xref: /freebsd-13-stable/sys/contrib/openzfs/module/zfs/dsl_bookmark.c (revision f193a24ec57067da831f732865e5871e311704af)
1 /*
2  * CDDL HEADER START
3  *
4  * This file and its contents are supplied under the terms of the
5  * Common Development and Distribution License ("CDDL"), version 1.0.
6  * You may only use this file in accordance with the terms of version
7  * 1.0 of the CDDL.
8  *
9  * A full copy of the text of the CDDL should have accompanied this
10  * source.  A copy of the CDDL is also available via the Internet at
11  * http://www.illumos.org/license/CDDL.
12  *
13  * CDDL HEADER END
14  */
15 
16 /*
17  * Copyright (c) 2013, 2018 by Delphix. All rights reserved.
18  * Copyright 2017 Nexenta Systems, Inc.
19  * Copyright 2019, 2020 by Christian Schwarz. All rights reserved.
20  */
21 
22 #include <sys/zfs_context.h>
23 #include <sys/dsl_dataset.h>
24 #include <sys/dsl_dir.h>
25 #include <sys/dsl_prop.h>
26 #include <sys/dsl_synctask.h>
27 #include <sys/dsl_destroy.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/arc.h>
31 #include <sys/zap.h>
32 #include <sys/zfeature.h>
33 #include <sys/spa.h>
34 #include <sys/dsl_bookmark.h>
35 #include <zfs_namecheck.h>
36 #include <sys/dmu_send.h>
37 
38 static int
dsl_bookmark_hold_ds(dsl_pool_t * dp,const char * fullname,dsl_dataset_t ** dsp,void * tag,char ** shortnamep)39 dsl_bookmark_hold_ds(dsl_pool_t *dp, const char *fullname,
40     dsl_dataset_t **dsp, void *tag, char **shortnamep)
41 {
42 	char buf[ZFS_MAX_DATASET_NAME_LEN];
43 	char *hashp;
44 
45 	if (strlen(fullname) >= ZFS_MAX_DATASET_NAME_LEN)
46 		return (SET_ERROR(ENAMETOOLONG));
47 	hashp = strchr(fullname, '#');
48 	if (hashp == NULL)
49 		return (SET_ERROR(EINVAL));
50 
51 	*shortnamep = hashp + 1;
52 	if (zfs_component_namecheck(*shortnamep, NULL, NULL))
53 		return (SET_ERROR(EINVAL));
54 	(void) strlcpy(buf, fullname, hashp - fullname + 1);
55 	return (dsl_dataset_hold(dp, buf, tag, dsp));
56 }
57 
58 /*
59  * When reading BOOKMARK_V1 bookmarks, the BOOKMARK_V2 fields are guaranteed
60  * to be zeroed.
61  *
62  * Returns ESRCH if bookmark is not found.
63  * Note, we need to use the ZAP rather than the AVL to look up bookmarks
64  * by name, because only the ZAP honors the casesensitivity setting.
65  */
66 int
dsl_bookmark_lookup_impl(dsl_dataset_t * ds,const char * shortname,zfs_bookmark_phys_t * bmark_phys)67 dsl_bookmark_lookup_impl(dsl_dataset_t *ds, const char *shortname,
68     zfs_bookmark_phys_t *bmark_phys)
69 {
70 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
71 	uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
72 	matchtype_t mt = 0;
73 	int err;
74 
75 	if (bmark_zapobj == 0)
76 		return (SET_ERROR(ESRCH));
77 
78 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
79 		mt = MT_NORMALIZE;
80 
81 	/*
82 	 * Zero out the bookmark in case the one stored on disk
83 	 * is in an older, shorter format.
84 	 */
85 	bzero(bmark_phys, sizeof (*bmark_phys));
86 
87 	err = zap_lookup_norm(mos, bmark_zapobj, shortname, sizeof (uint64_t),
88 	    sizeof (*bmark_phys) / sizeof (uint64_t), bmark_phys, mt, NULL, 0,
89 	    NULL);
90 
91 	return (err == ENOENT ? SET_ERROR(ESRCH) : err);
92 }
93 
94 /*
95  * If later_ds is non-NULL, this will return EXDEV if the specified bookmark
96  * does not represents an earlier point in later_ds's timeline.  However,
97  * bmp will still be filled in if we return EXDEV.
98  *
99  * Returns ENOENT if the dataset containing the bookmark does not exist.
100  * Returns ESRCH if the dataset exists but the bookmark was not found in it.
101  */
102 int
dsl_bookmark_lookup(dsl_pool_t * dp,const char * fullname,dsl_dataset_t * later_ds,zfs_bookmark_phys_t * bmp)103 dsl_bookmark_lookup(dsl_pool_t *dp, const char *fullname,
104     dsl_dataset_t *later_ds, zfs_bookmark_phys_t *bmp)
105 {
106 	char *shortname;
107 	dsl_dataset_t *ds;
108 	int error;
109 
110 	error = dsl_bookmark_hold_ds(dp, fullname, &ds, FTAG, &shortname);
111 	if (error != 0)
112 		return (error);
113 
114 	error = dsl_bookmark_lookup_impl(ds, shortname, bmp);
115 	if (error == 0 && later_ds != NULL) {
116 		if (!dsl_dataset_is_before(later_ds, ds, bmp->zbm_creation_txg))
117 			error = SET_ERROR(EXDEV);
118 	}
119 	dsl_dataset_rele(ds, FTAG);
120 	return (error);
121 }
122 
123 /*
124  * Validates that
125  * - bmark is a full dataset path of a bookmark (bookmark_namecheck)
126  * - source is a full path of a snapshot or bookmark
127  *   ({bookmark,snapshot}_namecheck)
128  *
129  * Returns 0 if valid, -1 otherwise.
130  */
131 static int
dsl_bookmark_create_nvl_validate_pair(const char * bmark,const char * source)132 dsl_bookmark_create_nvl_validate_pair(const char *bmark, const char *source)
133 {
134 	if (bookmark_namecheck(bmark, NULL, NULL) != 0)
135 		return (-1);
136 
137 	int is_bmark, is_snap;
138 	is_bmark = bookmark_namecheck(source, NULL, NULL) == 0;
139 	is_snap = snapshot_namecheck(source, NULL, NULL) == 0;
140 	if (!is_bmark && !is_snap)
141 		return (-1);
142 
143 	return (0);
144 }
145 
146 /*
147  * Check that the given nvlist corresponds to the following schema:
148  *  { newbookmark -> source, ... }
149  * where
150  * - each pair passes dsl_bookmark_create_nvl_validate_pair
151  * - all newbookmarks are in the same pool
152  * - all newbookmarks have unique names
153  *
154  * Note that this function is only validates above schema. Callers must ensure
155  * that the bookmarks can be created, e.g. that sources exist.
156  *
157  * Returns 0 if the nvlist adheres to above schema.
158  * Returns -1 if it doesn't.
159  */
160 int
dsl_bookmark_create_nvl_validate(nvlist_t * bmarks)161 dsl_bookmark_create_nvl_validate(nvlist_t *bmarks)
162 {
163 	char *first;
164 	size_t first_len;
165 
166 	first = NULL;
167 	for (nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
168 	    pair != NULL; pair = nvlist_next_nvpair(bmarks, pair)) {
169 
170 		char *bmark = nvpair_name(pair);
171 		char *source;
172 
173 		/* list structure: values must be snapshots XOR bookmarks */
174 		if (nvpair_value_string(pair, &source) != 0)
175 			return (-1);
176 		if (dsl_bookmark_create_nvl_validate_pair(bmark, source) != 0)
177 			return (-1);
178 
179 		/* same pool check */
180 		if (first == NULL) {
181 			char *cp = strpbrk(bmark, "/#");
182 			if (cp == NULL)
183 				return (-1);
184 			first = bmark;
185 			first_len = cp - bmark;
186 		}
187 		if (strncmp(first, bmark, first_len) != 0)
188 			return (-1);
189 		switch (*(bmark + first_len)) {
190 			case '/': /* fallthrough */
191 			case '#':
192 				break;
193 			default:
194 				return (-1);
195 		}
196 
197 		/* unique newbookmark names; todo: O(n^2) */
198 		for (nvpair_t *pair2 = nvlist_next_nvpair(bmarks, pair);
199 		    pair2 != NULL; pair2 = nvlist_next_nvpair(bmarks, pair2)) {
200 			if (strcmp(nvpair_name(pair), nvpair_name(pair2)) == 0)
201 				return (-1);
202 		}
203 
204 	}
205 	return (0);
206 }
207 
208 /*
209  * expects that newbm and source have been validated using
210  * dsl_bookmark_create_nvl_validate_pair
211  */
212 static int
dsl_bookmark_create_check_impl(dsl_pool_t * dp,const char * newbm,const char * source)213 dsl_bookmark_create_check_impl(dsl_pool_t *dp,
214     const char *newbm, const char *source)
215 {
216 	ASSERT0(dsl_bookmark_create_nvl_validate_pair(newbm, source));
217 	/* defer source namecheck until we know it's a snapshot or bookmark */
218 
219 	int error;
220 	dsl_dataset_t *newbm_ds;
221 	char *newbm_short;
222 	zfs_bookmark_phys_t bmark_phys;
223 
224 	error = dsl_bookmark_hold_ds(dp, newbm, &newbm_ds, FTAG, &newbm_short);
225 	if (error != 0)
226 		return (error);
227 
228 	/* Verify that the new bookmark does not already exist */
229 	error = dsl_bookmark_lookup_impl(newbm_ds, newbm_short, &bmark_phys);
230 	switch (error) {
231 	case ESRCH:
232 		/* happy path: new bmark doesn't exist, proceed after switch */
233 		error = 0;
234 		break;
235 	case 0:
236 		error = SET_ERROR(EEXIST);
237 		goto eholdnewbmds;
238 	default:
239 		/* dsl_bookmark_lookup_impl already did SET_ERROR */
240 		goto eholdnewbmds;
241 	}
242 
243 	/* error is retval of the following if-cascade */
244 	if (strchr(source, '@') != NULL) {
245 		dsl_dataset_t *source_snap_ds;
246 		ASSERT3S(snapshot_namecheck(source, NULL, NULL), ==, 0);
247 		error = dsl_dataset_hold(dp, source, FTAG, &source_snap_ds);
248 		if (error == 0) {
249 			VERIFY(source_snap_ds->ds_is_snapshot);
250 			/*
251 			 * Verify that source snapshot is an earlier point in
252 			 * newbm_ds's timeline (source may be newbm_ds's origin)
253 			 */
254 			if (!dsl_dataset_is_before(newbm_ds, source_snap_ds, 0))
255 				error = SET_ERROR(
256 				    ZFS_ERR_BOOKMARK_SOURCE_NOT_ANCESTOR);
257 			dsl_dataset_rele(source_snap_ds, FTAG);
258 		}
259 	} else if (strchr(source, '#') != NULL) {
260 		zfs_bookmark_phys_t source_phys;
261 		ASSERT3S(bookmark_namecheck(source, NULL, NULL), ==, 0);
262 		/*
263 		 * Source must exists and be an earlier point in newbm_ds's
264 		 * timeline (newbm_ds's origin may be a snap of source's ds)
265 		 */
266 		error = dsl_bookmark_lookup(dp, source, newbm_ds, &source_phys);
267 		switch (error) {
268 		case 0:
269 			break; /* happy path */
270 		case EXDEV:
271 			error = SET_ERROR(ZFS_ERR_BOOKMARK_SOURCE_NOT_ANCESTOR);
272 			break;
273 		default:
274 			/* dsl_bookmark_lookup already did SET_ERROR */
275 			break;
276 		}
277 	} else {
278 		/*
279 		 * dsl_bookmark_create_nvl_validate validates that source is
280 		 * either snapshot or bookmark
281 		 */
282 		panic("unreachable code: %s", source);
283 	}
284 
285 eholdnewbmds:
286 	dsl_dataset_rele(newbm_ds, FTAG);
287 	return (error);
288 }
289 
290 int
dsl_bookmark_create_check(void * arg,dmu_tx_t * tx)291 dsl_bookmark_create_check(void *arg, dmu_tx_t *tx)
292 {
293 	dsl_bookmark_create_arg_t *dbca = arg;
294 	int rv = 0;
295 	int schema_err = 0;
296 	ASSERT3P(dbca, !=, NULL);
297 	ASSERT3P(dbca->dbca_bmarks, !=, NULL);
298 	/* dbca->dbca_errors is allowed to be NULL */
299 
300 	dsl_pool_t *dp = dmu_tx_pool(tx);
301 
302 	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
303 		return (SET_ERROR(ENOTSUP));
304 
305 	if (dsl_bookmark_create_nvl_validate(dbca->dbca_bmarks) != 0)
306 		rv = schema_err = SET_ERROR(EINVAL);
307 
308 	for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
309 	    pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
310 		char *new = nvpair_name(pair);
311 
312 		int error = schema_err;
313 		if (error == 0) {
314 			char *source = fnvpair_value_string(pair);
315 			error = dsl_bookmark_create_check_impl(dp, new, source);
316 			if (error != 0)
317 				error = SET_ERROR(error);
318 		}
319 
320 		if (error != 0) {
321 			rv = error;
322 			if (dbca->dbca_errors != NULL)
323 				fnvlist_add_int32(dbca->dbca_errors,
324 				    new, error);
325 		}
326 	}
327 
328 	return (rv);
329 }
330 
331 static dsl_bookmark_node_t *
dsl_bookmark_node_alloc(char * shortname)332 dsl_bookmark_node_alloc(char *shortname)
333 {
334 	dsl_bookmark_node_t *dbn = kmem_alloc(sizeof (*dbn), KM_SLEEP);
335 	dbn->dbn_name = spa_strdup(shortname);
336 	dbn->dbn_dirty = B_FALSE;
337 	mutex_init(&dbn->dbn_lock, NULL, MUTEX_DEFAULT, NULL);
338 	return (dbn);
339 }
340 
341 /*
342  * Set the fields in the zfs_bookmark_phys_t based on the specified snapshot.
343  */
344 static void
dsl_bookmark_set_phys(zfs_bookmark_phys_t * zbm,dsl_dataset_t * snap)345 dsl_bookmark_set_phys(zfs_bookmark_phys_t *zbm, dsl_dataset_t *snap)
346 {
347 	spa_t *spa = dsl_dataset_get_spa(snap);
348 	objset_t *mos = spa_get_dsl(spa)->dp_meta_objset;
349 	dsl_dataset_phys_t *dsp = dsl_dataset_phys(snap);
350 
351 	memset(zbm, 0, sizeof (zfs_bookmark_phys_t));
352 	zbm->zbm_guid = dsp->ds_guid;
353 	zbm->zbm_creation_txg = dsp->ds_creation_txg;
354 	zbm->zbm_creation_time = dsp->ds_creation_time;
355 	zbm->zbm_redaction_obj = 0;
356 
357 	/*
358 	 * If the dataset is encrypted create a larger bookmark to
359 	 * accommodate the IVset guid. The IVset guid was added
360 	 * after the encryption feature to prevent a problem with
361 	 * raw sends. If we encounter an encrypted dataset without
362 	 * an IVset guid we fall back to a normal bookmark.
363 	 */
364 	if (snap->ds_dir->dd_crypto_obj != 0 &&
365 	    spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
366 		(void) zap_lookup(mos, snap->ds_object,
367 		    DS_FIELD_IVSET_GUID, sizeof (uint64_t), 1,
368 		    &zbm->zbm_ivset_guid);
369 	}
370 
371 	if (spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_WRITTEN)) {
372 		zbm->zbm_flags = ZBM_FLAG_SNAPSHOT_EXISTS | ZBM_FLAG_HAS_FBN;
373 		zbm->zbm_referenced_bytes_refd = dsp->ds_referenced_bytes;
374 		zbm->zbm_compressed_bytes_refd = dsp->ds_compressed_bytes;
375 		zbm->zbm_uncompressed_bytes_refd = dsp->ds_uncompressed_bytes;
376 
377 		dsl_dataset_t *nextds;
378 		VERIFY0(dsl_dataset_hold_obj(snap->ds_dir->dd_pool,
379 		    dsp->ds_next_snap_obj, FTAG, &nextds));
380 		dsl_deadlist_space(&nextds->ds_deadlist,
381 		    &zbm->zbm_referenced_freed_before_next_snap,
382 		    &zbm->zbm_compressed_freed_before_next_snap,
383 		    &zbm->zbm_uncompressed_freed_before_next_snap);
384 		dsl_dataset_rele(nextds, FTAG);
385 	}
386 }
387 
388 /*
389  * Add dsl_bookmark_node_t `dbn` to the given dataset and increment appropriate
390  * SPA feature counters.
391  */
392 void
dsl_bookmark_node_add(dsl_dataset_t * hds,dsl_bookmark_node_t * dbn,dmu_tx_t * tx)393 dsl_bookmark_node_add(dsl_dataset_t *hds, dsl_bookmark_node_t *dbn,
394     dmu_tx_t *tx)
395 {
396 	dsl_pool_t *dp = dmu_tx_pool(tx);
397 	objset_t *mos = dp->dp_meta_objset;
398 
399 	if (hds->ds_bookmarks_obj == 0) {
400 		hds->ds_bookmarks_obj = zap_create_norm(mos,
401 		    U8_TEXTPREP_TOUPPER, DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0,
402 		    tx);
403 		spa_feature_incr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
404 
405 		dsl_dataset_zapify(hds, tx);
406 		VERIFY0(zap_add(mos, hds->ds_object,
407 		    DS_FIELD_BOOKMARK_NAMES,
408 		    sizeof (hds->ds_bookmarks_obj), 1,
409 		    &hds->ds_bookmarks_obj, tx));
410 	}
411 
412 	avl_add(&hds->ds_bookmarks, dbn);
413 
414 	/*
415 	 * To maintain backwards compatibility with software that doesn't
416 	 * understand SPA_FEATURE_BOOKMARK_V2, we need to use the smallest
417 	 * possible bookmark size.
418 	 */
419 	uint64_t bookmark_phys_size = BOOKMARK_PHYS_SIZE_V1;
420 	if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARK_V2) &&
421 	    (dbn->dbn_phys.zbm_ivset_guid != 0 || dbn->dbn_phys.zbm_flags &
422 	    ZBM_FLAG_HAS_FBN || dbn->dbn_phys.zbm_redaction_obj != 0)) {
423 		bookmark_phys_size = BOOKMARK_PHYS_SIZE_V2;
424 		spa_feature_incr(dp->dp_spa, SPA_FEATURE_BOOKMARK_V2, tx);
425 	}
426 
427 	__attribute__((unused)) zfs_bookmark_phys_t zero_phys = { 0 };
428 	ASSERT0(bcmp(((char *)&dbn->dbn_phys) + bookmark_phys_size,
429 	    &zero_phys, sizeof (zfs_bookmark_phys_t) - bookmark_phys_size));
430 
431 	VERIFY0(zap_add(mos, hds->ds_bookmarks_obj, dbn->dbn_name,
432 	    sizeof (uint64_t), bookmark_phys_size / sizeof (uint64_t),
433 	    &dbn->dbn_phys, tx));
434 }
435 
436 /*
437  * If redaction_list is non-null, we create a redacted bookmark and redaction
438  * list, and store the object number of the redaction list in redact_obj.
439  */
440 static void
dsl_bookmark_create_sync_impl_snap(const char * bookmark,const char * snapshot,dmu_tx_t * tx,uint64_t num_redact_snaps,uint64_t * redact_snaps,void * tag,redaction_list_t ** redaction_list)441 dsl_bookmark_create_sync_impl_snap(const char *bookmark, const char *snapshot,
442     dmu_tx_t *tx, uint64_t num_redact_snaps, uint64_t *redact_snaps, void *tag,
443     redaction_list_t **redaction_list)
444 {
445 	dsl_pool_t *dp = dmu_tx_pool(tx);
446 	objset_t *mos = dp->dp_meta_objset;
447 	dsl_dataset_t *snapds, *bmark_fs;
448 	char *shortname;
449 	boolean_t bookmark_redacted;
450 	uint64_t *dsredactsnaps;
451 	uint64_t dsnumsnaps;
452 
453 	VERIFY0(dsl_dataset_hold(dp, snapshot, FTAG, &snapds));
454 	VERIFY0(dsl_bookmark_hold_ds(dp, bookmark, &bmark_fs, FTAG,
455 	    &shortname));
456 
457 	dsl_bookmark_node_t *dbn = dsl_bookmark_node_alloc(shortname);
458 	dsl_bookmark_set_phys(&dbn->dbn_phys, snapds);
459 
460 	bookmark_redacted = dsl_dataset_get_uint64_array_feature(snapds,
461 	    SPA_FEATURE_REDACTED_DATASETS, &dsnumsnaps, &dsredactsnaps);
462 	if (redaction_list != NULL || bookmark_redacted) {
463 		redaction_list_t *local_rl;
464 		if (bookmark_redacted) {
465 			redact_snaps = dsredactsnaps;
466 			num_redact_snaps = dsnumsnaps;
467 		}
468 		dbn->dbn_phys.zbm_redaction_obj = dmu_object_alloc(mos,
469 		    DMU_OTN_UINT64_METADATA, SPA_OLD_MAXBLOCKSIZE,
470 		    DMU_OTN_UINT64_METADATA, sizeof (redaction_list_phys_t) +
471 		    num_redact_snaps * sizeof (uint64_t), tx);
472 		spa_feature_incr(dp->dp_spa,
473 		    SPA_FEATURE_REDACTION_BOOKMARKS, tx);
474 
475 		VERIFY0(dsl_redaction_list_hold_obj(dp,
476 		    dbn->dbn_phys.zbm_redaction_obj, tag, &local_rl));
477 		dsl_redaction_list_long_hold(dp, local_rl, tag);
478 
479 		ASSERT3U((local_rl)->rl_dbuf->db_size, >=,
480 		    sizeof (redaction_list_phys_t) + num_redact_snaps *
481 		    sizeof (uint64_t));
482 		dmu_buf_will_dirty(local_rl->rl_dbuf, tx);
483 		bcopy(redact_snaps, local_rl->rl_phys->rlp_snaps,
484 		    sizeof (uint64_t) * num_redact_snaps);
485 		local_rl->rl_phys->rlp_num_snaps = num_redact_snaps;
486 		if (bookmark_redacted) {
487 			ASSERT3P(redaction_list, ==, NULL);
488 			local_rl->rl_phys->rlp_last_blkid = UINT64_MAX;
489 			local_rl->rl_phys->rlp_last_object = UINT64_MAX;
490 			dsl_redaction_list_long_rele(local_rl, tag);
491 			dsl_redaction_list_rele(local_rl, tag);
492 		} else {
493 			*redaction_list = local_rl;
494 		}
495 	}
496 
497 	if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
498 		spa_feature_incr(dp->dp_spa,
499 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
500 	}
501 
502 	dsl_bookmark_node_add(bmark_fs, dbn, tx);
503 
504 	spa_history_log_internal_ds(bmark_fs, "bookmark", tx,
505 	    "name=%s creation_txg=%llu target_snap=%llu redact_obj=%llu",
506 	    shortname, (longlong_t)dbn->dbn_phys.zbm_creation_txg,
507 	    (longlong_t)snapds->ds_object,
508 	    (longlong_t)dbn->dbn_phys.zbm_redaction_obj);
509 
510 	dsl_dataset_rele(bmark_fs, FTAG);
511 	dsl_dataset_rele(snapds, FTAG);
512 }
513 
514 
515 static void
dsl_bookmark_create_sync_impl_book(const char * new_name,const char * source_name,dmu_tx_t * tx)516 dsl_bookmark_create_sync_impl_book(
517     const char *new_name, const char *source_name, dmu_tx_t *tx)
518 {
519 	dsl_pool_t *dp = dmu_tx_pool(tx);
520 	dsl_dataset_t *bmark_fs_source, *bmark_fs_new;
521 	char *source_shortname, *new_shortname;
522 	zfs_bookmark_phys_t source_phys;
523 
524 	VERIFY0(dsl_bookmark_hold_ds(dp, source_name, &bmark_fs_source, FTAG,
525 	    &source_shortname));
526 	VERIFY0(dsl_bookmark_hold_ds(dp, new_name, &bmark_fs_new, FTAG,
527 	    &new_shortname));
528 
529 	/*
530 	 * create a copy of the source bookmark by copying most of its members
531 	 *
532 	 * Caveat: bookmarking a redaction bookmark yields a normal bookmark
533 	 * -----------------------------------------------------------------
534 	 * Reasoning:
535 	 * - The zbm_redaction_obj would be referred to by both source and new
536 	 *   bookmark, but would be destroyed once either source or new is
537 	 *   destroyed, resulting in use-after-free of the referred object.
538 	 * - User expectation when issuing the `zfs bookmark` command is that
539 	 *   a normal bookmark of the source is created
540 	 *
541 	 * Design Alternatives For Full Redaction Bookmark Copying:
542 	 * - reference-count the redaction object => would require on-disk
543 	 *   format change for existing redaction objects
544 	 * - Copy the redaction object => cannot be done in syncing context
545 	 *   because the redaction object might be too large
546 	 */
547 
548 	VERIFY0(dsl_bookmark_lookup_impl(bmark_fs_source, source_shortname,
549 	    &source_phys));
550 	dsl_bookmark_node_t *new_dbn = dsl_bookmark_node_alloc(new_shortname);
551 
552 	memcpy(&new_dbn->dbn_phys, &source_phys, sizeof (source_phys));
553 	new_dbn->dbn_phys.zbm_redaction_obj = 0;
554 
555 	/* update feature counters */
556 	if (new_dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
557 		spa_feature_incr(dp->dp_spa,
558 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
559 	}
560 	/* no need for redaction bookmark counter; nulled zbm_redaction_obj */
561 	/* dsl_bookmark_node_add bumps bookmarks and v2-bookmarks counter */
562 
563 	/*
564 	 * write new bookmark
565 	 *
566 	 * Note that dsl_bookmark_lookup_impl guarantees that, if source is a
567 	 * v1 bookmark, the v2-only fields are zeroed.
568 	 * And dsl_bookmark_node_add writes back a v1-sized bookmark if
569 	 * v2 bookmarks are disabled and/or v2-only fields are zeroed.
570 	 * => bookmark copying works on pre-bookmark-v2 pools
571 	 */
572 	dsl_bookmark_node_add(bmark_fs_new, new_dbn, tx);
573 
574 	spa_history_log_internal_ds(bmark_fs_source, "bookmark", tx,
575 	    "name=%s creation_txg=%llu source_guid=%llu",
576 	    new_shortname, (longlong_t)new_dbn->dbn_phys.zbm_creation_txg,
577 	    (longlong_t)source_phys.zbm_guid);
578 
579 	dsl_dataset_rele(bmark_fs_source, FTAG);
580 	dsl_dataset_rele(bmark_fs_new, FTAG);
581 }
582 
583 void
dsl_bookmark_create_sync(void * arg,dmu_tx_t * tx)584 dsl_bookmark_create_sync(void *arg, dmu_tx_t *tx)
585 {
586 	dsl_bookmark_create_arg_t *dbca = arg;
587 
588 	ASSERT(spa_feature_is_enabled(dmu_tx_pool(tx)->dp_spa,
589 	    SPA_FEATURE_BOOKMARKS));
590 
591 	for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
592 	    pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
593 
594 		char *new = nvpair_name(pair);
595 		char *source = fnvpair_value_string(pair);
596 
597 		if (strchr(source, '@') != NULL) {
598 			dsl_bookmark_create_sync_impl_snap(new, source, tx,
599 			    0, NULL, NULL, NULL);
600 		} else if (strchr(source, '#') != NULL) {
601 			dsl_bookmark_create_sync_impl_book(new, source, tx);
602 		} else {
603 			panic("unreachable code");
604 		}
605 
606 	}
607 }
608 
609 /*
610  * The bookmarks must all be in the same pool.
611  */
612 int
dsl_bookmark_create(nvlist_t * bmarks,nvlist_t * errors)613 dsl_bookmark_create(nvlist_t *bmarks, nvlist_t *errors)
614 {
615 	nvpair_t *pair;
616 	dsl_bookmark_create_arg_t dbca;
617 
618 	pair = nvlist_next_nvpair(bmarks, NULL);
619 	if (pair == NULL)
620 		return (0);
621 
622 	dbca.dbca_bmarks = bmarks;
623 	dbca.dbca_errors = errors;
624 
625 	return (dsl_sync_task(nvpair_name(pair), dsl_bookmark_create_check,
626 	    dsl_bookmark_create_sync, &dbca,
627 	    fnvlist_num_pairs(bmarks), ZFS_SPACE_CHECK_NORMAL));
628 }
629 
630 static int
dsl_bookmark_create_redacted_check(void * arg,dmu_tx_t * tx)631 dsl_bookmark_create_redacted_check(void *arg, dmu_tx_t *tx)
632 {
633 	dsl_bookmark_create_redacted_arg_t *dbcra = arg;
634 	dsl_pool_t *dp = dmu_tx_pool(tx);
635 	int rv = 0;
636 
637 	if (!spa_feature_is_enabled(dp->dp_spa,
638 	    SPA_FEATURE_REDACTION_BOOKMARKS))
639 		return (SET_ERROR(ENOTSUP));
640 	/*
641 	 * If the list of redact snaps will not fit in the bonus buffer with
642 	 * the furthest reached object and offset, fail.
643 	 */
644 	if (dbcra->dbcra_numsnaps > (dmu_bonus_max() -
645 	    sizeof (redaction_list_phys_t)) / sizeof (uint64_t))
646 		return (SET_ERROR(E2BIG));
647 
648 	if (dsl_bookmark_create_nvl_validate_pair(
649 	    dbcra->dbcra_bmark, dbcra->dbcra_snap) != 0)
650 		return (SET_ERROR(EINVAL));
651 
652 	rv = dsl_bookmark_create_check_impl(dp,
653 	    dbcra->dbcra_bmark, dbcra->dbcra_snap);
654 	return (rv);
655 }
656 
657 static void
dsl_bookmark_create_redacted_sync(void * arg,dmu_tx_t * tx)658 dsl_bookmark_create_redacted_sync(void *arg, dmu_tx_t *tx)
659 {
660 	dsl_bookmark_create_redacted_arg_t *dbcra = arg;
661 	dsl_bookmark_create_sync_impl_snap(dbcra->dbcra_bmark,
662 	    dbcra->dbcra_snap, tx, dbcra->dbcra_numsnaps, dbcra->dbcra_snaps,
663 	    dbcra->dbcra_tag, dbcra->dbcra_rl);
664 }
665 
666 int
dsl_bookmark_create_redacted(const char * bookmark,const char * snapshot,uint64_t numsnaps,uint64_t * snapguids,void * tag,redaction_list_t ** rl)667 dsl_bookmark_create_redacted(const char *bookmark, const char *snapshot,
668     uint64_t numsnaps, uint64_t *snapguids, void *tag, redaction_list_t **rl)
669 {
670 	dsl_bookmark_create_redacted_arg_t dbcra;
671 
672 	dbcra.dbcra_bmark = bookmark;
673 	dbcra.dbcra_snap = snapshot;
674 	dbcra.dbcra_rl = rl;
675 	dbcra.dbcra_numsnaps = numsnaps;
676 	dbcra.dbcra_snaps = snapguids;
677 	dbcra.dbcra_tag = tag;
678 
679 	return (dsl_sync_task(bookmark, dsl_bookmark_create_redacted_check,
680 	    dsl_bookmark_create_redacted_sync, &dbcra, 5,
681 	    ZFS_SPACE_CHECK_NORMAL));
682 }
683 
684 /*
685  * Retrieve the list of properties given in the 'props' nvlist for a bookmark.
686  * If 'props' is NULL, retrieves all properties.
687  */
688 static void
dsl_bookmark_fetch_props(dsl_pool_t * dp,zfs_bookmark_phys_t * bmark_phys,nvlist_t * props,nvlist_t * out_props)689 dsl_bookmark_fetch_props(dsl_pool_t *dp, zfs_bookmark_phys_t *bmark_phys,
690     nvlist_t *props, nvlist_t *out_props)
691 {
692 	ASSERT3P(dp, !=, NULL);
693 	ASSERT3P(bmark_phys, !=, NULL);
694 	ASSERT3P(out_props, !=, NULL);
695 	ASSERT(RRW_LOCK_HELD(&dp->dp_config_rwlock));
696 
697 	if (props == NULL || nvlist_exists(props,
698 	    zfs_prop_to_name(ZFS_PROP_GUID))) {
699 		dsl_prop_nvlist_add_uint64(out_props,
700 		    ZFS_PROP_GUID, bmark_phys->zbm_guid);
701 	}
702 	if (props == NULL || nvlist_exists(props,
703 	    zfs_prop_to_name(ZFS_PROP_CREATETXG))) {
704 		dsl_prop_nvlist_add_uint64(out_props,
705 		    ZFS_PROP_CREATETXG, bmark_phys->zbm_creation_txg);
706 	}
707 	if (props == NULL || nvlist_exists(props,
708 	    zfs_prop_to_name(ZFS_PROP_CREATION))) {
709 		dsl_prop_nvlist_add_uint64(out_props,
710 		    ZFS_PROP_CREATION, bmark_phys->zbm_creation_time);
711 	}
712 	if (props == NULL || nvlist_exists(props,
713 	    zfs_prop_to_name(ZFS_PROP_IVSET_GUID))) {
714 		dsl_prop_nvlist_add_uint64(out_props,
715 		    ZFS_PROP_IVSET_GUID, bmark_phys->zbm_ivset_guid);
716 	}
717 	if (bmark_phys->zbm_flags & ZBM_FLAG_HAS_FBN) {
718 		if (props == NULL || nvlist_exists(props,
719 		    zfs_prop_to_name(ZFS_PROP_REFERENCED))) {
720 			dsl_prop_nvlist_add_uint64(out_props,
721 			    ZFS_PROP_REFERENCED,
722 			    bmark_phys->zbm_referenced_bytes_refd);
723 		}
724 		if (props == NULL || nvlist_exists(props,
725 		    zfs_prop_to_name(ZFS_PROP_LOGICALREFERENCED))) {
726 			dsl_prop_nvlist_add_uint64(out_props,
727 			    ZFS_PROP_LOGICALREFERENCED,
728 			    bmark_phys->zbm_uncompressed_bytes_refd);
729 		}
730 		if (props == NULL || nvlist_exists(props,
731 		    zfs_prop_to_name(ZFS_PROP_REFRATIO))) {
732 			uint64_t ratio =
733 			    bmark_phys->zbm_compressed_bytes_refd == 0 ? 100 :
734 			    bmark_phys->zbm_uncompressed_bytes_refd * 100 /
735 			    bmark_phys->zbm_compressed_bytes_refd;
736 			dsl_prop_nvlist_add_uint64(out_props,
737 			    ZFS_PROP_REFRATIO, ratio);
738 		}
739 	}
740 
741 	if ((props == NULL || nvlist_exists(props, "redact_snaps") ||
742 	    nvlist_exists(props, "redact_complete")) &&
743 	    bmark_phys->zbm_redaction_obj != 0) {
744 		redaction_list_t *rl;
745 		int err = dsl_redaction_list_hold_obj(dp,
746 		    bmark_phys->zbm_redaction_obj, FTAG, &rl);
747 		if (err == 0) {
748 			if (nvlist_exists(props, "redact_snaps")) {
749 				nvlist_t *nvl;
750 				nvl = fnvlist_alloc();
751 				fnvlist_add_uint64_array(nvl, ZPROP_VALUE,
752 				    rl->rl_phys->rlp_snaps,
753 				    rl->rl_phys->rlp_num_snaps);
754 				fnvlist_add_nvlist(out_props, "redact_snaps",
755 				    nvl);
756 				nvlist_free(nvl);
757 			}
758 			if (nvlist_exists(props, "redact_complete")) {
759 				nvlist_t *nvl;
760 				nvl = fnvlist_alloc();
761 				fnvlist_add_boolean_value(nvl, ZPROP_VALUE,
762 				    rl->rl_phys->rlp_last_blkid == UINT64_MAX &&
763 				    rl->rl_phys->rlp_last_object == UINT64_MAX);
764 				fnvlist_add_nvlist(out_props, "redact_complete",
765 				    nvl);
766 				nvlist_free(nvl);
767 			}
768 			dsl_redaction_list_rele(rl, FTAG);
769 		}
770 	}
771 }
772 
773 int
dsl_get_bookmarks_impl(dsl_dataset_t * ds,nvlist_t * props,nvlist_t * outnvl)774 dsl_get_bookmarks_impl(dsl_dataset_t *ds, nvlist_t *props, nvlist_t *outnvl)
775 {
776 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
777 
778 	ASSERT(dsl_pool_config_held(dp));
779 
780 	if (dsl_dataset_is_snapshot(ds))
781 		return (SET_ERROR(EINVAL));
782 
783 	for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
784 	    dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
785 		nvlist_t *out_props = fnvlist_alloc();
786 
787 		dsl_bookmark_fetch_props(dp, &dbn->dbn_phys, props, out_props);
788 
789 		fnvlist_add_nvlist(outnvl, dbn->dbn_name, out_props);
790 		fnvlist_free(out_props);
791 	}
792 	return (0);
793 }
794 
795 /*
796  * Comparison func for ds_bookmarks AVL tree.  We sort the bookmarks by
797  * their TXG, then by their FBN-ness.  The "FBN-ness" component ensures
798  * that all bookmarks at the same TXG that HAS_FBN are adjacent, which
799  * dsl_bookmark_destroy_sync_impl() depends on.  Note that there may be
800  * multiple bookmarks at the same TXG (with the same FBN-ness).  In this
801  * case we differentiate them by an arbitrary metric (in this case,
802  * their names).
803  */
804 static int
dsl_bookmark_compare(const void * l,const void * r)805 dsl_bookmark_compare(const void *l, const void *r)
806 {
807 	const dsl_bookmark_node_t *ldbn = l;
808 	const dsl_bookmark_node_t *rdbn = r;
809 
810 	int64_t cmp = TREE_CMP(ldbn->dbn_phys.zbm_creation_txg,
811 	    rdbn->dbn_phys.zbm_creation_txg);
812 	if (likely(cmp))
813 		return (cmp);
814 	cmp = TREE_CMP((ldbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN),
815 	    (rdbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
816 	if (likely(cmp))
817 		return (cmp);
818 	cmp = strcmp(ldbn->dbn_name, rdbn->dbn_name);
819 	return (TREE_ISIGN(cmp));
820 }
821 
822 /*
823  * Cache this (head) dataset's bookmarks in the ds_bookmarks AVL tree.
824  */
825 int
dsl_bookmark_init_ds(dsl_dataset_t * ds)826 dsl_bookmark_init_ds(dsl_dataset_t *ds)
827 {
828 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
829 	objset_t *mos = dp->dp_meta_objset;
830 
831 	ASSERT(!ds->ds_is_snapshot);
832 
833 	avl_create(&ds->ds_bookmarks, dsl_bookmark_compare,
834 	    sizeof (dsl_bookmark_node_t),
835 	    offsetof(dsl_bookmark_node_t, dbn_node));
836 
837 	if (!dsl_dataset_is_zapified(ds))
838 		return (0);
839 
840 	int zaperr = zap_lookup(mos, ds->ds_object, DS_FIELD_BOOKMARK_NAMES,
841 	    sizeof (ds->ds_bookmarks_obj), 1, &ds->ds_bookmarks_obj);
842 	if (zaperr == ENOENT)
843 		return (0);
844 	if (zaperr != 0)
845 		return (zaperr);
846 
847 	if (ds->ds_bookmarks_obj == 0)
848 		return (0);
849 
850 	int err = 0;
851 	zap_cursor_t zc;
852 	zap_attribute_t attr;
853 
854 	for (zap_cursor_init(&zc, mos, ds->ds_bookmarks_obj);
855 	    (err = zap_cursor_retrieve(&zc, &attr)) == 0;
856 	    zap_cursor_advance(&zc)) {
857 		dsl_bookmark_node_t *dbn =
858 		    dsl_bookmark_node_alloc(attr.za_name);
859 
860 		err = dsl_bookmark_lookup_impl(ds,
861 		    dbn->dbn_name, &dbn->dbn_phys);
862 		ASSERT3U(err, !=, ENOENT);
863 		if (err != 0) {
864 			kmem_free(dbn, sizeof (*dbn));
865 			break;
866 		}
867 		avl_add(&ds->ds_bookmarks, dbn);
868 	}
869 	zap_cursor_fini(&zc);
870 	if (err == ENOENT)
871 		err = 0;
872 	return (err);
873 }
874 
875 void
dsl_bookmark_fini_ds(dsl_dataset_t * ds)876 dsl_bookmark_fini_ds(dsl_dataset_t *ds)
877 {
878 	void *cookie = NULL;
879 	dsl_bookmark_node_t *dbn;
880 
881 	if (ds->ds_is_snapshot)
882 		return;
883 
884 	while ((dbn = avl_destroy_nodes(&ds->ds_bookmarks, &cookie)) != NULL) {
885 		spa_strfree(dbn->dbn_name);
886 		mutex_destroy(&dbn->dbn_lock);
887 		kmem_free(dbn, sizeof (*dbn));
888 	}
889 	avl_destroy(&ds->ds_bookmarks);
890 }
891 
892 /*
893  * Retrieve the bookmarks that exist in the specified dataset, and the
894  * requested properties of each bookmark.
895  *
896  * The "props" nvlist specifies which properties are requested.
897  * See lzc_get_bookmarks() for the list of valid properties.
898  */
899 int
dsl_get_bookmarks(const char * dsname,nvlist_t * props,nvlist_t * outnvl)900 dsl_get_bookmarks(const char *dsname, nvlist_t *props, nvlist_t *outnvl)
901 {
902 	dsl_pool_t *dp;
903 	dsl_dataset_t *ds;
904 	int err;
905 
906 	err = dsl_pool_hold(dsname, FTAG, &dp);
907 	if (err != 0)
908 		return (err);
909 	err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
910 	if (err != 0) {
911 		dsl_pool_rele(dp, FTAG);
912 		return (err);
913 	}
914 
915 	err = dsl_get_bookmarks_impl(ds, props, outnvl);
916 
917 	dsl_dataset_rele(ds, FTAG);
918 	dsl_pool_rele(dp, FTAG);
919 	return (err);
920 }
921 
922 /*
923  * Retrieve all properties for a single bookmark in the given dataset.
924  */
925 int
dsl_get_bookmark_props(const char * dsname,const char * bmname,nvlist_t * props)926 dsl_get_bookmark_props(const char *dsname, const char *bmname, nvlist_t *props)
927 {
928 	dsl_pool_t *dp;
929 	dsl_dataset_t *ds;
930 	zfs_bookmark_phys_t bmark_phys = { 0 };
931 	int err;
932 
933 	err = dsl_pool_hold(dsname, FTAG, &dp);
934 	if (err != 0)
935 		return (err);
936 	err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
937 	if (err != 0) {
938 		dsl_pool_rele(dp, FTAG);
939 		return (err);
940 	}
941 
942 	err = dsl_bookmark_lookup_impl(ds, bmname, &bmark_phys);
943 	if (err != 0)
944 		goto out;
945 
946 	dsl_bookmark_fetch_props(dp, &bmark_phys, NULL, props);
947 out:
948 	dsl_dataset_rele(ds, FTAG);
949 	dsl_pool_rele(dp, FTAG);
950 	return (err);
951 }
952 
953 typedef struct dsl_bookmark_destroy_arg {
954 	nvlist_t *dbda_bmarks;
955 	nvlist_t *dbda_success;
956 	nvlist_t *dbda_errors;
957 } dsl_bookmark_destroy_arg_t;
958 
959 static void
dsl_bookmark_destroy_sync_impl(dsl_dataset_t * ds,const char * name,dmu_tx_t * tx)960 dsl_bookmark_destroy_sync_impl(dsl_dataset_t *ds, const char *name,
961     dmu_tx_t *tx)
962 {
963 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
964 	uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
965 	matchtype_t mt = 0;
966 	uint64_t int_size, num_ints;
967 	/*
968 	 * 'search' must be zeroed so that dbn_flags (which is used in
969 	 * dsl_bookmark_compare()) will be zeroed even if the on-disk
970 	 * (in ZAP) bookmark is shorter than offsetof(dbn_flags).
971 	 */
972 	dsl_bookmark_node_t search = { 0 };
973 	char realname[ZFS_MAX_DATASET_NAME_LEN];
974 
975 	/*
976 	 * Find the real name of this bookmark, which may be different
977 	 * from the given name if the dataset is case-insensitive.  Then
978 	 * use the real name to find the node in the ds_bookmarks AVL tree.
979 	 */
980 
981 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
982 		mt = MT_NORMALIZE;
983 
984 	VERIFY0(zap_length(mos, bmark_zapobj, name, &int_size, &num_ints));
985 
986 	ASSERT3U(int_size, ==, sizeof (uint64_t));
987 
988 	if (num_ints * int_size > BOOKMARK_PHYS_SIZE_V1) {
989 		spa_feature_decr(dmu_objset_spa(mos),
990 		    SPA_FEATURE_BOOKMARK_V2, tx);
991 	}
992 	VERIFY0(zap_lookup_norm(mos, bmark_zapobj, name, sizeof (uint64_t),
993 	    num_ints, &search.dbn_phys, mt, realname, sizeof (realname), NULL));
994 
995 	search.dbn_name = realname;
996 	dsl_bookmark_node_t *dbn = avl_find(&ds->ds_bookmarks, &search, NULL);
997 	ASSERT(dbn != NULL);
998 
999 	if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
1000 		/*
1001 		 * If this bookmark HAS_FBN, and it is before the most
1002 		 * recent snapshot, then its TXG is a key in the head's
1003 		 * deadlist (and all clones' heads' deadlists).  If this is
1004 		 * the last thing keeping the key (i.e. there are no more
1005 		 * bookmarks with HAS_FBN at this TXG, and there is no
1006 		 * snapshot at this TXG), then remove the key.
1007 		 *
1008 		 * Note that this algorithm depends on ds_bookmarks being
1009 		 * sorted such that all bookmarks at the same TXG with
1010 		 * HAS_FBN are adjacent (with no non-HAS_FBN bookmarks
1011 		 * at the same TXG in between them).  If this were not
1012 		 * the case, we would need to examine *all* bookmarks
1013 		 * at this TXG, rather than just the adjacent ones.
1014 		 */
1015 
1016 		dsl_bookmark_node_t *dbn_prev =
1017 		    AVL_PREV(&ds->ds_bookmarks, dbn);
1018 		dsl_bookmark_node_t *dbn_next =
1019 		    AVL_NEXT(&ds->ds_bookmarks, dbn);
1020 
1021 		boolean_t more_bookmarks_at_this_txg =
1022 		    (dbn_prev != NULL && dbn_prev->dbn_phys.zbm_creation_txg ==
1023 		    dbn->dbn_phys.zbm_creation_txg &&
1024 		    (dbn_prev->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) ||
1025 		    (dbn_next != NULL && dbn_next->dbn_phys.zbm_creation_txg ==
1026 		    dbn->dbn_phys.zbm_creation_txg &&
1027 		    (dbn_next->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
1028 
1029 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS) &&
1030 		    !more_bookmarks_at_this_txg &&
1031 		    dbn->dbn_phys.zbm_creation_txg <
1032 		    dsl_dataset_phys(ds)->ds_prev_snap_txg) {
1033 			dsl_dir_remove_clones_key(ds->ds_dir,
1034 			    dbn->dbn_phys.zbm_creation_txg, tx);
1035 			dsl_deadlist_remove_key(&ds->ds_deadlist,
1036 			    dbn->dbn_phys.zbm_creation_txg, tx);
1037 		}
1038 
1039 		spa_feature_decr(dmu_objset_spa(mos),
1040 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
1041 	}
1042 
1043 	if (dbn->dbn_phys.zbm_redaction_obj != 0) {
1044 		VERIFY0(dmu_object_free(mos,
1045 		    dbn->dbn_phys.zbm_redaction_obj, tx));
1046 		spa_feature_decr(dmu_objset_spa(mos),
1047 		    SPA_FEATURE_REDACTION_BOOKMARKS, tx);
1048 	}
1049 
1050 	avl_remove(&ds->ds_bookmarks, dbn);
1051 	spa_strfree(dbn->dbn_name);
1052 	mutex_destroy(&dbn->dbn_lock);
1053 	kmem_free(dbn, sizeof (*dbn));
1054 
1055 	VERIFY0(zap_remove_norm(mos, bmark_zapobj, name, mt, tx));
1056 }
1057 
1058 static int
dsl_bookmark_destroy_check(void * arg,dmu_tx_t * tx)1059 dsl_bookmark_destroy_check(void *arg, dmu_tx_t *tx)
1060 {
1061 	dsl_bookmark_destroy_arg_t *dbda = arg;
1062 	dsl_pool_t *dp = dmu_tx_pool(tx);
1063 	int rv = 0;
1064 
1065 	ASSERT(nvlist_empty(dbda->dbda_success));
1066 	ASSERT(nvlist_empty(dbda->dbda_errors));
1067 
1068 	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
1069 		return (0);
1070 
1071 	for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_bmarks, NULL);
1072 	    pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_bmarks, pair)) {
1073 		const char *fullname = nvpair_name(pair);
1074 		dsl_dataset_t *ds;
1075 		zfs_bookmark_phys_t bm;
1076 		int error;
1077 		char *shortname;
1078 
1079 		error = dsl_bookmark_hold_ds(dp, fullname, &ds,
1080 		    FTAG, &shortname);
1081 		if (error == ENOENT) {
1082 			/* ignore it; the bookmark is "already destroyed" */
1083 			continue;
1084 		}
1085 		if (error == 0) {
1086 			error = dsl_bookmark_lookup_impl(ds, shortname, &bm);
1087 			dsl_dataset_rele(ds, FTAG);
1088 			if (error == ESRCH) {
1089 				/*
1090 				 * ignore it; the bookmark is
1091 				 * "already destroyed"
1092 				 */
1093 				continue;
1094 			}
1095 			if (error == 0 && bm.zbm_redaction_obj != 0) {
1096 				redaction_list_t *rl = NULL;
1097 				error = dsl_redaction_list_hold_obj(tx->tx_pool,
1098 				    bm.zbm_redaction_obj, FTAG, &rl);
1099 				if (error == ENOENT) {
1100 					error = 0;
1101 				} else if (error == 0 &&
1102 				    dsl_redaction_list_long_held(rl)) {
1103 					error = SET_ERROR(EBUSY);
1104 				}
1105 				if (rl != NULL) {
1106 					dsl_redaction_list_rele(rl, FTAG);
1107 				}
1108 			}
1109 		}
1110 		if (error == 0) {
1111 			if (dmu_tx_is_syncing(tx)) {
1112 				fnvlist_add_boolean(dbda->dbda_success,
1113 				    fullname);
1114 			}
1115 		} else {
1116 			fnvlist_add_int32(dbda->dbda_errors, fullname, error);
1117 			rv = error;
1118 		}
1119 	}
1120 	return (rv);
1121 }
1122 
1123 static void
dsl_bookmark_destroy_sync(void * arg,dmu_tx_t * tx)1124 dsl_bookmark_destroy_sync(void *arg, dmu_tx_t *tx)
1125 {
1126 	dsl_bookmark_destroy_arg_t *dbda = arg;
1127 	dsl_pool_t *dp = dmu_tx_pool(tx);
1128 	objset_t *mos = dp->dp_meta_objset;
1129 
1130 	for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_success, NULL);
1131 	    pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_success, pair)) {
1132 		dsl_dataset_t *ds;
1133 		char *shortname;
1134 		uint64_t zap_cnt;
1135 
1136 		VERIFY0(dsl_bookmark_hold_ds(dp, nvpair_name(pair),
1137 		    &ds, FTAG, &shortname));
1138 		dsl_bookmark_destroy_sync_impl(ds, shortname, tx);
1139 
1140 		/*
1141 		 * If all of this dataset's bookmarks have been destroyed,
1142 		 * free the zap object and decrement the feature's use count.
1143 		 */
1144 		VERIFY0(zap_count(mos, ds->ds_bookmarks_obj, &zap_cnt));
1145 		if (zap_cnt == 0) {
1146 			dmu_buf_will_dirty(ds->ds_dbuf, tx);
1147 			VERIFY0(zap_destroy(mos, ds->ds_bookmarks_obj, tx));
1148 			ds->ds_bookmarks_obj = 0;
1149 			spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
1150 			VERIFY0(zap_remove(mos, ds->ds_object,
1151 			    DS_FIELD_BOOKMARK_NAMES, tx));
1152 		}
1153 
1154 		spa_history_log_internal_ds(ds, "remove bookmark", tx,
1155 		    "name=%s", shortname);
1156 
1157 		dsl_dataset_rele(ds, FTAG);
1158 	}
1159 }
1160 
1161 /*
1162  * The bookmarks must all be in the same pool.
1163  */
1164 int
dsl_bookmark_destroy(nvlist_t * bmarks,nvlist_t * errors)1165 dsl_bookmark_destroy(nvlist_t *bmarks, nvlist_t *errors)
1166 {
1167 	int rv;
1168 	dsl_bookmark_destroy_arg_t dbda;
1169 	nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
1170 	if (pair == NULL)
1171 		return (0);
1172 
1173 	dbda.dbda_bmarks = bmarks;
1174 	dbda.dbda_errors = errors;
1175 	dbda.dbda_success = fnvlist_alloc();
1176 
1177 	rv = dsl_sync_task(nvpair_name(pair), dsl_bookmark_destroy_check,
1178 	    dsl_bookmark_destroy_sync, &dbda, fnvlist_num_pairs(bmarks),
1179 	    ZFS_SPACE_CHECK_RESERVED);
1180 	fnvlist_free(dbda.dbda_success);
1181 	return (rv);
1182 }
1183 
1184 /* Return B_TRUE if there are any long holds on this dataset. */
1185 boolean_t
dsl_redaction_list_long_held(redaction_list_t * rl)1186 dsl_redaction_list_long_held(redaction_list_t *rl)
1187 {
1188 	return (!zfs_refcount_is_zero(&rl->rl_longholds));
1189 }
1190 
1191 void
dsl_redaction_list_long_hold(dsl_pool_t * dp,redaction_list_t * rl,void * tag)1192 dsl_redaction_list_long_hold(dsl_pool_t *dp, redaction_list_t *rl, void *tag)
1193 {
1194 	ASSERT(dsl_pool_config_held(dp));
1195 	(void) zfs_refcount_add(&rl->rl_longholds, tag);
1196 }
1197 
1198 void
dsl_redaction_list_long_rele(redaction_list_t * rl,void * tag)1199 dsl_redaction_list_long_rele(redaction_list_t *rl, void *tag)
1200 {
1201 	(void) zfs_refcount_remove(&rl->rl_longholds, tag);
1202 }
1203 
1204 static void
redaction_list_evict_sync(void * rlu)1205 redaction_list_evict_sync(void *rlu)
1206 {
1207 	redaction_list_t *rl = rlu;
1208 	zfs_refcount_destroy(&rl->rl_longholds);
1209 
1210 	kmem_free(rl, sizeof (redaction_list_t));
1211 }
1212 
1213 void
dsl_redaction_list_rele(redaction_list_t * rl,void * tag)1214 dsl_redaction_list_rele(redaction_list_t *rl, void *tag)
1215 {
1216 	dmu_buf_rele(rl->rl_dbuf, tag);
1217 }
1218 
1219 int
dsl_redaction_list_hold_obj(dsl_pool_t * dp,uint64_t rlobj,void * tag,redaction_list_t ** rlp)1220 dsl_redaction_list_hold_obj(dsl_pool_t *dp, uint64_t rlobj, void *tag,
1221     redaction_list_t **rlp)
1222 {
1223 	objset_t *mos = dp->dp_meta_objset;
1224 	dmu_buf_t *dbuf;
1225 	redaction_list_t *rl;
1226 	int err;
1227 
1228 	ASSERT(dsl_pool_config_held(dp));
1229 
1230 	err = dmu_bonus_hold(mos, rlobj, tag, &dbuf);
1231 	if (err != 0)
1232 		return (err);
1233 
1234 	rl = dmu_buf_get_user(dbuf);
1235 	if (rl == NULL) {
1236 		redaction_list_t *winner = NULL;
1237 
1238 		rl = kmem_zalloc(sizeof (redaction_list_t), KM_SLEEP);
1239 		rl->rl_dbuf = dbuf;
1240 		rl->rl_object = rlobj;
1241 		rl->rl_phys = dbuf->db_data;
1242 		rl->rl_mos = dp->dp_meta_objset;
1243 		zfs_refcount_create(&rl->rl_longholds);
1244 		dmu_buf_init_user(&rl->rl_dbu, redaction_list_evict_sync, NULL,
1245 		    &rl->rl_dbuf);
1246 		if ((winner = dmu_buf_set_user_ie(dbuf, &rl->rl_dbu)) != NULL) {
1247 			kmem_free(rl, sizeof (*rl));
1248 			rl = winner;
1249 		}
1250 	}
1251 	*rlp = rl;
1252 	return (0);
1253 }
1254 
1255 /*
1256  * Snapshot ds is being destroyed.
1257  *
1258  * Adjust the "freed_before_next" of any bookmarks between this snap
1259  * and the previous snapshot, because their "next snapshot" is changing.
1260  *
1261  * If there are any bookmarks with HAS_FBN at this snapshot, remove
1262  * their HAS_SNAP flag (note: there can be at most one snapshot of
1263  * each filesystem at a given txg), and return B_TRUE.  In this case
1264  * the caller can not remove the key in the deadlist at this TXG, because
1265  * the HAS_FBN bookmarks require the key be there.
1266  *
1267  * Returns B_FALSE if there are no bookmarks with HAS_FBN at this
1268  * snapshot's TXG.  In this case the caller can remove the key in the
1269  * deadlist at this TXG.
1270  */
1271 boolean_t
dsl_bookmark_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)1272 dsl_bookmark_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
1273 {
1274 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1275 
1276 	dsl_dataset_t *head, *next;
1277 	VERIFY0(dsl_dataset_hold_obj(dp,
1278 	    dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj, FTAG, &head));
1279 	VERIFY0(dsl_dataset_hold_obj(dp,
1280 	    dsl_dataset_phys(ds)->ds_next_snap_obj, FTAG, &next));
1281 
1282 	/*
1283 	 * Find the first bookmark that HAS_FBN at or after the
1284 	 * previous snapshot.
1285 	 */
1286 	dsl_bookmark_node_t search = { 0 };
1287 	avl_index_t idx;
1288 	search.dbn_phys.zbm_creation_txg =
1289 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1290 	search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1291 	/*
1292 	 * The empty-string name can't be in the AVL, and it compares
1293 	 * before any entries with this TXG.
1294 	 */
1295 	search.dbn_name = "";
1296 	VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1297 	dsl_bookmark_node_t *dbn =
1298 	    avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1299 
1300 	/*
1301 	 * Iterate over all bookmarks that are at or after the previous
1302 	 * snapshot, and before this (being deleted) snapshot.  Adjust
1303 	 * their FBN based on their new next snapshot.
1304 	 */
1305 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg <
1306 	    dsl_dataset_phys(ds)->ds_creation_txg;
1307 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1308 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN))
1309 			continue;
1310 		/*
1311 		 * Increase our FBN by the amount of space that was live
1312 		 * (referenced) at the time of this bookmark (i.e.
1313 		 * birth <= zbm_creation_txg), and killed between this
1314 		 * (being deleted) snapshot and the next snapshot (i.e.
1315 		 * on the next snapshot's deadlist).  (Space killed before
1316 		 * this are already on our FBN.)
1317 		 */
1318 		uint64_t referenced, compressed, uncompressed;
1319 		dsl_deadlist_space_range(&next->ds_deadlist,
1320 		    0, dbn->dbn_phys.zbm_creation_txg,
1321 		    &referenced, &compressed, &uncompressed);
1322 		dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1323 		    referenced;
1324 		dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1325 		    compressed;
1326 		dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1327 		    uncompressed;
1328 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1329 		    dbn->dbn_name, sizeof (uint64_t),
1330 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1331 		    &dbn->dbn_phys, tx));
1332 	}
1333 	dsl_dataset_rele(next, FTAG);
1334 
1335 	/*
1336 	 * There may be several bookmarks at this txg (the TXG of the
1337 	 * snapshot being deleted).  We need to clear the SNAPSHOT_EXISTS
1338 	 * flag on all of them, and return TRUE if there is at least 1
1339 	 * bookmark here with HAS_FBN (thus preventing the deadlist
1340 	 * key from being removed).
1341 	 */
1342 	boolean_t rv = B_FALSE;
1343 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1344 	    dsl_dataset_phys(ds)->ds_creation_txg;
1345 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1346 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1347 			ASSERT(!(dbn->dbn_phys.zbm_flags &
1348 			    ZBM_FLAG_SNAPSHOT_EXISTS));
1349 			continue;
1350 		}
1351 		ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS);
1352 		dbn->dbn_phys.zbm_flags &= ~ZBM_FLAG_SNAPSHOT_EXISTS;
1353 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1354 		    dbn->dbn_name, sizeof (uint64_t),
1355 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1356 		    &dbn->dbn_phys, tx));
1357 		rv = B_TRUE;
1358 	}
1359 	dsl_dataset_rele(head, FTAG);
1360 	return (rv);
1361 }
1362 
1363 /*
1364  * A snapshot is being created of this (head) dataset.
1365  *
1366  * We don't keep keys in the deadlist for the most recent snapshot, or any
1367  * bookmarks at or after it, because there can't be any blocks on the
1368  * deadlist in this range.  Now that the most recent snapshot is after
1369  * all bookmarks, we need to add these keys.  Note that the caller always
1370  * adds a key at the previous snapshot, so we only add keys for bookmarks
1371  * after that.
1372  */
1373 void
dsl_bookmark_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)1374 dsl_bookmark_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
1375 {
1376 	uint64_t last_key_added = UINT64_MAX;
1377 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1378 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >
1379 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1380 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1381 		uint64_t creation_txg = dbn->dbn_phys.zbm_creation_txg;
1382 		ASSERT3U(creation_txg, <=, last_key_added);
1383 		/*
1384 		 * Note, there may be multiple bookmarks at this TXG,
1385 		 * and we only want to add the key for this TXG once.
1386 		 * The ds_bookmarks AVL is sorted by TXG, so we will visit
1387 		 * these bookmarks in sequence.
1388 		 */
1389 		if ((dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) &&
1390 		    creation_txg != last_key_added) {
1391 			dsl_deadlist_add_key(&ds->ds_deadlist,
1392 			    creation_txg, tx);
1393 			last_key_added = creation_txg;
1394 		}
1395 	}
1396 }
1397 
1398 /*
1399  * The next snapshot of the origin dataset has changed, due to
1400  * promote or clone swap.  If there are any bookmarks at this dataset,
1401  * we need to update their zbm_*_freed_before_next_snap to reflect this.
1402  * The head dataset has the relevant bookmarks in ds_bookmarks.
1403  */
1404 void
dsl_bookmark_next_changed(dsl_dataset_t * head,dsl_dataset_t * origin,dmu_tx_t * tx)1405 dsl_bookmark_next_changed(dsl_dataset_t *head, dsl_dataset_t *origin,
1406     dmu_tx_t *tx)
1407 {
1408 	dsl_pool_t *dp = dmu_tx_pool(tx);
1409 
1410 	/*
1411 	 * Find the first bookmark that HAS_FBN at the origin snapshot.
1412 	 */
1413 	dsl_bookmark_node_t search = { 0 };
1414 	avl_index_t idx;
1415 	search.dbn_phys.zbm_creation_txg =
1416 	    dsl_dataset_phys(origin)->ds_creation_txg;
1417 	search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1418 	/*
1419 	 * The empty-string name can't be in the AVL, and it compares
1420 	 * before any entries with this TXG.
1421 	 */
1422 	search.dbn_name = "";
1423 	VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1424 	dsl_bookmark_node_t *dbn =
1425 	    avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1426 
1427 	/*
1428 	 * Iterate over all bookmarks that are at the origin txg.
1429 	 * Adjust their FBN based on their new next snapshot.
1430 	 */
1431 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1432 	    dsl_dataset_phys(origin)->ds_creation_txg &&
1433 	    (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1434 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1435 
1436 		/*
1437 		 * Bookmark is at the origin, therefore its
1438 		 * "next dataset" is changing, so we need
1439 		 * to reset its FBN by recomputing it in
1440 		 * dsl_bookmark_set_phys().
1441 		 */
1442 		ASSERT3U(dbn->dbn_phys.zbm_guid, ==,
1443 		    dsl_dataset_phys(origin)->ds_guid);
1444 		ASSERT3U(dbn->dbn_phys.zbm_referenced_bytes_refd, ==,
1445 		    dsl_dataset_phys(origin)->ds_referenced_bytes);
1446 		ASSERT(dbn->dbn_phys.zbm_flags &
1447 		    ZBM_FLAG_SNAPSHOT_EXISTS);
1448 		/*
1449 		 * Save and restore the zbm_redaction_obj, which
1450 		 * is zeroed by dsl_bookmark_set_phys().
1451 		 */
1452 		uint64_t redaction_obj =
1453 		    dbn->dbn_phys.zbm_redaction_obj;
1454 		dsl_bookmark_set_phys(&dbn->dbn_phys, origin);
1455 		dbn->dbn_phys.zbm_redaction_obj = redaction_obj;
1456 
1457 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1458 		    dbn->dbn_name, sizeof (uint64_t),
1459 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1460 		    &dbn->dbn_phys, tx));
1461 	}
1462 }
1463 
1464 /*
1465  * This block is no longer referenced by this (head) dataset.
1466  *
1467  * Adjust the FBN of any bookmarks that reference this block, whose "next"
1468  * is the head dataset.
1469  */
1470 void
dsl_bookmark_block_killed(dsl_dataset_t * ds,const blkptr_t * bp,dmu_tx_t * tx)1471 dsl_bookmark_block_killed(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
1472 {
1473 	(void) tx;
1474 
1475 	/*
1476 	 * Iterate over bookmarks whose "next" is the head dataset.
1477 	 */
1478 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1479 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1480 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1481 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1482 		/*
1483 		 * If the block was live (referenced) at the time of this
1484 		 * bookmark, add its space to the bookmark's FBN.
1485 		 */
1486 		if (bp->blk_birth <= dbn->dbn_phys.zbm_creation_txg &&
1487 		    (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1488 			mutex_enter(&dbn->dbn_lock);
1489 			dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1490 			    bp_get_dsize_sync(dsl_dataset_get_spa(ds), bp);
1491 			dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1492 			    BP_GET_PSIZE(bp);
1493 			dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1494 			    BP_GET_UCSIZE(bp);
1495 			/*
1496 			 * Changing the ZAP object here would be too
1497 			 * expensive.  Also, we may be called from the zio
1498 			 * interrupt thread, which can't block on i/o.
1499 			 * Therefore, we mark this bookmark as dirty and
1500 			 * modify the ZAP once per txg, in
1501 			 * dsl_bookmark_sync_done().
1502 			 */
1503 			dbn->dbn_dirty = B_TRUE;
1504 			mutex_exit(&dbn->dbn_lock);
1505 		}
1506 	}
1507 }
1508 
1509 void
dsl_bookmark_sync_done(dsl_dataset_t * ds,dmu_tx_t * tx)1510 dsl_bookmark_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
1511 {
1512 	dsl_pool_t *dp = dmu_tx_pool(tx);
1513 
1514 	if (dsl_dataset_is_snapshot(ds))
1515 		return;
1516 
1517 	/*
1518 	 * We only dirty bookmarks that are at or after the most recent
1519 	 * snapshot.  We can't create snapshots between
1520 	 * dsl_bookmark_block_killed() and dsl_bookmark_sync_done(), so we
1521 	 * don't need to look at any bookmarks before ds_prev_snap_txg.
1522 	 */
1523 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1524 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1525 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1526 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1527 		if (dbn->dbn_dirty) {
1528 			/*
1529 			 * We only dirty nodes with HAS_FBN, therefore
1530 			 * we can always use the current bookmark struct size.
1531 			 */
1532 			ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1533 			VERIFY0(zap_update(dp->dp_meta_objset,
1534 			    ds->ds_bookmarks_obj,
1535 			    dbn->dbn_name, sizeof (uint64_t),
1536 			    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1537 			    &dbn->dbn_phys, tx));
1538 			dbn->dbn_dirty = B_FALSE;
1539 		}
1540 	}
1541 #ifdef ZFS_DEBUG
1542 	for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
1543 	    dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
1544 		ASSERT(!dbn->dbn_dirty);
1545 	}
1546 #endif
1547 }
1548 
1549 /*
1550  * Return the TXG of the most recent bookmark (or 0 if there are no bookmarks).
1551  */
1552 uint64_t
dsl_bookmark_latest_txg(dsl_dataset_t * ds)1553 dsl_bookmark_latest_txg(dsl_dataset_t *ds)
1554 {
1555 	ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool));
1556 	dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1557 	if (dbn == NULL)
1558 		return (0);
1559 	return (dbn->dbn_phys.zbm_creation_txg);
1560 }
1561 
1562 /*
1563  * Compare the redact_block_phys_t to the bookmark. If the last block in the
1564  * redact_block_phys_t is before the bookmark, return -1.  If the first block in
1565  * the redact_block_phys_t is after the bookmark, return 1.  Otherwise, the
1566  * bookmark is inside the range of the redact_block_phys_t, and we return 0.
1567  */
1568 static int
redact_block_zb_compare(redact_block_phys_t * first,zbookmark_phys_t * second)1569 redact_block_zb_compare(redact_block_phys_t *first,
1570     zbookmark_phys_t *second)
1571 {
1572 	/*
1573 	 * If the block_phys is for a previous object, or the last block in the
1574 	 * block_phys is strictly before the block in the bookmark, the
1575 	 * block_phys is earlier.
1576 	 */
1577 	if (first->rbp_object < second->zb_object ||
1578 	    (first->rbp_object == second->zb_object &&
1579 	    first->rbp_blkid + (redact_block_get_count(first) - 1) <
1580 	    second->zb_blkid)) {
1581 		return (-1);
1582 	}
1583 
1584 	/*
1585 	 * If the bookmark is for a previous object, or the block in the
1586 	 * bookmark is strictly before the first block in the block_phys, the
1587 	 * bookmark is earlier.
1588 	 */
1589 	if (first->rbp_object > second->zb_object ||
1590 	    (first->rbp_object == second->zb_object &&
1591 	    first->rbp_blkid > second->zb_blkid)) {
1592 		return (1);
1593 	}
1594 
1595 	return (0);
1596 }
1597 
1598 /*
1599  * Traverse the redaction list in the provided object, and call the callback for
1600  * each entry we find. Don't call the callback for any records before resume.
1601  */
1602 int
dsl_redaction_list_traverse(redaction_list_t * rl,zbookmark_phys_t * resume,rl_traverse_callback_t cb,void * arg)1603 dsl_redaction_list_traverse(redaction_list_t *rl, zbookmark_phys_t *resume,
1604     rl_traverse_callback_t cb, void *arg)
1605 {
1606 	objset_t *mos = rl->rl_mos;
1607 	int err = 0;
1608 
1609 	if (rl->rl_phys->rlp_last_object != UINT64_MAX ||
1610 	    rl->rl_phys->rlp_last_blkid != UINT64_MAX) {
1611 		/*
1612 		 * When we finish a send, we update the last object and offset
1613 		 * to UINT64_MAX.  If a send fails partway through, the last
1614 		 * object and offset will have some other value, indicating how
1615 		 * far the send got. The redaction list must be complete before
1616 		 * it can be traversed, so return EINVAL if the last object and
1617 		 * blkid are not set to UINT64_MAX.
1618 		 */
1619 		return (SET_ERROR(EINVAL));
1620 	}
1621 
1622 	/*
1623 	 * This allows us to skip the binary search and resume checking logic
1624 	 * below, if we're not resuming a redacted send.
1625 	 */
1626 	if (ZB_IS_ZERO(resume))
1627 		resume = NULL;
1628 
1629 	/*
1630 	 * Binary search for the point to resume from.
1631 	 */
1632 	uint64_t maxidx = rl->rl_phys->rlp_num_entries - 1;
1633 	uint64_t minidx = 0;
1634 	while (resume != NULL && maxidx > minidx) {
1635 		redact_block_phys_t rbp = { 0 };
1636 		ASSERT3U(maxidx, >, minidx);
1637 		uint64_t mididx = minidx + ((maxidx - minidx) / 2);
1638 		err = dmu_read(mos, rl->rl_object, mididx * sizeof (rbp),
1639 		    sizeof (rbp), &rbp, DMU_READ_NO_PREFETCH);
1640 		if (err != 0)
1641 			break;
1642 
1643 		int cmp = redact_block_zb_compare(&rbp, resume);
1644 
1645 		if (cmp == 0) {
1646 			minidx = mididx;
1647 			break;
1648 		} else if (cmp > 0) {
1649 			maxidx =
1650 			    (mididx == minidx ? minidx : mididx - 1);
1651 		} else {
1652 			minidx = mididx + 1;
1653 		}
1654 	}
1655 
1656 	unsigned int bufsize = SPA_OLD_MAXBLOCKSIZE;
1657 	redact_block_phys_t *buf = zio_data_buf_alloc(bufsize);
1658 
1659 	unsigned int entries_per_buf = bufsize / sizeof (redact_block_phys_t);
1660 	uint64_t start_block = minidx / entries_per_buf;
1661 	err = dmu_read(mos, rl->rl_object, start_block * bufsize, bufsize, buf,
1662 	    DMU_READ_PREFETCH);
1663 
1664 	for (uint64_t curidx = minidx;
1665 	    err == 0 && curidx < rl->rl_phys->rlp_num_entries;
1666 	    curidx++) {
1667 		/*
1668 		 * We read in the redaction list one block at a time.  Once we
1669 		 * finish with all the entries in a given block, we read in a
1670 		 * new one.  The predictive prefetcher will take care of any
1671 		 * prefetching, and this code shouldn't be the bottleneck, so we
1672 		 * don't need to do manual prefetching.
1673 		 */
1674 		if (curidx % entries_per_buf == 0) {
1675 			err = dmu_read(mos, rl->rl_object, curidx *
1676 			    sizeof (*buf), bufsize, buf,
1677 			    DMU_READ_PREFETCH);
1678 			if (err != 0)
1679 				break;
1680 		}
1681 		redact_block_phys_t *rb = &buf[curidx % entries_per_buf];
1682 		/*
1683 		 * If resume is non-null, we should either not send the data, or
1684 		 * null out resume so we don't have to keep doing these
1685 		 * comparisons.
1686 		 */
1687 		if (resume != NULL) {
1688 			/*
1689 			 * It is possible that after the binary search we got
1690 			 * a record before the resume point. There's two cases
1691 			 * where this can occur. If the record is the last
1692 			 * redaction record, and the resume point is after the
1693 			 * end of the redacted data, curidx will be the last
1694 			 * redaction record. In that case, the loop will end
1695 			 * after this iteration. The second case is if the
1696 			 * resume point is between two redaction records, the
1697 			 * binary search can return either the record before
1698 			 * or after the resume point. In that case, the next
1699 			 * iteration will be greater than the resume point.
1700 			 */
1701 			if (redact_block_zb_compare(rb, resume) < 0) {
1702 				ASSERT3U(curidx, ==, minidx);
1703 				continue;
1704 			} else {
1705 				/*
1706 				 * If the place to resume is in the middle of
1707 				 * the range described by this
1708 				 * redact_block_phys, then modify the
1709 				 * redact_block_phys in memory so we generate
1710 				 * the right records.
1711 				 */
1712 				if (resume->zb_object == rb->rbp_object &&
1713 				    resume->zb_blkid > rb->rbp_blkid) {
1714 					uint64_t diff = resume->zb_blkid -
1715 					    rb->rbp_blkid;
1716 					rb->rbp_blkid = resume->zb_blkid;
1717 					redact_block_set_count(rb,
1718 					    redact_block_get_count(rb) - diff);
1719 				}
1720 				resume = NULL;
1721 			}
1722 		}
1723 
1724 		if (cb(rb, arg) != 0) {
1725 			err = EINTR;
1726 			break;
1727 		}
1728 	}
1729 
1730 	zio_data_buf_free(buf, bufsize);
1731 	return (err);
1732 }
1733