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, 2014 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 = dbuf_find(dn, old_toplvl, i);
  80
  81                 if (child == NULL)
  82                         continue;
  83 #ifdef  DEBUG
  84                 DB_DNODE_ENTER(child);
  85                 ASSERT3P(DB_DNODE(child), ==, dn);
  86                 DB_DNODE_EXIT(child);
  87 #endif  /* DEBUG */
  88                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  89                         ASSERT(child->db_parent->db_level == db->db_level);
  90                         ASSERT(child->db_blkptr !=
  91                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  92                         mutex_exit(&child->db_mtx);
  93                         continue;
  94                 }
  95                 ASSERT(child->db_parent == NULL ||
  96                     child->db_parent == dn->dn_dbuf);
  97
  98                 child->db_parent = db;
  99                 dbuf_add_ref(db, child);
 100                 if (db->db.db_data)
 101                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
 102                 else
 103                         child->db_blkptr = NULL;
 104                 dprintf_dbuf_bp(child, child->db_blkptr,
 105                     "changed db_blkptr to new indirect %s", "");
 106
 107                 mutex_exit(&child->db_mtx);
 108         }
 109
 110         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 111
 112         dbuf_rele(db, FTAG);
 113
 114         rw_exit(&dn->dn_struct_rwlock);
 115 }
 116
 117 static void
 118 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 119 {
 120         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 121         uint64_t bytesfreed = 0;
 122         int i;
 123
 124         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 125
 126         for (i = 0; i < num; i++, bp++) {
 127                 uint64_t lsize, lvl;
 128                 dmu_object_type_t type;
 129
 130                 if (BP_IS_HOLE(bp))
 131                         continue;
 132
 133                 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
 134                 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
 135
 136                 /*
 137                  * Save some useful information on the holes being
 138                  * punched, including logical size, type, and indirection
 139                  * level. Retaining birth time enables detection of when
 140                  * holes are punched for reducing the number of free
 141                  * records transmitted during a zfs send.
 142                  */
 143
 144                 lsize = BP_GET_LSIZE(bp);
 145                 type = BP_GET_TYPE(bp);
 146                 lvl = BP_GET_LEVEL(bp);
 147
 148                 bzero(bp, sizeof (blkptr_t));
 149
 150                 if (spa_feature_is_active(dn->dn_objset->os_spa,
 151                     SPA_FEATURE_HOLE_BIRTH)) {
 152                         BP_SET_LSIZE(bp, lsize);
 153                         BP_SET_TYPE(bp, type);
 154                         BP_SET_LEVEL(bp, lvl);
 155                         BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
 156                 }
 157         }
 158         dnode_diduse_space(dn, -bytesfreed);
 159 }
 160
 161 #ifdef ZFS_DEBUG
 162 static void
 163 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 164 {
 165         int off, num;
 166         int i, err, epbs;
 167         uint64_t txg = tx->tx_txg;
 168         dnode_t *dn;
 169
 170         DB_DNODE_ENTER(db);
 171         dn = DB_DNODE(db);
 172         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 173         off = start - (db->db_blkid * 1<<epbs);
 174         num = end - start + 1;
 175
 176         ASSERT3U(off, >=, 0);
 177         ASSERT3U(num, >=, 0);
 178         ASSERT3U(db->db_level, >, 0);
 179         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 180         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 181         ASSERT(db->db_blkptr != NULL);
 182
 183         for (i = off; i < off+num; i++) {
 184                 uint64_t *buf;
 185                 dmu_buf_impl_t *child;
 186                 dbuf_dirty_record_t *dr;
 187                 int j;
 188
 189                 ASSERT(db->db_level == 1);
 190
 191                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 192                 err = dbuf_hold_impl(dn, db->db_level-1,
 193                     (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
 194                 rw_exit(&dn->dn_struct_rwlock);
 195                 if (err == ENOENT)
 196                         continue;
 197                 ASSERT(err == 0);
 198                 ASSERT(child->db_level == 0);
 199                 dr = child->db_last_dirty;
 200                 while (dr && dr->dr_txg > txg)
 201                         dr = dr->dr_next;
 202                 ASSERT(dr == NULL || dr->dr_txg == txg);
 203
 204                 /* data_old better be zeroed */
 205                 if (dr) {
 206                         buf = dr->dt.dl.dr_data->b_data;
 207                         for (j = 0; j < child->db.db_size >> 3; j++) {
 208                                 if (buf[j] != 0) {
 209                                         panic("freed data not zero: "
 210                                             "child=%p i=%d off=%d num=%d\n",
 211                                             (void *)child, i, off, num);
 212                                 }
 213                         }
 214                 }
 215
 216                 /*
 217                  * db_data better be zeroed unless it's dirty in a
 218                  * future txg.
 219                  */
 220                 mutex_enter(&child->db_mtx);
 221                 buf = child->db.db_data;
 222                 if (buf != NULL && child->db_state != DB_FILL &&
 223                     child->db_last_dirty == NULL) {
 224                         for (j = 0; j < child->db.db_size >> 3; j++) {
 225                                 if (buf[j] != 0) {
 226                                         panic("freed data not zero: "
 227                                             "child=%p i=%d off=%d num=%d\n",
 228                                             (void *)child, i, off, num);
 229                                 }
 230                         }
 231                 }
 232                 mutex_exit(&child->db_mtx);
 233
 234                 dbuf_rele(child, FTAG);
 235         }
 236         DB_DNODE_EXIT(db);
 237 }
 238 #endif
 239
 240 static void
 241 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 242     dmu_tx_t *tx)
 243 {
 244         dnode_t *dn;
 245         blkptr_t *bp;
 246         dmu_buf_impl_t *subdb;
 247         uint64_t start, end, dbstart, dbend, i;
 248         int epbs, shift;
 249
 250         /*
 251          * There is a small possibility that this block will not be cached:
 252          *   1 - if level > 1 and there are no children with level <= 1
 253          *   2 - if this block was evicted since we read it from
 254          *       dmu_tx_hold_free().
 255          */
 256         if (db->db_state != DB_CACHED)
 257                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 258
 259         dbuf_release_bp(db);
 260         bp = db->db.db_data;
 261
 262         DB_DNODE_ENTER(db);
 263         dn = DB_DNODE(db);
 264         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 265         shift = (db->db_level - 1) * epbs;
 266         dbstart = db->db_blkid << epbs;
 267         start = blkid >> shift;
 268         if (dbstart < start) {
 269                 bp += start - dbstart;
 270         } else {
 271                 start = dbstart;
 272         }
 273         dbend = ((db->db_blkid + 1) << epbs) - 1;
 274         end = (blkid + nblks - 1) >> shift;
 275         if (dbend <= end)
 276                 end = dbend;
 277
 278         ASSERT3U(start, <=, end);
 279
 280         if (db->db_level == 1) {
 281                 FREE_VERIFY(db, start, end, tx);
 282                 free_blocks(dn, bp, end-start+1, tx);
 283         } else {
 284                 for (i = start; i <= end; i++, bp++) {
 285                         if (BP_IS_HOLE(bp))
 286                                 continue;
 287                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 288                         VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
 289                             i, B_TRUE, FTAG, &subdb));
 290                         rw_exit(&dn->dn_struct_rwlock);
 291                         ASSERT3P(bp, ==, subdb->db_blkptr);
 292
 293                         free_children(subdb, blkid, nblks, tx);
 294                         dbuf_rele(subdb, FTAG);
 295                 }
 296         }
 297
 298         /* If this whole block is free, free ourself too. */
 299         for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
 300                 if (!BP_IS_HOLE(bp))
 301                         break;
 302         }
 303         if (i == 1 << epbs) {
 304                 /* didn't find any non-holes */
 305                 bzero(db->db.db_data, db->db.db_size);
 306                 free_blocks(dn, db->db_blkptr, 1, tx);
 307         } else {
 308                 /*
 309                  * Partial block free; must be marked dirty so that it
 310                  * will be written out.
 311                  */
 312                 ASSERT(db->db_dirtycnt > 0);
 313         }
 314
 315         DB_DNODE_EXIT(db);
 316         arc_buf_freeze(db->db_buf);
 317 }
 318
 319 /*
 320  * Traverse the indicated range of the provided file
 321  * and "free" all the blocks contained there.
 322  */
 323 static void
 324 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
 325     dmu_tx_t *tx)
 326 {
 327         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 328         int dnlevel = dn->dn_phys->dn_nlevels;
 329         boolean_t trunc = B_FALSE;
 330
 331         if (blkid > dn->dn_phys->dn_maxblkid)
 332                 return;
 333
 334         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 335         if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
 336                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 337                 trunc = B_TRUE;
 338         }
 339
 340         /* There are no indirect blocks in the object */
 341         if (dnlevel == 1) {
 342                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 343                         /* this range was never made persistent */
 344                         return;
 345                 }
 346                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 347                 free_blocks(dn, bp + blkid, nblks, tx);
 348         } else {
 349                 int shift = (dnlevel - 1) *
 350                     (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 351                 int start = blkid >> shift;
 352                 int end = (blkid + nblks - 1) >> shift;
 353                 dmu_buf_impl_t *db;
 354                 int i;
 355
 356                 ASSERT(start < dn->dn_phys->dn_nblkptr);
 357                 bp += start;
 358                 for (i = start; i <= end; i++, bp++) {
 359                         if (BP_IS_HOLE(bp))
 360                                 continue;
 361                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 362                         VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
 363                             TRUE, FTAG, &db));
 364                         rw_exit(&dn->dn_struct_rwlock);
 365
 366                         free_children(db, blkid, nblks, tx);
 367                         dbuf_rele(db, FTAG);
 368                 }
 369         }
 370
 371         if (trunc) {
 372                 ASSERTV(uint64_t off);
 373                 dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
 374
 375                 ASSERTV(off = (dn->dn_phys->dn_maxblkid + 1) *
 376                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT));
 377                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 378                     dn->dn_phys->dn_maxblkid == 0 ||
 379                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 380         }
 381 }
 382
 383 typedef struct dnode_sync_free_range_arg {
 384         dnode_t *dsfra_dnode;
 385         dmu_tx_t *dsfra_tx;
 386 } dnode_sync_free_range_arg_t;
 387
 388 static void
 389 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
 390 {
 391         dnode_sync_free_range_arg_t *dsfra = arg;
 392         dnode_t *dn = dsfra->dsfra_dnode;
 393
 394         mutex_exit(&dn->dn_mtx);
 395         dnode_sync_free_range_impl(dn, blkid, nblks, dsfra->dsfra_tx);
 396         mutex_enter(&dn->dn_mtx);
 397 }
 398
 399 /*
 400  * Try to kick all the dnode's dbufs out of the cache...
 401  */
 402 void
 403 dnode_evict_dbufs(dnode_t *dn)
 404 {
 405         int progress;
 406         int pass = 0;
 407
 408         do {
 409                 dmu_buf_impl_t *db, marker;
 410                 int evicting = FALSE;
 411
 412                 progress = FALSE;
 413                 mutex_enter(&dn->dn_dbufs_mtx);
 414                 list_insert_tail(&dn->dn_dbufs, &marker);
 415                 db = list_head(&dn->dn_dbufs);
 416                 for (; db != &marker; db = list_head(&dn->dn_dbufs)) {
 417                         list_remove(&dn->dn_dbufs, db);
 418                         list_insert_tail(&dn->dn_dbufs, db);
 419 #ifdef  DEBUG
 420                         DB_DNODE_ENTER(db);
 421                         ASSERT3P(DB_DNODE(db), ==, dn);
 422                         DB_DNODE_EXIT(db);
 423 #endif  /* DEBUG */
 424
 425                         mutex_enter(&db->db_mtx);
 426                         if (db->db_state == DB_EVICTING) {
 427                                 progress = TRUE;
 428                                 evicting = TRUE;
 429                                 mutex_exit(&db->db_mtx);
 430                         } else if (refcount_is_zero(&db->db_holds)) {
 431                                 progress = TRUE;
 432                                 dbuf_clear(db); /* exits db_mtx for us */
 433                         } else {
 434                                 mutex_exit(&db->db_mtx);
 435                         }
 436
 437                 }
 438                 list_remove(&dn->dn_dbufs, &marker);
 439                 /*
 440                  * NB: we need to drop dn_dbufs_mtx between passes so
 441                  * that any DB_EVICTING dbufs can make progress.
 442                  * Ideally, we would have some cv we could wait on, but
 443                  * since we don't, just wait a bit to give the other
 444                  * thread a chance to run.
 445                  */
 446                 mutex_exit(&dn->dn_dbufs_mtx);
 447                 if (evicting)
 448                         delay(1);
 449                 pass++;
 450                 if ((pass % 100) == 0)
 451                         dprintf("Exceeded %d passes evicting dbufs\n", pass);
 452         } while (progress);
 453
 454         if (pass >= 100)
 455                 dprintf("Required %d passes to evict dbufs\n", pass);
 456
 457         dnode_evict_bonus(dn);
 458 }
 459
 460 void
 461 dnode_evict_bonus(dnode_t *dn)
 462 {
 463         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 464         if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
 465                 mutex_enter(&dn->dn_bonus->db_mtx);
 466                 dbuf_evict(dn->dn_bonus);
 467                 dn->dn_bonus = NULL;
 468         }
 469         rw_exit(&dn->dn_struct_rwlock);
 470 }
 471
 472 static void
 473 dnode_undirty_dbufs(list_t *list)
 474 {
 475         dbuf_dirty_record_t *dr;
 476
 477         while ((dr = list_head(list))) {
 478                 dmu_buf_impl_t *db = dr->dr_dbuf;
 479                 uint64_t txg = dr->dr_txg;
 480
 481                 if (db->db_level != 0)
 482                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 483
 484                 mutex_enter(&db->db_mtx);
 485                 /* XXX - use dbuf_undirty()? */
 486                 list_remove(list, dr);
 487                 ASSERT(db->db_last_dirty == dr);
 488                 db->db_last_dirty = NULL;
 489                 db->db_dirtycnt -= 1;
 490                 if (db->db_level == 0) {
 491                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 492                             dr->dt.dl.dr_data == db->db_buf);
 493                         dbuf_unoverride(dr);
 494                 } else {
 495                         mutex_destroy(&dr->dt.di.dr_mtx);
 496                         list_destroy(&dr->dt.di.dr_children);
 497                 }
 498                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 499                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
 500         }
 501 }
 502
 503 static void
 504 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 505 {
 506         int txgoff = tx->tx_txg & TXG_MASK;
 507
 508         ASSERT(dmu_tx_is_syncing(tx));
 509
 510         /*
 511          * Our contents should have been freed in dnode_sync() by the
 512          * free range record inserted by the caller of dnode_free().
 513          */
 514         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 515         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 516
 517         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 518         dnode_evict_dbufs(dn);
 519         ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
 520         ASSERT3P(dn->dn_bonus, ==, NULL);
 521
 522         /*
 523          * XXX - It would be nice to assert this, but we may still
 524          * have residual holds from async evictions from the arc...
 525          *
 526          * zfs_obj_to_path() also depends on this being
 527          * commented out.
 528          *
 529          * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
 530          */
 531
 532         /* Undirty next bits */
 533         dn->dn_next_nlevels[txgoff] = 0;
 534         dn->dn_next_indblkshift[txgoff] = 0;
 535         dn->dn_next_blksz[txgoff] = 0;
 536
 537         /* ASSERT(blkptrs are zero); */
 538         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 539         ASSERT(dn->dn_type != DMU_OT_NONE);
 540
 541         ASSERT(dn->dn_free_txg > 0);
 542         if (dn->dn_allocated_txg != dn->dn_free_txg)
 543                 dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
 544         bzero(dn->dn_phys, sizeof (dnode_phys_t));
 545
 546         mutex_enter(&dn->dn_mtx);
 547         dn->dn_type = DMU_OT_NONE;
 548         dn->dn_maxblkid = 0;
 549         dn->dn_allocated_txg = 0;
 550         dn->dn_free_txg = 0;
 551         dn->dn_have_spill = B_FALSE;
 552         mutex_exit(&dn->dn_mtx);
 553
 554         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 555
 556         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 557         /*
 558          * Now that we've released our hold, the dnode may
 559          * be evicted, so we musn't access it.
 560          */
 561 }
 562
 563 /*
 564  * Write out the dnode's dirty buffers.
 565  */
 566 void
 567 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 568 {
 569         dnode_phys_t *dnp = dn->dn_phys;
 570         int txgoff = tx->tx_txg & TXG_MASK;
 571         list_t *list = &dn->dn_dirty_records[txgoff];
 572         boolean_t kill_spill = B_FALSE;
 573         boolean_t freeing_dnode;
 574         ASSERTV(static const dnode_phys_t zerodn = { 0 });
 575
 576         ASSERT(dmu_tx_is_syncing(tx));
 577         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 578         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 579             bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
 580         DNODE_VERIFY(dn);
 581
 582         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 583
 584         if (dmu_objset_userused_enabled(dn->dn_objset) &&
 585             !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 586                 mutex_enter(&dn->dn_mtx);
 587                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 588                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 589                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 590                 mutex_exit(&dn->dn_mtx);
 591                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 592         } else {
 593                 /* Once we account for it, we should always account for it. */
 594                 ASSERT(!(dn->dn_phys->dn_flags &
 595                     DNODE_FLAG_USERUSED_ACCOUNTED));
 596         }
 597
 598         mutex_enter(&dn->dn_mtx);
 599         if (dn->dn_allocated_txg == tx->tx_txg) {
 600                 /* The dnode is newly allocated or reallocated */
 601                 if (dnp->dn_type == DMU_OT_NONE) {
 602                         /* this is a first alloc, not a realloc */
 603                         dnp->dn_nlevels = 1;
 604                         dnp->dn_nblkptr = dn->dn_nblkptr;
 605                 }
 606
 607                 dnp->dn_type = dn->dn_type;
 608                 dnp->dn_bonustype = dn->dn_bonustype;
 609                 dnp->dn_bonuslen = dn->dn_bonuslen;
 610         }
 611         ASSERT(dnp->dn_nlevels > 1 ||
 612             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 613             BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
 614             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 615             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 616         ASSERT(dnp->dn_nlevels < 2 ||
 617             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 618             BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
 619
 620         if (dn->dn_next_type[txgoff] != 0) {
 621                 dnp->dn_type = dn->dn_type;
 622                 dn->dn_next_type[txgoff] = 0;
 623         }
 624
 625         if (dn->dn_next_blksz[txgoff] != 0) {
 626                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 627                     SPA_MINBLOCKSIZE) == 0);
 628                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 629                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 630                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 631                     dnp->dn_datablkszsec ||
 632                     range_tree_space(dn->dn_free_ranges[txgoff]) != 0);
 633                 dnp->dn_datablkszsec =
 634                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 635                 dn->dn_next_blksz[txgoff] = 0;
 636         }
 637
 638         if (dn->dn_next_bonuslen[txgoff] != 0) {
 639                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 640                         dnp->dn_bonuslen = 0;
 641                 else
 642                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 643                 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
 644                 dn->dn_next_bonuslen[txgoff] = 0;
 645         }
 646
 647         if (dn->dn_next_bonustype[txgoff] != 0) {
 648                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 649                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 650                 dn->dn_next_bonustype[txgoff] = 0;
 651         }
 652
 653         freeing_dnode = dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg;
 654
 655         /*
 656          * We will either remove a spill block when a file is being removed
 657          * or we have been asked to remove it.
 658          */
 659         if (dn->dn_rm_spillblk[txgoff] ||
 660             ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) && freeing_dnode)) {
 661                 if ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
 662                         kill_spill = B_TRUE;
 663                 dn->dn_rm_spillblk[txgoff] = 0;
 664         }
 665
 666         if (dn->dn_next_indblkshift[txgoff] != 0) {
 667                 ASSERT(dnp->dn_nlevels == 1);
 668                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 669                 dn->dn_next_indblkshift[txgoff] = 0;
 670         }
 671
 672         /*
 673          * Just take the live (open-context) values for checksum and compress.
 674          * Strictly speaking it's a future leak, but nothing bad happens if we
 675          * start using the new checksum or compress algorithm a little early.
 676          */
 677         dnp->dn_checksum = dn->dn_checksum;
 678         dnp->dn_compress = dn->dn_compress;
 679
 680         mutex_exit(&dn->dn_mtx);
 681
 682         if (kill_spill) {
 683                 free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
 684                 mutex_enter(&dn->dn_mtx);
 685                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 686                 mutex_exit(&dn->dn_mtx);
 687         }
 688
 689         /* process all the "freed" ranges in the file */
 690         if (dn->dn_free_ranges[txgoff] != NULL) {
 691                 dnode_sync_free_range_arg_t dsfra;
 692                 dsfra.dsfra_dnode = dn;
 693                 dsfra.dsfra_tx = tx;
 694                 mutex_enter(&dn->dn_mtx);
 695                 range_tree_vacate(dn->dn_free_ranges[txgoff],
 696                     dnode_sync_free_range, &dsfra);
 697                 range_tree_destroy(dn->dn_free_ranges[txgoff]);
 698                 dn->dn_free_ranges[txgoff] = NULL;
 699                 mutex_exit(&dn->dn_mtx);
 700         }
 701
 702         if (freeing_dnode) {
 703                 dnode_sync_free(dn, tx);
 704                 return;
 705         }
 706
 707         if (dn->dn_next_nlevels[txgoff]) {
 708                 dnode_increase_indirection(dn, tx);
 709                 dn->dn_next_nlevels[txgoff] = 0;
 710         }
 711
 712         if (dn->dn_next_nblkptr[txgoff]) {
 713                 /* this should only happen on a realloc */
 714                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 715                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 716                         /* zero the new blkptrs we are gaining */
 717                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 718                             sizeof (blkptr_t) *
 719                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 720 #ifdef ZFS_DEBUG
 721                 } else {
 722                         int i;
 723                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 724                         /* the blkptrs we are losing better be unallocated */
 725                         for (i = 0; i < dnp->dn_nblkptr; i++) {
 726                                 if (i >= dn->dn_next_nblkptr[txgoff])
 727                                         ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 728                         }
 729 #endif
 730                 }
 731                 mutex_enter(&dn->dn_mtx);
 732                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 733                 dn->dn_next_nblkptr[txgoff] = 0;
 734                 mutex_exit(&dn->dn_mtx);
 735         }
 736
 737         dbuf_sync_list(list, tx);
 738
 739         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 740                 ASSERT3P(list_head(list), ==, NULL);
 741                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 742         }
 743
 744         /*
 745          * Although we have dropped our reference to the dnode, it
 746          * can't be evicted until its written, and we haven't yet
 747          * initiated the IO for the dnode's dbuf.
 748          */
 749 }