xref: /freebsd/sys/contrib/openzfs/module/zfs/dsl_bookmark.c (revision 7a7741af18d6c8a804cc643cb7ecda9d730c6aa6)
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 #include <sys/dbuf.h>
38 
39 static int
dsl_bookmark_hold_ds(dsl_pool_t * dp,const char * fullname,dsl_dataset_t ** dsp,const void * tag,char ** shortnamep)40 dsl_bookmark_hold_ds(dsl_pool_t *dp, const char *fullname,
41     dsl_dataset_t **dsp, const void *tag, char **shortnamep)
42 {
43 	char buf[ZFS_MAX_DATASET_NAME_LEN];
44 	char *hashp;
45 
46 	if (strlen(fullname) >= ZFS_MAX_DATASET_NAME_LEN)
47 		return (SET_ERROR(ENAMETOOLONG));
48 	hashp = strchr(fullname, '#');
49 	if (hashp == NULL)
50 		return (SET_ERROR(EINVAL));
51 
52 	*shortnamep = hashp + 1;
53 	if (zfs_component_namecheck(*shortnamep, NULL, NULL))
54 		return (SET_ERROR(EINVAL));
55 	(void) strlcpy(buf, fullname, hashp - fullname + 1);
56 	return (dsl_dataset_hold(dp, buf, tag, dsp));
57 }
58 
59 /*
60  * When reading BOOKMARK_V1 bookmarks, the BOOKMARK_V2 fields are guaranteed
61  * to be zeroed.
62  *
63  * Returns ESRCH if bookmark is not found.
64  * Note, we need to use the ZAP rather than the AVL to look up bookmarks
65  * by name, because only the ZAP honors the casesensitivity setting.
66  */
67 int
dsl_bookmark_lookup_impl(dsl_dataset_t * ds,const char * shortname,zfs_bookmark_phys_t * bmark_phys)68 dsl_bookmark_lookup_impl(dsl_dataset_t *ds, const char *shortname,
69     zfs_bookmark_phys_t *bmark_phys)
70 {
71 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
72 	uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
73 	matchtype_t mt = 0;
74 	int err;
75 
76 	if (bmark_zapobj == 0)
77 		return (SET_ERROR(ESRCH));
78 
79 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
80 		mt = MT_NORMALIZE;
81 
82 	/*
83 	 * Zero out the bookmark in case the one stored on disk
84 	 * is in an older, shorter format.
85 	 */
86 	memset(bmark_phys, 0, sizeof (*bmark_phys));
87 
88 	err = zap_lookup_norm(mos, bmark_zapobj, shortname, sizeof (uint64_t),
89 	    sizeof (*bmark_phys) / sizeof (uint64_t), bmark_phys, mt, NULL, 0,
90 	    NULL);
91 
92 	return (err == ENOENT ? SET_ERROR(ESRCH) : err);
93 }
94 
95 /*
96  * If later_ds is non-NULL, this will return EXDEV if the specified bookmark
97  * does not represents an earlier point in later_ds's timeline.  However,
98  * bmp will still be filled in if we return EXDEV.
99  *
100  * Returns ENOENT if the dataset containing the bookmark does not exist.
101  * Returns ESRCH if the dataset exists but the bookmark was not found in it.
102  */
103 int
dsl_bookmark_lookup(dsl_pool_t * dp,const char * fullname,dsl_dataset_t * later_ds,zfs_bookmark_phys_t * bmp)104 dsl_bookmark_lookup(dsl_pool_t *dp, const char *fullname,
105     dsl_dataset_t *later_ds, zfs_bookmark_phys_t *bmp)
106 {
107 	char *shortname;
108 	dsl_dataset_t *ds;
109 	int error;
110 
111 	error = dsl_bookmark_hold_ds(dp, fullname, &ds, FTAG, &shortname);
112 	if (error != 0)
113 		return (error);
114 
115 	error = dsl_bookmark_lookup_impl(ds, shortname, bmp);
116 	if (error == 0 && later_ds != NULL) {
117 		if (!dsl_dataset_is_before(later_ds, ds, bmp->zbm_creation_txg))
118 			error = SET_ERROR(EXDEV);
119 	}
120 	dsl_dataset_rele(ds, FTAG);
121 	return (error);
122 }
123 
124 /*
125  * Validates that
126  * - bmark is a full dataset path of a bookmark (bookmark_namecheck)
127  * - source is a full path of a snapshot or bookmark
128  *   ({bookmark,snapshot}_namecheck)
129  *
130  * Returns 0 if valid, -1 otherwise.
131  */
132 static int
dsl_bookmark_create_nvl_validate_pair(const char * bmark,const char * source)133 dsl_bookmark_create_nvl_validate_pair(const char *bmark, const char *source)
134 {
135 	if (bookmark_namecheck(bmark, NULL, NULL) != 0)
136 		return (-1);
137 
138 	int is_bmark, is_snap;
139 	is_bmark = bookmark_namecheck(source, NULL, NULL) == 0;
140 	is_snap = snapshot_namecheck(source, NULL, NULL) == 0;
141 	if (!is_bmark && !is_snap)
142 		return (-1);
143 
144 	return (0);
145 }
146 
147 /*
148  * Check that the given nvlist corresponds to the following schema:
149  *  { newbookmark -> source, ... }
150  * where
151  * - each pair passes dsl_bookmark_create_nvl_validate_pair
152  * - all newbookmarks are in the same pool
153  * - all newbookmarks have unique names
154  *
155  * Note that this function is only validates above schema. Callers must ensure
156  * that the bookmarks can be created, e.g. that sources exist.
157  *
158  * Returns 0 if the nvlist adheres to above schema.
159  * Returns -1 if it doesn't.
160  */
161 int
dsl_bookmark_create_nvl_validate(nvlist_t * bmarks)162 dsl_bookmark_create_nvl_validate(nvlist_t *bmarks)
163 {
164 	const char *first = NULL;
165 	size_t first_len = 0;
166 
167 	for (nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
168 	    pair != NULL; pair = nvlist_next_nvpair(bmarks, pair)) {
169 
170 		const char *bmark = nvpair_name(pair);
171 		const 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 			const 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 		break;
234 	case 0:
235 		error = SET_ERROR(EEXIST);
236 		goto eholdnewbmds;
237 	default:
238 		/* dsl_bookmark_lookup_impl already did SET_ERROR */
239 		goto eholdnewbmds;
240 	}
241 
242 	/* error is retval of the following if-cascade */
243 	if (strchr(source, '@') != NULL) {
244 		dsl_dataset_t *source_snap_ds;
245 		ASSERT3S(snapshot_namecheck(source, NULL, NULL), ==, 0);
246 		error = dsl_dataset_hold(dp, source, FTAG, &source_snap_ds);
247 		if (error == 0) {
248 			VERIFY(source_snap_ds->ds_is_snapshot);
249 			/*
250 			 * Verify that source snapshot is an earlier point in
251 			 * newbm_ds's timeline (source may be newbm_ds's origin)
252 			 */
253 			if (!dsl_dataset_is_before(newbm_ds, source_snap_ds, 0))
254 				error = SET_ERROR(
255 				    ZFS_ERR_BOOKMARK_SOURCE_NOT_ANCESTOR);
256 			dsl_dataset_rele(source_snap_ds, FTAG);
257 		}
258 	} else if (strchr(source, '#') != NULL) {
259 		zfs_bookmark_phys_t source_phys;
260 		ASSERT3S(bookmark_namecheck(source, NULL, NULL), ==, 0);
261 		/*
262 		 * Source must exists and be an earlier point in newbm_ds's
263 		 * timeline (newbm_ds's origin may be a snap of source's ds)
264 		 */
265 		error = dsl_bookmark_lookup(dp, source, newbm_ds, &source_phys);
266 		switch (error) {
267 		case 0:
268 			break; /* happy path */
269 		case EXDEV:
270 			error = SET_ERROR(ZFS_ERR_BOOKMARK_SOURCE_NOT_ANCESTOR);
271 			break;
272 		default:
273 			/* dsl_bookmark_lookup already did SET_ERROR */
274 			break;
275 		}
276 	} else {
277 		/*
278 		 * dsl_bookmark_create_nvl_validate validates that source is
279 		 * either snapshot or bookmark
280 		 */
281 		panic("unreachable code: %s", source);
282 	}
283 
284 eholdnewbmds:
285 	dsl_dataset_rele(newbm_ds, FTAG);
286 	return (error);
287 }
288 
289 int
dsl_bookmark_create_check(void * arg,dmu_tx_t * tx)290 dsl_bookmark_create_check(void *arg, dmu_tx_t *tx)
291 {
292 	dsl_bookmark_create_arg_t *dbca = arg;
293 	int rv = 0;
294 	int schema_err = 0;
295 	ASSERT3P(dbca, !=, NULL);
296 	ASSERT3P(dbca->dbca_bmarks, !=, NULL);
297 	/* dbca->dbca_errors is allowed to be NULL */
298 
299 	dsl_pool_t *dp = dmu_tx_pool(tx);
300 
301 	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
302 		return (SET_ERROR(ENOTSUP));
303 
304 	if (dsl_bookmark_create_nvl_validate(dbca->dbca_bmarks) != 0)
305 		rv = schema_err = SET_ERROR(EINVAL);
306 
307 	for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
308 	    pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
309 		const char *new = nvpair_name(pair);
310 
311 		int error = schema_err;
312 		if (error == 0) {
313 			const char *source = fnvpair_value_string(pair);
314 			error = dsl_bookmark_create_check_impl(dp, new, source);
315 			if (error != 0)
316 				error = SET_ERROR(error);
317 		}
318 
319 		if (error != 0) {
320 			rv = error;
321 			if (dbca->dbca_errors != NULL)
322 				fnvlist_add_int32(dbca->dbca_errors,
323 				    new, error);
324 		}
325 	}
326 
327 	return (rv);
328 }
329 
330 static dsl_bookmark_node_t *
dsl_bookmark_node_alloc(char * shortname)331 dsl_bookmark_node_alloc(char *shortname)
332 {
333 	dsl_bookmark_node_t *dbn = kmem_alloc(sizeof (*dbn), KM_SLEEP);
334 	dbn->dbn_name = spa_strdup(shortname);
335 	dbn->dbn_dirty = B_FALSE;
336 	mutex_init(&dbn->dbn_lock, NULL, MUTEX_DEFAULT, NULL);
337 	return (dbn);
338 }
339 
340 /*
341  * Set the fields in the zfs_bookmark_phys_t based on the specified snapshot.
342  */
343 static void
dsl_bookmark_set_phys(zfs_bookmark_phys_t * zbm,dsl_dataset_t * snap)344 dsl_bookmark_set_phys(zfs_bookmark_phys_t *zbm, dsl_dataset_t *snap)
345 {
346 	spa_t *spa = dsl_dataset_get_spa(snap);
347 	objset_t *mos = spa_get_dsl(spa)->dp_meta_objset;
348 	dsl_dataset_phys_t *dsp = dsl_dataset_phys(snap);
349 
350 	memset(zbm, 0, sizeof (zfs_bookmark_phys_t));
351 	zbm->zbm_guid = dsp->ds_guid;
352 	zbm->zbm_creation_txg = dsp->ds_creation_txg;
353 	zbm->zbm_creation_time = dsp->ds_creation_time;
354 	zbm->zbm_redaction_obj = 0;
355 
356 	/*
357 	 * If the dataset is encrypted create a larger bookmark to
358 	 * accommodate the IVset guid. The IVset guid was added
359 	 * after the encryption feature to prevent a problem with
360 	 * raw sends. If we encounter an encrypted dataset without
361 	 * an IVset guid we fall back to a normal bookmark.
362 	 */
363 	if (snap->ds_dir->dd_crypto_obj != 0 &&
364 	    spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
365 		(void) zap_lookup(mos, snap->ds_object,
366 		    DS_FIELD_IVSET_GUID, sizeof (uint64_t), 1,
367 		    &zbm->zbm_ivset_guid);
368 	}
369 
370 	if (spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_WRITTEN)) {
371 		zbm->zbm_flags = ZBM_FLAG_SNAPSHOT_EXISTS | ZBM_FLAG_HAS_FBN;
372 		zbm->zbm_referenced_bytes_refd = dsp->ds_referenced_bytes;
373 		zbm->zbm_compressed_bytes_refd = dsp->ds_compressed_bytes;
374 		zbm->zbm_uncompressed_bytes_refd = dsp->ds_uncompressed_bytes;
375 
376 		dsl_dataset_t *nextds;
377 		VERIFY0(dsl_dataset_hold_obj(snap->ds_dir->dd_pool,
378 		    dsp->ds_next_snap_obj, FTAG, &nextds));
379 		dsl_deadlist_space(&nextds->ds_deadlist,
380 		    &zbm->zbm_referenced_freed_before_next_snap,
381 		    &zbm->zbm_compressed_freed_before_next_snap,
382 		    &zbm->zbm_uncompressed_freed_before_next_snap);
383 		dsl_dataset_rele(nextds, FTAG);
384 	}
385 }
386 
387 /*
388  * Add dsl_bookmark_node_t `dbn` to the given dataset and increment appropriate
389  * SPA feature counters.
390  */
391 void
dsl_bookmark_node_add(dsl_dataset_t * hds,dsl_bookmark_node_t * dbn,dmu_tx_t * tx)392 dsl_bookmark_node_add(dsl_dataset_t *hds, dsl_bookmark_node_t *dbn,
393     dmu_tx_t *tx)
394 {
395 	dsl_pool_t *dp = dmu_tx_pool(tx);
396 	objset_t *mos = dp->dp_meta_objset;
397 
398 	if (hds->ds_bookmarks_obj == 0) {
399 		hds->ds_bookmarks_obj = zap_create_norm(mos,
400 		    U8_TEXTPREP_TOUPPER, DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0,
401 		    tx);
402 		spa_feature_incr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
403 
404 		dsl_dataset_zapify(hds, tx);
405 		VERIFY0(zap_add(mos, hds->ds_object,
406 		    DS_FIELD_BOOKMARK_NAMES,
407 		    sizeof (hds->ds_bookmarks_obj), 1,
408 		    &hds->ds_bookmarks_obj, tx));
409 	}
410 
411 	avl_add(&hds->ds_bookmarks, dbn);
412 
413 	/*
414 	 * To maintain backwards compatibility with software that doesn't
415 	 * understand SPA_FEATURE_BOOKMARK_V2, we need to use the smallest
416 	 * possible bookmark size.
417 	 */
418 	uint64_t bookmark_phys_size = BOOKMARK_PHYS_SIZE_V1;
419 	if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARK_V2) &&
420 	    (dbn->dbn_phys.zbm_ivset_guid != 0 || dbn->dbn_phys.zbm_flags &
421 	    ZBM_FLAG_HAS_FBN || dbn->dbn_phys.zbm_redaction_obj != 0)) {
422 		bookmark_phys_size = BOOKMARK_PHYS_SIZE_V2;
423 		spa_feature_incr(dp->dp_spa, SPA_FEATURE_BOOKMARK_V2, tx);
424 	}
425 
426 	zfs_bookmark_phys_t zero_phys = { 0 };
427 	ASSERT0(memcmp(((char *)&dbn->dbn_phys) + bookmark_phys_size,
428 	    &zero_phys, sizeof (zfs_bookmark_phys_t) - bookmark_phys_size));
429 
430 	VERIFY0(zap_add(mos, hds->ds_bookmarks_obj, dbn->dbn_name,
431 	    sizeof (uint64_t), bookmark_phys_size / sizeof (uint64_t),
432 	    &dbn->dbn_phys, tx));
433 }
434 
435 /*
436  * If redaction_list is non-null, we create a redacted bookmark and redaction
437  * list, and store the object number of the redaction list in redact_obj.
438  */
439 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)440 dsl_bookmark_create_sync_impl_snap(const char *bookmark, const char *snapshot,
441     dmu_tx_t *tx, uint64_t num_redact_snaps, uint64_t *redact_snaps,
442     const void *tag, redaction_list_t **redaction_list)
443 {
444 	dsl_pool_t *dp = dmu_tx_pool(tx);
445 	objset_t *mos = dp->dp_meta_objset;
446 	dsl_dataset_t *snapds, *bmark_fs;
447 	char *shortname;
448 	boolean_t bookmark_redacted;
449 	uint64_t *dsredactsnaps;
450 	uint64_t dsnumsnaps;
451 
452 	VERIFY0(dsl_dataset_hold(dp, snapshot, FTAG, &snapds));
453 	VERIFY0(dsl_bookmark_hold_ds(dp, bookmark, &bmark_fs, FTAG,
454 	    &shortname));
455 
456 	dsl_bookmark_node_t *dbn = dsl_bookmark_node_alloc(shortname);
457 	dsl_bookmark_set_phys(&dbn->dbn_phys, snapds);
458 
459 	bookmark_redacted = dsl_dataset_get_uint64_array_feature(snapds,
460 	    SPA_FEATURE_REDACTED_DATASETS, &dsnumsnaps, &dsredactsnaps);
461 	if (redaction_list != NULL || bookmark_redacted) {
462 		redaction_list_t *local_rl;
463 		boolean_t spill = B_FALSE;
464 		if (bookmark_redacted) {
465 			redact_snaps = dsredactsnaps;
466 			num_redact_snaps = dsnumsnaps;
467 		}
468 		int bonuslen = sizeof (redaction_list_phys_t) +
469 		    num_redact_snaps * sizeof (uint64_t);
470 		if (bonuslen > dmu_bonus_max())
471 			spill = B_TRUE;
472 		dbn->dbn_phys.zbm_redaction_obj = dmu_object_alloc(mos,
473 		    DMU_OTN_UINT64_METADATA, SPA_OLD_MAXBLOCKSIZE,
474 		    DMU_OTN_UINT64_METADATA, spill ? 0 : bonuslen, tx);
475 		spa_feature_incr(dp->dp_spa,
476 		    SPA_FEATURE_REDACTION_BOOKMARKS, tx);
477 		if (spill) {
478 			spa_feature_incr(dp->dp_spa,
479 			    SPA_FEATURE_REDACTION_LIST_SPILL, tx);
480 		}
481 
482 		VERIFY0(dsl_redaction_list_hold_obj(dp,
483 		    dbn->dbn_phys.zbm_redaction_obj, tag, &local_rl));
484 		dsl_redaction_list_long_hold(dp, local_rl, tag);
485 
486 		if (!spill) {
487 			ASSERT3U(local_rl->rl_bonus->db_size, >=, bonuslen);
488 			dmu_buf_will_dirty(local_rl->rl_bonus, tx);
489 		} else {
490 			dmu_buf_t *db;
491 			VERIFY0(dmu_spill_hold_by_bonus(local_rl->rl_bonus,
492 			    DB_RF_MUST_SUCCEED, FTAG, &db));
493 			dmu_buf_will_fill(db, tx, B_FALSE);
494 			VERIFY0(dbuf_spill_set_blksz(db, P2ROUNDUP(bonuslen,
495 			    SPA_MINBLOCKSIZE), tx));
496 			local_rl->rl_phys = db->db_data;
497 			local_rl->rl_dbuf = db;
498 		}
499 		memcpy(local_rl->rl_phys->rlp_snaps, redact_snaps,
500 		    sizeof (uint64_t) * num_redact_snaps);
501 		local_rl->rl_phys->rlp_num_snaps = num_redact_snaps;
502 		if (bookmark_redacted) {
503 			ASSERT3P(redaction_list, ==, NULL);
504 			local_rl->rl_phys->rlp_last_blkid = UINT64_MAX;
505 			local_rl->rl_phys->rlp_last_object = UINT64_MAX;
506 			dsl_redaction_list_long_rele(local_rl, tag);
507 			dsl_redaction_list_rele(local_rl, tag);
508 		} else {
509 			*redaction_list = local_rl;
510 		}
511 	}
512 
513 	if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
514 		spa_feature_incr(dp->dp_spa,
515 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
516 	}
517 
518 	dsl_bookmark_node_add(bmark_fs, dbn, tx);
519 
520 	spa_history_log_internal_ds(bmark_fs, "bookmark", tx,
521 	    "name=%s creation_txg=%llu target_snap=%llu redact_obj=%llu",
522 	    shortname, (longlong_t)dbn->dbn_phys.zbm_creation_txg,
523 	    (longlong_t)snapds->ds_object,
524 	    (longlong_t)dbn->dbn_phys.zbm_redaction_obj);
525 
526 	dsl_dataset_rele(bmark_fs, FTAG);
527 	dsl_dataset_rele(snapds, FTAG);
528 }
529 
530 
531 static void
dsl_bookmark_create_sync_impl_book(const char * new_name,const char * source_name,dmu_tx_t * tx)532 dsl_bookmark_create_sync_impl_book(
533     const char *new_name, const char *source_name, dmu_tx_t *tx)
534 {
535 	dsl_pool_t *dp = dmu_tx_pool(tx);
536 	dsl_dataset_t *bmark_fs_source, *bmark_fs_new;
537 	char *source_shortname, *new_shortname;
538 	zfs_bookmark_phys_t source_phys;
539 
540 	VERIFY0(dsl_bookmark_hold_ds(dp, source_name, &bmark_fs_source, FTAG,
541 	    &source_shortname));
542 	VERIFY0(dsl_bookmark_hold_ds(dp, new_name, &bmark_fs_new, FTAG,
543 	    &new_shortname));
544 
545 	/*
546 	 * create a copy of the source bookmark by copying most of its members
547 	 *
548 	 * Caveat: bookmarking a redaction bookmark yields a normal bookmark
549 	 * -----------------------------------------------------------------
550 	 * Reasoning:
551 	 * - The zbm_redaction_obj would be referred to by both source and new
552 	 *   bookmark, but would be destroyed once either source or new is
553 	 *   destroyed, resulting in use-after-free of the referred object.
554 	 * - User expectation when issuing the `zfs bookmark` command is that
555 	 *   a normal bookmark of the source is created
556 	 *
557 	 * Design Alternatives For Full Redaction Bookmark Copying:
558 	 * - reference-count the redaction object => would require on-disk
559 	 *   format change for existing redaction objects
560 	 * - Copy the redaction object => cannot be done in syncing context
561 	 *   because the redaction object might be too large
562 	 */
563 
564 	VERIFY0(dsl_bookmark_lookup_impl(bmark_fs_source, source_shortname,
565 	    &source_phys));
566 	dsl_bookmark_node_t *new_dbn = dsl_bookmark_node_alloc(new_shortname);
567 
568 	memcpy(&new_dbn->dbn_phys, &source_phys, sizeof (source_phys));
569 	new_dbn->dbn_phys.zbm_redaction_obj = 0;
570 
571 	/* update feature counters */
572 	if (new_dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
573 		spa_feature_incr(dp->dp_spa,
574 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
575 	}
576 	/* no need for redaction bookmark counter; nulled zbm_redaction_obj */
577 	/* dsl_bookmark_node_add bumps bookmarks and v2-bookmarks counter */
578 
579 	/*
580 	 * write new bookmark
581 	 *
582 	 * Note that dsl_bookmark_lookup_impl guarantees that, if source is a
583 	 * v1 bookmark, the v2-only fields are zeroed.
584 	 * And dsl_bookmark_node_add writes back a v1-sized bookmark if
585 	 * v2 bookmarks are disabled and/or v2-only fields are zeroed.
586 	 * => bookmark copying works on pre-bookmark-v2 pools
587 	 */
588 	dsl_bookmark_node_add(bmark_fs_new, new_dbn, tx);
589 
590 	spa_history_log_internal_ds(bmark_fs_source, "bookmark", tx,
591 	    "name=%s creation_txg=%llu source_guid=%llu",
592 	    new_shortname, (longlong_t)new_dbn->dbn_phys.zbm_creation_txg,
593 	    (longlong_t)source_phys.zbm_guid);
594 
595 	dsl_dataset_rele(bmark_fs_source, FTAG);
596 	dsl_dataset_rele(bmark_fs_new, FTAG);
597 }
598 
599 void
dsl_bookmark_create_sync(void * arg,dmu_tx_t * tx)600 dsl_bookmark_create_sync(void *arg, dmu_tx_t *tx)
601 {
602 	dsl_bookmark_create_arg_t *dbca = arg;
603 
604 	ASSERT(spa_feature_is_enabled(dmu_tx_pool(tx)->dp_spa,
605 	    SPA_FEATURE_BOOKMARKS));
606 
607 	for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
608 	    pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
609 
610 		const char *new = nvpair_name(pair);
611 		const char *source = fnvpair_value_string(pair);
612 
613 		if (strchr(source, '@') != NULL) {
614 			dsl_bookmark_create_sync_impl_snap(new, source, tx,
615 			    0, NULL, NULL, NULL);
616 		} else if (strchr(source, '#') != NULL) {
617 			dsl_bookmark_create_sync_impl_book(new, source, tx);
618 		} else {
619 			panic("unreachable code");
620 		}
621 
622 	}
623 }
624 
625 /*
626  * The bookmarks must all be in the same pool.
627  */
628 int
dsl_bookmark_create(nvlist_t * bmarks,nvlist_t * errors)629 dsl_bookmark_create(nvlist_t *bmarks, nvlist_t *errors)
630 {
631 	nvpair_t *pair;
632 	dsl_bookmark_create_arg_t dbca;
633 
634 	pair = nvlist_next_nvpair(bmarks, NULL);
635 	if (pair == NULL)
636 		return (0);
637 
638 	dbca.dbca_bmarks = bmarks;
639 	dbca.dbca_errors = errors;
640 
641 	return (dsl_sync_task(nvpair_name(pair), dsl_bookmark_create_check,
642 	    dsl_bookmark_create_sync, &dbca,
643 	    fnvlist_num_pairs(bmarks), ZFS_SPACE_CHECK_NORMAL));
644 }
645 
646 static int
dsl_bookmark_create_redacted_check(void * arg,dmu_tx_t * tx)647 dsl_bookmark_create_redacted_check(void *arg, dmu_tx_t *tx)
648 {
649 	dsl_bookmark_create_redacted_arg_t *dbcra = arg;
650 	dsl_pool_t *dp = dmu_tx_pool(tx);
651 	int rv = 0;
652 
653 	if (!spa_feature_is_enabled(dp->dp_spa,
654 	    SPA_FEATURE_REDACTION_BOOKMARKS))
655 		return (SET_ERROR(ENOTSUP));
656 	/*
657 	 * If the list of redact snaps will not fit in the bonus buffer (or
658 	 * spill block, with the REDACTION_LIST_SPILL feature) with the
659 	 * furthest reached object and offset, fail.
660 	 */
661 	uint64_t snaplimit = ((spa_feature_is_enabled(dp->dp_spa,
662 	    SPA_FEATURE_REDACTION_LIST_SPILL) ? spa_maxblocksize(dp->dp_spa) :
663 	    dmu_bonus_max()) -
664 	    sizeof (redaction_list_phys_t)) / sizeof (uint64_t);
665 	if (dbcra->dbcra_numsnaps > snaplimit)
666 		return (SET_ERROR(E2BIG));
667 
668 	if (dsl_bookmark_create_nvl_validate_pair(
669 	    dbcra->dbcra_bmark, dbcra->dbcra_snap) != 0)
670 		return (SET_ERROR(EINVAL));
671 
672 	rv = dsl_bookmark_create_check_impl(dp,
673 	    dbcra->dbcra_bmark, dbcra->dbcra_snap);
674 	return (rv);
675 }
676 
677 static void
dsl_bookmark_create_redacted_sync(void * arg,dmu_tx_t * tx)678 dsl_bookmark_create_redacted_sync(void *arg, dmu_tx_t *tx)
679 {
680 	dsl_bookmark_create_redacted_arg_t *dbcra = arg;
681 	dsl_bookmark_create_sync_impl_snap(dbcra->dbcra_bmark,
682 	    dbcra->dbcra_snap, tx, dbcra->dbcra_numsnaps, dbcra->dbcra_snaps,
683 	    dbcra->dbcra_tag, dbcra->dbcra_rl);
684 }
685 
686 int
dsl_bookmark_create_redacted(const char * bookmark,const char * snapshot,uint64_t numsnaps,uint64_t * snapguids,const void * tag,redaction_list_t ** rl)687 dsl_bookmark_create_redacted(const char *bookmark, const char *snapshot,
688     uint64_t numsnaps, uint64_t *snapguids, const void *tag,
689     redaction_list_t **rl)
690 {
691 	dsl_bookmark_create_redacted_arg_t dbcra;
692 
693 	dbcra.dbcra_bmark = bookmark;
694 	dbcra.dbcra_snap = snapshot;
695 	dbcra.dbcra_rl = rl;
696 	dbcra.dbcra_numsnaps = numsnaps;
697 	dbcra.dbcra_snaps = snapguids;
698 	dbcra.dbcra_tag = tag;
699 
700 	return (dsl_sync_task(bookmark, dsl_bookmark_create_redacted_check,
701 	    dsl_bookmark_create_redacted_sync, &dbcra, 5,
702 	    ZFS_SPACE_CHECK_NORMAL));
703 }
704 
705 /*
706  * Retrieve the list of properties given in the 'props' nvlist for a bookmark.
707  * If 'props' is NULL, retrieves all properties.
708  */
709 static void
dsl_bookmark_fetch_props(dsl_pool_t * dp,zfs_bookmark_phys_t * bmark_phys,nvlist_t * props,nvlist_t * out_props)710 dsl_bookmark_fetch_props(dsl_pool_t *dp, zfs_bookmark_phys_t *bmark_phys,
711     nvlist_t *props, nvlist_t *out_props)
712 {
713 	ASSERT3P(dp, !=, NULL);
714 	ASSERT3P(bmark_phys, !=, NULL);
715 	ASSERT3P(out_props, !=, NULL);
716 	ASSERT(RRW_LOCK_HELD(&dp->dp_config_rwlock));
717 
718 	if (props == NULL || nvlist_exists(props,
719 	    zfs_prop_to_name(ZFS_PROP_GUID))) {
720 		dsl_prop_nvlist_add_uint64(out_props,
721 		    ZFS_PROP_GUID, bmark_phys->zbm_guid);
722 	}
723 	if (props == NULL || nvlist_exists(props,
724 	    zfs_prop_to_name(ZFS_PROP_CREATETXG))) {
725 		dsl_prop_nvlist_add_uint64(out_props,
726 		    ZFS_PROP_CREATETXG, bmark_phys->zbm_creation_txg);
727 	}
728 	if (props == NULL || nvlist_exists(props,
729 	    zfs_prop_to_name(ZFS_PROP_CREATION))) {
730 		dsl_prop_nvlist_add_uint64(out_props,
731 		    ZFS_PROP_CREATION, bmark_phys->zbm_creation_time);
732 	}
733 	if (props == NULL || nvlist_exists(props,
734 	    zfs_prop_to_name(ZFS_PROP_IVSET_GUID))) {
735 		dsl_prop_nvlist_add_uint64(out_props,
736 		    ZFS_PROP_IVSET_GUID, bmark_phys->zbm_ivset_guid);
737 	}
738 	if (bmark_phys->zbm_flags & ZBM_FLAG_HAS_FBN) {
739 		if (props == NULL || nvlist_exists(props,
740 		    zfs_prop_to_name(ZFS_PROP_REFERENCED))) {
741 			dsl_prop_nvlist_add_uint64(out_props,
742 			    ZFS_PROP_REFERENCED,
743 			    bmark_phys->zbm_referenced_bytes_refd);
744 		}
745 		if (props == NULL || nvlist_exists(props,
746 		    zfs_prop_to_name(ZFS_PROP_LOGICALREFERENCED))) {
747 			dsl_prop_nvlist_add_uint64(out_props,
748 			    ZFS_PROP_LOGICALREFERENCED,
749 			    bmark_phys->zbm_uncompressed_bytes_refd);
750 		}
751 		if (props == NULL || nvlist_exists(props,
752 		    zfs_prop_to_name(ZFS_PROP_REFRATIO))) {
753 			uint64_t ratio =
754 			    bmark_phys->zbm_compressed_bytes_refd == 0 ? 100 :
755 			    bmark_phys->zbm_uncompressed_bytes_refd * 100 /
756 			    bmark_phys->zbm_compressed_bytes_refd;
757 			dsl_prop_nvlist_add_uint64(out_props,
758 			    ZFS_PROP_REFRATIO, ratio);
759 		}
760 	}
761 
762 	if ((props == NULL || nvlist_exists(props, "redact_snaps") ||
763 	    nvlist_exists(props, "redact_complete")) &&
764 	    bmark_phys->zbm_redaction_obj != 0) {
765 		redaction_list_t *rl;
766 		int err = dsl_redaction_list_hold_obj(dp,
767 		    bmark_phys->zbm_redaction_obj, FTAG, &rl);
768 		if (err == 0) {
769 			if (nvlist_exists(props, "redact_snaps")) {
770 				nvlist_t *nvl;
771 				nvl = fnvlist_alloc();
772 				fnvlist_add_uint64_array(nvl, ZPROP_VALUE,
773 				    rl->rl_phys->rlp_snaps,
774 				    rl->rl_phys->rlp_num_snaps);
775 				fnvlist_add_nvlist(out_props, "redact_snaps",
776 				    nvl);
777 				nvlist_free(nvl);
778 			}
779 			if (nvlist_exists(props, "redact_complete")) {
780 				nvlist_t *nvl;
781 				nvl = fnvlist_alloc();
782 				fnvlist_add_boolean_value(nvl, ZPROP_VALUE,
783 				    rl->rl_phys->rlp_last_blkid == UINT64_MAX &&
784 				    rl->rl_phys->rlp_last_object == UINT64_MAX);
785 				fnvlist_add_nvlist(out_props, "redact_complete",
786 				    nvl);
787 				nvlist_free(nvl);
788 			}
789 			dsl_redaction_list_rele(rl, FTAG);
790 		}
791 	}
792 }
793 
794 int
dsl_get_bookmarks_impl(dsl_dataset_t * ds,nvlist_t * props,nvlist_t * outnvl)795 dsl_get_bookmarks_impl(dsl_dataset_t *ds, nvlist_t *props, nvlist_t *outnvl)
796 {
797 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
798 
799 	ASSERT(dsl_pool_config_held(dp));
800 
801 	if (dsl_dataset_is_snapshot(ds))
802 		return (SET_ERROR(EINVAL));
803 
804 	for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
805 	    dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
806 		nvlist_t *out_props = fnvlist_alloc();
807 
808 		dsl_bookmark_fetch_props(dp, &dbn->dbn_phys, props, out_props);
809 
810 		fnvlist_add_nvlist(outnvl, dbn->dbn_name, out_props);
811 		fnvlist_free(out_props);
812 	}
813 	return (0);
814 }
815 
816 /*
817  * Comparison func for ds_bookmarks AVL tree.  We sort the bookmarks by
818  * their TXG, then by their FBN-ness.  The "FBN-ness" component ensures
819  * that all bookmarks at the same TXG that HAS_FBN are adjacent, which
820  * dsl_bookmark_destroy_sync_impl() depends on.  Note that there may be
821  * multiple bookmarks at the same TXG (with the same FBN-ness).  In this
822  * case we differentiate them by an arbitrary metric (in this case,
823  * their names).
824  */
825 static int
dsl_bookmark_compare(const void * l,const void * r)826 dsl_bookmark_compare(const void *l, const void *r)
827 {
828 	const dsl_bookmark_node_t *ldbn = l;
829 	const dsl_bookmark_node_t *rdbn = r;
830 
831 	int64_t cmp = TREE_CMP(ldbn->dbn_phys.zbm_creation_txg,
832 	    rdbn->dbn_phys.zbm_creation_txg);
833 	if (likely(cmp))
834 		return (cmp);
835 	cmp = TREE_CMP((ldbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN),
836 	    (rdbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
837 	if (likely(cmp))
838 		return (cmp);
839 	cmp = strcmp(ldbn->dbn_name, rdbn->dbn_name);
840 	return (TREE_ISIGN(cmp));
841 }
842 
843 /*
844  * Cache this (head) dataset's bookmarks in the ds_bookmarks AVL tree.
845  */
846 int
dsl_bookmark_init_ds(dsl_dataset_t * ds)847 dsl_bookmark_init_ds(dsl_dataset_t *ds)
848 {
849 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
850 	objset_t *mos = dp->dp_meta_objset;
851 
852 	ASSERT(!ds->ds_is_snapshot);
853 
854 	avl_create(&ds->ds_bookmarks, dsl_bookmark_compare,
855 	    sizeof (dsl_bookmark_node_t),
856 	    offsetof(dsl_bookmark_node_t, dbn_node));
857 
858 	if (!dsl_dataset_is_zapified(ds))
859 		return (0);
860 
861 	int zaperr = zap_lookup(mos, ds->ds_object, DS_FIELD_BOOKMARK_NAMES,
862 	    sizeof (ds->ds_bookmarks_obj), 1, &ds->ds_bookmarks_obj);
863 	if (zaperr == ENOENT)
864 		return (0);
865 	if (zaperr != 0)
866 		return (zaperr);
867 
868 	if (ds->ds_bookmarks_obj == 0)
869 		return (0);
870 
871 	int err = 0;
872 	zap_cursor_t zc;
873 	zap_attribute_t *attr;
874 
875 	attr = zap_attribute_alloc();
876 	for (zap_cursor_init(&zc, mos, ds->ds_bookmarks_obj);
877 	    (err = zap_cursor_retrieve(&zc, attr)) == 0;
878 	    zap_cursor_advance(&zc)) {
879 		dsl_bookmark_node_t *dbn =
880 		    dsl_bookmark_node_alloc(attr->za_name);
881 
882 		err = dsl_bookmark_lookup_impl(ds,
883 		    dbn->dbn_name, &dbn->dbn_phys);
884 		ASSERT3U(err, !=, ENOENT);
885 		if (err != 0) {
886 			kmem_free(dbn, sizeof (*dbn));
887 			break;
888 		}
889 		avl_add(&ds->ds_bookmarks, dbn);
890 	}
891 	zap_cursor_fini(&zc);
892 	zap_attribute_free(attr);
893 	if (err == ENOENT)
894 		err = 0;
895 	return (err);
896 }
897 
898 void
dsl_bookmark_fini_ds(dsl_dataset_t * ds)899 dsl_bookmark_fini_ds(dsl_dataset_t *ds)
900 {
901 	void *cookie = NULL;
902 	dsl_bookmark_node_t *dbn;
903 
904 	if (ds->ds_is_snapshot)
905 		return;
906 
907 	while ((dbn = avl_destroy_nodes(&ds->ds_bookmarks, &cookie)) != NULL) {
908 		spa_strfree(dbn->dbn_name);
909 		mutex_destroy(&dbn->dbn_lock);
910 		kmem_free(dbn, sizeof (*dbn));
911 	}
912 	avl_destroy(&ds->ds_bookmarks);
913 }
914 
915 /*
916  * Retrieve the bookmarks that exist in the specified dataset, and the
917  * requested properties of each bookmark.
918  *
919  * The "props" nvlist specifies which properties are requested.
920  * See lzc_get_bookmarks() for the list of valid properties.
921  */
922 int
dsl_get_bookmarks(const char * dsname,nvlist_t * props,nvlist_t * outnvl)923 dsl_get_bookmarks(const char *dsname, nvlist_t *props, nvlist_t *outnvl)
924 {
925 	dsl_pool_t *dp;
926 	dsl_dataset_t *ds;
927 	int err;
928 
929 	err = dsl_pool_hold(dsname, FTAG, &dp);
930 	if (err != 0)
931 		return (err);
932 	err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
933 	if (err != 0) {
934 		dsl_pool_rele(dp, FTAG);
935 		return (err);
936 	}
937 
938 	err = dsl_get_bookmarks_impl(ds, props, outnvl);
939 
940 	dsl_dataset_rele(ds, FTAG);
941 	dsl_pool_rele(dp, FTAG);
942 	return (err);
943 }
944 
945 /*
946  * Retrieve all properties for a single bookmark in the given dataset.
947  */
948 int
dsl_get_bookmark_props(const char * dsname,const char * bmname,nvlist_t * props)949 dsl_get_bookmark_props(const char *dsname, const char *bmname, nvlist_t *props)
950 {
951 	dsl_pool_t *dp;
952 	dsl_dataset_t *ds;
953 	zfs_bookmark_phys_t bmark_phys = { 0 };
954 	int err;
955 
956 	err = dsl_pool_hold(dsname, FTAG, &dp);
957 	if (err != 0)
958 		return (err);
959 	err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
960 	if (err != 0) {
961 		dsl_pool_rele(dp, FTAG);
962 		return (err);
963 	}
964 
965 	err = dsl_bookmark_lookup_impl(ds, bmname, &bmark_phys);
966 	if (err != 0)
967 		goto out;
968 
969 	dsl_bookmark_fetch_props(dp, &bmark_phys, NULL, props);
970 out:
971 	dsl_dataset_rele(ds, FTAG);
972 	dsl_pool_rele(dp, FTAG);
973 	return (err);
974 }
975 
976 typedef struct dsl_bookmark_destroy_arg {
977 	nvlist_t *dbda_bmarks;
978 	nvlist_t *dbda_success;
979 	nvlist_t *dbda_errors;
980 } dsl_bookmark_destroy_arg_t;
981 
982 static void
dsl_bookmark_destroy_sync_impl(dsl_dataset_t * ds,const char * name,dmu_tx_t * tx)983 dsl_bookmark_destroy_sync_impl(dsl_dataset_t *ds, const char *name,
984     dmu_tx_t *tx)
985 {
986 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
987 	uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
988 	matchtype_t mt = 0;
989 	uint64_t int_size, num_ints;
990 	/*
991 	 * 'search' must be zeroed so that dbn_flags (which is used in
992 	 * dsl_bookmark_compare()) will be zeroed even if the on-disk
993 	 * (in ZAP) bookmark is shorter than offsetof(dbn_flags).
994 	 */
995 	dsl_bookmark_node_t search = { 0 };
996 	char realname[ZFS_MAX_DATASET_NAME_LEN];
997 
998 	/*
999 	 * Find the real name of this bookmark, which may be different
1000 	 * from the given name if the dataset is case-insensitive.  Then
1001 	 * use the real name to find the node in the ds_bookmarks AVL tree.
1002 	 */
1003 
1004 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
1005 		mt = MT_NORMALIZE;
1006 
1007 	VERIFY0(zap_length(mos, bmark_zapobj, name, &int_size, &num_ints));
1008 
1009 	ASSERT3U(int_size, ==, sizeof (uint64_t));
1010 
1011 	if (num_ints * int_size > BOOKMARK_PHYS_SIZE_V1) {
1012 		spa_feature_decr(dmu_objset_spa(mos),
1013 		    SPA_FEATURE_BOOKMARK_V2, tx);
1014 	}
1015 	VERIFY0(zap_lookup_norm(mos, bmark_zapobj, name, sizeof (uint64_t),
1016 	    num_ints, &search.dbn_phys, mt, realname, sizeof (realname), NULL));
1017 
1018 	search.dbn_name = realname;
1019 	dsl_bookmark_node_t *dbn = avl_find(&ds->ds_bookmarks, &search, NULL);
1020 	ASSERT(dbn != NULL);
1021 
1022 	if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
1023 		/*
1024 		 * If this bookmark HAS_FBN, and it is before the most
1025 		 * recent snapshot, then its TXG is a key in the head's
1026 		 * deadlist (and all clones' heads' deadlists).  If this is
1027 		 * the last thing keeping the key (i.e. there are no more
1028 		 * bookmarks with HAS_FBN at this TXG, and there is no
1029 		 * snapshot at this TXG), then remove the key.
1030 		 *
1031 		 * Note that this algorithm depends on ds_bookmarks being
1032 		 * sorted such that all bookmarks at the same TXG with
1033 		 * HAS_FBN are adjacent (with no non-HAS_FBN bookmarks
1034 		 * at the same TXG in between them).  If this were not
1035 		 * the case, we would need to examine *all* bookmarks
1036 		 * at this TXG, rather than just the adjacent ones.
1037 		 */
1038 
1039 		dsl_bookmark_node_t *dbn_prev =
1040 		    AVL_PREV(&ds->ds_bookmarks, dbn);
1041 		dsl_bookmark_node_t *dbn_next =
1042 		    AVL_NEXT(&ds->ds_bookmarks, dbn);
1043 
1044 		boolean_t more_bookmarks_at_this_txg =
1045 		    (dbn_prev != NULL && dbn_prev->dbn_phys.zbm_creation_txg ==
1046 		    dbn->dbn_phys.zbm_creation_txg &&
1047 		    (dbn_prev->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) ||
1048 		    (dbn_next != NULL && dbn_next->dbn_phys.zbm_creation_txg ==
1049 		    dbn->dbn_phys.zbm_creation_txg &&
1050 		    (dbn_next->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
1051 
1052 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS) &&
1053 		    !more_bookmarks_at_this_txg &&
1054 		    dbn->dbn_phys.zbm_creation_txg <
1055 		    dsl_dataset_phys(ds)->ds_prev_snap_txg) {
1056 			dsl_dir_remove_clones_key(ds->ds_dir,
1057 			    dbn->dbn_phys.zbm_creation_txg, tx);
1058 			dsl_deadlist_remove_key(&ds->ds_deadlist,
1059 			    dbn->dbn_phys.zbm_creation_txg, tx);
1060 		}
1061 
1062 		spa_feature_decr(dmu_objset_spa(mos),
1063 		    SPA_FEATURE_BOOKMARK_WRITTEN, tx);
1064 	}
1065 
1066 	if (dbn->dbn_phys.zbm_redaction_obj != 0) {
1067 		dnode_t *rl;
1068 		VERIFY0(dnode_hold(mos,
1069 		    dbn->dbn_phys.zbm_redaction_obj, FTAG, &rl));
1070 		if (rl->dn_have_spill) {
1071 			spa_feature_decr(dmu_objset_spa(mos),
1072 			    SPA_FEATURE_REDACTION_LIST_SPILL, tx);
1073 		}
1074 		dnode_rele(rl, FTAG);
1075 		VERIFY0(dmu_object_free(mos,
1076 		    dbn->dbn_phys.zbm_redaction_obj, tx));
1077 		spa_feature_decr(dmu_objset_spa(mos),
1078 		    SPA_FEATURE_REDACTION_BOOKMARKS, tx);
1079 	}
1080 
1081 	avl_remove(&ds->ds_bookmarks, dbn);
1082 	spa_strfree(dbn->dbn_name);
1083 	mutex_destroy(&dbn->dbn_lock);
1084 	kmem_free(dbn, sizeof (*dbn));
1085 
1086 	VERIFY0(zap_remove_norm(mos, bmark_zapobj, name, mt, tx));
1087 }
1088 
1089 static int
dsl_bookmark_destroy_check(void * arg,dmu_tx_t * tx)1090 dsl_bookmark_destroy_check(void *arg, dmu_tx_t *tx)
1091 {
1092 	dsl_bookmark_destroy_arg_t *dbda = arg;
1093 	dsl_pool_t *dp = dmu_tx_pool(tx);
1094 	int rv = 0;
1095 
1096 	ASSERT(nvlist_empty(dbda->dbda_success));
1097 	ASSERT(nvlist_empty(dbda->dbda_errors));
1098 
1099 	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
1100 		return (0);
1101 
1102 	for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_bmarks, NULL);
1103 	    pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_bmarks, pair)) {
1104 		const char *fullname = nvpair_name(pair);
1105 		dsl_dataset_t *ds;
1106 		zfs_bookmark_phys_t bm;
1107 		int error;
1108 		char *shortname;
1109 
1110 		error = dsl_bookmark_hold_ds(dp, fullname, &ds,
1111 		    FTAG, &shortname);
1112 		if (error == ENOENT) {
1113 			/* ignore it; the bookmark is "already destroyed" */
1114 			continue;
1115 		}
1116 		if (error == 0) {
1117 			error = dsl_bookmark_lookup_impl(ds, shortname, &bm);
1118 			dsl_dataset_rele(ds, FTAG);
1119 			if (error == ESRCH) {
1120 				/*
1121 				 * ignore it; the bookmark is
1122 				 * "already destroyed"
1123 				 */
1124 				continue;
1125 			}
1126 			if (error == 0 && bm.zbm_redaction_obj != 0) {
1127 				redaction_list_t *rl = NULL;
1128 				error = dsl_redaction_list_hold_obj(tx->tx_pool,
1129 				    bm.zbm_redaction_obj, FTAG, &rl);
1130 				if (error == ENOENT) {
1131 					error = 0;
1132 				} else if (error == 0 &&
1133 				    dsl_redaction_list_long_held(rl)) {
1134 					error = SET_ERROR(EBUSY);
1135 				}
1136 				if (rl != NULL) {
1137 					dsl_redaction_list_rele(rl, FTAG);
1138 				}
1139 			}
1140 		}
1141 		if (error == 0) {
1142 			if (dmu_tx_is_syncing(tx)) {
1143 				fnvlist_add_boolean(dbda->dbda_success,
1144 				    fullname);
1145 			}
1146 		} else {
1147 			fnvlist_add_int32(dbda->dbda_errors, fullname, error);
1148 			rv = error;
1149 		}
1150 	}
1151 	return (rv);
1152 }
1153 
1154 static void
dsl_bookmark_destroy_sync(void * arg,dmu_tx_t * tx)1155 dsl_bookmark_destroy_sync(void *arg, dmu_tx_t *tx)
1156 {
1157 	dsl_bookmark_destroy_arg_t *dbda = arg;
1158 	dsl_pool_t *dp = dmu_tx_pool(tx);
1159 	objset_t *mos = dp->dp_meta_objset;
1160 
1161 	for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_success, NULL);
1162 	    pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_success, pair)) {
1163 		dsl_dataset_t *ds;
1164 		char *shortname;
1165 		uint64_t zap_cnt;
1166 
1167 		VERIFY0(dsl_bookmark_hold_ds(dp, nvpair_name(pair),
1168 		    &ds, FTAG, &shortname));
1169 		dsl_bookmark_destroy_sync_impl(ds, shortname, tx);
1170 
1171 		/*
1172 		 * If all of this dataset's bookmarks have been destroyed,
1173 		 * free the zap object and decrement the feature's use count.
1174 		 */
1175 		VERIFY0(zap_count(mos, ds->ds_bookmarks_obj, &zap_cnt));
1176 		if (zap_cnt == 0) {
1177 			dmu_buf_will_dirty(ds->ds_dbuf, tx);
1178 			VERIFY0(zap_destroy(mos, ds->ds_bookmarks_obj, tx));
1179 			ds->ds_bookmarks_obj = 0;
1180 			spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
1181 			VERIFY0(zap_remove(mos, ds->ds_object,
1182 			    DS_FIELD_BOOKMARK_NAMES, tx));
1183 		}
1184 
1185 		spa_history_log_internal_ds(ds, "remove bookmark", tx,
1186 		    "name=%s", shortname);
1187 
1188 		dsl_dataset_rele(ds, FTAG);
1189 	}
1190 }
1191 
1192 /*
1193  * The bookmarks must all be in the same pool.
1194  */
1195 int
dsl_bookmark_destroy(nvlist_t * bmarks,nvlist_t * errors)1196 dsl_bookmark_destroy(nvlist_t *bmarks, nvlist_t *errors)
1197 {
1198 	int rv;
1199 	dsl_bookmark_destroy_arg_t dbda;
1200 	nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
1201 	if (pair == NULL)
1202 		return (0);
1203 
1204 	dbda.dbda_bmarks = bmarks;
1205 	dbda.dbda_errors = errors;
1206 	dbda.dbda_success = fnvlist_alloc();
1207 
1208 	rv = dsl_sync_task(nvpair_name(pair), dsl_bookmark_destroy_check,
1209 	    dsl_bookmark_destroy_sync, &dbda, fnvlist_num_pairs(bmarks),
1210 	    ZFS_SPACE_CHECK_RESERVED);
1211 	fnvlist_free(dbda.dbda_success);
1212 	return (rv);
1213 }
1214 
1215 /* Return B_TRUE if there are any long holds on this dataset. */
1216 boolean_t
dsl_redaction_list_long_held(redaction_list_t * rl)1217 dsl_redaction_list_long_held(redaction_list_t *rl)
1218 {
1219 	return (!zfs_refcount_is_zero(&rl->rl_longholds));
1220 }
1221 
1222 void
dsl_redaction_list_long_hold(dsl_pool_t * dp,redaction_list_t * rl,const void * tag)1223 dsl_redaction_list_long_hold(dsl_pool_t *dp, redaction_list_t *rl,
1224     const void *tag)
1225 {
1226 	ASSERT(dsl_pool_config_held(dp));
1227 	(void) zfs_refcount_add(&rl->rl_longholds, tag);
1228 }
1229 
1230 void
dsl_redaction_list_long_rele(redaction_list_t * rl,const void * tag)1231 dsl_redaction_list_long_rele(redaction_list_t *rl, const void *tag)
1232 {
1233 	(void) zfs_refcount_remove(&rl->rl_longholds, tag);
1234 }
1235 
1236 static void
redaction_list_evict_sync(void * rlu)1237 redaction_list_evict_sync(void *rlu)
1238 {
1239 	redaction_list_t *rl = rlu;
1240 	zfs_refcount_destroy(&rl->rl_longholds);
1241 
1242 	kmem_free(rl, sizeof (redaction_list_t));
1243 }
1244 
1245 void
dsl_redaction_list_rele(redaction_list_t * rl,const void * tag)1246 dsl_redaction_list_rele(redaction_list_t *rl, const void *tag)
1247 {
1248 	if (rl->rl_bonus != rl->rl_dbuf)
1249 		dmu_buf_rele(rl->rl_dbuf, tag);
1250 	dmu_buf_rele(rl->rl_bonus, tag);
1251 }
1252 
1253 int
dsl_redaction_list_hold_obj(dsl_pool_t * dp,uint64_t rlobj,const void * tag,redaction_list_t ** rlp)1254 dsl_redaction_list_hold_obj(dsl_pool_t *dp, uint64_t rlobj, const void *tag,
1255     redaction_list_t **rlp)
1256 {
1257 	objset_t *mos = dp->dp_meta_objset;
1258 	dmu_buf_t *dbuf, *spill_dbuf;
1259 	redaction_list_t *rl;
1260 	int err;
1261 
1262 	ASSERT(dsl_pool_config_held(dp));
1263 
1264 	err = dmu_bonus_hold(mos, rlobj, tag, &dbuf);
1265 	if (err != 0)
1266 		return (err);
1267 
1268 	rl = dmu_buf_get_user(dbuf);
1269 	if (rl == NULL) {
1270 		redaction_list_t *winner = NULL;
1271 
1272 		rl = kmem_zalloc(sizeof (redaction_list_t), KM_SLEEP);
1273 		rl->rl_bonus = dbuf;
1274 		if (dmu_spill_hold_existing(dbuf, tag, &spill_dbuf) == 0) {
1275 			rl->rl_dbuf = spill_dbuf;
1276 		} else {
1277 			rl->rl_dbuf = dbuf;
1278 		}
1279 		rl->rl_object = rlobj;
1280 		rl->rl_phys = rl->rl_dbuf->db_data;
1281 		rl->rl_mos = dp->dp_meta_objset;
1282 		zfs_refcount_create(&rl->rl_longholds);
1283 		dmu_buf_init_user(&rl->rl_dbu, redaction_list_evict_sync, NULL,
1284 		    &rl->rl_bonus);
1285 		if ((winner = dmu_buf_set_user_ie(dbuf, &rl->rl_dbu)) != NULL) {
1286 			kmem_free(rl, sizeof (*rl));
1287 			rl = winner;
1288 		}
1289 	}
1290 	*rlp = rl;
1291 	return (0);
1292 }
1293 
1294 /*
1295  * Snapshot ds is being destroyed.
1296  *
1297  * Adjust the "freed_before_next" of any bookmarks between this snap
1298  * and the previous snapshot, because their "next snapshot" is changing.
1299  *
1300  * If there are any bookmarks with HAS_FBN at this snapshot, remove
1301  * their HAS_SNAP flag (note: there can be at most one snapshot of
1302  * each filesystem at a given txg), and return B_TRUE.  In this case
1303  * the caller can not remove the key in the deadlist at this TXG, because
1304  * the HAS_FBN bookmarks require the key be there.
1305  *
1306  * Returns B_FALSE if there are no bookmarks with HAS_FBN at this
1307  * snapshot's TXG.  In this case the caller can remove the key in the
1308  * deadlist at this TXG.
1309  */
1310 boolean_t
dsl_bookmark_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)1311 dsl_bookmark_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
1312 {
1313 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1314 
1315 	dsl_dataset_t *head, *next;
1316 	VERIFY0(dsl_dataset_hold_obj(dp,
1317 	    dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj, FTAG, &head));
1318 	VERIFY0(dsl_dataset_hold_obj(dp,
1319 	    dsl_dataset_phys(ds)->ds_next_snap_obj, FTAG, &next));
1320 
1321 	/*
1322 	 * Find the first bookmark that HAS_FBN at or after the
1323 	 * previous snapshot.
1324 	 */
1325 	dsl_bookmark_node_t search = { 0 };
1326 	avl_index_t idx;
1327 	search.dbn_phys.zbm_creation_txg =
1328 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1329 	search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1330 	/*
1331 	 * The empty-string name can't be in the AVL, and it compares
1332 	 * before any entries with this TXG.
1333 	 */
1334 	search.dbn_name = (char *)"";
1335 	VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1336 	dsl_bookmark_node_t *dbn =
1337 	    avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1338 
1339 	/*
1340 	 * Iterate over all bookmarks that are at or after the previous
1341 	 * snapshot, and before this (being deleted) snapshot.  Adjust
1342 	 * their FBN based on their new next snapshot.
1343 	 */
1344 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg <
1345 	    dsl_dataset_phys(ds)->ds_creation_txg;
1346 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1347 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN))
1348 			continue;
1349 		/*
1350 		 * Increase our FBN by the amount of space that was live
1351 		 * (referenced) at the time of this bookmark (i.e.
1352 		 * birth <= zbm_creation_txg), and killed between this
1353 		 * (being deleted) snapshot and the next snapshot (i.e.
1354 		 * on the next snapshot's deadlist).  (Space killed before
1355 		 * this are already on our FBN.)
1356 		 */
1357 		uint64_t referenced, compressed, uncompressed;
1358 		dsl_deadlist_space_range(&next->ds_deadlist,
1359 		    0, dbn->dbn_phys.zbm_creation_txg,
1360 		    &referenced, &compressed, &uncompressed);
1361 		dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1362 		    referenced;
1363 		dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1364 		    compressed;
1365 		dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1366 		    uncompressed;
1367 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1368 		    dbn->dbn_name, sizeof (uint64_t),
1369 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1370 		    &dbn->dbn_phys, tx));
1371 	}
1372 	dsl_dataset_rele(next, FTAG);
1373 
1374 	/*
1375 	 * There may be several bookmarks at this txg (the TXG of the
1376 	 * snapshot being deleted).  We need to clear the SNAPSHOT_EXISTS
1377 	 * flag on all of them, and return TRUE if there is at least 1
1378 	 * bookmark here with HAS_FBN (thus preventing the deadlist
1379 	 * key from being removed).
1380 	 */
1381 	boolean_t rv = B_FALSE;
1382 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1383 	    dsl_dataset_phys(ds)->ds_creation_txg;
1384 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1385 		if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1386 			ASSERT(!(dbn->dbn_phys.zbm_flags &
1387 			    ZBM_FLAG_SNAPSHOT_EXISTS));
1388 			continue;
1389 		}
1390 		ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS);
1391 		dbn->dbn_phys.zbm_flags &= ~ZBM_FLAG_SNAPSHOT_EXISTS;
1392 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1393 		    dbn->dbn_name, sizeof (uint64_t),
1394 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1395 		    &dbn->dbn_phys, tx));
1396 		rv = B_TRUE;
1397 	}
1398 	dsl_dataset_rele(head, FTAG);
1399 	return (rv);
1400 }
1401 
1402 /*
1403  * A snapshot is being created of this (head) dataset.
1404  *
1405  * We don't keep keys in the deadlist for the most recent snapshot, or any
1406  * bookmarks at or after it, because there can't be any blocks on the
1407  * deadlist in this range.  Now that the most recent snapshot is after
1408  * all bookmarks, we need to add these keys.  Note that the caller always
1409  * adds a key at the previous snapshot, so we only add keys for bookmarks
1410  * after that.
1411  */
1412 void
dsl_bookmark_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)1413 dsl_bookmark_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
1414 {
1415 	uint64_t last_key_added = UINT64_MAX;
1416 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1417 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >
1418 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1419 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1420 		uint64_t creation_txg = dbn->dbn_phys.zbm_creation_txg;
1421 		ASSERT3U(creation_txg, <=, last_key_added);
1422 		/*
1423 		 * Note, there may be multiple bookmarks at this TXG,
1424 		 * and we only want to add the key for this TXG once.
1425 		 * The ds_bookmarks AVL is sorted by TXG, so we will visit
1426 		 * these bookmarks in sequence.
1427 		 */
1428 		if ((dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) &&
1429 		    creation_txg != last_key_added) {
1430 			dsl_deadlist_add_key(&ds->ds_deadlist,
1431 			    creation_txg, tx);
1432 			last_key_added = creation_txg;
1433 		}
1434 	}
1435 }
1436 
1437 /*
1438  * The next snapshot of the origin dataset has changed, due to
1439  * promote or clone swap.  If there are any bookmarks at this dataset,
1440  * we need to update their zbm_*_freed_before_next_snap to reflect this.
1441  * The head dataset has the relevant bookmarks in ds_bookmarks.
1442  */
1443 void
dsl_bookmark_next_changed(dsl_dataset_t * head,dsl_dataset_t * origin,dmu_tx_t * tx)1444 dsl_bookmark_next_changed(dsl_dataset_t *head, dsl_dataset_t *origin,
1445     dmu_tx_t *tx)
1446 {
1447 	dsl_pool_t *dp = dmu_tx_pool(tx);
1448 
1449 	/*
1450 	 * Find the first bookmark that HAS_FBN at the origin snapshot.
1451 	 */
1452 	dsl_bookmark_node_t search = { 0 };
1453 	avl_index_t idx;
1454 	search.dbn_phys.zbm_creation_txg =
1455 	    dsl_dataset_phys(origin)->ds_creation_txg;
1456 	search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1457 	/*
1458 	 * The empty-string name can't be in the AVL, and it compares
1459 	 * before any entries with this TXG.
1460 	 */
1461 	search.dbn_name = (char *)"";
1462 	VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1463 	dsl_bookmark_node_t *dbn =
1464 	    avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1465 
1466 	/*
1467 	 * Iterate over all bookmarks that are at the origin txg.
1468 	 * Adjust their FBN based on their new next snapshot.
1469 	 */
1470 	for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1471 	    dsl_dataset_phys(origin)->ds_creation_txg &&
1472 	    (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1473 	    dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1474 
1475 		/*
1476 		 * Bookmark is at the origin, therefore its
1477 		 * "next dataset" is changing, so we need
1478 		 * to reset its FBN by recomputing it in
1479 		 * dsl_bookmark_set_phys().
1480 		 */
1481 		ASSERT3U(dbn->dbn_phys.zbm_guid, ==,
1482 		    dsl_dataset_phys(origin)->ds_guid);
1483 		ASSERT3U(dbn->dbn_phys.zbm_referenced_bytes_refd, ==,
1484 		    dsl_dataset_phys(origin)->ds_referenced_bytes);
1485 		ASSERT(dbn->dbn_phys.zbm_flags &
1486 		    ZBM_FLAG_SNAPSHOT_EXISTS);
1487 		/*
1488 		 * Save and restore the zbm_redaction_obj, which
1489 		 * is zeroed by dsl_bookmark_set_phys().
1490 		 */
1491 		uint64_t redaction_obj =
1492 		    dbn->dbn_phys.zbm_redaction_obj;
1493 		dsl_bookmark_set_phys(&dbn->dbn_phys, origin);
1494 		dbn->dbn_phys.zbm_redaction_obj = redaction_obj;
1495 
1496 		VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1497 		    dbn->dbn_name, sizeof (uint64_t),
1498 		    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1499 		    &dbn->dbn_phys, tx));
1500 	}
1501 }
1502 
1503 /*
1504  * This block is no longer referenced by this (head) dataset.
1505  *
1506  * Adjust the FBN of any bookmarks that reference this block, whose "next"
1507  * is the head dataset.
1508  */
1509 void
dsl_bookmark_block_killed(dsl_dataset_t * ds,const blkptr_t * bp,dmu_tx_t * tx)1510 dsl_bookmark_block_killed(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
1511 {
1512 	(void) tx;
1513 
1514 	/*
1515 	 * Iterate over bookmarks whose "next" is the head dataset.
1516 	 */
1517 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1518 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1519 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1520 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1521 		/*
1522 		 * If the block was live (referenced) at the time of this
1523 		 * bookmark, add its space to the bookmark's FBN.
1524 		 */
1525 		if (BP_GET_LOGICAL_BIRTH(bp) <=
1526 		    dbn->dbn_phys.zbm_creation_txg &&
1527 		    (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1528 			mutex_enter(&dbn->dbn_lock);
1529 			dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1530 			    bp_get_dsize_sync(dsl_dataset_get_spa(ds), bp);
1531 			dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1532 			    BP_GET_PSIZE(bp);
1533 			dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1534 			    BP_GET_UCSIZE(bp);
1535 			/*
1536 			 * Changing the ZAP object here would be too
1537 			 * expensive.  Also, we may be called from the zio
1538 			 * interrupt thread, which can't block on i/o.
1539 			 * Therefore, we mark this bookmark as dirty and
1540 			 * modify the ZAP once per txg, in
1541 			 * dsl_bookmark_sync_done().
1542 			 */
1543 			dbn->dbn_dirty = B_TRUE;
1544 			mutex_exit(&dbn->dbn_lock);
1545 		}
1546 	}
1547 }
1548 
1549 void
dsl_bookmark_sync_done(dsl_dataset_t * ds,dmu_tx_t * tx)1550 dsl_bookmark_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
1551 {
1552 	dsl_pool_t *dp = dmu_tx_pool(tx);
1553 
1554 	if (dsl_dataset_is_snapshot(ds))
1555 		return;
1556 
1557 	/*
1558 	 * We only dirty bookmarks that are at or after the most recent
1559 	 * snapshot.  We can't create snapshots between
1560 	 * dsl_bookmark_block_killed() and dsl_bookmark_sync_done(), so we
1561 	 * don't need to look at any bookmarks before ds_prev_snap_txg.
1562 	 */
1563 	for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1564 	    dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1565 	    dsl_dataset_phys(ds)->ds_prev_snap_txg;
1566 	    dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1567 		if (dbn->dbn_dirty) {
1568 			/*
1569 			 * We only dirty nodes with HAS_FBN, therefore
1570 			 * we can always use the current bookmark struct size.
1571 			 */
1572 			ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1573 			VERIFY0(zap_update(dp->dp_meta_objset,
1574 			    ds->ds_bookmarks_obj,
1575 			    dbn->dbn_name, sizeof (uint64_t),
1576 			    sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1577 			    &dbn->dbn_phys, tx));
1578 			dbn->dbn_dirty = B_FALSE;
1579 		}
1580 	}
1581 #ifdef ZFS_DEBUG
1582 	for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
1583 	    dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
1584 		ASSERT(!dbn->dbn_dirty);
1585 	}
1586 #endif
1587 }
1588 
1589 /*
1590  * Return the TXG of the most recent bookmark (or 0 if there are no bookmarks).
1591  */
1592 uint64_t
dsl_bookmark_latest_txg(dsl_dataset_t * ds)1593 dsl_bookmark_latest_txg(dsl_dataset_t *ds)
1594 {
1595 	ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool));
1596 	dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1597 	if (dbn == NULL)
1598 		return (0);
1599 	return (dbn->dbn_phys.zbm_creation_txg);
1600 }
1601 
1602 /*
1603  * Compare the redact_block_phys_t to the bookmark. If the last block in the
1604  * redact_block_phys_t is before the bookmark, return -1.  If the first block in
1605  * the redact_block_phys_t is after the bookmark, return 1.  Otherwise, the
1606  * bookmark is inside the range of the redact_block_phys_t, and we return 0.
1607  */
1608 static int
redact_block_zb_compare(redact_block_phys_t * first,zbookmark_phys_t * second)1609 redact_block_zb_compare(redact_block_phys_t *first,
1610     zbookmark_phys_t *second)
1611 {
1612 	/*
1613 	 * If the block_phys is for a previous object, or the last block in the
1614 	 * block_phys is strictly before the block in the bookmark, the
1615 	 * block_phys is earlier.
1616 	 */
1617 	if (first->rbp_object < second->zb_object ||
1618 	    (first->rbp_object == second->zb_object &&
1619 	    first->rbp_blkid + (redact_block_get_count(first) - 1) <
1620 	    second->zb_blkid)) {
1621 		return (-1);
1622 	}
1623 
1624 	/*
1625 	 * If the bookmark is for a previous object, or the block in the
1626 	 * bookmark is strictly before the first block in the block_phys, the
1627 	 * bookmark is earlier.
1628 	 */
1629 	if (first->rbp_object > second->zb_object ||
1630 	    (first->rbp_object == second->zb_object &&
1631 	    first->rbp_blkid > second->zb_blkid)) {
1632 		return (1);
1633 	}
1634 
1635 	return (0);
1636 }
1637 
1638 /*
1639  * Traverse the redaction list in the provided object, and call the callback for
1640  * each entry we find. Don't call the callback for any records before resume.
1641  */
1642 int
dsl_redaction_list_traverse(redaction_list_t * rl,zbookmark_phys_t * resume,rl_traverse_callback_t cb,void * arg)1643 dsl_redaction_list_traverse(redaction_list_t *rl, zbookmark_phys_t *resume,
1644     rl_traverse_callback_t cb, void *arg)
1645 {
1646 	objset_t *mos = rl->rl_mos;
1647 	int err = 0;
1648 
1649 	if (rl->rl_phys->rlp_last_object != UINT64_MAX ||
1650 	    rl->rl_phys->rlp_last_blkid != UINT64_MAX) {
1651 		/*
1652 		 * When we finish a send, we update the last object and offset
1653 		 * to UINT64_MAX.  If a send fails partway through, the last
1654 		 * object and offset will have some other value, indicating how
1655 		 * far the send got. The redaction list must be complete before
1656 		 * it can be traversed, so return EINVAL if the last object and
1657 		 * blkid are not set to UINT64_MAX.
1658 		 */
1659 		return (SET_ERROR(EINVAL));
1660 	}
1661 
1662 	/*
1663 	 * This allows us to skip the binary search and resume checking logic
1664 	 * below, if we're not resuming a redacted send.
1665 	 */
1666 	if (ZB_IS_ZERO(resume))
1667 		resume = NULL;
1668 
1669 	/*
1670 	 * Binary search for the point to resume from.
1671 	 */
1672 	uint64_t maxidx = rl->rl_phys->rlp_num_entries - 1;
1673 	uint64_t minidx = 0;
1674 	while (resume != NULL && maxidx > minidx) {
1675 		redact_block_phys_t rbp = { 0 };
1676 		ASSERT3U(maxidx, >, minidx);
1677 		uint64_t mididx = minidx + ((maxidx - minidx) / 2);
1678 		err = dmu_read(mos, rl->rl_object, mididx * sizeof (rbp),
1679 		    sizeof (rbp), &rbp, DMU_READ_NO_PREFETCH);
1680 		if (err != 0)
1681 			break;
1682 
1683 		int cmp = redact_block_zb_compare(&rbp, resume);
1684 
1685 		if (cmp == 0) {
1686 			minidx = mididx;
1687 			break;
1688 		} else if (cmp > 0) {
1689 			maxidx =
1690 			    (mididx == minidx ? minidx : mididx - 1);
1691 		} else {
1692 			minidx = mididx + 1;
1693 		}
1694 	}
1695 
1696 	unsigned int bufsize = SPA_OLD_MAXBLOCKSIZE;
1697 	redact_block_phys_t *buf = zio_data_buf_alloc(bufsize);
1698 
1699 	unsigned int entries_per_buf = bufsize / sizeof (redact_block_phys_t);
1700 	uint64_t start_block = minidx / entries_per_buf;
1701 	err = dmu_read(mos, rl->rl_object, start_block * bufsize, bufsize, buf,
1702 	    DMU_READ_PREFETCH);
1703 
1704 	for (uint64_t curidx = minidx;
1705 	    err == 0 && curidx < rl->rl_phys->rlp_num_entries;
1706 	    curidx++) {
1707 		/*
1708 		 * We read in the redaction list one block at a time.  Once we
1709 		 * finish with all the entries in a given block, we read in a
1710 		 * new one.  The predictive prefetcher will take care of any
1711 		 * prefetching, and this code shouldn't be the bottleneck, so we
1712 		 * don't need to do manual prefetching.
1713 		 */
1714 		if (curidx % entries_per_buf == 0) {
1715 			err = dmu_read(mos, rl->rl_object, curidx *
1716 			    sizeof (*buf), bufsize, buf,
1717 			    DMU_READ_PREFETCH);
1718 			if (err != 0)
1719 				break;
1720 		}
1721 		redact_block_phys_t *rb = &buf[curidx % entries_per_buf];
1722 		/*
1723 		 * If resume is non-null, we should either not send the data, or
1724 		 * null out resume so we don't have to keep doing these
1725 		 * comparisons.
1726 		 */
1727 		if (resume != NULL) {
1728 			/*
1729 			 * It is possible that after the binary search we got
1730 			 * a record before the resume point. There's two cases
1731 			 * where this can occur. If the record is the last
1732 			 * redaction record, and the resume point is after the
1733 			 * end of the redacted data, curidx will be the last
1734 			 * redaction record. In that case, the loop will end
1735 			 * after this iteration. The second case is if the
1736 			 * resume point is between two redaction records, the
1737 			 * binary search can return either the record before
1738 			 * or after the resume point. In that case, the next
1739 			 * iteration will be greater than the resume point.
1740 			 */
1741 			if (redact_block_zb_compare(rb, resume) < 0) {
1742 				ASSERT3U(curidx, ==, minidx);
1743 				continue;
1744 			} else {
1745 				/*
1746 				 * If the place to resume is in the middle of
1747 				 * the range described by this
1748 				 * redact_block_phys, then modify the
1749 				 * redact_block_phys in memory so we generate
1750 				 * the right records.
1751 				 */
1752 				if (resume->zb_object == rb->rbp_object &&
1753 				    resume->zb_blkid > rb->rbp_blkid) {
1754 					uint64_t diff = resume->zb_blkid -
1755 					    rb->rbp_blkid;
1756 					rb->rbp_blkid = resume->zb_blkid;
1757 					redact_block_set_count(rb,
1758 					    redact_block_get_count(rb) - diff);
1759 				}
1760 				resume = NULL;
1761 			}
1762 		}
1763 
1764 		if (cb(rb, arg) != 0) {
1765 			err = EINTR;
1766 			break;
1767 		}
1768 	}
1769 
1770 	zio_data_buf_free(buf, bufsize);
1771 	return (err);
1772 }
1773