2 * Copyright (c) 2001, 2002 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
27 * This implements a hash-based lookup scheme for UFS directories.
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
41 #include <sys/mutex.h>
42 #include <sys/malloc.h>
43 #include <sys/fnv_hash.h>
47 #include <sys/vnode.h>
48 #include <sys/mount.h>
49 #include <sys/refcount.h>
50 #include <sys/sysctl.h>
54 #include <ufs/ufs/quota.h>
55 #include <ufs/ufs/inode.h>
56 #include <ufs/ufs/dir.h>
57 #include <ufs/ufs/dirhash.h>
58 #include <ufs/ufs/extattr.h>
59 #include <ufs/ufs/ufsmount.h>
60 #include <ufs/ufs/ufs_extern.h>
62 #define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1))
63 #define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1))
64 #define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0)
65 #define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n))
67 static MALLOC_DEFINE(M_DIRHASH, "ufs_dirhash", "UFS directory hash tables");
69 static SYSCTL_NODE(_vfs, OID_AUTO, ufs, CTLFLAG_RD, 0, "UFS filesystem");
71 static int ufs_mindirhashsize = DIRBLKSIZ * 5;
72 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_minsize, CTLFLAG_RW,
74 0, "minimum directory size in bytes for which to use hashed lookup");
75 static int ufs_dirhashmaxmem = 2 * 1024 * 1024;
76 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem,
77 0, "maximum allowed dirhash memory usage");
78 static int ufs_dirhashmem;
79 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem,
80 0, "current dirhash memory usage");
81 static int ufs_dirhashcheck = 0;
82 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck,
83 0, "enable extra sanity tests");
86 static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen);
87 static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff);
88 static void ufsdirhash_delslot(struct dirhash *dh, int slot);
89 static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen,
91 static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset);
92 static int ufsdirhash_recycle(int wanted);
93 static void ufsdirhash_free_locked(struct inode *ip);
95 static uma_zone_t ufsdirhash_zone;
97 #define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx)
98 #define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx)
99 #define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK)
100 #define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr))
101 #define DIRHASH_ASSERT_LOCKED(dh) \
102 sx_assert(&(dh)->dh_lock, SA_LOCKED)
104 /* Dirhash list; recently-used entries are near the tail. */
105 static TAILQ_HEAD(, dirhash) ufsdirhash_list;
107 /* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */
108 static struct mtx ufsdirhash_mtx;
113 * The relationship between inode and dirhash is protected either by an
114 * exclusive vnode lock or the vnode interlock where a shared vnode lock
115 * may be used. The dirhash_mtx is acquired after the dirhash lock. To
116 * handle teardown races, code wishing to lock the dirhash for an inode
117 * when using a shared vnode lock must obtain a private reference on the
118 * dirhash while holding the vnode interlock. They can drop it once they
119 * have obtained the dirhash lock and verified that the dirhash wasn't
120 * recycled while they waited for the dirhash lock.
122 * ufsdirhash_build() acquires a shared lock on the dirhash when it is
123 * successful. This lock is released after a call to ufsdirhash_lookup().
125 * Functions requiring exclusive access use ufsdirhash_acquire() which may
126 * free a dirhash structure that was recycled by ufsdirhash_recycle().
128 * The dirhash lock may be held across io operations.
130 * WITNESS reports a lock order reversal between the "bufwait" lock
131 * and the "dirhash" lock. However, this specific reversal will not
132 * cause a deadlock. To get a deadlock, one would have to lock a
133 * buffer followed by the dirhash while a second thread locked a
134 * buffer while holding the dirhash lock. The second order can happen
135 * under a shared or exclusive vnode lock for the associated directory
136 * in lookup(). The first order, however, can only happen under an
137 * exclusive vnode lock (e.g. unlink(), rename(), etc.). Thus, for
138 * a thread to be doing a "bufwait" -> "dirhash" order, it has to hold
139 * an exclusive vnode lock. That exclusive vnode lock will prevent
140 * any other threads from doing a "dirhash" -> "bufwait" order.
144 ufsdirhash_hold(struct dirhash *dh)
147 refcount_acquire(&dh->dh_refcount);
151 ufsdirhash_drop(struct dirhash *dh)
154 if (refcount_release(&dh->dh_refcount)) {
155 sx_destroy(&dh->dh_lock);
161 * Release the lock on a dirhash.
164 ufsdirhash_release(struct dirhash *dh)
167 sx_unlock(&dh->dh_lock);
171 * Either acquire an existing hash locked shared or create a new hash and
172 * return it exclusively locked. May return NULL if the allocation fails.
174 * The vnode interlock is used to protect the i_dirhash pointer from
175 * simultaneous access while only a shared vnode lock is held.
177 static struct dirhash *
178 ufsdirhash_create(struct inode *ip)
189 /* Racy check for i_dirhash to prefetch an dirhash structure. */
190 if (ip->i_dirhash == NULL && ndh == NULL) {
191 MALLOC(ndh, struct dirhash *, sizeof *dh, M_DIRHASH,
195 refcount_init(&ndh->dh_refcount, 1);
196 sx_init(&ndh->dh_lock, "dirhash");
197 sx_xlock(&ndh->dh_lock);
200 * Check i_dirhash. If it's NULL just try to use a
201 * preallocated structure. If none exists loop and try again.
215 /* Acquire a shared lock on existing hashes. */
216 sx_slock(&dh->dh_lock);
218 /* The hash could've been recycled while we were waiting. */
220 if (ip->i_dirhash != dh) {
222 ufsdirhash_release(dh);
229 /* If the hash is still valid we've succeeded. */
230 if (dh->dh_hash != NULL)
233 * If the hash is NULL it has been recycled. Try to upgrade
234 * so we can recreate it. If we fail the upgrade, drop our
235 * lock and try again.
237 if (sx_try_upgrade(&dh->dh_lock))
239 sx_sunlock(&dh->dh_lock);
241 /* Free the preallocated structure if it was not necessary. */
243 ufsdirhash_release(ndh);
244 ufsdirhash_drop(ndh);
250 * Acquire an exclusive lock on an existing hash. Requires an exclusive
251 * vnode lock to protect the i_dirhash pointer. hashes that have been
252 * recycled are reclaimed here and NULL is returned.
254 static struct dirhash *
255 ufsdirhash_acquire(struct inode *ip)
260 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__);
266 sx_xlock(&dh->dh_lock);
267 if (dh->dh_hash != NULL)
269 ufsdirhash_free_locked(ip);
274 * Acquire exclusively and free the hash pointed to by ip. Works with a
275 * shared or exclusive vnode lock.
278 ufsdirhash_free(struct inode *ip)
285 /* Grab a reference on this inode's dirhash if it has one. */
295 /* Exclusively lock the dirhash. */
296 sx_xlock(&dh->dh_lock);
298 /* If this dirhash still belongs to this inode, then free it. */
300 if (ip->i_dirhash == dh) {
308 * This inode's dirhash has changed while we were
309 * waiting for the dirhash lock, so try again.
311 ufsdirhash_release(dh);
314 ufsdirhash_free_locked(ip);
318 * Attempt to build up a hash table for the directory contents in
319 * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
322 ufsdirhash_build(struct inode *ip)
325 struct buf *bp = NULL;
329 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
331 /* Take care of a decreased sysctl value. */
332 while (ufs_dirhashmem > ufs_dirhashmaxmem)
333 if (ufsdirhash_recycle(0) != 0)
336 /* Check if we can/should use dirhash. */
337 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) ||
338 ip->i_effnlink == 0) {
343 dh = ufsdirhash_create(ip);
346 if (dh->dh_hash != NULL)
350 /* Allocate 50% more entries than this dir size could ever need. */
351 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size"));
352 nslots = ip->i_size / DIRECTSIZ(1);
353 nslots = (nslots * 3 + 1) / 2;
354 narrays = howmany(nslots, DH_NBLKOFF);
355 nslots = narrays * DH_NBLKOFF;
356 dirblocks = howmany(ip->i_size, DIRBLKSIZ);
357 nblocks = (dirblocks * 3 + 1) / 2;
358 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
359 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
360 nblocks * sizeof(*dh->dh_blkfree);
362 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
363 DIRHASHLIST_UNLOCK();
364 if (memreqd > ufs_dirhashmaxmem / 2)
366 /* Try to free some space. */
367 if (ufsdirhash_recycle(memreqd) != 0)
369 /* Enough was freed, and list has been locked. */
371 ufs_dirhashmem += memreqd;
372 DIRHASHLIST_UNLOCK();
374 /* Initialise the hash table and block statistics. */
375 dh->dh_memreq = memreqd;
376 dh->dh_narrays = narrays;
377 dh->dh_hlen = nslots;
378 dh->dh_nblk = nblocks;
379 dh->dh_dirblks = dirblocks;
380 for (i = 0; i < DH_NFSTATS; i++)
381 dh->dh_firstfree[i] = -1;
382 dh->dh_firstfree[DH_NFSTATS] = 0;
386 dh->dh_score = DH_SCOREINIT;
389 * Use non-blocking mallocs so that we will revert to a linear
390 * lookup on failure rather than potentially blocking forever.
392 MALLOC(dh->dh_hash, doff_t **, narrays * sizeof(dh->dh_hash[0]),
393 M_DIRHASH, M_NOWAIT | M_ZERO);
394 if (dh->dh_hash == NULL)
396 MALLOC(dh->dh_blkfree, u_int8_t *, nblocks * sizeof(dh->dh_blkfree[0]),
397 M_DIRHASH, M_NOWAIT);
398 if (dh->dh_blkfree == NULL)
400 for (i = 0; i < narrays; i++) {
401 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL)
403 for (j = 0; j < DH_NBLKOFF; j++)
404 dh->dh_hash[i][j] = DIRHASH_EMPTY;
406 for (i = 0; i < dirblocks; i++)
407 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN;
408 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
410 while (pos < ip->i_size) {
411 /* If necessary, get the next directory block. */
412 if ((pos & bmask) == 0) {
415 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
419 /* Add this entry to the hash. */
420 ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
421 if (ep->d_reclen == 0 || ep->d_reclen >
422 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) {
423 /* Corrupted directory. */
427 if (ep->d_ino != 0) {
428 /* Add the entry (simplified ufsdirhash_add). */
429 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
430 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
431 slot = WRAPINCR(slot, dh->dh_hlen);
433 DH_ENTRY(dh, slot) = pos;
434 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep));
442 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
444 DIRHASHLIST_UNLOCK();
445 sx_downgrade(&dh->dh_lock);
449 ufsdirhash_free_locked(ip);
454 * Free any hash table associated with inode 'ip'.
457 ufsdirhash_free_locked(struct inode *ip)
463 DIRHASH_ASSERT_LOCKED(ip->i_dirhash);
466 * Clear the pointer in the inode to prevent new threads from
467 * finding the dead structure.
472 ip->i_dirhash = NULL;
476 * Remove the hash from the list since we are going to free its
481 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
482 ufs_dirhashmem -= dh->dh_memreq;
483 DIRHASHLIST_UNLOCK();
486 * At this point, any waiters for the lock should hold their
487 * own reference on the dirhash structure. They will drop
488 * that reference once they grab the vnode interlock and see
489 * that ip->i_dirhash is NULL.
491 sx_xunlock(&dh->dh_lock);
494 * Handle partially recycled as well as fully constructed hashes.
496 if (dh->dh_hash != NULL) {
497 for (i = 0; i < dh->dh_narrays; i++)
498 if (dh->dh_hash[i] != NULL)
499 DIRHASH_BLKFREE(dh->dh_hash[i]);
500 FREE(dh->dh_hash, M_DIRHASH);
501 if (dh->dh_blkfree != NULL)
502 FREE(dh->dh_blkfree, M_DIRHASH);
506 * Drop the inode's reference to the data structure.
512 * Find the offset of the specified name within the given inode.
513 * Returns 0 on success, ENOENT if the entry does not exist, or
514 * EJUSTRETURN if the caller should revert to a linear search.
516 * If successful, the directory offset is stored in *offp, and a
517 * pointer to a struct buf containing the entry is stored in *bpp. If
518 * prevoffp is non-NULL, the offset of the previous entry within
519 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
520 * is the first in a block, the start of the block is used).
522 * Must be called with the hash locked. Returns with the hash unlocked.
525 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp,
526 struct buf **bpp, doff_t *prevoffp)
528 struct dirhash *dh, *dh_next;
532 doff_t blkoff, bmask, offset, prevoff;
537 KASSERT(dh != NULL && dh->dh_hash != NULL,
538 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh));
539 DIRHASH_ASSERT_LOCKED(dh);
541 * Move this dirhash towards the end of the list if it has a
542 * score higher than the next entry, and acquire the dh_lock.
545 if (TAILQ_NEXT(dh, dh_list) != NULL) {
547 * If the new score will be greater than that of the next
548 * entry, then move this entry past it. With both mutexes
549 * held, dh_next won't go away, but its dh_score could
550 * change; that's not important since it is just a hint.
552 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
553 dh->dh_score >= dh_next->dh_score) {
554 KASSERT(dh->dh_onlist, ("dirhash: not on list"));
555 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
556 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
560 /* Update the score. */
561 if (dh->dh_score < DH_SCOREMAX)
563 DIRHASHLIST_UNLOCK();
566 bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
570 slot = ufsdirhash_hash(dh, name, namelen);
574 * Sequential access optimisation. dh_seqoff contains the
575 * offset of the directory entry immediately following
576 * the last entry that was looked up. Check if this offset
577 * appears in the hash chain for the name we are looking for.
579 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
580 i = WRAPINCR(i, dh->dh_hlen))
581 if (offset == dh->dh_seqoff)
583 if (offset == dh->dh_seqoff) {
585 * We found an entry with the expected offset. This
586 * is probably the entry we want, but if not, the
587 * code below will turn off seqopt and retry.
594 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
595 slot = WRAPINCR(slot, dh->dh_hlen)) {
596 if (offset == DIRHASH_DEL)
598 if (offset < 0 || offset >= ip->i_size)
599 panic("ufsdirhash_lookup: bad offset in hash array");
600 if ((offset & ~bmask) != blkoff) {
603 blkoff = offset & ~bmask;
604 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) {
609 dp = (struct direct *)(bp->b_data + (offset & bmask));
610 if (dp->d_reclen == 0 || dp->d_reclen >
611 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
612 /* Corrupted directory. */
616 if (dp->d_namlen == namelen &&
617 bcmp(dp->d_name, name, namelen) == 0) {
618 /* Found. Get the prev offset if needed. */
619 if (prevoffp != NULL) {
620 if (offset & (DIRBLKSIZ - 1)) {
621 prevoff = ufsdirhash_getprev(dp,
632 /* Check for sequential access, and update offset. */
633 if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset)
635 dh->dh_seqoff = offset + DIRSIZ(0, dp);
638 ufsdirhash_release(dh);
643 * When the name doesn't match in the seqopt case, go back
644 * and search normally.
653 ufsdirhash_release(dh);
660 * Find a directory block with room for 'slotneeded' bytes. Returns
661 * the offset of the directory entry that begins the free space.
662 * This will either be the offset of an existing entry that has free
663 * space at the end, or the offset of an entry with d_ino == 0 at
664 * the start of a DIRBLKSIZ block.
666 * To use the space, the caller may need to compact existing entries in
667 * the directory. The total number of bytes in all of the entries involved
668 * in the compaction is stored in *slotsize. In other words, all of
669 * the entries that must be compacted are exactly contained in the
670 * region beginning at the returned offset and spanning *slotsize bytes.
672 * Returns -1 if no space was found, indicating that the directory
676 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
681 doff_t pos, slotstart;
682 int dirblock, error, freebytes, i;
685 KASSERT(dh != NULL && dh->dh_hash != NULL,
686 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh));
687 DIRHASH_ASSERT_LOCKED(dh);
689 /* Find a directory block with the desired free space. */
691 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
692 if ((dirblock = dh->dh_firstfree[i]) != -1)
697 KASSERT(dirblock < dh->dh_nblk &&
698 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
699 ("ufsdirhash_findfree: bad stats"));
700 pos = dirblock * DIRBLKSIZ;
701 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp);
705 /* Find the first entry with free space. */
706 for (i = 0; i < DIRBLKSIZ; ) {
707 if (dp->d_reclen == 0) {
711 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
714 dp = (struct direct *)((char *)dp + dp->d_reclen);
722 /* Find the range of entries needed to get enough space */
724 while (i < DIRBLKSIZ && freebytes < slotneeded) {
725 freebytes += dp->d_reclen;
727 freebytes -= DIRSIZ(0, dp);
728 if (dp->d_reclen == 0) {
733 dp = (struct direct *)((char *)dp + dp->d_reclen);
739 if (freebytes < slotneeded)
740 panic("ufsdirhash_findfree: free mismatch");
742 *slotsize = pos + i - slotstart;
747 * Return the start of the unused space at the end of a directory, or
748 * -1 if there are no trailing unused blocks.
751 ufsdirhash_enduseful(struct inode *ip)
758 DIRHASH_ASSERT_LOCKED(dh);
759 KASSERT(dh != NULL && dh->dh_hash != NULL,
760 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh));
762 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN)
765 for (i = dh->dh_dirblks - 1; i >= 0; i--)
766 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
769 return ((doff_t)(i + 1) * DIRBLKSIZ);
773 * Insert information into the hash about a new directory entry. dirp
774 * points to a struct direct containing the entry, and offset specifies
775 * the offset of this entry.
778 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
783 if ((dh = ufsdirhash_acquire(ip)) == NULL)
786 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
787 ("ufsdirhash_add: bad offset"));
789 * Normal hash usage is < 66%. If the usage gets too high then
790 * remove the hash entirely and let it be rebuilt later.
792 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
793 ufsdirhash_free_locked(ip);
797 /* Find a free hash slot (empty or deleted), and add the entry. */
798 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
799 while (DH_ENTRY(dh, slot) >= 0)
800 slot = WRAPINCR(slot, dh->dh_hlen);
801 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
803 DH_ENTRY(dh, slot) = offset;
805 /* Update the per-block summary info. */
806 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp));
807 ufsdirhash_release(dh);
811 * Remove the specified directory entry from the hash. The entry to remove
812 * is defined by the name in `dirp', which must exist at the specified
813 * `offset' within the directory.
816 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
821 if ((dh = ufsdirhash_acquire(ip)) == NULL)
824 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
825 ("ufsdirhash_remove: bad offset"));
827 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
829 /* Remove the hash entry. */
830 ufsdirhash_delslot(dh, slot);
832 /* Update the per-block summary info. */
833 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp));
834 ufsdirhash_release(dh);
838 * Change the offset associated with a directory entry in the hash. Used
839 * when compacting directory blocks.
842 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
848 if ((dh = ufsdirhash_acquire(ip)) == NULL)
851 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ &&
852 newoff < dh->dh_dirblks * DIRBLKSIZ,
853 ("ufsdirhash_move: bad offset"));
854 /* Find the entry, and update the offset. */
855 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
856 DH_ENTRY(dh, slot) = newoff;
857 ufsdirhash_release(dh);
861 * Inform dirhash that the directory has grown by one block that
862 * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
865 ufsdirhash_newblk(struct inode *ip, doff_t offset)
870 if ((dh = ufsdirhash_acquire(ip)) == NULL)
873 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ,
874 ("ufsdirhash_newblk: bad offset"));
875 block = offset / DIRBLKSIZ;
876 if (block >= dh->dh_nblk) {
877 /* Out of space; must rebuild. */
878 ufsdirhash_free_locked(ip);
881 dh->dh_dirblks = block + 1;
883 /* Account for the new free block. */
884 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN;
885 if (dh->dh_firstfree[DH_NFSTATS] == -1)
886 dh->dh_firstfree[DH_NFSTATS] = block;
887 ufsdirhash_release(dh);
891 * Inform dirhash that the directory is being truncated.
894 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
899 if ((dh = ufsdirhash_acquire(ip)) == NULL)
902 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ,
903 ("ufsdirhash_dirtrunc: bad offset"));
904 block = howmany(offset, DIRBLKSIZ);
906 * If the directory shrinks to less than 1/8 of dh_nblk blocks
907 * (about 20% of its original size due to the 50% extra added in
908 * ufsdirhash_build) then free it, and let the caller rebuild
911 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
912 ufsdirhash_free_locked(ip);
917 * Remove any `first free' information pertaining to the
918 * truncated blocks. All blocks we're removing should be
921 if (dh->dh_firstfree[DH_NFSTATS] >= block)
922 dh->dh_firstfree[DH_NFSTATS] = -1;
923 for (i = block; i < dh->dh_dirblks; i++)
924 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
925 panic("ufsdirhash_dirtrunc: blocks in use");
926 for (i = 0; i < DH_NFSTATS; i++)
927 if (dh->dh_firstfree[i] >= block)
928 panic("ufsdirhash_dirtrunc: first free corrupt");
929 dh->dh_dirblks = block;
930 ufsdirhash_release(dh);
934 * Debugging function to check that the dirhash information about
935 * a directory block matches its actual contents. Panics if a mismatch
938 * On entry, `buf' should point to the start of an in-core
939 * DIRBLKSIZ-sized directory block, and `offset' should contain the
940 * offset from the start of the directory of that block.
943 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
947 int block, ffslot, i, nfree;
949 if (!ufs_dirhashcheck)
951 if ((dh = ufsdirhash_acquire(ip)) == NULL)
954 block = offset / DIRBLKSIZ;
955 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks)
956 panic("ufsdirhash_checkblock: bad offset");
959 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) {
960 dp = (struct direct *)(buf + i);
961 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ)
962 panic("ufsdirhash_checkblock: bad dir");
964 if (dp->d_ino == 0) {
967 * XXX entries with d_ino == 0 should only occur
968 * at the start of a DIRBLKSIZ block. However the
969 * ufs code is tolerant of such entries at other
970 * offsets, and fsck does not fix them.
973 panic("ufsdirhash_checkblock: bad dir inode");
975 nfree += dp->d_reclen;
979 /* Check that the entry exists (will panic if it doesn't). */
980 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
982 nfree += dp->d_reclen - DIRSIZ(0, dp);
985 panic("ufsdirhash_checkblock: bad dir end");
987 if (dh->dh_blkfree[block] * DIRALIGN != nfree)
988 panic("ufsdirhash_checkblock: bad free count");
990 ffslot = BLKFREE2IDX(nfree / DIRALIGN);
991 for (i = 0; i <= DH_NFSTATS; i++)
992 if (dh->dh_firstfree[i] == block && i != ffslot)
993 panic("ufsdirhash_checkblock: bad first-free");
994 if (dh->dh_firstfree[ffslot] == -1)
995 panic("ufsdirhash_checkblock: missing first-free entry");
996 ufsdirhash_release(dh);
1000 * Hash the specified filename into a dirhash slot.
1003 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen)
1008 * We hash the name and then some other bit of data that is
1009 * invariant over the dirhash's lifetime. Otherwise names
1010 * differing only in the last byte are placed close to one
1011 * another in the table, which is bad for linear probing.
1013 hash = fnv_32_buf(name, namelen, FNV1_32_INIT);
1014 hash = fnv_32_buf(&dh, sizeof(dh), hash);
1015 return (hash % dh->dh_hlen);
1019 * Adjust the number of free bytes in the block containing `offset'
1020 * by the value specified by `diff'.
1022 * The caller must ensure we have exclusive access to `dh'; normally
1023 * that means that dh_lock should be held, but this is also called
1024 * from ufsdirhash_build() where exclusive access can be assumed.
1027 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff)
1029 int block, i, nfidx, ofidx;
1031 /* Update the per-block summary info. */
1032 block = offset / DIRBLKSIZ;
1033 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks,
1034 ("dirhash bad offset"));
1035 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1036 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
1037 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1039 /* Update the `first free' list if necessary. */
1040 if (ofidx != nfidx) {
1041 /* If removing, scan forward for the next block. */
1042 if (dh->dh_firstfree[ofidx] == block) {
1043 for (i = block + 1; i < dh->dh_dirblks; i++)
1044 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
1046 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
1049 /* Make this the new `first free' if necessary */
1050 if (dh->dh_firstfree[nfidx] > block ||
1051 dh->dh_firstfree[nfidx] == -1)
1052 dh->dh_firstfree[nfidx] = block;
1057 * Find the specified name which should have the specified offset.
1058 * Returns a slot number, and panics on failure.
1060 * `dh' must be locked on entry and remains so on return.
1063 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset)
1067 DIRHASH_ASSERT_LOCKED(dh);
1069 /* Find the entry. */
1070 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full"));
1071 slot = ufsdirhash_hash(dh, name, namelen);
1072 while (DH_ENTRY(dh, slot) != offset &&
1073 DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
1074 slot = WRAPINCR(slot, dh->dh_hlen);
1075 if (DH_ENTRY(dh, slot) != offset)
1076 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
1082 * Remove the entry corresponding to the specified slot from the hash array.
1084 * `dh' must be locked on entry and remains so on return.
1087 ufsdirhash_delslot(struct dirhash *dh, int slot)
1091 DIRHASH_ASSERT_LOCKED(dh);
1093 /* Mark the entry as deleted. */
1094 DH_ENTRY(dh, slot) = DIRHASH_DEL;
1096 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
1097 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
1098 i = WRAPINCR(i, dh->dh_hlen);
1099 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
1100 i = WRAPDECR(i, dh->dh_hlen);
1101 while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
1102 DH_ENTRY(dh, i) = DIRHASH_EMPTY;
1104 i = WRAPDECR(i, dh->dh_hlen);
1106 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen"));
1111 * Given a directory entry and its offset, find the offset of the
1112 * previous entry in the same DIRBLKSIZ-sized block. Returns an
1113 * offset, or -1 if there is no previous entry in the block or some
1114 * other problem occurred.
1117 ufsdirhash_getprev(struct direct *dirp, doff_t offset)
1121 doff_t blkoff, prevoff;
1124 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */
1125 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */
1126 blkbuf = (char *)dirp - entrypos;
1129 /* If `offset' is the start of a block, there is no previous entry. */
1133 /* Scan from the start of the block until we get to the entry. */
1134 for (i = 0; i < entrypos; i += dp->d_reclen) {
1135 dp = (struct direct *)(blkbuf + i);
1136 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
1137 return (-1); /* Corrupted directory. */
1138 prevoff = blkoff + i;
1144 * Try to free up `wanted' bytes by stealing memory from existing
1145 * dirhashes. Returns zero with list locked if successful.
1148 ufsdirhash_recycle(int wanted)
1153 int i, mem, narrays;
1156 dh = TAILQ_FIRST(&ufsdirhash_list);
1157 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
1158 /* Decrement the score; only recycle if it becomes zero. */
1159 if (dh == NULL || --dh->dh_score > 0) {
1160 DIRHASHLIST_UNLOCK();
1164 * If we can't lock it it's in use and we don't want to
1165 * recycle it anyway.
1167 if (!sx_try_xlock(&dh->dh_lock)) {
1168 dh = TAILQ_NEXT(dh, dh_list);
1171 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list"));
1173 /* Remove it from the list and detach its memory. */
1174 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
1178 blkfree = dh->dh_blkfree;
1179 dh->dh_blkfree = NULL;
1180 narrays = dh->dh_narrays;
1181 mem = dh->dh_memreq;
1184 /* Unlock everything, free the detached memory. */
1185 ufsdirhash_release(dh);
1186 DIRHASHLIST_UNLOCK();
1187 for (i = 0; i < narrays; i++)
1188 DIRHASH_BLKFREE(hash[i]);
1189 FREE(hash, M_DIRHASH);
1190 FREE(blkfree, M_DIRHASH);
1192 /* Account for the returned memory, and repeat if necessary. */
1194 ufs_dirhashmem -= mem;
1195 dh = TAILQ_FIRST(&ufsdirhash_list);
1197 /* Success; return with list locked. */
1205 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t),
1206 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1207 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF);
1208 TAILQ_INIT(&ufsdirhash_list);
1214 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit"));
1215 uma_zdestroy(ufsdirhash_zone);
1216 mtx_destroy(&ufsdirhash_mtx);
1219 #endif /* UFS_DIRHASH */