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, 2015 by Delphix. All rights reserved.
  25  */
  26
  27 #include <sys/zfs_context.h>
  28 #include <sys/dbuf.h>
  29 #include <sys/dnode.h>
  30 #include <sys/dmu.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/dmu_objset.h>
  33 #include <sys/dsl_dataset.h>
  34 #include <sys/spa.h>
  35 #include <sys/range_tree.h>
  36 #include <sys/zfeature.h>
  37
  38 static void
  39 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
  40 {
  41         dmu_buf_impl_t *db;
  42         int txgoff = tx->tx_txg & TXG_MASK;
  43         int nblkptr = dn->dn_phys->dn_nblkptr;
  44         int old_toplvl = dn->dn_phys->dn_nlevels - 1;
  45         int new_level = dn->dn_next_nlevels[txgoff];
  46         int i;
  47
  48         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
  49
  50         /* this dnode can't be paged out because it's dirty */
  51         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
  52         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
  53         ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
  54
  55         db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
  56         ASSERT(db != NULL);
  57
  58         dn->dn_phys->dn_nlevels = new_level;
  59         dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
  60             dn->dn_object, dn->dn_phys->dn_nlevels);
  61
  62         /* check for existing blkptrs in the dnode */
  63         for (i = 0; i < nblkptr; i++)
  64                 if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
  65                         break;
  66         if (i != nblkptr) {
  67                 /* transfer dnode's block pointers to new indirect block */
  68                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
  69                 ASSERT(db->db.db_data);
  70                 ASSERT(arc_released(db->db_buf));
  71                 ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
  72                 bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
  73                     sizeof (blkptr_t) * nblkptr);
  74                 arc_buf_freeze(db->db_buf);
  75         }
  76
  77         /* set dbuf's parent pointers to new indirect buf */
  78         for (i = 0; i < nblkptr; i++) {
  79                 dmu_buf_impl_t *child =
  80                     dbuf_find(dn->dn_objset, dn->dn_object, old_toplvl, i);
  81
  82                 if (child == NULL)
  83                         continue;
  84 #ifdef  DEBUG
  85                 DB_DNODE_ENTER(child);
  86                 ASSERT3P(DB_DNODE(child), ==, dn);
  87                 DB_DNODE_EXIT(child);
  88 #endif  /* DEBUG */
  89                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  90                         ASSERT(child->db_parent->db_level == db->db_level);
  91                         ASSERT(child->db_blkptr !=
  92                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  93                         mutex_exit(&child->db_mtx);
  94                         continue;
  95                 }
  96                 ASSERT(child->db_parent == NULL ||
  97                     child->db_parent == dn->dn_dbuf);
  98
  99                 child->db_parent = db;
 100                 dbuf_add_ref(db, child);
 101                 if (db->db.db_data)
 102                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
 103                 else
 104                         child->db_blkptr = NULL;
 105                 dprintf_dbuf_bp(child, child->db_blkptr,
 106                     "changed db_blkptr to new indirect %s", "");
 107
 108                 mutex_exit(&child->db_mtx);
 109         }
 110
 111         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 112
 113         dbuf_rele(db, FTAG);
 114
 115         rw_exit(&dn->dn_struct_rwlock);
 116 }
 117
 118 static void
 119 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 120 {
 121         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 122         uint64_t bytesfreed = 0;
 123
 124         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 125
 126         for (int i = 0; i < num; i++, bp++) {
 127                 if (BP_IS_HOLE(bp))
 128                         continue;
 129
 130                 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
 131                 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
 132
 133                 /*
 134                  * Save some useful information on the holes being
 135                  * punched, including logical size, type, and indirection
 136                  * level. Retaining birth time enables detection of when
 137                  * holes are punched for reducing the number of free
 138                  * records transmitted during a zfs send.
 139                  */
 140
 141                 uint64_t lsize = BP_GET_LSIZE(bp);
 142                 dmu_object_type_t type = BP_GET_TYPE(bp);
 143                 uint64_t lvl = BP_GET_LEVEL(bp);
 144
 145                 bzero(bp, sizeof (blkptr_t));
 146
 147                 if (spa_feature_is_active(dn->dn_objset->os_spa,
 148                     SPA_FEATURE_HOLE_BIRTH)) {
 149                         BP_SET_LSIZE(bp, lsize);
 150                         BP_SET_TYPE(bp, type);
 151                         BP_SET_LEVEL(bp, lvl);
 152                         BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
 153                 }
 154         }
 155         dnode_diduse_space(dn, -bytesfreed);
 156 }
 157
 158 #ifdef ZFS_DEBUG
 159 static void
 160 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 161 {
 162         int off, num;
 163         int i, err, epbs;
 164         uint64_t txg = tx->tx_txg;
 165         dnode_t *dn;
 166
 167         DB_DNODE_ENTER(db);
 168         dn = DB_DNODE(db);
 169         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 170         off = start - (db->db_blkid * 1<<epbs);
 171         num = end - start + 1;
 172
 173         ASSERT3U(off, >=, 0);
 174         ASSERT3U(num, >=, 0);
 175         ASSERT3U(db->db_level, >, 0);
 176         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 177         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 178         ASSERT(db->db_blkptr != NULL);
 179
 180         for (i = off; i < off+num; i++) {
 181                 uint64_t *buf;
 182                 dmu_buf_impl_t *child;
 183                 dbuf_dirty_record_t *dr;
 184                 int j;
 185
 186                 ASSERT(db->db_level == 1);
 187
 188                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 189                 err = dbuf_hold_impl(dn, db->db_level-1,
 190                     (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
 191                 rw_exit(&dn->dn_struct_rwlock);
 192                 if (err == ENOENT)
 193                         continue;
 194                 ASSERT(err == 0);
 195                 ASSERT(child->db_level == 0);
 196                 dr = child->db_last_dirty;
 197                 while (dr && dr->dr_txg > txg)
 198                         dr = dr->dr_next;
 199                 ASSERT(dr == NULL || dr->dr_txg == txg);
 200
 201                 /* data_old better be zeroed */
 202                 if (dr) {
 203                         buf = dr->dt.dl.dr_data->b_data;
 204                         for (j = 0; j < child->db.db_size >> 3; j++) {
 205                                 if (buf[j] != 0) {
 206                                         panic("freed data not zero: "
 207                                             "child=%p i=%d off=%d num=%d\n",
 208                                             (void *)child, i, off, num);
 209                                 }
 210                         }
 211                 }
 212
 213                 /*
 214                  * db_data better be zeroed unless it's dirty in a
 215                  * future txg.
 216                  */
 217                 mutex_enter(&child->db_mtx);
 218                 buf = child->db.db_data;
 219                 if (buf != NULL && child->db_state != DB_FILL &&
 220                     child->db_last_dirty == NULL) {
 221                         for (j = 0; j < child->db.db_size >> 3; j++) {
 222                                 if (buf[j] != 0) {
 223                                         panic("freed data not zero: "
 224                                             "child=%p i=%d off=%d num=%d\n",
 225                                             (void *)child, i, off, num);
 226                                 }
 227                         }
 228                 }
 229                 mutex_exit(&child->db_mtx);
 230
 231                 dbuf_rele(child, FTAG);
 232         }
 233         DB_DNODE_EXIT(db);
 234 }
 235 #endif
 236
 237 static void
 238 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 239     dmu_tx_t *tx)
 240 {
 241         dnode_t *dn;
 242         blkptr_t *bp;
 243         dmu_buf_impl_t *subdb;
 244         uint64_t start, end, dbstart, dbend, i;
 245         int epbs, shift;
 246
 247         /*
 248          * There is a small possibility that this block will not be cached:
 249          *   1 - if level > 1 and there are no children with level <= 1
 250          *   2 - if this block was evicted since we read it from
 251          *       dmu_tx_hold_free().
 252          */
 253         if (db->db_state != DB_CACHED)
 254                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 255
 256         dbuf_release_bp(db);
 257         bp = db->db.db_data;
 258
 259         DB_DNODE_ENTER(db);
 260         dn = DB_DNODE(db);
 261         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 262         shift = (db->db_level - 1) * epbs;
 263         dbstart = db->db_blkid << epbs;
 264         start = blkid >> shift;
 265         if (dbstart < start) {
 266                 bp += start - dbstart;
 267         } else {
 268                 start = dbstart;
 269         }
 270         dbend = ((db->db_blkid + 1) << epbs) - 1;
 271         end = (blkid + nblks - 1) >> shift;
 272         if (dbend <= end)
 273                 end = dbend;
 274
 275         ASSERT3U(start, <=, end);
 276
 277         if (db->db_level == 1) {
 278                 FREE_VERIFY(db, start, end, tx);
 279                 free_blocks(dn, bp, end-start+1, tx);
 280         } else {
 281                 for (i = start; i <= end; i++, bp++) {
 282                         if (BP_IS_HOLE(bp))
 283                                 continue;
 284                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 285                         VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
 286                             i, B_TRUE, FTAG, &subdb));
 287                         rw_exit(&dn->dn_struct_rwlock);
 288                         ASSERT3P(bp, ==, subdb->db_blkptr);
 289
 290                         free_children(subdb, blkid, nblks, tx);
 291                         dbuf_rele(subdb, FTAG);
 292                 }
 293         }
 294
 295         /* If this whole block is free, free ourself too. */
 296         for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
 297                 if (!BP_IS_HOLE(bp))
 298                         break;
 299         }
 300         if (i == 1 << epbs) {
 301                 /* didn't find any non-holes */
 302                 bzero(db->db.db_data, db->db.db_size);
 303                 free_blocks(dn, db->db_blkptr, 1, tx);
 304         } else {
 305                 /*
 306                  * Partial block free; must be marked dirty so that it
 307                  * will be written out.
 308                  */
 309                 ASSERT(db->db_dirtycnt > 0);
 310         }
 311
 312         DB_DNODE_EXIT(db);
 313         arc_buf_freeze(db->db_buf);
 314 }
 315
 316 /*
 317  * Traverse the indicated range of the provided file
 318  * and "free" all the blocks contained there.
 319  */
 320 static void
 321 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
 322     dmu_tx_t *tx)
 323 {
 324         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 325         int dnlevel = dn->dn_phys->dn_nlevels;
 326         boolean_t trunc = B_FALSE;
 327
 328         if (blkid > dn->dn_phys->dn_maxblkid)
 329                 return;
 330
 331         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 332         if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
 333                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 334                 trunc = B_TRUE;
 335         }
 336
 337         /* There are no indirect blocks in the object */
 338         if (dnlevel == 1) {
 339                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 340                         /* this range was never made persistent */
 341                         return;
 342                 }
 343                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 344                 free_blocks(dn, bp + blkid, nblks, tx);
 345         } else {
 346                 int shift = (dnlevel - 1) *
 347                     (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 348                 int start = blkid >> shift;
 349                 int end = (blkid + nblks - 1) >> shift;
 350                 dmu_buf_impl_t *db;
 351
 352                 ASSERT(start < dn->dn_phys->dn_nblkptr);
 353                 bp += start;
 354                 for (int i = start; i <= end; i++, bp++) {
 355                         if (BP_IS_HOLE(bp))
 356                                 continue;
 357                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 358                         VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
 359                             TRUE, FTAG, &db));
 360                         rw_exit(&dn->dn_struct_rwlock);
 361
 362                         free_children(db, blkid, nblks, tx);
 363                         dbuf_rele(db, FTAG);
 364                 }
 365         }
 366
 367         if (trunc) {
 368                 dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
 369
 370                 uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
 371                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 372                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 373                     dn->dn_phys->dn_maxblkid == 0 ||
 374                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 375         }
 376 }
 377
 378 typedef struct dnode_sync_free_range_arg {
 379         dnode_t *dsfra_dnode;
 380         dmu_tx_t *dsfra_tx;
 381 } dnode_sync_free_range_arg_t;
 382
 383 static void
 384 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
 385 {
 386         dnode_sync_free_range_arg_t *dsfra = arg;
 387         dnode_t *dn = dsfra->dsfra_dnode;
 388
 389         mutex_exit(&dn->dn_mtx);
 390         dnode_sync_free_range_impl(dn, blkid, nblks, dsfra->dsfra_tx);
 391         mutex_enter(&dn->dn_mtx);
 392 }
 393
 394 /*
 395  * Try to kick all the dnode's dbufs out of the cache...
 396  */
 397 void
 398 dnode_evict_dbufs(dnode_t *dn)
 399 {
 400         int progress;
 401         int pass = 0;
 402
 403         do {
 404                 dmu_buf_impl_t *db, *db_next;
 405                 int evicting = FALSE;
 406
 407                 progress = FALSE;
 408                 mutex_enter(&dn->dn_dbufs_mtx);
 409                 for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
 410                         db_next = AVL_NEXT(&dn->dn_dbufs, db);
 411 #ifdef  DEBUG
 412                         DB_DNODE_ENTER(db);
 413                         ASSERT3P(DB_DNODE(db), ==, dn);
 414                         DB_DNODE_EXIT(db);
 415 #endif  /* DEBUG */
 416
 417                         mutex_enter(&db->db_mtx);
 418                         if (db->db_state == DB_EVICTING) {
 419                                 progress = TRUE;
 420                                 evicting = TRUE;
 421                                 mutex_exit(&db->db_mtx);
 422                         } else if (refcount_is_zero(&db->db_holds)) {
 423                                 progress = TRUE;
 424                                 dbuf_clear(db); /* exits db_mtx for us */
 425                         } else {
 426                                 mutex_exit(&db->db_mtx);
 427                         }
 428
 429                 }
 430                 /*
 431                  * NB: we need to drop dn_dbufs_mtx between passes so
 432                  * that any DB_EVICTING dbufs can make progress.
 433                  * Ideally, we would have some cv we could wait on, but
 434                  * since we don't, just wait a bit to give the other
 435                  * thread a chance to run.
 436                  */
 437                 mutex_exit(&dn->dn_dbufs_mtx);
 438                 if (evicting)
 439                         delay(1);
 440                 pass++;
 441                 ASSERT(pass < 100); /* sanity check */
 442         } while (progress);
 443
 444         dnode_evict_bonus(dn);
 445 }
 446
 447 void
 448 dnode_evict_bonus(dnode_t *dn)
 449 {
 450         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 451         if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
 452                 mutex_enter(&dn->dn_bonus->db_mtx);
 453                 dbuf_evict(dn->dn_bonus);
 454                 dn->dn_bonus = NULL;
 455         }
 456         rw_exit(&dn->dn_struct_rwlock);
 457 }
 458
 459 static void
 460 dnode_undirty_dbufs(list_t *list)
 461 {
 462         dbuf_dirty_record_t *dr;
 463
 464         while (dr = list_head(list)) {
 465                 dmu_buf_impl_t *db = dr->dr_dbuf;
 466                 uint64_t txg = dr->dr_txg;
 467
 468                 if (db->db_level != 0)
 469                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 470
 471                 mutex_enter(&db->db_mtx);
 472                 /* XXX - use dbuf_undirty()? */
 473                 list_remove(list, dr);
 474                 ASSERT(db->db_last_dirty == dr);
 475                 db->db_last_dirty = NULL;
 476                 db->db_dirtycnt -= 1;
 477                 if (db->db_level == 0) {
 478                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 479                             dr->dt.dl.dr_data == db->db_buf);
 480                         dbuf_unoverride(dr);
 481                 } else {
 482                         mutex_destroy(&dr->dt.di.dr_mtx);
 483                         list_destroy(&dr->dt.di.dr_children);
 484                 }
 485                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 486                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
 487         }
 488 }
 489
 490 static void
 491 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 492 {
 493         int txgoff = tx->tx_txg & TXG_MASK;
 494
 495         ASSERT(dmu_tx_is_syncing(tx));
 496
 497         /*
 498          * Our contents should have been freed in dnode_sync() by the
 499          * free range record inserted by the caller of dnode_free().
 500          */
 501         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 502         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 503
 504         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 505         dnode_evict_dbufs(dn);
 506         ASSERT(avl_is_empty(&dn->dn_dbufs));
 507         ASSERT3P(dn->dn_bonus, ==, NULL);
 508
 509         /*
 510          * XXX - It would be nice to assert this, but we may still
 511          * have residual holds from async evictions from the arc...
 512          *
 513          * zfs_obj_to_path() also depends on this being
 514          * commented out.
 515          *
 516          * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
 517          */
 518
 519         /* Undirty next bits */
 520         dn->dn_next_nlevels[txgoff] = 0;
 521         dn->dn_next_indblkshift[txgoff] = 0;
 522         dn->dn_next_blksz[txgoff] = 0;
 523
 524         /* ASSERT(blkptrs are zero); */
 525         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 526         ASSERT(dn->dn_type != DMU_OT_NONE);
 527
 528         ASSERT(dn->dn_free_txg > 0);
 529         if (dn->dn_allocated_txg != dn->dn_free_txg)
 530                 dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
 531         bzero(dn->dn_phys, sizeof (dnode_phys_t));
 532
 533         mutex_enter(&dn->dn_mtx);
 534         dn->dn_type = DMU_OT_NONE;
 535         dn->dn_maxblkid = 0;
 536         dn->dn_allocated_txg = 0;
 537         dn->dn_free_txg = 0;
 538         dn->dn_have_spill = B_FALSE;
 539         mutex_exit(&dn->dn_mtx);
 540
 541         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 542
 543         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 544         /*
 545          * Now that we've released our hold, the dnode may
 546          * be evicted, so we musn't access it.
 547          */
 548 }
 549
 550 /*
 551  * Write out the dnode's dirty buffers.
 552  */
 553 void
 554 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 555 {
 556         dnode_phys_t *dnp = dn->dn_phys;
 557         int txgoff = tx->tx_txg & TXG_MASK;
 558         list_t *list = &dn->dn_dirty_records[txgoff];
 559         static const dnode_phys_t zerodn = { 0 };
 560         boolean_t kill_spill = B_FALSE;
 561
 562         ASSERT(dmu_tx_is_syncing(tx));
 563         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 564         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 565             bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
 566         DNODE_VERIFY(dn);
 567
 568         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 569
 570         if (dmu_objset_userused_enabled(dn->dn_objset) &&
 571             !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 572                 mutex_enter(&dn->dn_mtx);
 573                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 574                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 575                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 576                 mutex_exit(&dn->dn_mtx);
 577                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 578         } else {
 579                 /* Once we account for it, we should always account for it. */
 580                 ASSERT(!(dn->dn_phys->dn_flags &
 581                     DNODE_FLAG_USERUSED_ACCOUNTED));
 582         }
 583
 584         mutex_enter(&dn->dn_mtx);
 585         if (dn->dn_allocated_txg == tx->tx_txg) {
 586                 /* The dnode is newly allocated or reallocated */
 587                 if (dnp->dn_type == DMU_OT_NONE) {
 588                         /* this is a first alloc, not a realloc */
 589                         dnp->dn_nlevels = 1;
 590                         dnp->dn_nblkptr = dn->dn_nblkptr;
 591                 }
 592
 593                 dnp->dn_type = dn->dn_type;
 594                 dnp->dn_bonustype = dn->dn_bonustype;
 595                 dnp->dn_bonuslen = dn->dn_bonuslen;
 596         }
 597         ASSERT(dnp->dn_nlevels > 1 ||
 598             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 599             BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
 600             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 601             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 602         ASSERT(dnp->dn_nlevels < 2 ||
 603             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 604             BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
 605
 606         if (dn->dn_next_type[txgoff] != 0) {
 607                 dnp->dn_type = dn->dn_type;
 608                 dn->dn_next_type[txgoff] = 0;
 609         }
 610
 611         if (dn->dn_next_blksz[txgoff] != 0) {
 612                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 613                     SPA_MINBLOCKSIZE) == 0);
 614                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 615                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 616                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 617                     dnp->dn_datablkszsec ||
 618                     range_tree_space(dn->dn_free_ranges[txgoff]) != 0);
 619                 dnp->dn_datablkszsec =
 620                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 621                 dn->dn_next_blksz[txgoff] = 0;
 622         }
 623
 624         if (dn->dn_next_bonuslen[txgoff] != 0) {
 625                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 626                         dnp->dn_bonuslen = 0;
 627                 else
 628                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 629                 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
 630                 dn->dn_next_bonuslen[txgoff] = 0;
 631         }
 632
 633         if (dn->dn_next_bonustype[txgoff] != 0) {
 634                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 635                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 636                 dn->dn_next_bonustype[txgoff] = 0;
 637         }
 638
 639         boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
 640             dn->dn_free_txg <= tx->tx_txg;
 641
 642         /*
 643          * Remove the spill block if we have been explicitly asked to
 644          * remove it, or if the object is being removed.
 645          */
 646         if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
 647                 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 648                         kill_spill = B_TRUE;
 649                 dn->dn_rm_spillblk[txgoff] = 0;
 650         }
 651
 652         if (dn->dn_next_indblkshift[txgoff] != 0) {
 653                 ASSERT(dnp->dn_nlevels == 1);
 654                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 655                 dn->dn_next_indblkshift[txgoff] = 0;
 656         }
 657
 658         /*
 659          * Just take the live (open-context) values for checksum and compress.
 660          * Strictly speaking it's a future leak, but nothing bad happens if we
 661          * start using the new checksum or compress algorithm a little early.
 662          */
 663         dnp->dn_checksum = dn->dn_checksum;
 664         dnp->dn_compress = dn->dn_compress;
 665
 666         mutex_exit(&dn->dn_mtx);
 667
 668         if (kill_spill) {
 669                 free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
 670                 mutex_enter(&dn->dn_mtx);
 671                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 672                 mutex_exit(&dn->dn_mtx);
 673         }
 674
 675         /* process all the "freed" ranges in the file */
 676         if (dn->dn_free_ranges[txgoff] != NULL) {
 677                 dnode_sync_free_range_arg_t dsfra;
 678                 dsfra.dsfra_dnode = dn;
 679                 dsfra.dsfra_tx = tx;
 680                 mutex_enter(&dn->dn_mtx);
 681                 range_tree_vacate(dn->dn_free_ranges[txgoff],
 682                     dnode_sync_free_range, &dsfra);
 683                 range_tree_destroy(dn->dn_free_ranges[txgoff]);
 684                 dn->dn_free_ranges[txgoff] = NULL;
 685                 mutex_exit(&dn->dn_mtx);
 686         }
 687
 688         if (freeing_dnode) {
 689                 dnode_sync_free(dn, tx);
 690                 return;
 691         }
 692
 693         if (dn->dn_next_nlevels[txgoff]) {
 694                 dnode_increase_indirection(dn, tx);
 695                 dn->dn_next_nlevels[txgoff] = 0;
 696         }
 697
 698         if (dn->dn_next_nblkptr[txgoff]) {
 699                 /* this should only happen on a realloc */
 700                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 701                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 702                         /* zero the new blkptrs we are gaining */
 703                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 704                             sizeof (blkptr_t) *
 705                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 706 #ifdef ZFS_DEBUG
 707                 } else {
 708                         int i;
 709                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 710                         /* the blkptrs we are losing better be unallocated */
 711                         for (i = dn->dn_next_nblkptr[txgoff];
 712                             i < dnp->dn_nblkptr; i++)
 713                                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 714 #endif
 715                 }
 716                 mutex_enter(&dn->dn_mtx);
 717                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 718                 dn->dn_next_nblkptr[txgoff] = 0;
 719                 mutex_exit(&dn->dn_mtx);
 720         }
 721
 722         dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
 723
 724         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 725                 ASSERT3P(list_head(list), ==, NULL);
 726                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 727         }
 728
 729         /*
 730          * Although we have dropped our reference to the dnode, it
 731          * can't be evicted until its written, and we haven't yet
 732          * initiated the IO for the dnode's dbuf.
 733          */
 734 }