Re: [PATCH 2/2] aer: add support aer interrupt with none MSI/MSI-X/INTx mode

From: Bjorn Helgaas
Date: Fri Jun 03 2016 - 23:49:16 EST


On Fri, Jun 03, 2016 at 01:31:11PM -0400, Murali Karicheri wrote:
> Po,
>
> Sorry to hijack your discussion, but the problem seems to be same for
> Keystone PCI controller which is also designware (old version) based.
>
> On 06/03/2016 12:09 AM, Bjorn Helgaas wrote:
> > On Thu, Jun 02, 2016 at 11:37:28AM -0400, Murali Karicheri wrote:
> >> On 06/02/2016 09:55 AM, Bjorn Helgaas wrote:
> >>> On Thu, Jun 02, 2016 at 05:01:19AM +0000, Po Liu wrote:
> >>>>> -----Original Message-----
> >>>>> From: Bjorn Helgaas [mailto:helgaas@xxxxxxxxxx]
> >>>>> Sent: Thursday, June 02, 2016 11:48 AM
> >>>>> To: Po Liu
> >>>>> Cc: linux-pci@xxxxxxxxxxxxxxx; linux-arm-kernel@xxxxxxxxxxxxxxxxxxx;
> >>>>> linux-kernel@xxxxxxxxxxxxxxx; devicetree@xxxxxxxxxxxxxxx; Arnd Bergmann;
> >>>>> Roy Zang; Marc Zyngier; Stuart Yoder; Yang-Leo Li; Minghuan Lian; Bjorn
> >>>>> Helgaas; Shawn Guo; Mingkai Hu; Rob Herring
> >>>>> Subject: Re: [PATCH 2/2] aer: add support aer interrupt with none
> >>>>> MSI/MSI-X/INTx mode
> >>>>>
> >>>>> [+cc Rob]
> >>>>>
> >>>>> Hi Po,
> >>>>>
> >>>>> On Thu, May 26, 2016 at 02:00:06PM +0800, Po Liu wrote:
> >>>>> > On some platforms, root port doesn't support MSI/MSI-X/INTx in RC mode.
> >>>>> > When chip support the aer interrupt with none MSI/MSI-X/INTx mode,
> >>>>> > maybe there is interrupt line for aer pme etc. Search the interrupt
> >>>>> > number in the fdt file.
> >>>>>
> >>>>> My understanding is that AER interrupt signaling can be done via INTx,
> >>>>> MSI, or MSI-X (PCIe spec r3.0, sec 6.2.4.1.2). Apparently your device
> >>>>> doesn't support MSI or MSI-X. Are you saying it doesn't support INTx
> >>>>> either? How is the interrupt you're requesting here different from INTx?
> >>>>
> >>>> Layerscape use none of MSI or MSI-X or INTx to indicate the devices
> >>>> or root error in RC mode. But use an independent SPI interrupt(arm
> >>>> interrupt controller) line.
> >>>
> >>> The Root Port is a PCI device and should follow the normal PCI rules
> >>> for interrupts. As far as I understand, that means it should use MSI,
> >>> MSI-X, or INTx. If your Root Port doesn't use MSI or MSI-X, it should
> >>> use INTx, the PCI_INTERRUPT_PIN register should tell us which (INTA/
> >>> INTB/etc.), and PCI_COMMAND_INTX_DISABLE should work to disable it.
> >>> That's all from the PCI point of view, of course.
> >>
> >> I am faced with the same issue on Keystone PCI hardware and it has been
> >> on my TODO list for quite some time. Keystone PCI hardware also doesn't
> >> use MSI or MSI-X or INTx for reporting errors received at the root port,
> >> but use a platform interrupt instead (not complaint to PCI standard as
> >> per PCI base spec). So I would need similar change to have the error
> >> interrupt passed to the aer driver. So there are hardware out there
> >> like Keystone which requires to support this through platform IRQ.
> >
> > This is not a new area of the spec, and it's hard for me to believe
> > that these two new PCIe controllers are both broken the same way
> > (although I guess both are DesignWare-based, so maybe this is the same
> > underlying problem in both cases?). I think it's more likely that we
> > just haven't figured out the right way to describe this in the DT.
>
> Keystone is using an older version of the designware IP and it implements
> all of the interrupts in the application register space unlike other
> newer version of the hardware. So I assume, the version used on Layerscape
> is also an older version and the both have same issue in terms of
> non standard platform interrupt used for error reporting.
>
> > I assume you have a Root Port with an AER capability, no MSI
> > capability, and no MSI-X capability, right?
>
> Has AER capability and both MSI and INTx (legacy) capability
>
> > What does its Interrupt
> > Pin register contain? If it's zero, it doesn't use INTx either, so
> > according to the spec it should generate no interrupts.
> >
> At address offset 0x3C by default has a value of 1, but it is writable
> by software. So default is INTx A.

0x3c is the Interrupt *Line*, which is read/write. The Interrupt
*Pin* is at 0x3d and should be read-only.

Does your Keystone driver support MSI? If so, since your Root Port
supports MSI, I would think we would use that by default, and the INTx
stuff wouldn't even matter.

> > But if Interrupt Pin is non-zero, that means the Root Port should be
> > able to generate virtual INTx interrupts. Presumably the Root Complex
> > connects those interrupts to something; maybe to your platform
> > interrupt?
>
> Probably that is what is happening. Both Power management and Error
> interrupts are raised on platform interrupt lines.
>
> 12 Error Interrupts
>
> [0] System error (OR of fatal, nonfatal, correctable errors) (RC mode only)
> [1] PCIe fatal error (RC mode only)
> [2] PCIe non-fatal error (RC mode only)
> [3] PCIe correctable error (RC mode only)
> [4] AXI Error due to fatal condition in AXI bridge (EP/RC modes)
> [5] PCIe advanced error (RC mode only)
>
> 13 Power management and reset event interrupts
>
> [0] Power management turn-off message interrupt (EP mode only)
> [1] Power management ack message interrupt (RC mode only)
> [2] Power management event interrupt (RC mode only)
> [3] Link request reset interrupt (hot reset or link down) (RC mode only)
>
> > PCI doesn't say anything about an interrupt after it leaves the Root
> > Complex, so the fact that it's connected to a "platform interrupt" or
> > "SPI interrupt" or "IOAPIC interrupt" doesn't make it non-compliant.
> > Shouldn't we be able to use the interrupt-map and related DT
> > properties to express the fact that Root Port virtual INTx is routed
> > to platform interrupt Y, even without any special-case code in
> > portdrv?
>
> My understanding is if RC also raise interrupt on INTx, then below
> map make sense, where either EP or RC can raise interrupt and the
> line will be muxed for interrupt from EP or RC port.

I'm sorry, I didn't quite catch your meaning here, so let me try to
clarify some terminology. Maybe we'll eventually blunder into a
common understanding :)

INTx is a PCI concept and only means something in the PCI hierarchy.
The RC would *receive* virtual INTx interrupts from the PCI hierarchy
and turn them into some platform-specific interrupt (not INTx) on the
upstream side.

So strictly speaking, the RC might raise platform-specific interrupts
when it receives INTx interrupts, but it would not raise any INTx
interrupts itself.

> Here is the DT entry in PCIE keystone for Legacy interrupt mapping
> to host interrupt.
> #interrupt-cells = <1>;
> interrupt-map-mask = <0 0 0 7>;
> interrupt-map = <0 0 0 1 &pcie_intc0 0>, /* INT A */
> <0 0 0 2 &pcie_intc0 1>, /* INT B */
> <0 0 0 3 &pcie_intc0 2>, /* INT C */
> <0 0 0 4 &pcie_intc0 3>; /* INT D */

If I understand correctly, this is the mapping from the PCI world
(INTA, INTB, etc.) to the platform-specific world (pcie_intc0 0, etc.)

If a Root Port raises a virtual INTA, the RC should turn it into the
corresponding platform interrupt, i.e., GIC_SPI 48 from the
description below.

> And then
>
> pcie_intc0: legacy-interrupt-controller {
> interrupt-controller;
> #interrupt-cells = <1>;
> interrupt-parent = <&gic>;
> interrupts = <GIC_SPI 48 IRQ_TYPE_EDGE_RISING>,
> <GIC_SPI 49 IRQ_TYPE_EDGE_RISING>,
> <GIC_SPI 50 IRQ_TYPE_EDGE_RISING>,
> <GIC_SPI 51 IRQ_TYPE_EDGE_RISING>;
> };
>
> So if RC interrupt for Power management and Error interrupt is a separate
> line, how can software describe that using the above interrupt map?

I don't know anything about interrupts the RC might generate on its
own behalf. A lot of the RC behavior is not specified by the PCIe
spec because the RC is on the border between the upstream
platform-specific stuff and the downstream PCIe stuff. Is there
something in the PCIe spec about the power management and error
interrupts you're talking about? Of maybe you can point me to a
section of the Keystone spec that talks about interrupts generated by
the RC?

In any event, the AER interrupts we're looking for in portdrv are from
the Root Port, not from the RC. For INTx, my understanding is that
the RC *transforms* virtual INTx messages from the Root Port in the
PCIe domain into GIC transactions in the platform domain.

Bjorn