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  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  */
  24
  25 #include <sys/zfs_context.h>
  26 #include <sys/dbuf.h>
  27 #include <sys/dnode.h>
  28 #include <sys/dmu.h>
  29 #include <sys/dmu_tx.h>
  30 #include <sys/dmu_objset.h>
  31 #include <sys/dsl_dataset.h>
  32 #include <sys/spa.h>
  33
  34 static void
  35 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
  36 {
  37         dmu_buf_impl_t *db;
  38         int txgoff = tx->tx_txg & TXG_MASK;
  39         int nblkptr = dn->dn_phys->dn_nblkptr;
  40         int old_toplvl = dn->dn_phys->dn_nlevels - 1;
  41         int new_level = dn->dn_next_nlevels[txgoff];
  42         int i;
  43
  44         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
  45
  46         /* this dnode can't be paged out because it's dirty */
  47         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
  48         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
  49         ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
  50
  51         db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
  52         ASSERT(db != NULL);
  53
  54         dn->dn_phys->dn_nlevels = new_level;
  55         dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
  56             dn->dn_object, dn->dn_phys->dn_nlevels);
  57
  58         /* check for existing blkptrs in the dnode */
  59         for (i = 0; i < nblkptr; i++)
  60                 if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
  61                         break;
  62         if (i != nblkptr) {
  63                 /* transfer dnode's block pointers to new indirect block */
  64                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
  65                 ASSERT(db->db.db_data);
  66                 ASSERT(arc_released(db->db_buf));
  67                 ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
  68                 bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
  69                     sizeof (blkptr_t) * nblkptr);
  70                 arc_buf_freeze(db->db_buf);
  71         }
  72
  73         /* set dbuf's parent pointers to new indirect buf */
  74         for (i = 0; i < nblkptr; i++) {
  75                 dmu_buf_impl_t *child = dbuf_find(dn, old_toplvl, i);
  76
  77                 if (child == NULL)
  78                         continue;
  79 #ifdef  DEBUG
  80                 DB_DNODE_ENTER(child);
  81                 ASSERT3P(DB_DNODE(child), ==, dn);
  82                 DB_DNODE_EXIT(child);
  83 #endif  /* DEBUG */
  84                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  85                         ASSERT(child->db_parent->db_level == db->db_level);
  86                         ASSERT(child->db_blkptr !=
  87                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  88                         mutex_exit(&child->db_mtx);
  89                         continue;
  90                 }
  91                 ASSERT(child->db_parent == NULL ||
  92                     child->db_parent == dn->dn_dbuf);
  93
  94                 child->db_parent = db;
  95                 dbuf_add_ref(db, child);
  96                 if (db->db.db_data)
  97                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
  98                 else
  99                         child->db_blkptr = NULL;
 100                 dprintf_dbuf_bp(child, child->db_blkptr,
 101                     "changed db_blkptr to new indirect %s", "");
 102
 103                 mutex_exit(&child->db_mtx);
 104         }
 105
 106         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 107
 108         dbuf_rele(db, FTAG);
 109
 110         rw_exit(&dn->dn_struct_rwlock);
 111 }
 112
 113 static int
 114 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 115 {
 116         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 117         uint64_t bytesfreed = 0;
 118         int i, blocks_freed = 0;
 119
 120         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 121
 122         for (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                 bzero(bp, sizeof (blkptr_t));
 129                 blocks_freed += 1;
 130         }
 131         dnode_diduse_space(dn, -bytesfreed);
 132         return (blocks_freed);
 133 }
 134
 135 #ifdef ZFS_DEBUG
 136 static void
 137 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 138 {
 139         int off, num;
 140         int i, err, epbs;
 141         uint64_t txg = tx->tx_txg;
 142         dnode_t *dn;
 143
 144         DB_DNODE_ENTER(db);
 145         dn = DB_DNODE(db);
 146         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 147         off = start - (db->db_blkid * 1<<epbs);
 148         num = end - start + 1;
 149
 150         ASSERT3U(off, >=, 0);
 151         ASSERT3U(num, >=, 0);
 152         ASSERT3U(db->db_level, >, 0);
 153         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 154         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 155         ASSERT(db->db_blkptr != NULL);
 156
 157         for (i = off; i < off+num; i++) {
 158                 uint64_t *buf;
 159                 dmu_buf_impl_t *child;
 160                 dbuf_dirty_record_t *dr;
 161                 int j;
 162
 163                 ASSERT(db->db_level == 1);
 164
 165                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 166                 err = dbuf_hold_impl(dn, db->db_level-1,
 167                     (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
 168                 rw_exit(&dn->dn_struct_rwlock);
 169                 if (err == ENOENT)
 170                         continue;
 171                 ASSERT(err == 0);
 172                 ASSERT(child->db_level == 0);
 173                 dr = child->db_last_dirty;
 174                 while (dr && dr->dr_txg > txg)
 175                         dr = dr->dr_next;
 176                 ASSERT(dr == NULL || dr->dr_txg == txg);
 177
 178                 /* data_old better be zeroed */
 179                 if (dr) {
 180                         buf = dr->dt.dl.dr_data->b_data;
 181                         for (j = 0; j < child->db.db_size >> 3; j++) {
 182                                 if (buf[j] != 0) {
 183                                         panic("freed data not zero: "
 184                                             "child=%p i=%d off=%d num=%d\n",
 185                                             (void *)child, i, off, num);
 186                                 }
 187                         }
 188                 }
 189
 190                 /*
 191                  * db_data better be zeroed unless it's dirty in a
 192                  * future txg.
 193                  */
 194                 mutex_enter(&child->db_mtx);
 195                 buf = child->db.db_data;
 196                 if (buf != NULL && child->db_state != DB_FILL &&
 197                     child->db_last_dirty == NULL) {
 198                         for (j = 0; j < child->db.db_size >> 3; j++) {
 199                                 if (buf[j] != 0) {
 200                                         panic("freed data not zero: "
 201                                             "child=%p i=%d off=%d num=%d\n",
 202                                             (void *)child, i, off, num);
 203                                 }
 204                         }
 205                 }
 206                 mutex_exit(&child->db_mtx);
 207
 208                 dbuf_rele(child, FTAG);
 209         }
 210         DB_DNODE_EXIT(db);
 211 }
 212 #endif
 213
 214 #define ALL -1
 215
 216 static int
 217 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks, int trunc,
 218     dmu_tx_t *tx)
 219 {
 220         dnode_t *dn;
 221         blkptr_t *bp;
 222         dmu_buf_impl_t *subdb;
 223         uint64_t start, end, dbstart, dbend, i;
 224         int epbs, shift, err;
 225         int all = TRUE;
 226         int blocks_freed = 0;
 227
 228         /*
 229          * There is a small possibility that this block will not be cached:
 230          *   1 - if level > 1 and there are no children with level <= 1
 231          *   2 - if we didn't get a dirty hold (because this block had just
 232          *       finished being written -- and so had no holds), and then this
 233          *       block got evicted before we got here.
 234          */
 235         if (db->db_state != DB_CACHED)
 236                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 237
 238         dbuf_release_bp(db);
 239         bp = (blkptr_t *)db->db.db_data;
 240
 241         DB_DNODE_ENTER(db);
 242         dn = DB_DNODE(db);
 243         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 244         shift = (db->db_level - 1) * epbs;
 245         dbstart = db->db_blkid << epbs;
 246         start = blkid >> shift;
 247         if (dbstart < start) {
 248                 bp += start - dbstart;
 249                 all = FALSE;
 250         } else {
 251                 start = dbstart;
 252         }
 253         dbend = ((db->db_blkid + 1) << epbs) - 1;
 254         end = (blkid + nblks - 1) >> shift;
 255         if (dbend <= end)
 256                 end = dbend;
 257         else if (all)
 258                 all = trunc;
 259         ASSERT3U(start, <=, end);
 260
 261         if (db->db_level == 1) {
 262                 FREE_VERIFY(db, start, end, tx);
 263                 blocks_freed = free_blocks(dn, bp, end-start+1, tx);
 264                 arc_buf_freeze(db->db_buf);
 265                 ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
 266                 DB_DNODE_EXIT(db);
 267                 return (all ? ALL : blocks_freed);
 268         }
 269
 270         for (i = start; i <= end; i++, bp++) {
 271                 if (BP_IS_HOLE(bp))
 272                         continue;
 273                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 274                 err = dbuf_hold_impl(dn, db->db_level-1, i, TRUE, FTAG, &subdb);
 275                 ASSERT3U(err, ==, 0);
 276                 rw_exit(&dn->dn_struct_rwlock);
 277
 278                 if (free_children(subdb, blkid, nblks, trunc, tx) == ALL) {
 279                         ASSERT3P(subdb->db_blkptr, ==, bp);
 280                         blocks_freed += free_blocks(dn, bp, 1, tx);
 281                 } else {
 282                         all = FALSE;
 283                 }
 284                 dbuf_rele(subdb, FTAG);
 285         }
 286         DB_DNODE_EXIT(db);
 287         arc_buf_freeze(db->db_buf);
 288 #ifdef ZFS_DEBUG
 289         bp -= (end-start)+1;
 290         for (i = start; i <= end; i++, bp++) {
 291                 if (i == start && blkid != 0)
 292                         continue;
 293                 else if (i == end && !trunc)
 294                         continue;
 295                 ASSERT3U(bp->blk_birth, ==, 0);
 296         }
 297 #endif
 298         ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
 299         return (all ? ALL : blocks_freed);
 300 }
 301
 302 /*
 303  * free_range: Traverse the indicated range of the provided file
 304  * and "free" all the blocks contained there.
 305  */
 306 static void
 307 dnode_sync_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
 308 {
 309         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 310         dmu_buf_impl_t *db;
 311         int trunc, start, end, shift, i, err;
 312         int dnlevel = dn->dn_phys->dn_nlevels;
 313
 314         if (blkid > dn->dn_phys->dn_maxblkid)
 315                 return;
 316
 317         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 318         trunc = blkid + nblks > dn->dn_phys->dn_maxblkid;
 319         if (trunc)
 320                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 321
 322         /* There are no indirect blocks in the object */
 323         if (dnlevel == 1) {
 324                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 325                         /* this range was never made persistent */
 326                         return;
 327                 }
 328                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 329                 (void) free_blocks(dn, bp + blkid, nblks, tx);
 330                 if (trunc) {
 331                         ASSERTV(uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
 332                             (dn->dn_phys->dn_datablkszsec<<SPA_MINBLOCKSHIFT));
 333                         dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
 334                         ASSERT(off < dn->dn_phys->dn_maxblkid ||
 335                             dn->dn_phys->dn_maxblkid == 0 ||
 336                             dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 337                 }
 338                 return;
 339         }
 340
 341         shift = (dnlevel - 1) * (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 342         start = blkid >> shift;
 343         ASSERT(start < dn->dn_phys->dn_nblkptr);
 344         end = (blkid + nblks - 1) >> shift;
 345         bp += start;
 346         for (i = start; i <= end; i++, bp++) {
 347                 if (BP_IS_HOLE(bp))
 348                         continue;
 349                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 350                 err = dbuf_hold_impl(dn, dnlevel-1, i, TRUE, FTAG, &db);
 351                 ASSERT3U(err, ==, 0);
 352                 rw_exit(&dn->dn_struct_rwlock);
 353
 354                 if (free_children(db, blkid, nblks, trunc, tx) == ALL) {
 355                         ASSERT3P(db->db_blkptr, ==, bp);
 356                         (void) free_blocks(dn, bp, 1, tx);
 357                 }
 358                 dbuf_rele(db, FTAG);
 359         }
 360         if (trunc) {
 361                 ASSERTV(uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
 362                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT));
 363                 dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
 364                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 365                     dn->dn_phys->dn_maxblkid == 0 ||
 366                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 367         }
 368 }
 369
 370 /*
 371  * Try to kick all the dnodes dbufs out of the cache...
 372  */
 373 void
 374 dnode_evict_dbufs(dnode_t *dn)
 375 {
 376         int progress;
 377         int pass = 0;
 378
 379         do {
 380                 dmu_buf_impl_t *db, marker;
 381                 int evicting = FALSE;
 382
 383                 progress = FALSE;
 384                 mutex_enter(&dn->dn_dbufs_mtx);
 385                 list_insert_tail(&dn->dn_dbufs, &marker);
 386                 db = list_head(&dn->dn_dbufs);
 387                 for (; db != &marker; db = list_head(&dn->dn_dbufs)) {
 388                         list_remove(&dn->dn_dbufs, db);
 389                         list_insert_tail(&dn->dn_dbufs, db);
 390 #ifdef  DEBUG
 391                         DB_DNODE_ENTER(db);
 392                         ASSERT3P(DB_DNODE(db), ==, dn);
 393                         DB_DNODE_EXIT(db);
 394 #endif  /* DEBUG */
 395
 396                         mutex_enter(&db->db_mtx);
 397                         if (db->db_state == DB_EVICTING) {
 398                                 progress = TRUE;
 399                                 evicting = TRUE;
 400                                 mutex_exit(&db->db_mtx);
 401                         } else if (refcount_is_zero(&db->db_holds)) {
 402                                 progress = TRUE;
 403                                 dbuf_clear(db); /* exits db_mtx for us */
 404                         } else {
 405                                 mutex_exit(&db->db_mtx);
 406                         }
 407
 408                 }
 409                 list_remove(&dn->dn_dbufs, &marker);
 410                 /*
 411                  * NB: we need to drop dn_dbufs_mtx between passes so
 412                  * that any DB_EVICTING dbufs can make progress.
 413                  * Ideally, we would have some cv we could wait on, but
 414                  * since we don't, just wait a bit to give the other
 415                  * thread a chance to run.
 416                  */
 417                 mutex_exit(&dn->dn_dbufs_mtx);
 418                 if (evicting)
 419                         delay(1);
 420                 pass++;
 421                 if ((pass % 100) == 0)
 422                         dprintf("Exceeded %d passes evicting dbufs\n", pass);
 423         } while (progress);
 424
 425         if (pass >= 100)
 426                 dprintf("Required %d passes to evict dbufs\n", pass);
 427
 428         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 429         if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
 430                 mutex_enter(&dn->dn_bonus->db_mtx);
 431                 dbuf_evict(dn->dn_bonus);
 432                 dn->dn_bonus = NULL;
 433         }
 434         rw_exit(&dn->dn_struct_rwlock);
 435 }
 436
 437 static void
 438 dnode_undirty_dbufs(list_t *list)
 439 {
 440         dbuf_dirty_record_t *dr;
 441
 442         while ((dr = list_head(list))) {
 443                 dmu_buf_impl_t *db = dr->dr_dbuf;
 444                 uint64_t txg = dr->dr_txg;
 445
 446                 if (db->db_level != 0)
 447                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 448
 449                 mutex_enter(&db->db_mtx);
 450                 /* XXX - use dbuf_undirty()? */
 451                 list_remove(list, dr);
 452                 ASSERT(db->db_last_dirty == dr);
 453                 db->db_last_dirty = NULL;
 454                 db->db_dirtycnt -= 1;
 455                 if (db->db_level == 0) {
 456                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 457                             dr->dt.dl.dr_data == db->db_buf);
 458                         dbuf_unoverride(dr);
 459                 }
 460                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 461                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
 462         }
 463 }
 464
 465 static void
 466 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 467 {
 468         int txgoff = tx->tx_txg & TXG_MASK;
 469
 470         ASSERT(dmu_tx_is_syncing(tx));
 471
 472         /*
 473          * Our contents should have been freed in dnode_sync() by the
 474          * free range record inserted by the caller of dnode_free().
 475          */
 476         ASSERT3U(DN_USED_BYTES(dn->dn_phys), ==, 0);
 477         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 478
 479         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 480         dnode_evict_dbufs(dn);
 481         ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
 482
 483         /*
 484          * XXX - It would be nice to assert this, but we may still
 485          * have residual holds from async evictions from the arc...
 486          *
 487          * zfs_obj_to_path() also depends on this being
 488          * commented out.
 489          *
 490          * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
 491          */
 492
 493         /* Undirty next bits */
 494         dn->dn_next_nlevels[txgoff] = 0;
 495         dn->dn_next_indblkshift[txgoff] = 0;
 496         dn->dn_next_blksz[txgoff] = 0;
 497
 498         /* ASSERT(blkptrs are zero); */
 499         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 500         ASSERT(dn->dn_type != DMU_OT_NONE);
 501
 502         ASSERT(dn->dn_free_txg > 0);
 503         if (dn->dn_allocated_txg != dn->dn_free_txg)
 504                 dbuf_will_dirty(dn->dn_dbuf, tx);
 505         bzero(dn->dn_phys, sizeof (dnode_phys_t));
 506
 507         mutex_enter(&dn->dn_mtx);
 508         dn->dn_type = DMU_OT_NONE;
 509         dn->dn_maxblkid = 0;
 510         dn->dn_allocated_txg = 0;
 511         dn->dn_free_txg = 0;
 512         dn->dn_have_spill = B_FALSE;
 513         mutex_exit(&dn->dn_mtx);
 514
 515         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 516
 517         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 518         /*
 519          * Now that we've released our hold, the dnode may
 520          * be evicted, so we musn't access it.
 521          */
 522 }
 523
 524 /*
 525  * Write out the dnode's dirty buffers.
 526  */
 527 void
 528 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 529 {
 530         free_range_t *rp;
 531         dnode_phys_t *dnp = dn->dn_phys;
 532         int txgoff = tx->tx_txg & TXG_MASK;
 533         list_t *list = &dn->dn_dirty_records[txgoff];
 534         boolean_t kill_spill = B_FALSE;
 535         ASSERTV(static const dnode_phys_t zerodn = { 0 });
 536
 537         ASSERT(dmu_tx_is_syncing(tx));
 538         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 539         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 540             bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
 541         DNODE_VERIFY(dn);
 542
 543         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 544
 545         if (dmu_objset_userused_enabled(dn->dn_objset) &&
 546             !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 547                 mutex_enter(&dn->dn_mtx);
 548                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 549                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 550                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 551                 mutex_exit(&dn->dn_mtx);
 552                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 553         } else {
 554                 /* Once we account for it, we should always account for it. */
 555                 ASSERT(!(dn->dn_phys->dn_flags &
 556                     DNODE_FLAG_USERUSED_ACCOUNTED));
 557         }
 558
 559         mutex_enter(&dn->dn_mtx);
 560         if (dn->dn_allocated_txg == tx->tx_txg) {
 561                 /* The dnode is newly allocated or reallocated */
 562                 if (dnp->dn_type == DMU_OT_NONE) {
 563                         /* this is a first alloc, not a realloc */
 564                         dnp->dn_nlevels = 1;
 565                         dnp->dn_nblkptr = dn->dn_nblkptr;
 566                 }
 567
 568                 dnp->dn_type = dn->dn_type;
 569                 dnp->dn_bonustype = dn->dn_bonustype;
 570                 dnp->dn_bonuslen = dn->dn_bonuslen;
 571         }
 572
 573         ASSERT(dnp->dn_nlevels > 1 ||
 574             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 575             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 576             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 577
 578         if (dn->dn_next_blksz[txgoff]) {
 579                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 580                     SPA_MINBLOCKSIZE) == 0);
 581                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 582                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 583                     avl_last(&dn->dn_ranges[txgoff]) ||
 584                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 585                     dnp->dn_datablkszsec);
 586                 dnp->dn_datablkszsec =
 587                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 588                 dn->dn_next_blksz[txgoff] = 0;
 589         }
 590
 591         if (dn->dn_next_bonuslen[txgoff]) {
 592                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 593                         dnp->dn_bonuslen = 0;
 594                 else
 595                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 596                 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
 597                 dn->dn_next_bonuslen[txgoff] = 0;
 598         }
 599
 600         if (dn->dn_next_bonustype[txgoff]) {
 601                 ASSERT(dn->dn_next_bonustype[txgoff] < DMU_OT_NUMTYPES);
 602                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 603                 dn->dn_next_bonustype[txgoff] = 0;
 604         }
 605
 606         /*
 607          * We will either remove a spill block when a file is being removed
 608          * or we have been asked to remove it.
 609          */
 610         if (dn->dn_rm_spillblk[txgoff] ||
 611             ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) &&
 612             dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg)) {
 613                 if ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
 614                         kill_spill = B_TRUE;
 615                 dn->dn_rm_spillblk[txgoff] = 0;
 616         }
 617
 618         if (dn->dn_next_indblkshift[txgoff]) {
 619                 ASSERT(dnp->dn_nlevels == 1);
 620                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 621                 dn->dn_next_indblkshift[txgoff] = 0;
 622         }
 623
 624         /*
 625          * Just take the live (open-context) values for checksum and compress.
 626          * Strictly speaking it's a future leak, but nothing bad happens if we
 627          * start using the new checksum or compress algorithm a little early.
 628          */
 629         dnp->dn_checksum = dn->dn_checksum;
 630         dnp->dn_compress = dn->dn_compress;
 631
 632         mutex_exit(&dn->dn_mtx);
 633
 634         if (kill_spill) {
 635                 (void) free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
 636                 mutex_enter(&dn->dn_mtx);
 637                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 638                 mutex_exit(&dn->dn_mtx);
 639         }
 640
 641         /* process all the "freed" ranges in the file */
 642         while ((rp = avl_last(&dn->dn_ranges[txgoff]))) {
 643                 dnode_sync_free_range(dn, rp->fr_blkid, rp->fr_nblks, tx);
 644                 /* grab the mutex so we don't race with dnode_block_freed() */
 645                 mutex_enter(&dn->dn_mtx);
 646                 avl_remove(&dn->dn_ranges[txgoff], rp);
 647                 mutex_exit(&dn->dn_mtx);
 648                 kmem_free(rp, sizeof (free_range_t));
 649         }
 650
 651         if (dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg) {
 652                 dnode_sync_free(dn, tx);
 653                 return;
 654         }
 655
 656         if (dn->dn_next_nblkptr[txgoff]) {
 657                 /* this should only happen on a realloc */
 658                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 659                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 660                         /* zero the new blkptrs we are gaining */
 661                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 662                             sizeof (blkptr_t) *
 663                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 664 #ifdef ZFS_DEBUG
 665                 } else {
 666                         int i;
 667                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 668                         /* the blkptrs we are losing better be unallocated */
 669                         for (i = dn->dn_next_nblkptr[txgoff];
 670                             i < dnp->dn_nblkptr; i++)
 671                                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 672 #endif
 673                 }
 674                 mutex_enter(&dn->dn_mtx);
 675                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 676                 dn->dn_next_nblkptr[txgoff] = 0;
 677                 mutex_exit(&dn->dn_mtx);
 678         }
 679
 680         if (dn->dn_next_nlevels[txgoff]) {
 681                 dnode_increase_indirection(dn, tx);
 682                 dn->dn_next_nlevels[txgoff] = 0;
 683         }
 684
 685         dbuf_sync_list(list, tx);
 686
 687         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 688                 ASSERT3P(list_head(list), ==, NULL);
 689                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 690         }
 691
 692         /*
 693          * Although we have dropped our reference to the dnode, it
 694          * can't be evicted until its written, and we haven't yet
 695          * initiated the IO for the dnode's dbuf.
 696          */
 697 }