Re: [PATCH 1/2] regmap: teach regmap to use raw spinlocks if requested in the config

From: Thomas Gleixner <tglx@linutronix.de>
To: Mark Brown <broonie@kernel.org>
Cc: Vladimir Oltean <vladimir.oltean@nxp.com>,
	Rasmus Villemoes <linux@rasmusvillemoes.dk>,
	Lee Jones <lee.jones@linaro.org>, Arnd Bergmann <arnd@arndb.de>,
	Marc Zyngier <maz@kernel.org>,
	Hou Zhiqiang <Zhiqiang.Hou@nxp.com>, Biwen Li <biwen.li@nxp.com>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	"Rafael J. Wysocki" <rafael@kernel.org>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 1/2] regmap: teach regmap to use raw spinlocks if requested in the config
Date: Thu, 02 Sep 2021 10:35:01 +0200	[thread overview]
Message-ID: <874kb3cr4a.ffs@tglx> (raw)
In-Reply-To: <20210901160555.GG5976@sirena.org.uk>

On Wed, Sep 01 2021 at 17:05, Mark Brown wrote:
> On Mon, Aug 30, 2021 at 04:16:04PM +0200, Thomas Gleixner wrote:
>> On Mon, Aug 30 2021 at 13:19, Mark Brown wrote:
>
>> > That probably does make sense, I think we're just using regular
>> > spinlocks for spinlocks mainly because they're the default rather
>> > than because anyone put huge amounts of thought into it.  IIRC
>> > the first users were using spinlocks for their locking when they
>> > were converted.
>
>> So if the actual spinlock protected operations are not doing any other
>> business than accessing preallocated cache memory and a few MMIO
>> operations then converting them to raw spinlocks should have no real
>> impact on RT.
>
> I think Vladimir's point that something might try to use one of the APIs
> that can do multiple register writes atomically to generate a very long
> register write sequence is valid here.  It's far from the common case
> but it'd be hard to audit, it's going to be a lot easier to handle going
> to raw spinlocks in the cases where it's specifically needed than to
> keep on top of ensuring that none of the users are causing issues or
> start causing issues in the future.  This does make me feel it's a bit
> safer to leave the default the way it is since if you get it wrong then
> lockdep will tend to notice very quickly while it's less likely that
> we'd get tooling spotting issues the other way around.

Fair enough.

>> One way to do that is obviously starting with the patch from Vladimir
>> and then convert them one by one, so the assumption that they are not
>> doing anything nasty (in the RT sense) can be validated.
>
> Vladimir's patch is in Linus' tree now so users that can safely do so
> can start using raw spinlocks.

ok