sys/contrib/openzfs/module/zfs/dmu_traverse.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) 2012, 2018 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/zfs_context.h>
  27 #include <sys/dmu_objset.h>
  28 #include <sys/dmu_traverse.h>
  29 #include <sys/dsl_dataset.h>
  30 #include <sys/dsl_dir.h>
  31 #include <sys/dsl_pool.h>
  32 #include <sys/dnode.h>
  33 #include <sys/spa.h>
  34 #include <sys/spa_impl.h>
  35 #include <sys/zio.h>
  36 #include <sys/dmu_impl.h>
  37 #include <sys/sa.h>
  38 #include <sys/sa_impl.h>
  39 #include <sys/callb.h>
  40 #include <sys/zfeature.h>
  41
  42 static int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;     /* 50MB */
  43 static int32_t send_holes_without_birth_time = 1;
  44 static int32_t zfs_traverse_indirect_prefetch_limit = 32;
  45
  46 typedef struct prefetch_data {
  47         kmutex_t pd_mtx;
  48         kcondvar_t pd_cv;
  49         int32_t pd_bytes_fetched;
  50         int pd_flags;
  51         boolean_t pd_cancel;
  52         boolean_t pd_exited;
  53         zbookmark_phys_t pd_resume;
  54 } prefetch_data_t;
  55
  56 typedef struct traverse_data {
  57         spa_t *td_spa;
  58         uint64_t td_objset;
  59         blkptr_t *td_rootbp;
  60         uint64_t td_min_txg;
  61         zbookmark_phys_t *td_resume;
  62         int td_flags;
  63         prefetch_data_t *td_pfd;
  64         boolean_t td_paused;
  65         uint64_t td_hole_birth_enabled_txg;
  66         blkptr_cb_t *td_func;
  67         void *td_arg;
  68         boolean_t td_realloc_possible;
  69 } traverse_data_t;
  70
  71 static int traverse_dnode(traverse_data_t *td, const blkptr_t *bp,
  72     const dnode_phys_t *dnp, uint64_t objset, uint64_t object);
  73 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
  74     uint64_t objset, uint64_t object);
  75
  76 static int
  77 traverse_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
  78     uint64_t claim_txg)
  79 {
  80         traverse_data_t *td = arg;
  81         zbookmark_phys_t zb;
  82
  83         if (BP_IS_HOLE(bp))
  84                 return (0);
  85
  86         if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa))
  87                 return (-1);
  88
  89         SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
  90             bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
  91
  92         (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
  93
  94         return (0);
  95 }
  96
  97 static int
  98 traverse_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
  99     uint64_t claim_txg)
 100 {
 101         traverse_data_t *td = arg;
 102
 103         if (lrc->lrc_txtype == TX_WRITE) {
 104                 lr_write_t *lr = (lr_write_t *)lrc;
 105                 blkptr_t *bp = &lr->lr_blkptr;
 106                 zbookmark_phys_t zb;
 107
 108                 if (BP_IS_HOLE(bp))
 109                         return (0);
 110
 111                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 112                         return (0);
 113
 114                 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
 115                     ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
 116
 117                 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
 118                     td->td_arg);
 119         }
 120         return (0);
 121 }
 122
 123 static void
 124 traverse_zil(traverse_data_t *td, zil_header_t *zh)
 125 {
 126         uint64_t claim_txg = zh->zh_claim_txg;
 127
 128         /*
 129          * We only want to visit blocks that have been claimed but not yet
 130          * replayed; plus blocks that are already stable in read-only mode.
 131          */
 132         if (claim_txg == 0 && spa_writeable(td->td_spa))
 133                 return;
 134
 135         zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
 136         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
 137             claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
 138         zil_free(zilog);
 139 }
 140
 141 typedef enum resume_skip {
 142         RESUME_SKIP_ALL,
 143         RESUME_SKIP_NONE,
 144         RESUME_SKIP_CHILDREN
 145 } resume_skip_t;
 146
 147 /*
 148  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
 149  * the block indicated by zb does not need to be visited at all. Returns
 150  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
 151  * resume point. This indicates that this block should be visited but not its
 152  * children (since they must have been visited in a previous traversal).
 153  * Otherwise returns RESUME_SKIP_NONE.
 154  */
 155 static resume_skip_t
 156 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
 157     const zbookmark_phys_t *zb)
 158 {
 159         if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
 160                 /*
 161                  * If we already visited this bp & everything below,
 162                  * don't bother doing it again.
 163                  */
 164                 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
 165                         return (RESUME_SKIP_ALL);
 166
 167                 /*
 168                  * If we found the block we're trying to resume from, zero
 169                  * the bookmark out to indicate that we have resumed.
 170                  */
 171                 if (memcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
 172                         memset(td->td_resume, 0, sizeof (*zb));
 173                         if (td->td_flags & TRAVERSE_POST)
 174                                 return (RESUME_SKIP_CHILDREN);
 175                 }
 176         }
 177         return (RESUME_SKIP_NONE);
 178 }
 179
 180 /*
 181  * Returns B_TRUE, if prefetch read is issued, otherwise B_FALSE.
 182  */
 183 static boolean_t
 184 traverse_prefetch_metadata(traverse_data_t *td,
 185     const blkptr_t *bp, const zbookmark_phys_t *zb)
 186 {
 187         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 188         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
 189
 190         if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
 191                 return (B_FALSE);
 192         /*
 193          * If we are in the process of resuming, don't prefetch, because
 194          * some children will not be needed (and in fact may have already
 195          * been freed).
 196          */
 197         if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
 198                 return (B_FALSE);
 199         if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
 200                 return (B_FALSE);
 201         if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
 202                 return (B_FALSE);
 203         ASSERT(!BP_IS_REDACTED(bp));
 204
 205         if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 206                 zio_flags |= ZIO_FLAG_RAW;
 207
 208         (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
 209             ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 210         return (B_TRUE);
 211 }
 212
 213 static boolean_t
 214 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
 215 {
 216         ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
 217         if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
 218             BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp))
 219                 return (B_FALSE);
 220         return (B_TRUE);
 221 }
 222
 223 static int
 224 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
 225     const blkptr_t *bp, const zbookmark_phys_t *zb)
 226 {
 227         int err = 0;
 228         arc_buf_t *buf = NULL;
 229         prefetch_data_t *pd = td->td_pfd;
 230
 231         switch (resume_skip_check(td, dnp, zb)) {
 232         case RESUME_SKIP_ALL:
 233                 return (0);
 234         case RESUME_SKIP_CHILDREN:
 235                 goto post;
 236         case RESUME_SKIP_NONE:
 237                 break;
 238         default:
 239                 ASSERT(0);
 240         }
 241
 242         if (bp->blk_birth == 0) {
 243                 /*
 244                  * Since this block has a birth time of 0 it must be one of
 245                  * two things: a hole created before the
 246                  * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
 247                  * which has always been a hole in an object.
 248                  *
 249                  * If a file is written sparsely, then the unwritten parts of
 250                  * the file were "always holes" -- that is, they have been
 251                  * holes since this object was allocated.  However, we (and
 252                  * our callers) can not necessarily tell when an object was
 253                  * allocated.  Therefore, if it's possible that this object
 254                  * was freed and then its object number reused, we need to
 255                  * visit all the holes with birth==0.
 256                  *
 257                  * If it isn't possible that the object number was reused,
 258                  * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
 259                  * all the blocks we will visit as part of this traversal,
 260                  * then this hole must have always existed, so we can skip
 261                  * it.  We visit blocks born after (exclusive) td_min_txg.
 262                  *
 263                  * Note that the meta-dnode cannot be reallocated.
 264                  */
 265                 if (!send_holes_without_birth_time &&
 266                     (!td->td_realloc_possible ||
 267                     zb->zb_object == DMU_META_DNODE_OBJECT) &&
 268                     td->td_hole_birth_enabled_txg <= td->td_min_txg)
 269                         return (0);
 270         } else if (bp->blk_birth <= td->td_min_txg) {
 271                 return (0);
 272         }
 273
 274         if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
 275                 uint64_t size = BP_GET_LSIZE(bp);
 276                 mutex_enter(&pd->pd_mtx);
 277                 ASSERT(pd->pd_bytes_fetched >= 0);
 278                 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
 279                         cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
 280                 pd->pd_bytes_fetched -= size;
 281                 cv_broadcast(&pd->pd_cv);
 282                 mutex_exit(&pd->pd_mtx);
 283         }
 284
 285         if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
 286                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 287                 if (err != 0)
 288                         goto post;
 289                 return (0);
 290         }
 291
 292         if (td->td_flags & TRAVERSE_PRE) {
 293                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
 294                     td->td_arg);
 295                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 296                         return (0);
 297                 if (err != 0)
 298                         goto post;
 299         }
 300
 301         if (BP_GET_LEVEL(bp) > 0) {
 302                 uint32_t flags = ARC_FLAG_WAIT;
 303                 int32_t i, ptidx, pidx;
 304                 uint32_t prefetchlimit;
 305                 int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 306                 zbookmark_phys_t *czb;
 307
 308                 ASSERT(!BP_IS_PROTECTED(bp));
 309
 310                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 311                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 312                 if (err != 0)
 313                         goto post;
 314
 315                 czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
 316
 317                 /*
 318                  * When performing a traversal it is beneficial to
 319                  * asynchronously read-ahead the upcoming indirect
 320                  * blocks since they will be needed shortly. However,
 321                  * since a 128k indirect (non-L0) block may contain up
 322                  * to 1024 128-byte block pointers, its preferable to not
 323                  * prefetch them all at once. Issuing a large number of
 324                  * async reads may effect performance, and the earlier
 325                  * the indirect blocks are prefetched the less likely
 326                  * they are to still be resident in the ARC when needed.
 327                  * Therefore, prefetching indirect blocks is limited to
 328                  * zfs_traverse_indirect_prefetch_limit=32 blocks by
 329                  * default.
 330                  *
 331                  * pidx: Index for which next prefetch to be issued.
 332                  * ptidx: Index at which next prefetch to be triggered.
 333                  */
 334                 ptidx = 0;
 335                 pidx = 1;
 336                 prefetchlimit = zfs_traverse_indirect_prefetch_limit;
 337                 for (i = 0; i < epb; i++) {
 338                         if (prefetchlimit && i == ptidx) {
 339                                 ASSERT3S(ptidx, <=, pidx);
 340                                 for (uint32_t  prefetched = 0; pidx < epb &&
 341                                     prefetched < prefetchlimit; pidx++) {
 342                                         SET_BOOKMARK(czb, zb->zb_objset,
 343                                             zb->zb_object, zb->zb_level - 1,
 344                                             zb->zb_blkid * epb + pidx);
 345                                         if (traverse_prefetch_metadata(td,
 346                                             &((blkptr_t *)buf->b_data)[pidx],
 347                                             czb) == B_TRUE) {
 348                                                 prefetched++;
 349                                                 if (prefetched ==
 350                                                     MAX(prefetchlimit / 2, 1))
 351                                                         ptidx = pidx;
 352                                         }
 353                                 }
 354                         }
 355
 356                         /* recursively visitbp() blocks below this */
 357                         SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
 358                             zb->zb_level - 1,
 359                             zb->zb_blkid * epb + i);
 360                         err = traverse_visitbp(td, dnp,
 361                             &((blkptr_t *)buf->b_data)[i], czb);
 362                         if (err != 0)
 363                                 break;
 364                 }
 365
 366                 kmem_free(czb, sizeof (zbookmark_phys_t));
 367
 368         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 369                 uint32_t flags = ARC_FLAG_WAIT;
 370                 uint32_t zio_flags = ZIO_FLAG_CANFAIL;
 371                 int32_t i;
 372                 int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 373                 dnode_phys_t *child_dnp;
 374
 375                 /*
 376                  * dnode blocks might have their bonus buffers encrypted, so
 377                  * we must be careful to honor TRAVERSE_NO_DECRYPT
 378                  */
 379                 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 380                         zio_flags |= ZIO_FLAG_RAW;
 381
 382                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 383                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 384                 if (err != 0)
 385                         goto post;
 386
 387                 child_dnp = buf->b_data;
 388
 389                 for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
 390                         prefetch_dnode_metadata(td, &child_dnp[i],
 391                             zb->zb_objset, zb->zb_blkid * epb + i);
 392                 }
 393
 394                 /* recursively visitbp() blocks below this */
 395                 for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
 396                         err = traverse_dnode(td, bp, &child_dnp[i],
 397                             zb->zb_objset, zb->zb_blkid * epb + i);
 398                         if (err != 0)
 399                                 break;
 400                 }
 401         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 402                 uint32_t zio_flags = ZIO_FLAG_CANFAIL;
 403                 arc_flags_t flags = ARC_FLAG_WAIT;
 404                 objset_phys_t *osp;
 405
 406                 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 407                         zio_flags |= ZIO_FLAG_RAW;
 408
 409                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 410                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 411                 if (err != 0)
 412                         goto post;
 413
 414                 osp = buf->b_data;
 415                 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
 416                     DMU_META_DNODE_OBJECT);
 417                 /*
 418                  * See the block comment above for the goal of this variable.
 419                  * If the maxblkid of the meta-dnode is 0, then we know that
 420                  * we've never had more than DNODES_PER_BLOCK objects in the
 421                  * dataset, which means we can't have reused any object ids.
 422                  */
 423                 if (osp->os_meta_dnode.dn_maxblkid == 0)
 424                         td->td_realloc_possible = B_FALSE;
 425
 426                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 427                         if (OBJSET_BUF_HAS_PROJECTUSED(buf))
 428                                 prefetch_dnode_metadata(td,
 429                                     &osp->os_projectused_dnode,
 430                                     zb->zb_objset, DMU_PROJECTUSED_OBJECT);
 431                         prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
 432                             zb->zb_objset, DMU_GROUPUSED_OBJECT);
 433                         prefetch_dnode_metadata(td, &osp->os_userused_dnode,
 434                             zb->zb_objset, DMU_USERUSED_OBJECT);
 435                 }
 436
 437                 err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset,
 438                     DMU_META_DNODE_OBJECT);
 439                 if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
 440                         if (OBJSET_BUF_HAS_PROJECTUSED(buf))
 441                                 err = traverse_dnode(td, bp,
 442                                     &osp->os_projectused_dnode, zb->zb_objset,
 443                                     DMU_PROJECTUSED_OBJECT);
 444                         if (err == 0)
 445                                 err = traverse_dnode(td, bp,
 446                                     &osp->os_groupused_dnode, zb->zb_objset,
 447                                     DMU_GROUPUSED_OBJECT);
 448                         if (err == 0)
 449                                 err = traverse_dnode(td, bp,
 450                                     &osp->os_userused_dnode, zb->zb_objset,
 451                                     DMU_USERUSED_OBJECT);
 452                 }
 453         }
 454
 455         if (buf)
 456                 arc_buf_destroy(buf, &buf);
 457
 458 post:
 459         if (err == 0 && (td->td_flags & TRAVERSE_POST))
 460                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 461
 462         if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
 463                 /*
 464                  * Ignore this disk error as requested by the HARD flag,
 465                  * and continue traversal.
 466                  */
 467                 err = 0;
 468         }
 469
 470         /*
 471          * If we are stopping here, set td_resume.
 472          */
 473         if (td->td_resume != NULL && err != 0 && !td->td_paused) {
 474                 td->td_resume->zb_objset = zb->zb_objset;
 475                 td->td_resume->zb_object = zb->zb_object;
 476                 td->td_resume->zb_level = 0;
 477                 /*
 478                  * If we have stopped on an indirect block (e.g. due to
 479                  * i/o error), we have not visited anything below it.
 480                  * Set the bookmark to the first level-0 block that we need
 481                  * to visit.  This way, the resuming code does not need to
 482                  * deal with resuming from indirect blocks.
 483                  *
 484                  * Note, if zb_level <= 0, dnp may be NULL, so we don't want
 485                  * to dereference it.
 486                  */
 487                 td->td_resume->zb_blkid = zb->zb_blkid;
 488                 if (zb->zb_level > 0) {
 489                         td->td_resume->zb_blkid <<= zb->zb_level *
 490                             (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
 491                 }
 492                 td->td_paused = B_TRUE;
 493         }
 494
 495         return (err);
 496 }
 497
 498 static void
 499 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
 500     uint64_t objset, uint64_t object)
 501 {
 502         int j;
 503         zbookmark_phys_t czb;
 504
 505         for (j = 0; j < dnp->dn_nblkptr; j++) {
 506                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 507                 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
 508         }
 509
 510         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 511                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 512                 traverse_prefetch_metadata(td, DN_SPILL_BLKPTR(dnp), &czb);
 513         }
 514 }
 515
 516 static int
 517 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp,
 518     uint64_t objset, uint64_t object)
 519 {
 520         int j, err = 0;
 521         zbookmark_phys_t czb;
 522
 523         if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
 524             object < td->td_resume->zb_object)
 525                 return (0);
 526
 527         if (td->td_flags & TRAVERSE_PRE) {
 528                 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
 529                     ZB_DNODE_BLKID);
 530                 err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
 531                     td->td_arg);
 532                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 533                         return (0);
 534                 if (err != 0)
 535                         return (err);
 536         }
 537
 538         for (j = 0; j < dnp->dn_nblkptr; j++) {
 539                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 540                 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
 541                 if (err != 0)
 542                         break;
 543         }
 544
 545         if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
 546                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 547                 err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
 548         }
 549
 550         if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
 551                 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
 552                     ZB_DNODE_BLKID);
 553                 err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
 554                     td->td_arg);
 555                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 556                         return (0);
 557                 if (err != 0)
 558                         return (err);
 559         }
 560         return (err);
 561 }
 562
 563 static int
 564 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 565     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 566 {
 567         (void) zilog, (void) dnp;
 568         prefetch_data_t *pfd = arg;
 569         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
 570         arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
 571             ARC_FLAG_PRESCIENT_PREFETCH;
 572
 573         ASSERT(pfd->pd_bytes_fetched >= 0);
 574         if (zb->zb_level == ZB_DNODE_LEVEL)
 575                 return (0);
 576         if (pfd->pd_cancel)
 577                 return (SET_ERROR(EINTR));
 578
 579         if (!prefetch_needed(pfd, bp))
 580                 return (0);
 581
 582         mutex_enter(&pfd->pd_mtx);
 583         while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
 584                 cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx);
 585         pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
 586         cv_broadcast(&pfd->pd_cv);
 587         mutex_exit(&pfd->pd_mtx);
 588
 589         if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 590                 zio_flags |= ZIO_FLAG_RAW;
 591
 592         (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 593             zio_flags, &aflags, zb);
 594
 595         return (0);
 596 }
 597
 598 static void
 599 traverse_prefetch_thread(void *arg)
 600 {
 601         traverse_data_t *td_main = arg;
 602         traverse_data_t td = *td_main;
 603         zbookmark_phys_t czb;
 604         fstrans_cookie_t cookie = spl_fstrans_mark();
 605
 606         td.td_func = traverse_prefetcher;
 607         td.td_arg = td_main->td_pfd;
 608         td.td_pfd = NULL;
 609         td.td_resume = &td_main->td_pfd->pd_resume;
 610
 611         SET_BOOKMARK(&czb, td.td_objset,
 612             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 613         (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
 614
 615         mutex_enter(&td_main->td_pfd->pd_mtx);
 616         td_main->td_pfd->pd_exited = B_TRUE;
 617         cv_broadcast(&td_main->td_pfd->pd_cv);
 618         mutex_exit(&td_main->td_pfd->pd_mtx);
 619         spl_fstrans_unmark(cookie);
 620 }
 621
 622 /*
 623  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 624  * in syncing context).
 625  */
 626 static int
 627 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
 628     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 629     blkptr_cb_t func, void *arg)
 630 {
 631         traverse_data_t *td;
 632         prefetch_data_t *pd;
 633         zbookmark_phys_t *czb;
 634         int err;
 635
 636         ASSERT(ds == NULL || objset == ds->ds_object);
 637         ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
 638
 639         td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP);
 640         pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP);
 641         czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
 642
 643         td->td_spa = spa;
 644         td->td_objset = objset;
 645         td->td_rootbp = rootbp;
 646         td->td_min_txg = txg_start;
 647         td->td_resume = resume;
 648         td->td_func = func;
 649         td->td_arg = arg;
 650         td->td_pfd = pd;
 651         td->td_flags = flags;
 652         td->td_paused = B_FALSE;
 653         td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
 654
 655         if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
 656                 VERIFY(spa_feature_enabled_txg(spa,
 657                     SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg));
 658         } else {
 659                 td->td_hole_birth_enabled_txg = UINT64_MAX;
 660         }
 661
 662         pd->pd_flags = flags;
 663         if (resume != NULL)
 664                 pd->pd_resume = *resume;
 665         mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
 666         cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
 667
 668         SET_BOOKMARK(czb, td->td_objset,
 669             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 670
 671         /* See comment on ZIL traversal in dsl_scan_visitds. */
 672         if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
 673                 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
 674                 uint32_t flags = ARC_FLAG_WAIT;
 675                 objset_phys_t *osp;
 676                 arc_buf_t *buf;
 677                 ASSERT(!BP_IS_REDACTED(rootbp));
 678
 679                 if ((td->td_flags & TRAVERSE_NO_DECRYPT) &&
 680                     BP_IS_PROTECTED(rootbp))
 681                         zio_flags |= ZIO_FLAG_RAW;
 682
 683                 err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func,
 684                     &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb);
 685                 if (err != 0) {
 686                         /*
 687                          * If both TRAVERSE_HARD and TRAVERSE_PRE are set,
 688                          * continue to visitbp so that td_func can be called
 689                          * in pre stage, and err will reset to zero.
 690                          */
 691                         if (!(td->td_flags & TRAVERSE_HARD) ||
 692                             !(td->td_flags & TRAVERSE_PRE))
 693                                 goto out;
 694                 } else {
 695                         osp = buf->b_data;
 696                         traverse_zil(td, &osp->os_zil_header);
 697                         arc_buf_destroy(buf, &buf);
 698                 }
 699         }
 700
 701         if (!(flags & TRAVERSE_PREFETCH_DATA) ||
 702             taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread,
 703             td, TQ_NOQUEUE) == TASKQID_INVALID)
 704                 pd->pd_exited = B_TRUE;
 705
 706         err = traverse_visitbp(td, NULL, rootbp, czb);
 707
 708         mutex_enter(&pd->pd_mtx);
 709         pd->pd_cancel = B_TRUE;
 710         cv_broadcast(&pd->pd_cv);
 711         while (!pd->pd_exited)
 712                 cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
 713         mutex_exit(&pd->pd_mtx);
 714 out:
 715         mutex_destroy(&pd->pd_mtx);
 716         cv_destroy(&pd->pd_cv);
 717
 718         kmem_free(czb, sizeof (zbookmark_phys_t));
 719         kmem_free(pd, sizeof (struct prefetch_data));
 720         kmem_free(td, sizeof (struct traverse_data));
 721
 722         return (err);
 723 }
 724
 725 /*
 726  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 727  * in syncing context).
 728  */
 729 int
 730 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
 731     zbookmark_phys_t *resume,
 732     int flags, blkptr_cb_t func, void *arg)
 733 {
 734         return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
 735             &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
 736 }
 737
 738 int
 739 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
 740     int flags, blkptr_cb_t func, void *arg)
 741 {
 742         return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
 743 }
 744
 745 int
 746 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
 747     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 748     blkptr_cb_t func, void *arg)
 749 {
 750         return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
 751             blkptr, txg_start, resume, flags, func, arg));
 752 }
 753
 754 /*
 755  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
 756  */
 757 int
 758 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
 759     blkptr_cb_t func, void *arg)
 760 {
 761         int err;
 762         dsl_pool_t *dp = spa_get_dsl(spa);
 763         objset_t *mos = dp->dp_meta_objset;
 764         boolean_t hard = (flags & TRAVERSE_HARD);
 765
 766         /* visit the MOS */
 767         err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
 768             txg_start, NULL, flags, func, arg);
 769         if (err != 0)
 770                 return (err);
 771
 772         /* visit each dataset */
 773         for (uint64_t obj = 1; err == 0;
 774             err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
 775                 dmu_object_info_t doi;
 776
 777                 err = dmu_object_info(mos, obj, &doi);
 778                 if (err != 0) {
 779                         if (hard)
 780                                 continue;
 781                         break;
 782                 }
 783
 784                 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
 785                         dsl_dataset_t *ds;
 786                         uint64_t txg = txg_start;
 787
 788                         dsl_pool_config_enter(dp, FTAG);
 789                         err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
 790                         dsl_pool_config_exit(dp, FTAG);
 791                         if (err != 0) {
 792                                 if (hard)
 793                                         continue;
 794                                 break;
 795                         }
 796                         if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
 797                                 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
 798                         err = traverse_dataset(ds, txg, flags, func, arg);
 799                         dsl_dataset_rele(ds, FTAG);
 800                         if (err != 0)
 801                                 break;
 802                 }
 803         }
 804         if (err == ESRCH)
 805                 err = 0;
 806         return (err);
 807 }
 808
 809 EXPORT_SYMBOL(traverse_dataset);
 810 EXPORT_SYMBOL(traverse_pool);
 811
 812 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW,
 813         "Max number of bytes to prefetch");
 814
 815 ZFS_MODULE_PARAM(zfs, zfs_, traverse_indirect_prefetch_limit, INT, ZMOD_RW,
 816         "Traverse prefetch number of blocks pointed by indirect block");
 817
 818 #if defined(_KERNEL)
 819 module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644);
 820 MODULE_PARM_DESC(ignore_hole_birth,
 821         "Alias for send_holes_without_birth_time");
 822 #endif
 823
 824 /* CSTYLED */
 825 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW,
 826         "Ignore hole_birth txg for zfs send");