ufs_dirhash.c revision 107868
179561Siedowse/* 2107868Siedowse * Copyright (c) 2001, 2002 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 107868 2002-12-14 17:16:16Z iedowse $ 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> 4892768Sjeff#include <vm/uma.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)) 5992807Sdwmalone#define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1)) 6079561Siedowse#define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0) 6192098Siedowse#define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n)) 6279561Siedowse 6379561Siedowsestatic MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables"); 6479561Siedowse 6579561SiedowseSYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem"); 6679561Siedowse 6779561Siedowsestatic int ufs_mindirhashsize = DIRBLKSIZ * 5; 6879561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW, 6979561Siedowse &ufs_mindirhashsize, 7079561Siedowse 0, "minimum directory size in bytes for which to use hashed lookup"); 7179561Siedowsestatic int ufs_dirhashmaxmem = 2 * 1024 * 1024; 7279561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem, 7379561Siedowse 0, "maximum allowed dirhash memory usage"); 7479561Siedowsestatic int ufs_dirhashmem; 7579561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem, 7679561Siedowse 0, "current dirhash memory usage"); 7785512Siedowsestatic int ufs_dirhashcheck = 0; 7879561SiedowseSYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck, 7979561Siedowse 0, "enable extra sanity tests"); 8079561Siedowse 8179561Siedowse 8279561Siedowsestatic int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen); 8379561Siedowsestatic void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff); 8479561Siedowsestatic void ufsdirhash_delslot(struct dirhash *dh, int slot); 8579561Siedowsestatic int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, 8679561Siedowse doff_t offset); 8779561Siedowsestatic doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset); 8879561Siedowsestatic int ufsdirhash_recycle(int wanted); 8979561Siedowse 9092768Sjeffstatic uma_zone_t ufsdirhash_zone; 9179561Siedowse 9279561Siedowse/* Dirhash list; recently-used entries are near the tail. */ 9379561Siedowsestatic TAILQ_HEAD(, dirhash) ufsdirhash_list; 9479561Siedowse 9579561Siedowse/* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */ 9679561Siedowsestatic struct mtx ufsdirhash_mtx; 9779561Siedowse 9879561Siedowse/* 9979561Siedowse * Locking order: 10079561Siedowse * ufsdirhash_mtx 10179561Siedowse * dh_mtx 10279561Siedowse * 103107868Siedowse * The dh_mtx mutex should be acquired either via the inode lock, or via 10479561Siedowse * ufsdirhash_mtx. Only the owner of the inode may free the associated 10579561Siedowse * dirhash, but anything can steal its memory and set dh_hash to NULL. 10679561Siedowse */ 10779561Siedowse 10879561Siedowse/* 10979561Siedowse * Attempt to build up a hash table for the directory contents in 11079561Siedowse * inode 'ip'. Returns 0 on success, or -1 of the operation failed. 11179561Siedowse */ 11279561Siedowseint 11379561Siedowseufsdirhash_build(struct inode *ip) 11479561Siedowse{ 11579561Siedowse struct dirhash *dh; 11679561Siedowse struct buf *bp = NULL; 11779561Siedowse struct direct *ep; 11879561Siedowse struct vnode *vp; 11979561Siedowse doff_t bmask, pos; 12079561Siedowse int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot; 12179561Siedowse 12279561Siedowse /* Check if we can/should use dirhash. */ 12379561Siedowse if (ip->i_dirhash == NULL) { 12479561Siedowse if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode)) 12579561Siedowse return (-1); 12679561Siedowse } else { 12779561Siedowse /* Hash exists, but sysctls could have changed. */ 12879561Siedowse if (ip->i_size < ufs_mindirhashsize || 12979561Siedowse ufs_dirhashmem > ufs_dirhashmaxmem) { 13079561Siedowse ufsdirhash_free(ip); 13179561Siedowse return (-1); 13279561Siedowse } 13379561Siedowse /* Check if hash exists and is intact (note: unlocked read). */ 13479561Siedowse if (ip->i_dirhash->dh_hash != NULL) 13579561Siedowse return (0); 13679561Siedowse /* Free the old, recycled hash and build a new one. */ 13779561Siedowse ufsdirhash_free(ip); 13879561Siedowse } 13979561Siedowse 14082364Siedowse /* Don't hash removed directories. */ 14182364Siedowse if (ip->i_effnlink == 0) 14282364Siedowse return (-1); 14382364Siedowse 14479561Siedowse vp = ip->i_vnode; 14579561Siedowse /* Allocate 50% more entries than this dir size could ever need. */ 14679561Siedowse KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size")); 14779561Siedowse nslots = ip->i_size / DIRECTSIZ(1); 14879561Siedowse nslots = (nslots * 3 + 1) / 2; 14979561Siedowse narrays = howmany(nslots, DH_NBLKOFF); 15079561Siedowse nslots = narrays * DH_NBLKOFF; 15179561Siedowse dirblocks = howmany(ip->i_size, DIRBLKSIZ); 15279561Siedowse nblocks = (dirblocks * 3 + 1) / 2; 15379561Siedowse 15479561Siedowse memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) + 15579561Siedowse narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 15679561Siedowse nblocks * sizeof(*dh->dh_blkfree); 15779561Siedowse mtx_lock(&ufsdirhash_mtx); 15879561Siedowse if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) { 15979561Siedowse mtx_unlock(&ufsdirhash_mtx); 16079561Siedowse if (memreqd > ufs_dirhashmaxmem / 2) 16179561Siedowse return (-1); 16279561Siedowse 16379561Siedowse /* Try to free some space. */ 16479561Siedowse if (ufsdirhash_recycle(memreqd) != 0) 16579561Siedowse return (-1); 16679561Siedowse /* Enough was freed, and ufsdirhash_mtx has been locked. */ 16779561Siedowse } 16879561Siedowse ufs_dirhashmem += memreqd; 16979561Siedowse mtx_unlock(&ufsdirhash_mtx); 17079561Siedowse 17179561Siedowse /* 17279561Siedowse * Use non-blocking mallocs so that we will revert to a linear 17379561Siedowse * lookup on failure rather than potentially blocking forever. 17479561Siedowse */ 17579561Siedowse MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO); 176107868Siedowse if (dh == NULL) { 177107868Siedowse mtx_lock(&ufsdirhash_mtx); 178107868Siedowse ufs_dirhashmem -= memreqd; 179107868Siedowse mtx_unlock(&ufsdirhash_mtx); 18079561Siedowse return (-1); 181107868Siedowse } 18279561Siedowse MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]), 18379561Siedowse M_DIRHASH, M_NOWAIT | M_ZERO); 18479561Siedowse MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]), 18579561Siedowse M_DIRHASH, M_NOWAIT); 18679561Siedowse if (dh->dh_hash == NULL || dh->dh_blkfree == NULL) 18779561Siedowse goto fail; 18879561Siedowse for (i = 0; i < narrays; i++) { 18992768Sjeff if ((dh->dh_hash[i] = uma_zalloc(ufsdirhash_zone, 19092768Sjeff M_WAITOK)) == NULL) 19179561Siedowse goto fail; 19279561Siedowse for (j = 0; j < DH_NBLKOFF; j++) 19379561Siedowse dh->dh_hash[i][j] = DIRHASH_EMPTY; 19479561Siedowse } 19579561Siedowse 19679561Siedowse /* Initialise the hash table and block statistics. */ 19793818Sjhb mtx_init(&dh->dh_mtx, "dirhash", NULL, MTX_DEF); 19879561Siedowse dh->dh_narrays = narrays; 19979561Siedowse dh->dh_hlen = nslots; 20079561Siedowse dh->dh_nblk = nblocks; 20179561Siedowse dh->dh_dirblks = dirblocks; 20279561Siedowse for (i = 0; i < dirblocks; i++) 20379561Siedowse dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN; 20479561Siedowse for (i = 0; i < DH_NFSTATS; i++) 20579561Siedowse dh->dh_firstfree[i] = -1; 20679561Siedowse dh->dh_firstfree[DH_NFSTATS] = 0; 20779561Siedowse dh->dh_seqopt = 0; 20879561Siedowse dh->dh_seqoff = 0; 20979561Siedowse dh->dh_score = DH_SCOREINIT; 21079561Siedowse ip->i_dirhash = dh; 21179561Siedowse 21279561Siedowse bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 21379561Siedowse pos = 0; 21479561Siedowse while (pos < ip->i_size) { 21579561Siedowse /* If necessary, get the next directory block. */ 21679561Siedowse if ((pos & bmask) == 0) { 21779561Siedowse if (bp != NULL) 21879561Siedowse brelse(bp); 21979561Siedowse if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0) 22079561Siedowse goto fail; 22179561Siedowse } 22279561Siedowse 22379561Siedowse /* Add this entry to the hash. */ 22479561Siedowse ep = (struct direct *)((char *)bp->b_data + (pos & bmask)); 22579561Siedowse if (ep->d_reclen == 0 || ep->d_reclen > 22679561Siedowse DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) { 22779561Siedowse /* Corrupted directory. */ 22879561Siedowse brelse(bp); 22979561Siedowse goto fail; 23079561Siedowse } 23179561Siedowse if (ep->d_ino != 0) { 23279561Siedowse /* Add the entry (simplified ufsdirhash_add). */ 23379561Siedowse slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen); 23479561Siedowse while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 23579561Siedowse slot = WRAPINCR(slot, dh->dh_hlen); 23679561Siedowse dh->dh_hused++; 23779561Siedowse DH_ENTRY(dh, slot) = pos; 23879561Siedowse ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep)); 23979561Siedowse } 24079561Siedowse pos += ep->d_reclen; 24179561Siedowse } 24279561Siedowse 24379561Siedowse if (bp != NULL) 24479561Siedowse brelse(bp); 24579561Siedowse mtx_lock(&ufsdirhash_mtx); 24679561Siedowse TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list); 24779561Siedowse dh->dh_onlist = 1; 24879561Siedowse mtx_unlock(&ufsdirhash_mtx); 24979561Siedowse return (0); 25079561Siedowse 25179561Siedowsefail: 25279561Siedowse if (dh->dh_hash != NULL) { 25379561Siedowse for (i = 0; i < narrays; i++) 25479561Siedowse if (dh->dh_hash[i] != NULL) 25592768Sjeff uma_zfree(ufsdirhash_zone, dh->dh_hash[i]); 25679561Siedowse FREE(dh->dh_hash, M_DIRHASH); 25779561Siedowse } 25879561Siedowse if (dh->dh_blkfree != NULL) 25979561Siedowse FREE(dh->dh_blkfree, M_DIRHASH); 26079561Siedowse FREE(dh, M_DIRHASH); 26179561Siedowse ip->i_dirhash = NULL; 26279561Siedowse mtx_lock(&ufsdirhash_mtx); 26379561Siedowse ufs_dirhashmem -= memreqd; 26479561Siedowse mtx_unlock(&ufsdirhash_mtx); 26579561Siedowse return (-1); 26679561Siedowse} 26779561Siedowse 26879561Siedowse/* 26979561Siedowse * Free any hash table associated with inode 'ip'. 27079561Siedowse */ 27179561Siedowsevoid 27279561Siedowseufsdirhash_free(struct inode *ip) 27379561Siedowse{ 27479561Siedowse struct dirhash *dh; 27579561Siedowse int i, mem; 27679561Siedowse 27779561Siedowse if ((dh = ip->i_dirhash) == NULL) 27879561Siedowse return; 27979561Siedowse mtx_lock(&ufsdirhash_mtx); 28079561Siedowse mtx_lock(&dh->dh_mtx); 28179561Siedowse if (dh->dh_onlist) 28279561Siedowse TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 28379561Siedowse mtx_unlock(&dh->dh_mtx); 28479561Siedowse mtx_unlock(&ufsdirhash_mtx); 28579561Siedowse 28679561Siedowse /* The dirhash pointed to by 'dh' is exclusively ours now. */ 28779561Siedowse 28879561Siedowse mem = sizeof(*dh); 28979561Siedowse if (dh->dh_hash != NULL) { 29079561Siedowse for (i = 0; i < dh->dh_narrays; i++) 29192768Sjeff uma_zfree(ufsdirhash_zone, dh->dh_hash[i]); 29279561Siedowse FREE(dh->dh_hash, M_DIRHASH); 29379561Siedowse FREE(dh->dh_blkfree, M_DIRHASH); 29479561Siedowse mem += dh->dh_narrays * sizeof(*dh->dh_hash) + 29579561Siedowse dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 29679561Siedowse dh->dh_nblk * sizeof(*dh->dh_blkfree); 29779561Siedowse } 29879561Siedowse mtx_destroy(&dh->dh_mtx); 29979561Siedowse FREE(dh, M_DIRHASH); 30079561Siedowse ip->i_dirhash = NULL; 30179561Siedowse 30279561Siedowse mtx_lock(&ufsdirhash_mtx); 30379561Siedowse ufs_dirhashmem -= mem; 30479561Siedowse mtx_unlock(&ufsdirhash_mtx); 30579561Siedowse} 30679561Siedowse 30779561Siedowse/* 30879561Siedowse * Find the offset of the specified name within the given inode. 30979561Siedowse * Returns 0 on success, ENOENT if the entry does not exist, or 31079561Siedowse * EJUSTRETURN if the caller should revert to a linear search. 31179561Siedowse * 31279690Siedowse * If successful, the directory offset is stored in *offp, and a 31379690Siedowse * pointer to a struct buf containing the entry is stored in *bpp. If 31479561Siedowse * prevoffp is non-NULL, the offset of the previous entry within 31579561Siedowse * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry 31679561Siedowse * is the first in a block, the start of the block is used). 31779561Siedowse */ 31879561Siedowseint 31979561Siedowseufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp, 32079690Siedowse struct buf **bpp, doff_t *prevoffp) 32179561Siedowse{ 32279561Siedowse struct dirhash *dh, *dh_next; 32379561Siedowse struct direct *dp; 32479561Siedowse struct vnode *vp; 32579561Siedowse struct buf *bp; 32679561Siedowse doff_t blkoff, bmask, offset, prevoff; 32779561Siedowse int i, slot; 32879561Siedowse 32979561Siedowse if ((dh = ip->i_dirhash) == NULL) 33079561Siedowse return (EJUSTRETURN); 33179561Siedowse /* 33279561Siedowse * Move this dirhash towards the end of the list if it has a 333107868Siedowse * score higher than the next entry, and acquire the dh_mtx. 33479561Siedowse * Optimise the case where it's already the last by performing 33579561Siedowse * an unlocked read of the TAILQ_NEXT pointer. 33679561Siedowse * 33779561Siedowse * In both cases, end up holding just dh_mtx. 33879561Siedowse */ 33979561Siedowse if (TAILQ_NEXT(dh, dh_list) != NULL) { 34079561Siedowse mtx_lock(&ufsdirhash_mtx); 34179561Siedowse mtx_lock(&dh->dh_mtx); 34279561Siedowse /* 34379561Siedowse * If the new score will be greater than that of the next 34479561Siedowse * entry, then move this entry past it. With both mutexes 34579561Siedowse * held, dh_next won't go away, but its dh_score could 34679561Siedowse * change; that's not important since it is just a hint. 34779561Siedowse */ 34879561Siedowse if (dh->dh_hash != NULL && 34979561Siedowse (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL && 35079561Siedowse dh->dh_score >= dh_next->dh_score) { 35179561Siedowse KASSERT(dh->dh_onlist, ("dirhash: not on list")); 35279561Siedowse TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 35379561Siedowse TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh, 35479561Siedowse dh_list); 35579561Siedowse } 35679561Siedowse mtx_unlock(&ufsdirhash_mtx); 35779561Siedowse } else { 35879561Siedowse /* Already the last, though that could change as we wait. */ 35979561Siedowse mtx_lock(&dh->dh_mtx); 36079561Siedowse } 36179561Siedowse if (dh->dh_hash == NULL) { 36279561Siedowse mtx_unlock(&dh->dh_mtx); 36379561Siedowse ufsdirhash_free(ip); 36479561Siedowse return (EJUSTRETURN); 36579561Siedowse } 36679561Siedowse 36779561Siedowse /* Update the score. */ 36879561Siedowse if (dh->dh_score < DH_SCOREMAX) 36979561Siedowse dh->dh_score++; 37079561Siedowse 37179561Siedowse vp = ip->i_vnode; 37279561Siedowse bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1; 37379561Siedowse blkoff = -1; 37479561Siedowse bp = NULL; 37579561Siedowserestart: 37679561Siedowse slot = ufsdirhash_hash(dh, name, namelen); 37779561Siedowse 37879561Siedowse if (dh->dh_seqopt) { 37979561Siedowse /* 38079561Siedowse * Sequential access optimisation. dh_seqoff contains the 38179561Siedowse * offset of the directory entry immediately following 38279561Siedowse * the last entry that was looked up. Check if this offset 38379561Siedowse * appears in the hash chain for the name we are looking for. 38479561Siedowse */ 38579561Siedowse for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY; 38679561Siedowse i = WRAPINCR(i, dh->dh_hlen)) 38786350Siedowse if (offset == dh->dh_seqoff) 38879561Siedowse break; 38979561Siedowse if (offset == dh->dh_seqoff) { 39079561Siedowse /* 39179561Siedowse * We found an entry with the expected offset. This 39279561Siedowse * is probably the entry we want, but if not, the 39379561Siedowse * code below will turn off seqoff and retry. 39479561Siedowse */ 39579561Siedowse slot = i; 39679561Siedowse } else 39779561Siedowse dh->dh_seqopt = 0; 39879561Siedowse } 39979561Siedowse 40079561Siedowse for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY; 40179561Siedowse slot = WRAPINCR(slot, dh->dh_hlen)) { 40279561Siedowse if (offset == DIRHASH_DEL) 40379561Siedowse continue; 40479561Siedowse mtx_unlock(&dh->dh_mtx); 40579561Siedowse 40679561Siedowse if (offset < 0 || offset >= ip->i_size) 40779561Siedowse panic("ufsdirhash_lookup: bad offset in hash array"); 40879561Siedowse if ((offset & ~bmask) != blkoff) { 40979561Siedowse if (bp != NULL) 41079561Siedowse brelse(bp); 41179561Siedowse blkoff = offset & ~bmask; 41279561Siedowse if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) 41379561Siedowse return (EJUSTRETURN); 41479561Siedowse } 41579561Siedowse dp = (struct direct *)(bp->b_data + (offset & bmask)); 41679561Siedowse if (dp->d_reclen == 0 || dp->d_reclen > 41779561Siedowse DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) { 41879561Siedowse /* Corrupted directory. */ 41979561Siedowse brelse(bp); 42079561Siedowse return (EJUSTRETURN); 42179561Siedowse } 42279561Siedowse if (dp->d_namlen == namelen && 42379561Siedowse bcmp(dp->d_name, name, namelen) == 0) { 42479561Siedowse /* Found. Get the prev offset if needed. */ 42579561Siedowse if (prevoffp != NULL) { 42679561Siedowse if (offset & (DIRBLKSIZ - 1)) { 42779561Siedowse prevoff = ufsdirhash_getprev(dp, 42879561Siedowse offset); 42979561Siedowse if (prevoff == -1) { 43079561Siedowse brelse(bp); 43179561Siedowse return (EJUSTRETURN); 43279561Siedowse } 43379561Siedowse } else 43479561Siedowse prevoff = offset; 43579561Siedowse *prevoffp = prevoff; 43679561Siedowse } 43779561Siedowse 43879561Siedowse /* Check for sequential access, and update offset. */ 43979561Siedowse if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset) 44079561Siedowse dh->dh_seqopt = 1; 44179561Siedowse dh->dh_seqoff = offset + DIRSIZ(0, dp); 44279561Siedowse 44379690Siedowse *bpp = bp; 44479561Siedowse *offp = offset; 44579561Siedowse return (0); 44679561Siedowse } 44779561Siedowse 44879561Siedowse mtx_lock(&dh->dh_mtx); 44979561Siedowse if (dh->dh_hash == NULL) { 45079561Siedowse mtx_unlock(&dh->dh_mtx); 45179561Siedowse if (bp != NULL) 45279561Siedowse brelse(bp); 45379561Siedowse ufsdirhash_free(ip); 45479561Siedowse return (EJUSTRETURN); 45579561Siedowse } 45679561Siedowse /* 45779561Siedowse * When the name doesn't match in the seqopt case, go back 45879561Siedowse * and search normally. 45979561Siedowse */ 46079561Siedowse if (dh->dh_seqopt) { 46179561Siedowse dh->dh_seqopt = 0; 46279561Siedowse goto restart; 46379561Siedowse } 46479561Siedowse } 46579561Siedowse mtx_unlock(&dh->dh_mtx); 46679561Siedowse if (bp != NULL) 46779561Siedowse brelse(bp); 46879561Siedowse return (ENOENT); 46979561Siedowse} 47079561Siedowse 47179561Siedowse/* 47279561Siedowse * Find a directory block with room for 'slotneeded' bytes. Returns 47379561Siedowse * the offset of the directory entry that begins the free space. 47479561Siedowse * This will either be the offset of an existing entry that has free 47579561Siedowse * space at the end, or the offset of an entry with d_ino == 0 at 47679561Siedowse * the start of a DIRBLKSIZ block. 47779561Siedowse * 47879561Siedowse * To use the space, the caller may need to compact existing entries in 47979561Siedowse * the directory. The total number of bytes in all of the entries involved 48079561Siedowse * in the compaction is stored in *slotsize. In other words, all of 48179561Siedowse * the entries that must be compacted are exactly contained in the 48279561Siedowse * region beginning at the returned offset and spanning *slotsize bytes. 48379561Siedowse * 48479561Siedowse * Returns -1 if no space was found, indicating that the directory 48579561Siedowse * must be extended. 48679561Siedowse */ 48779561Siedowsedoff_t 48879561Siedowseufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize) 48979561Siedowse{ 49079561Siedowse struct direct *dp; 49179561Siedowse struct dirhash *dh; 49279561Siedowse struct buf *bp; 49379561Siedowse doff_t pos, slotstart; 49479561Siedowse int dirblock, error, freebytes, i; 49579561Siedowse 49679561Siedowse if ((dh = ip->i_dirhash) == NULL) 49779561Siedowse return (-1); 49879561Siedowse mtx_lock(&dh->dh_mtx); 49979561Siedowse if (dh->dh_hash == NULL) { 50079561Siedowse mtx_unlock(&dh->dh_mtx); 50179561Siedowse ufsdirhash_free(ip); 50279561Siedowse return (-1); 50379561Siedowse } 50479561Siedowse 50579561Siedowse /* Find a directory block with the desired free space. */ 50679561Siedowse dirblock = -1; 50779561Siedowse for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++) 50879561Siedowse if ((dirblock = dh->dh_firstfree[i]) != -1) 50979561Siedowse break; 51079561Siedowse if (dirblock == -1) { 51179561Siedowse mtx_unlock(&dh->dh_mtx); 51279561Siedowse return (-1); 51379561Siedowse } 51479561Siedowse 51579561Siedowse KASSERT(dirblock < dh->dh_nblk && 51679561Siedowse dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN), 51779561Siedowse ("ufsdirhash_findfree: bad stats")); 51879561Siedowse mtx_unlock(&dh->dh_mtx); 51979561Siedowse pos = dirblock * DIRBLKSIZ; 52079561Siedowse error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp); 52179561Siedowse if (error) 52279561Siedowse return (-1); 52379561Siedowse 52479561Siedowse /* Find the first entry with free space. */ 52579561Siedowse for (i = 0; i < DIRBLKSIZ; ) { 52679561Siedowse if (dp->d_reclen == 0) { 52779561Siedowse brelse(bp); 52879561Siedowse return (-1); 52979561Siedowse } 53079561Siedowse if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp)) 53179561Siedowse break; 53279561Siedowse i += dp->d_reclen; 53379561Siedowse dp = (struct direct *)((char *)dp + dp->d_reclen); 53479561Siedowse } 53579561Siedowse if (i > DIRBLKSIZ) { 53679561Siedowse brelse(bp); 53779561Siedowse return (-1); 53879561Siedowse } 53979561Siedowse slotstart = pos + i; 54079561Siedowse 54179561Siedowse /* Find the range of entries needed to get enough space */ 54279561Siedowse freebytes = 0; 54379561Siedowse while (i < DIRBLKSIZ && freebytes < slotneeded) { 54479561Siedowse freebytes += dp->d_reclen; 54579561Siedowse if (dp->d_ino != 0) 54679561Siedowse freebytes -= DIRSIZ(0, dp); 54779561Siedowse if (dp->d_reclen == 0) { 54879561Siedowse brelse(bp); 54979561Siedowse return (-1); 55079561Siedowse } 55179561Siedowse i += dp->d_reclen; 55279561Siedowse dp = (struct direct *)((char *)dp + dp->d_reclen); 55379561Siedowse } 55479561Siedowse if (i > DIRBLKSIZ) { 55579561Siedowse brelse(bp); 55679561Siedowse return (-1); 55779561Siedowse } 55879561Siedowse if (freebytes < slotneeded) 55979561Siedowse panic("ufsdirhash_findfree: free mismatch"); 56079561Siedowse brelse(bp); 56179561Siedowse *slotsize = pos + i - slotstart; 56279561Siedowse return (slotstart); 56379561Siedowse} 56479561Siedowse 56579561Siedowse/* 56679561Siedowse * Return the start of the unused space at the end of a directory, or 56779561Siedowse * -1 if there are no trailing unused blocks. 56879561Siedowse */ 56979561Siedowsedoff_t 57079561Siedowseufsdirhash_enduseful(struct inode *ip) 57179561Siedowse{ 57279561Siedowse 57379561Siedowse struct dirhash *dh; 57479561Siedowse int i; 57579561Siedowse 57679561Siedowse if ((dh = ip->i_dirhash) == NULL) 57779561Siedowse return (-1); 57879561Siedowse mtx_lock(&dh->dh_mtx); 57979561Siedowse if (dh->dh_hash == NULL) { 58079561Siedowse mtx_unlock(&dh->dh_mtx); 58179561Siedowse ufsdirhash_free(ip); 58279561Siedowse return (-1); 58379561Siedowse } 58479561Siedowse 58579561Siedowse if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) { 58679561Siedowse mtx_unlock(&dh->dh_mtx); 58779561Siedowse return (-1); 58879561Siedowse } 58979561Siedowse 59079561Siedowse for (i = dh->dh_dirblks - 1; i >= 0; i--) 59179561Siedowse if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 59279561Siedowse break; 59379561Siedowse mtx_unlock(&dh->dh_mtx); 59479561Siedowse return ((doff_t)(i + 1) * DIRBLKSIZ); 59579561Siedowse} 59679561Siedowse 59779561Siedowse/* 59879561Siedowse * Insert information into the hash about a new directory entry. dirp 59979561Siedowse * points to a struct direct containing the entry, and offset specifies 60079561Siedowse * the offset of this entry. 60179561Siedowse */ 60279561Siedowsevoid 60379561Siedowseufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset) 60479561Siedowse{ 60579561Siedowse struct dirhash *dh; 60679561Siedowse int slot; 60779561Siedowse 60879561Siedowse if ((dh = ip->i_dirhash) == NULL) 60979561Siedowse return; 61079561Siedowse mtx_lock(&dh->dh_mtx); 61179561Siedowse if (dh->dh_hash == NULL) { 61279561Siedowse mtx_unlock(&dh->dh_mtx); 61379561Siedowse ufsdirhash_free(ip); 61479561Siedowse return; 61579561Siedowse } 61679561Siedowse 61779561Siedowse KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 61879561Siedowse ("ufsdirhash_add: bad offset")); 61979561Siedowse /* 62079561Siedowse * Normal hash usage is < 66%. If the usage gets too high then 62179561Siedowse * remove the hash entirely and let it be rebuilt later. 62279561Siedowse */ 62379561Siedowse if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) { 62479561Siedowse mtx_unlock(&dh->dh_mtx); 62579561Siedowse ufsdirhash_free(ip); 62679561Siedowse return; 62779561Siedowse } 62879561Siedowse 62979561Siedowse /* Find a free hash slot (empty or deleted), and add the entry. */ 63079561Siedowse slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen); 63179561Siedowse while (DH_ENTRY(dh, slot) >= 0) 63279561Siedowse slot = WRAPINCR(slot, dh->dh_hlen); 63379561Siedowse if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY) 63479561Siedowse dh->dh_hused++; 63579561Siedowse DH_ENTRY(dh, slot) = offset; 63679561Siedowse 63779561Siedowse /* Update the per-block summary info. */ 63879561Siedowse ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp)); 63979561Siedowse mtx_unlock(&dh->dh_mtx); 64079561Siedowse} 64179561Siedowse 64279561Siedowse/* 64379561Siedowse * Remove the specified directory entry from the hash. The entry to remove 64479561Siedowse * is defined by the name in `dirp', which must exist at the specified 64579561Siedowse * `offset' within the directory. 64679561Siedowse */ 64779561Siedowsevoid 64879561Siedowseufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset) 64979561Siedowse{ 65079561Siedowse struct dirhash *dh; 65179561Siedowse int slot; 65279561Siedowse 65379561Siedowse if ((dh = ip->i_dirhash) == NULL) 65479561Siedowse return; 65579561Siedowse mtx_lock(&dh->dh_mtx); 65679561Siedowse if (dh->dh_hash == NULL) { 65779561Siedowse mtx_unlock(&dh->dh_mtx); 65879561Siedowse ufsdirhash_free(ip); 65979561Siedowse return; 66079561Siedowse } 66179561Siedowse 66279561Siedowse KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ, 66379561Siedowse ("ufsdirhash_remove: bad offset")); 66479561Siedowse /* Find the entry */ 66579561Siedowse slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset); 66679561Siedowse 66779561Siedowse /* Remove the hash entry. */ 66879561Siedowse ufsdirhash_delslot(dh, slot); 66979561Siedowse 67079561Siedowse /* Update the per-block summary info. */ 67179561Siedowse ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp)); 67279561Siedowse mtx_unlock(&dh->dh_mtx); 67379561Siedowse} 67479561Siedowse 67579561Siedowse/* 67679561Siedowse * Change the offset associated with a directory entry in the hash. Used 67779561Siedowse * when compacting directory blocks. 67879561Siedowse */ 67979561Siedowsevoid 68079561Siedowseufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff, 68179561Siedowse doff_t newoff) 68279561Siedowse{ 68379561Siedowse struct dirhash *dh; 68479561Siedowse int slot; 68579561Siedowse 68679561Siedowse if ((dh = ip->i_dirhash) == NULL) 68779561Siedowse return; 68879561Siedowse mtx_lock(&dh->dh_mtx); 68979561Siedowse if (dh->dh_hash == NULL) { 69079561Siedowse mtx_unlock(&dh->dh_mtx); 69179561Siedowse ufsdirhash_free(ip); 69279561Siedowse return; 69379561Siedowse } 69479561Siedowse 69579561Siedowse KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ && 69679561Siedowse newoff < dh->dh_dirblks * DIRBLKSIZ, 69779561Siedowse ("ufsdirhash_move: bad offset")); 69879561Siedowse /* Find the entry, and update the offset. */ 69979561Siedowse slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff); 70079561Siedowse DH_ENTRY(dh, slot) = newoff; 70179561Siedowse mtx_unlock(&dh->dh_mtx); 70279561Siedowse} 70379561Siedowse 70479561Siedowse/* 70579561Siedowse * Inform dirhash that the directory has grown by one block that 70679561Siedowse * begins at offset (i.e. the new length is offset + DIRBLKSIZ). 70779561Siedowse */ 70879561Siedowsevoid 70979561Siedowseufsdirhash_newblk(struct inode *ip, doff_t offset) 71079561Siedowse{ 71179561Siedowse struct dirhash *dh; 71279561Siedowse int block; 71379561Siedowse 71479561Siedowse if ((dh = ip->i_dirhash) == NULL) 71579561Siedowse return; 71679561Siedowse mtx_lock(&dh->dh_mtx); 71779561Siedowse if (dh->dh_hash == NULL) { 71879561Siedowse mtx_unlock(&dh->dh_mtx); 71979561Siedowse ufsdirhash_free(ip); 72079561Siedowse return; 72179561Siedowse } 72279561Siedowse 72379561Siedowse KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ, 72479561Siedowse ("ufsdirhash_newblk: bad offset")); 72579561Siedowse block = offset / DIRBLKSIZ; 72679561Siedowse if (block >= dh->dh_nblk) { 72779561Siedowse /* Out of space; must rebuild. */ 72879561Siedowse mtx_unlock(&dh->dh_mtx); 72979561Siedowse ufsdirhash_free(ip); 73079561Siedowse return; 73179561Siedowse } 73279561Siedowse dh->dh_dirblks = block + 1; 73379561Siedowse 73479561Siedowse /* Account for the new free block. */ 73579561Siedowse dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN; 73679561Siedowse if (dh->dh_firstfree[DH_NFSTATS] == -1) 73779561Siedowse dh->dh_firstfree[DH_NFSTATS] = block; 73879561Siedowse mtx_unlock(&dh->dh_mtx); 73979561Siedowse} 74079561Siedowse 74179561Siedowse/* 74279561Siedowse * Inform dirhash that the directory is being truncated. 74379561Siedowse */ 74479561Siedowsevoid 74579561Siedowseufsdirhash_dirtrunc(struct inode *ip, doff_t offset) 74679561Siedowse{ 74779561Siedowse struct dirhash *dh; 74879561Siedowse int block, i; 74979561Siedowse 75079561Siedowse if ((dh = ip->i_dirhash) == NULL) 75179561Siedowse return; 75279561Siedowse mtx_lock(&dh->dh_mtx); 75379561Siedowse if (dh->dh_hash == NULL) { 75479561Siedowse mtx_unlock(&dh->dh_mtx); 75579561Siedowse ufsdirhash_free(ip); 75679561Siedowse return; 75779561Siedowse } 75879561Siedowse 75979561Siedowse KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ, 76079561Siedowse ("ufsdirhash_dirtrunc: bad offset")); 76179561Siedowse block = howmany(offset, DIRBLKSIZ); 76279561Siedowse /* 76379561Siedowse * If the directory shrinks to less than 1/8 of dh_nblk blocks 76479561Siedowse * (about 20% of its original size due to the 50% extra added in 76579561Siedowse * ufsdirhash_build) then free it, and let the caller rebuild 76679561Siedowse * if necessary. 76779561Siedowse */ 76879561Siedowse if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) { 76979561Siedowse mtx_unlock(&dh->dh_mtx); 77079561Siedowse ufsdirhash_free(ip); 77179561Siedowse return; 77279561Siedowse } 77379561Siedowse 77479561Siedowse /* 77579561Siedowse * Remove any `first free' information pertaining to the 77679561Siedowse * truncated blocks. All blocks we're removing should be 77779561Siedowse * completely unused. 77879561Siedowse */ 77979561Siedowse if (dh->dh_firstfree[DH_NFSTATS] >= block) 78079561Siedowse dh->dh_firstfree[DH_NFSTATS] = -1; 78179561Siedowse for (i = block; i < dh->dh_dirblks; i++) 78279561Siedowse if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN) 78379561Siedowse panic("ufsdirhash_dirtrunc: blocks in use"); 78479561Siedowse for (i = 0; i < DH_NFSTATS; i++) 78579561Siedowse if (dh->dh_firstfree[i] >= block) 78679561Siedowse panic("ufsdirhash_dirtrunc: first free corrupt"); 78779561Siedowse dh->dh_dirblks = block; 78879561Siedowse mtx_unlock(&dh->dh_mtx); 78979561Siedowse} 79079561Siedowse 79179561Siedowse/* 79279561Siedowse * Debugging function to check that the dirhash information about 79379561Siedowse * a directory block matches its actual contents. Panics if a mismatch 79479561Siedowse * is detected. 79579561Siedowse * 79679561Siedowse * On entry, `buf' should point to the start of an in-core 79779561Siedowse * DIRBLKSIZ-sized directory block, and `offset' should contain the 79879561Siedowse * offset from the start of the directory of that block. 79979561Siedowse */ 80079561Siedowsevoid 80179561Siedowseufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset) 80279561Siedowse{ 80379561Siedowse struct dirhash *dh; 80479561Siedowse struct direct *dp; 80579561Siedowse int block, ffslot, i, nfree; 80679561Siedowse 80779561Siedowse if (!ufs_dirhashcheck) 80879561Siedowse return; 80979561Siedowse if ((dh = ip->i_dirhash) == NULL) 81079561Siedowse return; 81179561Siedowse mtx_lock(&dh->dh_mtx); 81279561Siedowse if (dh->dh_hash == NULL) { 81379561Siedowse mtx_unlock(&dh->dh_mtx); 81479561Siedowse ufsdirhash_free(ip); 81579561Siedowse return; 81679561Siedowse } 81779561Siedowse 81879561Siedowse block = offset / DIRBLKSIZ; 81979561Siedowse if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks) 82079561Siedowse panic("ufsdirhash_checkblock: bad offset"); 82179561Siedowse 82279561Siedowse nfree = 0; 82379561Siedowse for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) { 82479561Siedowse dp = (struct direct *)(buf + i); 82579561Siedowse if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ) 82679561Siedowse panic("ufsdirhash_checkblock: bad dir"); 82779561Siedowse 82879561Siedowse if (dp->d_ino == 0) { 82980456Siedowse#if 0 83080456Siedowse /* 83180456Siedowse * XXX entries with d_ino == 0 should only occur 83280456Siedowse * at the start of a DIRBLKSIZ block. However the 83380456Siedowse * ufs code is tolerant of such entries at other 83480456Siedowse * offsets, and fsck does not fix them. 83580456Siedowse */ 83679561Siedowse if (i != 0) 83779561Siedowse panic("ufsdirhash_checkblock: bad dir inode"); 83880456Siedowse#endif 83979561Siedowse nfree += dp->d_reclen; 84079561Siedowse continue; 84179561Siedowse } 84279561Siedowse 84379561Siedowse /* Check that the entry exists (will panic if it doesn't). */ 84479561Siedowse ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i); 84579561Siedowse 84679561Siedowse nfree += dp->d_reclen - DIRSIZ(0, dp); 84779561Siedowse } 84879561Siedowse if (i != DIRBLKSIZ) 84979561Siedowse panic("ufsdirhash_checkblock: bad dir end"); 85079561Siedowse 85179561Siedowse if (dh->dh_blkfree[block] * DIRALIGN != nfree) 85279561Siedowse panic("ufsdirhash_checkblock: bad free count"); 85379561Siedowse 85492098Siedowse ffslot = BLKFREE2IDX(nfree / DIRALIGN); 85579561Siedowse for (i = 0; i <= DH_NFSTATS; i++) 85679561Siedowse if (dh->dh_firstfree[i] == block && i != ffslot) 85779561Siedowse panic("ufsdirhash_checkblock: bad first-free"); 85892098Siedowse if (dh->dh_firstfree[ffslot] == -1) 85992098Siedowse panic("ufsdirhash_checkblock: missing first-free entry"); 86079561Siedowse mtx_unlock(&dh->dh_mtx); 86179561Siedowse} 86279561Siedowse 86379561Siedowse/* 86479561Siedowse * Hash the specified filename into a dirhash slot. 86579561Siedowse */ 86679561Siedowsestatic int 86779561Siedowseufsdirhash_hash(struct dirhash *dh, char *name, int namelen) 86879561Siedowse{ 86992807Sdwmalone u_int32_t hash; 87092807Sdwmalone 87192807Sdwmalone /* 872107868Siedowse * We hash the name and then some other bit of data that is 873107868Siedowse * invariant over the dirhash's lifetime. Otherwise names 87492807Sdwmalone * differing only in the last byte are placed close to one 87592807Sdwmalone * another in the table, which is bad for linear probing. 87692807Sdwmalone */ 87792807Sdwmalone hash = fnv_32_buf(name, namelen, FNV1_32_INIT); 87892807Sdwmalone hash = fnv_32_buf(dh, sizeof(dh), hash); 87992807Sdwmalone return (hash % dh->dh_hlen); 88079561Siedowse} 88179561Siedowse 88279561Siedowse/* 88379561Siedowse * Adjust the number of free bytes in the block containing `offset' 88479561Siedowse * by the value specified by `diff'. 88579561Siedowse * 88679561Siedowse * The caller must ensure we have exclusive access to `dh'; normally 88779561Siedowse * that means that dh_mtx should be held, but this is also called 88879561Siedowse * from ufsdirhash_build() where exclusive access can be assumed. 88979561Siedowse */ 89079561Siedowsestatic void 89179561Siedowseufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff) 89279561Siedowse{ 89379561Siedowse int block, i, nfidx, ofidx; 89479561Siedowse 89579561Siedowse /* Update the per-block summary info. */ 89679561Siedowse block = offset / DIRBLKSIZ; 89779561Siedowse KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks, 89879561Siedowse ("dirhash bad offset")); 89992098Siedowse ofidx = BLKFREE2IDX(dh->dh_blkfree[block]); 90079561Siedowse dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN); 90192098Siedowse nfidx = BLKFREE2IDX(dh->dh_blkfree[block]); 90279561Siedowse 90379561Siedowse /* Update the `first free' list if necessary. */ 90479561Siedowse if (ofidx != nfidx) { 90579561Siedowse /* If removing, scan forward for the next block. */ 90679561Siedowse if (dh->dh_firstfree[ofidx] == block) { 90779561Siedowse for (i = block + 1; i < dh->dh_dirblks; i++) 90892098Siedowse if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx) 90979561Siedowse break; 91079561Siedowse dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1; 91179561Siedowse } 91279561Siedowse 91379561Siedowse /* Make this the new `first free' if necessary */ 91479561Siedowse if (dh->dh_firstfree[nfidx] > block || 91579561Siedowse dh->dh_firstfree[nfidx] == -1) 91679561Siedowse dh->dh_firstfree[nfidx] = block; 91779561Siedowse } 91879561Siedowse} 91979561Siedowse 92079561Siedowse/* 92179561Siedowse * Find the specified name which should have the specified offset. 92279561Siedowse * Returns a slot number, and panics on failure. 92379561Siedowse * 92479561Siedowse * `dh' must be locked on entry and remains so on return. 92579561Siedowse */ 92679561Siedowsestatic int 92779561Siedowseufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset) 92879561Siedowse{ 92979561Siedowse int slot; 93079561Siedowse 93179561Siedowse mtx_assert(&dh->dh_mtx, MA_OWNED); 93279561Siedowse 93379561Siedowse /* Find the entry. */ 93479561Siedowse KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full")); 93579561Siedowse slot = ufsdirhash_hash(dh, name, namelen); 93679561Siedowse while (DH_ENTRY(dh, slot) != offset && 93779561Siedowse DH_ENTRY(dh, slot) != DIRHASH_EMPTY) 93879561Siedowse slot = WRAPINCR(slot, dh->dh_hlen); 93979561Siedowse if (DH_ENTRY(dh, slot) != offset) 94079561Siedowse panic("ufsdirhash_findslot: '%.*s' not found", namelen, name); 94179561Siedowse 94279561Siedowse return (slot); 94379561Siedowse} 94479561Siedowse 94579561Siedowse/* 94679561Siedowse * Remove the entry corresponding to the specified slot from the hash array. 94779561Siedowse * 94879561Siedowse * `dh' must be locked on entry and remains so on return. 94979561Siedowse */ 95079561Siedowsestatic void 95179561Siedowseufsdirhash_delslot(struct dirhash *dh, int slot) 95279561Siedowse{ 95379561Siedowse int i; 95479561Siedowse 95579561Siedowse mtx_assert(&dh->dh_mtx, MA_OWNED); 95679561Siedowse 95779561Siedowse /* Mark the entry as deleted. */ 95879561Siedowse DH_ENTRY(dh, slot) = DIRHASH_DEL; 95979561Siedowse 96079561Siedowse /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */ 96179561Siedowse for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; ) 96279561Siedowse i = WRAPINCR(i, dh->dh_hlen); 96379561Siedowse if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) { 96492807Sdwmalone i = WRAPDECR(i, dh->dh_hlen); 96592807Sdwmalone while (DH_ENTRY(dh, i) == DIRHASH_DEL) { 96679561Siedowse DH_ENTRY(dh, i) = DIRHASH_EMPTY; 96779561Siedowse dh->dh_hused--; 96892807Sdwmalone i = WRAPDECR(i, dh->dh_hlen); 96979561Siedowse } 97079561Siedowse KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen")); 97179561Siedowse } 97279561Siedowse} 97379561Siedowse 97479561Siedowse/* 97579561Siedowse * Given a directory entry and its offset, find the offset of the 97679561Siedowse * previous entry in the same DIRBLKSIZ-sized block. Returns an 97779561Siedowse * offset, or -1 if there is no previous entry in the block or some 97879561Siedowse * other problem occurred. 97979561Siedowse */ 98079561Siedowsestatic doff_t 98179561Siedowseufsdirhash_getprev(struct direct *dirp, doff_t offset) 98279561Siedowse{ 98379561Siedowse struct direct *dp; 98479561Siedowse char *blkbuf; 98579561Siedowse doff_t blkoff, prevoff; 98679561Siedowse int entrypos, i; 98779561Siedowse 98879561Siedowse blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */ 98979561Siedowse entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */ 99079561Siedowse blkbuf = (char *)dirp - entrypos; 99179561Siedowse prevoff = blkoff; 99279561Siedowse 99379561Siedowse /* If `offset' is the start of a block, there is no previous entry. */ 99479561Siedowse if (entrypos == 0) 99579561Siedowse return (-1); 99679561Siedowse 99779561Siedowse /* Scan from the start of the block until we get to the entry. */ 99879561Siedowse for (i = 0; i < entrypos; i += dp->d_reclen) { 99979561Siedowse dp = (struct direct *)(blkbuf + i); 100079561Siedowse if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos) 100179561Siedowse return (-1); /* Corrupted directory. */ 100279561Siedowse prevoff = blkoff + i; 100379561Siedowse } 100479561Siedowse return (prevoff); 100579561Siedowse} 100679561Siedowse 100779561Siedowse/* 100879561Siedowse * Try to free up `wanted' bytes by stealing memory from existing 100979561Siedowse * dirhashes. Returns zero with ufsdirhash_mtx locked if successful. 101079561Siedowse */ 101179561Siedowsestatic int 101279561Siedowseufsdirhash_recycle(int wanted) 101379561Siedowse{ 101479561Siedowse struct dirhash *dh; 101579561Siedowse doff_t **hash; 101679561Siedowse u_int8_t *blkfree; 101779561Siedowse int i, mem, narrays; 101879561Siedowse 101979561Siedowse mtx_lock(&ufsdirhash_mtx); 102079561Siedowse while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) { 102179561Siedowse /* Find a dirhash, and lock it. */ 102279561Siedowse if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) { 102379561Siedowse mtx_unlock(&ufsdirhash_mtx); 102479561Siedowse return (-1); 102579561Siedowse } 102679561Siedowse mtx_lock(&dh->dh_mtx); 102779561Siedowse KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list")); 102879561Siedowse 102979561Siedowse /* Decrement the score; only recycle if it becomes zero. */ 103079561Siedowse if (--dh->dh_score > 0) { 103179561Siedowse mtx_unlock(&dh->dh_mtx); 103279561Siedowse mtx_unlock(&ufsdirhash_mtx); 103379561Siedowse return (-1); 103479561Siedowse } 103579561Siedowse 103679561Siedowse /* Remove it from the list and detach its memory. */ 103779561Siedowse TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list); 103879561Siedowse dh->dh_onlist = 0; 103979561Siedowse hash = dh->dh_hash; 104079561Siedowse dh->dh_hash = NULL; 104179561Siedowse blkfree = dh->dh_blkfree; 104279561Siedowse dh->dh_blkfree = NULL; 104379561Siedowse narrays = dh->dh_narrays; 104479561Siedowse mem = narrays * sizeof(*dh->dh_hash) + 104579561Siedowse narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) + 104679561Siedowse dh->dh_nblk * sizeof(*dh->dh_blkfree); 104779561Siedowse 104879561Siedowse /* Unlock everything, free the detached memory. */ 104979561Siedowse mtx_unlock(&dh->dh_mtx); 105079561Siedowse mtx_unlock(&ufsdirhash_mtx); 105179561Siedowse for (i = 0; i < narrays; i++) 105292768Sjeff uma_zfree(ufsdirhash_zone, hash[i]); 105379561Siedowse FREE(hash, M_DIRHASH); 105479561Siedowse FREE(blkfree, M_DIRHASH); 105579561Siedowse 105679561Siedowse /* Account for the returned memory, and repeat if necessary. */ 105779561Siedowse mtx_lock(&ufsdirhash_mtx); 105879561Siedowse ufs_dirhashmem -= mem; 105979561Siedowse } 106079561Siedowse /* Success; return with ufsdirhash_mtx locked. */ 106179561Siedowse return (0); 106279561Siedowse} 106379561Siedowse 106479561Siedowse 106599101Siedowsevoid 106679561Siedowseufsdirhash_init() 106779561Siedowse{ 106896874Siedowse ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t), 106992768Sjeff NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 107093818Sjhb mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF); 107179561Siedowse TAILQ_INIT(&ufsdirhash_list); 107279561Siedowse} 107379561Siedowse 107499101Siedowsevoid 107599101Siedowseufsdirhash_uninit() 107699101Siedowse{ 107799101Siedowse KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit")); 107899101Siedowse uma_zdestroy(ufsdirhash_zone); 107999101Siedowse mtx_destroy(&ufsdirhash_mtx); 108099101Siedowse} 108179561Siedowse 108279561Siedowse#endif /* UFS_DIRHASH */ 1083