liblttng-ust/lttng-context-perf-counters.c

   1 /*
   2  * lttng-context-perf-counters.c
   3  *
   4  * LTTng UST performance monitoring counters (perf-counters) integration.
   5  *
   6  * Copyright (C) 2009-2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  *
   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.
  12  *
  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.
  17  *
  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
  21  */
  22
  23 #define _LGPL_SOURCE
  24 #include <sys/types.h>
  25 #include <unistd.h>
  26 #include <string.h>
  27 #include <stdlib.h>
  28 #include <stdio.h>
  29 #include <stdbool.h>
  30 #include <sys/mman.h>
  31 #include <sys/syscall.h>
  32 #include <lttng/ust-events.h>
  33 #include <lttng/ust-tracer.h>
  34 #include <lttng/ringbuffer-config.h>
  35 #include <urcu/system.h>
  36 #include <urcu/arch.h>
  37 #include <urcu/rculist.h>
  38 #include <helper.h>
  39 #include <urcu/ref.h>
  40 #include <usterr-signal-safe.h>
  41 #include <signal.h>
  42 #include <urcu/tls-compat.h>
  43 #include "perf_event.h"
  44 #include "lttng-tracer-core.h"
  45
  46 /*
  47  * We use a global perf counter key and iterate on per-thread RCU lists
  48  * of fields in the fast path, even though this is not strictly speaking
  49  * what would provide the best fast-path complexity, to ensure teardown
  50  * of sessions vs thread exit is handled racelessly.
  51  *
  52  * Updates and traversals of thread_list are protected by UST lock.
  53  * Updates to rcu_field_list are protected by UST lock.
  54  */
  55
  56 struct lttng_perf_counter_thread_field {
  57         struct lttng_perf_counter_field *field; /* Back reference */
  58         struct perf_event_mmap_page *pc;
  59         struct cds_list_head thread_field_node; /* Per-field list of thread fields (node) */
  60         struct cds_list_head rcu_field_node;    /* RCU per-thread list of fields (node) */
  61         int fd;                                 /* Perf FD */
  62 };
  63
  64 struct lttng_perf_counter_thread {
  65         struct cds_list_head rcu_field_list;    /* RCU per-thread list of fields */
  66 };
  67
  68 struct lttng_perf_counter_field {
  69         struct perf_event_attr attr;
  70         struct cds_list_head thread_field_list; /* Per-field list of thread fields */
  71 };
  72
  73 static pthread_key_t perf_counter_key;
  74
  75 /*
  76  * lttng_perf_lock - Protect lttng-ust perf counter data structures
  77  *
  78  * Nests within the ust_lock, and therefore within the libc dl lock.
  79  * Therefore, we need to fixup the TLS before nesting into this lock.
  80  * Nests inside RCU bp read-side lock. Protects against concurrent
  81  * fork.
  82  */
  83 static pthread_mutex_t ust_perf_mutex = PTHREAD_MUTEX_INITIALIZER;
  84
  85 /*
  86  * Cancel state when grabbing the ust_perf_mutex. Saved when locking,
  87  * restored on unlock. Protected by ust_perf_mutex.
  88  */
  89 static int ust_perf_saved_cancelstate;
  90
  91 /*
  92  * Track whether we are tracing from a signal handler nested on an
  93  * application thread.
  94  */
  95 static DEFINE_URCU_TLS(int, ust_perf_mutex_nest);
  96
  97 /*
  98  * Force a read (imply TLS fixup for dlopen) of TLS variables.
  99  */
 100 void lttng_ust_fixup_perf_counter_tls(void)
 101 {
 102         asm volatile ("" : : "m" (URCU_TLS(ust_perf_mutex_nest)));
 103 }
 104
 105 void lttng_perf_lock(void)
 106 {
 107         sigset_t sig_all_blocked, orig_mask;
 108         int ret, oldstate;
 109
 110         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 111         if (ret) {
 112                 ERR("pthread_setcancelstate: %s", strerror(ret));
 113         }
 114         sigfillset(&sig_all_blocked);
 115         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 116         if (ret) {
 117                 ERR("pthread_sigmask: %s", strerror(ret));
 118         }
 119         if (!URCU_TLS(ust_perf_mutex_nest)++) {
 120                 /*
 121                  * Ensure the compiler don't move the store after the close()
 122                  * call in case close() would be marked as leaf.
 123                  */
 124                 cmm_barrier();
 125                 pthread_mutex_lock(&ust_perf_mutex);
 126                 ust_perf_saved_cancelstate = oldstate;
 127         }
 128         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 129         if (ret) {
 130                 ERR("pthread_sigmask: %s", strerror(ret));
 131         }
 132 }
 133
 134 void lttng_perf_unlock(void)
 135 {
 136         sigset_t sig_all_blocked, orig_mask;
 137         int ret, newstate, oldstate;
 138         bool restore_cancel = false;
 139
 140         sigfillset(&sig_all_blocked);
 141         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 142         if (ret) {
 143                 ERR("pthread_sigmask: %s", strerror(ret));
 144         }
 145         /*
 146          * Ensure the compiler don't move the store before the close()
 147          * call, in case close() would be marked as leaf.
 148          */
 149         cmm_barrier();
 150         if (!--URCU_TLS(ust_perf_mutex_nest)) {
 151                 newstate = ust_perf_saved_cancelstate;
 152                 restore_cancel = true;
 153                 pthread_mutex_unlock(&ust_perf_mutex);
 154         }
 155         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 156         if (ret) {
 157                 ERR("pthread_sigmask: %s", strerror(ret));
 158         }
 159         if (restore_cancel) {
 160                 ret = pthread_setcancelstate(newstate, &oldstate);
 161                 if (ret) {
 162                         ERR("pthread_setcancelstate: %s", strerror(ret));
 163                 }
 164         }
 165 }
 166
 167 static
 168 size_t perf_counter_get_size(struct lttng_ctx_field *field, size_t offset)
 169 {
 170         size_t size = 0;
 171
 172         size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t));
 173         size += sizeof(uint64_t);
 174         return size;
 175 }
 176
 177 static
 178 uint64_t read_perf_counter_syscall(
 179                 struct lttng_perf_counter_thread_field *thread_field)
 180 {
 181         uint64_t count;
 182
 183         if (caa_unlikely(thread_field->fd < 0))
 184                 return 0;
 185
 186         if (caa_unlikely(read(thread_field->fd, &count, sizeof(count))
 187                                 < sizeof(count)))
 188                 return 0;
 189
 190         return count;
 191 }
 192
 193 #if defined(__x86_64__) || defined(__i386__)
 194
 195 static
 196 uint64_t rdpmc(unsigned int counter)
 197 {
 198         unsigned int low, high;
 199
 200         asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter));
 201
 202         return low | ((uint64_t) high) << 32;
 203 }
 204
 205 static
 206 bool has_rdpmc(struct perf_event_mmap_page *pc)
 207 {
 208         if (caa_unlikely(!pc->cap_bit0_is_deprecated))
 209                 return false;
 210         /* Since Linux kernel 3.12. */
 211         return pc->cap_user_rdpmc;
 212 }
 213
 214 static
 215 uint64_t arch_read_perf_counter(
 216                 struct lttng_perf_counter_thread_field *thread_field)
 217 {
 218         uint32_t seq, idx;
 219         uint64_t count;
 220         struct perf_event_mmap_page *pc = thread_field->pc;
 221
 222         if (caa_unlikely(!pc))
 223                 return 0;
 224
 225         do {
 226                 seq = CMM_LOAD_SHARED(pc->lock);
 227                 cmm_barrier();
 228
 229                 idx = pc->index;
 230                 if (caa_likely(has_rdpmc(pc) && idx)) {
 231                         int64_t pmcval;
 232
 233                         pmcval = rdpmc(idx - 1);
 234                         /* Sign-extend the pmc register result. */
 235                         pmcval <<= 64 - pc->pmc_width;
 236                         pmcval >>= 64 - pc->pmc_width;
 237                         count = pc->offset + pmcval;
 238                 } else {
 239                         /* Fall-back on system call if rdpmc cannot be used. */
 240                         return read_perf_counter_syscall(thread_field);
 241                 }
 242                 cmm_barrier();
 243         } while (CMM_LOAD_SHARED(pc->lock) != seq);
 244
 245         return count;
 246 }
 247
 248 static
 249 int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field)
 250 {
 251         struct perf_event_mmap_page *pc = thread_field->pc;
 252
 253         if (!pc)
 254                 return 0;
 255         return !has_rdpmc(pc);
 256 }
 257
 258 #else
 259
 260 /* Generic (slow) implementation using a read system call. */
 261 static
 262 uint64_t arch_read_perf_counter(
 263                 struct lttng_perf_counter_thread_field *thread_field)
 264 {
 265         return read_perf_counter_syscall(thread_field);
 266 }
 267
 268 static
 269 int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field)
 270 {
 271         return 1;
 272 }
 273
 274 #endif
 275
 276 static
 277 int sys_perf_event_open(struct perf_event_attr *attr,
 278                 pid_t pid, int cpu, int group_fd,
 279                 unsigned long flags)
 280 {
 281         return syscall(SYS_perf_event_open, attr, pid, cpu,
 282                         group_fd, flags);
 283 }
 284
 285 static
 286 int open_perf_fd(struct perf_event_attr *attr)
 287 {
 288         int fd;
 289
 290         fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 291         if (fd < 0)
 292                 return -1;
 293
 294         return fd;
 295 }
 296
 297 static
 298 void close_perf_fd(int fd)
 299 {
 300         int ret;
 301
 302         if (fd < 0)
 303                 return;
 304
 305         ret = close(fd);
 306         if (ret) {
 307                 perror("Error closing LTTng-UST perf memory mapping FD");
 308         }
 309 }
 310
 311 static void setup_perf(struct lttng_perf_counter_thread_field *thread_field)
 312 {
 313         void *perf_addr;
 314
 315         perf_addr = mmap(NULL, sizeof(struct perf_event_mmap_page),
 316                         PROT_READ, MAP_SHARED, thread_field->fd, 0);
 317         if (perf_addr == MAP_FAILED)
 318                 perf_addr = NULL;
 319         thread_field->pc = perf_addr;
 320
 321         if (!arch_perf_keep_fd(thread_field)) {
 322                 close_perf_fd(thread_field->fd);
 323                 thread_field->fd = -1;
 324         }
 325 }
 326
 327 static
 328 void unmap_perf_page(struct perf_event_mmap_page *pc)
 329 {
 330         int ret;
 331
 332         if (!pc)
 333                 return;
 334         ret = munmap(pc, sizeof(struct perf_event_mmap_page));
 335         if (ret < 0) {
 336                 PERROR("Error in munmap");
 337                 abort();
 338         }
 339 }
 340
 341 static
 342 struct lttng_perf_counter_thread *alloc_perf_counter_thread(void)
 343 {
 344         struct lttng_perf_counter_thread *perf_thread;
 345         sigset_t newmask, oldmask;
 346         int ret;
 347
 348         ret = sigfillset(&newmask);
 349         if (ret)
 350                 abort();
 351         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 352         if (ret)
 353                 abort();
 354         /* Check again with signals disabled */
 355         perf_thread = pthread_getspecific(perf_counter_key);
 356         if (perf_thread)
 357                 goto skip;
 358         perf_thread = zmalloc(sizeof(*perf_thread));
 359         if (!perf_thread)
 360                 abort();
 361         CDS_INIT_LIST_HEAD(&perf_thread->rcu_field_list);
 362         ret = pthread_setspecific(perf_counter_key, perf_thread);
 363         if (ret)
 364                 abort();
 365 skip:
 366         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 367         if (ret)
 368                 abort();
 369         return perf_thread;
 370 }
 371
 372 static
 373 struct lttng_perf_counter_thread_field *
 374         add_thread_field(struct lttng_perf_counter_field *perf_field,
 375                 struct lttng_perf_counter_thread *perf_thread)
 376 {
 377         struct lttng_perf_counter_thread_field *thread_field;
 378         sigset_t newmask, oldmask;
 379         int ret;
 380
 381         ret = sigfillset(&newmask);
 382         if (ret)
 383                 abort();
 384         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 385         if (ret)
 386                 abort();
 387         /* Check again with signals disabled */
 388         cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list,
 389                         rcu_field_node) {
 390                 if (thread_field->field == perf_field)
 391                         goto skip;
 392         }
 393         thread_field = zmalloc(sizeof(*thread_field));
 394         if (!thread_field)
 395                 abort();
 396         thread_field->field = perf_field;
 397         thread_field->fd = open_perf_fd(&perf_field->attr);
 398         if (thread_field->fd >= 0)
 399                 setup_perf(thread_field);
 400         /*
 401          * Note: thread_field->pc can be NULL if setup_perf() fails.
 402          * Also, thread_field->fd can be -1 if open_perf_fd() fails.
 403          */
 404         lttng_perf_lock();
 405         cds_list_add_rcu(&thread_field->rcu_field_node,
 406                         &perf_thread->rcu_field_list);
 407         cds_list_add(&thread_field->thread_field_node,
 408                         &perf_field->thread_field_list);
 409         lttng_perf_unlock();
 410 skip:
 411         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 412         if (ret)
 413                 abort();
 414         return thread_field;
 415 }
 416
 417 static
 418 struct lttng_perf_counter_thread_field *
 419                 get_thread_field(struct lttng_perf_counter_field *field)
 420 {
 421         struct lttng_perf_counter_thread *perf_thread;
 422         struct lttng_perf_counter_thread_field *thread_field;
 423
 424         perf_thread = pthread_getspecific(perf_counter_key);
 425         if (!perf_thread)
 426                 perf_thread = alloc_perf_counter_thread();
 427         cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list,
 428                         rcu_field_node) {
 429                 if (thread_field->field == field)
 430                         return thread_field;
 431         }
 432         /* perf_counter_thread_field not found, need to add one */
 433         return add_thread_field(field, perf_thread);
 434 }
 435
 436 static
 437 uint64_t wrapper_perf_counter_read(struct lttng_ctx_field *field)
 438 {
 439         struct lttng_perf_counter_field *perf_field;
 440         struct lttng_perf_counter_thread_field *perf_thread_field;
 441
 442         perf_field = field->u.perf_counter;
 443         perf_thread_field = get_thread_field(perf_field);
 444         return arch_read_perf_counter(perf_thread_field);
 445 }
 446
 447 static
 448 void perf_counter_record(struct lttng_ctx_field *field,
 449                  struct lttng_ust_lib_ring_buffer_ctx *ctx,
 450                  struct lttng_channel *chan)
 451 {
 452         uint64_t value;
 453
 454         value = wrapper_perf_counter_read(field);
 455         lib_ring_buffer_align_ctx(ctx, lttng_alignof(value));
 456         chan->ops->event_write(ctx, &value, sizeof(value));
 457 }
 458
 459 static
 460 void perf_counter_get_value(struct lttng_ctx_field *field,
 461                 struct lttng_ctx_value *value)
 462 {
 463         value->u.s64 = wrapper_perf_counter_read(field);
 464 }
 465
 466 /* Called with perf lock held */
 467 static
 468 void lttng_destroy_perf_thread_field(
 469                 struct lttng_perf_counter_thread_field *thread_field)
 470 {
 471         close_perf_fd(thread_field->fd);
 472         unmap_perf_page(thread_field->pc);
 473         cds_list_del_rcu(&thread_field->rcu_field_node);
 474         cds_list_del(&thread_field->thread_field_node);
 475         free(thread_field);
 476 }
 477
 478 static
 479 void lttng_destroy_perf_thread_key(void *_key)
 480 {
 481         struct lttng_perf_counter_thread *perf_thread = _key;
 482         struct lttng_perf_counter_thread_field *pos, *p;
 483
 484         lttng_perf_lock();
 485         cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list,
 486                         rcu_field_node)
 487                 lttng_destroy_perf_thread_field(pos);
 488         lttng_perf_unlock();
 489         free(perf_thread);
 490 }
 491
 492 /* Called with UST lock held */
 493 static
 494 void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
 495 {
 496         struct lttng_perf_counter_field *perf_field;
 497         struct lttng_perf_counter_thread_field *pos, *p;
 498
 499         free((char *) field->event_field.name);
 500         perf_field = field->u.perf_counter;
 501         /*
 502          * This put is performed when no threads can concurrently
 503          * perform a "get" concurrently, thanks to urcu-bp grace
 504          * period. Holding the lttng perf lock protects against
 505          * concurrent modification of the per-thread thread field
 506          * list.
 507          */
 508         lttng_perf_lock();
 509         cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list,
 510                         thread_field_node)
 511                 lttng_destroy_perf_thread_field(pos);
 512         lttng_perf_unlock();
 513         free(perf_field);
 514 }
 515
 516 #ifdef __ARM_ARCH_7A__
 517
 518 static
 519 int perf_get_exclude_kernel(void)
 520 {
 521         return 0;
 522 }
 523
 524 #else /* __ARM_ARCH_7A__ */
 525
 526 static
 527 int perf_get_exclude_kernel(void)
 528 {
 529         return 1;
 530 }
 531
 532 #endif /* __ARM_ARCH_7A__ */
 533
 534 /* Called with UST lock held */
 535 int lttng_add_perf_counter_to_ctx(uint32_t type,
 536                                 uint64_t config,
 537                                 const char *name,
 538                                 struct lttng_ctx **ctx)
 539 {
 540         struct lttng_ctx_field *field;
 541         struct lttng_perf_counter_field *perf_field;
 542         char *name_alloc;
 543         int ret;
 544
 545         name_alloc = strdup(name);
 546         if (!name_alloc) {
 547                 ret = -ENOMEM;
 548                 goto name_alloc_error;
 549         }
 550         perf_field = zmalloc(sizeof(*perf_field));
 551         if (!perf_field) {
 552                 ret = -ENOMEM;
 553                 goto perf_field_alloc_error;
 554         }
 555         field = lttng_append_context(ctx);
 556         if (!field) {
 557                 ret = -ENOMEM;
 558                 goto append_context_error;
 559         }
 560         if (lttng_find_context(*ctx, name_alloc)) {
 561                 ret = -EEXIST;
 562                 goto find_error;
 563         }
 564
 565         field->destroy = lttng_destroy_perf_counter_field;
 566
 567         field->event_field.name = name_alloc;
 568         field->event_field.type.atype = atype_integer;
 569         field->event_field.type.u.basic.integer.size =
 570                         sizeof(uint64_t) * CHAR_BIT;
 571         field->event_field.type.u.basic.integer.alignment =
 572                         lttng_alignof(uint64_t) * CHAR_BIT;
 573         field->event_field.type.u.basic.integer.signedness =
 574                         lttng_is_signed_type(uint64_t);
 575         field->event_field.type.u.basic.integer.reverse_byte_order = 0;
 576         field->event_field.type.u.basic.integer.base = 10;
 577         field->event_field.type.u.basic.integer.encoding = lttng_encode_none;
 578         field->get_size = perf_counter_get_size;
 579         field->record = perf_counter_record;
 580         field->get_value = perf_counter_get_value;
 581
 582         perf_field->attr.type = type;
 583         perf_field->attr.config = config;
 584         perf_field->attr.exclude_kernel = perf_get_exclude_kernel();
 585         CDS_INIT_LIST_HEAD(&perf_field->thread_field_list);
 586         field->u.perf_counter = perf_field;
 587
 588         /* Ensure that this perf counter can be used in this process. */
 589         ret = open_perf_fd(&perf_field->attr);
 590         if (ret < 0) {
 591                 ret = -ENODEV;
 592                 goto setup_error;
 593         }
 594         close_perf_fd(ret);
 595
 596         /*
 597          * Contexts can only be added before tracing is started, so we
 598          * don't have to synchronize against concurrent threads using
 599          * the field here.
 600          */
 601
 602         lttng_context_update(*ctx);
 603         return 0;
 604
 605 setup_error:
 606 find_error:
 607         lttng_remove_context_field(ctx, field);
 608 append_context_error:
 609         free(perf_field);
 610 perf_field_alloc_error:
 611         free(name_alloc);
 612 name_alloc_error:
 613         return ret;
 614 }
 615
 616 int lttng_perf_counter_init(void)
 617 {
 618         int ret;
 619
 620         ret = pthread_key_create(&perf_counter_key,
 621                         lttng_destroy_perf_thread_key);
 622         if (ret)
 623                 ret = -ret;
 624         return ret;
 625 }
 626
 627 void lttng_perf_counter_exit(void)
 628 {
 629         int ret;
 630
 631         ret = pthread_key_delete(perf_counter_key);
 632         if (ret) {
 633                 errno = ret;
 634                 PERROR("Error in pthread_key_delete");
 635         }
 636 }