Skip to contentSkip to conversation

perf/x86: Restore event pointer setup in x86_pmu_start()

PATCH Superseded Breno Leitao 3 participants 4 messages Mar 9 – Mar 10

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf/x86: Restore event pointer setup in x86_pmu_start()

A production AMD EPYC system crashed with a NULL pointer dereference in the PMU NMI handler:

  BUG: kernel NULL pointer dereference, address: 0000000000000198
  RIP: x86_perf_event_update+0xc/0xa0
  Call Trace:
   <NMI>
   amd_pmu_v2_handle_irq+0x1a6/0x390
   perf_event_nmi_handler+0x24/0x40

The faulting instruction is cmpq $0x0, 0x198(%rdi) with RDI=0, corresponding to the if (unlikely(!hwc->event_base)) check in x86_perf_event_update() where hwc = &event->hw and event is NULL.

drgn inspection of the vmcore on CPU 106 showed a mismatch between cpuc->active_mask and cpuc->events[]:

active_mask: 0x1e (bits 1, 2, 3, 4)

  events[1]:   0xff1100136cbd4f38  (valid)
  events[2]:   0x0                 (NULL, but active_mask bit 2 set)
  events[3]:   0xff1100076fd2cf38  (valid)
  events[4]:   0xff1100079e990a90  (valid)

The event that should occupy events[2] was found in event_list[2] with hw.idx=2 and hw.state=0x0, confirming x86_pmu_start() had run (which clears hw.state and sets active_mask) but events[2] was never populated.

Another event (event_list[0]) had hw.state=0x7 (STOPPED|UPTODATE|ARCH), showing it was stopped when the PMU rescheduled events, confirming the throttle-then-reschedule sequence occurred.

The root cause is commit 7e772a93eb61 ("perf/x86: Fix NULL event access and potential PEBS record loss") which moved the cpuc->events[idx] assignment out of x86_pmu_start() and into x86_pmu_enable(). This broke any path that calls pmu->start() without going through x86_pmu_enable() -- specifically the unthrottle path:

perf_adjust_freq_unthr_events()

    -> perf_event_unthrottle_group()
      -> perf_event_unthrottle()
        -> event->pmu->start(event, 0)
          -> x86_pmu_start()     // sets active_mask but not events[]

The race sequence is:

1. A group of perf events overflows, triggering group throttle via

     perf_event_throttle_group(). All events are stopped: active_mask
     bits cleared, events[] preserved (x86_pmu_stop no longer clears
     events[] after commit 7e772a93eb61).

2. While still throttled (PERF_HES_STOPPED), x86_pmu_enable() runs due to other scheduling activity. Stopped events that need to move counters get PERF_HES_ARCH set and events[old_idx] cleared. In step 2 of x86_pmu_enable(), PERF_HES_ARCH causes these events to be skipped -- events[new_idx] is never set.

3. The timer tick unthrottles the group via pmu->start(). Since commit 7e772a93eb61 removed the events[] assignment from x86_pmu_start(), active_mask[new_idx] is set but events[new_idx] remains NULL.

4. A PMC overflow NMI fires. The handler iterates active counters, finds active_mask[2] set, reads events[2] which is NULL, and crashes dereferencing it.

Restore cpuc->events[idx] = event in x86_pmu_start() so that every caller of pmu->start() correctly populates events[] before setting active_mask. This does not reintroduce the PEBS issue that commit 7e772a93eb61 fixed, because that fix also moved the events[] = NULL clearing from x86_pmu_stop() to x86_pmu_del() -- throttle/unthrottle cycles no longer clear events[].

Fixes 7e772a93eb61 ("perf/x86: Fix NULL event access and potential PEBS record loss")
---
 arch/x86/events/core.c | 5 +++++
 1 file changed, 5 insertions(+)
arch/x86/events/core.c +5 -0
index 03ce1bc7ef2ea..fd82d1427b335 100644
@@ -1546,6 +1546,11 @@ static void x86_pmu_start(struct perf_event *event, int flags)
 
 	event->hw.state = 0;
 
+	/*
+	 * Ensure events[idx] is set before active_mask, so NMI handlers
+	 * never see an active counter with a NULL event pointer.
+	 */
+	cpuc->events[idx] = event;
 	__set_bit(idx, cpuc->active_mask);
 	static_call(x86_pmu_enable)(event);
 	perf_event_update_userpage(event);

---
base-commit: 0bcac7b11262557c990da1ac564d45777eb6b005
change-id: 20260309-perf-fd32da0317a8

Best regards,

Signature · 1 line
Breno Leitao <[email protected]>
maintainer
#1 lore raw
Quoted 51 lines · Breno Leitao

A production AMD EPYC system crashed with a NULL pointer dereference in the PMU NMI handler:

  BUG: kernel NULL pointer dereference, address: 0000000000000198
  RIP: x86_perf_event_update+0xc/0xa0
  Call Trace:
   <NMI>
   amd_pmu_v2_handle_irq+0x1a6/0x390
   perf_event_nmi_handler+0x24/0x40

The faulting instruction is cmpq $0x0, 0x198(%rdi) with RDI=0, corresponding to the if (unlikely(!hwc->event_base)) check in x86_perf_event_update() where hwc = &event->hw and event is NULL.

drgn inspection of the vmcore on CPU 106 showed a mismatch between cpuc->active_mask and cpuc->events[]:

active_mask: 0x1e (bits 1, 2, 3, 4)

  events[1]:   0xff1100136cbd4f38  (valid)
  events[2]:   0x0                 (NULL, but active_mask bit 2 set)
  events[3]:   0xff1100076fd2cf38  (valid)
  events[4]:   0xff1100079e990a90  (valid)

The event that should occupy events[2] was found in event_list[2] with hw.idx=2 and hw.state=0x0, confirming x86_pmu_start() had run (which clears hw.state and sets active_mask) but events[2] was never populated.

Another event (event_list[0]) had hw.state=0x7 (STOPPED|UPTODATE|ARCH), showing it was stopped when the PMU rescheduled events, confirming the throttle-then-reschedule sequence occurred.

The root cause is commit 7e772a93eb61 ("perf/x86: Fix NULL event access and potential PEBS record loss") which moved the cpuc->events[idx] assignment out of x86_pmu_start() and into x86_pmu_enable(). This broke any path that calls pmu->start() without going through x86_pmu_enable() -- specifically the unthrottle path:

perf_adjust_freq_unthr_events()

    -> perf_event_unthrottle_group()
      -> perf_event_unthrottle()
        -> event->pmu->start(event, 0)
          -> x86_pmu_start()     // sets active_mask but not events[]

The race sequence is:

1. A group of perf events overflows, triggering group throttle via

     perf_event_throttle_group(). All events are stopped: active_mask
     bits cleared, events[] preserved (x86_pmu_stop no longer clears
     events[] after commit 7e772a93eb61).
Quoted 5 lines

2. While still throttled (PERF_HES_STOPPED), x86_pmu_enable() runs due to other scheduling activity. Stopped events that need to move counters get PERF_HES_ARCH set and events[old_idx] cleared. In step 2 of x86_pmu_enable(), PERF_HES_ARCH causes these events to be skipped -- events[new_idx] is never set.

So why not just move this then? Having less sites that set that value is more better, no?

---

arch/x86/events/core.c +2 -1
index 03ce1bc7ef2e..54b4c315d927 100644
@@ -1372,6 +1372,8 @@ static void x86_pmu_enable(struct pmu *pmu)
 			else if (i < n_running)
 				continue;
 
+			cpuc->events[hwc->idx] = event;
+
 			if (hwc->state & PERF_HES_ARCH)
 				continue;
 
@@ -1379,7 +1381,6 @@ static void x86_pmu_enable(struct pmu *pmu)
 			 * if cpuc->enabled = 0, then no wrmsr as
 			 * per x86_pmu_enable_event()
 			 */
-			cpuc->events[hwc->idx] = event;
 			x86_pmu_start(event, PERF_EF_RELOAD);
 		}
 		cpuc->n_added = 0;
#2 lore raw
Quoted 5 lines · Peter Zijlstra

On Mon, Mar 09, 2026 at 07:40:56AM -0700, Breno Leitao wrote:

So why not just move this then? Having less sites that set that value is more better, no?

Sure, let me update.

Thanks for the review,
--breno

#3 lore raw
In reply to Breno Leitao
Quoted 57 lines · Breno Leitao

A production AMD EPYC system crashed with a NULL pointer dereference in the PMU NMI handler:

  BUG: kernel NULL pointer dereference, address: 0000000000000198
  RIP: x86_perf_event_update+0xc/0xa0
  Call Trace:
   <NMI>
   amd_pmu_v2_handle_irq+0x1a6/0x390
   perf_event_nmi_handler+0x24/0x40

The faulting instruction is cmpq $0x0, 0x198(%rdi) with RDI=0, corresponding to the if (unlikely(!hwc->event_base)) check in x86_perf_event_update() where hwc = &event->hw and event is NULL.

drgn inspection of the vmcore on CPU 106 showed a mismatch between cpuc->active_mask and cpuc->events[]:

active_mask: 0x1e (bits 1, 2, 3, 4)

  events[1]:   0xff1100136cbd4f38  (valid)
  events[2]:   0x0                 (NULL, but active_mask bit 2 set)
  events[3]:   0xff1100076fd2cf38  (valid)
  events[4]:   0xff1100079e990a90  (valid)

The event that should occupy events[2] was found in event_list[2] with hw.idx=2 and hw.state=0x0, confirming x86_pmu_start() had run (which clears hw.state and sets active_mask) but events[2] was never populated.

Another event (event_list[0]) had hw.state=0x7 (STOPPED|UPTODATE|ARCH), showing it was stopped when the PMU rescheduled events, confirming the throttle-then-reschedule sequence occurred.

The root cause is commit 7e772a93eb61 ("perf/x86: Fix NULL event access and potential PEBS record loss") which moved the cpuc->events[idx] assignment out of x86_pmu_start() and into x86_pmu_enable(). This broke any path that calls pmu->start() without going through x86_pmu_enable() -- specifically the unthrottle path:

perf_adjust_freq_unthr_events()

    -> perf_event_unthrottle_group()
      -> perf_event_unthrottle()
        -> event->pmu->start(event, 0)
          -> x86_pmu_start()     // sets active_mask but not events[]

The race sequence is:

1. A group of perf events overflows, triggering group throttle via

     perf_event_throttle_group(). All events are stopped: active_mask
     bits cleared, events[] preserved (x86_pmu_stop no longer clears
     events[] after commit 7e772a93eb61).

2. While still throttled (PERF_HES_STOPPED), x86_pmu_enable() runs due to other scheduling activity. Stopped events that need to move counters get PERF_HES_ARCH set and events[old_idx] cleared. In step 2 of x86_pmu_enable(), PERF_HES_ARCH causes these events to be skipped -- events[new_idx] is never set.

Quoted 8 lines

3. The timer tick unthrottles the group via pmu->start(). Since commit 7e772a93eb61 removed the events[] assignment from x86_pmu_start(), active_mask[new_idx] is set but events[new_idx] remains NULL.

4. A PMC overflow NMI fires. The handler iterates active counters, finds active_mask[2] set, reads events[2] which is NULL, and crashes dereferencing it.

Thanks for fixing this issue. Better add an "Cc: [email protected]" tag as well.

Quoted 40 lines

Restore cpuc->events[idx] = event in x86_pmu_start() so that every caller of pmu->start() correctly populates events[] before setting active_mask. This does not reintroduce the PEBS issue that commit 7e772a93eb61 fixed, because that fix also moved the events[] = NULL clearing from x86_pmu_stop() to x86_pmu_del() -- throttle/unthrottle cycles no longer clear events[].

Fixes 7e772a93eb61 ("perf/x86: Fix NULL event access and potential PEBS record loss")
---
 arch/x86/events/core.c | 5 +++++
 1 file changed, 5 insertions(+)
arch/x86/events/core.c +5 -1
index 03ce1bc7ef2ea..fd82d1427b335 100644
@@ -1546,6 +1546,11 @@ static void x86_pmu_start(struct perf_event *event, int flags)
 
 	event->hw.state = 0;
 
+	/*
+	 * Ensure events[idx] is set before active_mask, so NMI handlers
+	 * never see an active counter with a NULL event pointer.
+	 */
+	cpuc->events[idx] = event;
 	__set_bit(idx, cpuc->active_mask);
 	static_call(x86_pmu_enable)(event);
 	perf_event_update_userpage(event);
 
---

base-commit: 0bcac7b11262557c990da1ac564d45777eb6b005
change-id: 20260309-perf-fd32da0317a8

Best regards,

Signature · 1 line
Breno Leitao <[email protected]>