[PATCH V4 5/5] riscv: atomic: Optimize LRSC-pairs atomic ops with .aqrl annotation

[Dan Lustig]

On 7/6/2022 8:03 PM, Boqun Feng wrote:
> On Sat, Jun 25, 2022 at 01:29:50PM +0800, Guo Ren wrote:
>> On Fri, Jun 24, 2022 at 1:09 AM Dan Lustig <dlustig at nvidia.com> wrote:
>>>
>>> On 6/22/2022 11:31 PM, Boqun Feng wrote:
>>>> Hi,
>>>>
>>>> On Tue, Jun 14, 2022 at 01:03:47PM +0200, Andrea Parri wrote:
>>>> [...]
>>>>>> 5ce6c1f3535f ("riscv/atomic: Strengthen implementations with fences")
>>>>>> is about fixup wrong spinlock/unlock implementation and not relate to
>>>>>> this patch.
>>>>>
>>>>> No.  The commit in question is evidence of the fact that the changes
>>>>> you are presenting here (as an optimization) were buggy/incorrect at
>>>>> the time in which that commit was worked out.
>>>>>
>>>>>
>>>>>> Actually, sc.w.aqrl is very strong and the same with:
>>>>>> fence rw, rw
>>>>>> sc.w
>>>>>> fence rw,rw
>>>>>>
>>>>>> So "which do not give full-ordering with .aqrl" is not writen in
>>>>>> RISC-V ISA and we could use sc.w/d.aqrl with LKMM.
>>>>>>
>>>>>>>
>>>>>>>>> describes the issue more specifically, that's when we added these
>>>>>>>>> fences.  There have certainly been complains that these fences are too
>>>>>>>>> heavyweight for the HW to go fast, but IIUC it's the best option we have
>>>>>>>> Yeah, it would reduce the performance on D1 and our next-generation
>>>>>>>> processor has optimized fence performance a lot.
>>>>>>>
>>>>>>> Definately a bummer that the fences make the HW go slow, but I don't
>>>>>>> really see any other way to go about this.  If you think these mappings
>>>>>>> are valid for LKMM and RVWMO then we should figure this out, but trying
>>>>>>> to drop fences to make HW go faster in ways that violate the memory
>>>>>>> model is going to lead to insanity.
>>>>>> Actually, this patch is okay with the ISA spec, and Dan also thought
>>>>>> it was valid.
>>>>>>
>>>>>> ref: https://lore.kernel.org/lkml/41e01514-74ca-84f2-f5cc-2645c444fd8e@nvidia.com/raw
>>>>>
>>>>> "Thoughts" on this regard have _changed_.  Please compare that quote
>>>>> with, e.g.
>>>>>
>>>>>   https://lore.kernel.org/linux-riscv/ddd5ca34-805b-60c4-bf2a-d6a9d95d89e7@nvidia.com/
>>>>>
>>>>> So here's a suggestion:
>>>>>
>>>>> Reviewers of your patches have asked:  How come that code we used to
>>>>> consider as buggy is now considered "an optimization" (correct)?
>>>>>
>>>>> Denying the evidence or going around it is not making their job (and
>>>>> this upstreaming) easier, so why don't you address it?  Take time to
>>>>> review previous works and discussions in this area, understand them,
>>>>> and integrate such knowledge in future submissions.
>>>>>
>>>>
>>>> I agree with Andrea.
>>>>
>>>> And I actually took a look into this, and I think I find some
>>>> explanation. There are two versions of RISV memory model here:
>>>>
>>>> Model 2017: released at Dec 1, 2017 as a draft
>>>>
>>>>       https://groups.google.com/a/groups.riscv.org/g/isa-dev/c/hKywNHBkAXM/m/QzUtxEWLBQAJ
>>>>
>>>> Model 2018: released at May 2, 2018
>>>>
>>>>       https://groups.google.com/a/groups.riscv.org/g/isa-dev/c/xW03vmfmPuA/m/bMPk3UCWAgAJ
>>>>
>>>> Noted that previous conversation about commit 5ce6c1f3535f happened at
>>>> March 2018. So the timeline is roughly:
>>>>
>>>>       Model 2017 -> commit 5ce6c1f3535f -> Model 2018
>>>>
>>>> And in the email thread of Model 2018, the commit related to model
>>>> changes also got mentioned:
>>>>
>>>>       https://github.com/riscv/riscv-isa-manual/commit/b875fe417948635ed68b9644ffdf718cb343a81a
>>>>
>>>> in that commit, we can see the changes related to sc.aqrl are:
>>>>
>>>>        to have occurred between the LR and a successful SC.  The LR/SC
>>>>        sequence can be given acquire semantics by setting the {\em aq} bit on
>>>>       -the SC instruction.  The LR/SC sequence can be given release semantics
>>>>       -by setting the {\em rl} bit on the LR instruction.  Setting both {\em
>>>>       -  aq} and {\em rl} bits on the LR instruction, and setting the {\em
>>>>       -  aq} bit on the SC instruction makes the LR/SC sequence sequentially
>>>>       -consistent with respect to other sequentially consistent atomic
>>>>       -operations.
>>>>       +the LR instruction.  The LR/SC sequence can be given release semantics
>>>>       +by setting the {\em rl} bit on the SC instruction.  Setting the {\em
>>>>       +  aq} bit on the LR instruction, and setting both the {\em aq} and the {\em
>>>>       +  rl} bit on the SC instruction makes the LR/SC sequence sequentially
>>>>       +consistent, meaning that it cannot be reordered with earlier or
>>>>       +later memory operations from the same hart.
>>>>
>>>> note that Model 2018 explicitly says that "ld.aq+sc.aqrl" is ordered
>>>> against "earlier or later memory operations from the same hart", and
>>>> this statement was not in Model 2017.
>>>>
>>>> So my understanding of the story is that at some point between March and
>>>> May 2018, RISV memory model folks decided to add this rule, which does
>>>> look more consistent with other parts of the model and is useful.
>>>>
>>>> And this is why (and when) "ld.aq+sc.aqrl" can be used as a fully-ordered
>>>> barrier ;-)
>>>>
>>>> Now if my understanding is correct, to move forward, it's better that 1)
>>>> this patch gets resend with the above information (better rewording a
>>>> bit), and 2) gets an Acked-by from Dan to confirm this is a correct
>>>> history ;-)
>>>
>>> I'm a bit lost as to why digging into RISC-V mailing list history is
>>> relevant here...what's relevant is what was ratified in the RVWMO
>>> chapter of the RISC-V spec, and whether the code you're proposing
>>> is the most optimized code that is correct wrt RVWMO.
>>>
>>> Is your claim that the code you're proposing to fix was based on a
>>> pre-RVWMO RISC-V memory model definition, and you're updating it to
>>> be more RVWMO-compliant?
>> Could "lr + beq + sc.aqrl" provides a conditional RCsc here with
>> current spec? I only found "lr.aq + sc.aqrl" despcriton which is
>> un-conditional RCsc.
>>
> 
> /me put the temporary RISCV memory model hat on and pretend to be a
> RISCV memory expert.
> 
> I think the answer is yes, it's actually quite straightforwards given
> that RISCV treats PPO (Preserved Program Order) as part of GMO (Global
> Memory Order), considering the following (A and B are memory accesses):
> 
> 	A
> 	..
> 	sc.aqrl // M
> 	..
> 	B
> 
> , A has a ->ppo ordering to M since "sc.aqrl" is a RELEASE, and M has
> a ->ppo ordeing to B since "sc.aqrl" is an AQUIRE, so
> 
> 	A ->ppo M ->ppo B
> 
> And since RISCV describes that PPO is part of GMO:
> 
> """
> The subset of program order that must be respected by the global memory
> order is known as preserved program order.
> """
> 
> also in the herd model:
> 
> 	(* Main model axiom *)
> 	acyclic co | rfe | fr | ppo as Model
> 
> , therefore the ordering between A and B is GMO and GMO should be
> respected by all harts.
> 
> Regards,
> Boqun

I agree with Boqun's reasoning, at least for the case where there
is no branch.

But to confirm, was the original question about also having a branch,
I assume to the instruction immediately after the sc?  If so, then
yes, that would make the .aqrl effect conditional.

Dan

> 
>>>
>>> Dan
>>>
>>>> Regards,
>>>> Boqun
>>>>
>>>>>   Andrea
>>>>>
>>>>>
>>>> [...]
>>
>>
>>
>> --
>> Best Regards
>>  Guo Ren
>>
>> ML: https://lore.kernel.org/linux-csky/