On Wed, Mar 31, 2021 at 07:33:50PM +0200, Mickaël Salaün wrote:
> > +static inline u64 unmask_layers(
> > + const struct landlock_ruleset *const domain,
> > + const struct path *const path, const u32 access_request,
> > + u64 layer_mask)
> > +{
> > + const struct landlock_rule *rule;
> > + const struct inode *inode;
> > + size_t i;
> > +
> > + if (d_is_negative(path->dentry))
> > + /* Continues to walk while there is no mapped inode. */
^^^^^
Odd comment, that...
> > +static int check_access_path(const struct landlock_ruleset *const domain,
> > + const struct path *const path, u32 access_request)
> > +{
> > + walker_path = *path;
> > + path_get(&walker_path);
> > + while (true) {
> > + struct dentry *parent_dentry;
> > +
> > + layer_mask = unmask_layers(domain, &walker_path,
> > + access_request, layer_mask);
> > + if (layer_mask == 0) {
> > + /* Stops when a rule from each layer grants access. */
> > + allowed = true;
> > + break;
> > + }
> > +
> > +jump_up:
> > + if (walker_path.dentry == walker_path.mnt->mnt_root) {
> > + if (follow_up(&walker_path)) {
> > + /* Ignores hidden mount points. */
> > + goto jump_up;
> > + } else {
> > + /*
> > + * Stops at the real root. Denies access
> > + * because not all layers have granted access.
> > + */
> > + allowed = false;
> > + break;
> > + }
> > + }
> > + if (unlikely(IS_ROOT(walker_path.dentry))) {
> > + /*
> > + * Stops at disconnected root directories. Only allows
> > + * access to internal filesystems (e.g. nsfs, which is
> > + * reachable through /proc/<pid>/ns/<namespace>).
> > + */
> > + allowed = !!(walker_path.mnt->mnt_flags & MNT_INTERNAL);
> > + break;
> > + }
> > + parent_dentry = dget_parent(walker_path.dentry);
> > + dput(walker_path.dentry);
> > + walker_path.dentry = parent_dentry;
> > + }
> > + path_put(&walker_path);
> > + return allowed ? 0 : -EACCES;
That's a whole lot of grabbing/dropping references... I realize that it's
an utterly tactless question, but... how costly it is? IOW, do you have
profiling data?
> > +/*
> > + * pivot_root(2), like mount(2), changes the current mount namespace. It must
> > + * then be forbidden for a landlocked process.
... and cross-directory rename(2) can change the tree topology. Do you ban that
as well?
[snip]
> > +static int hook_path_rename(const struct path *const old_dir,
> > + struct dentry *const old_dentry,
> > + const struct path *const new_dir,
> > + struct dentry *const new_dentry)
> > +{
> > + const struct landlock_ruleset *const dom =
> > + landlock_get_current_domain();
> > +
> > + if (!dom)
> > + return 0;
> > + /* The mount points are the same for old and new paths, cf. EXDEV. */
> > + if (old_dir->dentry != new_dir->dentry)
> > + /* For now, forbids reparenting. */
> > + return -EACCES;
You do, apparently, and not in a way that would have the userland fall
back to copy+unlink. Lovely... Does e.g. git survive such restriction?
Same question for your average package build...
