metag: Perf

Add Perf support for metag.
Signed-off-by: James Hogan <james.hogan@imgtec.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>

arch/metag/Kconfig

@@ -22,6 +22,7 @@ config METAG
 	select HAVE_MEMBLOCK
 	select HAVE_MEMBLOCK_NODE_MAP
 	select HAVE_MOD_ARCH_SPECIFIC
+	select HAVE_PERF_EVENTS
 	select HAVE_SYSCALL_TRACEPOINTS
 	select IRQ_DOMAIN
 	select MODULES_USE_ELF_RELA

arch/metag/include/asm/perf_event.h

+#ifndef __ASM_METAG_PERF_EVENT_H
+#define __ASM_METAG_PERF_EVENT_H
+
+#endif /* __ASM_METAG_PERF_EVENT_H */

arch/metag/kernel/Makefile

@@ -25,6 +25,8 @@ obj-y	+= topology.o
 obj-y	+= traps.o
 obj-y	+= user_gateway.o
 
+obj-$(CONFIG_PERF_EVENTS)		+= perf/
+
 obj-$(CONFIG_METAG_COREMEM)		+= coremem.o
 obj-$(CONFIG_DYNAMIC_FTRACE)		+= ftrace.o
 obj-$(CONFIG_FUNCTION_TRACER)		+= ftrace_stub.o

arch/metag/kernel/perf/Makefile

+# Makefile for performance event core
+
+obj-y += perf_event.o

arch/metag/kernel/perf/perf_event.c

+/*
+ * Meta performance counter support.
+ *  Copyright (C) 2012 Imagination Technologies Ltd
+ *
+ * This code is based on the sh pmu code:
+ *  Copyright (C) 2009 Paul Mundt
+ *
+ * and on the arm pmu code:
+ *  Copyright (C) 2009 picoChip Designs, Ltd., James Iles
+ *  Copyright (C) 2010 ARM Ltd., Will Deacon <will.deacon@arm.com>
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ */
+
+#include <linux/atomic.h>
+#include <linux/export.h>
+#include <linux/init.h>
+#include <linux/irqchip/metag.h>
+#include <linux/perf_event.h>
+#include <linux/slab.h>
+
+#include <asm/core_reg.h>
+#include <asm/hwthread.h>
+#include <asm/io.h>
+#include <asm/irq.h>
+
+#include "perf_event.h"
+
+static int _hw_perf_event_init(struct perf_event *);
+static void _hw_perf_event_destroy(struct perf_event *);
+
+/* Determines which core type we are */
+static struct metag_pmu *metag_pmu __read_mostly;
+
+/* Processor specific data */
+static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
+
+/* PMU admin */
+const char *perf_pmu_name(void)
+{
+	if (metag_pmu)
+		return metag_pmu->pmu.name;
+
+	return NULL;
+}
+EXPORT_SYMBOL_GPL(perf_pmu_name);
+
+int perf_num_counters(void)
+{
+	if (metag_pmu)
+		return metag_pmu->max_events;
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(perf_num_counters);
+
+static inline int metag_pmu_initialised(void)
+{
+	return !!metag_pmu;
+}
+
+static void release_pmu_hardware(void)
+{
+	int irq;
+	unsigned int version = (metag_pmu->version &
+			(METAC_ID_MINOR_BITS | METAC_ID_REV_BITS)) >>
+			METAC_ID_REV_S;
+
+	/* Early cores don't have overflow interrupts */
+	if (version < 0x0104)
+		return;
+
+	irq = internal_irq_map(17);
+	if (irq >= 0)
+		free_irq(irq, (void *)1);
+
+	irq = internal_irq_map(16);
+	if (irq >= 0)
+		free_irq(irq, (void *)0);
+}
+
+static int reserve_pmu_hardware(void)
+{
+	int err = 0, irq[2];
+	unsigned int version = (metag_pmu->version &
+			(METAC_ID_MINOR_BITS | METAC_ID_REV_BITS)) >>
+			METAC_ID_REV_S;
+
+	/* Early cores don't have overflow interrupts */
+	if (version < 0x0104)
+		goto out;
+
+	/*
+	 * Bit 16 on HWSTATMETA is the interrupt for performance counter 0;
+	 * similarly, 17 is the interrupt for performance counter 1.
+	 * We can't (yet) interrupt on the cycle counter, because it's a
+	 * register, however it holds a 32-bit value as opposed to 24-bit.
+	 */
+	irq[0] = internal_irq_map(16);
+	if (irq[0] < 0) {
+		pr_err("unable to map internal IRQ %d\n", 16);
+		goto out;
+	}
+	err = request_irq(irq[0], metag_pmu->handle_irq, IRQF_NOBALANCING,
+			"metagpmu0", (void *)0);
+	if (err) {
+		pr_err("unable to request IRQ%d for metag PMU counters\n",
+				irq[0]);
+		goto out;
+	}
+
+	irq[1] = internal_irq_map(17);
+	if (irq[1] < 0) {
+		pr_err("unable to map internal IRQ %d\n", 17);
+		goto out_irq1;
+	}
+	err = request_irq(irq[1], metag_pmu->handle_irq, IRQF_NOBALANCING,
+			"metagpmu1", (void *)1);
+	if (err) {
+		pr_err("unable to request IRQ%d for metag PMU counters\n",
+				irq[1]);
+		goto out_irq1;
+	}
+
+	return 0;
+
+out_irq1:
+	free_irq(irq[0], (void *)0);
+out:
+	return err;
+}
+
+/* PMU operations */
+static void metag_pmu_enable(struct pmu *pmu)
+{
+}
+
+static void metag_pmu_disable(struct pmu *pmu)
+{
+}
+
+static int metag_pmu_event_init(struct perf_event *event)
+{
+	int err = 0;
+	atomic_t *active_events = &metag_pmu->active_events;
+
+	if (!metag_pmu_initialised()) {
+		err = -ENODEV;
+		goto out;
+	}
+
+	if (has_branch_stack(event))
+		return -EOPNOTSUPP;
+
+	event->destroy = _hw_perf_event_destroy;
+
+	if (!atomic_inc_not_zero(active_events)) {
+		mutex_lock(&metag_pmu->reserve_mutex);
+		if (atomic_read(active_events) == 0)
+			err = reserve_pmu_hardware();
+
+		if (!err)
+			atomic_inc(active_events);
+
+		mutex_unlock(&metag_pmu->reserve_mutex);
+	}
+
+	/* Hardware and caches counters */
+	switch (event->attr.type) {
+	case PERF_TYPE_HARDWARE:
+	case PERF_TYPE_HW_CACHE:
+		err = _hw_perf_event_init(event);
+		break;
+
+	default:
+		return -ENOENT;
+	}
+
+	if (err)
+		event->destroy(event);
+
+out:
+	return err;
+}
+
+void metag_pmu_event_update(struct perf_event *event,
+		struct hw_perf_event *hwc, int idx)
+{
+	u64 prev_raw_count, new_raw_count;
+	s64 delta;
+
+	/*
+	 * If this counter is chained, it may be that the previous counter
+	 * value has been changed beneath us.
+	 *
+	 * To get around this, we read and exchange the new raw count, then
+	 * add the delta (new - prev) to the generic counter atomically.
+	 *
+	 * Without interrupts, this is the simplest approach.
+	 */
+again:
+	prev_raw_count = local64_read(&hwc->prev_count);
+	new_raw_count = metag_pmu->read(idx);
+
+	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
+			new_raw_count) != prev_raw_count)
+		goto again;
+
+	/*
+	 * Calculate the delta and add it to the counter.
+	 */
+	delta = new_raw_count - prev_raw_count;
+
+	local64_add(delta, &event->count);
+}
+
+int metag_pmu_event_set_period(struct perf_event *event,
+		struct hw_perf_event *hwc, int idx)
+{
+	s64 left = local64_read(&hwc->period_left);
+	s64 period = hwc->sample_period;
+	int ret = 0;
+
+	if (unlikely(left <= -period)) {
+		left = period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (unlikely(left <= 0)) {
+		left += period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (left > (s64)metag_pmu->max_period)
+		left = metag_pmu->max_period;
+
+	if (metag_pmu->write)
+		metag_pmu->write(idx, (u64)(-left) & MAX_PERIOD);
+
+	perf_event_update_userpage(event);
+
+	return ret;
+}
+
+static void metag_pmu_start(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	/*
+	 * We always have to reprogram the period, so ignore PERF_EF_RELOAD.
+	 */
+	if (flags & PERF_EF_RELOAD)
+		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
+
+	hwc->state = 0;
+
+	/*
+	 * Reset the period.
+	 * Some counters can't be stopped (i.e. are core global), so when the
+	 * counter was 'stopped' we merely disabled the IRQ. If we don't reset
+	 * the period, then we'll either: a) get an overflow too soon;
+	 * or b) too late if the overflow happened since disabling.
+	 * Obviously, this has little bearing on cores without the overflow
+	 * interrupt, as the performance counter resets to zero on write
+	 * anyway.
+	 */
+	if (metag_pmu->max_period)
+		metag_pmu_event_set_period(event, hwc, hwc->idx);
+	cpuc->events[idx] = event;
+	metag_pmu->enable(hwc, idx);
+}
+
+static void metag_pmu_stop(struct perf_event *event, int flags)
+{
+	struct hw_perf_event *hwc = &event->hw;
+
+	/*
+	 * We should always update the counter on stop; see comment above
+	 * why.
+	 */
+	if (!(hwc->state & PERF_HES_STOPPED)) {
+		metag_pmu_event_update(event, hwc, hwc->idx);
+		metag_pmu->disable(hwc, hwc->idx);
+		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
+	}
+}
+
+static int metag_pmu_add(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = 0, ret = 0;
+
+	perf_pmu_disable(event->pmu);
+
+	/* check whether we're counting instructions */
+	if (hwc->config == 0x100) {
+		if (__test_and_set_bit(METAG_INST_COUNTER,
+				cpuc->used_mask)) {
+			ret = -EAGAIN;
+			goto out;
+		}
+		idx = METAG_INST_COUNTER;
+	} else {
+		/* Check whether we have a spare counter */
+		idx = find_first_zero_bit(cpuc->used_mask,
+				atomic_read(&metag_pmu->active_events));
+		if (idx >= METAG_INST_COUNTER) {
+			ret = -EAGAIN;
+			goto out;
+		}
+
+		__set_bit(idx, cpuc->used_mask);
+	}
+	hwc->idx = idx;
+
+	/* Make sure the counter is disabled */
+	metag_pmu->disable(hwc, idx);
+
+	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
+	if (flags & PERF_EF_START)
+		metag_pmu_start(event, PERF_EF_RELOAD);
+
+	perf_event_update_userpage(event);
+out:
+	perf_pmu_enable(event->pmu);
+	return ret;
+}
+
+static void metag_pmu_del(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+
+	WARN_ON(idx < 0);
+	metag_pmu_stop(event, PERF_EF_UPDATE);
+	cpuc->events[idx] = NULL;
+	__clear_bit(idx, cpuc->used_mask);
+
+	perf_event_update_userpage(event);
+}
+
+static void metag_pmu_read(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+
+	/* Don't read disabled counters! */
+	if (hwc->idx < 0)
+		return;
+
+	metag_pmu_event_update(event, hwc, hwc->idx);
+}
+
+static struct pmu pmu = {
+	.pmu_enable	= metag_pmu_enable,
+	.pmu_disable	= metag_pmu_disable,
+
+	.event_init	= metag_pmu_event_init,
+
+	.add		= metag_pmu_add,
+	.del		= metag_pmu_del,
+	.start		= metag_pmu_start,
+	.stop		= metag_pmu_stop,
+	.read		= metag_pmu_read,
+};
+
+/* Core counter specific functions */
+static const int metag_general_events[] = {
+	[PERF_COUNT_HW_CPU_CYCLES] = 0x03,
+	[PERF_COUNT_HW_INSTRUCTIONS] = 0x100,
+	[PERF_COUNT_HW_CACHE_REFERENCES] = -1,
+	[PERF_COUNT_HW_CACHE_MISSES] = -1,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
+	[PERF_COUNT_HW_BRANCH_MISSES] = -1,
+	[PERF_COUNT_HW_BUS_CYCLES] = -1,
+	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = -1,
+	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = -1,
+	[PERF_COUNT_HW_REF_CPU_CYCLES] = -1,
+};
+
+static const int metag_pmu_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
+	[C(L1D)] = {
+		[C(OP_READ)] = {
+			[C(RESULT_ACCESS)] = 0x08,
+			[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,
+		},
+	},
+	[C(L1I)] = {
+		[C(OP_READ)] = {
+			[C(RESULT_ACCESS)] = 0x09,
+			[C(RESULT_MISS)] = 0x0a,
+		},
+		[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(LL)] = {
+		[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,
+		},
+	},
+	[C(DTLB)] = {
+		[C(OP_READ)] = {
+			[C(RESULT_ACCESS)] = 0xd0,
+			[C(RESULT_MISS)] = 0xd2,
+		},
+		[C(OP_WRITE)] = {
+			[C(RESULT_ACCESS)] = 0xd4,
+			[C(RESULT_MISS)] = 0xd5,
+		},
+		[C(OP_PREFETCH)] = {
+			[C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+			[C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+		},
+	},
+	[C(ITLB)] = {
+		[C(OP_READ)] = {
+			[C(RESULT_ACCESS)] = 0xd1,
+			[C(RESULT_MISS)] = 0xd3,
+		},
+		[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,
+		},
+	},
+	[C(NODE)] = {
+		[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 void _hw_perf_event_destroy(struct perf_event *event)
+{
+	atomic_t *active_events = &metag_pmu->active_events;
+	struct mutex *pmu_mutex = &metag_pmu->reserve_mutex;
+
+	if (atomic_dec_and_mutex_lock(active_events, pmu_mutex)) {
+		release_pmu_hardware();
+		mutex_unlock(pmu_mutex);
+	}
+}
+
+static int _hw_perf_cache_event(int config, int *evp)
+{
+	unsigned long type, op, result;
+	int ev;
+
+	if (!metag_pmu->cache_events)
+		return -EINVAL;
+
+	/* Unpack config */
+	type = config & 0xff;
+	op = (config >> 8) & 0xff;
+	result = (config >> 16) & 0xff;
+
+	if (type >= PERF_COUNT_HW_CACHE_MAX ||
+			op >= PERF_COUNT_HW_CACHE_OP_MAX ||
+			result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+		return -EINVAL;
+
+	ev = (*metag_pmu->cache_events)[type][op][result];
+	if (ev == 0)
+		return -EOPNOTSUPP;
+	if (ev == -1)
+		return -EINVAL;
+	*evp = ev;
+	return 0;
+}
+
+static int _hw_perf_event_init(struct perf_event *event)
+{
+	struct perf_event_attr *attr = &event->attr;
+	struct hw_perf_event *hwc = &event->hw;
+	int mapping = 0, err;
+
+	switch (attr->type) {
+	case PERF_TYPE_HARDWARE:
+		if (attr->config >= PERF_COUNT_HW_MAX)
+			return -EINVAL;
+
+		mapping = metag_pmu->event_map(attr->config);
+		break;
+
+	case PERF_TYPE_HW_CACHE:
+		err = _hw_perf_cache_event(attr->config, &mapping);
+		if (err)
+			return err;
+		break;
+	}
+
+	/* Return early if the event is unsupported */
+	if (mapping == -1)
+		return -EINVAL;
+
+	/*
+	 * Early cores have "limited" counters - they have no overflow
+	 * interrupts - and so are unable to do sampling without extra work
+	 * and timer assistance.
+	 */
+	if (metag_pmu->max_period == 0) {
+		if (hwc->sample_period)
+			return -EINVAL;
+	}
+
+	/*
+	 * Don't assign an index until the event is placed into the hardware.
+	 * -1 signifies that we're still deciding where to put it. On SMP
+	 * systems each core has its own set of counters, so we can't do any
+	 * constraint checking yet.
+	 */
+	hwc->idx = -1;
+
+	/* Store the event encoding */
+	hwc->config |= (unsigned long)mapping;
+
+	/*
+	 * For non-sampling runs, limit the sample_period to half of the
+	 * counter width. This way, the new counter value should be less
+	 * likely to overtake the previous one (unless there are IRQ latency
+	 * issues...)
+	 */
+	if (metag_pmu->max_period) {
+		if (!hwc->sample_period) {
+			hwc->sample_period = metag_pmu->max_period >> 1;
+			hwc->last_period = hwc->sample_period;
+			local64_set(&hwc->period_left, hwc->sample_period);
+		}
+	}
+
+	return 0;
+}
+
+static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx)
+{
+	struct cpu_hw_events *events = &__get_cpu_var(cpu_hw_events);
+	unsigned int config = event->config;
+	unsigned int tmp = config & 0xf0;
+	unsigned long flags;
+
+	raw_spin_lock_irqsave(&events->pmu_lock, flags);
+
+	/*
+	 * Check if we're enabling the instruction counter (index of
+	 * MAX_HWEVENTS - 1)
+	 */
+	if (METAG_INST_COUNTER == idx) {
+		WARN_ONCE((config != 0x100),
+			"invalid configuration (%d) for counter (%d)\n",
+			config, idx);
+
+		/* Reset the cycle count */
+		__core_reg_set(TXTACTCYC, 0);
+		goto unlock;
+	}
+
+	/* Check for a core internal or performance channel event. */
+	if (tmp) {
+		void *perf_addr = (void *)PERF_COUNT(idx);
+
+		/*
+		 * Anything other than a cycle count will write the low-
+		 * nibble to the correct counter register.
+		 */
+		switch (tmp) {
+		case 0xd0:
+			perf_addr = (void *)PERF_ICORE(idx);
+			break;
+
+		case 0xf0:
+			perf_addr = (void *)PERF_CHAN(idx);
+			break;
+		}
+
+		metag_out32((tmp & 0x0f), perf_addr);
+
+		/*
+		 * Now we use the high nibble as the performance event to
+		 * to count.
+		 */
+		config = tmp >> 4;
+	}
+
+	/*
+	 * Enabled counters start from 0. Early cores clear the count on
+	 * write but newer cores don't, so we make sure that the count is
+	 * set to 0.
+	 */
+	tmp = ((config & 0xf) << 28) |
+			((1 << 24) << cpu_2_hwthread_id[get_cpu()]);
+	metag_out32(tmp, PERF_COUNT(idx));
+unlock:
+	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
+}
+
+static void metag_pmu_disable_counter(struct hw_perf_event *event, int idx)
+{
+	struct cpu_hw_events *events = &__get_cpu_var(cpu_hw_events);
+	unsigned int tmp = 0;
+	unsigned long flags;
+
+	/*
+	 * The cycle counter can't be disabled per se, as it's a hardware
+	 * thread register which is always counting. We merely return if this
+	 * is the counter we're attempting to disable.
+	 */
+	if (METAG_INST_COUNTER == idx)
+		return;
+
+	/*
+	 * The counter value _should_ have been read prior to disabling,
+	 * as if we're running on an early core then the value gets reset to
+	 * 0, and any read after that would be useless. On the newer cores,
+	 * however, it's better to read-modify-update this for purposes of
+	 * the overflow interrupt.
+	 * Here we remove the thread id AND the event nibble (there are at
+	 * least two events that count events that are core global and ignore
+	 * the thread id mask). This only works because we don't mix thread
+	 * performance counts, and event 0x00 requires a thread id mask!
+	 */
+	raw_spin_lock_irqsave(&events->pmu_lock, flags);
+
+	tmp = metag_in32(PERF_COUNT(idx));
+	tmp &= 0x00ffffff;
+	metag_out32(tmp, PERF_COUNT(idx));
+
+	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
+}
+
+static u64 metag_pmu_read_counter(int idx)
+{
+	u32 tmp = 0;
+
+	/* The act of reading the cycle counter also clears it */
+	if (METAG_INST_COUNTER == idx) {
+		__core_reg_swap(TXTACTCYC, tmp);
+		goto out;
+	}
+
+	tmp = metag_in32(PERF_COUNT(idx)) & 0x00ffffff;
+out:
+	return tmp;
+}
+
+static void metag_pmu_write_counter(int idx, u32 val)
+{
+	struct cpu_hw_events *events = &__get_cpu_var(cpu_hw_events);
+	u32 tmp = 0;
+	unsigned long flags;
+
+	/*
+	 * This _shouldn't_ happen, but if it does, then we can just
+	 * ignore the write, as the register is read-only and clear-on-write.
+	 */
+	if (METAG_INST_COUNTER == idx)
+		return;
+
+	/*
+	 * We'll keep the thread mask and event id, and just update the
+	 * counter itself. Also , we should bound the value to 24-bits.
+	 */
+	raw_spin_lock_irqsave(&events->pmu_lock, flags);
+
+	val &= 0x00ffffff;
+	tmp = metag_in32(PERF_COUNT(idx)) & 0xff000000;
+	val |= tmp;
+	metag_out32(val, PERF_COUNT(idx));
+
+	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
+}
+
+static int metag_pmu_event_map(int idx)
+{
+	return metag_general_events[idx];
+}
+
+static irqreturn_t metag_pmu_counter_overflow(int irq, void *dev)
+{
+	int idx = (int)dev;
+	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+	struct perf_event *event = cpuhw->events[idx];
+	struct hw_perf_event *hwc = &event->hw;
+	struct pt_regs *regs = get_irq_regs();
+	struct perf_sample_data sampledata;
+	unsigned long flags;
+	u32 counter = 0;
+
+	/*
+	 * We need to stop the core temporarily from generating another
+	 * interrupt while we disable this counter. However, we don't want
+	 * to flag the counter as free
+	 */
+	__global_lock2(flags);
+	counter = metag_in32(PERF_COUNT(idx));
+	metag_out32((counter & 0x00ffffff), PERF_COUNT(idx));
+	__global_unlock2(flags);
+
+	/* Update the counts and reset the sample period */
+	metag_pmu_event_update(event, hwc, idx);
+	perf_sample_data_init(&sampledata, 0, hwc->last_period);
+	metag_pmu_event_set_period(event, hwc, idx);
+
+	/*
+	 * Enable the counter again once core overflow processing has
+	 * completed.
+	 */
+	if (!perf_event_overflow(event, &sampledata, regs))
+		metag_out32(counter, PERF_COUNT(idx));
+
+	return IRQ_HANDLED;
+}
+
+static struct metag_pmu _metag_pmu = {
+	.handle_irq	= metag_pmu_counter_overflow,
+	.enable		= metag_pmu_enable_counter,
+	.disable	= metag_pmu_disable_counter,
+	.read		= metag_pmu_read_counter,
+	.write		= metag_pmu_write_counter,
+	.event_map	= metag_pmu_event_map,
+	.cache_events	= &metag_pmu_cache_events,
+	.max_period	= MAX_PERIOD,
+	.max_events	= MAX_HWEVENTS,
+};
+
+/* PMU CPU hotplug notifier */
+static int __cpuinit metag_pmu_cpu_notify(struct notifier_block *b,
+		unsigned long action, void *hcpu)
+{
+	unsigned int cpu = (unsigned int)hcpu;
+	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
+
+	if ((action & ~CPU_TASKS_FROZEN) != CPU_STARTING)
+		return NOTIFY_DONE;
+
+	memset(cpuc, 0, sizeof(struct cpu_hw_events));
+	raw_spin_lock_init(&cpuc->pmu_lock);
+
+	return NOTIFY_OK;
+}
+
+static struct notifier_block __cpuinitdata metag_pmu_notifier = {
+	.notifier_call = metag_pmu_cpu_notify,
+};
+
+/* PMU Initialisation */
+static int __init init_hw_perf_events(void)
+{
+	int ret = 0, cpu;
+	u32 version = *(u32 *)METAC_ID;
+	int major = (version & METAC_ID_MAJOR_BITS) >> METAC_ID_MAJOR_S;
+	int min_rev = (version & (METAC_ID_MINOR_BITS | METAC_ID_REV_BITS))
+			>> METAC_ID_REV_S;
+
+	/* Not a Meta 2 core, then not supported */
+	if (0x02 > major) {
+		pr_info("no hardware counter support available\n");
+		goto out;
+	} else if (0x02 == major) {
+		metag_pmu = &_metag_pmu;
+
+		if (min_rev < 0x0104) {
+			/*
+			 * A core without overflow interrupts, and clear-on-
+			 * write counters.
+			 */
+			metag_pmu->handle_irq = NULL;
+			metag_pmu->write = NULL;
+			metag_pmu->max_period = 0;
+		}
+
+		metag_pmu->name = "Meta 2";
+		metag_pmu->version = version;
+		metag_pmu->pmu = pmu;
+	}
+
+	pr_info("enabled with %s PMU driver, %d counters available\n",
+			metag_pmu->name, metag_pmu->max_events);
+
+	/* Initialise the active events and reservation mutex */
+	atomic_set(&metag_pmu->active_events, 0);
+	mutex_init(&metag_pmu->reserve_mutex);
+
+	/* Clear the counters */
+	metag_out32(0, PERF_COUNT(0));
+	metag_out32(0, PERF_COUNT(1));
+
+	for_each_possible_cpu(cpu) {
+		struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
+
+		memset(cpuc, 0, sizeof(struct cpu_hw_events));
+		raw_spin_lock_init(&cpuc->pmu_lock);
+	}
+
+	register_cpu_notifier(&metag_pmu_notifier);
+	ret = perf_pmu_register(&pmu, (char *)metag_pmu->name, PERF_TYPE_RAW);
+out:
+	return ret;
+}
+early_initcall(init_hw_perf_events);

arch/metag/kernel/perf/perf_event.h

+/*
+ * Meta performance counter support.
+ *  Copyright (C) 2012 Imagination Technologies Ltd
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ */
+
+#ifndef METAG_PERF_EVENT_H_
+#define METAG_PERF_EVENT_H_
+
+#include <linux/kernel.h>
+#include <linux/interrupt.h>
+#include <linux/perf_event.h>
+
+/* For performance counter definitions */
+#include <asm/metag_mem.h>
+
+/*
+ * The Meta core has two performance counters, with 24-bit resolution. Newer
+ * cores generate an overflow interrupt on transition from 0xffffff to 0.
+ *
+ * Each counter consists of the counter id, hardware thread id, and the count
+ * itself; each counter can be assigned to multiple hardware threads at any
+ * one time, with the returned count being an aggregate of events. A small
+ * number of events are thread global, i.e. they count the aggregate of all
+ * threads' events, regardless of the thread selected.
+ *
+ * Newer cores can store an arbitrary 24-bit number in the counter, whereas
+ * older cores will clear the counter bits on write.
+ *
+ * We also have a pseudo-counter in the form of the thread active cycles
+ * counter (which, incidentally, is also bound to
+ */
+
+#define MAX_HWEVENTS		3
+#define MAX_PERIOD		((1UL << 24) - 1)
+#define METAG_INST_COUNTER	(MAX_HWEVENTS - 1)
+
+/**
+ * struct cpu_hw_events - a processor core's performance events
+ * @events:	an array of perf_events active for a given index.
+ * @used_mask:	a bitmap of in-use counters.
+ * @pmu_lock:	a perf counter lock
+ *
+ * This is a per-cpu/core structure that maintains a record of its
+ * performance counters' state.
+ */
+struct cpu_hw_events {
+	struct perf_event	*events[MAX_HWEVENTS];
+	unsigned long		used_mask[BITS_TO_LONGS(MAX_HWEVENTS)];
+	raw_spinlock_t		pmu_lock;
+};
+
+/**
+ * struct metag_pmu - the Meta PMU structure
+ * @pmu:		core pmu structure
+ * @name:		pmu name
+ * @version:		core version
+ * @handle_irq:		overflow interrupt handler
+ * @enable:		enable a counter
+ * @disable:		disable a counter
+ * @read:		read the value of a counter
+ * @write:		write a value to a counter
+ * @event_map:		kernel event to counter event id map
+ * @cache_events:	kernel cache counter to core cache counter map
+ * @max_period:		maximum value of the counter before overflow
+ * @max_events:		maximum number of counters available at any one time
+ * @active_events:	number of active counters
+ * @reserve_mutex:	counter reservation mutex
+ *
+ * This describes the main functionality and data used by the performance
+ * event core.
+ */
+struct metag_pmu {
+	struct pmu	pmu;
+	const char	*name;
+	u32		version;
+	irqreturn_t	(*handle_irq)(int irq_num, void *dev);
+	void		(*enable)(struct hw_perf_event *evt, int idx);
+	void		(*disable)(struct hw_perf_event *evt, int idx);
+	u64		(*read)(int idx);
+	void		(*write)(int idx, u32 val);
+	int		(*event_map)(int idx);
+	const int	(*cache_events)[PERF_COUNT_HW_CACHE_MAX]
+				[PERF_COUNT_HW_CACHE_OP_MAX]
+				[PERF_COUNT_HW_CACHE_RESULT_MAX];
+	u32		max_period;
+	int		max_events;
+	atomic_t	active_events;
+	struct mutex	reserve_mutex;
+};
+
+/* Convenience macros for accessing the perf counters */
+/* Define some convenience accessors */
+#define PERF_COUNT(x)	(PERF_COUNT0 + (sizeof(u64) * (x)))
+#define PERF_ICORE(x)	(PERF_ICORE0 + (sizeof(u64) * (x)))
+#define PERF_CHAN(x)	(PERF_CHAN0 + (sizeof(u64) * (x)))
+
+/* Cache index macros */
+#define C(x) PERF_COUNT_HW_CACHE_##x
+#define CACHE_OP_UNSUPPORTED	0xfffe
+#define CACHE_OP_NONSENSE	0xffff
+
+#endif

arch/metag/kernel/perf_callchain.c

+/*
+ * Perf callchain handling code.
+ *
+ *   Based on the ARM perf implementation.
+ */
+
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/perf_event.h>
+#include <linux/uaccess.h>
+#include <asm/ptrace.h>
+#include <asm/stacktrace.h>
+
+static bool is_valid_call(unsigned long calladdr)
+{
+	unsigned int callinsn;
+
+	/* Check the possible return address is aligned. */
+	if (!(calladdr & 0x3)) {
+		if (!get_user(callinsn, (unsigned int *)calladdr)) {
+			/* Check for CALLR or SWAP PC,D1RtP. */
+			if ((callinsn & 0xff000000) == 0xab000000 ||
+			    callinsn == 0xa3200aa0)
+				return true;
+		}
+	}
+	return false;
+}
+
+static struct metag_frame __user *
+user_backtrace(struct metag_frame __user *user_frame,
+	       struct perf_callchain_entry *entry)
+{
+	struct metag_frame frame;
+	unsigned long calladdr;
+
+	/* We cannot rely on having frame pointers in user code. */
+	while (1) {
+		/* Also check accessibility of one struct frame beyond */
+		if (!access_ok(VERIFY_READ, user_frame, sizeof(frame)))
+			return 0;
+		if (__copy_from_user_inatomic(&frame, user_frame,
+					      sizeof(frame)))
+			return 0;
+
+		--user_frame;
+
+		calladdr = frame.lr - 4;
+		if (is_valid_call(calladdr)) {
+			perf_callchain_store(entry, calladdr);
+			return user_frame;
+		}
+	}
+
+	return 0;
+}
+
+void
+perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
+{
+	unsigned long sp = regs->ctx.AX[0].U0;
+	struct metag_frame __user *frame;
+
+	frame = (struct metag_frame __user *)sp;
+
+	--frame;
+
+	while ((entry->nr < PERF_MAX_STACK_DEPTH) && frame)
+		frame = user_backtrace(frame, entry);
+}
+
+/*
+ * Gets called by walk_stackframe() for every stackframe. This will be called
+ * whist unwinding the stackframe and is like a subroutine return so we use
+ * the PC.
+ */
+static int
+callchain_trace(struct stackframe *fr,
+		void *data)
+{
+	struct perf_callchain_entry *entry = data;
+	perf_callchain_store(entry, fr->pc);
+	return 0;
+}
+
+void
+perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
+{
+	struct stackframe fr;
+
+	fr.fp = regs->ctx.AX[1].U0;
+	fr.sp = regs->ctx.AX[0].U0;
+	fr.lr = regs->ctx.DX[4].U1;
+	fr.pc = regs->ctx.CurrPC;
+	walk_stackframe(&fr, callchain_trace, entry);
+}