granicus.if.org Git - zfs/blob - module/zfs/dnode_sync.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21
  22 /*
  23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
  25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  26  */
  27
  28 #include <sys/zfs_context.h>
  29 #include <sys/dbuf.h>
  30 #include <sys/dnode.h>
  31 #include <sys/dmu.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dmu_objset.h>
  34 #include <sys/dmu_recv.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 /*
 234  * We don't usually free the indirect blocks here.  If in one txg we have a
 235  * free_range and a write to the same indirect block, it's important that we
 236  * preserve the hole's birth times. Therefore, we don't free any any indirect
 237  * blocks in free_children().  If an indirect block happens to turn into all
 238  * holes, it will be freed by dbuf_write_children_ready, which happens at a
 239  * point in the syncing process where we know for certain the contents of the
 240  * indirect block.
 241  *
 242  * However, if we're freeing a dnode, its space accounting must go to zero
 243  * before we actually try to free the dnode, or we will trip an assertion. In
 244  * addition, we know the case described above cannot occur, because the dnode is
 245  * being freed.  Therefore, we free the indirect blocks immediately in that
 246  * case.
 247  */
 248 static void
 249 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 250     boolean_t free_indirects, dmu_tx_t *tx)
 251 {
 252         dnode_t *dn;
 253         blkptr_t *bp;
 254         dmu_buf_impl_t *subdb;
 255         uint64_t start, end, dbstart, dbend;
 256         unsigned int epbs, shift, i;
 257
 258         /*
 259          * There is a small possibility that this block will not be cached:
 260          *   1 - if level > 1 and there are no children with level <= 1
 261          *   2 - if this block was evicted since we read it from
 262          *       dmu_tx_hold_free().
 263          */
 264         if (db->db_state != DB_CACHED)
 265                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
 266
 267         /*
 268          * If we modify this indirect block, and we are not freeing the
 269          * dnode (!free_indirects), then this indirect block needs to get
 270          * written to disk by dbuf_write().  If it is dirty, we know it will
 271          * be written (otherwise, we would have incorrect on-disk state
 272          * because the space would be freed but still referenced by the BP
 273          * in this indirect block).  Therefore we VERIFY that it is
 274          * dirty.
 275          *
 276          * Our VERIFY covers some cases that do not actually have to be
 277          * dirty, but the open-context code happens to dirty.  E.g. if the
 278          * blocks we are freeing are all holes, because in that case, we
 279          * are only freeing part of this indirect block, so it is an
 280          * ancestor of the first or last block to be freed.  The first and
 281          * last L1 indirect blocks are always dirtied by dnode_free_range().
 282          */
 283         VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
 284
 285         dbuf_release_bp(db);
 286         bp = db->db.db_data;
 287
 288         DB_DNODE_ENTER(db);
 289         dn = DB_DNODE(db);
 290         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
 291         ASSERT3U(epbs, <, 31);
 292         shift = (db->db_level - 1) * epbs;
 293         dbstart = db->db_blkid << epbs;
 294         start = blkid >> shift;
 295         if (dbstart < start) {
 296                 bp += start - dbstart;
 297         } else {
 298                 start = dbstart;
 299         }
 300         dbend = ((db->db_blkid + 1) << epbs) - 1;
 301         end = (blkid + nblks - 1) >> shift;
 302         if (dbend <= end)
 303                 end = dbend;
 304
 305         ASSERT3U(start, <=, end);
 306
 307         if (db->db_level == 1) {
 308                 FREE_VERIFY(db, start, end, tx);
 309                 free_blocks(dn, bp, end-start+1, tx);
 310         } else {
 311                 for (uint64_t id = start; id <= end; id++, bp++) {
 312                         if (BP_IS_HOLE(bp))
 313                                 continue;
 314                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 315                         VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
 316                             id, TRUE, FALSE, FTAG, &subdb));
 317                         rw_exit(&dn->dn_struct_rwlock);
 318                         ASSERT3P(bp, ==, subdb->db_blkptr);
 319
 320                         free_children(subdb, blkid, nblks, free_indirects, tx);
 321                         dbuf_rele(subdb, FTAG);
 322                 }
 323         }
 324
 325         if (free_indirects) {
 326                 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
 327                         ASSERT(BP_IS_HOLE(bp));
 328                 bzero(db->db.db_data, db->db.db_size);
 329                 free_blocks(dn, db->db_blkptr, 1, tx);
 330         }
 331
 332         DB_DNODE_EXIT(db);
 333         arc_buf_freeze(db->db_buf);
 334 }
 335
 336 /*
 337  * Traverse the indicated range of the provided file
 338  * and "free" all the blocks contained there.
 339  */
 340 static void
 341 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
 342     boolean_t free_indirects, dmu_tx_t *tx)
 343 {
 344         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 345         int dnlevel = dn->dn_phys->dn_nlevels;
 346         boolean_t trunc = B_FALSE;
 347
 348         if (blkid > dn->dn_phys->dn_maxblkid)
 349                 return;
 350
 351         ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
 352         if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
 353                 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
 354                 trunc = B_TRUE;
 355         }
 356
 357         /* There are no indirect blocks in the object */
 358         if (dnlevel == 1) {
 359                 if (blkid >= dn->dn_phys->dn_nblkptr) {
 360                         /* this range was never made persistent */
 361                         return;
 362                 }
 363                 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
 364                 free_blocks(dn, bp + blkid, nblks, tx);
 365         } else {
 366                 int shift = (dnlevel - 1) *
 367                     (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
 368                 int start = blkid >> shift;
 369                 int end = (blkid + nblks - 1) >> shift;
 370                 dmu_buf_impl_t *db;
 371
 372                 ASSERT(start < dn->dn_phys->dn_nblkptr);
 373                 bp += start;
 374                 for (int i = start; i <= end; i++, bp++) {
 375                         if (BP_IS_HOLE(bp))
 376                                 continue;
 377                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 378                         VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
 379                             TRUE, FALSE, FTAG, &db));
 380                         rw_exit(&dn->dn_struct_rwlock);
 381
 382                         free_children(db, blkid, nblks, free_indirects, tx);
 383                         dbuf_rele(db, FTAG);
 384                 }
 385         }
 386
 387         /*
 388          * Do not truncate the maxblkid if we are performing a raw
 389          * receive. The raw receive sets the mablkid manually and
 390          * must not be overriden.
 391          */
 392         if (trunc && !dn->dn_objset->os_raw_receive) {
 393                 ASSERTV(uint64_t off);
 394                 dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
 395
 396                 ASSERTV(off = (dn->dn_phys->dn_maxblkid + 1) *
 397                     (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT));
 398                 ASSERT(off < dn->dn_phys->dn_maxblkid ||
 399                     dn->dn_phys->dn_maxblkid == 0 ||
 400                     dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
 401         }
 402 }
 403
 404 typedef struct dnode_sync_free_range_arg {
 405         dnode_t *dsfra_dnode;
 406         dmu_tx_t *dsfra_tx;
 407         boolean_t dsfra_free_indirects;
 408 } dnode_sync_free_range_arg_t;
 409
 410 static void
 411 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
 412 {
 413         dnode_sync_free_range_arg_t *dsfra = arg;
 414         dnode_t *dn = dsfra->dsfra_dnode;
 415
 416         mutex_exit(&dn->dn_mtx);
 417         dnode_sync_free_range_impl(dn, blkid, nblks,
 418             dsfra->dsfra_free_indirects, dsfra->dsfra_tx);
 419         mutex_enter(&dn->dn_mtx);
 420 }
 421
 422 /*
 423  * Try to kick all the dnode's dbufs out of the cache...
 424  */
 425 void
 426 dnode_evict_dbufs(dnode_t *dn)
 427 {
 428         dmu_buf_impl_t *db_marker;
 429         dmu_buf_impl_t *db, *db_next;
 430
 431         db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
 432
 433         mutex_enter(&dn->dn_dbufs_mtx);
 434         for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
 435
 436 #ifdef  DEBUG
 437                 DB_DNODE_ENTER(db);
 438                 ASSERT3P(DB_DNODE(db), ==, dn);
 439                 DB_DNODE_EXIT(db);
 440 #endif  /* DEBUG */
 441
 442                 mutex_enter(&db->db_mtx);
 443                 if (db->db_state != DB_EVICTING &&
 444                     zfs_refcount_is_zero(&db->db_holds)) {
 445                         db_marker->db_level = db->db_level;
 446                         db_marker->db_blkid = db->db_blkid;
 447                         db_marker->db_state = DB_SEARCH;
 448                         avl_insert_here(&dn->dn_dbufs, db_marker, db,
 449                             AVL_BEFORE);
 450
 451                         /*
 452                          * We need to use the "marker" dbuf rather than
 453                          * simply getting the next dbuf, because
 454                          * dbuf_destroy() may actually remove multiple dbufs.
 455                          * It can call itself recursively on the parent dbuf,
 456                          * which may also be removed from dn_dbufs.  The code
 457                          * flow would look like:
 458                          *
 459                          * dbuf_destroy():
 460                          *   dnode_rele_and_unlock(parent_dbuf, evicting=TRUE):
 461                          *      if (!cacheable || pending_evict)
 462                          *        dbuf_destroy()
 463                          */
 464                         dbuf_destroy(db);
 465
 466                         db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
 467                         avl_remove(&dn->dn_dbufs, db_marker);
 468                 } else {
 469                         db->db_pending_evict = TRUE;
 470                         mutex_exit(&db->db_mtx);
 471                         db_next = AVL_NEXT(&dn->dn_dbufs, db);
 472                 }
 473         }
 474         mutex_exit(&dn->dn_dbufs_mtx);
 475
 476         kmem_free(db_marker, sizeof (dmu_buf_impl_t));
 477
 478         dnode_evict_bonus(dn);
 479 }
 480
 481 void
 482 dnode_evict_bonus(dnode_t *dn)
 483 {
 484         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 485         if (dn->dn_bonus != NULL) {
 486                 if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
 487                         mutex_enter(&dn->dn_bonus->db_mtx);
 488                         dbuf_destroy(dn->dn_bonus);
 489                         dn->dn_bonus = NULL;
 490                 } else {
 491                         dn->dn_bonus->db_pending_evict = TRUE;
 492                 }
 493         }
 494         rw_exit(&dn->dn_struct_rwlock);
 495 }
 496
 497 static void
 498 dnode_undirty_dbufs(list_t *list)
 499 {
 500         dbuf_dirty_record_t *dr;
 501
 502         while ((dr = list_head(list))) {
 503                 dmu_buf_impl_t *db = dr->dr_dbuf;
 504                 uint64_t txg = dr->dr_txg;
 505
 506                 if (db->db_level != 0)
 507                         dnode_undirty_dbufs(&dr->dt.di.dr_children);
 508
 509                 mutex_enter(&db->db_mtx);
 510                 /* XXX - use dbuf_undirty()? */
 511                 list_remove(list, dr);
 512                 ASSERT(db->db_last_dirty == dr);
 513                 db->db_last_dirty = NULL;
 514                 db->db_dirtycnt -= 1;
 515                 if (db->db_level == 0) {
 516                         ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 517                             dr->dt.dl.dr_data == db->db_buf);
 518                         dbuf_unoverride(dr);
 519                 } else {
 520                         mutex_destroy(&dr->dt.di.dr_mtx);
 521                         list_destroy(&dr->dt.di.dr_children);
 522                 }
 523                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
 524                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
 525         }
 526 }
 527
 528 static void
 529 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
 530 {
 531         int txgoff = tx->tx_txg & TXG_MASK;
 532
 533         ASSERT(dmu_tx_is_syncing(tx));
 534
 535         /*
 536          * Our contents should have been freed in dnode_sync() by the
 537          * free range record inserted by the caller of dnode_free().
 538          */
 539         ASSERT0(DN_USED_BYTES(dn->dn_phys));
 540         ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
 541
 542         dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
 543         dnode_evict_dbufs(dn);
 544
 545         /*
 546          * XXX - It would be nice to assert this, but we may still
 547          * have residual holds from async evictions from the arc...
 548          *
 549          * zfs_obj_to_path() also depends on this being
 550          * commented out.
 551          *
 552          * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
 553          */
 554
 555         /* Undirty next bits */
 556         dn->dn_next_nlevels[txgoff] = 0;
 557         dn->dn_next_indblkshift[txgoff] = 0;
 558         dn->dn_next_blksz[txgoff] = 0;
 559         dn->dn_next_maxblkid[txgoff] = 0;
 560
 561         /* ASSERT(blkptrs are zero); */
 562         ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
 563         ASSERT(dn->dn_type != DMU_OT_NONE);
 564
 565         ASSERT(dn->dn_free_txg > 0);
 566         if (dn->dn_allocated_txg != dn->dn_free_txg)
 567                 dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
 568         bzero(dn->dn_phys, sizeof (dnode_phys_t) * dn->dn_num_slots);
 569         dnode_free_interior_slots(dn);
 570
 571         mutex_enter(&dn->dn_mtx);
 572         dn->dn_type = DMU_OT_NONE;
 573         dn->dn_maxblkid = 0;
 574         dn->dn_allocated_txg = 0;
 575         dn->dn_free_txg = 0;
 576         dn->dn_have_spill = B_FALSE;
 577         dn->dn_num_slots = 1;
 578         mutex_exit(&dn->dn_mtx);
 579
 580         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
 581
 582         dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 583         /*
 584          * Now that we've released our hold, the dnode may
 585          * be evicted, so we mustn't access it.
 586          */
 587 }
 588
 589 /*
 590  * Write out the dnode's dirty buffers.
 591  */
 592 void
 593 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
 594 {
 595         objset_t *os = dn->dn_objset;
 596         dnode_phys_t *dnp = dn->dn_phys;
 597         int txgoff = tx->tx_txg & TXG_MASK;
 598         list_t *list = &dn->dn_dirty_records[txgoff];
 599         ASSERTV(static const dnode_phys_t zerodn = { 0 });
 600         boolean_t kill_spill = B_FALSE;
 601
 602         ASSERT(dmu_tx_is_syncing(tx));
 603         ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
 604         ASSERT(dnp->dn_type != DMU_OT_NONE ||
 605             bcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
 606         DNODE_VERIFY(dn);
 607
 608         ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
 609
 610         /*
 611          * Do user accounting if it is enabled and this is not
 612          * an encrypted receive.
 613          */
 614         if (dmu_objset_userused_enabled(os) &&
 615             !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
 616             (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
 617                 mutex_enter(&dn->dn_mtx);
 618                 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
 619                 dn->dn_oldflags = dn->dn_phys->dn_flags;
 620                 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
 621                 if (dmu_objset_userobjused_enabled(dn->dn_objset))
 622                         dn->dn_phys->dn_flags |=
 623                             DNODE_FLAG_USEROBJUSED_ACCOUNTED;
 624                 mutex_exit(&dn->dn_mtx);
 625                 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
 626         } else {
 627                 /* Once we account for it, we should always account for it */
 628                 ASSERT(!(dn->dn_phys->dn_flags &
 629                     DNODE_FLAG_USERUSED_ACCOUNTED));
 630                 ASSERT(!(dn->dn_phys->dn_flags &
 631                     DNODE_FLAG_USEROBJUSED_ACCOUNTED));
 632         }
 633
 634         mutex_enter(&dn->dn_mtx);
 635         if (dn->dn_allocated_txg == tx->tx_txg) {
 636                 /* The dnode is newly allocated or reallocated */
 637                 if (dnp->dn_type == DMU_OT_NONE) {
 638                         /* this is a first alloc, not a realloc */
 639                         dnp->dn_nlevels = 1;
 640                         dnp->dn_nblkptr = dn->dn_nblkptr;
 641                 }
 642
 643                 dnp->dn_type = dn->dn_type;
 644                 dnp->dn_bonustype = dn->dn_bonustype;
 645                 dnp->dn_bonuslen = dn->dn_bonuslen;
 646         }
 647
 648         dnp->dn_extra_slots = dn->dn_num_slots - 1;
 649
 650         ASSERT(dnp->dn_nlevels > 1 ||
 651             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 652             BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
 653             BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
 654             dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 655         ASSERT(dnp->dn_nlevels < 2 ||
 656             BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 657             BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
 658
 659         if (dn->dn_next_type[txgoff] != 0) {
 660                 dnp->dn_type = dn->dn_type;
 661                 dn->dn_next_type[txgoff] = 0;
 662         }
 663
 664         if (dn->dn_next_blksz[txgoff] != 0) {
 665                 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
 666                     SPA_MINBLOCKSIZE) == 0);
 667                 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
 668                     dn->dn_maxblkid == 0 || list_head(list) != NULL ||
 669                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
 670                     dnp->dn_datablkszsec ||
 671                     !range_tree_is_empty(dn->dn_free_ranges[txgoff]));
 672                 dnp->dn_datablkszsec =
 673                     dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
 674                 dn->dn_next_blksz[txgoff] = 0;
 675         }
 676
 677         if (dn->dn_next_bonuslen[txgoff] != 0) {
 678                 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
 679                         dnp->dn_bonuslen = 0;
 680                 else
 681                         dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
 682                 ASSERT(dnp->dn_bonuslen <=
 683                     DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
 684                 dn->dn_next_bonuslen[txgoff] = 0;
 685         }
 686
 687         if (dn->dn_next_bonustype[txgoff] != 0) {
 688                 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
 689                 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
 690                 dn->dn_next_bonustype[txgoff] = 0;
 691         }
 692
 693         boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
 694             dn->dn_free_txg <= tx->tx_txg;
 695
 696         /*
 697          * Remove the spill block if we have been explicitly asked to
 698          * remove it, or if the object is being removed.
 699          */
 700         if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
 701                 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 702                         kill_spill = B_TRUE;
 703                 dn->dn_rm_spillblk[txgoff] = 0;
 704         }
 705
 706         if (dn->dn_next_indblkshift[txgoff] != 0) {
 707                 ASSERT(dnp->dn_nlevels == 1);
 708                 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
 709                 dn->dn_next_indblkshift[txgoff] = 0;
 710         }
 711
 712         /*
 713          * Just take the live (open-context) values for checksum and compress.
 714          * Strictly speaking it's a future leak, but nothing bad happens if we
 715          * start using the new checksum or compress algorithm a little early.
 716          */
 717         dnp->dn_checksum = dn->dn_checksum;
 718         dnp->dn_compress = dn->dn_compress;
 719
 720         mutex_exit(&dn->dn_mtx);
 721
 722         if (kill_spill) {
 723                 free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
 724                 mutex_enter(&dn->dn_mtx);
 725                 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
 726                 mutex_exit(&dn->dn_mtx);
 727         }
 728
 729         /* process all the "freed" ranges in the file */
 730         if (dn->dn_free_ranges[txgoff] != NULL) {
 731                 dnode_sync_free_range_arg_t dsfra;
 732                 dsfra.dsfra_dnode = dn;
 733                 dsfra.dsfra_tx = tx;
 734                 dsfra.dsfra_free_indirects = freeing_dnode;
 735                 if (freeing_dnode) {
 736                         ASSERT(range_tree_contains(dn->dn_free_ranges[txgoff],
 737                             0, dn->dn_maxblkid + 1));
 738                 }
 739                 mutex_enter(&dn->dn_mtx);
 740                 range_tree_vacate(dn->dn_free_ranges[txgoff],
 741                     dnode_sync_free_range, &dsfra);
 742                 range_tree_destroy(dn->dn_free_ranges[txgoff]);
 743                 dn->dn_free_ranges[txgoff] = NULL;
 744                 mutex_exit(&dn->dn_mtx);
 745         }
 746
 747         if (freeing_dnode) {
 748                 dn->dn_objset->os_freed_dnodes++;
 749                 dnode_sync_free(dn, tx);
 750                 return;
 751         }
 752
 753         if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
 754                 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 755                 mutex_enter(&ds->ds_lock);
 756                 ds->ds_feature_activation[SPA_FEATURE_LARGE_DNODE] =
 757                     (void *)B_TRUE;
 758                 mutex_exit(&ds->ds_lock);
 759         }
 760
 761         if (dn->dn_next_nlevels[txgoff]) {
 762                 dnode_increase_indirection(dn, tx);
 763                 dn->dn_next_nlevels[txgoff] = 0;
 764         }
 765
 766         /*
 767          * This must be done after dnode_sync_free_range()
 768          * and dnode_increase_indirection().
 769          */
 770         if (dn->dn_next_maxblkid[txgoff]) {
 771                 mutex_enter(&dn->dn_mtx);
 772                 dnp->dn_maxblkid = dn->dn_next_maxblkid[txgoff];
 773                 dn->dn_next_maxblkid[txgoff] = 0;
 774                 mutex_exit(&dn->dn_mtx);
 775         }
 776
 777         if (dn->dn_next_nblkptr[txgoff]) {
 778                 /* this should only happen on a realloc */
 779                 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
 780                 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
 781                         /* zero the new blkptrs we are gaining */
 782                         bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
 783                             sizeof (blkptr_t) *
 784                             (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
 785 #ifdef ZFS_DEBUG
 786                 } else {
 787                         int i;
 788                         ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
 789                         /* the blkptrs we are losing better be unallocated */
 790                         for (i = 0; i < dnp->dn_nblkptr; i++) {
 791                                 if (i >= dn->dn_next_nblkptr[txgoff])
 792                                         ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
 793                         }
 794 #endif
 795                 }
 796                 mutex_enter(&dn->dn_mtx);
 797                 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
 798                 dn->dn_next_nblkptr[txgoff] = 0;
 799                 mutex_exit(&dn->dn_mtx);
 800         }
 801
 802         dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
 803
 804         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 805                 ASSERT3P(list_head(list), ==, NULL);
 806                 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
 807         }
 808
 809         /*
 810          * Although we have dropped our reference to the dnode, it
 811          * can't be evicted until its written, and we haven't yet
 812          * initiated the IO for the dnode's dbuf.
 813          */
 814 }