module/zfs/dsl_scrub.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 2009 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #include <sys/dsl_pool.h>
  27 #include <sys/dsl_dataset.h>
  28 #include <sys/dsl_prop.h>
  29 #include <sys/dsl_dir.h>
  30 #include <sys/dsl_synctask.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dmu_objset.h>
  34 #include <sys/arc.h>
  35 #include <sys/zap.h>
  36 #include <sys/zio.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/fs/zfs.h>
  39 #include <sys/zfs_znode.h>
  40 #include <sys/spa_impl.h>
  41 #include <sys/vdev_impl.h>
  42 #include <sys/zil_impl.h>
  43
  44 typedef int (scrub_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  45
  46 static scrub_cb_t dsl_pool_scrub_clean_cb;
  47 static dsl_syncfunc_t dsl_pool_scrub_cancel_sync;
  48
  49 int zfs_scrub_min_time = 1; /* scrub for at least 1 sec each txg */
  50 int zfs_resilver_min_time = 3; /* resilver for at least 3 sec each txg */
  51 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  52
  53 extern int zfs_txg_timeout;
  54
  55 static scrub_cb_t *scrub_funcs[SCRUB_FUNC_NUMFUNCS] = {
  56         NULL,
  57         dsl_pool_scrub_clean_cb
  58 };
  59
  60 #define SET_BOOKMARK(zb, objset, object, level, blkid)  \
  61 {                                                       \
  62         (zb)->zb_objset = objset;                       \
  63         (zb)->zb_object = object;                       \
  64         (zb)->zb_level = level;                         \
  65         (zb)->zb_blkid = blkid;                         \
  66 }
  67
  68 /* ARGSUSED */
  69 static void
  70 dsl_pool_scrub_setup_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
  71 {
  72         dsl_pool_t *dp = arg1;
  73         enum scrub_func *funcp = arg2;
  74         dmu_object_type_t ot = 0;
  75         boolean_t complete = B_FALSE;
  76
  77         dsl_pool_scrub_cancel_sync(dp, &complete, cr, tx);
  78
  79         ASSERT(dp->dp_scrub_func == SCRUB_FUNC_NONE);
  80         ASSERT(*funcp > SCRUB_FUNC_NONE);
  81         ASSERT(*funcp < SCRUB_FUNC_NUMFUNCS);
  82
  83         dp->dp_scrub_min_txg = 0;
  84         dp->dp_scrub_max_txg = tx->tx_txg;
  85
  86         if (*funcp == SCRUB_FUNC_CLEAN) {
  87                 vdev_t *rvd = dp->dp_spa->spa_root_vdev;
  88
  89                 /* rewrite all disk labels */
  90                 vdev_config_dirty(rvd);
  91
  92                 if (vdev_resilver_needed(rvd,
  93                     &dp->dp_scrub_min_txg, &dp->dp_scrub_max_txg)) {
  94                         spa_event_notify(dp->dp_spa, NULL,
  95                             ESC_ZFS_RESILVER_START);
  96                         dp->dp_scrub_max_txg = MIN(dp->dp_scrub_max_txg,
  97                             tx->tx_txg);
  98                 } else {
  99                         spa_event_notify(dp->dp_spa, NULL,
 100                             ESC_ZFS_SCRUB_START);
 101                 }
 102
 103                 /* zero out the scrub stats in all vdev_stat_t's */
 104                 vdev_scrub_stat_update(rvd,
 105                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 106                     POOL_SCRUB_EVERYTHING, B_FALSE);
 107
 108                 dp->dp_spa->spa_scrub_started = B_TRUE;
 109         }
 110
 111         /* back to the generic stuff */
 112
 113         if (dp->dp_blkstats == NULL) {
 114                 dp->dp_blkstats =
 115                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 116         }
 117         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 118
 119         if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB)
 120                 ot = DMU_OT_ZAP_OTHER;
 121
 122         dp->dp_scrub_func = *funcp;
 123         dp->dp_scrub_queue_obj = zap_create(dp->dp_meta_objset,
 124             ot ? ot : DMU_OT_SCRUB_QUEUE, DMU_OT_NONE, 0, tx);
 125         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 126         dp->dp_scrub_restart = B_FALSE;
 127         dp->dp_spa->spa_scrub_errors = 0;
 128
 129         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 130             DMU_POOL_SCRUB_FUNC, sizeof (uint32_t), 1,
 131             &dp->dp_scrub_func, tx));
 132         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 133             DMU_POOL_SCRUB_QUEUE, sizeof (uint64_t), 1,
 134             &dp->dp_scrub_queue_obj, tx));
 135         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 136             DMU_POOL_SCRUB_MIN_TXG, sizeof (uint64_t), 1,
 137             &dp->dp_scrub_min_txg, tx));
 138         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 139             DMU_POOL_SCRUB_MAX_TXG, sizeof (uint64_t), 1,
 140             &dp->dp_scrub_max_txg, tx));
 141         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 142             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 143             &dp->dp_scrub_bookmark, tx));
 144         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 145             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 146             &dp->dp_spa->spa_scrub_errors, tx));
 147
 148         spa_history_internal_log(LOG_POOL_SCRUB, dp->dp_spa, tx, cr,
 149             "func=%u mintxg=%llu maxtxg=%llu",
 150             *funcp, dp->dp_scrub_min_txg, dp->dp_scrub_max_txg);
 151 }
 152
 153 int
 154 dsl_pool_scrub_setup(dsl_pool_t *dp, enum scrub_func func)
 155 {
 156         return (dsl_sync_task_do(dp, NULL,
 157             dsl_pool_scrub_setup_sync, dp, &func, 0));
 158 }
 159
 160 /* ARGSUSED */
 161 static void
 162 dsl_pool_scrub_cancel_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
 163 {
 164         dsl_pool_t *dp = arg1;
 165         boolean_t *completep = arg2;
 166
 167         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 168                 return;
 169
 170         mutex_enter(&dp->dp_scrub_cancel_lock);
 171
 172         if (dp->dp_scrub_restart) {
 173                 dp->dp_scrub_restart = B_FALSE;
 174                 *completep = B_FALSE;
 175         }
 176
 177         /* XXX this is scrub-clean specific */
 178         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 179         while (dp->dp_spa->spa_scrub_inflight > 0) {
 180                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 181                     &dp->dp_spa->spa_scrub_lock);
 182         }
 183         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 184         dp->dp_spa->spa_scrub_started = B_FALSE;
 185         dp->dp_spa->spa_scrub_active = B_FALSE;
 186
 187         dp->dp_scrub_func = SCRUB_FUNC_NONE;
 188         VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 189             dp->dp_scrub_queue_obj, tx));
 190         dp->dp_scrub_queue_obj = 0;
 191         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 192
 193         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 194             DMU_POOL_SCRUB_QUEUE, tx));
 195         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 196             DMU_POOL_SCRUB_MIN_TXG, tx));
 197         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 198             DMU_POOL_SCRUB_MAX_TXG, tx));
 199         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 200             DMU_POOL_SCRUB_BOOKMARK, tx));
 201         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 202             DMU_POOL_SCRUB_FUNC, tx));
 203         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 204             DMU_POOL_SCRUB_ERRORS, tx));
 205
 206         spa_history_internal_log(LOG_POOL_SCRUB_DONE, dp->dp_spa, tx, cr,
 207             "complete=%u", *completep);
 208
 209         /* below is scrub-clean specific */
 210         vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev, POOL_SCRUB_NONE,
 211             *completep);
 212         /*
 213          * If the scrub/resilver completed, update all DTLs to reflect this.
 214          * Whether it succeeded or not, vacate all temporary scrub DTLs.
 215          */
 216         vdev_dtl_reassess(dp->dp_spa->spa_root_vdev, tx->tx_txg,
 217             *completep ? dp->dp_scrub_max_txg : 0, B_TRUE);
 218         if (*completep)
 219                 spa_event_notify(dp->dp_spa, NULL, dp->dp_scrub_min_txg ?
 220                     ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 221         spa_errlog_rotate(dp->dp_spa);
 222
 223         /*
 224          * We may have finished replacing a device.
 225          * Let the async thread assess this and handle the detach.
 226          */
 227         spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER_DONE);
 228
 229         dp->dp_scrub_min_txg = dp->dp_scrub_max_txg = 0;
 230         mutex_exit(&dp->dp_scrub_cancel_lock);
 231 }
 232
 233 int
 234 dsl_pool_scrub_cancel(dsl_pool_t *dp)
 235 {
 236         boolean_t complete = B_FALSE;
 237
 238         return (dsl_sync_task_do(dp, NULL,
 239             dsl_pool_scrub_cancel_sync, dp, &complete, 3));
 240 }
 241
 242 int
 243 dsl_free(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp,
 244     zio_done_func_t *done, void *private, uint32_t arc_flags)
 245 {
 246         /*
 247          * This function will be used by bp-rewrite wad to intercept frees.
 248          */
 249         return (arc_free(pio, dp->dp_spa, txg, (blkptr_t *)bpp,
 250             done, private, arc_flags));
 251 }
 252
 253 static boolean_t
 254 bookmark_is_zero(const zbookmark_t *zb)
 255 {
 256         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 257             zb->zb_level == 0 && zb->zb_blkid == 0);
 258 }
 259
 260 /* dnp is the dnode for zb1->zb_object */
 261 static boolean_t
 262 bookmark_is_before(dnode_phys_t *dnp, const zbookmark_t *zb1,
 263     const zbookmark_t *zb2)
 264 {
 265         uint64_t zb1nextL0, zb2thisobj;
 266
 267         ASSERT(zb1->zb_objset == zb2->zb_objset);
 268         ASSERT(zb1->zb_object != -1ULL);
 269         ASSERT(zb2->zb_level == 0);
 270
 271         /*
 272          * A bookmark in the deadlist is considered to be after
 273          * everything else.
 274          */
 275         if (zb2->zb_object == -1ULL)
 276                 return (B_TRUE);
 277
 278         /* The objset_phys_t isn't before anything. */
 279         if (dnp == NULL)
 280                 return (B_FALSE);
 281
 282         zb1nextL0 = (zb1->zb_blkid + 1) <<
 283             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 284
 285         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 286             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 287
 288         if (zb1->zb_object == 0) {
 289                 uint64_t nextobj = zb1nextL0 *
 290                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 291                 return (nextobj <= zb2thisobj);
 292         }
 293
 294         if (zb1->zb_object < zb2thisobj)
 295                 return (B_TRUE);
 296         if (zb1->zb_object > zb2thisobj)
 297                 return (B_FALSE);
 298         if (zb2->zb_object == 0)
 299                 return (B_FALSE);
 300         return (zb1nextL0 <= zb2->zb_blkid);
 301 }
 302
 303 static boolean_t
 304 scrub_pause(dsl_pool_t *dp, const zbookmark_t *zb)
 305 {
 306         int elapsed_ticks;
 307         int mintime;
 308
 309         if (dp->dp_scrub_pausing)
 310                 return (B_TRUE); /* we're already pausing */
 311
 312         if (!bookmark_is_zero(&dp->dp_scrub_bookmark))
 313                 return (B_FALSE); /* we're resuming */
 314
 315         /* We only know how to resume from level-0 blocks. */
 316         if (zb->zb_level != 0)
 317                 return (B_FALSE);
 318
 319         mintime = dp->dp_scrub_isresilver ? zfs_resilver_min_time :
 320             zfs_scrub_min_time;
 321         elapsed_ticks = lbolt64 - dp->dp_scrub_start_time;
 322         if (elapsed_ticks > hz * zfs_txg_timeout ||
 323             (elapsed_ticks > hz * mintime && txg_sync_waiting(dp))) {
 324                 dprintf("pausing at %llx/%llx/%llx/%llx\n",
 325                     (longlong_t)zb->zb_objset, (longlong_t)zb->zb_object,
 326                     (longlong_t)zb->zb_level, (longlong_t)zb->zb_blkid);
 327                 dp->dp_scrub_pausing = B_TRUE;
 328                 dp->dp_scrub_bookmark = *zb;
 329                 return (B_TRUE);
 330         }
 331         return (B_FALSE);
 332 }
 333
 334 typedef struct zil_traverse_arg {
 335         dsl_pool_t      *zta_dp;
 336         zil_header_t    *zta_zh;
 337 } zil_traverse_arg_t;
 338
 339 /* ARGSUSED */
 340 static void
 341 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 342 {
 343         zil_traverse_arg_t *zta = arg;
 344         dsl_pool_t *dp = zta->zta_dp;
 345         zil_header_t *zh = zta->zta_zh;
 346         zbookmark_t zb;
 347
 348         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 349                 return;
 350
 351         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 352                 return;
 353
 354         zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 355         zb.zb_object = 0;
 356         zb.zb_level = -1;
 357         zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
 358         VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 359 }
 360
 361 /* ARGSUSED */
 362 static void
 363 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 364 {
 365         if (lrc->lrc_txtype == TX_WRITE) {
 366                 zil_traverse_arg_t *zta = arg;
 367                 dsl_pool_t *dp = zta->zta_dp;
 368                 zil_header_t *zh = zta->zta_zh;
 369                 lr_write_t *lr = (lr_write_t *)lrc;
 370                 blkptr_t *bp = &lr->lr_blkptr;
 371                 zbookmark_t zb;
 372
 373                 if (bp->blk_birth <= dp->dp_scrub_min_txg)
 374                         return;
 375
 376                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 377                         return;
 378
 379                 zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 380                 zb.zb_object = lr->lr_foid;
 381                 zb.zb_level = BP_GET_LEVEL(bp);
 382                 zb.zb_blkid = lr->lr_offset / BP_GET_LSIZE(bp);
 383                 VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 384         }
 385 }
 386
 387 static void
 388 traverse_zil(dsl_pool_t *dp, zil_header_t *zh)
 389 {
 390         uint64_t claim_txg = zh->zh_claim_txg;
 391         zil_traverse_arg_t zta = { dp, zh };
 392         zilog_t *zilog;
 393
 394         /*
 395          * We only want to visit blocks that have been claimed but not yet
 396          * replayed (or, in read-only mode, blocks that *would* be claimed).
 397          */
 398         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 399                 return;
 400
 401         zilog = zil_alloc(dp->dp_meta_objset, zh);
 402
 403         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, &zta,
 404             claim_txg);
 405
 406         zil_free(zilog);
 407 }
 408
 409 static void
 410 scrub_visitbp(dsl_pool_t *dp, dnode_phys_t *dnp,
 411     arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb)
 412 {
 413         int err;
 414         arc_buf_t *buf = NULL;
 415
 416         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 417                 return;
 418
 419         if (scrub_pause(dp, zb))
 420                 return;
 421
 422         if (!bookmark_is_zero(&dp->dp_scrub_bookmark)) {
 423                 /*
 424                  * If we already visited this bp & everything below (in
 425                  * a prior txg), don't bother doing it again.
 426                  */
 427                 if (bookmark_is_before(dnp, zb, &dp->dp_scrub_bookmark))
 428                         return;
 429
 430                 /*
 431                  * If we found the block we're trying to resume from, or
 432                  * we went past it to a different object, zero it out to
 433                  * indicate that it's OK to start checking for pausing
 434                  * again.
 435                  */
 436                 if (bcmp(zb, &dp->dp_scrub_bookmark, sizeof (*zb)) == 0 ||
 437                     zb->zb_object > dp->dp_scrub_bookmark.zb_object) {
 438                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 439                             (longlong_t)zb->zb_objset,
 440                             (longlong_t)zb->zb_object,
 441                             (longlong_t)zb->zb_level,
 442                             (longlong_t)zb->zb_blkid);
 443                         bzero(&dp->dp_scrub_bookmark, sizeof (*zb));
 444                 }
 445         }
 446
 447         if (BP_GET_LEVEL(bp) > 0) {
 448                 uint32_t flags = ARC_WAIT;
 449                 int i;
 450                 blkptr_t *cbp;
 451                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 452
 453                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 454                     arc_getbuf_func, &buf,
 455                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 456                 if (err) {
 457                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 458                         dp->dp_spa->spa_scrub_errors++;
 459                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 460                         return;
 461                 }
 462                 cbp = buf->b_data;
 463
 464                 for (i = 0; i < epb; i++, cbp++) {
 465                         zbookmark_t czb;
 466
 467                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 468                             zb->zb_level - 1,
 469                             zb->zb_blkid * epb + i);
 470                         scrub_visitbp(dp, dnp, buf, cbp, &czb);
 471                 }
 472         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 473                 uint32_t flags = ARC_WAIT;
 474                 dnode_phys_t *child_dnp;
 475                 int i, j;
 476                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 477
 478                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 479                     arc_getbuf_func, &buf,
 480                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 481                 if (err) {
 482                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 483                         dp->dp_spa->spa_scrub_errors++;
 484                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 485                         return;
 486                 }
 487                 child_dnp = buf->b_data;
 488
 489                 for (i = 0; i < epb; i++, child_dnp++) {
 490                         for (j = 0; j < child_dnp->dn_nblkptr; j++) {
 491                                 zbookmark_t czb;
 492
 493                                 SET_BOOKMARK(&czb, zb->zb_objset,
 494                                     zb->zb_blkid * epb + i,
 495                                     child_dnp->dn_nlevels - 1, j);
 496                                 scrub_visitbp(dp, child_dnp, buf,
 497                                     &child_dnp->dn_blkptr[j], &czb);
 498                         }
 499                 }
 500         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 501                 uint32_t flags = ARC_WAIT;
 502                 objset_phys_t *osp;
 503                 int j;
 504
 505                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 506                     arc_getbuf_func, &buf,
 507                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 508                 if (err) {
 509                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 510                         dp->dp_spa->spa_scrub_errors++;
 511                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 512                         return;
 513                 }
 514
 515                 osp = buf->b_data;
 516
 517                 traverse_zil(dp, &osp->os_zil_header);
 518
 519                 for (j = 0; j < osp->os_meta_dnode.dn_nblkptr; j++) {
 520                         zbookmark_t czb;
 521
 522                         SET_BOOKMARK(&czb, zb->zb_objset, 0,
 523                             osp->os_meta_dnode.dn_nlevels - 1, j);
 524                         scrub_visitbp(dp, &osp->os_meta_dnode, buf,
 525                             &osp->os_meta_dnode.dn_blkptr[j], &czb);
 526                 }
 527         }
 528
 529         (void) scrub_funcs[dp->dp_scrub_func](dp, bp, zb);
 530         if (buf)
 531                 (void) arc_buf_remove_ref(buf, &buf);
 532 }
 533
 534 static void
 535 scrub_visit_rootbp(dsl_pool_t *dp, dsl_dataset_t *ds, blkptr_t *bp)
 536 {
 537         zbookmark_t zb;
 538
 539         SET_BOOKMARK(&zb, ds ? ds->ds_object : 0, 0, -1, 0);
 540         scrub_visitbp(dp, NULL, NULL, bp, &zb);
 541 }
 542
 543 void
 544 dsl_pool_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 545 {
 546         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 547
 548         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 549                 return;
 550
 551         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 552                 SET_BOOKMARK(&dp->dp_scrub_bookmark, -1, 0, 0, 0);
 553         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 554             ds->ds_object, tx) != 0) {
 555                 return;
 556         }
 557
 558         if (ds->ds_phys->ds_next_snap_obj != 0) {
 559                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 560                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 561         }
 562         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 563 }
 564
 565 void
 566 dsl_pool_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 567 {
 568         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 569
 570         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 571                 return;
 572
 573         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 574
 575         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 576                 dp->dp_scrub_bookmark.zb_objset =
 577                     ds->ds_phys->ds_prev_snap_obj;
 578         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 579             ds->ds_object, tx) == 0) {
 580                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 581                     ds->ds_phys->ds_prev_snap_obj, tx) == 0);
 582         }
 583 }
 584
 585 void
 586 dsl_pool_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 587 {
 588         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 589
 590         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 591                 return;
 592
 593         if (dp->dp_scrub_bookmark.zb_objset == ds1->ds_object) {
 594                 dp->dp_scrub_bookmark.zb_objset = ds2->ds_object;
 595         } else if (dp->dp_scrub_bookmark.zb_objset == ds2->ds_object) {
 596                 dp->dp_scrub_bookmark.zb_objset = ds1->ds_object;
 597         }
 598
 599         if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 600             ds1->ds_object, tx) == 0) {
 601                 int err = zap_add_int(dp->dp_meta_objset,
 602                     dp->dp_scrub_queue_obj, ds2->ds_object, tx);
 603                 VERIFY(err == 0 || err == EEXIST);
 604                 if (err == EEXIST) {
 605                         /* Both were there to begin with */
 606                         VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 607                             dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 608                 }
 609         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 610             ds2->ds_object, tx) == 0) {
 611                 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 612                     dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 613         }
 614 }
 615
 616 struct enqueue_clones_arg {
 617         dmu_tx_t *tx;
 618         uint64_t originobj;
 619 };
 620
 621 /* ARGSUSED */
 622 static int
 623 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 624 {
 625         struct enqueue_clones_arg *eca = arg;
 626         dsl_dataset_t *ds;
 627         int err;
 628         dsl_pool_t *dp;
 629
 630         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 631         if (err)
 632                 return (err);
 633         dp = ds->ds_dir->dd_pool;
 634
 635         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 636                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 637                         dsl_dataset_t *prev;
 638                         err = dsl_dataset_hold_obj(dp,
 639                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 640
 641                         dsl_dataset_rele(ds, FTAG);
 642                         if (err)
 643                                 return (err);
 644                         ds = prev;
 645                 }
 646                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 647                     ds->ds_object, eca->tx) == 0);
 648         }
 649         dsl_dataset_rele(ds, FTAG);
 650         return (0);
 651 }
 652
 653 static void
 654 scrub_visitds(dsl_pool_t *dp, uint64_t dsobj, dmu_tx_t *tx)
 655 {
 656         dsl_dataset_t *ds;
 657         uint64_t min_txg_save;
 658
 659         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 660
 661         /*
 662          * Iterate over the bps in this ds.
 663          */
 664         min_txg_save = dp->dp_scrub_min_txg;
 665         dp->dp_scrub_min_txg =
 666             MAX(dp->dp_scrub_min_txg, ds->ds_phys->ds_prev_snap_txg);
 667         scrub_visit_rootbp(dp, ds, &ds->ds_phys->ds_bp);
 668         dp->dp_scrub_min_txg = min_txg_save;
 669
 670         if (dp->dp_scrub_pausing)
 671                 goto out;
 672
 673         /*
 674          * Add descendent datasets to work queue.
 675          */
 676         if (ds->ds_phys->ds_next_snap_obj != 0) {
 677                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 678                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 679         }
 680         if (ds->ds_phys->ds_num_children > 1) {
 681                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
 682                         struct enqueue_clones_arg eca;
 683                         eca.tx = tx;
 684                         eca.originobj = ds->ds_object;
 685
 686                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
 687                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
 688                 } else {
 689                         VERIFY(zap_join(dp->dp_meta_objset,
 690                             ds->ds_phys->ds_next_clones_obj,
 691                             dp->dp_scrub_queue_obj, tx) == 0);
 692                 }
 693         }
 694
 695 out:
 696         dsl_dataset_rele(ds, FTAG);
 697 }
 698
 699 /* ARGSUSED */
 700 static int
 701 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 702 {
 703         dmu_tx_t *tx = arg;
 704         dsl_dataset_t *ds;
 705         int err;
 706         dsl_pool_t *dp;
 707
 708         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 709         if (err)
 710                 return (err);
 711
 712         dp = ds->ds_dir->dd_pool;
 713
 714         while (ds->ds_phys->ds_prev_snap_obj != 0) {
 715                 dsl_dataset_t *prev;
 716                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
 717                     FTAG, &prev);
 718                 if (err) {
 719                         dsl_dataset_rele(ds, FTAG);
 720                         return (err);
 721                 }
 722
 723                 /*
 724                  * If this is a clone, we don't need to worry about it for now.
 725                  */
 726                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
 727                         dsl_dataset_rele(ds, FTAG);
 728                         dsl_dataset_rele(prev, FTAG);
 729                         return (0);
 730                 }
 731                 dsl_dataset_rele(ds, FTAG);
 732                 ds = prev;
 733         }
 734
 735         VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 736             ds->ds_object, tx) == 0);
 737         dsl_dataset_rele(ds, FTAG);
 738         return (0);
 739 }
 740
 741 void
 742 dsl_pool_scrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
 743 {
 744         spa_t *spa = dp->dp_spa;
 745         zap_cursor_t zc;
 746         zap_attribute_t za;
 747         boolean_t complete = B_TRUE;
 748
 749         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 750                 return;
 751
 752         /*
 753          * If the pool is not loaded, or is trying to unload, leave it alone.
 754          */
 755         if (spa->spa_load_state != SPA_LOAD_NONE || spa_shutting_down(spa))
 756                 return;
 757
 758         if (dp->dp_scrub_restart) {
 759                 enum scrub_func func = dp->dp_scrub_func;
 760                 dp->dp_scrub_restart = B_FALSE;
 761                 dsl_pool_scrub_setup_sync(dp, &func, kcred, tx);
 762         }
 763
 764         if (spa->spa_root_vdev->vdev_stat.vs_scrub_type == 0) {
 765                 /*
 766                  * We must have resumed after rebooting; reset the vdev
 767                  * stats to know that we're doing a scrub (although it
 768                  * will think we're just starting now).
 769                  */
 770                 vdev_scrub_stat_update(spa->spa_root_vdev,
 771                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 772                     POOL_SCRUB_EVERYTHING, B_FALSE);
 773         }
 774
 775         dp->dp_scrub_pausing = B_FALSE;
 776         dp->dp_scrub_start_time = lbolt64;
 777         dp->dp_scrub_isresilver = (dp->dp_scrub_min_txg != 0);
 778         spa->spa_scrub_active = B_TRUE;
 779
 780         if (dp->dp_scrub_bookmark.zb_objset == 0) {
 781                 /* First do the MOS & ORIGIN */
 782                 scrub_visit_rootbp(dp, NULL, &dp->dp_meta_rootbp);
 783                 if (dp->dp_scrub_pausing)
 784                         goto out;
 785
 786                 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) {
 787                         VERIFY(0 == dmu_objset_find_spa(spa,
 788                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
 789                 } else {
 790                         scrub_visitds(dp, dp->dp_origin_snap->ds_object, tx);
 791                 }
 792                 ASSERT(!dp->dp_scrub_pausing);
 793         } else if (dp->dp_scrub_bookmark.zb_objset != -1ULL) {
 794                 /*
 795                  * If we were paused, continue from here.  Note if the
 796                  * ds we were paused on was deleted, the zb_objset will
 797                  * be -1, so we will skip this and find a new objset
 798                  * below.
 799                  */
 800                 scrub_visitds(dp, dp->dp_scrub_bookmark.zb_objset, tx);
 801                 if (dp->dp_scrub_pausing)
 802                         goto out;
 803         }
 804
 805         /*
 806          * In case we were paused right at the end of the ds, zero the
 807          * bookmark so we don't think that we're still trying to resume.
 808          */
 809         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 810
 811         /* keep pulling things out of the zap-object-as-queue */
 812         while (zap_cursor_init(&zc, dp->dp_meta_objset, dp->dp_scrub_queue_obj),
 813             zap_cursor_retrieve(&zc, &za) == 0) {
 814                 VERIFY(0 == zap_remove(dp->dp_meta_objset,
 815                     dp->dp_scrub_queue_obj, za.za_name, tx));
 816                 scrub_visitds(dp, za.za_first_integer, tx);
 817                 if (dp->dp_scrub_pausing)
 818                         break;
 819                 zap_cursor_fini(&zc);
 820         }
 821         zap_cursor_fini(&zc);
 822         if (dp->dp_scrub_pausing)
 823                 goto out;
 824
 825         /* done. */
 826
 827         dsl_pool_scrub_cancel_sync(dp, &complete, kcred, tx);
 828         return;
 829 out:
 830         VERIFY(0 == zap_update(dp->dp_meta_objset,
 831             DMU_POOL_DIRECTORY_OBJECT,
 832             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 833             &dp->dp_scrub_bookmark, tx));
 834         VERIFY(0 == zap_update(dp->dp_meta_objset,
 835             DMU_POOL_DIRECTORY_OBJECT,
 836             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 837             &spa->spa_scrub_errors, tx));
 838
 839         /* XXX this is scrub-clean specific */
 840         mutex_enter(&spa->spa_scrub_lock);
 841         while (spa->spa_scrub_inflight > 0)
 842                 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 843         mutex_exit(&spa->spa_scrub_lock);
 844 }
 845
 846 void
 847 dsl_pool_scrub_restart(dsl_pool_t *dp)
 848 {
 849         mutex_enter(&dp->dp_scrub_cancel_lock);
 850         dp->dp_scrub_restart = B_TRUE;
 851         mutex_exit(&dp->dp_scrub_cancel_lock);
 852 }
 853
 854 /*
 855  * scrub consumers
 856  */
 857
 858 static void
 859 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
 860 {
 861         int i;
 862
 863         /*
 864          * If we resume after a reboot, zab will be NULL; don't record
 865          * incomplete stats in that case.
 866          */
 867         if (zab == NULL)
 868                 return;
 869
 870         for (i = 0; i < 4; i++) {
 871                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
 872                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
 873                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
 874                 int equal;
 875
 876                 zb->zb_count++;
 877                 zb->zb_asize += BP_GET_ASIZE(bp);
 878                 zb->zb_lsize += BP_GET_LSIZE(bp);
 879                 zb->zb_psize += BP_GET_PSIZE(bp);
 880                 zb->zb_gangs += BP_COUNT_GANG(bp);
 881
 882                 switch (BP_GET_NDVAS(bp)) {
 883                 case 2:
 884                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 885                             DVA_GET_VDEV(&bp->blk_dva[1]))
 886                                 zb->zb_ditto_2_of_2_samevdev++;
 887                         break;
 888                 case 3:
 889                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 890                             DVA_GET_VDEV(&bp->blk_dva[1])) +
 891                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 892                             DVA_GET_VDEV(&bp->blk_dva[2])) +
 893                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
 894                             DVA_GET_VDEV(&bp->blk_dva[2]));
 895                         if (equal == 1)
 896                                 zb->zb_ditto_2_of_3_samevdev++;
 897                         else if (equal == 3)
 898                                 zb->zb_ditto_3_of_3_samevdev++;
 899                         break;
 900                 }
 901         }
 902 }
 903
 904 static void
 905 dsl_pool_scrub_clean_done(zio_t *zio)
 906 {
 907         spa_t *spa = zio->io_spa;
 908
 909         zio_data_buf_free(zio->io_data, zio->io_size);
 910
 911         mutex_enter(&spa->spa_scrub_lock);
 912         spa->spa_scrub_inflight--;
 913         cv_broadcast(&spa->spa_scrub_io_cv);
 914
 915         if (zio->io_error && (zio->io_error != ECKSUM ||
 916             !(zio->io_flags & ZIO_FLAG_SPECULATIVE)))
 917                 spa->spa_scrub_errors++;
 918         mutex_exit(&spa->spa_scrub_lock);
 919 }
 920
 921 static int
 922 dsl_pool_scrub_clean_cb(dsl_pool_t *dp,
 923     const blkptr_t *bp, const zbookmark_t *zb)
 924 {
 925         size_t size = BP_GET_PSIZE(bp);
 926         spa_t *spa = dp->dp_spa;
 927         boolean_t needs_io;
 928         int zio_flags = ZIO_FLAG_SCRUB_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
 929         int zio_priority;
 930
 931         ASSERT(bp->blk_birth > dp->dp_scrub_min_txg);
 932
 933         if (bp->blk_birth >= dp->dp_scrub_max_txg)
 934                 return (0);
 935
 936         count_block(dp->dp_blkstats, bp);
 937
 938         if (dp->dp_scrub_isresilver == 0) {
 939                 /* It's a scrub */
 940                 zio_flags |= ZIO_FLAG_SCRUB;
 941                 zio_priority = ZIO_PRIORITY_SCRUB;
 942                 needs_io = B_TRUE;
 943         } else {
 944                 /* It's a resilver */
 945                 zio_flags |= ZIO_FLAG_RESILVER;
 946                 zio_priority = ZIO_PRIORITY_RESILVER;
 947                 needs_io = B_FALSE;
 948         }
 949
 950         /* If it's an intent log block, failure is expected. */
 951         if (zb->zb_level == -1 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)
 952                 zio_flags |= ZIO_FLAG_SPECULATIVE;
 953
 954         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
 955                 vdev_t *vd = vdev_lookup_top(spa,
 956                     DVA_GET_VDEV(&bp->blk_dva[d]));
 957
 958                 /*
 959                  * Keep track of how much data we've examined so that
 960                  * zpool(1M) status can make useful progress reports.
 961                  */
 962                 mutex_enter(&vd->vdev_stat_lock);
 963                 vd->vdev_stat.vs_scrub_examined +=
 964                     DVA_GET_ASIZE(&bp->blk_dva[d]);
 965                 mutex_exit(&vd->vdev_stat_lock);
 966
 967                 /* if it's a resilver, this may not be in the target range */
 968                 if (!needs_io) {
 969                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
 970                                 /*
 971                                  * Gang members may be spread across multiple
 972                                  * vdevs, so the best estimate we have is the
 973                                  * scrub range, which has already been checked.
 974                                  * XXX -- it would be better to change our
 975                                  * allocation policy to ensure that all
 976                                  * gang members reside on the same vdev.
 977                                  */
 978                                 needs_io = B_TRUE;
 979                         } else {
 980                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
 981                                     bp->blk_birth, 1);
 982                         }
 983                 }
 984         }
 985
 986         if (needs_io && !zfs_no_scrub_io) {
 987                 void *data = zio_data_buf_alloc(size);
 988
 989                 mutex_enter(&spa->spa_scrub_lock);
 990                 while (spa->spa_scrub_inflight >= spa->spa_scrub_maxinflight)
 991                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 992                 spa->spa_scrub_inflight++;
 993                 mutex_exit(&spa->spa_scrub_lock);
 994
 995                 zio_nowait(zio_read(NULL, spa, bp, data, size,
 996                     dsl_pool_scrub_clean_done, NULL, zio_priority,
 997                     zio_flags, zb));
 998         }
 999
1000         /* do not relocate this block */
1001         return (0);
1002 }
1003
1004 int
1005 dsl_pool_scrub_clean(dsl_pool_t *dp)
1006 {
1007         /*
1008          * Purge all vdev caches.  We do this here rather than in sync
1009          * context because this requires a writer lock on the spa_config
1010          * lock, which we can't do from sync context.  The
1011          * spa_scrub_reopen flag indicates that vdev_open() should not
1012          * attempt to start another scrub.
1013          */
1014         spa_config_enter(dp->dp_spa, SCL_ALL, FTAG, RW_WRITER);
1015         dp->dp_spa->spa_scrub_reopen = B_TRUE;
1016         vdev_reopen(dp->dp_spa->spa_root_vdev);
1017         dp->dp_spa->spa_scrub_reopen = B_FALSE;
1018         spa_config_exit(dp->dp_spa, SCL_ALL, FTAG);
1019
1020         return (dsl_pool_scrub_setup(dp, SCRUB_FUNC_CLEAN));
1021 }