RE: [PATCH 3/3] PCI: Xilinx NWL PCIe: Fix Error for multi function device for legacy interrupts.

From: Bharat Kumar Gogada
Date: Wed Sep 14 2016 - 01:34:36 EST


> [+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?
>
Hi Bjorn,

xilinx (non-NWL) will also work with all 4 INTA, INTB, INTc and INTD, but I haven't sent patches
for this yet, because by that time marc has already raised issue regarding the fix w.r.t this. So
once a proper fix was agreed upon for Xilinx-nwl, I will send fix for this also.

Thanks & Regards,
Bharat