1 /* 2 * Copyright (c) 2001 Ian Dowse. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 * $FreeBSD: src/sys/ufs/ufs/ufs_dirhash.c,v 1.3.2.6 2002/04/10 21:41:14 dwmalone Exp $ 26 * $DragonFly: src/sys/vfs/ufs/ufs_dirhash.c,v 1.2 2003/06/17 04:28:59 dillon Exp $ 27 */ 28 /* 29 * This implements a hash-based lookup scheme for UFS directories. 30 */ 31 32 #include "opt_ufs.h" 33 34 #ifdef UFS_DIRHASH 35 36 #include <sys/param.h> 37 #include <sys/systm.h> 38 #include <sys/kernel.h> 39 #include <sys/malloc.h> 40 #include <sys/fnv_hash.h> 41 #include <sys/proc.h> 42 #include <sys/buf.h> 43 #include <sys/vnode.h> 44 #include <sys/mount.h> 45 #include <sys/sysctl.h> 46 #include <vm/vm_zone.h> 47 48 #include <ufs/ufs/quota.h> 49 #include <ufs/ufs/inode.h> 50 #include <ufs/ufs/dir.h> 51 #include <ufs/ufs/dirhash.h> 52 #include <ufs/ufs/ufsmount.h> 53 #include <ufs/ufs/ufs_extern.h> 54 55 #define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1)) 56 #define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1)) 57 #define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 58 #define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n)) 59 60 static MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables"); 61 62 SYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem"); 63 64 static int ufs_mindirhashsize = DIRBLKSIZ * 5; 65 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW, 66 &ufs_mindirhashsize, 67 0, "minimum directory size in bytes for which to use hashed lookup"); 68 static int ufs_dirhashmaxmem = 2 * 1024 * 1024; 69 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 70 0, "maximum allowed dirhash memory usage"); 71 static int ufs_dirhashmem; 72 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 73 0, "current dirhash memory usage"); 74 static int ufs_dirhashcheck = 0; 75 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 76 0, "enable extra sanity tests"); 77 78 79 static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 80 static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 81 static void ufsdirhash_delslot(struct dirhash *dh, int slot); 82 static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 83 doff_t offset); 84 static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 85 static void ufsdirhash_init(void); 86 static int ufsdirhash_recycle(int wanted); 87 88 static vm_zone_t ufsdirhash_zone; 89 90 /* Dirhash list; recently-used entries are near the tail. */ 91 static TAILQ_HEAD(, dirhash) ufsdirhash_list; 92 93 /* 94 * Attempt to build up a hash table for the directory contents in 95 * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 96 */ 97 int 98 ufsdirhash_build(struct inode *ip) 99 { 100 struct dirhash *dh; 101 struct buf *bp = NULL; 102 struct direct *ep; 103 struct vnode *vp; 104 doff_t bmask, pos; 105 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 106 107 /* Check if we can/should use dirhash. */ 108 if (ip->i_dirhash == NULL) { 109 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode)) 110 return (-1); 111 } else { 112 /* Hash exists, but sysctls could have changed. */ 113 if (ip->i_size < ufs_mindirhashsize || 114 ufs_dirhashmem > ufs_dirhashmaxmem) { 115 ufsdirhash_free(ip); 116 return (-1); 117 } 118 /* Check if hash exists and is intact (note: unlocked read). */ 119 if (ip->i_dirhash->dh_hash != NULL) 120 return (0); 121 /* Free the old, recycled hash and build a new one. */ 122 ufsdirhash_free(ip); 123 } 124 125 /* Don't hash removed directories. */ 126 if (ip->i_effnlink == 0) 127 return (-1); 128 129 vp = ip->i_vnode; 130 /* Allocate 50% more entries than this dir size could ever need. */ 131 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 132 nslots = ip->i_size / DIRECTSIZ(1); 133 nslots = (nslots * 3 + 1) / 2; 134 narrays = howmany(nslots, DH_NBLKOFF); 135 nslots = narrays * DH_NBLKOFF; 136 dirblocks = howmany(ip->i_size, DIRBLKSIZ); 137 nblocks = (dirblocks * 3 + 1) / 2; 138 139 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 140 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 141 nblocks * sizeof(*dh->dh_blkfree); 142 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 143 if (memreqd > ufs_dirhashmaxmem / 2) 144 return (-1); 145 146 /* Try to free some space. */ 147 if (ufsdirhash_recycle(memreqd) != 0) 148 return (-1); 149 /* Enough was freed. */ 150 } 151 ufs_dirhashmem += memreqd; 152 153 /* 154 * Use non-blocking mallocs so that we will revert to a linear 155 * lookup on failure rather than potentially blocking forever. 156 */ 157 MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO); 158 if (dh == NULL) 159 return (-1); 160 MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]), 161 M_DIRHASH, M_NOWAIT | M_ZERO); 162 MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]), 163 M_DIRHASH, M_NOWAIT); 164 if (dh->dh_hash == NULL || dh->dh_blkfree == NULL) 165 goto fail; 166 for (i = 0; i < narrays; i++) { 167 if ((dh->dh_hash[i] = zalloc(ufsdirhash_zone)) == NULL) 168 goto fail; 169 for (j = 0; j < DH_NBLKOFF; j++) 170 dh->dh_hash[i][j] = DIRHASH_EMPTY; 171 } 172 173 /* Initialise the hash table and block statistics. */ 174 dh->dh_narrays = narrays; 175 dh->dh_hlen = nslots; 176 dh->dh_nblk = nblocks; 177 dh->dh_dirblks = dirblocks; 178 for (i = 0; i < dirblocks; i++) 179 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 180 for (i = 0; i < DH_NFSTATS; i++) 181 dh->dh_firstfree[i] = -1; 182 dh->dh_firstfree[DH_NFSTATS] = 0; 183 dh->dh_seqopt = 0; 184 dh->dh_seqoff = 0; 185 dh->dh_score = DH_SCOREINIT; 186 ip->i_dirhash = dh; 187 188 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 189 pos = 0; 190 while (pos < ip->i_size) { 191 /* If necessary, get the next directory block. */ 192 if ((pos & bmask) == 0) { 193 if (bp != NULL) 194 brelse(bp); 195 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 196 goto fail; 197 } 198 199 /* Add this entry to the hash. */ 200 ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 201 if (ep->d_reclen == 0 || ep->d_reclen > 202 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 203 /* Corrupted directory. */ 204 brelse(bp); 205 goto fail; 206 } 207 if (ep->d_ino != 0) { 208 /* Add the entry (simplified ufsdirhash_add). */ 209 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 210 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 211 slot = WRAPINCR(slot, dh->dh_hlen); 212 dh->dh_hused++; 213 DH_ENTRY(dh, slot) = pos; 214 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 215 } 216 pos += ep->d_reclen; 217 } 218 219 if (bp != NULL) 220 brelse(bp); 221 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 222 dh->dh_onlist = 1; 223 return (0); 224 225 fail: 226 if (dh->dh_hash != NULL) { 227 for (i = 0; i < narrays; i++) 228 if (dh->dh_hash[i] != NULL) 229 zfree(ufsdirhash_zone, dh->dh_hash[i]); 230 FREE(dh->dh_hash, M_DIRHASH); 231 } 232 if (dh->dh_blkfree != NULL) 233 FREE(dh->dh_blkfree, M_DIRHASH); 234 FREE(dh, M_DIRHASH); 235 ip->i_dirhash = NULL; 236 ufs_dirhashmem -= memreqd; 237 return (-1); 238 } 239 240 /* 241 * Free any hash table associated with inode 'ip'. 242 */ 243 void 244 ufsdirhash_free(struct inode *ip) 245 { 246 struct dirhash *dh; 247 int i, mem; 248 249 if ((dh = ip->i_dirhash) == NULL) 250 return; 251 if (dh->dh_onlist) 252 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 253 254 /* The dirhash pointed to by 'dh' is exclusively ours now. */ 255 256 mem = sizeof(*dh); 257 if (dh->dh_hash != NULL) { 258 for (i = 0; i < dh->dh_narrays; i++) 259 zfree(ufsdirhash_zone, dh->dh_hash[i]); 260 FREE(dh->dh_hash, M_DIRHASH); 261 FREE(dh->dh_blkfree, M_DIRHASH); 262 mem += dh->dh_narrays * sizeof(*dh->dh_hash) + 263 dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 264 dh->dh_nblk * sizeof(*dh->dh_blkfree); 265 } 266 FREE(dh, M_DIRHASH); 267 ip->i_dirhash = NULL; 268 269 ufs_dirhashmem -= mem; 270 } 271 272 /* 273 * Find the offset of the specified name within the given inode. 274 * Returns 0 on success, ENOENT if the entry does not exist, or 275 * EJUSTRETURN if the caller should revert to a linear search. 276 * 277 * If successful, the directory offset is stored in *offp, and a 278 * pointer to a struct buf containing the entry is stored in *bpp. If 279 * prevoffp is non-NULL, the offset of the previous entry within 280 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 281 * is the first in a block, the start of the block is used). 282 */ 283 int 284 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 285 struct buf **bpp, doff_t *prevoffp) 286 { 287 struct dirhash *dh, *dh_next; 288 struct direct *dp; 289 struct vnode *vp; 290 struct buf *bp; 291 doff_t blkoff, bmask, offset, prevoff; 292 int i, slot; 293 294 if ((dh = ip->i_dirhash) == NULL) 295 return (EJUSTRETURN); 296 /* 297 * Move this dirhash towards the end of the list if it has a 298 * score higher than the next entry. 299 * Optimise the case where it's already the last by performing 300 * an unlocked read of the TAILQ_NEXT pointer. 301 */ 302 if (TAILQ_NEXT(dh, dh_list) != NULL) { 303 /* 304 * If the new score will be greater than that of the next 305 * entry, then move this entry past it. With both mutexes 306 * held, dh_next won't go away, but its dh_score could 307 * change; that's not important since it is just a hint. 308 */ 309 if (dh->dh_hash != NULL && 310 (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 311 dh->dh_score >= dh_next->dh_score) { 312 KASSERT(dh->dh_onlist, ("dirhash: not on list")); 313 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 314 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 315 dh_list); 316 } 317 } 318 if (dh->dh_hash == NULL) { 319 ufsdirhash_free(ip); 320 return (EJUSTRETURN); 321 } 322 323 /* Update the score. */ 324 if (dh->dh_score < DH_SCOREMAX) 325 dh->dh_score++; 326 327 vp = ip->i_vnode; 328 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 329 blkoff = -1; 330 bp = NULL; 331 restart: 332 slot = ufsdirhash_hash(dh, name, namelen); 333 334 if (dh->dh_seqopt) { 335 /* 336 * Sequential access optimisation. dh_seqoff contains the 337 * offset of the directory entry immediately following 338 * the last entry that was looked up. Check if this offset 339 * appears in the hash chain for the name we are looking for. 340 */ 341 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 342 i = WRAPINCR(i, dh->dh_hlen)) 343 if (offset == dh->dh_seqoff) 344 break; 345 if (offset == dh->dh_seqoff) { 346 /* 347 * We found an entry with the expected offset. This 348 * is probably the entry we want, but if not, the 349 * code below will turn off seqoff and retry. 350 */ 351 slot = i; 352 } else 353 dh->dh_seqopt = 0; 354 } 355 356 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 357 slot = WRAPINCR(slot, dh->dh_hlen)) { 358 if (offset == DIRHASH_DEL) 359 continue; 360 361 if (offset < 0 || offset >= ip->i_size) 362 panic("ufsdirhash_lookup: bad offset in hash array"); 363 if ((offset & ~bmask) != blkoff) { 364 if (bp != NULL) 365 brelse(bp); 366 blkoff = offset & ~bmask; 367 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) 368 return (EJUSTRETURN); 369 } 370 dp = (struct direct *)(bp->b_data + (offset & bmask)); 371 if (dp->d_reclen == 0 || dp->d_reclen > 372 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 373 /* Corrupted directory. */ 374 brelse(bp); 375 return (EJUSTRETURN); 376 } 377 if (dp->d_namlen == namelen && 378 bcmp(dp->d_name, name, namelen) == 0) { 379 /* Found. Get the prev offset if needed. */ 380 if (prevoffp != NULL) { 381 if (offset & (DIRBLKSIZ - 1)) { 382 prevoff = ufsdirhash_getprev(dp, 383 offset); 384 if (prevoff == -1) { 385 brelse(bp); 386 return (EJUSTRETURN); 387 } 388 } else 389 prevoff = offset; 390 *prevoffp = prevoff; 391 } 392 393 /* Check for sequential access, and update offset. */ 394 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 395 dh->dh_seqopt = 1; 396 dh->dh_seqoff = offset + DIRSIZ(0, dp); 397 398 *bpp = bp; 399 *offp = offset; 400 return (0); 401 } 402 403 if (dh->dh_hash == NULL) { 404 if (bp != NULL) 405 brelse(bp); 406 ufsdirhash_free(ip); 407 return (EJUSTRETURN); 408 } 409 /* 410 * When the name doesn't match in the seqopt case, go back 411 * and search normally. 412 */ 413 if (dh->dh_seqopt) { 414 dh->dh_seqopt = 0; 415 goto restart; 416 } 417 } 418 if (bp != NULL) 419 brelse(bp); 420 return (ENOENT); 421 } 422 423 /* 424 * Find a directory block with room for 'slotneeded' bytes. Returns 425 * the offset of the directory entry that begins the free space. 426 * This will either be the offset of an existing entry that has free 427 * space at the end, or the offset of an entry with d_ino == 0 at 428 * the start of a DIRBLKSIZ block. 429 * 430 * To use the space, the caller may need to compact existing entries in 431 * the directory. The total number of bytes in all of the entries involved 432 * in the compaction is stored in *slotsize. In other words, all of 433 * the entries that must be compacted are exactly contained in the 434 * region beginning at the returned offset and spanning *slotsize bytes. 435 * 436 * Returns -1 if no space was found, indicating that the directory 437 * must be extended. 438 */ 439 doff_t 440 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 441 { 442 struct direct *dp; 443 struct dirhash *dh; 444 struct buf *bp; 445 doff_t pos, slotstart; 446 int dirblock, error, freebytes, i; 447 448 if ((dh = ip->i_dirhash) == NULL) 449 return (-1); 450 if (dh->dh_hash == NULL) { 451 ufsdirhash_free(ip); 452 return (-1); 453 } 454 455 /* Find a directory block with the desired free space. */ 456 dirblock = -1; 457 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 458 if ((dirblock = dh->dh_firstfree[i]) != -1) 459 break; 460 if (dirblock == -1) { 461 return (-1); 462 } 463 464 KASSERT(dirblock < dh->dh_nblk && 465 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 466 ("ufsdirhash_findfree: bad stats")); 467 pos = dirblock * DIRBLKSIZ; 468 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 469 if (error) 470 return (-1); 471 472 /* Find the first entry with free space. */ 473 for (i = 0; i < DIRBLKSIZ; ) { 474 if (dp->d_reclen == 0) { 475 brelse(bp); 476 return (-1); 477 } 478 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 479 break; 480 i += dp->d_reclen; 481 dp = (struct direct *)((char *)dp + dp->d_reclen); 482 } 483 if (i > DIRBLKSIZ) { 484 brelse(bp); 485 return (-1); 486 } 487 slotstart = pos + i; 488 489 /* Find the range of entries needed to get enough space */ 490 freebytes = 0; 491 while (i < DIRBLKSIZ && freebytes < slotneeded) { 492 freebytes += dp->d_reclen; 493 if (dp->d_ino != 0) 494 freebytes -= DIRSIZ(0, dp); 495 if (dp->d_reclen == 0) { 496 brelse(bp); 497 return (-1); 498 } 499 i += dp->d_reclen; 500 dp = (struct direct *)((char *)dp + dp->d_reclen); 501 } 502 if (i > DIRBLKSIZ) { 503 brelse(bp); 504 return (-1); 505 } 506 if (freebytes < slotneeded) 507 panic("ufsdirhash_findfree: free mismatch"); 508 brelse(bp); 509 *slotsize = pos + i - slotstart; 510 return (slotstart); 511 } 512 513 /* 514 * Return the start of the unused space at the end of a directory, or 515 * -1 if there are no trailing unused blocks. 516 */ 517 doff_t 518 ufsdirhash_enduseful(struct inode *ip) 519 { 520 521 struct dirhash *dh; 522 int i; 523 524 if ((dh = ip->i_dirhash) == NULL) 525 return (-1); 526 if (dh->dh_hash == NULL) { 527 ufsdirhash_free(ip); 528 return (-1); 529 } 530 531 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) { 532 return (-1); 533 } 534 535 for (i = dh->dh_dirblks - 1; i >= 0; i--) 536 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 537 break; 538 return ((doff_t)(i + 1) * DIRBLKSIZ); 539 } 540 541 /* 542 * Insert information into the hash about a new directory entry. dirp 543 * points to a struct direct containing the entry, and offset specifies 544 * the offset of this entry. 545 */ 546 void 547 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 548 { 549 struct dirhash *dh; 550 int slot; 551 552 if ((dh = ip->i_dirhash) == NULL) 553 return; 554 if (dh->dh_hash == NULL) { 555 ufsdirhash_free(ip); 556 return; 557 } 558 559 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 560 ("ufsdirhash_add: bad offset")); 561 /* 562 * Normal hash usage is < 66%. If the usage gets too high then 563 * remove the hash entirely and let it be rebuilt later. 564 */ 565 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 566 ufsdirhash_free(ip); 567 return; 568 } 569 570 /* Find a free hash slot (empty or deleted), and add the entry. */ 571 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 572 while (DH_ENTRY(dh, slot) >= 0) 573 slot = WRAPINCR(slot, dh->dh_hlen); 574 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 575 dh->dh_hused++; 576 DH_ENTRY(dh, slot) = offset; 577 578 /* Update the per-block summary info. */ 579 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 580 } 581 582 /* 583 * Remove the specified directory entry from the hash. The entry to remove 584 * is defined by the name in `dirp', which must exist at the specified 585 * `offset' within the directory. 586 */ 587 void 588 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 589 { 590 struct dirhash *dh; 591 int slot; 592 593 if ((dh = ip->i_dirhash) == NULL) 594 return; 595 if (dh->dh_hash == NULL) { 596 ufsdirhash_free(ip); 597 return; 598 } 599 600 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 601 ("ufsdirhash_remove: bad offset")); 602 /* Find the entry */ 603 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 604 605 /* Remove the hash entry. */ 606 ufsdirhash_delslot(dh, slot); 607 608 /* Update the per-block summary info. */ 609 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 610 } 611 612 /* 613 * Change the offset associated with a directory entry in the hash. Used 614 * when compacting directory blocks. 615 */ 616 void 617 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 618 doff_t newoff) 619 { 620 struct dirhash *dh; 621 int slot; 622 623 if ((dh = ip->i_dirhash) == NULL) 624 return; 625 if (dh->dh_hash == NULL) { 626 ufsdirhash_free(ip); 627 return; 628 } 629 630 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 631 newoff < dh->dh_dirblks * DIRBLKSIZ, 632 ("ufsdirhash_move: bad offset")); 633 /* Find the entry, and update the offset. */ 634 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 635 DH_ENTRY(dh, slot) = newoff; 636 } 637 638 /* 639 * Inform dirhash that the directory has grown by one block that 640 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 641 */ 642 void 643 ufsdirhash_newblk(struct inode *ip, doff_t offset) 644 { 645 struct dirhash *dh; 646 int block; 647 648 if ((dh = ip->i_dirhash) == NULL) 649 return; 650 if (dh->dh_hash == NULL) { 651 ufsdirhash_free(ip); 652 return; 653 } 654 655 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 656 ("ufsdirhash_newblk: bad offset")); 657 block = offset / DIRBLKSIZ; 658 if (block >= dh->dh_nblk) { 659 /* Out of space; must rebuild. */ 660 ufsdirhash_free(ip); 661 return; 662 } 663 dh->dh_dirblks = block + 1; 664 665 /* Account for the new free block. */ 666 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 667 if (dh->dh_firstfree[DH_NFSTATS] == -1) 668 dh->dh_firstfree[DH_NFSTATS] = block; 669 } 670 671 /* 672 * Inform dirhash that the directory is being truncated. 673 */ 674 void 675 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 676 { 677 struct dirhash *dh; 678 int block, i; 679 680 if ((dh = ip->i_dirhash) == NULL) 681 return; 682 if (dh->dh_hash == NULL) { 683 ufsdirhash_free(ip); 684 return; 685 } 686 687 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 688 ("ufsdirhash_dirtrunc: bad offset")); 689 block = howmany(offset, DIRBLKSIZ); 690 /* 691 * If the directory shrinks to less than 1/8 of dh_nblk blocks 692 * (about 20% of its original size due to the 50% extra added in 693 * ufsdirhash_build) then free it, and let the caller rebuild 694 * if necessary. 695 */ 696 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 697 ufsdirhash_free(ip); 698 return; 699 } 700 701 /* 702 * Remove any `first free' information pertaining to the 703 * truncated blocks. All blocks we're removing should be 704 * completely unused. 705 */ 706 if (dh->dh_firstfree[DH_NFSTATS] >= block) 707 dh->dh_firstfree[DH_NFSTATS] = -1; 708 for (i = block; i < dh->dh_dirblks; i++) 709 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 710 panic("ufsdirhash_dirtrunc: blocks in use"); 711 for (i = 0; i < DH_NFSTATS; i++) 712 if (dh->dh_firstfree[i] >= block) 713 panic("ufsdirhash_dirtrunc: first free corrupt"); 714 dh->dh_dirblks = block; 715 } 716 717 /* 718 * Debugging function to check that the dirhash information about 719 * a directory block matches its actual contents. Panics if a mismatch 720 * is detected. 721 * 722 * On entry, `buf' should point to the start of an in-core 723 * DIRBLKSIZ-sized directory block, and `offset' should contain the 724 * offset from the start of the directory of that block. 725 */ 726 void 727 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 728 { 729 struct dirhash *dh; 730 struct direct *dp; 731 int block, ffslot, i, nfree; 732 733 if (!ufs_dirhashcheck) 734 return; 735 if ((dh = ip->i_dirhash) == NULL) 736 return; 737 if (dh->dh_hash == NULL) { 738 ufsdirhash_free(ip); 739 return; 740 } 741 742 block = offset / DIRBLKSIZ; 743 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 744 panic("ufsdirhash_checkblock: bad offset"); 745 746 nfree = 0; 747 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 748 dp = (struct direct *)(buf + i); 749 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 750 panic("ufsdirhash_checkblock: bad dir"); 751 752 if (dp->d_ino == 0) { 753 #if 0 754 /* 755 * XXX entries with d_ino == 0 should only occur 756 * at the start of a DIRBLKSIZ block. However the 757 * ufs code is tolerant of such entries at other 758 * offsets, and fsck does not fix them. 759 */ 760 if (i != 0) 761 panic("ufsdirhash_checkblock: bad dir inode"); 762 #endif 763 nfree += dp->d_reclen; 764 continue; 765 } 766 767 /* Check that the entry exists (will panic if it doesn't). */ 768 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 769 770 nfree += dp->d_reclen - DIRSIZ(0, dp); 771 } 772 if (i != DIRBLKSIZ) 773 panic("ufsdirhash_checkblock: bad dir end"); 774 775 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 776 panic("ufsdirhash_checkblock: bad free count"); 777 778 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 779 for (i = 0; i <= DH_NFSTATS; i++) 780 if (dh->dh_firstfree[i] == block && i != ffslot) 781 panic("ufsdirhash_checkblock: bad first-free"); 782 if (dh->dh_firstfree[ffslot] == -1) 783 panic("ufsdirhash_checkblock: missing first-free entry"); 784 } 785 786 /* 787 * Hash the specified filename into a dirhash slot. 788 */ 789 static int 790 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 791 { 792 u_int32_t hash; 793 794 /* 795 * We hash the name and then some ofther bit of data which is 796 * invarient over the dirhash's lifetime. Otherwise names 797 * differing only in the last byte are placed close to one 798 * another in the table, which is bad for linear probing. 799 */ 800 hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 801 hash = fnv_32_buf(dh, sizeof(dh), hash); 802 return (hash % dh->dh_hlen); 803 } 804 805 /* 806 * Adjust the number of free bytes in the block containing `offset' 807 * by the value specified by `diff'. 808 * 809 * The caller must ensure we have exclusive access to `dh'. 810 */ 811 static void 812 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 813 { 814 int block, i, nfidx, ofidx; 815 816 /* Update the per-block summary info. */ 817 block = offset / DIRBLKSIZ; 818 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 819 ("dirhash bad offset")); 820 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 821 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 822 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 823 824 /* Update the `first free' list if necessary. */ 825 if (ofidx != nfidx) { 826 /* If removing, scan forward for the next block. */ 827 if (dh->dh_firstfree[ofidx] == block) { 828 for (i = block + 1; i < dh->dh_dirblks; i++) 829 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 830 break; 831 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 832 } 833 834 /* Make this the new `first free' if necessary */ 835 if (dh->dh_firstfree[nfidx] > block || 836 dh->dh_firstfree[nfidx] == -1) 837 dh->dh_firstfree[nfidx] = block; 838 } 839 } 840 841 /* 842 * Find the specified name which should have the specified offset. 843 * Returns a slot number, and panics on failure. 844 * 845 * `dh' must be locked on entry and remains so on return. 846 */ 847 static int 848 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 849 { 850 int slot; 851 852 /* Find the entry. */ 853 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 854 slot = ufsdirhash_hash(dh, name, namelen); 855 while (DH_ENTRY(dh, slot) != offset && 856 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 857 slot = WRAPINCR(slot, dh->dh_hlen); 858 if (DH_ENTRY(dh, slot) != offset) 859 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 860 861 return (slot); 862 } 863 864 /* 865 * Remove the entry corresponding to the specified slot from the hash array. 866 * 867 * `dh' must be locked on entry and remains so on return. 868 */ 869 static void 870 ufsdirhash_delslot(struct dirhash *dh, int slot) 871 { 872 int i; 873 874 /* Mark the entry as deleted. */ 875 DH_ENTRY(dh, slot) = DIRHASH_DEL; 876 877 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 878 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 879 i = WRAPINCR(i, dh->dh_hlen); 880 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 881 i = WRAPDECR(i, dh->dh_hlen); 882 while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 883 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 884 dh->dh_hused--; 885 i = WRAPDECR(i, dh->dh_hlen); 886 } 887 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 888 } 889 } 890 891 /* 892 * Given a directory entry and its offset, find the offset of the 893 * previous entry in the same DIRBLKSIZ-sized block. Returns an 894 * offset, or -1 if there is no previous entry in the block or some 895 * other problem occurred. 896 */ 897 static doff_t 898 ufsdirhash_getprev(struct direct *dirp, doff_t offset) 899 { 900 struct direct *dp; 901 char *blkbuf; 902 doff_t blkoff, prevoff; 903 int entrypos, i; 904 905 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 906 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 907 blkbuf = (char *)dirp - entrypos; 908 prevoff = blkoff; 909 910 /* If `offset' is the start of a block, there is no previous entry. */ 911 if (entrypos == 0) 912 return (-1); 913 914 /* Scan from the start of the block until we get to the entry. */ 915 for (i = 0; i < entrypos; i += dp->d_reclen) { 916 dp = (struct direct *)(blkbuf + i); 917 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 918 return (-1); /* Corrupted directory. */ 919 prevoff = blkoff + i; 920 } 921 return (prevoff); 922 } 923 924 /* 925 * Try to free up `wanted' bytes by stealing memory from existing 926 * dirhashes. Returns zero if successful. 927 */ 928 static int 929 ufsdirhash_recycle(int wanted) 930 { 931 struct dirhash *dh; 932 doff_t **hash; 933 u_int8_t *blkfree; 934 int i, mem, narrays; 935 936 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 937 /* Find a dirhash, and lock it. */ 938 if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) { 939 return (-1); 940 } 941 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 942 943 /* Decrement the score; only recycle if it becomes zero. */ 944 if (--dh->dh_score > 0) { 945 return (-1); 946 } 947 948 /* Remove it from the list and detach its memory. */ 949 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 950 dh->dh_onlist = 0; 951 hash = dh->dh_hash; 952 dh->dh_hash = NULL; 953 blkfree = dh->dh_blkfree; 954 dh->dh_blkfree = NULL; 955 narrays = dh->dh_narrays; 956 mem = narrays * sizeof(*dh->dh_hash) + 957 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 958 dh->dh_nblk * sizeof(*dh->dh_blkfree); 959 960 /* Free the detached memory. */ 961 for (i = 0; i < narrays; i++) 962 zfree(ufsdirhash_zone, hash[i]); 963 FREE(hash, M_DIRHASH); 964 FREE(blkfree, M_DIRHASH); 965 966 /* Account for the returned memory, and repeat if necessary. */ 967 ufs_dirhashmem -= mem; 968 } 969 /* Success. */ 970 return (0); 971 } 972 973 974 static void 975 ufsdirhash_init() 976 { 977 ufsdirhash_zone = zinit("DIRHASH", DH_NBLKOFF * sizeof(daddr_t), 0, 978 0, 1); 979 TAILQ_INIT(&ufsdirhash_list); 980 } 981 SYSINIT(ufsdirhash, SI_SUB_PSEUDO, SI_ORDER_ANY, ufsdirhash_init, NULL) 982 983 984 #endif /* UFS_DIRHASH */ 985