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 MD |
71 | |
72 | struct shm_object_table *shm_object_table_create(size_t max_nb_obj) | |
73 | { | |
74 | struct shm_object_table *table; | |
75 | ||
76 | table = zmalloc(sizeof(struct shm_object_table) + | |
77 | max_nb_obj * sizeof(table->objects[0])); | |
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, |
5ea386c3 | 87 | int stream_fd) |
1d498196 | 88 | { |
5ea386c3 | 89 | int shmfd, waitfd[2], ret, i; |
1d498196 MD |
90 | struct shm_object *obj; |
91 | char *memory_map; | |
92 | ||
5ea386c3 MD |
93 | if (stream_fd < 0) |
94 | return NULL; | |
1d498196 MD |
95 | if (table->allocated_len >= table->size) |
96 | return NULL; | |
7a9c21bd | 97 | obj = &table->objects[table->allocated_len]; |
1d498196 MD |
98 | |
99 | /* wait_fd: create pipe */ | |
8a8c2117 | 100 | ret = pipe2(waitfd, O_CLOEXEC); |
1d498196 MD |
101 | if (ret < 0) { |
102 | PERROR("pipe"); | |
103 | goto error_pipe; | |
104 | } | |
5d61a504 MD |
105 | /* The write end of the pipe needs to be non-blocking */ |
106 | ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK); | |
107 | if (ret < 0) { | |
108 | PERROR("fcntl"); | |
109 | goto error_fcntl; | |
110 | } | |
7a9c21bd | 111 | memcpy(obj->wait_fd, waitfd, sizeof(waitfd)); |
1d498196 | 112 | |
053e6e24 MJ |
113 | /* |
114 | * Set POSIX shared memory object size | |
115 | * | |
116 | * First, use ftruncate() to set its size, some implementations won't | |
117 | * allow writes past the size set by ftruncate. | |
118 | * Then, use write() to fill it with zeros, this allows us to fully | |
119 | * allocate it and detect a shortage of shm space without dealing with | |
120 | * a SIGBUS. | |
121 | */ | |
a9ff648c | 122 | |
5ea386c3 | 123 | shmfd = stream_fd; |
1d498196 MD |
124 | ret = ftruncate(shmfd, memory_map_size); |
125 | if (ret) { | |
126 | PERROR("ftruncate"); | |
127 | goto error_ftruncate; | |
128 | } | |
053e6e24 MJ |
129 | ret = zero_file(shmfd, memory_map_size); |
130 | if (ret) { | |
131 | PERROR("zero_file"); | |
132 | goto error_zero_file; | |
133 | } | |
71be0c53 | 134 | |
d0f6cf57 MD |
135 | /* |
136 | * Also ensure the file metadata is synced with the storage by using | |
71be0c53 MJ |
137 | * fsync(2). Some platforms don't allow fsync on POSIX shm fds, ignore |
138 | * EINVAL accordingly. | |
d0f6cf57 MD |
139 | */ |
140 | ret = fsync(shmfd); | |
71be0c53 | 141 | if (ret && errno != EINVAL) { |
d0f6cf57 MD |
142 | PERROR("fsync"); |
143 | goto error_fsync; | |
144 | } | |
5ea386c3 | 145 | obj->shm_fd_ownership = 0; |
1d498196 MD |
146 | obj->shm_fd = shmfd; |
147 | ||
148 | /* memory_map: mmap */ | |
149 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, | |
4d4838ba | 150 | MAP_SHARED | LTTNG_MAP_POPULATE, shmfd, 0); |
1d498196 MD |
151 | if (memory_map == MAP_FAILED) { |
152 | PERROR("mmap"); | |
153 | goto error_mmap; | |
154 | } | |
74d81a6c | 155 | obj->type = SHM_OBJECT_SHM; |
1d498196 MD |
156 | obj->memory_map = memory_map; |
157 | obj->memory_map_size = memory_map_size; | |
158 | obj->allocated_len = 0; | |
dc613eb9 | 159 | obj->index = table->allocated_len++; |
7a9c21bd | 160 | |
1d498196 MD |
161 | return obj; |
162 | ||
163 | error_mmap: | |
d0f6cf57 | 164 | error_fsync: |
1d498196 | 165 | error_ftruncate: |
3a81f31d | 166 | error_zero_file: |
1d498196 MD |
167 | error_fcntl: |
168 | for (i = 0; i < 2; i++) { | |
169 | ret = close(waitfd[i]); | |
170 | if (ret) { | |
171 | PERROR("close"); | |
172 | assert(0); | |
173 | } | |
174 | } | |
175 | error_pipe: | |
1d498196 | 176 | return NULL; |
1d498196 MD |
177 | } |
178 | ||
74d81a6c MD |
179 | static |
180 | struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table, | |
181 | size_t memory_map_size) | |
182 | { | |
183 | struct shm_object *obj; | |
184 | void *memory_map; | |
ff0f5728 | 185 | int waitfd[2], i, ret; |
74d81a6c MD |
186 | |
187 | if (table->allocated_len >= table->size) | |
188 | return NULL; | |
189 | obj = &table->objects[table->allocated_len]; | |
190 | ||
191 | memory_map = zmalloc(memory_map_size); | |
192 | if (!memory_map) | |
193 | goto alloc_error; | |
194 | ||
ff0f5728 | 195 | /* wait_fd: create pipe */ |
8a8c2117 | 196 | ret = pipe2(waitfd, O_CLOEXEC); |
ff0f5728 MD |
197 | if (ret < 0) { |
198 | PERROR("pipe"); | |
199 | goto error_pipe; | |
200 | } | |
ff0f5728 MD |
201 | /* The write end of the pipe needs to be non-blocking */ |
202 | ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK); | |
203 | if (ret < 0) { | |
204 | PERROR("fcntl"); | |
205 | goto error_fcntl; | |
206 | } | |
207 | memcpy(obj->wait_fd, waitfd, sizeof(waitfd)); | |
208 | ||
209 | /* no shm_fd */ | |
74d81a6c | 210 | obj->shm_fd = -1; |
5ea386c3 | 211 | obj->shm_fd_ownership = 0; |
74d81a6c MD |
212 | |
213 | obj->type = SHM_OBJECT_MEM; | |
214 | obj->memory_map = memory_map; | |
215 | obj->memory_map_size = memory_map_size; | |
216 | obj->allocated_len = 0; | |
217 | obj->index = table->allocated_len++; | |
218 | ||
219 | return obj; | |
220 | ||
ff0f5728 MD |
221 | error_fcntl: |
222 | for (i = 0; i < 2; i++) { | |
223 | ret = close(waitfd[i]); | |
224 | if (ret) { | |
225 | PERROR("close"); | |
226 | assert(0); | |
227 | } | |
228 | } | |
229 | error_pipe: | |
230 | free(memory_map); | |
74d81a6c MD |
231 | alloc_error: |
232 | return NULL; | |
233 | } | |
234 | ||
8a208943 MD |
235 | /* |
236 | * libnuma prints errors on the console even for numa_available(). | |
237 | * Work-around this limitation by using get_mempolicy() directly to | |
238 | * check whether the kernel supports mempolicy. | |
239 | */ | |
240 | #ifdef HAVE_LIBNUMA | |
241 | static bool lttng_is_numa_available(void) | |
242 | { | |
243 | int ret; | |
244 | ||
245 | ret = get_mempolicy(NULL, NULL, 0, NULL, 0); | |
246 | if (ret && errno == ENOSYS) { | |
247 | return false; | |
248 | } | |
249 | return numa_available() > 0; | |
250 | } | |
251 | #endif | |
252 | ||
c494c0f1 | 253 | #ifdef HAVE_LIBNUMA |
74d81a6c MD |
254 | struct shm_object *shm_object_table_alloc(struct shm_object_table *table, |
255 | size_t memory_map_size, | |
a9ff648c | 256 | enum shm_object_type type, |
4b68c31f MD |
257 | int stream_fd, |
258 | int cpu) | |
c494c0f1 MJ |
259 | #else |
260 | struct shm_object *shm_object_table_alloc(struct shm_object_table *table, | |
261 | size_t memory_map_size, | |
262 | enum shm_object_type type, | |
263 | int stream_fd, | |
264 | int cpu __attribute__((unused))) | |
265 | #endif | |
74d81a6c | 266 | { |
4b68c31f | 267 | struct shm_object *shm_object; |
bfcda6ce | 268 | #ifdef HAVE_LIBNUMA |
8a208943 MD |
269 | int oldnode = 0, node; |
270 | bool numa_avail; | |
4b68c31f | 271 | |
8a208943 MD |
272 | numa_avail = lttng_is_numa_available(); |
273 | if (numa_avail) { | |
274 | oldnode = numa_preferred(); | |
275 | if (cpu >= 0) { | |
276 | node = numa_node_of_cpu(cpu); | |
277 | if (node >= 0) | |
278 | numa_set_preferred(node); | |
279 | } | |
280 | if (cpu < 0 || node < 0) | |
281 | numa_set_localalloc(); | |
4b68c31f | 282 | } |
bfcda6ce | 283 | #endif /* HAVE_LIBNUMA */ |
74d81a6c MD |
284 | switch (type) { |
285 | case SHM_OBJECT_SHM: | |
4b68c31f | 286 | shm_object = _shm_object_table_alloc_shm(table, memory_map_size, |
5ea386c3 | 287 | stream_fd); |
4b68c31f | 288 | break; |
74d81a6c | 289 | case SHM_OBJECT_MEM: |
4b68c31f MD |
290 | shm_object = _shm_object_table_alloc_mem(table, memory_map_size); |
291 | break; | |
74d81a6c MD |
292 | default: |
293 | assert(0); | |
294 | } | |
bfcda6ce | 295 | #ifdef HAVE_LIBNUMA |
8a208943 MD |
296 | if (numa_avail) |
297 | numa_set_preferred(oldnode); | |
bfcda6ce | 298 | #endif /* HAVE_LIBNUMA */ |
4b68c31f | 299 | return shm_object; |
74d81a6c MD |
300 | } |
301 | ||
302 | struct shm_object *shm_object_table_append_shm(struct shm_object_table *table, | |
303 | int shm_fd, int wakeup_fd, uint32_t stream_nr, | |
304 | size_t memory_map_size) | |
193183fb MD |
305 | { |
306 | struct shm_object *obj; | |
307 | char *memory_map; | |
74d81a6c | 308 | int ret; |
193183fb MD |
309 | |
310 | if (table->allocated_len >= table->size) | |
311 | return NULL; | |
74d81a6c MD |
312 | /* streams _must_ be received in sequential order, else fail. */ |
313 | if (stream_nr + 1 != table->allocated_len) | |
314 | return NULL; | |
315 | ||
193183fb MD |
316 | obj = &table->objects[table->allocated_len]; |
317 | ||
74d81a6c MD |
318 | /* wait_fd: set write end of the pipe. */ |
319 | obj->wait_fd[0] = -1; /* read end is unset */ | |
320 | obj->wait_fd[1] = wakeup_fd; | |
193183fb | 321 | obj->shm_fd = shm_fd; |
5ea386c3 | 322 | obj->shm_fd_ownership = 1; |
193183fb | 323 | |
74d81a6c MD |
324 | /* The write end of the pipe needs to be non-blocking */ |
325 | ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK); | |
326 | if (ret < 0) { | |
327 | PERROR("fcntl"); | |
328 | goto error_fcntl; | |
329 | } | |
330 | ||
193183fb MD |
331 | /* memory_map: mmap */ |
332 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, | |
4d4838ba | 333 | MAP_SHARED | LTTNG_MAP_POPULATE, shm_fd, 0); |
193183fb MD |
334 | if (memory_map == MAP_FAILED) { |
335 | PERROR("mmap"); | |
336 | goto error_mmap; | |
337 | } | |
74d81a6c | 338 | obj->type = SHM_OBJECT_SHM; |
193183fb MD |
339 | obj->memory_map = memory_map; |
340 | obj->memory_map_size = memory_map_size; | |
341 | obj->allocated_len = memory_map_size; | |
342 | obj->index = table->allocated_len++; | |
343 | ||
344 | return obj; | |
345 | ||
74d81a6c | 346 | error_fcntl: |
193183fb MD |
347 | error_mmap: |
348 | return NULL; | |
349 | } | |
350 | ||
74d81a6c MD |
351 | /* |
352 | * Passing ownership of mem to object. | |
353 | */ | |
354 | struct shm_object *shm_object_table_append_mem(struct shm_object_table *table, | |
ff0f5728 | 355 | void *mem, size_t memory_map_size, int wakeup_fd) |
74d81a6c MD |
356 | { |
357 | struct shm_object *obj; | |
ff0f5728 | 358 | int ret; |
74d81a6c MD |
359 | |
360 | if (table->allocated_len >= table->size) | |
361 | return NULL; | |
362 | obj = &table->objects[table->allocated_len]; | |
363 | ||
ff0f5728 MD |
364 | obj->wait_fd[0] = -1; /* read end is unset */ |
365 | obj->wait_fd[1] = wakeup_fd; | |
74d81a6c | 366 | obj->shm_fd = -1; |
5ea386c3 | 367 | obj->shm_fd_ownership = 0; |
74d81a6c | 368 | |
ff0f5728 MD |
369 | /* The write end of the pipe needs to be non-blocking */ |
370 | ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK); | |
371 | if (ret < 0) { | |
372 | PERROR("fcntl"); | |
373 | goto error_fcntl; | |
374 | } | |
375 | ||
74d81a6c MD |
376 | obj->type = SHM_OBJECT_MEM; |
377 | obj->memory_map = mem; | |
378 | obj->memory_map_size = memory_map_size; | |
379 | obj->allocated_len = memory_map_size; | |
380 | obj->index = table->allocated_len++; | |
381 | ||
382 | return obj; | |
ff0f5728 MD |
383 | |
384 | error_fcntl: | |
385 | return NULL; | |
74d81a6c MD |
386 | } |
387 | ||
1d498196 | 388 | static |
6548fca4 | 389 | void shmp_object_destroy(struct shm_object *obj, int consumer) |
1d498196 | 390 | { |
74d81a6c MD |
391 | switch (obj->type) { |
392 | case SHM_OBJECT_SHM: | |
393 | { | |
394 | int ret, i; | |
1d498196 | 395 | |
7a784989 MD |
396 | ret = munmap(obj->memory_map, obj->memory_map_size); |
397 | if (ret) { | |
398 | PERROR("umnmap"); | |
399 | assert(0); | |
400 | } | |
6548fca4 | 401 | |
5ea386c3 | 402 | if (obj->shm_fd_ownership) { |
6548fca4 MD |
403 | /* Delete FDs only if called from app (not consumer). */ |
404 | if (!consumer) { | |
405 | lttng_ust_lock_fd_tracker(); | |
406 | ret = close(obj->shm_fd); | |
407 | if (!ret) { | |
408 | lttng_ust_delete_fd_from_tracker(obj->shm_fd); | |
409 | } else { | |
410 | PERROR("close"); | |
411 | assert(0); | |
412 | } | |
413 | lttng_ust_unlock_fd_tracker(); | |
414 | } else { | |
415 | ret = close(obj->shm_fd); | |
416 | if (ret) { | |
417 | PERROR("close"); | |
418 | assert(0); | |
419 | } | |
a9ff648c MD |
420 | } |
421 | } | |
74d81a6c MD |
422 | for (i = 0; i < 2; i++) { |
423 | if (obj->wait_fd[i] < 0) | |
424 | continue; | |
6548fca4 MD |
425 | if (!consumer) { |
426 | lttng_ust_lock_fd_tracker(); | |
427 | ret = close(obj->wait_fd[i]); | |
428 | if (!ret) { | |
429 | lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]); | |
430 | } else { | |
431 | PERROR("close"); | |
432 | assert(0); | |
433 | } | |
434 | lttng_ust_unlock_fd_tracker(); | |
435 | } else { | |
436 | ret = close(obj->wait_fd[i]); | |
437 | if (ret) { | |
438 | PERROR("close"); | |
439 | assert(0); | |
440 | } | |
74d81a6c | 441 | } |
1d498196 | 442 | } |
74d81a6c MD |
443 | break; |
444 | } | |
445 | case SHM_OBJECT_MEM: | |
ff0f5728 MD |
446 | { |
447 | int ret, i; | |
448 | ||
449 | for (i = 0; i < 2; i++) { | |
450 | if (obj->wait_fd[i] < 0) | |
451 | continue; | |
6548fca4 MD |
452 | if (!consumer) { |
453 | lttng_ust_lock_fd_tracker(); | |
454 | ret = close(obj->wait_fd[i]); | |
455 | if (!ret) { | |
456 | lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]); | |
457 | } else { | |
458 | PERROR("close"); | |
459 | assert(0); | |
460 | } | |
461 | lttng_ust_unlock_fd_tracker(); | |
462 | } else { | |
463 | ret = close(obj->wait_fd[i]); | |
464 | if (ret) { | |
465 | PERROR("close"); | |
466 | assert(0); | |
467 | } | |
ff0f5728 MD |
468 | } |
469 | } | |
74d81a6c MD |
470 | free(obj->memory_map); |
471 | break; | |
ff0f5728 | 472 | } |
74d81a6c MD |
473 | default: |
474 | assert(0); | |
1d498196 MD |
475 | } |
476 | } | |
477 | ||
6548fca4 | 478 | void shm_object_table_destroy(struct shm_object_table *table, int consumer) |
1d498196 MD |
479 | { |
480 | int i; | |
481 | ||
482 | for (i = 0; i < table->allocated_len; i++) | |
6548fca4 | 483 | shmp_object_destroy(&table->objects[i], consumer); |
1d498196 MD |
484 | free(table); |
485 | } | |
486 | ||
487 | /* | |
488 | * zalloc_shm - allocate memory within a shm object. | |
489 | * | |
490 | * Shared memory is already zeroed by shmget. | |
491 | * *NOT* multithread-safe (should be protected by mutex). | |
492 | * Returns a -1, -1 tuple on error. | |
493 | */ | |
494 | struct shm_ref zalloc_shm(struct shm_object *obj, size_t len) | |
495 | { | |
496 | struct shm_ref ref; | |
497 | struct shm_ref shm_ref_error = { -1, -1 }; | |
498 | ||
499 | if (obj->memory_map_size - obj->allocated_len < len) | |
500 | return shm_ref_error; | |
501 | ref.index = obj->index; | |
502 | ref.offset = obj->allocated_len; | |
503 | obj->allocated_len += len; | |
504 | return ref; | |
505 | } | |
506 | ||
507 | void align_shm(struct shm_object *obj, size_t align) | |
508 | { | |
b72687b8 | 509 | size_t offset_len = lttng_ust_offset_align(obj->allocated_len, align); |
1d498196 MD |
510 | obj->allocated_len += offset_len; |
511 | } |