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, 2017 by Delphix. All rights reserved.
  24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  25  * Copyright (c) 2014 Integros [integros.com]
  26  * Copyright 2017 RackTop Systems.
  27  */
  28 
  29 #include <sys/zfs_context.h>
  30 #include <sys/dbuf.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu.h>
  33 #include <sys/dmu_impl.h>
  34 #include <sys/dmu_tx.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/dsl_dir.h>
  37 #include <sys/dsl_dataset.h>
  38 #include <sys/spa.h>
  39 #include <sys/zio.h>
  40 #include <sys/dmu_zfetch.h>
  41 #include <sys/range_tree.h>
  42 
  43 static kmem_cache_t *dnode_cache;
  44 /*
  45  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
  46  * turned on when DEBUG is also defined.
  47  */
  48 #ifdef  DEBUG
  49 #define DNODE_STATS
  50 #endif  /* DEBUG */
  51 
  52 #ifdef  DNODE_STATS
  53 #define DNODE_STAT_ADD(stat)                    ((stat)++)
  54 #else
  55 #define DNODE_STAT_ADD(stat)                    /* nothing */
  56 #endif  /* DNODE_STATS */
  57 
  58 static dnode_phys_t dnode_phys_zero;
  59 
  60 int zfs_default_bs = SPA_MINBLOCKSHIFT;
  61 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
  62 
  63 #ifdef  _KERNEL
  64 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
  65 #endif  /* _KERNEL */
  66 
  67 static int
  68 dbuf_compare(const void *x1, const void *x2)
  69 {
  70         const dmu_buf_impl_t *d1 = x1;
  71         const dmu_buf_impl_t *d2 = x2;
  72 
  73         if (d1->db_level < d2->db_level) {
  74                 return (-1);
  75         }
  76         if (d1->db_level > d2->db_level) {
  77                 return (1);
  78         }
  79 
  80         if (d1->db_blkid < d2->db_blkid) {
  81                 return (-1);
  82         }
  83         if (d1->db_blkid > d2->db_blkid) {
  84                 return (1);
  85         }
  86 
  87         if (d1->db_state == DB_SEARCH) {
  88                 ASSERT3S(d2->db_state, !=, DB_SEARCH);
  89                 return (-1);
  90         } else if (d2->db_state == DB_SEARCH) {
  91                 ASSERT3S(d1->db_state, !=, DB_SEARCH);
  92                 return (1);
  93         }
  94 
  95         if ((uintptr_t)d1 < (uintptr_t)d2) {
  96                 return (-1);
  97         }
  98         if ((uintptr_t)d1 > (uintptr_t)d2) {
  99                 return (1);
 100         }
 101         return (0);
 102 }
 103 
 104 /* ARGSUSED */
 105 static int
 106 dnode_cons(void *arg, void *unused, int kmflag)
 107 {
 108         dnode_t *dn = arg;
 109         int i;
 110 
 111         rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
 112         mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
 113         mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
 114         cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
 115 
 116         /*
 117          * Every dbuf has a reference, and dropping a tracked reference is
 118          * O(number of references), so don't track dn_holds.
 119          */
 120         refcount_create_untracked(&dn->dn_holds);
 121         refcount_create(&dn->dn_tx_holds);
 122         list_link_init(&dn->dn_link);
 123 
 124         bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
 125         bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
 126         bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
 127         bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
 128         bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
 129         bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
 130         bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
 131 
 132         for (i = 0; i < TXG_SIZE; i++) {
 133                 list_link_init(&dn->dn_dirty_link[i]);
 134                 dn->dn_free_ranges[i] = NULL;
 135                 list_create(&dn->dn_dirty_records[i],
 136                     sizeof (dbuf_dirty_record_t),
 137                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 138         }
 139 
 140         dn->dn_allocated_txg = 0;
 141         dn->dn_free_txg = 0;
 142         dn->dn_assigned_txg = 0;
 143         dn->dn_dirtyctx = 0;
 144         dn->dn_dirtyctx_firstset = NULL;
 145         dn->dn_bonus = NULL;
 146         dn->dn_have_spill = B_FALSE;
 147         dn->dn_zio = NULL;
 148         dn->dn_oldused = 0;
 149         dn->dn_oldflags = 0;
 150         dn->dn_olduid = 0;
 151         dn->dn_oldgid = 0;
 152         dn->dn_newuid = 0;
 153         dn->dn_newgid = 0;
 154         dn->dn_id_flags = 0;
 155 
 156         dn->dn_dbufs_count = 0;
 157         avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
 158             offsetof(dmu_buf_impl_t, db_link));
 159 
 160         dn->dn_moved = 0;
 161         return (0);
 162 }
 163 
 164 /* ARGSUSED */
 165 static void
 166 dnode_dest(void *arg, void *unused)
 167 {
 168         int i;
 169         dnode_t *dn = arg;
 170 
 171         rw_destroy(&dn->dn_struct_rwlock);
 172         mutex_destroy(&dn->dn_mtx);
 173         mutex_destroy(&dn->dn_dbufs_mtx);
 174         cv_destroy(&dn->dn_notxholds);
 175         refcount_destroy(&dn->dn_holds);
 176         refcount_destroy(&dn->dn_tx_holds);
 177         ASSERT(!list_link_active(&dn->dn_link));
 178 
 179         for (i = 0; i < TXG_SIZE; i++) {
 180                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 181                 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
 182                 list_destroy(&dn->dn_dirty_records[i]);
 183                 ASSERT0(dn->dn_next_nblkptr[i]);
 184                 ASSERT0(dn->dn_next_nlevels[i]);
 185                 ASSERT0(dn->dn_next_indblkshift[i]);
 186                 ASSERT0(dn->dn_next_bonustype[i]);
 187                 ASSERT0(dn->dn_rm_spillblk[i]);
 188                 ASSERT0(dn->dn_next_bonuslen[i]);
 189                 ASSERT0(dn->dn_next_blksz[i]);
 190         }
 191 
 192         ASSERT0(dn->dn_allocated_txg);
 193         ASSERT0(dn->dn_free_txg);
 194         ASSERT0(dn->dn_assigned_txg);
 195         ASSERT0(dn->dn_dirtyctx);
 196         ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
 197         ASSERT3P(dn->dn_bonus, ==, NULL);
 198         ASSERT(!dn->dn_have_spill);
 199         ASSERT3P(dn->dn_zio, ==, NULL);
 200         ASSERT0(dn->dn_oldused);
 201         ASSERT0(dn->dn_oldflags);
 202         ASSERT0(dn->dn_olduid);
 203         ASSERT0(dn->dn_oldgid);
 204         ASSERT0(dn->dn_newuid);
 205         ASSERT0(dn->dn_newgid);
 206         ASSERT0(dn->dn_id_flags);
 207 
 208         ASSERT0(dn->dn_dbufs_count);
 209         avl_destroy(&dn->dn_dbufs);
 210 }
 211 
 212 void
 213 dnode_init(void)
 214 {
 215         ASSERT(dnode_cache == NULL);
 216         dnode_cache = kmem_cache_create("dnode_t",
 217             sizeof (dnode_t),
 218             0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
 219 #ifdef  _KERNEL
 220         kmem_cache_set_move(dnode_cache, dnode_move);
 221 #endif  /* _KERNEL */
 222 }
 223 
 224 void
 225 dnode_fini(void)
 226 {
 227         kmem_cache_destroy(dnode_cache);
 228         dnode_cache = NULL;
 229 }
 230 
 231 
 232 #ifdef ZFS_DEBUG
 233 void
 234 dnode_verify(dnode_t *dn)
 235 {
 236         int drop_struct_lock = FALSE;
 237 
 238         ASSERT(dn->dn_phys);
 239         ASSERT(dn->dn_objset);
 240         ASSERT(dn->dn_handle->dnh_dnode == dn);
 241 
 242         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 243 
 244         if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
 245                 return;
 246 
 247         if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
 248                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 249                 drop_struct_lock = TRUE;
 250         }
 251         if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
 252                 int i;
 253                 ASSERT3U(dn->dn_indblkshift, >=, 0);
 254                 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
 255                 if (dn->dn_datablkshift) {
 256                         ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
 257                         ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
 258                         ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
 259                 }
 260                 ASSERT3U(dn->dn_nlevels, <=, 30);
 261                 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
 262                 ASSERT3U(dn->dn_nblkptr, >=, 1);
 263                 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 264                 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 265                 ASSERT3U(dn->dn_datablksz, ==,
 266                     dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 267                 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
 268                 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
 269                     dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 270                 for (i = 0; i < TXG_SIZE; i++) {
 271                         ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
 272                 }
 273         }
 274         if (dn->dn_phys->dn_type != DMU_OT_NONE)
 275                 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
 276         ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
 277         if (dn->dn_dbuf != NULL) {
 278                 ASSERT3P(dn->dn_phys, ==,
 279                     (dnode_phys_t *)dn->dn_dbuf->db.db_data +
 280                     (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
 281         }
 282         if (drop_struct_lock)
 283                 rw_exit(&dn->dn_struct_rwlock);
 284 }
 285 #endif
 286 
 287 void
 288 dnode_byteswap(dnode_phys_t *dnp)
 289 {
 290         uint64_t *buf64 = (void*)&dnp->dn_blkptr;
 291         int i;
 292 
 293         if (dnp->dn_type == DMU_OT_NONE) {
 294                 bzero(dnp, sizeof (dnode_phys_t));
 295                 return;
 296         }
 297 
 298         dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
 299         dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
 300         dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
 301         dnp->dn_used = BSWAP_64(dnp->dn_used);
 302 
 303         /*
 304          * dn_nblkptr is only one byte, so it's OK to read it in either
 305          * byte order.  We can't read dn_bouslen.
 306          */
 307         ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
 308         ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
 309         for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
 310                 buf64[i] = BSWAP_64(buf64[i]);
 311 
 312         /*
 313          * OK to check dn_bonuslen for zero, because it won't matter if
 314          * we have the wrong byte order.  This is necessary because the
 315          * dnode dnode is smaller than a regular dnode.
 316          */
 317         if (dnp->dn_bonuslen != 0) {
 318                 /*
 319                  * Note that the bonus length calculated here may be
 320                  * longer than the actual bonus buffer.  This is because
 321                  * we always put the bonus buffer after the last block
 322                  * pointer (instead of packing it against the end of the
 323                  * dnode buffer).
 324                  */
 325                 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
 326                 size_t len = DN_MAX_BONUSLEN - off;
 327                 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
 328                 dmu_object_byteswap_t byteswap =
 329                     DMU_OT_BYTESWAP(dnp->dn_bonustype);
 330                 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
 331         }
 332 
 333         /* Swap SPILL block if we have one */
 334         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 335                 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
 336 
 337 }
 338 
 339 void
 340 dnode_buf_byteswap(void *vbuf, size_t size)
 341 {
 342         dnode_phys_t *buf = vbuf;
 343         int i;
 344 
 345         ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
 346         ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
 347 
 348         size >>= DNODE_SHIFT;
 349         for (i = 0; i < size; i++) {
 350                 dnode_byteswap(buf);
 351                 buf++;
 352         }
 353 }
 354 
 355 void
 356 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
 357 {
 358         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 359 
 360         dnode_setdirty(dn, tx);
 361         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 362         ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
 363             (dn->dn_nblkptr-1) * sizeof (blkptr_t));
 364         dn->dn_bonuslen = newsize;
 365         if (newsize == 0)
 366                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
 367         else
 368                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 369         rw_exit(&dn->dn_struct_rwlock);
 370 }
 371 
 372 void
 373 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
 374 {
 375         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 376         dnode_setdirty(dn, tx);
 377         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 378         dn->dn_bonustype = newtype;
 379         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 380         rw_exit(&dn->dn_struct_rwlock);
 381 }
 382 
 383 void
 384 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
 385 {
 386         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 387         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
 388         dnode_setdirty(dn, tx);
 389         dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
 390         dn->dn_have_spill = B_FALSE;
 391 }
 392 
 393 static void
 394 dnode_setdblksz(dnode_t *dn, int size)
 395 {
 396         ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
 397         ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
 398         ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
 399         ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
 400             1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
 401         dn->dn_datablksz = size;
 402         dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
 403         dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
 404 }
 405 
 406 static dnode_t *
 407 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
 408     uint64_t object, dnode_handle_t *dnh)
 409 {
 410         dnode_t *dn;
 411 
 412         dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
 413 #ifdef _KERNEL
 414         ASSERT(!POINTER_IS_VALID(dn->dn_objset));
 415 #endif /* _KERNEL */
 416         dn->dn_moved = 0;
 417 
 418         /*
 419          * Defer setting dn_objset until the dnode is ready to be a candidate
 420          * for the dnode_move() callback.
 421          */
 422         dn->dn_object = object;
 423         dn->dn_dbuf = db;
 424         dn->dn_handle = dnh;
 425         dn->dn_phys = dnp;
 426 
 427         if (dnp->dn_datablkszsec) {
 428                 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 429         } else {
 430                 dn->dn_datablksz = 0;
 431                 dn->dn_datablkszsec = 0;
 432                 dn->dn_datablkshift = 0;
 433         }
 434         dn->dn_indblkshift = dnp->dn_indblkshift;
 435         dn->dn_nlevels = dnp->dn_nlevels;
 436         dn->dn_type = dnp->dn_type;
 437         dn->dn_nblkptr = dnp->dn_nblkptr;
 438         dn->dn_checksum = dnp->dn_checksum;
 439         dn->dn_compress = dnp->dn_compress;
 440         dn->dn_bonustype = dnp->dn_bonustype;
 441         dn->dn_bonuslen = dnp->dn_bonuslen;
 442         dn->dn_maxblkid = dnp->dn_maxblkid;
 443         dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
 444         dn->dn_id_flags = 0;
 445 
 446         dmu_zfetch_init(&dn->dn_zfetch, dn);
 447 
 448         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 449 
 450         mutex_enter(&os->os_lock);
 451         if (dnh->dnh_dnode != NULL) {
 452                 /* Lost the allocation race. */
 453                 mutex_exit(&os->os_lock);
 454                 kmem_cache_free(dnode_cache, dn);
 455                 return (dnh->dnh_dnode);
 456         }
 457 
 458         /*
 459          * Exclude special dnodes from os_dnodes so an empty os_dnodes
 460          * signifies that the special dnodes have no references from
 461          * their children (the entries in os_dnodes).  This allows
 462          * dnode_destroy() to easily determine if the last child has
 463          * been removed and then complete eviction of the objset.
 464          */
 465         if (!DMU_OBJECT_IS_SPECIAL(object))
 466                 list_insert_head(&os->os_dnodes, dn);
 467         membar_producer();
 468 
 469         /*
 470          * Everything else must be valid before assigning dn_objset
 471          * makes the dnode eligible for dnode_move().
 472          */
 473         dn->dn_objset = os;
 474 
 475         dnh->dnh_dnode = dn;
 476         mutex_exit(&os->os_lock);
 477 
 478         arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
 479         return (dn);
 480 }
 481 
 482 /*
 483  * Caller must be holding the dnode handle, which is released upon return.
 484  */
 485 static void
 486 dnode_destroy(dnode_t *dn)
 487 {
 488         objset_t *os = dn->dn_objset;
 489         boolean_t complete_os_eviction = B_FALSE;
 490 
 491         ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
 492 
 493         mutex_enter(&os->os_lock);
 494         POINTER_INVALIDATE(&dn->dn_objset);
 495         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 496                 list_remove(&os->os_dnodes, dn);
 497                 complete_os_eviction =
 498                     list_is_empty(&os->os_dnodes) &&
 499                     list_link_active(&os->os_evicting_node);
 500         }
 501         mutex_exit(&os->os_lock);
 502 
 503         /* the dnode can no longer move, so we can release the handle */
 504         zrl_remove(&dn->dn_handle->dnh_zrlock);
 505 
 506         dn->dn_allocated_txg = 0;
 507         dn->dn_free_txg = 0;
 508         dn->dn_assigned_txg = 0;
 509 
 510         dn->dn_dirtyctx = 0;
 511         if (dn->dn_dirtyctx_firstset != NULL) {
 512                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 513                 dn->dn_dirtyctx_firstset = NULL;
 514         }
 515         if (dn->dn_bonus != NULL) {
 516                 mutex_enter(&dn->dn_bonus->db_mtx);
 517                 dbuf_destroy(dn->dn_bonus);
 518                 dn->dn_bonus = NULL;
 519         }
 520         dn->dn_zio = NULL;
 521 
 522         dn->dn_have_spill = B_FALSE;
 523         dn->dn_oldused = 0;
 524         dn->dn_oldflags = 0;
 525         dn->dn_olduid = 0;
 526         dn->dn_oldgid = 0;
 527         dn->dn_newuid = 0;
 528         dn->dn_newgid = 0;
 529         dn->dn_id_flags = 0;
 530 
 531         dmu_zfetch_fini(&dn->dn_zfetch);
 532         kmem_cache_free(dnode_cache, dn);
 533         arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
 534 
 535         if (complete_os_eviction)
 536                 dmu_objset_evict_done(os);
 537 }
 538 
 539 void
 540 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
 541     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 542 {
 543         int i;
 544 
 545         ASSERT3U(blocksize, <=,
 546             spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 547         if (blocksize == 0)
 548                 blocksize = 1 << zfs_default_bs;
 549         else
 550                 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
 551 
 552         if (ibs == 0)
 553                 ibs = zfs_default_ibs;
 554 
 555         ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
 556 
 557         dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
 558             dn->dn_object, tx->tx_txg, blocksize, ibs);
 559 
 560         ASSERT(dn->dn_type == DMU_OT_NONE);
 561         ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
 562         ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
 563         ASSERT(ot != DMU_OT_NONE);
 564         ASSERT(DMU_OT_IS_VALID(ot));
 565         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 566             (bonustype == DMU_OT_SA && bonuslen == 0) ||
 567             (bonustype != DMU_OT_NONE && bonuslen != 0));
 568         ASSERT(DMU_OT_IS_VALID(bonustype));
 569         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 570         ASSERT(dn->dn_type == DMU_OT_NONE);
 571         ASSERT0(dn->dn_maxblkid);
 572         ASSERT0(dn->dn_allocated_txg);
 573         ASSERT0(dn->dn_assigned_txg);
 574         ASSERT(refcount_is_zero(&dn->dn_tx_holds));
 575         ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
 576         ASSERT(avl_is_empty(&dn->dn_dbufs));
 577 
 578         for (i = 0; i < TXG_SIZE; i++) {
 579                 ASSERT0(dn->dn_next_nblkptr[i]);
 580                 ASSERT0(dn->dn_next_nlevels[i]);
 581                 ASSERT0(dn->dn_next_indblkshift[i]);
 582                 ASSERT0(dn->dn_next_bonuslen[i]);
 583                 ASSERT0(dn->dn_next_bonustype[i]);
 584                 ASSERT0(dn->dn_rm_spillblk[i]);
 585                 ASSERT0(dn->dn_next_blksz[i]);
 586                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 587                 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
 588                 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
 589         }
 590 
 591         dn->dn_type = ot;
 592         dnode_setdblksz(dn, blocksize);
 593         dn->dn_indblkshift = ibs;
 594         dn->dn_nlevels = 1;
 595         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 596                 dn->dn_nblkptr = 1;
 597         else
 598                 dn->dn_nblkptr = 1 +
 599                     ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 600         dn->dn_bonustype = bonustype;
 601         dn->dn_bonuslen = bonuslen;
 602         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 603         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 604         dn->dn_dirtyctx = 0;
 605 
 606         dn->dn_free_txg = 0;
 607         if (dn->dn_dirtyctx_firstset) {
 608                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 609                 dn->dn_dirtyctx_firstset = NULL;
 610         }
 611 
 612         dn->dn_allocated_txg = tx->tx_txg;
 613         dn->dn_id_flags = 0;
 614 
 615         dnode_setdirty(dn, tx);
 616         dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
 617         dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 618         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 619         dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
 620 }
 621 
 622 void
 623 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
 624     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 625 {
 626         int nblkptr;
 627 
 628         ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
 629         ASSERT3U(blocksize, <=,
 630             spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 631         ASSERT0(blocksize % SPA_MINBLOCKSIZE);
 632         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
 633         ASSERT(tx->tx_txg != 0);
 634         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 635             (bonustype != DMU_OT_NONE && bonuslen != 0) ||
 636             (bonustype == DMU_OT_SA && bonuslen == 0));
 637         ASSERT(DMU_OT_IS_VALID(bonustype));
 638         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 639 
 640         /* clean up any unreferenced dbufs */
 641         dnode_evict_dbufs(dn);
 642 
 643         dn->dn_id_flags = 0;
 644 
 645         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 646         dnode_setdirty(dn, tx);
 647         if (dn->dn_datablksz != blocksize) {
 648                 /* change blocksize */
 649                 ASSERT(dn->dn_maxblkid == 0 &&
 650                     (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
 651                     dnode_block_freed(dn, 0)));
 652                 dnode_setdblksz(dn, blocksize);
 653                 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
 654         }
 655         if (dn->dn_bonuslen != bonuslen)
 656                 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
 657 
 658         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 659                 nblkptr = 1;
 660         else
 661                 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 662         if (dn->dn_bonustype != bonustype)
 663                 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
 664         if (dn->dn_nblkptr != nblkptr)
 665                 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
 666         if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 667                 dbuf_rm_spill(dn, tx);
 668                 dnode_rm_spill(dn, tx);
 669         }
 670         rw_exit(&dn->dn_struct_rwlock);
 671 
 672         /* change type */
 673         dn->dn_type = ot;
 674 
 675         /* change bonus size and type */
 676         mutex_enter(&dn->dn_mtx);
 677         dn->dn_bonustype = bonustype;
 678         dn->dn_bonuslen = bonuslen;
 679         dn->dn_nblkptr = nblkptr;
 680         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 681         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 682         ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 683 
 684         /* fix up the bonus db_size */
 685         if (dn->dn_bonus) {
 686                 dn->dn_bonus->db.db_size =
 687                     DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
 688                 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
 689         }
 690 
 691         dn->dn_allocated_txg = tx->tx_txg;
 692         mutex_exit(&dn->dn_mtx);
 693 }
 694 
 695 #ifdef  DNODE_STATS
 696 static struct {
 697         uint64_t dms_dnode_invalid;
 698         uint64_t dms_dnode_recheck1;
 699         uint64_t dms_dnode_recheck2;
 700         uint64_t dms_dnode_special;
 701         uint64_t dms_dnode_handle;
 702         uint64_t dms_dnode_rwlock;
 703         uint64_t dms_dnode_active;
 704 } dnode_move_stats;
 705 #endif  /* DNODE_STATS */
 706 
 707 #ifdef  _KERNEL
 708 static void
 709 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
 710 {
 711         int i;
 712 
 713         ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
 714         ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
 715         ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
 716         ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
 717 
 718         /* Copy fields. */
 719         ndn->dn_objset = odn->dn_objset;
 720         ndn->dn_object = odn->dn_object;
 721         ndn->dn_dbuf = odn->dn_dbuf;
 722         ndn->dn_handle = odn->dn_handle;
 723         ndn->dn_phys = odn->dn_phys;
 724         ndn->dn_type = odn->dn_type;
 725         ndn->dn_bonuslen = odn->dn_bonuslen;
 726         ndn->dn_bonustype = odn->dn_bonustype;
 727         ndn->dn_nblkptr = odn->dn_nblkptr;
 728         ndn->dn_checksum = odn->dn_checksum;
 729         ndn->dn_compress = odn->dn_compress;
 730         ndn->dn_nlevels = odn->dn_nlevels;
 731         ndn->dn_indblkshift = odn->dn_indblkshift;
 732         ndn->dn_datablkshift = odn->dn_datablkshift;
 733         ndn->dn_datablkszsec = odn->dn_datablkszsec;
 734         ndn->dn_datablksz = odn->dn_datablksz;
 735         ndn->dn_maxblkid = odn->dn_maxblkid;
 736         bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
 737             sizeof (odn->dn_next_nblkptr));
 738         bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
 739             sizeof (odn->dn_next_nlevels));
 740         bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
 741             sizeof (odn->dn_next_indblkshift));
 742         bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
 743             sizeof (odn->dn_next_bonustype));
 744         bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
 745             sizeof (odn->dn_rm_spillblk));
 746         bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
 747             sizeof (odn->dn_next_bonuslen));
 748         bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
 749             sizeof (odn->dn_next_blksz));
 750         for (i = 0; i < TXG_SIZE; i++) {
 751                 list_move_tail(&ndn->dn_dirty_records[i],
 752                     &odn->dn_dirty_records[i]);
 753         }
 754         bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
 755             sizeof (odn->dn_free_ranges));
 756         ndn->dn_allocated_txg = odn->dn_allocated_txg;
 757         ndn->dn_free_txg = odn->dn_free_txg;
 758         ndn->dn_assigned_txg = odn->dn_assigned_txg;
 759         ndn->dn_dirtyctx = odn->dn_dirtyctx;
 760         ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
 761         ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
 762         refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
 763         ASSERT(avl_is_empty(&ndn->dn_dbufs));
 764         avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
 765         ndn->dn_dbufs_count = odn->dn_dbufs_count;
 766         ndn->dn_bonus = odn->dn_bonus;
 767         ndn->dn_have_spill = odn->dn_have_spill;
 768         ndn->dn_zio = odn->dn_zio;
 769         ndn->dn_oldused = odn->dn_oldused;
 770         ndn->dn_oldflags = odn->dn_oldflags;
 771         ndn->dn_olduid = odn->dn_olduid;
 772         ndn->dn_oldgid = odn->dn_oldgid;
 773         ndn->dn_newuid = odn->dn_newuid;
 774         ndn->dn_newgid = odn->dn_newgid;
 775         ndn->dn_id_flags = odn->dn_id_flags;
 776         dmu_zfetch_init(&ndn->dn_zfetch, NULL);
 777         list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
 778         ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
 779 
 780         /*
 781          * Update back pointers. Updating the handle fixes the back pointer of
 782          * every descendant dbuf as well as the bonus dbuf.
 783          */
 784         ASSERT(ndn->dn_handle->dnh_dnode == odn);
 785         ndn->dn_handle->dnh_dnode = ndn;
 786         if (ndn->dn_zfetch.zf_dnode == odn) {
 787                 ndn->dn_zfetch.zf_dnode = ndn;
 788         }
 789 
 790         /*
 791          * Invalidate the original dnode by clearing all of its back pointers.
 792          */
 793         odn->dn_dbuf = NULL;
 794         odn->dn_handle = NULL;
 795         avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
 796             offsetof(dmu_buf_impl_t, db_link));
 797         odn->dn_dbufs_count = 0;
 798         odn->dn_bonus = NULL;
 799         odn->dn_zfetch.zf_dnode = NULL;
 800 
 801         /*
 802          * Set the low bit of the objset pointer to ensure that dnode_move()
 803          * recognizes the dnode as invalid in any subsequent callback.
 804          */
 805         POINTER_INVALIDATE(&odn->dn_objset);
 806 
 807         /*
 808          * Satisfy the destructor.
 809          */
 810         for (i = 0; i < TXG_SIZE; i++) {
 811                 list_create(&odn->dn_dirty_records[i],
 812                     sizeof (dbuf_dirty_record_t),
 813                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 814                 odn->dn_free_ranges[i] = NULL;
 815                 odn->dn_next_nlevels[i] = 0;
 816                 odn->dn_next_indblkshift[i] = 0;
 817                 odn->dn_next_bonustype[i] = 0;
 818                 odn->dn_rm_spillblk[i] = 0;
 819                 odn->dn_next_bonuslen[i] = 0;
 820                 odn->dn_next_blksz[i] = 0;
 821         }
 822         odn->dn_allocated_txg = 0;
 823         odn->dn_free_txg = 0;
 824         odn->dn_assigned_txg = 0;
 825         odn->dn_dirtyctx = 0;
 826         odn->dn_dirtyctx_firstset = NULL;
 827         odn->dn_have_spill = B_FALSE;
 828         odn->dn_zio = NULL;
 829         odn->dn_oldused = 0;
 830         odn->dn_oldflags = 0;
 831         odn->dn_olduid = 0;
 832         odn->dn_oldgid = 0;
 833         odn->dn_newuid = 0;
 834         odn->dn_newgid = 0;
 835         odn->dn_id_flags = 0;
 836 
 837         /*
 838          * Mark the dnode.
 839          */
 840         ndn->dn_moved = 1;
 841         odn->dn_moved = (uint8_t)-1;
 842 }
 843 
 844 /*ARGSUSED*/
 845 static kmem_cbrc_t
 846 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
 847 {
 848         dnode_t *odn = buf, *ndn = newbuf;
 849         objset_t *os;
 850         int64_t refcount;
 851         uint32_t dbufs;
 852 
 853         /*
 854          * The dnode is on the objset's list of known dnodes if the objset
 855          * pointer is valid. We set the low bit of the objset pointer when
 856          * freeing the dnode to invalidate it, and the memory patterns written
 857          * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
 858          * A newly created dnode sets the objset pointer last of all to indicate
 859          * that the dnode is known and in a valid state to be moved by this
 860          * function.
 861          */
 862         os = odn->dn_objset;
 863         if (!POINTER_IS_VALID(os)) {
 864                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
 865                 return (KMEM_CBRC_DONT_KNOW);
 866         }
 867 
 868         /*
 869          * Ensure that the objset does not go away during the move.
 870          */
 871         rw_enter(&os_lock, RW_WRITER);
 872         if (os != odn->dn_objset) {
 873                 rw_exit(&os_lock);
 874                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
 875                 return (KMEM_CBRC_DONT_KNOW);
 876         }
 877 
 878         /*
 879          * If the dnode is still valid, then so is the objset. We know that no
 880          * valid objset can be freed while we hold os_lock, so we can safely
 881          * ensure that the objset remains in use.
 882          */
 883         mutex_enter(&os->os_lock);
 884 
 885         /*
 886          * Recheck the objset pointer in case the dnode was removed just before
 887          * acquiring the lock.
 888          */
 889         if (os != odn->dn_objset) {
 890                 mutex_exit(&os->os_lock);
 891                 rw_exit(&os_lock);
 892                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
 893                 return (KMEM_CBRC_DONT_KNOW);
 894         }
 895 
 896         /*
 897          * At this point we know that as long as we hold os->os_lock, the dnode
 898          * cannot be freed and fields within the dnode can be safely accessed.
 899          * The objset listing this dnode cannot go away as long as this dnode is
 900          * on its list.
 901          */
 902         rw_exit(&os_lock);
 903         if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
 904                 mutex_exit(&os->os_lock);
 905                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
 906                 return (KMEM_CBRC_NO);
 907         }
 908         ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
 909 
 910         /*
 911          * Lock the dnode handle to prevent the dnode from obtaining any new
 912          * holds. This also prevents the descendant dbufs and the bonus dbuf
 913          * from accessing the dnode, so that we can discount their holds. The
 914          * handle is safe to access because we know that while the dnode cannot
 915          * go away, neither can its handle. Once we hold dnh_zrlock, we can
 916          * safely move any dnode referenced only by dbufs.
 917          */
 918         if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
 919                 mutex_exit(&os->os_lock);
 920                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
 921                 return (KMEM_CBRC_LATER);
 922         }
 923 
 924         /*
 925          * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
 926          * We need to guarantee that there is a hold for every dbuf in order to
 927          * determine whether the dnode is actively referenced. Falsely matching
 928          * a dbuf to an active hold would lead to an unsafe move. It's possible
 929          * that a thread already having an active dnode hold is about to add a
 930          * dbuf, and we can't compare hold and dbuf counts while the add is in
 931          * progress.
 932          */
 933         if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
 934                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 935                 mutex_exit(&os->os_lock);
 936                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
 937                 return (KMEM_CBRC_LATER);
 938         }
 939 
 940         /*
 941          * A dbuf may be removed (evicted) without an active dnode hold. In that
 942          * case, the dbuf count is decremented under the handle lock before the
 943          * dbuf's hold is released. This order ensures that if we count the hold
 944          * after the dbuf is removed but before its hold is released, we will
 945          * treat the unmatched hold as active and exit safely. If we count the
 946          * hold before the dbuf is removed, the hold is discounted, and the
 947          * removal is blocked until the move completes.
 948          */
 949         refcount = refcount_count(&odn->dn_holds);
 950         ASSERT(refcount >= 0);
 951         dbufs = odn->dn_dbufs_count;
 952 
 953         /* We can't have more dbufs than dnode holds. */
 954         ASSERT3U(dbufs, <=, refcount);
 955         DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
 956             uint32_t, dbufs);
 957 
 958         if (refcount > dbufs) {
 959                 rw_exit(&odn->dn_struct_rwlock);
 960                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 961                 mutex_exit(&os->os_lock);
 962                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
 963                 return (KMEM_CBRC_LATER);
 964         }
 965 
 966         rw_exit(&odn->dn_struct_rwlock);
 967 
 968         /*
 969          * At this point we know that anyone with a hold on the dnode is not
 970          * actively referencing it. The dnode is known and in a valid state to
 971          * move. We're holding the locks needed to execute the critical section.
 972          */
 973         dnode_move_impl(odn, ndn);
 974 
 975         list_link_replace(&odn->dn_link, &ndn->dn_link);
 976         /* If the dnode was safe to move, the refcount cannot have changed. */
 977         ASSERT(refcount == refcount_count(&ndn->dn_holds));
 978         ASSERT(dbufs == ndn->dn_dbufs_count);
 979         zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
 980         mutex_exit(&os->os_lock);
 981 
 982         return (KMEM_CBRC_YES);
 983 }
 984 #endif  /* _KERNEL */
 985 
 986 void
 987 dnode_special_close(dnode_handle_t *dnh)
 988 {
 989         dnode_t *dn = dnh->dnh_dnode;
 990 
 991         /*
 992          * Wait for final references to the dnode to clear.  This can
 993          * only happen if the arc is asyncronously evicting state that
 994          * has a hold on this dnode while we are trying to evict this
 995          * dnode.
 996          */
 997         while (refcount_count(&dn->dn_holds) > 0)
 998                 delay(1);
 999         ASSERT(dn->dn_dbuf == NULL ||
1000             dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1001         zrl_add(&dnh->dnh_zrlock);
1002         dnode_destroy(dn); /* implicit zrl_remove() */
1003         zrl_destroy(&dnh->dnh_zrlock);
1004         dnh->dnh_dnode = NULL;
1005 }
1006 
1007 void
1008 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1009     dnode_handle_t *dnh)
1010 {
1011         dnode_t *dn;
1012 
1013         dn = dnode_create(os, dnp, NULL, object, dnh);
1014         zrl_init(&dnh->dnh_zrlock);
1015         DNODE_VERIFY(dn);
1016 }
1017 
1018 static void
1019 dnode_buf_evict_async(void *dbu)
1020 {
1021         dnode_children_t *children_dnodes = dbu;
1022         int i;
1023 
1024         for (i = 0; i < children_dnodes->dnc_count; i++) {
1025                 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1026                 dnode_t *dn;
1027 
1028                 /*
1029                  * The dnode handle lock guards against the dnode moving to
1030                  * another valid address, so there is no need here to guard
1031                  * against changes to or from NULL.
1032                  */
1033                 if (dnh->dnh_dnode == NULL) {
1034                         zrl_destroy(&dnh->dnh_zrlock);
1035                         continue;
1036                 }
1037 
1038                 zrl_add(&dnh->dnh_zrlock);
1039                 dn = dnh->dnh_dnode;
1040                 /*
1041                  * If there are holds on this dnode, then there should
1042                  * be holds on the dnode's containing dbuf as well; thus
1043                  * it wouldn't be eligible for eviction and this function
1044                  * would not have been called.
1045                  */
1046                 ASSERT(refcount_is_zero(&dn->dn_holds));
1047                 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1048 
1049                 dnode_destroy(dn); /* implicit zrl_remove() */
1050                 zrl_destroy(&dnh->dnh_zrlock);
1051                 dnh->dnh_dnode = NULL;
1052         }
1053         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1054             children_dnodes->dnc_count * sizeof (dnode_handle_t));
1055 }
1056 
1057 /*
1058  * errors:
1059  * EINVAL - invalid object number.
1060  * EIO - i/o error.
1061  * succeeds even for free dnodes.
1062  */
1063 int
1064 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1065     void *tag, dnode_t **dnp)
1066 {
1067         int epb, idx, err;
1068         int drop_struct_lock = FALSE;
1069         int type;
1070         uint64_t blk;
1071         dnode_t *mdn, *dn;
1072         dmu_buf_impl_t *db;
1073         dnode_children_t *children_dnodes;
1074         dnode_handle_t *dnh;
1075 
1076         /*
1077          * If you are holding the spa config lock as writer, you shouldn't
1078          * be asking the DMU to do *anything* unless it's the root pool
1079          * which may require us to read from the root filesystem while
1080          * holding some (not all) of the locks as writer.
1081          */
1082         ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1083             (spa_is_root(os->os_spa) &&
1084             spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1085 
1086         if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1087                 dn = (object == DMU_USERUSED_OBJECT) ?
1088                     DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1089                 if (dn == NULL)
1090                         return (SET_ERROR(ENOENT));
1091                 type = dn->dn_type;
1092                 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1093                         return (SET_ERROR(ENOENT));
1094                 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1095                         return (SET_ERROR(EEXIST));
1096                 DNODE_VERIFY(dn);
1097                 (void) refcount_add(&dn->dn_holds, tag);
1098                 *dnp = dn;
1099                 return (0);
1100         }
1101 
1102         if (object == 0 || object >= DN_MAX_OBJECT)
1103                 return (SET_ERROR(EINVAL));
1104 
1105         mdn = DMU_META_DNODE(os);
1106         ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1107 
1108         DNODE_VERIFY(mdn);
1109 
1110         if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1111                 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1112                 drop_struct_lock = TRUE;
1113         }
1114 
1115         blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1116 
1117         db = dbuf_hold(mdn, blk, FTAG);
1118         if (drop_struct_lock)
1119                 rw_exit(&mdn->dn_struct_rwlock);
1120         if (db == NULL)
1121                 return (SET_ERROR(EIO));
1122         err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1123         if (err) {
1124                 dbuf_rele(db, FTAG);
1125                 return (err);
1126         }
1127 
1128         ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1129         epb = db->db.db_size >> DNODE_SHIFT;
1130 
1131         idx = object & (epb-1);
1132 
1133         ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1134         children_dnodes = dmu_buf_get_user(&db->db);
1135         if (children_dnodes == NULL) {
1136                 int i;
1137                 dnode_children_t *winner;
1138                 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1139                     epb * sizeof (dnode_handle_t), KM_SLEEP);
1140                 children_dnodes->dnc_count = epb;
1141                 dnh = &children_dnodes->dnc_children[0];
1142                 for (i = 0; i < epb; i++) {
1143                         zrl_init(&dnh[i].dnh_zrlock);
1144                 }
1145                 dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1146                     dnode_buf_evict_async, NULL);
1147                 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1148                 if (winner != NULL) {
1149 
1150                         for (i = 0; i < epb; i++) {
1151                                 zrl_destroy(&dnh[i].dnh_zrlock);
1152                         }
1153 
1154                         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1155                             epb * sizeof (dnode_handle_t));
1156                         children_dnodes = winner;
1157                 }
1158         }
1159         ASSERT(children_dnodes->dnc_count == epb);
1160 
1161         dnh = &children_dnodes->dnc_children[idx];
1162         zrl_add(&dnh->dnh_zrlock);
1163         dn = dnh->dnh_dnode;
1164         if (dn == NULL) {
1165                 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1166 
1167                 dn = dnode_create(os, phys, db, object, dnh);
1168         }
1169 
1170         mutex_enter(&dn->dn_mtx);
1171         type = dn->dn_type;
1172         if (dn->dn_free_txg ||
1173             ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1174             ((flag & DNODE_MUST_BE_FREE) &&
1175             (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1176                 mutex_exit(&dn->dn_mtx);
1177                 zrl_remove(&dnh->dnh_zrlock);
1178                 dbuf_rele(db, FTAG);
1179                 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1180         }
1181         if (refcount_add(&dn->dn_holds, tag) == 1)
1182                 dbuf_add_ref(db, dnh);
1183         mutex_exit(&dn->dn_mtx);
1184 
1185         /* Now we can rely on the hold to prevent the dnode from moving. */
1186         zrl_remove(&dnh->dnh_zrlock);
1187 
1188         DNODE_VERIFY(dn);
1189         ASSERT3P(dn->dn_dbuf, ==, db);
1190         ASSERT3U(dn->dn_object, ==, object);
1191         dbuf_rele(db, FTAG);
1192 
1193         *dnp = dn;
1194         return (0);
1195 }
1196 
1197 /*
1198  * Return held dnode if the object is allocated, NULL if not.
1199  */
1200 int
1201 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1202 {
1203         return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1204 }
1205 
1206 /*
1207  * Can only add a reference if there is already at least one
1208  * reference on the dnode.  Returns FALSE if unable to add a
1209  * new reference.
1210  */
1211 boolean_t
1212 dnode_add_ref(dnode_t *dn, void *tag)
1213 {
1214         mutex_enter(&dn->dn_mtx);
1215         if (refcount_is_zero(&dn->dn_holds)) {
1216                 mutex_exit(&dn->dn_mtx);
1217                 return (FALSE);
1218         }
1219         VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1220         mutex_exit(&dn->dn_mtx);
1221         return (TRUE);
1222 }
1223 
1224 void
1225 dnode_rele(dnode_t *dn, void *tag)
1226 {
1227         mutex_enter(&dn->dn_mtx);
1228         dnode_rele_and_unlock(dn, tag);
1229 }
1230 
1231 void
1232 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1233 {
1234         uint64_t refs;
1235         /* Get while the hold prevents the dnode from moving. */
1236         dmu_buf_impl_t *db = dn->dn_dbuf;
1237         dnode_handle_t *dnh = dn->dn_handle;
1238 
1239         refs = refcount_remove(&dn->dn_holds, tag);
1240         mutex_exit(&dn->dn_mtx);
1241 
1242         /*
1243          * It's unsafe to release the last hold on a dnode by dnode_rele() or
1244          * indirectly by dbuf_rele() while relying on the dnode handle to
1245          * prevent the dnode from moving, since releasing the last hold could
1246          * result in the dnode's parent dbuf evicting its dnode handles. For
1247          * that reason anyone calling dnode_rele() or dbuf_rele() without some
1248          * other direct or indirect hold on the dnode must first drop the dnode
1249          * handle.
1250          */
1251         ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1252 
1253         /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1254         if (refs == 0 && db != NULL) {
1255                 /*
1256                  * Another thread could add a hold to the dnode handle in
1257                  * dnode_hold_impl() while holding the parent dbuf. Since the
1258                  * hold on the parent dbuf prevents the handle from being
1259                  * destroyed, the hold on the handle is OK. We can't yet assert
1260                  * that the handle has zero references, but that will be
1261                  * asserted anyway when the handle gets destroyed.
1262                  */
1263                 dbuf_rele(db, dnh);
1264         }
1265 }
1266 
1267 void
1268 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1269 {
1270         objset_t *os = dn->dn_objset;
1271         uint64_t txg = tx->tx_txg;
1272 
1273         if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1274                 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1275                 return;
1276         }
1277 
1278         DNODE_VERIFY(dn);
1279 
1280 #ifdef ZFS_DEBUG
1281         mutex_enter(&dn->dn_mtx);
1282         ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1283         ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1284         mutex_exit(&dn->dn_mtx);
1285 #endif
1286 
1287         /*
1288          * Determine old uid/gid when necessary
1289          */
1290         dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1291 
1292         multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1293         multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1294 
1295         /*
1296          * If we are already marked dirty, we're done.
1297          */
1298         if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1299                 multilist_sublist_unlock(mls);
1300                 return;
1301         }
1302 
1303         ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1304             !avl_is_empty(&dn->dn_dbufs));
1305         ASSERT(dn->dn_datablksz != 0);
1306         ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1307         ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1308         ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1309 
1310         dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1311             dn->dn_object, txg);
1312 
1313         multilist_sublist_insert_head(mls, dn);
1314 
1315         multilist_sublist_unlock(mls);
1316 
1317         /*
1318          * The dnode maintains a hold on its containing dbuf as
1319          * long as there are holds on it.  Each instantiated child
1320          * dbuf maintains a hold on the dnode.  When the last child
1321          * drops its hold, the dnode will drop its hold on the
1322          * containing dbuf. We add a "dirty hold" here so that the
1323          * dnode will hang around after we finish processing its
1324          * children.
1325          */
1326         VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1327 
1328         (void) dbuf_dirty(dn->dn_dbuf, tx);
1329 
1330         dsl_dataset_dirty(os->os_dsl_dataset, tx);
1331 }
1332 
1333 void
1334 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1335 {
1336         mutex_enter(&dn->dn_mtx);
1337         if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1338                 mutex_exit(&dn->dn_mtx);
1339                 return;
1340         }
1341         dn->dn_free_txg = tx->tx_txg;
1342         mutex_exit(&dn->dn_mtx);
1343 
1344         dnode_setdirty(dn, tx);
1345 }
1346 
1347 /*
1348  * Try to change the block size for the indicated dnode.  This can only
1349  * succeed if there are no blocks allocated or dirty beyond first block
1350  */
1351 int
1352 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1353 {
1354         dmu_buf_impl_t *db;
1355         int err;
1356 
1357         ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1358         if (size == 0)
1359                 size = SPA_MINBLOCKSIZE;
1360         else
1361                 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1362 
1363         if (ibs == dn->dn_indblkshift)
1364                 ibs = 0;
1365 
1366         if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1367                 return (0);
1368 
1369         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1370 
1371         /* Check for any allocated blocks beyond the first */
1372         if (dn->dn_maxblkid != 0)
1373                 goto fail;
1374 
1375         mutex_enter(&dn->dn_dbufs_mtx);
1376         for (db = avl_first(&dn->dn_dbufs); db != NULL;
1377             db = AVL_NEXT(&dn->dn_dbufs, db)) {
1378                 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1379                     db->db_blkid != DMU_SPILL_BLKID) {
1380                         mutex_exit(&dn->dn_dbufs_mtx);
1381                         goto fail;
1382                 }
1383         }
1384         mutex_exit(&dn->dn_dbufs_mtx);
1385 
1386         if (ibs && dn->dn_nlevels != 1)
1387                 goto fail;
1388 
1389         /* resize the old block */
1390         err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1391         if (err == 0)
1392                 dbuf_new_size(db, size, tx);
1393         else if (err != ENOENT)
1394                 goto fail;
1395 
1396         dnode_setdblksz(dn, size);
1397         dnode_setdirty(dn, tx);
1398         dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1399         if (ibs) {
1400                 dn->dn_indblkshift = ibs;
1401                 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1402         }
1403         /* rele after we have fixed the blocksize in the dnode */
1404         if (db)
1405                 dbuf_rele(db, FTAG);
1406 
1407         rw_exit(&dn->dn_struct_rwlock);
1408         return (0);
1409 
1410 fail:
1411         rw_exit(&dn->dn_struct_rwlock);
1412         return (SET_ERROR(ENOTSUP));
1413 }
1414 
1415 /* read-holding callers must not rely on the lock being continuously held */
1416 void
1417 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1418 {
1419         uint64_t txgoff = tx->tx_txg & TXG_MASK;
1420         int epbs, new_nlevels;
1421         uint64_t sz;
1422 
1423         ASSERT(blkid != DMU_BONUS_BLKID);
1424 
1425         ASSERT(have_read ?
1426             RW_READ_HELD(&dn->dn_struct_rwlock) :
1427             RW_WRITE_HELD(&dn->dn_struct_rwlock));
1428 
1429         /*
1430          * if we have a read-lock, check to see if we need to do any work
1431          * before upgrading to a write-lock.
1432          */
1433         if (have_read) {
1434                 if (blkid <= dn->dn_maxblkid)
1435                         return;
1436 
1437                 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1438                         rw_exit(&dn->dn_struct_rwlock);
1439                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1440                 }
1441         }
1442 
1443         if (blkid <= dn->dn_maxblkid)
1444                 goto out;
1445 
1446         dn->dn_maxblkid = blkid;
1447 
1448         /*
1449          * Compute the number of levels necessary to support the new maxblkid.
1450          */
1451         new_nlevels = 1;
1452         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1453         for (sz = dn->dn_nblkptr;
1454             sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1455                 new_nlevels++;
1456 
1457         if (new_nlevels > dn->dn_nlevels) {
1458                 int old_nlevels = dn->dn_nlevels;
1459                 dmu_buf_impl_t *db;
1460                 list_t *list;
1461                 dbuf_dirty_record_t *new, *dr, *dr_next;
1462 
1463                 dn->dn_nlevels = new_nlevels;
1464 
1465                 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1466                 dn->dn_next_nlevels[txgoff] = new_nlevels;
1467 
1468                 /* dirty the left indirects */
1469                 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1470                 ASSERT(db != NULL);
1471                 new = dbuf_dirty(db, tx);
1472                 dbuf_rele(db, FTAG);
1473 
1474                 /* transfer the dirty records to the new indirect */
1475                 mutex_enter(&dn->dn_mtx);
1476                 mutex_enter(&new->dt.di.dr_mtx);
1477                 list = &dn->dn_dirty_records[txgoff];
1478                 for (dr = list_head(list); dr; dr = dr_next) {
1479                         dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1480                         if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1481                             dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1482                             dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1483                                 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1484                                 list_remove(&dn->dn_dirty_records[txgoff], dr);
1485                                 list_insert_tail(&new->dt.di.dr_children, dr);
1486                                 dr->dr_parent = new;
1487                         }
1488                 }
1489                 mutex_exit(&new->dt.di.dr_mtx);
1490                 mutex_exit(&dn->dn_mtx);
1491         }
1492 
1493 out:
1494         if (have_read)
1495                 rw_downgrade(&dn->dn_struct_rwlock);
1496 }
1497 
1498 static void
1499 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1500 {
1501         dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1502         if (db != NULL) {
1503                 dmu_buf_will_dirty(&db->db, tx);
1504                 dbuf_rele(db, FTAG);
1505         }
1506 }
1507 
1508 void
1509 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1510 {
1511         dmu_buf_impl_t *db;
1512         uint64_t blkoff, blkid, nblks;
1513         int blksz, blkshift, head, tail;
1514         int trunc = FALSE;
1515         int epbs;
1516 
1517         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1518         blksz = dn->dn_datablksz;
1519         blkshift = dn->dn_datablkshift;
1520         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1521 
1522         if (len == DMU_OBJECT_END) {
1523                 len = UINT64_MAX - off;
1524                 trunc = TRUE;
1525         }
1526 
1527         /*
1528          * First, block align the region to free:
1529          */
1530         if (ISP2(blksz)) {
1531                 head = P2NPHASE(off, blksz);
1532                 blkoff = P2PHASE(off, blksz);
1533                 if ((off >> blkshift) > dn->dn_maxblkid)
1534                         goto out;
1535         } else {
1536                 ASSERT(dn->dn_maxblkid == 0);
1537                 if (off == 0 && len >= blksz) {
1538                         /*
1539                          * Freeing the whole block; fast-track this request.
1540                          * Note that we won't dirty any indirect blocks,
1541                          * which is fine because we will be freeing the entire
1542                          * file and thus all indirect blocks will be freed
1543                          * by free_children().
1544                          */
1545                         blkid = 0;
1546                         nblks = 1;
1547                         goto done;
1548                 } else if (off >= blksz) {
1549                         /* Freeing past end-of-data */
1550                         goto out;
1551                 } else {
1552                         /* Freeing part of the block. */
1553                         head = blksz - off;
1554                         ASSERT3U(head, >, 0);
1555                 }
1556                 blkoff = off;
1557         }
1558         /* zero out any partial block data at the start of the range */
1559         if (head) {
1560                 ASSERT3U(blkoff + head, ==, blksz);
1561                 if (len < head)
1562                         head = len;
1563                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1564                     TRUE, FALSE, FTAG, &db) == 0) {
1565                         caddr_t data;
1566 
1567                         /* don't dirty if it isn't on disk and isn't dirty */
1568                         if (db->db_last_dirty ||
1569                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1570                                 rw_exit(&dn->dn_struct_rwlock);
1571                                 dmu_buf_will_dirty(&db->db, tx);
1572                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1573                                 data = db->db.db_data;
1574                                 bzero(data + blkoff, head);
1575                         }
1576                         dbuf_rele(db, FTAG);
1577                 }
1578                 off += head;
1579                 len -= head;
1580         }
1581 
1582         /* If the range was less than one block, we're done */
1583         if (len == 0)
1584                 goto out;
1585 
1586         /* If the remaining range is past end of file, we're done */
1587         if ((off >> blkshift) > dn->dn_maxblkid)
1588                 goto out;
1589 
1590         ASSERT(ISP2(blksz));
1591         if (trunc)
1592                 tail = 0;
1593         else
1594                 tail = P2PHASE(len, blksz);
1595 
1596         ASSERT0(P2PHASE(off, blksz));
1597         /* zero out any partial block data at the end of the range */
1598         if (tail) {
1599                 if (len < tail)
1600                         tail = len;
1601                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1602                     TRUE, FALSE, FTAG, &db) == 0) {
1603                         /* don't dirty if not on disk and not dirty */
1604                         if (db->db_last_dirty ||
1605                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1606                                 rw_exit(&dn->dn_struct_rwlock);
1607                                 dmu_buf_will_dirty(&db->db, tx);
1608                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1609                                 bzero(db->db.db_data, tail);
1610                         }
1611                         dbuf_rele(db, FTAG);
1612                 }
1613                 len -= tail;
1614         }
1615 
1616         /* If the range did not include a full block, we are done */
1617         if (len == 0)
1618                 goto out;
1619 
1620         ASSERT(IS_P2ALIGNED(off, blksz));
1621         ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1622         blkid = off >> blkshift;
1623         nblks = len >> blkshift;
1624         if (trunc)
1625                 nblks += 1;
1626 
1627         /*
1628          * Dirty all the indirect blocks in this range.  Note that only
1629          * the first and last indirect blocks can actually be written
1630          * (if they were partially freed) -- they must be dirtied, even if
1631          * they do not exist on disk yet.  The interior blocks will
1632          * be freed by free_children(), so they will not actually be written.
1633          * Even though these interior blocks will not be written, we
1634          * dirty them for two reasons:
1635          *
1636          *  - It ensures that the indirect blocks remain in memory until
1637          *    syncing context.  (They have already been prefetched by
1638          *    dmu_tx_hold_free(), so we don't have to worry about reading
1639          *    them serially here.)
1640          *
1641          *  - The dirty space accounting will put pressure on the txg sync
1642          *    mechanism to begin syncing, and to delay transactions if there
1643          *    is a large amount of freeing.  Even though these indirect
1644          *    blocks will not be written, we could need to write the same
1645          *    amount of space if we copy the freed BPs into deadlists.
1646          */
1647         if (dn->dn_nlevels > 1) {
1648                 uint64_t first, last;
1649 
1650                 first = blkid >> epbs;
1651                 dnode_dirty_l1(dn, first, tx);
1652                 if (trunc)
1653                         last = dn->dn_maxblkid >> epbs;
1654                 else
1655                         last = (blkid + nblks - 1) >> epbs;
1656                 if (last != first)
1657                         dnode_dirty_l1(dn, last, tx);
1658 
1659                 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1660                     SPA_BLKPTRSHIFT;
1661                 for (uint64_t i = first + 1; i < last; i++) {
1662                         /*
1663                          * Set i to the blockid of the next non-hole
1664                          * level-1 indirect block at or after i.  Note
1665                          * that dnode_next_offset() operates in terms of
1666                          * level-0-equivalent bytes.
1667                          */
1668                         uint64_t ibyte = i << shift;
1669                         int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1670                             &ibyte, 2, 1, 0);
1671                         i = ibyte >> shift;
1672                         if (i >= last)
1673                                 break;
1674 
1675                         /*
1676                          * Normally we should not see an error, either
1677                          * from dnode_next_offset() or dbuf_hold_level()
1678                          * (except for ESRCH from dnode_next_offset).
1679                          * If there is an i/o error, then when we read
1680                          * this block in syncing context, it will use
1681                          * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1682                          * to the "failmode" property.  dnode_next_offset()
1683                          * doesn't have a flag to indicate MUSTSUCCEED.
1684                          */
1685                         if (err != 0)
1686                                 break;
1687 
1688                         dnode_dirty_l1(dn, i, tx);
1689                 }
1690         }
1691 
1692 done:
1693         /*
1694          * Add this range to the dnode range list.
1695          * We will finish up this free operation in the syncing phase.
1696          */
1697         mutex_enter(&dn->dn_mtx);
1698         int txgoff = tx->tx_txg & TXG_MASK;
1699         if (dn->dn_free_ranges[txgoff] == NULL) {
1700                 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
1701         }
1702         range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1703         range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1704         dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1705             blkid, nblks, tx->tx_txg);
1706         mutex_exit(&dn->dn_mtx);
1707 
1708         dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1709         dnode_setdirty(dn, tx);
1710 out:
1711 
1712         rw_exit(&dn->dn_struct_rwlock);
1713 }
1714 
1715 static boolean_t
1716 dnode_spill_freed(dnode_t *dn)
1717 {
1718         int i;
1719 
1720         mutex_enter(&dn->dn_mtx);
1721         for (i = 0; i < TXG_SIZE; i++) {
1722                 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1723                         break;
1724         }
1725         mutex_exit(&dn->dn_mtx);
1726         return (i < TXG_SIZE);
1727 }
1728 
1729 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1730 uint64_t
1731 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1732 {
1733         void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1734         int i;
1735 
1736         if (blkid == DMU_BONUS_BLKID)
1737                 return (FALSE);
1738 
1739         /*
1740          * If we're in the process of opening the pool, dp will not be
1741          * set yet, but there shouldn't be anything dirty.
1742          */
1743         if (dp == NULL)
1744                 return (FALSE);
1745 
1746         if (dn->dn_free_txg)
1747                 return (TRUE);
1748 
1749         if (blkid == DMU_SPILL_BLKID)
1750                 return (dnode_spill_freed(dn));
1751 
1752         mutex_enter(&dn->dn_mtx);
1753         for (i = 0; i < TXG_SIZE; i++) {
1754                 if (dn->dn_free_ranges[i] != NULL &&
1755                     range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1756                         break;
1757         }
1758         mutex_exit(&dn->dn_mtx);
1759         return (i < TXG_SIZE);
1760 }
1761 
1762 /* call from syncing context when we actually write/free space for this dnode */
1763 void
1764 dnode_diduse_space(dnode_t *dn, int64_t delta)
1765 {
1766         uint64_t space;
1767         dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1768             dn, dn->dn_phys,
1769             (u_longlong_t)dn->dn_phys->dn_used,
1770             (longlong_t)delta);
1771 
1772         mutex_enter(&dn->dn_mtx);
1773         space = DN_USED_BYTES(dn->dn_phys);
1774         if (delta > 0) {
1775                 ASSERT3U(space + delta, >=, space); /* no overflow */
1776         } else {
1777                 ASSERT3U(space, >=, -delta); /* no underflow */
1778         }
1779         space += delta;
1780         if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1781                 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1782                 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1783                 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1784         } else {
1785                 dn->dn_phys->dn_used = space;
1786                 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1787         }
1788         mutex_exit(&dn->dn_mtx);
1789 }
1790 
1791 /*
1792  * Scans a block at the indicated "level" looking for a hole or data,
1793  * depending on 'flags'.
1794  *
1795  * If level > 0, then we are scanning an indirect block looking at its
1796  * pointers.  If level == 0, then we are looking at a block of dnodes.
1797  *
1798  * If we don't find what we are looking for in the block, we return ESRCH.
1799  * Otherwise, return with *offset pointing to the beginning (if searching
1800  * forwards) or end (if searching backwards) of the range covered by the
1801  * block pointer we matched on (or dnode).
1802  *
1803  * The basic search algorithm used below by dnode_next_offset() is to
1804  * use this function to search up the block tree (widen the search) until
1805  * we find something (i.e., we don't return ESRCH) and then search back
1806  * down the tree (narrow the search) until we reach our original search
1807  * level.
1808  */
1809 static int
1810 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1811     int lvl, uint64_t blkfill, uint64_t txg)
1812 {
1813         dmu_buf_impl_t *db = NULL;
1814         void *data = NULL;
1815         uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1816         uint64_t epb = 1ULL << epbs;
1817         uint64_t minfill, maxfill;
1818         boolean_t hole;
1819         int i, inc, error, span;
1820 
1821         dprintf("probing object %llu offset %llx level %d of %u\n",
1822             dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1823 
1824         hole = ((flags & DNODE_FIND_HOLE) != 0);
1825         inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1826         ASSERT(txg == 0 || !hole);
1827 
1828         if (lvl == dn->dn_phys->dn_nlevels) {
1829                 error = 0;
1830                 epb = dn->dn_phys->dn_nblkptr;
1831                 data = dn->dn_phys->dn_blkptr;
1832         } else {
1833                 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1834                 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1835                 if (error) {
1836                         if (error != ENOENT)
1837                                 return (error);
1838                         if (hole)
1839                                 return (0);
1840                         /*
1841                          * This can only happen when we are searching up
1842                          * the block tree for data.  We don't really need to
1843                          * adjust the offset, as we will just end up looking
1844                          * at the pointer to this block in its parent, and its
1845                          * going to be unallocated, so we will skip over it.
1846                          */
1847                         return (SET_ERROR(ESRCH));
1848                 }
1849                 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1850                 if (error) {
1851                         dbuf_rele(db, FTAG);
1852                         return (error);
1853                 }
1854                 data = db->db.db_data;
1855         }
1856 
1857 
1858         if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1859             db->db_blkptr->blk_birth <= txg ||
1860             BP_IS_HOLE(db->db_blkptr))) {
1861                 /*
1862                  * This can only happen when we are searching up the tree
1863                  * and these conditions mean that we need to keep climbing.
1864                  */
1865                 error = SET_ERROR(ESRCH);
1866         } else if (lvl == 0) {
1867                 dnode_phys_t *dnp = data;
1868                 span = DNODE_SHIFT;
1869                 ASSERT(dn->dn_type == DMU_OT_DNODE);
1870 
1871                 for (i = (*offset >> span) & (blkfill - 1);
1872                     i >= 0 && i < blkfill; i += inc) {
1873                         if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1874                                 break;
1875                         *offset += (1ULL << span) * inc;
1876                 }
1877                 if (i < 0 || i == blkfill)
1878                         error = SET_ERROR(ESRCH);
1879         } else {
1880                 blkptr_t *bp = data;
1881                 uint64_t start = *offset;
1882                 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1883                 minfill = 0;
1884                 maxfill = blkfill << ((lvl - 1) * epbs);
1885 
1886                 if (hole)
1887                         maxfill--;
1888                 else
1889                         minfill++;
1890 
1891                 *offset = *offset >> span;
1892                 for (i = BF64_GET(*offset, 0, epbs);
1893                     i >= 0 && i < epb; i += inc) {
1894                         if (BP_GET_FILL(&bp[i]) >= minfill &&
1895                             BP_GET_FILL(&bp[i]) <= maxfill &&
1896                             (hole || bp[i].blk_birth > txg))
1897                                 break;
1898                         if (inc > 0 || *offset > 0)
1899                                 *offset += inc;
1900                 }
1901                 *offset = *offset << span;
1902                 if (inc < 0) {
1903                         /* traversing backwards; position offset at the end */
1904                         ASSERT3U(*offset, <=, start);
1905                         *offset = MIN(*offset + (1ULL << span) - 1, start);
1906                 } else if (*offset < start) {
1907                         *offset = start;
1908                 }
1909                 if (i < 0 || i >= epb)
1910                         error = SET_ERROR(ESRCH);
1911         }
1912 
1913         if (db)
1914                 dbuf_rele(db, FTAG);
1915 
1916         return (error);
1917 }
1918 
1919 /*
1920  * Find the next hole, data, or sparse region at or after *offset.
1921  * The value 'blkfill' tells us how many items we expect to find
1922  * in an L0 data block; this value is 1 for normal objects,
1923  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1924  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1925  *
1926  * Examples:
1927  *
1928  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1929  *      Finds the next/previous hole/data in a file.
1930  *      Used in dmu_offset_next().
1931  *
1932  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1933  *      Finds the next free/allocated dnode an objset's meta-dnode.
1934  *      Only finds objects that have new contents since txg (ie.
1935  *      bonus buffer changes and content removal are ignored).
1936  *      Used in dmu_object_next().
1937  *
1938  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1939  *      Finds the next L2 meta-dnode bp that's at most 1/4 full.
1940  *      Used in dmu_object_alloc().
1941  */
1942 int
1943 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1944     int minlvl, uint64_t blkfill, uint64_t txg)
1945 {
1946         uint64_t initial_offset = *offset;
1947         int lvl, maxlvl;
1948         int error = 0;
1949 
1950         if (!(flags & DNODE_FIND_HAVELOCK))
1951                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1952 
1953         if (dn->dn_phys->dn_nlevels == 0) {
1954                 error = SET_ERROR(ESRCH);
1955                 goto out;
1956         }
1957 
1958         if (dn->dn_datablkshift == 0) {
1959                 if (*offset < dn->dn_datablksz) {
1960                         if (flags & DNODE_FIND_HOLE)
1961                                 *offset = dn->dn_datablksz;
1962                 } else {
1963                         error = SET_ERROR(ESRCH);
1964                 }
1965                 goto out;
1966         }
1967 
1968         maxlvl = dn->dn_phys->dn_nlevels;
1969 
1970         for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1971                 error = dnode_next_offset_level(dn,
1972                     flags, offset, lvl, blkfill, txg);
1973                 if (error != ESRCH)
1974                         break;
1975         }
1976 
1977         while (error == 0 && --lvl >= minlvl) {
1978                 error = dnode_next_offset_level(dn,
1979                     flags, offset, lvl, blkfill, txg);
1980         }
1981 
1982         /*
1983          * There's always a "virtual hole" at the end of the object, even
1984          * if all BP's which physically exist are non-holes.
1985          */
1986         if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1987             minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1988                 error = 0;
1989         }
1990 
1991         if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1992             initial_offset < *offset : initial_offset > *offset))
1993                 error = SET_ERROR(ESRCH);
1994 out:
1995         if (!(flags & DNODE_FIND_HAVELOCK))
1996                 rw_exit(&dn->dn_struct_rwlock);
1997 
1998         return (error);
1999 }