Re: [PATCH v2 1/1] ppp: Fix one deadlock issue of PPP when send frame

Hi Guillaume,

Thanks your detail analyses.
Now I think it is a good solution that just drop the packet and print
error log instead my original solution that supports reentrant. This
solution will not bring any side effects.

I will send one update according to this new solution.


Regards
Feng


On Fri, Aug 19, 2016 at 12:13 AM, Guillaume Nault [off-list ref] wrote:

On Thu, Aug 18, 2016 at 08:58:31AM +0800, Feng Gao wrote:

quoted

On Thu, Aug 18, 2016 at 1:42 AM, Guillaume Nault [off-list ref] wrote:

quoted

On Tue, Aug 16, 2016 at 07:33:38PM +0800, fgao@ikuai8.com wrote:

quoted

From: Gao Feng <redacted>

PPP channel holds one spinlock before send frame. But the skb may
select the same PPP channel with wrong route policy. As a result,
the skb reaches the same channel path. It tries to get the same
spinlock which is held before. Bang, the deadlock comes out.

Unless I misunderstood the problem you're trying to solve, this patch
doesn't really help: deadlock still occurs if the same IP is used for
L2TP and PPP's peer address.

The deadlock happens because the same cpu try to hold the spinlock
which is already held before by itself.
Now the PPP_CHANNEL_LOCK_BH sets the lock owner after hold lock, then
when the same cpu
invokes PPP_CHANNEL_LOCK_BH again. The cl.owner equals current cpu id,
so it only increases
the lock_cnt without trying to hold the lock again.
So it avoids the deadlock.

I'm sorry but, again, it just _moves_ the deadlock down to L2TP. The
kernel still oops because, now, l2tp_xmit_skb() is called recursively
while holding its tunnel socket.

quoted

quoted

quoted

Now add one lock owner to avoid it like xmit_lock_owner of
netdev_queue. Check the lock owner before try to get the spinlock.
If the current cpu is already the owner, needn't lock again. When
PPP channel holds the spinlock at the first time, it sets owner
with current CPU ID.

I think you should forbid lock recursion entirely, and drop the packet
if the owner tries to re-acquire the channel lock. Otherwise you just
move the deadlock down the stack (l2tp_xmit_skb() can't be called
recursively).

The reason that fix it in ppp is that there are other layer on the ppp module.
We resolve it in ppp module, it could avoid all similar potential issues.

Not sure if I understand your point here.
The xmit path of PPP and its sub-layers hasn't been designed to be
reentrant. Allowing recursive sends thus require to review the full
path.

Beyond the L2TP issue discussed above, just consider the locking
dependencies used in PPP: ppp->wlock has to be held before
channel->downl. Sending a packet directly on a channel will lock
channel->downl. If this packet is routed back to the parent unit then
ppp_xmit_process() will lock ppp->wlock, effectively leading to lock
inversion and potential deadlock.

So we have two options: adapt the whole xmit path to handle recursive
sends or forbid recursion entirely. Unfortunately none of these options
looks easy to achieve:

  * Making PPP xmit path reentrant will be hard and error prone because
    of all the locking dependencies. Looks like simplifying PPP's
    locking scheme will be required first.

  * I can't see any way to reliably prevent settings where a packet sent
    on a given channel would be routed back to the parent unit.


OTOH, we can try to limit the impact of recursive sends for simple
cases:

  * Following your approach, either adapt the lower layers
    (like l2tp_xmit_skb() for L2TP), or drop the packet when
    cl.owner == smp_processor_id(). This is very limited in scope and
    doesn't address issues like locking inversions. But it may let the
    system survive long enough for the PPP to time out.

  * In the lower layer, check where the packet is going to be enqueued
    and drop it if it's the parent device. That should reliably handle
    simple and common cases. However this requires to update at least
    L2TP and PPTP and to get a way to access the parent device. Also,
    it doesn't prevent recursion with stacked interfaces.