1236884Smm/* 2236884Smm * CDDL HEADER START 3236884Smm * 4236884Smm * The contents of this file are subject to the terms of the 5236884Smm * Common Development and Distribution License (the "License"). 6236884Smm * You may not use this file except in compliance with the License. 7236884Smm * 8236884Smm * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9236884Smm * or http://www.opensolaris.org/os/licensing. 10236884Smm * See the License for the specific language governing permissions 11236884Smm * and limitations under the License. 12236884Smm * 13236884Smm * When distributing Covered Code, include this CDDL HEADER in each 14236884Smm * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15236884Smm * If applicable, add the following below this CDDL HEADER, with the 16236884Smm * fields enclosed by brackets "[]" replaced with your own identifying 17236884Smm * information: Portions Copyright [yyyy] [name of copyright owner] 18236884Smm * 19236884Smm * CDDL HEADER END 20236884Smm */ 21236884Smm 22236884Smm/* 23268657Sdelphij * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24236884Smm */ 25236884Smm 26236884Smm#include <sys/arc.h> 27236884Smm#include <sys/bptree.h> 28236884Smm#include <sys/dmu.h> 29236884Smm#include <sys/dmu_objset.h> 30236884Smm#include <sys/dmu_tx.h> 31236884Smm#include <sys/dmu_traverse.h> 32236884Smm#include <sys/dsl_dataset.h> 33236884Smm#include <sys/dsl_dir.h> 34236884Smm#include <sys/dsl_pool.h> 35236884Smm#include <sys/dnode.h> 36236884Smm#include <sys/refcount.h> 37236884Smm#include <sys/spa.h> 38236884Smm 39236884Smm/* 40236884Smm * A bptree is a queue of root block pointers from destroyed datasets. When a 41236884Smm * dataset is destroyed its root block pointer is put on the end of the pool's 42236884Smm * bptree queue so the dataset's blocks can be freed asynchronously by 43236884Smm * dsl_scan_sync. This allows the delete operation to finish without traversing 44236884Smm * all the dataset's blocks. 45236884Smm * 46251631Sdelphij * Note that while bt_begin and bt_end are only ever incremented in this code, 47236884Smm * they are effectively reset to 0 every time the entire bptree is freed because 48236884Smm * the bptree's object is destroyed and re-created. 49236884Smm */ 50236884Smm 51236884Smmstruct bptree_args { 52236884Smm bptree_phys_t *ba_phys; /* data in bonus buffer, dirtied if freeing */ 53236884Smm boolean_t ba_free; /* true if freeing during traversal */ 54236884Smm 55236884Smm bptree_itor_t *ba_func; /* function to call for each blockpointer */ 56236884Smm void *ba_arg; /* caller supplied argument to ba_func */ 57236884Smm dmu_tx_t *ba_tx; /* caller supplied tx, NULL if not freeing */ 58236884Smm} bptree_args_t; 59236884Smm 60236884Smmuint64_t 61236884Smmbptree_alloc(objset_t *os, dmu_tx_t *tx) 62236884Smm{ 63236884Smm uint64_t obj; 64236884Smm dmu_buf_t *db; 65236884Smm bptree_phys_t *bt; 66236884Smm 67236884Smm obj = dmu_object_alloc(os, DMU_OTN_UINT64_METADATA, 68236884Smm SPA_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA, 69236884Smm sizeof (bptree_phys_t), tx); 70236884Smm 71236884Smm /* 72236884Smm * Bonus buffer contents are already initialized to 0, but for 73236884Smm * readability we make it explicit. 74236884Smm */ 75236884Smm VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db)); 76236884Smm dmu_buf_will_dirty(db, tx); 77236884Smm bt = db->db_data; 78236884Smm bt->bt_begin = 0; 79236884Smm bt->bt_end = 0; 80236884Smm bt->bt_bytes = 0; 81236884Smm bt->bt_comp = 0; 82236884Smm bt->bt_uncomp = 0; 83236884Smm dmu_buf_rele(db, FTAG); 84236884Smm 85236884Smm return (obj); 86236884Smm} 87236884Smm 88236884Smmint 89236884Smmbptree_free(objset_t *os, uint64_t obj, dmu_tx_t *tx) 90236884Smm{ 91236884Smm dmu_buf_t *db; 92236884Smm bptree_phys_t *bt; 93236884Smm 94236884Smm VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db)); 95236884Smm bt = db->db_data; 96236884Smm ASSERT3U(bt->bt_begin, ==, bt->bt_end); 97240415Smm ASSERT0(bt->bt_bytes); 98240415Smm ASSERT0(bt->bt_comp); 99240415Smm ASSERT0(bt->bt_uncomp); 100236884Smm dmu_buf_rele(db, FTAG); 101236884Smm 102236884Smm return (dmu_object_free(os, obj, tx)); 103236884Smm} 104236884Smm 105268650Sdelphijboolean_t 106268650Sdelphijbptree_is_empty(objset_t *os, uint64_t obj) 107268650Sdelphij{ 108268650Sdelphij dmu_buf_t *db; 109268650Sdelphij bptree_phys_t *bt; 110268650Sdelphij boolean_t rv; 111268650Sdelphij 112268650Sdelphij VERIFY0(dmu_bonus_hold(os, obj, FTAG, &db)); 113268650Sdelphij bt = db->db_data; 114268650Sdelphij rv = (bt->bt_begin == bt->bt_end); 115268650Sdelphij dmu_buf_rele(db, FTAG); 116268650Sdelphij return (rv); 117268650Sdelphij} 118268650Sdelphij 119236884Smmvoid 120236884Smmbptree_add(objset_t *os, uint64_t obj, blkptr_t *bp, uint64_t birth_txg, 121236884Smm uint64_t bytes, uint64_t comp, uint64_t uncomp, dmu_tx_t *tx) 122236884Smm{ 123236884Smm dmu_buf_t *db; 124236884Smm bptree_phys_t *bt; 125268650Sdelphij bptree_entry_phys_t bte = { 0 }; 126236884Smm 127236884Smm /* 128236884Smm * bptree objects are in the pool mos, therefore they can only be 129236884Smm * modified in syncing context. Furthermore, this is only modified 130236884Smm * by the sync thread, so no locking is necessary. 131236884Smm */ 132236884Smm ASSERT(dmu_tx_is_syncing(tx)); 133236884Smm 134236884Smm VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db)); 135236884Smm bt = db->db_data; 136236884Smm 137236884Smm bte.be_birth_txg = birth_txg; 138236884Smm bte.be_bp = *bp; 139236884Smm dmu_write(os, obj, bt->bt_end * sizeof (bte), sizeof (bte), &bte, tx); 140236884Smm 141236884Smm dmu_buf_will_dirty(db, tx); 142236884Smm bt->bt_end++; 143236884Smm bt->bt_bytes += bytes; 144236884Smm bt->bt_comp += comp; 145236884Smm bt->bt_uncomp += uncomp; 146236884Smm dmu_buf_rele(db, FTAG); 147236884Smm} 148236884Smm 149236884Smm/* ARGSUSED */ 150236884Smmstatic int 151246666Smmbptree_visit_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 152268657Sdelphij const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) 153236884Smm{ 154236884Smm int err; 155236884Smm struct bptree_args *ba = arg; 156236884Smm 157263397Sdelphij if (BP_IS_HOLE(bp)) 158236884Smm return (0); 159236884Smm 160236884Smm err = ba->ba_func(ba->ba_arg, bp, ba->ba_tx); 161236884Smm if (err == 0 && ba->ba_free) { 162236884Smm ba->ba_phys->bt_bytes -= bp_get_dsize_sync(spa, bp); 163236884Smm ba->ba_phys->bt_comp -= BP_GET_PSIZE(bp); 164236884Smm ba->ba_phys->bt_uncomp -= BP_GET_UCSIZE(bp); 165236884Smm } 166236884Smm return (err); 167236884Smm} 168236884Smm 169268650Sdelphij/* 170268650Sdelphij * If "free" is set: 171268650Sdelphij * - It is assumed that "func" will be freeing the block pointers. 172268650Sdelphij * - If "func" returns nonzero, the bookmark will be remembered and 173268650Sdelphij * iteration will be restarted from this point on next invocation. 174268650Sdelphij * - If an i/o error is encountered (e.g. "func" returns EIO or ECKSUM), 175268650Sdelphij * bptree_iterate will remember the bookmark, continue traversing 176268650Sdelphij * any additional entries, and return 0. 177268650Sdelphij * 178268650Sdelphij * If "free" is not set, traversal will stop and return an error if 179268650Sdelphij * an i/o error is encountered. 180268650Sdelphij * 181268650Sdelphij * In either case, if zfs_free_leak_on_eio is set, i/o errors will be 182268650Sdelphij * ignored and traversal will continue (i.e. TRAVERSE_HARD will be passed to 183268650Sdelphij * traverse_dataset_destroyed()). 184268650Sdelphij */ 185236884Smmint 186236884Smmbptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func, 187236884Smm void *arg, dmu_tx_t *tx) 188236884Smm{ 189268650Sdelphij boolean_t ioerr = B_FALSE; 190236884Smm int err; 191236884Smm uint64_t i; 192236884Smm dmu_buf_t *db; 193236884Smm struct bptree_args ba; 194236884Smm 195236884Smm ASSERT(!free || dmu_tx_is_syncing(tx)); 196236884Smm 197236884Smm err = dmu_bonus_hold(os, obj, FTAG, &db); 198236884Smm if (err != 0) 199236884Smm return (err); 200236884Smm 201236884Smm if (free) 202236884Smm dmu_buf_will_dirty(db, tx); 203236884Smm 204236884Smm ba.ba_phys = db->db_data; 205236884Smm ba.ba_free = free; 206236884Smm ba.ba_func = func; 207236884Smm ba.ba_arg = arg; 208236884Smm ba.ba_tx = tx; 209236884Smm 210236884Smm err = 0; 211236884Smm for (i = ba.ba_phys->bt_begin; i < ba.ba_phys->bt_end; i++) { 212236884Smm bptree_entry_phys_t bte; 213262120Savg int flags = TRAVERSE_PREFETCH_METADATA | TRAVERSE_POST; 214236884Smm 215236884Smm err = dmu_read(os, obj, i * sizeof (bte), sizeof (bte), 216236884Smm &bte, DMU_READ_NO_PREFETCH); 217236884Smm if (err != 0) 218236884Smm break; 219236884Smm 220268650Sdelphij if (zfs_free_leak_on_eio) 221262120Savg flags |= TRAVERSE_HARD; 222268650Sdelphij zfs_dbgmsg("bptree index %d: traversing from min_txg=%lld " 223268650Sdelphij "bookmark %lld/%lld/%lld/%lld", 224268650Sdelphij i, (longlong_t)bte.be_birth_txg, 225268650Sdelphij (longlong_t)bte.be_zb.zb_objset, 226268650Sdelphij (longlong_t)bte.be_zb.zb_object, 227268650Sdelphij (longlong_t)bte.be_zb.zb_level, 228268650Sdelphij (longlong_t)bte.be_zb.zb_blkid); 229236884Smm err = traverse_dataset_destroyed(os->os_spa, &bte.be_bp, 230262120Savg bte.be_birth_txg, &bte.be_zb, flags, 231236884Smm bptree_visit_cb, &ba); 232236884Smm if (free) { 233268650Sdelphij /* 234268650Sdelphij * The callback has freed the visited block pointers. 235268650Sdelphij * Record our traversal progress on disk, either by 236268650Sdelphij * updating this record's bookmark, or by logically 237268650Sdelphij * removing this record by advancing bt_begin. 238268650Sdelphij */ 239268650Sdelphij if (err != 0) { 240236884Smm /* save bookmark for future resume */ 241236884Smm ASSERT3U(bte.be_zb.zb_objset, ==, 242236884Smm ZB_DESTROYED_OBJSET); 243240415Smm ASSERT0(bte.be_zb.zb_level); 244236884Smm dmu_write(os, obj, i * sizeof (bte), 245236884Smm sizeof (bte), &bte, tx); 246268650Sdelphij if (err == EIO || err == ECKSUM || 247268650Sdelphij err == ENXIO) { 248268650Sdelphij /* 249268650Sdelphij * Skip the rest of this tree and 250268650Sdelphij * continue on to the next entry. 251268650Sdelphij */ 252268650Sdelphij err = 0; 253268650Sdelphij ioerr = B_TRUE; 254268650Sdelphij } else { 255268650Sdelphij break; 256268650Sdelphij } 257268650Sdelphij } else if (ioerr) { 258262120Savg /* 259268650Sdelphij * This entry is finished, but there were 260268650Sdelphij * i/o errors on previous entries, so we 261268650Sdelphij * can't adjust bt_begin. Set this entry's 262268650Sdelphij * be_birth_txg such that it will be 263268650Sdelphij * treated as a no-op in future traversals. 264262120Savg */ 265268650Sdelphij bte.be_birth_txg = UINT64_MAX; 266268650Sdelphij dmu_write(os, obj, i * sizeof (bte), 267268650Sdelphij sizeof (bte), &bte, tx); 268262120Savg } 269262120Savg 270268650Sdelphij if (!ioerr) { 271268650Sdelphij ba.ba_phys->bt_begin++; 272268650Sdelphij (void) dmu_free_range(os, obj, 273268650Sdelphij i * sizeof (bte), sizeof (bte), tx); 274268650Sdelphij } 275268650Sdelphij } else if (err != 0) { 276268650Sdelphij break; 277236884Smm } 278236884Smm } 279236884Smm 280268650Sdelphij ASSERT(!free || err != 0 || ioerr || 281268650Sdelphij ba.ba_phys->bt_begin == ba.ba_phys->bt_end); 282236884Smm 283236884Smm /* if all blocks are free there should be no used space */ 284236884Smm if (ba.ba_phys->bt_begin == ba.ba_phys->bt_end) { 285268650Sdelphij if (zfs_free_leak_on_eio) { 286268650Sdelphij ba.ba_phys->bt_bytes = 0; 287268650Sdelphij ba.ba_phys->bt_comp = 0; 288268650Sdelphij ba.ba_phys->bt_uncomp = 0; 289268650Sdelphij } 290268650Sdelphij 291240415Smm ASSERT0(ba.ba_phys->bt_bytes); 292240415Smm ASSERT0(ba.ba_phys->bt_comp); 293240415Smm ASSERT0(ba.ba_phys->bt_uncomp); 294236884Smm } 295236884Smm 296236884Smm dmu_buf_rele(db, FTAG); 297236884Smm 298236884Smm return (err); 299236884Smm} 300