Re: [PATCH 2/2] x86/asm: Fix inline asm call constraints for clang

[Dmitry Vyukov]

On Thu, Sep 21, 2017 at 1:52 PM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
>>>>>>> I think we need just the frame itself and RSP pointing below this
>>>>>>> frame. If we don't have a frame, CALL instruction will smash whatever
>>>>>>> RSP happens to point to. Compiler doesn't have to setup RSP to point
>>>>>>> below used part of stack in leaf functions.
>>>>>>>
>>>>>>
>>>>>> In the kernel it does.  Redzoning is not allowed in the kernel, because
>>>>>> interrupts or exceptions would also smash the redzone.
>>>>>
>>>>> I see... But it's the same for user-space signals, the first thing a
>>>>> signal should do is to skip the redzone. I guess interrupt handlers
>>>>> should switch to interrupt stack which avoids smashing redzone
>>>>> altogether. Do you mean nested interrupts/exceptions in interrupts?
>>>>> In my experience frames in leaf functions can have pretty large
>>>>> performance penalty. Wonder if we have we considered changing
>>>>> interrupt/exception handlers to avoid smashing redzones and disable
>>>>> leaf frames?
>>>>
>>>> Currently, on x86-64, I believe all exceptions have their own dedicated
>>>> stacks in the kernel, but IRQs still come in on the task's kernel stack.
>>>>
>>>> Andy, do you know if there's a reason why IRQs don't use a dedicated IST
>>>> stack?
>>>>
>>>
>>> Because IST is awful due to recursion issues.  We immediately switch to an IRQ stack, though.
>>>
>>> If the kernel wanted a redzone, it would have to use IST for everything, which would entail a bunch of unpleasant hackery.
>>
>> Thanks.
>>
>> I guess it must be finite recursion, because we could not handle
>> infinite with finite stack. I thing that solves it is simply:
>>
>> sub $256, %rsp
>> ... do stuff ...
>> add $256, %rsp
>>
>> Don't know if it's applicable to interrupts or not.
>
> No, it is not.  The processor pushes 5 or 6 words of data on the stack
> (the IRET frame plus an error code for certain exceptions) before the
> interrupt handler gets control.  So without using the IST for stack
> switching on every interrupt, the redzone cannot be used in the kernel
> as it will get smashed by the IRET frame.  In addition, since the
> kernel's stack is limited in size, skipping 128 bytes on every
> interrupt would overrun the stack faster.  The small gain from using
> the redzone in the kernel is outweighed by these limitations.

I see, thanks for educating.