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: stable/11/sys/ufs/ufs/ufs_dirhash.c 326845 2017-12-14 11:41:12Z kib $"); 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; /* NOTE: initial value. It is 76 tuned in ufsdirhash_init() */ 77SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 78 0, "maximum allowed dirhash memory usage"); 79static int ufs_dirhashmem; 80SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 81 0, "current dirhash memory usage"); 82static int ufs_dirhashcheck = 0; 83SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 84 0, "enable extra sanity tests"); 85static int ufs_dirhashlowmemcount = 0; 86SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_lowmemcount, CTLFLAG_RD, 87 &ufs_dirhashlowmemcount, 0, "number of times low memory hook called"); 88static int ufs_dirhashreclaimpercent = 10; 89static int ufsdirhash_set_reclaimpercent(SYSCTL_HANDLER_ARGS); 90SYSCTL_PROC(_vfs_ufs, OID_AUTO, dirhash_reclaimpercent, 91 CTLTYPE_INT | CTLFLAG_RW, 0, 0, ufsdirhash_set_reclaimpercent, "I", 92 "set percentage of dirhash cache to be removed in low VM events"); 93 94 95static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 96static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 97static void ufsdirhash_delslot(struct dirhash *dh, int slot); 98static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 99 doff_t offset); 100static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 101static int ufsdirhash_recycle(int wanted); 102static void ufsdirhash_lowmem(void); 103static void ufsdirhash_free_locked(struct inode *ip); 104 105static uma_zone_t ufsdirhash_zone; 106 107#define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx) 108#define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx) 109#define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK) 110#define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr)) 111#define DIRHASH_ASSERT_LOCKED(dh) \ 112 sx_assert(&(dh)->dh_lock, SA_LOCKED) 113 114/* Dirhash list; recently-used entries are near the tail. */ 115static TAILQ_HEAD(, dirhash) ufsdirhash_list; 116 117/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 118static struct mtx ufsdirhash_mtx; 119 120/* 121 * Locking: 122 * 123 * The relationship between inode and dirhash is protected either by an 124 * exclusive vnode lock or the vnode interlock where a shared vnode lock 125 * may be used. The dirhash_mtx is acquired after the dirhash lock. To 126 * handle teardown races, code wishing to lock the dirhash for an inode 127 * when using a shared vnode lock must obtain a private reference on the 128 * dirhash while holding the vnode interlock. They can drop it once they 129 * have obtained the dirhash lock and verified that the dirhash wasn't 130 * recycled while they waited for the dirhash lock. 131 * 132 * ufsdirhash_build() acquires a shared lock on the dirhash when it is 133 * successful. This lock is released after a call to ufsdirhash_lookup(). 134 * 135 * Functions requiring exclusive access use ufsdirhash_acquire() which may 136 * free a dirhash structure that was recycled by ufsdirhash_recycle(). 137 * 138 * The dirhash lock may be held across io operations. 139 * 140 * WITNESS reports a lock order reversal between the "bufwait" lock 141 * and the "dirhash" lock. However, this specific reversal will not 142 * cause a deadlock. To get a deadlock, one would have to lock a 143 * buffer followed by the dirhash while a second thread locked a 144 * buffer while holding the dirhash lock. The second order can happen 145 * under a shared or exclusive vnode lock for the associated directory 146 * in lookup(). The first order, however, can only happen under an 147 * exclusive vnode lock (e.g. unlink(), rename(), etc.). Thus, for 148 * a thread to be doing a "bufwait" -> "dirhash" order, it has to hold 149 * an exclusive vnode lock. That exclusive vnode lock will prevent 150 * any other threads from doing a "dirhash" -> "bufwait" order. 151 */ 152 153static void 154ufsdirhash_hold(struct dirhash *dh) 155{ 156 157 refcount_acquire(&dh->dh_refcount); 158} 159 160static void 161ufsdirhash_drop(struct dirhash *dh) 162{ 163 164 if (refcount_release(&dh->dh_refcount)) { 165 sx_destroy(&dh->dh_lock); 166 free(dh, M_DIRHASH); 167 } 168} 169 170/* 171 * Release the lock on a dirhash. 172 */ 173static void 174ufsdirhash_release(struct dirhash *dh) 175{ 176 177 sx_unlock(&dh->dh_lock); 178} 179 180/* 181 * Either acquire an existing hash locked shared or create a new hash and 182 * return it exclusively locked. May return NULL if the allocation fails. 183 * 184 * The vnode interlock is used to protect the i_dirhash pointer from 185 * simultaneous access while only a shared vnode lock is held. 186 */ 187static struct dirhash * 188ufsdirhash_create(struct inode *ip) 189{ 190 struct dirhash *ndh; 191 struct dirhash *dh; 192 struct vnode *vp; 193 bool excl; 194 195 ndh = dh = NULL; 196 vp = ip->i_vnode; 197 excl = false; 198 for (;;) { 199 /* Racy check for i_dirhash to prefetch a dirhash structure. */ 200 if (ip->i_dirhash == NULL && ndh == NULL) { 201 ndh = malloc(sizeof *dh, M_DIRHASH, 202 M_NOWAIT | M_ZERO); 203 if (ndh == NULL) 204 return (NULL); 205 refcount_init(&ndh->dh_refcount, 1); 206 207 /* 208 * The DUPOK is to prevent warnings from the 209 * sx_slock() a few lines down which is safe 210 * since the duplicate lock in that case is 211 * the one for this dirhash we are creating 212 * now which has no external references until 213 * after this function returns. 214 */ 215 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK); 216 sx_xlock(&ndh->dh_lock); 217 } 218 /* 219 * Check i_dirhash. If it's NULL just try to use a 220 * preallocated structure. If none exists loop and try again. 221 */ 222 VI_LOCK(vp); 223 dh = ip->i_dirhash; 224 if (dh == NULL) { 225 ip->i_dirhash = ndh; 226 VI_UNLOCK(vp); 227 if (ndh == NULL) 228 continue; 229 return (ndh); 230 } 231 ufsdirhash_hold(dh); 232 VI_UNLOCK(vp); 233 234 /* Acquire a lock on existing hashes. */ 235 if (excl) 236 sx_xlock(&dh->dh_lock); 237 else 238 sx_slock(&dh->dh_lock); 239 240 /* The hash could've been recycled while we were waiting. */ 241 VI_LOCK(vp); 242 if (ip->i_dirhash != dh) { 243 VI_UNLOCK(vp); 244 ufsdirhash_release(dh); 245 ufsdirhash_drop(dh); 246 continue; 247 } 248 VI_UNLOCK(vp); 249 ufsdirhash_drop(dh); 250 251 /* If the hash is still valid we've succeeded. */ 252 if (dh->dh_hash != NULL) 253 break; 254 /* 255 * If the hash is NULL it has been recycled. Try to upgrade 256 * so we can recreate it. If we fail the upgrade, drop our 257 * lock and try again. 258 */ 259 if (excl || sx_try_upgrade(&dh->dh_lock)) 260 break; 261 sx_sunlock(&dh->dh_lock); 262 excl = true; 263 } 264 /* Free the preallocated structure if it was not necessary. */ 265 if (ndh) { 266 ufsdirhash_release(ndh); 267 ufsdirhash_drop(ndh); 268 } 269 return (dh); 270} 271 272/* 273 * Acquire an exclusive lock on an existing hash. Requires an exclusive 274 * vnode lock to protect the i_dirhash pointer. hashes that have been 275 * recycled are reclaimed here and NULL is returned. 276 */ 277static struct dirhash * 278ufsdirhash_acquire(struct inode *ip) 279{ 280 struct dirhash *dh; 281 282 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__); 283 284 dh = ip->i_dirhash; 285 if (dh == NULL) 286 return (NULL); 287 sx_xlock(&dh->dh_lock); 288 if (dh->dh_hash != NULL) 289 return (dh); 290 ufsdirhash_free_locked(ip); 291 return (NULL); 292} 293 294/* 295 * Acquire exclusively and free the hash pointed to by ip. Works with a 296 * shared or exclusive vnode lock. 297 */ 298void 299ufsdirhash_free(struct inode *ip) 300{ 301 struct dirhash *dh; 302 struct vnode *vp; 303 304 vp = ip->i_vnode; 305 for (;;) { 306 /* Grab a reference on this inode's dirhash if it has one. */ 307 VI_LOCK(vp); 308 dh = ip->i_dirhash; 309 if (dh == NULL) { 310 VI_UNLOCK(vp); 311 return; 312 } 313 ufsdirhash_hold(dh); 314 VI_UNLOCK(vp); 315 316 /* Exclusively lock the dirhash. */ 317 sx_xlock(&dh->dh_lock); 318 319 /* If this dirhash still belongs to this inode, then free it. */ 320 VI_LOCK(vp); 321 if (ip->i_dirhash == dh) { 322 VI_UNLOCK(vp); 323 ufsdirhash_drop(dh); 324 break; 325 } 326 VI_UNLOCK(vp); 327 328 /* 329 * This inode's dirhash has changed while we were 330 * waiting for the dirhash lock, so try again. 331 */ 332 ufsdirhash_release(dh); 333 ufsdirhash_drop(dh); 334 } 335 ufsdirhash_free_locked(ip); 336} 337 338/* 339 * Attempt to build up a hash table for the directory contents in 340 * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 341 */ 342int 343ufsdirhash_build(struct inode *ip) 344{ 345 struct dirhash *dh; 346 struct buf *bp = NULL; 347 struct direct *ep; 348 struct vnode *vp; 349 doff_t bmask, pos; 350 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 351 352 /* Take care of a decreased sysctl value. */ 353 while (ufs_dirhashmem > ufs_dirhashmaxmem) { 354 if (ufsdirhash_recycle(0) != 0) 355 return (-1); 356 /* Recycled enough memory, so unlock the list. */ 357 DIRHASHLIST_UNLOCK(); 358 } 359 360 /* Check if we can/should use dirhash. */ 361 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) || 362 ip->i_effnlink == 0) { 363 if (ip->i_dirhash) 364 ufsdirhash_free(ip); 365 return (-1); 366 } 367 dh = ufsdirhash_create(ip); 368 if (dh == NULL) 369 return (-1); 370 if (dh->dh_hash != NULL) 371 return (0); 372 373 vp = ip->i_vnode; 374 /* Allocate 50% more entries than this dir size could ever need. */ 375 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 376 nslots = ip->i_size / DIRECTSIZ(1); 377 nslots = (nslots * 3 + 1) / 2; 378 narrays = howmany(nslots, DH_NBLKOFF); 379 nslots = narrays * DH_NBLKOFF; 380 dirblocks = howmany(ip->i_size, DIRBLKSIZ); 381 nblocks = (dirblocks * 3 + 1) / 2; 382 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 383 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 384 nblocks * sizeof(*dh->dh_blkfree); 385 DIRHASHLIST_LOCK(); 386 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 387 DIRHASHLIST_UNLOCK(); 388 if (memreqd > ufs_dirhashmaxmem / 2) 389 goto fail; 390 /* Try to free some space. */ 391 if (ufsdirhash_recycle(memreqd) != 0) 392 goto fail; 393 /* Enough was freed, and list has been locked. */ 394 } 395 ufs_dirhashmem += memreqd; 396 DIRHASHLIST_UNLOCK(); 397 398 /* Initialise the hash table and block statistics. */ 399 dh->dh_memreq = memreqd; 400 dh->dh_narrays = narrays; 401 dh->dh_hlen = nslots; 402 dh->dh_nblk = nblocks; 403 dh->dh_dirblks = dirblocks; 404 for (i = 0; i < DH_NFSTATS; i++) 405 dh->dh_firstfree[i] = -1; 406 dh->dh_firstfree[DH_NFSTATS] = 0; 407 dh->dh_hused = 0; 408 dh->dh_seqoff = -1; 409 dh->dh_score = DH_SCOREINIT; 410 dh->dh_lastused = time_second; 411 412 /* 413 * Use non-blocking mallocs so that we will revert to a linear 414 * lookup on failure rather than potentially blocking forever. 415 */ 416 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]), 417 M_DIRHASH, M_NOWAIT | M_ZERO); 418 if (dh->dh_hash == NULL) 419 goto fail; 420 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]), 421 M_DIRHASH, M_NOWAIT); 422 if (dh->dh_blkfree == NULL) 423 goto fail; 424 for (i = 0; i < narrays; i++) { 425 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL) 426 goto fail; 427 for (j = 0; j < DH_NBLKOFF; j++) 428 dh->dh_hash[i][j] = DIRHASH_EMPTY; 429 } 430 for (i = 0; i < dirblocks; i++) 431 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 432 bmask = vp->v_mount->mnt_stat.f_iosize - 1; 433 pos = 0; 434 while (pos < ip->i_size) { 435 /* If necessary, get the next directory block. */ 436 if ((pos & bmask) == 0) { 437 if (bp != NULL) 438 brelse(bp); 439 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 440 goto fail; 441 } 442 443 /* Add this entry to the hash. */ 444 ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 445 if (ep->d_reclen == 0 || ep->d_reclen > 446 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 447 /* Corrupted directory. */ 448 brelse(bp); 449 goto fail; 450 } 451 if (ep->d_ino != 0) { 452 /* Add the entry (simplified ufsdirhash_add). */ 453 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 454 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 455 slot = WRAPINCR(slot, dh->dh_hlen); 456 dh->dh_hused++; 457 DH_ENTRY(dh, slot) = pos; 458 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 459 } 460 pos += ep->d_reclen; 461 } 462 463 if (bp != NULL) 464 brelse(bp); 465 DIRHASHLIST_LOCK(); 466 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 467 dh->dh_onlist = 1; 468 DIRHASHLIST_UNLOCK(); 469 sx_downgrade(&dh->dh_lock); 470 return (0); 471 472fail: 473 ufsdirhash_free_locked(ip); 474 return (-1); 475} 476 477/* 478 * Free any hash table associated with inode 'ip'. 479 */ 480static void 481ufsdirhash_free_locked(struct inode *ip) 482{ 483 struct dirhash *dh; 484 struct vnode *vp; 485 int i; 486 487 DIRHASH_ASSERT_LOCKED(ip->i_dirhash); 488 489 /* 490 * Clear the pointer in the inode to prevent new threads from 491 * finding the dead structure. 492 */ 493 vp = ip->i_vnode; 494 VI_LOCK(vp); 495 dh = ip->i_dirhash; 496 ip->i_dirhash = NULL; 497 VI_UNLOCK(vp); 498 499 /* 500 * Remove the hash from the list since we are going to free its 501 * memory. 502 */ 503 DIRHASHLIST_LOCK(); 504 if (dh->dh_onlist) 505 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 506 ufs_dirhashmem -= dh->dh_memreq; 507 DIRHASHLIST_UNLOCK(); 508 509 /* 510 * At this point, any waiters for the lock should hold their 511 * own reference on the dirhash structure. They will drop 512 * that reference once they grab the vnode interlock and see 513 * that ip->i_dirhash is NULL. 514 */ 515 sx_xunlock(&dh->dh_lock); 516 517 /* 518 * Handle partially recycled as well as fully constructed hashes. 519 */ 520 if (dh->dh_hash != NULL) { 521 for (i = 0; i < dh->dh_narrays; i++) 522 if (dh->dh_hash[i] != NULL) 523 DIRHASH_BLKFREE(dh->dh_hash[i]); 524 free(dh->dh_hash, M_DIRHASH); 525 if (dh->dh_blkfree != NULL) 526 free(dh->dh_blkfree, M_DIRHASH); 527 } 528 529 /* 530 * Drop the inode's reference to the data structure. 531 */ 532 ufsdirhash_drop(dh); 533} 534 535/* 536 * Find the offset of the specified name within the given inode. 537 * Returns 0 on success, ENOENT if the entry does not exist, or 538 * EJUSTRETURN if the caller should revert to a linear search. 539 * 540 * If successful, the directory offset is stored in *offp, and a 541 * pointer to a struct buf containing the entry is stored in *bpp. If 542 * prevoffp is non-NULL, the offset of the previous entry within 543 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 544 * is the first in a block, the start of the block is used). 545 * 546 * Must be called with the hash locked. Returns with the hash unlocked. 547 */ 548int 549ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 550 struct buf **bpp, doff_t *prevoffp) 551{ 552 struct dirhash *dh, *dh_next; 553 struct direct *dp; 554 struct vnode *vp; 555 struct buf *bp; 556 doff_t blkoff, bmask, offset, prevoff, seqoff; 557 int i, slot; 558 int error; 559 560 dh = ip->i_dirhash; 561 KASSERT(dh != NULL && dh->dh_hash != NULL, 562 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh)); 563 DIRHASH_ASSERT_LOCKED(dh); 564 /* 565 * Move this dirhash towards the end of the list if it has a 566 * score higher than the next entry, and acquire the dh_lock. 567 */ 568 DIRHASHLIST_LOCK(); 569 if (TAILQ_NEXT(dh, dh_list) != NULL) { 570 /* 571 * If the new score will be greater than that of the next 572 * entry, then move this entry past it. With both mutexes 573 * held, dh_next won't go away, but its dh_score could 574 * change; that's not important since it is just a hint. 575 */ 576 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 577 dh->dh_score >= dh_next->dh_score) { 578 KASSERT(dh->dh_onlist, ("dirhash: not on list")); 579 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 580 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 581 dh_list); 582 } 583 } 584 /* Update the score. */ 585 if (dh->dh_score < DH_SCOREMAX) 586 dh->dh_score++; 587 588 /* Update last used time. */ 589 dh->dh_lastused = time_second; 590 DIRHASHLIST_UNLOCK(); 591 592 vp = ip->i_vnode; 593 bmask = vp->v_mount->mnt_stat.f_iosize - 1; 594 blkoff = -1; 595 bp = NULL; 596 seqoff = dh->dh_seqoff; 597restart: 598 slot = ufsdirhash_hash(dh, name, namelen); 599 600 if (seqoff != -1) { 601 /* 602 * Sequential access optimisation. seqoff contains the 603 * offset of the directory entry immediately following 604 * the last entry that was looked up. Check if this offset 605 * appears in the hash chain for the name we are looking for. 606 */ 607 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 608 i = WRAPINCR(i, dh->dh_hlen)) 609 if (offset == seqoff) 610 break; 611 if (offset == seqoff) { 612 /* 613 * We found an entry with the expected offset. This 614 * is probably the entry we want, but if not, the 615 * code below will retry. 616 */ 617 slot = i; 618 } else 619 seqoff = -1; 620 } 621 622 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 623 slot = WRAPINCR(slot, dh->dh_hlen)) { 624 if (offset == DIRHASH_DEL) 625 continue; 626 if (offset < 0 || offset >= ip->i_size) 627 panic("ufsdirhash_lookup: bad offset in hash array"); 628 if ((offset & ~bmask) != blkoff) { 629 if (bp != NULL) 630 brelse(bp); 631 blkoff = offset & ~bmask; 632 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) { 633 error = EJUSTRETURN; 634 goto fail; 635 } 636 } 637 KASSERT(bp != NULL, ("no buffer allocated")); 638 dp = (struct direct *)(bp->b_data + (offset & bmask)); 639 if (dp->d_reclen == 0 || dp->d_reclen > 640 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 641 /* Corrupted directory. */ 642 error = EJUSTRETURN; 643 goto fail; 644 } 645 if (dp->d_namlen == namelen && 646 bcmp(dp->d_name, name, namelen) == 0) { 647 /* Found. Get the prev offset if needed. */ 648 if (prevoffp != NULL) { 649 if (offset & (DIRBLKSIZ - 1)) { 650 prevoff = ufsdirhash_getprev(dp, 651 offset); 652 if (prevoff == -1) { 653 error = EJUSTRETURN; 654 goto fail; 655 } 656 } else 657 prevoff = offset; 658 *prevoffp = prevoff; 659 } 660 661 /* Update offset. */ 662 dh->dh_seqoff = offset + DIRSIZ(0, dp); 663 *bpp = bp; 664 *offp = offset; 665 ufsdirhash_release(dh); 666 return (0); 667 } 668 669 /* 670 * When the name doesn't match in the sequential 671 * optimization case, go back and search normally. 672 */ 673 if (seqoff != -1) { 674 seqoff = -1; 675 goto restart; 676 } 677 } 678 error = ENOENT; 679fail: 680 ufsdirhash_release(dh); 681 if (bp != NULL) 682 brelse(bp); 683 return (error); 684} 685 686/* 687 * Find a directory block with room for 'slotneeded' bytes. Returns 688 * the offset of the directory entry that begins the free space. 689 * This will either be the offset of an existing entry that has free 690 * space at the end, or the offset of an entry with d_ino == 0 at 691 * the start of a DIRBLKSIZ block. 692 * 693 * To use the space, the caller may need to compact existing entries in 694 * the directory. The total number of bytes in all of the entries involved 695 * in the compaction is stored in *slotsize. In other words, all of 696 * the entries that must be compacted are exactly contained in the 697 * region beginning at the returned offset and spanning *slotsize bytes. 698 * 699 * Returns -1 if no space was found, indicating that the directory 700 * must be extended. 701 */ 702doff_t 703ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 704{ 705 struct direct *dp; 706 struct dirhash *dh; 707 struct buf *bp; 708 doff_t pos, slotstart; 709 int dirblock, error, freebytes, i; 710 711 dh = ip->i_dirhash; 712 KASSERT(dh != NULL && dh->dh_hash != NULL, 713 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh)); 714 DIRHASH_ASSERT_LOCKED(dh); 715 716 /* Find a directory block with the desired free space. */ 717 dirblock = -1; 718 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 719 if ((dirblock = dh->dh_firstfree[i]) != -1) 720 break; 721 if (dirblock == -1) 722 return (-1); 723 724 KASSERT(dirblock < dh->dh_nblk && 725 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 726 ("ufsdirhash_findfree: bad stats")); 727 pos = dirblock * DIRBLKSIZ; 728 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 729 if (error) 730 return (-1); 731 732 /* Find the first entry with free space. */ 733 for (i = 0; i < DIRBLKSIZ; ) { 734 if (dp->d_reclen == 0) { 735 brelse(bp); 736 return (-1); 737 } 738 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 739 break; 740 i += dp->d_reclen; 741 dp = (struct direct *)((char *)dp + dp->d_reclen); 742 } 743 if (i > DIRBLKSIZ) { 744 brelse(bp); 745 return (-1); 746 } 747 slotstart = pos + i; 748 749 /* Find the range of entries needed to get enough space */ 750 freebytes = 0; 751 while (i < DIRBLKSIZ && freebytes < slotneeded) { 752 freebytes += dp->d_reclen; 753 if (dp->d_ino != 0) 754 freebytes -= DIRSIZ(0, dp); 755 if (dp->d_reclen == 0) { 756 brelse(bp); 757 return (-1); 758 } 759 i += dp->d_reclen; 760 dp = (struct direct *)((char *)dp + dp->d_reclen); 761 } 762 if (i > DIRBLKSIZ) { 763 brelse(bp); 764 return (-1); 765 } 766 if (freebytes < slotneeded) 767 panic("ufsdirhash_findfree: free mismatch"); 768 brelse(bp); 769 *slotsize = pos + i - slotstart; 770 return (slotstart); 771} 772 773/* 774 * Return the start of the unused space at the end of a directory, or 775 * -1 if there are no trailing unused blocks. 776 */ 777doff_t 778ufsdirhash_enduseful(struct inode *ip) 779{ 780 781 struct dirhash *dh; 782 int i; 783 784 dh = ip->i_dirhash; 785 DIRHASH_ASSERT_LOCKED(dh); 786 KASSERT(dh != NULL && dh->dh_hash != NULL, 787 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh)); 788 789 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) 790 return (-1); 791 792 for (i = dh->dh_dirblks - 1; i >= 0; i--) 793 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 794 break; 795 796 return ((doff_t)(i + 1) * DIRBLKSIZ); 797} 798 799/* 800 * Insert information into the hash about a new directory entry. dirp 801 * points to a struct direct containing the entry, and offset specifies 802 * the offset of this entry. 803 */ 804void 805ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 806{ 807 struct dirhash *dh; 808 int slot; 809 810 if ((dh = ufsdirhash_acquire(ip)) == NULL) 811 return; 812 813 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 814 ("ufsdirhash_add: bad offset")); 815 /* 816 * Normal hash usage is < 66%. If the usage gets too high then 817 * remove the hash entirely and let it be rebuilt later. 818 */ 819 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 820 ufsdirhash_free_locked(ip); 821 return; 822 } 823 824 /* Find a free hash slot (empty or deleted), and add the entry. */ 825 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 826 while (DH_ENTRY(dh, slot) >= 0) 827 slot = WRAPINCR(slot, dh->dh_hlen); 828 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 829 dh->dh_hused++; 830 DH_ENTRY(dh, slot) = offset; 831 832 /* Update last used time. */ 833 dh->dh_lastused = time_second; 834 835 /* Update the per-block summary info. */ 836 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 837 ufsdirhash_release(dh); 838} 839 840/* 841 * Remove the specified directory entry from the hash. The entry to remove 842 * is defined by the name in `dirp', which must exist at the specified 843 * `offset' within the directory. 844 */ 845void 846ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 847{ 848 struct dirhash *dh; 849 int slot; 850 851 if ((dh = ufsdirhash_acquire(ip)) == NULL) 852 return; 853 854 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 855 ("ufsdirhash_remove: bad offset")); 856 /* Find the entry */ 857 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 858 859 /* Remove the hash entry. */ 860 ufsdirhash_delslot(dh, slot); 861 862 /* Update the per-block summary info. */ 863 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 864 ufsdirhash_release(dh); 865} 866 867/* 868 * Change the offset associated with a directory entry in the hash. Used 869 * when compacting directory blocks. 870 */ 871void 872ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 873 doff_t newoff) 874{ 875 struct dirhash *dh; 876 int slot; 877 878 if ((dh = ufsdirhash_acquire(ip)) == NULL) 879 return; 880 881 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 882 newoff < dh->dh_dirblks * DIRBLKSIZ, 883 ("ufsdirhash_move: bad offset")); 884 /* Find the entry, and update the offset. */ 885 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 886 DH_ENTRY(dh, slot) = newoff; 887 ufsdirhash_release(dh); 888} 889 890/* 891 * Inform dirhash that the directory has grown by one block that 892 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 893 */ 894void 895ufsdirhash_newblk(struct inode *ip, doff_t offset) 896{ 897 struct dirhash *dh; 898 int block; 899 900 if ((dh = ufsdirhash_acquire(ip)) == NULL) 901 return; 902 903 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 904 ("ufsdirhash_newblk: bad offset")); 905 block = offset / DIRBLKSIZ; 906 if (block >= dh->dh_nblk) { 907 /* Out of space; must rebuild. */ 908 ufsdirhash_free_locked(ip); 909 return; 910 } 911 dh->dh_dirblks = block + 1; 912 913 /* Account for the new free block. */ 914 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 915 if (dh->dh_firstfree[DH_NFSTATS] == -1) 916 dh->dh_firstfree[DH_NFSTATS] = block; 917 ufsdirhash_release(dh); 918} 919 920/* 921 * Inform dirhash that the directory is being truncated. 922 */ 923void 924ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 925{ 926 struct dirhash *dh; 927 int block, i; 928 929 if ((dh = ufsdirhash_acquire(ip)) == NULL) 930 return; 931 932 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 933 ("ufsdirhash_dirtrunc: bad offset")); 934 block = howmany(offset, DIRBLKSIZ); 935 /* 936 * If the directory shrinks to less than 1/8 of dh_nblk blocks 937 * (about 20% of its original size due to the 50% extra added in 938 * ufsdirhash_build) then free it, and let the caller rebuild 939 * if necessary. 940 */ 941 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 942 ufsdirhash_free_locked(ip); 943 return; 944 } 945 946 /* 947 * Remove any `first free' information pertaining to the 948 * truncated blocks. All blocks we're removing should be 949 * completely unused. 950 */ 951 if (dh->dh_firstfree[DH_NFSTATS] >= block) 952 dh->dh_firstfree[DH_NFSTATS] = -1; 953 for (i = block; i < dh->dh_dirblks; i++) 954 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 955 panic("ufsdirhash_dirtrunc: blocks in use"); 956 for (i = 0; i < DH_NFSTATS; i++) 957 if (dh->dh_firstfree[i] >= block) 958 panic("ufsdirhash_dirtrunc: first free corrupt"); 959 dh->dh_dirblks = block; 960 ufsdirhash_release(dh); 961} 962 963/* 964 * Debugging function to check that the dirhash information about 965 * a directory block matches its actual contents. Panics if a mismatch 966 * is detected. 967 * 968 * On entry, `buf' should point to the start of an in-core 969 * DIRBLKSIZ-sized directory block, and `offset' should contain the 970 * offset from the start of the directory of that block. 971 */ 972void 973ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 974{ 975 struct dirhash *dh; 976 struct direct *dp; 977 int block, ffslot, i, nfree; 978 979 if (!ufs_dirhashcheck) 980 return; 981 if ((dh = ufsdirhash_acquire(ip)) == NULL) 982 return; 983 984 block = offset / DIRBLKSIZ; 985 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 986 panic("ufsdirhash_checkblock: bad offset"); 987 988 nfree = 0; 989 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 990 dp = (struct direct *)(buf + i); 991 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 992 panic("ufsdirhash_checkblock: bad dir"); 993 994 if (dp->d_ino == 0) { 995#if 0 996 /* 997 * XXX entries with d_ino == 0 should only occur 998 * at the start of a DIRBLKSIZ block. However the 999 * ufs code is tolerant of such entries at other 1000 * offsets, and fsck does not fix them. 1001 */ 1002 if (i != 0) 1003 panic("ufsdirhash_checkblock: bad dir inode"); 1004#endif 1005 nfree += dp->d_reclen; 1006 continue; 1007 } 1008 1009 /* Check that the entry exists (will panic if it doesn't). */ 1010 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 1011 1012 nfree += dp->d_reclen - DIRSIZ(0, dp); 1013 } 1014 if (i != DIRBLKSIZ) 1015 panic("ufsdirhash_checkblock: bad dir end"); 1016 1017 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 1018 panic("ufsdirhash_checkblock: bad free count"); 1019 1020 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 1021 for (i = 0; i <= DH_NFSTATS; i++) 1022 if (dh->dh_firstfree[i] == block && i != ffslot) 1023 panic("ufsdirhash_checkblock: bad first-free"); 1024 if (dh->dh_firstfree[ffslot] == -1) 1025 panic("ufsdirhash_checkblock: missing first-free entry"); 1026 ufsdirhash_release(dh); 1027} 1028 1029/* 1030 * Hash the specified filename into a dirhash slot. 1031 */ 1032static int 1033ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 1034{ 1035 u_int32_t hash; 1036 1037 /* 1038 * We hash the name and then some other bit of data that is 1039 * invariant over the dirhash's lifetime. Otherwise names 1040 * differing only in the last byte are placed close to one 1041 * another in the table, which is bad for linear probing. 1042 */ 1043 hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 1044 hash = fnv_32_buf(&dh, sizeof(dh), hash); 1045 return (hash % dh->dh_hlen); 1046} 1047 1048/* 1049 * Adjust the number of free bytes in the block containing `offset' 1050 * by the value specified by `diff'. 1051 * 1052 * The caller must ensure we have exclusive access to `dh'; normally 1053 * that means that dh_lock should be held, but this is also called 1054 * from ufsdirhash_build() where exclusive access can be assumed. 1055 */ 1056static void 1057ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 1058{ 1059 int block, i, nfidx, ofidx; 1060 1061 /* Update the per-block summary info. */ 1062 block = offset / DIRBLKSIZ; 1063 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 1064 ("dirhash bad offset")); 1065 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1066 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 1067 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 1068 1069 /* Update the `first free' list if necessary. */ 1070 if (ofidx != nfidx) { 1071 /* If removing, scan forward for the next block. */ 1072 if (dh->dh_firstfree[ofidx] == block) { 1073 for (i = block + 1; i < dh->dh_dirblks; i++) 1074 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 1075 break; 1076 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 1077 } 1078 1079 /* Make this the new `first free' if necessary */ 1080 if (dh->dh_firstfree[nfidx] > block || 1081 dh->dh_firstfree[nfidx] == -1) 1082 dh->dh_firstfree[nfidx] = block; 1083 } 1084} 1085 1086/* 1087 * Find the specified name which should have the specified offset. 1088 * Returns a slot number, and panics on failure. 1089 * 1090 * `dh' must be locked on entry and remains so on return. 1091 */ 1092static int 1093ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 1094{ 1095 int slot; 1096 1097 DIRHASH_ASSERT_LOCKED(dh); 1098 1099 /* Find the entry. */ 1100 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 1101 slot = ufsdirhash_hash(dh, name, namelen); 1102 while (DH_ENTRY(dh, slot) != offset && 1103 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 1104 slot = WRAPINCR(slot, dh->dh_hlen); 1105 if (DH_ENTRY(dh, slot) != offset) 1106 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 1107 1108 return (slot); 1109} 1110 1111/* 1112 * Remove the entry corresponding to the specified slot from the hash array. 1113 * 1114 * `dh' must be locked on entry and remains so on return. 1115 */ 1116static void 1117ufsdirhash_delslot(struct dirhash *dh, int slot) 1118{ 1119 int i; 1120 1121 DIRHASH_ASSERT_LOCKED(dh); 1122 1123 /* Mark the entry as deleted. */ 1124 DH_ENTRY(dh, slot) = DIRHASH_DEL; 1125 1126 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 1127 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 1128 i = WRAPINCR(i, dh->dh_hlen); 1129 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 1130 i = WRAPDECR(i, dh->dh_hlen); 1131 while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 1132 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 1133 dh->dh_hused--; 1134 i = WRAPDECR(i, dh->dh_hlen); 1135 } 1136 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 1137 } 1138} 1139 1140/* 1141 * Given a directory entry and its offset, find the offset of the 1142 * previous entry in the same DIRBLKSIZ-sized block. Returns an 1143 * offset, or -1 if there is no previous entry in the block or some 1144 * other problem occurred. 1145 */ 1146static doff_t 1147ufsdirhash_getprev(struct direct *dirp, doff_t offset) 1148{ 1149 struct direct *dp; 1150 char *blkbuf; 1151 doff_t blkoff, prevoff; 1152 int entrypos, i; 1153 1154 blkoff = rounddown2(offset, DIRBLKSIZ); /* offset of start of block */ 1155 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 1156 blkbuf = (char *)dirp - entrypos; 1157 prevoff = blkoff; 1158 1159 /* If `offset' is the start of a block, there is no previous entry. */ 1160 if (entrypos == 0) 1161 return (-1); 1162 1163 /* Scan from the start of the block until we get to the entry. */ 1164 for (i = 0; i < entrypos; i += dp->d_reclen) { 1165 dp = (struct direct *)(blkbuf + i); 1166 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 1167 return (-1); /* Corrupted directory. */ 1168 prevoff = blkoff + i; 1169 } 1170 return (prevoff); 1171} 1172 1173/* 1174 * Delete the given dirhash and reclaim its memory. Assumes that 1175 * ufsdirhash_list is locked, and leaves it locked. Also assumes 1176 * that dh is locked. Returns the amount of memory freed. 1177 */ 1178static int 1179ufsdirhash_destroy(struct dirhash *dh) 1180{ 1181 doff_t **hash; 1182 u_int8_t *blkfree; 1183 int i, mem, narrays; 1184 1185 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 1186 1187 /* Remove it from the list and detach its memory. */ 1188 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 1189 dh->dh_onlist = 0; 1190 hash = dh->dh_hash; 1191 dh->dh_hash = NULL; 1192 blkfree = dh->dh_blkfree; 1193 dh->dh_blkfree = NULL; 1194 narrays = dh->dh_narrays; 1195 mem = dh->dh_memreq; 1196 dh->dh_memreq = 0; 1197 1198 /* Unlock dirhash and free the detached memory. */ 1199 ufsdirhash_release(dh); 1200 for (i = 0; i < narrays; i++) 1201 DIRHASH_BLKFREE(hash[i]); 1202 free(hash, M_DIRHASH); 1203 free(blkfree, M_DIRHASH); 1204 1205 /* Account for the returned memory. */ 1206 ufs_dirhashmem -= mem; 1207 1208 return (mem); 1209} 1210 1211/* 1212 * Try to free up `wanted' bytes by stealing memory from existing 1213 * dirhashes. Returns zero with list locked if successful. 1214 */ 1215static int 1216ufsdirhash_recycle(int wanted) 1217{ 1218 struct dirhash *dh; 1219 1220 DIRHASHLIST_LOCK(); 1221 dh = TAILQ_FIRST(&ufsdirhash_list); 1222 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 1223 /* Decrement the score; only recycle if it becomes zero. */ 1224 if (dh == NULL || --dh->dh_score > 0) { 1225 DIRHASHLIST_UNLOCK(); 1226 return (-1); 1227 } 1228 /* 1229 * If we can't lock it it's in use and we don't want to 1230 * recycle it anyway. 1231 */ 1232 if (!sx_try_xlock(&dh->dh_lock)) { 1233 dh = TAILQ_NEXT(dh, dh_list); 1234 continue; 1235 } 1236 1237 ufsdirhash_destroy(dh); 1238 1239 /* Repeat if necessary. */ 1240 dh = TAILQ_FIRST(&ufsdirhash_list); 1241 } 1242 /* Success; return with list locked. */ 1243 return (0); 1244} 1245 1246/* 1247 * Callback that frees some dirhashes when the system is low on virtual memory. 1248 */ 1249static void 1250ufsdirhash_lowmem() 1251{ 1252 struct dirhash *dh, *dh_temp; 1253 int memfreed, memwanted; 1254 1255 ufs_dirhashlowmemcount++; 1256 memfreed = 0; 1257 memwanted = ufs_dirhashmem * ufs_dirhashreclaimpercent / 100; 1258 1259 DIRHASHLIST_LOCK(); 1260 1261 /* 1262 * Reclaim up to memwanted from the oldest dirhashes. This will allow 1263 * us to make some progress when the system is running out of memory 1264 * without compromising the dinamicity of maximum age. If the situation 1265 * does not improve lowmem will be eventually retriggered and free some 1266 * other entry in the cache. The entries on the head of the list should 1267 * be the oldest. If during list traversal we can't get a lock on the 1268 * dirhash, it will be skipped. 1269 */ 1270 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) { 1271 if (sx_try_xlock(&dh->dh_lock)) 1272 memfreed += ufsdirhash_destroy(dh); 1273 if (memfreed >= memwanted) 1274 break; 1275 } 1276 DIRHASHLIST_UNLOCK(); 1277} 1278 1279static int 1280ufsdirhash_set_reclaimpercent(SYSCTL_HANDLER_ARGS) 1281{ 1282 int error, v; 1283 1284 v = ufs_dirhashreclaimpercent; 1285 error = sysctl_handle_int(oidp, &v, v, req); 1286 if (error) 1287 return (error); 1288 if (req->newptr == NULL) 1289 return (error); 1290 if (v == ufs_dirhashreclaimpercent) 1291 return (0); 1292 1293 /* Refuse invalid percentages */ 1294 if (v < 0 || v > 100) 1295 return (EINVAL); 1296 ufs_dirhashreclaimpercent = v; 1297 return (0); 1298} 1299 1300void 1301ufsdirhash_init() 1302{ 1303 ufs_dirhashmaxmem = lmax(roundup(hibufspace / 64, PAGE_SIZE), 1304 2 * 1024 * 1024); 1305 1306 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t), 1307 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1308 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF); 1309 TAILQ_INIT(&ufsdirhash_list); 1310 1311 /* Register a callback function to handle low memory signals */ 1312 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL, 1313 EVENTHANDLER_PRI_FIRST); 1314} 1315 1316void 1317ufsdirhash_uninit() 1318{ 1319 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit")); 1320 uma_zdestroy(ufsdirhash_zone); 1321 mtx_destroy(&ufsdirhash_mtx); 1322} 1323 1324#endif /* UFS_DIRHASH */ 1325