 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bo hwc->sample_period = sample_period;
if (local64_read(&hwc->period_left) > 8*sample_period) { - if (disable) - event->pmu->stop(event, PERF_EF_UPDATE); + if (disable) { + int flags = PERF_EF_UPDATE; + + if (nowait) + flags |= PERF_EF_NO_WAIT; + event->pmu->stop(event, flags); + }
local64_set(&hwc->period_left, 0);
@@ -3645,7 +3651,7 @@ static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, * twice. */ if (delta > 0) - perf_adjust_period(event, period, delta, false); + perf_adjust_period(event, period, delta, false, false);
event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); next: @@ -7681,7 +7687,8 @@ static void perf_log_itrace_start(struct perf_event *event) }
static int -__perf_event_account_interrupt(struct perf_event *event, int throttle) +__perf_event_account_interrupt(struct perf_event *event, int throttle, + bool nowait) { struct hw_perf_event *hwc = &event->hw; int ret = 0; @@ -7710,7 +7717,8 @@ __perf_event_account_interrupt(struct perf_event *event, int throttle) hwc->freq_time_stamp = now;
if (delta > 0 && delta < 2*TICK_NSEC) - perf_adjust_period(event, delta, hwc->last_period, true); + perf_adjust_period(event, delta, hwc->last_period, true, + nowait); }
return ret; @@ -7718,7 +7726,7 @@ __perf_event_account_interrupt(struct perf_event *event, int throttle)
int perf_event_account_interrupt(struct perf_event *event) { - return __perf_event_account_interrupt(event, 1); + return __perf_event_account_interrupt(event, 1, false); }
/* @@ -7727,7 +7735,7 @@ int perf_event_account_interrupt(struct perf_event *event)
static int __perf_event_overflow(struct perf_event *event, int throttle, struct perf_sample_data *data, - struct pt_regs *regs) + struct pt_regs *regs, bool nowait) { int events = atomic_read(&event->event_limit); int ret = 0; @@ -7739,7 +7747,7 @@ static int __perf_event_overflow(struct perf_event *event, if (unlikely(!is_sampling_event(event))) return 0;
- ret = __perf_event_account_interrupt(event, throttle); + ret = __perf_event_account_interrupt(event, throttle, nowait);
/* * XXX event_limit might not quite work as expected on inherited @@ -7768,7 +7776,7 @@ int perf_event_overflow(struct perf_event *event, struct perf_sample_data *data, struct pt_regs *regs) { - return __perf_event_overflow(event, 1, data, regs); + return __perf_event_overflow(event, 1, data, regs, true); }
/* @@ -7831,7 +7839,7 @@ static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
for (; overflow; overflow--) { if (__perf_event_overflow(event, throttle, - data, regs)) { + data, regs, false)) { /* * We inhibit the overflow from happening when * hwc->interrupts == MAX_INTERRUPTS. @@ -9110,7 +9118,7 @@ static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
if (regs && !perf_exclude_event(event, regs)) { if (!(event->attr.exclude_idle && is_idle_task(current))) - if (__perf_event_overflow(event, 1, &data, regs)) + if (__perf_event_overflow(event, 1, &data, regs, true)) ret = HRTIMER_NORESTART; }
@@ -9141,7 +9149,7 @@ static void perf_swevent_start_hrtimer(struct perf_event *event) HRTIMER_MODE_REL_PINNED); }
-static void perf_swevent_cancel_hrtimer(struct perf_event *event) +static void perf_swevent_cancel_hrtimer(struct perf_event *event, bool sync) { struct hw_perf_event *hwc = &event->hw;
@@ -9149,7 +9157,10 @@ static void perf_swevent_cancel_hrtimer(struct perf_event *event) ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); local64_set(&hwc->period_left, ktime_to_ns(remaining));
- hrtimer_cancel(&hwc->hrtimer); + if (sync) + hrtimer_cancel(&hwc->hrtimer); + else + hrtimer_try_to_cancel(&hwc->hrtimer); } }
@@ -9200,7 +9211,7 @@ static void cpu_clock_event_start(struct perf_event *event, int flags)
static void cpu_clock_event_stop(struct perf_event *event, int flags) { - perf_swevent_cancel_hrtimer(event); + perf_swevent_cancel_hrtimer(event, flags & PERF_EF_NO_WAIT); cpu_clock_event_update(event); }
@@ -9277,7 +9288,7 @@ static void task_clock_event_start(struct perf_event *event, int flags)
static void task_clock_event_stop(struct perf_event *event, int flags) { - perf_swevent_cancel_hrtimer(event); + perf_swevent_cancel_hrtimer(event, flags & PERF_EF_NO_WAIT); task_clock_event_update(event, event->ctx->time); }