From mboxrd@z Thu Jan  1 00:00:00 1970
From: marek.vasut@gmail.com (Marek Vasut)
Date: Mon, 19 Dec 2011 12:57:05 +0100
Subject: [PATCH] MXS: Convert mutexes in clock.c to spinlocks
In-Reply-To: <20111219082225.GC14542@n2100.arm.linux.org.uk>
References: <1324217174-6574-1-git-send-email-marek.vasut@gmail.com>
 <201112190503.46668.marek.vasut@gmail.com>
 <20111219082225.GC14542@n2100.arm.linux.org.uk>
Message-ID: <201112191257.06145.marek.vasut@gmail.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

> On Mon, Dec 19, 2011 at 05:03:45AM +0100, Marek Vasut wrote:
> > > On Sun, Dec 18, 2011 at 03:06:13PM +0100, Marek Vasut wrote:
> > > > The mutexes can't be safely used under certain circumstances. I
> > > > noticed this
> > > 
> > > > issue during some network instability at home:
> > > Yes, this is a known issue.  And there was some discussion[1] about
> > > why mutex is needed.
> > 
> > Thanks for pointing this out, I was unaware of it.
> > 
> > > But I really have not thought about why we can
> > > not use spinlock only, since using mutex only leads to the issue we
> > > are seeing here, and using spinlock in enable/disable and mutex in
> > > rate/parent will not work, because the mxs clocks have enable/disable
> > > and rate/parent functions access the same register.  I know it's not
> > > good to hold spinlock in rate/parent functions for a long time, but
> > > do we have a way around rather than using spinlock for both sets of
> > > functions?
> > 
> > Yea, spinlock is not good either. On the other hand, is it really held
> > for so long ?
> 
> There is another solution to this, which I've pointed out before when
> this has come up:
> 
> 1. Convert all your drivers to _also_ use clk_prepare()/clk_unprepare().
>    You need to do this anyway as it will become mandatory for the common
>    clk stuff.
> 
> 2. Rename your existing clk_enable()/clk_disable() implementation to
>    clk_prepare()/clk_unprepare().  Ensure CONFIG_HAVE_CLK_PREPARE is
>    selected.
> 
> 3. Provide a new no-op clk_enable()/clk_disable() functions.

Well, I'm still unsure how'd you then enable/disable the clock ? 
clk_prepare/clk_unprepare is good, but how would that help in avoiding the 
mutex/spinlock?

> 
> This fixes the issue because clk_prepare()/clk_unprepare() must only be
> called from process contexts, whereas clk_enable()/clk_disable() may be
> called from atomic contexts as well.

Sure, but I need to enable the clock in atomic context ...

M