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