Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Wed, 10 Mar 2021 13:33:23 -0700
From:   Nathan Chancellor <nathan@kernel.org>
To:     Jens Axboe <axboe@kernel.dk>
Cc:     linux-block@vger.kernel.org, linux-kernel@vger.kernel.org,
        clang-built-linux@googlegroups.com
Subject: Re: -Walign-mismatch in block/blk-mq.c
Message-ID: <20210310203323.35w2q7tlnxe23ukg@Ryzen-9-3900X.localdomain>
References: <20210310182307.zzcbi5w5jrmveld4@archlinux-ax161>
 <99cf90ea-81c0-e110-4815-dd1f7df36cb4@kernel.dk>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <99cf90ea-81c0-e110-4815-dd1f7df36cb4@kernel.dk>
Precedence: bulk

On Wed, Mar 10, 2021 at 01:21:52PM -0700, Jens Axboe wrote:
> On 3/10/21 11:23 AM, Nathan Chancellor wrote:
> > Hi Jens,
> > 
> > There is a new clang warning added in the development branch,
> > -Walign-mismatch, which shows an instance in block/blk-mq.c:
> > 
> > block/blk-mq.c:630:39: warning: passing 8-byte aligned argument to
> > 32-byte aligned parameter 2 of 'smp_call_function_single_async' may
> > result in an unaligned pointer access [-Walign-mismatch]
> >                 smp_call_function_single_async(cpu, &rq->csd);
> >                                                     ^
> > 1 warning generated.
> > 
> > There appears to be some history here as I can see that this member was
> > purposefully unaligned in commit 4ccafe032005 ("block: unalign
> > call_single_data in struct request"). However, I later see a change in
> > commit 7c3fb70f0341 ("block: rearrange a few request fields for better
> > cache layout") that seems somewhat related. Is it possible to get back
> > the alignment by rearranging the structure again? This seems to be the
> > only solution for the warning aside from just outright disabling it,
> > which would be a shame since it seems like it could be useful for
> > architectures that cannot handle unaligned accesses well, unless I am
> > missing something obvious :)
> 
> It should not be hard to ensure that alignment without re-introducing
> the bloat. Is there some background on why 32-byte alignment is
> required?
> 

This alignment requirement was introduced in commit 966a967116e6 ("smp:
Avoid using two cache lines for struct call_single_data") and it looks
like there was a thread between you and Peter Zijlstra that has some
more information on this:
https://lore.kernel.org/r/a9beb452-7344-9e2d-fc80-094d8f5a0394@kernel.dk/

Cheers,
Nathan