dmu_tx.c revision 240133
1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22219089Spjd * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23226512Smm * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24236884Smm * Copyright (c) 2012 by Delphix. All rights reserved. 25226512Smm */ 26168404Spjd 27168404Spjd#include <sys/dmu.h> 28168404Spjd#include <sys/dmu_impl.h> 29168404Spjd#include <sys/dbuf.h> 30168404Spjd#include <sys/dmu_tx.h> 31168404Spjd#include <sys/dmu_objset.h> 32168404Spjd#include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */ 33168404Spjd#include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */ 34168404Spjd#include <sys/dsl_pool.h> 35168404Spjd#include <sys/zap_impl.h> /* for fzap_default_block_shift */ 36168404Spjd#include <sys/spa.h> 37219089Spjd#include <sys/sa.h> 38219089Spjd#include <sys/sa_impl.h> 39168404Spjd#include <sys/zfs_context.h> 40219089Spjd#include <sys/varargs.h> 41168404Spjd 42168404Spjdtypedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn, 43168404Spjd uint64_t arg1, uint64_t arg2); 44168404Spjd 45168404Spjd 46168404Spjddmu_tx_t * 47168404Spjddmu_tx_create_dd(dsl_dir_t *dd) 48168404Spjd{ 49168404Spjd dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP); 50168404Spjd tx->tx_dir = dd; 51168404Spjd if (dd) 52168404Spjd tx->tx_pool = dd->dd_pool; 53168404Spjd list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t), 54168404Spjd offsetof(dmu_tx_hold_t, txh_node)); 55219089Spjd list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t), 56219089Spjd offsetof(dmu_tx_callback_t, dcb_node)); 57168404Spjd#ifdef ZFS_DEBUG 58168404Spjd refcount_create(&tx->tx_space_written); 59168404Spjd refcount_create(&tx->tx_space_freed); 60168404Spjd#endif 61168404Spjd return (tx); 62168404Spjd} 63168404Spjd 64168404Spjddmu_tx_t * 65168404Spjddmu_tx_create(objset_t *os) 66168404Spjd{ 67219089Spjd dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir); 68168404Spjd tx->tx_objset = os; 69219089Spjd tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset); 70168404Spjd return (tx); 71168404Spjd} 72168404Spjd 73168404Spjddmu_tx_t * 74168404Spjddmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg) 75168404Spjd{ 76168404Spjd dmu_tx_t *tx = dmu_tx_create_dd(NULL); 77168404Spjd 78168404Spjd ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg); 79168404Spjd tx->tx_pool = dp; 80168404Spjd tx->tx_txg = txg; 81168404Spjd tx->tx_anyobj = TRUE; 82168404Spjd 83168404Spjd return (tx); 84168404Spjd} 85168404Spjd 86168404Spjdint 87168404Spjddmu_tx_is_syncing(dmu_tx_t *tx) 88168404Spjd{ 89168404Spjd return (tx->tx_anyobj); 90168404Spjd} 91168404Spjd 92168404Spjdint 93168404Spjddmu_tx_private_ok(dmu_tx_t *tx) 94168404Spjd{ 95168404Spjd return (tx->tx_anyobj); 96168404Spjd} 97168404Spjd 98168404Spjdstatic dmu_tx_hold_t * 99168404Spjddmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object, 100168404Spjd enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2) 101168404Spjd{ 102168404Spjd dmu_tx_hold_t *txh; 103168404Spjd dnode_t *dn = NULL; 104168404Spjd int err; 105168404Spjd 106168404Spjd if (object != DMU_NEW_OBJECT) { 107219089Spjd err = dnode_hold(os, object, tx, &dn); 108168404Spjd if (err) { 109168404Spjd tx->tx_err = err; 110168404Spjd return (NULL); 111168404Spjd } 112168404Spjd 113168404Spjd if (err == 0 && tx->tx_txg != 0) { 114168404Spjd mutex_enter(&dn->dn_mtx); 115168404Spjd /* 116168404Spjd * dn->dn_assigned_txg == tx->tx_txg doesn't pose a 117168404Spjd * problem, but there's no way for it to happen (for 118168404Spjd * now, at least). 119168404Spjd */ 120168404Spjd ASSERT(dn->dn_assigned_txg == 0); 121168404Spjd dn->dn_assigned_txg = tx->tx_txg; 122168404Spjd (void) refcount_add(&dn->dn_tx_holds, tx); 123168404Spjd mutex_exit(&dn->dn_mtx); 124168404Spjd } 125168404Spjd } 126168404Spjd 127168404Spjd txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP); 128168404Spjd txh->txh_tx = tx; 129168404Spjd txh->txh_dnode = dn; 130168404Spjd#ifdef ZFS_DEBUG 131168404Spjd txh->txh_type = type; 132168404Spjd txh->txh_arg1 = arg1; 133168404Spjd txh->txh_arg2 = arg2; 134168404Spjd#endif 135168404Spjd list_insert_tail(&tx->tx_holds, txh); 136168404Spjd 137168404Spjd return (txh); 138168404Spjd} 139168404Spjd 140168404Spjdvoid 141168404Spjddmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object) 142168404Spjd{ 143168404Spjd /* 144168404Spjd * If we're syncing, they can manipulate any object anyhow, and 145168404Spjd * the hold on the dnode_t can cause problems. 146168404Spjd */ 147168404Spjd if (!dmu_tx_is_syncing(tx)) { 148168404Spjd (void) dmu_tx_hold_object_impl(tx, os, 149168404Spjd object, THT_NEWOBJECT, 0, 0); 150168404Spjd } 151168404Spjd} 152168404Spjd 153168404Spjdstatic int 154168404Spjddmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid) 155168404Spjd{ 156168404Spjd int err; 157168404Spjd dmu_buf_impl_t *db; 158168404Spjd 159168404Spjd rw_enter(&dn->dn_struct_rwlock, RW_READER); 160168404Spjd db = dbuf_hold_level(dn, level, blkid, FTAG); 161168404Spjd rw_exit(&dn->dn_struct_rwlock); 162168404Spjd if (db == NULL) 163168404Spjd return (EIO); 164185029Spjd err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH); 165168404Spjd dbuf_rele(db, FTAG); 166168404Spjd return (err); 167168404Spjd} 168168404Spjd 169209962Smmstatic void 170219089Spjddmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db, 171219089Spjd int level, uint64_t blkid, boolean_t freeable, uint64_t *history) 172209962Smm{ 173219089Spjd objset_t *os = dn->dn_objset; 174219089Spjd dsl_dataset_t *ds = os->os_dsl_dataset; 175219089Spjd int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 176219089Spjd dmu_buf_impl_t *parent = NULL; 177219089Spjd blkptr_t *bp = NULL; 178219089Spjd uint64_t space; 179209962Smm 180219089Spjd if (level >= dn->dn_nlevels || history[level] == blkid) 181209962Smm return; 182209962Smm 183219089Spjd history[level] = blkid; 184209962Smm 185219089Spjd space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift); 186219089Spjd 187219089Spjd if (db == NULL || db == dn->dn_dbuf) { 188219089Spjd ASSERT(level != 0); 189219089Spjd db = NULL; 190219089Spjd } else { 191219089Spjd ASSERT(DB_DNODE(db) == dn); 192219089Spjd ASSERT(db->db_level == level); 193219089Spjd ASSERT(db->db.db_size == space); 194219089Spjd ASSERT(db->db_blkid == blkid); 195219089Spjd bp = db->db_blkptr; 196219089Spjd parent = db->db_parent; 197209962Smm } 198209962Smm 199219089Spjd freeable = (bp && (freeable || 200219089Spjd dsl_dataset_block_freeable(ds, bp, bp->blk_birth))); 201209962Smm 202219089Spjd if (freeable) 203219089Spjd txh->txh_space_tooverwrite += space; 204219089Spjd else 205219089Spjd txh->txh_space_towrite += space; 206219089Spjd if (bp) 207219089Spjd txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp); 208219089Spjd 209219089Spjd dmu_tx_count_twig(txh, dn, parent, level + 1, 210219089Spjd blkid >> epbs, freeable, history); 211209962Smm} 212209962Smm 213168404Spjd/* ARGSUSED */ 214168404Spjdstatic void 215168404Spjddmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 216168404Spjd{ 217168404Spjd dnode_t *dn = txh->txh_dnode; 218168404Spjd uint64_t start, end, i; 219168404Spjd int min_bs, max_bs, min_ibs, max_ibs, epbs, bits; 220168404Spjd int err = 0; 221168404Spjd 222168404Spjd if (len == 0) 223168404Spjd return; 224168404Spjd 225168404Spjd min_bs = SPA_MINBLOCKSHIFT; 226168404Spjd max_bs = SPA_MAXBLOCKSHIFT; 227168404Spjd min_ibs = DN_MIN_INDBLKSHIFT; 228168404Spjd max_ibs = DN_MAX_INDBLKSHIFT; 229168404Spjd 230209962Smm if (dn) { 231219089Spjd uint64_t history[DN_MAX_LEVELS]; 232209962Smm int nlvls = dn->dn_nlevels; 233209962Smm int delta; 234168404Spjd 235209962Smm /* 236209962Smm * For i/o error checking, read the first and last level-0 237209962Smm * blocks (if they are not aligned), and all the level-1 blocks. 238209962Smm */ 239168404Spjd if (dn->dn_maxblkid == 0) { 240209962Smm delta = dn->dn_datablksz; 241209962Smm start = (off < dn->dn_datablksz) ? 0 : 1; 242209962Smm end = (off+len <= dn->dn_datablksz) ? 0 : 1; 243209962Smm if (start == 0 && (off > 0 || len < dn->dn_datablksz)) { 244209962Smm err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 245209962Smm if (err) 246209962Smm goto out; 247209962Smm delta -= off; 248209962Smm } 249168404Spjd } else { 250168404Spjd zio_t *zio = zio_root(dn->dn_objset->os_spa, 251168404Spjd NULL, NULL, ZIO_FLAG_CANFAIL); 252168404Spjd 253168404Spjd /* first level-0 block */ 254168404Spjd start = off >> dn->dn_datablkshift; 255168404Spjd if (P2PHASE(off, dn->dn_datablksz) || 256168404Spjd len < dn->dn_datablksz) { 257168404Spjd err = dmu_tx_check_ioerr(zio, dn, 0, start); 258168404Spjd if (err) 259168404Spjd goto out; 260168404Spjd } 261168404Spjd 262168404Spjd /* last level-0 block */ 263168404Spjd end = (off+len-1) >> dn->dn_datablkshift; 264219089Spjd if (end != start && end <= dn->dn_maxblkid && 265168404Spjd P2PHASE(off+len, dn->dn_datablksz)) { 266168404Spjd err = dmu_tx_check_ioerr(zio, dn, 0, end); 267168404Spjd if (err) 268168404Spjd goto out; 269168404Spjd } 270168404Spjd 271168404Spjd /* level-1 blocks */ 272209962Smm if (nlvls > 1) { 273209962Smm int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 274209962Smm for (i = (start>>shft)+1; i < end>>shft; i++) { 275168404Spjd err = dmu_tx_check_ioerr(zio, dn, 1, i); 276168404Spjd if (err) 277168404Spjd goto out; 278168404Spjd } 279168404Spjd } 280168404Spjd 281168404Spjd err = zio_wait(zio); 282168404Spjd if (err) 283168404Spjd goto out; 284209962Smm delta = P2NPHASE(off, dn->dn_datablksz); 285168404Spjd } 286168404Spjd 287209962Smm if (dn->dn_maxblkid > 0) { 288209962Smm /* 289209962Smm * The blocksize can't change, 290209962Smm * so we can make a more precise estimate. 291209962Smm */ 292209962Smm ASSERT(dn->dn_datablkshift != 0); 293168404Spjd min_bs = max_bs = dn->dn_datablkshift; 294209962Smm min_ibs = max_ibs = dn->dn_indblkshift; 295209962Smm } else if (dn->dn_indblkshift > max_ibs) { 296209962Smm /* 297209962Smm * This ensures that if we reduce DN_MAX_INDBLKSHIFT, 298209962Smm * the code will still work correctly on older pools. 299209962Smm */ 300209962Smm min_ibs = max_ibs = dn->dn_indblkshift; 301209962Smm } 302209962Smm 303209962Smm /* 304209962Smm * If this write is not off the end of the file 305209962Smm * we need to account for overwrites/unref. 306209962Smm */ 307219089Spjd if (start <= dn->dn_maxblkid) { 308219089Spjd for (int l = 0; l < DN_MAX_LEVELS; l++) 309219089Spjd history[l] = -1ULL; 310219089Spjd } 311209962Smm while (start <= dn->dn_maxblkid) { 312209962Smm dmu_buf_impl_t *db; 313209962Smm 314209962Smm rw_enter(&dn->dn_struct_rwlock, RW_READER); 315219089Spjd err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db); 316209962Smm rw_exit(&dn->dn_struct_rwlock); 317219089Spjd 318219089Spjd if (err) { 319219089Spjd txh->txh_tx->tx_err = err; 320219089Spjd return; 321209962Smm } 322219089Spjd 323219089Spjd dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE, 324219089Spjd history); 325209962Smm dbuf_rele(db, FTAG); 326209962Smm if (++start > end) { 327209962Smm /* 328209962Smm * Account for new indirects appearing 329209962Smm * before this IO gets assigned into a txg. 330209962Smm */ 331209962Smm bits = 64 - min_bs; 332209962Smm epbs = min_ibs - SPA_BLKPTRSHIFT; 333209962Smm for (bits -= epbs * (nlvls - 1); 334209962Smm bits >= 0; bits -= epbs) 335209962Smm txh->txh_fudge += 1ULL << max_ibs; 336209962Smm goto out; 337209962Smm } 338209962Smm off += delta; 339209962Smm if (len >= delta) 340209962Smm len -= delta; 341209962Smm delta = dn->dn_datablksz; 342209962Smm } 343168404Spjd } 344168404Spjd 345168404Spjd /* 346168404Spjd * 'end' is the last thing we will access, not one past. 347168404Spjd * This way we won't overflow when accessing the last byte. 348168404Spjd */ 349168404Spjd start = P2ALIGN(off, 1ULL << max_bs); 350168404Spjd end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1; 351168404Spjd txh->txh_space_towrite += end - start + 1; 352168404Spjd 353168404Spjd start >>= min_bs; 354168404Spjd end >>= min_bs; 355168404Spjd 356168404Spjd epbs = min_ibs - SPA_BLKPTRSHIFT; 357168404Spjd 358168404Spjd /* 359168404Spjd * The object contains at most 2^(64 - min_bs) blocks, 360168404Spjd * and each indirect level maps 2^epbs. 361168404Spjd */ 362168404Spjd for (bits = 64 - min_bs; bits >= 0; bits -= epbs) { 363168404Spjd start >>= epbs; 364168404Spjd end >>= epbs; 365209962Smm ASSERT3U(end, >=, start); 366209962Smm txh->txh_space_towrite += (end - start + 1) << max_ibs; 367209962Smm if (start != 0) { 368209962Smm /* 369209962Smm * We also need a new blkid=0 indirect block 370209962Smm * to reference any existing file data. 371209962Smm */ 372168404Spjd txh->txh_space_towrite += 1ULL << max_ibs; 373209962Smm } 374168404Spjd } 375168404Spjd 376209962Smmout: 377209962Smm if (txh->txh_space_towrite + txh->txh_space_tooverwrite > 378209962Smm 2 * DMU_MAX_ACCESS) 379209962Smm err = EFBIG; 380168404Spjd 381168404Spjd if (err) 382168404Spjd txh->txh_tx->tx_err = err; 383168404Spjd} 384168404Spjd 385168404Spjdstatic void 386168404Spjddmu_tx_count_dnode(dmu_tx_hold_t *txh) 387168404Spjd{ 388168404Spjd dnode_t *dn = txh->txh_dnode; 389219089Spjd dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset); 390168404Spjd uint64_t space = mdn->dn_datablksz + 391168404Spjd ((mdn->dn_nlevels-1) << mdn->dn_indblkshift); 392168404Spjd 393168404Spjd if (dn && dn->dn_dbuf->db_blkptr && 394168404Spjd dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, 395219089Spjd dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) { 396168404Spjd txh->txh_space_tooverwrite += space; 397209962Smm txh->txh_space_tounref += space; 398168404Spjd } else { 399168404Spjd txh->txh_space_towrite += space; 400185029Spjd if (dn && dn->dn_dbuf->db_blkptr) 401185029Spjd txh->txh_space_tounref += space; 402168404Spjd } 403168404Spjd} 404168404Spjd 405168404Spjdvoid 406168404Spjddmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len) 407168404Spjd{ 408168404Spjd dmu_tx_hold_t *txh; 409168404Spjd 410168404Spjd ASSERT(tx->tx_txg == 0); 411168404Spjd ASSERT(len < DMU_MAX_ACCESS); 412168404Spjd ASSERT(len == 0 || UINT64_MAX - off >= len - 1); 413168404Spjd 414168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 415168404Spjd object, THT_WRITE, off, len); 416168404Spjd if (txh == NULL) 417168404Spjd return; 418168404Spjd 419168404Spjd dmu_tx_count_write(txh, off, len); 420168404Spjd dmu_tx_count_dnode(txh); 421168404Spjd} 422168404Spjd 423168404Spjdstatic void 424168404Spjddmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 425168404Spjd{ 426185029Spjd uint64_t blkid, nblks, lastblk; 427185029Spjd uint64_t space = 0, unref = 0, skipped = 0; 428168404Spjd dnode_t *dn = txh->txh_dnode; 429168404Spjd dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 430168404Spjd spa_t *spa = txh->txh_tx->tx_pool->dp_spa; 431185029Spjd int epbs; 432240133Smm uint64_t l0span = 0, nl1blks = 0; 433168404Spjd 434185029Spjd if (dn->dn_nlevels == 0) 435168404Spjd return; 436168404Spjd 437168404Spjd /* 438185029Spjd * The struct_rwlock protects us against dn_nlevels 439168404Spjd * changing, in case (against all odds) we manage to dirty & 440168404Spjd * sync out the changes after we check for being dirty. 441219089Spjd * Also, dbuf_hold_impl() wants us to have the struct_rwlock. 442168404Spjd */ 443168404Spjd rw_enter(&dn->dn_struct_rwlock, RW_READER); 444185029Spjd epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 445185029Spjd if (dn->dn_maxblkid == 0) { 446168404Spjd if (off == 0 && len >= dn->dn_datablksz) { 447168404Spjd blkid = 0; 448168404Spjd nblks = 1; 449168404Spjd } else { 450168404Spjd rw_exit(&dn->dn_struct_rwlock); 451168404Spjd return; 452168404Spjd } 453168404Spjd } else { 454168404Spjd blkid = off >> dn->dn_datablkshift; 455185029Spjd nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift; 456168404Spjd 457185029Spjd if (blkid >= dn->dn_maxblkid) { 458168404Spjd rw_exit(&dn->dn_struct_rwlock); 459168404Spjd return; 460168404Spjd } 461185029Spjd if (blkid + nblks > dn->dn_maxblkid) 462185029Spjd nblks = dn->dn_maxblkid - blkid; 463168404Spjd 464168404Spjd } 465240133Smm l0span = nblks; /* save for later use to calc level > 1 overhead */ 466185029Spjd if (dn->dn_nlevels == 1) { 467168404Spjd int i; 468168404Spjd for (i = 0; i < nblks; i++) { 469168404Spjd blkptr_t *bp = dn->dn_phys->dn_blkptr; 470185029Spjd ASSERT3U(blkid + i, <, dn->dn_nblkptr); 471168404Spjd bp += blkid + i; 472219089Spjd if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) { 473168404Spjd dprintf_bp(bp, "can free old%s", ""); 474219089Spjd space += bp_get_dsize(spa, bp); 475168404Spjd } 476185029Spjd unref += BP_GET_ASIZE(bp); 477168404Spjd } 478240133Smm nl1blks = 1; 479168404Spjd nblks = 0; 480168404Spjd } 481168404Spjd 482185029Spjd lastblk = blkid + nblks - 1; 483168404Spjd while (nblks) { 484168404Spjd dmu_buf_impl_t *dbuf; 485185029Spjd uint64_t ibyte, new_blkid; 486185029Spjd int epb = 1 << epbs; 487185029Spjd int err, i, blkoff, tochk; 488185029Spjd blkptr_t *bp; 489168404Spjd 490185029Spjd ibyte = blkid << dn->dn_datablkshift; 491185029Spjd err = dnode_next_offset(dn, 492185029Spjd DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0); 493185029Spjd new_blkid = ibyte >> dn->dn_datablkshift; 494185029Spjd if (err == ESRCH) { 495185029Spjd skipped += (lastblk >> epbs) - (blkid >> epbs) + 1; 496185029Spjd break; 497185029Spjd } 498185029Spjd if (err) { 499185029Spjd txh->txh_tx->tx_err = err; 500185029Spjd break; 501185029Spjd } 502185029Spjd if (new_blkid > lastblk) { 503185029Spjd skipped += (lastblk >> epbs) - (blkid >> epbs) + 1; 504185029Spjd break; 505185029Spjd } 506168404Spjd 507185029Spjd if (new_blkid > blkid) { 508185029Spjd ASSERT((new_blkid >> epbs) > (blkid >> epbs)); 509185029Spjd skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1; 510185029Spjd nblks -= new_blkid - blkid; 511185029Spjd blkid = new_blkid; 512185029Spjd } 513185029Spjd blkoff = P2PHASE(blkid, epb); 514185029Spjd tochk = MIN(epb - blkoff, nblks); 515168404Spjd 516219089Spjd err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf); 517219089Spjd if (err) { 518219089Spjd txh->txh_tx->tx_err = err; 519219089Spjd break; 520219089Spjd } 521168404Spjd 522185029Spjd txh->txh_memory_tohold += dbuf->db.db_size; 523219089Spjd 524219089Spjd /* 525219089Spjd * We don't check memory_tohold against DMU_MAX_ACCESS because 526219089Spjd * memory_tohold is an over-estimation (especially the >L1 527219089Spjd * indirect blocks), so it could fail. Callers should have 528219089Spjd * already verified that they will not be holding too much 529219089Spjd * memory. 530219089Spjd */ 531219089Spjd 532185029Spjd err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL); 533185029Spjd if (err != 0) { 534168404Spjd txh->txh_tx->tx_err = err; 535185029Spjd dbuf_rele(dbuf, FTAG); 536168404Spjd break; 537168404Spjd } 538168404Spjd 539185029Spjd bp = dbuf->db.db_data; 540185029Spjd bp += blkoff; 541185029Spjd 542185029Spjd for (i = 0; i < tochk; i++) { 543219089Spjd if (dsl_dataset_block_freeable(ds, &bp[i], 544219089Spjd bp[i].blk_birth)) { 545185029Spjd dprintf_bp(&bp[i], "can free old%s", ""); 546219089Spjd space += bp_get_dsize(spa, &bp[i]); 547185029Spjd } 548185029Spjd unref += BP_GET_ASIZE(bp); 549185029Spjd } 550185029Spjd dbuf_rele(dbuf, FTAG); 551185029Spjd 552240133Smm ++nl1blks; 553168404Spjd blkid += tochk; 554168404Spjd nblks -= tochk; 555168404Spjd } 556168404Spjd rw_exit(&dn->dn_struct_rwlock); 557168404Spjd 558240133Smm /* 559240133Smm * Add in memory requirements of higher-level indirects. 560240133Smm * This assumes a worst-possible scenario for dn_nlevels and a 561240133Smm * worst-possible distribution of l1-blocks over the region to free. 562240133Smm */ 563240133Smm { 564240133Smm uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs); 565240133Smm int level = 2; 566240133Smm /* 567240133Smm * Here we don't use DN_MAX_LEVEL, but calculate it with the 568240133Smm * given datablkshift and indblkshift. This makes the 569240133Smm * difference between 19 and 8 on large files. 570240133Smm */ 571240133Smm int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) / 572240133Smm (dn->dn_indblkshift - SPA_BLKPTRSHIFT); 573240133Smm 574240133Smm while (level++ < maxlevel) { 575240133Smm txh->txh_memory_tohold += MIN(blkcnt, (nl1blks >> epbs)) 576240133Smm << dn->dn_indblkshift; 577240133Smm blkcnt = 1 + (blkcnt >> epbs); 578240133Smm } 579240133Smm } 580240133Smm 581185029Spjd /* account for new level 1 indirect blocks that might show up */ 582185029Spjd if (skipped > 0) { 583185029Spjd txh->txh_fudge += skipped << dn->dn_indblkshift; 584185029Spjd skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs); 585185029Spjd txh->txh_memory_tohold += skipped << dn->dn_indblkshift; 586185029Spjd } 587168404Spjd txh->txh_space_tofree += space; 588185029Spjd txh->txh_space_tounref += unref; 589168404Spjd} 590168404Spjd 591168404Spjdvoid 592168404Spjddmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len) 593168404Spjd{ 594168404Spjd dmu_tx_hold_t *txh; 595168404Spjd dnode_t *dn; 596168404Spjd uint64_t start, end, i; 597168404Spjd int err, shift; 598168404Spjd zio_t *zio; 599168404Spjd 600168404Spjd ASSERT(tx->tx_txg == 0); 601168404Spjd 602168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 603168404Spjd object, THT_FREE, off, len); 604168404Spjd if (txh == NULL) 605168404Spjd return; 606168404Spjd dn = txh->txh_dnode; 607168404Spjd 608168404Spjd /* first block */ 609168404Spjd if (off != 0) 610168404Spjd dmu_tx_count_write(txh, off, 1); 611168404Spjd /* last block */ 612168404Spjd if (len != DMU_OBJECT_END) 613168404Spjd dmu_tx_count_write(txh, off+len, 1); 614168404Spjd 615219089Spjd dmu_tx_count_dnode(txh); 616219089Spjd 617168404Spjd if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz) 618168404Spjd return; 619168404Spjd if (len == DMU_OBJECT_END) 620168404Spjd len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off; 621168404Spjd 622168404Spjd /* 623168404Spjd * For i/o error checking, read the first and last level-0 624168404Spjd * blocks, and all the level-1 blocks. The above count_write's 625185029Spjd * have already taken care of the level-0 blocks. 626168404Spjd */ 627168404Spjd if (dn->dn_nlevels > 1) { 628168404Spjd shift = dn->dn_datablkshift + dn->dn_indblkshift - 629168404Spjd SPA_BLKPTRSHIFT; 630168404Spjd start = off >> shift; 631168404Spjd end = dn->dn_datablkshift ? ((off+len) >> shift) : 0; 632168404Spjd 633168404Spjd zio = zio_root(tx->tx_pool->dp_spa, 634168404Spjd NULL, NULL, ZIO_FLAG_CANFAIL); 635168404Spjd for (i = start; i <= end; i++) { 636168404Spjd uint64_t ibyte = i << shift; 637185029Spjd err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0); 638168404Spjd i = ibyte >> shift; 639168404Spjd if (err == ESRCH) 640168404Spjd break; 641168404Spjd if (err) { 642168404Spjd tx->tx_err = err; 643168404Spjd return; 644168404Spjd } 645168404Spjd 646168404Spjd err = dmu_tx_check_ioerr(zio, dn, 1, i); 647168404Spjd if (err) { 648168404Spjd tx->tx_err = err; 649168404Spjd return; 650168404Spjd } 651168404Spjd } 652168404Spjd err = zio_wait(zio); 653168404Spjd if (err) { 654168404Spjd tx->tx_err = err; 655168404Spjd return; 656168404Spjd } 657168404Spjd } 658168404Spjd 659168404Spjd dmu_tx_count_free(txh, off, len); 660168404Spjd} 661168404Spjd 662168404Spjdvoid 663209962Smmdmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name) 664168404Spjd{ 665168404Spjd dmu_tx_hold_t *txh; 666168404Spjd dnode_t *dn; 667168404Spjd uint64_t nblocks; 668168404Spjd int epbs, err; 669168404Spjd 670168404Spjd ASSERT(tx->tx_txg == 0); 671168404Spjd 672168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 673168404Spjd object, THT_ZAP, add, (uintptr_t)name); 674168404Spjd if (txh == NULL) 675168404Spjd return; 676168404Spjd dn = txh->txh_dnode; 677168404Spjd 678168404Spjd dmu_tx_count_dnode(txh); 679168404Spjd 680168404Spjd if (dn == NULL) { 681168404Spjd /* 682168404Spjd * We will be able to fit a new object's entries into one leaf 683168404Spjd * block. So there will be at most 2 blocks total, 684168404Spjd * including the header block. 685168404Spjd */ 686168404Spjd dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift); 687168404Spjd return; 688168404Spjd } 689168404Spjd 690236884Smm ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP); 691168404Spjd 692168404Spjd if (dn->dn_maxblkid == 0 && !add) { 693226512Smm blkptr_t *bp; 694226512Smm 695168404Spjd /* 696168404Spjd * If there is only one block (i.e. this is a micro-zap) 697168404Spjd * and we are not adding anything, the accounting is simple. 698168404Spjd */ 699168404Spjd err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 700168404Spjd if (err) { 701168404Spjd tx->tx_err = err; 702168404Spjd return; 703168404Spjd } 704168404Spjd 705168404Spjd /* 706168404Spjd * Use max block size here, since we don't know how much 707168404Spjd * the size will change between now and the dbuf dirty call. 708168404Spjd */ 709226512Smm bp = &dn->dn_phys->dn_blkptr[0]; 710168404Spjd if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, 711226512Smm bp, bp->blk_birth)) 712168404Spjd txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE; 713226512Smm else 714168404Spjd txh->txh_space_towrite += SPA_MAXBLOCKSIZE; 715226512Smm if (!BP_IS_HOLE(bp)) 716209093Smm txh->txh_space_tounref += SPA_MAXBLOCKSIZE; 717168404Spjd return; 718168404Spjd } 719168404Spjd 720168404Spjd if (dn->dn_maxblkid > 0 && name) { 721168404Spjd /* 722168404Spjd * access the name in this fat-zap so that we'll check 723168404Spjd * for i/o errors to the leaf blocks, etc. 724168404Spjd */ 725219089Spjd err = zap_lookup(dn->dn_objset, dn->dn_object, name, 726168404Spjd 8, 0, NULL); 727168404Spjd if (err == EIO) { 728168404Spjd tx->tx_err = err; 729168404Spjd return; 730168404Spjd } 731168404Spjd } 732168404Spjd 733219089Spjd err = zap_count_write(dn->dn_objset, dn->dn_object, name, add, 734209962Smm &txh->txh_space_towrite, &txh->txh_space_tooverwrite); 735168404Spjd 736168404Spjd /* 737168404Spjd * If the modified blocks are scattered to the four winds, 738168404Spjd * we'll have to modify an indirect twig for each. 739168404Spjd */ 740168404Spjd epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 741168404Spjd for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs) 742209962Smm if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj) 743209962Smm txh->txh_space_towrite += 3 << dn->dn_indblkshift; 744209962Smm else 745209962Smm txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift; 746168404Spjd} 747168404Spjd 748168404Spjdvoid 749168404Spjddmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object) 750168404Spjd{ 751168404Spjd dmu_tx_hold_t *txh; 752168404Spjd 753168404Spjd ASSERT(tx->tx_txg == 0); 754168404Spjd 755168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 756168404Spjd object, THT_BONUS, 0, 0); 757168404Spjd if (txh) 758168404Spjd dmu_tx_count_dnode(txh); 759168404Spjd} 760168404Spjd 761168404Spjdvoid 762168404Spjddmu_tx_hold_space(dmu_tx_t *tx, uint64_t space) 763168404Spjd{ 764168404Spjd dmu_tx_hold_t *txh; 765168404Spjd ASSERT(tx->tx_txg == 0); 766168404Spjd 767168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 768168404Spjd DMU_NEW_OBJECT, THT_SPACE, space, 0); 769168404Spjd 770168404Spjd txh->txh_space_towrite += space; 771168404Spjd} 772168404Spjd 773168404Spjdint 774168404Spjddmu_tx_holds(dmu_tx_t *tx, uint64_t object) 775168404Spjd{ 776168404Spjd dmu_tx_hold_t *txh; 777168404Spjd int holds = 0; 778168404Spjd 779168404Spjd /* 780168404Spjd * By asserting that the tx is assigned, we're counting the 781168404Spjd * number of dn_tx_holds, which is the same as the number of 782168404Spjd * dn_holds. Otherwise, we'd be counting dn_holds, but 783168404Spjd * dn_tx_holds could be 0. 784168404Spjd */ 785168404Spjd ASSERT(tx->tx_txg != 0); 786168404Spjd 787168404Spjd /* if (tx->tx_anyobj == TRUE) */ 788168404Spjd /* return (0); */ 789168404Spjd 790168404Spjd for (txh = list_head(&tx->tx_holds); txh; 791168404Spjd txh = list_next(&tx->tx_holds, txh)) { 792168404Spjd if (txh->txh_dnode && txh->txh_dnode->dn_object == object) 793168404Spjd holds++; 794168404Spjd } 795168404Spjd 796168404Spjd return (holds); 797168404Spjd} 798168404Spjd 799168404Spjd#ifdef ZFS_DEBUG 800168404Spjdvoid 801168404Spjddmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db) 802168404Spjd{ 803168404Spjd dmu_tx_hold_t *txh; 804168404Spjd int match_object = FALSE, match_offset = FALSE; 805219089Spjd dnode_t *dn; 806168404Spjd 807219089Spjd DB_DNODE_ENTER(db); 808219089Spjd dn = DB_DNODE(db); 809168404Spjd ASSERT(tx->tx_txg != 0); 810219089Spjd ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset); 811168404Spjd ASSERT3U(dn->dn_object, ==, db->db.db_object); 812168404Spjd 813219089Spjd if (tx->tx_anyobj) { 814219089Spjd DB_DNODE_EXIT(db); 815168404Spjd return; 816219089Spjd } 817168404Spjd 818168404Spjd /* XXX No checking on the meta dnode for now */ 819219089Spjd if (db->db.db_object == DMU_META_DNODE_OBJECT) { 820219089Spjd DB_DNODE_EXIT(db); 821168404Spjd return; 822219089Spjd } 823168404Spjd 824168404Spjd for (txh = list_head(&tx->tx_holds); txh; 825168404Spjd txh = list_next(&tx->tx_holds, txh)) { 826168404Spjd ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg); 827168404Spjd if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT) 828168404Spjd match_object = TRUE; 829168404Spjd if (txh->txh_dnode == NULL || txh->txh_dnode == dn) { 830168404Spjd int datablkshift = dn->dn_datablkshift ? 831168404Spjd dn->dn_datablkshift : SPA_MAXBLOCKSHIFT; 832168404Spjd int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 833168404Spjd int shift = datablkshift + epbs * db->db_level; 834168404Spjd uint64_t beginblk = shift >= 64 ? 0 : 835168404Spjd (txh->txh_arg1 >> shift); 836168404Spjd uint64_t endblk = shift >= 64 ? 0 : 837168404Spjd ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift); 838168404Spjd uint64_t blkid = db->db_blkid; 839168404Spjd 840168404Spjd /* XXX txh_arg2 better not be zero... */ 841168404Spjd 842168404Spjd dprintf("found txh type %x beginblk=%llx endblk=%llx\n", 843168404Spjd txh->txh_type, beginblk, endblk); 844168404Spjd 845168404Spjd switch (txh->txh_type) { 846168404Spjd case THT_WRITE: 847168404Spjd if (blkid >= beginblk && blkid <= endblk) 848168404Spjd match_offset = TRUE; 849168404Spjd /* 850168404Spjd * We will let this hold work for the bonus 851219089Spjd * or spill buffer so that we don't need to 852219089Spjd * hold it when creating a new object. 853168404Spjd */ 854219089Spjd if (blkid == DMU_BONUS_BLKID || 855219089Spjd blkid == DMU_SPILL_BLKID) 856168404Spjd match_offset = TRUE; 857168404Spjd /* 858168404Spjd * They might have to increase nlevels, 859168404Spjd * thus dirtying the new TLIBs. Or the 860168404Spjd * might have to change the block size, 861168404Spjd * thus dirying the new lvl=0 blk=0. 862168404Spjd */ 863168404Spjd if (blkid == 0) 864168404Spjd match_offset = TRUE; 865168404Spjd break; 866168404Spjd case THT_FREE: 867185029Spjd /* 868185029Spjd * We will dirty all the level 1 blocks in 869185029Spjd * the free range and perhaps the first and 870185029Spjd * last level 0 block. 871185029Spjd */ 872185029Spjd if (blkid >= beginblk && (blkid <= endblk || 873185029Spjd txh->txh_arg2 == DMU_OBJECT_END)) 874168404Spjd match_offset = TRUE; 875168404Spjd break; 876219089Spjd case THT_SPILL: 877219089Spjd if (blkid == DMU_SPILL_BLKID) 878219089Spjd match_offset = TRUE; 879219089Spjd break; 880168404Spjd case THT_BONUS: 881219089Spjd if (blkid == DMU_BONUS_BLKID) 882168404Spjd match_offset = TRUE; 883168404Spjd break; 884168404Spjd case THT_ZAP: 885168404Spjd match_offset = TRUE; 886168404Spjd break; 887168404Spjd case THT_NEWOBJECT: 888168404Spjd match_object = TRUE; 889168404Spjd break; 890168404Spjd default: 891168404Spjd ASSERT(!"bad txh_type"); 892168404Spjd } 893168404Spjd } 894219089Spjd if (match_object && match_offset) { 895219089Spjd DB_DNODE_EXIT(db); 896168404Spjd return; 897219089Spjd } 898168404Spjd } 899219089Spjd DB_DNODE_EXIT(db); 900168404Spjd panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n", 901168404Spjd (u_longlong_t)db->db.db_object, db->db_level, 902168404Spjd (u_longlong_t)db->db_blkid); 903168404Spjd} 904168404Spjd#endif 905168404Spjd 906168404Spjdstatic int 907168404Spjddmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how) 908168404Spjd{ 909168404Spjd dmu_tx_hold_t *txh; 910185029Spjd spa_t *spa = tx->tx_pool->dp_spa; 911185029Spjd uint64_t memory, asize, fsize, usize; 912185029Spjd uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge; 913168404Spjd 914168404Spjd ASSERT3U(tx->tx_txg, ==, 0); 915185029Spjd 916168404Spjd if (tx->tx_err) 917168404Spjd return (tx->tx_err); 918168404Spjd 919185029Spjd if (spa_suspended(spa)) { 920185029Spjd /* 921185029Spjd * If the user has indicated a blocking failure mode 922185029Spjd * then return ERESTART which will block in dmu_tx_wait(). 923185029Spjd * Otherwise, return EIO so that an error can get 924185029Spjd * propagated back to the VOP calls. 925185029Spjd * 926185029Spjd * Note that we always honor the txg_how flag regardless 927185029Spjd * of the failuremode setting. 928185029Spjd */ 929185029Spjd if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE && 930185029Spjd txg_how != TXG_WAIT) 931185029Spjd return (EIO); 932185029Spjd 933185029Spjd return (ERESTART); 934185029Spjd } 935185029Spjd 936168404Spjd tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh); 937168404Spjd tx->tx_needassign_txh = NULL; 938168404Spjd 939168404Spjd /* 940168404Spjd * NB: No error returns are allowed after txg_hold_open, but 941168404Spjd * before processing the dnode holds, due to the 942168404Spjd * dmu_tx_unassign() logic. 943168404Spjd */ 944168404Spjd 945185029Spjd towrite = tofree = tooverwrite = tounref = tohold = fudge = 0; 946168404Spjd for (txh = list_head(&tx->tx_holds); txh; 947168404Spjd txh = list_next(&tx->tx_holds, txh)) { 948168404Spjd dnode_t *dn = txh->txh_dnode; 949168404Spjd if (dn != NULL) { 950168404Spjd mutex_enter(&dn->dn_mtx); 951168404Spjd if (dn->dn_assigned_txg == tx->tx_txg - 1) { 952168404Spjd mutex_exit(&dn->dn_mtx); 953168404Spjd tx->tx_needassign_txh = txh; 954168404Spjd return (ERESTART); 955168404Spjd } 956168404Spjd if (dn->dn_assigned_txg == 0) 957168404Spjd dn->dn_assigned_txg = tx->tx_txg; 958168404Spjd ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 959168404Spjd (void) refcount_add(&dn->dn_tx_holds, tx); 960168404Spjd mutex_exit(&dn->dn_mtx); 961168404Spjd } 962168404Spjd towrite += txh->txh_space_towrite; 963168404Spjd tofree += txh->txh_space_tofree; 964168404Spjd tooverwrite += txh->txh_space_tooverwrite; 965185029Spjd tounref += txh->txh_space_tounref; 966185029Spjd tohold += txh->txh_memory_tohold; 967185029Spjd fudge += txh->txh_fudge; 968168404Spjd } 969168404Spjd 970168404Spjd /* 971168404Spjd * NB: This check must be after we've held the dnodes, so that 972168404Spjd * the dmu_tx_unassign() logic will work properly 973168404Spjd */ 974168404Spjd if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg) 975168404Spjd return (ERESTART); 976168404Spjd 977168404Spjd /* 978168404Spjd * If a snapshot has been taken since we made our estimates, 979168404Spjd * assume that we won't be able to free or overwrite anything. 980168404Spjd */ 981168404Spjd if (tx->tx_objset && 982219089Spjd dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) > 983168404Spjd tx->tx_lastsnap_txg) { 984168404Spjd towrite += tooverwrite; 985168404Spjd tooverwrite = tofree = 0; 986168404Spjd } 987168404Spjd 988185029Spjd /* needed allocation: worst-case estimate of write space */ 989185029Spjd asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite); 990185029Spjd /* freed space estimate: worst-case overwrite + free estimate */ 991168404Spjd fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree; 992185029Spjd /* convert unrefd space to worst-case estimate */ 993185029Spjd usize = spa_get_asize(tx->tx_pool->dp_spa, tounref); 994185029Spjd /* calculate memory footprint estimate */ 995185029Spjd memory = towrite + tooverwrite + tohold; 996168404Spjd 997168404Spjd#ifdef ZFS_DEBUG 998185029Spjd /* 999185029Spjd * Add in 'tohold' to account for our dirty holds on this memory 1000185029Spjd * XXX - the "fudge" factor is to account for skipped blocks that 1001185029Spjd * we missed because dnode_next_offset() misses in-core-only blocks. 1002185029Spjd */ 1003185029Spjd tx->tx_space_towrite = asize + 1004185029Spjd spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge); 1005168404Spjd tx->tx_space_tofree = tofree; 1006168404Spjd tx->tx_space_tooverwrite = tooverwrite; 1007185029Spjd tx->tx_space_tounref = tounref; 1008168404Spjd#endif 1009168404Spjd 1010168404Spjd if (tx->tx_dir && asize != 0) { 1011185029Spjd int err = dsl_dir_tempreserve_space(tx->tx_dir, memory, 1012185029Spjd asize, fsize, usize, &tx->tx_tempreserve_cookie, tx); 1013168404Spjd if (err) 1014168404Spjd return (err); 1015168404Spjd } 1016168404Spjd 1017168404Spjd return (0); 1018168404Spjd} 1019168404Spjd 1020168404Spjdstatic void 1021168404Spjddmu_tx_unassign(dmu_tx_t *tx) 1022168404Spjd{ 1023168404Spjd dmu_tx_hold_t *txh; 1024168404Spjd 1025168404Spjd if (tx->tx_txg == 0) 1026168404Spjd return; 1027168404Spjd 1028168404Spjd txg_rele_to_quiesce(&tx->tx_txgh); 1029168404Spjd 1030168404Spjd for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh; 1031168404Spjd txh = list_next(&tx->tx_holds, txh)) { 1032168404Spjd dnode_t *dn = txh->txh_dnode; 1033168404Spjd 1034168404Spjd if (dn == NULL) 1035168404Spjd continue; 1036168404Spjd mutex_enter(&dn->dn_mtx); 1037168404Spjd ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1038168404Spjd 1039168404Spjd if (refcount_remove(&dn->dn_tx_holds, tx) == 0) { 1040168404Spjd dn->dn_assigned_txg = 0; 1041168404Spjd cv_broadcast(&dn->dn_notxholds); 1042168404Spjd } 1043168404Spjd mutex_exit(&dn->dn_mtx); 1044168404Spjd } 1045168404Spjd 1046168404Spjd txg_rele_to_sync(&tx->tx_txgh); 1047168404Spjd 1048168404Spjd tx->tx_lasttried_txg = tx->tx_txg; 1049168404Spjd tx->tx_txg = 0; 1050168404Spjd} 1051168404Spjd 1052168404Spjd/* 1053168404Spjd * Assign tx to a transaction group. txg_how can be one of: 1054168404Spjd * 1055168404Spjd * (1) TXG_WAIT. If the current open txg is full, waits until there's 1056168404Spjd * a new one. This should be used when you're not holding locks. 1057168404Spjd * If will only fail if we're truly out of space (or over quota). 1058168404Spjd * 1059168404Spjd * (2) TXG_NOWAIT. If we can't assign into the current open txg without 1060168404Spjd * blocking, returns immediately with ERESTART. This should be used 1061168404Spjd * whenever you're holding locks. On an ERESTART error, the caller 1062168404Spjd * should drop locks, do a dmu_tx_wait(tx), and try again. 1063168404Spjd * 1064168404Spjd * (3) A specific txg. Use this if you need to ensure that multiple 1065168404Spjd * transactions all sync in the same txg. Like TXG_NOWAIT, it 1066168404Spjd * returns ERESTART if it can't assign you into the requested txg. 1067168404Spjd */ 1068168404Spjdint 1069168404Spjddmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how) 1070168404Spjd{ 1071168404Spjd int err; 1072168404Spjd 1073168404Spjd ASSERT(tx->tx_txg == 0); 1074168404Spjd ASSERT(txg_how != 0); 1075168404Spjd ASSERT(!dsl_pool_sync_context(tx->tx_pool)); 1076168404Spjd 1077168404Spjd while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) { 1078168404Spjd dmu_tx_unassign(tx); 1079168404Spjd 1080168404Spjd if (err != ERESTART || txg_how != TXG_WAIT) 1081168404Spjd return (err); 1082168404Spjd 1083168404Spjd dmu_tx_wait(tx); 1084168404Spjd } 1085168404Spjd 1086168404Spjd txg_rele_to_quiesce(&tx->tx_txgh); 1087168404Spjd 1088168404Spjd return (0); 1089168404Spjd} 1090168404Spjd 1091168404Spjdvoid 1092168404Spjddmu_tx_wait(dmu_tx_t *tx) 1093168404Spjd{ 1094185029Spjd spa_t *spa = tx->tx_pool->dp_spa; 1095185029Spjd 1096168404Spjd ASSERT(tx->tx_txg == 0); 1097168404Spjd 1098185029Spjd /* 1099185029Spjd * It's possible that the pool has become active after this thread 1100185029Spjd * has tried to obtain a tx. If that's the case then his 1101185029Spjd * tx_lasttried_txg would not have been assigned. 1102185029Spjd */ 1103185029Spjd if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) { 1104185029Spjd txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1); 1105185029Spjd } else if (tx->tx_needassign_txh) { 1106168404Spjd dnode_t *dn = tx->tx_needassign_txh->txh_dnode; 1107168404Spjd 1108168404Spjd mutex_enter(&dn->dn_mtx); 1109168404Spjd while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1) 1110168404Spjd cv_wait(&dn->dn_notxholds, &dn->dn_mtx); 1111168404Spjd mutex_exit(&dn->dn_mtx); 1112168404Spjd tx->tx_needassign_txh = NULL; 1113168404Spjd } else { 1114168404Spjd txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1); 1115168404Spjd } 1116168404Spjd} 1117168404Spjd 1118168404Spjdvoid 1119168404Spjddmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta) 1120168404Spjd{ 1121168404Spjd#ifdef ZFS_DEBUG 1122168404Spjd if (tx->tx_dir == NULL || delta == 0) 1123168404Spjd return; 1124168404Spjd 1125168404Spjd if (delta > 0) { 1126168404Spjd ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=, 1127168404Spjd tx->tx_space_towrite); 1128168404Spjd (void) refcount_add_many(&tx->tx_space_written, delta, NULL); 1129168404Spjd } else { 1130168404Spjd (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL); 1131168404Spjd } 1132168404Spjd#endif 1133168404Spjd} 1134168404Spjd 1135168404Spjdvoid 1136168404Spjddmu_tx_commit(dmu_tx_t *tx) 1137168404Spjd{ 1138168404Spjd dmu_tx_hold_t *txh; 1139168404Spjd 1140168404Spjd ASSERT(tx->tx_txg != 0); 1141168404Spjd 1142168404Spjd while (txh = list_head(&tx->tx_holds)) { 1143168404Spjd dnode_t *dn = txh->txh_dnode; 1144168404Spjd 1145168404Spjd list_remove(&tx->tx_holds, txh); 1146168404Spjd kmem_free(txh, sizeof (dmu_tx_hold_t)); 1147168404Spjd if (dn == NULL) 1148168404Spjd continue; 1149168404Spjd mutex_enter(&dn->dn_mtx); 1150168404Spjd ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1151168404Spjd 1152168404Spjd if (refcount_remove(&dn->dn_tx_holds, tx) == 0) { 1153168404Spjd dn->dn_assigned_txg = 0; 1154168404Spjd cv_broadcast(&dn->dn_notxholds); 1155168404Spjd } 1156168404Spjd mutex_exit(&dn->dn_mtx); 1157168404Spjd dnode_rele(dn, tx); 1158168404Spjd } 1159168404Spjd 1160168404Spjd if (tx->tx_tempreserve_cookie) 1161168404Spjd dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx); 1162168404Spjd 1163219089Spjd if (!list_is_empty(&tx->tx_callbacks)) 1164219089Spjd txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks); 1165219089Spjd 1166168404Spjd if (tx->tx_anyobj == FALSE) 1167168404Spjd txg_rele_to_sync(&tx->tx_txgh); 1168219089Spjd 1169219089Spjd list_destroy(&tx->tx_callbacks); 1170185029Spjd list_destroy(&tx->tx_holds); 1171168404Spjd#ifdef ZFS_DEBUG 1172168404Spjd dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n", 1173168404Spjd tx->tx_space_towrite, refcount_count(&tx->tx_space_written), 1174168404Spjd tx->tx_space_tofree, refcount_count(&tx->tx_space_freed)); 1175168404Spjd refcount_destroy_many(&tx->tx_space_written, 1176168404Spjd refcount_count(&tx->tx_space_written)); 1177168404Spjd refcount_destroy_many(&tx->tx_space_freed, 1178168404Spjd refcount_count(&tx->tx_space_freed)); 1179168404Spjd#endif 1180168404Spjd kmem_free(tx, sizeof (dmu_tx_t)); 1181168404Spjd} 1182168404Spjd 1183168404Spjdvoid 1184168404Spjddmu_tx_abort(dmu_tx_t *tx) 1185168404Spjd{ 1186168404Spjd dmu_tx_hold_t *txh; 1187168404Spjd 1188168404Spjd ASSERT(tx->tx_txg == 0); 1189168404Spjd 1190168404Spjd while (txh = list_head(&tx->tx_holds)) { 1191168404Spjd dnode_t *dn = txh->txh_dnode; 1192168404Spjd 1193168404Spjd list_remove(&tx->tx_holds, txh); 1194168404Spjd kmem_free(txh, sizeof (dmu_tx_hold_t)); 1195168404Spjd if (dn != NULL) 1196168404Spjd dnode_rele(dn, tx); 1197168404Spjd } 1198219089Spjd 1199219089Spjd /* 1200219089Spjd * Call any registered callbacks with an error code. 1201219089Spjd */ 1202219089Spjd if (!list_is_empty(&tx->tx_callbacks)) 1203219089Spjd dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED); 1204219089Spjd 1205219089Spjd list_destroy(&tx->tx_callbacks); 1206185029Spjd list_destroy(&tx->tx_holds); 1207168404Spjd#ifdef ZFS_DEBUG 1208168404Spjd refcount_destroy_many(&tx->tx_space_written, 1209168404Spjd refcount_count(&tx->tx_space_written)); 1210168404Spjd refcount_destroy_many(&tx->tx_space_freed, 1211168404Spjd refcount_count(&tx->tx_space_freed)); 1212168404Spjd#endif 1213168404Spjd kmem_free(tx, sizeof (dmu_tx_t)); 1214168404Spjd} 1215168404Spjd 1216168404Spjduint64_t 1217168404Spjddmu_tx_get_txg(dmu_tx_t *tx) 1218168404Spjd{ 1219168404Spjd ASSERT(tx->tx_txg != 0); 1220168404Spjd return (tx->tx_txg); 1221168404Spjd} 1222219089Spjd 1223219089Spjdvoid 1224219089Spjddmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data) 1225219089Spjd{ 1226219089Spjd dmu_tx_callback_t *dcb; 1227219089Spjd 1228219089Spjd dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP); 1229219089Spjd 1230219089Spjd dcb->dcb_func = func; 1231219089Spjd dcb->dcb_data = data; 1232219089Spjd 1233219089Spjd list_insert_tail(&tx->tx_callbacks, dcb); 1234219089Spjd} 1235219089Spjd 1236219089Spjd/* 1237219089Spjd * Call all the commit callbacks on a list, with a given error code. 1238219089Spjd */ 1239219089Spjdvoid 1240219089Spjddmu_tx_do_callbacks(list_t *cb_list, int error) 1241219089Spjd{ 1242219089Spjd dmu_tx_callback_t *dcb; 1243219089Spjd 1244219089Spjd while (dcb = list_head(cb_list)) { 1245219089Spjd list_remove(cb_list, dcb); 1246219089Spjd dcb->dcb_func(dcb->dcb_data, error); 1247219089Spjd kmem_free(dcb, sizeof (dmu_tx_callback_t)); 1248219089Spjd } 1249219089Spjd} 1250219089Spjd 1251219089Spjd/* 1252219089Spjd * Interface to hold a bunch of attributes. 1253219089Spjd * used for creating new files. 1254219089Spjd * attrsize is the total size of all attributes 1255219089Spjd * to be added during object creation 1256219089Spjd * 1257219089Spjd * For updating/adding a single attribute dmu_tx_hold_sa() should be used. 1258219089Spjd */ 1259219089Spjd 1260219089Spjd/* 1261219089Spjd * hold necessary attribute name for attribute registration. 1262219089Spjd * should be a very rare case where this is needed. If it does 1263219089Spjd * happen it would only happen on the first write to the file system. 1264219089Spjd */ 1265219089Spjdstatic void 1266219089Spjddmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx) 1267219089Spjd{ 1268219089Spjd int i; 1269219089Spjd 1270219089Spjd if (!sa->sa_need_attr_registration) 1271219089Spjd return; 1272219089Spjd 1273219089Spjd for (i = 0; i != sa->sa_num_attrs; i++) { 1274219089Spjd if (!sa->sa_attr_table[i].sa_registered) { 1275219089Spjd if (sa->sa_reg_attr_obj) 1276219089Spjd dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj, 1277219089Spjd B_TRUE, sa->sa_attr_table[i].sa_name); 1278219089Spjd else 1279219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, 1280219089Spjd B_TRUE, sa->sa_attr_table[i].sa_name); 1281219089Spjd } 1282219089Spjd } 1283219089Spjd} 1284219089Spjd 1285219089Spjd 1286219089Spjdvoid 1287219089Spjddmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object) 1288219089Spjd{ 1289219089Spjd dnode_t *dn; 1290219089Spjd dmu_tx_hold_t *txh; 1291219089Spjd 1292219089Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object, 1293219089Spjd THT_SPILL, 0, 0); 1294219089Spjd 1295219089Spjd dn = txh->txh_dnode; 1296219089Spjd 1297219089Spjd if (dn == NULL) 1298219089Spjd return; 1299219089Spjd 1300219089Spjd /* If blkptr doesn't exist then add space to towrite */ 1301219089Spjd if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { 1302219089Spjd txh->txh_space_towrite += SPA_MAXBLOCKSIZE; 1303219089Spjd } else { 1304226512Smm blkptr_t *bp; 1305226512Smm 1306219089Spjd bp = &dn->dn_phys->dn_spill; 1307219089Spjd if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, 1308219089Spjd bp, bp->blk_birth)) 1309219089Spjd txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE; 1310219089Spjd else 1311219089Spjd txh->txh_space_towrite += SPA_MAXBLOCKSIZE; 1312226512Smm if (!BP_IS_HOLE(bp)) 1313219089Spjd txh->txh_space_tounref += SPA_MAXBLOCKSIZE; 1314219089Spjd } 1315219089Spjd} 1316219089Spjd 1317219089Spjdvoid 1318219089Spjddmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize) 1319219089Spjd{ 1320219089Spjd sa_os_t *sa = tx->tx_objset->os_sa; 1321219089Spjd 1322219089Spjd dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT); 1323219089Spjd 1324219089Spjd if (tx->tx_objset->os_sa->sa_master_obj == 0) 1325219089Spjd return; 1326219089Spjd 1327219089Spjd if (tx->tx_objset->os_sa->sa_layout_attr_obj) 1328219089Spjd dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL); 1329219089Spjd else { 1330219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS); 1331219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY); 1332219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1333219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1334219089Spjd } 1335219089Spjd 1336219089Spjd dmu_tx_sa_registration_hold(sa, tx); 1337219089Spjd 1338219089Spjd if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill) 1339219089Spjd return; 1340219089Spjd 1341219089Spjd (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT, 1342219089Spjd THT_SPILL, 0, 0); 1343219089Spjd} 1344219089Spjd 1345219089Spjd/* 1346219089Spjd * Hold SA attribute 1347219089Spjd * 1348219089Spjd * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size) 1349219089Spjd * 1350219089Spjd * variable_size is the total size of all variable sized attributes 1351219089Spjd * passed to this function. It is not the total size of all 1352219089Spjd * variable size attributes that *may* exist on this object. 1353219089Spjd */ 1354219089Spjdvoid 1355219089Spjddmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow) 1356219089Spjd{ 1357219089Spjd uint64_t object; 1358219089Spjd sa_os_t *sa = tx->tx_objset->os_sa; 1359219089Spjd 1360219089Spjd ASSERT(hdl != NULL); 1361219089Spjd 1362219089Spjd object = sa_handle_object(hdl); 1363219089Spjd 1364219089Spjd dmu_tx_hold_bonus(tx, object); 1365219089Spjd 1366219089Spjd if (tx->tx_objset->os_sa->sa_master_obj == 0) 1367219089Spjd return; 1368219089Spjd 1369219089Spjd if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 || 1370219089Spjd tx->tx_objset->os_sa->sa_layout_attr_obj == 0) { 1371219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS); 1372219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY); 1373219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1374219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1375219089Spjd } 1376219089Spjd 1377219089Spjd dmu_tx_sa_registration_hold(sa, tx); 1378219089Spjd 1379219089Spjd if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj) 1380219089Spjd dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL); 1381219089Spjd 1382219089Spjd if (sa->sa_force_spill || may_grow || hdl->sa_spill) { 1383219089Spjd ASSERT(tx->tx_txg == 0); 1384219089Spjd dmu_tx_hold_spill(tx, object); 1385219089Spjd } else { 1386219089Spjd dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus; 1387219089Spjd dnode_t *dn; 1388219089Spjd 1389219089Spjd DB_DNODE_ENTER(db); 1390219089Spjd dn = DB_DNODE(db); 1391219089Spjd if (dn->dn_have_spill) { 1392219089Spjd ASSERT(tx->tx_txg == 0); 1393219089Spjd dmu_tx_hold_spill(tx, object); 1394219089Spjd } 1395219089Spjd DB_DNODE_EXIT(db); 1396219089Spjd } 1397219089Spjd} 1398