3740 Poor ZFS send / receive performance due to snapshot hold / release processing
Submitted by: Steven Hartland <steven.hartland@multiplay.co.uk>
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 */
25
26 /*
27 * LibZFS_Core (lzc) is intended to replace most functionality in libzfs.
28 * It has the following characteristics:
29 *
30 * - Thread Safe. libzfs_core is accessible concurrently from multiple
31 * threads. This is accomplished primarily by avoiding global data
32 * (e.g. caching). Since it's thread-safe, there is no reason for a
33 * process to have multiple libzfs "instances". Therefore, we store
34 * our few pieces of data (e.g. the file descriptor) in global
35 * variables. The fd is reference-counted so that the libzfs_core
36 * library can be "initialized" multiple times (e.g. by different
37 * consumers within the same process).
38 *
39 * - Committed Interface. The libzfs_core interface will be committed,
40 * therefore consumers can compile against it and be confident that
41 * their code will continue to work on future releases of this code.
42 * Currently, the interface is Evolving (not Committed), but we intend
43 * to commit to it once it is more complete and we determine that it
44 * meets the needs of all consumers.
45 *
46 * - Programatic Error Handling. libzfs_core communicates errors with
47 * defined error numbers, and doesn't print anything to stdout/stderr.
48 *
49 * - Thin Layer. libzfs_core is a thin layer, marshaling arguments
50 * to/from the kernel ioctls. There is generally a 1:1 correspondence
51 * between libzfs_core functions and ioctls to /dev/zfs.
52 *
53 * - Clear Atomicity. Because libzfs_core functions are generally 1:1
54 * with kernel ioctls, and kernel ioctls are general atomic, each
55 * libzfs_core function is atomic. For example, creating multiple
56 * snapshots with a single call to lzc_snapshot() is atomic -- it
57 * can't fail with only some of the requested snapshots created, even
58 * in the event of power loss or system crash.
59 *
60 * - Continued libzfs Support. Some higher-level operations (e.g.
61 * support for "zfs send -R") are too complicated to fit the scope of
62 * libzfs_core. This functionality will continue to live in libzfs.
63 * Where appropriate, libzfs will use the underlying atomic operations
64 * of libzfs_core. For example, libzfs may implement "zfs send -R |
65 * zfs receive" by using individual "send one snapshot", rename,
66 * destroy, and "receive one snapshot" operations in libzfs_core.
67 * /sbin/zfs and /zbin/zpool will link with both libzfs and
68 * libzfs_core. Other consumers should aim to use only libzfs_core,
69 * since that will be the supported, stable interface going forwards.
70 */
71
72 #include <libzfs_core.h>
73 #include <ctype.h>
74 #include <unistd.h>
75 #include <stdlib.h>
76 #include <string.h>
77 #include <errno.h>
78 #include <fcntl.h>
79 #include <pthread.h>
80 #include <sys/nvpair.h>
81 #include <sys/param.h>
82 #include <sys/types.h>
83 #include <sys/stat.h>
84 #include <sys/zfs_ioctl.h>
85
86 static int g_fd;
87 static pthread_mutex_t g_lock = PTHREAD_MUTEX_INITIALIZER;
88 static int g_refcount;
89
90 int
91 libzfs_core_init(void)
92 {
93 (void) pthread_mutex_lock(&g_lock);
94 if (g_refcount == 0) {
95 g_fd = open("/dev/zfs", O_RDWR);
96 if (g_fd < 0) {
97 (void) pthread_mutex_unlock(&g_lock);
98 return (errno);
99 }
100 }
101 g_refcount++;
102 (void) pthread_mutex_unlock(&g_lock);
103 return (0);
104 }
105
106 void
107 libzfs_core_fini(void)
108 {
109 (void) pthread_mutex_lock(&g_lock);
110 ASSERT3S(g_refcount, >, 0);
111 g_refcount--;
112 if (g_refcount == 0)
113 (void) close(g_fd);
114 (void) pthread_mutex_unlock(&g_lock);
115 }
116
117 static int
118 lzc_ioctl(zfs_ioc_t ioc, const char *name,
119 nvlist_t *source, nvlist_t **resultp)
120 {
121 zfs_cmd_t zc = { 0 };
122 int error = 0;
123 char *packed;
124 size_t size;
125
126 ASSERT3S(g_refcount, >, 0);
127
128 (void) strlcpy(zc.zc_name, name, sizeof (zc.zc_name));
129
130 packed = fnvlist_pack(source, &size);
131 zc.zc_nvlist_src = (uint64_t)(uintptr_t)packed;
132 zc.zc_nvlist_src_size = size;
133
134 if (resultp != NULL) {
135 *resultp = NULL;
136 zc.zc_nvlist_dst_size = MAX(size * 2, 128 * 1024);
137 zc.zc_nvlist_dst = (uint64_t)(uintptr_t)
138 malloc(zc.zc_nvlist_dst_size);
139 if (zc.zc_nvlist_dst == NULL) {
140 error = ENOMEM;
141 goto out;
142 }
143 }
144
145 while (ioctl(g_fd, ioc, &zc) != 0) {
146 if (errno == ENOMEM && resultp != NULL) {
147 free((void *)(uintptr_t)zc.zc_nvlist_dst);
148 zc.zc_nvlist_dst_size *= 2;
149 zc.zc_nvlist_dst = (uint64_t)(uintptr_t)
150 malloc(zc.zc_nvlist_dst_size);
151 if (zc.zc_nvlist_dst == NULL) {
152 error = ENOMEM;
153 goto out;
154 }
155 } else {
156 error = errno;
157 break;
158 }
159 }
160 if (zc.zc_nvlist_dst_filled) {
161 *resultp = fnvlist_unpack((void *)(uintptr_t)zc.zc_nvlist_dst,
162 zc.zc_nvlist_dst_size);
163 }
164
165 out:
166 fnvlist_pack_free(packed, size);
167 free((void *)(uintptr_t)zc.zc_nvlist_dst);
168 return (error);
169 }
170
171 int
172 lzc_create(const char *fsname, dmu_objset_type_t type, nvlist_t *props)
173 {
174 int error;
175 nvlist_t *args = fnvlist_alloc();
176 fnvlist_add_int32(args, "type", type);
177 if (props != NULL)
178 fnvlist_add_nvlist(args, "props", props);
179 error = lzc_ioctl(ZFS_IOC_CREATE, fsname, args, NULL);
180 nvlist_free(args);
181 return (error);
182 }
183
184 int
185 lzc_clone(const char *fsname, const char *origin,
186 nvlist_t *props)
187 {
188 int error;
189 nvlist_t *args = fnvlist_alloc();
190 fnvlist_add_string(args, "origin", origin);
191 if (props != NULL)
192 fnvlist_add_nvlist(args, "props", props);
193 error = lzc_ioctl(ZFS_IOC_CLONE, fsname, args, NULL);
194 nvlist_free(args);
195 return (error);
196 }
197
198 /*
199 * Creates snapshots.
200 *
201 * The keys in the snaps nvlist are the snapshots to be created.
202 * They must all be in the same pool.
203 *
204 * The props nvlist is properties to set. Currently only user properties
205 * are supported. { user:prop_name -> string value }
206 *
207 * The returned results nvlist will have an entry for each snapshot that failed.
208 * The value will be the (int32) error code.
209 *
210 * The return value will be 0 if all snapshots were created, otherwise it will
211 * be the errno of a (unspecified) snapshot that failed.
212 */
213 int
214 lzc_snapshot(nvlist_t *snaps, nvlist_t *props, nvlist_t **errlist)
215 {
216 nvpair_t *elem;
217 nvlist_t *args;
218 int error;
219 char pool[MAXNAMELEN];
220
221 *errlist = NULL;
222
223 /* determine the pool name */
224 elem = nvlist_next_nvpair(snaps, NULL);
225 if (elem == NULL)
226 return (0);
227 (void) strlcpy(pool, nvpair_name(elem), sizeof (pool));
228 pool[strcspn(pool, "/@")] = '\0';
229
230 args = fnvlist_alloc();
231 fnvlist_add_nvlist(args, "snaps", snaps);
232 if (props != NULL)
233 fnvlist_add_nvlist(args, "props", props);
234
235 error = lzc_ioctl(ZFS_IOC_SNAPSHOT, pool, args, errlist);
236 nvlist_free(args);
237
238 return (error);
239 }
240
241 /*
242 * Destroys snapshots.
243 *
244 * The keys in the snaps nvlist are the snapshots to be destroyed.
245 * They must all be in the same pool.
246 *
247 * Snapshots that do not exist will be silently ignored.
248 *
249 * If 'defer' is not set, and a snapshot has user holds or clones, the
250 * destroy operation will fail and none of the snapshots will be
251 * destroyed.
252 *
253 * If 'defer' is set, and a snapshot has user holds or clones, it will be
254 * marked for deferred destruction, and will be destroyed when the last hold
255 * or clone is removed/destroyed.
256 *
257 * The return value will be 0 if all snapshots were destroyed (or marked for
258 * later destruction if 'defer' is set) or didn't exist to begin with.
259 *
260 * Otherwise the return value will be the errno of a (unspecified) snapshot
261 * that failed, no snapshots will be destroyed, and the errlist will have an
262 * entry for each snapshot that failed. The value in the errlist will be
263 * the (int32) error code.
264 */
265 int
266 lzc_destroy_snaps(nvlist_t *snaps, boolean_t defer, nvlist_t **errlist)
267 {
268 nvpair_t *elem;
269 nvlist_t *args;
270 int error;
271 char pool[MAXNAMELEN];
272
273 /* determine the pool name */
274 elem = nvlist_next_nvpair(snaps, NULL);
275 if (elem == NULL)
276 return (0);
277 (void) strlcpy(pool, nvpair_name(elem), sizeof (pool));
278 pool[strcspn(pool, "/@")] = '\0';
279
280 args = fnvlist_alloc();
281 fnvlist_add_nvlist(args, "snaps", snaps);
282 if (defer)
283 fnvlist_add_boolean(args, "defer");
284
285 error = lzc_ioctl(ZFS_IOC_DESTROY_SNAPS, pool, args, errlist);
286 nvlist_free(args);
287
288 return (error);
289
290 }
291
292 int
293 lzc_snaprange_space(const char *firstsnap, const char *lastsnap,
294 uint64_t *usedp)
295 {
296 nvlist_t *args;
297 nvlist_t *result;
298 int err;
299 char fs[MAXNAMELEN];
300 char *atp;
301
302 /* determine the fs name */
303 (void) strlcpy(fs, firstsnap, sizeof (fs));
304 atp = strchr(fs, '@');
305 if (atp == NULL)
306 return (EINVAL);
307 *atp = '\0';
308
309 args = fnvlist_alloc();
310 fnvlist_add_string(args, "firstsnap", firstsnap);
311
312 err = lzc_ioctl(ZFS_IOC_SPACE_SNAPS, lastsnap, args, &result);
313 nvlist_free(args);
314 if (err == 0)
315 *usedp = fnvlist_lookup_uint64(result, "used");
316 fnvlist_free(result);
317
318 return (err);
319 }
320
321 boolean_t
322 lzc_exists(const char *dataset)
323 {
324 /*
325 * The objset_stats ioctl is still legacy, so we need to construct our
326 * own zfs_cmd_t rather than using zfsc_ioctl().
327 */
328 zfs_cmd_t zc = { 0 };
329
330 (void) strlcpy(zc.zc_name, dataset, sizeof (zc.zc_name));
331 return (ioctl(g_fd, ZFS_IOC_OBJSET_STATS, &zc) == 0);
332 }
333
334 /*
335 * Create "user holds" on snapshots. If there is a hold on a snapshot,
336 * the snapshot can not be destroyed. (However, it can be marked for deletion
337 * by lzc_destroy_snaps(defer=B_TRUE).)
338 *
339 * The keys in the nvlist are snapshot names.
340 * The snapshots must all be in the same pool.
341 * The value is the name of the hold (string type).
342 *
343 * If cleanup_fd is not -1, it must be the result of open("/dev/zfs", O_EXCL).
344 * In this case, when the cleanup_fd is closed (including on process
345 * termination), the holds will be released. If the system is shut down
346 * uncleanly, the holds will be released when the pool is next opened
347 * or imported.
348 *
349 * Holds for snapshots which don't exist will be skipped and have an entry
350 * added to errlist, but will not cause an overall failure, except in the
351 * case that all holds where skipped.
352 *
353 * The return value will be 0 if the nvl holds was empty or all holds, for
354 * snapshots that existed, were succesfully created and at least one hold
355 * was created.
356 *
357 * If none of the snapshots for the requested holds existed ENOENT will be
358 * returned.
359 *
360 * Otherwise the return value will be the errno of a (unspecified) hold that
361 * failed, no holds will be created.
362 *
363 * In all cases the errlist will have an entry for each hold that failed
364 * (name = snapshot), with its value being the error code (int32).
365 */
366 int
367 lzc_hold(nvlist_t *holds, int cleanup_fd, nvlist_t **errlist)
368 {
369 char pool[MAXNAMELEN];
370 nvlist_t *args;
371 nvpair_t *elem;
372 int error;
373
374 /* determine the pool name */
375 elem = nvlist_next_nvpair(holds, NULL);
376 if (elem == NULL)
377 return (0);
378 (void) strlcpy(pool, nvpair_name(elem), sizeof (pool));
379 pool[strcspn(pool, "/@")] = '\0';
380
381 args = fnvlist_alloc();
382 fnvlist_add_nvlist(args, "holds", holds);
383 if (cleanup_fd != -1)
384 fnvlist_add_int32(args, "cleanup_fd", cleanup_fd);
385
386 error = lzc_ioctl(ZFS_IOC_HOLD, pool, args, errlist);
387 nvlist_free(args);
388 return (error);
389 }
390
391 /*
392 * Release "user holds" on snapshots. If the snapshot has been marked for
393 * deferred destroy (by lzc_destroy_snaps(defer=B_TRUE)), it does not have
394 * any clones, and all the user holds are removed, then the snapshot will be
395 * destroyed.
396 *
397 * The keys in the nvlist are snapshot names.
398 * The snapshots must all be in the same pool.
399 * The value is a nvlist whose keys are the holds to remove.
400 *
401 * Holds which failed to release because they didn't exist will have an entry
402 * added to errlist, but will not cause an overall failure.
403 *
404 * The return value will be 0 if the nvl holds was empty or all holds, that
405 * existed, were succesfully removed and at least one hold was removed.
406 *
407 * If none of the holds specified existed ENOENT will be returned.
408 *
409 * Otherwise the return value will be the errno of a (unspecified) hold that
410 * failed to release and no holds will be released.
411 *
412 * In all cases the errlist will have an entry for each hold that failed to
413 * to release.
414 */
415 int
416 lzc_release(nvlist_t *holds, nvlist_t **errlist)
417 {
418 char pool[MAXNAMELEN];
419 nvpair_t *elem;
420
421 /* determine the pool name */
422 elem = nvlist_next_nvpair(holds, NULL);
423 if (elem == NULL)
424 return (0);
425 (void) strlcpy(pool, nvpair_name(elem), sizeof (pool));
426 pool[strcspn(pool, "/@")] = '\0';
427
428 return (lzc_ioctl(ZFS_IOC_RELEASE, pool, holds, errlist));
429 }
430
431 /*
432 * Retrieve list of user holds on the specified snapshot.
433 *
434 * On success, *holdsp will be set to a nvlist which the caller must free.
435 * The keys are the names of the holds, and the value is the creation time
436 * of the hold (uint64) in seconds since the epoch.
437 */
438 int
439 lzc_get_holds(const char *snapname, nvlist_t **holdsp)
440 {
441 int error;
442 nvlist_t *innvl = fnvlist_alloc();
443 error = lzc_ioctl(ZFS_IOC_GET_HOLDS, snapname, innvl, holdsp);
444 fnvlist_free(innvl);
445 return (error);
446 }
447
448 /*
449 * If fromsnap is NULL, a full (non-incremental) stream will be sent.
450 */
451 int
452 lzc_send(const char *snapname, const char *fromsnap, int fd)
453 {
454 nvlist_t *args;
455 int err;
456
457 args = fnvlist_alloc();
458 fnvlist_add_int32(args, "fd", fd);
459 if (fromsnap != NULL)
460 fnvlist_add_string(args, "fromsnap", fromsnap);
461 err = lzc_ioctl(ZFS_IOC_SEND_NEW, snapname, args, NULL);
462 nvlist_free(args);
463 return (err);
464 }
465
466 /*
467 * If fromsnap is NULL, a full (non-incremental) stream will be estimated.
468 */
469 int
470 lzc_send_space(const char *snapname, const char *fromsnap, uint64_t *spacep)
471 {
472 nvlist_t *args;
473 nvlist_t *result;
474 int err;
475
476 args = fnvlist_alloc();
477 if (fromsnap != NULL)
478 fnvlist_add_string(args, "fromsnap", fromsnap);
479 err = lzc_ioctl(ZFS_IOC_SEND_SPACE, snapname, args, &result);
480 nvlist_free(args);
481 if (err == 0)
482 *spacep = fnvlist_lookup_uint64(result, "space");
483 nvlist_free(result);
484 return (err);
485 }
486
487 static int
488 recv_read(int fd, void *buf, int ilen)
489 {
490 char *cp = buf;
491 int rv;
492 int len = ilen;
493
494 do {
495 rv = read(fd, cp, len);
496 cp += rv;
497 len -= rv;
498 } while (rv > 0);
499
500 if (rv < 0 || len != 0)
501 return (EIO);
502
503 return (0);
504 }
505
506 /*
507 * The simplest receive case: receive from the specified fd, creating the
508 * specified snapshot. Apply the specified properties a "received" properties
509 * (which can be overridden by locally-set properties). If the stream is a
510 * clone, its origin snapshot must be specified by 'origin'. The 'force'
511 * flag will cause the target filesystem to be rolled back or destroyed if
512 * necessary to receive.
513 *
514 * Return 0 on success or an errno on failure.
515 *
516 * Note: this interface does not work on dedup'd streams
517 * (those with DMU_BACKUP_FEATURE_DEDUP).
518 */
519 int
520 lzc_receive(const char *snapname, nvlist_t *props, const char *origin,
521 boolean_t force, int fd)
522 {
523 /*
524 * The receive ioctl is still legacy, so we need to construct our own
525 * zfs_cmd_t rather than using zfsc_ioctl().
526 */
527 zfs_cmd_t zc = { 0 };
528 char *atp;
529 char *packed = NULL;
530 size_t size;
531 dmu_replay_record_t drr;
532 int error;
533
534 ASSERT3S(g_refcount, >, 0);
535
536 /* zc_name is name of containing filesystem */
537 (void) strlcpy(zc.zc_name, snapname, sizeof (zc.zc_name));
538 atp = strchr(zc.zc_name, '@');
539 if (atp == NULL)
540 return (EINVAL);
541 *atp = '\0';
542
543 /* if the fs does not exist, try its parent. */
544 if (!lzc_exists(zc.zc_name)) {
545 char *slashp = strrchr(zc.zc_name, '/');
546 if (slashp == NULL)
547 return (ENOENT);
548 *slashp = '\0';
549
550 }
551
552 /* zc_value is full name of the snapshot to create */
553 (void) strlcpy(zc.zc_value, snapname, sizeof (zc.zc_value));
554
555 if (props != NULL) {
556 /* zc_nvlist_src is props to set */
557 packed = fnvlist_pack(props, &size);
558 zc.zc_nvlist_src = (uint64_t)(uintptr_t)packed;
559 zc.zc_nvlist_src_size = size;
560 }
561
562 /* zc_string is name of clone origin (if DRR_FLAG_CLONE) */
563 if (origin != NULL)
564 (void) strlcpy(zc.zc_string, origin, sizeof (zc.zc_string));
565
566 /* zc_begin_record is non-byteswapped BEGIN record */
567 error = recv_read(fd, &drr, sizeof (drr));
568 if (error != 0)
569 goto out;
570 zc.zc_begin_record = drr.drr_u.drr_begin;
571
572 /* zc_cookie is fd to read from */
573 zc.zc_cookie = fd;
574
575 /* zc guid is force flag */
576 zc.zc_guid = force;
577
578 /* zc_cleanup_fd is unused */
579 zc.zc_cleanup_fd = -1;
580
581 error = ioctl(g_fd, ZFS_IOC_RECV, &zc);
582 if (error != 0)
583 error = errno;
584
585 out:
586 if (packed != NULL)
587 fnvlist_pack_free(packed, size);
588 free((void*)(uintptr_t)zc.zc_nvlist_dst);
589 return (error);
590 }
--- EOF ---