[PATCH 08/12] rtc: omap: restore irq state after reading TC registers

[Felipe Balbi]

On Wed, Oct 22, 2014 at 12:50:11PM +0200, Johan Hovold wrote:
> On Sat, Oct 11, 2014 at 07:47:58PM -0500, Felipe Balbi wrote:
> > On Sat, Oct 11, 2014 at 12:12:01PM +0200, Johan Hovold wrote:
> > > On Fri, Oct 10, 2014 at 01:02:31PM -0500, Felipe Balbi wrote:
> > > > Hi,
> > > > 
> > > > On Thu, Oct 09, 2014 at 09:06:30PM +0200, Johan Hovold wrote:
> > > > > Make sure to restore local irq state when reading the timer/calendar
> > > > > (TC) registers, so that omap_rtc_read_time() can be called with
> > > > > interrupts disabled.
> > > > > 
> > > > > Signed-off-by: Johan Hovold <johan at kernel.org>
> > > > > ---
> > > > >  drivers/rtc/rtc-omap.c | 6 ++++--
> > > > >  1 file changed, 4 insertions(+), 2 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/rtc/rtc-omap.c b/drivers/rtc/rtc-omap.c
> > > > > index 0ef016553a97..62e2e9a9887a 100644
> > > > > --- a/drivers/rtc/rtc-omap.c
> > > > > +++ b/drivers/rtc/rtc-omap.c
> > > > > @@ -239,8 +239,10 @@ static void bcd2tm(struct rtc_time *tm)
> > > > >  
> > > > >  static int omap_rtc_read_time(struct device *dev, struct rtc_time *tm)
> > > > >  {
> > > > > +	unsigned long flags;
> > > > > +
> > > > >  	/* we don't report wday/yday/isdst ... */
> > > > > -	local_irq_disable();
> > > > > +	local_irq_save(flags);
> > > > 
> > > > you should really convert these to a real spin_lock_irq*(), that's
> > > > because local_irq* do not get re-written with RT patchset, so this
> > > > pretty much "breaks" RT.
> > > 
> > > The driver uses local_irq* throughout to guarantee registers are not
> > > read or written during an update event.
> > > 
> > > In fact, at least on AM33xx, this is not even necessary when reading (as
> > > opposed to writing) the TC registers, but I did not dare change that
> > > without knowing how the legacy platforms work in this respect.
> > > 
> > > Do you suggest doing this conversion as part of, or on top of, this
> > > series?
> > 
> > probably on top of is. Safer that way.
> 
> I didn't include this conversion in v2 as you may have noticed. In fact,
> I'm not even sure it should be done.
> 
> Using local_irq* "breaks" RT, but if we change it to spin_lock_irq* and
> RT preempts us, we may miss the 15us access window which would break the
> driver:
> 
> 	"If the ARM accesses the TC registers outside of the access
> 	 period, then the access is not guaranteed."
> 
> On am33xx this only affects updating the TC (timer/calendar) or alarm
> registers, but the driver currently disables interrupts also on reading
> the time (possibly because the early IP revisions did not have the
> shadow read registers).

that's alright, better not to change the behavior in this case. But as a
future patch, you might want to convert this to a raw_spinlock. That
won't get preempted.

-- 
balbi
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20141023/d6b894e5/attachment.sig>