ufs_dirhash.c revision 207141
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 207141 2010-04-24 07:05:35Z jeff $"); 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 69static MALLOC_DEFINE(M_DIRHASH, "ufs_dirhash", "UFS directory hash tables"); 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"); 84static int ufs_dirhashlowmemcount = 0; 85SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_lowmemcount, CTLFLAG_RD, 86 &ufs_dirhashlowmemcount, 0, "number of times low memory hook called"); 87static int ufs_dirhashreclaimage = 5; 88SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_reclaimage, CTLFLAG_RW, 89 &ufs_dirhashreclaimage, 0, 90 "max time in seconds of hash inactivity before deletion in low VM events"); 91 92 93static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 94static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 95static void ufsdirhash_delslot(struct dirhash *dh, int slot); 96static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 97 doff_t offset); 98static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 99static int ufsdirhash_recycle(int wanted); 100static void ufsdirhash_lowmem(void); 101static void ufsdirhash_free_locked(struct inode *ip); 102 103static uma_zone_t ufsdirhash_zone; 104 105#define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx) 106#define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx) 107#define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK) 108#define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr)) 109#define DIRHASH_ASSERT_LOCKED(dh) \ 110 sx_assert(&(dh)->dh_lock, SA_LOCKED) 111 112/* Dirhash list; recently-used entries are near the tail. */ 113static TAILQ_HEAD(, dirhash) ufsdirhash_list; 114 115/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 116static struct mtx ufsdirhash_mtx; 117 118/* 119 * Locking: 120 * 121 * The relationship between inode and dirhash is protected either by an 122 * exclusive vnode lock or the vnode interlock where a shared vnode lock 123 * may be used. The dirhash_mtx is acquired after the dirhash lock. To 124 * handle teardown races, code wishing to lock the dirhash for an inode 125 * when using a shared vnode lock must obtain a private reference on the 126 * dirhash while holding the vnode interlock. They can drop it once they 127 * have obtained the dirhash lock and verified that the dirhash wasn't 128 * recycled while they waited for the dirhash lock. 129 * 130 * ufsdirhash_build() acquires a shared lock on the dirhash when it is 131 * successful. This lock is released after a call to ufsdirhash_lookup(). 132 * 133 * Functions requiring exclusive access use ufsdirhash_acquire() which may 134 * free a dirhash structure that was recycled by ufsdirhash_recycle(). 135 * 136 * The dirhash lock may be held across io operations. 137 * 138 * WITNESS reports a lock order reversal between the "bufwait" lock 139 * and the "dirhash" lock. However, this specific reversal will not 140 * cause a deadlock. To get a deadlock, one would have to lock a 141 * buffer followed by the dirhash while a second thread locked a 142 * buffer while holding the dirhash lock. The second order can happen 143 * under a shared or exclusive vnode lock for the associated directory 144 * in lookup(). The first order, however, can only happen under an 145 * exclusive vnode lock (e.g. unlink(), rename(), etc.). Thus, for 146 * a thread to be doing a "bufwait" -> "dirhash" order, it has to hold 147 * an exclusive vnode lock. That exclusive vnode lock will prevent 148 * any other threads from doing a "dirhash" -> "bufwait" order. 149 */ 150 151static void 152ufsdirhash_hold(struct dirhash *dh) 153{ 154 155 refcount_acquire(&dh->dh_refcount); 156} 157 158static void 159ufsdirhash_drop(struct dirhash *dh) 160{ 161 162 if (refcount_release(&dh->dh_refcount)) { 163 sx_destroy(&dh->dh_lock); 164 free(dh, M_DIRHASH); 165 } 166} 167 168/* 169 * Release the lock on a dirhash. 170 */ 171static void 172ufsdirhash_release(struct dirhash *dh) 173{ 174 175 sx_unlock(&dh->dh_lock); 176} 177 178/* 179 * Either acquire an existing hash locked shared or create a new hash and 180 * return it exclusively locked. May return NULL if the allocation fails. 181 * 182 * The vnode interlock is used to protect the i_dirhash pointer from 183 * simultaneous access while only a shared vnode lock is held. 184 */ 185static struct dirhash * 186ufsdirhash_create(struct inode *ip) 187{ 188 struct dirhash *ndh; 189 struct dirhash *dh; 190 struct vnode *vp; 191 int error; 192 193 error = 0; 194 ndh = dh = NULL; 195 vp = ip->i_vnode; 196 for (;;) { 197 /* Racy check for i_dirhash to prefetch a dirhash structure. */ 198 if (ip->i_dirhash == NULL && ndh == NULL) { 199 ndh = malloc(sizeof *dh, M_DIRHASH, 200 M_NOWAIT | M_ZERO); 201 if (ndh == NULL) 202 return (NULL); 203 refcount_init(&ndh->dh_refcount, 1); 204 205 /* 206 * The DUPOK is to prevent warnings from the 207 * sx_slock() a few lines down which is safe 208 * since the duplicate lock in that case is 209 * the one for this dirhash we are creating 210 * now which has no external references until 211 * after this function returns. 212 */ 213 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK); 214 sx_xlock(&ndh->dh_lock); 215 } 216 /* 217 * Check i_dirhash. If it's NULL just try to use a 218 * preallocated structure. If none exists loop and try again. 219 */ 220 VI_LOCK(vp); 221 dh = ip->i_dirhash; 222 if (dh == NULL) { 223 ip->i_dirhash = ndh; 224 VI_UNLOCK(vp); 225 if (ndh == NULL) 226 continue; 227 return (ndh); 228 } 229 ufsdirhash_hold(dh); 230 VI_UNLOCK(vp); 231 232 /* Acquire a shared lock on existing hashes. */ 233 sx_slock(&dh->dh_lock); 234 235 /* The hash could've been recycled while we were waiting. */ 236 VI_LOCK(vp); 237 if (ip->i_dirhash != dh) { 238 VI_UNLOCK(vp); 239 ufsdirhash_release(dh); 240 ufsdirhash_drop(dh); 241 continue; 242 } 243 VI_UNLOCK(vp); 244 ufsdirhash_drop(dh); 245 246 /* If the hash is still valid we've succeeded. */ 247 if (dh->dh_hash != NULL) 248 break; 249 /* 250 * If the hash is NULL it has been recycled. Try to upgrade 251 * so we can recreate it. If we fail the upgrade, drop our 252 * lock and try again. 253 */ 254 if (sx_try_upgrade(&dh->dh_lock)) 255 break; 256 sx_sunlock(&dh->dh_lock); 257 } 258 /* Free the preallocated structure if it was not necessary. */ 259 if (ndh) { 260 ufsdirhash_release(ndh); 261 ufsdirhash_drop(ndh); 262 } 263 return (dh); 264} 265 266/* 267 * Acquire an exclusive lock on an existing hash. Requires an exclusive 268 * vnode lock to protect the i_dirhash pointer. hashes that have been 269 * recycled are reclaimed here and NULL is returned. 270 */ 271static struct dirhash * 272ufsdirhash_acquire(struct inode *ip) 273{ 274 struct dirhash *dh; 275 struct vnode *vp; 276 277 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__); 278 279 vp = ip->i_vnode; 280 dh = ip->i_dirhash; 281 if (dh == NULL) 282 return (NULL); 283 sx_xlock(&dh->dh_lock); 284 if (dh->dh_hash != NULL) 285 return (dh); 286 ufsdirhash_free_locked(ip); 287 return (NULL); 288} 289 290/* 291 * Acquire exclusively and free the hash pointed to by ip. Works with a 292 * shared or exclusive vnode lock. 293 */ 294void 295ufsdirhash_free(struct inode *ip) 296{ 297 struct dirhash *dh; 298 struct vnode *vp; 299 300 vp = ip->i_vnode; 301 for (;;) { 302 /* Grab a reference on this inode's dirhash if it has one. */ 303 VI_LOCK(vp); 304 dh = ip->i_dirhash; 305 if (dh == NULL) { 306 VI_UNLOCK(vp); 307 return; 308 } 309 ufsdirhash_hold(dh); 310 VI_UNLOCK(vp); 311 312 /* Exclusively lock the dirhash. */ 313 sx_xlock(&dh->dh_lock); 314 315 /* If this dirhash still belongs to this inode, then free it. */ 316 VI_LOCK(vp); 317 if (ip->i_dirhash == dh) { 318 VI_UNLOCK(vp); 319 ufsdirhash_drop(dh); 320 break; 321 } 322 VI_UNLOCK(vp); 323 324 /* 325 * This inode's dirhash has changed while we were 326 * waiting for the dirhash lock, so try again. 327 */ 328 ufsdirhash_release(dh); 329 ufsdirhash_drop(dh); 330 } 331 ufsdirhash_free_locked(ip); 332} 333 334/* 335 * Attempt to build up a hash table for the directory contents in 336 * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 337 */ 338int 339ufsdirhash_build(struct inode *ip) 340{ 341 struct dirhash *dh; 342 struct buf *bp = NULL; 343 struct direct *ep; 344 struct vnode *vp; 345 doff_t bmask, pos; 346 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 347 348 /* Take care of a decreased sysctl value. */ 349 while (ufs_dirhashmem > ufs_dirhashmaxmem) { 350 if (ufsdirhash_recycle(0) != 0) 351 return (-1); 352 /* Recycled enough memory, so unlock the list. */ 353 DIRHASHLIST_UNLOCK(); 354 } 355 356 /* Check if we can/should use dirhash. */ 357 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) || 358 ip->i_effnlink == 0) { 359 if (ip->i_dirhash) 360 ufsdirhash_free(ip); 361 return (-1); 362 } 363 dh = ufsdirhash_create(ip); 364 if (dh == NULL) 365 return (-1); 366 if (dh->dh_hash != NULL) 367 return (0); 368 369 vp = ip->i_vnode; 370 /* Allocate 50% more entries than this dir size could ever need. */ 371 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 372 nslots = ip->i_size / DIRECTSIZ(1); 373 nslots = (nslots * 3 + 1) / 2; 374 narrays = howmany(nslots, DH_NBLKOFF); 375 nslots = narrays * DH_NBLKOFF; 376 dirblocks = howmany(ip->i_size, DIRBLKSIZ); 377 nblocks = (dirblocks * 3 + 1) / 2; 378 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 379 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 380 nblocks * sizeof(*dh->dh_blkfree); 381 DIRHASHLIST_LOCK(); 382 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 383 DIRHASHLIST_UNLOCK(); 384 if (memreqd > ufs_dirhashmaxmem / 2) 385 goto fail; 386 /* Try to free some space. */ 387 if (ufsdirhash_recycle(memreqd) != 0) 388 goto fail; 389 /* Enough was freed, and list has been locked. */ 390 } 391 ufs_dirhashmem += memreqd; 392 DIRHASHLIST_UNLOCK(); 393 394 /* Initialise the hash table and block statistics. */ 395 dh->dh_memreq = memreqd; 396 dh->dh_narrays = narrays; 397 dh->dh_hlen = nslots; 398 dh->dh_nblk = nblocks; 399 dh->dh_dirblks = dirblocks; 400 for (i = 0; i < DH_NFSTATS; i++) 401 dh->dh_firstfree[i] = -1; 402 dh->dh_firstfree[DH_NFSTATS] = 0; 403 dh->dh_hused = 0; 404 dh->dh_seqopt = 0; 405 dh->dh_seqoff = 0; 406 dh->dh_score = DH_SCOREINIT; 407 dh->dh_lastused = time_second; 408 409 /* 410 * Use non-blocking mallocs so that we will revert to a linear 411 * lookup on failure rather than potentially blocking forever. 412 */ 413 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]), 414 M_DIRHASH, M_NOWAIT | M_ZERO); 415 if (dh->dh_hash == NULL) 416 goto fail; 417 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]), 418 M_DIRHASH, M_NOWAIT); 419 if (dh->dh_blkfree == NULL) 420 goto fail; 421 for (i = 0; i < narrays; i++) { 422 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL) 423 goto fail; 424 for (j = 0; j < DH_NBLKOFF; j++) 425 dh->dh_hash[i][j] = DIRHASH_EMPTY; 426 } 427 for (i = 0; i < dirblocks; i++) 428 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 429 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 430 pos = 0; 431 while (pos < ip->i_size) { 432 /* If necessary, get the next directory block. */ 433 if ((pos & bmask) == 0) { 434 if (bp != NULL) 435 brelse(bp); 436 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 437 goto fail; 438 } 439 440 /* Add this entry to the hash. */ 441 ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 442 if (ep->d_reclen == 0 || ep->d_reclen > 443 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 444 /* Corrupted directory. */ 445 brelse(bp); 446 goto fail; 447 } 448 if (ep->d_ino != 0) { 449 /* Add the entry (simplified ufsdirhash_add). */ 450 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 451 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 452 slot = WRAPINCR(slot, dh->dh_hlen); 453 dh->dh_hused++; 454 DH_ENTRY(dh, slot) = pos; 455 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 456 } 457 pos += ep->d_reclen; 458 } 459 460 if (bp != NULL) 461 brelse(bp); 462 DIRHASHLIST_LOCK(); 463 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 464 dh->dh_onlist = 1; 465 DIRHASHLIST_UNLOCK(); 466 sx_downgrade(&dh->dh_lock); 467 return (0); 468 469fail: 470 ufsdirhash_free_locked(ip); 471 return (-1); 472} 473 474/* 475 * Free any hash table associated with inode 'ip'. 476 */ 477static void 478ufsdirhash_free_locked(struct inode *ip) 479{ 480 struct dirhash *dh; 481 struct vnode *vp; 482 int i; 483 484 DIRHASH_ASSERT_LOCKED(ip->i_dirhash); 485 486 /* 487 * Clear the pointer in the inode to prevent new threads from 488 * finding the dead structure. 489 */ 490 vp = ip->i_vnode; 491 VI_LOCK(vp); 492 dh = ip->i_dirhash; 493 ip->i_dirhash = NULL; 494 VI_UNLOCK(vp); 495 496 /* 497 * Remove the hash from the list since we are going to free its 498 * memory. 499 */ 500 DIRHASHLIST_LOCK(); 501 if (dh->dh_onlist) 502 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 503 ufs_dirhashmem -= dh->dh_memreq; 504 DIRHASHLIST_UNLOCK(); 505 506 /* 507 * At this point, any waiters for the lock should hold their 508 * own reference on the dirhash structure. They will drop 509 * that reference once they grab the vnode interlock and see 510 * that ip->i_dirhash is NULL. 511 */ 512 sx_xunlock(&dh->dh_lock); 513 514 /* 515 * Handle partially recycled as well as fully constructed hashes. 516 */ 517 if (dh->dh_hash != NULL) { 518 for (i = 0; i < dh->dh_narrays; i++) 519 if (dh->dh_hash[i] != NULL) 520 DIRHASH_BLKFREE(dh->dh_hash[i]); 521 free(dh->dh_hash, M_DIRHASH); 522 if (dh->dh_blkfree != NULL) 523 free(dh->dh_blkfree, M_DIRHASH); 524 } 525 526 /* 527 * Drop the inode's reference to the data structure. 528 */ 529 ufsdirhash_drop(dh); 530} 531 532/* 533 * Find the offset of the specified name within the given inode. 534 * Returns 0 on success, ENOENT if the entry does not exist, or 535 * EJUSTRETURN if the caller should revert to a linear search. 536 * 537 * If successful, the directory offset is stored in *offp, and a 538 * pointer to a struct buf containing the entry is stored in *bpp. If 539 * prevoffp is non-NULL, the offset of the previous entry within 540 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 541 * is the first in a block, the start of the block is used). 542 * 543 * Must be called with the hash locked. Returns with the hash unlocked. 544 */ 545int 546ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 547 struct buf **bpp, doff_t *prevoffp) 548{ 549 struct dirhash *dh, *dh_next; 550 struct direct *dp; 551 struct vnode *vp; 552 struct buf *bp; 553 doff_t blkoff, bmask, offset, prevoff; 554 int i, slot; 555 int error; 556 557 dh = ip->i_dirhash; 558 KASSERT(dh != NULL && dh->dh_hash != NULL, 559 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh)); 560 DIRHASH_ASSERT_LOCKED(dh); 561 /* 562 * Move this dirhash towards the end of the list if it has a 563 * score higher than the next entry, and acquire the dh_lock. 564 */ 565 DIRHASHLIST_LOCK(); 566 if (TAILQ_NEXT(dh, dh_list) != NULL) { 567 /* 568 * If the new score will be greater than that of the next 569 * entry, then move this entry past it. With both mutexes 570 * held, dh_next won't go away, but its dh_score could 571 * change; that's not important since it is just a hint. 572 */ 573 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 574 dh->dh_score >= dh_next->dh_score) { 575 KASSERT(dh->dh_onlist, ("dirhash: not on list")); 576 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 577 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 578 dh_list); 579 } 580 } 581 /* Update the score. */ 582 if (dh->dh_score < DH_SCOREMAX) 583 dh->dh_score++; 584 585 /* Update last used time. */ 586 dh->dh_lastused = time_second; 587 DIRHASHLIST_UNLOCK(); 588 589 vp = ip->i_vnode; 590 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 591 blkoff = -1; 592 bp = NULL; 593restart: 594 slot = ufsdirhash_hash(dh, name, namelen); 595 596 if (dh->dh_seqopt) { 597 /* 598 * Sequential access optimisation. dh_seqoff contains the 599 * offset of the directory entry immediately following 600 * the last entry that was looked up. Check if this offset 601 * appears in the hash chain for the name we are looking for. 602 */ 603 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 604 i = WRAPINCR(i, dh->dh_hlen)) 605 if (offset == dh->dh_seqoff) 606 break; 607 if (offset == dh->dh_seqoff) { 608 /* 609 * We found an entry with the expected offset. This 610 * is probably the entry we want, but if not, the 611 * code below will turn off seqopt and retry. 612 */ 613 slot = i; 614 } else 615 dh->dh_seqopt = 0; 616 } 617 618 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 619 slot = WRAPINCR(slot, dh->dh_hlen)) { 620 if (offset == DIRHASH_DEL) 621 continue; 622 if (offset < 0 || offset >= ip->i_size) 623 panic("ufsdirhash_lookup: bad offset in hash array"); 624 if ((offset & ~bmask) != blkoff) { 625 if (bp != NULL) 626 brelse(bp); 627 blkoff = offset & ~bmask; 628 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) { 629 error = EJUSTRETURN; 630 goto fail; 631 } 632 } 633 KASSERT(bp != NULL, ("no buffer allocated")); 634 dp = (struct direct *)(bp->b_data + (offset & bmask)); 635 if (dp->d_reclen == 0 || dp->d_reclen > 636 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 637 /* Corrupted directory. */ 638 error = EJUSTRETURN; 639 goto fail; 640 } 641 if (dp->d_namlen == namelen && 642 bcmp(dp->d_name, name, namelen) == 0) { 643 /* Found. Get the prev offset if needed. */ 644 if (prevoffp != NULL) { 645 if (offset & (DIRBLKSIZ - 1)) { 646 prevoff = ufsdirhash_getprev(dp, 647 offset); 648 if (prevoff == -1) { 649 error = EJUSTRETURN; 650 goto fail; 651 } 652 } else 653 prevoff = offset; 654 *prevoffp = prevoff; 655 } 656 657 /* Check for sequential access, and update offset. */ 658 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 659 dh->dh_seqopt = 1; 660 dh->dh_seqoff = offset + DIRSIZ(0, dp); 661 *bpp = bp; 662 *offp = offset; 663 ufsdirhash_release(dh); 664 return (0); 665 } 666 667 /* 668 * When the name doesn't match in the seqopt case, go back 669 * and search normally. 670 */ 671 if (dh->dh_seqopt) { 672 dh->dh_seqopt = 0; 673 goto restart; 674 } 675 } 676 error = ENOENT; 677fail: 678 ufsdirhash_release(dh); 679 if (bp != NULL) 680 brelse(bp); 681 return (error); 682} 683 684/* 685 * Find a directory block with room for 'slotneeded' bytes. Returns 686 * the offset of the directory entry that begins the free space. 687 * This will either be the offset of an existing entry that has free 688 * space at the end, or the offset of an entry with d_ino == 0 at 689 * the start of a DIRBLKSIZ block. 690 * 691 * To use the space, the caller may need to compact existing entries in 692 * the directory. The total number of bytes in all of the entries involved 693 * in the compaction is stored in *slotsize. In other words, all of 694 * the entries that must be compacted are exactly contained in the 695 * region beginning at the returned offset and spanning *slotsize bytes. 696 * 697 * Returns -1 if no space was found, indicating that the directory 698 * must be extended. 699 */ 700doff_t 701ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 702{ 703 struct direct *dp; 704 struct dirhash *dh; 705 struct buf *bp; 706 doff_t pos, slotstart; 707 int dirblock, error, freebytes, i; 708 709 dh = ip->i_dirhash; 710 KASSERT(dh != NULL && dh->dh_hash != NULL, 711 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh)); 712 DIRHASH_ASSERT_LOCKED(dh); 713 714 /* Find a directory block with the desired free space. */ 715 dirblock = -1; 716 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 717 if ((dirblock = dh->dh_firstfree[i]) != -1) 718 break; 719 if (dirblock == -1) 720 return (-1); 721 722 KASSERT(dirblock < dh->dh_nblk && 723 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 724 ("ufsdirhash_findfree: bad stats")); 725 pos = dirblock * DIRBLKSIZ; 726 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 727 if (error) 728 return (-1); 729 730 /* Find the first entry with free space. */ 731 for (i = 0; i < DIRBLKSIZ; ) { 732 if (dp->d_reclen == 0) { 733 brelse(bp); 734 return (-1); 735 } 736 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 737 break; 738 i += dp->d_reclen; 739 dp = (struct direct *)((char *)dp + dp->d_reclen); 740 } 741 if (i > DIRBLKSIZ) { 742 brelse(bp); 743 return (-1); 744 } 745 slotstart = pos + i; 746 747 /* Find the range of entries needed to get enough space */ 748 freebytes = 0; 749 while (i < DIRBLKSIZ && freebytes < slotneeded) { 750 freebytes += dp->d_reclen; 751 if (dp->d_ino != 0) 752 freebytes -= DIRSIZ(0, dp); 753 if (dp->d_reclen == 0) { 754 brelse(bp); 755 return (-1); 756 } 757 i += dp->d_reclen; 758 dp = (struct direct *)((char *)dp + dp->d_reclen); 759 } 760 if (i > DIRBLKSIZ) { 761 brelse(bp); 762 return (-1); 763 } 764 if (freebytes < slotneeded) 765 panic("ufsdirhash_findfree: free mismatch"); 766 brelse(bp); 767 *slotsize = pos + i - slotstart; 768 return (slotstart); 769} 770 771/* 772 * Return the start of the unused space at the end of a directory, or 773 * -1 if there are no trailing unused blocks. 774 */ 775doff_t 776ufsdirhash_enduseful(struct inode *ip) 777{ 778 779 struct dirhash *dh; 780 int i; 781 782 dh = ip->i_dirhash; 783 DIRHASH_ASSERT_LOCKED(dh); 784 KASSERT(dh != NULL && dh->dh_hash != NULL, 785 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh)); 786 787 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) 788 return (-1); 789 790 for (i = dh->dh_dirblks - 1; i >= 0; i--) 791 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 792 break; 793 794 return ((doff_t)(i + 1) * DIRBLKSIZ); 795} 796 797/* 798 * Insert information into the hash about a new directory entry. dirp 799 * points to a struct direct containing the entry, and offset specifies 800 * the offset of this entry. 801 */ 802void 803ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 804{ 805 struct dirhash *dh; 806 int slot; 807 808 if ((dh = ufsdirhash_acquire(ip)) == NULL) 809 return; 810 811 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 812 ("ufsdirhash_add: bad offset")); 813 /* 814 * Normal hash usage is < 66%. If the usage gets too high then 815 * remove the hash entirely and let it be rebuilt later. 816 */ 817 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 818 ufsdirhash_free_locked(ip); 819 return; 820 } 821 822 /* Find a free hash slot (empty or deleted), and add the entry. */ 823 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 824 while (DH_ENTRY(dh, slot) >= 0) 825 slot = WRAPINCR(slot, dh->dh_hlen); 826 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 827 dh->dh_hused++; 828 DH_ENTRY(dh, slot) = offset; 829 830 /* Update last used time. */ 831 dh->dh_lastused = time_second; 832 833 /* Update the per-block summary info. */ 834 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 835 ufsdirhash_release(dh); 836} 837 838/* 839 * Remove the specified directory entry from the hash. The entry to remove 840 * is defined by the name in `dirp', which must exist at the specified 841 * `offset' within the directory. 842 */ 843void 844ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 845{ 846 struct dirhash *dh; 847 int slot; 848 849 if ((dh = ufsdirhash_acquire(ip)) == NULL) 850 return; 851 852 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 853 ("ufsdirhash_remove: bad offset")); 854 /* Find the entry */ 855 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 856 857 /* Remove the hash entry. */ 858 ufsdirhash_delslot(dh, slot); 859 860 /* Update the per-block summary info. */ 861 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 862 ufsdirhash_release(dh); 863} 864 865/* 866 * Change the offset associated with a directory entry in the hash. Used 867 * when compacting directory blocks. 868 */ 869void 870ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 871 doff_t newoff) 872{ 873 struct dirhash *dh; 874 int slot; 875 876 if ((dh = ufsdirhash_acquire(ip)) == NULL) 877 return; 878 879 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 880 newoff < dh->dh_dirblks * DIRBLKSIZ, 881 ("ufsdirhash_move: bad offset")); 882 /* Find the entry, and update the offset. */ 883 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 884 DH_ENTRY(dh, slot) = newoff; 885 ufsdirhash_release(dh); 886} 887 888/* 889 * Inform dirhash that the directory has grown by one block that 890 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 891 */ 892void 893ufsdirhash_newblk(struct inode *ip, doff_t offset) 894{ 895 struct dirhash *dh; 896 int block; 897 898 if ((dh = ufsdirhash_acquire(ip)) == NULL) 899 return; 900 901 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 902 ("ufsdirhash_newblk: bad offset")); 903 block = offset / DIRBLKSIZ; 904 if (block >= dh->dh_nblk) { 905 /* Out of space; must rebuild. */ 906 ufsdirhash_free_locked(ip); 907 return; 908 } 909 dh->dh_dirblks = block + 1; 910 911 /* Account for the new free block. */ 912 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 913 if (dh->dh_firstfree[DH_NFSTATS] == -1) 914 dh->dh_firstfree[DH_NFSTATS] = block; 915 ufsdirhash_release(dh); 916} 917 918/* 919 * Inform dirhash that the directory is being truncated. 920 */ 921void 922ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 923{ 924 struct dirhash *dh; 925 int block, i; 926 927 if ((dh = ufsdirhash_acquire(ip)) == NULL) 928 return; 929 930 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 931 ("ufsdirhash_dirtrunc: bad offset")); 932 block = howmany(offset, DIRBLKSIZ); 933 /* 934 * If the directory shrinks to less than 1/8 of dh_nblk blocks 935 * (about 20% of its original size due to the 50% extra added in 936 * ufsdirhash_build) then free it, and let the caller rebuild 937 * if necessary. 938 */ 939 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 940 ufsdirhash_free_locked(ip); 941 return; 942 } 943 944 /* 945 * Remove any `first free' information pertaining to the 946 * truncated blocks. All blocks we're removing should be 947 * completely unused. 948 */ 949 if (dh->dh_firstfree[DH_NFSTATS] >= block) 950 dh->dh_firstfree[DH_NFSTATS] = -1; 951 for (i = block; i < dh->dh_dirblks; i++) 952 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 953 panic("ufsdirhash_dirtrunc: blocks in use"); 954 for (i = 0; i < DH_NFSTATS; i++) 955 if (dh->dh_firstfree[i] >= block) 956 panic("ufsdirhash_dirtrunc: first free corrupt"); 957 dh->dh_dirblks = block; 958 ufsdirhash_release(dh); 959} 960 961/* 962 * Debugging function to check that the dirhash information about 963 * a directory block matches its actual contents. Panics if a mismatch 964 * is detected. 965 * 966 * On entry, `buf' should point to the start of an in-core 967 * DIRBLKSIZ-sized directory block, and `offset' should contain the 968 * offset from the start of the directory of that block. 969 */ 970void 971ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 972{ 973 struct dirhash *dh; 974 struct direct *dp; 975 int block, ffslot, i, nfree; 976 977 if (!ufs_dirhashcheck) 978 return; 979 if ((dh = ufsdirhash_acquire(ip)) == NULL) 980 return; 981 982 block = offset / DIRBLKSIZ; 983 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 984 panic("ufsdirhash_checkblock: bad offset"); 985 986 nfree = 0; 987 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 988 dp = (struct direct *)(buf + i); 989 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 990 panic("ufsdirhash_checkblock: bad dir"); 991 992 if (dp->d_ino == 0) { 993#if 0 994 /* 995 * XXX entries with d_ino == 0 should only occur 996 * at the start of a DIRBLKSIZ block. However the 997 * ufs code is tolerant of such entries at other 998 * offsets, and fsck does not fix them. 999 */ 1000 if (i != 0) 1001 panic("ufsdirhash_checkblock: bad dir inode"); 1002#endif 1003 nfree += dp->d_reclen; 1004 continue; 1005 } 1006 1007 /* Check that the entry exists (will panic if it doesn't). */ 1008 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 1009 1010 nfree += dp->d_reclen - DIRSIZ(0, dp); 1011 } 1012 if (i != DIRBLKSIZ) 1013 panic("ufsdirhash_checkblock: bad dir end"); 1014 1015 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 1016 panic("ufsdirhash_checkblock: bad free count"); 1017 1018 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 1019 for (i = 0; i <= DH_NFSTATS; i++) 1020 if (dh->dh_firstfree[i] == block && i != ffslot) 1021 panic("ufsdirhash_checkblock: bad first-free"); 1022 if (dh->dh_firstfree[ffslot] == -1) 1023 panic("ufsdirhash_checkblock: missing first-free entry"); 1024 ufsdirhash_release(dh); 1025} 1026 1027/* 1028 * Hash the specified filename into a dirhash slot. 1029 */ 1030static int 1031ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 1032{ 1033 u_int32_t hash; 1034 1035 /* 1036 * We hash the name and then some other bit of data that is 1037 * invariant over the dirhash's lifetime. Otherwise names 1038 * differing only in the last byte are placed close to one 1039 * another in the table, which is bad for linear probing. 1040 */ 1041 hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 1042 hash = fnv_32_buf(&dh, sizeof(dh), hash); 1043 return (hash % dh->dh_hlen); 1044} 1045 1046/* 1047 * Adjust the number of free bytes in the block containing `offset' 1048 * by the value specified by `diff'. 1049 * 1050 * The caller must ensure we have exclusive access to `dh'; normally 1051 * that means that dh_lock should be held, but this is also called 1052 * from ufsdirhash_build() where exclusive access can be assumed. 1053 */ 1054static void 1055ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 1056{ 1057 int block, i, nfidx, ofidx; 1058 1059 /* Update the per-block summary info. */ 1060 block = offset / DIRBLKSIZ; 1061 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 1062 ("dirhash bad offset")); 1063 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1064 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 1065 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1066 1067 /* Update the `first free' list if necessary. */ 1068 if (ofidx != nfidx) { 1069 /* If removing, scan forward for the next block. */ 1070 if (dh->dh_firstfree[ofidx] == block) { 1071 for (i = block + 1; i < dh->dh_dirblks; i++) 1072 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 1073 break; 1074 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 1075 } 1076 1077 /* Make this the new `first free' if necessary */ 1078 if (dh->dh_firstfree[nfidx] > block || 1079 dh->dh_firstfree[nfidx] == -1) 1080 dh->dh_firstfree[nfidx] = block; 1081 } 1082} 1083 1084/* 1085 * Find the specified name which should have the specified offset. 1086 * Returns a slot number, and panics on failure. 1087 * 1088 * `dh' must be locked on entry and remains so on return. 1089 */ 1090static int 1091ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 1092{ 1093 int slot; 1094 1095 DIRHASH_ASSERT_LOCKED(dh); 1096 1097 /* Find the entry. */ 1098 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 1099 slot = ufsdirhash_hash(dh, name, namelen); 1100 while (DH_ENTRY(dh, slot) != offset && 1101 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 1102 slot = WRAPINCR(slot, dh->dh_hlen); 1103 if (DH_ENTRY(dh, slot) != offset) 1104 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 1105 1106 return (slot); 1107} 1108 1109/* 1110 * Remove the entry corresponding to the specified slot from the hash array. 1111 * 1112 * `dh' must be locked on entry and remains so on return. 1113 */ 1114static void 1115ufsdirhash_delslot(struct dirhash *dh, int slot) 1116{ 1117 int i; 1118 1119 DIRHASH_ASSERT_LOCKED(dh); 1120 1121 /* Mark the entry as deleted. */ 1122 DH_ENTRY(dh, slot) = DIRHASH_DEL; 1123 1124 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 1125 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 1126 i = WRAPINCR(i, dh->dh_hlen); 1127 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 1128 i = WRAPDECR(i, dh->dh_hlen); 1129 while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 1130 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 1131 dh->dh_hused--; 1132 i = WRAPDECR(i, dh->dh_hlen); 1133 } 1134 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 1135 } 1136} 1137 1138/* 1139 * Given a directory entry and its offset, find the offset of the 1140 * previous entry in the same DIRBLKSIZ-sized block. Returns an 1141 * offset, or -1 if there is no previous entry in the block or some 1142 * other problem occurred. 1143 */ 1144static doff_t 1145ufsdirhash_getprev(struct direct *dirp, doff_t offset) 1146{ 1147 struct direct *dp; 1148 char *blkbuf; 1149 doff_t blkoff, prevoff; 1150 int entrypos, i; 1151 1152 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 1153 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 1154 blkbuf = (char *)dirp - entrypos; 1155 prevoff = blkoff; 1156 1157 /* If `offset' is the start of a block, there is no previous entry. */ 1158 if (entrypos == 0) 1159 return (-1); 1160 1161 /* Scan from the start of the block until we get to the entry. */ 1162 for (i = 0; i < entrypos; i += dp->d_reclen) { 1163 dp = (struct direct *)(blkbuf + i); 1164 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 1165 return (-1); /* Corrupted directory. */ 1166 prevoff = blkoff + i; 1167 } 1168 return (prevoff); 1169} 1170 1171/* 1172 * Delete the given dirhash and reclaim its memory. Assumes that 1173 * ufsdirhash_list is locked, and leaves it locked. Also assumes 1174 * that dh is locked. Returns the amount of memory freed. 1175 */ 1176static int 1177ufsdirhash_destroy(struct dirhash *dh) 1178{ 1179 doff_t **hash; 1180 u_int8_t *blkfree; 1181 int i, mem, narrays; 1182 1183 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 1184 1185 /* Remove it from the list and detach its memory. */ 1186 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 1187 dh->dh_onlist = 0; 1188 hash = dh->dh_hash; 1189 dh->dh_hash = NULL; 1190 blkfree = dh->dh_blkfree; 1191 dh->dh_blkfree = NULL; 1192 narrays = dh->dh_narrays; 1193 mem = dh->dh_memreq; 1194 dh->dh_memreq = 0; 1195 1196 /* Unlock dirhash and free the detached memory. */ 1197 ufsdirhash_release(dh); 1198 for (i = 0; i < narrays; i++) 1199 DIRHASH_BLKFREE(hash[i]); 1200 free(hash, M_DIRHASH); 1201 free(blkfree, M_DIRHASH); 1202 1203 /* Account for the returned memory. */ 1204 ufs_dirhashmem -= mem; 1205 1206 return (mem); 1207} 1208 1209/* 1210 * Try to free up `wanted' bytes by stealing memory from existing 1211 * dirhashes. Returns zero with list locked if successful. 1212 */ 1213static int 1214ufsdirhash_recycle(int wanted) 1215{ 1216 struct dirhash *dh; 1217 1218 DIRHASHLIST_LOCK(); 1219 dh = TAILQ_FIRST(&ufsdirhash_list); 1220 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 1221 /* Decrement the score; only recycle if it becomes zero. */ 1222 if (dh == NULL || --dh->dh_score > 0) { 1223 DIRHASHLIST_UNLOCK(); 1224 return (-1); 1225 } 1226 /* 1227 * If we can't lock it it's in use and we don't want to 1228 * recycle it anyway. 1229 */ 1230 if (!sx_try_xlock(&dh->dh_lock)) { 1231 dh = TAILQ_NEXT(dh, dh_list); 1232 continue; 1233 } 1234 1235 ufsdirhash_destroy(dh); 1236 1237 /* Repeat if necessary. */ 1238 dh = TAILQ_FIRST(&ufsdirhash_list); 1239 } 1240 /* Success; return with list locked. */ 1241 return (0); 1242} 1243 1244/* 1245 * Callback that frees some dirhashes when the system is low on virtual memory. 1246 */ 1247static void 1248ufsdirhash_lowmem() 1249{ 1250 struct dirhash *dh, *dh_temp; 1251 int memfreed = 0; 1252 /* XXX: this 10% may need to be adjusted */ 1253 int memwanted = ufs_dirhashmem / 10; 1254 1255 ufs_dirhashlowmemcount++; 1256 1257 DIRHASHLIST_LOCK(); 1258 /* 1259 * Delete dirhashes not used for more than ufs_dirhashreclaimage 1260 * seconds. If we can't get a lock on the dirhash, it will be skipped. 1261 */ 1262 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) { 1263 if (!sx_try_xlock(&dh->dh_lock)) 1264 continue; 1265 if (time_second - dh->dh_lastused > ufs_dirhashreclaimage) 1266 memfreed += ufsdirhash_destroy(dh); 1267 /* Unlock if we didn't delete the dirhash */ 1268 else 1269 ufsdirhash_release(dh); 1270 } 1271 1272 /* 1273 * If not enough memory was freed, keep deleting hashes from the head 1274 * of the dirhash list. The ones closest to the head should be the 1275 * oldest. 1276 */ 1277 if (memfreed < memwanted) { 1278 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) { 1279 if (!sx_try_xlock(&dh->dh_lock)) 1280 continue; 1281 memfreed += ufsdirhash_destroy(dh); 1282 if (memfreed >= memwanted) 1283 break; 1284 } 1285 } 1286 DIRHASHLIST_UNLOCK(); 1287} 1288 1289 1290void 1291ufsdirhash_init() 1292{ 1293 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t), 1294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1295 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF); 1296 TAILQ_INIT(&ufsdirhash_list); 1297 1298 /* Register a callback function to handle low memory signals */ 1299 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL, 1300 EVENTHANDLER_PRI_FIRST); 1301} 1302 1303void 1304ufsdirhash_uninit() 1305{ 1306 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit")); 1307 uma_zdestroy(ufsdirhash_zone); 1308 mtx_destroy(&ufsdirhash_mtx); 1309} 1310 1311#endif /* UFS_DIRHASH */ 1312