Messages in this thread | | | Subject | Re: [RFC PATCH v1 2/4] irqchip: GICv3: set non-percpu irqs status with _IRQ_MOVE_PCNTXT | From | Jiang Liu <> | Date | Tue, 8 Sep 2015 00:33:28 +0800 |
| |
On 2015/9/7 22:56, Marc Zyngier wrote: > Hi Thomas, > > On 07/09/15 14:24, Thomas Gleixner wrote: >> On Mon, 7 Sep 2015, Marc Zyngier wrote: >>> On 06/09/15 06:56, Jiang Liu wrote: >>>> On 2015/9/6 12:23, Yang Yingliang wrote: >>>>> Use irq_settings_set_move_pcntxt() helper irqs status with >>>>> _IRQ_MOVE_PCNTXT. So that it can do set affinity when calling >>>>> irq_set_affinity_locked(). >>>> Hi Yingliang, >>>> We could only set _IRQ_MOVE_PCNTCT flag to enable migrating >>>> IRQ in process context if your hardware platform supports atomically >>>> change IRQ configuration. Not sure whether that's true for GICv3. >>>> If GICv3 doesn't support atomically change irq configuration, this >>>> change may cause trouble. >>> >>> I think it boils down to what exactly "process context" means here. If >>> this means "we do not need to mask the interrupt" while moving it, then >>> it should be fine (the GIC architecture guarantees that a pending >>> interrupt will be migrated). >>> >>> Is there any other requirement for this flag? >> >> The history of this flag is as follows: >> >> On x86 interrupts can only be safely migrated while the interrupt is >> handled. > > Woa! That's creative! :-) I suppose this doesn't work very well with CPU > hotplug though... X86 has special handling of this case when hot-removing a CPU. Basically, it does: 1) mask an irq 2) migrate irq to other cpus with set_affinity 3) redirect(retrigger) irq to other CPUs if it's pending on the CPU to be removed. Thanks! Gerry
> >> With the introduction of IRQ remapping this requirement >> changed. Remapped interrupts can be migrated in any context. >> >> If you look at irq_set_affinity_locked() >> >> if (irq_can_move_pcntxt(data) { >> irq_do_set_affinity(data,...) >> chip->irq_set_affinity(data,...); >> } else { >> irqd_set_move_pending(data); >> } >> >> So if IRQ_MOVE_PCNTXT is not set, we handle the migration of the >> interrupt from next the interrupt. If it's set set_affinity() is >> called right away. > > OK, that is now starting to make more sense. > >> All architectures which do not select GENERIC_PENDING_IRQ are using >> the direct method. > > Right. On ARM, only the direct method makes sense so far (we have no > constraint such as the one you describe above). > > So I wonder why we bother introducing the IRQ_MOVE_PCNTXT flag on ARM at > all. Is that just because migration.c is only compiled when > GENERIC_PENDING_IRQ is set? > > Thanks, > > M. >
| |