1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5f65e61c0Sahrens * Common Development and Distribution License (the "License"). 6f65e61c0Sahrens * 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 /* 22e7437265Sahrens * 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 29fa9e4066Sahrens /* 30fa9e4066Sahrens * This file contains the top half of the zfs directory structure 31fa9e4066Sahrens * implementation. The bottom half is in zap_leaf.c. 32fa9e4066Sahrens * 33fa9e4066Sahrens * The zdir is an extendable hash data structure. There is a table of 34fa9e4066Sahrens * pointers to buckets (zap_t->zd_data->zd_leafs). The buckets are 35fa9e4066Sahrens * each a constant size and hold a variable number of directory entries. 36fa9e4066Sahrens * The buckets (aka "leaf nodes") are implemented in zap_leaf.c. 37fa9e4066Sahrens * 38fa9e4066Sahrens * The pointer table holds a power of 2 number of pointers. 39fa9e4066Sahrens * (1<<zap_t->zd_data->zd_phys->zd_prefix_len). The bucket pointed to 40fa9e4066Sahrens * by the pointer at index i in the table holds entries whose hash value 41fa9e4066Sahrens * has a zd_prefix_len - bit prefix 42fa9e4066Sahrens */ 43fa9e4066Sahrens 44fa9e4066Sahrens #include <sys/spa.h> 45fa9e4066Sahrens #include <sys/dmu.h> 46fa9e4066Sahrens #include <sys/zfs_context.h> 47fa9e4066Sahrens #include <sys/zap.h> 48ea8dc4b6Seschrock #include <sys/refcount.h> 49fa9e4066Sahrens #include <sys/zap_impl.h> 50fa9e4066Sahrens #include <sys/zap_leaf.h> 51fa9e4066Sahrens 52f65e61c0Sahrens int fzap_default_block_shift = 14; /* 16k blocksize */ 53fa9e4066Sahrens 54fa9e4066Sahrens static void zap_leaf_pageout(dmu_buf_t *db, void *vl); 5566328dd3Sahrens static uint64_t zap_allocate_blocks(zap_t *zap, int nblocks); 56fa9e4066Sahrens 57fa9e4066Sahrens 58fa9e4066Sahrens void 59fa9e4066Sahrens fzap_byteswap(void *vbuf, size_t size) 60fa9e4066Sahrens { 61fa9e4066Sahrens uint64_t block_type; 62fa9e4066Sahrens 63fa9e4066Sahrens block_type = *(uint64_t *)vbuf; 64fa9e4066Sahrens 655ad82045Snd if (block_type == ZBT_LEAF || block_type == BSWAP_64(ZBT_LEAF)) 66f65e61c0Sahrens zap_leaf_byteswap(vbuf, size); 675ad82045Snd else { 68fa9e4066Sahrens /* it's a ptrtbl block */ 69f65e61c0Sahrens byteswap_uint64_array(vbuf, size); 70fa9e4066Sahrens } 71fa9e4066Sahrens } 72fa9e4066Sahrens 73fa9e4066Sahrens void 74fa9e4066Sahrens fzap_upgrade(zap_t *zap, dmu_tx_t *tx) 75fa9e4066Sahrens { 76fa9e4066Sahrens dmu_buf_t *db; 77fa9e4066Sahrens zap_leaf_t *l; 78fa9e4066Sahrens int i; 79fa9e4066Sahrens zap_phys_t *zp; 80fa9e4066Sahrens 81fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 82fa9e4066Sahrens zap->zap_ismicro = FALSE; 83fa9e4066Sahrens 84fa9e4066Sahrens (void) dmu_buf_update_user(zap->zap_dbuf, zap, zap, 85455d5089Sahrens &zap->zap_f.zap_phys, zap_evict); 86fa9e4066Sahrens 87fa9e4066Sahrens mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, 0, 0); 88f65e61c0Sahrens zap->zap_f.zap_block_shift = highbit(zap->zap_dbuf->db_size) - 1; 89fa9e4066Sahrens 90fa9e4066Sahrens zp = zap->zap_f.zap_phys; 91fa9e4066Sahrens /* 92fa9e4066Sahrens * explicitly zero it since it might be coming from an 93fa9e4066Sahrens * initialized microzap 94fa9e4066Sahrens */ 95f65e61c0Sahrens bzero(zap->zap_dbuf->db_data, zap->zap_dbuf->db_size); 96fa9e4066Sahrens zp->zap_block_type = ZBT_HEADER; 97fa9e4066Sahrens zp->zap_magic = ZAP_MAGIC; 98fa9e4066Sahrens 99f65e61c0Sahrens zp->zap_ptrtbl.zt_shift = ZAP_EMBEDDED_PTRTBL_SHIFT(zap); 100fa9e4066Sahrens 101fa9e4066Sahrens zp->zap_freeblk = 2; /* block 1 will be the first leaf */ 102fa9e4066Sahrens zp->zap_num_leafs = 1; 103fa9e4066Sahrens zp->zap_num_entries = 0; 104fa9e4066Sahrens zp->zap_salt = zap->zap_salt; 105da6c28aaSamw zp->zap_normflags = zap->zap_normflags; 106fa9e4066Sahrens 107f65e61c0Sahrens /* block 1 will be the first leaf */ 108f65e61c0Sahrens for (i = 0; i < (1<<zp->zap_ptrtbl.zt_shift); i++) 109f65e61c0Sahrens ZAP_EMBEDDED_PTRTBL_ENT(zap, i) = 1; 110fa9e4066Sahrens 111fa9e4066Sahrens /* 112fa9e4066Sahrens * set up block 1 - the first leaf 113fa9e4066Sahrens */ 114ea8dc4b6Seschrock VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object, 115ea8dc4b6Seschrock 1<<FZAP_BLOCK_SHIFT(zap), FTAG, &db)); 116fa9e4066Sahrens dmu_buf_will_dirty(db, tx); 117fa9e4066Sahrens 118fa9e4066Sahrens l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP); 119fa9e4066Sahrens l->l_dbuf = db; 120fa9e4066Sahrens l->l_phys = db->db_data; 121fa9e4066Sahrens 122da6c28aaSamw zap_leaf_init(l, spa_version(dmu_objset_spa(zap->zap_objset))); 123fa9e4066Sahrens 124fa9e4066Sahrens kmem_free(l, sizeof (zap_leaf_t)); 125ea8dc4b6Seschrock dmu_buf_rele(db, FTAG); 126fa9e4066Sahrens } 127fa9e4066Sahrens 128fa9e4066Sahrens static int 129fa9e4066Sahrens zap_tryupgradedir(zap_t *zap, dmu_tx_t *tx) 130fa9e4066Sahrens { 131fa9e4066Sahrens if (RW_WRITE_HELD(&zap->zap_rwlock)) 132fa9e4066Sahrens return (1); 133fa9e4066Sahrens if (rw_tryupgrade(&zap->zap_rwlock)) { 134fa9e4066Sahrens dmu_buf_will_dirty(zap->zap_dbuf, tx); 135fa9e4066Sahrens return (1); 136fa9e4066Sahrens } 137fa9e4066Sahrens return (0); 138fa9e4066Sahrens } 139fa9e4066Sahrens 140fa9e4066Sahrens /* 141fa9e4066Sahrens * Generic routines for dealing with the pointer & cookie tables. 142fa9e4066Sahrens */ 143fa9e4066Sahrens 14466328dd3Sahrens static int 145fa9e4066Sahrens zap_table_grow(zap_t *zap, zap_table_phys_t *tbl, 146fa9e4066Sahrens void (*transfer_func)(const uint64_t *src, uint64_t *dst, int n), 147fa9e4066Sahrens dmu_tx_t *tx) 148fa9e4066Sahrens { 149fa9e4066Sahrens uint64_t b, newblk; 150fa9e4066Sahrens dmu_buf_t *db_old, *db_new; 151ea8dc4b6Seschrock int err; 152f65e61c0Sahrens int bs = FZAP_BLOCK_SHIFT(zap); 153f65e61c0Sahrens int hepb = 1<<(bs-4); 154fa9e4066Sahrens /* hepb = half the number of entries in a block */ 155fa9e4066Sahrens 156fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 157fa9e4066Sahrens ASSERT(tbl->zt_blk != 0); 158fa9e4066Sahrens ASSERT(tbl->zt_numblks > 0); 159fa9e4066Sahrens 160fa9e4066Sahrens if (tbl->zt_nextblk != 0) { 161fa9e4066Sahrens newblk = tbl->zt_nextblk; 162fa9e4066Sahrens } else { 16366328dd3Sahrens newblk = zap_allocate_blocks(zap, tbl->zt_numblks * 2); 164fa9e4066Sahrens tbl->zt_nextblk = newblk; 165fa9e4066Sahrens ASSERT3U(tbl->zt_blks_copied, ==, 0); 166fa9e4066Sahrens dmu_prefetch(zap->zap_objset, zap->zap_object, 167f65e61c0Sahrens tbl->zt_blk << bs, tbl->zt_numblks << bs); 168fa9e4066Sahrens } 169fa9e4066Sahrens 170fa9e4066Sahrens /* 17166328dd3Sahrens * Copy the ptrtbl from the old to new location. 172fa9e4066Sahrens */ 173fa9e4066Sahrens 174fa9e4066Sahrens b = tbl->zt_blks_copied; 175ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 176ea8dc4b6Seschrock (tbl->zt_blk + b) << bs, FTAG, &db_old); 177ea8dc4b6Seschrock if (err) 17866328dd3Sahrens return (err); 179fa9e4066Sahrens 180fa9e4066Sahrens /* first half of entries in old[b] go to new[2*b+0] */ 181ea8dc4b6Seschrock VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object, 182ea8dc4b6Seschrock (newblk + 2*b+0) << bs, FTAG, &db_new)); 183fa9e4066Sahrens dmu_buf_will_dirty(db_new, tx); 184fa9e4066Sahrens transfer_func(db_old->db_data, db_new->db_data, hepb); 185ea8dc4b6Seschrock dmu_buf_rele(db_new, FTAG); 186fa9e4066Sahrens 187fa9e4066Sahrens /* second half of entries in old[b] go to new[2*b+1] */ 188ea8dc4b6Seschrock VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object, 189ea8dc4b6Seschrock (newblk + 2*b+1) << bs, FTAG, &db_new)); 190fa9e4066Sahrens dmu_buf_will_dirty(db_new, tx); 191fa9e4066Sahrens transfer_func((uint64_t *)db_old->db_data + hepb, 192fa9e4066Sahrens db_new->db_data, hepb); 193ea8dc4b6Seschrock dmu_buf_rele(db_new, FTAG); 194fa9e4066Sahrens 195ea8dc4b6Seschrock dmu_buf_rele(db_old, FTAG); 196fa9e4066Sahrens 197fa9e4066Sahrens tbl->zt_blks_copied++; 198fa9e4066Sahrens 199fa9e4066Sahrens dprintf("copied block %llu of %llu\n", 200fa9e4066Sahrens tbl->zt_blks_copied, tbl->zt_numblks); 201fa9e4066Sahrens 202fa9e4066Sahrens if (tbl->zt_blks_copied == tbl->zt_numblks) { 203ea8dc4b6Seschrock (void) dmu_free_range(zap->zap_objset, zap->zap_object, 204f65e61c0Sahrens tbl->zt_blk << bs, tbl->zt_numblks << bs, tx); 205fa9e4066Sahrens 206fa9e4066Sahrens tbl->zt_blk = newblk; 207fa9e4066Sahrens tbl->zt_numblks *= 2; 208fa9e4066Sahrens tbl->zt_shift++; 209fa9e4066Sahrens tbl->zt_nextblk = 0; 210fa9e4066Sahrens tbl->zt_blks_copied = 0; 211fa9e4066Sahrens 212fa9e4066Sahrens dprintf("finished; numblocks now %llu (%lluk entries)\n", 213fa9e4066Sahrens tbl->zt_numblks, 1<<(tbl->zt_shift-10)); 214fa9e4066Sahrens } 21566328dd3Sahrens 21666328dd3Sahrens return (0); 217fa9e4066Sahrens } 218fa9e4066Sahrens 219ea8dc4b6Seschrock static int 220fa9e4066Sahrens zap_table_store(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t val, 221fa9e4066Sahrens dmu_tx_t *tx) 222fa9e4066Sahrens { 223ea8dc4b6Seschrock int err; 224ea8dc4b6Seschrock uint64_t blk, off; 225f65e61c0Sahrens int bs = FZAP_BLOCK_SHIFT(zap); 226ea8dc4b6Seschrock dmu_buf_t *db; 227fa9e4066Sahrens 228fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 229fa9e4066Sahrens ASSERT(tbl->zt_blk != 0); 230fa9e4066Sahrens 231fa9e4066Sahrens dprintf("storing %llx at index %llx\n", val, idx); 232fa9e4066Sahrens 233f65e61c0Sahrens blk = idx >> (bs-3); 234f65e61c0Sahrens off = idx & ((1<<(bs-3))-1); 235fa9e4066Sahrens 236ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 237ea8dc4b6Seschrock (tbl->zt_blk + blk) << bs, FTAG, &db); 238ea8dc4b6Seschrock if (err) 239ea8dc4b6Seschrock return (err); 240fa9e4066Sahrens dmu_buf_will_dirty(db, tx); 241fa9e4066Sahrens 242fa9e4066Sahrens if (tbl->zt_nextblk != 0) { 243ea8dc4b6Seschrock uint64_t idx2 = idx * 2; 244ea8dc4b6Seschrock uint64_t blk2 = idx2 >> (bs-3); 245ea8dc4b6Seschrock uint64_t off2 = idx2 & ((1<<(bs-3))-1); 246ea8dc4b6Seschrock dmu_buf_t *db2; 247ea8dc4b6Seschrock 248ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 249ea8dc4b6Seschrock (tbl->zt_nextblk + blk2) << bs, FTAG, &db2); 250ea8dc4b6Seschrock if (err) { 251ea8dc4b6Seschrock dmu_buf_rele(db, FTAG); 252ea8dc4b6Seschrock return (err); 253ea8dc4b6Seschrock } 254ea8dc4b6Seschrock dmu_buf_will_dirty(db2, tx); 255ea8dc4b6Seschrock ((uint64_t *)db2->db_data)[off2] = val; 256ea8dc4b6Seschrock ((uint64_t *)db2->db_data)[off2+1] = val; 257ea8dc4b6Seschrock dmu_buf_rele(db2, FTAG); 258fa9e4066Sahrens } 259fa9e4066Sahrens 260ea8dc4b6Seschrock ((uint64_t *)db->db_data)[off] = val; 261ea8dc4b6Seschrock dmu_buf_rele(db, FTAG); 262ea8dc4b6Seschrock 263ea8dc4b6Seschrock return (0); 264fa9e4066Sahrens } 265fa9e4066Sahrens 266ea8dc4b6Seschrock static int 267ea8dc4b6Seschrock zap_table_load(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t *valp) 268fa9e4066Sahrens { 269ea8dc4b6Seschrock uint64_t blk, off; 270ea8dc4b6Seschrock int err; 271fa9e4066Sahrens dmu_buf_t *db; 272f65e61c0Sahrens int bs = FZAP_BLOCK_SHIFT(zap); 273fa9e4066Sahrens 274fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 275fa9e4066Sahrens 276f65e61c0Sahrens blk = idx >> (bs-3); 277f65e61c0Sahrens off = idx & ((1<<(bs-3))-1); 278fa9e4066Sahrens 279ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 280ea8dc4b6Seschrock (tbl->zt_blk + blk) << bs, FTAG, &db); 281ea8dc4b6Seschrock if (err) 282ea8dc4b6Seschrock return (err); 283ea8dc4b6Seschrock *valp = ((uint64_t *)db->db_data)[off]; 284ea8dc4b6Seschrock dmu_buf_rele(db, FTAG); 285ea8dc4b6Seschrock 286ea8dc4b6Seschrock if (tbl->zt_nextblk != 0) { 287ea8dc4b6Seschrock /* 288ea8dc4b6Seschrock * read the nextblk for the sake of i/o error checking, 289ea8dc4b6Seschrock * so that zap_table_load() will catch errors for 290ea8dc4b6Seschrock * zap_table_store. 291ea8dc4b6Seschrock */ 292ea8dc4b6Seschrock blk = (idx*2) >> (bs-3); 293ea8dc4b6Seschrock 294ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 295ea8dc4b6Seschrock (tbl->zt_nextblk + blk) << bs, FTAG, &db); 296ea8dc4b6Seschrock dmu_buf_rele(db, FTAG); 297ea8dc4b6Seschrock } 298ea8dc4b6Seschrock return (err); 299fa9e4066Sahrens } 300fa9e4066Sahrens 301fa9e4066Sahrens /* 302fa9e4066Sahrens * Routines for growing the ptrtbl. 303fa9e4066Sahrens */ 304fa9e4066Sahrens 305fa9e4066Sahrens static void 306fa9e4066Sahrens zap_ptrtbl_transfer(const uint64_t *src, uint64_t *dst, int n) 307fa9e4066Sahrens { 308fa9e4066Sahrens int i; 309fa9e4066Sahrens for (i = 0; i < n; i++) { 310fa9e4066Sahrens uint64_t lb = src[i]; 311fa9e4066Sahrens dst[2*i+0] = lb; 312fa9e4066Sahrens dst[2*i+1] = lb; 313fa9e4066Sahrens } 314fa9e4066Sahrens } 315fa9e4066Sahrens 31666328dd3Sahrens static int 317fa9e4066Sahrens zap_grow_ptrtbl(zap_t *zap, dmu_tx_t *tx) 318fa9e4066Sahrens { 31966328dd3Sahrens /* In case things go horribly wrong. */ 32066328dd3Sahrens if (zap->zap_f.zap_phys->zap_ptrtbl.zt_shift >= ZAP_HASHBITS-2) 32166328dd3Sahrens return (ENOSPC); 322fa9e4066Sahrens 323fa9e4066Sahrens if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) { 324fa9e4066Sahrens /* 325f65e61c0Sahrens * We are outgrowing the "embedded" ptrtbl (the one 326f65e61c0Sahrens * stored in the header block). Give it its own entire 327f65e61c0Sahrens * block, which will double the size of the ptrtbl. 328fa9e4066Sahrens */ 329fa9e4066Sahrens uint64_t newblk; 330fa9e4066Sahrens dmu_buf_t *db_new; 331ea8dc4b6Seschrock int err; 332fa9e4066Sahrens 333fa9e4066Sahrens ASSERT3U(zap->zap_f.zap_phys->zap_ptrtbl.zt_shift, ==, 334f65e61c0Sahrens ZAP_EMBEDDED_PTRTBL_SHIFT(zap)); 335fa9e4066Sahrens ASSERT3U(zap->zap_f.zap_phys->zap_ptrtbl.zt_blk, ==, 0); 336fa9e4066Sahrens 33766328dd3Sahrens newblk = zap_allocate_blocks(zap, 1); 338ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 339ea8dc4b6Seschrock newblk << FZAP_BLOCK_SHIFT(zap), FTAG, &db_new); 340ea8dc4b6Seschrock if (err) 34166328dd3Sahrens return (err); 342fa9e4066Sahrens dmu_buf_will_dirty(db_new, tx); 343f65e61c0Sahrens zap_ptrtbl_transfer(&ZAP_EMBEDDED_PTRTBL_ENT(zap, 0), 344f65e61c0Sahrens db_new->db_data, 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap)); 345ea8dc4b6Seschrock dmu_buf_rele(db_new, FTAG); 346fa9e4066Sahrens 347fa9e4066Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_blk = newblk; 348fa9e4066Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks = 1; 349fa9e4066Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_shift++; 350fa9e4066Sahrens 351fa9e4066Sahrens ASSERT3U(1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift, ==, 352fa9e4066Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks << 353f65e61c0Sahrens (FZAP_BLOCK_SHIFT(zap)-3)); 35466328dd3Sahrens 35566328dd3Sahrens return (0); 356fa9e4066Sahrens } else { 35766328dd3Sahrens return (zap_table_grow(zap, &zap->zap_f.zap_phys->zap_ptrtbl, 35866328dd3Sahrens zap_ptrtbl_transfer, tx)); 359fa9e4066Sahrens } 360fa9e4066Sahrens } 361fa9e4066Sahrens 362fa9e4066Sahrens static void 363fa9e4066Sahrens zap_increment_num_entries(zap_t *zap, int delta, dmu_tx_t *tx) 364fa9e4066Sahrens { 365fa9e4066Sahrens dmu_buf_will_dirty(zap->zap_dbuf, tx); 366fa9e4066Sahrens mutex_enter(&zap->zap_f.zap_num_entries_mtx); 367fa9e4066Sahrens ASSERT(delta > 0 || zap->zap_f.zap_phys->zap_num_entries >= -delta); 368fa9e4066Sahrens zap->zap_f.zap_phys->zap_num_entries += delta; 369fa9e4066Sahrens mutex_exit(&zap->zap_f.zap_num_entries_mtx); 370fa9e4066Sahrens } 371fa9e4066Sahrens 37266328dd3Sahrens static uint64_t 37366328dd3Sahrens zap_allocate_blocks(zap_t *zap, int nblocks) 374fa9e4066Sahrens { 375fa9e4066Sahrens uint64_t newblk; 37666328dd3Sahrens ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 37766328dd3Sahrens newblk = zap->zap_f.zap_phys->zap_freeblk; 37866328dd3Sahrens zap->zap_f.zap_phys->zap_freeblk += nblocks; 379fa9e4066Sahrens return (newblk); 380fa9e4066Sahrens } 381fa9e4066Sahrens 38266328dd3Sahrens static zap_leaf_t * 383fa9e4066Sahrens zap_create_leaf(zap_t *zap, dmu_tx_t *tx) 384fa9e4066Sahrens { 385fa9e4066Sahrens void *winner; 386fa9e4066Sahrens zap_leaf_t *l = kmem_alloc(sizeof (zap_leaf_t), KM_SLEEP); 387fa9e4066Sahrens 388fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 389fa9e4066Sahrens 390fa9e4066Sahrens rw_init(&l->l_rwlock, 0, 0, 0); 391fa9e4066Sahrens rw_enter(&l->l_rwlock, RW_WRITER); 39266328dd3Sahrens l->l_blkid = zap_allocate_blocks(zap, 1); 393fa9e4066Sahrens l->l_dbuf = NULL; 394fa9e4066Sahrens l->l_phys = NULL; 395fa9e4066Sahrens 396ea8dc4b6Seschrock VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object, 397ea8dc4b6Seschrock l->l_blkid << FZAP_BLOCK_SHIFT(zap), NULL, &l->l_dbuf)); 398fa9e4066Sahrens winner = dmu_buf_set_user(l->l_dbuf, l, &l->l_phys, zap_leaf_pageout); 399fa9e4066Sahrens ASSERT(winner == NULL); 400fa9e4066Sahrens dmu_buf_will_dirty(l->l_dbuf, tx); 401fa9e4066Sahrens 402da6c28aaSamw zap_leaf_init(l, spa_version(dmu_objset_spa(zap->zap_objset))); 403fa9e4066Sahrens 40466328dd3Sahrens zap->zap_f.zap_phys->zap_num_leafs++; 405fa9e4066Sahrens 40666328dd3Sahrens return (l); 407fa9e4066Sahrens } 408fa9e4066Sahrens 409fa9e4066Sahrens int 410fa9e4066Sahrens fzap_count(zap_t *zap, uint64_t *count) 411fa9e4066Sahrens { 412fa9e4066Sahrens ASSERT(!zap->zap_ismicro); 413fa9e4066Sahrens mutex_enter(&zap->zap_f.zap_num_entries_mtx); /* unnecessary */ 414fa9e4066Sahrens *count = zap->zap_f.zap_phys->zap_num_entries; 415fa9e4066Sahrens mutex_exit(&zap->zap_f.zap_num_entries_mtx); 416fa9e4066Sahrens return (0); 417fa9e4066Sahrens } 418fa9e4066Sahrens 419fa9e4066Sahrens /* 420fa9e4066Sahrens * Routines for obtaining zap_leaf_t's 421fa9e4066Sahrens */ 422fa9e4066Sahrens 42387e5029aSahrens void 424fa9e4066Sahrens zap_put_leaf(zap_leaf_t *l) 425fa9e4066Sahrens { 426fa9e4066Sahrens rw_exit(&l->l_rwlock); 427ea8dc4b6Seschrock dmu_buf_rele(l->l_dbuf, NULL); 428fa9e4066Sahrens } 429fa9e4066Sahrens 430fa9e4066Sahrens _NOTE(ARGSUSED(0)) 431fa9e4066Sahrens static void 432fa9e4066Sahrens zap_leaf_pageout(dmu_buf_t *db, void *vl) 433fa9e4066Sahrens { 434fa9e4066Sahrens zap_leaf_t *l = vl; 435fa9e4066Sahrens 436fa9e4066Sahrens rw_destroy(&l->l_rwlock); 437fa9e4066Sahrens kmem_free(l, sizeof (zap_leaf_t)); 438fa9e4066Sahrens } 439fa9e4066Sahrens 440fa9e4066Sahrens static zap_leaf_t * 441fa9e4066Sahrens zap_open_leaf(uint64_t blkid, dmu_buf_t *db) 442fa9e4066Sahrens { 443fa9e4066Sahrens zap_leaf_t *l, *winner; 444fa9e4066Sahrens 445fa9e4066Sahrens ASSERT(blkid != 0); 446fa9e4066Sahrens 447fa9e4066Sahrens l = kmem_alloc(sizeof (zap_leaf_t), KM_SLEEP); 448fa9e4066Sahrens rw_init(&l->l_rwlock, 0, 0, 0); 449fa9e4066Sahrens rw_enter(&l->l_rwlock, RW_WRITER); 450fa9e4066Sahrens l->l_blkid = blkid; 451f65e61c0Sahrens l->l_bs = highbit(db->db_size)-1; 452fa9e4066Sahrens l->l_dbuf = db; 453fa9e4066Sahrens l->l_phys = NULL; 454fa9e4066Sahrens 455fa9e4066Sahrens winner = dmu_buf_set_user(db, l, &l->l_phys, zap_leaf_pageout); 456fa9e4066Sahrens 457fa9e4066Sahrens rw_exit(&l->l_rwlock); 458fa9e4066Sahrens if (winner != NULL) { 459fa9e4066Sahrens /* someone else set it first */ 460fa9e4066Sahrens zap_leaf_pageout(NULL, l); 461fa9e4066Sahrens l = winner; 462fa9e4066Sahrens } 463fa9e4066Sahrens 46466328dd3Sahrens /* 46566328dd3Sahrens * lhr_pad was previously used for the next leaf in the leaf 46666328dd3Sahrens * chain. There should be no chained leafs (as we have removed 46766328dd3Sahrens * support for them). 46866328dd3Sahrens */ 46966328dd3Sahrens ASSERT3U(l->l_phys->l_hdr.lh_pad1, ==, 0); 47066328dd3Sahrens 471f65e61c0Sahrens /* 472f65e61c0Sahrens * There should be more hash entries than there can be 473f65e61c0Sahrens * chunks to put in the hash table 474f65e61c0Sahrens */ 475f65e61c0Sahrens ASSERT3U(ZAP_LEAF_HASH_NUMENTRIES(l), >, ZAP_LEAF_NUMCHUNKS(l) / 3); 476f65e61c0Sahrens 477f65e61c0Sahrens /* The chunks should begin at the end of the hash table */ 478f65e61c0Sahrens ASSERT3P(&ZAP_LEAF_CHUNK(l, 0), ==, 479f65e61c0Sahrens &l->l_phys->l_hash[ZAP_LEAF_HASH_NUMENTRIES(l)]); 480f65e61c0Sahrens 481f65e61c0Sahrens /* The chunks should end at the end of the block */ 482f65e61c0Sahrens ASSERT3U((uintptr_t)&ZAP_LEAF_CHUNK(l, ZAP_LEAF_NUMCHUNKS(l)) - 483f65e61c0Sahrens (uintptr_t)l->l_phys, ==, l->l_dbuf->db_size); 484f65e61c0Sahrens 485fa9e4066Sahrens return (l); 486fa9e4066Sahrens } 487fa9e4066Sahrens 488ea8dc4b6Seschrock static int 48966328dd3Sahrens zap_get_leaf_byblk(zap_t *zap, uint64_t blkid, dmu_tx_t *tx, krw_t lt, 490ea8dc4b6Seschrock zap_leaf_t **lp) 491fa9e4066Sahrens { 492fa9e4066Sahrens dmu_buf_t *db; 493fa9e4066Sahrens zap_leaf_t *l; 494f65e61c0Sahrens int bs = FZAP_BLOCK_SHIFT(zap); 495ea8dc4b6Seschrock int err; 496fa9e4066Sahrens 497fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 498fa9e4066Sahrens 499ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 500ea8dc4b6Seschrock blkid << bs, NULL, &db); 501ea8dc4b6Seschrock if (err) 502ea8dc4b6Seschrock return (err); 503fa9e4066Sahrens 504fa9e4066Sahrens ASSERT3U(db->db_object, ==, zap->zap_object); 505f65e61c0Sahrens ASSERT3U(db->db_offset, ==, blkid << bs); 506f65e61c0Sahrens ASSERT3U(db->db_size, ==, 1 << bs); 507fa9e4066Sahrens ASSERT(blkid != 0); 508fa9e4066Sahrens 509fa9e4066Sahrens l = dmu_buf_get_user(db); 510fa9e4066Sahrens 511fa9e4066Sahrens if (l == NULL) 512fa9e4066Sahrens l = zap_open_leaf(blkid, db); 513fa9e4066Sahrens 514fa9e4066Sahrens rw_enter(&l->l_rwlock, lt); 515fa9e4066Sahrens /* 516fa9e4066Sahrens * Must lock before dirtying, otherwise l->l_phys could change, 517fa9e4066Sahrens * causing ASSERT below to fail. 518fa9e4066Sahrens */ 519fa9e4066Sahrens if (lt == RW_WRITER) 520fa9e4066Sahrens dmu_buf_will_dirty(db, tx); 521fa9e4066Sahrens ASSERT3U(l->l_blkid, ==, blkid); 522fa9e4066Sahrens ASSERT3P(l->l_dbuf, ==, db); 523fa9e4066Sahrens ASSERT3P(l->l_phys, ==, l->l_dbuf->db_data); 52466328dd3Sahrens ASSERT3U(l->l_phys->l_hdr.lh_block_type, ==, ZBT_LEAF); 52566328dd3Sahrens ASSERT3U(l->l_phys->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC); 526fa9e4066Sahrens 527ea8dc4b6Seschrock *lp = l; 528ea8dc4b6Seschrock return (0); 529fa9e4066Sahrens } 530fa9e4066Sahrens 531ea8dc4b6Seschrock static int 532ea8dc4b6Seschrock zap_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t *valp) 533fa9e4066Sahrens { 534fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 535fa9e4066Sahrens 536fa9e4066Sahrens if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) { 537fa9e4066Sahrens ASSERT3U(idx, <, 538fa9e4066Sahrens (1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift)); 539ea8dc4b6Seschrock *valp = ZAP_EMBEDDED_PTRTBL_ENT(zap, idx); 540ea8dc4b6Seschrock return (0); 541fa9e4066Sahrens } else { 542fa9e4066Sahrens return (zap_table_load(zap, &zap->zap_f.zap_phys->zap_ptrtbl, 543ea8dc4b6Seschrock idx, valp)); 544fa9e4066Sahrens } 545fa9e4066Sahrens } 546fa9e4066Sahrens 547ea8dc4b6Seschrock static int 548fa9e4066Sahrens zap_set_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t blk, dmu_tx_t *tx) 549fa9e4066Sahrens { 550fa9e4066Sahrens ASSERT(tx != NULL); 551fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 552fa9e4066Sahrens 553fa9e4066Sahrens if (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk == 0) { 554f65e61c0Sahrens ZAP_EMBEDDED_PTRTBL_ENT(zap, idx) = blk; 555ea8dc4b6Seschrock return (0); 556fa9e4066Sahrens } else { 557ea8dc4b6Seschrock return (zap_table_store(zap, &zap->zap_f.zap_phys->zap_ptrtbl, 558ea8dc4b6Seschrock idx, blk, tx)); 559fa9e4066Sahrens } 560fa9e4066Sahrens } 561fa9e4066Sahrens 562ea8dc4b6Seschrock static int 563ea8dc4b6Seschrock zap_deref_leaf(zap_t *zap, uint64_t h, dmu_tx_t *tx, krw_t lt, zap_leaf_t **lp) 564fa9e4066Sahrens { 565ea8dc4b6Seschrock uint64_t idx, blk; 566ea8dc4b6Seschrock int err; 567fa9e4066Sahrens 568fa9e4066Sahrens ASSERT(zap->zap_dbuf == NULL || 569fa9e4066Sahrens zap->zap_f.zap_phys == zap->zap_dbuf->db_data); 570fa9e4066Sahrens ASSERT3U(zap->zap_f.zap_phys->zap_magic, ==, ZAP_MAGIC); 571fa9e4066Sahrens idx = ZAP_HASH_IDX(h, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift); 572ea8dc4b6Seschrock err = zap_idx_to_blk(zap, idx, &blk); 573ea8dc4b6Seschrock if (err != 0) 574ea8dc4b6Seschrock return (err); 575ea8dc4b6Seschrock err = zap_get_leaf_byblk(zap, blk, tx, lt, lp); 576fa9e4066Sahrens 57766328dd3Sahrens ASSERT(err || ZAP_HASH_IDX(h, (*lp)->l_phys->l_hdr.lh_prefix_len) == 57866328dd3Sahrens (*lp)->l_phys->l_hdr.lh_prefix); 579ea8dc4b6Seschrock return (err); 580fa9e4066Sahrens } 581fa9e4066Sahrens 582ea8dc4b6Seschrock static int 583ea8dc4b6Seschrock zap_expand_leaf(zap_t *zap, zap_leaf_t *l, uint64_t hash, dmu_tx_t *tx, 584ea8dc4b6Seschrock zap_leaf_t **lp) 585fa9e4066Sahrens { 586fa9e4066Sahrens zap_leaf_t *nl; 587fa9e4066Sahrens int prefix_diff, i, err; 588fa9e4066Sahrens uint64_t sibling; 58966328dd3Sahrens int old_prefix_len = l->l_phys->l_hdr.lh_prefix_len; 590fa9e4066Sahrens 59166328dd3Sahrens ASSERT3U(old_prefix_len, <=, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift); 592fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 593fa9e4066Sahrens 59466328dd3Sahrens ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==, 59566328dd3Sahrens l->l_phys->l_hdr.lh_prefix); 596fa9e4066Sahrens 59766328dd3Sahrens if (zap_tryupgradedir(zap, tx) == 0 || 59866328dd3Sahrens old_prefix_len == zap->zap_f.zap_phys->zap_ptrtbl.zt_shift) { 59966328dd3Sahrens /* We failed to upgrade, or need to grow the pointer table */ 600fa9e4066Sahrens objset_t *os = zap->zap_objset; 601fa9e4066Sahrens uint64_t object = zap->zap_object; 602fa9e4066Sahrens 603fa9e4066Sahrens zap_put_leaf(l); 604fa9e4066Sahrens zap_unlockdir(zap); 605*c5f9e431Sahrens err = zap_lockdir(os, object, tx, RW_WRITER, 606*c5f9e431Sahrens FALSE, FALSE, &zap); 60766328dd3Sahrens if (err) 60866328dd3Sahrens return (err); 609fa9e4066Sahrens ASSERT(!zap->zap_ismicro); 610fa9e4066Sahrens 61166328dd3Sahrens while (old_prefix_len == 61266328dd3Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_shift) { 61366328dd3Sahrens err = zap_grow_ptrtbl(zap, tx); 61466328dd3Sahrens if (err) 61566328dd3Sahrens return (err); 61666328dd3Sahrens } 61766328dd3Sahrens 61866328dd3Sahrens err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l); 61966328dd3Sahrens if (err) 62066328dd3Sahrens return (err); 62166328dd3Sahrens 62266328dd3Sahrens if (l->l_phys->l_hdr.lh_prefix_len != old_prefix_len) { 623fa9e4066Sahrens /* it split while our locks were down */ 624ea8dc4b6Seschrock *lp = l; 625ea8dc4b6Seschrock return (0); 626ea8dc4b6Seschrock } 627fa9e4066Sahrens } 628fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 62966328dd3Sahrens ASSERT3U(old_prefix_len, <, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift); 63066328dd3Sahrens ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==, 63166328dd3Sahrens l->l_phys->l_hdr.lh_prefix); 632fa9e4066Sahrens 63366328dd3Sahrens prefix_diff = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift - 63466328dd3Sahrens (old_prefix_len + 1); 63566328dd3Sahrens sibling = (ZAP_HASH_IDX(hash, old_prefix_len + 1) | 1) << prefix_diff; 636ea8dc4b6Seschrock 637ea8dc4b6Seschrock /* check for i/o errors before doing zap_leaf_split */ 638fa9e4066Sahrens for (i = 0; i < (1ULL<<prefix_diff); i++) { 639ea8dc4b6Seschrock uint64_t blk; 640ea8dc4b6Seschrock err = zap_idx_to_blk(zap, sibling+i, &blk); 641ea8dc4b6Seschrock if (err) 642ea8dc4b6Seschrock return (err); 643ea8dc4b6Seschrock ASSERT3U(blk, ==, l->l_blkid); 644ea8dc4b6Seschrock } 645ea8dc4b6Seschrock 64666328dd3Sahrens nl = zap_create_leaf(zap, tx); 647da6c28aaSamw zap_leaf_split(l, nl, spa_version(dmu_objset_spa(zap->zap_objset))); 648ea8dc4b6Seschrock 64966328dd3Sahrens /* set sibling pointers */ 650ea8dc4b6Seschrock for (i = 0; i < (1ULL<<prefix_diff); i++) { 651ea8dc4b6Seschrock err = zap_set_idx_to_blk(zap, sibling+i, nl->l_blkid, tx); 652ea8dc4b6Seschrock ASSERT3U(err, ==, 0); /* we checked for i/o errors above */ 653fa9e4066Sahrens } 654fa9e4066Sahrens 65566328dd3Sahrens if (hash & (1ULL << (64 - l->l_phys->l_hdr.lh_prefix_len))) { 656fa9e4066Sahrens /* we want the sibling */ 657fa9e4066Sahrens zap_put_leaf(l); 65866328dd3Sahrens *lp = nl; 659fa9e4066Sahrens } else { 660fa9e4066Sahrens zap_put_leaf(nl); 66166328dd3Sahrens *lp = l; 662fa9e4066Sahrens } 663fa9e4066Sahrens 664ea8dc4b6Seschrock return (0); 665fa9e4066Sahrens } 666fa9e4066Sahrens 667fa9e4066Sahrens static void 668f65e61c0Sahrens zap_put_leaf_maybe_grow_ptrtbl(zap_t *zap, zap_leaf_t *l, dmu_tx_t *tx) 669fa9e4066Sahrens { 67066328dd3Sahrens int shift = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift; 67166328dd3Sahrens int leaffull = (l->l_phys->l_hdr.lh_prefix_len == shift && 67266328dd3Sahrens l->l_phys->l_hdr.lh_nfree < ZAP_LEAF_LOW_WATER); 673fa9e4066Sahrens 67466328dd3Sahrens zap_put_leaf(l); 675fa9e4066Sahrens 67666328dd3Sahrens if (leaffull || zap->zap_f.zap_phys->zap_ptrtbl.zt_nextblk) { 67766328dd3Sahrens int err; 678fa9e4066Sahrens 67966328dd3Sahrens /* 68066328dd3Sahrens * We are in the middle of growing the pointer table, or 68166328dd3Sahrens * this leaf will soon make us grow it. 68266328dd3Sahrens */ 683fa9e4066Sahrens if (zap_tryupgradedir(zap, tx) == 0) { 684fa9e4066Sahrens objset_t *os = zap->zap_objset; 685fa9e4066Sahrens uint64_t zapobj = zap->zap_object; 686fa9e4066Sahrens 687fa9e4066Sahrens zap_unlockdir(zap); 688fa9e4066Sahrens err = zap_lockdir(os, zapobj, tx, 689*c5f9e431Sahrens RW_WRITER, FALSE, FALSE, &zap); 69066328dd3Sahrens if (err) 69166328dd3Sahrens return; 692fa9e4066Sahrens } 693fa9e4066Sahrens 69466328dd3Sahrens /* could have finished growing while our locks were down */ 69566328dd3Sahrens if (zap->zap_f.zap_phys->zap_ptrtbl.zt_shift == shift) 69666328dd3Sahrens (void) zap_grow_ptrtbl(zap, tx); 697fa9e4066Sahrens } 698fa9e4066Sahrens } 699fa9e4066Sahrens 700fa9e4066Sahrens 701fa9e4066Sahrens static int 70266328dd3Sahrens fzap_checksize(const char *name, uint64_t integer_size, uint64_t num_integers) 703fa9e4066Sahrens { 70466328dd3Sahrens if (name && strlen(name) > ZAP_MAXNAMELEN) 70566328dd3Sahrens return (E2BIG); 70666328dd3Sahrens 707fa9e4066Sahrens /* Only integer sizes supported by C */ 708fa9e4066Sahrens switch (integer_size) { 709fa9e4066Sahrens case 1: 710fa9e4066Sahrens case 2: 711fa9e4066Sahrens case 4: 712fa9e4066Sahrens case 8: 713fa9e4066Sahrens break; 714fa9e4066Sahrens default: 715fa9e4066Sahrens return (EINVAL); 716fa9e4066Sahrens } 717fa9e4066Sahrens 71866328dd3Sahrens if (integer_size * num_integers > ZAP_MAXVALUELEN) 71966328dd3Sahrens return (E2BIG); 720fa9e4066Sahrens 721fa9e4066Sahrens return (0); 722fa9e4066Sahrens } 723fa9e4066Sahrens 724fa9e4066Sahrens /* 725da6c28aaSamw * Routines for manipulating attributes. 726fa9e4066Sahrens */ 727fa9e4066Sahrens int 728da6c28aaSamw fzap_lookup(zap_name_t *zn, 729da6c28aaSamw uint64_t integer_size, uint64_t num_integers, void *buf, 730da6c28aaSamw char *realname, int rn_len, boolean_t *ncp) 731fa9e4066Sahrens { 732fa9e4066Sahrens zap_leaf_t *l; 733fa9e4066Sahrens int err; 734fa9e4066Sahrens zap_entry_handle_t zeh; 735fa9e4066Sahrens 736da6c28aaSamw err = fzap_checksize(zn->zn_name_orij, integer_size, num_integers); 737fa9e4066Sahrens if (err != 0) 738fa9e4066Sahrens return (err); 739fa9e4066Sahrens 740da6c28aaSamw err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l); 741ea8dc4b6Seschrock if (err != 0) 742ea8dc4b6Seschrock return (err); 743da6c28aaSamw err = zap_leaf_lookup(l, zn, &zeh); 744da6c28aaSamw if (err == 0) { 74566328dd3Sahrens err = zap_entry_read(&zeh, integer_size, num_integers, buf); 746da6c28aaSamw (void) zap_entry_read_name(&zeh, rn_len, realname); 747da6c28aaSamw if (ncp) { 748da6c28aaSamw *ncp = zap_entry_normalization_conflict(&zeh, 749da6c28aaSamw zn, NULL, zn->zn_zap); 750da6c28aaSamw } 751da6c28aaSamw } 75266328dd3Sahrens 753fa9e4066Sahrens zap_put_leaf(l); 754fa9e4066Sahrens return (err); 755fa9e4066Sahrens } 756fa9e4066Sahrens 757fa9e4066Sahrens int 758da6c28aaSamw fzap_add_cd(zap_name_t *zn, 759fa9e4066Sahrens uint64_t integer_size, uint64_t num_integers, 760ea8dc4b6Seschrock const void *val, uint32_t cd, dmu_tx_t *tx) 761fa9e4066Sahrens { 762fa9e4066Sahrens zap_leaf_t *l; 763fa9e4066Sahrens int err; 764fa9e4066Sahrens zap_entry_handle_t zeh; 765da6c28aaSamw zap_t *zap = zn->zn_zap; 766fa9e4066Sahrens 767fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 768fa9e4066Sahrens ASSERT(!zap->zap_ismicro); 769da6c28aaSamw ASSERT(fzap_checksize(zn->zn_name_orij, 770da6c28aaSamw integer_size, num_integers) == 0); 771fa9e4066Sahrens 772da6c28aaSamw err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l); 773ea8dc4b6Seschrock if (err != 0) 774ea8dc4b6Seschrock return (err); 775fa9e4066Sahrens retry: 776da6c28aaSamw err = zap_leaf_lookup(l, zn, &zeh); 777fa9e4066Sahrens if (err == 0) { 778fa9e4066Sahrens err = EEXIST; 779fa9e4066Sahrens goto out; 780fa9e4066Sahrens } 781ea8dc4b6Seschrock if (err != ENOENT) 782ea8dc4b6Seschrock goto out; 783fa9e4066Sahrens 784da6c28aaSamw err = zap_entry_create(l, zn->zn_name_orij, zn->zn_hash, cd, 785fa9e4066Sahrens integer_size, num_integers, val, &zeh); 786fa9e4066Sahrens 787fa9e4066Sahrens if (err == 0) { 788fa9e4066Sahrens zap_increment_num_entries(zap, 1, tx); 789fa9e4066Sahrens } else if (err == EAGAIN) { 790da6c28aaSamw err = zap_expand_leaf(zap, l, zn->zn_hash, tx, &l); 79166328dd3Sahrens if (err == 0) 79266328dd3Sahrens goto retry; 793fa9e4066Sahrens } 794fa9e4066Sahrens 795fa9e4066Sahrens out: 796ea8dc4b6Seschrock zap_put_leaf_maybe_grow_ptrtbl(zap, l, tx); 797fa9e4066Sahrens return (err); 798fa9e4066Sahrens } 799fa9e4066Sahrens 800fa9e4066Sahrens int 801da6c28aaSamw fzap_add(zap_name_t *zn, 802fa9e4066Sahrens uint64_t integer_size, uint64_t num_integers, 803fa9e4066Sahrens const void *val, dmu_tx_t *tx) 804fa9e4066Sahrens { 805da6c28aaSamw int err = fzap_checksize(zn->zn_name_orij, integer_size, num_integers); 806fa9e4066Sahrens if (err != 0) 807fa9e4066Sahrens return (err); 808fa9e4066Sahrens 809da6c28aaSamw return (fzap_add_cd(zn, integer_size, num_integers, 81066328dd3Sahrens val, ZAP_MAXCD, tx)); 811fa9e4066Sahrens } 812fa9e4066Sahrens 813fa9e4066Sahrens int 814da6c28aaSamw fzap_update(zap_name_t *zn, 815fa9e4066Sahrens int integer_size, uint64_t num_integers, const void *val, dmu_tx_t *tx) 816fa9e4066Sahrens { 817fa9e4066Sahrens zap_leaf_t *l; 818fa9e4066Sahrens int err, create; 819fa9e4066Sahrens zap_entry_handle_t zeh; 820da6c28aaSamw zap_t *zap = zn->zn_zap; 821fa9e4066Sahrens 822fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 823da6c28aaSamw err = fzap_checksize(zn->zn_name_orij, integer_size, num_integers); 824fa9e4066Sahrens if (err != 0) 825fa9e4066Sahrens return (err); 826fa9e4066Sahrens 827da6c28aaSamw err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l); 828ea8dc4b6Seschrock if (err != 0) 829ea8dc4b6Seschrock return (err); 830fa9e4066Sahrens retry: 831da6c28aaSamw err = zap_leaf_lookup(l, zn, &zeh); 832fa9e4066Sahrens create = (err == ENOENT); 833fa9e4066Sahrens ASSERT(err == 0 || err == ENOENT); 834fa9e4066Sahrens 835fa9e4066Sahrens if (create) { 836da6c28aaSamw err = zap_entry_create(l, zn->zn_name_orij, zn->zn_hash, 837da6c28aaSamw ZAP_MAXCD, integer_size, num_integers, val, &zeh); 838fa9e4066Sahrens if (err == 0) 839fa9e4066Sahrens zap_increment_num_entries(zap, 1, tx); 840fa9e4066Sahrens } else { 841fa9e4066Sahrens err = zap_entry_update(&zeh, integer_size, num_integers, val); 842fa9e4066Sahrens } 843fa9e4066Sahrens 844fa9e4066Sahrens if (err == EAGAIN) { 845da6c28aaSamw err = zap_expand_leaf(zap, l, zn->zn_hash, tx, &l); 84666328dd3Sahrens if (err == 0) 84766328dd3Sahrens goto retry; 848fa9e4066Sahrens } 849fa9e4066Sahrens 850fa9e4066Sahrens zap_put_leaf_maybe_grow_ptrtbl(zap, l, tx); 851fa9e4066Sahrens return (err); 852fa9e4066Sahrens } 853fa9e4066Sahrens 854fa9e4066Sahrens int 855da6c28aaSamw fzap_length(zap_name_t *zn, 856fa9e4066Sahrens uint64_t *integer_size, uint64_t *num_integers) 857fa9e4066Sahrens { 858fa9e4066Sahrens zap_leaf_t *l; 859fa9e4066Sahrens int err; 860fa9e4066Sahrens zap_entry_handle_t zeh; 861fa9e4066Sahrens 862da6c28aaSamw err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l); 863ea8dc4b6Seschrock if (err != 0) 864ea8dc4b6Seschrock return (err); 865da6c28aaSamw err = zap_leaf_lookup(l, zn, &zeh); 866fa9e4066Sahrens if (err != 0) 867fa9e4066Sahrens goto out; 868fa9e4066Sahrens 869fa9e4066Sahrens if (integer_size) 870fa9e4066Sahrens *integer_size = zeh.zeh_integer_size; 871fa9e4066Sahrens if (num_integers) 872fa9e4066Sahrens *num_integers = zeh.zeh_num_integers; 873fa9e4066Sahrens out: 874fa9e4066Sahrens zap_put_leaf(l); 875fa9e4066Sahrens return (err); 876fa9e4066Sahrens } 877fa9e4066Sahrens 878fa9e4066Sahrens int 879da6c28aaSamw fzap_remove(zap_name_t *zn, dmu_tx_t *tx) 880fa9e4066Sahrens { 881fa9e4066Sahrens zap_leaf_t *l; 882fa9e4066Sahrens int err; 883fa9e4066Sahrens zap_entry_handle_t zeh; 884fa9e4066Sahrens 885da6c28aaSamw err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, tx, RW_WRITER, &l); 886ea8dc4b6Seschrock if (err != 0) 887ea8dc4b6Seschrock return (err); 888da6c28aaSamw err = zap_leaf_lookup(l, zn, &zeh); 889fa9e4066Sahrens if (err == 0) { 890fa9e4066Sahrens zap_entry_remove(&zeh); 891da6c28aaSamw zap_increment_num_entries(zn->zn_zap, -1, tx); 892fa9e4066Sahrens } 893fa9e4066Sahrens zap_put_leaf(l); 894fa9e4066Sahrens return (err); 895fa9e4066Sahrens } 896fa9e4066Sahrens 897fa9e4066Sahrens int 898e7437265Sahrens zap_value_search(objset_t *os, uint64_t zapobj, uint64_t value, uint64_t mask, 899e7437265Sahrens char *name) 900fa9e4066Sahrens { 901fa9e4066Sahrens zap_cursor_t zc; 902fa9e4066Sahrens zap_attribute_t *za; 903fa9e4066Sahrens int err; 904fa9e4066Sahrens 905e7437265Sahrens if (mask == 0) 906e7437265Sahrens mask = -1ULL; 907e7437265Sahrens 908fa9e4066Sahrens za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP); 909fa9e4066Sahrens for (zap_cursor_init(&zc, os, zapobj); 910fa9e4066Sahrens (err = zap_cursor_retrieve(&zc, za)) == 0; 911fa9e4066Sahrens zap_cursor_advance(&zc)) { 912e7437265Sahrens if ((za->za_first_integer & mask) == (value & mask)) { 913fa9e4066Sahrens (void) strcpy(name, za->za_name); 914fa9e4066Sahrens break; 915fa9e4066Sahrens } 916fa9e4066Sahrens } 91787e5029aSahrens zap_cursor_fini(&zc); 918fa9e4066Sahrens kmem_free(za, sizeof (zap_attribute_t)); 919fa9e4066Sahrens return (err); 920fa9e4066Sahrens } 921fa9e4066Sahrens 922fa9e4066Sahrens 923fa9e4066Sahrens /* 924fa9e4066Sahrens * Routines for iterating over the attributes. 925fa9e4066Sahrens */ 926fa9e4066Sahrens 927fa9e4066Sahrens int 928fa9e4066Sahrens fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za) 929fa9e4066Sahrens { 930fa9e4066Sahrens int err = ENOENT; 931fa9e4066Sahrens zap_entry_handle_t zeh; 932fa9e4066Sahrens zap_leaf_t *l; 933fa9e4066Sahrens 934fa9e4066Sahrens /* retrieve the next entry at or after zc_hash/zc_cd */ 935fa9e4066Sahrens /* if no entry, return ENOENT */ 936fa9e4066Sahrens 93787e5029aSahrens if (zc->zc_leaf && 93866328dd3Sahrens (ZAP_HASH_IDX(zc->zc_hash, 93966328dd3Sahrens zc->zc_leaf->l_phys->l_hdr.lh_prefix_len) != 94066328dd3Sahrens zc->zc_leaf->l_phys->l_hdr.lh_prefix)) { 94187e5029aSahrens rw_enter(&zc->zc_leaf->l_rwlock, RW_READER); 94287e5029aSahrens zap_put_leaf(zc->zc_leaf); 94387e5029aSahrens zc->zc_leaf = NULL; 94487e5029aSahrens } 94587e5029aSahrens 946fa9e4066Sahrens again: 94787e5029aSahrens if (zc->zc_leaf == NULL) { 948ea8dc4b6Seschrock err = zap_deref_leaf(zap, zc->zc_hash, NULL, RW_READER, 949ea8dc4b6Seschrock &zc->zc_leaf); 950ea8dc4b6Seschrock if (err != 0) 951ea8dc4b6Seschrock return (err); 95287e5029aSahrens } else { 95387e5029aSahrens rw_enter(&zc->zc_leaf->l_rwlock, RW_READER); 95487e5029aSahrens } 95587e5029aSahrens l = zc->zc_leaf; 95687e5029aSahrens 957fa9e4066Sahrens err = zap_leaf_lookup_closest(l, zc->zc_hash, zc->zc_cd, &zeh); 958fa9e4066Sahrens 959fa9e4066Sahrens if (err == ENOENT) { 96066328dd3Sahrens uint64_t nocare = 96166328dd3Sahrens (1ULL << (64 - l->l_phys->l_hdr.lh_prefix_len)) - 1; 962fa9e4066Sahrens zc->zc_hash = (zc->zc_hash & ~nocare) + nocare + 1; 963fa9e4066Sahrens zc->zc_cd = 0; 96466328dd3Sahrens if (l->l_phys->l_hdr.lh_prefix_len == 0 || zc->zc_hash == 0) { 965fa9e4066Sahrens zc->zc_hash = -1ULL; 966fa9e4066Sahrens } else { 96787e5029aSahrens zap_put_leaf(zc->zc_leaf); 96887e5029aSahrens zc->zc_leaf = NULL; 969fa9e4066Sahrens goto again; 970fa9e4066Sahrens } 971fa9e4066Sahrens } 972fa9e4066Sahrens 973fa9e4066Sahrens if (err == 0) { 974fa9e4066Sahrens zc->zc_hash = zeh.zeh_hash; 975fa9e4066Sahrens zc->zc_cd = zeh.zeh_cd; 976fa9e4066Sahrens za->za_integer_length = zeh.zeh_integer_size; 977fa9e4066Sahrens za->za_num_integers = zeh.zeh_num_integers; 978fa9e4066Sahrens if (zeh.zeh_num_integers == 0) { 979fa9e4066Sahrens za->za_first_integer = 0; 980fa9e4066Sahrens } else { 981fa9e4066Sahrens err = zap_entry_read(&zeh, 8, 1, &za->za_first_integer); 982fa9e4066Sahrens ASSERT(err == 0 || err == EOVERFLOW); 983fa9e4066Sahrens } 984fa9e4066Sahrens err = zap_entry_read_name(&zeh, 985fa9e4066Sahrens sizeof (za->za_name), za->za_name); 986fa9e4066Sahrens ASSERT(err == 0); 987da6c28aaSamw 988da6c28aaSamw za->za_normalization_conflict = 989da6c28aaSamw zap_entry_normalization_conflict(&zeh, 990da6c28aaSamw NULL, za->za_name, zap); 991fa9e4066Sahrens } 99287e5029aSahrens rw_exit(&zc->zc_leaf->l_rwlock); 993fa9e4066Sahrens return (err); 994fa9e4066Sahrens } 995fa9e4066Sahrens 996fa9e4066Sahrens 997fa9e4066Sahrens static void 998fa9e4066Sahrens zap_stats_ptrtbl(zap_t *zap, uint64_t *tbl, int len, zap_stats_t *zs) 999fa9e4066Sahrens { 1000ea8dc4b6Seschrock int i, err; 1001fa9e4066Sahrens uint64_t lastblk = 0; 1002fa9e4066Sahrens 1003fa9e4066Sahrens /* 1004fa9e4066Sahrens * NB: if a leaf has more pointers than an entire ptrtbl block 1005fa9e4066Sahrens * can hold, then it'll be accounted for more than once, since 1006fa9e4066Sahrens * we won't have lastblk. 1007fa9e4066Sahrens */ 1008fa9e4066Sahrens for (i = 0; i < len; i++) { 1009fa9e4066Sahrens zap_leaf_t *l; 1010fa9e4066Sahrens 1011fa9e4066Sahrens if (tbl[i] == lastblk) 1012fa9e4066Sahrens continue; 1013fa9e4066Sahrens lastblk = tbl[i]; 1014fa9e4066Sahrens 1015ea8dc4b6Seschrock err = zap_get_leaf_byblk(zap, tbl[i], NULL, RW_READER, &l); 1016ea8dc4b6Seschrock if (err == 0) { 101766328dd3Sahrens zap_leaf_stats(zap, l, zs); 1018ea8dc4b6Seschrock zap_put_leaf(l); 1019ea8dc4b6Seschrock } 1020fa9e4066Sahrens } 1021fa9e4066Sahrens } 1022fa9e4066Sahrens 1023fa9e4066Sahrens void 1024fa9e4066Sahrens fzap_get_stats(zap_t *zap, zap_stats_t *zs) 1025fa9e4066Sahrens { 1026f65e61c0Sahrens int bs = FZAP_BLOCK_SHIFT(zap); 1027f65e61c0Sahrens zs->zs_blocksize = 1ULL << bs; 10288248818dSnd 10298248818dSnd /* 10308248818dSnd * Set zap_phys_t fields 10318248818dSnd */ 1032fa9e4066Sahrens zs->zs_num_leafs = zap->zap_f.zap_phys->zap_num_leafs; 1033fa9e4066Sahrens zs->zs_num_entries = zap->zap_f.zap_phys->zap_num_entries; 1034fa9e4066Sahrens zs->zs_num_blocks = zap->zap_f.zap_phys->zap_freeblk; 10358248818dSnd zs->zs_block_type = zap->zap_f.zap_phys->zap_block_type; 10368248818dSnd zs->zs_magic = zap->zap_f.zap_phys->zap_magic; 10378248818dSnd zs->zs_salt = zap->zap_f.zap_phys->zap_salt; 10388248818dSnd 10398248818dSnd /* 10408248818dSnd * Set zap_ptrtbl fields 10418248818dSnd */ 10428248818dSnd zs->zs_ptrtbl_len = 1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift; 10438248818dSnd zs->zs_ptrtbl_nextblk = zap->zap_f.zap_phys->zap_ptrtbl.zt_nextblk; 10448248818dSnd zs->zs_ptrtbl_blks_copied = 10458248818dSnd zap->zap_f.zap_phys->zap_ptrtbl.zt_blks_copied; 10468248818dSnd zs->zs_ptrtbl_zt_blk = zap->zap_f.zap_phys->zap_ptrtbl.zt_blk; 10478248818dSnd zs->zs_ptrtbl_zt_numblks = zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks; 10488248818dSnd zs->zs_ptrtbl_zt_shift = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift; 1049fa9e4066Sahrens 1050fa9e4066Sahrens if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) { 1051fa9e4066Sahrens /* the ptrtbl is entirely in the header block. */ 1052f65e61c0Sahrens zap_stats_ptrtbl(zap, &ZAP_EMBEDDED_PTRTBL_ENT(zap, 0), 1053f65e61c0Sahrens 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap), zs); 1054fa9e4066Sahrens } else { 1055fa9e4066Sahrens int b; 1056fa9e4066Sahrens 1057fa9e4066Sahrens dmu_prefetch(zap->zap_objset, zap->zap_object, 1058f65e61c0Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_blk << bs, 1059f65e61c0Sahrens zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks << bs); 1060fa9e4066Sahrens 1061fa9e4066Sahrens for (b = 0; b < zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks; 1062fa9e4066Sahrens b++) { 1063fa9e4066Sahrens dmu_buf_t *db; 1064ea8dc4b6Seschrock int err; 1065ea8dc4b6Seschrock 1066ea8dc4b6Seschrock err = dmu_buf_hold(zap->zap_objset, zap->zap_object, 1067ea8dc4b6Seschrock (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk + b) << bs, 1068ea8dc4b6Seschrock FTAG, &db); 1069ea8dc4b6Seschrock if (err == 0) { 1070ea8dc4b6Seschrock zap_stats_ptrtbl(zap, db->db_data, 1071ea8dc4b6Seschrock 1<<(bs-3), zs); 1072ea8dc4b6Seschrock dmu_buf_rele(db, FTAG); 1073ea8dc4b6Seschrock } 1074fa9e4066Sahrens } 1075fa9e4066Sahrens } 1076fa9e4066Sahrens } 1077