Commit | Line | Data |
---|---|---|
1d498196 | 1 | /* |
c0c0989a | 2 | * SPDX-License-Identifier: LGPL-2.1-only |
1d498196 | 3 | * |
e92f3e28 | 4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
1d498196 MD |
5 | */ |
6 | ||
3fbec7dc | 7 | #define _LGPL_SOURCE |
1d498196 MD |
8 | #include "shm.h" |
9 | #include <unistd.h> | |
10 | #include <fcntl.h> | |
11 | #include <sys/mman.h> | |
a9ff648c | 12 | #include <sys/types.h> |
1d498196 MD |
13 | #include <sys/stat.h> /* For mode constants */ |
14 | #include <fcntl.h> /* For O_* constants */ | |
15 | #include <assert.h> | |
8da6cd6d MD |
16 | #include <stdio.h> |
17 | #include <signal.h> | |
18 | #include <dirent.h> | |
96e80018 | 19 | #include <limits.h> |
8a208943 | 20 | #include <stdbool.h> |
fb31eb73 | 21 | #include <stdint.h> |
3d3a2bb8 | 22 | |
bfcda6ce | 23 | #ifdef HAVE_LIBNUMA |
4b68c31f | 24 | #include <numa.h> |
8a208943 | 25 | #include <numaif.h> |
bfcda6ce | 26 | #endif |
3d3a2bb8 | 27 | |
eae3c729 | 28 | #include <lttng/ust-utils.h> |
3d3a2bb8 | 29 | |
9d315d6d MJ |
30 | #include "common/macros.h" |
31 | #include "common/ust-fd.h" | |
1be43539 | 32 | #include "common/compat/mmap.h" |
3a81f31d MD |
33 | |
34 | /* | |
35 | * Ensure we have the required amount of space available by writing 0 | |
36 | * into the entire buffer. Not doing so can trigger SIGBUS when going | |
37 | * beyond the available shm space. | |
38 | */ | |
39 | static | |
40 | int zero_file(int fd, size_t len) | |
41 | { | |
42 | ssize_t retlen; | |
43 | size_t written = 0; | |
44 | char *zeropage; | |
45 | long pagelen; | |
46 | int ret; | |
47 | ||
48 | pagelen = sysconf(_SC_PAGESIZE); | |
49 | if (pagelen < 0) | |
50 | return (int) pagelen; | |
51 | zeropage = calloc(pagelen, 1); | |
52 | if (!zeropage) | |
53 | return -ENOMEM; | |
54 | ||
55 | while (len > written) { | |
56 | do { | |
57 | retlen = write(fd, zeropage, | |
58 | min_t(size_t, pagelen, len - written)); | |
59 | } while (retlen == -1UL && errno == EINTR); | |
60 | if (retlen < 0) { | |
61 | ret = (int) retlen; | |
62 | goto error; | |
63 | } | |
64 | written += retlen; | |
65 | } | |
66 | ret = 0; | |
67 | error: | |
68 | free(zeropage); | |
69 | return ret; | |
70 | } | |
1d498196 | 71 | |
97572c04 | 72 | struct shm_object_table *shm_object_table_create(size_t max_nb_obj, bool populate) |
1d498196 MD |
73 | { |
74 | struct shm_object_table *table; | |
75 | ||
97572c04 MD |
76 | table = zmalloc_populate(sizeof(struct shm_object_table) + |
77 | max_nb_obj * sizeof(table->objects[0]), populate); | |
74d48abe MD |
78 | if (!table) |
79 | return NULL; | |
1d498196 MD |
80 | table->size = max_nb_obj; |
81 | return table; | |
82 | } | |
83 | ||
74d81a6c MD |
84 | static |
85 | struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table, | |
a9ff648c | 86 | size_t memory_map_size, |
97572c04 MD |
87 | int stream_fd, |
88 | bool populate) | |
1d498196 | 89 | { |
5ea386c3 | 90 | int shmfd, waitfd[2], ret, i; |
97572c04 | 91 | int flags = MAP_SHARED; |
1d498196 MD |
92 | struct shm_object *obj; |
93 | char *memory_map; | |
94 | ||
5ea386c3 MD |
95 | if (stream_fd < 0) |
96 | return NULL; | |
1d498196 MD |
97 | if (table->allocated_len >= table->size) |
98 | return NULL; | |
7a9c21bd | 99 | obj = &table->objects[table->allocated_len]; |
1d498196 MD |
100 | |
101 | /* wait_fd: create pipe */ | |
8a8c2117 | 102 | ret = pipe2(waitfd, O_CLOEXEC); |
1d498196 MD |
103 | if (ret < 0) { |
104 | PERROR("pipe"); | |
105 | goto error_pipe; | |
106 | } | |
5d61a504 MD |
107 | /* The write end of the pipe needs to be non-blocking */ |
108 | ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK); | |
109 | if (ret < 0) { | |
110 | PERROR("fcntl"); | |
111 | goto error_fcntl; | |
112 | } | |
7a9c21bd | 113 | memcpy(obj->wait_fd, waitfd, sizeof(waitfd)); |
1d498196 | 114 | |
053e6e24 MJ |
115 | /* |
116 | * Set POSIX shared memory object size | |
117 | * | |
118 | * First, use ftruncate() to set its size, some implementations won't | |
119 | * allow writes past the size set by ftruncate. | |
120 | * Then, use write() to fill it with zeros, this allows us to fully | |
121 | * allocate it and detect a shortage of shm space without dealing with | |
122 | * a SIGBUS. | |
123 | */ | |
a9ff648c | 124 | |
5ea386c3 | 125 | shmfd = stream_fd; |
1d498196 MD |
126 | ret = ftruncate(shmfd, memory_map_size); |
127 | if (ret) { | |
128 | PERROR("ftruncate"); | |
129 | goto error_ftruncate; | |
130 | } | |
053e6e24 MJ |
131 | ret = zero_file(shmfd, memory_map_size); |
132 | if (ret) { | |
133 | PERROR("zero_file"); | |
134 | goto error_zero_file; | |
135 | } | |
71be0c53 | 136 | |
d0f6cf57 MD |
137 | /* |
138 | * Also ensure the file metadata is synced with the storage by using | |
71be0c53 MJ |
139 | * fsync(2). Some platforms don't allow fsync on POSIX shm fds, ignore |
140 | * EINVAL accordingly. | |
d0f6cf57 MD |
141 | */ |
142 | ret = fsync(shmfd); | |
71be0c53 | 143 | if (ret && errno != EINVAL) { |
d0f6cf57 MD |
144 | PERROR("fsync"); |
145 | goto error_fsync; | |
146 | } | |
5ea386c3 | 147 | obj->shm_fd_ownership = 0; |
1d498196 MD |
148 | obj->shm_fd = shmfd; |
149 | ||
97572c04 MD |
150 | if (populate) |
151 | flags |= LTTNG_MAP_POPULATE; | |
1d498196 MD |
152 | /* memory_map: mmap */ |
153 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, | |
97572c04 | 154 | flags, shmfd, 0); |
1d498196 MD |
155 | if (memory_map == MAP_FAILED) { |
156 | PERROR("mmap"); | |
157 | goto error_mmap; | |
158 | } | |
74d81a6c | 159 | obj->type = SHM_OBJECT_SHM; |
1d498196 MD |
160 | obj->memory_map = memory_map; |
161 | obj->memory_map_size = memory_map_size; | |
162 | obj->allocated_len = 0; | |
dc613eb9 | 163 | obj->index = table->allocated_len++; |
7a9c21bd | 164 | |
1d498196 MD |
165 | return obj; |
166 | ||
167 | error_mmap: | |
d0f6cf57 | 168 | error_fsync: |
1d498196 | 169 | error_ftruncate: |
3a81f31d | 170 | error_zero_file: |
1d498196 MD |
171 | error_fcntl: |
172 | for (i = 0; i < 2; i++) { | |
173 | ret = close(waitfd[i]); | |
174 | if (ret) { | |
175 | PERROR("close"); | |
176 | assert(0); | |
177 | } | |
178 | } | |
179 | error_pipe: | |
1d498196 | 180 | return NULL; |
1d498196 MD |
181 | } |
182 | ||
74d81a6c MD |
183 | static |
184 | struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table, | |
97572c04 | 185 | size_t memory_map_size, bool populate) |
74d81a6c MD |
186 | { |
187 | struct shm_object *obj; | |
188 | void *memory_map; | |
ff0f5728 | 189 | int waitfd[2], i, ret; |
74d81a6c MD |
190 | |
191 | if (table->allocated_len >= table->size) | |
192 | return NULL; | |
193 | obj = &table->objects[table->allocated_len]; | |
194 | ||
97572c04 | 195 | memory_map = zmalloc_populate(memory_map_size, populate); |
74d81a6c MD |
196 | if (!memory_map) |
197 | goto alloc_error; | |
198 | ||
ff0f5728 | 199 | /* wait_fd: create pipe */ |
8a8c2117 | 200 | ret = pipe2(waitfd, O_CLOEXEC); |
ff0f5728 MD |
201 | if (ret < 0) { |
202 | PERROR("pipe"); | |
203 | goto error_pipe; | |
204 | } | |
ff0f5728 MD |
205 | /* The write end of the pipe needs to be non-blocking */ |
206 | ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK); | |
207 | if (ret < 0) { | |
208 | PERROR("fcntl"); | |
209 | goto error_fcntl; | |
210 | } | |
211 | memcpy(obj->wait_fd, waitfd, sizeof(waitfd)); | |
212 | ||
213 | /* no shm_fd */ | |
74d81a6c | 214 | obj->shm_fd = -1; |
5ea386c3 | 215 | obj->shm_fd_ownership = 0; |
74d81a6c MD |
216 | |
217 | obj->type = SHM_OBJECT_MEM; | |
218 | obj->memory_map = memory_map; | |
219 | obj->memory_map_size = memory_map_size; | |
220 | obj->allocated_len = 0; | |
221 | obj->index = table->allocated_len++; | |
222 | ||
223 | return obj; | |
224 | ||
ff0f5728 MD |
225 | error_fcntl: |
226 | for (i = 0; i < 2; i++) { | |
227 | ret = close(waitfd[i]); | |
228 | if (ret) { | |
229 | PERROR("close"); | |
230 | assert(0); | |
231 | } | |
232 | } | |
233 | error_pipe: | |
234 | free(memory_map); | |
74d81a6c MD |
235 | alloc_error: |
236 | return NULL; | |
237 | } | |
238 | ||
8a208943 MD |
239 | /* |
240 | * libnuma prints errors on the console even for numa_available(). | |
241 | * Work-around this limitation by using get_mempolicy() directly to | |
242 | * check whether the kernel supports mempolicy. | |
243 | */ | |
244 | #ifdef HAVE_LIBNUMA | |
245 | static bool lttng_is_numa_available(void) | |
246 | { | |
247 | int ret; | |
248 | ||
249 | ret = get_mempolicy(NULL, NULL, 0, NULL, 0); | |
250 | if (ret && errno == ENOSYS) { | |
251 | return false; | |
252 | } | |
253 | return numa_available() > 0; | |
254 | } | |
255 | #endif | |
256 | ||
c494c0f1 | 257 | #ifdef HAVE_LIBNUMA |
74d81a6c MD |
258 | struct shm_object *shm_object_table_alloc(struct shm_object_table *table, |
259 | size_t memory_map_size, | |
a9ff648c | 260 | enum shm_object_type type, |
4b68c31f | 261 | int stream_fd, |
97572c04 MD |
262 | int cpu, |
263 | bool populate) | |
c494c0f1 MJ |
264 | #else |
265 | struct shm_object *shm_object_table_alloc(struct shm_object_table *table, | |
266 | size_t memory_map_size, | |
267 | enum shm_object_type type, | |
268 | int stream_fd, | |
97572c04 MD |
269 | int cpu __attribute__((unused)), |
270 | bool populate) | |
c494c0f1 | 271 | #endif |
74d81a6c | 272 | { |
4b68c31f | 273 | struct shm_object *shm_object; |
bfcda6ce | 274 | #ifdef HAVE_LIBNUMA |
8a208943 MD |
275 | int oldnode = 0, node; |
276 | bool numa_avail; | |
4b68c31f | 277 | |
8a208943 MD |
278 | numa_avail = lttng_is_numa_available(); |
279 | if (numa_avail) { | |
280 | oldnode = numa_preferred(); | |
281 | if (cpu >= 0) { | |
282 | node = numa_node_of_cpu(cpu); | |
283 | if (node >= 0) | |
284 | numa_set_preferred(node); | |
285 | } | |
286 | if (cpu < 0 || node < 0) | |
287 | numa_set_localalloc(); | |
4b68c31f | 288 | } |
bfcda6ce | 289 | #endif /* HAVE_LIBNUMA */ |
74d81a6c MD |
290 | switch (type) { |
291 | case SHM_OBJECT_SHM: | |
4b68c31f | 292 | shm_object = _shm_object_table_alloc_shm(table, memory_map_size, |
97572c04 | 293 | stream_fd, populate); |
4b68c31f | 294 | break; |
74d81a6c | 295 | case SHM_OBJECT_MEM: |
97572c04 MD |
296 | shm_object = _shm_object_table_alloc_mem(table, memory_map_size, |
297 | populate); | |
4b68c31f | 298 | break; |
74d81a6c MD |
299 | default: |
300 | assert(0); | |
301 | } | |
bfcda6ce | 302 | #ifdef HAVE_LIBNUMA |
8a208943 MD |
303 | if (numa_avail) |
304 | numa_set_preferred(oldnode); | |
bfcda6ce | 305 | #endif /* HAVE_LIBNUMA */ |
4b68c31f | 306 | return shm_object; |
74d81a6c MD |
307 | } |
308 | ||
309 | struct shm_object *shm_object_table_append_shm(struct shm_object_table *table, | |
310 | int shm_fd, int wakeup_fd, uint32_t stream_nr, | |
97572c04 | 311 | size_t memory_map_size, bool populate) |
193183fb | 312 | { |
97572c04 | 313 | int flags = MAP_SHARED; |
193183fb MD |
314 | struct shm_object *obj; |
315 | char *memory_map; | |
74d81a6c | 316 | int ret; |
193183fb MD |
317 | |
318 | if (table->allocated_len >= table->size) | |
319 | return NULL; | |
74d81a6c MD |
320 | /* streams _must_ be received in sequential order, else fail. */ |
321 | if (stream_nr + 1 != table->allocated_len) | |
322 | return NULL; | |
323 | ||
193183fb MD |
324 | obj = &table->objects[table->allocated_len]; |
325 | ||
74d81a6c MD |
326 | /* wait_fd: set write end of the pipe. */ |
327 | obj->wait_fd[0] = -1; /* read end is unset */ | |
328 | obj->wait_fd[1] = wakeup_fd; | |
193183fb | 329 | obj->shm_fd = shm_fd; |
5ea386c3 | 330 | obj->shm_fd_ownership = 1; |
193183fb | 331 | |
74d81a6c MD |
332 | /* The write end of the pipe needs to be non-blocking */ |
333 | ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK); | |
334 | if (ret < 0) { | |
335 | PERROR("fcntl"); | |
336 | goto error_fcntl; | |
337 | } | |
338 | ||
97572c04 MD |
339 | if (populate) |
340 | flags |= LTTNG_MAP_POPULATE; | |
193183fb MD |
341 | /* memory_map: mmap */ |
342 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, | |
97572c04 | 343 | flags, shm_fd, 0); |
193183fb MD |
344 | if (memory_map == MAP_FAILED) { |
345 | PERROR("mmap"); | |
346 | goto error_mmap; | |
347 | } | |
74d81a6c | 348 | obj->type = SHM_OBJECT_SHM; |
193183fb MD |
349 | obj->memory_map = memory_map; |
350 | obj->memory_map_size = memory_map_size; | |
351 | obj->allocated_len = memory_map_size; | |
352 | obj->index = table->allocated_len++; | |
353 | ||
354 | return obj; | |
355 | ||
74d81a6c | 356 | error_fcntl: |
193183fb MD |
357 | error_mmap: |
358 | return NULL; | |
359 | } | |
360 | ||
74d81a6c MD |
361 | /* |
362 | * Passing ownership of mem to object. | |
363 | */ | |
364 | struct shm_object *shm_object_table_append_mem(struct shm_object_table *table, | |
ff0f5728 | 365 | void *mem, size_t memory_map_size, int wakeup_fd) |
74d81a6c MD |
366 | { |
367 | struct shm_object *obj; | |
ff0f5728 | 368 | int ret; |
74d81a6c MD |
369 | |
370 | if (table->allocated_len >= table->size) | |
371 | return NULL; | |
372 | obj = &table->objects[table->allocated_len]; | |
373 | ||
ff0f5728 MD |
374 | obj->wait_fd[0] = -1; /* read end is unset */ |
375 | obj->wait_fd[1] = wakeup_fd; | |
74d81a6c | 376 | obj->shm_fd = -1; |
5ea386c3 | 377 | obj->shm_fd_ownership = 0; |
74d81a6c | 378 | |
ff0f5728 MD |
379 | /* The write end of the pipe needs to be non-blocking */ |
380 | ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK); | |
381 | if (ret < 0) { | |
382 | PERROR("fcntl"); | |
383 | goto error_fcntl; | |
384 | } | |
385 | ||
74d81a6c MD |
386 | obj->type = SHM_OBJECT_MEM; |
387 | obj->memory_map = mem; | |
388 | obj->memory_map_size = memory_map_size; | |
389 | obj->allocated_len = memory_map_size; | |
390 | obj->index = table->allocated_len++; | |
391 | ||
392 | return obj; | |
ff0f5728 MD |
393 | |
394 | error_fcntl: | |
395 | return NULL; | |
74d81a6c MD |
396 | } |
397 | ||
1d498196 | 398 | static |
6548fca4 | 399 | void shmp_object_destroy(struct shm_object *obj, int consumer) |
1d498196 | 400 | { |
74d81a6c MD |
401 | switch (obj->type) { |
402 | case SHM_OBJECT_SHM: | |
403 | { | |
404 | int ret, i; | |
1d498196 | 405 | |
7a784989 MD |
406 | ret = munmap(obj->memory_map, obj->memory_map_size); |
407 | if (ret) { | |
408 | PERROR("umnmap"); | |
409 | assert(0); | |
410 | } | |
6548fca4 | 411 | |
5ea386c3 | 412 | if (obj->shm_fd_ownership) { |
6548fca4 MD |
413 | /* Delete FDs only if called from app (not consumer). */ |
414 | if (!consumer) { | |
415 | lttng_ust_lock_fd_tracker(); | |
416 | ret = close(obj->shm_fd); | |
417 | if (!ret) { | |
418 | lttng_ust_delete_fd_from_tracker(obj->shm_fd); | |
419 | } else { | |
420 | PERROR("close"); | |
421 | assert(0); | |
422 | } | |
423 | lttng_ust_unlock_fd_tracker(); | |
424 | } else { | |
425 | ret = close(obj->shm_fd); | |
426 | if (ret) { | |
427 | PERROR("close"); | |
428 | assert(0); | |
429 | } | |
a9ff648c MD |
430 | } |
431 | } | |
74d81a6c MD |
432 | for (i = 0; i < 2; i++) { |
433 | if (obj->wait_fd[i] < 0) | |
434 | continue; | |
6548fca4 MD |
435 | if (!consumer) { |
436 | lttng_ust_lock_fd_tracker(); | |
437 | ret = close(obj->wait_fd[i]); | |
438 | if (!ret) { | |
439 | lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]); | |
440 | } else { | |
441 | PERROR("close"); | |
442 | assert(0); | |
443 | } | |
444 | lttng_ust_unlock_fd_tracker(); | |
445 | } else { | |
446 | ret = close(obj->wait_fd[i]); | |
447 | if (ret) { | |
448 | PERROR("close"); | |
449 | assert(0); | |
450 | } | |
74d81a6c | 451 | } |
1d498196 | 452 | } |
74d81a6c MD |
453 | break; |
454 | } | |
455 | case SHM_OBJECT_MEM: | |
ff0f5728 MD |
456 | { |
457 | int ret, i; | |
458 | ||
459 | for (i = 0; i < 2; i++) { | |
460 | if (obj->wait_fd[i] < 0) | |
461 | continue; | |
6548fca4 MD |
462 | if (!consumer) { |
463 | lttng_ust_lock_fd_tracker(); | |
464 | ret = close(obj->wait_fd[i]); | |
465 | if (!ret) { | |
466 | lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]); | |
467 | } else { | |
468 | PERROR("close"); | |
469 | assert(0); | |
470 | } | |
471 | lttng_ust_unlock_fd_tracker(); | |
472 | } else { | |
473 | ret = close(obj->wait_fd[i]); | |
474 | if (ret) { | |
475 | PERROR("close"); | |
476 | assert(0); | |
477 | } | |
ff0f5728 MD |
478 | } |
479 | } | |
74d81a6c MD |
480 | free(obj->memory_map); |
481 | break; | |
ff0f5728 | 482 | } |
74d81a6c MD |
483 | default: |
484 | assert(0); | |
1d498196 MD |
485 | } |
486 | } | |
487 | ||
6548fca4 | 488 | void shm_object_table_destroy(struct shm_object_table *table, int consumer) |
1d498196 MD |
489 | { |
490 | int i; | |
491 | ||
492 | for (i = 0; i < table->allocated_len; i++) | |
6548fca4 | 493 | shmp_object_destroy(&table->objects[i], consumer); |
1d498196 MD |
494 | free(table); |
495 | } | |
496 | ||
497 | /* | |
498 | * zalloc_shm - allocate memory within a shm object. | |
499 | * | |
500 | * Shared memory is already zeroed by shmget. | |
501 | * *NOT* multithread-safe (should be protected by mutex). | |
502 | * Returns a -1, -1 tuple on error. | |
503 | */ | |
504 | struct shm_ref zalloc_shm(struct shm_object *obj, size_t len) | |
505 | { | |
506 | struct shm_ref ref; | |
507 | struct shm_ref shm_ref_error = { -1, -1 }; | |
508 | ||
509 | if (obj->memory_map_size - obj->allocated_len < len) | |
510 | return shm_ref_error; | |
511 | ref.index = obj->index; | |
512 | ref.offset = obj->allocated_len; | |
513 | obj->allocated_len += len; | |
514 | return ref; | |
515 | } | |
516 | ||
517 | void align_shm(struct shm_object *obj, size_t align) | |
518 | { | |
b72687b8 | 519 | size_t offset_len = lttng_ust_offset_align(obj->allocated_len, align); |
1d498196 MD |
520 | obj->allocated_len += offset_len; |
521 | } |