2 * lttng-context-perf-counters.c
4 * LTTng UST performance monitoring counters (perf-counters) integration.
6 * Copyright (C) 2009-2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; only
11 * version 2.1 of the License.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #include <sys/types.h>
29 #include <sys/syscall.h>
30 #include <linux/perf_event.h>
31 #include <lttng/ust-events.h>
32 #include <lttng/ust-tracer.h>
33 #include <lttng/ringbuffer-config.h>
34 #include <urcu/system.h>
35 #include <urcu/arch.h>
36 #include <urcu/rculist.h>
39 #include <usterr-signal-safe.h>
41 #include "lttng-tracer-core.h"
44 * We use a global perf counter key and iterate on per-thread RCU lists
45 * of fields in the fast path, even though this is not strictly speaking
46 * what would provide the best fast-path complexity, to ensure teardown
47 * of sessions vs thread exit is handled racelessly.
49 * Updates and traversals of thread_list are protected by UST lock.
50 * Updates to rcu_field_list are protected by UST lock.
53 struct lttng_perf_counter_thread_field
{
54 struct lttng_perf_counter_field
*field
; /* Back reference */
55 struct perf_event_mmap_page
*pc
;
56 struct cds_list_head thread_field_node
; /* Per-field list of thread fields (node) */
57 struct cds_list_head rcu_field_node
; /* RCU per-thread list of fields (node) */
61 struct lttng_perf_counter_thread
{
62 struct cds_list_head rcu_field_list
; /* RCU per-thread list of fields */
65 struct lttng_perf_counter_field
{
66 struct perf_event_attr attr
;
67 struct cds_list_head thread_field_list
; /* Per-field list of thread fields */
70 static pthread_key_t perf_counter_key
;
73 size_t perf_counter_get_size(struct lttng_ctx_field
*field
, size_t offset
)
77 size
+= lib_ring_buffer_align(offset
, lttng_alignof(uint64_t));
78 size
+= sizeof(uint64_t);
82 #if defined(__x86_64__) || defined(__i386__)
85 uint64_t rdpmc(unsigned int counter
)
87 unsigned int low
, high
;
89 asm volatile("rdpmc" : "=a" (low
), "=d" (high
) : "c" (counter
));
91 return low
| ((uint64_t) high
) << 32;
94 static bool arch_perf_use_read(void)
100 uint64_t read_perf_counter(
101 struct lttng_perf_counter_thread_field
*thread_field
)
105 struct perf_event_mmap_page
*pc
= thread_field
->pc
;
107 if (caa_unlikely(!pc
))
111 seq
= CMM_LOAD_SHARED(pc
->lock
);
116 count
= pc
->offset
+ rdpmc(idx
- 1);
121 } while (CMM_LOAD_SHARED(pc
->lock
) != seq
);
126 #elif defined (__ARM_ARCH_7A__)
128 static bool arch_perf_use_read(void)
134 uint64_t read_perf_counter(
135 struct lttng_perf_counter_thread_field
*thread_field
)
139 if (caa_unlikely(thread_field
->fd
< 0))
142 if (caa_unlikely(read(thread_field
->fd
, &count
, sizeof(count
))
149 #else /* defined(__x86_64__) || defined(__i386__) || defined(__ARM_ARCH_7A__) */
151 #error "Perf event counters are only supported on x86 and ARMv7 so far."
153 #endif /* #else defined(__x86_64__) || defined(__i386__) || defined(__ARM_ARCH_7A__) */
156 int sys_perf_event_open(struct perf_event_attr
*attr
,
157 pid_t pid
, int cpu
, int group_fd
,
160 return syscall(SYS_perf_event_open
, attr
, pid
, cpu
,
165 int open_perf_fd(struct perf_event_attr
*attr
)
169 fd
= sys_perf_event_open(attr
, 0, -1, -1, 0);
177 void close_perf_fd(int fd
)
186 perror("Error closing LTTng-UST perf memory mapping FD");
191 struct perf_event_mmap_page
*setup_perf(
192 struct lttng_perf_counter_thread_field
*thread_field
)
196 perf_addr
= mmap(NULL
, sizeof(struct perf_event_mmap_page
),
197 PROT_READ
, MAP_SHARED
, thread_field
->fd
, 0);
198 if (perf_addr
== MAP_FAILED
)
201 if (!arch_perf_use_read()) {
202 close_perf_fd(thread_field
->fd
);
203 thread_field
->fd
= -1;
210 void unmap_perf_page(struct perf_event_mmap_page
*pc
)
216 ret
= munmap(pc
, sizeof(struct perf_event_mmap_page
));
218 PERROR("Error in munmap");
224 struct lttng_perf_counter_thread
*alloc_perf_counter_thread(void)
226 struct lttng_perf_counter_thread
*perf_thread
;
227 sigset_t newmask
, oldmask
;
230 ret
= sigfillset(&newmask
);
233 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
236 /* Check again with signals disabled */
237 perf_thread
= pthread_getspecific(perf_counter_key
);
240 perf_thread
= zmalloc(sizeof(*perf_thread
));
243 CDS_INIT_LIST_HEAD(&perf_thread
->rcu_field_list
);
244 ret
= pthread_setspecific(perf_counter_key
, perf_thread
);
248 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
255 struct lttng_perf_counter_thread_field
*
256 add_thread_field(struct lttng_perf_counter_field
*perf_field
,
257 struct lttng_perf_counter_thread
*perf_thread
)
259 struct lttng_perf_counter_thread_field
*thread_field
;
260 sigset_t newmask
, oldmask
;
263 ret
= sigfillset(&newmask
);
266 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
269 /* Check again with signals disabled */
270 cds_list_for_each_entry_rcu(thread_field
, &perf_thread
->rcu_field_list
,
272 if (thread_field
->field
== perf_field
)
275 thread_field
= zmalloc(sizeof(*thread_field
));
278 thread_field
->field
= perf_field
;
279 thread_field
->fd
= open_perf_fd(&perf_field
->attr
);
280 if (thread_field
->fd
>= 0)
281 thread_field
->pc
= setup_perf(thread_field
);
283 * Note: thread_field->pc can be NULL if setup_perf() fails.
284 * Also, thread_field->fd can be -1 if open_perf_fd() fails.
287 cds_list_add_rcu(&thread_field
->rcu_field_node
,
288 &perf_thread
->rcu_field_list
);
289 cds_list_add(&thread_field
->thread_field_node
,
290 &perf_field
->thread_field_list
);
293 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
300 struct lttng_perf_counter_thread_field
*
301 get_thread_field(struct lttng_perf_counter_field
*field
)
303 struct lttng_perf_counter_thread
*perf_thread
;
304 struct lttng_perf_counter_thread_field
*thread_field
;
306 perf_thread
= pthread_getspecific(perf_counter_key
);
308 perf_thread
= alloc_perf_counter_thread();
309 cds_list_for_each_entry_rcu(thread_field
, &perf_thread
->rcu_field_list
,
311 if (thread_field
->field
== field
)
314 /* perf_counter_thread_field not found, need to add one */
315 return add_thread_field(field
, perf_thread
);
319 uint64_t wrapper_perf_counter_read(struct lttng_ctx_field
*field
)
321 struct lttng_perf_counter_field
*perf_field
;
322 struct lttng_perf_counter_thread_field
*perf_thread_field
;
324 perf_field
= field
->u
.perf_counter
;
325 perf_thread_field
= get_thread_field(perf_field
);
326 return read_perf_counter(perf_thread_field
);
330 void perf_counter_record(struct lttng_ctx_field
*field
,
331 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
332 struct lttng_channel
*chan
)
336 value
= wrapper_perf_counter_read(field
);
337 lib_ring_buffer_align_ctx(ctx
, lttng_alignof(value
));
338 chan
->ops
->event_write(ctx
, &value
, sizeof(value
));
342 void perf_counter_get_value(struct lttng_ctx_field
*field
,
343 struct lttng_ctx_value
*value
)
347 v
= wrapper_perf_counter_read(field
);
351 /* Called with UST lock held */
353 void lttng_destroy_perf_thread_field(
354 struct lttng_perf_counter_thread_field
*thread_field
)
356 close_perf_fd(thread_field
->fd
);
357 unmap_perf_page(thread_field
->pc
);
358 cds_list_del_rcu(&thread_field
->rcu_field_node
);
359 cds_list_del(&thread_field
->thread_field_node
);
364 void lttng_destroy_perf_thread_key(void *_key
)
366 struct lttng_perf_counter_thread
*perf_thread
= _key
;
367 struct lttng_perf_counter_thread_field
*pos
, *p
;
370 cds_list_for_each_entry_safe(pos
, p
, &perf_thread
->rcu_field_list
,
372 lttng_destroy_perf_thread_field(pos
);
377 /* Called with UST lock held */
379 void lttng_destroy_perf_counter_field(struct lttng_ctx_field
*field
)
381 struct lttng_perf_counter_field
*perf_field
;
382 struct lttng_perf_counter_thread_field
*pos
, *p
;
384 free((char *) field
->event_field
.name
);
385 perf_field
= field
->u
.perf_counter
;
387 * This put is performed when no threads can concurrently
388 * perform a "get" concurrently, thanks to urcu-bp grace
391 cds_list_for_each_entry_safe(pos
, p
, &perf_field
->thread_field_list
,
393 lttng_destroy_perf_thread_field(pos
);
397 #ifdef __ARM_ARCH_7A__
400 int perf_get_exclude_kernel(void)
405 #else /* __ARM_ARCH_7A__ */
408 int perf_get_exclude_kernel(void)
413 #endif /* __ARM_ARCH_7A__ */
415 /* Called with UST lock held */
416 int lttng_add_perf_counter_to_ctx(uint32_t type
,
419 struct lttng_ctx
**ctx
)
421 struct lttng_ctx_field
*field
;
422 struct lttng_perf_counter_field
*perf_field
;
426 name_alloc
= strdup(name
);
429 goto name_alloc_error
;
431 perf_field
= zmalloc(sizeof(*perf_field
));
434 goto perf_field_alloc_error
;
436 field
= lttng_append_context(ctx
);
439 goto append_context_error
;
441 if (lttng_find_context(*ctx
, name_alloc
)) {
446 field
->destroy
= lttng_destroy_perf_counter_field
;
448 field
->event_field
.name
= name_alloc
;
449 field
->event_field
.type
.atype
= atype_integer
;
450 field
->event_field
.type
.u
.basic
.integer
.size
=
451 sizeof(uint64_t) * CHAR_BIT
;
452 field
->event_field
.type
.u
.basic
.integer
.alignment
=
453 lttng_alignof(uint64_t) * CHAR_BIT
;
454 field
->event_field
.type
.u
.basic
.integer
.signedness
=
455 lttng_is_signed_type(uint64_t);
456 field
->event_field
.type
.u
.basic
.integer
.reverse_byte_order
= 0;
457 field
->event_field
.type
.u
.basic
.integer
.base
= 10;
458 field
->event_field
.type
.u
.basic
.integer
.encoding
= lttng_encode_none
;
459 field
->get_size
= perf_counter_get_size
;
460 field
->record
= perf_counter_record
;
461 field
->get_value
= perf_counter_get_value
;
463 perf_field
->attr
.type
= type
;
464 perf_field
->attr
.config
= config
;
465 perf_field
->attr
.exclude_kernel
= perf_get_exclude_kernel();
466 CDS_INIT_LIST_HEAD(&perf_field
->thread_field_list
);
467 field
->u
.perf_counter
= perf_field
;
469 /* Ensure that this perf counter can be used in this process. */
470 ret
= open_perf_fd(&perf_field
->attr
);
478 * Contexts can only be added before tracing is started, so we
479 * don't have to synchronize against concurrent threads using
483 lttng_context_update(*ctx
);
488 lttng_remove_context_field(ctx
, field
);
489 append_context_error
:
491 perf_field_alloc_error
:
497 int lttng_perf_counter_init(void)
501 ret
= pthread_key_create(&perf_counter_key
,
502 lttng_destroy_perf_thread_key
);
508 void lttng_perf_counter_exit(void)
512 ret
= pthread_key_delete(perf_counter_key
);
515 PERROR("Error in pthread_key_delete");