sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dnode_sync.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21
  22 /*
  23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright (c) 2012 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
  36 static void
  37 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
  38 {
  39         dmu_buf_impl_t *db;
  40         int txgoff = tx->tx_txg & TXG_MASK;
  41         int nblkptr = dn->dn_phys->dn_nblkptr;
  42         int old_toplvl = dn->dn_phys->dn_nlevels - 1;
  43         int new_level = dn->dn_next_nlevels[txgoff];
  44         int i;
  45
  46         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
  47
  48         /* this dnode can't be paged out because it's dirty */
  49         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
  50         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
  51         ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
  52
  53         db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
  54         ASSERT(db != NULL);
  55
  56         dn->dn_phys->dn_nlevels = new_level;
  57         dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
  58             dn->dn_object, dn->dn_phys->dn_nlevels);
  59
  60         /* check for existing blkptrs in the dnode */
  61         for (i = 0; i < nblkptr; i++)
  62                 if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
  63                         break;
  64         if (i != nblkptr) {
  65                 /* transfer dnode's block pointers to new indirect block */
  66                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
  67                 ASSERT(db->db.db_data);
  68                 ASSERT(arc_released(db->db_buf));
  69                 ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
  70                 bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
  71                     sizeof (blkptr_t) * nblkptr);
  72                 arc_buf_freeze(db->db_buf);
  73         }
  74
  75         /* set dbuf's parent pointers to new indirect buf */
  76         for (i = 0; i < nblkptr; i++) {
  77                 dmu_buf_impl_t *child = dbuf_find(dn, old_toplvl, i);
  78
  79                 if (child == NULL)
  80                         continue;
  81 #ifdef  DEBUG
  82                 DB_DNODE_ENTER(child);
  83                 ASSERT3P(DB_DNODE(child), ==, dn);
  84                 DB_DNODE_EXIT(child);
  85 #endif  /* DEBUG */
  86                 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
  87                         ASSERT(child->db_parent->db_level == db->db_level);
  88                         ASSERT(child->db_blkptr !=
  89                             &dn->dn_phys->dn_blkptr[child->db_blkid]);
  90                         mutex_exit(&child->db_mtx);
  91                         continue;
  92                 }
  93                 ASSERT(child->db_parent == NULL ||
  94                     child->db_parent == dn->dn_dbuf);
  95
  96                 child->db_parent = db;
  97                 dbuf_add_ref(db, child);
  98                 if (db->db.db_data)
  99                         child->db_blkptr = (blkptr_t *)db->db.db_data + i;
 100                 else
 101                         child->db_blkptr = NULL;
 102                 dprintf_dbuf_bp(child, child->db_blkptr,
 103                     "changed db_blkptr to new indirect %s", "");
 104
 105                 mutex_exit(&child->db_mtx);
 106         }
 107
 108         bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
 109
 110         dbuf_rele(db, FTAG);
 111
 112         rw_exit(&dn->dn_struct_rwlock);
 113 }
 114
 115 static int
 116 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
 117 {
 118         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 119         uint64_t bytesfreed = 0;
 120         int i, blocks_freed = 0;
 121
 122         dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
 123
 124         for (i = 0; i < num; i++, bp++) {
 125                 if (BP_IS_HOLE(bp))
 126                         continue;
 127
 128                 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
 129                 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
 130                 bzero(bp, sizeof (blkptr_t));
 131                 blocks_freed += 1;
 132         }
 133         dnode_diduse_space(dn, -bytesfreed);
 134         return (blocks_freed);
 135 }
 136
 137 #ifdef ZFS_DEBUG
 138 static void
 139 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 140 {
 141         int off, num;
 142         int i, err, epbs;
 143         uint64_t txg = tx->tx_txg;
 144         dnode_t *dn;
 145
 146         DB_DNODE_ENTER(db);
 147         dn = DB_DNODE(db);
 148         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 149         off = start - (db->db_blkid * 1<<epbs);
 150         num = end - start + 1;
 151
 152         ASSERT3U(off, >=, 0);
 153         ASSERT3U(num, >=, 0);
 154         ASSERT3U(db->db_level, >, 0);
 155         ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
 156         ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
 157         ASSERT(db->db_blkptr != NULL);
 158
 159         for (i = off; i < off+num; i++) {
 160                 uint64_t *buf;
 161                 dmu_buf_impl_t *child;
 162                 dbuf_dirty_record_t *dr;
 163                 int j;
 164
 165                 ASSERT(db->db_level == 1);
 166
 167                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 168                 err = dbuf_hold_impl(dn, db->db_level-1,
 169                     (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
 170                 rw_exit(&dn->dn_struct_rwlock);
 171                 if (err == ENOENT)
 172                         continue;
 173                 ASSERT(err == 0);
 174                 ASSERT(child->db_level == 0);
 175                 dr = child->db_last_dirty;
 176                 while (dr && dr->dr_txg > txg)
 177                         dr = dr->dr_next;
 178                 ASSERT(dr == NULL || dr->dr_txg == txg);
 179
 180                 /* data_old better be zeroed */
 181                 if (dr) {
 182                         buf = dr->dt.dl.dr_data->b_data;
 183                         for (j = 0; j < child->db.db_size >> 3; j++) {
 184                                 if (buf[j] != 0) {
 185                                         panic("freed data not zero: "
 186                                             "child=%p i=%d off=%d num=%d\n",
 187                                             (void *)child, i, off, num);
 188                                 }
 189                         }
 190                 }
 191
 192                 /*
 193                  * db_data better be zeroed unless it's dirty in a
 194                  * future txg.
 195                  */
 196                 mutex_enter(&child->db_mtx);
 197                 buf = child->db.db_data;
 198                 if (buf != NULL && child->db_state != DB_FILL &&
 199                     child->db_last_dirty == NULL) {
 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                 mutex_exit(&child->db_mtx);
 209
 210                 dbuf_rele(child, FTAG);
 211         }
 212         DB_DNODE_EXIT(db);
 213 }
 214 #endif
 215
 216 #define ALL -1
 217
 218 static int
 219 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks, int trunc,
 220     dmu_tx_t *tx)
 221 {
 222         dnode_t *dn;
 223         blkptr_t *bp;
 224         dmu_buf_impl_t *subdb;
 225         uint64_t start, end, dbstart, dbend, i;
 226         int epbs, shift, err;
 227         int all = TRUE;
 228         int blocks_freed = 0;
 229
 230         /*
 231          * There is a small possibility that this block will not be cached:
 232          *   1 - if level > 1 and there are no children with level <= 1
 233          *   2 - if we didn't get a dirty hold (because this block had just
 234          *       finished being written -- and so had no holds), and then this
 235          *       block got evicted before we got here.
 236          */
 237         if (db->db_state != DB_CACHED)
 238                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 239
 240         dbuf_release_bp(db);
 241         bp = (blkptr_t *)db->db.db_data;
 242
 243         DB_DNODE_ENTER(db);
 244         dn = DB_DNODE(db);
 245         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 246         shift = (db->db_level - 1) * epbs;
 247         dbstart = db->db_blkid << epbs;
 248         start = blkid >> shift;
 249         if (dbstart < start) {
 250                 bp += start - dbstart;
 251                 all = FALSE;
 252         } else {
 253                 start = dbstart;
 254         }
 255         dbend = ((db->db_blkid + 1) << epbs) - 1;
 256         end = (blkid + nblks - 1) >> shift;
 257         if (dbend <= end)
 258                 end = dbend;
 259         else if (all)
 260                 all = trunc;
 261         ASSERT3U(start, <=, end);
 262
 263         if (db->db_level == 1) {
 264                 FREE_VERIFY(db, start, end, tx);
 265                 blocks_freed = free_blocks(dn, bp, end-start+1, tx);
 266                 arc_buf_freeze(db->db_buf);
 267                 ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
 268                 DB_DNODE_EXIT(db);
 269                 return (all ? ALL : blocks_freed);
 270         }
 271
 272         for (i = start; i <= end; i++, bp++) {
 273                 if (BP_IS_HOLE(bp))
 274                         continue;
 275                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 276                 err = dbuf_hold_impl(dn, db->db_level-1, i, TRUE, FTAG, &subdb);
 277                 ASSERT0(err);
 278                 rw_exit(&dn->dn_struct_rwlock);
 279
 280                 if (free_children(subdb, blkid, nblks, trunc, tx) == ALL) {
 281                         ASSERT3P(subdb->db_blkptr, ==, bp);
 282                         blocks_freed += free_blocks(dn, bp, 1, tx);
 283                 } else {
 284                         all = FALSE;
 285                 }
 286                 dbuf_rele(subdb, FTAG);
 287         }
 288         DB_DNODE_EXIT(db);
 289         arc_buf_freeze(db->db_buf);
 290 #ifdef ZFS_DEBUG
 291         bp -= (end-start)+1;
 292         for (i = start; i <= end; i++, bp++) {
 293                 if (i == start && blkid != 0)
 294                         continue;
 295                 else if (i == end && !trunc)
 296                         continue;
 297                 ASSERT0(bp->blk_birth);
 298         }
 299 #endif
 300         ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
 301         return (all ? ALL : blocks_freed);
 302 }
 303
 304 /*
 305  * Traverse the indicated range of the provided file
 306  * and "free" all the blocks contained there.
 307  */
 308 static void
 309 dnode_sync_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
 310 {
 311         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 312         dmu_buf_impl_t *db;
 313         int trunc, start, end, shift, i, err;
 314         int dnlevel = dn->dn_phys->dn_nlevels;
 315
 316         if (blkid > dn->dn_phys->dn_maxblkid)
 317                 return;
 318
 319         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 320         trunc = blkid + nblks > dn->dn_phys->dn_maxblkid;
 321         if (trunc)
 322                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 323
 324         /* There are no indirect blocks in the object */
 325         if (dnlevel == 1) {
 326                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 327                         /* this range was never made persistent */
 328                         return;
 329                 }
 330                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 331                 (void) free_blocks(dn, bp + blkid, nblks, tx);
 332                 if (trunc) {
 333                         uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
 334                             (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 335                         dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
 336                         ASSERT(off < dn->dn_phys->dn_maxblkid ||
 337                             dn->dn_phys->dn_maxblkid == 0 ||
 338                             dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 339                 }
 340                 return;
 341         }
 342
 343         shift = (dnlevel - 1) * (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 344         start = blkid >> shift;
 345         ASSERT(start < dn->dn_phys->dn_nblkptr);
 346         end = (blkid + nblks - 1) >> shift;
 347         bp += start;
 348         for (i = start; i <= end; i++, bp++) {
 349                 if (BP_IS_HOLE(bp))
 350                         continue;
 351                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 352                 err = dbuf_hold_impl(dn, dnlevel-1, i, TRUE, FTAG, &db);
 353                 ASSERT0(err);
 354                 rw_exit(&dn->dn_struct_rwlock);
 355
 356                 if (free_children(db, blkid, nblks, trunc, tx) == ALL) {
 357                         ASSERT3P(db->db_blkptr, ==, bp);
 358                         (void) free_blocks(dn, bp, 1, tx);
 359                 }
 360                 dbuf_rele(db, FTAG);
 361         }
 362         if (trunc) {
 363                 uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
 364                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 365                 dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
 366                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 367                     dn->dn_phys->dn_maxblkid == 0 ||
 368                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 369         }
 370 }
 371
 372 /*
 373  * Try to kick all the dnode's dbufs out of the cache...
 374  */
 375 void
 376 dnode_evict_dbufs(dnode_t *dn)
 377 {
 378         int progress;
 379         int pass = 0;
 380
 381         do {
 382                 dmu_buf_impl_t *db, marker;
 383                 int evicting = FALSE;
 384
 385                 progress = FALSE;
 386                 mutex_enter(&dn->dn_dbufs_mtx);
 387                 list_insert_tail(&dn->dn_dbufs, &marker);
 388                 db = list_head(&dn->dn_dbufs);
 389                 for (; db != &marker; db = list_head(&dn->dn_dbufs)) {
 390                         list_remove(&dn->dn_dbufs, db);
 391                         list_insert_tail(&dn->dn_dbufs, db);
 392 #ifdef  DEBUG
 393                         DB_DNODE_ENTER(db);
 394                         ASSERT3P(DB_DNODE(db), ==, dn);
 395                         DB_DNODE_EXIT(db);
 396 #endif  /* DEBUG */
 397
 398                         mutex_enter(&db->db_mtx);
 399                         if (db->db_state == DB_EVICTING) {
 400                                 progress = TRUE;
 401                                 evicting = TRUE;
 402                                 mutex_exit(&db->db_mtx);
 403                         } else if (refcount_is_zero(&db->db_holds)) {
 404                                 progress = TRUE;
 405                                 dbuf_clear(db); /* exits db_mtx for us */
 406                         } else {
 407                                 mutex_exit(&db->db_mtx);
 408                         }
 409
 410                 }
 411                 list_remove(&dn->dn_dbufs, &marker);
 412                 /*
 413                  * NB: we need to drop dn_dbufs_mtx between passes so
 414                  * that any DB_EVICTING dbufs can make progress.
 415                  * Ideally, we would have some cv we could wait on, but
 416                  * since we don't, just wait a bit to give the other
 417                  * thread a chance to run.
 418                  */
 419                 mutex_exit(&dn->dn_dbufs_mtx);
 420                 if (evicting)
 421                         delay(1);
 422                 pass++;
 423                 ASSERT(pass < 100); /* sanity check */
 424         } while (progress);
 425
 426         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 427         if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
 428                 mutex_enter(&dn->dn_bonus->db_mtx);
 429                 dbuf_evict(dn->dn_bonus);
 430                 dn->dn_bonus = NULL;
 431         }
 432         rw_exit(&dn->dn_struct_rwlock);
 433 }
 434
 435 static void
 436 dnode_undirty_dbufs(list_t *list)
 437 {
 438         dbuf_dirty_record_t *dr;
 439
 440         while (dr = list_head(list)) {
 441                 dmu_buf_impl_t *db = dr->dr_dbuf;
 442                 uint64_t txg = dr->dr_txg;
 443
 444                 if (db->db_level != 0)
 445                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 446
 447                 mutex_enter(&db->db_mtx);
 448                 /* XXX - use dbuf_undirty()? */
 449                 list_remove(list, dr);
 450                 ASSERT(db->db_last_dirty == dr);
 451                 db->db_last_dirty = NULL;
 452                 db->db_dirtycnt -= 1;
 453                 if (db->db_level == 0) {
 454                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 455                             dr->dt.dl.dr_data == db->db_buf);
 456                         dbuf_unoverride(dr);
 457                 } else {
 458                         list_destroy(&dr->dt.di.dr_children);
 459                         mutex_destroy(&dr->dt.di.dr_mtx);
 460                 }
 461                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 462                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
 463         }
 464 }
 465
 466 static void
 467 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 468 {
 469         int txgoff = tx->tx_txg & TXG_MASK;
 470
 471         ASSERT(dmu_tx_is_syncing(tx));
 472
 473         /*
 474          * Our contents should have been freed in dnode_sync() by the
 475          * free range record inserted by the caller of dnode_free().
 476          */
 477         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 478         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 479
 480         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 481         dnode_evict_dbufs(dn);
 482         ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
 483         ASSERT3P(dn->dn_bonus, ==, NULL);
 484
 485         /*
 486          * XXX - It would be nice to assert this, but we may still
 487          * have residual holds from async evictions from the arc...
 488          *
 489          * zfs_obj_to_path() also depends on this being
 490          * commented out.
 491          *
 492          * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
 493          */
 494
 495         /* Undirty next bits */
 496         dn->dn_next_nlevels[txgoff] = 0;
 497         dn->dn_next_indblkshift[txgoff] = 0;
 498         dn->dn_next_blksz[txgoff] = 0;
 499
 500         /* ASSERT(blkptrs are zero); */
 501         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 502         ASSERT(dn->dn_type != DMU_OT_NONE);
 503
 504         ASSERT(dn->dn_free_txg > 0);
 505         if (dn->dn_allocated_txg != dn->dn_free_txg)
 506                 dbuf_will_dirty(dn->dn_dbuf, tx);
 507         bzero(dn->dn_phys, sizeof (dnode_phys_t));
 508
 509         mutex_enter(&dn->dn_mtx);
 510         dn->dn_type = DMU_OT_NONE;
 511         dn->dn_maxblkid = 0;
 512         dn->dn_allocated_txg = 0;
 513         dn->dn_free_txg = 0;
 514         dn->dn_have_spill = B_FALSE;
 515         mutex_exit(&dn->dn_mtx);
 516
 517         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 518
 519         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 520         /*
 521          * Now that we've released our hold, the dnode may
 522          * be evicted, so we musn't access it.
 523          */
 524 }
 525
 526 /*
 527  * Write out the dnode's dirty buffers.
 528  */
 529 void
 530 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 531 {
 532         free_range_t *rp;
 533         dnode_phys_t *dnp = dn->dn_phys;
 534         int txgoff = tx->tx_txg & TXG_MASK;
 535         list_t *list = &dn->dn_dirty_records[txgoff];
 536         static const dnode_phys_t zerodn = { 0 };
 537         boolean_t kill_spill = B_FALSE;
 538
 539         ASSERT(dmu_tx_is_syncing(tx));
 540         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 541         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 542             bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
 543         DNODE_VERIFY(dn);
 544
 545         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 546
 547         if (dmu_objset_userused_enabled(dn->dn_objset) &&
 548             !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 549                 mutex_enter(&dn->dn_mtx);
 550                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 551                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 552                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 553                 mutex_exit(&dn->dn_mtx);
 554                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 555         } else {
 556                 /* Once we account for it, we should always account for it. */
 557                 ASSERT(!(dn->dn_phys->dn_flags &
 558                     DNODE_FLAG_USERUSED_ACCOUNTED));
 559         }
 560
 561         mutex_enter(&dn->dn_mtx);
 562         if (dn->dn_allocated_txg == tx->tx_txg) {
 563                 /* The dnode is newly allocated or reallocated */
 564                 if (dnp->dn_type == DMU_OT_NONE) {
 565                         /* this is a first alloc, not a realloc */
 566                         dnp->dn_nlevels = 1;
 567                         dnp->dn_nblkptr = dn->dn_nblkptr;
 568                 }
 569
 570                 dnp->dn_type = dn->dn_type;
 571                 dnp->dn_bonustype = dn->dn_bonustype;
 572                 dnp->dn_bonuslen = dn->dn_bonuslen;
 573         }
 574
 575         ASSERT(dnp->dn_nlevels > 1 ||
 576             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 577             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 578             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 579
 580         if (dn->dn_next_blksz[txgoff]) {
 581                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 582                     SPA_MINBLOCKSIZE) == 0);
 583                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 584                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 585                     avl_last(&dn->dn_ranges[txgoff]) ||
 586                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 587                     dnp->dn_datablkszsec);
 588                 dnp->dn_datablkszsec =
 589                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 590                 dn->dn_next_blksz[txgoff] = 0;
 591         }
 592
 593         if (dn->dn_next_bonuslen[txgoff]) {
 594                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 595                         dnp->dn_bonuslen = 0;
 596                 else
 597                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 598                 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
 599                 dn->dn_next_bonuslen[txgoff] = 0;
 600         }
 601
 602         if (dn->dn_next_bonustype[txgoff]) {
 603                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 604                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 605                 dn->dn_next_bonustype[txgoff] = 0;
 606         }
 607
 608         /*
 609          * We will either remove a spill block when a file is being removed
 610          * or we have been asked to remove it.
 611          */
 612         if (dn->dn_rm_spillblk[txgoff] ||
 613             ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) &&
 614             dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg)) {
 615                 if ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
 616                         kill_spill = B_TRUE;
 617                 dn->dn_rm_spillblk[txgoff] = 0;
 618         }
 619
 620         if (dn->dn_next_indblkshift[txgoff]) {
 621                 ASSERT(dnp->dn_nlevels == 1);
 622                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 623                 dn->dn_next_indblkshift[txgoff] = 0;
 624         }
 625
 626         /*
 627          * Just take the live (open-context) values for checksum and compress.
 628          * Strictly speaking it's a future leak, but nothing bad happens if we
 629          * start using the new checksum or compress algorithm a little early.
 630          */
 631         dnp->dn_checksum = dn->dn_checksum;
 632         dnp->dn_compress = dn->dn_compress;
 633
 634         mutex_exit(&dn->dn_mtx);
 635
 636         if (kill_spill) {
 637                 (void) free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
 638                 mutex_enter(&dn->dn_mtx);
 639                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 640                 mutex_exit(&dn->dn_mtx);
 641         }
 642
 643         /* process all the "freed" ranges in the file */
 644         while (rp = avl_last(&dn->dn_ranges[txgoff])) {
 645                 dnode_sync_free_range(dn, rp->fr_blkid, rp->fr_nblks, tx);
 646                 /* grab the mutex so we don't race with dnode_block_freed() */
 647                 mutex_enter(&dn->dn_mtx);
 648                 avl_remove(&dn->dn_ranges[txgoff], rp);
 649                 mutex_exit(&dn->dn_mtx);
 650                 kmem_free(rp, sizeof (free_range_t));
 651         }
 652
 653         if (dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg) {
 654                 dnode_sync_free(dn, tx);
 655                 return;
 656         }
 657
 658         if (dn->dn_next_nblkptr[txgoff]) {
 659                 /* this should only happen on a realloc */
 660                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 661                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 662                         /* zero the new blkptrs we are gaining */
 663                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 664                             sizeof (blkptr_t) *
 665                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 666 #ifdef ZFS_DEBUG
 667                 } else {
 668                         int i;
 669                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 670                         /* the blkptrs we are losing better be unallocated */
 671                         for (i = dn->dn_next_nblkptr[txgoff];
 672                             i < dnp->dn_nblkptr; i++)
 673                                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 674 #endif
 675                 }
 676                 mutex_enter(&dn->dn_mtx);
 677                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 678                 dn->dn_next_nblkptr[txgoff] = 0;
 679                 mutex_exit(&dn->dn_mtx);
 680         }
 681
 682         if (dn->dn_next_nlevels[txgoff]) {
 683                 dnode_increase_indirection(dn, tx);
 684                 dn->dn_next_nlevels[txgoff] = 0;
 685         }
 686
 687         dbuf_sync_list(list, tx);
 688
 689         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 690                 ASSERT3P(list_head(list), ==, NULL);
 691                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 692         }
 693
 694         /*
 695          * Although we have dropped our reference to the dnode, it
 696          * can't be evicted until its written, and we haven't yet
 697          * initiated the IO for the dnode's dbuf.
 698          */
 699 }