Re: [PATCH v2 0/5] perf, bts: Fallout from the fuzzer for perf/urgent

[Alexander Shishkin <alexander.shishkin@linux.intel.com>]

Ingo Molnar <mingo@kernel.org> writes:

> * Alexander Shishkin <alexander.shishkin@linux.intel.com> wrote:
>
>> Hi,
>> 
>> There were more bugs since the previous version, plus the BTS barriers got 
>> fixed. With these patches, my testcase keeps running and no spurious NMI 
>> warnings pop up any more.
>
> Could you please also run the fuzzer that Vince uses, does it now pass on hardware 
> you have access to?

Sure. And yes, I did catch a warning, which calls for one more patch
(below). Also one unrelated thing in PEBS that Peter fixed.

> I'd like to make "passes the fuzzer" a standard requirement before new changes are 
> accepted to perf core.

Let's make it so.

For the sake of consistency, this one needs to go before 3/5. I'll
re-send the whole series, though, if need be. I've got 2 perf_fuzzers
running on this meanwhile.

>From c170c607bfdc3804578033faa43f342a5d95eb6c Mon Sep 17 00:00:00 2001
From: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Date: Wed, 7 Sep 2016 18:05:08 +0300
Subject: [PATCH] perf/x86/intel: Don't disable "intel_bts" around "intel"
 event batching

At the moment, intel_bts events get disabled from intel PMU's disable
callback, which includes event scheduling transactions of said PMU,
which have nothing to do with intel_bts events.

We do want to keep intel_bts events off inside the PMI handler to
avoid filling up their buffer too soon.

This patch moves intel_bts enabling/disabling directly to the PMI
handler.

Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
---
 arch/x86/events/intel/core.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index 88792f846d..e2d71513c9 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -1730,9 +1730,11 @@ static __initconst const u64 knl_hw_cache_extra_regs
  * disabled state if called consecutively.
  *
  * During consecutive calls, the same disable value will be written to related
- * registers, so the PMU state remains unchanged. hw.state in
- * intel_bts_disable_local will remain PERF_HES_STOPPED too in consecutive
- * calls.
+ * registers, so the PMU state remains unchanged.
+ *
+ * intel_bts events don't coexist with intel pmu's BTS events because of
+ * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them
+ * disabled around intel pmu's event batching etc, only inside the PMI handler.
  */
 static void __intel_pmu_disable_all(void)
 {
@@ -1742,8 +1744,6 @@ static void __intel_pmu_disable_all(void)
 
 	if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
 		intel_pmu_disable_bts();
-	else
-		intel_bts_disable_local();
 
 	intel_pmu_pebs_disable_all();
 }
@@ -1771,8 +1771,7 @@ static void __intel_pmu_enable_all(int added, bool pmi)
 			return;
 
 		intel_pmu_enable_bts(event->hw.config);
-	} else
-		intel_bts_enable_local();
+	}
 }
 
 static void intel_pmu_enable_all(int added)
@@ -2076,6 +2075,7 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
 	 */
 	if (!x86_pmu.late_ack)
 		apic_write(APIC_LVTPC, APIC_DM_NMI);
+	intel_bts_disable_local();
 	__intel_pmu_disable_all();
 	handled = intel_pmu_drain_bts_buffer();
 	handled += intel_bts_interrupt();
@@ -2175,6 +2175,7 @@ done:
 	/* Only restore PMU state when it's active. See x86_pmu_disable(). */
 	if (cpuc->enabled)
 		__intel_pmu_enable_all(0, true);
+	intel_bts_enable_local();
 
 	/*
 	 * Only unmask the NMI after the overflow counters
-- 
2.9.3