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, bool check_messup, 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 (check_messup) { 32 /* 33 * If we're further behind than half the buffer, there's a chance 34 * the writer will bite our tail and mess up the samples under us. 35 * 36 * If we somehow ended up ahead of the 'end', we got messed up. 37 * 38 * In either case, truncate and restart at 'end'. 39 */ 40 if (diff > map->mask / 2 || diff < 0) { 41 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n"); 42 43 /* 44 * 'end' points to a known good entry, start there. 45 */ 46 start = end; 47 diff = 0; 48 } 49 } 50 51 if (diff >= (int)sizeof(event->header)) { 52 size_t size; 53 54 event = (union perf_event *)&data[start & map->mask]; 55 size = event->header.size; 56 57 if (size < sizeof(event->header) || diff < (int)size) { 58 event = NULL; 59 goto broken_event; 60 } 61 62 /* 63 * Event straddles the mmap boundary -- header should always 64 * be inside due to u64 alignment of output. 65 */ 66 if ((start & map->mask) + size != ((start + size) & map->mask)) { 67 unsigned int offset = start; 68 unsigned int len = min(sizeof(*event), size), cpy; 69 void *dst = map->event_copy; 70 71 do { 72 cpy = min(map->mask + 1 - (offset & map->mask), len); 73 memcpy(dst, &data[offset & map->mask], cpy); 74 offset += cpy; 75 dst += cpy; 76 len -= cpy; 77 } while (len); 78 79 event = (union perf_event *)map->event_copy; 80 } 81 82 start += size; 83 } 84 85 broken_event: 86 if (prev) 87 *prev = start; 88 89 return event; 90 } 91 92 union perf_event *perf_mmap__read_forward(struct perf_mmap *map, bool check_messup) 93 { 94 u64 head; 95 u64 old = map->prev; 96 97 /* 98 * Check if event was unmapped due to a POLLHUP/POLLERR. 99 */ 100 if (!refcount_read(&map->refcnt)) 101 return NULL; 102 103 head = perf_mmap__read_head(map); 104 105 return perf_mmap__read(map, check_messup, old, head, &map->prev); 106 } 107 108 union perf_event *perf_mmap__read_backward(struct perf_mmap *map) 109 { 110 u64 head, end; 111 u64 start = map->prev; 112 113 /* 114 * Check if event was unmapped due to a POLLHUP/POLLERR. 115 */ 116 if (!refcount_read(&map->refcnt)) 117 return NULL; 118 119 head = perf_mmap__read_head(map); 120 if (!head) 121 return NULL; 122 123 /* 124 * 'head' pointer starts from 0. Kernel minus sizeof(record) form 125 * it each time when kernel writes to it, so in fact 'head' is 126 * negative. 'end' pointer is made manually by adding the size of 127 * the ring buffer to 'head' pointer, means the validate data can 128 * read is the whole ring buffer. If 'end' is positive, the ring 129 * buffer has not fully filled, so we must adjust 'end' to 0. 130 * 131 * However, since both 'head' and 'end' is unsigned, we can't 132 * simply compare 'end' against 0. Here we compare '-head' and 133 * the size of the ring buffer, where -head is the number of bytes 134 * kernel write to the ring buffer. 135 */ 136 if (-head < (u64)(map->mask + 1)) 137 end = 0; 138 else 139 end = head + map->mask + 1; 140 141 return perf_mmap__read(map, false, start, end, &map->prev); 142 } 143 144 void perf_mmap__read_catchup(struct perf_mmap *map) 145 { 146 u64 head; 147 148 if (!refcount_read(&map->refcnt)) 149 return; 150 151 head = perf_mmap__read_head(map); 152 map->prev = head; 153 } 154 155 static bool perf_mmap__empty(struct perf_mmap *map) 156 { 157 return perf_mmap__read_head(map) == map->prev && !map->auxtrace_mmap.base; 158 } 159 160 void perf_mmap__get(struct perf_mmap *map) 161 { 162 refcount_inc(&map->refcnt); 163 } 164 165 void perf_mmap__put(struct perf_mmap *map) 166 { 167 BUG_ON(map->base && refcount_read(&map->refcnt) == 0); 168 169 if (refcount_dec_and_test(&map->refcnt)) 170 perf_mmap__munmap(map); 171 } 172 173 void perf_mmap__consume(struct perf_mmap *map, bool overwrite) 174 { 175 if (!overwrite) { 176 u64 old = map->prev; 177 178 perf_mmap__write_tail(map, old); 179 } 180 181 if (refcount_read(&map->refcnt) == 1 && perf_mmap__empty(map)) 182 perf_mmap__put(map); 183 } 184 185 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused, 186 struct auxtrace_mmap_params *mp __maybe_unused, 187 void *userpg __maybe_unused, 188 int fd __maybe_unused) 189 { 190 return 0; 191 } 192 193 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused) 194 { 195 } 196 197 void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused, 198 off_t auxtrace_offset __maybe_unused, 199 unsigned int auxtrace_pages __maybe_unused, 200 bool auxtrace_overwrite __maybe_unused) 201 { 202 } 203 204 void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused, 205 struct perf_evlist *evlist __maybe_unused, 206 int idx __maybe_unused, 207 bool per_cpu __maybe_unused) 208 { 209 } 210 211 void perf_mmap__munmap(struct perf_mmap *map) 212 { 213 if (map->base != NULL) { 214 munmap(map->base, perf_mmap__mmap_len(map)); 215 map->base = NULL; 216 map->fd = -1; 217 refcount_set(&map->refcnt, 0); 218 } 219 auxtrace_mmap__munmap(&map->auxtrace_mmap); 220 } 221 222 int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd) 223 { 224 /* 225 * The last one will be done at perf_evlist__mmap_consume(), so that we 226 * make sure we don't prevent tools from consuming every last event in 227 * the ring buffer. 228 * 229 * I.e. we can get the POLLHUP meaning that the fd doesn't exist 230 * anymore, but the last events for it are still in the ring buffer, 231 * waiting to be consumed. 232 * 233 * Tools can chose to ignore this at their own discretion, but the 234 * evlist layer can't just drop it when filtering events in 235 * perf_evlist__filter_pollfd(). 236 */ 237 refcount_set(&map->refcnt, 2); 238 map->prev = 0; 239 map->mask = mp->mask; 240 map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot, 241 MAP_SHARED, fd, 0); 242 if (map->base == MAP_FAILED) { 243 pr_debug2("failed to mmap perf event ring buffer, error %d\n", 244 errno); 245 map->base = NULL; 246 return -1; 247 } 248 map->fd = fd; 249 250 if (auxtrace_mmap__mmap(&map->auxtrace_mmap, 251 &mp->auxtrace_mp, map->base, fd)) 252 return -1; 253 254 return 0; 255 } 256 257 static int backward_rb_find_range(void *buf, int mask, u64 head, u64 *start, u64 *end) 258 { 259 struct perf_event_header *pheader; 260 u64 evt_head = head; 261 int size = mask + 1; 262 263 pr_debug2("backward_rb_find_range: buf=%p, head=%"PRIx64"\n", buf, head); 264 pheader = (struct perf_event_header *)(buf + (head & mask)); 265 *start = head; 266 while (true) { 267 if (evt_head - head >= (unsigned int)size) { 268 pr_debug("Finished reading backward ring buffer: rewind\n"); 269 if (evt_head - head > (unsigned int)size) 270 evt_head -= pheader->size; 271 *end = evt_head; 272 return 0; 273 } 274 275 pheader = (struct perf_event_header *)(buf + (evt_head & mask)); 276 277 if (pheader->size == 0) { 278 pr_debug("Finished reading backward ring buffer: get start\n"); 279 *end = evt_head; 280 return 0; 281 } 282 283 evt_head += pheader->size; 284 pr_debug3("move evt_head: %"PRIx64"\n", evt_head); 285 } 286 WARN_ONCE(1, "Shouldn't get here\n"); 287 return -1; 288 } 289 290 static int rb_find_range(void *data, int mask, u64 head, u64 old, 291 u64 *start, u64 *end, bool backward) 292 { 293 if (!backward) { 294 *start = old; 295 *end = head; 296 return 0; 297 } 298 299 return backward_rb_find_range(data, mask, head, start, end); 300 } 301 302 int perf_mmap__push(struct perf_mmap *md, bool overwrite, bool backward, 303 void *to, int push(void *to, void *buf, size_t size)) 304 { 305 u64 head = perf_mmap__read_head(md); 306 u64 old = md->prev; 307 u64 end = head, start = old; 308 unsigned char *data = md->base + page_size; 309 unsigned long size; 310 void *buf; 311 int rc = 0; 312 313 if (rb_find_range(data, md->mask, head, old, &start, &end, backward)) 314 return -1; 315 316 if (start == end) 317 return 0; 318 319 size = end - start; 320 if (size > (unsigned long)(md->mask) + 1) { 321 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n"); 322 323 md->prev = head; 324 perf_mmap__consume(md, overwrite || backward); 325 return 0; 326 } 327 328 if ((start & md->mask) + size != (end & md->mask)) { 329 buf = &data[start & md->mask]; 330 size = md->mask + 1 - (start & md->mask); 331 start += size; 332 333 if (push(to, buf, size) < 0) { 334 rc = -1; 335 goto out; 336 } 337 } 338 339 buf = &data[start & md->mask]; 340 size = end - start; 341 start += size; 342 343 if (push(to, buf, size) < 0) { 344 rc = -1; 345 goto out; 346 } 347 348 md->prev = head; 349 perf_mmap__consume(md, overwrite || backward); 350 out: 351 return rc; 352 } 353