> -----Original Message----- > From: Or Gerlitz [mailto:gerlitz...@gmail.com] > Sent: Monday, April 20, 2015 9:38 PM > > On Mon, Apr 20, 2015 at 7:41 PM, Jason Gunthorpe > <jguntho...@obsidianresearch.com> wrote: > > On Mon, Apr 20, 2015 at 12:03:32PM +0300, Haggai Eran wrote: > >> From: Yotam Kenneth <yota...@mellanox.com> > >> > >> When accepting a new connection with the listener being IPv6, the > >> family of the new connection is set as IPv6. This causes > cma_zero_addr > >> function to return true on an non-zero address. As a result, the > wrong > >> code path is taken. This causes the connection request to be > rejected, > >> as the RDMA-CM code looks for the wrong type of device. > > > > This description doesn't really make sense as to what the problem is. > > > >> @@ -866,12 +866,12 @@ static void cma_save_ip4_info(struct rdma_cm_id > *id, struct rdma_cm_id *listen_i > >> > >> listen4 = (struct sockaddr_in *) &listen_id- > >route.addr.src_addr; > >> ip4 = (struct sockaddr_in *) &id->route.addr.src_addr; > >> - ip4->sin_family = listen4->sin_family; > >> + ip4->sin_family = AF_INET; > > > > If listen_id->route.addr.src_addr.ss_family != AF_INET then it is > > invalid to cast to sockaddr_in. > > > > So listen4->sin_family MUST be AF_INET or this function MUST NOT be > > called. > > > > Forcing to AF_INET cannot be correct here. > > Jason, could you take a look @ this thread > http://marc.info/?t=141589395000004&r=1&w=2 where the authors > addressed some comments from Sean and he eventually Acked the patch? > > > What does this patch have to do with this series? > > I believe it's either a pre-patch to address some assumption or > something they stepped on while testing >
We stepped upon this issue while testing the containers support we are Submitting here. When creating a new network namespace, the kernel set net->ipv6.sysctl.bindv6only to 0. As a result, we got the IPv6 listening ID accepting IPv4 connection. This is fixed by the above patch. Thanks, --Shachar
