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 #include <perf/mmap.h>
  17 #ifdef HAVE_LIBNUMA_SUPPORT
  18 #include <numaif.h>
  19 #endif
  20 #include "cpumap.h"
  21 #include "debug.h"
  22 #include "event.h"
  23 #include "mmap.h"
  24 #include "../perf.h"
  25 #include <internal/lib.h> /* page_size */
  26 #include <linux/bitmap.h>
  27
  28 #define MASK_SIZE 1023
  29 void mmap_cpu_mask__scnprintf(struct mmap_cpu_mask *mask, const char *tag)
  30 {
  31         char buf[MASK_SIZE + 1];
  32         size_t len;
  33
  34         len = bitmap_scnprintf(mask->bits, mask->nbits, buf, MASK_SIZE);
  35         buf[len] = '\0';
  36         pr_debug("%p: %s mask[%zd]: %s\n", mask, tag, mask->nbits, buf);
  37 }
  38
  39 size_t mmap__mmap_len(struct mmap *map)
  40 {
  41         return perf_mmap__mmap_len(&map->core);
  42 }
  43
  44 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
  45                                struct auxtrace_mmap_params *mp __maybe_unused,
  46                                void *userpg __maybe_unused,
  47                                int fd __maybe_unused)
  48 {
  49         return 0;
  50 }
  51
  52 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
  53 {
  54 }
  55
  56 void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
  57                                        off_t auxtrace_offset __maybe_unused,
  58                                        unsigned int auxtrace_pages __maybe_unused,
  59                                        bool auxtrace_overwrite __maybe_unused)
  60 {
  61 }
  62
  63 void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
  64                                           struct evlist *evlist __maybe_unused,
  65                                           int idx __maybe_unused,
  66                                           bool per_cpu __maybe_unused)
  67 {
  68 }
  69
  70 #ifdef HAVE_AIO_SUPPORT
  71 static int perf_mmap__aio_enabled(struct mmap *map)
  72 {
  73         return map->aio.nr_cblocks > 0;
  74 }
  75
  76 #ifdef HAVE_LIBNUMA_SUPPORT
  77 static int perf_mmap__aio_alloc(struct mmap *map, int idx)
  78 {
  79         map->aio.data[idx] = mmap(NULL, mmap__mmap_len(map), PROT_READ|PROT_WRITE,
  80                                   MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  81         if (map->aio.data[idx] == MAP_FAILED) {
  82                 map->aio.data[idx] = NULL;
  83                 return -1;
  84         }
  85
  86         return 0;
  87 }
  88
  89 static void perf_mmap__aio_free(struct mmap *map, int idx)
  90 {
  91         if (map->aio.data[idx]) {
  92                 munmap(map->aio.data[idx], mmap__mmap_len(map));
  93                 map->aio.data[idx] = NULL;
  94         }
  95 }
  96
  97 static int perf_mmap__aio_bind(struct mmap *map, int idx, int cpu, int affinity)
  98 {
  99         void *data;
 100         size_t mmap_len;
 101         unsigned long node_mask;
 102
 103         if (affinity != PERF_AFFINITY_SYS && cpu__max_node() > 1) {
 104                 data = map->aio.data[idx];
 105                 mmap_len = mmap__mmap_len(map);
 106                 node_mask = 1UL << cpu__get_node(cpu);
 107                 if (mbind(data, mmap_len, MPOL_BIND, &node_mask, 1, 0)) {
 108                         pr_err("Failed to bind [%p-%p] AIO buffer to node %d: error %m\n",
 109                                 data, data + mmap_len, cpu__get_node(cpu));
 110                         return -1;
 111                 }
 112         }
 113
 114         return 0;
 115 }
 116 #else /* !HAVE_LIBNUMA_SUPPORT */
 117 static int perf_mmap__aio_alloc(struct mmap *map, int idx)
 118 {
 119         map->aio.data[idx] = malloc(mmap__mmap_len(map));
 120         if (map->aio.data[idx] == NULL)
 121                 return -1;
 122
 123         return 0;
 124 }
 125
 126 static void perf_mmap__aio_free(struct mmap *map, int idx)
 127 {
 128         zfree(&(map->aio.data[idx]));
 129 }
 130
 131 static int perf_mmap__aio_bind(struct mmap *map __maybe_unused, int idx __maybe_unused,
 132                 int cpu __maybe_unused, int affinity __maybe_unused)
 133 {
 134         return 0;
 135 }
 136 #endif
 137
 138 static int perf_mmap__aio_mmap(struct mmap *map, struct mmap_params *mp)
 139 {
 140         int delta_max, i, prio, ret;
 141
 142         map->aio.nr_cblocks = mp->nr_cblocks;
 143         if (map->aio.nr_cblocks) {
 144                 map->aio.aiocb = calloc(map->aio.nr_cblocks, sizeof(struct aiocb *));
 145                 if (!map->aio.aiocb) {
 146                         pr_debug2("failed to allocate aiocb for data buffer, error %m\n");
 147                         return -1;
 148                 }
 149                 map->aio.cblocks = calloc(map->aio.nr_cblocks, sizeof(struct aiocb));
 150                 if (!map->aio.cblocks) {
 151                         pr_debug2("failed to allocate cblocks for data buffer, error %m\n");
 152                         return -1;
 153                 }
 154                 map->aio.data = calloc(map->aio.nr_cblocks, sizeof(void *));
 155                 if (!map->aio.data) {
 156                         pr_debug2("failed to allocate data buffer, error %m\n");
 157                         return -1;
 158                 }
 159                 delta_max = sysconf(_SC_AIO_PRIO_DELTA_MAX);
 160                 for (i = 0; i < map->aio.nr_cblocks; ++i) {
 161                         ret = perf_mmap__aio_alloc(map, i);
 162                         if (ret == -1) {
 163                                 pr_debug2("failed to allocate data buffer area, error %m");
 164                                 return -1;
 165                         }
 166                         ret = perf_mmap__aio_bind(map, i, map->core.cpu, mp->affinity);
 167                         if (ret == -1)
 168                                 return -1;
 169                         /*
 170                          * Use cblock.aio_fildes value different from -1
 171                          * to denote started aio write operation on the
 172                          * cblock so it requires explicit record__aio_sync()
 173                          * call prior the cblock may be reused again.
 174                          */
 175                         map->aio.cblocks[i].aio_fildes = -1;
 176                         /*
 177                          * Allocate cblocks with priority delta to have
 178                          * faster aio write system calls because queued requests
 179                          * are kept in separate per-prio queues and adding
 180                          * a new request will iterate thru shorter per-prio
 181                          * list. Blocks with numbers higher than
 182                          *  _SC_AIO_PRIO_DELTA_MAX go with priority 0.
 183                          */
 184                         prio = delta_max - i;
 185                         map->aio.cblocks[i].aio_reqprio = prio >= 0 ? prio : 0;
 186                 }
 187         }
 188
 189         return 0;
 190 }
 191
 192 static void perf_mmap__aio_munmap(struct mmap *map)
 193 {
 194         int i;
 195
 196         for (i = 0; i < map->aio.nr_cblocks; ++i)
 197                 perf_mmap__aio_free(map, i);
 198         if (map->aio.data)
 199                 zfree(&map->aio.data);
 200         zfree(&map->aio.cblocks);
 201         zfree(&map->aio.aiocb);
 202 }
 203 #else /* !HAVE_AIO_SUPPORT */
 204 static int perf_mmap__aio_enabled(struct mmap *map __maybe_unused)
 205 {
 206         return 0;
 207 }
 208
 209 static int perf_mmap__aio_mmap(struct mmap *map __maybe_unused,
 210                                struct mmap_params *mp __maybe_unused)
 211 {
 212         return 0;
 213 }
 214
 215 static void perf_mmap__aio_munmap(struct mmap *map __maybe_unused)
 216 {
 217 }
 218 #endif
 219
 220 void mmap__munmap(struct mmap *map)
 221 {
 222         bitmap_free(map->affinity_mask.bits);
 223
 224         perf_mmap__aio_munmap(map);
 225         if (map->data != NULL) {
 226                 munmap(map->data, mmap__mmap_len(map));
 227                 map->data = NULL;
 228         }
 229         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 230 }
 231
 232 static void build_node_mask(int node, struct mmap_cpu_mask *mask)
 233 {
 234         int c, cpu, nr_cpus;
 235         const struct perf_cpu_map *cpu_map = NULL;
 236
 237         cpu_map = cpu_map__online();
 238         if (!cpu_map)
 239                 return;
 240
 241         nr_cpus = perf_cpu_map__nr(cpu_map);
 242         for (c = 0; c < nr_cpus; c++) {
 243                 cpu = cpu_map->map[c]; /* map c index to online cpu index */
 244                 if (cpu__get_node(cpu) == node)
 245                         set_bit(cpu, mask->bits);
 246         }
 247 }
 248
 249 static int perf_mmap__setup_affinity_mask(struct mmap *map, struct mmap_params *mp)
 250 {
 251         map->affinity_mask.nbits = cpu__max_cpu();
 252         map->affinity_mask.bits = bitmap_alloc(map->affinity_mask.nbits);
 253         if (!map->affinity_mask.bits)
 254                 return -1;
 255
 256         if (mp->affinity == PERF_AFFINITY_NODE && cpu__max_node() > 1)
 257                 build_node_mask(cpu__get_node(map->core.cpu), &map->affinity_mask);
 258         else if (mp->affinity == PERF_AFFINITY_CPU)
 259                 set_bit(map->core.cpu, map->affinity_mask.bits);
 260
 261         return 0;
 262 }
 263
 264 int mmap__mmap(struct mmap *map, struct mmap_params *mp, int fd, int cpu)
 265 {
 266         if (perf_mmap__mmap(&map->core, &mp->core, fd, cpu)) {
 267                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 268                           errno);
 269                 return -1;
 270         }
 271
 272         if (mp->affinity != PERF_AFFINITY_SYS &&
 273                 perf_mmap__setup_affinity_mask(map, mp)) {
 274                 pr_debug2("failed to alloc mmap affinity mask, error %d\n",
 275                           errno);
 276                 return -1;
 277         }
 278
 279         if (verbose == 2)
 280                 mmap_cpu_mask__scnprintf(&map->affinity_mask, "mmap");
 281
 282         map->core.flush = mp->flush;
 283
 284         map->comp_level = mp->comp_level;
 285
 286         if (map->comp_level && !perf_mmap__aio_enabled(map)) {
 287                 map->data = mmap(NULL, mmap__mmap_len(map), PROT_READ|PROT_WRITE,
 288                                  MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 289                 if (map->data == MAP_FAILED) {
 290                         pr_debug2("failed to mmap data buffer, error %d\n",
 291                                         errno);
 292                         map->data = NULL;
 293                         return -1;
 294                 }
 295         }
 296
 297         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
 298                                 &mp->auxtrace_mp, map->core.base, fd))
 299                 return -1;
 300
 301         return perf_mmap__aio_mmap(map, mp);
 302 }
 303
 304 int perf_mmap__push(struct mmap *md, void *to,
 305                     int push(struct mmap *map, void *to, void *buf, size_t size))
 306 {
 307         u64 head = perf_mmap__read_head(&md->core);
 308         unsigned char *data = md->core.base + page_size;
 309         unsigned long size;
 310         void *buf;
 311         int rc = 0;
 312
 313         rc = perf_mmap__read_init(&md->core);
 314         if (rc < 0)
 315                 return (rc == -EAGAIN) ? 1 : -1;
 316
 317         size = md->core.end - md->core.start;
 318
 319         if ((md->core.start & md->core.mask) + size != (md->core.end & md->core.mask)) {
 320                 buf = &data[md->core.start & md->core.mask];
 321                 size = md->core.mask + 1 - (md->core.start & md->core.mask);
 322                 md->core.start += size;
 323
 324                 if (push(md, to, buf, size) < 0) {
 325                         rc = -1;
 326                         goto out;
 327                 }
 328         }
 329
 330         buf = &data[md->core.start & md->core.mask];
 331         size = md->core.end - md->core.start;
 332         md->core.start += size;
 333
 334         if (push(md, to, buf, size) < 0) {
 335                 rc = -1;
 336                 goto out;
 337         }
 338
 339         md->core.prev = head;
 340         perf_mmap__consume(&md->core);
 341 out:
 342         return rc;
 343 }