sys/cddl/contrib/opensolaris/uts/common/fs/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 2008 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                 }
  99
 100                 /* zero out the scrub stats in all vdev_stat_t's */
 101                 vdev_scrub_stat_update(rvd,
 102                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 103                     POOL_SCRUB_EVERYTHING, B_FALSE);
 104
 105                 dp->dp_spa->spa_scrub_started = B_TRUE;
 106         }
 107
 108         /* back to the generic stuff */
 109
 110         if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB)
 111                 ot = DMU_OT_ZAP_OTHER;
 112
 113         dp->dp_scrub_func = *funcp;
 114         dp->dp_scrub_queue_obj = zap_create(dp->dp_meta_objset,
 115             ot ? ot : DMU_OT_SCRUB_QUEUE, DMU_OT_NONE, 0, tx);
 116         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 117         dp->dp_scrub_restart = B_FALSE;
 118         dp->dp_spa->spa_scrub_errors = 0;
 119
 120         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 121             DMU_POOL_SCRUB_FUNC, sizeof (uint32_t), 1,
 122             &dp->dp_scrub_func, tx));
 123         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 124             DMU_POOL_SCRUB_QUEUE, sizeof (uint64_t), 1,
 125             &dp->dp_scrub_queue_obj, tx));
 126         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 127             DMU_POOL_SCRUB_MIN_TXG, sizeof (uint64_t), 1,
 128             &dp->dp_scrub_min_txg, tx));
 129         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 130             DMU_POOL_SCRUB_MAX_TXG, sizeof (uint64_t), 1,
 131             &dp->dp_scrub_max_txg, tx));
 132         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 133             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 134             &dp->dp_scrub_bookmark, tx));
 135         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 136             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 137             &dp->dp_spa->spa_scrub_errors, tx));
 138
 139         spa_history_internal_log(LOG_POOL_SCRUB, dp->dp_spa, tx, cr,
 140             "func=%u mintxg=%llu maxtxg=%llu",
 141             *funcp, dp->dp_scrub_min_txg, dp->dp_scrub_max_txg);
 142 }
 143
 144 int
 145 dsl_pool_scrub_setup(dsl_pool_t *dp, enum scrub_func func)
 146 {
 147         return (dsl_sync_task_do(dp, NULL,
 148             dsl_pool_scrub_setup_sync, dp, &func, 0));
 149 }
 150
 151 /* ARGSUSED */
 152 static void
 153 dsl_pool_scrub_cancel_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
 154 {
 155         dsl_pool_t *dp = arg1;
 156         boolean_t *completep = arg2;
 157
 158         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 159                 return;
 160
 161         mutex_enter(&dp->dp_scrub_cancel_lock);
 162
 163         if (dp->dp_scrub_restart) {
 164                 dp->dp_scrub_restart = B_FALSE;
 165                 *completep = B_FALSE;
 166         }
 167
 168         /* XXX this is scrub-clean specific */
 169         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 170         while (dp->dp_spa->spa_scrub_inflight > 0) {
 171                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 172                     &dp->dp_spa->spa_scrub_lock);
 173         }
 174         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 175         dp->dp_spa->spa_scrub_started = B_FALSE;
 176         dp->dp_spa->spa_scrub_active = B_FALSE;
 177
 178         dp->dp_scrub_func = SCRUB_FUNC_NONE;
 179         VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 180             dp->dp_scrub_queue_obj, tx));
 181         dp->dp_scrub_queue_obj = 0;
 182         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 183
 184         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 185             DMU_POOL_SCRUB_QUEUE, tx));
 186         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 187             DMU_POOL_SCRUB_MIN_TXG, tx));
 188         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 189             DMU_POOL_SCRUB_MAX_TXG, tx));
 190         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 191             DMU_POOL_SCRUB_BOOKMARK, tx));
 192         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 193             DMU_POOL_SCRUB_FUNC, tx));
 194         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 195             DMU_POOL_SCRUB_ERRORS, tx));
 196
 197         spa_history_internal_log(LOG_POOL_SCRUB_DONE, dp->dp_spa, tx, cr,
 198             "complete=%u", *completep);
 199
 200         /* below is scrub-clean specific */
 201         vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev, POOL_SCRUB_NONE,
 202             *completep);
 203         /*
 204          * If the scrub/resilver completed, update all DTLs to reflect this.
 205          * Whether it succeeded or not, vacate all temporary scrub DTLs.
 206          */
 207         vdev_dtl_reassess(dp->dp_spa->spa_root_vdev, tx->tx_txg,
 208             *completep ? dp->dp_scrub_max_txg : 0, B_TRUE);
 209         if (dp->dp_scrub_min_txg && *completep)
 210                 spa_event_notify(dp->dp_spa, NULL, ESC_ZFS_RESILVER_FINISH);
 211         spa_errlog_rotate(dp->dp_spa);
 212
 213         /*
 214          * We may have finished replacing a device.
 215          * Let the async thread assess this and handle the detach.
 216          */
 217         spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER_DONE);
 218
 219         dp->dp_scrub_min_txg = dp->dp_scrub_max_txg = 0;
 220         mutex_exit(&dp->dp_scrub_cancel_lock);
 221 }
 222
 223 int
 224 dsl_pool_scrub_cancel(dsl_pool_t *dp)
 225 {
 226         boolean_t complete = B_FALSE;
 227
 228         return (dsl_sync_task_do(dp, NULL,
 229             dsl_pool_scrub_cancel_sync, dp, &complete, 3));
 230 }
 231
 232 int
 233 dsl_free(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp,
 234     zio_done_func_t *done, void *private, uint32_t arc_flags)
 235 {
 236         /*
 237          * This function will be used by bp-rewrite wad to intercept frees.
 238          */
 239         return (arc_free(pio, dp->dp_spa, txg, (blkptr_t *)bpp,
 240             done, private, arc_flags));
 241 }
 242
 243 static boolean_t
 244 bookmark_is_zero(const zbookmark_t *zb)
 245 {
 246         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 247             zb->zb_level == 0 && zb->zb_blkid == 0);
 248 }
 249
 250 /* dnp is the dnode for zb1->zb_object */
 251 static boolean_t
 252 bookmark_is_before(dnode_phys_t *dnp, const zbookmark_t *zb1,
 253     const zbookmark_t *zb2)
 254 {
 255         uint64_t zb1nextL0, zb2thisobj;
 256
 257         ASSERT(zb1->zb_objset == zb2->zb_objset);
 258         ASSERT(zb1->zb_object != -1ULL);
 259         ASSERT(zb2->zb_level == 0);
 260
 261         /*
 262          * A bookmark in the deadlist is considered to be after
 263          * everything else.
 264          */
 265         if (zb2->zb_object == -1ULL)
 266                 return (B_TRUE);
 267
 268         /* The objset_phys_t isn't before anything. */
 269         if (dnp == NULL)
 270                 return (B_FALSE);
 271
 272         zb1nextL0 = (zb1->zb_blkid + 1) <<
 273             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 274
 275         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 276             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 277
 278         if (zb1->zb_object == 0) {
 279                 uint64_t nextobj = zb1nextL0 *
 280                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 281                 return (nextobj <= zb2thisobj);
 282         }
 283
 284         if (zb1->zb_object < zb2thisobj)
 285                 return (B_TRUE);
 286         if (zb1->zb_object > zb2thisobj)
 287                 return (B_FALSE);
 288         if (zb2->zb_object == 0)
 289                 return (B_FALSE);
 290         return (zb1nextL0 <= zb2->zb_blkid);
 291 }
 292
 293 static boolean_t
 294 scrub_pause(dsl_pool_t *dp, const zbookmark_t *zb)
 295 {
 296         int elapsed_ticks;
 297         int mintime;
 298
 299         if (dp->dp_scrub_pausing)
 300                 return (B_TRUE); /* we're already pausing */
 301
 302         if (!bookmark_is_zero(&dp->dp_scrub_bookmark))
 303                 return (B_FALSE); /* we're resuming */
 304
 305         /* We only know how to resume from level-0 blocks. */
 306         if (zb->zb_level != 0)
 307                 return (B_FALSE);
 308
 309         mintime = dp->dp_scrub_isresilver ? zfs_resilver_min_time :
 310             zfs_scrub_min_time;
 311         elapsed_ticks = lbolt64 - dp->dp_scrub_start_time;
 312         if (elapsed_ticks > hz * zfs_txg_timeout ||
 313             (elapsed_ticks > hz * mintime && txg_sync_waiting(dp))) {
 314                 dprintf("pausing at %llx/%llx/%llx/%llx\n",
 315                     (longlong_t)zb->zb_objset, (longlong_t)zb->zb_object,
 316                     (longlong_t)zb->zb_level, (longlong_t)zb->zb_blkid);
 317                 dp->dp_scrub_pausing = B_TRUE;
 318                 dp->dp_scrub_bookmark = *zb;
 319                 return (B_TRUE);
 320         }
 321         return (B_FALSE);
 322 }
 323
 324 typedef struct zil_traverse_arg {
 325         dsl_pool_t      *zta_dp;
 326         zil_header_t    *zta_zh;
 327 } zil_traverse_arg_t;
 328
 329 /* ARGSUSED */
 330 static void
 331 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 332 {
 333         zil_traverse_arg_t *zta = arg;
 334         dsl_pool_t *dp = zta->zta_dp;
 335         zil_header_t *zh = zta->zta_zh;
 336         zbookmark_t zb;
 337
 338         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 339                 return;
 340
 341         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 342                 return;
 343
 344         zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 345         zb.zb_object = 0;
 346         zb.zb_level = -1;
 347         zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
 348         VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 349 }
 350
 351 /* ARGSUSED */
 352 static void
 353 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 354 {
 355         if (lrc->lrc_txtype == TX_WRITE) {
 356                 zil_traverse_arg_t *zta = arg;
 357                 dsl_pool_t *dp = zta->zta_dp;
 358                 zil_header_t *zh = zta->zta_zh;
 359                 lr_write_t *lr = (lr_write_t *)lrc;
 360                 blkptr_t *bp = &lr->lr_blkptr;
 361                 zbookmark_t zb;
 362
 363                 if (bp->blk_birth <= dp->dp_scrub_min_txg)
 364                         return;
 365
 366                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 367                         return;
 368
 369                 zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 370                 zb.zb_object = lr->lr_foid;
 371                 zb.zb_level = BP_GET_LEVEL(bp);
 372                 zb.zb_blkid = lr->lr_offset / BP_GET_LSIZE(bp);
 373                 VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 374         }
 375 }
 376
 377 static void
 378 traverse_zil(dsl_pool_t *dp, zil_header_t *zh)
 379 {
 380         uint64_t claim_txg = zh->zh_claim_txg;
 381         zil_traverse_arg_t zta = { dp, zh };
 382         zilog_t *zilog;
 383
 384         /*
 385          * We only want to visit blocks that have been claimed but not yet
 386          * replayed (or, in read-only mode, blocks that *would* be claimed).
 387          */
 388         if (claim_txg == 0 && (spa_mode & FWRITE))
 389                 return;
 390
 391         zilog = zil_alloc(dp->dp_meta_objset, zh);
 392
 393         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, &zta,
 394             claim_txg);
 395
 396         zil_free(zilog);
 397 }
 398
 399 static void
 400 scrub_visitbp(dsl_pool_t *dp, dnode_phys_t *dnp,
 401     arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb)
 402 {
 403         int err;
 404         arc_buf_t *buf = NULL;
 405
 406         if (bp->blk_birth == 0)
 407                 return;
 408
 409         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 410                 return;
 411
 412         if (scrub_pause(dp, zb))
 413                 return;
 414
 415         if (!bookmark_is_zero(&dp->dp_scrub_bookmark)) {
 416                 /*
 417                  * If we already visited this bp & everything below (in
 418                  * a prior txg), don't bother doing it again.
 419                  */
 420                 if (bookmark_is_before(dnp, zb, &dp->dp_scrub_bookmark))
 421                         return;
 422
 423                 /*
 424                  * If we found the block we're trying to resume from, or
 425                  * we went past it to a different object, zero it out to
 426                  * indicate that it's OK to start checking for pausing
 427                  * again.
 428                  */
 429                 if (bcmp(zb, &dp->dp_scrub_bookmark, sizeof (*zb)) == 0 ||
 430                     zb->zb_object > dp->dp_scrub_bookmark.zb_object) {
 431                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 432                             (longlong_t)zb->zb_objset,
 433                             (longlong_t)zb->zb_object,
 434                             (longlong_t)zb->zb_level,
 435                             (longlong_t)zb->zb_blkid);
 436                         bzero(&dp->dp_scrub_bookmark, sizeof (*zb));
 437                 }
 438         }
 439
 440         if (BP_GET_LEVEL(bp) > 0) {
 441                 uint32_t flags = ARC_WAIT;
 442                 int i;
 443                 blkptr_t *cbp;
 444                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 445
 446                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 447                     arc_getbuf_func, &buf,
 448                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 449                 if (err) {
 450                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 451                         dp->dp_spa->spa_scrub_errors++;
 452                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 453                         return;
 454                 }
 455                 cbp = buf->b_data;
 456
 457                 for (i = 0; i < epb; i++, cbp++) {
 458                         zbookmark_t czb;
 459
 460                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 461                             zb->zb_level - 1,
 462                             zb->zb_blkid * epb + i);
 463                         scrub_visitbp(dp, dnp, buf, cbp, &czb);
 464                 }
 465         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 466                 uint32_t flags = ARC_WAIT;
 467                 dnode_phys_t *child_dnp;
 468                 int i, j;
 469                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 470
 471                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 472                     arc_getbuf_func, &buf,
 473                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 474                 if (err) {
 475                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 476                         dp->dp_spa->spa_scrub_errors++;
 477                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 478                         return;
 479                 }
 480                 child_dnp = buf->b_data;
 481
 482                 for (i = 0; i < epb; i++, child_dnp++) {
 483                         for (j = 0; j < child_dnp->dn_nblkptr; j++) {
 484                                 zbookmark_t czb;
 485
 486                                 SET_BOOKMARK(&czb, zb->zb_objset,
 487                                     zb->zb_blkid * epb + i,
 488                                     child_dnp->dn_nlevels - 1, j);
 489                                 scrub_visitbp(dp, child_dnp, buf,
 490                                     &child_dnp->dn_blkptr[j], &czb);
 491                         }
 492                 }
 493         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 494                 uint32_t flags = ARC_WAIT;
 495                 objset_phys_t *osp;
 496                 int j;
 497
 498                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 499                     arc_getbuf_func, &buf,
 500                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 501                 if (err) {
 502                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 503                         dp->dp_spa->spa_scrub_errors++;
 504                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 505                         return;
 506                 }
 507
 508                 osp = buf->b_data;
 509
 510                 traverse_zil(dp, &osp->os_zil_header);
 511
 512                 for (j = 0; j < osp->os_meta_dnode.dn_nblkptr; j++) {
 513                         zbookmark_t czb;
 514
 515                         SET_BOOKMARK(&czb, zb->zb_objset, 0,
 516                             osp->os_meta_dnode.dn_nlevels - 1, j);
 517                         scrub_visitbp(dp, &osp->os_meta_dnode, buf,
 518                             &osp->os_meta_dnode.dn_blkptr[j], &czb);
 519                 }
 520         }
 521
 522         (void) scrub_funcs[dp->dp_scrub_func](dp, bp, zb);
 523         if (buf)
 524                 (void) arc_buf_remove_ref(buf, &buf);
 525 }
 526
 527 static void
 528 scrub_visit_rootbp(dsl_pool_t *dp, dsl_dataset_t *ds, blkptr_t *bp)
 529 {
 530         zbookmark_t zb;
 531
 532         SET_BOOKMARK(&zb, ds ? ds->ds_object : 0, 0, -1, 0);
 533         scrub_visitbp(dp, NULL, NULL, bp, &zb);
 534 }
 535
 536 void
 537 dsl_pool_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 538 {
 539         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 540
 541         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 542                 return;
 543
 544         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 545                 SET_BOOKMARK(&dp->dp_scrub_bookmark, -1, 0, 0, 0);
 546         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 547             ds->ds_object, tx) != 0) {
 548                 return;
 549         }
 550
 551         if (ds->ds_phys->ds_next_snap_obj != 0) {
 552                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 553                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 554         }
 555         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 556 }
 557
 558 void
 559 dsl_pool_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 560 {
 561         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 562
 563         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 564                 return;
 565
 566         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 567
 568         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 569                 dp->dp_scrub_bookmark.zb_objset =
 570                     ds->ds_phys->ds_prev_snap_obj;
 571         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 572             ds->ds_object, tx) == 0) {
 573                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 574                     ds->ds_phys->ds_prev_snap_obj, tx) == 0);
 575         }
 576 }
 577
 578 struct enqueue_clones_arg {
 579         dmu_tx_t *tx;
 580         uint64_t originobj;
 581 };
 582
 583 /* ARGSUSED */
 584 static int
 585 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 586 {
 587         struct enqueue_clones_arg *eca = arg;
 588         dsl_dataset_t *ds;
 589         int err;
 590         dsl_pool_t *dp;
 591
 592         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 593         if (err)
 594                 return (err);
 595         dp = ds->ds_dir->dd_pool;
 596
 597         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 598                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 599                         dsl_dataset_t *prev;
 600                         err = dsl_dataset_hold_obj(dp,
 601                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 602
 603                         dsl_dataset_rele(ds, FTAG);
 604                         if (err)
 605                                 return (err);
 606                         ds = prev;
 607                 }
 608                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 609                     ds->ds_object, eca->tx) == 0);
 610         }
 611         dsl_dataset_rele(ds, FTAG);
 612         return (0);
 613 }
 614
 615 static void
 616 scrub_visitds(dsl_pool_t *dp, uint64_t dsobj, dmu_tx_t *tx)
 617 {
 618         dsl_dataset_t *ds;
 619         uint64_t min_txg_save;
 620
 621         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 622
 623         /*
 624          * Iterate over the bps in this ds.
 625          */
 626         min_txg_save = dp->dp_scrub_min_txg;
 627         dp->dp_scrub_min_txg =
 628             MAX(dp->dp_scrub_min_txg, ds->ds_phys->ds_prev_snap_txg);
 629         scrub_visit_rootbp(dp, ds, &ds->ds_phys->ds_bp);
 630         dp->dp_scrub_min_txg = min_txg_save;
 631
 632         if (dp->dp_scrub_pausing)
 633                 goto out;
 634
 635         /*
 636          * Add descendent datasets to work queue.
 637          */
 638         if (ds->ds_phys->ds_next_snap_obj != 0) {
 639                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 640                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 641         }
 642         if (ds->ds_phys->ds_num_children > 1) {
 643                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
 644                         struct enqueue_clones_arg eca;
 645                         eca.tx = tx;
 646                         eca.originobj = ds->ds_object;
 647
 648                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
 649                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
 650                 } else {
 651                         VERIFY(zap_join(dp->dp_meta_objset,
 652                             ds->ds_phys->ds_next_clones_obj,
 653                             dp->dp_scrub_queue_obj, tx) == 0);
 654                 }
 655         }
 656
 657 out:
 658         dsl_dataset_rele(ds, FTAG);
 659 }
 660
 661 /* ARGSUSED */
 662 static int
 663 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 664 {
 665         dmu_tx_t *tx = arg;
 666         dsl_dataset_t *ds;
 667         int err;
 668         dsl_pool_t *dp;
 669
 670         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 671         if (err)
 672                 return (err);
 673
 674         dp = ds->ds_dir->dd_pool;
 675
 676         while (ds->ds_phys->ds_prev_snap_obj != 0) {
 677                 dsl_dataset_t *prev;
 678                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
 679                     FTAG, &prev);
 680                 if (err) {
 681                         dsl_dataset_rele(ds, FTAG);
 682                         return (err);
 683                 }
 684
 685                 /*
 686                  * If this is a clone, we don't need to worry about it for now.
 687                  */
 688                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
 689                         dsl_dataset_rele(ds, FTAG);
 690                         dsl_dataset_rele(prev, FTAG);
 691                         return (0);
 692                 }
 693                 dsl_dataset_rele(ds, FTAG);
 694                 ds = prev;
 695         }
 696
 697         VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 698             ds->ds_object, tx) == 0);
 699         dsl_dataset_rele(ds, FTAG);
 700         return (0);
 701 }
 702
 703 void
 704 dsl_pool_scrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
 705 {
 706         zap_cursor_t zc;
 707         zap_attribute_t za;
 708         boolean_t complete = B_TRUE;
 709
 710         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 711                 return;
 712
 713         /* If the spa is not fully loaded, don't bother. */
 714         if (dp->dp_spa->spa_load_state != SPA_LOAD_NONE)
 715                 return;
 716
 717         if (dp->dp_scrub_restart) {
 718                 enum scrub_func func = dp->dp_scrub_func;
 719                 dp->dp_scrub_restart = B_FALSE;
 720                 dsl_pool_scrub_setup_sync(dp, &func, kcred, tx);
 721         }
 722
 723         if (dp->dp_spa->spa_root_vdev->vdev_stat.vs_scrub_type == 0) {
 724                 /*
 725                  * We must have resumed after rebooting; reset the vdev
 726                  * stats to know that we're doing a scrub (although it
 727                  * will think we're just starting now).
 728                  */
 729                 vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev,
 730                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 731                     POOL_SCRUB_EVERYTHING, B_FALSE);
 732         }
 733
 734         dp->dp_scrub_pausing = B_FALSE;
 735         dp->dp_scrub_start_time = lbolt64;
 736         dp->dp_scrub_isresilver = (dp->dp_scrub_min_txg != 0);
 737         dp->dp_spa->spa_scrub_active = B_TRUE;
 738
 739         if (dp->dp_scrub_bookmark.zb_objset == 0) {
 740                 /* First do the MOS & ORIGIN */
 741                 scrub_visit_rootbp(dp, NULL, &dp->dp_meta_rootbp);
 742                 if (dp->dp_scrub_pausing)
 743                         goto out;
 744
 745                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
 746                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
 747                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
 748                 } else {
 749                         scrub_visitds(dp, dp->dp_origin_snap->ds_object, tx);
 750                 }
 751                 ASSERT(!dp->dp_scrub_pausing);
 752         } else if (dp->dp_scrub_bookmark.zb_objset != -1ULL) {
 753                 /*
 754                  * If we were paused, continue from here.  Note if the
 755                  * ds we were paused on was deleted, the zb_objset will
 756                  * be -1, so we will skip this and find a new objset
 757                  * below.
 758                  */
 759                 scrub_visitds(dp, dp->dp_scrub_bookmark.zb_objset, tx);
 760                 if (dp->dp_scrub_pausing)
 761                         goto out;
 762         }
 763
 764         /*
 765          * In case we were paused right at the end of the ds, zero the
 766          * bookmark so we don't think that we're still trying to resume.
 767          */
 768         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 769
 770         /* keep pulling things out of the zap-object-as-queue */
 771         while (zap_cursor_init(&zc, dp->dp_meta_objset, dp->dp_scrub_queue_obj),
 772             zap_cursor_retrieve(&zc, &za) == 0) {
 773                 VERIFY(0 == zap_remove(dp->dp_meta_objset,
 774                     dp->dp_scrub_queue_obj, za.za_name, tx));
 775                 scrub_visitds(dp, za.za_first_integer, tx);
 776                 if (dp->dp_scrub_pausing)
 777                         break;
 778                 zap_cursor_fini(&zc);
 779         }
 780         zap_cursor_fini(&zc);
 781         if (dp->dp_scrub_pausing)
 782                 goto out;
 783
 784         /* done. */
 785
 786         dsl_pool_scrub_cancel_sync(dp, &complete, kcred, tx);
 787         return;
 788 out:
 789         VERIFY(0 == zap_update(dp->dp_meta_objset,
 790             DMU_POOL_DIRECTORY_OBJECT,
 791             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 792             &dp->dp_scrub_bookmark, tx));
 793         VERIFY(0 == zap_update(dp->dp_meta_objset,
 794             DMU_POOL_DIRECTORY_OBJECT,
 795             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 796             &dp->dp_spa->spa_scrub_errors, tx));
 797
 798         /* XXX this is scrub-clean specific */
 799         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 800         while (dp->dp_spa->spa_scrub_inflight > 0) {
 801                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 802                     &dp->dp_spa->spa_scrub_lock);
 803         }
 804         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 805 }
 806
 807 void
 808 dsl_pool_scrub_restart(dsl_pool_t *dp)
 809 {
 810         mutex_enter(&dp->dp_scrub_cancel_lock);
 811         dp->dp_scrub_restart = B_TRUE;
 812         mutex_exit(&dp->dp_scrub_cancel_lock);
 813 }
 814
 815 /*
 816  * scrub consumers
 817  */
 818
 819 static void
 820 dsl_pool_scrub_clean_done(zio_t *zio)
 821 {
 822         spa_t *spa = zio->io_spa;
 823
 824         zio_data_buf_free(zio->io_data, zio->io_size);
 825
 826         mutex_enter(&spa->spa_scrub_lock);
 827         spa->spa_scrub_inflight--;
 828         cv_broadcast(&spa->spa_scrub_io_cv);
 829
 830         if (zio->io_error && (zio->io_error != ECKSUM ||
 831             !(zio->io_flags & ZIO_FLAG_SPECULATIVE)))
 832                 spa->spa_scrub_errors++;
 833         mutex_exit(&spa->spa_scrub_lock);
 834 }
 835
 836 static int
 837 dsl_pool_scrub_clean_cb(dsl_pool_t *dp,
 838     const blkptr_t *bp, const zbookmark_t *zb)
 839 {
 840         size_t size = BP_GET_LSIZE(bp);
 841         int d;
 842         spa_t *spa = dp->dp_spa;
 843         boolean_t needs_io;
 844         int zio_flags = ZIO_FLAG_SCRUB_THREAD | ZIO_FLAG_CANFAIL;
 845         int zio_priority;
 846
 847         if (dp->dp_scrub_isresilver == 0) {
 848                 /* It's a scrub */
 849                 zio_flags |= ZIO_FLAG_SCRUB;
 850                 zio_priority = ZIO_PRIORITY_SCRUB;
 851                 needs_io = B_TRUE;
 852         } else {
 853                 /* It's a resilver */
 854                 zio_flags |= ZIO_FLAG_RESILVER;
 855                 zio_priority = ZIO_PRIORITY_RESILVER;
 856                 needs_io = B_FALSE;
 857         }
 858
 859         /* If it's an intent log block, failure is expected. */
 860         if (zb->zb_level == -1 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)
 861                 zio_flags |= ZIO_FLAG_SPECULATIVE;
 862
 863         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
 864                 vdev_t *vd = vdev_lookup_top(spa,
 865                     DVA_GET_VDEV(&bp->blk_dva[d]));
 866
 867                 /*
 868                  * Keep track of how much data we've examined so that
 869                  * zpool(1M) status can make useful progress reports.
 870                  */
 871                 mutex_enter(&vd->vdev_stat_lock);
 872                 vd->vdev_stat.vs_scrub_examined +=
 873                     DVA_GET_ASIZE(&bp->blk_dva[d]);
 874                 mutex_exit(&vd->vdev_stat_lock);
 875
 876                 /* if it's a resilver, this may not be in the target range */
 877                 if (!needs_io) {
 878                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
 879                                 /*
 880                                  * Gang members may be spread across multiple
 881                                  * vdevs, so the best we can do is look at the
 882                                  * pool-wide DTL.
 883                                  * XXX -- it would be better to change our
 884                                  * allocation policy to ensure that this can't
 885                                  * happen.
 886                                  */
 887                                 vd = spa->spa_root_vdev;
 888                         }
 889                         needs_io = vdev_dtl_contains(&vd->vdev_dtl_map,
 890                             bp->blk_birth, 1);
 891                 }
 892         }
 893
 894         if (needs_io && !zfs_no_scrub_io) {
 895                 void *data = zio_data_buf_alloc(size);
 896
 897                 mutex_enter(&spa->spa_scrub_lock);
 898                 while (spa->spa_scrub_inflight >= spa->spa_scrub_maxinflight)
 899                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 900                 spa->spa_scrub_inflight++;
 901                 mutex_exit(&spa->spa_scrub_lock);
 902
 903                 zio_nowait(zio_read(NULL, spa, bp, data, size,
 904                     dsl_pool_scrub_clean_done, NULL, zio_priority,
 905                     zio_flags, zb));
 906         }
 907
 908         /* do not relocate this block */
 909         return (0);
 910 }
 911
 912 int
 913 dsl_pool_scrub_clean(dsl_pool_t *dp)
 914 {
 915         /*
 916          * Purge all vdev caches.  We do this here rather than in sync
 917          * context because this requires a writer lock on the spa_config
 918          * lock, which we can't do from sync context.  The
 919          * spa_scrub_reopen flag indicates that vdev_open() should not
 920          * attempt to start another scrub.
 921          */
 922         spa_config_enter(dp->dp_spa, SCL_ALL, FTAG, RW_WRITER);
 923         dp->dp_spa->spa_scrub_reopen = B_TRUE;
 924         vdev_reopen(dp->dp_spa->spa_root_vdev);
 925         dp->dp_spa->spa_scrub_reopen = B_FALSE;
 926         spa_config_exit(dp->dp_spa, SCL_ALL, FTAG);
 927
 928         return (dsl_pool_scrub_setup(dp, SCRUB_FUNC_CLEAN));
 929 }