1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5ea8dc4b6Seschrock * Common Development and Distribution License (the "License"). 6ea8dc4b6Seschrock * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 22*55434c77Sek * Copyright 2007 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 26fa9e4066Sahrens #pragma ident "%Z%%M% %I% %E% SMI" 27fa9e4066Sahrens 28fa9e4066Sahrens #include <sys/zfs_context.h> 29fa9e4066Sahrens #include <sys/dbuf.h> 30fa9e4066Sahrens #include <sys/dnode.h> 31fa9e4066Sahrens #include <sys/dmu.h> 32fa9e4066Sahrens #include <sys/dmu_tx.h> 33fa9e4066Sahrens #include <sys/dmu_objset.h> 34fa9e4066Sahrens #include <sys/dsl_dataset.h> 35fa9e4066Sahrens #include <sys/spa.h> 36fa9e4066Sahrens #include <sys/zio.h> 37fa9e4066Sahrens 38fa9e4066Sahrens static void 39fa9e4066Sahrens dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx) 40fa9e4066Sahrens { 41fa9e4066Sahrens dmu_buf_impl_t *db; 42fa9e4066Sahrens int i; 43fa9e4066Sahrens uint64_t txg = tx->tx_txg; 44fa9e4066Sahrens 45fa9e4066Sahrens ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE); 46fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 47fa9e4066Sahrens /* this dnode can't be paged out because it's dirty */ 48fa9e4066Sahrens 49fa9e4066Sahrens db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG); 50ea8dc4b6Seschrock ASSERT(db != NULL); 51fa9e4066Sahrens for (i = 0; i < dn->dn_phys->dn_nblkptr; i++) 52fa9e4066Sahrens if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i])) 53fa9e4066Sahrens break; 54fa9e4066Sahrens if (i != dn->dn_phys->dn_nblkptr) { 55fa9e4066Sahrens ASSERT(list_link_active(&db->db_dirty_node[txg&TXG_MASK])); 56fa9e4066Sahrens 57ea8dc4b6Seschrock (void) dbuf_read(db, NULL, 58ea8dc4b6Seschrock DB_RF_HAVESTRUCT | DB_RF_MUST_SUCCEED); 59fa9e4066Sahrens arc_release(db->db_buf, db); 60fa9e4066Sahrens /* copy dnode's block pointers to new indirect block */ 61fa9e4066Sahrens ASSERT3U(sizeof (blkptr_t) * dn->dn_phys->dn_nblkptr, <=, 62fa9e4066Sahrens db->db.db_size); 63fa9e4066Sahrens bcopy(dn->dn_phys->dn_blkptr, db->db.db_data, 64fa9e4066Sahrens sizeof (blkptr_t) * dn->dn_phys->dn_nblkptr); 656b4acc8bSahrens arc_buf_freeze(db->db_buf); 66fa9e4066Sahrens } 67fa9e4066Sahrens 68fa9e4066Sahrens dn->dn_phys->dn_nlevels += 1; 69fa9e4066Sahrens dprintf("os=%p obj=%llu, increase to %d\n", 70fa9e4066Sahrens dn->dn_objset, dn->dn_object, 71fa9e4066Sahrens dn->dn_phys->dn_nlevels); 72fa9e4066Sahrens 73fa9e4066Sahrens /* set dbuf's parent pointers to new indirect buf */ 74fa9e4066Sahrens for (i = 0; i < dn->dn_phys->dn_nblkptr; i++) { 75fa9e4066Sahrens dmu_buf_impl_t *child = 76fa9e4066Sahrens dbuf_find(dn, dn->dn_phys->dn_nlevels-2, i); 77fa9e4066Sahrens if (child == NULL) 78fa9e4066Sahrens continue; 79fa9e4066Sahrens if (child->db_dnode == NULL) { 80fa9e4066Sahrens mutex_exit(&child->db_mtx); 81fa9e4066Sahrens continue; 82fa9e4066Sahrens } 83fa9e4066Sahrens 84fa9e4066Sahrens if (child->db_parent == NULL || 85fa9e4066Sahrens child->db_parent == dn->dn_dbuf) { 86fa9e4066Sahrens dprintf_dbuf_bp(child, child->db_blkptr, 87fa9e4066Sahrens "changing db_blkptr to new indirect %s", ""); 88fa9e4066Sahrens child->db_parent = db; 89fa9e4066Sahrens dbuf_add_ref(db, child); 90fa9e4066Sahrens if (db->db.db_data) { 91fa9e4066Sahrens child->db_blkptr = 92fa9e4066Sahrens (blkptr_t *)db->db.db_data + i; 93fa9e4066Sahrens } else { 94fa9e4066Sahrens child->db_blkptr = NULL; 95fa9e4066Sahrens } 96fa9e4066Sahrens dprintf_dbuf_bp(child, child->db_blkptr, 97fa9e4066Sahrens "changed db_blkptr to new indirect %s", ""); 98fa9e4066Sahrens } 99fa9e4066Sahrens ASSERT3P(child->db_parent, ==, db); 100fa9e4066Sahrens 101fa9e4066Sahrens mutex_exit(&child->db_mtx); 102fa9e4066Sahrens } 103fa9e4066Sahrens 104fa9e4066Sahrens bzero(dn->dn_phys->dn_blkptr, 105fa9e4066Sahrens sizeof (blkptr_t) * dn->dn_phys->dn_nblkptr); 106fa9e4066Sahrens 107ea8dc4b6Seschrock dbuf_rele(db, FTAG); 108fa9e4066Sahrens } 109fa9e4066Sahrens 110fa9e4066Sahrens static void 111fa9e4066Sahrens free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx) 112fa9e4066Sahrens { 113fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 114fa9e4066Sahrens uint64_t bytesfreed = 0; 115fa9e4066Sahrens int i; 116fa9e4066Sahrens 117fa9e4066Sahrens dprintf("os=%p obj=%llx num=%d\n", os, dn->dn_object, num); 118fa9e4066Sahrens 119fa9e4066Sahrens for (i = 0; i < num; i++, bp++) { 120fa9e4066Sahrens if (BP_IS_HOLE(bp)) 121fa9e4066Sahrens continue; 122fa9e4066Sahrens 12399653d4eSeschrock bytesfreed += bp_get_dasize(os->os_spa, bp); 12499653d4eSeschrock ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys)); 125fa9e4066Sahrens dsl_dataset_block_kill(os->os_dsl_dataset, bp, tx); 126fa9e4066Sahrens } 127fa9e4066Sahrens dnode_diduse_space(dn, -bytesfreed); 128fa9e4066Sahrens } 129fa9e4066Sahrens 1309c9dc39aSek #ifdef ZFS_DEBUG 131fa9e4066Sahrens static void 132fa9e4066Sahrens free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx) 133fa9e4066Sahrens { 134fa9e4066Sahrens int off, num; 135fa9e4066Sahrens int i, err, epbs; 136fa9e4066Sahrens uint64_t txg = tx->tx_txg; 137fa9e4066Sahrens 138fa9e4066Sahrens epbs = db->db_dnode->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 139fa9e4066Sahrens off = start - (db->db_blkid * 1<<epbs); 140fa9e4066Sahrens num = end - start + 1; 141fa9e4066Sahrens 142fa9e4066Sahrens ASSERT3U(off, >=, 0); 143fa9e4066Sahrens ASSERT3U(num, >=, 0); 144fa9e4066Sahrens ASSERT3U(db->db_level, >, 0); 145fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, 1<<db->db_dnode->dn_phys->dn_indblkshift); 146fa9e4066Sahrens ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT); 147fa9e4066Sahrens ASSERT(db->db_blkptr != NULL); 148fa9e4066Sahrens 149fa9e4066Sahrens for (i = off; i < off+num; i++) { 150fa9e4066Sahrens uint64_t *buf; 151fa9e4066Sahrens int j; 152fa9e4066Sahrens dmu_buf_impl_t *child; 153fa9e4066Sahrens 154fa9e4066Sahrens ASSERT(db->db_level == 1); 155fa9e4066Sahrens 156fa9e4066Sahrens rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER); 157fa9e4066Sahrens err = dbuf_hold_impl(db->db_dnode, db->db_level-1, 158fa9e4066Sahrens (db->db_blkid << epbs) + i, TRUE, FTAG, &child); 159fa9e4066Sahrens rw_exit(&db->db_dnode->dn_struct_rwlock); 160fa9e4066Sahrens if (err == ENOENT) 161fa9e4066Sahrens continue; 162fa9e4066Sahrens ASSERT(err == 0); 163fa9e4066Sahrens ASSERT(child->db_level == 0); 164fa9e4066Sahrens ASSERT(!list_link_active(&child->db_dirty_node[txg&TXG_MASK])); 165fa9e4066Sahrens 166fa9e4066Sahrens /* db_data_old better be zeroed */ 167fa9e4066Sahrens if (child->db_d.db_data_old[txg & TXG_MASK]) { 1686b4acc8bSahrens buf = child->db_d.db_data_old[txg & TXG_MASK]->b_data; 169fa9e4066Sahrens for (j = 0; j < child->db.db_size >> 3; j++) { 170fa9e4066Sahrens if (buf[j] != 0) { 171fa9e4066Sahrens panic("freed data not zero: " 172fa9e4066Sahrens "child=%p i=%d off=%d num=%d\n", 173fa9e4066Sahrens child, i, off, num); 174fa9e4066Sahrens } 175fa9e4066Sahrens } 176fa9e4066Sahrens } 177fa9e4066Sahrens 178fa9e4066Sahrens /* 179fa9e4066Sahrens * db_data better be zeroed unless it's dirty in a 180fa9e4066Sahrens * future txg. 181fa9e4066Sahrens */ 182fa9e4066Sahrens mutex_enter(&child->db_mtx); 183fa9e4066Sahrens buf = child->db.db_data; 184fa9e4066Sahrens if (buf != NULL && child->db_state != DB_FILL && 185fa9e4066Sahrens !list_link_active(&child->db_dirty_node 186fa9e4066Sahrens [(txg+1) & TXG_MASK]) && 187fa9e4066Sahrens !list_link_active(&child->db_dirty_node 188fa9e4066Sahrens [(txg+2) & TXG_MASK])) { 189fa9e4066Sahrens for (j = 0; j < child->db.db_size >> 3; j++) { 190fa9e4066Sahrens if (buf[j] != 0) { 191fa9e4066Sahrens panic("freed data not zero: " 192fa9e4066Sahrens "child=%p i=%d off=%d num=%d\n", 193fa9e4066Sahrens child, i, off, num); 194fa9e4066Sahrens } 195fa9e4066Sahrens } 196fa9e4066Sahrens } 197fa9e4066Sahrens mutex_exit(&child->db_mtx); 198fa9e4066Sahrens 199ea8dc4b6Seschrock dbuf_rele(child, FTAG); 200fa9e4066Sahrens } 201fa9e4066Sahrens } 2029c9dc39aSek #endif 203fa9e4066Sahrens 204fa9e4066Sahrens static int 205fa9e4066Sahrens free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks, int trunc, 206fa9e4066Sahrens dmu_tx_t *tx) 207fa9e4066Sahrens { 208fa9e4066Sahrens dnode_t *dn = db->db_dnode; 209fa9e4066Sahrens blkptr_t *bp; 210fa9e4066Sahrens dmu_buf_impl_t *subdb; 211fa9e4066Sahrens uint64_t start, end, dbstart, dbend, i; 212fa9e4066Sahrens int epbs, shift, err; 21323b11526Smaybee int txgoff = tx->tx_txg & TXG_MASK; 214fa9e4066Sahrens int all = TRUE; 215fa9e4066Sahrens 216ea8dc4b6Seschrock (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 217fa9e4066Sahrens arc_release(db->db_buf, db); 218fa9e4066Sahrens bp = (blkptr_t *)db->db.db_data; 219fa9e4066Sahrens 220fa9e4066Sahrens epbs = db->db_dnode->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 221fa9e4066Sahrens shift = (db->db_level - 1) * epbs; 222fa9e4066Sahrens dbstart = db->db_blkid << epbs; 223fa9e4066Sahrens start = blkid >> shift; 224fa9e4066Sahrens if (dbstart < start) { 225fa9e4066Sahrens bp += start - dbstart; 226fa9e4066Sahrens all = FALSE; 227fa9e4066Sahrens } else { 228fa9e4066Sahrens start = dbstart; 229fa9e4066Sahrens } 230fa9e4066Sahrens dbend = ((db->db_blkid + 1) << epbs) - 1; 231fa9e4066Sahrens end = (blkid + nblks - 1) >> shift; 232fa9e4066Sahrens if (dbend <= end) 233fa9e4066Sahrens end = dbend; 234fa9e4066Sahrens else if (all) 235fa9e4066Sahrens all = trunc; 236fa9e4066Sahrens ASSERT3U(start, <=, end); 237fa9e4066Sahrens 238fa9e4066Sahrens if (db->db_level == 1) { 2399c9dc39aSek FREE_VERIFY(db, start, end, tx); 240fa9e4066Sahrens free_blocks(dn, bp, end-start+1, tx); 2416b4acc8bSahrens arc_buf_freeze(db->db_buf); 24223b11526Smaybee ASSERT(all || list_link_active(&db->db_dirty_node[txgoff])); 243fa9e4066Sahrens return (all); 244fa9e4066Sahrens } 245fa9e4066Sahrens 246fa9e4066Sahrens for (i = start; i <= end; i++, bp++) { 247fa9e4066Sahrens if (BP_IS_HOLE(bp)) 248fa9e4066Sahrens continue; 249fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 250fa9e4066Sahrens err = dbuf_hold_impl(dn, db->db_level-1, i, TRUE, FTAG, &subdb); 251fa9e4066Sahrens ASSERT3U(err, ==, 0); 252fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 253fa9e4066Sahrens 254fa9e4066Sahrens if (free_children(subdb, blkid, nblks, trunc, tx)) { 255fa9e4066Sahrens ASSERT3P(subdb->db_blkptr, ==, bp); 256fa9e4066Sahrens free_blocks(dn, bp, 1, tx); 25723b11526Smaybee } else { 25823b11526Smaybee all = FALSE; 259fa9e4066Sahrens } 260ea8dc4b6Seschrock dbuf_rele(subdb, FTAG); 261fa9e4066Sahrens } 2626b4acc8bSahrens arc_buf_freeze(db->db_buf); 263fa9e4066Sahrens #ifdef ZFS_DEBUG 264fa9e4066Sahrens bp -= (end-start)+1; 265fa9e4066Sahrens for (i = start; i <= end; i++, bp++) { 266fa9e4066Sahrens if (i == start && blkid != 0) 267fa9e4066Sahrens continue; 268fa9e4066Sahrens else if (i == end && !trunc) 269fa9e4066Sahrens continue; 270fa9e4066Sahrens ASSERT3U(bp->blk_birth, ==, 0); 271fa9e4066Sahrens } 272fa9e4066Sahrens #endif 27323b11526Smaybee ASSERT(all || list_link_active(&db->db_dirty_node[txgoff])); 274fa9e4066Sahrens return (all); 275fa9e4066Sahrens } 276fa9e4066Sahrens 277fa9e4066Sahrens /* 278fa9e4066Sahrens * free_range: Traverse the indicated range of the provided file 279fa9e4066Sahrens * and "free" all the blocks contained there. 280fa9e4066Sahrens */ 281fa9e4066Sahrens static void 282fa9e4066Sahrens dnode_sync_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx) 283fa9e4066Sahrens { 284fa9e4066Sahrens blkptr_t *bp = dn->dn_phys->dn_blkptr; 285fa9e4066Sahrens dmu_buf_impl_t *db; 286fa9e4066Sahrens int trunc, start, end, shift, i, err; 287fa9e4066Sahrens int dnlevel = dn->dn_phys->dn_nlevels; 288fa9e4066Sahrens 289fa9e4066Sahrens if (blkid > dn->dn_phys->dn_maxblkid) 290fa9e4066Sahrens return; 291fa9e4066Sahrens 292fa9e4066Sahrens ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX); 293fa9e4066Sahrens trunc = blkid + nblks > dn->dn_phys->dn_maxblkid; 294fa9e4066Sahrens if (trunc) 295fa9e4066Sahrens nblks = dn->dn_phys->dn_maxblkid - blkid + 1; 296fa9e4066Sahrens 297fa9e4066Sahrens /* There are no indirect blocks in the object */ 298fa9e4066Sahrens if (dnlevel == 1) { 299fa9e4066Sahrens if (blkid >= dn->dn_phys->dn_nblkptr) { 300fa9e4066Sahrens /* this range was never made persistent */ 301fa9e4066Sahrens return; 302fa9e4066Sahrens } 303fa9e4066Sahrens ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr); 304fa9e4066Sahrens free_blocks(dn, bp + blkid, nblks, tx); 305fa9e4066Sahrens if (trunc) { 306fa9e4066Sahrens uint64_t off = (dn->dn_phys->dn_maxblkid + 1) * 307fa9e4066Sahrens (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT); 308fa9e4066Sahrens dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0); 309fa9e4066Sahrens ASSERT(off < dn->dn_phys->dn_maxblkid || 310fa9e4066Sahrens dn->dn_phys->dn_maxblkid == 0 || 3116754306eSahrens dnode_next_offset(dn, FALSE, &off, 3126754306eSahrens 1, 1, 0) != 0); 313fa9e4066Sahrens } 314fa9e4066Sahrens return; 315fa9e4066Sahrens } 316fa9e4066Sahrens 317fa9e4066Sahrens shift = (dnlevel - 1) * (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT); 318fa9e4066Sahrens start = blkid >> shift; 319fa9e4066Sahrens ASSERT(start < dn->dn_phys->dn_nblkptr); 320fa9e4066Sahrens end = (blkid + nblks - 1) >> shift; 321fa9e4066Sahrens bp += start; 322fa9e4066Sahrens for (i = start; i <= end; i++, bp++) { 323fa9e4066Sahrens if (BP_IS_HOLE(bp)) 324fa9e4066Sahrens continue; 325fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 326fa9e4066Sahrens err = dbuf_hold_impl(dn, dnlevel-1, i, TRUE, FTAG, &db); 327fa9e4066Sahrens ASSERT3U(err, ==, 0); 328fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 329fa9e4066Sahrens 330fa9e4066Sahrens if (free_children(db, blkid, nblks, trunc, tx)) { 331fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, bp); 332fa9e4066Sahrens free_blocks(dn, bp, 1, tx); 333fa9e4066Sahrens } 334ea8dc4b6Seschrock dbuf_rele(db, FTAG); 335fa9e4066Sahrens } 336fa9e4066Sahrens if (trunc) { 337fa9e4066Sahrens uint64_t off = (dn->dn_phys->dn_maxblkid + 1) * 338fa9e4066Sahrens (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT); 339fa9e4066Sahrens dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0); 340fa9e4066Sahrens ASSERT(off < dn->dn_phys->dn_maxblkid || 341fa9e4066Sahrens dn->dn_phys->dn_maxblkid == 0 || 3426754306eSahrens dnode_next_offset(dn, FALSE, &off, 1, 1, 0) != 0); 343fa9e4066Sahrens } 344fa9e4066Sahrens } 345fa9e4066Sahrens 346ea8dc4b6Seschrock /* 347ea8dc4b6Seschrock * Try to kick all the dnodes dbufs out of the cache... 348ea8dc4b6Seschrock */ 349436b2950Sperrin int 350436b2950Sperrin dnode_evict_dbufs(dnode_t *dn, int try) 351ea8dc4b6Seschrock { 352c543ec06Sahrens int progress; 353c543ec06Sahrens int pass = 0; 354c543ec06Sahrens 355c543ec06Sahrens do { 356d90a49d6Smaybee dmu_buf_impl_t *db, marker; 357c543ec06Sahrens int evicting = FALSE; 358c543ec06Sahrens 359c543ec06Sahrens progress = FALSE; 360c543ec06Sahrens mutex_enter(&dn->dn_dbufs_mtx); 361d90a49d6Smaybee list_insert_tail(&dn->dn_dbufs, &marker); 362d90a49d6Smaybee db = list_head(&dn->dn_dbufs); 363d90a49d6Smaybee for (; db != ▮ db = list_head(&dn->dn_dbufs)) { 364d90a49d6Smaybee list_remove(&dn->dn_dbufs, db); 365d90a49d6Smaybee list_insert_tail(&dn->dn_dbufs, db); 366ea8dc4b6Seschrock 367ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 368c543ec06Sahrens if (db->db_state == DB_EVICTING) { 369c543ec06Sahrens progress = TRUE; 370c543ec06Sahrens evicting = TRUE; 371c543ec06Sahrens mutex_exit(&db->db_mtx); 372c543ec06Sahrens } else if (refcount_is_zero(&db->db_holds)) { 373c543ec06Sahrens progress = TRUE; 374c543ec06Sahrens ASSERT(!arc_released(db->db_buf)); 375c543ec06Sahrens dbuf_clear(db); /* exits db_mtx for us */ 376c543ec06Sahrens } else { 377c543ec06Sahrens mutex_exit(&db->db_mtx); 378c543ec06Sahrens } 379c543ec06Sahrens 380ea8dc4b6Seschrock } 381d90a49d6Smaybee list_remove(&dn->dn_dbufs, &marker); 382c543ec06Sahrens /* 383c543ec06Sahrens * NB: we need to drop dn_dbufs_mtx between passes so 384c543ec06Sahrens * that any DB_EVICTING dbufs can make progress. 385c543ec06Sahrens * Ideally, we would have some cv we could wait on, but 386c543ec06Sahrens * since we don't, just wait a bit to give the other 387c543ec06Sahrens * thread a chance to run. 388c543ec06Sahrens */ 389c543ec06Sahrens mutex_exit(&dn->dn_dbufs_mtx); 390c543ec06Sahrens if (evicting) 391c543ec06Sahrens delay(1); 392c543ec06Sahrens pass++; 393c543ec06Sahrens ASSERT(pass < 100); /* sanity check */ 394c543ec06Sahrens } while (progress); 395c543ec06Sahrens 396c543ec06Sahrens /* 397436b2950Sperrin * This function works fine even if it can't evict everything. 398436b2950Sperrin * If were only asked to try to evict everything then 399436b2950Sperrin * return an error if we can't. Otherwise panic as the caller 400436b2950Sperrin * expects total eviction. 401c543ec06Sahrens */ 402c543ec06Sahrens if (list_head(&dn->dn_dbufs) != NULL) { 403436b2950Sperrin if (try) { 404436b2950Sperrin return (1); 405436b2950Sperrin } else { 406436b2950Sperrin panic("dangling dbufs (dn=%p, dbuf=%p)\n", 407436b2950Sperrin dn, list_head(&dn->dn_dbufs)); 408436b2950Sperrin } 409ea8dc4b6Seschrock } 410c543ec06Sahrens 411ea8dc4b6Seschrock rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 412ea8dc4b6Seschrock if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) { 413ea8dc4b6Seschrock mutex_enter(&dn->dn_bonus->db_mtx); 414ea8dc4b6Seschrock dbuf_evict(dn->dn_bonus); 415ea8dc4b6Seschrock dn->dn_bonus = NULL; 416ea8dc4b6Seschrock } 417ea8dc4b6Seschrock rw_exit(&dn->dn_struct_rwlock); 418436b2950Sperrin return (0); 419ea8dc4b6Seschrock } 420ea8dc4b6Seschrock 421fa9e4066Sahrens static int 422fa9e4066Sahrens dnode_sync_free(dnode_t *dn, dmu_tx_t *tx) 423fa9e4066Sahrens { 424fa9e4066Sahrens dmu_buf_impl_t *db; 425fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 426fa9e4066Sahrens 427fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 428fa9e4066Sahrens 429fa9e4066Sahrens /* Undirty all buffers */ 430fa9e4066Sahrens while (db = list_head(&dn->dn_dirty_dbufs[txgoff])) { 431fa9e4066Sahrens mutex_enter(&db->db_mtx); 432fa9e4066Sahrens /* XXX - use dbuf_undirty()? */ 433fa9e4066Sahrens list_remove(&dn->dn_dirty_dbufs[txgoff], db); 434fa9e4066Sahrens if (db->db_level == 0) { 435ea8dc4b6Seschrock ASSERT(db->db_blkid == DB_BONUS_BLKID || 436ea8dc4b6Seschrock db->db_d.db_data_old[txgoff] == db->db_buf); 437fa9e4066Sahrens if (db->db_d.db_overridden_by[txgoff]) 438fa9e4066Sahrens dbuf_unoverride(db, tx->tx_txg); 439fa9e4066Sahrens db->db_d.db_data_old[txgoff] = NULL; 440fa9e4066Sahrens } 441fa9e4066Sahrens db->db_dirtycnt -= 1; 442fa9e4066Sahrens mutex_exit(&db->db_mtx); 443ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)tx->tx_txg); 444fa9e4066Sahrens } 445fa9e4066Sahrens 446436b2950Sperrin (void) dnode_evict_dbufs(dn, 0); 447ea8dc4b6Seschrock ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL); 448ea8dc4b6Seschrock 449ea8dc4b6Seschrock /* 450ea8dc4b6Seschrock * XXX - It would be nice to assert this, but we may still 451ea8dc4b6Seschrock * have residual holds from async evictions from the arc... 452ea8dc4b6Seschrock * 453*55434c77Sek * zfs_obj_to_path() also depends on this being 454*55434c77Sek * commented out. 455*55434c77Sek * 456ea8dc4b6Seschrock * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); 457ea8dc4b6Seschrock */ 458fa9e4066Sahrens 459fa9e4066Sahrens /* Undirty next bits */ 460fa9e4066Sahrens dn->dn_next_nlevels[txgoff] = 0; 461fa9e4066Sahrens dn->dn_next_indblkshift[txgoff] = 0; 462c543ec06Sahrens dn->dn_next_blksz[txgoff] = 0; 463fa9e4066Sahrens 464fa9e4066Sahrens /* free up all the blocks in the file. */ 465fa9e4066Sahrens dnode_sync_free_range(dn, 0, dn->dn_phys->dn_maxblkid+1, tx); 46699653d4eSeschrock ASSERT3U(DN_USED_BYTES(dn->dn_phys), ==, 0); 467fa9e4066Sahrens 468fa9e4066Sahrens /* ASSERT(blkptrs are zero); */ 469fa9e4066Sahrens ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE); 470fa9e4066Sahrens ASSERT(dn->dn_type != DMU_OT_NONE); 471fa9e4066Sahrens 472fa9e4066Sahrens ASSERT(dn->dn_free_txg > 0); 473fa9e4066Sahrens if (dn->dn_allocated_txg != dn->dn_free_txg) 474fa9e4066Sahrens dbuf_will_dirty(dn->dn_dbuf, tx); 475fa9e4066Sahrens bzero(dn->dn_phys, sizeof (dnode_phys_t)); 476fa9e4066Sahrens 477fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 478fa9e4066Sahrens dn->dn_type = DMU_OT_NONE; 479fa9e4066Sahrens dn->dn_maxblkid = 0; 480fa9e4066Sahrens dn->dn_allocated_txg = 0; 481fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 482fa9e4066Sahrens 483ea8dc4b6Seschrock ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 484fa9e4066Sahrens 485fa9e4066Sahrens dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg); 486fa9e4066Sahrens /* 487fa9e4066Sahrens * Now that we've released our hold, the dnode may 488fa9e4066Sahrens * be evicted, so we musn't access it. 489fa9e4066Sahrens */ 490fa9e4066Sahrens return (1); 491fa9e4066Sahrens } 492fa9e4066Sahrens 493fa9e4066Sahrens /* 494fa9e4066Sahrens * Write out the dnode's dirty buffers at the specified level. 495fa9e4066Sahrens * This may create more dirty buffers at the next level up. 496fa9e4066Sahrens * 497fa9e4066Sahrens * NOTE: The dnode is kept in memory by being dirty. Once the 498fa9e4066Sahrens * dirty bit is cleared, it may be evicted. Beware of this! 499fa9e4066Sahrens */ 500fa9e4066Sahrens int 501fa9e4066Sahrens dnode_sync(dnode_t *dn, int level, zio_t *zio, dmu_tx_t *tx) 502fa9e4066Sahrens { 503fa9e4066Sahrens free_range_t *rp; 504fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 505fa9e4066Sahrens dnode_phys_t *dnp = dn->dn_phys; 506fa9e4066Sahrens 507fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 508fa9e4066Sahrens ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg); 5099c9dc39aSek DNODE_VERIFY(dn); 510c543ec06Sahrens 511fa9e4066Sahrens /* 512fa9e4066Sahrens * Make sure the dbuf for the dn_phys is released before we modify it. 513fa9e4066Sahrens */ 514fa9e4066Sahrens if (dn->dn_dbuf) 515fa9e4066Sahrens arc_release(dn->dn_dbuf->db_buf, dn->dn_dbuf); 516fa9e4066Sahrens 517fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 518fa9e4066Sahrens if (dn->dn_allocated_txg == tx->tx_txg) { 519fa9e4066Sahrens /* The dnode is newly allocated or reallocated */ 520fa9e4066Sahrens if (dnp->dn_type == DMU_OT_NONE) { 521fa9e4066Sahrens /* this is a first alloc, not a realloc */ 522fa9e4066Sahrens /* XXX shouldn't the phys already be zeroed? */ 523fa9e4066Sahrens bzero(dnp, DNODE_CORE_SIZE); 524fa9e4066Sahrens dnp->dn_nlevels = 1; 525fa9e4066Sahrens } 526fa9e4066Sahrens 527fa9e4066Sahrens if (dn->dn_nblkptr > dnp->dn_nblkptr) { 528fa9e4066Sahrens /* zero the new blkptrs we are gaining */ 529fa9e4066Sahrens bzero(dnp->dn_blkptr + dnp->dn_nblkptr, 530fa9e4066Sahrens sizeof (blkptr_t) * 531fa9e4066Sahrens (dn->dn_nblkptr - dnp->dn_nblkptr)); 532fa9e4066Sahrens } 533fa9e4066Sahrens dnp->dn_type = dn->dn_type; 534fa9e4066Sahrens dnp->dn_bonustype = dn->dn_bonustype; 535fa9e4066Sahrens dnp->dn_bonuslen = dn->dn_bonuslen; 536fa9e4066Sahrens dnp->dn_nblkptr = dn->dn_nblkptr; 537fa9e4066Sahrens } 538fa9e4066Sahrens 539f676ed34Sahrens ASSERT(level != 0 || dnp->dn_nlevels > 1 || 540f676ed34Sahrens BP_IS_HOLE(&dnp->dn_blkptr[0]) || 541f676ed34Sahrens BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 542f676ed34Sahrens dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT); 543f676ed34Sahrens 544c543ec06Sahrens if (dn->dn_next_blksz[txgoff]) { 545c543ec06Sahrens ASSERT(P2PHASE(dn->dn_next_blksz[txgoff], 546fa9e4066Sahrens SPA_MINBLOCKSIZE) == 0); 547f676ed34Sahrens ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) || 548347a31bcSahrens list_head(&dn->dn_dirty_dbufs[txgoff]) != NULL || 549347a31bcSahrens dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT == 550347a31bcSahrens dnp->dn_datablkszsec); 551fa9e4066Sahrens dnp->dn_datablkszsec = 552c543ec06Sahrens dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT; 553c543ec06Sahrens dn->dn_next_blksz[txgoff] = 0; 554fa9e4066Sahrens } 555fa9e4066Sahrens 556fa9e4066Sahrens if (dn->dn_next_indblkshift[txgoff]) { 557fa9e4066Sahrens ASSERT(dnp->dn_nlevels == 1); 558fa9e4066Sahrens dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff]; 559fa9e4066Sahrens dn->dn_next_indblkshift[txgoff] = 0; 560fa9e4066Sahrens } 561fa9e4066Sahrens 562fa9e4066Sahrens /* 563fa9e4066Sahrens * Just take the live (open-context) values for checksum and compress. 564fa9e4066Sahrens * Strictly speaking it's a future leak, but nothing bad happens if we 565fa9e4066Sahrens * start using the new checksum or compress algorithm a little early. 566fa9e4066Sahrens */ 567fa9e4066Sahrens dnp->dn_checksum = dn->dn_checksum; 568fa9e4066Sahrens dnp->dn_compress = dn->dn_compress; 569fa9e4066Sahrens 570fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 571fa9e4066Sahrens 572fa9e4066Sahrens /* process all the "freed" ranges in the file */ 573fa9e4066Sahrens if (dn->dn_free_txg == 0 || dn->dn_free_txg > tx->tx_txg) { 57423b11526Smaybee for (rp = avl_last(&dn->dn_ranges[txgoff]); rp != NULL; 57523b11526Smaybee rp = AVL_PREV(&dn->dn_ranges[txgoff], rp)) 576fa9e4066Sahrens dnode_sync_free_range(dn, 577fa9e4066Sahrens rp->fr_blkid, rp->fr_nblks, tx); 578fa9e4066Sahrens } 579fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 580fa9e4066Sahrens for (rp = avl_first(&dn->dn_ranges[txgoff]); rp; ) { 581fa9e4066Sahrens free_range_t *last = rp; 582fa9e4066Sahrens rp = AVL_NEXT(&dn->dn_ranges[txgoff], rp); 583fa9e4066Sahrens avl_remove(&dn->dn_ranges[txgoff], last); 584fa9e4066Sahrens kmem_free(last, sizeof (free_range_t)); 585fa9e4066Sahrens } 586fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 587fa9e4066Sahrens 588fa9e4066Sahrens if (dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg) { 589fa9e4066Sahrens ASSERT3U(level, ==, 0); 590fa9e4066Sahrens return (dnode_sync_free(dn, tx)); 591fa9e4066Sahrens } 592fa9e4066Sahrens 593fa9e4066Sahrens if (dn->dn_next_nlevels[txgoff]) { 594fa9e4066Sahrens int new_lvl = dn->dn_next_nlevels[txgoff]; 595fa9e4066Sahrens 596fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 597fa9e4066Sahrens while (new_lvl > dnp->dn_nlevels) 598fa9e4066Sahrens dnode_increase_indirection(dn, tx); 599fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 600fa9e4066Sahrens dn->dn_next_nlevels[txgoff] = 0; 601fa9e4066Sahrens } 602fa9e4066Sahrens 603fa9e4066Sahrens if (level == dnp->dn_nlevels) { 604fa9e4066Sahrens uint64_t off = (dn->dn_phys->dn_maxblkid + 1) * 605fa9e4066Sahrens (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT); 606fa9e4066Sahrens 607fa9e4066Sahrens /* we've already synced out all data and indirect blocks */ 608fa9e4066Sahrens /* there are no more dirty dbufs under this dnode */ 609fa9e4066Sahrens ASSERT3P(list_head(&dn->dn_dirty_dbufs[txgoff]), ==, NULL); 610fa9e4066Sahrens ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= tx->tx_txg); 611fa9e4066Sahrens 612fa9e4066Sahrens /* NB: the "off < maxblkid" is to catch overflow */ 613fa9e4066Sahrens /* 614fa9e4066Sahrens * NB: if blocksize is changing, we could get confused, 615fa9e4066Sahrens * so only bother if there are multiple blocks and thus 616fa9e4066Sahrens * it can't be changing. 617fa9e4066Sahrens */ 6186754306eSahrens ASSERT(off < dn->dn_phys->dn_maxblkid || 619fa9e4066Sahrens dn->dn_phys->dn_maxblkid == 0 || 6206754306eSahrens dnode_next_offset(dn, FALSE, &off, 1, 1, 0) != 0); 621fa9e4066Sahrens 622f676ed34Sahrens ASSERT(dnp->dn_nlevels > 1 || 623f676ed34Sahrens BP_IS_HOLE(&dnp->dn_blkptr[0]) || 624f676ed34Sahrens BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 625f676ed34Sahrens dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT); 626f676ed34Sahrens 627ea8dc4b6Seschrock if (dn->dn_object != DMU_META_DNODE_OBJECT) { 628fa9e4066Sahrens dbuf_will_dirty(dn->dn_dbuf, tx); 629fa9e4066Sahrens dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg); 630fa9e4066Sahrens } 631fa9e4066Sahrens 632fa9e4066Sahrens /* 633fa9e4066Sahrens * Now that we've dropped the reference, the dnode may 634fa9e4066Sahrens * be evicted, so we musn't access it. 635fa9e4066Sahrens */ 636fa9e4066Sahrens return (1); 637fa9e4066Sahrens } else { 638fa9e4066Sahrens dmu_buf_impl_t *db, *db_next; 639fa9e4066Sahrens list_t *list = &dn->dn_dirty_dbufs[txgoff]; 640fa9e4066Sahrens /* 641fa9e4066Sahrens * Iterate over the list, removing and sync'ing dbufs 642fa9e4066Sahrens * which are on the level we want, and leaving others. 643fa9e4066Sahrens */ 644fa9e4066Sahrens for (db = list_head(list); db; db = db_next) { 645fa9e4066Sahrens db_next = list_next(list, db); 646fa9e4066Sahrens if (db->db_level == level) { 647fa9e4066Sahrens list_remove(list, db); 648fa9e4066Sahrens dbuf_sync(db, zio, tx); 649fa9e4066Sahrens } 650fa9e4066Sahrens } 651fa9e4066Sahrens return (0); 652fa9e4066Sahrens } 653fa9e4066Sahrens } 654