2 * Copyright (c) 2001 Ian Dowse. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * $FreeBSD: src/sys/ufs/ufs/ufs_dirhash.c,v 1.3.2.6 2002/04/10 21:41:14 dwmalone Exp $
28 * This implements a hash-based lookup scheme for UFS directories.
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/malloc.h>
39 #include <sys/fnv_hash.h>
42 #include <sys/vnode.h>
43 #include <sys/mount.h>
44 #include <sys/sysctl.h>
45 #include <vm/vm_zone.h>
47 #include <ufs/ufs/quota.h>
48 #include <ufs/ufs/inode.h>
49 #include <ufs/ufs/dir.h>
50 #include <ufs/ufs/dirhash.h>
51 #include <ufs/ufs/ufsmount.h>
52 #include <ufs/ufs/ufs_extern.h>
54 #define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1))
55 #define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1))
56 #define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0)
57 #define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n))
59 static MALLOC_DEFINE(M_DIRHASH, "UFS dirhash", "UFS directory hash tables");
61 SYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem");
63 static int ufs_mindirhashsize = DIRBLKSIZ * 5;
64 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW,
66 0, "minimum directory size in bytes for which to use hashed lookup");
67 static int ufs_dirhashmaxmem = 2 * 1024 * 1024;
68 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem,
69 0, "maximum allowed dirhash memory usage");
70 static int ufs_dirhashmem;
71 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem,
72 0, "current dirhash memory usage");
73 static int ufs_dirhashcheck = 0;
74 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck,
75 0, "enable extra sanity tests");
78 static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen);
79 static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff);
80 static void ufsdirhash_delslot(struct dirhash *dh, int slot);
81 static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen,
83 static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset);
84 static void ufsdirhash_init(void);
85 static int ufsdirhash_recycle(int wanted);
87 static vm_zone_t ufsdirhash_zone;
89 /* Dirhash list; recently-used entries are near the tail. */
90 static TAILQ_HEAD(, dirhash) ufsdirhash_list;
93 * Attempt to build up a hash table for the directory contents in
94 * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
97 ufsdirhash_build(struct inode *ip)
100 struct buf *bp = NULL;
104 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
106 /* Check if we can/should use dirhash. */
107 if (ip->i_dirhash == NULL) {
108 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode))
111 /* Hash exists, but sysctls could have changed. */
112 if (ip->i_size < ufs_mindirhashsize ||
113 ufs_dirhashmem > ufs_dirhashmaxmem) {
117 /* Check if hash exists and is intact (note: unlocked read). */
118 if (ip->i_dirhash->dh_hash != NULL)
120 /* Free the old, recycled hash and build a new one. */
124 /* Don't hash removed directories. */
125 if (ip->i_effnlink == 0)
129 /* Allocate 50% more entries than this dir size could ever need. */
130 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size"));
131 nslots = ip->i_size / DIRECTSIZ(1);
132 nslots = (nslots * 3 + 1) / 2;
133 narrays = howmany(nslots, DH_NBLKOFF);
134 nslots = narrays * DH_NBLKOFF;
135 dirblocks = howmany(ip->i_size, DIRBLKSIZ);
136 nblocks = (dirblocks * 3 + 1) / 2;
138 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
139 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
140 nblocks * sizeof(*dh->dh_blkfree);
141 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
142 if (memreqd > ufs_dirhashmaxmem / 2)
145 /* Try to free some space. */
146 if (ufsdirhash_recycle(memreqd) != 0)
148 /* Enough was freed. */
150 ufs_dirhashmem += memreqd;
153 * Use non-blocking mallocs so that we will revert to a linear
154 * lookup on failure rather than potentially blocking forever.
156 MALLOC(dh, struct dirhash *, sizeof *dh, M_DIRHASH, M_NOWAIT | M_ZERO);
159 MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]),
160 M_DIRHASH, M_NOWAIT | M_ZERO);
161 MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]),
162 M_DIRHASH, M_NOWAIT);
163 if (dh->dh_hash == NULL || dh->dh_blkfree == NULL)
165 for (i = 0; i < narrays; i++) {
166 if ((dh->dh_hash[i] = zalloc(ufsdirhash_zone)) == NULL)
168 for (j = 0; j < DH_NBLKOFF; j++)
169 dh->dh_hash[i][j] = DIRHASH_EMPTY;
172 /* Initialise the hash table and block statistics. */
173 dh->dh_narrays = narrays;
174 dh->dh_hlen = nslots;
175 dh->dh_nblk = nblocks;
176 dh->dh_dirblks = dirblocks;
177 for (i = 0; i < dirblocks; i++)
178 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN;
179 for (i = 0; i < DH_NFSTATS; i++)
180 dh->dh_firstfree[i] = -1;
181 dh->dh_firstfree[DH_NFSTATS] = 0;
184 dh->dh_score = DH_SCOREINIT;
187 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
189 while (pos < ip->i_size) {
190 /* If necessary, get the next directory block. */
191 if ((pos & bmask) == 0) {
194 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
198 /* Add this entry to the hash. */
199 ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
200 if (ep->d_reclen == 0 || ep->d_reclen >
201 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) {
202 /* Corrupted directory. */
206 if (ep->d_ino != 0) {
207 /* Add the entry (simplified ufsdirhash_add). */
208 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
209 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
210 slot = WRAPINCR(slot, dh->dh_hlen);
212 DH_ENTRY(dh, slot) = pos;
213 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep));
220 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
225 if (dh->dh_hash != NULL) {
226 for (i = 0; i < narrays; i++)
227 if (dh->dh_hash[i] != NULL)
228 zfree(ufsdirhash_zone, dh->dh_hash[i]);
229 FREE(dh->dh_hash, M_DIRHASH);
231 if (dh->dh_blkfree != NULL)
232 FREE(dh->dh_blkfree, M_DIRHASH);
234 ip->i_dirhash = NULL;
235 ufs_dirhashmem -= memreqd;
240 * Free any hash table associated with inode 'ip'.
243 ufsdirhash_free(struct inode *ip)
248 if ((dh = ip->i_dirhash) == NULL)
251 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
253 /* The dirhash pointed to by 'dh' is exclusively ours now. */
256 if (dh->dh_hash != NULL) {
257 for (i = 0; i < dh->dh_narrays; i++)
258 zfree(ufsdirhash_zone, dh->dh_hash[i]);
259 FREE(dh->dh_hash, M_DIRHASH);
260 FREE(dh->dh_blkfree, M_DIRHASH);
261 mem += dh->dh_narrays * sizeof(*dh->dh_hash) +
262 dh->dh_narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
263 dh->dh_nblk * sizeof(*dh->dh_blkfree);
266 ip->i_dirhash = NULL;
268 ufs_dirhashmem -= mem;
272 * Find the offset of the specified name within the given inode.
273 * Returns 0 on success, ENOENT if the entry does not exist, or
274 * EJUSTRETURN if the caller should revert to a linear search.
276 * If successful, the directory offset is stored in *offp, and a
277 * pointer to a struct buf containing the entry is stored in *bpp. If
278 * prevoffp is non-NULL, the offset of the previous entry within
279 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
280 * is the first in a block, the start of the block is used).
283 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp,
284 struct buf **bpp, doff_t *prevoffp)
286 struct dirhash *dh, *dh_next;
290 doff_t blkoff, bmask, offset, prevoff;
293 if ((dh = ip->i_dirhash) == NULL)
294 return (EJUSTRETURN);
296 * Move this dirhash towards the end of the list if it has a
297 * score higher than the next entry.
298 * Optimise the case where it's already the last by performing
299 * an unlocked read of the TAILQ_NEXT pointer.
301 if (TAILQ_NEXT(dh, dh_list) != NULL) {
303 * If the new score will be greater than that of the next
304 * entry, then move this entry past it. With both mutexes
305 * held, dh_next won't go away, but its dh_score could
306 * change; that's not important since it is just a hint.
308 if (dh->dh_hash != NULL &&
309 (dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
310 dh->dh_score >= dh_next->dh_score) {
311 KASSERT(dh->dh_onlist, ("dirhash: not on list"));
312 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
313 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
317 if (dh->dh_hash == NULL) {
319 return (EJUSTRETURN);
322 /* Update the score. */
323 if (dh->dh_score < DH_SCOREMAX)
327 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
331 slot = ufsdirhash_hash(dh, name, namelen);
335 * Sequential access optimisation. dh_seqoff contains the
336 * offset of the directory entry immediately following
337 * the last entry that was looked up. Check if this offset
338 * appears in the hash chain for the name we are looking for.
340 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
341 i = WRAPINCR(i, dh->dh_hlen))
342 if (offset == dh->dh_seqoff)
344 if (offset == dh->dh_seqoff) {
346 * We found an entry with the expected offset. This
347 * is probably the entry we want, but if not, the
348 * code below will turn off seqoff and retry.
355 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
356 slot = WRAPINCR(slot, dh->dh_hlen)) {
357 if (offset == DIRHASH_DEL)
360 if (offset < 0 || offset >= ip->i_size)
361 panic("ufsdirhash_lookup: bad offset in hash array");
362 if ((offset & ~bmask) != blkoff) {
365 blkoff = offset & ~bmask;
366 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0)
367 return (EJUSTRETURN);
369 dp = (struct direct *)(bp->b_data + (offset & bmask));
370 if (dp->d_reclen == 0 || dp->d_reclen >
371 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
372 /* Corrupted directory. */
374 return (EJUSTRETURN);
376 if (dp->d_namlen == namelen &&
377 bcmp(dp->d_name, name, namelen) == 0) {
378 /* Found. Get the prev offset if needed. */
379 if (prevoffp != NULL) {
380 if (offset & (DIRBLKSIZ - 1)) {
381 prevoff = ufsdirhash_getprev(dp,
385 return (EJUSTRETURN);
392 /* Check for sequential access, and update offset. */
393 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset)
395 dh->dh_seqoff = offset + DIRSIZ(0, dp);
402 if (dh->dh_hash == NULL) {
406 return (EJUSTRETURN);
409 * When the name doesn't match in the seqopt case, go back
410 * and search normally.
423 * Find a directory block with room for 'slotneeded' bytes. Returns
424 * the offset of the directory entry that begins the free space.
425 * This will either be the offset of an existing entry that has free
426 * space at the end, or the offset of an entry with d_ino == 0 at
427 * the start of a DIRBLKSIZ block.
429 * To use the space, the caller may need to compact existing entries in
430 * the directory. The total number of bytes in all of the entries involved
431 * in the compaction is stored in *slotsize. In other words, all of
432 * the entries that must be compacted are exactly contained in the
433 * region beginning at the returned offset and spanning *slotsize bytes.
435 * Returns -1 if no space was found, indicating that the directory
439 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
444 doff_t pos, slotstart;
445 int dirblock, error, freebytes, i;
447 if ((dh = ip->i_dirhash) == NULL)
449 if (dh->dh_hash == NULL) {
454 /* Find a directory block with the desired free space. */
456 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
457 if ((dirblock = dh->dh_firstfree[i]) != -1)
459 if (dirblock == -1) {
463 KASSERT(dirblock < dh->dh_nblk &&
464 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
465 ("ufsdirhash_findfree: bad stats"));
466 pos = dirblock * DIRBLKSIZ;
467 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp);
471 /* Find the first entry with free space. */
472 for (i = 0; i < DIRBLKSIZ; ) {
473 if (dp->d_reclen == 0) {
477 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
480 dp = (struct direct *)((char *)dp + dp->d_reclen);
488 /* Find the range of entries needed to get enough space */
490 while (i < DIRBLKSIZ && freebytes < slotneeded) {
491 freebytes += dp->d_reclen;
493 freebytes -= DIRSIZ(0, dp);
494 if (dp->d_reclen == 0) {
499 dp = (struct direct *)((char *)dp + dp->d_reclen);
505 if (freebytes < slotneeded)
506 panic("ufsdirhash_findfree: free mismatch");
508 *slotsize = pos + i - slotstart;
513 * Return the start of the unused space at the end of a directory, or
514 * -1 if there are no trailing unused blocks.
517 ufsdirhash_enduseful(struct inode *ip)
523 if ((dh = ip->i_dirhash) == NULL)
525 if (dh->dh_hash == NULL) {
530 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN) {
534 for (i = dh->dh_dirblks - 1; i >= 0; i--)
535 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
537 return ((doff_t)(i + 1) * DIRBLKSIZ);
541 * Insert information into the hash about a new directory entry. dirp
542 * points to a struct direct containing the entry, and offset specifies
543 * the offset of this entry.
546 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
551 if ((dh = ip->i_dirhash) == NULL)
553 if (dh->dh_hash == NULL) {
558 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
559 ("ufsdirhash_add: bad offset"));
561 * Normal hash usage is < 66%. If the usage gets too high then
562 * remove the hash entirely and let it be rebuilt later.
564 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
569 /* Find a free hash slot (empty or deleted), and add the entry. */
570 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
571 while (DH_ENTRY(dh, slot) >= 0)
572 slot = WRAPINCR(slot, dh->dh_hlen);
573 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
575 DH_ENTRY(dh, slot) = offset;
577 /* Update the per-block summary info. */
578 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp));
582 * Remove the specified directory entry from the hash. The entry to remove
583 * is defined by the name in `dirp', which must exist at the specified
584 * `offset' within the directory.
587 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
592 if ((dh = ip->i_dirhash) == NULL)
594 if (dh->dh_hash == NULL) {
599 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
600 ("ufsdirhash_remove: bad offset"));
602 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
604 /* Remove the hash entry. */
605 ufsdirhash_delslot(dh, slot);
607 /* Update the per-block summary info. */
608 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp));
612 * Change the offset associated with a directory entry in the hash. Used
613 * when compacting directory blocks.
616 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
622 if ((dh = ip->i_dirhash) == NULL)
624 if (dh->dh_hash == NULL) {
629 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ &&
630 newoff < dh->dh_dirblks * DIRBLKSIZ,
631 ("ufsdirhash_move: bad offset"));
632 /* Find the entry, and update the offset. */
633 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
634 DH_ENTRY(dh, slot) = newoff;
638 * Inform dirhash that the directory has grown by one block that
639 * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
642 ufsdirhash_newblk(struct inode *ip, doff_t offset)
647 if ((dh = ip->i_dirhash) == NULL)
649 if (dh->dh_hash == NULL) {
654 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ,
655 ("ufsdirhash_newblk: bad offset"));
656 block = offset / DIRBLKSIZ;
657 if (block >= dh->dh_nblk) {
658 /* Out of space; must rebuild. */
662 dh->dh_dirblks = block + 1;
664 /* Account for the new free block. */
665 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN;
666 if (dh->dh_firstfree[DH_NFSTATS] == -1)
667 dh->dh_firstfree[DH_NFSTATS] = block;
671 * Inform dirhash that the directory is being truncated.
674 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
679 if ((dh = ip->i_dirhash) == NULL)
681 if (dh->dh_hash == NULL) {
686 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ,
687 ("ufsdirhash_dirtrunc: bad offset"));
688 block = howmany(offset, DIRBLKSIZ);
690 * If the directory shrinks to less than 1/8 of dh_nblk blocks
691 * (about 20% of its original size due to the 50% extra added in
692 * ufsdirhash_build) then free it, and let the caller rebuild
695 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
701 * Remove any `first free' information pertaining to the
702 * truncated blocks. All blocks we're removing should be
705 if (dh->dh_firstfree[DH_NFSTATS] >= block)
706 dh->dh_firstfree[DH_NFSTATS] = -1;
707 for (i = block; i < dh->dh_dirblks; i++)
708 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
709 panic("ufsdirhash_dirtrunc: blocks in use");
710 for (i = 0; i < DH_NFSTATS; i++)
711 if (dh->dh_firstfree[i] >= block)
712 panic("ufsdirhash_dirtrunc: first free corrupt");
713 dh->dh_dirblks = block;
717 * Debugging function to check that the dirhash information about
718 * a directory block matches its actual contents. Panics if a mismatch
721 * On entry, `buf' should point to the start of an in-core
722 * DIRBLKSIZ-sized directory block, and `offset' should contain the
723 * offset from the start of the directory of that block.
726 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
730 int block, ffslot, i, nfree;
732 if (!ufs_dirhashcheck)
734 if ((dh = ip->i_dirhash) == NULL)
736 if (dh->dh_hash == NULL) {
741 block = offset / DIRBLKSIZ;
742 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks)
743 panic("ufsdirhash_checkblock: bad offset");
746 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) {
747 dp = (struct direct *)(buf + i);
748 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ)
749 panic("ufsdirhash_checkblock: bad dir");
751 if (dp->d_ino == 0) {
754 * XXX entries with d_ino == 0 should only occur
755 * at the start of a DIRBLKSIZ block. However the
756 * ufs code is tolerant of such entries at other
757 * offsets, and fsck does not fix them.
760 panic("ufsdirhash_checkblock: bad dir inode");
762 nfree += dp->d_reclen;
766 /* Check that the entry exists (will panic if it doesn't). */
767 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
769 nfree += dp->d_reclen - DIRSIZ(0, dp);
772 panic("ufsdirhash_checkblock: bad dir end");
774 if (dh->dh_blkfree[block] * DIRALIGN != nfree)
775 panic("ufsdirhash_checkblock: bad free count");
777 ffslot = BLKFREE2IDX(nfree / DIRALIGN);
778 for (i = 0; i <= DH_NFSTATS; i++)
779 if (dh->dh_firstfree[i] == block && i != ffslot)
780 panic("ufsdirhash_checkblock: bad first-free");
781 if (dh->dh_firstfree[ffslot] == -1)
782 panic("ufsdirhash_checkblock: missing first-free entry");
786 * Hash the specified filename into a dirhash slot.
789 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen)
794 * We hash the name and then some ofther bit of data which is
795 * invarient over the dirhash's lifetime. Otherwise names
796 * differing only in the last byte are placed close to one
797 * another in the table, which is bad for linear probing.
799 hash = fnv_32_buf(name, namelen, FNV1_32_INIT);
800 hash = fnv_32_buf(dh, sizeof(dh), hash);
801 return (hash % dh->dh_hlen);
805 * Adjust the number of free bytes in the block containing `offset'
806 * by the value specified by `diff'.
808 * The caller must ensure we have exclusive access to `dh'.
811 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff)
813 int block, i, nfidx, ofidx;
815 /* Update the per-block summary info. */
816 block = offset / DIRBLKSIZ;
817 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks,
818 ("dirhash bad offset"));
819 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
820 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
821 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
823 /* Update the `first free' list if necessary. */
824 if (ofidx != nfidx) {
825 /* If removing, scan forward for the next block. */
826 if (dh->dh_firstfree[ofidx] == block) {
827 for (i = block + 1; i < dh->dh_dirblks; i++)
828 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
830 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
833 /* Make this the new `first free' if necessary */
834 if (dh->dh_firstfree[nfidx] > block ||
835 dh->dh_firstfree[nfidx] == -1)
836 dh->dh_firstfree[nfidx] = block;
841 * Find the specified name which should have the specified offset.
842 * Returns a slot number, and panics on failure.
844 * `dh' must be locked on entry and remains so on return.
847 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset)
851 /* Find the entry. */
852 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full"));
853 slot = ufsdirhash_hash(dh, name, namelen);
854 while (DH_ENTRY(dh, slot) != offset &&
855 DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
856 slot = WRAPINCR(slot, dh->dh_hlen);
857 if (DH_ENTRY(dh, slot) != offset)
858 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
864 * Remove the entry corresponding to the specified slot from the hash array.
866 * `dh' must be locked on entry and remains so on return.
869 ufsdirhash_delslot(struct dirhash *dh, int slot)
873 /* Mark the entry as deleted. */
874 DH_ENTRY(dh, slot) = DIRHASH_DEL;
876 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
877 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
878 i = WRAPINCR(i, dh->dh_hlen);
879 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
880 i = WRAPDECR(i, dh->dh_hlen);
881 while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
882 DH_ENTRY(dh, i) = DIRHASH_EMPTY;
884 i = WRAPDECR(i, dh->dh_hlen);
886 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen"));
891 * Given a directory entry and its offset, find the offset of the
892 * previous entry in the same DIRBLKSIZ-sized block. Returns an
893 * offset, or -1 if there is no previous entry in the block or some
894 * other problem occurred.
897 ufsdirhash_getprev(struct direct *dirp, doff_t offset)
901 doff_t blkoff, prevoff;
904 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */
905 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */
906 blkbuf = (char *)dirp - entrypos;
909 /* If `offset' is the start of a block, there is no previous entry. */
913 /* Scan from the start of the block until we get to the entry. */
914 for (i = 0; i < entrypos; i += dp->d_reclen) {
915 dp = (struct direct *)(blkbuf + i);
916 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
917 return (-1); /* Corrupted directory. */
918 prevoff = blkoff + i;
924 * Try to free up `wanted' bytes by stealing memory from existing
925 * dirhashes. Returns zero if successful.
928 ufsdirhash_recycle(int wanted)
935 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
936 /* Find a dirhash, and lock it. */
937 if ((dh = TAILQ_FIRST(&ufsdirhash_list)) == NULL) {
940 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list"));
942 /* Decrement the score; only recycle if it becomes zero. */
943 if (--dh->dh_score > 0) {
947 /* Remove it from the list and detach its memory. */
948 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
952 blkfree = dh->dh_blkfree;
953 dh->dh_blkfree = NULL;
954 narrays = dh->dh_narrays;
955 mem = narrays * sizeof(*dh->dh_hash) +
956 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
957 dh->dh_nblk * sizeof(*dh->dh_blkfree);
959 /* Free the detached memory. */
960 for (i = 0; i < narrays; i++)
961 zfree(ufsdirhash_zone, hash[i]);
962 FREE(hash, M_DIRHASH);
963 FREE(blkfree, M_DIRHASH);
965 /* Account for the returned memory, and repeat if necessary. */
966 ufs_dirhashmem -= mem;
976 ufsdirhash_zone = zinit("DIRHASH", DH_NBLKOFF * sizeof(daddr_t), 0,
978 TAILQ_INIT(&ufsdirhash_list);
980 SYSINIT(ufsdirhash, SI_SUB_PSEUDO, SI_ORDER_ANY, ufsdirhash_init, NULL)
983 #endif /* UFS_DIRHASH */