> From: linux-rdma-ow...@vger.kernel.org [mailto:linux-rdma- > ow...@vger.kernel.org] On Behalf Of Doug Ledford
> These patches add the concept of duplicate GIDs that are differentiated by > their RoCE version (also called network type). So, now, an incoming packet > could match a couple different gid_indexes and we need additional > information to get back to the definitive 1:1 mapping. The submitted patches > are designed around a lazy resolution of the namespace, preferring to defer > the work of mapping the incoming packet to a unique namespace until that > information is actually needed. To enable this lazy resolution, it provides > the > network_type so that the resolution can be done. > > This is a fair assessment of the current state of things and what these > patches do, yes? Hi Doug, Yes, your description nails the crux of the discussion. Thanks. > > Jason's objections are this: > > 1) The lazy resolution is wrong. > 2) The use of network_type as the additional information to get to the > unique namespace is vendor specific cruft that shouldn't be part of the core > kernel API. > Which is totally wrong. RoCE versions and the network type attribute are part of the RoCE specification. There is nothing vendor specific. In fact, this patch initially came from Avago!!! > Jason's preference would be that the above issues be resolved by skipping > the lazy resolution and instead doing proactive resolution on receipt of a > packet and then probably just pass the namespace around instead of passing > around the information needed to resolve the namespace. Or, at a > minimum, at least make the information added to the core API not > something vendor specific like network_type, which is a detail of the > Mellanox implementation. Again, there is nothing here specific to the Mellanox implementation. This argument is invalid. > > Jason, is this accurate for your position? > > If everyone agrees that this is a fair statement of where we stand, then I'll > continue my response. If not, please correct anything I have wrong above > and I'll take that into my continued response. > Since we are following the standard, and: - This is *not* a vendor specific issue - Lazy resolution adheres to the current flow of the stack (which IMO, is the right thing to do anyway) - There are no security issues whatsoever (see below) - The code could be easily extended to support namespaces in the future - We should "release early release often", and not hold progress because of endless philosophical discussions then let's apply these patches. I perfectly understand Jason's proposal and its merits. I just don't agree that it is a better approach, and in any case this is not the time to continue discussing it. > 1 - Actually, for any received packet with associated IP address information. > We've only enabled net namespaces for IP connections between user space > applications, for direct IB connections or for kernel connections there is not > yet any namespace support. Correct. That's why we are not risking anything in the KAPI with regard to namespaces; there is no security issue. --Liran > > -- > Doug Ledford <dledf...@redhat.com> > GPG KeyID: 0E572FDD > N�����r��y����b�X��ǧv�^�){.n�+����{��ٚ�{ay�ʇڙ�,j��f���h���z��w��� ���j:+v���w�j�m��������zZ+�����ݢj"��!�i