1 files changed, 537 insertions, 40 deletions

diff --git a/arch/sparc/kernel/perf_event.c b/arch/sparc/kernel/perf_event.c
index 2d6a1b1..04db927 100644
--- a/arch/sparc/kernel/perf_event.c
+++ b/arch/sparc/kernel/perf_event.c
@@ -56,7 +56,8 @@ struct cpu_hw_events {
 	struct perf_event	*events[MAX_HWEVENTS];
 	unsigned long		used_mask[BITS_TO_LONGS(MAX_HWEVENTS)];
 	unsigned long		active_mask[BITS_TO_LONGS(MAX_HWEVENTS)];
-	int enabled;
+	u64			pcr;
+	int			enabled;
 };
 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
 
@@ -68,8 +69,30 @@ struct perf_event_map {
 #define PIC_LOWER	0x02
 };
 
+static unsigned long perf_event_encode(const struct perf_event_map *pmap)
+{
+	return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
+}
+
+static void perf_event_decode(unsigned long val, u16 *enc, u8 *msk)
+{
+	*msk = val & 0xff;
+	*enc = val >> 16;
+}
+
+#define C(x) PERF_COUNT_HW_CACHE_##x
+
+#define CACHE_OP_UNSUPPORTED	0xfffe
+#define CACHE_OP_NONSENSE	0xffff
+
+typedef struct perf_event_map cache_map_t
+				[PERF_COUNT_HW_CACHE_MAX]
+				[PERF_COUNT_HW_CACHE_OP_MAX]
+				[PERF_COUNT_HW_CACHE_RESULT_MAX];
+
 struct sparc_pmu {
 	const struct perf_event_map	*(*event_map)(int);
+	const cache_map_t		*cache_map;
 	int				max_events;
 	int				upper_shift;
 	int				lower_shift;
@@ -80,21 +103,109 @@ struct sparc_pmu {
 	int				lower_nop;
 };
 
-static const struct perf_event_map ultra3i_perfmon_event_map[] = {
+static const struct perf_event_map ultra3_perfmon_event_map[] = {
 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
 };
 
-static const struct perf_event_map *ultra3i_event_map(int event_id)
+static const struct perf_event_map *ultra3_event_map(int event_id)
 {
-	return &ultra3i_perfmon_event_map[event_id];
+	return &ultra3_perfmon_event_map[event_id];
 }
 
-static const struct sparc_pmu ultra3i_pmu = {
-	.event_map	= ultra3i_event_map,
-	.max_events	= ARRAY_SIZE(ultra3i_perfmon_event_map),
+static const cache_map_t ultra3_cache_map = {
+[C(L1D)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
+		[C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
+		[C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+};
+
+static const struct sparc_pmu ultra3_pmu = {
+	.event_map	= ultra3_event_map,
+	.cache_map	= &ultra3_cache_map,
+	.max_events	= ARRAY_SIZE(ultra3_perfmon_event_map),
 	.upper_shift	= 11,
 	.lower_shift	= 4,
 	.event_mask	= 0x3f,
@@ -102,6 +213,121 @@ static const struct sparc_pmu ultra3i_pmu = {
 	.lower_nop	= 0x14,
 };
 
+/* Niagara1 is very limited.  The upper PIC is hard-locked to count
+ * only instructions, so it is free running which creates all kinds of
+ * problems.  Some hardware designs make one wonder if the creator
+ * even looked at how this stuff gets used by software.
+ */
+static const struct perf_event_map niagara1_perfmon_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
+	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
+	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
+	[PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
+};
+
+static const struct perf_event_map *niagara1_event_map(int event_id)
+{
+	return &niagara1_perfmon_event_map[event_id];
+}
+
+static const cache_map_t niagara1_cache_map = {
+[C(L1D)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
+		[C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+};
+
+static const struct sparc_pmu niagara1_pmu = {
+	.event_map	= niagara1_event_map,
+	.cache_map	= &niagara1_cache_map,
+	.max_events	= ARRAY_SIZE(niagara1_perfmon_event_map),
+	.upper_shift	= 0,
+	.lower_shift	= 4,
+	.event_mask	= 0x7,
+	.upper_nop	= 0x0,
+	.lower_nop	= 0x0,
+};
+
 static const struct perf_event_map niagara2_perfmon_event_map[] = {
 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
@@ -116,8 +342,96 @@ static const struct perf_event_map *niagara2_event_map(int event_id)
 	return &niagara2_perfmon_event_map[event_id];
 }
 
+static const cache_map_t niagara2_cache_map = {
+[C(L1D)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
+		[C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
+		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
+		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
+	},
+},
+};
+
 static const struct sparc_pmu niagara2_pmu = {
 	.event_map	= niagara2_event_map,
+	.cache_map	= &niagara2_cache_map,
 	.max_events	= ARRAY_SIZE(niagara2_perfmon_event_map),
 	.upper_shift	= 19,
 	.lower_shift	= 6,
@@ -151,23 +465,30 @@ static u64 nop_for_index(int idx)
 			      sparc_pmu->lower_nop, idx);
 }
 
-static inline void sparc_pmu_enable_event(struct hw_perf_event *hwc,
-					    int idx)
+static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
 {
 	u64 val, mask = mask_for_index(idx);
 
-	val = pcr_ops->read();
-	pcr_ops->write((val & ~mask) | hwc->config);
+	val = cpuc->pcr;
+	val &= ~mask;
+	val |= hwc->config;
+	cpuc->pcr = val;
+
+	pcr_ops->write(cpuc->pcr);
 }
 
-static inline void sparc_pmu_disable_event(struct hw_perf_event *hwc,
-					     int idx)
+static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
 {
 	u64 mask = mask_for_index(idx);
 	u64 nop = nop_for_index(idx);
-	u64 val = pcr_ops->read();
+	u64 val;
 
-	pcr_ops->write((val & ~mask) | nop);
+	val = cpuc->pcr;
+	val &= ~mask;
+	val |= nop;
+	cpuc->pcr = val;
+
+	pcr_ops->write(cpuc->pcr);
 }
 
 void hw_perf_enable(void)
@@ -182,7 +503,7 @@ void hw_perf_enable(void)
 	cpuc->enabled = 1;
 	barrier();
 
-	val = pcr_ops->read();
+	val = cpuc->pcr;
 
 	for (i = 0; i < MAX_HWEVENTS; i++) {
 		struct perf_event *cp = cpuc->events[i];
@@ -194,7 +515,9 @@ void hw_perf_enable(void)
 		val |= hwc->config_base;
 	}
 
-	pcr_ops->write(val);
+	cpuc->pcr = val;
+
+	pcr_ops->write(cpuc->pcr);
 }
 
 void hw_perf_disable(void)
@@ -207,10 +530,12 @@ void hw_perf_disable(void)
 
 	cpuc->enabled = 0;
 
-	val = pcr_ops->read();
+	val = cpuc->pcr;
 	val &= ~(PCR_UTRACE | PCR_STRACE |
 		 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
-	pcr_ops->write(val);
+	cpuc->pcr = val;
+
+	pcr_ops->write(cpuc->pcr);
 }
 
 static u32 read_pmc(int idx)
@@ -242,7 +567,7 @@ static void write_pmc(int idx, u64 val)
 }
 
 static int sparc_perf_event_set_period(struct perf_event *event,
-					 struct hw_perf_event *hwc, int idx)
+				       struct hw_perf_event *hwc, int idx)
 {
 	s64 left = atomic64_read(&hwc->period_left);
 	s64 period = hwc->sample_period;
@@ -282,19 +607,19 @@ static int sparc_pmu_enable(struct perf_event *event)
 	if (test_and_set_bit(idx, cpuc->used_mask))
 		return -EAGAIN;
 
-	sparc_pmu_disable_event(hwc, idx);
+	sparc_pmu_disable_event(cpuc, hwc, idx);
 
 	cpuc->events[idx] = event;
 	set_bit(idx, cpuc->active_mask);
 
 	sparc_perf_event_set_period(event, hwc, idx);
-	sparc_pmu_enable_event(hwc, idx);
+	sparc_pmu_enable_event(cpuc, hwc, idx);
 	perf_event_update_userpage(event);
 	return 0;
 }
 
 static u64 sparc_perf_event_update(struct perf_event *event,
-				     struct hw_perf_event *hwc, int idx)
+				   struct hw_perf_event *hwc, int idx)
 {
 	int shift = 64 - 32;
 	u64 prev_raw_count, new_raw_count;
@@ -324,7 +649,7 @@ static void sparc_pmu_disable(struct perf_event *event)
 	int idx = hwc->idx;
 
 	clear_bit(idx, cpuc->active_mask);
-	sparc_pmu_disable_event(hwc, idx);
+	sparc_pmu_disable_event(cpuc, hwc, idx);
 
 	barrier();
 
@@ -338,18 +663,29 @@ static void sparc_pmu_disable(struct perf_event *event)
 static void sparc_pmu_read(struct perf_event *event)
 {
 	struct hw_perf_event *hwc = &event->hw;
+
 	sparc_perf_event_update(event, hwc, hwc->idx);
 }
 
 static void sparc_pmu_unthrottle(struct perf_event *event)
 {
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 	struct hw_perf_event *hwc = &event->hw;
-	sparc_pmu_enable_event(hwc, hwc->idx);
+
+	sparc_pmu_enable_event(cpuc, hwc, hwc->idx);
 }
 
 static atomic_t active_events = ATOMIC_INIT(0);
 static DEFINE_MUTEX(pmc_grab_mutex);
 
+static void perf_stop_nmi_watchdog(void *unused)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+
+	stop_nmi_watchdog(NULL);
+	cpuc->pcr = pcr_ops->read();
+}
+
 void perf_event_grab_pmc(void)
 {
 	if (atomic_inc_not_zero(&active_events))
@@ -358,7 +694,7 @@ void perf_event_grab_pmc(void)
 	mutex_lock(&pmc_grab_mutex);
 	if (atomic_read(&active_events) == 0) {
 		if (atomic_read(&nmi_active) > 0) {
-			on_each_cpu(stop_nmi_watchdog, NULL, 1);
+			on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
 			BUG_ON(atomic_read(&nmi_active) != 0);
 		}
 		atomic_inc(&active_events);
@@ -375,30 +711,160 @@ void perf_event_release_pmc(void)
 	}
 }
 
+static const struct perf_event_map *sparc_map_cache_event(u64 config)
+{
+	unsigned int cache_type, cache_op, cache_result;
+	const struct perf_event_map *pmap;
+
+	if (!sparc_pmu->cache_map)
+		return ERR_PTR(-ENOENT);
+
+	cache_type = (config >>  0) & 0xff;
+	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
+		return ERR_PTR(-EINVAL);
+
+	cache_op = (config >>  8) & 0xff;
+	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
+		return ERR_PTR(-EINVAL);
+
+	cache_result = (config >> 16) & 0xff;
+	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+		return ERR_PTR(-EINVAL);
+
+	pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
+
+	if (pmap->encoding == CACHE_OP_UNSUPPORTED)
+		return ERR_PTR(-ENOENT);
+
+	if (pmap->encoding == CACHE_OP_NONSENSE)
+		return ERR_PTR(-EINVAL);
+
+	return pmap;
+}
+
 static void hw_perf_event_destroy(struct perf_event *event)
 {
 	perf_event_release_pmc();
 }
 
+/* Make sure all events can be scheduled into the hardware at
+ * the same time.  This is simplified by the fact that we only
+ * need to support 2 simultaneous HW events.
+ */
+static int sparc_check_constraints(unsigned long *events, int n_ev)
+{
+	if (n_ev <= perf_max_events) {
+		u8 msk1, msk2;
+		u16 dummy;
+
+		if (n_ev == 1)
+			return 0;
+		BUG_ON(n_ev != 2);
+		perf_event_decode(events[0], &dummy, &msk1);
+		perf_event_decode(events[1], &dummy, &msk2);
+
+		/* If both events can go on any counter, OK.  */
+		if (msk1 == (PIC_UPPER | PIC_LOWER) &&
+		    msk2 == (PIC_UPPER | PIC_LOWER))
+			return 0;
+
+		/* If one event is limited to a specific counter,
+		 * and the other can go on both, OK.
+		 */
+		if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
+		    msk2 == (PIC_UPPER | PIC_LOWER))
+			return 0;
+		if ((msk2 == PIC_UPPER || msk2 == PIC_LOWER) &&
+		    msk1 == (PIC_UPPER | PIC_LOWER))
+			return 0;
+
+		/* If the events are fixed to different counters, OK.  */
+		if ((msk1 == PIC_UPPER && msk2 == PIC_LOWER) ||
+		    (msk1 == PIC_LOWER && msk2 == PIC_UPPER))
+			return 0;
+
+		/* Otherwise, there is a conflict.  */
+	}
+
+	return -1;
+}
+
+static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
+{
+	int eu = 0, ek = 0, eh = 0;
+	struct perf_event *event;
+	int i, n, first;
+
+	n = n_prev + n_new;
+	if (n <= 1)
+		return 0;
+
+	first = 1;
+	for (i = 0; i < n; i++) {
+		event = evts[i];
+		if (first) {
+			eu = event->attr.exclude_user;
+			ek = event->attr.exclude_kernel;
+			eh = event->attr.exclude_hv;
+			first = 0;
+		} else if (event->attr.exclude_user != eu ||
+			   event->attr.exclude_kernel != ek ||
+			   event->attr.exclude_hv != eh) {
+			return -EAGAIN;
+		}
+	}
+
+	return 0;
+}
+
+static int collect_events(struct perf_event *group, int max_count,
+			  struct perf_event *evts[], unsigned long *events)
+{
+	struct perf_event *event;
+	int n = 0;
+
+	if (!is_software_event(group)) {
+		if (n >= max_count)
+			return -1;
+		evts[n] = group;
+		events[n++] = group->hw.event_base;
+	}
+	list_for_each_entry(event, &group->sibling_list, group_entry) {
+		if (!is_software_event(event) &&
+		    event->state != PERF_EVENT_STATE_OFF) {
+			if (n >= max_count)
+				return -1;
+			evts[n] = event;
+			events[n++] = event->hw.event_base;
+		}
+	}
+	return n;
+}
+
 static int __hw_perf_event_init(struct perf_event *event)
 {
 	struct perf_event_attr *attr = &event->attr;
+	struct perf_event *evts[MAX_HWEVENTS];
 	struct hw_perf_event *hwc = &event->hw;
+	unsigned long events[MAX_HWEVENTS];
 	const struct perf_event_map *pmap;
 	u64 enc;
+	int n;
 
 	if (atomic_read(&nmi_active) < 0)
 		return -ENODEV;
 
-	if (attr->type != PERF_TYPE_HARDWARE)
+	if (attr->type == PERF_TYPE_HARDWARE) {
+		if (attr->config >= sparc_pmu->max_events)
+			return -EINVAL;
+		pmap = sparc_pmu->event_map(attr->config);
+	} else if (attr->type == PERF_TYPE_HW_CACHE) {
+		pmap = sparc_map_cache_event(attr->config);
+		if (IS_ERR(pmap))
+			return PTR_ERR(pmap);
+	} else
 		return -EOPNOTSUPP;
 
-	if (attr->config >= sparc_pmu->max_events)
-		return -EINVAL;
-
-	perf_event_grab_pmc();
-	event->destroy = hw_perf_event_destroy;
-
 	/* We save the enable bits in the config_base.  So to
 	 * turn off sampling just write 'config', and to enable
 	 * things write 'config | config_base'.
@@ -411,15 +877,39 @@ static int __hw_perf_event_init(struct perf_event *event)
 	if (!attr->exclude_hv)
 		hwc->config_base |= sparc_pmu->hv_bit;
 
+	hwc->event_base = perf_event_encode(pmap);
+
+	enc = pmap->encoding;
+
+	n = 0;
+	if (event->group_leader != event) {
+		n = collect_events(event->group_leader,
+				   perf_max_events - 1,
+				   evts, events);
+		if (n < 0)
+			return -EINVAL;
+	}
+	events[n] = hwc->event_base;
+	evts[n] = event;
+
+	if (check_excludes(evts, n, 1))
+		return -EINVAL;
+
+	if (sparc_check_constraints(events, n + 1))
+		return -EINVAL;
+
+	/* Try to do all error checking before this point, as unwinding
+	 * state after grabbing the PMC is difficult.
+	 */
+	perf_event_grab_pmc();
+	event->destroy = hw_perf_event_destroy;
+
 	if (!hwc->sample_period) {
 		hwc->sample_period = MAX_PERIOD;
 		hwc->last_period = hwc->sample_period;
 		atomic64_set(&hwc->period_left, hwc->sample_period);
 	}
 
-	pmap = sparc_pmu->event_map(attr->config);
-
-	enc = pmap->encoding;
 	if (pmap->pic_mask & PIC_UPPER) {
 		hwc->idx = PIC_UPPER_INDEX;
 		enc <<= sparc_pmu->upper_shift;
@@ -472,7 +962,7 @@ void perf_event_print_debug(void)
 }
 
 static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
-					      unsigned long cmd, void *__args)
+					    unsigned long cmd, void *__args)
 {
 	struct die_args *args = __args;
 	struct perf_sample_data data;
@@ -513,7 +1003,7 @@ static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
 			continue;
 
 		if (perf_event_overflow(event, 1, &data, regs))
-			sparc_pmu_disable_event(hwc, idx);
+			sparc_pmu_disable_event(cpuc, hwc, idx);
 	}
 
 	return NOTIFY_STOP;
@@ -525,8 +1015,15 @@ static __read_mostly struct notifier_block perf_event_nmi_notifier = {
 
 static bool __init supported_pmu(void)
 {
-	if (!strcmp(sparc_pmu_type, "ultra3i")) {
-		sparc_pmu = &ultra3i_pmu;
+	if (!strcmp(sparc_pmu_type, "ultra3") ||
+	    !strcmp(sparc_pmu_type, "ultra3+") ||
+	    !strcmp(sparc_pmu_type, "ultra3i") ||
+	    !strcmp(sparc_pmu_type, "ultra4+")) {
+		sparc_pmu = &ultra3_pmu;
+		return true;
+	}
+	if (!strcmp(sparc_pmu_type, "niagara")) {
+		sparc_pmu = &niagara1_pmu;
 		return true;
 	}
 	if (!strcmp(sparc_pmu_type, "niagara2")) {