On 21/05/2021 12:37, Will Deacon wrote: > On Fri, May 21, 2021 at 10:39:32AM +0200, Juri Lelli wrote: >> On 21/05/21 08:15, Quentin Perret wrote: >>> On Friday 21 May 2021 at 07:25:51 (+0200), Juri Lelli wrote: >>>> On 20/05/21 19:01, Will Deacon wrote: >>>>> On Thu, May 20, 2021 at 02:38:55PM +0200, Daniel Bristot de Oliveira wrote: >>>>>> On 5/20/21 12:33 PM, Quentin Perret wrote: >>>>>>> On Thursday 20 May 2021 at 11:16:41 (+0100), Will Deacon wrote: >>>>>>>> Ok, thanks for the insight. In which case, I'll go with what we discussed: >>>>>>>> require admission control to be disabled for sched_setattr() but allow >>>>>>>> execve() to a 32-bit task from a 64-bit deadline task with a warning (this >>>>>>>> is probably similar to CPU hotplug?). >>>>>>> >>>>>>> Still not sure that we can let execve go through ... It will break AC >>>>>>> all the same, so it should probably fail as well if AC is on IMO >>>>>>> >>>>>> >>>>>> If the cpumask of the 32-bit task is != of the 64-bit task that is executing it, >>>>>> the admission control needs to be re-executed, and it could fail. So I see this >>>>>> operation equivalent to sched_setaffinity(). This will likely be true for future >>>>>> schedulers that will allow arbitrary affinities (AC should run on affinity >>>>>> change, and could fail). >>>>>> >>>>>> I would vote with Juri: "I'd go with fail hard if AC is on, let it >>>>>> pass if AC is off (supposedly the user knows what to do)," (also hope nobody >>>>>> complains until we add better support for affinity, and use this as a motivation >>>>>> to get back on this front). >>>>> >>>>> I can have a go at implementing it, but I don't think it's a great solution >>>>> and here's why: >>>>> >>>>> Failing an execve() is _very_ likely to be fatal to the application. It's >>>>> also very likely that the task calling execve() doesn't know whether the >>>>> program it's trying to execute is 32-bit or not. Consequently, if we go >>>>> with failing execve() then all that will happen is that people will disable >>>>> admission control altogether. >>> >>> Right, but only on these dumb 32bit asymmetric systems, and only if we >>> care about running 32bits deadline tasks -- which I seriously doubt for >>> the Android use-case. >>> >>> Note that running deadline tasks is also a privileged operation, it >>> can't be done by random apps. >>> >>>>> That has a negative impact on "pure" 64-bit >>>>> applications and so I think we end up with the tail wagging the dog because >>>>> admission control will be disabled for everybody just because there is a >>>>> handful of 32-bit programs which may get executed. I understand that it >>>>> also means that RT throttling would be disabled. >>>> >>>> Completely understand your perplexity. But how can the kernel still give >>>> guarantees to "pure" 64-bit applications if there are 32-bit >>>> applications around that essentially broke admission control when they >>>> were restricted to a subset of cores? >>>> >>>>> Allowing the execve() to continue with a warning is very similar to the >>>>> case in which all the 64-bit CPUs are hot-unplugged at the point of >>>>> execve(), and this is much closer to the illusion that this patch series >>>>> intends to provide. >>>> >>>> So, for hotplug we currently have a check that would make hotplug >>>> operations fail if removing a CPU would mean not enough bandwidth to run >>>> the currently admitted set of DEADLINE tasks. >>> >>> Aha, wasn't aware. Any pointers to that check for my education? >> >> Hotplug ends up calling dl_cpu_busy() (after the cpu being hotplugged out >> got removed), IIRC. So, if that fails the operation in undone. > > Interesting, thanks. Thinking about this some more, it strikes me that with > these silly asymmetric systems there could be an interesting additional > problem with hotplug and deadline tasks. Imagine the following sequence of > events: > > 1. All online CPUs are 32-bit-capable > 2. sched_setattr() admits a 32-bit deadline task > 3. A 64-bit-only CPU is onlined > 4. Some of the 32-bit-capable CPUs are offlined > > I wonder if we can get into a situation where we think we have enough > bandwidth available, but in reality the 32-bit task is in trouble because > it can't make use of the 64-bit-only CPU. > > If so, then it seems to me that admission control is really just > "best-effort" for 32-bit deadline tasks on these systems because it's based > on a snapshot in time of the available resources. IMHO DL AC is per root domain (rd). So if we have e.g. an 8 CPU system with aarch32_el0 eq. [0-3] then we would need 2 exclusive cpusets ([0-3] and [4-7]) to admit 32-bit DL tasks into [0-3] (i.e. to pass the `if (!cpumask_subset(span, p->cpus_ptr) ...` test in __sched_setscheduler(). Trying to admit too many 32-bit DL tasks or trying to hp out a CPU[0-3] would lead to `Device or resource busy` in case the rd bw wouldn't be sufficient anymore for the set of admitted tasks. But the [0-3] DL AC wouldn't care about hp on CPU[4-7].
