On Tue, Aug 11, 2026 at 2:21 PM JinRui <[email protected]> wrote:
>
> From: jinrui <[email protected]>
>
> When the cross-compiler defaults to an -march that includes the V
> (vector) extension, -O2 auto-vectorization generates vector
> instructions (e.g. vsetvli, vadd.vv) in Guest binary code. If the
> Guest executes a vector instruction while sstatus.VS is Off, an
> EXC_INST_ILLEGAL (scause=2) is raised. KVM's hedeleg delegates this
> exception to the Guest, but the selftest has no way to handle it as a
> bare-metal program, causing all Guest tests to fail.
>
> In contrast, a real OS kernel handles this via riscv_v_first_use_handler(),
> which detects the vector instruction, sets sstatus.VS to Initial, and
> srets to re-execute.
>
> Fix this with four changes in processor.c:
>
> 1. Delete the now-unused guest_unexp_trap() handler, which is replaced
> by the full exception vector table.
>
> 2. In vm_arch_vcpu_add(), notify KVM that the Guest is allowed to use
> the V extension via __vcpu_set_reg(V, 1) (best-effort, silently
> ignores errors on hardware without V). Also replace the raw stvec
> handler with the full exception vector table, which provides
> save_context/restore_context for safe lazy enablement.
>
> 3. In route_exception(), add a lazy V enablement check that runs
> before any test-registered handler. When the cause is a non-IRQ
> EXC_INST_ILLEGAL and sstatus.VS is Off, set VS to Initial and
> return so the faulting instruction is re-executed via sret.
> Track the epc per-vCPU via a flexible array member to avoid
> infinite loops on hardware without V and eliminate races between
> concurrent vCPUs.
>
> 4. Make vm_init_vector_tables() idempotent by checking vm->handlers
> before allocation, so tests that manually call it (ebreak_test,
> arch_timer, sbi_pmu_test) do not leak memory.
>
> The check runs before test-registered EXC_INST_ILLEGAL handlers
> (e.g. sbi_pmu_test) to ensure V enablement takes priority.
>
> Tested on a riscv64 host with KVM enabled: all 20 KVM selftest
> binaries pass (arch_timer, ebreak_test, steal_time, get-reg-list,
> sbi_pmu_test, etc.).
>
> Signed-off-by: jinrui <[email protected]>
> ---
> Changes in v9:
> - Size the per-vCPU epc array by KVM_CAP_MAX_VCPU_ID instead of
> KVM_CAP_MAX_VCPUS, since vcpu_id may be sparse and exceed the vCPU
> count (found by sashiko-bot review).
>
> .../selftests/kvm/lib/riscv/processor.c | 115 +++++++++++++++---
> 1 file changed, 100 insertions(+), 15 deletions(-)
>
> diff --git a/tools/testing/selftests/kvm/lib/riscv/processor.c
> b/tools/testing/selftests/kvm/lib/riscv/processor.c
> index ded5429f3..05e20ef40 100644
> --- a/tools/testing/selftests/kvm/lib/riscv/processor.c
> +++ b/tools/testing/selftests/kvm/lib/riscv/processor.c
> @@ -17,6 +17,13 @@
>
> static gva_t exception_handlers;
>
> +struct handlers {
> + exception_handler_fn exception_handlers[NR_VECTORS][NR_EXCEPTIONS];
> + bool v_available;
> + unsigned int v_epc_capacity;
> + unsigned long v_epc[];
> +};
> +
> bool __vcpu_has_ext(struct kvm_vcpu *vcpu, u64 ext)
> {
> unsigned long value = 0;
> @@ -298,13 +305,6 @@ void vcpu_arch_dump(FILE *stream, struct kvm_vcpu *vcpu,
> u8 indent)
> core.regs.t3, core.regs.t4, core.regs.t5, core.regs.t6);
> }
>
> -static void __aligned(16) guest_unexp_trap(void)
> -{
> - sbi_ecall(KVM_RISCV_SELFTESTS_SBI_EXT,
> - KVM_RISCV_SELFTESTS_SBI_UNEXP,
> - 0, 0, 0, 0, 0, 0);
> -}
> -
> void vcpu_arch_set_entry_point(struct kvm_vcpu *vcpu, void *guest_code)
> {
> vcpu_set_reg(vcpu, RISCV_CORE_REG(regs.pc), (unsigned
> long)guest_code);
> @@ -348,8 +348,33 @@ struct kvm_vcpu *vm_arch_vcpu_add(struct kvm_vm *vm, u32
> vcpu_id)
> /* Setup sscratch for guest_get_vcpuid() */
> vcpu_set_reg(vcpu, RISCV_GENERAL_CSR_REG(sscratch), vcpu_id);
>
> - /* Setup default exception vector of guest */
> - vcpu_set_reg(vcpu, RISCV_GENERAL_CSR_REG(stvec), (unsigned
> long)guest_unexp_trap);
> + /*
> + * Enable the V (vector) extension in KVM so that the compiler can
> + * safely generate vector instructions (e.g. via -O2 auto-
> + * vectorization). Silently ignore errors; the test will still work
> + * without V.
> + */
> + __vcpu_set_reg(vcpu, RISCV_ISA_EXT_REG(KVM_RISCV_ISA_EXT_V), 1);
> +
> + /*
> + * Use the full exception vector table (which provides lazy V
> + * extension enablement for EXC_INST_ILLEGAL in route_exception)
> + * as the default exception handler. vm_init_vector_tables() is
> + * idempotent; tests that call it again will get a no-op.
> + */
> + vm_init_vector_tables(vm);
> + vcpu_init_vector_tables(vcpu);
> +
> + /*
> + * Record V extension availability in the handlers struct so that
> + * route_exception() (called from Guest context) can check it
> + * without relying on a host-side global variable.
> + */
> + {
> + struct handlers *h = addr_gva2hva(vm, vm->handlers);
> +
> + h->v_available = __vcpu_has_isa_ext(vcpu,
> KVM_RISCV_ISA_EXT_V);
> + }
>
> return vcpu;
> }
> @@ -408,19 +433,17 @@ void assert_on_unhandled_exception(struct kvm_vcpu
> *vcpu)
> struct ucall uc;
>
> if (get_ucall(vcpu, &uc) == UCALL_UNHANDLED) {
> + vcpu_dump(stderr, vcpu, 2);
> TEST_FAIL("Unexpected exception (vector:0x%lx, ec:0x%lx)",
> uc.args[0], uc.args[1]);
> }
> }
>
> -struct handlers {
> - exception_handler_fn exception_handlers[NR_VECTORS][NR_EXCEPTIONS];
> -};
> -
> void route_exception(struct pt_regs *regs)
> {
> struct handlers *handlers = (struct handlers *)exception_handlers;
> - int vector = 0, ec;
> + int vector = 0;
> + unsigned long ec;
>
> ec = regs->cause & ~CAUSE_IRQ_FLAG;
> if (ec >= NR_EXCEPTIONS)
> @@ -432,6 +455,46 @@ void route_exception(struct pt_regs *regs)
> ec = 0;
> }
>
> + /*
> + * Handle V (vector) extension lazy enablement before any
> + * registered handler. The compiler's default march may include
> + * V, and auto-vectorization generates vector instructions that
> + * trigger EXC_INST_ILLEGAL when VS (Vector Status) in sstatus
> + * is Off. Enable VS to Initial and re-execute the faulting
> + * instruction, mimicking what a real OS kernel does.
> + *
> + * This check runs before any test-registered handler, so tests
> + * that install their own EXC_INST_ILLEGAL handler (e.g.
> + * sbi_pmu_test) are not affected.
> + */
> + if (!(regs->cause & CAUSE_IRQ_FLAG) && ec == EXC_INST_ILLEGAL) {
> + /*
> + * If KVM supports the V extension for this Guest and VS
> + * (Vector Status) is Off in the saved sstatus, set it to
> + * Initial and re-execute the faulting instruction.
> + *
> + * Use regs->status (saved at exception entry) rather than
> + * reading the live CSR to avoid a TOCTOU race.
> + *
> + * Track the epc per-vCPU to avoid an infinite loop when
> + * V is disabled or the hardware rejects the VS change.
> + * Using a per-vCPU array avoids races between concurrent
> + * vCPUs that would occur with a single shared (epc, vcpu)
> + * pair.
> + */
> + if (handlers && handlers->v_available && !(regs->status &
> SR_VS)) {
Treating all illegal exceptions as V instruction traps is not the right the way.
We may have genuinely unexpected traps for which guest_unexp_trap()
should be called.
> + unsigned int vcpu_id;
> +
> + asm volatile("csrr %0, sscratch" : "=r" (vcpu_id));
> + if (vcpu_id < handlers->v_epc_capacity &&
> + handlers->v_epc[vcpu_id] != regs->epc) {
> + handlers->v_epc[vcpu_id] = regs->epc;
> + regs->status |= SR_VS_INITIAL;
> + return;
> + }
> + }
> + }
> +
> if (handlers && handlers->exception_handlers[vector][ec])
> return handlers->exception_handlers[vector][ec](regs);
>
> @@ -448,9 +511,31 @@ void vcpu_init_vector_tables(struct kvm_vcpu *vcpu)
>
> void vm_init_vector_tables(struct kvm_vm *vm)
> {
> - vm->handlers = __vm_alloc(vm, sizeof(struct handlers), vm->page_size,
> + unsigned int max_vcpu_id;
> + size_t size;
> +
> + if (vm->handlers)
> + return;
> +
> + max_vcpu_id = kvm_check_cap(KVM_CAP_MAX_VCPU_ID);
> + if (max_vcpu_id == 0)
> + max_vcpu_id = 512;
> +
> + /*
> + * vcpu_id may be sparse and ranges from 0 to KVM_CAP_MAX_VCPU_ID
> + * (which can be much larger than KVM_CAP_MAX_VCPUS), so size the
> + * per-vCPU epc array accordingly.
> + */
> + size = sizeof(struct handlers) + (max_vcpu_id + 1) * sizeof(unsigned
> long);
> + vm->handlers = __vm_alloc(vm, size, vm->page_size,
> MEM_REGION_DATA);
>
> + {
> + struct handlers *h = addr_gva2hva(vm, vm->handlers);
> +
> + h->v_epc_capacity = max_vcpu_id + 1;
> + }
> +
> *(gva_t *)addr_gva2hva(vm, (gva_t)(&exception_handlers)) =
> vm->handlers;
> }
>
> --
> 2.53.0
>
Regards,
Anup
