1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
  24  */
  25 
  26 #include <sys/zfs_context.h>
  27 #include <sys/dbuf.h>
  28 #include <sys/dnode.h>
  29 #include <sys/dmu.h>
  30 #include <sys/dmu_impl.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/dmu_objset.h>
  33 #include <sys/dsl_dir.h>
  34 #include <sys/dsl_dataset.h>
  35 #include <sys/spa.h>
  36 #include <sys/zio.h>
  37 #include <sys/dmu_zfetch.h>
  38 #include <sys/range_tree.h>
  39 
  40 static kmem_cache_t *dnode_cache;
  41 /*
  42  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
  43  * turned on when DEBUG is also defined.
  44  */
  45 #ifdef  DEBUG
  46 #define DNODE_STATS
  47 #endif  /* DEBUG */
  48 
  49 #ifdef  DNODE_STATS
  50 #define DNODE_STAT_ADD(stat)                    ((stat)++)
  51 #else
  52 #define DNODE_STAT_ADD(stat)                    /* nothing */
  53 #endif  /* DNODE_STATS */
  54 
  55 static dnode_phys_t dnode_phys_zero;
  56 
  57 int zfs_default_bs = SPA_MINBLOCKSHIFT;
  58 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
  59 
  60 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
  61 
  62 static int
  63 dbuf_compare(const void *x1, const void *x2)
  64 {
  65         const dmu_buf_impl_t *d1 = x1;
  66         const dmu_buf_impl_t *d2 = x2;
  67 
  68         if (d1->db_level < d2->db_level) {
  69                 return (-1);
  70         }
  71         if (d1->db_level > d2->db_level) {
  72                 return (1);
  73         }
  74 
  75         if (d1->db_blkid < d2->db_blkid) {
  76                 return (-1);
  77         }
  78         if (d1->db_blkid > d2->db_blkid) {
  79                 return (1);
  80         }
  81 
  82         if (d1->db_state < d2->db_state) {
  83                 return (-1);
  84         }
  85         if (d1->db_state > d2->db_state) {
  86                 return (1);
  87         }
  88 
  89         ASSERT3S(d1->db_state, !=, DB_SEARCH);
  90         ASSERT3S(d2->db_state, !=, DB_SEARCH);
  91 
  92         if ((uintptr_t)d1 < (uintptr_t)d2) {
  93                 return (-1);
  94         }
  95         if ((uintptr_t)d1 > (uintptr_t)d2) {
  96                 return (1);
  97         }
  98         return (0);
  99 }
 100 
 101 /* ARGSUSED */
 102 static int
 103 dnode_cons(void *arg, void *unused, int kmflag)
 104 {
 105         dnode_t *dn = arg;
 106         int i;
 107 
 108         rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
 109         mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
 110         mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
 111         cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
 112 
 113         /*
 114          * Every dbuf has a reference, and dropping a tracked reference is
 115          * O(number of references), so don't track dn_holds.
 116          */
 117         refcount_create_untracked(&dn->dn_holds);
 118         refcount_create(&dn->dn_tx_holds);
 119         list_link_init(&dn->dn_link);
 120 
 121         bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
 122         bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
 123         bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
 124         bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
 125         bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
 126         bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
 127         bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
 128 
 129         for (i = 0; i < TXG_SIZE; i++) {
 130                 list_link_init(&dn->dn_dirty_link[i]);
 131                 dn->dn_free_ranges[i] = NULL;
 132                 list_create(&dn->dn_dirty_records[i],
 133                     sizeof (dbuf_dirty_record_t),
 134                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 135         }
 136 
 137         dn->dn_allocated_txg = 0;
 138         dn->dn_free_txg = 0;
 139         dn->dn_assigned_txg = 0;
 140         dn->dn_dirtyctx = 0;
 141         dn->dn_dirtyctx_firstset = NULL;
 142         dn->dn_bonus = NULL;
 143         dn->dn_have_spill = B_FALSE;
 144         dn->dn_zio = NULL;
 145         dn->dn_oldused = 0;
 146         dn->dn_oldflags = 0;
 147         dn->dn_olduid = 0;
 148         dn->dn_oldgid = 0;
 149         dn->dn_newuid = 0;
 150         dn->dn_newgid = 0;
 151         dn->dn_id_flags = 0;
 152 
 153         dn->dn_dbufs_count = 0;
 154         dn->dn_unlisted_l0_blkid = 0;
 155         avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
 156             offsetof(dmu_buf_impl_t, db_link));
 157 
 158         dn->dn_moved = 0;
 159         return (0);
 160 }
 161 
 162 /* ARGSUSED */
 163 static void
 164 dnode_dest(void *arg, void *unused)
 165 {
 166         int i;
 167         dnode_t *dn = arg;
 168 
 169         rw_destroy(&dn->dn_struct_rwlock);
 170         mutex_destroy(&dn->dn_mtx);
 171         mutex_destroy(&dn->dn_dbufs_mtx);
 172         cv_destroy(&dn->dn_notxholds);
 173         refcount_destroy(&dn->dn_holds);
 174         refcount_destroy(&dn->dn_tx_holds);
 175         ASSERT(!list_link_active(&dn->dn_link));
 176 
 177         for (i = 0; i < TXG_SIZE; i++) {
 178                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 179                 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
 180                 list_destroy(&dn->dn_dirty_records[i]);
 181                 ASSERT0(dn->dn_next_nblkptr[i]);
 182                 ASSERT0(dn->dn_next_nlevels[i]);
 183                 ASSERT0(dn->dn_next_indblkshift[i]);
 184                 ASSERT0(dn->dn_next_bonustype[i]);
 185                 ASSERT0(dn->dn_rm_spillblk[i]);
 186                 ASSERT0(dn->dn_next_bonuslen[i]);
 187                 ASSERT0(dn->dn_next_blksz[i]);
 188         }
 189 
 190         ASSERT0(dn->dn_allocated_txg);
 191         ASSERT0(dn->dn_free_txg);
 192         ASSERT0(dn->dn_assigned_txg);
 193         ASSERT0(dn->dn_dirtyctx);
 194         ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
 195         ASSERT3P(dn->dn_bonus, ==, NULL);
 196         ASSERT(!dn->dn_have_spill);
 197         ASSERT3P(dn->dn_zio, ==, NULL);
 198         ASSERT0(dn->dn_oldused);
 199         ASSERT0(dn->dn_oldflags);
 200         ASSERT0(dn->dn_olduid);
 201         ASSERT0(dn->dn_oldgid);
 202         ASSERT0(dn->dn_newuid);
 203         ASSERT0(dn->dn_newgid);
 204         ASSERT0(dn->dn_id_flags);
 205 
 206         ASSERT0(dn->dn_dbufs_count);
 207         ASSERT0(dn->dn_unlisted_l0_blkid);
 208         avl_destroy(&dn->dn_dbufs);
 209 }
 210 
 211 void
 212 dnode_init(void)
 213 {
 214         ASSERT(dnode_cache == NULL);
 215         dnode_cache = kmem_cache_create("dnode_t",
 216             sizeof (dnode_t),
 217             0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
 218         kmem_cache_set_move(dnode_cache, dnode_move);
 219 }
 220 
 221 void
 222 dnode_fini(void)
 223 {
 224         kmem_cache_destroy(dnode_cache);
 225         dnode_cache = NULL;
 226 }
 227 
 228 
 229 #ifdef ZFS_DEBUG
 230 void
 231 dnode_verify(dnode_t *dn)
 232 {
 233         int drop_struct_lock = FALSE;
 234 
 235         ASSERT(dn->dn_phys);
 236         ASSERT(dn->dn_objset);
 237         ASSERT(dn->dn_handle->dnh_dnode == dn);
 238 
 239         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 240 
 241         if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
 242                 return;
 243 
 244         if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
 245                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 246                 drop_struct_lock = TRUE;
 247         }
 248         if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
 249                 int i;
 250                 ASSERT3U(dn->dn_indblkshift, >=, 0);
 251                 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
 252                 if (dn->dn_datablkshift) {
 253                         ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
 254                         ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
 255                         ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
 256                 }
 257                 ASSERT3U(dn->dn_nlevels, <=, 30);
 258                 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
 259                 ASSERT3U(dn->dn_nblkptr, >=, 1);
 260                 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 261                 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 262                 ASSERT3U(dn->dn_datablksz, ==,
 263                     dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 264                 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
 265                 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
 266                     dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 267                 for (i = 0; i < TXG_SIZE; i++) {
 268                         ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
 269                 }
 270         }
 271         if (dn->dn_phys->dn_type != DMU_OT_NONE)
 272                 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
 273         ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
 274         if (dn->dn_dbuf != NULL) {
 275                 ASSERT3P(dn->dn_phys, ==,
 276                     (dnode_phys_t *)dn->dn_dbuf->db.db_data +
 277                     (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
 278         }
 279         if (drop_struct_lock)
 280                 rw_exit(&dn->dn_struct_rwlock);
 281 }
 282 #endif
 283 
 284 void
 285 dnode_byteswap(dnode_phys_t *dnp)
 286 {
 287         uint64_t *buf64 = (void*)&dnp->dn_blkptr;
 288         int i;
 289 
 290         if (dnp->dn_type == DMU_OT_NONE) {
 291                 bzero(dnp, sizeof (dnode_phys_t));
 292                 return;
 293         }
 294 
 295         dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
 296         dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
 297         dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
 298         dnp->dn_used = BSWAP_64(dnp->dn_used);
 299 
 300         /*
 301          * dn_nblkptr is only one byte, so it's OK to read it in either
 302          * byte order.  We can't read dn_bouslen.
 303          */
 304         ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
 305         ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
 306         for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
 307                 buf64[i] = BSWAP_64(buf64[i]);
 308 
 309         /*
 310          * OK to check dn_bonuslen for zero, because it won't matter if
 311          * we have the wrong byte order.  This is necessary because the
 312          * dnode dnode is smaller than a regular dnode.
 313          */
 314         if (dnp->dn_bonuslen != 0) {
 315                 /*
 316                  * Note that the bonus length calculated here may be
 317                  * longer than the actual bonus buffer.  This is because
 318                  * we always put the bonus buffer after the last block
 319                  * pointer (instead of packing it against the end of the
 320                  * dnode buffer).
 321                  */
 322                 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
 323                 size_t len = DN_MAX_BONUSLEN - off;
 324                 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
 325                 dmu_object_byteswap_t byteswap =
 326                     DMU_OT_BYTESWAP(dnp->dn_bonustype);
 327                 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
 328         }
 329 
 330         /* Swap SPILL block if we have one */
 331         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 332                 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
 333 
 334 }
 335 
 336 void
 337 dnode_buf_byteswap(void *vbuf, size_t size)
 338 {
 339         dnode_phys_t *buf = vbuf;
 340         int i;
 341 
 342         ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
 343         ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
 344 
 345         size >>= DNODE_SHIFT;
 346         for (i = 0; i < size; i++) {
 347                 dnode_byteswap(buf);
 348                 buf++;
 349         }
 350 }
 351 
 352 void
 353 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
 354 {
 355         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 356 
 357         dnode_setdirty(dn, tx);
 358         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 359         ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
 360             (dn->dn_nblkptr-1) * sizeof (blkptr_t));
 361         dn->dn_bonuslen = newsize;
 362         if (newsize == 0)
 363                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
 364         else
 365                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 366         rw_exit(&dn->dn_struct_rwlock);
 367 }
 368 
 369 void
 370 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
 371 {
 372         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 373         dnode_setdirty(dn, tx);
 374         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 375         dn->dn_bonustype = newtype;
 376         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 377         rw_exit(&dn->dn_struct_rwlock);
 378 }
 379 
 380 void
 381 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
 382 {
 383         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 384         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
 385         dnode_setdirty(dn, tx);
 386         dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
 387         dn->dn_have_spill = B_FALSE;
 388 }
 389 
 390 static void
 391 dnode_setdblksz(dnode_t *dn, int size)
 392 {
 393         ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
 394         ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
 395         ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
 396         ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
 397             1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
 398         dn->dn_datablksz = size;
 399         dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
 400         dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
 401 }
 402 
 403 static dnode_t *
 404 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
 405     uint64_t object, dnode_handle_t *dnh)
 406 {
 407         dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
 408 
 409         ASSERT(!POINTER_IS_VALID(dn->dn_objset));
 410         dn->dn_moved = 0;
 411 
 412         /*
 413          * Defer setting dn_objset until the dnode is ready to be a candidate
 414          * for the dnode_move() callback.
 415          */
 416         dn->dn_object = object;
 417         dn->dn_dbuf = db;
 418         dn->dn_handle = dnh;
 419         dn->dn_phys = dnp;
 420 
 421         if (dnp->dn_datablkszsec) {
 422                 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 423         } else {
 424                 dn->dn_datablksz = 0;
 425                 dn->dn_datablkszsec = 0;
 426                 dn->dn_datablkshift = 0;
 427         }
 428         dn->dn_indblkshift = dnp->dn_indblkshift;
 429         dn->dn_nlevels = dnp->dn_nlevels;
 430         dn->dn_type = dnp->dn_type;
 431         dn->dn_nblkptr = dnp->dn_nblkptr;
 432         dn->dn_checksum = dnp->dn_checksum;
 433         dn->dn_compress = dnp->dn_compress;
 434         dn->dn_bonustype = dnp->dn_bonustype;
 435         dn->dn_bonuslen = dnp->dn_bonuslen;
 436         dn->dn_maxblkid = dnp->dn_maxblkid;
 437         dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
 438         dn->dn_id_flags = 0;
 439 
 440         dmu_zfetch_init(&dn->dn_zfetch, dn);
 441 
 442         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 443 
 444         mutex_enter(&os->os_lock);
 445         list_insert_head(&os->os_dnodes, dn);
 446         membar_producer();
 447         /*
 448          * Everything else must be valid before assigning dn_objset makes the
 449          * dnode eligible for dnode_move().
 450          */
 451         dn->dn_objset = os;
 452         mutex_exit(&os->os_lock);
 453 
 454         arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
 455         return (dn);
 456 }
 457 
 458 /*
 459  * Caller must be holding the dnode handle, which is released upon return.
 460  */
 461 static void
 462 dnode_destroy(dnode_t *dn)
 463 {
 464         objset_t *os = dn->dn_objset;
 465 
 466         ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
 467 
 468         mutex_enter(&os->os_lock);
 469         POINTER_INVALIDATE(&dn->dn_objset);
 470         list_remove(&os->os_dnodes, dn);
 471         mutex_exit(&os->os_lock);
 472 
 473         /* the dnode can no longer move, so we can release the handle */
 474         zrl_remove(&dn->dn_handle->dnh_zrlock);
 475 
 476         dn->dn_allocated_txg = 0;
 477         dn->dn_free_txg = 0;
 478         dn->dn_assigned_txg = 0;
 479 
 480         dn->dn_dirtyctx = 0;
 481         if (dn->dn_dirtyctx_firstset != NULL) {
 482                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 483                 dn->dn_dirtyctx_firstset = NULL;
 484         }
 485         if (dn->dn_bonus != NULL) {
 486                 mutex_enter(&dn->dn_bonus->db_mtx);
 487                 dbuf_evict(dn->dn_bonus);
 488                 dn->dn_bonus = NULL;
 489         }
 490         dn->dn_zio = NULL;
 491 
 492         dn->dn_have_spill = B_FALSE;
 493         dn->dn_oldused = 0;
 494         dn->dn_oldflags = 0;
 495         dn->dn_olduid = 0;
 496         dn->dn_oldgid = 0;
 497         dn->dn_newuid = 0;
 498         dn->dn_newgid = 0;
 499         dn->dn_id_flags = 0;
 500         dn->dn_unlisted_l0_blkid = 0;
 501 
 502         dmu_zfetch_rele(&dn->dn_zfetch);
 503         kmem_cache_free(dnode_cache, dn);
 504         arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
 505 }
 506 
 507 void
 508 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
 509     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 510 {
 511         int i;
 512 
 513         if (blocksize == 0)
 514                 blocksize = 1 << zfs_default_bs;
 515         else if (blocksize > SPA_MAXBLOCKSIZE)
 516                 blocksize = SPA_MAXBLOCKSIZE;
 517         else
 518                 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
 519 
 520         if (ibs == 0)
 521                 ibs = zfs_default_ibs;
 522 
 523         ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
 524 
 525         dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
 526             dn->dn_object, tx->tx_txg, blocksize, ibs);
 527 
 528         ASSERT(dn->dn_type == DMU_OT_NONE);
 529         ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
 530         ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
 531         ASSERT(ot != DMU_OT_NONE);
 532         ASSERT(DMU_OT_IS_VALID(ot));
 533         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 534             (bonustype == DMU_OT_SA && bonuslen == 0) ||
 535             (bonustype != DMU_OT_NONE && bonuslen != 0));
 536         ASSERT(DMU_OT_IS_VALID(bonustype));
 537         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 538         ASSERT(dn->dn_type == DMU_OT_NONE);
 539         ASSERT0(dn->dn_maxblkid);
 540         ASSERT0(dn->dn_allocated_txg);
 541         ASSERT0(dn->dn_assigned_txg);
 542         ASSERT(refcount_is_zero(&dn->dn_tx_holds));
 543         ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
 544         ASSERT(avl_is_empty(&dn->dn_dbufs));
 545 
 546         for (i = 0; i < TXG_SIZE; i++) {
 547                 ASSERT0(dn->dn_next_nblkptr[i]);
 548                 ASSERT0(dn->dn_next_nlevels[i]);
 549                 ASSERT0(dn->dn_next_indblkshift[i]);
 550                 ASSERT0(dn->dn_next_bonuslen[i]);
 551                 ASSERT0(dn->dn_next_bonustype[i]);
 552                 ASSERT0(dn->dn_rm_spillblk[i]);
 553                 ASSERT0(dn->dn_next_blksz[i]);
 554                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 555                 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
 556                 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
 557         }
 558 
 559         dn->dn_type = ot;
 560         dnode_setdblksz(dn, blocksize);
 561         dn->dn_indblkshift = ibs;
 562         dn->dn_nlevels = 1;
 563         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 564                 dn->dn_nblkptr = 1;
 565         else
 566                 dn->dn_nblkptr = 1 +
 567                     ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 568         dn->dn_bonustype = bonustype;
 569         dn->dn_bonuslen = bonuslen;
 570         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 571         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 572         dn->dn_dirtyctx = 0;
 573 
 574         dn->dn_free_txg = 0;
 575         if (dn->dn_dirtyctx_firstset) {
 576                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 577                 dn->dn_dirtyctx_firstset = NULL;
 578         }
 579 
 580         dn->dn_allocated_txg = tx->tx_txg;
 581         dn->dn_id_flags = 0;
 582 
 583         dnode_setdirty(dn, tx);
 584         dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
 585         dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 586         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 587         dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
 588 }
 589 
 590 void
 591 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
 592     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 593 {
 594         int nblkptr;
 595 
 596         ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
 597         ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
 598         ASSERT0(blocksize % SPA_MINBLOCKSIZE);
 599         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
 600         ASSERT(tx->tx_txg != 0);
 601         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 602             (bonustype != DMU_OT_NONE && bonuslen != 0) ||
 603             (bonustype == DMU_OT_SA && bonuslen == 0));
 604         ASSERT(DMU_OT_IS_VALID(bonustype));
 605         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 606 
 607         /* clean up any unreferenced dbufs */
 608         dnode_evict_dbufs(dn);
 609 
 610         dn->dn_id_flags = 0;
 611 
 612         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 613         dnode_setdirty(dn, tx);
 614         if (dn->dn_datablksz != blocksize) {
 615                 /* change blocksize */
 616                 ASSERT(dn->dn_maxblkid == 0 &&
 617                     (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
 618                     dnode_block_freed(dn, 0)));
 619                 dnode_setdblksz(dn, blocksize);
 620                 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
 621         }
 622         if (dn->dn_bonuslen != bonuslen)
 623                 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
 624 
 625         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 626                 nblkptr = 1;
 627         else
 628                 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 629         if (dn->dn_bonustype != bonustype)
 630                 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
 631         if (dn->dn_nblkptr != nblkptr)
 632                 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
 633         if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 634                 dbuf_rm_spill(dn, tx);
 635                 dnode_rm_spill(dn, tx);
 636         }
 637         rw_exit(&dn->dn_struct_rwlock);
 638 
 639         /* change type */
 640         dn->dn_type = ot;
 641 
 642         /* change bonus size and type */
 643         mutex_enter(&dn->dn_mtx);
 644         dn->dn_bonustype = bonustype;
 645         dn->dn_bonuslen = bonuslen;
 646         dn->dn_nblkptr = nblkptr;
 647         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 648         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 649         ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 650 
 651         /* fix up the bonus db_size */
 652         if (dn->dn_bonus) {
 653                 dn->dn_bonus->db.db_size =
 654                     DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
 655                 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
 656         }
 657 
 658         dn->dn_allocated_txg = tx->tx_txg;
 659         mutex_exit(&dn->dn_mtx);
 660 }
 661 
 662 #ifdef  DNODE_STATS
 663 static struct {
 664         uint64_t dms_dnode_invalid;
 665         uint64_t dms_dnode_recheck1;
 666         uint64_t dms_dnode_recheck2;
 667         uint64_t dms_dnode_special;
 668         uint64_t dms_dnode_handle;
 669         uint64_t dms_dnode_rwlock;
 670         uint64_t dms_dnode_active;
 671 } dnode_move_stats;
 672 #endif  /* DNODE_STATS */
 673 
 674 static void
 675 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
 676 {
 677         int i;
 678 
 679         ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
 680         ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
 681         ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
 682         ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
 683 
 684         /* Copy fields. */
 685         ndn->dn_objset = odn->dn_objset;
 686         ndn->dn_object = odn->dn_object;
 687         ndn->dn_dbuf = odn->dn_dbuf;
 688         ndn->dn_handle = odn->dn_handle;
 689         ndn->dn_phys = odn->dn_phys;
 690         ndn->dn_type = odn->dn_type;
 691         ndn->dn_bonuslen = odn->dn_bonuslen;
 692         ndn->dn_bonustype = odn->dn_bonustype;
 693         ndn->dn_nblkptr = odn->dn_nblkptr;
 694         ndn->dn_checksum = odn->dn_checksum;
 695         ndn->dn_compress = odn->dn_compress;
 696         ndn->dn_nlevels = odn->dn_nlevels;
 697         ndn->dn_indblkshift = odn->dn_indblkshift;
 698         ndn->dn_datablkshift = odn->dn_datablkshift;
 699         ndn->dn_datablkszsec = odn->dn_datablkszsec;
 700         ndn->dn_datablksz = odn->dn_datablksz;
 701         ndn->dn_maxblkid = odn->dn_maxblkid;
 702         bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
 703             sizeof (odn->dn_next_nblkptr));
 704         bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
 705             sizeof (odn->dn_next_nlevels));
 706         bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
 707             sizeof (odn->dn_next_indblkshift));
 708         bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
 709             sizeof (odn->dn_next_bonustype));
 710         bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
 711             sizeof (odn->dn_rm_spillblk));
 712         bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
 713             sizeof (odn->dn_next_bonuslen));
 714         bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
 715             sizeof (odn->dn_next_blksz));
 716         for (i = 0; i < TXG_SIZE; i++) {
 717                 list_move_tail(&ndn->dn_dirty_records[i],
 718                     &odn->dn_dirty_records[i]);
 719         }
 720         bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
 721             sizeof (odn->dn_free_ranges));
 722         ndn->dn_allocated_txg = odn->dn_allocated_txg;
 723         ndn->dn_free_txg = odn->dn_free_txg;
 724         ndn->dn_assigned_txg = odn->dn_assigned_txg;
 725         ndn->dn_dirtyctx = odn->dn_dirtyctx;
 726         ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
 727         ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
 728         refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
 729         ASSERT(avl_is_empty(&ndn->dn_dbufs));
 730         avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
 731         ndn->dn_dbufs_count = odn->dn_dbufs_count;
 732         ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
 733         ndn->dn_bonus = odn->dn_bonus;
 734         ndn->dn_have_spill = odn->dn_have_spill;
 735         ndn->dn_zio = odn->dn_zio;
 736         ndn->dn_oldused = odn->dn_oldused;
 737         ndn->dn_oldflags = odn->dn_oldflags;
 738         ndn->dn_olduid = odn->dn_olduid;
 739         ndn->dn_oldgid = odn->dn_oldgid;
 740         ndn->dn_newuid = odn->dn_newuid;
 741         ndn->dn_newgid = odn->dn_newgid;
 742         ndn->dn_id_flags = odn->dn_id_flags;
 743         dmu_zfetch_init(&ndn->dn_zfetch, NULL);
 744         list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
 745         ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
 746         ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
 747         ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
 748 
 749         /*
 750          * Update back pointers. Updating the handle fixes the back pointer of
 751          * every descendant dbuf as well as the bonus dbuf.
 752          */
 753         ASSERT(ndn->dn_handle->dnh_dnode == odn);
 754         ndn->dn_handle->dnh_dnode = ndn;
 755         if (ndn->dn_zfetch.zf_dnode == odn) {
 756                 ndn->dn_zfetch.zf_dnode = ndn;
 757         }
 758 
 759         /*
 760          * Invalidate the original dnode by clearing all of its back pointers.
 761          */
 762         odn->dn_dbuf = NULL;
 763         odn->dn_handle = NULL;
 764         avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
 765             offsetof(dmu_buf_impl_t, db_link));
 766         odn->dn_dbufs_count = 0;
 767         odn->dn_unlisted_l0_blkid = 0;
 768         odn->dn_bonus = NULL;
 769         odn->dn_zfetch.zf_dnode = NULL;
 770 
 771         /*
 772          * Set the low bit of the objset pointer to ensure that dnode_move()
 773          * recognizes the dnode as invalid in any subsequent callback.
 774          */
 775         POINTER_INVALIDATE(&odn->dn_objset);
 776 
 777         /*
 778          * Satisfy the destructor.
 779          */
 780         for (i = 0; i < TXG_SIZE; i++) {
 781                 list_create(&odn->dn_dirty_records[i],
 782                     sizeof (dbuf_dirty_record_t),
 783                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 784                 odn->dn_free_ranges[i] = NULL;
 785                 odn->dn_next_nlevels[i] = 0;
 786                 odn->dn_next_indblkshift[i] = 0;
 787                 odn->dn_next_bonustype[i] = 0;
 788                 odn->dn_rm_spillblk[i] = 0;
 789                 odn->dn_next_bonuslen[i] = 0;
 790                 odn->dn_next_blksz[i] = 0;
 791         }
 792         odn->dn_allocated_txg = 0;
 793         odn->dn_free_txg = 0;
 794         odn->dn_assigned_txg = 0;
 795         odn->dn_dirtyctx = 0;
 796         odn->dn_dirtyctx_firstset = NULL;
 797         odn->dn_have_spill = B_FALSE;
 798         odn->dn_zio = NULL;
 799         odn->dn_oldused = 0;
 800         odn->dn_oldflags = 0;
 801         odn->dn_olduid = 0;
 802         odn->dn_oldgid = 0;
 803         odn->dn_newuid = 0;
 804         odn->dn_newgid = 0;
 805         odn->dn_id_flags = 0;
 806 
 807         /*
 808          * Mark the dnode.
 809          */
 810         ndn->dn_moved = 1;
 811         odn->dn_moved = (uint8_t)-1;
 812 }
 813 
 814 #ifdef  _KERNEL
 815 /*ARGSUSED*/
 816 static kmem_cbrc_t
 817 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
 818 {
 819         dnode_t *odn = buf, *ndn = newbuf;
 820         objset_t *os;
 821         int64_t refcount;
 822         uint32_t dbufs;
 823 
 824         /*
 825          * The dnode is on the objset's list of known dnodes if the objset
 826          * pointer is valid. We set the low bit of the objset pointer when
 827          * freeing the dnode to invalidate it, and the memory patterns written
 828          * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
 829          * A newly created dnode sets the objset pointer last of all to indicate
 830          * that the dnode is known and in a valid state to be moved by this
 831          * function.
 832          */
 833         os = odn->dn_objset;
 834         if (!POINTER_IS_VALID(os)) {
 835                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
 836                 return (KMEM_CBRC_DONT_KNOW);
 837         }
 838 
 839         /*
 840          * Ensure that the objset does not go away during the move.
 841          */
 842         rw_enter(&os_lock, RW_WRITER);
 843         if (os != odn->dn_objset) {
 844                 rw_exit(&os_lock);
 845                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
 846                 return (KMEM_CBRC_DONT_KNOW);
 847         }
 848 
 849         /*
 850          * If the dnode is still valid, then so is the objset. We know that no
 851          * valid objset can be freed while we hold os_lock, so we can safely
 852          * ensure that the objset remains in use.
 853          */
 854         mutex_enter(&os->os_lock);
 855 
 856         /*
 857          * Recheck the objset pointer in case the dnode was removed just before
 858          * acquiring the lock.
 859          */
 860         if (os != odn->dn_objset) {
 861                 mutex_exit(&os->os_lock);
 862                 rw_exit(&os_lock);
 863                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
 864                 return (KMEM_CBRC_DONT_KNOW);
 865         }
 866 
 867         /*
 868          * At this point we know that as long as we hold os->os_lock, the dnode
 869          * cannot be freed and fields within the dnode can be safely accessed.
 870          * The objset listing this dnode cannot go away as long as this dnode is
 871          * on its list.
 872          */
 873         rw_exit(&os_lock);
 874         if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
 875                 mutex_exit(&os->os_lock);
 876                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
 877                 return (KMEM_CBRC_NO);
 878         }
 879         ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
 880 
 881         /*
 882          * Lock the dnode handle to prevent the dnode from obtaining any new
 883          * holds. This also prevents the descendant dbufs and the bonus dbuf
 884          * from accessing the dnode, so that we can discount their holds. The
 885          * handle is safe to access because we know that while the dnode cannot
 886          * go away, neither can its handle. Once we hold dnh_zrlock, we can
 887          * safely move any dnode referenced only by dbufs.
 888          */
 889         if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
 890                 mutex_exit(&os->os_lock);
 891                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
 892                 return (KMEM_CBRC_LATER);
 893         }
 894 
 895         /*
 896          * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
 897          * We need to guarantee that there is a hold for every dbuf in order to
 898          * determine whether the dnode is actively referenced. Falsely matching
 899          * a dbuf to an active hold would lead to an unsafe move. It's possible
 900          * that a thread already having an active dnode hold is about to add a
 901          * dbuf, and we can't compare hold and dbuf counts while the add is in
 902          * progress.
 903          */
 904         if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
 905                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 906                 mutex_exit(&os->os_lock);
 907                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
 908                 return (KMEM_CBRC_LATER);
 909         }
 910 
 911         /*
 912          * A dbuf may be removed (evicted) without an active dnode hold. In that
 913          * case, the dbuf count is decremented under the handle lock before the
 914          * dbuf's hold is released. This order ensures that if we count the hold
 915          * after the dbuf is removed but before its hold is released, we will
 916          * treat the unmatched hold as active and exit safely. If we count the
 917          * hold before the dbuf is removed, the hold is discounted, and the
 918          * removal is blocked until the move completes.
 919          */
 920         refcount = refcount_count(&odn->dn_holds);
 921         ASSERT(refcount >= 0);
 922         dbufs = odn->dn_dbufs_count;
 923 
 924         /* We can't have more dbufs than dnode holds. */
 925         ASSERT3U(dbufs, <=, refcount);
 926         DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
 927             uint32_t, dbufs);
 928 
 929         if (refcount > dbufs) {
 930                 rw_exit(&odn->dn_struct_rwlock);
 931                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 932                 mutex_exit(&os->os_lock);
 933                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
 934                 return (KMEM_CBRC_LATER);
 935         }
 936 
 937         rw_exit(&odn->dn_struct_rwlock);
 938 
 939         /*
 940          * At this point we know that anyone with a hold on the dnode is not
 941          * actively referencing it. The dnode is known and in a valid state to
 942          * move. We're holding the locks needed to execute the critical section.
 943          */
 944         dnode_move_impl(odn, ndn);
 945 
 946         list_link_replace(&odn->dn_link, &ndn->dn_link);
 947         /* If the dnode was safe to move, the refcount cannot have changed. */
 948         ASSERT(refcount == refcount_count(&ndn->dn_holds));
 949         ASSERT(dbufs == ndn->dn_dbufs_count);
 950         zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
 951         mutex_exit(&os->os_lock);
 952 
 953         return (KMEM_CBRC_YES);
 954 }
 955 #endif  /* _KERNEL */
 956 
 957 void
 958 dnode_special_close(dnode_handle_t *dnh)
 959 {
 960         dnode_t *dn = dnh->dnh_dnode;
 961 
 962         /*
 963          * Wait for final references to the dnode to clear.  This can
 964          * only happen if the arc is asyncronously evicting state that
 965          * has a hold on this dnode while we are trying to evict this
 966          * dnode.
 967          */
 968         while (refcount_count(&dn->dn_holds) > 0)
 969                 delay(1);
 970         zrl_add(&dnh->dnh_zrlock);
 971         dnode_destroy(dn); /* implicit zrl_remove() */
 972         zrl_destroy(&dnh->dnh_zrlock);
 973         dnh->dnh_dnode = NULL;
 974 }
 975 
 976 dnode_t *
 977 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
 978     dnode_handle_t *dnh)
 979 {
 980         dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
 981         dnh->dnh_dnode = dn;
 982         zrl_init(&dnh->dnh_zrlock);
 983         DNODE_VERIFY(dn);
 984         return (dn);
 985 }
 986 
 987 static void
 988 dnode_buf_pageout(dmu_buf_t *db, void *arg)
 989 {
 990         dnode_children_t *children_dnodes = arg;
 991         int i;
 992         int epb = db->db_size >> DNODE_SHIFT;
 993 
 994         ASSERT(epb == children_dnodes->dnc_count);
 995 
 996         for (i = 0; i < epb; i++) {
 997                 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
 998                 dnode_t *dn;
 999 
1000                 /*
1001                  * The dnode handle lock guards against the dnode moving to
1002                  * another valid address, so there is no need here to guard
1003                  * against changes to or from NULL.
1004                  */
1005                 if (dnh->dnh_dnode == NULL) {
1006                         zrl_destroy(&dnh->dnh_zrlock);
1007                         continue;
1008                 }
1009 
1010                 zrl_add(&dnh->dnh_zrlock);
1011                 dn = dnh->dnh_dnode;
1012                 /*
1013                  * If there are holds on this dnode, then there should
1014                  * be holds on the dnode's containing dbuf as well; thus
1015                  * it wouldn't be eligible for eviction and this function
1016                  * would not have been called.
1017                  */
1018                 ASSERT(refcount_is_zero(&dn->dn_holds));
1019                 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1020 
1021                 dnode_destroy(dn); /* implicit zrl_remove() */
1022                 zrl_destroy(&dnh->dnh_zrlock);
1023                 dnh->dnh_dnode = NULL;
1024         }
1025         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1026             epb * sizeof (dnode_handle_t));
1027 }
1028 
1029 /*
1030  * errors:
1031  * EINVAL - invalid object number.
1032  * EIO - i/o error.
1033  * succeeds even for free dnodes.
1034  */
1035 int
1036 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1037     void *tag, dnode_t **dnp)
1038 {
1039         int epb, idx, err;
1040         int drop_struct_lock = FALSE;
1041         int type;
1042         uint64_t blk;
1043         dnode_t *mdn, *dn;
1044         dmu_buf_impl_t *db;
1045         dnode_children_t *children_dnodes;
1046         dnode_handle_t *dnh;
1047 
1048         /*
1049          * If you are holding the spa config lock as writer, you shouldn't
1050          * be asking the DMU to do *anything* unless it's the root pool
1051          * which may require us to read from the root filesystem while
1052          * holding some (not all) of the locks as writer.
1053          */
1054         ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1055             (spa_is_root(os->os_spa) &&
1056             spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1057 
1058         if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1059                 dn = (object == DMU_USERUSED_OBJECT) ?
1060                     DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1061                 if (dn == NULL)
1062                         return (SET_ERROR(ENOENT));
1063                 type = dn->dn_type;
1064                 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1065                         return (SET_ERROR(ENOENT));
1066                 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1067                         return (SET_ERROR(EEXIST));
1068                 DNODE_VERIFY(dn);
1069                 (void) refcount_add(&dn->dn_holds, tag);
1070                 *dnp = dn;
1071                 return (0);
1072         }
1073 
1074         if (object == 0 || object >= DN_MAX_OBJECT)
1075                 return (SET_ERROR(EINVAL));
1076 
1077         mdn = DMU_META_DNODE(os);
1078         ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1079 
1080         DNODE_VERIFY(mdn);
1081 
1082         if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1083                 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1084                 drop_struct_lock = TRUE;
1085         }
1086 
1087         blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1088 
1089         db = dbuf_hold(mdn, blk, FTAG);
1090         if (drop_struct_lock)
1091                 rw_exit(&mdn->dn_struct_rwlock);
1092         if (db == NULL)
1093                 return (SET_ERROR(EIO));
1094         err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1095         if (err) {
1096                 dbuf_rele(db, FTAG);
1097                 return (err);
1098         }
1099 
1100         ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1101         epb = db->db.db_size >> DNODE_SHIFT;
1102 
1103         idx = object & (epb-1);
1104 
1105         ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1106         children_dnodes = dmu_buf_get_user(&db->db);
1107         if (children_dnodes == NULL) {
1108                 int i;
1109                 dnode_children_t *winner;
1110                 children_dnodes = kmem_alloc(sizeof (dnode_children_t) +
1111                     epb * sizeof (dnode_handle_t), KM_SLEEP);
1112                 children_dnodes->dnc_count = epb;
1113                 dnh = &children_dnodes->dnc_children[0];
1114                 for (i = 0; i < epb; i++) {
1115                         zrl_init(&dnh[i].dnh_zrlock);
1116                         dnh[i].dnh_dnode = NULL;
1117                 }
1118                 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1119                     dnode_buf_pageout)) {
1120 
1121                         for (i = 0; i < epb; i++) {
1122                                 zrl_destroy(&dnh[i].dnh_zrlock);
1123                         }
1124 
1125                         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1126                             epb * sizeof (dnode_handle_t));
1127                         children_dnodes = winner;
1128                 }
1129         }
1130         ASSERT(children_dnodes->dnc_count == epb);
1131 
1132         dnh = &children_dnodes->dnc_children[idx];
1133         zrl_add(&dnh->dnh_zrlock);
1134         if ((dn = dnh->dnh_dnode) == NULL) {
1135                 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1136                 dnode_t *winner;
1137 
1138                 dn = dnode_create(os, phys, db, object, dnh);
1139                 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1140                 if (winner != NULL) {
1141                         zrl_add(&dnh->dnh_zrlock);
1142                         dnode_destroy(dn); /* implicit zrl_remove() */
1143                         dn = winner;
1144                 }
1145         }
1146 
1147         mutex_enter(&dn->dn_mtx);
1148         type = dn->dn_type;
1149         if (dn->dn_free_txg ||
1150             ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1151             ((flag & DNODE_MUST_BE_FREE) &&
1152             (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1153                 mutex_exit(&dn->dn_mtx);
1154                 zrl_remove(&dnh->dnh_zrlock);
1155                 dbuf_rele(db, FTAG);
1156                 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1157         }
1158         mutex_exit(&dn->dn_mtx);
1159 
1160         if (refcount_add(&dn->dn_holds, tag) == 1)
1161                 dbuf_add_ref(db, dnh);
1162         /* Now we can rely on the hold to prevent the dnode from moving. */
1163         zrl_remove(&dnh->dnh_zrlock);
1164 
1165         DNODE_VERIFY(dn);
1166         ASSERT3P(dn->dn_dbuf, ==, db);
1167         ASSERT3U(dn->dn_object, ==, object);
1168         dbuf_rele(db, FTAG);
1169 
1170         *dnp = dn;
1171         return (0);
1172 }
1173 
1174 /*
1175  * Return held dnode if the object is allocated, NULL if not.
1176  */
1177 int
1178 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1179 {
1180         return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1181 }
1182 
1183 /*
1184  * Can only add a reference if there is already at least one
1185  * reference on the dnode.  Returns FALSE if unable to add a
1186  * new reference.
1187  */
1188 boolean_t
1189 dnode_add_ref(dnode_t *dn, void *tag)
1190 {
1191         mutex_enter(&dn->dn_mtx);
1192         if (refcount_is_zero(&dn->dn_holds)) {
1193                 mutex_exit(&dn->dn_mtx);
1194                 return (FALSE);
1195         }
1196         VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1197         mutex_exit(&dn->dn_mtx);
1198         return (TRUE);
1199 }
1200 
1201 void
1202 dnode_rele(dnode_t *dn, void *tag)
1203 {
1204         uint64_t refs;
1205         /* Get while the hold prevents the dnode from moving. */
1206         dmu_buf_impl_t *db = dn->dn_dbuf;
1207         dnode_handle_t *dnh = dn->dn_handle;
1208 
1209         mutex_enter(&dn->dn_mtx);
1210         refs = refcount_remove(&dn->dn_holds, tag);
1211         mutex_exit(&dn->dn_mtx);
1212 
1213         /*
1214          * It's unsafe to release the last hold on a dnode by dnode_rele() or
1215          * indirectly by dbuf_rele() while relying on the dnode handle to
1216          * prevent the dnode from moving, since releasing the last hold could
1217          * result in the dnode's parent dbuf evicting its dnode handles. For
1218          * that reason anyone calling dnode_rele() or dbuf_rele() without some
1219          * other direct or indirect hold on the dnode must first drop the dnode
1220          * handle.
1221          */
1222         ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1223 
1224         /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1225         if (refs == 0 && db != NULL) {
1226                 /*
1227                  * Another thread could add a hold to the dnode handle in
1228                  * dnode_hold_impl() while holding the parent dbuf. Since the
1229                  * hold on the parent dbuf prevents the handle from being
1230                  * destroyed, the hold on the handle is OK. We can't yet assert
1231                  * that the handle has zero references, but that will be
1232                  * asserted anyway when the handle gets destroyed.
1233                  */
1234                 dbuf_rele(db, dnh);
1235         }
1236 }
1237 
1238 void
1239 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1240 {
1241         objset_t *os = dn->dn_objset;
1242         uint64_t txg = tx->tx_txg;
1243 
1244         if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1245                 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1246                 return;
1247         }
1248 
1249         DNODE_VERIFY(dn);
1250 
1251 #ifdef ZFS_DEBUG
1252         mutex_enter(&dn->dn_mtx);
1253         ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1254         ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1255         mutex_exit(&dn->dn_mtx);
1256 #endif
1257 
1258         /*
1259          * Determine old uid/gid when necessary
1260          */
1261         dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1262 
1263         mutex_enter(&os->os_lock);
1264 
1265         /*
1266          * If we are already marked dirty, we're done.
1267          */
1268         if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1269                 mutex_exit(&os->os_lock);
1270                 return;
1271         }
1272 
1273         ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1274             !avl_is_empty(&dn->dn_dbufs));
1275         ASSERT(dn->dn_datablksz != 0);
1276         ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1277         ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1278         ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1279 
1280         dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1281             dn->dn_object, txg);
1282 
1283         if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1284                 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1285         } else {
1286                 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1287         }
1288 
1289         mutex_exit(&os->os_lock);
1290 
1291         /*
1292          * The dnode maintains a hold on its containing dbuf as
1293          * long as there are holds on it.  Each instantiated child
1294          * dbuf maintains a hold on the dnode.  When the last child
1295          * drops its hold, the dnode will drop its hold on the
1296          * containing dbuf. We add a "dirty hold" here so that the
1297          * dnode will hang around after we finish processing its
1298          * children.
1299          */
1300         VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1301 
1302         (void) dbuf_dirty(dn->dn_dbuf, tx, B_FALSE);
1303 
1304         dsl_dataset_dirty(os->os_dsl_dataset, tx);
1305 }
1306 
1307 void
1308 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1309 {
1310         int txgoff = tx->tx_txg & TXG_MASK;
1311 
1312         dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1313 
1314         /* we should be the only holder... hopefully */
1315         /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1316 
1317         mutex_enter(&dn->dn_mtx);
1318         if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1319                 mutex_exit(&dn->dn_mtx);
1320                 return;
1321         }
1322         dn->dn_free_txg = tx->tx_txg;
1323         mutex_exit(&dn->dn_mtx);
1324 
1325         /*
1326          * If the dnode is already dirty, it needs to be moved from
1327          * the dirty list to the free list.
1328          */
1329         mutex_enter(&dn->dn_objset->os_lock);
1330         if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1331                 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1332                 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1333                 mutex_exit(&dn->dn_objset->os_lock);
1334         } else {
1335                 mutex_exit(&dn->dn_objset->os_lock);
1336                 dnode_setdirty(dn, tx);
1337         }
1338 }
1339 
1340 /*
1341  * Try to change the block size for the indicated dnode.  This can only
1342  * succeed if there are no blocks allocated or dirty beyond first block
1343  */
1344 int
1345 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1346 {
1347         dmu_buf_impl_t *db;
1348         int err;
1349 
1350         if (size == 0)
1351                 size = SPA_MINBLOCKSIZE;
1352         if (size > SPA_MAXBLOCKSIZE)
1353                 size = SPA_MAXBLOCKSIZE;
1354         else
1355                 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1356 
1357         if (ibs == dn->dn_indblkshift)
1358                 ibs = 0;
1359 
1360         if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1361                 return (0);
1362 
1363         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1364 
1365         /* Check for any allocated blocks beyond the first */
1366         if (dn->dn_maxblkid != 0)
1367                 goto fail;
1368 
1369         mutex_enter(&dn->dn_dbufs_mtx);
1370         for (db = avl_first(&dn->dn_dbufs); db != NULL;
1371             db = AVL_NEXT(&dn->dn_dbufs, db)) {
1372                 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1373                     db->db_blkid != DMU_SPILL_BLKID) {
1374                         mutex_exit(&dn->dn_dbufs_mtx);
1375                         goto fail;
1376                 }
1377         }
1378         mutex_exit(&dn->dn_dbufs_mtx);
1379 
1380         if (ibs && dn->dn_nlevels != 1)
1381                 goto fail;
1382 
1383         /* resize the old block */
1384         err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1385         if (err == 0)
1386                 dbuf_new_size(db, size, tx);
1387         else if (err != ENOENT)
1388                 goto fail;
1389 
1390         dnode_setdblksz(dn, size);
1391         dnode_setdirty(dn, tx);
1392         dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1393         if (ibs) {
1394                 dn->dn_indblkshift = ibs;
1395                 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1396         }
1397         /* rele after we have fixed the blocksize in the dnode */
1398         if (db)
1399                 dbuf_rele(db, FTAG);
1400 
1401         rw_exit(&dn->dn_struct_rwlock);
1402         return (0);
1403 
1404 fail:
1405         rw_exit(&dn->dn_struct_rwlock);
1406         return (SET_ERROR(ENOTSUP));
1407 }
1408 
1409 /* read-holding callers must not rely on the lock being continuously held */
1410 void
1411 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1412 {
1413         uint64_t txgoff = tx->tx_txg & TXG_MASK;
1414         int epbs, new_nlevels;
1415         uint64_t sz;
1416 
1417         ASSERT(blkid != DMU_BONUS_BLKID);
1418 
1419         ASSERT(have_read ?
1420             RW_READ_HELD(&dn->dn_struct_rwlock) :
1421             RW_WRITE_HELD(&dn->dn_struct_rwlock));
1422 
1423         /*
1424          * if we have a read-lock, check to see if we need to do any work
1425          * before upgrading to a write-lock.
1426          */
1427         if (have_read) {
1428                 if (blkid <= dn->dn_maxblkid)
1429                         return;
1430 
1431                 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1432                         rw_exit(&dn->dn_struct_rwlock);
1433                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1434                 }
1435         }
1436 
1437         if (blkid <= dn->dn_maxblkid)
1438                 goto out;
1439 
1440         dn->dn_maxblkid = blkid;
1441 
1442         /*
1443          * Compute the number of levels necessary to support the new maxblkid.
1444          */
1445         new_nlevels = 1;
1446         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1447         for (sz = dn->dn_nblkptr;
1448             sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1449                 new_nlevels++;
1450 
1451         if (new_nlevels > dn->dn_nlevels) {
1452                 int old_nlevels = dn->dn_nlevels;
1453                 dmu_buf_impl_t *db;
1454                 list_t *list;
1455                 dbuf_dirty_record_t *new, *dr, *dr_next;
1456 
1457                 dn->dn_nlevels = new_nlevels;
1458 
1459                 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1460                 dn->dn_next_nlevels[txgoff] = new_nlevels;
1461 
1462                 /* dirty the left indirects */
1463                 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1464                 ASSERT(db != NULL);
1465                 new = dbuf_dirty(db, tx, B_FALSE);
1466                 dbuf_rele(db, FTAG);
1467 
1468                 /* transfer the dirty records to the new indirect */
1469                 mutex_enter(&dn->dn_mtx);
1470                 mutex_enter(&new->dt.di.dr_mtx);
1471                 list = &dn->dn_dirty_records[txgoff];
1472                 for (dr = list_head(list); dr; dr = dr_next) {
1473                         dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1474                         if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1475                             dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1476                             dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1477                                 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1478                                 list_remove(&dn->dn_dirty_records[txgoff], dr);
1479                                 list_insert_tail(&new->dt.di.dr_children, dr);
1480                                 dr->dr_parent = new;
1481                         }
1482                 }
1483                 mutex_exit(&new->dt.di.dr_mtx);
1484                 mutex_exit(&dn->dn_mtx);
1485         }
1486 
1487 out:
1488         if (have_read)
1489                 rw_downgrade(&dn->dn_struct_rwlock);
1490 }
1491 
1492 void
1493 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1494 {
1495         dmu_buf_impl_t *db;
1496         uint64_t blkoff, blkid, nblks;
1497         int blksz, blkshift, head, tail;
1498         int trunc = FALSE;
1499         int epbs;
1500 
1501         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1502         blksz = dn->dn_datablksz;
1503         blkshift = dn->dn_datablkshift;
1504         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1505 
1506         if (len == DMU_OBJECT_END) {
1507                 len = UINT64_MAX - off;
1508                 trunc = TRUE;
1509         }
1510 
1511         /*
1512          * First, block align the region to free:
1513          */
1514         if (ISP2(blksz)) {
1515                 head = P2NPHASE(off, blksz);
1516                 blkoff = P2PHASE(off, blksz);
1517                 if ((off >> blkshift) > dn->dn_maxblkid)
1518                         goto out;
1519         } else {
1520                 ASSERT(dn->dn_maxblkid == 0);
1521                 if (off == 0 && len >= blksz) {
1522                         /*
1523                          * Freeing the whole block; fast-track this request.
1524                          * Note that we won't dirty any indirect blocks,
1525                          * which is fine because we will be freeing the entire
1526                          * file and thus all indirect blocks will be freed
1527                          * by free_children().
1528                          */
1529                         blkid = 0;
1530                         nblks = 1;
1531                         goto done;
1532                 } else if (off >= blksz) {
1533                         /* Freeing past end-of-data */
1534                         goto out;
1535                 } else {
1536                         /* Freeing part of the block. */
1537                         head = blksz - off;
1538                         ASSERT3U(head, >, 0);
1539                 }
1540                 blkoff = off;
1541         }
1542         /* zero out any partial block data at the start of the range */
1543         if (head) {
1544                 ASSERT3U(blkoff + head, ==, blksz);
1545                 if (len < head)
1546                         head = len;
1547                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1548                     FTAG, &db) == 0) {
1549                         caddr_t data;
1550 
1551                         /* don't dirty if it isn't on disk and isn't dirty */
1552                         if (db->db_last_dirty ||
1553                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1554                                 rw_exit(&dn->dn_struct_rwlock);
1555                                 dmu_buf_will_dirty(&db->db, tx);
1556                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1557                                 data = db->db.db_data;
1558                                 bzero(data + blkoff, head);
1559                         }
1560                         dbuf_rele(db, FTAG);
1561                 }
1562                 off += head;
1563                 len -= head;
1564         }
1565 
1566         /* If the range was less than one block, we're done */
1567         if (len == 0)
1568                 goto out;
1569 
1570         /* If the remaining range is past end of file, we're done */
1571         if ((off >> blkshift) > dn->dn_maxblkid)
1572                 goto out;
1573 
1574         ASSERT(ISP2(blksz));
1575         if (trunc)
1576                 tail = 0;
1577         else
1578                 tail = P2PHASE(len, blksz);
1579 
1580         ASSERT0(P2PHASE(off, blksz));
1581         /* zero out any partial block data at the end of the range */
1582         if (tail) {
1583                 if (len < tail)
1584                         tail = len;
1585                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1586                     TRUE, FTAG, &db) == 0) {
1587                         /* don't dirty if not on disk and not dirty */
1588                         if (db->db_last_dirty ||
1589                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1590                                 rw_exit(&dn->dn_struct_rwlock);
1591                                 dmu_buf_will_dirty(&db->db, tx);
1592                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1593                                 bzero(db->db.db_data, tail);
1594                         }
1595                         dbuf_rele(db, FTAG);
1596                 }
1597                 len -= tail;
1598         }
1599 
1600         /* If the range did not include a full block, we are done */
1601         if (len == 0)
1602                 goto out;
1603 
1604         ASSERT(IS_P2ALIGNED(off, blksz));
1605         ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1606         blkid = off >> blkshift;
1607         nblks = len >> blkshift;
1608         if (trunc)
1609                 nblks += 1;
1610 
1611         /*
1612          * Dirty the first and last indirect blocks, as they (and/or their
1613          * parents) will need to be written out if they were only
1614          * partially freed.  Interior indirect blocks will be themselves freed,
1615          * by free_children(), so they need not be dirtied.  Note that these
1616          * interior blocks have already been prefetched by dmu_tx_hold_free().
1617          */
1618         if (dn->dn_nlevels > 1) {
1619                 uint64_t first, last;
1620 
1621                 first = blkid >> epbs;
1622                 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1623                         dmu_buf_will_dirty(&db->db, tx);
1624                         dbuf_rele(db, FTAG);
1625                 }
1626                 if (trunc)
1627                         last = dn->dn_maxblkid >> epbs;
1628                 else
1629                         last = (blkid + nblks - 1) >> epbs;
1630                 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1631                         dmu_buf_will_dirty(&db->db, tx);
1632                         dbuf_rele(db, FTAG);
1633                 }
1634         }
1635 
1636 done:
1637         /*
1638          * Add this range to the dnode range list.
1639          * We will finish up this free operation in the syncing phase.
1640          */
1641         mutex_enter(&dn->dn_mtx);
1642         int txgoff = tx->tx_txg & TXG_MASK;
1643         if (dn->dn_free_ranges[txgoff] == NULL) {
1644                 dn->dn_free_ranges[txgoff] =
1645                     range_tree_create(NULL, NULL, &dn->dn_mtx);
1646         }
1647         range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1648         range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1649         dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1650             blkid, nblks, tx->tx_txg);
1651         mutex_exit(&dn->dn_mtx);
1652 
1653         dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1654         dnode_setdirty(dn, tx);
1655 out:
1656 
1657         rw_exit(&dn->dn_struct_rwlock);
1658 }
1659 
1660 static boolean_t
1661 dnode_spill_freed(dnode_t *dn)
1662 {
1663         int i;
1664 
1665         mutex_enter(&dn->dn_mtx);
1666         for (i = 0; i < TXG_SIZE; i++) {
1667                 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1668                         break;
1669         }
1670         mutex_exit(&dn->dn_mtx);
1671         return (i < TXG_SIZE);
1672 }
1673 
1674 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1675 uint64_t
1676 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1677 {
1678         void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1679         int i;
1680 
1681         if (blkid == DMU_BONUS_BLKID)
1682                 return (FALSE);
1683 
1684         /*
1685          * If we're in the process of opening the pool, dp will not be
1686          * set yet, but there shouldn't be anything dirty.
1687          */
1688         if (dp == NULL)
1689                 return (FALSE);
1690 
1691         if (dn->dn_free_txg)
1692                 return (TRUE);
1693 
1694         if (blkid == DMU_SPILL_BLKID)
1695                 return (dnode_spill_freed(dn));
1696 
1697         mutex_enter(&dn->dn_mtx);
1698         for (i = 0; i < TXG_SIZE; i++) {
1699                 if (dn->dn_free_ranges[i] != NULL &&
1700                     range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1701                         break;
1702         }
1703         mutex_exit(&dn->dn_mtx);
1704         return (i < TXG_SIZE);
1705 }
1706 
1707 /* call from syncing context when we actually write/free space for this dnode */
1708 void
1709 dnode_diduse_space(dnode_t *dn, int64_t delta)
1710 {
1711         uint64_t space;
1712         dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1713             dn, dn->dn_phys,
1714             (u_longlong_t)dn->dn_phys->dn_used,
1715             (longlong_t)delta);
1716 
1717         mutex_enter(&dn->dn_mtx);
1718         space = DN_USED_BYTES(dn->dn_phys);
1719         if (delta > 0) {
1720                 ASSERT3U(space + delta, >=, space); /* no overflow */
1721         } else {
1722                 ASSERT3U(space, >=, -delta); /* no underflow */
1723         }
1724         space += delta;
1725         if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1726                 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1727                 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1728                 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1729         } else {
1730                 dn->dn_phys->dn_used = space;
1731                 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1732         }
1733         mutex_exit(&dn->dn_mtx);
1734 }
1735 
1736 /*
1737  * Call when we think we're going to write/free space in open context to track
1738  * the amount of memory in use by the currently open txg.
1739  */
1740 void
1741 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1742 {
1743         objset_t *os = dn->dn_objset;
1744         dsl_dataset_t *ds = os->os_dsl_dataset;
1745         int64_t aspace = spa_get_asize(os->os_spa, space);
1746 
1747         if (ds != NULL) {
1748                 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1749                 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1750         }
1751 
1752         dmu_tx_willuse_space(tx, aspace);
1753 }
1754 
1755 /*
1756  * Scans a block at the indicated "level" looking for a hole or data,
1757  * depending on 'flags'.
1758  *
1759  * If level > 0, then we are scanning an indirect block looking at its
1760  * pointers.  If level == 0, then we are looking at a block of dnodes.
1761  *
1762  * If we don't find what we are looking for in the block, we return ESRCH.
1763  * Otherwise, return with *offset pointing to the beginning (if searching
1764  * forwards) or end (if searching backwards) of the range covered by the
1765  * block pointer we matched on (or dnode).
1766  *
1767  * The basic search algorithm used below by dnode_next_offset() is to
1768  * use this function to search up the block tree (widen the search) until
1769  * we find something (i.e., we don't return ESRCH) and then search back
1770  * down the tree (narrow the search) until we reach our original search
1771  * level.
1772  */
1773 static int
1774 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1775         int lvl, uint64_t blkfill, uint64_t txg)
1776 {
1777         dmu_buf_impl_t *db = NULL;
1778         void *data = NULL;
1779         uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1780         uint64_t epb = 1ULL << epbs;
1781         uint64_t minfill, maxfill;
1782         boolean_t hole;
1783         int i, inc, error, span;
1784 
1785         dprintf("probing object %llu offset %llx level %d of %u\n",
1786             dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1787 
1788         hole = ((flags & DNODE_FIND_HOLE) != 0);
1789         inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1790         ASSERT(txg == 0 || !hole);
1791 
1792         if (lvl == dn->dn_phys->dn_nlevels) {
1793                 error = 0;
1794                 epb = dn->dn_phys->dn_nblkptr;
1795                 data = dn->dn_phys->dn_blkptr;
1796         } else {
1797                 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1798                 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1799                 if (error) {
1800                         if (error != ENOENT)
1801                                 return (error);
1802                         if (hole)
1803                                 return (0);
1804                         /*
1805                          * This can only happen when we are searching up
1806                          * the block tree for data.  We don't really need to
1807                          * adjust the offset, as we will just end up looking
1808                          * at the pointer to this block in its parent, and its
1809                          * going to be unallocated, so we will skip over it.
1810                          */
1811                         return (SET_ERROR(ESRCH));
1812                 }
1813                 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1814                 if (error) {
1815                         dbuf_rele(db, FTAG);
1816                         return (error);
1817                 }
1818                 data = db->db.db_data;
1819         }
1820 
1821 
1822         if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1823             db->db_blkptr->blk_birth <= txg ||
1824             BP_IS_HOLE(db->db_blkptr))) {
1825                 /*
1826                  * This can only happen when we are searching up the tree
1827                  * and these conditions mean that we need to keep climbing.
1828                  */
1829                 error = SET_ERROR(ESRCH);
1830         } else if (lvl == 0) {
1831                 dnode_phys_t *dnp = data;
1832                 span = DNODE_SHIFT;
1833                 ASSERT(dn->dn_type == DMU_OT_DNODE);
1834 
1835                 for (i = (*offset >> span) & (blkfill - 1);
1836                     i >= 0 && i < blkfill; i += inc) {
1837                         if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1838                                 break;
1839                         *offset += (1ULL << span) * inc;
1840                 }
1841                 if (i < 0 || i == blkfill)
1842                         error = SET_ERROR(ESRCH);
1843         } else {
1844                 blkptr_t *bp = data;
1845                 uint64_t start = *offset;
1846                 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1847                 minfill = 0;
1848                 maxfill = blkfill << ((lvl - 1) * epbs);
1849 
1850                 if (hole)
1851                         maxfill--;
1852                 else
1853                         minfill++;
1854 
1855                 *offset = *offset >> span;
1856                 for (i = BF64_GET(*offset, 0, epbs);
1857                     i >= 0 && i < epb; i += inc) {
1858                         if (BP_GET_FILL(&bp[i]) >= minfill &&
1859                             BP_GET_FILL(&bp[i]) <= maxfill &&
1860                             (hole || bp[i].blk_birth > txg))
1861                                 break;
1862                         if (inc > 0 || *offset > 0)
1863                                 *offset += inc;
1864                 }
1865                 *offset = *offset << span;
1866                 if (inc < 0) {
1867                         /* traversing backwards; position offset at the end */
1868                         ASSERT3U(*offset, <=, start);
1869                         *offset = MIN(*offset + (1ULL << span) - 1, start);
1870                 } else if (*offset < start) {
1871                         *offset = start;
1872                 }
1873                 if (i < 0 || i >= epb)
1874                         error = SET_ERROR(ESRCH);
1875         }
1876 
1877         if (db)
1878                 dbuf_rele(db, FTAG);
1879 
1880         return (error);
1881 }
1882 
1883 /*
1884  * Find the next hole, data, or sparse region at or after *offset.
1885  * The value 'blkfill' tells us how many items we expect to find
1886  * in an L0 data block; this value is 1 for normal objects,
1887  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1888  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1889  *
1890  * Examples:
1891  *
1892  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1893  *      Finds the next/previous hole/data in a file.
1894  *      Used in dmu_offset_next().
1895  *
1896  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1897  *      Finds the next free/allocated dnode an objset's meta-dnode.
1898  *      Only finds objects that have new contents since txg (ie.
1899  *      bonus buffer changes and content removal are ignored).
1900  *      Used in dmu_object_next().
1901  *
1902  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1903  *      Finds the next L2 meta-dnode bp that's at most 1/4 full.
1904  *      Used in dmu_object_alloc().
1905  */
1906 int
1907 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1908     int minlvl, uint64_t blkfill, uint64_t txg)
1909 {
1910         uint64_t initial_offset = *offset;
1911         int lvl, maxlvl;
1912         int error = 0;
1913 
1914         if (!(flags & DNODE_FIND_HAVELOCK))
1915                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1916 
1917         if (dn->dn_phys->dn_nlevels == 0) {
1918                 error = SET_ERROR(ESRCH);
1919                 goto out;
1920         }
1921 
1922         if (dn->dn_datablkshift == 0) {
1923                 if (*offset < dn->dn_datablksz) {
1924                         if (flags & DNODE_FIND_HOLE)
1925                                 *offset = dn->dn_datablksz;
1926                 } else {
1927                         error = SET_ERROR(ESRCH);
1928                 }
1929                 goto out;
1930         }
1931 
1932         maxlvl = dn->dn_phys->dn_nlevels;
1933 
1934         for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1935                 error = dnode_next_offset_level(dn,
1936                     flags, offset, lvl, blkfill, txg);
1937                 if (error != ESRCH)
1938                         break;
1939         }
1940 
1941         while (error == 0 && --lvl >= minlvl) {
1942                 error = dnode_next_offset_level(dn,
1943                     flags, offset, lvl, blkfill, txg);
1944         }
1945 
1946         /*
1947          * There's always a "virtual hole" at the end of the object, even
1948          * if all BP's which physically exist are non-holes.
1949          */
1950         if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1951             minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1952                 error = 0;
1953         }
1954 
1955         if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1956             initial_offset < *offset : initial_offset > *offset))
1957                 error = SET_ERROR(ESRCH);
1958 out:
1959         if (!(flags & DNODE_FIND_HAVELOCK))
1960                 rw_exit(&dn->dn_struct_rwlock);
1961 
1962         return (error);
1963 }