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) 2013 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 
  39 static int free_range_compar(const void *node1, const void *node2);
  40 
  41 static kmem_cache_t *dnode_cache;
  42 /*
  43  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
  44  * turned on when DEBUG is also defined.
  45  */
  46 #ifdef  DEBUG
  47 #define DNODE_STATS
  48 #endif  /* DEBUG */
  49 
  50 #ifdef  DNODE_STATS
  51 #define DNODE_STAT_ADD(stat)                    ((stat)++)
  52 #else
  53 #define DNODE_STAT_ADD(stat)                    /* nothing */
  54 #endif  /* DNODE_STATS */
  55 
  56 static dnode_phys_t dnode_phys_zero;
  57 
  58 int zfs_default_bs = SPA_MINBLOCKSHIFT;
  59 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
  60 
  61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
  62 
  63 /* ARGSUSED */
  64 static int
  65 dnode_cons(void *arg, void *unused, int kmflag)
  66 {
  67         dnode_t *dn = arg;
  68         int i;
  69 
  70         rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
  71         mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
  72         mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
  73         cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
  74 
  75         /*
  76          * Every dbuf has a reference, and dropping a tracked reference is
  77          * O(number of references), so don't track dn_holds.
  78          */
  79         refcount_create_untracked(&dn->dn_holds);
  80         refcount_create(&dn->dn_tx_holds);
  81         list_link_init(&dn->dn_link);
  82 
  83         bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
  84         bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
  85         bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
  86         bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
  87         bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
  88         bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
  89         bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
  90 
  91         for (i = 0; i < TXG_SIZE; i++) {
  92                 list_link_init(&dn->dn_dirty_link[i]);
  93                 avl_create(&dn->dn_ranges[i], free_range_compar,
  94                     sizeof (free_range_t),
  95                     offsetof(struct free_range, fr_node));
  96                 list_create(&dn->dn_dirty_records[i],
  97                     sizeof (dbuf_dirty_record_t),
  98                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
  99         }
 100 
 101         dn->dn_allocated_txg = 0;
 102         dn->dn_free_txg = 0;
 103         dn->dn_assigned_txg = 0;
 104         dn->dn_dirtyctx = 0;
 105         dn->dn_dirtyctx_firstset = NULL;
 106         dn->dn_bonus = NULL;
 107         dn->dn_have_spill = B_FALSE;
 108         dn->dn_zio = NULL;
 109         dn->dn_oldused = 0;
 110         dn->dn_oldflags = 0;
 111         dn->dn_olduid = 0;
 112         dn->dn_oldgid = 0;
 113         dn->dn_newuid = 0;
 114         dn->dn_newgid = 0;
 115         dn->dn_id_flags = 0;
 116 
 117         dn->dn_dbufs_count = 0;
 118         dn->dn_unlisted_l0_blkid = 0;
 119         list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
 120             offsetof(dmu_buf_impl_t, db_link));
 121 
 122         dn->dn_moved = 0;
 123         return (0);
 124 }
 125 
 126 /* ARGSUSED */
 127 static void
 128 dnode_dest(void *arg, void *unused)
 129 {
 130         int i;
 131         dnode_t *dn = arg;
 132 
 133         rw_destroy(&dn->dn_struct_rwlock);
 134         mutex_destroy(&dn->dn_mtx);
 135         mutex_destroy(&dn->dn_dbufs_mtx);
 136         cv_destroy(&dn->dn_notxholds);
 137         refcount_destroy(&dn->dn_holds);
 138         refcount_destroy(&dn->dn_tx_holds);
 139         ASSERT(!list_link_active(&dn->dn_link));
 140 
 141         for (i = 0; i < TXG_SIZE; i++) {
 142                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 143                 avl_destroy(&dn->dn_ranges[i]);
 144                 list_destroy(&dn->dn_dirty_records[i]);
 145                 ASSERT0(dn->dn_next_nblkptr[i]);
 146                 ASSERT0(dn->dn_next_nlevels[i]);
 147                 ASSERT0(dn->dn_next_indblkshift[i]);
 148                 ASSERT0(dn->dn_next_bonustype[i]);
 149                 ASSERT0(dn->dn_rm_spillblk[i]);
 150                 ASSERT0(dn->dn_next_bonuslen[i]);
 151                 ASSERT0(dn->dn_next_blksz[i]);
 152         }
 153 
 154         ASSERT0(dn->dn_allocated_txg);
 155         ASSERT0(dn->dn_free_txg);
 156         ASSERT0(dn->dn_assigned_txg);
 157         ASSERT0(dn->dn_dirtyctx);
 158         ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
 159         ASSERT3P(dn->dn_bonus, ==, NULL);
 160         ASSERT(!dn->dn_have_spill);
 161         ASSERT3P(dn->dn_zio, ==, NULL);
 162         ASSERT0(dn->dn_oldused);
 163         ASSERT0(dn->dn_oldflags);
 164         ASSERT0(dn->dn_olduid);
 165         ASSERT0(dn->dn_oldgid);
 166         ASSERT0(dn->dn_newuid);
 167         ASSERT0(dn->dn_newgid);
 168         ASSERT0(dn->dn_id_flags);
 169 
 170         ASSERT0(dn->dn_dbufs_count);
 171         ASSERT0(dn->dn_unlisted_l0_blkid);
 172         list_destroy(&dn->dn_dbufs);
 173 }
 174 
 175 void
 176 dnode_init(void)
 177 {
 178         ASSERT(dnode_cache == NULL);
 179         dnode_cache = kmem_cache_create("dnode_t",
 180             sizeof (dnode_t),
 181             0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
 182         kmem_cache_set_move(dnode_cache, dnode_move);
 183 }
 184 
 185 void
 186 dnode_fini(void)
 187 {
 188         kmem_cache_destroy(dnode_cache);
 189         dnode_cache = NULL;
 190 }
 191 
 192 
 193 #ifdef ZFS_DEBUG
 194 void
 195 dnode_verify(dnode_t *dn)
 196 {
 197         int drop_struct_lock = FALSE;
 198 
 199         ASSERT(dn->dn_phys);
 200         ASSERT(dn->dn_objset);
 201         ASSERT(dn->dn_handle->dnh_dnode == dn);
 202 
 203         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 204 
 205         if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
 206                 return;
 207 
 208         if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
 209                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 210                 drop_struct_lock = TRUE;
 211         }
 212         if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
 213                 int i;
 214                 ASSERT3U(dn->dn_indblkshift, >=, 0);
 215                 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
 216                 if (dn->dn_datablkshift) {
 217                         ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
 218                         ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
 219                         ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
 220                 }
 221                 ASSERT3U(dn->dn_nlevels, <=, 30);
 222                 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
 223                 ASSERT3U(dn->dn_nblkptr, >=, 1);
 224                 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 225                 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 226                 ASSERT3U(dn->dn_datablksz, ==,
 227                     dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 228                 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
 229                 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
 230                     dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 231                 for (i = 0; i < TXG_SIZE; i++) {
 232                         ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
 233                 }
 234         }
 235         if (dn->dn_phys->dn_type != DMU_OT_NONE)
 236                 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
 237         ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
 238         if (dn->dn_dbuf != NULL) {
 239                 ASSERT3P(dn->dn_phys, ==,
 240                     (dnode_phys_t *)dn->dn_dbuf->db.db_data +
 241                     (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
 242         }
 243         if (drop_struct_lock)
 244                 rw_exit(&dn->dn_struct_rwlock);
 245 }
 246 #endif
 247 
 248 void
 249 dnode_byteswap(dnode_phys_t *dnp)
 250 {
 251         uint64_t *buf64 = (void*)&dnp->dn_blkptr;
 252         int i;
 253 
 254         if (dnp->dn_type == DMU_OT_NONE) {
 255                 bzero(dnp, sizeof (dnode_phys_t));
 256                 return;
 257         }
 258 
 259         dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
 260         dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
 261         dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
 262         dnp->dn_used = BSWAP_64(dnp->dn_used);
 263 
 264         /*
 265          * dn_nblkptr is only one byte, so it's OK to read it in either
 266          * byte order.  We can't read dn_bouslen.
 267          */
 268         ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
 269         ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
 270         for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
 271                 buf64[i] = BSWAP_64(buf64[i]);
 272 
 273         /*
 274          * OK to check dn_bonuslen for zero, because it won't matter if
 275          * we have the wrong byte order.  This is necessary because the
 276          * dnode dnode is smaller than a regular dnode.
 277          */
 278         if (dnp->dn_bonuslen != 0) {
 279                 /*
 280                  * Note that the bonus length calculated here may be
 281                  * longer than the actual bonus buffer.  This is because
 282                  * we always put the bonus buffer after the last block
 283                  * pointer (instead of packing it against the end of the
 284                  * dnode buffer).
 285                  */
 286                 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
 287                 size_t len = DN_MAX_BONUSLEN - off;
 288                 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
 289                 dmu_object_byteswap_t byteswap =
 290                     DMU_OT_BYTESWAP(dnp->dn_bonustype);
 291                 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
 292         }
 293 
 294         /* Swap SPILL block if we have one */
 295         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 296                 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
 297 
 298 }
 299 
 300 void
 301 dnode_buf_byteswap(void *vbuf, size_t size)
 302 {
 303         dnode_phys_t *buf = vbuf;
 304         int i;
 305 
 306         ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
 307         ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
 308 
 309         size >>= DNODE_SHIFT;
 310         for (i = 0; i < size; i++) {
 311                 dnode_byteswap(buf);
 312                 buf++;
 313         }
 314 }
 315 
 316 static int
 317 free_range_compar(const void *node1, const void *node2)
 318 {
 319         const free_range_t *rp1 = node1;
 320         const free_range_t *rp2 = node2;
 321 
 322         if (rp1->fr_blkid < rp2->fr_blkid)
 323                 return (-1);
 324         else if (rp1->fr_blkid > rp2->fr_blkid)
 325                 return (1);
 326         else return (0);
 327 }
 328 
 329 void
 330 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
 331 {
 332         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 333 
 334         dnode_setdirty(dn, tx);
 335         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 336         ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
 337             (dn->dn_nblkptr-1) * sizeof (blkptr_t));
 338         dn->dn_bonuslen = newsize;
 339         if (newsize == 0)
 340                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
 341         else
 342                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 343         rw_exit(&dn->dn_struct_rwlock);
 344 }
 345 
 346 void
 347 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
 348 {
 349         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 350         dnode_setdirty(dn, tx);
 351         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 352         dn->dn_bonustype = newtype;
 353         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 354         rw_exit(&dn->dn_struct_rwlock);
 355 }
 356 
 357 void
 358 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
 359 {
 360         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 361         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
 362         dnode_setdirty(dn, tx);
 363         dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
 364         dn->dn_have_spill = B_FALSE;
 365 }
 366 
 367 static void
 368 dnode_setdblksz(dnode_t *dn, int size)
 369 {
 370         ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
 371         ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
 372         ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
 373         ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
 374             1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
 375         dn->dn_datablksz = size;
 376         dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
 377         dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
 378 }
 379 
 380 static dnode_t *
 381 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
 382     uint64_t object, dnode_handle_t *dnh)
 383 {
 384         dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
 385 
 386         ASSERT(!POINTER_IS_VALID(dn->dn_objset));
 387         dn->dn_moved = 0;
 388 
 389         /*
 390          * Defer setting dn_objset until the dnode is ready to be a candidate
 391          * for the dnode_move() callback.
 392          */
 393         dn->dn_object = object;
 394         dn->dn_dbuf = db;
 395         dn->dn_handle = dnh;
 396         dn->dn_phys = dnp;
 397 
 398         if (dnp->dn_datablkszsec) {
 399                 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 400         } else {
 401                 dn->dn_datablksz = 0;
 402                 dn->dn_datablkszsec = 0;
 403                 dn->dn_datablkshift = 0;
 404         }
 405         dn->dn_indblkshift = dnp->dn_indblkshift;
 406         dn->dn_nlevels = dnp->dn_nlevels;
 407         dn->dn_type = dnp->dn_type;
 408         dn->dn_nblkptr = dnp->dn_nblkptr;
 409         dn->dn_checksum = dnp->dn_checksum;
 410         dn->dn_compress = dnp->dn_compress;
 411         dn->dn_bonustype = dnp->dn_bonustype;
 412         dn->dn_bonuslen = dnp->dn_bonuslen;
 413         dn->dn_maxblkid = dnp->dn_maxblkid;
 414         dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
 415         dn->dn_id_flags = 0;
 416 
 417         dmu_zfetch_init(&dn->dn_zfetch, dn);
 418 
 419         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 420 
 421         mutex_enter(&os->os_lock);
 422         list_insert_head(&os->os_dnodes, dn);
 423         membar_producer();
 424         /*
 425          * Everything else must be valid before assigning dn_objset makes the
 426          * dnode eligible for dnode_move().
 427          */
 428         dn->dn_objset = os;
 429         mutex_exit(&os->os_lock);
 430 
 431         arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
 432         return (dn);
 433 }
 434 
 435 /*
 436  * Caller must be holding the dnode handle, which is released upon return.
 437  */
 438 static void
 439 dnode_destroy(dnode_t *dn)
 440 {
 441         objset_t *os = dn->dn_objset;
 442 
 443         ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
 444 
 445         mutex_enter(&os->os_lock);
 446         POINTER_INVALIDATE(&dn->dn_objset);
 447         list_remove(&os->os_dnodes, dn);
 448         mutex_exit(&os->os_lock);
 449 
 450         /* the dnode can no longer move, so we can release the handle */
 451         zrl_remove(&dn->dn_handle->dnh_zrlock);
 452 
 453         dn->dn_allocated_txg = 0;
 454         dn->dn_free_txg = 0;
 455         dn->dn_assigned_txg = 0;
 456 
 457         dn->dn_dirtyctx = 0;
 458         if (dn->dn_dirtyctx_firstset != NULL) {
 459                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 460                 dn->dn_dirtyctx_firstset = NULL;
 461         }
 462         if (dn->dn_bonus != NULL) {
 463                 mutex_enter(&dn->dn_bonus->db_mtx);
 464                 dbuf_evict(dn->dn_bonus);
 465                 dn->dn_bonus = NULL;
 466         }
 467         dn->dn_zio = NULL;
 468 
 469         dn->dn_have_spill = B_FALSE;
 470         dn->dn_oldused = 0;
 471         dn->dn_oldflags = 0;
 472         dn->dn_olduid = 0;
 473         dn->dn_oldgid = 0;
 474         dn->dn_newuid = 0;
 475         dn->dn_newgid = 0;
 476         dn->dn_id_flags = 0;
 477         dn->dn_unlisted_l0_blkid = 0;
 478 
 479         dmu_zfetch_rele(&dn->dn_zfetch);
 480         kmem_cache_free(dnode_cache, dn);
 481         arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
 482 }
 483 
 484 void
 485 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
 486     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 487 {
 488         int i;
 489 
 490         if (blocksize == 0)
 491                 blocksize = 1 << zfs_default_bs;
 492         else if (blocksize > SPA_MAXBLOCKSIZE)
 493                 blocksize = SPA_MAXBLOCKSIZE;
 494         else
 495                 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
 496 
 497         if (ibs == 0)
 498                 ibs = zfs_default_ibs;
 499 
 500         ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
 501 
 502         dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
 503             dn->dn_object, tx->tx_txg, blocksize, ibs);
 504 
 505         ASSERT(dn->dn_type == DMU_OT_NONE);
 506         ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
 507         ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
 508         ASSERT(ot != DMU_OT_NONE);
 509         ASSERT(DMU_OT_IS_VALID(ot));
 510         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 511             (bonustype == DMU_OT_SA && bonuslen == 0) ||
 512             (bonustype != DMU_OT_NONE && bonuslen != 0));
 513         ASSERT(DMU_OT_IS_VALID(bonustype));
 514         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 515         ASSERT(dn->dn_type == DMU_OT_NONE);
 516         ASSERT0(dn->dn_maxblkid);
 517         ASSERT0(dn->dn_allocated_txg);
 518         ASSERT0(dn->dn_assigned_txg);
 519         ASSERT(refcount_is_zero(&dn->dn_tx_holds));
 520         ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
 521         ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
 522 
 523         for (i = 0; i < TXG_SIZE; i++) {
 524                 ASSERT0(dn->dn_next_nblkptr[i]);
 525                 ASSERT0(dn->dn_next_nlevels[i]);
 526                 ASSERT0(dn->dn_next_indblkshift[i]);
 527                 ASSERT0(dn->dn_next_bonuslen[i]);
 528                 ASSERT0(dn->dn_next_bonustype[i]);
 529                 ASSERT0(dn->dn_rm_spillblk[i]);
 530                 ASSERT0(dn->dn_next_blksz[i]);
 531                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 532                 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
 533                 ASSERT0(avl_numnodes(&dn->dn_ranges[i]));
 534         }
 535 
 536         dn->dn_type = ot;
 537         dnode_setdblksz(dn, blocksize);
 538         dn->dn_indblkshift = ibs;
 539         dn->dn_nlevels = 1;
 540         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 541                 dn->dn_nblkptr = 1;
 542         else
 543                 dn->dn_nblkptr = 1 +
 544                     ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 545         dn->dn_bonustype = bonustype;
 546         dn->dn_bonuslen = bonuslen;
 547         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 548         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 549         dn->dn_dirtyctx = 0;
 550 
 551         dn->dn_free_txg = 0;
 552         if (dn->dn_dirtyctx_firstset) {
 553                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 554                 dn->dn_dirtyctx_firstset = NULL;
 555         }
 556 
 557         dn->dn_allocated_txg = tx->tx_txg;
 558         dn->dn_id_flags = 0;
 559 
 560         dnode_setdirty(dn, tx);
 561         dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
 562         dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 563         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 564         dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
 565 }
 566 
 567 void
 568 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
 569     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 570 {
 571         int nblkptr;
 572 
 573         ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
 574         ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
 575         ASSERT0(blocksize % SPA_MINBLOCKSIZE);
 576         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
 577         ASSERT(tx->tx_txg != 0);
 578         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 579             (bonustype != DMU_OT_NONE && bonuslen != 0) ||
 580             (bonustype == DMU_OT_SA && bonuslen == 0));
 581         ASSERT(DMU_OT_IS_VALID(bonustype));
 582         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 583 
 584         /* clean up any unreferenced dbufs */
 585         dnode_evict_dbufs(dn);
 586 
 587         dn->dn_id_flags = 0;
 588 
 589         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 590         dnode_setdirty(dn, tx);
 591         if (dn->dn_datablksz != blocksize) {
 592                 /* change blocksize */
 593                 ASSERT(dn->dn_maxblkid == 0 &&
 594                     (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
 595                     dnode_block_freed(dn, 0)));
 596                 dnode_setdblksz(dn, blocksize);
 597                 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
 598         }
 599         if (dn->dn_bonuslen != bonuslen)
 600                 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
 601 
 602         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 603                 nblkptr = 1;
 604         else
 605                 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 606         if (dn->dn_bonustype != bonustype)
 607                 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
 608         if (dn->dn_nblkptr != nblkptr)
 609                 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
 610         if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 611                 dbuf_rm_spill(dn, tx);
 612                 dnode_rm_spill(dn, tx);
 613         }
 614         rw_exit(&dn->dn_struct_rwlock);
 615 
 616         /* change type */
 617         dn->dn_type = ot;
 618 
 619         /* change bonus size and type */
 620         mutex_enter(&dn->dn_mtx);
 621         dn->dn_bonustype = bonustype;
 622         dn->dn_bonuslen = bonuslen;
 623         dn->dn_nblkptr = nblkptr;
 624         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 625         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 626         ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 627 
 628         /* fix up the bonus db_size */
 629         if (dn->dn_bonus) {
 630                 dn->dn_bonus->db.db_size =
 631                     DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
 632                 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
 633         }
 634 
 635         dn->dn_allocated_txg = tx->tx_txg;
 636         mutex_exit(&dn->dn_mtx);
 637 }
 638 
 639 #ifdef  DNODE_STATS
 640 static struct {
 641         uint64_t dms_dnode_invalid;
 642         uint64_t dms_dnode_recheck1;
 643         uint64_t dms_dnode_recheck2;
 644         uint64_t dms_dnode_special;
 645         uint64_t dms_dnode_handle;
 646         uint64_t dms_dnode_rwlock;
 647         uint64_t dms_dnode_active;
 648 } dnode_move_stats;
 649 #endif  /* DNODE_STATS */
 650 
 651 static void
 652 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
 653 {
 654         int i;
 655 
 656         ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
 657         ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
 658         ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
 659         ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
 660 
 661         /* Copy fields. */
 662         ndn->dn_objset = odn->dn_objset;
 663         ndn->dn_object = odn->dn_object;
 664         ndn->dn_dbuf = odn->dn_dbuf;
 665         ndn->dn_handle = odn->dn_handle;
 666         ndn->dn_phys = odn->dn_phys;
 667         ndn->dn_type = odn->dn_type;
 668         ndn->dn_bonuslen = odn->dn_bonuslen;
 669         ndn->dn_bonustype = odn->dn_bonustype;
 670         ndn->dn_nblkptr = odn->dn_nblkptr;
 671         ndn->dn_checksum = odn->dn_checksum;
 672         ndn->dn_compress = odn->dn_compress;
 673         ndn->dn_nlevels = odn->dn_nlevels;
 674         ndn->dn_indblkshift = odn->dn_indblkshift;
 675         ndn->dn_datablkshift = odn->dn_datablkshift;
 676         ndn->dn_datablkszsec = odn->dn_datablkszsec;
 677         ndn->dn_datablksz = odn->dn_datablksz;
 678         ndn->dn_maxblkid = odn->dn_maxblkid;
 679         bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
 680             sizeof (odn->dn_next_nblkptr));
 681         bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
 682             sizeof (odn->dn_next_nlevels));
 683         bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
 684             sizeof (odn->dn_next_indblkshift));
 685         bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
 686             sizeof (odn->dn_next_bonustype));
 687         bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
 688             sizeof (odn->dn_rm_spillblk));
 689         bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
 690             sizeof (odn->dn_next_bonuslen));
 691         bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
 692             sizeof (odn->dn_next_blksz));
 693         for (i = 0; i < TXG_SIZE; i++) {
 694                 list_move_tail(&ndn->dn_dirty_records[i],
 695                     &odn->dn_dirty_records[i]);
 696         }
 697         bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges));
 698         ndn->dn_allocated_txg = odn->dn_allocated_txg;
 699         ndn->dn_free_txg = odn->dn_free_txg;
 700         ndn->dn_assigned_txg = odn->dn_assigned_txg;
 701         ndn->dn_dirtyctx = odn->dn_dirtyctx;
 702         ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
 703         ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
 704         refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
 705         ASSERT(list_is_empty(&ndn->dn_dbufs));
 706         list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
 707         ndn->dn_dbufs_count = odn->dn_dbufs_count;
 708         ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
 709         ndn->dn_bonus = odn->dn_bonus;
 710         ndn->dn_have_spill = odn->dn_have_spill;
 711         ndn->dn_zio = odn->dn_zio;
 712         ndn->dn_oldused = odn->dn_oldused;
 713         ndn->dn_oldflags = odn->dn_oldflags;
 714         ndn->dn_olduid = odn->dn_olduid;
 715         ndn->dn_oldgid = odn->dn_oldgid;
 716         ndn->dn_newuid = odn->dn_newuid;
 717         ndn->dn_newgid = odn->dn_newgid;
 718         ndn->dn_id_flags = odn->dn_id_flags;
 719         dmu_zfetch_init(&ndn->dn_zfetch, NULL);
 720         list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
 721         ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
 722         ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
 723         ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
 724 
 725         /*
 726          * Update back pointers. Updating the handle fixes the back pointer of
 727          * every descendant dbuf as well as the bonus dbuf.
 728          */
 729         ASSERT(ndn->dn_handle->dnh_dnode == odn);
 730         ndn->dn_handle->dnh_dnode = ndn;
 731         if (ndn->dn_zfetch.zf_dnode == odn) {
 732                 ndn->dn_zfetch.zf_dnode = ndn;
 733         }
 734 
 735         /*
 736          * Invalidate the original dnode by clearing all of its back pointers.
 737          */
 738         odn->dn_dbuf = NULL;
 739         odn->dn_handle = NULL;
 740         list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
 741             offsetof(dmu_buf_impl_t, db_link));
 742         odn->dn_dbufs_count = 0;
 743         odn->dn_unlisted_l0_blkid = 0;
 744         odn->dn_bonus = NULL;
 745         odn->dn_zfetch.zf_dnode = NULL;
 746 
 747         /*
 748          * Set the low bit of the objset pointer to ensure that dnode_move()
 749          * recognizes the dnode as invalid in any subsequent callback.
 750          */
 751         POINTER_INVALIDATE(&odn->dn_objset);
 752 
 753         /*
 754          * Satisfy the destructor.
 755          */
 756         for (i = 0; i < TXG_SIZE; i++) {
 757                 list_create(&odn->dn_dirty_records[i],
 758                     sizeof (dbuf_dirty_record_t),
 759                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 760                 odn->dn_ranges[i].avl_root = NULL;
 761                 odn->dn_ranges[i].avl_numnodes = 0;
 762                 odn->dn_next_nlevels[i] = 0;
 763                 odn->dn_next_indblkshift[i] = 0;
 764                 odn->dn_next_bonustype[i] = 0;
 765                 odn->dn_rm_spillblk[i] = 0;
 766                 odn->dn_next_bonuslen[i] = 0;
 767                 odn->dn_next_blksz[i] = 0;
 768         }
 769         odn->dn_allocated_txg = 0;
 770         odn->dn_free_txg = 0;
 771         odn->dn_assigned_txg = 0;
 772         odn->dn_dirtyctx = 0;
 773         odn->dn_dirtyctx_firstset = NULL;
 774         odn->dn_have_spill = B_FALSE;
 775         odn->dn_zio = NULL;
 776         odn->dn_oldused = 0;
 777         odn->dn_oldflags = 0;
 778         odn->dn_olduid = 0;
 779         odn->dn_oldgid = 0;
 780         odn->dn_newuid = 0;
 781         odn->dn_newgid = 0;
 782         odn->dn_id_flags = 0;
 783 
 784         /*
 785          * Mark the dnode.
 786          */
 787         ndn->dn_moved = 1;
 788         odn->dn_moved = (uint8_t)-1;
 789 }
 790 
 791 #ifdef  _KERNEL
 792 /*ARGSUSED*/
 793 static kmem_cbrc_t
 794 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
 795 {
 796         dnode_t *odn = buf, *ndn = newbuf;
 797         objset_t *os;
 798         int64_t refcount;
 799         uint32_t dbufs;
 800 
 801         /*
 802          * The dnode is on the objset's list of known dnodes if the objset
 803          * pointer is valid. We set the low bit of the objset pointer when
 804          * freeing the dnode to invalidate it, and the memory patterns written
 805          * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
 806          * A newly created dnode sets the objset pointer last of all to indicate
 807          * that the dnode is known and in a valid state to be moved by this
 808          * function.
 809          */
 810         os = odn->dn_objset;
 811         if (!POINTER_IS_VALID(os)) {
 812                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
 813                 return (KMEM_CBRC_DONT_KNOW);
 814         }
 815 
 816         /*
 817          * Ensure that the objset does not go away during the move.
 818          */
 819         rw_enter(&os_lock, RW_WRITER);
 820         if (os != odn->dn_objset) {
 821                 rw_exit(&os_lock);
 822                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
 823                 return (KMEM_CBRC_DONT_KNOW);
 824         }
 825 
 826         /*
 827          * If the dnode is still valid, then so is the objset. We know that no
 828          * valid objset can be freed while we hold os_lock, so we can safely
 829          * ensure that the objset remains in use.
 830          */
 831         mutex_enter(&os->os_lock);
 832 
 833         /*
 834          * Recheck the objset pointer in case the dnode was removed just before
 835          * acquiring the lock.
 836          */
 837         if (os != odn->dn_objset) {
 838                 mutex_exit(&os->os_lock);
 839                 rw_exit(&os_lock);
 840                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
 841                 return (KMEM_CBRC_DONT_KNOW);
 842         }
 843 
 844         /*
 845          * At this point we know that as long as we hold os->os_lock, the dnode
 846          * cannot be freed and fields within the dnode can be safely accessed.
 847          * The objset listing this dnode cannot go away as long as this dnode is
 848          * on its list.
 849          */
 850         rw_exit(&os_lock);
 851         if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
 852                 mutex_exit(&os->os_lock);
 853                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
 854                 return (KMEM_CBRC_NO);
 855         }
 856         ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
 857 
 858         /*
 859          * Lock the dnode handle to prevent the dnode from obtaining any new
 860          * holds. This also prevents the descendant dbufs and the bonus dbuf
 861          * from accessing the dnode, so that we can discount their holds. The
 862          * handle is safe to access because we know that while the dnode cannot
 863          * go away, neither can its handle. Once we hold dnh_zrlock, we can
 864          * safely move any dnode referenced only by dbufs.
 865          */
 866         if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
 867                 mutex_exit(&os->os_lock);
 868                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
 869                 return (KMEM_CBRC_LATER);
 870         }
 871 
 872         /*
 873          * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
 874          * We need to guarantee that there is a hold for every dbuf in order to
 875          * determine whether the dnode is actively referenced. Falsely matching
 876          * a dbuf to an active hold would lead to an unsafe move. It's possible
 877          * that a thread already having an active dnode hold is about to add a
 878          * dbuf, and we can't compare hold and dbuf counts while the add is in
 879          * progress.
 880          */
 881         if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
 882                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 883                 mutex_exit(&os->os_lock);
 884                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
 885                 return (KMEM_CBRC_LATER);
 886         }
 887 
 888         /*
 889          * A dbuf may be removed (evicted) without an active dnode hold. In that
 890          * case, the dbuf count is decremented under the handle lock before the
 891          * dbuf's hold is released. This order ensures that if we count the hold
 892          * after the dbuf is removed but before its hold is released, we will
 893          * treat the unmatched hold as active and exit safely. If we count the
 894          * hold before the dbuf is removed, the hold is discounted, and the
 895          * removal is blocked until the move completes.
 896          */
 897         refcount = refcount_count(&odn->dn_holds);
 898         ASSERT(refcount >= 0);
 899         dbufs = odn->dn_dbufs_count;
 900 
 901         /* We can't have more dbufs than dnode holds. */
 902         ASSERT3U(dbufs, <=, refcount);
 903         DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
 904             uint32_t, dbufs);
 905 
 906         if (refcount > dbufs) {
 907                 rw_exit(&odn->dn_struct_rwlock);
 908                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 909                 mutex_exit(&os->os_lock);
 910                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
 911                 return (KMEM_CBRC_LATER);
 912         }
 913 
 914         rw_exit(&odn->dn_struct_rwlock);
 915 
 916         /*
 917          * At this point we know that anyone with a hold on the dnode is not
 918          * actively referencing it. The dnode is known and in a valid state to
 919          * move. We're holding the locks needed to execute the critical section.
 920          */
 921         dnode_move_impl(odn, ndn);
 922 
 923         list_link_replace(&odn->dn_link, &ndn->dn_link);
 924         /* If the dnode was safe to move, the refcount cannot have changed. */
 925         ASSERT(refcount == refcount_count(&ndn->dn_holds));
 926         ASSERT(dbufs == ndn->dn_dbufs_count);
 927         zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
 928         mutex_exit(&os->os_lock);
 929 
 930         return (KMEM_CBRC_YES);
 931 }
 932 #endif  /* _KERNEL */
 933 
 934 void
 935 dnode_special_close(dnode_handle_t *dnh)
 936 {
 937         dnode_t *dn = dnh->dnh_dnode;
 938 
 939         /*
 940          * Wait for final references to the dnode to clear.  This can
 941          * only happen if the arc is asyncronously evicting state that
 942          * has a hold on this dnode while we are trying to evict this
 943          * dnode.
 944          */
 945         while (refcount_count(&dn->dn_holds) > 0)
 946                 delay(1);
 947         zrl_add(&dnh->dnh_zrlock);
 948         dnode_destroy(dn); /* implicit zrl_remove() */
 949         zrl_destroy(&dnh->dnh_zrlock);
 950         dnh->dnh_dnode = NULL;
 951 }
 952 
 953 dnode_t *
 954 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
 955     dnode_handle_t *dnh)
 956 {
 957         dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
 958         dnh->dnh_dnode = dn;
 959         zrl_init(&dnh->dnh_zrlock);
 960         DNODE_VERIFY(dn);
 961         return (dn);
 962 }
 963 
 964 static void
 965 dnode_buf_pageout(dmu_buf_t *db, void *arg)
 966 {
 967         dnode_children_t *children_dnodes = arg;
 968         int i;
 969         int epb = db->db_size >> DNODE_SHIFT;
 970 
 971         ASSERT(epb == children_dnodes->dnc_count);
 972 
 973         for (i = 0; i < epb; i++) {
 974                 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
 975                 dnode_t *dn;
 976 
 977                 /*
 978                  * The dnode handle lock guards against the dnode moving to
 979                  * another valid address, so there is no need here to guard
 980                  * against changes to or from NULL.
 981                  */
 982                 if (dnh->dnh_dnode == NULL) {
 983                         zrl_destroy(&dnh->dnh_zrlock);
 984                         continue;
 985                 }
 986 
 987                 zrl_add(&dnh->dnh_zrlock);
 988                 dn = dnh->dnh_dnode;
 989                 /*
 990                  * If there are holds on this dnode, then there should
 991                  * be holds on the dnode's containing dbuf as well; thus
 992                  * it wouldn't be eligible for eviction and this function
 993                  * would not have been called.
 994                  */
 995                 ASSERT(refcount_is_zero(&dn->dn_holds));
 996                 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
 997 
 998                 dnode_destroy(dn); /* implicit zrl_remove() */
 999                 zrl_destroy(&dnh->dnh_zrlock);
1000                 dnh->dnh_dnode = NULL;
1001         }
1002         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1003             (epb - 1) * sizeof (dnode_handle_t));
1004 }
1005 
1006 /*
1007  * errors:
1008  * EINVAL - invalid object number.
1009  * EIO - i/o error.
1010  * succeeds even for free dnodes.
1011  */
1012 int
1013 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1014     void *tag, dnode_t **dnp)
1015 {
1016         int epb, idx, err;
1017         int drop_struct_lock = FALSE;
1018         int type;
1019         uint64_t blk;
1020         dnode_t *mdn, *dn;
1021         dmu_buf_impl_t *db;
1022         dnode_children_t *children_dnodes;
1023         dnode_handle_t *dnh;
1024 
1025         /*
1026          * If you are holding the spa config lock as writer, you shouldn't
1027          * be asking the DMU to do *anything* unless it's the root pool
1028          * which may require us to read from the root filesystem while
1029          * holding some (not all) of the locks as writer.
1030          */
1031         ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1032             (spa_is_root(os->os_spa) &&
1033             spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1034 
1035         if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1036                 dn = (object == DMU_USERUSED_OBJECT) ?
1037                     DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1038                 if (dn == NULL)
1039                         return (SET_ERROR(ENOENT));
1040                 type = dn->dn_type;
1041                 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1042                         return (SET_ERROR(ENOENT));
1043                 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1044                         return (SET_ERROR(EEXIST));
1045                 DNODE_VERIFY(dn);
1046                 (void) refcount_add(&dn->dn_holds, tag);
1047                 *dnp = dn;
1048                 return (0);
1049         }
1050 
1051         if (object == 0 || object >= DN_MAX_OBJECT)
1052                 return (SET_ERROR(EINVAL));
1053 
1054         mdn = DMU_META_DNODE(os);
1055         ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1056 
1057         DNODE_VERIFY(mdn);
1058 
1059         if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1060                 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1061                 drop_struct_lock = TRUE;
1062         }
1063 
1064         blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1065 
1066         db = dbuf_hold(mdn, blk, FTAG);
1067         if (drop_struct_lock)
1068                 rw_exit(&mdn->dn_struct_rwlock);
1069         if (db == NULL)
1070                 return (SET_ERROR(EIO));
1071         err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1072         if (err) {
1073                 dbuf_rele(db, FTAG);
1074                 return (err);
1075         }
1076 
1077         ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1078         epb = db->db.db_size >> DNODE_SHIFT;
1079 
1080         idx = object & (epb-1);
1081 
1082         ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1083         children_dnodes = dmu_buf_get_user(&db->db);
1084         if (children_dnodes == NULL) {
1085                 int i;
1086                 dnode_children_t *winner;
1087                 children_dnodes = kmem_alloc(sizeof (dnode_children_t) +
1088                     (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1089                 children_dnodes->dnc_count = epb;
1090                 dnh = &children_dnodes->dnc_children[0];
1091                 for (i = 0; i < epb; i++) {
1092                         zrl_init(&dnh[i].dnh_zrlock);
1093                         dnh[i].dnh_dnode = NULL;
1094                 }
1095                 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1096                     dnode_buf_pageout)) {
1097                         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1098                             (epb - 1) * sizeof (dnode_handle_t));
1099                         children_dnodes = winner;
1100                 }
1101         }
1102         ASSERT(children_dnodes->dnc_count == epb);
1103 
1104         dnh = &children_dnodes->dnc_children[idx];
1105         zrl_add(&dnh->dnh_zrlock);
1106         if ((dn = dnh->dnh_dnode) == NULL) {
1107                 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1108                 dnode_t *winner;
1109 
1110                 dn = dnode_create(os, phys, db, object, dnh);
1111                 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1112                 if (winner != NULL) {
1113                         zrl_add(&dnh->dnh_zrlock);
1114                         dnode_destroy(dn); /* implicit zrl_remove() */
1115                         dn = winner;
1116                 }
1117         }
1118 
1119         mutex_enter(&dn->dn_mtx);
1120         type = dn->dn_type;
1121         if (dn->dn_free_txg ||
1122             ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1123             ((flag & DNODE_MUST_BE_FREE) &&
1124             (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1125                 mutex_exit(&dn->dn_mtx);
1126                 zrl_remove(&dnh->dnh_zrlock);
1127                 dbuf_rele(db, FTAG);
1128                 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1129         }
1130         mutex_exit(&dn->dn_mtx);
1131 
1132         if (refcount_add(&dn->dn_holds, tag) == 1)
1133                 dbuf_add_ref(db, dnh);
1134         /* Now we can rely on the hold to prevent the dnode from moving. */
1135         zrl_remove(&dnh->dnh_zrlock);
1136 
1137         DNODE_VERIFY(dn);
1138         ASSERT3P(dn->dn_dbuf, ==, db);
1139         ASSERT3U(dn->dn_object, ==, object);
1140         dbuf_rele(db, FTAG);
1141 
1142         *dnp = dn;
1143         return (0);
1144 }
1145 
1146 /*
1147  * Return held dnode if the object is allocated, NULL if not.
1148  */
1149 int
1150 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1151 {
1152         return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1153 }
1154 
1155 /*
1156  * Can only add a reference if there is already at least one
1157  * reference on the dnode.  Returns FALSE if unable to add a
1158  * new reference.
1159  */
1160 boolean_t
1161 dnode_add_ref(dnode_t *dn, void *tag)
1162 {
1163         mutex_enter(&dn->dn_mtx);
1164         if (refcount_is_zero(&dn->dn_holds)) {
1165                 mutex_exit(&dn->dn_mtx);
1166                 return (FALSE);
1167         }
1168         VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1169         mutex_exit(&dn->dn_mtx);
1170         return (TRUE);
1171 }
1172 
1173 void
1174 dnode_rele(dnode_t *dn, void *tag)
1175 {
1176         uint64_t refs;
1177         /* Get while the hold prevents the dnode from moving. */
1178         dmu_buf_impl_t *db = dn->dn_dbuf;
1179         dnode_handle_t *dnh = dn->dn_handle;
1180 
1181         mutex_enter(&dn->dn_mtx);
1182         refs = refcount_remove(&dn->dn_holds, tag);
1183         mutex_exit(&dn->dn_mtx);
1184 
1185         /*
1186          * It's unsafe to release the last hold on a dnode by dnode_rele() or
1187          * indirectly by dbuf_rele() while relying on the dnode handle to
1188          * prevent the dnode from moving, since releasing the last hold could
1189          * result in the dnode's parent dbuf evicting its dnode handles. For
1190          * that reason anyone calling dnode_rele() or dbuf_rele() without some
1191          * other direct or indirect hold on the dnode must first drop the dnode
1192          * handle.
1193          */
1194         ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1195 
1196         /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1197         if (refs == 0 && db != NULL) {
1198                 /*
1199                  * Another thread could add a hold to the dnode handle in
1200                  * dnode_hold_impl() while holding the parent dbuf. Since the
1201                  * hold on the parent dbuf prevents the handle from being
1202                  * destroyed, the hold on the handle is OK. We can't yet assert
1203                  * that the handle has zero references, but that will be
1204                  * asserted anyway when the handle gets destroyed.
1205                  */
1206                 dbuf_rele(db, dnh);
1207         }
1208 }
1209 
1210 void
1211 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1212 {
1213         objset_t *os = dn->dn_objset;
1214         uint64_t txg = tx->tx_txg;
1215 
1216         if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1217                 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1218                 return;
1219         }
1220 
1221         DNODE_VERIFY(dn);
1222 
1223 #ifdef ZFS_DEBUG
1224         mutex_enter(&dn->dn_mtx);
1225         ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1226         ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1227         mutex_exit(&dn->dn_mtx);
1228 #endif
1229 
1230         /*
1231          * Determine old uid/gid when necessary
1232          */
1233         dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1234 
1235         mutex_enter(&os->os_lock);
1236 
1237         /*
1238          * If we are already marked dirty, we're done.
1239          */
1240         if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1241                 mutex_exit(&os->os_lock);
1242                 return;
1243         }
1244 
1245         ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1246         ASSERT(dn->dn_datablksz != 0);
1247         ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1248         ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1249         ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1250 
1251         dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1252             dn->dn_object, txg);
1253 
1254         if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1255                 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1256         } else {
1257                 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1258         }
1259 
1260         mutex_exit(&os->os_lock);
1261 
1262         /*
1263          * The dnode maintains a hold on its containing dbuf as
1264          * long as there are holds on it.  Each instantiated child
1265          * dbuf maintains a hold on the dnode.  When the last child
1266          * drops its hold, the dnode will drop its hold on the
1267          * containing dbuf. We add a "dirty hold" here so that the
1268          * dnode will hang around after we finish processing its
1269          * children.
1270          */
1271         VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1272 
1273         (void) dbuf_dirty(dn->dn_dbuf, tx);
1274 
1275         dsl_dataset_dirty(os->os_dsl_dataset, tx);
1276 }
1277 
1278 void
1279 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1280 {
1281         int txgoff = tx->tx_txg & TXG_MASK;
1282 
1283         dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1284 
1285         /* we should be the only holder... hopefully */
1286         /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1287 
1288         mutex_enter(&dn->dn_mtx);
1289         if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1290                 mutex_exit(&dn->dn_mtx);
1291                 return;
1292         }
1293         dn->dn_free_txg = tx->tx_txg;
1294         mutex_exit(&dn->dn_mtx);
1295 
1296         /*
1297          * If the dnode is already dirty, it needs to be moved from
1298          * the dirty list to the free list.
1299          */
1300         mutex_enter(&dn->dn_objset->os_lock);
1301         if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1302                 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1303                 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1304                 mutex_exit(&dn->dn_objset->os_lock);
1305         } else {
1306                 mutex_exit(&dn->dn_objset->os_lock);
1307                 dnode_setdirty(dn, tx);
1308         }
1309 }
1310 
1311 /*
1312  * Try to change the block size for the indicated dnode.  This can only
1313  * succeed if there are no blocks allocated or dirty beyond first block
1314  */
1315 int
1316 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1317 {
1318         dmu_buf_impl_t *db, *db_next;
1319         int err;
1320 
1321         if (size == 0)
1322                 size = SPA_MINBLOCKSIZE;
1323         if (size > SPA_MAXBLOCKSIZE)
1324                 size = SPA_MAXBLOCKSIZE;
1325         else
1326                 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1327 
1328         if (ibs == dn->dn_indblkshift)
1329                 ibs = 0;
1330 
1331         if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1332                 return (0);
1333 
1334         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1335 
1336         /* Check for any allocated blocks beyond the first */
1337         if (dn->dn_phys->dn_maxblkid != 0)
1338                 goto fail;
1339 
1340         mutex_enter(&dn->dn_dbufs_mtx);
1341         for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1342                 db_next = list_next(&dn->dn_dbufs, db);
1343 
1344                 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1345                     db->db_blkid != DMU_SPILL_BLKID) {
1346                         mutex_exit(&dn->dn_dbufs_mtx);
1347                         goto fail;
1348                 }
1349         }
1350         mutex_exit(&dn->dn_dbufs_mtx);
1351 
1352         if (ibs && dn->dn_nlevels != 1)
1353                 goto fail;
1354 
1355         /* resize the old block */
1356         err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1357         if (err == 0)
1358                 dbuf_new_size(db, size, tx);
1359         else if (err != ENOENT)
1360                 goto fail;
1361 
1362         dnode_setdblksz(dn, size);
1363         dnode_setdirty(dn, tx);
1364         dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1365         if (ibs) {
1366                 dn->dn_indblkshift = ibs;
1367                 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1368         }
1369         /* rele after we have fixed the blocksize in the dnode */
1370         if (db)
1371                 dbuf_rele(db, FTAG);
1372 
1373         rw_exit(&dn->dn_struct_rwlock);
1374         return (0);
1375 
1376 fail:
1377         rw_exit(&dn->dn_struct_rwlock);
1378         return (SET_ERROR(ENOTSUP));
1379 }
1380 
1381 /* read-holding callers must not rely on the lock being continuously held */
1382 void
1383 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1384 {
1385         uint64_t txgoff = tx->tx_txg & TXG_MASK;
1386         int epbs, new_nlevels;
1387         uint64_t sz;
1388 
1389         ASSERT(blkid != DMU_BONUS_BLKID);
1390 
1391         ASSERT(have_read ?
1392             RW_READ_HELD(&dn->dn_struct_rwlock) :
1393             RW_WRITE_HELD(&dn->dn_struct_rwlock));
1394 
1395         /*
1396          * if we have a read-lock, check to see if we need to do any work
1397          * before upgrading to a write-lock.
1398          */
1399         if (have_read) {
1400                 if (blkid <= dn->dn_maxblkid)
1401                         return;
1402 
1403                 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1404                         rw_exit(&dn->dn_struct_rwlock);
1405                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1406                 }
1407         }
1408 
1409         if (blkid <= dn->dn_maxblkid)
1410                 goto out;
1411 
1412         dn->dn_maxblkid = blkid;
1413 
1414         /*
1415          * Compute the number of levels necessary to support the new maxblkid.
1416          */
1417         new_nlevels = 1;
1418         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1419         for (sz = dn->dn_nblkptr;
1420             sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1421                 new_nlevels++;
1422 
1423         if (new_nlevels > dn->dn_nlevels) {
1424                 int old_nlevels = dn->dn_nlevels;
1425                 dmu_buf_impl_t *db;
1426                 list_t *list;
1427                 dbuf_dirty_record_t *new, *dr, *dr_next;
1428 
1429                 dn->dn_nlevels = new_nlevels;
1430 
1431                 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1432                 dn->dn_next_nlevels[txgoff] = new_nlevels;
1433 
1434                 /* dirty the left indirects */
1435                 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1436                 ASSERT(db != NULL);
1437                 new = dbuf_dirty(db, tx);
1438                 dbuf_rele(db, FTAG);
1439 
1440                 /* transfer the dirty records to the new indirect */
1441                 mutex_enter(&dn->dn_mtx);
1442                 mutex_enter(&new->dt.di.dr_mtx);
1443                 list = &dn->dn_dirty_records[txgoff];
1444                 for (dr = list_head(list); dr; dr = dr_next) {
1445                         dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1446                         if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1447                             dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1448                             dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1449                                 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1450                                 list_remove(&dn->dn_dirty_records[txgoff], dr);
1451                                 list_insert_tail(&new->dt.di.dr_children, dr);
1452                                 dr->dr_parent = new;
1453                         }
1454                 }
1455                 mutex_exit(&new->dt.di.dr_mtx);
1456                 mutex_exit(&dn->dn_mtx);
1457         }
1458 
1459 out:
1460         if (have_read)
1461                 rw_downgrade(&dn->dn_struct_rwlock);
1462 }
1463 
1464 void
1465 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
1466 {
1467         avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1468         avl_index_t where;
1469         free_range_t *rp;
1470         free_range_t rp_tofind;
1471         uint64_t endblk = blkid + nblks;
1472 
1473         ASSERT(MUTEX_HELD(&dn->dn_mtx));
1474         ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
1475 
1476         dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1477             blkid, nblks, tx->tx_txg);
1478         rp_tofind.fr_blkid = blkid;
1479         rp = avl_find(tree, &rp_tofind, &where);
1480         if (rp == NULL)
1481                 rp = avl_nearest(tree, where, AVL_BEFORE);
1482         if (rp == NULL)
1483                 rp = avl_nearest(tree, where, AVL_AFTER);
1484 
1485         while (rp && (rp->fr_blkid <= blkid + nblks)) {
1486                 uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
1487                 free_range_t *nrp = AVL_NEXT(tree, rp);
1488 
1489                 if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
1490                         /* clear this entire range */
1491                         avl_remove(tree, rp);
1492                         kmem_free(rp, sizeof (free_range_t));
1493                 } else if (blkid <= rp->fr_blkid &&
1494                     endblk > rp->fr_blkid && endblk < fr_endblk) {
1495                         /* clear the beginning of this range */
1496                         rp->fr_blkid = endblk;
1497                         rp->fr_nblks = fr_endblk - endblk;
1498                 } else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
1499                     endblk >= fr_endblk) {
1500                         /* clear the end of this range */
1501                         rp->fr_nblks = blkid - rp->fr_blkid;
1502                 } else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
1503                         /* clear a chunk out of this range */
1504                         free_range_t *new_rp =
1505                             kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1506 
1507                         new_rp->fr_blkid = endblk;
1508                         new_rp->fr_nblks = fr_endblk - endblk;
1509                         avl_insert_here(tree, new_rp, rp, AVL_AFTER);
1510                         rp->fr_nblks = blkid - rp->fr_blkid;
1511                 }
1512                 /* there may be no overlap */
1513                 rp = nrp;
1514         }
1515 }
1516 
1517 void
1518 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1519 {
1520         dmu_buf_impl_t *db;
1521         uint64_t blkoff, blkid, nblks;
1522         int blksz, blkshift, head, tail;
1523         int trunc = FALSE;
1524         int epbs;
1525 
1526         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1527         blksz = dn->dn_datablksz;
1528         blkshift = dn->dn_datablkshift;
1529         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1530 
1531         if (len == DMU_OBJECT_END) {
1532                 len = UINT64_MAX - off;
1533                 trunc = TRUE;
1534         }
1535 
1536         /*
1537          * First, block align the region to free:
1538          */
1539         if (ISP2(blksz)) {
1540                 head = P2NPHASE(off, blksz);
1541                 blkoff = P2PHASE(off, blksz);
1542                 if ((off >> blkshift) > dn->dn_maxblkid)
1543                         goto out;
1544         } else {
1545                 ASSERT(dn->dn_maxblkid == 0);
1546                 if (off == 0 && len >= blksz) {
1547                         /* Freeing the whole block; fast-track this request */
1548                         blkid = 0;
1549                         nblks = 1;
1550                         goto done;
1551                 } else if (off >= blksz) {
1552                         /* Freeing past end-of-data */
1553                         goto out;
1554                 } else {
1555                         /* Freeing part of the block. */
1556                         head = blksz - off;
1557                         ASSERT3U(head, >, 0);
1558                 }
1559                 blkoff = off;
1560         }
1561         /* zero out any partial block data at the start of the range */
1562         if (head) {
1563                 ASSERT3U(blkoff + head, ==, blksz);
1564                 if (len < head)
1565                         head = len;
1566                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1567                     FTAG, &db) == 0) {
1568                         caddr_t data;
1569 
1570                         /* don't dirty if it isn't on disk and isn't dirty */
1571                         if (db->db_last_dirty ||
1572                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1573                                 rw_exit(&dn->dn_struct_rwlock);
1574                                 dbuf_will_dirty(db, tx);
1575                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1576                                 data = db->db.db_data;
1577                                 bzero(data + blkoff, head);
1578                         }
1579                         dbuf_rele(db, FTAG);
1580                 }
1581                 off += head;
1582                 len -= head;
1583         }
1584 
1585         /* If the range was less than one block, we're done */
1586         if (len == 0)
1587                 goto out;
1588 
1589         /* If the remaining range is past end of file, we're done */
1590         if ((off >> blkshift) > dn->dn_maxblkid)
1591                 goto out;
1592 
1593         ASSERT(ISP2(blksz));
1594         if (trunc)
1595                 tail = 0;
1596         else
1597                 tail = P2PHASE(len, blksz);
1598 
1599         ASSERT0(P2PHASE(off, blksz));
1600         /* zero out any partial block data at the end of the range */
1601         if (tail) {
1602                 if (len < tail)
1603                         tail = len;
1604                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1605                     TRUE, FTAG, &db) == 0) {
1606                         /* don't dirty if not on disk and not dirty */
1607                         if (db->db_last_dirty ||
1608                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1609                                 rw_exit(&dn->dn_struct_rwlock);
1610                                 dbuf_will_dirty(db, tx);
1611                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1612                                 bzero(db->db.db_data, tail);
1613                         }
1614                         dbuf_rele(db, FTAG);
1615                 }
1616                 len -= tail;
1617         }
1618 
1619         /* If the range did not include a full block, we are done */
1620         if (len == 0)
1621                 goto out;
1622 
1623         ASSERT(IS_P2ALIGNED(off, blksz));
1624         ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1625         blkid = off >> blkshift;
1626         nblks = len >> blkshift;
1627         if (trunc)
1628                 nblks += 1;
1629 
1630         /*
1631          * Read in and mark all the level-1 indirects dirty,
1632          * so that they will stay in memory until syncing phase.
1633          * Always dirty the first and last indirect to make sure
1634          * we dirty all the partial indirects.
1635          */
1636         if (dn->dn_nlevels > 1) {
1637                 uint64_t i, first, last;
1638                 int shift = epbs + dn->dn_datablkshift;
1639 
1640                 first = blkid >> epbs;
1641                 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1642                         dbuf_will_dirty(db, tx);
1643                         dbuf_rele(db, FTAG);
1644                 }
1645                 if (trunc)
1646                         last = dn->dn_maxblkid >> epbs;
1647                 else
1648                         last = (blkid + nblks - 1) >> epbs;
1649                 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1650                         dbuf_will_dirty(db, tx);
1651                         dbuf_rele(db, FTAG);
1652                 }
1653                 for (i = first + 1; i < last; i++) {
1654                         uint64_t ibyte = i << shift;
1655                         int err;
1656 
1657                         err = dnode_next_offset(dn,
1658                             DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0);
1659                         i = ibyte >> shift;
1660                         if (err == ESRCH || i >= last)
1661                                 break;
1662                         ASSERT(err == 0);
1663                         db = dbuf_hold_level(dn, 1, i, FTAG);
1664                         if (db) {
1665                                 dbuf_will_dirty(db, tx);
1666                                 dbuf_rele(db, FTAG);
1667                         }
1668                 }
1669         }
1670 done:
1671         /*
1672          * Add this range to the dnode range list.
1673          * We will finish up this free operation in the syncing phase.
1674          */
1675         mutex_enter(&dn->dn_mtx);
1676         dnode_clear_range(dn, blkid, nblks, tx);
1677         {
1678                 free_range_t *rp, *found;
1679                 avl_index_t where;
1680                 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1681 
1682                 /* Add new range to dn_ranges */
1683                 rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1684                 rp->fr_blkid = blkid;
1685                 rp->fr_nblks = nblks;
1686                 found = avl_find(tree, rp, &where);
1687                 ASSERT(found == NULL);
1688                 avl_insert(tree, rp, where);
1689                 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1690                     blkid, nblks, tx->tx_txg);
1691         }
1692         mutex_exit(&dn->dn_mtx);
1693 
1694         dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1695         dnode_setdirty(dn, tx);
1696 out:
1697         if (trunc && dn->dn_maxblkid >= (off >> blkshift))
1698                 dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0);
1699 
1700         rw_exit(&dn->dn_struct_rwlock);
1701 }
1702 
1703 static boolean_t
1704 dnode_spill_freed(dnode_t *dn)
1705 {
1706         int i;
1707 
1708         mutex_enter(&dn->dn_mtx);
1709         for (i = 0; i < TXG_SIZE; i++) {
1710                 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1711                         break;
1712         }
1713         mutex_exit(&dn->dn_mtx);
1714         return (i < TXG_SIZE);
1715 }
1716 
1717 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1718 uint64_t
1719 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1720 {
1721         free_range_t range_tofind;
1722         void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1723         int i;
1724 
1725         if (blkid == DMU_BONUS_BLKID)
1726                 return (FALSE);
1727 
1728         /*
1729          * If we're in the process of opening the pool, dp will not be
1730          * set yet, but there shouldn't be anything dirty.
1731          */
1732         if (dp == NULL)
1733                 return (FALSE);
1734 
1735         if (dn->dn_free_txg)
1736                 return (TRUE);
1737 
1738         if (blkid == DMU_SPILL_BLKID)
1739                 return (dnode_spill_freed(dn));
1740 
1741         range_tofind.fr_blkid = blkid;
1742         mutex_enter(&dn->dn_mtx);
1743         for (i = 0; i < TXG_SIZE; i++) {
1744                 free_range_t *range_found;
1745                 avl_index_t idx;
1746 
1747                 range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1748                 if (range_found) {
1749                         ASSERT(range_found->fr_nblks > 0);
1750                         break;
1751                 }
1752                 range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1753                 if (range_found &&
1754                     range_found->fr_blkid + range_found->fr_nblks > blkid)
1755                         break;
1756         }
1757         mutex_exit(&dn->dn_mtx);
1758         return (i < TXG_SIZE);
1759 }
1760 
1761 /* call from syncing context when we actually write/free space for this dnode */
1762 void
1763 dnode_diduse_space(dnode_t *dn, int64_t delta)
1764 {
1765         uint64_t space;
1766         dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1767             dn, dn->dn_phys,
1768             (u_longlong_t)dn->dn_phys->dn_used,
1769             (longlong_t)delta);
1770 
1771         mutex_enter(&dn->dn_mtx);
1772         space = DN_USED_BYTES(dn->dn_phys);
1773         if (delta > 0) {
1774                 ASSERT3U(space + delta, >=, space); /* no overflow */
1775         } else {
1776                 ASSERT3U(space, >=, -delta); /* no underflow */
1777         }
1778         space += delta;
1779         if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1780                 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1781                 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1782                 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1783         } else {
1784                 dn->dn_phys->dn_used = space;
1785                 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1786         }
1787         mutex_exit(&dn->dn_mtx);
1788 }
1789 
1790 /*
1791  * Call when we think we're going to write/free space in open context.
1792  * Be conservative (ie. OK to write less than this or free more than
1793  * this, but don't write more or free less).
1794  */
1795 void
1796 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1797 {
1798         objset_t *os = dn->dn_objset;
1799         dsl_dataset_t *ds = os->os_dsl_dataset;
1800 
1801         if (space > 0)
1802                 space = spa_get_asize(os->os_spa, space);
1803 
1804         if (ds)
1805                 dsl_dir_willuse_space(ds->ds_dir, space, tx);
1806 
1807         dmu_tx_willuse_space(tx, space);
1808 }
1809 
1810 /*
1811  * Scans a block at the indicated "level" looking for a hole or data,
1812  * depending on 'flags'.
1813  *
1814  * If level > 0, then we are scanning an indirect block looking at its
1815  * pointers.  If level == 0, then we are looking at a block of dnodes.
1816  *
1817  * If we don't find what we are looking for in the block, we return ESRCH.
1818  * Otherwise, return with *offset pointing to the beginning (if searching
1819  * forwards) or end (if searching backwards) of the range covered by the
1820  * block pointer we matched on (or dnode).
1821  *
1822  * The basic search algorithm used below by dnode_next_offset() is to
1823  * use this function to search up the block tree (widen the search) until
1824  * we find something (i.e., we don't return ESRCH) and then search back
1825  * down the tree (narrow the search) until we reach our original search
1826  * level.
1827  */
1828 static int
1829 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1830         int lvl, uint64_t blkfill, uint64_t txg)
1831 {
1832         dmu_buf_impl_t *db = NULL;
1833         void *data = NULL;
1834         uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1835         uint64_t epb = 1ULL << epbs;
1836         uint64_t minfill, maxfill;
1837         boolean_t hole;
1838         int i, inc, error, span;
1839 
1840         dprintf("probing object %llu offset %llx level %d of %u\n",
1841             dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1842 
1843         hole = ((flags & DNODE_FIND_HOLE) != 0);
1844         inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1845         ASSERT(txg == 0 || !hole);
1846 
1847         if (lvl == dn->dn_phys->dn_nlevels) {
1848                 error = 0;
1849                 epb = dn->dn_phys->dn_nblkptr;
1850                 data = dn->dn_phys->dn_blkptr;
1851         } else {
1852                 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1853                 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1854                 if (error) {
1855                         if (error != ENOENT)
1856                                 return (error);
1857                         if (hole)
1858                                 return (0);
1859                         /*
1860                          * This can only happen when we are searching up
1861                          * the block tree for data.  We don't really need to
1862                          * adjust the offset, as we will just end up looking
1863                          * at the pointer to this block in its parent, and its
1864                          * going to be unallocated, so we will skip over it.
1865                          */
1866                         return (SET_ERROR(ESRCH));
1867                 }
1868                 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1869                 if (error) {
1870                         dbuf_rele(db, FTAG);
1871                         return (error);
1872                 }
1873                 data = db->db.db_data;
1874         }
1875 
1876         if (db && txg &&
1877             (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1878                 /*
1879                  * This can only happen when we are searching up the tree
1880                  * and these conditions mean that we need to keep climbing.
1881                  */
1882                 error = SET_ERROR(ESRCH);
1883         } else if (lvl == 0) {
1884                 dnode_phys_t *dnp = data;
1885                 span = DNODE_SHIFT;
1886                 ASSERT(dn->dn_type == DMU_OT_DNODE);
1887 
1888                 for (i = (*offset >> span) & (blkfill - 1);
1889                     i >= 0 && i < blkfill; i += inc) {
1890                         if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1891                                 break;
1892                         *offset += (1ULL << span) * inc;
1893                 }
1894                 if (i < 0 || i == blkfill)
1895                         error = SET_ERROR(ESRCH);
1896         } else {
1897                 blkptr_t *bp = data;
1898                 uint64_t start = *offset;
1899                 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1900                 minfill = 0;
1901                 maxfill = blkfill << ((lvl - 1) * epbs);
1902 
1903                 if (hole)
1904                         maxfill--;
1905                 else
1906                         minfill++;
1907 
1908                 *offset = *offset >> span;
1909                 for (i = BF64_GET(*offset, 0, epbs);
1910                     i >= 0 && i < epb; i += inc) {
1911                         if (bp[i].blk_fill >= minfill &&
1912                             bp[i].blk_fill <= maxfill &&
1913                             (hole || bp[i].blk_birth > txg))
1914                                 break;
1915                         if (inc > 0 || *offset > 0)
1916                                 *offset += inc;
1917                 }
1918                 *offset = *offset << span;
1919                 if (inc < 0) {
1920                         /* traversing backwards; position offset at the end */
1921                         ASSERT3U(*offset, <=, start);
1922                         *offset = MIN(*offset + (1ULL << span) - 1, start);
1923                 } else if (*offset < start) {
1924                         *offset = start;
1925                 }
1926                 if (i < 0 || i >= epb)
1927                         error = SET_ERROR(ESRCH);
1928         }
1929 
1930         if (db)
1931                 dbuf_rele(db, FTAG);
1932 
1933         return (error);
1934 }
1935 
1936 /*
1937  * Find the next hole, data, or sparse region at or after *offset.
1938  * The value 'blkfill' tells us how many items we expect to find
1939  * in an L0 data block; this value is 1 for normal objects,
1940  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1941  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1942  *
1943  * Examples:
1944  *
1945  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1946  *      Finds the next/previous hole/data in a file.
1947  *      Used in dmu_offset_next().
1948  *
1949  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1950  *      Finds the next free/allocated dnode an objset's meta-dnode.
1951  *      Only finds objects that have new contents since txg (ie.
1952  *      bonus buffer changes and content removal are ignored).
1953  *      Used in dmu_object_next().
1954  *
1955  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1956  *      Finds the next L2 meta-dnode bp that's at most 1/4 full.
1957  *      Used in dmu_object_alloc().
1958  */
1959 int
1960 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1961     int minlvl, uint64_t blkfill, uint64_t txg)
1962 {
1963         uint64_t initial_offset = *offset;
1964         int lvl, maxlvl;
1965         int error = 0;
1966 
1967         if (!(flags & DNODE_FIND_HAVELOCK))
1968                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1969 
1970         if (dn->dn_phys->dn_nlevels == 0) {
1971                 error = SET_ERROR(ESRCH);
1972                 goto out;
1973         }
1974 
1975         if (dn->dn_datablkshift == 0) {
1976                 if (*offset < dn->dn_datablksz) {
1977                         if (flags & DNODE_FIND_HOLE)
1978                                 *offset = dn->dn_datablksz;
1979                 } else {
1980                         error = SET_ERROR(ESRCH);
1981                 }
1982                 goto out;
1983         }
1984 
1985         maxlvl = dn->dn_phys->dn_nlevels;
1986 
1987         for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1988                 error = dnode_next_offset_level(dn,
1989                     flags, offset, lvl, blkfill, txg);
1990                 if (error != ESRCH)
1991                         break;
1992         }
1993 
1994         while (error == 0 && --lvl >= minlvl) {
1995                 error = dnode_next_offset_level(dn,
1996                     flags, offset, lvl, blkfill, txg);
1997         }
1998 
1999         if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2000             initial_offset < *offset : initial_offset > *offset))
2001                 error = SET_ERROR(ESRCH);
2002 out:
2003         if (!(flags & DNODE_FIND_HAVELOCK))
2004                 rw_exit(&dn->dn_struct_rwlock);
2005 
2006         return (error);
2007 }