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;
Date:   Sat, 13 May 2023 13:10:56 -0700
From:   Davidlohr Bueso <dave@stgolabs.net>
To:     Matthew Wilcox <willy@infradead.org>
Cc:     Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Darren Hart <dvhart@infradead.org>,
        =?utf-8?B?QW5kcsOvwr/CvQ==?= Almeida <andrealmeid@igalia.com>,
        linux-kernel@vger.kernel.org, Jason Gunthorpe <jgg@nvidia.com>,
        linux-mm@kvack.org
Subject: Re: Futexes & Folios
Message-ID: <4b5rvoi6w6nhiihmog6xjftoveohd47oooodekqedo5qhflmrt@jemw5ss5xps3>
Mail-Followup-To: Matthew Wilcox <willy@infradead.org>, 
        Thomas Gleixner <tglx@linutronix.de>, Ingo Molnar <mingo@redhat.com>, 
        Peter Zijlstra <peterz@infradead.org>, Darren Hart <dvhart@infradead.org>, 
        =?utf-8?B?QW5kcsOvwr/CvQ==?= Almeida <andrealmeid@igalia.com>, linux-kernel@vger.kernel.org, Jason Gunthorpe <jgg@nvidia.com>, 
        linux-mm@kvack.org
References: <ZF+3xpmouwur4sVE@casper.infradead.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Disposition: inline
In-Reply-To: <ZF+3xpmouwur4sVE@casper.infradead.org>
User-Agent: NeoMutt/20230407
Precedence: bulk

On Sat, 13 May 2023, Matthew Wilcox wrote:

>Here's a patch which converts the futex code from using pages to using
>folios.  I do have some questions.
>
> - I assume we really do want to base the key on the page offset of the
>   futex in the file.  That is, it's not good enough to base the key
>   on the folio's index because the folio can be split after setting up
>   the key, and if it is we'd then fail to find the futex later.
> - In hugetlbfs, it doesn't matter whether the pgoff is based on base
>   pages or huge page size -- as long as it's consistent between
>   invocations, everything will work

As tglx already mentioned, yes to both.

>(I have ideas about a get_user_folio() but those are not represented in
>this patch)

Btw, regarding some of the concerns about the gup overhead for futexes
in the physr discussion at lsfmm; 'perf bench futex hash' would be a good
way of measuring any impact at a micro level and going from there. This
is only necessary of course for shared futexes, fast/private ones just
use the address space.

Thanks,
Davidlohr