Re: [PATCH V2 09/12] net/eipoib: Add main driver functionality

Eric W. Biederman [off-list ref] wrote:

Or Gerlitz [off-list ref] writes:

quoted

as you wrote here, when performance and ease-of-use is under the spot,
VM deployments tend to not to use routing.
This is b/c it involves more over-head on the packet forwarding, and
more administration work, for example, for setting routing rules that
involve the VM IP address, something which AFAIK the hypervisor have
no clue on, also its unclear to me if/how live migration can work in such setting.

All you need to make proxy-arp essentially pain free is a smart dhcp
relay, that sets up the routes.

So dhcp relay would help to maybe avoid some of the pain, however,
can you elaborate if/how live migration is supported under this scheme?

quoted

From this exact reason, there's a bunch of use-cases by tools and
cloud stacks (such as open stack, ovirt, more) which do use bridged
mode and the rest of the Ethernet envelope, such as using virtual L2
vlan domains, ebtables based rules, etc etc. Where they and are not
application to ipoib, but are working file ith eipoib.

Yes I am certain all of their IPv6 traffic works fine.

I'm not sure to follow this comment.

Regardless those are open source projects and can be modified to add
support to cleanly support inifiniband.

open source/code can be modified indeed, but since exposing IPoIB link
layer to tools/emulators and VMs doesn't really make sense (see below),
we brought that this approach as a way to go for allowing people to use
bridging mode when the fabric is IB.

quoted

You mentioned that bridging mode doesn't apply to environment such as
802.11, and hence routing mode is used, we are trying to make a pointn
here that bridging mode applies to ipoib with the approach suggested by eipoib.

You are completely failing.  Every time I look I see something about
eIPoIB that is even more broken.  Given that eIPoIB is a NAT
implementation that isn't really a surprise but still.

eIPoIB imposes enough overhead that I expect that routing is cheaper,
so your performance advantges go right out the window.

eIPoIB is seriously incompatible with ethernet breaking almost
everything and barely allowing IPv4 to work.

I don't agree on this incompatiblity statement, you had a claim
on DHCP and I addressed it, beyond that, you don't like the eIPoIB
basic idea/design but this can't base an incompatiblity argument.

quoted

Also, if we extend the discussion a bit, there are two more aspects to throw in:
The first is the performance thing we have already started to mention
-- specifically, the approach for RX zero copy (into the VM buffer),
use designs such as vhost + macvtap NIC in passthrough mode which is
likey to be set over a per VM hypervisor NIC, e.g such as the ones
provided by VMDQ patches John Fastabend started to post (see
http://marc.info/?l=linux-netdev&m=134264998405581&w=2) -- the ib0.N
clone child are IPoIB VMDQ NICs if you like, and setting an eipoib NIC
on top of each they can be plugged to that design.

If you care about performance link-layer NAT is not the way to go.
Teach the pieces you care about how to talk infiniband.

quoted

The 2nd aspect, is NON VM environments where a NIC with Ethernet look
and feel is required for IP traffic, but this have to live within an
echo-system that fully uses IPoIB.
In other words, a use case where IPoIB has to be below the cover for
set of some specific apps, or nodes but do IP interaction with other
apps/nodes and gateways who use IPoIB, the eIPoIB driver provides that
functionality.

ip link add type dummy.

There now you have an interface with ethernet look and feel, and
routing can happily avoid it.

again, not sure to follow, you mean "routing can happily use it", correct? that
is do routing between the dummy interface to IPoIB interface?

quoted

So, to sum up, routing / proxy-arp seems to be off where we are targeting.

My condolences.
The existence of router / proxy-arp means that solutions do exist
(unlike your previous claim) you just don't like the idea of deploying them.

Don't like them from set of arguments, which we are covering here, re
manageability it still needs to clarified if/how live migration work and what
does it mean to always mandate dhcp relay.

Infiniband is standard enough you could quite easily implement virtual
infiniband bridging as an alternative to ethernet bridging.

Not really, as Michael indicated in his response over this thread
http://marc.info/?l=linux-netdev&m=134419288218373&w=2
IPoIB link layer addresses use IB HW constructs for which soft
hardware address setting isn't supported, and this interferes
with live migration.

Or.