module/zfs/dsl_bookmark.c

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