Re: [Linux-kernel-mentees] [PATCH] net: rose: Fix Null pointer dereference in rose_send_frame()

[Greg KH <gregkh@linuxfoundation.org>]

On Fri, Oct 30, 2020 at 04:24:13PM +0530, Anmol Karn wrote:
> On Thu, Oct 15, 2020 at 05:50:51PM +0200, Greg KH wrote:
> > On Thu, Oct 15, 2020 at 07:40:12PM +0530, Anmol Karn wrote:
> > > On Thu, Oct 15, 2020 at 07:12:25AM +0200, Greg KH wrote:
> > > > On Thu, Oct 15, 2020 at 05:47:12AM +0530, Anmol Karn wrote:
> > > > > In rose_send_frame(), when comparing two ax.25 addresses, it assigns rose_call to 
> > > > > either global ROSE callsign or default port, but when the former block triggers and 
> > > > > rose_call is assigned by (ax25_address *)neigh->dev->dev_addr, a NULL pointer is 
> > > > > dereferenced by 'neigh' when dereferencing 'dev'.
> > > > > 
> > > > > - net/rose/rose_link.c
> > > > > This bug seems to get triggered in this line:
> > > > > 
> > > > > rose_call = (ax25_address *)neigh->dev->dev_addr;
> > > > > 
> > > > > Prevent it by checking NULL condition for neigh->dev before comparing addressed for 
> > > > > rose_call initialization.
> > > > > 
> > > > > Reported-by: syzbot+a1c743815982d9496393@syzkaller.appspotmail.com 
> > > > > Link: https://syzkaller.appspot.com/bug?id=9d2a7ca8c7f2e4b682c97578dfa3f236258300b3 
> > > > > Signed-off-by: Anmol Karn <anmol.karan123@gmail.com>
> > > > > ---
> > > > > I am bit sceptical about the error return code, please suggest if anything else is 
> > > > > appropriate in place of '-ENODEV'.
> > > > > 
> > > > >  net/rose/rose_link.c | 3 +++
> > > > >  1 file changed, 3 insertions(+)
> > > > > 
> > > > > diff --git a/net/rose/rose_link.c b/net/rose/rose_link.c
> > > > > index f6102e6f5161..92ea6a31d575 100644
> > > > > --- a/net/rose/rose_link.c
> > > > > +++ b/net/rose/rose_link.c
> > > > > @@ -97,6 +97,9 @@ static int rose_send_frame(struct sk_buff *skb, struct rose_neigh *neigh)
> > > > >  	ax25_address *rose_call;
> > > > >  	ax25_cb *ax25s;
> > > > >  
> > > > > +	if (!neigh->dev)
> > > > > +		return -ENODEV;
> > > > 
> > > > How can ->dev not be set at this point in time?  Shouldn't that be
> > > > fixed, because it could change right after you check this, right?
> > > > 
> > > > thanks,
> > > > 
> > > > greg k-h
> > > 
> > > Hello Sir,
> > > 
> > > Thanks for the review,
> > > After following the call trace i thought, if neigh->dev is NULL it should
> > > be checked, but I will figure out what is going on with the crash reproducer,
> > > and I think rose_loopback_timer() is the place where problem started. 
> > > 
> > > Also, I have created a diff for checking neigh->dev before assigning ROSE callsign
> > > , please give your suggestions on this.
> > > 
> > > 
> > > diff --git a/net/rose/rose_link.c b/net/rose/rose_link.c
> > > index f6102e6f5161..2ddd5e559442 100644
> > > --- a/net/rose/rose_link.c
> > > +++ b/net/rose/rose_link.c
> > > @@ -97,10 +97,14 @@ static int rose_send_frame(struct sk_buff *skb, struct rose_neigh *neigh)
> > >         ax25_address *rose_call;
> > >         ax25_cb *ax25s;
> > >  
> > > -       if (ax25cmp(&rose_callsign, &null_ax25_address) == 0)
> > > -               rose_call = (ax25_address *)neigh->dev->dev_addr;
> > > -       else
> > > -               rose_call = &rose_callsign;
> > > +       if (neigh->dev) {
> > > +               if (ax25cmp(&rose_callsign, &null_ax25_address) == 0)
> > > +                       rose_call = (ax25_address *)neigh->dev->dev_addr;
> > > +               else
> > > +                       rose_call = &rose_callsign;
> > > +       } else {
> > > +               return -ENODEV;
> > > +       }
> > 
> > The point I am trying to make is that if someone else is setting ->dev
> > to NULL in some other thread/context/whatever, while this is running,
> > checking for it like this will not work.
> > 
> > What is the lifetime rules of that pointer?  Who initializes it, and who
> > sets it to NULL.  Figure that out first please to determine how to check
> > for this properly.
> > 
> > thanks,
> > 
> > greg k-h
> 
> Hello All,
> 
> I investigated further on this,
> 
> Here is some things i noticed:
> 
> When I followed the call trace,
> 
> [ 84.241331][ C3] Call Trace:
> [ 84.241331][ C3] rose_transmit_clear_request ($SOURCE/net/rose/rose_link.c:255)
> [ 84.241331][ C3] ? lockdep_hardirqs_on ($SOURCE/kernel/locking/lockdep.c:4161)
> [ 84.241331][ C3] rose_rx_call_request ($SOURCE/net/rose/af_rose.c:999)
> [ 84.241331][ C3] ? rose_release ($SOURCE/net/rose/af_rose.c:970)
> [ 84.241331][ C3] rose_loopback_timer ($SOURCE/net/rose/rose_loopback.c:100)
> [ 84.241331][ C3] ? rose_transmit_link ($SOURCE/net/rose/rose_loopback.c:60)
> 
> in the rose_send_frame() it dereferenced `neigh->dev` when called from 
> rose_transmit_clear_request(), and the first occurance of the `neigh`
> is in rose_loopback_timer() as `rose_loopback_neigh`, and it is initialized 
> in rose_add_loopback_neighh() as NULL.
> 
> - net/rose/rose_route.c:381
> 
> void rose_add_loopback_neigh(void)
> {
> 	struct rose_neigh *sn;
> 
> 	rose_loopback_neigh = kmalloc(sizeof(struct rose_neigh), GFP_KERNEL);
> 	if (!rose_loopback_neigh)
> 		return;
> 	sn = rose_loopback_neigh;
> 
> 	sn->callsign  = null_ax25_address;
> 	sn->digipeat  = NULL;
> 	sn->ax25      = NULL;
> 	sn->dev       = NULL;
> 	^^^^^^^^^^^^^^^^^^^^^
> 
> i.e when `rose_loopback_neigh` used in rose_loopback_timer() its `->dev` was
> still NULL and rose_loopback_timer() was calling rose_rx_call_request() 
> without checking for NULL.
> 
> 
> I have created the following patch to check for NULL pointer.
> 
> diff --git a/net/rose/rose_loopback.c b/net/rose/rose_loopback.c
> index 7b094275ea8b..cd7774cb1d07 100644
> --- a/net/rose/rose_loopback.c
> +++ b/net/rose/rose_loopback.c
> @@ -96,7 +96,7 @@ static void rose_loopback_timer(struct timer_list *unused)
>                 }
>  
>                 if (frametype == ROSE_CALL_REQUEST) {
> -                       if ((dev = rose_dev_get(dest)) != NULL) {
> +                       if (rose_loopback_neigh->dev && (dev = rose_dev_get(dest)) != NULL) {
>                                 if (rose_rx_call_request(skb, dev, rose_loopback_neigh, lci_o) == 0)
>                                         kfree_skb(skb);
>                         } else {
> 
> 
> 
> Please, review it and give me suggestions whether i am going right or not.

That seems better, does it solve the syzbot test?

thanks,

greg k-h