when register_netdevice(dev) fails we should check whether struct
veth_rq has been allocated via ndo_init callback and free it, because,
depending on the code path, register_netdevice() might not call
priv_destructor() callback
Reported-and-tested-by: [email protected]
Link: https://syzkaller.appspot.com/bug?extid=59ef240dd8f0ed7598a8
Signed-off-by: Rustam Kovhaev <[email protected]>
---
drivers/net/veth.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index a475f48d43c4..e40ca62a046a 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -1394,7 +1394,9 @@ static int veth_newlink(struct net *src_net, struct net_device *dev,
return 0;
err_register_dev:
- /* nothing to do */
+ priv = netdev_priv(dev);
+ if (priv->rq)
+ veth_dev_free(dev);
err_configure_peer:
unregister_netdevice(peer);
return err;
--
2.28.0
On 2020/08/30 22:13, Rustam Kovhaev wrote:
> when register_netdevice(dev) fails we should check whether struct
> veth_rq has been allocated via ndo_init callback and free it, because,
> depending on the code path, register_netdevice() might not call
> priv_destructor() callback
AFAICS, register_netdevice() always goto err_uninit and calls priv_destructor()
on failure after ndo_init() succeeded.
So I could not find such a code path.
Would you elaborate on it?
Thanks,
Toshiaki Makita
On Mon, Aug 31, 2020 at 09:16:32AM +0900, Toshiaki Makita wrote:
> On 2020/08/30 22:13, Rustam Kovhaev wrote:
> > when register_netdevice(dev) fails we should check whether struct
> > veth_rq has been allocated via ndo_init callback and free it, because,
> > depending on the code path, register_netdevice() might not call
> > priv_destructor() callback
>
> AFAICS, register_netdevice() always goto err_uninit and calls priv_destructor()
> on failure after ndo_init() succeeded.
> So I could not find such a code path.
> Would you elaborate on it?
in net/core/dev.c:9863, where register_netdevice() calls rollback_registered(),
which does not call priv_destructor(), then register_netdevice() returns error
net/core/dev.c:9884
On 2020/08/31 9:51, Rustam Kovhaev wrote:
> On Mon, Aug 31, 2020 at 09:16:32AM +0900, Toshiaki Makita wrote:
>> On 2020/08/30 22:13, Rustam Kovhaev wrote:
>>> when register_netdevice(dev) fails we should check whether struct
>>> veth_rq has been allocated via ndo_init callback and free it, because,
>>> depending on the code path, register_netdevice() might not call
>>> priv_destructor() callback
>>
>> AFAICS, register_netdevice() always goto err_uninit and calls priv_destructor()
>> on failure after ndo_init() succeeded.
>> So I could not find such a code path.
>> Would you elaborate on it?
>
> in net/core/dev.c:9863, where register_netdevice() calls rollback_registered(),
> which does not call priv_destructor(), then register_netdevice() returns error
> net/core/dev.c:9884
Thank you, now I see the code path.
But then all devices which allocate something in ndo_init() and free them in
priv_destructor() are affected? E.g. loopback and ifb seem to do such thing.
Why not calling priv_destructor() after invocation of rollback_registered()?
It looks weird that only that path does not call priv_destructor().
Toshiaki Makita
From: Rustam Kovhaev <[email protected]>
Date: Sun, 30 Aug 2020 06:13:36 -0700
> when register_netdevice(dev) fails we should check whether struct
> veth_rq has been allocated via ndo_init callback and free it, because,
> depending on the code path, register_netdevice() might not call
> priv_destructor() callback
>
> Reported-and-tested-by: [email protected]
> Link: https://syzkaller.appspot.com/bug?extid=59ef240dd8f0ed7598a8
> Signed-off-by: Rustam Kovhaev <[email protected]>
I think I agree with Toshiaki here. There is no reason why the
rollback_registered() path of register_netdevice() should behave
differently from the normal control flow.
Any code path that invokes ->ndo_uninit() should probably also
invoke the priv destructor.
The question is why does the err_uninit: label of register_netdevice
behave differently from rollback_registered()? If there is a reason,
it should be documented in a comment or similar. If it is wrong,
it should be corrected.
On Tue, Sep 01, 2020 at 01:01:27PM -0700, David Miller wrote:
> From: Rustam Kovhaev <[email protected]>
> Date: Sun, 30 Aug 2020 06:13:36 -0700
>
> > when register_netdevice(dev) fails we should check whether struct
> > veth_rq has been allocated via ndo_init callback and free it, because,
> > depending on the code path, register_netdevice() might not call
> > priv_destructor() callback
> >
> > Reported-and-tested-by: [email protected]
> > Link: https://syzkaller.appspot.com/bug?extid=59ef240dd8f0ed7598a8
> > Signed-off-by: Rustam Kovhaev <[email protected]>
>
> I think I agree with Toshiaki here. There is no reason why the
> rollback_registered() path of register_netdevice() should behave
> differently from the normal control flow.
>
> Any code path that invokes ->ndo_uninit() should probably also
> invoke the priv destructor.
hi David, thank you for the review!
>
> The question is why does the err_uninit: label of register_netdevice
> behave differently from rollback_registered()? If there is a reason,
> it should be documented in a comment or similar. If it is wrong,
> it should be corrected.
good question, that i do not know, i'll review it