On Fri, 9 Dec 2022 at 16:00, Mark Rutland <mark.rutland@xxxxxxx> wrote: > > On Fri, Dec 09, 2022 at 03:46:48PM +0100, Ard Biesheuvel wrote: > > On Fri, 9 Dec 2022 at 15:37, Mark Rutland <mark.rutland@xxxxxxx> wrote: > > > > > > On Fri, Dec 09, 2022 at 02:34:14PM +0100, Ard Biesheuvel wrote: > > > > The EFI runtime services run from a dedicated stack now, and so the > > > > stack unwinder needs to be informed about this. > > > > > > > > Signed-off-by: Ard Biesheuvel <ardb@xxxxxxxxxx> > > > > --- > > > > > > > > I realised while looking into this that comparing current_work() against > > > > efi_rts_work.work is not sufficient to decide whether current is running > > > > EFI code, given that the ACPI subsystem will call efi_call_virt_pointer() > > > > directly. > > > > > > > > So instead, we can check whether the stashed thread stack pointer value > > > > matches current's thread stack if the EFI runtime stack is currently in > > > > use: > > > > > > > > #define current_in_efi() \ > > > > (!preemptible() && spin_is_locked(&efi_rt_lock) && \ > > > > on_task_stack(current, efi_rt_stack_top[-1], 1)) > > > > > > Unless you're overwriting task_struct::stack (which seems scary to me), that > > > doesn't look right; on_task_stack() checks whether a given base + size is on > > > the stack allocated for the task (i.e. task_struct::stack + THREAD_SIZE), not > > > the stack the task is currently using. > > > > > > > Note the [-1]. > > > > efi_rt_stack_top[-1] contains the value the stack pointer had before > > switching to the EFI runtime stack. If that value is an address > > covered by current's thread stack, current must be the task that has a > > live call frame inside the EFI code at the time the call stack is > > captured. > > Ah, I had missed that subtlety. > > Would you mind if we add that first sentence as a comment for that code, i.e. > > | /* > | * efi_rt_stack_top[-1] contains the value the stack pointer had before > | * switching to the EFI runtime stack. > | */ > | #define current_in_efi() \ > | (!preemptible() && spin_is_locked(&efi_rt_lock) && \ > | on_task_stack(current, efi_rt_stack_top[-1], 1)) > > ... that way when I look at this in 3 to 6 months time I won't fall into the > same trap. :) > Will do. > I assume that the EFI trampoline code clobbers the value on the way out so it > doesn't spruriously match later. > Not currently, no. But that's easily added. > > > I would expect this to be something like: > > > > > > #define current_in_efi() \ > > > (!preemptible() && spin_is_locked(&efi_rt_lock) && \ > > > stackinfo_on_stack(stackinfo_get_efi(), current_stack_pointer, 1)) > > > > > > ... or an inline function given this is sufficiently painful as a macro. > > > > current_stack_pointer is the actual value of SP at the time this code > > is called. So if we are unwinding from a sync exception taken while > > handling an IRQ that arrived while running the EFI code, that SP value > > has nothing to do with the EFI stack. > > Yes, good point. > > > > ... unless I've confused myself? > > > > > > > I think you might have ... :-) > > :) > > Mark.
