Re: [PATCH 10/11] PCI: mvebu: Implement support for legacy INTx interrupts

[Marc Zyngier]

On Thu, 06 Jan 2022 15:44:47 +0000,
Pali Rohár <pali@xxxxxxxxxx> wrote:
> 
> On Thursday 06 January 2022 15:28:20 Marc Zyngier wrote:
> > On Wed, 05 Jan 2022 15:02:38 +0000,
> > Pali Rohár <pali@xxxxxxxxxx> wrote:
> > > 
> > > This adds support for legacy INTx interrupts received from other PCIe
> > > devices and which are reported by a new INTx irq chip.
> > > 
> > > With this change, kernel can distinguish between INTA, INTB, INTC and INTD
> > > interrupts.
> > > 
> > > Note that for this support, device tree files has to be properly adjusted
> > > to provide "interrupts" or "interrupts-extended" property with intx
> > > interrupt source, "interrupt-names" property with "intx" string and also
> > > 'interrupt-controller' subnode must be defined.
> > > 
> > > If device tree files do not provide these nodes then driver would work as
> > > before.
> > > 
> > > Signed-off-by: Pali Rohár <pali@xxxxxxxxxx>
> > > ---
> > >  drivers/pci/controller/pci-mvebu.c | 182 +++++++++++++++++++++++++++--
> > >  1 file changed, 174 insertions(+), 8 deletions(-)
> > > 
> > > diff --git a/drivers/pci/controller/pci-mvebu.c b/drivers/pci/controller/pci-mvebu.c
> > > index 1e90ab888075..04bcdd7b7a6d 100644
> > > --- a/drivers/pci/controller/pci-mvebu.c
> > > +++ b/drivers/pci/controller/pci-mvebu.c
> > > @@ -54,9 +54,10 @@
> > >  	 PCIE_CONF_ADDR_EN)
> > >  #define PCIE_CONF_DATA_OFF	0x18fc
> > >  #define PCIE_INT_CAUSE_OFF	0x1900
> > > +#define PCIE_INT_UNMASK_OFF	0x1910
> > > +#define  PCIE_INT_INTX(i)		BIT(24+i)
> > >  #define  PCIE_INT_PM_PME		BIT(28)
> > > -#define PCIE_MASK_OFF		0x1910
> > > -#define  PCIE_MASK_ENABLE_INTS          0x0f000000
> > > +#define  PCIE_INT_ALL_MASK		GENMASK(31, 0)
> > >  #define PCIE_CTRL_OFF		0x1a00
> > >  #define  PCIE_CTRL_X1_MODE		0x0001
> > >  #define  PCIE_CTRL_RC_MODE		BIT(1)
> > > @@ -110,6 +111,10 @@ struct mvebu_pcie_port {
> > >  	struct mvebu_pcie_window iowin;
> > >  	u32 saved_pcie_stat;
> > >  	struct resource regs;
> > > +	struct irq_domain *intx_irq_domain;
> > > +	struct irq_chip intx_irq_chip;
> > 
> > Why is this structure per port? It really should be global. Printing
> > the port number in the name isn't enough of a reason.
> 
> Because each port has its own independent set of INTA-INTD
> interrupts.

That doesn't warrant a copy of an irq_chip structure that contains the
exact same callbacks, and only differs by *a string*. And the use of
this string is only to end-up in /proc/interrupts, which is totally
pointless.

> 
> > > +	raw_spinlock_t irq_lock;
> > > +	int intx_irq;
> > >  };
> > >  
> > >  static inline void mvebu_writel(struct mvebu_pcie_port *port, u32 val, u32 reg)
> > > @@ -235,7 +240,7 @@ static void mvebu_pcie_setup_wins(struct mvebu_pcie_port *port)
> > >  
> > >  static void mvebu_pcie_setup_hw(struct mvebu_pcie_port *port)
> > >  {
> > > -	u32 ctrl, lnkcap, cmd, dev_rev, mask;
> > > +	u32 ctrl, lnkcap, cmd, dev_rev, unmask;
> > >  
> > >  	/* Setup PCIe controller to Root Complex mode. */
> > >  	ctrl = mvebu_readl(port, PCIE_CTRL_OFF);
> > > @@ -288,10 +293,30 @@ static void mvebu_pcie_setup_hw(struct mvebu_pcie_port *port)
> > >  	/* Point PCIe unit MBUS decode windows to DRAM space. */
> > >  	mvebu_pcie_setup_wins(port);
> > >  
> > > -	/* Enable interrupt lines A-D. */
> > > -	mask = mvebu_readl(port, PCIE_MASK_OFF);
> > > -	mask |= PCIE_MASK_ENABLE_INTS;
> > > -	mvebu_writel(port, mask, PCIE_MASK_OFF);
> > > +	/* Mask all interrupt sources. */
> > > +	mvebu_writel(port, ~PCIE_INT_ALL_MASK, PCIE_INT_UNMASK_OFF);
> > > +
> > > +	/* Clear all interrupt causes. */
> > > +	mvebu_writel(port, ~PCIE_INT_ALL_MASK, PCIE_INT_CAUSE_OFF);
> > > +
> > > +	if (port->intx_irq <= 0) {
> > > +		/*
> > > +		 * When neither "summary" interrupt, nor "intx" interrupt was
> > > +		 * specified in DT then unmask all legacy INTx interrupts as in
> > > +		 * this case driver does not provide a way for masking and
> > > +		 * unmasking of individual legacy INTx interrupts. In this case
> > > +		 * all interrupts, including legacy INTx are reported via one
> > > +		 * shared GIC source and therefore kernel cannot distinguish
> > > +		 * which individual legacy INTx was triggered. These interrupts
> > > +		 * are shared, so it should not cause any issue. Just
> > > +		 * performance penalty as every PCIe interrupt handler needs to
> > > +		 * be called when some interrupt is triggered.
> > > +		 */
> > > +		unmask = mvebu_readl(port, PCIE_INT_UNMASK_OFF);
> > > +		unmask |= PCIE_INT_INTX(0) | PCIE_INT_INTX(1) |
> > > +			  PCIE_INT_INTX(2) | PCIE_INT_INTX(3);
> > > +		mvebu_writel(port, unmask, PCIE_INT_UNMASK_OFF);
> > 
> > Maybe worth printing a warning here, so that the user knows they are
> > on thin ice.
> 
> Ok. I can add it here. Anyway, this is default current state without
> this patch.
> 
> > > +	}
> > >  }
> > >  
> > >  static struct mvebu_pcie_port *mvebu_pcie_find_port(struct mvebu_pcie *pcie,
> > > @@ -924,6 +949,109 @@ static struct pci_ops mvebu_pcie_ops = {
> > >  	.write = mvebu_pcie_wr_conf,
> > >  };
> > >  
> > > +static void mvebu_pcie_intx_irq_mask(struct irq_data *d)
> > > +{
> > > +	struct mvebu_pcie_port *port = d->domain->host_data;
> > > +	irq_hw_number_t hwirq = irqd_to_hwirq(d);
> > > +	unsigned long flags;
> > > +	u32 unmask;
> > > +
> > > +	raw_spin_lock_irqsave(&port->irq_lock, flags);
> > > +	unmask = mvebu_readl(port, PCIE_INT_UNMASK_OFF);
> > > +	unmask &= ~PCIE_INT_INTX(hwirq);
> > > +	mvebu_writel(port, unmask, PCIE_INT_UNMASK_OFF);
> > > +	raw_spin_unlock_irqrestore(&port->irq_lock, flags);
> > > +}
> > > +
> > > +static void mvebu_pcie_intx_irq_unmask(struct irq_data *d)
> > > +{
> > > +	struct mvebu_pcie_port *port = d->domain->host_data;
> > > +	irq_hw_number_t hwirq = irqd_to_hwirq(d);
> > > +	unsigned long flags;
> > > +	u32 unmask;
> > > +
> > > +	raw_spin_lock_irqsave(&port->irq_lock, flags);
> > > +	unmask = mvebu_readl(port, PCIE_INT_UNMASK_OFF);
> > > +	unmask |= PCIE_INT_INTX(hwirq);
> > > +	mvebu_writel(port, unmask, PCIE_INT_UNMASK_OFF);
> > > +	raw_spin_unlock_irqrestore(&port->irq_lock, flags);
> > > +}
> > > +
> > > +static int mvebu_pcie_intx_irq_map(struct irq_domain *h,
> > > +				   unsigned int virq, irq_hw_number_t hwirq)
> > > +{
> > > +	struct mvebu_pcie_port *port = h->host_data;
> > > +
> > > +	irq_set_status_flags(virq, IRQ_LEVEL);
> > > +	irq_set_chip_and_handler(virq, &port->intx_irq_chip, handle_level_irq);
> > > +	irq_set_chip_data(virq, port);
> > > +
> > > +	return 0;
> > > +}
> > > +
> > > +static const struct irq_domain_ops mvebu_pcie_intx_irq_domain_ops = {
> > > +	.map = mvebu_pcie_intx_irq_map,
> > > +	.xlate = irq_domain_xlate_onecell,
> > > +};
> > > +
> > > +static int mvebu_pcie_init_irq_domain(struct mvebu_pcie_port *port)
> > > +{
> > > +	struct device *dev = &port->pcie->pdev->dev;
> > > +	struct device_node *pcie_intc_node;
> > > +
> > > +	raw_spin_lock_init(&port->irq_lock);
> > > +
> > > +	port->intx_irq_chip.name = devm_kasprintf(dev, GFP_KERNEL,
> > > +						  "mvebu-%s-INTx",
> > > +						  port->name);
> > 
> > That's exactly what I really don't want to see. It prevents sharing of
> > the irq_chip structure, and gets in the way of making it const in the
> > future. Yes, I know that some drivers do that. I can't fix those,
> > because /proc/interrupts is ABI. But I really don't want to see more
> > of these.
> 
> Well, I do not understand why it should be shared and with who. HW has N
> independent IRQ chips for legacy interrupts. And each one will be
> specified in DT per HW layout / design.

If you have multiple ports, all the ports can share the irq_chip
structure. Actually scratch that. They *MUST* share the structure. The
only reason you're not sharing it is to be able to print this useless
string in /proc/interrupts.

	M.

-- 
Without deviation from the norm, progress is not possible.