[deliverable/lttng-modules.git] / lttng-context-perf-counters.c

/*
 * (C) Copyright	2009-2011 -
 * 		Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 *
 * LTTng performance monitoring counters (perf-counters) integration module.
 *
 * Dual LGPL v2.1/GPL v2 license.
 */

#include <linux/module.h>
#include <linux/slab.h>
#include <linux/perf_event.h>
#include <linux/list.h>
#include <linux/string.h>
#include "ltt-events.h"
#include "wrapper/ringbuffer/frontend_types.h"
#include "wrapper/vmalloc.h"
#include "wrapper/perf.h"
#include "ltt-tracer.h"

static
size_t perf_counter_get_size(size_t offset)
{
	size_t size = 0;

	size += lib_ring_buffer_align(offset, ltt_alignof(uint64_t));
	size += sizeof(uint64_t);
	return size;
}

static
void perf_counter_record(struct lttng_ctx_field *field,
			 struct lib_ring_buffer_ctx *ctx,
			 struct ltt_channel *chan)
{
	struct perf_event *event;
	uint64_t value;

	event = field->u.perf_counter->e[ctx->cpu];
	if (likely(event)) {
		if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) {
			value = 0;
		} else {
			event->pmu->read(event);
			value = local64_read(&event->count);
		}
	} else {
		/*
		 * Perf chooses not to be clever and not to support enabling a
		 * perf counter before the cpu is brought up. Therefore, we need
		 * to support having events coming (e.g. scheduler events)
		 * before the counter is setup. Write an arbitrary 0 in this
		 * case.
		 */
		value = 0;
	}
	lib_ring_buffer_align_ctx(ctx, ltt_alignof(value));
	chan->ops->event_write(ctx, &value, sizeof(value));
}

#if defined(CONFIG_PERF_EVENTS) && (LINUX_VERSION_CODE >= KERNEL_VERSION(3,0,99))
static
void overflow_callback(struct perf_event *event,
		       struct perf_sample_data *data,
		       struct pt_regs *regs)
{
}
#else
static
void overflow_callback(struct perf_event *event, int nmi,
		       struct perf_sample_data *data,
		       struct pt_regs *regs)
{
}
#endif

static
void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
{
	struct perf_event **events = field->u.perf_counter->e;
	int cpu;

	get_online_cpus();
	for_each_online_cpu(cpu)
		perf_event_release_kernel(events[cpu]);
	put_online_cpus();
#ifdef CONFIG_HOTPLUG_CPU
	unregister_cpu_notifier(&field->u.perf_counter->nb);
#endif
	kfree(field->event_field.name);
	kfree(field->u.perf_counter->attr);
	kfree(events);
	kfree(field->u.perf_counter);
}

#ifdef CONFIG_HOTPLUG_CPU

/**
 *	lttng_perf_counter_hp_callback - CPU hotplug callback
 *	@nb: notifier block
 *	@action: hotplug action to take
 *	@hcpu: CPU number
 *
 *	Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
 *
 * We can setup perf counters when the cpu is online (up prepare seems to be too
 * soon).
 */
static
int __cpuinit lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb,
						 unsigned long action,
						 void *hcpu)
{
	unsigned int cpu = (unsigned long) hcpu;
	struct lttng_perf_counter_field *perf_field =
		container_of(nb, struct lttng_perf_counter_field, nb);
	struct perf_event **events = perf_field->e;
	struct perf_event_attr *attr = perf_field->attr;
	struct perf_event *pevent;

	if (!perf_field->hp_enable)
		return NOTIFY_OK;

	switch (action) {
	case CPU_ONLINE:
	case CPU_ONLINE_FROZEN:
		pevent = wrapper_perf_event_create_kernel_counter(attr,
				cpu, NULL, overflow_callback);
		if (!pevent || IS_ERR(pevent))
			return NOTIFY_BAD;
		if (pevent->state == PERF_EVENT_STATE_ERROR) {
			perf_event_release_kernel(pevent);
			return NOTIFY_BAD;
		}
		barrier();	/* Create perf counter before setting event */
		events[cpu] = pevent;
		break;
	case CPU_UP_CANCELED:
	case CPU_UP_CANCELED_FROZEN:
	case CPU_DEAD:
	case CPU_DEAD_FROZEN:
		pevent = events[cpu];
		events[cpu] = NULL;
		barrier();	/* NULLify event before perf counter teardown */
		perf_event_release_kernel(pevent);
		break;
	}
	return NOTIFY_OK;
}

#endif

int lttng_add_perf_counter_to_ctx(uint32_t type,
				  uint64_t config,
				  const char *name,
				  struct lttng_ctx **ctx)
{
	struct lttng_ctx_field *field;
	struct lttng_perf_counter_field *perf_field;
	struct perf_event **events;
	struct perf_event_attr *attr;
	int ret;
	int cpu;
	char *name_alloc;

	events = kzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL);
	if (!events)
		return -ENOMEM;

	attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL);
	if (!attr) {
		ret = -ENOMEM;
		goto error_attr;
	}

	attr->type = type;
	attr->config = config;
	attr->size = sizeof(struct perf_event_attr);
	attr->pinned = 1;
	attr->disabled = 0;

	perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL);
	if (!perf_field) {
		ret = -ENOMEM;
		goto error_alloc_perf_field;
	}
	perf_field->e = events;
	perf_field->attr = attr;

	name_alloc = kstrdup(name, GFP_KERNEL);
	if (!name_alloc) {
		ret = -ENOMEM;
		goto name_alloc_error;
	}

	field = lttng_append_context(ctx);
	if (!field) {
		ret = -ENOMEM;
		goto append_context_error;
	}
	if (lttng_find_context(*ctx, name_alloc)) {
		ret = -EEXIST;
		goto find_error;
	}

#ifdef CONFIG_HOTPLUG_CPU
	perf_field->nb.notifier_call =
		lttng_perf_counter_cpu_hp_callback;
	perf_field->nb.priority = 0;
	register_cpu_notifier(&perf_field->nb);
#endif

	get_online_cpus();
	for_each_online_cpu(cpu) {
		events[cpu] = wrapper_perf_event_create_kernel_counter(attr,
					cpu, NULL, overflow_callback);
		if (!events[cpu] || IS_ERR(events[cpu])) {
			ret = -EINVAL;
			goto counter_error;
		}
		if (events[cpu]->state == PERF_EVENT_STATE_ERROR) {
			ret = -EBUSY;
			goto counter_busy;
		}
	}
	put_online_cpus();

	field->destroy = lttng_destroy_perf_counter_field;

	field->event_field.name = name_alloc;
	field->event_field.type.atype = atype_integer;
	field->event_field.type.u.basic.integer.size = sizeof(unsigned long) * CHAR_BIT;
	field->event_field.type.u.basic.integer.alignment = ltt_alignof(unsigned long) * CHAR_BIT;
	field->event_field.type.u.basic.integer.signedness = is_signed_type(unsigned long);
	field->event_field.type.u.basic.integer.reverse_byte_order = 0;
	field->event_field.type.u.basic.integer.base = 10;
	field->event_field.type.u.basic.integer.encoding = lttng_encode_none;
	field->get_size = perf_counter_get_size;
	field->record = perf_counter_record;
	field->u.perf_counter = perf_field;
	perf_field->hp_enable = 1;

	wrapper_vmalloc_sync_all();
	return 0;

counter_busy:
counter_error:
	for_each_online_cpu(cpu) {
		if (events[cpu] && !IS_ERR(events[cpu]))
			perf_event_release_kernel(events[cpu]);
	}
	put_online_cpus();
#ifdef CONFIG_HOTPLUG_CPU
	unregister_cpu_notifier(&perf_field->nb);
#endif
find_error:
	lttng_remove_context_field(ctx, field);
append_context_error:
	kfree(name_alloc);
name_alloc_error:
	kfree(perf_field);
error_alloc_perf_field:
	kfree(attr);
error_attr:
	kfree(events);
	return ret;
}

MODULE_LICENSE("GPL and additional rights");
MODULE_AUTHOR("Mathieu Desnoyers");
MODULE_DESCRIPTION("Linux Trace Toolkit Perf Support");
Commit	Line	Data
833ad6a0 MD	1	/*
	2	* (C) Copyright 2009-2011 -
	3	* Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	4	*
	5	* LTTng performance monitoring counters (perf-counters) integration module.
	6	*
	7	* Dual LGPL v2.1/GPL v2 license.
	8	*/
	9
	10	#include <linux/module.h>
	11	#include <linux/slab.h>
	12	#include <linux/perf_event.h>
	13	#include <linux/list.h>
c24a0d71 MD	14	#include <linux/string.h>
	15	#include "ltt-events.h"
	16	#include "wrapper/ringbuffer/frontend_types.h"
	17	#include "wrapper/vmalloc.h"
90f5546c	18	#include "wrapper/perf.h"
c24a0d71	19	#include "ltt-tracer.h"
833ad6a0	20
f1676205 MD	21	static
	22	size_t perf_counter_get_size(size_t offset)
	23	{
	24	size_t size = 0;
	25
	26	size += lib_ring_buffer_align(offset, ltt_alignof(uint64_t));
	27	size += sizeof(uint64_t);
	28	return size;
	29	}
	30
833ad6a0 MD	31	static
	32	void perf_counter_record(struct lttng_ctx_field *field,
	33	struct lib_ring_buffer_ctx *ctx,
	34	struct ltt_channel *chan)
	35	{
	36	struct perf_event *event;
	37	uint64_t value;
	38
2001023e	39	event = field->u.perf_counter->e[ctx->cpu];
0478c519	40	if (likely(event)) {
7b745a96 MD	41	if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) {
	42	value = 0;
	43	} else {
	44	event->pmu->read(event);
	45	value = local64_read(&event->count);
	46	}
f91fd73b MD	47	} else {
	48	/*
	49	* Perf chooses not to be clever and not to support enabling a
	50	* perf counter before the cpu is brought up. Therefore, we need
	51	* to support having events coming (e.g. scheduler events)
	52	* before the counter is setup. Write an arbitrary 0 in this
	53	* case.
	54	*/
	55	value = 0;
	56	}
9e7e4892	57	lib_ring_buffer_align_ctx(ctx, ltt_alignof(value));
833ad6a0 MD	58	chan->ops->event_write(ctx, &value, sizeof(value));
	59	}
	60
90f5546c MD	61	#if defined(CONFIG_PERF_EVENTS) && (LINUX_VERSION_CODE >= KERNEL_VERSION(3,0,99))
	62	static
	63	void overflow_callback(struct perf_event *event,
	64	struct perf_sample_data *data,
	65	struct pt_regs *regs)
	66	{
	67	}
	68	#else
833ad6a0 MD	69	static
	70	void overflow_callback(struct perf_event *event, int nmi,
	71	struct perf_sample_data *data,
	72	struct pt_regs *regs)
	73	{
	74	}
90f5546c	75	#endif
833ad6a0	76
2dccf128 MD	77	static
	78	void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
	79	{
2001023e	80	struct perf_event **events = field->u.perf_counter->e;
2dccf128 MD	81	int cpu;
2dccf128 MD	82
8289661d	83	get_online_cpus();
2dccf128 MD	84	for_each_online_cpu(cpu)
2dccf128 MD	85	perf_event_release_kernel(events[cpu]);
8289661d MD	86	put_online_cpus();
8289661d MD	87	#ifdef CONFIG_HOTPLUG_CPU
2001023e	88	unregister_cpu_notifier(&field->u.perf_counter->nb);
8289661d	89	#endif
c24a0d71	90	kfree(field->event_field.name);
2001023e	91	kfree(field->u.perf_counter->attr);
2dccf128	92	kfree(events);
2001023e	93	kfree(field->u.perf_counter);
2dccf128 MD	94	}
2dccf128 MD	95
8289661d MD	96	#ifdef CONFIG_HOTPLUG_CPU
	97
	98	/**
	99	* lttng_perf_counter_hp_callback - CPU hotplug callback
	100	* @nb: notifier block
	101	* @action: hotplug action to take
	102	* @hcpu: CPU number
	103	*
	104	* Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
	105	*
	106	* We can setup perf counters when the cpu is online (up prepare seems to be too
	107	* soon).
	108	*/
	109	static
	110	int __cpuinit lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb,
	111	unsigned long action,
	112	void *hcpu)
	113	{
	114	unsigned int cpu = (unsigned long) hcpu;
2001023e MD	115	struct lttng_perf_counter_field *perf_field =
	116	container_of(nb, struct lttng_perf_counter_field, nb);
	117	struct perf_event **events = perf_field->e;
	118	struct perf_event_attr *attr = perf_field->attr;
f91fd73b	119	struct perf_event *pevent;
8289661d	120
2001023e	121	if (!perf_field->hp_enable)
8289661d MD	122	return NOTIFY_OK;
	123
	124	switch (action) {
	125	case CPU_ONLINE:
	126	case CPU_ONLINE_FROZEN:
90f5546c	127	pevent = wrapper_perf_event_create_kernel_counter(attr,
8289661d	128	cpu, NULL, overflow_callback);
0478c519	129	if (!pevent \|\| IS_ERR(pevent))
8289661d	130	return NOTIFY_BAD;
7b745a96 MD	131	if (pevent->state == PERF_EVENT_STATE_ERROR) {
	132	perf_event_release_kernel(pevent);
	133	return NOTIFY_BAD;
	134	}
f91fd73b MD	135	barrier(); /* Create perf counter before setting event */
f91fd73b MD	136	events[cpu] = pevent;
8289661d MD	137	break;
	138	case CPU_UP_CANCELED:
	139	case CPU_UP_CANCELED_FROZEN:
	140	case CPU_DEAD:
	141	case CPU_DEAD_FROZEN:
f91fd73b MD	142	pevent = events[cpu];
	143	events[cpu] = NULL;
	144	barrier(); /* NULLify event before perf counter teardown */
	145	perf_event_release_kernel(pevent);
8289661d MD	146	break;
	147	}
	148	return NOTIFY_OK;
	149	}
	150
	151	#endif
	152
833ad6a0 MD	153	int lttng_add_perf_counter_to_ctx(uint32_t type,
833ad6a0 MD	154	uint64_t config,
c24a0d71	155	const char *name,
2dccf128	156	struct lttng_ctx **ctx)
833ad6a0 MD	157	{
833ad6a0 MD	158	struct lttng_ctx_field *field;
2001023e	159	struct lttng_perf_counter_field *perf_field;
833ad6a0 MD	160	struct perf_event **events;
	161	struct perf_event_attr *attr;
	162	int ret;
	163	int cpu;
c24a0d71	164	char *name_alloc;
833ad6a0 MD	165
	166	events = kzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL);
	167	if (!events)
	168	return -ENOMEM;
	169
2001023e	170	attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL);
833ad6a0 MD	171	if (!attr) {
	172	ret = -ENOMEM;
	173	goto error_attr;
	174	}
	175
	176	attr->type = type;
	177	attr->config = config;
	178	attr->size = sizeof(struct perf_event_attr);
	179	attr->pinned = 1;
	180	attr->disabled = 0;
	181
2001023e MD	182	perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL);
	183	if (!perf_field) {
	184	ret = -ENOMEM;
	185	goto error_alloc_perf_field;
	186	}
	187	perf_field->e = events;
	188	perf_field->attr = attr;
	189
c24a0d71	190	name_alloc = kstrdup(name, GFP_KERNEL);
bef96e48 MD	191	if (!name_alloc) {
bef96e48 MD	192	ret = -ENOMEM;
c24a0d71	193	goto name_alloc_error;
bef96e48	194	}
c24a0d71	195
2dccf128 MD	196	field = lttng_append_context(ctx);
	197	if (!field) {
	198	ret = -ENOMEM;
8289661d	199	goto append_context_error;
833ad6a0	200	}
2001023e	201	if (lttng_find_context(*ctx, name_alloc)) {
44252f0f MD	202	ret = -EEXIST;
	203	goto find_error;
	204	}
8289661d MD	205
8289661d MD	206	#ifdef CONFIG_HOTPLUG_CPU
2001023e	207	perf_field->nb.notifier_call =
8289661d	208	lttng_perf_counter_cpu_hp_callback;
2001023e MD	209	perf_field->nb.priority = 0;
2001023e MD	210	register_cpu_notifier(&perf_field->nb);
8289661d MD	211	#endif
	212
	213	get_online_cpus();
	214	for_each_online_cpu(cpu) {
90f5546c	215	events[cpu] = wrapper_perf_event_create_kernel_counter(attr,
8289661d	216	cpu, NULL, overflow_callback);
0478c519	217	if (!events[cpu] \|\| IS_ERR(events[cpu])) {
8289661d MD	218	ret = -EINVAL;
	219	goto counter_error;
	220	}
7b745a96 MD	221	if (events[cpu]->state == PERF_EVENT_STATE_ERROR) {
	222	ret = -EBUSY;
	223	goto counter_busy;
	224	}
8289661d MD	225	}
	226	put_online_cpus();
	227
2dccf128	228	field->destroy = lttng_destroy_perf_counter_field;
833ad6a0	229
c24a0d71	230	field->event_field.name = name_alloc;
8070f5c0 MD	231	field->event_field.type.atype = atype_integer;
	232	field->event_field.type.u.basic.integer.size = sizeof(unsigned long) * CHAR_BIT;
	233	field->event_field.type.u.basic.integer.alignment = ltt_alignof(unsigned long) * CHAR_BIT;
	234	field->event_field.type.u.basic.integer.signedness = is_signed_type(unsigned long);
	235	field->event_field.type.u.basic.integer.reverse_byte_order = 0;
	236	field->event_field.type.u.basic.integer.base = 10;
	237	field->event_field.type.u.basic.integer.encoding = lttng_encode_none;
f1676205 MD	238	field->get_size = perf_counter_get_size;
f1676205 MD	239	field->record = perf_counter_record;
2001023e MD	240	field->u.perf_counter = perf_field;
2001023e MD	241	perf_field->hp_enable = 1;
833ad6a0 MD	242
	243	wrapper_vmalloc_sync_all();
	244	return 0;
	245
7b745a96	246	counter_busy:
8289661d	247	counter_error:
833ad6a0	248	for_each_online_cpu(cpu) {
0478c519	249	if (events[cpu] && !IS_ERR(events[cpu]))
833ad6a0 MD	250	perf_event_release_kernel(events[cpu]);
833ad6a0 MD	251	}
8289661d MD	252	put_online_cpus();
8289661d MD	253	#ifdef CONFIG_HOTPLUG_CPU
2001023e	254	unregister_cpu_notifier(&perf_field->nb);
8289661d	255	#endif
44252f0f	256	find_error:
8289661d MD	257	lttng_remove_context_field(ctx, field);
	258	append_context_error:
	259	kfree(name_alloc);
	260	name_alloc_error:
2001023e MD	261	kfree(perf_field);
2001023e MD	262	error_alloc_perf_field:
833ad6a0 MD	263	kfree(attr);
	264	error_attr:
	265	kfree(events);
	266	return ret;
	267	}
	268
833ad6a0 MD	269	MODULE_LICENSE("GPL and additional rights");
	270	MODULE_AUTHOR("Mathieu Desnoyers");
	271	MODULE_DESCRIPTION("Linux Trace Toolkit Perf Support");