On Thu, Aug 09, 2018 at 01:38:12PM +0100, Will Deacon wrote: > On Thu, Aug 09, 2018 at 01:34:57PM +0100, Dave Martin wrote: > > On Wed, Aug 08, 2018 at 01:34:09PM +0100, Catalin Marinas wrote: > > > On Tue, Aug 07, 2018 at 11:24:34AM +0100, Marc Zyngier wrote: > > > > On 07/08/18 11:05, Dave Martin wrote: > > > > > On Tue, Aug 07, 2018 at 10:33:26AM +0100, Marc Zyngier wrote: > > > > >> It recently came to light that userspace can execute WFI, and that > > > > >> the arm64 kernel doesn trap this event. This sounds rather benign, > > > > > > Nitpick: "doesn't". > > > > > > > >> but the kernel should decide when it wants to wait for an interrupt, > > > > >> and not userspace. > > > > >> > > > > >> Let's trap WFI and treat it as a way to yield the CPU to another > > > > >> process. > > > [...] > > > > > I can't think of a legitimate reason for userspace to execute WFI > > > > > however. Userspace doesn't have interrupts under Linux, so it makes > > > > > no sense to wait for one. > > > > > > > > > > Have we seen anybody using WFI in userspace? It may be cleaner to > > > > > map this to SIGILL rather than be permissive and regret it later. > > > > > > > > I couldn't find any user, and I'm happy to just send userspace to hell > > > > in that case. But it could also been said that since it was never > > > > prevented, it is a de-facto ABI. > > > > > > I wouldn't really go as far as SIGILL on WFI. I think the patch is fine > > > as it is. In case Will plans to merge it: > > > > For practical purposes I agree, because we can't control the binary > > blobs out there: I just wanted to bang the drum because we are creating > > semantics here and there is not an obvious correct answer to what they > > should be. > > > > I'd still like to see rationale for why this should map to schedule() > > (which userspace currently has no direct way to trigger) as opposed to > > sched_yield() or something like that. > > A better idea might just be to do pc +=4 and return. If there's work > pending, we'll hit it on the return path (just like any other ret_to_user > call). > > I initially thought about sched_yield(), but it's not clear whether that > creates a problem if, e.g. seccomp has been used to restrict that syscall.
Indeed. I can't see why that might be restricted, but there's presumably nothing to stop people doing that today. Other than putting the task to sleep for 1ms or something, I don't know what to suggest ;) Perhaps we can patch a NOP into .text, like Marc's BX trick :P Cheers ---Dave
