On Fri, Oct 30, 2020 at 01:29:56PM +0530, Sai Prakash Ranjan wrote: > Hello guys, > > On 2020-10-24 02:07, Mathieu Poirier wrote: > > On Fri, Oct 23, 2020 at 03:44:16PM +0200, Peter Zijlstra wrote: > > > On Fri, Oct 23, 2020 at 02:29:54PM +0100, Suzuki Poulose wrote: > > > > On 10/23/20 2:16 PM, Peter Zijlstra wrote: > > > > > On Fri, Oct 23, 2020 at 01:56:47PM +0100, Suzuki Poulose wrote: > > > > > > > > > That way another session could use the same sink if it is free. i.e > > > > > > > > > > > > perf record -e cs_etm/@sink0/u --per-thread app1 > > > > > > > > > > > > and > > > > > > > > > > > > perf record -e cs_etm/@sink0/u --per-thread app2 > > > > > > > > > > > > both can work as long as the sink is not used by the other session. > > > > > > > > > > Like said above, if sink is shared between CPUs, that's going to be a > > > > > trainwreck :/ Why do you want that? > > > > > > > > That ship has sailed. That is how the current generation of systems are, > > > > unfortunately. But as I said, this is changing and there are guidelines > > > > in place to avoid these kind of topologies. With the future > > > > technologies, this will be completely gone. > > > > > > I understand that the hardware is like that, but why do you want to > > > support this insanity in software? > > > > > > If you only allow a single sink user (group) at the same time, your > > > problem goes away. Simply disallow the above scenario, do not allow > > > concurrent sink users if sinks are shared like this. > > > > > > Have the perf-record of app2 above fail because the sink is in-user > > > already. > > > > I agree with you that --per-thread scenarios are easy to deal with, but > > to > > support cpu-wide scenarios events must share a sink (because there is > > one event > > per CPU). CPU-wide support can't be removed because it has been around > > for close to a couple of years and heavily used. I also think using the > > pid of > > the process that created the events, i.e perf, is a good idea. We just > > need to > > agree on how to gain access to it. > > > > In Sai's patch you objected to the following: > > > > > + struct task_struct *task = READ_ONCE(event->owner); > > > + > > > + if (!task || is_kernel_event(event)) > > > > Would it be better to use task_nr_pid(current) instead of event->owner? > > The end > > result will be exactly the same. There is also no need to check the > > validity of > > @current since it is a user process. > > > > We have devices deployed where these crashes are seen consistently, > so for some immediate relief, could we atleast get some fix in this > cycle without major design overhaul which would likely take more time. > Perhaps my first patch [1] without any check for owner or > I can post a new version as Suzuki suggested [2] dropping the export > of is_kernel_event(). Then we can always work on top of it based on the > conclusion of this discussion, we will atleast not have the systems > crash in the meantime, thoughts? For the time being I think [1], exactly the way it is, is a reasonable way forward. Regards, Mathieu > > [1] https://lore.kernel.org/patchwork/patch/1318098/ > [2] > https://lore.kernel.org/lkml/fa6cdf34-88a0-1050-b9ea-556d0a9438cb@xxxxxxx/ > > Thanks, > Sai > > -- > QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member > of Code Aurora Forum, hosted by The Linux Foundation
