Re: Q: put_user_try & co (Was: [PATCH 1/5] Split wait_noreap_copyout())

[Oleg Nesterov <oleg@redhat.com>]

On 05/20, Roland McGrath wrote:
>
> > 	#define __put_user_asm_ex(...)				\
> > 		asm volatile(					\
> > 			"1:	mov ..."			\
> > 			_ASM_EXTABLE(1b, &__efault_label)	\
> > 		     	: : ...)
>
> You mean &&__efault_label here (it's a funny syntax, but that's how it is).
> &&label is a GCC extension that I'm not sure the kernel has used before.
>
> I think it can be touchy to have an asm jump into compiled code that way.
> e.g., perhaps the compiler produced:
>
> 	mov reg, 40(sp)
> 	mov $123, reg
> 	#APP
> 	... inside of your asm ...
> 	#NO_APP
> 	mov 40(sp), reg
>
> or some such thing.  If you jump away from inside the asm, you won't ever
> do "mov 40(sp), reg".  But the compiler might think that reg has its
> original value at the __efault_label: code location.
>
> Perhaps more important than any particular compiler-confusion scenario we
> can come up with is simply that this would be an obscure corner of code
> generation in the compiler that the kernel has not evoked before.  There
> might be bugs or oddities in various compilers of various vintages, that
> we don't know about because they never came up before.

Yes, agreed. Thanks to all for replies.

Oleg.