ufs_dirhash.c revision 84811
179561Siedowse/* 279561Siedowse * Copyright (c) 2001 Ian Dowse. All rights reserved. 379561Siedowse * 479561Siedowse * Redistribution and use in source and binary forms, with or without 579561Siedowse * modification, are permitted provided that the following conditions 679561Siedowse * are met: 779561Siedowse * 1. Redistributions of source code must retain the above copyright 879561Siedowse * notice, this list of conditions and the following disclaimer. 979561Siedowse * 2. Redistributions in binary form must reproduce the above copyright 1079561Siedowse * notice, this list of conditions and the following disclaimer in the 1179561Siedowse * documentation and/or other materials provided with the distribution. 1279561Siedowse * 1379561Siedowse * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 1479561Siedowse * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1579561Siedowse * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1679561Siedowse * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 1779561Siedowse * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 1879561Siedowse * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 1979561Siedowse * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2079561Siedowse * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2179561Siedowse * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2279561Siedowse * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2379561Siedowse * SUCH DAMAGE. 2479561Siedowse * 2579561Siedowse * $FreeBSD: head/sys/ufs/ufs/ufs_dirhash.c 84811 2001-10-11 17:52:20Z jhb $ 2679561Siedowse */ 2779561Siedowse/* 2879561Siedowse * This implements a hash-based lookup scheme for UFS directories. 2979561Siedowse */ 3079561Siedowse 3179561Siedowse#include "opt_ufs.h" 3279561Siedowse 3379561Siedowse#ifdef UFS_DIRHASH 3479561Siedowse 3579561Siedowse#include <sys/param.h> 3679561Siedowse#include <sys/systm.h> 3779561Siedowse#include <sys/kernel.h> 3884811Sjhb#include <sys/lock.h> 3979561Siedowse#include <sys/mutex.h> 4079561Siedowse#include <sys/malloc.h> 4179561Siedowse#include <sys/fnv_hash.h> 4279561Siedowse#include <sys/proc.h> 4379561Siedowse#include <sys/bio.h> 4479561Siedowse#include <sys/buf.h> 4579561Siedowse#include <sys/vnode.h> 4679561Siedowse#include <sys/mount.h> 4779561Siedowse#include <sys/sysctl.h> 4879561Siedowse#include <vm/vm_zone.h> 4979561Siedowse 5079561Siedowse#include <ufs/ufs/quota.h> 5179561Siedowse#include <ufs/ufs/inode.h> 5279561Siedowse#include <ufs/ufs/dir.h> 5379561Siedowse#include <ufs/ufs/dirhash.h> 5479561Siedowse#include <ufs/ufs/extattr.h> 5579561Siedowse#include <ufs/ufs/ufsmount.h> 5679561Siedowse#include <ufs/ufs/ufs_extern.h> 5779561Siedowse 5879561Siedowse#define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1)) 5979561Siedowse#define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 6079561Siedowse 6179561Siedowsestatic MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables"); 6279561Siedowse 6379561SiedowseSYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem"); 6479561Siedowse 6579561Siedowsestatic int ufs_mindirhashsize = DIRBLKSIZ * 5; 6679561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW, 6779561Siedowse &ufs_mindirhashsize, 6879561Siedowse 0, "minimum directory size in bytes for which to use hashed lookup"); 6979561Siedowsestatic int ufs_dirhashmaxmem = 2 * 1024 * 1024; 7079561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 7179561Siedowse 0, "maximum allowed dirhash memory usage"); 7279561Siedowsestatic int ufs_dirhashmem; 7379561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 7479561Siedowse 0, "current dirhash memory usage"); 7579561Siedowsestatic int ufs_dirhashcheck = 1; 7679561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 7779561Siedowse 0, "enable extra sanity tests"); 7879561Siedowse 7979561Siedowse 8079561Siedowsestatic int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 8179561Siedowsestatic void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 8279561Siedowsestatic void ufsdirhash_delslot(struct dirhash *dh, int slot); 8379561Siedowsestatic int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 8479561Siedowse doff_t offset); 8579561Siedowsestatic doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 8679561Siedowsestatic void ufsdirhash_init(void); 8779561Siedowsestatic int ufsdirhash_recycle(int wanted); 8879561Siedowse 8979561Siedowsestatic vm_zone_t ufsdirhash_zone; 9079561Siedowse 9179561Siedowse/* Dirhash list; recently-used entries are near the tail. */ 9279561Siedowsestatic TAILQ_HEAD(, dirhash) ufsdirhash_list; 9379561Siedowse 9479561Siedowse/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 9579561Siedowsestatic struct mtx ufsdirhash_mtx; 9679561Siedowse 9779561Siedowse/* 9879561Siedowse * Locking order: 9979561Siedowse * ufsdirhash_mtx 10079561Siedowse * dh_mtx 10179561Siedowse * 10279561Siedowse * The dh_mtx mutex should be aquired either via the inode lock, or via 10379561Siedowse * ufsdirhash_mtx. Only the owner of the inode may free the associated 10479561Siedowse * dirhash, but anything can steal its memory and set dh_hash to NULL. 10579561Siedowse */ 10679561Siedowse 10779561Siedowse/* 10879561Siedowse * Attempt to build up a hash table for the directory contents in 10979561Siedowse * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 11079561Siedowse */ 11179561Siedowseint 11279561Siedowseufsdirhash_build(struct inode *ip) 11379561Siedowse{ 11479561Siedowse struct dirhash *dh; 11579561Siedowse struct buf *bp = NULL; 11679561Siedowse struct direct *ep; 11779561Siedowse struct vnode *vp; 11879561Siedowse doff_t bmask, pos; 11979561Siedowse int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 12079561Siedowse 12179561Siedowse /* Check if we can/should use dirhash. */ 12279561Siedowse if (ip->i_dirhash == NULL) { 12379561Siedowse if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode)) 12479561Siedowse return (-1); 12579561Siedowse } else { 12679561Siedowse /* Hash exists, but sysctls could have changed. */ 12779561Siedowse if (ip->i_size < ufs_mindirhashsize || 12879561Siedowse ufs_dirhashmem > ufs_dirhashmaxmem) { 12979561Siedowse ufsdirhash_free(ip); 13079561Siedowse return (-1); 13179561Siedowse } 13279561Siedowse /* Check if hash exists and is intact (note: unlocked read). */ 13379561Siedowse if (ip->i_dirhash->dh_hash != NULL) 13479561Siedowse return (0); 13579561Siedowse /* Free the old, recycled hash and build a new one. */ 13679561Siedowse ufsdirhash_free(ip); 13779561Siedowse } 13879561Siedowse 13982364Siedowse /* Don't hash removed directories. */ 14082364Siedowse if (ip->i_effnlink == 0) 14182364Siedowse return (-1); 14282364Siedowse 14379561Siedowse vp = ip->i_vnode; 14479561Siedowse /* Allocate 50% more entries than this dir size could ever need. */ 14579561Siedowse KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 14679561Siedowse nslots = ip->i_size / DIRECTSIZ(1); 14779561Siedowse nslots = (nslots * 3 + 1) / 2; 14879561Siedowse narrays = howmany(nslots, DH_NBLKOFF); 14979561Siedowse nslots = narrays * DH_NBLKOFF; 15079561Siedowse dirblocks = howmany(ip->i_size, DIRBLKSIZ); 15179561Siedowse nblocks = (dirblocks * 3 + 1) / 2; 15279561Siedowse 15379561Siedowse memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 15479561Siedowse narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 15579561Siedowse nblocks * sizeof(*dh->dh_blkfree); 15679561Siedowse mtx_lock(&ufsdirhash_mtx); 15779561Siedowse if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 15879561Siedowse mtx_unlock(&ufsdirhash_mtx); 15979561Siedowse if (memreqd > ufs_dirhashmaxmem / 2) 16079561Siedowse return (-1); 16179561Siedowse 16279561Siedowse /* Try to free some space. */ 16379561Siedowse if (ufsdirhash_recycle(memreqd) != 0) 16479561Siedowse return (-1); 16579561Siedowse /* Enough was freed, and ufsdirhash_mtx has been locked. */ 16679561Siedowse } 16779561Siedowse ufs_dirhashmem += memreqd; 16879561Siedowse mtx_unlock(&ufsdirhash_mtx); 16979561Siedowse 17079561Siedowse /* 17179561Siedowse * Use non-blocking mallocs so that we will revert to a linear 17279561Siedowse * lookup on failure rather than potentially blocking forever. 17379561Siedowse */ 17479561Siedowse MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO); 17579561Siedowse if (dh == NULL) 17679561Siedowse return (-1); 17779561Siedowse MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]), 17879561Siedowse M_DIRHASH, M_NOWAIT | M_ZERO); 17979561Siedowse MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]), 18079561Siedowse M_DIRHASH, M_NOWAIT); 18179561Siedowse if (dh->dh_hash == NULL || dh->dh_blkfree == NULL) 18279561Siedowse goto fail; 18379561Siedowse for (i = 0; i < narrays; i++) { 18479561Siedowse if ((dh->dh_hash[i] = zalloc(ufsdirhash_zone)) == NULL) 18579561Siedowse goto fail; 18679561Siedowse for (j = 0; j < DH_NBLKOFF; j++) 18779561Siedowse dh->dh_hash[i][j] = DIRHASH_EMPTY; 18879561Siedowse } 18979561Siedowse 19079561Siedowse /* Initialise the hash table and block statistics. */ 19179561Siedowse mtx_init(&dh->dh_mtx, "dirhash", MTX_DEF); 19279561Siedowse dh->dh_narrays = narrays; 19379561Siedowse dh->dh_hlen = nslots; 19479561Siedowse dh->dh_nblk = nblocks; 19579561Siedowse dh->dh_dirblks = dirblocks; 19679561Siedowse for (i = 0; i < dirblocks; i++) 19779561Siedowse dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 19879561Siedowse for (i = 0; i < DH_NFSTATS; i++) 19979561Siedowse dh->dh_firstfree[i] = -1; 20079561Siedowse dh->dh_firstfree[DH_NFSTATS] = 0; 20179561Siedowse dh->dh_seqopt = 0; 20279561Siedowse dh->dh_seqoff = 0; 20379561Siedowse dh->dh_score = DH_SCOREINIT; 20479561Siedowse ip->i_dirhash = dh; 20579561Siedowse 20679561Siedowse bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 20779561Siedowse pos = 0; 20879561Siedowse while (pos < ip->i_size) { 20979561Siedowse /* If necessary, get the next directory block. */ 21079561Siedowse if ((pos & bmask) == 0) { 21179561Siedowse if (bp != NULL) 21279561Siedowse brelse(bp); 21379561Siedowse if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 21479561Siedowse goto fail; 21579561Siedowse } 21679561Siedowse 21779561Siedowse /* Add this entry to the hash. */ 21879561Siedowse ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 21979561Siedowse if (ep->d_reclen == 0 || ep->d_reclen > 22079561Siedowse DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 22179561Siedowse /* Corrupted directory. */ 22279561Siedowse brelse(bp); 22379561Siedowse goto fail; 22479561Siedowse } 22579561Siedowse if (ep->d_ino != 0) { 22679561Siedowse /* Add the entry (simplified ufsdirhash_add). */ 22779561Siedowse slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 22879561Siedowse while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 22979561Siedowse slot = WRAPINCR(slot, dh->dh_hlen); 23079561Siedowse dh->dh_hused++; 23179561Siedowse DH_ENTRY(dh, slot) = pos; 23279561Siedowse ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 23379561Siedowse } 23479561Siedowse pos += ep->d_reclen; 23579561Siedowse } 23679561Siedowse 23779561Siedowse if (bp != NULL) 23879561Siedowse brelse(bp); 23979561Siedowse mtx_lock(&ufsdirhash_mtx); 24079561Siedowse TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 24179561Siedowse dh->dh_onlist = 1; 24279561Siedowse mtx_unlock(&ufsdirhash_mtx); 24379561Siedowse return (0); 24479561Siedowse 24579561Siedowsefail: 24679561Siedowse if (dh->dh_hash != NULL) { 24779561Siedowse for (i = 0; i < narrays; i++) 24879561Siedowse if (dh->dh_hash[i] != NULL) 24979561Siedowse zfree(ufsdirhash_zone, dh->dh_hash[i]); 25079561Siedowse FREE(dh->dh_hash, M_DIRHASH); 25179561Siedowse } 25279561Siedowse if (dh->dh_blkfree != NULL) 25379561Siedowse FREE(dh->dh_blkfree, M_DIRHASH); 25479561Siedowse FREE(dh, M_DIRHASH); 25579561Siedowse ip->i_dirhash = NULL; 25679561Siedowse mtx_lock(&ufsdirhash_mtx); 25779561Siedowse ufs_dirhashmem -= memreqd; 25879561Siedowse mtx_unlock(&ufsdirhash_mtx); 25979561Siedowse return (-1); 26079561Siedowse} 26179561Siedowse 26279561Siedowse/* 26379561Siedowse * Free any hash table associated with inode 'ip'. 26479561Siedowse */ 26579561Siedowsevoid 26679561Siedowseufsdirhash_free(struct inode *ip) 26779561Siedowse{ 26879561Siedowse struct dirhash *dh; 26979561Siedowse int i, mem; 27079561Siedowse 27179561Siedowse if ((dh = ip->i_dirhash) == NULL) 27279561Siedowse return; 27379561Siedowse mtx_lock(&ufsdirhash_mtx); 27479561Siedowse mtx_lock(&dh->dh_mtx); 27579561Siedowse if (dh->dh_onlist) 27679561Siedowse TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 27779561Siedowse mtx_unlock(&dh->dh_mtx); 27879561Siedowse mtx_unlock(&ufsdirhash_mtx); 27979561Siedowse 28079561Siedowse /* The dirhash pointed to by 'dh' is exclusively ours now. */ 28179561Siedowse 28279561Siedowse mem = sizeof(*dh); 28379561Siedowse if (dh->dh_hash != NULL) { 28479561Siedowse for (i = 0; i < dh->dh_narrays; i++) 28579561Siedowse zfree(ufsdirhash_zone, dh->dh_hash[i]); 28679561Siedowse FREE(dh->dh_hash, M_DIRHASH); 28779561Siedowse FREE(dh->dh_blkfree, M_DIRHASH); 28879561Siedowse mem += dh->dh_narrays * sizeof(*dh->dh_hash) + 28979561Siedowse dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 29079561Siedowse dh->dh_nblk * sizeof(*dh->dh_blkfree); 29179561Siedowse } 29279561Siedowse mtx_destroy(&dh->dh_mtx); 29379561Siedowse FREE(dh, M_DIRHASH); 29479561Siedowse ip->i_dirhash = NULL; 29579561Siedowse 29679561Siedowse mtx_lock(&ufsdirhash_mtx); 29779561Siedowse ufs_dirhashmem -= mem; 29879561Siedowse mtx_unlock(&ufsdirhash_mtx); 29979561Siedowse} 30079561Siedowse 30179561Siedowse/* 30279561Siedowse * Find the offset of the specified name within the given inode. 30379561Siedowse * Returns 0 on success, ENOENT if the entry does not exist, or 30479561Siedowse * EJUSTRETURN if the caller should revert to a linear search. 30579561Siedowse * 30679690Siedowse * If successful, the directory offset is stored in *offp, and a 30779690Siedowse * pointer to a struct buf containing the entry is stored in *bpp. If 30879561Siedowse * prevoffp is non-NULL, the offset of the previous entry within 30979561Siedowse * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 31079561Siedowse * is the first in a block, the start of the block is used). 31179561Siedowse */ 31279561Siedowseint 31379561Siedowseufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 31479690Siedowse struct buf **bpp, doff_t *prevoffp) 31579561Siedowse{ 31679561Siedowse struct dirhash *dh, *dh_next; 31779561Siedowse struct direct *dp; 31879561Siedowse struct vnode *vp; 31979561Siedowse struct buf *bp; 32079561Siedowse doff_t blkoff, bmask, offset, prevoff; 32179561Siedowse int i, slot; 32279561Siedowse 32379561Siedowse if ((dh = ip->i_dirhash) == NULL) 32479561Siedowse return (EJUSTRETURN); 32579561Siedowse /* 32679561Siedowse * Move this dirhash towards the end of the list if it has a 32779561Siedowse * score higher than the next entry, and aquire the dh_mtx. 32879561Siedowse * Optimise the case where it's already the last by performing 32979561Siedowse * an unlocked read of the TAILQ_NEXT pointer. 33079561Siedowse * 33179561Siedowse * In both cases, end up holding just dh_mtx. 33279561Siedowse */ 33379561Siedowse if (TAILQ_NEXT(dh, dh_list) != NULL) { 33479561Siedowse mtx_lock(&ufsdirhash_mtx); 33579561Siedowse mtx_lock(&dh->dh_mtx); 33679561Siedowse /* 33779561Siedowse * If the new score will be greater than that of the next 33879561Siedowse * entry, then move this entry past it. With both mutexes 33979561Siedowse * held, dh_next won't go away, but its dh_score could 34079561Siedowse * change; that's not important since it is just a hint. 34179561Siedowse */ 34279561Siedowse if (dh->dh_hash != NULL && 34379561Siedowse (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 34479561Siedowse dh->dh_score >= dh_next->dh_score) { 34579561Siedowse KASSERT(dh->dh_onlist, ("dirhash: not on list")); 34679561Siedowse TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 34779561Siedowse TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 34879561Siedowse dh_list); 34979561Siedowse } 35079561Siedowse mtx_unlock(&ufsdirhash_mtx); 35179561Siedowse } else { 35279561Siedowse /* Already the last, though that could change as we wait. */ 35379561Siedowse mtx_lock(&dh->dh_mtx); 35479561Siedowse } 35579561Siedowse if (dh->dh_hash == NULL) { 35679561Siedowse mtx_unlock(&dh->dh_mtx); 35779561Siedowse ufsdirhash_free(ip); 35879561Siedowse return (EJUSTRETURN); 35979561Siedowse } 36079561Siedowse 36179561Siedowse /* Update the score. */ 36279561Siedowse if (dh->dh_score < DH_SCOREMAX) 36379561Siedowse dh->dh_score++; 36479561Siedowse 36579561Siedowse vp = ip->i_vnode; 36679561Siedowse bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 36779561Siedowse blkoff = -1; 36879561Siedowse bp = NULL; 36979561Siedowserestart: 37079561Siedowse slot = ufsdirhash_hash(dh, name, namelen); 37179561Siedowse 37279561Siedowse if (dh->dh_seqopt) { 37379561Siedowse /* 37479561Siedowse * Sequential access optimisation. dh_seqoff contains the 37579561Siedowse * offset of the directory entry immediately following 37679561Siedowse * the last entry that was looked up. Check if this offset 37779561Siedowse * appears in the hash chain for the name we are looking for. 37879561Siedowse */ 37979561Siedowse for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 38079561Siedowse i = WRAPINCR(i, dh->dh_hlen)) 38179561Siedowse if (offset == dh->dh_seqopt) 38279561Siedowse break; 38379561Siedowse if (offset == dh->dh_seqoff) { 38479561Siedowse /* 38579561Siedowse * We found an entry with the expected offset. This 38679561Siedowse * is probably the entry we want, but if not, the 38779561Siedowse * code below will turn off seqoff and retry. 38879561Siedowse */ 38979561Siedowse slot = i; 39079561Siedowse } else 39179561Siedowse dh->dh_seqopt = 0; 39279561Siedowse } 39379561Siedowse 39479561Siedowse for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 39579561Siedowse slot = WRAPINCR(slot, dh->dh_hlen)) { 39679561Siedowse if (offset == DIRHASH_DEL) 39779561Siedowse continue; 39879561Siedowse mtx_unlock(&dh->dh_mtx); 39979561Siedowse 40079561Siedowse if (offset < 0 || offset >= ip->i_size) 40179561Siedowse panic("ufsdirhash_lookup: bad offset in hash array"); 40279561Siedowse if ((offset & ~bmask) != blkoff) { 40379561Siedowse if (bp != NULL) 40479561Siedowse brelse(bp); 40579561Siedowse blkoff = offset & ~bmask; 40679561Siedowse if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) 40779561Siedowse return (EJUSTRETURN); 40879561Siedowse } 40979561Siedowse dp = (struct direct *)(bp->b_data + (offset & bmask)); 41079561Siedowse if (dp->d_reclen == 0 || dp->d_reclen > 41179561Siedowse DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 41279561Siedowse /* Corrupted directory. */ 41379561Siedowse brelse(bp); 41479561Siedowse return (EJUSTRETURN); 41579561Siedowse } 41679561Siedowse if (dp->d_namlen == namelen && 41779561Siedowse bcmp(dp->d_name, name, namelen) == 0) { 41879561Siedowse /* Found. Get the prev offset if needed. */ 41979561Siedowse if (prevoffp != NULL) { 42079561Siedowse if (offset & (DIRBLKSIZ - 1)) { 42179561Siedowse prevoff = ufsdirhash_getprev(dp, 42279561Siedowse offset); 42379561Siedowse if (prevoff == -1) { 42479561Siedowse brelse(bp); 42579561Siedowse return (EJUSTRETURN); 42679561Siedowse } 42779561Siedowse } else 42879561Siedowse prevoff = offset; 42979561Siedowse *prevoffp = prevoff; 43079561Siedowse } 43179561Siedowse 43279561Siedowse /* Check for sequential access, and update offset. */ 43379561Siedowse if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 43479561Siedowse dh->dh_seqopt = 1; 43579561Siedowse dh->dh_seqoff = offset + DIRSIZ(0, dp); 43679561Siedowse 43779690Siedowse *bpp = bp; 43879561Siedowse *offp = offset; 43979561Siedowse return (0); 44079561Siedowse } 44179561Siedowse 44279561Siedowse mtx_lock(&dh->dh_mtx); 44379561Siedowse if (dh->dh_hash == NULL) { 44479561Siedowse mtx_unlock(&dh->dh_mtx); 44579561Siedowse if (bp != NULL) 44679561Siedowse brelse(bp); 44779561Siedowse ufsdirhash_free(ip); 44879561Siedowse return (EJUSTRETURN); 44979561Siedowse } 45079561Siedowse /* 45179561Siedowse * When the name doesn't match in the seqopt case, go back 45279561Siedowse * and search normally. 45379561Siedowse */ 45479561Siedowse if (dh->dh_seqopt) { 45579561Siedowse dh->dh_seqopt = 0; 45679561Siedowse goto restart; 45779561Siedowse } 45879561Siedowse } 45979561Siedowse mtx_unlock(&dh->dh_mtx); 46079561Siedowse if (bp != NULL) 46179561Siedowse brelse(bp); 46279561Siedowse return (ENOENT); 46379561Siedowse} 46479561Siedowse 46579561Siedowse/* 46679561Siedowse * Find a directory block with room for 'slotneeded' bytes. Returns 46779561Siedowse * the offset of the directory entry that begins the free space. 46879561Siedowse * This will either be the offset of an existing entry that has free 46979561Siedowse * space at the end, or the offset of an entry with d_ino == 0 at 47079561Siedowse * the start of a DIRBLKSIZ block. 47179561Siedowse * 47279561Siedowse * To use the space, the caller may need to compact existing entries in 47379561Siedowse * the directory. The total number of bytes in all of the entries involved 47479561Siedowse * in the compaction is stored in *slotsize. In other words, all of 47579561Siedowse * the entries that must be compacted are exactly contained in the 47679561Siedowse * region beginning at the returned offset and spanning *slotsize bytes. 47779561Siedowse * 47879561Siedowse * Returns -1 if no space was found, indicating that the directory 47979561Siedowse * must be extended. 48079561Siedowse */ 48179561Siedowsedoff_t 48279561Siedowseufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 48379561Siedowse{ 48479561Siedowse struct direct *dp; 48579561Siedowse struct dirhash *dh; 48679561Siedowse struct buf *bp; 48779561Siedowse doff_t pos, slotstart; 48879561Siedowse int dirblock, error, freebytes, i; 48979561Siedowse 49079561Siedowse if ((dh = ip->i_dirhash) == NULL) 49179561Siedowse return (-1); 49279561Siedowse mtx_lock(&dh->dh_mtx); 49379561Siedowse if (dh->dh_hash == NULL) { 49479561Siedowse mtx_unlock(&dh->dh_mtx); 49579561Siedowse ufsdirhash_free(ip); 49679561Siedowse return (-1); 49779561Siedowse } 49879561Siedowse 49979561Siedowse /* Find a directory block with the desired free space. */ 50079561Siedowse dirblock = -1; 50179561Siedowse for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 50279561Siedowse if ((dirblock = dh->dh_firstfree[i]) != -1) 50379561Siedowse break; 50479561Siedowse if (dirblock == -1) { 50579561Siedowse mtx_unlock(&dh->dh_mtx); 50679561Siedowse return (-1); 50779561Siedowse } 50879561Siedowse 50979561Siedowse KASSERT(dirblock < dh->dh_nblk && 51079561Siedowse dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 51179561Siedowse ("ufsdirhash_findfree: bad stats")); 51279561Siedowse mtx_unlock(&dh->dh_mtx); 51379561Siedowse pos = dirblock * DIRBLKSIZ; 51479561Siedowse error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 51579561Siedowse if (error) 51679561Siedowse return (-1); 51779561Siedowse 51879561Siedowse /* Find the first entry with free space. */ 51979561Siedowse for (i = 0; i < DIRBLKSIZ; ) { 52079561Siedowse if (dp->d_reclen == 0) { 52179561Siedowse brelse(bp); 52279561Siedowse return (-1); 52379561Siedowse } 52479561Siedowse if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 52579561Siedowse break; 52679561Siedowse i += dp->d_reclen; 52779561Siedowse dp = (struct direct *)((char *)dp + dp->d_reclen); 52879561Siedowse } 52979561Siedowse if (i > DIRBLKSIZ) { 53079561Siedowse brelse(bp); 53179561Siedowse return (-1); 53279561Siedowse } 53379561Siedowse slotstart = pos + i; 53479561Siedowse 53579561Siedowse /* Find the range of entries needed to get enough space */ 53679561Siedowse freebytes = 0; 53779561Siedowse while (i < DIRBLKSIZ && freebytes < slotneeded) { 53879561Siedowse freebytes += dp->d_reclen; 53979561Siedowse if (dp->d_ino != 0) 54079561Siedowse freebytes -= DIRSIZ(0, dp); 54179561Siedowse if (dp->d_reclen == 0) { 54279561Siedowse brelse(bp); 54379561Siedowse return (-1); 54479561Siedowse } 54579561Siedowse i += dp->d_reclen; 54679561Siedowse dp = (struct direct *)((char *)dp + dp->d_reclen); 54779561Siedowse } 54879561Siedowse if (i > DIRBLKSIZ) { 54979561Siedowse brelse(bp); 55079561Siedowse return (-1); 55179561Siedowse } 55279561Siedowse if (freebytes < slotneeded) 55379561Siedowse panic("ufsdirhash_findfree: free mismatch"); 55479561Siedowse brelse(bp); 55579561Siedowse *slotsize = pos + i - slotstart; 55679561Siedowse return (slotstart); 55779561Siedowse} 55879561Siedowse 55979561Siedowse/* 56079561Siedowse * Return the start of the unused space at the end of a directory, or 56179561Siedowse * -1 if there are no trailing unused blocks. 56279561Siedowse */ 56379561Siedowsedoff_t 56479561Siedowseufsdirhash_enduseful(struct inode *ip) 56579561Siedowse{ 56679561Siedowse 56779561Siedowse struct dirhash *dh; 56879561Siedowse int i; 56979561Siedowse 57079561Siedowse if ((dh = ip->i_dirhash) == NULL) 57179561Siedowse return (-1); 57279561Siedowse mtx_lock(&dh->dh_mtx); 57379561Siedowse if (dh->dh_hash == NULL) { 57479561Siedowse mtx_unlock(&dh->dh_mtx); 57579561Siedowse ufsdirhash_free(ip); 57679561Siedowse return (-1); 57779561Siedowse } 57879561Siedowse 57979561Siedowse if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) { 58079561Siedowse mtx_unlock(&dh->dh_mtx); 58179561Siedowse return (-1); 58279561Siedowse } 58379561Siedowse 58479561Siedowse for (i = dh->dh_dirblks - 1; i >= 0; i--) 58579561Siedowse if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 58679561Siedowse break; 58779561Siedowse mtx_unlock(&dh->dh_mtx); 58879561Siedowse return ((doff_t)(i + 1) * DIRBLKSIZ); 58979561Siedowse} 59079561Siedowse 59179561Siedowse/* 59279561Siedowse * Insert information into the hash about a new directory entry. dirp 59379561Siedowse * points to a struct direct containing the entry, and offset specifies 59479561Siedowse * the offset of this entry. 59579561Siedowse */ 59679561Siedowsevoid 59779561Siedowseufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 59879561Siedowse{ 59979561Siedowse struct dirhash *dh; 60079561Siedowse int slot; 60179561Siedowse 60279561Siedowse if ((dh = ip->i_dirhash) == NULL) 60379561Siedowse return; 60479561Siedowse mtx_lock(&dh->dh_mtx); 60579561Siedowse if (dh->dh_hash == NULL) { 60679561Siedowse mtx_unlock(&dh->dh_mtx); 60779561Siedowse ufsdirhash_free(ip); 60879561Siedowse return; 60979561Siedowse } 61079561Siedowse 61179561Siedowse KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 61279561Siedowse ("ufsdirhash_add: bad offset")); 61379561Siedowse /* 61479561Siedowse * Normal hash usage is < 66%. If the usage gets too high then 61579561Siedowse * remove the hash entirely and let it be rebuilt later. 61679561Siedowse */ 61779561Siedowse if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 61879561Siedowse mtx_unlock(&dh->dh_mtx); 61979561Siedowse ufsdirhash_free(ip); 62079561Siedowse return; 62179561Siedowse } 62279561Siedowse 62379561Siedowse /* Find a free hash slot (empty or deleted), and add the entry. */ 62479561Siedowse slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 62579561Siedowse while (DH_ENTRY(dh, slot) >= 0) 62679561Siedowse slot = WRAPINCR(slot, dh->dh_hlen); 62779561Siedowse if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 62879561Siedowse dh->dh_hused++; 62979561Siedowse DH_ENTRY(dh, slot) = offset; 63079561Siedowse 63179561Siedowse /* Update the per-block summary info. */ 63279561Siedowse ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 63379561Siedowse mtx_unlock(&dh->dh_mtx); 63479561Siedowse} 63579561Siedowse 63679561Siedowse/* 63779561Siedowse * Remove the specified directory entry from the hash. The entry to remove 63879561Siedowse * is defined by the name in `dirp', which must exist at the specified 63979561Siedowse * `offset' within the directory. 64079561Siedowse */ 64179561Siedowsevoid 64279561Siedowseufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 64379561Siedowse{ 64479561Siedowse struct dirhash *dh; 64579561Siedowse int slot; 64679561Siedowse 64779561Siedowse if ((dh = ip->i_dirhash) == NULL) 64879561Siedowse return; 64979561Siedowse mtx_lock(&dh->dh_mtx); 65079561Siedowse if (dh->dh_hash == NULL) { 65179561Siedowse mtx_unlock(&dh->dh_mtx); 65279561Siedowse ufsdirhash_free(ip); 65379561Siedowse return; 65479561Siedowse } 65579561Siedowse 65679561Siedowse KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 65779561Siedowse ("ufsdirhash_remove: bad offset")); 65879561Siedowse /* Find the entry */ 65979561Siedowse slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 66079561Siedowse 66179561Siedowse /* Remove the hash entry. */ 66279561Siedowse ufsdirhash_delslot(dh, slot); 66379561Siedowse 66479561Siedowse /* Update the per-block summary info. */ 66579561Siedowse ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 66679561Siedowse mtx_unlock(&dh->dh_mtx); 66779561Siedowse} 66879561Siedowse 66979561Siedowse/* 67079561Siedowse * Change the offset associated with a directory entry in the hash. Used 67179561Siedowse * when compacting directory blocks. 67279561Siedowse */ 67379561Siedowsevoid 67479561Siedowseufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 67579561Siedowse doff_t newoff) 67679561Siedowse{ 67779561Siedowse struct dirhash *dh; 67879561Siedowse int slot; 67979561Siedowse 68079561Siedowse if ((dh = ip->i_dirhash) == NULL) 68179561Siedowse return; 68279561Siedowse mtx_lock(&dh->dh_mtx); 68379561Siedowse if (dh->dh_hash == NULL) { 68479561Siedowse mtx_unlock(&dh->dh_mtx); 68579561Siedowse ufsdirhash_free(ip); 68679561Siedowse return; 68779561Siedowse } 68879561Siedowse 68979561Siedowse KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 69079561Siedowse newoff < dh->dh_dirblks * DIRBLKSIZ, 69179561Siedowse ("ufsdirhash_move: bad offset")); 69279561Siedowse /* Find the entry, and update the offset. */ 69379561Siedowse slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 69479561Siedowse DH_ENTRY(dh, slot) = newoff; 69579561Siedowse mtx_unlock(&dh->dh_mtx); 69679561Siedowse} 69779561Siedowse 69879561Siedowse/* 69979561Siedowse * Inform dirhash that the directory has grown by one block that 70079561Siedowse * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 70179561Siedowse */ 70279561Siedowsevoid 70379561Siedowseufsdirhash_newblk(struct inode *ip, doff_t offset) 70479561Siedowse{ 70579561Siedowse struct dirhash *dh; 70679561Siedowse int block; 70779561Siedowse 70879561Siedowse if ((dh = ip->i_dirhash) == NULL) 70979561Siedowse return; 71079561Siedowse mtx_lock(&dh->dh_mtx); 71179561Siedowse if (dh->dh_hash == NULL) { 71279561Siedowse mtx_unlock(&dh->dh_mtx); 71379561Siedowse ufsdirhash_free(ip); 71479561Siedowse return; 71579561Siedowse } 71679561Siedowse 71779561Siedowse KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 71879561Siedowse ("ufsdirhash_newblk: bad offset")); 71979561Siedowse block = offset / DIRBLKSIZ; 72079561Siedowse if (block >= dh->dh_nblk) { 72179561Siedowse /* Out of space; must rebuild. */ 72279561Siedowse mtx_unlock(&dh->dh_mtx); 72379561Siedowse ufsdirhash_free(ip); 72479561Siedowse return; 72579561Siedowse } 72679561Siedowse dh->dh_dirblks = block + 1; 72779561Siedowse 72879561Siedowse /* Account for the new free block. */ 72979561Siedowse dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 73079561Siedowse if (dh->dh_firstfree[DH_NFSTATS] == -1) 73179561Siedowse dh->dh_firstfree[DH_NFSTATS] = block; 73279561Siedowse mtx_unlock(&dh->dh_mtx); 73379561Siedowse} 73479561Siedowse 73579561Siedowse/* 73679561Siedowse * Inform dirhash that the directory is being truncated. 73779561Siedowse */ 73879561Siedowsevoid 73979561Siedowseufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 74079561Siedowse{ 74179561Siedowse struct dirhash *dh; 74279561Siedowse int block, i; 74379561Siedowse 74479561Siedowse if ((dh = ip->i_dirhash) == NULL) 74579561Siedowse return; 74679561Siedowse mtx_lock(&dh->dh_mtx); 74779561Siedowse if (dh->dh_hash == NULL) { 74879561Siedowse mtx_unlock(&dh->dh_mtx); 74979561Siedowse ufsdirhash_free(ip); 75079561Siedowse return; 75179561Siedowse } 75279561Siedowse 75379561Siedowse KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 75479561Siedowse ("ufsdirhash_dirtrunc: bad offset")); 75579561Siedowse block = howmany(offset, DIRBLKSIZ); 75679561Siedowse /* 75779561Siedowse * If the directory shrinks to less than 1/8 of dh_nblk blocks 75879561Siedowse * (about 20% of its original size due to the 50% extra added in 75979561Siedowse * ufsdirhash_build) then free it, and let the caller rebuild 76079561Siedowse * if necessary. 76179561Siedowse */ 76279561Siedowse if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 76379561Siedowse mtx_unlock(&dh->dh_mtx); 76479561Siedowse ufsdirhash_free(ip); 76579561Siedowse return; 76679561Siedowse } 76779561Siedowse 76879561Siedowse /* 76979561Siedowse * Remove any `first free' information pertaining to the 77079561Siedowse * truncated blocks. All blocks we're removing should be 77179561Siedowse * completely unused. 77279561Siedowse */ 77379561Siedowse if (dh->dh_firstfree[DH_NFSTATS] >= block) 77479561Siedowse dh->dh_firstfree[DH_NFSTATS] = -1; 77579561Siedowse for (i = block; i < dh->dh_dirblks; i++) 77679561Siedowse if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 77779561Siedowse panic("ufsdirhash_dirtrunc: blocks in use"); 77879561Siedowse for (i = 0; i < DH_NFSTATS; i++) 77979561Siedowse if (dh->dh_firstfree[i] >= block) 78079561Siedowse panic("ufsdirhash_dirtrunc: first free corrupt"); 78179561Siedowse dh->dh_dirblks = block; 78279561Siedowse mtx_unlock(&dh->dh_mtx); 78379561Siedowse} 78479561Siedowse 78579561Siedowse/* 78679561Siedowse * Debugging function to check that the dirhash information about 78779561Siedowse * a directory block matches its actual contents. Panics if a mismatch 78879561Siedowse * is detected. 78979561Siedowse * 79079561Siedowse * On entry, `buf' should point to the start of an in-core 79179561Siedowse * DIRBLKSIZ-sized directory block, and `offset' should contain the 79279561Siedowse * offset from the start of the directory of that block. 79379561Siedowse */ 79479561Siedowsevoid 79579561Siedowseufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 79679561Siedowse{ 79779561Siedowse struct dirhash *dh; 79879561Siedowse struct direct *dp; 79979561Siedowse int block, ffslot, i, nfree; 80079561Siedowse 80179561Siedowse if (!ufs_dirhashcheck) 80279561Siedowse return; 80379561Siedowse if ((dh = ip->i_dirhash) == NULL) 80479561Siedowse return; 80579561Siedowse mtx_lock(&dh->dh_mtx); 80679561Siedowse if (dh->dh_hash == NULL) { 80779561Siedowse mtx_unlock(&dh->dh_mtx); 80879561Siedowse ufsdirhash_free(ip); 80979561Siedowse return; 81079561Siedowse } 81179561Siedowse 81279561Siedowse block = offset / DIRBLKSIZ; 81379561Siedowse if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 81479561Siedowse panic("ufsdirhash_checkblock: bad offset"); 81579561Siedowse 81679561Siedowse nfree = 0; 81779561Siedowse for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 81879561Siedowse dp = (struct direct *)(buf + i); 81979561Siedowse if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 82079561Siedowse panic("ufsdirhash_checkblock: bad dir"); 82179561Siedowse 82279561Siedowse if (dp->d_ino == 0) { 82380456Siedowse#if 0 82480456Siedowse /* 82580456Siedowse * XXX entries with d_ino == 0 should only occur 82680456Siedowse * at the start of a DIRBLKSIZ block. However the 82780456Siedowse * ufs code is tolerant of such entries at other 82880456Siedowse * offsets, and fsck does not fix them. 82980456Siedowse */ 83079561Siedowse if (i != 0) 83179561Siedowse panic("ufsdirhash_checkblock: bad dir inode"); 83280456Siedowse#endif 83379561Siedowse nfree += dp->d_reclen; 83479561Siedowse continue; 83579561Siedowse } 83679561Siedowse 83779561Siedowse /* Check that the entry exists (will panic if it doesn't). */ 83879561Siedowse ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 83979561Siedowse 84079561Siedowse nfree += dp->d_reclen - DIRSIZ(0, dp); 84179561Siedowse } 84279561Siedowse if (i != DIRBLKSIZ) 84379561Siedowse panic("ufsdirhash_checkblock: bad dir end"); 84479561Siedowse 84579561Siedowse if (dh->dh_blkfree[block] * DIRALIGN != nfree) 84679561Siedowse panic("ufsdirhash_checkblock: bad free count"); 84779561Siedowse 84879561Siedowse ffslot = nfree / DIRALIGN; 84979561Siedowse if (ffslot > DH_NFSTATS) 85079561Siedowse ffslot = DH_NFSTATS; 85179561Siedowse for (i = 0; i <= DH_NFSTATS; i++) 85279561Siedowse if (dh->dh_firstfree[i] == block && i != ffslot) 85379561Siedowse panic("ufsdirhash_checkblock: bad first-free"); 85479561Siedowse mtx_unlock(&dh->dh_mtx); 85579561Siedowse} 85679561Siedowse 85779561Siedowse/* 85879561Siedowse * Hash the specified filename into a dirhash slot. 85979561Siedowse */ 86079561Siedowsestatic int 86179561Siedowseufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 86279561Siedowse{ 86379561Siedowse return (fnv_32_buf(name, namelen, FNV1_32_INIT) % dh->dh_hlen); 86479561Siedowse} 86579561Siedowse 86679561Siedowse/* 86779561Siedowse * Adjust the number of free bytes in the block containing `offset' 86879561Siedowse * by the value specified by `diff'. 86979561Siedowse * 87079561Siedowse * The caller must ensure we have exclusive access to `dh'; normally 87179561Siedowse * that means that dh_mtx should be held, but this is also called 87279561Siedowse * from ufsdirhash_build() where exclusive access can be assumed. 87379561Siedowse */ 87479561Siedowsestatic void 87579561Siedowseufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 87679561Siedowse{ 87779561Siedowse int block, i, nfidx, ofidx; 87879561Siedowse 87979561Siedowse /* Update the per-block summary info. */ 88079561Siedowse block = offset / DIRBLKSIZ; 88179561Siedowse KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 88279561Siedowse ("dirhash bad offset")); 88379561Siedowse ofidx = dh->dh_blkfree[block]; 88479561Siedowse if (ofidx > DH_NFSTATS) 88579561Siedowse ofidx = DH_NFSTATS; 88679561Siedowse dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 88779561Siedowse nfidx = dh->dh_blkfree[block]; 88879561Siedowse if (nfidx > DH_NFSTATS) 88979561Siedowse nfidx = DH_NFSTATS; 89079561Siedowse 89179561Siedowse /* Update the `first free' list if necessary. */ 89279561Siedowse if (ofidx != nfidx) { 89379561Siedowse /* If removing, scan forward for the next block. */ 89479561Siedowse if (dh->dh_firstfree[ofidx] == block) { 89579561Siedowse for (i = block + 1; i < dh->dh_dirblks; i++) 89679561Siedowse if (dh->dh_blkfree[i] == ofidx) 89779561Siedowse break; 89879561Siedowse dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 89979561Siedowse } 90079561Siedowse 90179561Siedowse /* Make this the new `first free' if necessary */ 90279561Siedowse if (dh->dh_firstfree[nfidx] > block || 90379561Siedowse dh->dh_firstfree[nfidx] == -1) 90479561Siedowse dh->dh_firstfree[nfidx] = block; 90579561Siedowse } 90679561Siedowse} 90779561Siedowse 90879561Siedowse/* 90979561Siedowse * Find the specified name which should have the specified offset. 91079561Siedowse * Returns a slot number, and panics on failure. 91179561Siedowse * 91279561Siedowse * `dh' must be locked on entry and remains so on return. 91379561Siedowse */ 91479561Siedowsestatic int 91579561Siedowseufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 91679561Siedowse{ 91779561Siedowse int slot; 91879561Siedowse 91979561Siedowse mtx_assert(&dh->dh_mtx, MA_OWNED); 92079561Siedowse 92179561Siedowse /* Find the entry. */ 92279561Siedowse KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 92379561Siedowse slot = ufsdirhash_hash(dh, name, namelen); 92479561Siedowse while (DH_ENTRY(dh, slot) != offset && 92579561Siedowse DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 92679561Siedowse slot = WRAPINCR(slot, dh->dh_hlen); 92779561Siedowse if (DH_ENTRY(dh, slot) != offset) 92879561Siedowse panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 92979561Siedowse 93079561Siedowse return (slot); 93179561Siedowse} 93279561Siedowse 93379561Siedowse/* 93479561Siedowse * Remove the entry corresponding to the specified slot from the hash array. 93579561Siedowse * 93679561Siedowse * `dh' must be locked on entry and remains so on return. 93779561Siedowse */ 93879561Siedowsestatic void 93979561Siedowseufsdirhash_delslot(struct dirhash *dh, int slot) 94079561Siedowse{ 94179561Siedowse int i; 94279561Siedowse 94379561Siedowse mtx_assert(&dh->dh_mtx, MA_OWNED); 94479561Siedowse 94579561Siedowse /* Mark the entry as deleted. */ 94679561Siedowse DH_ENTRY(dh, slot) = DIRHASH_DEL; 94779561Siedowse 94879561Siedowse /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 94979561Siedowse for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 95079561Siedowse i = WRAPINCR(i, dh->dh_hlen); 95179561Siedowse if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 95279561Siedowse for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) { 95379561Siedowse DH_ENTRY(dh, i) = DIRHASH_EMPTY; 95479561Siedowse dh->dh_hused--; 95579561Siedowse i = WRAPINCR(i, dh->dh_hlen); 95679561Siedowse } 95779561Siedowse KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 95879561Siedowse } 95979561Siedowse} 96079561Siedowse 96179561Siedowse/* 96279561Siedowse * Given a directory entry and its offset, find the offset of the 96379561Siedowse * previous entry in the same DIRBLKSIZ-sized block. Returns an 96479561Siedowse * offset, or -1 if there is no previous entry in the block or some 96579561Siedowse * other problem occurred. 96679561Siedowse */ 96779561Siedowsestatic doff_t 96879561Siedowseufsdirhash_getprev(struct direct *dirp, doff_t offset) 96979561Siedowse{ 97079561Siedowse struct direct *dp; 97179561Siedowse char *blkbuf; 97279561Siedowse doff_t blkoff, prevoff; 97379561Siedowse int entrypos, i; 97479561Siedowse 97579561Siedowse blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 97679561Siedowse entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 97779561Siedowse blkbuf = (char *)dirp - entrypos; 97879561Siedowse prevoff = blkoff; 97979561Siedowse 98079561Siedowse /* If `offset' is the start of a block, there is no previous entry. */ 98179561Siedowse if (entrypos == 0) 98279561Siedowse return (-1); 98379561Siedowse 98479561Siedowse /* Scan from the start of the block until we get to the entry. */ 98579561Siedowse for (i = 0; i < entrypos; i += dp->d_reclen) { 98679561Siedowse dp = (struct direct *)(blkbuf + i); 98779561Siedowse if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 98879561Siedowse return (-1); /* Corrupted directory. */ 98979561Siedowse prevoff = blkoff + i; 99079561Siedowse } 99179561Siedowse return (prevoff); 99279561Siedowse} 99379561Siedowse 99479561Siedowse/* 99579561Siedowse * Try to free up `wanted' bytes by stealing memory from existing 99679561Siedowse * dirhashes. Returns zero with ufsdirhash_mtx locked if successful. 99779561Siedowse */ 99879561Siedowsestatic int 99979561Siedowseufsdirhash_recycle(int wanted) 100079561Siedowse{ 100179561Siedowse struct dirhash *dh; 100279561Siedowse doff_t **hash; 100379561Siedowse u_int8_t *blkfree; 100479561Siedowse int i, mem, narrays; 100579561Siedowse 100679561Siedowse mtx_lock(&ufsdirhash_mtx); 100779561Siedowse while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 100879561Siedowse /* Find a dirhash, and lock it. */ 100979561Siedowse if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) { 101079561Siedowse mtx_unlock(&ufsdirhash_mtx); 101179561Siedowse return (-1); 101279561Siedowse } 101379561Siedowse mtx_lock(&dh->dh_mtx); 101479561Siedowse KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 101579561Siedowse 101679561Siedowse /* Decrement the score; only recycle if it becomes zero. */ 101779561Siedowse if (--dh->dh_score > 0) { 101879561Siedowse mtx_unlock(&dh->dh_mtx); 101979561Siedowse mtx_unlock(&ufsdirhash_mtx); 102079561Siedowse return (-1); 102179561Siedowse } 102279561Siedowse 102379561Siedowse /* Remove it from the list and detach its memory. */ 102479561Siedowse TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 102579561Siedowse dh->dh_onlist = 0; 102679561Siedowse hash = dh->dh_hash; 102779561Siedowse dh->dh_hash = NULL; 102879561Siedowse blkfree = dh->dh_blkfree; 102979561Siedowse dh->dh_blkfree = NULL; 103079561Siedowse narrays = dh->dh_narrays; 103179561Siedowse mem = narrays * sizeof(*dh->dh_hash) + 103279561Siedowse narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 103379561Siedowse dh->dh_nblk * sizeof(*dh->dh_blkfree); 103479561Siedowse 103579561Siedowse /* Unlock everything, free the detached memory. */ 103679561Siedowse mtx_unlock(&dh->dh_mtx); 103779561Siedowse mtx_unlock(&ufsdirhash_mtx); 103879561Siedowse for (i = 0; i < narrays; i++) 103979561Siedowse zfree(ufsdirhash_zone, hash[i]); 104079561Siedowse FREE(hash, M_DIRHASH); 104179561Siedowse FREE(blkfree, M_DIRHASH); 104279561Siedowse 104379561Siedowse /* Account for the returned memory, and repeat if necessary. */ 104479561Siedowse mtx_lock(&ufsdirhash_mtx); 104579561Siedowse ufs_dirhashmem -= mem; 104679561Siedowse } 104779561Siedowse /* Success; return with ufsdirhash_mtx locked. */ 104879561Siedowse return (0); 104979561Siedowse} 105079561Siedowse 105179561Siedowse 105279561Siedowsestatic void 105379561Siedowseufsdirhash_init() 105479561Siedowse{ 105579561Siedowse ufsdirhash_zone = zinit("DIRHASH", DH_NBLKOFF * sizeof(daddr_t), 0, 105679561Siedowse 0, 1); 105779561Siedowse mtx_init(&ufsdirhash_mtx, "dirhash list", MTX_DEF); 105879561Siedowse TAILQ_INIT(&ufsdirhash_list); 105979561Siedowse} 106079561SiedowseSYSINIT(ufsdirhash, SI_SUB_PSEUDO, SI_ORDER_ANY, ufsdirhash_init, NULL) 106179561Siedowse 106279561Siedowse 106379561Siedowse#endif /* UFS_DIRHASH */ 1064