On Sat, Feb 08, 2020 at 11:25:18AM +0100, Ulf Hansson wrote: > On Fri, 7 Feb 2020 at 17:15, Sudeep Holla <sudeep.holla@xxxxxxx> wrote: > > > > On Fri, Feb 07, 2020 at 04:52:52PM +0100, Ulf Hansson wrote: > > > On Fri, 7 Feb 2020 at 15:48, Lorenzo Pieralisi > > > <lorenzo.pieralisi@xxxxxxx> wrote: > > > > > > > > On Fri, Feb 07, 2020 at 01:32:28PM +0100, Ulf Hansson wrote: > > > > > [...] > > > > > > > > > > > > I understand the arguments for using PC vs OSI and agree with it. But > > > > > > > what in PSCI is against Linux knowing when the last core is powering > > > > > > > down when the PSCI is configured to do only Platform Cordinated. > > > > > > > > > > > > Nothing :D. But knowing the evolution and reasons for adding OSI in the > > > > > > PSCI specification and having argued about benefits of OSI over PC for > > > > > > years and finally when we have it in mainline, this argument of using > > > > > > PC for exact reasons why OSI evolved is something I can't understand > > > > > > and I am confused. > > > > > > > > > > > > > There should not be any objection to drivers knowing when all the cores > > > > > > > are powered down, be it reference counting CPU PM notifications or using > > > > > > > a cleaner approach like this where GendPD framwork does everything > > > > > > > cleanly and gives a nice callback. ARM architecture allows for different > > > > > > > aspects of CPU access be handled at different levels. I see this as an > > > > > > > extension of that approach. > > > > > > > > > > > > > > > > > > > One thing that was repeatedly pointed out during OSI patch review was no > > > > > > extra overhead for PC mode where firmware can make decisions. So, just > > > > > > use OSI now and let us be done with this discussion of OSI vs PC. If PC > > > > > > is what you think you need for future, we can revert all OSI changes and > > > > > > start discussing again :-) > > > > > > > > > > Just to make it clear, I fully agree with you in regards to overhead > > > > > for PC-mode. This is especially critical for ARM SoCs with lots of > > > > > cores, I assume. > > > > > > > > > > However, the overhead you refer to, is *only* going to be present in > > > > > case when the DTS has the hierarchical CPU topology description with > > > > > "power-domains". Because, that is *optional* to use, I am expecting > > > > > only those SoC/platforms that needs to manage last-man activities to > > > > > use this layout, the others will remain unaffected. > > > > > > > > In PC mode not only there is no need but it is wrong to manage > > > > any last-man activity in the kernel. I wonder why we are still > > > > talking about this to be honest. > > > > > > I guess the discussion is here because there is a use case to consider now. > > > > > > > If this is what Bjorn presented in his email, I have responded to that. > > If it's any different, please let us know the complete details. > > > > > For sure, we agree on what is the best solution. But this is rather > > > about what can we do to improve the current situation, if we should do > > > anything. > > > > > > > Sure, and I haven't found a reason to do that in OSPM yet(as part of the > > discussion in this thread) > > > > > > > > > > Code to handle PSCI platform coordinated mode has been/is in > > > > the kernel today and that's all is needed according to the PSCI > > > > specifications. > > > > > > PSCI specifies CPU power management, not SoC power management. If > > > these things were completely decoupled, I would agree with you, but > > > that's not the case. Maybe SCMI, etc, helps with this in future. > > > > > > > Why does that not work even if they are not decoupled. The IO/device > > that share with CPU votes from OSPM and the CPU/Cluster from PSCI in > > PC mode. There is no argument there, but why it needs to be done in OSPM > > is the objection here. > > That implies the votes from I/O devices needs to reach the FW > immediately when the vote is done. No caching or other optimizations > can be done at OSPM. > > In principle, the FW needs to have an always up to date view of the > votes, etc. That sounds highly inefficient, both from energy and > latency point of view, at least in my opinion. > Sorry but I need to re-iterate, use OSI if you need all those fancy caching and other optimizations. > > > > > Anyway, my fear is that not many ARM vendors implements OSI support, > > > but still they have "last-man-activities" to deal with. This is not > > > only QCOM. > > > > > > > I am interested to hear from them. And the same question to same too as > > above. > > I have been talking to some of them. But, yes, we need to hear more from them. > > > > > > I guess an option would be to add OSI support to the public ARM > > > Trusted Firmware, then we could more easily point to that - rather > > > than trying to mitigate the problem on the kernel side. > > > > > > > I would say go for it. But don't mix responsibility of OSPM in PC vs OSI. > > We have discussed this for years and I hope this discussion ends ASAP. > > I don't see any point in dragging this any further. > > Okay. > I keep saying that but still responding to the discussions. I must stop ;-) -- Regards, Sudeep
