On Wed, Aug 31, 2016 at 11:05:10AM -0500, Bjorn Helgaas wrote: > On Wed, Aug 31, 2016 at 02:34:54PM +0100, Lorenzo Pieralisi wrote: > > On Wed, Aug 31, 2016 at 08:05:06AM -0500, Bjorn Helgaas wrote: > > > On Tue, Aug 30, 2016 at 11:08:34AM +0100, Lorenzo Pieralisi wrote: > > > > On Fri, Aug 26, 2016 at 06:53:29PM -0400, Sinan Kaya wrote: > > > > > > > > > > >> Let me throw option d here. > > > > > >> > > > > > >> I know Bjorn wants to keep ACTIVE_LOW in the code for common code but > > > > > >> can't we override this in an arch specific way (arm64's pci.c) while > > > > > >> creating the root bridge? > > > > > > > > > > > > On what basis ? You were not copied in from the beginning, but that > > > > > > is not different from Duc's initial proposal, which Marc discarded > > > > > > because it should not be done at arch level, it depends on the interrupt > > > > > > controller. > > > > > > > > > > I happen to watch the linux-pci and linux-acpi mail-lists. I also saw > > > > > Duc's initial proposal. > > > > > > > > > > I can't imagine someone building an ACPI compliant ARM64 platform > > > > > without a GIC interrupt controller. > > > > > > > > > > The SBSA spec already mentions what kind of compatibility should be > > > > > maintained with respect to GIC. You can't have an ACPI system that's > > > > > SBSA compliant and not using GIC. > > > > > > > > > > Can't we just hard code this to ACTIVE_HIGH in arch directory if ACPI > > > > > is defined. Why do we have to reach out to the interrupt controller? > > > > > > > > Patch below (horrible but no solution will be shiny either). > > > > > > > > > https://lists.linaro.org/pipermail/linaro-acpi/2015-November/005973.html > > > > > > > > [...] > > > > > > > > > If you look at my email above, I tried getting rid of PCI Link object > > > > > and I couldn't. I sticked to only thing that works. > > > > > > > > That's what I object to. If the ACPI bindings do not work for ARM > > > > we do not work around issues, we upgrade the specs because what may work > > > > for you has to work on all ARM platforms (and all FW developers have > > > > to be aware of that). Granted, this is a tiny snag, but the point is > > > > that no one knows what's the correct way of describing PCI legacy IRQs > > > > on ARM and we need that rectified. > > > > > > > > This does the trick for me (I can turn it into a function/look-up > > > > that returns the polarity), I am sure it will ruffle feathers but > > > > we have to find a solution so here it is (that acpi_irq_model gem > > > > is already used in generic code drivers/acpi/pci_link.c ;-)) > > > > > > > > -- >8 -- > > > > diff --git a/drivers/acpi/pci_irq.c b/drivers/acpi/pci_irq.c > > > > index 2c45dd3..c9b8c85 100644 > > > > --- a/drivers/acpi/pci_irq.c > > > > +++ b/drivers/acpi/pci_irq.c > > > > @@ -411,7 +411,8 @@ int acpi_pci_irq_enable(struct pci_dev *dev) > > > > int gsi; > > > > u8 pin; > > > > int triggering = ACPI_LEVEL_SENSITIVE; > > > > - int polarity = ACPI_ACTIVE_LOW; > > > > + int polarity = acpi_irq_model == ACPI_IRQ_MODEL_GIC ? > > > > + ACPI_ACTIVE_HIGH : ACPI_ACTIVE_LOW; > > > > char *link = NULL; > > > > char link_desc[16]; > > > > int rc; > > > > > > This still seems weird to me. > > > > > > If I understand correctly, this GIC has several inputs, all active > > > high. Some of those inputs are connected to inverters and then to PCI > > > INTx wires. > > > > > > A generic device driver knows about the hardware it drives, including > > > the properties of its interrupt wires. PCI drivers and the ACPI/PCI > > > core know that conventional PCI device INTx wires are active low. > > > These drivers, being generic, do not know about the GIC inverters. > > > > > > The patch above basically says "if ACPI tells us about a PCI interrupt > > > connected to a GIC, *assume* there is an inverter on the input." But > > > there's no actual description of that inverter anywhere in ACPI or a > > > device tree. Shouldn't that be made explicit somewhere? > > > > It is explicit for all IRQs other than PCI legacy IRQs ;-), that's > > the message I wanted to get across and I failed so far. > > > > For "normal" IRQs we can use Extended Interrupt Descriptors, that allow > > us to describe polarity. For PCI legacy IRQs we could use extended > > interrupt descriptors if we were allowed to use PCI interrupt links, > > except that, according to current ACPI specs, PCI interrupt links can > > only be used for *configurable* interrupt pins. > > > > So, some ARM vendors stuck to the static/hardwired configuration > > in the _PRT, that does not give us a chance to describe polarity. > > > > I think that we should be allowed to use interrupt links, but that would > > not comply with the specs (and on top of that there is FW that is > > already shipped in ACPI tables that we can't change anymore). > > > > > If we connect a non-PCI device to a GIC, we need to know whether > > > there's an inverter. How could we figure that out? > > > > Through an Extended Interrupt Descriptor. How are we solving this ? > > OK, I think I'm convinced. What in the spec says you can't use PCI > Interrupt Links for this case? I see the example in ACPI 5.0 sec > 6.2.12 that only shows them changing interrupt numbers with Interrupt > descriptors. Is there something that prohibits Extended Interrupt > descriptors for PNP0C0F devices? Is there something in the code that > doesn't handle that? ACPI 6.1 (6.2.13, page 335), that describes the two ways _PRT can be used. The problem is not about whether we can use Extended Interrupt Descriptors for PNP0C0F, we *can* use them in their (CRS,SRS,PRS), the problem is that by our specs reading, PNP0C0F PCI interrupt link devices can *only* describe interrupt pins that are configurable; the GIC pins are not (ie every PCI INTx is connected to a specific GIC pin and can't be routed dynamically - it has an inverter in the path though ;-)). I suspect that's because a) PNP0C0F devices can have a _PRS (that can be empty) b) that's how IRQ chips work on x86. What (some) ARM vendors did, given the above, is hardcode the _PRT source field to 0 and add the GSI in the source index (as per specs); that does not work because it implies level low, and this thread is the outcome. So, there are two niggles to sort out: 1) Kernel code handles PCI interrupt links and we do not need anything on top of that; we do need to update the specs to allow ARM platforms to use them though 2) We still need a point hack (as the one I inlined) to cater for platforms that in current FW do NOT describe legacy IRQs with PCI interrupt link devices, is the patch I inlined ok ? Thanks ! Lorenzo -- To unsubscribe from this list: send the line "unsubscribe linux-acpi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
