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 2015 Nexenta Systems, Inc.  All rights reserved.
  24  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
  25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  26  * Copyright (c) 2014 Integros [integros.com]
  27  * Copyright 2017 RackTop Systems.
  28  */
  29 
  30 #include <sys/zfs_context.h>
  31 #include <sys/dbuf.h>
  32 #include <sys/dnode.h>
  33 #include <sys/dmu.h>
  34 #include <sys/dmu_impl.h>
  35 #include <sys/dmu_tx.h>
  36 #include <sys/dmu_objset.h>
  37 #include <sys/dsl_dir.h>
  38 #include <sys/dsl_dataset.h>
  39 #include <sys/spa.h>
  40 #include <sys/zio.h>
  41 #include <sys/dmu_zfetch.h>
  42 #include <sys/range_tree.h>
  43 
  44 static void smartcomp_check_comp(dnode_smartcomp_t *sc);
  45 
  46 static kmem_cache_t *dnode_cache;
  47 /*
  48  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
  49  * turned on when DEBUG is also defined.
  50  */
  51 #ifdef  DEBUG
  52 #define DNODE_STATS
  53 #endif  /* DEBUG */
  54 
  55 #ifdef  DNODE_STATS
  56 #define DNODE_STAT_ADD(stat)                    ((stat)++)
  57 #else
  58 #define DNODE_STAT_ADD(stat)                    /* nothing */
  59 #endif  /* DNODE_STATS */
  60 
  61 static dnode_phys_t dnode_phys_zero;
  62 
  63 int zfs_default_bs = SPA_MINBLOCKSHIFT;
  64 int zfs_default_ibs = DN_DFL_INDBLKSHIFT;
  65 
  66 #ifdef  _KERNEL
  67 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
  68 #endif  /* _KERNEL */
  69 
  70 static int
  71 dbuf_compare(const void *x1, const void *x2)
  72 {
  73         const dmu_buf_impl_t *d1 = x1;
  74         const dmu_buf_impl_t *d2 = x2;
  75 
  76         if (d1->db_level < d2->db_level) {
  77                 return (-1);
  78         }
  79         if (d1->db_level > d2->db_level) {
  80                 return (1);
  81         }
  82 
  83         if (d1->db_blkid < d2->db_blkid) {
  84                 return (-1);
  85         }
  86         if (d1->db_blkid > d2->db_blkid) {
  87                 return (1);
  88         }
  89 
  90         if (d1->db_state == DB_SEARCH) {
  91                 ASSERT3S(d2->db_state, !=, DB_SEARCH);
  92                 return (-1);
  93         } else if (d2->db_state == DB_SEARCH) {
  94                 ASSERT3S(d1->db_state, !=, DB_SEARCH);
  95                 return (1);
  96         }
  97 
  98         if ((uintptr_t)d1 < (uintptr_t)d2) {
  99                 return (-1);
 100         }
 101         if ((uintptr_t)d1 > (uintptr_t)d2) {
 102                 return (1);
 103         }
 104         return (0);
 105 }
 106 
 107 /* ARGSUSED */
 108 static int
 109 dnode_cons(void *arg, void *unused, int kmflag)
 110 {
 111         dnode_t *dn = arg;
 112         int i;
 113 
 114         rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
 115         mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
 116         mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
 117         cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
 118 
 119         /*
 120          * Every dbuf has a reference, and dropping a tracked reference is
 121          * O(number of references), so don't track dn_holds.
 122          */
 123         refcount_create_untracked(&dn->dn_holds);
 124         refcount_create(&dn->dn_tx_holds);
 125         list_link_init(&dn->dn_link);
 126 
 127         bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
 128         bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
 129         bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
 130         bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
 131         bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
 132         bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
 133         bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
 134 
 135         for (i = 0; i < TXG_SIZE; i++) {
 136                 list_link_init(&dn->dn_dirty_link[i]);
 137                 dn->dn_free_ranges[i] = NULL;
 138                 list_create(&dn->dn_dirty_records[i],
 139                     sizeof (dbuf_dirty_record_t),
 140                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 141         }
 142 
 143         dn->dn_allocated_txg = 0;
 144         dn->dn_free_txg = 0;
 145         dn->dn_assigned_txg = 0;
 146         dn->dn_dirtyctx = 0;
 147         dn->dn_dirtyctx_firstset = NULL;
 148         dn->dn_bonus = NULL;
 149         dn->dn_have_spill = B_FALSE;
 150         dn->dn_zio = NULL;
 151         dn->dn_oldused = 0;
 152         dn->dn_oldflags = 0;
 153         dn->dn_olduid = 0;
 154         dn->dn_oldgid = 0;
 155         dn->dn_newuid = 0;
 156         dn->dn_newgid = 0;
 157         dn->dn_id_flags = 0;
 158 
 159         dn->dn_dbufs_count = 0;
 160         avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
 161             offsetof(dmu_buf_impl_t, db_link));
 162 
 163         dn->dn_moved = 0;
 164 
 165         bzero(&dn->dn_smartcomp, sizeof (dn->dn_smartcomp));
 166         mutex_init(&dn->dn_smartcomp.sc_lock, NULL, MUTEX_DEFAULT, NULL);
 167 
 168         return (0);
 169 }
 170 
 171 /* ARGSUSED */
 172 static void
 173 dnode_dest(void *arg, void *unused)
 174 {
 175         int i;
 176         dnode_t *dn = arg;
 177 
 178         mutex_destroy(&dn->dn_smartcomp.sc_lock);
 179 
 180         rw_destroy(&dn->dn_struct_rwlock);
 181         mutex_destroy(&dn->dn_mtx);
 182         mutex_destroy(&dn->dn_dbufs_mtx);
 183         cv_destroy(&dn->dn_notxholds);
 184         refcount_destroy(&dn->dn_holds);
 185         refcount_destroy(&dn->dn_tx_holds);
 186         ASSERT(!list_link_active(&dn->dn_link));
 187 
 188         for (i = 0; i < TXG_SIZE; i++) {
 189                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 190                 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
 191                 list_destroy(&dn->dn_dirty_records[i]);
 192                 ASSERT0(dn->dn_next_nblkptr[i]);
 193                 ASSERT0(dn->dn_next_nlevels[i]);
 194                 ASSERT0(dn->dn_next_indblkshift[i]);
 195                 ASSERT0(dn->dn_next_bonustype[i]);
 196                 ASSERT0(dn->dn_rm_spillblk[i]);
 197                 ASSERT0(dn->dn_next_bonuslen[i]);
 198                 ASSERT0(dn->dn_next_blksz[i]);
 199         }
 200 
 201         ASSERT0(dn->dn_allocated_txg);
 202         ASSERT0(dn->dn_free_txg);
 203         ASSERT0(dn->dn_assigned_txg);
 204         ASSERT0(dn->dn_dirtyctx);
 205         ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
 206         ASSERT3P(dn->dn_bonus, ==, NULL);
 207         ASSERT(!dn->dn_have_spill);
 208         ASSERT3P(dn->dn_zio, ==, NULL);
 209         ASSERT0(dn->dn_oldused);
 210         ASSERT0(dn->dn_oldflags);
 211         ASSERT0(dn->dn_olduid);
 212         ASSERT0(dn->dn_oldgid);
 213         ASSERT0(dn->dn_newuid);
 214         ASSERT0(dn->dn_newgid);
 215         ASSERT0(dn->dn_id_flags);
 216 
 217         ASSERT0(dn->dn_dbufs_count);
 218         avl_destroy(&dn->dn_dbufs);
 219 }
 220 
 221 void
 222 dnode_init(void)
 223 {
 224         ASSERT(dnode_cache == NULL);
 225         dnode_cache = kmem_cache_create("dnode_t",
 226             sizeof (dnode_t),
 227             0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
 228 #ifdef  _KERNEL
 229         kmem_cache_set_move(dnode_cache, dnode_move);
 230 #endif  /* _KERNEL */
 231 }
 232 
 233 void
 234 dnode_fini(void)
 235 {
 236         kmem_cache_destroy(dnode_cache);
 237         dnode_cache = NULL;
 238 }
 239 
 240 
 241 #ifdef ZFS_DEBUG
 242 void
 243 dnode_verify(dnode_t *dn)
 244 {
 245         int drop_struct_lock = FALSE;
 246 
 247         ASSERT(dn->dn_phys);
 248         ASSERT(dn->dn_objset);
 249         ASSERT(dn->dn_handle->dnh_dnode == dn);
 250 
 251         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 252 
 253         if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
 254                 return;
 255 
 256         if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
 257                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 258                 drop_struct_lock = TRUE;
 259         }
 260         if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
 261                 int i;
 262                 ASSERT3U(dn->dn_indblkshift, >=, 0);
 263                 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
 264                 if (dn->dn_datablkshift) {
 265                         ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
 266                         ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
 267                         ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
 268                 }
 269                 ASSERT3U(dn->dn_nlevels, <=, 30);
 270                 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
 271                 ASSERT3U(dn->dn_nblkptr, >=, 1);
 272                 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 273                 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 274                 ASSERT3U(dn->dn_datablksz, ==,
 275                     dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 276                 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
 277                 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
 278                     dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
 279                 for (i = 0; i < TXG_SIZE; i++) {
 280                         ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
 281                 }
 282         }
 283         if (dn->dn_phys->dn_type != DMU_OT_NONE)
 284                 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
 285         ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
 286         if (dn->dn_dbuf != NULL) {
 287                 ASSERT3P(dn->dn_phys, ==,
 288                     (dnode_phys_t *)dn->dn_dbuf->db.db_data +
 289                     (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
 290         }
 291         if (drop_struct_lock)
 292                 rw_exit(&dn->dn_struct_rwlock);
 293 }
 294 #endif
 295 
 296 void
 297 dnode_byteswap(dnode_phys_t *dnp)
 298 {
 299         uint64_t *buf64 = (void*)&dnp->dn_blkptr;
 300         int i;
 301 
 302         if (dnp->dn_type == DMU_OT_NONE) {
 303                 bzero(dnp, sizeof (dnode_phys_t));
 304                 return;
 305         }
 306 
 307         dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
 308         dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
 309         dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
 310         dnp->dn_used = BSWAP_64(dnp->dn_used);
 311 
 312         /*
 313          * dn_nblkptr is only one byte, so it's OK to read it in either
 314          * byte order.  We can't read dn_bouslen.
 315          */
 316         ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
 317         ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
 318         for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
 319                 buf64[i] = BSWAP_64(buf64[i]);
 320 
 321         /*
 322          * OK to check dn_bonuslen for zero, because it won't matter if
 323          * we have the wrong byte order.  This is necessary because the
 324          * dnode dnode is smaller than a regular dnode.
 325          */
 326         if (dnp->dn_bonuslen != 0) {
 327                 /*
 328                  * Note that the bonus length calculated here may be
 329                  * longer than the actual bonus buffer.  This is because
 330                  * we always put the bonus buffer after the last block
 331                  * pointer (instead of packing it against the end of the
 332                  * dnode buffer).
 333                  */
 334                 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
 335                 size_t len = DN_MAX_BONUSLEN - off;
 336                 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
 337                 dmu_object_byteswap_t byteswap =
 338                     DMU_OT_BYTESWAP(dnp->dn_bonustype);
 339                 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
 340         }
 341 
 342         /* Swap SPILL block if we have one */
 343         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
 344                 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
 345 
 346 }
 347 
 348 void
 349 dnode_buf_byteswap(void *vbuf, size_t size)
 350 {
 351         dnode_phys_t *buf = vbuf;
 352         int i;
 353 
 354         ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
 355         ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
 356 
 357         size >>= DNODE_SHIFT;
 358         for (i = 0; i < size; i++) {
 359                 dnode_byteswap(buf);
 360                 buf++;
 361         }
 362 }
 363 
 364 void
 365 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
 366 {
 367         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 368 
 369         dnode_setdirty(dn, tx);
 370         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 371         ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
 372             (dn->dn_nblkptr-1) * sizeof (blkptr_t));
 373         dn->dn_bonuslen = newsize;
 374         if (newsize == 0)
 375                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
 376         else
 377                 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 378         rw_exit(&dn->dn_struct_rwlock);
 379 }
 380 
 381 void
 382 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
 383 {
 384         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 385         dnode_setdirty(dn, tx);
 386         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 387         dn->dn_bonustype = newtype;
 388         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 389         rw_exit(&dn->dn_struct_rwlock);
 390 }
 391 
 392 void
 393 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
 394 {
 395         ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
 396         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
 397         dnode_setdirty(dn, tx);
 398         dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
 399         dn->dn_have_spill = B_FALSE;
 400 }
 401 
 402 static void
 403 dnode_setdblksz(dnode_t *dn, int size)
 404 {
 405         ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
 406         ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
 407         ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
 408         ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
 409             1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
 410         dn->dn_datablksz = size;
 411         dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
 412         dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
 413 }
 414 
 415 static dnode_t *
 416 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
 417     uint64_t object, dnode_handle_t *dnh)
 418 {
 419         dnode_t *dn;
 420 
 421         dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
 422 #ifdef _KERNEL
 423         ASSERT(!POINTER_IS_VALID(dn->dn_objset));
 424 #endif /* _KERNEL */
 425         dn->dn_moved = 0;
 426 
 427         /*
 428          * Defer setting dn_objset until the dnode is ready to be a candidate
 429          * for the dnode_move() callback.
 430          */
 431         dn->dn_object = object;
 432         dn->dn_dbuf = db;
 433         dn->dn_handle = dnh;
 434         dn->dn_phys = dnp;
 435 
 436         if (dnp->dn_datablkszsec) {
 437                 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 438         } else {
 439                 dn->dn_datablksz = 0;
 440                 dn->dn_datablkszsec = 0;
 441                 dn->dn_datablkshift = 0;
 442         }
 443         dn->dn_indblkshift = dnp->dn_indblkshift;
 444         dn->dn_nlevels = dnp->dn_nlevels;
 445         dn->dn_type = dnp->dn_type;
 446         dn->dn_nblkptr = dnp->dn_nblkptr;
 447         dn->dn_checksum = dnp->dn_checksum;
 448         dn->dn_compress = dnp->dn_compress;
 449         dn->dn_bonustype = dnp->dn_bonustype;
 450         dn->dn_bonuslen = dnp->dn_bonuslen;
 451         dn->dn_maxblkid = dnp->dn_maxblkid;
 452         dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
 453         dn->dn_id_flags = 0;
 454 
 455         dmu_zfetch_init(&dn->dn_zfetch, dn);
 456 
 457         ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
 458 
 459         mutex_enter(&os->os_lock);
 460         if (dnh->dnh_dnode != NULL) {
 461                 /* Lost the allocation race. */
 462                 mutex_exit(&os->os_lock);
 463                 kmem_cache_free(dnode_cache, dn);
 464                 return (dnh->dnh_dnode);
 465         }
 466 
 467         /*
 468          * Exclude special dnodes from os_dnodes so an empty os_dnodes
 469          * signifies that the special dnodes have no references from
 470          * their children (the entries in os_dnodes).  This allows
 471          * dnode_destroy() to easily determine if the last child has
 472          * been removed and then complete eviction of the objset.
 473          */
 474         if (!DMU_OBJECT_IS_SPECIAL(object))
 475                 list_insert_head(&os->os_dnodes, dn);
 476         membar_producer();
 477 
 478         /*
 479          * Everything else must be valid before assigning dn_objset
 480          * makes the dnode eligible for dnode_move().
 481          */
 482         dn->dn_objset = os;
 483 
 484         dnh->dnh_dnode = dn;
 485         mutex_exit(&os->os_lock);
 486 
 487         arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
 488         return (dn);
 489 }
 490 
 491 /*
 492  * Caller must be holding the dnode handle, which is released upon return.
 493  */
 494 static void
 495 dnode_destroy(dnode_t *dn)
 496 {
 497         objset_t *os = dn->dn_objset;
 498         boolean_t complete_os_eviction = B_FALSE;
 499 
 500         ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
 501 
 502         mutex_enter(&os->os_lock);
 503         POINTER_INVALIDATE(&dn->dn_objset);
 504         if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
 505                 list_remove(&os->os_dnodes, dn);
 506                 complete_os_eviction =
 507                     list_is_empty(&os->os_dnodes) &&
 508                     list_link_active(&os->os_evicting_node);
 509         }
 510         mutex_exit(&os->os_lock);
 511 
 512         /* the dnode can no longer move, so we can release the handle */
 513         zrl_remove(&dn->dn_handle->dnh_zrlock);
 514 
 515         dn->dn_allocated_txg = 0;
 516         dn->dn_free_txg = 0;
 517         dn->dn_assigned_txg = 0;
 518 
 519         dn->dn_dirtyctx = 0;
 520         if (dn->dn_dirtyctx_firstset != NULL) {
 521                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 522                 dn->dn_dirtyctx_firstset = NULL;
 523         }
 524         if (dn->dn_bonus != NULL) {
 525                 mutex_enter(&dn->dn_bonus->db_mtx);
 526                 dbuf_destroy(dn->dn_bonus);
 527                 dn->dn_bonus = NULL;
 528         }
 529         dn->dn_zio = NULL;
 530 
 531         dn->dn_have_spill = B_FALSE;
 532         dn->dn_oldused = 0;
 533         dn->dn_oldflags = 0;
 534         dn->dn_olduid = 0;
 535         dn->dn_oldgid = 0;
 536         dn->dn_newuid = 0;
 537         dn->dn_newgid = 0;
 538         dn->dn_id_flags = 0;
 539 
 540         dmu_zfetch_fini(&dn->dn_zfetch);
 541         kmem_cache_free(dnode_cache, dn);
 542         arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
 543 
 544         if (complete_os_eviction)
 545                 dmu_objset_evict_done(os);
 546 }
 547 
 548 void
 549 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
 550     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 551 {
 552         int i;
 553 
 554         ASSERT3U(blocksize, <=,
 555             spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 556         if (blocksize == 0)
 557                 blocksize = 1 << zfs_default_bs;
 558         else
 559                 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
 560 
 561         if (ibs == 0)
 562                 ibs = zfs_default_ibs;
 563 
 564         ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
 565 
 566         dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
 567             dn->dn_object, tx->tx_txg, blocksize, ibs);
 568 
 569         ASSERT(dn->dn_type == DMU_OT_NONE);
 570         ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
 571         ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
 572         ASSERT(ot != DMU_OT_NONE);
 573         ASSERT(DMU_OT_IS_VALID(ot));
 574         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 575             (bonustype == DMU_OT_SA && bonuslen == 0) ||
 576             (bonustype != DMU_OT_NONE && bonuslen != 0));
 577         ASSERT(DMU_OT_IS_VALID(bonustype));
 578         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 579         ASSERT(dn->dn_type == DMU_OT_NONE);
 580         ASSERT0(dn->dn_maxblkid);
 581         ASSERT0(dn->dn_allocated_txg);
 582         ASSERT0(dn->dn_assigned_txg);
 583         ASSERT(refcount_is_zero(&dn->dn_tx_holds));
 584         ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
 585         ASSERT(avl_is_empty(&dn->dn_dbufs));
 586 
 587         for (i = 0; i < TXG_SIZE; i++) {
 588                 ASSERT0(dn->dn_next_nblkptr[i]);
 589                 ASSERT0(dn->dn_next_nlevels[i]);
 590                 ASSERT0(dn->dn_next_indblkshift[i]);
 591                 ASSERT0(dn->dn_next_bonuslen[i]);
 592                 ASSERT0(dn->dn_next_bonustype[i]);
 593                 ASSERT0(dn->dn_rm_spillblk[i]);
 594                 ASSERT0(dn->dn_next_blksz[i]);
 595                 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
 596                 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
 597                 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
 598         }
 599 
 600         dn->dn_type = ot;
 601         dnode_setdblksz(dn, blocksize);
 602         dn->dn_indblkshift = ibs;
 603         dn->dn_nlevels = 1;
 604         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 605                 dn->dn_nblkptr = 1;
 606         else
 607                 dn->dn_nblkptr = 1 +
 608                     ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 609         dn->dn_bonustype = bonustype;
 610         dn->dn_bonuslen = bonuslen;
 611         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 612         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 613         dn->dn_dirtyctx = 0;
 614 
 615         dn->dn_free_txg = 0;
 616         if (dn->dn_dirtyctx_firstset) {
 617                 kmem_free(dn->dn_dirtyctx_firstset, 1);
 618                 dn->dn_dirtyctx_firstset = NULL;
 619         }
 620 
 621         dn->dn_allocated_txg = tx->tx_txg;
 622         dn->dn_id_flags = 0;
 623 
 624         dnode_setdirty(dn, tx);
 625         dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
 626         dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
 627         dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
 628         dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
 629 }
 630 
 631 void
 632 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
 633     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
 634 {
 635         int nblkptr;
 636 
 637         ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
 638         ASSERT3U(blocksize, <=,
 639             spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 640         ASSERT0(blocksize % SPA_MINBLOCKSIZE);
 641         ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
 642         ASSERT(tx->tx_txg != 0);
 643         ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
 644             (bonustype != DMU_OT_NONE && bonuslen != 0) ||
 645             (bonustype == DMU_OT_SA && bonuslen == 0));
 646         ASSERT(DMU_OT_IS_VALID(bonustype));
 647         ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
 648 
 649         /* clean up any unreferenced dbufs */
 650         dnode_evict_dbufs(dn, DBUF_EVICT_ALL);
 651 
 652         dn->dn_id_flags = 0;
 653 
 654         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 655         dnode_setdirty(dn, tx);
 656         if (dn->dn_datablksz != blocksize) {
 657                 /* change blocksize */
 658                 ASSERT(dn->dn_maxblkid == 0 &&
 659                     (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
 660                     dnode_block_freed(dn, 0)));
 661                 dnode_setdblksz(dn, blocksize);
 662                 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
 663         }
 664         if (dn->dn_bonuslen != bonuslen)
 665                 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
 666 
 667         if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
 668                 nblkptr = 1;
 669         else
 670                 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
 671         if (dn->dn_bonustype != bonustype)
 672                 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
 673         if (dn->dn_nblkptr != nblkptr)
 674                 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
 675         if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 676                 dbuf_rm_spill(dn, tx);
 677                 dnode_rm_spill(dn, tx);
 678         }
 679         rw_exit(&dn->dn_struct_rwlock);
 680 
 681         /* change type */
 682         dn->dn_type = ot;
 683 
 684         /* change bonus size and type */
 685         mutex_enter(&dn->dn_mtx);
 686         dn->dn_bonustype = bonustype;
 687         dn->dn_bonuslen = bonuslen;
 688         dn->dn_nblkptr = nblkptr;
 689         dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
 690         dn->dn_compress = ZIO_COMPRESS_INHERIT;
 691         ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
 692 
 693         /* fix up the bonus db_size */
 694         if (dn->dn_bonus) {
 695                 dn->dn_bonus->db.db_size =
 696                     DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
 697                 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
 698         }
 699 
 700         dn->dn_allocated_txg = tx->tx_txg;
 701         mutex_exit(&dn->dn_mtx);
 702 }
 703 
 704 #ifdef  DNODE_STATS
 705 static struct {
 706         uint64_t dms_dnode_invalid;
 707         uint64_t dms_dnode_recheck1;
 708         uint64_t dms_dnode_recheck2;
 709         uint64_t dms_dnode_special;
 710         uint64_t dms_dnode_handle;
 711         uint64_t dms_dnode_rwlock;
 712         uint64_t dms_dnode_active;
 713 } dnode_move_stats;
 714 #endif  /* DNODE_STATS */
 715 
 716 #ifdef  _KERNEL
 717 static void
 718 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
 719 {
 720         int i;
 721 
 722         ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
 723         ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
 724         ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
 725         ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
 726 
 727         /* Copy fields. */
 728         ndn->dn_objset = odn->dn_objset;
 729         ndn->dn_object = odn->dn_object;
 730         ndn->dn_dbuf = odn->dn_dbuf;
 731         ndn->dn_handle = odn->dn_handle;
 732         ndn->dn_phys = odn->dn_phys;
 733         ndn->dn_type = odn->dn_type;
 734         ndn->dn_bonuslen = odn->dn_bonuslen;
 735         ndn->dn_bonustype = odn->dn_bonustype;
 736         ndn->dn_nblkptr = odn->dn_nblkptr;
 737         ndn->dn_checksum = odn->dn_checksum;
 738         ndn->dn_compress = odn->dn_compress;
 739         ndn->dn_nlevels = odn->dn_nlevels;
 740         ndn->dn_indblkshift = odn->dn_indblkshift;
 741         ndn->dn_datablkshift = odn->dn_datablkshift;
 742         ndn->dn_datablkszsec = odn->dn_datablkszsec;
 743         ndn->dn_datablksz = odn->dn_datablksz;
 744         ndn->dn_maxblkid = odn->dn_maxblkid;
 745         bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
 746             sizeof (odn->dn_next_nblkptr));
 747         bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
 748             sizeof (odn->dn_next_nlevels));
 749         bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
 750             sizeof (odn->dn_next_indblkshift));
 751         bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
 752             sizeof (odn->dn_next_bonustype));
 753         bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
 754             sizeof (odn->dn_rm_spillblk));
 755         bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
 756             sizeof (odn->dn_next_bonuslen));
 757         bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
 758             sizeof (odn->dn_next_blksz));
 759         for (i = 0; i < TXG_SIZE; i++) {
 760                 list_move_tail(&ndn->dn_dirty_records[i],
 761                     &odn->dn_dirty_records[i]);
 762         }
 763         bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
 764             sizeof (odn->dn_free_ranges));
 765         ndn->dn_allocated_txg = odn->dn_allocated_txg;
 766         ndn->dn_free_txg = odn->dn_free_txg;
 767         ndn->dn_assigned_txg = odn->dn_assigned_txg;
 768         ndn->dn_dirtyctx = odn->dn_dirtyctx;
 769         ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
 770         ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
 771         refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
 772         ASSERT(avl_is_empty(&ndn->dn_dbufs));
 773         avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
 774         ndn->dn_dbufs_count = odn->dn_dbufs_count;
 775         ndn->dn_bonus = odn->dn_bonus;
 776         ndn->dn_have_spill = odn->dn_have_spill;
 777         ndn->dn_zio = odn->dn_zio;
 778         ndn->dn_oldused = odn->dn_oldused;
 779         ndn->dn_oldflags = odn->dn_oldflags;
 780         ndn->dn_olduid = odn->dn_olduid;
 781         ndn->dn_oldgid = odn->dn_oldgid;
 782         ndn->dn_newuid = odn->dn_newuid;
 783         ndn->dn_newgid = odn->dn_newgid;
 784         ndn->dn_id_flags = odn->dn_id_flags;
 785         dmu_zfetch_init(&ndn->dn_zfetch, NULL);
 786         list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
 787         ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
 788 
 789         /*
 790          * Update back pointers. Updating the handle fixes the back pointer of
 791          * every descendant dbuf as well as the bonus dbuf.
 792          */
 793         ASSERT(ndn->dn_handle->dnh_dnode == odn);
 794         ndn->dn_handle->dnh_dnode = ndn;
 795         if (ndn->dn_zfetch.zf_dnode == odn) {
 796                 ndn->dn_zfetch.zf_dnode = ndn;
 797         }
 798 
 799         /*
 800          * Invalidate the original dnode by clearing all of its back pointers.
 801          */
 802         odn->dn_dbuf = NULL;
 803         odn->dn_handle = NULL;
 804         avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
 805             offsetof(dmu_buf_impl_t, db_link));
 806         odn->dn_dbufs_count = 0;
 807         odn->dn_bonus = NULL;
 808         odn->dn_zfetch.zf_dnode = NULL;
 809 
 810         /*
 811          * Set the low bit of the objset pointer to ensure that dnode_move()
 812          * recognizes the dnode as invalid in any subsequent callback.
 813          */
 814         POINTER_INVALIDATE(&odn->dn_objset);
 815 
 816         /*
 817          * Satisfy the destructor.
 818          */
 819         for (i = 0; i < TXG_SIZE; i++) {
 820                 list_create(&odn->dn_dirty_records[i],
 821                     sizeof (dbuf_dirty_record_t),
 822                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
 823                 odn->dn_free_ranges[i] = NULL;
 824                 odn->dn_next_nlevels[i] = 0;
 825                 odn->dn_next_indblkshift[i] = 0;
 826                 odn->dn_next_bonustype[i] = 0;
 827                 odn->dn_rm_spillblk[i] = 0;
 828                 odn->dn_next_bonuslen[i] = 0;
 829                 odn->dn_next_blksz[i] = 0;
 830         }
 831         odn->dn_allocated_txg = 0;
 832         odn->dn_free_txg = 0;
 833         odn->dn_assigned_txg = 0;
 834         odn->dn_dirtyctx = 0;
 835         odn->dn_dirtyctx_firstset = NULL;
 836         odn->dn_have_spill = B_FALSE;
 837         odn->dn_zio = NULL;
 838         odn->dn_oldused = 0;
 839         odn->dn_oldflags = 0;
 840         odn->dn_olduid = 0;
 841         odn->dn_oldgid = 0;
 842         odn->dn_newuid = 0;
 843         odn->dn_newgid = 0;
 844         odn->dn_id_flags = 0;
 845 
 846         /*
 847          * Mark the dnode.
 848          */
 849         ndn->dn_moved = 1;
 850         odn->dn_moved = (uint8_t)-1;
 851 }
 852 
 853 /*ARGSUSED*/
 854 static kmem_cbrc_t
 855 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
 856 {
 857         dnode_t *odn = buf, *ndn = newbuf;
 858         objset_t *os;
 859         int64_t refcount;
 860         uint32_t dbufs;
 861 
 862         /*
 863          * The dnode is on the objset's list of known dnodes if the objset
 864          * pointer is valid. We set the low bit of the objset pointer when
 865          * freeing the dnode to invalidate it, and the memory patterns written
 866          * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
 867          * A newly created dnode sets the objset pointer last of all to indicate
 868          * that the dnode is known and in a valid state to be moved by this
 869          * function.
 870          */
 871         os = odn->dn_objset;
 872         if (!POINTER_IS_VALID(os)) {
 873                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
 874                 return (KMEM_CBRC_DONT_KNOW);
 875         }
 876 
 877         /*
 878          * Ensure that the objset does not go away during the move.
 879          */
 880         rw_enter(&os_lock, RW_WRITER);
 881         if (os != odn->dn_objset) {
 882                 rw_exit(&os_lock);
 883                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
 884                 return (KMEM_CBRC_DONT_KNOW);
 885         }
 886 
 887         /*
 888          * If the dnode is still valid, then so is the objset. We know that no
 889          * valid objset can be freed while we hold os_lock, so we can safely
 890          * ensure that the objset remains in use.
 891          */
 892         mutex_enter(&os->os_lock);
 893 
 894         /*
 895          * Recheck the objset pointer in case the dnode was removed just before
 896          * acquiring the lock.
 897          */
 898         if (os != odn->dn_objset) {
 899                 mutex_exit(&os->os_lock);
 900                 rw_exit(&os_lock);
 901                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
 902                 return (KMEM_CBRC_DONT_KNOW);
 903         }
 904 
 905         /*
 906          * At this point we know that as long as we hold os->os_lock, the dnode
 907          * cannot be freed and fields within the dnode can be safely accessed.
 908          * The objset listing this dnode cannot go away as long as this dnode is
 909          * on its list.
 910          */
 911         rw_exit(&os_lock);
 912         if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
 913                 mutex_exit(&os->os_lock);
 914                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
 915                 return (KMEM_CBRC_NO);
 916         }
 917         ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
 918 
 919         /*
 920          * Lock the dnode handle to prevent the dnode from obtaining any new
 921          * holds. This also prevents the descendant dbufs and the bonus dbuf
 922          * from accessing the dnode, so that we can discount their holds. The
 923          * handle is safe to access because we know that while the dnode cannot
 924          * go away, neither can its handle. Once we hold dnh_zrlock, we can
 925          * safely move any dnode referenced only by dbufs.
 926          */
 927         if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
 928                 mutex_exit(&os->os_lock);
 929                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
 930                 return (KMEM_CBRC_LATER);
 931         }
 932 
 933         /*
 934          * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
 935          * We need to guarantee that there is a hold for every dbuf in order to
 936          * determine whether the dnode is actively referenced. Falsely matching
 937          * a dbuf to an active hold would lead to an unsafe move. It's possible
 938          * that a thread already having an active dnode hold is about to add a
 939          * dbuf, and we can't compare hold and dbuf counts while the add is in
 940          * progress.
 941          */
 942         if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
 943                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 944                 mutex_exit(&os->os_lock);
 945                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
 946                 return (KMEM_CBRC_LATER);
 947         }
 948 
 949         /*
 950          * A dbuf may be removed (evicted) without an active dnode hold. In that
 951          * case, the dbuf count is decremented under the handle lock before the
 952          * dbuf's hold is released. This order ensures that if we count the hold
 953          * after the dbuf is removed but before its hold is released, we will
 954          * treat the unmatched hold as active and exit safely. If we count the
 955          * hold before the dbuf is removed, the hold is discounted, and the
 956          * removal is blocked until the move completes.
 957          */
 958         refcount = refcount_count(&odn->dn_holds);
 959         ASSERT(refcount >= 0);
 960         dbufs = odn->dn_dbufs_count;
 961 
 962         /* We can't have more dbufs than dnode holds. */
 963         ASSERT3U(dbufs, <=, refcount);
 964         DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
 965             uint32_t, dbufs);
 966 
 967         if (refcount > dbufs) {
 968                 rw_exit(&odn->dn_struct_rwlock);
 969                 zrl_exit(&odn->dn_handle->dnh_zrlock);
 970                 mutex_exit(&os->os_lock);
 971                 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
 972                 return (KMEM_CBRC_LATER);
 973         }
 974 
 975         rw_exit(&odn->dn_struct_rwlock);
 976 
 977         /*
 978          * At this point we know that anyone with a hold on the dnode is not
 979          * actively referencing it. The dnode is known and in a valid state to
 980          * move. We're holding the locks needed to execute the critical section.
 981          */
 982         dnode_move_impl(odn, ndn);
 983 
 984         list_link_replace(&odn->dn_link, &ndn->dn_link);
 985         /* If the dnode was safe to move, the refcount cannot have changed. */
 986         ASSERT(refcount == refcount_count(&ndn->dn_holds));
 987         ASSERT(dbufs == ndn->dn_dbufs_count);
 988         zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
 989         mutex_exit(&os->os_lock);
 990 
 991         return (KMEM_CBRC_YES);
 992 }
 993 #endif  /* _KERNEL */
 994 
 995 void
 996 dnode_special_close(dnode_handle_t *dnh)
 997 {
 998         dnode_t *dn = dnh->dnh_dnode;
 999 
1000         /*
1001          * Wait for final references to the dnode to clear.  This can
1002          * only happen if the arc is asyncronously evicting state that
1003          * has a hold on this dnode while we are trying to evict this
1004          * dnode.
1005          */
1006         while (refcount_count(&dn->dn_holds) > 0)
1007                 delay(1);
1008         ASSERT(dn->dn_dbuf == NULL ||
1009             dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1010         zrl_add(&dnh->dnh_zrlock);
1011         dnode_destroy(dn); /* implicit zrl_remove() */
1012         zrl_destroy(&dnh->dnh_zrlock);
1013         dnh->dnh_dnode = NULL;
1014 }
1015 
1016 void
1017 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1018     dnode_handle_t *dnh)
1019 {
1020         dnode_t *dn;
1021 
1022         dn = dnode_create(os, dnp, NULL, object, dnh);
1023         zrl_init(&dnh->dnh_zrlock);
1024         DNODE_VERIFY(dn);
1025 }
1026 
1027 static void
1028 dnode_buf_evict_async(void *dbu)
1029 {
1030         dnode_children_t *children_dnodes = dbu;
1031         int i;
1032 
1033         for (i = 0; i < children_dnodes->dnc_count; i++) {
1034                 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1035                 dnode_t *dn;
1036 
1037                 /*
1038                  * The dnode handle lock guards against the dnode moving to
1039                  * another valid address, so there is no need here to guard
1040                  * against changes to or from NULL.
1041                  */
1042                 if (dnh->dnh_dnode == NULL) {
1043                         zrl_destroy(&dnh->dnh_zrlock);
1044                         continue;
1045                 }
1046 
1047                 zrl_add(&dnh->dnh_zrlock);
1048                 dn = dnh->dnh_dnode;
1049                 /*
1050                  * If there are holds on this dnode, then there should
1051                  * be holds on the dnode's containing dbuf as well; thus
1052                  * it wouldn't be eligible for eviction and this function
1053                  * would not have been called.
1054                  */
1055                 ASSERT(refcount_is_zero(&dn->dn_holds));
1056                 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1057 
1058                 dnode_destroy(dn); /* implicit zrl_remove() */
1059                 zrl_destroy(&dnh->dnh_zrlock);
1060                 dnh->dnh_dnode = NULL;
1061         }
1062         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1063             children_dnodes->dnc_count * sizeof (dnode_handle_t));
1064 }
1065 
1066 /*
1067  * errors:
1068  * EINVAL - invalid object number.
1069  * EIO - i/o error.
1070  * succeeds even for free dnodes.
1071  */
1072 int
1073 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1074     void *tag, dnode_t **dnp)
1075 {
1076         int epb, idx, err;
1077         int drop_struct_lock = FALSE;
1078         int type;
1079         uint64_t blk;
1080         dnode_t *mdn, *dn;
1081         dmu_buf_impl_t *db;
1082         dnode_children_t *children_dnodes;
1083         dnode_handle_t *dnh;
1084 
1085         /*
1086          * If you are holding the spa config lock as writer, you shouldn't
1087          * be asking the DMU to do *anything* unless it's the root pool
1088          * which may require us to read from the root filesystem while
1089          * holding some (not all) of the locks as writer.
1090          */
1091         ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1092             (spa_is_root(os->os_spa) &&
1093             spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1094 
1095         if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1096                 dn = (object == DMU_USERUSED_OBJECT) ?
1097                     DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1098                 if (dn == NULL)
1099                         return (SET_ERROR(ENOENT));
1100                 type = dn->dn_type;
1101                 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1102                         return (SET_ERROR(ENOENT));
1103                 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1104                         return (SET_ERROR(EEXIST));
1105                 DNODE_VERIFY(dn);
1106                 (void) refcount_add(&dn->dn_holds, tag);
1107                 *dnp = dn;
1108                 return (0);
1109         }
1110 
1111         if (object == 0 || object >= DN_MAX_OBJECT)
1112                 return (SET_ERROR(EINVAL));
1113 
1114         mdn = DMU_META_DNODE(os);
1115         ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1116 
1117         DNODE_VERIFY(mdn);
1118 
1119         if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1120                 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1121                 drop_struct_lock = TRUE;
1122         }
1123 
1124         blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1125 
1126         db = dbuf_hold(mdn, blk, FTAG);
1127         if (drop_struct_lock)
1128                 rw_exit(&mdn->dn_struct_rwlock);
1129         if (db == NULL)
1130                 return (SET_ERROR(EIO));
1131         err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1132         if (err) {
1133                 dbuf_rele(db, FTAG);
1134                 return (err);
1135         }
1136 
1137         ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1138         epb = db->db.db_size >> DNODE_SHIFT;
1139 
1140         idx = object & (epb-1);
1141 
1142         ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1143         children_dnodes = dmu_buf_get_user(&db->db);
1144         if (children_dnodes == NULL) {
1145                 int i;
1146                 dnode_children_t *winner;
1147                 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1148                     epb * sizeof (dnode_handle_t), KM_SLEEP);
1149                 children_dnodes->dnc_count = epb;
1150                 dnh = &children_dnodes->dnc_children[0];
1151                 for (i = 0; i < epb; i++) {
1152                         zrl_init(&dnh[i].dnh_zrlock);
1153                 }
1154                 dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1155                     dnode_buf_evict_async, NULL);
1156                 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1157                 if (winner != NULL) {
1158 
1159                         for (i = 0; i < epb; i++) {
1160                                 zrl_destroy(&dnh[i].dnh_zrlock);
1161                         }
1162 
1163                         kmem_free(children_dnodes, sizeof (dnode_children_t) +
1164                             epb * sizeof (dnode_handle_t));
1165                         children_dnodes = winner;
1166                 }
1167         }
1168         ASSERT(children_dnodes->dnc_count == epb);
1169 
1170         dnh = &children_dnodes->dnc_children[idx];
1171         zrl_add(&dnh->dnh_zrlock);
1172         dn = dnh->dnh_dnode;
1173         if (dn == NULL) {
1174                 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1175 
1176                 dn = dnode_create(os, phys, db, object, dnh);
1177         }
1178 
1179         mutex_enter(&dn->dn_mtx);
1180         type = dn->dn_type;
1181         if (dn->dn_free_txg ||
1182             ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1183             ((flag & DNODE_MUST_BE_FREE) &&
1184             (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1185                 mutex_exit(&dn->dn_mtx);
1186                 zrl_remove(&dnh->dnh_zrlock);
1187                 dbuf_rele(db, FTAG);
1188                 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1189         }
1190         if (refcount_add(&dn->dn_holds, tag) == 1)
1191                 dbuf_add_ref(db, dnh);
1192         mutex_exit(&dn->dn_mtx);
1193 
1194         /* Now we can rely on the hold to prevent the dnode from moving. */
1195         zrl_remove(&dnh->dnh_zrlock);
1196 
1197         DNODE_VERIFY(dn);
1198         ASSERT3P(dn->dn_dbuf, ==, db);
1199         ASSERT3U(dn->dn_object, ==, object);
1200         dbuf_rele(db, FTAG);
1201 
1202         *dnp = dn;
1203         return (0);
1204 }
1205 
1206 /*
1207  * Return held dnode if the object is allocated, NULL if not.
1208  */
1209 int
1210 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1211 {
1212         return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1213 }
1214 
1215 /*
1216  * Can only add a reference if there is already at least one
1217  * reference on the dnode.  Returns FALSE if unable to add a
1218  * new reference.
1219  */
1220 boolean_t
1221 dnode_add_ref(dnode_t *dn, void *tag)
1222 {
1223         mutex_enter(&dn->dn_mtx);
1224         if (refcount_is_zero(&dn->dn_holds)) {
1225                 mutex_exit(&dn->dn_mtx);
1226                 return (FALSE);
1227         }
1228         VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1229         mutex_exit(&dn->dn_mtx);
1230         return (TRUE);
1231 }
1232 
1233 void
1234 dnode_rele(dnode_t *dn, void *tag)
1235 {
1236         mutex_enter(&dn->dn_mtx);
1237         dnode_rele_and_unlock(dn, tag);
1238 }
1239 
1240 void
1241 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1242 {
1243         uint64_t refs;
1244         /* Get while the hold prevents the dnode from moving. */
1245         dmu_buf_impl_t *db = dn->dn_dbuf;
1246         dnode_handle_t *dnh = dn->dn_handle;
1247 
1248         refs = refcount_remove(&dn->dn_holds, tag);
1249         mutex_exit(&dn->dn_mtx);
1250 
1251         /*
1252          * It's unsafe to release the last hold on a dnode by dnode_rele() or
1253          * indirectly by dbuf_rele() while relying on the dnode handle to
1254          * prevent the dnode from moving, since releasing the last hold could
1255          * result in the dnode's parent dbuf evicting its dnode handles. For
1256          * that reason anyone calling dnode_rele() or dbuf_rele() without some
1257          * other direct or indirect hold on the dnode must first drop the dnode
1258          * handle.
1259          */
1260         ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1261 
1262         /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1263         if (refs == 0 && db != NULL) {
1264                 /*
1265                  * Another thread could add a hold to the dnode handle in
1266                  * dnode_hold_impl() while holding the parent dbuf. Since the
1267                  * hold on the parent dbuf prevents the handle from being
1268                  * destroyed, the hold on the handle is OK. We can't yet assert
1269                  * that the handle has zero references, but that will be
1270                  * asserted anyway when the handle gets destroyed.
1271                  */
1272                 dbuf_rele(db, dnh);
1273         }
1274 }
1275 
1276 void
1277 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1278 {
1279         dnode_setdirty_sc(dn, tx, B_TRUE);
1280 }
1281 
1282 void
1283 dnode_setdirty_sc(dnode_t *dn, dmu_tx_t *tx, boolean_t usesc)
1284 {
1285         objset_t *os = dn->dn_objset;
1286         uint64_t txg = tx->tx_txg;
1287 
1288         if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1289                 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1290                 return;
1291         }
1292 
1293         DNODE_VERIFY(dn);
1294 
1295 #ifdef ZFS_DEBUG
1296         mutex_enter(&dn->dn_mtx);
1297         ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1298         ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1299         mutex_exit(&dn->dn_mtx);
1300 #endif
1301 
1302         /*
1303          * Determine old uid/gid when necessary
1304          */
1305         dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1306 
1307         multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1308         multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1309 
1310         /*
1311          * If we are already marked dirty, we're done.
1312          */
1313         if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1314                 multilist_sublist_unlock(mls);
1315                 return;
1316         }
1317 
1318         ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1319             !avl_is_empty(&dn->dn_dbufs));
1320         ASSERT(dn->dn_datablksz != 0);
1321         ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1322         ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1323         ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1324 
1325         dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1326             dn->dn_object, txg);
1327 
1328         multilist_sublist_insert_head(mls, dn);
1329 
1330         multilist_sublist_unlock(mls);
1331 
1332         /*
1333          * The dnode maintains a hold on its containing dbuf as
1334          * long as there are holds on it.  Each instantiated child
1335          * dbuf maintains a hold on the dnode.  When the last child
1336          * drops its hold, the dnode will drop its hold on the
1337          * containing dbuf. We add a "dirty hold" here so that the
1338          * dnode will hang around after we finish processing its
1339          * children.
1340          */
1341         VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1342 
1343         (void) dbuf_dirty_sc(dn->dn_dbuf, tx, usesc);
1344         dsl_dataset_dirty(os->os_dsl_dataset, tx);
1345 }
1346 
1347 void
1348 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1349 {
1350         mutex_enter(&dn->dn_mtx);
1351         if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1352                 mutex_exit(&dn->dn_mtx);
1353                 return;
1354         }
1355         dn->dn_free_txg = tx->tx_txg;
1356         mutex_exit(&dn->dn_mtx);
1357 
1358         dnode_setdirty(dn, tx);
1359 }
1360 
1361 /*
1362  * Try to change the block size for the indicated dnode.  This can only
1363  * succeed if there are no blocks allocated or dirty beyond first block
1364  */
1365 int
1366 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1367 {
1368         dmu_buf_impl_t *db;
1369         int err;
1370 
1371         ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1372         if (size == 0)
1373                 size = SPA_MINBLOCKSIZE;
1374         else
1375                 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1376 
1377         if (ibs == dn->dn_indblkshift)
1378                 ibs = 0;
1379 
1380         if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1381                 return (0);
1382 
1383         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1384 
1385         /* Check for any allocated blocks beyond the first */
1386         if (dn->dn_maxblkid != 0)
1387                 goto fail;
1388 
1389         mutex_enter(&dn->dn_dbufs_mtx);
1390         for (db = avl_first(&dn->dn_dbufs); db != NULL;
1391             db = AVL_NEXT(&dn->dn_dbufs, db)) {
1392                 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1393                     db->db_blkid != DMU_SPILL_BLKID) {
1394                         mutex_exit(&dn->dn_dbufs_mtx);
1395                         goto fail;
1396                 }
1397         }
1398         mutex_exit(&dn->dn_dbufs_mtx);
1399 
1400         if (ibs && dn->dn_nlevels != 1)
1401                 goto fail;
1402 
1403         /* resize the old block */
1404         err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1405         if (err == 0)
1406                 dbuf_new_size(db, size, tx);
1407         else if (err != ENOENT)
1408                 goto fail;
1409 
1410         dnode_setdblksz(dn, size);
1411         dnode_setdirty(dn, tx);
1412         dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1413         if (ibs) {
1414                 dn->dn_indblkshift = ibs;
1415                 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1416         }
1417         /* rele after we have fixed the blocksize in the dnode */
1418         if (db)
1419                 dbuf_rele(db, FTAG);
1420 
1421         rw_exit(&dn->dn_struct_rwlock);
1422         return (0);
1423 
1424 fail:
1425         rw_exit(&dn->dn_struct_rwlock);
1426         return (SET_ERROR(ENOTSUP));
1427 }
1428 
1429 /* read-holding callers must not rely on the lock being continuously held */
1430 void
1431 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx,
1432     boolean_t usesc, boolean_t have_read)
1433 {
1434         uint64_t txgoff = tx->tx_txg & TXG_MASK;
1435         int epbs, new_nlevels;
1436         uint64_t sz;
1437 
1438         ASSERT(blkid != DMU_BONUS_BLKID);
1439 
1440         ASSERT(have_read ?
1441             RW_READ_HELD(&dn->dn_struct_rwlock) :
1442             RW_WRITE_HELD(&dn->dn_struct_rwlock));
1443 
1444         /*
1445          * if we have a read-lock, check to see if we need to do any work
1446          * before upgrading to a write-lock.
1447          */
1448         if (have_read) {
1449                 if (blkid <= dn->dn_maxblkid)
1450                         return;
1451 
1452                 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1453                         rw_exit(&dn->dn_struct_rwlock);
1454                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1455                 }
1456         }
1457 
1458         if (blkid <= dn->dn_maxblkid)
1459                 goto out;
1460 
1461         dn->dn_maxblkid = blkid;
1462 
1463         /*
1464          * Compute the number of levels necessary to support the new maxblkid.
1465          */
1466         new_nlevels = 1;
1467         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1468         for (sz = dn->dn_nblkptr;
1469             sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1470                 new_nlevels++;
1471 
1472         if (new_nlevels > dn->dn_nlevels) {
1473                 int old_nlevels = dn->dn_nlevels;
1474                 dmu_buf_impl_t *db;
1475                 list_t *list;
1476                 dbuf_dirty_record_t *new, *dr, *dr_next;
1477 
1478                 dn->dn_nlevels = new_nlevels;
1479 
1480                 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1481                 dn->dn_next_nlevels[txgoff] = new_nlevels;
1482 
1483                 /* dirty the left indirects */
1484                 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1485                 ASSERT(db != NULL);
1486                 new = dbuf_dirty_sc(db, tx, usesc);
1487                 dbuf_rele(db, FTAG);
1488 
1489                 /* transfer the dirty records to the new indirect */
1490                 mutex_enter(&dn->dn_mtx);
1491                 mutex_enter(&new->dt.di.dr_mtx);
1492                 list = &dn->dn_dirty_records[txgoff];
1493                 for (dr = list_head(list); dr; dr = dr_next) {
1494                         dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1495                         if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1496                             dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1497                             dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1498                                 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1499                                 list_remove(&dn->dn_dirty_records[txgoff], dr);
1500                                 list_insert_tail(&new->dt.di.dr_children, dr);
1501                                 dr->dr_parent = new;
1502                         }
1503                 }
1504                 mutex_exit(&new->dt.di.dr_mtx);
1505                 mutex_exit(&dn->dn_mtx);
1506         }
1507 
1508 out:
1509         if (have_read)
1510                 rw_downgrade(&dn->dn_struct_rwlock);
1511 }
1512 
1513 static void
1514 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1515 {
1516         dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1517         if (db != NULL) {
1518                 dmu_buf_will_dirty(&db->db, tx);
1519                 dbuf_rele(db, FTAG);
1520         }
1521 }
1522 
1523 void
1524 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1525 {
1526         dmu_buf_impl_t *db;
1527         uint64_t blkoff, blkid, nblks;
1528         int blksz, blkshift, head, tail;
1529         int trunc = FALSE;
1530         int epbs;
1531 
1532         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1533         blksz = dn->dn_datablksz;
1534         blkshift = dn->dn_datablkshift;
1535         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1536 
1537         if (len == DMU_OBJECT_END) {
1538                 len = UINT64_MAX - off;
1539                 trunc = TRUE;
1540         }
1541 
1542         /*
1543          * First, block align the region to free:
1544          */
1545         if (ISP2(blksz)) {
1546                 head = P2NPHASE(off, blksz);
1547                 blkoff = P2PHASE(off, blksz);
1548                 if ((off >> blkshift) > dn->dn_maxblkid)
1549                         goto out;
1550         } else {
1551                 ASSERT(dn->dn_maxblkid == 0);
1552                 if (off == 0 && len >= blksz) {
1553                         /*
1554                          * Freeing the whole block; fast-track this request.
1555                          * Note that we won't dirty any indirect blocks,
1556                          * which is fine because we will be freeing the entire
1557                          * file and thus all indirect blocks will be freed
1558                          * by free_children().
1559                          */
1560                         blkid = 0;
1561                         nblks = 1;
1562                         goto done;
1563                 } else if (off >= blksz) {
1564                         /* Freeing past end-of-data */
1565                         goto out;
1566                 } else {
1567                         /* Freeing part of the block. */
1568                         head = blksz - off;
1569                         ASSERT3U(head, >, 0);
1570                 }
1571                 blkoff = off;
1572         }
1573         /* zero out any partial block data at the start of the range */
1574         if (head) {
1575                 ASSERT3U(blkoff + head, ==, blksz);
1576                 if (len < head)
1577                         head = len;
1578                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1579                     TRUE, FALSE, FTAG, &db) == 0) {
1580                         caddr_t data;
1581 
1582                         /* don't dirty if it isn't on disk and isn't dirty */
1583                         if (db->db_last_dirty ||
1584                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1585                                 rw_exit(&dn->dn_struct_rwlock);
1586                                 dmu_buf_will_dirty(&db->db, tx);
1587                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1588                                 data = db->db.db_data;
1589                                 bzero(data + blkoff, head);
1590                         }
1591                         dbuf_rele(db, FTAG);
1592                 }
1593                 off += head;
1594                 len -= head;
1595         }
1596 
1597         /* If the range was less than one block, we're done */
1598         if (len == 0)
1599                 goto out;
1600 
1601         /* If the remaining range is past end of file, we're done */
1602         if ((off >> blkshift) > dn->dn_maxblkid)
1603                 goto out;
1604 
1605         ASSERT(ISP2(blksz));
1606         if (trunc)
1607                 tail = 0;
1608         else
1609                 tail = P2PHASE(len, blksz);
1610 
1611         ASSERT0(P2PHASE(off, blksz));
1612         /* zero out any partial block data at the end of the range */
1613         if (tail) {
1614                 if (len < tail)
1615                         tail = len;
1616                 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1617                     TRUE, FALSE, FTAG, &db) == 0) {
1618                         /* don't dirty if not on disk and not dirty */
1619                         if (db->db_last_dirty ||
1620                             (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1621                                 rw_exit(&dn->dn_struct_rwlock);
1622                                 dmu_buf_will_dirty(&db->db, tx);
1623                                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1624                                 bzero(db->db.db_data, tail);
1625                         }
1626                         dbuf_rele(db, FTAG);
1627                 }
1628                 len -= tail;
1629         }
1630 
1631         /* If the range did not include a full block, we are done */
1632         if (len == 0)
1633                 goto out;
1634 
1635         ASSERT(IS_P2ALIGNED(off, blksz));
1636         ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1637         blkid = off >> blkshift;
1638         nblks = len >> blkshift;
1639         if (trunc)
1640                 nblks += 1;
1641 
1642         /*
1643          * Dirty all the indirect blocks in this range.  Note that only
1644          * the first and last indirect blocks can actually be written
1645          * (if they were partially freed) -- they must be dirtied, even if
1646          * they do not exist on disk yet.  The interior blocks will
1647          * be freed by free_children(), so they will not actually be written.
1648          * Even though these interior blocks will not be written, we
1649          * dirty them for two reasons:
1650          *
1651          *  - It ensures that the indirect blocks remain in memory until
1652          *    syncing context.  (They have already been prefetched by
1653          *    dmu_tx_hold_free(), so we don't have to worry about reading
1654          *    them serially here.)
1655          *
1656          *  - The dirty space accounting will put pressure on the txg sync
1657          *    mechanism to begin syncing, and to delay transactions if there
1658          *    is a large amount of freeing.  Even though these indirect
1659          *    blocks will not be written, we could need to write the same
1660          *    amount of space if we copy the freed BPs into deadlists.
1661          */
1662         if (dn->dn_nlevels > 1) {
1663                 uint64_t first, last;
1664 
1665                 first = blkid >> epbs;
1666                 dnode_dirty_l1(dn, first, tx);
1667                 if (trunc)
1668                         last = dn->dn_maxblkid >> epbs;
1669                 else
1670                         last = (blkid + nblks - 1) >> epbs;
1671                 if (last != first)
1672                         dnode_dirty_l1(dn, last, tx);
1673 
1674                 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1675                     SPA_BLKPTRSHIFT;
1676                 for (uint64_t i = first + 1; i < last; i++) {
1677                         /*
1678                          * Set i to the blockid of the next non-hole
1679                          * level-1 indirect block at or after i.  Note
1680                          * that dnode_next_offset() operates in terms of
1681                          * level-0-equivalent bytes.
1682                          */
1683                         uint64_t ibyte = i << shift;
1684                         int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1685                             &ibyte, 2, 1, 0);
1686                         i = ibyte >> shift;
1687                         if (i >= last)
1688                                 break;
1689 
1690                         /*
1691                          * Normally we should not see an error, either
1692                          * from dnode_next_offset() or dbuf_hold_level()
1693                          * (except for ESRCH from dnode_next_offset).
1694                          * If there is an i/o error, then when we read
1695                          * this block in syncing context, it will use
1696                          * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1697                          * to the "failmode" property.  dnode_next_offset()
1698                          * doesn't have a flag to indicate MUSTSUCCEED.
1699                          */
1700                         if (err != 0)
1701                                 break;
1702 
1703                         dnode_dirty_l1(dn, i, tx);
1704                 }
1705         }
1706 
1707 done:
1708         /*
1709          * Add this range to the dnode range list.
1710          * We will finish up this free operation in the syncing phase.
1711          */
1712         mutex_enter(&dn->dn_mtx);
1713         int txgoff = tx->tx_txg & TXG_MASK;
1714         if (dn->dn_free_ranges[txgoff] == NULL) {
1715                 dn->dn_free_ranges[txgoff] =
1716                     range_tree_create(NULL, NULL, &dn->dn_mtx);
1717         }
1718         range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1719         range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1720         dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1721             blkid, nblks, tx->tx_txg);
1722         mutex_exit(&dn->dn_mtx);
1723 
1724         dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1725         dnode_setdirty(dn, tx);
1726 out:
1727 
1728         rw_exit(&dn->dn_struct_rwlock);
1729 }
1730 
1731 static boolean_t
1732 dnode_spill_freed(dnode_t *dn)
1733 {
1734         int i;
1735 
1736         mutex_enter(&dn->dn_mtx);
1737         for (i = 0; i < TXG_SIZE; i++) {
1738                 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1739                         break;
1740         }
1741         mutex_exit(&dn->dn_mtx);
1742         return (i < TXG_SIZE);
1743 }
1744 
1745 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1746 uint64_t
1747 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1748 {
1749         void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1750         int i;
1751 
1752         if (blkid == DMU_BONUS_BLKID)
1753                 return (FALSE);
1754 
1755         /*
1756          * If we're in the process of opening the pool, dp will not be
1757          * set yet, but there shouldn't be anything dirty.
1758          */
1759         if (dp == NULL)
1760                 return (FALSE);
1761 
1762         if (dn->dn_free_txg)
1763                 return (TRUE);
1764 
1765         if (blkid == DMU_SPILL_BLKID)
1766                 return (dnode_spill_freed(dn));
1767 
1768         mutex_enter(&dn->dn_mtx);
1769         for (i = 0; i < TXG_SIZE; i++) {
1770                 if (dn->dn_free_ranges[i] != NULL &&
1771                     range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1772                         break;
1773         }
1774         mutex_exit(&dn->dn_mtx);
1775         return (i < TXG_SIZE);
1776 }
1777 
1778 /* call from syncing context when we actually write/free space for this dnode */
1779 void
1780 dnode_diduse_space(dnode_t *dn, int64_t delta)
1781 {
1782         uint64_t space;
1783         dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1784             dn, dn->dn_phys,
1785             (u_longlong_t)dn->dn_phys->dn_used,
1786             (longlong_t)delta);
1787 
1788         mutex_enter(&dn->dn_mtx);
1789         space = DN_USED_BYTES(dn->dn_phys);
1790         if (delta > 0) {
1791                 ASSERT3U(space + delta, >=, space); /* no overflow */
1792         } else {
1793                 ASSERT3U(space, >=, -delta); /* no underflow */
1794         }
1795         space += delta;
1796         if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1797                 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1798                 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1799                 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1800         } else {
1801                 dn->dn_phys->dn_used = space;
1802                 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1803         }
1804         mutex_exit(&dn->dn_mtx);
1805 }
1806 
1807 /*
1808  * Scans a block at the indicated "level" looking for a hole or data,
1809  * depending on 'flags'.
1810  *
1811  * If level > 0, then we are scanning an indirect block looking at its
1812  * pointers.  If level == 0, then we are looking at a block of dnodes.
1813  *
1814  * If we don't find what we are looking for in the block, we return ESRCH.
1815  * Otherwise, return with *offset pointing to the beginning (if searching
1816  * forwards) or end (if searching backwards) of the range covered by the
1817  * block pointer we matched on (or dnode).
1818  *
1819  * The basic search algorithm used below by dnode_next_offset() is to
1820  * use this function to search up the block tree (widen the search) until
1821  * we find something (i.e., we don't return ESRCH) and then search back
1822  * down the tree (narrow the search) until we reach our original search
1823  * level.
1824  */
1825 static int
1826 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1827     int lvl, uint64_t blkfill, uint64_t txg)
1828 {
1829         dmu_buf_impl_t *db = NULL;
1830         void *data = NULL;
1831         uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1832         uint64_t epb = 1ULL << epbs;
1833         uint64_t minfill, maxfill;
1834         boolean_t hole;
1835         int i, inc, error, span;
1836 
1837         dprintf("probing object %llu offset %llx level %d of %u\n",
1838             dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1839 
1840         hole = ((flags & DNODE_FIND_HOLE) != 0);
1841         inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1842         ASSERT(txg == 0 || !hole);
1843 
1844         if (lvl == dn->dn_phys->dn_nlevels) {
1845                 error = 0;
1846                 epb = dn->dn_phys->dn_nblkptr;
1847                 data = dn->dn_phys->dn_blkptr;
1848         } else {
1849                 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1850                 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1851                 if (error) {
1852                         if (error != ENOENT)
1853                                 return (error);
1854                         if (hole)
1855                                 return (0);
1856                         /*
1857                          * This can only happen when we are searching up
1858                          * the block tree for data.  We don't really need to
1859                          * adjust the offset, as we will just end up looking
1860                          * at the pointer to this block in its parent, and its
1861                          * going to be unallocated, so we will skip over it.
1862                          */
1863                         return (SET_ERROR(ESRCH));
1864                 }
1865                 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1866                 if (error) {
1867                         dbuf_rele(db, FTAG);
1868                         return (error);
1869                 }
1870                 data = db->db.db_data;
1871         }
1872 
1873 
1874         if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1875             db->db_blkptr->blk_birth <= txg ||
1876             BP_IS_HOLE(db->db_blkptr))) {
1877                 /*
1878                  * This can only happen when we are searching up the tree
1879                  * and these conditions mean that we need to keep climbing.
1880                  */
1881                 error = SET_ERROR(ESRCH);
1882         } else if (lvl == 0) {
1883                 dnode_phys_t *dnp = data;
1884                 span = DNODE_SHIFT;
1885                 ASSERT(dn->dn_type == DMU_OT_DNODE);
1886 
1887                 for (i = (*offset >> span) & (blkfill - 1);
1888                     i >= 0 && i < blkfill; i += inc) {
1889                         if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1890                                 break;
1891                         *offset += (1ULL << span) * inc;
1892                 }
1893                 if (i < 0 || i == blkfill)
1894                         error = SET_ERROR(ESRCH);
1895         } else {
1896                 blkptr_t *bp = data;
1897                 uint64_t start = *offset;
1898                 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1899                 minfill = 0;
1900                 maxfill = blkfill << ((lvl - 1) * epbs);
1901 
1902                 if (hole)
1903                         maxfill--;
1904                 else
1905                         minfill++;
1906 
1907                 *offset = *offset >> span;
1908                 for (i = BF64_GET(*offset, 0, epbs);
1909                     i >= 0 && i < epb; i += inc) {
1910                         if (BP_GET_FILL(&bp[i]) >= minfill &&
1911                             BP_GET_FILL(&bp[i]) <= maxfill &&
1912                             (hole || bp[i].blk_birth > txg))
1913                                 break;
1914                         if (inc > 0 || *offset > 0)
1915                                 *offset += inc;
1916                 }
1917                 *offset = *offset << span;
1918                 if (inc < 0) {
1919                         /* traversing backwards; position offset at the end */
1920                         ASSERT3U(*offset, <=, start);
1921                         *offset = MIN(*offset + (1ULL << span) - 1, start);
1922                 } else if (*offset < start) {
1923                         *offset = start;
1924                 }
1925                 if (i < 0 || i >= epb)
1926                         error = SET_ERROR(ESRCH);
1927         }
1928 
1929         if (db)
1930                 dbuf_rele(db, FTAG);
1931 
1932         return (error);
1933 }
1934 
1935 /*
1936  * Find the next hole, data, or sparse region at or after *offset.
1937  * The value 'blkfill' tells us how many items we expect to find
1938  * in an L0 data block; this value is 1 for normal objects,
1939  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1940  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1941  *
1942  * Examples:
1943  *
1944  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1945  *      Finds the next/previous hole/data in a file.
1946  *      Used in dmu_offset_next().
1947  *
1948  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1949  *      Finds the next free/allocated dnode an objset's meta-dnode.
1950  *      Only finds objects that have new contents since txg (ie.
1951  *      bonus buffer changes and content removal are ignored).
1952  *      Used in dmu_object_next().
1953  *
1954  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1955  *      Finds the next L2 meta-dnode bp that's at most 1/4 full.
1956  *      Used in dmu_object_alloc().
1957  */
1958 int
1959 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1960     int minlvl, uint64_t blkfill, uint64_t txg)
1961 {
1962         uint64_t initial_offset = *offset;
1963         int lvl, maxlvl;
1964         int error = 0;
1965 
1966         if (!(flags & DNODE_FIND_HAVELOCK))
1967                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1968 
1969         if (dn->dn_phys->dn_nlevels == 0) {
1970                 error = SET_ERROR(ESRCH);
1971                 goto out;
1972         }
1973 
1974         if (dn->dn_datablkshift == 0) {
1975                 if (*offset < dn->dn_datablksz) {
1976                         if (flags & DNODE_FIND_HOLE)
1977                                 *offset = dn->dn_datablksz;
1978                 } else {
1979                         error = SET_ERROR(ESRCH);
1980                 }
1981                 goto out;
1982         }
1983 
1984         maxlvl = dn->dn_phys->dn_nlevels;
1985 
1986         for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1987                 error = dnode_next_offset_level(dn,
1988                     flags, offset, lvl, blkfill, txg);
1989                 if (error != ESRCH)
1990                         break;
1991         }
1992 
1993         while (error == 0 && --lvl >= minlvl) {
1994                 error = dnode_next_offset_level(dn,
1995                     flags, offset, lvl, blkfill, txg);
1996         }
1997 
1998         /*
1999          * There's always a "virtual hole" at the end of the object, even
2000          * if all BP's which physically exist are non-holes.
2001          */
2002         if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2003             minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2004                 error = 0;
2005         }
2006 
2007         if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2008             initial_offset < *offset : initial_offset > *offset))
2009                 error = SET_ERROR(ESRCH);
2010 out:
2011         if (!(flags & DNODE_FIND_HAVELOCK))
2012                 rw_exit(&dn->dn_struct_rwlock);
2013 
2014         return (error);
2015 }
2016 
2017 /*
2018  * When in the compressing phase, we check our results every 1 MiB. If
2019  * compression ratio drops below the threshold factor, we give up trying
2020  * to compress the file for a while. The length of the interval is
2021  * calculated from this interval value according to the algorithm in
2022  * smartcomp_check_comp.
2023  */
2024 uint64_t zfs_smartcomp_interval = 1 * 1024 * 1024;
2025 
2026 /*
2027  * Minimum compression factor is 12.5% (100% / factor) - below that we
2028  * consider compression to have failed.
2029  */
2030 uint64_t zfs_smartcomp_threshold_factor = 8;
2031 
2032 /*
2033  * Maximum power-of-2 exponent on the deny interval and consequently
2034  * the maximum number of compression successes and failures we track.
2035  * Successive compression failures extend the deny interval, whereas
2036  * repeated successes makes the algorithm more hesitant to start denying.
2037  */
2038 int64_t zfs_smartcomp_interval_exp = 5;
2039 
2040 /*
2041  * Callback invoked by the zio machinery when it wants to compress a data
2042  * block. If we are in the denying compression phase, we add the amount of
2043  * data written to our stats and check if we've denied enough data to
2044  * transition back in to the compression phase again.
2045  */
2046 boolean_t
2047 dnode_smartcomp_ask_cb(void *userinfo, const zio_t *zio)
2048 {
2049         dnode_t *dn = userinfo;
2050         dnode_smartcomp_t *sc;
2051         dnode_smartcomp_state_t old_state;
2052 
2053         ASSERT(dn != NULL);
2054 
2055         sc = &dn->dn_smartcomp;
2056         mutex_enter(&sc->sc_lock);
2057         old_state = sc->sc_state;
2058         if (sc->sc_state == DNODE_SMARTCOMP_DENYING) {
2059                 sc->sc_orig_size += zio->io_orig_size;
2060                 if (sc->sc_orig_size >= sc->sc_deny_interval) {
2061                         /* time to retry compression on next call */
2062                         sc->sc_state = DNODE_SMARTCOMP_COMPRESSING;
2063                         sc->sc_size = 0;
2064                         sc->sc_orig_size = 0;
2065                 }
2066         }
2067         mutex_exit(&sc->sc_lock);
2068 
2069         return (old_state != DNODE_SMARTCOMP_DENYING);
2070 }
2071 
2072 /*
2073  * Callback invoked after compression has been performed to allow us to
2074  * monitor compression performance. If we're in a compressing phase, we
2075  * add the uncompressed and compressed data volumes to our state counters
2076  * and see if we need to recheck compression performance in
2077  * smartcomp_check_comp.
2078  */
2079 void
2080 dnode_smartcomp_result_cb(void *userinfo, const zio_t *zio)
2081 {
2082         dnode_t *dn = userinfo;
2083         dnode_smartcomp_t *sc;
2084         uint64_t io_size = zio->io_size, io_orig_size = zio->io_orig_size;
2085 
2086         ASSERT(dn != NULL);
2087         sc = &dn->dn_smartcomp;
2088 
2089         if (io_orig_size == 0)
2090                 /* XXX: is this valid anyway? */
2091                 return;
2092 
2093         mutex_enter(&sc->sc_lock);
2094         if (sc->sc_state == DNODE_SMARTCOMP_COMPRESSING) {
2095                 /* add last block's compression performance to our stats */
2096                 sc->sc_size += io_size;
2097                 sc->sc_orig_size += io_orig_size;
2098                 /* time to recheck compression performance? */
2099                 if (sc->sc_orig_size >= zfs_smartcomp_interval)
2100                         smartcomp_check_comp(sc);
2101         }
2102         mutex_exit(&sc->sc_lock);
2103 }
2104 
2105 /*
2106  * This function checks whether the compression we've been getting is above
2107  * the threshold value. If it is, we decrement the sc_comp_failures counter
2108  * to indicate compression success. If it isn't we increment the same
2109  * counter and potentially start a compression deny phase.
2110  */
2111 static void
2112 smartcomp_check_comp(dnode_smartcomp_t *sc)
2113 {
2114         uint64_t threshold = sc->sc_orig_size -
2115             sc->sc_orig_size / zfs_smartcomp_threshold_factor;
2116 
2117         ASSERT(MUTEX_HELD(&sc->sc_lock));
2118         if (sc->sc_size > threshold) {
2119                 sc->sc_comp_failures =
2120                     MIN(sc->sc_comp_failures + 1, zfs_smartcomp_interval_exp);
2121                 if (sc->sc_comp_failures > 0) {
2122                         /* consistently getting too little compression, stop */
2123                         sc->sc_state = DNODE_SMARTCOMP_DENYING;
2124                         sc->sc_deny_interval =
2125                             zfs_smartcomp_interval << sc->sc_comp_failures;
2126                         /* randomize the interval by +-10% to avoid patterns */
2127                         sc->sc_deny_interval = (sc->sc_deny_interval -
2128                             (sc->sc_deny_interval / 10)) +
2129                             spa_get_random(sc->sc_deny_interval / 5 + 1);
2130                 }
2131         } else {
2132                 if (sc->sc_comp_failures > 0) {
2133                         /*
2134                          * We're biased for compression, so any success makes
2135                          * us forget the file's past incompressibility.
2136                          */
2137                         sc->sc_comp_failures = 0;
2138                 } else {
2139                         sc->sc_comp_failures = MAX(sc->sc_comp_failures - 1,
2140                             -zfs_smartcomp_interval_exp);
2141                 }
2142         }
2143         /* reset state counters */
2144         sc->sc_size = 0;
2145         sc->sc_orig_size = 0;
2146 }
2147 
2148 /*
2149  * Prepares a zio_smartcomp_info_t structure for passing to zio_write or
2150  * arc_write depending on whether smart compression should be applied to
2151  * the specified objset, dnode and buffer.
2152  */
2153 extern void
2154 dnode_setup_zio_smartcomp(dmu_buf_impl_t *db, zio_smartcomp_info_t *sc)
2155 {
2156         dnode_t *dn = DB_DNODE(db);
2157         objset_t *os = dn->dn_objset;
2158 
2159         /* Only do smart compression on user data of plain files. */
2160         if (dn->dn_type == DMU_OT_PLAIN_FILE_CONTENTS && db->db_level == 0 &&
2161             os->os_smartcomp_enabled && os->os_compress != ZIO_COMPRESS_OFF) {
2162                 sc->sc_ask = dnode_smartcomp_ask_cb;
2163                 sc->sc_result = dnode_smartcomp_result_cb;
2164                 sc->sc_userinfo = dn;
2165         } else {
2166                 /*
2167                  * Zeroing out the structure passed to zio_write will turn
2168                  * smart compression off.
2169                  */
2170                 bzero(sc, sizeof (*sc));
2171         }
2172 }