[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 "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)) {
		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));
}

static
void overflow_callback(struct perf_event *event, int nmi,
		       struct perf_sample_data *data,
		       struct pt_regs *regs)
{
}

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);
}

#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_ctx_field *field =
		container_of(nb, struct lttng_ctx_field, u.perf_counter.nb);
	struct perf_event **events = field->u.perf_counter.e;
	struct perf_event_attr *attr = field->u.perf_counter.attr;
	struct perf_event *pevent;

	if (!field->u.perf_counter.hp_enable)
		return NOTIFY_OK;

	switch (action) {
	case CPU_ONLINE:
	case CPU_ONLINE_FROZEN:
		pevent = perf_event_create_kernel_counter(attr,
				cpu, NULL, overflow_callback);
		if (!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 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(*field->u.perf_counter.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;

	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;
	}

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

	get_online_cpus();
	for_each_online_cpu(cpu) {
		events[cpu] = perf_event_create_kernel_counter(attr,
					cpu, NULL, overflow_callback);
		if (!events[cpu]) {
			ret = -EINVAL;
			goto counter_error;
		}
	}
	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.e = events;
	field->u.perf_counter.attr = attr;
	field->u.perf_counter.hp_enable = 1;

	wrapper_vmalloc_sync_all();
	return 0;

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