> The most optimal way of alternatively calling two functions would be
> something like this, IMO:
>
> alternative_call(memcpy, __mcsafe_copy, X86_FEATURE_MCRECOVERY,
> ASM_OUTPUT2("=a" (mcsafe_ret.trapnr), "=d"
> (mcsafe_ret.remain)),
> "D" (dst), "S" (src), "d" (len));
>
> I hope I've not messed up the calling convention but you want the inputs
> in %rdi, %rsi, %rdx and the outputs in %rax, %rdx, respectively. Just
> check the asm gcc generates and do not trust me :)
>
> The other thing you probably would need to do is create our own
> __memcpy() which returns struct mcsafe_ret so that the signatures of
> both functions match.
>
> Yeah, it is a bit of jumping through hoops but this way we do a CALL
> <func_ptr> directly in asm, without any JMPs or NOPs padding the other
> alternatives methods add.
>
> But if you don't care about a small JMP and that is not a hot path, you
> could do the simpler:
>
> if (static_cpu_has(X86_FEATURE_MCRECOVERY))
> return __mcsafe_copy(...);
>
> return memcpy();
>
> which adds a JMP or a 5-byte NOP depending on the X86_FEATURE_MCRECOVERY
> setting.
Dan,
What do you want the API to look like at the point you make a call
in the libnvdimm code? Something like:
r = nvcopy(dst, src, len);
where the innards of nvcopy() does the check for X86_FEATURE_MCE_RECOVERY?
What is useful to you in the return value? The low level __mcsafe_copy() returns
both a remainder and a trap number. But in your case I don't think you
need the trap
number (if the remaining count is not zero, then there must have been a #MC. #PF
isn't an option for you, right?
-Tony
