perf/x86/intel: Fix race in intel_pmu_disable_event()

New race in x86_pmu_stop() was introduced by replacing the
atomic __test_and_clear_bit() of cpuc->active_mask by separate
test_bit() and __clear_bit() calls in the following commit:

  3966c3feca ("x86/perf/amd: Remove need to check "running" bit in NMI handler")

The race causes panic for PEBS events with enabled callchains:

  BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
  ...
  RIP: 0010:perf_prepare_sample+0x8c/0x530
  Call Trace:
   <NMI>
   perf_event_output_forward+0x2a/0x80
   __perf_event_overflow+0x51/0xe0
   handle_pmi_common+0x19e/0x240
   intel_pmu_handle_irq+0xad/0x170
   perf_event_nmi_handler+0x2e/0x50
   nmi_handle+0x69/0x110
   default_do_nmi+0x3e/0x100
   do_nmi+0x11a/0x180
   end_repeat_nmi+0x16/0x1a
  RIP: 0010:native_write_msr+0x6/0x20
  ...
   </NMI>
   intel_pmu_disable_event+0x98/0xf0
   x86_pmu_stop+0x6e/0xb0
   x86_pmu_del+0x46/0x140
   event_sched_out.isra.97+0x7e/0x160
  ...

The event is configured to make samples from PEBS drain code,
but when it's disabled, we'll go through NMI path instead,
where data->callchain will not get allocated and we'll crash:

          x86_pmu_stop
            test_bit(hwc->idx, cpuc->active_mask)
            intel_pmu_disable_event(event)
            {
              ...
              intel_pmu_pebs_disable(event);
              ...

EVENT OVERFLOW ->  <NMI>
                     intel_pmu_handle_irq
                       handle_pmi_common
   TEST PASSES ->        test_bit(bit, cpuc->active_mask))
                           perf_event_overflow
                             perf_prepare_sample
                             {
                               ...
                               if (!(sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
                                     data->callchain = perf_callchain(event, regs);

         CRASH ->              size += data->callchain->nr;
                             }
                   </NMI>
              ...
              x86_pmu_disable_event(event)
            }

            __clear_bit(hwc->idx, cpuc->active_mask);

Fixing this by disabling the event itself before setting
off the PEBS bit.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Arcari <darcari@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Lendacky Thomas <Thomas.Lendacky@amd.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Fixes: 3966c3feca ("x86/perf/amd: Remove need to check "running" bit in NMI handler")
Link: http://lkml.kernel.org/r/20190504151556.31031-1-jolsa@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/events/intel/core.c

@@ -2091,15 +2091,19 @@ static void intel_pmu_disable_event(struct perf_event *event)
 	cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
 	cpuc->intel_cp_status &= ~(1ull << hwc->idx);
 
-	if (unlikely(event->attr.precise_ip))
-		intel_pmu_pebs_disable(event);
-
 	if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
 		intel_pmu_disable_fixed(hwc);
 		return;
 	}
 
 	x86_pmu_disable_event(event);
+
+	/*
+	 * Needs to be called after x86_pmu_disable_event,
+	 * so we don't trigger the event without PEBS bit set.
+	 */
+	if (unlikely(event->attr.precise_ip))
+		intel_pmu_pebs_disable(event);
 }
 
 static void intel_pmu_del_event(struct perf_event *event)