@@ -130,6 +130,7 @@ TEST_GEN_PROGS_x86_64 += x86_64/max_vcpuid_cap_test
TEST_GEN_PROGS_x86_64 += x86_64/triple_fault_event_test
TEST_GEN_PROGS_x86_64 += x86_64/recalc_apic_map_test
TEST_GEN_PROGS_x86_64 += access_tracking_perf_test
+TEST_GEN_PROGS_x86_64 += coalesced_io_test
TEST_GEN_PROGS_x86_64 += demand_paging_test
TEST_GEN_PROGS_x86_64 += dirty_log_test
TEST_GEN_PROGS_x86_64 += dirty_log_perf_test
@@ -165,6 +166,7 @@ TEST_GEN_PROGS_aarch64 += aarch64/vgic_lpi_stress
TEST_GEN_PROGS_aarch64 += aarch64/vpmu_counter_access
TEST_GEN_PROGS_aarch64 += access_tracking_perf_test
TEST_GEN_PROGS_aarch64 += arch_timer
+TEST_GEN_PROGS_aarch64 += coalesced_io_test
TEST_GEN_PROGS_aarch64 += demand_paging_test
TEST_GEN_PROGS_aarch64 += dirty_log_test
TEST_GEN_PROGS_aarch64 += dirty_log_perf_test
@@ -198,6 +200,7 @@ TEST_GEN_PROGS_s390x += kvm_binary_stats_test
TEST_GEN_PROGS_riscv += riscv/sbi_pmu_test
TEST_GEN_PROGS_riscv += riscv/ebreak_test
TEST_GEN_PROGS_riscv += arch_timer
+TEST_GEN_PROGS_riscv += coalesced_io_test
TEST_GEN_PROGS_riscv += demand_paging_test
TEST_GEN_PROGS_riscv += dirty_log_test
TEST_GEN_PROGS_riscv += get-reg-list
new file mode 100644
@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+
+#include <linux/sizes.h>
+
+#include <kvm_util.h>
+#include <processor.h>
+
+#include "ucall_common.h"
+
+struct kvm_coalesced_io {
+ struct kvm_coalesced_mmio_ring *ring;
+ uint32_t ring_size;
+ uint64_t mmio_gpa;
+ uint64_t *mmio;
+
+ /*
+ * x86-only, but define pio_port for all architectures to minimize the
+ * amount of #ifdeffery and complexity, without having to sacrifice
+ * verbose error messages.
+ */
+ uint8_t pio_port;
+};
+
+static struct kvm_coalesced_io kvm_builtin_io_ring;
+
+#ifdef __x86_64__
+static const int has_pio = 1;
+#else
+static const int has_pio = 0;
+#endif
+
+static void guest_code(struct kvm_coalesced_io *io)
+{
+ int i, j;
+
+ for (;;) {
+ for (j = 0; j < 1 + has_pio; j++) {
+ /*
+ * KVM always leaves one free entry, i.e. exits to
+ * userspace before the last entry is filled.
+ */
+ for (i = 0; i < io->ring_size - 1; i++) {
+#ifdef __x86_64__
+ if (i & 1)
+ outl(io->pio_port, io->pio_port + i);
+ else
+#endif
+ WRITE_ONCE(*io->mmio, io->mmio_gpa + i);
+ }
+#ifdef __x86_64__
+ if (j & 1)
+ outl(io->pio_port, io->pio_port + i);
+ else
+#endif
+ WRITE_ONCE(*io->mmio, io->mmio_gpa + i);
+ }
+ GUEST_SYNC(0);
+
+ WRITE_ONCE(*io->mmio, io->mmio_gpa + i);
+#ifdef __x86_64__
+ outl(io->pio_port, io->pio_port + i);
+#endif
+ }
+}
+
+static void vcpu_run_and_verify_io_exit(struct kvm_vcpu *vcpu,
+ struct kvm_coalesced_io *io,
+ uint32_t ring_start,
+ uint32_t expected_exit)
+{
+ const bool want_pio = expected_exit == KVM_EXIT_IO;
+ struct kvm_coalesced_mmio_ring *ring = io->ring;
+ struct kvm_run *run = vcpu->run;
+ uint32_t pio_value;
+
+ WRITE_ONCE(ring->first, ring_start);
+ WRITE_ONCE(ring->last, ring_start);
+
+ vcpu_run(vcpu);
+
+ /*
+ * Annoyingly, reading PIO data is safe only for PIO exits, otherwise
+ * data_offset is garbage, e.g. an MMIO gpa.
+ */
+ if (run->exit_reason == KVM_EXIT_IO)
+ pio_value = *(uint32_t *)((void *)run + run->io.data_offset);
+ else
+ pio_value = 0;
+
+ TEST_ASSERT((!want_pio && (run->exit_reason == KVM_EXIT_MMIO && run->mmio.is_write &&
+ run->mmio.phys_addr == io->mmio_gpa && run->mmio.len == 8 &&
+ *(uint64_t *)run->mmio.data == io->mmio_gpa + io->ring_size - 1)) ||
+ (want_pio && (run->exit_reason == KVM_EXIT_IO && run->io.port == io->pio_port &&
+ run->io.direction == KVM_EXIT_IO_OUT && run->io.count == 1 &&
+ pio_value == io->pio_port + io->ring_size - 1)),
+ "For start = %u, expected exit on %u-byte %s write 0x%llx = %lx, got exit_reason = %u (%s)\n "
+ "(MMIO addr = 0x%llx, write = %u, len = %u, data = %lx)\n "
+ "(PIO port = 0x%x, write = %u, len = %u, count = %u, data = %x",
+ ring_start, want_pio ? 4 : 8, want_pio ? "PIO" : "MMIO",
+ want_pio ? (unsigned long long)io->pio_port : io->mmio_gpa,
+ (want_pio ? io->pio_port : io->mmio_gpa) + io->ring_size - 1, run->exit_reason,
+ run->exit_reason == KVM_EXIT_MMIO ? "MMIO" : run->exit_reason == KVM_EXIT_IO ? "PIO" : "other",
+ run->mmio.phys_addr, run->mmio.is_write, run->mmio.len, *(uint64_t *)run->mmio.data,
+ run->io.port, run->io.direction, run->io.size, run->io.count, pio_value);
+}
+
+static void vcpu_run_and_verify_coalesced_io(struct kvm_vcpu *vcpu,
+ struct kvm_coalesced_io *io,
+ uint32_t ring_start,
+ uint32_t expected_exit)
+{
+ struct kvm_coalesced_mmio_ring *ring = io->ring;
+ int i;
+
+ vcpu_run_and_verify_io_exit(vcpu, io, ring_start, expected_exit);
+
+ TEST_ASSERT((ring->last + 1) % io->ring_size == ring->first,
+ "Expected ring to be full (minus 1), first = %u, last = %u, max = %u, start = %u",
+ ring->first, ring->last, io->ring_size, ring_start);
+
+ for (i = 0; i < io->ring_size - 1; i++) {
+ uint32_t idx = (ring->first + i) % io->ring_size;
+ struct kvm_coalesced_mmio *entry = &ring->coalesced_mmio[idx];
+
+#ifdef __x86_64__
+ if (i & 1)
+ TEST_ASSERT(entry->phys_addr == io->pio_port &&
+ entry->len == 4 && entry->pio &&
+ *(uint32_t *)entry->data == io->pio_port + i,
+ "Wanted 4-byte port I/O 0x%x = 0x%x in entry %u, got %u-byte %s 0x%llx = 0x%x",
+ io->pio_port, io->pio_port + i, i,
+ entry->len, entry->pio ? "PIO" : "MMIO",
+ entry->phys_addr, *(uint32_t *)entry->data);
+ else
+#endif
+ TEST_ASSERT(entry->phys_addr == io->mmio_gpa &&
+ entry->len == 8 && !entry->pio,
+ "Wanted 8-byte MMIO to 0x%lx = %lx in entry %u, got %u-byte %s 0x%llx = 0x%lx",
+ io->mmio_gpa, io->mmio_gpa + i, i,
+ entry->len, entry->pio ? "PIO" : "MMIO",
+ entry->phys_addr, *(uint64_t *)entry->data);
+ }
+}
+
+static void test_coalesced_io(struct kvm_vcpu *vcpu,
+ struct kvm_coalesced_io *io, uint32_t ring_start)
+{
+ struct kvm_coalesced_mmio_ring *ring = io->ring;
+
+ kvm_vm_register_coalesced_io(vcpu->vm, io->mmio_gpa, 8, false /* pio */);
+#ifdef __x86_64__
+ kvm_vm_register_coalesced_io(vcpu->vm, io->pio_port, 8, true /* pio */);
+#endif
+
+ vcpu_run_and_verify_coalesced_io(vcpu, io, ring_start, KVM_EXIT_MMIO);
+#ifdef __x86_64__
+ vcpu_run_and_verify_coalesced_io(vcpu, io, ring_start, KVM_EXIT_IO);
+#endif
+
+ /*
+ * Verify ucall, which may use non-coalesced MMIO or PIO, generates an
+ * immediate exit.
+ */
+ WRITE_ONCE(ring->first, ring_start);
+ WRITE_ONCE(ring->last, ring_start);
+ vcpu_run(vcpu);
+ TEST_ASSERT_EQ(get_ucall(vcpu, NULL), UCALL_SYNC);
+ TEST_ASSERT_EQ(ring->first, ring_start);
+ TEST_ASSERT_EQ(ring->last, ring_start);
+
+ /* Verify that non-coalesced MMIO/PIO generates an exit to userspace. */
+ kvm_vm_unregister_coalesced_io(vcpu->vm, io->mmio_gpa, 8, false /* pio */);
+ vcpu_run_and_verify_io_exit(vcpu, io, ring_start, KVM_EXIT_MMIO);
+
+#ifdef __x86_64__
+ kvm_vm_unregister_coalesced_io(vcpu->vm, io->pio_port, 8, true /* pio */);
+ vcpu_run_and_verify_io_exit(vcpu, io, ring_start, KVM_EXIT_IO);
+#endif
+}
+
+int main(int argc, char *argv[])
+{
+ struct kvm_vcpu *vcpu;
+ struct kvm_vm *vm;
+ int i;
+
+ TEST_REQUIRE(kvm_has_cap(KVM_CAP_COALESCED_MMIO));
+
+#ifdef __x86_64__
+ TEST_REQUIRE(kvm_has_cap(KVM_CAP_COALESCED_PIO));
+#endif
+
+ vm = vm_create_with_one_vcpu(&vcpu, guest_code);
+
+ kvm_builtin_io_ring = (struct kvm_coalesced_io) {
+ /*
+ * The I/O ring is a kernel-allocated page whose address is
+ * relative to each vCPU's run page, with the page offset
+ * provided by KVM in the return of KVM_CAP_COALESCED_MMIO.
+ */
+ .ring = (void *)vcpu->run +
+ (kvm_check_cap(KVM_CAP_COALESCED_MMIO) * getpagesize()),
+
+ /*
+ * The size of the I/O ring is fixed, but KVM defines the sized
+ * based on the kernel's PAGE_SIZE. Thus, userspace must query
+ * the host's page size at runtime to compute the ring size.
+ */
+ .ring_size = (getpagesize() - sizeof(struct kvm_coalesced_mmio_ring)) /
+ sizeof(struct kvm_coalesced_mmio),
+
+ /*
+ * Arbitrary address+port (MMIO mustn't overlap memslots), with
+ * the MMIO GPA identity mapped in the guest.
+ */
+ .mmio_gpa = 4ull * SZ_1G,
+ .mmio = (uint64_t *)(4ull * SZ_1G),
+ .pio_port = 0x80,
+ };
+
+ virt_map(vm, (uint64_t)kvm_builtin_io_ring.mmio, kvm_builtin_io_ring.mmio_gpa, 1);
+
+ sync_global_to_guest(vm, kvm_builtin_io_ring);
+ vcpu_args_set(vcpu, 1, &kvm_builtin_io_ring);
+
+ for (i = 0; i < kvm_builtin_io_ring.ring_size; i++)
+ test_coalesced_io(vcpu, &kvm_builtin_io_ring, i);
+
+ kvm_vm_free(vm);
+ return 0;
+}
@@ -460,6 +460,32 @@ static inline uint32_t kvm_vm_reset_dirty_ring(struct kvm_vm *vm)
return __vm_ioctl(vm, KVM_RESET_DIRTY_RINGS, NULL);
}
+static inline void kvm_vm_register_coalesced_io(struct kvm_vm *vm,
+ uint64_t address,
+ uint64_t size, bool pio)
+{
+ struct kvm_coalesced_mmio_zone zone = {
+ .addr = address,
+ .size = size,
+ .pio = pio,
+ };
+
+ vm_ioctl(vm, KVM_REGISTER_COALESCED_MMIO, &zone);
+}
+
+static inline void kvm_vm_unregister_coalesced_io(struct kvm_vm *vm,
+ uint64_t address,
+ uint64_t size, bool pio)
+{
+ struct kvm_coalesced_mmio_zone zone = {
+ .addr = address,
+ .size = size,
+ .pio = pio,
+ };
+
+ vm_ioctl(vm, KVM_UNREGISTER_COALESCED_MMIO, &zone);
+}
+
static inline int vm_get_stats_fd(struct kvm_vm *vm)
{
int fd = __vm_ioctl(vm, KVM_GET_STATS_FD, NULL);
Add a test to verify that KVM correctly exits (or not) when a vCPU's coalesced I/O ring is full (or isn't). Iterate over all legal starting points in the ring (with an empty ring), and verify that KVM doesn't exit until the ring is full. Opportunistically verify that KVM exits immediately on non-coalesced I/O, either because the MMIO/PIO region was never registered, or because a previous region was unregistered. This is a regression test for a KVM bug where KVM would prematurely exit due to bad math resulting in a false positive if the first entry in the ring was before the halfway mark. See commit 92f6d4130497 ("KVM: Fix coalesced_mmio_has_room() to avoid premature userspace exit"). Enable the test for x86, arm64, and risc-v, i.e. all architectures except s390, which doesn't have MMIO. On x86, which has both MMIO and PIO, interleave MMIO and PIO into the same ring, as KVM shouldn't exit until a non-coalesced I/O is encountered, regardless of whether the ring is filled with MMIO, PIO, or both. Lastly, wrap the coalesced I/O ring in a structure to prepare for a potential future where KVM supports multiple ring buffers beyond KVM's "default" built-in buffer. Link: https://lore.kernel.org/all/20240820133333.1724191-1-ilstam@amazon.com Cc: Ilias Stamatis <ilstam@amazon.com> Cc: Marc Zyngier <maz@kernel.org> Cc: Oliver Upton <oliver.upton@linux.dev> Cc: Anup Patel <anup@brainfault.org> Signed-off-by: Sean Christopherson <seanjc@google.com> --- tools/testing/selftests/kvm/Makefile | 3 + .../testing/selftests/kvm/coalesced_io_test.c | 236 ++++++++++++++++++ .../testing/selftests/kvm/include/kvm_util.h | 26 ++ 3 files changed, 265 insertions(+) create mode 100644 tools/testing/selftests/kvm/coalesced_io_test.c