On Fri, Oct 09, 2020 at 09:31:47AM +0100, Will Deacon wrote: > On Fri, Oct 09, 2020 at 10:13:12AM +0200, Morten Rasmussen wrote: > > On Fri, Oct 09, 2020 at 09:29:43AM +0200, Peter Zijlstra wrote: > > > On Thu, Oct 08, 2020 at 07:16:41PM +0100, Qais Yousef wrote: > > > > diff --git a/arch/arm64/kernel/signal.c b/arch/arm64/kernel/signal.c > > > > index cf94cc248fbe..7e97f1589f33 100644 > > > > --- a/arch/arm64/kernel/signal.c > > > > +++ b/arch/arm64/kernel/signal.c > > > > @@ -908,13 +908,28 @@ static void do_signal(struct pt_regs *regs) > > > > restore_saved_sigmask(); > > > > } > > > > > > > > +static void set_32bit_cpus_allowed(void) > > > > { > > > > + cpumask_var_t cpus_allowed; > > > > + int ret = 0; > > > > + > > > > + if (cpumask_subset(current->cpus_ptr, &aarch32_el0_mask)) > > > > + return; > > > > + > > > > /* > > > > + * On asym aarch32 systems, if the task has invalid cpus in its mask, > > > > + * we try to fix it by removing the invalid ones. > > > > */ > > > > + if (!alloc_cpumask_var(&cpus_allowed, GFP_ATOMIC)) { > > > > + ret = -ENOMEM; > > > > + } else { > > > > + cpumask_and(cpus_allowed, current->cpus_ptr, &aarch32_el0_mask); > > > > + ret = set_cpus_allowed_ptr(current, cpus_allowed); > > > > + free_cpumask_var(cpus_allowed); > > > > + } > > > > + > > > > + if (ret) { > > > > + pr_warn_once("Failed to fixup affinity of running 32-bit task\n"); > > > > force_sig(SIGKILL); > > > > } > > > > } > > > > > > Yeah, no. Not going to happen. > > > > > > Fundamentally, you're not supposed to change the userspace provided > > > affinity mask. If we want to do something like this, we'll have to teach > > > the scheduler about this second mask such that it can compute an > > > effective mask as the intersection between the 'feature' and user mask. > > > > I agree that we shouldn't mess wit the user-space mask directly. Would it > > be unthinkable to go down the route of maintaining a new mask which is > > the intersection of the feature mask (controlled and updated by arch > > code) and the user-space mask? > > > > It shouldn't add overhead in the scheduler as it would use the > > intersection mask instead of the user-space mask, the main complexity > > would be around making sure the intersection mask is updated correctly > > (cpusets, hotplug, ...). > > > > Like the above tweak, this won't help if the intersection mask is empty, > > task will still get killed but it will allow tasks to survive > > user-space masks including some non-compatible CPUs. If we want to > > prevent task killing in all cases (ignoring hotplug) it gets more ugly > > as we would have to ignore the user-space mask in some cases. > > Honestly, I don't understand why we're trying to hide this asymmetry from > userspace by playing games with affinity masks in the kernel. Userspace > is likely to want to move things about _anyway_ because even amongst the > 32-bit capable cores, you may well have different clock frequencies to > contend with. > > So I'd be *much* happier to let the schesduler do its thing, and if one > of these 32-bit tasks ends up on a core that can't deal with it, then > tough, it gets killed. Give userspace the information it needs to avoid > that happening in the first place, rather than implicitly limit the mask. > > That way, the kernel support really boils down to two parts: > > 1. Remove the sanity checks we have to prevent 32-bit applications running > on asymmetric systems > > 2. Tell userspace about the problem This works for me as well as long as it is default off with a knob to turn it on. I'd prefer a sysctl (which can be driven from the command line in recent kernels IIRC) so that one can play with it a run-time. This way it's also a userspace choice and not an admin or whoever controls the cmdline (well, that's rather theoretical since the target is Android). -- Catalin