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_dirhashreclaimpercent = 10;
89 static int ufsdirhash_set_reclaimpercent(SYSCTL_HANDLER_ARGS);
90 SYSCTL_PROC(_vfs_ufs, OID_AUTO, dirhash_reclaimpercent,
91 CTLTYPE_INT | CTLFLAG_RW, 0, 0, ufsdirhash_set_reclaimpercent, "I",
92 "set percentage of dirhash cache to be removed in low VM events");
95 static int ufsdirhash_hash(struct dirhash *dh, char *name, int namelen);
96 static void ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff);
97 static void ufsdirhash_delslot(struct dirhash *dh, int slot);
98 static int ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen,
100 static doff_t ufsdirhash_getprev(struct direct *dp, doff_t offset);
101 static int ufsdirhash_recycle(int wanted);
102 static void ufsdirhash_lowmem(void);
103 static void ufsdirhash_free_locked(struct inode *ip);
105 static uma_zone_t ufsdirhash_zone;
107 #define DIRHASHLIST_LOCK() mtx_lock(&ufsdirhash_mtx)
108 #define DIRHASHLIST_UNLOCK() mtx_unlock(&ufsdirhash_mtx)
109 #define DIRHASH_BLKALLOC_WAITOK() uma_zalloc(ufsdirhash_zone, M_WAITOK)
110 #define DIRHASH_BLKFREE(ptr) uma_zfree(ufsdirhash_zone, (ptr))
111 #define DIRHASH_ASSERT_LOCKED(dh) \
112 sx_assert(&(dh)->dh_lock, SA_LOCKED)
114 /* Dirhash list; recently-used entries are near the tail. */
115 static TAILQ_HEAD(, dirhash) ufsdirhash_list;
117 /* Protects: ufsdirhash_list, `dh_list' field, ufs_dirhashmem. */
118 static struct mtx ufsdirhash_mtx;
123 * The relationship between inode and dirhash is protected either by an
124 * exclusive vnode lock or the vnode interlock where a shared vnode lock
125 * may be used. The dirhash_mtx is acquired after the dirhash lock. To
126 * handle teardown races, code wishing to lock the dirhash for an inode
127 * when using a shared vnode lock must obtain a private reference on the
128 * dirhash while holding the vnode interlock. They can drop it once they
129 * have obtained the dirhash lock and verified that the dirhash wasn't
130 * recycled while they waited for the dirhash lock.
132 * ufsdirhash_build() acquires a shared lock on the dirhash when it is
133 * successful. This lock is released after a call to ufsdirhash_lookup().
135 * Functions requiring exclusive access use ufsdirhash_acquire() which may
136 * free a dirhash structure that was recycled by ufsdirhash_recycle().
138 * The dirhash lock may be held across io operations.
140 * WITNESS reports a lock order reversal between the "bufwait" lock
141 * and the "dirhash" lock. However, this specific reversal will not
142 * cause a deadlock. To get a deadlock, one would have to lock a
143 * buffer followed by the dirhash while a second thread locked a
144 * buffer while holding the dirhash lock. The second order can happen
145 * under a shared or exclusive vnode lock for the associated directory
146 * in lookup(). The first order, however, can only happen under an
147 * exclusive vnode lock (e.g. unlink(), rename(), etc.). Thus, for
148 * a thread to be doing a "bufwait" -> "dirhash" order, it has to hold
149 * an exclusive vnode lock. That exclusive vnode lock will prevent
150 * any other threads from doing a "dirhash" -> "bufwait" order.
154 ufsdirhash_hold(struct dirhash *dh)
157 refcount_acquire(&dh->dh_refcount);
161 ufsdirhash_drop(struct dirhash *dh)
164 if (refcount_release(&dh->dh_refcount)) {
165 sx_destroy(&dh->dh_lock);
171 * Release the lock on a dirhash.
174 ufsdirhash_release(struct dirhash *dh)
177 sx_unlock(&dh->dh_lock);
181 * Either acquire an existing hash locked shared or create a new hash and
182 * return it exclusively locked. May return NULL if the allocation fails.
184 * The vnode interlock is used to protect the i_dirhash pointer from
185 * simultaneous access while only a shared vnode lock is held.
187 static struct dirhash *
188 ufsdirhash_create(struct inode *ip)
199 /* Racy check for i_dirhash to prefetch a dirhash structure. */
200 if (ip->i_dirhash == NULL && ndh == NULL) {
201 ndh = malloc(sizeof *dh, M_DIRHASH,
205 refcount_init(&ndh->dh_refcount, 1);
208 * The DUPOK is to prevent warnings from the
209 * sx_slock() a few lines down which is safe
210 * since the duplicate lock in that case is
211 * the one for this dirhash we are creating
212 * now which has no external references until
213 * after this function returns.
215 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK);
216 sx_xlock(&ndh->dh_lock);
219 * Check i_dirhash. If it's NULL just try to use a
220 * preallocated structure. If none exists loop and try again.
234 /* Acquire a shared lock on existing hashes. */
235 sx_slock(&dh->dh_lock);
237 /* The hash could've been recycled while we were waiting. */
239 if (ip->i_dirhash != dh) {
241 ufsdirhash_release(dh);
248 /* If the hash is still valid we've succeeded. */
249 if (dh->dh_hash != NULL)
252 * If the hash is NULL it has been recycled. Try to upgrade
253 * so we can recreate it. If we fail the upgrade, drop our
254 * lock and try again.
256 if (sx_try_upgrade(&dh->dh_lock))
258 sx_sunlock(&dh->dh_lock);
260 /* Free the preallocated structure if it was not necessary. */
262 ufsdirhash_release(ndh);
263 ufsdirhash_drop(ndh);
269 * Acquire an exclusive lock on an existing hash. Requires an exclusive
270 * vnode lock to protect the i_dirhash pointer. hashes that have been
271 * recycled are reclaimed here and NULL is returned.
273 static struct dirhash *
274 ufsdirhash_acquire(struct inode *ip)
279 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__);
285 sx_xlock(&dh->dh_lock);
286 if (dh->dh_hash != NULL)
288 ufsdirhash_free_locked(ip);
293 * Acquire exclusively and free the hash pointed to by ip. Works with a
294 * shared or exclusive vnode lock.
297 ufsdirhash_free(struct inode *ip)
304 /* Grab a reference on this inode's dirhash if it has one. */
314 /* Exclusively lock the dirhash. */
315 sx_xlock(&dh->dh_lock);
317 /* If this dirhash still belongs to this inode, then free it. */
319 if (ip->i_dirhash == dh) {
327 * This inode's dirhash has changed while we were
328 * waiting for the dirhash lock, so try again.
330 ufsdirhash_release(dh);
333 ufsdirhash_free_locked(ip);
337 * Attempt to build up a hash table for the directory contents in
338 * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
341 ufsdirhash_build(struct inode *ip)
344 struct buf *bp = NULL;
348 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
350 /* Take care of a decreased sysctl value. */
351 while (ufs_dirhashmem > ufs_dirhashmaxmem) {
352 if (ufsdirhash_recycle(0) != 0)
354 /* Recycled enough memory, so unlock the list. */
355 DIRHASHLIST_UNLOCK();
358 /* Check if we can/should use dirhash. */
359 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) ||
360 ip->i_effnlink == 0) {
365 dh = ufsdirhash_create(ip);
368 if (dh->dh_hash != NULL)
372 /* Allocate 50% more entries than this dir size could ever need. */
373 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size"));
374 nslots = ip->i_size / DIRECTSIZ(1);
375 nslots = (nslots * 3 + 1) / 2;
376 narrays = howmany(nslots, DH_NBLKOFF);
377 nslots = narrays * DH_NBLKOFF;
378 dirblocks = howmany(ip->i_size, DIRBLKSIZ);
379 nblocks = (dirblocks * 3 + 1) / 2;
380 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
381 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
382 nblocks * sizeof(*dh->dh_blkfree);
384 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
385 DIRHASHLIST_UNLOCK();
386 if (memreqd > ufs_dirhashmaxmem / 2)
388 /* Try to free some space. */
389 if (ufsdirhash_recycle(memreqd) != 0)
391 /* Enough was freed, and list has been locked. */
393 ufs_dirhashmem += memreqd;
394 DIRHASHLIST_UNLOCK();
396 /* Initialise the hash table and block statistics. */
397 dh->dh_memreq = memreqd;
398 dh->dh_narrays = narrays;
399 dh->dh_hlen = nslots;
400 dh->dh_nblk = nblocks;
401 dh->dh_dirblks = dirblocks;
402 for (i = 0; i < DH_NFSTATS; i++)
403 dh->dh_firstfree[i] = -1;
404 dh->dh_firstfree[DH_NFSTATS] = 0;
407 dh->dh_score = DH_SCOREINIT;
408 dh->dh_lastused = time_second;
411 * Use non-blocking mallocs so that we will revert to a linear
412 * lookup on failure rather than potentially blocking forever.
414 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]),
415 M_DIRHASH, M_NOWAIT | M_ZERO);
416 if (dh->dh_hash == NULL)
418 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]),
419 M_DIRHASH, M_NOWAIT);
420 if (dh->dh_blkfree == NULL)
422 for (i = 0; i < narrays; i++) {
423 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL)
425 for (j = 0; j < DH_NBLKOFF; j++)
426 dh->dh_hash[i][j] = DIRHASH_EMPTY;
428 for (i = 0; i < dirblocks; i++)
429 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN;
430 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
432 while (pos < ip->i_size) {
433 /* If necessary, get the next directory block. */
434 if ((pos & bmask) == 0) {
437 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
441 /* Add this entry to the hash. */
442 ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
443 if (ep->d_reclen == 0 || ep->d_reclen >
444 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) {
445 /* Corrupted directory. */
449 if (ep->d_ino != 0) {
450 /* Add the entry (simplified ufsdirhash_add). */
451 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
452 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
453 slot = WRAPINCR(slot, dh->dh_hlen);
455 DH_ENTRY(dh, slot) = pos;
456 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep));
464 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
466 DIRHASHLIST_UNLOCK();
467 sx_downgrade(&dh->dh_lock);
471 ufsdirhash_free_locked(ip);
476 * Free any hash table associated with inode 'ip'.
479 ufsdirhash_free_locked(struct inode *ip)
485 DIRHASH_ASSERT_LOCKED(ip->i_dirhash);
488 * Clear the pointer in the inode to prevent new threads from
489 * finding the dead structure.
494 ip->i_dirhash = NULL;
498 * Remove the hash from the list since we are going to free its
503 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
504 ufs_dirhashmem -= dh->dh_memreq;
505 DIRHASHLIST_UNLOCK();
508 * At this point, any waiters for the lock should hold their
509 * own reference on the dirhash structure. They will drop
510 * that reference once they grab the vnode interlock and see
511 * that ip->i_dirhash is NULL.
513 sx_xunlock(&dh->dh_lock);
516 * Handle partially recycled as well as fully constructed hashes.
518 if (dh->dh_hash != NULL) {
519 for (i = 0; i < dh->dh_narrays; i++)
520 if (dh->dh_hash[i] != NULL)
521 DIRHASH_BLKFREE(dh->dh_hash[i]);
522 free(dh->dh_hash, M_DIRHASH);
523 if (dh->dh_blkfree != NULL)
524 free(dh->dh_blkfree, M_DIRHASH);
528 * Drop the inode's reference to the data structure.
534 * Find the offset of the specified name within the given inode.
535 * Returns 0 on success, ENOENT if the entry does not exist, or
536 * EJUSTRETURN if the caller should revert to a linear search.
538 * If successful, the directory offset is stored in *offp, and a
539 * pointer to a struct buf containing the entry is stored in *bpp. If
540 * prevoffp is non-NULL, the offset of the previous entry within
541 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
542 * is the first in a block, the start of the block is used).
544 * Must be called with the hash locked. Returns with the hash unlocked.
547 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp,
548 struct buf **bpp, doff_t *prevoffp)
550 struct dirhash *dh, *dh_next;
554 doff_t blkoff, bmask, offset, prevoff, seqoff;
559 KASSERT(dh != NULL && dh->dh_hash != NULL,
560 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh));
561 DIRHASH_ASSERT_LOCKED(dh);
563 * Move this dirhash towards the end of the list if it has a
564 * score higher than the next entry, and acquire the dh_lock.
567 if (TAILQ_NEXT(dh, dh_list) != NULL) {
569 * If the new score will be greater than that of the next
570 * entry, then move this entry past it. With both mutexes
571 * held, dh_next won't go away, but its dh_score could
572 * change; that's not important since it is just a hint.
574 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
575 dh->dh_score >= dh_next->dh_score) {
576 KASSERT(dh->dh_onlist, ("dirhash: not on list"));
577 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
578 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
582 /* Update the score. */
583 if (dh->dh_score < DH_SCOREMAX)
586 /* Update last used time. */
587 dh->dh_lastused = time_second;
588 DIRHASHLIST_UNLOCK();
591 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
594 seqoff = dh->dh_seqoff;
596 slot = ufsdirhash_hash(dh, name, namelen);
600 * Sequential access optimisation. seqoff contains the
601 * offset of the directory entry immediately following
602 * the last entry that was looked up. Check if this offset
603 * appears in the hash chain for the name we are looking for.
605 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
606 i = WRAPINCR(i, dh->dh_hlen))
607 if (offset == seqoff)
609 if (offset == seqoff) {
611 * We found an entry with the expected offset. This
612 * is probably the entry we want, but if not, the
613 * code below will retry.
620 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
621 slot = WRAPINCR(slot, dh->dh_hlen)) {
622 if (offset == DIRHASH_DEL)
624 if (offset < 0 || offset >= ip->i_size)
625 panic("ufsdirhash_lookup: bad offset in hash array");
626 if ((offset & ~bmask) != blkoff) {
629 blkoff = offset & ~bmask;
630 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) {
635 KASSERT(bp != NULL, ("no buffer allocated"));
636 dp = (struct direct *)(bp->b_data + (offset & bmask));
637 if (dp->d_reclen == 0 || dp->d_reclen >
638 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
639 /* Corrupted directory. */
643 if (dp->d_namlen == namelen &&
644 bcmp(dp->d_name, name, namelen) == 0) {
645 /* Found. Get the prev offset if needed. */
646 if (prevoffp != NULL) {
647 if (offset & (DIRBLKSIZ - 1)) {
648 prevoff = ufsdirhash_getprev(dp,
660 dh->dh_seqoff = offset + DIRSIZ(0, dp);
663 ufsdirhash_release(dh);
668 * When the name doesn't match in the sequential
669 * optimization case, go back and search normally.
678 ufsdirhash_release(dh);
685 * Find a directory block with room for 'slotneeded' bytes. Returns
686 * the offset of the directory entry that begins the free space.
687 * This will either be the offset of an existing entry that has free
688 * space at the end, or the offset of an entry with d_ino == 0 at
689 * the start of a DIRBLKSIZ block.
691 * To use the space, the caller may need to compact existing entries in
692 * the directory. The total number of bytes in all of the entries involved
693 * in the compaction is stored in *slotsize. In other words, all of
694 * the entries that must be compacted are exactly contained in the
695 * region beginning at the returned offset and spanning *slotsize bytes.
697 * Returns -1 if no space was found, indicating that the directory
701 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
706 doff_t pos, slotstart;
707 int dirblock, error, freebytes, i;
710 KASSERT(dh != NULL && dh->dh_hash != NULL,
711 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh));
712 DIRHASH_ASSERT_LOCKED(dh);
714 /* Find a directory block with the desired free space. */
716 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
717 if ((dirblock = dh->dh_firstfree[i]) != -1)
722 KASSERT(dirblock < dh->dh_nblk &&
723 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
724 ("ufsdirhash_findfree: bad stats"));
725 pos = dirblock * DIRBLKSIZ;
726 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp);
730 /* Find the first entry with free space. */
731 for (i = 0; i < DIRBLKSIZ; ) {
732 if (dp->d_reclen == 0) {
736 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
739 dp = (struct direct *)((char *)dp + dp->d_reclen);
747 /* Find the range of entries needed to get enough space */
749 while (i < DIRBLKSIZ && freebytes < slotneeded) {
750 freebytes += dp->d_reclen;
752 freebytes -= DIRSIZ(0, dp);
753 if (dp->d_reclen == 0) {
758 dp = (struct direct *)((char *)dp + dp->d_reclen);
764 if (freebytes < slotneeded)
765 panic("ufsdirhash_findfree: free mismatch");
767 *slotsize = pos + i - slotstart;
772 * Return the start of the unused space at the end of a directory, or
773 * -1 if there are no trailing unused blocks.
776 ufsdirhash_enduseful(struct inode *ip)
783 DIRHASH_ASSERT_LOCKED(dh);
784 KASSERT(dh != NULL && dh->dh_hash != NULL,
785 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh));
787 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN)
790 for (i = dh->dh_dirblks - 1; i >= 0; i--)
791 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
794 return ((doff_t)(i + 1) * DIRBLKSIZ);
798 * Insert information into the hash about a new directory entry. dirp
799 * points to a struct direct containing the entry, and offset specifies
800 * the offset of this entry.
803 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
808 if ((dh = ufsdirhash_acquire(ip)) == NULL)
811 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
812 ("ufsdirhash_add: bad offset"));
814 * Normal hash usage is < 66%. If the usage gets too high then
815 * remove the hash entirely and let it be rebuilt later.
817 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
818 ufsdirhash_free_locked(ip);
822 /* Find a free hash slot (empty or deleted), and add the entry. */
823 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
824 while (DH_ENTRY(dh, slot) >= 0)
825 slot = WRAPINCR(slot, dh->dh_hlen);
826 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
828 DH_ENTRY(dh, slot) = offset;
830 /* Update last used time. */
831 dh->dh_lastused = time_second;
833 /* Update the per-block summary info. */
834 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp));
835 ufsdirhash_release(dh);
839 * Remove the specified directory entry from the hash. The entry to remove
840 * is defined by the name in `dirp', which must exist at the specified
841 * `offset' within the directory.
844 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
849 if ((dh = ufsdirhash_acquire(ip)) == NULL)
852 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
853 ("ufsdirhash_remove: bad offset"));
855 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
857 /* Remove the hash entry. */
858 ufsdirhash_delslot(dh, slot);
860 /* Update the per-block summary info. */
861 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp));
862 ufsdirhash_release(dh);
866 * Change the offset associated with a directory entry in the hash. Used
867 * when compacting directory blocks.
870 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
876 if ((dh = ufsdirhash_acquire(ip)) == NULL)
879 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ &&
880 newoff < dh->dh_dirblks * DIRBLKSIZ,
881 ("ufsdirhash_move: bad offset"));
882 /* Find the entry, and update the offset. */
883 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
884 DH_ENTRY(dh, slot) = newoff;
885 ufsdirhash_release(dh);
889 * Inform dirhash that the directory has grown by one block that
890 * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
893 ufsdirhash_newblk(struct inode *ip, doff_t offset)
898 if ((dh = ufsdirhash_acquire(ip)) == NULL)
901 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ,
902 ("ufsdirhash_newblk: bad offset"));
903 block = offset / DIRBLKSIZ;
904 if (block >= dh->dh_nblk) {
905 /* Out of space; must rebuild. */
906 ufsdirhash_free_locked(ip);
909 dh->dh_dirblks = block + 1;
911 /* Account for the new free block. */
912 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN;
913 if (dh->dh_firstfree[DH_NFSTATS] == -1)
914 dh->dh_firstfree[DH_NFSTATS] = block;
915 ufsdirhash_release(dh);
919 * Inform dirhash that the directory is being truncated.
922 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
927 if ((dh = ufsdirhash_acquire(ip)) == NULL)
930 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ,
931 ("ufsdirhash_dirtrunc: bad offset"));
932 block = howmany(offset, DIRBLKSIZ);
934 * If the directory shrinks to less than 1/8 of dh_nblk blocks
935 * (about 20% of its original size due to the 50% extra added in
936 * ufsdirhash_build) then free it, and let the caller rebuild
939 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
940 ufsdirhash_free_locked(ip);
945 * Remove any `first free' information pertaining to the
946 * truncated blocks. All blocks we're removing should be
949 if (dh->dh_firstfree[DH_NFSTATS] >= block)
950 dh->dh_firstfree[DH_NFSTATS] = -1;
951 for (i = block; i < dh->dh_dirblks; i++)
952 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
953 panic("ufsdirhash_dirtrunc: blocks in use");
954 for (i = 0; i < DH_NFSTATS; i++)
955 if (dh->dh_firstfree[i] >= block)
956 panic("ufsdirhash_dirtrunc: first free corrupt");
957 dh->dh_dirblks = block;
958 ufsdirhash_release(dh);
962 * Debugging function to check that the dirhash information about
963 * a directory block matches its actual contents. Panics if a mismatch
966 * On entry, `buf' should point to the start of an in-core
967 * DIRBLKSIZ-sized directory block, and `offset' should contain the
968 * offset from the start of the directory of that block.
971 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
975 int block, ffslot, i, nfree;
977 if (!ufs_dirhashcheck)
979 if ((dh = ufsdirhash_acquire(ip)) == NULL)
982 block = offset / DIRBLKSIZ;
983 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks)
984 panic("ufsdirhash_checkblock: bad offset");
987 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) {
988 dp = (struct direct *)(buf + i);
989 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ)
990 panic("ufsdirhash_checkblock: bad dir");
992 if (dp->d_ino == 0) {
995 * XXX entries with d_ino == 0 should only occur
996 * at the start of a DIRBLKSIZ block. However the
997 * ufs code is tolerant of such entries at other
998 * offsets, and fsck does not fix them.
1001 panic("ufsdirhash_checkblock: bad dir inode");
1003 nfree += dp->d_reclen;
1007 /* Check that the entry exists (will panic if it doesn't). */
1008 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
1010 nfree += dp->d_reclen - DIRSIZ(0, dp);
1013 panic("ufsdirhash_checkblock: bad dir end");
1015 if (dh->dh_blkfree[block] * DIRALIGN != nfree)
1016 panic("ufsdirhash_checkblock: bad free count");
1018 ffslot = BLKFREE2IDX(nfree / DIRALIGN);
1019 for (i = 0; i <= DH_NFSTATS; i++)
1020 if (dh->dh_firstfree[i] == block && i != ffslot)
1021 panic("ufsdirhash_checkblock: bad first-free");
1022 if (dh->dh_firstfree[ffslot] == -1)
1023 panic("ufsdirhash_checkblock: missing first-free entry");
1024 ufsdirhash_release(dh);
1028 * Hash the specified filename into a dirhash slot.
1031 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen)
1036 * We hash the name and then some other bit of data that is
1037 * invariant over the dirhash's lifetime. Otherwise names
1038 * differing only in the last byte are placed close to one
1039 * another in the table, which is bad for linear probing.
1041 hash = fnv_32_buf(name, namelen, FNV1_32_INIT);
1042 hash = fnv_32_buf(&dh, sizeof(dh), hash);
1043 return (hash % dh->dh_hlen);
1047 * Adjust the number of free bytes in the block containing `offset'
1048 * by the value specified by `diff'.
1050 * The caller must ensure we have exclusive access to `dh'; normally
1051 * that means that dh_lock should be held, but this is also called
1052 * from ufsdirhash_build() where exclusive access can be assumed.
1055 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff)
1057 int block, i, nfidx, ofidx;
1059 /* Update the per-block summary info. */
1060 block = offset / DIRBLKSIZ;
1061 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks,
1062 ("dirhash bad offset"));
1063 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1064 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
1065 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1067 /* Update the `first free' list if necessary. */
1068 if (ofidx != nfidx) {
1069 /* If removing, scan forward for the next block. */
1070 if (dh->dh_firstfree[ofidx] == block) {
1071 for (i = block + 1; i < dh->dh_dirblks; i++)
1072 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
1074 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
1077 /* Make this the new `first free' if necessary */
1078 if (dh->dh_firstfree[nfidx] > block ||
1079 dh->dh_firstfree[nfidx] == -1)
1080 dh->dh_firstfree[nfidx] = block;
1085 * Find the specified name which should have the specified offset.
1086 * Returns a slot number, and panics on failure.
1088 * `dh' must be locked on entry and remains so on return.
1091 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset)
1095 DIRHASH_ASSERT_LOCKED(dh);
1097 /* Find the entry. */
1098 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full"));
1099 slot = ufsdirhash_hash(dh, name, namelen);
1100 while (DH_ENTRY(dh, slot) != offset &&
1101 DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
1102 slot = WRAPINCR(slot, dh->dh_hlen);
1103 if (DH_ENTRY(dh, slot) != offset)
1104 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
1110 * Remove the entry corresponding to the specified slot from the hash array.
1112 * `dh' must be locked on entry and remains so on return.
1115 ufsdirhash_delslot(struct dirhash *dh, int slot)
1119 DIRHASH_ASSERT_LOCKED(dh);
1121 /* Mark the entry as deleted. */
1122 DH_ENTRY(dh, slot) = DIRHASH_DEL;
1124 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
1125 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
1126 i = WRAPINCR(i, dh->dh_hlen);
1127 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
1128 i = WRAPDECR(i, dh->dh_hlen);
1129 while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
1130 DH_ENTRY(dh, i) = DIRHASH_EMPTY;
1132 i = WRAPDECR(i, dh->dh_hlen);
1134 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen"));
1139 * Given a directory entry and its offset, find the offset of the
1140 * previous entry in the same DIRBLKSIZ-sized block. Returns an
1141 * offset, or -1 if there is no previous entry in the block or some
1142 * other problem occurred.
1145 ufsdirhash_getprev(struct direct *dirp, doff_t offset)
1149 doff_t blkoff, prevoff;
1152 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */
1153 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */
1154 blkbuf = (char *)dirp - entrypos;
1157 /* If `offset' is the start of a block, there is no previous entry. */
1161 /* Scan from the start of the block until we get to the entry. */
1162 for (i = 0; i < entrypos; i += dp->d_reclen) {
1163 dp = (struct direct *)(blkbuf + i);
1164 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
1165 return (-1); /* Corrupted directory. */
1166 prevoff = blkoff + i;
1172 * Delete the given dirhash and reclaim its memory. Assumes that
1173 * ufsdirhash_list is locked, and leaves it locked. Also assumes
1174 * that dh is locked. Returns the amount of memory freed.
1177 ufsdirhash_destroy(struct dirhash *dh)
1181 int i, mem, narrays;
1183 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list"));
1185 /* Remove it from the list and detach its memory. */
1186 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
1190 blkfree = dh->dh_blkfree;
1191 dh->dh_blkfree = NULL;
1192 narrays = dh->dh_narrays;
1193 mem = dh->dh_memreq;
1196 /* Unlock dirhash and free the detached memory. */
1197 ufsdirhash_release(dh);
1198 for (i = 0; i < narrays; i++)
1199 DIRHASH_BLKFREE(hash[i]);
1200 free(hash, M_DIRHASH);
1201 free(blkfree, M_DIRHASH);
1203 /* Account for the returned memory. */
1204 ufs_dirhashmem -= mem;
1210 * Try to free up `wanted' bytes by stealing memory from existing
1211 * dirhashes. Returns zero with list locked if successful.
1214 ufsdirhash_recycle(int wanted)
1219 dh = TAILQ_FIRST(&ufsdirhash_list);
1220 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
1221 /* Decrement the score; only recycle if it becomes zero. */
1222 if (dh == NULL || --dh->dh_score > 0) {
1223 DIRHASHLIST_UNLOCK();
1227 * If we can't lock it it's in use and we don't want to
1228 * recycle it anyway.
1230 if (!sx_try_xlock(&dh->dh_lock)) {
1231 dh = TAILQ_NEXT(dh, dh_list);
1235 ufsdirhash_destroy(dh);
1237 /* Repeat if necessary. */
1238 dh = TAILQ_FIRST(&ufsdirhash_list);
1240 /* Success; return with list locked. */
1245 * Callback that frees some dirhashes when the system is low on virtual memory.
1250 struct dirhash *dh, *dh_temp;
1251 int memfreed, memwanted;
1253 ufs_dirhashlowmemcount++;
1255 memwanted = ufs_dirhashmem * ufs_dirhashreclaimpercent / 100;
1260 * Reclaim up to memwanted from the oldest dirhashes. This will allow
1261 * us to make some progress when the system is running out of memory
1262 * without compromising the dinamicity of maximum age. If the situation
1263 * does not improve lowmem will be eventually retriggered and free some
1264 * other entry in the cache. The entries on the head of the list should
1265 * be the oldest. If during list traversal we can't get a lock on the
1266 * dirhash, it will be skipped.
1268 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) {
1269 if (sx_try_xlock(&dh->dh_lock))
1270 memfreed += ufsdirhash_destroy(dh);
1271 if (memfreed >= memwanted)
1274 DIRHASHLIST_UNLOCK();
1278 ufsdirhash_set_reclaimpercent(SYSCTL_HANDLER_ARGS)
1282 v = ufs_dirhashreclaimpercent;
1283 error = sysctl_handle_int(oidp, &v, v, req);
1286 if (req->newptr == NULL)
1288 if (v == ufs_dirhashreclaimpercent)
1291 /* Refuse invalid percentages */
1292 if (v < 0 || v > 100)
1294 ufs_dirhashreclaimpercent = v;
1301 ufs_dirhashmaxmem = lmax(roundup(hibufspace / 64, PAGE_SIZE),
1304 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t),
1305 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1306 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF);
1307 TAILQ_INIT(&ufsdirhash_list);
1309 /* Register a callback function to handle low memory signals */
1310 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL,
1311 EVENTHANDLER_PRI_FIRST);
1317 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit"));
1318 uma_zdestroy(ufsdirhash_zone);
1319 mtx_destroy(&ufsdirhash_mtx);
1322 #endif /* UFS_DIRHASH */