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 https://opensource.org/licenses/CDDL-1.0.
  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 uint_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                 ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
 115                 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
 116                     ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
 117
 118                 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
 119                     td->td_arg);
 120         }
 121         return (0);
 122 }
 123
 124 static void
 125 traverse_zil(traverse_data_t *td, zil_header_t *zh)
 126 {
 127         uint64_t claim_txg = zh->zh_claim_txg;
 128
 129         /*
 130          * We only want to visit blocks that have been claimed but not yet
 131          * replayed; plus blocks that are already stable in read-only mode.
 132          */
 133         if (claim_txg == 0 && spa_writeable(td->td_spa))
 134                 return;
 135
 136         zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
 137         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
 138             claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
 139         zil_free(zilog);
 140 }
 141
 142 typedef enum resume_skip {
 143         RESUME_SKIP_ALL,
 144         RESUME_SKIP_NONE,
 145         RESUME_SKIP_CHILDREN
 146 } resume_skip_t;
 147
 148 /*
 149  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
 150  * the block indicated by zb does not need to be visited at all. Returns
 151  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
 152  * resume point. This indicates that this block should be visited but not its
 153  * children (since they must have been visited in a previous traversal).
 154  * Otherwise returns RESUME_SKIP_NONE.
 155  */
 156 static resume_skip_t
 157 resume_skip_check(const traverse_data_t *td, const dnode_phys_t *dnp,
 158     const zbookmark_phys_t *zb)
 159 {
 160         if (td->td_resume != NULL) {
 161                 /*
 162                  * If we already visited this bp & everything below,
 163                  * don't bother doing it again.
 164                  */
 165                 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
 166                         return (RESUME_SKIP_ALL);
 167
 168                 if (memcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
 169                         if (td->td_flags & TRAVERSE_POST)
 170                                 return (RESUME_SKIP_CHILDREN);
 171                 }
 172         }
 173         return (RESUME_SKIP_NONE);
 174 }
 175
 176 /*
 177  * Returns B_TRUE, if prefetch read is issued, otherwise B_FALSE.
 178  */
 179 static boolean_t
 180 traverse_prefetch_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
 181     const blkptr_t *bp, const zbookmark_phys_t *zb)
 182 {
 183         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
 184             ARC_FLAG_PRESCIENT_PREFETCH;
 185         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
 186
 187         if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
 188                 return (B_FALSE);
 189         /*
 190          * If this bp is before the resume point, it may have already been
 191          * freed.
 192          */
 193         if (resume_skip_check(td, dnp, zb) != RESUME_SKIP_NONE)
 194                 return (B_FALSE);
 195         if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
 196                 return (B_FALSE);
 197         if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
 198                 return (B_FALSE);
 199         ASSERT(!BP_IS_REDACTED(bp));
 200
 201         if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 202                 zio_flags |= ZIO_FLAG_RAW;
 203
 204         (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
 205             ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 206         return (B_TRUE);
 207 }
 208
 209 static boolean_t
 210 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
 211 {
 212         ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
 213         if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
 214             BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp))
 215                 return (B_FALSE);
 216         return (B_TRUE);
 217 }
 218
 219 static int
 220 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
 221     const blkptr_t *bp, const zbookmark_phys_t *zb)
 222 {
 223         int err = 0;
 224         arc_buf_t *buf = NULL;
 225         prefetch_data_t *pd = td->td_pfd;
 226
 227         switch (resume_skip_check(td, dnp, zb)) {
 228         case RESUME_SKIP_ALL:
 229                 return (0);
 230         case RESUME_SKIP_CHILDREN:
 231                 goto post;
 232         case RESUME_SKIP_NONE:
 233                 break;
 234         default:
 235                 ASSERT(0);
 236         }
 237
 238         if (bp->blk_birth == 0) {
 239                 /*
 240                  * Since this block has a birth time of 0 it must be one of
 241                  * two things: a hole created before the
 242                  * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
 243                  * which has always been a hole in an object.
 244                  *
 245                  * If a file is written sparsely, then the unwritten parts of
 246                  * the file were "always holes" -- that is, they have been
 247                  * holes since this object was allocated.  However, we (and
 248                  * our callers) can not necessarily tell when an object was
 249                  * allocated.  Therefore, if it's possible that this object
 250                  * was freed and then its object number reused, we need to
 251                  * visit all the holes with birth==0.
 252                  *
 253                  * If it isn't possible that the object number was reused,
 254                  * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
 255                  * all the blocks we will visit as part of this traversal,
 256                  * then this hole must have always existed, so we can skip
 257                  * it.  We visit blocks born after (exclusive) td_min_txg.
 258                  *
 259                  * Note that the meta-dnode cannot be reallocated.
 260                  */
 261                 if (!send_holes_without_birth_time &&
 262                     (!td->td_realloc_possible ||
 263                     zb->zb_object == DMU_META_DNODE_OBJECT) &&
 264                     td->td_hole_birth_enabled_txg <= td->td_min_txg)
 265                         return (0);
 266         } else if (bp->blk_birth <= td->td_min_txg) {
 267                 return (0);
 268         }
 269
 270         if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
 271                 uint64_t size = BP_GET_LSIZE(bp);
 272                 mutex_enter(&pd->pd_mtx);
 273                 ASSERT(pd->pd_bytes_fetched >= 0);
 274                 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
 275                         cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
 276                 pd->pd_bytes_fetched -= size;
 277                 cv_broadcast(&pd->pd_cv);
 278                 mutex_exit(&pd->pd_mtx);
 279         }
 280
 281         if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
 282                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 283                 if (err != 0)
 284                         goto post;
 285                 return (0);
 286         }
 287
 288         if (td->td_flags & TRAVERSE_PRE) {
 289                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
 290                     td->td_arg);
 291                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 292                         return (0);
 293                 if (err != 0)
 294                         goto post;
 295         }
 296
 297         if (BP_GET_LEVEL(bp) > 0) {
 298                 uint32_t flags = ARC_FLAG_WAIT;
 299                 int32_t i, ptidx, pidx;
 300                 uint32_t prefetchlimit;
 301                 int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 302                 zbookmark_phys_t *czb;
 303
 304                 ASSERT(!BP_IS_PROTECTED(bp));
 305
 306                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 307                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 308                 if (err != 0)
 309                         goto post;
 310
 311                 czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
 312
 313                 /*
 314                  * When performing a traversal it is beneficial to
 315                  * asynchronously read-ahead the upcoming indirect
 316                  * blocks since they will be needed shortly. However,
 317                  * since a 128k indirect (non-L0) block may contain up
 318                  * to 1024 128-byte block pointers, its preferable to not
 319                  * prefetch them all at once. Issuing a large number of
 320                  * async reads may effect performance, and the earlier
 321                  * the indirect blocks are prefetched the less likely
 322                  * they are to still be resident in the ARC when needed.
 323                  * Therefore, prefetching indirect blocks is limited to
 324                  * zfs_traverse_indirect_prefetch_limit=32 blocks by
 325                  * default.
 326                  *
 327                  * pidx: Index for which next prefetch to be issued.
 328                  * ptidx: Index at which next prefetch to be triggered.
 329                  */
 330                 ptidx = 0;
 331                 pidx = 1;
 332                 prefetchlimit = zfs_traverse_indirect_prefetch_limit;
 333                 for (i = 0; i < epb; i++) {
 334                         if (prefetchlimit && i == ptidx) {
 335                                 ASSERT3S(ptidx, <=, pidx);
 336                                 for (uint32_t  prefetched = 0; pidx < epb &&
 337                                     prefetched < prefetchlimit; pidx++) {
 338                                         SET_BOOKMARK(czb, zb->zb_objset,
 339                                             zb->zb_object, zb->zb_level - 1,
 340                                             zb->zb_blkid * epb + pidx);
 341                                         if (traverse_prefetch_metadata(td, dnp,
 342                                             &((blkptr_t *)buf->b_data)[pidx],
 343                                             czb) == B_TRUE) {
 344                                                 prefetched++;
 345                                                 if (prefetched ==
 346                                                     MAX(prefetchlimit / 2, 1))
 347                                                         ptidx = pidx;
 348                                         }
 349                                 }
 350                         }
 351
 352                         /* recursively visitbp() blocks below this */
 353                         SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
 354                             zb->zb_level - 1,
 355                             zb->zb_blkid * epb + i);
 356                         err = traverse_visitbp(td, dnp,
 357                             &((blkptr_t *)buf->b_data)[i], czb);
 358                         if (err != 0)
 359                                 break;
 360                 }
 361
 362                 kmem_free(czb, sizeof (zbookmark_phys_t));
 363
 364         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 365                 uint32_t flags = ARC_FLAG_WAIT;
 366                 uint32_t zio_flags = ZIO_FLAG_CANFAIL;
 367                 int32_t i;
 368                 int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 369                 dnode_phys_t *child_dnp;
 370
 371                 /*
 372                  * dnode blocks might have their bonus buffers encrypted, so
 373                  * we must be careful to honor TRAVERSE_NO_DECRYPT
 374                  */
 375                 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 376                         zio_flags |= ZIO_FLAG_RAW;
 377
 378                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 379                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 380                 if (err != 0)
 381                         goto post;
 382
 383                 child_dnp = buf->b_data;
 384
 385                 for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
 386                         prefetch_dnode_metadata(td, &child_dnp[i],
 387                             zb->zb_objset, zb->zb_blkid * epb + i);
 388                 }
 389
 390                 /* recursively visitbp() blocks below this */
 391                 for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
 392                         err = traverse_dnode(td, bp, &child_dnp[i],
 393                             zb->zb_objset, zb->zb_blkid * epb + i);
 394                         if (err != 0)
 395                                 break;
 396                 }
 397         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 398                 uint32_t zio_flags = ZIO_FLAG_CANFAIL;
 399                 arc_flags_t flags = ARC_FLAG_WAIT;
 400                 objset_phys_t *osp;
 401
 402                 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 403                         zio_flags |= ZIO_FLAG_RAW;
 404
 405                 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
 406                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 407                 if (err != 0)
 408                         goto post;
 409
 410                 osp = buf->b_data;
 411                 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
 412                     DMU_META_DNODE_OBJECT);
 413                 /*
 414                  * See the block comment above for the goal of this variable.
 415                  * If the maxblkid of the meta-dnode is 0, then we know that
 416                  * we've never had more than DNODES_PER_BLOCK objects in the
 417                  * dataset, which means we can't have reused any object ids.
 418                  */
 419                 if (osp->os_meta_dnode.dn_maxblkid == 0)
 420                         td->td_realloc_possible = B_FALSE;
 421
 422                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 423                         if (OBJSET_BUF_HAS_PROJECTUSED(buf))
 424                                 prefetch_dnode_metadata(td,
 425                                     &osp->os_projectused_dnode,
 426                                     zb->zb_objset, DMU_PROJECTUSED_OBJECT);
 427                         prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
 428                             zb->zb_objset, DMU_GROUPUSED_OBJECT);
 429                         prefetch_dnode_metadata(td, &osp->os_userused_dnode,
 430                             zb->zb_objset, DMU_USERUSED_OBJECT);
 431                 }
 432
 433                 err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset,
 434                     DMU_META_DNODE_OBJECT);
 435                 if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
 436                         if (OBJSET_BUF_HAS_PROJECTUSED(buf))
 437                                 err = traverse_dnode(td, bp,
 438                                     &osp->os_projectused_dnode, zb->zb_objset,
 439                                     DMU_PROJECTUSED_OBJECT);
 440                         if (err == 0)
 441                                 err = traverse_dnode(td, bp,
 442                                     &osp->os_groupused_dnode, zb->zb_objset,
 443                                     DMU_GROUPUSED_OBJECT);
 444                         if (err == 0)
 445                                 err = traverse_dnode(td, bp,
 446                                     &osp->os_userused_dnode, zb->zb_objset,
 447                                     DMU_USERUSED_OBJECT);
 448                 }
 449         }
 450
 451         if (buf)
 452                 arc_buf_destroy(buf, &buf);
 453
 454 post:
 455         if (err == 0 && (td->td_flags & TRAVERSE_POST))
 456                 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 457
 458         if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
 459                 /*
 460                  * Ignore this disk error as requested by the HARD flag,
 461                  * and continue traversal.
 462                  */
 463                 err = 0;
 464         }
 465
 466         /*
 467          * If we are stopping here, set td_resume.
 468          */
 469         if (td->td_resume != NULL && err != 0 && !td->td_paused) {
 470                 td->td_resume->zb_objset = zb->zb_objset;
 471                 td->td_resume->zb_object = zb->zb_object;
 472                 td->td_resume->zb_level = 0;
 473                 /*
 474                  * If we have stopped on an indirect block (e.g. due to
 475                  * i/o error), we have not visited anything below it.
 476                  * Set the bookmark to the first level-0 block that we need
 477                  * to visit.  This way, the resuming code does not need to
 478                  * deal with resuming from indirect blocks.
 479                  *
 480                  * Note, if zb_level <= 0, dnp may be NULL, so we don't want
 481                  * to dereference it.
 482                  */
 483                 td->td_resume->zb_blkid = zb->zb_blkid;
 484                 if (zb->zb_level > 0) {
 485                         td->td_resume->zb_blkid <<= zb->zb_level *
 486                             (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
 487                 }
 488                 td->td_paused = B_TRUE;
 489         }
 490
 491         return (err);
 492 }
 493
 494 static void
 495 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
 496     uint64_t objset, uint64_t object)
 497 {
 498         int j;
 499         zbookmark_phys_t czb;
 500
 501         for (j = 0; j < dnp->dn_nblkptr; j++) {
 502                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 503                 traverse_prefetch_metadata(td, dnp, &dnp->dn_blkptr[j], &czb);
 504         }
 505
 506         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 507                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 508                 traverse_prefetch_metadata(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
 509         }
 510 }
 511
 512 static int
 513 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp,
 514     uint64_t objset, uint64_t object)
 515 {
 516         int j, err = 0;
 517         zbookmark_phys_t czb;
 518
 519         if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
 520             object < td->td_resume->zb_object)
 521                 return (0);
 522
 523         if (td->td_flags & TRAVERSE_PRE) {
 524                 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
 525                     ZB_DNODE_BLKID);
 526                 err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
 527                     td->td_arg);
 528                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 529                         return (0);
 530                 if (err != 0)
 531                         return (err);
 532         }
 533
 534         for (j = 0; j < dnp->dn_nblkptr; j++) {
 535                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 536                 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
 537                 if (err != 0)
 538                         break;
 539         }
 540
 541         if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
 542                 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 543                 err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
 544         }
 545
 546         if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
 547                 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
 548                     ZB_DNODE_BLKID);
 549                 err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
 550                     td->td_arg);
 551                 if (err == TRAVERSE_VISIT_NO_CHILDREN)
 552                         return (0);
 553                 if (err != 0)
 554                         return (err);
 555         }
 556         return (err);
 557 }
 558
 559 static int
 560 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 561     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 562 {
 563         (void) zilog, (void) dnp;
 564         prefetch_data_t *pfd = arg;
 565         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
 566         arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
 567             ARC_FLAG_PRESCIENT_PREFETCH;
 568
 569         ASSERT(pfd->pd_bytes_fetched >= 0);
 570         if (zb->zb_level == ZB_DNODE_LEVEL)
 571                 return (0);
 572         if (pfd->pd_cancel)
 573                 return (SET_ERROR(EINTR));
 574
 575         if (!prefetch_needed(pfd, bp))
 576                 return (0);
 577
 578         mutex_enter(&pfd->pd_mtx);
 579         while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
 580                 cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx);
 581         pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
 582         cv_broadcast(&pfd->pd_cv);
 583         mutex_exit(&pfd->pd_mtx);
 584
 585         if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
 586                 zio_flags |= ZIO_FLAG_RAW;
 587
 588         (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 589             zio_flags, &aflags, zb);
 590
 591         return (0);
 592 }
 593
 594 static void
 595 traverse_prefetch_thread(void *arg)
 596 {
 597         traverse_data_t *td_main = arg;
 598         traverse_data_t td = *td_main;
 599         zbookmark_phys_t czb;
 600         fstrans_cookie_t cookie = spl_fstrans_mark();
 601
 602         td.td_func = traverse_prefetcher;
 603         td.td_arg = td_main->td_pfd;
 604         td.td_pfd = NULL;
 605         td.td_resume = &td_main->td_pfd->pd_resume;
 606
 607         SET_BOOKMARK(&czb, td.td_objset,
 608             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 609         (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
 610
 611         mutex_enter(&td_main->td_pfd->pd_mtx);
 612         td_main->td_pfd->pd_exited = B_TRUE;
 613         cv_broadcast(&td_main->td_pfd->pd_cv);
 614         mutex_exit(&td_main->td_pfd->pd_mtx);
 615         spl_fstrans_unmark(cookie);
 616 }
 617
 618 /*
 619  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 620  * in syncing context).
 621  */
 622 static int
 623 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
 624     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 625     blkptr_cb_t func, void *arg)
 626 {
 627         traverse_data_t *td;
 628         prefetch_data_t *pd;
 629         zbookmark_phys_t *czb;
 630         int err;
 631
 632         ASSERT(ds == NULL || objset == ds->ds_object);
 633         ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
 634
 635         td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP);
 636         pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP);
 637         czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
 638
 639         td->td_spa = spa;
 640         td->td_objset = objset;
 641         td->td_rootbp = rootbp;
 642         td->td_min_txg = txg_start;
 643         td->td_resume = resume;
 644         td->td_func = func;
 645         td->td_arg = arg;
 646         td->td_pfd = pd;
 647         td->td_flags = flags;
 648         td->td_paused = B_FALSE;
 649         td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
 650
 651         if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
 652                 VERIFY(spa_feature_enabled_txg(spa,
 653                     SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg));
 654         } else {
 655                 td->td_hole_birth_enabled_txg = UINT64_MAX;
 656         }
 657
 658         pd->pd_flags = flags;
 659         if (resume != NULL)
 660                 pd->pd_resume = *resume;
 661         mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
 662         cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
 663
 664         SET_BOOKMARK(czb, td->td_objset,
 665             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 666
 667         /* See comment on ZIL traversal in dsl_scan_visitds. */
 668         if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
 669                 zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
 670                 uint32_t flags = ARC_FLAG_WAIT;
 671                 objset_phys_t *osp;
 672                 arc_buf_t *buf;
 673                 ASSERT(!BP_IS_REDACTED(rootbp));
 674
 675                 if ((td->td_flags & TRAVERSE_NO_DECRYPT) &&
 676                     BP_IS_PROTECTED(rootbp))
 677                         zio_flags |= ZIO_FLAG_RAW;
 678
 679                 err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func,
 680                     &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb);
 681                 if (err != 0) {
 682                         /*
 683                          * If both TRAVERSE_HARD and TRAVERSE_PRE are set,
 684                          * continue to visitbp so that td_func can be called
 685                          * in pre stage, and err will reset to zero.
 686                          */
 687                         if (!(td->td_flags & TRAVERSE_HARD) ||
 688                             !(td->td_flags & TRAVERSE_PRE))
 689                                 goto out;
 690                 } else {
 691                         osp = buf->b_data;
 692                         traverse_zil(td, &osp->os_zil_header);
 693                         arc_buf_destroy(buf, &buf);
 694                 }
 695         }
 696
 697         if (!(flags & TRAVERSE_PREFETCH_DATA) ||
 698             taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread,
 699             td, TQ_NOQUEUE) == TASKQID_INVALID)
 700                 pd->pd_exited = B_TRUE;
 701
 702         err = traverse_visitbp(td, NULL, rootbp, czb);
 703
 704         mutex_enter(&pd->pd_mtx);
 705         pd->pd_cancel = B_TRUE;
 706         cv_broadcast(&pd->pd_cv);
 707         while (!pd->pd_exited)
 708                 cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
 709         mutex_exit(&pd->pd_mtx);
 710 out:
 711         mutex_destroy(&pd->pd_mtx);
 712         cv_destroy(&pd->pd_cv);
 713
 714         kmem_free(czb, sizeof (zbookmark_phys_t));
 715         kmem_free(pd, sizeof (struct prefetch_data));
 716         kmem_free(td, sizeof (struct traverse_data));
 717
 718         return (err);
 719 }
 720
 721 /*
 722  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
 723  * in syncing context).
 724  */
 725 int
 726 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
 727     zbookmark_phys_t *resume,
 728     int flags, blkptr_cb_t func, void *arg)
 729 {
 730         return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
 731             &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
 732 }
 733
 734 int
 735 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
 736     int flags, blkptr_cb_t func, void *arg)
 737 {
 738         return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
 739 }
 740
 741 int
 742 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
 743     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
 744     blkptr_cb_t func, void *arg)
 745 {
 746         return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
 747             blkptr, txg_start, resume, flags, func, arg));
 748 }
 749
 750 /*
 751  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
 752  */
 753 int
 754 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
 755     blkptr_cb_t func, void *arg)
 756 {
 757         int err;
 758         dsl_pool_t *dp = spa_get_dsl(spa);
 759         objset_t *mos = dp->dp_meta_objset;
 760         boolean_t hard = (flags & TRAVERSE_HARD);
 761
 762         /* visit the MOS */
 763         err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
 764             txg_start, NULL, flags, func, arg);
 765         if (err != 0)
 766                 return (err);
 767
 768         /* visit each dataset */
 769         for (uint64_t obj = 1; err == 0;
 770             err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
 771                 dmu_object_info_t doi;
 772
 773                 err = dmu_object_info(mos, obj, &doi);
 774                 if (err != 0) {
 775                         if (hard)
 776                                 continue;
 777                         break;
 778                 }
 779
 780                 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
 781                         dsl_dataset_t *ds;
 782                         uint64_t txg = txg_start;
 783
 784                         dsl_pool_config_enter(dp, FTAG);
 785                         err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
 786                         dsl_pool_config_exit(dp, FTAG);
 787                         if (err != 0) {
 788                                 if (hard)
 789                                         continue;
 790                                 break;
 791                         }
 792                         if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
 793                                 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
 794                         err = traverse_dataset(ds, txg, flags, func, arg);
 795                         dsl_dataset_rele(ds, FTAG);
 796                         if (err != 0)
 797                                 break;
 798                 }
 799         }
 800         if (err == ESRCH)
 801                 err = 0;
 802         return (err);
 803 }
 804
 805 EXPORT_SYMBOL(traverse_dataset);
 806 EXPORT_SYMBOL(traverse_pool);
 807
 808 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW,
 809         "Max number of bytes to prefetch");
 810
 811 ZFS_MODULE_PARAM(zfs, zfs_, traverse_indirect_prefetch_limit, UINT, ZMOD_RW,
 812         "Traverse prefetch number of blocks pointed by indirect block");
 813
 814 #if defined(_KERNEL)
 815 module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644);
 816 MODULE_PARM_DESC(ignore_hole_birth,
 817         "Alias for send_holes_without_birth_time");
 818 #endif
 819
 820 /* CSTYLED */
 821 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW,
 822         "Ignore hole_birth txg for zfs send");