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 (dp->dp_blkstats == NULL) {
 111                 dp->dp_blkstats =
 112                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 113         }
 114         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 115
 116         if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB)
 117                 ot = DMU_OT_ZAP_OTHER;
 118
 119         dp->dp_scrub_func = *funcp;
 120         dp->dp_scrub_queue_obj = zap_create(dp->dp_meta_objset,
 121             ot ? ot : DMU_OT_SCRUB_QUEUE, DMU_OT_NONE, 0, tx);
 122         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 123         dp->dp_scrub_restart = B_FALSE;
 124         dp->dp_spa->spa_scrub_errors = 0;
 125
 126         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 127             DMU_POOL_SCRUB_FUNC, sizeof (uint32_t), 1,
 128             &dp->dp_scrub_func, tx));
 129         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 130             DMU_POOL_SCRUB_QUEUE, sizeof (uint64_t), 1,
 131             &dp->dp_scrub_queue_obj, tx));
 132         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 133             DMU_POOL_SCRUB_MIN_TXG, sizeof (uint64_t), 1,
 134             &dp->dp_scrub_min_txg, tx));
 135         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 136             DMU_POOL_SCRUB_MAX_TXG, sizeof (uint64_t), 1,
 137             &dp->dp_scrub_max_txg, tx));
 138         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 139             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 140             &dp->dp_scrub_bookmark, tx));
 141         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 142             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 143             &dp->dp_spa->spa_scrub_errors, tx));
 144
 145         spa_history_internal_log(LOG_POOL_SCRUB, dp->dp_spa, tx, cr,
 146             "func=%u mintxg=%llu maxtxg=%llu",
 147             *funcp, dp->dp_scrub_min_txg, dp->dp_scrub_max_txg);
 148 }
 149
 150 int
 151 dsl_pool_scrub_setup(dsl_pool_t *dp, enum scrub_func func)
 152 {
 153         return (dsl_sync_task_do(dp, NULL,
 154             dsl_pool_scrub_setup_sync, dp, &func, 0));
 155 }
 156
 157 /* ARGSUSED */
 158 static void
 159 dsl_pool_scrub_cancel_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
 160 {
 161         dsl_pool_t *dp = arg1;
 162         boolean_t *completep = arg2;
 163
 164         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 165                 return;
 166
 167         mutex_enter(&dp->dp_scrub_cancel_lock);
 168
 169         if (dp->dp_scrub_restart) {
 170                 dp->dp_scrub_restart = B_FALSE;
 171                 *completep = B_FALSE;
 172         }
 173
 174         /* XXX this is scrub-clean specific */
 175         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 176         while (dp->dp_spa->spa_scrub_inflight > 0) {
 177                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 178                     &dp->dp_spa->spa_scrub_lock);
 179         }
 180         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 181         dp->dp_spa->spa_scrub_started = B_FALSE;
 182         dp->dp_spa->spa_scrub_active = B_FALSE;
 183
 184         dp->dp_scrub_func = SCRUB_FUNC_NONE;
 185         VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 186             dp->dp_scrub_queue_obj, tx));
 187         dp->dp_scrub_queue_obj = 0;
 188         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 189
 190         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 191             DMU_POOL_SCRUB_QUEUE, tx));
 192         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 193             DMU_POOL_SCRUB_MIN_TXG, tx));
 194         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 195             DMU_POOL_SCRUB_MAX_TXG, tx));
 196         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 197             DMU_POOL_SCRUB_BOOKMARK, tx));
 198         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 199             DMU_POOL_SCRUB_FUNC, tx));
 200         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 201             DMU_POOL_SCRUB_ERRORS, tx));
 202
 203         spa_history_internal_log(LOG_POOL_SCRUB_DONE, dp->dp_spa, tx, cr,
 204             "complete=%u", *completep);
 205
 206         /* below is scrub-clean specific */
 207         vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev, POOL_SCRUB_NONE,
 208             *completep);
 209         /*
 210          * If the scrub/resilver completed, update all DTLs to reflect this.
 211          * Whether it succeeded or not, vacate all temporary scrub DTLs.
 212          */
 213         vdev_dtl_reassess(dp->dp_spa->spa_root_vdev, tx->tx_txg,
 214             *completep ? dp->dp_scrub_max_txg : 0, B_TRUE);
 215         if (dp->dp_scrub_min_txg && *completep)
 216                 spa_event_notify(dp->dp_spa, NULL, ESC_ZFS_RESILVER_FINISH);
 217         spa_errlog_rotate(dp->dp_spa);
 218
 219         /*
 220          * We may have finished replacing a device.
 221          * Let the async thread assess this and handle the detach.
 222          */
 223         spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER_DONE);
 224
 225         dp->dp_scrub_min_txg = dp->dp_scrub_max_txg = 0;
 226         mutex_exit(&dp->dp_scrub_cancel_lock);
 227 }
 228
 229 int
 230 dsl_pool_scrub_cancel(dsl_pool_t *dp)
 231 {
 232         boolean_t complete = B_FALSE;
 233
 234         return (dsl_sync_task_do(dp, NULL,
 235             dsl_pool_scrub_cancel_sync, dp, &complete, 3));
 236 }
 237
 238 int
 239 dsl_free(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp,
 240     zio_done_func_t *done, void *private, uint32_t arc_flags)
 241 {
 242         /*
 243          * This function will be used by bp-rewrite wad to intercept frees.
 244          */
 245         return (arc_free(pio, dp->dp_spa, txg, (blkptr_t *)bpp,
 246             done, private, arc_flags));
 247 }
 248
 249 static boolean_t
 250 bookmark_is_zero(const zbookmark_t *zb)
 251 {
 252         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 253             zb->zb_level == 0 && zb->zb_blkid == 0);
 254 }
 255
 256 /* dnp is the dnode for zb1->zb_object */
 257 static boolean_t
 258 bookmark_is_before(dnode_phys_t *dnp, const zbookmark_t *zb1,
 259     const zbookmark_t *zb2)
 260 {
 261         uint64_t zb1nextL0, zb2thisobj;
 262
 263         ASSERT(zb1->zb_objset == zb2->zb_objset);
 264         ASSERT(zb1->zb_object != -1ULL);
 265         ASSERT(zb2->zb_level == 0);
 266
 267         /*
 268          * A bookmark in the deadlist is considered to be after
 269          * everything else.
 270          */
 271         if (zb2->zb_object == -1ULL)
 272                 return (B_TRUE);
 273
 274         /* The objset_phys_t isn't before anything. */
 275         if (dnp == NULL)
 276                 return (B_FALSE);
 277
 278         zb1nextL0 = (zb1->zb_blkid + 1) <<
 279             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 280
 281         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 282             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 283
 284         if (zb1->zb_object == 0) {
 285                 uint64_t nextobj = zb1nextL0 *
 286                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 287                 return (nextobj <= zb2thisobj);
 288         }
 289
 290         if (zb1->zb_object < zb2thisobj)
 291                 return (B_TRUE);
 292         if (zb1->zb_object > zb2thisobj)
 293                 return (B_FALSE);
 294         if (zb2->zb_object == 0)
 295                 return (B_FALSE);
 296         return (zb1nextL0 <= zb2->zb_blkid);
 297 }
 298
 299 static boolean_t
 300 scrub_pause(dsl_pool_t *dp, const zbookmark_t *zb)
 301 {
 302         int elapsed_ticks;
 303         int mintime;
 304
 305         if (dp->dp_scrub_pausing)
 306                 return (B_TRUE); /* we're already pausing */
 307
 308         if (!bookmark_is_zero(&dp->dp_scrub_bookmark))
 309                 return (B_FALSE); /* we're resuming */
 310
 311         /* We only know how to resume from level-0 blocks. */
 312         if (zb->zb_level != 0)
 313                 return (B_FALSE);
 314
 315         mintime = dp->dp_scrub_isresilver ? zfs_resilver_min_time :
 316             zfs_scrub_min_time;
 317         elapsed_ticks = lbolt64 - dp->dp_scrub_start_time;
 318         if (elapsed_ticks > hz * zfs_txg_timeout ||
 319             (elapsed_ticks > hz * mintime && txg_sync_waiting(dp))) {
 320                 dprintf("pausing at %llx/%llx/%llx/%llx\n",
 321                     (longlong_t)zb->zb_objset, (longlong_t)zb->zb_object,
 322                     (longlong_t)zb->zb_level, (longlong_t)zb->zb_blkid);
 323                 dp->dp_scrub_pausing = B_TRUE;
 324                 dp->dp_scrub_bookmark = *zb;
 325                 return (B_TRUE);
 326         }
 327         return (B_FALSE);
 328 }
 329
 330 typedef struct zil_traverse_arg {
 331         dsl_pool_t      *zta_dp;
 332         zil_header_t    *zta_zh;
 333 } zil_traverse_arg_t;
 334
 335 /* ARGSUSED */
 336 static void
 337 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 338 {
 339         zil_traverse_arg_t *zta = arg;
 340         dsl_pool_t *dp = zta->zta_dp;
 341         zil_header_t *zh = zta->zta_zh;
 342         zbookmark_t zb;
 343
 344         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 345                 return;
 346
 347         /*
 348          * One block ("stumpy") can be allocated a long time ago; we
 349          * want to visit that one because it has been allocated
 350          * (on-disk) even if it hasn't been claimed (even though for
 351          * plain scrub there's nothing to do to it).
 352          */
 353         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 354                 return;
 355
 356         zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 357         zb.zb_object = 0;
 358         zb.zb_level = -1;
 359         zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
 360         VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 361 }
 362
 363 /* ARGSUSED */
 364 static void
 365 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 366 {
 367         if (lrc->lrc_txtype == TX_WRITE) {
 368                 zil_traverse_arg_t *zta = arg;
 369                 dsl_pool_t *dp = zta->zta_dp;
 370                 zil_header_t *zh = zta->zta_zh;
 371                 lr_write_t *lr = (lr_write_t *)lrc;
 372                 blkptr_t *bp = &lr->lr_blkptr;
 373                 zbookmark_t zb;
 374
 375                 if (bp->blk_birth <= dp->dp_scrub_min_txg)
 376                         return;
 377
 378                 /*
 379                  * birth can be < claim_txg if this record's txg is
 380                  * already txg sync'ed (but this log block contains
 381                  * other records that are not synced)
 382                  */
 383                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 384                         return;
 385
 386                 zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 387                 zb.zb_object = lr->lr_foid;
 388                 zb.zb_level = BP_GET_LEVEL(bp);
 389                 zb.zb_blkid = lr->lr_offset / BP_GET_LSIZE(bp);
 390                 VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 391         }
 392 }
 393
 394 static void
 395 traverse_zil(dsl_pool_t *dp, zil_header_t *zh)
 396 {
 397         uint64_t claim_txg = zh->zh_claim_txg;
 398         zil_traverse_arg_t zta = { dp, zh };
 399         zilog_t *zilog;
 400
 401         /*
 402          * We only want to visit blocks that have been claimed but not yet
 403          * replayed (or, in read-only mode, blocks that *would* be claimed).
 404          */
 405         if (claim_txg == 0 && (spa_mode & FWRITE))
 406                 return;
 407
 408         zilog = zil_alloc(dp->dp_meta_objset, zh);
 409
 410         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, &zta,
 411             claim_txg);
 412
 413         zil_free(zilog);
 414 }
 415
 416 static void
 417 scrub_visitbp(dsl_pool_t *dp, dnode_phys_t *dnp,
 418     arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb)
 419 {
 420         int err;
 421         arc_buf_t *buf = NULL;
 422
 423         if (bp->blk_birth == 0)
 424                 return;
 425
 426         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 427                 return;
 428
 429         if (scrub_pause(dp, zb))
 430                 return;
 431
 432         if (!bookmark_is_zero(&dp->dp_scrub_bookmark)) {
 433                 /*
 434                  * If we already visited this bp & everything below (in
 435                  * a prior txg), don't bother doing it again.
 436                  */
 437                 if (bookmark_is_before(dnp, zb, &dp->dp_scrub_bookmark))
 438                         return;
 439
 440                 /*
 441                  * If we found the block we're trying to resume from, or
 442                  * we went past it to a different object, zero it out to
 443                  * indicate that it's OK to start checking for pausing
 444                  * again.
 445                  */
 446                 if (bcmp(zb, &dp->dp_scrub_bookmark, sizeof (*zb)) == 0 ||
 447                     zb->zb_object > dp->dp_scrub_bookmark.zb_object) {
 448                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 449                             (longlong_t)zb->zb_objset,
 450                             (longlong_t)zb->zb_object,
 451                             (longlong_t)zb->zb_level,
 452                             (longlong_t)zb->zb_blkid);
 453                         bzero(&dp->dp_scrub_bookmark, sizeof (*zb));
 454                 }
 455         }
 456
 457         if (BP_GET_LEVEL(bp) > 0) {
 458                 uint32_t flags = ARC_WAIT;
 459                 int i;
 460                 blkptr_t *cbp;
 461                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 462
 463                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 464                     arc_getbuf_func, &buf,
 465                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 466                 if (err) {
 467                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 468                         dp->dp_spa->spa_scrub_errors++;
 469                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 470                         return;
 471                 }
 472                 cbp = buf->b_data;
 473
 474                 for (i = 0; i < epb; i++, cbp++) {
 475                         zbookmark_t czb;
 476
 477                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 478                             zb->zb_level - 1,
 479                             zb->zb_blkid * epb + i);
 480                         scrub_visitbp(dp, dnp, buf, cbp, &czb);
 481                 }
 482         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 483                 uint32_t flags = ARC_WAIT;
 484                 dnode_phys_t *child_dnp;
 485                 int i, j;
 486                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 487
 488                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 489                     arc_getbuf_func, &buf,
 490                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 491                 if (err) {
 492                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 493                         dp->dp_spa->spa_scrub_errors++;
 494                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 495                         return;
 496                 }
 497                 child_dnp = buf->b_data;
 498
 499                 for (i = 0; i < epb; i++, child_dnp++) {
 500                         for (j = 0; j < child_dnp->dn_nblkptr; j++) {
 501                                 zbookmark_t czb;
 502
 503                                 SET_BOOKMARK(&czb, zb->zb_objset,
 504                                     zb->zb_blkid * epb + i,
 505                                     child_dnp->dn_nlevels - 1, j);
 506                                 scrub_visitbp(dp, child_dnp, buf,
 507                                     &child_dnp->dn_blkptr[j], &czb);
 508                         }
 509                 }
 510         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 511                 uint32_t flags = ARC_WAIT;
 512                 objset_phys_t *osp;
 513                 int j;
 514
 515                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 516                     arc_getbuf_func, &buf,
 517                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 518                 if (err) {
 519                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 520                         dp->dp_spa->spa_scrub_errors++;
 521                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 522                         return;
 523                 }
 524
 525                 osp = buf->b_data;
 526
 527                 traverse_zil(dp, &osp->os_zil_header);
 528
 529                 for (j = 0; j < osp->os_meta_dnode.dn_nblkptr; j++) {
 530                         zbookmark_t czb;
 531
 532                         SET_BOOKMARK(&czb, zb->zb_objset, 0,
 533                             osp->os_meta_dnode.dn_nlevels - 1, j);
 534                         scrub_visitbp(dp, &osp->os_meta_dnode, buf,
 535                             &osp->os_meta_dnode.dn_blkptr[j], &czb);
 536                 }
 537         }
 538
 539         (void) scrub_funcs[dp->dp_scrub_func](dp, bp, zb);
 540         if (buf)
 541                 (void) arc_buf_remove_ref(buf, &buf);
 542 }
 543
 544 static void
 545 scrub_visit_rootbp(dsl_pool_t *dp, dsl_dataset_t *ds, blkptr_t *bp)
 546 {
 547         zbookmark_t zb;
 548
 549         SET_BOOKMARK(&zb, ds ? ds->ds_object : 0, 0, -1, 0);
 550         scrub_visitbp(dp, NULL, NULL, bp, &zb);
 551 }
 552
 553 void
 554 dsl_pool_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 555 {
 556         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 557
 558         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 559                 return;
 560
 561         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 562                 SET_BOOKMARK(&dp->dp_scrub_bookmark, -1, 0, 0, 0);
 563         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 564             ds->ds_object, tx) != 0) {
 565                 return;
 566         }
 567
 568         if (ds->ds_phys->ds_next_snap_obj != 0) {
 569                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 570                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 571         }
 572         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 573 }
 574
 575 void
 576 dsl_pool_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 577 {
 578         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 579
 580         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 581                 return;
 582
 583         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 584
 585         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 586                 dp->dp_scrub_bookmark.zb_objset =
 587                     ds->ds_phys->ds_prev_snap_obj;
 588         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 589             ds->ds_object, tx) == 0) {
 590                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 591                     ds->ds_phys->ds_prev_snap_obj, tx) == 0);
 592         }
 593 }
 594
 595 void
 596 dsl_pool_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 597 {
 598         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 599
 600         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 601                 return;
 602
 603         if (dp->dp_scrub_bookmark.zb_objset == ds1->ds_object) {
 604                 dp->dp_scrub_bookmark.zb_objset = ds2->ds_object;
 605         } else if (dp->dp_scrub_bookmark.zb_objset == ds2->ds_object) {
 606                 dp->dp_scrub_bookmark.zb_objset = ds1->ds_object;
 607         }
 608
 609         if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 610             ds1->ds_object, tx) == 0) {
 611                 int err = zap_add_int(dp->dp_meta_objset,
 612                     dp->dp_scrub_queue_obj, ds2->ds_object, tx);
 613                 VERIFY(err == 0 || err == EEXIST);
 614                 if (err == EEXIST) {
 615                         /* Both were there to begin with */
 616                         VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 617                             dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 618                 }
 619         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 620             ds2->ds_object, tx) == 0) {
 621                 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 622                     dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 623         }
 624 }
 625
 626 struct enqueue_clones_arg {
 627         dmu_tx_t *tx;
 628         uint64_t originobj;
 629 };
 630
 631 /* ARGSUSED */
 632 static int
 633 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 634 {
 635         struct enqueue_clones_arg *eca = arg;
 636         dsl_dataset_t *ds;
 637         int err;
 638         dsl_pool_t *dp;
 639
 640         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 641         if (err)
 642                 return (err);
 643         dp = ds->ds_dir->dd_pool;
 644
 645         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 646                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 647                         dsl_dataset_t *prev;
 648                         err = dsl_dataset_hold_obj(dp,
 649                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 650
 651                         dsl_dataset_rele(ds, FTAG);
 652                         if (err)
 653                                 return (err);
 654                         ds = prev;
 655                 }
 656                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 657                     ds->ds_object, eca->tx) == 0);
 658         }
 659         dsl_dataset_rele(ds, FTAG);
 660         return (0);
 661 }
 662
 663 static void
 664 scrub_visitds(dsl_pool_t *dp, uint64_t dsobj, dmu_tx_t *tx)
 665 {
 666         dsl_dataset_t *ds;
 667         uint64_t min_txg_save;
 668
 669         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 670
 671         /*
 672          * Iterate over the bps in this ds.
 673          */
 674         min_txg_save = dp->dp_scrub_min_txg;
 675         dp->dp_scrub_min_txg =
 676             MAX(dp->dp_scrub_min_txg, ds->ds_phys->ds_prev_snap_txg);
 677         scrub_visit_rootbp(dp, ds, &ds->ds_phys->ds_bp);
 678         dp->dp_scrub_min_txg = min_txg_save;
 679
 680         if (dp->dp_scrub_pausing)
 681                 goto out;
 682
 683         /*
 684          * Add descendent datasets to work queue.
 685          */
 686         if (ds->ds_phys->ds_next_snap_obj != 0) {
 687                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 688                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 689         }
 690         if (ds->ds_phys->ds_num_children > 1) {
 691                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
 692                         struct enqueue_clones_arg eca;
 693                         eca.tx = tx;
 694                         eca.originobj = ds->ds_object;
 695
 696                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
 697                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
 698                 } else {
 699                         VERIFY(zap_join(dp->dp_meta_objset,
 700                             ds->ds_phys->ds_next_clones_obj,
 701                             dp->dp_scrub_queue_obj, tx) == 0);
 702                 }
 703         }
 704
 705 out:
 706         dsl_dataset_rele(ds, FTAG);
 707 }
 708
 709 /* ARGSUSED */
 710 static int
 711 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 712 {
 713         dmu_tx_t *tx = arg;
 714         dsl_dataset_t *ds;
 715         int err;
 716         dsl_pool_t *dp;
 717
 718         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 719         if (err)
 720                 return (err);
 721
 722         dp = ds->ds_dir->dd_pool;
 723
 724         while (ds->ds_phys->ds_prev_snap_obj != 0) {
 725                 dsl_dataset_t *prev;
 726                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
 727                     FTAG, &prev);
 728                 if (err) {
 729                         dsl_dataset_rele(ds, FTAG);
 730                         return (err);
 731                 }
 732
 733                 /*
 734                  * If this is a clone, we don't need to worry about it for now.
 735                  */
 736                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
 737                         dsl_dataset_rele(ds, FTAG);
 738                         dsl_dataset_rele(prev, FTAG);
 739                         return (0);
 740                 }
 741                 dsl_dataset_rele(ds, FTAG);
 742                 ds = prev;
 743         }
 744
 745         VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 746             ds->ds_object, tx) == 0);
 747         dsl_dataset_rele(ds, FTAG);
 748         return (0);
 749 }
 750
 751 void
 752 dsl_pool_scrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
 753 {
 754         zap_cursor_t zc;
 755         zap_attribute_t za;
 756         boolean_t complete = B_TRUE;
 757
 758         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 759                 return;
 760
 761         /* If the spa is not fully loaded, don't bother. */
 762         if (dp->dp_spa->spa_load_state != SPA_LOAD_NONE)
 763                 return;
 764
 765         if (dp->dp_scrub_restart) {
 766                 enum scrub_func func = dp->dp_scrub_func;
 767                 dp->dp_scrub_restart = B_FALSE;
 768                 dsl_pool_scrub_setup_sync(dp, &func, kcred, tx);
 769         }
 770
 771         if (dp->dp_spa->spa_root_vdev->vdev_stat.vs_scrub_type == 0) {
 772                 /*
 773                  * We must have resumed after rebooting; reset the vdev
 774                  * stats to know that we're doing a scrub (although it
 775                  * will think we're just starting now).
 776                  */
 777                 vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev,
 778                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 779                     POOL_SCRUB_EVERYTHING, B_FALSE);
 780         }
 781
 782         dp->dp_scrub_pausing = B_FALSE;
 783         dp->dp_scrub_start_time = lbolt64;
 784         dp->dp_scrub_isresilver = (dp->dp_scrub_min_txg != 0);
 785         dp->dp_spa->spa_scrub_active = B_TRUE;
 786
 787         if (dp->dp_scrub_bookmark.zb_objset == 0) {
 788                 /* First do the MOS & ORIGIN */
 789                 scrub_visit_rootbp(dp, NULL, &dp->dp_meta_rootbp);
 790                 if (dp->dp_scrub_pausing)
 791                         goto out;
 792
 793                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
 794                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
 795                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
 796                 } else {
 797                         scrub_visitds(dp, dp->dp_origin_snap->ds_object, tx);
 798                 }
 799                 ASSERT(!dp->dp_scrub_pausing);
 800         } else if (dp->dp_scrub_bookmark.zb_objset != -1ULL) {
 801                 /*
 802                  * If we were paused, continue from here.  Note if the
 803                  * ds we were paused on was deleted, the zb_objset will
 804                  * be -1, so we will skip this and find a new objset
 805                  * below.
 806                  */
 807                 scrub_visitds(dp, dp->dp_scrub_bookmark.zb_objset, tx);
 808                 if (dp->dp_scrub_pausing)
 809                         goto out;
 810         }
 811
 812         /*
 813          * In case we were paused right at the end of the ds, zero the
 814          * bookmark so we don't think that we're still trying to resume.
 815          */
 816         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 817
 818         /* keep pulling things out of the zap-object-as-queue */
 819         while (zap_cursor_init(&zc, dp->dp_meta_objset, dp->dp_scrub_queue_obj),
 820             zap_cursor_retrieve(&zc, &za) == 0) {
 821                 VERIFY(0 == zap_remove(dp->dp_meta_objset,
 822                     dp->dp_scrub_queue_obj, za.za_name, tx));
 823                 scrub_visitds(dp, za.za_first_integer, tx);
 824                 if (dp->dp_scrub_pausing)
 825                         break;
 826                 zap_cursor_fini(&zc);
 827         }
 828         zap_cursor_fini(&zc);
 829         if (dp->dp_scrub_pausing)
 830                 goto out;
 831
 832         /* done. */
 833
 834         dsl_pool_scrub_cancel_sync(dp, &complete, kcred, tx);
 835         return;
 836 out:
 837         VERIFY(0 == zap_update(dp->dp_meta_objset,
 838             DMU_POOL_DIRECTORY_OBJECT,
 839             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 840             &dp->dp_scrub_bookmark, tx));
 841         VERIFY(0 == zap_update(dp->dp_meta_objset,
 842             DMU_POOL_DIRECTORY_OBJECT,
 843             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 844             &dp->dp_spa->spa_scrub_errors, tx));
 845
 846         /* XXX this is scrub-clean specific */
 847         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 848         while (dp->dp_spa->spa_scrub_inflight > 0) {
 849                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 850                     &dp->dp_spa->spa_scrub_lock);
 851         }
 852         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 853 }
 854
 855 void
 856 dsl_pool_scrub_restart(dsl_pool_t *dp)
 857 {
 858         mutex_enter(&dp->dp_scrub_cancel_lock);
 859         dp->dp_scrub_restart = B_TRUE;
 860         mutex_exit(&dp->dp_scrub_cancel_lock);
 861 }
 862
 863 /*
 864  * scrub consumers
 865  */
 866
 867 static void
 868 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
 869 {
 870         int i;
 871
 872         /*
 873          * If we resume after a reboot, zab will be NULL; don't record
 874          * incomplete stats in that case.
 875          */
 876         if (zab == NULL)
 877                 return;
 878
 879         for (i = 0; i < 4; i++) {
 880                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
 881                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
 882                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
 883                 int equal;
 884
 885                 zb->zb_count++;
 886                 zb->zb_asize += BP_GET_ASIZE(bp);
 887                 zb->zb_lsize += BP_GET_LSIZE(bp);
 888                 zb->zb_psize += BP_GET_PSIZE(bp);
 889                 zb->zb_gangs += BP_COUNT_GANG(bp);
 890
 891                 switch (BP_GET_NDVAS(bp)) {
 892                 case 2:
 893                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 894                             DVA_GET_VDEV(&bp->blk_dva[1]))
 895                                 zb->zb_ditto_2_of_2_samevdev++;
 896                         break;
 897                 case 3:
 898                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 899                             DVA_GET_VDEV(&bp->blk_dva[1])) +
 900                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 901                             DVA_GET_VDEV(&bp->blk_dva[2])) +
 902                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
 903                             DVA_GET_VDEV(&bp->blk_dva[2]));
 904                         if (equal == 1)
 905                                 zb->zb_ditto_2_of_3_samevdev++;
 906                         else if (equal == 3)
 907                                 zb->zb_ditto_3_of_3_samevdev++;
 908                         break;
 909                 }
 910         }
 911 }
 912
 913 static void
 914 dsl_pool_scrub_clean_done(zio_t *zio)
 915 {
 916         spa_t *spa = zio->io_spa;
 917
 918         zio_data_buf_free(zio->io_data, zio->io_size);
 919
 920         mutex_enter(&spa->spa_scrub_lock);
 921         spa->spa_scrub_inflight--;
 922         cv_broadcast(&spa->spa_scrub_io_cv);
 923
 924         if (zio->io_error && (zio->io_error != ECKSUM ||
 925             !(zio->io_flags & ZIO_FLAG_SPECULATIVE)))
 926                 spa->spa_scrub_errors++;
 927         mutex_exit(&spa->spa_scrub_lock);
 928 }
 929
 930 static int
 931 dsl_pool_scrub_clean_cb(dsl_pool_t *dp,
 932     const blkptr_t *bp, const zbookmark_t *zb)
 933 {
 934         size_t size = BP_GET_LSIZE(bp);
 935         int d;
 936         spa_t *spa = dp->dp_spa;
 937         boolean_t needs_io;
 938         int zio_flags = ZIO_FLAG_SCRUB_THREAD | ZIO_FLAG_CANFAIL;
 939         int zio_priority;
 940
 941         count_block(dp->dp_blkstats, bp);
 942
 943         if (dp->dp_scrub_isresilver == 0) {
 944                 /* It's a scrub */
 945                 zio_flags |= ZIO_FLAG_SCRUB;
 946                 zio_priority = ZIO_PRIORITY_SCRUB;
 947                 needs_io = B_TRUE;
 948         } else {
 949                 /* It's a resilver */
 950                 zio_flags |= ZIO_FLAG_RESILVER;
 951                 zio_priority = ZIO_PRIORITY_RESILVER;
 952                 needs_io = B_FALSE;
 953         }
 954
 955         /* If it's an intent log block, failure is expected. */
 956         if (zb->zb_level == -1 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)
 957                 zio_flags |= ZIO_FLAG_SPECULATIVE;
 958
 959         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
 960                 vdev_t *vd = vdev_lookup_top(spa,
 961                     DVA_GET_VDEV(&bp->blk_dva[d]));
 962
 963                 /*
 964                  * Keep track of how much data we've examined so that
 965                  * zpool(1M) status can make useful progress reports.
 966                  */
 967                 mutex_enter(&vd->vdev_stat_lock);
 968                 vd->vdev_stat.vs_scrub_examined +=
 969                     DVA_GET_ASIZE(&bp->blk_dva[d]);
 970                 mutex_exit(&vd->vdev_stat_lock);
 971
 972                 /* if it's a resilver, this may not be in the target range */
 973                 if (!needs_io) {
 974                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
 975                                 /*
 976                                  * Gang members may be spread across multiple
 977                                  * vdevs, so the best we can do is look at the
 978                                  * pool-wide DTL.
 979                                  * XXX -- it would be better to change our
 980                                  * allocation policy to ensure that this can't
 981                                  * happen.
 982                                  */
 983                                 vd = spa->spa_root_vdev;
 984                         }
 985                         needs_io = vdev_dtl_contains(&vd->vdev_dtl_map,
 986                             bp->blk_birth, 1);
 987                 }
 988         }
 989
 990         if (needs_io && !zfs_no_scrub_io) {
 991                 void *data = zio_data_buf_alloc(size);
 992
 993                 mutex_enter(&spa->spa_scrub_lock);
 994                 while (spa->spa_scrub_inflight >= spa->spa_scrub_maxinflight)
 995                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 996                 spa->spa_scrub_inflight++;
 997                 mutex_exit(&spa->spa_scrub_lock);
 998
 999                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1000                     dsl_pool_scrub_clean_done, NULL, zio_priority,
1001                     zio_flags, zb));
1002         }
1003
1004         /* do not relocate this block */
1005         return (0);
1006 }
1007
1008 int
1009 dsl_pool_scrub_clean(dsl_pool_t *dp)
1010 {
1011         /*
1012          * Purge all vdev caches.  We do this here rather than in sync
1013          * context because this requires a writer lock on the spa_config
1014          * lock, which we can't do from sync context.  The
1015          * spa_scrub_reopen flag indicates that vdev_open() should not
1016          * attempt to start another scrub.
1017          */
1018         spa_config_enter(dp->dp_spa, SCL_ALL, FTAG, RW_WRITER);
1019         dp->dp_spa->spa_scrub_reopen = B_TRUE;
1020         vdev_reopen(dp->dp_spa->spa_root_vdev);
1021         dp->dp_spa->spa_scrub_reopen = B_FALSE;
1022         spa_config_exit(dp->dp_spa, SCL_ALL, FTAG);
1023
1024         return (dsl_pool_scrub_setup(dp, SCRUB_FUNC_CLEAN));
1025 }