Re: [PATCH] mailbox: fix a UAF bug in msg_submit()

[Xianting TIan]

在 2021/8/17 下午1:58, Xianting TIan 写道:
> 在 2021/8/17 下午12:29, Jassi Brar 写道:
> > On Fri, Aug 6, 2021 at 7:15 AM Xianting Tian
> > <xianting.tian@xxxxxxxxxxxxxxxxx> wrote:
> > > We met a UAF issue during our mailbox testing.
> > >
> > > In synchronous mailbox, we use mbox_send_message() to send a message
> > > and wait for completion. mbox_send_message() calls msg_submit() to
> > > send the message for the first time, if timeout, it will send the
> > > message in tx_tick() for the second time.
> > Seems like your controller's  .send_data() returns error. Can you
> > please explain why it does so? Because
> > send_data() only _accepts_ data for further transmission... which
> > should seldom be a problem.
>
> Thanks for the comments,
>
> We developed virtio-mailbox for heterogeneous virtualization system.
>
> virtio-mailbox is based on the mialbox framework.
>
> In virtio framework, its send func 'virtqueue_add_outbuf()' may return 
> error, which caused .send_data() return error.  And also contains 
> other csenarios.
>
> But I think mailbox framework shouldn't depend on .send_data() always 
> return OK,  as .send_data() is implemented by mailbox hardware 
> manufacturer, which is not controlled by mailbox framework itself.
>
> You said 'seldom',  but it still possible we can meet such issue. sucn 
> as flexrm_send_data() of drivers/mailbox/bcm-flexrm-mailbox.c.
>
> I think mailbox framework should be work normaly no matter 
> .send_data() returns ok or not ok.  Do you think so? thanks

Another solution is to ignore the return value of .send_data() in 
msg_submit(),

change

        err = chan->mbox->ops->send_data(chan, data);
        if (!err) {
                chan->active_req = data;
                chan->msg_count--;
        }

to

        chan->mbox->ops->send_data(chan, data);
        chan->active_req = data;
        chan->msg_count--;

But the side effect of the solution is obvious, as if send failed in the 
first time, it will have no chance to sent it in tx_tick() for the 
second time. That is to say, no retry.

So I think the solution in this patch is better.

Looking forward to your further comments, thanks

> > thanks.