svn commit: r236456 - in head/sys: amd64/include i386/include

Attilio Rao attilio at freebsd.org
Mon Jun 4 15:59:23 UTC 2012


2012/6/4 Konstantin Belousov <kostikbel at gmail.com>:
> On Mon, Jun 04, 2012 at 02:58:57PM +0100, Attilio Rao wrote:
>> 2012/6/4 Konstantin Belousov <kostikbel at gmail.com>:
>> > On Mon, Jun 04, 2012 at 12:00:27PM +0200, Tijl Coosemans wrote:
>> >> On 02-06-2012 20:10, Konstantin Belousov wrote:
>> >> > Author: kib
>> >> > Date: Sat Jun  2 18:10:16 2012
>> >> > New Revision: 236456
>> >> > URL: http://svn.freebsd.org/changeset/base/236456
>> >> >
>> >> > Log:
>> >> >   Use plain store for atomic_store_rel on x86, instead of implicitly
>> >> >   locked xchg instruction.  IA32 memory model guarantees that store has
>> >> >   release semantic, since stores cannot pass loads or stores.
>> >>
>> >> They can pass non-temporal stores can't they?
>> > Sure. But (our) barriers only work for WB memory accesses, in respect to other
>> > WB memory accesses.
>> >
>> > The atomic(9) contains not quite explicit mention of the requirement,
>> > for ia32 and more direct notion for ia64. It could probably be reworded to
>> > mention memory access type explicitely for ia32 too.
>>
>> I don't think this is right.
>> What if I want to use NTI in a block of code locked? What if I want to
>> use CLFLUSH? I simply cannot do that now because of the reordering
>> requirement.
> Assuming that NTI means "Non Temporal Instruction", Intel explicit
> requirement is to use fence barrier if order shall be ensured. This,
> as well as CLFLUSH use, is somewhat commonly documented. More, CLFLUSH
> is documented by Intel to _not_ serialize with any other fencing or
> serialization instruction, except MFENCE. So xchg-based _store_rel is
> not different from mov-based _store_rel for CLFLUSH and non-temporal
> ops.
>
> I do not see how you note is relevant.
>
>> Also, there is the more worrisome case of the string operations. If
>> gcc/clang optimize the code in order to do string operations between
>> locked path, this is not valid anymore as they can be reordered
>> against the _rel() barrier.
> They cannot. Fast string operation volatile store order only among
> string operation itself, the operation cannot pass sequential store.
> The store used in _store_rel thus cannot be passed by fast string
> optimizations.
>
> I do not see how you note is relevant there, again.

I'm not sure why but I thought that the string writes could be
re-ordered against "external" writes too, but  I re-read the manual
and this is not the case.

>
>>
>> However, we should consider atomic(9) as a script for MI requirement
>> of our locking primitives among the architectures. Right now too many
>> things live on assumptions of people doing patches (like this case)
>> rather than actually working on a common policy of what we can easilly
>> support and what we can't.
>>
>> I also wondered often if we should use *fence on architectures
>> supporting them, by default, because of the possibility to use FPU now
>> (which wasn't present back in the day) and thus we cannot really
>> guarantee memory ordering over stores of memory areas bigger than a
>> quad-word. If we don't want to add the burden, we should explicitely
>> mention that in atomic(9) or any other place.
> The proposal to use fence explicitely contradicts recommendations from
> the AMD Optimization Guide, which, JFYI, I cited in the updated comment
> in the patch.
>
> How is FPU relevant to the memory model discussion, I left out of the
> answer.

I'm not saying to use *fence, I'm saying that we should document MD
cases where this is required (we can use agnostic terms like "simple
barrier") in atomic(9).
Also, about FPU, I'm saying that before we had no need to consider
FPU/XMM into consideration from a model perspective but now we should
do that because we can use FPU within the kernel.

>> Definitively: I think this patch violates some edge cases. Please back it out.
> No. I explicitely inform you that I consider the backout request
> as frivolous, technically unfounded, and that I will not back it out.

My main motivation for asking the backout was for the ordering issue
with the string operations, but it doesn't seem to be the case, than I
withdraw my request, sorry.
I'm fine with using explicit memory barriers in MD code that use NTI
(I just assume gcc/clang won't over-optimize things on their own as
thinking they want to reduce cache traffic by using NTI).

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein


More information about the svn-src-all mailing list