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 #include <sys/types.h>
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <stdlib.h>
  27 #include <stdio.h>
  28 #include <sys/mman.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>
  37 #include <helper.h>
  38 #include <urcu/ref.h>
  39 #include <usterr-signal-safe.h>
  40 #include <signal.h>
  41 #include "lttng-tracer-core.h"
  42
  43 /*
  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.
  48  *
  49  * Updates and traversals of thread_list are protected by UST lock.
  50  * Updates to rcu_field_list are protected by UST lock.
  51  */
  52
  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) */
  58 };
  59
  60 struct lttng_perf_counter_thread {
  61         struct cds_list_head rcu_field_list;    /* RCU per-thread list of fields */
  62 };
  63
  64 struct lttng_perf_counter_field {
  65         struct perf_event_attr attr;
  66         struct cds_list_head thread_field_list; /* Per-field list of thread fields */
  67 };
  68
  69 static pthread_key_t perf_counter_key;
  70
  71 static
  72 size_t perf_counter_get_size(size_t offset)
  73 {
  74         size_t size = 0;
  75
  76         size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t));
  77         size += sizeof(uint64_t);
  78         return size;
  79 }
  80
  81 #if defined(__x86_64__) || defined(__i386__)
  82
  83 static
  84 uint64_t rdpmc(unsigned int counter)
  85 {
  86         unsigned int low, high;
  87
  88         asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter));
  89
  90         return low | ((uint64_t) high) << 32;
  91 }
  92
  93 #else /* defined(__x86_64__) || defined(__i386__) */
  94
  95 #error "Perf event counters are only supported on x86 so far."
  96
  97 #endif /* #else defined(__x86_64__) || defined(__i386__) */
  98
  99 static
 100 uint64_t read_perf_counter(struct perf_event_mmap_page *pc)
 101 {
 102         uint32_t seq, idx;
 103         uint64_t count;
 104
 105         if (caa_unlikely(!pc))
 106                 return 0;
 107
 108         do {
 109                 seq = CMM_LOAD_SHARED(pc->lock);
 110                 cmm_barrier();
 111
 112                 idx = pc->index;
 113                 if (idx)
 114                         count = pc->offset + rdpmc(idx - 1);
 115                 else
 116                         count = 0;
 117
 118                 cmm_barrier();
 119         } while (CMM_LOAD_SHARED(pc->lock) != seq);
 120
 121         return count;
 122 }
 123
 124 static
 125 int sys_perf_event_open(struct perf_event_attr *attr,
 126                 pid_t pid, int cpu, int group_fd,
 127                 unsigned long flags)
 128 {
 129         return syscall(SYS_perf_event_open, attr, pid, cpu,
 130                         group_fd, flags);
 131 }
 132
 133 static
 134 struct perf_event_mmap_page *setup_perf(struct perf_event_attr *attr)
 135 {
 136         void *perf_addr;
 137         int fd;
 138
 139         fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 140         if (fd < 0)
 141                 return NULL;
 142
 143         perf_addr = mmap(NULL, sizeof(struct perf_event_mmap_page),
 144                         PROT_READ, MAP_SHARED, fd, 0);
 145         if (perf_addr == MAP_FAILED)
 146                 return NULL;
 147         close(fd);
 148         return perf_addr;
 149 }
 150
 151 static
 152 void unmap_perf_page(struct perf_event_mmap_page *pc)
 153 {
 154         int ret;
 155
 156         if (!pc)
 157                 return;
 158         ret = munmap(pc, sizeof(struct perf_event_mmap_page));
 159         if (ret < 0) {
 160                 PERROR("Error in munmap");
 161                 abort();
 162         }
 163 }
 164
 165 static
 166 struct lttng_perf_counter_thread *alloc_perf_counter_thread(void)
 167 {
 168         struct lttng_perf_counter_thread *perf_thread;
 169         sigset_t newmask, oldmask;
 170         int ret;
 171
 172         ret = sigfillset(&newmask);
 173         if (ret)
 174                 abort();
 175         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 176         if (ret)
 177                 abort();
 178         /* Check again with signals disabled */
 179         perf_thread = pthread_getspecific(perf_counter_key);
 180         if (perf_thread)
 181                 goto skip;
 182         perf_thread = zmalloc(sizeof(*perf_thread));
 183         if (!perf_thread)
 184                 abort();
 185         CDS_INIT_LIST_HEAD(&perf_thread->rcu_field_list);
 186         ret = pthread_setspecific(perf_counter_key, perf_thread);
 187         if (ret)
 188                 abort();
 189 skip:
 190         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 191         if (ret)
 192                 abort();
 193         return perf_thread;
 194 }
 195
 196 static
 197 struct lttng_perf_counter_thread_field *
 198         add_thread_field(struct lttng_perf_counter_field *perf_field,
 199                 struct lttng_perf_counter_thread *perf_thread)
 200 {
 201         struct lttng_perf_counter_thread_field *thread_field;
 202         sigset_t newmask, oldmask;
 203         int ret;
 204
 205         ret = sigfillset(&newmask);
 206         if (ret)
 207                 abort();
 208         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 209         if (ret)
 210                 abort();
 211         /* Check again with signals disabled */
 212         cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list,
 213                         rcu_field_node) {
 214                 if (thread_field->field == perf_field)
 215                         goto skip;
 216         }
 217         thread_field = zmalloc(sizeof(*thread_field));
 218         if (!thread_field)
 219                 abort();
 220         thread_field->field = perf_field;
 221         thread_field->pc = setup_perf(&perf_field->attr);
 222         /* Note: thread_field->pc can be NULL if setup_perf() fails. */
 223         ust_lock_nocheck();
 224         cds_list_add_rcu(&thread_field->rcu_field_node,
 225                         &perf_thread->rcu_field_list);
 226         cds_list_add(&thread_field->thread_field_node,
 227                         &perf_field->thread_field_list);
 228         ust_unlock();
 229 skip:
 230         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 231         if (ret)
 232                 abort();
 233         return thread_field;
 234 }
 235
 236 static
 237 struct lttng_perf_counter_thread_field *
 238                 get_thread_field(struct lttng_perf_counter_field *field)
 239 {
 240         struct lttng_perf_counter_thread *perf_thread;
 241         struct lttng_perf_counter_thread_field *thread_field;
 242
 243         perf_thread = pthread_getspecific(perf_counter_key);
 244         if (!perf_thread)
 245                 perf_thread = alloc_perf_counter_thread();
 246         cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list,
 247                         rcu_field_node) {
 248                 if (thread_field->field == field)
 249                         return thread_field;
 250         }
 251         /* perf_counter_thread_field not found, need to add one */
 252         return add_thread_field(field, perf_thread);
 253 }
 254
 255 static
 256 uint64_t wrapper_perf_counter_read(struct lttng_ctx_field *field)
 257 {
 258         struct lttng_perf_counter_field *perf_field;
 259         struct lttng_perf_counter_thread_field *perf_thread_field;
 260
 261         perf_field = field->u.perf_counter;
 262         perf_thread_field = get_thread_field(perf_field);
 263         return read_perf_counter(perf_thread_field->pc);
 264 }
 265
 266 static
 267 void perf_counter_record(struct lttng_ctx_field *field,
 268                  struct lttng_ust_lib_ring_buffer_ctx *ctx,
 269                  struct lttng_channel *chan)
 270 {
 271         uint64_t value;
 272
 273         value = wrapper_perf_counter_read(field);
 274         lib_ring_buffer_align_ctx(ctx, lttng_alignof(value));
 275         chan->ops->event_write(ctx, &value, sizeof(value));
 276 }
 277
 278 static
 279 void perf_counter_get_value(struct lttng_ctx_field *field,
 280                 union lttng_ctx_value *value)
 281 {
 282         uint64_t v;
 283
 284         v = wrapper_perf_counter_read(field);
 285         value->s64 = v;
 286 }
 287
 288 /* Called with UST lock held */
 289 static
 290 void lttng_destroy_perf_thread_field(
 291                 struct lttng_perf_counter_thread_field *thread_field)
 292 {
 293         unmap_perf_page(thread_field->pc);
 294         cds_list_del_rcu(&thread_field->rcu_field_node);
 295         cds_list_del(&thread_field->thread_field_node);
 296         free(thread_field);
 297 }
 298
 299 static
 300 void lttng_destroy_perf_thread_key(void *_key)
 301 {
 302         struct lttng_perf_counter_thread *perf_thread = _key;
 303         struct lttng_perf_counter_thread_field *pos, *p;
 304
 305         ust_lock_nocheck();
 306         cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list,
 307                         rcu_field_node)
 308                 lttng_destroy_perf_thread_field(pos);
 309         ust_unlock();
 310         free(perf_thread);
 311 }
 312
 313 /* Called with UST lock held */
 314 static
 315 void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
 316 {
 317         struct lttng_perf_counter_field *perf_field;
 318         struct lttng_perf_counter_thread_field *pos, *p;
 319
 320         free((char *) field->event_field.name);
 321         perf_field = field->u.perf_counter;
 322         /*
 323          * This put is performed when no threads can concurrently
 324          * perform a "get" concurrently, thanks to urcu-bp grace
 325          * period.
 326          */
 327         cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list,
 328                         thread_field_node)
 329                 lttng_destroy_perf_thread_field(pos);
 330         free(perf_field);
 331 }
 332
 333 /* Called with UST lock held */
 334 int lttng_add_perf_counter_to_ctx(uint32_t type,
 335                                 uint64_t config,
 336                                 const char *name,
 337                                 struct lttng_ctx **ctx)
 338 {
 339         struct lttng_ctx_field *field;
 340         struct lttng_perf_counter_field *perf_field;
 341         struct perf_event_mmap_page *tmp_pc;
 342         char *name_alloc;
 343         int ret;
 344
 345         name_alloc = strdup(name);
 346         if (!name_alloc) {
 347                 ret = -ENOMEM;
 348                 goto name_alloc_error;
 349         }
 350         perf_field = zmalloc(sizeof(*perf_field));
 351         if (!perf_field) {
 352                 ret = -ENOMEM;
 353                 goto perf_field_alloc_error;
 354         }
 355         field = lttng_append_context(ctx);
 356         if (!field) {
 357                 ret = -ENOMEM;
 358                 goto append_context_error;
 359         }
 360         if (lttng_find_context(*ctx, name_alloc)) {
 361                 ret = -EEXIST;
 362                 goto find_error;
 363         }
 364
 365         field->destroy = lttng_destroy_perf_counter_field;
 366
 367         field->event_field.name = name_alloc;
 368         field->event_field.type.atype = atype_integer;
 369         field->event_field.type.u.basic.integer.size =
 370                         sizeof(uint64_t) * CHAR_BIT;
 371         field->event_field.type.u.basic.integer.alignment =
 372                         lttng_alignof(uint64_t) * CHAR_BIT;
 373         field->event_field.type.u.basic.integer.signedness =
 374                         lttng_is_signed_type(uint64_t);
 375         field->event_field.type.u.basic.integer.reverse_byte_order = 0;
 376         field->event_field.type.u.basic.integer.base = 10;
 377         field->event_field.type.u.basic.integer.encoding = lttng_encode_none;
 378         field->get_size = perf_counter_get_size;
 379         field->record = perf_counter_record;
 380         field->get_value = perf_counter_get_value;
 381
 382         perf_field->attr.type = type;
 383         perf_field->attr.config = config;
 384         perf_field->attr.exclude_kernel = 1;
 385         CDS_INIT_LIST_HEAD(&perf_field->thread_field_list);
 386         field->u.perf_counter = perf_field;
 387
 388         /* Ensure that this perf counter can be used in this process. */
 389         tmp_pc = setup_perf(&perf_field->attr);
 390         if (!tmp_pc) {
 391                 ret = -ENODEV;
 392                 goto setup_error;
 393         }
 394         unmap_perf_page(tmp_pc);
 395
 396         /*
 397          * Contexts can only be added before tracing is started, so we
 398          * don't have to synchronize against concurrent threads using
 399          * the field here.
 400          */
 401
 402         return 0;
 403
 404 setup_error:
 405 find_error:
 406         lttng_remove_context_field(ctx, field);
 407 append_context_error:
 408         free(perf_field);
 409 perf_field_alloc_error:
 410         free(name_alloc);
 411 name_alloc_error:
 412         return ret;
 413 }
 414
 415 int lttng_perf_counter_init(void)
 416 {
 417         int ret;
 418
 419         ret = pthread_key_create(&perf_counter_key,
 420                         lttng_destroy_perf_thread_key);
 421         if (ret)
 422                 ret = -ret;
 423         return ret;
 424 }
 425
 426 void lttng_perf_counter_exit(void)
 427 {
 428         int ret;
 429
 430         ret = pthread_key_delete(perf_counter_key);
 431         if (ret) {
 432                 errno = ret;
 433                 PERROR("Error in pthread_key_delete");
 434         }
 435 }