Skip to contentSkip to conversation

perf/x86: fix spurious NMI with PEBS Load Latency event

PATCH In review [email protected] 3 participants 3 messages Apr 4, 2017 – Apr 14, 2017

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 perf/x86: fix spurious NMI with PEBS Load Latency event

From: Kan Liang <[email protected]>

Spurious NMIs will be observed when applying the following command.
while true ; do sudo perf record -b -a -e "cpu/umask=0x01,event=0xcd,ldlat=0x80/pp,cpu/umask=0x03,event=0x0/,
cpu/umask=0x02,event=0x0/,cycles,branches,cache-misses, cache-references" -- sleep 10 ; done

The issue was introduced by
commit 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+")

The previous patch clear the status's bits for the counters used for PEBS events, by masking the whole 64 bits pebs_enabled. However, only the lower 32 bits of both status and pebs_enabled are reserved for PEBS-able counters.
For status, the first three bits of upper 32 bits are fixed counter overflow bit.
For pebs_enabled, the first three bits of upper 32 bits are for PEBS Load Latency event.
In the test case, the PEBS Load Latency event and fixed counter event could be overflowed at the same time. The fixed counter overflow bit will
be cleared by mistake. Once it is cleared, the fixed counter overflow never be processed, which finally trigger spurious NMI.

Correct the PEBS enabled mask by ignoring the non-PEBS bits.

Fixes 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL

status on HSW+")

---

Change since V1:
- using a macros for (1ULL << MAX_PEBS_EVENTS) - 1)

 arch/x86/events/intel/core.c | 2 +-
 arch/x86/events/intel/ds.c   | 2 +-
 arch/x86/events/perf_event.h | 1 +
 3 files changed, 3 insertions(+), 2 deletions(-)
arch/x86/events/intel/core.c +1 -1
index 319da60..3411e79 100644
@@ -2151,7 +2151,7 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
 	 * counters from the GLOBAL_STATUS mask and we always process PEBS
 	 * events via drain_pebs().
 	 */
-	status &= ~cpuc->pebs_enabled;
+	status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
 
 	/*
 	 * PEBS overflow sets bit 62 in the global status register
arch/x86/events/intel/ds.c +1 -1
index 9dfeeec..c6d23ff 100644
@@ -1222,7 +1222,7 @@ get_next_pebs_record_by_bit(void *base, void *top, int bit)
 
 			/* clear non-PEBS bit and re-check */
 			pebs_status = p->status & cpuc->pebs_enabled;
-			pebs_status &= (1ULL << MAX_PEBS_EVENTS) - 1;
+			pebs_status &= PEBS_COUNTER_MASK;
 			if (pebs_status == (1 << bit))
 				return at;
 		}
arch/x86/events/perf_event.h +1 -0
index 416f565..2469fba 100644
@@ -79,6 +79,7 @@ struct amd_nb {
 
 /* The maximal number of PEBS events: */
 #define MAX_PEBS_EVENTS		8
+#define PEBS_COUNTER_MASK	((1ULL << MAX_PEBS_EVENTS) - 1)
 
 /*
  * Flags PEBS can handle without an PMI.
Signature · 1 line
2.4.3
maintainer
#1 lore raw
Quoted 26 lines · On Tue, Apr 04, 2017 at 03:14:06PM -0400, [email protected] wrote:

From: Kan Liang <[email protected]>

Spurious NMIs will be observed when applying the following command.
while true ; do sudo perf record -b -a -e "cpu/umask=0x01,event=0xcd,ldlat=0x80/pp,cpu/umask=0x03,event=0x0/,
cpu/umask=0x02,event=0x0/,cycles,branches,cache-misses, cache-references" -- sleep 10 ; done

The issue was introduced by
commit 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+")

The previous patch clear the status's bits for the counters used for PEBS events, by masking the whole 64 bits pebs_enabled. However, only the lower 32 bits of both status and pebs_enabled are reserved for PEBS-able counters.
For status, the first three bits of upper 32 bits are fixed counter overflow bit.
For pebs_enabled, the first three bits of upper 32 bits are for PEBS Load Latency event.
In the test case, the PEBS Load Latency event and fixed counter event could be overflowed at the same time. The fixed counter overflow bit will
be cleared by mistake. Once it is cleared, the fixed counter overflow never be processed, which finally trigger spurious NMI.

Quoted 5 lines

Correct the PEBS enabled mask by ignoring the non-PEBS bits.

Fixes 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL

status on HSW+")

That is atrocious unreadable garbage; in a large part due to the random line breaks.

Any half sane editor can reflow text; use that.

---
Subject: perf/x86: Fix spurious NMI with PEBS Load Latency event
From: Kan Liang <[email protected]>
Date: Tue, 4 Apr 2017 15:14:06 -0400

Spurious NMIs will be observed with the following command:

while :; do

    perf record -bae "cpu/umask=0x01,event=0xcd,ldlat=0x80/pp"
                  -e "cpu/umask=0x03,event=0x0/"
                  -e "cpu/umask=0x02,event=0x0/"
                  -e cycles,branches,cache-misses
                  -e cache-references -- sleep 10

done

The issue was introduced by commit:

8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+")

That commit clears the status bits for the counters used for PEBS events, by masking the whole 64 bits pebs_enabled. However, only the low 32 bits of both status and pebs_enabled are reserved for PEBS-able counters.

For status bits 32-34 are fixed counter overflow bits. For pebe_enabled bits 32-34 are for PEBS Load Latency.

In the test case, the PEBS Load Latency event and fixed counter event could overflow at the same time. The fixed counter overflow bit will be cleared by mistake. Once it is cleared, the fixed counter overflow never be processed, which finally trigger spurious NMI.

Correct the PEBS enabled mask by ignoring the non-PEBS bits.

Fixes 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+")
---
 arch/x86/events/intel/core.c |    2 +-
 arch/x86/events/intel/ds.c   |    2 +-
 arch/x86/events/perf_event.h |    1 +
 3 files changed, 3 insertions(+), 2 deletions(-)
arch/x86/events/intel/core.c +1 -1
@@ -2151,7 +2151,7 @@ static int intel_pmu_handle_irq(struct p
 	 * counters from the GLOBAL_STATUS mask and we always process PEBS
 	 * events via drain_pebs().
 	 */
-	status &= ~cpuc->pebs_enabled;
+	status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
 
 	/*
 	 * PEBS overflow sets bit 62 in the global status register
arch/x86/events/intel/ds.c +1 -1
@@ -1222,7 +1222,7 @@ get_next_pebs_record_by_bit(void *base,
 
 			/* clear non-PEBS bit and re-check */
 			pebs_status = p->status & cpuc->pebs_enabled;
-			pebs_status &= (1ULL << MAX_PEBS_EVENTS) - 1;
+			pebs_status &= PEBS_COUNTER_MASK;
 			if (pebs_status == (1 << bit))
 				return at;
 		}
arch/x86/events/perf_event.h +1 -0
@@ -79,6 +79,7 @@ struct amd_nb {
 
 /* The maximal number of PEBS events: */
 #define MAX_PEBS_EVENTS		8
+#define PEBS_COUNTER_MASK	((1ULL << MAX_PEBS_EVENTS) - 1)
 
 /*
  * Flags PEBS can handle without an PMI.
#2 lore raw

[tip:perf/core] perf/x86: Fix spurious NMI with PEBS Load Latency event

In reply to [email protected]
perf/x86: Fix spurious NMI with PEBS Load Latency event
Commit-ID:  fd583ad1563bec5f00140e1f2444adbcd331caad
Gitweb:     http://git.kernel.org/tip/fd583ad1563bec5f00140e1f2444adbcd331caad
Author:     Kan Liang <[email protected]>
AuthorDate: Tue, 4 Apr 2017 15:14:06 -0400
Committer:  Ingo Molnar <[email protected]>
CommitDate: Fri, 14 Apr 2017 10:31:39 +0200

perf/x86: Fix spurious NMI with PEBS Load Latency event

Spurious NMIs will be observed with the following command:

while :; do

    perf record -bae "cpu/umask=0x01,event=0xcd,ldlat=0x80/pp"
                  -e "cpu/umask=0x03,event=0x0/"
                  -e "cpu/umask=0x02,event=0x0/"
                  -e cycles,branches,cache-misses
                  -e cache-references -- sleep 10

done

The bug was introduced by commit:

8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+")

That commit clears the status bits for the counters used for PEBS events, by masking the whole 64 bits pebs_enabled. However, only the low 32 bits of both status and pebs_enabled are reserved for PEBS-able counters.

For status bits 32-34 are fixed counter overflow bits. For pebs_enabled bits 32-34 are for PEBS Load Latency.

In the test case, the PEBS Load Latency event and fixed counter event could overflow at the same time. The fixed counter overflow bit will be cleared by mistake. Once it is cleared, the fixed counter overflow never be processed, which finally trigger spurious NMI.

Correct the PEBS enabled mask by ignoring the non-PEBS bits.

Cc Alexander Shishkin <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Stephane Eranian <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Vince Weaver <[email protected]>
Fixes 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+")
---
 arch/x86/events/intel/core.c | 2 +-
 arch/x86/events/intel/ds.c   | 2 +-
 arch/x86/events/perf_event.h | 1 +
 3 files changed, 3 insertions(+), 2 deletions(-)
arch/x86/events/intel/core.c +1 -1
index 4244bed..a6d91d4 100644
@@ -2151,7 +2151,7 @@ again:
 	 * counters from the GLOBAL_STATUS mask and we always process PEBS
 	 * events via drain_pebs().
 	 */
-	status &= ~cpuc->pebs_enabled;
+	status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
 
 	/*
 	 * PEBS overflow sets bit 62 in the global status register
arch/x86/events/intel/ds.c +1 -1
index 9dfeeec..c6d23ff 100644
@@ -1222,7 +1222,7 @@ get_next_pebs_record_by_bit(void *base, void *top, int bit)
 
 			/* clear non-PEBS bit and re-check */
 			pebs_status = p->status & cpuc->pebs_enabled;
-			pebs_status &= (1ULL << MAX_PEBS_EVENTS) - 1;
+			pebs_status &= PEBS_COUNTER_MASK;
 			if (pebs_status == (1 << bit))
 				return at;
 		}
arch/x86/events/perf_event.h +1 -0
index bcbb1d2..be3d362 100644
@@ -79,6 +79,7 @@ struct amd_nb {
 
 /* The maximal number of PEBS events: */
 #define MAX_PEBS_EVENTS		8
+#define PEBS_COUNTER_MASK	((1ULL << MAX_PEBS_EVENTS) - 1)
 
 /*
  * Flags PEBS can handle without an PMI.