On Mon, Jul 19, 2021 at 06:49:49PM +0000, Oliver Upton wrote:
> Add a test case for counter emulation on arm64. A side effect of how KVM
> handles physical counter offsetting on non-ECV systems is that the
> virtual counter will always hit hardware and the physical could be
> emulated. Force emulation by writing a nonzero offset to the physical
> counter and compare the elapsed cycles to a direct read of the hardware
> register.
>
> Reviewed-by: Ricardo Koller <ricarkol@xxxxxxxxxx>
> Signed-off-by: Oliver Upton <oupton@xxxxxxxxxx>
> ---
> tools/testing/selftests/kvm/.gitignore | 1 +
> tools/testing/selftests/kvm/Makefile | 1 +
> .../kvm/aarch64/counter_emulation_benchmark.c | 215 ++++++++++++++++++
> 3 files changed, 217 insertions(+)
> create mode 100644 tools/testing/selftests/kvm/aarch64/counter_emulation_benchmark.c
>
> diff --git a/tools/testing/selftests/kvm/.gitignore b/tools/testing/selftests/kvm/.gitignore
> index 2752813d5090..1d811c6a769b 100644
> --- a/tools/testing/selftests/kvm/.gitignore
> +++ b/tools/testing/selftests/kvm/.gitignore
> @@ -1,5 +1,6 @@
> # SPDX-License-Identifier: GPL-2.0-only
> /aarch64/debug-exceptions
> +/aarch64/counter_emulation_benchmark
alphabetic order please
> /aarch64/get-reg-list
> /aarch64/vgic_init
> /s390x/memop
> diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile
> index d89908108c97..e560a3e74bc2 100644
> --- a/tools/testing/selftests/kvm/Makefile
> +++ b/tools/testing/selftests/kvm/Makefile
> @@ -86,6 +86,7 @@ TEST_GEN_PROGS_x86_64 += kvm_binary_stats_test
> TEST_GEN_PROGS_x86_64 += system_counter_offset_test
>
> TEST_GEN_PROGS_aarch64 += aarch64/debug-exceptions
> +TEST_GEN_PROGS_aarch64 += aarch64/counter_emulation_benchmark
alphabetic order please
> TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list
> TEST_GEN_PROGS_aarch64 += aarch64/vgic_init
> TEST_GEN_PROGS_aarch64 += demand_paging_test
> diff --git a/tools/testing/selftests/kvm/aarch64/counter_emulation_benchmark.c b/tools/testing/selftests/kvm/aarch64/counter_emulation_benchmark.c
> new file mode 100644
> index 000000000000..73aeb6cdebfe
> --- /dev/null
> +++ b/tools/testing/selftests/kvm/aarch64/counter_emulation_benchmark.c
> @@ -0,0 +1,215 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * counter_emulation_benchmark.c -- test to measure the effects of counter
> + * emulation on guest reads of the physical counter.
> + *
> + * Copyright (c) 2021, Google LLC.
> + */
> +
> +#define _GNU_SOURCE
> +#include <asm/kvm.h>
> +#include <linux/kvm.h>
> +#include <stdio.h>
> +#include <stdint.h>
> +#include <stdlib.h>
> +#include <unistd.h>
> +
> +#include "kvm_util.h"
> +#include "processor.h"
> +#include "test_util.h"
> +
> +#define VCPU_ID 0
> +
> +static struct counter_values {
> + uint64_t cntvct_start;
> + uint64_t cntpct;
> + uint64_t cntvct_end;
> +} counter_values;
> +
> +static uint64_t nr_iterations = 1000;
> +
> +static void do_test(void)
> +{
> + /*
> + * Open-coded approach instead of using helper methods to keep a tight
> + * interval around the physical counter read.
> + */
> + asm volatile("isb\n\t"
> + "mrs %[cntvct_start], cntvct_el0\n\t"
> + "isb\n\t"
> + "mrs %[cntpct], cntpct_el0\n\t"
> + "isb\n\t"
> + "mrs %[cntvct_end], cntvct_el0\n\t"
> + "isb\n\t"
> + : [cntvct_start] "=r"(counter_values.cntvct_start),
> + [cntpct] "=r"(counter_values.cntpct),
> + [cntvct_end] "=r"(counter_values.cntvct_end));
> +}
> +
> +static void guest_main(void)
> +{
> + int i;
> +
> + for (i = 0; i < nr_iterations; i++) {
> + do_test();
> + GUEST_SYNC(i);
> + }
> +
> + for (i = 0; i < nr_iterations; i++) {
> + do_test();
> + GUEST_SYNC(i);
> + }
> +
> + GUEST_DONE();
> +}
> +
> +static bool enter_guest(struct kvm_vm *vm)
> +{
> + struct ucall uc;
> +
> + vcpu_ioctl(vm, VCPU_ID, KVM_RUN, NULL);
> +
> + switch (get_ucall(vm, VCPU_ID, &uc)) {
> + case UCALL_DONE:
> + return true;
> + case UCALL_SYNC:
> + break;
> + case UCALL_ABORT:
> + TEST_ASSERT(false, "%s at %s:%ld", (const char *)uc.args[0],
> + __FILE__, uc.args[1]);
> + break;
> + default:
> + TEST_ASSERT(false, "unexpected exit: %s",
> + exit_reason_str(vcpu_state(vm, VCPU_ID)->exit_reason));
> + break;
> + }
> +
> + /* more work to do in the guest */
> + return false;
> +}
> +
> +static double counter_frequency(void)
> +{
> + uint32_t freq;
> +
> + asm volatile("mrs %0, cntfrq_el0"
> + : "=r" (freq));
> +
> + return freq / 1000000.0;
> +}
> +
> +static void log_csv(FILE *csv, bool trapped)
> +{
> + double freq = counter_frequency();
> +
> + fprintf(csv, "%s,%.02f,%lu,%lu,%lu\n",
> + trapped ? "true" : "false", freq,
> + counter_values.cntvct_start,
> + counter_values.cntpct,
> + counter_values.cntvct_end);
> +}
> +
> +static double run_loop(struct kvm_vm *vm, FILE *csv, bool trapped)
> +{
> + double avg = 0;
> + int i;
> +
> + for (i = 0; i < nr_iterations; i++) {
> + uint64_t delta;
> +
> + TEST_ASSERT(!enter_guest(vm), "guest exited unexpectedly");
UCALL_DONE would assert, but we never do a UCALL_DONE because we're
entering the guest nr_iterations times but [would] exit it
2 * nr_iterations times before GUEST_DONE() gets called. IOW, the
logic of the run loop looks like it could use some cleanup.
> + sync_global_from_guest(vm, counter_values);
> +
> + if (csv)
> + log_csv(csv, trapped);
> +
> + delta = counter_values.cntvct_end - counter_values.cntvct_start;
> + avg = ((avg * i) + delta) / (i + 1);
> + }
> +
> + return avg;
> +}
> +
> +static void setup_counter(struct kvm_vm *vm, uint64_t offset)
> +{
> + vcpu_access_device_attr(vm, VCPU_ID, KVM_ARM_VCPU_TIMER_CTRL,
> + KVM_ARM_VCPU_TIMER_OFFSET_PTIMER, &offset,
> + true);
> +}
> +
> +static void run_tests(struct kvm_vm *vm, FILE *csv)
> +{
> + double avg_trapped, avg_native, freq;
> +
> + freq = counter_frequency();
> +
> + if (csv)
> + fputs("trapped,freq_mhz,cntvct_start,cntpct,cntvct_end\n", csv);
> +
> + /* no physical offsetting; kvm allows reads of cntpct_el0 */
> + setup_counter(vm, 0);
> + avg_native = run_loop(vm, csv, false);
> +
> + /* force emulation of the physical counter */
> + setup_counter(vm, 1);
> + avg_trapped = run_loop(vm, csv, true);
> +
> + TEST_ASSERT(enter_guest(vm), "guest didn't run to completion");
> + pr_info("%lu iterations: average cycles (@%.02fMHz) native: %.02f, trapped: %.02f\n",
> + nr_iterations, freq, avg_native, avg_trapped);
> +}
> +
> +static void usage(const char *program_name)
> +{
> + fprintf(stderr,
> + "Usage: %s [-h] [-o csv_file] [-n iterations]\n"
> + " -h prints this message\n"
> + " -n number of test iterations (default: %lu)\n"
> + " -o csv file to write data\n",
> + program_name, nr_iterations);
> +}
> +
> +int main(int argc, char **argv)
> +{
> + struct kvm_vm *vm;
> + FILE *csv = NULL;
> + int opt;
> +
> + while ((opt = getopt(argc, argv, "hn:o:")) != -1) {
> + switch (opt) {
> + case 'o':
> + csv = fopen(optarg, "w");
> + if (!csv) {
> + fprintf(stderr, "failed to open file '%s': %d\n",
> + optarg, errno);
> + exit(1);
> + }
> + break;
> + case 'n':
> + nr_iterations = strtoul(optarg, NULL, 0);
> + break;
> + default:
> + fprintf(stderr, "unrecognized option: '-%c'\n", opt);
> + /* fallthrough */
> + case 'h':
> + usage(argv[0]);
> + exit(1);
> + }
> + }
> +
> + vm = vm_create_default(VCPU_ID, 0, guest_main);
> + sync_global_to_guest(vm, nr_iterations);
> + ucall_init(vm, NULL);
> +
> + if (_vcpu_has_device_attr(vm, VCPU_ID, KVM_ARM_VCPU_TIMER_CTRL,
> + KVM_ARM_VCPU_TIMER_OFFSET_PTIMER)) {
> + print_skip("KVM_ARM_VCPU_TIMER_OFFSET_PTIMER not supported.");
> + exit(KSFT_SKIP);
> + }
> +
> + run_tests(vm, csv);
> + kvm_vm_free(vm);
> +
> + if (csv)
> + fclose(csv);
> +}
> --
> 2.32.0.402.g57bb445576-goog
>
> _______________________________________________
> kvmarm mailing list
> kvmarm@xxxxxxxxxxxxxxxxxxxxx
> https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
>
Thanks,
drew
