Received-SPF: pass (google.com: domain of linux-kernel+bounces-187638-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Message-ID: <a25f9654-e681-1bad-47ae-ddc519610504@huaweicloud.com>
Date: Thu, 23 May 2024 16:26:23 +0200
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101
 Thunderbird/102.13.0
Subject: Re: LKMM: Making RMW barriers explicit
To: Alan Stern <stern@rowland.harvard.edu>,
 Jonas Oberhauser <jonas.oberhauser@huaweicloud.com>
Cc: "Paul E. McKenney" <paulmck@kernel.org>, linux-kernel@vger.kernel.org,
 linux-arch@vger.kernel.org, kernel-team@meta.com, parri.andrea@gmail.com,
 boqun.feng@gmail.com, j.alglave@ucl.ac.uk, luc.maranget@inria.fr,
 Joel Fernandes <joel@joelfernandes.org>
References: <72c804c8-2511-4349-a823-bc1de8bb729e@rowland.harvard.edu>
 <e030f7a4-97e7-4e91-bbae-230ee5c97763@huaweicloud.com>
 <a9bf972c-b5ee-f1c2-36bf-30ba62f419d7@huaweicloud.com>
 <2f20e7cf-7c67-4ad3-8a0c-3c1d01257ae4@rowland.harvard.edu>
 <0c309dd3-f8c1-4945-b8f1-154b2a775216@huaweicloud.com>
 <4286e5b2-5954-4c77-a815-c1c2735d9509@rowland.harvard.edu>
 <58042cf3-e515-4e5f-ab48-1d0d6123c9e9@huaweicloud.com>
 <6174fd09-b287-49ae-b117-c3a36ef3800a@rowland.harvard.edu>
 <7bd31eca-3cf3-4377-a747-ec224262bd2e@huaweicloud.com>
 <35b3fd07-fa85-4244-b9cb-50ea54d9de6a@rowland.harvard.edu>
Content-Language: en-US
From: Hernan Ponce de Leon <hernan.poncedeleon@huaweicloud.com>
In-Reply-To: <35b3fd07-fa85-4244-b9cb-50ea54d9de6a@rowland.harvard.edu>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit

On 5/23/2024 4:05 PM, Alan Stern wrote:
> On Thu, May 23, 2024 at 02:54:05PM +0200, Jonas Oberhauser wrote:
>>
>>
>> Am 5/22/2024 um 4:20 PM schrieb Alan Stern:
>>> It would be better if there was a way to tell herd7 not to add the 'mb
>>> tag to failed instructions in the first place.  This approach is
>>> brittle; see below.
>>
>> Hernan told me that in fact that is actually currently the case in herd7.
>> Failing RMW get assigned the Once tag implicitly.
>> Another thing that I'd suggest to change.
> 
> Indeed.
> 
>>> An alternative would be to have a way for the .cat file to remove the
>>> 'mb tag from a failed RMW instruction.  But I don't know if this is
>>> feasible.
>>
>> For Mb it's feasible, as there is no Mb read or Mb store.
>>
>> Mb = Mb & (~M | dom(rmw) | range(rmw))
>>
>> However one would want to do the same for Acq and Rel.
>>
>> For that one would need to distinguish e.g. between a read that comes from a
>> failed rmw instruction, and where the tag would disappear, or a normal
>> standalone read.
>>
>> For example, by using two different acquire tags, 'acquire and 'rmw-acquire,
>> and defining
>>
>> Acquire = Acquire | Rmw-acquire & (dom(rmw) | range(rmw))
>>
>> Anyways we can do this change independently. So for now, we don't need
>> RMW_MB.
> 
> Overall, it seems better to have herd7 assign the right tag, but change
> the way the .def file works so that it can tell herd7 which tag to use
> in each of the success and failure cases.

I am not fully sure how herd7 uses the .def file, but I guess something 
like adding a second memory tag to __cmpxchg could work

cmpxchg(X,V,W) __cmpxchg{mb, once}(X,V,W)
cmpxchg_relaxed(X,V,W) __cmpxchg{once, once}(X,V,W)
cmpxchg_acquire(X,V,W) __cmpxchg{acquire, acquire}(X,V,W)
cmpxchg_release(X,V,W) __cmpxchg{release, release}(X,V,W)

Hernan

> 
>>> 	[M] ; po ; [RMW_MB]
>>>
>>> 	[RMW_MB] ; po ; [M]
>>>
>>> This is because events tagged with RMW_MB always are memory accesses,
>>> and accesses that aren't part of the RMW are already covered by the
>>> fencerel(Mb) thing above.
>>
>> This has exactly the issue mentioned above - it will cause the rmw to have
>> an internal strong fence that on powerpc probably isn't there.
> 
> Oops, that's right.  Silly oversight on my part.  But at least you
> understood what I meant.
> 
>> We could do (with the assumption that Mb applies only to successful rmw):
>>
>>   	[M] ; po ; [Mb & R]
>>   	[Mb & W] ; po ; [M]
> 
> That works.
> 
> Alan