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);
vnode_destroy_vobject(vp);
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;
}
/*
* Caller must hold an exclusive node lock.
*/
int
tmpfs_truncate(struct vnode *vp, off_t length)
{
int error;
struct tmpfs_node *node;
node = VP_TO_TMPFS_NODE(vp);
if (length < 0) {
error = EINVAL;
goto out;
}
if (node->tn_size == length) {
error = 0;
goto out;
}
if (length > VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize)
return (EFBIG);
error = tmpfs_reg_resize(vp, length, 1);
if (error == 0)
node->tn_status |= TMPFS_NODE_CHANGED | TMPFS_NODE_MODIFIED;
out:
tmpfs_update(vp);
return error;
}
static int
tmpfs_pathconf(struct vop_pathconf_args *ap)
{
struct vnode *vp = ap->a_vp;
int name = ap->a_name;
register_t *retval = ap->a_retval;
struct tmpfs_mount *tmp;
int error;
error = 0;
switch (name) {
case _PC_CHOWN_RESTRICTED:
*retval = 1;
break;
case _PC_FILESIZEBITS:
tmp = VFS_TO_TMPFS(vp->v_mount);
*retval = max(32, flsll(tmp->tm_pages_max * PAGE_SIZE) + 1);
break;
case _PC_LINK_MAX:
*retval = LINK_MAX;
break;
case _PC_NAME_MAX:
*retval = NAME_MAX;
break;
case _PC_NO_TRUNC:
*retval = 1;
break;
case _PC_PATH_MAX:
*retval = PATH_MAX;
break;
case _PC_PIPE_BUF:
*retval = PIPE_BUF;
break;
case _PC_SYNC_IO:
*retval = 1;
break;
case _PC_2_SYMLINKS:
*retval = 1;
break;
default:
error = EINVAL;
}
return error;
}
static int
tmpfs_fifo_close(struct vop_close_args *v)
{
struct tmpfs_node *node;
node = VP_TO_TMPFS_NODE(v->a_vp);
tmpfs_set_status(VFS_TO_TMPFS(v->a_vp->v_mount), node,
TMPFS_NODE_ACCESSED);
tmpfs_update(v->a_vp);
return (fifo_specops.vop_close(v));
}
static int
tmpfs_write(struct vop_write_args *v)
{
struct vnode *vp;
struct uio *uio;
struct tmpfs_node *node;
off_t oldsize;
int error, ioflag;
boolean_t extended;
vp = v->a_vp;
uio = v->a_uio;
ioflag = v->a_ioflag;
error = 0;
node = VP_TO_TMPFS_NODE(vp);
oldsize = node->tn_size;
if (uio->uio_offset < 0 || vp->v_type != VREG)
return (EINVAL);
if (uio->uio_resid == 0)
return (0);
if (ioflag & IO_APPEND)
uio->uio_offset = node->tn_size;
if (uio->uio_offset + uio->uio_resid >
VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize)
return (EFBIG);
if (vn_rlimit_fsize(vp, uio, uio->uio_td))
return (EFBIG);
extended = uio->uio_offset + uio->uio_resid > node->tn_size;
if (extended) {
error = tmpfs_reg_resize(vp, uio->uio_offset + uio->uio_resid,
FALSE);
if (error != 0)
goto out;
}
error = uiomove_object(node->tn_reg.tn_aobj, node->tn_size, uio);
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED |
(extended ? TMPFS_NODE_CHANGED : 0);
if (node->tn_mode & (S_ISUID | S_ISGID)) {
if (priv_check_cred(v->a_cred, PRIV_VFS_RETAINSUGID, 0))
node->tn_mode &= ~(S_ISUID | S_ISGID);
}
if (error != 0)
(void)tmpfs_reg_resize(vp, oldsize, TRUE);
out:
MPASS(IMPLIES(error == 0, uio->uio_resid == 0));
MPASS(IMPLIES(error != 0, oldsize == node->tn_size));
return (error);
}
static int
tmpfs_remove(struct vop_remove_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));
if (vp->v_type == VDIR) {
error = EISDIR;
goto out;
}
dnode = VP_TO_TMPFS_DIR(dvp);
node = VP_TO_TMPFS_NODE(vp);
tmp = VFS_TO_TMPFS(vp->v_mount);
de = tmpfs_dir_lookup(dnode, node, v->a_cnp);
MPASS(de != NULL);
/* 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(dvp, de);
if (v->a_cnp->cn_flags & DOWHITEOUT)
tmpfs_dir_whiteout_add(dvp, v->a_cnp);
/* 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);
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED;
error = 0;
out:
return error;
}
static int
tmpfs_link(struct vop_link_args *v)
{
struct vnode *dvp = v->a_tdvp;
struct vnode *vp = v->a_vp;
struct componentname *cnp = v->a_cnp;
int error;
struct tmpfs_dirent *de;
struct tmpfs_node *node;
MPASS(VOP_ISLOCKED(dvp));
MPASS(cnp->cn_flags & HASBUF);
MPASS(dvp != vp); /* XXX When can this be false? */
node = VP_TO_TMPFS_NODE(vp);
/* Ensure that we do not overflow the maximum number of links imposed
* by the system. */
MPASS(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,
cnp->cn_nameptr, cnp->cn_namelen, &de);
if (error != 0)
goto out;
/* Insert the new directory entry into the appropriate directory. */
if (cnp->cn_flags & ISWHITEOUT)
tmpfs_dir_whiteout_remove(dvp, cnp);
tmpfs_dir_attach(dvp, de);
/* vp link count has changed, so update node times. */
node->tn_status |= TMPFS_NODE_CHANGED;
tmpfs_update(vp);
error = 0;
out:
return error;
}
/*
* fsync is usually a NOP, but we must take action when unmounting or
* when recycling.
*/
static int
tmpfs_fsync(struct vop_fsync_args *v)
{
struct tmpfs_mount *tmp;
struct tmpfs_node *node;
struct vnode *vp = v->a_vp;
tmp = VFS_TO_TMPFS(vp->v_mount);
node = VP_TO_TMPFS_NODE(vp);
tmpfs_update(vp);
if (vp->v_type == VREG) {
if (vp->v_flag & VRECLAIMED) {
if (node->tn_links == 0)
tmpfs_truncate(vp, 0);
else
vfsync(v->a_vp, v->a_waitfor, 1, NULL, NULL);
}
}
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;
}
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_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_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;
}
/*
* Resizes the aobj associated to the regular file pointed to by vp to
* the size newsize. 'vp' must point to a vnode that represents a regular
* file. 'newsize' must be positive.
*
* pass trivial as 1 when buf content will be overwritten, otherwise set 0
* to be zero filled.
*
* Returns zero on success or an appropriate error code on failure.
*
* Caller must hold the node exclusively locked.
*/
int
tmpfs_reg_resize(struct vnode *vp, off_t newsize, int trivial)
{
int error;
vm_pindex_t newpages, oldpages;
struct tmpfs_mount *tmp;
struct tmpfs_node *node;
off_t oldsize;
#ifdef INVARIANTS
KKASSERT(vp->v_type == VREG);
KKASSERT(newsize >= 0);
#endif
node = VP_TO_TMPFS_NODE(vp);
tmp = VFS_TO_TMPFS(vp->v_mount);
/*
* Convert the old and new sizes to the number of pages needed to
* store them. It may happen that we do not need to do anything
* because the last allocated page can accommodate the change on
* its own.
*/
oldsize = node->tn_size;
oldpages = round_page64(oldsize) / PAGE_SIZE;
KKASSERT(oldpages == node->tn_reg.tn_aobj_pages);
newpages = round_page64(newsize) / PAGE_SIZE;
if (newpages > oldpages &&
tmp->tm_pages_used + newpages - oldpages > tmp->tm_pages_max) {
error = ENOSPC;
goto out;
}
node->tn_reg.tn_aobj_pages = newpages;
node->tn_size = newsize;
if (newpages != oldpages)
atomic_add_long(&tmp->tm_pages_used, (newpages - oldpages));
/*
* When adjusting the vnode filesize and its VM object we must
* also adjust our backing VM object (aobj). The blocksize
* used must match the block sized we use for the buffer cache.
*
* The backing VM object may contain VM pages as well as swap
* assignments if we previously renamed main object pages into
* it during deactivation.
*/
if (newsize < oldsize) {
vm_pindex_t osize;
vm_pindex_t nsize;
vm_object_t aobj;
error = nvtruncbuf(vp, newsize, TMPFS_BLKSIZE, -1, 0);
aobj = node->tn_reg.tn_aobj;
if (aobj) {
osize = aobj->size;
nsize = vp->v_object->size;
if (nsize < osize) {
aobj->size = osize;
swap_pager_freespace(aobj, nsize,
osize - nsize);
vm_object_page_remove(aobj, nsize, osize,
FALSE);
}
}
} else {
vm_object_t aobj;
error = nvextendbuf(vp, oldsize, newsize,
TMPFS_BLKSIZE, TMPFS_BLKSIZE,
-1, -1, trivial);
aobj = node->tn_reg.tn_aobj;
if (aobj)
aobj->size = vp->v_object->size;
}
out:
return error;
}
/*
* 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;
}
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_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)
goto outok;
cnt++;
}
if (uio->uio_offset == TMPFS_DIRCOOKIE_DOTDOT) {
error = tmpfs_dir_getdotdotdent(tmp, node, uio);
if (error != 0)
goto outok;
cnt++;
}
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 = TAILQ_FIRST(&node->tn_dir.tn_dirhead);
} else if (de != NULL) {
de = TAILQ_NEXT(de, td_entries);
} else {
de = tmpfs_dir_lookupbycookie(node,
off);
KKASSERT(de != NULL);
de = TAILQ_NEXT(de, td_entries);
}
if (de == NULL)
off = TMPFS_DIRCOOKIE_EOF;
else
off = tmpfs_dircookie(de);
}
(*cookies)[i] = off;
}
KKASSERT(uio->uio_offset == off);
}
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_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_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;
}
static int
tmpfs_write(struct vop_write_args *v)
{
struct vnode *vp = v->a_vp;
struct uio *uio = v->a_uio;
int ioflag = v->a_ioflag;
boolean_t extended;
int error = 0;
off_t oldsize;
struct tmpfs_node *node;
vm_object_t uobj;
size_t len;
int resid;
node = VP_TO_TMPFS_NODE(vp);
oldsize = node->tn_size;
if (uio->uio_offset < 0 || vp->v_type != VREG) {
error = EINVAL;
goto out;
}
if (uio->uio_resid == 0) {
error = 0;
goto out;
}
if (ioflag & IO_APPEND)
uio->uio_offset = node->tn_size;
if (uio->uio_offset + uio->uio_resid >
VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize)
return (EFBIG);
if (vn_rlimit_fsize(vp, uio, uio->uio_td))
return (EFBIG);
extended = uio->uio_offset + uio->uio_resid > node->tn_size;
if (extended) {
error = tmpfs_reg_resize(vp, uio->uio_offset + uio->uio_resid,
FALSE);
if (error != 0)
goto out;
}
uobj = node->tn_reg.tn_aobj;
while ((resid = uio->uio_resid) > 0) {
if (node->tn_size <= uio->uio_offset)
break;
len = MIN(node->tn_size - uio->uio_offset, resid);
if (len == 0)
break;
error = tmpfs_mappedwrite(vp->v_object, uobj, len, uio);
if ((error != 0) || (resid == uio->uio_resid))
break;
}
node->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED |
(extended ? TMPFS_NODE_CHANGED : 0);
if (node->tn_mode & (S_ISUID | S_ISGID)) {
if (priv_check_cred(v->a_cred, PRIV_VFS_RETAINSUGID, 0))
node->tn_mode &= ~(S_ISUID | S_ISGID);
}
if (error != 0)
(void)tmpfs_reg_resize(vp, oldsize, TRUE);
out:
MPASS(IMPLIES(error == 0, uio->uio_resid == 0));
MPASS(IMPLIES(error != 0, oldsize == node->tn_size));
return error;
}
