fix(pci): disable decoding while sizing BARs

src/arch/x86_64/kernel/apic.rs

@@ -613,6 +613,26 @@ fn ioapic_set_interrupt(irq: u8, apicid: u8, enabled: bool) {
 	ioapic_write(IOAPIC_REG_TABLE + off + 1, ioredirect_upper);
 }
 
+/// Reconfigures an interrupt line for a legacy PCI interrupt.
+///
+/// [`init_ioapic`] sets up every line for the edge triggered, active high
+/// behavior of the ISA interrupts. PCI signals its interrupts level triggered
+/// and active low instead.
+#[cfg(feature = "pci")]
+pub(crate) fn ioapic_set_pci_interrupt(irq: u8, apicid: u8) {
+	assert!(irq <= 24);
+
+	let off = u32::from(irq * 2);
+	let ioredirect_upper = u32::from(apicid) << 24;
+	// Vector, active low (bit 13) and level triggered (bit 15).
+	let ioredirect_lower = u32::from(0x20 + irq) | (1 << 13) | (1 << 15);
+
+	debug!("Configuring irq {irq} as level triggered, active low");
+
+	ioapic_write(IOAPIC_REG_TABLE + off, ioredirect_lower);
+	ioapic_write(IOAPIC_REG_TABLE + off + 1, ioredirect_upper);
+}
+
 pub fn init_local_apic() {
 	// Mask out all interrupts we don't need right now.
 	local_apic_write(IA32_X2APIC_LVT_TIMER, APIC_LVT_MASK);

src/drivers/pci.rs

@@ -73,9 +73,23 @@ impl<T: ConfigRegionAccess> PciDevice<T> {
 
 	/// Returns the bar at bar-register `slot`.
 	pub fn get_bar(&self, slot: u8) -> Option<Bar> {
-		let header = self.header();
-		let endpoint = EndpointHeader::from_header(header, &self.access)?;
-		endpoint.bar(slot, &self.access)
+		let endpoint = EndpointHeader::from_header(self.header(), &self.access)?;
+		let mut header = self.header();
+
+		// Determining the size of a bar writes all-ones into it and restores the
+		// old value afterwards. While decoding is enabled, the device decodes
+		// these intermediate values as real addresses and relocates itself. On a
+		// passed-through device the host follows that relocation and tries to map
+		// the bar at an address outside of the guest, which kills the VM. The PCI
+		// specification requires decoding to be disabled while sizing a bar.
+		let command = header.command(&self.access);
+		header.update_command(&self.access, |command| {
+			command & !(CommandRegister::IO_ENABLE | CommandRegister::MEMORY_ENABLE)
+		});
+		let bar = endpoint.bar(slot, &self.access);
+		header.update_command(&self.access, |_| command);
+
+		bar
 	}
 
 	/// Configure the bar at register `slot`
@@ -181,6 +195,15 @@ impl<T: ConfigRegionAccess> PciDevice<T> {
 					return None;
 				}
 
+				// A legacy PCI interrupt is signaled level triggered and active
+				// low, while the IOAPIC defaults to the edge triggered, active
+				// high behavior of the ISA interrupts. Without the correction a
+				// passed-through device stops delivering interrupts after its
+				// first one, because the host only unmasks the interrupt once the
+				// guest has completed a level triggered one.
+				#[cfg(target_arch = "x86_64")]
+				crate::arch::kernel::apic::ioapic_set_pci_interrupt(line, 0);
+
 				Some(line)
 			}
 			5.. => {
@@ -268,7 +291,7 @@ impl<T: ConfigRegionAccess> fmt::Display for PciDevice<T> {
 
 			let mut slot: u8 = 0;
 			while usize::from(slot) < MAX_BARS {
-				if let Some(pci_bar) = endpoint.bar(slot, &self.access) {
+				if let Some(pci_bar) = self.get_bar(slot) {
 					match pci_bar {
 						Bar::Memory64 {
 							address,