On Sat, May 13, 2017 at 12:52:47AM -0600, David Ahern wrote: > On 5/12/17 3:41 PM, Jan Moskyto Matejka wrote: > > On Fri, May 12, 2017 at 10:26:08AM -0700, David Ahern wrote: > >> On 5/12/17 8:24 AM, David Miller wrote: > >>> From: Jan Moskyto Matejka <m...@ucw.cz> > >>> Date: Fri, 12 May 2017 13:15:10 +0200 > >>> > >>>> -int rt6_dump_route(struct rt6_info *rt, void *p_arg); > >>>> +int rt6_dump_route(struct rt6_info *rt, void *p_arg, int truncate); > >>> > >>> Please use "bool" and "true"/"false" for boolean values. > >>> > >>> What does ipv4 do in this situation? > >>> > >>> I'm hesitant to be OK with adding a new nlmsg flag just for this case > >>> if we solve this problem differently and using existing mechanisms > >>> elsewhere. > >>> > >> > >> I'll take a look at this later today or this weekend; we can't just > >> truncate the route returned to userspace. > > > > Agreed. My favourite would be skb realloc somewhere inside the dump loop > > ... but I don't know whether it's feasible. > > > > MQ > > > > Here is what is happening: user initiates a route dump and specifies a > buffer size for receiving the message which becomes max_recvmsg_len. > This buffer size dictates the skb length allocated by netlink_dump. > > The dump is interrupted when a route does not fit in the skb and > returns. Subsequent call picks up with the next route to be dumped - the > one that overflowed. All good and normal so far. > > If the next route is larger than max_recvmsg_len, then the route can not > be put in the buffer, nothing is returned to the user which causes the > dump to abort showing the abbreviated output. This problem occurs with > IPv4 and IPv6. You can see this with modest size routes by just dropping > the buffer size to something really small (e.g., with iproute2, change > buf size in rtnl_dump_filter_l to say 2048).
Ergo no skb realloc is possible. > I see 2 problems: > 1. the kernel is not telling the user the supplied buffer is too small > (ie., if a single route does not fit in the skb then it should fail and > return an error code to the user), Definitely. I want just to note that this condition usually occurs somewhere during route dump. To know it before starting output, we would have to walk the FIB once before dump to calculate max route len. > 2. multipath routes for IPv4 and IPv6 do not have a limit. > > Should the kernel put a limit on the number of nexthops? I recently put > a cap on MPLS route size as 4096 bytes, but I think this should be > revisited in terms of a limit on number of nexthops to create a > consistent limit even if struct sizes change. And, the limit on the > number of nexthops should be consistent across address families (same > limit for IPv4, IPv6, and MPLS). > > From discussions I have had, 32 nexthops for a single route is on the > laughably high side, but some people do crazy things. How about a limit > of 256 nexthops? 256 should be OK even for a crazy developer of BIRD. It would be nice to have if the returned error were somehow useful for the userspace -- to know what is happening, not only something like "impossible to add / append route". Thanks for looking into that! MQ
