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