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, 2014 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 int zfs_pd_blks_max = 100;
  42
  43 typedef struct prefetch_data {
  44         kmutex_t pd_mtx;
  45         kcondvar_t pd_cv;
  46         int pd_blks_max;
  47         int pd_blks_fetched;
  48         int pd_flags;
  49         boolean_t pd_cancel;
  50         boolean_t pd_exited;
  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         blkptr_cb_t *td_func;
  63         void *td_arg;
  64 } traverse_data_t;
  65
  66 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
  67     uint64_t objset, uint64_t object);
  68 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
  69     uint64_t objset, uint64_t object);
  70
  71 static int
  72 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
  73 {
  74         traverse_data_t *td = arg;
  75         zbookmark_phys_t zb;
  76
  77         if (BP_IS_HOLE(bp))
  78                 return (0);
  79
  80         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
  81                 return (0);
  82
  83         SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
  84             bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
  85
  86         (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
  87
  88         return (0);
  89 }
  90
  91 static int
  92 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
  93 {
  94         traverse_data_t *td = arg;
  95
  96         if (lrc->lrc_txtype == TX_WRITE) {
  97                 lr_write_t *lr = (lr_write_t *)lrc;
  98                 blkptr_t *bp = &lr->lr_blkptr;
  99                 zbookmark_phys_t zb;
 100
 101                 if (BP_IS_HOLE(bp))
 102                         return (0);
 103
 104                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 105                         return (0);
 106
 107                 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
 108                     ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
 109
 110                 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
 111                     td->td_arg);
 112         }
 113         return (0);
 114 }
 115
 116 static void
 117 traverse_zil(traverse_data_t *td, zil_header_t *zh)
 118 {
 119         uint64_t claim_txg = zh->zh_claim_txg;
 120         zilog_t *zilog;
 121
 122         /*
 123          * We only want to visit blocks that have been claimed but not yet
 124          * replayed; plus, in read-only mode, blocks that are already stable.
 125          */
 126         if (claim_txg == 0 && spa_writeable(td->td_spa))
 127                 return;
 128
 129         zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
 130
 131         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
 132             claim_txg);
 133
 134         zil_free(zilog);
 135 }
 136
 137 typedef enum resume_skip {
 138         RESUME_SKIP_ALL,
 139         RESUME_SKIP_NONE,
 140         RESUME_SKIP_CHILDREN
 141 } resume_skip_t;
 142
 143 /*
 144  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
 145  * the block indicated by zb does not need to be visited at all. Returns
 146  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
 147  * resume point. This indicates that this block should be visited but not its
 148  * children (since they must have been visited in a previous traversal).
 149  * Otherwise returns RESUME_SKIP_NONE.
 150  */
 151 static resume_skip_t
 152 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
 153     const zbookmark_phys_t *zb)
 154 {
 155         if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
 156                 /*
 157                  * If we already visited this bp & everything below,
 158                  * don't bother doing it again.
 159                  */
 160                 if (zbookmark_is_before(dnp, zb, td->td_resume))
 161                         return (RESUME_SKIP_ALL);
 162
 163                 /*
 164                  * If we found the block we're trying to resume from, zero
 165                  * the bookmark out to indicate that we have resumed.
 166                  */
 167                 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
 168                         bzero(td->td_resume, sizeof (*zb));
 169                         if (td->td_flags & TRAVERSE_POST)
 170                                 return (RESUME_SKIP_CHILDREN);
 171                 }
 172         }
 173         return (RESUME_SKIP_NONE);
 174 }
 175
 176 static void
 177 traverse_prefetch_metadata(traverse_data_t *td,
 178     const blkptr_t *bp, const zbookmark_phys_t *zb)
 179 {
 180         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 181
 182         if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
 183                 return;
 184         /*
 185          * If we are in the process of resuming, don't prefetch, because
 186          * some children will not be needed (and in fact may have already
 187          * been freed).
 188          */
 189         if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
 190                 return;
 191         if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
 192                 return;
 193         if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
 194                 return;
 195
 196         (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
 197             ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 198 }
 199
 200 static boolean_t
 201 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
 202 {
 203         ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
 204         if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
 205             BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
 206                 return (B_FALSE);
 207         return (B_TRUE);
 208 }
 209
 210 static int
 211 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
 212     const blkptr_t *bp, const zbookmark_phys_t *zb)
 213 {
 214         int err = 0;
 215         arc_buf_t *buf = NULL;
 216
 217         switch (resume_skip_check(td, dnp, zb)) {
 218         case RESUME_SKIP_ALL:
 219                 return (0);
 220         case RESUME_SKIP_CHILDREN:
 221                 goto post;
 222         case RESUME_SKIP_NONE:
 223                 break;
 224         default:
 225                 ASSERT(0);
 226         }
 227
 228         if (bp->blk_birth == 0) {
 229                 if (spa_feature_is_active(td->td_spa, SPA_FEATURE_HOLE_BIRTH)) {
 230                         /*
 231                          * Since this block has a birth time of 0 it must be a
 232                          * hole created before the SPA_FEATURE_HOLE_BIRTH
 233                          * feature was enabled.  If SPA_FEATURE_HOLE_BIRTH
 234                          * was enabled before the min_txg for this traveral we
 235                          * know the hole must have been created before the
 236                          * min_txg for this traveral, so we can skip it. If
 237                          * SPA_FEATURE_HOLE_BIRTH was enabled after the min_txg
 238                          * for this traveral we cannot tell if the hole was
 239                          * created before or after the min_txg for this
 240                          * traversal, so we cannot skip it.
 241                          */
 242                         uint64_t hole_birth_enabled_txg;
 243                         VERIFY(spa_feature_enabled_txg(td->td_spa,
 244                             SPA_FEATURE_HOLE_BIRTH, &hole_birth_enabled_txg));
 245                         if (hole_birth_enabled_txg < td->td_min_txg)
 246                                 return (0);
 247                 }
 248         } else if (bp->blk_birth <= td->td_min_txg) {
 249                 return (0);
 250         }
 251
 252         if (td->td_pfd != NULL && !td->td_pfd->pd_exited &&
 253             prefetch_needed(td->td_pfd, bp)) {
 254                 mutex_enter(&td->td_pfd->pd_mtx);
 255                 ASSERT(td->td_pfd->pd_blks_fetched >= 0);
 256                 while (td->td_pfd->pd_blks_fetched == 0 &&
 257                     !td->td_pfd->pd_exited)
 258                         cv_wait(&td->td_pfd->pd_cv, &td->td_pfd->pd_mtx);
 259                 td->td_pfd->pd_blks_fetched--;
 260                 cv_broadcast(&td->td_pfd->pd_cv);
 261                 mutex_exit(&td->td_pfd->pd_mtx);
 262         }
 263
 264         if (BP_IS_HOLE(bp)) {
 265                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 266                 if (err != 0)
 267                         goto post;
 268                 return (0);
 269         }
 270
 271         if (td->td_flags & TRAVERSE_PRE) {
 272                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
 273                     td->td_arg);
 274                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 275                         return (0);
 276                 if (err != 0)
 277                         goto post;
 278         }
 279
 280         if (BP_GET_LEVEL(bp) > 0) {
 281                 uint32_t flags = ARC_WAIT;
 282                 int32_t i;
 283                 int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 284                 zbookmark_phys_t *czb;
 285
 286                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 287                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 288                 if (err != 0)
 289                         goto post;
 290
 291                 czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_PUSHPAGE);
 292
 293                 for (i = 0; i < epb; i++) {
 294                         SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
 295                             zb->zb_level - 1,
 296                             zb->zb_blkid * epb + i);
 297                         traverse_prefetch_metadata(td,
 298                             &((blkptr_t *)buf->b_data)[i], czb);
 299                 }
 300
 301                 /* recursively visitbp() blocks below this */
 302                 for (i = 0; i < epb; i++) {
 303                         SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
 304                             zb->zb_level - 1,
 305                             zb->zb_blkid * epb + i);
 306                         err = traverse_visitbp(td, dnp,
 307                             &((blkptr_t *)buf->b_data)[i], czb);
 308                         if (err != 0)
 309                                 break;
 310                 }
 311
 312                 kmem_free(czb, sizeof (zbookmark_phys_t));
 313
 314         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 315                 uint32_t flags = ARC_WAIT;
 316                 int32_t i;
 317                 int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 318
 319                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 320                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 321                 if (err != 0)
 322                         goto post;
 323                 dnp = buf->b_data;
 324
 325                 for (i = 0; i < epb; i++) {
 326                         prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset,
 327                             zb->zb_blkid * epb + i);
 328                 }
 329
 330                 /* recursively visitbp() blocks below this */
 331                 for (i = 0; i < epb; i++) {
 332                         err = traverse_dnode(td, &dnp[i], zb->zb_objset,
 333                             zb->zb_blkid * epb + i);
 334                         if (err != 0)
 335                                 break;
 336                 }
 337         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 338                 uint32_t flags = ARC_WAIT;
 339                 objset_phys_t *osp;
 340                 dnode_phys_t *dnp;
 341
 342                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 343                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 344                 if (err != 0)
 345                         goto post;
 346
 347                 osp = buf->b_data;
 348                 dnp = &osp->os_meta_dnode;
 349                 prefetch_dnode_metadata(td, dnp, zb->zb_objset,
 350                     DMU_META_DNODE_OBJECT);
 351                 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 352                         prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
 353                             zb->zb_objset, DMU_GROUPUSED_OBJECT);
 354                         prefetch_dnode_metadata(td, &osp->os_userused_dnode,
 355                             zb->zb_objset, DMU_USERUSED_OBJECT);
 356                 }
 357
 358                 err = traverse_dnode(td, dnp, zb->zb_objset,
 359                     DMU_META_DNODE_OBJECT);
 360                 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 361                         dnp = &osp->os_groupused_dnode;
 362                         err = traverse_dnode(td, dnp, zb->zb_objset,
 363                             DMU_GROUPUSED_OBJECT);
 364                 }
 365                 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 366                         dnp = &osp->os_userused_dnode;
 367                         err = traverse_dnode(td, dnp, zb->zb_objset,
 368                             DMU_USERUSED_OBJECT);
 369                 }
 370         }
 371
 372         if (buf)
 373                 (void) arc_buf_remove_ref(buf, &buf);
 374
 375 post:
 376         if (err == 0 && (td->td_flags & TRAVERSE_POST))
 377                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 378
 379         if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
 380                 /*
 381                  * Ignore this disk error as requested by the HARD flag,
 382                  * and continue traversal.
 383                  */
 384                 err = 0;
 385         }
 386
 387         /*
 388          * If we are stopping here, set td_resume.
 389          */
 390         if (td->td_resume != NULL && err != 0 && !td->td_paused) {
 391                 td->td_resume->zb_objset = zb->zb_objset;
 392                 td->td_resume->zb_object = zb->zb_object;
 393                 td->td_resume->zb_level = 0;
 394                 /*
 395                  * If we have stopped on an indirect block (e.g. due to
 396                  * i/o error), we have not visited anything below it.
 397                  * Set the bookmark to the first level-0 block that we need
 398                  * to visit.  This way, the resuming code does not need to
 399                  * deal with resuming from indirect blocks.
 400                  */
 401                 td->td_resume->zb_blkid = zb->zb_blkid <<
 402                     (zb->zb_level * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 403                 td->td_paused = B_TRUE;
 404         }
 405
 406         return (err);
 407 }
 408
 409 static void
 410 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
 411     uint64_t objset, uint64_t object)
 412 {
 413         int j;
 414         zbookmark_phys_t czb;
 415
 416         for (j = 0; j < dnp->dn_nblkptr; j++) {
 417                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 418                 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
 419         }
 420
 421         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 422                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 423                 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
 424         }
 425 }
 426
 427 static int
 428 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
 429     uint64_t objset, uint64_t object)
 430 {
 431         int j, err = 0;
 432         zbookmark_phys_t czb;
 433
 434         for (j = 0; j < dnp->dn_nblkptr; j++) {
 435                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 436                 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
 437                 if (err != 0)
 438                         break;
 439         }
 440
 441         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 442                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 443                 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
 444         }
 445         return (err);
 446 }
 447
 448 /* ARGSUSED */
 449 static int
 450 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 451     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 452 {
 453         prefetch_data_t *pfd = arg;
 454         uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
 455
 456         ASSERT(pfd->pd_blks_fetched >= 0);
 457         if (pfd->pd_cancel)
 458                 return (SET_ERROR(EINTR));
 459
 460         if (!prefetch_needed(pfd, bp))
 461                 return (0);
 462
 463         mutex_enter(&pfd->pd_mtx);
 464         while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
 465                 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
 466         pfd->pd_blks_fetched++;
 467         cv_broadcast(&pfd->pd_cv);
 468         mutex_exit(&pfd->pd_mtx);
 469
 470         (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 471             ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
 472
 473         return (0);
 474 }
 475
 476 static void
 477 traverse_prefetch_thread(void *arg)
 478 {
 479         traverse_data_t *td_main = arg;
 480         traverse_data_t td = *td_main;
 481         zbookmark_phys_t czb;
 482
 483         td.td_func = traverse_prefetcher;
 484         td.td_arg = td_main->td_pfd;
 485         td.td_pfd = NULL;
 486
 487         SET_BOOKMARK(&czb, td.td_objset,
 488             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 489         (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
 490
 491         mutex_enter(&td_main->td_pfd->pd_mtx);
 492         td_main->td_pfd->pd_exited = B_TRUE;
 493         cv_broadcast(&td_main->td_pfd->pd_cv);
 494         mutex_exit(&td_main->td_pfd->pd_mtx);
 495 }
 496
 497 /*
 498  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 499  * in syncing context).
 500  */
 501 static int
 502 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
 503     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 504     blkptr_cb_t func, void *arg)
 505 {
 506         traverse_data_t *td;
 507         prefetch_data_t *pd;
 508         zbookmark_phys_t *czb;
 509         int err;
 510
 511         ASSERT(ds == NULL || objset == ds->ds_object);
 512         ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
 513
 514         /*
 515          * The data prefetching mechanism (the prefetch thread) is incompatible
 516          * with resuming from a bookmark.
 517          */
 518         ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
 519
 520         td = kmem_alloc(sizeof (traverse_data_t), KM_PUSHPAGE);
 521         pd = kmem_zalloc(sizeof (prefetch_data_t), KM_PUSHPAGE);
 522         czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_PUSHPAGE);
 523
 524         td->td_spa = spa;
 525         td->td_objset = objset;
 526         td->td_rootbp = rootbp;
 527         td->td_min_txg = txg_start;
 528         td->td_resume = resume;
 529         td->td_func = func;
 530         td->td_arg = arg;
 531         td->td_pfd = pd;
 532         td->td_flags = flags;
 533         td->td_paused = B_FALSE;
 534
 535         pd->pd_blks_max = zfs_pd_blks_max;
 536         pd->pd_flags = flags;
 537         mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
 538         cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
 539
 540         SET_BOOKMARK(czb, td->td_objset,
 541             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 542
 543         /* See comment on ZIL traversal in dsl_scan_visitds. */
 544         if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
 545                 uint32_t flags = ARC_WAIT;
 546                 objset_phys_t *osp;
 547                 arc_buf_t *buf;
 548
 549                 err = arc_read(NULL, td->td_spa, rootbp,
 550                     arc_getbuf_func, &buf,
 551                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, czb);
 552                 if (err != 0)
 553                         return (err);
 554
 555                 osp = buf->b_data;
 556                 traverse_zil(td, &osp->os_zil_header);
 557                 (void) arc_buf_remove_ref(buf, &buf);
 558         }
 559
 560         if (!(flags & TRAVERSE_PREFETCH_DATA) ||
 561             0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
 562             td, TQ_NOQUEUE))
 563                 pd->pd_exited = B_TRUE;
 564
 565         err = traverse_visitbp(td, NULL, rootbp, czb);
 566
 567         mutex_enter(&pd->pd_mtx);
 568         pd->pd_cancel = B_TRUE;
 569         cv_broadcast(&pd->pd_cv);
 570         while (!pd->pd_exited)
 571                 cv_wait(&pd->pd_cv, &pd->pd_mtx);
 572         mutex_exit(&pd->pd_mtx);
 573
 574         mutex_destroy(&pd->pd_mtx);
 575         cv_destroy(&pd->pd_cv);
 576
 577         kmem_free(czb, sizeof (zbookmark_phys_t));
 578         kmem_free(pd, sizeof (struct prefetch_data));
 579         kmem_free(td, sizeof (struct traverse_data));
 580
 581         return (err);
 582 }
 583
 584 /*
 585  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 586  * in syncing context).
 587  */
 588 int
 589 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
 590     blkptr_cb_t func, void *arg)
 591 {
 592         return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
 593             &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg));
 594 }
 595
 596 int
 597 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
 598     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 599     blkptr_cb_t func, void *arg)
 600 {
 601         return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
 602             blkptr, txg_start, resume, flags, func, arg));
 603 }
 604
 605 /*
 606  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
 607  */
 608 int
 609 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
 610     blkptr_cb_t func, void *arg)
 611 {
 612         int err;
 613         uint64_t obj;
 614         dsl_pool_t *dp = spa_get_dsl(spa);
 615         objset_t *mos = dp->dp_meta_objset;
 616         boolean_t hard = (flags & TRAVERSE_HARD);
 617
 618         /* visit the MOS */
 619         err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
 620             txg_start, NULL, flags, func, arg);
 621         if (err != 0)
 622                 return (err);
 623
 624         /* visit each dataset */
 625         for (obj = 1; err == 0;
 626             err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
 627                 dmu_object_info_t doi;
 628
 629                 err = dmu_object_info(mos, obj, &doi);
 630                 if (err != 0) {
 631                         if (hard)
 632                                 continue;
 633                         break;
 634                 }
 635
 636                 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
 637                         dsl_dataset_t *ds;
 638                         uint64_t txg = txg_start;
 639
 640                         dsl_pool_config_enter(dp, FTAG);
 641                         err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
 642                         dsl_pool_config_exit(dp, FTAG);
 643                         if (err != 0) {
 644                                 if (hard)
 645                                         continue;
 646                                 break;
 647                         }
 648                         if (ds->ds_phys->ds_prev_snap_txg > txg)
 649                                 txg = ds->ds_phys->ds_prev_snap_txg;
 650                         err = traverse_dataset(ds, txg, flags, func, arg);
 651                         dsl_dataset_rele(ds, FTAG);
 652                         if (err != 0)
 653                                 break;
 654                 }
 655         }
 656         if (err == ESRCH)
 657                 err = 0;
 658         return (err);
 659 }
 660
 661 #if defined(_KERNEL) && defined(HAVE_SPL)
 662 EXPORT_SYMBOL(traverse_dataset);
 663 EXPORT_SYMBOL(traverse_pool);
 664
 665 module_param(zfs_pd_blks_max, int, 0644);
 666 MODULE_PARM_DESC(zfs_pd_blks_max, "Max number of blocks to prefetch");
 667 #endif