1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5ea8dc4b6Seschrock * Common Development and Distribution License (the "License"). 6ea8dc4b6Seschrock * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 2294d1a210STim Haley * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23fa9e4066Sahrens */ 24fa9e4066Sahrens 25fa9e4066Sahrens #include <sys/dmu.h> 26fa9e4066Sahrens #include <sys/dmu_impl.h> 27fa9e4066Sahrens #include <sys/dmu_tx.h> 28fa9e4066Sahrens #include <sys/dbuf.h> 29fa9e4066Sahrens #include <sys/dnode.h> 30fa9e4066Sahrens #include <sys/zfs_context.h> 31fa9e4066Sahrens #include <sys/dmu_objset.h> 32fa9e4066Sahrens #include <sys/dmu_traverse.h> 33fa9e4066Sahrens #include <sys/dsl_dataset.h> 34fa9e4066Sahrens #include <sys/dsl_dir.h> 35fa9e4066Sahrens #include <sys/dsl_pool.h> 361d452cf5Sahrens #include <sys/dsl_synctask.h> 37a2eea2e1Sahrens #include <sys/dsl_prop.h> 38fa9e4066Sahrens #include <sys/dmu_zfetch.h> 39fa9e4066Sahrens #include <sys/zfs_ioctl.h> 40fa9e4066Sahrens #include <sys/zap.h> 41ea8dc4b6Seschrock #include <sys/zio_checksum.h> 420a586ceaSMark Shellenbaum #include <sys/sa.h> 4344eda4d7Smaybee #ifdef _KERNEL 4444eda4d7Smaybee #include <sys/vmsystm.h> 450fab61baSJonathan W Adams #include <sys/zfs_znode.h> 4644eda4d7Smaybee #endif 47fa9e4066Sahrens 48fa9e4066Sahrens const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 49fa9e4066Sahrens { byteswap_uint8_array, TRUE, "unallocated" }, 50fa9e4066Sahrens { zap_byteswap, TRUE, "object directory" }, 51fa9e4066Sahrens { byteswap_uint64_array, TRUE, "object array" }, 52fa9e4066Sahrens { byteswap_uint8_array, TRUE, "packed nvlist" }, 53fa9e4066Sahrens { byteswap_uint64_array, TRUE, "packed nvlist size" }, 54fa9e4066Sahrens { byteswap_uint64_array, TRUE, "bplist" }, 55fa9e4066Sahrens { byteswap_uint64_array, TRUE, "bplist header" }, 56fa9e4066Sahrens { byteswap_uint64_array, TRUE, "SPA space map header" }, 57fa9e4066Sahrens { byteswap_uint64_array, TRUE, "SPA space map" }, 58fa9e4066Sahrens { byteswap_uint64_array, TRUE, "ZIL intent log" }, 59fa9e4066Sahrens { dnode_buf_byteswap, TRUE, "DMU dnode" }, 60fa9e4066Sahrens { dmu_objset_byteswap, TRUE, "DMU objset" }, 61fa9e4066Sahrens { byteswap_uint64_array, TRUE, "DSL directory" }, 62fa9e4066Sahrens { zap_byteswap, TRUE, "DSL directory child map"}, 63fa9e4066Sahrens { zap_byteswap, TRUE, "DSL dataset snap map" }, 64fa9e4066Sahrens { zap_byteswap, TRUE, "DSL props" }, 65fa9e4066Sahrens { byteswap_uint64_array, TRUE, "DSL dataset" }, 66fa9e4066Sahrens { zfs_znode_byteswap, TRUE, "ZFS znode" }, 67da6c28aaSamw { zfs_oldacl_byteswap, TRUE, "ZFS V0 ACL" }, 68fa9e4066Sahrens { byteswap_uint8_array, FALSE, "ZFS plain file" }, 69fa9e4066Sahrens { zap_byteswap, TRUE, "ZFS directory" }, 70fa9e4066Sahrens { zap_byteswap, TRUE, "ZFS master node" }, 71fa9e4066Sahrens { zap_byteswap, TRUE, "ZFS delete queue" }, 72fa9e4066Sahrens { byteswap_uint8_array, FALSE, "zvol object" }, 73fa9e4066Sahrens { zap_byteswap, TRUE, "zvol prop" }, 74fa9e4066Sahrens { byteswap_uint8_array, FALSE, "other uint8[]" }, 75fa9e4066Sahrens { byteswap_uint64_array, FALSE, "other uint64[]" }, 76fa9e4066Sahrens { zap_byteswap, TRUE, "other ZAP" }, 77ea8dc4b6Seschrock { zap_byteswap, TRUE, "persistent error log" }, 7806eeb2adSek110237 { byteswap_uint8_array, TRUE, "SPA history" }, 7906eeb2adSek110237 { byteswap_uint64_array, TRUE, "SPA history offsets" }, 80b1b8ab34Slling { zap_byteswap, TRUE, "Pool properties" }, 81da6c28aaSamw { zap_byteswap, TRUE, "DSL permissions" }, 82da6c28aaSamw { zfs_acl_byteswap, TRUE, "ZFS ACL" }, 83da6c28aaSamw { byteswap_uint8_array, TRUE, "ZFS SYSACL" }, 84da6c28aaSamw { byteswap_uint8_array, TRUE, "FUID table" }, 85add89791Smarks { byteswap_uint64_array, TRUE, "FUID table size" }, 86088f3894Sahrens { zap_byteswap, TRUE, "DSL dataset next clones"}, 87*3f9d6ad7SLin Ling { zap_byteswap, TRUE, "scan work queue" }, 8814843421SMatthew Ahrens { zap_byteswap, TRUE, "ZFS user/group used" }, 8914843421SMatthew Ahrens { zap_byteswap, TRUE, "ZFS user/group quota" }, 90842727c2SChris Kirby { zap_byteswap, TRUE, "snapshot refcount tags"}, 91b24ab676SJeff Bonwick { zap_byteswap, TRUE, "DDT ZAP algorithm" }, 92b24ab676SJeff Bonwick { zap_byteswap, TRUE, "DDT statistics" }, 930a586ceaSMark Shellenbaum { byteswap_uint8_array, TRUE, "System attributes" }, 940a586ceaSMark Shellenbaum { zap_byteswap, TRUE, "SA master node" }, 950a586ceaSMark Shellenbaum { zap_byteswap, TRUE, "SA attr registration" }, 96*3f9d6ad7SLin Ling { zap_byteswap, TRUE, "SA attr layouts" }, 97*3f9d6ad7SLin Ling { zap_byteswap, TRUE, "scan translations" }, 98*3f9d6ad7SLin Ling { byteswap_uint8_array, FALSE, "deduplicated block" }, 99*3f9d6ad7SLin Ling }; 100fa9e4066Sahrens 101fa9e4066Sahrens int 102ea8dc4b6Seschrock dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 10347cb52daSJeff Bonwick void *tag, dmu_buf_t **dbp, int flags) 104fa9e4066Sahrens { 105fa9e4066Sahrens dnode_t *dn; 106fa9e4066Sahrens uint64_t blkid; 107fa9e4066Sahrens dmu_buf_impl_t *db; 108ea8dc4b6Seschrock int err; 10947cb52daSJeff Bonwick int db_flags = DB_RF_CANFAIL; 11047cb52daSJeff Bonwick 11147cb52daSJeff Bonwick if (flags & DMU_READ_NO_PREFETCH) 11247cb52daSJeff Bonwick db_flags |= DB_RF_NOPREFETCH; 113fa9e4066Sahrens 114503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 115ea8dc4b6Seschrock if (err) 116ea8dc4b6Seschrock return (err); 117fa9e4066Sahrens blkid = dbuf_whichblock(dn, offset); 118fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 119ea8dc4b6Seschrock db = dbuf_hold(dn, blkid, tag); 120fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 121ea8dc4b6Seschrock if (db == NULL) { 122ea8dc4b6Seschrock err = EIO; 123ea8dc4b6Seschrock } else { 12447cb52daSJeff Bonwick err = dbuf_read(db, NULL, db_flags); 125ea8dc4b6Seschrock if (err) { 126ea8dc4b6Seschrock dbuf_rele(db, tag); 127ea8dc4b6Seschrock db = NULL; 128ea8dc4b6Seschrock } 129fa9e4066Sahrens } 130fa9e4066Sahrens 131ea8dc4b6Seschrock dnode_rele(dn, FTAG); 132ea8dc4b6Seschrock *dbp = &db->db; 133ea8dc4b6Seschrock return (err); 134fa9e4066Sahrens } 135fa9e4066Sahrens 136fa9e4066Sahrens int 137fa9e4066Sahrens dmu_bonus_max(void) 138fa9e4066Sahrens { 139fa9e4066Sahrens return (DN_MAX_BONUSLEN); 140fa9e4066Sahrens } 141fa9e4066Sahrens 1421934e92fSmaybee int 1431934e92fSmaybee dmu_set_bonus(dmu_buf_t *db, int newsize, dmu_tx_t *tx) 1441934e92fSmaybee { 1451934e92fSmaybee dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode; 1461934e92fSmaybee 1471934e92fSmaybee if (dn->dn_bonus != (dmu_buf_impl_t *)db) 1481934e92fSmaybee return (EINVAL); 1491934e92fSmaybee if (newsize < 0 || newsize > db->db_size) 1501934e92fSmaybee return (EINVAL); 1511934e92fSmaybee dnode_setbonuslen(dn, newsize, tx); 1521934e92fSmaybee return (0); 1531934e92fSmaybee } 1541934e92fSmaybee 1550a586ceaSMark Shellenbaum int 1560a586ceaSMark Shellenbaum dmu_set_bonustype(dmu_buf_t *db, dmu_object_type_t type, dmu_tx_t *tx) 1570a586ceaSMark Shellenbaum { 1580a586ceaSMark Shellenbaum dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode; 1590a586ceaSMark Shellenbaum 1600a586ceaSMark Shellenbaum if (type > DMU_OT_NUMTYPES) 1610a586ceaSMark Shellenbaum return (EINVAL); 1620a586ceaSMark Shellenbaum 1630a586ceaSMark Shellenbaum if (dn->dn_bonus != (dmu_buf_impl_t *)db) 1640a586ceaSMark Shellenbaum return (EINVAL); 1650a586ceaSMark Shellenbaum 1660a586ceaSMark Shellenbaum dnode_setbonus_type(dn, type, tx); 1670a586ceaSMark Shellenbaum return (0); 1680a586ceaSMark Shellenbaum } 1690a586ceaSMark Shellenbaum 1700a586ceaSMark Shellenbaum int 1710a586ceaSMark Shellenbaum dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 1720a586ceaSMark Shellenbaum { 1730a586ceaSMark Shellenbaum dnode_t *dn; 1740a586ceaSMark Shellenbaum int error; 1750a586ceaSMark Shellenbaum 1760a586ceaSMark Shellenbaum error = dnode_hold(os, object, FTAG, &dn); 1770a586ceaSMark Shellenbaum dbuf_rm_spill(dn, tx); 17806e0070dSMark Shellenbaum rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 17906e0070dSMark Shellenbaum dnode_rm_spill(dn, tx); 18006e0070dSMark Shellenbaum rw_exit(&dn->dn_struct_rwlock); 1810a586ceaSMark Shellenbaum dnode_rele(dn, FTAG); 1820a586ceaSMark Shellenbaum return (error); 1830a586ceaSMark Shellenbaum } 1840a586ceaSMark Shellenbaum 185fa9e4066Sahrens /* 186ea8dc4b6Seschrock * returns ENOENT, EIO, or 0. 187fa9e4066Sahrens */ 188ea8dc4b6Seschrock int 189ea8dc4b6Seschrock dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 190fa9e4066Sahrens { 191ea8dc4b6Seschrock dnode_t *dn; 192fa9e4066Sahrens dmu_buf_impl_t *db; 1931934e92fSmaybee int error; 194fa9e4066Sahrens 195503ad85cSMatthew Ahrens error = dnode_hold(os, object, FTAG, &dn); 1961934e92fSmaybee if (error) 1971934e92fSmaybee return (error); 198fa9e4066Sahrens 199fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 200ea8dc4b6Seschrock if (dn->dn_bonus == NULL) { 201fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 202ea8dc4b6Seschrock rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 203ea8dc4b6Seschrock if (dn->dn_bonus == NULL) 2041934e92fSmaybee dbuf_create_bonus(dn); 205fa9e4066Sahrens } 206ea8dc4b6Seschrock db = dn->dn_bonus; 207ea8dc4b6Seschrock rw_exit(&dn->dn_struct_rwlock); 2081934e92fSmaybee 2091934e92fSmaybee /* as long as the bonus buf is held, the dnode will be held */ 2101934e92fSmaybee if (refcount_add(&db->db_holds, tag) == 1) 2111934e92fSmaybee VERIFY(dnode_add_ref(dn, db)); 2121934e92fSmaybee 213fa9e4066Sahrens dnode_rele(dn, FTAG); 214ea8dc4b6Seschrock 21547cb52daSJeff Bonwick VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 216ea8dc4b6Seschrock 217ea8dc4b6Seschrock *dbp = &db->db; 218ea8dc4b6Seschrock return (0); 219fa9e4066Sahrens } 220fa9e4066Sahrens 22113506d1eSmaybee /* 2220a586ceaSMark Shellenbaum * returns ENOENT, EIO, or 0. 2230a586ceaSMark Shellenbaum * 2240a586ceaSMark Shellenbaum * This interface will allocate a blank spill dbuf when a spill blk 2250a586ceaSMark Shellenbaum * doesn't already exist on the dnode. 2260a586ceaSMark Shellenbaum * 2270a586ceaSMark Shellenbaum * if you only want to find an already existing spill db, then 2280a586ceaSMark Shellenbaum * dmu_spill_hold_existing() should be used. 2290a586ceaSMark Shellenbaum */ 2300a586ceaSMark Shellenbaum int 2310a586ceaSMark Shellenbaum dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 2320a586ceaSMark Shellenbaum { 2330a586ceaSMark Shellenbaum dmu_buf_impl_t *db = NULL; 2340a586ceaSMark Shellenbaum int err; 2350a586ceaSMark Shellenbaum 2360a586ceaSMark Shellenbaum if ((flags & DB_RF_HAVESTRUCT) == 0) 2370a586ceaSMark Shellenbaum rw_enter(&dn->dn_struct_rwlock, RW_READER); 2380a586ceaSMark Shellenbaum 2390a586ceaSMark Shellenbaum db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 2400a586ceaSMark Shellenbaum 2410a586ceaSMark Shellenbaum if ((flags & DB_RF_HAVESTRUCT) == 0) 2420a586ceaSMark Shellenbaum rw_exit(&dn->dn_struct_rwlock); 2430a586ceaSMark Shellenbaum 2440a586ceaSMark Shellenbaum ASSERT(db != NULL); 2450a586ceaSMark Shellenbaum err = dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | flags); 2460a586ceaSMark Shellenbaum *dbp = &db->db; 2470a586ceaSMark Shellenbaum return (err); 2480a586ceaSMark Shellenbaum } 2490a586ceaSMark Shellenbaum 2500a586ceaSMark Shellenbaum int 2510a586ceaSMark Shellenbaum dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 2520a586ceaSMark Shellenbaum { 2530a586ceaSMark Shellenbaum dnode_t *dn = ((dmu_buf_impl_t *)bonus)->db_dnode; 2540a586ceaSMark Shellenbaum int err; 2550a586ceaSMark Shellenbaum 2560a586ceaSMark Shellenbaum if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) 2570a586ceaSMark Shellenbaum return (EINVAL); 2580a586ceaSMark Shellenbaum rw_enter(&dn->dn_struct_rwlock, RW_READER); 2590a586ceaSMark Shellenbaum 2600a586ceaSMark Shellenbaum if (!dn->dn_have_spill) { 2610a586ceaSMark Shellenbaum rw_exit(&dn->dn_struct_rwlock); 2620a586ceaSMark Shellenbaum return (ENOENT); 2630a586ceaSMark Shellenbaum } 2640a586ceaSMark Shellenbaum err = dmu_spill_hold_by_dnode(dn, DB_RF_HAVESTRUCT, tag, dbp); 2650a586ceaSMark Shellenbaum rw_exit(&dn->dn_struct_rwlock); 2660a586ceaSMark Shellenbaum return (err); 2670a586ceaSMark Shellenbaum } 2680a586ceaSMark Shellenbaum 2690a586ceaSMark Shellenbaum int 2700a586ceaSMark Shellenbaum dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 2710a586ceaSMark Shellenbaum { 2720a586ceaSMark Shellenbaum return (dmu_spill_hold_by_dnode(((dmu_buf_impl_t *)bonus)->db_dnode, 2730a586ceaSMark Shellenbaum 0, tag, dbp)); 2740a586ceaSMark Shellenbaum } 2750a586ceaSMark Shellenbaum 2760a586ceaSMark Shellenbaum /* 27713506d1eSmaybee * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 27813506d1eSmaybee * to take a held dnode rather than <os, object> -- the lookup is wasteful, 27913506d1eSmaybee * and can induce severe lock contention when writing to several files 28013506d1eSmaybee * whose dnodes are in the same block. 28113506d1eSmaybee */ 28213506d1eSmaybee static int 2837bfdf011SNeil Perrin dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 2847bfdf011SNeil Perrin int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 285fa9e4066Sahrens { 28605715f94SMark Maybee dsl_pool_t *dp = NULL; 287fa9e4066Sahrens dmu_buf_t **dbp; 288fa9e4066Sahrens uint64_t blkid, nblks, i; 2897bfdf011SNeil Perrin uint32_t dbuf_flags; 290ea8dc4b6Seschrock int err; 291ea8dc4b6Seschrock zio_t *zio; 29205715f94SMark Maybee hrtime_t start; 293ea8dc4b6Seschrock 294ea8dc4b6Seschrock ASSERT(length <= DMU_MAX_ACCESS); 295fa9e4066Sahrens 296c87b8fc5SMark J Musante dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT; 2977bfdf011SNeil Perrin if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz) 2987bfdf011SNeil Perrin dbuf_flags |= DB_RF_NOPREFETCH; 299ea8dc4b6Seschrock 300fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 301fa9e4066Sahrens if (dn->dn_datablkshift) { 302fa9e4066Sahrens int blkshift = dn->dn_datablkshift; 303fa9e4066Sahrens nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) - 304fa9e4066Sahrens P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift; 305fa9e4066Sahrens } else { 3060125049cSahrens if (offset + length > dn->dn_datablksz) { 3070125049cSahrens zfs_panic_recover("zfs: accessing past end of object " 3080125049cSahrens "%llx/%llx (size=%u access=%llu+%llu)", 3090125049cSahrens (longlong_t)dn->dn_objset-> 3100125049cSahrens os_dsl_dataset->ds_object, 3110125049cSahrens (longlong_t)dn->dn_object, dn->dn_datablksz, 3120125049cSahrens (longlong_t)offset, (longlong_t)length); 313c87b8fc5SMark J Musante rw_exit(&dn->dn_struct_rwlock); 3140125049cSahrens return (EIO); 3150125049cSahrens } 316fa9e4066Sahrens nblks = 1; 317fa9e4066Sahrens } 318ea8dc4b6Seschrock dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 319fa9e4066Sahrens 32005715f94SMark Maybee if (dn->dn_objset->os_dsl_dataset) 32105715f94SMark Maybee dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool; 32205715f94SMark Maybee if (dp && dsl_pool_sync_context(dp)) 32305715f94SMark Maybee start = gethrtime(); 324e14bb325SJeff Bonwick zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 325fa9e4066Sahrens blkid = dbuf_whichblock(dn, offset); 326fa9e4066Sahrens for (i = 0; i < nblks; i++) { 327ea8dc4b6Seschrock dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag); 328ea8dc4b6Seschrock if (db == NULL) { 329ea8dc4b6Seschrock rw_exit(&dn->dn_struct_rwlock); 330ea8dc4b6Seschrock dmu_buf_rele_array(dbp, nblks, tag); 331ea8dc4b6Seschrock zio_nowait(zio); 332ea8dc4b6Seschrock return (EIO); 333ea8dc4b6Seschrock } 334ea8dc4b6Seschrock /* initiate async i/o */ 33513506d1eSmaybee if (read) { 3367bfdf011SNeil Perrin (void) dbuf_read(db, zio, dbuf_flags); 337ea8dc4b6Seschrock } 338ea8dc4b6Seschrock dbp[i] = &db->db; 339fa9e4066Sahrens } 340fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 341fa9e4066Sahrens 342ea8dc4b6Seschrock /* wait for async i/o */ 343ea8dc4b6Seschrock err = zio_wait(zio); 34405715f94SMark Maybee /* track read overhead when we are in sync context */ 34505715f94SMark Maybee if (dp && dsl_pool_sync_context(dp)) 34605715f94SMark Maybee dp->dp_read_overhead += gethrtime() - start; 347ea8dc4b6Seschrock if (err) { 348ea8dc4b6Seschrock dmu_buf_rele_array(dbp, nblks, tag); 349ea8dc4b6Seschrock return (err); 350ea8dc4b6Seschrock } 351ea8dc4b6Seschrock 352ea8dc4b6Seschrock /* wait for other io to complete */ 353ea8dc4b6Seschrock if (read) { 354ea8dc4b6Seschrock for (i = 0; i < nblks; i++) { 355ea8dc4b6Seschrock dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 356ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 357ea8dc4b6Seschrock while (db->db_state == DB_READ || 358ea8dc4b6Seschrock db->db_state == DB_FILL) 359ea8dc4b6Seschrock cv_wait(&db->db_changed, &db->db_mtx); 360ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED) 361ea8dc4b6Seschrock err = EIO; 362ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 363ea8dc4b6Seschrock if (err) { 364ea8dc4b6Seschrock dmu_buf_rele_array(dbp, nblks, tag); 365ea8dc4b6Seschrock return (err); 366ea8dc4b6Seschrock } 367ea8dc4b6Seschrock } 368ea8dc4b6Seschrock } 369ea8dc4b6Seschrock 370ea8dc4b6Seschrock *numbufsp = nblks; 371ea8dc4b6Seschrock *dbpp = dbp; 372ea8dc4b6Seschrock return (0); 373fa9e4066Sahrens } 374fa9e4066Sahrens 375a2eea2e1Sahrens static int 37613506d1eSmaybee dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 37713506d1eSmaybee uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 37813506d1eSmaybee { 37913506d1eSmaybee dnode_t *dn; 38013506d1eSmaybee int err; 38113506d1eSmaybee 382503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 38313506d1eSmaybee if (err) 38413506d1eSmaybee return (err); 38513506d1eSmaybee 38613506d1eSmaybee err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 3877bfdf011SNeil Perrin numbufsp, dbpp, DMU_READ_PREFETCH); 38813506d1eSmaybee 38913506d1eSmaybee dnode_rele(dn, FTAG); 39013506d1eSmaybee 39113506d1eSmaybee return (err); 39213506d1eSmaybee } 39313506d1eSmaybee 39413506d1eSmaybee int 39513506d1eSmaybee dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset, 39613506d1eSmaybee uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 39713506d1eSmaybee { 39813506d1eSmaybee dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode; 39913506d1eSmaybee int err; 40013506d1eSmaybee 40113506d1eSmaybee err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 4027bfdf011SNeil Perrin numbufsp, dbpp, DMU_READ_PREFETCH); 40313506d1eSmaybee 40413506d1eSmaybee return (err); 40513506d1eSmaybee } 40613506d1eSmaybee 407fa9e4066Sahrens void 408ea8dc4b6Seschrock dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 409fa9e4066Sahrens { 410fa9e4066Sahrens int i; 411fa9e4066Sahrens dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 412fa9e4066Sahrens 413fa9e4066Sahrens if (numbufs == 0) 414fa9e4066Sahrens return; 415fa9e4066Sahrens 416ea8dc4b6Seschrock for (i = 0; i < numbufs; i++) { 417ea8dc4b6Seschrock if (dbp[i]) 418ea8dc4b6Seschrock dbuf_rele(dbp[i], tag); 419ea8dc4b6Seschrock } 420fa9e4066Sahrens 421fa9e4066Sahrens kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 422fa9e4066Sahrens } 423fa9e4066Sahrens 424fa9e4066Sahrens void 425fa9e4066Sahrens dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len) 426fa9e4066Sahrens { 427fa9e4066Sahrens dnode_t *dn; 428fa9e4066Sahrens uint64_t blkid; 429ea8dc4b6Seschrock int nblks, i, err; 430fa9e4066Sahrens 431416e0cd8Sek110237 if (zfs_prefetch_disable) 432416e0cd8Sek110237 return; 433416e0cd8Sek110237 434fa9e4066Sahrens if (len == 0) { /* they're interested in the bonus buffer */ 435503ad85cSMatthew Ahrens dn = os->os_meta_dnode; 436fa9e4066Sahrens 437fa9e4066Sahrens if (object == 0 || object >= DN_MAX_OBJECT) 438fa9e4066Sahrens return; 439fa9e4066Sahrens 440fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 441fa9e4066Sahrens blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t)); 442fa9e4066Sahrens dbuf_prefetch(dn, blkid); 443fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 444fa9e4066Sahrens return; 445fa9e4066Sahrens } 446fa9e4066Sahrens 447fa9e4066Sahrens /* 448fa9e4066Sahrens * XXX - Note, if the dnode for the requested object is not 449fa9e4066Sahrens * already cached, we will do a *synchronous* read in the 450fa9e4066Sahrens * dnode_hold() call. The same is true for any indirects. 451fa9e4066Sahrens */ 452503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 453ea8dc4b6Seschrock if (err != 0) 454fa9e4066Sahrens return; 455fa9e4066Sahrens 456fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 457fa9e4066Sahrens if (dn->dn_datablkshift) { 458fa9e4066Sahrens int blkshift = dn->dn_datablkshift; 459fa9e4066Sahrens nblks = (P2ROUNDUP(offset+len, 1<<blkshift) - 460fa9e4066Sahrens P2ALIGN(offset, 1<<blkshift)) >> blkshift; 461fa9e4066Sahrens } else { 462fa9e4066Sahrens nblks = (offset < dn->dn_datablksz); 463fa9e4066Sahrens } 464fa9e4066Sahrens 465fa9e4066Sahrens if (nblks != 0) { 466fa9e4066Sahrens blkid = dbuf_whichblock(dn, offset); 467fa9e4066Sahrens for (i = 0; i < nblks; i++) 468fa9e4066Sahrens dbuf_prefetch(dn, blkid+i); 469fa9e4066Sahrens } 470fa9e4066Sahrens 471fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 472fa9e4066Sahrens 473fa9e4066Sahrens dnode_rele(dn, FTAG); 474fa9e4066Sahrens } 475fa9e4066Sahrens 47676256205SMark Maybee /* 47776256205SMark Maybee * Get the next "chunk" of file data to free. We traverse the file from 47876256205SMark Maybee * the end so that the file gets shorter over time (if we crashes in the 47976256205SMark Maybee * middle, this will leave us in a better state). We find allocated file 48076256205SMark Maybee * data by simply searching the allocated level 1 indirects. 48176256205SMark Maybee */ 482cdb0ab79Smaybee static int 48376256205SMark Maybee get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit) 484cdb0ab79Smaybee { 48576256205SMark Maybee uint64_t len = *start - limit; 48676256205SMark Maybee uint64_t blkcnt = 0; 48776256205SMark Maybee uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1)); 48876256205SMark Maybee uint64_t iblkrange = 4891c8564a7SMark Maybee dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 490cdb0ab79Smaybee 49176256205SMark Maybee ASSERT(limit <= *start); 492cdb0ab79Smaybee 49376256205SMark Maybee if (len <= iblkrange * maxblks) { 49476256205SMark Maybee *start = limit; 495cdb0ab79Smaybee return (0); 496cdb0ab79Smaybee } 49776256205SMark Maybee ASSERT(ISP2(iblkrange)); 498cdb0ab79Smaybee 49976256205SMark Maybee while (*start > limit && blkcnt < maxblks) { 5001c8564a7SMark Maybee int err; 501cdb0ab79Smaybee 50276256205SMark Maybee /* find next allocated L1 indirect */ 503cdb0ab79Smaybee err = dnode_next_offset(dn, 50476256205SMark Maybee DNODE_FIND_BACKWARDS, start, 2, 1, 0); 505cdb0ab79Smaybee 50676256205SMark Maybee /* if there are no more, then we are done */ 50776256205SMark Maybee if (err == ESRCH) { 50876256205SMark Maybee *start = limit; 509cdb0ab79Smaybee return (0); 51076256205SMark Maybee } else if (err) { 511cdb0ab79Smaybee return (err); 51276256205SMark Maybee } 51376256205SMark Maybee blkcnt += 1; 514cdb0ab79Smaybee 51576256205SMark Maybee /* reset offset to end of "next" block back */ 51676256205SMark Maybee *start = P2ALIGN(*start, iblkrange); 51776256205SMark Maybee if (*start <= limit) 51876256205SMark Maybee *start = limit; 51976256205SMark Maybee else 52076256205SMark Maybee *start -= 1; 521cdb0ab79Smaybee } 522cdb0ab79Smaybee return (0); 523cdb0ab79Smaybee } 524cdb0ab79Smaybee 525cdb0ab79Smaybee static int 526cdb0ab79Smaybee dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 527cdb0ab79Smaybee uint64_t length, boolean_t free_dnode) 528cdb0ab79Smaybee { 529cdb0ab79Smaybee dmu_tx_t *tx; 530cdb0ab79Smaybee uint64_t object_size, start, end, len; 531cdb0ab79Smaybee boolean_t trunc = (length == DMU_OBJECT_END); 532cdb0ab79Smaybee int align, err; 533cdb0ab79Smaybee 534cdb0ab79Smaybee align = 1 << dn->dn_datablkshift; 535cdb0ab79Smaybee ASSERT(align > 0); 536cdb0ab79Smaybee object_size = align == 1 ? dn->dn_datablksz : 537cdb0ab79Smaybee (dn->dn_maxblkid + 1) << dn->dn_datablkshift; 538cdb0ab79Smaybee 53914843421SMatthew Ahrens end = offset + length; 54014843421SMatthew Ahrens if (trunc || end > object_size) 541cdb0ab79Smaybee end = object_size; 542cdb0ab79Smaybee if (end <= offset) 543cdb0ab79Smaybee return (0); 544cdb0ab79Smaybee length = end - offset; 545cdb0ab79Smaybee 546cdb0ab79Smaybee while (length) { 547cdb0ab79Smaybee start = end; 54814843421SMatthew Ahrens /* assert(offset <= start) */ 549cdb0ab79Smaybee err = get_next_chunk(dn, &start, offset); 550cdb0ab79Smaybee if (err) 551cdb0ab79Smaybee return (err); 552cdb0ab79Smaybee len = trunc ? DMU_OBJECT_END : end - start; 553cdb0ab79Smaybee 554cdb0ab79Smaybee tx = dmu_tx_create(os); 555cdb0ab79Smaybee dmu_tx_hold_free(tx, dn->dn_object, start, len); 556cdb0ab79Smaybee err = dmu_tx_assign(tx, TXG_WAIT); 557cdb0ab79Smaybee if (err) { 558cdb0ab79Smaybee dmu_tx_abort(tx); 559cdb0ab79Smaybee return (err); 560cdb0ab79Smaybee } 561cdb0ab79Smaybee 562cdb0ab79Smaybee dnode_free_range(dn, start, trunc ? -1 : len, tx); 563cdb0ab79Smaybee 5641c8564a7SMark Maybee if (start == 0 && free_dnode) { 5651c8564a7SMark Maybee ASSERT(trunc); 566cdb0ab79Smaybee dnode_free(dn, tx); 5671c8564a7SMark Maybee } 568cdb0ab79Smaybee 569cdb0ab79Smaybee length -= end - start; 570cdb0ab79Smaybee 571cdb0ab79Smaybee dmu_tx_commit(tx); 572cdb0ab79Smaybee end = start; 573cdb0ab79Smaybee } 574cdb0ab79Smaybee return (0); 575cdb0ab79Smaybee } 576cdb0ab79Smaybee 577cdb0ab79Smaybee int 578cdb0ab79Smaybee dmu_free_long_range(objset_t *os, uint64_t object, 579cdb0ab79Smaybee uint64_t offset, uint64_t length) 580cdb0ab79Smaybee { 581cdb0ab79Smaybee dnode_t *dn; 582cdb0ab79Smaybee int err; 583cdb0ab79Smaybee 584503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 585cdb0ab79Smaybee if (err != 0) 586cdb0ab79Smaybee return (err); 587cdb0ab79Smaybee err = dmu_free_long_range_impl(os, dn, offset, length, FALSE); 588cdb0ab79Smaybee dnode_rele(dn, FTAG); 589cdb0ab79Smaybee return (err); 590cdb0ab79Smaybee } 591cdb0ab79Smaybee 592cdb0ab79Smaybee int 593cdb0ab79Smaybee dmu_free_object(objset_t *os, uint64_t object) 594cdb0ab79Smaybee { 595cdb0ab79Smaybee dnode_t *dn; 596cdb0ab79Smaybee dmu_tx_t *tx; 597cdb0ab79Smaybee int err; 598cdb0ab79Smaybee 599503ad85cSMatthew Ahrens err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 600cdb0ab79Smaybee FTAG, &dn); 601cdb0ab79Smaybee if (err != 0) 602cdb0ab79Smaybee return (err); 603cdb0ab79Smaybee if (dn->dn_nlevels == 1) { 604cdb0ab79Smaybee tx = dmu_tx_create(os); 605cdb0ab79Smaybee dmu_tx_hold_bonus(tx, object); 606cdb0ab79Smaybee dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END); 607cdb0ab79Smaybee err = dmu_tx_assign(tx, TXG_WAIT); 608cdb0ab79Smaybee if (err == 0) { 609cdb0ab79Smaybee dnode_free_range(dn, 0, DMU_OBJECT_END, tx); 610cdb0ab79Smaybee dnode_free(dn, tx); 611cdb0ab79Smaybee dmu_tx_commit(tx); 612cdb0ab79Smaybee } else { 613cdb0ab79Smaybee dmu_tx_abort(tx); 614cdb0ab79Smaybee } 615cdb0ab79Smaybee } else { 616cdb0ab79Smaybee err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE); 617cdb0ab79Smaybee } 618cdb0ab79Smaybee dnode_rele(dn, FTAG); 619cdb0ab79Smaybee return (err); 620cdb0ab79Smaybee } 621cdb0ab79Smaybee 622ea8dc4b6Seschrock int 623fa9e4066Sahrens dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 624fa9e4066Sahrens uint64_t size, dmu_tx_t *tx) 625fa9e4066Sahrens { 626ea8dc4b6Seschrock dnode_t *dn; 627503ad85cSMatthew Ahrens int err = dnode_hold(os, object, FTAG, &dn); 628ea8dc4b6Seschrock if (err) 629ea8dc4b6Seschrock return (err); 630fa9e4066Sahrens ASSERT(offset < UINT64_MAX); 631fa9e4066Sahrens ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 632fa9e4066Sahrens dnode_free_range(dn, offset, size, tx); 633fa9e4066Sahrens dnode_rele(dn, FTAG); 634ea8dc4b6Seschrock return (0); 635fa9e4066Sahrens } 636fa9e4066Sahrens 637ea8dc4b6Seschrock int 638ea8dc4b6Seschrock dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 6397bfdf011SNeil Perrin void *buf, uint32_t flags) 640fa9e4066Sahrens { 641fa9e4066Sahrens dnode_t *dn; 642fa9e4066Sahrens dmu_buf_t **dbp; 643c87b8fc5SMark J Musante int numbufs, err; 644fa9e4066Sahrens 645503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 646ea8dc4b6Seschrock if (err) 647ea8dc4b6Seschrock return (err); 648feb08c6bSbillm 649feb08c6bSbillm /* 650feb08c6bSbillm * Deal with odd block sizes, where there can't be data past the first 651feb08c6bSbillm * block. If we ever do the tail block optimization, we will need to 652feb08c6bSbillm * handle that here as well. 653feb08c6bSbillm */ 654c87b8fc5SMark J Musante if (dn->dn_maxblkid == 0) { 655fa9e4066Sahrens int newsz = offset > dn->dn_datablksz ? 0 : 656fa9e4066Sahrens MIN(size, dn->dn_datablksz - offset); 657fa9e4066Sahrens bzero((char *)buf + newsz, size - newsz); 658fa9e4066Sahrens size = newsz; 659fa9e4066Sahrens } 660fa9e4066Sahrens 661fa9e4066Sahrens while (size > 0) { 662fa9e4066Sahrens uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 663c87b8fc5SMark J Musante int i; 664fa9e4066Sahrens 665fa9e4066Sahrens /* 666fa9e4066Sahrens * NB: we could do this block-at-a-time, but it's nice 667fa9e4066Sahrens * to be reading in parallel. 668fa9e4066Sahrens */ 669a2eea2e1Sahrens err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 6707bfdf011SNeil Perrin TRUE, FTAG, &numbufs, &dbp, flags); 671ea8dc4b6Seschrock if (err) 6721934e92fSmaybee break; 673fa9e4066Sahrens 674fa9e4066Sahrens for (i = 0; i < numbufs; i++) { 675fa9e4066Sahrens int tocpy; 676fa9e4066Sahrens int bufoff; 677fa9e4066Sahrens dmu_buf_t *db = dbp[i]; 678fa9e4066Sahrens 679fa9e4066Sahrens ASSERT(size > 0); 680fa9e4066Sahrens 681fa9e4066Sahrens bufoff = offset - db->db_offset; 682fa9e4066Sahrens tocpy = (int)MIN(db->db_size - bufoff, size); 683fa9e4066Sahrens 684fa9e4066Sahrens bcopy((char *)db->db_data + bufoff, buf, tocpy); 685fa9e4066Sahrens 686fa9e4066Sahrens offset += tocpy; 687fa9e4066Sahrens size -= tocpy; 688fa9e4066Sahrens buf = (char *)buf + tocpy; 689fa9e4066Sahrens } 690ea8dc4b6Seschrock dmu_buf_rele_array(dbp, numbufs, FTAG); 691fa9e4066Sahrens } 692a2eea2e1Sahrens dnode_rele(dn, FTAG); 6931934e92fSmaybee return (err); 694fa9e4066Sahrens } 695fa9e4066Sahrens 696fa9e4066Sahrens void 697fa9e4066Sahrens dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 698fa9e4066Sahrens const void *buf, dmu_tx_t *tx) 699fa9e4066Sahrens { 700fa9e4066Sahrens dmu_buf_t **dbp; 701fa9e4066Sahrens int numbufs, i; 702fa9e4066Sahrens 70313506d1eSmaybee if (size == 0) 70413506d1eSmaybee return; 70513506d1eSmaybee 706ea8dc4b6Seschrock VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 707ea8dc4b6Seschrock FALSE, FTAG, &numbufs, &dbp)); 708fa9e4066Sahrens 709fa9e4066Sahrens for (i = 0; i < numbufs; i++) { 710fa9e4066Sahrens int tocpy; 711fa9e4066Sahrens int bufoff; 712fa9e4066Sahrens dmu_buf_t *db = dbp[i]; 713fa9e4066Sahrens 714fa9e4066Sahrens ASSERT(size > 0); 715fa9e4066Sahrens 716fa9e4066Sahrens bufoff = offset - db->db_offset; 717fa9e4066Sahrens tocpy = (int)MIN(db->db_size - bufoff, size); 718fa9e4066Sahrens 719fa9e4066Sahrens ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 720fa9e4066Sahrens 721fa9e4066Sahrens if (tocpy == db->db_size) 722fa9e4066Sahrens dmu_buf_will_fill(db, tx); 723fa9e4066Sahrens else 724fa9e4066Sahrens dmu_buf_will_dirty(db, tx); 725fa9e4066Sahrens 726fa9e4066Sahrens bcopy(buf, (char *)db->db_data + bufoff, tocpy); 727fa9e4066Sahrens 728fa9e4066Sahrens if (tocpy == db->db_size) 729fa9e4066Sahrens dmu_buf_fill_done(db, tx); 730fa9e4066Sahrens 731fa9e4066Sahrens offset += tocpy; 732fa9e4066Sahrens size -= tocpy; 733fa9e4066Sahrens buf = (char *)buf + tocpy; 734fa9e4066Sahrens } 735ea8dc4b6Seschrock dmu_buf_rele_array(dbp, numbufs, FTAG); 736fa9e4066Sahrens } 737fa9e4066Sahrens 73882c9918fSTim Haley void 73982c9918fSTim Haley dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 74082c9918fSTim Haley dmu_tx_t *tx) 74182c9918fSTim Haley { 74282c9918fSTim Haley dmu_buf_t **dbp; 74382c9918fSTim Haley int numbufs, i; 74482c9918fSTim Haley 74582c9918fSTim Haley if (size == 0) 74682c9918fSTim Haley return; 74782c9918fSTim Haley 74882c9918fSTim Haley VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 74982c9918fSTim Haley FALSE, FTAG, &numbufs, &dbp)); 75082c9918fSTim Haley 75182c9918fSTim Haley for (i = 0; i < numbufs; i++) { 75282c9918fSTim Haley dmu_buf_t *db = dbp[i]; 75382c9918fSTim Haley 75482c9918fSTim Haley dmu_buf_will_not_fill(db, tx); 75582c9918fSTim Haley } 75682c9918fSTim Haley dmu_buf_rele_array(dbp, numbufs, FTAG); 75782c9918fSTim Haley } 75882c9918fSTim Haley 759c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* 760c242f9a0Schunli zhang - Sun Microsystems - Irvine United States * DMU support for xuio 761c242f9a0Schunli zhang - Sun Microsystems - Irvine United States */ 762c242f9a0Schunli zhang - Sun Microsystems - Irvine United States kstat_t *xuio_ksp = NULL; 763c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 764c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int 765c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_init(xuio_t *xuio, int nblk) 766c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 767c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_t *priv; 768c242f9a0Schunli zhang - Sun Microsystems - Irvine United States uio_t *uio = &xuio->xu_uio; 769c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 770c242f9a0Schunli zhang - Sun Microsystems - Irvine United States uio->uio_iovcnt = nblk; 771c242f9a0Schunli zhang - Sun Microsystems - Irvine United States uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 772c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 773c242f9a0Schunli zhang - Sun Microsystems - Irvine United States priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 774c242f9a0Schunli zhang - Sun Microsystems - Irvine United States priv->cnt = nblk; 775c242f9a0Schunli zhang - Sun Microsystems - Irvine United States priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 776c242f9a0Schunli zhang - Sun Microsystems - Irvine United States priv->iovp = uio->uio_iov; 777c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIO_XUZC_PRIV(xuio) = priv; 778c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 779c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (XUIO_XUZC_RW(xuio) == UIO_READ) 780c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 781c242f9a0Schunli zhang - Sun Microsystems - Irvine United States else 782c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 783c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 784c242f9a0Schunli zhang - Sun Microsystems - Irvine United States return (0); 785c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 786c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 787c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 788c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_fini(xuio_t *xuio) 789c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 790c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 791c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int nblk = priv->cnt; 792c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 793c242f9a0Schunli zhang - Sun Microsystems - Irvine United States kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 794c242f9a0Schunli zhang - Sun Microsystems - Irvine United States kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 795c242f9a0Schunli zhang - Sun Microsystems - Irvine United States kmem_free(priv, sizeof (dmu_xuio_t)); 796c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 797c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (XUIO_XUZC_RW(xuio) == UIO_READ) 798c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 799c242f9a0Schunli zhang - Sun Microsystems - Irvine United States else 800c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 801c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 802c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 803c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* 804c242f9a0Schunli zhang - Sun Microsystems - Irvine United States * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 805c242f9a0Schunli zhang - Sun Microsystems - Irvine United States * and increase priv->next by 1. 806c242f9a0Schunli zhang - Sun Microsystems - Irvine United States */ 807c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int 808c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 809c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 810c242f9a0Schunli zhang - Sun Microsystems - Irvine United States struct iovec *iov; 811c242f9a0Schunli zhang - Sun Microsystems - Irvine United States uio_t *uio = &xuio->xu_uio; 812c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 813c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int i = priv->next++; 814c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 815c242f9a0Schunli zhang - Sun Microsystems - Irvine United States ASSERT(i < priv->cnt); 816c242f9a0Schunli zhang - Sun Microsystems - Irvine United States ASSERT(off + n <= arc_buf_size(abuf)); 817c242f9a0Schunli zhang - Sun Microsystems - Irvine United States iov = uio->uio_iov + i; 818c242f9a0Schunli zhang - Sun Microsystems - Irvine United States iov->iov_base = (char *)abuf->b_data + off; 819c242f9a0Schunli zhang - Sun Microsystems - Irvine United States iov->iov_len = n; 820c242f9a0Schunli zhang - Sun Microsystems - Irvine United States priv->bufs[i] = abuf; 821c242f9a0Schunli zhang - Sun Microsystems - Irvine United States return (0); 822c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 823c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 824c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int 825c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_cnt(xuio_t *xuio) 826c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 827c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 828c242f9a0Schunli zhang - Sun Microsystems - Irvine United States return (priv->cnt); 829c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 830c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 831c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t * 832c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_arcbuf(xuio_t *xuio, int i) 833c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 834c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 835c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 836c242f9a0Schunli zhang - Sun Microsystems - Irvine United States ASSERT(i < priv->cnt); 837c242f9a0Schunli zhang - Sun Microsystems - Irvine United States return (priv->bufs[i]); 838c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 839c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 840c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 841c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_clear(xuio_t *xuio, int i) 842c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 843c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 844c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 845c242f9a0Schunli zhang - Sun Microsystems - Irvine United States ASSERT(i < priv->cnt); 846c242f9a0Schunli zhang - Sun Microsystems - Irvine United States priv->bufs[i] = NULL; 847c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 848c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 849c242f9a0Schunli zhang - Sun Microsystems - Irvine United States static void 850c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_init(void) 851c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 852c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 853c242f9a0Schunli zhang - Sun Microsystems - Irvine United States KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 854c242f9a0Schunli zhang - Sun Microsystems - Irvine United States KSTAT_FLAG_VIRTUAL); 855c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (xuio_ksp != NULL) { 856c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_ksp->ks_data = &xuio_stats; 857c242f9a0Schunli zhang - Sun Microsystems - Irvine United States kstat_install(xuio_ksp); 858c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 859c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 860c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 861c242f9a0Schunli zhang - Sun Microsystems - Irvine United States static void 862c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_fini(void) 863c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 864c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (xuio_ksp != NULL) { 865c242f9a0Schunli zhang - Sun Microsystems - Irvine United States kstat_delete(xuio_ksp); 866c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_ksp = NULL; 867c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 868c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 869c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 870c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 871c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_wbuf_copied() 872c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 873c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_BUMP(xuiostat_wbuf_copied); 874c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 875c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 876c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 877c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_wbuf_nocopy() 878c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 879c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 880c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 881c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 882fa9e4066Sahrens #ifdef _KERNEL 883fa9e4066Sahrens int 884feb08c6bSbillm dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 885feb08c6bSbillm { 886feb08c6bSbillm dmu_buf_t **dbp; 887feb08c6bSbillm int numbufs, i, err; 888c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_t *xuio = NULL; 889feb08c6bSbillm 890feb08c6bSbillm /* 891feb08c6bSbillm * NB: we could do this block-at-a-time, but it's nice 892feb08c6bSbillm * to be reading in parallel. 893feb08c6bSbillm */ 894feb08c6bSbillm err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG, 895feb08c6bSbillm &numbufs, &dbp); 896feb08c6bSbillm if (err) 897feb08c6bSbillm return (err); 898feb08c6bSbillm 899c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (uio->uio_extflg == UIO_XUIO) 900c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio = (xuio_t *)uio; 901c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 902feb08c6bSbillm for (i = 0; i < numbufs; i++) { 903feb08c6bSbillm int tocpy; 904feb08c6bSbillm int bufoff; 905feb08c6bSbillm dmu_buf_t *db = dbp[i]; 906feb08c6bSbillm 907feb08c6bSbillm ASSERT(size > 0); 908feb08c6bSbillm 909feb08c6bSbillm bufoff = uio->uio_loffset - db->db_offset; 910feb08c6bSbillm tocpy = (int)MIN(db->db_size - bufoff, size); 911feb08c6bSbillm 912c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (xuio) { 913c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 914c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t *dbuf_abuf = dbi->db_buf; 915c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 916c242f9a0Schunli zhang - Sun Microsystems - Irvine United States err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 917c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (!err) { 918c242f9a0Schunli zhang - Sun Microsystems - Irvine United States uio->uio_resid -= tocpy; 919c242f9a0Schunli zhang - Sun Microsystems - Irvine United States uio->uio_loffset += tocpy; 920c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 921c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 922c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (abuf == dbuf_abuf) 923c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 924c242f9a0Schunli zhang - Sun Microsystems - Irvine United States else 925c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_BUMP(xuiostat_rbuf_copied); 926c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } else { 927feb08c6bSbillm err = uiomove((char *)db->db_data + bufoff, tocpy, 928feb08c6bSbillm UIO_READ, uio); 929c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 930feb08c6bSbillm if (err) 931feb08c6bSbillm break; 932feb08c6bSbillm 933feb08c6bSbillm size -= tocpy; 934feb08c6bSbillm } 935feb08c6bSbillm dmu_buf_rele_array(dbp, numbufs, FTAG); 936feb08c6bSbillm 937feb08c6bSbillm return (err); 938feb08c6bSbillm } 939feb08c6bSbillm 94094d1a210STim Haley static int 94194d1a210STim Haley dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 942fa9e4066Sahrens { 943fa9e4066Sahrens dmu_buf_t **dbp; 94494d1a210STim Haley int numbufs; 945fa9e4066Sahrens int err = 0; 94694d1a210STim Haley int i; 947fa9e4066Sahrens 94894d1a210STim Haley err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 94994d1a210STim Haley FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 950ea8dc4b6Seschrock if (err) 951ea8dc4b6Seschrock return (err); 952fa9e4066Sahrens 953fa9e4066Sahrens for (i = 0; i < numbufs; i++) { 954fa9e4066Sahrens int tocpy; 955fa9e4066Sahrens int bufoff; 956fa9e4066Sahrens dmu_buf_t *db = dbp[i]; 957fa9e4066Sahrens 958fa9e4066Sahrens ASSERT(size > 0); 959fa9e4066Sahrens 960feb08c6bSbillm bufoff = uio->uio_loffset - db->db_offset; 961fa9e4066Sahrens tocpy = (int)MIN(db->db_size - bufoff, size); 962fa9e4066Sahrens 963fa9e4066Sahrens ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 964fa9e4066Sahrens 965fa9e4066Sahrens if (tocpy == db->db_size) 966fa9e4066Sahrens dmu_buf_will_fill(db, tx); 967fa9e4066Sahrens else 968fa9e4066Sahrens dmu_buf_will_dirty(db, tx); 969fa9e4066Sahrens 970fa9e4066Sahrens /* 971fa9e4066Sahrens * XXX uiomove could block forever (eg. nfs-backed 972fa9e4066Sahrens * pages). There needs to be a uiolockdown() function 973fa9e4066Sahrens * to lock the pages in memory, so that uiomove won't 974fa9e4066Sahrens * block. 975fa9e4066Sahrens */ 976fa9e4066Sahrens err = uiomove((char *)db->db_data + bufoff, tocpy, 977fa9e4066Sahrens UIO_WRITE, uio); 978fa9e4066Sahrens 979fa9e4066Sahrens if (tocpy == db->db_size) 980fa9e4066Sahrens dmu_buf_fill_done(db, tx); 981fa9e4066Sahrens 982fa9e4066Sahrens if (err) 983fa9e4066Sahrens break; 984fa9e4066Sahrens 985fa9e4066Sahrens size -= tocpy; 986fa9e4066Sahrens } 98794d1a210STim Haley 988ea8dc4b6Seschrock dmu_buf_rele_array(dbp, numbufs, FTAG); 989fa9e4066Sahrens return (err); 990fa9e4066Sahrens } 99144eda4d7Smaybee 99244eda4d7Smaybee int 99394d1a210STim Haley dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 99494d1a210STim Haley dmu_tx_t *tx) 99594d1a210STim Haley { 99694d1a210STim Haley if (size == 0) 99794d1a210STim Haley return (0); 99894d1a210STim Haley 99994d1a210STim Haley return (dmu_write_uio_dnode(((dmu_buf_impl_t *)zdb)->db_dnode, 100094d1a210STim Haley uio, size, tx)); 100194d1a210STim Haley } 100294d1a210STim Haley 100394d1a210STim Haley int 100494d1a210STim Haley dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 100594d1a210STim Haley dmu_tx_t *tx) 100694d1a210STim Haley { 100794d1a210STim Haley dnode_t *dn; 100894d1a210STim Haley int err; 100994d1a210STim Haley 101094d1a210STim Haley if (size == 0) 101194d1a210STim Haley return (0); 101294d1a210STim Haley 101394d1a210STim Haley err = dnode_hold(os, object, FTAG, &dn); 101494d1a210STim Haley if (err) 101594d1a210STim Haley return (err); 101694d1a210STim Haley 101794d1a210STim Haley err = dmu_write_uio_dnode(dn, uio, size, tx); 101894d1a210STim Haley 101994d1a210STim Haley dnode_rele(dn, FTAG); 102094d1a210STim Haley 102194d1a210STim Haley return (err); 102294d1a210STim Haley } 102394d1a210STim Haley 102494d1a210STim Haley int 102544eda4d7Smaybee dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 102644eda4d7Smaybee page_t *pp, dmu_tx_t *tx) 102744eda4d7Smaybee { 102844eda4d7Smaybee dmu_buf_t **dbp; 102944eda4d7Smaybee int numbufs, i; 103044eda4d7Smaybee int err; 103144eda4d7Smaybee 103244eda4d7Smaybee if (size == 0) 103344eda4d7Smaybee return (0); 103444eda4d7Smaybee 103544eda4d7Smaybee err = dmu_buf_hold_array(os, object, offset, size, 103644eda4d7Smaybee FALSE, FTAG, &numbufs, &dbp); 103744eda4d7Smaybee if (err) 103844eda4d7Smaybee return (err); 103944eda4d7Smaybee 104044eda4d7Smaybee for (i = 0; i < numbufs; i++) { 104144eda4d7Smaybee int tocpy, copied, thiscpy; 104244eda4d7Smaybee int bufoff; 104344eda4d7Smaybee dmu_buf_t *db = dbp[i]; 104444eda4d7Smaybee caddr_t va; 104544eda4d7Smaybee 104644eda4d7Smaybee ASSERT(size > 0); 104744eda4d7Smaybee ASSERT3U(db->db_size, >=, PAGESIZE); 104844eda4d7Smaybee 104944eda4d7Smaybee bufoff = offset - db->db_offset; 105044eda4d7Smaybee tocpy = (int)MIN(db->db_size - bufoff, size); 105144eda4d7Smaybee 105244eda4d7Smaybee ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 105344eda4d7Smaybee 105444eda4d7Smaybee if (tocpy == db->db_size) 105544eda4d7Smaybee dmu_buf_will_fill(db, tx); 105644eda4d7Smaybee else 105744eda4d7Smaybee dmu_buf_will_dirty(db, tx); 105844eda4d7Smaybee 105944eda4d7Smaybee for (copied = 0; copied < tocpy; copied += PAGESIZE) { 106044eda4d7Smaybee ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 106144eda4d7Smaybee thiscpy = MIN(PAGESIZE, tocpy - copied); 10620fab61baSJonathan W Adams va = zfs_map_page(pp, S_READ); 106344eda4d7Smaybee bcopy(va, (char *)db->db_data + bufoff, thiscpy); 10640fab61baSJonathan W Adams zfs_unmap_page(pp, va); 106544eda4d7Smaybee pp = pp->p_next; 106644eda4d7Smaybee bufoff += PAGESIZE; 106744eda4d7Smaybee } 106844eda4d7Smaybee 106944eda4d7Smaybee if (tocpy == db->db_size) 107044eda4d7Smaybee dmu_buf_fill_done(db, tx); 107144eda4d7Smaybee 107244eda4d7Smaybee offset += tocpy; 107344eda4d7Smaybee size -= tocpy; 107444eda4d7Smaybee } 107544eda4d7Smaybee dmu_buf_rele_array(dbp, numbufs, FTAG); 107644eda4d7Smaybee return (err); 107744eda4d7Smaybee } 1078fa9e4066Sahrens #endif 1079fa9e4066Sahrens 10802fdbea25SAleksandr Guzovskiy /* 10812fdbea25SAleksandr Guzovskiy * Allocate a loaned anonymous arc buffer. 10822fdbea25SAleksandr Guzovskiy */ 10832fdbea25SAleksandr Guzovskiy arc_buf_t * 10842fdbea25SAleksandr Guzovskiy dmu_request_arcbuf(dmu_buf_t *handle, int size) 10852fdbea25SAleksandr Guzovskiy { 10862fdbea25SAleksandr Guzovskiy dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode; 10872fdbea25SAleksandr Guzovskiy 10882fdbea25SAleksandr Guzovskiy return (arc_loan_buf(dn->dn_objset->os_spa, size)); 10892fdbea25SAleksandr Guzovskiy } 10902fdbea25SAleksandr Guzovskiy 10912fdbea25SAleksandr Guzovskiy /* 10922fdbea25SAleksandr Guzovskiy * Free a loaned arc buffer. 10932fdbea25SAleksandr Guzovskiy */ 10942fdbea25SAleksandr Guzovskiy void 10952fdbea25SAleksandr Guzovskiy dmu_return_arcbuf(arc_buf_t *buf) 10962fdbea25SAleksandr Guzovskiy { 10972fdbea25SAleksandr Guzovskiy arc_return_buf(buf, FTAG); 10982fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(buf, FTAG) == 1); 10992fdbea25SAleksandr Guzovskiy } 11002fdbea25SAleksandr Guzovskiy 11012fdbea25SAleksandr Guzovskiy /* 11022fdbea25SAleksandr Guzovskiy * When possible directly assign passed loaned arc buffer to a dbuf. 11032fdbea25SAleksandr Guzovskiy * If this is not possible copy the contents of passed arc buf via 11042fdbea25SAleksandr Guzovskiy * dmu_write(). 11052fdbea25SAleksandr Guzovskiy */ 11062fdbea25SAleksandr Guzovskiy void 11072fdbea25SAleksandr Guzovskiy dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 11082fdbea25SAleksandr Guzovskiy dmu_tx_t *tx) 11092fdbea25SAleksandr Guzovskiy { 11102fdbea25SAleksandr Guzovskiy dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode; 11112fdbea25SAleksandr Guzovskiy dmu_buf_impl_t *db; 11122fdbea25SAleksandr Guzovskiy uint32_t blksz = (uint32_t)arc_buf_size(buf); 11132fdbea25SAleksandr Guzovskiy uint64_t blkid; 11142fdbea25SAleksandr Guzovskiy 11152fdbea25SAleksandr Guzovskiy rw_enter(&dn->dn_struct_rwlock, RW_READER); 11162fdbea25SAleksandr Guzovskiy blkid = dbuf_whichblock(dn, offset); 11172fdbea25SAleksandr Guzovskiy VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 11182fdbea25SAleksandr Guzovskiy rw_exit(&dn->dn_struct_rwlock); 11192fdbea25SAleksandr Guzovskiy 11202fdbea25SAleksandr Guzovskiy if (offset == db->db.db_offset && blksz == db->db.db_size) { 11212fdbea25SAleksandr Guzovskiy dbuf_assign_arcbuf(db, buf, tx); 11222fdbea25SAleksandr Guzovskiy dbuf_rele(db, FTAG); 11232fdbea25SAleksandr Guzovskiy } else { 11242fdbea25SAleksandr Guzovskiy dbuf_rele(db, FTAG); 1125503ad85cSMatthew Ahrens dmu_write(dn->dn_objset, dn->dn_object, offset, blksz, 11262fdbea25SAleksandr Guzovskiy buf->b_data, tx); 11272fdbea25SAleksandr Guzovskiy dmu_return_arcbuf(buf); 1128c242f9a0Schunli zhang - Sun Microsystems - Irvine United States XUIOSTAT_BUMP(xuiostat_wbuf_copied); 11292fdbea25SAleksandr Guzovskiy } 11302fdbea25SAleksandr Guzovskiy } 11312fdbea25SAleksandr Guzovskiy 1132c5c6ffa0Smaybee typedef struct { 1133b24ab676SJeff Bonwick dbuf_dirty_record_t *dsa_dr; 1134b24ab676SJeff Bonwick dmu_sync_cb_t *dsa_done; 1135b24ab676SJeff Bonwick zgd_t *dsa_zgd; 1136b24ab676SJeff Bonwick dmu_tx_t *dsa_tx; 1137c717a561Smaybee } dmu_sync_arg_t; 1138c5c6ffa0Smaybee 1139c5c6ffa0Smaybee /* ARGSUSED */ 1140c5c6ffa0Smaybee static void 1141e14bb325SJeff Bonwick dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1142e14bb325SJeff Bonwick { 1143b24ab676SJeff Bonwick dmu_sync_arg_t *dsa = varg; 1144b24ab676SJeff Bonwick dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1145b24ab676SJeff Bonwick dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode; 1146e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 1147975c32a0SNeil Perrin 1148b24ab676SJeff Bonwick if (zio->io_error == 0) { 1149b24ab676SJeff Bonwick if (BP_IS_HOLE(bp)) { 1150b24ab676SJeff Bonwick /* 1151b24ab676SJeff Bonwick * A block of zeros may compress to a hole, but the 1152b24ab676SJeff Bonwick * block size still needs to be known for replay. 1153b24ab676SJeff Bonwick */ 1154b24ab676SJeff Bonwick BP_SET_LSIZE(bp, db->db_size); 1155b24ab676SJeff Bonwick } else { 1156b24ab676SJeff Bonwick ASSERT(BP_GET_TYPE(bp) == dn->dn_type); 1157e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(bp) == 0); 1158e14bb325SJeff Bonwick bp->blk_fill = 1; 1159e14bb325SJeff Bonwick } 1160e14bb325SJeff Bonwick } 1161b24ab676SJeff Bonwick } 1162b24ab676SJeff Bonwick 1163b24ab676SJeff Bonwick static void 1164b24ab676SJeff Bonwick dmu_sync_late_arrival_ready(zio_t *zio) 1165b24ab676SJeff Bonwick { 1166b24ab676SJeff Bonwick dmu_sync_ready(zio, NULL, zio->io_private); 1167b24ab676SJeff Bonwick } 1168e14bb325SJeff Bonwick 1169e14bb325SJeff Bonwick /* ARGSUSED */ 1170e14bb325SJeff Bonwick static void 1171c5c6ffa0Smaybee dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1172c5c6ffa0Smaybee { 1173b24ab676SJeff Bonwick dmu_sync_arg_t *dsa = varg; 1174b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = dsa->dsa_dr; 1175c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 1176c5c6ffa0Smaybee 1177b50a0fe0SNeil Perrin mutex_enter(&db->db_mtx); 1178b50a0fe0SNeil Perrin ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1179b24ab676SJeff Bonwick if (zio->io_error == 0) { 1180b24ab676SJeff Bonwick dr->dt.dl.dr_overridden_by = *zio->io_bp; 1181b24ab676SJeff Bonwick dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1182b24ab676SJeff Bonwick dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1183b50a0fe0SNeil Perrin if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) 1184b50a0fe0SNeil Perrin BP_ZERO(&dr->dt.dl.dr_overridden_by); 1185b24ab676SJeff Bonwick } else { 1186b24ab676SJeff Bonwick dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1187b24ab676SJeff Bonwick } 1188c5c6ffa0Smaybee cv_broadcast(&db->db_changed); 1189b50a0fe0SNeil Perrin mutex_exit(&db->db_mtx); 1190b50a0fe0SNeil Perrin 1191b24ab676SJeff Bonwick dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1192c717a561Smaybee 1193b24ab676SJeff Bonwick kmem_free(dsa, sizeof (*dsa)); 1194b24ab676SJeff Bonwick } 1195b24ab676SJeff Bonwick 1196b24ab676SJeff Bonwick static void 1197b24ab676SJeff Bonwick dmu_sync_late_arrival_done(zio_t *zio) 1198b24ab676SJeff Bonwick { 1199b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 1200b24ab676SJeff Bonwick dmu_sync_arg_t *dsa = zio->io_private; 1201b24ab676SJeff Bonwick 1202b24ab676SJeff Bonwick if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1203b24ab676SJeff Bonwick ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1204b24ab676SJeff Bonwick ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1205b24ab676SJeff Bonwick zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1206b24ab676SJeff Bonwick } 1207b24ab676SJeff Bonwick 1208b24ab676SJeff Bonwick dmu_tx_commit(dsa->dsa_tx); 1209b24ab676SJeff Bonwick 1210b24ab676SJeff Bonwick dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1211b24ab676SJeff Bonwick 1212b24ab676SJeff Bonwick kmem_free(dsa, sizeof (*dsa)); 1213b24ab676SJeff Bonwick } 1214b24ab676SJeff Bonwick 1215b24ab676SJeff Bonwick static int 1216b24ab676SJeff Bonwick dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1217b24ab676SJeff Bonwick zio_prop_t *zp, zbookmark_t *zb) 1218b24ab676SJeff Bonwick { 1219b24ab676SJeff Bonwick dmu_sync_arg_t *dsa; 1220b24ab676SJeff Bonwick dmu_tx_t *tx; 1221b24ab676SJeff Bonwick 1222b24ab676SJeff Bonwick tx = dmu_tx_create(os); 1223b24ab676SJeff Bonwick dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 12246e1f5caaSNeil Perrin if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1225b24ab676SJeff Bonwick dmu_tx_abort(tx); 1226b24ab676SJeff Bonwick return (EIO); /* Make zl_get_data do txg_waited_synced() */ 1227b24ab676SJeff Bonwick } 1228b24ab676SJeff Bonwick 1229b24ab676SJeff Bonwick dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1230b24ab676SJeff Bonwick dsa->dsa_dr = NULL; 1231b24ab676SJeff Bonwick dsa->dsa_done = done; 1232b24ab676SJeff Bonwick dsa->dsa_zgd = zgd; 1233b24ab676SJeff Bonwick dsa->dsa_tx = tx; 1234b24ab676SJeff Bonwick 1235b24ab676SJeff Bonwick zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1236b24ab676SJeff Bonwick zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1237b24ab676SJeff Bonwick dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa, 1238b24ab676SJeff Bonwick ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1239b24ab676SJeff Bonwick 1240b24ab676SJeff Bonwick return (0); 1241c5c6ffa0Smaybee } 1242c5c6ffa0Smaybee 1243fa9e4066Sahrens /* 1244c5c6ffa0Smaybee * Intent log support: sync the block associated with db to disk. 1245c5c6ffa0Smaybee * N.B. and XXX: the caller is responsible for making sure that the 1246c5c6ffa0Smaybee * data isn't changing while dmu_sync() is writing it. 1247fa9e4066Sahrens * 1248fa9e4066Sahrens * Return values: 1249fa9e4066Sahrens * 1250c5c6ffa0Smaybee * EEXIST: this txg has already been synced, so there's nothing to to. 1251fa9e4066Sahrens * The caller should not log the write. 1252fa9e4066Sahrens * 1253fa9e4066Sahrens * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1254fa9e4066Sahrens * The caller should not log the write. 1255fa9e4066Sahrens * 1256c5c6ffa0Smaybee * EALREADY: this block is already in the process of being synced. 1257c5c6ffa0Smaybee * The caller should track its progress (somehow). 1258fa9e4066Sahrens * 1259b24ab676SJeff Bonwick * EIO: could not do the I/O. 1260b24ab676SJeff Bonwick * The caller should do a txg_wait_synced(). 1261fa9e4066Sahrens * 1262b24ab676SJeff Bonwick * 0: the I/O has been initiated. 1263b24ab676SJeff Bonwick * The caller should log this blkptr in the done callback. 1264b24ab676SJeff Bonwick * It is possible that the I/O will fail, in which case 1265b24ab676SJeff Bonwick * the error will be reported to the done callback and 1266b24ab676SJeff Bonwick * propagated to pio from zio_done(). 1267fa9e4066Sahrens */ 1268fa9e4066Sahrens int 1269b24ab676SJeff Bonwick dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1270fa9e4066Sahrens { 1271b24ab676SJeff Bonwick blkptr_t *bp = zgd->zgd_bp; 1272b24ab676SJeff Bonwick dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1273503ad85cSMatthew Ahrens objset_t *os = db->db_objset; 1274b24ab676SJeff Bonwick dsl_dataset_t *ds = os->os_dsl_dataset; 1275c717a561Smaybee dbuf_dirty_record_t *dr; 1276b24ab676SJeff Bonwick dmu_sync_arg_t *dsa; 1277ea8dc4b6Seschrock zbookmark_t zb; 1278b24ab676SJeff Bonwick zio_prop_t zp; 1279fa9e4066Sahrens 1280b24ab676SJeff Bonwick ASSERT(pio != NULL); 1281fa9e4066Sahrens ASSERT(BP_IS_HOLE(bp)); 1282fa9e4066Sahrens ASSERT(txg != 0); 1283fa9e4066Sahrens 1284b24ab676SJeff Bonwick SET_BOOKMARK(&zb, ds->ds_object, 1285b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 1286b24ab676SJeff Bonwick 1287b24ab676SJeff Bonwick dmu_write_policy(os, db->db_dnode, db->db_level, WP_DMU_SYNC, &zp); 1288fa9e4066Sahrens 1289fa9e4066Sahrens /* 1290b24ab676SJeff Bonwick * If we're frozen (running ziltest), we always need to generate a bp. 1291ea8dc4b6Seschrock */ 1292b24ab676SJeff Bonwick if (txg > spa_freeze_txg(os->os_spa)) 1293b24ab676SJeff Bonwick return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1294ea8dc4b6Seschrock 1295ea8dc4b6Seschrock /* 1296b24ab676SJeff Bonwick * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1297b24ab676SJeff Bonwick * and us. If we determine that this txg is not yet syncing, 1298b24ab676SJeff Bonwick * but it begins to sync a moment later, that's OK because the 1299b24ab676SJeff Bonwick * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1300fa9e4066Sahrens */ 1301b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 1302b24ab676SJeff Bonwick 1303b24ab676SJeff Bonwick if (txg <= spa_last_synced_txg(os->os_spa)) { 1304fa9e4066Sahrens /* 1305b24ab676SJeff Bonwick * This txg has already synced. There's nothing to do. 1306fa9e4066Sahrens */ 1307b24ab676SJeff Bonwick mutex_exit(&db->db_mtx); 1308c5c6ffa0Smaybee return (EEXIST); 1309fa9e4066Sahrens } 1310fa9e4066Sahrens 1311b24ab676SJeff Bonwick if (txg <= spa_syncing_txg(os->os_spa)) { 1312c5c6ffa0Smaybee /* 1313b24ab676SJeff Bonwick * This txg is currently syncing, so we can't mess with 1314b24ab676SJeff Bonwick * the dirty record anymore; just write a new log block. 1315c5c6ffa0Smaybee */ 131613506d1eSmaybee mutex_exit(&db->db_mtx); 1317b24ab676SJeff Bonwick return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1318c5c6ffa0Smaybee } 1319c5c6ffa0Smaybee 1320c717a561Smaybee dr = db->db_last_dirty; 1321b24ab676SJeff Bonwick while (dr && dr->dr_txg != txg) 1322c717a561Smaybee dr = dr->dr_next; 1323b24ab676SJeff Bonwick 1324b24ab676SJeff Bonwick if (dr == NULL) { 1325c5c6ffa0Smaybee /* 1326b24ab676SJeff Bonwick * There's no dr for this dbuf, so it must have been freed. 1327fa9e4066Sahrens * There's no need to log writes to freed blocks, so we're done. 1328fa9e4066Sahrens */ 1329fa9e4066Sahrens mutex_exit(&db->db_mtx); 1330fa9e4066Sahrens return (ENOENT); 1331fa9e4066Sahrens } 1332fa9e4066Sahrens 1333c717a561Smaybee ASSERT(dr->dr_txg == txg); 1334b24ab676SJeff Bonwick if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1335b24ab676SJeff Bonwick dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1336c5c6ffa0Smaybee /* 1337b24ab676SJeff Bonwick * We have already issued a sync write for this buffer, 1338b24ab676SJeff Bonwick * or this buffer has already been synced. It could not 1339c717a561Smaybee * have been dirtied since, or we would have cleared the state. 1340c717a561Smaybee */ 1341c717a561Smaybee mutex_exit(&db->db_mtx); 1342b24ab676SJeff Bonwick return (EALREADY); 1343c717a561Smaybee } 1344c717a561Smaybee 1345b24ab676SJeff Bonwick ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1346c717a561Smaybee dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1347fa9e4066Sahrens mutex_exit(&db->db_mtx); 1348fa9e4066Sahrens 1349b24ab676SJeff Bonwick dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1350b24ab676SJeff Bonwick dsa->dsa_dr = dr; 1351b24ab676SJeff Bonwick dsa->dsa_done = done; 1352b24ab676SJeff Bonwick dsa->dsa_zgd = zgd; 1353b24ab676SJeff Bonwick dsa->dsa_tx = NULL; 1354e14bb325SJeff Bonwick 1355b24ab676SJeff Bonwick zio_nowait(arc_write(pio, os->os_spa, txg, 1356b24ab676SJeff Bonwick bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp, 1357b24ab676SJeff Bonwick dmu_sync_ready, dmu_sync_done, dsa, 1358b24ab676SJeff Bonwick ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1359e14bb325SJeff Bonwick 1360b24ab676SJeff Bonwick return (0); 1361fa9e4066Sahrens } 1362fa9e4066Sahrens 1363fa9e4066Sahrens int 1364fa9e4066Sahrens dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1365fa9e4066Sahrens dmu_tx_t *tx) 1366fa9e4066Sahrens { 1367ea8dc4b6Seschrock dnode_t *dn; 1368ea8dc4b6Seschrock int err; 1369ea8dc4b6Seschrock 1370503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 1371ea8dc4b6Seschrock if (err) 1372ea8dc4b6Seschrock return (err); 1373ea8dc4b6Seschrock err = dnode_set_blksz(dn, size, ibs, tx); 1374fa9e4066Sahrens dnode_rele(dn, FTAG); 1375fa9e4066Sahrens return (err); 1376fa9e4066Sahrens } 1377fa9e4066Sahrens 1378fa9e4066Sahrens void 1379fa9e4066Sahrens dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1380fa9e4066Sahrens dmu_tx_t *tx) 1381fa9e4066Sahrens { 1382ea8dc4b6Seschrock dnode_t *dn; 1383ea8dc4b6Seschrock 1384ea8dc4b6Seschrock /* XXX assumes dnode_hold will not get an i/o error */ 1385503ad85cSMatthew Ahrens (void) dnode_hold(os, object, FTAG, &dn); 1386fa9e4066Sahrens ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS); 1387fa9e4066Sahrens dn->dn_checksum = checksum; 1388fa9e4066Sahrens dnode_setdirty(dn, tx); 1389fa9e4066Sahrens dnode_rele(dn, FTAG); 1390fa9e4066Sahrens } 1391fa9e4066Sahrens 1392fa9e4066Sahrens void 1393fa9e4066Sahrens dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1394fa9e4066Sahrens dmu_tx_t *tx) 1395fa9e4066Sahrens { 1396ea8dc4b6Seschrock dnode_t *dn; 1397ea8dc4b6Seschrock 1398ea8dc4b6Seschrock /* XXX assumes dnode_hold will not get an i/o error */ 1399503ad85cSMatthew Ahrens (void) dnode_hold(os, object, FTAG, &dn); 1400fa9e4066Sahrens ASSERT(compress < ZIO_COMPRESS_FUNCTIONS); 1401fa9e4066Sahrens dn->dn_compress = compress; 1402fa9e4066Sahrens dnode_setdirty(dn, tx); 1403fa9e4066Sahrens dnode_rele(dn, FTAG); 1404fa9e4066Sahrens } 1405fa9e4066Sahrens 1406b24ab676SJeff Bonwick int zfs_mdcomp_disable = 0; 1407b24ab676SJeff Bonwick 1408b24ab676SJeff Bonwick void 1409b24ab676SJeff Bonwick dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1410b24ab676SJeff Bonwick { 1411b24ab676SJeff Bonwick dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1412b24ab676SJeff Bonwick boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata); 1413b24ab676SJeff Bonwick enum zio_checksum checksum = os->os_checksum; 1414b24ab676SJeff Bonwick enum zio_compress compress = os->os_compress; 1415b24ab676SJeff Bonwick enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1416b24ab676SJeff Bonwick boolean_t dedup; 1417b24ab676SJeff Bonwick boolean_t dedup_verify = os->os_dedup_verify; 1418b24ab676SJeff Bonwick int copies = os->os_copies; 1419b24ab676SJeff Bonwick 1420b24ab676SJeff Bonwick /* 1421b24ab676SJeff Bonwick * Determine checksum setting. 1422b24ab676SJeff Bonwick */ 1423b24ab676SJeff Bonwick if (ismd) { 1424b24ab676SJeff Bonwick /* 1425b24ab676SJeff Bonwick * Metadata always gets checksummed. If the data 1426b24ab676SJeff Bonwick * checksum is multi-bit correctable, and it's not a 1427b24ab676SJeff Bonwick * ZBT-style checksum, then it's suitable for metadata 1428b24ab676SJeff Bonwick * as well. Otherwise, the metadata checksum defaults 1429b24ab676SJeff Bonwick * to fletcher4. 1430b24ab676SJeff Bonwick */ 1431b24ab676SJeff Bonwick if (zio_checksum_table[checksum].ci_correctable < 1 || 14326e1f5caaSNeil Perrin zio_checksum_table[checksum].ci_eck) 1433b24ab676SJeff Bonwick checksum = ZIO_CHECKSUM_FLETCHER_4; 1434b24ab676SJeff Bonwick } else { 1435b24ab676SJeff Bonwick checksum = zio_checksum_select(dn->dn_checksum, checksum); 1436b24ab676SJeff Bonwick } 1437b24ab676SJeff Bonwick 1438b24ab676SJeff Bonwick /* 1439b24ab676SJeff Bonwick * Determine compression setting. 1440b24ab676SJeff Bonwick */ 1441b24ab676SJeff Bonwick if (ismd) { 1442b24ab676SJeff Bonwick /* 1443b24ab676SJeff Bonwick * XXX -- we should design a compression algorithm 1444b24ab676SJeff Bonwick * that specializes in arrays of bps. 1445b24ab676SJeff Bonwick */ 1446b24ab676SJeff Bonwick compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 1447b24ab676SJeff Bonwick ZIO_COMPRESS_LZJB; 1448b24ab676SJeff Bonwick } else { 1449b24ab676SJeff Bonwick compress = zio_compress_select(dn->dn_compress, compress); 1450b24ab676SJeff Bonwick } 1451b24ab676SJeff Bonwick 1452b24ab676SJeff Bonwick /* 1453b24ab676SJeff Bonwick * Determine dedup setting. If we are in dmu_sync(), we won't 1454b24ab676SJeff Bonwick * actually dedup now because that's all done in syncing context; 1455b24ab676SJeff Bonwick * but we do want to use the dedup checkum. If the checksum is not 1456b24ab676SJeff Bonwick * strong enough to ensure unique signatures, force dedup_verify. 1457b24ab676SJeff Bonwick */ 1458b24ab676SJeff Bonwick dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF); 1459b24ab676SJeff Bonwick if (dedup) { 1460b24ab676SJeff Bonwick checksum = dedup_checksum; 1461b24ab676SJeff Bonwick if (!zio_checksum_table[checksum].ci_dedup) 1462b24ab676SJeff Bonwick dedup_verify = 1; 1463b24ab676SJeff Bonwick } 1464b24ab676SJeff Bonwick 1465b24ab676SJeff Bonwick if (wp & WP_DMU_SYNC) 1466b24ab676SJeff Bonwick dedup = 0; 1467b24ab676SJeff Bonwick 1468b24ab676SJeff Bonwick if (wp & WP_NOFILL) { 1469b24ab676SJeff Bonwick ASSERT(!ismd && level == 0); 1470b24ab676SJeff Bonwick checksum = ZIO_CHECKSUM_OFF; 1471b24ab676SJeff Bonwick compress = ZIO_COMPRESS_OFF; 1472b24ab676SJeff Bonwick dedup = B_FALSE; 1473b24ab676SJeff Bonwick } 1474b24ab676SJeff Bonwick 1475b24ab676SJeff Bonwick zp->zp_checksum = checksum; 1476b24ab676SJeff Bonwick zp->zp_compress = compress; 14770a586ceaSMark Shellenbaum zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1478b24ab676SJeff Bonwick zp->zp_level = level; 1479b24ab676SJeff Bonwick zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa)); 1480b24ab676SJeff Bonwick zp->zp_dedup = dedup; 1481b24ab676SJeff Bonwick zp->zp_dedup_verify = dedup && dedup_verify; 1482b24ab676SJeff Bonwick } 1483b24ab676SJeff Bonwick 148444cd46caSbillm int 1485fa9e4066Sahrens dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1486fa9e4066Sahrens { 1487fa9e4066Sahrens dnode_t *dn; 1488fa9e4066Sahrens int i, err; 1489fa9e4066Sahrens 1490503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 1491ea8dc4b6Seschrock if (err) 1492ea8dc4b6Seschrock return (err); 1493fa9e4066Sahrens /* 1494fa9e4066Sahrens * Sync any current changes before 1495fa9e4066Sahrens * we go trundling through the block pointers. 1496fa9e4066Sahrens */ 1497fa9e4066Sahrens for (i = 0; i < TXG_SIZE; i++) { 1498c543ec06Sahrens if (list_link_active(&dn->dn_dirty_link[i])) 1499fa9e4066Sahrens break; 1500fa9e4066Sahrens } 1501fa9e4066Sahrens if (i != TXG_SIZE) { 1502fa9e4066Sahrens dnode_rele(dn, FTAG); 1503fa9e4066Sahrens txg_wait_synced(dmu_objset_pool(os), 0); 1504503ad85cSMatthew Ahrens err = dnode_hold(os, object, FTAG, &dn); 1505ea8dc4b6Seschrock if (err) 1506ea8dc4b6Seschrock return (err); 1507fa9e4066Sahrens } 1508fa9e4066Sahrens 1509cdb0ab79Smaybee err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1510fa9e4066Sahrens dnode_rele(dn, FTAG); 1511fa9e4066Sahrens 1512fa9e4066Sahrens return (err); 1513fa9e4066Sahrens } 1514fa9e4066Sahrens 1515fa9e4066Sahrens void 1516fa9e4066Sahrens dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1517fa9e4066Sahrens { 1518b24ab676SJeff Bonwick dnode_phys_t *dnp; 1519b24ab676SJeff Bonwick 1520fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 1521fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1522fa9e4066Sahrens 1523b24ab676SJeff Bonwick dnp = dn->dn_phys; 1524b24ab676SJeff Bonwick 1525fa9e4066Sahrens doi->doi_data_block_size = dn->dn_datablksz; 1526fa9e4066Sahrens doi->doi_metadata_block_size = dn->dn_indblkshift ? 1527fa9e4066Sahrens 1ULL << dn->dn_indblkshift : 0; 1528b24ab676SJeff Bonwick doi->doi_type = dn->dn_type; 1529b24ab676SJeff Bonwick doi->doi_bonus_type = dn->dn_bonustype; 1530b24ab676SJeff Bonwick doi->doi_bonus_size = dn->dn_bonuslen; 1531fa9e4066Sahrens doi->doi_indirection = dn->dn_nlevels; 1532fa9e4066Sahrens doi->doi_checksum = dn->dn_checksum; 1533fa9e4066Sahrens doi->doi_compress = dn->dn_compress; 1534b24ab676SJeff Bonwick doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 1535b24ab676SJeff Bonwick doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz; 1536b24ab676SJeff Bonwick doi->doi_fill_count = 0; 1537b24ab676SJeff Bonwick for (int i = 0; i < dnp->dn_nblkptr; i++) 1538b24ab676SJeff Bonwick doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill; 1539fa9e4066Sahrens 1540fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1541fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1542fa9e4066Sahrens } 1543fa9e4066Sahrens 1544fa9e4066Sahrens /* 1545fa9e4066Sahrens * Get information on a DMU object. 1546fa9e4066Sahrens * If doi is NULL, just indicates whether the object exists. 1547fa9e4066Sahrens */ 1548fa9e4066Sahrens int 1549fa9e4066Sahrens dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 1550fa9e4066Sahrens { 1551ea8dc4b6Seschrock dnode_t *dn; 1552503ad85cSMatthew Ahrens int err = dnode_hold(os, object, FTAG, &dn); 1553fa9e4066Sahrens 1554ea8dc4b6Seschrock if (err) 1555ea8dc4b6Seschrock return (err); 1556fa9e4066Sahrens 1557fa9e4066Sahrens if (doi != NULL) 1558fa9e4066Sahrens dmu_object_info_from_dnode(dn, doi); 1559fa9e4066Sahrens 1560fa9e4066Sahrens dnode_rele(dn, FTAG); 1561fa9e4066Sahrens return (0); 1562fa9e4066Sahrens } 1563fa9e4066Sahrens 1564fa9e4066Sahrens /* 1565fa9e4066Sahrens * As above, but faster; can be used when you have a held dbuf in hand. 1566fa9e4066Sahrens */ 1567fa9e4066Sahrens void 1568fa9e4066Sahrens dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi) 1569fa9e4066Sahrens { 1570fa9e4066Sahrens dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi); 1571fa9e4066Sahrens } 1572fa9e4066Sahrens 1573fa9e4066Sahrens /* 1574fa9e4066Sahrens * Faster still when you only care about the size. 1575fa9e4066Sahrens * This is specifically optimized for zfs_getattr(). 1576fa9e4066Sahrens */ 1577fa9e4066Sahrens void 1578fa9e4066Sahrens dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512) 1579fa9e4066Sahrens { 1580fa9e4066Sahrens dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode; 1581fa9e4066Sahrens 1582fa9e4066Sahrens *blksize = dn->dn_datablksz; 158399653d4eSeschrock /* add 1 for dnode space */ 158499653d4eSeschrock *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 158599653d4eSeschrock SPA_MINBLOCKSHIFT) + 1; 1586fa9e4066Sahrens } 1587fa9e4066Sahrens 1588fa9e4066Sahrens void 1589fa9e4066Sahrens byteswap_uint64_array(void *vbuf, size_t size) 1590fa9e4066Sahrens { 1591fa9e4066Sahrens uint64_t *buf = vbuf; 1592fa9e4066Sahrens size_t count = size >> 3; 1593fa9e4066Sahrens int i; 1594fa9e4066Sahrens 1595fa9e4066Sahrens ASSERT((size & 7) == 0); 1596fa9e4066Sahrens 1597fa9e4066Sahrens for (i = 0; i < count; i++) 1598fa9e4066Sahrens buf[i] = BSWAP_64(buf[i]); 1599fa9e4066Sahrens } 1600fa9e4066Sahrens 1601fa9e4066Sahrens void 1602fa9e4066Sahrens byteswap_uint32_array(void *vbuf, size_t size) 1603fa9e4066Sahrens { 1604fa9e4066Sahrens uint32_t *buf = vbuf; 1605fa9e4066Sahrens size_t count = size >> 2; 1606fa9e4066Sahrens int i; 1607fa9e4066Sahrens 1608fa9e4066Sahrens ASSERT((size & 3) == 0); 1609fa9e4066Sahrens 1610fa9e4066Sahrens for (i = 0; i < count; i++) 1611fa9e4066Sahrens buf[i] = BSWAP_32(buf[i]); 1612fa9e4066Sahrens } 1613fa9e4066Sahrens 1614fa9e4066Sahrens void 1615fa9e4066Sahrens byteswap_uint16_array(void *vbuf, size_t size) 1616fa9e4066Sahrens { 1617fa9e4066Sahrens uint16_t *buf = vbuf; 1618fa9e4066Sahrens size_t count = size >> 1; 1619fa9e4066Sahrens int i; 1620fa9e4066Sahrens 1621fa9e4066Sahrens ASSERT((size & 1) == 0); 1622fa9e4066Sahrens 1623fa9e4066Sahrens for (i = 0; i < count; i++) 1624fa9e4066Sahrens buf[i] = BSWAP_16(buf[i]); 1625fa9e4066Sahrens } 1626fa9e4066Sahrens 1627fa9e4066Sahrens /* ARGSUSED */ 1628fa9e4066Sahrens void 1629fa9e4066Sahrens byteswap_uint8_array(void *vbuf, size_t size) 1630fa9e4066Sahrens { 1631fa9e4066Sahrens } 1632fa9e4066Sahrens 1633fa9e4066Sahrens void 1634fa9e4066Sahrens dmu_init(void) 1635fa9e4066Sahrens { 1636*3f9d6ad7SLin Ling zfs_dbgmsg_init(); 1637fa9e4066Sahrens dbuf_init(); 1638fa9e4066Sahrens dnode_init(); 16397cbf8b43SRich Morris zfetch_init(); 1640fa9e4066Sahrens arc_init(); 1641fa94a07fSbrendan l2arc_init(); 1642c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_init(); 16430a586ceaSMark Shellenbaum sa_cache_init(); 1644fa9e4066Sahrens } 1645fa9e4066Sahrens 1646fa9e4066Sahrens void 1647fa9e4066Sahrens dmu_fini(void) 1648fa9e4066Sahrens { 1649fa9e4066Sahrens arc_fini(); 16507cbf8b43SRich Morris zfetch_fini(); 1651fa9e4066Sahrens dnode_fini(); 1652fa9e4066Sahrens dbuf_fini(); 1653fa94a07fSbrendan l2arc_fini(); 1654c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_fini(); 16550a586ceaSMark Shellenbaum sa_cache_fini(); 1656*3f9d6ad7SLin Ling zfs_dbgmsg_fini(); 1657fa9e4066Sahrens } 1658