ufs_dirhash.c revision 92768
11638Srgrimes/* 250476Speter * Copyright (c) 2001 Ian Dowse. All rights reserved. 31638Srgrimes * 4115103Strhodes * Redistribution and use in source and binary forms, with or without 5115103Strhodes * modification, are permitted provided that the following conditions 6115103Strhodes * are met: 7115103Strhodes * 1. Redistributions of source code must retain the above copyright 83470Srgrimes * notice, this list of conditions and the following disclaimer. 9115103Strhodes * 2. Redistributions in binary form must reproduce the above copyright 10115103Strhodes * notice, this list of conditions and the following disclaimer in the 111638Srgrimes * documentation and/or other materials provided with the distribution. 12115103Strhodes * 13115103Strhodes * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14115103Strhodes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15115103Strhodes * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16115103Strhodes * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17115103Strhodes * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18115103Strhodes * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19115103Strhodes * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20291786Sbdrewery * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2123559Swosch * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2223559Swosch * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23115103Strhodes * SUCH DAMAGE. 24115103Strhodes * 2523559Swosch * $FreeBSD: head/sys/ufs/ufs/ufs_dirhash.c 92768 2002-03-20 08:48:07Z jeff $ 26115103Strhodes */ 2723559Swosch/* 2823559Swosch * This implements a hash-based lookup scheme for UFS directories. 2923559Swosch */ 30115103Strhodes 31115103Strhodes#include "opt_ufs.h" 32115103Strhodes 3323559Swosch#ifdef UFS_DIRHASH 3423559Swosch 3523559Swosch#include <sys/param.h> 36115103Strhodes#include <sys/systm.h> 37115103Strhodes#include <sys/kernel.h> 3823559Swosch#include <sys/lock.h> 3923559Swosch#include <sys/mutex.h> 40152265Sharti#include <sys/malloc.h> 4123559Swosch#include <sys/fnv_hash.h> 42115103Strhodes#include <sys/proc.h> 43264483Sjmmv#include <sys/bio.h> 44115103Strhodes#include <sys/buf.h> 4523559Swosch#include <sys/vnode.h> 46115103Strhodes#include <sys/mount.h> 4723559Swosch#include <sys/sysctl.h> 48115103Strhodes#include <vm/uma.h> 49115103Strhodes 501638Srgrimes#include <ufs/ufs/quota.h> 51115103Strhodes#include <ufs/ufs/inode.h> 5223578Swosch#include <ufs/ufs/dir.h> 53115103Strhodes#include <ufs/ufs/dirhash.h> 54115103Strhodes#include <ufs/ufs/extattr.h> 55115103Strhodes#include <ufs/ufs/ufsmount.h> 56115103Strhodes#include <ufs/ufs/ufs_extern.h> 57115103Strhodes 58115103Strhodes#define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1)) 59115103Strhodes#define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 60115103Strhodes#define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n)) 61115103Strhodes 62115103Strhodesstatic MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables"); 63115103Strhodes 641638SrgrimesSYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem"); 651638Srgrimes 661638Srgrimesstatic int ufs_mindirhashsize = DIRBLKSIZ * 5; 671638SrgrimesSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW, 68115103Strhodes &ufs_mindirhashsize, 691638Srgrimes 0, "minimum directory size in bytes for which to use hashed lookup"); 701638Srgrimesstatic int ufs_dirhashmaxmem = 2 * 1024 * 1024; 711638SrgrimesSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 721638Srgrimes 0, "maximum allowed dirhash memory usage"); 731638Srgrimesstatic int ufs_dirhashmem; 741638SrgrimesSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 751638Srgrimes 0, "current dirhash memory usage"); 761638Srgrimesstatic int ufs_dirhashcheck = 0; 771638SrgrimesSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 781638Srgrimes 0, "enable extra sanity tests"); 791638Srgrimes 801638Srgrimes 811638Srgrimesstatic int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 821638Srgrimesstatic void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 831638Srgrimesstatic void ufsdirhash_delslot(struct dirhash *dh, int slot); 841638Srgrimesstatic int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 851638Srgrimes doff_t offset); 861638Srgrimesstatic doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 871638Srgrimesstatic void ufsdirhash_init(void); 881638Srgrimesstatic int ufsdirhash_recycle(int wanted); 891638Srgrimes 901638Srgrimesstatic uma_zone_t ufsdirhash_zone; 911638Srgrimes 921638Srgrimes/* Dirhash list; recently-used entries are near the tail. */ 931638Srgrimesstatic TAILQ_HEAD(, dirhash) ufsdirhash_list; 941638Srgrimes 951638Srgrimes/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 96208420Smaximstatic struct mtx ufsdirhash_mtx; 97208420Smaxim 98208420Smaxim/* 99208420Smaxim * Locking order: 100208420Smaxim * ufsdirhash_mtx 101208417Smaxim * dh_mtx 1021638Srgrimes * 1031638Srgrimes * The dh_mtx mutex should be aquired either via the inode lock, or via 1041638Srgrimes * ufsdirhash_mtx. Only the owner of the inode may free the associated 1051638Srgrimes * dirhash, but anything can steal its memory and set dh_hash to NULL. 1061638Srgrimes */ 1071638Srgrimes 1081638Srgrimes/* 1091638Srgrimes * Attempt to build up a hash table for the directory contents in 1101638Srgrimes * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 1111638Srgrimes */ 1121638Srgrimesint 1131638Srgrimesufsdirhash_build(struct inode *ip) 1141638Srgrimes{ 1151638Srgrimes struct dirhash *dh; 1161638Srgrimes struct buf *bp = NULL; 1171638Srgrimes struct direct *ep; 1181638Srgrimes struct vnode *vp; 1191638Srgrimes doff_t bmask, pos; 1201638Srgrimes int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 1211638Srgrimes 1221638Srgrimes /* Check if we can/should use dirhash. */ 1231638Srgrimes if (ip->i_dirhash == NULL) { 12483075Sru if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode)) 1251638Srgrimes return (-1); 12683075Sru } else { 12783075Sru /* Hash exists, but sysctls could have changed. */ 1281638Srgrimes if (ip->i_size < ufs_mindirhashsize || 12983075Sru ufs_dirhashmem > ufs_dirhashmaxmem) { 13083075Sru ufsdirhash_free(ip); 13183075Sru return (-1); 1321638Srgrimes } 1331638Srgrimes /* Check if hash exists and is intact (note: unlocked read). */ 1341638Srgrimes if (ip->i_dirhash->dh_hash != NULL) 1351638Srgrimes return (0); 1361638Srgrimes /* Free the old, recycled hash and build a new one. */ 1371638Srgrimes ufsdirhash_free(ip); 1381638Srgrimes } 1391638Srgrimes 1401638Srgrimes /* Don't hash removed directories. */ 1411638Srgrimes if (ip->i_effnlink == 0) 1421638Srgrimes return (-1); 1431638Srgrimes 1441638Srgrimes vp = ip->i_vnode; 1451638Srgrimes /* Allocate 50% more entries than this dir size could ever need. */ 14674942Sru KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 1471638Srgrimes nslots = ip->i_size / DIRECTSIZ(1); 14858494Sru nslots = (nslots * 3 + 1) / 2; 1491638Srgrimes narrays = howmany(nslots, DH_NBLKOFF); 1501638Srgrimes nslots = narrays * DH_NBLKOFF; 1511638Srgrimes dirblocks = howmany(ip->i_size, DIRBLKSIZ); 15274942Sru nblocks = (dirblocks * 3 + 1) / 2; 15374942Sru 1541638Srgrimes memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 1551638Srgrimes narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 1561638Srgrimes nblocks * sizeof(*dh->dh_blkfree); 1571638Srgrimes mtx_lock(&ufsdirhash_mtx); 1581638Srgrimes if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 1591638Srgrimes mtx_unlock(&ufsdirhash_mtx); 1601638Srgrimes if (memreqd > ufs_dirhashmaxmem / 2) 1611638Srgrimes return (-1); 1621638Srgrimes 1631638Srgrimes /* Try to free some space. */ 1641638Srgrimes if (ufsdirhash_recycle(memreqd) != 0) 1651638Srgrimes return (-1); 1661638Srgrimes /* Enough was freed, and ufsdirhash_mtx has been locked. */ 1671638Srgrimes } 1681638Srgrimes ufs_dirhashmem += memreqd; 1691638Srgrimes mtx_unlock(&ufsdirhash_mtx); 1701638Srgrimes 1711638Srgrimes /* 1721638Srgrimes * Use non-blocking mallocs so that we will revert to a linear 1731638Srgrimes * lookup on failure rather than potentially blocking forever. 1741638Srgrimes */ 1751638Srgrimes MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO); 1761638Srgrimes if (dh == NULL) 1771638Srgrimes return (-1); 1781638Srgrimes MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]), 1791638Srgrimes M_DIRHASH, M_NOWAIT | M_ZERO); 1801638Srgrimes MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]), 1811638Srgrimes M_DIRHASH, M_NOWAIT); 1821638Srgrimes if (dh->dh_hash == NULL || dh->dh_blkfree == NULL) 1831638Srgrimes goto fail; 1841638Srgrimes for (i = 0; i < narrays; i++) { 1851638Srgrimes if ((dh->dh_hash[i] = uma_zalloc(ufsdirhash_zone, 1861638Srgrimes M_WAITOK)) == NULL) 1871638Srgrimes goto fail; 1881638Srgrimes for (j = 0; j < DH_NBLKOFF; j++) 1891638Srgrimes dh->dh_hash[i][j] = DIRHASH_EMPTY; 1901638Srgrimes } 1911638Srgrimes 192159721Syar /* Initialise the hash table and block statistics. */ 1931638Srgrimes mtx_init(&dh->dh_mtx, "dirhash", MTX_DEF); 1941638Srgrimes dh->dh_narrays = narrays; 1951638Srgrimes dh->dh_hlen = nslots; 1961638Srgrimes dh->dh_nblk = nblocks; 1971638Srgrimes dh->dh_dirblks = dirblocks; 1981638Srgrimes for (i = 0; i < dirblocks; i++) 1991638Srgrimes dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 2001638Srgrimes for (i = 0; i < DH_NFSTATS; i++) 2011638Srgrimes dh->dh_firstfree[i] = -1; 2021638Srgrimes dh->dh_firstfree[DH_NFSTATS] = 0; 2031638Srgrimes dh->dh_seqopt = 0; 2041638Srgrimes dh->dh_seqoff = 0; 2051638Srgrimes dh->dh_score = DH_SCOREINIT; 2061638Srgrimes ip->i_dirhash = dh; 2071638Srgrimes 2081638Srgrimes bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 2091638Srgrimes pos = 0; 2101638Srgrimes while (pos < ip->i_size) { 2111638Srgrimes /* If necessary, get the next directory block. */ 2121638Srgrimes if ((pos & bmask) == 0) { 2131638Srgrimes if (bp != NULL) 2141638Srgrimes brelse(bp); 2151638Srgrimes if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 2161638Srgrimes goto fail; 2171638Srgrimes } 21816437Sphk 21916437Sphk /* Add this entry to the hash. */ 22016437Sphk ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 2211638Srgrimes if (ep->d_reclen == 0 || ep->d_reclen > 222133653Sru DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 2231638Srgrimes /* Corrupted directory. */ 22488055Sru brelse(bp); 22588055Sru goto fail; 22688055Sru } 22788055Sru if (ep->d_ino != 0) { 22888055Sru /* Add the entry (simplified ufsdirhash_add). */ 2291638Srgrimes slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 2301638Srgrimes while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 2311638Srgrimes slot = WRAPINCR(slot, dh->dh_hlen); 2321638Srgrimes dh->dh_hused++; 233235928Smarcel DH_ENTRY(dh, slot) = pos; 2341638Srgrimes ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 2351638Srgrimes } 2361638Srgrimes pos += ep->d_reclen; 2371638Srgrimes } 2381638Srgrimes 2391638Srgrimes if (bp != NULL) 2401638Srgrimes brelse(bp); 2411638Srgrimes mtx_lock(&ufsdirhash_mtx); 2421638Srgrimes TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 2431638Srgrimes dh->dh_onlist = 1; 24474942Sru mtx_unlock(&ufsdirhash_mtx); 24574942Sru return (0); 2461638Srgrimes 2471638Srgrimesfail: 2481638Srgrimes if (dh->dh_hash != NULL) { 2491638Srgrimes for (i = 0; i < narrays; i++) 25094424Sru if (dh->dh_hash[i] != NULL) 25194424Sru uma_zfree(ufsdirhash_zone, dh->dh_hash[i]); 25294424Sru FREE(dh->dh_hash, M_DIRHASH); 25394424Sru } 25494424Sru if (dh->dh_blkfree != NULL) 255289052Sbdrewery FREE(dh->dh_blkfree, M_DIRHASH); 256289052Sbdrewery FREE(dh, M_DIRHASH); 257289052Sbdrewery ip->i_dirhash = NULL; 258289052Sbdrewery mtx_lock(&ufsdirhash_mtx); 259289052Sbdrewery ufs_dirhashmem -= memreqd; 260289052Sbdrewery mtx_unlock(&ufsdirhash_mtx); 261289052Sbdrewery return (-1); 262289052Sbdrewery} 263289052Sbdrewery 264289052Sbdrewery/* 265289052Sbdrewery * Free any hash table associated with inode 'ip'. 266289052Sbdrewery */ 267289052Sbdreweryvoid 268289052Sbdreweryufsdirhash_free(struct inode *ip) 26975083Sru{ 27075083Sru struct dirhash *dh; 27175083Sru int i, mem; 27253965Smharo 27394424Sru if ((dh = ip->i_dirhash) == NULL) 27494424Sru return; 2751638Srgrimes mtx_lock(&ufsdirhash_mtx); 2761638Srgrimes mtx_lock(&dh->dh_mtx); 2771638Srgrimes if (dh->dh_onlist) 2781638Srgrimes TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 2791638Srgrimes mtx_unlock(&dh->dh_mtx); 280235928Smarcel mtx_unlock(&ufsdirhash_mtx); 2811638Srgrimes 28214701Sbde /* The dirhash pointed to by 'dh' is exclusively ours now. */ 28314701Sbde 28414701Sbde mem = sizeof(*dh); 28514701Sbde if (dh->dh_hash != NULL) { 2861638Srgrimes for (i = 0; i < dh->dh_narrays; i++) 28714701Sbde uma_zfree(ufsdirhash_zone, dh->dh_hash[i]); 28814701Sbde FREE(dh->dh_hash, M_DIRHASH); 28914701Sbde FREE(dh->dh_blkfree, M_DIRHASH); 29014701Sbde mem += dh->dh_narrays * sizeof(*dh->dh_hash) + 29114701Sbde dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 29214701Sbde dh->dh_nblk * sizeof(*dh->dh_blkfree); 29314701Sbde } 29414701Sbde mtx_destroy(&dh->dh_mtx); 2951638Srgrimes FREE(dh, M_DIRHASH); 2961638Srgrimes ip->i_dirhash = NULL; 297125493Sru 298125493Sru mtx_lock(&ufsdirhash_mtx); 299125493Sru ufs_dirhashmem -= mem; 3001638Srgrimes mtx_unlock(&ufsdirhash_mtx); 3011638Srgrimes} 3021638Srgrimes 3031638Srgrimes/* 3041638Srgrimes * Find the offset of the specified name within the given inode. 30575284Sru * Returns 0 on success, ENOENT if the entry does not exist, or 30675284Sru * EJUSTRETURN if the caller should revert to a linear search. 30775284Sru * 30875284Sru * If successful, the directory offset is stored in *offp, and a 30975284Sru * pointer to a struct buf containing the entry is stored in *bpp. If 31074942Sru * prevoffp is non-NULL, the offset of the previous entry within 31174942Sru * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 3121638Srgrimes * is the first in a block, the start of the block is used). 3131638Srgrimes */ 3141638Srgrimesint 3151638Srgrimesufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 3161638Srgrimes struct buf **bpp, doff_t *prevoffp) 3171638Srgrimes{ 3181638Srgrimes struct dirhash *dh, *dh_next; 3191638Srgrimes struct direct *dp; 3201638Srgrimes struct vnode *vp; 3211638Srgrimes struct buf *bp; 3221638Srgrimes doff_t blkoff, bmask, offset, prevoff; 323156743Sru int i, slot; 3241638Srgrimes 3251638Srgrimes if ((dh = ip->i_dirhash) == NULL) 3261638Srgrimes return (EJUSTRETURN); 327276486Sngie /* 3281638Srgrimes * Move this dirhash towards the end of the list if it has a 3291638Srgrimes * score higher than the next entry, and aquire the dh_mtx. 3301638Srgrimes * Optimise the case where it's already the last by performing 3311638Srgrimes * an unlocked read of the TAILQ_NEXT pointer. 3321638Srgrimes * 3331638Srgrimes * In both cases, end up holding just dh_mtx. 3341638Srgrimes */ 3351638Srgrimes if (TAILQ_NEXT(dh, dh_list) != NULL) { 3361638Srgrimes mtx_lock(&ufsdirhash_mtx); 3371638Srgrimes mtx_lock(&dh->dh_mtx); 3381638Srgrimes /* 3391638Srgrimes * If the new score will be greater than that of the next 3401638Srgrimes * entry, then move this entry past it. With both mutexes 3411638Srgrimes * held, dh_next won't go away, but its dh_score could 3421638Srgrimes * change; that's not important since it is just a hint. 3431638Srgrimes */ 3441638Srgrimes if (dh->dh_hash != NULL && 3451638Srgrimes (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 3461638Srgrimes dh->dh_score >= dh_next->dh_score) { 3471638Srgrimes KASSERT(dh->dh_onlist, ("dirhash: not on list")); 3481638Srgrimes TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 3491638Srgrimes TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 3501638Srgrimes dh_list); 3511638Srgrimes } 352281056Sbdrewery mtx_unlock(&ufsdirhash_mtx); 353281056Sbdrewery } else { 354281056Sbdrewery /* Already the last, though that could change as we wait. */ 355281056Sbdrewery mtx_lock(&dh->dh_mtx); 356281056Sbdrewery } 357281056Sbdrewery if (dh->dh_hash == NULL) { 358281056Sbdrewery mtx_unlock(&dh->dh_mtx); 3591638Srgrimes ufsdirhash_free(ip); 3601638Srgrimes return (EJUSTRETURN); 3611638Srgrimes } 3621638Srgrimes 3631638Srgrimes /* Update the score. */ 3641638Srgrimes if (dh->dh_score < DH_SCOREMAX) 3651638Srgrimes dh->dh_score++; 3661638Srgrimes 3671638Srgrimes vp = ip->i_vnode; 3681638Srgrimes bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 3691638Srgrimes blkoff = -1; 3701638Srgrimes bp = NULL; 37174942Srurestart: 3721638Srgrimes slot = ufsdirhash_hash(dh, name, namelen); 3731638Srgrimes 3741638Srgrimes if (dh->dh_seqopt) { 3751638Srgrimes /* 3761638Srgrimes * Sequential access optimisation. dh_seqoff contains the 3771638Srgrimes * offset of the directory entry immediately following 378235122Sjlh * the last entry that was looked up. Check if this offset 379235122Sjlh * appears in the hash chain for the name we are looking for. 380235122Sjlh */ 381235122Sjlh for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 382245561Sbrooks i = WRAPINCR(i, dh->dh_hlen)) 383245561Sbrooks if (offset == dh->dh_seqoff) 38474942Sru break; 38574942Sru if (offset == dh->dh_seqoff) { 3861638Srgrimes /* 3871638Srgrimes * We found an entry with the expected offset. This 3881638Srgrimes * is probably the entry we want, but if not, the 3891638Srgrimes * code below will turn off seqoff and retry. 3901638Srgrimes */ 391264483Sjmmv slot = i; 392264483Sjmmv } else 393264483Sjmmv dh->dh_seqopt = 0; 394264483Sjmmv } 395264483Sjmmv 396264483Sjmmv for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 397264483Sjmmv slot = WRAPINCR(slot, dh->dh_hlen)) { 398264483Sjmmv if (offset == DIRHASH_DEL) 399264483Sjmmv continue; 400264483Sjmmv mtx_unlock(&dh->dh_mtx); 401264483Sjmmv 402264483Sjmmv if (offset < 0 || offset >= ip->i_size) 403264483Sjmmv panic("ufsdirhash_lookup: bad offset in hash array"); 404264483Sjmmv if ((offset & ~bmask) != blkoff) { 405264483Sjmmv if (bp != NULL) 406264483Sjmmv brelse(bp); 407264483Sjmmv blkoff = offset & ~bmask; 408264483Sjmmv if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) 409264483Sjmmv return (EJUSTRETURN); 410264483Sjmmv } 411264483Sjmmv dp = (struct direct *)(bp->b_data + (offset & bmask)); 412264483Sjmmv if (dp->d_reclen == 0 || dp->d_reclen > 413264483Sjmmv DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 414264483Sjmmv /* Corrupted directory. */ 415264483Sjmmv brelse(bp); 416264483Sjmmv return (EJUSTRETURN); 417264483Sjmmv } 418264483Sjmmv if (dp->d_namlen == namelen && 419264483Sjmmv bcmp(dp->d_name, name, namelen) == 0) { 420264483Sjmmv /* Found. Get the prev offset if needed. */ 421264483Sjmmv if (prevoffp != NULL) { 422264483Sjmmv if (offset & (DIRBLKSIZ - 1)) { 423264483Sjmmv prevoff = ufsdirhash_getprev(dp, 424264483Sjmmv offset); 425264483Sjmmv if (prevoff == -1) { 426264483Sjmmv brelse(bp); 427264483Sjmmv return (EJUSTRETURN); 428292278Sngie } 429292278Sngie } else 430292278Sngie prevoff = offset; 431264483Sjmmv *prevoffp = prevoff; 432264483Sjmmv } 433264483Sjmmv 434292278Sngie /* Check for sequential access, and update offset. */ 435292278Sngie if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 436292278Sngie dh->dh_seqopt = 1; 437292278Sngie dh->dh_seqoff = offset + DIRSIZ(0, dp); 438264483Sjmmv 439264483Sjmmv *bpp = bp; 440264483Sjmmv *offp = offset; 441264483Sjmmv return (0); 442264483Sjmmv } 443264483Sjmmv 444264483Sjmmv mtx_lock(&dh->dh_mtx); 445292812Sngie if (dh->dh_hash == NULL) { 446292812Sngie mtx_unlock(&dh->dh_mtx); 447292812Sngie if (bp != NULL) 448292812Sngie brelse(bp); 449264483Sjmmv ufsdirhash_free(ip); 450264483Sjmmv return (EJUSTRETURN); 451264483Sjmmv } 452264483Sjmmv /* 453264483Sjmmv * When the name doesn't match in the seqopt case, go back 454264483Sjmmv * and search normally. 455264483Sjmmv */ 456264483Sjmmv if (dh->dh_seqopt) { 457264483Sjmmv dh->dh_seqopt = 0; 458264483Sjmmv goto restart; 459264483Sjmmv } 460264483Sjmmv } 461264483Sjmmv mtx_unlock(&dh->dh_mtx); 462264483Sjmmv if (bp != NULL) 463264483Sjmmv brelse(bp); 464264483Sjmmv return (ENOENT); 465264483Sjmmv} 466264483Sjmmv 467264483Sjmmv/* 468264483Sjmmv * Find a directory block with room for 'slotneeded' bytes. Returns 469264483Sjmmv * the offset of the directory entry that begins the free space. 470264483Sjmmv * This will either be the offset of an existing entry that has free 471264483Sjmmv * space at the end, or the offset of an entry with d_ino == 0 at 472264483Sjmmv * the start of a DIRBLKSIZ block. 473264483Sjmmv * 474264483Sjmmv * To use the space, the caller may need to compact existing entries in 475264483Sjmmv * the directory. The total number of bytes in all of the entries involved 476264483Sjmmv * in the compaction is stored in *slotsize. In other words, all of 477264483Sjmmv * the entries that must be compacted are exactly contained in the 478264483Sjmmv * region beginning at the returned offset and spanning *slotsize bytes. 479264483Sjmmv * 480264483Sjmmv * Returns -1 if no space was found, indicating that the directory 481264483Sjmmv * must be extended. 482264483Sjmmv */ 483264483Sjmmvdoff_t 484264483Sjmmvufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 485264483Sjmmv{ 486264483Sjmmv struct direct *dp; 487264483Sjmmv struct dirhash *dh; 488264483Sjmmv struct buf *bp; 489264483Sjmmv doff_t pos, slotstart; 490264483Sjmmv int dirblock, error, freebytes, i; 491264483Sjmmv 492264483Sjmmv if ((dh = ip->i_dirhash) == NULL) 493264483Sjmmv return (-1); 494264483Sjmmv mtx_lock(&dh->dh_mtx); 495 if (dh->dh_hash == NULL) { 496 mtx_unlock(&dh->dh_mtx); 497 ufsdirhash_free(ip); 498 return (-1); 499 } 500 501 /* Find a directory block with the desired free space. */ 502 dirblock = -1; 503 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 504 if ((dirblock = dh->dh_firstfree[i]) != -1) 505 break; 506 if (dirblock == -1) { 507 mtx_unlock(&dh->dh_mtx); 508 return (-1); 509 } 510 511 KASSERT(dirblock < dh->dh_nblk && 512 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 513 ("ufsdirhash_findfree: bad stats")); 514 mtx_unlock(&dh->dh_mtx); 515 pos = dirblock * DIRBLKSIZ; 516 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 517 if (error) 518 return (-1); 519 520 /* Find the first entry with free space. */ 521 for (i = 0; i < DIRBLKSIZ; ) { 522 if (dp->d_reclen == 0) { 523 brelse(bp); 524 return (-1); 525 } 526 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 527 break; 528 i += dp->d_reclen; 529 dp = (struct direct *)((char *)dp + dp->d_reclen); 530 } 531 if (i > DIRBLKSIZ) { 532 brelse(bp); 533 return (-1); 534 } 535 slotstart = pos + i; 536 537 /* Find the range of entries needed to get enough space */ 538 freebytes = 0; 539 while (i < DIRBLKSIZ && freebytes < slotneeded) { 540 freebytes += dp->d_reclen; 541 if (dp->d_ino != 0) 542 freebytes -= DIRSIZ(0, dp); 543 if (dp->d_reclen == 0) { 544 brelse(bp); 545 return (-1); 546 } 547 i += dp->d_reclen; 548 dp = (struct direct *)((char *)dp + dp->d_reclen); 549 } 550 if (i > DIRBLKSIZ) { 551 brelse(bp); 552 return (-1); 553 } 554 if (freebytes < slotneeded) 555 panic("ufsdirhash_findfree: free mismatch"); 556 brelse(bp); 557 *slotsize = pos + i - slotstart; 558 return (slotstart); 559} 560 561/* 562 * Return the start of the unused space at the end of a directory, or 563 * -1 if there are no trailing unused blocks. 564 */ 565doff_t 566ufsdirhash_enduseful(struct inode *ip) 567{ 568 569 struct dirhash *dh; 570 int i; 571 572 if ((dh = ip->i_dirhash) == NULL) 573 return (-1); 574 mtx_lock(&dh->dh_mtx); 575 if (dh->dh_hash == NULL) { 576 mtx_unlock(&dh->dh_mtx); 577 ufsdirhash_free(ip); 578 return (-1); 579 } 580 581 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) { 582 mtx_unlock(&dh->dh_mtx); 583 return (-1); 584 } 585 586 for (i = dh->dh_dirblks - 1; i >= 0; i--) 587 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 588 break; 589 mtx_unlock(&dh->dh_mtx); 590 return ((doff_t)(i + 1) * DIRBLKSIZ); 591} 592 593/* 594 * Insert information into the hash about a new directory entry. dirp 595 * points to a struct direct containing the entry, and offset specifies 596 * the offset of this entry. 597 */ 598void 599ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 600{ 601 struct dirhash *dh; 602 int slot; 603 604 if ((dh = ip->i_dirhash) == NULL) 605 return; 606 mtx_lock(&dh->dh_mtx); 607 if (dh->dh_hash == NULL) { 608 mtx_unlock(&dh->dh_mtx); 609 ufsdirhash_free(ip); 610 return; 611 } 612 613 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 614 ("ufsdirhash_add: bad offset")); 615 /* 616 * Normal hash usage is < 66%. If the usage gets too high then 617 * remove the hash entirely and let it be rebuilt later. 618 */ 619 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 620 mtx_unlock(&dh->dh_mtx); 621 ufsdirhash_free(ip); 622 return; 623 } 624 625 /* Find a free hash slot (empty or deleted), and add the entry. */ 626 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 627 while (DH_ENTRY(dh, slot) >= 0) 628 slot = WRAPINCR(slot, dh->dh_hlen); 629 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 630 dh->dh_hused++; 631 DH_ENTRY(dh, slot) = offset; 632 633 /* Update the per-block summary info. */ 634 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 635 mtx_unlock(&dh->dh_mtx); 636} 637 638/* 639 * Remove the specified directory entry from the hash. The entry to remove 640 * is defined by the name in `dirp', which must exist at the specified 641 * `offset' within the directory. 642 */ 643void 644ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 645{ 646 struct dirhash *dh; 647 int slot; 648 649 if ((dh = ip->i_dirhash) == NULL) 650 return; 651 mtx_lock(&dh->dh_mtx); 652 if (dh->dh_hash == NULL) { 653 mtx_unlock(&dh->dh_mtx); 654 ufsdirhash_free(ip); 655 return; 656 } 657 658 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 659 ("ufsdirhash_remove: bad offset")); 660 /* Find the entry */ 661 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 662 663 /* Remove the hash entry. */ 664 ufsdirhash_delslot(dh, slot); 665 666 /* Update the per-block summary info. */ 667 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 668 mtx_unlock(&dh->dh_mtx); 669} 670 671/* 672 * Change the offset associated with a directory entry in the hash. Used 673 * when compacting directory blocks. 674 */ 675void 676ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 677 doff_t newoff) 678{ 679 struct dirhash *dh; 680 int slot; 681 682 if ((dh = ip->i_dirhash) == NULL) 683 return; 684 mtx_lock(&dh->dh_mtx); 685 if (dh->dh_hash == NULL) { 686 mtx_unlock(&dh->dh_mtx); 687 ufsdirhash_free(ip); 688 return; 689 } 690 691 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 692 newoff < dh->dh_dirblks * DIRBLKSIZ, 693 ("ufsdirhash_move: bad offset")); 694 /* Find the entry, and update the offset. */ 695 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 696 DH_ENTRY(dh, slot) = newoff; 697 mtx_unlock(&dh->dh_mtx); 698} 699 700/* 701 * Inform dirhash that the directory has grown by one block that 702 * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 703 */ 704void 705ufsdirhash_newblk(struct inode *ip, doff_t offset) 706{ 707 struct dirhash *dh; 708 int block; 709 710 if ((dh = ip->i_dirhash) == NULL) 711 return; 712 mtx_lock(&dh->dh_mtx); 713 if (dh->dh_hash == NULL) { 714 mtx_unlock(&dh->dh_mtx); 715 ufsdirhash_free(ip); 716 return; 717 } 718 719 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 720 ("ufsdirhash_newblk: bad offset")); 721 block = offset / DIRBLKSIZ; 722 if (block >= dh->dh_nblk) { 723 /* Out of space; must rebuild. */ 724 mtx_unlock(&dh->dh_mtx); 725 ufsdirhash_free(ip); 726 return; 727 } 728 dh->dh_dirblks = block + 1; 729 730 /* Account for the new free block. */ 731 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 732 if (dh->dh_firstfree[DH_NFSTATS] == -1) 733 dh->dh_firstfree[DH_NFSTATS] = block; 734 mtx_unlock(&dh->dh_mtx); 735} 736 737/* 738 * Inform dirhash that the directory is being truncated. 739 */ 740void 741ufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 742{ 743 struct dirhash *dh; 744 int block, i; 745 746 if ((dh = ip->i_dirhash) == NULL) 747 return; 748 mtx_lock(&dh->dh_mtx); 749 if (dh->dh_hash == NULL) { 750 mtx_unlock(&dh->dh_mtx); 751 ufsdirhash_free(ip); 752 return; 753 } 754 755 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 756 ("ufsdirhash_dirtrunc: bad offset")); 757 block = howmany(offset, DIRBLKSIZ); 758 /* 759 * If the directory shrinks to less than 1/8 of dh_nblk blocks 760 * (about 20% of its original size due to the 50% extra added in 761 * ufsdirhash_build) then free it, and let the caller rebuild 762 * if necessary. 763 */ 764 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 765 mtx_unlock(&dh->dh_mtx); 766 ufsdirhash_free(ip); 767 return; 768 } 769 770 /* 771 * Remove any `first free' information pertaining to the 772 * truncated blocks. All blocks we're removing should be 773 * completely unused. 774 */ 775 if (dh->dh_firstfree[DH_NFSTATS] >= block) 776 dh->dh_firstfree[DH_NFSTATS] = -1; 777 for (i = block; i < dh->dh_dirblks; i++) 778 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 779 panic("ufsdirhash_dirtrunc: blocks in use"); 780 for (i = 0; i < DH_NFSTATS; i++) 781 if (dh->dh_firstfree[i] >= block) 782 panic("ufsdirhash_dirtrunc: first free corrupt"); 783 dh->dh_dirblks = block; 784 mtx_unlock(&dh->dh_mtx); 785} 786 787/* 788 * Debugging function to check that the dirhash information about 789 * a directory block matches its actual contents. Panics if a mismatch 790 * is detected. 791 * 792 * On entry, `buf' should point to the start of an in-core 793 * DIRBLKSIZ-sized directory block, and `offset' should contain the 794 * offset from the start of the directory of that block. 795 */ 796void 797ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 798{ 799 struct dirhash *dh; 800 struct direct *dp; 801 int block, ffslot, i, nfree; 802 803 if (!ufs_dirhashcheck) 804 return; 805 if ((dh = ip->i_dirhash) == NULL) 806 return; 807 mtx_lock(&dh->dh_mtx); 808 if (dh->dh_hash == NULL) { 809 mtx_unlock(&dh->dh_mtx); 810 ufsdirhash_free(ip); 811 return; 812 } 813 814 block = offset / DIRBLKSIZ; 815 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 816 panic("ufsdirhash_checkblock: bad offset"); 817 818 nfree = 0; 819 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 820 dp = (struct direct *)(buf + i); 821 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 822 panic("ufsdirhash_checkblock: bad dir"); 823 824 if (dp->d_ino == 0) { 825#if 0 826 /* 827 * XXX entries with d_ino == 0 should only occur 828 * at the start of a DIRBLKSIZ block. However the 829 * ufs code is tolerant of such entries at other 830 * offsets, and fsck does not fix them. 831 */ 832 if (i != 0) 833 panic("ufsdirhash_checkblock: bad dir inode"); 834#endif 835 nfree += dp->d_reclen; 836 continue; 837 } 838 839 /* Check that the entry exists (will panic if it doesn't). */ 840 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 841 842 nfree += dp->d_reclen - DIRSIZ(0, dp); 843 } 844 if (i != DIRBLKSIZ) 845 panic("ufsdirhash_checkblock: bad dir end"); 846 847 if (dh->dh_blkfree[block] * DIRALIGN != nfree) 848 panic("ufsdirhash_checkblock: bad free count"); 849 850 ffslot = BLKFREE2IDX(nfree / DIRALIGN); 851 for (i = 0; i <= DH_NFSTATS; i++) 852 if (dh->dh_firstfree[i] == block && i != ffslot) 853 panic("ufsdirhash_checkblock: bad first-free"); 854 if (dh->dh_firstfree[ffslot] == -1) 855 panic("ufsdirhash_checkblock: missing first-free entry"); 856 mtx_unlock(&dh->dh_mtx); 857} 858 859/* 860 * Hash the specified filename into a dirhash slot. 861 */ 862static int 863ufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 864{ 865 return (fnv_32_buf(name, namelen, FNV1_32_INIT) % dh->dh_hlen); 866} 867 868/* 869 * Adjust the number of free bytes in the block containing `offset' 870 * by the value specified by `diff'. 871 * 872 * The caller must ensure we have exclusive access to `dh'; normally 873 * that means that dh_mtx should be held, but this is also called 874 * from ufsdirhash_build() where exclusive access can be assumed. 875 */ 876static void 877ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 878{ 879 int block, i, nfidx, ofidx; 880 881 /* Update the per-block summary info. */ 882 block = offset / DIRBLKSIZ; 883 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 884 ("dirhash bad offset")); 885 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 886 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 887 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 888 889 /* Update the `first free' list if necessary. */ 890 if (ofidx != nfidx) { 891 /* If removing, scan forward for the next block. */ 892 if (dh->dh_firstfree[ofidx] == block) { 893 for (i = block + 1; i < dh->dh_dirblks; i++) 894 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 895 break; 896 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 897 } 898 899 /* Make this the new `first free' if necessary */ 900 if (dh->dh_firstfree[nfidx] > block || 901 dh->dh_firstfree[nfidx] == -1) 902 dh->dh_firstfree[nfidx] = block; 903 } 904} 905 906/* 907 * Find the specified name which should have the specified offset. 908 * Returns a slot number, and panics on failure. 909 * 910 * `dh' must be locked on entry and remains so on return. 911 */ 912static int 913ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 914{ 915 int slot; 916 917 mtx_assert(&dh->dh_mtx, MA_OWNED); 918 919 /* Find the entry. */ 920 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 921 slot = ufsdirhash_hash(dh, name, namelen); 922 while (DH_ENTRY(dh, slot) != offset && 923 DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 924 slot = WRAPINCR(slot, dh->dh_hlen); 925 if (DH_ENTRY(dh, slot) != offset) 926 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 927 928 return (slot); 929} 930 931/* 932 * Remove the entry corresponding to the specified slot from the hash array. 933 * 934 * `dh' must be locked on entry and remains so on return. 935 */ 936static void 937ufsdirhash_delslot(struct dirhash *dh, int slot) 938{ 939 int i; 940 941 mtx_assert(&dh->dh_mtx, MA_OWNED); 942 943 /* Mark the entry as deleted. */ 944 DH_ENTRY(dh, slot) = DIRHASH_DEL; 945 946 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 947 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 948 i = WRAPINCR(i, dh->dh_hlen); 949 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 950 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) { 951 DH_ENTRY(dh, i) = DIRHASH_EMPTY; 952 dh->dh_hused--; 953 i = WRAPINCR(i, dh->dh_hlen); 954 } 955 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 956 } 957} 958 959/* 960 * Given a directory entry and its offset, find the offset of the 961 * previous entry in the same DIRBLKSIZ-sized block. Returns an 962 * offset, or -1 if there is no previous entry in the block or some 963 * other problem occurred. 964 */ 965static doff_t 966ufsdirhash_getprev(struct direct *dirp, doff_t offset) 967{ 968 struct direct *dp; 969 char *blkbuf; 970 doff_t blkoff, prevoff; 971 int entrypos, i; 972 973 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 974 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 975 blkbuf = (char *)dirp - entrypos; 976 prevoff = blkoff; 977 978 /* If `offset' is the start of a block, there is no previous entry. */ 979 if (entrypos == 0) 980 return (-1); 981 982 /* Scan from the start of the block until we get to the entry. */ 983 for (i = 0; i < entrypos; i += dp->d_reclen) { 984 dp = (struct direct *)(blkbuf + i); 985 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 986 return (-1); /* Corrupted directory. */ 987 prevoff = blkoff + i; 988 } 989 return (prevoff); 990} 991 992/* 993 * Try to free up `wanted' bytes by stealing memory from existing 994 * dirhashes. Returns zero with ufsdirhash_mtx locked if successful. 995 */ 996static int 997ufsdirhash_recycle(int wanted) 998{ 999 struct dirhash *dh; 1000 doff_t **hash; 1001 u_int8_t *blkfree; 1002 int i, mem, narrays; 1003 1004 mtx_lock(&ufsdirhash_mtx); 1005 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 1006 /* Find a dirhash, and lock it. */ 1007 if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) { 1008 mtx_unlock(&ufsdirhash_mtx); 1009 return (-1); 1010 } 1011 mtx_lock(&dh->dh_mtx); 1012 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 1013 1014 /* Decrement the score; only recycle if it becomes zero. */ 1015 if (--dh->dh_score > 0) { 1016 mtx_unlock(&dh->dh_mtx); 1017 mtx_unlock(&ufsdirhash_mtx); 1018 return (-1); 1019 } 1020 1021 /* Remove it from the list and detach its memory. */ 1022 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 1023 dh->dh_onlist = 0; 1024 hash = dh->dh_hash; 1025 dh->dh_hash = NULL; 1026 blkfree = dh->dh_blkfree; 1027 dh->dh_blkfree = NULL; 1028 narrays = dh->dh_narrays; 1029 mem = narrays * sizeof(*dh->dh_hash) + 1030 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 1031 dh->dh_nblk * sizeof(*dh->dh_blkfree); 1032 1033 /* Unlock everything, free the detached memory. */ 1034 mtx_unlock(&dh->dh_mtx); 1035 mtx_unlock(&ufsdirhash_mtx); 1036 for (i = 0; i < narrays; i++) 1037 uma_zfree(ufsdirhash_zone, hash[i]); 1038 FREE(hash, M_DIRHASH); 1039 FREE(blkfree, M_DIRHASH); 1040 1041 /* Account for the returned memory, and repeat if necessary. */ 1042 mtx_lock(&ufsdirhash_mtx); 1043 ufs_dirhashmem -= mem; 1044 } 1045 /* Success; return with ufsdirhash_mtx locked. */ 1046 return (0); 1047} 1048 1049 1050static void 1051ufsdirhash_init() 1052{ 1053 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(daddr_t), 1054 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1055 mtx_init(&ufsdirhash_mtx, "dirhash list", MTX_DEF); 1056 TAILQ_INIT(&ufsdirhash_list); 1057} 1058SYSINIT(ufsdirhash, SI_SUB_PSEUDO, SI_ORDER_ANY, ufsdirhash_init, NULL) 1059 1060 1061#endif /* UFS_DIRHASH */ 1062