[ovs-discuss] [PATCH net-next] fast_hash: clobber registers correctly for inline function use

Jay Vosburgh jay.vosburgh at canonical.com
Fri Nov 14 20:15:15 UTC 2014

Hannes Frederic Sowa <hannes at stressinduktion.org> wrote:

>On Fr, 2014-11-14 at 13:38 -0500, David Miller wrote:
>> From: Hannes Frederic Sowa <hannes at stressinduktion.org>
>> Date: Fri, 14 Nov 2014 16:46:18 +0100
>> > I would still like to see the current proposed fix getting applied and
>> > we can do this on-top. The inline call after this patch reassembles a
>> > direct function call, so besides the long list of clobbers, it should
>> > still be pretty fast.
>> I would rather revert the change entirely until it is implemented
>> properly.
>> Also, I am strongly of the opinion that this is a mis-use of the
>> alternative call interface.  It was never intended to be used for
>> things that can make real function calls.
>I tend to disagree. Grepping e.g. shows
>        alternative_call_2(copy_user_generic_unrolled,
>                         copy_user_generic_string,
>                         X86_FEATURE_REP_GOOD,
>                         copy_user_enhanced_fast_string,
>                         X86_FEATURE_ERMS,
>                         ASM_OUTPUT2("=a" (ret), "=D" (to), "=S" (from),
>                                     "=d" (len)),
>                         "1" (to), "2" (from), "3" (len)
>                         : "memory", "rcx", "r8", "r9", "r10", "r11");
>(it has a few less clobbers because it has more output operands)

	As those functions (copy_user_generic_unrolled, et al) are all
in assembly language, presumably the list of clobbered registers can be
had via inspection.

	For the arch_fast_hash2 case, the functions (__intel_crc4_2_hash
and __jash2) are both written in C, so how would the clobber list be


>I just tried to come up with some macros which lets you abstract away
>the clobber list, but in the end it somehow has to look exactly like
>that. The double-colon syntax also makes it difficult to come up with
>something that let's us use varargs for that.
>> You can add a million clobbers, or a trampoline, it's still using a
>> facility in a manner for which it was not designed.
>The full clobber list for a function call which would always clear
>registers like we would have in a normal non-inlined function call would
>look like this:
>#define FUNC_CLOBBER LIST "memory", "cc", "rax", "rdi", "rsi", "rdx", "rcx", "r8", "r9", "r10", "r11"
>(reference in arch/x86/include/asm/calling.h).
>> This means a new interface with a new name and with capabilities
>> explicitly supporting this case are in order.
>It try to implicitly embed the clobber list, would something like that
>be ok?

	-Jay Vosburgh, jay.vosburgh at canonical.com

More information about the discuss mailing list