Commit | Line | Data |
---|---|---|
1dc7b960 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 | ||
18 | #include <urcu/ref.h> | |
19 | #include <urcu.h> | |
20 | #include <urcu/list.h> | |
21 | #include <urcu/rculfhash.h> | |
22 | ||
23 | #include <sys/stat.h> | |
24 | #include <sys/types.h> | |
25 | #include <fcntl.h> | |
26 | #include <stdbool.h> | |
27 | #include <pthread.h> | |
28 | #include <inttypes.h> | |
29 | ||
30 | #include "common/macros.h" | |
31 | #include "common/error.h" | |
32 | #include "common/defaults.h" | |
33 | #include "common/hashtable/utils.h" | |
34 | #include "common/hashtable/hashtable.h" | |
35 | ||
36 | #include "fd-tracker.h" | |
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; | |
88 | }; | |
89 | ||
90 | struct open_properties { | |
91 | char *path; | |
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; | |
116 | int fd; | |
117 | /* inode number of the file at the time of the handle's creation. */ | |
118 | uint64_t ino; | |
119 | bool in_use; | |
120 | /* Offset to which the file should be restored. */ | |
121 | off_t offset; | |
122 | struct cds_list_head handles_list_node; | |
123 | }; | |
124 | ||
125 | struct unsuspendable_fd { | |
126 | /* | |
127 | * Accesses are only performed through the tracker, which is protected | |
128 | * by its own lock. | |
129 | */ | |
130 | int fd; | |
131 | char *name; | |
132 | struct cds_lfht_node tracker_node; | |
133 | }; | |
134 | ||
135 | static struct { | |
136 | pthread_mutex_t lock; | |
137 | bool initialized; | |
138 | unsigned long value; | |
139 | } seed = { | |
140 | .lock = PTHREAD_MUTEX_INITIALIZER, | |
141 | }; | |
142 | ||
143 | static int match_fd(struct cds_lfht_node *node, const void *key); | |
144 | static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry); | |
145 | static struct unsuspendable_fd *unsuspendable_fd_create(const char *name, | |
146 | int fd); | |
147 | static int open_from_properties(struct open_properties *properties); | |
148 | ||
149 | static void fs_handle_log(struct fs_handle *handle); | |
150 | static int fs_handle_suspend(struct fs_handle *handle); | |
151 | static int fs_handle_restore(struct fs_handle *handle); | |
152 | ||
153 | static void fd_tracker_track(struct fd_tracker *tracker, | |
154 | struct fs_handle *handle); | |
155 | static void fd_tracker_untrack(struct fd_tracker *tracker, | |
156 | struct fs_handle *handle); | |
157 | static int fd_tracker_suspend_handles(struct fd_tracker *tracker, | |
158 | unsigned int count); | |
159 | static int fd_tracker_restore_handle(struct fd_tracker *tracker, | |
160 | struct fs_handle *handle); | |
161 | ||
162 | /* Match function of the tracker's unsuspendable_fds hash table. */ | |
163 | static | |
164 | int match_fd(struct cds_lfht_node *node, const void *key) | |
165 | { | |
166 | struct unsuspendable_fd *entry = | |
167 | caa_container_of(node, struct unsuspendable_fd, tracker_node); | |
168 | ||
169 | return hash_match_key_ulong((void *) (unsigned long) entry->fd, | |
170 | (void *) key); | |
171 | } | |
172 | ||
173 | static | |
174 | void unsuspendable_fd_destroy(struct unsuspendable_fd *entry) | |
175 | { | |
176 | if (!entry) { | |
177 | return; | |
178 | } | |
179 | free(entry->name); | |
180 | free(entry); | |
181 | } | |
182 | ||
183 | static | |
184 | struct unsuspendable_fd *unsuspendable_fd_create(const char *name, int fd) | |
185 | { | |
186 | struct unsuspendable_fd *entry = | |
187 | zmalloc(sizeof(*entry)); | |
188 | ||
189 | if (!entry) { | |
190 | goto error; | |
191 | } | |
192 | if (name) { | |
193 | entry->name = strdup(name); | |
194 | if (!entry->name) { | |
195 | goto error; | |
196 | } | |
197 | } | |
198 | cds_lfht_node_init(&entry->tracker_node); | |
199 | entry->fd = fd; | |
200 | return entry; | |
201 | error: | |
202 | unsuspendable_fd_destroy(entry); | |
203 | return NULL; | |
204 | } | |
205 | ||
206 | static | |
207 | void fs_handle_log(struct fs_handle *handle) | |
208 | { | |
209 | pthread_mutex_lock(&handle->lock); | |
210 | if (handle->fd >= 0) { | |
211 | DBG_NO_LOC(" %s [active, fd %d%s]", | |
212 | handle->properties.path, | |
213 | handle->fd, | |
214 | handle->in_use ? ", in use" : ""); | |
215 | } else { | |
216 | DBG_NO_LOC(" %s [suspended]", handle->properties.path); | |
217 | } | |
218 | pthread_mutex_unlock(&handle->lock); | |
219 | } | |
220 | ||
221 | static | |
222 | int fs_handle_suspend(struct fs_handle *handle) | |
223 | { | |
224 | int ret = 0; | |
225 | struct stat fs_stat; | |
226 | ||
227 | pthread_mutex_lock(&handle->lock); | |
228 | assert(handle->fd >= 0); | |
229 | if (handle->in_use) { | |
230 | /* This handle can't be suspended as it is currently in use. */ | |
231 | ret = -EAGAIN; | |
232 | goto end; | |
233 | } | |
234 | ||
235 | ret = stat(handle->properties.path, &fs_stat); | |
236 | if (ret) { | |
237 | PERROR("Filesystem handle to %s cannot be suspended as stat() failed", | |
238 | handle->properties.path); | |
239 | ret = -errno; | |
240 | goto end; | |
241 | } | |
242 | ||
243 | if (fs_stat.st_ino != handle->ino) { | |
244 | /* Don't suspend as the handle would not be restorable. */ | |
245 | WARN("Filesystem handle to %s cannot be suspended as its inode changed", | |
246 | handle->properties.path); | |
247 | ret = -ENOENT; | |
248 | goto end; | |
249 | } | |
250 | ||
251 | handle->offset = lseek(handle->fd, 0, SEEK_CUR); | |
252 | if (handle->offset == -1) { | |
253 | WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position", | |
254 | handle->properties.path); | |
255 | ret = -errno; | |
256 | goto end; | |
257 | } | |
258 | ||
259 | ret = close(handle->fd); | |
260 | if (ret) { | |
261 | PERROR("Filesystem handle to %s cannot be suspended as close() failed", | |
262 | handle->properties.path); | |
263 | ret = -errno; | |
264 | goto end; | |
265 | } | |
266 | DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64, | |
267 | handle->properties.path, handle->fd, handle->offset); | |
268 | handle->fd = -1; | |
269 | end: | |
270 | if (ret) { | |
271 | handle->tracker->stats.errors++; | |
272 | } | |
273 | pthread_mutex_unlock(&handle->lock); | |
274 | return ret; | |
275 | } | |
276 | ||
277 | /* Caller must hold the tracker and handle's locks. */ | |
278 | static | |
279 | int fs_handle_restore(struct fs_handle *handle) | |
280 | { | |
281 | int ret, fd = -1; | |
282 | ||
283 | assert(handle->fd == -1); | |
284 | ret = open_from_properties(&handle->properties); | |
285 | if (ret < 0) { | |
286 | PERROR("Failed to restore filesystem handle to %s, open() failed", | |
287 | handle->properties.path); | |
288 | ret = -errno; | |
289 | goto end; | |
290 | } | |
291 | fd = ret; | |
292 | ||
293 | ret = lseek(fd, handle->offset, SEEK_SET); | |
294 | if (ret < 0) { | |
295 | PERROR("Failed to restore filesystem handle to %s, lseek() failed", | |
296 | handle->properties.path); | |
297 | ret = -errno; | |
298 | goto end; | |
299 | } | |
300 | DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64, | |
301 | handle->properties.path, fd, handle->offset); | |
302 | ret = 0; | |
303 | handle->fd = fd; | |
304 | fd = -1; | |
305 | end: | |
306 | if (fd >= 0) { | |
307 | (void) close(fd); | |
308 | } | |
309 | return ret; | |
310 | } | |
311 | ||
312 | static | |
313 | int open_from_properties(struct open_properties *properties) | |
314 | { | |
315 | int ret; | |
316 | ||
317 | /* | |
318 | * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE | |
319 | * flags are set. O_TMPFILE would not make sense in the context of a | |
320 | * suspendable fs_handle as it would not be restorable (see OPEN(2)), | |
321 | * thus it is ignored here. | |
322 | */ | |
323 | if ((properties->flags & O_CREAT) && properties->mode.is_set) { | |
324 | ret = open(properties->path, properties->flags, | |
325 | properties->mode.value); | |
326 | } else { | |
327 | ret = open(properties->path, properties->flags); | |
328 | } | |
329 | /* | |
330 | * Some flags should not be used beyond the initial open() of a | |
331 | * restorable file system handle. O_CREAT and O_TRUNC must | |
332 | * be cleared since it would be unexpected to re-use them | |
333 | * when the handle is retored: | |
334 | * - O_CREAT should not be needed as the file has been created | |
335 | * on the initial call to open(), | |
336 | * - O_TRUNC would destroy the file's contents by truncating it | |
337 | * to length 0. | |
338 | */ | |
339 | properties->flags &= ~(O_CREAT | O_TRUNC); | |
340 | if (ret < 0) { | |
341 | ret = -errno; | |
342 | goto end; | |
343 | } | |
344 | end: | |
345 | return ret; | |
346 | } | |
347 | ||
348 | struct fd_tracker *fd_tracker_create(unsigned int capacity) | |
349 | { | |
350 | struct fd_tracker *tracker = zmalloc(sizeof(struct fd_tracker)); | |
351 | ||
352 | if (!tracker) { | |
353 | goto end; | |
354 | } | |
355 | ||
356 | pthread_mutex_lock(&seed.lock); | |
357 | if (!seed.initialized) { | |
358 | seed.value = (unsigned long) time(NULL); | |
359 | seed.initialized = true; | |
360 | } | |
361 | pthread_mutex_unlock(&seed.lock); | |
362 | ||
363 | CDS_INIT_LIST_HEAD(&tracker->active_handles); | |
364 | CDS_INIT_LIST_HEAD(&tracker->suspended_handles); | |
365 | tracker->capacity = capacity; | |
366 | tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0, | |
367 | CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL); | |
368 | end: | |
369 | return tracker; | |
370 | } | |
371 | ||
372 | void fd_tracker_log(struct fd_tracker *tracker) | |
373 | { | |
374 | struct fs_handle *handle; | |
375 | struct unsuspendable_fd *unsuspendable_fd; | |
376 | struct cds_lfht_iter iter; | |
377 | ||
378 | pthread_mutex_lock(&tracker->lock); | |
379 | DBG_NO_LOC("File descriptor tracker"); | |
380 | DBG_NO_LOC(" Stats:"); | |
381 | DBG_NO_LOC(" uses: %" PRIu64, tracker->stats.uses); | |
382 | DBG_NO_LOC(" misses: %" PRIu64, tracker->stats.misses); | |
383 | DBG_NO_LOC(" errors: %" PRIu64, tracker->stats.errors); | |
384 | DBG_NO_LOC(" Tracked: %u", TRACKED_COUNT(tracker)); | |
385 | DBG_NO_LOC(" active: %u", ACTIVE_COUNT(tracker)); | |
386 | DBG_NO_LOC(" suspendable: %u", SUSPENDABLE_COUNT(tracker)); | |
387 | DBG_NO_LOC(" unsuspendable: %u", UNSUSPENDABLE_COUNT(tracker)); | |
388 | DBG_NO_LOC(" suspended: %u", SUSPENDED_COUNT(tracker)); | |
389 | DBG_NO_LOC(" capacity: %u", tracker->capacity); | |
390 | ||
391 | DBG_NO_LOC(" Tracked suspendable file descriptors"); | |
392 | cds_list_for_each_entry(handle, &tracker->active_handles, | |
393 | handles_list_node) { | |
394 | fs_handle_log(handle); | |
395 | } | |
396 | cds_list_for_each_entry(handle, &tracker->suspended_handles, | |
397 | handles_list_node) { | |
398 | fs_handle_log(handle); | |
399 | } | |
400 | if (!SUSPENDABLE_COUNT(tracker)) { | |
401 | DBG_NO_LOC(" None"); | |
402 | } | |
403 | ||
404 | DBG_NO_LOC(" Tracked unsuspendable file descriptors"); | |
405 | rcu_read_lock(); | |
406 | cds_lfht_for_each_entry(tracker->unsuspendable_fds, &iter, | |
407 | unsuspendable_fd, tracker_node) { | |
408 | DBG_NO_LOC(" %s [active, fd %d]", unsuspendable_fd->name ? : "Unnamed", | |
409 | unsuspendable_fd->fd); | |
410 | } | |
411 | rcu_read_unlock(); | |
412 | if (!UNSUSPENDABLE_COUNT(tracker)) { | |
413 | DBG_NO_LOC(" None"); | |
414 | } | |
415 | ||
416 | pthread_mutex_unlock(&tracker->lock); | |
417 | } | |
418 | ||
419 | int fd_tracker_destroy(struct fd_tracker *tracker) | |
420 | { | |
421 | int ret = 0; | |
422 | ||
423 | /* | |
424 | * Refuse to destroy the tracker as fs_handles may still old | |
425 | * weak references to the tracker. | |
426 | */ | |
427 | pthread_mutex_lock(&tracker->lock); | |
428 | if (TRACKED_COUNT(tracker)) { | |
429 | ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked", | |
430 | TRACKED_COUNT(tracker)); | |
431 | pthread_mutex_unlock(&tracker->lock); | |
432 | fd_tracker_log(tracker); | |
433 | ret = -1; | |
434 | goto end; | |
435 | } | |
436 | pthread_mutex_unlock(&tracker->lock); | |
437 | ||
438 | ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL); | |
439 | assert(!ret); | |
440 | pthread_mutex_destroy(&tracker->lock); | |
441 | free(tracker); | |
442 | end: | |
443 | return ret; | |
444 | } | |
445 | ||
446 | struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker, | |
447 | const char *path, int flags, mode_t *mode) | |
448 | { | |
449 | int ret; | |
450 | struct fs_handle *handle = NULL; | |
451 | struct stat fd_stat; | |
452 | struct open_properties properties = { | |
453 | .path = strdup(path), | |
454 | .flags = flags, | |
455 | .mode.is_set = !!mode, | |
456 | .mode.value = mode ? *mode : 0, | |
457 | }; | |
458 | ||
459 | if (!properties.path) { | |
460 | goto end; | |
461 | } | |
462 | ||
463 | pthread_mutex_lock(&tracker->lock); | |
464 | if (ACTIVE_COUNT(tracker) == tracker->capacity) { | |
465 | if (tracker->count.suspendable.active > 0) { | |
466 | ret = fd_tracker_suspend_handles(tracker, 1); | |
467 | if (ret) { | |
468 | goto error_destroy; | |
469 | } | |
470 | } else { | |
471 | /* | |
472 | * There are not enough active suspendable file | |
473 | * descriptors to open a new fd and still accomodate the | |
474 | * tracker's capacity. | |
475 | */ | |
476 | WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)", | |
477 | tracker->count.unsuspendable); | |
478 | ret = -EMFILE; | |
479 | goto error_destroy; | |
480 | } | |
481 | } | |
482 | ||
483 | handle = zmalloc(sizeof(*handle)); | |
484 | if (!handle) { | |
485 | goto end; | |
486 | } | |
487 | ||
488 | ret = pthread_mutex_init(&handle->lock, NULL); | |
489 | if (ret) { | |
490 | PERROR("Failed to initialize handle mutex while creating fs handle"); | |
491 | free(handle); | |
492 | goto end; | |
493 | } | |
494 | ||
495 | handle->fd = open_from_properties(&properties); | |
496 | if (handle->fd < 0) { | |
497 | PERROR("Failed to open fs handle to %s, open() returned", path); | |
498 | ret = -errno; | |
499 | goto error_destroy; | |
500 | } | |
501 | ||
502 | /* | |
503 | * Clear the create flag from the open flags as it would make no sense | |
504 | * to use it when restoring a fs handle. | |
505 | */ | |
506 | properties.flags &= ~O_CREAT; | |
507 | handle->properties = properties; | |
508 | properties.path = NULL; | |
509 | ||
510 | if (fstat(handle->fd, &fd_stat)) { | |
511 | PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned"); | |
512 | ret = -errno; | |
513 | goto error_destroy; | |
514 | } | |
515 | handle->ino = fd_stat.st_ino; | |
516 | ||
517 | fd_tracker_track(tracker, handle); | |
518 | handle->tracker = tracker; | |
519 | pthread_mutex_unlock(&tracker->lock); | |
520 | end: | |
521 | free(properties.path); | |
522 | return handle; | |
523 | error_destroy: | |
524 | pthread_mutex_unlock(&tracker->lock); | |
525 | (void) fs_handle_close(handle); | |
526 | handle = NULL; | |
527 | goto end; | |
528 | } | |
529 | ||
530 | /* Caller must hold the tracker's lock. */ | |
531 | static | |
532 | int fd_tracker_suspend_handles(struct fd_tracker *tracker, | |
533 | unsigned int count) | |
534 | { | |
535 | unsigned int left_to_close = count; | |
536 | struct fs_handle *handle, *tmp; | |
537 | ||
538 | cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles, | |
539 | handles_list_node) { | |
540 | int ret; | |
541 | ||
542 | fd_tracker_untrack(tracker, handle); | |
543 | ret = fs_handle_suspend(handle); | |
544 | fd_tracker_track(tracker, handle); | |
545 | if (!ret) { | |
546 | left_to_close--; | |
547 | } | |
548 | ||
549 | if (!left_to_close) { | |
550 | break; | |
551 | } | |
552 | } | |
553 | return left_to_close ? -EMFILE : 0; | |
554 | } | |
555 | ||
556 | int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker, | |
557 | int *out_fds, const char **names, unsigned int fd_count, | |
558 | fd_open_cb open, void *user_data) | |
559 | { | |
560 | int ret, user_ret, i, fds_to_suspend; | |
561 | unsigned int active_fds; | |
562 | struct unsuspendable_fd *entries[fd_count]; | |
563 | ||
564 | memset(entries, 0, sizeof(entries)); | |
565 | ||
566 | pthread_mutex_lock(&tracker->lock); | |
567 | ||
568 | active_fds = ACTIVE_COUNT(tracker); | |
569 | fds_to_suspend = (int) active_fds + (int) fd_count - (int) tracker->capacity; | |
570 | if (fds_to_suspend > 0) { | |
571 | if (fds_to_suspend <= tracker->count.suspendable.active) { | |
572 | ret = fd_tracker_suspend_handles(tracker, fds_to_suspend); | |
573 | if (ret) { | |
574 | goto end; | |
575 | } | |
576 | } else { | |
577 | /* | |
578 | * There are not enough active suspendable file | |
579 | * descriptors to open a new fd and still accomodate the | |
580 | * tracker's capacity. | |
581 | */ | |
582 | WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)", | |
583 | tracker->count.unsuspendable); | |
584 | ret = -EMFILE; | |
585 | goto end; | |
586 | } | |
587 | } | |
588 | ||
589 | user_ret = open(user_data, out_fds); | |
590 | if (user_ret) { | |
591 | ret = user_ret; | |
592 | goto end; | |
593 | } | |
594 | ||
595 | /* | |
596 | * Add the fds returned by the user's callback to the hashtable | |
597 | * of unsuspendable fds. | |
598 | */ | |
599 | for (i = 0; i < fd_count; i++) { | |
600 | struct unsuspendable_fd *entry = | |
601 | unsuspendable_fd_create(names ? names[i] : NULL, | |
602 | out_fds[i]); | |
603 | ||
604 | if (!entry) { | |
605 | ret = -1; | |
606 | goto end_free_entries; | |
607 | } | |
608 | entries[i] = entry; | |
609 | } | |
610 | ||
611 | rcu_read_lock(); | |
612 | for (i = 0; i < fd_count; i++) { | |
613 | struct cds_lfht_node *node; | |
614 | struct unsuspendable_fd *entry = entries[i]; | |
615 | ||
616 | node = cds_lfht_add_unique( | |
617 | tracker->unsuspendable_fds, | |
618 | hash_key_ulong((void *) (unsigned long) out_fds[i], | |
619 | seed.value), | |
620 | match_fd, | |
621 | (void *) (unsigned long) out_fds[i], | |
622 | &entry->tracker_node); | |
623 | ||
624 | if (node != &entry->tracker_node) { | |
625 | ret = -EEXIST; | |
626 | rcu_read_unlock(); | |
627 | goto end_free_entries; | |
628 | } | |
629 | entries[i] = NULL; | |
630 | } | |
631 | tracker->count.unsuspendable += fd_count; | |
632 | rcu_read_unlock(); | |
633 | ret = user_ret; | |
634 | end: | |
635 | pthread_mutex_unlock(&tracker->lock); | |
636 | return ret; | |
637 | end_free_entries: | |
638 | for (i = 0; i < fd_count; i++) { | |
639 | unsuspendable_fd_destroy(entries[i]); | |
640 | } | |
641 | goto end; | |
642 | } | |
643 | ||
644 | int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker, | |
645 | int *fds_in, unsigned int fd_count, fd_close_cb close, | |
646 | void *user_data) | |
647 | { | |
648 | int i, ret, user_ret; | |
649 | int fds[fd_count]; | |
650 | ||
651 | /* | |
652 | * Maintain a local copy of fds_in as the user's callback may modify its | |
653 | * contents (e.g. setting the fd(s) to -1 after close). | |
654 | */ | |
655 | memcpy(fds, fds_in, sizeof(*fds) * fd_count); | |
656 | ||
657 | pthread_mutex_lock(&tracker->lock); | |
658 | rcu_read_lock(); | |
659 | ||
660 | /* Let the user close the file descriptors. */ | |
661 | user_ret = close(user_data, fds_in); | |
662 | if (user_ret) { | |
663 | ret = user_ret; | |
664 | goto end; | |
665 | } | |
666 | ||
667 | /* Untrack the fds that were just closed by the user's callback. */ | |
668 | for (i = 0; i < fd_count; i++) { | |
669 | struct cds_lfht_node *node; | |
670 | struct cds_lfht_iter iter; | |
671 | struct unsuspendable_fd *entry; | |
672 | ||
673 | cds_lfht_lookup(tracker->unsuspendable_fds, | |
674 | hash_key_ulong((void *) (unsigned long) fds[i], | |
675 | seed.value), | |
676 | match_fd, | |
677 | (void *) (unsigned long) fds[i], | |
678 | &iter); | |
679 | node = cds_lfht_iter_get_node(&iter); | |
680 | if (!node) { | |
681 | /* Unknown file descriptor. */ | |
682 | WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()", | |
683 | fds[i]); | |
684 | ret = -EINVAL; | |
685 | goto end; | |
686 | } | |
687 | entry = caa_container_of(node, | |
688 | struct unsuspendable_fd, | |
689 | tracker_node); | |
690 | ||
691 | cds_lfht_del(tracker->unsuspendable_fds, node); | |
692 | unsuspendable_fd_destroy(entry); | |
693 | fds[i] = -1; | |
694 | } | |
695 | ||
696 | tracker->count.unsuspendable -= fd_count; | |
697 | ret = 0; | |
698 | end: | |
699 | rcu_read_unlock(); | |
700 | pthread_mutex_unlock(&tracker->lock); | |
701 | return ret; | |
702 | } | |
703 | ||
704 | /* Caller must have taken the tracker's and handle's locks. */ | |
705 | static | |
706 | void fd_tracker_track(struct fd_tracker *tracker, struct fs_handle *handle) | |
707 | { | |
708 | if (handle->fd >= 0) { | |
709 | tracker->count.suspendable.active++; | |
710 | cds_list_add_tail(&handle->handles_list_node, | |
711 | &tracker->active_handles); | |
712 | } else { | |
713 | tracker->count.suspendable.suspended++; | |
714 | cds_list_add_tail(&handle->handles_list_node, | |
715 | &tracker->suspended_handles); | |
716 | } | |
717 | } | |
718 | ||
719 | /* Caller must have taken the tracker's and handle's locks. */ | |
720 | static | |
721 | void fd_tracker_untrack(struct fd_tracker *tracker, struct fs_handle *handle) | |
722 | { | |
723 | if (handle->fd >= 0) { | |
724 | tracker->count.suspendable.active--; | |
725 | } else { | |
726 | tracker->count.suspendable.suspended--; | |
727 | } | |
728 | cds_list_del(&handle->handles_list_node); | |
729 | } | |
730 | ||
731 | /* Caller must have taken the tracker's and handle's locks. */ | |
732 | static | |
733 | int fd_tracker_restore_handle(struct fd_tracker *tracker, | |
734 | struct fs_handle *handle) | |
735 | { | |
736 | int ret; | |
737 | ||
738 | fd_tracker_untrack(tracker, handle); | |
739 | if (ACTIVE_COUNT(tracker) >= tracker->capacity) { | |
740 | ret = fd_tracker_suspend_handles(tracker, 1); | |
741 | if (ret) { | |
742 | goto end; | |
743 | } | |
744 | } | |
745 | ret = fs_handle_restore(handle); | |
746 | end: | |
747 | fd_tracker_track(tracker, handle); | |
748 | return ret ? ret : handle->fd; | |
749 | } | |
750 | ||
751 | int fs_handle_get_fd(struct fs_handle *handle) | |
752 | { | |
753 | int ret; | |
754 | ||
755 | pthread_mutex_lock(&handle->tracker->lock); | |
756 | pthread_mutex_lock(&handle->lock); | |
757 | assert(!handle->in_use); | |
758 | ||
759 | handle->tracker->stats.uses++; | |
760 | if (handle->fd >= 0) { | |
761 | ret = handle->fd; | |
762 | /* Mark as most recently used. */ | |
763 | fd_tracker_untrack(handle->tracker, handle); | |
764 | fd_tracker_track(handle->tracker, handle); | |
765 | } else { | |
766 | handle->tracker->stats.misses++; | |
767 | ret = fd_tracker_restore_handle(handle->tracker, handle); | |
768 | if (ret < 0) { | |
769 | handle->tracker->stats.errors++; | |
770 | goto end; | |
771 | } | |
772 | } | |
773 | handle->in_use = true; | |
774 | end: | |
775 | pthread_mutex_unlock(&handle->lock); | |
776 | pthread_mutex_unlock(&handle->tracker->lock); | |
777 | return ret; | |
778 | } | |
779 | ||
780 | void fs_handle_put_fd(struct fs_handle *handle) | |
781 | { | |
782 | pthread_mutex_lock(&handle->lock); | |
783 | handle->in_use = false; | |
784 | pthread_mutex_unlock(&handle->lock); | |
785 | } | |
786 | ||
787 | int fs_handle_close(struct fs_handle *handle) | |
788 | { | |
789 | int ret = 0; | |
790 | ||
791 | if (!handle) { | |
792 | ret = -EINVAL; | |
793 | goto end; | |
794 | } | |
795 | ||
796 | pthread_mutex_lock(&handle->tracker->lock); | |
797 | pthread_mutex_lock(&handle->lock); | |
798 | fd_tracker_untrack(handle->tracker, handle); | |
799 | if (handle->fd >= 0) { | |
800 | /* | |
801 | * The return value of close() is not propagated as there | |
802 | * isn't much the user can do about it. | |
803 | */ | |
804 | if (close(handle->fd)) { | |
805 | PERROR("Failed to close the file descritptor (%d) of fs handle to %s, close() returned", | |
806 | handle->fd, handle->properties.path); | |
807 | } | |
808 | handle->fd = -1; | |
809 | } | |
810 | pthread_mutex_unlock(&handle->lock); | |
811 | pthread_mutex_destroy(&handle->lock); | |
812 | pthread_mutex_unlock(&handle->tracker->lock); | |
813 | free(handle->properties.path); | |
814 | free(handle); | |
815 | end: | |
816 | return ret; | |
817 | } |