[Qemu-devel] [PATCH 1/2] Add Interrupt Source Override Structure to BIOS

[Qemu-devel] [PATCH 1/2] Add Interrupt Source Override Structure to BIOS

[Beth Kon]

[-- Attachment #1: Type: text/plain, Size: 1557 bytes --]

Looking into a problem with getting the HPET to work for both Windows
and Linux revealed what appears to be a problem with QEMU's interrupt
generation.  

Since QEMU supports both APIC and dual 8259 mode (as indicated by bit 0
of the MADT flags - PCAT COMPAT), the BIOS should include the Interrupt
Source Override structure to inform the OS that IRQ0 is mapped to
interrupt 2 in APIC mode, which is the standard scheme for the PIT.

What must be happening is that Linux is abiding by what the BIOS says
and expecting INTI0 from the APIC (since there is no Interrupt Source
Override). Windows must be ignoring the Interrupt Source Override
information and just expecting to see the timer interrupt on INTI2. So
the HPET had to raise both 0 and 2 to keep them both happy.

Patch 1 adds the Interrupt Source Override to BIOS, and Patch 2 modifies
ioapic_set_irq to change vector 0 to 2.

Making this change allowed Windows and Linux to both work with HPET code
that just raised IRQ0. 

That said, I don't understand why the PIT was working with the code as
it was in the case of Windows, unless in that case windows was honoring
the Interrupt Source Override (or the lack of one, I should say). I'm
not sure these patches are the right solution for all cases, but they
work for the HPET. I should add that KVM does have an Interrupt Source
Override structure in their BIOS but they do not remap 0 to 2, so maybe
I'm missing something here. Comments appreciated.

-- 
Elizabeth Kon (Beth)
IBM Linux Technology Center
Open Hypervisor Team
email: eak@us.ibm.com

[-- Attachment #2: bios_int_override.patch --]
[-- Type: text/x-patch, Size: 1986 bytes --]

Index: bochs-2.3.7/bios/rombios32.c
===================================================================
--- bochs-2.3.7.orig/bios/rombios32.c	2008-10-24 08:07:00.000000000 -0500
+++ bochs-2.3.7/bios/rombios32.c	2008-10-27 09:36:17.000000000 -0500
@@ -1262,6 +1262,17 @@
 			  * lines start */
 };
 
+#ifdef BX_QEMU
+struct madt_int_override
+{
+	APIC_HEADER_DEF
+	uint8_t                bus;     /* Identifies ISA Bus */
+	uint8_t                source;  /* Bus-relative interrupt source */
+	uint32_t               gsi;     /* GSI that source will signal */
+	uint16_t               flags;   /* MPS INTI flags */
+};
+#endif
+
 #include "acpi-dsdt.hex"
 
 static inline uint16_t cpu_to_le16(uint16_t x)
@@ -1410,7 +1421,11 @@
     madt_addr = addr;
     madt_size = sizeof(*madt) +
         sizeof(struct madt_processor_apic) * smp_cpus +
+#ifdef BX_QEMU
+        sizeof(struct madt_io_apic) + sizeof(struct madt_int_override);
+#else
         sizeof(struct madt_io_apic);
+#endif
     madt = (void *)(addr);
     addr += madt_size;
 
@@ -1480,6 +1495,9 @@
     {
         struct madt_processor_apic *apic;
         struct madt_io_apic *io_apic;
+#ifdef BX_QEMU
+        struct madt_int_override *int_override;
+#endif
 
         memset(madt, 0, madt_size);
         madt->local_apic_address = cpu_to_le32(0xfee00000);
@@ -1499,6 +1517,17 @@
         io_apic->io_apic_id = smp_cpus;
         io_apic->address = cpu_to_le32(0xfec00000);
         io_apic->interrupt = cpu_to_le32(0);
+#ifdef BX_QEMU
+        io_apic++;
+
+        int_override = (void *)io_apic;
+        int_override->type = APIC_XRUPT_OVERRIDE;
+        int_override->length = sizeof(*int_override);
+        int_override->bus = cpu_to_le32(0);
+        int_override->source = cpu_to_le32(0);
+        int_override->gsi = cpu_to_le32(2);
+        int_override->flags = cpu_to_le32(0);
+#endif
 
         acpi_build_table_header((struct acpi_table_header *)madt,
                                 "APIC", madt_size, 1);

Re: [Qemu-devel] [PATCH 1/2] Add Interrupt Source Override Structure to BIOS

[Beth Kon]

On Mon, 2008-10-27 at 14:20 -0400, Beth Kon wrote:
> Looking into a problem with getting the HPET to work for both Windows
> and Linux revealed what appears to be a problem with QEMU's interrupt
> generation.  
> 
> Since QEMU supports both APIC and dual 8259 mode (as indicated by bit 0
> of the MADT flags - PCAT COMPAT), the BIOS should include the Interrupt
> Source Override structure to inform the OS that IRQ0 is mapped to
> interrupt 2 in APIC mode, which is the standard scheme for the PIT.
> 
> What must be happening is that Linux is abiding by what the BIOS says
> and expecting INTI0 from the APIC (since there is no Interrupt Source
> Override). Windows must be ignoring the Interrupt Source Override
> information and just expecting to see the timer interrupt on INTI2. So
> the HPET had to raise both 0 and 2 to keep them both happy.
> 
> Patch 1 adds the Interrupt Source Override to BIOS, and Patch 2 modifies
> ioapic_set_irq to change vector 0 to 2.
> 
> Making this change allowed Windows and Linux to both work with HPET code
> that just raised IRQ0. 
> 
> That said, I don't understand why the PIT was working with the code as
> it was in the case of Windows, unless in that case windows was honoring
> the Interrupt Source Override (or the lack of one, I should say). I'm
> not sure these patches are the right solution for all cases, but they
> work for the HPET. I should add that KVM does have an Interrupt Source
> Override structure in their BIOS but they do not remap 0 to 2, so maybe
> I'm missing something here. Comments appreciated.
> 
Forgot ....
Signed-off-by Beth Kon <bkon@us.ibm.com>
-- 
Elizabeth Kon (Beth)
IBM Linux Technology Center
Open Hypervisor Team
email: eak@us.ibm.com