/*
* Rewrite an existing directory entry to point at the inode
* supplied. The parameters describing the directory entry are
* set up by a call to namei.
*/
int
ufs_dirrewrite(struct inode *dp, struct inode *oip, ufsino_t newinum,
int newtype, int isrmdir)
{
struct buf *bp;
struct direct *ep;
struct vnode *vdp = ITOV(dp);
int error;
error = UFS_BUFATOFF(dp, (off_t)dp->i_offset, (char **)&ep, &bp);
if (error)
return (error);
ep->d_ino = newinum;
if (vdp->v_mount->mnt_maxsymlinklen > 0)
ep->d_type = newtype;
oip->i_effnlink--;
if (DOINGSOFTDEP(vdp)) {
softdep_change_linkcnt(oip, 0);
softdep_setup_directory_change(bp, dp, oip, newinum, isrmdir);
bdwrite(bp);
} else {
DIP_ADD(oip, nlink, -1);
oip->i_flag |= IN_CHANGE;
UFS_WAPBL_UPDATE(oip, MNT_WAIT);
if (DOINGASYNC(vdp)) {
bdwrite(bp);
error = 0;
} else {
error = VOP_BWRITE(bp);
}
}
dp->i_flag |= IN_CHANGE | IN_UPDATE;
UFS_WAPBL_UPDATE(dp, MNT_WAIT);
return (error);
}
/*
* Remove a directory entry after a call to namei, using
* the parameters which it left in nameidata. The entry
* dp->i_offset contains the offset into the directory of the
* entry to be eliminated. The dp->i_count field contains the
* size of the previous record in the directory. If this
* is 0, the first entry is being deleted, so we need only
* zero the inode number to mark the entry as free. If the
* entry is not the first in the directory, we must reclaim
* the space of the now empty record by adding the record size
* to the size of the previous entry.
*/
int
ufs_dirremove(struct vnode *dvp, struct inode *ip, int flags, int isrmdir)
{
struct inode *dp;
struct direct *ep;
struct buf *bp;
int error;
UFS_WAPBL_JLOCK_ASSERT(dvp->v_mount);
dp = VTOI(dvp);
if ((error = UFS_BUFATOFF(dp,
(off_t)(dp->i_offset - dp->i_count), (char **)&ep, &bp)) != 0)
return (error);
#ifdef UFS_DIRHASH
/*
* Remove the dirhash entry. This is complicated by the fact
* that `ep' is the previous entry when dp->i_count != 0.
*/
if (dp->i_dirhash != NULL)
ufsdirhash_remove(dp, (dp->i_count == 0) ? ep :
(struct direct *)((char *)ep + ep->d_reclen), dp->i_offset);
#endif
if (dp->i_count == 0) {
/*
* First entry in block: set d_ino to zero.
*/
ep->d_ino = 0;
} else {
/*
* Collapse new free space into previous entry.
*/
ep->d_reclen += dp->i_reclen;
}
#ifdef UFS_DIRHASH
if (dp->i_dirhash != NULL)
ufsdirhash_checkblock(dp, (char *)ep -
((dp->i_offset - dp->i_count) & (DIRBLKSIZ - 1)),
dp->i_offset & ~(DIRBLKSIZ - 1));
#endif
if (DOINGSOFTDEP(dvp)) {
if (ip) {
ip->i_effnlink--;
softdep_change_linkcnt(ip, 0);
softdep_setup_remove(bp, dp, ip, isrmdir);
}
if (softdep_slowdown(dvp)) {
error = bwrite(bp);
} else {
bdwrite(bp);
error = 0;
}
} else {
if (ip) {
ip->i_effnlink--;
DIP_ADD(ip, nlink, -1);
ip->i_flag |= IN_CHANGE;
UFS_WAPBL_UPDATE(ip, 0);
}
if (DOINGASYNC(dvp) && dp->i_count != 0) {
bdwrite(bp);
error = 0;
} else
error = bwrite(bp);
}
dp->i_flag |= IN_CHANGE | IN_UPDATE;
UFS_WAPBL_UPDATE(dp, 0);
return (error);
}
/*
* Write a directory entry after a call to namei, using the parameters
* that it left in nameidata. The argument dirp is the new directory
* entry contents. Dvp is a pointer to the directory to be written,
* which was left locked by namei. Remaining parameters (dp->i_offset,
* dp->i_count) indicate how the space for the new entry is to be obtained.
* Non-null bp indicates that a directory is being created (for the
* soft dependency code).
*/
int
ufs_direnter(struct vnode *dvp, struct vnode *tvp, struct direct *dirp,
struct componentname *cnp, struct buf *newdirbp)
{
struct ucred *cr;
struct proc *p;
int newentrysize;
struct inode *dp;
struct buf *bp;
u_int dsize;
struct direct *ep, *nep;
int error, ret, blkoff, loc, spacefree, flags;
char *dirbuf;
UFS_WAPBL_JLOCK_ASSERT(dvp->v_mount);
error = 0;
cr = cnp->cn_cred;
p = cnp->cn_proc;
dp = VTOI(dvp);
newentrysize = DIRSIZ(FSFMT(dvp), dirp);
if (dp->i_count == 0) {
/*
* If dp->i_count is 0, then namei could find no
* space in the directory. Here, dp->i_offset will
* be on a directory block boundary and we will write the
* new entry into a fresh block.
*/
if (dp->i_offset & (DIRBLKSIZ - 1))
panic("ufs_direnter: newblk");
flags = B_CLRBUF;
if (!DOINGSOFTDEP(dvp))
flags |= B_SYNC;
if ((error = UFS_BUF_ALLOC(dp, (off_t)dp->i_offset, DIRBLKSIZ,
cr, flags, &bp)) != 0) {
if (DOINGSOFTDEP(dvp) && newdirbp != NULL)
bdwrite(newdirbp);
return (error);
}
DIP_ASSIGN(dp, size, dp->i_offset + DIRBLKSIZ);
dp->i_flag |= IN_CHANGE | IN_UPDATE;
uvm_vnp_setsize(dvp, DIP(dp, size));
dirp->d_reclen = DIRBLKSIZ;
blkoff = dp->i_offset &
(VFSTOUFS(dvp->v_mount)->um_mountp->mnt_stat.f_iosize - 1);
memcpy(bp->b_data + blkoff, dirp, newentrysize);
#ifdef UFS_DIRHASH
if (dp->i_dirhash != NULL) {
ufsdirhash_newblk(dp, dp->i_offset);
ufsdirhash_add(dp, dirp, dp->i_offset);
ufsdirhash_checkblock(dp, (char *)bp->b_data + blkoff,
dp->i_offset);
}
#endif
if (DOINGSOFTDEP(dvp)) {
/*
* Ensure that the entire newly allocated block is a
* valid directory so that future growth within the
* block does not have to ensure that the block is
* written before the inode.
*/
blkoff += DIRBLKSIZ;
while (blkoff < bp->b_bcount) {
((struct direct *)
(bp->b_data + blkoff))->d_reclen = DIRBLKSIZ;
blkoff += DIRBLKSIZ;
}
if (softdep_setup_directory_add(bp, dp, dp->i_offset,
dirp->d_ino, newdirbp, 1) == 0) {
bdwrite(bp);
return (UFS_UPDATE(dp, 0));
}
/* We have just allocated a directory block in an
* indirect block. Rather than tracking when it gets
* claimed by the inode, we simply do a VOP_FSYNC
* now to ensure that it is there (in case the user
* does a future fsync). Note that we have to unlock
* the inode for the entry that we just entered, as
* the VOP_FSYNC may need to lock other inodes which
* can lead to deadlock if we also hold a lock on
* the newly entered node.
*/
if ((error = VOP_BWRITE(bp)))
return (error);
if (tvp != NULL)
VOP_UNLOCK(tvp, 0);
error = VOP_FSYNC(dvp, p->p_ucred, MNT_WAIT);
if (tvp != NULL)
vn_lock(tvp, LK_EXCLUSIVE | LK_RETRY, p);
return (error);
}
error = VOP_BWRITE(bp);
ret = UFS_UPDATE(dp, !DOINGSOFTDEP(dvp));
if (error == 0)
return (ret);
return (error);
}
/*
* If dp->i_count is non-zero, then namei found space for the new
* entry in the range dp->i_offset to dp->i_offset + dp->i_count
* in the directory. To use this space, we may have to compact
* the entries located there, by copying them together towards the
* beginning of the block, leaving the free space in one usable
* chunk at the end.
*/
/*
* Increase size of directory if entry eats into new space.
* This should never push the size past a new multiple of
* DIRBLKSIZE.
*
* N.B. - THIS IS AN ARTIFACT OF 4.2 AND SHOULD NEVER HAPPEN.
*/
if (dp->i_offset + dp->i_count > DIP(dp, size)) {
DIP_ASSIGN(dp, size, dp->i_offset + dp->i_count);
dp->i_flag |= IN_CHANGE | IN_UPDATE;
UFS_WAPBL_UPDATE(dp, MNT_WAIT);
}
/*
* Get the block containing the space for the new directory entry.
*/
if ((error = UFS_BUFATOFF(dp, (off_t)dp->i_offset, &dirbuf, &bp))
!= 0) {
if (DOINGSOFTDEP(dvp) && newdirbp != NULL)
bdwrite(newdirbp);
return (error);
}
/*
* Find space for the new entry. In the simple case, the entry at
* offset base will have the space. If it does not, then namei
* arranged that compacting the region dp->i_offset to
* dp->i_offset + dp->i_count would yield the space.
*/
ep = (struct direct *)dirbuf;
dsize = ep->d_ino ? DIRSIZ(FSFMT(dvp), ep) : 0;
spacefree = ep->d_reclen - dsize;
for (loc = ep->d_reclen; loc < dp->i_count; ) {
nep = (struct direct *)(dirbuf + loc);
/* Trim the existing slot (NB: dsize may be zero). */
ep->d_reclen = dsize;
ep = (struct direct *)((char *)ep + dsize);
/* Read nep->d_reclen now as the memmove() may clobber it. */
loc += nep->d_reclen;
if (nep->d_ino == 0) {
/*
* A mid-block unused entry. Such entries are
* never created by the kernel, but fsck_ffs
* can create them (and it doesn't fix them).
*
* Add up the free space, and initialise the
* relocated entry since we don't memmove it.
*/
spacefree += nep->d_reclen;
ep->d_ino = 0;
dsize = 0;
continue;
}
dsize = DIRSIZ(FSFMT(dvp), nep);
spacefree += nep->d_reclen - dsize;
#ifdef UFS_DIRHASH
if (dp->i_dirhash != NULL)
ufsdirhash_move(dp, nep,
dp->i_offset + ((char *)nep - dirbuf),
dp->i_offset + ((char *)ep - dirbuf));
#endif
if (DOINGSOFTDEP(dvp))
softdep_change_directoryentry_offset(dp, dirbuf,
(caddr_t)nep, (caddr_t)ep, dsize);
else
memmove(ep, nep, dsize);
}
/*
* Here, `ep' points to a directory entry containing `dsize' in-use
* bytes followed by `spacefree' unused bytes. If ep->d_ino == 0,
* then the entry is completely unused (dsize == 0). The value
* of ep->d_reclen is always indeterminate.
*
* Update the pointer fields in the previous entry (if any),
* copy in the new entry, and write out the block.
*/
if (ep->d_ino == 0) {
if (spacefree + dsize < newentrysize)
panic("ufs_direnter: compact1");
dirp->d_reclen = spacefree + dsize;
} else {
if (spacefree < newentrysize)
panic("ufs_direnter: compact2");
dirp->d_reclen = spacefree;
ep->d_reclen = dsize;
ep = (struct direct *)((char *)ep + dsize);
}
#ifdef UFS_DIRHASH
if (dp->i_dirhash != NULL && (ep->d_ino == 0 ||
dirp->d_reclen == spacefree))
ufsdirhash_add(dp, dirp, dp->i_offset + ((char *)ep - dirbuf));
#endif
memcpy(ep, dirp, newentrysize);
#ifdef UFS_DIRHASH
if (dp->i_dirhash != NULL)
ufsdirhash_checkblock(dp, dirbuf -
(dp->i_offset & (DIRBLKSIZ - 1)),
dp->i_offset & ~(DIRBLKSIZ - 1));
#endif
if (DOINGSOFTDEP(dvp)) {
(void)softdep_setup_directory_add(bp, dp,
dp->i_offset + (caddr_t)ep - dirbuf,
dirp->d_ino, newdirbp, 0);
bdwrite(bp);
} else {
error = VOP_BWRITE(bp);
}
dp->i_flag |= IN_CHANGE | IN_UPDATE;
/*
* If all went well, and the directory can be shortened, proceed
* with the truncation. Note that we have to unlock the inode for
* the entry that we just entered, as the truncation may need to
* lock other inodes which can lead to deadlock if we also hold a
* lock on the newly entered node.
*/
if (error == 0 && dp->i_endoff && dp->i_endoff < DIP(dp, size)) {
if (tvp != NULL)
VOP_UNLOCK(tvp, 0);
#ifdef UFS_DIRHASH
if (dp->i_dirhash != NULL)
ufsdirhash_dirtrunc(dp, dp->i_endoff);
#endif
error = UFS_TRUNCATE(dp, (off_t)dp->i_endoff, IO_SYNC, cr);
if (tvp != NULL)
vn_lock(tvp, LK_EXCLUSIVE | LK_RETRY, p);
}
UFS_WAPBL_UPDATE(dp, MNT_WAIT);
return (error);
}
/*
* Convert a component of a pathname into a pointer to a locked inode.
* This is a very central and rather complicated routine.
* If the file system is not maintained in a strict tree hierarchy,
* this can result in a deadlock situation (see comments in code below).
*
* The cnp->cn_nameiop argument is LOOKUP, CREATE, RENAME, or DELETE depending
* on whether the name is to be looked up, created, renamed, or deleted.
* When CREATE, RENAME, or DELETE is specified, information usable in
* creating, renaming, or deleting a directory entry may be calculated.
* If flag has LOCKPARENT or'ed into it and the target of the pathname
* exists, lookup returns both the target and its parent directory locked.
* When creating or renaming and LOCKPARENT is specified, the target may
* not be ".". When deleting and LOCKPARENT is specified, the target may
* be "."., but the caller must check to ensure it does an vrele and vput
* instead of two vputs.
*
* Overall outline of ufs_lookup:
*
* check accessibility of directory
* look for name in cache, if found, then if at end of path
* and deleting or creating, drop it, else return name
* search for name in directory, to found or notfound
* notfound:
* if creating, return locked directory, leaving info on available slots
* else return error
* found:
* if at end of path and deleting, return information to allow delete
* if at end of path and rewriting (RENAME and LOCKPARENT), lock target
* inode and return info to allow rewrite
* if not at end, add name to cache; if at end and neither creating
* nor deleting, add name to cache
*/
int
ufs_lookup(void *v)
{
struct vop_lookup_args *ap = v;
struct vnode *vdp; /* vnode for directory being searched */
struct inode *dp; /* inode for directory being searched */
struct buf *bp; /* a buffer of directory entries */
struct direct *ep; /* the current directory entry */
int entryoffsetinblock; /* offset of ep in bp's buffer */
enum {NONE, COMPACT, FOUND} slotstatus;
doff_t slotoffset; /* offset of area with free space */
int slotsize; /* size of area at slotoffset */
int slotfreespace; /* amount of space free in slot */
int slotneeded; /* size of the entry we're seeking */
int numdirpasses; /* strategy for directory search */
doff_t endsearch; /* offset to end directory search */
doff_t prevoff; /* prev entry dp->i_offset */
struct vnode *pdp; /* saved dp during symlink work */
struct vnode *tdp; /* returned by VFS_VGET */
doff_t enduseful; /* pointer past last used dir slot */
u_long bmask; /* block offset mask */
int lockparent; /* 1 => lockparent flag is set */
int wantparent; /* 1 => wantparent or lockparent flag */
int namlen, error;
struct vnode **vpp = ap->a_vpp;
struct componentname *cnp = ap->a_cnp;
struct ucred *cred = cnp->cn_cred;
int flags;
int nameiop = cnp->cn_nameiop;
struct proc *p = cnp->cn_proc;
cnp->cn_flags &= ~PDIRUNLOCK;
flags = cnp->cn_flags;
bp = NULL;
slotoffset = -1;
*vpp = NULL;
vdp = ap->a_dvp;
dp = VTOI(vdp);
lockparent = flags & LOCKPARENT;
wantparent = flags & (LOCKPARENT|WANTPARENT);
/*
* Check accessiblity of directory.
*/
if ((DIP(dp, mode) & IFMT) != IFDIR)
return (ENOTDIR);
if ((error = VOP_ACCESS(vdp, VEXEC, cred)) != 0)
return (error);
if ((flags & ISLASTCN) && (vdp->v_mount->mnt_flag & MNT_RDONLY) &&
(cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
return (EROFS);
/*
* We now have a segment name to search for, and a directory to search.
*
* Before tediously performing a linear scan of the directory,
* check the name cache to see if the directory/name pair
* we are looking for is known already.
*/
if ((error = cache_lookup(vdp, vpp, cnp)) >= 0)
return (error);
/*
* Suppress search for slots unless creating
* file and at end of pathname, in which case
* we watch for a place to put the new file in
* case it doesn't already exist.
*/
slotstatus = FOUND;
slotfreespace = slotsize = slotneeded = 0;
if ((nameiop == CREATE || nameiop == RENAME) &&
(flags & ISLASTCN)) {
slotstatus = NONE;
slotneeded = (sizeof(struct direct) - MAXNAMLEN +
cnp->cn_namelen + 3) &~ 3;
}
/*
* If there is cached information on a previous search of
* this directory, pick up where we last left off.
* We cache only lookups as these are the most common
* and have the greatest payoff. Caching CREATE has little
* benefit as it usually must search the entire directory
* to determine that the entry does not exist. Caching the
* location of the last DELETE or RENAME has not reduced
* profiling time and hence has been removed in the interest
* of simplicity.
*/
bmask = VFSTOUFS(vdp->v_mount)->um_mountp->mnt_stat.f_iosize - 1;
#ifdef UFS_DIRHASH
/*
* Use dirhash for fast operations on large directories. The logic
* to determine whether to hash the directory is contained within
* ufsdirhash_build(); a zero return means that it decided to hash
* this directory and it successfully built up the hash table.
*/
if (ufsdirhash_build(dp) == 0) {
/* Look for a free slot if needed. */
enduseful = DIP(dp, size);
if (slotstatus != FOUND) {
slotoffset = ufsdirhash_findfree(dp, slotneeded,
&slotsize);
if (slotoffset >= 0) {
slotstatus = COMPACT;
enduseful = ufsdirhash_enduseful(dp);
if (enduseful < 0)
enduseful = DIP(dp, size);
}
}
/* Look up the component. */
numdirpasses = 1;
entryoffsetinblock = 0; /* silence compiler warning */
switch (ufsdirhash_lookup(dp, cnp->cn_nameptr, cnp->cn_namelen,
&dp->i_offset, &bp, nameiop == DELETE ? &prevoff : NULL)) {
case 0:
ep = (struct direct *)((char *)bp->b_data +
(dp->i_offset & bmask));
goto foundentry;
case ENOENT:
#define roundup2(x, y) (((x)+((y)-1))&(~((y)-1))) /* if y is powers of two */
dp->i_offset = roundup2(DIP(dp, size), DIRBLKSIZ);
goto notfound;
default:
/* Something failed; just do a linear search. */
break;
}
}
#endif /* UFS_DIRHASH */
if (nameiop != LOOKUP || dp->i_diroff == 0 ||
dp->i_diroff >= DIP(dp, size)) {
entryoffsetinblock = 0;
dp->i_offset = 0;
numdirpasses = 1;
} else {
dp->i_offset = dp->i_diroff;
if ((entryoffsetinblock = dp->i_offset & bmask) &&
(error = UFS_BUFATOFF(dp, (off_t)dp->i_offset, NULL, &bp)))
return (error);
numdirpasses = 2;
nchstats.ncs_2passes++;
}
prevoff = dp->i_offset;
endsearch = roundup(DIP(dp, size), DIRBLKSIZ);
enduseful = 0;
searchloop:
while (dp->i_offset < endsearch) {
/*
* If necessary, get the next directory block.
*/
if ((dp->i_offset & bmask) == 0) {
if (bp != NULL)
brelse(bp);
error = UFS_BUFATOFF(dp, (off_t)dp->i_offset, NULL,
&bp);
if (error)
return (error);
entryoffsetinblock = 0;
}
/*
* If still looking for a slot, and at a DIRBLKSIZE
* boundary, have to start looking for free space again.
*/
if (slotstatus == NONE &&
(entryoffsetinblock & (DIRBLKSIZ - 1)) == 0) {
slotoffset = -1;
slotfreespace = 0;
}
/*
* Get pointer to next entry.
* Full validation checks are slow, so we only check
* enough to insure forward progress through the
* directory. Complete checks can be run by patching
* "dirchk" to be true.
*/
ep = (struct direct *)((char *)bp->b_data + entryoffsetinblock);
if (ep->d_reclen == 0 ||
(dirchk && ufs_dirbadentry(vdp, ep, entryoffsetinblock))) {
int i;
ufs_dirbad(dp, dp->i_offset, "mangled entry");
i = DIRBLKSIZ - (entryoffsetinblock & (DIRBLKSIZ - 1));
dp->i_offset += i;
entryoffsetinblock += i;
continue;
}
/*
* If an appropriate sized slot has not yet been found,
* check to see if one is available. Also accumulate space
* in the current block so that we can determine if
* compaction is viable.
*/
if (slotstatus != FOUND) {
int size = ep->d_reclen;
if (ep->d_ino != 0)
size -= DIRSIZ(FSFMT(vdp), ep);
if (size > 0) {
if (size >= slotneeded) {
slotstatus = FOUND;
slotoffset = dp->i_offset;
slotsize = ep->d_reclen;
} else if (slotstatus == NONE) {
slotfreespace += size;
if (slotoffset == -1)
slotoffset = dp->i_offset;
if (slotfreespace >= slotneeded) {
slotstatus = COMPACT;
slotsize = dp->i_offset +
ep->d_reclen - slotoffset;
}
}
}
}
/*
* Check for a name match.
*/
if (ep->d_ino) {
# if (BYTE_ORDER == LITTLE_ENDIAN)
if (vdp->v_mount->mnt_maxsymlinklen > 0)
namlen = ep->d_namlen;
else
namlen = ep->d_type;
# else
namlen = ep->d_namlen;
# endif
if (namlen == cnp->cn_namelen &&
!memcmp(cnp->cn_nameptr, ep->d_name, namlen)) {
#ifdef UFS_DIRHASH
foundentry:
#endif
/*
* Save directory entry's inode number and
* reclen in ndp->ni_ufs area, and release
* directory buffer.
*/
dp->i_ino = ep->d_ino;
dp->i_reclen = ep->d_reclen;
goto found;
}
}
prevoff = dp->i_offset;
dp->i_offset += ep->d_reclen;
entryoffsetinblock += ep->d_reclen;
if (ep->d_ino)
enduseful = dp->i_offset;
}
#ifdef UFS_DIRHASH
notfound:
#endif
/*
* If we started in the middle of the directory and failed
* to find our target, we must check the beginning as well.
*/
if (numdirpasses == 2) {
numdirpasses--;
dp->i_offset = 0;
endsearch = dp->i_diroff;
goto searchloop;
}
if (bp != NULL)
brelse(bp);
/*
* If creating, and at end of pathname and current
* directory has not been removed, then can consider
* allowing file to be created.
*/
if ((nameiop == CREATE || nameiop == RENAME) &&
(flags & ISLASTCN) && dp->i_effnlink != 0) {
/*
* Access for write is interpreted as allowing
* creation of files in the directory.
*/
error = VOP_ACCESS(vdp, VWRITE, cred);
if (error)
return (error);
/*
* Return an indication of where the new directory
* entry should be put. If we didn't find a slot,
* then set dp->i_count to 0 indicating
* that the new slot belongs at the end of the
* directory. If we found a slot, then the new entry
* can be put in the range from dp->i_offset to
* dp->i_offset + dp->i_count.
*/
if (slotstatus == NONE) {
dp->i_offset = roundup(DIP(dp, size), DIRBLKSIZ);
dp->i_count = 0;
enduseful = dp->i_offset;
} else if (nameiop == DELETE) {
dp->i_offset = slotoffset;
if ((dp->i_offset & (DIRBLKSIZ - 1)) == 0)
dp->i_count = 0;
else
dp->i_count = dp->i_offset - prevoff;
} else {
dp->i_offset = slotoffset;
dp->i_count = slotsize;
if (enduseful < slotoffset + slotsize)
enduseful = slotoffset + slotsize;
}
dp->i_endoff = roundup(enduseful, DIRBLKSIZ);
/*
* We return with the directory locked, so that
* the parameters we set up above will still be
* valid if we actually decide to do a direnter().
* We return ni_vp == NULL to indicate that the entry
* does not currently exist; we leave a pointer to
* the (locked) directory inode in ndp->ni_dvp.
* The pathname buffer is saved so that the name
* can be obtained later.
*
* NB - if the directory is unlocked, then this
* information cannot be used.
*/
cnp->cn_flags |= SAVENAME;
if (!lockparent) {
VOP_UNLOCK(vdp, 0);
cnp->cn_flags |= PDIRUNLOCK;
}
return (EJUSTRETURN);
}
/*
* Insert name into cache (as non-existent) if appropriate.
*/
if ((cnp->cn_flags & MAKEENTRY) && nameiop != CREATE)
cache_enter(vdp, *vpp, cnp);
return (ENOENT);
found:
if (numdirpasses == 2)
nchstats.ncs_pass2++;
/*
* Check that directory length properly reflects presence
* of this entry.
*/
if (dp->i_offset + DIRSIZ(FSFMT(vdp), ep) > DIP(dp, size)) {
ufs_dirbad(dp, dp->i_offset, "i_ffs_size too small");
DIP_ASSIGN(dp, size, dp->i_offset + DIRSIZ(FSFMT(vdp), ep));
dp->i_flag |= IN_CHANGE | IN_UPDATE;
UFS_WAPBL_UPDATE(dp, MNT_WAIT);
}
brelse(bp);
/*
* Found component in pathname.
* If the final component of path name, save information
* in the cache as to where the entry was found.
*/
if ((flags & ISLASTCN) && nameiop == LOOKUP)
dp->i_diroff = dp->i_offset &~ (DIRBLKSIZ - 1);
/*
* If deleting, and at end of pathname, return
* parameters which can be used to remove file.
* If the wantparent flag isn't set, we return only
* the directory (in ndp->ni_dvp), otherwise we go
* on and lock the inode, being careful with ".".
*/
if (nameiop == DELETE && (flags & ISLASTCN)) {
/*
* Write access to directory required to delete files.
*/
error = VOP_ACCESS(vdp, VWRITE, cred);
if (error)
return (error);
/*
* Return pointer to current entry in dp->i_offset,
* and distance past previous entry (if there
* is a previous entry in this block) in dp->i_count.
* Save directory inode pointer in ndp->ni_dvp for dirremove().
*/
if ((dp->i_offset & (DIRBLKSIZ - 1)) == 0)
dp->i_count = 0;
else
dp->i_count = dp->i_offset - prevoff;
if (dp->i_number == dp->i_ino) {
vref(vdp);
*vpp = vdp;
return (0);
}
error = VFS_VGET(vdp->v_mount, dp->i_ino, &tdp);
if (error)
return (error);
/*
* If directory is "sticky", then user must own
* the directory, or the file in it, else she
* may not delete it (unless she's root). This
* implements append-only directories.
*/
if ((DIP(dp, mode) & ISVTX) &&
cred->cr_uid != 0 &&
cred->cr_uid != DIP(dp, uid) &&
DIP(VTOI(tdp), uid) != cred->cr_uid) {
vput(tdp);
return (EPERM);
}
*vpp = tdp;
if (!lockparent) {
VOP_UNLOCK(vdp, 0);
cnp->cn_flags |= PDIRUNLOCK;
}
return (0);
}
/*
* If rewriting (RENAME), return the inode and the
* information required to rewrite the present directory
* Must get inode of directory entry to verify it's a
* regular file, or empty directory.
*/
if (nameiop == RENAME && wantparent &&
(flags & ISLASTCN)) {
error = VOP_ACCESS(vdp, VWRITE, cred);
if (error)
return (error);
/*
* Careful about locking second inode.
* This can only occur if the target is ".".
*/
if (dp->i_number == dp->i_ino)
return (EISDIR);
error = VFS_VGET(vdp->v_mount, dp->i_ino, &tdp);
if (error)
return (error);
*vpp = tdp;
cnp->cn_flags |= SAVENAME;
if (!lockparent) {
VOP_UNLOCK(vdp, 0);
cnp->cn_flags |= PDIRUNLOCK;
}
return (0);
}
/*
* Step through the translation in the name. We do not `vput' the
* directory because we may need it again if a symbolic link
* is relative to the current directory. Instead we save it
* unlocked as "pdp". We must get the target inode before unlocking
* the directory to insure that the inode will not be removed
* before we get it. We prevent deadlock by always fetching
* inodes from the root, moving down the directory tree. Thus
* when following backward pointers ".." we must unlock the
* parent directory before getting the requested directory.
* There is a potential race condition here if both the current
* and parent directories are removed before the VFS_VGET for the
* inode associated with ".." returns. We hope that this occurs
* infrequently since we cannot avoid this race condition without
* implementing a sophisticated deadlock detection algorithm.
* Note also that this simple deadlock detection scheme will not
* work if the file system has any hard links other than ".."
* that point backwards in the directory structure.
*/
pdp = vdp;
if (flags & ISDOTDOT) {
VOP_UNLOCK(pdp, 0); /* race to get the inode */
cnp->cn_flags |= PDIRUNLOCK;
error = VFS_VGET(vdp->v_mount, dp->i_ino, &tdp);
if (error) {
if (vn_lock(pdp, LK_EXCLUSIVE | LK_RETRY, p) == 0)
cnp->cn_flags &= ~PDIRUNLOCK;
return (error);
}
if (lockparent && (flags & ISLASTCN)) {
if ((error = vn_lock(pdp, LK_EXCLUSIVE, p))) {
vput(tdp);
return (error);
}
cnp->cn_flags &= ~PDIRUNLOCK;
}
*vpp = tdp;
} else if (dp->i_number == dp->i_ino) {
vref(vdp); /* we want ourself, ie "." */
*vpp = vdp;
} else {
error = VFS_VGET(vdp->v_mount, dp->i_ino, &tdp);
if (error)
return (error);
if (!lockparent || !(flags & ISLASTCN)) {
VOP_UNLOCK(pdp, 0);
cnp->cn_flags |= PDIRUNLOCK;
}
*vpp = tdp;
}
/*
* Insert name into cache if appropriate.
*/
if (cnp->cn_flags & MAKEENTRY)
cache_enter(vdp, *vpp, cnp);
return (0);
}
/*
* 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 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;
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);
}
DIRHASH_ASSERT(dirblock < dh->dh_nblk &&
dh->dh_blkfree[dirblock] >= howmany(slotneeded, DIRALIGN),
("ufsdirhash_findfree: bad stats"));
DIRHASH_UNLOCK(dh);
pos = dirblock * DIRBLKSIZ;
error = UFS_BUFATOFF(ip, (off_t)pos, (char **)&dp, &bp);
if (error)
return (-1);
/* Find the first entry with free space. */
for (i = 0; i < DIRBLKSIZ; ) {
if (dp->d_reclen == 0) {
brelse(bp);
return (-1);
}
if (dp->d_ino == 0 || dp->d_reclen > DIRSIZ(0, dp))
break;
i += dp->d_reclen;
dp = (struct direct *)((char *)dp + dp->d_reclen);
}
if (i > DIRBLKSIZ) {
brelse(bp);
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 -= DIRSIZ(0, dp);
if (dp->d_reclen == 0) {
brelse(bp);
return (-1);
}
i += dp->d_reclen;
dp = (struct direct *)((char *)dp + dp->d_reclen);
}
if (i > DIRBLKSIZ) {
brelse(bp);
return (-1);
}
if (freebytes < slotneeded)
panic("ufsdirhash_findfree: free mismatch");
brelse(bp);
*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 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, 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;
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_mtx.
* 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_mtx.
*/
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) {
DIRHASH_ASSERT(dh->dh_onlist, ("dirhash: not on list"));
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 seqopt 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;
DIRHASH_UNLOCK(dh);
if (offset < 0 || offset >= DIP(ip, size))
panic("ufsdirhash_lookup: bad offset in hash array");
if ((offset & ~bmask) != blkoff) {
if (bp != NULL)
brelse(bp);
blkoff = offset & ~bmask;
if (UFS_BUFATOFF(ip, (off_t)blkoff, NULL, &bp) != 0)
return (EJUSTRETURN);
}
dp = (struct direct *)(bp->b_data + (offset & bmask));
if (dp->d_reclen == 0 || dp->d_reclen >
DIRBLKSIZ - (offset & (DIRBLKSIZ - 1))) {
/* Corrupted directory. */
brelse(bp);
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);
if (prevoff == -1) {
brelse(bp);
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 + DIRSIZ(0, dp);
*bpp = bp;
*offp = offset;
return (0);
}
DIRHASH_LOCK(dh);
if (dh->dh_hash == NULL) {
DIRHASH_UNLOCK(dh);
if (bp != NULL)
brelse(bp);
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);
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;
/* Check if we can/should use dirhash. */
if (ip->i_dirhash == NULL) {
if (DIP(ip, size) < ufs_mindirhashsize || OFSFMT(ip))
return (-1);
} else {
/* Hash exists, but sysctls could have changed. */
if (DIP(ip, size) < ufs_mindirhashsize ||
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_effnlink == 0)
return (-1);
vp = ip->i_vnode;
/* Allocate 50% more entries than this dir size could ever need. */
DIRHASH_ASSERT(DIP(ip, size) >= DIRBLKSIZ, ("ufsdirhash_build size"));
nslots = DIP(ip, size) / DIRECTSIZ(1);
nslots = (nslots * 3 + 1) / 2;
narrays = howmany(nslots, DH_NBLKOFF);
nslots = narrays * DH_NBLKOFF;
dirblocks = howmany(DIP(ip, 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);
DIRHASHLIST_LOCK();
if (memreqd + ufs_dirhashmem > ufs_dirhashmaxmem) {
DIRHASHLIST_UNLOCK();
if (memreqd > ufs_dirhashmaxmem / 2)
return (-1);
/* Try to free some space. */
if (ufsdirhash_recycle(memreqd) != 0)
return (-1);
/* Enough was freed, and list has been locked. */
}
ufs_dirhashmem += memreqd;
DIRHASHLIST_UNLOCK();
/*
* Use non-blocking mallocs so that we will revert to a linear
* lookup on failure rather than potentially blocking forever.
*/
dh = malloc(sizeof(*dh), M_DIRHASH, M_NOWAIT|M_ZERO);
if (dh == NULL) {
DIRHASHLIST_LOCK();
ufs_dirhashmem -= memreqd;
DIRHASHLIST_UNLOCK();
return (-1);
}
dh->dh_hash = mallocarray(narrays, sizeof(dh->dh_hash[0]),
M_DIRHASH, M_NOWAIT|M_ZERO);
dh->dh_blkfree = mallocarray(nblocks, sizeof(dh->dh_blkfree[0]),
M_DIRHASH, M_NOWAIT | M_ZERO);
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. */
mtx_init(&dh->dh_mtx, IPL_NONE);
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 < DIP(ip, size)) {
/* If necessary, get the next directory block. */
if ((pos & bmask) == 0) {
if (bp != NULL)
brelse(bp);
if (UFS_BUFATOFF(ip, (off_t)pos, NULL, &bp) != 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);
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, -DIRSIZ(0, ep));
}
pos += ep->d_reclen;
}
if (bp != NULL)
brelse(bp);
DIRHASHLIST_LOCK();
TAILQ_INSERT_TAIL(&ufsdirhash_list, dh, dh_list);
dh->dh_onlist = 1;
DIRHASHLIST_UNLOCK();
return (0);
fail:
if (dh->dh_hash != NULL) {
for (i = 0; i < narrays; i++)
if (dh->dh_hash[i] != NULL)
DIRHASH_BLKFREE(dh->dh_hash[i]);
free(dh->dh_hash, M_DIRHASH, 0);
}
if (dh->dh_blkfree != NULL)
free(dh->dh_blkfree, M_DIRHASH, 0);
free(dh, M_DIRHASH, 0);
ip->i_dirhash = NULL;
DIRHASHLIST_LOCK();
ufs_dirhashmem -= memreqd;
DIRHASHLIST_UNLOCK();
return (-1);
}
