2 * Copyright © 2016 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #include <linux/prime_numbers.h>
27 #include "../i915_selftest.h"
29 #include "mock_context.h"
30 #include "mock_gem_device.h"
32 static int igt_add_request(void *arg)
34 struct drm_i915_private *i915 = arg;
35 struct drm_i915_gem_request *request;
38 /* Basic preliminary test to create a request and let it loose! */
40 mutex_lock(&i915->drm.struct_mutex);
41 request = mock_request(i915->engine[RCS],
47 i915_add_request(request);
51 mutex_unlock(&i915->drm.struct_mutex);
55 static int igt_wait_request(void *arg)
57 const long T = HZ / 4;
58 struct drm_i915_private *i915 = arg;
59 struct drm_i915_gem_request *request;
62 /* Submit a request, then wait upon it */
64 mutex_lock(&i915->drm.struct_mutex);
65 request = mock_request(i915->engine[RCS], i915->kernel_context, T);
71 if (i915_wait_request(request, I915_WAIT_LOCKED, 0) != -ETIME) {
72 pr_err("request wait (busy query) succeeded (expected timeout before submit!)\n");
76 if (i915_wait_request(request, I915_WAIT_LOCKED, T) != -ETIME) {
77 pr_err("request wait succeeded (expected timeout before submit!)\n");
81 if (i915_gem_request_completed(request)) {
82 pr_err("request completed before submit!!\n");
86 i915_add_request(request);
88 if (i915_wait_request(request, I915_WAIT_LOCKED, 0) != -ETIME) {
89 pr_err("request wait (busy query) succeeded (expected timeout after submit!)\n");
93 if (i915_gem_request_completed(request)) {
94 pr_err("request completed immediately!\n");
98 if (i915_wait_request(request, I915_WAIT_LOCKED, T / 2) != -ETIME) {
99 pr_err("request wait succeeded (expected timeout!)\n");
103 if (i915_wait_request(request, I915_WAIT_LOCKED, T) == -ETIME) {
104 pr_err("request wait timed out!\n");
108 if (!i915_gem_request_completed(request)) {
109 pr_err("request not complete after waiting!\n");
113 if (i915_wait_request(request, I915_WAIT_LOCKED, T) == -ETIME) {
114 pr_err("request wait timed out when already complete!\n");
120 mock_device_flush(i915);
121 mutex_unlock(&i915->drm.struct_mutex);
125 static int igt_fence_wait(void *arg)
127 const long T = HZ / 4;
128 struct drm_i915_private *i915 = arg;
129 struct drm_i915_gem_request *request;
132 /* Submit a request, treat it as a fence and wait upon it */
134 mutex_lock(&i915->drm.struct_mutex);
135 request = mock_request(i915->engine[RCS], i915->kernel_context, T);
140 mutex_unlock(&i915->drm.struct_mutex); /* safe as we are single user */
142 if (dma_fence_wait_timeout(&request->fence, false, T) != -ETIME) {
143 pr_err("fence wait success before submit (expected timeout)!\n");
147 mutex_lock(&i915->drm.struct_mutex);
148 i915_add_request(request);
149 mutex_unlock(&i915->drm.struct_mutex);
151 if (dma_fence_is_signaled(&request->fence)) {
152 pr_err("fence signaled immediately!\n");
156 if (dma_fence_wait_timeout(&request->fence, false, T / 2) != -ETIME) {
157 pr_err("fence wait success after submit (expected timeout)!\n");
161 if (dma_fence_wait_timeout(&request->fence, false, T) <= 0) {
162 pr_err("fence wait timed out (expected success)!\n");
166 if (!dma_fence_is_signaled(&request->fence)) {
167 pr_err("fence unsignaled after waiting!\n");
171 if (dma_fence_wait_timeout(&request->fence, false, T) <= 0) {
172 pr_err("fence wait timed out when complete (expected success)!\n");
178 mutex_lock(&i915->drm.struct_mutex);
180 mock_device_flush(i915);
181 mutex_unlock(&i915->drm.struct_mutex);
185 static int igt_request_rewind(void *arg)
187 struct drm_i915_private *i915 = arg;
188 struct drm_i915_gem_request *request, *vip;
189 struct i915_gem_context *ctx[2];
192 mutex_lock(&i915->drm.struct_mutex);
193 ctx[0] = mock_context(i915, "A");
194 request = mock_request(i915->engine[RCS], ctx[0], 2 * HZ);
200 i915_gem_request_get(request);
201 i915_add_request(request);
203 ctx[1] = mock_context(i915, "B");
204 vip = mock_request(i915->engine[RCS], ctx[1], 0);
210 /* Simulate preemption by manual reordering */
211 if (!mock_cancel_request(request)) {
212 pr_err("failed to cancel request (already executed)!\n");
213 i915_add_request(vip);
216 i915_gem_request_get(vip);
217 i915_add_request(vip);
218 request->engine->submit_request(request);
220 mutex_unlock(&i915->drm.struct_mutex);
222 if (i915_wait_request(vip, 0, HZ) == -ETIME) {
223 pr_err("timed out waiting for high priority request, vip.seqno=%d, current seqno=%d\n",
224 vip->global_seqno, intel_engine_get_seqno(i915->engine[RCS]));
228 if (i915_gem_request_completed(request)) {
229 pr_err("low priority request already completed\n");
235 i915_gem_request_put(vip);
236 mutex_lock(&i915->drm.struct_mutex);
238 mock_context_close(ctx[1]);
239 i915_gem_request_put(request);
241 mock_context_close(ctx[0]);
242 mock_device_flush(i915);
243 mutex_unlock(&i915->drm.struct_mutex);
247 int i915_gem_request_mock_selftests(void)
249 static const struct i915_subtest tests[] = {
250 SUBTEST(igt_add_request),
251 SUBTEST(igt_wait_request),
252 SUBTEST(igt_fence_wait),
253 SUBTEST(igt_request_rewind),
255 struct drm_i915_private *i915;
258 i915 = mock_gem_device();
262 err = i915_subtests(tests, i915);
263 drm_dev_unref(&i915->drm);
269 struct drm_i915_private *i915;
273 unsigned int reset_count;
276 static int begin_live_test(struct live_test *t,
277 struct drm_i915_private *i915,
287 err = i915_gem_wait_for_idle(i915, I915_WAIT_LOCKED);
289 pr_err("%s(%s): failed to idle before, with err=%d!",
294 i915->gpu_error.missed_irq_rings = 0;
295 t->reset_count = i915_reset_count(&i915->gpu_error);
300 static int end_live_test(struct live_test *t)
302 struct drm_i915_private *i915 = t->i915;
304 i915_gem_retire_requests(i915);
306 if (wait_for(intel_engines_are_idle(i915), 10)) {
307 pr_err("%s(%s): GPU not idle\n", t->func, t->name);
311 if (t->reset_count != i915_reset_count(&i915->gpu_error)) {
312 pr_err("%s(%s): GPU was reset %d times!\n",
314 i915_reset_count(&i915->gpu_error) - t->reset_count);
318 if (i915->gpu_error.missed_irq_rings) {
319 pr_err("%s(%s): Missed interrupts on engines %lx\n",
320 t->func, t->name, i915->gpu_error.missed_irq_rings);
327 static int live_nop_request(void *arg)
329 struct drm_i915_private *i915 = arg;
330 struct intel_engine_cs *engine;
335 /* Submit various sized batches of empty requests, to each engine
336 * (individually), and wait for the batch to complete. We can check
337 * the overhead of submitting requests to the hardware.
340 mutex_lock(&i915->drm.struct_mutex);
342 for_each_engine(engine, i915, id) {
343 IGT_TIMEOUT(end_time);
344 struct drm_i915_gem_request *request;
345 unsigned long n, prime;
346 ktime_t times[2] = {};
348 err = begin_live_test(&t, i915, __func__, engine->name);
352 for_each_prime_number_from(prime, 1, 8192) {
353 times[1] = ktime_get_raw();
355 for (n = 0; n < prime; n++) {
356 request = i915_gem_request_alloc(engine,
357 i915->kernel_context);
358 if (IS_ERR(request)) {
359 err = PTR_ERR(request);
363 /* This space is left intentionally blank.
365 * We do not actually want to perform any
366 * action with this request, we just want
367 * to measure the latency in allocation
368 * and submission of our breadcrumbs -
369 * ensuring that the bare request is sufficient
370 * for the system to work (i.e. proper HEAD
371 * tracking of the rings, interrupt handling,
372 * etc). It also gives us the lowest bounds
376 i915_add_request(request);
378 i915_wait_request(request,
380 MAX_SCHEDULE_TIMEOUT);
382 times[1] = ktime_sub(ktime_get_raw(), times[1]);
386 if (__igt_timeout(end_time, NULL))
390 err = end_live_test(&t);
394 pr_info("Request latencies on %s: 1 = %lluns, %lu = %lluns\n",
396 ktime_to_ns(times[0]),
397 prime, div64_u64(ktime_to_ns(times[1]), prime));
401 mutex_unlock(&i915->drm.struct_mutex);
405 static struct i915_vma *empty_batch(struct drm_i915_private *i915)
407 struct drm_i915_gem_object *obj;
408 struct i915_vma *vma;
412 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
414 return ERR_CAST(obj);
416 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
421 *cmd = MI_BATCH_BUFFER_END;
422 i915_gem_object_unpin_map(obj);
424 err = i915_gem_object_set_to_gtt_domain(obj, false);
428 vma = i915_vma_instance(obj, &i915->ggtt.base, NULL);
434 err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_GLOBAL);
441 i915_gem_object_put(obj);
445 static struct drm_i915_gem_request *
446 empty_request(struct intel_engine_cs *engine,
447 struct i915_vma *batch)
449 struct drm_i915_gem_request *request;
452 request = i915_gem_request_alloc(engine,
453 engine->i915->kernel_context);
457 err = engine->emit_flush(request, EMIT_INVALIDATE);
461 err = i915_switch_context(request);
465 err = engine->emit_bb_start(request,
468 I915_DISPATCH_SECURE);
473 __i915_add_request(request, err == 0);
474 return err ? ERR_PTR(err) : request;
477 static int live_empty_request(void *arg)
479 struct drm_i915_private *i915 = arg;
480 struct intel_engine_cs *engine;
482 struct i915_vma *batch;
486 /* Submit various sized batches of empty requests, to each engine
487 * (individually), and wait for the batch to complete. We can check
488 * the overhead of submitting requests to the hardware.
491 mutex_lock(&i915->drm.struct_mutex);
493 batch = empty_batch(i915);
495 err = PTR_ERR(batch);
499 for_each_engine(engine, i915, id) {
500 IGT_TIMEOUT(end_time);
501 struct drm_i915_gem_request *request;
502 unsigned long n, prime;
503 ktime_t times[2] = {};
505 err = begin_live_test(&t, i915, __func__, engine->name);
509 /* Warmup / preload */
510 request = empty_request(engine, batch);
511 if (IS_ERR(request)) {
512 err = PTR_ERR(request);
515 i915_wait_request(request,
517 MAX_SCHEDULE_TIMEOUT);
519 for_each_prime_number_from(prime, 1, 8192) {
520 times[1] = ktime_get_raw();
522 for (n = 0; n < prime; n++) {
523 request = empty_request(engine, batch);
524 if (IS_ERR(request)) {
525 err = PTR_ERR(request);
529 i915_wait_request(request,
531 MAX_SCHEDULE_TIMEOUT);
533 times[1] = ktime_sub(ktime_get_raw(), times[1]);
537 if (__igt_timeout(end_time, NULL))
541 err = end_live_test(&t);
545 pr_info("Batch latencies on %s: 1 = %lluns, %lu = %lluns\n",
547 ktime_to_ns(times[0]),
548 prime, div64_u64(ktime_to_ns(times[1]), prime));
552 i915_vma_unpin(batch);
555 mutex_unlock(&i915->drm.struct_mutex);
559 static struct i915_vma *recursive_batch(struct drm_i915_private *i915)
561 struct i915_gem_context *ctx = i915->kernel_context;
562 struct i915_address_space *vm = ctx->ppgtt ? &ctx->ppgtt->base : &i915->ggtt.base;
563 struct drm_i915_gem_object *obj;
564 const int gen = INTEL_GEN(i915);
565 struct i915_vma *vma;
569 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
571 return ERR_CAST(obj);
573 vma = i915_vma_instance(obj, vm, NULL);
579 err = i915_vma_pin(vma, 0, 0, PIN_USER);
583 err = i915_gem_object_set_to_wc_domain(obj, true);
587 cmd = i915_gem_object_pin_map(obj, I915_MAP_WC);
594 *cmd++ = MI_BATCH_BUFFER_START | 1 << 8 | 1;
595 *cmd++ = lower_32_bits(vma->node.start);
596 *cmd++ = upper_32_bits(vma->node.start);
597 } else if (gen >= 6) {
598 *cmd++ = MI_BATCH_BUFFER_START | 1 << 8;
599 *cmd++ = lower_32_bits(vma->node.start);
600 } else if (gen >= 4) {
601 *cmd++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
602 *cmd++ = lower_32_bits(vma->node.start);
604 *cmd++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT | 1;
605 *cmd++ = lower_32_bits(vma->node.start);
607 *cmd++ = MI_BATCH_BUFFER_END; /* terminate early in case of error */
610 i915_gem_object_unpin_map(obj);
615 i915_gem_object_put(obj);
619 static int recursive_batch_resolve(struct i915_vma *batch)
623 cmd = i915_gem_object_pin_map(batch->obj, I915_MAP_WC);
627 *cmd = MI_BATCH_BUFFER_END;
630 i915_gem_object_unpin_map(batch->obj);
635 static int live_all_engines(void *arg)
637 struct drm_i915_private *i915 = arg;
638 struct intel_engine_cs *engine;
639 struct drm_i915_gem_request *request[I915_NUM_ENGINES];
640 struct i915_vma *batch;
645 /* Check we can submit requests to all engines simultaneously. We
646 * send a recursive batch to each engine - checking that we don't
647 * block doing so, and that they don't complete too soon.
650 mutex_lock(&i915->drm.struct_mutex);
652 err = begin_live_test(&t, i915, __func__, "");
656 batch = recursive_batch(i915);
658 err = PTR_ERR(batch);
659 pr_err("%s: Unable to create batch, err=%d\n", __func__, err);
663 for_each_engine(engine, i915, id) {
664 request[id] = i915_gem_request_alloc(engine,
665 i915->kernel_context);
666 if (IS_ERR(request[id])) {
667 err = PTR_ERR(request[id]);
668 pr_err("%s: Request allocation failed with err=%d\n",
673 err = engine->emit_flush(request[id], EMIT_INVALIDATE);
676 err = i915_switch_context(request[id]);
679 err = engine->emit_bb_start(request[id],
684 request[id]->batch = batch;
686 if (!i915_gem_object_has_active_reference(batch->obj)) {
687 i915_gem_object_get(batch->obj);
688 i915_gem_object_set_active_reference(batch->obj);
691 i915_vma_move_to_active(batch, request[id], 0);
692 i915_gem_request_get(request[id]);
693 i915_add_request(request[id]);
696 for_each_engine(engine, i915, id) {
697 if (i915_gem_request_completed(request[id])) {
698 pr_err("%s(%s): request completed too early!\n",
699 __func__, engine->name);
705 err = recursive_batch_resolve(batch);
707 pr_err("%s: failed to resolve batch, err=%d\n", __func__, err);
711 for_each_engine(engine, i915, id) {
714 timeout = i915_wait_request(request[id],
716 MAX_SCHEDULE_TIMEOUT);
719 pr_err("%s: error waiting for request on %s, err=%d\n",
720 __func__, engine->name, err);
724 GEM_BUG_ON(!i915_gem_request_completed(request[id]));
725 i915_gem_request_put(request[id]);
729 err = end_live_test(&t);
732 for_each_engine(engine, i915, id)
734 i915_gem_request_put(request[id]);
735 i915_vma_unpin(batch);
738 mutex_unlock(&i915->drm.struct_mutex);
742 static int live_sequential_engines(void *arg)
744 struct drm_i915_private *i915 = arg;
745 struct drm_i915_gem_request *request[I915_NUM_ENGINES] = {};
746 struct drm_i915_gem_request *prev = NULL;
747 struct intel_engine_cs *engine;
752 /* Check we can submit requests to all engines sequentially, such
753 * that each successive request waits for the earlier ones. This
754 * tests that we don't execute requests out of order, even though
755 * they are running on independent engines.
758 mutex_lock(&i915->drm.struct_mutex);
760 err = begin_live_test(&t, i915, __func__, "");
764 for_each_engine(engine, i915, id) {
765 struct i915_vma *batch;
767 batch = recursive_batch(i915);
769 err = PTR_ERR(batch);
770 pr_err("%s: Unable to create batch for %s, err=%d\n",
771 __func__, engine->name, err);
775 request[id] = i915_gem_request_alloc(engine,
776 i915->kernel_context);
777 if (IS_ERR(request[id])) {
778 err = PTR_ERR(request[id]);
779 pr_err("%s: Request allocation failed for %s with err=%d\n",
780 __func__, engine->name, err);
785 err = i915_gem_request_await_dma_fence(request[id],
788 i915_add_request(request[id]);
789 pr_err("%s: Request await failed for %s with err=%d\n",
790 __func__, engine->name, err);
795 err = engine->emit_flush(request[id], EMIT_INVALIDATE);
798 err = i915_switch_context(request[id]);
801 err = engine->emit_bb_start(request[id],
806 request[id]->batch = batch;
808 i915_vma_move_to_active(batch, request[id], 0);
809 i915_gem_object_set_active_reference(batch->obj);
812 i915_gem_request_get(request[id]);
813 i915_add_request(request[id]);
818 for_each_engine(engine, i915, id) {
821 if (i915_gem_request_completed(request[id])) {
822 pr_err("%s(%s): request completed too early!\n",
823 __func__, engine->name);
828 err = recursive_batch_resolve(request[id]->batch);
830 pr_err("%s: failed to resolve batch, err=%d\n",
835 timeout = i915_wait_request(request[id],
837 MAX_SCHEDULE_TIMEOUT);
840 pr_err("%s: error waiting for request on %s, err=%d\n",
841 __func__, engine->name, err);
845 GEM_BUG_ON(!i915_gem_request_completed(request[id]));
848 err = end_live_test(&t);
851 for_each_engine(engine, i915, id) {
857 cmd = i915_gem_object_pin_map(request[id]->batch->obj,
860 *cmd = MI_BATCH_BUFFER_END;
862 i915_gem_object_unpin_map(request[id]->batch->obj);
865 i915_vma_put(request[id]->batch);
866 i915_gem_request_put(request[id]);
869 mutex_unlock(&i915->drm.struct_mutex);
873 int i915_gem_request_live_selftests(struct drm_i915_private *i915)
875 static const struct i915_subtest tests[] = {
876 SUBTEST(live_nop_request),
877 SUBTEST(live_all_engines),
878 SUBTEST(live_sequential_engines),
879 SUBTEST(live_empty_request),
881 return i915_subtests(tests, i915);