tools/perf/util/mmap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011-2017, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from evlist.c builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <sys/mman.h>
  10 #include <inttypes.h>
  11 #include <asm/bug.h>
  12 #include <linux/zalloc.h>
  13 #include <stdlib.h>
  14 #include <string.h>
  15 #include <unistd.h> // sysconf()
  16 #ifdef HAVE_LIBNUMA_SUPPORT
  17 #include <numaif.h>
  18 #endif
  19 #include "cpumap.h"
  20 #include "debug.h"
  21 #include "event.h"
  22 #include "mmap.h"
  23 #include "../perf.h"
  24 #include <internal/lib.h> /* page_size */
  25
  26 size_t perf_mmap__mmap_len(struct mmap *map)
  27 {
  28         return map->core.mask + 1 + page_size;
  29 }
  30
  31 /* When check_messup is true, 'end' must points to a good entry */
  32 static union perf_event *perf_mmap__read(struct mmap *map,
  33                                          u64 *startp, u64 end)
  34 {
  35         unsigned char *data = map->core.base + page_size;
  36         union perf_event *event = NULL;
  37         int diff = end - *startp;
  38
  39         if (diff >= (int)sizeof(event->header)) {
  40                 size_t size;
  41
  42                 event = (union perf_event *)&data[*startp & map->core.mask];
  43                 size = event->header.size;
  44
  45                 if (size < sizeof(event->header) || diff < (int)size)
  46                         return NULL;
  47
  48                 /*
  49                  * Event straddles the mmap boundary -- header should always
  50                  * be inside due to u64 alignment of output.
  51                  */
  52                 if ((*startp & map->core.mask) + size != ((*startp + size) & map->core.mask)) {
  53                         unsigned int offset = *startp;
  54                         unsigned int len = min(sizeof(*event), size), cpy;
  55                         void *dst = map->core.event_copy;
  56
  57                         do {
  58                                 cpy = min(map->core.mask + 1 - (offset & map->core.mask), len);
  59                                 memcpy(dst, &data[offset & map->core.mask], cpy);
  60                                 offset += cpy;
  61                                 dst += cpy;
  62                                 len -= cpy;
  63                         } while (len);
  64
  65                         event = (union perf_event *)map->core.event_copy;
  66                 }
  67
  68                 *startp += size;
  69         }
  70
  71         return event;
  72 }
  73
  74 /*
  75  * Read event from ring buffer one by one.
  76  * Return one event for each call.
  77  *
  78  * Usage:
  79  * perf_mmap__read_init()
  80  * while(event = perf_mmap__read_event()) {
  81  *      //process the event
  82  *      perf_mmap__consume()
  83  * }
  84  * perf_mmap__read_done()
  85  */
  86 union perf_event *perf_mmap__read_event(struct mmap *map)
  87 {
  88         union perf_event *event;
  89
  90         /*
  91          * Check if event was unmapped due to a POLLHUP/POLLERR.
  92          */
  93         if (!refcount_read(&map->core.refcnt))
  94                 return NULL;
  95
  96         /* non-overwirte doesn't pause the ringbuffer */
  97         if (!map->core.overwrite)
  98                 map->core.end = perf_mmap__read_head(map);
  99
 100         event = perf_mmap__read(map, &map->core.start, map->core.end);
 101
 102         if (!map->core.overwrite)
 103                 map->core.prev = map->core.start;
 104
 105         return event;
 106 }
 107
 108 static bool perf_mmap__empty(struct mmap *map)
 109 {
 110         return perf_mmap__read_head(map) == map->core.prev && !map->auxtrace_mmap.base;
 111 }
 112
 113 void perf_mmap__get(struct mmap *map)
 114 {
 115         refcount_inc(&map->core.refcnt);
 116 }
 117
 118 void perf_mmap__put(struct mmap *map)
 119 {
 120         BUG_ON(map->core.base && refcount_read(&map->core.refcnt) == 0);
 121
 122         if (refcount_dec_and_test(&map->core.refcnt))
 123                 perf_mmap__munmap(map);
 124 }
 125
 126 void perf_mmap__consume(struct mmap *map)
 127 {
 128         if (!map->core.overwrite) {
 129                 u64 old = map->core.prev;
 130
 131                 perf_mmap__write_tail(map, old);
 132         }
 133
 134         if (refcount_read(&map->core.refcnt) == 1 && perf_mmap__empty(map))
 135                 perf_mmap__put(map);
 136 }
 137
 138 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 139                                struct auxtrace_mmap_params *mp __maybe_unused,
 140                                void *userpg __maybe_unused,
 141                                int fd __maybe_unused)
 142 {
 143         return 0;
 144 }
 145
 146 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 147 {
 148 }
 149
 150 void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
 151                                        off_t auxtrace_offset __maybe_unused,
 152                                        unsigned int auxtrace_pages __maybe_unused,
 153                                        bool auxtrace_overwrite __maybe_unused)
 154 {
 155 }
 156
 157 void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
 158                                           struct evlist *evlist __maybe_unused,
 159                                           int idx __maybe_unused,
 160                                           bool per_cpu __maybe_unused)
 161 {
 162 }
 163
 164 #ifdef HAVE_AIO_SUPPORT
 165 static int perf_mmap__aio_enabled(struct mmap *map)
 166 {
 167         return map->aio.nr_cblocks > 0;
 168 }
 169
 170 #ifdef HAVE_LIBNUMA_SUPPORT
 171 static int perf_mmap__aio_alloc(struct mmap *map, int idx)
 172 {
 173         map->aio.data[idx] = mmap(NULL, perf_mmap__mmap_len(map), PROT_READ|PROT_WRITE,
 174                                   MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 175         if (map->aio.data[idx] == MAP_FAILED) {
 176                 map->aio.data[idx] = NULL;
 177                 return -1;
 178         }
 179
 180         return 0;
 181 }
 182
 183 static void perf_mmap__aio_free(struct mmap *map, int idx)
 184 {
 185         if (map->aio.data[idx]) {
 186                 munmap(map->aio.data[idx], perf_mmap__mmap_len(map));
 187                 map->aio.data[idx] = NULL;
 188         }
 189 }
 190
 191 static int perf_mmap__aio_bind(struct mmap *map, int idx, int cpu, int affinity)
 192 {
 193         void *data;
 194         size_t mmap_len;
 195         unsigned long node_mask;
 196
 197         if (affinity != PERF_AFFINITY_SYS && cpu__max_node() > 1) {
 198                 data = map->aio.data[idx];
 199                 mmap_len = perf_mmap__mmap_len(map);
 200                 node_mask = 1UL << cpu__get_node(cpu);
 201                 if (mbind(data, mmap_len, MPOL_BIND, &node_mask, 1, 0)) {
 202                         pr_err("Failed to bind [%p-%p] AIO buffer to node %d: error %m\n",
 203                                 data, data + mmap_len, cpu__get_node(cpu));
 204                         return -1;
 205                 }
 206         }
 207
 208         return 0;
 209 }
 210 #else /* !HAVE_LIBNUMA_SUPPORT */
 211 static int perf_mmap__aio_alloc(struct mmap *map, int idx)
 212 {
 213         map->aio.data[idx] = malloc(perf_mmap__mmap_len(map));
 214         if (map->aio.data[idx] == NULL)
 215                 return -1;
 216
 217         return 0;
 218 }
 219
 220 static void perf_mmap__aio_free(struct mmap *map, int idx)
 221 {
 222         zfree(&(map->aio.data[idx]));
 223 }
 224
 225 static int perf_mmap__aio_bind(struct mmap *map __maybe_unused, int idx __maybe_unused,
 226                 int cpu __maybe_unused, int affinity __maybe_unused)
 227 {
 228         return 0;
 229 }
 230 #endif
 231
 232 static int perf_mmap__aio_mmap(struct mmap *map, struct mmap_params *mp)
 233 {
 234         int delta_max, i, prio, ret;
 235
 236         map->aio.nr_cblocks = mp->nr_cblocks;
 237         if (map->aio.nr_cblocks) {
 238                 map->aio.aiocb = calloc(map->aio.nr_cblocks, sizeof(struct aiocb *));
 239                 if (!map->aio.aiocb) {
 240                         pr_debug2("failed to allocate aiocb for data buffer, error %m\n");
 241                         return -1;
 242                 }
 243                 map->aio.cblocks = calloc(map->aio.nr_cblocks, sizeof(struct aiocb));
 244                 if (!map->aio.cblocks) {
 245                         pr_debug2("failed to allocate cblocks for data buffer, error %m\n");
 246                         return -1;
 247                 }
 248                 map->aio.data = calloc(map->aio.nr_cblocks, sizeof(void *));
 249                 if (!map->aio.data) {
 250                         pr_debug2("failed to allocate data buffer, error %m\n");
 251                         return -1;
 252                 }
 253                 delta_max = sysconf(_SC_AIO_PRIO_DELTA_MAX);
 254                 for (i = 0; i < map->aio.nr_cblocks; ++i) {
 255                         ret = perf_mmap__aio_alloc(map, i);
 256                         if (ret == -1) {
 257                                 pr_debug2("failed to allocate data buffer area, error %m");
 258                                 return -1;
 259                         }
 260                         ret = perf_mmap__aio_bind(map, i, map->core.cpu, mp->affinity);
 261                         if (ret == -1)
 262                                 return -1;
 263                         /*
 264                          * Use cblock.aio_fildes value different from -1
 265                          * to denote started aio write operation on the
 266                          * cblock so it requires explicit record__aio_sync()
 267                          * call prior the cblock may be reused again.
 268                          */
 269                         map->aio.cblocks[i].aio_fildes = -1;
 270                         /*
 271                          * Allocate cblocks with priority delta to have
 272                          * faster aio write system calls because queued requests
 273                          * are kept in separate per-prio queues and adding
 274                          * a new request will iterate thru shorter per-prio
 275                          * list. Blocks with numbers higher than
 276                          *  _SC_AIO_PRIO_DELTA_MAX go with priority 0.
 277                          */
 278                         prio = delta_max - i;
 279                         map->aio.cblocks[i].aio_reqprio = prio >= 0 ? prio : 0;
 280                 }
 281         }
 282
 283         return 0;
 284 }
 285
 286 static void perf_mmap__aio_munmap(struct mmap *map)
 287 {
 288         int i;
 289
 290         for (i = 0; i < map->aio.nr_cblocks; ++i)
 291                 perf_mmap__aio_free(map, i);
 292         if (map->aio.data)
 293                 zfree(&map->aio.data);
 294         zfree(&map->aio.cblocks);
 295         zfree(&map->aio.aiocb);
 296 }
 297 #else /* !HAVE_AIO_SUPPORT */
 298 static int perf_mmap__aio_enabled(struct mmap *map __maybe_unused)
 299 {
 300         return 0;
 301 }
 302
 303 static int perf_mmap__aio_mmap(struct mmap *map __maybe_unused,
 304                                struct mmap_params *mp __maybe_unused)
 305 {
 306         return 0;
 307 }
 308
 309 static void perf_mmap__aio_munmap(struct mmap *map __maybe_unused)
 310 {
 311 }
 312 #endif
 313
 314 void perf_mmap__munmap(struct mmap *map)
 315 {
 316         perf_mmap__aio_munmap(map);
 317         if (map->data != NULL) {
 318                 munmap(map->data, perf_mmap__mmap_len(map));
 319                 map->data = NULL;
 320         }
 321         if (map->core.base != NULL) {
 322                 munmap(map->core.base, perf_mmap__mmap_len(map));
 323                 map->core.base = NULL;
 324                 map->core.fd = -1;
 325                 refcount_set(&map->core.refcnt, 0);
 326         }
 327         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 328 }
 329
 330 static void build_node_mask(int node, cpu_set_t *mask)
 331 {
 332         int c, cpu, nr_cpus;
 333         const struct perf_cpu_map *cpu_map = NULL;
 334
 335         cpu_map = cpu_map__online();
 336         if (!cpu_map)
 337                 return;
 338
 339         nr_cpus = perf_cpu_map__nr(cpu_map);
 340         for (c = 0; c < nr_cpus; c++) {
 341                 cpu = cpu_map->map[c]; /* map c index to online cpu index */
 342                 if (cpu__get_node(cpu) == node)
 343                         CPU_SET(cpu, mask);
 344         }
 345 }
 346
 347 static void perf_mmap__setup_affinity_mask(struct mmap *map, struct mmap_params *mp)
 348 {
 349         CPU_ZERO(&map->affinity_mask);
 350         if (mp->affinity == PERF_AFFINITY_NODE && cpu__max_node() > 1)
 351                 build_node_mask(cpu__get_node(map->core.cpu), &map->affinity_mask);
 352         else if (mp->affinity == PERF_AFFINITY_CPU)
 353                 CPU_SET(map->core.cpu, &map->affinity_mask);
 354 }
 355
 356 int perf_mmap__mmap(struct mmap *map, struct mmap_params *mp, int fd, int cpu)
 357 {
 358         /*
 359          * The last one will be done at perf_mmap__consume(), so that we
 360          * make sure we don't prevent tools from consuming every last event in
 361          * the ring buffer.
 362          *
 363          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 364          * anymore, but the last events for it are still in the ring buffer,
 365          * waiting to be consumed.
 366          *
 367          * Tools can chose to ignore this at their own discretion, but the
 368          * evlist layer can't just drop it when filtering events in
 369          * perf_evlist__filter_pollfd().
 370          */
 371         refcount_set(&map->core.refcnt, 2);
 372         map->core.prev = 0;
 373         map->core.mask = mp->mask;
 374         map->core.base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
 375                          MAP_SHARED, fd, 0);
 376         if (map->core.base == MAP_FAILED) {
 377                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 378                           errno);
 379                 map->core.base = NULL;
 380                 return -1;
 381         }
 382         map->core.fd = fd;
 383         map->core.cpu = cpu;
 384
 385         perf_mmap__setup_affinity_mask(map, mp);
 386
 387         map->core.flush = mp->flush;
 388
 389         map->comp_level = mp->comp_level;
 390
 391         if (map->comp_level && !perf_mmap__aio_enabled(map)) {
 392                 map->data = mmap(NULL, perf_mmap__mmap_len(map), PROT_READ|PROT_WRITE,
 393                                  MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 394                 if (map->data == MAP_FAILED) {
 395                         pr_debug2("failed to mmap data buffer, error %d\n",
 396                                         errno);
 397                         map->data = NULL;
 398                         return -1;
 399                 }
 400         }
 401
 402         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
 403                                 &mp->auxtrace_mp, map->core.base, fd))
 404                 return -1;
 405
 406         return perf_mmap__aio_mmap(map, mp);
 407 }
 408
 409 static int overwrite_rb_find_range(void *buf, int mask, u64 *start, u64 *end)
 410 {
 411         struct perf_event_header *pheader;
 412         u64 evt_head = *start;
 413         int size = mask + 1;
 414
 415         pr_debug2("%s: buf=%p, start=%"PRIx64"\n", __func__, buf, *start);
 416         pheader = (struct perf_event_header *)(buf + (*start & mask));
 417         while (true) {
 418                 if (evt_head - *start >= (unsigned int)size) {
 419                         pr_debug("Finished reading overwrite ring buffer: rewind\n");
 420                         if (evt_head - *start > (unsigned int)size)
 421                                 evt_head -= pheader->size;
 422                         *end = evt_head;
 423                         return 0;
 424                 }
 425
 426                 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
 427
 428                 if (pheader->size == 0) {
 429                         pr_debug("Finished reading overwrite ring buffer: get start\n");
 430                         *end = evt_head;
 431                         return 0;
 432                 }
 433
 434                 evt_head += pheader->size;
 435                 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
 436         }
 437         WARN_ONCE(1, "Shouldn't get here\n");
 438         return -1;
 439 }
 440
 441 /*
 442  * Report the start and end of the available data in ringbuffer
 443  */
 444 static int __perf_mmap__read_init(struct mmap *md)
 445 {
 446         u64 head = perf_mmap__read_head(md);
 447         u64 old = md->core.prev;
 448         unsigned char *data = md->core.base + page_size;
 449         unsigned long size;
 450
 451         md->core.start = md->core.overwrite ? head : old;
 452         md->core.end = md->core.overwrite ? old : head;
 453
 454         if ((md->core.end - md->core.start) < md->core.flush)
 455                 return -EAGAIN;
 456
 457         size = md->core.end - md->core.start;
 458         if (size > (unsigned long)(md->core.mask) + 1) {
 459                 if (!md->core.overwrite) {
 460                         WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
 461
 462                         md->core.prev = head;
 463                         perf_mmap__consume(md);
 464                         return -EAGAIN;
 465                 }
 466
 467                 /*
 468                  * Backward ring buffer is full. We still have a chance to read
 469                  * most of data from it.
 470                  */
 471                 if (overwrite_rb_find_range(data, md->core.mask, &md->core.start, &md->core.end))
 472                         return -EINVAL;
 473         }
 474
 475         return 0;
 476 }
 477
 478 int perf_mmap__read_init(struct mmap *map)
 479 {
 480         /*
 481          * Check if event was unmapped due to a POLLHUP/POLLERR.
 482          */
 483         if (!refcount_read(&map->core.refcnt))
 484                 return -ENOENT;
 485
 486         return __perf_mmap__read_init(map);
 487 }
 488
 489 int perf_mmap__push(struct mmap *md, void *to,
 490                     int push(struct mmap *map, void *to, void *buf, size_t size))
 491 {
 492         u64 head = perf_mmap__read_head(md);
 493         unsigned char *data = md->core.base + page_size;
 494         unsigned long size;
 495         void *buf;
 496         int rc = 0;
 497
 498         rc = perf_mmap__read_init(md);
 499         if (rc < 0)
 500                 return (rc == -EAGAIN) ? 1 : -1;
 501
 502         size = md->core.end - md->core.start;
 503
 504         if ((md->core.start & md->core.mask) + size != (md->core.end & md->core.mask)) {
 505                 buf = &data[md->core.start & md->core.mask];
 506                 size = md->core.mask + 1 - (md->core.start & md->core.mask);
 507                 md->core.start += size;
 508
 509                 if (push(md, to, buf, size) < 0) {
 510                         rc = -1;
 511                         goto out;
 512                 }
 513         }
 514
 515         buf = &data[md->core.start & md->core.mask];
 516         size = md->core.end - md->core.start;
 517         md->core.start += size;
 518
 519         if (push(md, to, buf, size) < 0) {
 520                 rc = -1;
 521                 goto out;
 522         }
 523
 524         md->core.prev = head;
 525         perf_mmap__consume(md);
 526 out:
 527         return rc;
 528 }
 529
 530 /*
 531  * Mandatory for overwrite mode
 532  * The direction of overwrite mode is backward.
 533  * The last perf_mmap__read() will set tail to map->core.prev.
 534  * Need to correct the map->core.prev to head which is the end of next read.
 535  */
 536 void perf_mmap__read_done(struct mmap *map)
 537 {
 538         /*
 539          * Check if event was unmapped due to a POLLHUP/POLLERR.
 540          */
 541         if (!refcount_read(&map->core.refcnt))
 542                 return;
 543
 544         map->core.prev = perf_mmap__read_head(map);
 545 }