[PATCH v2 1/2] powerpc/64s: system call scv tabort fix for corrupt irq soft-mask state

[Christophe Leroy]

Le 03/09/2021 à 00:20, Segher Boessenkool a écrit :
> On Thu, Sep 02, 2021 at 04:52:03PM -0500, Segher Boessenkool wrote:
>> On Thu, Sep 02, 2021 at 01:33:10PM +1000, Nicholas Piggin wrote:
>>> Excerpts from Christophe Leroy's message of September 2, 2021 3:21 am:
>>>>> -	/* Firstly we need to enable TM in the kernel */
>>>>> +	/* We need to enable TM in the kernel, and disable EE (for scv) */
>>>>>    	mfmsr	r10
>>>>>    	li	r9, 1
>>>>>    	rldimi	r10, r9, MSR_TM_LG, 63-MSR_TM_LG
>>>>> +	LOAD_REG_IMMEDIATE(r9, MSR_EE)
>>>>> +	andc	r10, r10, r9
>>>>
>>>> Why not use 'rlwinm' to mask out MSR_EE ?
>>>>
>>>> Something like
>>>>
>>>> 	rlwinm	r10, r10, 0, ~MSR_EE
>>>
>>> Mainly because I'm bad at powerpc assembly. Why do you think I'm trying
>>> to change as much as possible to C?
>>
>> The actual bit (bit 31, i.e. with value 1UL << 32) cannot be cleared
>> with rlwinm (only the low 32 bits can).  There are many ways to do it
>> using two insns of course.
> 
> Wow I misread that, you want to clear MSR[EE] really, not MSR[TM].
> 
> You cannot use rlwinm and keep the high 32 bits of the target register
> intact.  You either clear all to 0 or set them to a copy of the rotated
> value in the low 32 bits.
> 

Oops, my mistake. When I tested it in C to see what was generated by GCC I forgot the ~ so I got 
rlwinm r3,r3,0,16,16 and didn't realise it was different from rlwinm r3,r3,0,~(1<<15)

By the way it would be more explicit if objdump could display the mask instead of the mask 
boundaries. Is there a way to do that ?

Christophe