[Xenomai] Problem with ipipe locking PPIs and SGIs globally not per cpu
abel.vesa at nxp.com
Thu May 31 09:25:48 CEST 2018
The problem is the fact that the arm twd irq (for example) is masked on that
specific core inside gic, but the ipipe is locking it globally.
According to GIC architecture specification
Interrupt numbers ID0-ID31 are used for interrupts that are private to a CPU interface. These interrupts are
banked in the Distributor.
A banked interrupt is one where the Distributor can have multiple interrupts with the same ID. A banked
interrupt is identified uniquely by its ID number and its associated CPU interface number. Of the banked
— ID0-ID15 are used for SGIs
— ID16-ID31 are used for PPIs
In a multiprocessor system:
— A PPI is forwarded to a particular CPU interface, and is private to that interface. In prioritizing
interrupts for a CPU interface the Distributor does not consider PPIs that relate to other interfaces.
In our case, we're talking about ID29 (linux mapped 16). In kernel/ipipe/core.c,
in function __ipipe_lock_irq the control bits should either be set per cpu and
not globally if the hwirq is less than 31.
A suggestion would be to use the NR_CPUS bits from ipd->irqs[irq].control to
mark the lock on a per cpu basis and then depending on the type of the irq
set the corresponding bits for each (or all, in case SPIs) cpu. And then, when
checking if locked (in __ipipe_set_irq_pending and __ipipe_do_sync_stage)
include the cpu in the condition. That should be all.
I'll try to come up with a patch later. I just wanted to point out the problem
More information about the Xenomai