tools/perf/util/mmap.c

   1 /*
   2  * Copyright (C) 2011-2017, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from evlist.c builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <sys/mman.h>
  11 #include <inttypes.h>
  12 #include <asm/bug.h>
  13 #include "debug.h"
  14 #include "event.h"
  15 #include "mmap.h"
  16 #include "util.h" /* page_size */
  17
  18 size_t perf_mmap__mmap_len(struct perf_mmap *map)
  19 {
  20         return map->mask + 1 + page_size;
  21 }
  22
  23 /* When check_messup is true, 'end' must points to a good entry */
  24 static union perf_event *perf_mmap__read(struct perf_mmap *map,
  25                                          u64 *startp, u64 end)
  26 {
  27         unsigned char *data = map->base + page_size;
  28         union perf_event *event = NULL;
  29         int diff = end - *startp;
  30
  31         if (diff >= (int)sizeof(event->header)) {
  32                 size_t size;
  33
  34                 event = (union perf_event *)&data[*startp & map->mask];
  35                 size = event->header.size;
  36
  37                 if (size < sizeof(event->header) || diff < (int)size)
  38                         return NULL;
  39
  40                 /*
  41                  * Event straddles the mmap boundary -- header should always
  42                  * be inside due to u64 alignment of output.
  43                  */
  44                 if ((*startp & map->mask) + size != ((*startp + size) & map->mask)) {
  45                         unsigned int offset = *startp;
  46                         unsigned int len = min(sizeof(*event), size), cpy;
  47                         void *dst = map->event_copy;
  48
  49                         do {
  50                                 cpy = min(map->mask + 1 - (offset & map->mask), len);
  51                                 memcpy(dst, &data[offset & map->mask], cpy);
  52                                 offset += cpy;
  53                                 dst += cpy;
  54                                 len -= cpy;
  55                         } while (len);
  56
  57                         event = (union perf_event *)map->event_copy;
  58                 }
  59
  60                 *startp += size;
  61         }
  62
  63         return event;
  64 }
  65
  66 /*
  67  * legacy interface for mmap read.
  68  * Don't use it. Use perf_mmap__read_event().
  69  */
  70 union perf_event *perf_mmap__read_forward(struct perf_mmap *map)
  71 {
  72         u64 head;
  73
  74         /*
  75          * Check if event was unmapped due to a POLLHUP/POLLERR.
  76          */
  77         if (!refcount_read(&map->refcnt))
  78                 return NULL;
  79
  80         head = perf_mmap__read_head(map);
  81
  82         return perf_mmap__read(map, &map->prev, head);
  83 }
  84
  85 /*
  86  * Read event from ring buffer one by one.
  87  * Return one event for each call.
  88  *
  89  * Usage:
  90  * perf_mmap__read_init()
  91  * while(event = perf_mmap__read_event()) {
  92  *      //process the event
  93  *      perf_mmap__consume()
  94  * }
  95  * perf_mmap__read_done()
  96  */
  97 union perf_event *perf_mmap__read_event(struct perf_mmap *map,
  98                                         bool overwrite,
  99                                         u64 *startp, u64 end)
 100 {
 101         union perf_event *event;
 102
 103         /*
 104          * Check if event was unmapped due to a POLLHUP/POLLERR.
 105          */
 106         if (!refcount_read(&map->refcnt))
 107                 return NULL;
 108
 109         if (startp == NULL)
 110                 return NULL;
 111
 112         /* non-overwirte doesn't pause the ringbuffer */
 113         if (!overwrite)
 114                 end = perf_mmap__read_head(map);
 115
 116         event = perf_mmap__read(map, startp, end);
 117
 118         if (!overwrite)
 119                 map->prev = *startp;
 120
 121         return event;
 122 }
 123
 124 static bool perf_mmap__empty(struct perf_mmap *map)
 125 {
 126         return perf_mmap__read_head(map) == map->prev && !map->auxtrace_mmap.base;
 127 }
 128
 129 void perf_mmap__get(struct perf_mmap *map)
 130 {
 131         refcount_inc(&map->refcnt);
 132 }
 133
 134 void perf_mmap__put(struct perf_mmap *map)
 135 {
 136         BUG_ON(map->base && refcount_read(&map->refcnt) == 0);
 137
 138         if (refcount_dec_and_test(&map->refcnt))
 139                 perf_mmap__munmap(map);
 140 }
 141
 142 void perf_mmap__consume(struct perf_mmap *map, bool overwrite)
 143 {
 144         if (!overwrite) {
 145                 u64 old = map->prev;
 146
 147                 perf_mmap__write_tail(map, old);
 148         }
 149
 150         if (refcount_read(&map->refcnt) == 1 && perf_mmap__empty(map))
 151                 perf_mmap__put(map);
 152 }
 153
 154 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 155                                struct auxtrace_mmap_params *mp __maybe_unused,
 156                                void *userpg __maybe_unused,
 157                                int fd __maybe_unused)
 158 {
 159         return 0;
 160 }
 161
 162 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 163 {
 164 }
 165
 166 void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
 167                                        off_t auxtrace_offset __maybe_unused,
 168                                        unsigned int auxtrace_pages __maybe_unused,
 169                                        bool auxtrace_overwrite __maybe_unused)
 170 {
 171 }
 172
 173 void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
 174                                           struct perf_evlist *evlist __maybe_unused,
 175                                           int idx __maybe_unused,
 176                                           bool per_cpu __maybe_unused)
 177 {
 178 }
 179
 180 void perf_mmap__munmap(struct perf_mmap *map)
 181 {
 182         if (map->base != NULL) {
 183                 munmap(map->base, perf_mmap__mmap_len(map));
 184                 map->base = NULL;
 185                 map->fd = -1;
 186                 refcount_set(&map->refcnt, 0);
 187         }
 188         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 189 }
 190
 191 int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
 192 {
 193         /*
 194          * The last one will be done at perf_evlist__mmap_consume(), so that we
 195          * make sure we don't prevent tools from consuming every last event in
 196          * the ring buffer.
 197          *
 198          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 199          * anymore, but the last events for it are still in the ring buffer,
 200          * waiting to be consumed.
 201          *
 202          * Tools can chose to ignore this at their own discretion, but the
 203          * evlist layer can't just drop it when filtering events in
 204          * perf_evlist__filter_pollfd().
 205          */
 206         refcount_set(&map->refcnt, 2);
 207         map->prev = 0;
 208         map->mask = mp->mask;
 209         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
 210                          MAP_SHARED, fd, 0);
 211         if (map->base == MAP_FAILED) {
 212                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 213                           errno);
 214                 map->base = NULL;
 215                 return -1;
 216         }
 217         map->fd = fd;
 218
 219         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
 220                                 &mp->auxtrace_mp, map->base, fd))
 221                 return -1;
 222
 223         return 0;
 224 }
 225
 226 static int overwrite_rb_find_range(void *buf, int mask, u64 head, u64 *start, u64 *end)
 227 {
 228         struct perf_event_header *pheader;
 229         u64 evt_head = head;
 230         int size = mask + 1;
 231
 232         pr_debug2("overwrite_rb_find_range: buf=%p, head=%"PRIx64"\n", buf, head);
 233         pheader = (struct perf_event_header *)(buf + (head & mask));
 234         *start = head;
 235         while (true) {
 236                 if (evt_head - head >= (unsigned int)size) {
 237                         pr_debug("Finished reading overwrite ring buffer: rewind\n");
 238                         if (evt_head - head > (unsigned int)size)
 239                                 evt_head -= pheader->size;
 240                         *end = evt_head;
 241                         return 0;
 242                 }
 243
 244                 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
 245
 246                 if (pheader->size == 0) {
 247                         pr_debug("Finished reading overwrite ring buffer: get start\n");
 248                         *end = evt_head;
 249                         return 0;
 250                 }
 251
 252                 evt_head += pheader->size;
 253                 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
 254         }
 255         WARN_ONCE(1, "Shouldn't get here\n");
 256         return -1;
 257 }
 258
 259 /*
 260  * Report the start and end of the available data in ringbuffer
 261  */
 262 int perf_mmap__read_init(struct perf_mmap *md, bool overwrite,
 263                          u64 *startp, u64 *endp)
 264 {
 265         u64 head = perf_mmap__read_head(md);
 266         u64 old = md->prev;
 267         unsigned char *data = md->base + page_size;
 268         unsigned long size;
 269
 270         *startp = overwrite ? head : old;
 271         *endp = overwrite ? old : head;
 272
 273         if (*startp == *endp)
 274                 return -EAGAIN;
 275
 276         size = *endp - *startp;
 277         if (size > (unsigned long)(md->mask) + 1) {
 278                 if (!overwrite) {
 279                         WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
 280
 281                         md->prev = head;
 282                         perf_mmap__consume(md, overwrite);
 283                         return -EAGAIN;
 284                 }
 285
 286                 /*
 287                  * Backward ring buffer is full. We still have a chance to read
 288                  * most of data from it.
 289                  */
 290                 if (overwrite_rb_find_range(data, md->mask, head, startp, endp))
 291                         return -EINVAL;
 292         }
 293
 294         return 0;
 295 }
 296
 297 int perf_mmap__push(struct perf_mmap *md, bool overwrite,
 298                     void *to, int push(void *to, void *buf, size_t size))
 299 {
 300         u64 head = perf_mmap__read_head(md);
 301         u64 end, start;
 302         unsigned char *data = md->base + page_size;
 303         unsigned long size;
 304         void *buf;
 305         int rc = 0;
 306
 307         rc = perf_mmap__read_init(md, overwrite, &start, &end);
 308         if (rc < 0)
 309                 return (rc == -EAGAIN) ? 0 : -1;
 310
 311         size = end - start;
 312
 313         if ((start & md->mask) + size != (end & md->mask)) {
 314                 buf = &data[start & md->mask];
 315                 size = md->mask + 1 - (start & md->mask);
 316                 start += size;
 317
 318                 if (push(to, buf, size) < 0) {
 319                         rc = -1;
 320                         goto out;
 321                 }
 322         }
 323
 324         buf = &data[start & md->mask];
 325         size = end - start;
 326         start += size;
 327
 328         if (push(to, buf, size) < 0) {
 329                 rc = -1;
 330                 goto out;
 331         }
 332
 333         md->prev = head;
 334         perf_mmap__consume(md, overwrite);
 335 out:
 336         return rc;
 337 }
 338
 339 /*
 340  * Mandatory for overwrite mode
 341  * The direction of overwrite mode is backward.
 342  * The last perf_mmap__read() will set tail to map->prev.
 343  * Need to correct the map->prev to head which is the end of next read.
 344  */
 345 void perf_mmap__read_done(struct perf_mmap *map)
 346 {
 347         map->prev = perf_mmap__read_head(map);
 348 }