Commit | Line | Data |
---|---|---|
ebabbf58 | 1 | /* |
c0c0989a | 2 | * SPDX-License-Identifier: LGPL-2.1-only |
ebabbf58 MD |
3 | * |
4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
ebabbf58 MD |
5 | */ |
6 | ||
7 | #define _LGPL_SOURCE | |
8 | #include <config.h> | |
9 | #include "shm.h" | |
10 | #include <unistd.h> | |
11 | #include <fcntl.h> | |
12 | #include <sys/mman.h> | |
13 | #include <sys/types.h> | |
14 | #include <sys/stat.h> /* For mode constants */ | |
15 | #include <fcntl.h> /* For O_* constants */ | |
16 | #include <assert.h> | |
17 | #include <stdio.h> | |
18 | #include <signal.h> | |
19 | #include <dirent.h> | |
ebabbf58 MD |
20 | #include <limits.h> |
21 | #include <stdbool.h> | |
22 | #include <stdint.h> | |
3d3a2bb8 | 23 | |
ebabbf58 MD |
24 | #ifdef HAVE_LIBNUMA |
25 | #include <numa.h> | |
26 | #include <numaif.h> | |
27 | #endif | |
3d3a2bb8 MJ |
28 | |
29 | #include <lttng/ust-align.h> | |
30 | ||
864a1eda | 31 | #include <ust-helper.h> |
ebabbf58 MD |
32 | #include <ust-fd.h> |
33 | #include "../libringbuffer/mmap.h" | |
34 | ||
35 | /* | |
36 | * Ensure we have the required amount of space available by writing 0 | |
37 | * into the entire buffer. Not doing so can trigger SIGBUS when going | |
38 | * beyond the available shm space. | |
39 | */ | |
40 | static | |
41 | int zero_file(int fd, size_t len) | |
42 | { | |
43 | ssize_t retlen; | |
44 | size_t written = 0; | |
45 | char *zeropage; | |
46 | long pagelen; | |
47 | int ret; | |
48 | ||
49 | pagelen = sysconf(_SC_PAGESIZE); | |
50 | if (pagelen < 0) | |
51 | return (int) pagelen; | |
52 | zeropage = calloc(pagelen, 1); | |
53 | if (!zeropage) | |
54 | return -ENOMEM; | |
55 | ||
56 | while (len > written) { | |
57 | do { | |
58 | retlen = write(fd, zeropage, | |
59 | min_t(size_t, pagelen, len - written)); | |
60 | } while (retlen == -1UL && errno == EINTR); | |
61 | if (retlen < 0) { | |
62 | ret = (int) retlen; | |
63 | goto error; | |
64 | } | |
65 | written += retlen; | |
66 | } | |
67 | ret = 0; | |
68 | error: | |
69 | free(zeropage); | |
70 | return ret; | |
71 | } | |
72 | ||
73 | struct lttng_counter_shm_object_table *lttng_counter_shm_object_table_create(size_t max_nb_obj) | |
74 | { | |
75 | struct lttng_counter_shm_object_table *table; | |
76 | ||
77 | table = zmalloc(sizeof(struct lttng_counter_shm_object_table) + | |
78 | max_nb_obj * sizeof(table->objects[0])); | |
79 | if (!table) | |
80 | return NULL; | |
81 | table->size = max_nb_obj; | |
82 | return table; | |
83 | } | |
84 | ||
85 | static | |
86 | struct lttng_counter_shm_object *_lttng_counter_shm_object_table_alloc_shm(struct lttng_counter_shm_object_table *table, | |
87 | size_t memory_map_size, | |
88 | int cpu_fd) | |
89 | { | |
90 | int shmfd, ret; | |
91 | struct lttng_counter_shm_object *obj; | |
92 | char *memory_map; | |
93 | ||
94 | if (cpu_fd < 0) | |
95 | return NULL; | |
96 | if (table->allocated_len >= table->size) | |
97 | return NULL; | |
98 | obj = &table->objects[table->allocated_len]; | |
99 | ||
100 | /* create shm */ | |
101 | ||
102 | shmfd = cpu_fd; | |
103 | ret = zero_file(shmfd, memory_map_size); | |
104 | if (ret) { | |
105 | PERROR("zero_file"); | |
106 | goto error_zero_file; | |
107 | } | |
108 | ret = ftruncate(shmfd, memory_map_size); | |
109 | if (ret) { | |
110 | PERROR("ftruncate"); | |
111 | goto error_ftruncate; | |
112 | } | |
113 | /* | |
114 | * Also ensure the file metadata is synced with the storage by using | |
115 | * fsync(2). | |
116 | */ | |
117 | ret = fsync(shmfd); | |
118 | if (ret) { | |
119 | PERROR("fsync"); | |
120 | goto error_fsync; | |
121 | } | |
122 | obj->shm_fd_ownership = 0; | |
123 | obj->shm_fd = shmfd; | |
124 | ||
125 | /* memory_map: mmap */ | |
126 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, | |
127 | MAP_SHARED | LTTNG_MAP_POPULATE, shmfd, 0); | |
128 | if (memory_map == MAP_FAILED) { | |
129 | PERROR("mmap"); | |
130 | goto error_mmap; | |
131 | } | |
132 | obj->type = LTTNG_COUNTER_SHM_OBJECT_SHM; | |
133 | obj->memory_map = memory_map; | |
134 | obj->memory_map_size = memory_map_size; | |
135 | obj->allocated_len = 0; | |
136 | obj->index = table->allocated_len++; | |
137 | ||
138 | return obj; | |
139 | ||
140 | error_mmap: | |
141 | error_fsync: | |
142 | error_ftruncate: | |
143 | error_zero_file: | |
144 | return NULL; | |
145 | } | |
146 | ||
147 | static | |
148 | struct lttng_counter_shm_object *_lttng_counter_shm_object_table_alloc_mem(struct lttng_counter_shm_object_table *table, | |
149 | size_t memory_map_size) | |
150 | { | |
151 | struct lttng_counter_shm_object *obj; | |
152 | void *memory_map; | |
153 | ||
154 | if (table->allocated_len >= table->size) | |
155 | return NULL; | |
156 | obj = &table->objects[table->allocated_len]; | |
157 | ||
158 | memory_map = zmalloc(memory_map_size); | |
159 | if (!memory_map) | |
160 | goto alloc_error; | |
161 | ||
162 | /* no shm_fd */ | |
163 | obj->shm_fd = -1; | |
164 | obj->shm_fd_ownership = 0; | |
165 | ||
166 | obj->type = LTTNG_COUNTER_SHM_OBJECT_MEM; | |
167 | obj->memory_map = memory_map; | |
168 | obj->memory_map_size = memory_map_size; | |
169 | obj->allocated_len = 0; | |
170 | obj->index = table->allocated_len++; | |
171 | ||
172 | return obj; | |
173 | ||
174 | alloc_error: | |
175 | return NULL; | |
176 | } | |
177 | ||
178 | /* | |
179 | * libnuma prints errors on the console even for numa_available(). | |
180 | * Work-around this limitation by using get_mempolicy() directly to | |
181 | * check whether the kernel supports mempolicy. | |
182 | */ | |
183 | #ifdef HAVE_LIBNUMA | |
184 | static bool lttng_is_numa_available(void) | |
185 | { | |
186 | int ret; | |
187 | ||
188 | ret = get_mempolicy(NULL, NULL, 0, NULL, 0); | |
189 | if (ret && errno == ENOSYS) { | |
190 | return false; | |
191 | } | |
192 | return numa_available() > 0; | |
193 | } | |
194 | #endif | |
195 | ||
196 | struct lttng_counter_shm_object *lttng_counter_shm_object_table_alloc(struct lttng_counter_shm_object_table *table, | |
197 | size_t memory_map_size, | |
198 | enum lttng_counter_shm_object_type type, | |
199 | int cpu_fd, | |
200 | int cpu) | |
201 | { | |
202 | struct lttng_counter_shm_object *shm_object; | |
203 | #ifdef HAVE_LIBNUMA | |
204 | int oldnode = 0, node; | |
205 | bool numa_avail; | |
206 | ||
207 | numa_avail = lttng_is_numa_available(); | |
208 | if (numa_avail) { | |
209 | oldnode = numa_preferred(); | |
210 | if (cpu >= 0) { | |
211 | node = numa_node_of_cpu(cpu); | |
212 | if (node >= 0) | |
213 | numa_set_preferred(node); | |
214 | } | |
215 | if (cpu < 0 || node < 0) | |
216 | numa_set_localalloc(); | |
217 | } | |
218 | #endif /* HAVE_LIBNUMA */ | |
219 | switch (type) { | |
220 | case LTTNG_COUNTER_SHM_OBJECT_SHM: | |
221 | shm_object = _lttng_counter_shm_object_table_alloc_shm(table, memory_map_size, | |
222 | cpu_fd); | |
223 | break; | |
224 | case LTTNG_COUNTER_SHM_OBJECT_MEM: | |
225 | shm_object = _lttng_counter_shm_object_table_alloc_mem(table, memory_map_size); | |
226 | break; | |
227 | default: | |
228 | assert(0); | |
229 | } | |
230 | #ifdef HAVE_LIBNUMA | |
231 | if (numa_avail) | |
232 | numa_set_preferred(oldnode); | |
233 | #endif /* HAVE_LIBNUMA */ | |
234 | return shm_object; | |
235 | } | |
236 | ||
237 | struct lttng_counter_shm_object *lttng_counter_shm_object_table_append_shm(struct lttng_counter_shm_object_table *table, | |
238 | int shm_fd, | |
239 | size_t memory_map_size) | |
240 | { | |
241 | struct lttng_counter_shm_object *obj; | |
242 | char *memory_map; | |
243 | ||
244 | if (table->allocated_len >= table->size) | |
245 | return NULL; | |
246 | ||
247 | obj = &table->objects[table->allocated_len]; | |
248 | ||
249 | obj->shm_fd = shm_fd; | |
250 | obj->shm_fd_ownership = 1; | |
251 | ||
252 | /* memory_map: mmap */ | |
253 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, | |
254 | MAP_SHARED | LTTNG_MAP_POPULATE, shm_fd, 0); | |
255 | if (memory_map == MAP_FAILED) { | |
256 | PERROR("mmap"); | |
257 | goto error_mmap; | |
258 | } | |
259 | obj->type = LTTNG_COUNTER_SHM_OBJECT_SHM; | |
260 | obj->memory_map = memory_map; | |
261 | obj->memory_map_size = memory_map_size; | |
262 | obj->allocated_len = memory_map_size; | |
263 | obj->index = table->allocated_len++; | |
264 | ||
265 | return obj; | |
266 | ||
267 | error_mmap: | |
268 | return NULL; | |
269 | } | |
270 | ||
271 | /* | |
272 | * Passing ownership of mem to object. | |
273 | */ | |
274 | struct lttng_counter_shm_object *lttng_counter_shm_object_table_append_mem(struct lttng_counter_shm_object_table *table, | |
275 | void *mem, size_t memory_map_size) | |
276 | { | |
277 | struct lttng_counter_shm_object *obj; | |
278 | ||
279 | if (table->allocated_len >= table->size) | |
280 | return NULL; | |
281 | obj = &table->objects[table->allocated_len]; | |
282 | ||
283 | obj->shm_fd = -1; | |
284 | obj->shm_fd_ownership = 0; | |
285 | ||
286 | obj->type = LTTNG_COUNTER_SHM_OBJECT_MEM; | |
287 | obj->memory_map = mem; | |
288 | obj->memory_map_size = memory_map_size; | |
289 | obj->allocated_len = memory_map_size; | |
290 | obj->index = table->allocated_len++; | |
291 | ||
292 | return obj; | |
293 | ||
294 | return NULL; | |
295 | } | |
296 | ||
297 | static | |
298 | void lttng_counter_shmp_object_destroy(struct lttng_counter_shm_object *obj, int consumer) | |
299 | { | |
300 | switch (obj->type) { | |
301 | case LTTNG_COUNTER_SHM_OBJECT_SHM: | |
302 | { | |
303 | int ret; | |
304 | ||
305 | ret = munmap(obj->memory_map, obj->memory_map_size); | |
306 | if (ret) { | |
307 | PERROR("umnmap"); | |
308 | assert(0); | |
309 | } | |
310 | ||
311 | if (obj->shm_fd_ownership) { | |
312 | /* Delete FDs only if called from app (not consumer). */ | |
313 | if (!consumer) { | |
314 | lttng_ust_lock_fd_tracker(); | |
315 | ret = close(obj->shm_fd); | |
316 | if (!ret) { | |
317 | lttng_ust_delete_fd_from_tracker(obj->shm_fd); | |
318 | } else { | |
319 | PERROR("close"); | |
320 | assert(0); | |
321 | } | |
322 | lttng_ust_unlock_fd_tracker(); | |
323 | } else { | |
324 | ret = close(obj->shm_fd); | |
325 | if (ret) { | |
326 | PERROR("close"); | |
327 | assert(0); | |
328 | } | |
329 | } | |
330 | } | |
331 | break; | |
332 | } | |
333 | case LTTNG_COUNTER_SHM_OBJECT_MEM: | |
334 | { | |
335 | free(obj->memory_map); | |
336 | break; | |
337 | } | |
338 | default: | |
339 | assert(0); | |
340 | } | |
341 | } | |
342 | ||
343 | void lttng_counter_shm_object_table_destroy(struct lttng_counter_shm_object_table *table, int consumer) | |
344 | { | |
345 | int i; | |
346 | ||
347 | for (i = 0; i < table->allocated_len; i++) | |
348 | lttng_counter_shmp_object_destroy(&table->objects[i], consumer); | |
349 | free(table); | |
350 | } | |
351 | ||
352 | /* | |
353 | * lttng_counter_zalloc_shm - allocate memory within a shm object. | |
354 | * | |
355 | * Shared memory is already zeroed by shmget. | |
356 | * *NOT* multithread-safe (should be protected by mutex). | |
357 | * Returns a -1, -1 tuple on error. | |
358 | */ | |
359 | struct lttng_counter_shm_ref lttng_counter_zalloc_shm(struct lttng_counter_shm_object *obj, size_t len) | |
360 | { | |
361 | struct lttng_counter_shm_ref ref; | |
362 | struct lttng_counter_shm_ref shm_ref_error = { -1, -1 }; | |
363 | ||
364 | if (obj->memory_map_size - obj->allocated_len < len) | |
365 | return shm_ref_error; | |
366 | ref.index = obj->index; | |
367 | ref.offset = obj->allocated_len; | |
368 | obj->allocated_len += len; | |
369 | return ref; | |
370 | } | |
371 | ||
372 | void lttng_counter_align_shm(struct lttng_counter_shm_object *obj, size_t align) | |
373 | { | |
cd61d9bf | 374 | size_t offset_len = lttng_ust_offset_align(obj->allocated_len, align); |
ebabbf58 MD |
375 | obj->allocated_len += offset_len; |
376 | } |