ufs_dirhash.c revision 133837
1/* 2 * Copyright (c) 2001, 2002 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 26/* 27 * This implements a hash-based lookup scheme for UFS directories. 28 */ 29 30#include <sys/cdefs.h> 31__FBSDID("$FreeBSD: head/sys/ufs/ufs/ufs_dirhash.c 133837 2004-08-16 10:00:44Z dwmalone $"); 32 33#include "opt_ufs.h" 34 35#ifdef UFS_DIRHASH 36 37#include <sys/param.h> 38#include <sys/systm.h> 39#include <sys/kernel.h> 40#include <sys/lock.h> 41#include <sys/mutex.h> 42#include <sys/malloc.h> 43#include <sys/fnv_hash.h> 44#include <sys/proc.h> 45#include <sys/bio.h> 46#include <sys/buf.h> 47#include <sys/vnode.h> 48#include <sys/mount.h> 49#include <sys/sysctl.h> 50#include <vm/uma.h> 51 52#include <ufs/ufs/quota.h> 53#include <ufs/ufs/inode.h> 54#include <ufs/ufs/dir.h> 55#include <ufs/ufs/dirhash.h> 56#include <ufs/ufs/extattr.h> 57#include <ufs/ufs/ufsmount.h> 58#include <ufs/ufs/ufs_extern.h> 59 60#define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1)) 61#define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1)) 62#define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 63#define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n)) 64 65static MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables"); 66 67SYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem"); 68 69static int ufs_mindirhashsize = DIRBLKSIZ * 5; 70SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW, 71 &ufs_mindirhashsize, 72 0, "minimum directory size in bytes for which to use hashed lookup"); 73static int ufs_dirhashmaxmem = 2 * 1024 * 1024; 74SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 75 0, "maximum allowed dirhash memory usage"); 76static int ufs_dirhashmem; 77SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 78 0, "current dirhash memory usage"); 79static int ufs_dirhashcheck = 0; 80SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 81 0, "enable extra sanity tests"); 82 83 84static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 85static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 86static void ufsdirhash_delslot(struct dirhash *dh, int slot); 87static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 88 doff_t offset); 89static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 90static int ufsdirhash_recycle(int wanted); 91 92static uma_zone_t ufsdirhash_zone; 93 94#define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx) 95#define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx) 96#define DIRHASH_LOCK(dh) mtx_lock(&(dh)->dh_mtx) 97#define DIRHASH_UNLOCK(dh) mtx_unlock(&(dh)->dh_mtx) 98#define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK) 99#define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr)) 100 101/* Dirhash list; recently-used entries are near the tail. */ 102static TAILQ_HEAD(, dirhash) ufsdirhash_list; 103 104/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 105static struct mtx ufsdirhash_mtx; 106 107/* 108 * Locking order: 109 * ufsdirhash_mtx 110 * dh_mtx 111 * 112 * The dh_mtx mutex should be acquired either via the inode lock, or via 113 * ufsdirhash_mtx. Only the owner of the inode may free the associated 114 * dirhash, but anything can steal its memory and set dh_hash to NULL. 115 */ 116 117/* 118 * Attempt to build up a hash table for the directory contents in 119 * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 120 */ 121int 122ufsdirhash_build(struct inode *ip) 123{ 124 struct dirhash *dh; 125 struct buf *bp = NULL; 126 struct direct *ep; 127 struct vnode *vp; 128 doff_t bmask, pos; 129 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 130 131 /* Check if we can/should use dirhash. */ 132 if (ip->i_dirhash == NULL) { 133 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode)) 134 return (-1); 135 } else { 136 /* Hash exists, but sysctls could have changed. */ 137 if (ip->i_size < ufs_mindirhashsize || 138 ufs_dirhashmem > ufs_dirhashmaxmem) { 139 ufsdirhash_free(ip); 140 return (-1); 141 } 142 /* Check if hash exists and is intact (note: unlocked read). */ 143 if (ip->i_dirhash->dh_hash != NULL) 144 return (0); 145 /* Free the old, recycled hash and build a new one. */ 146 ufsdirhash_free(ip); 147 } 148 149 /* Don't hash removed directories. */ 150 if (ip->i_effnlink == 0) 151 return (-1); 152 153 vp = ip->i_vnode; 154 /* Allocate 50% more entries than this dir size could ever need. */ 155 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 156 nslots = ip->i_size / DIRECTSIZ(1); 157 nslots = (nslots * 3 + 1) / 2; 158 narrays = howmany(nslots, DH_NBLKOFF); 159 nslots = narrays * DH_NBLKOFF; 160 dirblocks = howmany(ip->i_size, DIRBLKSIZ); 161 nblocks = (dirblocks * 3 + 1) / 2; 162 163 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 164 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 165 nblocks * sizeof(*dh->dh_blkfree); 166 DIRHASHLIST_LOCK(); 167 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 168 DIRHASHLIST_UNLOCK(); 169 if (memreqd > ufs_dirhashmaxmem / 2) 170 return (-1); 171 172 /* Try to free some space. */ 173 if (ufsdirhash_recycle(memreqd) != 0) 174 return (-1); 175 /* Enough was freed, and list has been locked. */ 176 } 177 ufs_dirhashmem += memreqd; 178 DIRHASHLIST_UNLOCK(); 179 180 /* 181 * Use non-blocking mallocs so that we will revert to a linear 182 * lookup on failure rather than potentially blocking forever. 183 */ 184 MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO); 185 if (dh == NULL) { 186 DIRHASHLIST_LOCK(); 187 ufs_dirhashmem -= memreqd; 188 DIRHASHLIST_UNLOCK(); 189 return (-1); 190 } 191 MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]), 192 M_DIRHASH, M_NOWAIT | M_ZERO); 193 MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]), 194 M_DIRHASH, M_NOWAIT); 195 if (dh->dh_hash == NULL || dh->dh_blkfree == NULL) 196 goto fail; 197 for (i = 0; i < narrays; i++) { 198 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL) 199 goto fail; 200 for (j = 0; j < DH_NBLKOFF; j++) 201 dh->dh_hash[i][j] = DIRHASH_EMPTY; 202 } 203 204 /* Initialise the hash table and block statistics. */ 205 mtx_init(&dh->dh_mtx, "dirhash", NULL, MTX_DEF); 206 dh->dh_narrays = narrays; 207 dh->dh_hlen = nslots; 208 dh->dh_nblk = nblocks; 209 dh->dh_dirblks = dirblocks; 210 for (i = 0; i < dirblocks; i++) 211 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 212 for (i = 0; i < DH_NFSTATS; i++) 213 dh->dh_firstfree[i] = -1; 214 dh->dh_firstfree[DH_NFSTATS] = 0; 215 dh->dh_seqopt = 0; 216 dh->dh_seqoff = 0; 217 dh->dh_score = DH_SCOREINIT; 218 ip->i_dirhash = dh; 219 220 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 221 pos = 0; 222 while (pos < ip->i_size) { 223 /* If necessary, get the next directory block. */ 224 if ((pos & bmask) == 0) { 225 if (bp != NULL) 226 brelse(bp); 227 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 228 goto fail; 229 } 230 231 /* Add this entry to the hash. */ 232 ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 233 if (ep->d_reclen == 0 || ep->d_reclen > 234 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 235 /* Corrupted directory. */ 236 brelse(bp); 237 goto fail; 238 } 239 if (ep->d_ino != 0) { 240 /* Add the entry (simplified ufsdirhash_add). */ 241 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 242 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 243 slot = WRAPINCR(slot, dh->dh_hlen); 244 dh->dh_hused++; 245 DH_ENTRY(dh, slot) = pos; 246 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 247 } 248 pos += ep->d_reclen; 249 } 250 251 if (bp != NULL) 252 brelse(bp); 253 DIRHASHLIST_LOCK(); 254 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 255 dh->dh_onlist = 1; 256 DIRHASHLIST_UNLOCK(); 257 return (0); 258 259fail: 260 if (dh->dh_hash != NULL) { 261 for (i = 0; i < narrays; i++) 262 if (dh->dh_hash[i] != NULL) 263 DIRHASH_BLKFREE(dh->dh_hash[i]); 264 FREE(dh->dh_hash, M_DIRHASH); 265 } 266 if (dh->dh_blkfree != NULL) 267 FREE(dh->dh_blkfree, M_DIRHASH); 268 FREE(dh, M_DIRHASH); 269 ip->i_dirhash = NULL; 270 DIRHASHLIST_LOCK(); 271 ufs_dirhashmem -= memreqd; 272 DIRHASHLIST_UNLOCK(); 273 return (-1); 274} 275 276/* 277 * Free any hash table associated with inode 'ip'. 278 */ 279void 280ufsdirhash_free(struct inode *ip) 281{ 282 struct dirhash *dh; 283 int i, mem; 284 285 if ((dh = ip->i_dirhash) == NULL) 286 return; 287 DIRHASHLIST_LOCK(); 288 DIRHASH_LOCK(dh); 289 if (dh->dh_onlist) 290 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 291 DIRHASH_UNLOCK(dh); 292 DIRHASHLIST_UNLOCK(); 293 294 /* The dirhash pointed to by 'dh' is exclusively ours now. */ 295 296 mem = sizeof(*dh); 297 if (dh->dh_hash != NULL) { 298 for (i = 0; i < dh->dh_narrays; i++) 299 DIRHASH_BLKFREE(dh->dh_hash[i]); 300 FREE(dh->dh_hash, M_DIRHASH); 301 FREE(dh->dh_blkfree, M_DIRHASH); 302 mem += dh->dh_narrays * sizeof(*dh->dh_hash) + 303 dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 304 dh->dh_nblk * sizeof(*dh->dh_blkfree); 305 } 306 mtx_destroy(&dh->dh_mtx); 307 FREE(dh, M_DIRHASH); 308 ip->i_dirhash = NULL; 309 310 DIRHASHLIST_LOCK(); 311 ufs_dirhashmem -= mem; 312 DIRHASHLIST_UNLOCK(); 313} 314 315/* 316 * Find the offset of the specified name within the given inode. 317 * Returns 0 on success, ENOENT if the entry does not exist, or 318 * EJUSTRETURN if the caller should revert to a linear search. 319 * 320 * If successful, the directory offset is stored in *offp, and a 321 * pointer to a struct buf containing the entry is stored in *bpp. If 322 * prevoffp is non-NULL, the offset of the previous entry within 323 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 324 * is the first in a block, the start of the block is used). 325 */ 326int 327ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 328 struct buf **bpp, doff_t *prevoffp) 329{ 330 struct dirhash *dh, *dh_next; 331 struct direct *dp; 332 struct vnode *vp; 333 struct buf *bp; 334 doff_t blkoff, bmask, offset, prevoff; 335 int i, slot; 336 337 if ((dh = ip->i_dirhash) == NULL) 338 return (EJUSTRETURN); 339 /* 340 * Move this dirhash towards the end of the list if it has a 341 * score higher than the next entry, and acquire the dh_mtx. 342 * Optimise the case where it's already the last by performing 343 * an unlocked read of the TAILQ_NEXT pointer. 344 * 345 * In both cases, end up holding just dh_mtx. 346 */ 347 if (TAILQ_NEXT(dh, dh_list) != NULL) { 348 DIRHASHLIST_LOCK(); 349 DIRHASH_LOCK(dh); 350 /* 351 * If the new score will be greater than that of the next 352 * entry, then move this entry past it. With both mutexes 353 * held, dh_next won't go away, but its dh_score could 354 * change; that's not important since it is just a hint. 355 */ 356 if (dh->dh_hash != NULL && 357 (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 358 dh->dh_score >= dh_next->dh_score) { 359 KASSERT(dh->dh_onlist, ("dirhash: not on list")); 360 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 361 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 362 dh_list); 363 } 364 DIRHASHLIST_UNLOCK(); 365 } else { 366 /* Already the last, though that could change as we wait. */ 367 DIRHASH_LOCK(dh); 368 } 369 if (dh->dh_hash == NULL) { 370 DIRHASH_UNLOCK(dh); 371 ufsdirhash_free(ip); 372 return (EJUSTRETURN); 373 } 374 375 /* Update the score. */ 376 if (dh->dh_score < DH_SCOREMAX) 377 dh->dh_score++; 378 379 vp = ip->i_vnode; 380 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 381 blkoff = -1; 382 bp = NULL; 383restart: 384 slot = ufsdirhash_hash(dh, name, namelen); 385 386 if (dh->dh_seqopt) { 387 /* 388 * Sequential access optimisation. dh_seqoff contains the 389 * offset of the directory entry immediately following 390 * the last entry that was looked up. Check if this offset 391 * appears in the hash chain for the name we are looking for. 392 */ 393 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 394 i = WRAPINCR(i, dh->dh_hlen)) 395 if (offset == dh->dh_seqoff) 396 break; 397 if (offset == dh->dh_seqoff) { 398 /* 399 * We found an entry with the expected offset. This 400 * is probably the entry we want, but if not, the 401 * code below will turn off seqoff and retry. 402 */ 403 slot = i; 404 } else 405 dh->dh_seqopt = 0; 406 } 407 408 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 409 slot = WRAPINCR(slot, dh->dh_hlen)) { 410 if (offset == DIRHASH_DEL) 411 continue; 412 DIRHASH_UNLOCK(dh); 413 414 if (offset < 0 || offset >= ip->i_size) 415 panic("ufsdirhash_lookup: bad offset in hash array"); 416 if ((offset & ~bmask) != blkoff) { 417 if (bp != NULL) 418 brelse(bp); 419 blkoff = offset & ~bmask; 420 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) 421 return (EJUSTRETURN); 422 } 423 dp = (struct direct *)(bp->b_data + (offset & bmask)); 424 if (dp->d_reclen == 0 || dp->d_reclen > 425 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 426 /* Corrupted directory. */ 427 brelse(bp); 428 return (EJUSTRETURN); 429 } 430 if (dp->d_namlen == namelen && 431 bcmp(dp->d_name, name, namelen) == 0) { 432 /* Found. Get the prev offset if needed. */ 433 if (prevoffp != NULL) { 434 if (offset & (DIRBLKSIZ - 1)) { 435 prevoff = ufsdirhash_getprev(dp, 436 offset); 437 if (prevoff == -1) { 438 brelse(bp); 439 return (EJUSTRETURN); 440 } 441 } else 442 prevoff = offset; 443 *prevoffp = prevoff; 444 } 445 446 /* Check for sequential access, and update offset. */ 447 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 448 dh->dh_seqopt = 1; 449 dh->dh_seqoff = offset + DIRSIZ(0, dp); 450 451 *bpp = bp; 452 *offp = offset; 453 return (0); 454 } 455 456 DIRHASH_LOCK(dh); 457 if (dh->dh_hash == NULL) { 458 DIRHASH_UNLOCK(dh); 459 if (bp != NULL) 460 brelse(bp); 461 ufsdirhash_free(ip); 462 return (EJUSTRETURN); 463 } 464 /* 465 * When the name doesn't match in the seqopt case, go back 466 * and search normally. 467 */ 468 if (dh->dh_seqopt) { 469 dh->dh_seqopt = 0; 470 goto restart; 471 } 472 } 473 DIRHASH_UNLOCK(dh); 474 if (bp != NULL) 475 brelse(bp); 476 return (ENOENT); 477} 478 479/* 480 * Find a directory block with room for 'slotneeded' bytes. Returns 481 * the offset of the directory entry that begins the free space. 482 * This will either be the offset of an existing entry that has free 483 * space at the end, or the offset of an entry with d_ino == 0 at 484 * the start of a DIRBLKSIZ block. 485 * 486 * To use the space, the caller may need to compact existing entries in 487 * the directory. The total number of bytes in all of the entries involved 488 * in the compaction is stored in *slotsize. In other words, all of 489 * the entries that must be compacted are exactly contained in the 490 * region beginning at the returned offset and spanning *slotsize bytes. 491 * 492 * Returns -1 if no space was found, indicating that the directory 493 * must be extended. 494 */ 495doff_t 496ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 497{ 498 struct direct *dp; 499 struct dirhash *dh; 500 struct buf *bp; 501 doff_t pos, slotstart; 502 int dirblock, error, freebytes, i; 503 504 if ((dh = ip->i_dirhash) == NULL) 505 return (-1); 506 DIRHASH_LOCK(dh); 507 if (dh->dh_hash == NULL) { 508 DIRHASH_UNLOCK(dh); 509 ufsdirhash_free(ip); 510 return (-1); 511 } 512 513 /* Find a directory block with the desired free space. */ 514 dirblock = -1; 515 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 516 if ((dirblock = dh->dh_firstfree[i]) != -1) 517 break; 518 if (dirblock == -1) { 519 DIRHASH_UNLOCK(dh); 520 return (-1); 521 } 522 523 KASSERT(dirblock < dh->dh_nblk && 524 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 525 ("ufsdirhash_findfree: bad stats")); 526 DIRHASH_UNLOCK(dh); 527 pos = dirblock * DIRBLKSIZ; 528 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 529 if (error) 530 return (-1); 531 532 /* Find the first entry with free space. */ 533 for (i = 0; i < DIRBLKSIZ; ) { 534 if (dp->d_reclen == 0) { 535 brelse(bp); 536 return (-1); 537 } 538 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 539 break; 540 i += dp->d_reclen; 541 dp = (struct direct *)((char *)dp + dp->d_reclen); 542 } 543 if (i > DIRBLKSIZ) { 544 brelse(bp); 545 return (-1); 546 } 547 slotstart = pos + i; 548 549 /* Find the range of entries needed to get enough space */ 550 freebytes = 0; 551 while (i < DIRBLKSIZ && freebytes < slotneeded) { 552 freebytes += dp->d_reclen; 553 if (dp->d_ino != 0) 554 freebytes -= DIRSIZ(0, dp); 555 if (dp->d_reclen == 0) { 556 brelse(bp); 557 return (-1); 558 } 559 i += dp->d_reclen; 560 dp = (struct direct *)((char *)dp + dp->d_reclen); 561 } 562 if (i > DIRBLKSIZ) { 563 brelse(bp); 564 return (-1); 565 } 566 if (freebytes < slotneeded) 567 panic("ufsdirhash_findfree: free mismatch"); 568 brelse(bp); 569 *slotsize = pos + i - slotstart; 570 return (slotstart); 571} 572 573/* 574 * Return the start of the unused space at the end of a directory, or 575 * -1 if there are no trailing unused blocks. 576 */ 577doff_t 578ufsdirhash_enduseful(struct inode *ip) 579{ 580 581 struct dirhash *dh; 582 int i; 583 584 if ((dh = ip->i_dirhash) == NULL) 585 return (-1); 586 DIRHASH_LOCK(dh); 587 if (dh->dh_hash == NULL) { 588 DIRHASH_UNLOCK(dh); 589 ufsdirhash_free(ip); 590 return (-1); 591 } 592 593 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) { 594 DIRHASH_UNLOCK(dh); 595 return (-1); 596 } 597 598 for (i = dh->dh_dirblks - 1; i >= 0; i--) 599 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 600 break; 601 DIRHASH_UNLOCK(dh); 602 return ((doff_t)(i + 1) * DIRBLKSIZ); 603} 604 605/* 606 * Insert information into the hash about a new directory entry. dirp 607 * points to a struct direct containing the entry, and offset specifies 608 * the offset of this entry. 609 */ 610void 611ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 612{ 613 struct dirhash *dh; 614 int slot; 615 616 if ((dh = ip->i_dirhash) == NULL) 617 return; 618 DIRHASH_LOCK(dh); 619 if (dh->dh_hash == NULL) { 620 DIRHASH_UNLOCK(dh); 621 ufsdirhash_free(ip); 622 return; 623 } 624 625 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 626 ("ufsdirhash_add: bad offset")); 627 /* 628 * Normal hash usage is < 66%. If the usage gets too high then 629 * remove the hash entirely and let it be rebuilt later. 630 */ 631 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 632 DIRHASH_UNLOCK(dh); 633 ufsdirhash_free(ip); 634 return; 635 } 636 637 /* Find a free hash slot (empty or deleted), and add the entry. */ 638 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 639 while (DH_ENTRY(dh, slot) >= 0) 640 slot = WRAPINCR(slot, dh->dh_hlen); 641 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 642 dh->dh_hused++; 643 DH_ENTRY(dh, slot) = offset; 644 645 /* Update the per-block summary info. */ 646 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 647 DIRHASH_UNLOCK(dh); 648} 649 650/* 651 * Remove the specified directory entry from the hash. The entry to remove 652 * is defined by the name in `dirp', which must exist at the specified 653 * `offset' within the directory. 654 */ 655void 656ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 657{ 658 struct dirhash *dh; 659 int slot; 660 661 if ((dh = ip->i_dirhash) == NULL) 662 return; 663 DIRHASH_LOCK(dh); 664 if (dh->dh_hash == NULL) { 665 DIRHASH_UNLOCK(dh); 666 ufsdirhash_free(ip); 667 return; 668 } 669 670 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 671 ("ufsdirhash_remove: bad offset")); 672 /* Find the entry */ 673 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 674 675 /* Remove the hash entry. */ 676 ufsdirhash_delslot(dh, slot); 677 678 /* Update the per-block summary info. */ 679 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 680 DIRHASH_UNLOCK(dh); 681} 682 683/* 684 * Change the offset associated with a directory entry in the hash. Used 685 * when compacting directory blocks. 686 */ 687void 688ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 689 doff_t newoff) 690{ 691 struct dirhash *dh; 692 int slot; 693 694 if ((dh = ip->i_dirhash) == NULL) 695 return; 696 DIRHASH_LOCK(dh); 697 if (dh->dh_hash == NULL) { 698 DIRHASH_UNLOCK(dh); 699 ufsdirhash_free(ip); 700 return; 701 } 702 703 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 704 newoff < dh->dh_dirblks * DIRBLKSIZ, 705 ("ufsdirhash_move: bad offset")); 706 /* Find the entry, and update the offset. */ 707 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 708 DH_ENTRY(dh, slot) = newoff; 709 DIRHASH_UNLOCK(dh); 710} 711 712/* 713 * Inform dirhash that the directory has grown by one block that 714 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 715 */ 716void 717ufsdirhash_newblk(struct inode *ip, doff_t offset) 718{ 719 struct dirhash *dh; 720 int block; 721 722 if ((dh = ip->i_dirhash) == NULL) 723 return; 724 DIRHASH_LOCK(dh); 725 if (dh->dh_hash == NULL) { 726 DIRHASH_UNLOCK(dh); 727 ufsdirhash_free(ip); 728 return; 729 } 730 731 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 732 ("ufsdirhash_newblk: bad offset")); 733 block = offset / DIRBLKSIZ; 734 if (block >= dh->dh_nblk) { 735 /* Out of space; must rebuild. */ 736 DIRHASH_UNLOCK(dh); 737 ufsdirhash_free(ip); 738 return; 739 } 740 dh->dh_dirblks = block + 1; 741 742 /* Account for the new free block. */ 743 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 744 if (dh->dh_firstfree[DH_NFSTATS] == -1) 745 dh->dh_firstfree[DH_NFSTATS] = block; 746 DIRHASH_UNLOCK(dh); 747} 748 749/* 750 * Inform dirhash that the directory is being truncated. 751 */ 752void 753ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 754{ 755 struct dirhash *dh; 756 int block, i; 757 758 if ((dh = ip->i_dirhash) == NULL) 759 return; 760 DIRHASH_LOCK(dh); 761 if (dh->dh_hash == NULL) { 762 DIRHASH_UNLOCK(dh); 763 ufsdirhash_free(ip); 764 return; 765 } 766 767 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 768 ("ufsdirhash_dirtrunc: bad offset")); 769 block = howmany(offset, DIRBLKSIZ); 770 /* 771 * If the directory shrinks to less than 1/8 of dh_nblk blocks 772 * (about 20% of its original size due to the 50% extra added in 773 * ufsdirhash_build) then free it, and let the caller rebuild 774 * if necessary. 775 */ 776 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 777 DIRHASH_UNLOCK(dh); 778 ufsdirhash_free(ip); 779 return; 780 } 781 782 /* 783 * Remove any `first free' information pertaining to the 784 * truncated blocks. All blocks we're removing should be 785 * completely unused. 786 */ 787 if (dh->dh_firstfree[DH_NFSTATS] >= block) 788 dh->dh_firstfree[DH_NFSTATS] = -1; 789 for (i = block; i < dh->dh_dirblks; i++) 790 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 791 panic("ufsdirhash_dirtrunc: blocks in use"); 792 for (i = 0; i < DH_NFSTATS; i++) 793 if (dh->dh_firstfree[i] >= block) 794 panic("ufsdirhash_dirtrunc: first free corrupt"); 795 dh->dh_dirblks = block; 796 DIRHASH_UNLOCK(dh); 797} 798 799/* 800 * Debugging function to check that the dirhash information about 801 * a directory block matches its actual contents. Panics if a mismatch 802 * is detected. 803 * 804 * On entry, `buf' should point to the start of an in-core 805 * DIRBLKSIZ-sized directory block, and `offset' should contain the 806 * offset from the start of the directory of that block. 807 */ 808void 809ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 810{ 811 struct dirhash *dh; 812 struct direct *dp; 813 int block, ffslot, i, nfree; 814 815 if (!ufs_dirhashcheck) 816 return; 817 if ((dh = ip->i_dirhash) == NULL) 818 return; 819 DIRHASH_LOCK(dh); 820 if (dh->dh_hash == NULL) { 821 DIRHASH_UNLOCK(dh); 822 ufsdirhash_free(ip); 823 return; 824 } 825 826 block = offset / DIRBLKSIZ; 827 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 828 panic("ufsdirhash_checkblock: bad offset"); 829 830 nfree = 0; 831 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 832 dp = (struct direct *)(buf + i); 833 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 834 panic("ufsdirhash_checkblock: bad dir"); 835 836 if (dp->d_ino == 0) { 837#if 0 838 /* 839 * XXX entries with d_ino == 0 should only occur 840 * at the start of a DIRBLKSIZ block. However the 841 * ufs code is tolerant of such entries at other 842 * offsets, and fsck does not fix them. 843 */ 844 if (i != 0) 845 panic("ufsdirhash_checkblock: bad dir inode"); 846#endif 847 nfree += dp->d_reclen; 848 continue; 849 } 850 851 /* Check that the entry exists (will panic if it doesn't). */ 852 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 853 854 nfree += dp->d_reclen - DIRSIZ(0, dp); 855 } 856 if (i != DIRBLKSIZ) 857 panic("ufsdirhash_checkblock: bad dir end"); 858 859 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 860 panic("ufsdirhash_checkblock: bad free count"); 861 862 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 863 for (i = 0; i <= DH_NFSTATS; i++) 864 if (dh->dh_firstfree[i] == block && i != ffslot) 865 panic("ufsdirhash_checkblock: bad first-free"); 866 if (dh->dh_firstfree[ffslot] == -1) 867 panic("ufsdirhash_checkblock: missing first-free entry"); 868 DIRHASH_UNLOCK(dh); 869} 870 871/* 872 * Hash the specified filename into a dirhash slot. 873 */ 874static int 875ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 876{ 877 u_int32_t hash; 878 879 /* 880 * We hash the name and then some other bit of data that is 881 * invariant over the dirhash's lifetime. Otherwise names 882 * differing only in the last byte are placed close to one 883 * another in the table, which is bad for linear probing. 884 */ 885 hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 886 hash = fnv_32_buf(&dh, sizeof(dh), hash); 887 return (hash % dh->dh_hlen); 888} 889 890/* 891 * Adjust the number of free bytes in the block containing `offset' 892 * by the value specified by `diff'. 893 * 894 * The caller must ensure we have exclusive access to `dh'; normally 895 * that means that dh_mtx should be held, but this is also called 896 * from ufsdirhash_build() where exclusive access can be assumed. 897 */ 898static void 899ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 900{ 901 int block, i, nfidx, ofidx; 902 903 /* Update the per-block summary info. */ 904 block = offset / DIRBLKSIZ; 905 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 906 ("dirhash bad offset")); 907 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 908 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 909 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 910 911 /* Update the `first free' list if necessary. */ 912 if (ofidx != nfidx) { 913 /* If removing, scan forward for the next block. */ 914 if (dh->dh_firstfree[ofidx] == block) { 915 for (i = block + 1; i < dh->dh_dirblks; i++) 916 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 917 break; 918 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 919 } 920 921 /* Make this the new `first free' if necessary */ 922 if (dh->dh_firstfree[nfidx] > block || 923 dh->dh_firstfree[nfidx] == -1) 924 dh->dh_firstfree[nfidx] = block; 925 } 926} 927 928/* 929 * Find the specified name which should have the specified offset. 930 * Returns a slot number, and panics on failure. 931 * 932 * `dh' must be locked on entry and remains so on return. 933 */ 934static int 935ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 936{ 937 int slot; 938 939 mtx_assert(&dh->dh_mtx, MA_OWNED); 940 941 /* Find the entry. */ 942 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 943 slot = ufsdirhash_hash(dh, name, namelen); 944 while (DH_ENTRY(dh, slot) != offset && 945 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 946 slot = WRAPINCR(slot, dh->dh_hlen); 947 if (DH_ENTRY(dh, slot) != offset) 948 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 949 950 return (slot); 951} 952 953/* 954 * Remove the entry corresponding to the specified slot from the hash array. 955 * 956 * `dh' must be locked on entry and remains so on return. 957 */ 958static void 959ufsdirhash_delslot(struct dirhash *dh, int slot) 960{ 961 int i; 962 963 mtx_assert(&dh->dh_mtx, MA_OWNED); 964 965 /* Mark the entry as deleted. */ 966 DH_ENTRY(dh, slot) = DIRHASH_DEL; 967 968 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 969 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 970 i = WRAPINCR(i, dh->dh_hlen); 971 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 972 i = WRAPDECR(i, dh->dh_hlen); 973 while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 974 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 975 dh->dh_hused--; 976 i = WRAPDECR(i, dh->dh_hlen); 977 } 978 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 979 } 980} 981 982/* 983 * Given a directory entry and its offset, find the offset of the 984 * previous entry in the same DIRBLKSIZ-sized block. Returns an 985 * offset, or -1 if there is no previous entry in the block or some 986 * other problem occurred. 987 */ 988static doff_t 989ufsdirhash_getprev(struct direct *dirp, doff_t offset) 990{ 991 struct direct *dp; 992 char *blkbuf; 993 doff_t blkoff, prevoff; 994 int entrypos, i; 995 996 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 997 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 998 blkbuf = (char *)dirp - entrypos; 999 prevoff = blkoff; 1000 1001 /* If `offset' is the start of a block, there is no previous entry. */ 1002 if (entrypos == 0) 1003 return (-1); 1004 1005 /* Scan from the start of the block until we get to the entry. */ 1006 for (i = 0; i < entrypos; i += dp->d_reclen) { 1007 dp = (struct direct *)(blkbuf + i); 1008 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 1009 return (-1); /* Corrupted directory. */ 1010 prevoff = blkoff + i; 1011 } 1012 return (prevoff); 1013} 1014 1015/* 1016 * Try to free up `wanted' bytes by stealing memory from existing 1017 * dirhashes. Returns zero with list locked if successful. 1018 */ 1019static int 1020ufsdirhash_recycle(int wanted) 1021{ 1022 struct dirhash *dh; 1023 doff_t **hash; 1024 u_int8_t *blkfree; 1025 int i, mem, narrays; 1026 1027 DIRHASHLIST_LOCK(); 1028 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 1029 /* Find a dirhash, and lock it. */ 1030 if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) { 1031 DIRHASHLIST_UNLOCK(); 1032 return (-1); 1033 } 1034 DIRHASH_LOCK(dh); 1035 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 1036 1037 /* Decrement the score; only recycle if it becomes zero. */ 1038 if (--dh->dh_score > 0) { 1039 DIRHASH_UNLOCK(dh); 1040 DIRHASHLIST_UNLOCK(); 1041 return (-1); 1042 } 1043 1044 /* Remove it from the list and detach its memory. */ 1045 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 1046 dh->dh_onlist = 0; 1047 hash = dh->dh_hash; 1048 dh->dh_hash = NULL; 1049 blkfree = dh->dh_blkfree; 1050 dh->dh_blkfree = NULL; 1051 narrays = dh->dh_narrays; 1052 mem = narrays * sizeof(*dh->dh_hash) + 1053 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 1054 dh->dh_nblk * sizeof(*dh->dh_blkfree); 1055 1056 /* Unlock everything, free the detached memory. */ 1057 DIRHASH_UNLOCK(dh); 1058 DIRHASHLIST_UNLOCK(); 1059 for (i = 0; i < narrays; i++) 1060 DIRHASH_BLKFREE(hash[i]); 1061 FREE(hash, M_DIRHASH); 1062 FREE(blkfree, M_DIRHASH); 1063 1064 /* Account for the returned memory, and repeat if necessary. */ 1065 DIRHASHLIST_LOCK(); 1066 ufs_dirhashmem -= mem; 1067 } 1068 /* Success; return with list locked. */ 1069 return (0); 1070} 1071 1072 1073void 1074ufsdirhash_init() 1075{ 1076 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t), 1077 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1078 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF); 1079 TAILQ_INIT(&ufsdirhash_list); 1080} 1081 1082void 1083ufsdirhash_uninit() 1084{ 1085 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit")); 1086 uma_zdestroy(ufsdirhash_zone); 1087 mtx_destroy(&ufsdirhash_mtx); 1088} 1089 1090#endif /* UFS_DIRHASH */ 1091