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         bzero(bmark_phys, sizeof (*bmark_phys));
  86
  87         err = zap_lookup_norm(mos, bmark_zapobj, shortname, sizeof (uint64_t),
  88             sizeof (*bmark_phys) / sizeof (uint64_t), bmark_phys, mt, NULL, 0,
  89             NULL);
  90
  91         return (err == ENOENT ? SET_ERROR(ESRCH) : err);
  92 }
  93
  94 /*
  95  * If later_ds is non-NULL, this will return EXDEV if the specified bookmark
  96  * does not represents an earlier point in later_ds's timeline.  However,
  97  * bmp will still be filled in if we return EXDEV.
  98  *
  99  * Returns ENOENT if the dataset containing the bookmark does not exist.
 100  * Returns ESRCH if the dataset exists but the bookmark was not found in it.
 101  */
 102 int
 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;
 164         size_t first_len;
 165
 166         first = NULL;
 167         for (nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
 168             pair != NULL; pair = nvlist_next_nvpair(bmarks, pair)) {
 169
 170                 char *bmark = nvpair_name(pair);
 171                 char *source;
 172
 173                 /* list structure: values must be snapshots XOR bookmarks */
 174                 if (nvpair_value_string(pair, &source) != 0)
 175                         return (-1);
 176                 if (dsl_bookmark_create_nvl_validate_pair(bmark, source) != 0)
 177                         return (-1);
 178
 179                 /* same pool check */
 180                 if (first == NULL) {
 181                         char *cp = strpbrk(bmark, "/#");
 182                         if (cp == NULL)
 183                                 return (-1);
 184                         first = bmark;
 185                         first_len = cp - bmark;
 186                 }
 187                 if (strncmp(first, bmark, first_len) != 0)
 188                         return (-1);
 189                 switch (*(bmark + first_len)) {
 190                         case '/': /* fallthrough */
 191                         case '#':
 192                                 break;
 193                         default:
 194                                 return (-1);
 195                 }
 196
 197                 /* unique newbookmark names; todo: O(n^2) */
 198                 for (nvpair_t *pair2 = nvlist_next_nvpair(bmarks, pair);
 199                     pair2 != NULL; pair2 = nvlist_next_nvpair(bmarks, pair2)) {
 200                         if (strcmp(nvpair_name(pair), nvpair_name(pair2)) == 0)
 201                                 return (-1);
 202                 }
 203
 204         }
 205         return (0);
 206 }
 207
 208 /*
 209  * expects that newbm and source have been validated using
 210  * dsl_bookmark_create_nvl_validate_pair
 211  */
 212 static int
 213 dsl_bookmark_create_check_impl(dsl_pool_t *dp,
 214     const char *newbm, const char *source)
 215 {
 216         ASSERT0(dsl_bookmark_create_nvl_validate_pair(newbm, source));
 217         /* defer source namecheck until we know it's a snapshot or bookmark */
 218
 219         int error;
 220         dsl_dataset_t *newbm_ds;
 221         char *newbm_short;
 222         zfs_bookmark_phys_t bmark_phys;
 223
 224         error = dsl_bookmark_hold_ds(dp, newbm, &newbm_ds, FTAG, &newbm_short);
 225         if (error != 0)
 226                 return (error);
 227
 228         /* Verify that the new bookmark does not already exist */
 229         error = dsl_bookmark_lookup_impl(newbm_ds, newbm_short, &bmark_phys);
 230         switch (error) {
 231         case ESRCH:
 232                 /* happy path: new bmark doesn't exist, proceed after switch */
 233                 error = 0;
 234                 break;
 235         case 0:
 236                 error = SET_ERROR(EEXIST);
 237                 goto eholdnewbmds;
 238         default:
 239                 /* dsl_bookmark_lookup_impl already did SET_ERROR */
 240                 goto eholdnewbmds;
 241         }
 242
 243         /* error is retval of the following if-cascade */
 244         if (strchr(source, '@') != NULL) {
 245                 dsl_dataset_t *source_snap_ds;
 246                 ASSERT3S(snapshot_namecheck(source, NULL, NULL), ==, 0);
 247                 error = dsl_dataset_hold(dp, source, FTAG, &source_snap_ds);
 248                 if (error == 0) {
 249                         VERIFY(source_snap_ds->ds_is_snapshot);
 250                         /*
 251                          * Verify that source snapshot is an earlier point in
 252                          * newbm_ds's timeline (source may be newbm_ds's origin)
 253                          */
 254                         if (!dsl_dataset_is_before(newbm_ds, source_snap_ds, 0))
 255                                 error = SET_ERROR(
 256                                     ZFS_ERR_BOOKMARK_SOURCE_NOT_ANCESTOR);
 257                         dsl_dataset_rele(source_snap_ds, FTAG);
 258                 }
 259         } else if (strchr(source, '#') != NULL) {
 260                 zfs_bookmark_phys_t source_phys;
 261                 ASSERT3S(bookmark_namecheck(source, NULL, NULL), ==, 0);
 262                 /*
 263                  * Source must exists and be an earlier point in newbm_ds's
 264                  * timeline (newbm_ds's origin may be a snap of source's ds)
 265                  */
 266                 error = dsl_bookmark_lookup(dp, source, newbm_ds, &source_phys);
 267                 switch (error) {
 268                 case 0:
 269                         break; /* happy path */
 270                 case EXDEV:
 271                         error = SET_ERROR(ZFS_ERR_BOOKMARK_SOURCE_NOT_ANCESTOR);
 272                         break;
 273                 default:
 274                         /* dsl_bookmark_lookup already did SET_ERROR */
 275                         break;
 276                 }
 277         } else {
 278                 /*
 279                  * dsl_bookmark_create_nvl_validate validates that source is
 280                  * either snapshot or bookmark
 281                  */
 282                 panic("unreachable code: %s", source);
 283         }
 284
 285 eholdnewbmds:
 286         dsl_dataset_rele(newbm_ds, FTAG);
 287         return (error);
 288 }
 289
 290 int
 291 dsl_bookmark_create_check(void *arg, dmu_tx_t *tx)
 292 {
 293         dsl_bookmark_create_arg_t *dbca = arg;
 294         int rv = 0;
 295         int schema_err = 0;
 296         ASSERT3P(dbca, !=, NULL);
 297         ASSERT3P(dbca->dbca_bmarks, !=, NULL);
 298         /* dbca->dbca_errors is allowed to be NULL */
 299
 300         dsl_pool_t *dp = dmu_tx_pool(tx);
 301
 302         if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
 303                 return (SET_ERROR(ENOTSUP));
 304
 305         if (dsl_bookmark_create_nvl_validate(dbca->dbca_bmarks) != 0)
 306                 rv = schema_err = SET_ERROR(EINVAL);
 307
 308         for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
 309             pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
 310                 char *new = nvpair_name(pair);
 311
 312                 int error = schema_err;
 313                 if (error == 0) {
 314                         char *source = fnvpair_value_string(pair);
 315                         error = dsl_bookmark_create_check_impl(dp, new, source);
 316                         if (error != 0)
 317                                 error = SET_ERROR(error);
 318                 }
 319
 320                 if (error != 0) {
 321                         rv = error;
 322                         if (dbca->dbca_errors != NULL)
 323                                 fnvlist_add_int32(dbca->dbca_errors,
 324                                     new, error);
 325                 }
 326         }
 327
 328         return (rv);
 329 }
 330
 331 static dsl_bookmark_node_t *
 332 dsl_bookmark_node_alloc(char *shortname)
 333 {
 334         dsl_bookmark_node_t *dbn = kmem_alloc(sizeof (*dbn), KM_SLEEP);
 335         dbn->dbn_name = spa_strdup(shortname);
 336         dbn->dbn_dirty = B_FALSE;
 337         mutex_init(&dbn->dbn_lock, NULL, MUTEX_DEFAULT, NULL);
 338         return (dbn);
 339 }
 340
 341 /*
 342  * Set the fields in the zfs_bookmark_phys_t based on the specified snapshot.
 343  */
 344 static void
 345 dsl_bookmark_set_phys(zfs_bookmark_phys_t *zbm, dsl_dataset_t *snap)
 346 {
 347         spa_t *spa = dsl_dataset_get_spa(snap);
 348         objset_t *mos = spa_get_dsl(spa)->dp_meta_objset;
 349         dsl_dataset_phys_t *dsp = dsl_dataset_phys(snap);
 350         zbm->zbm_guid = dsp->ds_guid;
 351         zbm->zbm_creation_txg = dsp->ds_creation_txg;
 352         zbm->zbm_creation_time = dsp->ds_creation_time;
 353         zbm->zbm_redaction_obj = 0;
 354
 355         /*
 356          * If the dataset is encrypted create a larger bookmark to
 357          * accommodate the IVset guid. The IVset guid was added
 358          * after the encryption feature to prevent a problem with
 359          * raw sends. If we encounter an encrypted dataset without
 360          * an IVset guid we fall back to a normal bookmark.
 361          */
 362         if (snap->ds_dir->dd_crypto_obj != 0 &&
 363             spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
 364                 (void) zap_lookup(mos, snap->ds_object,
 365                     DS_FIELD_IVSET_GUID, sizeof (uint64_t), 1,
 366                     &zbm->zbm_ivset_guid);
 367         }
 368
 369         if (spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_WRITTEN)) {
 370                 zbm->zbm_flags = ZBM_FLAG_SNAPSHOT_EXISTS | ZBM_FLAG_HAS_FBN;
 371                 zbm->zbm_referenced_bytes_refd = dsp->ds_referenced_bytes;
 372                 zbm->zbm_compressed_bytes_refd = dsp->ds_compressed_bytes;
 373                 zbm->zbm_uncompressed_bytes_refd = dsp->ds_uncompressed_bytes;
 374
 375                 dsl_dataset_t *nextds;
 376                 VERIFY0(dsl_dataset_hold_obj(snap->ds_dir->dd_pool,
 377                     dsp->ds_next_snap_obj, FTAG, &nextds));
 378                 dsl_deadlist_space(&nextds->ds_deadlist,
 379                     &zbm->zbm_referenced_freed_before_next_snap,
 380                     &zbm->zbm_compressed_freed_before_next_snap,
 381                     &zbm->zbm_uncompressed_freed_before_next_snap);
 382                 dsl_dataset_rele(nextds, FTAG);
 383         } else {
 384                 bzero(&zbm->zbm_flags,
 385                     sizeof (zfs_bookmark_phys_t) -
 386                     offsetof(zfs_bookmark_phys_t, zbm_flags));
 387         }
 388 }
 389
 390 /*
 391  * Add dsl_bookmark_node_t `dbn` to the given dataset and increment appropriate
 392  * SPA feature counters.
 393  */
 394 void
 395 dsl_bookmark_node_add(dsl_dataset_t *hds, dsl_bookmark_node_t *dbn,
 396     dmu_tx_t *tx)
 397 {
 398         dsl_pool_t *dp = dmu_tx_pool(tx);
 399         objset_t *mos = dp->dp_meta_objset;
 400
 401         if (hds->ds_bookmarks_obj == 0) {
 402                 hds->ds_bookmarks_obj = zap_create_norm(mos,
 403                     U8_TEXTPREP_TOUPPER, DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0,
 404                     tx);
 405                 spa_feature_incr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
 406
 407                 dsl_dataset_zapify(hds, tx);
 408                 VERIFY0(zap_add(mos, hds->ds_object,
 409                     DS_FIELD_BOOKMARK_NAMES,
 410                     sizeof (hds->ds_bookmarks_obj), 1,
 411                     &hds->ds_bookmarks_obj, tx));
 412         }
 413
 414         avl_add(&hds->ds_bookmarks, dbn);
 415
 416         /*
 417          * To maintain backwards compatibility with software that doesn't
 418          * understand SPA_FEATURE_BOOKMARK_V2, we need to use the smallest
 419          * possible bookmark size.
 420          */
 421         uint64_t bookmark_phys_size = BOOKMARK_PHYS_SIZE_V1;
 422         if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARK_V2) &&
 423             (dbn->dbn_phys.zbm_ivset_guid != 0 || dbn->dbn_phys.zbm_flags &
 424             ZBM_FLAG_HAS_FBN || dbn->dbn_phys.zbm_redaction_obj != 0)) {
 425                 bookmark_phys_size = BOOKMARK_PHYS_SIZE_V2;
 426                 spa_feature_incr(dp->dp_spa, SPA_FEATURE_BOOKMARK_V2, tx);
 427         }
 428
 429         __attribute__((unused)) zfs_bookmark_phys_t zero_phys = { 0 };
 430         ASSERT0(bcmp(((char *)&dbn->dbn_phys) + bookmark_phys_size,
 431             &zero_phys, sizeof (zfs_bookmark_phys_t) - bookmark_phys_size));
 432
 433         VERIFY0(zap_add(mos, hds->ds_bookmarks_obj, dbn->dbn_name,
 434             sizeof (uint64_t), bookmark_phys_size / sizeof (uint64_t),
 435             &dbn->dbn_phys, tx));
 436 }
 437
 438 /*
 439  * If redaction_list is non-null, we create a redacted bookmark and redaction
 440  * list, and store the object number of the redaction list in redact_obj.
 441  */
 442 static void
 443 dsl_bookmark_create_sync_impl_snap(const char *bookmark, const char *snapshot,
 444     dmu_tx_t *tx, uint64_t num_redact_snaps, uint64_t *redact_snaps, void *tag,
 445     redaction_list_t **redaction_list)
 446 {
 447         dsl_pool_t *dp = dmu_tx_pool(tx);
 448         objset_t *mos = dp->dp_meta_objset;
 449         dsl_dataset_t *snapds, *bmark_fs;
 450         char *shortname;
 451         boolean_t bookmark_redacted;
 452         uint64_t *dsredactsnaps;
 453         uint64_t dsnumsnaps;
 454
 455         VERIFY0(dsl_dataset_hold(dp, snapshot, FTAG, &snapds));
 456         VERIFY0(dsl_bookmark_hold_ds(dp, bookmark, &bmark_fs, FTAG,
 457             &shortname));
 458
 459         dsl_bookmark_node_t *dbn = dsl_bookmark_node_alloc(shortname);
 460         dsl_bookmark_set_phys(&dbn->dbn_phys, snapds);
 461
 462         bookmark_redacted = dsl_dataset_get_uint64_array_feature(snapds,
 463             SPA_FEATURE_REDACTED_DATASETS, &dsnumsnaps, &dsredactsnaps);
 464         if (redaction_list != NULL || bookmark_redacted) {
 465                 redaction_list_t *local_rl;
 466                 if (bookmark_redacted) {
 467                         redact_snaps = dsredactsnaps;
 468                         num_redact_snaps = dsnumsnaps;
 469                 }
 470                 dbn->dbn_phys.zbm_redaction_obj = dmu_object_alloc(mos,
 471                     DMU_OTN_UINT64_METADATA, SPA_OLD_MAXBLOCKSIZE,
 472                     DMU_OTN_UINT64_METADATA, sizeof (redaction_list_phys_t) +
 473                     num_redact_snaps * sizeof (uint64_t), tx);
 474                 spa_feature_incr(dp->dp_spa,
 475                     SPA_FEATURE_REDACTION_BOOKMARKS, tx);
 476
 477                 VERIFY0(dsl_redaction_list_hold_obj(dp,
 478                     dbn->dbn_phys.zbm_redaction_obj, tag, &local_rl));
 479                 dsl_redaction_list_long_hold(dp, local_rl, tag);
 480
 481                 ASSERT3U((local_rl)->rl_dbuf->db_size, >=,
 482                     sizeof (redaction_list_phys_t) + num_redact_snaps *
 483                     sizeof (uint64_t));
 484                 dmu_buf_will_dirty(local_rl->rl_dbuf, tx);
 485                 bcopy(redact_snaps, local_rl->rl_phys->rlp_snaps,
 486                     sizeof (uint64_t) * num_redact_snaps);
 487                 local_rl->rl_phys->rlp_num_snaps = num_redact_snaps;
 488                 if (bookmark_redacted) {
 489                         ASSERT3P(redaction_list, ==, NULL);
 490                         local_rl->rl_phys->rlp_last_blkid = UINT64_MAX;
 491                         local_rl->rl_phys->rlp_last_object = UINT64_MAX;
 492                         dsl_redaction_list_long_rele(local_rl, tag);
 493                         dsl_redaction_list_rele(local_rl, tag);
 494                 } else {
 495                         *redaction_list = local_rl;
 496                 }
 497         }
 498
 499         if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
 500                 spa_feature_incr(dp->dp_spa,
 501                     SPA_FEATURE_BOOKMARK_WRITTEN, tx);
 502         }
 503
 504         dsl_bookmark_node_add(bmark_fs, dbn, tx);
 505
 506         spa_history_log_internal_ds(bmark_fs, "bookmark", tx,
 507             "name=%s creation_txg=%llu target_snap=%llu redact_obj=%llu",
 508             shortname, (longlong_t)dbn->dbn_phys.zbm_creation_txg,
 509             (longlong_t)snapds->ds_object,
 510             (longlong_t)dbn->dbn_phys.zbm_redaction_obj);
 511
 512         dsl_dataset_rele(bmark_fs, FTAG);
 513         dsl_dataset_rele(snapds, FTAG);
 514 }
 515
 516
 517 static void
 518 dsl_bookmark_create_sync_impl_book(
 519     const char *new_name, const char *source_name, dmu_tx_t *tx)
 520 {
 521         dsl_pool_t *dp = dmu_tx_pool(tx);
 522         dsl_dataset_t *bmark_fs_source, *bmark_fs_new;
 523         char *source_shortname, *new_shortname;
 524         zfs_bookmark_phys_t source_phys;
 525
 526         VERIFY0(dsl_bookmark_hold_ds(dp, source_name, &bmark_fs_source, FTAG,
 527             &source_shortname));
 528         VERIFY0(dsl_bookmark_hold_ds(dp, new_name, &bmark_fs_new, FTAG,
 529             &new_shortname));
 530
 531         /*
 532          * create a copy of the source bookmark by copying most of its members
 533          *
 534          * Caveat: bookmarking a redaction bookmark yields a normal bookmark
 535          * -----------------------------------------------------------------
 536          * Reasoning:
 537          * - The zbm_redaction_obj would be referred to by both source and new
 538          *   bookmark, but would be destroyed once either source or new is
 539          *   destroyed, resulting in use-after-free of the referred object.
 540          * - User expectation when issuing the `zfs bookmark` command is that
 541          *   a normal bookmark of the source is created
 542          *
 543          * Design Alternatives For Full Redaction Bookmark Copying:
 544          * - reference-count the redaction object => would require on-disk
 545          *   format change for existing redaction objects
 546          * - Copy the redaction object => cannot be done in syncing context
 547          *   because the redaction object might be too large
 548          */
 549
 550         VERIFY0(dsl_bookmark_lookup_impl(bmark_fs_source, source_shortname,
 551             &source_phys));
 552         dsl_bookmark_node_t *new_dbn = dsl_bookmark_node_alloc(new_shortname);
 553
 554         memcpy(&new_dbn->dbn_phys, &source_phys, sizeof (source_phys));
 555         new_dbn->dbn_phys.zbm_redaction_obj = 0;
 556
 557         /* update feature counters */
 558         if (new_dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
 559                 spa_feature_incr(dp->dp_spa,
 560                     SPA_FEATURE_BOOKMARK_WRITTEN, tx);
 561         }
 562         /* no need for redaction bookmark counter; nulled zbm_redaction_obj */
 563         /* dsl_bookmark_node_add bumps bookmarks and v2-bookmarks counter */
 564
 565         /*
 566          * write new bookmark
 567          *
 568          * Note that dsl_bookmark_lookup_impl guarantees that, if source is a
 569          * v1 bookmark, the v2-only fields are zeroed.
 570          * And dsl_bookmark_node_add writes back a v1-sized bookmark if
 571          * v2 bookmarks are disabled and/or v2-only fields are zeroed.
 572          * => bookmark copying works on pre-bookmark-v2 pools
 573          */
 574         dsl_bookmark_node_add(bmark_fs_new, new_dbn, tx);
 575
 576         spa_history_log_internal_ds(bmark_fs_source, "bookmark", tx,
 577             "name=%s creation_txg=%llu source_guid=%llu",
 578             new_shortname, (longlong_t)new_dbn->dbn_phys.zbm_creation_txg,
 579             (longlong_t)source_phys.zbm_guid);
 580
 581         dsl_dataset_rele(bmark_fs_source, FTAG);
 582         dsl_dataset_rele(bmark_fs_new, FTAG);
 583 }
 584
 585 void
 586 dsl_bookmark_create_sync(void *arg, dmu_tx_t *tx)
 587 {
 588         dsl_bookmark_create_arg_t *dbca = arg;
 589
 590         ASSERT(spa_feature_is_enabled(dmu_tx_pool(tx)->dp_spa,
 591             SPA_FEATURE_BOOKMARKS));
 592
 593         for (nvpair_t *pair = nvlist_next_nvpair(dbca->dbca_bmarks, NULL);
 594             pair != NULL; pair = nvlist_next_nvpair(dbca->dbca_bmarks, pair)) {
 595
 596                 char *new = nvpair_name(pair);
 597                 char *source = fnvpair_value_string(pair);
 598
 599                 if (strchr(source, '@') != NULL) {
 600                         dsl_bookmark_create_sync_impl_snap(new, source, tx,
 601                             0, NULL, NULL, NULL);
 602                 } else if (strchr(source, '#') != NULL) {
 603                         dsl_bookmark_create_sync_impl_book(new, source, tx);
 604                 } else {
 605                         panic("unreachable code");
 606                 }
 607
 608         }
 609 }
 610
 611 /*
 612  * The bookmarks must all be in the same pool.
 613  */
 614 int
 615 dsl_bookmark_create(nvlist_t *bmarks, nvlist_t *errors)
 616 {
 617         nvpair_t *pair;
 618         dsl_bookmark_create_arg_t dbca;
 619
 620         pair = nvlist_next_nvpair(bmarks, NULL);
 621         if (pair == NULL)
 622                 return (0);
 623
 624         dbca.dbca_bmarks = bmarks;
 625         dbca.dbca_errors = errors;
 626
 627         return (dsl_sync_task(nvpair_name(pair), dsl_bookmark_create_check,
 628             dsl_bookmark_create_sync, &dbca,
 629             fnvlist_num_pairs(bmarks), ZFS_SPACE_CHECK_NORMAL));
 630 }
 631
 632 static int
 633 dsl_bookmark_create_redacted_check(void *arg, dmu_tx_t *tx)
 634 {
 635         dsl_bookmark_create_redacted_arg_t *dbcra = arg;
 636         dsl_pool_t *dp = dmu_tx_pool(tx);
 637         int rv = 0;
 638
 639         if (!spa_feature_is_enabled(dp->dp_spa,
 640             SPA_FEATURE_REDACTION_BOOKMARKS))
 641                 return (SET_ERROR(ENOTSUP));
 642         /*
 643          * If the list of redact snaps will not fit in the bonus buffer with
 644          * the furthest reached object and offset, fail.
 645          */
 646         if (dbcra->dbcra_numsnaps > (dmu_bonus_max() -
 647             sizeof (redaction_list_phys_t)) / sizeof (uint64_t))
 648                 return (SET_ERROR(E2BIG));
 649
 650         if (dsl_bookmark_create_nvl_validate_pair(
 651             dbcra->dbcra_bmark, dbcra->dbcra_snap) != 0)
 652                 return (SET_ERROR(EINVAL));
 653
 654         rv = dsl_bookmark_create_check_impl(dp,
 655             dbcra->dbcra_bmark, dbcra->dbcra_snap);
 656         return (rv);
 657 }
 658
 659 static void
 660 dsl_bookmark_create_redacted_sync(void *arg, dmu_tx_t *tx)
 661 {
 662         dsl_bookmark_create_redacted_arg_t *dbcra = arg;
 663         dsl_bookmark_create_sync_impl_snap(dbcra->dbcra_bmark,
 664             dbcra->dbcra_snap, tx, dbcra->dbcra_numsnaps, dbcra->dbcra_snaps,
 665             dbcra->dbcra_tag, dbcra->dbcra_rl);
 666 }
 667
 668 int
 669 dsl_bookmark_create_redacted(const char *bookmark, const char *snapshot,
 670     uint64_t numsnaps, uint64_t *snapguids, void *tag, redaction_list_t **rl)
 671 {
 672         dsl_bookmark_create_redacted_arg_t dbcra;
 673
 674         dbcra.dbcra_bmark = bookmark;
 675         dbcra.dbcra_snap = snapshot;
 676         dbcra.dbcra_rl = rl;
 677         dbcra.dbcra_numsnaps = numsnaps;
 678         dbcra.dbcra_snaps = snapguids;
 679         dbcra.dbcra_tag = tag;
 680
 681         return (dsl_sync_task(bookmark, dsl_bookmark_create_redacted_check,
 682             dsl_bookmark_create_redacted_sync, &dbcra, 5,
 683             ZFS_SPACE_CHECK_NORMAL));
 684 }
 685
 686 /*
 687  * Retrieve the list of properties given in the 'props' nvlist for a bookmark.
 688  * If 'props' is NULL, retrieves all properties.
 689  */
 690 static void
 691 dsl_bookmark_fetch_props(dsl_pool_t *dp, zfs_bookmark_phys_t *bmark_phys,
 692     nvlist_t *props, nvlist_t *out_props)
 693 {
 694         ASSERT3P(dp, !=, NULL);
 695         ASSERT3P(bmark_phys, !=, NULL);
 696         ASSERT3P(out_props, !=, NULL);
 697         ASSERT(RRW_LOCK_HELD(&dp->dp_config_rwlock));
 698
 699         if (props == NULL || nvlist_exists(props,
 700             zfs_prop_to_name(ZFS_PROP_GUID))) {
 701                 dsl_prop_nvlist_add_uint64(out_props,
 702                     ZFS_PROP_GUID, bmark_phys->zbm_guid);
 703         }
 704         if (props == NULL || nvlist_exists(props,
 705             zfs_prop_to_name(ZFS_PROP_CREATETXG))) {
 706                 dsl_prop_nvlist_add_uint64(out_props,
 707                     ZFS_PROP_CREATETXG, bmark_phys->zbm_creation_txg);
 708         }
 709         if (props == NULL || nvlist_exists(props,
 710             zfs_prop_to_name(ZFS_PROP_CREATION))) {
 711                 dsl_prop_nvlist_add_uint64(out_props,
 712                     ZFS_PROP_CREATION, bmark_phys->zbm_creation_time);
 713         }
 714         if (props == NULL || nvlist_exists(props,
 715             zfs_prop_to_name(ZFS_PROP_IVSET_GUID))) {
 716                 dsl_prop_nvlist_add_uint64(out_props,
 717                     ZFS_PROP_IVSET_GUID, bmark_phys->zbm_ivset_guid);
 718         }
 719         if (bmark_phys->zbm_flags & ZBM_FLAG_HAS_FBN) {
 720                 if (props == NULL || nvlist_exists(props,
 721                     zfs_prop_to_name(ZFS_PROP_REFERENCED))) {
 722                         dsl_prop_nvlist_add_uint64(out_props,
 723                             ZFS_PROP_REFERENCED,
 724                             bmark_phys->zbm_referenced_bytes_refd);
 725                 }
 726                 if (props == NULL || nvlist_exists(props,
 727                     zfs_prop_to_name(ZFS_PROP_LOGICALREFERENCED))) {
 728                         dsl_prop_nvlist_add_uint64(out_props,
 729                             ZFS_PROP_LOGICALREFERENCED,
 730                             bmark_phys->zbm_uncompressed_bytes_refd);
 731                 }
 732                 if (props == NULL || nvlist_exists(props,
 733                     zfs_prop_to_name(ZFS_PROP_REFRATIO))) {
 734                         uint64_t ratio =
 735                             bmark_phys->zbm_compressed_bytes_refd == 0 ? 100 :
 736                             bmark_phys->zbm_uncompressed_bytes_refd * 100 /
 737                             bmark_phys->zbm_compressed_bytes_refd;
 738                         dsl_prop_nvlist_add_uint64(out_props,
 739                             ZFS_PROP_REFRATIO, ratio);
 740                 }
 741         }
 742
 743         if ((props == NULL || nvlist_exists(props, "redact_snaps") ||
 744             nvlist_exists(props, "redact_complete")) &&
 745             bmark_phys->zbm_redaction_obj != 0) {
 746                 redaction_list_t *rl;
 747                 int err = dsl_redaction_list_hold_obj(dp,
 748                     bmark_phys->zbm_redaction_obj, FTAG, &rl);
 749                 if (err == 0) {
 750                         if (nvlist_exists(props, "redact_snaps")) {
 751                                 nvlist_t *nvl;
 752                                 nvl = fnvlist_alloc();
 753                                 fnvlist_add_uint64_array(nvl, ZPROP_VALUE,
 754                                     rl->rl_phys->rlp_snaps,
 755                                     rl->rl_phys->rlp_num_snaps);
 756                                 fnvlist_add_nvlist(out_props, "redact_snaps",
 757                                     nvl);
 758                                 nvlist_free(nvl);
 759                         }
 760                         if (nvlist_exists(props, "redact_complete")) {
 761                                 nvlist_t *nvl;
 762                                 nvl = fnvlist_alloc();
 763                                 fnvlist_add_boolean_value(nvl, ZPROP_VALUE,
 764                                     rl->rl_phys->rlp_last_blkid == UINT64_MAX &&
 765                                     rl->rl_phys->rlp_last_object == UINT64_MAX);
 766                                 fnvlist_add_nvlist(out_props, "redact_complete",
 767                                     nvl);
 768                                 nvlist_free(nvl);
 769                         }
 770                         dsl_redaction_list_rele(rl, FTAG);
 771                 }
 772         }
 773 }
 774
 775 int
 776 dsl_get_bookmarks_impl(dsl_dataset_t *ds, nvlist_t *props, nvlist_t *outnvl)
 777 {
 778         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 779
 780         ASSERT(dsl_pool_config_held(dp));
 781
 782         if (dsl_dataset_is_snapshot(ds))
 783                 return (SET_ERROR(EINVAL));
 784
 785         for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
 786             dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
 787                 nvlist_t *out_props = fnvlist_alloc();
 788
 789                 dsl_bookmark_fetch_props(dp, &dbn->dbn_phys, props, out_props);
 790
 791                 fnvlist_add_nvlist(outnvl, dbn->dbn_name, out_props);
 792                 fnvlist_free(out_props);
 793         }
 794         return (0);
 795 }
 796
 797 /*
 798  * Comparison func for ds_bookmarks AVL tree.  We sort the bookmarks by
 799  * their TXG, then by their FBN-ness.  The "FBN-ness" component ensures
 800  * that all bookmarks at the same TXG that HAS_FBN are adjacent, which
 801  * dsl_bookmark_destroy_sync_impl() depends on.  Note that there may be
 802  * multiple bookmarks at the same TXG (with the same FBN-ness).  In this
 803  * case we differentiate them by an arbitrary metric (in this case,
 804  * their names).
 805  */
 806 static int
 807 dsl_bookmark_compare(const void *l, const void *r)
 808 {
 809         const dsl_bookmark_node_t *ldbn = l;
 810         const dsl_bookmark_node_t *rdbn = r;
 811
 812         int64_t cmp = TREE_CMP(ldbn->dbn_phys.zbm_creation_txg,
 813             rdbn->dbn_phys.zbm_creation_txg);
 814         if (likely(cmp))
 815                 return (cmp);
 816         cmp = TREE_CMP((ldbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN),
 817             (rdbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
 818         if (likely(cmp))
 819                 return (cmp);
 820         cmp = strcmp(ldbn->dbn_name, rdbn->dbn_name);
 821         return (TREE_ISIGN(cmp));
 822 }
 823
 824 /*
 825  * Cache this (head) dataset's bookmarks in the ds_bookmarks AVL tree.
 826  */
 827 int
 828 dsl_bookmark_init_ds(dsl_dataset_t *ds)
 829 {
 830         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 831         objset_t *mos = dp->dp_meta_objset;
 832
 833         ASSERT(!ds->ds_is_snapshot);
 834
 835         avl_create(&ds->ds_bookmarks, dsl_bookmark_compare,
 836             sizeof (dsl_bookmark_node_t),
 837             offsetof(dsl_bookmark_node_t, dbn_node));
 838
 839         if (!dsl_dataset_is_zapified(ds))
 840                 return (0);
 841
 842         int zaperr = zap_lookup(mos, ds->ds_object, DS_FIELD_BOOKMARK_NAMES,
 843             sizeof (ds->ds_bookmarks_obj), 1, &ds->ds_bookmarks_obj);
 844         if (zaperr == ENOENT)
 845                 return (0);
 846         if (zaperr != 0)
 847                 return (zaperr);
 848
 849         if (ds->ds_bookmarks_obj == 0)
 850                 return (0);
 851
 852         int err = 0;
 853         zap_cursor_t zc;
 854         zap_attribute_t attr;
 855
 856         for (zap_cursor_init(&zc, mos, ds->ds_bookmarks_obj);
 857             (err = zap_cursor_retrieve(&zc, &attr)) == 0;
 858             zap_cursor_advance(&zc)) {
 859                 dsl_bookmark_node_t *dbn =
 860                     dsl_bookmark_node_alloc(attr.za_name);
 861
 862                 err = dsl_bookmark_lookup_impl(ds,
 863                     dbn->dbn_name, &dbn->dbn_phys);
 864                 ASSERT3U(err, !=, ENOENT);
 865                 if (err != 0) {
 866                         kmem_free(dbn, sizeof (*dbn));
 867                         break;
 868                 }
 869                 avl_add(&ds->ds_bookmarks, dbn);
 870         }
 871         zap_cursor_fini(&zc);
 872         if (err == ENOENT)
 873                 err = 0;
 874         return (err);
 875 }
 876
 877 void
 878 dsl_bookmark_fini_ds(dsl_dataset_t *ds)
 879 {
 880         void *cookie = NULL;
 881         dsl_bookmark_node_t *dbn;
 882
 883         if (ds->ds_is_snapshot)
 884                 return;
 885
 886         while ((dbn = avl_destroy_nodes(&ds->ds_bookmarks, &cookie)) != NULL) {
 887                 spa_strfree(dbn->dbn_name);
 888                 mutex_destroy(&dbn->dbn_lock);
 889                 kmem_free(dbn, sizeof (*dbn));
 890         }
 891         avl_destroy(&ds->ds_bookmarks);
 892 }
 893
 894 /*
 895  * Retrieve the bookmarks that exist in the specified dataset, and the
 896  * requested properties of each bookmark.
 897  *
 898  * The "props" nvlist specifies which properties are requested.
 899  * See lzc_get_bookmarks() for the list of valid properties.
 900  */
 901 int
 902 dsl_get_bookmarks(const char *dsname, nvlist_t *props, nvlist_t *outnvl)
 903 {
 904         dsl_pool_t *dp;
 905         dsl_dataset_t *ds;
 906         int err;
 907
 908         err = dsl_pool_hold(dsname, FTAG, &dp);
 909         if (err != 0)
 910                 return (err);
 911         err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
 912         if (err != 0) {
 913                 dsl_pool_rele(dp, FTAG);
 914                 return (err);
 915         }
 916
 917         err = dsl_get_bookmarks_impl(ds, props, outnvl);
 918
 919         dsl_dataset_rele(ds, FTAG);
 920         dsl_pool_rele(dp, FTAG);
 921         return (err);
 922 }
 923
 924 /*
 925  * Retrieve all properties for a single bookmark in the given dataset.
 926  */
 927 int
 928 dsl_get_bookmark_props(const char *dsname, const char *bmname, nvlist_t *props)
 929 {
 930         dsl_pool_t *dp;
 931         dsl_dataset_t *ds;
 932         zfs_bookmark_phys_t bmark_phys = { 0 };
 933         int err;
 934
 935         err = dsl_pool_hold(dsname, FTAG, &dp);
 936         if (err != 0)
 937                 return (err);
 938         err = dsl_dataset_hold(dp, dsname, FTAG, &ds);
 939         if (err != 0) {
 940                 dsl_pool_rele(dp, FTAG);
 941                 return (err);
 942         }
 943
 944         err = dsl_bookmark_lookup_impl(ds, bmname, &bmark_phys);
 945         if (err != 0)
 946                 goto out;
 947
 948         dsl_bookmark_fetch_props(dp, &bmark_phys, NULL, props);
 949 out:
 950         dsl_dataset_rele(ds, FTAG);
 951         dsl_pool_rele(dp, FTAG);
 952         return (err);
 953 }
 954
 955 typedef struct dsl_bookmark_destroy_arg {
 956         nvlist_t *dbda_bmarks;
 957         nvlist_t *dbda_success;
 958         nvlist_t *dbda_errors;
 959 } dsl_bookmark_destroy_arg_t;
 960
 961 static void
 962 dsl_bookmark_destroy_sync_impl(dsl_dataset_t *ds, const char *name,
 963     dmu_tx_t *tx)
 964 {
 965         objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 966         uint64_t bmark_zapobj = ds->ds_bookmarks_obj;
 967         matchtype_t mt = 0;
 968         uint64_t int_size, num_ints;
 969         /*
 970          * 'search' must be zeroed so that dbn_flags (which is used in
 971          * dsl_bookmark_compare()) will be zeroed even if the on-disk
 972          * (in ZAP) bookmark is shorter than offsetof(dbn_flags).
 973          */
 974         dsl_bookmark_node_t search = { 0 };
 975         char realname[ZFS_MAX_DATASET_NAME_LEN];
 976
 977         /*
 978          * Find the real name of this bookmark, which may be different
 979          * from the given name if the dataset is case-insensitive.  Then
 980          * use the real name to find the node in the ds_bookmarks AVL tree.
 981          */
 982
 983         if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
 984                 mt = MT_NORMALIZE;
 985
 986         VERIFY0(zap_length(mos, bmark_zapobj, name, &int_size, &num_ints));
 987
 988         ASSERT3U(int_size, ==, sizeof (uint64_t));
 989
 990         if (num_ints * int_size > BOOKMARK_PHYS_SIZE_V1) {
 991                 spa_feature_decr(dmu_objset_spa(mos),
 992                     SPA_FEATURE_BOOKMARK_V2, tx);
 993         }
 994         VERIFY0(zap_lookup_norm(mos, bmark_zapobj, name, sizeof (uint64_t),
 995             num_ints, &search.dbn_phys, mt, realname, sizeof (realname), NULL));
 996
 997         search.dbn_name = realname;
 998         dsl_bookmark_node_t *dbn = avl_find(&ds->ds_bookmarks, &search, NULL);
 999         ASSERT(dbn != NULL);
1000
1001         if (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) {
1002                 /*
1003                  * If this bookmark HAS_FBN, and it is before the most
1004                  * recent snapshot, then its TXG is a key in the head's
1005                  * deadlist (and all clones' heads' deadlists).  If this is
1006                  * the last thing keeping the key (i.e. there are no more
1007                  * bookmarks with HAS_FBN at this TXG, and there is no
1008                  * snapshot at this TXG), then remove the key.
1009                  *
1010                  * Note that this algorithm depends on ds_bookmarks being
1011                  * sorted such that all bookmarks at the same TXG with
1012                  * HAS_FBN are adjacent (with no non-HAS_FBN bookmarks
1013                  * at the same TXG in between them).  If this were not
1014                  * the case, we would need to examine *all* bookmarks
1015                  * at this TXG, rather than just the adjacent ones.
1016                  */
1017
1018                 dsl_bookmark_node_t *dbn_prev =
1019                     AVL_PREV(&ds->ds_bookmarks, dbn);
1020                 dsl_bookmark_node_t *dbn_next =
1021                     AVL_NEXT(&ds->ds_bookmarks, dbn);
1022
1023                 boolean_t more_bookmarks_at_this_txg =
1024                     (dbn_prev != NULL && dbn_prev->dbn_phys.zbm_creation_txg ==
1025                     dbn->dbn_phys.zbm_creation_txg &&
1026                     (dbn_prev->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) ||
1027                     (dbn_next != NULL && dbn_next->dbn_phys.zbm_creation_txg ==
1028                     dbn->dbn_phys.zbm_creation_txg &&
1029                     (dbn_next->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN));
1030
1031                 if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS) &&
1032                     !more_bookmarks_at_this_txg &&
1033                     dbn->dbn_phys.zbm_creation_txg <
1034                     dsl_dataset_phys(ds)->ds_prev_snap_txg) {
1035                         dsl_dir_remove_clones_key(ds->ds_dir,
1036                             dbn->dbn_phys.zbm_creation_txg, tx);
1037                         dsl_deadlist_remove_key(&ds->ds_deadlist,
1038                             dbn->dbn_phys.zbm_creation_txg, tx);
1039                 }
1040
1041                 spa_feature_decr(dmu_objset_spa(mos),
1042                     SPA_FEATURE_BOOKMARK_WRITTEN, tx);
1043         }
1044
1045         if (dbn->dbn_phys.zbm_redaction_obj != 0) {
1046                 VERIFY0(dmu_object_free(mos,
1047                     dbn->dbn_phys.zbm_redaction_obj, tx));
1048                 spa_feature_decr(dmu_objset_spa(mos),
1049                     SPA_FEATURE_REDACTION_BOOKMARKS, tx);
1050         }
1051
1052         avl_remove(&ds->ds_bookmarks, dbn);
1053         spa_strfree(dbn->dbn_name);
1054         mutex_destroy(&dbn->dbn_lock);
1055         kmem_free(dbn, sizeof (*dbn));
1056
1057         VERIFY0(zap_remove_norm(mos, bmark_zapobj, name, mt, tx));
1058 }
1059
1060 static int
1061 dsl_bookmark_destroy_check(void *arg, dmu_tx_t *tx)
1062 {
1063         dsl_bookmark_destroy_arg_t *dbda = arg;
1064         dsl_pool_t *dp = dmu_tx_pool(tx);
1065         int rv = 0;
1066
1067         ASSERT(nvlist_empty(dbda->dbda_success));
1068         ASSERT(nvlist_empty(dbda->dbda_errors));
1069
1070         if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_BOOKMARKS))
1071                 return (0);
1072
1073         for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_bmarks, NULL);
1074             pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_bmarks, pair)) {
1075                 const char *fullname = nvpair_name(pair);
1076                 dsl_dataset_t *ds;
1077                 zfs_bookmark_phys_t bm;
1078                 int error;
1079                 char *shortname;
1080
1081                 error = dsl_bookmark_hold_ds(dp, fullname, &ds,
1082                     FTAG, &shortname);
1083                 if (error == ENOENT) {
1084                         /* ignore it; the bookmark is "already destroyed" */
1085                         continue;
1086                 }
1087                 if (error == 0) {
1088                         error = dsl_bookmark_lookup_impl(ds, shortname, &bm);
1089                         dsl_dataset_rele(ds, FTAG);
1090                         if (error == ESRCH) {
1091                                 /*
1092                                  * ignore it; the bookmark is
1093                                  * "already destroyed"
1094                                  */
1095                                 continue;
1096                         }
1097                         if (error == 0 && bm.zbm_redaction_obj != 0) {
1098                                 redaction_list_t *rl = NULL;
1099                                 error = dsl_redaction_list_hold_obj(tx->tx_pool,
1100                                     bm.zbm_redaction_obj, FTAG, &rl);
1101                                 if (error == ENOENT) {
1102                                         error = 0;
1103                                 } else if (error == 0 &&
1104                                     dsl_redaction_list_long_held(rl)) {
1105                                         error = SET_ERROR(EBUSY);
1106                                 }
1107                                 if (rl != NULL) {
1108                                         dsl_redaction_list_rele(rl, FTAG);
1109                                 }
1110                         }
1111                 }
1112                 if (error == 0) {
1113                         if (dmu_tx_is_syncing(tx)) {
1114                                 fnvlist_add_boolean(dbda->dbda_success,
1115                                     fullname);
1116                         }
1117                 } else {
1118                         fnvlist_add_int32(dbda->dbda_errors, fullname, error);
1119                         rv = error;
1120                 }
1121         }
1122         return (rv);
1123 }
1124
1125 static void
1126 dsl_bookmark_destroy_sync(void *arg, dmu_tx_t *tx)
1127 {
1128         dsl_bookmark_destroy_arg_t *dbda = arg;
1129         dsl_pool_t *dp = dmu_tx_pool(tx);
1130         objset_t *mos = dp->dp_meta_objset;
1131
1132         for (nvpair_t *pair = nvlist_next_nvpair(dbda->dbda_success, NULL);
1133             pair != NULL; pair = nvlist_next_nvpair(dbda->dbda_success, pair)) {
1134                 dsl_dataset_t *ds;
1135                 char *shortname;
1136                 uint64_t zap_cnt;
1137
1138                 VERIFY0(dsl_bookmark_hold_ds(dp, nvpair_name(pair),
1139                     &ds, FTAG, &shortname));
1140                 dsl_bookmark_destroy_sync_impl(ds, shortname, tx);
1141
1142                 /*
1143                  * If all of this dataset's bookmarks have been destroyed,
1144                  * free the zap object and decrement the feature's use count.
1145                  */
1146                 VERIFY0(zap_count(mos, ds->ds_bookmarks_obj, &zap_cnt));
1147                 if (zap_cnt == 0) {
1148                         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1149                         VERIFY0(zap_destroy(mos, ds->ds_bookmarks_obj, tx));
1150                         ds->ds_bookmarks_obj = 0;
1151                         spa_feature_decr(dp->dp_spa, SPA_FEATURE_BOOKMARKS, tx);
1152                         VERIFY0(zap_remove(mos, ds->ds_object,
1153                             DS_FIELD_BOOKMARK_NAMES, tx));
1154                 }
1155
1156                 spa_history_log_internal_ds(ds, "remove bookmark", tx,
1157                     "name=%s", shortname);
1158
1159                 dsl_dataset_rele(ds, FTAG);
1160         }
1161 }
1162
1163 /*
1164  * The bookmarks must all be in the same pool.
1165  */
1166 int
1167 dsl_bookmark_destroy(nvlist_t *bmarks, nvlist_t *errors)
1168 {
1169         int rv;
1170         dsl_bookmark_destroy_arg_t dbda;
1171         nvpair_t *pair = nvlist_next_nvpair(bmarks, NULL);
1172         if (pair == NULL)
1173                 return (0);
1174
1175         dbda.dbda_bmarks = bmarks;
1176         dbda.dbda_errors = errors;
1177         dbda.dbda_success = fnvlist_alloc();
1178
1179         rv = dsl_sync_task(nvpair_name(pair), dsl_bookmark_destroy_check,
1180             dsl_bookmark_destroy_sync, &dbda, fnvlist_num_pairs(bmarks),
1181             ZFS_SPACE_CHECK_RESERVED);
1182         fnvlist_free(dbda.dbda_success);
1183         return (rv);
1184 }
1185
1186 /* Return B_TRUE if there are any long holds on this dataset. */
1187 boolean_t
1188 dsl_redaction_list_long_held(redaction_list_t *rl)
1189 {
1190         return (!zfs_refcount_is_zero(&rl->rl_longholds));
1191 }
1192
1193 void
1194 dsl_redaction_list_long_hold(dsl_pool_t *dp, redaction_list_t *rl, void *tag)
1195 {
1196         ASSERT(dsl_pool_config_held(dp));
1197         (void) zfs_refcount_add(&rl->rl_longholds, tag);
1198 }
1199
1200 void
1201 dsl_redaction_list_long_rele(redaction_list_t *rl, void *tag)
1202 {
1203         (void) zfs_refcount_remove(&rl->rl_longholds, tag);
1204 }
1205
1206 /* ARGSUSED */
1207 static void
1208 redaction_list_evict_sync(void *rlu)
1209 {
1210         redaction_list_t *rl = rlu;
1211         zfs_refcount_destroy(&rl->rl_longholds);
1212
1213         kmem_free(rl, sizeof (redaction_list_t));
1214 }
1215
1216 void
1217 dsl_redaction_list_rele(redaction_list_t *rl, void *tag)
1218 {
1219         dmu_buf_rele(rl->rl_dbuf, tag);
1220 }
1221
1222 int
1223 dsl_redaction_list_hold_obj(dsl_pool_t *dp, uint64_t rlobj, void *tag,
1224     redaction_list_t **rlp)
1225 {
1226         objset_t *mos = dp->dp_meta_objset;
1227         dmu_buf_t *dbuf;
1228         redaction_list_t *rl;
1229         int err;
1230
1231         ASSERT(dsl_pool_config_held(dp));
1232
1233         err = dmu_bonus_hold(mos, rlobj, tag, &dbuf);
1234         if (err != 0)
1235                 return (err);
1236
1237         rl = dmu_buf_get_user(dbuf);
1238         if (rl == NULL) {
1239                 redaction_list_t *winner = NULL;
1240
1241                 rl = kmem_zalloc(sizeof (redaction_list_t), KM_SLEEP);
1242                 rl->rl_dbuf = dbuf;
1243                 rl->rl_object = rlobj;
1244                 rl->rl_phys = dbuf->db_data;
1245                 rl->rl_mos = dp->dp_meta_objset;
1246                 zfs_refcount_create(&rl->rl_longholds);
1247                 dmu_buf_init_user(&rl->rl_dbu, redaction_list_evict_sync, NULL,
1248                     &rl->rl_dbuf);
1249                 if ((winner = dmu_buf_set_user_ie(dbuf, &rl->rl_dbu)) != NULL) {
1250                         kmem_free(rl, sizeof (*rl));
1251                         rl = winner;
1252                 }
1253         }
1254         *rlp = rl;
1255         return (0);
1256 }
1257
1258 /*
1259  * Snapshot ds is being destroyed.
1260  *
1261  * Adjust the "freed_before_next" of any bookmarks between this snap
1262  * and the previous snapshot, because their "next snapshot" is changing.
1263  *
1264  * If there are any bookmarks with HAS_FBN at this snapshot, remove
1265  * their HAS_SNAP flag (note: there can be at most one snapshot of
1266  * each filesystem at a given txg), and return B_TRUE.  In this case
1267  * the caller can not remove the key in the deadlist at this TXG, because
1268  * the HAS_FBN bookmarks require the key be there.
1269  *
1270  * Returns B_FALSE if there are no bookmarks with HAS_FBN at this
1271  * snapshot's TXG.  In this case the caller can remove the key in the
1272  * deadlist at this TXG.
1273  */
1274 boolean_t
1275 dsl_bookmark_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
1276 {
1277         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1278
1279         dsl_dataset_t *head, *next;
1280         VERIFY0(dsl_dataset_hold_obj(dp,
1281             dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj, FTAG, &head));
1282         VERIFY0(dsl_dataset_hold_obj(dp,
1283             dsl_dataset_phys(ds)->ds_next_snap_obj, FTAG, &next));
1284
1285         /*
1286          * Find the first bookmark that HAS_FBN at or after the
1287          * previous snapshot.
1288          */
1289         dsl_bookmark_node_t search = { 0 };
1290         avl_index_t idx;
1291         search.dbn_phys.zbm_creation_txg =
1292             dsl_dataset_phys(ds)->ds_prev_snap_txg;
1293         search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1294         /*
1295          * The empty-string name can't be in the AVL, and it compares
1296          * before any entries with this TXG.
1297          */
1298         search.dbn_name = "";
1299         VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1300         dsl_bookmark_node_t *dbn =
1301             avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1302
1303         /*
1304          * Iterate over all bookmarks that are at or after the previous
1305          * snapshot, and before this (being deleted) snapshot.  Adjust
1306          * their FBN based on their new next snapshot.
1307          */
1308         for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg <
1309             dsl_dataset_phys(ds)->ds_creation_txg;
1310             dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1311                 if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN))
1312                         continue;
1313                 /*
1314                  * Increase our FBN by the amount of space that was live
1315                  * (referenced) at the time of this bookmark (i.e.
1316                  * birth <= zbm_creation_txg), and killed between this
1317                  * (being deleted) snapshot and the next snapshot (i.e.
1318                  * on the next snapshot's deadlist).  (Space killed before
1319                  * this are already on our FBN.)
1320                  */
1321                 uint64_t referenced, compressed, uncompressed;
1322                 dsl_deadlist_space_range(&next->ds_deadlist,
1323                     0, dbn->dbn_phys.zbm_creation_txg,
1324                     &referenced, &compressed, &uncompressed);
1325                 dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1326                     referenced;
1327                 dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1328                     compressed;
1329                 dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1330                     uncompressed;
1331                 VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1332                     dbn->dbn_name, sizeof (uint64_t),
1333                     sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1334                     &dbn->dbn_phys, tx));
1335         }
1336         dsl_dataset_rele(next, FTAG);
1337
1338         /*
1339          * There may be several bookmarks at this txg (the TXG of the
1340          * snapshot being deleted).  We need to clear the SNAPSHOT_EXISTS
1341          * flag on all of them, and return TRUE if there is at least 1
1342          * bookmark here with HAS_FBN (thus preventing the deadlist
1343          * key from being removed).
1344          */
1345         boolean_t rv = B_FALSE;
1346         for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1347             dsl_dataset_phys(ds)->ds_creation_txg;
1348             dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1349                 if (!(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1350                         ASSERT(!(dbn->dbn_phys.zbm_flags &
1351                             ZBM_FLAG_SNAPSHOT_EXISTS));
1352                         continue;
1353                 }
1354                 ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_SNAPSHOT_EXISTS);
1355                 dbn->dbn_phys.zbm_flags &= ~ZBM_FLAG_SNAPSHOT_EXISTS;
1356                 VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1357                     dbn->dbn_name, sizeof (uint64_t),
1358                     sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1359                     &dbn->dbn_phys, tx));
1360                 rv = B_TRUE;
1361         }
1362         dsl_dataset_rele(head, FTAG);
1363         return (rv);
1364 }
1365
1366 /*
1367  * A snapshot is being created of this (head) dataset.
1368  *
1369  * We don't keep keys in the deadlist for the most recent snapshot, or any
1370  * bookmarks at or after it, because there can't be any blocks on the
1371  * deadlist in this range.  Now that the most recent snapshot is after
1372  * all bookmarks, we need to add these keys.  Note that the caller always
1373  * adds a key at the previous snapshot, so we only add keys for bookmarks
1374  * after that.
1375  */
1376 void
1377 dsl_bookmark_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
1378 {
1379         uint64_t last_key_added = UINT64_MAX;
1380         for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1381             dbn != NULL && dbn->dbn_phys.zbm_creation_txg >
1382             dsl_dataset_phys(ds)->ds_prev_snap_txg;
1383             dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1384                 uint64_t creation_txg = dbn->dbn_phys.zbm_creation_txg;
1385                 ASSERT3U(creation_txg, <=, last_key_added);
1386                 /*
1387                  * Note, there may be multiple bookmarks at this TXG,
1388                  * and we only want to add the key for this TXG once.
1389                  * The ds_bookmarks AVL is sorted by TXG, so we will visit
1390                  * these bookmarks in sequence.
1391                  */
1392                 if ((dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN) &&
1393                     creation_txg != last_key_added) {
1394                         dsl_deadlist_add_key(&ds->ds_deadlist,
1395                             creation_txg, tx);
1396                         last_key_added = creation_txg;
1397                 }
1398         }
1399 }
1400
1401 /*
1402  * The next snapshot of the origin dataset has changed, due to
1403  * promote or clone swap.  If there are any bookmarks at this dataset,
1404  * we need to update their zbm_*_freed_before_next_snap to reflect this.
1405  * The head dataset has the relevant bookmarks in ds_bookmarks.
1406  */
1407 void
1408 dsl_bookmark_next_changed(dsl_dataset_t *head, dsl_dataset_t *origin,
1409     dmu_tx_t *tx)
1410 {
1411         dsl_pool_t *dp = dmu_tx_pool(tx);
1412
1413         /*
1414          * Find the first bookmark that HAS_FBN at the origin snapshot.
1415          */
1416         dsl_bookmark_node_t search = { 0 };
1417         avl_index_t idx;
1418         search.dbn_phys.zbm_creation_txg =
1419             dsl_dataset_phys(origin)->ds_creation_txg;
1420         search.dbn_phys.zbm_flags = ZBM_FLAG_HAS_FBN;
1421         /*
1422          * The empty-string name can't be in the AVL, and it compares
1423          * before any entries with this TXG.
1424          */
1425         search.dbn_name = "";
1426         VERIFY3P(avl_find(&head->ds_bookmarks, &search, &idx), ==, NULL);
1427         dsl_bookmark_node_t *dbn =
1428             avl_nearest(&head->ds_bookmarks, idx, AVL_AFTER);
1429
1430         /*
1431          * Iterate over all bookmarks that are at the origin txg.
1432          * Adjust their FBN based on their new next snapshot.
1433          */
1434         for (; dbn != NULL && dbn->dbn_phys.zbm_creation_txg ==
1435             dsl_dataset_phys(origin)->ds_creation_txg &&
1436             (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1437             dbn = AVL_NEXT(&head->ds_bookmarks, dbn)) {
1438
1439                 /*
1440                  * Bookmark is at the origin, therefore its
1441                  * "next dataset" is changing, so we need
1442                  * to reset its FBN by recomputing it in
1443                  * dsl_bookmark_set_phys().
1444                  */
1445                 ASSERT3U(dbn->dbn_phys.zbm_guid, ==,
1446                     dsl_dataset_phys(origin)->ds_guid);
1447                 ASSERT3U(dbn->dbn_phys.zbm_referenced_bytes_refd, ==,
1448                     dsl_dataset_phys(origin)->ds_referenced_bytes);
1449                 ASSERT(dbn->dbn_phys.zbm_flags &
1450                     ZBM_FLAG_SNAPSHOT_EXISTS);
1451                 /*
1452                  * Save and restore the zbm_redaction_obj, which
1453                  * is zeroed by dsl_bookmark_set_phys().
1454                  */
1455                 uint64_t redaction_obj =
1456                     dbn->dbn_phys.zbm_redaction_obj;
1457                 dsl_bookmark_set_phys(&dbn->dbn_phys, origin);
1458                 dbn->dbn_phys.zbm_redaction_obj = redaction_obj;
1459
1460                 VERIFY0(zap_update(dp->dp_meta_objset, head->ds_bookmarks_obj,
1461                     dbn->dbn_name, sizeof (uint64_t),
1462                     sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1463                     &dbn->dbn_phys, tx));
1464         }
1465 }
1466
1467 /*
1468  * This block is no longer referenced by this (head) dataset.
1469  *
1470  * Adjust the FBN of any bookmarks that reference this block, whose "next"
1471  * is the head dataset.
1472  */
1473 /* ARGSUSED */
1474 void
1475 dsl_bookmark_block_killed(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
1476 {
1477         /*
1478          * Iterate over bookmarks whose "next" is the head dataset.
1479          */
1480         for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1481             dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1482             dsl_dataset_phys(ds)->ds_prev_snap_txg;
1483             dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1484                 /*
1485                  * If the block was live (referenced) at the time of this
1486                  * bookmark, add its space to the bookmark's FBN.
1487                  */
1488                 if (bp->blk_birth <= dbn->dbn_phys.zbm_creation_txg &&
1489                     (dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN)) {
1490                         mutex_enter(&dbn->dbn_lock);
1491                         dbn->dbn_phys.zbm_referenced_freed_before_next_snap +=
1492                             bp_get_dsize_sync(dsl_dataset_get_spa(ds), bp);
1493                         dbn->dbn_phys.zbm_compressed_freed_before_next_snap +=
1494                             BP_GET_PSIZE(bp);
1495                         dbn->dbn_phys.zbm_uncompressed_freed_before_next_snap +=
1496                             BP_GET_UCSIZE(bp);
1497                         /*
1498                          * Changing the ZAP object here would be too
1499                          * expensive.  Also, we may be called from the zio
1500                          * interrupt thread, which can't block on i/o.
1501                          * Therefore, we mark this bookmark as dirty and
1502                          * modify the ZAP once per txg, in
1503                          * dsl_bookmark_sync_done().
1504                          */
1505                         dbn->dbn_dirty = B_TRUE;
1506                         mutex_exit(&dbn->dbn_lock);
1507                 }
1508         }
1509 }
1510
1511 void
1512 dsl_bookmark_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
1513 {
1514         dsl_pool_t *dp = dmu_tx_pool(tx);
1515
1516         if (dsl_dataset_is_snapshot(ds))
1517                 return;
1518
1519         /*
1520          * We only dirty bookmarks that are at or after the most recent
1521          * snapshot.  We can't create snapshots between
1522          * dsl_bookmark_block_killed() and dsl_bookmark_sync_done(), so we
1523          * don't need to look at any bookmarks before ds_prev_snap_txg.
1524          */
1525         for (dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1526             dbn != NULL && dbn->dbn_phys.zbm_creation_txg >=
1527             dsl_dataset_phys(ds)->ds_prev_snap_txg;
1528             dbn = AVL_PREV(&ds->ds_bookmarks, dbn)) {
1529                 if (dbn->dbn_dirty) {
1530                         /*
1531                          * We only dirty nodes with HAS_FBN, therefore
1532                          * we can always use the current bookmark struct size.
1533                          */
1534                         ASSERT(dbn->dbn_phys.zbm_flags & ZBM_FLAG_HAS_FBN);
1535                         VERIFY0(zap_update(dp->dp_meta_objset,
1536                             ds->ds_bookmarks_obj,
1537                             dbn->dbn_name, sizeof (uint64_t),
1538                             sizeof (zfs_bookmark_phys_t) / sizeof (uint64_t),
1539                             &dbn->dbn_phys, tx));
1540                         dbn->dbn_dirty = B_FALSE;
1541                 }
1542         }
1543 #ifdef ZFS_DEBUG
1544         for (dsl_bookmark_node_t *dbn = avl_first(&ds->ds_bookmarks);
1545             dbn != NULL; dbn = AVL_NEXT(&ds->ds_bookmarks, dbn)) {
1546                 ASSERT(!dbn->dbn_dirty);
1547         }
1548 #endif
1549 }
1550
1551 /*
1552  * Return the TXG of the most recent bookmark (or 0 if there are no bookmarks).
1553  */
1554 uint64_t
1555 dsl_bookmark_latest_txg(dsl_dataset_t *ds)
1556 {
1557         ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool));
1558         dsl_bookmark_node_t *dbn = avl_last(&ds->ds_bookmarks);
1559         if (dbn == NULL)
1560                 return (0);
1561         return (dbn->dbn_phys.zbm_creation_txg);
1562 }
1563
1564 /*
1565  * Compare the redact_block_phys_t to the bookmark. If the last block in the
1566  * redact_block_phys_t is before the bookmark, return -1.  If the first block in
1567  * the redact_block_phys_t is after the bookmark, return 1.  Otherwise, the
1568  * bookmark is inside the range of the redact_block_phys_t, and we return 0.
1569  */
1570 static int
1571 redact_block_zb_compare(redact_block_phys_t *first,
1572     zbookmark_phys_t *second)
1573 {
1574         /*
1575          * If the block_phys is for a previous object, or the last block in the
1576          * block_phys is strictly before the block in the bookmark, the
1577          * block_phys is earlier.
1578          */
1579         if (first->rbp_object < second->zb_object ||
1580             (first->rbp_object == second->zb_object &&
1581             first->rbp_blkid + (redact_block_get_count(first) - 1) <
1582             second->zb_blkid)) {
1583                 return (-1);
1584         }
1585
1586         /*
1587          * If the bookmark is for a previous object, or the block in the
1588          * bookmark is strictly before the first block in the block_phys, the
1589          * bookmark is earlier.
1590          */
1591         if (first->rbp_object > second->zb_object ||
1592             (first->rbp_object == second->zb_object &&
1593             first->rbp_blkid > second->zb_blkid)) {
1594                 return (1);
1595         }
1596
1597         return (0);
1598 }
1599
1600 /*
1601  * Traverse the redaction list in the provided object, and call the callback for
1602  * each entry we find. Don't call the callback for any records before resume.
1603  */
1604 int
1605 dsl_redaction_list_traverse(redaction_list_t *rl, zbookmark_phys_t *resume,
1606     rl_traverse_callback_t cb, void *arg)
1607 {
1608         objset_t *mos = rl->rl_mos;
1609         int err = 0;
1610
1611         if (rl->rl_phys->rlp_last_object != UINT64_MAX ||
1612             rl->rl_phys->rlp_last_blkid != UINT64_MAX) {
1613                 /*
1614                  * When we finish a send, we update the last object and offset
1615                  * to UINT64_MAX.  If a send fails partway through, the last
1616                  * object and offset will have some other value, indicating how
1617                  * far the send got. The redaction list must be complete before
1618                  * it can be traversed, so return EINVAL if the last object and
1619                  * blkid are not set to UINT64_MAX.
1620                  */
1621                 return (SET_ERROR(EINVAL));
1622         }
1623
1624         /*
1625          * This allows us to skip the binary search and resume checking logic
1626          * below, if we're not resuming a redacted send.
1627          */
1628         if (ZB_IS_ZERO(resume))
1629                 resume = NULL;
1630
1631         /*
1632          * Binary search for the point to resume from.
1633          */
1634         uint64_t maxidx = rl->rl_phys->rlp_num_entries - 1;
1635         uint64_t minidx = 0;
1636         while (resume != NULL && maxidx > minidx) {
1637                 redact_block_phys_t rbp = { 0 };
1638                 ASSERT3U(maxidx, >, minidx);
1639                 uint64_t mididx = minidx + ((maxidx - minidx) / 2);
1640                 err = dmu_read(mos, rl->rl_object, mididx * sizeof (rbp),
1641                     sizeof (rbp), &rbp, DMU_READ_NO_PREFETCH);
1642                 if (err != 0)
1643                         break;
1644
1645                 int cmp = redact_block_zb_compare(&rbp, resume);
1646
1647                 if (cmp == 0) {
1648                         minidx = mididx;
1649                         break;
1650                 } else if (cmp > 0) {
1651                         maxidx =
1652                             (mididx == minidx ? minidx : mididx - 1);
1653                 } else {
1654                         minidx = mididx + 1;
1655                 }
1656         }
1657
1658         unsigned int bufsize = SPA_OLD_MAXBLOCKSIZE;
1659         redact_block_phys_t *buf = zio_data_buf_alloc(bufsize);
1660
1661         unsigned int entries_per_buf = bufsize / sizeof (redact_block_phys_t);
1662         uint64_t start_block = minidx / entries_per_buf;
1663         err = dmu_read(mos, rl->rl_object, start_block * bufsize, bufsize, buf,
1664             DMU_READ_PREFETCH);
1665
1666         for (uint64_t curidx = minidx;
1667             err == 0 && curidx < rl->rl_phys->rlp_num_entries;
1668             curidx++) {
1669                 /*
1670                  * We read in the redaction list one block at a time.  Once we
1671                  * finish with all the entries in a given block, we read in a
1672                  * new one.  The predictive prefetcher will take care of any
1673                  * prefetching, and this code shouldn't be the bottleneck, so we
1674                  * don't need to do manual prefetching.
1675                  */
1676                 if (curidx % entries_per_buf == 0) {
1677                         err = dmu_read(mos, rl->rl_object, curidx *
1678                             sizeof (*buf), bufsize, buf,
1679                             DMU_READ_PREFETCH);
1680                         if (err != 0)
1681                                 break;
1682                 }
1683                 redact_block_phys_t *rb = &buf[curidx % entries_per_buf];
1684                 /*
1685                  * If resume is non-null, we should either not send the data, or
1686                  * null out resume so we don't have to keep doing these
1687                  * comparisons.
1688                  */
1689                 if (resume != NULL) {
1690                         /*
1691                          * It is possible that after the binary search we got
1692                          * a record before the resume point. There's two cases
1693                          * where this can occur. If the record is the last
1694                          * redaction record, and the resume point is after the
1695                          * end of the redacted data, curidx will be the last
1696                          * redaction record. In that case, the loop will end
1697                          * after this iteration. The second case is if the
1698                          * resume point is between two redaction records, the
1699                          * binary search can return either the record before
1700                          * or after the resume point. In that case, the next
1701                          * iteration will be greater than the resume point.
1702                          */
1703                         if (redact_block_zb_compare(rb, resume) < 0) {
1704                                 ASSERT3U(curidx, ==, minidx);
1705                                 continue;
1706                         } else {
1707                                 /*
1708                                  * If the place to resume is in the middle of
1709                                  * the range described by this
1710                                  * redact_block_phys, then modify the
1711                                  * redact_block_phys in memory so we generate
1712                                  * the right records.
1713                                  */
1714                                 if (resume->zb_object == rb->rbp_object &&
1715                                     resume->zb_blkid > rb->rbp_blkid) {
1716                                         uint64_t diff = resume->zb_blkid -
1717                                             rb->rbp_blkid;
1718                                         rb->rbp_blkid = resume->zb_blkid;
1719                                         redact_block_set_count(rb,
1720                                             redact_block_get_count(rb) - diff);
1721                                 }
1722                                 resume = NULL;
1723                         }
1724                 }
1725
1726                 if (cb(rb, arg) != 0) {
1727                         err = EINTR;
1728                         break;
1729                 }
1730         }
1731
1732         zio_data_buf_free(buf, bufsize);
1733         return (err);
1734 }