[tip:perf/core] perf/x86/pebs: Add workaround for broken OVFL status on HSW+

[tip-bot for Stephane Eranian]

Commit-ID:  8077eca079a212f26419c57226f28696b7100683
Gitweb:     http://git.kernel.org/tip/8077eca079a212f26419c57226f28696b7100683
Author:     Stephane Eranian <eranian@xxxxxxxxxx>
AuthorDate: Thu, 3 Mar 2016 20:50:41 +0100
Committer:  Ingo Molnar <mingo@xxxxxxxxxx>
CommitDate: Tue, 8 Mar 2016 12:18:35 +0100

perf/x86/pebs: Add workaround for broken OVFL status on HSW+

This patch fixes an issue with the GLOBAL_OVERFLOW_STATUS bits on
Haswell, Broadwell and Skylake processors when using PEBS.

The SDM stipulates that when the PEBS iterrupt threshold is crossed,
an interrupt is posted and the kernel is interrupted. The kernel will
find GLOBAL_OVF_SATUS bit 62 set indicating there are PEBS records to
drain. But the bits corresponding to the actual counters should NOT be
set. The kernel follows the SDM and assumes that all PEBS events are
processed in the drain_pebs() callback. The kernel then checks for
remaining overflows on any other (non-PEBS) events and processes these
in the for_each_bit_set(&status) loop.

As it turns out, under certain conditions on HSW and later processors,
on PEBS buffer interrupt, bit 62 is set but the counter bits may be
set as well. In that case, the kernel drains PEBS and generates
SAMPLES with the EXACT tag, then it processes the counter bits, and
generates normal (non-EXACT) SAMPLES.

I ran into this problem by trying to understand why on HSW sampling on
a PEBS event was sometimes returning SAMPLES without the EXACT tag.
This should not happen on user level code because HSW has the
eventing_ip which always point to the instruction that caused the
event.

The workaround in this patch simply ensures that the bits for the
counters used for PEBS events are cleared after the PEBS buffer has
been drained. With this fix 100% of the PEBS samples on my user code
report the EXACT tag.

Before:
  $ perf record -e cpu/event=0xd0,umask=0x81/upp ./multichase
  $ perf report -D | fgrep SAMPLES
  PERF_RECORD_SAMPLE(IP, 0x2): 11775/11775: 0x406de5 period: 73469 addr: 0 exact=Y
                           \--- EXACT tag is missing

After:
  $ perf record -e cpu/event=0xd0,umask=0x81/upp ./multichase
  $ perf report -D | fgrep SAMPLES
  PERF_RECORD_SAMPLE(IP, 0x4002): 11775/11775: 0x406de5 period: 73469 addr: 0 exact=Y
                           \--- EXACT tag is set

The problem tends to appear more often when multiple PEBS events are used.

Signed-off-by: Stephane Eranian <eranian@xxxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx>
Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Vince Weaver <vincent.weaver@xxxxxxxxx>
Cc: adrian.hunter@xxxxxxxxx
Cc: kan.liang@xxxxxxxxx
Cc: namhyung@xxxxxxxxxx
Link: http://lkml.kernel.org/r/1457034642-21837-3-git-send-email-eranian@xxxxxxxxxx
Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
---
 arch/x86/events/intel/core.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index b3f6349..6567c62 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -1892,6 +1892,16 @@ again:
 	if (__test_and_clear_bit(62, (unsigned long *)&status)) {
 		handled++;
 		x86_pmu.drain_pebs(regs);
+		/*
+		 * There are cases where, even though, the PEBS ovfl bit is set
+		 * in GLOBAL_OVF_STATUS, the PEBS events may also have their
+		 * overflow bits set for their counters. We must clear them
+		 * here because they have been processed as exact samples in
+		 * the drain_pebs() routine. They must not be processed again
+		 * in the for_each_bit_set() loop for regular samples below.
+		 */
+		status &= ~cpuc->pebs_enabled;
+		status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
 	}
 
 	/*