Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Subject: Re: [PATCH] net: bcmgenet: Use stronger register read/writes to
 assure ordering
To:     Jakub Kicinski <kuba@kernel.org>,
        Florian Fainelli <f.fainelli@gmail.com>
Cc:     Peter Robinson <pbrobinson@gmail.com>,
        Jeremy Linton <jeremy.linton@arm.com>, netdev@vger.kernel.org,
        opendmb@gmail.com, davem@davemloft.net,
        bcm-kernel-feedback-list@broadcom.com, linux-kernel@vger.kernel.org
References: <20220310045358.224350-1-jeremy.linton@arm.com>
 <f831a4c6-58c9-20bd-94e8-e221369609e8@gmail.com>
 <de28c0ec-56a7-bfff-0c41-72aeef097ee3@arm.com>
 <2167202d-327c-f87d-bded-702b39ae49e1@gmail.com>
 <CALeDE9MerhZWwJrkg+2OEaQ=_9C6PHYv7kQ_XEQ6Kp7aV2R31A@mail.gmail.com>
 <472540d2-3a61-beca-70df-d5f152e1cfd1@gmail.com>
 <20220318122017.24341eb1@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
From:   Florian Fainelli <f.fainelli@gmail.com>
Message-ID: <082a7743-2eb8-c067-e41d-2a3acca5c056@gmail.com>
Date:   Fri, 18 Mar 2022 14:26:36 -0700
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101
 Thunderbird/78.14.0
MIME-Version: 1.0
In-Reply-To: <20220318122017.24341eb1@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 3/18/22 12:20 PM, Jakub Kicinski wrote:
> On Fri, 18 Mar 2022 12:01:20 -0700 Florian Fainelli wrote:
>> Given the time crunch we should go with Jeremy's patch that uses
>> stronger I/O access method and then we will work with Jeremy offline to
>> make sure that our version of GCC12 is exactly the same as his, as well
>> as the compiler options (like -mtune/-march) to reproduce this.
>>
>> If we believe this is only a problem with GCC12 and 5.17 in Fedora, then
>> I would be inclined to remove the Fixes tag such that when we come up
>> with a more localized solution we do not have to revert "net: bcmgenet:
>> Use stronger register read/writes to assure ordering" from stable
>> branches. This would be mostly a courtesy to our future selves, but an
>> argument could be made that this probably has always existed, and that
>> different compilers could behave more or less like GCC12.
> 
> Are you expecting this patch to make 5.17? If Linus cuts final this
> weekend, as he most likely will, unless we do something special the
> patch in question will end up getting merged during the merge window.
> Without the Fixes tag you'll need to manually instruct Greg to pull 
> it in. Is that the plan?

Maybe I should have refrained from making that comment after all :)
Having the Fixes: tag dramatically helps with getting this patch applied
all the way to the relevant stable trees and surely correctness over
speed should prevail. If we want to restore the performance loss (with
the onus on Doug and I to prove that there is a performance drop), then
we could send a fix with the appropriate localized barrier followed by a
revert of Jeremy's patch. And if we cared about getting those two
patches applied to stable, we would tag them with the appropriate Fixes tag.

It looks like there are a few 'net' changes that showed up, are you
going to send a pull request to Linus before 5.17 final is cut?
-- 
Florian