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>
52 #include <sys/eventhandler.h>
56 #include <ufs/ufs/quota.h>
57 #include <ufs/ufs/inode.h>
58 #include <ufs/ufs/dir.h>
59 #include <ufs/ufs/dirhash.h>
60 #include <ufs/ufs/extattr.h>
61 #include <ufs/ufs/ufsmount.h>
62 #include <ufs/ufs/ufs_extern.h>
64 #define WRAPINCR(val, limit) (((val) + 1 == (limit)) ? 0 : ((val) + 1))
65 #define WRAPDECR(val, limit) (((val) == 0) ? ((limit) - 1) : ((val) - 1))
66 #define OFSFMT(vp) ((vp)->v_mount->mnt_maxsymlinklen <= 0)
67 #define BLKFREE2IDX(n) ((n) > DH_NFSTATS ? DH_NFSTATS : (n))
69 static MALLOC_DEFINE(M_DIRHASH, "ufs_dirhash", "UFS directory hash tables");
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; /* NOTE: initial value. It is
76 tuned in ufsdirhash_init() */
77 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_maxmem, CTLFLAG_RW, &ufs_dirhashmaxmem,
78 0, "maximum allowed dirhash memory usage");
79 static int ufs_dirhashmem;
80 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_mem, CTLFLAG_RD, &ufs_dirhashmem,
81 0, "current dirhash memory usage");
82 static int ufs_dirhashcheck = 0;
83 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_docheck, CTLFLAG_RW, &ufs_dirhashcheck,
84 0, "enable extra sanity tests");
85 static int ufs_dirhashlowmemcount = 0;
86 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_lowmemcount, CTLFLAG_RD,
87 &ufs_dirhashlowmemcount, 0, "number of times low memory hook called");
88 static int ufs_dirhashreclaimage = 60;
89 SYSCTL_INT(_vfs_ufs, OID_AUTO, dirhash_reclaimage, CTLFLAG_RW,
90 &ufs_dirhashreclaimage, 0,
91 "max time in seconds of hash inactivity before deletion in low VM events");
94 static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen);
95 static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff);
96 static void ufsdirhash_delslot(struct dirhash *dh, int slot);
97 static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen,
99 static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset);
100 static int ufsdirhash_recycle(int wanted);
101 static void ufsdirhash_lowmem(void);
102 static void ufsdirhash_free_locked(struct inode *ip);
104 static uma_zone_t ufsdirhash_zone;
106 #define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx)
107 #define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx)
108 #define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK)
109 #define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr))
110 #define DIRHASH_ASSERT_LOCKED(dh) \
111 sx_assert(&(dh)->dh_lock, SA_LOCKED)
113 /* Dirhash list; recently-used entries are near the tail. */
114 static TAILQ_HEAD(, dirhash) ufsdirhash_list;
116 /* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */
117 static struct mtx ufsdirhash_mtx;
122 * The relationship between inode and dirhash is protected either by an
123 * exclusive vnode lock or the vnode interlock where a shared vnode lock
124 * may be used. The dirhash_mtx is acquired after the dirhash lock. To
125 * handle teardown races, code wishing to lock the dirhash for an inode
126 * when using a shared vnode lock must obtain a private reference on the
127 * dirhash while holding the vnode interlock. They can drop it once they
128 * have obtained the dirhash lock and verified that the dirhash wasn't
129 * recycled while they waited for the dirhash lock.
131 * ufsdirhash_build() acquires a shared lock on the dirhash when it is
132 * successful. This lock is released after a call to ufsdirhash_lookup().
134 * Functions requiring exclusive access use ufsdirhash_acquire() which may
135 * free a dirhash structure that was recycled by ufsdirhash_recycle().
137 * The dirhash lock may be held across io operations.
139 * WITNESS reports a lock order reversal between the "bufwait" lock
140 * and the "dirhash" lock. However, this specific reversal will not
141 * cause a deadlock. To get a deadlock, one would have to lock a
142 * buffer followed by the dirhash while a second thread locked a
143 * buffer while holding the dirhash lock. The second order can happen
144 * under a shared or exclusive vnode lock for the associated directory
145 * in lookup(). The first order, however, can only happen under an
146 * exclusive vnode lock (e.g. unlink(), rename(), etc.). Thus, for
147 * a thread to be doing a "bufwait" -> "dirhash" order, it has to hold
148 * an exclusive vnode lock. That exclusive vnode lock will prevent
149 * any other threads from doing a "dirhash" -> "bufwait" order.
153 ufsdirhash_hold(struct dirhash *dh)
156 refcount_acquire(&dh->dh_refcount);
160 ufsdirhash_drop(struct dirhash *dh)
163 if (refcount_release(&dh->dh_refcount)) {
164 sx_destroy(&dh->dh_lock);
170 * Release the lock on a dirhash.
173 ufsdirhash_release(struct dirhash *dh)
176 sx_unlock(&dh->dh_lock);
180 * Either acquire an existing hash locked shared or create a new hash and
181 * return it exclusively locked. May return NULL if the allocation fails.
183 * The vnode interlock is used to protect the i_dirhash pointer from
184 * simultaneous access while only a shared vnode lock is held.
186 static struct dirhash *
187 ufsdirhash_create(struct inode *ip)
196 /* Racy check for i_dirhash to prefetch a dirhash structure. */
197 if (ip->i_dirhash == NULL && ndh == NULL) {
198 ndh = malloc(sizeof *dh, M_DIRHASH,
202 refcount_init(&ndh->dh_refcount, 1);
205 * The DUPOK is to prevent warnings from the
206 * sx_slock() a few lines down which is safe
207 * since the duplicate lock in that case is
208 * the one for this dirhash we are creating
209 * now which has no external references until
210 * after this function returns.
212 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK);
213 sx_xlock(&ndh->dh_lock);
216 * Check i_dirhash. If it's NULL just try to use a
217 * preallocated structure. If none exists loop and try again.
231 /* Acquire a shared lock on existing hashes. */
232 sx_slock(&dh->dh_lock);
234 /* The hash could've been recycled while we were waiting. */
236 if (ip->i_dirhash != dh) {
238 ufsdirhash_release(dh);
245 /* If the hash is still valid we've succeeded. */
246 if (dh->dh_hash != NULL)
249 * If the hash is NULL it has been recycled. Try to upgrade
250 * so we can recreate it. If we fail the upgrade, drop our
251 * lock and try again.
253 if (sx_try_upgrade(&dh->dh_lock))
255 sx_sunlock(&dh->dh_lock);
257 /* Free the preallocated structure if it was not necessary. */
259 ufsdirhash_release(ndh);
260 ufsdirhash_drop(ndh);
266 * Acquire an exclusive lock on an existing hash. Requires an exclusive
267 * vnode lock to protect the i_dirhash pointer. hashes that have been
268 * recycled are reclaimed here and NULL is returned.
270 static struct dirhash *
271 ufsdirhash_acquire(struct inode *ip)
275 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__);
280 sx_xlock(&dh->dh_lock);
281 if (dh->dh_hash != NULL)
283 ufsdirhash_free_locked(ip);
288 * Acquire exclusively and free the hash pointed to by ip. Works with a
289 * shared or exclusive vnode lock.
292 ufsdirhash_free(struct inode *ip)
299 /* Grab a reference on this inode's dirhash if it has one. */
309 /* Exclusively lock the dirhash. */
310 sx_xlock(&dh->dh_lock);
312 /* If this dirhash still belongs to this inode, then free it. */
314 if (ip->i_dirhash == dh) {
322 * This inode's dirhash has changed while we were
323 * waiting for the dirhash lock, so try again.
325 ufsdirhash_release(dh);
328 ufsdirhash_free_locked(ip);
332 * Attempt to build up a hash table for the directory contents in
333 * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
336 ufsdirhash_build(struct inode *ip)
339 struct buf *bp = NULL;
343 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
345 /* Take care of a decreased sysctl value. */
346 while (ufs_dirhashmem > ufs_dirhashmaxmem) {
347 if (ufsdirhash_recycle(0) != 0)
349 /* Recycled enough memory, so unlock the list. */
350 DIRHASHLIST_UNLOCK();
353 /* Check if we can/should use dirhash. */
354 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) ||
355 ip->i_effnlink == 0) {
360 dh = ufsdirhash_create(ip);
363 if (dh->dh_hash != NULL)
367 /* Allocate 50% more entries than this dir size could ever need. */
368 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size"));
369 nslots = ip->i_size / DIRECTSIZ(1);
370 nslots = (nslots * 3 + 1) / 2;
371 narrays = howmany(nslots, DH_NBLKOFF);
372 nslots = narrays * DH_NBLKOFF;
373 dirblocks = howmany(ip->i_size, DIRBLKSIZ);
374 nblocks = (dirblocks * 3 + 1) / 2;
375 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
376 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
377 nblocks * sizeof(*dh->dh_blkfree);
379 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
380 DIRHASHLIST_UNLOCK();
381 if (memreqd > ufs_dirhashmaxmem / 2)
383 /* Try to free some space. */
384 if (ufsdirhash_recycle(memreqd) != 0)
386 /* Enough was freed, and list has been locked. */
388 ufs_dirhashmem += memreqd;
389 DIRHASHLIST_UNLOCK();
391 /* Initialise the hash table and block statistics. */
392 dh->dh_memreq = memreqd;
393 dh->dh_narrays = narrays;
394 dh->dh_hlen = nslots;
395 dh->dh_nblk = nblocks;
396 dh->dh_dirblks = dirblocks;
397 for (i = 0; i < DH_NFSTATS; i++)
398 dh->dh_firstfree[i] = -1;
399 dh->dh_firstfree[DH_NFSTATS] = 0;
402 dh->dh_score = DH_SCOREINIT;
403 dh->dh_lastused = time_second;
406 * Use non-blocking mallocs so that we will revert to a linear
407 * lookup on failure rather than potentially blocking forever.
409 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]),
410 M_DIRHASH, M_NOWAIT | M_ZERO);
411 if (dh->dh_hash == NULL)
413 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]),
414 M_DIRHASH, M_NOWAIT);
415 if (dh->dh_blkfree == NULL)
417 for (i = 0; i < narrays; i++) {
418 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL)
420 for (j = 0; j < DH_NBLKOFF; j++)
421 dh->dh_hash[i][j] = DIRHASH_EMPTY;
423 for (i = 0; i < dirblocks; i++)
424 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN;
425 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
427 while (pos < ip->i_size) {
428 /* If necessary, get the next directory block. */
429 if ((pos & bmask) == 0) {
432 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
436 /* Add this entry to the hash. */
437 ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
438 if (ep->d_reclen == 0 || ep->d_reclen >
439 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) {
440 /* Corrupted directory. */
444 if (ep->d_ino != 0) {
445 /* Add the entry (simplified ufsdirhash_add). */
446 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
447 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
448 slot = WRAPINCR(slot, dh->dh_hlen);
450 DH_ENTRY(dh, slot) = pos;
451 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep));
459 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
461 DIRHASHLIST_UNLOCK();
462 sx_downgrade(&dh->dh_lock);
466 ufsdirhash_free_locked(ip);
471 * Free any hash table associated with inode 'ip'.
474 ufsdirhash_free_locked(struct inode *ip)
480 DIRHASH_ASSERT_LOCKED(ip->i_dirhash);
483 * Clear the pointer in the inode to prevent new threads from
484 * finding the dead structure.
489 ip->i_dirhash = NULL;
493 * Remove the hash from the list since we are going to free its
498 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
499 ufs_dirhashmem -= dh->dh_memreq;
500 DIRHASHLIST_UNLOCK();
503 * At this point, any waiters for the lock should hold their
504 * own reference on the dirhash structure. They will drop
505 * that reference once they grab the vnode interlock and see
506 * that ip->i_dirhash is NULL.
508 sx_xunlock(&dh->dh_lock);
511 * Handle partially recycled as well as fully constructed hashes.
513 if (dh->dh_hash != NULL) {
514 for (i = 0; i < dh->dh_narrays; i++)
515 if (dh->dh_hash[i] != NULL)
516 DIRHASH_BLKFREE(dh->dh_hash[i]);
517 free(dh->dh_hash, M_DIRHASH);
518 if (dh->dh_blkfree != NULL)
519 free(dh->dh_blkfree, M_DIRHASH);
523 * Drop the inode's reference to the data structure.
529 * Find the offset of the specified name within the given inode.
530 * Returns 0 on success, ENOENT if the entry does not exist, or
531 * EJUSTRETURN if the caller should revert to a linear search.
533 * If successful, the directory offset is stored in *offp, and a
534 * pointer to a struct buf containing the entry is stored in *bpp. If
535 * prevoffp is non-NULL, the offset of the previous entry within
536 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
537 * is the first in a block, the start of the block is used).
539 * Must be called with the hash locked. Returns with the hash unlocked.
542 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp,
543 struct buf **bpp, doff_t *prevoffp)
545 struct dirhash *dh, *dh_next;
549 doff_t blkoff, bmask, offset, prevoff, seqoff;
554 KASSERT(dh != NULL && dh->dh_hash != NULL,
555 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh));
556 DIRHASH_ASSERT_LOCKED(dh);
558 * Move this dirhash towards the end of the list if it has a
559 * score higher than the next entry, and acquire the dh_lock.
562 if (TAILQ_NEXT(dh, dh_list) != NULL) {
564 * If the new score will be greater than that of the next
565 * entry, then move this entry past it. With both mutexes
566 * held, dh_next won't go away, but its dh_score could
567 * change; that's not important since it is just a hint.
569 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
570 dh->dh_score >= dh_next->dh_score) {
571 KASSERT(dh->dh_onlist, ("dirhash: not on list"));
572 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
573 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
577 /* Update the score. */
578 if (dh->dh_score < DH_SCOREMAX)
581 /* Update last used time. */
582 dh->dh_lastused = time_second;
583 DIRHASHLIST_UNLOCK();
586 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
589 seqoff = dh->dh_seqoff;
591 slot = ufsdirhash_hash(dh, name, namelen);
595 * Sequential access optimisation. seqoff contains the
596 * offset of the directory entry immediately following
597 * the last entry that was looked up. Check if this offset
598 * appears in the hash chain for the name we are looking for.
600 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
601 i = WRAPINCR(i, dh->dh_hlen))
602 if (offset == seqoff)
604 if (offset == seqoff) {
606 * We found an entry with the expected offset. This
607 * is probably the entry we want, but if not, the
608 * code below will retry.
615 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
616 slot = WRAPINCR(slot, dh->dh_hlen)) {
617 if (offset == DIRHASH_DEL)
619 if (offset < 0 || offset >= ip->i_size)
620 panic("ufsdirhash_lookup: bad offset in hash array");
621 if ((offset & ~bmask) != blkoff) {
624 blkoff = offset & ~bmask;
625 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) {
630 KASSERT(bp != NULL, ("no buffer allocated"));
631 dp = (struct direct *)(bp->b_data + (offset & bmask));
632 if (dp->d_reclen == 0 || dp->d_reclen >
633 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
634 /* Corrupted directory. */
638 if (dp->d_namlen == namelen &&
639 bcmp(dp->d_name, name, namelen) == 0) {
640 /* Found. Get the prev offset if needed. */
641 if (prevoffp != NULL) {
642 if (offset & (DIRBLKSIZ - 1)) {
643 prevoff = ufsdirhash_getprev(dp,
655 dh->dh_seqoff = offset + DIRSIZ(0, dp);
658 ufsdirhash_release(dh);
663 * When the name doesn't match in the sequential
664 * optimization case, go back and search normally.
673 ufsdirhash_release(dh);
680 * Find a directory block with room for 'slotneeded' bytes. Returns
681 * the offset of the directory entry that begins the free space.
682 * This will either be the offset of an existing entry that has free
683 * space at the end, or the offset of an entry with d_ino == 0 at
684 * the start of a DIRBLKSIZ block.
686 * To use the space, the caller may need to compact existing entries in
687 * the directory. The total number of bytes in all of the entries involved
688 * in the compaction is stored in *slotsize. In other words, all of
689 * the entries that must be compacted are exactly contained in the
690 * region beginning at the returned offset and spanning *slotsize bytes.
692 * Returns -1 if no space was found, indicating that the directory
696 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
701 doff_t pos, slotstart;
702 int dirblock, error, freebytes, i;
705 KASSERT(dh != NULL && dh->dh_hash != NULL,
706 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh));
707 DIRHASH_ASSERT_LOCKED(dh);
709 /* Find a directory block with the desired free space. */
711 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
712 if ((dirblock = dh->dh_firstfree[i]) != -1)
717 KASSERT(dirblock < dh->dh_nblk &&
718 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
719 ("ufsdirhash_findfree: bad stats"));
720 pos = dirblock * DIRBLKSIZ;
721 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp);
725 /* Find the first entry with free space. */
726 for (i = 0; i < DIRBLKSIZ; ) {
727 if (dp->d_reclen == 0) {
731 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
734 dp = (struct direct *)((char *)dp + dp->d_reclen);
742 /* Find the range of entries needed to get enough space */
744 while (i < DIRBLKSIZ && freebytes < slotneeded) {
745 freebytes += dp->d_reclen;
747 freebytes -= DIRSIZ(0, dp);
748 if (dp->d_reclen == 0) {
753 dp = (struct direct *)((char *)dp + dp->d_reclen);
759 if (freebytes < slotneeded)
760 panic("ufsdirhash_findfree: free mismatch");
762 *slotsize = pos + i - slotstart;
767 * Return the start of the unused space at the end of a directory, or
768 * -1 if there are no trailing unused blocks.
771 ufsdirhash_enduseful(struct inode *ip)
778 DIRHASH_ASSERT_LOCKED(dh);
779 KASSERT(dh != NULL && dh->dh_hash != NULL,
780 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh));
782 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN)
785 for (i = dh->dh_dirblks - 1; i >= 0; i--)
786 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
789 return ((doff_t)(i + 1) * DIRBLKSIZ);
793 * Insert information into the hash about a new directory entry. dirp
794 * points to a struct direct containing the entry, and offset specifies
795 * the offset of this entry.
798 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
803 if ((dh = ufsdirhash_acquire(ip)) == NULL)
806 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
807 ("ufsdirhash_add: bad offset"));
809 * Normal hash usage is < 66%. If the usage gets too high then
810 * remove the hash entirely and let it be rebuilt later.
812 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
813 ufsdirhash_free_locked(ip);
817 /* Find a free hash slot (empty or deleted), and add the entry. */
818 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
819 while (DH_ENTRY(dh, slot) >= 0)
820 slot = WRAPINCR(slot, dh->dh_hlen);
821 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
823 DH_ENTRY(dh, slot) = offset;
825 /* Update last used time. */
826 dh->dh_lastused = time_second;
828 /* Update the per-block summary info. */
829 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp));
830 ufsdirhash_release(dh);
834 * Remove the specified directory entry from the hash. The entry to remove
835 * is defined by the name in `dirp', which must exist at the specified
836 * `offset' within the directory.
839 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
844 if ((dh = ufsdirhash_acquire(ip)) == NULL)
847 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
848 ("ufsdirhash_remove: bad offset"));
850 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
852 /* Remove the hash entry. */
853 ufsdirhash_delslot(dh, slot);
855 /* Update the per-block summary info. */
856 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp));
857 ufsdirhash_release(dh);
861 * Change the offset associated with a directory entry in the hash. Used
862 * when compacting directory blocks.
865 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
871 if ((dh = ufsdirhash_acquire(ip)) == NULL)
874 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ &&
875 newoff < dh->dh_dirblks * DIRBLKSIZ,
876 ("ufsdirhash_move: bad offset"));
877 /* Find the entry, and update the offset. */
878 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
879 DH_ENTRY(dh, slot) = newoff;
880 ufsdirhash_release(dh);
884 * Inform dirhash that the directory has grown by one block that
885 * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
888 ufsdirhash_newblk(struct inode *ip, doff_t offset)
893 if ((dh = ufsdirhash_acquire(ip)) == NULL)
896 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ,
897 ("ufsdirhash_newblk: bad offset"));
898 block = offset / DIRBLKSIZ;
899 if (block >= dh->dh_nblk) {
900 /* Out of space; must rebuild. */
901 ufsdirhash_free_locked(ip);
904 dh->dh_dirblks = block + 1;
906 /* Account for the new free block. */
907 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN;
908 if (dh->dh_firstfree[DH_NFSTATS] == -1)
909 dh->dh_firstfree[DH_NFSTATS] = block;
910 ufsdirhash_release(dh);
914 * Inform dirhash that the directory is being truncated.
917 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
922 if ((dh = ufsdirhash_acquire(ip)) == NULL)
925 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ,
926 ("ufsdirhash_dirtrunc: bad offset"));
927 block = howmany(offset, DIRBLKSIZ);
929 * If the directory shrinks to less than 1/8 of dh_nblk blocks
930 * (about 20% of its original size due to the 50% extra added in
931 * ufsdirhash_build) then free it, and let the caller rebuild
934 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
935 ufsdirhash_free_locked(ip);
940 * Remove any `first free' information pertaining to the
941 * truncated blocks. All blocks we're removing should be
944 if (dh->dh_firstfree[DH_NFSTATS] >= block)
945 dh->dh_firstfree[DH_NFSTATS] = -1;
946 for (i = block; i < dh->dh_dirblks; i++)
947 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
948 panic("ufsdirhash_dirtrunc: blocks in use");
949 for (i = 0; i < DH_NFSTATS; i++)
950 if (dh->dh_firstfree[i] >= block)
951 panic("ufsdirhash_dirtrunc: first free corrupt");
952 dh->dh_dirblks = block;
953 ufsdirhash_release(dh);
957 * Debugging function to check that the dirhash information about
958 * a directory block matches its actual contents. Panics if a mismatch
961 * On entry, `buf' should point to the start of an in-core
962 * DIRBLKSIZ-sized directory block, and `offset' should contain the
963 * offset from the start of the directory of that block.
966 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
970 int block, ffslot, i, nfree;
972 if (!ufs_dirhashcheck)
974 if ((dh = ufsdirhash_acquire(ip)) == NULL)
977 block = offset / DIRBLKSIZ;
978 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks)
979 panic("ufsdirhash_checkblock: bad offset");
982 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) {
983 dp = (struct direct *)(buf + i);
984 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ)
985 panic("ufsdirhash_checkblock: bad dir");
987 if (dp->d_ino == 0) {
990 * XXX entries with d_ino == 0 should only occur
991 * at the start of a DIRBLKSIZ block. However the
992 * ufs code is tolerant of such entries at other
993 * offsets, and fsck does not fix them.
996 panic("ufsdirhash_checkblock: bad dir inode");
998 nfree += dp->d_reclen;
1002 /* Check that the entry exists (will panic if it doesn't). */
1003 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
1005 nfree += dp->d_reclen - DIRSIZ(0, dp);
1008 panic("ufsdirhash_checkblock: bad dir end");
1010 if (dh->dh_blkfree[block] * DIRALIGN != nfree)
1011 panic("ufsdirhash_checkblock: bad free count");
1013 ffslot = BLKFREE2IDX(nfree / DIRALIGN);
1014 for (i = 0; i <= DH_NFSTATS; i++)
1015 if (dh->dh_firstfree[i] == block && i != ffslot)
1016 panic("ufsdirhash_checkblock: bad first-free");
1017 if (dh->dh_firstfree[ffslot] == -1)
1018 panic("ufsdirhash_checkblock: missing first-free entry");
1019 ufsdirhash_release(dh);
1023 * Hash the specified filename into a dirhash slot.
1026 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen)
1031 * We hash the name and then some other bit of data that is
1032 * invariant over the dirhash's lifetime. Otherwise names
1033 * differing only in the last byte are placed close to one
1034 * another in the table, which is bad for linear probing.
1036 hash = fnv_32_buf(name, namelen, FNV1_32_INIT);
1037 hash = fnv_32_buf(&dh, sizeof(dh), hash);
1038 return (hash % dh->dh_hlen);
1042 * Adjust the number of free bytes in the block containing `offset'
1043 * by the value specified by `diff'.
1045 * The caller must ensure we have exclusive access to `dh'; normally
1046 * that means that dh_lock should be held, but this is also called
1047 * from ufsdirhash_build() where exclusive access can be assumed.
1050 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff)
1052 int block, i, nfidx, ofidx;
1054 /* Update the per-block summary info. */
1055 block = offset / DIRBLKSIZ;
1056 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks,
1057 ("dirhash bad offset"));
1058 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1059 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
1060 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1062 /* Update the `first free' list if necessary. */
1063 if (ofidx != nfidx) {
1064 /* If removing, scan forward for the next block. */
1065 if (dh->dh_firstfree[ofidx] == block) {
1066 for (i = block + 1; i < dh->dh_dirblks; i++)
1067 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
1069 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
1072 /* Make this the new `first free' if necessary */
1073 if (dh->dh_firstfree[nfidx] > block ||
1074 dh->dh_firstfree[nfidx] == -1)
1075 dh->dh_firstfree[nfidx] = block;
1080 * Find the specified name which should have the specified offset.
1081 * Returns a slot number, and panics on failure.
1083 * `dh' must be locked on entry and remains so on return.
1086 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset)
1090 DIRHASH_ASSERT_LOCKED(dh);
1092 /* Find the entry. */
1093 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full"));
1094 slot = ufsdirhash_hash(dh, name, namelen);
1095 while (DH_ENTRY(dh, slot) != offset &&
1096 DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
1097 slot = WRAPINCR(slot, dh->dh_hlen);
1098 if (DH_ENTRY(dh, slot) != offset)
1099 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
1105 * Remove the entry corresponding to the specified slot from the hash array.
1107 * `dh' must be locked on entry and remains so on return.
1110 ufsdirhash_delslot(struct dirhash *dh, int slot)
1114 DIRHASH_ASSERT_LOCKED(dh);
1116 /* Mark the entry as deleted. */
1117 DH_ENTRY(dh, slot) = DIRHASH_DEL;
1119 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
1120 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
1121 i = WRAPINCR(i, dh->dh_hlen);
1122 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
1123 i = WRAPDECR(i, dh->dh_hlen);
1124 while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
1125 DH_ENTRY(dh, i) = DIRHASH_EMPTY;
1127 i = WRAPDECR(i, dh->dh_hlen);
1129 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen"));
1134 * Given a directory entry and its offset, find the offset of the
1135 * previous entry in the same DIRBLKSIZ-sized block. Returns an
1136 * offset, or -1 if there is no previous entry in the block or some
1137 * other problem occurred.
1140 ufsdirhash_getprev(struct direct *dirp, doff_t offset)
1144 doff_t blkoff, prevoff;
1147 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */
1148 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */
1149 blkbuf = (char *)dirp - entrypos;
1152 /* If `offset' is the start of a block, there is no previous entry. */
1156 /* Scan from the start of the block until we get to the entry. */
1157 for (i = 0; i < entrypos; i += dp->d_reclen) {
1158 dp = (struct direct *)(blkbuf + i);
1159 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
1160 return (-1); /* Corrupted directory. */
1161 prevoff = blkoff + i;
1167 * Delete the given dirhash and reclaim its memory. Assumes that
1168 * ufsdirhash_list is locked, and leaves it locked. Also assumes
1169 * that dh is locked. Returns the amount of memory freed.
1172 ufsdirhash_destroy(struct dirhash *dh)
1176 int i, mem, narrays;
1178 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list"));
1180 /* Remove it from the list and detach its memory. */
1181 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
1185 blkfree = dh->dh_blkfree;
1186 dh->dh_blkfree = NULL;
1187 narrays = dh->dh_narrays;
1188 mem = dh->dh_memreq;
1191 /* Unlock dirhash and free the detached memory. */
1192 ufsdirhash_release(dh);
1193 for (i = 0; i < narrays; i++)
1194 DIRHASH_BLKFREE(hash[i]);
1195 free(hash, M_DIRHASH);
1196 free(blkfree, M_DIRHASH);
1198 /* Account for the returned memory. */
1199 ufs_dirhashmem -= mem;
1205 * Try to free up `wanted' bytes by stealing memory from existing
1206 * dirhashes. Returns zero with list locked if successful.
1209 ufsdirhash_recycle(int wanted)
1214 dh = TAILQ_FIRST(&ufsdirhash_list);
1215 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
1216 /* Decrement the score; only recycle if it becomes zero. */
1217 if (dh == NULL || --dh->dh_score > 0) {
1218 DIRHASHLIST_UNLOCK();
1222 * If we can't lock it it's in use and we don't want to
1223 * recycle it anyway.
1225 if (!sx_try_xlock(&dh->dh_lock)) {
1226 dh = TAILQ_NEXT(dh, dh_list);
1230 ufsdirhash_destroy(dh);
1232 /* Repeat if necessary. */
1233 dh = TAILQ_FIRST(&ufsdirhash_list);
1235 /* Success; return with list locked. */
1240 * Callback that frees some dirhashes when the system is low on virtual memory.
1245 struct dirhash *dh, *dh_temp;
1248 * Will free a *minimum* of 10% of the dirhash, but possibly much
1249 * more (depending on dirhashreclaimage). System with large dirhashes
1250 * probably also need a much larger dirhashreclaimage.
1251 * XXX: this percentage may need to be adjusted.
1253 int memwanted = ufs_dirhashmem / 10;
1255 ufs_dirhashlowmemcount++;
1259 * Delete dirhashes not used for more than ufs_dirhashreclaimage
1260 * seconds. If we can't get a lock on the dirhash, it will be skipped.
1262 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) {
1263 if (!sx_try_xlock(&dh->dh_lock))
1265 if (time_second - dh->dh_lastused > ufs_dirhashreclaimage)
1266 memfreed += ufsdirhash_destroy(dh);
1267 /* Unlock if we didn't delete the dirhash */
1269 ufsdirhash_release(dh);
1273 * If not enough memory was freed, keep deleting hashes from the head
1274 * of the dirhash list. The ones closest to the head should be the
1277 if (memfreed < memwanted) {
1278 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) {
1279 if (!sx_try_xlock(&dh->dh_lock))
1281 memfreed += ufsdirhash_destroy(dh);
1282 if (memfreed >= memwanted)
1286 DIRHASHLIST_UNLOCK();
1293 ufs_dirhashmaxmem = lmax(roundup(hibufspace / 64, PAGE_SIZE),
1296 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t),
1297 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1298 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF);
1299 TAILQ_INIT(&ufsdirhash_list);
1301 /* Register a callback function to handle low memory signals */
1302 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL,
1303 EVENTHANDLER_PRI_FIRST);
1309 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit"));
1310 uma_zdestroy(ufsdirhash_zone);
1311 mtx_destroy(&ufsdirhash_mtx);
1314 #endif /* UFS_DIRHASH */