Commit | Line | Data |
---|---|---|
0ad3c26c JG |
1 | /* |
2 | * Copyright (C) 2018 - Jérémie Galarneau <jeremie.galarneau@efficios.com> | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify it | |
5 | * under the terms of the GNU General Public License, version 2 only, as | |
6 | * published by the Free Software Foundation. | |
7 | * | |
8 | * This program is distributed in the hope that it will be useful, but WITHOUT | |
9 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
10 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
11 | * more details. | |
12 | * | |
13 | * You should have received a copy of the GNU General Public License along with | |
14 | * this program; if not, write to the Free Software Foundation, Inc., 51 | |
15 | * Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. | |
16 | */ | |
17 | ||
0ad3c26c JG |
18 | #include <urcu.h> |
19 | #include <urcu/list.h> | |
20 | #include <urcu/rculfhash.h> | |
21 | ||
22 | #include <sys/stat.h> | |
23 | #include <sys/types.h> | |
24 | #include <fcntl.h> | |
25 | #include <stdbool.h> | |
26 | #include <pthread.h> | |
27 | #include <inttypes.h> | |
28 | ||
29 | #include "common/macros.h" | |
30 | #include "common/error.h" | |
31 | #include "common/defaults.h" | |
32 | #include "common/hashtable/utils.h" | |
33 | #include "common/hashtable/hashtable.h" | |
34 | ||
35 | #include "fd-tracker.h" | |
6f89b613 | 36 | #include "inode.h" |
0ad3c26c JG |
37 | |
38 | /* Tracker lock must be taken by the user. */ | |
39 | #define TRACKED_COUNT(tracker) \ | |
40 | (tracker->count.suspendable.active + \ | |
41 | tracker->count.suspendable.suspended + \ | |
42 | tracker->count.unsuspendable) | |
43 | ||
44 | /* Tracker lock must be taken by the user. */ | |
45 | #define ACTIVE_COUNT(tracker) \ | |
46 | (tracker->count.suspendable.active + \ | |
47 | tracker->count.unsuspendable) | |
48 | ||
49 | /* Tracker lock must be taken by the user. */ | |
50 | #define SUSPENDED_COUNT(tracker) \ | |
51 | (tracker->count.suspendable.suspended) | |
52 | ||
53 | /* Tracker lock must be taken by the user. */ | |
54 | #define SUSPENDABLE_COUNT(tracker) \ | |
55 | (tracker->count.suspendable.active + \ | |
56 | tracker->count.suspendable.suspended) | |
57 | ||
58 | /* Tracker lock must be taken by the user. */ | |
59 | #define UNSUSPENDABLE_COUNT(tracker) \ | |
60 | (tracker->count.unsuspendable) | |
61 | ||
62 | struct fd_tracker { | |
63 | pthread_mutex_t lock; | |
64 | struct { | |
65 | struct { | |
66 | unsigned int active; | |
67 | unsigned int suspended; | |
68 | } suspendable; | |
69 | unsigned int unsuspendable; | |
70 | } count; | |
71 | unsigned int capacity; | |
72 | struct { | |
73 | uint64_t uses; | |
74 | uint64_t misses; | |
75 | /* Failures to suspend or restore fs handles. */ | |
76 | uint64_t errors; | |
77 | } stats; | |
78 | /* | |
79 | * The head of the active_handles list is always the least recently | |
80 | * used active handle. When an handle is used, it is removed from the | |
81 | * list and added to the end. When a file has to be suspended, the | |
82 | * first element in the list is "popped", suspended, and added to the | |
83 | * list of suspended handles. | |
84 | */ | |
85 | struct cds_list_head active_handles; | |
86 | struct cds_list_head suspended_handles; | |
87 | struct cds_lfht *unsuspendable_fds; | |
6f89b613 | 88 | struct lttng_inode_registry *inode_registry; |
0ad3c26c JG |
89 | }; |
90 | ||
91 | struct open_properties { | |
0ad3c26c JG |
92 | int flags; |
93 | struct { | |
94 | bool is_set; | |
95 | mode_t value; | |
96 | } mode; | |
97 | }; | |
98 | ||
99 | /* | |
100 | * A fs_handle is not ref-counted. Therefore, it is assumed that a | |
101 | * handle is never in-use while it is being reclaimed. It can be | |
102 | * shared by multiple threads, but external synchronization is required | |
103 | * to ensure it is not still being used when it is reclaimed (close method). | |
104 | * In this respect, it is not different from a regular file descriptor. | |
105 | * | |
106 | * The fs_handle lock always nests _within_ the tracker's lock. | |
107 | */ | |
108 | struct fs_handle { | |
109 | pthread_mutex_t lock; | |
110 | /* | |
111 | * Weak reference to the tracker. All fs_handles are assumed to have | |
112 | * been closed at the moment of the destruction of the fd_tracker. | |
113 | */ | |
114 | struct fd_tracker *tracker; | |
115 | struct open_properties properties; | |
6f89b613 | 116 | struct lttng_inode *inode; |
0ad3c26c JG |
117 | int fd; |
118 | /* inode number of the file at the time of the handle's creation. */ | |
119 | uint64_t ino; | |
120 | bool in_use; | |
121 | /* Offset to which the file should be restored. */ | |
122 | off_t offset; | |
123 | struct cds_list_head handles_list_node; | |
124 | }; | |
125 | ||
126 | struct unsuspendable_fd { | |
127 | /* | |
128 | * Accesses are only performed through the tracker, which is protected | |
129 | * by its own lock. | |
130 | */ | |
131 | int fd; | |
132 | char *name; | |
133 | struct cds_lfht_node tracker_node; | |
89b4e38c | 134 | struct rcu_head rcu_head; |
0ad3c26c JG |
135 | }; |
136 | ||
137 | static struct { | |
138 | pthread_mutex_t lock; | |
139 | bool initialized; | |
140 | unsigned long value; | |
141 | } seed = { | |
142 | .lock = PTHREAD_MUTEX_INITIALIZER, | |
143 | }; | |
144 | ||
145 | static int match_fd(struct cds_lfht_node *node, const void *key); | |
146 | static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry); | |
147 | static struct unsuspendable_fd *unsuspendable_fd_create(const char *name, | |
148 | int fd); | |
6f89b613 JG |
149 | static int open_from_properties(const char *path, |
150 | struct open_properties *properties); | |
0ad3c26c JG |
151 | |
152 | static void fs_handle_log(struct fs_handle *handle); | |
153 | static int fs_handle_suspend(struct fs_handle *handle); | |
154 | static int fs_handle_restore(struct fs_handle *handle); | |
155 | ||
156 | static void fd_tracker_track(struct fd_tracker *tracker, | |
157 | struct fs_handle *handle); | |
158 | static void fd_tracker_untrack(struct fd_tracker *tracker, | |
159 | struct fs_handle *handle); | |
160 | static int fd_tracker_suspend_handles(struct fd_tracker *tracker, | |
161 | unsigned int count); | |
162 | static int fd_tracker_restore_handle(struct fd_tracker *tracker, | |
163 | struct fs_handle *handle); | |
164 | ||
165 | /* Match function of the tracker's unsuspendable_fds hash table. */ | |
166 | static | |
167 | int match_fd(struct cds_lfht_node *node, const void *key) | |
168 | { | |
169 | struct unsuspendable_fd *entry = | |
170 | caa_container_of(node, struct unsuspendable_fd, tracker_node); | |
171 | ||
172 | return hash_match_key_ulong((void *) (unsigned long) entry->fd, | |
173 | (void *) key); | |
174 | } | |
175 | ||
89b4e38c JG |
176 | static |
177 | void delete_unsuspendable_fd(struct rcu_head *head) | |
178 | { | |
179 | struct unsuspendable_fd *fd = caa_container_of(head, | |
180 | struct unsuspendable_fd, rcu_head); | |
181 | ||
182 | free(fd->name); | |
183 | free(fd); | |
184 | } | |
185 | ||
0ad3c26c JG |
186 | static |
187 | void unsuspendable_fd_destroy(struct unsuspendable_fd *entry) | |
188 | { | |
189 | if (!entry) { | |
190 | return; | |
191 | } | |
89b4e38c | 192 | call_rcu(&entry->rcu_head, delete_unsuspendable_fd); |
0ad3c26c JG |
193 | } |
194 | ||
195 | static | |
196 | struct unsuspendable_fd *unsuspendable_fd_create(const char *name, int fd) | |
197 | { | |
198 | struct unsuspendable_fd *entry = | |
199 | zmalloc(sizeof(*entry)); | |
200 | ||
201 | if (!entry) { | |
202 | goto error; | |
203 | } | |
204 | if (name) { | |
205 | entry->name = strdup(name); | |
206 | if (!entry->name) { | |
207 | goto error; | |
208 | } | |
209 | } | |
210 | cds_lfht_node_init(&entry->tracker_node); | |
211 | entry->fd = fd; | |
212 | return entry; | |
213 | error: | |
214 | unsuspendable_fd_destroy(entry); | |
215 | return NULL; | |
216 | } | |
217 | ||
218 | static | |
219 | void fs_handle_log(struct fs_handle *handle) | |
220 | { | |
6f89b613 JG |
221 | const char *path; |
222 | ||
0ad3c26c | 223 | pthread_mutex_lock(&handle->lock); |
6f89b613 JG |
224 | path = lttng_inode_get_path(handle->inode); |
225 | ||
0ad3c26c JG |
226 | if (handle->fd >= 0) { |
227 | DBG_NO_LOC(" %s [active, fd %d%s]", | |
6f89b613 | 228 | path, |
0ad3c26c JG |
229 | handle->fd, |
230 | handle->in_use ? ", in use" : ""); | |
231 | } else { | |
6f89b613 | 232 | DBG_NO_LOC(" %s [suspended]", path); |
0ad3c26c JG |
233 | } |
234 | pthread_mutex_unlock(&handle->lock); | |
235 | } | |
236 | ||
6f89b613 | 237 | /* Tracker lock must be held by the caller. */ |
0ad3c26c JG |
238 | static |
239 | int fs_handle_suspend(struct fs_handle *handle) | |
240 | { | |
241 | int ret = 0; | |
242 | struct stat fs_stat; | |
6f89b613 | 243 | const char *path; |
0ad3c26c JG |
244 | |
245 | pthread_mutex_lock(&handle->lock); | |
6f89b613 | 246 | path = lttng_inode_get_path(handle->inode); |
0ad3c26c JG |
247 | assert(handle->fd >= 0); |
248 | if (handle->in_use) { | |
249 | /* This handle can't be suspended as it is currently in use. */ | |
250 | ret = -EAGAIN; | |
251 | goto end; | |
252 | } | |
253 | ||
6f89b613 | 254 | ret = stat(path, &fs_stat); |
0ad3c26c JG |
255 | if (ret) { |
256 | PERROR("Filesystem handle to %s cannot be suspended as stat() failed", | |
6f89b613 | 257 | path); |
0ad3c26c JG |
258 | ret = -errno; |
259 | goto end; | |
260 | } | |
261 | ||
262 | if (fs_stat.st_ino != handle->ino) { | |
263 | /* Don't suspend as the handle would not be restorable. */ | |
264 | WARN("Filesystem handle to %s cannot be suspended as its inode changed", | |
6f89b613 | 265 | path); |
0ad3c26c JG |
266 | ret = -ENOENT; |
267 | goto end; | |
268 | } | |
269 | ||
270 | handle->offset = lseek(handle->fd, 0, SEEK_CUR); | |
271 | if (handle->offset == -1) { | |
272 | WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position", | |
6f89b613 | 273 | path); |
0ad3c26c JG |
274 | ret = -errno; |
275 | goto end; | |
276 | } | |
277 | ||
278 | ret = close(handle->fd); | |
279 | if (ret) { | |
280 | PERROR("Filesystem handle to %s cannot be suspended as close() failed", | |
6f89b613 | 281 | path); |
0ad3c26c JG |
282 | ret = -errno; |
283 | goto end; | |
284 | } | |
285 | DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64, | |
6f89b613 | 286 | path, handle->fd, handle->offset); |
0ad3c26c JG |
287 | handle->fd = -1; |
288 | end: | |
289 | if (ret) { | |
290 | handle->tracker->stats.errors++; | |
291 | } | |
292 | pthread_mutex_unlock(&handle->lock); | |
293 | return ret; | |
294 | } | |
295 | ||
296 | /* Caller must hold the tracker and handle's locks. */ | |
297 | static | |
298 | int fs_handle_restore(struct fs_handle *handle) | |
299 | { | |
300 | int ret, fd = -1; | |
6f89b613 | 301 | const char *path = lttng_inode_get_path(handle->inode); |
0ad3c26c JG |
302 | |
303 | assert(handle->fd == -1); | |
6f89b613 JG |
304 | assert(path); |
305 | ret = open_from_properties(path, | |
306 | &handle->properties); | |
0ad3c26c JG |
307 | if (ret < 0) { |
308 | PERROR("Failed to restore filesystem handle to %s, open() failed", | |
6f89b613 | 309 | path); |
0ad3c26c JG |
310 | ret = -errno; |
311 | goto end; | |
312 | } | |
313 | fd = ret; | |
314 | ||
315 | ret = lseek(fd, handle->offset, SEEK_SET); | |
316 | if (ret < 0) { | |
317 | PERROR("Failed to restore filesystem handle to %s, lseek() failed", | |
6f89b613 | 318 | path); |
0ad3c26c JG |
319 | ret = -errno; |
320 | goto end; | |
321 | } | |
322 | DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64, | |
6f89b613 | 323 | path, fd, handle->offset); |
0ad3c26c JG |
324 | ret = 0; |
325 | handle->fd = fd; | |
326 | fd = -1; | |
327 | end: | |
328 | if (fd >= 0) { | |
329 | (void) close(fd); | |
330 | } | |
331 | return ret; | |
332 | } | |
333 | ||
334 | static | |
6f89b613 | 335 | int open_from_properties(const char *path, struct open_properties *properties) |
0ad3c26c JG |
336 | { |
337 | int ret; | |
338 | ||
339 | /* | |
340 | * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE | |
341 | * flags are set. O_TMPFILE would not make sense in the context of a | |
342 | * suspendable fs_handle as it would not be restorable (see OPEN(2)), | |
343 | * thus it is ignored here. | |
344 | */ | |
345 | if ((properties->flags & O_CREAT) && properties->mode.is_set) { | |
6f89b613 | 346 | ret = open(path, properties->flags, |
0ad3c26c JG |
347 | properties->mode.value); |
348 | } else { | |
6f89b613 | 349 | ret = open(path, properties->flags); |
0ad3c26c JG |
350 | } |
351 | /* | |
352 | * Some flags should not be used beyond the initial open() of a | |
353 | * restorable file system handle. O_CREAT and O_TRUNC must | |
354 | * be cleared since it would be unexpected to re-use them | |
355 | * when the handle is retored: | |
356 | * - O_CREAT should not be needed as the file has been created | |
357 | * on the initial call to open(), | |
358 | * - O_TRUNC would destroy the file's contents by truncating it | |
359 | * to length 0. | |
360 | */ | |
361 | properties->flags &= ~(O_CREAT | O_TRUNC); | |
362 | if (ret < 0) { | |
363 | ret = -errno; | |
364 | goto end; | |
365 | } | |
366 | end: | |
367 | return ret; | |
368 | } | |
369 | ||
370 | struct fd_tracker *fd_tracker_create(unsigned int capacity) | |
371 | { | |
372 | struct fd_tracker *tracker = zmalloc(sizeof(struct fd_tracker)); | |
373 | ||
374 | if (!tracker) { | |
375 | goto end; | |
376 | } | |
377 | ||
378 | pthread_mutex_lock(&seed.lock); | |
379 | if (!seed.initialized) { | |
380 | seed.value = (unsigned long) time(NULL); | |
381 | seed.initialized = true; | |
382 | } | |
383 | pthread_mutex_unlock(&seed.lock); | |
384 | ||
385 | CDS_INIT_LIST_HEAD(&tracker->active_handles); | |
386 | CDS_INIT_LIST_HEAD(&tracker->suspended_handles); | |
387 | tracker->capacity = capacity; | |
388 | tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0, | |
389 | CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL); | |
6f89b613 JG |
390 | if (!tracker->unsuspendable_fds) { |
391 | ERR("Failed to create fd-tracker's unsuspendable_fds hash table"); | |
392 | goto error; | |
393 | } | |
394 | tracker->inode_registry = lttng_inode_registry_create(); | |
395 | if (!tracker->inode_registry) { | |
396 | ERR("Failed to create fd-tracker's inode registry"); | |
397 | goto error; | |
398 | } | |
a8a54a6f JG |
399 | DBG("File descriptor tracker created with a limit of %u simultaneously-opened FDs", |
400 | capacity); | |
0ad3c26c JG |
401 | end: |
402 | return tracker; | |
6f89b613 JG |
403 | error: |
404 | fd_tracker_destroy(tracker); | |
405 | return NULL; | |
0ad3c26c JG |
406 | } |
407 | ||
408 | void fd_tracker_log(struct fd_tracker *tracker) | |
409 | { | |
410 | struct fs_handle *handle; | |
411 | struct unsuspendable_fd *unsuspendable_fd; | |
412 | struct cds_lfht_iter iter; | |
413 | ||
414 | pthread_mutex_lock(&tracker->lock); | |
415 | DBG_NO_LOC("File descriptor tracker"); | |
416 | DBG_NO_LOC(" Stats:"); | |
417 | DBG_NO_LOC(" uses: %" PRIu64, tracker->stats.uses); | |
418 | DBG_NO_LOC(" misses: %" PRIu64, tracker->stats.misses); | |
419 | DBG_NO_LOC(" errors: %" PRIu64, tracker->stats.errors); | |
420 | DBG_NO_LOC(" Tracked: %u", TRACKED_COUNT(tracker)); | |
421 | DBG_NO_LOC(" active: %u", ACTIVE_COUNT(tracker)); | |
422 | DBG_NO_LOC(" suspendable: %u", SUSPENDABLE_COUNT(tracker)); | |
423 | DBG_NO_LOC(" unsuspendable: %u", UNSUSPENDABLE_COUNT(tracker)); | |
424 | DBG_NO_LOC(" suspended: %u", SUSPENDED_COUNT(tracker)); | |
425 | DBG_NO_LOC(" capacity: %u", tracker->capacity); | |
426 | ||
427 | DBG_NO_LOC(" Tracked suspendable file descriptors"); | |
428 | cds_list_for_each_entry(handle, &tracker->active_handles, | |
429 | handles_list_node) { | |
430 | fs_handle_log(handle); | |
431 | } | |
432 | cds_list_for_each_entry(handle, &tracker->suspended_handles, | |
433 | handles_list_node) { | |
434 | fs_handle_log(handle); | |
435 | } | |
436 | if (!SUSPENDABLE_COUNT(tracker)) { | |
437 | DBG_NO_LOC(" None"); | |
438 | } | |
439 | ||
440 | DBG_NO_LOC(" Tracked unsuspendable file descriptors"); | |
441 | rcu_read_lock(); | |
442 | cds_lfht_for_each_entry(tracker->unsuspendable_fds, &iter, | |
443 | unsuspendable_fd, tracker_node) { | |
444 | DBG_NO_LOC(" %s [active, fd %d]", unsuspendable_fd->name ? : "Unnamed", | |
445 | unsuspendable_fd->fd); | |
446 | } | |
447 | rcu_read_unlock(); | |
448 | if (!UNSUSPENDABLE_COUNT(tracker)) { | |
449 | DBG_NO_LOC(" None"); | |
450 | } | |
451 | ||
452 | pthread_mutex_unlock(&tracker->lock); | |
453 | } | |
454 | ||
455 | int fd_tracker_destroy(struct fd_tracker *tracker) | |
456 | { | |
457 | int ret = 0; | |
458 | ||
459 | /* | |
460 | * Refuse to destroy the tracker as fs_handles may still old | |
461 | * weak references to the tracker. | |
462 | */ | |
463 | pthread_mutex_lock(&tracker->lock); | |
464 | if (TRACKED_COUNT(tracker)) { | |
465 | ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked", | |
466 | TRACKED_COUNT(tracker)); | |
467 | pthread_mutex_unlock(&tracker->lock); | |
468 | fd_tracker_log(tracker); | |
469 | ret = -1; | |
470 | goto end; | |
471 | } | |
472 | pthread_mutex_unlock(&tracker->lock); | |
473 | ||
6f89b613 JG |
474 | if (tracker->unsuspendable_fds) { |
475 | ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL); | |
476 | assert(!ret); | |
477 | } | |
52803bf0 JG |
478 | |
479 | lttng_inode_registry_destroy(tracker->inode_registry); | |
0ad3c26c JG |
480 | pthread_mutex_destroy(&tracker->lock); |
481 | free(tracker); | |
482 | end: | |
483 | return ret; | |
484 | } | |
485 | ||
486 | struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker, | |
487 | const char *path, int flags, mode_t *mode) | |
488 | { | |
489 | int ret; | |
490 | struct fs_handle *handle = NULL; | |
491 | struct stat fd_stat; | |
492 | struct open_properties properties = { | |
0ad3c26c JG |
493 | .flags = flags, |
494 | .mode.is_set = !!mode, | |
495 | .mode.value = mode ? *mode : 0, | |
496 | }; | |
497 | ||
0ad3c26c JG |
498 | pthread_mutex_lock(&tracker->lock); |
499 | if (ACTIVE_COUNT(tracker) == tracker->capacity) { | |
500 | if (tracker->count.suspendable.active > 0) { | |
501 | ret = fd_tracker_suspend_handles(tracker, 1); | |
502 | if (ret) { | |
f5a2cd65 | 503 | goto end; |
0ad3c26c JG |
504 | } |
505 | } else { | |
506 | /* | |
507 | * There are not enough active suspendable file | |
f5a2cd65 JR |
508 | * descriptors to open a new fd and still accommodate |
509 | * the tracker's capacity. | |
0ad3c26c JG |
510 | */ |
511 | WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)", | |
512 | tracker->count.unsuspendable); | |
f5a2cd65 | 513 | goto end; |
0ad3c26c JG |
514 | } |
515 | } | |
516 | ||
517 | handle = zmalloc(sizeof(*handle)); | |
518 | if (!handle) { | |
519 | goto end; | |
520 | } | |
52803bf0 | 521 | handle->tracker = tracker; |
0ad3c26c JG |
522 | |
523 | ret = pthread_mutex_init(&handle->lock, NULL); | |
524 | if (ret) { | |
525 | PERROR("Failed to initialize handle mutex while creating fs handle"); | |
f5a2cd65 | 526 | goto error_mutex_init; |
0ad3c26c JG |
527 | } |
528 | ||
6f89b613 | 529 | handle->fd = open_from_properties(path, &properties); |
0ad3c26c JG |
530 | if (handle->fd < 0) { |
531 | PERROR("Failed to open fs handle to %s, open() returned", path); | |
f5a2cd65 | 532 | goto error; |
0ad3c26c JG |
533 | } |
534 | ||
0ad3c26c | 535 | handle->properties = properties; |
6f89b613 JG |
536 | |
537 | handle->inode = lttng_inode_registry_get_inode(tracker->inode_registry, | |
538 | handle->fd, path); | |
539 | if (!handle->inode) { | |
540 | ERR("Failed to get lttng_inode corresponding to file %s", | |
541 | path); | |
f5a2cd65 | 542 | goto error; |
6f89b613 | 543 | } |
0ad3c26c JG |
544 | |
545 | if (fstat(handle->fd, &fd_stat)) { | |
546 | PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned"); | |
f5a2cd65 | 547 | goto error; |
0ad3c26c JG |
548 | } |
549 | handle->ino = fd_stat.st_ino; | |
550 | ||
551 | fd_tracker_track(tracker, handle); | |
0ad3c26c | 552 | end: |
f5a2cd65 | 553 | pthread_mutex_unlock(&tracker->lock); |
0ad3c26c | 554 | return handle; |
f5a2cd65 | 555 | error: |
52803bf0 JG |
556 | if (handle->inode) { |
557 | lttng_inode_put(handle->inode); | |
558 | } | |
f5a2cd65 JR |
559 | pthread_mutex_destroy(&handle->lock); |
560 | error_mutex_init: | |
52803bf0 | 561 | free(handle); |
0ad3c26c JG |
562 | handle = NULL; |
563 | goto end; | |
564 | } | |
565 | ||
566 | /* Caller must hold the tracker's lock. */ | |
567 | static | |
568 | int fd_tracker_suspend_handles(struct fd_tracker *tracker, | |
569 | unsigned int count) | |
570 | { | |
571 | unsigned int left_to_close = count; | |
572 | struct fs_handle *handle, *tmp; | |
573 | ||
574 | cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles, | |
575 | handles_list_node) { | |
576 | int ret; | |
577 | ||
578 | fd_tracker_untrack(tracker, handle); | |
579 | ret = fs_handle_suspend(handle); | |
580 | fd_tracker_track(tracker, handle); | |
581 | if (!ret) { | |
582 | left_to_close--; | |
583 | } | |
584 | ||
585 | if (!left_to_close) { | |
586 | break; | |
587 | } | |
588 | } | |
589 | return left_to_close ? -EMFILE : 0; | |
590 | } | |
591 | ||
592 | int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker, | |
593 | int *out_fds, const char **names, unsigned int fd_count, | |
594 | fd_open_cb open, void *user_data) | |
595 | { | |
596 | int ret, user_ret, i, fds_to_suspend; | |
597 | unsigned int active_fds; | |
598 | struct unsuspendable_fd *entries[fd_count]; | |
599 | ||
600 | memset(entries, 0, sizeof(entries)); | |
601 | ||
602 | pthread_mutex_lock(&tracker->lock); | |
603 | ||
604 | active_fds = ACTIVE_COUNT(tracker); | |
605 | fds_to_suspend = (int) active_fds + (int) fd_count - (int) tracker->capacity; | |
606 | if (fds_to_suspend > 0) { | |
607 | if (fds_to_suspend <= tracker->count.suspendable.active) { | |
608 | ret = fd_tracker_suspend_handles(tracker, fds_to_suspend); | |
609 | if (ret) { | |
610 | goto end; | |
611 | } | |
612 | } else { | |
613 | /* | |
614 | * There are not enough active suspendable file | |
615 | * descriptors to open a new fd and still accomodate the | |
616 | * tracker's capacity. | |
617 | */ | |
618 | WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)", | |
619 | tracker->count.unsuspendable); | |
620 | ret = -EMFILE; | |
621 | goto end; | |
622 | } | |
623 | } | |
624 | ||
625 | user_ret = open(user_data, out_fds); | |
626 | if (user_ret) { | |
627 | ret = user_ret; | |
628 | goto end; | |
629 | } | |
630 | ||
631 | /* | |
632 | * Add the fds returned by the user's callback to the hashtable | |
633 | * of unsuspendable fds. | |
634 | */ | |
635 | for (i = 0; i < fd_count; i++) { | |
636 | struct unsuspendable_fd *entry = | |
637 | unsuspendable_fd_create(names ? names[i] : NULL, | |
638 | out_fds[i]); | |
639 | ||
640 | if (!entry) { | |
641 | ret = -1; | |
642 | goto end_free_entries; | |
643 | } | |
644 | entries[i] = entry; | |
645 | } | |
646 | ||
647 | rcu_read_lock(); | |
648 | for (i = 0; i < fd_count; i++) { | |
649 | struct cds_lfht_node *node; | |
650 | struct unsuspendable_fd *entry = entries[i]; | |
651 | ||
652 | node = cds_lfht_add_unique( | |
653 | tracker->unsuspendable_fds, | |
654 | hash_key_ulong((void *) (unsigned long) out_fds[i], | |
655 | seed.value), | |
656 | match_fd, | |
657 | (void *) (unsigned long) out_fds[i], | |
658 | &entry->tracker_node); | |
659 | ||
660 | if (node != &entry->tracker_node) { | |
661 | ret = -EEXIST; | |
662 | rcu_read_unlock(); | |
663 | goto end_free_entries; | |
664 | } | |
665 | entries[i] = NULL; | |
666 | } | |
667 | tracker->count.unsuspendable += fd_count; | |
668 | rcu_read_unlock(); | |
669 | ret = user_ret; | |
670 | end: | |
671 | pthread_mutex_unlock(&tracker->lock); | |
672 | return ret; | |
673 | end_free_entries: | |
674 | for (i = 0; i < fd_count; i++) { | |
675 | unsuspendable_fd_destroy(entries[i]); | |
676 | } | |
677 | goto end; | |
678 | } | |
679 | ||
680 | int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker, | |
681 | int *fds_in, unsigned int fd_count, fd_close_cb close, | |
682 | void *user_data) | |
683 | { | |
684 | int i, ret, user_ret; | |
685 | int fds[fd_count]; | |
686 | ||
687 | /* | |
688 | * Maintain a local copy of fds_in as the user's callback may modify its | |
689 | * contents (e.g. setting the fd(s) to -1 after close). | |
690 | */ | |
691 | memcpy(fds, fds_in, sizeof(*fds) * fd_count); | |
692 | ||
693 | pthread_mutex_lock(&tracker->lock); | |
694 | rcu_read_lock(); | |
695 | ||
696 | /* Let the user close the file descriptors. */ | |
697 | user_ret = close(user_data, fds_in); | |
698 | if (user_ret) { | |
699 | ret = user_ret; | |
700 | goto end; | |
701 | } | |
702 | ||
703 | /* Untrack the fds that were just closed by the user's callback. */ | |
704 | for (i = 0; i < fd_count; i++) { | |
705 | struct cds_lfht_node *node; | |
706 | struct cds_lfht_iter iter; | |
707 | struct unsuspendable_fd *entry; | |
708 | ||
709 | cds_lfht_lookup(tracker->unsuspendable_fds, | |
710 | hash_key_ulong((void *) (unsigned long) fds[i], | |
711 | seed.value), | |
712 | match_fd, | |
713 | (void *) (unsigned long) fds[i], | |
714 | &iter); | |
715 | node = cds_lfht_iter_get_node(&iter); | |
716 | if (!node) { | |
717 | /* Unknown file descriptor. */ | |
718 | WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()", | |
719 | fds[i]); | |
720 | ret = -EINVAL; | |
721 | goto end; | |
722 | } | |
723 | entry = caa_container_of(node, | |
724 | struct unsuspendable_fd, | |
725 | tracker_node); | |
726 | ||
727 | cds_lfht_del(tracker->unsuspendable_fds, node); | |
728 | unsuspendable_fd_destroy(entry); | |
729 | fds[i] = -1; | |
730 | } | |
731 | ||
732 | tracker->count.unsuspendable -= fd_count; | |
733 | ret = 0; | |
734 | end: | |
735 | rcu_read_unlock(); | |
736 | pthread_mutex_unlock(&tracker->lock); | |
737 | return ret; | |
738 | } | |
739 | ||
740 | /* Caller must have taken the tracker's and handle's locks. */ | |
741 | static | |
742 | void fd_tracker_track(struct fd_tracker *tracker, struct fs_handle *handle) | |
743 | { | |
744 | if (handle->fd >= 0) { | |
745 | tracker->count.suspendable.active++; | |
746 | cds_list_add_tail(&handle->handles_list_node, | |
747 | &tracker->active_handles); | |
748 | } else { | |
749 | tracker->count.suspendable.suspended++; | |
750 | cds_list_add_tail(&handle->handles_list_node, | |
751 | &tracker->suspended_handles); | |
752 | } | |
753 | } | |
754 | ||
755 | /* Caller must have taken the tracker's and handle's locks. */ | |
756 | static | |
757 | void fd_tracker_untrack(struct fd_tracker *tracker, struct fs_handle *handle) | |
758 | { | |
759 | if (handle->fd >= 0) { | |
760 | tracker->count.suspendable.active--; | |
761 | } else { | |
762 | tracker->count.suspendable.suspended--; | |
763 | } | |
764 | cds_list_del(&handle->handles_list_node); | |
765 | } | |
766 | ||
767 | /* Caller must have taken the tracker's and handle's locks. */ | |
768 | static | |
769 | int fd_tracker_restore_handle(struct fd_tracker *tracker, | |
770 | struct fs_handle *handle) | |
771 | { | |
772 | int ret; | |
773 | ||
774 | fd_tracker_untrack(tracker, handle); | |
775 | if (ACTIVE_COUNT(tracker) >= tracker->capacity) { | |
776 | ret = fd_tracker_suspend_handles(tracker, 1); | |
777 | if (ret) { | |
778 | goto end; | |
779 | } | |
780 | } | |
781 | ret = fs_handle_restore(handle); | |
782 | end: | |
783 | fd_tracker_track(tracker, handle); | |
784 | return ret ? ret : handle->fd; | |
785 | } | |
786 | ||
787 | int fs_handle_get_fd(struct fs_handle *handle) | |
788 | { | |
789 | int ret; | |
790 | ||
97c34b30 JG |
791 | /* |
792 | * TODO This should be optimized as it is a fairly hot path. | |
793 | * The fd-tracker's lock should only be taken when a fs_handle is | |
794 | * restored (slow path). On the fast path (fs_handle is active), | |
795 | * the only effect on the fd_tracker is marking the handle as the | |
796 | * most recently used. Currently, it is done by a call to the | |
797 | * track/untrack helpers, but it should be done atomically. | |
798 | * | |
799 | * Note that the lock's nesting order must still be respected here. | |
800 | * The handle's lock nests inside the tracker's lock. | |
801 | */ | |
0ad3c26c JG |
802 | pthread_mutex_lock(&handle->tracker->lock); |
803 | pthread_mutex_lock(&handle->lock); | |
804 | assert(!handle->in_use); | |
805 | ||
806 | handle->tracker->stats.uses++; | |
807 | if (handle->fd >= 0) { | |
808 | ret = handle->fd; | |
809 | /* Mark as most recently used. */ | |
810 | fd_tracker_untrack(handle->tracker, handle); | |
811 | fd_tracker_track(handle->tracker, handle); | |
812 | } else { | |
813 | handle->tracker->stats.misses++; | |
814 | ret = fd_tracker_restore_handle(handle->tracker, handle); | |
815 | if (ret < 0) { | |
816 | handle->tracker->stats.errors++; | |
817 | goto end; | |
818 | } | |
819 | } | |
820 | handle->in_use = true; | |
821 | end: | |
822 | pthread_mutex_unlock(&handle->lock); | |
823 | pthread_mutex_unlock(&handle->tracker->lock); | |
824 | return ret; | |
825 | } | |
826 | ||
827 | void fs_handle_put_fd(struct fs_handle *handle) | |
828 | { | |
829 | pthread_mutex_lock(&handle->lock); | |
830 | handle->in_use = false; | |
831 | pthread_mutex_unlock(&handle->lock); | |
832 | } | |
833 | ||
cf7f2fd5 JG |
834 | int fs_handle_unlink(struct fs_handle *handle) |
835 | { | |
836 | int ret; | |
837 | ||
838 | pthread_mutex_lock(&handle->tracker->lock); | |
839 | pthread_mutex_lock(&handle->lock); | |
840 | ret = lttng_inode_defer_unlink(handle->inode); | |
841 | pthread_mutex_unlock(&handle->lock); | |
842 | pthread_mutex_unlock(&handle->tracker->lock); | |
843 | return ret; | |
844 | } | |
845 | ||
0ad3c26c JG |
846 | int fs_handle_close(struct fs_handle *handle) |
847 | { | |
848 | int ret = 0; | |
52803bf0 | 849 | const char *path = NULL; |
0ad3c26c JG |
850 | |
851 | if (!handle) { | |
852 | ret = -EINVAL; | |
853 | goto end; | |
854 | } | |
855 | ||
856 | pthread_mutex_lock(&handle->tracker->lock); | |
857 | pthread_mutex_lock(&handle->lock); | |
52803bf0 JG |
858 | if (handle->inode) { |
859 | path = lttng_inode_get_path(handle->inode); | |
860 | } | |
0ad3c26c JG |
861 | fd_tracker_untrack(handle->tracker, handle); |
862 | if (handle->fd >= 0) { | |
bdafebb5 | 863 | assert(!handle->in_use); |
0ad3c26c JG |
864 | /* |
865 | * The return value of close() is not propagated as there | |
866 | * isn't much the user can do about it. | |
867 | */ | |
868 | if (close(handle->fd)) { | |
869 | PERROR("Failed to close the file descritptor (%d) of fs handle to %s, close() returned", | |
52803bf0 | 870 | handle->fd, path ? path : "Unknown"); |
0ad3c26c JG |
871 | } |
872 | handle->fd = -1; | |
873 | } | |
6f89b613 | 874 | lttng_inode_put(handle->inode); |
0ad3c26c JG |
875 | pthread_mutex_unlock(&handle->lock); |
876 | pthread_mutex_destroy(&handle->lock); | |
877 | pthread_mutex_unlock(&handle->tracker->lock); | |
0ad3c26c JG |
878 | free(handle); |
879 | end: | |
880 | return ret; | |
881 | } |