[+cc Marc, since he commented on this]
On Sat, Jul 09, 2022 at 04:31:51PM +0200, Pali Rohár wrote:
> On Friday 01 July 2022 16:29:41 Pali Rohár wrote:
> > On Thursday 23 June 2022 11:27:47 Bjorn Helgaas wrote:
> > > On Tue, May 24, 2022 at 02:28:17PM +0200, Pali Rohár wrote:
> > > > Same as in commit a3b69dd0ad62 ("Revert "PCI: aardvark: Rewrite IRQ code to
> > > > chained IRQ handler"") for pci-aardvark driver, use devm_request_irq()
> > > > instead of chained IRQ handler in pci-mvebu.c driver.
> > > >
> > > > This change fixes affinity support and allows to pin interrupts from
> > > > different PCIe controllers to different CPU cores.
> > >
> > > Several other drivers use irq_set_chained_handler_and_data(). Do any
> > > of them need similar changes? The commit log suggests that using
> > > chained IRQ handlers breaks affinity support. But perhaps that's not
> > > the case and the real culprit is some other difference between mvebu
> > > and the other drivers.
> >
> > And there is another reason to not use irq_set_chained_handler_and_data
> > and instead use devm_request_irq(). Armada XP has some interrupts
> > shared and it looks like that irq_set_chained_handler_and_data() API
> > does not handle shared interrupt sources too.
> >
> > I can update commit message to mention also this fact.
>
> Anything needed from me to improve this fix?
My impression from Marc's response [1] was that this patch would
"break the contract the kernel has with userspace" and he didn't think
this was acceptable. But maybe I'm not understanding it correctly.
In any event, I'm waiting for you to continue that discussion. Maybe
there's an argument for doing this even though it breaks some
userspace expectations. If so, that should be acknowledged and
explained. Or maybe there's an alternative implementation. Marc
gave a link to some suggestions [2], which I haven't looked into, but
maybe you could.
[1] https://lore.kernel.org/r/874k0bf7f7.wl-maz@xxxxxxxxxx
[2] https://lore.kernel.org/all/20220502102137.764606ee@thinkpad/
> > > > Fixes: ec075262648f ("PCI: mvebu: Implement support for legacy INTx interrupts")
> > > > Signed-off-by: Pali Rohár <pali@xxxxxxxxxx>
> > > > ---
> > > > Hello Bjorn! This is basically same issue as for pci-aardvark.c:
> > > > https://lore.kernel.org/linux-pci/20220515125815.30157-1-pali@xxxxxxxxxx/#t
> > > >
> > > > I tested this patch with pci=nomsi in cmdline (to force kernel to use
> > > > legacy intx instead of MSI) on A385 and checked that I can set affinity
> > > > via /proc/irq/XX/smp_affinity file for every mvebu pcie controller to
> > > > different CPU and legacy interrupts from different cards/controllers
> > > > were handled by different CPUs.
> > > >
> > > > I think that this is important on Armada XP platforms which have many
> > > > independent PCIe controllers (IIRC up to 10) and many cores (up to 4).
> > > > ---
> > > > drivers/pci/controller/pci-mvebu.c | 30 +++++++++++++++++-------------
> > > > 1 file changed, 17 insertions(+), 13 deletions(-)
> > > >
> > > > diff --git a/drivers/pci/controller/pci-mvebu.c b/drivers/pci/controller/pci-mvebu.c
> > > > index 8f76d4bda356..de67ea39fea5 100644
> > > > --- a/drivers/pci/controller/pci-mvebu.c
> > > > +++ b/drivers/pci/controller/pci-mvebu.c
> > > > @@ -1017,16 +1017,13 @@ static int mvebu_pcie_init_irq_domain(struct mvebu_pcie_port *port)
> > > > return 0;
> > > > }
> > > >
> > > > -static void mvebu_pcie_irq_handler(struct irq_desc *desc)
> > > > +static irqreturn_t mvebu_pcie_irq_handler(int irq, void *arg)
> > > > {
> > > > - struct mvebu_pcie_port *port = irq_desc_get_handler_data(desc);
> > > > - struct irq_chip *chip = irq_desc_get_chip(desc);
> > > > + struct mvebu_pcie_port *port = arg;
> > > > struct device *dev = &port->pcie->pdev->dev;
> > > > u32 cause, unmask, status;
> > > > int i;
> > > >
> > > > - chained_irq_enter(chip, desc);
> > > > -
> > > > cause = mvebu_readl(port, PCIE_INT_CAUSE_OFF);
> > > > unmask = mvebu_readl(port, PCIE_INT_UNMASK_OFF);
> > > > status = cause & unmask;
> > > > @@ -1040,7 +1037,7 @@ static void mvebu_pcie_irq_handler(struct irq_desc *desc)
> > > > dev_err_ratelimited(dev, "unexpected INT%c IRQ\n", (char)i+'A');
> > > > }
> > > >
> > > > - chained_irq_exit(chip, desc);
> > > > + return status ? IRQ_HANDLED : IRQ_NONE;
> > > > }
> > > >
> > > > static int mvebu_pcie_map_irq(const struct pci_dev *dev, u8 slot, u8 pin)
> > > > @@ -1490,9 +1487,20 @@ static int mvebu_pcie_probe(struct platform_device *pdev)
> > > > mvebu_pcie_powerdown(port);
> > > > continue;
> > > > }
> > > > - irq_set_chained_handler_and_data(irq,
> > > > - mvebu_pcie_irq_handler,
> > > > - port);
> > > > +
> > > > + ret = devm_request_irq(dev, irq, mvebu_pcie_irq_handler,
> > > > + IRQF_SHARED | IRQF_NO_THREAD,
> > > > + port->name, port);
> > > > + if (ret) {
> > > > + dev_err(dev, "%s: cannot register interrupt handler: %d\n",
> > > > + port->name, ret);
> > > > + irq_domain_remove(port->intx_irq_domain);
> > > > + pci_bridge_emul_cleanup(&port->bridge);
> > > > + devm_iounmap(dev, port->base);
> > > > + port->base = NULL;
> > > > + mvebu_pcie_powerdown(port);
> > > > + continue;
> > > > + }
> > > > }
> > > >
> > > > /*
> > > > @@ -1599,7 +1607,6 @@ static int mvebu_pcie_remove(struct platform_device *pdev)
> > > >
> > > > for (i = 0; i < pcie->nports; i++) {
> > > > struct mvebu_pcie_port *port = &pcie->ports[i];
> > > > - int irq = port->intx_irq;
> > > >
> > > > if (!port->base)
> > > > continue;
> > > > @@ -1615,9 +1622,6 @@ static int mvebu_pcie_remove(struct platform_device *pdev)
> > > > /* Clear all interrupt causes. */
> > > > mvebu_writel(port, ~PCIE_INT_ALL_MASK, PCIE_INT_CAUSE_OFF);
> > > >
> > > > - if (irq > 0)
> > > > - irq_set_chained_handler_and_data(irq, NULL, NULL);
> > > > -
> > > > /* Remove IRQ domains. */
> > > > if (port->intx_irq_domain)
> > > > irq_domain_remove(port->intx_irq_domain);
> > > > --
> > > > 2.20.1
> > > >
