On Wed, 2026-10-28 at 22:01 +0800, Tan Chi wrote: > KVM irqfd bypasses the PCI device's IOMMU AddressSpace, so MSI > addresses requiring translation may fail to reach the guest IMSIC. > > Translate MSI addresses through the device's IOMMU AddressSpace when > setting up KVM routes. Only direct mappings to an IMSIC MemoryRegion > are installed in KVM routes. > > Other IOMMU MSI remapping forms are not converted to KVM routes and > continue to use the existing userspace MSI delivery path. > > Select PCI for RISCV64 KVM builds, as QEMU currently enables the > RISC-V KVM target only for riscv64-softmmu builds and the MSI route > fixup requires PCI support. Also trace translation failures to help > diagnose route setup failures. > > Signed-off-by: Tan Chi <[email protected]> > Reviewed-by: Nutty Liu <[email protected]>
Acked-by: Alistair Francis <[email protected]> Alistair > --- > > Changes in v3: > - Remove the unused length variable and pass NULL to > address_space_translate(), following Nutty Liu's review. > > Changes in v2 (based on Alistair Francis's review): > - Select PCI for RISCV64 KVM builds. > - Add trace events for MSI translation failures. > > target/riscv/Kconfig | 2 ++ > target/riscv/kvm/kvm-cpu.c | 54 > ++++++++++++++++++++++++++++++++++++++ > target/riscv/trace-events | 4 +++ > 3 files changed, 60 insertions(+) > > diff --git a/target/riscv/Kconfig b/target/riscv/Kconfig > index 11bc09b414..0cf2c8e1c4 100644 > --- a/target/riscv/Kconfig > +++ b/target/riscv/Kconfig > @@ -7,3 +7,5 @@ config RISCV64 > bool > select ARM_COMPATIBLE_SEMIHOSTING if TCG > select DEVICE_TREE # needed by boot.c > + # kvm_arch_fixup_msi_route() needs to access PCIDevice > + select PCI if KVM > diff --git a/target/riscv/kvm/kvm-cpu.c b/target/riscv/kvm/kvm-cpu.c > index 45ad5e1525..a1eecd72b7 100644 > --- a/target/riscv/kvm/kvm-cpu.c > +++ b/target/riscv/kvm/kvm-cpu.c > @@ -1507,6 +1507,60 @@ int kvm_arch_release_virq_post(int virq) > int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route, > uint64_t address, uint32_t data, > PCIDevice *dev) > { > + AddressSpace *as; > + hwaddr xlat, translated; > + MemoryRegionSection section; > + MemoryRegion *mr; > + Object *owner; > + > + if (!dev) { > + return 0; > + } > + > + as = pci_device_iommu_address_space(dev); > + if (as == &address_space_memory) { > + return 0; > + } > + > + RCU_READ_LOCK_GUARD(); > + > + mr = address_space_translate(as, address, &xlat, NULL, true, > + MEMTXATTRS_UNSPECIFIED); > + > + if (!mr) { > + trace_kvm_riscv_fixup_msi_route_failed(address, > + "address translation > failed", > + "none"); > + return 1; > + } > + > + /* > + * Only direct IMSIC writes can be represented by a KVM MSI > route. > + * MSI page-table translations instead target the RISC-V IOMMU > trap > + * region, which requires further BASIC/MRIF processing. > + */ > + owner = memory_region_owner(mr); > + if (!owner || !object_dynamic_cast(owner, TYPE_RISCV_IMSIC)) { > + trace_kvm_riscv_fixup_msi_route_failed(address, > + "non-IMSIC target", > + > memory_region_name(mr)); > + return 1; > + } > + > + section = memory_region_find(mr, xlat, 1); > + if (!section.mr) { > + trace_kvm_riscv_fixup_msi_route_failed(address, > + "memory region lookup > failed", > + > memory_region_name(mr)); > + return 1; > + } > + > + translated = section.offset_within_address_space; > + memory_region_unref(section.mr); > + > + route->u.msi.address_lo = translated; > + route->u.msi.address_hi = translated >> 32; > + trace_kvm_riscv_fixup_msi_route_translated(address, translated); > return 0; > } > > diff --git a/target/riscv/trace-events b/target/riscv/trace-events > index 93837f82a1..50980c031d 100644 > --- a/target/riscv/trace-events > +++ b/target/riscv/trace-events > @@ -12,3 +12,7 @@ mseccfg_csr_write(uint64_t mhartid, uint64_t val) > "hart %" PRIu64 ": write msecc > > # op_helper.c > riscv_exception(uint32_t exception, const char *desc, uint64_t epc) > "%u (%s) on epc 0x%"PRIx64"" > + > +# kvm-cpu.c > +kvm_riscv_fixup_msi_route_translated(uint64_t iova, uint64_t gpa) > "MSI iova = 0x%"PRIx64" is translated into 0x%"PRIx64 > +kvm_riscv_fixup_msi_route_failed(uint64_t iova, const char *reason, > const char *region) "MSI iova = 0x%"PRIx64" failed: %s (region=%s)"
