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, 2015 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         /* check for existing blkptrs in the dnode */
  64         for (i = 0; i < nblkptr; i++)
  65                 if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
  66                         break;
  67         if (i != nblkptr) {
  68                 /* transfer dnode's block pointers to new indirect block */
  69                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
  70                 ASSERT(db->db.db_data);
  71                 ASSERT(arc_released(db->db_buf));
  72                 ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
  73                 bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
  74                     sizeof (blkptr_t) * nblkptr);
  75                 arc_buf_freeze(db->db_buf);
  76         }
  77
  78         /* set dbuf's parent pointers to new indirect buf */
  79         for (i = 0; i < nblkptr; i++) {
  80                 dmu_buf_impl_t *child =
  81                     dbuf_find(dn->dn_objset, dn->dn_object, old_toplvl, i);
  82
  83                 if (child == NULL)
  84                         continue;
  85 #ifdef  DEBUG
  86                 DB_DNODE_ENTER(child);
  87                 ASSERT3P(DB_DNODE(child), ==, dn);
  88                 DB_DNODE_EXIT(child);
  89 #endif  /* DEBUG */
  90                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  91                         ASSERT(child->db_parent->db_level == db->db_level);
  92                         ASSERT(child->db_blkptr !=
  93                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  94                         mutex_exit(&child->db_mtx);
  95                         continue;
  96                 }
  97                 ASSERT(child->db_parent == NULL ||
  98                     child->db_parent == dn->dn_dbuf);
  99
 100                 child->db_parent = db;
 101                 dbuf_add_ref(db, child);
 102                 if (db->db.db_data)
 103                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
 104                 else
 105                         child->db_blkptr = NULL;
 106                 dprintf_dbuf_bp(child, child->db_blkptr,
 107                     "changed db_blkptr to new indirect %s", "");
 108
 109                 mutex_exit(&child->db_mtx);
 110         }
 111
 112         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 113
 114         dbuf_rele(db, FTAG);
 115
 116         rw_exit(&dn->dn_struct_rwlock);
 117 }
 118
 119 static void
 120 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 121 {
 122         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 123         uint64_t bytesfreed = 0;
 124         int i;
 125
 126         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 127
 128         for (i = 0; i < num; i++, bp++) {
 129                 uint64_t lsize, lvl;
 130                 dmu_object_type_t type;
 131
 132                 if (BP_IS_HOLE(bp))
 133                         continue;
 134
 135                 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
 136                 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
 137
 138                 /*
 139                  * Save some useful information on the holes being
 140                  * punched, including logical size, type, and indirection
 141                  * level. Retaining birth time enables detection of when
 142                  * holes are punched for reducing the number of free
 143                  * records transmitted during a zfs send.
 144                  */
 145
 146                 lsize = BP_GET_LSIZE(bp);
 147                 type = BP_GET_TYPE(bp);
 148                 lvl = BP_GET_LEVEL(bp);
 149
 150                 bzero(bp, sizeof (blkptr_t));
 151
 152                 if (spa_feature_is_active(dn->dn_objset->os_spa,
 153                     SPA_FEATURE_HOLE_BIRTH)) {
 154                         BP_SET_LSIZE(bp, lsize);
 155                         BP_SET_TYPE(bp, type);
 156                         BP_SET_LEVEL(bp, lvl);
 157                         BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
 158                 }
 159         }
 160         dnode_diduse_space(dn, -bytesfreed);
 161 }
 162
 163 #ifdef ZFS_DEBUG
 164 static void
 165 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 166 {
 167         int off, num;
 168         int i, err, epbs;
 169         uint64_t txg = tx->tx_txg;
 170         dnode_t *dn;
 171
 172         DB_DNODE_ENTER(db);
 173         dn = DB_DNODE(db);
 174         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 175         off = start - (db->db_blkid * 1<<epbs);
 176         num = end - start + 1;
 177
 178         ASSERT3U(off, >=, 0);
 179         ASSERT3U(num, >=, 0);
 180         ASSERT3U(db->db_level, >, 0);
 181         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 182         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 183         ASSERT(db->db_blkptr != NULL);
 184
 185         for (i = off; i < off+num; i++) {
 186                 uint64_t *buf;
 187                 dmu_buf_impl_t *child;
 188                 dbuf_dirty_record_t *dr;
 189                 int j;
 190
 191                 ASSERT(db->db_level == 1);
 192
 193                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 194                 err = dbuf_hold_impl(dn, db->db_level-1,
 195                     (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
 196                 rw_exit(&dn->dn_struct_rwlock);
 197                 if (err == ENOENT)
 198                         continue;
 199                 ASSERT(err == 0);
 200                 ASSERT(child->db_level == 0);
 201                 dr = child->db_last_dirty;
 202                 while (dr && dr->dr_txg > txg)
 203                         dr = dr->dr_next;
 204                 ASSERT(dr == NULL || dr->dr_txg == txg);
 205
 206                 /* data_old better be zeroed */
 207                 if (dr) {
 208                         buf = dr->dt.dl.dr_data->b_data;
 209                         for (j = 0; j < child->db.db_size >> 3; j++) {
 210                                 if (buf[j] != 0) {
 211                                         panic("freed data not zero: "
 212                                             "child=%p i=%d off=%d num=%d\n",
 213                                             (void *)child, i, off, num);
 214                                 }
 215                         }
 216                 }
 217
 218                 /*
 219                  * db_data better be zeroed unless it's dirty in a
 220                  * future txg.
 221                  */
 222                 mutex_enter(&child->db_mtx);
 223                 buf = child->db.db_data;
 224                 if (buf != NULL && child->db_state != DB_FILL &&
 225                     child->db_last_dirty == NULL) {
 226                         for (j = 0; j < child->db.db_size >> 3; j++) {
 227                                 if (buf[j] != 0) {
 228                                         panic("freed data not zero: "
 229                                             "child=%p i=%d off=%d num=%d\n",
 230                                             (void *)child, i, off, num);
 231                                 }
 232                         }
 233                 }
 234                 mutex_exit(&child->db_mtx);
 235
 236                 dbuf_rele(child, FTAG);
 237         }
 238         DB_DNODE_EXIT(db);
 239 }
 240 #endif
 241
 242 static void
 243 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 244     dmu_tx_t *tx)
 245 {
 246         dnode_t *dn;
 247         blkptr_t *bp;
 248         dmu_buf_impl_t *subdb;
 249         uint64_t start, end, dbstart, dbend, i;
 250         int epbs, shift;
 251
 252         /*
 253          * There is a small possibility that this block will not be cached:
 254          *   1 - if level > 1 and there are no children with level <= 1
 255          *   2 - if this block was evicted since we read it from
 256          *       dmu_tx_hold_free().
 257          */
 258         if (db->db_state != DB_CACHED)
 259                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 260
 261         dbuf_release_bp(db);
 262         bp = db->db.db_data;
 263
 264         DB_DNODE_ENTER(db);
 265         dn = DB_DNODE(db);
 266         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 267         shift = (db->db_level - 1) * epbs;
 268         dbstart = db->db_blkid << epbs;
 269         start = blkid >> shift;
 270         if (dbstart < start) {
 271                 bp += start - dbstart;
 272         } else {
 273                 start = dbstart;
 274         }
 275         dbend = ((db->db_blkid + 1) << epbs) - 1;
 276         end = (blkid + nblks - 1) >> shift;
 277         if (dbend <= end)
 278                 end = dbend;
 279
 280         ASSERT3U(start, <=, end);
 281
 282         if (db->db_level == 1) {
 283                 FREE_VERIFY(db, start, end, tx);
 284                 free_blocks(dn, bp, end-start+1, tx);
 285         } else {
 286                 for (i = start; i <= end; i++, bp++) {
 287                         if (BP_IS_HOLE(bp))
 288                                 continue;
 289                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 290                         VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
 291                             i, TRUE, FALSE, FTAG, &subdb));
 292                         rw_exit(&dn->dn_struct_rwlock);
 293                         ASSERT3P(bp, ==, subdb->db_blkptr);
 294
 295                         free_children(subdb, blkid, nblks, tx);
 296                         dbuf_rele(subdb, FTAG);
 297                 }
 298         }
 299
 300         /* If this whole block is free, free ourself too. */
 301         for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
 302                 if (!BP_IS_HOLE(bp))
 303                         break;
 304         }
 305         if (i == 1 << epbs) {
 306                 /* didn't find any non-holes */
 307                 bzero(db->db.db_data, db->db.db_size);
 308                 free_blocks(dn, db->db_blkptr, 1, tx);
 309         } else {
 310                 /*
 311                  * Partial block free; must be marked dirty so that it
 312                  * will be written out.
 313                  */
 314                 ASSERT(db->db_dirtycnt > 0);
 315         }
 316
 317         DB_DNODE_EXIT(db);
 318         arc_buf_freeze(db->db_buf);
 319 }
 320
 321 /*
 322  * Traverse the indicated range of the provided file
 323  * and "free" all the blocks contained there.
 324  */
 325 static void
 326 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
 327     dmu_tx_t *tx)
 328 {
 329         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 330         int dnlevel = dn->dn_phys->dn_nlevels;
 331         boolean_t trunc = B_FALSE;
 332
 333         if (blkid > dn->dn_phys->dn_maxblkid)
 334                 return;
 335
 336         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 337         if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
 338                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 339                 trunc = B_TRUE;
 340         }
 341
 342         /* There are no indirect blocks in the object */
 343         if (dnlevel == 1) {
 344                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 345                         /* this range was never made persistent */
 346                         return;
 347                 }
 348                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 349                 free_blocks(dn, bp + blkid, nblks, tx);
 350         } else {
 351                 int shift = (dnlevel - 1) *
 352                     (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 353                 int start = blkid >> shift;
 354                 int end = (blkid + nblks - 1) >> shift;
 355                 dmu_buf_impl_t *db;
 356                 int i;
 357
 358                 ASSERT(start < dn->dn_phys->dn_nblkptr);
 359                 bp += start;
 360                 for (i = start; i <= end; i++, bp++) {
 361                         if (BP_IS_HOLE(bp))
 362                                 continue;
 363                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 364                         VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
 365                             TRUE, FALSE, FTAG, &db));
 366                         rw_exit(&dn->dn_struct_rwlock);
 367
 368                         free_children(db, blkid, nblks, tx);
 369                         dbuf_rele(db, FTAG);
 370                 }
 371         }
 372
 373         if (trunc) {
 374                 ASSERTV(uint64_t off);
 375                 dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
 376
 377                 ASSERTV(off = (dn->dn_phys->dn_maxblkid + 1) *
 378                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT));
 379                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 380                     dn->dn_phys->dn_maxblkid == 0 ||
 381                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 382         }
 383 }
 384
 385 typedef struct dnode_sync_free_range_arg {
 386         dnode_t *dsfra_dnode;
 387         dmu_tx_t *dsfra_tx;
 388 } dnode_sync_free_range_arg_t;
 389
 390 static void
 391 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
 392 {
 393         dnode_sync_free_range_arg_t *dsfra = arg;
 394         dnode_t *dn = dsfra->dsfra_dnode;
 395
 396         mutex_exit(&dn->dn_mtx);
 397         dnode_sync_free_range_impl(dn, blkid, nblks, dsfra->dsfra_tx);
 398         mutex_enter(&dn->dn_mtx);
 399 }
 400
 401 /*
 402  * Try to kick all the dnode's dbufs out of the cache...
 403  */
 404 void
 405 dnode_evict_dbufs(dnode_t *dn)
 406 {
 407         dmu_buf_impl_t *db_marker;
 408         dmu_buf_impl_t *db, *db_next;
 409
 410         db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
 411
 412         mutex_enter(&dn->dn_dbufs_mtx);
 413         for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
 414
 415 #ifdef  DEBUG
 416                 DB_DNODE_ENTER(db);
 417                 ASSERT3P(DB_DNODE(db), ==, dn);
 418                 DB_DNODE_EXIT(db);
 419 #endif  /* DEBUG */
 420
 421                 mutex_enter(&db->db_mtx);
 422                 if (db->db_state != DB_EVICTING &&
 423                     refcount_is_zero(&db->db_holds)) {
 424                         db_marker->db_level = db->db_level;
 425                         db_marker->db_blkid = db->db_blkid;
 426                         db_marker->db_state = DB_SEARCH;
 427                         avl_insert_here(&dn->dn_dbufs, db_marker, db,
 428                             AVL_BEFORE);
 429
 430                         dbuf_clear(db);
 431
 432                         db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
 433                         avl_remove(&dn->dn_dbufs, db_marker);
 434                 } else {
 435                         db->db_pending_evict = TRUE;
 436                         mutex_exit(&db->db_mtx);
 437                         db_next = AVL_NEXT(&dn->dn_dbufs, db);
 438                 }
 439         }
 440         mutex_exit(&dn->dn_dbufs_mtx);
 441
 442         kmem_free(db_marker, sizeof (dmu_buf_impl_t));
 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 != NULL) {
 452                 if (refcount_is_zero(&dn->dn_bonus->db_holds)) {
 453                         mutex_enter(&dn->dn_bonus->db_mtx);
 454                         dbuf_evict(dn->dn_bonus);
 455                         dn->dn_bonus = NULL;
 456                 } else {
 457                         dn->dn_bonus->db_pending_evict = TRUE;
 458                 }
 459         }
 460         rw_exit(&dn->dn_struct_rwlock);
 461 }
 462
 463 static void
 464 dnode_undirty_dbufs(list_t *list)
 465 {
 466         dbuf_dirty_record_t *dr;
 467
 468         while ((dr = list_head(list))) {
 469                 dmu_buf_impl_t *db = dr->dr_dbuf;
 470                 uint64_t txg = dr->dr_txg;
 471
 472                 if (db->db_level != 0)
 473                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 474
 475                 mutex_enter(&db->db_mtx);
 476                 /* XXX - use dbuf_undirty()? */
 477                 list_remove(list, dr);
 478                 ASSERT(db->db_last_dirty == dr);
 479                 db->db_last_dirty = NULL;
 480                 db->db_dirtycnt -= 1;
 481                 if (db->db_level == 0) {
 482                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 483                             dr->dt.dl.dr_data == db->db_buf);
 484                         dbuf_unoverride(dr);
 485                 } else {
 486                         mutex_destroy(&dr->dt.di.dr_mtx);
 487                         list_destroy(&dr->dt.di.dr_children);
 488                 }
 489                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 490                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
 491         }
 492 }
 493
 494 static void
 495 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 496 {
 497         int txgoff = tx->tx_txg & TXG_MASK;
 498
 499         ASSERT(dmu_tx_is_syncing(tx));
 500
 501         /*
 502          * Our contents should have been freed in dnode_sync() by the
 503          * free range record inserted by the caller of dnode_free().
 504          */
 505         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 506         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 507
 508         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 509         dnode_evict_dbufs(dn);
 510
 511         /*
 512          * XXX - It would be nice to assert this, but we may still
 513          * have residual holds from async evictions from the arc...
 514          *
 515          * zfs_obj_to_path() also depends on this being
 516          * commented out.
 517          *
 518          * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
 519          */
 520
 521         /* Undirty next bits */
 522         dn->dn_next_nlevels[txgoff] = 0;
 523         dn->dn_next_indblkshift[txgoff] = 0;
 524         dn->dn_next_blksz[txgoff] = 0;
 525
 526         /* ASSERT(blkptrs are zero); */
 527         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 528         ASSERT(dn->dn_type != DMU_OT_NONE);
 529
 530         ASSERT(dn->dn_free_txg > 0);
 531         if (dn->dn_allocated_txg != dn->dn_free_txg)
 532                 dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
 533         bzero(dn->dn_phys, sizeof (dnode_phys_t));
 534
 535         mutex_enter(&dn->dn_mtx);
 536         dn->dn_type = DMU_OT_NONE;
 537         dn->dn_maxblkid = 0;
 538         dn->dn_allocated_txg = 0;
 539         dn->dn_free_txg = 0;
 540         dn->dn_have_spill = B_FALSE;
 541         mutex_exit(&dn->dn_mtx);
 542
 543         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 544
 545         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 546         /*
 547          * Now that we've released our hold, the dnode may
 548          * be evicted, so we musn't access it.
 549          */
 550 }
 551
 552 /*
 553  * Write out the dnode's dirty buffers.
 554  */
 555 void
 556 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 557 {
 558         dnode_phys_t *dnp = dn->dn_phys;
 559         int txgoff = tx->tx_txg & TXG_MASK;
 560         list_t *list = &dn->dn_dirty_records[txgoff];
 561         boolean_t kill_spill = B_FALSE;
 562         boolean_t freeing_dnode;
 563         ASSERTV(static const dnode_phys_t zerodn = { 0 });
 564
 565         ASSERT(dmu_tx_is_syncing(tx));
 566         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 567         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 568             bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
 569         DNODE_VERIFY(dn);
 570
 571         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 572
 573         if (dmu_objset_userused_enabled(dn->dn_objset) &&
 574             !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 575                 mutex_enter(&dn->dn_mtx);
 576                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 577                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 578                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 579                 mutex_exit(&dn->dn_mtx);
 580                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 581         } else {
 582                 /* Once we account for it, we should always account for it. */
 583                 ASSERT(!(dn->dn_phys->dn_flags &
 584                     DNODE_FLAG_USERUSED_ACCOUNTED));
 585         }
 586
 587         mutex_enter(&dn->dn_mtx);
 588         if (dn->dn_allocated_txg == tx->tx_txg) {
 589                 /* The dnode is newly allocated or reallocated */
 590                 if (dnp->dn_type == DMU_OT_NONE) {
 591                         /* this is a first alloc, not a realloc */
 592                         dnp->dn_nlevels = 1;
 593                         dnp->dn_nblkptr = dn->dn_nblkptr;
 594                 }
 595
 596                 dnp->dn_type = dn->dn_type;
 597                 dnp->dn_bonustype = dn->dn_bonustype;
 598                 dnp->dn_bonuslen = dn->dn_bonuslen;
 599         }
 600         ASSERT(dnp->dn_nlevels > 1 ||
 601             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 602             BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
 603             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 604             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 605         ASSERT(dnp->dn_nlevels < 2 ||
 606             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 607             BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
 608
 609         if (dn->dn_next_type[txgoff] != 0) {
 610                 dnp->dn_type = dn->dn_type;
 611                 dn->dn_next_type[txgoff] = 0;
 612         }
 613
 614         if (dn->dn_next_blksz[txgoff] != 0) {
 615                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 616                     SPA_MINBLOCKSIZE) == 0);
 617                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 618                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 619                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 620                     dnp->dn_datablkszsec ||
 621                     range_tree_space(dn->dn_free_ranges[txgoff]) != 0);
 622                 dnp->dn_datablkszsec =
 623                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 624                 dn->dn_next_blksz[txgoff] = 0;
 625         }
 626
 627         if (dn->dn_next_bonuslen[txgoff] != 0) {
 628                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 629                         dnp->dn_bonuslen = 0;
 630                 else
 631                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 632                 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
 633                 dn->dn_next_bonuslen[txgoff] = 0;
 634         }
 635
 636         if (dn->dn_next_bonustype[txgoff] != 0) {
 637                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 638                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 639                 dn->dn_next_bonustype[txgoff] = 0;
 640         }
 641
 642         freeing_dnode = dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg;
 643
 644         /*
 645          * Remove the spill block if we have been explicitly asked to
 646          * remove it, or if the object is being removed.
 647          */
 648         if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
 649                 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 650                         kill_spill = B_TRUE;
 651                 dn->dn_rm_spillblk[txgoff] = 0;
 652         }
 653
 654         if (dn->dn_next_indblkshift[txgoff] != 0) {
 655                 ASSERT(dnp->dn_nlevels == 1);
 656                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 657                 dn->dn_next_indblkshift[txgoff] = 0;
 658         }
 659
 660         /*
 661          * Just take the live (open-context) values for checksum and compress.
 662          * Strictly speaking it's a future leak, but nothing bad happens if we
 663          * start using the new checksum or compress algorithm a little early.
 664          */
 665         dnp->dn_checksum = dn->dn_checksum;
 666         dnp->dn_compress = dn->dn_compress;
 667
 668         mutex_exit(&dn->dn_mtx);
 669
 670         if (kill_spill) {
 671                 free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
 672                 mutex_enter(&dn->dn_mtx);
 673                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 674                 mutex_exit(&dn->dn_mtx);
 675         }
 676
 677         /* process all the "freed" ranges in the file */
 678         if (dn->dn_free_ranges[txgoff] != NULL) {
 679                 dnode_sync_free_range_arg_t dsfra;
 680                 dsfra.dsfra_dnode = dn;
 681                 dsfra.dsfra_tx = tx;
 682                 mutex_enter(&dn->dn_mtx);
 683                 range_tree_vacate(dn->dn_free_ranges[txgoff],
 684                     dnode_sync_free_range, &dsfra);
 685                 range_tree_destroy(dn->dn_free_ranges[txgoff]);
 686                 dn->dn_free_ranges[txgoff] = NULL;
 687                 mutex_exit(&dn->dn_mtx);
 688         }
 689
 690         if (freeing_dnode) {
 691                 dnode_sync_free(dn, tx);
 692                 return;
 693         }
 694
 695         if (dn->dn_next_nlevels[txgoff]) {
 696                 dnode_increase_indirection(dn, tx);
 697                 dn->dn_next_nlevels[txgoff] = 0;
 698         }
 699
 700         if (dn->dn_next_nblkptr[txgoff]) {
 701                 /* this should only happen on a realloc */
 702                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 703                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 704                         /* zero the new blkptrs we are gaining */
 705                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 706                             sizeof (blkptr_t) *
 707                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 708 #ifdef ZFS_DEBUG
 709                 } else {
 710                         int i;
 711                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 712                         /* the blkptrs we are losing better be unallocated */
 713                         for (i = 0; i < dnp->dn_nblkptr; i++) {
 714                                 if (i >= dn->dn_next_nblkptr[txgoff])
 715                                         ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 716                         }
 717 #endif
 718                 }
 719                 mutex_enter(&dn->dn_mtx);
 720                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 721                 dn->dn_next_nblkptr[txgoff] = 0;
 722                 mutex_exit(&dn->dn_mtx);
 723         }
 724
 725         dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
 726
 727         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 728                 ASSERT3P(list_head(list), ==, NULL);
 729                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 730         }
 731
 732         /*
 733          * Although we have dropped our reference to the dnode, it
 734          * can't be evicted until its written, and we haven't yet
 735          * initiated the IO for the dnode's dbuf.
 736          */
 737 }