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 start, u64 end, u64 *prev)
  26 {
  27         unsigned char *data = map->base + page_size;
  28         union perf_event *event = NULL;
  29         int diff = end - start;
  30
  31         if (diff >= (int)sizeof(event->header)) {
  32                 size_t size;
  33
  34                 event = (union perf_event *)&data[start & map->mask];
  35                 size = event->header.size;
  36
  37                 if (size < sizeof(event->header) || diff < (int)size) {
  38                         event = NULL;
  39                         goto broken_event;
  40                 }
  41
  42                 /*
  43                  * Event straddles the mmap boundary -- header should always
  44                  * be inside due to u64 alignment of output.
  45                  */
  46                 if ((start & map->mask) + size != ((start + size) & map->mask)) {
  47                         unsigned int offset = start;
  48                         unsigned int len = min(sizeof(*event), size), cpy;
  49                         void *dst = map->event_copy;
  50
  51                         do {
  52                                 cpy = min(map->mask + 1 - (offset & map->mask), len);
  53                                 memcpy(dst, &data[offset & map->mask], cpy);
  54                                 offset += cpy;
  55                                 dst += cpy;
  56                                 len -= cpy;
  57                         } while (len);
  58
  59                         event = (union perf_event *)map->event_copy;
  60                 }
  61
  62                 start += size;
  63         }
  64
  65 broken_event:
  66         if (prev)
  67                 *prev = start;
  68
  69         return event;
  70 }
  71
  72 union perf_event *perf_mmap__read_forward(struct perf_mmap *map)
  73 {
  74         u64 head;
  75         u64 old = map->prev;
  76
  77         /*
  78          * Check if event was unmapped due to a POLLHUP/POLLERR.
  79          */
  80         if (!refcount_read(&map->refcnt))
  81                 return NULL;
  82
  83         head = perf_mmap__read_head(map);
  84
  85         return perf_mmap__read(map, old, head, &map->prev);
  86 }
  87
  88 union perf_event *perf_mmap__read_backward(struct perf_mmap *map)
  89 {
  90         u64 head, end;
  91         u64 start = map->prev;
  92
  93         /*
  94          * Check if event was unmapped due to a POLLHUP/POLLERR.
  95          */
  96         if (!refcount_read(&map->refcnt))
  97                 return NULL;
  98
  99         head = perf_mmap__read_head(map);
 100         if (!head)
 101                 return NULL;
 102
 103         /*
 104          * 'head' pointer starts from 0. Kernel minus sizeof(record) form
 105          * it each time when kernel writes to it, so in fact 'head' is
 106          * negative. 'end' pointer is made manually by adding the size of
 107          * the ring buffer to 'head' pointer, means the validate data can
 108          * read is the whole ring buffer. If 'end' is positive, the ring
 109          * buffer has not fully filled, so we must adjust 'end' to 0.
 110          *
 111          * However, since both 'head' and 'end' is unsigned, we can't
 112          * simply compare 'end' against 0. Here we compare '-head' and
 113          * the size of the ring buffer, where -head is the number of bytes
 114          * kernel write to the ring buffer.
 115          */
 116         if (-head < (u64)(map->mask + 1))
 117                 end = 0;
 118         else
 119                 end = head + map->mask + 1;
 120
 121         return perf_mmap__read(map, start, end, &map->prev);
 122 }
 123
 124 void perf_mmap__read_catchup(struct perf_mmap *map)
 125 {
 126         u64 head;
 127
 128         if (!refcount_read(&map->refcnt))
 129                 return;
 130
 131         head = perf_mmap__read_head(map);
 132         map->prev = head;
 133 }
 134
 135 static bool perf_mmap__empty(struct perf_mmap *map)
 136 {
 137         return perf_mmap__read_head(map) == map->prev && !map->auxtrace_mmap.base;
 138 }
 139
 140 void perf_mmap__get(struct perf_mmap *map)
 141 {
 142         refcount_inc(&map->refcnt);
 143 }
 144
 145 void perf_mmap__put(struct perf_mmap *map)
 146 {
 147         BUG_ON(map->base && refcount_read(&map->refcnt) == 0);
 148
 149         if (refcount_dec_and_test(&map->refcnt))
 150                 perf_mmap__munmap(map);
 151 }
 152
 153 void perf_mmap__consume(struct perf_mmap *map, bool overwrite)
 154 {
 155         if (!overwrite) {
 156                 u64 old = map->prev;
 157
 158                 perf_mmap__write_tail(map, old);
 159         }
 160
 161         if (refcount_read(&map->refcnt) == 1 && perf_mmap__empty(map))
 162                 perf_mmap__put(map);
 163 }
 164
 165 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 166                                struct auxtrace_mmap_params *mp __maybe_unused,
 167                                void *userpg __maybe_unused,
 168                                int fd __maybe_unused)
 169 {
 170         return 0;
 171 }
 172
 173 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 174 {
 175 }
 176
 177 void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
 178                                        off_t auxtrace_offset __maybe_unused,
 179                                        unsigned int auxtrace_pages __maybe_unused,
 180                                        bool auxtrace_overwrite __maybe_unused)
 181 {
 182 }
 183
 184 void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
 185                                           struct perf_evlist *evlist __maybe_unused,
 186                                           int idx __maybe_unused,
 187                                           bool per_cpu __maybe_unused)
 188 {
 189 }
 190
 191 void perf_mmap__munmap(struct perf_mmap *map)
 192 {
 193         if (map->base != NULL) {
 194                 munmap(map->base, perf_mmap__mmap_len(map));
 195                 map->base = NULL;
 196                 map->fd = -1;
 197                 refcount_set(&map->refcnt, 0);
 198         }
 199         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 200 }
 201
 202 int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
 203 {
 204         /*
 205          * The last one will be done at perf_evlist__mmap_consume(), so that we
 206          * make sure we don't prevent tools from consuming every last event in
 207          * the ring buffer.
 208          *
 209          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 210          * anymore, but the last events for it are still in the ring buffer,
 211          * waiting to be consumed.
 212          *
 213          * Tools can chose to ignore this at their own discretion, but the
 214          * evlist layer can't just drop it when filtering events in
 215          * perf_evlist__filter_pollfd().
 216          */
 217         refcount_set(&map->refcnt, 2);
 218         map->prev = 0;
 219         map->mask = mp->mask;
 220         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
 221                          MAP_SHARED, fd, 0);
 222         if (map->base == MAP_FAILED) {
 223                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 224                           errno);
 225                 map->base = NULL;
 226                 return -1;
 227         }
 228         map->fd = fd;
 229
 230         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
 231                                 &mp->auxtrace_mp, map->base, fd))
 232                 return -1;
 233
 234         return 0;
 235 }
 236
 237 static int overwrite_rb_find_range(void *buf, int mask, u64 head, u64 *start, u64 *end)
 238 {
 239         struct perf_event_header *pheader;
 240         u64 evt_head = head;
 241         int size = mask + 1;
 242
 243         pr_debug2("overwrite_rb_find_range: buf=%p, head=%"PRIx64"\n", buf, head);
 244         pheader = (struct perf_event_header *)(buf + (head & mask));
 245         *start = head;
 246         while (true) {
 247                 if (evt_head - head >= (unsigned int)size) {
 248                         pr_debug("Finished reading overwrite ring buffer: rewind\n");
 249                         if (evt_head - head > (unsigned int)size)
 250                                 evt_head -= pheader->size;
 251                         *end = evt_head;
 252                         return 0;
 253                 }
 254
 255                 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
 256
 257                 if (pheader->size == 0) {
 258                         pr_debug("Finished reading overwrite ring buffer: get start\n");
 259                         *end = evt_head;
 260                         return 0;
 261                 }
 262
 263                 evt_head += pheader->size;
 264                 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
 265         }
 266         WARN_ONCE(1, "Shouldn't get here\n");
 267         return -1;
 268 }
 269
 270 int perf_mmap__push(struct perf_mmap *md, bool overwrite,
 271                     void *to, int push(void *to, void *buf, size_t size))
 272 {
 273         u64 head = perf_mmap__read_head(md);
 274         u64 old = md->prev;
 275         u64 end = head, start = old;
 276         unsigned char *data = md->base + page_size;
 277         unsigned long size;
 278         void *buf;
 279         int rc = 0;
 280
 281         start = overwrite ? head : old;
 282         end = overwrite ? old : head;
 283
 284         if (start == end)
 285                 return 0;
 286
 287         size = end - start;
 288         if (size > (unsigned long)(md->mask) + 1) {
 289                 if (!overwrite) {
 290                         WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
 291
 292                         md->prev = head;
 293                         perf_mmap__consume(md, overwrite);
 294                         return 0;
 295                 }
 296
 297                 /*
 298                  * Backward ring buffer is full. We still have a chance to read
 299                  * most of data from it.
 300                  */
 301                 if (overwrite_rb_find_range(data, md->mask, head, &start, &end))
 302                         return -1;
 303         }
 304
 305         if ((start & md->mask) + size != (end & md->mask)) {
 306                 buf = &data[start & md->mask];
 307                 size = md->mask + 1 - (start & md->mask);
 308                 start += size;
 309
 310                 if (push(to, buf, size) < 0) {
 311                         rc = -1;
 312                         goto out;
 313                 }
 314         }
 315
 316         buf = &data[start & md->mask];
 317         size = end - start;
 318         start += size;
 319
 320         if (push(to, buf, size) < 0) {
 321                 rc = -1;
 322                 goto out;
 323         }
 324
 325         md->prev = head;
 326         perf_mmap__consume(md, overwrite);
 327 out:
 328         return rc;
 329 }