Hi, On Mon, Jul 22, 2013 at 08:25:05PM +0300, Illia Smyrnov wrote: > >please don't remove this code. It'll be good to have this around when we > >move the driver to threaded IRQs without IRQF_ONESHOT. In fact, it would > >be very simple to implement such a change, wanna take it up ? > > > >It should be doable in few patches: > > > >1) switch over to request_threaded_irq() > > > > just blind move to a thread, without hardirq handler, so > > IRQF_ONESHOT is mandatory. > > > >2) add hardirq handler > > > > read IRQSTATUS to check if our device has generated IRQs > > returning IRQ_WAKE_THREAD if true > > > >3) move 'IRQ masking logic' to hardirq handler, before returning > >IRQ_WAKE_THREAD > > > > this will let you remove IRQF_ONESHOT > > > >4) finally remove IRQF_ONESHOT > > > > this makes sure that IRQs aren't kept disabled until we have > > time to iterate over the entire keypad matrix. Only the keypad > > IRQ will be masked. > > > > Ok, but why we need to remove IRQF_ONESHOT flag for omap keypad driver? well, we might want to use this HW with the RT patchset. In that case, we want to run with IRQs disabled for as short time as possible. > The keypad IRQ isn't shared IRQ and in our case hardirq handler will > always return IRQ_WAKE_THREAD like default not entirely true, you still want to check if *this* HW generated to interrupt in case you get spurious IRQs and whatnot. > irq_default_primary_handler do. With IRQF_ONESHOT flag IRQ line will > be masked until the threaded > handler finished, but there is only keypad on this line. > > I tested two versions: > the first one - just threaded IRQs with IRQF_ONESHOT and without > specific hardirq handler. > the second version - threaded IRQs without IRQF_ONESHOT as you described. > Both versions was successfully tested on Blaze's keypad. you can't simply remove IRQF_ONESHOT, you need to mask this interrupt at the keypad level, basically you clear IRQ_ENABLE_SET (or write to IRQ_CLEAR or whatever it's called). Remember that the hardirq handler runs with IRQs disabled and what you need to guarantee is that once you go over your hardirq, you mask keypad's IRQs only. Currently, the RT patchset forces all IRQs to run as threads, but that's less than optimal. -- balbi
Attachment:
signature.asc
Description: Digital signature
