> -----Original Message----- > From: Jason Gunthorpe [mailto:jgg@xxxxxxxx] > Sent: Thursday, April 05, 2018 8:41 PM > To: Parav Pandit <parav@xxxxxxxxxxxx> > Cc: Leon Romanovsky <leon@xxxxxxxxxx>; Doug Ledford > <dledford@xxxxxxxxxx>; Leon Romanovsky <leonro@xxxxxxxxxxxx>; RDMA > mailing list <linux-rdma@xxxxxxxxxxxxxxx>; Mark Bloch <markb@xxxxxxxxxxxx> > Subject: Re: [PATCH rdma-next v1 0/7] GID refactoring cont 3. > > On Thu, Apr 05, 2018 at 10:32:45PM +0000, Parav Pandit wrote: > > > This series broke two scenarios as side effect described below. > > I am not sure how many users actually do such configuration. > > I am afraid that I cannot have quick fix in next 1 days (attending OFA). > > So I request to revert the series and subsequent qedr commit until I resolve it. > > I think it is possibly too big to revert.. > > Your discussion makes me think these situations are very fringe and not likely to > cause real problems. Meaning fixing them in the next two weeks time frame > should be OK. > > > Scenarios: > > 1. when IPv6 link local address is removed, its associated look alike > > GID is default GID, which gets removed. (Which is actually correct at > > least for RoCEv2 GID because we have to use IPv6 stack for resolving > > destination GID). When IPv6 same address is added back, GIDs are > > added as non-default GIDs. > > Realistically the ipv6 bound to the MAC of the device does not generally get > added/removed.. > > > I don't want to allow RoCEv2 IPv6 look-alike GID to exist when IPv6 > > address is removed, but code requires some changes to achieve this. > > I have a patch, when IPv6 addresses are added back, it adds those > > lookalike GID to default GID location. Let me know if you think that > > it is acceptable fixup patch. > > And even using a 'default gid' in either version of roce seems pretty fringe as the > CM doesn't do that. > > > 2. When primary netdev mac address is changed, GUID and default GID > > used to get updated because update_gid() silently used to update the > > GID entry without del_gid(), add_gid() combination (even though > > roce_gid_mgmt.c) is expected to do so in > > bond_delete_netdev_default_gids(), but due to bug in that function of > > comparison, it doesn't. This affects hns and usnic driver. Mlx4,5 > > currently are not expected to work with changing mac address. > > Changing MAC address is also a fairly rare situation.. > > So I think I'd prefer to proceed and expect you to fix this for the > rc1 time period, considering the OFA conference and all. > Ok. I was still aiming to avoid regression in this series. So made some progress without incurring massive code change in last few hours. Seems to solve the issues. I will get it reviewed and run some tests and come through Leon's queue. > Jason -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
