[Intel-gfx] [PATCH i-g-t 6/6] tests/i915_query: Engine queues tests
Lionel Landwerlin
lionel.g.landwerlin at intel.com
Wed Jun 6 14:33:30 UTC 2018
Just a few suggestions below. Otherwise looks good to me.
On 06/06/18 13:49, Tvrtko Ursulin wrote:
> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>
> Basic tests to cover engine queued/runnable/running metric as reported
> by the DRM_I915_QUERY_ENGINE_QUEUES query.
>
> v2:
> * Update ABI for i915 changes.
> * Use igt_spin_busywait_until_running.
> * Support no hardware contexts.
> * More comments. (Lionel Landwerlin)
> Chris Wilson:
> * Check for sw sync support.
> * Multiple contexts queued test.
> * Simplify context and bb allocation.
> * Fix asserts in the running subtest.
>
> v3:
> * Update for uAPI change.
>
> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> ---
> tests/i915_query.c | 442 +++++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 442 insertions(+)
>
> diff --git a/tests/i915_query.c b/tests/i915_query.c
> index c7de8cbd8371..588428749c50 100644
> --- a/tests/i915_query.c
> +++ b/tests/i915_query.c
> @@ -22,6 +22,7 @@
> */
>
> #include "igt.h"
> +#include "sw_sync.h"
>
> #include <limits.h>
>
> @@ -477,8 +478,414 @@ test_query_topology_known_pci_ids(int fd, int devid)
> free(topo_info);
> }
>
> +#define DRM_I915_QUERY_ENGINE_QUEUES 2
> +
> +struct drm_i915_query_engine_queues {
> + /** Engine class as in enum drm_i915_gem_engine_class. */
> + __u16 class;
> +
> + /** Engine instance number. */
> + __u16 instance;
> +
> + /** Number of requests with unresolved fences and dependencies. */
> + __u32 queued;
> +
> + /** Number of ready requests waiting on a slot on GPU. */
> + __u32 runnable;
> +
> + /** Number of requests executing on the GPU. */
> + __u32 running;
> +
> + __u32 rsvd[6];
> +};
I suppose this above would be in the updated i915_drm.h.
I thought you would have put it there as part of your first commit.
> +
> +static bool query_engine_queues_supported(int fd)
> +{
> + struct drm_i915_query_item item = {
> + .query_id = DRM_I915_QUERY_ENGINE_QUEUES,
> + };
> +
> + return __i915_query_items(fd, &item, 1) == 0 && item.length > 0;
> +}
> +
> +static void engine_queues_invalid(int fd)
> +{
> + struct drm_i915_query_engine_queues queues;
> + struct drm_i915_query_item item;
> + unsigned int len;
> + unsigned int i;
> +
> + /* Flags is MBZ. */
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.flags = 1;
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, -EINVAL);
> +
> + /* Length not zero and not greater or equal required size. */
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.length = 1;
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, -EINVAL);
> +
> + /* Query correct length. */
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + i915_query_items(fd, &item, 1);
> + igt_assert(item.length >= 0);
Could be item.length > 0?
> + len = item.length;
> +
> + /* Ivalid pointer. */
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.length = len;
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, -EFAULT);
Maybe verify ULONG_MAX too?
> +
> + /* Reserved fields are MBZ. */
> +
> + for (i = 0; i < ARRAY_SIZE(queues.rsvd); i++) {
> + memset(&queues, 0, sizeof(queues));
> + queues.rsvd[i] = 1;
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.length = len;
> + item.data_ptr = to_user_pointer(&queues);
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, -EINVAL);
> + }
> +
> + memset(&queues, 0, sizeof(queues));
> + queues.class = -1;
class = I915_ENGINE_CLASS_INVALID ?
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.length = len;
> + item.data_ptr = to_user_pointer(&queues);
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, -ENOENT);
> +
> + memset(&queues, 0, sizeof(queues));
> + queues.instance = -1;
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.length = len;
> + item.data_ptr = to_user_pointer(&queues);
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, -ENOENT);
> +}
> +
> +static void engine_queues(int fd, const struct intel_execution_engine2 *e)
> +{
> + struct drm_i915_query_engine_queues queues;
> + struct drm_i915_query_item item;
> + unsigned int len;
> +
> + /* Query required buffer length. */
> + memset(&queues, 0, sizeof(queues));
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.data_ptr = to_user_pointer(&queues);
> + i915_query_items(fd, &item, 1);
> + igt_assert(item.length >= 0);
item.length > 0?
> + igt_assert(item.length <= sizeof(queues));
I guess we can expect item.length == sizeof(queues) here.
> + len = item.length;
> +
> + /* Check length larger than required works and reports same length. */
> + memset(&queues, 0, sizeof(queues));
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.data_ptr = to_user_pointer(&queues);
> + item.length = len + 1;
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, len);
> +
> + /* Actual query. */
> + memset(&queues, 0, sizeof(queues));
> + queues.class = e->class;
> + queues.instance = e->instance;
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.data_ptr = to_user_pointer(&queues);
> + item.length = len;
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, len);
> +}
> +
> +static unsigned int e2ring(int gem_fd, const struct intel_execution_engine2 *e)
> +{
> + return gem_class_instance_to_eb_flags(gem_fd, e->class, e->instance);
> +}
> +
> +static void
> +__query_queues(int fd, const struct intel_execution_engine2 *e,
> + struct drm_i915_query_engine_queues *queues)
> +{
> + struct drm_i915_query_item item;
> +
> + memset(queues, 0, sizeof(*queues));
> + queues->class = e->class;
> + queues->instance = e->instance;
> + memset(&item, 0, sizeof(item));
> + item.query_id = DRM_I915_QUERY_ENGINE_QUEUES;
> + item.data_ptr = to_user_pointer(queues);
> + item.length = sizeof(*queues);
> + i915_query_items(fd, &item, 1);
> + igt_assert_eq(item.length, sizeof(*queues));
> +}
> +
> +#define TEST_CONTEXTS (1 << 0)
> +
> +/*
> + * Test that the reported number of queued (not ready for execution due fences
> + * or dependencies) requests on an engine is correct.
> + */
> +static void
> +engine_queued(int gem_fd, const struct intel_execution_engine2 *e,
> + unsigned int flags)
> +{
> + const unsigned long engine = e2ring(gem_fd, e);
> + struct drm_i915_query_engine_queues queues;
> + const uint32_t bbe = MI_BATCH_BUFFER_END;
> + const unsigned int max_rq = 10;
> + uint32_t queued[max_rq + 1];
> + unsigned int n, i;
> + uint32_t bo;
> +
> + igt_require_sw_sync();
> + if (flags & TEST_CONTEXTS)
> + gem_require_contexts(gem_fd);
> +
> + memset(queued, 0, sizeof(queued));
> +
> + bo = gem_create(gem_fd, 4096);
> + gem_write(gem_fd, bo, 4092, &bbe, sizeof(bbe));
> +
> + /* Create a specific queue depth of unready requests. */
> + for (n = 0; n <= max_rq; n++) {
> + int fence = -1;
> + IGT_CORK_FENCE(cork);
> +
> + gem_quiescent_gpu(gem_fd);
> +
> + /* Create a cork so we can create a dependency chain. */
> + if (n)
> + fence = igt_cork_plug(&cork, -1);
> +
> + /* Submit n unready requests depending on the cork. */
> + for (i = 0; i < n; i++) {
> + struct drm_i915_gem_exec_object2 obj = { };
> + struct drm_i915_gem_execbuffer2 eb = { };
> +
> + obj.handle = bo;
> +
> + eb.buffer_count = 1;
> + eb.buffers_ptr = to_user_pointer(&obj);
> +
> + eb.flags = engine | I915_EXEC_FENCE_IN;
> +
> + /*
> + * In context mode each submission is on a separate
> + * context.
> + */
> + if (flags & TEST_CONTEXTS)
> + eb.rsvd1 = gem_context_create(gem_fd);
> +
> + eb.rsvd2 = fence;
> +
> + gem_execbuf(gem_fd, &eb);
> +
> + if (flags & TEST_CONTEXTS)
> + gem_context_destroy(gem_fd, eb.rsvd1);
> + }
> +
> + /* Store reported queue depth to assert against later. */
> + __query_queues(gem_fd, e, &queues);
> + queued[n] = queues.queued;
> + igt_info("n=%u queued=%u\n", n, queued[n]);
> +
> + /* Unplug the queue and proceed to the next queue depth. */
> + if (fence >= 0)
> + igt_cork_unplug(&cork);
> +
> + gem_sync(gem_fd, bo);
> + }
> +
> + gem_close(gem_fd, bo);
> +
> + for (i = 0; i <= max_rq; i++)
> + igt_assert_eq(queued[i], i);
> +}
> +
> +static igt_spin_t * __spin_poll(int fd, uint32_t ctx, unsigned long flags)
> +{
> + if (gem_can_store_dword(fd, flags))
> + return __igt_spin_batch_new_poll(fd, ctx, flags);
> + else
> + return __igt_spin_batch_new(fd, ctx, flags, 0);
> +}
> +
> +static unsigned long __spin_wait(igt_spin_t *spin, unsigned int n)
> +{
> + struct timespec ts = { };
> + unsigned long t;
> +
> + igt_nsec_elapsed(&ts);
> +
> + if (spin->running) {
> + igt_spin_busywait_until_running(spin);
> + } else {
> + igt_debug("__spin_wait - usleep mode\n");
> + usleep(500e3); /* Better than nothing! */
> + }
> +
> + t = igt_nsec_elapsed(&ts);
> +
> + return spin->running ? t : 500e6 / n;
> +}
> +
> +/*
> + * Test that the number of requests ready for execution but waiting on space on
> + * GPU is correctly reported.
> + */
> +static void
> +engine_runnable(int gem_fd, const struct intel_execution_engine2 *e)
> +{
> + const unsigned long engine = e2ring(gem_fd, e);
> + struct drm_i915_query_engine_queues queues;
> + bool contexts = gem_has_contexts(gem_fd);
> + const unsigned int max_rq = 10;
> + igt_spin_t *spin[max_rq + 1];
> + uint32_t runnable[max_rq + 1];
> + uint32_t ctx[max_rq];
> + unsigned int n, i;
> +
> + memset(runnable, 0, sizeof(runnable));
> +
> + if (contexts) {
> + for (i = 0; i < max_rq; i++)
> + ctx[i] = gem_context_create(gem_fd);
> + }
> +
> + /*
> + * Submit different number of requests, potentially against different
> + * contexts, in order to provoke engine runnable metric returning
> + * different numbers.
> + */
> + for (n = 0; n <= max_rq; n++) {
> + gem_quiescent_gpu(gem_fd);
> +
> + for (i = 0; i < n; i++) {
> + uint32_t ctx_ = contexts ? ctx[i] : 0;
> +
> + if (i == 0)
> + spin[i] = __spin_poll(gem_fd, ctx_, engine);
> + else
> + spin[i] = __igt_spin_batch_new(gem_fd, ctx_,
> + engine, 0);
> + }
> +
> + if (n)
> + usleep(__spin_wait(spin[0], n) * n);
> +
> + /* Query and store for later checking. */
> + __query_queues(gem_fd, e, &queues);
> + runnable[n] = queues.runnable;
> + igt_info("n=%u runnable=%u\n", n, runnable[n]);
> +
> + for (i = 0; i < n; i++) {
> + igt_spin_batch_end(spin[i]);
> + gem_sync(gem_fd, spin[i]->handle);
> + igt_spin_batch_free(gem_fd, spin[i]);
> + }
> + }
> +
> + if (contexts) {
> + for (i = 0; i < max_rq; i++)
> + gem_context_destroy(gem_fd, ctx[i]);
> + }
> +
> + /*
> + * Check that the runnable metric is zero when nothing is submitted,
> + * and that it is greater than zero on the maximum queue depth.
> + *
> + * We cannot assert the exact value since we do not know how many
> + * requests can the submission backend consume.
> + */
> + igt_assert_eq(runnable[0], 0);
> + igt_assert(runnable[max_rq] > 0);
> +
> + /*
> + * We can only test that the runnable metric is growing by one if we
> + * have context support.
> + */
> + if (contexts)
> + igt_assert_eq(runnable[max_rq] - runnable[max_rq - 1], 1);
> +}
> +
> +/*
> + * Test that the number of requests currently executing on the GPU is correctly
> + * reported.
> + */
> +static void
> +engine_running(int gem_fd, const struct intel_execution_engine2 *e)
> +{
> + const unsigned long engine = e2ring(gem_fd, e);
> + struct drm_i915_query_engine_queues queues;
> + const unsigned int max_rq = 10;
> + igt_spin_t *spin[max_rq + 1];
> + uint32_t running[max_rq + 1];
> + unsigned int n, i;
> +
> + memset(running, 0, sizeof(running));
> + memset(spin, 0, sizeof(spin));
> +
> + /*
> + * Create various queue depths of requests against the same context to
> + * try and get submission backed execute one or more on the GPU.
> + */
> + for (n = 0; n <= max_rq; n++) {
> + gem_quiescent_gpu(gem_fd);
> +
> + for (i = 0; i < n; i++) {
> + if (i == 0)
> + spin[i] = __spin_poll(gem_fd, 0, engine);
> + else
> + spin[i] = __igt_spin_batch_new(gem_fd, 0,
> + engine, 0);
> + }
> +
> + if (n)
> + usleep(__spin_wait(spin[0], n) * n);
> +
> + /* Query and store for later checking. */
> + __query_queues(gem_fd, e, &queues);
> + running[n] = queues.running;
> + igt_info("n=%u running=%u\n", n, running[n]);
> +
> + for (i = 0; i < n; i++) {
> + igt_spin_batch_end(spin[i]);
> + gem_sync(gem_fd, spin[i]->handle);
> + igt_spin_batch_free(gem_fd, spin[i]);
> + }
> + }
> +
> + /*
> + * Check that the running metric is zero when nothing is submitted,
> + * one when one request is submitted, and at least one for any greater
> + * queue depth.
> + *
> + * We cannot assert the exact value since we do not know how many
> + * requests can the submission backend consume.
> + */
> + igt_assert_eq(running[0], 0);
> + for (i = 1; i <= max_rq; i++)
> + igt_assert(running[i] > 0);
> +}
> +
> igt_main
> {
> + const struct intel_execution_engine2 *e;
> int fd = -1;
> int devid;
>
> @@ -524,6 +931,41 @@ igt_main
> test_query_topology_known_pci_ids(fd, devid);
> }
>
> + igt_subtest_group {
> + igt_fixture {
> + igt_require(query_engine_queues_supported(fd));
> + }
> +
> + igt_subtest("engine-queues-invalid")
> + engine_queues_invalid(fd);
> +
> + __for_each_engine_class_instance(fd, e) {
> + igt_subtest_group {
> + igt_fixture {
> + gem_require_engine(fd,
> + e->class,
> + e->instance);
> + }
> +
> + igt_subtest_f("engine-queues-%s", e->name)
> + engine_queues(fd, e);
> +
> + igt_subtest_f("engine-queued-%s", e->name)
> + engine_queued(fd, e, 0);
> +
> + igt_subtest_f("engine-queued-contexts-%s",
> + e->name)
> + engine_queued(fd, e, TEST_CONTEXTS);
> +
> + igt_subtest_f("engine-runnable-%s", e->name)
> + engine_runnable(fd, e);
> +
> + igt_subtest_f("engine-running-%s", e->name)
> + engine_running(fd, e);
> + }
> + }
> + }
> +
> igt_fixture {
> close(fd);
> }
More information about the Intel-gfx
mailing list