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:   Tue, 06 Apr 2021 13:12:55 +0100
Message-ID: <87lf9vpq60.wl-maz@kernel.org>
From:   Marc Zyngier <maz@kernel.org>
To:     Christophe Leroy <christophe.leroy@csgroup.eu>
Cc:     linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        linux-mips@vger.kernel.org, linuxppc-dev@lists.ozlabs.org,
        linux-sh@vger.kernel.org,
        Thomas Bogendoerfer <tsbogend@alpha.franken.de>,
        Yoshinori Sato <ysato@users.sourceforge.jp>,
        Haojian Zhuang <haojian.zhuang@gmail.com>,
        Rich Felker <dalias@libc.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Robert Jarzmik <robert.jarzmik@free.fr>,
        Daniel Mack <daniel@zonque.org>
Subject: Re: [PATCH 1/9] irqdomain: Reimplement irq_linear_revmap() with irq_find_mapping()
In-Reply-To: <15be426f-4429-ebeb-1b4a-8342bce391e5@csgroup.eu>
References: <20210406093557.1073423-1-maz@kernel.org>
        <20210406093557.1073423-2-maz@kernel.org>
        <15be426f-4429-ebeb-1b4a-8342bce391e5@csgroup.eu>
User-Agent: Wanderlust/2.15.9 (Almost Unreal) SEMI-EPG/1.14.7 (Harue)
 FLIM-LB/1.14.9 (=?UTF-8?B?R29qxY0=?=) APEL-LB/10.8 EasyPG/1.0.0 Emacs/27.1
 (x86_64-pc-linux-gnu) MULE/6.0 (HANACHIRUSATO)
MIME-Version: 1.0 (generated by SEMI-EPG 1.14.7 - "Harue")
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8BIT
Precedence: bulk

Christophe,

On Tue, 06 Apr 2021 12:21:33 +0100,
Christophe Leroy <christophe.leroy@csgroup.eu> wrote:
> 
> 
> 
> Le 06/04/2021 à 11:35, Marc Zyngier a écrit :
> > irq_linear_revmap() is supposed to be a fast path for domain
> > lookups, but it only exposes low-level details of the irqdomain
> > implementation, details which are better kept private.
> 
> Can you elaborate with more details ?

Things like directly picking into the revmap are positively awful, and
doesn't work if the domain has been constructed using the radix
tree. Which on its own is totally broken, because things like
irq_domain_create_hierarchy() will pick an implementation or the
other.

> 
> > 
> > The *overhead* between the two is only a function call and
> > a couple of tests, so it is likely that noone can show any
> > meaningful difference compared to the cost of taking an
> > interrupt.
> 
> Do you have any measurement ?

I did measure things on arm64, and couldn't come up with any
difference other than noise.

> Can you make the "likely" a certitude ?

Of course not. You can always come up with an artificial CPU
implementation that has a very small exception entry overhead, and a
ridiculously slow memory subsystem. Do I care about these? No.

If you can come up with realistic platforms that show a regression
with this patch, I'm all ears.

> 
> > 
> > Reimplement irq_linear_revmap() with irq_find_mapping()
> > in order to preserve source code compatibility, and
> > rename the internal field for a measure.
> 
> This is in complete contradiction with commit https://github.com/torvalds/linux/commit/d3dcb436
> 
> At that time, irq_linear_revmap() was less complex than what
> irq_find_mapping() is today, and nevertheless it was considered worth
> restoring in as a fast path. What has changed since then ?

Over 8 years? Plenty. The use of irqdomains has been generalised, we
have domain hierarchies, and if anything, this commit introduces the
buggy behaviour I was mentioning above. I also don't see any mention
of actual performance in that commit.

And if we're worried about a fast path, being able to directly cache
the irq_data in the revmap, hence skipping the irq_desc lookup that
inevitable follows, is a much more interesting prospect than the "get
useless data quick" that irq_linear_revmap() implements.

This latter optimisation is what I am after.

> Can you also explain the reason for the renaming of "linear_revmap"
> into "revmap" ? What is that "measure" ?

To catch the potential direct use of the reverse map field.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.