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("$FreeBSD$"); 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 <sys/eventhandler.h> 53 #include <sys/time.h> 54 #include <vm/uma.h> 55 56 #include <ufs/ufs/quota.h> 57 #include <ufs/ufs/inode.h> 58 #include <ufs/ufs/dir.h> 59 #include <ufs/ufs/dirhash.h> 60 #include <ufs/ufs/extattr.h> 61 #include <ufs/ufs/ufsmount.h> 62 #include <ufs/ufs/ufs_extern.h> 63 64 #define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1)) 65 #define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1)) 66 #define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 67 #define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n)) 68 69 static MALLOC_DEFINE(M_DIRHASH, "ufs_dirhash", "UFS directory hash tables"); 70 71 static int ufs_mindirhashsize = DIRBLKSIZ * 5; 72 SYSCTL_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"); 75 static int ufs_dirhashmaxmem = 2 * 1024 * 1024; /* NOTE: initial value. It is 76 tuned in ufsdirhash_init() */ 77 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 78 0, "maximum allowed dirhash memory usage"); 79 static int ufs_dirhashmem; 80 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 81 0, "current dirhash memory usage"); 82 static int ufs_dirhashcheck = 0; 83 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 84 0, "enable extra sanity tests"); 85 static int ufs_dirhashlowmemcount = 0; 86 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_lowmemcount, CTLFLAG_RD, 87 &ufs_dirhashlowmemcount, 0, "number of times low memory hook called"); 88 static int ufs_dirhashreclaimage = 5; 89 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_reclaimage, CTLFLAG_RW, 90 &ufs_dirhashreclaimage, 0, 91 "max time in seconds of hash inactivity before deletion in low VM events"); 92 93 94 static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 95 static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 96 static void ufsdirhash_delslot(struct dirhash *dh, int slot); 97 static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 98 doff_t offset); 99 static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 100 static int ufsdirhash_recycle(int wanted); 101 static void ufsdirhash_lowmem(void); 102 static void ufsdirhash_free_locked(struct inode *ip); 103 104 static uma_zone_t ufsdirhash_zone; 105 106 #define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx) 107 #define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx) 108 #define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK) 109 #define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr)) 110 #define DIRHASH_ASSERT_LOCKED(dh) \ 111 sx_assert(&(dh)->dh_lock, SA_LOCKED) 112 113 /* Dirhash list; recently-used entries are near the tail. */ 114 static TAILQ_HEAD(, dirhash) ufsdirhash_list; 115 116 /* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 117 static struct mtx ufsdirhash_mtx; 118 119 /* 120 * Locking: 121 * 122 * The relationship between inode and dirhash is protected either by an 123 * exclusive vnode lock or the vnode interlock where a shared vnode lock 124 * may be used. The dirhash_mtx is acquired after the dirhash lock. To 125 * handle teardown races, code wishing to lock the dirhash for an inode 126 * when using a shared vnode lock must obtain a private reference on the 127 * dirhash while holding the vnode interlock. They can drop it once they 128 * have obtained the dirhash lock and verified that the dirhash wasn't 129 * recycled while they waited for the dirhash lock. 130 * 131 * ufsdirhash_build() acquires a shared lock on the dirhash when it is 132 * successful. This lock is released after a call to ufsdirhash_lookup(). 133 * 134 * Functions requiring exclusive access use ufsdirhash_acquire() which may 135 * free a dirhash structure that was recycled by ufsdirhash_recycle(). 136 * 137 * The dirhash lock may be held across io operations. 138 * 139 * WITNESS reports a lock order reversal between the "bufwait" lock 140 * and the "dirhash" lock. However, this specific reversal will not 141 * cause a deadlock. To get a deadlock, one would have to lock a 142 * buffer followed by the dirhash while a second thread locked a 143 * buffer while holding the dirhash lock. The second order can happen 144 * under a shared or exclusive vnode lock for the associated directory 145 * in lookup(). The first order, however, can only happen under an 146 * exclusive vnode lock (e.g. unlink(), rename(), etc.). Thus, for 147 * a thread to be doing a "bufwait" -> "dirhash" order, it has to hold 148 * an exclusive vnode lock. That exclusive vnode lock will prevent 149 * any other threads from doing a "dirhash" -> "bufwait" order. 150 */ 151 152 static void 153 ufsdirhash_hold(struct dirhash *dh) 154 { 155 156 refcount_acquire(&dh->dh_refcount); 157 } 158 159 static void 160 ufsdirhash_drop(struct dirhash *dh) 161 { 162 163 if (refcount_release(&dh->dh_refcount)) { 164 sx_destroy(&dh->dh_lock); 165 free(dh, M_DIRHASH); 166 } 167 } 168 169 /* 170 * Release the lock on a dirhash. 171 */ 172 static void 173 ufsdirhash_release(struct dirhash *dh) 174 { 175 176 sx_unlock(&dh->dh_lock); 177 } 178 179 /* 180 * Either acquire an existing hash locked shared or create a new hash and 181 * return it exclusively locked. May return NULL if the allocation fails. 182 * 183 * The vnode interlock is used to protect the i_dirhash pointer from 184 * simultaneous access while only a shared vnode lock is held. 185 */ 186 static struct dirhash * 187 ufsdirhash_create(struct inode *ip) 188 { 189 struct dirhash *ndh; 190 struct dirhash *dh; 191 struct vnode *vp; 192 int error; 193 194 error = 0; 195 ndh = dh = NULL; 196 vp = ip->i_vnode; 197 for (;;) { 198 /* Racy check for i_dirhash to prefetch a dirhash structure. */ 199 if (ip->i_dirhash == NULL && ndh == NULL) { 200 ndh = malloc(sizeof *dh, M_DIRHASH, 201 M_NOWAIT | M_ZERO); 202 if (ndh == NULL) 203 return (NULL); 204 refcount_init(&ndh->dh_refcount, 1); 205 206 /* 207 * The DUPOK is to prevent warnings from the 208 * sx_slock() a few lines down which is safe 209 * since the duplicate lock in that case is 210 * the one for this dirhash we are creating 211 * now which has no external references until 212 * after this function returns. 213 */ 214 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK); 215 sx_xlock(&ndh->dh_lock); 216 } 217 /* 218 * Check i_dirhash. If it's NULL just try to use a 219 * preallocated structure. If none exists loop and try again. 220 */ 221 VI_LOCK(vp); 222 dh = ip->i_dirhash; 223 if (dh == NULL) { 224 ip->i_dirhash = ndh; 225 VI_UNLOCK(vp); 226 if (ndh == NULL) 227 continue; 228 return (ndh); 229 } 230 ufsdirhash_hold(dh); 231 VI_UNLOCK(vp); 232 233 /* Acquire a shared lock on existing hashes. */ 234 sx_slock(&dh->dh_lock); 235 236 /* The hash could've been recycled while we were waiting. */ 237 VI_LOCK(vp); 238 if (ip->i_dirhash != dh) { 239 VI_UNLOCK(vp); 240 ufsdirhash_release(dh); 241 ufsdirhash_drop(dh); 242 continue; 243 } 244 VI_UNLOCK(vp); 245 ufsdirhash_drop(dh); 246 247 /* If the hash is still valid we've succeeded. */ 248 if (dh->dh_hash != NULL) 249 break; 250 /* 251 * If the hash is NULL it has been recycled. Try to upgrade 252 * so we can recreate it. If we fail the upgrade, drop our 253 * lock and try again. 254 */ 255 if (sx_try_upgrade(&dh->dh_lock)) 256 break; 257 sx_sunlock(&dh->dh_lock); 258 } 259 /* Free the preallocated structure if it was not necessary. */ 260 if (ndh) { 261 ufsdirhash_release(ndh); 262 ufsdirhash_drop(ndh); 263 } 264 return (dh); 265 } 266 267 /* 268 * Acquire an exclusive lock on an existing hash. Requires an exclusive 269 * vnode lock to protect the i_dirhash pointer. hashes that have been 270 * recycled are reclaimed here and NULL is returned. 271 */ 272 static struct dirhash * 273 ufsdirhash_acquire(struct inode *ip) 274 { 275 struct dirhash *dh; 276 struct vnode *vp; 277 278 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__); 279 280 vp = ip->i_vnode; 281 dh = ip->i_dirhash; 282 if (dh == NULL) 283 return (NULL); 284 sx_xlock(&dh->dh_lock); 285 if (dh->dh_hash != NULL) 286 return (dh); 287 ufsdirhash_free_locked(ip); 288 return (NULL); 289 } 290 291 /* 292 * Acquire exclusively and free the hash pointed to by ip. Works with a 293 * shared or exclusive vnode lock. 294 */ 295 void 296 ufsdirhash_free(struct inode *ip) 297 { 298 struct dirhash *dh; 299 struct vnode *vp; 300 301 vp = ip->i_vnode; 302 for (;;) { 303 /* Grab a reference on this inode's dirhash if it has one. */ 304 VI_LOCK(vp); 305 dh = ip->i_dirhash; 306 if (dh == NULL) { 307 VI_UNLOCK(vp); 308 return; 309 } 310 ufsdirhash_hold(dh); 311 VI_UNLOCK(vp); 312 313 /* Exclusively lock the dirhash. */ 314 sx_xlock(&dh->dh_lock); 315 316 /* If this dirhash still belongs to this inode, then free it. */ 317 VI_LOCK(vp); 318 if (ip->i_dirhash == dh) { 319 VI_UNLOCK(vp); 320 ufsdirhash_drop(dh); 321 break; 322 } 323 VI_UNLOCK(vp); 324 325 /* 326 * This inode's dirhash has changed while we were 327 * waiting for the dirhash lock, so try again. 328 */ 329 ufsdirhash_release(dh); 330 ufsdirhash_drop(dh); 331 } 332 ufsdirhash_free_locked(ip); 333 } 334 335 /* 336 * Attempt to build up a hash table for the directory contents in 337 * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 338 */ 339 int 340 ufsdirhash_build(struct inode *ip) 341 { 342 struct dirhash *dh; 343 struct buf *bp = NULL; 344 struct direct *ep; 345 struct vnode *vp; 346 doff_t bmask, pos; 347 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 348 349 /* Take care of a decreased sysctl value. */ 350 while (ufs_dirhashmem > ufs_dirhashmaxmem) { 351 if (ufsdirhash_recycle(0) != 0) 352 return (-1); 353 /* Recycled enough memory, so unlock the list. */ 354 DIRHASHLIST_UNLOCK(); 355 } 356 357 /* Check if we can/should use dirhash. */ 358 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) || 359 ip->i_effnlink == 0) { 360 if (ip->i_dirhash) 361 ufsdirhash_free(ip); 362 return (-1); 363 } 364 dh = ufsdirhash_create(ip); 365 if (dh == NULL) 366 return (-1); 367 if (dh->dh_hash != NULL) 368 return (0); 369 370 vp = ip->i_vnode; 371 /* Allocate 50% more entries than this dir size could ever need. */ 372 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 373 nslots = ip->i_size / DIRECTSIZ(1); 374 nslots = (nslots * 3 + 1) / 2; 375 narrays = howmany(nslots, DH_NBLKOFF); 376 nslots = narrays * DH_NBLKOFF; 377 dirblocks = howmany(ip->i_size, DIRBLKSIZ); 378 nblocks = (dirblocks * 3 + 1) / 2; 379 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 380 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 381 nblocks * sizeof(*dh->dh_blkfree); 382 DIRHASHLIST_LOCK(); 383 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 384 DIRHASHLIST_UNLOCK(); 385 if (memreqd > ufs_dirhashmaxmem / 2) 386 goto fail; 387 /* Try to free some space. */ 388 if (ufsdirhash_recycle(memreqd) != 0) 389 goto fail; 390 /* Enough was freed, and list has been locked. */ 391 } 392 ufs_dirhashmem += memreqd; 393 DIRHASHLIST_UNLOCK(); 394 395 /* Initialise the hash table and block statistics. */ 396 dh->dh_memreq = memreqd; 397 dh->dh_narrays = narrays; 398 dh->dh_hlen = nslots; 399 dh->dh_nblk = nblocks; 400 dh->dh_dirblks = dirblocks; 401 for (i = 0; i < DH_NFSTATS; i++) 402 dh->dh_firstfree[i] = -1; 403 dh->dh_firstfree[DH_NFSTATS] = 0; 404 dh->dh_hused = 0; 405 dh->dh_seqopt = 0; 406 dh->dh_seqoff = 0; 407 dh->dh_score = DH_SCOREINIT; 408 dh->dh_lastused = time_second; 409 410 /* 411 * Use non-blocking mallocs so that we will revert to a linear 412 * lookup on failure rather than potentially blocking forever. 413 */ 414 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]), 415 M_DIRHASH, M_NOWAIT | M_ZERO); 416 if (dh->dh_hash == NULL) 417 goto fail; 418 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]), 419 M_DIRHASH, M_NOWAIT); 420 if (dh->dh_blkfree == NULL) 421 goto fail; 422 for (i = 0; i < narrays; i++) { 423 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL) 424 goto fail; 425 for (j = 0; j < DH_NBLKOFF; j++) 426 dh->dh_hash[i][j] = DIRHASH_EMPTY; 427 } 428 for (i = 0; i < dirblocks; i++) 429 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 430 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 431 pos = 0; 432 while (pos < ip->i_size) { 433 /* If necessary, get the next directory block. */ 434 if ((pos & bmask) == 0) { 435 if (bp != NULL) 436 brelse(bp); 437 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 438 goto fail; 439 } 440 441 /* Add this entry to the hash. */ 442 ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 443 if (ep->d_reclen == 0 || ep->d_reclen > 444 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 445 /* Corrupted directory. */ 446 brelse(bp); 447 goto fail; 448 } 449 if (ep->d_ino != 0) { 450 /* Add the entry (simplified ufsdirhash_add). */ 451 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 452 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 453 slot = WRAPINCR(slot, dh->dh_hlen); 454 dh->dh_hused++; 455 DH_ENTRY(dh, slot) = pos; 456 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 457 } 458 pos += ep->d_reclen; 459 } 460 461 if (bp != NULL) 462 brelse(bp); 463 DIRHASHLIST_LOCK(); 464 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 465 dh->dh_onlist = 1; 466 DIRHASHLIST_UNLOCK(); 467 sx_downgrade(&dh->dh_lock); 468 return (0); 469 470 fail: 471 ufsdirhash_free_locked(ip); 472 return (-1); 473 } 474 475 /* 476 * Free any hash table associated with inode 'ip'. 477 */ 478 static void 479 ufsdirhash_free_locked(struct inode *ip) 480 { 481 struct dirhash *dh; 482 struct vnode *vp; 483 int i; 484 485 DIRHASH_ASSERT_LOCKED(ip->i_dirhash); 486 487 /* 488 * Clear the pointer in the inode to prevent new threads from 489 * finding the dead structure. 490 */ 491 vp = ip->i_vnode; 492 VI_LOCK(vp); 493 dh = ip->i_dirhash; 494 ip->i_dirhash = NULL; 495 VI_UNLOCK(vp); 496 497 /* 498 * Remove the hash from the list since we are going to free its 499 * memory. 500 */ 501 DIRHASHLIST_LOCK(); 502 if (dh->dh_onlist) 503 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 504 ufs_dirhashmem -= dh->dh_memreq; 505 DIRHASHLIST_UNLOCK(); 506 507 /* 508 * At this point, any waiters for the lock should hold their 509 * own reference on the dirhash structure. They will drop 510 * that reference once they grab the vnode interlock and see 511 * that ip->i_dirhash is NULL. 512 */ 513 sx_xunlock(&dh->dh_lock); 514 515 /* 516 * Handle partially recycled as well as fully constructed hashes. 517 */ 518 if (dh->dh_hash != NULL) { 519 for (i = 0; i < dh->dh_narrays; i++) 520 if (dh->dh_hash[i] != NULL) 521 DIRHASH_BLKFREE(dh->dh_hash[i]); 522 free(dh->dh_hash, M_DIRHASH); 523 if (dh->dh_blkfree != NULL) 524 free(dh->dh_blkfree, M_DIRHASH); 525 } 526 527 /* 528 * Drop the inode's reference to the data structure. 529 */ 530 ufsdirhash_drop(dh); 531 } 532 533 /* 534 * Find the offset of the specified name within the given inode. 535 * Returns 0 on success, ENOENT if the entry does not exist, or 536 * EJUSTRETURN if the caller should revert to a linear search. 537 * 538 * If successful, the directory offset is stored in *offp, and a 539 * pointer to a struct buf containing the entry is stored in *bpp. If 540 * prevoffp is non-NULL, the offset of the previous entry within 541 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 542 * is the first in a block, the start of the block is used). 543 * 544 * Must be called with the hash locked. Returns with the hash unlocked. 545 */ 546 int 547 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 548 struct buf **bpp, doff_t *prevoffp) 549 { 550 struct dirhash *dh, *dh_next; 551 struct direct *dp; 552 struct vnode *vp; 553 struct buf *bp; 554 doff_t blkoff, bmask, offset, prevoff; 555 int i, slot; 556 int error; 557 558 dh = ip->i_dirhash; 559 KASSERT(dh != NULL && dh->dh_hash != NULL, 560 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh)); 561 DIRHASH_ASSERT_LOCKED(dh); 562 /* 563 * Move this dirhash towards the end of the list if it has a 564 * score higher than the next entry, and acquire the dh_lock. 565 */ 566 DIRHASHLIST_LOCK(); 567 if (TAILQ_NEXT(dh, dh_list) != NULL) { 568 /* 569 * If the new score will be greater than that of the next 570 * entry, then move this entry past it. With both mutexes 571 * held, dh_next won't go away, but its dh_score could 572 * change; that's not important since it is just a hint. 573 */ 574 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 575 dh->dh_score >= dh_next->dh_score) { 576 KASSERT(dh->dh_onlist, ("dirhash: not on list")); 577 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 578 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 579 dh_list); 580 } 581 } 582 /* Update the score. */ 583 if (dh->dh_score < DH_SCOREMAX) 584 dh->dh_score++; 585 586 /* Update last used time. */ 587 dh->dh_lastused = time_second; 588 DIRHASHLIST_UNLOCK(); 589 590 vp = ip->i_vnode; 591 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 592 blkoff = -1; 593 bp = NULL; 594 restart: 595 slot = ufsdirhash_hash(dh, name, namelen); 596 597 if (dh->dh_seqopt) { 598 /* 599 * Sequential access optimisation. dh_seqoff contains the 600 * offset of the directory entry immediately following 601 * the last entry that was looked up. Check if this offset 602 * appears in the hash chain for the name we are looking for. 603 */ 604 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 605 i = WRAPINCR(i, dh->dh_hlen)) 606 if (offset == dh->dh_seqoff) 607 break; 608 if (offset == dh->dh_seqoff) { 609 /* 610 * We found an entry with the expected offset. This 611 * is probably the entry we want, but if not, the 612 * code below will turn off seqopt and retry. 613 */ 614 slot = i; 615 } else 616 dh->dh_seqopt = 0; 617 } 618 619 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 620 slot = WRAPINCR(slot, dh->dh_hlen)) { 621 if (offset == DIRHASH_DEL) 622 continue; 623 if (offset < 0 || offset >= ip->i_size) 624 panic("ufsdirhash_lookup: bad offset in hash array"); 625 if ((offset & ~bmask) != blkoff) { 626 if (bp != NULL) 627 brelse(bp); 628 blkoff = offset & ~bmask; 629 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) { 630 error = EJUSTRETURN; 631 goto fail; 632 } 633 } 634 KASSERT(bp != NULL, ("no buffer allocated")); 635 dp = (struct direct *)(bp->b_data + (offset & bmask)); 636 if (dp->d_reclen == 0 || dp->d_reclen > 637 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 638 /* Corrupted directory. */ 639 error = EJUSTRETURN; 640 goto fail; 641 } 642 if (dp->d_namlen == namelen && 643 bcmp(dp->d_name, name, namelen) == 0) { 644 /* Found. Get the prev offset if needed. */ 645 if (prevoffp != NULL) { 646 if (offset & (DIRBLKSIZ - 1)) { 647 prevoff = ufsdirhash_getprev(dp, 648 offset); 649 if (prevoff == -1) { 650 error = EJUSTRETURN; 651 goto fail; 652 } 653 } else 654 prevoff = offset; 655 *prevoffp = prevoff; 656 } 657 658 /* Check for sequential access, and update offset. */ 659 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 660 dh->dh_seqopt = 1; 661 dh->dh_seqoff = offset + DIRSIZ(0, dp); 662 *bpp = bp; 663 *offp = offset; 664 ufsdirhash_release(dh); 665 return (0); 666 } 667 668 /* 669 * When the name doesn't match in the seqopt case, go back 670 * and search normally. 671 */ 672 if (dh->dh_seqopt) { 673 dh->dh_seqopt = 0; 674 goto restart; 675 } 676 } 677 error = ENOENT; 678 fail: 679 ufsdirhash_release(dh); 680 if (bp != NULL) 681 brelse(bp); 682 return (error); 683 } 684 685 /* 686 * Find a directory block with room for 'slotneeded' bytes. Returns 687 * the offset of the directory entry that begins the free space. 688 * This will either be the offset of an existing entry that has free 689 * space at the end, or the offset of an entry with d_ino == 0 at 690 * the start of a DIRBLKSIZ block. 691 * 692 * To use the space, the caller may need to compact existing entries in 693 * the directory. The total number of bytes in all of the entries involved 694 * in the compaction is stored in *slotsize. In other words, all of 695 * the entries that must be compacted are exactly contained in the 696 * region beginning at the returned offset and spanning *slotsize bytes. 697 * 698 * Returns -1 if no space was found, indicating that the directory 699 * must be extended. 700 */ 701 doff_t 702 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 703 { 704 struct direct *dp; 705 struct dirhash *dh; 706 struct buf *bp; 707 doff_t pos, slotstart; 708 int dirblock, error, freebytes, i; 709 710 dh = ip->i_dirhash; 711 KASSERT(dh != NULL && dh->dh_hash != NULL, 712 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh)); 713 DIRHASH_ASSERT_LOCKED(dh); 714 715 /* Find a directory block with the desired free space. */ 716 dirblock = -1; 717 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 718 if ((dirblock = dh->dh_firstfree[i]) != -1) 719 break; 720 if (dirblock == -1) 721 return (-1); 722 723 KASSERT(dirblock < dh->dh_nblk && 724 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 725 ("ufsdirhash_findfree: bad stats")); 726 pos = dirblock * DIRBLKSIZ; 727 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 728 if (error) 729 return (-1); 730 731 /* Find the first entry with free space. */ 732 for (i = 0; i < DIRBLKSIZ; ) { 733 if (dp->d_reclen == 0) { 734 brelse(bp); 735 return (-1); 736 } 737 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 738 break; 739 i += dp->d_reclen; 740 dp = (struct direct *)((char *)dp + dp->d_reclen); 741 } 742 if (i > DIRBLKSIZ) { 743 brelse(bp); 744 return (-1); 745 } 746 slotstart = pos + i; 747 748 /* Find the range of entries needed to get enough space */ 749 freebytes = 0; 750 while (i < DIRBLKSIZ && freebytes < slotneeded) { 751 freebytes += dp->d_reclen; 752 if (dp->d_ino != 0) 753 freebytes -= DIRSIZ(0, dp); 754 if (dp->d_reclen == 0) { 755 brelse(bp); 756 return (-1); 757 } 758 i += dp->d_reclen; 759 dp = (struct direct *)((char *)dp + dp->d_reclen); 760 } 761 if (i > DIRBLKSIZ) { 762 brelse(bp); 763 return (-1); 764 } 765 if (freebytes < slotneeded) 766 panic("ufsdirhash_findfree: free mismatch"); 767 brelse(bp); 768 *slotsize = pos + i - slotstart; 769 return (slotstart); 770 } 771 772 /* 773 * Return the start of the unused space at the end of a directory, or 774 * -1 if there are no trailing unused blocks. 775 */ 776 doff_t 777 ufsdirhash_enduseful(struct inode *ip) 778 { 779 780 struct dirhash *dh; 781 int i; 782 783 dh = ip->i_dirhash; 784 DIRHASH_ASSERT_LOCKED(dh); 785 KASSERT(dh != NULL && dh->dh_hash != NULL, 786 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh)); 787 788 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) 789 return (-1); 790 791 for (i = dh->dh_dirblks - 1; i >= 0; i--) 792 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 793 break; 794 795 return ((doff_t)(i + 1) * DIRBLKSIZ); 796 } 797 798 /* 799 * Insert information into the hash about a new directory entry. dirp 800 * points to a struct direct containing the entry, and offset specifies 801 * the offset of this entry. 802 */ 803 void 804 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 805 { 806 struct dirhash *dh; 807 int slot; 808 809 if ((dh = ufsdirhash_acquire(ip)) == NULL) 810 return; 811 812 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 813 ("ufsdirhash_add: bad offset")); 814 /* 815 * Normal hash usage is < 66%. If the usage gets too high then 816 * remove the hash entirely and let it be rebuilt later. 817 */ 818 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 819 ufsdirhash_free_locked(ip); 820 return; 821 } 822 823 /* Find a free hash slot (empty or deleted), and add the entry. */ 824 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 825 while (DH_ENTRY(dh, slot) >= 0) 826 slot = WRAPINCR(slot, dh->dh_hlen); 827 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 828 dh->dh_hused++; 829 DH_ENTRY(dh, slot) = offset; 830 831 /* Update last used time. */ 832 dh->dh_lastused = time_second; 833 834 /* Update the per-block summary info. */ 835 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 836 ufsdirhash_release(dh); 837 } 838 839 /* 840 * Remove the specified directory entry from the hash. The entry to remove 841 * is defined by the name in `dirp', which must exist at the specified 842 * `offset' within the directory. 843 */ 844 void 845 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 846 { 847 struct dirhash *dh; 848 int slot; 849 850 if ((dh = ufsdirhash_acquire(ip)) == NULL) 851 return; 852 853 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 854 ("ufsdirhash_remove: bad offset")); 855 /* Find the entry */ 856 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 857 858 /* Remove the hash entry. */ 859 ufsdirhash_delslot(dh, slot); 860 861 /* Update the per-block summary info. */ 862 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 863 ufsdirhash_release(dh); 864 } 865 866 /* 867 * Change the offset associated with a directory entry in the hash. Used 868 * when compacting directory blocks. 869 */ 870 void 871 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 872 doff_t newoff) 873 { 874 struct dirhash *dh; 875 int slot; 876 877 if ((dh = ufsdirhash_acquire(ip)) == NULL) 878 return; 879 880 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 881 newoff < dh->dh_dirblks * DIRBLKSIZ, 882 ("ufsdirhash_move: bad offset")); 883 /* Find the entry, and update the offset. */ 884 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 885 DH_ENTRY(dh, slot) = newoff; 886 ufsdirhash_release(dh); 887 } 888 889 /* 890 * Inform dirhash that the directory has grown by one block that 891 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 892 */ 893 void 894 ufsdirhash_newblk(struct inode *ip, doff_t offset) 895 { 896 struct dirhash *dh; 897 int block; 898 899 if ((dh = ufsdirhash_acquire(ip)) == NULL) 900 return; 901 902 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 903 ("ufsdirhash_newblk: bad offset")); 904 block = offset / DIRBLKSIZ; 905 if (block >= dh->dh_nblk) { 906 /* Out of space; must rebuild. */ 907 ufsdirhash_free_locked(ip); 908 return; 909 } 910 dh->dh_dirblks = block + 1; 911 912 /* Account for the new free block. */ 913 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 914 if (dh->dh_firstfree[DH_NFSTATS] == -1) 915 dh->dh_firstfree[DH_NFSTATS] = block; 916 ufsdirhash_release(dh); 917 } 918 919 /* 920 * Inform dirhash that the directory is being truncated. 921 */ 922 void 923 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 924 { 925 struct dirhash *dh; 926 int block, i; 927 928 if ((dh = ufsdirhash_acquire(ip)) == NULL) 929 return; 930 931 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 932 ("ufsdirhash_dirtrunc: bad offset")); 933 block = howmany(offset, DIRBLKSIZ); 934 /* 935 * If the directory shrinks to less than 1/8 of dh_nblk blocks 936 * (about 20% of its original size due to the 50% extra added in 937 * ufsdirhash_build) then free it, and let the caller rebuild 938 * if necessary. 939 */ 940 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 941 ufsdirhash_free_locked(ip); 942 return; 943 } 944 945 /* 946 * Remove any `first free' information pertaining to the 947 * truncated blocks. All blocks we're removing should be 948 * completely unused. 949 */ 950 if (dh->dh_firstfree[DH_NFSTATS] >= block) 951 dh->dh_firstfree[DH_NFSTATS] = -1; 952 for (i = block; i < dh->dh_dirblks; i++) 953 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 954 panic("ufsdirhash_dirtrunc: blocks in use"); 955 for (i = 0; i < DH_NFSTATS; i++) 956 if (dh->dh_firstfree[i] >= block) 957 panic("ufsdirhash_dirtrunc: first free corrupt"); 958 dh->dh_dirblks = block; 959 ufsdirhash_release(dh); 960 } 961 962 /* 963 * Debugging function to check that the dirhash information about 964 * a directory block matches its actual contents. Panics if a mismatch 965 * is detected. 966 * 967 * On entry, `buf' should point to the start of an in-core 968 * DIRBLKSIZ-sized directory block, and `offset' should contain the 969 * offset from the start of the directory of that block. 970 */ 971 void 972 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 973 { 974 struct dirhash *dh; 975 struct direct *dp; 976 int block, ffslot, i, nfree; 977 978 if (!ufs_dirhashcheck) 979 return; 980 if ((dh = ufsdirhash_acquire(ip)) == NULL) 981 return; 982 983 block = offset / DIRBLKSIZ; 984 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 985 panic("ufsdirhash_checkblock: bad offset"); 986 987 nfree = 0; 988 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 989 dp = (struct direct *)(buf + i); 990 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 991 panic("ufsdirhash_checkblock: bad dir"); 992 993 if (dp->d_ino == 0) { 994 #if 0 995 /* 996 * XXX entries with d_ino == 0 should only occur 997 * at the start of a DIRBLKSIZ block. However the 998 * ufs code is tolerant of such entries at other 999 * offsets, and fsck does not fix them. 1000 */ 1001 if (i != 0) 1002 panic("ufsdirhash_checkblock: bad dir inode"); 1003 #endif 1004 nfree += dp->d_reclen; 1005 continue; 1006 } 1007 1008 /* Check that the entry exists (will panic if it doesn't). */ 1009 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 1010 1011 nfree += dp->d_reclen - DIRSIZ(0, dp); 1012 } 1013 if (i != DIRBLKSIZ) 1014 panic("ufsdirhash_checkblock: bad dir end"); 1015 1016 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 1017 panic("ufsdirhash_checkblock: bad free count"); 1018 1019 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 1020 for (i = 0; i <= DH_NFSTATS; i++) 1021 if (dh->dh_firstfree[i] == block && i != ffslot) 1022 panic("ufsdirhash_checkblock: bad first-free"); 1023 if (dh->dh_firstfree[ffslot] == -1) 1024 panic("ufsdirhash_checkblock: missing first-free entry"); 1025 ufsdirhash_release(dh); 1026 } 1027 1028 /* 1029 * Hash the specified filename into a dirhash slot. 1030 */ 1031 static int 1032 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 1033 { 1034 u_int32_t hash; 1035 1036 /* 1037 * We hash the name and then some other bit of data that is 1038 * invariant over the dirhash's lifetime. Otherwise names 1039 * differing only in the last byte are placed close to one 1040 * another in the table, which is bad for linear probing. 1041 */ 1042 hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 1043 hash = fnv_32_buf(&dh, sizeof(dh), hash); 1044 return (hash % dh->dh_hlen); 1045 } 1046 1047 /* 1048 * Adjust the number of free bytes in the block containing `offset' 1049 * by the value specified by `diff'. 1050 * 1051 * The caller must ensure we have exclusive access to `dh'; normally 1052 * that means that dh_lock should be held, but this is also called 1053 * from ufsdirhash_build() where exclusive access can be assumed. 1054 */ 1055 static void 1056 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 1057 { 1058 int block, i, nfidx, ofidx; 1059 1060 /* Update the per-block summary info. */ 1061 block = offset / DIRBLKSIZ; 1062 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 1063 ("dirhash bad offset")); 1064 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1065 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 1066 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1067 1068 /* Update the `first free' list if necessary. */ 1069 if (ofidx != nfidx) { 1070 /* If removing, scan forward for the next block. */ 1071 if (dh->dh_firstfree[ofidx] == block) { 1072 for (i = block + 1; i < dh->dh_dirblks; i++) 1073 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 1074 break; 1075 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 1076 } 1077 1078 /* Make this the new `first free' if necessary */ 1079 if (dh->dh_firstfree[nfidx] > block || 1080 dh->dh_firstfree[nfidx] == -1) 1081 dh->dh_firstfree[nfidx] = block; 1082 } 1083 } 1084 1085 /* 1086 * Find the specified name which should have the specified offset. 1087 * Returns a slot number, and panics on failure. 1088 * 1089 * `dh' must be locked on entry and remains so on return. 1090 */ 1091 static int 1092 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 1093 { 1094 int slot; 1095 1096 DIRHASH_ASSERT_LOCKED(dh); 1097 1098 /* Find the entry. */ 1099 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 1100 slot = ufsdirhash_hash(dh, name, namelen); 1101 while (DH_ENTRY(dh, slot) != offset && 1102 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 1103 slot = WRAPINCR(slot, dh->dh_hlen); 1104 if (DH_ENTRY(dh, slot) != offset) 1105 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 1106 1107 return (slot); 1108 } 1109 1110 /* 1111 * Remove the entry corresponding to the specified slot from the hash array. 1112 * 1113 * `dh' must be locked on entry and remains so on return. 1114 */ 1115 static void 1116 ufsdirhash_delslot(struct dirhash *dh, int slot) 1117 { 1118 int i; 1119 1120 DIRHASH_ASSERT_LOCKED(dh); 1121 1122 /* Mark the entry as deleted. */ 1123 DH_ENTRY(dh, slot) = DIRHASH_DEL; 1124 1125 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 1126 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 1127 i = WRAPINCR(i, dh->dh_hlen); 1128 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 1129 i = WRAPDECR(i, dh->dh_hlen); 1130 while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 1131 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 1132 dh->dh_hused--; 1133 i = WRAPDECR(i, dh->dh_hlen); 1134 } 1135 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 1136 } 1137 } 1138 1139 /* 1140 * Given a directory entry and its offset, find the offset of the 1141 * previous entry in the same DIRBLKSIZ-sized block. Returns an 1142 * offset, or -1 if there is no previous entry in the block or some 1143 * other problem occurred. 1144 */ 1145 static doff_t 1146 ufsdirhash_getprev(struct direct *dirp, doff_t offset) 1147 { 1148 struct direct *dp; 1149 char *blkbuf; 1150 doff_t blkoff, prevoff; 1151 int entrypos, i; 1152 1153 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 1154 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 1155 blkbuf = (char *)dirp - entrypos; 1156 prevoff = blkoff; 1157 1158 /* If `offset' is the start of a block, there is no previous entry. */ 1159 if (entrypos == 0) 1160 return (-1); 1161 1162 /* Scan from the start of the block until we get to the entry. */ 1163 for (i = 0; i < entrypos; i += dp->d_reclen) { 1164 dp = (struct direct *)(blkbuf + i); 1165 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 1166 return (-1); /* Corrupted directory. */ 1167 prevoff = blkoff + i; 1168 } 1169 return (prevoff); 1170 } 1171 1172 /* 1173 * Delete the given dirhash and reclaim its memory. Assumes that 1174 * ufsdirhash_list is locked, and leaves it locked. Also assumes 1175 * that dh is locked. Returns the amount of memory freed. 1176 */ 1177 static int 1178 ufsdirhash_destroy(struct dirhash *dh) 1179 { 1180 doff_t **hash; 1181 u_int8_t *blkfree; 1182 int i, mem, narrays; 1183 1184 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 1185 1186 /* Remove it from the list and detach its memory. */ 1187 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 1188 dh->dh_onlist = 0; 1189 hash = dh->dh_hash; 1190 dh->dh_hash = NULL; 1191 blkfree = dh->dh_blkfree; 1192 dh->dh_blkfree = NULL; 1193 narrays = dh->dh_narrays; 1194 mem = dh->dh_memreq; 1195 dh->dh_memreq = 0; 1196 1197 /* Unlock dirhash and free the detached memory. */ 1198 ufsdirhash_release(dh); 1199 for (i = 0; i < narrays; i++) 1200 DIRHASH_BLKFREE(hash[i]); 1201 free(hash, M_DIRHASH); 1202 free(blkfree, M_DIRHASH); 1203 1204 /* Account for the returned memory. */ 1205 ufs_dirhashmem -= mem; 1206 1207 return (mem); 1208 } 1209 1210 /* 1211 * Try to free up `wanted' bytes by stealing memory from existing 1212 * dirhashes. Returns zero with list locked if successful. 1213 */ 1214 static int 1215 ufsdirhash_recycle(int wanted) 1216 { 1217 struct dirhash *dh; 1218 1219 DIRHASHLIST_LOCK(); 1220 dh = TAILQ_FIRST(&ufsdirhash_list); 1221 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 1222 /* Decrement the score; only recycle if it becomes zero. */ 1223 if (dh == NULL || --dh->dh_score > 0) { 1224 DIRHASHLIST_UNLOCK(); 1225 return (-1); 1226 } 1227 /* 1228 * If we can't lock it it's in use and we don't want to 1229 * recycle it anyway. 1230 */ 1231 if (!sx_try_xlock(&dh->dh_lock)) { 1232 dh = TAILQ_NEXT(dh, dh_list); 1233 continue; 1234 } 1235 1236 ufsdirhash_destroy(dh); 1237 1238 /* Repeat if necessary. */ 1239 dh = TAILQ_FIRST(&ufsdirhash_list); 1240 } 1241 /* Success; return with list locked. */ 1242 return (0); 1243 } 1244 1245 /* 1246 * Callback that frees some dirhashes when the system is low on virtual memory. 1247 */ 1248 static void 1249 ufsdirhash_lowmem() 1250 { 1251 struct dirhash *dh, *dh_temp; 1252 int memfreed = 0; 1253 /* XXX: this 10% may need to be adjusted */ 1254 int memwanted = ufs_dirhashmem / 10; 1255 1256 ufs_dirhashlowmemcount++; 1257 1258 DIRHASHLIST_LOCK(); 1259 /* 1260 * Delete dirhashes not used for more than ufs_dirhashreclaimage 1261 * seconds. If we can't get a lock on the dirhash, it will be skipped. 1262 */ 1263 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) { 1264 if (!sx_try_xlock(&dh->dh_lock)) 1265 continue; 1266 if (time_second - dh->dh_lastused > ufs_dirhashreclaimage) 1267 memfreed += ufsdirhash_destroy(dh); 1268 /* Unlock if we didn't delete the dirhash */ 1269 else 1270 ufsdirhash_release(dh); 1271 } 1272 1273 /* 1274 * If not enough memory was freed, keep deleting hashes from the head 1275 * of the dirhash list. The ones closest to the head should be the 1276 * oldest. 1277 */ 1278 if (memfreed < memwanted) { 1279 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) { 1280 if (!sx_try_xlock(&dh->dh_lock)) 1281 continue; 1282 memfreed += ufsdirhash_destroy(dh); 1283 if (memfreed >= memwanted) 1284 break; 1285 } 1286 } 1287 DIRHASHLIST_UNLOCK(); 1288 } 1289 1290 1291 void 1292 ufsdirhash_init() 1293 { 1294 ufs_dirhashmaxmem = lmax(roundup(hibufspace / 64, PAGE_SIZE), 1295 2 * 1024 * 1024); 1296 1297 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t), 1298 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1299 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF); 1300 TAILQ_INIT(&ufsdirhash_list); 1301 1302 /* Register a callback function to handle low memory signals */ 1303 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL, 1304 EVENTHANDLER_PRI_FIRST); 1305 } 1306 1307 void 1308 ufsdirhash_uninit() 1309 { 1310 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit")); 1311 uma_zdestroy(ufsdirhash_zone); 1312 mtx_destroy(&ufsdirhash_mtx); 1313 } 1314 1315 #endif /* UFS_DIRHASH */ 1316