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, 2016 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/zio.h>
  35 #include <sys/dmu_impl.h>
  36 #include <sys/sa.h>
  37 #include <sys/sa_impl.h>
  38 #include <sys/callb.h>
  39 #include <sys/zfeature.h>
  40 
  41 int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;    /* 50MB */
  42 
  43 typedef struct prefetch_data {
  44         kmutex_t pd_mtx;
  45         kcondvar_t pd_cv;
  46         int32_t pd_bytes_fetched;
  47         int pd_flags;
  48         boolean_t pd_cancel;
  49         boolean_t pd_exited;
  50         zbookmark_phys_t pd_resume;
  51 } prefetch_data_t;
  52 
  53 typedef struct traverse_data {
  54         spa_t *td_spa;
  55         uint64_t td_objset;
  56         blkptr_t *td_rootbp;
  57         uint64_t td_min_txg;
  58         zbookmark_phys_t *td_resume;
  59         int td_flags;
  60         prefetch_data_t *td_pfd;
  61         boolean_t td_paused;
  62         uint64_t td_hole_birth_enabled_txg;
  63         blkptr_cb_t *td_func;
  64         void *td_arg;
  65         boolean_t td_realloc_possible;
  66 } traverse_data_t;
  67 
  68 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
  69     uint64_t objset, uint64_t object);
  70 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
  71     uint64_t objset, uint64_t object);
  72 
  73 static int
  74 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
  75 {
  76         traverse_data_t *td = arg;
  77         zbookmark_phys_t zb;
  78 
  79         if (BP_IS_HOLE(bp))
  80                 return (0);
  81 
  82         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
  83                 return (0);
  84 
  85         SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
  86             bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
  87 
  88         (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
  89 
  90         return (0);
  91 }
  92 
  93 static int
  94 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
  95 {
  96         traverse_data_t *td = arg;
  97 
  98         if (lrc->lrc_txtype == TX_WRITE) {
  99                 lr_write_t *lr = (lr_write_t *)lrc;
 100                 blkptr_t *bp = &lr->lr_blkptr;
 101                 zbookmark_phys_t zb;
 102 
 103                 if (BP_IS_HOLE(bp))
 104                         return (0);
 105 
 106                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 107                         return (0);
 108 
 109                 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
 110                     ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
 111 
 112                 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
 113                     td->td_arg);
 114         }
 115         return (0);
 116 }
 117 
 118 static void
 119 traverse_zil(traverse_data_t *td, zil_header_t *zh)
 120 {
 121         uint64_t claim_txg = zh->zh_claim_txg;
 122         zilog_t *zilog;
 123 
 124         /*
 125          * We only want to visit blocks that have been claimed but not yet
 126          * replayed; plus, in read-only mode, blocks that are already stable.
 127          */
 128         if (claim_txg == 0 && spa_writeable(td->td_spa))
 129                 return;
 130 
 131         zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
 132 
 133         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
 134             claim_txg);
 135 
 136         zil_free(zilog);
 137 }
 138 
 139 typedef enum resume_skip {
 140         RESUME_SKIP_ALL,
 141         RESUME_SKIP_NONE,
 142         RESUME_SKIP_CHILDREN
 143 } resume_skip_t;
 144 
 145 /*
 146  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
 147  * the block indicated by zb does not need to be visited at all. Returns
 148  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
 149  * resume point. This indicates that this block should be visited but not its
 150  * children (since they must have been visited in a previous traversal).
 151  * Otherwise returns RESUME_SKIP_NONE.
 152  */
 153 static resume_skip_t
 154 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
 155     const zbookmark_phys_t *zb)
 156 {
 157         if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
 158                 /*
 159                  * If we already visited this bp & everything below,
 160                  * don't bother doing it again.
 161                  */
 162                 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
 163                         return (RESUME_SKIP_ALL);
 164 
 165                 /*
 166                  * If we found the block we're trying to resume from, zero
 167                  * the bookmark out to indicate that we have resumed.
 168                  */
 169                 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
 170                         bzero(td->td_resume, sizeof (*zb));
 171                         if (td->td_flags & TRAVERSE_POST)
 172                                 return (RESUME_SKIP_CHILDREN);
 173                 }
 174         }
 175         return (RESUME_SKIP_NONE);
 176 }
 177 
 178 static void
 179 traverse_prefetch_metadata(traverse_data_t *td,
 180     const blkptr_t *bp, const zbookmark_phys_t *zb)
 181 {
 182         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 183 
 184         if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
 185                 return;
 186         /*
 187          * If we are in the process of resuming, don't prefetch, because
 188          * some children will not be needed (and in fact may have already
 189          * been freed).
 190          */
 191         if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
 192                 return;
 193         if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
 194                 return;
 195         if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
 196                 return;
 197 
 198         (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
 199             ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 200 }
 201 
 202 static boolean_t
 203 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
 204 {
 205         ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
 206         if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
 207             BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
 208                 return (B_FALSE);
 209         return (B_TRUE);
 210 }
 211 
 212 static int
 213 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
 214     const blkptr_t *bp, const zbookmark_phys_t *zb)
 215 {
 216         zbookmark_phys_t czb;
 217         int err = 0;
 218         arc_buf_t *buf = NULL;
 219         prefetch_data_t *pd = td->td_pfd;
 220         boolean_t hard = td->td_flags & TRAVERSE_HARD;
 221 
 222         switch (resume_skip_check(td, dnp, zb)) {
 223         case RESUME_SKIP_ALL:
 224                 return (0);
 225         case RESUME_SKIP_CHILDREN:
 226                 goto post;
 227         case RESUME_SKIP_NONE:
 228                 break;
 229         default:
 230                 ASSERT(0);
 231         }
 232 
 233         if (bp->blk_birth == 0) {
 234                 /*
 235                  * Since this block has a birth time of 0 it must be one of
 236                  * two things: a hole created before the
 237                  * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
 238                  * which has always been a hole in an object.
 239                  *
 240                  * If a file is written sparsely, then the unwritten parts of
 241                  * the file were "always holes" -- that is, they have been
 242                  * holes since this object was allocated.  However, we (and
 243                  * our callers) can not necessarily tell when an object was
 244                  * allocated.  Therefore, if it's possible that this object
 245                  * was freed and then its object number reused, we need to
 246                  * visit all the holes with birth==0.
 247                  *
 248                  * If it isn't possible that the object number was reused,
 249                  * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
 250                  * all the blocks we will visit as part of this traversal,
 251                  * then this hole must have always existed, so we can skip
 252                  * it.  We visit blocks born after (exclusive) td_min_txg.
 253                  *
 254                  * Note that the meta-dnode cannot be reallocated.
 255                  */
 256                 if ((!td->td_realloc_possible ||
 257                     zb->zb_object == DMU_META_DNODE_OBJECT) &&
 258                     td->td_hole_birth_enabled_txg <= td->td_min_txg)
 259                         return (0);
 260         } else if (bp->blk_birth <= td->td_min_txg) {
 261                 return (0);
 262         }
 263 
 264         if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
 265                 uint64_t size = BP_GET_LSIZE(bp);
 266                 mutex_enter(&pd->pd_mtx);
 267                 ASSERT(pd->pd_bytes_fetched >= 0);
 268                 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
 269                         cv_wait(&pd->pd_cv, &pd->pd_mtx);
 270                 pd->pd_bytes_fetched -= size;
 271                 cv_broadcast(&pd->pd_cv);
 272                 mutex_exit(&pd->pd_mtx);
 273         }
 274 
 275         if (BP_IS_HOLE(bp)) {
 276                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 277                 if (err != 0)
 278                         goto post;
 279                 return (0);
 280         }
 281 
 282         if (td->td_flags & TRAVERSE_PRE) {
 283                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
 284                     td->td_arg);
 285                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 286                         return (0);
 287                 if (err != 0)
 288                         goto post;
 289         }
 290 
 291         if (BP_GET_LEVEL(bp) > 0) {
 292                 arc_flags_t flags = ARC_FLAG_WAIT;
 293                 int i;
 294                 blkptr_t *cbp;
 295                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 296 
 297                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 298                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 299                 if (err != 0)
 300                         goto post;
 301                 cbp = buf->b_data;
 302 
 303                 for (i = 0; i < epb; i++) {
 304                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 305                             zb->zb_level - 1,
 306                             zb->zb_blkid * epb + i);
 307                         traverse_prefetch_metadata(td, &cbp[i], &czb);
 308                 }
 309 
 310                 /* recursively visitbp() blocks below this */
 311                 for (i = 0; i < epb; i++) {
 312                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 313                             zb->zb_level - 1,
 314                             zb->zb_blkid * epb + i);
 315                         err = traverse_visitbp(td, dnp, &cbp[i], &czb);
 316                         if (err != 0)
 317                                 break;
 318                 }
 319         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 320                 arc_flags_t flags = ARC_FLAG_WAIT;
 321                 int i;
 322                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 323 
 324                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 325                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 326                 if (err != 0)
 327                         goto post;
 328                 dnode_phys_t *child_dnp = buf->b_data;
 329 
 330                 for (i = 0; i < epb; i++) {
 331                         prefetch_dnode_metadata(td, &child_dnp[i],
 332                             zb->zb_objset, zb->zb_blkid * epb + i);
 333                 }
 334 
 335                 /* recursively visitbp() blocks below this */
 336                 for (i = 0; i < epb; i++) {
 337                         err = traverse_dnode(td, &child_dnp[i],
 338                             zb->zb_objset, zb->zb_blkid * epb + i);
 339                         if (err != 0)
 340                                 break;
 341                 }
 342         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 343                 arc_flags_t flags = ARC_FLAG_WAIT;
 344 
 345                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 346                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 347                 if (err != 0)
 348                         goto post;
 349 
 350                 objset_phys_t *osp = buf->b_data;
 351                 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
 352                     DMU_META_DNODE_OBJECT);
 353                 /*
 354                  * See the block comment above for the goal of this variable.
 355                  * If the maxblkid of the meta-dnode is 0, then we know that
 356                  * we've never had more than DNODES_PER_BLOCK objects in the
 357                  * dataset, which means we can't have reused any object ids.
 358                  */
 359                 if (osp->os_meta_dnode.dn_maxblkid == 0)
 360                         td->td_realloc_possible = B_FALSE;
 361 
 362                 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 363                         prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
 364                             zb->zb_objset, DMU_GROUPUSED_OBJECT);
 365                         prefetch_dnode_metadata(td, &osp->os_userused_dnode,
 366                             zb->zb_objset, DMU_USERUSED_OBJECT);
 367                 }
 368 
 369                 err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
 370                     DMU_META_DNODE_OBJECT);
 371                 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 372                         err = traverse_dnode(td, &osp->os_groupused_dnode,
 373                             zb->zb_objset, DMU_GROUPUSED_OBJECT);
 374                 }
 375                 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 376                         err = traverse_dnode(td, &osp->os_userused_dnode,
 377                             zb->zb_objset, DMU_USERUSED_OBJECT);
 378                 }
 379         }
 380 
 381         if (buf)
 382                 (void) arc_buf_remove_ref(buf, &buf);
 383 
 384 post:
 385         if (err == 0 && (td->td_flags & TRAVERSE_POST))
 386                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 387 
 388         if (hard && (err == EIO || err == ECKSUM)) {
 389                 /*
 390                  * Ignore this disk error as requested by the HARD flag,
 391                  * and continue traversal.
 392                  */
 393                 err = 0;
 394         }
 395 
 396         /*
 397          * If we are stopping here, set td_resume.
 398          */
 399         if (td->td_resume != NULL && err != 0 && !td->td_paused) {
 400                 td->td_resume->zb_objset = zb->zb_objset;
 401                 td->td_resume->zb_object = zb->zb_object;
 402                 td->td_resume->zb_level = 0;
 403                 /*
 404                  * If we have stopped on an indirect block (e.g. due to
 405                  * i/o error), we have not visited anything below it.
 406                  * Set the bookmark to the first level-0 block that we need
 407                  * to visit.  This way, the resuming code does not need to
 408                  * deal with resuming from indirect blocks.
 409                  *
 410                  * Note, if zb_level <= 0, dnp may be NULL, so we don't want
 411                  * to dereference it.
 412                  */
 413                 td->td_resume->zb_blkid = zb->zb_blkid;
 414                 if (zb->zb_level > 0) {
 415                         td->td_resume->zb_blkid <<= zb->zb_level *
 416                             (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
 417                 }
 418                 td->td_paused = B_TRUE;
 419         }
 420 
 421         return (err);
 422 }
 423 
 424 static void
 425 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
 426     uint64_t objset, uint64_t object)
 427 {
 428         int j;
 429         zbookmark_phys_t czb;
 430 
 431         for (j = 0; j < dnp->dn_nblkptr; j++) {
 432                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 433                 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
 434         }
 435 
 436         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 437                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 438                 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
 439         }
 440 }
 441 
 442 static int
 443 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
 444     uint64_t objset, uint64_t object)
 445 {
 446         int j, err = 0;
 447         zbookmark_phys_t czb;
 448 
 449         if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
 450             object < td->td_resume->zb_object)
 451                 return (0);
 452 
 453         if (td->td_flags & TRAVERSE_PRE) {
 454                 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
 455                     ZB_DNODE_BLKID);
 456                 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
 457                     td->td_arg);
 458                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 459                         return (0);
 460                 if (err != 0)
 461                         return (err);
 462         }
 463 
 464         for (j = 0; j < dnp->dn_nblkptr; j++) {
 465                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 466                 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
 467                 if (err != 0)
 468                         break;
 469         }
 470 
 471         if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
 472                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 473                 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
 474         }
 475 
 476         if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
 477                 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
 478                     ZB_DNODE_BLKID);
 479                 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
 480                     td->td_arg);
 481                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 482                         return (0);
 483                 if (err != 0)
 484                         return (err);
 485         }
 486         return (err);
 487 }
 488 
 489 /* ARGSUSED */
 490 static int
 491 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 492     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 493 {
 494         prefetch_data_t *pfd = arg;
 495         arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 496 
 497         ASSERT(pfd->pd_bytes_fetched >= 0);
 498         if (bp == NULL)
 499                 return (0);
 500         if (pfd->pd_cancel)
 501                 return (SET_ERROR(EINTR));
 502 
 503         if (!prefetch_needed(pfd, bp))
 504                 return (0);
 505 
 506         mutex_enter(&pfd->pd_mtx);
 507         while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
 508                 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
 509         pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
 510         cv_broadcast(&pfd->pd_cv);
 511         mutex_exit(&pfd->pd_mtx);
 512 
 513         (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 514             ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
 515 
 516         return (0);
 517 }
 518 
 519 static void
 520 traverse_prefetch_thread(void *arg)
 521 {
 522         traverse_data_t *td_main = arg;
 523         traverse_data_t td = *td_main;
 524         zbookmark_phys_t czb;
 525 
 526         td.td_func = traverse_prefetcher;
 527         td.td_arg = td_main->td_pfd;
 528         td.td_pfd = NULL;
 529         td.td_resume = &td_main->td_pfd->pd_resume;
 530 
 531         SET_BOOKMARK(&czb, td.td_objset,
 532             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 533         (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
 534 
 535         mutex_enter(&td_main->td_pfd->pd_mtx);
 536         td_main->td_pfd->pd_exited = B_TRUE;
 537         cv_broadcast(&td_main->td_pfd->pd_cv);
 538         mutex_exit(&td_main->td_pfd->pd_mtx);
 539 }
 540 
 541 /*
 542  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 543  * in syncing context).
 544  */
 545 static int
 546 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
 547     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 548     blkptr_cb_t func, void *arg)
 549 {
 550         traverse_data_t td;
 551         prefetch_data_t pd = { 0 };
 552         zbookmark_phys_t czb;
 553         int err;
 554 
 555         ASSERT(ds == NULL || objset == ds->ds_object);
 556         ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
 557 
 558         td.td_spa = spa;
 559         td.td_objset = objset;
 560         td.td_rootbp = rootbp;
 561         td.td_min_txg = txg_start;
 562         td.td_resume = resume;
 563         td.td_func = func;
 564         td.td_arg = arg;
 565         td.td_pfd = &pd;
 566         td.td_flags = flags;
 567         td.td_paused = B_FALSE;
 568         td.td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
 569 
 570         if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
 571                 VERIFY(spa_feature_enabled_txg(spa,
 572                     SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
 573         } else {
 574                 td.td_hole_birth_enabled_txg = UINT64_MAX;
 575         }
 576 
 577         pd.pd_flags = flags;
 578         if (resume != NULL)
 579                 pd.pd_resume = *resume;
 580         mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
 581         cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
 582 
 583         /* See comment on ZIL traversal in dsl_scan_visitds. */
 584         if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
 585                 arc_flags_t flags = ARC_FLAG_WAIT;
 586                 objset_phys_t *osp;
 587                 arc_buf_t *buf;
 588 
 589                 err = arc_read(NULL, td.td_spa, rootbp,
 590                     arc_getbuf_func, &buf,
 591                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
 592                 if (err != 0)
 593                         return (err);
 594 
 595                 osp = buf->b_data;
 596                 traverse_zil(&td, &osp->os_zil_header);
 597                 (void) arc_buf_remove_ref(buf, &buf);
 598         }
 599 
 600         if (!(flags & TRAVERSE_PREFETCH_DATA) ||
 601             0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
 602             &td, TQ_NOQUEUE))
 603                 pd.pd_exited = B_TRUE;
 604 
 605         SET_BOOKMARK(&czb, td.td_objset,
 606             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 607         err = traverse_visitbp(&td, NULL, rootbp, &czb);
 608 
 609         mutex_enter(&pd.pd_mtx);
 610         pd.pd_cancel = B_TRUE;
 611         cv_broadcast(&pd.pd_cv);
 612         while (!pd.pd_exited)
 613                 cv_wait(&pd.pd_cv, &pd.pd_mtx);
 614         mutex_exit(&pd.pd_mtx);
 615 
 616         mutex_destroy(&pd.pd_mtx);
 617         cv_destroy(&pd.pd_cv);
 618 
 619         return (err);
 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 int
 627 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
 628     zbookmark_phys_t *resume,
 629     int flags, blkptr_cb_t func, void *arg)
 630 {
 631         return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
 632             &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
 633 }
 634 
 635 int
 636 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
 637     int flags, blkptr_cb_t func, void *arg)
 638 {
 639         return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
 640 }
 641 
 642 int
 643 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
 644     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 645     blkptr_cb_t func, void *arg)
 646 {
 647         return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
 648             blkptr, txg_start, resume, flags, func, arg));
 649 }
 650 
 651 /*
 652  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
 653  */
 654 int
 655 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
 656     blkptr_cb_t func, void *arg)
 657 {
 658         int err;
 659         dsl_pool_t *dp = spa_get_dsl(spa);
 660         objset_t *mos = dp->dp_meta_objset;
 661         boolean_t hard = (flags & TRAVERSE_HARD);
 662 
 663         /* visit the MOS */
 664         err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
 665             txg_start, NULL, flags, func, arg);
 666         if (err != 0)
 667                 return (err);
 668 
 669         /* visit each dataset */
 670         for (uint64_t obj = 1; err == 0;
 671             err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
 672                 dmu_object_info_t doi;
 673 
 674                 err = dmu_object_info(mos, obj, &doi);
 675                 if (err != 0) {
 676                         if (hard)
 677                                 continue;
 678                         break;
 679                 }
 680 
 681                 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
 682                         dsl_dataset_t *ds;
 683                         uint64_t txg = txg_start;
 684 
 685                         dsl_pool_config_enter(dp, FTAG);
 686                         err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
 687                         dsl_pool_config_exit(dp, FTAG);
 688                         if (err != 0) {
 689                                 if (hard)
 690                                         continue;
 691                                 break;
 692                         }
 693                         if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
 694                                 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
 695                         err = traverse_dataset(ds, txg, flags, func, arg);
 696                         dsl_dataset_rele(ds, FTAG);
 697                         if (err != 0)
 698                                 break;
 699                 }
 700         }
 701         if (err == ESRCH)
 702                 err = 0;
 703         return (err);
 704 }