int
tmpfs_reclaim(struct vop_reclaim_args *v)
{
struct vnode *vp = v->a_vp;
struct tmpfs_mount *tmp;
struct tmpfs_node *node;
node = VP_TO_TMPFS_NODE(vp);
tmp = VFS_TO_TMPFS(vp->v_mount);
if (vp->v_type == VREG)
tmpfs_destroy_vobject(vp, node->tn_reg.tn_aobj);
else
vnode_destroy_vobject(vp);
vp->v_object = NULL;
cache_purge(vp);
TMPFS_NODE_LOCK(node);
TMPFS_ASSERT_ELOCKED(node);
tmpfs_free_vp(vp);
/* If the node referenced by this vnode was deleted by the user,
* we must free its associated data structures (now that the vnode
* is being reclaimed). */
if (node->tn_links == 0 &&
(node->tn_vpstate & TMPFS_VNODE_ALLOCATING) == 0) {
node->tn_vpstate = TMPFS_VNODE_DOOMED;
TMPFS_NODE_UNLOCK(node);
tmpfs_free_node(tmp, node);
} else
TMPFS_NODE_UNLOCK(node);
MPASS(vp->v_data == NULL);
return 0;
}
/*
* Frees a directory entry. It is the caller's responsibility to destroy
* the node referenced by it if needed.
*
* The link count of node is decreased by one to reflect the removal of an
* object that referenced it. This only happens if 'node_exists' is true;
* otherwise the function will not access the node referred to by the
* directory entry, as it may already have been released from the outside.
*/
void
tmpfs_free_dirent(struct tmpfs_mount *tmp, struct tmpfs_dirent *de)
{
struct tmpfs_node *node;
node = de->td_node;
TMPFS_NODE_LOCK(node);
TMPFS_ASSERT_ELOCKED(node);
KKASSERT(node->tn_links > 0);
node->tn_links--;
TMPFS_NODE_UNLOCK(node);
kfree(de->td_name, tmp->tm_name_zone);
de->td_namelen = 0;
de->td_name = NULL;
de->td_node = NULL;
objcache_put(tmp->tm_dirent_pool, de);
}
/*
* Change ownership of the given vnode. At least one of uid or gid must
* be different than VNOVAL. If one is set to that value, the attribute
* is unchanged.
* Caller should execute tmpfs_update on vp after a successful execution.
* The vnode must be locked on entry and remain locked on exit.
*/
int
tmpfs_chown(struct vnode *vp, uid_t uid, gid_t gid, struct ucred *cred)
{
mode_t cur_mode;
uid_t cur_uid;
gid_t cur_gid;
struct tmpfs_node *node;
int error;
KKASSERT(vn_islocked(vp));
node = VP_TO_TMPFS_NODE(vp);
/* Disallow this operation if the file system is mounted read-only. */
if (vp->v_mount->mnt_flag & MNT_RDONLY)
return EROFS;
/* Immutable or append-only files cannot be modified, either. */
if (node->tn_flags & (IMMUTABLE | APPEND))
return EPERM;
cur_uid = node->tn_uid;
cur_gid = node->tn_gid;
cur_mode = node->tn_mode;
error = vop_helper_chown(vp, uid, gid, cred,
&cur_uid, &cur_gid, &cur_mode);
if (error == 0) {
TMPFS_NODE_LOCK(node);
if (cur_uid != node->tn_uid ||
cur_gid != node->tn_gid ||
cur_mode != node->tn_mode) {
node->tn_uid = cur_uid;
node->tn_gid = cur_gid;
node->tn_mode = cur_mode;
node->tn_status |= TMPFS_NODE_CHANGED;
}
TMPFS_NODE_UNLOCK(node);
}
return error;
}
static int
tmpfs_readlink(struct vop_readlink_args *v)
{
struct vnode *vp = v->a_vp;
struct uio *uio = v->a_uio;
int error;
struct tmpfs_node *node;
KKASSERT(uio->uio_offset == 0);
KKASSERT(vp->v_type == VLNK);
node = VP_TO_TMPFS_NODE(vp);
error = uiomove(node->tn_link, MIN(node->tn_size, uio->uio_resid),
uio);
TMPFS_NODE_LOCK(node);
node->tn_status |= TMPFS_NODE_ACCESSED;
TMPFS_NODE_UNLOCK(node);
return error;
}
int
tmpfs_getattr(struct vop_getattr_args *v)
{
struct vnode *vp = v->a_vp;
struct vattr *vap = v->a_vap;
struct tmpfs_node *node;
node = VP_TO_TMPFS_NODE(vp);
tmpfs_update(vp);
TMPFS_NODE_LOCK_SH(node);
vap->va_type = vp->v_type;
vap->va_mode = node->tn_mode;
vap->va_nlink = node->tn_links;
vap->va_uid = node->tn_uid;
vap->va_gid = node->tn_gid;
vap->va_fsid = vp->v_mount->mnt_stat.f_fsid.val[0];
vap->va_fileid = node->tn_id;
vap->va_size = node->tn_size;
vap->va_blocksize = PAGE_SIZE;
vap->va_atime.tv_sec = node->tn_atime;
vap->va_atime.tv_nsec = node->tn_atimensec;
vap->va_mtime.tv_sec = node->tn_mtime;
vap->va_mtime.tv_nsec = node->tn_mtimensec;
vap->va_ctime.tv_sec = node->tn_ctime;
vap->va_ctime.tv_nsec = node->tn_ctimensec;
vap->va_gen = node->tn_gen;
vap->va_flags = node->tn_flags;
if (vp->v_type == VBLK || vp->v_type == VCHR) {
vap->va_rmajor = umajor(node->tn_rdev);
vap->va_rminor = uminor(node->tn_rdev);
}
vap->va_bytes = round_page(node->tn_size);
vap->va_filerev = 0;
TMPFS_NODE_UNLOCK(node);
return 0;
}
int
tmpfs_reclaim(struct vop_reclaim_args *v)
{
struct vnode *vp = v->a_vp;
struct tmpfs_mount *tmp;
struct tmpfs_node *node;
struct mount *mp;
mp = vp->v_mount;
lwkt_gettoken(&mp->mnt_token);
node = VP_TO_TMPFS_NODE(vp);
tmp = VFS_TO_TMPFS(vp->v_mount);
KKASSERT(mp == tmp->tm_mount);
tmpfs_free_vp(vp);
/*
* If the node referenced by this vnode was deleted by the
* user, we must free its associated data structures now that
* the vnode is being reclaimed.
*
* Directories have an extra link ref.
*/
TMPFS_NODE_LOCK(node);
if ((node->tn_vpstate & TMPFS_VNODE_ALLOCATING) == 0 &&
node->tn_links == 0) {
node->tn_vpstate = TMPFS_VNODE_DOOMED;
tmpfs_free_node(tmp, node);
/* eats the lock */
} else {
TMPFS_NODE_UNLOCK(node);
}
lwkt_reltoken(&mp->mnt_token);
KKASSERT(vp->v_data == NULL);
return 0;
}
/* Sync timestamps */
void
tmpfs_itimes(struct vnode *vp, const struct timespec *acc,
const struct timespec *mod)
{
struct tmpfs_node *node;
struct timespec now;
node = VP_TO_TMPFS_NODE(vp);
if ((node->tn_status & (TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED |
TMPFS_NODE_CHANGED)) == 0)
return;
vfs_timestamp(&now);
TMPFS_NODE_LOCK(node);
if (node->tn_status & TMPFS_NODE_ACCESSED) {
if (acc == NULL)
acc = &now;
node->tn_atime = acc->tv_sec;
node->tn_atimensec = acc->tv_nsec;
}
if (node->tn_status & TMPFS_NODE_MODIFIED) {
if (mod == NULL)
mod = &now;
node->tn_mtime = mod->tv_sec;
node->tn_mtimensec = mod->tv_nsec;
}
if (node->tn_status & TMPFS_NODE_CHANGED) {
node->tn_ctime = now.tv_sec;
node->tn_ctimensec = now.tv_nsec;
}
node->tn_status &= ~(TMPFS_NODE_ACCESSED |
TMPFS_NODE_MODIFIED |
TMPFS_NODE_CHANGED);
TMPFS_NODE_UNLOCK(node);
}
/*
* Change access mode on the given vnode.
* Caller should execute tmpfs_update on vp after a successful execution.
* The vnode must be locked on entry and remain locked on exit.
*/
int
tmpfs_chmod(struct vnode *vp, mode_t vamode, struct ucred *cred)
{
struct tmpfs_node *node;
mode_t cur_mode;
int error;
KKASSERT(vn_islocked(vp));
node = VP_TO_TMPFS_NODE(vp);
/* Disallow this operation if the file system is mounted read-only. */
if (vp->v_mount->mnt_flag & MNT_RDONLY)
return EROFS;
/* Immutable or append-only files cannot be modified, either. */
if (node->tn_flags & (IMMUTABLE | APPEND))
return EPERM;
cur_mode = node->tn_mode;
error = vop_helper_chmod(vp, vamode, cred, node->tn_uid, node->tn_gid,
&cur_mode);
if (error == 0 &&
(node->tn_mode & ALLPERMS) != (cur_mode & ALLPERMS)) {
TMPFS_NODE_LOCK(node);
node->tn_mode &= ~ALLPERMS;
node->tn_mode |= cur_mode & ALLPERMS;
node->tn_status |= TMPFS_NODE_CHANGED;
TMPFS_NODE_UNLOCK(node);
}
KKASSERT(vn_islocked(vp));
return 0;
}
static int
tmpfs_readdir(struct vop_readdir_args *v)
{
struct vnode *vp = v->a_vp;
struct uio *uio = v->a_uio;
int *eofflag = v->a_eofflag;
off_t **cookies = v->a_cookies;
int *ncookies = v->a_ncookies;
struct tmpfs_mount *tmp;
int error;
off_t startoff;
off_t cnt = 0;
struct tmpfs_node *node;
/* This operation only makes sense on directory nodes. */
if (vp->v_type != VDIR) {
return ENOTDIR;
}
tmp = VFS_TO_TMPFS(vp->v_mount);
node = VP_TO_TMPFS_DIR(vp);
startoff = uio->uio_offset;
if (uio->uio_offset == TMPFS_DIRCOOKIE_DOT) {
error = tmpfs_dir_getdotdent(node, uio);
if (error != 0) {
TMPFS_NODE_LOCK_SH(node);
goto outok;
}
cnt++;
}
if (uio->uio_offset == TMPFS_DIRCOOKIE_DOTDOT) {
/* may lock parent, cannot hold node lock */
error = tmpfs_dir_getdotdotdent(tmp, node, uio);
if (error != 0) {
TMPFS_NODE_LOCK_SH(node);
goto outok;
}
cnt++;
}
TMPFS_NODE_LOCK_SH(node);
error = tmpfs_dir_getdents(node, uio, &cnt);
outok:
KKASSERT(error >= -1);
if (error == -1)
error = 0;
if (eofflag != NULL)
*eofflag =
(error == 0 && uio->uio_offset == TMPFS_DIRCOOKIE_EOF);
/* Update NFS-related variables. */
if (error == 0 && cookies != NULL && ncookies != NULL) {
off_t i;
off_t off = startoff;
struct tmpfs_dirent *de = NULL;
*ncookies = cnt;
*cookies = kmalloc(cnt * sizeof(off_t), M_TEMP, M_WAITOK);
for (i = 0; i < cnt; i++) {
KKASSERT(off != TMPFS_DIRCOOKIE_EOF);
if (off == TMPFS_DIRCOOKIE_DOT) {
off = TMPFS_DIRCOOKIE_DOTDOT;
} else {
if (off == TMPFS_DIRCOOKIE_DOTDOT) {
de = RB_MIN(tmpfs_dirtree_cookie,
&node->tn_dir.tn_cookietree);
} else if (de != NULL) {
de = RB_NEXT(tmpfs_dirtree_cookie,
&node->tn_dir.tn_cookietree, de);
} else {
de = tmpfs_dir_lookupbycookie(node,
off);
KKASSERT(de != NULL);
de = RB_NEXT(tmpfs_dirtree_cookie,
&node->tn_dir.tn_cookietree, de);
}
if (de == NULL)
off = TMPFS_DIRCOOKIE_EOF;
else
off = tmpfs_dircookie(de);
}
(*cookies)[i] = off;
}
KKASSERT(uio->uio_offset == off);
}
TMPFS_NODE_UNLOCK(node);
if ((node->tn_status & TMPFS_NODE_ACCESSED) == 0) {
TMPFS_NODE_LOCK(node);
node->tn_status |= TMPFS_NODE_ACCESSED;
TMPFS_NODE_UNLOCK(node);
}
return error;
}
static int
tmpfs_nrmdir(struct vop_nrmdir_args *v)
{
struct vnode *dvp = v->a_dvp;
struct namecache *ncp = v->a_nch->ncp;
struct vnode *vp;
struct tmpfs_dirent *de;
struct tmpfs_mount *tmp;
struct tmpfs_node *dnode;
struct tmpfs_node *node;
struct mount *mp;
int error;
mp = dvp->v_mount;
/*
* We have to acquire the vp from v->a_nch because we will likely
* unresolve the namecache entry, and a vrele/vput is needed to
* trigger the tmpfs_inactive/tmpfs_reclaim sequence.
*
* We have to use vget to clear any inactive state on the vnode,
* otherwise the vnode may remain inactive and thus tmpfs_inactive
* will not get called when we release it.
*/
error = cache_vget(v->a_nch, v->a_cred, LK_SHARED, &vp);
KKASSERT(error == 0);
vn_unlock(vp);
/*
* Prevalidate so we don't hit an assertion later
*/
if (vp->v_type != VDIR) {
error = ENOTDIR;
goto out;
}
tmp = VFS_TO_TMPFS(dvp->v_mount);
dnode = VP_TO_TMPFS_DIR(dvp);
node = VP_TO_TMPFS_DIR(vp);
/*
* Directories with more than two entries ('.' and '..') cannot
* be removed.
*/
if (node->tn_size > 0) {
error = ENOTEMPTY;
goto out;
}
if ((dnode->tn_flags & APPEND)
|| (node->tn_flags & (NOUNLINK | IMMUTABLE | APPEND))) {
error = EPERM;
goto out;
}
/*
* This invariant holds only if we are not trying to
* remove "..". We checked for that above so this is safe now.
*/
KKASSERT(node->tn_dir.tn_parent == dnode);
/*
* Get the directory entry associated with node (vp). This
* was filled by tmpfs_lookup while looking up the entry.
*/
TMPFS_NODE_LOCK(dnode);
de = tmpfs_dir_lookup(dnode, node, ncp);
KKASSERT(TMPFS_DIRENT_MATCHES(de, ncp->nc_name, ncp->nc_nlen));
/* Check flags to see if we are allowed to remove the directory. */
if ((dnode->tn_flags & APPEND) ||
node->tn_flags & (NOUNLINK | IMMUTABLE | APPEND)) {
error = EPERM;
TMPFS_NODE_UNLOCK(dnode);
goto out;
}
/* Detach the directory entry from the directory (dnode). */
tmpfs_dir_detach(dnode, de);
TMPFS_NODE_UNLOCK(dnode);
/* No vnode should be allocated for this entry from this point */
TMPFS_NODE_LOCK(dnode);
TMPFS_ASSERT_ELOCKED(dnode);
TMPFS_NODE_LOCK(node);
TMPFS_ASSERT_ELOCKED(node);
/*
* Must set parent linkage to NULL (tested by ncreate to disallow
* the creation of new files/dirs in a deleted directory)
*/
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED |
TMPFS_NODE_MODIFIED;
dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED |
TMPFS_NODE_MODIFIED;
TMPFS_NODE_UNLOCK(node);
TMPFS_NODE_UNLOCK(dnode);
/* Free the directory entry we just deleted. Note that the node
* referred by it will not be removed until the vnode is really
* reclaimed. */
tmpfs_free_dirent(tmp, de);
/* Release the deleted vnode (will destroy the node, notify
* interested parties and clean it from the cache). */
TMPFS_NODE_LOCK(dnode);
dnode->tn_status |= TMPFS_NODE_CHANGED;
TMPFS_NODE_UNLOCK(dnode);
tmpfs_update(dvp);
cache_unlink(v->a_nch);
tmpfs_knote(dvp, NOTE_WRITE | NOTE_LINK);
error = 0;
out:
vrele(vp);
return error;
}
static int
tmpfs_nrename(struct vop_nrename_args *v)
{
struct vnode *fdvp = v->a_fdvp;
struct namecache *fncp = v->a_fnch->ncp;
struct vnode *fvp = fncp->nc_vp;
struct vnode *tdvp = v->a_tdvp;
struct namecache *tncp = v->a_tnch->ncp;
struct vnode *tvp;
struct tmpfs_dirent *de, *tde;
struct tmpfs_mount *tmp;
struct tmpfs_node *fdnode;
struct tmpfs_node *fnode;
struct tmpfs_node *tnode;
struct tmpfs_node *tdnode;
struct mount *mp;
char *newname;
char *oldname;
int error;
mp = fdvp->v_mount;
KKASSERT(fdvp->v_mount == fvp->v_mount);
/*
* Because tvp can get overwritten we have to vget it instead of
* just vref or use it, otherwise it's VINACTIVE flag may not get
* cleared and the node won't get destroyed.
*/
error = cache_vget(v->a_tnch, v->a_cred, LK_SHARED, &tvp);
if (error == 0) {
tnode = VP_TO_TMPFS_NODE(tvp);
vn_unlock(tvp);
} else {
tnode = NULL;
}
/* Disallow cross-device renames.
* XXX Why isn't this done by the caller? */
if (fvp->v_mount != tdvp->v_mount ||
(tvp != NULL && fvp->v_mount != tvp->v_mount)) {
error = EXDEV;
goto out;
}
tmp = VFS_TO_TMPFS(tdvp->v_mount);
tdnode = VP_TO_TMPFS_DIR(tdvp);
/* If source and target are the same file, there is nothing to do. */
if (fvp == tvp) {
error = 0;
goto out;
}
fdnode = VP_TO_TMPFS_DIR(fdvp);
fnode = VP_TO_TMPFS_NODE(fvp);
TMPFS_NODE_LOCK(fdnode);
de = tmpfs_dir_lookup(fdnode, fnode, fncp);
TMPFS_NODE_UNLOCK(fdnode); /* XXX depend on namecache lock */
/* Avoid manipulating '.' and '..' entries. */
if (de == NULL) {
error = ENOENT;
goto out_locked;
}
KKASSERT(de->td_node == fnode);
/*
* If replacing an entry in the target directory and that entry
* is a directory, it must be empty.
*
* Kern_rename gurantees the destination to be a directory
* if the source is one (it does?).
*/
if (tvp != NULL) {
KKASSERT(tnode != NULL);
if ((tnode->tn_flags & (NOUNLINK | IMMUTABLE | APPEND)) ||
(tdnode->tn_flags & (APPEND | IMMUTABLE))) {
error = EPERM;
goto out_locked;
}
if (fnode->tn_type == VDIR && tnode->tn_type == VDIR) {
if (tnode->tn_size > 0) {
error = ENOTEMPTY;
goto out_locked;
}
} else if (fnode->tn_type == VDIR && tnode->tn_type != VDIR) {
error = ENOTDIR;
goto out_locked;
} else if (fnode->tn_type != VDIR && tnode->tn_type == VDIR) {
error = EISDIR;
goto out_locked;
} else {
KKASSERT(fnode->tn_type != VDIR &&
tnode->tn_type != VDIR);
}
}
if ((fnode->tn_flags & (NOUNLINK | IMMUTABLE | APPEND)) ||
(fdnode->tn_flags & (APPEND | IMMUTABLE))) {
error = EPERM;
goto out_locked;
}
/*
* Ensure that we have enough memory to hold the new name, if it
* has to be changed.
*/
if (fncp->nc_nlen != tncp->nc_nlen ||
bcmp(fncp->nc_name, tncp->nc_name, fncp->nc_nlen) != 0) {
newname = kmalloc(tncp->nc_nlen + 1, tmp->tm_name_zone,
M_WAITOK | M_NULLOK);
if (newname == NULL) {
error = ENOSPC;
goto out_locked;
}
bcopy(tncp->nc_name, newname, tncp->nc_nlen);
newname[tncp->nc_nlen] = '\0';
} else {
newname = NULL;
}
/*
* Unlink entry from source directory. Note that the kernel has
* already checked for illegal recursion cases (renaming a directory
* into a subdirectory of itself).
*/
if (fdnode != tdnode) {
tmpfs_dir_detach(fdnode, de);
} else {
/* XXX depend on namecache lock */
TMPFS_NODE_LOCK(fdnode);
KKASSERT(de == tmpfs_dir_lookup(fdnode, fnode, fncp));
RB_REMOVE(tmpfs_dirtree, &fdnode->tn_dir.tn_dirtree, de);
RB_REMOVE(tmpfs_dirtree_cookie,
&fdnode->tn_dir.tn_cookietree, de);
TMPFS_NODE_UNLOCK(fdnode);
}
/*
* Handle any name change. Swap with newname, we will
* deallocate it at the end.
*/
if (newname != NULL) {
#if 0
TMPFS_NODE_LOCK(fnode);
fnode->tn_status |= TMPFS_NODE_CHANGED;
TMPFS_NODE_UNLOCK(fnode);
#endif
oldname = de->td_name;
de->td_name = newname;
de->td_namelen = (uint16_t)tncp->nc_nlen;
newname = oldname;
}
/*
* If we are overwriting an entry, we have to remove the old one
* from the target directory.
*/
if (tvp != NULL) {
/* Remove the old entry from the target directory. */
TMPFS_NODE_LOCK(tdnode);
tde = tmpfs_dir_lookup(tdnode, tnode, tncp);
tmpfs_dir_detach(tdnode, tde);
TMPFS_NODE_UNLOCK(tdnode);
tmpfs_knote(tdnode->tn_vnode, NOTE_DELETE);
/*
* Free the directory entry we just deleted. Note that the
* node referred by it will not be removed until the vnode is
* really reclaimed.
*/
tmpfs_free_dirent(VFS_TO_TMPFS(tvp->v_mount), tde);
/*cache_inval_vp(tvp, CINV_DESTROY);*/
}
/*
* Link entry to target directory. If the entry
* represents a directory move the parent linkage
* as well.
*/
if (fdnode != tdnode) {
if (de->td_node->tn_type == VDIR) {
TMPFS_VALIDATE_DIR(fnode);
}
tmpfs_dir_attach(tdnode, de);
} else {
TMPFS_NODE_LOCK(tdnode);
tdnode->tn_status |= TMPFS_NODE_MODIFIED;
RB_INSERT(tmpfs_dirtree, &tdnode->tn_dir.tn_dirtree, de);
RB_INSERT(tmpfs_dirtree_cookie,
&tdnode->tn_dir.tn_cookietree, de);
TMPFS_NODE_UNLOCK(tdnode);
}
/*
* Finish up
*/
if (newname) {
kfree(newname, tmp->tm_name_zone);
newname = NULL;
}
cache_rename(v->a_fnch, v->a_tnch);
tmpfs_knote(v->a_fdvp, NOTE_WRITE);
tmpfs_knote(v->a_tdvp, NOTE_WRITE);
if (fnode->tn_vnode)
tmpfs_knote(fnode->tn_vnode, NOTE_RENAME);
error = 0;
out_locked:
;
out:
if (tvp)
vrele(tvp);
return error;
}
static int
tmpfs_nremove(struct vop_nremove_args *v)
{
struct vnode *dvp = v->a_dvp;
struct namecache *ncp = v->a_nch->ncp;
struct vnode *vp;
int error;
struct tmpfs_dirent *de;
struct tmpfs_mount *tmp;
struct tmpfs_node *dnode;
struct tmpfs_node *node;
struct mount *mp;
mp = dvp->v_mount;
/*
* We have to acquire the vp from v->a_nch because we will likely
* unresolve the namecache entry, and a vrele/vput is needed to
* trigger the tmpfs_inactive/tmpfs_reclaim sequence.
*
* We have to use vget to clear any inactive state on the vnode,
* otherwise the vnode may remain inactive and thus tmpfs_inactive
* will not get called when we release it.
*/
error = cache_vget(v->a_nch, v->a_cred, LK_SHARED, &vp);
KKASSERT(vp->v_mount == dvp->v_mount);
KKASSERT(error == 0);
vn_unlock(vp);
if (vp->v_type == VDIR) {
error = EISDIR;
goto out2;
}
dnode = VP_TO_TMPFS_DIR(dvp);
node = VP_TO_TMPFS_NODE(vp);
tmp = VFS_TO_TMPFS(vp->v_mount);
TMPFS_NODE_LOCK(dnode);
de = tmpfs_dir_lookup(dnode, node, ncp);
if (de == NULL) {
error = ENOENT;
goto out;
}
/* Files marked as immutable or append-only cannot be deleted. */
if ((node->tn_flags & (IMMUTABLE | APPEND | NOUNLINK)) ||
(dnode->tn_flags & APPEND)) {
error = EPERM;
goto out;
}
/* Remove the entry from the directory; as it is a file, we do not
* have to change the number of hard links of the directory. */
tmpfs_dir_detach(dnode, de);
/* Free the directory entry we just deleted. Note that the node
* referred by it will not be removed until the vnode is really
* reclaimed. */
tmpfs_free_dirent(tmp, de);
if (node->tn_links > 0) {
TMPFS_NODE_LOCK(node);
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \
TMPFS_NODE_MODIFIED;
TMPFS_NODE_UNLOCK(node);
}
cache_unlink(v->a_nch);
tmpfs_knote(vp, NOTE_DELETE);
error = 0;
out:
TMPFS_NODE_UNLOCK(dnode);
if (error == 0)
tmpfs_knote(dvp, NOTE_WRITE);
out2:
vrele(vp);
return error;
}
static int
tmpfs_rmdir(struct vop_rmdir_args *v)
{
struct vnode *dvp = v->a_dvp;
struct vnode *vp = v->a_vp;
int error;
struct tmpfs_dirent *de;
struct tmpfs_mount *tmp;
struct tmpfs_node *dnode;
struct tmpfs_node *node;
MPASS(VOP_ISLOCKED(dvp));
MPASS(VOP_ISLOCKED(vp));
tmp = VFS_TO_TMPFS(dvp->v_mount);
dnode = VP_TO_TMPFS_DIR(dvp);
node = VP_TO_TMPFS_DIR(vp);
/* Directories with more than two entries ('.' and '..') cannot be
* removed. */
if (node->tn_size > 0) {
error = ENOTEMPTY;
goto out;
}
if ((dnode->tn_flags & APPEND)
|| (node->tn_flags & (NOUNLINK | IMMUTABLE | APPEND))) {
error = EPERM;
goto out;
}
/* This invariant holds only if we are not trying to remove "..".
* We checked for that above so this is safe now. */
MPASS(node->tn_dir.tn_parent == dnode);
/* Get the directory entry associated with node (vp). This was
* filled by tmpfs_lookup while looking up the entry. */
de = tmpfs_dir_lookup(dnode, node, v->a_cnp);
MPASS(TMPFS_DIRENT_MATCHES(de,
v->a_cnp->cn_nameptr,
v->a_cnp->cn_namelen));
/* Check flags to see if we are allowed to remove the directory. */
if (dnode->tn_flags & APPEND
|| node->tn_flags & (NOUNLINK | IMMUTABLE | APPEND)) {
error = EPERM;
goto out;
}
/* Detach the directory entry from the directory (dnode). */
tmpfs_dir_detach(dvp, de);
if (v->a_cnp->cn_flags & DOWHITEOUT)
tmpfs_dir_whiteout_add(dvp, v->a_cnp);
/* No vnode should be allocated for this entry from this point */
TMPFS_NODE_LOCK(node);
TMPFS_ASSERT_ELOCKED(node);
node->tn_links--;
node->tn_dir.tn_parent = NULL;
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \
TMPFS_NODE_MODIFIED;
TMPFS_NODE_UNLOCK(node);
TMPFS_NODE_LOCK(dnode);
TMPFS_ASSERT_ELOCKED(dnode);
dnode->tn_links--;
dnode->tn_status |= TMPFS_NODE_ACCESSED | \
TMPFS_NODE_CHANGED | TMPFS_NODE_MODIFIED;
TMPFS_NODE_UNLOCK(dnode);
cache_purge(dvp);
cache_purge(vp);
/* Free the directory entry we just deleted. Note that the node
* referred by it will not be removed until the vnode is really
* reclaimed. */
tmpfs_free_dirent(tmp, de, TRUE);
/* Release the deleted vnode (will destroy the node, notify
* interested parties and clean it from the cache). */
dnode->tn_status |= TMPFS_NODE_CHANGED;
tmpfs_update(dvp);
error = 0;
out:
return error;
}
static int
tmpfs_rename(struct vop_rename_args *v)
{
struct vnode *fdvp = v->a_fdvp;
struct vnode *fvp = v->a_fvp;
struct componentname *fcnp = v->a_fcnp;
struct vnode *tdvp = v->a_tdvp;
struct vnode *tvp = v->a_tvp;
struct componentname *tcnp = v->a_tcnp;
char *newname;
int error;
struct tmpfs_dirent *de;
struct tmpfs_mount *tmp;
struct tmpfs_node *fdnode;
struct tmpfs_node *fnode;
struct tmpfs_node *tnode;
struct tmpfs_node *tdnode;
MPASS(VOP_ISLOCKED(tdvp));
MPASS(IMPLIES(tvp != NULL, VOP_ISLOCKED(tvp)));
MPASS(fcnp->cn_flags & HASBUF);
MPASS(tcnp->cn_flags & HASBUF);
tnode = (tvp == NULL) ? NULL : VP_TO_TMPFS_NODE(tvp);
/* Disallow cross-device renames.
* XXX Why isn't this done by the caller? */
if (fvp->v_mount != tdvp->v_mount ||
(tvp != NULL && fvp->v_mount != tvp->v_mount)) {
error = EXDEV;
goto out;
}
tmp = VFS_TO_TMPFS(tdvp->v_mount);
tdnode = VP_TO_TMPFS_DIR(tdvp);
/* If source and target are the same file, there is nothing to do. */
if (fvp == tvp) {
error = 0;
goto out;
}
/* If we need to move the directory between entries, lock the
* source so that we can safely operate on it. */
if (fdvp != tdvp && fdvp != tvp)
vn_lock(fdvp, LK_EXCLUSIVE | LK_RETRY);
fdnode = VP_TO_TMPFS_DIR(fdvp);
fnode = VP_TO_TMPFS_NODE(fvp);
de = tmpfs_dir_lookup(fdnode, fnode, fcnp);
/* Entry can disappear before we lock fdvp,
* also avoid manipulating '.' and '..' entries. */
if (de == NULL) {
if ((fcnp->cn_flags & ISDOTDOT) != 0 ||
(fcnp->cn_namelen == 1 && fcnp->cn_nameptr[0] == '.'))
error = EINVAL;
else
error = ENOENT;
goto out_locked;
}
MPASS(de->td_node == fnode);
/* If re-naming a directory to another preexisting directory
* ensure that the target directory is empty so that its
* removal causes no side effects.
* Kern_rename gurantees the destination to be a directory
* if the source is one. */
if (tvp != NULL) {
MPASS(tnode != NULL);
if ((tnode->tn_flags & (NOUNLINK | IMMUTABLE | APPEND)) ||
(tdnode->tn_flags & (APPEND | IMMUTABLE))) {
error = EPERM;
goto out_locked;
}
if (fnode->tn_type == VDIR && tnode->tn_type == VDIR) {
if (tnode->tn_size > 0) {
error = ENOTEMPTY;
goto out_locked;
}
} else if (fnode->tn_type == VDIR && tnode->tn_type != VDIR) {
error = ENOTDIR;
goto out_locked;
} else if (fnode->tn_type != VDIR && tnode->tn_type == VDIR) {
error = EISDIR;
goto out_locked;
} else {
MPASS(fnode->tn_type != VDIR &&
tnode->tn_type != VDIR);
}
}
if ((fnode->tn_flags & (NOUNLINK | IMMUTABLE | APPEND))
|| (fdnode->tn_flags & (APPEND | IMMUTABLE))) {
error = EPERM;
goto out_locked;
}
/* Ensure that we have enough memory to hold the new name, if it
* has to be changed. */
if (fcnp->cn_namelen != tcnp->cn_namelen ||
bcmp(fcnp->cn_nameptr, tcnp->cn_nameptr, fcnp->cn_namelen) != 0) {
newname = malloc(tcnp->cn_namelen, M_TMPFSNAME, M_WAITOK);
} else
newname = NULL;
/* If the node is being moved to another directory, we have to do
* the move. */
if (fdnode != tdnode) {
/* In case we are moving a directory, we have to adjust its
* parent to point to the new parent. */
if (de->td_node->tn_type == VDIR) {
struct tmpfs_node *n;
/* Ensure the target directory is not a child of the
* directory being moved. Otherwise, we'd end up
* with stale nodes. */
n = tdnode;
/* TMPFS_LOCK garanties that no nodes are freed while
* traversing the list. Nodes can only be marked as
* removed: tn_parent == NULL. */
TMPFS_LOCK(tmp);
TMPFS_NODE_LOCK(n);
while (n != n->tn_dir.tn_parent) {
struct tmpfs_node *parent;
if (n == fnode) {
TMPFS_NODE_UNLOCK(n);
TMPFS_UNLOCK(tmp);
error = EINVAL;
if (newname != NULL)
free(newname, M_TMPFSNAME);
goto out_locked;
}
parent = n->tn_dir.tn_parent;
TMPFS_NODE_UNLOCK(n);
if (parent == NULL) {
n = NULL;
break;
}
TMPFS_NODE_LOCK(parent);
if (parent->tn_dir.tn_parent == NULL) {
TMPFS_NODE_UNLOCK(parent);
n = NULL;
break;
}
n = parent;
}
TMPFS_UNLOCK(tmp);
if (n == NULL) {
error = EINVAL;
if (newname != NULL)
free(newname, M_TMPFSNAME);
goto out_locked;
}
TMPFS_NODE_UNLOCK(n);
/* Adjust the parent pointer. */
TMPFS_VALIDATE_DIR(fnode);
TMPFS_NODE_LOCK(de->td_node);
de->td_node->tn_dir.tn_parent = tdnode;
TMPFS_NODE_UNLOCK(de->td_node);
/* As a result of changing the target of the '..'
* entry, the link count of the source and target
* directories has to be adjusted. */
TMPFS_NODE_LOCK(tdnode);
TMPFS_ASSERT_LOCKED(tdnode);
tdnode->tn_links++;
TMPFS_NODE_UNLOCK(tdnode);
TMPFS_NODE_LOCK(fdnode);
TMPFS_ASSERT_LOCKED(fdnode);
fdnode->tn_links--;
TMPFS_NODE_UNLOCK(fdnode);
}
/* Do the move: just remove the entry from the source directory
* and insert it into the target one. */
tmpfs_dir_detach(fdvp, de);
if (fcnp->cn_flags & DOWHITEOUT)
tmpfs_dir_whiteout_add(fdvp, fcnp);
if (tcnp->cn_flags & ISWHITEOUT)
tmpfs_dir_whiteout_remove(tdvp, tcnp);
tmpfs_dir_attach(tdvp, de);
}
/* If the name has changed, we need to make it effective by changing
* it in the directory entry. */
if (newname != NULL) {
MPASS(tcnp->cn_namelen <= MAXNAMLEN);
free(de->td_name, M_TMPFSNAME);
de->td_namelen = (uint16_t)tcnp->cn_namelen;
memcpy(newname, tcnp->cn_nameptr, tcnp->cn_namelen);
de->td_name = newname;
fnode->tn_status |= TMPFS_NODE_CHANGED;
tdnode->tn_status |= TMPFS_NODE_MODIFIED;
}
/* If we are overwriting an entry, we have to remove the old one
* from the target directory. */
if (tvp != NULL) {
/* Remove the old entry from the target directory. */
de = tmpfs_dir_lookup(tdnode, tnode, tcnp);
tmpfs_dir_detach(tdvp, de);
/* Free the directory entry we just deleted. Note that the
* node referred by it will not be removed until the vnode is
* really reclaimed. */
tmpfs_free_dirent(VFS_TO_TMPFS(tvp->v_mount), de, TRUE);
}
cache_purge(fvp);
error = 0;
out_locked:
if (fdvp != tdvp && fdvp != tvp)
VOP_UNLOCK(fdvp, 0);
out:
/* Release target nodes. */
/* XXX: I don't understand when tdvp can be the same as tvp, but
* other code takes care of this... */
if (tdvp == tvp)
vrele(tdvp);
else
vput(tdvp);
if (tvp != NULL)
vput(tvp);
/* Release source nodes. */
vrele(fdvp);
vrele(fvp);
return error;
}
/*
* Allocates a new node of type 'type' inside the 'tmp' mount point, with
* its owner set to 'uid', its group to 'gid' and its mode set to 'mode',
* using the credentials of the process 'p'.
*
* If the node type is set to 'VDIR', then the parent parameter must point
* to the parent directory of the node being created. It may only be NULL
* while allocating the root node.
*
* If the node type is set to 'VBLK' or 'VCHR', then the rdev parameter
* specifies the device the node represents.
*
* If the node type is set to 'VLNK', then the parameter target specifies
* the file name of the target file for the symbolic link that is being
* created.
*
* Note that new nodes are retrieved from the available list if it has
* items or, if it is empty, from the node pool as long as there is enough
* space to create them.
*
* Returns zero on success or an appropriate error code on failure.
*/
int
tmpfs_alloc_node(struct tmpfs_mount *tmp, enum vtype type,
uid_t uid, gid_t gid, mode_t mode,
char *target, int rmajor, int rminor,
struct tmpfs_node **node)
{
struct tmpfs_node *nnode;
struct timespec ts;
udev_t rdev;
KKASSERT(IFF(type == VLNK, target != NULL));
KKASSERT(IFF(type == VBLK || type == VCHR, rmajor != VNOVAL));
if (tmp->tm_nodes_inuse >= tmp->tm_nodes_max)
return (ENOSPC);
nnode = objcache_get(tmp->tm_node_pool, M_WAITOK | M_NULLOK);
if (nnode == NULL)
return (ENOSPC);
/* Generic initialization. */
nnode->tn_type = type;
vfs_timestamp(&ts);
nnode->tn_ctime = nnode->tn_mtime = nnode->tn_atime
= ts.tv_sec;
nnode->tn_ctimensec = nnode->tn_mtimensec = nnode->tn_atimensec
= ts.tv_nsec;
nnode->tn_uid = uid;
nnode->tn_gid = gid;
nnode->tn_mode = mode;
nnode->tn_id = tmpfs_fetch_ino(tmp);
nnode->tn_advlock.init_done = 0;
KKASSERT(nnode->tn_links == 0);
/* Type-specific initialization. */
switch (nnode->tn_type) {
case VBLK:
case VCHR:
rdev = makeudev(rmajor, rminor);
if (rdev == NOUDEV) {
objcache_put(tmp->tm_node_pool, nnode);
return(EINVAL);
}
nnode->tn_rdev = rdev;
break;
case VDIR:
RB_INIT(&nnode->tn_dir.tn_dirtree);
RB_INIT(&nnode->tn_dir.tn_cookietree);
nnode->tn_size = 0;
break;
case VFIFO:
/* FALLTHROUGH */
case VSOCK:
break;
case VLNK:
nnode->tn_size = strlen(target);
nnode->tn_link = kmalloc(nnode->tn_size + 1, tmp->tm_name_zone,
M_WAITOK | M_NULLOK);
if (nnode->tn_link == NULL) {
objcache_put(tmp->tm_node_pool, nnode);
return (ENOSPC);
}
bcopy(target, nnode->tn_link, nnode->tn_size);
nnode->tn_link[nnode->tn_size] = '\0';
break;
case VREG:
nnode->tn_reg.tn_aobj = swap_pager_alloc(NULL, 0,
VM_PROT_DEFAULT, 0);
nnode->tn_reg.tn_aobj_pages = 0;
nnode->tn_size = 0;
vm_object_set_flag(nnode->tn_reg.tn_aobj, OBJ_NOPAGEIN);
break;
default:
panic("tmpfs_alloc_node: type %p %d", nnode, (int)nnode->tn_type);
}
TMPFS_NODE_LOCK(nnode);
TMPFS_LOCK(tmp);
LIST_INSERT_HEAD(&tmp->tm_nodes_used, nnode, tn_entries);
tmp->tm_nodes_inuse++;
TMPFS_UNLOCK(tmp);
TMPFS_NODE_UNLOCK(nnode);
*node = nnode;
return 0;
}
/*
* Allocates a new file of type 'type' and adds it to the parent directory
* 'dvp'; this addition is done using the component name given in 'cnp'.
* The ownership of the new file is automatically assigned based on the
* credentials of the caller (through 'cnp'), the group is set based on
* the parent directory and the mode is determined from the 'vap' argument.
* If successful, *vpp holds a vnode to the newly created file and zero
* is returned. Otherwise *vpp is NULL and the function returns an
* appropriate error code.
*/
int
tmpfs_alloc_file(struct vnode *dvp, struct vnode **vpp, struct vattr *vap,
struct namecache *ncp, struct ucred *cred, char *target)
{
int error;
struct tmpfs_dirent *de;
struct tmpfs_mount *tmp;
struct tmpfs_node *dnode;
struct tmpfs_node *node;
tmp = VFS_TO_TMPFS(dvp->v_mount);
dnode = VP_TO_TMPFS_DIR(dvp);
*vpp = NULL;
/*
* If the directory was removed but a process was CD'd into it,
* we do not allow any more file/dir creation within it. Otherwise
* we will lose track of it.
*/
KKASSERT(dnode->tn_type == VDIR);
if (dnode != tmp->tm_root && dnode->tn_dir.tn_parent == NULL)
return ENOENT;
/*
* Make sure the link count does not overflow.
*/
if (vap->va_type == VDIR && dnode->tn_links >= LINK_MAX)
return EMLINK;
/* Allocate a node that represents the new file. */
error = tmpfs_alloc_node(tmp, vap->va_type, cred->cr_uid,
dnode->tn_gid, vap->va_mode, target,
vap->va_rmajor, vap->va_rminor, &node);
if (error != 0)
return error;
TMPFS_NODE_LOCK(node);
/* Allocate a directory entry that points to the new file. */
error = tmpfs_alloc_dirent(tmp, node, ncp->nc_name, ncp->nc_nlen, &de);
if (error != 0) {
tmpfs_free_node(tmp, node);
/* eats node lock */
return error;
}
/* Allocate a vnode for the new file. */
error = tmpfs_alloc_vp(dvp->v_mount, NULL, node, LK_EXCLUSIVE, vpp);
if (error != 0) {
tmpfs_free_dirent(tmp, de);
tmpfs_free_node(tmp, node);
/* eats node lock */
return error;
}
/*
* Now that all required items are allocated, we can proceed to
* insert the new node into the directory, an operation that
* cannot fail.
*/
tmpfs_dir_attach(dnode, de);
TMPFS_NODE_UNLOCK(node);
return error;
}
/*
* Allocates a new vnode for the node node or returns a new reference to
* an existing one if the node had already a vnode referencing it. The
* resulting locked vnode is returned in *vpp.
*
* Returns zero on success or an appropriate error code on failure.
*
* The caller must ensure that node cannot go away (usually by holding
* the related directory entry).
*
* If dnode is non-NULL this routine avoids deadlocking against it but
* can return EAGAIN. Caller must try again. The dnode lock will cycle
* in this case, it remains locked on return in all cases. dnode must
* be shared-locked.
*/
int
tmpfs_alloc_vp(struct mount *mp,
struct tmpfs_node *dnode, struct tmpfs_node *node, int lkflag,
struct vnode **vpp)
{
int error = 0;
struct vnode *vp;
loop:
/*
* Interlocked extraction from node. This can race many things.
* We have to get a soft reference on the vnode while we hold
* the node locked, then acquire it properly and check for races.
*/
TMPFS_NODE_LOCK(node);
if ((vp = node->tn_vnode) != NULL) {
KKASSERT((node->tn_vpstate & TMPFS_VNODE_DOOMED) == 0);
vhold(vp);
TMPFS_NODE_UNLOCK(node);
if (dnode) {
/*
* Special-case handling to avoid deadlocking against
* dnode. This case has been validated and occurs
* every so often during synth builds.
*/
if (vget(vp, (lkflag & ~LK_RETRY) |
LK_NOWAIT |
LK_EXCLUSIVE) != 0) {
TMPFS_NODE_UNLOCK(dnode);
if (vget(vp, (lkflag & ~LK_RETRY) |
LK_SLEEPFAIL |
LK_EXCLUSIVE) == 0) {
vn_unlock(vp);
}
vdrop(vp);
TMPFS_NODE_LOCK_SH(dnode);
return EAGAIN;
}
} else {
/*
* Normal path
*/
if (vget(vp, lkflag | LK_EXCLUSIVE) != 0) {
vdrop(vp);
goto loop;
}
}
if (node->tn_vnode != vp) {
vput(vp);
vdrop(vp);
goto loop;
}
vdrop(vp);
goto out;
}
/* vp is NULL */
/*
* This should never happen.
*/
if (node->tn_vpstate & TMPFS_VNODE_DOOMED) {
TMPFS_NODE_UNLOCK(node);
error = ENOENT;
goto out;
}
/*
* Interlock against other calls to tmpfs_alloc_vp() trying to
* allocate and assign a vp to node.
*/
if (node->tn_vpstate & TMPFS_VNODE_ALLOCATING) {
node->tn_vpstate |= TMPFS_VNODE_WANT;
error = tsleep(&node->tn_vpstate, PINTERLOCKED | PCATCH,
"tmpfs_alloc_vp", 0);
TMPFS_NODE_UNLOCK(node);
if (error)
return error;
goto loop;
}
node->tn_vpstate |= TMPFS_VNODE_ALLOCATING;
TMPFS_NODE_UNLOCK(node);
/*
* Allocate a new vnode (may block). The ALLOCATING flag should
* prevent a race against someone else assigning node->tn_vnode.
*/
error = getnewvnode(VT_TMPFS, mp, &vp, VLKTIMEOUT, LK_CANRECURSE);
if (error != 0)
goto unlock;
KKASSERT(node->tn_vnode == NULL);
KKASSERT(vp != NULL);
vp->v_data = node;
vp->v_type = node->tn_type;
/* Type-specific initialization. */
switch (node->tn_type) {
case VBLK:
/* FALLTHROUGH */
case VCHR:
/* FALLTHROUGH */
case VSOCK:
break;
case VREG:
/*
* VMIO is mandatory. Tmpfs also supports KVABIO
* for its tmpfs_strategy().
*/
vsetflags(vp, VKVABIO);
vinitvmio(vp, node->tn_size, TMPFS_BLKSIZE, -1);
break;
case VLNK:
break;
case VFIFO:
vp->v_ops = &mp->mnt_vn_fifo_ops;
break;
case VDIR:
break;
default:
panic("tmpfs_alloc_vp: type %p %d", node, (int)node->tn_type);
}
unlock:
TMPFS_NODE_LOCK(node);
KKASSERT(node->tn_vpstate & TMPFS_VNODE_ALLOCATING);
node->tn_vpstate &= ~TMPFS_VNODE_ALLOCATING;
node->tn_vnode = vp;
if (node->tn_vpstate & TMPFS_VNODE_WANT) {
node->tn_vpstate &= ~TMPFS_VNODE_WANT;
TMPFS_NODE_UNLOCK(node);
wakeup(&node->tn_vpstate);
} else {
TMPFS_NODE_UNLOCK(node);
}
out:
*vpp = vp;
KKASSERT(IFF(error == 0, *vpp != NULL && vn_islocked(*vpp)));
return error;
}
static int
tmpfs_read (struct vop_read_args *ap)
{
struct buf *bp;
struct vnode *vp = ap->a_vp;
struct uio *uio = ap->a_uio;
struct tmpfs_node *node;
off_t base_offset;
size_t offset;
size_t len;
size_t resid;
int error;
/*
* Check the basics
*/
if (uio->uio_offset < 0)
return (EINVAL);
if (vp->v_type != VREG)
return (EINVAL);
/*
* Extract node, try to shortcut the operation through
* the VM page cache, allowing us to avoid buffer cache
* overheads.
*/
node = VP_TO_TMPFS_NODE(vp);
resid = uio->uio_resid;
error = vop_helper_read_shortcut(ap);
if (error)
return error;
if (uio->uio_resid == 0) {
if (resid)
goto finished;
return error;
}
/*
* Fall-through to our normal read code.
*/
while (uio->uio_resid > 0 && uio->uio_offset < node->tn_size) {
/*
* Use buffer cache I/O (via tmpfs_strategy)
*/
offset = (size_t)uio->uio_offset & TMPFS_BLKMASK64;
base_offset = (off_t)uio->uio_offset - offset;
bp = getcacheblk(vp, base_offset, TMPFS_BLKSIZE, 0);
if (bp == NULL) {
error = bread(vp, base_offset, TMPFS_BLKSIZE, &bp);
if (error) {
brelse(bp);
kprintf("tmpfs_read bread error %d\n", error);
break;
}
/*
* tmpfs pretty much fiddles directly with the VM
* system, don't let it exhaust it or we won't play
* nice with other processes.
*
* Only do this if the VOP is coming from a normal
* read/write. The VM system handles the case for
* UIO_NOCOPY.
*/
if (uio->uio_segflg != UIO_NOCOPY)
vm_wait_nominal();
}
bp->b_flags |= B_CLUSTEROK;
/*
* Figure out how many bytes we can actually copy this loop.
*/
len = TMPFS_BLKSIZE - offset;
if (len > uio->uio_resid)
len = uio->uio_resid;
if (len > node->tn_size - uio->uio_offset)
len = (size_t)(node->tn_size - uio->uio_offset);
error = uiomovebp(bp, (char *)bp->b_data + offset, len, uio);
bqrelse(bp);
if (error) {
kprintf("tmpfs_read uiomove error %d\n", error);
break;
}
}
finished:
if ((node->tn_status & TMPFS_NODE_ACCESSED) == 0) {
TMPFS_NODE_LOCK(node);
node->tn_status |= TMPFS_NODE_ACCESSED;
TMPFS_NODE_UNLOCK(node);
}
return (error);
}
static int
tmpfs_read (struct vop_read_args *ap)
{
struct buf *bp;
struct vnode *vp = ap->a_vp;
struct uio *uio = ap->a_uio;
struct tmpfs_node *node;
off_t base_offset;
size_t offset;
size_t len;
int error;
error = 0;
if (uio->uio_resid == 0) {
return error;
}
node = VP_TO_TMPFS_NODE(vp);
if (uio->uio_offset < 0)
return (EINVAL);
if (vp->v_type != VREG)
return (EINVAL);
while (uio->uio_resid > 0 && uio->uio_offset < node->tn_size) {
/*
* Use buffer cache I/O (via tmpfs_strategy)
*/
offset = (size_t)uio->uio_offset & BMASK;
base_offset = (off_t)uio->uio_offset - offset;
bp = getcacheblk(vp, base_offset, BSIZE, 0);
if (bp == NULL) {
lwkt_gettoken(&vp->v_mount->mnt_token);
error = bread(vp, base_offset, BSIZE, &bp);
if (error) {
brelse(bp);
lwkt_reltoken(&vp->v_mount->mnt_token);
kprintf("tmpfs_read bread error %d\n", error);
break;
}
lwkt_reltoken(&vp->v_mount->mnt_token);
}
/*
* Figure out how many bytes we can actually copy this loop.
*/
len = BSIZE - offset;
if (len > uio->uio_resid)
len = uio->uio_resid;
if (len > node->tn_size - uio->uio_offset)
len = (size_t)(node->tn_size - uio->uio_offset);
error = uiomovebp(bp, (char *)bp->b_data + offset, len, uio);
bqrelse(bp);
if (error) {
kprintf("tmpfs_read uiomove error %d\n", error);
break;
}
}
TMPFS_NODE_LOCK(node);
node->tn_status |= TMPFS_NODE_ACCESSED;
TMPFS_NODE_UNLOCK(node);
return(error);
}
static int
tmpfs_write (struct vop_write_args *ap)
{
struct buf *bp;
struct vnode *vp = ap->a_vp;
struct uio *uio = ap->a_uio;
struct thread *td = uio->uio_td;
struct tmpfs_node *node;
boolean_t extended;
off_t oldsize;
int error;
off_t base_offset;
size_t offset;
size_t len;
struct rlimit limit;
int trivial = 0;
int kflags = 0;
int seqcount;
error = 0;
if (uio->uio_resid == 0) {
return error;
}
node = VP_TO_TMPFS_NODE(vp);
if (vp->v_type != VREG)
return (EINVAL);
seqcount = ap->a_ioflag >> 16;
TMPFS_NODE_LOCK(node);
oldsize = node->tn_size;
if (ap->a_ioflag & IO_APPEND)
uio->uio_offset = node->tn_size;
/*
* Check for illegal write offsets.
*/
if (uio->uio_offset + uio->uio_resid >
VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize) {
error = EFBIG;
goto done;
}
/*
* NOTE: Ignore if UIO does not come from a user thread (e.g. VN).
*/
if (vp->v_type == VREG && td != NULL && td->td_lwp != NULL) {
error = kern_getrlimit(RLIMIT_FSIZE, &limit);
if (error)
goto done;
if (uio->uio_offset + uio->uio_resid > limit.rlim_cur) {
ksignal(td->td_proc, SIGXFSZ);
error = EFBIG;
goto done;
}
}
/*
* Extend the file's size if necessary
*/
extended = ((uio->uio_offset + uio->uio_resid) > node->tn_size);
while (uio->uio_resid > 0) {
/*
* Don't completely blow out running buffer I/O
* when being hit from the pageout daemon.
*/
if (uio->uio_segflg == UIO_NOCOPY &&
(ap->a_ioflag & IO_RECURSE) == 0) {
bwillwrite(TMPFS_BLKSIZE);
}
/*
* Use buffer cache I/O (via tmpfs_strategy)
*/
offset = (size_t)uio->uio_offset & TMPFS_BLKMASK64;
base_offset = (off_t)uio->uio_offset - offset;
len = TMPFS_BLKSIZE - offset;
if (len > uio->uio_resid)
len = uio->uio_resid;
if ((uio->uio_offset + len) > node->tn_size) {
trivial = (uio->uio_offset <= node->tn_size);
error = tmpfs_reg_resize(vp, uio->uio_offset + len,
trivial);
if (error)
break;
}
/*
* Read to fill in any gaps. Theoretically we could
* optimize this if the write covers the entire buffer
* and is not a UIO_NOCOPY write, however this can lead
* to a security violation exposing random kernel memory
* (whatever junk was in the backing VM pages before).
*
* So just use bread() to do the right thing.
*/
error = bread(vp, base_offset, TMPFS_BLKSIZE, &bp);
error = uiomovebp(bp, (char *)bp->b_data + offset, len, uio);
if (error) {
kprintf("tmpfs_write uiomove error %d\n", error);
brelse(bp);
break;
}
if (uio->uio_offset > node->tn_size) {
node->tn_size = uio->uio_offset;
kflags |= NOTE_EXTEND;
}
kflags |= NOTE_WRITE;
/*
* Always try to flush the page in the UIO_NOCOPY case. This
* can come from the pageout daemon or during vnode eviction.
* It is not necessarily going to be marked IO_ASYNC/IO_SYNC.
*
* For the normal case we buwrite(), dirtying the underlying
* VM pages instead of dirtying the buffer and releasing the
* buffer as a clean buffer. This allows tmpfs to use
* essentially all available memory to cache file data.
* If we used bdwrite() the buffer cache would wind up
* flushing the data to swap too quickly.
*
* But because tmpfs can seriously load the VM system we
* fall-back to using bdwrite() when free memory starts
* to get low. This shifts the load away from the VM system
* and makes tmpfs act more like a normal filesystem with
* regards to disk activity.
*
* tmpfs pretty much fiddles directly with the VM
* system, don't let it exhaust it or we won't play
* nice with other processes. Only do this if the
* VOP is coming from a normal read/write. The VM system
* handles the case for UIO_NOCOPY.
*/
bp->b_flags |= B_CLUSTEROK;
if (uio->uio_segflg == UIO_NOCOPY) {
/*
* Flush from the pageout daemon, deal with
* potentially very heavy tmpfs write activity
* causing long stalls in the pageout daemon
* before pages get to free/cache.
*
* (a) Under severe pressure setting B_DIRECT will
* cause a buffer release to try to free the
* underlying pages.
*
* (b) Under modest memory pressure the B_RELBUF
* alone is sufficient to get the pages moved
* to the cache. We could also force this by
* setting B_NOTMETA but that might have other
* unintended side-effects (e.g. setting
* PG_NOTMETA on the VM page).
*
* Hopefully this will unblock the VM system more
* quickly under extreme tmpfs write load.
*/
if (vm_page_count_min(vm_page_free_hysteresis))
bp->b_flags |= B_DIRECT;
bp->b_flags |= B_AGE | B_RELBUF;
bp->b_act_count = 0; /* buffer->deactivate pgs */
cluster_awrite(bp);
} else if (vm_page_count_target()) {
/*
* Normal (userland) write but we are low on memory,
* run the buffer the buffer cache.
*/
bp->b_act_count = 0; /* buffer->deactivate pgs */
bdwrite(bp);
} else {
/*
* Otherwise run the buffer directly through to the
* backing VM store.
*/
buwrite(bp);
/*vm_wait_nominal();*/
}
if (bp->b_error) {
kprintf("tmpfs_write bwrite error %d\n", bp->b_error);
break;
}
}
if (error) {
if (extended) {
(void)tmpfs_reg_resize(vp, oldsize, trivial);
kflags &= ~NOTE_EXTEND;
}
goto done;
}
/*
* Currently we don't set the mtime on files modified via mmap()
* because we can't tell the difference between those modifications
* and an attempt by the pageout daemon to flush tmpfs pages to
* swap.
*
* This is because in order to defer flushes as long as possible
* buwrite() works by marking the underlying VM pages dirty in
* order to be able to dispose of the buffer cache buffer without
* flushing it.
*/
if (uio->uio_segflg != UIO_NOCOPY)
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED;
if (extended)
node->tn_status |= TMPFS_NODE_CHANGED;
if (node->tn_mode & (S_ISUID | S_ISGID)) {
if (priv_check_cred(ap->a_cred, PRIV_VFS_RETAINSUGID, 0))
node->tn_mode &= ~(S_ISUID | S_ISGID);
}
done:
TMPFS_NODE_UNLOCK(node);
if (kflags)
tmpfs_knote(vp, kflags);
return(error);
}
static int
tmpfs_write (struct vop_write_args *ap)
{
struct buf *bp;
struct vnode *vp = ap->a_vp;
struct uio *uio = ap->a_uio;
struct thread *td = uio->uio_td;
struct tmpfs_node *node;
boolean_t extended;
off_t oldsize;
int error;
off_t base_offset;
size_t offset;
size_t len;
struct rlimit limit;
int trivial = 0;
int kflags = 0;
error = 0;
if (uio->uio_resid == 0) {
return error;
}
node = VP_TO_TMPFS_NODE(vp);
if (vp->v_type != VREG)
return (EINVAL);
lwkt_gettoken(&vp->v_mount->mnt_token);
oldsize = node->tn_size;
if (ap->a_ioflag & IO_APPEND)
uio->uio_offset = node->tn_size;
/*
* Check for illegal write offsets.
*/
if (uio->uio_offset + uio->uio_resid >
VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize) {
lwkt_reltoken(&vp->v_mount->mnt_token);
return (EFBIG);
}
if (vp->v_type == VREG && td != NULL) {
error = kern_getrlimit(RLIMIT_FSIZE, &limit);
if (error != 0) {
lwkt_reltoken(&vp->v_mount->mnt_token);
return error;
}
if (uio->uio_offset + uio->uio_resid > limit.rlim_cur) {
ksignal(td->td_proc, SIGXFSZ);
lwkt_reltoken(&vp->v_mount->mnt_token);
return (EFBIG);
}
}
/*
* Extend the file's size if necessary
*/
extended = ((uio->uio_offset + uio->uio_resid) > node->tn_size);
while (uio->uio_resid > 0) {
/*
* Use buffer cache I/O (via tmpfs_strategy)
*/
offset = (size_t)uio->uio_offset & BMASK;
base_offset = (off_t)uio->uio_offset - offset;
len = BSIZE - offset;
if (len > uio->uio_resid)
len = uio->uio_resid;
if ((uio->uio_offset + len) > node->tn_size) {
trivial = (uio->uio_offset <= node->tn_size);
error = tmpfs_reg_resize(vp, uio->uio_offset + len, trivial);
if (error)
break;
}
/*
* Read to fill in any gaps. Theoretically we could
* optimize this if the write covers the entire buffer
* and is not a UIO_NOCOPY write, however this can lead
* to a security violation exposing random kernel memory
* (whatever junk was in the backing VM pages before).
*
* So just use bread() to do the right thing.
*/
error = bread(vp, base_offset, BSIZE, &bp);
error = uiomovebp(bp, (char *)bp->b_data + offset, len, uio);
if (error) {
kprintf("tmpfs_write uiomove error %d\n", error);
brelse(bp);
break;
}
if (uio->uio_offset > node->tn_size) {
node->tn_size = uio->uio_offset;
kflags |= NOTE_EXTEND;
}
kflags |= NOTE_WRITE;
/*
* Always try to flush the page if the request is coming
* from the pageout daemon (IO_ASYNC), else buwrite() the
* buffer.
*
* buwrite() dirties the underlying VM pages instead of
* dirtying the buffer, releasing the buffer as a clean
* buffer. This allows tmpfs to use essentially all
* available memory to cache file data. If we used bdwrite()
* the buffer cache would wind up flushing the data to
* swap too quickly.
*/
bp->b_flags |= B_AGE;
if (ap->a_ioflag & IO_ASYNC) {
bawrite(bp);
} else {
buwrite(bp);
}
if (bp->b_error) {
kprintf("tmpfs_write bwrite error %d\n", bp->b_error);
break;
}
}
if (error) {
if (extended) {
(void)tmpfs_reg_resize(vp, oldsize, trivial);
kflags &= ~NOTE_EXTEND;
}
goto done;
}
/*
* Currently we don't set the mtime on files modified via mmap()
* because we can't tell the difference between those modifications
* and an attempt by the pageout daemon to flush tmpfs pages to
* swap.
*
* This is because in order to defer flushes as long as possible
* buwrite() works by marking the underlying VM pages dirty in
* order to be able to dispose of the buffer cache buffer without
* flushing it.
*/
TMPFS_NODE_LOCK(node);
if (uio->uio_segflg != UIO_NOCOPY)
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED;
if (extended)
node->tn_status |= TMPFS_NODE_CHANGED;
if (node->tn_mode & (S_ISUID | S_ISGID)) {
if (priv_check_cred(ap->a_cred, PRIV_VFS_RETAINSUGID, 0))
node->tn_mode &= ~(S_ISUID | S_ISGID);
}
TMPFS_NODE_UNLOCK(node);
done:
tmpfs_knote(vp, kflags);
lwkt_reltoken(&vp->v_mount->mnt_token);
return(error);
}
static int
tmpfs_nlink(struct vop_nlink_args *v)
{
struct vnode *dvp = v->a_dvp;
struct vnode *vp = v->a_vp;
struct namecache *ncp = v->a_nch->ncp;
struct tmpfs_dirent *de;
struct tmpfs_node *node;
struct tmpfs_node *dnode;
struct mount *mp;
int error;
mp = dvp->v_mount;
KKASSERT(dvp != vp); /* XXX When can this be false? */
node = VP_TO_TMPFS_NODE(vp);
dnode = VP_TO_TMPFS_NODE(dvp);
TMPFS_NODE_LOCK(dnode);
/* XXX: Why aren't the following two tests done by the caller? */
/* Hard links of directories are forbidden. */
if (vp->v_type == VDIR) {
error = EPERM;
goto out;
}
/* Cannot create cross-device links. */
if (dvp->v_mount != vp->v_mount) {
error = EXDEV;
goto out;
}
/* Ensure that we do not overflow the maximum number of links imposed
* by the system. */
KKASSERT(node->tn_links <= LINK_MAX);
if (node->tn_links >= LINK_MAX) {
error = EMLINK;
goto out;
}
/* We cannot create links of files marked immutable or append-only. */
if (node->tn_flags & (IMMUTABLE | APPEND)) {
error = EPERM;
goto out;
}
/* Allocate a new directory entry to represent the node. */
error = tmpfs_alloc_dirent(VFS_TO_TMPFS(vp->v_mount), node,
ncp->nc_name, ncp->nc_nlen, &de);
if (error != 0)
goto out;
/* Insert the new directory entry into the appropriate directory. */
tmpfs_dir_attach(dnode, de);
/* vp link count has changed, so update node times. */
TMPFS_NODE_LOCK(node);
node->tn_status |= TMPFS_NODE_CHANGED;
TMPFS_NODE_UNLOCK(node);
tmpfs_update(vp);
tmpfs_knote(vp, NOTE_LINK);
cache_setunresolved(v->a_nch);
cache_setvp(v->a_nch, vp);
error = 0;
out:
TMPFS_NODE_UNLOCK(dnode);
if (error == 0)
tmpfs_knote(dvp, NOTE_WRITE);
return error;
}
/*
* Destroys the node pointed to by node from the file system 'tmp'.
* If the node does not belong to the given mount point, the results are
* unpredicted.
*
* If the node references a directory; no entries are allowed because
* their removal could need a recursive algorithm, something forbidden in
* kernel space. Furthermore, there is not need to provide such
* functionality (recursive removal) because the only primitives offered
* to the user are the removal of empty directories and the deletion of
* individual files.
*
* Note that nodes are not really deleted; in fact, when a node has been
* allocated, it cannot be deleted during the whole life of the file
* system. Instead, they are moved to the available list and remain there
* until reused.
*
* A caller must have TMPFS_NODE_LOCK(node) and this function unlocks it.
*/
void
tmpfs_free_node(struct tmpfs_mount *tmp, struct tmpfs_node *node)
{
vm_pindex_t pages = 0;
#ifdef INVARIANTS
TMPFS_ASSERT_ELOCKED(node);
KKASSERT(node->tn_vnode == NULL);
KKASSERT((node->tn_vpstate & TMPFS_VNODE_ALLOCATING) == 0);
#endif
TMPFS_LOCK(tmp);
LIST_REMOVE(node, tn_entries);
tmp->tm_nodes_inuse--;
TMPFS_UNLOCK(tmp);
TMPFS_NODE_UNLOCK(node); /* Caller has this lock */
switch (node->tn_type) {
case VNON:
/* Do not do anything. VNON is provided to let the
* allocation routine clean itself easily by avoiding
* duplicating code in it. */
/* FALLTHROUGH */
case VBLK:
/* FALLTHROUGH */
case VCHR:
/* FALLTHROUGH */
break;
case VDIR:
/*
* The parent link can be NULL if this is the root
* node or if it is a directory node that was rmdir'd.
*
* XXX what if node is a directory which still contains
* directory entries (e.g. due to a forced umount) ?
*/
node->tn_size = 0;
KKASSERT(node->tn_dir.tn_parent == NULL);
/*
* If the root node is being destroyed don't leave a
* dangling pointer in tmpfs_mount.
*/
if (node == tmp->tm_root)
tmp->tm_root = NULL;
break;
case VFIFO:
/* FALLTHROUGH */
case VSOCK:
break;
case VLNK:
kfree(node->tn_link, tmp->tm_name_zone);
node->tn_link = NULL;
node->tn_size = 0;
break;
case VREG:
if (node->tn_reg.tn_aobj != NULL)
vm_object_deallocate(node->tn_reg.tn_aobj);
node->tn_reg.tn_aobj = NULL;
pages = node->tn_reg.tn_aobj_pages;
break;
default:
panic("tmpfs_free_node: type %p %d", node, (int)node->tn_type);
}
/*
* Clean up fields for the next allocation. The objcache only ctors
* new allocations.
*/
tmpfs_node_ctor(node, NULL, 0);
objcache_put(tmp->tm_node_pool, node);
/* node is now invalid */
if (pages)
atomic_add_long(&tmp->tm_pages_used, -(long)pages);
}
