Re: [PATCH v2 2/2] PCI: controller: dwc: add UniPhier PCIe host controller support

[Lorenzo Pieralisi]

On Mon, Oct 08, 2018 at 11:15:59AM +0530, Kishon Vijay Abraham I wrote:
> Hi Lorenzo,
> 
> On Friday 28 September 2018 09:13 PM, Lorenzo Pieralisi wrote:
> > On Fri, Sep 28, 2018 at 02:17:16PM +0100, Marc Zyngier wrote:
> >> On 28/09/18 12:06, Lorenzo Pieralisi wrote:
> >>> [+Murali, Marc]
> >>>
> >>> On Thu, Sep 27, 2018 at 04:44:26PM +0900, Kunihiko Hayashi wrote:
> >>>> Hi Lorenzo, Gustavo,
> >>>>
> >>>> On Wed, 26 Sep 2018 21:31:36 +0900 <hayashi.kunihiko@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>>> Hi Lorenzo, Gustavo,
> >>>>>
> >>>>> Thank you for reviewing.
> >>>>>
> >>>>> On Tue, 25 Sep 2018 18:53:01 +0100
> >>>>> Gustavo Pimentel <gustavo.pimentel@xxxxxxxxxxxx> wrote:
> >>>>>
> >>>>>> On 25/09/2018 17:14, Lorenzo Pieralisi wrote:
> >>>>>>> [+Gustavo, please have a look at INTX/MSI management]
> >>>>>>>
> >>>>>>> On Thu, Sep 06, 2018 at 06:40:32PM +0900, Kunihiko Hayashi wrote:
> >>>>>>>> This introduces specific glue layer for UniPhier platform to support
> >>>>>>>> PCIe host controller that is based on the DesignWare PCIe core, and
> >>>>>>>> this driver supports Root Complex (host) mode.
> >>>>>>>
> >>>>>>> Please read this thread and apply it to next versions:
> >>>>>>>
> >>>>>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__marc.info_-3Fl-3Dlinux-2Dpci-26m-3D150905742808166-26w-3D2&d=DwIBAg&c=DPL6_X_6JkXFx7AXWqB0tg&r=bkWxpLoW-f-E3EdiDCCa0_h0PicsViasSlvIpzZvPxs&m=H8UNDDUGQnQnqfWr4CBios689dJcjxu4qeTTRGulLmU&s=CgcXc_2LThyOpW-4bCriJNo9H1lzROEdy_cG9p-Y5hU&e=
> >>>>>
> >>>>> I also found this thread in previous linux-pci, and I think it's helpful for me.
> >>>>> I'll check it carefully.
> >>>>
> >>>> [snip]
> >>>>
> >>>>>>>> +	ret = devm_request_irq(dev, pp->irq, uniphier_pcie_irq_handler,
> >>>>>>>> +			       IRQF_SHARED, "pcie", priv);
> >>>>>>>
> >>>>>>> This is wrong, you should set-up a chained IRQ for INTX.
> >>>>>>>
> >>>>>>> I *think* that
> >>>>>>>
> >>>>>>> ks_pcie_setup_interrupts()
> >>>>>>>
> >>>>>>> is a good example to start with but I wonder whether it is worth
> >>>>>>> generalizing the INTX approach to designware as a whole as it was
> >>>>>>> done for MSIs.
> >>>>>>>
> >>>>>>> Thoughts ?
> >>>>>>
> >>>>>> From what I understood this is for legacy IRQ, right?
> >>>>>
> >>>>> Yes. For legacy IRQ.
> >>>>>
> >>>>>> Like you (Lorenzo) said there is 2 drivers (pci-keystone-dw.c and pci-dra7xx.c)
> >>>>>> that uses it and can be use as a template for handling this type of interrupts.
> >>>>>>
> >>>>>> We can try to pass some kind of generic INTX function to the DesignWare host
> >>>>>> library to handling this, but this will require some help from keystone and
> >>>>>> dra7xx maintainers, since my setup doesn't have legacy IRQ HW support.
> >>>>>
> >>>>> Now I think it's difficult to make a template for INTX function,
> >>>>> and at first, I'll try to re-write this part with reference to pci-keystone-dw.c.
> >>>>
> >>>> I understand that there are 2 types of interrupt and the drivers.
> >>>>
> >>>> One like pci-keystone-dw.c is:
> >>>>
> >>>>  - there are 4 interrupts for legacy,
> >>>>  - invoke handlers for each interrupt, and handle the interrupt,
> >>>>  - call irq_set_chained_handler_and_data() to make a chain of the interrupts
> >>>>   when initializing
> >>>>
> >>>> The other like pci-dra7xx.c is:
> >>>>
> >>>>  - there is 1 IRQ for legacy as a parent,
> >>>>  - check an interrupt factor register, and handle the interrupt correspond
> >>>>    to the factor,
> >>>>  - call request_irq() for the parent IRQ and irq_domain_add_linear() for
> >>>>    the factor when initializing
> >>>>
> >>>> The pcie-uniphier.c is the same type as the latter (like pci-dra7xx.c).
> >>>>
> >>>> However, in pci-dra7xx.c, MSI and legacy IRQ share the same interrupt number,
> >>>> so the same handler is called and the handler divides these IRQs.
> >>>> (found in dra7xx_pcie_msi_irq_handler())
> >>>>
> >>>> In pcie-uniphier.c, MSI and legacy IRQ are independent.
> >>>> Therefore it's necessary to prepare the handler for the legacy IRQ.
> >>>>
> >>>> I think that it's difficult to apply the way of pci-keystone-dw.c, and
> >>>> uniphier_pcie_irq_handler() and calling devm_request_irq() are still
> >>>> necessary to handle legacy IRQ.
> >>>
> >>> I do not think it is difficult, the difference is that keystone has
> >>> 1 GIC irq line allocated per legacy IRQ, your set-up has one for
> >>> all INTX.
> >>>
> >>> *However*, I would like some clarifications from Murali on this code
> >>> in drivers/pci/controller/dwc/pci-keystone.c:
> >>>
> >>> static void ks_pcie_legacy_irq_handler(struct irq_desc *desc)
> >>> {
> >>> 	unsigned int irq = irq_desc_get_irq(desc);
> >>> 	struct keystone_pcie *ks_pcie = irq_desc_get_handler_data(desc);
> >>> 	struct dw_pcie *pci = ks_pcie->pci;
> >>> 	struct device *dev = pci->dev;
> >>> 	u32 irq_offset = irq - ks_pcie->legacy_host_irqs[0];
> >>>
> >>> Here the IRQ numbers are virtual IRQs, is it correct to consider
> >>> the virq numbers as sequential values ? The "offset" is used to
> >>> handle the PCI controller interrupt registers, so it must be a value
> >>> between 0-3 IIUC.
> >>
> >> There is absolutely no reason why virtual interrupt numbers should be
> >> contiguous. Shake the allocator hard enough, and you'll see gaps appearing.
> >>
> >> In general, the only thing that makes sense is to compute this offset based
> >> on the hwirq which is HW-specific.
> > 
> > That was my understanding and why I asked, which means that keystone
> > code can break (unless I read it wrong) and Murali will send me a fix as
> > soon as possible please to get it right (and Kunihiko will base his
> > code on this discussion).
> 
> I had cleaned up legacy interrupt handling in keystone driver [1] which was
> also required for TI's AM654 Platform.
> 
> But I guess the same issue will occur in MSI interrupt handling. I'll fix that
> up in the next version. Btw can you review [2] so that I can fix any other
> comments that you may have.

Hi Kishon,

yes I will, I am getting there (sorry for the delay), I don't think we
can make it v4.20 material but let me first have a look, maybe we can
split it up and simplify its merge.

Thanks,
Lorenzo