[PATCH -tip/master] perf, x86: P4 PMU - Fix unflagged overflows handlingv4

[Cyrill Gorcunov]

Don found that P4 PMU reads CCCR register instead of counter
itself (in attempt to catch unflagged event) this makes P4
NMI handler to consume all NMIs it observes. So the other
NMI users such as kgdb simply have no chance to get NMI
on their hands.

Side note: at moment there is no way to run nmi-watchdog
together with perf tool. This is because both 'perf top' and
nmi-watchdog use same event. So while nmi-watchdog reserves
one event/counter for own needs there is no room for perf tool
left (there is a way to disable nmi-watchdog on boot of course).

Ming has tested this patch with the following results

 | 1. watchdog disabled
 |
 | kgdb tests on boot OK
 | perf works OK
 |
 | 2. watchdog enabled, without patch perf-x86-p4-nmi-4
 |
 | kgdb tests on boot hang
 |
 | 3. watchdog enabled, without patch perf-x86-p4-nmi-4 and do not run kgdb
 | tests on boot
 |
 | "perf top" partialy works
 |   cpu-cycles            no
 |   instructions          yes
 |   cache-references      no
 |   cache-misses          no
 |   branch-instructions   no
 |   branch-misses         yes
 |   bus-cycles            no
 |
 | 4. watchdog enabled, with patch perf-x86-p4-nmi-4 applied
 |
 | kgdb tests on boot OK
 | perf does not work, NMI "Dazed and confused" messages show up
 |

Which means we still have problems with p4 box due to 'unknown'
nmi happens but at least it should fix kgdb test cases.

Reported-by: Jason Wessel <jason.wessel@xxxxxxxxxxxxx>
Reported-by: Don Zickus <dzickus@xxxxxxxxxx>
Signed-off-by: Cyrill Gorcunov <gorcunov@xxxxxxxxxx>
CC: Lin Ming <ming.m.lin@xxxxxxxxx>
CC: Stephane Eranian <eranian@xxxxxxxxxx>
CC: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
---

Don, your opinion?

 arch/x86/include/asm/perf_event_p4.h |    3 +++
 arch/x86/kernel/cpu/perf_event_p4.c  |   28 +++++++++++++++-------------
 2 files changed, 18 insertions(+), 13 deletions(-)

Index: linux-2.6.git/arch/x86/include/asm/perf_event_p4.h
=====================================================================
--- linux-2.6.git.orig/arch/x86/include/asm/perf_event_p4.h
+++ linux-2.6.git/arch/x86/include/asm/perf_event_p4.h
@@ -20,6 +20,9 @@
 #define ARCH_P4_MAX_ESCR	(ARCH_P4_TOTAL_ESCR - ARCH_P4_RESERVED_ESCR)
 #define ARCH_P4_MAX_CCCR	(18)

+#define ARCH_P4_CNTRVAL_BITS	(40)
+#define ARCH_P4_CNTRVAL_MASK	((1ULL << ARCH_P4_CNTRVAL_BITS) - 1)
+
 #define P4_ESCR_EVENT_MASK	0x7e000000U
 #define P4_ESCR_EVENT_SHIFT	25
 #define P4_ESCR_EVENTMASK_MASK	0x01fffe00U
Index: linux-2.6.git/arch/x86/kernel/cpu/perf_event_p4.c
=====================================================================
--- linux-2.6.git.orig/arch/x86/kernel/cpu/perf_event_p4.c
+++ linux-2.6.git/arch/x86/kernel/cpu/perf_event_p4.c
@@ -753,19 +753,21 @@ out:

 static inline int p4_pmu_clear_cccr_ovf(struct hw_perf_event *hwc)
 {
-	int overflow = 0;
-	u32 low, high;
+	u64 v;

-	rdmsr(hwc->config_base + hwc->idx, low, high);
-
-	/* we need to check high bit for unflagged overflows */
-	if ((low & P4_CCCR_OVF) || !(high & (1 << 31))) {
-		overflow = 1;
-		(void)checking_wrmsrl(hwc->config_base + hwc->idx,
-			((u64)low) & ~P4_CCCR_OVF);
+	/* an official way for overflow indication */
+	rdmsrl(hwc->config_base + hwc->idx, v);
+	if (v & P4_CCCR_OVF) {
+		wrmsrl(hwc->config_base + hwc->idx, v & ~P4_CCCR_OVF);
+		return 1;
 	}

-	return overflow;
+	/* it might be unflagged overflow */
+	rdmsrl(hwc->event_base + hwc->idx, v);
+	if (!(v & ARCH_P4_CNTRVAL_MASK))
+		return 1;
+
+	return 0;
 }

 static void p4_pmu_disable_pebs(void)
@@ -1152,9 +1154,9 @@ static __initconst const struct x86_pmu
 	 */
 	.num_counters		= ARCH_P4_MAX_CCCR,
 	.apic			= 1,
-	.cntval_bits		= 40,
-	.cntval_mask		= (1ULL << 40) - 1,
-	.max_period		= (1ULL << 39) - 1,
+	.cntval_bits		= ARCH_P4_CNTRVAL_BITS,
+	.cntval_mask		= ARCH_P4_CNTRVAL_MASK,
+	.max_period		= (1ULL << (ARCH_P4_CNTRVAL_BITS - 1)) - 1,
 	.hw_config		= p4_hw_config,
 	.schedule_events	= p4_pmu_schedule_events,
 	/*
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/