/*
* Remove the specified directory entry from the hash. The entry to remove
* is defined by the name in `dirp', which must exist at the specified
* `offset' within the directory.
*/
void
ufsdirhash_remove(struct inode *ip, struct direct *dirp, doff_t offset)
{
struct dirhash *dh;
int slot;
const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
int dirblksiz = ip->i_ump->um_dirblksiz;
if ((dh = ip->i_dirhash) == NULL)
return;
DIRHASH_LOCK(dh);
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
ufsdirhash_free(ip);
return;
}
KASSERT(offset < dh->dh_dirblks * dirblksiz);
/* Find the entry */
slot = ufsdirhash_findslot(dh, dirp->d_name, dirp->d_namlen, offset);
/* Remove the hash entry. */
ufsdirhash_delslot(dh, slot);
/* Update the per-block summary info. */
ufsdirhash_adjfree(dh, offset, UFS_DIRSIZ(0, dirp, needswap), dirblksiz);
DIRHASH_UNLOCK(dh);
}
static int
chnamefunc(struct inodesc *idesc)
{
struct direct *dirp = idesc->id_dirp;
struct direct testdir;
if (slotcount++ == desired) {
/* will name fit? */
testdir.d_namlen = strlen(idesc->id_name);
if (UFS_DIRSIZ(UFS_NEWDIRFMT, &testdir, 0) <= iswap16(dirp->d_reclen)) {
dirp->d_namlen = testdir.d_namlen;
strlcpy(dirp->d_name, idesc->id_name,
sizeof(dirp->d_name));
return STOP | ALTERED | FOUND;
} else
return STOP | FOUND; /* won't fit, so give up */
}
return KEEPON;
}
/*
* Insert information into the hash about a new directory entry. dirp
* points to a struct direct containing the entry, and offset specifies
* the offset of this entry.
*/
void
ufsdirhash_add(struct inode *ip, struct direct *dirp, doff_t offset)
{
struct dirhash *dh;
int slot;
const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
int dirblksiz = ip->i_ump->um_dirblksiz;
if ((dh = ip->i_dirhash) == NULL)
return;
DIRHASH_LOCK(dh);
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
ufsdirhash_free(ip);
return;
}
KASSERT(offset < dh->dh_dirblks * dirblksiz);
/*
* Normal hash usage is < 66%. If the usage gets too high then
* remove the hash entirely and let it be rebuilt later.
*/
if (dh->dh_hused >= (dh->dh_hlen * 3) / 4) {
DIRHASH_UNLOCK(dh);
ufsdirhash_free(ip);
return;
}
/* Find a free hash slot (empty or deleted), and add the entry. */
slot = ufsdirhash_hash(dh, dirp->d_name, dirp->d_namlen);
while (DH_ENTRY(dh, slot) >= 0)
slot = WRAPINCR(slot, dh->dh_hlen);
if (DH_ENTRY(dh, slot) == DIRHASH_EMPTY)
dh->dh_hused++;
DH_ENTRY(dh, slot) = offset;
/* Update the per-block summary info. */
ufsdirhash_adjfree(dh, offset, -UFS_DIRSIZ(0, dirp, needswap), dirblksiz);
DIRHASH_UNLOCK(dh);
}
/*
* Debugging function to check that the dirhash information about
* a directory block matches its actual contents. Panics if a mismatch
* is detected.
*
* On entry, `sbuf' should point to the start of an in-core
* DIRBLKSIZ-sized directory block, and `offset' should contain the
* offset from the start of the directory of that block.
*/
void
ufsdirhash_checkblock(struct inode *ip, char *sbuf, doff_t offset)
{
struct dirhash *dh;
struct direct *dp;
int block, ffslot, i, nfree;
const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
int dirblksiz = ip->i_ump->um_dirblksiz;
if (!ufs_dirhashcheck)
return;
if ((dh = ip->i_dirhash) == NULL)
return;
DIRHASH_LOCK(dh);
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
ufsdirhash_free(ip);
return;
}
block = offset / dirblksiz;
if ((offset & (dirblksiz - 1)) != 0 || block >= dh->dh_dirblks)
panic("ufsdirhash_checkblock: bad offset");
nfree = 0;
for (i = 0; i < dirblksiz; i += dp->d_reclen) {
dp = (struct direct *)(sbuf + i);
if (dp->d_reclen == 0 || i + dp->d_reclen > dirblksiz)
panic("ufsdirhash_checkblock: bad dir");
if (dp->d_ino == 0) {
#if 0
/*
* XXX entries with d_ino == 0 should only occur
* at the start of a DIRBLKSIZ block. However the
* ufs code is tolerant of such entries at other
* offsets, and fsck does not fix them.
*/
if (i != 0)
panic("ufsdirhash_checkblock: bad dir inode");
#endif
nfree += dp->d_reclen;
continue;
}
/* Check that the entry exists (will panic if it doesn't). */
ufsdirhash_findslot(dh, dp->d_name, dp->d_namlen, offset + i);
nfree += dp->d_reclen - UFS_DIRSIZ(0, dp, needswap);
}
if (i != dirblksiz)
panic("ufsdirhash_checkblock: bad dir end");
if (dh->dh_blkfree[block] * DIRALIGN != nfree)
panic("ufsdirhash_checkblock: bad free count");
ffslot = BLKFREE2IDX(nfree / DIRALIGN);
for (i = 0; i <= DH_NFSTATS; i++)
if (dh->dh_firstfree[i] == block && i != ffslot)
panic("ufsdirhash_checkblock: bad first-free");
if (dh->dh_firstfree[ffslot] == -1)
panic("ufsdirhash_checkblock: missing first-free entry");
DIRHASH_UNLOCK(dh);
}
/*
* Find a directory block with room for 'slotneeded' bytes. Returns
* the offset of the directory entry that begins the free space.
* This will either be the offset of an existing entry that has free
* space at the end, or the offset of an entry with d_ino == 0 at
* the start of a UFS_DIRBLKSIZ block.
*
* To use the space, the caller may need to compact existing entries in
* the directory. The total number of bytes in all of the entries involved
* in the compaction is stored in *slotsize. In other words, all of
* the entries that must be compacted are exactly contained in the
* region beginning at the returned offset and spanning *slotsize bytes.
*
* Returns -1 if no space was found, indicating that the directory
* must be extended.
*/
doff_t
ufsdirhash_findfree(struct inode *ip, int slotneeded, int *slotsize)
{
struct direct *dp;
struct dirhash *dh;
struct buf *bp;
doff_t pos, slotstart;
int dirblock, error, freebytes, i;
const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
int dirblksiz = ip->i_ump->um_dirblksiz;
if ((dh = ip->i_dirhash) == NULL)
return (-1);
DIRHASH_LOCK(dh);
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
ufsdirhash_free(ip);
return (-1);
}
/* Find a directory block with the desired free space. */
dirblock = -1;
for (i = howmany(slotneeded, DIRALIGN); i <= DH_NFSTATS; i++)
if ((dirblock = dh->dh_firstfree[i]) != -1)
break;
if (dirblock == -1) {
DIRHASH_UNLOCK(dh);
return (-1);
}
KASSERT(dirblock < dh->dh_nblk &&
dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN));
pos = dirblock * dirblksiz;
error = ufs_blkatoff(ip->i_vnode, (off_t)pos, (void *)&dp, &bp, false);
if (error) {
DIRHASH_UNLOCK(dh);
return (-1);
}
/* Find the first entry with free space. */
for (i = 0; i < dirblksiz; ) {
if (dp->d_reclen == 0) {
DIRHASH_UNLOCK(dh);
brelse(bp, 0);
return (-1);
}
if (dp->d_ino == 0 || dp->d_reclen > UFS_DIRSIZ(0, dp, needswap))
break;
i += dp->d_reclen;
dp = (struct direct *)((char *)dp + dp->d_reclen);
}
if (i > dirblksiz) {
DIRHASH_UNLOCK(dh);
brelse(bp, 0);
return (-1);
}
slotstart = pos + i;
/* Find the range of entries needed to get enough space */
freebytes = 0;
while (i < dirblksiz && freebytes < slotneeded) {
freebytes += dp->d_reclen;
if (dp->d_ino != 0)
freebytes -= UFS_DIRSIZ(0, dp, needswap);
if (dp->d_reclen == 0) {
DIRHASH_UNLOCK(dh);
brelse(bp, 0);
return (-1);
}
i += dp->d_reclen;
dp = (struct direct *)((char *)dp + dp->d_reclen);
}
if (i > dirblksiz) {
DIRHASH_UNLOCK(dh);
brelse(bp, 0);
return (-1);
}
if (freebytes < slotneeded)
panic("ufsdirhash_findfree: free mismatch");
DIRHASH_UNLOCK(dh);
brelse(bp, 0);
*slotsize = pos + i - slotstart;
return (slotstart);
}
/*
* Find the offset of the specified name within the given inode.
* Returns 0 on success, ENOENT if the entry does not exist, or
* EJUSTRETURN if the caller should revert to a linear search.
*
* If successful, the directory offset is stored in *offp, and a
* pointer to a struct buf containing the entry is stored in *bpp. If
* prevoffp is non-NULL, the offset of the previous entry within
* the UFS_DIRBLKSIZ-sized block is stored in *prevoffp (if the entry
* is the first in a block, the start of the block is used).
*/
int
ufsdirhash_lookup(struct inode *ip, const char *name, int namelen, doff_t *offp,
struct buf **bpp, doff_t *prevoffp)
{
struct dirhash *dh, *dh_next;
struct direct *dp;
struct vnode *vp;
struct buf *bp;
doff_t blkoff, bmask, offset, prevoff;
int i, slot;
const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
int dirblksiz = ip->i_ump->um_dirblksiz;
if ((dh = ip->i_dirhash) == NULL)
return (EJUSTRETURN);
/*
* Move this dirhash towards the end of the list if it has a
* score higher than the next entry, and acquire the dh_lock.
* Optimise the case where it's already the last by performing
* an unlocked read of the TAILQ_NEXT pointer.
*
* In both cases, end up holding just dh_lock.
*/
if (TAILQ_NEXT(dh, dh_list) != NULL) {
DIRHASHLIST_LOCK();
DIRHASH_LOCK(dh);
/*
* If the new score will be greater than that of the next
* entry, then move this entry past it. With both mutexes
* held, dh_next won't go away, but its dh_score could
* change; that's not important since it is just a hint.
*/
if (dh->dh_hash != NULL &&
(dh_next = TAILQ_NEXT(dh, dh_list)) != NULL &&
dh->dh_score >= dh_next->dh_score) {
KASSERT(dh->dh_onlist);
TAILQ_REMOVE(&ufsdirhash_list, dh, dh_list);
TAILQ_INSERT_AFTER(&ufsdirhash_list, dh_next, dh,
dh_list);
}
DIRHASHLIST_UNLOCK();
} else {
/* Already the last, though that could change as we wait. */
DIRHASH_LOCK(dh);
}
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
ufsdirhash_free(ip);
return (EJUSTRETURN);
}
/* Update the score. */
if (dh->dh_score < DH_SCOREMAX)
dh->dh_score++;
vp = ip->i_vnode;
bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
blkoff = -1;
bp = NULL;
restart:
slot = ufsdirhash_hash(dh, name, namelen);
if (dh->dh_seqopt) {
/*
* Sequential access optimisation. dh_seqoff contains the
* offset of the directory entry immediately following
* the last entry that was looked up. Check if this offset
* appears in the hash chain for the name we are looking for.
*/
for (i = slot; (offset = DH_ENTRY(dh, i)) != DIRHASH_EMPTY;
i = WRAPINCR(i, dh->dh_hlen))
if (offset == dh->dh_seqoff)
break;
if (offset == dh->dh_seqoff) {
/*
* We found an entry with the expected offset. This
* is probably the entry we want, but if not, the
* code below will turn off seqoff and retry.
*/
slot = i;
} else
dh->dh_seqopt = 0;
}
for (; (offset = DH_ENTRY(dh, slot)) != DIRHASH_EMPTY;
slot = WRAPINCR(slot, dh->dh_hlen)) {
if (offset == DIRHASH_DEL)
continue;
if (offset < 0 || offset >= ip->i_size)
panic("ufsdirhash_lookup: bad offset in hash array");
if ((offset & ~bmask) != blkoff) {
if (bp != NULL)
brelse(bp, 0);
blkoff = offset & ~bmask;
if (ufs_blkatoff(vp, (off_t)blkoff,
NULL, &bp, false) != 0) {
DIRHASH_UNLOCK(dh);
return (EJUSTRETURN);
}
}
dp = (struct direct *)((char *)bp->b_data + (offset & bmask));
if (dp->d_reclen == 0 || dp->d_reclen >
dirblksiz - (offset & (dirblksiz - 1))) {
/* Corrupted directory. */
DIRHASH_UNLOCK(dh);
brelse(bp, 0);
return (EJUSTRETURN);
}
if (dp->d_namlen == namelen &&
memcmp(dp->d_name, name, namelen) == 0) {
/* Found. Get the prev offset if needed. */
if (prevoffp != NULL) {
if (offset & (dirblksiz - 1)) {
prevoff = ufsdirhash_getprev(dp,
offset, dirblksiz);
if (prevoff == -1) {
brelse(bp, 0);
return (EJUSTRETURN);
}
} else
prevoff = offset;
*prevoffp = prevoff;
}
/* Check for sequential access, and update offset. */
if (dh->dh_seqopt == 0 && dh->dh_seqoff == offset)
dh->dh_seqopt = 1;
dh->dh_seqoff = offset + UFS_DIRSIZ(0, dp, needswap);
DIRHASH_UNLOCK(dh);
*bpp = bp;
*offp = offset;
return (0);
}
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
if (bp != NULL)
brelse(bp, 0);
ufsdirhash_free(ip);
return (EJUSTRETURN);
}
/*
* When the name doesn't match in the seqopt case, go back
* and search normally.
*/
if (dh->dh_seqopt) {
dh->dh_seqopt = 0;
goto restart;
}
}
DIRHASH_UNLOCK(dh);
if (bp != NULL)
brelse(bp, 0);
return (ENOENT);
}
/*
* Attempt to build up a hash table for the directory contents in
* inode 'ip'. Returns 0 on success, or -1 of the operation failed.
*/
int
ufsdirhash_build(struct inode *ip)
{
struct dirhash *dh;
struct buf *bp = NULL;
struct direct *ep;
struct vnode *vp;
doff_t bmask, pos;
int dirblocks, i, j, memreqd, nblocks, narrays, nslots, slot;
const int needswap = UFS_MPNEEDSWAP(ip->i_ump);
int dirblksiz = ip->i_ump->um_dirblksiz;
/* Check if we can/should use dirhash. */
if (ip->i_dirhash == NULL) {
if (ip->i_size < (ufs_dirhashminblks * dirblksiz) || OFSFMT(ip))
return (-1);
} else {
/* Hash exists, but sysctls could have changed. */
if (ip->i_size < (ufs_dirhashminblks * dirblksiz) ||
ufs_dirhashmem > ufs_dirhashmaxmem) {
ufsdirhash_free(ip);
return (-1);
}
/* Check if hash exists and is intact (note: unlocked read). */
if (ip->i_dirhash->dh_hash != NULL)
return (0);
/* Free the old, recycled hash and build a new one. */
ufsdirhash_free(ip);
}
/* Don't hash removed directories. */
if (ip->i_nlink == 0)
return (-1);
vp = ip->i_vnode;
/* Allocate 50% more entries than this dir size could ever need. */
KASSERT(ip->i_size >= dirblksiz);
nslots = ip->i_size / UFS_DIRECTSIZ(1);
nslots = (nslots * 3 + 1) / 2;
narrays = howmany(nslots, DH_NBLKOFF);
nslots = narrays * DH_NBLKOFF;
dirblocks = howmany(ip->i_size, dirblksiz);
nblocks = (dirblocks * 3 + 1) / 2;
memreqd = sizeof(*dh) + narrays * sizeof(*dh->dh_hash) +
narrays * DH_NBLKOFF * sizeof(**dh->dh_hash) +
nblocks * sizeof(*dh->dh_blkfree);
while (atomic_add_int_nv(&ufs_dirhashmem, memreqd) >
ufs_dirhashmaxmem) {
atomic_add_int(&ufs_dirhashmem, -memreqd);
if (memreqd > ufs_dirhashmaxmem / 2)
return (-1);
/* Try to free some space. */
if (ufsdirhash_recycle(memreqd) != 0)
return (-1);
else
DIRHASHLIST_UNLOCK();
}
/*
* Use non-blocking mallocs so that we will revert to a linear
* lookup on failure rather than potentially blocking forever.
*/
dh = pool_cache_get(ufsdirhash_cache, PR_NOWAIT);
if (dh == NULL) {
atomic_add_int(&ufs_dirhashmem, -memreqd);
return (-1);
}
memset(dh, 0, sizeof(*dh));
mutex_init(&dh->dh_lock, MUTEX_DEFAULT, IPL_NONE);
DIRHASH_LOCK(dh);
dh->dh_hashsz = narrays * sizeof(dh->dh_hash[0]);
dh->dh_hash = kmem_zalloc(dh->dh_hashsz, KM_NOSLEEP);
dh->dh_blkfreesz = nblocks * sizeof(dh->dh_blkfree[0]);
dh->dh_blkfree = kmem_zalloc(dh->dh_blkfreesz, KM_NOSLEEP);
if (dh->dh_hash == NULL || dh->dh_blkfree == NULL)
goto fail;
for (i = 0; i < narrays; i++) {
if ((dh->dh_hash[i] = DIRHASH_BLKALLOC()) == NULL)
goto fail;
for (j = 0; j < DH_NBLKOFF; j++)
dh->dh_hash[i][j] = DIRHASH_EMPTY;
}
/* Initialise the hash table and block statistics. */
dh->dh_narrays = narrays;
dh->dh_hlen = nslots;
dh->dh_nblk = nblocks;
dh->dh_dirblks = dirblocks;
for (i = 0; i < dirblocks; i++)
dh->dh_blkfree[i] = dirblksiz / DIRALIGN;
for (i = 0; i < DH_NFSTATS; i++)
dh->dh_firstfree[i] = -1;
dh->dh_firstfree[DH_NFSTATS] = 0;
dh->dh_seqopt = 0;
dh->dh_seqoff = 0;
dh->dh_score = DH_SCOREINIT;
ip->i_dirhash = dh;
bmask = VFSTOUFS(vp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
pos = 0;
while (pos < ip->i_size) {
if ((curcpu()->ci_schedstate.spc_flags & SPCF_SHOULDYIELD)
!= 0) {
preempt();
}
/* If necessary, get the next directory block. */
if ((pos & bmask) == 0) {
if (bp != NULL)
brelse(bp, 0);
if (ufs_blkatoff(vp, (off_t)pos, NULL, &bp, false) != 0)
goto fail;
}
/* Add this entry to the hash. */
ep = (struct direct *)((char *)bp->b_data + (pos & bmask));
if (ep->d_reclen == 0 || ep->d_reclen >
dirblksiz - (pos & (dirblksiz - 1))) {
/* Corrupted directory. */
brelse(bp, 0);
goto fail;
}
if (ep->d_ino != 0) {
/* Add the entry (simplified ufsdirhash_add). */
slot = ufsdirhash_hash(dh, ep->d_name, ep->d_namlen);
while (DH_ENTRY(dh, slot) != DIRHASH_EMPTY)
slot = WRAPINCR(slot, dh->dh_hlen);
dh->dh_hused++;
DH_ENTRY(dh, slot) = pos;
ufsdirhash_adjfree(dh, pos, -UFS_DIRSIZ(0, ep, needswap),
dirblksiz);
}
pos += ep->d_reclen;
}
if (bp != NULL)
brelse(bp, 0);
DIRHASHLIST_LOCK();
TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
dh->dh_onlist = 1;
DIRHASH_UNLOCK(dh);
DIRHASHLIST_UNLOCK();
return (0);
fail:
DIRHASH_UNLOCK(dh);
if (dh->dh_hash != NULL) {
for (i = 0; i < narrays; i++)
if (dh->dh_hash[i] != NULL)
DIRHASH_BLKFREE(dh->dh_hash[i]);
kmem_free(dh->dh_hash, dh->dh_hashsz);
}
if (dh->dh_blkfree != NULL)
kmem_free(dh->dh_blkfree, dh->dh_blkfreesz);
mutex_destroy(&dh->dh_lock);
pool_cache_put(ufsdirhash_cache, dh);
ip->i_dirhash = NULL;
atomic_add_int(&ufs_dirhashmem, -memreqd);
return (-1);
}
