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 uint32_t maxnamelen = zn->zn_normbuf_len; 836 uint64_t len = (uint64_t)zn->zn_key_orig_numints * zn->zn_key_intlen; 837 /* Only allow directory zap to have longname */ 838 if (len > maxnamelen || 839 (len > ZAP_MAXNAMELEN && 840 zn->zn_zap->zap_dnode->dn_type != DMU_OT_DIRECTORY_CONTENTS)) 841 return (SET_ERROR(ENAMETOOLONG)); 842 return (0); 843 } 844 845 static int 846 fzap_checksize(uint64_t integer_size, uint64_t num_integers) 847 { 848 /* Only integer sizes supported by C */ 849 switch (integer_size) { 850 case 1: 851 case 2: 852 case 4: 853 case 8: 854 break; 855 default: 856 return (SET_ERROR(EINVAL)); 857 } 858 859 if (integer_size * num_integers > ZAP_MAXVALUELEN) 860 return (SET_ERROR(E2BIG)); 861 862 return (0); 863 } 864 865 static int 866 fzap_check(zap_name_t *zn, uint64_t integer_size, uint64_t num_integers) 867 { 868 int err = fzap_checkname(zn); 869 if (err != 0) 870 return (err); 871 return (fzap_checksize(integer_size, num_integers)); 872 } 873 874 /* 875 * Routines for manipulating attributes. 876 */ 877 int 878 fzap_lookup(zap_name_t *zn, 879 uint64_t integer_size, uint64_t num_integers, void *buf, 880 char *realname, int rn_len, boolean_t *ncp) 881 { 882 zap_leaf_t *l; 883 zap_entry_handle_t zeh; 884 885 int err = fzap_checkname(zn); 886 if (err != 0) 887 return (err); 888 889 err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l); 890 if (err != 0) 891 return (err); 892 err = zap_leaf_lookup(l, zn, &zeh); 893 if (err == 0) { 894 if ((err = fzap_checksize(integer_size, num_integers)) != 0) { 895 zap_put_leaf(l); 896 return (err); 897 } 898 899 err = zap_entry_read(&zeh, integer_size, num_integers, buf); 900 (void) zap_entry_read_name(zn->zn_zap, &zeh, rn_len, realname); 901 if (ncp) { 902 *ncp = zap_entry_normalization_conflict(&zeh, 903 zn, NULL, zn->zn_zap); 904 } 905 } 906 907 zap_put_leaf(l); 908 return (err); 909 } 910 911 int 912 fzap_add_cd(zap_name_t *zn, 913 uint64_t integer_size, uint64_t num_integers, 914 const void *val, uint32_t cd, const void *tag, dmu_tx_t *tx) 915 { 916 zap_leaf_t *l; 917 int err; 918 zap_entry_handle_t zeh; 919 zap_t *zap = zn->zn_zap; 920 921 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 922 ASSERT(!zap->zap_ismicro); 923 ASSERT(fzap_check(zn, integer_size, num_integers) == 0); 924 925 err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l); 926 if (err != 0) 927 return (err); 928 retry: 929 err = zap_leaf_lookup(l, zn, &zeh); 930 if (err == 0) { 931 err = SET_ERROR(EEXIST); 932 goto out; 933 } 934 if (err != ENOENT) 935 goto out; 936 937 err = zap_entry_create(l, zn, cd, 938 integer_size, num_integers, val, &zeh); 939 940 if (err == 0) { 941 zap_increment_num_entries(zap, 1, tx); 942 } else if (err == EAGAIN) { 943 err = zap_expand_leaf(zn, l, tag, tx, &l); 944 zap = zn->zn_zap; /* zap_expand_leaf() may change zap */ 945 if (err == 0) 946 goto retry; 947 } 948 949 out: 950 if (l != NULL) { 951 if (err == ENOSPC) 952 zap_put_leaf(l); 953 else 954 zap_put_leaf_maybe_grow_ptrtbl(zn, l, tag, tx); 955 } 956 return (err); 957 } 958 959 int 960 fzap_add(zap_name_t *zn, 961 uint64_t integer_size, uint64_t num_integers, 962 const void *val, const void *tag, dmu_tx_t *tx) 963 { 964 int err = fzap_check(zn, integer_size, num_integers); 965 if (err != 0) 966 return (err); 967 968 return (fzap_add_cd(zn, integer_size, num_integers, 969 val, ZAP_NEED_CD, tag, tx)); 970 } 971 972 int 973 fzap_update(zap_name_t *zn, 974 int integer_size, uint64_t num_integers, const void *val, 975 const void *tag, dmu_tx_t *tx) 976 { 977 zap_leaf_t *l; 978 int err; 979 boolean_t create; 980 zap_entry_handle_t zeh; 981 zap_t *zap = zn->zn_zap; 982 983 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 984 err = fzap_check(zn, integer_size, num_integers); 985 if (err != 0) 986 return (err); 987 988 err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l); 989 if (err != 0) 990 return (err); 991 retry: 992 err = zap_leaf_lookup(l, zn, &zeh); 993 create = (err == ENOENT); 994 ASSERT(err == 0 || err == ENOENT); 995 996 if (create) { 997 err = zap_entry_create(l, zn, ZAP_NEED_CD, 998 integer_size, num_integers, val, &zeh); 999 if (err == 0) 1000 zap_increment_num_entries(zap, 1, tx); 1001 } else { 1002 err = zap_entry_update(&zeh, integer_size, num_integers, val); 1003 } 1004 1005 if (err == EAGAIN) { 1006 err = zap_expand_leaf(zn, l, tag, tx, &l); 1007 zap = zn->zn_zap; /* zap_expand_leaf() may change zap */ 1008 if (err == 0) 1009 goto retry; 1010 } 1011 1012 if (l != NULL) { 1013 if (err == ENOSPC) 1014 zap_put_leaf(l); 1015 else 1016 zap_put_leaf_maybe_grow_ptrtbl(zn, l, tag, tx); 1017 } 1018 return (err); 1019 } 1020 1021 int 1022 fzap_length(zap_name_t *zn, 1023 uint64_t *integer_size, uint64_t *num_integers) 1024 { 1025 zap_leaf_t *l; 1026 int err; 1027 zap_entry_handle_t zeh; 1028 1029 err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l); 1030 if (err != 0) 1031 return (err); 1032 err = zap_leaf_lookup(l, zn, &zeh); 1033 if (err != 0) 1034 goto out; 1035 1036 if (integer_size != NULL) 1037 *integer_size = zeh.zeh_integer_size; 1038 if (num_integers != NULL) 1039 *num_integers = zeh.zeh_num_integers; 1040 out: 1041 zap_put_leaf(l); 1042 return (err); 1043 } 1044 1045 int 1046 fzap_remove(zap_name_t *zn, dmu_tx_t *tx) 1047 { 1048 zap_leaf_t *l; 1049 int err; 1050 zap_entry_handle_t zeh; 1051 1052 err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, tx, RW_WRITER, &l); 1053 if (err != 0) 1054 return (err); 1055 err = zap_leaf_lookup(l, zn, &zeh); 1056 if (err == 0) { 1057 zap_entry_remove(&zeh); 1058 zap_increment_num_entries(zn->zn_zap, -1, tx); 1059 1060 if (zap_leaf_phys(l)->l_hdr.lh_nentries == 0 && 1061 zap_shrink_enabled) 1062 return (zap_shrink(zn, l, tx)); 1063 } 1064 zap_put_leaf(l); 1065 return (err); 1066 } 1067 1068 void 1069 fzap_prefetch(zap_name_t *zn) 1070 { 1071 uint64_t blk; 1072 zap_t *zap = zn->zn_zap; 1073 1074 uint64_t idx = ZAP_HASH_IDX(zn->zn_hash, 1075 zap_f_phys(zap)->zap_ptrtbl.zt_shift); 1076 if (zap_idx_to_blk(zap, idx, &blk) != 0) 1077 return; 1078 int bs = FZAP_BLOCK_SHIFT(zap); 1079 dmu_prefetch_by_dnode(zap->zap_dnode, 0, blk << bs, 1 << bs, 1080 ZIO_PRIORITY_SYNC_READ); 1081 } 1082 1083 /* 1084 * Helper functions for consumers. 1085 */ 1086 1087 uint64_t 1088 zap_create_link(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj, 1089 const char *name, dmu_tx_t *tx) 1090 { 1091 return (zap_create_link_dnsize(os, ot, parent_obj, name, 0, tx)); 1092 } 1093 1094 uint64_t 1095 zap_create_link_dnsize(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj, 1096 const char *name, int dnodesize, dmu_tx_t *tx) 1097 { 1098 uint64_t new_obj; 1099 1100 new_obj = zap_create_dnsize(os, ot, DMU_OT_NONE, 0, dnodesize, tx); 1101 VERIFY(new_obj != 0); 1102 VERIFY0(zap_add(os, parent_obj, name, sizeof (uint64_t), 1, &new_obj, 1103 tx)); 1104 1105 return (new_obj); 1106 } 1107 1108 int 1109 zap_value_search(objset_t *os, uint64_t zapobj, uint64_t value, uint64_t mask, 1110 char *name, uint64_t namelen) 1111 { 1112 zap_cursor_t zc; 1113 int err; 1114 1115 if (mask == 0) 1116 mask = -1ULL; 1117 1118 zap_attribute_t *za = zap_attribute_long_alloc(); 1119 for (zap_cursor_init(&zc, os, zapobj); 1120 (err = zap_cursor_retrieve(&zc, za)) == 0; 1121 zap_cursor_advance(&zc)) { 1122 if ((za->za_first_integer & mask) == (value & mask)) { 1123 if (strlcpy(name, za->za_name, namelen) >= namelen) 1124 err = SET_ERROR(ENAMETOOLONG); 1125 break; 1126 } 1127 } 1128 zap_cursor_fini(&zc); 1129 zap_attribute_free(za); 1130 return (err); 1131 } 1132 1133 int 1134 zap_join(objset_t *os, uint64_t fromobj, uint64_t intoobj, dmu_tx_t *tx) 1135 { 1136 zap_cursor_t zc; 1137 int err = 0; 1138 1139 zap_attribute_t *za = zap_attribute_long_alloc(); 1140 for (zap_cursor_init(&zc, os, fromobj); 1141 zap_cursor_retrieve(&zc, za) == 0; 1142 (void) zap_cursor_advance(&zc)) { 1143 if (za->za_integer_length != 8 || za->za_num_integers != 1) { 1144 err = SET_ERROR(EINVAL); 1145 break; 1146 } 1147 err = zap_add(os, intoobj, za->za_name, 1148 8, 1, &za->za_first_integer, tx); 1149 if (err != 0) 1150 break; 1151 } 1152 zap_cursor_fini(&zc); 1153 zap_attribute_free(za); 1154 return (err); 1155 } 1156 1157 int 1158 zap_join_key(objset_t *os, uint64_t fromobj, uint64_t intoobj, 1159 uint64_t value, dmu_tx_t *tx) 1160 { 1161 zap_cursor_t zc; 1162 int err = 0; 1163 1164 zap_attribute_t *za = zap_attribute_long_alloc(); 1165 for (zap_cursor_init(&zc, os, fromobj); 1166 zap_cursor_retrieve(&zc, za) == 0; 1167 (void) zap_cursor_advance(&zc)) { 1168 if (za->za_integer_length != 8 || za->za_num_integers != 1) { 1169 err = SET_ERROR(EINVAL); 1170 break; 1171 } 1172 err = zap_add(os, intoobj, za->za_name, 1173 8, 1, &value, tx); 1174 if (err != 0) 1175 break; 1176 } 1177 zap_cursor_fini(&zc); 1178 zap_attribute_free(za); 1179 return (err); 1180 } 1181 1182 int 1183 zap_join_increment(objset_t *os, uint64_t fromobj, uint64_t intoobj, 1184 dmu_tx_t *tx) 1185 { 1186 zap_cursor_t zc; 1187 int err = 0; 1188 1189 zap_attribute_t *za = zap_attribute_long_alloc(); 1190 for (zap_cursor_init(&zc, os, fromobj); 1191 zap_cursor_retrieve(&zc, za) == 0; 1192 (void) zap_cursor_advance(&zc)) { 1193 uint64_t delta = 0; 1194 1195 if (za->za_integer_length != 8 || za->za_num_integers != 1) { 1196 err = SET_ERROR(EINVAL); 1197 break; 1198 } 1199 1200 err = zap_lookup(os, intoobj, za->za_name, 8, 1, &delta); 1201 if (err != 0 && err != ENOENT) 1202 break; 1203 delta += za->za_first_integer; 1204 err = zap_update(os, intoobj, za->za_name, 8, 1, &delta, tx); 1205 if (err != 0) 1206 break; 1207 } 1208 zap_cursor_fini(&zc); 1209 zap_attribute_free(za); 1210 return (err); 1211 } 1212 1213 int 1214 zap_add_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx) 1215 { 1216 char name[20]; 1217 1218 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)value); 1219 return (zap_add(os, obj, name, 8, 1, &value, tx)); 1220 } 1221 1222 int 1223 zap_remove_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx) 1224 { 1225 char name[20]; 1226 1227 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)value); 1228 return (zap_remove(os, obj, name, tx)); 1229 } 1230 1231 int 1232 zap_lookup_int(objset_t *os, uint64_t obj, uint64_t value) 1233 { 1234 char name[20]; 1235 1236 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)value); 1237 return (zap_lookup(os, obj, name, 8, 1, &value)); 1238 } 1239 1240 int 1241 zap_add_int_key(objset_t *os, uint64_t obj, 1242 uint64_t key, uint64_t value, dmu_tx_t *tx) 1243 { 1244 char name[20]; 1245 1246 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1247 return (zap_add(os, obj, name, 8, 1, &value, tx)); 1248 } 1249 1250 int 1251 zap_update_int_key(objset_t *os, uint64_t obj, 1252 uint64_t key, uint64_t value, dmu_tx_t *tx) 1253 { 1254 char name[20]; 1255 1256 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1257 return (zap_update(os, obj, name, 8, 1, &value, tx)); 1258 } 1259 1260 int 1261 zap_lookup_int_key(objset_t *os, uint64_t obj, uint64_t key, uint64_t *valuep) 1262 { 1263 char name[20]; 1264 1265 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1266 return (zap_lookup(os, obj, name, 8, 1, valuep)); 1267 } 1268 1269 int 1270 zap_increment(objset_t *os, uint64_t obj, const char *name, int64_t delta, 1271 dmu_tx_t *tx) 1272 { 1273 uint64_t value = 0; 1274 1275 if (delta == 0) 1276 return (0); 1277 1278 int err = zap_lookup(os, obj, name, 8, 1, &value); 1279 if (err != 0 && err != ENOENT) 1280 return (err); 1281 value += delta; 1282 if (value == 0) 1283 err = zap_remove(os, obj, name, tx); 1284 else 1285 err = zap_update(os, obj, name, 8, 1, &value, tx); 1286 return (err); 1287 } 1288 1289 int 1290 zap_increment_int(objset_t *os, uint64_t obj, uint64_t key, int64_t delta, 1291 dmu_tx_t *tx) 1292 { 1293 char name[20]; 1294 1295 (void) snprintf(name, sizeof (name), "%llx", (longlong_t)key); 1296 return (zap_increment(os, obj, name, delta, tx)); 1297 } 1298 1299 /* 1300 * Routines for iterating over the attributes. 1301 */ 1302 1303 int 1304 fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za) 1305 { 1306 int err = ENOENT; 1307 zap_entry_handle_t zeh; 1308 zap_leaf_t *l; 1309 1310 /* retrieve the next entry at or after zc_hash/zc_cd */ 1311 /* if no entry, return ENOENT */ 1312 1313 /* 1314 * If we are reading from the beginning, we're almost certain to 1315 * iterate over the entire ZAP object. If there are multiple leaf 1316 * blocks (freeblk > 2), prefetch the whole object (up to 1317 * dmu_prefetch_max bytes), so that we read the leaf blocks 1318 * concurrently. (Unless noprefetch was requested via 1319 * zap_cursor_init_noprefetch()). 1320 */ 1321 if (zc->zc_hash == 0 && zap_iterate_prefetch && 1322 zc->zc_prefetch && zap_f_phys(zap)->zap_freeblk > 2) { 1323 dmu_prefetch_by_dnode(zap->zap_dnode, 0, 0, 1324 zap_f_phys(zap)->zap_freeblk << FZAP_BLOCK_SHIFT(zap), 1325 ZIO_PRIORITY_ASYNC_READ); 1326 } 1327 1328 if (zc->zc_leaf) { 1329 rw_enter(&zc->zc_leaf->l_rwlock, RW_READER); 1330 1331 /* 1332 * The leaf was either shrunk or split. 1333 */ 1334 if ((zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_block_type == 0) || 1335 (ZAP_HASH_IDX(zc->zc_hash, 1336 zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix_len) != 1337 zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix)) { 1338 zap_put_leaf(zc->zc_leaf); 1339 zc->zc_leaf = NULL; 1340 } 1341 } 1342 1343 again: 1344 if (zc->zc_leaf == NULL) { 1345 err = zap_deref_leaf(zap, zc->zc_hash, NULL, RW_READER, 1346 &zc->zc_leaf); 1347 if (err != 0) 1348 return (err); 1349 } 1350 l = zc->zc_leaf; 1351 1352 err = zap_leaf_lookup_closest(l, zc->zc_hash, zc->zc_cd, &zeh); 1353 1354 if (err == ENOENT) { 1355 if (zap_leaf_phys(l)->l_hdr.lh_prefix_len == 0) { 1356 zc->zc_hash = -1ULL; 1357 zc->zc_cd = 0; 1358 } else { 1359 uint64_t nocare = (1ULL << 1360 (64 - zap_leaf_phys(l)->l_hdr.lh_prefix_len)) - 1; 1361 1362 zc->zc_hash = (zc->zc_hash & ~nocare) + nocare + 1; 1363 zc->zc_cd = 0; 1364 1365 if (zc->zc_hash == 0) { 1366 zc->zc_hash = -1ULL; 1367 } else { 1368 zap_put_leaf(zc->zc_leaf); 1369 zc->zc_leaf = NULL; 1370 goto again; 1371 } 1372 } 1373 } 1374 1375 if (err == 0) { 1376 zc->zc_hash = zeh.zeh_hash; 1377 zc->zc_cd = zeh.zeh_cd; 1378 za->za_integer_length = zeh.zeh_integer_size; 1379 za->za_num_integers = zeh.zeh_num_integers; 1380 if (zeh.zeh_num_integers == 0) { 1381 za->za_first_integer = 0; 1382 } else { 1383 err = zap_entry_read(&zeh, 8, 1, &za->za_first_integer); 1384 ASSERT(err == 0 || err == EOVERFLOW); 1385 } 1386 err = zap_entry_read_name(zap, &zeh, 1387 za->za_name_len, za->za_name); 1388 ASSERT(err == 0); 1389 1390 za->za_normalization_conflict = 1391 zap_entry_normalization_conflict(&zeh, 1392 NULL, za->za_name, zap); 1393 } 1394 rw_exit(&zc->zc_leaf->l_rwlock); 1395 return (err); 1396 } 1397 1398 static void 1399 zap_stats_ptrtbl(zap_t *zap, uint64_t *tbl, int len, zap_stats_t *zs) 1400 { 1401 uint64_t lastblk = 0; 1402 1403 /* 1404 * NB: if a leaf has more pointers than an entire ptrtbl block 1405 * can hold, then it'll be accounted for more than once, since 1406 * we won't have lastblk. 1407 */ 1408 for (int i = 0; i < len; i++) { 1409 zap_leaf_t *l; 1410 1411 if (tbl[i] == lastblk) 1412 continue; 1413 lastblk = tbl[i]; 1414 1415 int err = zap_get_leaf_byblk(zap, tbl[i], NULL, RW_READER, &l); 1416 if (err == 0) { 1417 zap_leaf_stats(zap, l, zs); 1418 zap_put_leaf(l); 1419 } 1420 } 1421 } 1422 1423 void 1424 fzap_get_stats(zap_t *zap, zap_stats_t *zs) 1425 { 1426 int bs = FZAP_BLOCK_SHIFT(zap); 1427 zs->zs_blocksize = 1ULL << bs; 1428 1429 /* 1430 * Set zap_phys_t fields 1431 */ 1432 zs->zs_num_leafs = zap_f_phys(zap)->zap_num_leafs; 1433 zs->zs_num_entries = zap_f_phys(zap)->zap_num_entries; 1434 zs->zs_num_blocks = zap_f_phys(zap)->zap_freeblk; 1435 zs->zs_block_type = zap_f_phys(zap)->zap_block_type; 1436 zs->zs_magic = zap_f_phys(zap)->zap_magic; 1437 zs->zs_salt = zap_f_phys(zap)->zap_salt; 1438 1439 /* 1440 * Set zap_ptrtbl fields 1441 */ 1442 zs->zs_ptrtbl_len = 1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1443 zs->zs_ptrtbl_nextblk = zap_f_phys(zap)->zap_ptrtbl.zt_nextblk; 1444 zs->zs_ptrtbl_blks_copied = 1445 zap_f_phys(zap)->zap_ptrtbl.zt_blks_copied; 1446 zs->zs_ptrtbl_zt_blk = zap_f_phys(zap)->zap_ptrtbl.zt_blk; 1447 zs->zs_ptrtbl_zt_numblks = zap_f_phys(zap)->zap_ptrtbl.zt_numblks; 1448 zs->zs_ptrtbl_zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1449 1450 if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) { 1451 /* the ptrtbl is entirely in the header block. */ 1452 zap_stats_ptrtbl(zap, &ZAP_EMBEDDED_PTRTBL_ENT(zap, 0), 1453 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap), zs); 1454 } else { 1455 dmu_prefetch_by_dnode(zap->zap_dnode, 0, 1456 zap_f_phys(zap)->zap_ptrtbl.zt_blk << bs, 1457 zap_f_phys(zap)->zap_ptrtbl.zt_numblks << bs, 1458 ZIO_PRIORITY_SYNC_READ); 1459 1460 for (int b = 0; b < zap_f_phys(zap)->zap_ptrtbl.zt_numblks; 1461 b++) { 1462 dmu_buf_t *db; 1463 int err; 1464 1465 err = dmu_buf_hold_by_dnode(zap->zap_dnode, 1466 (zap_f_phys(zap)->zap_ptrtbl.zt_blk + b) << bs, 1467 FTAG, &db, DMU_READ_NO_PREFETCH); 1468 if (err == 0) { 1469 zap_stats_ptrtbl(zap, db->db_data, 1470 1<<(bs-3), zs); 1471 dmu_buf_rele(db, FTAG); 1472 } 1473 } 1474 } 1475 } 1476 1477 /* 1478 * Find last allocated block and update freeblk. 1479 */ 1480 static void 1481 zap_trunc(zap_t *zap) 1482 { 1483 uint64_t nentries; 1484 uint64_t lastblk; 1485 1486 ASSERT(RW_WRITE_HELD(&zap->zap_rwlock)); 1487 1488 if (zap_f_phys(zap)->zap_ptrtbl.zt_blk > 0) { 1489 /* External ptrtbl */ 1490 nentries = (1 << zap_f_phys(zap)->zap_ptrtbl.zt_shift); 1491 lastblk = zap_f_phys(zap)->zap_ptrtbl.zt_blk + 1492 zap_f_phys(zap)->zap_ptrtbl.zt_numblks - 1; 1493 } else { 1494 /* Embedded ptrtbl */ 1495 nentries = (1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap)); 1496 lastblk = 0; 1497 } 1498 1499 for (uint64_t idx = 0; idx < nentries; idx++) { 1500 uint64_t blk; 1501 if (zap_idx_to_blk(zap, idx, &blk) != 0) 1502 return; 1503 if (blk > lastblk) 1504 lastblk = blk; 1505 } 1506 1507 ASSERT3U(lastblk, <, zap_f_phys(zap)->zap_freeblk); 1508 1509 zap_f_phys(zap)->zap_freeblk = lastblk + 1; 1510 } 1511 1512 /* 1513 * ZAP shrinking algorithm. 1514 * 1515 * We shrink ZAP recuresively removing empty leaves. We can remove an empty leaf 1516 * only if it has a sibling. Sibling leaves have the same prefix length and 1517 * their prefixes differ only by the least significant (sibling) bit. We require 1518 * both siblings to be empty. This eliminates a need to rehash the non-empty 1519 * remaining leaf. When we have removed one of two empty sibling, we set ptrtbl 1520 * entries of the removed leaf to point out to the remaining leaf. Prefix length 1521 * of the remaining leaf is decremented. As a result, it has a new prefix and it 1522 * might have a new sibling. So, we repeat the process. 1523 * 1524 * Steps: 1525 * 1. Check if a sibling leaf (sl) exists and it is empty. 1526 * 2. Release the leaf (l) if it has the sibling bit (slbit) equal to 1. 1527 * 3. Release the sibling (sl) to derefer it again with WRITER lock. 1528 * 4. Upgrade zapdir lock to WRITER (once). 1529 * 5. Derefer released leaves again. 1530 * 6. If it is needed, recheck whether both leaves are still siblings and empty. 1531 * 7. Set ptrtbl pointers of the removed leaf (slbit 1) to point out to blkid of 1532 * the remaining leaf (slbit 0). 1533 * 8. Free disk block of the removed leaf (dmu_free_range). 1534 * 9. Decrement prefix_len of the remaining leaf. 1535 * 10. Repeat the steps. 1536 */ 1537 static int 1538 zap_shrink(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx) 1539 { 1540 zap_t *zap = zn->zn_zap; 1541 int64_t zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1542 uint64_t hash = zn->zn_hash; 1543 uint64_t prefix = zap_leaf_phys(l)->l_hdr.lh_prefix; 1544 uint64_t prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len; 1545 boolean_t trunc = B_FALSE; 1546 int err = 0; 1547 1548 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nentries, ==, 0); 1549 ASSERT3U(prefix_len, <=, zap_f_phys(zap)->zap_ptrtbl.zt_shift); 1550 ASSERT(RW_LOCK_HELD(&zap->zap_rwlock)); 1551 ASSERT3U(ZAP_HASH_IDX(hash, prefix_len), ==, prefix); 1552 1553 boolean_t writer = B_FALSE; 1554 1555 /* 1556 * To avoid deadlock always deref leaves in the same order - 1557 * sibling 0 first, then sibling 1. 1558 */ 1559 while (prefix_len) { 1560 zap_leaf_t *sl; 1561 int64_t prefix_diff = zt_shift - prefix_len; 1562 uint64_t sl_prefix = prefix ^ 1; 1563 uint64_t sl_hash = ZAP_PREFIX_HASH(sl_prefix, prefix_len); 1564 int slbit = prefix & 1; 1565 1566 ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nentries, ==, 0); 1567 1568 /* 1569 * Check if there is a sibling by reading ptrtbl ptrs. 1570 */ 1571 if (check_sibling_ptrtbl_range(zap, sl_prefix, prefix_len) == 0) 1572 break; 1573 1574 /* 1575 * sibling 1, unlock it - we haven't yet dereferenced sibling 0. 1576 */ 1577 if (slbit == 1) { 1578 zap_put_leaf(l); 1579 l = NULL; 1580 } 1581 1582 /* 1583 * Dereference sibling leaf and check if it is empty. 1584 */ 1585 if ((err = zap_deref_leaf(zap, sl_hash, tx, RW_READER, 1586 &sl)) != 0) 1587 break; 1588 1589 ASSERT3U(ZAP_HASH_IDX(sl_hash, prefix_len), ==, sl_prefix); 1590 1591 /* 1592 * Check if we have a sibling and it is empty. 1593 */ 1594 if (zap_leaf_phys(sl)->l_hdr.lh_prefix_len != prefix_len || 1595 zap_leaf_phys(sl)->l_hdr.lh_nentries != 0) { 1596 zap_put_leaf(sl); 1597 break; 1598 } 1599 1600 zap_put_leaf(sl); 1601 1602 /* 1603 * If there two empty sibling, we have work to do, so 1604 * we need to lock ZAP ptrtbl as WRITER. 1605 */ 1606 if (!writer && (writer = zap_tryupgradedir(zap, tx)) == 0) { 1607 /* We failed to upgrade */ 1608 if (l != NULL) { 1609 zap_put_leaf(l); 1610 l = NULL; 1611 } 1612 1613 /* 1614 * Usually, the right way to upgrade from a READER lock 1615 * to a WRITER lock is to call zap_unlockdir() and 1616 * zap_lockdir(), but we do not have a tag. Instead, 1617 * we do it in more sophisticated way. 1618 */ 1619 rw_exit(&zap->zap_rwlock); 1620 rw_enter(&zap->zap_rwlock, RW_WRITER); 1621 dmu_buf_will_dirty(zap->zap_dbuf, tx); 1622 1623 zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift; 1624 writer = B_TRUE; 1625 } 1626 1627 /* 1628 * Here we have WRITER lock for ptrtbl. 1629 * Now, we need a WRITER lock for both siblings leaves. 1630 * Also, we have to recheck if the leaves are still siblings 1631 * and still empty. 1632 */ 1633 if (l == NULL) { 1634 /* sibling 0 */ 1635 if ((err = zap_deref_leaf(zap, (slbit ? sl_hash : hash), 1636 tx, RW_WRITER, &l)) != 0) 1637 break; 1638 1639 /* 1640 * The leaf isn't empty anymore or 1641 * it was shrunk/split while our locks were down. 1642 */ 1643 if (zap_leaf_phys(l)->l_hdr.lh_nentries != 0 || 1644 zap_leaf_phys(l)->l_hdr.lh_prefix_len != prefix_len) 1645 break; 1646 } 1647 1648 /* sibling 1 */ 1649 if ((err = zap_deref_leaf(zap, (slbit ? hash : sl_hash), tx, 1650 RW_WRITER, &sl)) != 0) 1651 break; 1652 1653 /* 1654 * The leaf isn't empty anymore or 1655 * it was shrunk/split while our locks were down. 1656 */ 1657 if (zap_leaf_phys(sl)->l_hdr.lh_nentries != 0 || 1658 zap_leaf_phys(sl)->l_hdr.lh_prefix_len != prefix_len) { 1659 zap_put_leaf(sl); 1660 break; 1661 } 1662 1663 /* If we have gotten here, we have a leaf to collapse */ 1664 uint64_t idx = (slbit ? prefix : sl_prefix) << prefix_diff; 1665 uint64_t nptrs = (1ULL << prefix_diff); 1666 uint64_t sl_blkid = sl->l_blkid; 1667 1668 /* 1669 * Set ptrtbl entries to point out to the slibling 0 blkid 1670 */ 1671 if ((err = zap_set_idx_range_to_blk(zap, idx, nptrs, l->l_blkid, 1672 tx)) != 0) { 1673 zap_put_leaf(sl); 1674 break; 1675 } 1676 1677 /* 1678 * Free sibling 1 disk block. 1679 */ 1680 int bs = FZAP_BLOCK_SHIFT(zap); 1681 if (sl_blkid == zap_f_phys(zap)->zap_freeblk - 1) 1682 trunc = B_TRUE; 1683 1684 (void) dmu_free_range(zap->zap_objset, zap->zap_object, 1685 sl_blkid << bs, 1 << bs, tx); 1686 zap_put_leaf(sl); 1687 1688 zap_f_phys(zap)->zap_num_leafs--; 1689 1690 /* 1691 * Update prefix and prefix_len. 1692 */ 1693 zap_leaf_phys(l)->l_hdr.lh_prefix >>= 1; 1694 zap_leaf_phys(l)->l_hdr.lh_prefix_len--; 1695 1696 prefix = zap_leaf_phys(l)->l_hdr.lh_prefix; 1697 prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len; 1698 } 1699 1700 if (trunc) 1701 zap_trunc(zap); 1702 1703 if (l != NULL) 1704 zap_put_leaf(l); 1705 1706 return (err); 1707 } 1708 1709 ZFS_MODULE_PARAM(zfs, , zap_iterate_prefetch, INT, ZMOD_RW, 1710 "When iterating ZAP object, prefetch it"); 1711 1712 ZFS_MODULE_PARAM(zfs, , zap_shrink_enabled, INT, ZMOD_RW, 1713 "Enable ZAP shrinking"); 1714