[+cc Ley Foon (altera), Thomas (aardvark), Kishon (dra7xx), Murali (keystone)]
On Tue, Sep 13, 2016 at 10:05:11AM -0500, Bjorn Helgaas wrote:
> On Tue, Sep 13, 2016 at 08:41:28AM +0100, Marc Zyngier wrote:
> > On 12/09/16 23:02, Bjorn Helgaas wrote:
> > > On Thu, Sep 01, 2016 at 05:19:55AM +0000, Bharat Kumar Gogada wrote:
> > >>>>>>> Hi Bharat,
> > >>>>>>>> @@ -561,7 +561,7 @@ static int nwl_pcie_init_irq_domain(struct
> > >>>>>>>> nwl_pcie
> > >>>>>>> *pcie)
> > >>>>>>>> }
> > >>>>>>>>
> > >>>>>>>> pcie->legacy_irq_domain = irq_domain_add_linear(legacy_intc_node,
> > >>>>>>>> - INTX_NUM,
> > >>>>>>>> + INTX_NUM + 1,
> > >>>>>>>> &legacy_domain_ops,
> > >>>>>>>> pcie);
> > >>>>>>>
> > >>>>>>> This feels like the wrong thing to do. You have INTX_NUM irqs, so
> > >>>>>>> the domain allocation should reflect this. On the other hand, the
> > >>>>>>> way the driver currently deals with mappings is quite broken
> > >>>>>>> (consistently adding 1 to
> > >>>>> the HW interrupt).
> > >>>>>>>
> > >>>>>> Hi Marc,
> > >>>>>>
> > >>>>>> Without above change I get following crash in kernel while booting.
> > >>>>>>
> > >>>>>> [ 2.441684] error: hwirq 0x4 is too large for dummy
> > >>>>>>
> > >>>>>> [ 2.441694] ------------[ cut here ]------------
> > >>>>>>
> > >>>>>> [ 2.441698] WARNING: at kernel/irq/irqdomain.c:344
> > >>>>>>
> > >>>>>> [ 2.441702] Modules linked in:
> > >>>>>>
> > >>>>>> [ 2.441706]
> > >>>>>>
> > >>>>>> [ 2.441714] CPU: 1 PID: 1 Comm: swapper/0 Not tainted 4.4.0 #8
> > >>>>>>
> > >>>>>> [ 2.441718] Hardware name: xlnx,zynqmp (DT)
> > >>>>>>
> > >>>>>> [ 2.441723] task: ffffffc071886b80 ti: ffffffc071888000 task.ti:
> > >>>>> ffffffc071888000
> > >>>>>>
> > >>>>>> [ 2.441732] PC is at irq_domain_associate+0x138/0x1c0
> > >>>>>>
> > >>>>>> [ 2.441738] LR is at irq_domain_associate+0x138/0x1c0
> > >>>>>>
> > >>>>>> In kernel/irq/irqdomain.c function irq_domain_associate
> > >>>>>>
> > >>>>>> if (WARN(hwirq >= domain->hwirq_max,
> > >>>>>> "error: hwirq 0x%x is too large for %s\n", (int)hwirq, domain-
> > >>>> name))
> > >>>>>> return -EINVAL;
> > >>>>>>
> > >>>>>> Here the hwirq and hwirq_max are equal to 4 without the above
> > >>>>>> condition
> > >>>>> (INTX_NUM + 1) due to which crash is coming.
> > >>>>>> This is happening as the legacy interrupts are starting from 1 (INTA).
> > >>>>>
> > >>>>> I understood that. I'm still persisting in saying that you have the wrong fix.
> > >>>>>
> > >>>>> Your domain should always allocate many interrupts as you have
> > >>>>> interrupt sources. These interrupts (hwirq) should be numbered from 0 to (n-
> > >>> 1).
> > >>>>
> > >>>> Agreed, but here comes the problem the hwirq for legacy interrupts
> > >>>> will start at 0x1 to 0x4 (INTA to INTD) and these values are as per
> > >>>> PCIe specification for legacy interrupts. So these cannot be numbered
> > >>>> from 0. So when 0x4 (INTD) for a multi-function device comes the crash
> > >>>> occurs.
> > >>>
> > >>> So who provides this hwirq? Who calls irq_domain_associate() with hwirq set to
> > >>> 4?
> > >>>
> > >> PCIe subsystem invokes pcibios_add_device function in arch/arm64/kernel/pci.c for every pci device.
> > >> The purpose of this function is to assign dev->irq using of_irq_parse_and_map_pci.
> > >> of_irq_parse_and_map_pci invokes of_irq_parse_pci where it reads PCI_INTERRUPT_PIN from configuration space and saves it
> > >> in parameter of struct of_phandle_args.
> > >> This structure is passed to irq_create_of_mapping where it invokes irq_create_fwspec_mapping.
> > >> irq_create_fwspec_mapping invokes irq_domain_translate and gets hwirq, here the above saved PCI_INTERRUPT_PIN value is assigned
> > >> to hwirq (*hwirq = fwspec->param[0]).
> > >> And then using this hwirq irq_create_mapping -> irq_domain_associate were invoked and mapping is created for virtual irq with this hwirq.
> > >> So for any end point PCI_INTERRUPT_PIN value starts from 0x1 to 0x4 and so hwirq starts from 0x1 to 0x4.
> > >>
> > >> So the values are more generic w.r.t to protocol, that's why hwirq will range from 0x1 to 0x4.
> > >> And then if you check pcie-altera.c they are doing this adding one in their handler and while creating legacy domain.
> > >
> > > Is this resolved yet? Marc, are you happy, or should we iterate on this
> > > again?
> >
> > Ah, sorry to have dropped the ball on this patch.
>
> No problem, I wasn't making forward progress anyway.
>
> > I guess that given that the infrastructure imposes the hwirq range on
> > the host drivers, Bharat's approach is the only way (and a number of
> > other host drivers are already slightly broken). I'll try and have a
> > look at solving this at the generic level. In the meantime:
> >
> > Acked-by: Marc Zyngier <marc.zyngier@xxxxxxx>
>
> After looking at this myself, I'm not happy with this either. It feels
> like there are bugs lurking here and we're just hiding one of them.
>
> Here are the callers of irq_domain_add_linear() for legacy INTx in
> drivers/pci/host:
>
> advk_pcie_init_irq_domain LEGACY_IRQ_NUM (4)
> dra7xx_pcie_init_irq_domain 4
> ks_dw_pcie_host_init MAX_LEGACY_IRQS (4)
> altera_pcie_init_irq_domain INTX_NUM + 1 (5)
> nwl_pcie_init_irq_domain INTX_NUM + 1 (5)
> xilinx_pcie_init_irq_domain 4
The altera change corresponding to this was 99496bd2971f ("PCI: altera: Fix
error when INTx is 4"). I should have noticed this inconsistency back
then.
Are aardvark, dra7xx, keystone, and xilinx (non-NWL) broken because they
only request 4 IRQs and only INTA, INTB, and INTC work?
> I think all of these use the of_irq_parse_and_map_pci() path you
> mentioned, so if the problem is in the way that path works, I would
> think these should *all* be requesting the same number of interrupts
> in the domain.
>
> I agree with Marc that we should request 4 IRQs, because that's what
> we need. If we can't do that for some reason, we ought to at least
> make all these callers the same.
>
> Bjorn
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
