Skip to contentSkip to conversation

perf/x86/intel: Init early callchain for bts event

PATCH In review Jiri Olsa 4 participants 4 messages Nov 11, 2018 – Nov 12, 2018

Subsystems (2)
x86 perf events
lore
Share & tools
lore

Jiri Olsa

author reviewer
#0 lore raw

PATCH 1/1 perf/x86/intel: Init early callchain for bts event

Vince reported crash in bts flush code when touching the callchain data, which was supposed to be initialized as an 'early' callchain data.

  BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
  ...
  Call Trace:
   <IRQ>
   intel_pmu_drain_bts_buffer+0x151/0x220
   ? intel_get_event_constraints+0x219/0x360
   ? perf_assign_events+0xe2/0x2a0
   ? select_idle_sibling+0x22/0x3a0
   ? __update_load_avg_se+0x1ec/0x270
   ? enqueue_task_fair+0x377/0xdd0
   ? cpumask_next_and+0x19/0x20
   ? load_balance+0x134/0x950
   ? check_preempt_curr+0x7a/0x90
   ? ttwu_do_wakeup+0x19/0x140
   x86_pmu_stop+0x3b/0x90
   x86_pmu_del+0x57/0x160
   event_sched_out.isra.106+0x81/0x170
   group_sched_out.part.108+0x51/0xc0
   __perf_event_disable+0x7f/0x160
   event_function+0x8c/0xd0
   remote_function+0x3c/0x50
   flush_smp_call_function_queue+0x35/0xe0
   smp_call_function_single_interrupt+0x3a/0xd0
   call_function_single_interrupt+0xf/0x20
   </IRQ>

It was triggered by fuzzer by can be easilt reproduced by:

  # perf record -e cpu/branch-instructions/p -g -c 1

The problem is that bts drain code does not initialize sample's early callchain data and calls perf_prepare_sample with NULL sample->callchain, even if it's expected to exist via __PERF_SAMPLE_CALLCHAIN_EARLY sample type bit.

The fix initializes the callchain data for this case with empty callchain, because bts does not report callchains.

Cc Vince Weaver <[email protected]>
Fixes 6cbc304f2f36 ("perf/x86/intel: Fix unwind errors from PEBS entries (mk-II)")
Reported-by Vince Weaver <[email protected]>
Signed-off-by Jiri Olsa <[email protected]>
---
 arch/x86/events/intel/ds.c | 5 +++++
 1 file changed, 5 insertions(+)
arch/x86/events/intel/ds.c +5 -0
index b7b01d762d32..1049b547fdfe 100644
@@ -577,6 +577,8 @@ void intel_pmu_disable_bts(void)
 	update_debugctlmsr(debugctlmsr);
 }
 
+static struct perf_callchain_entry __empty_callchain = { .nr = 0, };
+
 int intel_pmu_drain_bts_buffer(void)
 {
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);

@@ -612,6 +614,9 @@ int intel_pmu_drain_bts_buffer(void)
 
 	perf_sample_data_init(&data, 0, event->hw.last_period);
 
+	if (event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY)
+		data.callchain = &__empty_callchain;
+
 	/*
 	 * BTS leaks kernel addresses in branches across the cpl boundary,
 	 * such as traps or system calls, so unless the user is asking for
Signature · 1 line
2.17.2
maintainer
#1 lore raw
Quoted 7 lines · Jiri Olsa

Vince reported crash in bts flush code when touching the callchain data, which was supposed to be initialized as an 'early' callchain data.

  BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
  ...
Quoted 7 lines

It was triggered by fuzzer by can be easilt reproduced by:

  # perf record -e cpu/branch-instructions/p -g -c 1

The problem is that bts drain code does not initialize sample's early callchain data and calls perf_prepare_sample with NULL sample->callchain, even if it's expected to exist via __PERF_SAMPLE_CALLCHAIN_EARLY sample type bit.

Not sure that is the actual problem, nor that this:

arch/x86/events/intel/ds.c int intel_pmu_drain_bts_buffer(void) · line 622

Quoted 10 lines
arch/x86/events/intel/ds.c +3 -0
@@ -612,6 +614,9 @@ int intel_pmu_drain_bts_buffer(void)
 
 	perf_sample_data_init(&data, 0, event->hw.last_period);
 
+	if (event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY)
+		data.callchain = &__empty_callchain;
+
 	/*
 	 * BTS leaks kernel addresses in branches across the cpl boundary,
 	 * such as traps or system calls, so unless the user is asking for

is the right fix.

If you look at commit:

6cbc304f2f36 ("perf/x86/intel: Fix unwind errors from PEBS entries (mk-II)")

Then the right fix would be to do perf_callchain() from the BTS drain code -- if '/p'.

Because prior to that commit, we would do a perf_callchain() in intel_pmu_drain_bts_buffer()'s call to perf_prepare_sample(), which would do an actual stack unwind for a branch entry.

With your patch, we get an empty stack for every entry.

Which is a change in behaviour...

Now arguably, this is really stupid behaviour. Who in his right mind wants callchain output on BTS entries. And even if they do, BTS + precise_ip is nonsensical.

So in my mind disallowing precise_ip on BTS would be the simplest fix.

Hmm?

#2 lore raw
Quoted 4 lines

So in my mind disallowing precise_ip on BTS would be the simplest fix.

Hmm?

Both seem reasonable to me.

-Andi

#3 lore raw
In reply to Peter Zijlstra
Quoted 38 lines · Peter Zijlstra
Quoted 7 lines · Jiri Olsa

Vince reported crash in bts flush code when touching the callchain data, which was supposed to be initialized as an 'early' callchain data.

  BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
  ...
Quoted 7 lines

It was triggered by fuzzer by can be easilt reproduced by:

  # perf record -e cpu/branch-instructions/p -g -c 1

The problem is that bts drain code does not initialize sample's early callchain data and calls perf_prepare_sample with NULL sample->callchain, even if it's expected to exist via __PERF_SAMPLE_CALLCHAIN_EARLY sample type bit.

Not sure that is the actual problem, nor that this:

arch/x86/events/intel/ds.c int intel_pmu_drain_bts_buffer(void) · line 622

Quoted 10 lines
arch/x86/events/intel/ds.c +3 -0
@@ -612,6 +614,9 @@ int intel_pmu_drain_bts_buffer(void)
 
 	perf_sample_data_init(&data, 0, event->hw.last_period);
 
+	if (event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY)
+		data.callchain = &__empty_callchain;
+
 	/*
 	 * BTS leaks kernel addresses in branches across the cpl boundary,
 	 * such as traps or system calls, so unless the user is asking for

is the right fix.

If you look at commit:

6cbc304f2f36 ("perf/x86/intel: Fix unwind errors from PEBS entries (mk-II)")

Then the right fix would be to do perf_callchain() from the BTS drain code -- if '/p'.

Quoted 7 lines

Because prior to that commit, we would do a perf_callchain() in intel_pmu_drain_bts_buffer()'s call to perf_prepare_sample(), which would do an actual stack unwind for a branch entry.

With your patch, we get an empty stack for every entry.

Which is a change in behaviour...

I thought there's no callchain anyway, because we use zero-ed regs

Quoted 8 lines

Now arguably, this is really stupid behaviour. Who in his right mind wants callchain output on BTS entries. And even if they do, BTS + precise_ip is nonsensical.

So in my mind disallowing precise_ip on BTS would be the simplest fix.

Hmm?

sounds ok, will post it

thanks,
jirka