Re: [PATCH for 4.18 2/6] rseq: use get_user/put_user rather than __get_user/__put_user

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Jul 10, 2018, at 2:16 AM, Michael Ellerman mpe@ellerman.id.au wrote:

> Mathieu Desnoyers <mathieu.desnoyers@efficios.com> writes:
>> ----- On Jul 8, 2018, at 5:03 PM, Mathieu Desnoyers
>> mathieu.desnoyers@efficios.com wrote:
>>
>>> In preparation to use __u64 for the rseq_cs pointer field, 32-bit
>>> architectures need to read this 64-bit value located in user-space
>>> addresses.
>>> 
>>> __get_user is used to read this value, given that its access check has
>>> already been performed with access_ok() on rseq registration.
>>> 
>>> arm does not implement 8-byte __get_user. Rather than trying to
>>> improve __get_user on ARM, use get_user/put_user across rseq instead.
>>> 
>>> If those end up showing up in benchmarks, the proper approach would be to
>>> use user_access_begin() / unsafe_get/put_user() / user_access_end()
>>> anyway.
>>
>> So, another twist to this story: ppc32 does not implement u64 get_user().
> 
> Or __get_user() for that matter.
> 
> But we should just fix it.
> 
> We have the asm to do it, it's just the fact that __gu_val is unsigned
> long causes the size > sizeof(x) check here to fail:
> 
> #define __get_user_size(x, ptr, size, retval)			\
> do {								\
>	retval = 0;						\
>	__chk_user_ptr(ptr);					\
>	if (size > sizeof(x))					\
>		(x) = __get_user_bad();				\
> 
> 
> 
> We seem to be able to fix that with the __inttype() trick that x86 uses.
> 
> That's probably not 4.18 material though. But if you want to go with
> copy_from_user() for now you could then switch to get_user() for 4.19.

I agree. Let's use copy_from_user() for 4.18. Once get_user() ends up supporting
u64 on ppc32 for 4.19, rseq will happily move back to it.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com