sys/contrib/openzfs/module/zfs/bpobj.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
  24  * Copyright (c) 2017 Datto Inc.
  25  */
  26
  27 #include <sys/bpobj.h>
  28 #include <sys/zfs_context.h>
  29 #include <sys/zfs_refcount.h>
  30 #include <sys/dsl_pool.h>
  31 #include <sys/zfeature.h>
  32 #include <sys/zap.h>
  33
  34 /*
  35  * Return an empty bpobj, preferably the empty dummy one (dp_empty_bpobj).
  36  */
  37 uint64_t
  38 bpobj_alloc_empty(objset_t *os, int blocksize, dmu_tx_t *tx)
  39 {
  40         spa_t *spa = dmu_objset_spa(os);
  41         dsl_pool_t *dp = dmu_objset_pool(os);
  42
  43         if (spa_feature_is_enabled(spa, SPA_FEATURE_EMPTY_BPOBJ)) {
  44                 if (!spa_feature_is_active(spa, SPA_FEATURE_EMPTY_BPOBJ)) {
  45                         ASSERT0(dp->dp_empty_bpobj);
  46                         dp->dp_empty_bpobj =
  47                             bpobj_alloc(os, SPA_OLD_MAXBLOCKSIZE, tx);
  48                         VERIFY(zap_add(os,
  49                             DMU_POOL_DIRECTORY_OBJECT,
  50                             DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
  51                             &dp->dp_empty_bpobj, tx) == 0);
  52                 }
  53                 spa_feature_incr(spa, SPA_FEATURE_EMPTY_BPOBJ, tx);
  54                 ASSERT(dp->dp_empty_bpobj != 0);
  55                 return (dp->dp_empty_bpobj);
  56         } else {
  57                 return (bpobj_alloc(os, blocksize, tx));
  58         }
  59 }
  60
  61 void
  62 bpobj_decr_empty(objset_t *os, dmu_tx_t *tx)
  63 {
  64         dsl_pool_t *dp = dmu_objset_pool(os);
  65
  66         spa_feature_decr(dmu_objset_spa(os), SPA_FEATURE_EMPTY_BPOBJ, tx);
  67         if (!spa_feature_is_active(dmu_objset_spa(os),
  68             SPA_FEATURE_EMPTY_BPOBJ)) {
  69                 VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
  70                     DMU_POOL_DIRECTORY_OBJECT,
  71                     DMU_POOL_EMPTY_BPOBJ, tx));
  72                 VERIFY3U(0, ==, dmu_object_free(os, dp->dp_empty_bpobj, tx));
  73                 dp->dp_empty_bpobj = 0;
  74         }
  75 }
  76
  77 uint64_t
  78 bpobj_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
  79 {
  80         int size;
  81
  82         if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_BPOBJ_ACCOUNT)
  83                 size = BPOBJ_SIZE_V0;
  84         else if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_DEADLISTS)
  85                 size = BPOBJ_SIZE_V1;
  86         else if (!spa_feature_is_active(dmu_objset_spa(os),
  87             SPA_FEATURE_LIVELIST))
  88                 size = BPOBJ_SIZE_V2;
  89         else
  90                 size = sizeof (bpobj_phys_t);
  91
  92         return (dmu_object_alloc(os, DMU_OT_BPOBJ, blocksize,
  93             DMU_OT_BPOBJ_HDR, size, tx));
  94 }
  95
  96 void
  97 bpobj_free(objset_t *os, uint64_t obj, dmu_tx_t *tx)
  98 {
  99         int64_t i;
 100         bpobj_t bpo;
 101         dmu_object_info_t doi;
 102         int epb;
 103         dmu_buf_t *dbuf = NULL;
 104
 105         ASSERT(obj != dmu_objset_pool(os)->dp_empty_bpobj);
 106         VERIFY3U(0, ==, bpobj_open(&bpo, os, obj));
 107
 108         mutex_enter(&bpo.bpo_lock);
 109
 110         if (!bpo.bpo_havesubobj || bpo.bpo_phys->bpo_subobjs == 0)
 111                 goto out;
 112
 113         VERIFY3U(0, ==, dmu_object_info(os, bpo.bpo_phys->bpo_subobjs, &doi));
 114         epb = doi.doi_data_block_size / sizeof (uint64_t);
 115
 116         for (i = bpo.bpo_phys->bpo_num_subobjs - 1; i >= 0; i--) {
 117                 uint64_t *objarray;
 118                 uint64_t offset, blkoff;
 119
 120                 offset = i * sizeof (uint64_t);
 121                 blkoff = P2PHASE(i, epb);
 122
 123                 if (dbuf == NULL || dbuf->db_offset > offset) {
 124                         if (dbuf)
 125                                 dmu_buf_rele(dbuf, FTAG);
 126                         VERIFY3U(0, ==, dmu_buf_hold(os,
 127                             bpo.bpo_phys->bpo_subobjs, offset, FTAG, &dbuf, 0));
 128                 }
 129
 130                 ASSERT3U(offset, >=, dbuf->db_offset);
 131                 ASSERT3U(offset, <, dbuf->db_offset + dbuf->db_size);
 132
 133                 objarray = dbuf->db_data;
 134                 bpobj_free(os, objarray[blkoff], tx);
 135         }
 136         if (dbuf) {
 137                 dmu_buf_rele(dbuf, FTAG);
 138                 dbuf = NULL;
 139         }
 140         VERIFY3U(0, ==, dmu_object_free(os, bpo.bpo_phys->bpo_subobjs, tx));
 141
 142 out:
 143         mutex_exit(&bpo.bpo_lock);
 144         bpobj_close(&bpo);
 145
 146         VERIFY3U(0, ==, dmu_object_free(os, obj, tx));
 147 }
 148
 149 int
 150 bpobj_open(bpobj_t *bpo, objset_t *os, uint64_t object)
 151 {
 152         dmu_object_info_t doi;
 153         int err;
 154
 155         err = dmu_object_info(os, object, &doi);
 156         if (err)
 157                 return (err);
 158
 159         bzero(bpo, sizeof (*bpo));
 160         mutex_init(&bpo->bpo_lock, NULL, MUTEX_DEFAULT, NULL);
 161
 162         ASSERT(bpo->bpo_dbuf == NULL);
 163         ASSERT(bpo->bpo_phys == NULL);
 164         ASSERT(object != 0);
 165         ASSERT3U(doi.doi_type, ==, DMU_OT_BPOBJ);
 166         ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_BPOBJ_HDR);
 167
 168         err = dmu_bonus_hold(os, object, bpo, &bpo->bpo_dbuf);
 169         if (err)
 170                 return (err);
 171
 172         bpo->bpo_os = os;
 173         bpo->bpo_object = object;
 174         bpo->bpo_epb = doi.doi_data_block_size >> SPA_BLKPTRSHIFT;
 175         bpo->bpo_havecomp = (doi.doi_bonus_size > BPOBJ_SIZE_V0);
 176         bpo->bpo_havesubobj = (doi.doi_bonus_size > BPOBJ_SIZE_V1);
 177         bpo->bpo_havefreed = (doi.doi_bonus_size > BPOBJ_SIZE_V2);
 178         bpo->bpo_phys = bpo->bpo_dbuf->db_data;
 179         return (0);
 180 }
 181
 182 boolean_t
 183 bpobj_is_open(const bpobj_t *bpo)
 184 {
 185         return (bpo->bpo_object != 0);
 186 }
 187
 188 void
 189 bpobj_close(bpobj_t *bpo)
 190 {
 191         /* Lame workaround for closing a bpobj that was never opened. */
 192         if (bpo->bpo_object == 0)
 193                 return;
 194
 195         dmu_buf_rele(bpo->bpo_dbuf, bpo);
 196         if (bpo->bpo_cached_dbuf != NULL)
 197                 dmu_buf_rele(bpo->bpo_cached_dbuf, bpo);
 198         bpo->bpo_dbuf = NULL;
 199         bpo->bpo_phys = NULL;
 200         bpo->bpo_cached_dbuf = NULL;
 201         bpo->bpo_object = 0;
 202
 203         mutex_destroy(&bpo->bpo_lock);
 204 }
 205
 206 static boolean_t
 207 bpobj_is_empty_impl(bpobj_t *bpo)
 208 {
 209         ASSERT(MUTEX_HELD(&bpo->bpo_lock));
 210         return (bpo->bpo_phys->bpo_num_blkptrs == 0 &&
 211             (!bpo->bpo_havesubobj || bpo->bpo_phys->bpo_num_subobjs == 0));
 212 }
 213
 214 boolean_t
 215 bpobj_is_empty(bpobj_t *bpo)
 216 {
 217         mutex_enter(&bpo->bpo_lock);
 218         boolean_t is_empty = bpobj_is_empty_impl(bpo);
 219         mutex_exit(&bpo->bpo_lock);
 220         return (is_empty);
 221 }
 222
 223 /*
 224  * A recursive iteration of the bpobjs would be nice here but we run the risk
 225  * of overflowing function stack space.  Instead, find each subobj and add it
 226  * to the head of our list so it can be scanned for subjobjs.  Like a
 227  * recursive implementation, the "deepest" subobjs will be freed first.
 228  * When a subobj is found to have no additional subojs, free it.
 229  */
 230 typedef struct bpobj_info {
 231         bpobj_t *bpi_bpo;
 232         /*
 233          * This object is a subobj of bpi_parent,
 234          * at bpi_index in its subobj array.
 235          */
 236         struct bpobj_info *bpi_parent;
 237         uint64_t bpi_index;
 238         /* How many of our subobj's are left to process. */
 239         uint64_t bpi_unprocessed_subobjs;
 240         /* True after having visited this bpo's directly referenced BPs. */
 241         boolean_t bpi_visited;
 242         list_node_t bpi_node;
 243 } bpobj_info_t;
 244
 245 static bpobj_info_t *
 246 bpi_alloc(bpobj_t *bpo, bpobj_info_t *parent, uint64_t index)
 247 {
 248         bpobj_info_t *bpi = kmem_zalloc(sizeof (bpobj_info_t), KM_SLEEP);
 249         bpi->bpi_bpo = bpo;
 250         bpi->bpi_parent = parent;
 251         bpi->bpi_index = index;
 252         if (bpo->bpo_havesubobj && bpo->bpo_phys->bpo_subobjs != 0) {
 253                 bpi->bpi_unprocessed_subobjs = bpo->bpo_phys->bpo_num_subobjs;
 254         }
 255         return (bpi);
 256 }
 257
 258 /*
 259  * Update bpobj and all of its parents with new space accounting.
 260  */
 261 static void
 262 propagate_space_reduction(bpobj_info_t *bpi, int64_t freed,
 263     int64_t comp_freed, int64_t uncomp_freed, dmu_tx_t *tx)
 264 {
 265
 266         for (; bpi != NULL; bpi = bpi->bpi_parent) {
 267                 bpobj_t *p = bpi->bpi_bpo;
 268                 ASSERT(dmu_buf_is_dirty(p->bpo_dbuf, tx));
 269                 p->bpo_phys->bpo_bytes -= freed;
 270                 ASSERT3S(p->bpo_phys->bpo_bytes, >=, 0);
 271                 if (p->bpo_havecomp) {
 272                         p->bpo_phys->bpo_comp -= comp_freed;
 273                         p->bpo_phys->bpo_uncomp -= uncomp_freed;
 274                 }
 275         }
 276 }
 277
 278 static int
 279 bpobj_iterate_blkptrs(bpobj_info_t *bpi, bpobj_itor_t func, void *arg,
 280     int64_t start, dmu_tx_t *tx, boolean_t free)
 281 {
 282         int err = 0;
 283         int64_t freed = 0, comp_freed = 0, uncomp_freed = 0;
 284         dmu_buf_t *dbuf = NULL;
 285         bpobj_t *bpo = bpi->bpi_bpo;
 286
 287         for (int64_t i = bpo->bpo_phys->bpo_num_blkptrs - 1; i >= start; i--) {
 288                 uint64_t offset = i * sizeof (blkptr_t);
 289                 uint64_t blkoff = P2PHASE(i, bpo->bpo_epb);
 290
 291                 if (dbuf == NULL || dbuf->db_offset > offset) {
 292                         if (dbuf)
 293                                 dmu_buf_rele(dbuf, FTAG);
 294                         err = dmu_buf_hold(bpo->bpo_os, bpo->bpo_object,
 295                             offset, FTAG, &dbuf, 0);
 296                         if (err)
 297                                 break;
 298                 }
 299
 300                 ASSERT3U(offset, >=, dbuf->db_offset);
 301                 ASSERT3U(offset, <, dbuf->db_offset + dbuf->db_size);
 302
 303                 blkptr_t *bparray = dbuf->db_data;
 304                 blkptr_t *bp = &bparray[blkoff];
 305
 306                 boolean_t bp_freed = BP_GET_FREE(bp);
 307                 err = func(arg, bp, bp_freed, tx);
 308                 if (err)
 309                         break;
 310
 311                 if (free) {
 312                         int sign = bp_freed ? -1 : +1;
 313                         spa_t *spa = dmu_objset_spa(bpo->bpo_os);
 314                         freed += sign * bp_get_dsize_sync(spa, bp);
 315                         comp_freed += sign * BP_GET_PSIZE(bp);
 316                         uncomp_freed += sign * BP_GET_UCSIZE(bp);
 317                         ASSERT(dmu_buf_is_dirty(bpo->bpo_dbuf, tx));
 318                         bpo->bpo_phys->bpo_num_blkptrs--;
 319                         ASSERT3S(bpo->bpo_phys->bpo_num_blkptrs, >=, 0);
 320                         if (bp_freed) {
 321                                 ASSERT(bpo->bpo_havefreed);
 322                                 bpo->bpo_phys->bpo_num_freed--;
 323                                 ASSERT3S(bpo->bpo_phys->bpo_num_freed, >=, 0);
 324                         }
 325                 }
 326         }
 327         if (free) {
 328                 propagate_space_reduction(bpi, freed, comp_freed,
 329                     uncomp_freed, tx);
 330                 VERIFY0(dmu_free_range(bpo->bpo_os,
 331                     bpo->bpo_object,
 332                     bpo->bpo_phys->bpo_num_blkptrs * sizeof (blkptr_t),
 333                     DMU_OBJECT_END, tx));
 334         }
 335         if (dbuf) {
 336                 dmu_buf_rele(dbuf, FTAG);
 337                 dbuf = NULL;
 338         }
 339         return (err);
 340 }
 341
 342 /*
 343  * Given an initial bpo, start by freeing the BPs that are directly referenced
 344  * by that bpo. If the bpo has subobjs, read in its last subobj and push the
 345  * subobj to our stack. By popping items off our stack, eventually we will
 346  * encounter a bpo that has no subobjs.  We can free its bpobj_info_t, and if
 347  * requested also free the now-empty bpo from disk and decrement
 348  * its parent's subobj count. We continue popping each subobj from our stack,
 349  * visiting its last subobj until they too have no more subobjs, and so on.
 350  */
 351 static int
 352 bpobj_iterate_impl(bpobj_t *initial_bpo, bpobj_itor_t func, void *arg,
 353     dmu_tx_t *tx, boolean_t free, uint64_t *bpobj_size)
 354 {
 355         list_t stack;
 356         bpobj_info_t *bpi;
 357         int err = 0;
 358
 359         /*
 360          * Create a "stack" for us to work with without worrying about
 361          * stack overflows. Initialize it with the initial_bpo.
 362          */
 363         list_create(&stack, sizeof (bpobj_info_t),
 364             offsetof(bpobj_info_t, bpi_node));
 365         mutex_enter(&initial_bpo->bpo_lock);
 366
 367         if (bpobj_size != NULL)
 368                 *bpobj_size = initial_bpo->bpo_phys->bpo_num_blkptrs;
 369
 370         list_insert_head(&stack, bpi_alloc(initial_bpo, NULL, 0));
 371
 372         while ((bpi = list_head(&stack)) != NULL) {
 373                 bpobj_t *bpo = bpi->bpi_bpo;
 374
 375                 ASSERT3P(bpo, !=, NULL);
 376                 ASSERT(MUTEX_HELD(&bpo->bpo_lock));
 377                 ASSERT(bpobj_is_open(bpo));
 378
 379                 if (free)
 380                         dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
 381
 382                 if (bpi->bpi_visited == B_FALSE) {
 383                         err = bpobj_iterate_blkptrs(bpi, func, arg, 0, tx,
 384                             free);
 385                         bpi->bpi_visited = B_TRUE;
 386                         if (err != 0)
 387                                 break;
 388                 }
 389                 /*
 390                  * We've finished with this bpo's directly-referenced BP's and
 391                  * it has no more unprocessed subobjs. We can free its
 392                  * bpobj_info_t (unless it is the topmost, initial_bpo).
 393                  * If we are freeing from disk, we can also do that.
 394                  */
 395                 if (bpi->bpi_unprocessed_subobjs == 0) {
 396                         /*
 397                          * If there are no entries, there should
 398                          * be no bytes.
 399                          */
 400                         if (bpobj_is_empty_impl(bpo)) {
 401                                 ASSERT0(bpo->bpo_phys->bpo_bytes);
 402                                 ASSERT0(bpo->bpo_phys->bpo_comp);
 403                                 ASSERT0(bpo->bpo_phys->bpo_uncomp);
 404                         }
 405
 406                         /* The initial_bpo has no parent and is not closed. */
 407                         if (bpi->bpi_parent != NULL) {
 408                                 if (free) {
 409                                         bpobj_t *p = bpi->bpi_parent->bpi_bpo;
 410
 411                                         ASSERT0(bpo->bpo_phys->bpo_num_blkptrs);
 412                                         ASSERT3U(p->bpo_phys->bpo_num_subobjs,
 413                                             >, 0);
 414                                         ASSERT3U(bpi->bpi_index, ==,
 415                                             p->bpo_phys->bpo_num_subobjs - 1);
 416                                         ASSERT(dmu_buf_is_dirty(bpo->bpo_dbuf,
 417                                             tx));
 418
 419                                         p->bpo_phys->bpo_num_subobjs--;
 420
 421                                         VERIFY0(dmu_free_range(p->bpo_os,
 422                                             p->bpo_phys->bpo_subobjs,
 423                                             bpi->bpi_index * sizeof (uint64_t),
 424                                             sizeof (uint64_t), tx));
 425
 426                                         /* eliminate the empty subobj list */
 427                                         if (bpo->bpo_havesubobj &&
 428                                             bpo->bpo_phys->bpo_subobjs != 0) {
 429                                                 ASSERT0(bpo->bpo_phys->
 430                                                     bpo_num_subobjs);
 431                                                 err = dmu_object_free(
 432                                                     bpo->bpo_os,
 433                                                     bpo->bpo_phys->bpo_subobjs,
 434                                                     tx);
 435                                                 if (err)
 436                                                         break;
 437                                                 bpo->bpo_phys->bpo_subobjs = 0;
 438                                         }
 439                                         err = dmu_object_free(p->bpo_os,
 440                                             bpo->bpo_object, tx);
 441                                         if (err)
 442                                                 break;
 443                                 }
 444
 445                                 mutex_exit(&bpo->bpo_lock);
 446                                 bpobj_close(bpo);
 447                                 kmem_free(bpo, sizeof (bpobj_t));
 448                         } else {
 449                                 mutex_exit(&bpo->bpo_lock);
 450                         }
 451
 452                         /*
 453                          * Finished processing this bpo. Unlock, and free
 454                          * our "stack" info.
 455                          */
 456                         list_remove_head(&stack);
 457                         kmem_free(bpi, sizeof (bpobj_info_t));
 458                 } else {
 459                         /*
 460                          * We have unprocessed subobjs. Process the next one.
 461                          */
 462                         ASSERT(bpo->bpo_havecomp);
 463                         ASSERT3P(bpobj_size, ==, NULL);
 464
 465                         /* Add the last subobj to stack. */
 466                         int64_t i = bpi->bpi_unprocessed_subobjs - 1;
 467                         uint64_t offset = i * sizeof (uint64_t);
 468
 469                         uint64_t obj_from_sublist;
 470                         err = dmu_read(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs,
 471                             offset, sizeof (uint64_t), &obj_from_sublist,
 472                             DMU_READ_PREFETCH);
 473                         if (err)
 474                                 break;
 475                         bpobj_t *sublist = kmem_alloc(sizeof (bpobj_t),
 476                             KM_SLEEP);
 477
 478                         err = bpobj_open(sublist, bpo->bpo_os,
 479                             obj_from_sublist);
 480                         if (err)
 481                                 break;
 482
 483                         list_insert_head(&stack, bpi_alloc(sublist, bpi, i));
 484                         mutex_enter(&sublist->bpo_lock);
 485                         bpi->bpi_unprocessed_subobjs--;
 486                 }
 487         }
 488         /*
 489          * Cleanup anything left on the "stack" after we left the loop.
 490          * Every bpo on the stack is locked so we must remember to undo
 491          * that now (in LIFO order).
 492          */
 493         while ((bpi = list_remove_head(&stack)) != NULL) {
 494                 bpobj_t *bpo = bpi->bpi_bpo;
 495                 ASSERT(err != 0);
 496                 ASSERT3P(bpo, !=, NULL);
 497
 498                 mutex_exit(&bpo->bpo_lock);
 499
 500                 /* do not free the initial_bpo */
 501                 if (bpi->bpi_parent != NULL) {
 502                         bpobj_close(bpi->bpi_bpo);
 503                         kmem_free(bpi->bpi_bpo, sizeof (bpobj_t));
 504                 }
 505                 kmem_free(bpi, sizeof (bpobj_info_t));
 506         }
 507
 508         list_destroy(&stack);
 509
 510         return (err);
 511 }
 512
 513 /*
 514  * Iterate and remove the entries.  If func returns nonzero, iteration
 515  * will stop and that entry will not be removed.
 516  */
 517 int
 518 bpobj_iterate(bpobj_t *bpo, bpobj_itor_t func, void *arg, dmu_tx_t *tx)
 519 {
 520         return (bpobj_iterate_impl(bpo, func, arg, tx, B_TRUE, NULL));
 521 }
 522
 523 /*
 524  * Iterate the entries.  If func returns nonzero, iteration will stop.
 525  *
 526  * If there are no subobjs:
 527  *
 528  * *bpobj_size can be used to return the number of block pointers in the
 529  * bpobj.  Note that this may be different from the number of block pointers
 530  * that are iterated over, if iteration is terminated early (e.g. by the func
 531  * returning nonzero).
 532  *
 533  * If there are concurrent (or subsequent) modifications to the bpobj then the
 534  * returned *bpobj_size can be passed as "start" to
 535  * livelist_bpobj_iterate_from_nofree() to iterate the newly added entries.
 536  */
 537 int
 538 bpobj_iterate_nofree(bpobj_t *bpo, bpobj_itor_t func, void *arg,
 539     uint64_t *bpobj_size)
 540 {
 541         return (bpobj_iterate_impl(bpo, func, arg, NULL, B_FALSE, bpobj_size));
 542 }
 543
 544 /*
 545  * Iterate over the blkptrs in the bpobj beginning at index start. If func
 546  * returns nonzero, iteration will stop. This is a livelist specific function
 547  * since it assumes that there are no subobjs present.
 548  */
 549 int
 550 livelist_bpobj_iterate_from_nofree(bpobj_t *bpo, bpobj_itor_t func, void *arg,
 551     int64_t start)
 552 {
 553         if (bpo->bpo_havesubobj)
 554                 VERIFY0(bpo->bpo_phys->bpo_subobjs);
 555         bpobj_info_t *bpi = bpi_alloc(bpo, NULL, 0);
 556         int err = bpobj_iterate_blkptrs(bpi, func, arg, start, NULL, B_FALSE);
 557         kmem_free(bpi, sizeof (bpobj_info_t));
 558         return (err);
 559 }
 560
 561 /*
 562  * Logically add subobj's contents to the parent bpobj.
 563  *
 564  * In the most general case, this is accomplished in constant time by adding
 565  * a reference to subobj.  This case is used when enqueuing a large subobj:
 566  * +--------------+                        +--------------+
 567  * | bpobj        |----------------------->| subobj list  |
 568  * +----+----+----+----+----+              +-----+-----+--+--+
 569  * | bp | bp | bp | bp | bp |              | obj | obj | obj |
 570  * +----+----+----+----+----+              +-----+-----+-----+
 571  *
 572  * +--------------+                        +--------------+
 573  * | sub-bpobj    |----------------------> | subsubobj    |
 574  * +----+----+----+----+---------+----+    +-----+-----+--+--------+-----+
 575  * | bp | bp | bp | bp |   ...   | bp |    | obj | obj |    ...    | obj |
 576  * +----+----+----+----+---------+----+    +-----+-----+-----------+-----+
 577  *
 578  * Result: sub-bpobj added to parent's subobj list.
 579  * +--------------+                        +--------------+
 580  * | bpobj        |----------------------->| subobj list  |
 581  * +----+----+----+----+----+              +-----+-----+--+--+-----+
 582  * | bp | bp | bp | bp | bp |              | obj | obj | obj | OBJ |
 583  * +----+----+----+----+----+              +-----+-----+-----+--|--+
 584  *                                                              |
 585  *       /-----------------------------------------------------/
 586  *       v
 587  * +--------------+                        +--------------+
 588  * | sub-bpobj    |----------------------> | subsubobj    |
 589  * +----+----+----+----+---------+----+    +-----+-----+--+--------+-----+
 590  * | bp | bp | bp | bp |   ...   | bp |    | obj | obj |    ...    | obj |
 591  * +----+----+----+----+---------+----+    +-----+-----+-----------+-----+
 592  *
 593  *
 594  * In a common case, the subobj is small: its bp's and its list of subobj's
 595  * are each stored in a single block.  In this case we copy the subobj's
 596  * contents to the parent:
 597  * +--------------+                        +--------------+
 598  * | bpobj        |----------------------->| subobj list  |
 599  * +----+----+----+----+----+              +-----+-----+--+--+
 600  * | bp | bp | bp | bp | bp |              | obj | obj | obj |
 601  * +----+----+----+----+----+              +-----+-----+-----+
 602  *                          ^                                ^
 603  * +--------------+         |              +--------------+  |
 604  * | sub-bpobj    |---------^------------> | subsubobj    |  ^
 605  * +----+----+----+         |              +-----+-----+--+  |
 606  * | BP | BP |-->-->-->-->-/               | OBJ | OBJ |-->-/
 607  * +----+----+                             +-----+-----+
 608  *
 609  * Result: subobj destroyed, contents copied to parent:
 610  * +--------------+                        +--------------+
 611  * | bpobj        |----------------------->| subobj list  |
 612  * +----+----+----+----+----+----+----+    +-----+-----+--+--+-----+-----+
 613  * | bp | bp | bp | bp | bp | BP | BP |    | obj | obj | obj | OBJ | OBJ |
 614  * +----+----+----+----+----+----+----+    +-----+-----+-----+-----+-----+
 615  *
 616  *
 617  * If the subobj has many BP's but few subobj's, we can copy the sub-subobj's
 618  * but retain the sub-bpobj:
 619  * +--------------+                        +--------------+
 620  * | bpobj        |----------------------->| subobj list  |
 621  * +----+----+----+----+----+              +-----+-----+--+--+
 622  * | bp | bp | bp | bp | bp |              | obj | obj | obj |
 623  * +----+----+----+----+----+              +-----+-----+-----+
 624  *                                                           ^
 625  * +--------------+                        +--------------+  |
 626  * | sub-bpobj    |----------------------> | subsubobj    |  ^
 627  * +----+----+----+----+---------+----+    +-----+-----+--+  |
 628  * | bp | bp | bp | bp |   ...   | bp |    | OBJ | OBJ |-->-/
 629  * +----+----+----+----+---------+----+    +-----+-----+
 630  *
 631  * Result: sub-sub-bpobjs and subobj added to parent's subobj list.
 632  * +--------------+                     +--------------+
 633  * | bpobj        |-------------------->| subobj list  |
 634  * +----+----+----+----+----+           +-----+-----+--+--+-----+-----+------+
 635  * | bp | bp | bp | bp | bp |           | obj | obj | obj | OBJ | OBJ | OBJ* |
 636  * +----+----+----+----+----+           +-----+-----+-----+-----+-----+--|---+
 637  *                                                                       |
 638  *       /--------------------------------------------------------------/
 639  *       v
 640  * +--------------+
 641  * | sub-bpobj    |
 642  * +----+----+----+----+---------+----+
 643  * | bp | bp | bp | bp |   ...   | bp |
 644  * +----+----+----+----+---------+----+
 645  */
 646 void
 647 bpobj_enqueue_subobj(bpobj_t *bpo, uint64_t subobj, dmu_tx_t *tx)
 648 {
 649         bpobj_t subbpo;
 650         uint64_t used, comp, uncomp, subsubobjs;
 651         boolean_t copy_subsub = B_TRUE;
 652         boolean_t copy_bps = B_TRUE;
 653
 654         ASSERT(bpobj_is_open(bpo));
 655         ASSERT(subobj != 0);
 656         ASSERT(bpo->bpo_havesubobj);
 657         ASSERT(bpo->bpo_havecomp);
 658         ASSERT(bpo->bpo_object != dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj);
 659
 660         if (subobj == dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj) {
 661                 bpobj_decr_empty(bpo->bpo_os, tx);
 662                 return;
 663         }
 664
 665         VERIFY3U(0, ==, bpobj_open(&subbpo, bpo->bpo_os, subobj));
 666         VERIFY3U(0, ==, bpobj_space(&subbpo, &used, &comp, &uncomp));
 667
 668         if (bpobj_is_empty(&subbpo)) {
 669                 /* No point in having an empty subobj. */
 670                 bpobj_close(&subbpo);
 671                 bpobj_free(bpo->bpo_os, subobj, tx);
 672                 return;
 673         }
 674
 675         mutex_enter(&bpo->bpo_lock);
 676         dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
 677
 678         dmu_object_info_t doi;
 679
 680         if (bpo->bpo_phys->bpo_subobjs != 0) {
 681                 ASSERT0(dmu_object_info(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs,
 682                     &doi));
 683                 ASSERT3U(doi.doi_type, ==, DMU_OT_BPOBJ_SUBOBJ);
 684         }
 685
 686         /*
 687          * If subobj has only one block of subobjs, then move subobj's
 688          * subobjs to bpo's subobj list directly.  This reduces recursion in
 689          * bpobj_iterate due to nested subobjs.
 690          */
 691         subsubobjs = subbpo.bpo_phys->bpo_subobjs;
 692         if (subsubobjs != 0) {
 693                 VERIFY0(dmu_object_info(bpo->bpo_os, subsubobjs, &doi));
 694                 if (doi.doi_max_offset > doi.doi_data_block_size) {
 695                         copy_subsub = B_FALSE;
 696                 }
 697         }
 698
 699         /*
 700          * If, in addition to having only one block of subobj's, subobj has
 701          * only one block of bp's, then move subobj's bp's to bpo's bp list
 702          * directly. This reduces recursion in bpobj_iterate due to nested
 703          * subobjs.
 704          */
 705         VERIFY3U(0, ==, dmu_object_info(bpo->bpo_os, subobj, &doi));
 706         if (doi.doi_max_offset > doi.doi_data_block_size || !copy_subsub) {
 707                 copy_bps = B_FALSE;
 708         }
 709
 710         if (copy_subsub && subsubobjs != 0) {
 711                 dmu_buf_t *subdb;
 712                 uint64_t numsubsub = subbpo.bpo_phys->bpo_num_subobjs;
 713
 714                 VERIFY0(dmu_buf_hold(bpo->bpo_os, subsubobjs,
 715                     0, FTAG, &subdb, 0));
 716                 /*
 717                  * Make sure that we are not asking dmu_write()
 718                  * to write more data than we have in our buffer.
 719                  */
 720                 VERIFY3U(subdb->db_size, >=,
 721                     numsubsub * sizeof (subobj));
 722                 if (bpo->bpo_phys->bpo_subobjs == 0) {
 723                         bpo->bpo_phys->bpo_subobjs =
 724                             dmu_object_alloc(bpo->bpo_os,
 725                             DMU_OT_BPOBJ_SUBOBJ, SPA_OLD_MAXBLOCKSIZE,
 726                             DMU_OT_NONE, 0, tx);
 727                 }
 728                 dmu_write(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs,
 729                     bpo->bpo_phys->bpo_num_subobjs * sizeof (subobj),
 730                     numsubsub * sizeof (subobj), subdb->db_data, tx);
 731                 dmu_buf_rele(subdb, FTAG);
 732                 bpo->bpo_phys->bpo_num_subobjs += numsubsub;
 733
 734                 dmu_buf_will_dirty(subbpo.bpo_dbuf, tx);
 735                 subbpo.bpo_phys->bpo_subobjs = 0;
 736                 VERIFY0(dmu_object_free(bpo->bpo_os, subsubobjs, tx));
 737         }
 738
 739         if (copy_bps) {
 740                 dmu_buf_t *bps;
 741                 uint64_t numbps = subbpo.bpo_phys->bpo_num_blkptrs;
 742
 743                 ASSERT(copy_subsub);
 744                 VERIFY0(dmu_buf_hold(bpo->bpo_os, subobj,
 745                     0, FTAG, &bps, 0));
 746
 747                 /*
 748                  * Make sure that we are not asking dmu_write()
 749                  * to write more data than we have in our buffer.
 750                  */
 751                 VERIFY3U(bps->db_size, >=, numbps * sizeof (blkptr_t));
 752                 dmu_write(bpo->bpo_os, bpo->bpo_object,
 753                     bpo->bpo_phys->bpo_num_blkptrs * sizeof (blkptr_t),
 754                     numbps * sizeof (blkptr_t),
 755                     bps->db_data, tx);
 756                 dmu_buf_rele(bps, FTAG);
 757                 bpo->bpo_phys->bpo_num_blkptrs += numbps;
 758
 759                 bpobj_close(&subbpo);
 760                 VERIFY0(dmu_object_free(bpo->bpo_os, subobj, tx));
 761         } else {
 762                 bpobj_close(&subbpo);
 763                 if (bpo->bpo_phys->bpo_subobjs == 0) {
 764                         bpo->bpo_phys->bpo_subobjs =
 765                             dmu_object_alloc(bpo->bpo_os,
 766                             DMU_OT_BPOBJ_SUBOBJ, SPA_OLD_MAXBLOCKSIZE,
 767                             DMU_OT_NONE, 0, tx);
 768                 }
 769
 770                 dmu_write(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs,
 771                     bpo->bpo_phys->bpo_num_subobjs * sizeof (subobj),
 772                     sizeof (subobj), &subobj, tx);
 773                 bpo->bpo_phys->bpo_num_subobjs++;
 774         }
 775
 776         bpo->bpo_phys->bpo_bytes += used;
 777         bpo->bpo_phys->bpo_comp += comp;
 778         bpo->bpo_phys->bpo_uncomp += uncomp;
 779         mutex_exit(&bpo->bpo_lock);
 780
 781 }
 782
 783 void
 784 bpobj_enqueue(bpobj_t *bpo, const blkptr_t *bp, boolean_t bp_freed,
 785     dmu_tx_t *tx)
 786 {
 787         blkptr_t stored_bp = *bp;
 788         uint64_t offset;
 789         int blkoff;
 790         blkptr_t *bparray;
 791
 792         ASSERT(bpobj_is_open(bpo));
 793         ASSERT(!BP_IS_HOLE(bp));
 794         ASSERT(bpo->bpo_object != dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj);
 795
 796         if (BP_IS_EMBEDDED(bp)) {
 797                 /*
 798                  * The bpobj will compress better without the payload.
 799                  *
 800                  * Note that we store EMBEDDED bp's because they have an
 801                  * uncompressed size, which must be accounted for.  An
 802                  * alternative would be to add their size to bpo_uncomp
 803                  * without storing the bp, but that would create additional
 804                  * complications: bpo_uncomp would be inconsistent with the
 805                  * set of BP's stored, and bpobj_iterate() wouldn't visit
 806                  * all the space accounted for in the bpobj.
 807                  */
 808                 bzero(&stored_bp, sizeof (stored_bp));
 809                 stored_bp.blk_prop = bp->blk_prop;
 810                 stored_bp.blk_birth = bp->blk_birth;
 811         } else if (!BP_GET_DEDUP(bp)) {
 812                 /* The bpobj will compress better without the checksum */
 813                 bzero(&stored_bp.blk_cksum, sizeof (stored_bp.blk_cksum));
 814         }
 815
 816         stored_bp.blk_fill = 0;
 817         BP_SET_FREE(&stored_bp, bp_freed);
 818
 819         mutex_enter(&bpo->bpo_lock);
 820
 821         offset = bpo->bpo_phys->bpo_num_blkptrs * sizeof (stored_bp);
 822         blkoff = P2PHASE(bpo->bpo_phys->bpo_num_blkptrs, bpo->bpo_epb);
 823
 824         if (bpo->bpo_cached_dbuf == NULL ||
 825             offset < bpo->bpo_cached_dbuf->db_offset ||
 826             offset >= bpo->bpo_cached_dbuf->db_offset +
 827             bpo->bpo_cached_dbuf->db_size) {
 828                 if (bpo->bpo_cached_dbuf)
 829                         dmu_buf_rele(bpo->bpo_cached_dbuf, bpo);
 830                 VERIFY3U(0, ==, dmu_buf_hold(bpo->bpo_os, bpo->bpo_object,
 831                     offset, bpo, &bpo->bpo_cached_dbuf, 0));
 832         }
 833
 834         dmu_buf_will_dirty(bpo->bpo_cached_dbuf, tx);
 835         bparray = bpo->bpo_cached_dbuf->db_data;
 836         bparray[blkoff] = stored_bp;
 837
 838         dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
 839         bpo->bpo_phys->bpo_num_blkptrs++;
 840         int sign = bp_freed ? -1 : +1;
 841         bpo->bpo_phys->bpo_bytes += sign *
 842             bp_get_dsize_sync(dmu_objset_spa(bpo->bpo_os), bp);
 843         if (bpo->bpo_havecomp) {
 844                 bpo->bpo_phys->bpo_comp += sign * BP_GET_PSIZE(bp);
 845                 bpo->bpo_phys->bpo_uncomp += sign * BP_GET_UCSIZE(bp);
 846         }
 847         if (bp_freed) {
 848                 ASSERT(bpo->bpo_havefreed);
 849                 bpo->bpo_phys->bpo_num_freed++;
 850         }
 851         mutex_exit(&bpo->bpo_lock);
 852 }
 853
 854 struct space_range_arg {
 855         spa_t *spa;
 856         uint64_t mintxg;
 857         uint64_t maxtxg;
 858         uint64_t used;
 859         uint64_t comp;
 860         uint64_t uncomp;
 861 };
 862
 863 static int
 864 space_range_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, dmu_tx_t *tx)
 865 {
 866         (void) bp_freed, (void) tx;
 867         struct space_range_arg *sra = arg;
 868
 869         if (bp->blk_birth > sra->mintxg && bp->blk_birth <= sra->maxtxg) {
 870                 if (dsl_pool_sync_context(spa_get_dsl(sra->spa)))
 871                         sra->used += bp_get_dsize_sync(sra->spa, bp);
 872                 else
 873                         sra->used += bp_get_dsize(sra->spa, bp);
 874                 sra->comp += BP_GET_PSIZE(bp);
 875                 sra->uncomp += BP_GET_UCSIZE(bp);
 876         }
 877         return (0);
 878 }
 879
 880 int
 881 bpobj_space(bpobj_t *bpo, uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
 882 {
 883         ASSERT(bpobj_is_open(bpo));
 884         mutex_enter(&bpo->bpo_lock);
 885
 886         *usedp = bpo->bpo_phys->bpo_bytes;
 887         if (bpo->bpo_havecomp) {
 888                 *compp = bpo->bpo_phys->bpo_comp;
 889                 *uncompp = bpo->bpo_phys->bpo_uncomp;
 890                 mutex_exit(&bpo->bpo_lock);
 891                 return (0);
 892         } else {
 893                 mutex_exit(&bpo->bpo_lock);
 894                 return (bpobj_space_range(bpo, 0, UINT64_MAX,
 895                     usedp, compp, uncompp));
 896         }
 897 }
 898
 899 /*
 900  * Return the amount of space in the bpobj which is:
 901  * mintxg < blk_birth <= maxtxg
 902  */
 903 int
 904 bpobj_space_range(bpobj_t *bpo, uint64_t mintxg, uint64_t maxtxg,
 905     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
 906 {
 907         struct space_range_arg sra = { 0 };
 908         int err;
 909
 910         ASSERT(bpobj_is_open(bpo));
 911
 912         /*
 913          * As an optimization, if they want the whole txg range, just
 914          * get bpo_bytes rather than iterating over the bps.
 915          */
 916         if (mintxg < TXG_INITIAL && maxtxg == UINT64_MAX && bpo->bpo_havecomp)
 917                 return (bpobj_space(bpo, usedp, compp, uncompp));
 918
 919         sra.spa = dmu_objset_spa(bpo->bpo_os);
 920         sra.mintxg = mintxg;
 921         sra.maxtxg = maxtxg;
 922
 923         err = bpobj_iterate_nofree(bpo, space_range_cb, &sra, NULL);
 924         *usedp = sra.used;
 925         *compp = sra.comp;
 926         *uncompp = sra.uncomp;
 927         return (err);
 928 }
 929
 930 /*
 931  * A bpobj_itor_t to append blkptrs to a bplist. Note that while blkptrs in a
 932  * bpobj are designated as free or allocated that information is not preserved
 933  * in bplists.
 934  */
 935 int
 936 bplist_append_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
 937     dmu_tx_t *tx)
 938 {
 939         (void) bp_freed, (void) tx;
 940         bplist_t *bpl = arg;
 941         bplist_append(bpl, bp);
 942         return (0);
 943 }