On 07/10/2015 02:39 PM, Wolfram Sang wrote: > >> 60 s sounds way too much and actually I simply don't believe this is >> the root cause. If I take a look into the driver, then I see, that > > I agree, this is just a workaround. > Yes, this is a workaround. I thought this is simpler change and can go into -rc while I work on the better fix. As you can see, the other suggestions need quite a significant change to the isr code. >> the design is not really the best. The whole IRQ handling could be >> actually performed in hard IRQ handler, without threading overhead. >> Putting even 2 bytes in the controller FIFO should not be too heavy >> for the hard IRQ handler. Then these ridiculous spin_lock()s. What >> is the reason behind? The IRQ is flagged with ONESHOT, so thread and >> hardirq handler are anyway mutually excluded. But if this thread >> ever runs longer than it's allowed in IRQ context, then it anyway >> produces this IRQ latency because it locks spin_lock_irqsave() for >> the whole time! So the whole point of threaded interrupt is missing. > > Furthermore, this combination of threaded_irq and struct completion seems > bogus to me. If you just want to ensure the irq happened before timeout, > you just complete when the irq happened and do the "bottom half" after the > completion returned? This sounds good to me. I will try to implement this option. Thanks for the suggestion. > >> I would propose you to throw away spinlocks. Convert threaded IRQ to >> just one hardirq handler. And continue debugging. You will reduce the >> load of the system with the above measures, maybe it will not happen >> any more, maybe you'll figure out that problem is somewhere else. > > Or this. I am not convinced with moving entire code at hardirq context. I believe its better to keep hardirq as small as possible. -- Regards Vignesh -- To unsubscribe from this list: send the line "unsubscribe linux-i2c" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
