Re: [GIT PULL] virtio: fixes, features

[Michael Ellerman <mpe@ellerman.id.au>]

[ Cc += Bjorn & linux-pci ]

"Michael S. Tsirkin" <mst@redhat.com> writes:
> On Wed, Oct 12, 2022 at 05:21:24PM +1100, Michael Ellerman wrote:
>> "Michael S. Tsirkin" <mst@redhat.com> writes:
...
>> > ----------------------------------------------------------------
>> > virtio: fixes, features
>> >
>> > 9k mtu perf improvements
>> > vdpa feature provisioning
>> > virtio blk SECURE ERASE support
>> >
>> > Fixes, cleanups all over the place.
>> >
>> > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
>> >
>> > ----------------------------------------------------------------
>> > Alvaro Karsz (1):
>> >       virtio_blk: add SECURE ERASE command support
>> >
>> > Angus Chen (1):
>> >       virtio_pci: don't try to use intxif pin is zero
>> 
>> This commit breaks virtio_pci for me on powerpc, when running as a qemu
>> guest.
>> 
>> vp_find_vqs() bails out because pci_dev->pin == 0.
>> 
>> But pci_dev->irq is populated correctly, so vp_find_vqs_intx() would
>> succeed if we called it - which is what the code used to do.
>> 
>> I think this happens because pci_dev->pin is not populated in
>> pci_assign_irq().
>> 
>> I would absolutely believe this is bug in our PCI code, but I think it
>> may also affect other platforms that use of_irq_parse_and_map_pci().
>
> How about fixing this in of_irq_parse_and_map_pci then?
> Something like the below maybe?
> 
> diff --git a/drivers/pci/of.c b/drivers/pci/of.c
> index 196834ed44fe..504c4d75c83f 100644
> --- a/drivers/pci/of.c
> +++ b/drivers/pci/of.c
> @@ -446,6 +446,8 @@ static int of_irq_parse_pci(const struct pci_dev *pdev, struct of_phandle_args *
>  	if (pin == 0)
>  		return -ENODEV;
>  
> +	pdev->pin = pin;
> +
>  	/* Local interrupt-map in the device node? Use it! */
>  	if (of_get_property(dn, "interrupt-map", NULL)) {
>  		pin = pci_swizzle_interrupt_pin(pdev, pin);

That doesn't fix it in all cases, because there's an early return if
there's a struct device_node associated with the pci_dev, before we even
read the pin.

Also the pci_dev is const, and removing the const would propagate to a
few other places.

The other obvious place to fix it would be in pci_assign_irq(), as
below. That fixes this bug for me, but is otherwise very lightly tested.

cheers


diff --git a/drivers/pci/setup-irq.c b/drivers/pci/setup-irq.c
index cc7d26b015f3..0135413b33af 100644
--- a/drivers/pci/setup-irq.c
+++ b/drivers/pci/setup-irq.c
@@ -22,6 +22,15 @@ void pci_assign_irq(struct pci_dev *dev)
 	int irq = 0;
 	struct pci_host_bridge *hbrg = pci_find_host_bridge(dev->bus);
 
+	/* Make sure dev->pin is populated */
+	pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &pin);
+
+	/* Cope with illegal. */
+	if (pin > 4)
+		pin = 1;
+
+	dev->pin = pin;
+
 	if (!(hbrg->map_irq)) {
 		pci_dbg(dev, "runtime IRQ mapping not provided by arch\n");
 		return;
@@ -34,11 +43,6 @@ void pci_assign_irq(struct pci_dev *dev)
 	 * time the interrupt line passes through a PCI-PCI bridge we must
 	 * apply the swizzle function.
 	 */
-	pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &pin);
-	/* Cope with illegal. */
-	if (pin > 4)
-		pin = 1;
-
 	if (pin) {
 		/* Follow the chain of bridges, swizzling as we go. */
 		if (hbrg->swizzle_irq)