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)
197 /* Racy check for i_dirhash to prefetch a dirhash structure. */
198 if (ip->i_dirhash == NULL && ndh == NULL) {
199 ndh = malloc(sizeof *dh, M_DIRHASH,
203 refcount_init(&ndh->dh_refcount, 1);
206 * The DUPOK is to prevent warnings from the
207 * sx_slock() a few lines down which is safe
208 * since the duplicate lock in that case is
209 * the one for this dirhash we are creating
210 * now which has no external references until
211 * after this function returns.
213 sx_init_flags(&ndh->dh_lock, "dirhash", SX_DUPOK);
214 sx_xlock(&ndh->dh_lock);
217 * Check i_dirhash. If it's NULL just try to use a
218 * preallocated structure. If none exists loop and try again.
232 /* Acquire a shared lock on existing hashes. */
233 sx_slock(&dh->dh_lock);
235 /* The hash could've been recycled while we were waiting. */
237 if (ip->i_dirhash != dh) {
239 ufsdirhash_release(dh);
246 /* If the hash is still valid we've succeeded. */
247 if (dh->dh_hash != NULL)
250 * If the hash is NULL it has been recycled. Try to upgrade
251 * so we can recreate it. If we fail the upgrade, drop our
252 * lock and try again.
254 if (sx_try_upgrade(&dh->dh_lock))
256 sx_sunlock(&dh->dh_lock);
258 /* Free the preallocated structure if it was not necessary. */
260 ufsdirhash_release(ndh);
261 ufsdirhash_drop(ndh);
267 * Acquire an exclusive lock on an existing hash. Requires an exclusive
268 * vnode lock to protect the i_dirhash pointer. hashes that have been
269 * recycled are reclaimed here and NULL is returned.
271 static struct dirhash *
272 ufsdirhash_acquire(struct inode *ip)
276 ASSERT_VOP_ELOCKED(ip->i_vnode, __FUNCTION__);
281 sx_xlock(&dh->dh_lock);
282 if (dh->dh_hash != NULL)
284 ufsdirhash_free_locked(ip);
289 * Acquire exclusively and free the hash pointed to by ip. Works with a
290 * shared or exclusive vnode lock.
293 ufsdirhash_free(struct inode *ip)
300 /* Grab a reference on this inode's dirhash if it has one. */
310 /* Exclusively lock the dirhash. */
311 sx_xlock(&dh->dh_lock);
313 /* If this dirhash still belongs to this inode, then free it. */
315 if (ip->i_dirhash == dh) {
323 * This inode's dirhash has changed while we were
324 * waiting for the dirhash lock, so try again.
326 ufsdirhash_release(dh);
329 ufsdirhash_free_locked(ip);
333 * Attempt to build up a hash table for the directory contents in
334 * inode 'ip'. Returns 0 on success, or -1 of the operation failed.
337 ufsdirhash_build(struct inode *ip)
340 struct buf *bp = NULL;
344 int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
346 /* Take care of a decreased sysctl value. */
347 while (ufs_dirhashmem > ufs_dirhashmaxmem) {
348 if (ufsdirhash_recycle(0) != 0)
350 /* Recycled enough memory, so unlock the list. */
351 DIRHASHLIST_UNLOCK();
354 /* Check if we can/should use dirhash. */
355 if (ip->i_size < ufs_mindirhashsize || OFSFMT(ip->i_vnode) ||
356 ip->i_effnlink == 0) {
361 dh = ufsdirhash_create(ip);
364 if (dh->dh_hash != NULL)
368 /* Allocate 50% more entries than this dir size could ever need. */
369 KASSERT(ip->i_size >= DIRBLKSIZ, ("ufsdirhash_build size"));
370 nslots = ip->i_size / DIRECTSIZ(1);
371 nslots = (nslots * 3 + 1) / 2;
372 narrays = howmany(nslots, DH_NBLKOFF);
373 nslots = narrays * DH_NBLKOFF;
374 dirblocks = howmany(ip->i_size, DIRBLKSIZ);
375 nblocks = (dirblocks * 3 + 1) / 2;
376 memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
377 narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
378 nblocks * sizeof(*dh->dh_blkfree);
380 if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
381 DIRHASHLIST_UNLOCK();
382 if (memreqd > ufs_dirhashmaxmem / 2)
384 /* Try to free some space. */
385 if (ufsdirhash_recycle(memreqd) != 0)
387 /* Enough was freed, and list has been locked. */
389 ufs_dirhashmem += memreqd;
390 DIRHASHLIST_UNLOCK();
392 /* Initialise the hash table and block statistics. */
393 dh->dh_memreq = memreqd;
394 dh->dh_narrays = narrays;
395 dh->dh_hlen = nslots;
396 dh->dh_nblk = nblocks;
397 dh->dh_dirblks = dirblocks;
398 for (i = 0; i < DH_NFSTATS; i++)
399 dh->dh_firstfree[i] = -1;
400 dh->dh_firstfree[DH_NFSTATS] = 0;
403 dh->dh_score = DH_SCOREINIT;
404 dh->dh_lastused = time_second;
407 * Use non-blocking mallocs so that we will revert to a linear
408 * lookup on failure rather than potentially blocking forever.
410 dh->dh_hash = malloc(narrays * sizeof(dh->dh_hash[0]),
411 M_DIRHASH, M_NOWAIT | M_ZERO);
412 if (dh->dh_hash == NULL)
414 dh->dh_blkfree = malloc(nblocks * sizeof(dh->dh_blkfree[0]),
415 M_DIRHASH, M_NOWAIT);
416 if (dh->dh_blkfree == NULL)
418 for (i = 0; i < narrays; i++) {
419 if ((dh->dh_hash[i] = DIRHASH_BLKALLOC_WAITOK()) == NULL)
421 for (j = 0; j < DH_NBLKOFF; j++)
422 dh->dh_hash[i][j] = DIRHASH_EMPTY;
424 for (i = 0; i < dirblocks; i++)
425 dh->dh_blkfree[i] = DIRBLKSIZ / DIRALIGN;
426 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
428 while (pos < ip->i_size) {
429 /* If necessary, get the next directory block. */
430 if ((pos & bmask) == 0) {
433 if (UFS_BLKATOFF(vp, (off_t)pos, NULL, &bp) != 0)
437 /* Add this entry to the hash. */
438 ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
439 if (ep->d_reclen == 0 || ep->d_reclen >
440 DIRBLKSIZ - (pos & (DIRBLKSIZ - 1))) {
441 /* Corrupted directory. */
445 if (ep->d_ino != 0) {
446 /* Add the entry (simplified ufsdirhash_add). */
447 slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
448 while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
449 slot = WRAPINCR(slot, dh->dh_hlen);
451 DH_ENTRY(dh, slot) = pos;
452 ufsdirhash_adjfree(dh, pos, -DIRSIZ(0, ep));
460 TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
462 DIRHASHLIST_UNLOCK();
463 sx_downgrade(&dh->dh_lock);
467 ufsdirhash_free_locked(ip);
472 * Free any hash table associated with inode 'ip'.
475 ufsdirhash_free_locked(struct inode *ip)
481 DIRHASH_ASSERT_LOCKED(ip->i_dirhash);
484 * Clear the pointer in the inode to prevent new threads from
485 * finding the dead structure.
490 ip->i_dirhash = NULL;
494 * Remove the hash from the list since we are going to free its
499 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
500 ufs_dirhashmem -= dh->dh_memreq;
501 DIRHASHLIST_UNLOCK();
504 * At this point, any waiters for the lock should hold their
505 * own reference on the dirhash structure. They will drop
506 * that reference once they grab the vnode interlock and see
507 * that ip->i_dirhash is NULL.
509 sx_xunlock(&dh->dh_lock);
512 * Handle partially recycled as well as fully constructed hashes.
514 if (dh->dh_hash != NULL) {
515 for (i = 0; i < dh->dh_narrays; i++)
516 if (dh->dh_hash[i] != NULL)
517 DIRHASH_BLKFREE(dh->dh_hash[i]);
518 free(dh->dh_hash, M_DIRHASH);
519 if (dh->dh_blkfree != NULL)
520 free(dh->dh_blkfree, M_DIRHASH);
524 * Drop the inode's reference to the data structure.
530 * Find the offset of the specified name within the given inode.
531 * Returns 0 on success, ENOENT if the entry does not exist, or
532 * EJUSTRETURN if the caller should revert to a linear search.
534 * If successful, the directory offset is stored in *offp, and a
535 * pointer to a struct buf containing the entry is stored in *bpp. If
536 * prevoffp is non-NULL, the offset of the previous entry within
537 * the DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
538 * is the first in a block, the start of the block is used).
540 * Must be called with the hash locked. Returns with the hash unlocked.
543 ufsdirhash_lookup(struct inode *ip, char *name, int namelen, doff_t *offp,
544 struct buf **bpp, doff_t *prevoffp)
546 struct dirhash *dh, *dh_next;
550 doff_t blkoff, bmask, offset, prevoff, seqoff;
555 KASSERT(dh != NULL && dh->dh_hash != NULL,
556 ("ufsdirhash_lookup: Invalid dirhash %p\n", dh));
557 DIRHASH_ASSERT_LOCKED(dh);
559 * Move this dirhash towards the end of the list if it has a
560 * score higher than the next entry, and acquire the dh_lock.
563 if (TAILQ_NEXT(dh, dh_list) != NULL) {
565 * If the new score will be greater than that of the next
566 * entry, then move this entry past it. With both mutexes
567 * held, dh_next won't go away, but its dh_score could
568 * change; that's not important since it is just a hint.
570 if ((dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
571 dh->dh_score >= dh_next->dh_score) {
572 KASSERT(dh->dh_onlist, ("dirhash: not on list"));
573 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
574 TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
578 /* Update the score. */
579 if (dh->dh_score < DH_SCOREMAX)
582 /* Update last used time. */
583 dh->dh_lastused = time_second;
584 DIRHASHLIST_UNLOCK();
587 bmask = vp->v_mount->mnt_stat.f_iosize - 1;
590 seqoff = dh->dh_seqoff;
592 slot = ufsdirhash_hash(dh, name, namelen);
596 * Sequential access optimisation. seqoff contains the
597 * offset of the directory entry immediately following
598 * the last entry that was looked up. Check if this offset
599 * appears in the hash chain for the name we are looking for.
601 for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
602 i = WRAPINCR(i, dh->dh_hlen))
603 if (offset == seqoff)
605 if (offset == seqoff) {
607 * We found an entry with the expected offset. This
608 * is probably the entry we want, but if not, the
609 * code below will retry.
616 for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
617 slot = WRAPINCR(slot, dh->dh_hlen)) {
618 if (offset == DIRHASH_DEL)
620 if (offset < 0 || offset >= ip->i_size)
621 panic("ufsdirhash_lookup: bad offset in hash array");
622 if ((offset & ~bmask) != blkoff) {
625 blkoff = offset & ~bmask;
626 if (UFS_BLKATOFF(vp, (off_t)blkoff, NULL, &bp) != 0) {
631 KASSERT(bp != NULL, ("no buffer allocated"));
632 dp = (struct direct *)(bp->b_data + (offset & bmask));
633 if (dp->d_reclen == 0 || dp->d_reclen >
634 DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
635 /* Corrupted directory. */
639 if (dp->d_namlen == namelen &&
640 bcmp(dp->d_name, name, namelen) == 0) {
641 /* Found. Get the prev offset if needed. */
642 if (prevoffp != NULL) {
643 if (offset & (DIRBLKSIZ - 1)) {
644 prevoff = ufsdirhash_getprev(dp,
656 dh->dh_seqoff = offset + DIRSIZ(0, dp);
659 ufsdirhash_release(dh);
664 * When the name doesn't match in the sequential
665 * optimization case, go back and search normally.
674 ufsdirhash_release(dh);
681 * Find a directory block with room for 'slotneeded' bytes. Returns
682 * the offset of the directory entry that begins the free space.
683 * This will either be the offset of an existing entry that has free
684 * space at the end, or the offset of an entry with d_ino == 0 at
685 * the start of a DIRBLKSIZ block.
687 * To use the space, the caller may need to compact existing entries in
688 * the directory. The total number of bytes in all of the entries involved
689 * in the compaction is stored in *slotsize. In other words, all of
690 * the entries that must be compacted are exactly contained in the
691 * region beginning at the returned offset and spanning *slotsize bytes.
693 * Returns -1 if no space was found, indicating that the directory
697 ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
702 doff_t pos, slotstart;
703 int dirblock, error, freebytes, i;
706 KASSERT(dh != NULL && dh->dh_hash != NULL,
707 ("ufsdirhash_findfree: Invalid dirhash %p\n", dh));
708 DIRHASH_ASSERT_LOCKED(dh);
710 /* Find a directory block with the desired free space. */
712 for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
713 if ((dirblock = dh->dh_firstfree[i]) != -1)
718 KASSERT(dirblock < dh->dh_nblk &&
719 dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
720 ("ufsdirhash_findfree: bad stats"));
721 pos = dirblock * DIRBLKSIZ;
722 error = UFS_BLKATOFF(ip->i_vnode, (off_t)pos, (char **)&dp, &bp);
726 /* Find the first entry with free space. */
727 for (i = 0; i < DIRBLKSIZ; ) {
728 if (dp->d_reclen == 0) {
732 if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
735 dp = (struct direct *)((char *)dp + dp->d_reclen);
743 /* Find the range of entries needed to get enough space */
745 while (i < DIRBLKSIZ && freebytes < slotneeded) {
746 freebytes += dp->d_reclen;
748 freebytes -= DIRSIZ(0, dp);
749 if (dp->d_reclen == 0) {
754 dp = (struct direct *)((char *)dp + dp->d_reclen);
760 if (freebytes < slotneeded)
761 panic("ufsdirhash_findfree: free mismatch");
763 *slotsize = pos + i - slotstart;
768 * Return the start of the unused space at the end of a directory, or
769 * -1 if there are no trailing unused blocks.
772 ufsdirhash_enduseful(struct inode *ip)
779 DIRHASH_ASSERT_LOCKED(dh);
780 KASSERT(dh != NULL && dh->dh_hash != NULL,
781 ("ufsdirhash_enduseful: Invalid dirhash %p\n", dh));
783 if (dh->dh_blkfree[dh->dh_dirblks - 1] != DIRBLKSIZ / DIRALIGN)
786 for (i = dh->dh_dirblks - 1; i >= 0; i--)
787 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
790 return ((doff_t)(i + 1) * DIRBLKSIZ);
794 * Insert information into the hash about a new directory entry. dirp
795 * points to a struct direct containing the entry, and offset specifies
796 * the offset of this entry.
799 ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
804 if ((dh = ufsdirhash_acquire(ip)) == NULL)
807 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
808 ("ufsdirhash_add: bad offset"));
810 * Normal hash usage is < 66%. If the usage gets too high then
811 * remove the hash entirely and let it be rebuilt later.
813 if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
814 ufsdirhash_free_locked(ip);
818 /* Find a free hash slot (empty or deleted), and add the entry. */
819 slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
820 while (DH_ENTRY(dh, slot) >= 0)
821 slot = WRAPINCR(slot, dh->dh_hlen);
822 if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
824 DH_ENTRY(dh, slot) = offset;
826 /* Update last used time. */
827 dh->dh_lastused = time_second;
829 /* Update the per-block summary info. */
830 ufsdirhash_adjfree(dh, offset, -DIRSIZ(0, dirp));
831 ufsdirhash_release(dh);
835 * Remove the specified directory entry from the hash. The entry to remove
836 * is defined by the name in `dirp', which must exist at the specified
837 * `offset' within the directory.
840 ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
845 if ((dh = ufsdirhash_acquire(ip)) == NULL)
848 KASSERT(offset < dh->dh_dirblks * DIRBLKSIZ,
849 ("ufsdirhash_remove: bad offset"));
851 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
853 /* Remove the hash entry. */
854 ufsdirhash_delslot(dh, slot);
856 /* Update the per-block summary info. */
857 ufsdirhash_adjfree(dh, offset, DIRSIZ(0, dirp));
858 ufsdirhash_release(dh);
862 * Change the offset associated with a directory entry in the hash. Used
863 * when compacting directory blocks.
866 ufsdirhash_move(struct inode *ip, struct direct *dirp, doff_t oldoff,
872 if ((dh = ufsdirhash_acquire(ip)) == NULL)
875 KASSERT(oldoff < dh->dh_dirblks * DIRBLKSIZ &&
876 newoff < dh->dh_dirblks * DIRBLKSIZ,
877 ("ufsdirhash_move: bad offset"));
878 /* Find the entry, and update the offset. */
879 slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, oldoff);
880 DH_ENTRY(dh, slot) = newoff;
881 ufsdirhash_release(dh);
885 * Inform dirhash that the directory has grown by one block that
886 * begins at offset (i.e. the new length is offset + DIRBLKSIZ).
889 ufsdirhash_newblk(struct inode *ip, doff_t offset)
894 if ((dh = ufsdirhash_acquire(ip)) == NULL)
897 KASSERT(offset == dh->dh_dirblks * DIRBLKSIZ,
898 ("ufsdirhash_newblk: bad offset"));
899 block = offset / DIRBLKSIZ;
900 if (block >= dh->dh_nblk) {
901 /* Out of space; must rebuild. */
902 ufsdirhash_free_locked(ip);
905 dh->dh_dirblks = block + 1;
907 /* Account for the new free block. */
908 dh->dh_blkfree[block] = DIRBLKSIZ / DIRALIGN;
909 if (dh->dh_firstfree[DH_NFSTATS] == -1)
910 dh->dh_firstfree[DH_NFSTATS] = block;
911 ufsdirhash_release(dh);
915 * Inform dirhash that the directory is being truncated.
918 ufsdirhash_dirtrunc(struct inode *ip, doff_t offset)
923 if ((dh = ufsdirhash_acquire(ip)) == NULL)
926 KASSERT(offset <= dh->dh_dirblks * DIRBLKSIZ,
927 ("ufsdirhash_dirtrunc: bad offset"));
928 block = howmany(offset, DIRBLKSIZ);
930 * If the directory shrinks to less than 1/8 of dh_nblk blocks
931 * (about 20% of its original size due to the 50% extra added in
932 * ufsdirhash_build) then free it, and let the caller rebuild
935 if (block < dh->dh_nblk / 8 && dh->dh_narrays > 1) {
936 ufsdirhash_free_locked(ip);
941 * Remove any `first free' information pertaining to the
942 * truncated blocks. All blocks we're removing should be
945 if (dh->dh_firstfree[DH_NFSTATS] >= block)
946 dh->dh_firstfree[DH_NFSTATS] = -1;
947 for (i = block; i < dh->dh_dirblks; i++)
948 if (dh->dh_blkfree[i] != DIRBLKSIZ / DIRALIGN)
949 panic("ufsdirhash_dirtrunc: blocks in use");
950 for (i = 0; i < DH_NFSTATS; i++)
951 if (dh->dh_firstfree[i] >= block)
952 panic("ufsdirhash_dirtrunc: first free corrupt");
953 dh->dh_dirblks = block;
954 ufsdirhash_release(dh);
958 * Debugging function to check that the dirhash information about
959 * a directory block matches its actual contents. Panics if a mismatch
962 * On entry, `buf' should point to the start of an in-core
963 * DIRBLKSIZ-sized directory block, and `offset' should contain the
964 * offset from the start of the directory of that block.
967 ufsdirhash_checkblock(struct inode *ip, char *buf, doff_t offset)
971 int block, ffslot, i, nfree;
973 if (!ufs_dirhashcheck)
975 if ((dh = ufsdirhash_acquire(ip)) == NULL)
978 block = offset / DIRBLKSIZ;
979 if ((offset & (DIRBLKSIZ - 1)) != 0 || block >= dh->dh_dirblks)
980 panic("ufsdirhash_checkblock: bad offset");
983 for (i = 0; i < DIRBLKSIZ; i += dp->d_reclen) {
984 dp = (struct direct *)(buf + i);
985 if (dp->d_reclen == 0 || i + dp->d_reclen > DIRBLKSIZ)
986 panic("ufsdirhash_checkblock: bad dir");
988 if (dp->d_ino == 0) {
991 * XXX entries with d_ino == 0 should only occur
992 * at the start of a DIRBLKSIZ block. However the
993 * ufs code is tolerant of such entries at other
994 * offsets, and fsck does not fix them.
997 panic("ufsdirhash_checkblock: bad dir inode");
999 nfree += dp->d_reclen;
1003 /* Check that the entry exists (will panic if it doesn't). */
1004 ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
1006 nfree += dp->d_reclen - DIRSIZ(0, dp);
1009 panic("ufsdirhash_checkblock: bad dir end");
1011 if (dh->dh_blkfree[block] * DIRALIGN != nfree)
1012 panic("ufsdirhash_checkblock: bad free count");
1014 ffslot = BLKFREE2IDX(nfree / DIRALIGN);
1015 for (i = 0; i <= DH_NFSTATS; i++)
1016 if (dh->dh_firstfree[i] == block && i != ffslot)
1017 panic("ufsdirhash_checkblock: bad first-free");
1018 if (dh->dh_firstfree[ffslot] == -1)
1019 panic("ufsdirhash_checkblock: missing first-free entry");
1020 ufsdirhash_release(dh);
1024 * Hash the specified filename into a dirhash slot.
1027 ufsdirhash_hash(struct dirhash *dh, char *name, int namelen)
1032 * We hash the name and then some other bit of data that is
1033 * invariant over the dirhash's lifetime. Otherwise names
1034 * differing only in the last byte are placed close to one
1035 * another in the table, which is bad for linear probing.
1037 hash = fnv_32_buf(name, namelen, FNV1_32_INIT);
1038 hash = fnv_32_buf(&dh, sizeof(dh), hash);
1039 return (hash % dh->dh_hlen);
1043 * Adjust the number of free bytes in the block containing `offset'
1044 * by the value specified by `diff'.
1046 * The caller must ensure we have exclusive access to `dh'; normally
1047 * that means that dh_lock should be held, but this is also called
1048 * from ufsdirhash_build() where exclusive access can be assumed.
1051 ufsdirhash_adjfree(struct dirhash *dh, doff_t offset, int diff)
1053 int block, i, nfidx, ofidx;
1055 /* Update the per-block summary info. */
1056 block = offset / DIRBLKSIZ;
1057 KASSERT(block < dh->dh_nblk && block < dh->dh_dirblks,
1058 ("dirhash bad offset"));
1059 ofidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1060 dh->dh_blkfree[block] = (int)dh->dh_blkfree[block] + (diff / DIRALIGN);
1061 nfidx = BLKFREE2IDX(dh->dh_blkfree[block]);
1063 /* Update the `first free' list if necessary. */
1064 if (ofidx != nfidx) {
1065 /* If removing, scan forward for the next block. */
1066 if (dh->dh_firstfree[ofidx] == block) {
1067 for (i = block + 1; i < dh->dh_dirblks; i++)
1068 if (BLKFREE2IDX(dh->dh_blkfree[i]) == ofidx)
1070 dh->dh_firstfree[ofidx] = (i < dh->dh_dirblks) ? i : -1;
1073 /* Make this the new `first free' if necessary */
1074 if (dh->dh_firstfree[nfidx] > block ||
1075 dh->dh_firstfree[nfidx] == -1)
1076 dh->dh_firstfree[nfidx] = block;
1081 * Find the specified name which should have the specified offset.
1082 * Returns a slot number, and panics on failure.
1084 * `dh' must be locked on entry and remains so on return.
1087 ufsdirhash_findslot(struct dirhash *dh, char *name, int namelen, doff_t offset)
1091 DIRHASH_ASSERT_LOCKED(dh);
1093 /* Find the entry. */
1094 KASSERT(dh->dh_hused < dh->dh_hlen, ("dirhash find full"));
1095 slot = ufsdirhash_hash(dh, name, namelen);
1096 while (DH_ENTRY(dh, slot) != offset &&
1097 DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
1098 slot = WRAPINCR(slot, dh->dh_hlen);
1099 if (DH_ENTRY(dh, slot) != offset)
1100 panic("ufsdirhash_findslot: '%.*s' not found", namelen, name);
1106 * Remove the entry corresponding to the specified slot from the hash array.
1108 * `dh' must be locked on entry and remains so on return.
1111 ufsdirhash_delslot(struct dirhash *dh, int slot)
1115 DIRHASH_ASSERT_LOCKED(dh);
1117 /* Mark the entry as deleted. */
1118 DH_ENTRY(dh, slot) = DIRHASH_DEL;
1120 /* If this is the end of a chain of DIRHASH_DEL slots, remove them. */
1121 for (i = slot; DH_ENTRY(dh, i) == DIRHASH_DEL; )
1122 i = WRAPINCR(i, dh->dh_hlen);
1123 if (DH_ENTRY(dh, i) == DIRHASH_EMPTY) {
1124 i = WRAPDECR(i, dh->dh_hlen);
1125 while (DH_ENTRY(dh, i) == DIRHASH_DEL) {
1126 DH_ENTRY(dh, i) = DIRHASH_EMPTY;
1128 i = WRAPDECR(i, dh->dh_hlen);
1130 KASSERT(dh->dh_hused >= 0, ("ufsdirhash_delslot neg hlen"));
1135 * Given a directory entry and its offset, find the offset of the
1136 * previous entry in the same DIRBLKSIZ-sized block. Returns an
1137 * offset, or -1 if there is no previous entry in the block or some
1138 * other problem occurred.
1141 ufsdirhash_getprev(struct direct *dirp, doff_t offset)
1145 doff_t blkoff, prevoff;
1148 blkoff = offset & ~(DIRBLKSIZ - 1); /* offset of start of block */
1149 entrypos = offset & (DIRBLKSIZ - 1); /* entry relative to block */
1150 blkbuf = (char *)dirp - entrypos;
1153 /* If `offset' is the start of a block, there is no previous entry. */
1157 /* Scan from the start of the block until we get to the entry. */
1158 for (i = 0; i < entrypos; i += dp->d_reclen) {
1159 dp = (struct direct *)(blkbuf + i);
1160 if (dp->d_reclen == 0 || i + dp->d_reclen > entrypos)
1161 return (-1); /* Corrupted directory. */
1162 prevoff = blkoff + i;
1168 * Delete the given dirhash and reclaim its memory. Assumes that
1169 * ufsdirhash_list is locked, and leaves it locked. Also assumes
1170 * that dh is locked. Returns the amount of memory freed.
1173 ufsdirhash_destroy(struct dirhash *dh)
1177 int i, mem, narrays;
1179 KASSERT(dh->dh_hash != NULL, ("dirhash: NULL hash on list"));
1181 /* Remove it from the list and detach its memory. */
1182 TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
1186 blkfree = dh->dh_blkfree;
1187 dh->dh_blkfree = NULL;
1188 narrays = dh->dh_narrays;
1189 mem = dh->dh_memreq;
1192 /* Unlock dirhash and free the detached memory. */
1193 ufsdirhash_release(dh);
1194 for (i = 0; i < narrays; i++)
1195 DIRHASH_BLKFREE(hash[i]);
1196 free(hash, M_DIRHASH);
1197 free(blkfree, M_DIRHASH);
1199 /* Account for the returned memory. */
1200 ufs_dirhashmem -= mem;
1206 * Try to free up `wanted' bytes by stealing memory from existing
1207 * dirhashes. Returns zero with list locked if successful.
1210 ufsdirhash_recycle(int wanted)
1215 dh = TAILQ_FIRST(&ufsdirhash_list);
1216 while (wanted + ufs_dirhashmem > ufs_dirhashmaxmem) {
1217 /* Decrement the score; only recycle if it becomes zero. */
1218 if (dh == NULL || --dh->dh_score > 0) {
1219 DIRHASHLIST_UNLOCK();
1223 * If we can't lock it it's in use and we don't want to
1224 * recycle it anyway.
1226 if (!sx_try_xlock(&dh->dh_lock)) {
1227 dh = TAILQ_NEXT(dh, dh_list);
1231 ufsdirhash_destroy(dh);
1233 /* Repeat if necessary. */
1234 dh = TAILQ_FIRST(&ufsdirhash_list);
1236 /* Success; return with list locked. */
1241 * Callback that frees some dirhashes when the system is low on virtual memory.
1246 struct dirhash *dh, *dh_temp;
1247 int memfreed, memwanted;
1249 ufs_dirhashlowmemcount++;
1251 memwanted = ufs_dirhashmem * ufs_dirhashreclaimpercent / 100;
1256 * Reclaim up to memwanted from the oldest dirhashes. This will allow
1257 * us to make some progress when the system is running out of memory
1258 * without compromising the dinamicity of maximum age. If the situation
1259 * does not improve lowmem will be eventually retriggered and free some
1260 * other entry in the cache. The entries on the head of the list should
1261 * be the oldest. If during list traversal we can't get a lock on the
1262 * dirhash, it will be skipped.
1264 TAILQ_FOREACH_SAFE(dh, &ufsdirhash_list, dh_list, dh_temp) {
1265 if (sx_try_xlock(&dh->dh_lock))
1266 memfreed += ufsdirhash_destroy(dh);
1267 if (memfreed >= memwanted)
1270 DIRHASHLIST_UNLOCK();
1274 ufsdirhash_set_reclaimpercent(SYSCTL_HANDLER_ARGS)
1278 v = ufs_dirhashreclaimpercent;
1279 error = sysctl_handle_int(oidp, &v, v, req);
1282 if (req->newptr == NULL)
1284 if (v == ufs_dirhashreclaimpercent)
1287 /* Refuse invalid percentages */
1288 if (v < 0 || v > 100)
1290 ufs_dirhashreclaimpercent = v;
1297 ufs_dirhashmaxmem = lmax(roundup(hibufspace / 64, PAGE_SIZE),
1300 ufsdirhash_zone = uma_zcreate("DIRHASH", DH_NBLKOFF * sizeof(doff_t),
1301 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1302 mtx_init(&ufsdirhash_mtx, "dirhash list", NULL, MTX_DEF);
1303 TAILQ_INIT(&ufsdirhash_list);
1305 /* Register a callback function to handle low memory signals */
1306 EVENTHANDLER_REGISTER(vm_lowmem, ufsdirhash_lowmem, NULL,
1307 EVENTHANDLER_PRI_FIRST);
1313 KASSERT(TAILQ_EMPTY(&ufsdirhash_list), ("ufsdirhash_uninit"));
1314 uma_zdestroy(ufsdirhash_zone);
1315 mtx_destroy(&ufsdirhash_mtx);
1318 #endif /* UFS_DIRHASH */