On  Thu, Dec 03, 2020 at 01:23:18 +0100, Guillaume Nault wrote:

On Tue, Dec 01, 2020 at 11:52:49AM +0000, Tom Parkin wrote:

quoted

+static int ppp_bridge_channels(struct channel *pch, struct channel *pchb)
+{
+	write_lock_bh(&pch->upl);
+	if (pch->ppp || pch->bridge) {

Since ->bridge is RCU protected, it should be dereferenced with
rcu_dereference_protected() here:
rcu_dereference_protected(pch->bridge, lockdep_is_held(&pch->upl)).

Ack, thanks.

Ditto for the other callsites which should also be using 
rcu_dereference_protected for access to the rcu-protected pointer.

<snip>

quoted

+	if (!pchb) {
+		write_unlock_bh(&pch->upl);
+		return -EINVAL;

I'm not sure I'd consider this case as an error.

To be honest I'd probably tend agree with you, but I was seeking to
maintain consistency with how PPPIOCCONNECT/PPPIOCDISCONN behave.  The
latter returns EINVAL if the channel isn't connected to an interface.

If you feel strongly I'm happy to change it but IMO it's better to be
consistent with existing ioctl calls.

If there's no bridged channel, there's just nothing to do.
Furthermore, there might be situations where this is not really an
error (see the possible race below).

quoted

+	}
+	RCU_INIT_POINTER(pch->bridge, NULL);
+	write_unlock_bh(&pch->upl);
+
+	write_lock_bh(&pchb->upl);
+	RCU_INIT_POINTER(pchb->bridge, NULL);
+	write_unlock_bh(&pchb->upl);
+
+	synchronize_rcu();
+
+	if (refcount_dec_and_test(&pch->file.refcnt))
+		ppp_destroy_channel(pch);

I think that we could have a situation where pchb->bridge could be
different from pch.
If ppp_unbridge_channels() was also called concurrently on pchb, then
pchb->bridge might have been already reset. And it might have dropped
the reference it had on pch. In this case, we'd erroneously decrement
the refcnt again.

In theory, pchb->bridge might even have been reassigned to a different
channel. So we'd reset pchb->bridge, but without decrementing the
refcnt of the channel it pointed to (and again, we'd erroneously
decrement pch's refcount instead).

So I think we need to save pchb->bridge to a local variable while we
hold pchb->upl, and then drop the refcount of that channel, instead of
assuming that it's equal to pch.

Ack, yes.

The v1 series protected against this, although by nesting locks :-|

I think in the case that pchb->bridge != pch, we probably want to
leave pchb alone, so:

 1. Don't unset the pchb->bridge pointer
 2. Don't drop the pch reference (pchb doesn't hold a reference on pch
    because pchb->bridge != pch)

This is on the assumption that pchb has been reassigned -- in that
scenario we don't want to mess with pchb at all since it'll break the
other bridge instance.