On 02/09/15 11:48, Qais Yousef wrote:
On 09/02/2015 10:55 AM, Marc Zyngier wrote:
On 02/09/15 10:33, Qais Yousef wrote:
On 08/28/2015 03:22 PM, Thomas Gleixner wrote:
On Fri, 28 Aug 2015, Qais Yousef wrote:
Thanks a lot for the detailed explanation. I wasn't looking for a quick and dirty solution but my view of the problem is much simpler than yours so my idea of a solution would look quick and dirty. I have a better appreciation of the problem now and a way to approach it :-)
From DT point of view are we OK with this form then
coprocessor { interrupt-source = <&intc INT_SPEC COP_HWAFFINITY>; interrupt-sink = <&intc INT_SPEC CPU_HWAFFINITY>; }and if the root controller sends normal IPI as it sends normal device interrupts then interrupt-sink can be a standard interrupts property (like in my case)
coprocessor { interrupt-source = <&intc INT_SPEC COP_HWAFFINITY>; interrupts = <INT_SPEC>; }Does this look right to you? Is there something else that needs to be covered still?
I'm not an DT wizard. I leave that to the DT experts.
Hi Marc Zyngier, Mark Rutland,
Any comments about the DT binding for the IPIs?
To recap, the proposal which is based on Marc Zyngier's is to use interrupt-source to represent an IPI from Linux CPU to a coprocessor and interrupt-sink to receive an IPI from coprocessor to Linux CPU. Hopefully the description above is self explanatory. Please let me know if you need more info. Thomas covered the routing, synthesising, and requesting parts in the core code. The remaining (high level) issue is how to describe the IPIs in DT.
I'm definitely *not* a DT expert! ;-) My initial binding proposal was only for wired interrupts, not for IPIs. There is definitely some common aspects, except for one part:
Who decides on the IPI number? So far, we've avoided encoding IPI numbers in the DT just like we don't encode MSIs, because they are programmable things. My feeling is that we shouldn't put the IPI number in the DT because the rest of the kernel uses them as well and could decide to use this particular IPI number for its own use: *clash*.
I think this is covered in Thomas proposal to reserve IPIs. His thoughts is to use a separate irq-domain for IPIs and use irq_reserve_ipi() and irq_destroy_ipi() to get and release IPIs.
The way I see it would be to have a pool of IPI numbers that the kernel requests for its own use first, leaving whatever remains to drivers.
That's what Thomas thinks too and he covered this by using irq_reserve_ipi() and irq_destroy_ipi().
https://lkml.org/lkml/2015/8/26/713
Ah, I missed that, sorry for the noise. This looks very sensible.
It's worth noting in the light of this that INT_SPEC should be optional since for hardware similar to mine there's not much to tell the controller if it's all dynamic except where we want the IPI to be routed to - the INT_SPEC is implicitly defined by the notion it's an IPI.
Well, I'd think that the INT_SPEC should say that it is an IPI, and I don't believe we should omit it. On the ARM GIC side, our interrupts are typed (type 0 is a normal wired interrupt, type 1 a per-cpu interrupt, and we could allocate type 2 to identify an IPI).
But we do need to identify it properly, as we should be able to cover both IPIs and normal wired interrupts.
Thanks,
M.