Re: [PATCH V7 2/3] ACPI: Add support for ResourceSource/IRQ domain mapping

[Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>]

Hi Agustin,

On Mon, Nov 28, 2016 at 05:40:24PM -0500, Agustin Vega-Frias wrote:
> Hi Rafael,
> 
> Can you chime in on Lorenzo's feedback and the discussion below?
> It would be great if you can comment on the reason ACPI does things
> in a certain way.
> 
> Hi Lorenzo,
> 
> On 2016-11-25 06:40, Lorenzo Pieralisi wrote:
> >Hi Agustin,
> >
> >On Thu, Nov 24, 2016 at 04:15:48PM +0000, Lorenzo Pieralisi wrote:
> >
> >[...]
> >
> >>> @@ -448,6 +449,7 @@ bool acpi_dev_resource_interrupt(struct acpi_resource *ares, int index,
> >>>  {
> >>>  	struct acpi_resource_irq *irq;
> >>>  	struct acpi_resource_extended_irq *ext_irq;
> >>> +	struct fwnode_handle *src;
> >>>
> >>>  	switch (ares->type) {
> >>>  	case ACPI_RESOURCE_TYPE_IRQ:
> >>> @@ -460,7 +462,7 @@ bool acpi_dev_resource_interrupt(struct acpi_resource *ares, int index,
> >>>  			acpi_dev_irqresource_disabled(res, 0);
> >>>  			return false;
> >>>  		}
> >>> -		acpi_dev_get_irqresource(res, irq->interrupts[index],
> >>> +		acpi_dev_get_irqresource(res, irq->interrupts[index], NULL,
> >>>  					 irq->triggering, irq->polarity,
> >>>  					 irq->sharable, true);
> >>>  		break;
> >>> @@ -470,7 +472,8 @@ bool acpi_dev_resource_interrupt(struct acpi_resource *ares, int index,
> >>>  			acpi_dev_irqresource_disabled(res, 0);
> >>>  			return false;
> >>>  		}
> >>> -		acpi_dev_get_irqresource(res, ext_irq->interrupts[index],
> >>> +		src = acpi_get_irq_source_fwhandle(&ext_irq->resource_source);
> >>
> >>Is there a reason why we need to do the domain look-up here ?
> 
> Because we need to pass the resource down to acpi_dev_get_irqresource
> which does the mapping through acpi_register_irq/acpi_register_gsi.
> 
> >>
> >>I would like to understand if, by reshuffling the code (and by
> >>returning
> >>the resource_source to the calling code - somehow), it would be
> >>possible
> >>to just mirror what the OF code does in of_irq_get(), namely:
> >>
> >>(1) parse the irq entry -> of_irq_parse_one()
> >>(2) look the domain up -> irq_find_host()
> >>(3) create the mapping -> irq_create_of_mapping()
> >>
> >>You wrote the code already, I think it is just a matter of shuffling
> >>it around (well, minus returning the resource_source to the caller
> >>which is phandle equivalent in DT).
> 
> This is one area in which DT and ACPI are fundamentally different. In DT
> once the flattened blob is expanded the data is fixed. In ACPI the data
> returned by a method can change. In reality most methods like CRS return
> constants, but given that per-spec they are methods the interpreter has
> to be involved, which makes it an expensive operation. I believe that is
> the reason the resource parsing code in ACPI attempts all mappings
> during
> the bus scan. Rafael can you comment on this?
> 
> One way to do what you suggest would be to defer IRQ mapping by, e.g.,
> populating res->start with the HW IRQ number and res->end with the
> fwnode.
> That way we can avoid having to walk the resource buffer when a mapping
> is needed. I don't think that approach would deviate much more from
> the spec from what the current ahead-of-time mapping does, but it would
> require more changes in the core code. An alternative would be to do
> that only for resources that fail to map.
> 
> >>
> >>You abstracted away (2) and (3) behind acpi_register_irq(), that
> >>on anything than does not use ACPI_GENERIC_GSI is just glue code
> >>to acpi_register_gsi().
> >>
> >>Also, it is not a question on this patch but I ask it here because it
> >>is related. On ACPI you are doing the reverse of what is done in
> >>DT in platform_get_irq():
> >>
> >>- get the resources already parsed -> platform_get_resource()
> >>- if they are disabled -> acpi_irq_get()
> >>
> >>and I think the ordering is tied to my question above because
> >>you carry out the domain look up in acpi_dev_resource_interrupt()
> >>so that if for any reason it fails the corresponding resource
> >>is disabled so that we try to get it again through acpi_irq_get().
> >>
> >>I suspect you did it this way to make sure:
> >>
> >>a) keep the current ACPI IRQ parsing interface changes to a mininum
> >>b) avoid changing the behaviour on x86/ia64; in particular, calling
> >>   acpi_register_gsi() for the _same_ mapping (an IRQ that was already
> >>   registered at device creation resource parsing) multiple times can
> >>   trigger issues on x86/ia64
> 
> You are correct about my reasons. I wanted to keep ACPI core code
> changes
> to a minimum, and I also needed to work within the current
> implementation
> which uses the pre-converted IRQ resources.
> 
> >>
> >>I think that's a reasonable approach but I wanted to get these
> >>clarifications, I do not think you are far from getting this
> >>done but since it is a significant change I think it is worth
> >>discussing the points I raised above because I think the DT code
> >>sequence in of_irq_get() (1-2-3 above) is cleaner from an IRQ
> >>layer perspective (instead of having the domain look-up buried
> >>inside the ACPI IRQ resource parsing API).
> >
> >I had another look and to achieve the above one way of doing that is to
> >implement acpi_irq_get() only for ACPI_GENERIC_GSI and stub it out for
> >!ACPI_GENERIC_GSI (ie return an error code so that on !ACPI_GENERIC_GSI
> >we would fall back to current solution for ACPI). Within acpi_irq_get()
> >you can easily carry out the same steps (1->2->3) above in ACPI
> >you have
> >the code already there I think it is easy to change the
> >acpi_irq_get_cb() interface to return a filled in struct irq_fwspec and
> >the interface would become identical to of_irq_get() that is an
> >advantage to maintain it from an IRQ maintainership perspective I
> >think,
> >that's my opinion.
> 
> I think I get what you mean. I'll take a stab at implementing
> acpi_irq_get()
> in the way you suggest.
> 
> >
> >There is still a nagging snag though. When platform devices are
> >created, core ACPI code parse the resources through:
> >
> >acpi_dev_get_resources()
> >
> >and we _have_ to have way to avoid initializing IRQ resources that
> >have a dependency (ie there is a resource_source pointer that is valid
> >in their descriptors) that's easy to do if we think that's the right
> >thing to do and can hardly break current code (which ignores the
> >resource_source altogether).
> 
> I'd rather keep the core code as-is with regard to the ahead-of-time
> conversion. Whether a resource source is available at the time of
> the bus
> scan should be transparent to the code in drivers/acpi/resource.c, and
> we need the initialization as a disabled resource to signal the need
> to retry anyway.

Yes, exactly that's the nub. Your current code works, I am trying to
make it more modular and similar to the DT/irqdomain IRQ look-up path,
which has its advantages.

There are two options IMHO:

- always disable the resource if it has a resource_source dependency and defer
  its parsing to acpi_irq_get() (where you can easily implement steps 1-2-3 above).
  What I wanted to say is that, by disabling the resource if it has a
  resource_source dependency you can't break x86/ia64 (it is ignored at
  present - hopefully there is nothing that we are not aware of behind
  that choice). On x86/ia64 acpi_irq_get() would be an empty stub.
  This way you would keep the irqdomain look-up out of the ACPI resource
  parsing API, correct ?
- keep code as-is

Your point on _CRS being _current_ resource setting is perfectly valid
so platform_get_resource() in platform_get_irq() must always take
precedence over acpi_irq_get() (which should just apply to disabled
resources), I am not sure that doing it the other way around is safe.

> Rafael, do you have any other suggestions/feedback on how to go about
> doing this?

Yes, comments very appreciated, these changes are not trivial and need
agreement.

Thanks,
Lorenzo

> 
> Thanks,
> Agustin
> 
> >
> >It is an important difference with DT probing, where the IRQ
> >resources are only created if the domain reference (ie interrupt
> >controller phandle) is satisfied at of_device_alloc() time
> >(see of_device_alloc()).
> >
> >Thoughts ? Please let me know, the code to implement what I say
> >is already in these patches, it is just a matter of reshuffling it.
> >
> >Thanks !
> >Lorenzo
> 
> -- 
> Qualcomm Datacenter Technologies, Inc. on behalf of the Qualcomm
> Technologies, Inc.
> Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a
> Linux Foundation Collaborative Project.