Samuel Thibault, le lun. 02 oct. 2023 01:56:51 +0200, a ecrit: > Samuel Thibault, le lun. 02 oct. 2023 01:50:14 +0200, a ecrit: > > Samuel Thibault, le lun. 02 oct. 2023 01:43:48 +0200, a ecrit: > > > Damien Zammit, le dim. 01 oct. 2023 23:26:02 +0000, a ecrit: > > > > I think the logic for this should be: > > > > > > > > When we get irq N, first we mask irq N, then EOI irq N. > > > > Then call the handler. If there is a user handler for irq N, let the > > > > irq_ack > > > > unmask irq N, otherwise we need to unmask irq N now. > > > > But don't EOI in the user handlers anymore. > > > > > > > > What do you think? > > > > > > That can be a plan, yes, but what about the legacy irqs, and their > > > in-kernel handlers? We need to fix them too. One keypoint there is that > > > we want to unmask with IF cleared, so that we don't get interrupted > > > again on unmask, but on iret (so we don't nest). > > > > For the legacy irqs and their in-kernel handlers, we don't really need > > to mask actually, we can just keep IF cleared, i.e. the spl level raise > > should be enough. But we want to keep IF cleared when lowering it again, > > IIRC that's already handled so. > > So in the end, I'd tend to think that it's up to queue_intr to do the > unmasking, like it does now, while spl is still high and thus we don't > risk nesting. That way it's the in-kernel intr handler that knows > whether to mask/unmask or not. I.e. it'd be like this: > > - interrupt.S raises spl (thus IF cleared) > - interrupt.S EOI > - interrupt.S calls the handler > - for pure in-kernel handlers, they do whatever they want with IF > cleared. > - when a userland handler is registers, queue_intr masks the irq. > - interrupt.S lowers spl with splx_cli, thus IF still cleared > - iret, that clears the IF > > - later on, userland acks the IRQ, that unmasks the irq > > Is that not already very close to what we have currently?
(of course we want to record that reasoning in the source code) Samuel
