[PATCH i-g-t, v4 1/2] tests/intel/xe_sysfs_preempt_timeout: Porting sysfs preempt test in xe
Riana Tauro
riana.tauro at intel.com
Thu Aug 1 14:10:10 UTC 2024
Hi Nakshtra
On 7/31/2024 8:00 PM, nakshtra.goyal at intel.com wrote:
> From: Nakshtra Goyal <nakshtra.goyal at intel.com>
>
> Using single exec, removing unnecessary headers,using variables of same
> type in single line according to comments in v3
Add a brief description of the test and version history separately
>
> Signed-off-by: Nakshtra Goyal <nakshtra.goyal at intel.com>
> ---
> tests/intel/xe_sysfs_preempt_timeout.c | 221 +++++++++++++++++++++++++
> tests/meson.build | 1 +
> 2 files changed, 222 insertions(+)
> create mode 100644 tests/intel/xe_sysfs_preempt_timeout.c
>
> diff --git a/tests/intel/xe_sysfs_preempt_timeout.c b/tests/intel/xe_sysfs_preempt_timeout.c
> new file mode 100644
> index 000000000..a77741fcf
> --- /dev/null
> +++ b/tests/intel/xe_sysfs_preempt_timeout.c
> @@ -0,0 +1,221 @@
> +// SPDX-License-Identifier: MIT
> +/*
> + * Copyright © 2024 Intel Corporation
> + */
> +
> +/**
> + * TEST: sysfs preempt timeout
> + * Category: Core
> + * Mega feature: SysMan
> + * Sub-category: SysMan tests
> + * Functionality: sysfs preempt timeout
> + * Feature: SMI, context
> + * Test category: SysMan
> + *
> + * SUBTEST: %s-timeout
> + * Description: Test to measure the delay from requestion the preemption to its
%s/requestion/request
Sentence seems wrong
> + * completion. Send down some non-preemptable workloads and then > + * request a switch to a higher priority context. The HW will not
> + * be able to respond, so the kernel will be forced to reset the hog.
> + * Test category: functionality test
> + *
> + * arg[1]:
> + *
> + * @preempt_timeout_us: preempt timeout us
> + */
> +
> +#include "xe/xe_spin.h"
> +#include "igt_syncobj.h"
> +#include "xe/xe_ioctl.h"
> +
> +#include "igt.h"
> +#include "igt_sysfs.h"
> +
order alphabetically
> +#define ATTR "preempt_timeout_us"
> +
> +static void set_preempt_timeout(int engine, unsigned int value)
> +{
> + unsigned int delay;
> +
> + igt_assert_lte(0, igt_sysfs_printf(engine, ATTR, "%u", value));
> + igt_sysfs_scanf(engine, ATTR, "%u", &delay);
> + igt_assert_eq(delay, value);
> +}
> +
> +static uint64_t __test_timeout(int xe, int engine, unsigned int timeout)
> +{
> + struct drm_xe_sync sync = {
> + .handle = syncobj_create(xe, 0),
> + .type = DRM_XE_SYNC_TYPE_SYNCOBJ,
> + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> + };
> +
> + struct drm_xe_exec exec = {
> + .num_batch_buffer = 1,
> + .num_syncs = 1,
> + .syncs = to_user_pointer(&sync),
> + };
> +/* high priority property */
indentation
> + struct drm_xe_ext_set_property ext = {
> + .base.next_extension = 0,
> + .base.name = DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY,
> + .property = DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY,
> + .value = 2, /* High priority */
> + };
> + struct drm_xe_engine_class_instance *hwe;
> + uint64_t ahnd = 0;
> + uint32_t exec_queues[2];
> + uint32_t vm[2];
> + uint32_t bo[2];
> + size_t bo_size;
> + struct xe_spin *spin[2];
> + struct timespec ts = {};
> + double elapsed;
> + uint64_t addr1 = 0x1a0000, addr2 = 0x100000;
> + int i, n_engines;
> +
> + n_engines = 0;
> + /* select an random engine */
> + i = rand() % xe_number_engines(xe);
> + xe_for_each_engine(xe, hwe) {
> + if (i == n_engines++)
> + break;
> + }
> +
> + /* set preempt timeout*/
> + set_preempt_timeout(engine, timeout);
> + vm[0] = xe_vm_create(xe, 0, 0);
> + vm[1] = xe_vm_create(xe, 0, 0);
> + exec_queues[0] = xe_exec_queue_create(xe, vm[0], hwe, 0);
> + exec_queues[1] = xe_exec_queue_create(xe, vm[1], hwe, to_user_pointer(&ext));
> + ahnd = intel_allocator_open(xe, 0, INTEL_ALLOCATOR_RELOC);
> + bo_size = xe_bb_size(xe, sizeof(*spin));
> + bo[0] = xe_bo_create(xe, vm[0], bo_size, vram_if_possible(xe, 0), 0);
> + spin[0] = xe_bo_map(xe, bo[0], bo_size);
> + xe_vm_bind_async(xe, vm[0], 0, bo[0], 0, addr1, bo_size, &sync, 1);
> + xe_spin_init_opts(spin[0], .addr = addr1,
> + .preempt = false);
> + exec.address = addr1;
> + exec.exec_queue_id = exec_queues[0];
> + xe_exec(xe, &exec);
> + xe_spin_wait_started(spin[0]);
> +
> + igt_nsec_elapsed(&ts);
> + bo[1] = xe_bo_create(xe, vm[1], bo_size, vram_if_possible(xe, 0), 0);
> + spin[1] = xe_bo_map(xe, bo[1], bo_size);
> + xe_vm_bind_sync(xe, vm[1], bo[1], 0, addr2, bo_size);
> + xe_spin_init_opts(spin[1], .addr = addr2);
> + exec.address = addr2;
> + exec.exec_queue_id = exec_queues[1];
> + xe_exec(xe, &exec);
> + xe_spin_wait_started(spin[1]);
> + elapsed = igt_nsec_elapsed(&ts);
> + xe_spin_end(spin[1]);
> +
> + xe_vm_unbind_async(xe, vm[0], 0, 0, addr1, bo_size, &sync, 1);
> + igt_assert(syncobj_wait(xe, &sync.handle, 1, INT64_MAX, 0, NULL));
> +
> + xe_spin_end(spin[0]);
> + xe_vm_unbind_sync(xe, vm[1], 0, addr2, bo_size);
> + syncobj_destroy(xe, sync.handle);
> +
> + xe_exec_queue_destroy(xe, exec_queues[0]);
> + xe_vm_destroy(xe, vm[0]);
> + xe_exec_queue_destroy(xe, exec_queues[1]);
> + xe_vm_destroy(xe, vm[1]);
> +
> + put_ahnd(ahnd);
> + return elapsed;
> +}
> +
> +static void test_timeout(int xe, int engine, const char **property)
> +{
> + int delays[] = { 1000, 50000, 100000, 500000 };
> + unsigned int saved;
> + uint64_t elapsed;
> + int epsilon;
> +
> + /*
> + * Send down some non-preemptable workloads and then request a
> + * switch to a higher priority context. The HW will not be able to
> + * respond, so the kernel will be forced to reset the hog. This
> + * timeout should match our specification, and so we can measure
> + * the delay from requesting the preemption to its completion.
> + */
> +
> + igt_assert(igt_sysfs_scanf(engine, ATTR, "%u", &saved) == 1);
> + igt_debug("Initial %s:%u\n", ATTR, saved);
> +
> + elapsed = __test_timeout(xe, engine, 1000);
> + epsilon = 2 * elapsed / 1000;
> + if (epsilon < 50000)
> + epsilon = 50000;
> + igt_info("Minimum timeout measured as %.3fus; setting error threshold to %dus\n",
> + elapsed * 1e-3, epsilon);
> + igt_require(epsilon < 10000000);
> +
> + for (int i = 0; i < ARRAY_SIZE(delays); i++) {
> + elapsed = __test_timeout(xe, engine, delays[i]);
> + igt_info("%s:%d, elapsed=%.3fus\n",
> + ATTR, delays[i], elapsed * 1e-3);
> +
> + /*
> + * We need to give a couple of jiffies slack for the scheduler
> + * timeouts and then a little more slack fr the overhead in
> + * submitting and measuring.
> + */
> + igt_assert_f(elapsed / 1000 / 1000 < delays[i] + epsilon,
> + "Forced preemption timeout exceeded request!\n");
indentation
> + }
> +
> + set_preempt_timeout(engine, saved);
> +}
> +
> +igt_main
> +{
> + static const struct {
> + const char *name;
> + void (*fn)(int, int, const char **);
> + } tests[] = {
> + { "timeout", test_timeout },
> + { }
> + };
There is only single test. Do we need the array?
> +
> + const char *property[][3] = { {"preempt_timeout_us",
> + "preempt_timeout_min",
> + "preempt_timeout_max"}, };
indentation
> + int count = sizeof(property) / sizeof(property[0]);
> + int xe = -1, sys_fd, gt;
> +
> + igt_fixture {
> + xe = drm_open_driver(DRIVER_XE);
> + xe_device_get(xe);
> +
> + sys_fd = igt_sysfs_open(xe);
> + igt_require(sys_fd != -1);
> + close(sys_fd);
Is this required?
> + }
> + int engines_fd = -1, gt_fd = -1;
> +
> + for (int i = 0; i < count; i++) {
> + for (typeof(*tests) *t = tests; t->name; t++) {
> + igt_subtest_with_dynamic_f("%s-%s", property[i][0], t->name) {
> + xe_for_each_gt(xe, gt) {
remove extra line
> +
> + gt_fd = xe_sysfs_gt_open(xe, gt);
> + igt_require(gt_fd != -1);
> + engines_fd = openat(gt_fd, "engines", O_RDONLY);
> + igt_require(engines_fd != -1);
> +
> + igt_sysfs_engines(xe, engines_fd, property[i], t->fn);
> + close(engines_fd);
> + close(gt_fd);
> + }
> + }
> + }
> + }
> + igt_fixture {
> + xe_device_put(xe);
> + close(xe);
> + }
> +}
> diff --git a/tests/meson.build b/tests/meson.build
> index e649466be..335c8b837 100644
> --- a/tests/meson.build
> +++ b/tests/meson.build
> @@ -315,6 +315,7 @@ intel_xe_progs = [
> 'xe_spin_batch',
> 'xe_sysfs_defaults',
> 'xe_sysfs_scheduler',
> + 'xe_sysfs_preempt_timeout',
alphabetical
Thanks,
Riana
> ]
>
> chamelium_progs = [
More information about the igt-dev
mailing list