On 06/15/2017 06:00 AM, Nikolay Aleksandrov wrote: > On 15/06/17 14:44, Nikolay Aleksandrov wrote: >> On 15/06/17 14:33, Nikolay Aleksandrov wrote: >>> On 15/06/17 00:51, Julien Gomes wrote: >>>> Hi Nikolay, >>>> >>>> On 06/14/2017 05:04 AM, Nikolay Aleksandrov wrote: >>>> >>>>> This has been on our todo list and I'm definitely interested in the >>>>> implementation. >>>>> A few things that need careful consideration from my POV. First are the >>>>> security >>>>> implications - this sends rtnl multicast messages but the rtnl socket has >>>>> the NL_CFG_F_NONROOT_RECV flag thus allowing any user on the system to >>>>> listen in. >>>>> This would allow them to see the full packets and all reports (granted >>>>> they can see >>>>> the notifications even now), but the full packet is like giving them the >>>>> opportunity >>>>> to tcpdump the PIM traffic. >>>> I definitely see how this can be an issue. >>>> From what I see, this means that either the packet should be >>>> transmitted another way, or another Netlink family should be used. >>>> >>>> NETLINK_ROUTE looks to be the logical family to choose though, >>>> but then I do not see a proper other way to handle this. >>> Right, currently me neither, unless it provides a bind callback when >>> registering >>> the kernel socket. >>> >>>> However I may just not be looking into the right direction, >>>> maybe you currently have another approach in mind? >>> I haven't gotten around to make (or even try) them but I was thinking about >>> 2 options >>> ending up with a similar result: >>> >>> 1) genetlink >>> It also has the NONROOT_RECV flag, but it also allows for a callback - >>> mcast_bind() >>> which can be used to filter. >>> >>> or >>> >>> 2) Providing a bind callback to the NETLINK_ROUTE socket. >>> >> Ah nevermind, these cannot be used for filtering currently, so it seems >> the netlink interface would need to be extended too if going down this road. >> > Sorry for the multiple emails, just to be thorough - again if going down this > road all of these would obviously require a different group to bind to in > order > to be able to filter on it, because users must keep receiving their > notifications > for the ipmr one.
Actually, using a bind callback for NETLINK_ROUTE with a new group, without netlink interface extension, could work. I quickly tested something like this: > static int rtnetlink_bind(struct net *net, int group) > { > switch (group) { > case RTNLGRP_IPV4_MROUTE_R: > case RTNLGRP_IPV6_MROUTE_R: > if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) > return -EPERM; > break; > } > return 0; > } With the addition of one/two groups this does restrict the reports' potential listeners. The group names here are just placeholders, I am not especially fixed on these ones. It is not perfect as this would introduce groups with specific requirements in NETLINK_ROUTE, but I think it can be decent. What do you think about this? -- Julien Gomes