1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012 by Delphix. All rights reserved.
  24  */
  25 
  26 #include <sys/dmu.h>
  27 #include <sys/dmu_impl.h>
  28 #include <sys/dmu_tx.h>
  29 #include <sys/dbuf.h>
  30 #include <sys/dnode.h>
  31 #include <sys/zfs_context.h>
  32 #include <sys/dmu_objset.h>
  33 #include <sys/dmu_traverse.h>
  34 #include <sys/dsl_dataset.h>
  35 #include <sys/dsl_dir.h>
  36 #include <sys/dsl_pool.h>
  37 #include <sys/dsl_synctask.h>
  38 #include <sys/dsl_prop.h>
  39 #include <sys/dmu_zfetch.h>
  40 #include <sys/zfs_ioctl.h>
  41 #include <sys/zap.h>
  42 #include <sys/zio_checksum.h>
  43 #include <sys/sa.h>
  44 #ifdef _KERNEL
  45 #include <sys/vmsystm.h>
  46 #include <sys/zfs_znode.h>
  47 #endif
  48 
  49 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
  50         {       DMU_BSWAP_UINT8,        TRUE,   "unallocated"           },
  51         {       DMU_BSWAP_ZAP,          TRUE,   "object directory"      },
  52         {       DMU_BSWAP_UINT64,       TRUE,   "object array"          },
  53         {       DMU_BSWAP_UINT8,        TRUE,   "packed nvlist"         },
  54         {       DMU_BSWAP_UINT64,       TRUE,   "packed nvlist size"    },
  55         {       DMU_BSWAP_UINT64,       TRUE,   "bpobj"                 },
  56         {       DMU_BSWAP_UINT64,       TRUE,   "bpobj header"          },
  57         {       DMU_BSWAP_UINT64,       TRUE,   "SPA space map header"  },
  58         {       DMU_BSWAP_UINT64,       TRUE,   "SPA space map"         },
  59         {       DMU_BSWAP_UINT64,       TRUE,   "ZIL intent log"        },
  60         {       DMU_BSWAP_DNODE,        TRUE,   "DMU dnode"             },
  61         {       DMU_BSWAP_OBJSET,       TRUE,   "DMU objset"            },
  62         {       DMU_BSWAP_UINT64,       TRUE,   "DSL directory"         },
  63         {       DMU_BSWAP_ZAP,          TRUE,   "DSL directory child map"},
  64         {       DMU_BSWAP_ZAP,          TRUE,   "DSL dataset snap map"  },
  65         {       DMU_BSWAP_ZAP,          TRUE,   "DSL props"             },
  66         {       DMU_BSWAP_UINT64,       TRUE,   "DSL dataset"           },
  67         {       DMU_BSWAP_ZNODE,        TRUE,   "ZFS znode"             },
  68         {       DMU_BSWAP_OLDACL,       TRUE,   "ZFS V0 ACL"            },
  69         {       DMU_BSWAP_UINT8,        FALSE,  "ZFS plain file"        },
  70         {       DMU_BSWAP_ZAP,          TRUE,   "ZFS directory"         },
  71         {       DMU_BSWAP_ZAP,          TRUE,   "ZFS master node"       },
  72         {       DMU_BSWAP_ZAP,          TRUE,   "ZFS delete queue"      },
  73         {       DMU_BSWAP_UINT8,        FALSE,  "zvol object"           },
  74         {       DMU_BSWAP_ZAP,          TRUE,   "zvol prop"             },
  75         {       DMU_BSWAP_UINT8,        FALSE,  "other uint8[]"         },
  76         {       DMU_BSWAP_UINT64,       FALSE,  "other uint64[]"        },
  77         {       DMU_BSWAP_ZAP,          TRUE,   "other ZAP"             },
  78         {       DMU_BSWAP_ZAP,          TRUE,   "persistent error log"  },
  79         {       DMU_BSWAP_UINT8,        TRUE,   "SPA history"           },
  80         {       DMU_BSWAP_UINT64,       TRUE,   "SPA history offsets"   },
  81         {       DMU_BSWAP_ZAP,          TRUE,   "Pool properties"       },
  82         {       DMU_BSWAP_ZAP,          TRUE,   "DSL permissions"       },
  83         {       DMU_BSWAP_ACL,          TRUE,   "ZFS ACL"               },
  84         {       DMU_BSWAP_UINT8,        TRUE,   "ZFS SYSACL"            },
  85         {       DMU_BSWAP_UINT8,        TRUE,   "FUID table"            },
  86         {       DMU_BSWAP_UINT64,       TRUE,   "FUID table size"       },
  87         {       DMU_BSWAP_ZAP,          TRUE,   "DSL dataset next clones"},
  88         {       DMU_BSWAP_ZAP,          TRUE,   "scan work queue"       },
  89         {       DMU_BSWAP_ZAP,          TRUE,   "ZFS user/group used"   },
  90         {       DMU_BSWAP_ZAP,          TRUE,   "ZFS user/group quota"  },
  91         {       DMU_BSWAP_ZAP,          TRUE,   "snapshot refcount tags"},
  92         {       DMU_BSWAP_ZAP,          TRUE,   "DDT ZAP algorithm"     },
  93         {       DMU_BSWAP_ZAP,          TRUE,   "DDT statistics"        },
  94         {       DMU_BSWAP_UINT8,        TRUE,   "System attributes"     },
  95         {       DMU_BSWAP_ZAP,          TRUE,   "SA master node"        },
  96         {       DMU_BSWAP_ZAP,          TRUE,   "SA attr registration"  },
  97         {       DMU_BSWAP_ZAP,          TRUE,   "SA attr layouts"       },
  98         {       DMU_BSWAP_ZAP,          TRUE,   "scan translations"     },
  99         {       DMU_BSWAP_UINT8,        FALSE,  "deduplicated block"    },
 100         {       DMU_BSWAP_ZAP,          TRUE,   "DSL deadlist map"      },
 101         {       DMU_BSWAP_UINT64,       TRUE,   "DSL deadlist map hdr"  },
 102         {       DMU_BSWAP_ZAP,          TRUE,   "DSL dir clones"        },
 103         {       DMU_BSWAP_UINT64,       TRUE,   "bpobj subobj"          }
 104 };
 105 
 106 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
 107         {       byteswap_uint8_array,   "uint8"         },
 108         {       byteswap_uint16_array,  "uint16"        },
 109         {       byteswap_uint32_array,  "uint32"        },
 110         {       byteswap_uint64_array,  "uint64"        },
 111         {       zap_byteswap,           "zap"           },
 112         {       dnode_buf_byteswap,     "dnode"         },
 113         {       dmu_objset_byteswap,    "objset"        },
 114         {       zfs_znode_byteswap,     "znode"         },
 115         {       zfs_oldacl_byteswap,    "oldacl"        },
 116         {       zfs_acl_byteswap,       "acl"           }
 117 };
 118 
 119 int
 120 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
 121     void *tag, dmu_buf_t **dbp, int flags)
 122 {
 123         dnode_t *dn;
 124         uint64_t blkid;
 125         dmu_buf_impl_t *db;
 126         int err;
 127         int db_flags = DB_RF_CANFAIL;
 128 
 129         if (flags & DMU_READ_NO_PREFETCH)
 130                 db_flags |= DB_RF_NOPREFETCH;
 131 
 132         err = dnode_hold(os, object, FTAG, &dn);
 133         if (err)
 134                 return (err);
 135         blkid = dbuf_whichblock(dn, offset);
 136         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 137         db = dbuf_hold(dn, blkid, tag);
 138         rw_exit(&dn->dn_struct_rwlock);
 139         if (db == NULL) {
 140                 err = EIO;
 141         } else {
 142                 err = dbuf_read(db, NULL, db_flags);
 143                 if (err) {
 144                         dbuf_rele(db, tag);
 145                         db = NULL;
 146                 }
 147         }
 148 
 149         dnode_rele(dn, FTAG);
 150         *dbp = &db->db; /* NULL db plus first field offset is NULL */
 151         return (err);
 152 }
 153 
 154 int
 155 dmu_bonus_max(void)
 156 {
 157         return (DN_MAX_BONUSLEN);
 158 }
 159 
 160 int
 161 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
 162 {
 163         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
 164         dnode_t *dn;
 165         int error;
 166 
 167         DB_DNODE_ENTER(db);
 168         dn = DB_DNODE(db);
 169 
 170         if (dn->dn_bonus != db) {
 171                 error = EINVAL;
 172         } else if (newsize < 0 || newsize > db_fake->db_size) {
 173                 error = EINVAL;
 174         } else {
 175                 dnode_setbonuslen(dn, newsize, tx);
 176                 error = 0;
 177         }
 178 
 179         DB_DNODE_EXIT(db);
 180         return (error);
 181 }
 182 
 183 int
 184 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
 185 {
 186         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
 187         dnode_t *dn;
 188         int error;
 189 
 190         DB_DNODE_ENTER(db);
 191         dn = DB_DNODE(db);
 192 
 193         if (!DMU_OT_IS_VALID(type)) {
 194                 error = EINVAL;
 195         } else if (dn->dn_bonus != db) {
 196                 error = EINVAL;
 197         } else {
 198                 dnode_setbonus_type(dn, type, tx);
 199                 error = 0;
 200         }
 201 
 202         DB_DNODE_EXIT(db);
 203         return (error);
 204 }
 205 
 206 dmu_object_type_t
 207 dmu_get_bonustype(dmu_buf_t *db_fake)
 208 {
 209         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
 210         dnode_t *dn;
 211         dmu_object_type_t type;
 212 
 213         DB_DNODE_ENTER(db);
 214         dn = DB_DNODE(db);
 215         type = dn->dn_bonustype;
 216         DB_DNODE_EXIT(db);
 217 
 218         return (type);
 219 }
 220 
 221 int
 222 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
 223 {
 224         dnode_t *dn;
 225         int error;
 226 
 227         error = dnode_hold(os, object, FTAG, &dn);
 228         dbuf_rm_spill(dn, tx);
 229         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 230         dnode_rm_spill(dn, tx);
 231         rw_exit(&dn->dn_struct_rwlock);
 232         dnode_rele(dn, FTAG);
 233         return (error);
 234 }
 235 
 236 /*
 237  * returns ENOENT, EIO, or 0.
 238  */
 239 int
 240 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
 241 {
 242         dnode_t *dn;
 243         dmu_buf_impl_t *db;
 244         int error;
 245 
 246         error = dnode_hold(os, object, FTAG, &dn);
 247         if (error)
 248                 return (error);
 249 
 250         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 251         if (dn->dn_bonus == NULL) {
 252                 rw_exit(&dn->dn_struct_rwlock);
 253                 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
 254                 if (dn->dn_bonus == NULL)
 255                         dbuf_create_bonus(dn);
 256         }
 257         db = dn->dn_bonus;
 258 
 259         /* as long as the bonus buf is held, the dnode will be held */
 260         if (refcount_add(&db->db_holds, tag) == 1) {
 261                 VERIFY(dnode_add_ref(dn, db));
 262                 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
 263         }
 264 
 265         /*
 266          * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
 267          * hold and incrementing the dbuf count to ensure that dnode_move() sees
 268          * a dnode hold for every dbuf.
 269          */
 270         rw_exit(&dn->dn_struct_rwlock);
 271 
 272         dnode_rele(dn, FTAG);
 273 
 274         VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
 275 
 276         *dbp = &db->db;
 277         return (0);
 278 }
 279 
 280 /*
 281  * returns ENOENT, EIO, or 0.
 282  *
 283  * This interface will allocate a blank spill dbuf when a spill blk
 284  * doesn't already exist on the dnode.
 285  *
 286  * if you only want to find an already existing spill db, then
 287  * dmu_spill_hold_existing() should be used.
 288  */
 289 int
 290 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
 291 {
 292         dmu_buf_impl_t *db = NULL;
 293         int err;
 294 
 295         if ((flags & DB_RF_HAVESTRUCT) == 0)
 296                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 297 
 298         db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
 299 
 300         if ((flags & DB_RF_HAVESTRUCT) == 0)
 301                 rw_exit(&dn->dn_struct_rwlock);
 302 
 303         ASSERT(db != NULL);
 304         err = dbuf_read(db, NULL, flags);
 305         if (err == 0)
 306                 *dbp = &db->db;
 307         else
 308                 dbuf_rele(db, tag);
 309         return (err);
 310 }
 311 
 312 int
 313 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
 314 {
 315         dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
 316         dnode_t *dn;
 317         int err;
 318 
 319         DB_DNODE_ENTER(db);
 320         dn = DB_DNODE(db);
 321 
 322         if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
 323                 err = EINVAL;
 324         } else {
 325                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 326 
 327                 if (!dn->dn_have_spill) {
 328                         err = ENOENT;
 329                 } else {
 330                         err = dmu_spill_hold_by_dnode(dn,
 331                             DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
 332                 }
 333 
 334                 rw_exit(&dn->dn_struct_rwlock);
 335         }
 336 
 337         DB_DNODE_EXIT(db);
 338         return (err);
 339 }
 340 
 341 int
 342 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
 343 {
 344         dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
 345         dnode_t *dn;
 346         int err;
 347 
 348         DB_DNODE_ENTER(db);
 349         dn = DB_DNODE(db);
 350         err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
 351         DB_DNODE_EXIT(db);
 352 
 353         return (err);
 354 }
 355 
 356 /*
 357  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
 358  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
 359  * and can induce severe lock contention when writing to several files
 360  * whose dnodes are in the same block.
 361  */
 362 static int
 363 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 364     int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
 365 {
 366         dsl_pool_t *dp = NULL;
 367         dmu_buf_t **dbp;
 368         uint64_t blkid, nblks, i;
 369         uint32_t dbuf_flags;
 370         int err;
 371         zio_t *zio;
 372         hrtime_t start;
 373 
 374         ASSERT(length <= DMU_MAX_ACCESS);
 375 
 376         dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
 377         if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
 378                 dbuf_flags |= DB_RF_NOPREFETCH;
 379 
 380         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 381         if (dn->dn_datablkshift) {
 382                 int blkshift = dn->dn_datablkshift;
 383                 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
 384                     P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
 385         } else {
 386                 if (offset + length > dn->dn_datablksz) {
 387                         zfs_panic_recover("zfs: accessing past end of object "
 388                             "%llx/%llx (size=%u access=%llu+%llu)",
 389                             (longlong_t)dn->dn_objset->
 390                             os_dsl_dataset->ds_object,
 391                             (longlong_t)dn->dn_object, dn->dn_datablksz,
 392                             (longlong_t)offset, (longlong_t)length);
 393                         rw_exit(&dn->dn_struct_rwlock);
 394                         return (EIO);
 395                 }
 396                 nblks = 1;
 397         }
 398         dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
 399 
 400         if (dn->dn_objset->os_dsl_dataset)
 401                 dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
 402         if (dp && dsl_pool_sync_context(dp))
 403                 start = gethrtime();
 404         zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
 405         blkid = dbuf_whichblock(dn, offset);
 406         for (i = 0; i < nblks; i++) {
 407                 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
 408                 if (db == NULL) {
 409                         rw_exit(&dn->dn_struct_rwlock);
 410                         dmu_buf_rele_array(dbp, nblks, tag);
 411                         zio_nowait(zio);
 412                         return (EIO);
 413                 }
 414                 /* initiate async i/o */
 415                 if (read) {
 416                         (void) dbuf_read(db, zio, dbuf_flags);
 417                 }
 418                 dbp[i] = &db->db;
 419         }
 420         rw_exit(&dn->dn_struct_rwlock);
 421 
 422         /* wait for async i/o */
 423         err = zio_wait(zio);
 424         /* track read overhead when we are in sync context */
 425         if (dp && dsl_pool_sync_context(dp))
 426                 dp->dp_read_overhead += gethrtime() - start;
 427         if (err) {
 428                 dmu_buf_rele_array(dbp, nblks, tag);
 429                 return (err);
 430         }
 431 
 432         /* wait for other io to complete */
 433         if (read) {
 434                 for (i = 0; i < nblks; i++) {
 435                         dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
 436                         mutex_enter(&db->db_mtx);
 437                         while (db->db_state == DB_READ ||
 438                             db->db_state == DB_FILL)
 439                                 cv_wait(&db->db_changed, &db->db_mtx);
 440                         if (db->db_state == DB_UNCACHED)
 441                                 err = EIO;
 442                         mutex_exit(&db->db_mtx);
 443                         if (err) {
 444                                 dmu_buf_rele_array(dbp, nblks, tag);
 445                                 return (err);
 446                         }
 447                 }
 448         }
 449 
 450         *numbufsp = nblks;
 451         *dbpp = dbp;
 452         return (0);
 453 }
 454 
 455 static int
 456 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
 457     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
 458 {
 459         dnode_t *dn;
 460         int err;
 461 
 462         err = dnode_hold(os, object, FTAG, &dn);
 463         if (err)
 464                 return (err);
 465 
 466         err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
 467             numbufsp, dbpp, DMU_READ_PREFETCH);
 468 
 469         dnode_rele(dn, FTAG);
 470 
 471         return (err);
 472 }
 473 
 474 int
 475 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
 476     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
 477 {
 478         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
 479         dnode_t *dn;
 480         int err;
 481 
 482         DB_DNODE_ENTER(db);
 483         dn = DB_DNODE(db);
 484         err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
 485             numbufsp, dbpp, DMU_READ_PREFETCH);
 486         DB_DNODE_EXIT(db);
 487 
 488         return (err);
 489 }
 490 
 491 void
 492 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
 493 {
 494         int i;
 495         dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
 496 
 497         if (numbufs == 0)
 498                 return;
 499 
 500         for (i = 0; i < numbufs; i++) {
 501                 if (dbp[i])
 502                         dbuf_rele(dbp[i], tag);
 503         }
 504 
 505         kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
 506 }
 507 
 508 void
 509 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
 510 {
 511         dnode_t *dn;
 512         uint64_t blkid;
 513         int nblks, i, err;
 514 
 515         if (zfs_prefetch_disable)
 516                 return;
 517 
 518         if (len == 0) {  /* they're interested in the bonus buffer */
 519                 dn = DMU_META_DNODE(os);
 520 
 521                 if (object == 0 || object >= DN_MAX_OBJECT)
 522                         return;
 523 
 524                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 525                 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
 526                 dbuf_prefetch(dn, blkid);
 527                 rw_exit(&dn->dn_struct_rwlock);
 528                 return;
 529         }
 530 
 531         /*
 532          * XXX - Note, if the dnode for the requested object is not
 533          * already cached, we will do a *synchronous* read in the
 534          * dnode_hold() call.  The same is true for any indirects.
 535          */
 536         err = dnode_hold(os, object, FTAG, &dn);
 537         if (err != 0)
 538                 return;
 539 
 540         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 541         if (dn->dn_datablkshift) {
 542                 int blkshift = dn->dn_datablkshift;
 543                 nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
 544                     P2ALIGN(offset, 1<<blkshift)) >> blkshift;
 545         } else {
 546                 nblks = (offset < dn->dn_datablksz);
 547         }
 548 
 549         if (nblks != 0) {
 550                 blkid = dbuf_whichblock(dn, offset);
 551                 for (i = 0; i < nblks; i++)
 552                         dbuf_prefetch(dn, blkid+i);
 553         }
 554 
 555         rw_exit(&dn->dn_struct_rwlock);
 556 
 557         dnode_rele(dn, FTAG);
 558 }
 559 
 560 /*
 561  * Get the next "chunk" of file data to free.  We traverse the file from
 562  * the end so that the file gets shorter over time (if we crashes in the
 563  * middle, this will leave us in a better state).  We find allocated file
 564  * data by simply searching the allocated level 1 indirects.
 565  */
 566 static int
 567 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
 568 {
 569         uint64_t len = *start - limit;
 570         uint64_t blkcnt = 0;
 571         uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
 572         uint64_t iblkrange =
 573             dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
 574 
 575         ASSERT(limit <= *start);
 576 
 577         if (len <= iblkrange * maxblks) {
 578                 *start = limit;
 579                 return (0);
 580         }
 581         ASSERT(ISP2(iblkrange));
 582 
 583         while (*start > limit && blkcnt < maxblks) {
 584                 int err;
 585 
 586                 /* find next allocated L1 indirect */
 587                 err = dnode_next_offset(dn,
 588                     DNODE_FIND_BACKWARDS, start, 2, 1, 0);
 589 
 590                 /* if there are no more, then we are done */
 591                 if (err == ESRCH) {
 592                         *start = limit;
 593                         return (0);
 594                 } else if (err) {
 595                         return (err);
 596                 }
 597                 blkcnt += 1;
 598 
 599                 /* reset offset to end of "next" block back */
 600                 *start = P2ALIGN(*start, iblkrange);
 601                 if (*start <= limit)
 602                         *start = limit;
 603                 else
 604                         *start -= 1;
 605         }
 606         return (0);
 607 }
 608 
 609 static int
 610 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
 611     uint64_t length, boolean_t free_dnode)
 612 {
 613         dmu_tx_t *tx;
 614         uint64_t object_size, start, end, len;
 615         boolean_t trunc = (length == DMU_OBJECT_END);
 616         int align, err;
 617 
 618         align = 1 << dn->dn_datablkshift;
 619         ASSERT(align > 0);
 620         object_size = align == 1 ? dn->dn_datablksz :
 621             (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
 622 
 623         end = offset + length;
 624         if (trunc || end > object_size)
 625                 end = object_size;
 626         if (end <= offset)
 627                 return (0);
 628         length = end - offset;
 629 
 630         while (length) {
 631                 start = end;
 632                 /* assert(offset <= start) */
 633                 err = get_next_chunk(dn, &start, offset);
 634                 if (err)
 635                         return (err);
 636                 len = trunc ? DMU_OBJECT_END : end - start;
 637 
 638                 tx = dmu_tx_create(os);
 639                 dmu_tx_hold_free(tx, dn->dn_object, start, len);
 640                 err = dmu_tx_assign(tx, TXG_WAIT);
 641                 if (err) {
 642                         dmu_tx_abort(tx);
 643                         return (err);
 644                 }
 645 
 646                 dnode_free_range(dn, start, trunc ? -1 : len, tx);
 647 
 648                 if (start == 0 && free_dnode) {
 649                         ASSERT(trunc);
 650                         dnode_free(dn, tx);
 651                 }
 652 
 653                 length -= end - start;
 654 
 655                 dmu_tx_commit(tx);
 656                 end = start;
 657         }
 658         return (0);
 659 }
 660 
 661 int
 662 dmu_free_long_range(objset_t *os, uint64_t object,
 663     uint64_t offset, uint64_t length)
 664 {
 665         dnode_t *dn;
 666         int err;
 667 
 668         err = dnode_hold(os, object, FTAG, &dn);
 669         if (err != 0)
 670                 return (err);
 671         err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
 672         dnode_rele(dn, FTAG);
 673         return (err);
 674 }
 675 
 676 int
 677 dmu_free_object(objset_t *os, uint64_t object)
 678 {
 679         dnode_t *dn;
 680         dmu_tx_t *tx;
 681         int err;
 682 
 683         err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
 684             FTAG, &dn);
 685         if (err != 0)
 686                 return (err);
 687         if (dn->dn_nlevels == 1) {
 688                 tx = dmu_tx_create(os);
 689                 dmu_tx_hold_bonus(tx, object);
 690                 dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
 691                 err = dmu_tx_assign(tx, TXG_WAIT);
 692                 if (err == 0) {
 693                         dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
 694                         dnode_free(dn, tx);
 695                         dmu_tx_commit(tx);
 696                 } else {
 697                         dmu_tx_abort(tx);
 698                 }
 699         } else {
 700                 err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
 701         }
 702         dnode_rele(dn, FTAG);
 703         return (err);
 704 }
 705 
 706 int
 707 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
 708     uint64_t size, dmu_tx_t *tx)
 709 {
 710         dnode_t *dn;
 711         int err = dnode_hold(os, object, FTAG, &dn);
 712         if (err)
 713                 return (err);
 714         ASSERT(offset < UINT64_MAX);
 715         ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
 716         dnode_free_range(dn, offset, size, tx);
 717         dnode_rele(dn, FTAG);
 718         return (0);
 719 }
 720 
 721 int
 722 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
 723     void *buf, uint32_t flags)
 724 {
 725         dnode_t *dn;
 726         dmu_buf_t **dbp;
 727         int numbufs, err;
 728 
 729         err = dnode_hold(os, object, FTAG, &dn);
 730         if (err)
 731                 return (err);
 732 
 733         /*
 734          * Deal with odd block sizes, where there can't be data past the first
 735          * block.  If we ever do the tail block optimization, we will need to
 736          * handle that here as well.
 737          */
 738         if (dn->dn_maxblkid == 0) {
 739                 int newsz = offset > dn->dn_datablksz ? 0 :
 740                     MIN(size, dn->dn_datablksz - offset);
 741                 bzero((char *)buf + newsz, size - newsz);
 742                 size = newsz;
 743         }
 744 
 745         while (size > 0) {
 746                 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
 747                 int i;
 748 
 749                 /*
 750                  * NB: we could do this block-at-a-time, but it's nice
 751                  * to be reading in parallel.
 752                  */
 753                 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
 754                     TRUE, FTAG, &numbufs, &dbp, flags);
 755                 if (err)
 756                         break;
 757 
 758                 for (i = 0; i < numbufs; i++) {
 759                         int tocpy;
 760                         int bufoff;
 761                         dmu_buf_t *db = dbp[i];
 762 
 763                         ASSERT(size > 0);
 764 
 765                         bufoff = offset - db->db_offset;
 766                         tocpy = (int)MIN(db->db_size - bufoff, size);
 767 
 768                         bcopy((char *)db->db_data + bufoff, buf, tocpy);
 769 
 770                         offset += tocpy;
 771                         size -= tocpy;
 772                         buf = (char *)buf + tocpy;
 773                 }
 774                 dmu_buf_rele_array(dbp, numbufs, FTAG);
 775         }
 776         dnode_rele(dn, FTAG);
 777         return (err);
 778 }
 779 
 780 void
 781 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
 782     const void *buf, dmu_tx_t *tx)
 783 {
 784         dmu_buf_t **dbp;
 785         int numbufs, i;
 786 
 787         if (size == 0)
 788                 return;
 789 
 790         VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
 791             FALSE, FTAG, &numbufs, &dbp));
 792 
 793         for (i = 0; i < numbufs; i++) {
 794                 int tocpy;
 795                 int bufoff;
 796                 dmu_buf_t *db = dbp[i];
 797 
 798                 ASSERT(size > 0);
 799 
 800                 bufoff = offset - db->db_offset;
 801                 tocpy = (int)MIN(db->db_size - bufoff, size);
 802 
 803                 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
 804 
 805                 if (tocpy == db->db_size)
 806                         dmu_buf_will_fill(db, tx);
 807                 else
 808                         dmu_buf_will_dirty(db, tx);
 809 
 810                 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
 811 
 812                 if (tocpy == db->db_size)
 813                         dmu_buf_fill_done(db, tx);
 814 
 815                 offset += tocpy;
 816                 size -= tocpy;
 817                 buf = (char *)buf + tocpy;
 818         }
 819         dmu_buf_rele_array(dbp, numbufs, FTAG);
 820 }
 821 
 822 void
 823 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
 824     dmu_tx_t *tx)
 825 {
 826         dmu_buf_t **dbp;
 827         int numbufs, i;
 828 
 829         if (size == 0)
 830                 return;
 831 
 832         VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
 833             FALSE, FTAG, &numbufs, &dbp));
 834 
 835         for (i = 0; i < numbufs; i++) {
 836                 dmu_buf_t *db = dbp[i];
 837 
 838                 dmu_buf_will_not_fill(db, tx);
 839         }
 840         dmu_buf_rele_array(dbp, numbufs, FTAG);
 841 }
 842 
 843 /*
 844  * DMU support for xuio
 845  */
 846 kstat_t *xuio_ksp = NULL;
 847 
 848 int
 849 dmu_xuio_init(xuio_t *xuio, int nblk)
 850 {
 851         dmu_xuio_t *priv;
 852         uio_t *uio = &xuio->xu_uio;
 853 
 854         uio->uio_iovcnt = nblk;
 855         uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
 856 
 857         priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
 858         priv->cnt = nblk;
 859         priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
 860         priv->iovp = uio->uio_iov;
 861         XUIO_XUZC_PRIV(xuio) = priv;
 862 
 863         if (XUIO_XUZC_RW(xuio) == UIO_READ)
 864                 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
 865         else
 866                 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
 867 
 868         return (0);
 869 }
 870 
 871 void
 872 dmu_xuio_fini(xuio_t *xuio)
 873 {
 874         dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
 875         int nblk = priv->cnt;
 876 
 877         kmem_free(priv->iovp, nblk * sizeof (iovec_t));
 878         kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
 879         kmem_free(priv, sizeof (dmu_xuio_t));
 880 
 881         if (XUIO_XUZC_RW(xuio) == UIO_READ)
 882                 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
 883         else
 884                 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
 885 }
 886 
 887 /*
 888  * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
 889  * and increase priv->next by 1.
 890  */
 891 int
 892 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
 893 {
 894         struct iovec *iov;
 895         uio_t *uio = &xuio->xu_uio;
 896         dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
 897         int i = priv->next++;
 898 
 899         ASSERT(i < priv->cnt);
 900         ASSERT(off + n <= arc_buf_size(abuf));
 901         iov = uio->uio_iov + i;
 902         iov->iov_base = (char *)abuf->b_data + off;
 903         iov->iov_len = n;
 904         priv->bufs[i] = abuf;
 905         return (0);
 906 }
 907 
 908 int
 909 dmu_xuio_cnt(xuio_t *xuio)
 910 {
 911         dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
 912         return (priv->cnt);
 913 }
 914 
 915 arc_buf_t *
 916 dmu_xuio_arcbuf(xuio_t *xuio, int i)
 917 {
 918         dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
 919 
 920         ASSERT(i < priv->cnt);
 921         return (priv->bufs[i]);
 922 }
 923 
 924 void
 925 dmu_xuio_clear(xuio_t *xuio, int i)
 926 {
 927         dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
 928 
 929         ASSERT(i < priv->cnt);
 930         priv->bufs[i] = NULL;
 931 }
 932 
 933 static void
 934 xuio_stat_init(void)
 935 {
 936         xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
 937             KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
 938             KSTAT_FLAG_VIRTUAL);
 939         if (xuio_ksp != NULL) {
 940                 xuio_ksp->ks_data = &xuio_stats;
 941                 kstat_install(xuio_ksp);
 942         }
 943 }
 944 
 945 static void
 946 xuio_stat_fini(void)
 947 {
 948         if (xuio_ksp != NULL) {
 949                 kstat_delete(xuio_ksp);
 950                 xuio_ksp = NULL;
 951         }
 952 }
 953 
 954 void
 955 xuio_stat_wbuf_copied()
 956 {
 957         XUIOSTAT_BUMP(xuiostat_wbuf_copied);
 958 }
 959 
 960 void
 961 xuio_stat_wbuf_nocopy()
 962 {
 963         XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
 964 }
 965 
 966 #ifdef _KERNEL
 967 int
 968 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
 969 {
 970         dmu_buf_t **dbp;
 971         int numbufs, i, err;
 972         xuio_t *xuio = NULL;
 973 
 974         /*
 975          * NB: we could do this block-at-a-time, but it's nice
 976          * to be reading in parallel.
 977          */
 978         err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
 979             &numbufs, &dbp);
 980         if (err)
 981                 return (err);
 982 
 983         if (uio->uio_extflg == UIO_XUIO)
 984                 xuio = (xuio_t *)uio;
 985 
 986         for (i = 0; i < numbufs; i++) {
 987                 int tocpy;
 988                 int bufoff;
 989                 dmu_buf_t *db = dbp[i];
 990 
 991                 ASSERT(size > 0);
 992 
 993                 bufoff = uio->uio_loffset - db->db_offset;
 994                 tocpy = (int)MIN(db->db_size - bufoff, size);
 995 
 996                 if (xuio) {
 997                         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
 998                         arc_buf_t *dbuf_abuf = dbi->db_buf;
 999                         arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1000                         err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1001                         if (!err) {
1002                                 uio->uio_resid -= tocpy;
1003                                 uio->uio_loffset += tocpy;
1004                         }
1005 
1006                         if (abuf == dbuf_abuf)
1007                                 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1008                         else
1009                                 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1010                 } else {
1011                         err = uiomove((char *)db->db_data + bufoff, tocpy,
1012                             UIO_READ, uio);
1013                 }
1014                 if (err)
1015                         break;
1016 
1017                 size -= tocpy;
1018         }
1019         dmu_buf_rele_array(dbp, numbufs, FTAG);
1020 
1021         return (err);
1022 }
1023 
1024 static int
1025 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1026 {
1027         dmu_buf_t **dbp;
1028         int numbufs;
1029         int err = 0;
1030         int i;
1031 
1032         err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1033             FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1034         if (err)
1035                 return (err);
1036 
1037         for (i = 0; i < numbufs; i++) {
1038                 int tocpy;
1039                 int bufoff;
1040                 dmu_buf_t *db = dbp[i];
1041 
1042                 ASSERT(size > 0);
1043 
1044                 bufoff = uio->uio_loffset - db->db_offset;
1045                 tocpy = (int)MIN(db->db_size - bufoff, size);
1046 
1047                 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1048 
1049                 if (tocpy == db->db_size)
1050                         dmu_buf_will_fill(db, tx);
1051                 else
1052                         dmu_buf_will_dirty(db, tx);
1053 
1054                 /*
1055                  * XXX uiomove could block forever (eg. nfs-backed
1056                  * pages).  There needs to be a uiolockdown() function
1057                  * to lock the pages in memory, so that uiomove won't
1058                  * block.
1059                  */
1060                 err = uiomove((char *)db->db_data + bufoff, tocpy,
1061                     UIO_WRITE, uio);
1062 
1063                 if (tocpy == db->db_size)
1064                         dmu_buf_fill_done(db, tx);
1065 
1066                 if (err)
1067                         break;
1068 
1069                 size -= tocpy;
1070         }
1071 
1072         dmu_buf_rele_array(dbp, numbufs, FTAG);
1073         return (err);
1074 }
1075 
1076 int
1077 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1078     dmu_tx_t *tx)
1079 {
1080         dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1081         dnode_t *dn;
1082         int err;
1083 
1084         if (size == 0)
1085                 return (0);
1086 
1087         DB_DNODE_ENTER(db);
1088         dn = DB_DNODE(db);
1089         err = dmu_write_uio_dnode(dn, uio, size, tx);
1090         DB_DNODE_EXIT(db);
1091 
1092         return (err);
1093 }
1094 
1095 int
1096 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1097     dmu_tx_t *tx)
1098 {
1099         dnode_t *dn;
1100         int err;
1101 
1102         if (size == 0)
1103                 return (0);
1104 
1105         err = dnode_hold(os, object, FTAG, &dn);
1106         if (err)
1107                 return (err);
1108 
1109         err = dmu_write_uio_dnode(dn, uio, size, tx);
1110 
1111         dnode_rele(dn, FTAG);
1112 
1113         return (err);
1114 }
1115 
1116 int
1117 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1118     page_t *pp, dmu_tx_t *tx)
1119 {
1120         dmu_buf_t **dbp;
1121         int numbufs, i;
1122         int err;
1123 
1124         if (size == 0)
1125                 return (0);
1126 
1127         err = dmu_buf_hold_array(os, object, offset, size,
1128             FALSE, FTAG, &numbufs, &dbp);
1129         if (err)
1130                 return (err);
1131 
1132         for (i = 0; i < numbufs; i++) {
1133                 int tocpy, copied, thiscpy;
1134                 int bufoff;
1135                 dmu_buf_t *db = dbp[i];
1136                 caddr_t va;
1137 
1138                 ASSERT(size > 0);
1139                 ASSERT3U(db->db_size, >=, PAGESIZE);
1140 
1141                 bufoff = offset - db->db_offset;
1142                 tocpy = (int)MIN(db->db_size - bufoff, size);
1143 
1144                 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1145 
1146                 if (tocpy == db->db_size)
1147                         dmu_buf_will_fill(db, tx);
1148                 else
1149                         dmu_buf_will_dirty(db, tx);
1150 
1151                 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1152                         ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1153                         thiscpy = MIN(PAGESIZE, tocpy - copied);
1154                         va = zfs_map_page(pp, S_READ);
1155                         bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1156                         zfs_unmap_page(pp, va);
1157                         pp = pp->p_next;
1158                         bufoff += PAGESIZE;
1159                 }
1160 
1161                 if (tocpy == db->db_size)
1162                         dmu_buf_fill_done(db, tx);
1163 
1164                 offset += tocpy;
1165                 size -= tocpy;
1166         }
1167         dmu_buf_rele_array(dbp, numbufs, FTAG);
1168         return (err);
1169 }
1170 #endif
1171 
1172 /*
1173  * Allocate a loaned anonymous arc buffer.
1174  */
1175 arc_buf_t *
1176 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1177 {
1178         dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1179         spa_t *spa;
1180 
1181         DB_GET_SPA(&spa, db);
1182         return (arc_loan_buf(spa, size));
1183 }
1184 
1185 /*
1186  * Free a loaned arc buffer.
1187  */
1188 void
1189 dmu_return_arcbuf(arc_buf_t *buf)
1190 {
1191         arc_return_buf(buf, FTAG);
1192         VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1193 }
1194 
1195 /*
1196  * When possible directly assign passed loaned arc buffer to a dbuf.
1197  * If this is not possible copy the contents of passed arc buf via
1198  * dmu_write().
1199  */
1200 void
1201 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1202     dmu_tx_t *tx)
1203 {
1204         dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1205         dnode_t *dn;
1206         dmu_buf_impl_t *db;
1207         uint32_t blksz = (uint32_t)arc_buf_size(buf);
1208         uint64_t blkid;
1209 
1210         DB_DNODE_ENTER(dbuf);
1211         dn = DB_DNODE(dbuf);
1212         rw_enter(&dn->dn_struct_rwlock, RW_READER);
1213         blkid = dbuf_whichblock(dn, offset);
1214         VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1215         rw_exit(&dn->dn_struct_rwlock);
1216         DB_DNODE_EXIT(dbuf);
1217 
1218         if (offset == db->db.db_offset && blksz == db->db.db_size) {
1219                 dbuf_assign_arcbuf(db, buf, tx);
1220                 dbuf_rele(db, FTAG);
1221         } else {
1222                 objset_t *os;
1223                 uint64_t object;
1224 
1225                 DB_DNODE_ENTER(dbuf);
1226                 dn = DB_DNODE(dbuf);
1227                 os = dn->dn_objset;
1228                 object = dn->dn_object;
1229                 DB_DNODE_EXIT(dbuf);
1230 
1231                 dbuf_rele(db, FTAG);
1232                 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1233                 dmu_return_arcbuf(buf);
1234                 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1235         }
1236 }
1237 
1238 typedef struct {
1239         dbuf_dirty_record_t     *dsa_dr;
1240         dmu_sync_cb_t           *dsa_done;
1241         zgd_t                   *dsa_zgd;
1242         dmu_tx_t                *dsa_tx;
1243 } dmu_sync_arg_t;
1244 
1245 /* ARGSUSED */
1246 static void
1247 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1248 {
1249         dmu_sync_arg_t *dsa = varg;
1250         dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1251         blkptr_t *bp = zio->io_bp;
1252 
1253         if (zio->io_error == 0) {
1254                 if (BP_IS_HOLE(bp)) {
1255                         /*
1256                          * A block of zeros may compress to a hole, but the
1257                          * block size still needs to be known for replay.
1258                          */
1259                         BP_SET_LSIZE(bp, db->db_size);
1260                 } else {
1261                         ASSERT(BP_GET_LEVEL(bp) == 0);
1262                         bp->blk_fill = 1;
1263                 }
1264         }
1265 }
1266 
1267 static void
1268 dmu_sync_late_arrival_ready(zio_t *zio)
1269 {
1270         dmu_sync_ready(zio, NULL, zio->io_private);
1271 }
1272 
1273 /* ARGSUSED */
1274 static void
1275 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1276 {
1277         dmu_sync_arg_t *dsa = varg;
1278         dbuf_dirty_record_t *dr = dsa->dsa_dr;
1279         dmu_buf_impl_t *db = dr->dr_dbuf;
1280 
1281         mutex_enter(&db->db_mtx);
1282         ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1283         if (zio->io_error == 0) {
1284                 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1285                 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1286                 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1287                 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1288                         BP_ZERO(&dr->dt.dl.dr_overridden_by);
1289         } else {
1290                 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1291         }
1292         cv_broadcast(&db->db_changed);
1293         mutex_exit(&db->db_mtx);
1294 
1295         dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1296 
1297         kmem_free(dsa, sizeof (*dsa));
1298 }
1299 
1300 static void
1301 dmu_sync_late_arrival_done(zio_t *zio)
1302 {
1303         blkptr_t *bp = zio->io_bp;
1304         dmu_sync_arg_t *dsa = zio->io_private;
1305 
1306         if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1307                 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1308                 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1309                 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1310         }
1311 
1312         dmu_tx_commit(dsa->dsa_tx);
1313 
1314         dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1315 
1316         kmem_free(dsa, sizeof (*dsa));
1317 }
1318 
1319 static int
1320 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1321     zio_prop_t *zp, zbookmark_t *zb)
1322 {
1323         dmu_sync_arg_t *dsa;
1324         dmu_tx_t *tx;
1325 
1326         tx = dmu_tx_create(os);
1327         dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1328         if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1329                 dmu_tx_abort(tx);
1330                 return (EIO);   /* Make zl_get_data do txg_waited_synced() */
1331         }
1332 
1333         dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1334         dsa->dsa_dr = NULL;
1335         dsa->dsa_done = done;
1336         dsa->dsa_zgd = zgd;
1337         dsa->dsa_tx = tx;
1338 
1339         zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1340             zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1341             dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1342             ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1343 
1344         return (0);
1345 }
1346 
1347 /*
1348  * Intent log support: sync the block associated with db to disk.
1349  * N.B. and XXX: the caller is responsible for making sure that the
1350  * data isn't changing while dmu_sync() is writing it.
1351  *
1352  * Return values:
1353  *
1354  *      EEXIST: this txg has already been synced, so there's nothing to to.
1355  *              The caller should not log the write.
1356  *
1357  *      ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1358  *              The caller should not log the write.
1359  *
1360  *      EALREADY: this block is already in the process of being synced.
1361  *              The caller should track its progress (somehow).
1362  *
1363  *      EIO: could not do the I/O.
1364  *              The caller should do a txg_wait_synced().
1365  *
1366  *      0: the I/O has been initiated.
1367  *              The caller should log this blkptr in the done callback.
1368  *              It is possible that the I/O will fail, in which case
1369  *              the error will be reported to the done callback and
1370  *              propagated to pio from zio_done().
1371  */
1372 int
1373 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1374 {
1375         blkptr_t *bp = zgd->zgd_bp;
1376         dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1377         objset_t *os = db->db_objset;
1378         dsl_dataset_t *ds = os->os_dsl_dataset;
1379         dbuf_dirty_record_t *dr;
1380         dmu_sync_arg_t *dsa;
1381         zbookmark_t zb;
1382         zio_prop_t zp;
1383         dnode_t *dn;
1384 
1385         ASSERT(pio != NULL);
1386         ASSERT(BP_IS_HOLE(bp));
1387         ASSERT(txg != 0);
1388 
1389         SET_BOOKMARK(&zb, ds->ds_object,
1390             db->db.db_object, db->db_level, db->db_blkid);
1391 
1392         DB_DNODE_ENTER(db);
1393         dn = DB_DNODE(db);
1394         dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1395         DB_DNODE_EXIT(db);
1396 
1397         /*
1398          * If we're frozen (running ziltest), we always need to generate a bp.
1399          */
1400         if (txg > spa_freeze_txg(os->os_spa))
1401                 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1402 
1403         /*
1404          * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1405          * and us.  If we determine that this txg is not yet syncing,
1406          * but it begins to sync a moment later, that's OK because the
1407          * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1408          */
1409         mutex_enter(&db->db_mtx);
1410 
1411         if (txg <= spa_last_synced_txg(os->os_spa)) {
1412                 /*
1413                  * This txg has already synced.  There's nothing to do.
1414                  */
1415                 mutex_exit(&db->db_mtx);
1416                 return (EEXIST);
1417         }
1418 
1419         if (txg <= spa_syncing_txg(os->os_spa)) {
1420                 /*
1421                  * This txg is currently syncing, so we can't mess with
1422                  * the dirty record anymore; just write a new log block.
1423                  */
1424                 mutex_exit(&db->db_mtx);
1425                 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1426         }
1427 
1428         dr = db->db_last_dirty;
1429         while (dr && dr->dr_txg != txg)
1430                 dr = dr->dr_next;
1431 
1432         if (dr == NULL) {
1433                 /*
1434                  * There's no dr for this dbuf, so it must have been freed.
1435                  * There's no need to log writes to freed blocks, so we're done.
1436                  */
1437                 mutex_exit(&db->db_mtx);
1438                 return (ENOENT);
1439         }
1440 
1441         ASSERT(dr->dr_txg == txg);
1442         if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1443             dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1444                 /*
1445                  * We have already issued a sync write for this buffer,
1446                  * or this buffer has already been synced.  It could not
1447                  * have been dirtied since, or we would have cleared the state.
1448                  */
1449                 mutex_exit(&db->db_mtx);
1450                 return (EALREADY);
1451         }
1452 
1453         ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1454         dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1455         mutex_exit(&db->db_mtx);
1456 
1457         dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1458         dsa->dsa_dr = dr;
1459         dsa->dsa_done = done;
1460         dsa->dsa_zgd = zgd;
1461         dsa->dsa_tx = NULL;
1462 
1463         zio_nowait(arc_write(pio, os->os_spa, txg,
1464             bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1465             dmu_sync_ready, dmu_sync_done, dsa,
1466             ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1467 
1468         return (0);
1469 }
1470 
1471 int
1472 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1473         dmu_tx_t *tx)
1474 {
1475         dnode_t *dn;
1476         int err;
1477 
1478         err = dnode_hold(os, object, FTAG, &dn);
1479         if (err)
1480                 return (err);
1481         err = dnode_set_blksz(dn, size, ibs, tx);
1482         dnode_rele(dn, FTAG);
1483         return (err);
1484 }
1485 
1486 void
1487 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1488         dmu_tx_t *tx)
1489 {
1490         dnode_t *dn;
1491 
1492         /* XXX assumes dnode_hold will not get an i/o error */
1493         (void) dnode_hold(os, object, FTAG, &dn);
1494         ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1495         dn->dn_checksum = checksum;
1496         dnode_setdirty(dn, tx);
1497         dnode_rele(dn, FTAG);
1498 }
1499 
1500 void
1501 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1502         dmu_tx_t *tx)
1503 {
1504         dnode_t *dn;
1505 
1506         /* XXX assumes dnode_hold will not get an i/o error */
1507         (void) dnode_hold(os, object, FTAG, &dn);
1508         ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1509         dn->dn_compress = compress;
1510         dnode_setdirty(dn, tx);
1511         dnode_rele(dn, FTAG);
1512 }
1513 
1514 int zfs_mdcomp_disable = 0;
1515 
1516 void
1517 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1518 {
1519         dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1520         boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1521             (wp & WP_SPILL));
1522         enum zio_checksum checksum = os->os_checksum;
1523         enum zio_compress compress = os->os_compress;
1524         enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1525         boolean_t dedup;
1526         boolean_t dedup_verify = os->os_dedup_verify;
1527         int copies = os->os_copies;
1528 
1529         /*
1530          * Determine checksum setting.
1531          */
1532         if (ismd) {
1533                 /*
1534                  * Metadata always gets checksummed.  If the data
1535                  * checksum is multi-bit correctable, and it's not a
1536                  * ZBT-style checksum, then it's suitable for metadata
1537                  * as well.  Otherwise, the metadata checksum defaults
1538                  * to fletcher4.
1539                  */
1540                 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1541                     zio_checksum_table[checksum].ci_eck)
1542                         checksum = ZIO_CHECKSUM_FLETCHER_4;
1543         } else {
1544                 checksum = zio_checksum_select(dn->dn_checksum, checksum);
1545         }
1546 
1547         /*
1548          * Determine compression setting.
1549          */
1550         if (ismd) {
1551                 /*
1552                  * XXX -- we should design a compression algorithm
1553                  * that specializes in arrays of bps.
1554                  */
1555                 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1556                     ZIO_COMPRESS_LZJB;
1557         } else {
1558                 compress = zio_compress_select(dn->dn_compress, compress);
1559         }
1560 
1561         /*
1562          * Determine dedup setting.  If we are in dmu_sync(), we won't
1563          * actually dedup now because that's all done in syncing context;
1564          * but we do want to use the dedup checkum.  If the checksum is not
1565          * strong enough to ensure unique signatures, force dedup_verify.
1566          */
1567         dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1568         if (dedup) {
1569                 checksum = dedup_checksum;
1570                 if (!zio_checksum_table[checksum].ci_dedup)
1571                         dedup_verify = 1;
1572         }
1573 
1574         if (wp & WP_DMU_SYNC)
1575                 dedup = 0;
1576 
1577         if (wp & WP_NOFILL) {
1578                 ASSERT(!ismd && level == 0);
1579                 checksum = ZIO_CHECKSUM_OFF;
1580                 compress = ZIO_COMPRESS_OFF;
1581                 dedup = B_FALSE;
1582         }
1583 
1584         zp->zp_checksum = checksum;
1585         zp->zp_compress = compress;
1586         zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1587         zp->zp_level = level;
1588         zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1589         zp->zp_dedup = dedup;
1590         zp->zp_dedup_verify = dedup && dedup_verify;
1591 }
1592 
1593 int
1594 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1595 {
1596         dnode_t *dn;
1597         int i, err;
1598 
1599         err = dnode_hold(os, object, FTAG, &dn);
1600         if (err)
1601                 return (err);
1602         /*
1603          * Sync any current changes before
1604          * we go trundling through the block pointers.
1605          */
1606         for (i = 0; i < TXG_SIZE; i++) {
1607                 if (list_link_active(&dn->dn_dirty_link[i]))
1608                         break;
1609         }
1610         if (i != TXG_SIZE) {
1611                 dnode_rele(dn, FTAG);
1612                 txg_wait_synced(dmu_objset_pool(os), 0);
1613                 err = dnode_hold(os, object, FTAG, &dn);
1614                 if (err)
1615                         return (err);
1616         }
1617 
1618         err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1619         dnode_rele(dn, FTAG);
1620 
1621         return (err);
1622 }
1623 
1624 void
1625 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1626 {
1627         dnode_phys_t *dnp;
1628 
1629         rw_enter(&dn->dn_struct_rwlock, RW_READER);
1630         mutex_enter(&dn->dn_mtx);
1631 
1632         dnp = dn->dn_phys;
1633 
1634         doi->doi_data_block_size = dn->dn_datablksz;
1635         doi->doi_metadata_block_size = dn->dn_indblkshift ?
1636             1ULL << dn->dn_indblkshift : 0;
1637         doi->doi_type = dn->dn_type;
1638         doi->doi_bonus_type = dn->dn_bonustype;
1639         doi->doi_bonus_size = dn->dn_bonuslen;
1640         doi->doi_indirection = dn->dn_nlevels;
1641         doi->doi_checksum = dn->dn_checksum;
1642         doi->doi_compress = dn->dn_compress;
1643         doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1644         doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1645         doi->doi_fill_count = 0;
1646         for (int i = 0; i < dnp->dn_nblkptr; i++)
1647                 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1648 
1649         mutex_exit(&dn->dn_mtx);
1650         rw_exit(&dn->dn_struct_rwlock);
1651 }
1652 
1653 /*
1654  * Get information on a DMU object.
1655  * If doi is NULL, just indicates whether the object exists.
1656  */
1657 int
1658 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1659 {
1660         dnode_t *dn;
1661         int err = dnode_hold(os, object, FTAG, &dn);
1662 
1663         if (err)
1664                 return (err);
1665 
1666         if (doi != NULL)
1667                 dmu_object_info_from_dnode(dn, doi);
1668 
1669         dnode_rele(dn, FTAG);
1670         return (0);
1671 }
1672 
1673 /*
1674  * As above, but faster; can be used when you have a held dbuf in hand.
1675  */
1676 void
1677 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1678 {
1679         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1680 
1681         DB_DNODE_ENTER(db);
1682         dmu_object_info_from_dnode(DB_DNODE(db), doi);
1683         DB_DNODE_EXIT(db);
1684 }
1685 
1686 /*
1687  * Faster still when you only care about the size.
1688  * This is specifically optimized for zfs_getattr().
1689  */
1690 void
1691 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1692     u_longlong_t *nblk512)
1693 {
1694         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1695         dnode_t *dn;
1696 
1697         DB_DNODE_ENTER(db);
1698         dn = DB_DNODE(db);
1699 
1700         *blksize = dn->dn_datablksz;
1701         /* add 1 for dnode space */
1702         *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1703             SPA_MINBLOCKSHIFT) + 1;
1704         DB_DNODE_EXIT(db);
1705 }
1706 
1707 void
1708 byteswap_uint64_array(void *vbuf, size_t size)
1709 {
1710         uint64_t *buf = vbuf;
1711         size_t count = size >> 3;
1712         int i;
1713 
1714         ASSERT((size & 7) == 0);
1715 
1716         for (i = 0; i < count; i++)
1717                 buf[i] = BSWAP_64(buf[i]);
1718 }
1719 
1720 void
1721 byteswap_uint32_array(void *vbuf, size_t size)
1722 {
1723         uint32_t *buf = vbuf;
1724         size_t count = size >> 2;
1725         int i;
1726 
1727         ASSERT((size & 3) == 0);
1728 
1729         for (i = 0; i < count; i++)
1730                 buf[i] = BSWAP_32(buf[i]);
1731 }
1732 
1733 void
1734 byteswap_uint16_array(void *vbuf, size_t size)
1735 {
1736         uint16_t *buf = vbuf;
1737         size_t count = size >> 1;
1738         int i;
1739 
1740         ASSERT((size & 1) == 0);
1741 
1742         for (i = 0; i < count; i++)
1743                 buf[i] = BSWAP_16(buf[i]);
1744 }
1745 
1746 /* ARGSUSED */
1747 void
1748 byteswap_uint8_array(void *vbuf, size_t size)
1749 {
1750 }
1751 
1752 void
1753 dmu_init(void)
1754 {
1755         zfs_dbgmsg_init();
1756         sa_cache_init();
1757         xuio_stat_init();
1758         dmu_objset_init();
1759         dnode_init();
1760         dbuf_init();
1761         zfetch_init();
1762         arc_init();
1763         l2arc_init();
1764 }
1765 
1766 void
1767 dmu_fini(void)
1768 {
1769         l2arc_fini();
1770         arc_fini();
1771         zfetch_fini();
1772         dbuf_fini();
1773         dnode_fini();
1774         dmu_objset_fini();
1775         xuio_stat_fini();
1776         sa_cache_fini();
1777         zfs_dbgmsg_fini();
1778 }