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) 2014 Nexenta Systems, Inc. All rights reserved.
  25  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
  26  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  27  */
  28 
  29 #include <sys/zfs_context.h>
  30 #include <sys/dbuf.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/dsl_dataset.h>
  36 #include <sys/spa.h>
  37 #include <sys/range_tree.h>
  38 #include <sys/zfeature.h>
  39 
  40 static void
  41 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
  42 {
  43         dmu_buf_impl_t *db;
  44         int txgoff = tx->tx_txg & TXG_MASK;
  45         int nblkptr = dn->dn_phys->dn_nblkptr;
  46         int old_toplvl = dn->dn_phys->dn_nlevels - 1;
  47         int new_level = dn->dn_next_nlevels[txgoff];
  48         int i;
  49 
  50         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
  51 
  52         /* this dnode can't be paged out because it's dirty */
  53         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
  54         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
  55         ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
  56 
  57         db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
  58         ASSERT(db != NULL);
  59 
  60         dn->dn_phys->dn_nlevels = new_level;
  61         dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
  62             dn->dn_object, dn->dn_phys->dn_nlevels);
  63 
  64         /* transfer dnode's block pointers to new indirect block */
  65         (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
  66         ASSERT(db->db.db_data);
  67         ASSERT(arc_released(db->db_buf));
  68         ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
  69         bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
  70             sizeof (blkptr_t) * nblkptr);
  71         arc_buf_freeze(db->db_buf);
  72 
  73         /* set dbuf's parent pointers to new indirect buf */
  74         for (i = 0; i < nblkptr; i++) {
  75                 dmu_buf_impl_t *child =
  76                     dbuf_find(dn->dn_objset, dn->dn_object, old_toplvl, i);
  77 
  78                 if (child == NULL)
  79                         continue;
  80 #ifdef  DEBUG
  81                 DB_DNODE_ENTER(child);
  82                 ASSERT3P(DB_DNODE(child), ==, dn);
  83                 DB_DNODE_EXIT(child);
  84 #endif  /* DEBUG */
  85                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  86                         ASSERT(child->db_parent->db_level == db->db_level);
  87                         ASSERT(child->db_blkptr !=
  88                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  89                         mutex_exit(&child->db_mtx);
  90                         continue;
  91                 }
  92                 ASSERT(child->db_parent == NULL ||
  93                     child->db_parent == dn->dn_dbuf);
  94 
  95                 child->db_parent = db;
  96                 dbuf_add_ref(db, child);
  97                 if (db->db.db_data)
  98                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
  99                 else
 100                         child->db_blkptr = NULL;
 101                 dprintf_dbuf_bp(child, child->db_blkptr,
 102                     "changed db_blkptr to new indirect %s", "");
 103 
 104                 mutex_exit(&child->db_mtx);
 105         }
 106 
 107         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 108 
 109         dbuf_rele(db, FTAG);
 110 
 111         rw_exit(&dn->dn_struct_rwlock);
 112 }
 113 
 114 static void
 115 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 116 {
 117         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 118         uint64_t bytesfreed = 0;
 119 
 120         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 121 
 122         for (int i = 0; i < num; i++, bp++) {
 123                 if (BP_IS_HOLE(bp))
 124                         continue;
 125 
 126                 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
 127                 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
 128 
 129                 /*
 130                  * Save some useful information on the holes being
 131                  * punched, including logical size, type, and indirection
 132                  * level. Retaining birth time enables detection of when
 133                  * holes are punched for reducing the number of free
 134                  * records transmitted during a zfs send.
 135                  */
 136 
 137                 uint64_t lsize = BP_GET_LSIZE(bp);
 138                 dmu_object_type_t type = BP_GET_TYPE(bp);
 139                 uint64_t lvl = BP_GET_LEVEL(bp);
 140 
 141                 bzero(bp, sizeof (blkptr_t));
 142 
 143                 if (spa_feature_is_active(dn->dn_objset->os_spa,
 144                     SPA_FEATURE_HOLE_BIRTH)) {
 145                         BP_SET_LSIZE(bp, lsize);
 146                         BP_SET_TYPE(bp, type);
 147                         BP_SET_LEVEL(bp, lvl);
 148                         BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
 149                 }
 150         }
 151         dnode_diduse_space(dn, -bytesfreed);
 152 }
 153 
 154 #ifdef ZFS_DEBUG
 155 static void
 156 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 157 {
 158         int off, num;
 159         int i, err, epbs;
 160         uint64_t txg = tx->tx_txg;
 161         dnode_t *dn;
 162 
 163         DB_DNODE_ENTER(db);
 164         dn = DB_DNODE(db);
 165         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 166         off = start - (db->db_blkid * 1<<epbs);
 167         num = end - start + 1;
 168 
 169         ASSERT3U(off, >=, 0);
 170         ASSERT3U(num, >=, 0);
 171         ASSERT3U(db->db_level, >, 0);
 172         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 173         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 174         ASSERT(db->db_blkptr != NULL);
 175 
 176         for (i = off; i < off+num; i++) {
 177                 uint64_t *buf;
 178                 dmu_buf_impl_t *child;
 179                 dbuf_dirty_record_t *dr;
 180                 int j;
 181 
 182                 ASSERT(db->db_level == 1);
 183 
 184                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 185                 err = dbuf_hold_impl(dn, db->db_level-1,
 186                     (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
 187                 rw_exit(&dn->dn_struct_rwlock);
 188                 if (err == ENOENT)
 189                         continue;
 190                 ASSERT(err == 0);
 191                 ASSERT(child->db_level == 0);
 192                 dr = child->db_last_dirty;
 193                 while (dr && dr->dr_txg > txg)
 194                         dr = dr->dr_next;
 195                 ASSERT(dr == NULL || dr->dr_txg == txg);
 196 
 197                 /* data_old better be zeroed */
 198                 if (dr) {
 199                         buf = dr->dt.dl.dr_data->b_data;
 200                         for (j = 0; j < child->db.db_size >> 3; j++) {
 201                                 if (buf[j] != 0) {
 202                                         panic("freed data not zero: "
 203                                             "child=%p i=%d off=%d num=%d\n",
 204                                             (void *)child, i, off, num);
 205                                 }
 206                         }
 207                 }
 208 
 209                 /*
 210                  * db_data better be zeroed unless it's dirty in a
 211                  * future txg.
 212                  */
 213                 mutex_enter(&child->db_mtx);
 214                 buf = child->db.db_data;
 215                 if (buf != NULL && child->db_state != DB_FILL &&
 216                     child->db_last_dirty == NULL) {
 217                         for (j = 0; j < child->db.db_size >> 3; j++) {
 218                                 if (buf[j] != 0) {
 219                                         panic("freed data not zero: "
 220                                             "child=%p i=%d off=%d num=%d\n",
 221                                             (void *)child, i, off, num);
 222                                 }
 223                         }
 224                 }
 225                 mutex_exit(&child->db_mtx);
 226 
 227                 dbuf_rele(child, FTAG);
 228         }
 229         DB_DNODE_EXIT(db);
 230 }
 231 #endif
 232 
 233 static void
 234 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 235     dmu_tx_t *tx)
 236 {
 237         dnode_t *dn;
 238         blkptr_t *bp;
 239         dmu_buf_impl_t *subdb;
 240         uint64_t start, end, dbstart, dbend;
 241         unsigned int epbs, shift, i;
 242 
 243         /*
 244          * There is a small possibility that this block will not be cached:
 245          *   1 - if level > 1 and there are no children with level <= 1
 246          *   2 - if this block was evicted since we read it from
 247          *       dmu_tx_hold_free().
 248          */
 249         if (db->db_state != DB_CACHED)
 250                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 251 
 252         dbuf_release_bp(db);
 253         bp = db->db.db_data;
 254 
 255         DB_DNODE_ENTER(db);
 256         dn = DB_DNODE(db);
 257         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 258         ASSERT3U(epbs, <, 31);
 259         shift = (db->db_level - 1) * epbs;
 260         dbstart = db->db_blkid << epbs;
 261         start = blkid >> shift;
 262         if (dbstart < start) {
 263                 bp += start - dbstart;
 264         } else {
 265                 start = dbstart;
 266         }
 267         dbend = ((db->db_blkid + 1) << epbs) - 1;
 268         end = (blkid + nblks - 1) >> shift;
 269         if (dbend <= end)
 270                 end = dbend;
 271 
 272         ASSERT3U(start, <=, end);
 273 
 274         if (db->db_level == 1) {
 275                 FREE_VERIFY(db, start, end, tx);
 276                 free_blocks(dn, bp, end-start+1, tx);
 277         } else {
 278                 for (uint64_t id = start; id <= end; id++, bp++) {
 279                         if (BP_IS_HOLE(bp))
 280                                 continue;
 281                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 282                         VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
 283                             id, TRUE, FALSE, FTAG, &subdb));
 284                         rw_exit(&dn->dn_struct_rwlock);
 285                         ASSERT3P(bp, ==, subdb->db_blkptr);
 286 
 287                         free_children(subdb, blkid, nblks, tx);
 288                         dbuf_rele(subdb, FTAG);
 289                 }
 290         }
 291 
 292         /* If this whole block is free, free ourself too. */
 293         for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
 294                 if (!BP_IS_HOLE(bp))
 295                         break;
 296         }
 297         if (i == 1 << epbs) {
 298                 /*
 299                  * We only found holes. Grab the rwlock to prevent
 300                  * anybody from reading the blocks we're about to
 301                  * zero out.
 302                  */
 303                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 304                 bzero(db->db.db_data, db->db.db_size);
 305                 rw_exit(&dn->dn_struct_rwlock);
 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 
 355                 ASSERT(start < dn->dn_phys->dn_nblkptr);
 356                 bp += start;
 357                 for (int i = start; i <= end; i++, bp++) {
 358                         if (BP_IS_HOLE(bp))
 359                                 continue;
 360                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 361                         VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
 362                             TRUE, FALSE, FTAG, &db));
 363                         rw_exit(&dn->dn_struct_rwlock);
 364 
 365                         free_children(db, blkid, nblks, tx);
 366                         dbuf_rele(db, FTAG);
 367                 }
 368         }
 369 
 370         if (trunc) {
 371                 dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
 372 
 373                 uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
 374                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 375                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 376                     dn->dn_phys->dn_maxblkid == 0 ||
 377                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 378         }
 379 }
 380 
 381 typedef struct dnode_sync_free_range_arg {
 382         dnode_t *dsfra_dnode;
 383         dmu_tx_t *dsfra_tx;
 384 } dnode_sync_free_range_arg_t;
 385 
 386 static void
 387 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
 388 {
 389         dnode_sync_free_range_arg_t *dsfra = arg;
 390         dnode_t *dn = dsfra->dsfra_dnode;
 391 
 392         mutex_exit(&dn->dn_mtx);
 393         dnode_sync_free_range_impl(dn, blkid, nblks, dsfra->dsfra_tx);
 394         mutex_enter(&dn->dn_mtx);
 395 }
 396 
 397 /*
 398  * Try to kick all the dnode's dbufs out of the cache...
 399  */
 400 void
 401 dnode_evict_dbufs(dnode_t *dn, int level)
 402 {
 403         dmu_buf_impl_t db_marker;
 404         dmu_buf_impl_t *db, *db_next;
 405 
 406         mutex_enter(&dn->dn_dbufs_mtx);
 407         for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
 408 
 409 #ifdef  DEBUG
 410                 DB_DNODE_ENTER(db);
 411                 ASSERT3P(DB_DNODE(db), ==, dn);
 412                 DB_DNODE_EXIT(db);
 413 #endif  /* DEBUG */
 414 
 415                 mutex_enter(&db->db_mtx);
 416                 if (level != DBUF_EVICT_ALL && db->db_level != level) {
 417                         mutex_exit(&db->db_mtx);
 418                         db_next = AVL_NEXT(&dn->dn_dbufs, db);
 419                         continue;
 420                 }
 421 
 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_destroy(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         dnode_evict_bonus(dn);
 443 }
 444 
 445 void
 446 dnode_evict_bonus(dnode_t *dn)
 447 {
 448         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 449         if (dn->dn_bonus != NULL) {
 450                 if (refcount_is_zero(&dn->dn_bonus->db_holds)) {
 451                         mutex_enter(&dn->dn_bonus->db_mtx);
 452                         dbuf_destroy(dn->dn_bonus);
 453                         dn->dn_bonus = NULL;
 454                 } else {
 455                         dn->dn_bonus->db_pending_evict = TRUE;
 456                 }
 457         }
 458         rw_exit(&dn->dn_struct_rwlock);
 459 }
 460 
 461 static void
 462 dnode_undirty_dbufs(list_t *list)
 463 {
 464         dbuf_dirty_record_t *dr;
 465 
 466         while (dr = list_head(list)) {
 467                 dmu_buf_impl_t *db = dr->dr_dbuf;
 468                 uint64_t txg = dr->dr_txg;
 469 
 470                 if (db->db_level != 0)
 471                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 472 
 473                 mutex_enter(&db->db_mtx);
 474                 /* XXX - use dbuf_undirty()? */
 475                 list_remove(list, dr);
 476                 ASSERT(db->db_last_dirty == dr);
 477                 db->db_last_dirty = NULL;
 478                 db->db_dirtycnt -= 1;
 479                 if (db->db_level == 0) {
 480                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 481                             dr->dt.dl.dr_data == db->db_buf);
 482                         dbuf_unoverride(dr);
 483                 } else {
 484                         mutex_destroy(&dr->dt.di.dr_mtx);
 485                         list_destroy(&dr->dt.di.dr_children);
 486                 }
 487                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 488                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
 489         }
 490 }
 491 
 492 static void
 493 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 494 {
 495         int txgoff = tx->tx_txg & TXG_MASK;
 496 
 497         ASSERT(dmu_tx_is_syncing(tx));
 498 
 499         /*
 500          * Our contents should have been freed in dnode_sync() by the
 501          * free range record inserted by the caller of dnode_free().
 502          */
 503         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 504         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 505 
 506         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 507         dnode_evict_dbufs(dn, DBUF_EVICT_ALL);
 508         ASSERT(avl_is_empty(&dn->dn_dbufs));
 509 
 510         /*
 511          * XXX - It would be nice to assert this, but we may still
 512          * have residual holds from async evictions from the arc...
 513          *
 514          * zfs_obj_to_path() also depends on this being
 515          * commented out.
 516          *
 517          * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
 518          */
 519 
 520         /* Undirty next bits */
 521         dn->dn_next_nlevels[txgoff] = 0;
 522         dn->dn_next_indblkshift[txgoff] = 0;
 523         dn->dn_next_blksz[txgoff] = 0;
 524 
 525         /* ASSERT(blkptrs are zero); */
 526         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 527         ASSERT(dn->dn_type != DMU_OT_NONE);
 528 
 529         ASSERT(dn->dn_free_txg > 0);
 530         if (dn->dn_allocated_txg != dn->dn_free_txg)
 531                 dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
 532         bzero(dn->dn_phys, sizeof (dnode_phys_t));
 533 
 534         mutex_enter(&dn->dn_mtx);
 535         dn->dn_type = DMU_OT_NONE;
 536         dn->dn_maxblkid = 0;
 537         dn->dn_allocated_txg = 0;
 538         dn->dn_free_txg = 0;
 539         dn->dn_have_spill = B_FALSE;
 540         mutex_exit(&dn->dn_mtx);
 541 
 542         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 543 
 544         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 545         /*
 546          * Now that we've released our hold, the dnode may
 547          * be evicted, so we musn't access it.
 548          */
 549 }
 550 
 551 /*
 552  * Write out the dnode's dirty buffers.
 553  */
 554 void
 555 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 556 {
 557         dnode_phys_t *dnp = dn->dn_phys;
 558         int txgoff = tx->tx_txg & TXG_MASK;
 559         list_t *list = &dn->dn_dirty_records[txgoff];
 560         static const dnode_phys_t zerodn = { 0 };
 561         boolean_t kill_spill = B_FALSE;
 562 
 563         ASSERT(dmu_tx_is_syncing(tx));
 564         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 565         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 566             bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
 567         DNODE_VERIFY(dn);
 568 
 569         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 570 
 571         if (dmu_objset_userused_enabled(dn->dn_objset) &&
 572             !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 573                 mutex_enter(&dn->dn_mtx);
 574                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 575                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 576                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 577                 mutex_exit(&dn->dn_mtx);
 578                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 579         } else {
 580                 /* Once we account for it, we should always account for it. */
 581                 ASSERT(!(dn->dn_phys->dn_flags &
 582                     DNODE_FLAG_USERUSED_ACCOUNTED));
 583         }
 584 
 585         mutex_enter(&dn->dn_mtx);
 586         if (dn->dn_allocated_txg == tx->tx_txg) {
 587                 /* The dnode is newly allocated or reallocated */
 588                 if (dnp->dn_type == DMU_OT_NONE) {
 589                         /* this is a first alloc, not a realloc */
 590                         dnp->dn_nlevels = 1;
 591                         dnp->dn_nblkptr = dn->dn_nblkptr;
 592                 }
 593 
 594                 dnp->dn_type = dn->dn_type;
 595                 dnp->dn_bonustype = dn->dn_bonustype;
 596                 dnp->dn_bonuslen = dn->dn_bonuslen;
 597         }
 598         ASSERT(dnp->dn_nlevels > 1 ||
 599             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 600             BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
 601             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 602             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 603         ASSERT(dnp->dn_nlevels < 2 ||
 604             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 605             BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
 606 
 607         if (dn->dn_next_type[txgoff] != 0) {
 608                 dnp->dn_type = dn->dn_type;
 609                 dn->dn_next_type[txgoff] = 0;
 610         }
 611 
 612         if (dn->dn_next_blksz[txgoff] != 0) {
 613                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 614                     SPA_MINBLOCKSIZE) == 0);
 615                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 616                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 617                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 618                     dnp->dn_datablkszsec ||
 619                     range_tree_space(dn->dn_free_ranges[txgoff]) != 0);
 620                 dnp->dn_datablkszsec =
 621                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 622                 dn->dn_next_blksz[txgoff] = 0;
 623         }
 624 
 625         if (dn->dn_next_bonuslen[txgoff] != 0) {
 626                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 627                         dnp->dn_bonuslen = 0;
 628                 else
 629                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 630                 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
 631                 dn->dn_next_bonuslen[txgoff] = 0;
 632         }
 633 
 634         if (dn->dn_next_bonustype[txgoff] != 0) {
 635                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 636                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 637                 dn->dn_next_bonustype[txgoff] = 0;
 638         }
 639 
 640         boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
 641             dn->dn_free_txg <= tx->tx_txg;
 642 
 643         /*
 644          * Remove the spill block if we have been explicitly asked to
 645          * remove it, or if the object is being removed.
 646          */
 647         if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
 648                 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 649                         kill_spill = B_TRUE;
 650                 dn->dn_rm_spillblk[txgoff] = 0;
 651         }
 652 
 653         if (dn->dn_next_indblkshift[txgoff] != 0) {
 654                 ASSERT(dnp->dn_nlevels == 1);
 655                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 656                 dn->dn_next_indblkshift[txgoff] = 0;
 657         }
 658 
 659         /*
 660          * Just take the live (open-context) values for checksum and compress.
 661          * Strictly speaking it's a future leak, but nothing bad happens if we
 662          * start using the new checksum or compress algorithm a little early.
 663          */
 664         dnp->dn_checksum = dn->dn_checksum;
 665         dnp->dn_compress = dn->dn_compress;
 666 
 667         mutex_exit(&dn->dn_mtx);
 668 
 669         if (kill_spill) {
 670                 free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
 671                 mutex_enter(&dn->dn_mtx);
 672                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 673                 mutex_exit(&dn->dn_mtx);
 674         }
 675 
 676         /* process all the "freed" ranges in the file */
 677         if (dn->dn_free_ranges[txgoff] != NULL) {
 678                 dnode_sync_free_range_arg_t dsfra;
 679                 dsfra.dsfra_dnode = dn;
 680                 dsfra.dsfra_tx = tx;
 681                 mutex_enter(&dn->dn_mtx);
 682                 range_tree_vacate(dn->dn_free_ranges[txgoff],
 683                     dnode_sync_free_range, &dsfra);
 684                 range_tree_destroy(dn->dn_free_ranges[txgoff]);
 685                 dn->dn_free_ranges[txgoff] = NULL;
 686                 mutex_exit(&dn->dn_mtx);
 687         }
 688 
 689         if (freeing_dnode) {
 690                 dn->dn_objset->os_freed_dnodes++;
 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 = dn->dn_next_nblkptr[txgoff];
 714                             i < dnp->dn_nblkptr; i++)
 715                                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 716 #endif
 717                 }
 718                 mutex_enter(&dn->dn_mtx);
 719                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 720                 dn->dn_next_nblkptr[txgoff] = 0;
 721                 mutex_exit(&dn->dn_mtx);
 722         }
 723 
 724         dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
 725 
 726         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 727                 ASSERT3P(list_head(list), ==, NULL);
 728                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 729         }
 730 
 731         /*
 732          * Although we have dropped our reference to the dnode, it
 733          * can't be evicted until its written, and we haven't yet
 734          * initiated the IO for the dnode's dbuf.
 735          */
 736 }