Re: [PATCH v2 2/4] pinctrl: rockchip: convert to raw spinlock

[Julia Cartwright <julia@ni.com>]

[-- Attachment #1: Type: text/plain, Size: 2463 bytes --]

On Wed, Mar 15, 2017 at 07:16:53PM +0100, Heiko Stuebner wrote:
> Am Mittwoch, 15. März 2017, 18:08:06 CET schrieb John Keeping:
> > On Wed, 15 Mar 2017 13:01:37 -0500, Julia Cartwright wrote:
> > > On Wed, Mar 15, 2017 at 05:46:52PM +0000, John Keeping wrote:
> > > > This lock is used from rockchip_irq_set_type() which is part of the
> > > > irq_chip implementation and thus must use raw_spinlock_t as documented
> > > > in Documentation/gpio/driver.txt.
> > > > 
> > > > Signed-off-by: John Keeping <john@metanate.com>
> > > > Reviewed-by: Heiko Stuebner <heiko@sntech.de>
> > > > Tested-by: Heiko Stuebner <heiko@sntech.de>
> > > > ---
> > > > v2: unchanged
> > > > ---
> > > > 
> > > >  drivers/pinctrl/pinctrl-rockchip.c | 30 +++++++++++++++---------------
> > > >  1 file changed, 15 insertions(+), 15 deletions(-)
> > > > 
> > > > diff --git a/drivers/pinctrl/pinctrl-rockchip.c
> > > > b/drivers/pinctrl/pinctrl-rockchip.c index 128c383ea7ba..8c1cae6d78d7
> > > > 100644
[..]
> > > > @@ -1295,14 +1295,14 @@ static int rockchip_set_pull(struct
> > > > rockchip_pin_bank *bank,> > 
> > > >  	switch (ctrl->type) {
> > > > 
> > > >  	case RK2928:
> > > > -		spin_lock_irqsave(&bank->slock, flags);
> > > > +		raw_spin_lock_irqsave(&bank->slock, flags);
> > > > 
> > > >  		data = BIT(bit + 16);
> > > >  		if (pull == PIN_CONFIG_BIAS_DISABLE)
> > > >  		
> > > >  			data |= BIT(bit);
> > > 
> > > This should be lifted out from under the lock.
> > > 
> > > >  		ret = regmap_write(regmap, reg, data);
> > > 
> > > How is this legal?  The regmap_write() here is going to end up acquiring
> > > the regmap mutex.
> > 
> > It's not, the spinlock can be deleted here.  I only have RK3288 hardware
> > to test and I missed this when checking the uses of slock.
> 
> That part could very well also use regmap_update_bits like the other parts.
> Not really sure, why we use regmap_write here, but I'm also not sure, if it 
> matters at all.

regmap_update_bits also acquires the regmap lock, which would similarly
be a problem here.[1]

But, if we could pull this entire operation out of the lock (and
convince ourselves that it's okay to do so), then even better!

   Julia

1: Why is this a problem?  Because we're in the middle of a
   raw_spinlock_t protected critical region: if there were contention on
   the nested mutex (the "regmap mutex"), then we'd attempt to sleep in
   atomic context.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]