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)
198 /* Racy check for i_dirhash to prefetch a dirhash structure. */
199 if (ip->i_dirhash == NULL && ndh == NULL) {
200 ndh = malloc(sizeof *dh, M_DIRHASH,
204 refcount_init(&ndh->dh_refcount, 1);
207 * The DUPOK is to prevent warnings from the
208 * sx_slock() a few lines down which is safe
209 * since the duplicate lock in that case is
210 * the one for this dirhash we are creating
211 * now which has no external references until
212 * after this function returns.
214 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK);
215 sx_xlock(&ndh->dh_lock);
218 * Check i_dirhash. If it's NULL just try to use a
219 * preallocated structure. If none exists loop and try again.
233 /* Acquire a lock on existing hashes. */
235 sx_xlock(&dh->dh_lock);
237 sx_slock(&dh->dh_lock);
239 /* The hash could've been recycled while we were waiting. */
241 if (ip->i_dirhash != dh) {
243 ufsdirhash_release(dh);
250 /* If the hash is still valid we've succeeded. */
251 if (dh->dh_hash != NULL)
254 * If the hash is NULL it has been recycled. Try to upgrade
255 * so we can recreate it. If we fail the upgrade, drop our
256 * lock and try again.
258 if (excl || sx_try_upgrade(&dh->dh_lock))
260 sx_sunlock(&dh->dh_lock);
263 /* Free the preallocated structure if it was not necessary. */
265 ufsdirhash_release(ndh);
266 ufsdirhash_drop(ndh);
272 * Acquire an exclusive lock on an existing hash. Requires an exclusive
273 * vnode lock to protect the i_dirhash pointer. hashes that have been
274 * recycled are reclaimed here and NULL is returned.
276 static struct dirhash *
277 ufsdirhash_acquire(struct inode *ip)
281 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__);
286 sx_xlock(&dh->dh_lock);
287 if (dh->dh_hash != NULL)
289 ufsdirhash_free_locked(ip);
294 * Acquire exclusively and free the hash pointed to by ip. Works with a
295 * shared or exclusive vnode lock.
298 ufsdirhash_free(struct inode *ip)
305 /* Grab a reference on this inode's dirhash if it has one. */
315 /* Exclusively lock the dirhash. */
316 sx_xlock(&dh->dh_lock);
318 /* If this dirhash still belongs to this inode, then free it. */
320 if (ip->i_dirhash == dh) {
328 * This inode's dirhash has changed while we were
329 * waiting for the dirhash lock, so try again.
331 ufsdirhash_release(dh);
334 ufsdirhash_free_locked(ip);
338 * Attempt to build up a hash table for the directory contents in
339 * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
342 ufsdirhash_build(struct inode *ip)
345 struct buf *bp = NULL;
349 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
351 /* Take care of a decreased sysctl value. */
352 while (ufs_dirhashmem > ufs_dirhashmaxmem) {
353 if (ufsdirhash_recycle(0) != 0)
355 /* Recycled enough memory, so unlock the list. */
356 DIRHASHLIST_UNLOCK();
359 /* Check if we can/should use dirhash. */
360 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) ||
361 ip->i_effnlink == 0) {
366 dh = ufsdirhash_create(ip);
369 if (dh->dh_hash != NULL)
373 /* Allocate 50% more entries than this dir size could ever need. */
374 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size"));
375 nslots = ip->i_size / DIRECTSIZ(1);
376 nslots = (nslots * 3 + 1) / 2;
377 narrays = howmany(nslots, DH_NBLKOFF);
378 nslots = narrays * DH_NBLKOFF;
379 dirblocks = howmany(ip->i_size, DIRBLKSIZ);
380 nblocks = (dirblocks * 3 + 1) / 2;
381 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
382 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
383 nblocks * sizeof(*dh->dh_blkfree);
385 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
386 DIRHASHLIST_UNLOCK();
387 if (memreqd > ufs_dirhashmaxmem / 2)
389 /* Try to free some space. */
390 if (ufsdirhash_recycle(memreqd) != 0)
392 /* Enough was freed, and list has been locked. */
394 ufs_dirhashmem += memreqd;
395 DIRHASHLIST_UNLOCK();
397 /* Initialise the hash table and block statistics. */
398 dh->dh_memreq = memreqd;
399 dh->dh_narrays = narrays;
400 dh->dh_hlen = nslots;
401 dh->dh_nblk = nblocks;
402 dh->dh_dirblks = dirblocks;
403 for (i = 0; i < DH_NFSTATS; i++)
404 dh->dh_firstfree[i] = -1;
405 dh->dh_firstfree[DH_NFSTATS] = 0;
408 dh->dh_score = DH_SCOREINIT;
409 dh->dh_lastused = time_second;
412 * Use non-blocking mallocs so that we will revert to a linear
413 * lookup on failure rather than potentially blocking forever.
415 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]),
416 M_DIRHASH, M_NOWAIT | M_ZERO);
417 if (dh->dh_hash == NULL)
419 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]),
420 M_DIRHASH, M_NOWAIT);
421 if (dh->dh_blkfree == NULL)
423 for (i = 0; i < narrays; i++) {
424 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL)
426 for (j = 0; j < DH_NBLKOFF; j++)
427 dh->dh_hash[i][j] = DIRHASH_EMPTY;
429 for (i = 0; i < dirblocks; i++)
430 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN;
431 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
433 while (pos < ip->i_size) {
434 /* If necessary, get the next directory block. */
435 if ((pos & bmask) == 0) {
438 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
442 /* Add this entry to the hash. */
443 ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
444 if (ep->d_reclen == 0 || ep->d_reclen >
445 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) {
446 /* Corrupted directory. */
450 if (ep->d_ino != 0) {
451 /* Add the entry (simplified ufsdirhash_add). */
452 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
453 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
454 slot = WRAPINCR(slot, dh->dh_hlen);
456 DH_ENTRY(dh, slot) = pos;
457 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep));
465 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
467 DIRHASHLIST_UNLOCK();
468 sx_downgrade(&dh->dh_lock);
472 ufsdirhash_free_locked(ip);
477 * Free any hash table associated with inode 'ip'.
480 ufsdirhash_free_locked(struct inode *ip)
486 DIRHASH_ASSERT_LOCKED(ip->i_dirhash);
489 * Clear the pointer in the inode to prevent new threads from
490 * finding the dead structure.
495 ip->i_dirhash = NULL;
499 * Remove the hash from the list since we are going to free its
504 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
505 ufs_dirhashmem -= dh->dh_memreq;
506 DIRHASHLIST_UNLOCK();
509 * At this point, any waiters for the lock should hold their
510 * own reference on the dirhash structure. They will drop
511 * that reference once they grab the vnode interlock and see
512 * that ip->i_dirhash is NULL.
514 sx_xunlock(&dh->dh_lock);
517 * Handle partially recycled as well as fully constructed hashes.
519 if (dh->dh_hash != NULL) {
520 for (i = 0; i < dh->dh_narrays; i++)
521 if (dh->dh_hash[i] != NULL)
522 DIRHASH_BLKFREE(dh->dh_hash[i]);
523 free(dh->dh_hash, M_DIRHASH);
524 if (dh->dh_blkfree != NULL)
525 free(dh->dh_blkfree, M_DIRHASH);
529 * Drop the inode's reference to the data structure.
535 * Find the offset of the specified name within the given inode.
536 * Returns 0 on success, ENOENT if the entry does not exist, or
537 * EJUSTRETURN if the caller should revert to a linear search.
539 * If successful, the directory offset is stored in *offp, and a
540 * pointer to a struct buf containing the entry is stored in *bpp. If
541 * prevoffp is non-NULL, the offset of the previous entry within
542 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
543 * is the first in a block, the start of the block is used).
545 * Must be called with the hash locked. Returns with the hash unlocked.
548 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp,
549 struct buf **bpp, doff_t *prevoffp)
551 struct dirhash *dh, *dh_next;
555 doff_t blkoff, bmask, offset, prevoff, seqoff;
560 KASSERT(dh != NULL && dh->dh_hash != NULL,
561 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh));
562 DIRHASH_ASSERT_LOCKED(dh);
564 * Move this dirhash towards the end of the list if it has a
565 * score higher than the next entry, and acquire the dh_lock.
568 if (TAILQ_NEXT(dh, dh_list) != NULL) {
570 * If the new score will be greater than that of the next
571 * entry, then move this entry past it. With both mutexes
572 * held, dh_next won't go away, but its dh_score could
573 * change; that's not important since it is just a hint.
575 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
576 dh->dh_score >= dh_next->dh_score) {
577 KASSERT(dh->dh_onlist, ("dirhash: not on list"));
578 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
579 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
583 /* Update the score. */
584 if (dh->dh_score < DH_SCOREMAX)
587 /* Update last used time. */
588 dh->dh_lastused = time_second;
589 DIRHASHLIST_UNLOCK();
592 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
595 seqoff = dh->dh_seqoff;
597 slot = ufsdirhash_hash(dh, name, namelen);
601 * Sequential access optimisation. seqoff contains the
602 * offset of the directory entry immediately following
603 * the last entry that was looked up. Check if this offset
604 * appears in the hash chain for the name we are looking for.
606 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
607 i = WRAPINCR(i, dh->dh_hlen))
608 if (offset == seqoff)
610 if (offset == seqoff) {
612 * We found an entry with the expected offset. This
613 * is probably the entry we want, but if not, the
614 * code below will retry.
621 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
622 slot = WRAPINCR(slot, dh->dh_hlen)) {
623 if (offset == DIRHASH_DEL)
625 if (offset < 0 || offset >= ip->i_size)
626 panic("ufsdirhash_lookup: bad offset in hash array");
627 if ((offset & ~bmask) != blkoff) {
630 blkoff = offset & ~bmask;
631 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) {
636 KASSERT(bp != NULL, ("no buffer allocated"));
637 dp = (struct direct *)(bp->b_data + (offset & bmask));
638 if (dp->d_reclen == 0 || dp->d_reclen >
639 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
640 /* Corrupted directory. */
644 if (dp->d_namlen == namelen &&
645 bcmp(dp->d_name, name, namelen) == 0) {
646 /* Found. Get the prev offset if needed. */
647 if (prevoffp != NULL) {
648 if (offset & (DIRBLKSIZ - 1)) {
649 prevoff = ufsdirhash_getprev(dp,
661 dh->dh_seqoff = offset + DIRSIZ(0, dp);
664 ufsdirhash_release(dh);
669 * When the name doesn't match in the sequential
670 * optimization case, go back and search normally.
679 ufsdirhash_release(dh);
686 * Find a directory block with room for 'slotneeded' bytes. Returns
687 * the offset of the directory entry that begins the free space.
688 * This will either be the offset of an existing entry that has free
689 * space at the end, or the offset of an entry with d_ino == 0 at
690 * the start of a DIRBLKSIZ block.
692 * To use the space, the caller may need to compact existing entries in
693 * the directory. The total number of bytes in all of the entries involved
694 * in the compaction is stored in *slotsize. In other words, all of
695 * the entries that must be compacted are exactly contained in the
696 * region beginning at the returned offset and spanning *slotsize bytes.
698 * Returns -1 if no space was found, indicating that the directory
702 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
707 doff_t pos, slotstart;
708 int dirblock, error, freebytes, i;
711 KASSERT(dh != NULL && dh->dh_hash != NULL,
712 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh));
713 DIRHASH_ASSERT_LOCKED(dh);
715 /* Find a directory block with the desired free space. */
717 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
718 if ((dirblock = dh->dh_firstfree[i]) != -1)
723 KASSERT(dirblock < dh->dh_nblk &&
724 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
725 ("ufsdirhash_findfree: bad stats"));
726 pos = dirblock * DIRBLKSIZ;
727 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp);
731 /* Find the first entry with free space. */
732 for (i = 0; i < DIRBLKSIZ; ) {
733 if (dp->d_reclen == 0) {
737 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
740 dp = (struct direct *)((char *)dp + dp->d_reclen);
748 /* Find the range of entries needed to get enough space */
750 while (i < DIRBLKSIZ && freebytes < slotneeded) {
751 freebytes += dp->d_reclen;
753 freebytes -= DIRSIZ(0, dp);
754 if (dp->d_reclen == 0) {
759 dp = (struct direct *)((char *)dp + dp->d_reclen);
765 if (freebytes < slotneeded)
766 panic("ufsdirhash_findfree: free mismatch");
768 *slotsize = pos + i - slotstart;
773 * Return the start of the unused space at the end of a directory, or
774 * -1 if there are no trailing unused blocks.
777 ufsdirhash_enduseful(struct inode *ip)
784 DIRHASH_ASSERT_LOCKED(dh);
785 KASSERT(dh != NULL && dh->dh_hash != NULL,
786 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh));
788 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN)
791 for (i = dh->dh_dirblks - 1; i >= 0; i--)
792 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
795 return ((doff_t)(i + 1) * DIRBLKSIZ);
799 * Insert information into the hash about a new directory entry. dirp
800 * points to a struct direct containing the entry, and offset specifies
801 * the offset of this entry.
804 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
809 if ((dh = ufsdirhash_acquire(ip)) == NULL)
812 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
813 ("ufsdirhash_add: bad offset"));
815 * Normal hash usage is < 66%. If the usage gets too high then
816 * remove the hash entirely and let it be rebuilt later.
818 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
819 ufsdirhash_free_locked(ip);
823 /* Find a free hash slot (empty or deleted), and add the entry. */
824 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
825 while (DH_ENTRY(dh, slot) >= 0)
826 slot = WRAPINCR(slot, dh->dh_hlen);
827 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
829 DH_ENTRY(dh, slot) = offset;
831 /* Update last used time. */
832 dh->dh_lastused = time_second;
834 /* Update the per-block summary info. */
835 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp));
836 ufsdirhash_release(dh);
840 * Remove the specified directory entry from the hash. The entry to remove
841 * is defined by the name in `dirp', which must exist at the specified
842 * `offset' within the directory.
845 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
850 if ((dh = ufsdirhash_acquire(ip)) == NULL)
853 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
854 ("ufsdirhash_remove: bad offset"));
856 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
858 /* Remove the hash entry. */
859 ufsdirhash_delslot(dh, slot);
861 /* Update the per-block summary info. */
862 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp));
863 ufsdirhash_release(dh);
867 * Change the offset associated with a directory entry in the hash. Used
868 * when compacting directory blocks.
871 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
877 if ((dh = ufsdirhash_acquire(ip)) == NULL)
880 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ &&
881 newoff < dh->dh_dirblks * DIRBLKSIZ,
882 ("ufsdirhash_move: bad offset"));
883 /* Find the entry, and update the offset. */
884 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
885 DH_ENTRY(dh, slot) = newoff;
886 ufsdirhash_release(dh);
890 * Inform dirhash that the directory has grown by one block that
891 * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
894 ufsdirhash_newblk(struct inode *ip, doff_t offset)
899 if ((dh = ufsdirhash_acquire(ip)) == NULL)
902 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ,
903 ("ufsdirhash_newblk: bad offset"));
904 block = offset / DIRBLKSIZ;
905 if (block >= dh->dh_nblk) {
906 /* Out of space; must rebuild. */
907 ufsdirhash_free_locked(ip);
910 dh->dh_dirblks = block + 1;
912 /* Account for the new free block. */
913 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN;
914 if (dh->dh_firstfree[DH_NFSTATS] == -1)
915 dh->dh_firstfree[DH_NFSTATS] = block;
916 ufsdirhash_release(dh);
920 * Inform dirhash that the directory is being truncated.
923 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
928 if ((dh = ufsdirhash_acquire(ip)) == NULL)
931 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ,
932 ("ufsdirhash_dirtrunc: bad offset"));
933 block = howmany(offset, DIRBLKSIZ);
935 * If the directory shrinks to less than 1/8 of dh_nblk blocks
936 * (about 20% of its original size due to the 50% extra added in
937 * ufsdirhash_build) then free it, and let the caller rebuild
940 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
941 ufsdirhash_free_locked(ip);
946 * Remove any `first free' information pertaining to the
947 * truncated blocks. All blocks we're removing should be
950 if (dh->dh_firstfree[DH_NFSTATS] >= block)
951 dh->dh_firstfree[DH_NFSTATS] = -1;
952 for (i = block; i < dh->dh_dirblks; i++)
953 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
954 panic("ufsdirhash_dirtrunc: blocks in use");
955 for (i = 0; i < DH_NFSTATS; i++)
956 if (dh->dh_firstfree[i] >= block)
957 panic("ufsdirhash_dirtrunc: first free corrupt");
958 dh->dh_dirblks = block;
959 ufsdirhash_release(dh);
963 * Debugging function to check that the dirhash information about
964 * a directory block matches its actual contents. Panics if a mismatch
967 * On entry, `buf' should point to the start of an in-core
968 * DIRBLKSIZ-sized directory block, and `offset' should contain the
969 * offset from the start of the directory of that block.
972 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
976 int block, ffslot, i, nfree;
978 if (!ufs_dirhashcheck)
980 if ((dh = ufsdirhash_acquire(ip)) == NULL)
983 block = offset / DIRBLKSIZ;
984 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks)
985 panic("ufsdirhash_checkblock: bad offset");
988 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) {
989 dp = (struct direct *)(buf + i);
990 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ)
991 panic("ufsdirhash_checkblock: bad dir");
993 if (dp->d_ino == 0) {
996 * XXX entries with d_ino == 0 should only occur
997 * at the start of a DIRBLKSIZ block. However the
998 * ufs code is tolerant of such entries at other
999 * offsets, and fsck does not fix them.
1002 panic("ufsdirhash_checkblock: bad dir inode");
1004 nfree += dp->d_reclen;
1008 /* Check that the entry exists (will panic if it doesn't). */
1009 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
1011 nfree += dp->d_reclen - DIRSIZ(0, dp);
1014 panic("ufsdirhash_checkblock: bad dir end");
1016 if (dh->dh_blkfree[block] * DIRALIGN != nfree)
1017 panic("ufsdirhash_checkblock: bad free count");
1019 ffslot = BLKFREE2IDX(nfree / DIRALIGN);
1020 for (i = 0; i <= DH_NFSTATS; i++)
1021 if (dh->dh_firstfree[i] == block && i != ffslot)
1022 panic("ufsdirhash_checkblock: bad first-free");
1023 if (dh->dh_firstfree[ffslot] == -1)
1024 panic("ufsdirhash_checkblock: missing first-free entry");
1025 ufsdirhash_release(dh);
1029 * Hash the specified filename into a dirhash slot.
1032 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen)
1037 * We hash the name and then some other bit of data that is
1038 * invariant over the dirhash's lifetime. Otherwise names
1039 * differing only in the last byte are placed close to one
1040 * another in the table, which is bad for linear probing.
1042 hash = fnv_32_buf(name, namelen, FNV1_32_INIT);
1043 hash = fnv_32_buf(&dh, sizeof(dh), hash);
1044 return (hash % dh->dh_hlen);
1048 * Adjust the number of free bytes in the block containing `offset'
1049 * by the value specified by `diff'.
1051 * The caller must ensure we have exclusive access to `dh'; normally
1052 * that means that dh_lock should be held, but this is also called
1053 * from ufsdirhash_build() where exclusive access can be assumed.
1056 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff)
1058 int block, i, nfidx, ofidx;
1060 /* Update the per-block summary info. */
1061 block = offset / DIRBLKSIZ;
1062 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks,
1063 ("dirhash bad offset"));
1064 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1065 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
1066 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1068 /* Update the `first free' list if necessary. */
1069 if (ofidx != nfidx) {
1070 /* If removing, scan forward for the next block. */
1071 if (dh->dh_firstfree[ofidx] == block) {
1072 for (i = block + 1; i < dh->dh_dirblks; i++)
1073 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
1075 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
1078 /* Make this the new `first free' if necessary */
1079 if (dh->dh_firstfree[nfidx] > block ||
1080 dh->dh_firstfree[nfidx] == -1)
1081 dh->dh_firstfree[nfidx] = block;
1086 * Find the specified name which should have the specified offset.
1087 * Returns a slot number, and panics on failure.
1089 * `dh' must be locked on entry and remains so on return.
1092 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset)
1096 DIRHASH_ASSERT_LOCKED(dh);
1098 /* Find the entry. */
1099 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full"));
1100 slot = ufsdirhash_hash(dh, name, namelen);
1101 while (DH_ENTRY(dh, slot) != offset &&
1102 DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
1103 slot = WRAPINCR(slot, dh->dh_hlen);
1104 if (DH_ENTRY(dh, slot) != offset)
1105 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
1111 * Remove the entry corresponding to the specified slot from the hash array.
1113 * `dh' must be locked on entry and remains so on return.
1116 ufsdirhash_delslot(struct dirhash *dh, int slot)
1120 DIRHASH_ASSERT_LOCKED(dh);
1122 /* Mark the entry as deleted. */
1123 DH_ENTRY(dh, slot) = DIRHASH_DEL;
1125 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
1126 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
1127 i = WRAPINCR(i, dh->dh_hlen);
1128 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
1129 i = WRAPDECR(i, dh->dh_hlen);
1130 while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
1131 DH_ENTRY(dh, i) = DIRHASH_EMPTY;
1133 i = WRAPDECR(i, dh->dh_hlen);
1135 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen"));
1140 * Given a directory entry and its offset, find the offset of the
1141 * previous entry in the same DIRBLKSIZ-sized block. Returns an
1142 * offset, or -1 if there is no previous entry in the block or some
1143 * other problem occurred.
1146 ufsdirhash_getprev(struct direct *dirp, doff_t offset)
1150 doff_t blkoff, prevoff;
1153 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */
1154 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */
1155 blkbuf = (char *)dirp - entrypos;
1158 /* If `offset' is the start of a block, there is no previous entry. */
1162 /* Scan from the start of the block until we get to the entry. */
1163 for (i = 0; i < entrypos; i += dp->d_reclen) {
1164 dp = (struct direct *)(blkbuf + i);
1165 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
1166 return (-1); /* Corrupted directory. */
1167 prevoff = blkoff + i;
1173 * Delete the given dirhash and reclaim its memory. Assumes that
1174 * ufsdirhash_list is locked, and leaves it locked. Also assumes
1175 * that dh is locked. Returns the amount of memory freed.
1178 ufsdirhash_destroy(struct dirhash *dh)
1182 int i, mem, narrays;
1184 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list"));
1186 /* Remove it from the list and detach its memory. */
1187 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
1191 blkfree = dh->dh_blkfree;
1192 dh->dh_blkfree = NULL;
1193 narrays = dh->dh_narrays;
1194 mem = dh->dh_memreq;
1197 /* Unlock dirhash and free the detached memory. */
1198 ufsdirhash_release(dh);
1199 for (i = 0; i < narrays; i++)
1200 DIRHASH_BLKFREE(hash[i]);
1201 free(hash, M_DIRHASH);
1202 free(blkfree, M_DIRHASH);
1204 /* Account for the returned memory. */
1205 ufs_dirhashmem -= mem;
1211 * Try to free up `wanted' bytes by stealing memory from existing
1212 * dirhashes. Returns zero with list locked if successful.
1215 ufsdirhash_recycle(int wanted)
1220 dh = TAILQ_FIRST(&ufsdirhash_list);
1221 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
1222 /* Decrement the score; only recycle if it becomes zero. */
1223 if (dh == NULL || --dh->dh_score > 0) {
1224 DIRHASHLIST_UNLOCK();
1228 * If we can't lock it it's in use and we don't want to
1229 * recycle it anyway.
1231 if (!sx_try_xlock(&dh->dh_lock)) {
1232 dh = TAILQ_NEXT(dh, dh_list);
1236 ufsdirhash_destroy(dh);
1238 /* Repeat if necessary. */
1239 dh = TAILQ_FIRST(&ufsdirhash_list);
1241 /* Success; return with list locked. */
1246 * Callback that frees some dirhashes when the system is low on virtual memory.
1251 struct dirhash *dh, *dh_temp;
1254 * Will free a *minimum* of 10% of the dirhash, but possibly much
1255 * more (depending on dirhashreclaimage). System with large dirhashes
1256 * probably also need a much larger dirhashreclaimage.
1257 * XXX: this percentage may need to be adjusted.
1259 int memwanted = ufs_dirhashmem / 10;
1261 ufs_dirhashlowmemcount++;
1265 * Delete dirhashes not used for more than ufs_dirhashreclaimage
1266 * seconds. If we can't get a lock on the dirhash, it will be skipped.
1268 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) {
1269 if (!sx_try_xlock(&dh->dh_lock))
1271 if (time_second - dh->dh_lastused > ufs_dirhashreclaimage)
1272 memfreed += ufsdirhash_destroy(dh);
1273 /* Unlock if we didn't delete the dirhash */
1275 ufsdirhash_release(dh);
1279 * If not enough memory was freed, keep deleting hashes from the head
1280 * of the dirhash list. The ones closest to the head should be the
1283 if (memfreed < memwanted) {
1284 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) {
1285 if (!sx_try_xlock(&dh->dh_lock))
1287 memfreed += ufsdirhash_destroy(dh);
1288 if (memfreed >= memwanted)
1292 DIRHASHLIST_UNLOCK();
1299 ufs_dirhashmaxmem = lmax(roundup(hibufspace / 64, PAGE_SIZE),
1302 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t),
1303 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1304 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF);
1305 TAILQ_INIT(&ufsdirhash_list);
1307 /* Register a callback function to handle low memory signals */
1308 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL,
1309 EVENTHANDLER_PRI_FIRST);
1315 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit"));
1316 uma_zdestroy(ufsdirhash_zone);
1317 mtx_destroy(&ufsdirhash_mtx);
1320 #endif /* UFS_DIRHASH */