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. 23226944Smm * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24247406Smm * Copyright (c) 2013 by Delphix. All rights reserved. 25226944Smm */ 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; 51249643Smm if (dd != NULL) 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)); 57260764Savg tx->tx_start = gethrtime(); 58168404Spjd#ifdef ZFS_DEBUG 59168404Spjd refcount_create(&tx->tx_space_written); 60168404Spjd refcount_create(&tx->tx_space_freed); 61168404Spjd#endif 62168404Spjd return (tx); 63168404Spjd} 64168404Spjd 65168404Spjddmu_tx_t * 66168404Spjddmu_tx_create(objset_t *os) 67168404Spjd{ 68219089Spjd dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir); 69168404Spjd tx->tx_objset = os; 70219089Spjd tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset); 71168404Spjd return (tx); 72168404Spjd} 73168404Spjd 74168404Spjddmu_tx_t * 75168404Spjddmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg) 76168404Spjd{ 77168404Spjd dmu_tx_t *tx = dmu_tx_create_dd(NULL); 78168404Spjd 79168404Spjd ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg); 80168404Spjd tx->tx_pool = dp; 81168404Spjd tx->tx_txg = txg; 82168404Spjd tx->tx_anyobj = TRUE; 83168404Spjd 84168404Spjd return (tx); 85168404Spjd} 86168404Spjd 87168404Spjdint 88168404Spjddmu_tx_is_syncing(dmu_tx_t *tx) 89168404Spjd{ 90168404Spjd return (tx->tx_anyobj); 91168404Spjd} 92168404Spjd 93168404Spjdint 94168404Spjddmu_tx_private_ok(dmu_tx_t *tx) 95168404Spjd{ 96168404Spjd return (tx->tx_anyobj); 97168404Spjd} 98168404Spjd 99168404Spjdstatic dmu_tx_hold_t * 100168404Spjddmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object, 101168404Spjd enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2) 102168404Spjd{ 103168404Spjd dmu_tx_hold_t *txh; 104168404Spjd dnode_t *dn = NULL; 105168404Spjd int err; 106168404Spjd 107168404Spjd if (object != DMU_NEW_OBJECT) { 108219089Spjd err = dnode_hold(os, object, tx, &dn); 109168404Spjd if (err) { 110168404Spjd tx->tx_err = err; 111168404Spjd return (NULL); 112168404Spjd } 113168404Spjd 114168404Spjd if (err == 0 && tx->tx_txg != 0) { 115168404Spjd mutex_enter(&dn->dn_mtx); 116168404Spjd /* 117168404Spjd * dn->dn_assigned_txg == tx->tx_txg doesn't pose a 118168404Spjd * problem, but there's no way for it to happen (for 119168404Spjd * now, at least). 120168404Spjd */ 121168404Spjd ASSERT(dn->dn_assigned_txg == 0); 122168404Spjd dn->dn_assigned_txg = tx->tx_txg; 123168404Spjd (void) refcount_add(&dn->dn_tx_holds, tx); 124168404Spjd mutex_exit(&dn->dn_mtx); 125168404Spjd } 126168404Spjd } 127168404Spjd 128168404Spjd txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP); 129168404Spjd txh->txh_tx = tx; 130168404Spjd txh->txh_dnode = dn; 131168404Spjd#ifdef ZFS_DEBUG 132168404Spjd txh->txh_type = type; 133168404Spjd txh->txh_arg1 = arg1; 134168404Spjd txh->txh_arg2 = arg2; 135168404Spjd#endif 136168404Spjd list_insert_tail(&tx->tx_holds, txh); 137168404Spjd 138168404Spjd return (txh); 139168404Spjd} 140168404Spjd 141168404Spjdvoid 142168404Spjddmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object) 143168404Spjd{ 144168404Spjd /* 145168404Spjd * If we're syncing, they can manipulate any object anyhow, and 146168404Spjd * the hold on the dnode_t can cause problems. 147168404Spjd */ 148168404Spjd if (!dmu_tx_is_syncing(tx)) { 149168404Spjd (void) dmu_tx_hold_object_impl(tx, os, 150168404Spjd object, THT_NEWOBJECT, 0, 0); 151168404Spjd } 152168404Spjd} 153168404Spjd 154168404Spjdstatic int 155168404Spjddmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid) 156168404Spjd{ 157168404Spjd int err; 158168404Spjd dmu_buf_impl_t *db; 159168404Spjd 160168404Spjd rw_enter(&dn->dn_struct_rwlock, RW_READER); 161168404Spjd db = dbuf_hold_level(dn, level, blkid, FTAG); 162168404Spjd rw_exit(&dn->dn_struct_rwlock); 163168404Spjd if (db == NULL) 164249643Smm return (SET_ERROR(EIO)); 165185029Spjd err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH); 166168404Spjd dbuf_rele(db, FTAG); 167168404Spjd return (err); 168168404Spjd} 169168404Spjd 170209962Smmstatic void 171219089Spjddmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db, 172219089Spjd int level, uint64_t blkid, boolean_t freeable, uint64_t *history) 173209962Smm{ 174219089Spjd objset_t *os = dn->dn_objset; 175219089Spjd dsl_dataset_t *ds = os->os_dsl_dataset; 176219089Spjd int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 177219089Spjd dmu_buf_impl_t *parent = NULL; 178219089Spjd blkptr_t *bp = NULL; 179219089Spjd uint64_t space; 180209962Smm 181219089Spjd if (level >= dn->dn_nlevels || history[level] == blkid) 182209962Smm return; 183209962Smm 184219089Spjd history[level] = blkid; 185209962Smm 186219089Spjd space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift); 187219089Spjd 188219089Spjd if (db == NULL || db == dn->dn_dbuf) { 189219089Spjd ASSERT(level != 0); 190219089Spjd db = NULL; 191219089Spjd } else { 192219089Spjd ASSERT(DB_DNODE(db) == dn); 193219089Spjd ASSERT(db->db_level == level); 194219089Spjd ASSERT(db->db.db_size == space); 195219089Spjd ASSERT(db->db_blkid == blkid); 196219089Spjd bp = db->db_blkptr; 197219089Spjd parent = db->db_parent; 198209962Smm } 199209962Smm 200219089Spjd freeable = (bp && (freeable || 201219089Spjd dsl_dataset_block_freeable(ds, bp, bp->blk_birth))); 202209962Smm 203219089Spjd if (freeable) 204219089Spjd txh->txh_space_tooverwrite += space; 205219089Spjd else 206219089Spjd txh->txh_space_towrite += space; 207219089Spjd if (bp) 208219089Spjd txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp); 209219089Spjd 210219089Spjd dmu_tx_count_twig(txh, dn, parent, level + 1, 211219089Spjd blkid >> epbs, freeable, history); 212209962Smm} 213209962Smm 214168404Spjd/* ARGSUSED */ 215168404Spjdstatic void 216168404Spjddmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 217168404Spjd{ 218168404Spjd dnode_t *dn = txh->txh_dnode; 219168404Spjd uint64_t start, end, i; 220168404Spjd int min_bs, max_bs, min_ibs, max_ibs, epbs, bits; 221168404Spjd int err = 0; 222168404Spjd 223168404Spjd if (len == 0) 224168404Spjd return; 225168404Spjd 226168404Spjd min_bs = SPA_MINBLOCKSHIFT; 227168404Spjd max_bs = SPA_MAXBLOCKSHIFT; 228168404Spjd min_ibs = DN_MIN_INDBLKSHIFT; 229168404Spjd max_ibs = DN_MAX_INDBLKSHIFT; 230168404Spjd 231209962Smm if (dn) { 232219089Spjd uint64_t history[DN_MAX_LEVELS]; 233209962Smm int nlvls = dn->dn_nlevels; 234209962Smm int delta; 235168404Spjd 236209962Smm /* 237209962Smm * For i/o error checking, read the first and last level-0 238209962Smm * blocks (if they are not aligned), and all the level-1 blocks. 239209962Smm */ 240168404Spjd if (dn->dn_maxblkid == 0) { 241209962Smm delta = dn->dn_datablksz; 242209962Smm start = (off < dn->dn_datablksz) ? 0 : 1; 243209962Smm end = (off+len <= dn->dn_datablksz) ? 0 : 1; 244209962Smm if (start == 0 && (off > 0 || len < dn->dn_datablksz)) { 245209962Smm err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 246209962Smm if (err) 247209962Smm goto out; 248209962Smm delta -= off; 249209962Smm } 250168404Spjd } else { 251168404Spjd zio_t *zio = zio_root(dn->dn_objset->os_spa, 252168404Spjd NULL, NULL, ZIO_FLAG_CANFAIL); 253168404Spjd 254168404Spjd /* first level-0 block */ 255168404Spjd start = off >> dn->dn_datablkshift; 256168404Spjd if (P2PHASE(off, dn->dn_datablksz) || 257168404Spjd len < dn->dn_datablksz) { 258168404Spjd err = dmu_tx_check_ioerr(zio, dn, 0, start); 259168404Spjd if (err) 260168404Spjd goto out; 261168404Spjd } 262168404Spjd 263168404Spjd /* last level-0 block */ 264168404Spjd end = (off+len-1) >> dn->dn_datablkshift; 265219089Spjd if (end != start && end <= dn->dn_maxblkid && 266168404Spjd P2PHASE(off+len, dn->dn_datablksz)) { 267168404Spjd err = dmu_tx_check_ioerr(zio, dn, 0, end); 268168404Spjd if (err) 269168404Spjd goto out; 270168404Spjd } 271168404Spjd 272168404Spjd /* level-1 blocks */ 273209962Smm if (nlvls > 1) { 274209962Smm int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 275209962Smm for (i = (start>>shft)+1; i < end>>shft; i++) { 276168404Spjd err = dmu_tx_check_ioerr(zio, dn, 1, i); 277168404Spjd if (err) 278168404Spjd goto out; 279168404Spjd } 280168404Spjd } 281168404Spjd 282168404Spjd err = zio_wait(zio); 283168404Spjd if (err) 284168404Spjd goto out; 285209962Smm delta = P2NPHASE(off, dn->dn_datablksz); 286168404Spjd } 287168404Spjd 288247406Smm min_ibs = max_ibs = dn->dn_indblkshift; 289209962Smm if (dn->dn_maxblkid > 0) { 290209962Smm /* 291209962Smm * The blocksize can't change, 292209962Smm * so we can make a more precise estimate. 293209962Smm */ 294209962Smm ASSERT(dn->dn_datablkshift != 0); 295168404Spjd min_bs = max_bs = dn->dn_datablkshift; 296209962Smm } 297209962Smm 298209962Smm /* 299209962Smm * If this write is not off the end of the file 300209962Smm * we need to account for overwrites/unref. 301209962Smm */ 302219089Spjd if (start <= dn->dn_maxblkid) { 303219089Spjd for (int l = 0; l < DN_MAX_LEVELS; l++) 304219089Spjd history[l] = -1ULL; 305219089Spjd } 306209962Smm while (start <= dn->dn_maxblkid) { 307209962Smm dmu_buf_impl_t *db; 308209962Smm 309209962Smm rw_enter(&dn->dn_struct_rwlock, RW_READER); 310219089Spjd err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db); 311209962Smm rw_exit(&dn->dn_struct_rwlock); 312219089Spjd 313219089Spjd if (err) { 314219089Spjd txh->txh_tx->tx_err = err; 315219089Spjd return; 316209962Smm } 317219089Spjd 318219089Spjd dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE, 319219089Spjd history); 320209962Smm dbuf_rele(db, FTAG); 321209962Smm if (++start > end) { 322209962Smm /* 323209962Smm * Account for new indirects appearing 324209962Smm * before this IO gets assigned into a txg. 325209962Smm */ 326209962Smm bits = 64 - min_bs; 327209962Smm epbs = min_ibs - SPA_BLKPTRSHIFT; 328209962Smm for (bits -= epbs * (nlvls - 1); 329209962Smm bits >= 0; bits -= epbs) 330209962Smm txh->txh_fudge += 1ULL << max_ibs; 331209962Smm goto out; 332209962Smm } 333209962Smm off += delta; 334209962Smm if (len >= delta) 335209962Smm len -= delta; 336209962Smm delta = dn->dn_datablksz; 337209962Smm } 338168404Spjd } 339168404Spjd 340168404Spjd /* 341168404Spjd * 'end' is the last thing we will access, not one past. 342168404Spjd * This way we won't overflow when accessing the last byte. 343168404Spjd */ 344168404Spjd start = P2ALIGN(off, 1ULL << max_bs); 345168404Spjd end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1; 346168404Spjd txh->txh_space_towrite += end - start + 1; 347168404Spjd 348168404Spjd start >>= min_bs; 349168404Spjd end >>= min_bs; 350168404Spjd 351168404Spjd epbs = min_ibs - SPA_BLKPTRSHIFT; 352168404Spjd 353168404Spjd /* 354168404Spjd * The object contains at most 2^(64 - min_bs) blocks, 355168404Spjd * and each indirect level maps 2^epbs. 356168404Spjd */ 357168404Spjd for (bits = 64 - min_bs; bits >= 0; bits -= epbs) { 358168404Spjd start >>= epbs; 359168404Spjd end >>= epbs; 360209962Smm ASSERT3U(end, >=, start); 361209962Smm txh->txh_space_towrite += (end - start + 1) << max_ibs; 362209962Smm if (start != 0) { 363209962Smm /* 364209962Smm * We also need a new blkid=0 indirect block 365209962Smm * to reference any existing file data. 366209962Smm */ 367168404Spjd txh->txh_space_towrite += 1ULL << max_ibs; 368209962Smm } 369168404Spjd } 370168404Spjd 371209962Smmout: 372209962Smm if (txh->txh_space_towrite + txh->txh_space_tooverwrite > 373209962Smm 2 * DMU_MAX_ACCESS) 374249643Smm err = SET_ERROR(EFBIG); 375168404Spjd 376168404Spjd if (err) 377168404Spjd txh->txh_tx->tx_err = err; 378168404Spjd} 379168404Spjd 380168404Spjdstatic void 381168404Spjddmu_tx_count_dnode(dmu_tx_hold_t *txh) 382168404Spjd{ 383168404Spjd dnode_t *dn = txh->txh_dnode; 384219089Spjd dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset); 385168404Spjd uint64_t space = mdn->dn_datablksz + 386168404Spjd ((mdn->dn_nlevels-1) << mdn->dn_indblkshift); 387168404Spjd 388168404Spjd if (dn && dn->dn_dbuf->db_blkptr && 389168404Spjd dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, 390219089Spjd dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) { 391168404Spjd txh->txh_space_tooverwrite += space; 392209962Smm txh->txh_space_tounref += space; 393168404Spjd } else { 394168404Spjd txh->txh_space_towrite += space; 395185029Spjd if (dn && dn->dn_dbuf->db_blkptr) 396185029Spjd txh->txh_space_tounref += space; 397168404Spjd } 398168404Spjd} 399168404Spjd 400168404Spjdvoid 401168404Spjddmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len) 402168404Spjd{ 403168404Spjd dmu_tx_hold_t *txh; 404168404Spjd 405168404Spjd ASSERT(tx->tx_txg == 0); 406168404Spjd ASSERT(len < DMU_MAX_ACCESS); 407168404Spjd ASSERT(len == 0 || UINT64_MAX - off >= len - 1); 408168404Spjd 409168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 410168404Spjd object, THT_WRITE, off, len); 411168404Spjd if (txh == NULL) 412168404Spjd return; 413168404Spjd 414168404Spjd dmu_tx_count_write(txh, off, len); 415168404Spjd dmu_tx_count_dnode(txh); 416168404Spjd} 417168404Spjd 418168404Spjdstatic void 419168404Spjddmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) 420168404Spjd{ 421185029Spjd uint64_t blkid, nblks, lastblk; 422185029Spjd uint64_t space = 0, unref = 0, skipped = 0; 423168404Spjd dnode_t *dn = txh->txh_dnode; 424168404Spjd dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 425168404Spjd spa_t *spa = txh->txh_tx->tx_pool->dp_spa; 426185029Spjd int epbs; 427243674Smm uint64_t l0span = 0, nl1blks = 0; 428168404Spjd 429185029Spjd if (dn->dn_nlevels == 0) 430168404Spjd return; 431168404Spjd 432168404Spjd /* 433185029Spjd * The struct_rwlock protects us against dn_nlevels 434168404Spjd * changing, in case (against all odds) we manage to dirty & 435168404Spjd * sync out the changes after we check for being dirty. 436219089Spjd * Also, dbuf_hold_impl() wants us to have the struct_rwlock. 437168404Spjd */ 438168404Spjd rw_enter(&dn->dn_struct_rwlock, RW_READER); 439185029Spjd epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 440185029Spjd if (dn->dn_maxblkid == 0) { 441168404Spjd if (off == 0 && len >= dn->dn_datablksz) { 442168404Spjd blkid = 0; 443168404Spjd nblks = 1; 444168404Spjd } else { 445168404Spjd rw_exit(&dn->dn_struct_rwlock); 446168404Spjd return; 447168404Spjd } 448168404Spjd } else { 449168404Spjd blkid = off >> dn->dn_datablkshift; 450185029Spjd nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift; 451168404Spjd 452262084Savg if (blkid > dn->dn_maxblkid) { 453168404Spjd rw_exit(&dn->dn_struct_rwlock); 454168404Spjd return; 455168404Spjd } 456185029Spjd if (blkid + nblks > dn->dn_maxblkid) 457262084Savg nblks = dn->dn_maxblkid - blkid + 1; 458168404Spjd 459168404Spjd } 460243674Smm l0span = nblks; /* save for later use to calc level > 1 overhead */ 461185029Spjd if (dn->dn_nlevels == 1) { 462168404Spjd int i; 463168404Spjd for (i = 0; i < nblks; i++) { 464168404Spjd blkptr_t *bp = dn->dn_phys->dn_blkptr; 465185029Spjd ASSERT3U(blkid + i, <, dn->dn_nblkptr); 466168404Spjd bp += blkid + i; 467219089Spjd if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) { 468168404Spjd dprintf_bp(bp, "can free old%s", ""); 469219089Spjd space += bp_get_dsize(spa, bp); 470168404Spjd } 471185029Spjd unref += BP_GET_ASIZE(bp); 472168404Spjd } 473243674Smm nl1blks = 1; 474168404Spjd nblks = 0; 475168404Spjd } 476168404Spjd 477185029Spjd lastblk = blkid + nblks - 1; 478168404Spjd while (nblks) { 479168404Spjd dmu_buf_impl_t *dbuf; 480185029Spjd uint64_t ibyte, new_blkid; 481185029Spjd int epb = 1 << epbs; 482185029Spjd int err, i, blkoff, tochk; 483185029Spjd blkptr_t *bp; 484168404Spjd 485185029Spjd ibyte = blkid << dn->dn_datablkshift; 486185029Spjd err = dnode_next_offset(dn, 487185029Spjd DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0); 488185029Spjd new_blkid = ibyte >> dn->dn_datablkshift; 489185029Spjd if (err == ESRCH) { 490185029Spjd skipped += (lastblk >> epbs) - (blkid >> epbs) + 1; 491185029Spjd break; 492185029Spjd } 493185029Spjd if (err) { 494185029Spjd txh->txh_tx->tx_err = err; 495185029Spjd break; 496185029Spjd } 497185029Spjd if (new_blkid > lastblk) { 498185029Spjd skipped += (lastblk >> epbs) - (blkid >> epbs) + 1; 499185029Spjd break; 500185029Spjd } 501168404Spjd 502185029Spjd if (new_blkid > blkid) { 503185029Spjd ASSERT((new_blkid >> epbs) > (blkid >> epbs)); 504185029Spjd skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1; 505185029Spjd nblks -= new_blkid - blkid; 506185029Spjd blkid = new_blkid; 507185029Spjd } 508185029Spjd blkoff = P2PHASE(blkid, epb); 509185029Spjd tochk = MIN(epb - blkoff, nblks); 510168404Spjd 511219089Spjd err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf); 512219089Spjd if (err) { 513219089Spjd txh->txh_tx->tx_err = err; 514219089Spjd break; 515219089Spjd } 516168404Spjd 517185029Spjd txh->txh_memory_tohold += dbuf->db.db_size; 518219089Spjd 519219089Spjd /* 520219089Spjd * We don't check memory_tohold against DMU_MAX_ACCESS because 521219089Spjd * memory_tohold is an over-estimation (especially the >L1 522219089Spjd * indirect blocks), so it could fail. Callers should have 523219089Spjd * already verified that they will not be holding too much 524219089Spjd * memory. 525219089Spjd */ 526219089Spjd 527185029Spjd err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL); 528185029Spjd if (err != 0) { 529168404Spjd txh->txh_tx->tx_err = err; 530185029Spjd dbuf_rele(dbuf, FTAG); 531168404Spjd break; 532168404Spjd } 533168404Spjd 534185029Spjd bp = dbuf->db.db_data; 535185029Spjd bp += blkoff; 536185029Spjd 537185029Spjd for (i = 0; i < tochk; i++) { 538219089Spjd if (dsl_dataset_block_freeable(ds, &bp[i], 539219089Spjd bp[i].blk_birth)) { 540185029Spjd dprintf_bp(&bp[i], "can free old%s", ""); 541219089Spjd space += bp_get_dsize(spa, &bp[i]); 542185029Spjd } 543185029Spjd unref += BP_GET_ASIZE(bp); 544185029Spjd } 545185029Spjd dbuf_rele(dbuf, FTAG); 546185029Spjd 547243674Smm ++nl1blks; 548168404Spjd blkid += tochk; 549168404Spjd nblks -= tochk; 550168404Spjd } 551168404Spjd rw_exit(&dn->dn_struct_rwlock); 552168404Spjd 553243674Smm /* 554243674Smm * Add in memory requirements of higher-level indirects. 555243674Smm * This assumes a worst-possible scenario for dn_nlevels and a 556243674Smm * worst-possible distribution of l1-blocks over the region to free. 557243674Smm */ 558243674Smm { 559243674Smm uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs); 560243674Smm int level = 2; 561243674Smm /* 562243674Smm * Here we don't use DN_MAX_LEVEL, but calculate it with the 563243674Smm * given datablkshift and indblkshift. This makes the 564243674Smm * difference between 19 and 8 on large files. 565243674Smm */ 566243674Smm int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) / 567243674Smm (dn->dn_indblkshift - SPA_BLKPTRSHIFT); 568243674Smm 569243674Smm while (level++ < maxlevel) { 570243674Smm txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1) 571243674Smm << dn->dn_indblkshift; 572243674Smm blkcnt = 1 + (blkcnt >> epbs); 573243674Smm } 574243674Smm } 575243674Smm 576185029Spjd /* account for new level 1 indirect blocks that might show up */ 577185029Spjd if (skipped > 0) { 578185029Spjd txh->txh_fudge += skipped << dn->dn_indblkshift; 579185029Spjd skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs); 580185029Spjd txh->txh_memory_tohold += skipped << dn->dn_indblkshift; 581185029Spjd } 582168404Spjd txh->txh_space_tofree += space; 583185029Spjd txh->txh_space_tounref += unref; 584168404Spjd} 585168404Spjd 586168404Spjdvoid 587168404Spjddmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len) 588168404Spjd{ 589168404Spjd dmu_tx_hold_t *txh; 590168404Spjd dnode_t *dn; 591260722Savg int err; 592168404Spjd zio_t *zio; 593168404Spjd 594168404Spjd ASSERT(tx->tx_txg == 0); 595168404Spjd 596168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 597168404Spjd object, THT_FREE, off, len); 598168404Spjd if (txh == NULL) 599168404Spjd return; 600168404Spjd dn = txh->txh_dnode; 601260764Savg dmu_tx_count_dnode(txh); 602168404Spjd 603168404Spjd if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz) 604168404Spjd return; 605168404Spjd if (len == DMU_OBJECT_END) 606168404Spjd len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off; 607168404Spjd 608260722Savg 609168404Spjd /* 610260722Savg * For i/o error checking, we read the first and last level-0 611260722Savg * blocks if they are not aligned, and all the level-1 blocks. 612260722Savg * 613260722Savg * Note: dbuf_free_range() assumes that we have not instantiated 614260722Savg * any level-0 dbufs that will be completely freed. Therefore we must 615260722Savg * exercise care to not read or count the first and last blocks 616260722Savg * if they are blocksize-aligned. 617168404Spjd */ 618260722Savg if (dn->dn_datablkshift == 0) { 619260722Savg if (off != 0 || len < dn->dn_datablksz) 620260722Savg dmu_tx_count_write(txh, 0, dn->dn_datablksz); 621260722Savg } else { 622260722Savg /* first block will be modified if it is not aligned */ 623260722Savg if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift)) 624260722Savg dmu_tx_count_write(txh, off, 1); 625260722Savg /* last block will be modified if it is not aligned */ 626260722Savg if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift)) 627260722Savg dmu_tx_count_write(txh, off+len, 1); 628260722Savg } 629260722Savg 630260722Savg /* 631260722Savg * Check level-1 blocks. 632260722Savg */ 633168404Spjd if (dn->dn_nlevels > 1) { 634260722Savg int shift = dn->dn_datablkshift + dn->dn_indblkshift - 635168404Spjd SPA_BLKPTRSHIFT; 636260722Savg uint64_t start = off >> shift; 637260722Savg uint64_t end = (off + len) >> shift; 638168404Spjd 639260722Savg ASSERT(dn->dn_indblkshift != 0); 640260722Savg 641262180Savg /* 642262180Savg * dnode_reallocate() can result in an object with indirect 643262180Savg * blocks having an odd data block size. In this case, 644262180Savg * just check the single block. 645262180Savg */ 646262180Savg if (dn->dn_datablkshift == 0) 647262180Savg start = end = 0; 648262180Savg 649168404Spjd zio = zio_root(tx->tx_pool->dp_spa, 650168404Spjd NULL, NULL, ZIO_FLAG_CANFAIL); 651260722Savg for (uint64_t i = start; i <= end; i++) { 652168404Spjd uint64_t ibyte = i << shift; 653185029Spjd err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0); 654168404Spjd i = ibyte >> shift; 655168404Spjd if (err == ESRCH) 656168404Spjd break; 657168404Spjd if (err) { 658168404Spjd tx->tx_err = err; 659168404Spjd return; 660168404Spjd } 661168404Spjd 662168404Spjd err = dmu_tx_check_ioerr(zio, dn, 1, i); 663168404Spjd if (err) { 664168404Spjd tx->tx_err = err; 665168404Spjd return; 666168404Spjd } 667168404Spjd } 668168404Spjd err = zio_wait(zio); 669168404Spjd if (err) { 670168404Spjd tx->tx_err = err; 671168404Spjd return; 672168404Spjd } 673168404Spjd } 674168404Spjd 675168404Spjd dmu_tx_count_free(txh, off, len); 676168404Spjd} 677168404Spjd 678168404Spjdvoid 679209962Smmdmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name) 680168404Spjd{ 681168404Spjd dmu_tx_hold_t *txh; 682168404Spjd dnode_t *dn; 683168404Spjd uint64_t nblocks; 684168404Spjd int epbs, err; 685168404Spjd 686168404Spjd ASSERT(tx->tx_txg == 0); 687168404Spjd 688168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 689168404Spjd object, THT_ZAP, add, (uintptr_t)name); 690168404Spjd if (txh == NULL) 691168404Spjd return; 692168404Spjd dn = txh->txh_dnode; 693168404Spjd 694168404Spjd dmu_tx_count_dnode(txh); 695168404Spjd 696168404Spjd if (dn == NULL) { 697168404Spjd /* 698168404Spjd * We will be able to fit a new object's entries into one leaf 699168404Spjd * block. So there will be at most 2 blocks total, 700168404Spjd * including the header block. 701168404Spjd */ 702168404Spjd dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift); 703168404Spjd return; 704168404Spjd } 705168404Spjd 706243674Smm ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP); 707168404Spjd 708168404Spjd if (dn->dn_maxblkid == 0 && !add) { 709226944Smm blkptr_t *bp; 710226944Smm 711168404Spjd /* 712168404Spjd * If there is only one block (i.e. this is a micro-zap) 713168404Spjd * and we are not adding anything, the accounting is simple. 714168404Spjd */ 715168404Spjd err = dmu_tx_check_ioerr(NULL, dn, 0, 0); 716168404Spjd if (err) { 717168404Spjd tx->tx_err = err; 718168404Spjd return; 719168404Spjd } 720168404Spjd 721168404Spjd /* 722168404Spjd * Use max block size here, since we don't know how much 723168404Spjd * the size will change between now and the dbuf dirty call. 724168404Spjd */ 725226944Smm bp = &dn->dn_phys->dn_blkptr[0]; 726168404Spjd if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, 727226944Smm bp, bp->blk_birth)) 728168404Spjd txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE; 729226944Smm else 730168404Spjd txh->txh_space_towrite += SPA_MAXBLOCKSIZE; 731226944Smm if (!BP_IS_HOLE(bp)) 732209093Smm txh->txh_space_tounref += SPA_MAXBLOCKSIZE; 733168404Spjd return; 734168404Spjd } 735168404Spjd 736168404Spjd if (dn->dn_maxblkid > 0 && name) { 737168404Spjd /* 738168404Spjd * access the name in this fat-zap so that we'll check 739168404Spjd * for i/o errors to the leaf blocks, etc. 740168404Spjd */ 741219089Spjd err = zap_lookup(dn->dn_objset, dn->dn_object, name, 742168404Spjd 8, 0, NULL); 743168404Spjd if (err == EIO) { 744168404Spjd tx->tx_err = err; 745168404Spjd return; 746168404Spjd } 747168404Spjd } 748168404Spjd 749219089Spjd err = zap_count_write(dn->dn_objset, dn->dn_object, name, add, 750209962Smm &txh->txh_space_towrite, &txh->txh_space_tooverwrite); 751168404Spjd 752168404Spjd /* 753168404Spjd * If the modified blocks are scattered to the four winds, 754168404Spjd * we'll have to modify an indirect twig for each. 755168404Spjd */ 756168404Spjd epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 757168404Spjd for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs) 758209962Smm if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj) 759209962Smm txh->txh_space_towrite += 3 << dn->dn_indblkshift; 760209962Smm else 761209962Smm txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift; 762168404Spjd} 763168404Spjd 764168404Spjdvoid 765168404Spjddmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object) 766168404Spjd{ 767168404Spjd dmu_tx_hold_t *txh; 768168404Spjd 769168404Spjd ASSERT(tx->tx_txg == 0); 770168404Spjd 771168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 772168404Spjd object, THT_BONUS, 0, 0); 773168404Spjd if (txh) 774168404Spjd dmu_tx_count_dnode(txh); 775168404Spjd} 776168404Spjd 777168404Spjdvoid 778168404Spjddmu_tx_hold_space(dmu_tx_t *tx, uint64_t space) 779168404Spjd{ 780168404Spjd dmu_tx_hold_t *txh; 781168404Spjd ASSERT(tx->tx_txg == 0); 782168404Spjd 783168404Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, 784168404Spjd DMU_NEW_OBJECT, THT_SPACE, space, 0); 785168404Spjd 786168404Spjd txh->txh_space_towrite += space; 787168404Spjd} 788168404Spjd 789168404Spjdint 790168404Spjddmu_tx_holds(dmu_tx_t *tx, uint64_t object) 791168404Spjd{ 792168404Spjd dmu_tx_hold_t *txh; 793168404Spjd int holds = 0; 794168404Spjd 795168404Spjd /* 796168404Spjd * By asserting that the tx is assigned, we're counting the 797168404Spjd * number of dn_tx_holds, which is the same as the number of 798168404Spjd * dn_holds. Otherwise, we'd be counting dn_holds, but 799168404Spjd * dn_tx_holds could be 0. 800168404Spjd */ 801168404Spjd ASSERT(tx->tx_txg != 0); 802168404Spjd 803168404Spjd /* if (tx->tx_anyobj == TRUE) */ 804168404Spjd /* return (0); */ 805168404Spjd 806168404Spjd for (txh = list_head(&tx->tx_holds); txh; 807168404Spjd txh = list_next(&tx->tx_holds, txh)) { 808168404Spjd if (txh->txh_dnode && txh->txh_dnode->dn_object == object) 809168404Spjd holds++; 810168404Spjd } 811168404Spjd 812168404Spjd return (holds); 813168404Spjd} 814168404Spjd 815168404Spjd#ifdef ZFS_DEBUG 816168404Spjdvoid 817168404Spjddmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db) 818168404Spjd{ 819168404Spjd dmu_tx_hold_t *txh; 820168404Spjd int match_object = FALSE, match_offset = FALSE; 821219089Spjd dnode_t *dn; 822168404Spjd 823219089Spjd DB_DNODE_ENTER(db); 824219089Spjd dn = DB_DNODE(db); 825168404Spjd ASSERT(tx->tx_txg != 0); 826219089Spjd ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset); 827168404Spjd ASSERT3U(dn->dn_object, ==, db->db.db_object); 828168404Spjd 829219089Spjd if (tx->tx_anyobj) { 830219089Spjd DB_DNODE_EXIT(db); 831168404Spjd return; 832219089Spjd } 833168404Spjd 834168404Spjd /* XXX No checking on the meta dnode for now */ 835219089Spjd if (db->db.db_object == DMU_META_DNODE_OBJECT) { 836219089Spjd DB_DNODE_EXIT(db); 837168404Spjd return; 838219089Spjd } 839168404Spjd 840168404Spjd for (txh = list_head(&tx->tx_holds); txh; 841168404Spjd txh = list_next(&tx->tx_holds, txh)) { 842168404Spjd ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg); 843168404Spjd if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT) 844168404Spjd match_object = TRUE; 845168404Spjd if (txh->txh_dnode == NULL || txh->txh_dnode == dn) { 846168404Spjd int datablkshift = dn->dn_datablkshift ? 847168404Spjd dn->dn_datablkshift : SPA_MAXBLOCKSHIFT; 848168404Spjd int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 849168404Spjd int shift = datablkshift + epbs * db->db_level; 850168404Spjd uint64_t beginblk = shift >= 64 ? 0 : 851168404Spjd (txh->txh_arg1 >> shift); 852168404Spjd uint64_t endblk = shift >= 64 ? 0 : 853168404Spjd ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift); 854168404Spjd uint64_t blkid = db->db_blkid; 855168404Spjd 856168404Spjd /* XXX txh_arg2 better not be zero... */ 857168404Spjd 858168404Spjd dprintf("found txh type %x beginblk=%llx endblk=%llx\n", 859168404Spjd txh->txh_type, beginblk, endblk); 860168404Spjd 861168404Spjd switch (txh->txh_type) { 862168404Spjd case THT_WRITE: 863168404Spjd if (blkid >= beginblk && blkid <= endblk) 864168404Spjd match_offset = TRUE; 865168404Spjd /* 866168404Spjd * We will let this hold work for the bonus 867219089Spjd * or spill buffer so that we don't need to 868219089Spjd * hold it when creating a new object. 869168404Spjd */ 870219089Spjd if (blkid == DMU_BONUS_BLKID || 871219089Spjd blkid == DMU_SPILL_BLKID) 872168404Spjd match_offset = TRUE; 873168404Spjd /* 874168404Spjd * They might have to increase nlevels, 875168404Spjd * thus dirtying the new TLIBs. Or the 876168404Spjd * might have to change the block size, 877168404Spjd * thus dirying the new lvl=0 blk=0. 878168404Spjd */ 879168404Spjd if (blkid == 0) 880168404Spjd match_offset = TRUE; 881168404Spjd break; 882168404Spjd case THT_FREE: 883185029Spjd /* 884185029Spjd * We will dirty all the level 1 blocks in 885185029Spjd * the free range and perhaps the first and 886185029Spjd * last level 0 block. 887185029Spjd */ 888185029Spjd if (blkid >= beginblk && (blkid <= endblk || 889185029Spjd txh->txh_arg2 == DMU_OBJECT_END)) 890168404Spjd match_offset = TRUE; 891168404Spjd break; 892219089Spjd case THT_SPILL: 893219089Spjd if (blkid == DMU_SPILL_BLKID) 894219089Spjd match_offset = TRUE; 895219089Spjd break; 896168404Spjd case THT_BONUS: 897219089Spjd if (blkid == DMU_BONUS_BLKID) 898168404Spjd match_offset = TRUE; 899168404Spjd break; 900168404Spjd case THT_ZAP: 901168404Spjd match_offset = TRUE; 902168404Spjd break; 903168404Spjd case THT_NEWOBJECT: 904168404Spjd match_object = TRUE; 905168404Spjd break; 906168404Spjd default: 907168404Spjd ASSERT(!"bad txh_type"); 908168404Spjd } 909168404Spjd } 910219089Spjd if (match_object && match_offset) { 911219089Spjd DB_DNODE_EXIT(db); 912168404Spjd return; 913219089Spjd } 914168404Spjd } 915219089Spjd DB_DNODE_EXIT(db); 916168404Spjd panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n", 917168404Spjd (u_longlong_t)db->db.db_object, db->db_level, 918168404Spjd (u_longlong_t)db->db_blkid); 919168404Spjd} 920168404Spjd#endif 921168404Spjd 922260764Savg/* 923260764Savg * If we can't do 10 iops, something is wrong. Let us go ahead 924260764Savg * and hit zfs_dirty_data_max. 925260764Savg */ 926260764Savghrtime_t zfs_delay_max_ns = MSEC2NSEC(100); 927260764Savgint zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */ 928260764Savg 929260764Savg/* 930260764Savg * We delay transactions when we've determined that the backend storage 931260764Savg * isn't able to accommodate the rate of incoming writes. 932260764Savg * 933260764Savg * If there is already a transaction waiting, we delay relative to when 934260764Savg * that transaction finishes waiting. This way the calculated min_time 935260764Savg * is independent of the number of threads concurrently executing 936260764Savg * transactions. 937260764Savg * 938260764Savg * If we are the only waiter, wait relative to when the transaction 939260764Savg * started, rather than the current time. This credits the transaction for 940260764Savg * "time already served", e.g. reading indirect blocks. 941260764Savg * 942260764Savg * The minimum time for a transaction to take is calculated as: 943260764Savg * min_time = scale * (dirty - min) / (max - dirty) 944260764Savg * min_time is then capped at zfs_delay_max_ns. 945260764Savg * 946260764Savg * The delay has two degrees of freedom that can be adjusted via tunables. 947260764Savg * The percentage of dirty data at which we start to delay is defined by 948260764Savg * zfs_delay_min_dirty_percent. This should typically be at or above 949260764Savg * zfs_vdev_async_write_active_max_dirty_percent so that we only start to 950260764Savg * delay after writing at full speed has failed to keep up with the incoming 951260764Savg * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly 952260764Savg * speaking, this variable determines the amount of delay at the midpoint of 953260764Savg * the curve. 954260764Savg * 955260764Savg * delay 956260764Savg * 10ms +-------------------------------------------------------------*+ 957260764Savg * | *| 958260764Savg * 9ms + *+ 959260764Savg * | *| 960260764Savg * 8ms + *+ 961260764Savg * | * | 962260764Savg * 7ms + * + 963260764Savg * | * | 964260764Savg * 6ms + * + 965260764Savg * | * | 966260764Savg * 5ms + * + 967260764Savg * | * | 968260764Savg * 4ms + * + 969260764Savg * | * | 970260764Savg * 3ms + * + 971260764Savg * | * | 972260764Savg * 2ms + (midpoint) * + 973260764Savg * | | ** | 974260764Savg * 1ms + v *** + 975260764Savg * | zfs_delay_scale ----------> ******** | 976260764Savg * 0 +-------------------------------------*********----------------+ 977260764Savg * 0% <- zfs_dirty_data_max -> 100% 978260764Savg * 979260764Savg * Note that since the delay is added to the outstanding time remaining on the 980260764Savg * most recent transaction, the delay is effectively the inverse of IOPS. 981260764Savg * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve 982260764Savg * was chosen such that small changes in the amount of accumulated dirty data 983260764Savg * in the first 3/4 of the curve yield relatively small differences in the 984260764Savg * amount of delay. 985260764Savg * 986260764Savg * The effects can be easier to understand when the amount of delay is 987260764Savg * represented on a log scale: 988260764Savg * 989260764Savg * delay 990260764Savg * 100ms +-------------------------------------------------------------++ 991260764Savg * + + 992260764Savg * | | 993260764Savg * + *+ 994260764Savg * 10ms + *+ 995260764Savg * + ** + 996260764Savg * | (midpoint) ** | 997260764Savg * + | ** + 998260764Savg * 1ms + v **** + 999260764Savg * + zfs_delay_scale ----------> ***** + 1000260764Savg * | **** | 1001260764Savg * + **** + 1002260764Savg * 100us + ** + 1003260764Savg * + * + 1004260764Savg * | * | 1005260764Savg * + * + 1006260764Savg * 10us + * + 1007260764Savg * + + 1008260764Savg * | | 1009260764Savg * + + 1010260764Savg * +--------------------------------------------------------------+ 1011260764Savg * 0% <- zfs_dirty_data_max -> 100% 1012260764Savg * 1013260764Savg * Note here that only as the amount of dirty data approaches its limit does 1014260764Savg * the delay start to increase rapidly. The goal of a properly tuned system 1015260764Savg * should be to keep the amount of dirty data out of that range by first 1016260764Savg * ensuring that the appropriate limits are set for the I/O scheduler to reach 1017260764Savg * optimal throughput on the backend storage, and then by changing the value 1018260764Savg * of zfs_delay_scale to increase the steepness of the curve. 1019260764Savg */ 1020260764Savgstatic void 1021260764Savgdmu_tx_delay(dmu_tx_t *tx, uint64_t dirty) 1022260764Savg{ 1023260764Savg dsl_pool_t *dp = tx->tx_pool; 1024260764Savg uint64_t delay_min_bytes = 1025260764Savg zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100; 1026260764Savg hrtime_t wakeup, min_tx_time, now; 1027260764Savg 1028260764Savg if (dirty <= delay_min_bytes) 1029260764Savg return; 1030260764Savg 1031260764Savg /* 1032260764Savg * The caller has already waited until we are under the max. 1033260764Savg * We make them pass us the amount of dirty data so we don't 1034260764Savg * have to handle the case of it being >= the max, which could 1035260764Savg * cause a divide-by-zero if it's == the max. 1036260764Savg */ 1037260764Savg ASSERT3U(dirty, <, zfs_dirty_data_max); 1038260764Savg 1039260764Savg now = gethrtime(); 1040260764Savg min_tx_time = zfs_delay_scale * 1041260764Savg (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty); 1042260764Savg if (now > tx->tx_start + min_tx_time) 1043260764Savg return; 1044260764Savg 1045260764Savg min_tx_time = MIN(min_tx_time, zfs_delay_max_ns); 1046260764Savg 1047260764Savg DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty, 1048260764Savg uint64_t, min_tx_time); 1049260764Savg 1050260764Savg mutex_enter(&dp->dp_lock); 1051260764Savg wakeup = MAX(tx->tx_start + min_tx_time, 1052260764Savg dp->dp_last_wakeup + min_tx_time); 1053260764Savg dp->dp_last_wakeup = wakeup; 1054260764Savg mutex_exit(&dp->dp_lock); 1055260764Savg 1056260764Savg#ifdef _KERNEL 1057260764Savg#ifdef illumos 1058260764Savg mutex_enter(&curthread->t_delay_lock); 1059260764Savg while (cv_timedwait_hires(&curthread->t_delay_cv, 1060260764Savg &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns, 1061260764Savg CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0) 1062260764Savg continue; 1063260764Savg mutex_exit(&curthread->t_delay_lock); 1064260764Savg#else 1065264505Sjhb int timo; 1066264505Sjhb 1067260764Savg /* XXX High resolution callouts are not available */ 1068260764Savg ASSERT(wakeup >= now); 1069264505Sjhb timo = NSEC_TO_TICK(wakeup - now); 1070264505Sjhb if (timo != 0) 1071264505Sjhb pause("dmu_tx_delay", timo); 1072260764Savg#endif 1073260764Savg#else 1074260764Savg hrtime_t delta = wakeup - gethrtime(); 1075260764Savg struct timespec ts; 1076260764Savg ts.tv_sec = delta / NANOSEC; 1077260764Savg ts.tv_nsec = delta % NANOSEC; 1078260764Savg (void) nanosleep(&ts, NULL); 1079260764Savg#endif 1080260764Savg} 1081260764Savg 1082168404Spjdstatic int 1083249643Smmdmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how) 1084168404Spjd{ 1085168404Spjd dmu_tx_hold_t *txh; 1086185029Spjd spa_t *spa = tx->tx_pool->dp_spa; 1087185029Spjd uint64_t memory, asize, fsize, usize; 1088185029Spjd uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge; 1089168404Spjd 1090243674Smm ASSERT0(tx->tx_txg); 1091185029Spjd 1092168404Spjd if (tx->tx_err) 1093168404Spjd return (tx->tx_err); 1094168404Spjd 1095185029Spjd if (spa_suspended(spa)) { 1096185029Spjd /* 1097185029Spjd * If the user has indicated a blocking failure mode 1098185029Spjd * then return ERESTART which will block in dmu_tx_wait(). 1099185029Spjd * Otherwise, return EIO so that an error can get 1100185029Spjd * propagated back to the VOP calls. 1101185029Spjd * 1102185029Spjd * Note that we always honor the txg_how flag regardless 1103185029Spjd * of the failuremode setting. 1104185029Spjd */ 1105185029Spjd if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE && 1106185029Spjd txg_how != TXG_WAIT) 1107249643Smm return (SET_ERROR(EIO)); 1108185029Spjd 1109249643Smm return (SET_ERROR(ERESTART)); 1110185029Spjd } 1111185029Spjd 1112260764Savg if (!tx->tx_waited && 1113260764Savg dsl_pool_need_dirty_delay(tx->tx_pool)) { 1114260764Savg tx->tx_wait_dirty = B_TRUE; 1115260764Savg return (SET_ERROR(ERESTART)); 1116260764Savg } 1117260764Savg 1118168404Spjd tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh); 1119168404Spjd tx->tx_needassign_txh = NULL; 1120168404Spjd 1121168404Spjd /* 1122168404Spjd * NB: No error returns are allowed after txg_hold_open, but 1123168404Spjd * before processing the dnode holds, due to the 1124168404Spjd * dmu_tx_unassign() logic. 1125168404Spjd */ 1126168404Spjd 1127185029Spjd towrite = tofree = tooverwrite = tounref = tohold = fudge = 0; 1128168404Spjd for (txh = list_head(&tx->tx_holds); txh; 1129168404Spjd txh = list_next(&tx->tx_holds, txh)) { 1130168404Spjd dnode_t *dn = txh->txh_dnode; 1131168404Spjd if (dn != NULL) { 1132168404Spjd mutex_enter(&dn->dn_mtx); 1133168404Spjd if (dn->dn_assigned_txg == tx->tx_txg - 1) { 1134168404Spjd mutex_exit(&dn->dn_mtx); 1135168404Spjd tx->tx_needassign_txh = txh; 1136249643Smm return (SET_ERROR(ERESTART)); 1137168404Spjd } 1138168404Spjd if (dn->dn_assigned_txg == 0) 1139168404Spjd dn->dn_assigned_txg = tx->tx_txg; 1140168404Spjd ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1141168404Spjd (void) refcount_add(&dn->dn_tx_holds, tx); 1142168404Spjd mutex_exit(&dn->dn_mtx); 1143168404Spjd } 1144168404Spjd towrite += txh->txh_space_towrite; 1145168404Spjd tofree += txh->txh_space_tofree; 1146168404Spjd tooverwrite += txh->txh_space_tooverwrite; 1147185029Spjd tounref += txh->txh_space_tounref; 1148185029Spjd tohold += txh->txh_memory_tohold; 1149185029Spjd fudge += txh->txh_fudge; 1150168404Spjd } 1151168404Spjd 1152168404Spjd /* 1153168404Spjd * If a snapshot has been taken since we made our estimates, 1154168404Spjd * assume that we won't be able to free or overwrite anything. 1155168404Spjd */ 1156168404Spjd if (tx->tx_objset && 1157219089Spjd dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) > 1158168404Spjd tx->tx_lastsnap_txg) { 1159168404Spjd towrite += tooverwrite; 1160168404Spjd tooverwrite = tofree = 0; 1161168404Spjd } 1162168404Spjd 1163185029Spjd /* needed allocation: worst-case estimate of write space */ 1164185029Spjd asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite); 1165185029Spjd /* freed space estimate: worst-case overwrite + free estimate */ 1166168404Spjd fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree; 1167185029Spjd /* convert unrefd space to worst-case estimate */ 1168185029Spjd usize = spa_get_asize(tx->tx_pool->dp_spa, tounref); 1169185029Spjd /* calculate memory footprint estimate */ 1170185029Spjd memory = towrite + tooverwrite + tohold; 1171168404Spjd 1172168404Spjd#ifdef ZFS_DEBUG 1173185029Spjd /* 1174185029Spjd * Add in 'tohold' to account for our dirty holds on this memory 1175185029Spjd * XXX - the "fudge" factor is to account for skipped blocks that 1176185029Spjd * we missed because dnode_next_offset() misses in-core-only blocks. 1177185029Spjd */ 1178185029Spjd tx->tx_space_towrite = asize + 1179185029Spjd spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge); 1180168404Spjd tx->tx_space_tofree = tofree; 1181168404Spjd tx->tx_space_tooverwrite = tooverwrite; 1182185029Spjd tx->tx_space_tounref = tounref; 1183168404Spjd#endif 1184168404Spjd 1185168404Spjd if (tx->tx_dir && asize != 0) { 1186185029Spjd int err = dsl_dir_tempreserve_space(tx->tx_dir, memory, 1187185029Spjd asize, fsize, usize, &tx->tx_tempreserve_cookie, tx); 1188168404Spjd if (err) 1189168404Spjd return (err); 1190168404Spjd } 1191168404Spjd 1192168404Spjd return (0); 1193168404Spjd} 1194168404Spjd 1195168404Spjdstatic void 1196168404Spjddmu_tx_unassign(dmu_tx_t *tx) 1197168404Spjd{ 1198168404Spjd dmu_tx_hold_t *txh; 1199168404Spjd 1200168404Spjd if (tx->tx_txg == 0) 1201168404Spjd return; 1202168404Spjd 1203168404Spjd txg_rele_to_quiesce(&tx->tx_txgh); 1204168404Spjd 1205252749Sdelphij /* 1206252749Sdelphij * Walk the transaction's hold list, removing the hold on the 1207252749Sdelphij * associated dnode, and notifying waiters if the refcount drops to 0. 1208252749Sdelphij */ 1209168404Spjd for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh; 1210168404Spjd txh = list_next(&tx->tx_holds, txh)) { 1211168404Spjd dnode_t *dn = txh->txh_dnode; 1212168404Spjd 1213168404Spjd if (dn == NULL) 1214168404Spjd continue; 1215168404Spjd mutex_enter(&dn->dn_mtx); 1216168404Spjd ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1217168404Spjd 1218168404Spjd if (refcount_remove(&dn->dn_tx_holds, tx) == 0) { 1219168404Spjd dn->dn_assigned_txg = 0; 1220168404Spjd cv_broadcast(&dn->dn_notxholds); 1221168404Spjd } 1222168404Spjd mutex_exit(&dn->dn_mtx); 1223168404Spjd } 1224168404Spjd 1225168404Spjd txg_rele_to_sync(&tx->tx_txgh); 1226168404Spjd 1227168404Spjd tx->tx_lasttried_txg = tx->tx_txg; 1228168404Spjd tx->tx_txg = 0; 1229168404Spjd} 1230168404Spjd 1231168404Spjd/* 1232168404Spjd * Assign tx to a transaction group. txg_how can be one of: 1233168404Spjd * 1234168404Spjd * (1) TXG_WAIT. If the current open txg is full, waits until there's 1235168404Spjd * a new one. This should be used when you're not holding locks. 1236249643Smm * It will only fail if we're truly out of space (or over quota). 1237168404Spjd * 1238168404Spjd * (2) TXG_NOWAIT. If we can't assign into the current open txg without 1239168404Spjd * blocking, returns immediately with ERESTART. This should be used 1240168404Spjd * whenever you're holding locks. On an ERESTART error, the caller 1241168404Spjd * should drop locks, do a dmu_tx_wait(tx), and try again. 1242260764Savg * 1243260764Savg * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait() 1244260764Savg * has already been called on behalf of this operation (though 1245260764Savg * most likely on a different tx). 1246168404Spjd */ 1247168404Spjdint 1248249643Smmdmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how) 1249168404Spjd{ 1250168404Spjd int err; 1251168404Spjd 1252168404Spjd ASSERT(tx->tx_txg == 0); 1253260764Savg ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT || 1254260764Savg txg_how == TXG_WAITED); 1255168404Spjd ASSERT(!dsl_pool_sync_context(tx->tx_pool)); 1256168404Spjd 1257249643Smm /* If we might wait, we must not hold the config lock. */ 1258249643Smm ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool)); 1259249643Smm 1260260764Savg if (txg_how == TXG_WAITED) 1261260764Savg tx->tx_waited = B_TRUE; 1262260764Savg 1263168404Spjd while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) { 1264168404Spjd dmu_tx_unassign(tx); 1265168404Spjd 1266168404Spjd if (err != ERESTART || txg_how != TXG_WAIT) 1267168404Spjd return (err); 1268168404Spjd 1269168404Spjd dmu_tx_wait(tx); 1270168404Spjd } 1271168404Spjd 1272168404Spjd txg_rele_to_quiesce(&tx->tx_txgh); 1273168404Spjd 1274168404Spjd return (0); 1275168404Spjd} 1276168404Spjd 1277168404Spjdvoid 1278168404Spjddmu_tx_wait(dmu_tx_t *tx) 1279168404Spjd{ 1280185029Spjd spa_t *spa = tx->tx_pool->dp_spa; 1281260764Savg dsl_pool_t *dp = tx->tx_pool; 1282185029Spjd 1283168404Spjd ASSERT(tx->tx_txg == 0); 1284249643Smm ASSERT(!dsl_pool_config_held(tx->tx_pool)); 1285168404Spjd 1286260764Savg if (tx->tx_wait_dirty) { 1287260764Savg /* 1288260764Savg * dmu_tx_try_assign() has determined that we need to wait 1289260764Savg * because we've consumed much or all of the dirty buffer 1290260764Savg * space. 1291260764Savg */ 1292260764Savg mutex_enter(&dp->dp_lock); 1293260764Savg while (dp->dp_dirty_total >= zfs_dirty_data_max) 1294260764Savg cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock); 1295260764Savg uint64_t dirty = dp->dp_dirty_total; 1296260764Savg mutex_exit(&dp->dp_lock); 1297260764Savg 1298260764Savg dmu_tx_delay(tx, dirty); 1299260764Savg 1300260764Savg tx->tx_wait_dirty = B_FALSE; 1301260764Savg 1302260764Savg /* 1303260764Savg * Note: setting tx_waited only has effect if the caller 1304260764Savg * used TX_WAIT. Otherwise they are going to destroy 1305260764Savg * this tx and try again. The common case, zfs_write(), 1306260764Savg * uses TX_WAIT. 1307260764Savg */ 1308260764Savg tx->tx_waited = B_TRUE; 1309260764Savg } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) { 1310260764Savg /* 1311260764Savg * If the pool is suspended we need to wait until it 1312260764Savg * is resumed. Note that it's possible that the pool 1313260764Savg * has become active after this thread has tried to 1314260764Savg * obtain a tx. If that's the case then tx_lasttried_txg 1315260764Savg * would not have been set. 1316260764Savg */ 1317260764Savg txg_wait_synced(dp, spa_last_synced_txg(spa) + 1); 1318185029Spjd } else if (tx->tx_needassign_txh) { 1319260764Savg /* 1320260764Savg * A dnode is assigned to the quiescing txg. Wait for its 1321260764Savg * transaction to complete. 1322260764Savg */ 1323168404Spjd dnode_t *dn = tx->tx_needassign_txh->txh_dnode; 1324168404Spjd 1325168404Spjd mutex_enter(&dn->dn_mtx); 1326168404Spjd while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1) 1327168404Spjd cv_wait(&dn->dn_notxholds, &dn->dn_mtx); 1328168404Spjd mutex_exit(&dn->dn_mtx); 1329168404Spjd tx->tx_needassign_txh = NULL; 1330168404Spjd } else { 1331168404Spjd txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1); 1332168404Spjd } 1333168404Spjd} 1334168404Spjd 1335168404Spjdvoid 1336168404Spjddmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta) 1337168404Spjd{ 1338168404Spjd#ifdef ZFS_DEBUG 1339168404Spjd if (tx->tx_dir == NULL || delta == 0) 1340168404Spjd return; 1341168404Spjd 1342168404Spjd if (delta > 0) { 1343168404Spjd ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=, 1344168404Spjd tx->tx_space_towrite); 1345168404Spjd (void) refcount_add_many(&tx->tx_space_written, delta, NULL); 1346168404Spjd } else { 1347168404Spjd (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL); 1348168404Spjd } 1349168404Spjd#endif 1350168404Spjd} 1351168404Spjd 1352168404Spjdvoid 1353168404Spjddmu_tx_commit(dmu_tx_t *tx) 1354168404Spjd{ 1355168404Spjd dmu_tx_hold_t *txh; 1356168404Spjd 1357168404Spjd ASSERT(tx->tx_txg != 0); 1358168404Spjd 1359252749Sdelphij /* 1360252749Sdelphij * Go through the transaction's hold list and remove holds on 1361252749Sdelphij * associated dnodes, notifying waiters if no holds remain. 1362252749Sdelphij */ 1363168404Spjd while (txh = list_head(&tx->tx_holds)) { 1364168404Spjd dnode_t *dn = txh->txh_dnode; 1365168404Spjd 1366168404Spjd list_remove(&tx->tx_holds, txh); 1367168404Spjd kmem_free(txh, sizeof (dmu_tx_hold_t)); 1368168404Spjd if (dn == NULL) 1369168404Spjd continue; 1370168404Spjd mutex_enter(&dn->dn_mtx); 1371168404Spjd ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); 1372168404Spjd 1373168404Spjd if (refcount_remove(&dn->dn_tx_holds, tx) == 0) { 1374168404Spjd dn->dn_assigned_txg = 0; 1375168404Spjd cv_broadcast(&dn->dn_notxholds); 1376168404Spjd } 1377168404Spjd mutex_exit(&dn->dn_mtx); 1378168404Spjd dnode_rele(dn, tx); 1379168404Spjd } 1380168404Spjd 1381168404Spjd if (tx->tx_tempreserve_cookie) 1382168404Spjd dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx); 1383168404Spjd 1384219089Spjd if (!list_is_empty(&tx->tx_callbacks)) 1385219089Spjd txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks); 1386219089Spjd 1387168404Spjd if (tx->tx_anyobj == FALSE) 1388168404Spjd txg_rele_to_sync(&tx->tx_txgh); 1389219089Spjd 1390219089Spjd list_destroy(&tx->tx_callbacks); 1391185029Spjd list_destroy(&tx->tx_holds); 1392168404Spjd#ifdef ZFS_DEBUG 1393168404Spjd dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n", 1394168404Spjd tx->tx_space_towrite, refcount_count(&tx->tx_space_written), 1395168404Spjd tx->tx_space_tofree, refcount_count(&tx->tx_space_freed)); 1396168404Spjd refcount_destroy_many(&tx->tx_space_written, 1397168404Spjd refcount_count(&tx->tx_space_written)); 1398168404Spjd refcount_destroy_many(&tx->tx_space_freed, 1399168404Spjd refcount_count(&tx->tx_space_freed)); 1400168404Spjd#endif 1401168404Spjd kmem_free(tx, sizeof (dmu_tx_t)); 1402168404Spjd} 1403168404Spjd 1404168404Spjdvoid 1405168404Spjddmu_tx_abort(dmu_tx_t *tx) 1406168404Spjd{ 1407168404Spjd dmu_tx_hold_t *txh; 1408168404Spjd 1409168404Spjd ASSERT(tx->tx_txg == 0); 1410168404Spjd 1411168404Spjd while (txh = list_head(&tx->tx_holds)) { 1412168404Spjd dnode_t *dn = txh->txh_dnode; 1413168404Spjd 1414168404Spjd list_remove(&tx->tx_holds, txh); 1415168404Spjd kmem_free(txh, sizeof (dmu_tx_hold_t)); 1416168404Spjd if (dn != NULL) 1417168404Spjd dnode_rele(dn, tx); 1418168404Spjd } 1419219089Spjd 1420219089Spjd /* 1421219089Spjd * Call any registered callbacks with an error code. 1422219089Spjd */ 1423219089Spjd if (!list_is_empty(&tx->tx_callbacks)) 1424219089Spjd dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED); 1425219089Spjd 1426219089Spjd list_destroy(&tx->tx_callbacks); 1427185029Spjd list_destroy(&tx->tx_holds); 1428168404Spjd#ifdef ZFS_DEBUG 1429168404Spjd refcount_destroy_many(&tx->tx_space_written, 1430168404Spjd refcount_count(&tx->tx_space_written)); 1431168404Spjd refcount_destroy_many(&tx->tx_space_freed, 1432168404Spjd refcount_count(&tx->tx_space_freed)); 1433168404Spjd#endif 1434168404Spjd kmem_free(tx, sizeof (dmu_tx_t)); 1435168404Spjd} 1436168404Spjd 1437168404Spjduint64_t 1438168404Spjddmu_tx_get_txg(dmu_tx_t *tx) 1439168404Spjd{ 1440168404Spjd ASSERT(tx->tx_txg != 0); 1441168404Spjd return (tx->tx_txg); 1442168404Spjd} 1443219089Spjd 1444249643Smmdsl_pool_t * 1445249643Smmdmu_tx_pool(dmu_tx_t *tx) 1446249643Smm{ 1447249643Smm ASSERT(tx->tx_pool != NULL); 1448249643Smm return (tx->tx_pool); 1449249643Smm} 1450249643Smm 1451249643Smm 1452219089Spjdvoid 1453219089Spjddmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data) 1454219089Spjd{ 1455219089Spjd dmu_tx_callback_t *dcb; 1456219089Spjd 1457219089Spjd dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP); 1458219089Spjd 1459219089Spjd dcb->dcb_func = func; 1460219089Spjd dcb->dcb_data = data; 1461219089Spjd 1462219089Spjd list_insert_tail(&tx->tx_callbacks, dcb); 1463219089Spjd} 1464219089Spjd 1465219089Spjd/* 1466219089Spjd * Call all the commit callbacks on a list, with a given error code. 1467219089Spjd */ 1468219089Spjdvoid 1469219089Spjddmu_tx_do_callbacks(list_t *cb_list, int error) 1470219089Spjd{ 1471219089Spjd dmu_tx_callback_t *dcb; 1472219089Spjd 1473219089Spjd while (dcb = list_head(cb_list)) { 1474219089Spjd list_remove(cb_list, dcb); 1475219089Spjd dcb->dcb_func(dcb->dcb_data, error); 1476219089Spjd kmem_free(dcb, sizeof (dmu_tx_callback_t)); 1477219089Spjd } 1478219089Spjd} 1479219089Spjd 1480219089Spjd/* 1481219089Spjd * Interface to hold a bunch of attributes. 1482219089Spjd * used for creating new files. 1483219089Spjd * attrsize is the total size of all attributes 1484219089Spjd * to be added during object creation 1485219089Spjd * 1486219089Spjd * For updating/adding a single attribute dmu_tx_hold_sa() should be used. 1487219089Spjd */ 1488219089Spjd 1489219089Spjd/* 1490219089Spjd * hold necessary attribute name for attribute registration. 1491219089Spjd * should be a very rare case where this is needed. If it does 1492219089Spjd * happen it would only happen on the first write to the file system. 1493219089Spjd */ 1494219089Spjdstatic void 1495219089Spjddmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx) 1496219089Spjd{ 1497219089Spjd int i; 1498219089Spjd 1499219089Spjd if (!sa->sa_need_attr_registration) 1500219089Spjd return; 1501219089Spjd 1502219089Spjd for (i = 0; i != sa->sa_num_attrs; i++) { 1503219089Spjd if (!sa->sa_attr_table[i].sa_registered) { 1504219089Spjd if (sa->sa_reg_attr_obj) 1505219089Spjd dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj, 1506219089Spjd B_TRUE, sa->sa_attr_table[i].sa_name); 1507219089Spjd else 1508219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, 1509219089Spjd B_TRUE, sa->sa_attr_table[i].sa_name); 1510219089Spjd } 1511219089Spjd } 1512219089Spjd} 1513219089Spjd 1514219089Spjd 1515219089Spjdvoid 1516219089Spjddmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object) 1517219089Spjd{ 1518219089Spjd dnode_t *dn; 1519219089Spjd dmu_tx_hold_t *txh; 1520219089Spjd 1521219089Spjd txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object, 1522219089Spjd THT_SPILL, 0, 0); 1523219089Spjd 1524219089Spjd dn = txh->txh_dnode; 1525219089Spjd 1526219089Spjd if (dn == NULL) 1527219089Spjd return; 1528219089Spjd 1529219089Spjd /* If blkptr doesn't exist then add space to towrite */ 1530219089Spjd if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { 1531219089Spjd txh->txh_space_towrite += SPA_MAXBLOCKSIZE; 1532219089Spjd } else { 1533226944Smm blkptr_t *bp; 1534226944Smm 1535219089Spjd bp = &dn->dn_phys->dn_spill; 1536219089Spjd if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, 1537219089Spjd bp, bp->blk_birth)) 1538219089Spjd txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE; 1539219089Spjd else 1540219089Spjd txh->txh_space_towrite += SPA_MAXBLOCKSIZE; 1541226944Smm if (!BP_IS_HOLE(bp)) 1542219089Spjd txh->txh_space_tounref += SPA_MAXBLOCKSIZE; 1543219089Spjd } 1544219089Spjd} 1545219089Spjd 1546219089Spjdvoid 1547219089Spjddmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize) 1548219089Spjd{ 1549219089Spjd sa_os_t *sa = tx->tx_objset->os_sa; 1550219089Spjd 1551219089Spjd dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT); 1552219089Spjd 1553219089Spjd if (tx->tx_objset->os_sa->sa_master_obj == 0) 1554219089Spjd return; 1555219089Spjd 1556219089Spjd if (tx->tx_objset->os_sa->sa_layout_attr_obj) 1557219089Spjd dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL); 1558219089Spjd else { 1559219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS); 1560219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY); 1561219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1562219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1563219089Spjd } 1564219089Spjd 1565219089Spjd dmu_tx_sa_registration_hold(sa, tx); 1566219089Spjd 1567219089Spjd if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill) 1568219089Spjd return; 1569219089Spjd 1570219089Spjd (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT, 1571219089Spjd THT_SPILL, 0, 0); 1572219089Spjd} 1573219089Spjd 1574219089Spjd/* 1575219089Spjd * Hold SA attribute 1576219089Spjd * 1577219089Spjd * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size) 1578219089Spjd * 1579219089Spjd * variable_size is the total size of all variable sized attributes 1580219089Spjd * passed to this function. It is not the total size of all 1581219089Spjd * variable size attributes that *may* exist on this object. 1582219089Spjd */ 1583219089Spjdvoid 1584219089Spjddmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow) 1585219089Spjd{ 1586219089Spjd uint64_t object; 1587219089Spjd sa_os_t *sa = tx->tx_objset->os_sa; 1588219089Spjd 1589219089Spjd ASSERT(hdl != NULL); 1590219089Spjd 1591219089Spjd object = sa_handle_object(hdl); 1592219089Spjd 1593219089Spjd dmu_tx_hold_bonus(tx, object); 1594219089Spjd 1595219089Spjd if (tx->tx_objset->os_sa->sa_master_obj == 0) 1596219089Spjd return; 1597219089Spjd 1598219089Spjd if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 || 1599219089Spjd tx->tx_objset->os_sa->sa_layout_attr_obj == 0) { 1600219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS); 1601219089Spjd dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY); 1602219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1603219089Spjd dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL); 1604219089Spjd } 1605219089Spjd 1606219089Spjd dmu_tx_sa_registration_hold(sa, tx); 1607219089Spjd 1608219089Spjd if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj) 1609219089Spjd dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL); 1610219089Spjd 1611219089Spjd if (sa->sa_force_spill || may_grow || hdl->sa_spill) { 1612219089Spjd ASSERT(tx->tx_txg == 0); 1613219089Spjd dmu_tx_hold_spill(tx, object); 1614219089Spjd } else { 1615219089Spjd dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus; 1616219089Spjd dnode_t *dn; 1617219089Spjd 1618219089Spjd DB_DNODE_ENTER(db); 1619219089Spjd dn = DB_DNODE(db); 1620219089Spjd if (dn->dn_have_spill) { 1621219089Spjd ASSERT(tx->tx_txg == 0); 1622219089Spjd dmu_tx_hold_spill(tx, object); 1623219089Spjd } 1624219089Spjd DB_DNODE_EXIT(db); 1625219089Spjd } 1626219089Spjd} 1627