On 28.04.22 02:33, Lai Jiangshan wrote:
> On Thu, Apr 28, 2022 at 1:45 AM Borislav Petkov <bp@alien8.de> wrote:
>>
>> On Thu, Apr 21, 2022 at 10:10:50PM +0800, Lai Jiangshan wrote:
>>> From: Lai Jiangshan <jiangshan.ljs@antgroup.com>
>>>
>>> The macro idtentry calls error_entry() unconditionally even on XENPV.
>>> But the code XENPV needs in error_entry() is PUSH_AND_CLEAR_REGS only.
>>> And error_entry() also calls sync_regs() which has to deal with the
>>> case of XENPV via an extra branch so that it doesn't copy the pt_regs.
>>
>> What extra branch?
>>
>> Do you mean the
>>
>>          if (regs != eregs)
>>
>> test in sync_regs()?
> 
> Hello, Borislav
> 
> Yes.
> 
>>
>> I'm confused. Are you, per chance, aiming to optimize XENPV here or
>> what's up?
> 
> 
> The branch in sync_regs() can be optimized out for the non-XENPV case
> since XENPV doesn't call sync_regs() after patch5 which makes XENPV
> not call error_entry().
> 
> The aim of this patch and most of the patchset is to make
> error_entry() be able to be converted to C.  And XENPV cases can
> also be optimized in the patchset although it is not the major main.
> 
>>
>>> And PUSH_AND_CLEAR_REGS in error_entry() makes the stack not return to
>>> its original place when the function returns, which means it is not
>>> possible to convert it to a C function.
>>>
>>> Move PUSH_AND_CLEAR_REGS out of error_entry(), add a function to wrap
>>> PUSH_AND_CLEAR_REGS and call it before error_entry().
>>>
>>> The new function call adds two instructions (CALL and RET) for every
>>> interrupt or exception.
>>
>> Not only - it pushes all the regs in PUSH_AND_CLEAR_REGS too. I don't
>> understand why that matters here? It was done in error_entry anyway.
>>
> 
> Compared to the original code, adding the new function call adds two
> instructions (CALL and RET) for every interrupt or exception.
> 
> PUSH_AND_CLEAR_REGS is not extra instructions added here.
> 
> Since this patch adds extra overhead (CALL and RET), the changelog
> has to explain why it is worth it not just for converting ASM to C.
> 
> The explanation in the changelog is that it can be offsetted by later
> reduced overhead.

I think you could avoid the extra call/ret by doing something like:

SYM_CODE_START_LOCAL(error_exit_push_and_save)
	UNWIND_HINT_FUNC
	PUSH_AND_CLEAR_REGS save_ret=1
	ENCODE_FRAME_POINTER 8
	jmp error_exit
SYM_CODE_END(error_exit_push_and_save)

... and use this instead of patch 5:

	ALTERNATIVE "call error_entry_push_and_save; movq %rax, %rsp", \
		"call push_and_clear_regs", X86_FEATURE_XENPV


Juergen