xref: /freebsd/sys/contrib/openzfs/module/zfs/dsl_bookmark.c (revision 61145dc2b94f12f6a47344fb9aac702321880e43)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * This file and its contents are supplied under the terms of the
6  * Common Development and Distribution License ("CDDL"), version 1.0.
7  * You may only use this file in accordance with the terms of version
8  * 1.0 of the CDDL.
9  *
10  * A full copy of the text of the CDDL should have accompanied this
11  * source.  A copy of the CDDL is also available via the Internet at
12  * http://www.illumos.org/license/CDDL.
13  *
14  * CDDL HEADER END
15  */
16 
17 /*
18  * Copyright (c) 2013, 2018 by Delphix. All rights reserved.
19  * Copyright 2017 Nexenta Systems, Inc.
20  * Copyright 2019, 2020 by Christian Schwarz. All rights reserved.
21  */
22 
23 #include <sys/zfs_context.h>
24 #include <sys/dsl_dataset.h>
25 #include <sys/dsl_dir.h>
26 #include <sys/dsl_prop.h>
27 #include <sys/dsl_synctask.h>
28 #include <sys/dsl_destroy.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/arc.h>
32 #include <sys/zap.h>
33 #include <sys/zfeature.h>
34 #include <sys/spa.h>
35 #include <sys/dsl_bookmark.h>
36 #include <zfs_namecheck.h>
37 #include <sys/dmu_send.h>
38 #include <sys/dbuf.h>
39 
40 static int
dsl_bookmark_hold_ds(dsl_pool_t * dp,const char * fullname,dsl_dataset_t ** dsp,const void * tag,char ** shortnamep)41 dsl_bookmark_hold_ds(dsl_pool_t *dp, const char *fullname,
42     dsl_dataset_t **dsp, const void *tag, char **shortnamep)
43 {
44 	char buf[ZFS_MAX_DATASET_NAME_LEN];
45 	char *hashp;
46 
47 	if (strlen(fullname) >= ZFS_MAX_DATASET_NAME_LEN)
48 		return (SET_ERROR(ENAMETOOLONG));
49 	hashp = strchr(fullname, '#');
50 	if (hashp == NULL)
51 		return (SET_ERROR(EINVAL));
52 
53 	*shortnamep = hashp + 1;
54 	if (zfs_component_namecheck(*shortnamep, NULL, NULL))
55 		return (SET_ERROR(EINVAL));
56 	(void) strlcpy(buf, fullname, hashp - fullname + 1);
57 	return (dsl_dataset_hold(dp, buf, tag, dsp));
58 }
59 
60 /*
61  * When reading BOOKMARK_V1 bookmarks, the BOOKMARK_V2 fields are guaranteed
62  * to be zeroed.
63  *
64  * Returns ESRCH if bookmark is not found.
65  * Note, we need to use the ZAP rather than the AVL to look up bookmarks
66  * by name, because only the ZAP honors the casesensitivity setting.
67  */
68 int
dsl_bookmark_lookup_impl(dsl_dataset_t * ds,const char * shortname,zfs_bookmark_phys_t * bmark_phys)69 dsl_bookmark_lookup_impl(dsl_dataset_t *ds, const char *shortname,
70     zfs_bookmark_phys_t *bmark_phys)
71 {
72 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
73 	uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
74 	matchtype_t mt = 0;
75 	int err;
76 
77 	if (bmark_zapobj == 0)
78 		return (SET_ERROR(ESRCH));
79 
80 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
81 		mt = MT_NORMALIZE;
82 
83 	/*
84 	 * Zero out the bookmark in case the one stored on disk
85 	 * is in an older, shorter format.
86 	 */
87 	memset(bmark_phys, 0, sizeof (*bmark_phys));
88 
89 	err = zap_lookup_norm(mos, bmark_zapobj, shortname, sizeof (uint64_t),
90 	    sizeof (*bmark_phys) / sizeof (uint64_t), bmark_phys, mt, NULL, 0,
91 	    NULL);
92 
93 	return (err == ENOENT ? SET_ERROR(ESRCH) : err);
94 }
95 
96 /*
97  * If later_ds is non-NULL, this will return EXDEV if the specified bookmark
98  * does not represents an earlier point in later_ds's timeline.  However,
99  * bmp will still be filled in if we return EXDEV.
100  *
101  * Returns ENOENT if the dataset containing the bookmark does not exist.
102  * Returns ESRCH if the dataset exists but the bookmark was not found in it.
103  */
104 int
dsl_bookmark_lookup(dsl_pool_t * dp,const char * fullname,dsl_dataset_t * later_ds,zfs_bookmark_phys_t * bmp)105 dsl_bookmark_lookup(dsl_pool_t *dp, const char *fullname,
106     dsl_dataset_t *later_ds, zfs_bookmark_phys_t *bmp)
107 {
108 	char *shortname;
109 	dsl_dataset_t *ds;
110 	int error;
111 
112 	error = dsl_bookmark_hold_ds(dp, fullname, &ds, FTAG, &shortname);
113 	if (error != 0)
114 		return (error);
115 
116 	error = dsl_bookmark_lookup_impl(ds, shortname, bmp);
117 	if (error == 0 && later_ds != NULL) {
118 		if (!dsl_dataset_is_before(later_ds, ds, bmp->zbm_creation_txg))
119 			error = SET_ERROR(EXDEV);
120 	}
121 	dsl_dataset_rele(ds, FTAG);
122 	return (error);
123 }
124 
125 /*
126  * Validates that
127  * - bmark is a full dataset path of a bookmark (bookmark_namecheck)
128  * - source is a full path of a snapshot or bookmark
129  *   ({bookmark,snapshot}_namecheck)
130  *
131  * Returns 0 if valid, -1 otherwise.
132  */
133 static int
dsl_bookmark_create_nvl_validate_pair(const char * bmark,const char * source)134 dsl_bookmark_create_nvl_validate_pair(const char *bmark, const char *source)
135 {
136 	if (bookmark_namecheck(bmark, NULL, NULL) != 0)
137 		return (-1);
138 
139 	int is_bmark, is_snap;
140 	is_bmark = bookmark_namecheck(source, NULL, NULL) == 0;
141 	is_snap = snapshot_namecheck(source, NULL, NULL) == 0;
142 	if (!is_bmark && !is_snap)
143 		return (-1);
144 
145 	return (0);
146 }
147 
148 /*
149  * Check that the given nvlist corresponds to the following schema:
150  *  { newbookmark -> source, ... }
151  * where
152  * - each pair passes dsl_bookmark_create_nvl_validate_pair
153  * - all newbookmarks are in the same pool
154  * - all newbookmarks have unique names
155  *
156  * Note that this function is only validates above schema. Callers must ensure
157  * that the bookmarks can be created, e.g. that sources exist.
158  *
159  * Returns 0 if the nvlist adheres to above schema.
160  * Returns -1 if it doesn't.
161  */
162 int
dsl_bookmark_create_nvl_validate(nvlist_t * bmarks)163 dsl_bookmark_create_nvl_validate(nvlist_t *bmarks)
164 {
165 	const char *first = NULL;
166 	size_t first_len = 0;
167 
168 	for (nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
169 	    pair != NULL; pair = nvlist_next_nvpair(bmarks, pair)) {
170 
171 		const char *bmark = nvpair_name(pair);
172 		const char *source;
173 
174 		/* list structure: values must be snapshots XOR bookmarks */
175 		if (nvpair_value_string(pair, &source) != 0)
176 			return (-1);
177 		if (dsl_bookmark_create_nvl_validate_pair(bmark, source) != 0)
178 			return (-1);
179 
180 		/* same pool check */
181 		if (first == NULL) {
182 			const char *cp = strpbrk(bmark, "/#");
183 			if (cp == NULL)
184 				return (-1);
185 			first = bmark;
186 			first_len = cp - bmark;
187 		}
188 		if (strncmp(first, bmark, first_len) != 0)
189 			return (-1);
190 		switch (*(bmark + first_len)) {
191 			case '/': /* fallthrough */
192 			case '#':
193 				break;
194 			default:
195 				return (-1);
196 		}
197 
198 		/* unique newbookmark names; todo: O(n^2) */
199 		for (nvpair_t *pair2 = nvlist_next_nvpair(bmarks, pair);
200 		    pair2 != NULL; pair2 = nvlist_next_nvpair(bmarks, pair2)) {
201 			if (strcmp(nvpair_name(pair), nvpair_name(pair2)) == 0)
202 				return (-1);
203 		}
204 
205 	}
206 	return (0);
207 }
208 
209 /*
210  * expects that newbm and source have been validated using
211  * dsl_bookmark_create_nvl_validate_pair
212  */
213 static int
dsl_bookmark_create_check_impl(dsl_pool_t * dp,const char * newbm,const char * source)214 dsl_bookmark_create_check_impl(dsl_pool_t *dp,
215     const char *newbm, const char *source)
216 {
217 	ASSERT0(dsl_bookmark_create_nvl_validate_pair(newbm, source));
218 	/* defer source namecheck until we know it's a snapshot or bookmark */
219 
220 	int error;
221 	dsl_dataset_t *newbm_ds;
222 	char *newbm_short;
223 	zfs_bookmark_phys_t bmark_phys;
224 
225 	error = dsl_bookmark_hold_ds(dp, newbm, &newbm_ds, FTAG, &newbm_short);
226 	if (error != 0)
227 		return (error);
228 
229 	/* Verify that the new bookmark does not already exist */
230 	error = dsl_bookmark_lookup_impl(newbm_ds, newbm_short, &bmark_phys);
231 	switch (error) {
232 	case ESRCH:
233 		/* happy path: new bmark doesn't exist, proceed after switch */
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 		const char *new = nvpair_name(pair);
311 
312 		int error = schema_err;
313 		if (error == 0) {
314 			const 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 	zfs_bookmark_phys_t zero_phys = { 0 };
428 	ASSERT0(memcmp(((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,const 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,
443     const void *tag, 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 		boolean_t spill = B_FALSE;
465 		if (bookmark_redacted) {
466 			redact_snaps = dsredactsnaps;
467 			num_redact_snaps = dsnumsnaps;
468 		}
469 		int bonuslen = sizeof (redaction_list_phys_t) +
470 		    num_redact_snaps * sizeof (uint64_t);
471 		if (bonuslen > dmu_bonus_max())
472 			spill = B_TRUE;
473 		dbn->dbn_phys.zbm_redaction_obj = dmu_object_alloc(mos,
474 		    DMU_OTN_UINT64_METADATA, SPA_OLD_MAXBLOCKSIZE,
475 		    DMU_OTN_UINT64_METADATA, spill ? 0 : bonuslen, tx);
476 		spa_feature_incr(dp->dp_spa,
477 		    SPA_FEATURE_REDACTION_BOOKMARKS, tx);
478 		if (spill) {
479 			spa_feature_incr(dp->dp_spa,
480 			    SPA_FEATURE_REDACTION_LIST_SPILL, tx);
481 		}
482 
483 		VERIFY0(dsl_redaction_list_hold_obj(dp,
484 		    dbn->dbn_phys.zbm_redaction_obj, tag, &local_rl));
485 		dsl_redaction_list_long_hold(dp, local_rl, tag);
486 
487 		if (!spill) {
488 			ASSERT3U(local_rl->rl_bonus->db_size, >=, bonuslen);
489 			dmu_buf_will_dirty(local_rl->rl_bonus, tx);
490 		} else {
491 			dmu_buf_t *db;
492 			VERIFY0(dmu_spill_hold_by_bonus(local_rl->rl_bonus,
493 			    DB_RF_MUST_SUCCEED, FTAG, &db));
494 			dmu_buf_will_fill(db, tx, B_FALSE);
495 			VERIFY0(dbuf_spill_set_blksz(db, P2ROUNDUP(bonuslen,
496 			    SPA_MINBLOCKSIZE), tx));
497 			local_rl->rl_phys = db->db_data;
498 			local_rl->rl_dbuf = db;
499 		}
500 		memcpy(local_rl->rl_phys->rlp_snaps, redact_snaps,
501 		    sizeof (uint64_t) * num_redact_snaps);
502 		local_rl->rl_phys->rlp_num_snaps = num_redact_snaps;
503 		if (bookmark_redacted) {
504 			ASSERT3P(redaction_list, ==, NULL);
505 			local_rl->rl_phys->rlp_last_blkid = UINT64_MAX;
506 			local_rl->rl_phys->rlp_last_object = UINT64_MAX;
507 			dsl_redaction_list_long_rele(local_rl, tag);
508 			dsl_redaction_list_rele(local_rl, tag);
509 		} else {
510 			*redaction_list = local_rl;
511 		}
512 	}
513 
514 	if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
515 		spa_feature_incr(dp->dp_spa,
516 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
517 	}
518 
519 	dsl_bookmark_node_add(bmark_fs, dbn, tx);
520 
521 	spa_history_log_internal_ds(bmark_fs, "bookmark", tx,
522 	    "name=%s creation_txg=%llu target_snap=%llu redact_obj=%llu",
523 	    shortname, (longlong_t)dbn->dbn_phys.zbm_creation_txg,
524 	    (longlong_t)snapds->ds_object,
525 	    (longlong_t)dbn->dbn_phys.zbm_redaction_obj);
526 
527 	dsl_dataset_rele(bmark_fs, FTAG);
528 	dsl_dataset_rele(snapds, FTAG);
529 }
530 
531 
532 static void
dsl_bookmark_create_sync_impl_book(const char * new_name,const char * source_name,dmu_tx_t * tx)533 dsl_bookmark_create_sync_impl_book(
534     const char *new_name, const char *source_name, dmu_tx_t *tx)
535 {
536 	dsl_pool_t *dp = dmu_tx_pool(tx);
537 	dsl_dataset_t *bmark_fs_source, *bmark_fs_new;
538 	char *source_shortname, *new_shortname;
539 	zfs_bookmark_phys_t source_phys;
540 
541 	VERIFY0(dsl_bookmark_hold_ds(dp, source_name, &bmark_fs_source, FTAG,
542 	    &source_shortname));
543 	VERIFY0(dsl_bookmark_hold_ds(dp, new_name, &bmark_fs_new, FTAG,
544 	    &new_shortname));
545 
546 	/*
547 	 * create a copy of the source bookmark by copying most of its members
548 	 *
549 	 * Caveat: bookmarking a redaction bookmark yields a normal bookmark
550 	 * -----------------------------------------------------------------
551 	 * Reasoning:
552 	 * - The zbm_redaction_obj would be referred to by both source and new
553 	 *   bookmark, but would be destroyed once either source or new is
554 	 *   destroyed, resulting in use-after-free of the referred object.
555 	 * - User expectation when issuing the `zfs bookmark` command is that
556 	 *   a normal bookmark of the source is created
557 	 *
558 	 * Design Alternatives For Full Redaction Bookmark Copying:
559 	 * - reference-count the redaction object => would require on-disk
560 	 *   format change for existing redaction objects
561 	 * - Copy the redaction object => cannot be done in syncing context
562 	 *   because the redaction object might be too large
563 	 */
564 
565 	VERIFY0(dsl_bookmark_lookup_impl(bmark_fs_source, source_shortname,
566 	    &source_phys));
567 	dsl_bookmark_node_t *new_dbn = dsl_bookmark_node_alloc(new_shortname);
568 
569 	memcpy(&new_dbn->dbn_phys, &source_phys, sizeof (source_phys));
570 	new_dbn->dbn_phys.zbm_redaction_obj = 0;
571 
572 	/* update feature counters */
573 	if (new_dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
574 		spa_feature_incr(dp->dp_spa,
575 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
576 	}
577 	/* no need for redaction bookmark counter; nulled zbm_redaction_obj */
578 	/* dsl_bookmark_node_add bumps bookmarks and v2-bookmarks counter */
579 
580 	/*
581 	 * write new bookmark
582 	 *
583 	 * Note that dsl_bookmark_lookup_impl guarantees that, if source is a
584 	 * v1 bookmark, the v2-only fields are zeroed.
585 	 * And dsl_bookmark_node_add writes back a v1-sized bookmark if
586 	 * v2 bookmarks are disabled and/or v2-only fields are zeroed.
587 	 * => bookmark copying works on pre-bookmark-v2 pools
588 	 */
589 	dsl_bookmark_node_add(bmark_fs_new, new_dbn, tx);
590 
591 	spa_history_log_internal_ds(bmark_fs_source, "bookmark", tx,
592 	    "name=%s creation_txg=%llu source_guid=%llu",
593 	    new_shortname, (longlong_t)new_dbn->dbn_phys.zbm_creation_txg,
594 	    (longlong_t)source_phys.zbm_guid);
595 
596 	dsl_dataset_rele(bmark_fs_source, FTAG);
597 	dsl_dataset_rele(bmark_fs_new, FTAG);
598 }
599 
600 void
dsl_bookmark_create_sync(void * arg,dmu_tx_t * tx)601 dsl_bookmark_create_sync(void *arg, dmu_tx_t *tx)
602 {
603 	dsl_bookmark_create_arg_t *dbca = arg;
604 
605 	ASSERT(spa_feature_is_enabled(dmu_tx_pool(tx)->dp_spa,
606 	    SPA_FEATURE_BOOKMARKS));
607 
608 	for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
609 	    pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
610 
611 		const char *new = nvpair_name(pair);
612 		const char *source = fnvpair_value_string(pair);
613 
614 		if (strchr(source, '@') != NULL) {
615 			dsl_bookmark_create_sync_impl_snap(new, source, tx,
616 			    0, NULL, NULL, NULL);
617 		} else if (strchr(source, '#') != NULL) {
618 			dsl_bookmark_create_sync_impl_book(new, source, tx);
619 		} else {
620 			panic("unreachable code");
621 		}
622 
623 	}
624 }
625 
626 /*
627  * The bookmarks must all be in the same pool.
628  */
629 int
dsl_bookmark_create(nvlist_t * bmarks,nvlist_t * errors)630 dsl_bookmark_create(nvlist_t *bmarks, nvlist_t *errors)
631 {
632 	nvpair_t *pair;
633 	dsl_bookmark_create_arg_t dbca;
634 
635 	pair = nvlist_next_nvpair(bmarks, NULL);
636 	if (pair == NULL)
637 		return (0);
638 
639 	dbca.dbca_bmarks = bmarks;
640 	dbca.dbca_errors = errors;
641 
642 	return (dsl_sync_task(nvpair_name(pair), dsl_bookmark_create_check,
643 	    dsl_bookmark_create_sync, &dbca,
644 	    fnvlist_num_pairs(bmarks), ZFS_SPACE_CHECK_NORMAL));
645 }
646 
647 static int
dsl_bookmark_create_redacted_check(void * arg,dmu_tx_t * tx)648 dsl_bookmark_create_redacted_check(void *arg, dmu_tx_t *tx)
649 {
650 	dsl_bookmark_create_redacted_arg_t *dbcra = arg;
651 	dsl_pool_t *dp = dmu_tx_pool(tx);
652 	int rv = 0;
653 
654 	if (!spa_feature_is_enabled(dp->dp_spa,
655 	    SPA_FEATURE_REDACTION_BOOKMARKS))
656 		return (SET_ERROR(ENOTSUP));
657 	/*
658 	 * If the list of redact snaps will not fit in the bonus buffer (or
659 	 * spill block, with the REDACTION_LIST_SPILL feature) with the
660 	 * furthest reached object and offset, fail.
661 	 */
662 	uint64_t snaplimit = ((spa_feature_is_enabled(dp->dp_spa,
663 	    SPA_FEATURE_REDACTION_LIST_SPILL) ? spa_maxblocksize(dp->dp_spa) :
664 	    dmu_bonus_max()) -
665 	    sizeof (redaction_list_phys_t)) / sizeof (uint64_t);
666 	if (dbcra->dbcra_numsnaps > snaplimit)
667 		return (SET_ERROR(E2BIG));
668 
669 	if (dsl_bookmark_create_nvl_validate_pair(
670 	    dbcra->dbcra_bmark, dbcra->dbcra_snap) != 0)
671 		return (SET_ERROR(EINVAL));
672 
673 	rv = dsl_bookmark_create_check_impl(dp,
674 	    dbcra->dbcra_bmark, dbcra->dbcra_snap);
675 	return (rv);
676 }
677 
678 static void
dsl_bookmark_create_redacted_sync(void * arg,dmu_tx_t * tx)679 dsl_bookmark_create_redacted_sync(void *arg, dmu_tx_t *tx)
680 {
681 	dsl_bookmark_create_redacted_arg_t *dbcra = arg;
682 	dsl_bookmark_create_sync_impl_snap(dbcra->dbcra_bmark,
683 	    dbcra->dbcra_snap, tx, dbcra->dbcra_numsnaps, dbcra->dbcra_snaps,
684 	    dbcra->dbcra_tag, dbcra->dbcra_rl);
685 }
686 
687 int
dsl_bookmark_create_redacted(const char * bookmark,const char * snapshot,uint64_t numsnaps,uint64_t * snapguids,const void * tag,redaction_list_t ** rl)688 dsl_bookmark_create_redacted(const char *bookmark, const char *snapshot,
689     uint64_t numsnaps, uint64_t *snapguids, const void *tag,
690     redaction_list_t **rl)
691 {
692 	dsl_bookmark_create_redacted_arg_t dbcra;
693 
694 	dbcra.dbcra_bmark = bookmark;
695 	dbcra.dbcra_snap = snapshot;
696 	dbcra.dbcra_rl = rl;
697 	dbcra.dbcra_numsnaps = numsnaps;
698 	dbcra.dbcra_snaps = snapguids;
699 	dbcra.dbcra_tag = tag;
700 
701 	return (dsl_sync_task(bookmark, dsl_bookmark_create_redacted_check,
702 	    dsl_bookmark_create_redacted_sync, &dbcra, 5,
703 	    ZFS_SPACE_CHECK_NORMAL));
704 }
705 
706 /*
707  * Retrieve the list of properties given in the 'props' nvlist for a bookmark.
708  * If 'props' is NULL, retrieves all properties.
709  */
710 static void
dsl_bookmark_fetch_props(dsl_pool_t * dp,zfs_bookmark_phys_t * bmark_phys,nvlist_t * props,nvlist_t * out_props)711 dsl_bookmark_fetch_props(dsl_pool_t *dp, zfs_bookmark_phys_t *bmark_phys,
712     nvlist_t *props, nvlist_t *out_props)
713 {
714 	ASSERT3P(dp, !=, NULL);
715 	ASSERT3P(bmark_phys, !=, NULL);
716 	ASSERT3P(out_props, !=, NULL);
717 	ASSERT(RRW_LOCK_HELD(&dp->dp_config_rwlock));
718 
719 	if (props == NULL || nvlist_exists(props,
720 	    zfs_prop_to_name(ZFS_PROP_GUID))) {
721 		dsl_prop_nvlist_add_uint64(out_props,
722 		    ZFS_PROP_GUID, bmark_phys->zbm_guid);
723 	}
724 	if (props == NULL || nvlist_exists(props,
725 	    zfs_prop_to_name(ZFS_PROP_CREATETXG))) {
726 		dsl_prop_nvlist_add_uint64(out_props,
727 		    ZFS_PROP_CREATETXG, bmark_phys->zbm_creation_txg);
728 	}
729 	if (props == NULL || nvlist_exists(props,
730 	    zfs_prop_to_name(ZFS_PROP_CREATION))) {
731 		dsl_prop_nvlist_add_uint64(out_props,
732 		    ZFS_PROP_CREATION, bmark_phys->zbm_creation_time);
733 	}
734 	if (props == NULL || nvlist_exists(props,
735 	    zfs_prop_to_name(ZFS_PROP_IVSET_GUID))) {
736 		dsl_prop_nvlist_add_uint64(out_props,
737 		    ZFS_PROP_IVSET_GUID, bmark_phys->zbm_ivset_guid);
738 	}
739 	if (bmark_phys->zbm_flags & ZBM_FLAG_HAS_FBN) {
740 		if (props == NULL || nvlist_exists(props,
741 		    zfs_prop_to_name(ZFS_PROP_REFERENCED))) {
742 			dsl_prop_nvlist_add_uint64(out_props,
743 			    ZFS_PROP_REFERENCED,
744 			    bmark_phys->zbm_referenced_bytes_refd);
745 		}
746 		if (props == NULL || nvlist_exists(props,
747 		    zfs_prop_to_name(ZFS_PROP_LOGICALREFERENCED))) {
748 			dsl_prop_nvlist_add_uint64(out_props,
749 			    ZFS_PROP_LOGICALREFERENCED,
750 			    bmark_phys->zbm_uncompressed_bytes_refd);
751 		}
752 		if (props == NULL || nvlist_exists(props,
753 		    zfs_prop_to_name(ZFS_PROP_REFRATIO))) {
754 			uint64_t ratio =
755 			    bmark_phys->zbm_compressed_bytes_refd == 0 ? 100 :
756 			    bmark_phys->zbm_uncompressed_bytes_refd * 100 /
757 			    bmark_phys->zbm_compressed_bytes_refd;
758 			dsl_prop_nvlist_add_uint64(out_props,
759 			    ZFS_PROP_REFRATIO, ratio);
760 		}
761 	}
762 
763 	if ((props == NULL || nvlist_exists(props, "redact_snaps") ||
764 	    nvlist_exists(props, "redact_complete")) &&
765 	    bmark_phys->zbm_redaction_obj != 0) {
766 		redaction_list_t *rl;
767 		int err = dsl_redaction_list_hold_obj(dp,
768 		    bmark_phys->zbm_redaction_obj, FTAG, &rl);
769 		if (err == 0) {
770 			if (nvlist_exists(props, "redact_snaps")) {
771 				nvlist_t *nvl;
772 				nvl = fnvlist_alloc();
773 				fnvlist_add_uint64_array(nvl, ZPROP_VALUE,
774 				    rl->rl_phys->rlp_snaps,
775 				    rl->rl_phys->rlp_num_snaps);
776 				fnvlist_add_nvlist(out_props, "redact_snaps",
777 				    nvl);
778 				nvlist_free(nvl);
779 			}
780 			if (nvlist_exists(props, "redact_complete")) {
781 				nvlist_t *nvl;
782 				nvl = fnvlist_alloc();
783 				fnvlist_add_boolean_value(nvl, ZPROP_VALUE,
784 				    rl->rl_phys->rlp_last_blkid == UINT64_MAX &&
785 				    rl->rl_phys->rlp_last_object == UINT64_MAX);
786 				fnvlist_add_nvlist(out_props, "redact_complete",
787 				    nvl);
788 				nvlist_free(nvl);
789 			}
790 			dsl_redaction_list_rele(rl, FTAG);
791 		}
792 	}
793 }
794 
795 int
dsl_get_bookmarks_impl(dsl_dataset_t * ds,nvlist_t * props,nvlist_t * outnvl)796 dsl_get_bookmarks_impl(dsl_dataset_t *ds, nvlist_t *props, nvlist_t *outnvl)
797 {
798 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
799 
800 	ASSERT(dsl_pool_config_held(dp));
801 
802 	if (dsl_dataset_is_snapshot(ds))
803 		return (SET_ERROR(EINVAL));
804 
805 	for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
806 	    dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
807 		nvlist_t *out_props = fnvlist_alloc();
808 
809 		dsl_bookmark_fetch_props(dp, &dbn->dbn_phys, props, out_props);
810 
811 		fnvlist_add_nvlist(outnvl, dbn->dbn_name, out_props);
812 		fnvlist_free(out_props);
813 	}
814 	return (0);
815 }
816 
817 /*
818  * Comparison func for ds_bookmarks AVL tree.  We sort the bookmarks by
819  * their TXG, then by their FBN-ness.  The "FBN-ness" component ensures
820  * that all bookmarks at the same TXG that HAS_FBN are adjacent, which
821  * dsl_bookmark_destroy_sync_impl() depends on.  Note that there may be
822  * multiple bookmarks at the same TXG (with the same FBN-ness).  In this
823  * case we differentiate them by an arbitrary metric (in this case,
824  * their names).
825  */
826 static int
dsl_bookmark_compare(const void * l,const void * r)827 dsl_bookmark_compare(const void *l, const void *r)
828 {
829 	const dsl_bookmark_node_t *ldbn = l;
830 	const dsl_bookmark_node_t *rdbn = r;
831 
832 	int64_t cmp = TREE_CMP(ldbn->dbn_phys.zbm_creation_txg,
833 	    rdbn->dbn_phys.zbm_creation_txg);
834 	if (likely(cmp))
835 		return (cmp);
836 	cmp = TREE_CMP((ldbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN),
837 	    (rdbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
838 	if (likely(cmp))
839 		return (cmp);
840 	cmp = strcmp(ldbn->dbn_name, rdbn->dbn_name);
841 	return (TREE_ISIGN(cmp));
842 }
843 
844 /*
845  * Cache this (head) dataset's bookmarks in the ds_bookmarks AVL tree.
846  */
847 int
dsl_bookmark_init_ds(dsl_dataset_t * ds)848 dsl_bookmark_init_ds(dsl_dataset_t *ds)
849 {
850 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
851 	objset_t *mos = dp->dp_meta_objset;
852 
853 	ASSERT(!ds->ds_is_snapshot);
854 
855 	avl_create(&ds->ds_bookmarks, dsl_bookmark_compare,
856 	    sizeof (dsl_bookmark_node_t),
857 	    offsetof(dsl_bookmark_node_t, dbn_node));
858 
859 	if (!dsl_dataset_is_zapified(ds))
860 		return (0);
861 
862 	int zaperr = zap_lookup(mos, ds->ds_object, DS_FIELD_BOOKMARK_NAMES,
863 	    sizeof (ds->ds_bookmarks_obj), 1, &ds->ds_bookmarks_obj);
864 	if (zaperr == ENOENT)
865 		return (0);
866 	if (zaperr != 0)
867 		return (zaperr);
868 
869 	if (ds->ds_bookmarks_obj == 0)
870 		return (0);
871 
872 	int err = 0;
873 	zap_cursor_t zc;
874 	zap_attribute_t *attr;
875 
876 	attr = zap_attribute_alloc();
877 	for (zap_cursor_init(&zc, mos, ds->ds_bookmarks_obj);
878 	    (err = zap_cursor_retrieve(&zc, attr)) == 0;
879 	    zap_cursor_advance(&zc)) {
880 		dsl_bookmark_node_t *dbn =
881 		    dsl_bookmark_node_alloc(attr->za_name);
882 
883 		err = dsl_bookmark_lookup_impl(ds,
884 		    dbn->dbn_name, &dbn->dbn_phys);
885 		ASSERT3U(err, !=, ENOENT);
886 		if (err != 0) {
887 			kmem_free(dbn, sizeof (*dbn));
888 			break;
889 		}
890 		avl_add(&ds->ds_bookmarks, dbn);
891 	}
892 	zap_cursor_fini(&zc);
893 	zap_attribute_free(attr);
894 	if (err == ENOENT)
895 		err = 0;
896 	return (err);
897 }
898 
899 void
dsl_bookmark_fini_ds(dsl_dataset_t * ds)900 dsl_bookmark_fini_ds(dsl_dataset_t *ds)
901 {
902 	void *cookie = NULL;
903 	dsl_bookmark_node_t *dbn;
904 
905 	if (ds->ds_is_snapshot)
906 		return;
907 
908 	while ((dbn = avl_destroy_nodes(&ds->ds_bookmarks, &cookie)) != NULL) {
909 		spa_strfree(dbn->dbn_name);
910 		mutex_destroy(&dbn->dbn_lock);
911 		kmem_free(dbn, sizeof (*dbn));
912 	}
913 	avl_destroy(&ds->ds_bookmarks);
914 }
915 
916 /*
917  * Retrieve the bookmarks that exist in the specified dataset, and the
918  * requested properties of each bookmark.
919  *
920  * The "props" nvlist specifies which properties are requested.
921  * See lzc_get_bookmarks() for the list of valid properties.
922  */
923 int
dsl_get_bookmarks(const char * dsname,nvlist_t * props,nvlist_t * outnvl)924 dsl_get_bookmarks(const char *dsname, nvlist_t *props, nvlist_t *outnvl)
925 {
926 	dsl_pool_t *dp;
927 	dsl_dataset_t *ds;
928 	int err;
929 
930 	err = dsl_pool_hold(dsname, FTAG, &dp);
931 	if (err != 0)
932 		return (err);
933 	err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
934 	if (err != 0) {
935 		dsl_pool_rele(dp, FTAG);
936 		return (err);
937 	}
938 
939 	err = dsl_get_bookmarks_impl(ds, props, outnvl);
940 
941 	dsl_dataset_rele(ds, FTAG);
942 	dsl_pool_rele(dp, FTAG);
943 	return (err);
944 }
945 
946 /*
947  * Retrieve all properties for a single bookmark in the given dataset.
948  */
949 int
dsl_get_bookmark_props(const char * dsname,const char * bmname,nvlist_t * props)950 dsl_get_bookmark_props(const char *dsname, const char *bmname, nvlist_t *props)
951 {
952 	dsl_pool_t *dp;
953 	dsl_dataset_t *ds;
954 	zfs_bookmark_phys_t bmark_phys = { 0 };
955 	int err;
956 
957 	err = dsl_pool_hold(dsname, FTAG, &dp);
958 	if (err != 0)
959 		return (err);
960 	err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
961 	if (err != 0) {
962 		dsl_pool_rele(dp, FTAG);
963 		return (err);
964 	}
965 
966 	err = dsl_bookmark_lookup_impl(ds, bmname, &bmark_phys);
967 	if (err != 0)
968 		goto out;
969 
970 	dsl_bookmark_fetch_props(dp, &bmark_phys, NULL, props);
971 out:
972 	dsl_dataset_rele(ds, FTAG);
973 	dsl_pool_rele(dp, FTAG);
974 	return (err);
975 }
976 
977 typedef struct dsl_bookmark_destroy_arg {
978 	nvlist_t *dbda_bmarks;
979 	nvlist_t *dbda_success;
980 	nvlist_t *dbda_errors;
981 } dsl_bookmark_destroy_arg_t;
982 
983 static void
dsl_bookmark_destroy_sync_impl(dsl_dataset_t * ds,const char * name,dmu_tx_t * tx)984 dsl_bookmark_destroy_sync_impl(dsl_dataset_t *ds, const char *name,
985     dmu_tx_t *tx)
986 {
987 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
988 	uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
989 	matchtype_t mt = 0;
990 	uint64_t int_size, num_ints;
991 	/*
992 	 * 'search' must be zeroed so that dbn_flags (which is used in
993 	 * dsl_bookmark_compare()) will be zeroed even if the on-disk
994 	 * (in ZAP) bookmark is shorter than offsetof(dbn_flags).
995 	 */
996 	dsl_bookmark_node_t search = { 0 };
997 	char realname[ZFS_MAX_DATASET_NAME_LEN];
998 
999 	/*
1000 	 * Find the real name of this bookmark, which may be different
1001 	 * from the given name if the dataset is case-insensitive.  Then
1002 	 * use the real name to find the node in the ds_bookmarks AVL tree.
1003 	 */
1004 
1005 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
1006 		mt = MT_NORMALIZE;
1007 
1008 	VERIFY0(zap_length(mos, bmark_zapobj, name, &int_size, &num_ints));
1009 
1010 	ASSERT3U(int_size, ==, sizeof (uint64_t));
1011 
1012 	if (num_ints * int_size > BOOKMARK_PHYS_SIZE_V1) {
1013 		spa_feature_decr(dmu_objset_spa(mos),
1014 		    SPA_FEATURE_BOOKMARK_V2, tx);
1015 	}
1016 	VERIFY0(zap_lookup_norm(mos, bmark_zapobj, name, sizeof (uint64_t),
1017 	    num_ints, &search.dbn_phys, mt, realname, sizeof (realname), NULL));
1018 
1019 	search.dbn_name = realname;
1020 	dsl_bookmark_node_t *dbn = avl_find(&ds->ds_bookmarks, &search, NULL);
1021 	ASSERT(dbn != NULL);
1022 
1023 	if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
1024 		/*
1025 		 * If this bookmark HAS_FBN, and it is before the most
1026 		 * recent snapshot, then its TXG is a key in the head's
1027 		 * deadlist (and all clones' heads' deadlists).  If this is
1028 		 * the last thing keeping the key (i.e. there are no more
1029 		 * bookmarks with HAS_FBN at this TXG, and there is no
1030 		 * snapshot at this TXG), then remove the key.
1031 		 *
1032 		 * Note that this algorithm depends on ds_bookmarks being
1033 		 * sorted such that all bookmarks at the same TXG with
1034 		 * HAS_FBN are adjacent (with no non-HAS_FBN bookmarks
1035 		 * at the same TXG in between them).  If this were not
1036 		 * the case, we would need to examine *all* bookmarks
1037 		 * at this TXG, rather than just the adjacent ones.
1038 		 */
1039 
1040 		dsl_bookmark_node_t *dbn_prev =
1041 		    AVL_PREV(&ds->ds_bookmarks, dbn);
1042 		dsl_bookmark_node_t *dbn_next =
1043 		    AVL_NEXT(&ds->ds_bookmarks, dbn);
1044 
1045 		boolean_t more_bookmarks_at_this_txg =
1046 		    (dbn_prev != NULL && dbn_prev->dbn_phys.zbm_creation_txg ==
1047 		    dbn->dbn_phys.zbm_creation_txg &&
1048 		    (dbn_prev->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) ||
1049 		    (dbn_next != NULL && dbn_next->dbn_phys.zbm_creation_txg ==
1050 		    dbn->dbn_phys.zbm_creation_txg &&
1051 		    (dbn_next->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
1052 
1053 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS) &&
1054 		    !more_bookmarks_at_this_txg &&
1055 		    dbn->dbn_phys.zbm_creation_txg <
1056 		    dsl_dataset_phys(ds)->ds_prev_snap_txg) {
1057 			dsl_dir_remove_clones_key(ds->ds_dir,
1058 			    dbn->dbn_phys.zbm_creation_txg, tx);
1059 			dsl_deadlist_remove_key(&ds->ds_deadlist,
1060 			    dbn->dbn_phys.zbm_creation_txg, tx);
1061 		}
1062 
1063 		spa_feature_decr(dmu_objset_spa(mos),
1064 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
1065 	}
1066 
1067 	if (dbn->dbn_phys.zbm_redaction_obj != 0) {
1068 		dnode_t *rl;
1069 		VERIFY0(dnode_hold(mos,
1070 		    dbn->dbn_phys.zbm_redaction_obj, FTAG, &rl));
1071 		if (rl->dn_have_spill) {
1072 			spa_feature_decr(dmu_objset_spa(mos),
1073 			    SPA_FEATURE_REDACTION_LIST_SPILL, tx);
1074 		}
1075 		dnode_rele(rl, FTAG);
1076 		VERIFY0(dmu_object_free(mos,
1077 		    dbn->dbn_phys.zbm_redaction_obj, tx));
1078 		spa_feature_decr(dmu_objset_spa(mos),
1079 		    SPA_FEATURE_REDACTION_BOOKMARKS, tx);
1080 	}
1081 
1082 	avl_remove(&ds->ds_bookmarks, dbn);
1083 	spa_strfree(dbn->dbn_name);
1084 	mutex_destroy(&dbn->dbn_lock);
1085 	kmem_free(dbn, sizeof (*dbn));
1086 
1087 	VERIFY0(zap_remove_norm(mos, bmark_zapobj, name, mt, tx));
1088 }
1089 
1090 static int
dsl_bookmark_destroy_check(void * arg,dmu_tx_t * tx)1091 dsl_bookmark_destroy_check(void *arg, dmu_tx_t *tx)
1092 {
1093 	dsl_bookmark_destroy_arg_t *dbda = arg;
1094 	dsl_pool_t *dp = dmu_tx_pool(tx);
1095 	int rv = 0;
1096 
1097 	ASSERT(nvlist_empty(dbda->dbda_success));
1098 	ASSERT(nvlist_empty(dbda->dbda_errors));
1099 
1100 	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
1101 		return (0);
1102 
1103 	for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_bmarks, NULL);
1104 	    pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_bmarks, pair)) {
1105 		const char *fullname = nvpair_name(pair);
1106 		dsl_dataset_t *ds;
1107 		zfs_bookmark_phys_t bm;
1108 		int error;
1109 		char *shortname;
1110 
1111 		error = dsl_bookmark_hold_ds(dp, fullname, &ds,
1112 		    FTAG, &shortname);
1113 		if (error == ENOENT) {
1114 			/* ignore it; the bookmark is "already destroyed" */
1115 			continue;
1116 		}
1117 		if (error == 0) {
1118 			error = dsl_bookmark_lookup_impl(ds, shortname, &bm);
1119 			dsl_dataset_rele(ds, FTAG);
1120 			if (error == ESRCH) {
1121 				/*
1122 				 * ignore it; the bookmark is
1123 				 * "already destroyed"
1124 				 */
1125 				continue;
1126 			}
1127 			if (error == 0 && bm.zbm_redaction_obj != 0) {
1128 				redaction_list_t *rl = NULL;
1129 				error = dsl_redaction_list_hold_obj(tx->tx_pool,
1130 				    bm.zbm_redaction_obj, FTAG, &rl);
1131 				if (error == ENOENT) {
1132 					error = 0;
1133 				} else if (error == 0 &&
1134 				    dsl_redaction_list_long_held(rl)) {
1135 					error = SET_ERROR(EBUSY);
1136 				}
1137 				if (rl != NULL) {
1138 					dsl_redaction_list_rele(rl, FTAG);
1139 				}
1140 			}
1141 		}
1142 		if (error == 0) {
1143 			if (dmu_tx_is_syncing(tx)) {
1144 				fnvlist_add_boolean(dbda->dbda_success,
1145 				    fullname);
1146 			}
1147 		} else {
1148 			fnvlist_add_int32(dbda->dbda_errors, fullname, error);
1149 			rv = error;
1150 		}
1151 	}
1152 	return (rv);
1153 }
1154 
1155 static void
dsl_bookmark_destroy_sync(void * arg,dmu_tx_t * tx)1156 dsl_bookmark_destroy_sync(void *arg, dmu_tx_t *tx)
1157 {
1158 	dsl_bookmark_destroy_arg_t *dbda = arg;
1159 	dsl_pool_t *dp = dmu_tx_pool(tx);
1160 	objset_t *mos = dp->dp_meta_objset;
1161 
1162 	for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_success, NULL);
1163 	    pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_success, pair)) {
1164 		dsl_dataset_t *ds;
1165 		char *shortname;
1166 		uint64_t zap_cnt;
1167 
1168 		VERIFY0(dsl_bookmark_hold_ds(dp, nvpair_name(pair),
1169 		    &ds, FTAG, &shortname));
1170 		dsl_bookmark_destroy_sync_impl(ds, shortname, tx);
1171 
1172 		/*
1173 		 * If all of this dataset's bookmarks have been destroyed,
1174 		 * free the zap object and decrement the feature's use count.
1175 		 */
1176 		VERIFY0(zap_count(mos, ds->ds_bookmarks_obj, &zap_cnt));
1177 		if (zap_cnt == 0) {
1178 			dmu_buf_will_dirty(ds->ds_dbuf, tx);
1179 			VERIFY0(zap_destroy(mos, ds->ds_bookmarks_obj, tx));
1180 			ds->ds_bookmarks_obj = 0;
1181 			spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
1182 			VERIFY0(zap_remove(mos, ds->ds_object,
1183 			    DS_FIELD_BOOKMARK_NAMES, tx));
1184 		}
1185 
1186 		spa_history_log_internal_ds(ds, "remove bookmark", tx,
1187 		    "name=%s", shortname);
1188 
1189 		dsl_dataset_rele(ds, FTAG);
1190 	}
1191 }
1192 
1193 /*
1194  * The bookmarks must all be in the same pool.
1195  */
1196 int
dsl_bookmark_destroy(nvlist_t * bmarks,nvlist_t * errors)1197 dsl_bookmark_destroy(nvlist_t *bmarks, nvlist_t *errors)
1198 {
1199 	int rv;
1200 	dsl_bookmark_destroy_arg_t dbda;
1201 	nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
1202 	if (pair == NULL)
1203 		return (0);
1204 
1205 	dbda.dbda_bmarks = bmarks;
1206 	dbda.dbda_errors = errors;
1207 	dbda.dbda_success = fnvlist_alloc();
1208 
1209 	rv = dsl_sync_task(nvpair_name(pair), dsl_bookmark_destroy_check,
1210 	    dsl_bookmark_destroy_sync, &dbda, fnvlist_num_pairs(bmarks),
1211 	    ZFS_SPACE_CHECK_RESERVED);
1212 	fnvlist_free(dbda.dbda_success);
1213 	return (rv);
1214 }
1215 
1216 /* Return B_TRUE if there are any long holds on this dataset. */
1217 boolean_t
dsl_redaction_list_long_held(redaction_list_t * rl)1218 dsl_redaction_list_long_held(redaction_list_t *rl)
1219 {
1220 	return (!zfs_refcount_is_zero(&rl->rl_longholds));
1221 }
1222 
1223 void
dsl_redaction_list_long_hold(dsl_pool_t * dp,redaction_list_t * rl,const void * tag)1224 dsl_redaction_list_long_hold(dsl_pool_t *dp, redaction_list_t *rl,
1225     const void *tag)
1226 {
1227 	ASSERT(dsl_pool_config_held(dp));
1228 	(void) zfs_refcount_add(&rl->rl_longholds, tag);
1229 }
1230 
1231 void
dsl_redaction_list_long_rele(redaction_list_t * rl,const void * tag)1232 dsl_redaction_list_long_rele(redaction_list_t *rl, const void *tag)
1233 {
1234 	(void) zfs_refcount_remove(&rl->rl_longholds, tag);
1235 }
1236 
1237 static void
redaction_list_evict_sync(void * rlu)1238 redaction_list_evict_sync(void *rlu)
1239 {
1240 	redaction_list_t *rl = rlu;
1241 	zfs_refcount_destroy(&rl->rl_longholds);
1242 
1243 	kmem_free(rl, sizeof (redaction_list_t));
1244 }
1245 
1246 void
dsl_redaction_list_rele(redaction_list_t * rl,const void * tag)1247 dsl_redaction_list_rele(redaction_list_t *rl, const void *tag)
1248 {
1249 	if (rl->rl_bonus != rl->rl_dbuf)
1250 		dmu_buf_rele(rl->rl_dbuf, tag);
1251 	dmu_buf_rele(rl->rl_bonus, tag);
1252 }
1253 
1254 int
dsl_redaction_list_hold_obj(dsl_pool_t * dp,uint64_t rlobj,const void * tag,redaction_list_t ** rlp)1255 dsl_redaction_list_hold_obj(dsl_pool_t *dp, uint64_t rlobj, const void *tag,
1256     redaction_list_t **rlp)
1257 {
1258 	objset_t *mos = dp->dp_meta_objset;
1259 	dmu_buf_t *dbuf, *spill_dbuf;
1260 	redaction_list_t *rl;
1261 	int err;
1262 
1263 	ASSERT(dsl_pool_config_held(dp));
1264 
1265 	err = dmu_bonus_hold(mos, rlobj, tag, &dbuf);
1266 	if (err != 0)
1267 		return (err);
1268 
1269 	rl = dmu_buf_get_user(dbuf);
1270 	if (rl == NULL) {
1271 		redaction_list_t *winner = NULL;
1272 
1273 		rl = kmem_zalloc(sizeof (redaction_list_t), KM_SLEEP);
1274 		rl->rl_bonus = dbuf;
1275 		if (dmu_spill_hold_existing(dbuf, tag, &spill_dbuf) == 0) {
1276 			rl->rl_dbuf = spill_dbuf;
1277 		} else {
1278 			rl->rl_dbuf = dbuf;
1279 		}
1280 		rl->rl_object = rlobj;
1281 		rl->rl_phys = rl->rl_dbuf->db_data;
1282 		rl->rl_mos = dp->dp_meta_objset;
1283 		zfs_refcount_create(&rl->rl_longholds);
1284 		dmu_buf_init_user(&rl->rl_dbu, redaction_list_evict_sync, NULL,
1285 		    &rl->rl_bonus);
1286 		if ((winner = dmu_buf_set_user_ie(dbuf, &rl->rl_dbu)) != NULL) {
1287 			kmem_free(rl, sizeof (*rl));
1288 			rl = winner;
1289 		}
1290 	}
1291 	*rlp = rl;
1292 	return (0);
1293 }
1294 
1295 /*
1296  * Snapshot ds is being destroyed.
1297  *
1298  * Adjust the "freed_before_next" of any bookmarks between this snap
1299  * and the previous snapshot, because their "next snapshot" is changing.
1300  *
1301  * If there are any bookmarks with HAS_FBN at this snapshot, remove
1302  * their HAS_SNAP flag (note: there can be at most one snapshot of
1303  * each filesystem at a given txg), and return B_TRUE.  In this case
1304  * the caller can not remove the key in the deadlist at this TXG, because
1305  * the HAS_FBN bookmarks require the key be there.
1306  *
1307  * Returns B_FALSE if there are no bookmarks with HAS_FBN at this
1308  * snapshot's TXG.  In this case the caller can remove the key in the
1309  * deadlist at this TXG.
1310  */
1311 boolean_t
dsl_bookmark_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)1312 dsl_bookmark_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
1313 {
1314 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1315 
1316 	dsl_dataset_t *head, *next;
1317 	VERIFY0(dsl_dataset_hold_obj(dp,
1318 	    dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj, FTAG, &head));
1319 	VERIFY0(dsl_dataset_hold_obj(dp,
1320 	    dsl_dataset_phys(ds)->ds_next_snap_obj, FTAG, &next));
1321 
1322 	/*
1323 	 * Find the first bookmark that HAS_FBN at or after the
1324 	 * previous snapshot.
1325 	 */
1326 	dsl_bookmark_node_t search = { 0 };
1327 	avl_index_t idx;
1328 	search.dbn_phys.zbm_creation_txg =
1329 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1330 	search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1331 	/*
1332 	 * The empty-string name can't be in the AVL, and it compares
1333 	 * before any entries with this TXG.
1334 	 */
1335 	search.dbn_name = (char *)"";
1336 	VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1337 	dsl_bookmark_node_t *dbn =
1338 	    avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1339 
1340 	/*
1341 	 * Iterate over all bookmarks that are at or after the previous
1342 	 * snapshot, and before this (being deleted) snapshot.  Adjust
1343 	 * their FBN based on their new next snapshot.
1344 	 */
1345 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg <
1346 	    dsl_dataset_phys(ds)->ds_creation_txg;
1347 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1348 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN))
1349 			continue;
1350 		/*
1351 		 * Increase our FBN by the amount of space that was live
1352 		 * (referenced) at the time of this bookmark (i.e.
1353 		 * birth <= zbm_creation_txg), and killed between this
1354 		 * (being deleted) snapshot and the next snapshot (i.e.
1355 		 * on the next snapshot's deadlist).  (Space killed before
1356 		 * this are already on our FBN.)
1357 		 */
1358 		uint64_t referenced, compressed, uncompressed;
1359 		dsl_deadlist_space_range(&next->ds_deadlist,
1360 		    0, dbn->dbn_phys.zbm_creation_txg,
1361 		    &referenced, &compressed, &uncompressed);
1362 		dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1363 		    referenced;
1364 		dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1365 		    compressed;
1366 		dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1367 		    uncompressed;
1368 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1369 		    dbn->dbn_name, sizeof (uint64_t),
1370 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1371 		    &dbn->dbn_phys, tx));
1372 	}
1373 	dsl_dataset_rele(next, FTAG);
1374 
1375 	/*
1376 	 * There may be several bookmarks at this txg (the TXG of the
1377 	 * snapshot being deleted).  We need to clear the SNAPSHOT_EXISTS
1378 	 * flag on all of them, and return TRUE if there is at least 1
1379 	 * bookmark here with HAS_FBN (thus preventing the deadlist
1380 	 * key from being removed).
1381 	 */
1382 	boolean_t rv = B_FALSE;
1383 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1384 	    dsl_dataset_phys(ds)->ds_creation_txg;
1385 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1386 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1387 			ASSERT(!(dbn->dbn_phys.zbm_flags &
1388 			    ZBM_FLAG_SNAPSHOT_EXISTS));
1389 			continue;
1390 		}
1391 		ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS);
1392 		dbn->dbn_phys.zbm_flags &= ~ZBM_FLAG_SNAPSHOT_EXISTS;
1393 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1394 		    dbn->dbn_name, sizeof (uint64_t),
1395 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1396 		    &dbn->dbn_phys, tx));
1397 		rv = B_TRUE;
1398 	}
1399 	dsl_dataset_rele(head, FTAG);
1400 	return (rv);
1401 }
1402 
1403 /*
1404  * A snapshot is being created of this (head) dataset.
1405  *
1406  * We don't keep keys in the deadlist for the most recent snapshot, or any
1407  * bookmarks at or after it, because there can't be any blocks on the
1408  * deadlist in this range.  Now that the most recent snapshot is after
1409  * all bookmarks, we need to add these keys.  Note that the caller always
1410  * adds a key at the previous snapshot, so we only add keys for bookmarks
1411  * after that.
1412  */
1413 void
dsl_bookmark_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)1414 dsl_bookmark_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
1415 {
1416 	uint64_t last_key_added = UINT64_MAX;
1417 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1418 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >
1419 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1420 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1421 		uint64_t creation_txg = dbn->dbn_phys.zbm_creation_txg;
1422 		ASSERT3U(creation_txg, <=, last_key_added);
1423 		/*
1424 		 * Note, there may be multiple bookmarks at this TXG,
1425 		 * and we only want to add the key for this TXG once.
1426 		 * The ds_bookmarks AVL is sorted by TXG, so we will visit
1427 		 * these bookmarks in sequence.
1428 		 */
1429 		if ((dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) &&
1430 		    creation_txg != last_key_added) {
1431 			dsl_deadlist_add_key(&ds->ds_deadlist,
1432 			    creation_txg, tx);
1433 			last_key_added = creation_txg;
1434 		}
1435 	}
1436 }
1437 
1438 /*
1439  * The next snapshot of the origin dataset has changed, due to
1440  * promote or clone swap.  If there are any bookmarks at this dataset,
1441  * we need to update their zbm_*_freed_before_next_snap to reflect this.
1442  * The head dataset has the relevant bookmarks in ds_bookmarks.
1443  */
1444 void
dsl_bookmark_next_changed(dsl_dataset_t * head,dsl_dataset_t * origin,dmu_tx_t * tx)1445 dsl_bookmark_next_changed(dsl_dataset_t *head, dsl_dataset_t *origin,
1446     dmu_tx_t *tx)
1447 {
1448 	dsl_pool_t *dp = dmu_tx_pool(tx);
1449 
1450 	/*
1451 	 * Find the first bookmark that HAS_FBN at the origin snapshot.
1452 	 */
1453 	dsl_bookmark_node_t search = { 0 };
1454 	avl_index_t idx;
1455 	search.dbn_phys.zbm_creation_txg =
1456 	    dsl_dataset_phys(origin)->ds_creation_txg;
1457 	search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1458 	/*
1459 	 * The empty-string name can't be in the AVL, and it compares
1460 	 * before any entries with this TXG.
1461 	 */
1462 	search.dbn_name = (char *)"";
1463 	VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1464 	dsl_bookmark_node_t *dbn =
1465 	    avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1466 
1467 	/*
1468 	 * Iterate over all bookmarks that are at the origin txg.
1469 	 * Adjust their FBN based on their new next snapshot.
1470 	 */
1471 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1472 	    dsl_dataset_phys(origin)->ds_creation_txg &&
1473 	    (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1474 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1475 
1476 		/*
1477 		 * Bookmark is at the origin, therefore its
1478 		 * "next dataset" is changing, so we need
1479 		 * to reset its FBN by recomputing it in
1480 		 * dsl_bookmark_set_phys().
1481 		 */
1482 		ASSERT3U(dbn->dbn_phys.zbm_guid, ==,
1483 		    dsl_dataset_phys(origin)->ds_guid);
1484 		ASSERT3U(dbn->dbn_phys.zbm_referenced_bytes_refd, ==,
1485 		    dsl_dataset_phys(origin)->ds_referenced_bytes);
1486 		ASSERT(dbn->dbn_phys.zbm_flags &
1487 		    ZBM_FLAG_SNAPSHOT_EXISTS);
1488 		/*
1489 		 * Save and restore the zbm_redaction_obj, which
1490 		 * is zeroed by dsl_bookmark_set_phys().
1491 		 */
1492 		uint64_t redaction_obj =
1493 		    dbn->dbn_phys.zbm_redaction_obj;
1494 		dsl_bookmark_set_phys(&dbn->dbn_phys, origin);
1495 		dbn->dbn_phys.zbm_redaction_obj = redaction_obj;
1496 
1497 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1498 		    dbn->dbn_name, sizeof (uint64_t),
1499 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1500 		    &dbn->dbn_phys, tx));
1501 	}
1502 }
1503 
1504 /*
1505  * This block is no longer referenced by this (head) dataset.
1506  *
1507  * Adjust the FBN of any bookmarks that reference this block, whose "next"
1508  * is the head dataset.
1509  */
1510 void
dsl_bookmark_block_killed(dsl_dataset_t * ds,const blkptr_t * bp,dmu_tx_t * tx)1511 dsl_bookmark_block_killed(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
1512 {
1513 	(void) tx;
1514 
1515 	/*
1516 	 * Iterate over bookmarks whose "next" is the head dataset.
1517 	 */
1518 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1519 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1520 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1521 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1522 		/*
1523 		 * If the block was live (referenced) at the time of this
1524 		 * bookmark, add its space to the bookmark's FBN.
1525 		 */
1526 		if (BP_GET_LOGICAL_BIRTH(bp) <=
1527 		    dbn->dbn_phys.zbm_creation_txg &&
1528 		    (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1529 			mutex_enter(&dbn->dbn_lock);
1530 			dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1531 			    bp_get_dsize_sync(dsl_dataset_get_spa(ds), bp);
1532 			dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1533 			    BP_GET_PSIZE(bp);
1534 			dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1535 			    BP_GET_UCSIZE(bp);
1536 			/*
1537 			 * Changing the ZAP object here would be too
1538 			 * expensive.  Also, we may be called from the zio
1539 			 * interrupt thread, which can't block on i/o.
1540 			 * Therefore, we mark this bookmark as dirty and
1541 			 * modify the ZAP once per txg, in
1542 			 * dsl_bookmark_sync_done().
1543 			 */
1544 			dbn->dbn_dirty = B_TRUE;
1545 			mutex_exit(&dbn->dbn_lock);
1546 		}
1547 	}
1548 }
1549 
1550 void
dsl_bookmark_sync_done(dsl_dataset_t * ds,dmu_tx_t * tx)1551 dsl_bookmark_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
1552 {
1553 	dsl_pool_t *dp = dmu_tx_pool(tx);
1554 
1555 	if (dsl_dataset_is_snapshot(ds))
1556 		return;
1557 
1558 	/*
1559 	 * We only dirty bookmarks that are at or after the most recent
1560 	 * snapshot.  We can't create snapshots between
1561 	 * dsl_bookmark_block_killed() and dsl_bookmark_sync_done(), so we
1562 	 * don't need to look at any bookmarks before ds_prev_snap_txg.
1563 	 */
1564 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1565 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1566 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1567 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1568 		if (dbn->dbn_dirty) {
1569 			/*
1570 			 * We only dirty nodes with HAS_FBN, therefore
1571 			 * we can always use the current bookmark struct size.
1572 			 */
1573 			ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1574 			VERIFY0(zap_update(dp->dp_meta_objset,
1575 			    ds->ds_bookmarks_obj,
1576 			    dbn->dbn_name, sizeof (uint64_t),
1577 			    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1578 			    &dbn->dbn_phys, tx));
1579 			dbn->dbn_dirty = B_FALSE;
1580 		}
1581 	}
1582 #ifdef ZFS_DEBUG
1583 	for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
1584 	    dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
1585 		ASSERT(!dbn->dbn_dirty);
1586 	}
1587 #endif
1588 }
1589 
1590 /*
1591  * Return the TXG of the most recent bookmark (or 0 if there are no bookmarks).
1592  */
1593 uint64_t
dsl_bookmark_latest_txg(dsl_dataset_t * ds)1594 dsl_bookmark_latest_txg(dsl_dataset_t *ds)
1595 {
1596 	ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool));
1597 	dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1598 	if (dbn == NULL)
1599 		return (0);
1600 	return (dbn->dbn_phys.zbm_creation_txg);
1601 }
1602 
1603 /*
1604  * Compare the redact_block_phys_t to the bookmark. If the last block in the
1605  * redact_block_phys_t is before the bookmark, return -1.  If the first block in
1606  * the redact_block_phys_t is after the bookmark, return 1.  Otherwise, the
1607  * bookmark is inside the range of the redact_block_phys_t, and we return 0.
1608  */
1609 static int
redact_block_zb_compare(redact_block_phys_t * first,zbookmark_phys_t * second)1610 redact_block_zb_compare(redact_block_phys_t *first,
1611     zbookmark_phys_t *second)
1612 {
1613 	/*
1614 	 * If the block_phys is for a previous object, or the last block in the
1615 	 * block_phys is strictly before the block in the bookmark, the
1616 	 * block_phys is earlier.
1617 	 */
1618 	if (first->rbp_object < second->zb_object ||
1619 	    (first->rbp_object == second->zb_object &&
1620 	    first->rbp_blkid + (redact_block_get_count(first) - 1) <
1621 	    second->zb_blkid)) {
1622 		return (-1);
1623 	}
1624 
1625 	/*
1626 	 * If the bookmark is for a previous object, or the block in the
1627 	 * bookmark is strictly before the first block in the block_phys, the
1628 	 * bookmark is earlier.
1629 	 */
1630 	if (first->rbp_object > second->zb_object ||
1631 	    (first->rbp_object == second->zb_object &&
1632 	    first->rbp_blkid > second->zb_blkid)) {
1633 		return (1);
1634 	}
1635 
1636 	return (0);
1637 }
1638 
1639 /*
1640  * Traverse the redaction list in the provided object, and call the callback for
1641  * each entry we find. Don't call the callback for any records before resume.
1642  */
1643 int
dsl_redaction_list_traverse(redaction_list_t * rl,zbookmark_phys_t * resume,rl_traverse_callback_t cb,void * arg)1644 dsl_redaction_list_traverse(redaction_list_t *rl, zbookmark_phys_t *resume,
1645     rl_traverse_callback_t cb, void *arg)
1646 {
1647 	objset_t *mos = rl->rl_mos;
1648 	int err = 0;
1649 
1650 	if (rl->rl_phys->rlp_last_object != UINT64_MAX ||
1651 	    rl->rl_phys->rlp_last_blkid != UINT64_MAX) {
1652 		/*
1653 		 * When we finish a send, we update the last object and offset
1654 		 * to UINT64_MAX.  If a send fails partway through, the last
1655 		 * object and offset will have some other value, indicating how
1656 		 * far the send got. The redaction list must be complete before
1657 		 * it can be traversed, so return EINVAL if the last object and
1658 		 * blkid are not set to UINT64_MAX.
1659 		 */
1660 		return (SET_ERROR(EINVAL));
1661 	}
1662 
1663 	/*
1664 	 * This allows us to skip the binary search and resume checking logic
1665 	 * below, if we're not resuming a redacted send.
1666 	 */
1667 	if (ZB_IS_ZERO(resume))
1668 		resume = NULL;
1669 
1670 	/*
1671 	 * Binary search for the point to resume from.
1672 	 */
1673 	uint64_t maxidx = rl->rl_phys->rlp_num_entries - 1;
1674 	uint64_t minidx = 0;
1675 	while (resume != NULL && maxidx > minidx) {
1676 		redact_block_phys_t rbp = { 0 };
1677 		ASSERT3U(maxidx, >, minidx);
1678 		uint64_t mididx = minidx + ((maxidx - minidx) / 2);
1679 		err = dmu_read(mos, rl->rl_object, mididx * sizeof (rbp),
1680 		    sizeof (rbp), &rbp, DMU_READ_NO_PREFETCH);
1681 		if (err != 0)
1682 			break;
1683 
1684 		int cmp = redact_block_zb_compare(&rbp, resume);
1685 
1686 		if (cmp == 0) {
1687 			minidx = mididx;
1688 			break;
1689 		} else if (cmp > 0) {
1690 			maxidx =
1691 			    (mididx == minidx ? minidx : mididx - 1);
1692 		} else {
1693 			minidx = mididx + 1;
1694 		}
1695 	}
1696 
1697 	unsigned int bufsize = SPA_OLD_MAXBLOCKSIZE;
1698 	redact_block_phys_t *buf = zio_data_buf_alloc(bufsize);
1699 
1700 	unsigned int entries_per_buf = bufsize / sizeof (redact_block_phys_t);
1701 	uint64_t start_block = minidx / entries_per_buf;
1702 	err = dmu_read(mos, rl->rl_object, start_block * bufsize, bufsize, buf,
1703 	    DMU_READ_PREFETCH);
1704 
1705 	for (uint64_t curidx = minidx;
1706 	    err == 0 && curidx < rl->rl_phys->rlp_num_entries;
1707 	    curidx++) {
1708 		/*
1709 		 * We read in the redaction list one block at a time.  Once we
1710 		 * finish with all the entries in a given block, we read in a
1711 		 * new one.  The predictive prefetcher will take care of any
1712 		 * prefetching, and this code shouldn't be the bottleneck, so we
1713 		 * don't need to do manual prefetching.
1714 		 */
1715 		if (curidx % entries_per_buf == 0) {
1716 			err = dmu_read(mos, rl->rl_object, curidx *
1717 			    sizeof (*buf), bufsize, buf,
1718 			    DMU_READ_PREFETCH);
1719 			if (err != 0)
1720 				break;
1721 		}
1722 		redact_block_phys_t *rb = &buf[curidx % entries_per_buf];
1723 		/*
1724 		 * If resume is non-null, we should either not send the data, or
1725 		 * null out resume so we don't have to keep doing these
1726 		 * comparisons.
1727 		 */
1728 		if (resume != NULL) {
1729 			/*
1730 			 * It is possible that after the binary search we got
1731 			 * a record before the resume point. There's two cases
1732 			 * where this can occur. If the record is the last
1733 			 * redaction record, and the resume point is after the
1734 			 * end of the redacted data, curidx will be the last
1735 			 * redaction record. In that case, the loop will end
1736 			 * after this iteration. The second case is if the
1737 			 * resume point is between two redaction records, the
1738 			 * binary search can return either the record before
1739 			 * or after the resume point. In that case, the next
1740 			 * iteration will be greater than the resume point.
1741 			 */
1742 			if (redact_block_zb_compare(rb, resume) < 0) {
1743 				ASSERT3U(curidx, ==, minidx);
1744 				continue;
1745 			} else {
1746 				/*
1747 				 * If the place to resume is in the middle of
1748 				 * the range described by this
1749 				 * redact_block_phys, then modify the
1750 				 * redact_block_phys in memory so we generate
1751 				 * the right records.
1752 				 */
1753 				if (resume->zb_object == rb->rbp_object &&
1754 				    resume->zb_blkid > rb->rbp_blkid) {
1755 					uint64_t diff = resume->zb_blkid -
1756 					    rb->rbp_blkid;
1757 					rb->rbp_blkid = resume->zb_blkid;
1758 					redact_block_set_count(rb,
1759 					    redact_block_get_count(rb) - diff);
1760 				}
1761 				resume = NULL;
1762 			}
1763 		}
1764 
1765 		if (cb(rb, arg) != 0) {
1766 			err = EINTR;
1767 			break;
1768 		}
1769 	}
1770 
1771 	zio_data_buf_free(buf, bufsize);
1772 	return (err);
1773 }
1774