From: YASUOKA Masahiko <yasuoka@openbsd.org>
Subject: Re: PCI amd64: Enables MSI/MSI-X interrupts on QEMU i440fx chipset emulation
To: mark.kettenis@xs4all.nl, jsg@jsg.id.au, sf@openbsd.org
Cc: naito.yuichiro@gmail.com, kettenis@openbsd.org, tech@openbsd.org
Date: Mon, 07 Oct 2024 17:04:26 +0900

On Mon, 07 Oct 2024 09:06:37 +0200
Mark Kettenis <mark.kettenis@xs4all.nl> wrote:
>> Date: Mon, 07 Oct 2024 10:54:14 +0900 (JST)
>> From: Yuichiro NAITO <naito.yuichiro@gmail.com>
>> 
>> Hi, this patch sets 'PCI_FLAGS_MSI_ENABLED' if the kernel runs on a QEMU
>> virtual machine with i440fx chipset emulation. The i440fx is the default
>> chipset on the QEMU so users will use it first.
>> And then ixv(4) and iavf(4) will fail because MSI-X interrupts are not
>> supported by the kernel.
>> 
>> The i440fx is quite old and the ACPI table is revision 1. We cannot know
>> if it has MSI capability from the ACPI table. So I add the flag as a quirk.
>> 
>> The same quirk is implemented in NetBSD.
>> 
>> https://github.com/NetBSD/src/blob/affb7619d18b17a184eede63a4823f629a2893fc/sys/arch/x86/pci/pci_machdep.c#L273
>> 
>> FreeBSD seems to have the quirk but is used as the exception for blacklist
>> devices. Probably they see the PCI device capability header of MSI/MSI-X.
>> Until we see the same capability, my patch will be useful for the VF driver
>> users.
>> 
>> OK?
> 
> I think this should be done using the subsystem vendir/id stuff like
> was suggested by someone else in the thread.

That was by sf@.
I added the check to jsg's diff.  I tested the diff fixes the problem.


Index: sys/arch/amd64/pci/acpipci.c
===================================================================
RCS file: /disk/cvs/openbsd/src/sys/arch/amd64/pci/acpipci.c,v
diff -u -p -u -p -r1.8 acpipci.c
--- sys/arch/amd64/pci/acpipci.c	13 May 2024 01:15:50 -0000	1.8
+++ sys/arch/amd64/pci/acpipci.c	7 Oct 2024 07:51:39 -0000
@@ -176,6 +176,7 @@ acpipci_attach_bus(struct device *parent
 	struct pcibus_attach_args pba;
 	pcitag_t tag;
 	pcireg_t id, class;
+	int device, maxdevs;
 
 	memset(&pba, 0, sizeof(pba));
 	pba.pba_busname = "pci";
@@ -193,6 +194,22 @@ acpipci_attach_bus(struct device *parent
 	if (sc->sc_acpi->sc_fadt->hdr.revision >= 2 &&
 	    (sc->sc_acpi->sc_fadt->iapc_boot_arch & FADT_NO_MSI) == 0)
 		pba.pba_flags |= PCI_FLAGS_MSI_ENABLED;
+
+	/* For hypervisors claiming ACPI 1.0. */
+	if (sc->sc_acpi->sc_fadt->hdr.revision == 1 &&
+	    cpu_ecxfeature & CPUIDECX_HV) {
+		/* Enable MSI if it's QEMU */
+		maxdevs = pci_bus_maxdevs(pba.pba_pc, sc->sc_bus);
+		for (device = 0; device < maxdevs; device++) {
+			tag = pci_make_tag(pba.pba_pc, sc->sc_bus, device,
+			    0);
+			id = pci_conf_read(pba.pba_pc, tag, PCI_ID_REG);
+			if (PCI_VENDOR(id) == PCI_VENDOR_QUMRANET) {
+				pba.pba_flags |= PCI_FLAGS_MSI_ENABLED;
+				break;
+			}
+		}
+	}
 
 	/*
 	 * Don't enable MSI on chipsets from low-end manufacturers