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 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 static void scrub_visitdnode(dsl_pool_t *dp, dnode_phys_t *dnp, arc_buf_t *buf,
  49     uint64_t objset, uint64_t object);
  50
  51 int zfs_scrub_min_time = 1; /* scrub for at least 1 sec each txg */
  52 int zfs_resilver_min_time = 3; /* resilver for at least 3 sec each txg */
  53 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  54
  55 extern int zfs_txg_timeout;
  56
  57 static scrub_cb_t *scrub_funcs[SCRUB_FUNC_NUMFUNCS] = {
  58         NULL,
  59         dsl_pool_scrub_clean_cb
  60 };
  61
  62 #define SET_BOOKMARK(zb, objset, object, level, blkid)  \
  63 {                                                       \
  64         (zb)->zb_objset = objset;                       \
  65         (zb)->zb_object = object;                       \
  66         (zb)->zb_level = level;                         \
  67         (zb)->zb_blkid = blkid;                         \
  68 }
  69
  70 /* ARGSUSED */
  71 static void
  72 dsl_pool_scrub_setup_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
  73 {
  74         dsl_pool_t *dp = arg1;
  75         enum scrub_func *funcp = arg2;
  76         dmu_object_type_t ot = 0;
  77         boolean_t complete = B_FALSE;
  78
  79         dsl_pool_scrub_cancel_sync(dp, &complete, cr, tx);
  80
  81         ASSERT(dp->dp_scrub_func == SCRUB_FUNC_NONE);
  82         ASSERT(*funcp > SCRUB_FUNC_NONE);
  83         ASSERT(*funcp < SCRUB_FUNC_NUMFUNCS);
  84
  85         dp->dp_scrub_min_txg = 0;
  86         dp->dp_scrub_max_txg = tx->tx_txg;
  87
  88         if (*funcp == SCRUB_FUNC_CLEAN) {
  89                 vdev_t *rvd = dp->dp_spa->spa_root_vdev;
  90
  91                 /* rewrite all disk labels */
  92                 vdev_config_dirty(rvd);
  93
  94                 if (vdev_resilver_needed(rvd,
  95                     &dp->dp_scrub_min_txg, &dp->dp_scrub_max_txg)) {
  96                         spa_event_notify(dp->dp_spa, NULL,
  97                             ESC_ZFS_RESILVER_START);
  98                         dp->dp_scrub_max_txg = MIN(dp->dp_scrub_max_txg,
  99                             tx->tx_txg);
 100                 } else {
 101                         spa_event_notify(dp->dp_spa, NULL,
 102                             ESC_ZFS_SCRUB_START);
 103                 }
 104
 105                 /* zero out the scrub stats in all vdev_stat_t's */
 106                 vdev_scrub_stat_update(rvd,
 107                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 108                     POOL_SCRUB_EVERYTHING, B_FALSE);
 109
 110                 dp->dp_spa->spa_scrub_started = B_TRUE;
 111         }
 112
 113         /* back to the generic stuff */
 114
 115         if (dp->dp_blkstats == NULL) {
 116                 dp->dp_blkstats =
 117                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 118         }
 119         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 120
 121         if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB)
 122                 ot = DMU_OT_ZAP_OTHER;
 123
 124         dp->dp_scrub_func = *funcp;
 125         dp->dp_scrub_queue_obj = zap_create(dp->dp_meta_objset,
 126             ot ? ot : DMU_OT_SCRUB_QUEUE, DMU_OT_NONE, 0, tx);
 127         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 128         dp->dp_scrub_restart = B_FALSE;
 129         dp->dp_spa->spa_scrub_errors = 0;
 130
 131         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 132             DMU_POOL_SCRUB_FUNC, sizeof (uint32_t), 1,
 133             &dp->dp_scrub_func, tx));
 134         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 135             DMU_POOL_SCRUB_QUEUE, sizeof (uint64_t), 1,
 136             &dp->dp_scrub_queue_obj, tx));
 137         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 138             DMU_POOL_SCRUB_MIN_TXG, sizeof (uint64_t), 1,
 139             &dp->dp_scrub_min_txg, tx));
 140         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 141             DMU_POOL_SCRUB_MAX_TXG, sizeof (uint64_t), 1,
 142             &dp->dp_scrub_max_txg, tx));
 143         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 144             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 145             &dp->dp_scrub_bookmark, tx));
 146         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 147             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 148             &dp->dp_spa->spa_scrub_errors, tx));
 149
 150         spa_history_internal_log(LOG_POOL_SCRUB, dp->dp_spa, tx, cr,
 151             "func=%u mintxg=%llu maxtxg=%llu",
 152             *funcp, dp->dp_scrub_min_txg, dp->dp_scrub_max_txg);
 153 }
 154
 155 int
 156 dsl_pool_scrub_setup(dsl_pool_t *dp, enum scrub_func func)
 157 {
 158         return (dsl_sync_task_do(dp, NULL,
 159             dsl_pool_scrub_setup_sync, dp, &func, 0));
 160 }
 161
 162 /* ARGSUSED */
 163 static void
 164 dsl_pool_scrub_cancel_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
 165 {
 166         dsl_pool_t *dp = arg1;
 167         boolean_t *completep = arg2;
 168
 169         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 170                 return;
 171
 172         mutex_enter(&dp->dp_scrub_cancel_lock);
 173
 174         if (dp->dp_scrub_restart) {
 175                 dp->dp_scrub_restart = B_FALSE;
 176                 *completep = B_FALSE;
 177         }
 178
 179         /* XXX this is scrub-clean specific */
 180         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 181         while (dp->dp_spa->spa_scrub_inflight > 0) {
 182                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 183                     &dp->dp_spa->spa_scrub_lock);
 184         }
 185         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 186         dp->dp_spa->spa_scrub_started = B_FALSE;
 187         dp->dp_spa->spa_scrub_active = B_FALSE;
 188
 189         dp->dp_scrub_func = SCRUB_FUNC_NONE;
 190         VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 191             dp->dp_scrub_queue_obj, tx));
 192         dp->dp_scrub_queue_obj = 0;
 193         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 194
 195         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 196             DMU_POOL_SCRUB_QUEUE, tx));
 197         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 198             DMU_POOL_SCRUB_MIN_TXG, tx));
 199         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 200             DMU_POOL_SCRUB_MAX_TXG, tx));
 201         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 202             DMU_POOL_SCRUB_BOOKMARK, tx));
 203         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 204             DMU_POOL_SCRUB_FUNC, tx));
 205         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 206             DMU_POOL_SCRUB_ERRORS, tx));
 207
 208         spa_history_internal_log(LOG_POOL_SCRUB_DONE, dp->dp_spa, tx, cr,
 209             "complete=%u", *completep);
 210
 211         /* below is scrub-clean specific */
 212         vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev, POOL_SCRUB_NONE,
 213             *completep);
 214         /*
 215          * If the scrub/resilver completed, update all DTLs to reflect this.
 216          * Whether it succeeded or not, vacate all temporary scrub DTLs.
 217          */
 218         vdev_dtl_reassess(dp->dp_spa->spa_root_vdev, tx->tx_txg,
 219             *completep ? dp->dp_scrub_max_txg : 0, B_TRUE);
 220         if (*completep)
 221                 spa_event_notify(dp->dp_spa, NULL, dp->dp_scrub_min_txg ?
 222                     ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 223         spa_errlog_rotate(dp->dp_spa);
 224
 225         /*
 226          * We may have finished replacing a device.
 227          * Let the async thread assess this and handle the detach.
 228          */
 229         spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER_DONE);
 230
 231         dp->dp_scrub_min_txg = dp->dp_scrub_max_txg = 0;
 232         mutex_exit(&dp->dp_scrub_cancel_lock);
 233 }
 234
 235 int
 236 dsl_pool_scrub_cancel(dsl_pool_t *dp)
 237 {
 238         boolean_t complete = B_FALSE;
 239
 240         return (dsl_sync_task_do(dp, NULL,
 241             dsl_pool_scrub_cancel_sync, dp, &complete, 3));
 242 }
 243
 244 int
 245 dsl_free(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp,
 246     zio_done_func_t *done, void *private, uint32_t arc_flags)
 247 {
 248         /*
 249          * This function will be used by bp-rewrite wad to intercept frees.
 250          */
 251         return (arc_free(pio, dp->dp_spa, txg, (blkptr_t *)bpp,
 252             done, private, arc_flags));
 253 }
 254
 255 static boolean_t
 256 bookmark_is_zero(const zbookmark_t *zb)
 257 {
 258         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 259             zb->zb_level == 0 && zb->zb_blkid == 0);
 260 }
 261
 262 /* dnp is the dnode for zb1->zb_object */
 263 static boolean_t
 264 bookmark_is_before(dnode_phys_t *dnp, const zbookmark_t *zb1,
 265     const zbookmark_t *zb2)
 266 {
 267         uint64_t zb1nextL0, zb2thisobj;
 268
 269         ASSERT(zb1->zb_objset == zb2->zb_objset);
 270         ASSERT(zb1->zb_object != -1ULL);
 271         ASSERT(zb2->zb_level == 0);
 272
 273         /*
 274          * A bookmark in the deadlist is considered to be after
 275          * everything else.
 276          */
 277         if (zb2->zb_object == -1ULL)
 278                 return (B_TRUE);
 279
 280         /* The objset_phys_t isn't before anything. */
 281         if (dnp == NULL)
 282                 return (B_FALSE);
 283
 284         zb1nextL0 = (zb1->zb_blkid + 1) <<
 285             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 286
 287         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 288             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 289
 290         if (zb1->zb_object == 0) {
 291                 uint64_t nextobj = zb1nextL0 *
 292                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 293                 return (nextobj <= zb2thisobj);
 294         }
 295
 296         if (zb1->zb_object < zb2thisobj)
 297                 return (B_TRUE);
 298         if (zb1->zb_object > zb2thisobj)
 299                 return (B_FALSE);
 300         if (zb2->zb_object == 0)
 301                 return (B_FALSE);
 302         return (zb1nextL0 <= zb2->zb_blkid);
 303 }
 304
 305 static boolean_t
 306 scrub_pause(dsl_pool_t *dp, const zbookmark_t *zb)
 307 {
 308         int elapsed_ticks;
 309         int mintime;
 310
 311         if (dp->dp_scrub_pausing)
 312                 return (B_TRUE); /* we're already pausing */
 313
 314         if (!bookmark_is_zero(&dp->dp_scrub_bookmark))
 315                 return (B_FALSE); /* we're resuming */
 316
 317         /* We only know how to resume from level-0 blocks. */
 318         if (zb->zb_level != 0)
 319                 return (B_FALSE);
 320
 321         mintime = dp->dp_scrub_isresilver ? zfs_resilver_min_time :
 322             zfs_scrub_min_time;
 323         elapsed_ticks = lbolt64 - dp->dp_scrub_start_time;
 324         if (elapsed_ticks > hz * zfs_txg_timeout ||
 325             (elapsed_ticks > hz * mintime && txg_sync_waiting(dp))) {
 326                 dprintf("pausing at %llx/%llx/%llx/%llx\n",
 327                     (longlong_t)zb->zb_objset, (longlong_t)zb->zb_object,
 328                     (longlong_t)zb->zb_level, (longlong_t)zb->zb_blkid);
 329                 dp->dp_scrub_pausing = B_TRUE;
 330                 dp->dp_scrub_bookmark = *zb;
 331                 return (B_TRUE);
 332         }
 333         return (B_FALSE);
 334 }
 335
 336 typedef struct zil_traverse_arg {
 337         dsl_pool_t      *zta_dp;
 338         zil_header_t    *zta_zh;
 339 } zil_traverse_arg_t;
 340
 341 /* ARGSUSED */
 342 static void
 343 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 344 {
 345         zil_traverse_arg_t *zta = arg;
 346         dsl_pool_t *dp = zta->zta_dp;
 347         zil_header_t *zh = zta->zta_zh;
 348         zbookmark_t zb;
 349
 350         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 351                 return;
 352
 353         /*
 354          * One block ("stumpy") can be allocated a long time ago; we
 355          * want to visit that one because it has been allocated
 356          * (on-disk) even if it hasn't been claimed (even though for
 357          * plain scrub there's nothing to do to it).
 358          */
 359         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 360                 return;
 361
 362         zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 363         zb.zb_object = 0;
 364         zb.zb_level = -1;
 365         zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
 366         VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 367 }
 368
 369 /* ARGSUSED */
 370 static void
 371 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 372 {
 373         if (lrc->lrc_txtype == TX_WRITE) {
 374                 zil_traverse_arg_t *zta = arg;
 375                 dsl_pool_t *dp = zta->zta_dp;
 376                 zil_header_t *zh = zta->zta_zh;
 377                 lr_write_t *lr = (lr_write_t *)lrc;
 378                 blkptr_t *bp = &lr->lr_blkptr;
 379                 zbookmark_t zb;
 380
 381                 if (bp->blk_birth <= dp->dp_scrub_min_txg)
 382                         return;
 383
 384                 /*
 385                  * birth can be < claim_txg if this record's txg is
 386                  * already txg sync'ed (but this log block contains
 387                  * other records that are not synced)
 388                  */
 389                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 390                         return;
 391
 392                 zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 393                 zb.zb_object = lr->lr_foid;
 394                 zb.zb_level = BP_GET_LEVEL(bp);
 395                 zb.zb_blkid = lr->lr_offset / BP_GET_LSIZE(bp);
 396                 VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 397         }
 398 }
 399
 400 static void
 401 traverse_zil(dsl_pool_t *dp, zil_header_t *zh)
 402 {
 403         uint64_t claim_txg = zh->zh_claim_txg;
 404         zil_traverse_arg_t zta = { dp, zh };
 405         zilog_t *zilog;
 406
 407         /*
 408          * We only want to visit blocks that have been claimed but not yet
 409          * replayed (or, in read-only mode, blocks that *would* be claimed).
 410          */
 411         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 412                 return;
 413
 414         zilog = zil_alloc(dp->dp_meta_objset, zh);
 415
 416         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, &zta,
 417             claim_txg);
 418
 419         zil_free(zilog);
 420 }
 421
 422 static void
 423 scrub_visitbp(dsl_pool_t *dp, dnode_phys_t *dnp,
 424     arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb)
 425 {
 426         int err;
 427         arc_buf_t *buf = NULL;
 428
 429         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 430                 return;
 431
 432         if (scrub_pause(dp, zb))
 433                 return;
 434
 435         if (!bookmark_is_zero(&dp->dp_scrub_bookmark)) {
 436                 /*
 437                  * If we already visited this bp & everything below (in
 438                  * a prior txg), don't bother doing it again.
 439                  */
 440                 if (bookmark_is_before(dnp, zb, &dp->dp_scrub_bookmark))
 441                         return;
 442
 443                 /*
 444                  * If we found the block we're trying to resume from, or
 445                  * we went past it to a different object, zero it out to
 446                  * indicate that it's OK to start checking for pausing
 447                  * again.
 448                  */
 449                 if (bcmp(zb, &dp->dp_scrub_bookmark, sizeof (*zb)) == 0 ||
 450                     zb->zb_object > dp->dp_scrub_bookmark.zb_object) {
 451                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 452                             (longlong_t)zb->zb_objset,
 453                             (longlong_t)zb->zb_object,
 454                             (longlong_t)zb->zb_level,
 455                             (longlong_t)zb->zb_blkid);
 456                         bzero(&dp->dp_scrub_bookmark, sizeof (*zb));
 457                 }
 458         }
 459
 460         if (BP_GET_LEVEL(bp) > 0) {
 461                 uint32_t flags = ARC_WAIT;
 462                 int i;
 463                 blkptr_t *cbp;
 464                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 465
 466                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 467                     arc_getbuf_func, &buf,
 468                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 469                 if (err) {
 470                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 471                         dp->dp_spa->spa_scrub_errors++;
 472                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 473                         return;
 474                 }
 475                 cbp = buf->b_data;
 476
 477                 for (i = 0; i < epb; i++, cbp++) {
 478                         zbookmark_t czb;
 479
 480                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 481                             zb->zb_level - 1,
 482                             zb->zb_blkid * epb + i);
 483                         scrub_visitbp(dp, dnp, buf, cbp, &czb);
 484                 }
 485         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 486                 uint32_t flags = ARC_WAIT;
 487                 dnode_phys_t *child_dnp;
 488                 int i;
 489                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 490
 491                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 492                     arc_getbuf_func, &buf,
 493                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 494                 if (err) {
 495                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 496                         dp->dp_spa->spa_scrub_errors++;
 497                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 498                         return;
 499                 }
 500                 child_dnp = buf->b_data;
 501
 502                 for (i = 0; i < epb; i++, child_dnp++) {
 503                         scrub_visitdnode(dp, child_dnp, buf, zb->zb_objset,
 504                             zb->zb_blkid * epb + i);
 505                 }
 506         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 507                 uint32_t flags = ARC_WAIT;
 508                 objset_phys_t *osp;
 509
 510                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 511                     arc_getbuf_func, &buf,
 512                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 513                 if (err) {
 514                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 515                         dp->dp_spa->spa_scrub_errors++;
 516                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 517                         return;
 518                 }
 519
 520                 osp = buf->b_data;
 521
 522                 traverse_zil(dp, &osp->os_zil_header);
 523
 524                 scrub_visitdnode(dp, &osp->os_meta_dnode,
 525                     buf, zb->zb_objset, 0);
 526                 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
 527                         scrub_visitdnode(dp, &osp->os_userused_dnode,
 528                             buf, zb->zb_objset, 0);
 529                         scrub_visitdnode(dp, &osp->os_groupused_dnode,
 530                             buf, zb->zb_objset, 0);
 531                 }
 532         }
 533
 534         (void) scrub_funcs[dp->dp_scrub_func](dp, bp, zb);
 535         if (buf)
 536                 (void) arc_buf_remove_ref(buf, &buf);
 537 }
 538
 539 static void
 540 scrub_visitdnode(dsl_pool_t *dp, dnode_phys_t *dnp, arc_buf_t *buf,
 541     uint64_t objset, uint64_t object)
 542 {
 543         int j;
 544
 545         for (j = 0; j < dnp->dn_nblkptr; j++) {
 546                 zbookmark_t czb;
 547
 548                 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 549                 scrub_visitbp(dp, dnp, buf, &dnp->dn_blkptr[j], &czb);
 550         }
 551
 552 }
 553
 554 static void
 555 scrub_visit_rootbp(dsl_pool_t *dp, dsl_dataset_t *ds, blkptr_t *bp)
 556 {
 557         zbookmark_t zb;
 558
 559         SET_BOOKMARK(&zb, ds ? ds->ds_object : 0, 0, -1, 0);
 560         scrub_visitbp(dp, NULL, NULL, bp, &zb);
 561 }
 562
 563 void
 564 dsl_pool_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 565 {
 566         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 567
 568         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 569                 return;
 570
 571         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 572                 SET_BOOKMARK(&dp->dp_scrub_bookmark, -1, 0, 0, 0);
 573         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 574             ds->ds_object, tx) != 0) {
 575                 return;
 576         }
 577
 578         if (ds->ds_phys->ds_next_snap_obj != 0) {
 579                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 580                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 581         }
 582         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 583 }
 584
 585 void
 586 dsl_pool_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 587 {
 588         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 589
 590         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 591                 return;
 592
 593         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 594
 595         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 596                 dp->dp_scrub_bookmark.zb_objset =
 597                     ds->ds_phys->ds_prev_snap_obj;
 598         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 599             ds->ds_object, tx) == 0) {
 600                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 601                     ds->ds_phys->ds_prev_snap_obj, tx) == 0);
 602         }
 603 }
 604
 605 void
 606 dsl_pool_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 607 {
 608         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 609
 610         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 611                 return;
 612
 613         if (dp->dp_scrub_bookmark.zb_objset == ds1->ds_object) {
 614                 dp->dp_scrub_bookmark.zb_objset = ds2->ds_object;
 615         } else if (dp->dp_scrub_bookmark.zb_objset == ds2->ds_object) {
 616                 dp->dp_scrub_bookmark.zb_objset = ds1->ds_object;
 617         }
 618
 619         if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 620             ds1->ds_object, tx) == 0) {
 621                 int err = zap_add_int(dp->dp_meta_objset,
 622                     dp->dp_scrub_queue_obj, ds2->ds_object, tx);
 623                 VERIFY(err == 0 || err == EEXIST);
 624                 if (err == EEXIST) {
 625                         /* Both were there to begin with */
 626                         VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 627                             dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 628                 }
 629         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 630             ds2->ds_object, tx) == 0) {
 631                 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 632                     dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 633         }
 634 }
 635
 636 struct enqueue_clones_arg {
 637         dmu_tx_t *tx;
 638         uint64_t originobj;
 639 };
 640
 641 /* ARGSUSED */
 642 static int
 643 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 644 {
 645         struct enqueue_clones_arg *eca = arg;
 646         dsl_dataset_t *ds;
 647         int err;
 648         dsl_pool_t *dp;
 649
 650         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 651         if (err)
 652                 return (err);
 653         dp = ds->ds_dir->dd_pool;
 654
 655         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 656                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 657                         dsl_dataset_t *prev;
 658                         err = dsl_dataset_hold_obj(dp,
 659                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 660
 661                         dsl_dataset_rele(ds, FTAG);
 662                         if (err)
 663                                 return (err);
 664                         ds = prev;
 665                 }
 666                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 667                     ds->ds_object, eca->tx) == 0);
 668         }
 669         dsl_dataset_rele(ds, FTAG);
 670         return (0);
 671 }
 672
 673 static void
 674 scrub_visitds(dsl_pool_t *dp, uint64_t dsobj, dmu_tx_t *tx)
 675 {
 676         dsl_dataset_t *ds;
 677         uint64_t min_txg_save;
 678
 679         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 680
 681         /*
 682          * Iterate over the bps in this ds.
 683          */
 684         min_txg_save = dp->dp_scrub_min_txg;
 685         dp->dp_scrub_min_txg =
 686             MAX(dp->dp_scrub_min_txg, ds->ds_phys->ds_prev_snap_txg);
 687         scrub_visit_rootbp(dp, ds, &ds->ds_phys->ds_bp);
 688         dp->dp_scrub_min_txg = min_txg_save;
 689
 690         if (dp->dp_scrub_pausing)
 691                 goto out;
 692
 693         /*
 694          * Add descendent datasets to work queue.
 695          */
 696         if (ds->ds_phys->ds_next_snap_obj != 0) {
 697                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 698                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 699         }
 700         if (ds->ds_phys->ds_num_children > 1) {
 701                 boolean_t usenext = B_FALSE;
 702                 if (ds->ds_phys->ds_next_clones_obj != 0) {
 703                         uint64_t count;
 704                         /*
 705                          * A bug in a previous version of the code could
 706                          * cause upgrade_clones_cb() to not set
 707                          * ds_next_snap_obj when it should, leading to a
 708                          * missing entry.  Therefore we can only use the
 709                          * next_clones_obj when its count is correct.
 710                          */
 711                         int err = zap_count(dp->dp_meta_objset,
 712                             ds->ds_phys->ds_next_clones_obj, &count);
 713                         if (err == 0 &&
 714                             count == ds->ds_phys->ds_num_children - 1)
 715                                 usenext = B_TRUE;
 716                 }
 717
 718                 if (usenext) {
 719                         VERIFY(zap_join(dp->dp_meta_objset,
 720                             ds->ds_phys->ds_next_clones_obj,
 721                             dp->dp_scrub_queue_obj, tx) == 0);
 722                 } else {
 723                         struct enqueue_clones_arg eca;
 724                         eca.tx = tx;
 725                         eca.originobj = ds->ds_object;
 726
 727                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
 728                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
 729                 }
 730         }
 731
 732 out:
 733         dsl_dataset_rele(ds, FTAG);
 734 }
 735
 736 /* ARGSUSED */
 737 static int
 738 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 739 {
 740         dmu_tx_t *tx = arg;
 741         dsl_dataset_t *ds;
 742         int err;
 743         dsl_pool_t *dp;
 744
 745         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 746         if (err)
 747                 return (err);
 748
 749         dp = ds->ds_dir->dd_pool;
 750
 751         while (ds->ds_phys->ds_prev_snap_obj != 0) {
 752                 dsl_dataset_t *prev;
 753                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
 754                     FTAG, &prev);
 755                 if (err) {
 756                         dsl_dataset_rele(ds, FTAG);
 757                         return (err);
 758                 }
 759
 760                 /*
 761                  * If this is a clone, we don't need to worry about it for now.
 762                  */
 763                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
 764                         dsl_dataset_rele(ds, FTAG);
 765                         dsl_dataset_rele(prev, FTAG);
 766                         return (0);
 767                 }
 768                 dsl_dataset_rele(ds, FTAG);
 769                 ds = prev;
 770         }
 771
 772         VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 773             ds->ds_object, tx) == 0);
 774         dsl_dataset_rele(ds, FTAG);
 775         return (0);
 776 }
 777
 778 void
 779 dsl_pool_scrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
 780 {
 781         spa_t *spa = dp->dp_spa;
 782         zap_cursor_t zc;
 783         zap_attribute_t za;
 784         boolean_t complete = B_TRUE;
 785
 786         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 787                 return;
 788
 789         /*
 790          * If the pool is not loaded, or is trying to unload, leave it alone.
 791          */
 792         if (spa->spa_load_state != SPA_LOAD_NONE || spa_shutting_down(spa))
 793                 return;
 794
 795         if (dp->dp_scrub_restart) {
 796                 enum scrub_func func = dp->dp_scrub_func;
 797                 dp->dp_scrub_restart = B_FALSE;
 798                 dsl_pool_scrub_setup_sync(dp, &func, kcred, tx);
 799         }
 800
 801         if (spa->spa_root_vdev->vdev_stat.vs_scrub_type == 0) {
 802                 /*
 803                  * We must have resumed after rebooting; reset the vdev
 804                  * stats to know that we're doing a scrub (although it
 805                  * will think we're just starting now).
 806                  */
 807                 vdev_scrub_stat_update(spa->spa_root_vdev,
 808                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 809                     POOL_SCRUB_EVERYTHING, B_FALSE);
 810         }
 811
 812         dp->dp_scrub_pausing = B_FALSE;
 813         dp->dp_scrub_start_time = lbolt64;
 814         dp->dp_scrub_isresilver = (dp->dp_scrub_min_txg != 0);
 815         spa->spa_scrub_active = B_TRUE;
 816
 817         if (dp->dp_scrub_bookmark.zb_objset == 0) {
 818                 /* First do the MOS & ORIGIN */
 819                 scrub_visit_rootbp(dp, NULL, &dp->dp_meta_rootbp);
 820                 if (dp->dp_scrub_pausing)
 821                         goto out;
 822
 823                 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) {
 824                         VERIFY(0 == dmu_objset_find_spa(spa,
 825                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
 826                 } else {
 827                         scrub_visitds(dp, dp->dp_origin_snap->ds_object, tx);
 828                 }
 829                 ASSERT(!dp->dp_scrub_pausing);
 830         } else if (dp->dp_scrub_bookmark.zb_objset != -1ULL) {
 831                 /*
 832                  * If we were paused, continue from here.  Note if the
 833                  * ds we were paused on was deleted, the zb_objset will
 834                  * be -1, so we will skip this and find a new objset
 835                  * below.
 836                  */
 837                 scrub_visitds(dp, dp->dp_scrub_bookmark.zb_objset, tx);
 838                 if (dp->dp_scrub_pausing)
 839                         goto out;
 840         }
 841
 842         /*
 843          * In case we were paused right at the end of the ds, zero the
 844          * bookmark so we don't think that we're still trying to resume.
 845          */
 846         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 847
 848         /* keep pulling things out of the zap-object-as-queue */
 849         while (zap_cursor_init(&zc, dp->dp_meta_objset, dp->dp_scrub_queue_obj),
 850             zap_cursor_retrieve(&zc, &za) == 0) {
 851                 VERIFY(0 == zap_remove(dp->dp_meta_objset,
 852                     dp->dp_scrub_queue_obj, za.za_name, tx));
 853                 scrub_visitds(dp, za.za_first_integer, tx);
 854                 if (dp->dp_scrub_pausing)
 855                         break;
 856                 zap_cursor_fini(&zc);
 857         }
 858         zap_cursor_fini(&zc);
 859         if (dp->dp_scrub_pausing)
 860                 goto out;
 861
 862         /* done. */
 863
 864         dsl_pool_scrub_cancel_sync(dp, &complete, kcred, tx);
 865         return;
 866 out:
 867         VERIFY(0 == zap_update(dp->dp_meta_objset,
 868             DMU_POOL_DIRECTORY_OBJECT,
 869             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 870             &dp->dp_scrub_bookmark, tx));
 871         VERIFY(0 == zap_update(dp->dp_meta_objset,
 872             DMU_POOL_DIRECTORY_OBJECT,
 873             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 874             &spa->spa_scrub_errors, tx));
 875
 876         /* XXX this is scrub-clean specific */
 877         mutex_enter(&spa->spa_scrub_lock);
 878         while (spa->spa_scrub_inflight > 0)
 879                 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 880         mutex_exit(&spa->spa_scrub_lock);
 881 }
 882
 883 void
 884 dsl_pool_scrub_restart(dsl_pool_t *dp)
 885 {
 886         mutex_enter(&dp->dp_scrub_cancel_lock);
 887         dp->dp_scrub_restart = B_TRUE;
 888         mutex_exit(&dp->dp_scrub_cancel_lock);
 889 }
 890
 891 /*
 892  * scrub consumers
 893  */
 894
 895 static void
 896 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
 897 {
 898         int i;
 899
 900         /*
 901          * If we resume after a reboot, zab will be NULL; don't record
 902          * incomplete stats in that case.
 903          */
 904         if (zab == NULL)
 905                 return;
 906
 907         for (i = 0; i < 4; i++) {
 908                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
 909                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
 910                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
 911                 int equal;
 912
 913                 zb->zb_count++;
 914                 zb->zb_asize += BP_GET_ASIZE(bp);
 915                 zb->zb_lsize += BP_GET_LSIZE(bp);
 916                 zb->zb_psize += BP_GET_PSIZE(bp);
 917                 zb->zb_gangs += BP_COUNT_GANG(bp);
 918
 919                 switch (BP_GET_NDVAS(bp)) {
 920                 case 2:
 921                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 922                             DVA_GET_VDEV(&bp->blk_dva[1]))
 923                                 zb->zb_ditto_2_of_2_samevdev++;
 924                         break;
 925                 case 3:
 926                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 927                             DVA_GET_VDEV(&bp->blk_dva[1])) +
 928                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 929                             DVA_GET_VDEV(&bp->blk_dva[2])) +
 930                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
 931                             DVA_GET_VDEV(&bp->blk_dva[2]));
 932                         if (equal == 1)
 933                                 zb->zb_ditto_2_of_3_samevdev++;
 934                         else if (equal == 3)
 935                                 zb->zb_ditto_3_of_3_samevdev++;
 936                         break;
 937                 }
 938         }
 939 }
 940
 941 static void
 942 dsl_pool_scrub_clean_done(zio_t *zio)
 943 {
 944         spa_t *spa = zio->io_spa;
 945
 946         zio_data_buf_free(zio->io_data, zio->io_size);
 947
 948         mutex_enter(&spa->spa_scrub_lock);
 949         spa->spa_scrub_inflight--;
 950         cv_broadcast(&spa->spa_scrub_io_cv);
 951
 952         if (zio->io_error && (zio->io_error != ECKSUM ||
 953             !(zio->io_flags & ZIO_FLAG_SPECULATIVE)))
 954                 spa->spa_scrub_errors++;
 955         mutex_exit(&spa->spa_scrub_lock);
 956 }
 957
 958 static int
 959 dsl_pool_scrub_clean_cb(dsl_pool_t *dp,
 960     const blkptr_t *bp, const zbookmark_t *zb)
 961 {
 962         size_t size = BP_GET_PSIZE(bp);
 963         spa_t *spa = dp->dp_spa;
 964         boolean_t needs_io;
 965         int zio_flags = ZIO_FLAG_SCRUB_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
 966         int zio_priority;
 967
 968         ASSERT(bp->blk_birth > dp->dp_scrub_min_txg);
 969
 970         if (bp->blk_birth >= dp->dp_scrub_max_txg)
 971                 return (0);
 972
 973         count_block(dp->dp_blkstats, bp);
 974
 975         if (dp->dp_scrub_isresilver == 0) {
 976                 /* It's a scrub */
 977                 zio_flags |= ZIO_FLAG_SCRUB;
 978                 zio_priority = ZIO_PRIORITY_SCRUB;
 979                 needs_io = B_TRUE;
 980         } else {
 981                 /* It's a resilver */
 982                 zio_flags |= ZIO_FLAG_RESILVER;
 983                 zio_priority = ZIO_PRIORITY_RESILVER;
 984                 needs_io = B_FALSE;
 985         }
 986
 987         /* If it's an intent log block, failure is expected. */
 988         if (zb->zb_level == -1 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)
 989                 zio_flags |= ZIO_FLAG_SPECULATIVE;
 990
 991         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
 992                 vdev_t *vd = vdev_lookup_top(spa,
 993                     DVA_GET_VDEV(&bp->blk_dva[d]));
 994
 995                 /*
 996                  * Keep track of how much data we've examined so that
 997                  * zpool(1M) status can make useful progress reports.
 998                  */
 999                 mutex_enter(&vd->vdev_stat_lock);
1000                 vd->vdev_stat.vs_scrub_examined +=
1001                     DVA_GET_ASIZE(&bp->blk_dva[d]);
1002                 mutex_exit(&vd->vdev_stat_lock);
1003
1004                 /* if it's a resilver, this may not be in the target range */
1005                 if (!needs_io) {
1006                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1007                                 /*
1008                                  * Gang members may be spread across multiple
1009                                  * vdevs, so the best estimate we have is the
1010                                  * scrub range, which has already been checked.
1011                                  * XXX -- it would be better to change our
1012                                  * allocation policy to ensure that all
1013                                  * gang members reside on the same vdev.
1014                                  */
1015                                 needs_io = B_TRUE;
1016                         } else {
1017                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1018                                     bp->blk_birth, 1);
1019                         }
1020                 }
1021         }
1022
1023         if (needs_io && !zfs_no_scrub_io) {
1024                 void *data = zio_data_buf_alloc(size);
1025
1026                 mutex_enter(&spa->spa_scrub_lock);
1027                 while (spa->spa_scrub_inflight >= spa->spa_scrub_maxinflight)
1028                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1029                 spa->spa_scrub_inflight++;
1030                 mutex_exit(&spa->spa_scrub_lock);
1031
1032                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1033                     dsl_pool_scrub_clean_done, NULL, zio_priority,
1034                     zio_flags, zb));
1035         }
1036
1037         /* do not relocate this block */
1038         return (0);
1039 }
1040
1041 int
1042 dsl_pool_scrub_clean(dsl_pool_t *dp)
1043 {
1044         spa_t *spa = dp->dp_spa;
1045
1046         /*
1047          * Purge all vdev caches.  We do this here rather than in sync
1048          * context because this requires a writer lock on the spa_config
1049          * lock, which we can't do from sync context.  The
1050          * spa_scrub_reopen flag indicates that vdev_open() should not
1051          * attempt to start another scrub.
1052          */
1053         spa_vdev_state_enter(spa);
1054         spa->spa_scrub_reopen = B_TRUE;
1055         vdev_reopen(spa->spa_root_vdev);
1056         spa->spa_scrub_reopen = B_FALSE;
1057         (void) spa_vdev_state_exit(spa, NULL, 0);
1058
1059         return (dsl_pool_scrub_setup(dp, SCRUB_FUNC_CLEAN));
1060 }