On 08/17/2017 03:54 PM, Oliver Hartkopp wrote:
> On 08/17/2017 03:17 PM, Marc Kleine-Budde wrote:
> 
>> The code gets complicated, as during NETDEV_REGISTER the malloc may
>> fail. But we can cancel the register with return
>> notifier_from_errno(-ENOMEM);
> 
> So this would lead to a netdevice completely failing to register then?
> Nice!
> 
>>> With NETDEV_REGISTER the memory would be allocated. Ok.
>>>
>>> An how would the memory be free'd then?
>>
>> NETDEV_UNREGISTER... But we have a problem, as j1939 needs per device
>> memory, too. Which is kfree()ed in a NETDEV_UNREGISTER hook, too. The
>> order of the hooks in not guaranteed.
> 
> Yes. That's the same problem I was working around with 
> 'd->remove_on_zero_entries' %-/

Using proper refcounting ...but it gets quite complicatd with j1939.

> Just to get an impression about the j1939 data structures:
> 
> Would the per-device memory increase significantly when CAN_J1939 is 
> enabled? Or is it some array of pointers which would be added?

An array with 256 members

> 	struct addr_ent {
> 		ktime_t rxtime;
> 		struct j1939_ecu *ecu;
> 		/* count users, to help transport protocol */
> 		int nusers;
> 	} ents[256];

and some more stuff. Sums up to sizeof(priv)=4184 Byte on armv5 (32bit).

Marc

-- 
Pengutronix e.K.                  | Marc Kleine-Budde           |
Industrial Linux Solutions        | Phone: +49-231-2826-924     |
Vertretung West/Dortmund          | Fax:   +49-5121-206917-5555 |
Amtsgericht Hildesheim, HRA 2686  | http://www.pengutronix.de   |