[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 04 Apr 2024 17:24:08 +0200
From: Niklas Schnelle <schnelle@...ux.ibm.com>
To: Wen Gu <guwen@...ux.alibaba.com>, Gerd Bayer <gbayer@...ux.ibm.com>,
wintera@...ux.ibm.com, twinkler@...ux.ibm.com, hca@...ux.ibm.com,
gor@...ux.ibm.com, agordeev@...ux.ibm.com, davem@...emloft.net,
edumazet@...gle.com, kuba@...nel.org, pabeni@...hat.com,
wenjia@...ux.ibm.com, jaka@...ux.ibm.com
Cc: borntraeger@...ux.ibm.com, svens@...ux.ibm.com, alibuda@...ux.alibaba.com,
tonylu@...ux.alibaba.com, linux-kernel@...r.kernel.org,
linux-s390@...r.kernel.org, netdev@...r.kernel.org
Subject: Re: [RFC PATCH net-next v5 05/11] net/smc: implement DMB-related
operations of loopback-ism
On Thu, 2024-04-04 at 21:44 +0800, Wen Gu wrote:
>
> On 2024/4/4 19:27, Niklas Schnelle wrote:
> > On Thu, 2024-04-04 at 18:20 +0800, Wen Gu wrote:
> > >
> > > On 2024/4/4 01:20, Gerd Bayer wrote:
> > > > On Sun, 2024-03-24 at 21:55 +0800, Wen Gu wrote:
> > > >
> > > > When I instrumented this to see, why I still see tons of my other
> > > > temporary instrumentation messages from the "ism" driver, I found that
> > > > in my setup loopback-ism is used rather infrequently.
> > > >
> > > > I suspect this is due to how the SMC proposals are constructed in
> > > > net/smc/af_smc.c and net/smc/smc_pnet.c - and later evaluated in
> > > > smc_check_ism_v2_match() - where there is a first-come-first-serve
> > > > selection.
> > > >
> > > > I wonder if one should change that to favour loopback-ism over "real"
> > > > ISM devices - and how this could be achieved elegantly.
> > > >
> > > > Just some food for thought... Probably little you can do on x86.
> > > >
> > >
> > > Yes, it is about the priority of available ISM devices, and now it
> > > is decided by their order in the smcd_dev_list. The later registered
> > > ISMv2 devices(without pnetid) will be added to the beginning of the
> > > list (see smcd_register_dev()). So there is a probability that
> > > loopback-ism will not be ranked first, since it is added into list
> > > earlier during smc_init().
> > >
> > > If we have the runtime switch of loopback-ism, we can re-active the
> > > loopback-ism, that make it be re-added into the beginning of the dev
> > > list and be chosen first. Or a new netlink command to adjust the slot
> > > order of available ISM devices in the list. As we discussed before,
> > > that could be tasks in stage 1 or stage 2.
> > >
> > > Thanks!
> >
> > Maybe when adding the ISM devices we could instead make sure that all
> > ISM devices are added after loopback and loopback is added in the
> > beginning. I think loopback should always be preferred and would
> > consider it a bug if it isn't faster too. Between virtio-ism and ISM it
> > may be less clear so maybe for stage 2 we would want a priority setting
> > and then insert ordered by priority. Thoughts?
> I have no objection. If we all agree, I will keep it at the beginning of the list.
>
> Thanks!
>
>
I think this is a decision that needs to be made by the SMC maintainers
and Wenjia will be back next week. That said the current code basically
prioritizes ISM devices without that having been a conscious and
documented decision. Also note that due to hotplug an LPAR could be
using loopback-ism happily until suddenly an ISM device is hot plugged
and new connections suddenly switch to ISM.
Powered by blists - more mailing lists