[PATCH 9/9] ARM: tegra: fix sleeping while atomic in CPU idle

[Michał Mirosław]

On Thu, Jul 20, 2017 at 01:45:12PM +0100, Jon Hunter wrote:
> 
> On 20/07/17 01:29, Michał Mirosław wrote:
> > This removes unnecessary lock causing following BUG splat:
> > 
> > BUG: sleeping function called from invalid context at /linux/kernel/locking/mutex.c:747
> > in_atomic(): 1, irqs_disabled(): 128, pid: 0, name: swapper/0
> > no locks held by swapper/0/0.
> > CPU: 0 PID: 0 Comm: swapper/0 Tainted: G        W 4.13.0-rc1mq-00016-gbf3f627c6b2b #1
> > Hardware name: NVIDIA Tegra SoC (Flattened Device Tree)
> > [<b0110a14>] (unwind_backtrace) from [<b010c710>] (show_stack+0x18/0x1c)
> > [<b010c710>] (show_stack) from [<b068ada8>] (dump_stack+0x84/0x98)
> > [<b068ada8>] (dump_stack) from [<b014ff54>] (___might_sleep+0x158/0x17c)
> > [<b014ff54>] (___might_sleep) from [<b06a15c4>] (__mutex_lock+0x34/0x9a0)
> > [<b06a15c4>] (__mutex_lock) from [<b06a30dc>] (mutex_lock_nested+0x24/0x2c)
> > [<b06a30dc>] (mutex_lock_nested) from [<b041f6d8>] (tegra_powergate_is_powered+0x40/0xa4)
> > [<b041f6d8>] (tegra_powergate_is_powered) from [<b04197e8>] (tegra30_cpu_rail_off_ready+0x28/0x6c)
> > [<b04197e8>] (tegra30_cpu_rail_off_ready) from [<b011c73c>] (tegra30_idle_lp2+0x70/0x114)
> > [<b011c73c>] (tegra30_idle_lp2) from [<b0516080>] (cpuidle_enter_state+0x12c/0x47c)
> > [<b0516080>] (cpuidle_enter_state) from [<b0163394>] (do_idle+0x1b4/0x204)
> > [<b0163394>] (do_idle) from [<b016368c>] (cpu_startup_entry+0x20/0x24)
> > [<b016368c>] (cpu_startup_entry) from [<b0900dd8>] (start_kernel+0x39c/0x3a8)
> > 
> > Signed-off-by: Michał Mirosław <mirq-linux at rere.qmqm.pl>
> > ---
> >  drivers/soc/tegra/pmc.c | 2 --
> >  1 file changed, 2 deletions(-)
> > 
> > diff --git a/drivers/soc/tegra/pmc.c b/drivers/soc/tegra/pmc.c
> > index e233dd5dcab3..3e6ec9fdba41 100644
> > --- a/drivers/soc/tegra/pmc.c
> > +++ b/drivers/soc/tegra/pmc.c
> > @@ -526,9 +526,7 @@ int tegra_powergate_is_powered(unsigned int id)
> >  	if (!tegra_powergate_is_valid(id))
> >  		return -EINVAL;
> >  
> > -	mutex_lock(&pmc->powergates_lock);
> >  	status = tegra_powergate_state(id);
> > -	mutex_unlock(&pmc->powergates_lock);
> >  
> >  	return status;
> >  }
> 
> Thanks for the fix. However, I would prefer that we fix this the following
> way ...
> 
> diff --git a/drivers/clk/tegra/clk-tegra30.c b/drivers/clk/tegra/clk-tegra30.c
> index a2d163f759b4..546d2b121c39 100644
> --- a/drivers/clk/tegra/clk-tegra30.c
> +++ b/drivers/clk/tegra/clk-tegra30.c
> @@ -1152,9 +1152,9 @@ static bool tegra30_cpu_rail_off_ready(void)
>  
>         cpu_rst_status = readl(clk_base +
>                                 TEGRA30_CLK_RST_CONTROLLER_CPU_CMPLX_STATUS);
> -       cpu_pwr_status = tegra_powergate_is_powered(TEGRA_POWERGATE_CPU1) ||
> -                        tegra_powergate_is_powered(TEGRA_POWERGATE_CPU2) ||
> -                        tegra_powergate_is_powered(TEGRA_POWERGATE_CPU3);
> +       cpu_pwr_status = tegra_pmc_cpu_is_powered(1) ||
> +                        tegra_pmc_cpu_is_powered(2) ||
> +                        tegra_pmc_cpu_is_powered(3);
>  
>         if (((cpu_rst_status & 0xE) != 0xE) || cpu_pwr_status)
>                 return false;
> diff --git a/drivers/soc/tegra/pmc.c b/drivers/soc/tegra/pmc.c
> index e233dd5dcab3..f61371ea3fe0 100644
> --- a/drivers/soc/tegra/pmc.c
> +++ b/drivers/soc/tegra/pmc.c
> @@ -605,7 +605,7 @@ bool tegra_pmc_cpu_is_powered(unsigned int cpuid)
>         if (id < 0)
>                 return false;
>  
> -       return tegra_powergate_is_powered(id);
> +       return tegra_powergate_state(id);
>  }
> 
> 
> Could you try the above and make sure that this works? 

Something ate TABs in your patch, but yes, it works.

> I would prefer to leave the mutex in tegra_powergate_is_powered() so it could
> be used by any driver. Or maybe we could even get rid of
> tegra_powergate_is_powered() if we don't really need it.

This mutex does not protect anything here, though - there is only one register
read inside the locked section, and after unlock other CPU might change
it before the read value gets returned.

Code size-wise it would be better to remove tegra_pmc_cpu_is_powered()
instead.

Best Regards,
Michał Mirosław