Re: [PATCH 25/24] x86/traps: Rewrite native_load_gs_index in C code

[Lai Jiangshan <laijs@linux.alibaba.com>]

On 2021/9/14 04:01, Andy Lutomirski wrote:
> 
> 
> On Thu, Sep 2, 2021, at 3:50 AM, Lai Jiangshan wrote:
>> From: Lai Jiangshan <laijs@linux.alibaba.com>
>>
>> There is no constrain/limition to force native_load_gs_index() to be in
>> ASM code.
>>
>> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
> 
>>   
>>   #ifdef CONFIG_X86_64
>> +
>> +/*
>> + * Reload gs selector with exception handling
>> + * selector:  new selector
>> + *
>> + * Is noinstr as it shouldn't be instrumented.
>> + */
>> +noinstr void native_load_gs_index(unsigned int selector)
>> +{
>> +	unsigned long flags;
>> +
>> +	local_irq_save(flags);
> 
> This patch would be a bit less alarming if you moved the swapgs into asm.  

Emmm, this patch is not so clean and persuadable in C.

I think Peter is still working on reworking the patchset and may be
including improving this patch.  I'm Okay if this patch is dropped.

> Also, this needs a comment explaining why skipping the swapgs back to kernel gs in the exception path is correct.
> 

I think it is all known that the exception handler in ASM_EXTABLE is running in kernel context where kernel gs is active.

It does need a comment explaining why the label asm_load_gs_index_gs_change is needed, how does it help the 
error_entry() restores back to the kernel gs.

Since the C-version error_entry() has to check asm_load_gs_index_gs_change, I think other way to handle the fault of 
"mov %gs" is just doing the %gs fixup in the C-version error_entry(). (see patch 11).  it would be more directly, 
simple, and self-documented.

Thank you for reviewing.

>> +	native_swapgs();
>> +	asm volatile(
>> +		".global asm_load_gs_index_gs_change \n"
>> +		"asm_load_gs_index_gs_change: \n"
>> +		"1: movl %0, %%gs \n"
>> +		"   swapgs \n"
>> +		"2: \n"
>> +		_ASM_EXTABLE_HANDLE(1b, 2b, ex_handler_clear_gs)
>> +		:: "r" (selector) : "memory");
>> +	alternative("", "mfence", X86_BUG_SWAPGS_FENCE);
>> +	local_irq_restore(flags);
>> +}
>> +