Re: [PATCH] PM / OPP: Protect updates to list_dev with mutex

From: Viresh Kumar <viresh.kumar@linaro.org>
To: Stephen Boyd <sboyd@codeaurora.org>
Cc: Rafael Wysocki <rjw@rjwysocki.net>,
	mturquette@baylibre.com, linaro-kernel@lists.linaro.org,
	linux-pm@vger.kernel.org,
	Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>,
	Dan Carpenter <dan.carpenter@oracle.com>,
	Dmitry Torokhov <dtor@chromium.org>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Len Brown <len.brown@intel.com>,
	open list <linux-kernel@vger.kernel.org>,
	Nishanth Menon <nm@ti.com>, Pavel Machek <pavel@ucw.cz>
Subject: Re: [PATCH] PM / OPP: Protect updates to list_dev with mutex
Date: Sat, 31 Oct 2015 07:44:09 +0530	[thread overview]
Message-ID: <20151031021409.GT3716@ubuntu> (raw)
In-Reply-To: <20151030170604.GD19782@codeaurora.org>

On 30-10-15, 10:06, Stephen Boyd wrote:
> On 10/30, Viresh Kumar wrote:
> > dev_opp_list_lock is used everywhere to protect device and OPP lists,
> > but dev_pm_opp_set_sharing_cpus() is missed somehow. And instead we used
> > rcu-lock, which wouldn't help here as we are adding a new list_dev.
> > 
> > This also fixes a problem where we have called kzalloc(..., GFP_KERNEL)
> > from within rcu-lock, which isn't allowed as kzalloc can sleep when
> > called with GFP_KERNEL.
> 
> Care to share the splat here?

I don't know what is wrong (or right) with my exynos 5250 board, but I
didn't got any splat here even with the right config options (yes I
should have mentioned that earlier). I have seen this at other times
as well, while we were running after some cpufreq traces..

But, the case in hand is pretty straight forward and Mike T. did get a
splat as that's what he told me. We are calling a sleep-able function
from rcu_lock and that's obviously wrong.

> > diff --git a/drivers/base/power/opp/cpu.c b/drivers/base/power/opp/cpu.c
> > index 7654c5606307..91f15b2e25ee 100644
> > --- a/drivers/base/power/opp/cpu.c
> > +++ b/drivers/base/power/opp/cpu.c
> > @@ -124,12 +124,12 @@ int dev_pm_opp_set_sharing_cpus(struct device *cpu_dev, cpumask_var_t cpumask)
> >  	struct device *dev;
> >  	int cpu, ret = 0;
> >  
> > -	rcu_read_lock();
> > +	mutex_lock(&dev_opp_list_lock);
> >  
> >  	dev_opp = _find_device_opp(cpu_dev);
> 
> So does _find_device_opp() need to be called with rcu_read_lock()
> held or not? The comment above the function makes it sound like
> we need RCU, but we don't do that here anymore.

That is more for the readers, as this function is going to return a
pointer to the device OPP, and to make sure it isn't freed behind
their back, they need to take the RCU lock.

There are other writer code paths as well, like add-opp, where we just
take the mutex as there can't be anything stronger than that :)

-- 
viresh