sparc64: Initial hw perf counter support.

Only supports one simple counter and only UltraSPARC-IIIi chips.

Signed-off-by: David S. Miller <davem@davemloft.net>

5 files changed, 538 insertions, 2 deletions

diff --git a/arch/sparc/include/asm/nmi.h b/arch/sparc/include/asm/nmi.h
index c7d11e435df9..72e6500e7ab0 100644
--- a/arch/sparc/include/asm/nmi.h
+++ b/arch/sparc/include/asm/nmi.h
@@ -7,4 +7,7 @@ extern void nmi_adjust_hz(unsigned int new_hz);
 
 extern atomic_t nmi_active;
 
+extern void start_nmi_watchdog(void *unused);
+extern void stop_nmi_watchdog(void *unused);
+
 #endif /* __NMI_H */
diff --git a/arch/sparc/include/asm/perf_counter.h b/arch/sparc/include/asm/perf_counter.h
index 38d644546435..5d7a8ca0e491 100644
--- a/arch/sparc/include/asm/perf_counter.h
+++ b/arch/sparc/include/asm/perf_counter.h
@@ -5,4 +5,10 @@ extern void set_perf_counter_pending(void);
 
 #define	PERF_COUNTER_INDEX_OFFSET	0
 
+#ifdef CONFIG_PERF_COUNTERS
+extern void init_hw_perf_counters(void);
+#else
+static inline void init_hw_perf_counters(void)	{ }
+#endif
+
 #endif
diff --git a/arch/sparc/kernel/Makefile b/arch/sparc/kernel/Makefile
index b0041756c9f0..f96dc5761f74 100644
--- a/arch/sparc/kernel/Makefile
+++ b/arch/sparc/kernel/Makefile
@@ -103,3 +103,6 @@ obj-$(CONFIG_SUN_LDOMS) += ldc.o vio.o viohs.o ds.o
 obj-$(CONFIG_AUDIT)     += audit.o
 audit--$(CONFIG_AUDIT)  := compat_audit.o
 obj-$(CONFIG_COMPAT)    += $(audit--y)
+
+pc--$(CONFIG_PERF_COUNTERS) := perf_counter.o
+obj-$(CONFIG_SPARC64)	+= $(pc--y)
diff --git a/arch/sparc/kernel/nmi.c b/arch/sparc/kernel/nmi.c
index 7de19dd30f40..391a6ed9a184 100644
--- a/arch/sparc/kernel/nmi.c
+++ b/arch/sparc/kernel/nmi.c
@@ -19,6 +19,7 @@
 #include <linux/delay.h>
 #include <linux/smp.h>
 
+#include <asm/perf_counter.h>
 #include <asm/ptrace.h>
 #include <asm/local.h>
 #include <asm/pcr.h>
@@ -156,7 +157,7 @@ static void report_broken_nmi(int cpu, int *prev_nmi_count)
 	atomic_dec(&nmi_active);
 }
 
-static void stop_nmi_watchdog(void *unused)
+void stop_nmi_watchdog(void *unused)
 {
 	pcr_ops->write(PCR_PIC_PRIV);
 	__get_cpu_var(wd_enabled) = 0;
@@ -210,7 +211,7 @@ error:
 	return err;
 }
 
-static void start_nmi_watchdog(void *unused)
+void start_nmi_watchdog(void *unused)
 {
 	__get_cpu_var(wd_enabled) = 1;
 	atomic_inc(&nmi_active);
@@ -263,6 +264,9 @@ int __init nmi_init(void)
 			atomic_set(&nmi_active, -1);
 		}
 	}
+	if (!err)
+		init_hw_perf_counters();
+
 	return err;
 }
 
diff --git a/arch/sparc/kernel/perf_counter.c b/arch/sparc/kernel/perf_counter.c
new file mode 100644
index 000000000000..f2c781450d08
--- /dev/null
+++ b/arch/sparc/kernel/perf_counter.c
@@ -0,0 +1,520 @@
+/* Performance counter support for sparc64.
+ *
+ * Copyright (C) 2009 David S. Miller <davem@davemloft.net>
+ *
+ * This code is based almost entirely upon the x86 perf counter
+ * code, which is:
+ *
+ *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
+ *  Copyright (C) 2009 Jaswinder Singh Rajput
+ *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
+ */
+
+#include <linux/perf_counter.h>
+#include <linux/kprobes.h>
+#include <linux/kernel.h>
+#include <linux/kdebug.h>
+#include <linux/mutex.h>
+
+#include <asm/cpudata.h>
+#include <asm/atomic.h>
+#include <asm/nmi.h>
+#include <asm/pcr.h>
+
+/* Sparc64 chips have two performance counters, 32-bits each, with
+ * overflow interrupts generated on transition from 0xffffffff to 0.
+ * The counters are accessed in one go using a 64-bit register.
+ *
+ * Both counters are controlled using a single control register.  The
+ * only way to stop all sampling is to clear all of the context (user,
+ * supervisor, hypervisor) sampling enable bits.  But these bits apply
+ * to both counters, thus the two counters can't be enabled/disabled
+ * individually.
+ *
+ * The control register has two event fields, one for each of the two
+ * counters.  It's thus nearly impossible to have one counter going
+ * while keeping the other one stopped.  Therefore it is possible to
+ * get overflow interrupts for counters not currently "in use" and
+ * that condition must be checked in the overflow interrupt handler.
+ *
+ * So we use a hack, in that we program inactive counters with the
+ * "sw_count0" and "sw_count1" events.  These count how many times
+ * the instruction "sethi %hi(0xfc000), %g0" is executed.  It's an
+ * unusual way to encode a NOP and therefore will not trigger in
+ * normal code.
+ */
+
+#define MAX_HWCOUNTERS			2
+#define MAX_PERIOD			((1UL << 32) - 1)
+
+#define PIC_UPPER_INDEX			0
+#define PIC_LOWER_INDEX			1
+
+#define PIC_UPPER_NOP			0x1c
+#define PIC_LOWER_NOP			0x14
+
+struct cpu_hw_counters {
+	struct perf_counter	*counters[MAX_HWCOUNTERS];
+	unsigned long		used_mask[BITS_TO_LONGS(MAX_HWCOUNTERS)];
+	unsigned long		active_mask[BITS_TO_LONGS(MAX_HWCOUNTERS)];
+	int enabled;
+};
+DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = { .enabled = 1, };
+
+struct perf_event_map {
+	u16	encoding;
+	u8	pic_mask;
+#define PIC_NONE	0x00
+#define PIC_UPPER	0x01
+#define PIC_LOWER	0x02
+};
+
+struct sparc_pmu {
+	const struct perf_event_map	*(*event_map)(int);
+	int				max_events;
+	int				upper_shift;
+	int				lower_shift;
+	int				event_mask;
+};
+
+static const struct perf_event_map ultra3i_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)
+{
+	return &ultra3i_perfmon_event_map[event];
+}
+
+static const struct sparc_pmu ultra3i_pmu = {
+	.event_map	= ultra3i_event_map,
+	.max_events	= ARRAY_SIZE(ultra3i_perfmon_event_map),
+	.upper_shift	= 11,
+	.lower_shift	= 4,
+	.event_mask	= 0x3f,
+};
+
+static const struct sparc_pmu *sparc_pmu __read_mostly;
+
+static u64 event_encoding(u64 event, int idx)
+{
+	if (idx == PIC_UPPER_INDEX)
+		event <<= sparc_pmu->upper_shift;
+	else
+		event <<= sparc_pmu->lower_shift;
+	return event;
+}
+
+static u64 mask_for_index(int idx)
+{
+	return event_encoding(sparc_pmu->event_mask, idx);
+}
+
+static u64 nop_for_index(int idx)
+{
+	return event_encoding(idx == PIC_UPPER_INDEX ?
+			      PIC_UPPER_NOP : PIC_LOWER_NOP, idx);
+}
+
+static inline void sparc_pmu_enable_counter(struct hw_perf_counter *hwc,
+					    int idx)
+{
+	u64 val, mask = mask_for_index(idx);
+
+	val = pcr_ops->read();
+	pcr_ops->write((val & ~mask) | hwc->config);
+}
+
+static inline void sparc_pmu_disable_counter(struct hw_perf_counter *hwc,
+					     int idx)
+{
+	u64 mask = mask_for_index(idx);
+	u64 nop = nop_for_index(idx);
+	u64 val = pcr_ops->read();
+
+	pcr_ops->write((val & ~mask) | nop);
+}
+
+void hw_perf_enable(void)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	u64 val;
+	int i;
+
+	if (cpuc->enabled)
+		return;
+
+	cpuc->enabled = 1;
+	barrier();
+
+	val = pcr_ops->read();
+
+	for (i = 0; i < MAX_HWCOUNTERS; i++) {
+		struct perf_counter *cp = cpuc->counters[i];
+		struct hw_perf_counter *hwc;
+
+		if (!cp)
+			continue;
+		hwc = &cp->hw;
+		val |= hwc->config_base;
+	}
+
+	pcr_ops->write(val);
+}
+
+void hw_perf_disable(void)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	u64 val;
+
+	if (!cpuc->enabled)
+		return;
+
+	cpuc->enabled = 0;
+
+	val = pcr_ops->read();
+	val &= ~(PCR_UTRACE | PCR_STRACE);
+	pcr_ops->write(val);
+}
+
+static u32 read_pmc(int idx)
+{
+	u64 val;
+
+	read_pic(val);
+	if (idx == PIC_UPPER_INDEX)
+		val >>= 32;
+
+	return val & 0xffffffff;
+}
+
+static void write_pmc(int idx, u64 val)
+{
+	u64 shift, mask, pic;
+
+	shift = 0;
+	if (idx == PIC_UPPER_INDEX)
+		shift = 32;
+
+	mask = ((u64) 0xffffffff) << shift;
+	val <<= shift;
+
+	read_pic(pic);
+	pic &= ~mask;
+	pic |= val;
+	write_pic(pic);
+}
+
+static int sparc_perf_counter_set_period(struct perf_counter *counter,
+					 struct hw_perf_counter *hwc, int idx)
+{
+	s64 left = atomic64_read(&hwc->period_left);
+	s64 period = hwc->sample_period;
+	int ret = 0;
+
+	if (unlikely(left <= -period)) {
+		left = period;
+		atomic64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (unlikely(left <= 0)) {
+		left += period;
+		atomic64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+	if (left > MAX_PERIOD)
+		left = MAX_PERIOD;
+
+	atomic64_set(&hwc->prev_count, (u64)-left);
+
+	write_pmc(idx, (u64)(-left) & 0xffffffff);
+
+	perf_counter_update_userpage(counter);
+
+	return ret;
+}
+
+static int sparc_pmu_enable(struct perf_counter *counter)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	struct hw_perf_counter *hwc = &counter->hw;
+	int idx = hwc->idx;
+
+	if (test_and_set_bit(idx, cpuc->used_mask))
+		return -EAGAIN;
+
+	sparc_pmu_disable_counter(hwc, idx);
+
+	cpuc->counters[idx] = counter;
+	set_bit(idx, cpuc->active_mask);
+
+	sparc_perf_counter_set_period(counter, hwc, idx);
+	sparc_pmu_enable_counter(hwc, idx);
+	perf_counter_update_userpage(counter);
+	return 0;
+}
+
+static u64 sparc_perf_counter_update(struct perf_counter *counter,
+				     struct hw_perf_counter *hwc, int idx)
+{
+	int shift = 64 - 32;
+	u64 prev_raw_count, new_raw_count;
+	s64 delta;
+
+again:
+	prev_raw_count = atomic64_read(&hwc->prev_count);
+	new_raw_count = read_pmc(idx);
+
+	if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
+			     new_raw_count) != prev_raw_count)
+		goto again;
+
+	delta = (new_raw_count << shift) - (prev_raw_count << shift);
+	delta >>= shift;
+
+	atomic64_add(delta, &counter->count);
+	atomic64_sub(delta, &hwc->period_left);
+
+	return new_raw_count;
+}
+
+static void sparc_pmu_disable(struct perf_counter *counter)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	struct hw_perf_counter *hwc = &counter->hw;
+	int idx = hwc->idx;
+
+	clear_bit(idx, cpuc->active_mask);
+	sparc_pmu_disable_counter(hwc, idx);
+
+	barrier();
+
+	sparc_perf_counter_update(counter, hwc, idx);
+	cpuc->counters[idx] = NULL;
+	clear_bit(idx, cpuc->used_mask);
+
+	perf_counter_update_userpage(counter);
+}
+
+static void sparc_pmu_read(struct perf_counter *counter)
+{
+	struct hw_perf_counter *hwc = &counter->hw;
+	sparc_perf_counter_update(counter, hwc, hwc->idx);
+}
+
+static void sparc_pmu_unthrottle(struct perf_counter *counter)
+{
+	struct hw_perf_counter *hwc = &counter->hw;
+	sparc_pmu_enable_counter(hwc, hwc->idx);
+}
+
+static atomic_t active_counters = ATOMIC_INIT(0);
+static DEFINE_MUTEX(pmc_grab_mutex);
+
+void perf_counter_grab_pmc(void)
+{
+	if (atomic_inc_not_zero(&active_counters))
+		return;
+
+	mutex_lock(&pmc_grab_mutex);
+	if (atomic_read(&active_counters) == 0) {
+		if (atomic_read(&nmi_active) > 0) {
+			on_each_cpu(stop_nmi_watchdog, NULL, 1);
+			BUG_ON(atomic_read(&nmi_active) != 0);
+		}
+		atomic_inc(&active_counters);
+	}
+	mutex_unlock(&pmc_grab_mutex);
+}
+
+void perf_counter_release_pmc(void)
+{
+	if (atomic_dec_and_mutex_lock(&active_counters, &pmc_grab_mutex)) {
+		if (atomic_read(&nmi_active) == 0)
+			on_each_cpu(start_nmi_watchdog, NULL, 1);
+		mutex_unlock(&pmc_grab_mutex);
+	}
+}
+
+static void hw_perf_counter_destroy(struct perf_counter *counter)
+{
+	perf_counter_release_pmc();
+}
+
+static int __hw_perf_counter_init(struct perf_counter *counter)
+{
+	struct perf_counter_attr *attr = &counter->attr;
+	struct hw_perf_counter *hwc = &counter->hw;
+	const struct perf_event_map *pmap;
+	u64 enc;
+
+	if (atomic_read(&nmi_active) < 0)
+		return -ENODEV;
+
+	if (attr->type != PERF_TYPE_HARDWARE)
+		return -EOPNOTSUPP;
+
+	if (attr->config >= sparc_pmu->max_events)
+		return -EINVAL;
+
+	perf_counter_grab_pmc();
+	counter->destroy = hw_perf_counter_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'.
+	 */
+	hwc->config_base = 0;
+	if (!attr->exclude_user)
+		hwc->config_base |= PCR_UTRACE;
+	if (!attr->exclude_kernel)
+		hwc->config_base |= PCR_STRACE;
+
+	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;
+	} else {
+		hwc->idx = PIC_LOWER_INDEX;
+		enc <<= sparc_pmu->lower_shift;
+	}
+
+	hwc->config |= enc;
+	return 0;
+}
+
+static const struct pmu pmu = {
+	.enable		= sparc_pmu_enable,
+	.disable	= sparc_pmu_disable,
+	.read		= sparc_pmu_read,
+	.unthrottle	= sparc_pmu_unthrottle,
+};
+
+const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
+{
+	int err = __hw_perf_counter_init(counter);
+
+	if (err)
+		return ERR_PTR(err);
+	return &pmu;
+}
+
+void perf_counter_print_debug(void)
+{
+	unsigned long flags;
+	u64 pcr, pic;
+	int cpu;
+
+	if (!sparc_pmu)
+		return;
+
+	local_irq_save(flags);
+
+	cpu = smp_processor_id();
+
+	pcr = pcr_ops->read();
+	read_pic(pic);
+
+	pr_info("\n");
+	pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n",
+		cpu, pcr, pic);
+
+	local_irq_restore(flags);
+}
+
+static int __kprobes perf_counter_nmi_handler(struct notifier_block *self,
+					      unsigned long cmd, void *__args)
+{
+	struct die_args *args = __args;
+	struct perf_sample_data data;
+	struct cpu_hw_counters *cpuc;
+	struct pt_regs *regs;
+	int idx;
+
+	if (!atomic_read(&active_counters))
+		return NOTIFY_DONE;
+
+	switch (cmd) {
+	case DIE_NMI:
+		break;
+
+	default:
+		return NOTIFY_DONE;
+	}
+
+	regs = args->regs;
+
+	data.regs = regs;
+	data.addr = 0;
+
+	cpuc = &__get_cpu_var(cpu_hw_counters);
+	for (idx = 0; idx < MAX_HWCOUNTERS; idx++) {
+		struct perf_counter *counter = cpuc->counters[idx];
+		struct hw_perf_counter *hwc;
+		u64 val;
+
+		if (!test_bit(idx, cpuc->active_mask))
+			continue;
+		hwc = &counter->hw;
+		val = sparc_perf_counter_update(counter, hwc, idx);
+		if (val & (1ULL << 31))
+			continue;
+
+		data.period = counter->hw.last_period;
+		if (!sparc_perf_counter_set_period(counter, hwc, idx))
+			continue;
+
+		if (perf_counter_overflow(counter, 1, &data))
+			sparc_pmu_disable_counter(hwc, idx);
+	}
+
+	return NOTIFY_STOP;
+}
+
+static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
+	.notifier_call		= perf_counter_nmi_handler,
+};
+
+static bool __init supported_pmu(void)
+{
+	if (!strcmp(sparc_pmu_type, "ultra3i")) {
+		sparc_pmu = &ultra3i_pmu;
+		return true;
+	}
+	return false;
+}
+
+void __init init_hw_perf_counters(void)
+{
+	pr_info("Performance counters: ");
+
+	if (!supported_pmu()) {
+		pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
+		return;
+	}
+
+	pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
+
+	/* All sparc64 PMUs currently have 2 counters.  But this simple
+	 * driver only supports one active counter at a time.
+	 */
+	perf_max_counters = 1;
+
+	register_die_notifier(&perf_counter_nmi_notifier);
+}