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