ufs_dirhash.c revision 183080
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 183080 2008-09-16 16:23:56Z jhb $"); 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/refcount.h> 50#include <sys/sysctl.h> 51#include <sys/sx.h> 52#include <vm/uma.h> 53 54#include <ufs/ufs/quota.h> 55#include <ufs/ufs/inode.h> 56#include <ufs/ufs/dir.h> 57#include <ufs/ufs/dirhash.h> 58#include <ufs/ufs/extattr.h> 59#include <ufs/ufs/ufsmount.h> 60#include <ufs/ufs/ufs_extern.h> 61 62#define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1)) 63#define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1)) 64#define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 65#define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n)) 66 67static MALLOC_DEFINE(M_DIRHASH, "ufs_dirhash", "UFS directory hash tables"); 68 69static SYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem"); 70 71static int ufs_mindirhashsize = DIRBLKSIZ * 5; 72SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW, 73 &ufs_mindirhashsize, 74 0, "minimum directory size in bytes for which to use hashed lookup"); 75static int ufs_dirhashmaxmem = 2 * 1024 * 1024; 76SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 77 0, "maximum allowed dirhash memory usage"); 78static int ufs_dirhashmem; 79SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 80 0, "current dirhash memory usage"); 81static int ufs_dirhashcheck = 0; 82SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 83 0, "enable extra sanity tests"); 84 85 86static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 87static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 88static void ufsdirhash_delslot(struct dirhash *dh, int slot); 89static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 90 doff_t offset); 91static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 92static int ufsdirhash_recycle(int wanted); 93static void ufsdirhash_free_locked(struct inode *ip); 94 95static uma_zone_t ufsdirhash_zone; 96 97#define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx) 98#define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx) 99#define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK) 100#define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr)) 101#define DIRHASH_ASSERT_LOCKED(dh) \ 102 sx_assert(&(dh)->dh_lock, SA_LOCKED) 103 104/* Dirhash list; recently-used entries are near the tail. */ 105static TAILQ_HEAD(, dirhash) ufsdirhash_list; 106 107/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 108static struct mtx ufsdirhash_mtx; 109 110/* 111 * Locking: 112 * 113 * The relationship between inode and dirhash is protected either by an 114 * exclusive vnode lock or the vnode interlock where a shared vnode lock 115 * may be used. The dirhash_mtx is acquired after the dirhash lock. To 116 * handle teardown races, code wishing to lock the dirhash for an inode 117 * when using a shared vnode lock must obtain a private reference on the 118 * dirhash while holding the vnode interlock. They can drop it once they 119 * have obtained the dirhash lock and verified that the dirhash wasn't 120 * recycled while they waited for the dirhash lock. 121 * 122 * ufsdirhash_build() acquires a shared lock on the dirhash when it is 123 * successful. This lock is released after a call to ufsdirhash_lookup(). 124 * 125 * Functions requiring exclusive access use ufsdirhash_acquire() which may 126 * free a dirhash structure that was recycled by ufsdirhash_recycle(). 127 * 128 * The dirhash lock may be held across io operations. 129 */ 130 131static void 132ufsdirhash_hold(struct dirhash *dh) 133{ 134 135 refcount_acquire(&dh->dh_refcount); 136} 137 138static void 139ufsdirhash_drop(struct dirhash *dh) 140{ 141 142 if (refcount_release(&dh->dh_refcount)) { 143 sx_destroy(&dh->dh_lock); 144 free(dh, M_DIRHASH); 145 } 146} 147 148/* 149 * Release the lock on a dirhash. 150 */ 151static void 152ufsdirhash_release(struct dirhash *dh) 153{ 154 155 sx_unlock(&dh->dh_lock); 156} 157 158/* 159 * Either acquire an existing hash locked shared or create a new hash and 160 * return it exclusively locked. May return NULL if the allocation fails. 161 * 162 * The vnode interlock is used to protect the i_dirhash pointer from 163 * simultaneous access while only a shared vnode lock is held. 164 */ 165static struct dirhash * 166ufsdirhash_create(struct inode *ip) 167{ 168 struct dirhash *ndh; 169 struct dirhash *dh; 170 struct vnode *vp; 171 int error; 172 173 error = 0; 174 ndh = dh = NULL; 175 vp = ip->i_vnode; 176 for (;;) { 177 /* Racy check for i_dirhash to prefetch an dirhash structure. */ 178 if (ip->i_dirhash == NULL && ndh == NULL) { 179 MALLOC(ndh, struct dirhash *, sizeof *dh, M_DIRHASH, 180 M_NOWAIT | M_ZERO); 181 if (ndh == NULL) 182 return (NULL); 183 refcount_init(&ndh->dh_refcount, 1); 184 sx_init(&ndh->dh_lock, "dirhash"); 185 sx_xlock(&ndh->dh_lock); 186 } 187 /* 188 * Check i_dirhash. If it's NULL just try to use a 189 * preallocated structure. If none exists loop and try again. 190 */ 191 VI_LOCK(vp); 192 dh = ip->i_dirhash; 193 if (dh == NULL) { 194 ip->i_dirhash = ndh; 195 VI_UNLOCK(vp); 196 if (ndh == NULL) 197 continue; 198 return (ndh); 199 } 200 ufsdirhash_hold(dh); 201 VI_UNLOCK(vp); 202 203 /* Acquire a shared lock on existing hashes. */ 204 sx_slock(&dh->dh_lock); 205 206 /* The hash could've been recycled while we were waiting. */ 207 VI_LOCK(vp); 208 if (ip->i_dirhash != dh) { 209 VI_UNLOCK(vp); 210 ufsdirhash_release(dh); 211 ufsdirhash_drop(dh); 212 continue; 213 } 214 VI_UNLOCK(vp); 215 ufsdirhash_drop(dh); 216 217 /* If the hash is still valid we've succeeded. */ 218 if (dh->dh_hash != NULL) 219 break; 220 /* 221 * If the hash is NULL it has been recycled. Try to upgrade 222 * so we can recreate it. If we fail the upgrade, drop our 223 * lock and try again. 224 */ 225 if (sx_try_upgrade(&dh->dh_lock)) 226 break; 227 sx_sunlock(&dh->dh_lock); 228 } 229 /* Free the preallocated structure if it was not necessary. */ 230 if (ndh) { 231 ufsdirhash_release(ndh); 232 ufsdirhash_drop(ndh); 233 } 234 return (dh); 235} 236 237/* 238 * Acquire an exclusive lock on an existing hash. Requires an exclusive 239 * vnode lock to protect the i_dirhash pointer. hashes that have been 240 * recycled are reclaimed here and NULL is returned. 241 */ 242static struct dirhash * 243ufsdirhash_acquire(struct inode *ip) 244{ 245 struct dirhash *dh; 246 struct vnode *vp; 247 248 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__); 249 250 vp = ip->i_vnode; 251 dh = ip->i_dirhash; 252 if (dh == NULL) 253 return (NULL); 254 sx_xlock(&dh->dh_lock); 255 if (dh->dh_hash != NULL) 256 return (dh); 257 ufsdirhash_free_locked(ip); 258 return (NULL); 259} 260 261/* 262 * Acquire exclusively and free the hash pointed to by ip. Works with a 263 * shared or exclusive vnode lock. 264 */ 265void 266ufsdirhash_free(struct inode *ip) 267{ 268 struct dirhash *dh; 269 struct vnode *vp; 270 271 vp = ip->i_vnode; 272 for (;;) { 273 /* Grab a reference on this inode's dirhash if it has one. */ 274 VI_LOCK(vp); 275 dh = ip->i_dirhash; 276 if (dh == NULL) { 277 VI_UNLOCK(vp); 278 return; 279 } 280 ufsdirhash_hold(dh); 281 VI_UNLOCK(vp); 282 283 /* Exclusively lock the dirhash. */ 284 sx_xlock(&dh->dh_lock); 285 286 /* If this dirhash still belongs to this inode, then free it. */ 287 VI_LOCK(vp); 288 if (ip->i_dirhash == dh) { 289 VI_UNLOCK(vp); 290 ufsdirhash_drop(dh); 291 break; 292 } 293 VI_UNLOCK(vp); 294 295 /* 296 * This inode's dirhash has changed while we were 297 * waiting for the dirhash lock, so try again. 298 */ 299 ufsdirhash_release(dh); 300 ufsdirhash_drop(dh); 301 } 302 ufsdirhash_free_locked(ip); 303} 304 305/* 306 * Attempt to build up a hash table for the directory contents in 307 * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 308 */ 309int 310ufsdirhash_build(struct inode *ip) 311{ 312 struct dirhash *dh; 313 struct buf *bp = NULL; 314 struct direct *ep; 315 struct vnode *vp; 316 doff_t bmask, pos; 317 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 318 319 /* Take care of a decreased sysctl value. */ 320 while (ufs_dirhashmem > ufs_dirhashmaxmem) 321 if (ufsdirhash_recycle(0) != 0) 322 return (-1); 323 324 /* Check if we can/should use dirhash. */ 325 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) || 326 ip->i_effnlink == 0) { 327 if (ip->i_dirhash) 328 ufsdirhash_free(ip); 329 return (-1); 330 } 331 dh = ufsdirhash_create(ip); 332 if (dh == NULL) 333 return (-1); 334 if (dh->dh_hash != NULL) 335 return (0); 336 337 vp = ip->i_vnode; 338 /* Allocate 50% more entries than this dir size could ever need. */ 339 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 340 nslots = ip->i_size / DIRECTSIZ(1); 341 nslots = (nslots * 3 + 1) / 2; 342 narrays = howmany(nslots, DH_NBLKOFF); 343 nslots = narrays * DH_NBLKOFF; 344 dirblocks = howmany(ip->i_size, DIRBLKSIZ); 345 nblocks = (dirblocks * 3 + 1) / 2; 346 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 347 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 348 nblocks * sizeof(*dh->dh_blkfree); 349 DIRHASHLIST_LOCK(); 350 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 351 DIRHASHLIST_UNLOCK(); 352 if (memreqd > ufs_dirhashmaxmem / 2) 353 goto fail; 354 /* Try to free some space. */ 355 if (ufsdirhash_recycle(memreqd) != 0) 356 goto fail; 357 /* Enough was freed, and list has been locked. */ 358 } 359 ufs_dirhashmem += memreqd; 360 DIRHASHLIST_UNLOCK(); 361 362 /* Initialise the hash table and block statistics. */ 363 dh->dh_memreq = memreqd; 364 dh->dh_narrays = narrays; 365 dh->dh_hlen = nslots; 366 dh->dh_nblk = nblocks; 367 dh->dh_dirblks = dirblocks; 368 for (i = 0; i < DH_NFSTATS; i++) 369 dh->dh_firstfree[i] = -1; 370 dh->dh_firstfree[DH_NFSTATS] = 0; 371 dh->dh_hused = 0; 372 dh->dh_seqopt = 0; 373 dh->dh_seqoff = 0; 374 dh->dh_score = DH_SCOREINIT; 375 376 /* 377 * Use non-blocking mallocs so that we will revert to a linear 378 * lookup on failure rather than potentially blocking forever. 379 */ 380 MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]), 381 M_DIRHASH, M_NOWAIT | M_ZERO); 382 if (dh->dh_hash == NULL) 383 goto fail; 384 MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]), 385 M_DIRHASH, M_NOWAIT); 386 if (dh->dh_blkfree == NULL) 387 goto fail; 388 for (i = 0; i < narrays; i++) { 389 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL) 390 goto fail; 391 for (j = 0; j < DH_NBLKOFF; j++) 392 dh->dh_hash[i][j] = DIRHASH_EMPTY; 393 } 394 for (i = 0; i < dirblocks; i++) 395 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 396 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 397 pos = 0; 398 while (pos < ip->i_size) { 399 /* If necessary, get the next directory block. */ 400 if ((pos & bmask) == 0) { 401 if (bp != NULL) 402 brelse(bp); 403 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 404 goto fail; 405 } 406 407 /* Add this entry to the hash. */ 408 ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 409 if (ep->d_reclen == 0 || ep->d_reclen > 410 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 411 /* Corrupted directory. */ 412 brelse(bp); 413 goto fail; 414 } 415 if (ep->d_ino != 0) { 416 /* Add the entry (simplified ufsdirhash_add). */ 417 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 418 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 419 slot = WRAPINCR(slot, dh->dh_hlen); 420 dh->dh_hused++; 421 DH_ENTRY(dh, slot) = pos; 422 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 423 } 424 pos += ep->d_reclen; 425 } 426 427 if (bp != NULL) 428 brelse(bp); 429 DIRHASHLIST_LOCK(); 430 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 431 dh->dh_onlist = 1; 432 DIRHASHLIST_UNLOCK(); 433 sx_downgrade(&dh->dh_lock); 434 return (0); 435 436fail: 437 ufsdirhash_free_locked(ip); 438 return (-1); 439} 440 441/* 442 * Free any hash table associated with inode 'ip'. 443 */ 444static void 445ufsdirhash_free_locked(struct inode *ip) 446{ 447 struct dirhash *dh; 448 struct vnode *vp; 449 int i; 450 451 DIRHASH_ASSERT_LOCKED(ip->i_dirhash); 452 453 /* 454 * Clear the pointer in the inode to prevent new threads from 455 * finding the dead structure. 456 */ 457 vp = ip->i_vnode; 458 VI_LOCK(vp); 459 dh = ip->i_dirhash; 460 ip->i_dirhash = NULL; 461 VI_UNLOCK(vp); 462 463 /* 464 * At this point, any waiters for the lock should hold their 465 * own reference on the dirhash structure. They will drop 466 * that reference once they grab the vnode interlock and see 467 * that ip->i_dirhash is NULL. 468 */ 469 sx_xunlock(&dh->dh_lock); 470 471 /* 472 * Handle partially recycled as well as fully constructed hashes. 473 */ 474 if (dh->dh_hash != NULL) { 475 for (i = 0; i < dh->dh_narrays; i++) 476 if (dh->dh_hash[i] != NULL) 477 DIRHASH_BLKFREE(dh->dh_hash[i]); 478 FREE(dh->dh_hash, M_DIRHASH); 479 if (dh->dh_blkfree != NULL) 480 FREE(dh->dh_blkfree, M_DIRHASH); 481 } 482 DIRHASHLIST_LOCK(); 483 if (dh->dh_onlist) 484 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 485 ufs_dirhashmem -= dh->dh_memreq; 486 DIRHASHLIST_UNLOCK(); 487 488 /* 489 * Drop the inode's reference to the data structure. 490 */ 491 ufsdirhash_drop(dh); 492} 493 494/* 495 * Find the offset of the specified name within the given inode. 496 * Returns 0 on success, ENOENT if the entry does not exist, or 497 * EJUSTRETURN if the caller should revert to a linear search. 498 * 499 * If successful, the directory offset is stored in *offp, and a 500 * pointer to a struct buf containing the entry is stored in *bpp. If 501 * prevoffp is non-NULL, the offset of the previous entry within 502 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 503 * is the first in a block, the start of the block is used). 504 * 505 * Must be called with the hash locked. Returns with the hash unlocked. 506 */ 507int 508ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 509 struct buf **bpp, doff_t *prevoffp) 510{ 511 struct dirhash *dh, *dh_next; 512 struct direct *dp; 513 struct vnode *vp; 514 struct buf *bp; 515 doff_t blkoff, bmask, offset, prevoff; 516 int i, slot; 517 int error; 518 519 dh = ip->i_dirhash; 520 KASSERT(dh != NULL && dh->dh_hash != NULL, 521 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh)); 522 DIRHASH_ASSERT_LOCKED(dh); 523 /* 524 * Move this dirhash towards the end of the list if it has a 525 * score higher than the next entry, and acquire the dh_lock. 526 */ 527 DIRHASHLIST_LOCK(); 528 if (TAILQ_NEXT(dh, dh_list) != NULL) { 529 /* 530 * If the new score will be greater than that of the next 531 * entry, then move this entry past it. With both mutexes 532 * held, dh_next won't go away, but its dh_score could 533 * change; that's not important since it is just a hint. 534 */ 535 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 536 dh->dh_score >= dh_next->dh_score) { 537 KASSERT(dh->dh_onlist, ("dirhash: not on list")); 538 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 539 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 540 dh_list); 541 } 542 } 543 /* Update the score. */ 544 if (dh->dh_score < DH_SCOREMAX) 545 dh->dh_score++; 546 DIRHASHLIST_UNLOCK(); 547 548 vp = ip->i_vnode; 549 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 550 blkoff = -1; 551 bp = NULL; 552restart: 553 slot = ufsdirhash_hash(dh, name, namelen); 554 555 if (dh->dh_seqopt) { 556 /* 557 * Sequential access optimisation. dh_seqoff contains the 558 * offset of the directory entry immediately following 559 * the last entry that was looked up. Check if this offset 560 * appears in the hash chain for the name we are looking for. 561 */ 562 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 563 i = WRAPINCR(i, dh->dh_hlen)) 564 if (offset == dh->dh_seqoff) 565 break; 566 if (offset == dh->dh_seqoff) { 567 /* 568 * We found an entry with the expected offset. This 569 * is probably the entry we want, but if not, the 570 * code below will turn off seqopt and retry. 571 */ 572 slot = i; 573 } else 574 dh->dh_seqopt = 0; 575 } 576 577 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 578 slot = WRAPINCR(slot, dh->dh_hlen)) { 579 if (offset == DIRHASH_DEL) 580 continue; 581 if (offset < 0 || offset >= ip->i_size) 582 panic("ufsdirhash_lookup: bad offset in hash array"); 583 if ((offset & ~bmask) != blkoff) { 584 if (bp != NULL) 585 brelse(bp); 586 blkoff = offset & ~bmask; 587 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) { 588 error = EJUSTRETURN; 589 goto fail; 590 } 591 } 592 dp = (struct direct *)(bp->b_data + (offset & bmask)); 593 if (dp->d_reclen == 0 || dp->d_reclen > 594 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 595 /* Corrupted directory. */ 596 error = EJUSTRETURN; 597 goto fail; 598 } 599 if (dp->d_namlen == namelen && 600 bcmp(dp->d_name, name, namelen) == 0) { 601 /* Found. Get the prev offset if needed. */ 602 if (prevoffp != NULL) { 603 if (offset & (DIRBLKSIZ - 1)) { 604 prevoff = ufsdirhash_getprev(dp, 605 offset); 606 if (prevoff == -1) { 607 error = EJUSTRETURN; 608 goto fail; 609 } 610 } else 611 prevoff = offset; 612 *prevoffp = prevoff; 613 } 614 615 /* Check for sequential access, and update offset. */ 616 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 617 dh->dh_seqopt = 1; 618 dh->dh_seqoff = offset + DIRSIZ(0, dp); 619 *bpp = bp; 620 *offp = offset; 621 ufsdirhash_release(dh); 622 return (0); 623 } 624 625 /* 626 * When the name doesn't match in the seqopt case, go back 627 * and search normally. 628 */ 629 if (dh->dh_seqopt) { 630 dh->dh_seqopt = 0; 631 goto restart; 632 } 633 } 634 error = ENOENT; 635fail: 636 ufsdirhash_release(dh); 637 if (bp != NULL) 638 brelse(bp); 639 return (error); 640} 641 642/* 643 * Find a directory block with room for 'slotneeded' bytes. Returns 644 * the offset of the directory entry that begins the free space. 645 * This will either be the offset of an existing entry that has free 646 * space at the end, or the offset of an entry with d_ino == 0 at 647 * the start of a DIRBLKSIZ block. 648 * 649 * To use the space, the caller may need to compact existing entries in 650 * the directory. The total number of bytes in all of the entries involved 651 * in the compaction is stored in *slotsize. In other words, all of 652 * the entries that must be compacted are exactly contained in the 653 * region beginning at the returned offset and spanning *slotsize bytes. 654 * 655 * Returns -1 if no space was found, indicating that the directory 656 * must be extended. 657 */ 658doff_t 659ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 660{ 661 struct direct *dp; 662 struct dirhash *dh; 663 struct buf *bp; 664 doff_t pos, slotstart; 665 int dirblock, error, freebytes, i; 666 667 dh = ip->i_dirhash; 668 KASSERT(dh != NULL && dh->dh_hash != NULL, 669 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh)); 670 DIRHASH_ASSERT_LOCKED(dh); 671 672 /* Find a directory block with the desired free space. */ 673 dirblock = -1; 674 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 675 if ((dirblock = dh->dh_firstfree[i]) != -1) 676 break; 677 if (dirblock == -1) 678 return (-1); 679 680 KASSERT(dirblock < dh->dh_nblk && 681 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 682 ("ufsdirhash_findfree: bad stats")); 683 pos = dirblock * DIRBLKSIZ; 684 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 685 if (error) 686 return (-1); 687 688 /* Find the first entry with free space. */ 689 for (i = 0; i < DIRBLKSIZ; ) { 690 if (dp->d_reclen == 0) { 691 brelse(bp); 692 return (-1); 693 } 694 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 695 break; 696 i += dp->d_reclen; 697 dp = (struct direct *)((char *)dp + dp->d_reclen); 698 } 699 if (i > DIRBLKSIZ) { 700 brelse(bp); 701 return (-1); 702 } 703 slotstart = pos + i; 704 705 /* Find the range of entries needed to get enough space */ 706 freebytes = 0; 707 while (i < DIRBLKSIZ && freebytes < slotneeded) { 708 freebytes += dp->d_reclen; 709 if (dp->d_ino != 0) 710 freebytes -= DIRSIZ(0, dp); 711 if (dp->d_reclen == 0) { 712 brelse(bp); 713 return (-1); 714 } 715 i += dp->d_reclen; 716 dp = (struct direct *)((char *)dp + dp->d_reclen); 717 } 718 if (i > DIRBLKSIZ) { 719 brelse(bp); 720 return (-1); 721 } 722 if (freebytes < slotneeded) 723 panic("ufsdirhash_findfree: free mismatch"); 724 brelse(bp); 725 *slotsize = pos + i - slotstart; 726 return (slotstart); 727} 728 729/* 730 * Return the start of the unused space at the end of a directory, or 731 * -1 if there are no trailing unused blocks. 732 */ 733doff_t 734ufsdirhash_enduseful(struct inode *ip) 735{ 736 737 struct dirhash *dh; 738 int i; 739 740 dh = ip->i_dirhash; 741 DIRHASH_ASSERT_LOCKED(dh); 742 KASSERT(dh != NULL && dh->dh_hash != NULL, 743 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh)); 744 745 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) 746 return (-1); 747 748 for (i = dh->dh_dirblks - 1; i >= 0; i--) 749 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 750 break; 751 752 return ((doff_t)(i + 1) * DIRBLKSIZ); 753} 754 755/* 756 * Insert information into the hash about a new directory entry. dirp 757 * points to a struct direct containing the entry, and offset specifies 758 * the offset of this entry. 759 */ 760void 761ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 762{ 763 struct dirhash *dh; 764 int slot; 765 766 if ((dh = ufsdirhash_acquire(ip)) == NULL) 767 return; 768 769 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 770 ("ufsdirhash_add: bad offset")); 771 /* 772 * Normal hash usage is < 66%. If the usage gets too high then 773 * remove the hash entirely and let it be rebuilt later. 774 */ 775 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 776 ufsdirhash_free_locked(ip); 777 return; 778 } 779 780 /* Find a free hash slot (empty or deleted), and add the entry. */ 781 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 782 while (DH_ENTRY(dh, slot) >= 0) 783 slot = WRAPINCR(slot, dh->dh_hlen); 784 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 785 dh->dh_hused++; 786 DH_ENTRY(dh, slot) = offset; 787 788 /* Update the per-block summary info. */ 789 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 790 ufsdirhash_release(dh); 791} 792 793/* 794 * Remove the specified directory entry from the hash. The entry to remove 795 * is defined by the name in `dirp', which must exist at the specified 796 * `offset' within the directory. 797 */ 798void 799ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 800{ 801 struct dirhash *dh; 802 int slot; 803 804 if ((dh = ufsdirhash_acquire(ip)) == NULL) 805 return; 806 807 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 808 ("ufsdirhash_remove: bad offset")); 809 /* Find the entry */ 810 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 811 812 /* Remove the hash entry. */ 813 ufsdirhash_delslot(dh, slot); 814 815 /* Update the per-block summary info. */ 816 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 817 ufsdirhash_release(dh); 818} 819 820/* 821 * Change the offset associated with a directory entry in the hash. Used 822 * when compacting directory blocks. 823 */ 824void 825ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 826 doff_t newoff) 827{ 828 struct dirhash *dh; 829 int slot; 830 831 if ((dh = ufsdirhash_acquire(ip)) == NULL) 832 return; 833 834 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 835 newoff < dh->dh_dirblks * DIRBLKSIZ, 836 ("ufsdirhash_move: bad offset")); 837 /* Find the entry, and update the offset. */ 838 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 839 DH_ENTRY(dh, slot) = newoff; 840 ufsdirhash_release(dh); 841} 842 843/* 844 * Inform dirhash that the directory has grown by one block that 845 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 846 */ 847void 848ufsdirhash_newblk(struct inode *ip, doff_t offset) 849{ 850 struct dirhash *dh; 851 int block; 852 853 if ((dh = ufsdirhash_acquire(ip)) == NULL) 854 return; 855 856 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 857 ("ufsdirhash_newblk: bad offset")); 858 block = offset / DIRBLKSIZ; 859 if (block >= dh->dh_nblk) { 860 /* Out of space; must rebuild. */ 861 ufsdirhash_free_locked(ip); 862 return; 863 } 864 dh->dh_dirblks = block + 1; 865 866 /* Account for the new free block. */ 867 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 868 if (dh->dh_firstfree[DH_NFSTATS] == -1) 869 dh->dh_firstfree[DH_NFSTATS] = block; 870 ufsdirhash_release(dh); 871} 872 873/* 874 * Inform dirhash that the directory is being truncated. 875 */ 876void 877ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 878{ 879 struct dirhash *dh; 880 int block, i; 881 882 if ((dh = ufsdirhash_acquire(ip)) == NULL) 883 return; 884 885 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 886 ("ufsdirhash_dirtrunc: bad offset")); 887 block = howmany(offset, DIRBLKSIZ); 888 /* 889 * If the directory shrinks to less than 1/8 of dh_nblk blocks 890 * (about 20% of its original size due to the 50% extra added in 891 * ufsdirhash_build) then free it, and let the caller rebuild 892 * if necessary. 893 */ 894 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 895 ufsdirhash_free_locked(ip); 896 return; 897 } 898 899 /* 900 * Remove any `first free' information pertaining to the 901 * truncated blocks. All blocks we're removing should be 902 * completely unused. 903 */ 904 if (dh->dh_firstfree[DH_NFSTATS] >= block) 905 dh->dh_firstfree[DH_NFSTATS] = -1; 906 for (i = block; i < dh->dh_dirblks; i++) 907 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 908 panic("ufsdirhash_dirtrunc: blocks in use"); 909 for (i = 0; i < DH_NFSTATS; i++) 910 if (dh->dh_firstfree[i] >= block) 911 panic("ufsdirhash_dirtrunc: first free corrupt"); 912 dh->dh_dirblks = block; 913 ufsdirhash_release(dh); 914} 915 916/* 917 * Debugging function to check that the dirhash information about 918 * a directory block matches its actual contents. Panics if a mismatch 919 * is detected. 920 * 921 * On entry, `buf' should point to the start of an in-core 922 * DIRBLKSIZ-sized directory block, and `offset' should contain the 923 * offset from the start of the directory of that block. 924 */ 925void 926ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 927{ 928 struct dirhash *dh; 929 struct direct *dp; 930 int block, ffslot, i, nfree; 931 932 if (!ufs_dirhashcheck) 933 return; 934 if ((dh = ufsdirhash_acquire(ip)) == NULL) 935 return; 936 937 block = offset / DIRBLKSIZ; 938 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 939 panic("ufsdirhash_checkblock: bad offset"); 940 941 nfree = 0; 942 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 943 dp = (struct direct *)(buf + i); 944 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 945 panic("ufsdirhash_checkblock: bad dir"); 946 947 if (dp->d_ino == 0) { 948#if 0 949 /* 950 * XXX entries with d_ino == 0 should only occur 951 * at the start of a DIRBLKSIZ block. However the 952 * ufs code is tolerant of such entries at other 953 * offsets, and fsck does not fix them. 954 */ 955 if (i != 0) 956 panic("ufsdirhash_checkblock: bad dir inode"); 957#endif 958 nfree += dp->d_reclen; 959 continue; 960 } 961 962 /* Check that the entry exists (will panic if it doesn't). */ 963 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 964 965 nfree += dp->d_reclen - DIRSIZ(0, dp); 966 } 967 if (i != DIRBLKSIZ) 968 panic("ufsdirhash_checkblock: bad dir end"); 969 970 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 971 panic("ufsdirhash_checkblock: bad free count"); 972 973 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 974 for (i = 0; i <= DH_NFSTATS; i++) 975 if (dh->dh_firstfree[i] == block && i != ffslot) 976 panic("ufsdirhash_checkblock: bad first-free"); 977 if (dh->dh_firstfree[ffslot] == -1) 978 panic("ufsdirhash_checkblock: missing first-free entry"); 979 ufsdirhash_release(dh); 980} 981 982/* 983 * Hash the specified filename into a dirhash slot. 984 */ 985static int 986ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 987{ 988 u_int32_t hash; 989 990 /* 991 * We hash the name and then some other bit of data that is 992 * invariant over the dirhash's lifetime. Otherwise names 993 * differing only in the last byte are placed close to one 994 * another in the table, which is bad for linear probing. 995 */ 996 hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 997 hash = fnv_32_buf(&dh, sizeof(dh), hash); 998 return (hash % dh->dh_hlen); 999} 1000 1001/* 1002 * Adjust the number of free bytes in the block containing `offset' 1003 * by the value specified by `diff'. 1004 * 1005 * The caller must ensure we have exclusive access to `dh'; normally 1006 * that means that dh_lock should be held, but this is also called 1007 * from ufsdirhash_build() where exclusive access can be assumed. 1008 */ 1009static void 1010ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 1011{ 1012 int block, i, nfidx, ofidx; 1013 1014 /* Update the per-block summary info. */ 1015 block = offset / DIRBLKSIZ; 1016 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 1017 ("dirhash bad offset")); 1018 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1019 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 1020 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1021 1022 /* Update the `first free' list if necessary. */ 1023 if (ofidx != nfidx) { 1024 /* If removing, scan forward for the next block. */ 1025 if (dh->dh_firstfree[ofidx] == block) { 1026 for (i = block + 1; i < dh->dh_dirblks; i++) 1027 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 1028 break; 1029 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 1030 } 1031 1032 /* Make this the new `first free' if necessary */ 1033 if (dh->dh_firstfree[nfidx] > block || 1034 dh->dh_firstfree[nfidx] == -1) 1035 dh->dh_firstfree[nfidx] = block; 1036 } 1037} 1038 1039/* 1040 * Find the specified name which should have the specified offset. 1041 * Returns a slot number, and panics on failure. 1042 * 1043 * `dh' must be locked on entry and remains so on return. 1044 */ 1045static int 1046ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 1047{ 1048 int slot; 1049 1050 DIRHASH_ASSERT_LOCKED(dh); 1051 1052 /* Find the entry. */ 1053 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 1054 slot = ufsdirhash_hash(dh, name, namelen); 1055 while (DH_ENTRY(dh, slot) != offset && 1056 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 1057 slot = WRAPINCR(slot, dh->dh_hlen); 1058 if (DH_ENTRY(dh, slot) != offset) 1059 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 1060 1061 return (slot); 1062} 1063 1064/* 1065 * Remove the entry corresponding to the specified slot from the hash array. 1066 * 1067 * `dh' must be locked on entry and remains so on return. 1068 */ 1069static void 1070ufsdirhash_delslot(struct dirhash *dh, int slot) 1071{ 1072 int i; 1073 1074 DIRHASH_ASSERT_LOCKED(dh); 1075 1076 /* Mark the entry as deleted. */ 1077 DH_ENTRY(dh, slot) = DIRHASH_DEL; 1078 1079 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 1080 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 1081 i = WRAPINCR(i, dh->dh_hlen); 1082 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 1083 i = WRAPDECR(i, dh->dh_hlen); 1084 while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 1085 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 1086 dh->dh_hused--; 1087 i = WRAPDECR(i, dh->dh_hlen); 1088 } 1089 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 1090 } 1091} 1092 1093/* 1094 * Given a directory entry and its offset, find the offset of the 1095 * previous entry in the same DIRBLKSIZ-sized block. Returns an 1096 * offset, or -1 if there is no previous entry in the block or some 1097 * other problem occurred. 1098 */ 1099static doff_t 1100ufsdirhash_getprev(struct direct *dirp, doff_t offset) 1101{ 1102 struct direct *dp; 1103 char *blkbuf; 1104 doff_t blkoff, prevoff; 1105 int entrypos, i; 1106 1107 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 1108 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 1109 blkbuf = (char *)dirp - entrypos; 1110 prevoff = blkoff; 1111 1112 /* If `offset' is the start of a block, there is no previous entry. */ 1113 if (entrypos == 0) 1114 return (-1); 1115 1116 /* Scan from the start of the block until we get to the entry. */ 1117 for (i = 0; i < entrypos; i += dp->d_reclen) { 1118 dp = (struct direct *)(blkbuf + i); 1119 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 1120 return (-1); /* Corrupted directory. */ 1121 prevoff = blkoff + i; 1122 } 1123 return (prevoff); 1124} 1125 1126/* 1127 * Try to free up `wanted' bytes by stealing memory from existing 1128 * dirhashes. Returns zero with list locked if successful. 1129 */ 1130static int 1131ufsdirhash_recycle(int wanted) 1132{ 1133 struct dirhash *dh; 1134 doff_t **hash; 1135 u_int8_t *blkfree; 1136 int i, mem, narrays; 1137 1138 DIRHASHLIST_LOCK(); 1139 dh = TAILQ_FIRST(&ufsdirhash_list); 1140 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 1141 /* Decrement the score; only recycle if it becomes zero. */ 1142 if (dh == NULL || --dh->dh_score > 0) { 1143 DIRHASHLIST_UNLOCK(); 1144 return (-1); 1145 } 1146 /* 1147 * If we can't lock it it's in use and we don't want to 1148 * recycle it anyway. 1149 */ 1150 if (!sx_try_xlock(&dh->dh_lock)) { 1151 dh = TAILQ_NEXT(dh, dh_list); 1152 continue; 1153 } 1154 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 1155 1156 /* Remove it from the list and detach its memory. */ 1157 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 1158 dh->dh_onlist = 0; 1159 hash = dh->dh_hash; 1160 dh->dh_hash = NULL; 1161 blkfree = dh->dh_blkfree; 1162 dh->dh_blkfree = NULL; 1163 narrays = dh->dh_narrays; 1164 mem = dh->dh_memreq; 1165 dh->dh_memreq = 0; 1166 1167 /* Unlock everything, free the detached memory. */ 1168 ufsdirhash_release(dh); 1169 DIRHASHLIST_UNLOCK(); 1170 for (i = 0; i < narrays; i++) 1171 DIRHASH_BLKFREE(hash[i]); 1172 FREE(hash, M_DIRHASH); 1173 FREE(blkfree, M_DIRHASH); 1174 1175 /* Account for the returned memory, and repeat if necessary. */ 1176 DIRHASHLIST_LOCK(); 1177 ufs_dirhashmem -= mem; 1178 dh = TAILQ_FIRST(&ufsdirhash_list); 1179 } 1180 /* Success; return with list locked. */ 1181 return (0); 1182} 1183 1184 1185void 1186ufsdirhash_init() 1187{ 1188 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t), 1189 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1190 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF); 1191 TAILQ_INIT(&ufsdirhash_list); 1192} 1193 1194void 1195ufsdirhash_uninit() 1196{ 1197 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit")); 1198 uma_zdestroy(ufsdirhash_zone); 1199 mtx_destroy(&ufsdirhash_mtx); 1200} 1201 1202#endif /* UFS_DIRHASH */ 1203