RE: net/smc and the RDMA core

Hi Bart, Ursula, Dave,

I am particularly concerned about SMC as address family.
It should not be treated as address family, but rather an additional protocol similar for socket type SOCK_STREAM.
While doing performance benchmarking last month and while porting few database application,

I encountered a major hurdle where user space library heavily depend on AF_INET and AF_INET6 family through get_addrinfo and other friend functions.
Adding or treating AF_SMC as AF_INET just doesn't sound right.

Most user space code doesn't care for the protocol field, but do handle domain field.

I personally believe it's not too late to modify SMC to drop expose AF_SMC and have it exposed through new protocol that can be exposed through socket() API.

Parav

-----Original Message-----
From: linux-rdma-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org [mailto:linux-rdma-
owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org] On Behalf Of Bart Van Assche
Sent: Monday, May 1, 2017 12:30 PM
To: hch-jcswGhMUV9g@public.gmane.org; davem-fT/PcQaiUtIeIZ0/mPfg9Q@public.gmane.org; ubraun-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org
Cc: netdev-u79uwXL29TY76Z2rM5mHXA@public.gmane.org; linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
Subject: Re: net/smc and the RDMA core

On Mon, 2017-05-01 at 18:33 +0200, Christoph Hellwig wrote:

quoted

Hi Ursual, hi netdev reviewers,

how did the smc protocol manage to get merged without any review on
linux-rdma at all?  As the results it seems it's very substandard in
terms of RDMA API usage, e.g. it neither uses the proper CQ API nor
the RDMA R/W API, and other will probably find additional issues as
well.

Hello Dave and Ursula,

It seems very rude to me to have merged the SMC protocol driver without
having involved the linux-rdma community. Anyway, I have the following
questions for Dave and Ursula:
* Since the Linux kernel is standards based: where can we find the standard
  that defines the SMC wire protocol? If this protocol has not been
  standardized yet: in what file (other than *.[ch]) in the Linux kernel
  tree has this protocol been documented?
* What are the differences between the SMC protocol, the SDP protocol and
  the rsockets protocol? How do existing implementations for these protocols
  compare to each other from a performance point of view? If no performance
  comparison between these protocols is available, shouldn't the performance
  of these protocols have been compared with each other before a review of
  the SMC driver even started?
* What are the reasons why the SDP driver was never accepted upstream? Do
  the arguments why SDP was not accepted upstream also apply to the SMC
  driver (SDP = Sockets Direct Protocol)?
* Since SMC has to be selected by specifying AF_SMC, how are users expected
  to specify whether AF_INET, AF_INET6 or yet another address family should
  be used to set up a connection between SMC endpoints?
* Is the SMC driver limited to RoCE? Are you aware that the rsockets library
  supports multiple transport layers (RoCE, IB and iWARP)?
* Since functionality that is similar what the SMC driver provides already
  exists in user space (rsockets), why has this functionality been
  reimplemented as a kernel driver (SMC)?

Bart.--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body
of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org More majordomo info at
http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html