summaryrefslogtreecommitdiff
path: root/sys
diff options
context:
space:
mode:
authorrumble <rumble@NetBSD.org>2005-01-23 19:37:05 +0000
committerrumble <rumble@NetBSD.org>2005-01-23 19:37:05 +0000
commit32386a4e992ee104dcbd699f62a5afc9c87ce41e (patch)
treec3a7bcaddca443a9ac87ed8e2ae2a598209a97b1 /sys
parent5ed37446b2ec6d4fb06aecd4d594ce18c4579f8a (diff)
Bring in Ian Dowse's Dirhash from FreeBSD. Hash tables of
directories are created on the fly and used to increase performance by circumventing ufs_lookup's linear search. Dirhash is enabled by the UFS_DIRHASH option, but not by default.
Diffstat (limited to 'sys')
-rw-r--r--sys/ufs/files.ufs3
-rw-r--r--sys/ufs/ufs/Makefile5
-rw-r--r--sys/ufs/ufs/dirhash.h124
-rw-r--r--sys/ufs/ufs/inode.h4
-rw-r--r--sys/ufs/ufs/ufs_dirhash.c1121
-rw-r--r--sys/ufs/ufs/ufs_inode.c11
-rw-r--r--sys/ufs/ufs/ufs_lookup.c100
-rw-r--r--sys/ufs/ufs/ufs_vfsops.c13
-rw-r--r--sys/ufs/ufs/ufs_vnops.c11
9 files changed, 1380 insertions, 12 deletions
diff --git a/sys/ufs/files.ufs b/sys/ufs/files.ufs
index 47409074d78..9010483c4a5 100644
--- a/sys/ufs/files.ufs
+++ b/sys/ufs/files.ufs
@@ -1,4 +1,4 @@
-# $NetBSD: files.ufs,v 1.3 2004/05/25 14:54:58 hannken Exp $
+# $NetBSD: files.ufs,v 1.4 2005/01/23 19:37:05 rumble Exp $
deffs FFS
deffs EXT2FS
@@ -50,6 +50,7 @@ file ufs/mfs/mfs_vfsops.c mfs
file ufs/mfs/mfs_vnops.c mfs
file ufs/ufs/ufs_bmap.c ffs | lfs | mfs | ext2fs
+file ufs/ufs/ufs_dirhash.c ffs | lfs | mfs | ext2fs
file ufs/ufs/ufs_ihash.c ffs | lfs | mfs | ext2fs
file ufs/ufs/ufs_inode.c ffs | lfs | mfs
file ufs/ufs/ufs_lookup.c ffs | lfs | mfs | ext2fs
diff --git a/sys/ufs/ufs/Makefile b/sys/ufs/ufs/Makefile
index d8ed0ea130b..1b42c51af58 100644
--- a/sys/ufs/ufs/Makefile
+++ b/sys/ufs/ufs/Makefile
@@ -1,7 +1,8 @@
-# $NetBSD: Makefile,v 1.1 1998/06/12 23:23:12 cgd Exp $
+# $NetBSD: Makefile,v 1.2 2005/01/23 19:37:05 rumble Exp $
INCSDIR= /usr/include/ufs/ufs
-INCS= dinode.h dir.h inode.h quota.h ufs_bswap.h ufs_extern.h ufsmount.h
+INCS= dinode.h dir.h dirhash.h inode.h quota.h ufs_bswap.h ufs_extern.h \
+ ufsmount.h
.include <bsd.kinc.mk>
diff --git a/sys/ufs/ufs/dirhash.h b/sys/ufs/ufs/dirhash.h
new file mode 100644
index 00000000000..7d974243caa
--- /dev/null
+++ b/sys/ufs/ufs/dirhash.h
@@ -0,0 +1,124 @@
+/* $NetBSD: dirhash.h,v 1.1 2005/01/23 19:37:05 rumble Exp $ */
+
+/*
+ * Copyright (c) 2001 Ian Dowse. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ *
+ * $FreeBSD: src/sys/ufs/ufs/dirhash.h,v 1.2.2.2 2004/12/08 11:54:13 dwmalone Exp $
+ */
+
+#ifndef _UFS_UFS_DIRHASH_H_
+#define _UFS_UFS_DIRHASH_H_
+
+/*
+ * For fast operations on large directories, we maintain a hash
+ * that maps the file name to the offset of the directory entry within
+ * the directory file.
+ *
+ * The hashing uses a dumb spillover to the next free slot on
+ * collisions, so we must keep the utilisation low to avoid
+ * long linear searches. Deleted entries that are not the last
+ * in a chain must be marked DIRHASH_DEL.
+ *
+ * We also maintain information about free space in each block
+ * to speed up creations.
+ */
+#define DIRHASH_EMPTY (-1) /* entry unused */
+#define DIRHASH_DEL (-2) /* deleted entry; may be part of chain */
+
+#define DIRALIGN 4
+#define DH_NFSTATS (DIRECTSIZ(MAXNAMLEN + 1) / DIRALIGN)
+ /* max DIRALIGN words in a directory entry */
+
+/*
+ * Dirhash uses a score mechanism to achieve a hybrid between a
+ * least-recently-used and a least-often-used algorithm for entry
+ * recycling. The score is incremented when a directory is used, and
+ * decremented when the directory is a candidate for recycling. When
+ * the score reaches zero, the hash is recycled. Hashes are linked
+ * together on a TAILQ list, and hashes with higher scores filter
+ * towards the tail (most recently used) end of the list.
+ *
+ * New hash entries are given an inital score of DH_SCOREINIT and are
+ * placed at the most-recently-used end of the list. This helps a lot
+ * in the worst-case case scenario where every directory access is
+ * to a directory that is not hashed (i.e. the working set of hash
+ * candidates is much larger than the configured memry limit). In this
+ * case it limits the number of hash builds to 1/DH_SCOREINIT of the
+ * number of accesses.
+ */
+#define DH_SCOREINIT 8 /* initial dh_score when dirhash built */
+#define DH_SCOREMAX 64 /* max dh_score value */
+
+/*
+ * The main hash table has 2 levels. It is an array of pointers to
+ * blocks of DH_NBLKOFF offsets.
+ */
+#define DH_BLKOFFSHIFT 8
+#define DH_NBLKOFF (1 << DH_BLKOFFSHIFT)
+#define DH_BLKOFFMASK (DH_NBLKOFF - 1)
+
+#define DH_ENTRY(dh, slot) \
+ ((dh)->dh_hash[(slot) >> DH_BLKOFFSHIFT][(slot) & DH_BLKOFFMASK])
+
+struct dirhash {
+ doff_t **dh_hash; /* the hash array (2-level) */
+ int dh_narrays; /* number of entries in dh_hash */
+ int dh_hlen; /* total slots in the 2-level hash array */
+ int dh_hused; /* entries in use */
+
+ u_int8_t *dh_blkfree; /* free DIRALIGN words in each dir block */
+ int dh_nblk; /* size of dh_blkfree array */
+ int dh_dirblks; /* number of DIRBLKSIZ blocks in dir */
+ int dh_firstfree[DH_NFSTATS + 1]; /* first blk with N words free */
+
+ int dh_seqopt; /* sequential access optimisation enabled */
+ doff_t dh_seqoff; /* sequential access optimisation offset */
+
+ int dh_score; /* access count for this dirhash */
+
+ int dh_onlist; /* true if on the ufsdirhash_list chain */
+
+ TAILQ_ENTRY(dirhash) dh_list; /* chain of all dirhashes */
+};
+
+
+/*
+ * Dirhash functions.
+ */
+int ufsdirhash_build(struct inode *);
+doff_t ufsdirhash_findfree(struct inode *, int, int *);
+doff_t ufsdirhash_enduseful(struct inode *);
+int ufsdirhash_lookup(struct inode *, const char *, int, doff_t *,
+ struct buf **, doff_t *);
+void ufsdirhash_newblk(struct inode *, doff_t);
+void ufsdirhash_add(struct inode *, struct direct *, doff_t);
+void ufsdirhash_remove(struct inode *, struct direct *, doff_t);
+void ufsdirhash_move(struct inode *, struct direct *, doff_t, doff_t);
+void ufsdirhash_dirtrunc(struct inode *, doff_t);
+void ufsdirhash_free(struct inode *);
+void ufsdirhash_checkblock(struct inode *, char *, doff_t);
+void ufsdirhash_init(void);
+void ufsdirhash_done(void);
+
+#endif /* !_UFS_UFS_DIRHASH_H_ */
diff --git a/sys/ufs/ufs/inode.h b/sys/ufs/ufs/inode.h
index 30457860058..91f740ced2d 100644
--- a/sys/ufs/ufs/inode.h
+++ b/sys/ufs/ufs/inode.h
@@ -1,4 +1,4 @@
-/* $NetBSD: inode.h,v 1.39 2004/08/14 14:32:04 mycroft Exp $ */
+/* $NetBSD: inode.h,v 1.40 2005/01/23 19:37:05 rumble Exp $ */
/*
* Copyright (c) 1982, 1989, 1993
@@ -129,6 +129,8 @@ struct inode {
u_int32_t i_uid; /* File owner. */
u_int32_t i_gid; /* File group. */
+ struct dirhash *i_dirhash; /* Hashing for large directories */
+
/*
* The on-disk dinode itself.
*/
diff --git a/sys/ufs/ufs/ufs_dirhash.c b/sys/ufs/ufs/ufs_dirhash.c
new file mode 100644
index 00000000000..dc85f2a8b1e
--- /dev/null
+++ b/sys/ufs/ufs/ufs_dirhash.c
@@ -0,0 +1,1121 @@
+/* $NetBSD: ufs_dirhash.c,v 1.1 2005/01/23 19:37:05 rumble Exp $ */
+
+/*
+ * Copyright (c) 2001, 2002 Ian Dowse. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ *
+ * $FreeBSD: src/sys/ufs/ufs/ufs_dirhash.c,v 1.3.2.8 2004/12/08 11:54:13 dwmalone Exp $
+ */
+
+/*
+ * This implements a hash-based lookup scheme for UFS directories.
+ */
+
+#ifdef UFS_DIRHASH
+
+#include <sys/param.h>
+#include <sys/systm.h>
+#include <sys/kernel.h>
+#include <sys/malloc.h>
+#include <sys/types.h>
+#include <sys/hash.h>
+#include <sys/proc.h>
+#include <sys/buf.h>
+#include <sys/vnode.h>
+#include <sys/mount.h>
+#include <sys/pool.h>
+#include <sys/sysctl.h>
+
+#include <ufs/ufs/quota.h>
+#include <ufs/ufs/inode.h>
+#include <ufs/ufs/dir.h>
+#include <ufs/ufs/dirhash.h>
+#include <ufs/ufs/ufsmount.h>
+#include <ufs/ufs/ufs_bswap.h>
+#include <ufs/ufs/ufs_extern.h>
+
+#define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1))
+#define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1))
+#define OFSFMT(ip) ((ip)->i_ump->um_maxsymlinklen <= 0)
+#define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n))
+
+static MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables");
+
+static int ufs_dirhashminblks = 5;
+static int ufs_dirhashmaxmem = 2 * 1024 * 1024;
+static int ufs_dirhashmem;
+static int ufs_dirhashcheck = 0;
+
+static int ufsdirhash_hash(struct dirhash *dh, const char *name, int namelen);
+static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff,
+ int dirblksiz);
+static void ufsdirhash_delslot(struct dirhash *dh, int slot);
+static int ufsdirhash_findslot(struct dirhash *dh, const char *name,
+ int namelen, doff_t offset);
+static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset,
+ int dirblksiz);
+static int ufsdirhash_recycle(int wanted);
+
+POOL_INIT(ufsdirhash_pool, DH_NBLKOFF * sizeof(daddr_t), 0, 0, 0, "ufsdirhash",
+ &pool_allocator_nointr);
+
+#define DIRHASHLIST_LOCK() do { } while (0)
+#define DIRHASHLIST_UNLOCK() do { } while (0)
+#define DIRHASH_LOCK(dh) do { } while (0)
+#define DIRHASH_UNLOCK(dh) do { } while (0)
+#define DIRHASH_BLKALLOC_WAITOK() pool_get(&ufsdirhash_pool, PR_WAITOK)
+#define DIRHASH_BLKFREE(ptr) pool_put(&ufsdirhash_pool, ptr)
+
+/* Dirhash list; recently-used entries are near the tail. */
+static TAILQ_HEAD(, dirhash) ufsdirhash_list;
+
+/*
+ * Attempt to build up a hash table for the directory contents in
+ * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
+ */
+int
+ufsdirhash_build(struct inode *ip)
+{
+ struct dirhash *dh;
+ struct buf *bp = NULL;
+ struct direct *ep;
+ struct vnode *vp;
+ doff_t bmask, pos;
+ int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
+ const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ /* Check if we can/should use dirhash. */
+ if (ip->i_dirhash == NULL) {
+ if (ip->i_size < (ufs_dirhashminblks * dirblksiz) || OFSFMT(ip))
+ return (-1);
+ } else {
+ /* Hash exists, but sysctls could have changed. */
+ if (ip->i_size < (ufs_dirhashminblks * dirblksiz) ||
+ ufs_dirhashmem > ufs_dirhashmaxmem) {
+ ufsdirhash_free(ip);
+ return (-1);
+ }
+ /* Check if hash exists and is intact (note: unlocked read). */
+ if (ip->i_dirhash->dh_hash != NULL)
+ return (0);
+ /* Free the old, recycled hash and build a new one. */
+ ufsdirhash_free(ip);
+ }
+
+ /* Don't hash removed directories. */
+ if (ip->i_ffs_effnlink == 0)
+ return (-1);
+
+ vp = ip->i_vnode;
+ /* Allocate 50% more entries than this dir size could ever need. */
+ KASSERT(ip->i_size >= dirblksiz);
+ nslots = ip->i_size / DIRECTSIZ(1);
+ nslots = (nslots * 3 + 1) / 2;
+ narrays = howmany(nslots, DH_NBLKOFF);
+ nslots = narrays * DH_NBLKOFF;
+ dirblocks = howmany(ip->i_size, dirblksiz);
+ nblocks = (dirblocks * 3 + 1) / 2;
+
+ memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
+ narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
+ nblocks * sizeof(*dh->dh_blkfree);
+ DIRHASHLIST_LOCK();
+ if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
+ DIRHASHLIST_UNLOCK();
+ if (memreqd > ufs_dirhashmaxmem / 2)
+ return (-1);
+
+ /* Try to free some space. */
+ if (ufsdirhash_recycle(memreqd) != 0)
+ return (-1);
+ /* Enough was freed, and list has been locked. */
+ }
+ ufs_dirhashmem += memreqd;
+ DIRHASHLIST_UNLOCK();
+
+ /*
+ * Use non-blocking mallocs so that we will revert to a linear
+ * lookup on failure rather than potentially blocking forever.
+ */
+ MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO);
+ if (dh == NULL) {
+ DIRHASHLIST_LOCK();
+ ufs_dirhashmem -= memreqd;
+ DIRHASHLIST_UNLOCK();
+ return (-1);
+ }
+ MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]),
+ M_DIRHASH, M_NOWAIT | M_ZERO);
+ MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]),
+ M_DIRHASH, M_NOWAIT);
+ if (dh->dh_hash == NULL || dh->dh_blkfree == NULL)
+ goto fail;
+ for (i = 0; i < narrays; i++) {
+ if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL)
+ goto fail;
+ for (j = 0; j < DH_NBLKOFF; j++)
+ dh->dh_hash[i][j] = DIRHASH_EMPTY;
+ }
+
+ /* Initialise the hash table and block statistics. */
+ dh->dh_narrays = narrays;
+ dh->dh_hlen = nslots;
+ dh->dh_nblk = nblocks;
+ dh->dh_dirblks = dirblocks;
+ for (i = 0; i < dirblocks; i++)
+ dh->dh_blkfree[i] = dirblksiz / DIRALIGN;
+ for (i = 0; i < DH_NFSTATS; i++)
+ dh->dh_firstfree[i] = -1;
+ dh->dh_firstfree[DH_NFSTATS] = 0;
+ dh->dh_seqopt = 0;
+ dh->dh_seqoff = 0;
+ dh->dh_score = DH_SCOREINIT;
+ ip->i_dirhash = dh;
+
+ bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
+ pos = 0;
+ while (pos < ip->i_size) {
+ /* If necessary, get the next directory block. */
+ if ((pos & bmask) == 0) {
+ if (bp != NULL)
+ brelse(bp);
+ if (VOP_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
+ goto fail;
+ }
+
+ /* Add this entry to the hash. */
+ ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
+ if (ep->d_reclen == 0 || ep->d_reclen >
+ dirblksiz - (pos & (dirblksiz - 1))) {
+ /* Corrupted directory. */
+ brelse(bp);
+ goto fail;
+ }
+ if (ep->d_ino != 0) {
+ /* Add the entry (simplified ufsdirhash_add). */
+ slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
+ while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
+ slot = WRAPINCR(slot, dh->dh_hlen);
+ dh->dh_hused++;
+ DH_ENTRY(dh, slot) = pos;
+ ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep, needswap),
+ dirblksiz);
+ }
+ pos += ep->d_reclen;
+ }
+
+ if (bp != NULL)
+ brelse(bp);
+ DIRHASHLIST_LOCK();
+ TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
+ dh->dh_onlist = 1;
+ DIRHASHLIST_UNLOCK();
+ return (0);
+
+fail:
+ if (dh->dh_hash != NULL) {
+ for (i = 0; i < narrays; i++)
+ if (dh->dh_hash[i] != NULL)
+ DIRHASH_BLKFREE(dh->dh_hash[i]);
+ FREE(dh->dh_hash, M_DIRHASH);
+ }
+ if (dh->dh_blkfree != NULL)
+ FREE(dh->dh_blkfree, M_DIRHASH);
+ FREE(dh, M_DIRHASH);
+ ip->i_dirhash = NULL;
+ DIRHASHLIST_LOCK();
+ ufs_dirhashmem -= memreqd;
+ DIRHASHLIST_UNLOCK();
+ return (-1);
+}
+
+/*
+ * Free any hash table associated with inode 'ip'.
+ */
+void
+ufsdirhash_free(struct inode *ip)
+{
+ struct dirhash *dh;
+ int i, mem;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASHLIST_LOCK();
+ DIRHASH_LOCK(dh);
+ if (dh->dh_onlist)
+ TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
+ DIRHASH_UNLOCK(dh);
+ DIRHASHLIST_UNLOCK();
+
+ /* The dirhash pointed to by 'dh' is exclusively ours now. */
+
+ mem = sizeof(*dh);
+ if (dh->dh_hash != NULL) {
+ for (i = 0; i < dh->dh_narrays; i++)
+ DIRHASH_BLKFREE(dh->dh_hash[i]);
+ FREE(dh->dh_hash, M_DIRHASH);
+ FREE(dh->dh_blkfree, M_DIRHASH);
+ mem += dh->dh_narrays * sizeof(*dh->dh_hash) +
+ dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
+ dh->dh_nblk * sizeof(*dh->dh_blkfree);
+ }
+ FREE(dh, M_DIRHASH);
+ ip->i_dirhash = NULL;
+
+ DIRHASHLIST_LOCK();
+ ufs_dirhashmem -= mem;
+ DIRHASHLIST_UNLOCK();
+}
+
+/*
+ * Find the offset of the specified name within the given inode.
+ * Returns 0 on success, ENOENT if the entry does not exist, or
+ * EJUSTRETURN if the caller should revert to a linear search.
+ *
+ * If successful, the directory offset is stored in *offp, and a
+ * pointer to a struct buf containing the entry is stored in *bpp. If
+ * prevoffp is non-NULL, the offset of the previous entry within
+ * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
+ * is the first in a block, the start of the block is used).
+ */
+int
+ufsdirhash_lookup(struct inode *ip, const char *name, int namelen, doff_t *offp,
+ struct buf **bpp, doff_t *prevoffp)
+{
+ struct dirhash *dh, *dh_next;
+ struct direct *dp;
+ struct vnode *vp;
+ struct buf *bp;
+ doff_t blkoff, bmask, offset, prevoff;
+ int i, slot;
+ const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return (EJUSTRETURN);
+ /*
+ * Move this dirhash towards the end of the list if it has a
+ * score higher than the next entry, and acquire the dh_mtx.
+ * Optimise the case where it's already the last by performing
+ * an unlocked read of the TAILQ_NEXT pointer.
+ *
+ * In both cases, end up holding just dh_mtx.
+ */
+ if (TAILQ_NEXT(dh, dh_list) != NULL) {
+ DIRHASHLIST_LOCK();
+ DIRHASH_LOCK(dh);
+ /*
+ * If the new score will be greater than that of the next
+ * entry, then move this entry past it. With both mutexes
+ * held, dh_next won't go away, but its dh_score could
+ * change; that's not important since it is just a hint.
+ */
+ if (dh->dh_hash != NULL &&
+ (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
+ dh->dh_score >= dh_next->dh_score) {
+ KASSERT(dh->dh_onlist);
+ TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
+ TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
+ dh_list);
+ }
+ DIRHASHLIST_UNLOCK();
+ } else {
+ /* Already the last, though that could change as we wait. */
+ DIRHASH_LOCK(dh);
+ }
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return (EJUSTRETURN);
+ }
+
+ /* Update the score. */
+ if (dh->dh_score < DH_SCOREMAX)
+ dh->dh_score++;
+
+ vp = ip->i_vnode;
+ bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
+ blkoff = -1;
+ bp = NULL;
+restart:
+ slot = ufsdirhash_hash(dh, name, namelen);
+
+ if (dh->dh_seqopt) {
+ /*
+ * Sequential access optimisation. dh_seqoff contains the
+ * offset of the directory entry immediately following
+ * the last entry that was looked up. Check if this offset
+ * appears in the hash chain for the name we are looking for.
+ */
+ for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
+ i = WRAPINCR(i, dh->dh_hlen))
+ if (offset == dh->dh_seqoff)
+ break;
+ if (offset == dh->dh_seqoff) {
+ /*
+ * We found an entry with the expected offset. This
+ * is probably the entry we want, but if not, the
+ * code below will turn off seqoff and retry.
+ */
+ slot = i;
+ } else
+ dh->dh_seqopt = 0;
+ }
+
+ for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
+ slot = WRAPINCR(slot, dh->dh_hlen)) {
+ if (offset == DIRHASH_DEL)
+ continue;
+ DIRHASH_UNLOCK(dh);
+
+ if (offset < 0 || offset >= ip->i_size)
+ panic("ufsdirhash_lookup: bad offset in hash array");
+ if ((offset & ~bmask) != blkoff) {
+ if (bp != NULL)
+ brelse(bp);
+ blkoff = offset & ~bmask;
+ if (VOP_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0)
+ return (EJUSTRETURN);
+ }
+ dp = (struct direct *)(bp->b_data + (offset & bmask));
+ if (dp->d_reclen == 0 || dp->d_reclen >
+ dirblksiz - (offset & (dirblksiz - 1))) {
+ /* Corrupted directory. */
+ brelse(bp);
+ return (EJUSTRETURN);
+ }
+ if (dp->d_namlen == namelen &&
+ bcmp(dp->d_name, name, namelen) == 0) {
+ /* Found. Get the prev offset if needed. */
+ if (prevoffp != NULL) {
+ if (offset & (dirblksiz - 1)) {
+ prevoff = ufsdirhash_getprev(dp,
+ offset, dirblksiz);
+ if (prevoff == -1) {
+ brelse(bp);
+ return (EJUSTRETURN);
+ }
+ } else
+ prevoff = offset;
+ *prevoffp = prevoff;
+ }
+
+ /* Check for sequential access, and update offset. */
+ if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset)
+ dh->dh_seqopt = 1;
+ dh->dh_seqoff = offset + DIRSIZ(0, dp, needswap);
+
+ *bpp = bp;
+ *offp = offset;
+ return (0);
+ }
+
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ if (bp != NULL)
+ brelse(bp);
+ ufsdirhash_free(ip);
+ return (EJUSTRETURN);
+ }
+ /*
+ * When the name doesn't match in the seqopt case, go back
+ * and search normally.
+ */
+ if (dh->dh_seqopt) {
+ dh->dh_seqopt = 0;
+ goto restart;
+ }
+ }
+ DIRHASH_UNLOCK(dh);
+ if (bp != NULL)
+ brelse(bp);
+ return (ENOENT);
+}
+
+/*
+ * Find a directory block with room for 'slotneeded' bytes. Returns
+ * the offset of the directory entry that begins the free space.
+ * This will either be the offset of an existing entry that has free
+ * space at the end, or the offset of an entry with d_ino == 0 at
+ * the start of a DIRBLKSIZ block.
+ *
+ * To use the space, the caller may need to compact existing entries in
+ * the directory. The total number of bytes in all of the entries involved
+ * in the compaction is stored in *slotsize. In other words, all of
+ * the entries that must be compacted are exactly contained in the
+ * region beginning at the returned offset and spanning *slotsize bytes.
+ *
+ * Returns -1 if no space was found, indicating that the directory
+ * must be extended.
+ */
+doff_t
+ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
+{
+ struct direct *dp;
+ struct dirhash *dh;
+ struct buf *bp;
+ doff_t pos, slotstart;
+ int dirblock, error, freebytes, i;
+ const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return (-1);
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return (-1);
+ }
+
+ /* Find a directory block with the desired free space. */
+ dirblock = -1;
+ for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
+ if ((dirblock = dh->dh_firstfree[i]) != -1)
+ break;
+ if (dirblock == -1) {
+ DIRHASH_UNLOCK(dh);
+ return (-1);
+ }
+
+ KASSERT(dirblock < dh->dh_nblk &&
+ dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN));
+ DIRHASH_UNLOCK(dh);
+ pos = dirblock * dirblksiz;
+ error = VOP_BLKATOFF(ip->i_vnode, (off_t)pos, (void *)&dp, &bp);
+ if (error)
+ return (-1);
+ /* Find the first entry with free space. */
+ for (i = 0; i < dirblksiz; ) {
+ if (dp->d_reclen == 0) {
+ brelse(bp);
+ return (-1);
+ }
+ if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp, needswap))
+ break;
+ i += dp->d_reclen;
+ dp = (struct direct *)((char *)dp + dp->d_reclen);
+ }
+ if (i > dirblksiz) {
+ brelse(bp);
+ return (-1);
+ }
+ slotstart = pos + i;
+
+ /* Find the range of entries needed to get enough space */
+ freebytes = 0;
+ while (i < dirblksiz && freebytes < slotneeded) {
+ freebytes += dp->d_reclen;
+ if (dp->d_ino != 0)
+ freebytes -= DIRSIZ(0, dp, needswap);
+ if (dp->d_reclen == 0) {
+ brelse(bp);
+ return (-1);
+ }
+ i += dp->d_reclen;
+ dp = (struct direct *)((char *)dp + dp->d_reclen);
+ }
+ if (i > dirblksiz) {
+ brelse(bp);
+ return (-1);
+ }
+ if (freebytes < slotneeded)
+ panic("ufsdirhash_findfree: free mismatch");
+ brelse(bp);
+ *slotsize = pos + i - slotstart;
+ return (slotstart);
+}
+
+/*
+ * Return the start of the unused space at the end of a directory, or
+ * -1 if there are no trailing unused blocks.
+ */
+doff_t
+ufsdirhash_enduseful(struct inode *ip)
+{
+ struct dirhash *dh;
+ int i;
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return (-1);
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return (-1);
+ }
+
+ if (dh->dh_blkfree[dh->dh_dirblks - 1] != dirblksiz / DIRALIGN) {
+ DIRHASH_UNLOCK(dh);
+ return (-1);
+ }
+
+ for (i = dh->dh_dirblks - 1; i >= 0; i--)
+ if (dh->dh_blkfree[i] != dirblksiz / DIRALIGN)
+ break;
+ DIRHASH_UNLOCK(dh);
+ return ((doff_t)(i + 1) * dirblksiz);
+}
+
+/*
+ * Insert information into the hash about a new directory entry. dirp
+ * points to a struct direct containing the entry, and offset specifies
+ * the offset of this entry.
+ */
+void
+ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
+{
+ struct dirhash *dh;
+ int slot;
+ const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ KASSERT(offset < dh->dh_dirblks * dirblksiz);
+ /*
+ * Normal hash usage is < 66%. If the usage gets too high then
+ * remove the hash entirely and let it be rebuilt later.
+ */
+ if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ /* Find a free hash slot (empty or deleted), and add the entry. */
+ slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
+ while (DH_ENTRY(dh, slot) >= 0)
+ slot = WRAPINCR(slot, dh->dh_hlen);
+ if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
+ dh->dh_hused++;
+ DH_ENTRY(dh, slot) = offset;
+
+ /* Update the per-block summary info. */
+ ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp, needswap), dirblksiz);
+ DIRHASH_UNLOCK(dh);
+}
+
+/*
+ * Remove the specified directory entry from the hash. The entry to remove
+ * is defined by the name in `dirp', which must exist at the specified
+ * `offset' within the directory.
+ */
+void
+ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
+{
+ struct dirhash *dh;
+ int slot;
+ const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ KASSERT(offset < dh->dh_dirblks * dirblksiz);
+ /* Find the entry */
+ slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
+
+ /* Remove the hash entry. */
+ ufsdirhash_delslot(dh, slot);
+
+ /* Update the per-block summary info. */
+ ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp, needswap), dirblksiz);
+ DIRHASH_UNLOCK(dh);
+}
+
+/*
+ * Change the offset associated with a directory entry in the hash. Used
+ * when compacting directory blocks.
+ */
+void
+ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
+ doff_t newoff)
+{
+ struct dirhash *dh;
+ int slot;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ KASSERT(oldoff < dh->dh_dirblks * ip->i_ump->um_dirblksiz &&
+ newoff < dh->dh_dirblks * ip->i_ump->um_dirblksiz);
+ /* Find the entry, and update the offset. */
+ slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
+ DH_ENTRY(dh, slot) = newoff;
+ DIRHASH_UNLOCK(dh);
+}
+
+/*
+ * Inform dirhash that the directory has grown by one block that
+ * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
+ */
+void
+ufsdirhash_newblk(struct inode *ip, doff_t offset)
+{
+ struct dirhash *dh;
+ int block;
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ KASSERT(offset == dh->dh_dirblks * dirblksiz);
+ block = offset / dirblksiz;
+ if (block >= dh->dh_nblk) {
+ /* Out of space; must rebuild. */
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+ dh->dh_dirblks = block + 1;
+
+ /* Account for the new free block. */
+ dh->dh_blkfree[block] = dirblksiz / DIRALIGN;
+ if (dh->dh_firstfree[DH_NFSTATS] == -1)
+ dh->dh_firstfree[DH_NFSTATS] = block;
+ DIRHASH_UNLOCK(dh);
+}
+
+/*
+ * Inform dirhash that the directory is being truncated.
+ */
+void
+ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
+{
+ struct dirhash *dh;
+ int block, i;
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ KASSERT(offset <= dh->dh_dirblks * dirblksiz);
+ block = howmany(offset, dirblksiz);
+ /*
+ * If the directory shrinks to less than 1/8 of dh_nblk blocks
+ * (about 20% of its original size due to the 50% extra added in
+ * ufsdirhash_build) then free it, and let the caller rebuild
+ * if necessary.
+ */
+ if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ /*
+ * Remove any `first free' information pertaining to the
+ * truncated blocks. All blocks we're removing should be
+ * completely unused.
+ */
+ if (dh->dh_firstfree[DH_NFSTATS] >= block)
+ dh->dh_firstfree[DH_NFSTATS] = -1;
+ for (i = block; i < dh->dh_dirblks; i++)
+ if (dh->dh_blkfree[i] != dirblksiz / DIRALIGN)
+ panic("ufsdirhash_dirtrunc: blocks in use");
+ for (i = 0; i < DH_NFSTATS; i++)
+ if (dh->dh_firstfree[i] >= block)
+ panic("ufsdirhash_dirtrunc: first free corrupt");
+ dh->dh_dirblks = block;
+ DIRHASH_UNLOCK(dh);
+}
+
+/*
+ * Debugging function to check that the dirhash information about
+ * a directory block matches its actual contents. Panics if a mismatch
+ * is detected.
+ *
+ * On entry, `buf' should point to the start of an in-core
+ * DIRBLKSIZ-sized directory block, and `offset' should contain the
+ * offset from the start of the directory of that block.
+ */
+void
+ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
+{
+ struct dirhash *dh;
+ struct direct *dp;
+ int block, ffslot, i, nfree;
+ const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+
+ if (!ufs_dirhashcheck)
+ return;
+ if ((dh = ip->i_dirhash) == NULL)
+ return;
+ DIRHASH_LOCK(dh);
+ if (dh->dh_hash == NULL) {
+ DIRHASH_UNLOCK(dh);
+ ufsdirhash_free(ip);
+ return;
+ }
+
+ block = offset / dirblksiz;
+ if ((offset & (dirblksiz - 1)) != 0 || block >= dh->dh_dirblks)
+ panic("ufsdirhash_checkblock: bad offset");
+
+ nfree = 0;
+ for (i = 0; i < dirblksiz; i += dp->d_reclen) {
+ dp = (struct direct *)(buf + i);
+ if (dp->d_reclen == 0 || i + dp->d_reclen > dirblksiz)
+ panic("ufsdirhash_checkblock: bad dir");
+
+ if (dp->d_ino == 0) {
+#if 0
+ /*
+ * XXX entries with d_ino == 0 should only occur
+ * at the start of a DIRBLKSIZ block. However the
+ * ufs code is tolerant of such entries at other
+ * offsets, and fsck does not fix them.
+ */
+ if (i != 0)
+ panic("ufsdirhash_checkblock: bad dir inode");
+#endif
+ nfree += dp->d_reclen;
+ continue;
+ }
+
+ /* Check that the entry exists (will panic if it doesn't). */
+ ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
+
+ nfree += dp->d_reclen - DIRSIZ(0, dp, needswap);
+ }
+ if (i != dirblksiz)
+ panic("ufsdirhash_checkblock: bad dir end");
+
+ if (dh->dh_blkfree[block] * DIRALIGN != nfree)
+ panic("ufsdirhash_checkblock: bad free count");
+
+ ffslot = BLKFREE2IDX(nfree / DIRALIGN);
+ for (i = 0; i <= DH_NFSTATS; i++)
+ if (dh->dh_firstfree[i] == block && i != ffslot)
+ panic("ufsdirhash_checkblock: bad first-free");
+ if (dh->dh_firstfree[ffslot] == -1)
+ panic("ufsdirhash_checkblock: missing first-free entry");
+ DIRHASH_UNLOCK(dh);
+}
+
+/*
+ * Hash the specified filename into a dirhash slot.
+ */
+static int
+ufsdirhash_hash(struct dirhash *dh, const char *name, int namelen)
+{
+ u_int32_t hash;
+
+ /*
+ * We hash the name and then some other bit of data that is
+ * invariant over the dirhash's lifetime. Otherwise names
+ * differing only in the last byte are placed close to one
+ * another in the table, which is bad for linear probing.
+ */
+ hash = hash32_buf(name, namelen, HASH32_BUF_INIT);
+ hash = hash32_buf(&dh, sizeof(dh), hash);
+ return (hash % dh->dh_hlen);
+}
+
+/*
+ * Adjust the number of free bytes in the block containing `offset'
+ * by the value specified by `diff'.
+ *
+ * The caller must ensure we have exclusive access to `dh'; normally
+ * that means that dh_mtx should be held, but this is also called
+ * from ufsdirhash_build() where exclusive access can be assumed.
+ */
+static void
+ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff, int dirblksiz)
+{
+ int block, i, nfidx, ofidx;
+
+ /* Update the per-block summary info. */
+ block = offset / dirblksiz;
+ KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks);
+ ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
+ dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
+ nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
+
+ /* Update the `first free' list if necessary. */
+ if (ofidx != nfidx) {
+ /* If removing, scan forward for the next block. */
+ if (dh->dh_firstfree[ofidx] == block) {
+ for (i = block + 1; i < dh->dh_dirblks; i++)
+ if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
+ break;
+ dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
+ }
+
+ /* Make this the new `first free' if necessary */
+ if (dh->dh_firstfree[nfidx] > block ||
+ dh->dh_firstfree[nfidx] == -1)
+ dh->dh_firstfree[nfidx] = block;
+ }
+}
+
+/*
+ * Find the specified name which should have the specified offset.
+ * Returns a slot number, and panics on failure.
+ *
+ * `dh' must be locked on entry and remains so on return.
+ */
+static int
+ufsdirhash_findslot(struct dirhash *dh, const char *name, int namelen,
+ doff_t offset)
+{
+ int slot;
+
+ /* Find the entry. */
+ KASSERT(dh->dh_hused < dh->dh_hlen);
+ slot = ufsdirhash_hash(dh, name, namelen);
+ while (DH_ENTRY(dh, slot) != offset &&
+ DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
+ slot = WRAPINCR(slot, dh->dh_hlen);
+ if (DH_ENTRY(dh, slot) != offset)
+ panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
+
+ return (slot);
+}
+
+/*
+ * Remove the entry corresponding to the specified slot from the hash array.
+ *
+ * `dh' must be locked on entry and remains so on return.
+ */
+static void
+ufsdirhash_delslot(struct dirhash *dh, int slot)
+{
+ int i;
+
+ /* Mark the entry as deleted. */
+ DH_ENTRY(dh, slot) = DIRHASH_DEL;
+
+ /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
+ for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
+ i = WRAPINCR(i, dh->dh_hlen);
+ if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
+ i = WRAPDECR(i, dh->dh_hlen);
+ while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
+ DH_ENTRY(dh, i) = DIRHASH_EMPTY;
+ dh->dh_hused--;
+ i = WRAPDECR(i, dh->dh_hlen);
+ }
+ KASSERT(dh->dh_hused >= 0);
+ }
+}
+
+/*
+ * Given a directory entry and its offset, find the offset of the
+ * previous entry in the same DIRBLKSIZ-sized block. Returns an
+ * offset, or -1 if there is no previous entry in the block or some
+ * other problem occurred.
+ */
+static doff_t
+ufsdirhash_getprev(struct direct *dirp, doff_t offset, int dirblksiz)
+{
+ struct direct *dp;
+ char *blkbuf;
+ doff_t blkoff, prevoff;
+ int entrypos, i;
+
+ blkoff = offset & ~(dirblksiz - 1); /* offset of start of block */
+ entrypos = offset & (dirblksiz - 1); /* entry relative to block */
+ blkbuf = (char *)dirp - entrypos;
+ prevoff = blkoff;
+
+ /* If `offset' is the start of a block, there is no previous entry. */
+ if (entrypos == 0)
+ return (-1);
+
+ /* Scan from the start of the block until we get to the entry. */
+ for (i = 0; i < entrypos; i += dp->d_reclen) {
+ dp = (struct direct *)(blkbuf + i);
+ if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
+ return (-1); /* Corrupted directory. */
+ prevoff = blkoff + i;
+ }
+ return (prevoff);
+}
+
+/*
+ * Try to free up `wanted' bytes by stealing memory from existing
+ * dirhashes. Returns zero with list locked if successful.
+ */
+static int
+ufsdirhash_recycle(int wanted)
+{
+ struct dirhash *dh;
+ doff_t **hash;
+ u_int8_t *blkfree;
+ int i, mem, narrays;
+
+ DIRHASHLIST_LOCK();
+ while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
+ /* Find a dirhash, and lock it. */
+ if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) {
+ DIRHASHLIST_UNLOCK();
+ return (-1);
+ }
+ DIRHASH_LOCK(dh);
+ KASSERT(dh->dh_hash != NULL);
+
+ /* Decrement the score; only recycle if it becomes zero. */
+ if (--dh->dh_score > 0) {
+ DIRHASH_UNLOCK(dh);
+ DIRHASHLIST_UNLOCK();
+ return (-1);
+ }
+
+ /* Remove it from the list and detach its memory. */
+ TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
+ dh->dh_onlist = 0;
+ hash = dh->dh_hash;
+ dh->dh_hash = NULL;
+ blkfree = dh->dh_blkfree;
+ dh->dh_blkfree = NULL;
+ narrays = dh->dh_narrays;
+ mem = narrays * sizeof(*dh->dh_hash) +
+ narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
+ dh->dh_nblk * sizeof(*dh->dh_blkfree);
+
+ /* Unlock everything, free the detached memory. */
+ DIRHASH_UNLOCK(dh);
+ DIRHASHLIST_UNLOCK();
+ for (i = 0; i < narrays; i++)
+ DIRHASH_BLKFREE(hash[i]);
+ FREE(hash, M_DIRHASH);
+ FREE(blkfree, M_DIRHASH);
+
+ /* Account for the returned memory, and repeat if necessary. */
+ DIRHASHLIST_LOCK();
+ ufs_dirhashmem -= mem;
+ }
+ /* Success; return with list locked. */
+ return (0);
+}
+
+void
+ufsdirhash_init()
+{
+#ifdef _LKM
+ pool_init(&ufsdirhash_pool, DH_NBLKOFF * sizeof(daddr_t), 0, 0, 0,
+ "ufsdirhash", &pool_allocator_nointr);
+#endif
+ TAILQ_INIT(&ufsdirhash_list);
+}
+
+void
+ufsdirhash_done()
+{
+ KASSERT(TAILQ_EMPTY(&ufsdirhash_list));
+#ifdef _LKM
+ pool_destroy(&ufsdirhash_pool);
+#endif
+}
+
+SYSCTL_SETUP(sysctl_vfs_ufs_setup, "sysctl vfs.ufs.dirhash subtree setup")
+{
+ struct sysctlnode *rnode, *cnode;
+
+ sysctl_createv(clog, 0, NULL, &rnode,
+ CTLFLAG_PERMANENT,
+ CTLTYPE_NODE, "vfs", NULL,
+ NULL, 0, NULL, 0,
+ CTL_VFS, CTL_EOL);
+
+ sysctl_createv(clog, 0, &rnode, &rnode,
+ CTLFLAG_PERMANENT,
+ CTLTYPE_NODE, "ufs",
+ SYSCTL_DESCR("ufs"),
+ NULL, 0, NULL, 0,
+ CTL_CREATE, CTL_EOL);
+
+ sysctl_createv(clog, 0, &rnode, &rnode,
+ CTLFLAG_PERMANENT,
+ CTLTYPE_NODE, "dirhash",
+ SYSCTL_DESCR("dirhash"),
+ NULL, 0, NULL, 0,
+ CTL_CREATE, CTL_EOL);
+
+ sysctl_createv(clog, 0, &rnode, &cnode,
+ CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
+ CTLTYPE_INT, "minblocks",
+ SYSCTL_DESCR("minimum hashed directory size in blocks"),
+ NULL, 0, &ufs_dirhashminblks, 0,
+ CTL_CREATE, CTL_EOL);
+
+ sysctl_createv(clog, 0, &rnode, &cnode,
+ CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
+ CTLTYPE_INT, "maxmem",
+ SYSCTL_DESCR("maximum dirhash memory usage"),
+ NULL, 0, &ufs_dirhashmaxmem, 0,
+ CTL_CREATE, CTL_EOL);
+
+ sysctl_createv(clog, 0, &rnode, &cnode,
+ CTLFLAG_PERMANENT|CTLFLAG_READONLY,
+ CTLTYPE_INT, "memused",
+ SYSCTL_DESCR("current dirhash memory usage"),
+ NULL, 0, &ufs_dirhashmem, 0,
+ CTL_CREATE, CTL_EOL);
+
+ sysctl_createv(clog, 0, &rnode, &cnode,
+ CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
+ CTLTYPE_INT, "docheck",
+ SYSCTL_DESCR("enable extra sanity checks"),
+ NULL, 0, &ufs_dirhashcheck, 0,
+ CTL_CREATE, CTL_EOL);
+}
+
+#endif /* UFS_DIRHASH */
diff --git a/sys/ufs/ufs/ufs_inode.c b/sys/ufs/ufs/ufs_inode.c
index 711a48628ae..9ab8150d16f 100644
--- a/sys/ufs/ufs/ufs_inode.c
+++ b/sys/ufs/ufs/ufs_inode.c
@@ -1,4 +1,4 @@
-/* $NetBSD: ufs_inode.c,v 1.46 2004/12/20 03:12:20 dbj Exp $ */
+/* $NetBSD: ufs_inode.c,v 1.47 2005/01/23 19:37:05 rumble Exp $ */
/*
* Copyright (c) 1991, 1993
@@ -37,7 +37,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: ufs_inode.c,v 1.46 2004/12/20 03:12:20 dbj Exp $");
+__KERNEL_RCSID(0, "$NetBSD: ufs_inode.c,v 1.47 2005/01/23 19:37:05 rumble Exp $");
#if defined(_KERNEL_OPT)
#include "opt_quota.h"
@@ -55,6 +55,9 @@ __KERNEL_RCSID(0, "$NetBSD: ufs_inode.c,v 1.46 2004/12/20 03:12:20 dbj Exp $");
#include <ufs/ufs/inode.h>
#include <ufs/ufs/ufsmount.h>
#include <ufs/ufs/ufs_extern.h>
+#ifdef UFS_DIRHASH
+#include <ufs/ufs/dirhash.h>
+#endif
#include <uvm/uvm.h>
@@ -173,6 +176,10 @@ ufs_reclaim(vp, p)
}
}
#endif
+#ifdef UFS_DIRHASH
+ if (ip->i_dirhash != NULL)
+ ufsdirhash_free(ip);
+#endif
return (0);
}
diff --git a/sys/ufs/ufs/ufs_lookup.c b/sys/ufs/ufs/ufs_lookup.c
index 65440dca55d..db072ed280c 100644
--- a/sys/ufs/ufs/ufs_lookup.c
+++ b/sys/ufs/ufs/ufs_lookup.c
@@ -1,4 +1,4 @@
-/* $NetBSD: ufs_lookup.c,v 1.60 2004/09/17 14:11:27 skrll Exp $ */
+/* $NetBSD: ufs_lookup.c,v 1.61 2005/01/23 19:37:05 rumble Exp $ */
/*
* Copyright (c) 1989, 1993
@@ -37,7 +37,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: ufs_lookup.c,v 1.60 2004/09/17 14:11:27 skrll Exp $");
+__KERNEL_RCSID(0, "$NetBSD: ufs_lookup.c,v 1.61 2005/01/23 19:37:05 rumble Exp $");
#include <sys/param.h>
#include <sys/systm.h>
@@ -51,6 +51,9 @@ __KERNEL_RCSID(0, "$NetBSD: ufs_lookup.c,v 1.60 2004/09/17 14:11:27 skrll Exp $"
#include <ufs/ufs/inode.h>
#include <ufs/ufs/dir.h>
+#ifdef UFS_DIRHASH
+#include <ufs/ufs/dirhash.h>
+#endif
#include <ufs/ufs/ufsmount.h>
#include <ufs/ufs/ufs_extern.h>
#include <ufs/ufs/ufs_bswap.h>
@@ -143,6 +146,7 @@ ufs_lookup(v)
*vpp = NULL;
lockparent = flags & LOCKPARENT;
wantparent = flags & (LOCKPARENT|WANTPARENT);
+ endsearch = 0; /* silence compiler warning */
/*
* Check accessiblity of directory.
*/
@@ -189,6 +193,46 @@ ufs_lookup(v)
* of simplicity.
*/
bmask = vdp->v_mount->mnt_stat.f_iosize - 1;
+
+#ifdef UFS_DIRHASH
+ /*
+ * Use dirhash for fast operations on large directories. The logic
+ * to determine whether to hash the directory is contained within
+ * ufsdirhash_build(); a zero return means that it decided to hash
+ * this directory and it successfully built up the hash table.
+ */
+ if (ufsdirhash_build(dp) == 0) {
+ /* Look for a free slot if needed. */
+ enduseful = dp->i_size;
+ if (slotstatus != FOUND) {
+ slotoffset = ufsdirhash_findfree(dp, slotneeded,
+ &slotsize);
+ if (slotoffset >= 0) {
+ slotstatus = COMPACT;
+ enduseful = ufsdirhash_enduseful(dp);
+ if (enduseful < 0)
+ enduseful = dp->i_size;
+ }
+ }
+ /* Look up the component. */
+ numdirpasses = 1;
+ entryoffsetinblock = 0; /* silence compiler warning */
+ switch (ufsdirhash_lookup(dp, cnp->cn_nameptr, cnp->cn_namelen,
+ &dp->i_offset, &bp, nameiop == DELETE ? &prevoff : NULL)) {
+ case 0:
+ ep = (struct direct *)((char *)bp->b_data +
+ (dp->i_offset & bmask));
+ goto foundentry;
+ case ENOENT:
+ dp->i_offset = roundup(dp->i_size, dirblksiz);
+ goto notfound;
+ default:
+ /* Something failed; just do a linear search. */
+ break;
+ }
+ }
+#endif /* UFS_DIRHASH */
+
if (nameiop != LOOKUP || dp->i_diroff == 0 ||
dp->i_diroff >= dp->i_size) {
entryoffsetinblock = 0;
@@ -300,6 +344,9 @@ searchloop:
if (namlen == cnp->cn_namelen &&
!memcmp(cnp->cn_nameptr, ep->d_name,
(unsigned)namlen)) {
+#ifdef UFS_DIRHASH
+foundentry:
+#endif
/*
* Save directory entry's inode number and
* reclen in ndp->ni_ufs area, and release
@@ -776,6 +823,14 @@ ufs_direnter(dvp, tvp, dirp, cnp, newdirbp)
blkoff = dp->i_offset & (ump->um_mountp->mnt_stat.f_iosize - 1);
memcpy((caddr_t)bp->b_data + blkoff, (caddr_t)dirp,
newentrysize);
+#ifdef UFS_DIRHASH
+ if (dp->i_dirhash != NULL) {
+ ufsdirhash_newblk(dp, dp->i_offset);
+ ufsdirhash_add(dp, dirp, dp->i_offset);
+ ufsdirhash_checkblock(dp, (char *)bp->b_data + blkoff,
+ dp->i_offset);
+ }
+#endif
if (DOINGSOFTDEP(dvp)) {
/*
* Ensure that the entire newly allocated block is a
@@ -876,6 +931,12 @@ ufs_direnter(dvp, tvp, dirp, cnp, newdirbp)
dsize = DIRSIZ(FSFMT(dvp), nep, needswap);
spacefree += ufs_rw16(nep->d_reclen, needswap) - dsize;
loc += ufs_rw16(nep->d_reclen, needswap);
+#ifdef UFS_DIRHASH
+ if (dp->i_dirhash != NULL)
+ ufsdirhash_move(dp, nep,
+ dp->i_offset + ((char *)nep - dirbuf),
+ dp->i_offset + ((char *)ep - dirbuf));
+#endif
if (DOINGSOFTDEP(dvp))
softdep_change_directoryentry_offset(dp, dirbuf,
(caddr_t)nep, (caddr_t)ep, dsize);
@@ -912,7 +973,18 @@ ufs_direnter(dvp, tvp, dirp, cnp, newdirbp)
dirp->d_type = tmp;
}
}
+#ifdef UFS_DIRHASH
+ if (dp->i_dirhash != NULL && (ep->d_ino == 0 ||
+ dirp->d_reclen == spacefree))
+ ufsdirhash_add(dp, dirp, dp->i_offset + ((char *)ep - dirbuf));
+#endif
memcpy((caddr_t)ep, (caddr_t)dirp, (u_int)newentrysize);
+#ifdef UFS_DIRHASH
+ if (dp->i_dirhash != NULL)
+ ufsdirhash_checkblock(dp, dirbuf -
+ (dp->i_offset & (dirblksiz - 1)),
+ dp->i_offset & ~(dirblksiz - 1));
+#endif
if (DOINGSOFTDEP(dvp)) {
softdep_setup_directory_add(bp, dp,
dp->i_offset + (caddr_t)ep - dirbuf,
@@ -932,6 +1004,10 @@ ufs_direnter(dvp, tvp, dirp, cnp, newdirbp)
if (error == 0 && dp->i_endoff && dp->i_endoff < dp->i_size) {
if (DOINGSOFTDEP(dvp) && (tvp != NULL))
VOP_UNLOCK(tvp, 0);
+#ifdef UFS_DIRHASH
+ if (dp->i_dirhash != NULL)
+ ufsdirhash_dirtrunc(dp, dp->i_endoff);
+#endif
(void) VOP_TRUNCATE(dvp, (off_t)dp->i_endoff, IO_SYNC, cr, p);
if (DOINGSOFTDEP(dvp) && (tvp != NULL))
vn_lock(tvp, LK_EXCLUSIVE | LK_RETRY);
@@ -983,6 +1059,16 @@ ufs_dirremove(dvp, ip, flags, isrmdir)
(off_t)(dp->i_offset - dp->i_count), (void *)&ep, &bp)) != 0)
return (error);
+#ifdef UFS_DIRHASH
+ /*
+ * Remove the dirhash entry. This is complicated by the fact
+ * that `ep' is the previous entry when dp->i_count != 0.
+ */
+ if (dp->i_dirhash != NULL)
+ ufsdirhash_remove(dp, (dp->i_count == 0) ? ep :
+ (struct direct *)((char *)ep + ep->d_reclen), dp->i_offset);
+#endif
+
if (dp->i_count == 0) {
/*
* First entry in block: set d_ino to zero.
@@ -996,6 +1082,16 @@ ufs_dirremove(dvp, ip, flags, isrmdir)
ufs_rw16(ufs_rw16(ep->d_reclen, needswap) + dp->i_reclen,
needswap);
}
+
+#ifdef UFS_DIRHASH
+ if (dp->i_dirhash != NULL) {
+ int dirblksiz = ip->i_ump->um_dirblksiz;
+ ufsdirhash_checkblock(dp, (char *)ep -
+ ((dp->i_offset - dp->i_count) & (dirblksiz - 1)),
+ dp->i_offset & ~(dirblksiz - 1));
+ }
+#endif
+
out:
if (DOINGSOFTDEP(dvp)) {
if (ip) {
diff --git a/sys/ufs/ufs/ufs_vfsops.c b/sys/ufs/ufs/ufs_vfsops.c
index 60746ef137d..827d14276c6 100644
--- a/sys/ufs/ufs/ufs_vfsops.c
+++ b/sys/ufs/ufs/ufs_vfsops.c
@@ -1,4 +1,4 @@
-/* $NetBSD: ufs_vfsops.c,v 1.21 2004/12/20 03:12:20 dbj Exp $ */
+/* $NetBSD: ufs_vfsops.c,v 1.22 2005/01/23 19:37:05 rumble Exp $ */
/*
* Copyright (c) 1991, 1993, 1994
@@ -37,7 +37,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: ufs_vfsops.c,v 1.21 2004/12/20 03:12:20 dbj Exp $");
+__KERNEL_RCSID(0, "$NetBSD: ufs_vfsops.c,v 1.22 2005/01/23 19:37:05 rumble Exp $");
#if defined(_KERNEL_OPT)
#include "opt_quota.h"
@@ -57,6 +57,9 @@ __KERNEL_RCSID(0, "$NetBSD: ufs_vfsops.c,v 1.21 2004/12/20 03:12:20 dbj Exp $");
#include <ufs/ufs/inode.h>
#include <ufs/ufs/ufsmount.h>
#include <ufs/ufs/ufs_extern.h>
+#ifdef UFS_DIRHASH
+#include <ufs/ufs/dirhash.h>
+#endif
/* how many times ufs_init() was called */
int ufs_initcount = 0;
@@ -241,6 +244,9 @@ ufs_init()
#ifdef QUOTA
dqinit();
#endif
+#ifdef UFS_DIRHASH
+ ufsdirhash_init();
+#endif
}
void
@@ -268,4 +274,7 @@ ufs_done()
#ifdef _LKM
pool_destroy(&ufs_direct_pool);
#endif
+#ifdef UFS_DIRHASH
+ ufsdirhash_done();
+#endif
}
diff --git a/sys/ufs/ufs/ufs_vnops.c b/sys/ufs/ufs/ufs_vnops.c
index be2090e55cf..f1d6330813b 100644
--- a/sys/ufs/ufs/ufs_vnops.c
+++ b/sys/ufs/ufs/ufs_vnops.c
@@ -1,4 +1,4 @@
-/* $NetBSD: ufs_vnops.c,v 1.123 2004/09/21 03:10:36 thorpej Exp $ */
+/* $NetBSD: ufs_vnops.c,v 1.124 2005/01/23 19:37:05 rumble Exp $ */
/*
* Copyright (c) 1982, 1986, 1989, 1993, 1995
@@ -37,7 +37,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: ufs_vnops.c,v 1.123 2004/09/21 03:10:36 thorpej Exp $");
+__KERNEL_RCSID(0, "$NetBSD: ufs_vnops.c,v 1.124 2005/01/23 19:37:05 rumble Exp $");
#ifndef _LKM
#include "opt_quota.h"
@@ -68,6 +68,9 @@ __KERNEL_RCSID(0, "$NetBSD: ufs_vnops.c,v 1.123 2004/09/21 03:10:36 thorpej Exp
#include <ufs/ufs/ufsmount.h>
#include <ufs/ufs/ufs_bswap.h>
#include <ufs/ufs/ufs_extern.h>
+#ifdef UFS_DIRHASH
+#include <ufs/ufs/dirhash.h>
+#endif
#include <ufs/ext2fs/ext2fs_extern.h>
#include <ufs/lfs/lfs_extern.h>
@@ -1497,6 +1500,10 @@ ufs_rmdir(void *v)
cnp->cn_proc);
}
cache_purge(vp);
+#ifdef UFS_DIRHASH
+ if (ip->i_dirhash != NULL)
+ ufsdirhash_free(ip);
+#endif
out:
VN_KNOTE(vp, NOTE_DELETE);
vput(dvp);