fix(msi): allocate GSI for MSI and "legacy IRQ" from different ranges

Currently, we're limited to 24 GSI lines, which is too little for PCI
devices. Keep the current ranges as "legacy GSI", and create a new range
for "MSI GSI" that goes up to the kvm theoretical maximum of 4096 lines.

Signed-off-by: Riccardo Mancini <[email protected]>

Author: Manciukic

src/vmm/src/arch/aarch64/gic/gicv2/mod.rs

@@ -135,9 +135,9 @@ impl GICv2 {
         // On arm there are 3 types of interrupts: SGI (0-15), PPI (16-31), SPI (32-1020).
         // SPIs are used to signal interrupts from various peripherals accessible across
         // the whole system so these are the ones that we increment when adding a new virtio device.
-        // KVM_DEV_ARM_VGIC_GRP_NR_IRQS sets the highest SPI number. Consequently, we will have a
-        // total of `super::layout::IRQ_MAX - 32` usable SPIs in our microVM.
-        let nr_irqs: u32 = super::layout::IRQ_MAX;
+        // KVM_DEV_ARM_VGIC_GRP_NR_IRQS sets the number of interrupts (SGI, PPI, and SPI).
+        // Consequently, we need to add 32 to the number of SPIs ("legacy GSI").
+        let nr_irqs: u32 = crate::arch::GSI_LEGACY_NUM + super::layout::SPI_START;
         let nr_irqs_ptr = &nr_irqs as *const u32;
         Self::set_device_attribute(
             gic_device.device_fd(),

src/vmm/src/arch/aarch64/gic/gicv2/regs/dist_regs.rs

@@ -8,7 +8,7 @@ use kvm_ioctls::DeviceFd;
 
 use crate::arch::aarch64::gic::GicError;
 use crate::arch::aarch64::gic::regs::{GicRegState, MmioReg, SimpleReg, VgicRegEngine};
-use crate::arch::{IRQ_BASE, IRQ_MAX};
+use crate::arch::{GSI_LEGACY_NUM, SPI_START};
 
 // Distributor registers as detailed at page 75 from
 // https://developer.arm.com/documentation/ihi0048/latest/.
@@ -62,9 +62,9 @@ impl MmioReg for SharedIrqReg {
         // read-as-zero/write-ignore (RAZ/WI) policy.
         // The first part of a shared-irq register, the one corresponding to the
         // SGI and PPI IRQs (0-32) is RAZ/WI, so we skip it.
-        let start = self.offset + u64::from(IRQ_BASE) * u64::from(self.bits_per_irq) / 8;
+        let start = self.offset + u64::from(SPI_START) * u64::from(self.bits_per_irq) / 8;
 
-        let size_in_bits = u64::from(self.bits_per_irq) * u64::from(IRQ_MAX - IRQ_BASE);
+        let size_in_bits = u64::from(self.bits_per_irq) * u64::from(GSI_LEGACY_NUM);
         let mut size_in_bytes = size_in_bits / 8;
         if size_in_bits % 8 > 0 {
             size_in_bytes += 1;

src/vmm/src/arch/aarch64/gic/gicv3/mod.rs

@@ -184,9 +184,9 @@ impl GICv3 {
         // On arm there are 3 types of interrupts: SGI (0-15), PPI (16-31), SPI (32-1020).
         // SPIs are used to signal interrupts from various peripherals accessible across
         // the whole system so these are the ones that we increment when adding a new virtio device.
-        // KVM_DEV_ARM_VGIC_GRP_NR_IRQS sets the highest SPI number. Consequently, we will have a
-        // total of `super::layout::IRQ_MAX - 32` usable SPIs in our microVM.
-        let nr_irqs: u32 = super::layout::IRQ_MAX;
+        // KVM_DEV_ARM_VGIC_GRP_NR_IRQS sets the number of interrupts (SGI, PPI, and SPI).
+        // Consequently, we need to add 32 to the number of SPIs ("legacy GSI").
+        let nr_irqs: u32 = crate::arch::GSI_LEGACY_NUM + super::layout::SPI_START;
         let nr_irqs_ptr = &nr_irqs as *const u32;
         Self::set_device_attribute(
             gic_device.device_fd(),

src/vmm/src/arch/aarch64/gic/gicv3/regs/dist_regs.rs

@@ -8,7 +8,7 @@ use kvm_ioctls::DeviceFd;
 
 use crate::arch::aarch64::gic::GicError;
 use crate::arch::aarch64::gic::regs::{GicRegState, MmioReg, SimpleReg, VgicRegEngine};
-use crate::arch::{IRQ_BASE, IRQ_MAX};
+use crate::arch::{GSI_LEGACY_NUM, SPI_START};
 
 // Distributor registers as detailed at page 456 from
 // https://static.docs.arm.com/ihi0069/c/IHI0069C_gic_architecture_specification.pdf.
@@ -64,9 +64,9 @@ impl MmioReg for SharedIrqReg {
         // read-as-zero/write-ignore (RAZ/WI) policy.
         // The first part of a shared-irq register, the one corresponding to the
         // SGI and PPI IRQs (0-32) is RAZ/WI, so we skip it.
-        let start = self.offset + u64::from(IRQ_BASE) * u64::from(self.bits_per_irq) / 8;
+        let start = self.offset + u64::from(SPI_START) * u64::from(self.bits_per_irq) / 8;
 
-        let size_in_bits = u64::from(self.bits_per_irq) * u64::from(IRQ_MAX - IRQ_BASE);
+        let size_in_bits = u64::from(self.bits_per_irq) * u64::from(GSI_LEGACY_NUM);
         let mut size_in_bytes = size_in_bits / 8;
         if size_in_bits % 8 > 0 {
             size_in_bytes += 1;

src/vmm/src/arch/aarch64/layout.rs

@@ -76,19 +76,25 @@ pub const FDT_MAX_SIZE: usize = 0x20_0000;
 // * bigger than 32
 // * less than 1023 and
 // * a multiple of 32.
-/// The highest usable SPI on aarch64.
-pub const IRQ_MAX: u32 = 128;
-
-/// First usable interrupt on aarch64.
-pub const IRQ_BASE: u32 = 32;
-
-// The Linux kernel automatically shifts the GSI by 32 if it is an SPI,
-// allowing us to start numbering from 0 instead of 32.
-/// The first usable GSI on aarch64.
-pub const GSI_BASE: u32 = 0;
-
-/// The maximum usable GSI on aarch64.
-pub const GSI_MAX: u32 = IRQ_MAX - IRQ_BASE - 1;
+// The first 32 SPIs are reserved, but KVM already shifts the gsi we
+// pass, so we go from 0 to 95 for legacy gsis ("irq") and the remaining
+// we use for MSI.
+/// Offset of first SPI in the GIC
+pub const SPI_START: u32 = 32;
+/// Last possible SPI in the GIC (128 total SPIs)
+pub const SPI_END: u32 = 127;
+/// First usable GSI id on aarch64 (corresponds to SPI #32).
+pub const GSI_LEGACY_START: u32 = 0;
+/// There are 128 SPIs available, but the first 32 are reserved
+pub const GSI_LEGACY_NUM: u32 = SPI_END - SPI_START + 1;
+/// Last available GSI
+pub const GSI_LEGACY_END: u32 = GSI_LEGACY_START + GSI_LEGACY_NUM - 1;
+/// First GSI used by MSI after legacy GSI
+pub const GSI_MSI_START: u32 = GSI_LEGACY_END + 1;
+/// The highest available GSI in KVM (KVM_MAX_IRQ_ROUTES=4096)
+pub const GSI_MSI_END: u32 = 4095;
+/// Number of GSI available for MSI.
+pub const GSI_MSI_NUM: u32 = GSI_MSI_END - GSI_MSI_START + 1;
 
 /// The start of the memory area reserved for MMIO 32-bit accesses.
 /// Below this address will reside the GIC, above this address will reside the MMIO devices.

src/vmm/src/arch/mod.rs

@@ -21,13 +21,14 @@ pub use aarch64::vm::{ArchVm, ArchVmError, VmState};
 #[cfg(target_arch = "aarch64")]
 pub use aarch64::{
     ConfigurationError, arch_memory_regions, configure_system_for_boot, get_kernel_start,
-    initrd_load_addr, layout::BOOT_DEVICE_MEM_START, layout::CMDLINE_MAX_SIZE, layout::GSI_BASE,
-    layout::GSI_MAX, layout::IRQ_BASE, layout::IRQ_MAX, layout::MEM_32BIT_DEVICES_SIZE,
+    initrd_load_addr, layout::BOOT_DEVICE_MEM_START, layout::CMDLINE_MAX_SIZE,
+    layout::GSI_LEGACY_END, layout::GSI_LEGACY_NUM, layout::GSI_LEGACY_START, layout::GSI_MSI_END,
+    layout::GSI_MSI_NUM, layout::GSI_MSI_START, layout::MEM_32BIT_DEVICES_SIZE,
     layout::MEM_32BIT_DEVICES_START, layout::MEM_64BIT_DEVICES_SIZE,
     layout::MEM_64BIT_DEVICES_START, layout::MMIO32_MEM_SIZE, layout::MMIO32_MEM_START,
     layout::PCI_MMCONFIG_SIZE, layout::PCI_MMCONFIG_START,
     layout::PCI_MMIO_CONFIG_SIZE_PER_SEGMENT, layout::RTC_MEM_START, layout::SERIAL_MEM_START,
-    layout::SYSTEM_MEM_SIZE, layout::SYSTEM_MEM_START, load_kernel,
+    layout::SPI_START, layout::SYSTEM_MEM_SIZE, layout::SYSTEM_MEM_START, load_kernel,
 };
 
 /// Module for x86_64 related functionality.
@@ -45,7 +46,8 @@ pub use x86_64::vm::{ArchVm, ArchVmError, VmState};
 pub use crate::arch::x86_64::{
     ConfigurationError, arch_memory_regions, configure_system_for_boot, get_kernel_start,
     initrd_load_addr, layout::APIC_ADDR, layout::BOOT_DEVICE_MEM_START, layout::CMDLINE_MAX_SIZE,
-    layout::GSI_BASE, layout::GSI_MAX, layout::IOAPIC_ADDR, layout::IRQ_BASE, layout::IRQ_MAX,
+    layout::GSI_LEGACY_END, layout::GSI_LEGACY_NUM, layout::GSI_LEGACY_START, layout::GSI_MSI_END,
+    layout::GSI_MSI_NUM, layout::GSI_MSI_START, layout::IOAPIC_ADDR,
     layout::MEM_32BIT_DEVICES_SIZE, layout::MEM_32BIT_DEVICES_START,
     layout::MEM_64BIT_DEVICES_SIZE, layout::MEM_64BIT_DEVICES_START, layout::MMIO32_MEM_SIZE,
     layout::MMIO32_MEM_START, layout::PCI_MMCONFIG_SIZE, layout::PCI_MMCONFIG_START,

src/vmm/src/arch/x86_64/layout.rs

@@ -21,17 +21,21 @@ pub const CMDLINE_MAX_SIZE: usize = 2048;
 /// Start of the high memory.
 pub const HIMEM_START: u64 = 0x0010_0000; // 1 MB.
 
-// Typically, on x86 systems 24 IRQs are used (0-23).
-/// First usable IRQ ID for virtio device interrupts on x86_64.
-pub const IRQ_BASE: u32 = 5;
-/// Last usable IRQ ID for virtio device interrupts on x86_64.
-pub const IRQ_MAX: u32 = 23;
-
-/// The first usable GSI on x86_64 is the same as the first usable IRQ ID.
-pub const GSI_BASE: u32 = IRQ_BASE;
-
-/// The maximum usable GSI on x86_64 is the same as the last usable IRQ ID.
-pub const GSI_MAX: u32 = IRQ_MAX;
+// Typically, on x86 systems 24 IRQs are used for legacy devices (0-23).
+// However, the first 5 are reserved.
+// We allocate the remaining GSIs to MSIs.
+/// First usable GSI for legacy interrupts (IRQ) on x86_64.
+pub const GSI_LEGACY_START: u32 = 5;
+/// Last usable GSI for legacy interrupts (IRQ) on x86_64.
+pub const GSI_LEGACY_END: u32 = 23;
+/// Number of legacy GSI (IRQ) available on x86_64.
+pub const GSI_LEGACY_NUM: u32 = GSI_LEGACY_END - GSI_LEGACY_START + 1;
+/// First GSI used by MSI after legacy GSI.
+pub const GSI_MSI_START: u32 = GSI_LEGACY_END + 1;
+/// The highest available GSI in KVM (KVM_MAX_IRQ_ROUTES=4096).
+pub const GSI_MSI_END: u32 = 4095;
+/// Number of GSI available for MSI.
+pub const GSI_MSI_NUM: u32 = GSI_MSI_END - GSI_MSI_START + 1;
 
 /// Address for the TSS setup.
 pub const KVM_TSS_ADDRESS: u64 = 0xfffb_d000;

src/vmm/src/arch/x86_64/mptable.rs

@@ -13,7 +13,7 @@ use libc::c_char;
 use log::debug;
 use vm_allocator::AllocPolicy;
 
-use crate::arch::IRQ_MAX;
+use crate::arch::GSI_LEGACY_END;
 use crate::arch::x86_64::generated::mpspec;
 use crate::vstate::memory::{
     Address, ByteValued, Bytes, GuestAddress, GuestMemory, GuestMemoryMmap,
@@ -109,7 +109,7 @@ fn compute_mp_size(num_cpus: u8) -> usize {
         + mem::size_of::<mpspec::mpc_cpu>() * (num_cpus as usize)
         + mem::size_of::<mpspec::mpc_ioapic>()
         + mem::size_of::<mpspec::mpc_bus>()
-        + mem::size_of::<mpspec::mpc_intsrc>() * (IRQ_MAX as usize + 1)
+        + mem::size_of::<mpspec::mpc_intsrc>() * (GSI_LEGACY_END as usize + 1)
         + mem::size_of::<mpspec::mpc_lintsrc>() * 2
 }
 
@@ -225,7 +225,7 @@ pub fn setup_mptable(
         mp_num_entries += 1;
     }
     // Per kvm_setup_default_irq_routing() in kernel
-    for i in 0..=u8::try_from(IRQ_MAX).map_err(|_| MptableError::TooManyIrqs)? {
+    for i in 0..=u8::try_from(GSI_LEGACY_END).map_err(|_| MptableError::TooManyIrqs)? {
         let size = mem::size_of::<mpspec::mpc_intsrc>() as u64;
         let mpc_intsrc = mpspec::mpc_intsrc {
             type_: mpspec::MP_INTSRC.try_into().unwrap(),
@@ -406,7 +406,7 @@ mod tests {
             // ISA Bus
             + 1
             // IRQ
-            + u16::try_from(IRQ_MAX).unwrap() + 1
+            + u16::try_from(GSI_LEGACY_END).unwrap() + 1
             // Interrupt source ExtINT
             + 1
             // Interrupt source NMI

src/vmm/src/device_manager/acpi.rs

@@ -64,7 +64,7 @@ impl Aml for ACPIDeviceManager {
                                 // We know that the maximum IRQ number fits in a u8. We have up to
                                 // 32 IRQs in x86 and up to 128 in
                                 // ARM (look into
-                                // `vmm::crate::arch::layout::IRQ_MAX`)
+                                // `vmm::crate::arch::layout::GSI_LEGACY_END`)
                                 #[allow(clippy::cast_possible_truncation)]
                                 &aml::Equal::new(&aml::Arg(0), &(vmgenid.gsi as u8)),
                                 vec![&aml::Notify::new(

src/vmm/src/device_manager/mmio.rs

@@ -82,12 +82,12 @@ fn add_virtio_aml(
     dsdt_data: &mut Vec<u8>,
     addr: u64,
     len: u64,
-    irq: u32,
+    gsi: u32,
 ) -> Result<(), aml::AmlError> {
-    let dev_id = irq - crate::arch::GSI_BASE;
+    let dev_id = gsi - crate::arch::GSI_LEGACY_START;
     debug!(
-        "acpi: Building AML for VirtIO device _SB_.V{:03}. memory range: {:#010x}:{} irq: {}",
-        dev_id, addr, len, irq
+        "acpi: Building AML for VirtIO device _SB_.V{:03}. memory range: {:#010x}:{} gsi: {}",
+        dev_id, addr, len, gsi
     );
     aml::Device::new(
         format!("V{:03}", dev_id).as_str().try_into()?,
@@ -103,7 +103,7 @@ fn add_virtio_aml(
                         addr.try_into().unwrap(),
                         len.try_into().unwrap(),
                     ),
-                    &aml::Interrupt::new(true, true, false, false, irq),
+                    &aml::Interrupt::new(true, true, false, false, gsi),
                 ]),
             )?,
         ],
@@ -156,9 +156,9 @@ impl MMIODeviceManager {
         resource_allocator: &mut ResourceAllocator,
         irq_count: u32,
     ) -> Result<MMIODeviceInfo, MmioError> {
-        let irq = match resource_allocator.allocate_gsi(irq_count)?[..] {
+        let gsi = match resource_allocator.allocate_gsi_legacy(irq_count)?[..] {
             [] => None,
-            [irq] => Some(irq),
+            [gsi] => Some(gsi),
             _ => return Err(MmioError::InvalidIrqConfig),
         };
 
@@ -169,7 +169,7 @@ impl MMIODeviceManager {
                 AllocPolicy::FirstMatch,
             )?,
             len: MMIO_LEN,
-            irq,
+            irq: gsi,
         };
         Ok(device_info)
     }
@@ -183,7 +183,7 @@ impl MMIODeviceManager {
     ) -> Result<(), MmioError> {
         // Our virtio devices are currently hardcoded to use a single IRQ.
         // Validate that requirement.
-        let irq = device.resources.irq.ok_or(MmioError::InvalidIrqConfig)?;
+        let gsi = device.resources.irq.ok_or(MmioError::InvalidIrqConfig)?;
         let identifier;
         {
             let mmio_device = device.inner.lock().expect("Poisoned lock");
@@ -197,7 +197,7 @@ impl MMIODeviceManager {
                     .register_ioevent(queue_evt, &io_addr, u32::try_from(i).unwrap())
                     .map_err(MmioError::RegisterIoEvent)?;
             }
-            vm.register_irq(&mmio_device.interrupt.irq_evt, irq)
+            vm.register_irq(&mmio_device.interrupt.irq_evt, gsi)
                 .map_err(MmioError::RegisterIrqFd)?;
         }
 
@@ -276,7 +276,7 @@ impl MMIODeviceManager {
         let device_info = if let Some(device_info) = device_info_opt {
             device_info
         } else {
-            let gsi = vm.resource_allocator().allocate_gsi(1)?;
+            let gsi = vm.resource_allocator().allocate_gsi_legacy(1)?;
             MMIODeviceInfo {
                 addr: SERIAL_MEM_START,
                 len: MMIO_LEN,
@@ -335,7 +335,7 @@ impl MMIODeviceManager {
         let device_info = if let Some(device_info) = device_info_opt {
             device_info
         } else {
-            let gsi = vm.resource_allocator().allocate_gsi(1)?;
+            let gsi = vm.resource_allocator().allocate_gsi_legacy(1)?;
             MMIODeviceInfo {
                 addr: RTC_MEM_START,
                 len: MMIO_LEN,
@@ -612,15 +612,15 @@ pub(crate) mod tests {
         let dev = device_manager.get_virtio_device(0, "dummy").unwrap();
         assert_eq!(dev.resources.addr, arch::MEM_32BIT_DEVICES_START);
         assert_eq!(dev.resources.len, MMIO_LEN);
-        assert_eq!(dev.resources.irq, Some(arch::GSI_BASE));
+        assert_eq!(dev.resources.irq, Some(arch::GSI_LEGACY_START));
 
         device_manager
             .for_each_virtio_device(|virtio_type, device_id, mmio_device| {
                 assert_eq!(*virtio_type, 0);
                 assert_eq!(device_id, "dummy");
                 assert_eq!(mmio_device.resources.addr, arch::MEM_32BIT_DEVICES_START);
                 assert_eq!(mmio_device.resources.len, MMIO_LEN);
-                assert_eq!(mmio_device.resources.irq, Some(arch::GSI_BASE));
+                assert_eq!(mmio_device.resources.irq, Some(arch::GSI_LEGACY_START));
                 Ok::<(), ()>(())
             })
             .unwrap();
@@ -643,7 +643,7 @@ pub(crate) mod tests {
         #[cfg(target_arch = "aarch64")]
         vm.setup_irqchip(1).unwrap();
 
-        for _i in crate::arch::GSI_BASE..=crate::arch::GSI_MAX {
+        for _i in crate::arch::GSI_LEGACY_START..=crate::arch::GSI_LEGACY_END {
             device_manager
                 .register_virtio_test_device(
                     &vm,
@@ -711,7 +711,7 @@ pub(crate) mod tests {
                 .addr
         );
         assert_eq!(
-            crate::arch::GSI_BASE,
+            crate::arch::GSI_LEGACY_START,
             device_manager.virtio_devices[&(type_id, id)]
                 .resources
                 .irq
@@ -762,7 +762,7 @@ pub(crate) mod tests {
         let device_info = device_manager
             .allocate_mmio_resources(&mut resource_allocator, 1)
             .unwrap();
-        assert_eq!(device_info.irq.unwrap(), crate::arch::GSI_BASE);
+        assert_eq!(device_info.irq.unwrap(), crate::arch::GSI_LEGACY_START);
     }
 
     #[test]