1 // SPDX-License-Identifier: CDDL-1.0 2 /* 3 * CDDL HEADER START 4 * 5 * The contents of this file are subject to the terms of the 6 * Common Development and Distribution License (the "License"). 7 * You may not use this file except in compliance with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or https://opensource.org/licenses/CDDL-1.0. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2012, 2018 by Delphix. All rights reserved. 25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 26 * Copyright 2023 Alexander Stetsenko <alex.stetsenko@gmail.com> 27 * Copyright (c) 2023, Klara Inc. 28 */ 29 30 /* 31 * This file contains the top half of the zfs directory structure 32 * implementation. The bottom half is in zap_leaf.c. 33 * 34 * The zdir is an extendable hash data structure. There is a table of 35 * pointers to buckets (zap_t->zd_data->zd_leafs). The buckets are 36 * each a constant size and hold a variable number of directory entries. 37 * The buckets (aka "leaf nodes") are implemented in zap_leaf.c. 38 * 39 * The pointer table holds a power of 2 number of pointers. 40 * (1<<zap_t->zd_data->zd_phys->zd_prefix_len). The bucket pointed to 41 * by the pointer at index i in the table holds entries whose hash value 42 * has a zd_prefix_len - bit prefix 43 */ 44 45 #include <sys/spa.h> 46 #include <sys/dmu.h> 47 #include <sys/dnode.h> 48 #include <sys/zfs_context.h> 49 #include <sys/zfs_znode.h> 50 #include <sys/fs/zfs.h> 51 #include <sys/zap.h> 52 #include <sys/zap_impl.h> 53 #include <sys/zap_leaf.h> 54 55 /* 56 * If zap_iterate_prefetch is set, we will prefetch the entire ZAP object 57 * (all leaf blocks) when we start iterating over it. 58 * 59 * For zap_cursor_init(), the callers all intend to iterate through all the 60 * entries. There are a few cases where an error (typically i/o error) could 61 * cause it to bail out early. 62 * 63 * For zap_cursor_init_serialized(), there are callers that do the iteration 64 * outside of ZFS. Typically they would iterate over everything, but we 65 * don't have control of that. E.g. zfs_ioc_snapshot_list_next(), 66 * zcp_snapshots_iter(), and other iterators over things in the MOS - these 67 * are called by /sbin/zfs and channel programs. The other example is 68 * zfs_readdir() which iterates over directory entries for the getdents() 69 * syscall. /sbin/ls iterates to the end (unless it receives a signal), but 70 * userland doesn't have to. 71 * 72 * Given that the ZAP entries aren't returned in a specific order, the only 73 * legitimate use cases for partial iteration would be: 74 * 75 * 1. Pagination: e.g. you only want to display 100 entries at a time, so you 76 * get the first 100 and then wait for the user to hit "next page", which 77 * they may never do). 78 * 79 * 2. You want to know if there are more than X entries, without relying on 80 * the zfs-specific implementation of the directory's st_size (which is 81 * the number of entries). 82 */ 83 static int zap_iterate_prefetch = B_TRUE; 84 85 /* 86 * Enable ZAP shrinking. When enabled, empty sibling leaf blocks will be 87 * collapsed into a single block. 88 */ 89 int zap_shrink_enabled = B_TRUE; 90 91 int fzap_default_block_shift = 14; /* 16k blocksize */ 92 93 static uint64_t zap_allocate_blocks(zap_t *zap, int nblocks); 94 static int zap_shrink(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx); 95 96 void 97 fzap_byteswap(void *vbuf, size_t size) 98 { 99 uint64_t block_type = *(uint64_t *)vbuf; 100 101 if (block_type == ZBT_LEAF || block_type == BSWAP_64(ZBT_LEAF)) 102 zap_leaf_byteswap(vbuf, size); 103 else { 104 /* it's a ptrtbl block */ 105 byteswap_uint64_array(vbuf, size); 106 } 107 } 108 109 void 110 fzap_upgrade(zap_t *zap, dmu_tx_t *tx, zap_flags_t flags) 111 { 112 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 113 zap->zap_ismicro = FALSE; 114 115 zap->zap_dbu.dbu_evict_func_sync = zap_evict_sync; 116 zap->zap_dbu.dbu_evict_func_async = NULL; 117 118 mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, MUTEX_DEFAULT, 0); 119 zap->zap_f.zap_block_shift = highbit64(zap->zap_dbuf->db_size) - 1; 120 121 zap_phys_t *zp = zap_f_phys(zap); 122 /* 123 * explicitly zero it since it might be coming from an 124 * initialized microzap 125 */ 126 memset(zap->zap_dbuf->db_data, 0, zap->zap_dbuf->db_size); 127 zp->zap_block_type = ZBT_HEADER; 128 zp->zap_magic = ZAP_MAGIC; 129 130 zp->zap_ptrtbl.zt_shift = ZAP_EMBEDDED_PTRTBL_SHIFT(zap); 131 132 zp->zap_freeblk = 2; /* block 1 will be the first leaf */ 133 zp->zap_num_leafs = 1; 134 zp->zap_num_entries = 0; 135 zp->zap_salt = zap->zap_salt; 136 zp->zap_normflags = zap->zap_normflags; 137 zp->zap_flags = flags; 138 139 /* block 1 will be the first leaf */ 140 for (int i = 0; i < (1<<zp->zap_ptrtbl.zt_shift); i++) 141 ZAP_EMBEDDED_PTRTBL_ENT(zap, i) = 1; 142 143 /* 144 * set up block 1 - the first leaf 145 */ 146 dmu_buf_t *db; 147 VERIFY0(dmu_buf_hold_by_dnode(zap->zap_dnode, 148 1<<FZAP_BLOCK_SHIFT(zap), FTAG, &db, DMU_READ_NO_PREFETCH)); 149 dmu_buf_will_dirty(db, tx); 150 151 zap_leaf_t *l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP); 152 l->l_dbuf = db; 153 154 zap_leaf_init(l, zp->zap_normflags != 0); 155 156 kmem_free(l, sizeof (zap_leaf_t)); 157 dmu_buf_rele(db, FTAG); 158 } 159 160 static int 161 zap_tryupgradedir(zap_t *zap, dmu_tx_t *tx) 162 { 163 if (RW_WRITE_HELD(&zap->zap_rwlock)) 164 return (1); 165 if (rw_tryupgrade(&zap->zap_rwlock)) { 166 dmu_buf_will_dirty(zap->zap_dbuf, tx); 167 return (1); 168 } 169 return (0); 170 } 171 172 /* 173 * Generic routines for dealing with the pointer & cookie tables. 174 */ 175 176 static int 177 zap_table_grow(zap_t *zap, zap_table_phys_t *tbl, 178 void (*transfer_func)(const uint64_t *src, uint64_t *dst, int n), 179 dmu_tx_t *tx) 180 { 181 uint64_t newblk; 182 int bs = FZAP_BLOCK_SHIFT(zap); 183 int hepb = 1<<(bs-4); 184 /* hepb = half the number of entries in a block */ 185 186 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 187 ASSERT(tbl->zt_blk != 0); 188 ASSERT(tbl->zt_numblks > 0); 189 190 if (tbl->zt_nextblk != 0) { 191 newblk = tbl->zt_nextblk; 192 } else { 193 newblk = zap_allocate_blocks(zap, tbl->zt_numblks * 2); 194 tbl->zt_nextblk = newblk; 195 ASSERT0(tbl->zt_blks_copied); 196 dmu_prefetch_by_dnode(zap->zap_dnode, 0, 197 tbl->zt_blk << bs, tbl->zt_numblks << bs, 198 ZIO_PRIORITY_SYNC_READ); 199 } 200 201 /* 202 * Copy the ptrtbl from the old to new location. 203 */ 204 205 uint64_t b = tbl->zt_blks_copied; 206 dmu_buf_t *db_old; 207 int err = dmu_buf_hold_by_dnode(zap->zap_dnode, 208 (tbl->zt_blk + b) << bs, FTAG, &db_old, DMU_READ_NO_PREFETCH); 209 if (err != 0) 210 return (err); 211 212 /* first half of entries in old[b] go to new[2*b+0] */ 213 dmu_buf_t *db_new; 214 VERIFY0(dmu_buf_hold_by_dnode(zap->zap_dnode, 215 (newblk + 2*b+0) << bs, FTAG, &db_new, DMU_READ_NO_PREFETCH)); 216 dmu_buf_will_dirty(db_new, tx); 217 transfer_func(db_old->db_data, db_new->db_data, hepb); 218 dmu_buf_rele(db_new, FTAG); 219 220 /* second half of entries in old[b] go to new[2*b+1] */ 221 VERIFY0(dmu_buf_hold_by_dnode(zap->zap_dnode, 222 (newblk + 2*b+1) << bs, FTAG, &db_new, DMU_READ_NO_PREFETCH)); 223 dmu_buf_will_dirty(db_new, tx); 224 transfer_func((uint64_t *)db_old->db_data + hepb, 225 db_new->db_data, hepb); 226 dmu_buf_rele(db_new, FTAG); 227 228 dmu_buf_rele(db_old, FTAG); 229 230 tbl->zt_blks_copied++; 231 232 dprintf("copied block %llu of %llu\n", 233 (u_longlong_t)tbl->zt_blks_copied, 234 (u_longlong_t)tbl->zt_numblks); 235 236 if (tbl->zt_blks_copied == tbl->zt_numblks) { 237 (void) dmu_free_range(zap->zap_objset, zap->zap_object, 238 tbl->zt_blk << bs, tbl->zt_numblks << bs, tx); 239 240 tbl->zt_blk = newblk; 241 tbl->zt_numblks *= 2; 242 tbl->zt_shift++; 243 tbl->zt_nextblk = 0; 244 tbl->zt_blks_copied = 0; 245 246 dprintf("finished; numblocks now %llu (%uk entries)\n", 247 (u_longlong_t)tbl->zt_numblks, 1<<(tbl->zt_shift-10)); 248 } 249 250 return (0); 251 } 252 253 static int 254 zap_table_store(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t val, 255 dmu_tx_t *tx) 256 { 257 int bs = FZAP_BLOCK_SHIFT(zap); 258 259 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 260 ASSERT(tbl->zt_blk != 0); 261 262 dprintf("storing %llx at index %llx\n", (u_longlong_t)val, 263 (u_longlong_t)idx); 264 265 uint64_t blk = idx >> (bs-3); 266 uint64_t off = idx & ((1<<(bs-3))-1); 267 268 dmu_buf_t *db; 269 int err = dmu_buf_hold_by_dnode(zap->zap_dnode, 270 (tbl->zt_blk + blk) << bs, FTAG, &db, DMU_READ_NO_PREFETCH); 271 if (err != 0) 272 return (err); 273 dmu_buf_will_dirty(db, tx); 274 275 if (tbl->zt_nextblk != 0) { 276 uint64_t idx2 = idx * 2; 277 uint64_t blk2 = idx2 >> (bs-3); 278 uint64_t off2 = idx2 & ((1<<(bs-3))-1); 279 dmu_buf_t *db2; 280 281 err = dmu_buf_hold_by_dnode(zap->zap_dnode, 282 (tbl->zt_nextblk + blk2) << bs, FTAG, &db2, 283 DMU_READ_NO_PREFETCH); 284 if (err != 0) { 285 dmu_buf_rele(db, FTAG); 286 return (err); 287 } 288 dmu_buf_will_dirty(db2, tx); 289 ((uint64_t *)db2->db_data)[off2] = val; 290 ((uint64_t *)db2->db_data)[off2+1] = val; 291 dmu_buf_rele(db2, FTAG); 292 } 293 294 ((uint64_t *)db->db_data)[off] = val; 295 dmu_buf_rele(db, FTAG); 296 297 return (0); 298 } 299 300 static int 301 zap_table_load(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t *valp) 302 { 303 int bs = FZAP_BLOCK_SHIFT(zap); 304 305 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 306 307 uint64_t blk = idx >> (bs-3); 308 uint64_t off = idx & ((1<<(bs-3))-1); 309 310 dmu_buf_t *db; 311 int err = dmu_buf_hold_by_dnode(zap->zap_dnode, 312 (tbl->zt_blk + blk) << bs, FTAG, &db, DMU_READ_NO_PREFETCH); 313 if (err != 0) 314 return (err); 315 *valp = ((uint64_t *)db->db_data)[off]; 316 dmu_buf_rele(db, FTAG); 317 318 if (tbl->zt_nextblk != 0) { 319 /* 320 * read the nextblk for the sake of i/o error checking, 321 * so that zap_table_load() will catch errors for 322 * zap_table_store. 323 */ 324 blk = (idx*2) >> (bs-3); 325 326 err = dmu_buf_hold_by_dnode(zap->zap_dnode, 327 (tbl->zt_nextblk + blk) << bs, FTAG, &db, 328 DMU_READ_NO_PREFETCH); 329 if (err == 0) 330 dmu_buf_rele(db, FTAG); 331 } 332 return (err); 333 } 334 335 /* 336 * Routines for growing the ptrtbl. 337 */ 338 339 static void 340 zap_ptrtbl_transfer(const uint64_t *src, uint64_t *dst, int n) 341 { 342 for (int i = 0; i < n; i++) { 343 uint64_t lb = src[i]; 344 dst[2 * i + 0] = lb; 345 dst[2 * i + 1] = lb; 346 } 347 } 348 349 static int 350 zap_grow_ptrtbl(zap_t *zap, dmu_tx_t *tx) 351 { 352 /* 353 * The pointer table should never use more hash bits than we 354 * have (otherwise we'd be using useless zero bits to index it). 355 * If we are within 2 bits of running out, stop growing, since 356 * this is already an aberrant condition. 357 */ 358 if (zap_f_phys(zap)->zap_ptrtbl.zt_shift >= zap_hashbits(zap) - 2) 359 return (SET_ERROR(ENOSPC)); 360 361 if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) { 362 /* 363 * We are outgrowing the "embedded" ptrtbl (the one 364 * stored in the header block). Give it its own entire 365 * block, which will double the size of the ptrtbl. 366 */ 367 ASSERT3U(zap_f_phys(zap)->zap_ptrtbl.zt_shift, ==, 368 ZAP_EMBEDDED_PTRTBL_SHIFT(zap)); 369 ASSERT0(zap_f_phys(zap)->zap_ptrtbl.zt_blk); 370 371 uint64_t newblk = zap_allocate_blocks(zap, 1); 372 dmu_buf_t *db_new; 373 int err = dmu_buf_hold_by_dnode(zap->zap_dnode, 374 newblk << FZAP_BLOCK_SHIFT(zap), FTAG, &db_new, 375 DMU_READ_NO_PREFETCH); 376 if (err != 0) 377 return (err); 378 dmu_buf_will_dirty(db_new, tx); 379 zap_ptrtbl_transfer(&ZAP_EMBEDDED_PTRTBL_ENT(zap, 0), 380 db_new->db_data, 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap)); 381 dmu_buf_rele(db_new, FTAG); 382 383 zap_f_phys(zap)->zap_ptrtbl.zt_blk = newblk; 384 zap_f_phys(zap)->zap_ptrtbl.zt_numblks = 1; 385 zap_f_phys(zap)->zap_ptrtbl.zt_shift++; 386 387 ASSERT3U(1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift, ==, 388 zap_f_phys(zap)->zap_ptrtbl.zt_numblks << 389 (FZAP_BLOCK_SHIFT(zap)-3)); 390 391 return (0); 392 } else { 393 return (zap_table_grow(zap, &zap_f_phys(zap)->zap_ptrtbl, 394 zap_ptrtbl_transfer, tx)); 395 } 396 } 397 398 static void 399 zap_increment_num_entries(zap_t *zap, int delta, dmu_tx_t *tx) 400 { 401 dmu_buf_will_dirty(zap->zap_dbuf, tx); 402 mutex_enter(&zap->zap_f.zap_num_entries_mtx); 403 ASSERT(delta > 0 || zap_f_phys(zap)->zap_num_entries >= -delta); 404 zap_f_phys(zap)->zap_num_entries += delta; 405 mutex_exit(&zap->zap_f.zap_num_entries_mtx); 406 } 407 408 static uint64_t 409 zap_allocate_blocks(zap_t *zap, int nblocks) 410 { 411 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 412 uint64_t newblk = zap_f_phys(zap)->zap_freeblk; 413 zap_f_phys(zap)->zap_freeblk += nblocks; 414 return (newblk); 415 } 416 417 static void 418 zap_leaf_evict_sync(void *dbu) 419 { 420 zap_leaf_t *l = dbu; 421 422 rw_destroy(&l->l_rwlock); 423 kmem_free(l, sizeof (zap_leaf_t)); 424 } 425 426 static zap_leaf_t * 427 zap_create_leaf(zap_t *zap, dmu_tx_t *tx) 428 { 429 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 430 431 uint64_t blkid = zap_allocate_blocks(zap, 1); 432 dmu_buf_t *db = NULL; 433 434 VERIFY0(dmu_buf_hold_by_dnode(zap->zap_dnode, 435 blkid << FZAP_BLOCK_SHIFT(zap), NULL, &db, 436 DMU_READ_NO_PREFETCH)); 437 438 /* 439 * Create the leaf structure and stash it on the dbuf. If zap was 440 * recent shrunk or truncated, the dbuf might have been sitting in the 441 * cache waiting to be evicted, and so still have the old leaf attached 442 * to it. If so, just reuse it. 443 */ 444 zap_leaf_t *l = dmu_buf_get_user(db); 445 if (l == NULL) { 446 l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP); 447 l->l_blkid = blkid; 448 l->l_dbuf = db; 449 rw_init(&l->l_rwlock, NULL, RW_NOLOCKDEP, NULL); 450 dmu_buf_init_user(&l->l_dbu, zap_leaf_evict_sync, NULL, 451 &l->l_dbuf); 452 dmu_buf_set_user(l->l_dbuf, &l->l_dbu); 453 } else { 454 ASSERT3U(l->l_blkid, ==, blkid); 455 ASSERT3P(l->l_dbuf, ==, db); 456 } 457 458 rw_enter(&l->l_rwlock, RW_WRITER); 459 dmu_buf_will_dirty(l->l_dbuf, tx); 460 461 zap_leaf_init(l, zap->zap_normflags != 0); 462 463 zap_f_phys(zap)->zap_num_leafs++; 464 465 return (l); 466 } 467 468 int 469 fzap_count(zap_t *zap, uint64_t *count) 470 { 471 ASSERT(!zap->zap_ismicro); 472 mutex_enter(&zap->zap_f.zap_num_entries_mtx); /* unnecessary */ 473 *count = zap_f_phys(zap)->zap_num_entries; 474 mutex_exit(&zap->zap_f.zap_num_entries_mtx); 475 return (0); 476 } 477 478 /* 479 * Routines for obtaining zap_leaf_t's 480 */ 481 482 void 483 zap_put_leaf(zap_leaf_t *l) 484 { 485 rw_exit(&l->l_rwlock); 486 dmu_buf_rele(l->l_dbuf, NULL); 487 } 488 489 static zap_leaf_t * 490 zap_open_leaf(uint64_t blkid, dmu_buf_t *db) 491 { 492 ASSERT(blkid != 0); 493 494 zap_leaf_t *l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP); 495 rw_init(&l->l_rwlock, NULL, RW_DEFAULT, NULL); 496 rw_enter(&l->l_rwlock, RW_WRITER); 497 l->l_blkid = blkid; 498 l->l_bs = highbit64(db->db_size) - 1; 499 l->l_dbuf = db; 500 501 dmu_buf_init_user(&l->l_dbu, zap_leaf_evict_sync, NULL, &l->l_dbuf); 502 zap_leaf_t *winner = dmu_buf_set_user(db, &l->l_dbu); 503 504 rw_exit(&l->l_rwlock); 505 if (winner != NULL) { 506 /* someone else set it first */ 507 zap_leaf_evict_sync(&l->l_dbu); 508 l = winner; 509 } 510 511 /* 512 * lhr_pad was previously used for the next leaf in the leaf 513 * chain. There should be no chained leafs (as we have removed 514 * support for them). 515 */ 516 ASSERT0(zap_leaf_phys(l)->l_hdr.lh_pad1); 517 518 /* 519 * There should be more hash entries than there can be 520 * chunks to put in the hash table 521 */ 522 ASSERT3U(ZAP_LEAF_HASH_NUMENTRIES(l), >, ZAP_LEAF_NUMCHUNKS(l) / 3); 523 524 /* The chunks should begin at the end of the hash table */ 525 ASSERT3P(&ZAP_LEAF_CHUNK(l, 0), ==, (zap_leaf_chunk_t *) 526 &zap_leaf_phys(l)->l_hash[ZAP_LEAF_HASH_NUMENTRIES(l)]); 527 528 /* The chunks should end at the end of the block */ 529 ASSERT3U((uintptr_t)&ZAP_LEAF_CHUNK(l, ZAP_LEAF_NUMCHUNKS(l)) - 530 (uintptr_t)zap_leaf_phys(l), ==, l->l_dbuf->db_size); 531 532 return (l); 533 } 534 535 static int 536 zap_get_leaf_byblk(zap_t *zap, uint64_t blkid, dmu_tx_t *tx, krw_t lt, 537 zap_leaf_t **lp) 538 { 539 dmu_buf_t *db; 540 541 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 542 543 /* 544 * If system crashed just after dmu_free_long_range in zfs_rmnode, we 545 * would be left with an empty xattr dir in delete queue. blkid=0 546 * would be passed in when doing zfs_purgedir. If that's the case we 547 * should just return immediately. The underlying objects should 548 * already be freed, so this should be perfectly fine. 549 */ 550 if (blkid == 0) 551 return (SET_ERROR(ENOENT)); 552 553 int bs = FZAP_BLOCK_SHIFT(zap); 554 int err = dmu_buf_hold_by_dnode(zap->zap_dnode, 555 blkid << bs, NULL, &db, DMU_READ_NO_PREFETCH); 556 if (err != 0) 557 return (err); 558 559 ASSERT3U(db->db_object, ==, zap->zap_object); 560 ASSERT3U(db->db_offset, ==, blkid << bs); 561 ASSERT3U(db->db_size, ==, 1 << bs); 562 ASSERT(blkid != 0); 563 564 zap_leaf_t *l = dmu_buf_get_user(db); 565 566 if (l == NULL) 567 l = zap_open_leaf(blkid, db); 568 569 rw_enter(&l->l_rwlock, lt); 570 /* 571 * Must lock before dirtying, otherwise zap_leaf_phys(l) could change, 572 * causing ASSERT below to fail. 573 */ 574 if (lt == RW_WRITER) 575 dmu_buf_will_dirty(db, tx); 576 ASSERT3U(l->l_blkid, ==, blkid); 577 ASSERT3P(l->l_dbuf, ==, db); 578 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_block_type, ==, ZBT_LEAF); 579 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC); 580 581 *lp = l; 582 return (0); 583 } 584 585 static int 586 zap_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t *valp) 587 { 588 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 589 590 if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) { 591 ASSERT3U(idx, <, 592 (1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift)); 593 *valp = ZAP_EMBEDDED_PTRTBL_ENT(zap, idx); 594 return (0); 595 } else { 596 return (zap_table_load(zap, &zap_f_phys(zap)->zap_ptrtbl, 597 idx, valp)); 598 } 599 } 600 601 static int 602 zap_set_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t blk, dmu_tx_t *tx) 603 { 604 ASSERT(tx != NULL); 605 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 606 607 if (zap_f_phys(zap)->zap_ptrtbl.zt_blk == 0) { 608 ZAP_EMBEDDED_PTRTBL_ENT(zap, idx) = blk; 609 return (0); 610 } else { 611 return (zap_table_store(zap, &zap_f_phys(zap)->zap_ptrtbl, 612 idx, blk, tx)); 613 } 614 } 615 616 static int 617 zap_set_idx_range_to_blk(zap_t *zap, uint64_t idx, uint64_t nptrs, uint64_t blk, 618 dmu_tx_t *tx) 619 { 620 int bs = FZAP_BLOCK_SHIFT(zap); 621 int epb = bs >> 3; /* entries per block */ 622 int err = 0; 623 624 ASSERT(tx != NULL); 625 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 626 627 /* 628 * Check for i/o errors 629 */ 630 for (int i = 0; i < nptrs; i += epb) { 631 uint64_t blk; 632 err = zap_idx_to_blk(zap, idx + i, &blk); 633 if (err != 0) { 634 return (err); 635 } 636 } 637 638 for (int i = 0; i < nptrs; i++) { 639 err = zap_set_idx_to_blk(zap, idx + i, blk, tx); 640 ASSERT0(err); /* we checked for i/o errors above */ 641 if (err != 0) 642 break; 643 } 644 645 return (err); 646 } 647 648 #define ZAP_PREFIX_HASH(pref, pref_len) ((pref) << (64 - (pref_len))) 649 650 /* 651 * Each leaf has single range of entries (block pointers) in the ZAP ptrtbl. 652 * If two leaves are siblings, their ranges are adjecent and contain the same 653 * number of entries. In order to find out if a leaf has a sibling, we need to 654 * check the range corresponding to the sibling leaf. There is no need to check 655 * all entries in the range, we only need to check the frist and the last one. 656 */ 657 static uint64_t 658 check_sibling_ptrtbl_range(zap_t *zap, uint64_t prefix, uint64_t prefix_len) 659 { 660 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 661 662 uint64_t h = ZAP_PREFIX_HASH(prefix, prefix_len); 663 uint64_t idx = ZAP_HASH_IDX(h, zap_f_phys(zap)->zap_ptrtbl.zt_shift); 664 uint64_t pref_diff = zap_f_phys(zap)->zap_ptrtbl.zt_shift - prefix_len; 665 uint64_t nptrs = (1 << pref_diff); 666 uint64_t first; 667 uint64_t last; 668 669 ASSERT3U(idx+nptrs, <=, (1UL << zap_f_phys(zap)->zap_ptrtbl.zt_shift)); 670 671 if (zap_idx_to_blk(zap, idx, &first) != 0) 672 return (0); 673 674 if (zap_idx_to_blk(zap, idx + nptrs - 1, &last) != 0) 675 return (0); 676 677 if (first != last) 678 return (0); 679 return (first); 680 } 681 682 static int 683 zap_deref_leaf(zap_t *zap, uint64_t h, dmu_tx_t *tx, krw_t lt, zap_leaf_t **lp) 684 { 685 uint64_t blk; 686 687 ASSERT(zap->zap_dbuf == NULL || 688 zap_f_phys(zap) == zap->zap_dbuf->db_data); 689 690 /* Reality check for corrupt zap objects (leaf or header). */ 691 if ((zap_f_phys(zap)->zap_block_type != ZBT_LEAF && 692 zap_f_phys(zap)->zap_block_type != ZBT_HEADER) || 693 zap_f_phys(zap)->zap_magic != ZAP_MAGIC) { 694 return (SET_ERROR(EIO)); 695 } 696 697 uint64_t idx = ZAP_HASH_IDX(h, zap_f_phys(zap)->zap_ptrtbl.zt_shift); 698 int err = zap_idx_to_blk(zap, idx, &blk); 699 if (err != 0) 700 return (err); 701 err = zap_get_leaf_byblk(zap, blk, tx, lt, lp); 702 703 ASSERT(err || 704 ZAP_HASH_IDX(h, zap_leaf_phys(*lp)->l_hdr.lh_prefix_len) == 705 zap_leaf_phys(*lp)->l_hdr.lh_prefix); 706 return (err); 707 } 708 709 static int 710 zap_expand_leaf(zap_name_t *zn, zap_leaf_t *l, 711 const void *tag, dmu_tx_t *tx, zap_leaf_t **lp) 712 { 713 zap_t *zap = zn->zn_zap; 714 uint64_t hash = zn->zn_hash; 715 int err; 716 int old_prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len; 717 718 ASSERT3U(old_prefix_len, <=, zap_f_phys(zap)->zap_ptrtbl.zt_shift); 719 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 720 721 ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==, 722 zap_leaf_phys(l)->l_hdr.lh_prefix); 723 724 if (zap_tryupgradedir(zap, tx) == 0 || 725 old_prefix_len == zap_f_phys(zap)->zap_ptrtbl.zt_shift) { 726 /* We failed to upgrade, or need to grow the pointer table */ 727 objset_t *os = zap->zap_objset; 728 uint64_t object = zap->zap_object; 729 730 zap_put_leaf(l); 731 *lp = l = NULL; 732 zap_unlockdir(zap, tag); 733 err = zap_lockdir(os, object, tx, RW_WRITER, 734 FALSE, FALSE, tag, &zn->zn_zap); 735 zap = zn->zn_zap; 736 if (err != 0) 737 return (err); 738 ASSERT(!zap->zap_ismicro); 739 740 while (old_prefix_len == 741 zap_f_phys(zap)->zap_ptrtbl.zt_shift) { 742 err = zap_grow_ptrtbl(zap, tx); 743 if (err != 0) 744 return (err); 745 } 746 747 err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l); 748 if (err != 0) 749 return (err); 750 751 if (zap_leaf_phys(l)->l_hdr.lh_prefix_len != old_prefix_len) { 752 /* it split while our locks were down */ 753 *lp = l; 754 return (0); 755 } 756 } 757 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 758 ASSERT3U(old_prefix_len, <, zap_f_phys(zap)->zap_ptrtbl.zt_shift); 759 ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==, 760 zap_leaf_phys(l)->l_hdr.lh_prefix); 761 762 int prefix_diff = zap_f_phys(zap)->zap_ptrtbl.zt_shift - 763 (old_prefix_len + 1); 764 uint64_t sibling = 765 (ZAP_HASH_IDX(hash, old_prefix_len + 1) | 1) << prefix_diff; 766 767 /* check for i/o errors before doing zap_leaf_split */ 768 for (int i = 0; i < (1ULL << prefix_diff); i++) { 769 uint64_t blk; 770 err = zap_idx_to_blk(zap, sibling + i, &blk); 771 if (err != 0) 772 return (err); 773 ASSERT3U(blk, ==, l->l_blkid); 774 } 775 776 zap_leaf_t *nl = zap_create_leaf(zap, tx); 777 zap_leaf_split(l, nl, zap->zap_normflags != 0); 778 779 /* set sibling pointers */ 780 for (int i = 0; i < (1ULL << prefix_diff); i++) { 781 err = zap_set_idx_to_blk(zap, sibling + i, nl->l_blkid, tx); 782 ASSERT0(err); /* we checked for i/o errors above */ 783 } 784 785 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_prefix_len, >, 0); 786 787 if (hash & (1ULL << (64 - zap_leaf_phys(l)->l_hdr.lh_prefix_len))) { 788 /* we want the sibling */ 789 zap_put_leaf(l); 790 *lp = nl; 791 } else { 792 zap_put_leaf(nl); 793 *lp = l; 794 } 795 796 return (0); 797 } 798 799 static void 800 zap_put_leaf_maybe_grow_ptrtbl(zap_name_t *zn, zap_leaf_t *l, 801 const void *tag, dmu_tx_t *tx) 802 { 803 zap_t *zap = zn->zn_zap; 804 int shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 805 int leaffull = (zap_leaf_phys(l)->l_hdr.lh_prefix_len == shift && 806 zap_leaf_phys(l)->l_hdr.lh_nfree < ZAP_LEAF_LOW_WATER); 807 808 zap_put_leaf(l); 809 810 if (leaffull || zap_f_phys(zap)->zap_ptrtbl.zt_nextblk) { 811 /* 812 * We are in the middle of growing the pointer table, or 813 * this leaf will soon make us grow it. 814 */ 815 if (zap_tryupgradedir(zap, tx) == 0) { 816 objset_t *os = zap->zap_objset; 817 uint64_t zapobj = zap->zap_object; 818 819 zap_unlockdir(zap, tag); 820 int err = zap_lockdir(os, zapobj, tx, 821 RW_WRITER, FALSE, FALSE, tag, &zn->zn_zap); 822 zap = zn->zn_zap; 823 if (err != 0) 824 return; 825 } 826 827 /* could have finished growing while our locks were down */ 828 if (zap_f_phys(zap)->zap_ptrtbl.zt_shift == shift) 829 (void) zap_grow_ptrtbl(zap, tx); 830 } 831 } 832 833 static int 834 fzap_checkname(zap_name_t *zn) 835 { 836 uint32_t maxnamelen = zn->zn_normbuf_len; 837 uint64_t len = (uint64_t)zn->zn_key_orig_numints * zn->zn_key_intlen; 838 /* Only allow directory zap to have longname */ 839 if (len > maxnamelen || 840 (len > ZAP_MAXNAMELEN && 841 zn->zn_zap->zap_dnode->dn_type != DMU_OT_DIRECTORY_CONTENTS)) 842 return (SET_ERROR(ENAMETOOLONG)); 843 return (0); 844 } 845 846 static int 847 fzap_checksize(uint64_t integer_size, uint64_t num_integers) 848 { 849 /* Only integer sizes supported by C */ 850 switch (integer_size) { 851 case 1: 852 case 2: 853 case 4: 854 case 8: 855 break; 856 default: 857 return (SET_ERROR(EINVAL)); 858 } 859 860 if (integer_size * num_integers > ZAP_MAXVALUELEN) 861 return (SET_ERROR(E2BIG)); 862 863 return (0); 864 } 865 866 static int 867 fzap_check(zap_name_t *zn, uint64_t integer_size, uint64_t num_integers) 868 { 869 int err = fzap_checkname(zn); 870 if (err != 0) 871 return (err); 872 return (fzap_checksize(integer_size, num_integers)); 873 } 874 875 /* 876 * Routines for manipulating attributes. 877 */ 878 int 879 fzap_lookup(zap_name_t *zn, 880 uint64_t integer_size, uint64_t num_integers, void *buf, 881 char *realname, int rn_len, boolean_t *ncp) 882 { 883 zap_leaf_t *l; 884 zap_entry_handle_t zeh; 885 886 int err = fzap_checkname(zn); 887 if (err != 0) 888 return (err); 889 890 err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l); 891 if (err != 0) 892 return (err); 893 err = zap_leaf_lookup(l, zn, &zeh); 894 if (err == 0) { 895 if ((err = fzap_checksize(integer_size, num_integers)) != 0) { 896 zap_put_leaf(l); 897 return (err); 898 } 899 900 err = zap_entry_read(&zeh, integer_size, num_integers, buf); 901 (void) zap_entry_read_name(zn->zn_zap, &zeh, rn_len, realname); 902 if (ncp) { 903 *ncp = zap_entry_normalization_conflict(&zeh, 904 zn, NULL, zn->zn_zap); 905 } 906 } 907 908 zap_put_leaf(l); 909 return (err); 910 } 911 912 int 913 fzap_add_cd(zap_name_t *zn, 914 uint64_t integer_size, uint64_t num_integers, 915 const void *val, uint32_t cd, const void *tag, dmu_tx_t *tx) 916 { 917 zap_leaf_t *l; 918 int err; 919 zap_entry_handle_t zeh; 920 zap_t *zap = zn->zn_zap; 921 922 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 923 ASSERT(!zap->zap_ismicro); 924 ASSERT(fzap_check(zn, integer_size, num_integers) == 0); 925 926 err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l); 927 if (err != 0) 928 return (err); 929 retry: 930 err = zap_leaf_lookup(l, zn, &zeh); 931 if (err == 0) { 932 err = SET_ERROR(EEXIST); 933 goto out; 934 } 935 if (err != ENOENT) 936 goto out; 937 938 err = zap_entry_create(l, zn, cd, 939 integer_size, num_integers, val, &zeh); 940 941 if (err == 0) { 942 zap_increment_num_entries(zap, 1, tx); 943 } else if (err == EAGAIN) { 944 err = zap_expand_leaf(zn, l, tag, tx, &l); 945 zap = zn->zn_zap; /* zap_expand_leaf() may change zap */ 946 if (err == 0) 947 goto retry; 948 } 949 950 out: 951 if (l != NULL) { 952 if (err == ENOSPC) 953 zap_put_leaf(l); 954 else 955 zap_put_leaf_maybe_grow_ptrtbl(zn, l, tag, tx); 956 } 957 return (err); 958 } 959 960 int 961 fzap_add(zap_name_t *zn, 962 uint64_t integer_size, uint64_t num_integers, 963 const void *val, const void *tag, dmu_tx_t *tx) 964 { 965 int err = fzap_check(zn, integer_size, num_integers); 966 if (err != 0) 967 return (err); 968 969 return (fzap_add_cd(zn, integer_size, num_integers, 970 val, ZAP_NEED_CD, tag, tx)); 971 } 972 973 int 974 fzap_update(zap_name_t *zn, 975 int integer_size, uint64_t num_integers, const void *val, 976 const void *tag, dmu_tx_t *tx) 977 { 978 zap_leaf_t *l; 979 int err; 980 boolean_t create; 981 zap_entry_handle_t zeh; 982 zap_t *zap = zn->zn_zap; 983 984 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 985 err = fzap_check(zn, integer_size, num_integers); 986 if (err != 0) 987 return (err); 988 989 err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l); 990 if (err != 0) 991 return (err); 992 retry: 993 err = zap_leaf_lookup(l, zn, &zeh); 994 create = (err == ENOENT); 995 ASSERT(err == 0 || err == ENOENT); 996 997 if (create) { 998 err = zap_entry_create(l, zn, ZAP_NEED_CD, 999 integer_size, num_integers, val, &zeh); 1000 if (err == 0) 1001 zap_increment_num_entries(zap, 1, tx); 1002 } else { 1003 err = zap_entry_update(&zeh, integer_size, num_integers, val); 1004 } 1005 1006 if (err == EAGAIN) { 1007 err = zap_expand_leaf(zn, l, tag, tx, &l); 1008 zap = zn->zn_zap; /* zap_expand_leaf() may change zap */ 1009 if (err == 0) 1010 goto retry; 1011 } 1012 1013 if (l != NULL) { 1014 if (err == ENOSPC) 1015 zap_put_leaf(l); 1016 else 1017 zap_put_leaf_maybe_grow_ptrtbl(zn, l, tag, tx); 1018 } 1019 return (err); 1020 } 1021 1022 int 1023 fzap_length(zap_name_t *zn, 1024 uint64_t *integer_size, uint64_t *num_integers) 1025 { 1026 zap_leaf_t *l; 1027 int err; 1028 zap_entry_handle_t zeh; 1029 1030 err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l); 1031 if (err != 0) 1032 return (err); 1033 err = zap_leaf_lookup(l, zn, &zeh); 1034 if (err != 0) 1035 goto out; 1036 1037 if (integer_size != NULL) 1038 *integer_size = zeh.zeh_integer_size; 1039 if (num_integers != NULL) 1040 *num_integers = zeh.zeh_num_integers; 1041 out: 1042 zap_put_leaf(l); 1043 return (err); 1044 } 1045 1046 int 1047 fzap_remove(zap_name_t *zn, dmu_tx_t *tx) 1048 { 1049 zap_leaf_t *l; 1050 int err; 1051 zap_entry_handle_t zeh; 1052 1053 err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, tx, RW_WRITER, &l); 1054 if (err != 0) 1055 return (err); 1056 err = zap_leaf_lookup(l, zn, &zeh); 1057 if (err == 0) { 1058 zap_entry_remove(&zeh); 1059 zap_increment_num_entries(zn->zn_zap, -1, tx); 1060 1061 if (zap_leaf_phys(l)->l_hdr.lh_nentries == 0 && 1062 zap_shrink_enabled) 1063 return (zap_shrink(zn, l, tx)); 1064 } 1065 zap_put_leaf(l); 1066 return (err); 1067 } 1068 1069 void 1070 fzap_prefetch(zap_name_t *zn) 1071 { 1072 uint64_t blk; 1073 zap_t *zap = zn->zn_zap; 1074 1075 uint64_t idx = ZAP_HASH_IDX(zn->zn_hash, 1076 zap_f_phys(zap)->zap_ptrtbl.zt_shift); 1077 if (zap_idx_to_blk(zap, idx, &blk) != 0) 1078 return; 1079 int bs = FZAP_BLOCK_SHIFT(zap); 1080 dmu_prefetch_by_dnode(zap->zap_dnode, 0, blk << bs, 1 << bs, 1081 ZIO_PRIORITY_SYNC_READ); 1082 } 1083 1084 /* 1085 * Helper functions for consumers. 1086 */ 1087 1088 uint64_t 1089 zap_create_link(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj, 1090 const char *name, dmu_tx_t *tx) 1091 { 1092 return (zap_create_link_dnsize(os, ot, parent_obj, name, 0, tx)); 1093 } 1094 1095 uint64_t 1096 zap_create_link_dnsize(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj, 1097 const char *name, int dnodesize, dmu_tx_t *tx) 1098 { 1099 uint64_t new_obj; 1100 1101 new_obj = zap_create_dnsize(os, ot, DMU_OT_NONE, 0, dnodesize, tx); 1102 VERIFY(new_obj != 0); 1103 VERIFY0(zap_add(os, parent_obj, name, sizeof (uint64_t), 1, &new_obj, 1104 tx)); 1105 1106 return (new_obj); 1107 } 1108 1109 int 1110 zap_value_search(objset_t *os, uint64_t zapobj, uint64_t value, uint64_t mask, 1111 char *name, uint64_t namelen) 1112 { 1113 zap_cursor_t zc; 1114 int err; 1115 1116 if (mask == 0) 1117 mask = -1ULL; 1118 1119 zap_attribute_t *za = zap_attribute_long_alloc(); 1120 for (zap_cursor_init(&zc, os, zapobj); 1121 (err = zap_cursor_retrieve(&zc, za)) == 0; 1122 zap_cursor_advance(&zc)) { 1123 if ((za->za_first_integer & mask) == (value & mask)) { 1124 if (strlcpy(name, za->za_name, namelen) >= namelen) 1125 err = SET_ERROR(ENAMETOOLONG); 1126 break; 1127 } 1128 } 1129 zap_cursor_fini(&zc); 1130 zap_attribute_free(za); 1131 return (err); 1132 } 1133 1134 int 1135 zap_join(objset_t *os, uint64_t fromobj, uint64_t intoobj, dmu_tx_t *tx) 1136 { 1137 zap_cursor_t zc; 1138 int err = 0; 1139 1140 zap_attribute_t *za = zap_attribute_long_alloc(); 1141 for (zap_cursor_init(&zc, os, fromobj); 1142 zap_cursor_retrieve(&zc, za) == 0; 1143 (void) zap_cursor_advance(&zc)) { 1144 if (za->za_integer_length != 8 || za->za_num_integers != 1) { 1145 err = SET_ERROR(EINVAL); 1146 break; 1147 } 1148 err = zap_add(os, intoobj, za->za_name, 1149 8, 1, &za->za_first_integer, tx); 1150 if (err != 0) 1151 break; 1152 } 1153 zap_cursor_fini(&zc); 1154 zap_attribute_free(za); 1155 return (err); 1156 } 1157 1158 int 1159 zap_join_key(objset_t *os, uint64_t fromobj, uint64_t intoobj, 1160 uint64_t value, dmu_tx_t *tx) 1161 { 1162 zap_cursor_t zc; 1163 int err = 0; 1164 1165 zap_attribute_t *za = zap_attribute_long_alloc(); 1166 for (zap_cursor_init(&zc, os, fromobj); 1167 zap_cursor_retrieve(&zc, za) == 0; 1168 (void) zap_cursor_advance(&zc)) { 1169 if (za->za_integer_length != 8 || za->za_num_integers != 1) { 1170 err = SET_ERROR(EINVAL); 1171 break; 1172 } 1173 err = zap_add(os, intoobj, za->za_name, 1174 8, 1, &value, tx); 1175 if (err != 0) 1176 break; 1177 } 1178 zap_cursor_fini(&zc); 1179 zap_attribute_free(za); 1180 return (err); 1181 } 1182 1183 int 1184 zap_join_increment(objset_t *os, uint64_t fromobj, uint64_t intoobj, 1185 dmu_tx_t *tx) 1186 { 1187 zap_cursor_t zc; 1188 int err = 0; 1189 1190 zap_attribute_t *za = zap_attribute_long_alloc(); 1191 for (zap_cursor_init(&zc, os, fromobj); 1192 zap_cursor_retrieve(&zc, za) == 0; 1193 (void) zap_cursor_advance(&zc)) { 1194 uint64_t delta = 0; 1195 1196 if (za->za_integer_length != 8 || za->za_num_integers != 1) { 1197 err = SET_ERROR(EINVAL); 1198 break; 1199 } 1200 1201 err = zap_lookup(os, intoobj, za->za_name, 8, 1, &delta); 1202 if (err != 0 && err != ENOENT) 1203 break; 1204 delta += za->za_first_integer; 1205 err = zap_update(os, intoobj, za->za_name, 8, 1, &delta, tx); 1206 if (err != 0) 1207 break; 1208 } 1209 zap_cursor_fini(&zc); 1210 zap_attribute_free(za); 1211 return (err); 1212 } 1213 1214 int 1215 zap_add_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx) 1216 { 1217 char name[20]; 1218 1219 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)value); 1220 return (zap_add(os, obj, name, 8, 1, &value, tx)); 1221 } 1222 1223 int 1224 zap_remove_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx) 1225 { 1226 char name[20]; 1227 1228 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)value); 1229 return (zap_remove(os, obj, name, tx)); 1230 } 1231 1232 int 1233 zap_lookup_int(objset_t *os, uint64_t obj, uint64_t value) 1234 { 1235 char name[20]; 1236 1237 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)value); 1238 return (zap_lookup(os, obj, name, 8, 1, &value)); 1239 } 1240 1241 int 1242 zap_add_int_key(objset_t *os, uint64_t obj, 1243 uint64_t key, uint64_t value, dmu_tx_t *tx) 1244 { 1245 char name[20]; 1246 1247 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1248 return (zap_add(os, obj, name, 8, 1, &value, tx)); 1249 } 1250 1251 int 1252 zap_update_int_key(objset_t *os, uint64_t obj, 1253 uint64_t key, uint64_t value, dmu_tx_t *tx) 1254 { 1255 char name[20]; 1256 1257 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1258 return (zap_update(os, obj, name, 8, 1, &value, tx)); 1259 } 1260 1261 int 1262 zap_lookup_int_key(objset_t *os, uint64_t obj, uint64_t key, uint64_t *valuep) 1263 { 1264 char name[20]; 1265 1266 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1267 return (zap_lookup(os, obj, name, 8, 1, valuep)); 1268 } 1269 1270 int 1271 zap_increment(objset_t *os, uint64_t obj, const char *name, int64_t delta, 1272 dmu_tx_t *tx) 1273 { 1274 uint64_t value = 0; 1275 1276 if (delta == 0) 1277 return (0); 1278 1279 int err = zap_lookup(os, obj, name, 8, 1, &value); 1280 if (err != 0 && err != ENOENT) 1281 return (err); 1282 value += delta; 1283 if (value == 0) 1284 err = zap_remove(os, obj, name, tx); 1285 else 1286 err = zap_update(os, obj, name, 8, 1, &value, tx); 1287 return (err); 1288 } 1289 1290 int 1291 zap_increment_int(objset_t *os, uint64_t obj, uint64_t key, int64_t delta, 1292 dmu_tx_t *tx) 1293 { 1294 char name[20]; 1295 1296 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1297 return (zap_increment(os, obj, name, delta, tx)); 1298 } 1299 1300 /* 1301 * Routines for iterating over the attributes. 1302 */ 1303 1304 int 1305 fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za) 1306 { 1307 int err = ENOENT; 1308 zap_entry_handle_t zeh; 1309 zap_leaf_t *l; 1310 1311 /* retrieve the next entry at or after zc_hash/zc_cd */ 1312 /* if no entry, return ENOENT */ 1313 1314 /* 1315 * If we are reading from the beginning, we're almost certain to 1316 * iterate over the entire ZAP object. If there are multiple leaf 1317 * blocks (freeblk > 2), prefetch the whole object (up to 1318 * dmu_prefetch_max bytes), so that we read the leaf blocks 1319 * concurrently. (Unless noprefetch was requested via 1320 * zap_cursor_init_noprefetch()). 1321 */ 1322 if (zc->zc_hash == 0 && zap_iterate_prefetch && 1323 zc->zc_prefetch && zap_f_phys(zap)->zap_freeblk > 2) { 1324 dmu_prefetch_by_dnode(zap->zap_dnode, 0, 0, 1325 zap_f_phys(zap)->zap_freeblk << FZAP_BLOCK_SHIFT(zap), 1326 ZIO_PRIORITY_ASYNC_READ); 1327 } 1328 1329 if (zc->zc_leaf) { 1330 rw_enter(&zc->zc_leaf->l_rwlock, RW_READER); 1331 1332 /* 1333 * The leaf was either shrunk or split. 1334 */ 1335 if ((zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_block_type == 0) || 1336 (ZAP_HASH_IDX(zc->zc_hash, 1337 zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix_len) != 1338 zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix)) { 1339 zap_put_leaf(zc->zc_leaf); 1340 zc->zc_leaf = NULL; 1341 } 1342 } 1343 1344 again: 1345 if (zc->zc_leaf == NULL) { 1346 err = zap_deref_leaf(zap, zc->zc_hash, NULL, RW_READER, 1347 &zc->zc_leaf); 1348 if (err != 0) 1349 return (err); 1350 } 1351 l = zc->zc_leaf; 1352 1353 err = zap_leaf_lookup_closest(l, zc->zc_hash, zc->zc_cd, &zeh); 1354 1355 if (err == ENOENT) { 1356 if (zap_leaf_phys(l)->l_hdr.lh_prefix_len == 0) { 1357 zc->zc_hash = -1ULL; 1358 zc->zc_cd = 0; 1359 } else { 1360 uint64_t nocare = (1ULL << 1361 (64 - zap_leaf_phys(l)->l_hdr.lh_prefix_len)) - 1; 1362 1363 zc->zc_hash = (zc->zc_hash & ~nocare) + nocare + 1; 1364 zc->zc_cd = 0; 1365 1366 if (zc->zc_hash == 0) { 1367 zc->zc_hash = -1ULL; 1368 } else { 1369 zap_put_leaf(zc->zc_leaf); 1370 zc->zc_leaf = NULL; 1371 goto again; 1372 } 1373 } 1374 } 1375 1376 if (err == 0) { 1377 zc->zc_hash = zeh.zeh_hash; 1378 zc->zc_cd = zeh.zeh_cd; 1379 za->za_integer_length = zeh.zeh_integer_size; 1380 za->za_num_integers = zeh.zeh_num_integers; 1381 if (zeh.zeh_num_integers == 0) { 1382 za->za_first_integer = 0; 1383 } else { 1384 err = zap_entry_read(&zeh, 8, 1, &za->za_first_integer); 1385 ASSERT(err == 0 || err == EOVERFLOW); 1386 } 1387 err = zap_entry_read_name(zap, &zeh, 1388 za->za_name_len, za->za_name); 1389 ASSERT(err == 0); 1390 1391 za->za_normalization_conflict = 1392 zap_entry_normalization_conflict(&zeh, 1393 NULL, za->za_name, zap); 1394 } 1395 rw_exit(&zc->zc_leaf->l_rwlock); 1396 return (err); 1397 } 1398 1399 static void 1400 zap_stats_ptrtbl(zap_t *zap, uint64_t *tbl, int len, zap_stats_t *zs) 1401 { 1402 uint64_t lastblk = 0; 1403 1404 /* 1405 * NB: if a leaf has more pointers than an entire ptrtbl block 1406 * can hold, then it'll be accounted for more than once, since 1407 * we won't have lastblk. 1408 */ 1409 for (int i = 0; i < len; i++) { 1410 zap_leaf_t *l; 1411 1412 if (tbl[i] == lastblk) 1413 continue; 1414 lastblk = tbl[i]; 1415 1416 int err = zap_get_leaf_byblk(zap, tbl[i], NULL, RW_READER, &l); 1417 if (err == 0) { 1418 zap_leaf_stats(zap, l, zs); 1419 zap_put_leaf(l); 1420 } 1421 } 1422 } 1423 1424 void 1425 fzap_get_stats(zap_t *zap, zap_stats_t *zs) 1426 { 1427 int bs = FZAP_BLOCK_SHIFT(zap); 1428 zs->zs_blocksize = 1ULL << bs; 1429 1430 /* 1431 * Set zap_phys_t fields 1432 */ 1433 zs->zs_num_leafs = zap_f_phys(zap)->zap_num_leafs; 1434 zs->zs_num_entries = zap_f_phys(zap)->zap_num_entries; 1435 zs->zs_num_blocks = zap_f_phys(zap)->zap_freeblk; 1436 zs->zs_block_type = zap_f_phys(zap)->zap_block_type; 1437 zs->zs_magic = zap_f_phys(zap)->zap_magic; 1438 zs->zs_salt = zap_f_phys(zap)->zap_salt; 1439 1440 /* 1441 * Set zap_ptrtbl fields 1442 */ 1443 zs->zs_ptrtbl_len = 1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1444 zs->zs_ptrtbl_nextblk = zap_f_phys(zap)->zap_ptrtbl.zt_nextblk; 1445 zs->zs_ptrtbl_blks_copied = 1446 zap_f_phys(zap)->zap_ptrtbl.zt_blks_copied; 1447 zs->zs_ptrtbl_zt_blk = zap_f_phys(zap)->zap_ptrtbl.zt_blk; 1448 zs->zs_ptrtbl_zt_numblks = zap_f_phys(zap)->zap_ptrtbl.zt_numblks; 1449 zs->zs_ptrtbl_zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1450 1451 if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) { 1452 /* the ptrtbl is entirely in the header block. */ 1453 zap_stats_ptrtbl(zap, &ZAP_EMBEDDED_PTRTBL_ENT(zap, 0), 1454 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap), zs); 1455 } else { 1456 dmu_prefetch_by_dnode(zap->zap_dnode, 0, 1457 zap_f_phys(zap)->zap_ptrtbl.zt_blk << bs, 1458 zap_f_phys(zap)->zap_ptrtbl.zt_numblks << bs, 1459 ZIO_PRIORITY_SYNC_READ); 1460 1461 for (int b = 0; b < zap_f_phys(zap)->zap_ptrtbl.zt_numblks; 1462 b++) { 1463 dmu_buf_t *db; 1464 int err; 1465 1466 err = dmu_buf_hold_by_dnode(zap->zap_dnode, 1467 (zap_f_phys(zap)->zap_ptrtbl.zt_blk + b) << bs, 1468 FTAG, &db, DMU_READ_NO_PREFETCH); 1469 if (err == 0) { 1470 zap_stats_ptrtbl(zap, db->db_data, 1471 1<<(bs-3), zs); 1472 dmu_buf_rele(db, FTAG); 1473 } 1474 } 1475 } 1476 } 1477 1478 /* 1479 * Find last allocated block and update freeblk. 1480 */ 1481 static void 1482 zap_trunc(zap_t *zap) 1483 { 1484 uint64_t nentries; 1485 uint64_t lastblk; 1486 1487 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 1488 1489 if (zap_f_phys(zap)->zap_ptrtbl.zt_blk > 0) { 1490 /* External ptrtbl */ 1491 nentries = (1 << zap_f_phys(zap)->zap_ptrtbl.zt_shift); 1492 lastblk = zap_f_phys(zap)->zap_ptrtbl.zt_blk + 1493 zap_f_phys(zap)->zap_ptrtbl.zt_numblks - 1; 1494 } else { 1495 /* Embedded ptrtbl */ 1496 nentries = (1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap)); 1497 lastblk = 0; 1498 } 1499 1500 for (uint64_t idx = 0; idx < nentries; idx++) { 1501 uint64_t blk; 1502 if (zap_idx_to_blk(zap, idx, &blk) != 0) 1503 return; 1504 if (blk > lastblk) 1505 lastblk = blk; 1506 } 1507 1508 ASSERT3U(lastblk, <, zap_f_phys(zap)->zap_freeblk); 1509 1510 zap_f_phys(zap)->zap_freeblk = lastblk + 1; 1511 } 1512 1513 /* 1514 * ZAP shrinking algorithm. 1515 * 1516 * We shrink ZAP recuresively removing empty leaves. We can remove an empty leaf 1517 * only if it has a sibling. Sibling leaves have the same prefix length and 1518 * their prefixes differ only by the least significant (sibling) bit. We require 1519 * both siblings to be empty. This eliminates a need to rehash the non-empty 1520 * remaining leaf. When we have removed one of two empty sibling, we set ptrtbl 1521 * entries of the removed leaf to point out to the remaining leaf. Prefix length 1522 * of the remaining leaf is decremented. As a result, it has a new prefix and it 1523 * might have a new sibling. So, we repeat the process. 1524 * 1525 * Steps: 1526 * 1. Check if a sibling leaf (sl) exists and it is empty. 1527 * 2. Release the leaf (l) if it has the sibling bit (slbit) equal to 1. 1528 * 3. Release the sibling (sl) to derefer it again with WRITER lock. 1529 * 4. Upgrade zapdir lock to WRITER (once). 1530 * 5. Derefer released leaves again. 1531 * 6. If it is needed, recheck whether both leaves are still siblings and empty. 1532 * 7. Set ptrtbl pointers of the removed leaf (slbit 1) to point out to blkid of 1533 * the remaining leaf (slbit 0). 1534 * 8. Free disk block of the removed leaf (dmu_free_range). 1535 * 9. Decrement prefix_len of the remaining leaf. 1536 * 10. Repeat the steps. 1537 */ 1538 static int 1539 zap_shrink(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx) 1540 { 1541 zap_t *zap = zn->zn_zap; 1542 int64_t zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1543 uint64_t hash = zn->zn_hash; 1544 uint64_t prefix = zap_leaf_phys(l)->l_hdr.lh_prefix; 1545 uint64_t prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len; 1546 boolean_t trunc = B_FALSE; 1547 int err = 0; 1548 1549 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nentries, ==, 0); 1550 ASSERT3U(prefix_len, <=, zap_f_phys(zap)->zap_ptrtbl.zt_shift); 1551 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 1552 ASSERT3U(ZAP_HASH_IDX(hash, prefix_len), ==, prefix); 1553 1554 boolean_t writer = B_FALSE; 1555 1556 /* 1557 * To avoid deadlock always deref leaves in the same order - 1558 * sibling 0 first, then sibling 1. 1559 */ 1560 while (prefix_len) { 1561 zap_leaf_t *sl; 1562 int64_t prefix_diff = zt_shift - prefix_len; 1563 uint64_t sl_prefix = prefix ^ 1; 1564 uint64_t sl_hash = ZAP_PREFIX_HASH(sl_prefix, prefix_len); 1565 int slbit = prefix & 1; 1566 1567 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nentries, ==, 0); 1568 1569 /* 1570 * Check if there is a sibling by reading ptrtbl ptrs. 1571 */ 1572 if (check_sibling_ptrtbl_range(zap, sl_prefix, prefix_len) == 0) 1573 break; 1574 1575 /* 1576 * sibling 1, unlock it - we haven't yet dereferenced sibling 0. 1577 */ 1578 if (slbit == 1) { 1579 zap_put_leaf(l); 1580 l = NULL; 1581 } 1582 1583 /* 1584 * Dereference sibling leaf and check if it is empty. 1585 */ 1586 if ((err = zap_deref_leaf(zap, sl_hash, tx, RW_READER, 1587 &sl)) != 0) 1588 break; 1589 1590 ASSERT3U(ZAP_HASH_IDX(sl_hash, prefix_len), ==, sl_prefix); 1591 1592 /* 1593 * Check if we have a sibling and it is empty. 1594 */ 1595 if (zap_leaf_phys(sl)->l_hdr.lh_prefix_len != prefix_len || 1596 zap_leaf_phys(sl)->l_hdr.lh_nentries != 0) { 1597 zap_put_leaf(sl); 1598 break; 1599 } 1600 1601 zap_put_leaf(sl); 1602 1603 /* 1604 * If there two empty sibling, we have work to do, so 1605 * we need to lock ZAP ptrtbl as WRITER. 1606 */ 1607 if (!writer && (writer = zap_tryupgradedir(zap, tx)) == 0) { 1608 /* We failed to upgrade */ 1609 if (l != NULL) { 1610 zap_put_leaf(l); 1611 l = NULL; 1612 } 1613 1614 /* 1615 * Usually, the right way to upgrade from a READER lock 1616 * to a WRITER lock is to call zap_unlockdir() and 1617 * zap_lockdir(), but we do not have a tag. Instead, 1618 * we do it in more sophisticated way. 1619 */ 1620 rw_exit(&zap->zap_rwlock); 1621 rw_enter(&zap->zap_rwlock, RW_WRITER); 1622 dmu_buf_will_dirty(zap->zap_dbuf, tx); 1623 1624 zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1625 writer = B_TRUE; 1626 } 1627 1628 /* 1629 * Here we have WRITER lock for ptrtbl. 1630 * Now, we need a WRITER lock for both siblings leaves. 1631 * Also, we have to recheck if the leaves are still siblings 1632 * and still empty. 1633 */ 1634 if (l == NULL) { 1635 /* sibling 0 */ 1636 if ((err = zap_deref_leaf(zap, (slbit ? sl_hash : hash), 1637 tx, RW_WRITER, &l)) != 0) 1638 break; 1639 1640 /* 1641 * The leaf isn't empty anymore or 1642 * it was shrunk/split while our locks were down. 1643 */ 1644 if (zap_leaf_phys(l)->l_hdr.lh_nentries != 0 || 1645 zap_leaf_phys(l)->l_hdr.lh_prefix_len != prefix_len) 1646 break; 1647 } 1648 1649 /* sibling 1 */ 1650 if ((err = zap_deref_leaf(zap, (slbit ? hash : sl_hash), tx, 1651 RW_WRITER, &sl)) != 0) 1652 break; 1653 1654 /* 1655 * The leaf isn't empty anymore or 1656 * it was shrunk/split while our locks were down. 1657 */ 1658 if (zap_leaf_phys(sl)->l_hdr.lh_nentries != 0 || 1659 zap_leaf_phys(sl)->l_hdr.lh_prefix_len != prefix_len) { 1660 zap_put_leaf(sl); 1661 break; 1662 } 1663 1664 /* If we have gotten here, we have a leaf to collapse */ 1665 uint64_t idx = (slbit ? prefix : sl_prefix) << prefix_diff; 1666 uint64_t nptrs = (1ULL << prefix_diff); 1667 uint64_t sl_blkid = sl->l_blkid; 1668 1669 /* 1670 * Set ptrtbl entries to point out to the slibling 0 blkid 1671 */ 1672 if ((err = zap_set_idx_range_to_blk(zap, idx, nptrs, l->l_blkid, 1673 tx)) != 0) { 1674 zap_put_leaf(sl); 1675 break; 1676 } 1677 1678 /* 1679 * Free sibling 1 disk block. 1680 */ 1681 int bs = FZAP_BLOCK_SHIFT(zap); 1682 if (sl_blkid == zap_f_phys(zap)->zap_freeblk - 1) 1683 trunc = B_TRUE; 1684 1685 (void) dmu_free_range(zap->zap_objset, zap->zap_object, 1686 sl_blkid << bs, 1 << bs, tx); 1687 zap_put_leaf(sl); 1688 1689 zap_f_phys(zap)->zap_num_leafs--; 1690 1691 /* 1692 * Update prefix and prefix_len. 1693 */ 1694 zap_leaf_phys(l)->l_hdr.lh_prefix >>= 1; 1695 zap_leaf_phys(l)->l_hdr.lh_prefix_len--; 1696 1697 prefix = zap_leaf_phys(l)->l_hdr.lh_prefix; 1698 prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len; 1699 } 1700 1701 if (trunc) 1702 zap_trunc(zap); 1703 1704 if (l != NULL) 1705 zap_put_leaf(l); 1706 1707 return (err); 1708 } 1709 1710 ZFS_MODULE_PARAM(zfs, , zap_iterate_prefetch, INT, ZMOD_RW, 1711 "When iterating ZAP object, prefetch it"); 1712 1713 ZFS_MODULE_PARAM(zfs, , zap_shrink_enabled, INT, ZMOD_RW, 1714 "Enable ZAP shrinking"); 1715