mirror of
https://github.com/Droid-VM/crosvm.git
synced 2026-11-17 06:58:03 +00:00
Gunyah has no ITS and takes a VM's irqfds only from the device tree it hands the resource manager at VM start, so a virtio-pci device was left with a shared, level INTx. --mbi-spis N gives the guest GICv3 message-based interrupts on N SPIs past the board's own: - their edge lines are registered before the VM starts (IrqChipAArch64::enable_mbi); a GICv2m-style doorbell frame (devices::MbiFrame, MSI_TYPER and MSI_SETSPI_NS) sits at 0x7000; - the GIC node gets msi-controller, mbi-ranges and mbi-alias, the PCI node msi-parent. Linux takes them as they are; edk2's GunyahMbiAcpiDxe turns them into the MADT's GIC MSI Frame for Windows; - an edge event registered on GunyahIrqChip after that is an MSI vector. route_irq records which SPI the guest's message (address = doorbell, data = INTID) names, and AddMsiRoute answers with that SPI's line (VmIrqResponse::MsiRouteEvent), so MsixConfig signals the hypervisor directly. A vector that still has its own irqfd (vhost-user) is waited on by the irq handler thread and relayed in service_irq_event. A Linux guest with virtio-blk-pci on MSI-X does 4K reads at QD1 43-87% faster than on INTx, at QD8 21-37%. On Windows NetKVM and vioinput take MSI-X; viostor on MSI-X matches its INTx-plus-poll throughput with fewer interrupts per I/O. |
History
|
|
|---|---|---|
| .. | ||
| src | ||
| Android.bp | ||
| Cargo.toml | ||