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