Re: [RFC] call_with_creds()

[Al Viro]

On Wed, Jul 18, 2018 at 11:19:18AM -0700, Linus Torvalds wrote:
> On Wed, Jul 18, 2018 at 11:13 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > Linus, David - do you have any objections to the above?
> 
> I damn well do.
> 
> I explained earlier why it's wrong and fragile, and why it can just
> cause the *reverse* security problem if you do it wrong. So now you
> take a subtle bug, and make it even more subtle, and encourage people
> to do this known-broken model of using creds at IO time.
> 
> No.
> 
> Some debugging option to just clear current->creds entirely and catch
> mis-uses, sure. But saying "we have shit buggy garbage in random write
> functions, so we'll just paper over it"? No.

Huh?  Nevermind ->write(), what about open()?  Here's a specific question
Miklos brought when I suggested to get rid of that override:
/*
 * These allocate and release file read/write context information.
 */
int nfs_open(struct inode *inode, struct file *filp)
{
        struct nfs_open_context *ctx;

        ctx = alloc_nfs_open_context(file_dentry(filp), filp->f_mode, filp);

struct nfs_open_context *alloc_nfs_open_context(struct dentry *dentry,
                                                fmode_t f_mode,
                                                struct file *filp)
{
        struct nfs_open_context *ctx;
        struct rpc_cred *cred = rpc_lookup_cred();

struct rpc_cred *rpc_lookup_cred(void)
{
        return rpcauth_lookupcred(&generic_auth, 0);

struct rpc_cred *
rpcauth_lookupcred(struct rpc_auth *auth, int flags)
{
        struct auth_cred acred;
        struct rpc_cred *ret;
        const struct cred *cred = current_cred();

How should we bring the cred passed to do_dentry_open() where open() has been
called to rpcauth_lookupcred() where we end up looking for rpc_cred by what
should've been the cred passed to do_dentry_open() and is, instead, current_cred()?

We can pass filp->f_cred to rpc_lookup_cred() variant that gets it as an explicit
argument and feed it down to rpcauth_lookupcred() variant that does the same.
We can basically ignore the ->f_cred here.  Or we can get current_cred() equal
to ->f_cred for the duration of open().

I'd probably prefer the first variant, but the last part of the question Miklos
asked
> Okay, so ->open() is a file op, and file ops should use file->f_cred,
> but how are we going to enforce this?
is not trivial - how do we find the places where that kind of thing happens and
what do we do in the meanwhile?  I don't see any quick answers - any suggestions
would be very welcome.  It's not just direct current_cred() callers; that stuff
gets called deep in call chains.  And lifting it all the way up means a lot of
methods that need to get an explicit struct cred * argument.  Are you OK with
going in that direction?

I'm honestly not sure - it's not an attempt to maneuver you into changing your
policy re ->write().  Do we care about ->f_cred at all and if we do, how do we
get it consistent across the filesystems?  I'd buy "it's a weird and obscure thing"
for overlayfs, but that example is on NFS...

We definitely do have bugs in that area - consider e.g.
static int ecryptfs_threadfn(void *ignored)
{
        set_freezable();
        while (1)  {
                struct ecryptfs_open_req *req;

                wait_event_freezable(
                        ecryptfs_kthread_ctl.wait,
                        (!list_empty(&ecryptfs_kthread_ctl.req_list)
                         || kthread_should_stop()));
                mutex_lock(&ecryptfs_kthread_ctl.mux);
                if (ecryptfs_kthread_ctl.flags & ECRYPTFS_KTHREAD_ZOMBIE) {
                        mutex_unlock(&ecryptfs_kthread_ctl.mux);
                        goto out;
                }
                while (!list_empty(&ecryptfs_kthread_ctl.req_list)) {
                        req = list_first_entry(&ecryptfs_kthread_ctl.req_list,
                                               struct ecryptfs_open_req,
                                               kthread_ctl_list);
                        list_del(&req->kthread_ctl_list);
                        *req->lower_file = dentry_open(&req->path,
                                (O_RDWR | O_LARGEFILE), current_cred());
                        complete(&req->done);
                }
                mutex_unlock(&ecryptfs_kthread_ctl.mux);
        }
out:
        return 0;
}

It's a kernel thread, so current_cred() looks bogus...