On Fri, Aug 02, 2024 at 01:13:19PM +0530, Manivannan Sadhasivam wrote:
> On Thu, Aug 01, 2024 at 12:23:08PM -0500, Bjorn Helgaas wrote:
> > On Wed, Jul 31, 2024 at 04:20:09PM +0530, Manivannan Sadhasivam via B4 Relay wrote:
> > > From: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
> > >
> > > Currently, the IRQ device name for both of these IRQs doesn't have Qcom
> > > specific prefix and PCIe domain number. This causes 2 issues:
> > >
> > > 1. Pollutes the global IRQ namespace since 'global' is a common name.
> > > 2. When more than one EP controller instance is present in the SoC, naming
> > > conflict will occur.
> > >
> > > Hence, add 'qcom_pcie_ep_' prefix and PCIe domain number suffix to the IRQ
> > > names to uniquely identify the IRQs and also to fix the above mentioned
> > > issues.
> > >
> > > Reviewed-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
> > > ---
> > > drivers/pci/controller/dwc/pcie-qcom-ep.c | 16 ++++++++++++++--
> > > 1 file changed, 14 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > index 0bb0a056dd8f..d0a27fa6fdc8 100644
> > > --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > @@ -711,8 +711,15 @@ static irqreturn_t qcom_pcie_ep_perst_irq_thread(int irq, void *data)
> > > static int qcom_pcie_ep_enable_irq_resources(struct platform_device *pdev,
> > > struct qcom_pcie_ep *pcie_ep)
> > > {
> > > + struct device *dev = pcie_ep->pci.dev;
> > > + char *name;
> > > int ret;
> > >
> > > + name = devm_kasprintf(dev, GFP_KERNEL, "qcom_pcie_ep_global_irq%d",
> > > + pcie_ep->pci.ep.epc->domain_nr);
> > > + if (!name)
> > > + return -ENOMEM;
> >
> > I assume this is what shows up in /proc/interrupts?
>
> Yes.
>
> > I always wonder
> > why it doesn't include dev_name(). A few drivers do that, but
> > apparently it's not universally desirable. It's sort of annoying
> > that, e.g., we get a bunch of "aerdrv" interrupts with no clue which
> > device they relate to.
>
> dev_name() can be big at times. I wouldn't recommend to include it
> unless there are no other ways to differentiate between IRQs.
> Luckily PCIe has the domain number that we can use to differentiate
> these IRQs.
/proc/interrupts is 159 characters wide even on my puny 8 CPU laptop,
so I don't think width is a strong argument, and having to use
per-device heuristics (instance number like dmarX, idma64.X, nvmeXqY,
domain number, etc) to find the related device is ... well, a hassle.
But like I said, obviously devm_request_threaded_irq() *could* have
been implemented to include dev_name() internally but wasn't, so I
acknowledge there must be good reasons not to, and I'm fine with this
patch as-is since it continues the existing practice.
> > > pcie_ep->global_irq = platform_get_irq_byname(pdev, "global");
> > > if (pcie_ep->global_irq < 0)
> > > return pcie_ep->global_irq;
> > > @@ -720,18 +727,23 @@ static int qcom_pcie_ep_enable_irq_resources(struct platform_device *pdev,
> > > ret = devm_request_threaded_irq(&pdev->dev, pcie_ep->global_irq, NULL,
> > > qcom_pcie_ep_global_irq_thread,
> > > IRQF_ONESHOT,
> > > - "global_irq", pcie_ep);
> > > + name, pcie_ep);
