granicus.if.org Git - zfs/blob - module/zfs/dnode_sync.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 /*
  23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
  25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  26  */
  27
  28 #include <sys/zfs_context.h>
  29 #include <sys/dbuf.h>
  30 #include <sys/dnode.h>
  31 #include <sys/dmu.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dmu_objset.h>
  34 #include <sys/dmu_recv.h>
  35 #include <sys/dsl_dataset.h>
  36 #include <sys/spa.h>
  37 #include <sys/range_tree.h>
  38 #include <sys/zfeature.h>
  39
  40 static void
  41 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
  42 {
  43         dmu_buf_impl_t *db;
  44         int txgoff = tx->tx_txg & TXG_MASK;
  45         int nblkptr = dn->dn_phys->dn_nblkptr;
  46         int old_toplvl = dn->dn_phys->dn_nlevels - 1;
  47         int new_level = dn->dn_next_nlevels[txgoff];
  48         int i;
  49
  50         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
  51
  52         /* this dnode can't be paged out because it's dirty */
  53         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
  54         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
  55         ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
  56
  57         db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
  58         ASSERT(db != NULL);
  59
  60         dn->dn_phys->dn_nlevels = new_level;
  61         dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
  62             dn->dn_object, dn->dn_phys->dn_nlevels);
  63
  64         /* transfer dnode's block pointers to new indirect block */
  65         (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
  66         ASSERT(db->db.db_data);
  67         ASSERT(arc_released(db->db_buf));
  68         ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
  69         bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
  70             sizeof (blkptr_t) * nblkptr);
  71         arc_buf_freeze(db->db_buf);
  72
  73         /* set dbuf's parent pointers to new indirect buf */
  74         for (i = 0; i < nblkptr; i++) {
  75                 dmu_buf_impl_t *child =
  76                     dbuf_find(dn->dn_objset, dn->dn_object, old_toplvl, i);
  77
  78                 if (child == NULL)
  79                         continue;
  80 #ifdef  DEBUG
  81                 DB_DNODE_ENTER(child);
  82                 ASSERT3P(DB_DNODE(child), ==, dn);
  83                 DB_DNODE_EXIT(child);
  84 #endif  /* DEBUG */
  85                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  86                         ASSERT(child->db_parent->db_level == db->db_level);
  87                         ASSERT(child->db_blkptr !=
  88                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  89                         mutex_exit(&child->db_mtx);
  90                         continue;
  91                 }
  92                 ASSERT(child->db_parent == NULL ||
  93                     child->db_parent == dn->dn_dbuf);
  94
  95                 child->db_parent = db;
  96                 dbuf_add_ref(db, child);
  97                 if (db->db.db_data)
  98                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
  99                 else
 100                         child->db_blkptr = NULL;
 101                 dprintf_dbuf_bp(child, child->db_blkptr,
 102                     "changed db_blkptr to new indirect %s", "");
 103
 104                 mutex_exit(&child->db_mtx);
 105         }
 106
 107         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 108
 109         dbuf_rele(db, FTAG);
 110
 111         rw_exit(&dn->dn_struct_rwlock);
 112 }
 113
 114 static void
 115 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 116 {
 117         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 118         uint64_t bytesfreed = 0;
 119
 120         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 121
 122         for (int i = 0; i < num; i++, bp++) {
 123                 if (BP_IS_HOLE(bp))
 124                         continue;
 125
 126                 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
 127                 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
 128
 129                 /*
 130                  * Save some useful information on the holes being
 131                  * punched, including logical size, type, and indirection
 132                  * level. Retaining birth time enables detection of when
 133                  * holes are punched for reducing the number of free
 134                  * records transmitted during a zfs send.
 135                  */
 136
 137                 uint64_t lsize = BP_GET_LSIZE(bp);
 138                 dmu_object_type_t type = BP_GET_TYPE(bp);
 139                 uint64_t lvl = BP_GET_LEVEL(bp);
 140
 141                 bzero(bp, sizeof (blkptr_t));
 142
 143                 if (spa_feature_is_active(dn->dn_objset->os_spa,
 144                     SPA_FEATURE_HOLE_BIRTH)) {
 145                         BP_SET_LSIZE(bp, lsize);
 146                         BP_SET_TYPE(bp, type);
 147                         BP_SET_LEVEL(bp, lvl);
 148                         BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
 149                 }
 150         }
 151         dnode_diduse_space(dn, -bytesfreed);
 152 }
 153
 154 #ifdef ZFS_DEBUG
 155 static void
 156 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 157 {
 158         int off, num;
 159         int i, err, epbs;
 160         uint64_t txg = tx->tx_txg;
 161         dnode_t *dn;
 162
 163         DB_DNODE_ENTER(db);
 164         dn = DB_DNODE(db);
 165         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 166         off = start - (db->db_blkid * 1<<epbs);
 167         num = end - start + 1;
 168
 169         ASSERT3U(off, >=, 0);
 170         ASSERT3U(num, >=, 0);
 171         ASSERT3U(db->db_level, >, 0);
 172         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 173         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 174         ASSERT(db->db_blkptr != NULL);
 175
 176         for (i = off; i < off+num; i++) {
 177                 uint64_t *buf;
 178                 dmu_buf_impl_t *child;
 179                 dbuf_dirty_record_t *dr;
 180                 int j;
 181
 182                 ASSERT(db->db_level == 1);
 183
 184                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 185                 err = dbuf_hold_impl(dn, db->db_level-1,
 186                     (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
 187                 rw_exit(&dn->dn_struct_rwlock);
 188                 if (err == ENOENT)
 189                         continue;
 190                 ASSERT(err == 0);
 191                 ASSERT(child->db_level == 0);
 192                 dr = child->db_last_dirty;
 193                 while (dr && dr->dr_txg > txg)
 194                         dr = dr->dr_next;
 195                 ASSERT(dr == NULL || dr->dr_txg == txg);
 196
 197                 /* data_old better be zeroed */
 198                 if (dr) {
 199                         buf = dr->dt.dl.dr_data->b_data;
 200                         for (j = 0; j < child->db.db_size >> 3; j++) {
 201                                 if (buf[j] != 0) {
 202                                         panic("freed data not zero: "
 203                                             "child=%p i=%d off=%d num=%d\n",
 204                                             (void *)child, i, off, num);
 205                                 }
 206                         }
 207                 }
 208
 209                 /*
 210                  * db_data better be zeroed unless it's dirty in a
 211                  * future txg.
 212                  */
 213                 mutex_enter(&child->db_mtx);
 214                 buf = child->db.db_data;
 215                 if (buf != NULL && child->db_state != DB_FILL &&
 216                     child->db_last_dirty == NULL) {
 217                         for (j = 0; j < child->db.db_size >> 3; j++) {
 218                                 if (buf[j] != 0) {
 219                                         panic("freed data not zero: "
 220                                             "child=%p i=%d off=%d num=%d\n",
 221                                             (void *)child, i, off, num);
 222                                 }
 223                         }
 224                 }
 225                 mutex_exit(&child->db_mtx);
 226
 227                 dbuf_rele(child, FTAG);
 228         }
 229         DB_DNODE_EXIT(db);
 230 }
 231 #endif
 232
 233 /*
 234  * We don't usually free the indirect blocks here.  If in one txg we have a
 235  * free_range and a write to the same indirect block, it's important that we
 236  * preserve the hole's birth times. Therefore, we don't free any any indirect
 237  * blocks in free_children().  If an indirect block happens to turn into all
 238  * holes, it will be freed by dbuf_write_children_ready, which happens at a
 239  * point in the syncing process where we know for certain the contents of the
 240  * indirect block.
 241  *
 242  * However, if we're freeing a dnode, its space accounting must go to zero
 243  * before we actually try to free the dnode, or we will trip an assertion. In
 244  * addition, we know the case described above cannot occur, because the dnode is
 245  * being freed.  Therefore, we free the indirect blocks immediately in that
 246  * case.
 247  */
 248 static void
 249 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 250     boolean_t free_indirects, dmu_tx_t *tx)
 251 {
 252         dnode_t *dn;
 253         blkptr_t *bp;
 254         dmu_buf_impl_t *subdb;
 255         uint64_t start, end, dbstart, dbend;
 256         unsigned int epbs, shift, i;
 257
 258         /*
 259          * There is a small possibility that this block will not be cached:
 260          *   1 - if level > 1 and there are no children with level <= 1
 261          *   2 - if this block was evicted since we read it from
 262          *       dmu_tx_hold_free().
 263          */
 264         if (db->db_state != DB_CACHED)
 265                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 266
 267         /*
 268          * If we modify this indirect block, and we are not freeing the
 269          * dnode (!free_indirects), then this indirect block needs to get
 270          * written to disk by dbuf_write().  If it is dirty, we know it will
 271          * be written (otherwise, we would have incorrect on-disk state
 272          * because the space would be freed but still referenced by the BP
 273          * in this indirect block).  Therefore we VERIFY that it is
 274          * dirty.
 275          *
 276          * Our VERIFY covers some cases that do not actually have to be
 277          * dirty, but the open-context code happens to dirty.  E.g. if the
 278          * blocks we are freeing are all holes, because in that case, we
 279          * are only freeing part of this indirect block, so it is an
 280          * ancestor of the first or last block to be freed.  The first and
 281          * last L1 indirect blocks are always dirtied by dnode_free_range().
 282          */
 283         VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
 284
 285         dbuf_release_bp(db);
 286         bp = db->db.db_data;
 287
 288         DB_DNODE_ENTER(db);
 289         dn = DB_DNODE(db);
 290         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 291         ASSERT3U(epbs, <, 31);
 292         shift = (db->db_level - 1) * epbs;
 293         dbstart = db->db_blkid << epbs;
 294         start = blkid >> shift;
 295         if (dbstart < start) {
 296                 bp += start - dbstart;
 297         } else {
 298                 start = dbstart;
 299         }
 300         dbend = ((db->db_blkid + 1) << epbs) - 1;
 301         end = (blkid + nblks - 1) >> shift;
 302         if (dbend <= end)
 303                 end = dbend;
 304
 305         ASSERT3U(start, <=, end);
 306
 307         if (db->db_level == 1) {
 308                 FREE_VERIFY(db, start, end, tx);
 309                 free_blocks(dn, bp, end-start+1, tx);
 310         } else {
 311                 for (uint64_t id = start; id <= end; id++, bp++) {
 312                         if (BP_IS_HOLE(bp))
 313                                 continue;
 314                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 315                         VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
 316                             id, TRUE, FALSE, FTAG, &subdb));
 317                         rw_exit(&dn->dn_struct_rwlock);
 318                         ASSERT3P(bp, ==, subdb->db_blkptr);
 319
 320                         free_children(subdb, blkid, nblks, free_indirects, tx);
 321                         dbuf_rele(subdb, FTAG);
 322                 }
 323         }
 324
 325         if (free_indirects) {
 326                 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
 327                         ASSERT(BP_IS_HOLE(bp));
 328                 bzero(db->db.db_data, db->db.db_size);
 329                 free_blocks(dn, db->db_blkptr, 1, tx);
 330         }
 331
 332         DB_DNODE_EXIT(db);
 333         arc_buf_freeze(db->db_buf);
 334 }
 335
 336 /*
 337  * Traverse the indicated range of the provided file
 338  * and "free" all the blocks contained there.
 339  */
 340 static void
 341 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
 342     boolean_t free_indirects, dmu_tx_t *tx)
 343 {
 344         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 345         int dnlevel = dn->dn_phys->dn_nlevels;
 346         boolean_t trunc = B_FALSE;
 347
 348         if (blkid > dn->dn_phys->dn_maxblkid)
 349                 return;
 350
 351         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 352         if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
 353                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 354                 trunc = B_TRUE;
 355         }
 356
 357         /* There are no indirect blocks in the object */
 358         if (dnlevel == 1) {
 359                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 360                         /* this range was never made persistent */
 361                         return;
 362                 }
 363                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 364                 free_blocks(dn, bp + blkid, nblks, tx);
 365         } else {
 366                 int shift = (dnlevel - 1) *
 367                     (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 368                 int start = blkid >> shift;
 369                 int end = (blkid + nblks - 1) >> shift;
 370                 dmu_buf_impl_t *db;
 371
 372                 ASSERT(start < dn->dn_phys->dn_nblkptr);
 373                 bp += start;
 374                 for (int i = start; i <= end; i++, bp++) {
 375                         if (BP_IS_HOLE(bp))
 376                                 continue;
 377                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 378                         VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
 379                             TRUE, FALSE, FTAG, &db));
 380                         rw_exit(&dn->dn_struct_rwlock);
 381
 382                         free_children(db, blkid, nblks, free_indirects, tx);
 383                         dbuf_rele(db, FTAG);
 384                 }
 385         }
 386
 387         if (trunc) {
 388                 ASSERTV(uint64_t off);
 389                 dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
 390
 391                 ASSERTV(off = (dn->dn_phys->dn_maxblkid + 1) *
 392                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT));
 393                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 394                     dn->dn_phys->dn_maxblkid == 0 ||
 395                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 396         }
 397 }
 398
 399 typedef struct dnode_sync_free_range_arg {
 400         dnode_t *dsfra_dnode;
 401         dmu_tx_t *dsfra_tx;
 402         boolean_t dsfra_free_indirects;
 403 } dnode_sync_free_range_arg_t;
 404
 405 static void
 406 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
 407 {
 408         dnode_sync_free_range_arg_t *dsfra = arg;
 409         dnode_t *dn = dsfra->dsfra_dnode;
 410
 411         mutex_exit(&dn->dn_mtx);
 412         dnode_sync_free_range_impl(dn, blkid, nblks,
 413             dsfra->dsfra_free_indirects, dsfra->dsfra_tx);
 414         mutex_enter(&dn->dn_mtx);
 415 }
 416
 417 /*
 418  * Try to kick all the dnode's dbufs out of the cache...
 419  */
 420 void
 421 dnode_evict_dbufs(dnode_t *dn)
 422 {
 423         dmu_buf_impl_t *db_marker;
 424         dmu_buf_impl_t *db, *db_next;
 425
 426         db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
 427
 428         mutex_enter(&dn->dn_dbufs_mtx);
 429         for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
 430
 431 #ifdef  DEBUG
 432                 DB_DNODE_ENTER(db);
 433                 ASSERT3P(DB_DNODE(db), ==, dn);
 434                 DB_DNODE_EXIT(db);
 435 #endif  /* DEBUG */
 436
 437                 mutex_enter(&db->db_mtx);
 438                 if (db->db_state != DB_EVICTING &&
 439                     zfs_refcount_is_zero(&db->db_holds)) {
 440                         db_marker->db_level = db->db_level;
 441                         db_marker->db_blkid = db->db_blkid;
 442                         db_marker->db_state = DB_SEARCH;
 443                         avl_insert_here(&dn->dn_dbufs, db_marker, db,
 444                             AVL_BEFORE);
 445
 446                         /*
 447                          * We need to use the "marker" dbuf rather than
 448                          * simply getting the next dbuf, because
 449                          * dbuf_destroy() may actually remove multiple dbufs.
 450                          * It can call itself recursively on the parent dbuf,
 451                          * which may also be removed from dn_dbufs.  The code
 452                          * flow would look like:
 453                          *
 454                          * dbuf_destroy():
 455                          *   dnode_rele_and_unlock(parent_dbuf, evicting=TRUE):
 456                          *      if (!cacheable || pending_evict)
 457                          *        dbuf_destroy()
 458                          */
 459                         dbuf_destroy(db);
 460
 461                         db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
 462                         avl_remove(&dn->dn_dbufs, db_marker);
 463                 } else {
 464                         db->db_pending_evict = TRUE;
 465                         mutex_exit(&db->db_mtx);
 466                         db_next = AVL_NEXT(&dn->dn_dbufs, db);
 467                 }
 468         }
 469         mutex_exit(&dn->dn_dbufs_mtx);
 470
 471         kmem_free(db_marker, sizeof (dmu_buf_impl_t));
 472
 473         dnode_evict_bonus(dn);
 474 }
 475
 476 void
 477 dnode_evict_bonus(dnode_t *dn)
 478 {
 479         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 480         if (dn->dn_bonus != NULL) {
 481                 if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
 482                         mutex_enter(&dn->dn_bonus->db_mtx);
 483                         dbuf_destroy(dn->dn_bonus);
 484                         dn->dn_bonus = NULL;
 485                 } else {
 486                         dn->dn_bonus->db_pending_evict = TRUE;
 487                 }
 488         }
 489         rw_exit(&dn->dn_struct_rwlock);
 490 }
 491
 492 static void
 493 dnode_undirty_dbufs(list_t *list)
 494 {
 495         dbuf_dirty_record_t *dr;
 496
 497         while ((dr = list_head(list))) {
 498                 dmu_buf_impl_t *db = dr->dr_dbuf;
 499                 uint64_t txg = dr->dr_txg;
 500
 501                 if (db->db_level != 0)
 502                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 503
 504                 mutex_enter(&db->db_mtx);
 505                 /* XXX - use dbuf_undirty()? */
 506                 list_remove(list, dr);
 507                 ASSERT(db->db_last_dirty == dr);
 508                 db->db_last_dirty = NULL;
 509                 db->db_dirtycnt -= 1;
 510                 if (db->db_level == 0) {
 511                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 512                             dr->dt.dl.dr_data == db->db_buf);
 513                         dbuf_unoverride(dr);
 514                 } else {
 515                         mutex_destroy(&dr->dt.di.dr_mtx);
 516                         list_destroy(&dr->dt.di.dr_children);
 517                 }
 518                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 519                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
 520         }
 521 }
 522
 523 static void
 524 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 525 {
 526         int txgoff = tx->tx_txg & TXG_MASK;
 527
 528         ASSERT(dmu_tx_is_syncing(tx));
 529
 530         /*
 531          * Our contents should have been freed in dnode_sync() by the
 532          * free range record inserted by the caller of dnode_free().
 533          */
 534         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 535         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 536
 537         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 538         dnode_evict_dbufs(dn);
 539
 540         /*
 541          * XXX - It would be nice to assert this, but we may still
 542          * have residual holds from async evictions from the arc...
 543          *
 544          * zfs_obj_to_path() also depends on this being
 545          * commented out.
 546          *
 547          * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
 548          */
 549
 550         /* Undirty next bits */
 551         dn->dn_next_nlevels[txgoff] = 0;
 552         dn->dn_next_indblkshift[txgoff] = 0;
 553         dn->dn_next_blksz[txgoff] = 0;
 554         dn->dn_next_maxblkid[txgoff] = 0;
 555
 556         /* ASSERT(blkptrs are zero); */
 557         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 558         ASSERT(dn->dn_type != DMU_OT_NONE);
 559
 560         ASSERT(dn->dn_free_txg > 0);
 561         if (dn->dn_allocated_txg != dn->dn_free_txg)
 562                 dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
 563         bzero(dn->dn_phys, sizeof (dnode_phys_t) * dn->dn_num_slots);
 564         dnode_free_interior_slots(dn);
 565
 566         mutex_enter(&dn->dn_mtx);
 567         dn->dn_type = DMU_OT_NONE;
 568         dn->dn_maxblkid = 0;
 569         dn->dn_allocated_txg = 0;
 570         dn->dn_free_txg = 0;
 571         dn->dn_have_spill = B_FALSE;
 572         dn->dn_num_slots = 1;
 573         mutex_exit(&dn->dn_mtx);
 574
 575         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 576
 577         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 578         /*
 579          * Now that we've released our hold, the dnode may
 580          * be evicted, so we mustn't access it.
 581          */
 582 }
 583
 584 /*
 585  * Write out the dnode's dirty buffers.
 586  */
 587 void
 588 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 589 {
 590         objset_t *os = dn->dn_objset;
 591         dnode_phys_t *dnp = dn->dn_phys;
 592         int txgoff = tx->tx_txg & TXG_MASK;
 593         list_t *list = &dn->dn_dirty_records[txgoff];
 594         ASSERTV(static const dnode_phys_t zerodn = { 0 });
 595         boolean_t kill_spill = B_FALSE;
 596
 597         ASSERT(dmu_tx_is_syncing(tx));
 598         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 599         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 600             bcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
 601         DNODE_VERIFY(dn);
 602
 603         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 604
 605         /*
 606          * Do user accounting if it is enabled and this is not
 607          * an encrypted receive.
 608          */
 609         if (dmu_objset_userused_enabled(os) &&
 610             !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
 611             (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
 612                 mutex_enter(&dn->dn_mtx);
 613                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 614                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 615                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 616                 if (dmu_objset_userobjused_enabled(dn->dn_objset))
 617                         dn->dn_phys->dn_flags |=
 618                             DNODE_FLAG_USEROBJUSED_ACCOUNTED;
 619                 mutex_exit(&dn->dn_mtx);
 620                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 621         } else {
 622                 /* Once we account for it, we should always account for it */
 623                 ASSERT(!(dn->dn_phys->dn_flags &
 624                     DNODE_FLAG_USERUSED_ACCOUNTED));
 625                 ASSERT(!(dn->dn_phys->dn_flags &
 626                     DNODE_FLAG_USEROBJUSED_ACCOUNTED));
 627         }
 628
 629         mutex_enter(&dn->dn_mtx);
 630         if (dn->dn_allocated_txg == tx->tx_txg) {
 631                 /* The dnode is newly allocated or reallocated */
 632                 if (dnp->dn_type == DMU_OT_NONE) {
 633                         /* this is a first alloc, not a realloc */
 634                         dnp->dn_nlevels = 1;
 635                         dnp->dn_nblkptr = dn->dn_nblkptr;
 636                 }
 637
 638                 dnp->dn_type = dn->dn_type;
 639                 dnp->dn_bonustype = dn->dn_bonustype;
 640                 dnp->dn_bonuslen = dn->dn_bonuslen;
 641         }
 642
 643         dnp->dn_extra_slots = dn->dn_num_slots - 1;
 644
 645         ASSERT(dnp->dn_nlevels > 1 ||
 646             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 647             BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
 648             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 649             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 650         ASSERT(dnp->dn_nlevels < 2 ||
 651             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 652             BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
 653
 654         if (dn->dn_next_type[txgoff] != 0) {
 655                 dnp->dn_type = dn->dn_type;
 656                 dn->dn_next_type[txgoff] = 0;
 657         }
 658
 659         if (dn->dn_next_blksz[txgoff] != 0) {
 660                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 661                     SPA_MINBLOCKSIZE) == 0);
 662                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 663                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 664                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 665                     dnp->dn_datablkszsec ||
 666                     !range_tree_is_empty(dn->dn_free_ranges[txgoff]));
 667                 dnp->dn_datablkszsec =
 668                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 669                 dn->dn_next_blksz[txgoff] = 0;
 670         }
 671
 672         if (dn->dn_next_bonuslen[txgoff] != 0) {
 673                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 674                         dnp->dn_bonuslen = 0;
 675                 else
 676                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 677                 ASSERT(dnp->dn_bonuslen <=
 678                     DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
 679                 dn->dn_next_bonuslen[txgoff] = 0;
 680         }
 681
 682         if (dn->dn_next_bonustype[txgoff] != 0) {
 683                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 684                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 685                 dn->dn_next_bonustype[txgoff] = 0;
 686         }
 687
 688         boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
 689             dn->dn_free_txg <= tx->tx_txg;
 690
 691         /*
 692          * Remove the spill block if we have been explicitly asked to
 693          * remove it, or if the object is being removed.
 694          */
 695         if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
 696                 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 697                         kill_spill = B_TRUE;
 698                 dn->dn_rm_spillblk[txgoff] = 0;
 699         }
 700
 701         if (dn->dn_next_indblkshift[txgoff] != 0) {
 702                 ASSERT(dnp->dn_nlevels == 1);
 703                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 704                 dn->dn_next_indblkshift[txgoff] = 0;
 705         }
 706
 707         /*
 708          * Just take the live (open-context) values for checksum and compress.
 709          * Strictly speaking it's a future leak, but nothing bad happens if we
 710          * start using the new checksum or compress algorithm a little early.
 711          */
 712         dnp->dn_checksum = dn->dn_checksum;
 713         dnp->dn_compress = dn->dn_compress;
 714
 715         mutex_exit(&dn->dn_mtx);
 716
 717         if (kill_spill) {
 718                 free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
 719                 mutex_enter(&dn->dn_mtx);
 720                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 721                 mutex_exit(&dn->dn_mtx);
 722         }
 723
 724         /* process all the "freed" ranges in the file */
 725         if (dn->dn_free_ranges[txgoff] != NULL) {
 726                 dnode_sync_free_range_arg_t dsfra;
 727                 dsfra.dsfra_dnode = dn;
 728                 dsfra.dsfra_tx = tx;
 729                 dsfra.dsfra_free_indirects = freeing_dnode;
 730                 if (freeing_dnode) {
 731                         ASSERT(range_tree_contains(dn->dn_free_ranges[txgoff],
 732                             0, dn->dn_maxblkid + 1));
 733                 }
 734                 mutex_enter(&dn->dn_mtx);
 735                 range_tree_vacate(dn->dn_free_ranges[txgoff],
 736                     dnode_sync_free_range, &dsfra);
 737                 range_tree_destroy(dn->dn_free_ranges[txgoff]);
 738                 dn->dn_free_ranges[txgoff] = NULL;
 739                 mutex_exit(&dn->dn_mtx);
 740         }
 741
 742         if (freeing_dnode) {
 743                 dn->dn_objset->os_freed_dnodes++;
 744                 dnode_sync_free(dn, tx);
 745                 return;
 746         }
 747
 748         if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
 749                 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 750                 mutex_enter(&ds->ds_lock);
 751                 ds->ds_feature_activation[SPA_FEATURE_LARGE_DNODE] =
 752                     (void *)B_TRUE;
 753                 mutex_exit(&ds->ds_lock);
 754         }
 755
 756         if (dn->dn_next_nlevels[txgoff]) {
 757                 dnode_increase_indirection(dn, tx);
 758                 dn->dn_next_nlevels[txgoff] = 0;
 759         }
 760
 761         /*
 762          * This must be done after dnode_sync_free_range()
 763          * and dnode_increase_indirection(). See dnode_new_blkid()
 764          * for an explanation of the high bit being set.
 765          */
 766         if (dn->dn_next_maxblkid[txgoff]) {
 767                 mutex_enter(&dn->dn_mtx);
 768                 dnp->dn_maxblkid =
 769                     dn->dn_next_maxblkid[txgoff] & ~DMU_NEXT_MAXBLKID_SET;
 770                 dn->dn_next_maxblkid[txgoff] = 0;
 771                 mutex_exit(&dn->dn_mtx);
 772         }
 773
 774         if (dn->dn_next_nblkptr[txgoff]) {
 775                 /* this should only happen on a realloc */
 776                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 777                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 778                         /* zero the new blkptrs we are gaining */
 779                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 780                             sizeof (blkptr_t) *
 781                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 782 #ifdef ZFS_DEBUG
 783                 } else {
 784                         int i;
 785                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 786                         /* the blkptrs we are losing better be unallocated */
 787                         for (i = 0; i < dnp->dn_nblkptr; i++) {
 788                                 if (i >= dn->dn_next_nblkptr[txgoff])
 789                                         ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 790                         }
 791 #endif
 792                 }
 793                 mutex_enter(&dn->dn_mtx);
 794                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 795                 dn->dn_next_nblkptr[txgoff] = 0;
 796                 mutex_exit(&dn->dn_mtx);
 797         }
 798
 799         dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
 800
 801         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 802                 ASSERT3P(list_head(list), ==, NULL);
 803                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 804         }
 805
 806         /*
 807          * Although we have dropped our reference to the dnode, it
 808          * can't be evicted until its written, and we haven't yet
 809          * initiated the IO for the dnode's dbuf.
 810          */
 811 }