2 * SPDX-License-Identifier: MIT
4 * Copyright © 2017 Intel Corporation
7 #include <linux/prime_numbers.h>
9 #include "gem/i915_gem_pm.h"
10 #include "gt/intel_engine_pm.h"
11 #include "gt/intel_gt.h"
12 #include "gt/intel_gt_requests.h"
13 #include "gt/intel_reset.h"
14 #include "i915_selftest.h"
16 #include "gem/selftests/igt_gem_utils.h"
17 #include "selftests/i915_random.h"
18 #include "selftests/igt_flush_test.h"
19 #include "selftests/igt_live_test.h"
20 #include "selftests/igt_reset.h"
21 #include "selftests/igt_spinner.h"
22 #include "selftests/mock_drm.h"
23 #include "selftests/mock_gem_device.h"
25 #include "huge_gem_object.h"
26 #include "igt_gem_utils.h"
28 #define DW_PER_PAGE (PAGE_SIZE / sizeof(u32))
30 static inline struct i915_address_space *ctx_vm(struct i915_gem_context *ctx)
32 /* single threaded, private ctx */
33 return rcu_dereference_protected(ctx->vm, true);
36 static int live_nop_switch(void *arg)
38 const unsigned int nctx = 1024;
39 struct drm_i915_private *i915 = arg;
40 struct intel_engine_cs *engine;
41 struct i915_gem_context **ctx;
42 struct igt_live_test t;
48 * Create as many contexts as we can feasibly get away with
49 * and check we can switch between them rapidly.
51 * Serves as very simple stress test for submission and HW switching
55 if (!DRIVER_CAPS(i915)->has_logical_contexts)
58 file = mock_file(i915);
62 ctx = kcalloc(nctx, sizeof(*ctx), GFP_KERNEL);
68 for (n = 0; n < nctx; n++) {
69 ctx[n] = live_context(i915, file);
71 err = PTR_ERR(ctx[n]);
76 for_each_uabi_engine(engine, i915) {
77 struct i915_request *rq = NULL;
78 unsigned long end_time, prime;
79 ktime_t times[2] = {};
81 times[0] = ktime_get_raw();
82 for (n = 0; n < nctx; n++) {
83 struct i915_request *this;
85 this = igt_request_alloc(ctx[n], engine);
91 i915_request_await_dma_fence(this, &rq->fence);
94 rq = i915_request_get(this);
95 i915_request_add(this);
97 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
98 pr_err("Failed to populated %d contexts\n", nctx);
99 intel_gt_set_wedged(&i915->gt);
100 i915_request_put(rq);
104 i915_request_put(rq);
106 times[1] = ktime_get_raw();
108 pr_info("Populated %d contexts on %s in %lluns\n",
109 nctx, engine->name, ktime_to_ns(times[1] - times[0]));
111 err = igt_live_test_begin(&t, i915, __func__, engine->name);
115 end_time = jiffies + i915_selftest.timeout_jiffies;
116 for_each_prime_number_from(prime, 2, 8192) {
117 times[1] = ktime_get_raw();
120 for (n = 0; n < prime; n++) {
121 struct i915_request *this;
123 this = igt_request_alloc(ctx[n % nctx], engine);
129 if (rq) { /* Force submission order */
130 i915_request_await_dma_fence(this, &rq->fence);
131 i915_request_put(rq);
135 * This space is left intentionally blank.
137 * We do not actually want to perform any
138 * action with this request, we just want
139 * to measure the latency in allocation
140 * and submission of our breadcrumbs -
141 * ensuring that the bare request is sufficient
142 * for the system to work (i.e. proper HEAD
143 * tracking of the rings, interrupt handling,
144 * etc). It also gives us the lowest bounds
148 rq = i915_request_get(this);
149 i915_request_add(this);
152 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
153 pr_err("Switching between %ld contexts timed out\n",
155 intel_gt_set_wedged(&i915->gt);
156 i915_request_put(rq);
159 i915_request_put(rq);
161 times[1] = ktime_sub(ktime_get_raw(), times[1]);
165 if (__igt_timeout(end_time, NULL))
169 err = igt_live_test_end(&t);
173 pr_info("Switch latencies on %s: 1 = %lluns, %lu = %lluns\n",
175 ktime_to_ns(times[0]),
176 prime - 1, div64_u64(ktime_to_ns(times[1]), prime - 1));
184 struct parallel_switch {
185 struct task_struct *tsk;
186 struct intel_context *ce[2];
189 static int __live_parallel_switch1(void *data)
191 struct parallel_switch *arg = data;
192 IGT_TIMEOUT(end_time);
197 struct i915_request *rq = NULL;
201 for (n = 0; !err && n < ARRAY_SIZE(arg->ce); n++) {
202 struct i915_request *prev = rq;
204 rq = i915_request_create(arg->ce[n]);
206 i915_request_put(prev);
210 i915_request_get(rq);
212 err = i915_request_await_dma_fence(rq, &prev->fence);
213 i915_request_put(prev);
216 i915_request_add(rq);
218 if (i915_request_wait(rq, 0, HZ / 5) < 0)
220 i915_request_put(rq);
225 } while (!__igt_timeout(end_time, NULL));
227 pr_info("%s: %lu switches (sync)\n", arg->ce[0]->engine->name, count);
231 static int __live_parallel_switchN(void *data)
233 struct parallel_switch *arg = data;
234 struct i915_request *rq = NULL;
235 IGT_TIMEOUT(end_time);
241 for (n = 0; n < ARRAY_SIZE(arg->ce); n++) {
242 struct i915_request *prev = rq;
245 rq = i915_request_create(arg->ce[n]);
247 i915_request_put(prev);
251 i915_request_get(rq);
253 err = i915_request_await_dma_fence(rq, &prev->fence);
254 i915_request_put(prev);
257 i915_request_add(rq);
259 i915_request_put(rq);
265 } while (!__igt_timeout(end_time, NULL));
266 i915_request_put(rq);
268 pr_info("%s: %lu switches (many)\n", arg->ce[0]->engine->name, count);
272 static int live_parallel_switch(void *arg)
274 struct drm_i915_private *i915 = arg;
275 static int (* const func[])(void *arg) = {
276 __live_parallel_switch1,
277 __live_parallel_switchN,
280 struct parallel_switch *data = NULL;
281 struct i915_gem_engines *engines;
282 struct i915_gem_engines_iter it;
283 int (* const *fn)(void *arg);
284 struct i915_gem_context *ctx;
285 struct intel_context *ce;
291 * Check we can process switches on all engines simultaneously.
294 if (!DRIVER_CAPS(i915)->has_logical_contexts)
297 file = mock_file(i915);
299 return PTR_ERR(file);
301 ctx = live_context(i915, file);
307 engines = i915_gem_context_lock_engines(ctx);
308 count = engines->num_engines;
310 data = kcalloc(count, sizeof(*data), GFP_KERNEL);
312 i915_gem_context_unlock_engines(ctx);
317 m = 0; /* Use the first context as our template for the engines */
318 for_each_gem_engine(ce, engines, it) {
319 err = intel_context_pin(ce);
321 i915_gem_context_unlock_engines(ctx);
324 data[m++].ce[0] = intel_context_get(ce);
326 i915_gem_context_unlock_engines(ctx);
328 /* Clone the same set of engines into the other contexts */
329 for (n = 1; n < ARRAY_SIZE(data->ce); n++) {
330 ctx = live_context(i915, file);
336 for (m = 0; m < count; m++) {
340 ce = intel_context_create(data[m].ce[0]->engine);
344 err = intel_context_pin(ce);
346 intel_context_put(ce);
354 for (fn = func; !err && *fn; fn++) {
355 struct igt_live_test t;
358 err = igt_live_test_begin(&t, i915, __func__, "");
362 for (n = 0; n < count; n++) {
366 data[n].tsk = kthread_run(*fn, &data[n],
368 data[n].ce[0]->engine->name);
369 if (IS_ERR(data[n].tsk)) {
370 err = PTR_ERR(data[n].tsk);
373 get_task_struct(data[n].tsk);
376 yield(); /* start all threads before we kthread_stop() */
378 for (n = 0; n < count; n++) {
381 if (IS_ERR_OR_NULL(data[n].tsk))
384 status = kthread_stop(data[n].tsk);
388 put_task_struct(data[n].tsk);
392 if (igt_live_test_end(&t))
397 for (n = 0; n < count; n++) {
398 for (m = 0; m < ARRAY_SIZE(data->ce); m++) {
402 intel_context_unpin(data[n].ce[m]);
403 intel_context_put(data[n].ce[m]);
412 static unsigned long real_page_count(struct drm_i915_gem_object *obj)
414 return huge_gem_object_phys_size(obj) >> PAGE_SHIFT;
417 static unsigned long fake_page_count(struct drm_i915_gem_object *obj)
419 return huge_gem_object_dma_size(obj) >> PAGE_SHIFT;
422 static int gpu_fill(struct intel_context *ce,
423 struct drm_i915_gem_object *obj,
426 struct i915_vma *vma;
429 GEM_BUG_ON(obj->base.size > ce->vm->total);
430 GEM_BUG_ON(!intel_engine_can_store_dword(ce->engine));
432 vma = i915_vma_instance(obj, ce->vm, NULL);
436 err = i915_vma_pin(vma, 0, 0, PIN_HIGH | PIN_USER);
441 * Within the GTT the huge objects maps every page onto
442 * its 1024 real pages (using phys_pfn = dma_pfn % 1024).
443 * We set the nth dword within the page using the nth
444 * mapping via the GTT - this should exercise the GTT mapping
445 * whilst checking that each context provides a unique view
448 err = igt_gpu_fill_dw(ce, vma,
449 (dw * real_page_count(obj)) << PAGE_SHIFT |
451 real_page_count(obj),
458 static int cpu_fill(struct drm_i915_gem_object *obj, u32 value)
460 const bool has_llc = HAS_LLC(to_i915(obj->base.dev));
461 unsigned int n, m, need_flush;
464 i915_gem_object_lock(obj, NULL);
465 err = i915_gem_object_prepare_write(obj, &need_flush);
469 for (n = 0; n < real_page_count(obj); n++) {
472 map = kmap_atomic(i915_gem_object_get_page(obj, n));
473 for (m = 0; m < DW_PER_PAGE; m++)
476 drm_clflush_virt_range(map, PAGE_SIZE);
480 i915_gem_object_finish_access(obj);
481 obj->read_domains = I915_GEM_DOMAIN_GTT | I915_GEM_DOMAIN_CPU;
482 obj->write_domain = 0;
484 i915_gem_object_unlock(obj);
488 static noinline int cpu_check(struct drm_i915_gem_object *obj,
489 unsigned int idx, unsigned int max)
491 unsigned int n, m, needs_flush;
494 i915_gem_object_lock(obj, NULL);
495 err = i915_gem_object_prepare_read(obj, &needs_flush);
499 for (n = 0; n < real_page_count(obj); n++) {
502 map = kmap_atomic(i915_gem_object_get_page(obj, n));
503 if (needs_flush & CLFLUSH_BEFORE)
504 drm_clflush_virt_range(map, PAGE_SIZE);
506 for (m = 0; m < max; m++) {
508 pr_err("%pS: Invalid value at object %d page %d/%ld, offset %d/%d: found %x expected %x\n",
509 __builtin_return_address(0), idx,
510 n, real_page_count(obj), m, max,
517 for (; m < DW_PER_PAGE; m++) {
518 if (map[m] != STACK_MAGIC) {
519 pr_err("%pS: Invalid value at object %d page %d, offset %d: found %x expected %x (uninitialised)\n",
520 __builtin_return_address(0), idx, n, m,
521 map[m], STACK_MAGIC);
533 i915_gem_object_finish_access(obj);
535 i915_gem_object_unlock(obj);
539 static int file_add_object(struct file *file, struct drm_i915_gem_object *obj)
543 GEM_BUG_ON(obj->base.handle_count);
545 /* tie the object to the drm_file for easy reaping */
546 err = idr_alloc(&to_drm_file(file)->object_idr,
547 &obj->base, 1, 0, GFP_KERNEL);
551 i915_gem_object_get(obj);
552 obj->base.handle_count++;
556 static struct drm_i915_gem_object *
557 create_test_object(struct i915_address_space *vm,
559 struct list_head *objects)
561 struct drm_i915_gem_object *obj;
565 /* Keep in GEM's good graces */
566 intel_gt_retire_requests(vm->gt);
568 size = min(vm->total / 2, 1024ull * DW_PER_PAGE * PAGE_SIZE);
569 size = round_down(size, DW_PER_PAGE * PAGE_SIZE);
571 obj = huge_gem_object(vm->i915, DW_PER_PAGE * PAGE_SIZE, size);
575 err = file_add_object(file, obj);
576 i915_gem_object_put(obj);
580 err = cpu_fill(obj, STACK_MAGIC);
582 pr_err("Failed to fill object with cpu, err=%d\n",
587 list_add_tail(&obj->st_link, objects);
591 static unsigned long max_dwords(struct drm_i915_gem_object *obj)
593 unsigned long npages = fake_page_count(obj);
595 GEM_BUG_ON(!IS_ALIGNED(npages, DW_PER_PAGE));
596 return npages / DW_PER_PAGE;
599 static void throttle_release(struct i915_request **q, int count)
603 for (i = 0; i < count; i++) {
604 if (IS_ERR_OR_NULL(q[i]))
607 i915_request_put(fetch_and_zero(&q[i]));
611 static int throttle(struct intel_context *ce,
612 struct i915_request **q, int count)
616 if (!IS_ERR_OR_NULL(q[0])) {
617 if (i915_request_wait(q[0],
618 I915_WAIT_INTERRUPTIBLE,
619 MAX_SCHEDULE_TIMEOUT) < 0)
622 i915_request_put(q[0]);
625 for (i = 0; i < count - 1; i++)
628 q[i] = intel_context_create_request(ce);
630 return PTR_ERR(q[i]);
632 i915_request_get(q[i]);
633 i915_request_add(q[i]);
638 static int igt_ctx_exec(void *arg)
640 struct drm_i915_private *i915 = arg;
641 struct intel_engine_cs *engine;
645 * Create a few different contexts (with different mm) and write
646 * through each ctx/mm using the GPU making sure those writes end
647 * up in the expected pages of our obj.
650 if (!DRIVER_CAPS(i915)->has_logical_contexts)
653 for_each_uabi_engine(engine, i915) {
654 struct drm_i915_gem_object *obj = NULL;
655 unsigned long ncontexts, ndwords, dw;
656 struct i915_request *tq[5] = {};
657 struct igt_live_test t;
658 IGT_TIMEOUT(end_time);
662 if (!intel_engine_can_store_dword(engine))
665 if (!engine->context_size)
666 continue; /* No logical context support in HW */
668 file = mock_file(i915);
670 return PTR_ERR(file);
672 err = igt_live_test_begin(&t, i915, __func__, engine->name);
679 while (!time_after(jiffies, end_time)) {
680 struct i915_gem_context *ctx;
681 struct intel_context *ce;
683 ctx = kernel_context(i915);
689 ce = i915_gem_context_get_engine(ctx, engine->legacy_idx);
690 GEM_BUG_ON(IS_ERR(ce));
693 obj = create_test_object(ce->vm, file, &objects);
696 intel_context_put(ce);
697 kernel_context_close(ctx);
702 err = gpu_fill(ce, obj, dw);
704 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) [full-ppgtt? %s], err=%d\n",
705 ndwords, dw, max_dwords(obj),
707 yesno(!!rcu_access_pointer(ctx->vm)),
709 intel_context_put(ce);
710 kernel_context_close(ctx);
714 err = throttle(ce, tq, ARRAY_SIZE(tq));
716 intel_context_put(ce);
717 kernel_context_close(ctx);
721 if (++dw == max_dwords(obj)) {
729 intel_context_put(ce);
730 kernel_context_close(ctx);
733 pr_info("Submitted %lu contexts to %s, filling %lu dwords\n",
734 ncontexts, engine->name, ndwords);
737 list_for_each_entry(obj, &objects, st_link) {
739 min_t(unsigned int, ndwords - dw, max_dwords(obj));
741 err = cpu_check(obj, ncontexts++, rem);
749 throttle_release(tq, ARRAY_SIZE(tq));
750 if (igt_live_test_end(&t))
757 i915_gem_drain_freed_objects(i915);
763 static int igt_shared_ctx_exec(void *arg)
765 struct drm_i915_private *i915 = arg;
766 struct i915_request *tq[5] = {};
767 struct i915_gem_context *parent;
768 struct intel_engine_cs *engine;
769 struct igt_live_test t;
774 * Create a few different contexts with the same mm and write
775 * through each ctx using the GPU making sure those writes end
776 * up in the expected pages of our obj.
778 if (!DRIVER_CAPS(i915)->has_logical_contexts)
781 file = mock_file(i915);
783 return PTR_ERR(file);
785 parent = live_context(i915, file);
786 if (IS_ERR(parent)) {
787 err = PTR_ERR(parent);
791 if (!parent->vm) { /* not full-ppgtt; nothing to share */
796 err = igt_live_test_begin(&t, i915, __func__, "");
800 for_each_uabi_engine(engine, i915) {
801 unsigned long ncontexts, ndwords, dw;
802 struct drm_i915_gem_object *obj = NULL;
803 IGT_TIMEOUT(end_time);
806 if (!intel_engine_can_store_dword(engine))
812 while (!time_after(jiffies, end_time)) {
813 struct i915_gem_context *ctx;
814 struct intel_context *ce;
816 ctx = kernel_context(i915);
822 mutex_lock(&ctx->mutex);
823 __assign_ppgtt(ctx, ctx_vm(parent));
824 mutex_unlock(&ctx->mutex);
826 ce = i915_gem_context_get_engine(ctx, engine->legacy_idx);
827 GEM_BUG_ON(IS_ERR(ce));
830 obj = create_test_object(ctx_vm(parent),
834 intel_context_put(ce);
835 kernel_context_close(ctx);
840 err = gpu_fill(ce, obj, dw);
842 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) [full-ppgtt? %s], err=%d\n",
843 ndwords, dw, max_dwords(obj),
845 yesno(!!rcu_access_pointer(ctx->vm)),
847 intel_context_put(ce);
848 kernel_context_close(ctx);
852 err = throttle(ce, tq, ARRAY_SIZE(tq));
854 intel_context_put(ce);
855 kernel_context_close(ctx);
859 if (++dw == max_dwords(obj)) {
867 intel_context_put(ce);
868 kernel_context_close(ctx);
870 pr_info("Submitted %lu contexts to %s, filling %lu dwords\n",
871 ncontexts, engine->name, ndwords);
874 list_for_each_entry(obj, &objects, st_link) {
876 min_t(unsigned int, ndwords - dw, max_dwords(obj));
878 err = cpu_check(obj, ncontexts++, rem);
885 i915_gem_drain_freed_objects(i915);
888 throttle_release(tq, ARRAY_SIZE(tq));
889 if (igt_live_test_end(&t))
896 static int rpcs_query_batch(struct drm_i915_gem_object *rpcs, struct i915_vma *vma)
900 GEM_BUG_ON(INTEL_GEN(vma->vm->i915) < 8);
902 cmd = i915_gem_object_pin_map(rpcs, I915_MAP_WB);
906 *cmd++ = MI_STORE_REGISTER_MEM_GEN8;
907 *cmd++ = i915_mmio_reg_offset(GEN8_R_PWR_CLK_STATE);
908 *cmd++ = lower_32_bits(vma->node.start);
909 *cmd++ = upper_32_bits(vma->node.start);
910 *cmd = MI_BATCH_BUFFER_END;
912 __i915_gem_object_flush_map(rpcs, 0, 64);
913 i915_gem_object_unpin_map(rpcs);
915 intel_gt_chipset_flush(vma->vm->gt);
921 emit_rpcs_query(struct drm_i915_gem_object *obj,
922 struct intel_context *ce,
923 struct i915_request **rq_out)
925 struct drm_i915_private *i915 = to_i915(obj->base.dev);
926 struct i915_request *rq;
927 struct i915_gem_ww_ctx ww;
928 struct i915_vma *batch;
929 struct i915_vma *vma;
930 struct drm_i915_gem_object *rpcs;
933 GEM_BUG_ON(!intel_engine_can_store_dword(ce->engine));
935 if (INTEL_GEN(i915) < 8)
938 vma = i915_vma_instance(obj, ce->vm, NULL);
942 rpcs = i915_gem_object_create_internal(i915, PAGE_SIZE);
944 return PTR_ERR(rpcs);
946 batch = i915_vma_instance(rpcs, ce->vm, NULL);
948 err = PTR_ERR(batch);
952 i915_gem_ww_ctx_init(&ww, false);
954 err = i915_gem_object_lock(obj, &ww);
956 err = i915_gem_object_lock(rpcs, &ww);
958 err = i915_gem_object_set_to_gtt_domain(obj, false);
960 err = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_USER);
964 err = i915_vma_pin_ww(batch, &ww, 0, 0, PIN_USER);
968 err = rpcs_query_batch(rpcs, vma);
972 rq = i915_request_create(ce);
978 err = i915_request_await_object(rq, batch->obj, false);
980 err = i915_vma_move_to_active(batch, rq, 0);
984 err = i915_request_await_object(rq, vma->obj, true);
986 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
990 if (rq->engine->emit_init_breadcrumb) {
991 err = rq->engine->emit_init_breadcrumb(rq);
996 err = rq->engine->emit_bb_start(rq,
997 batch->node.start, batch->node.size,
1002 *rq_out = i915_request_get(rq);
1006 i915_request_set_error_once(rq, err);
1007 i915_request_add(rq);
1009 i915_vma_unpin(batch);
1011 i915_vma_unpin(vma);
1013 if (err == -EDEADLK) {
1014 err = i915_gem_ww_ctx_backoff(&ww);
1018 i915_gem_ww_ctx_fini(&ww);
1019 i915_gem_object_put(rpcs);
1023 #define TEST_IDLE BIT(0)
1024 #define TEST_BUSY BIT(1)
1025 #define TEST_RESET BIT(2)
1028 __sseu_prepare(const char *name,
1030 struct intel_context *ce,
1031 struct igt_spinner **spin)
1033 struct i915_request *rq;
1037 if (!(flags & (TEST_BUSY | TEST_RESET)))
1040 *spin = kzalloc(sizeof(**spin), GFP_KERNEL);
1044 ret = igt_spinner_init(*spin, ce->engine->gt);
1048 rq = igt_spinner_create_request(*spin, ce, MI_NOOP);
1054 i915_request_add(rq);
1056 if (!igt_wait_for_spinner(*spin, rq)) {
1057 pr_err("%s: Spinner failed to start!\n", name);
1065 igt_spinner_end(*spin);
1067 igt_spinner_fini(*spin);
1069 kfree(fetch_and_zero(spin));
1074 __read_slice_count(struct intel_context *ce,
1075 struct drm_i915_gem_object *obj,
1076 struct igt_spinner *spin,
1079 struct i915_request *rq = NULL;
1080 u32 s_mask, s_shift;
1085 ret = emit_rpcs_query(obj, ce, &rq);
1090 igt_spinner_end(spin);
1092 ret = i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
1093 i915_request_put(rq);
1097 buf = i915_gem_object_pin_map(obj, I915_MAP_WB);
1103 if (INTEL_GEN(ce->engine->i915) >= 11) {
1104 s_mask = GEN11_RPCS_S_CNT_MASK;
1105 s_shift = GEN11_RPCS_S_CNT_SHIFT;
1107 s_mask = GEN8_RPCS_S_CNT_MASK;
1108 s_shift = GEN8_RPCS_S_CNT_SHIFT;
1112 cnt = (val & s_mask) >> s_shift;
1115 i915_gem_object_unpin_map(obj);
1121 __check_rpcs(const char *name, u32 rpcs, int slices, unsigned int expected,
1122 const char *prefix, const char *suffix)
1124 if (slices == expected)
1128 pr_err("%s: %s read slice count failed with %d%s\n",
1129 name, prefix, slices, suffix);
1133 pr_err("%s: %s slice count %d is not %u%s\n",
1134 name, prefix, slices, expected, suffix);
1136 pr_info("RPCS=0x%x; %u%sx%u%s\n",
1138 (rpcs & GEN8_RPCS_S_CNT_ENABLE) ? "*" : "",
1139 (rpcs & GEN8_RPCS_SS_CNT_MASK) >> GEN8_RPCS_SS_CNT_SHIFT,
1140 (rpcs & GEN8_RPCS_SS_CNT_ENABLE) ? "*" : "");
1146 __sseu_finish(const char *name,
1148 struct intel_context *ce,
1149 struct drm_i915_gem_object *obj,
1150 unsigned int expected,
1151 struct igt_spinner *spin)
1153 unsigned int slices = hweight32(ce->engine->sseu.slice_mask);
1157 if (flags & TEST_RESET) {
1158 ret = intel_engine_reset(ce->engine, "sseu");
1163 ret = __read_slice_count(ce, obj,
1164 flags & TEST_RESET ? NULL : spin, &rpcs);
1165 ret = __check_rpcs(name, rpcs, ret, expected, "Context", "!");
1169 ret = __read_slice_count(ce->engine->kernel_context, obj, NULL, &rpcs);
1170 ret = __check_rpcs(name, rpcs, ret, slices, "Kernel context", "!");
1174 igt_spinner_end(spin);
1176 if ((flags & TEST_IDLE) && ret == 0) {
1177 ret = igt_flush_test(ce->engine->i915);
1181 ret = __read_slice_count(ce, obj, NULL, &rpcs);
1182 ret = __check_rpcs(name, rpcs, ret, expected,
1183 "Context", " after idle!");
1190 __sseu_test(const char *name,
1192 struct intel_context *ce,
1193 struct drm_i915_gem_object *obj,
1194 struct intel_sseu sseu)
1196 struct igt_spinner *spin = NULL;
1199 intel_engine_pm_get(ce->engine);
1201 ret = __sseu_prepare(name, flags, ce, &spin);
1205 ret = intel_context_reconfigure_sseu(ce, sseu);
1209 ret = __sseu_finish(name, flags, ce, obj,
1210 hweight32(sseu.slice_mask), spin);
1214 igt_spinner_end(spin);
1215 igt_spinner_fini(spin);
1219 intel_engine_pm_put(ce->engine);
1224 __igt_ctx_sseu(struct drm_i915_private *i915,
1228 struct drm_i915_gem_object *obj;
1232 if (INTEL_GEN(i915) < 9)
1235 if (flags & TEST_RESET)
1236 igt_global_reset_lock(&i915->gt);
1238 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
1245 struct intel_engine_cs *engine;
1246 struct intel_context *ce;
1247 struct intel_sseu pg_sseu;
1249 engine = intel_engine_lookup_user(i915,
1250 I915_ENGINE_CLASS_RENDER,
1255 if (hweight32(engine->sseu.slice_mask) < 2)
1258 if (!engine->gt->info.sseu.has_slice_pg)
1262 * Gen11 VME friendly power-gated configuration with
1263 * half enabled sub-slices.
1265 pg_sseu = engine->sseu;
1266 pg_sseu.slice_mask = 1;
1267 pg_sseu.subslice_mask =
1268 ~(~0 << (hweight32(engine->sseu.subslice_mask) / 2));
1270 pr_info("%s: SSEU subtest '%s', flags=%x, def_slices=%u, pg_slices=%u\n",
1271 engine->name, name, flags,
1272 hweight32(engine->sseu.slice_mask),
1273 hweight32(pg_sseu.slice_mask));
1275 ce = intel_context_create(engine);
1281 ret = intel_context_pin(ce);
1285 /* First set the default mask. */
1286 ret = __sseu_test(name, flags, ce, obj, engine->sseu);
1290 /* Then set a power-gated configuration. */
1291 ret = __sseu_test(name, flags, ce, obj, pg_sseu);
1295 /* Back to defaults. */
1296 ret = __sseu_test(name, flags, ce, obj, engine->sseu);
1300 /* One last power-gated configuration for the road. */
1301 ret = __sseu_test(name, flags, ce, obj, pg_sseu);
1306 intel_context_unpin(ce);
1308 intel_context_put(ce);
1311 if (igt_flush_test(i915))
1315 i915_gem_object_put(obj);
1318 if (flags & TEST_RESET)
1319 igt_global_reset_unlock(&i915->gt);
1322 pr_err("%s: Failed with %d!\n", name, ret);
1327 static int igt_ctx_sseu(void *arg)
1332 } *phase, phases[] = {
1333 { .name = "basic", .flags = 0 },
1334 { .name = "idle", .flags = TEST_IDLE },
1335 { .name = "busy", .flags = TEST_BUSY },
1336 { .name = "busy-reset", .flags = TEST_BUSY | TEST_RESET },
1337 { .name = "busy-idle", .flags = TEST_BUSY | TEST_IDLE },
1338 { .name = "reset-idle", .flags = TEST_RESET | TEST_IDLE },
1343 for (i = 0, phase = phases; ret == 0 && i < ARRAY_SIZE(phases);
1345 ret = __igt_ctx_sseu(arg, phase->name, phase->flags);
1350 static int igt_ctx_readonly(void *arg)
1352 struct drm_i915_private *i915 = arg;
1353 unsigned long idx, ndwords, dw, num_engines;
1354 struct drm_i915_gem_object *obj = NULL;
1355 struct i915_request *tq[5] = {};
1356 struct i915_gem_engines_iter it;
1357 struct i915_address_space *vm;
1358 struct i915_gem_context *ctx;
1359 struct intel_context *ce;
1360 struct igt_live_test t;
1361 I915_RND_STATE(prng);
1362 IGT_TIMEOUT(end_time);
1368 * Create a few read-only objects (with the occasional writable object)
1369 * and try to write into these object checking that the GPU discards
1370 * any write to a read-only object.
1373 file = mock_file(i915);
1375 return PTR_ERR(file);
1377 err = igt_live_test_begin(&t, i915, __func__, "");
1381 ctx = live_context(i915, file);
1387 vm = ctx_vm(ctx) ?: &i915->ggtt.alias->vm;
1388 if (!vm || !vm->has_read_only) {
1394 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it)
1395 if (intel_engine_can_store_dword(ce->engine))
1397 i915_gem_context_unlock_engines(ctx);
1401 while (!time_after(jiffies, end_time)) {
1402 for_each_gem_engine(ce,
1403 i915_gem_context_lock_engines(ctx), it) {
1404 if (!intel_engine_can_store_dword(ce->engine))
1408 obj = create_test_object(ce->vm, file, &objects);
1411 i915_gem_context_unlock_engines(ctx);
1415 if (prandom_u32_state(&prng) & 1)
1416 i915_gem_object_set_readonly(obj);
1419 err = gpu_fill(ce, obj, dw);
1421 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) [full-ppgtt? %s], err=%d\n",
1422 ndwords, dw, max_dwords(obj),
1424 yesno(!!ctx_vm(ctx)),
1426 i915_gem_context_unlock_engines(ctx);
1430 err = throttle(ce, tq, ARRAY_SIZE(tq));
1432 i915_gem_context_unlock_engines(ctx);
1436 if (++dw == max_dwords(obj)) {
1442 i915_gem_context_unlock_engines(ctx);
1444 pr_info("Submitted %lu dwords (across %lu engines)\n",
1445 ndwords, num_engines);
1449 list_for_each_entry(obj, &objects, st_link) {
1451 min_t(unsigned int, ndwords - dw, max_dwords(obj));
1452 unsigned int num_writes;
1455 if (i915_gem_object_is_readonly(obj))
1458 err = cpu_check(obj, idx++, num_writes);
1466 throttle_release(tq, ARRAY_SIZE(tq));
1467 if (igt_live_test_end(&t))
1474 static int check_scratch(struct i915_address_space *vm, u64 offset)
1476 struct drm_mm_node *node;
1478 mutex_lock(&vm->mutex);
1479 node = __drm_mm_interval_first(&vm->mm,
1480 offset, offset + sizeof(u32) - 1);
1481 mutex_unlock(&vm->mutex);
1482 if (!node || node->start > offset)
1485 GEM_BUG_ON(offset >= node->start + node->size);
1487 pr_err("Target offset 0x%08x_%08x overlaps with a node in the mm!\n",
1488 upper_32_bits(offset), lower_32_bits(offset));
1492 static int write_to_scratch(struct i915_gem_context *ctx,
1493 struct intel_engine_cs *engine,
1494 u64 offset, u32 value)
1496 struct drm_i915_private *i915 = ctx->i915;
1497 struct drm_i915_gem_object *obj;
1498 struct i915_address_space *vm;
1499 struct i915_request *rq;
1500 struct i915_vma *vma;
1504 GEM_BUG_ON(offset < I915_GTT_PAGE_SIZE);
1506 err = check_scratch(ctx_vm(ctx), offset);
1510 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
1512 return PTR_ERR(obj);
1514 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
1520 *cmd++ = MI_STORE_DWORD_IMM_GEN4;
1521 if (INTEL_GEN(i915) >= 8) {
1522 *cmd++ = lower_32_bits(offset);
1523 *cmd++ = upper_32_bits(offset);
1529 *cmd = MI_BATCH_BUFFER_END;
1530 __i915_gem_object_flush_map(obj, 0, 64);
1531 i915_gem_object_unpin_map(obj);
1533 intel_gt_chipset_flush(engine->gt);
1535 vm = i915_gem_context_get_vm_rcu(ctx);
1536 vma = i915_vma_instance(obj, vm, NULL);
1542 err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_OFFSET_FIXED);
1546 rq = igt_request_alloc(ctx, engine);
1553 err = i915_request_await_object(rq, vma->obj, false);
1555 err = i915_vma_move_to_active(vma, rq, 0);
1556 i915_vma_unlock(vma);
1560 if (rq->engine->emit_init_breadcrumb) {
1561 err = rq->engine->emit_init_breadcrumb(rq);
1566 err = engine->emit_bb_start(rq, vma->node.start, vma->node.size, 0);
1570 i915_vma_unpin(vma);
1572 i915_request_add(rq);
1576 i915_request_set_error_once(rq, err);
1577 i915_request_add(rq);
1579 i915_vma_unpin(vma);
1583 i915_gem_object_put(obj);
1587 static int read_from_scratch(struct i915_gem_context *ctx,
1588 struct intel_engine_cs *engine,
1589 u64 offset, u32 *value)
1591 struct drm_i915_private *i915 = ctx->i915;
1592 struct drm_i915_gem_object *obj;
1593 struct i915_address_space *vm;
1594 const u32 result = 0x100;
1595 struct i915_request *rq;
1596 struct i915_vma *vma;
1601 GEM_BUG_ON(offset < I915_GTT_PAGE_SIZE);
1603 err = check_scratch(ctx_vm(ctx), offset);
1607 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
1609 return PTR_ERR(obj);
1611 if (INTEL_GEN(i915) >= 8) {
1612 const u32 GPR0 = engine->mmio_base + 0x600;
1614 vm = i915_gem_context_get_vm_rcu(ctx);
1615 vma = i915_vma_instance(obj, vm, NULL);
1621 err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_OFFSET_FIXED);
1625 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
1631 memset(cmd, POISON_INUSE, PAGE_SIZE);
1632 *cmd++ = MI_LOAD_REGISTER_MEM_GEN8;
1634 *cmd++ = lower_32_bits(offset);
1635 *cmd++ = upper_32_bits(offset);
1636 *cmd++ = MI_STORE_REGISTER_MEM_GEN8;
1640 *cmd = MI_BATCH_BUFFER_END;
1642 i915_gem_object_flush_map(obj);
1643 i915_gem_object_unpin_map(obj);
1647 const u32 reg = engine->mmio_base + 0x420;
1649 /* hsw: register access even to 3DPRIM! is protected */
1650 vm = i915_vm_get(&engine->gt->ggtt->vm);
1651 vma = i915_vma_instance(obj, vm, NULL);
1657 err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1661 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
1667 memset(cmd, POISON_INUSE, PAGE_SIZE);
1668 *cmd++ = MI_LOAD_REGISTER_MEM;
1671 *cmd++ = MI_STORE_REGISTER_MEM | MI_USE_GGTT;
1673 *cmd++ = vma->node.start + result;
1674 *cmd = MI_BATCH_BUFFER_END;
1676 i915_gem_object_flush_map(obj);
1677 i915_gem_object_unpin_map(obj);
1679 flags = I915_DISPATCH_SECURE;
1682 intel_gt_chipset_flush(engine->gt);
1684 rq = igt_request_alloc(ctx, engine);
1691 err = i915_request_await_object(rq, vma->obj, true);
1693 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
1694 i915_vma_unlock(vma);
1698 if (rq->engine->emit_init_breadcrumb) {
1699 err = rq->engine->emit_init_breadcrumb(rq);
1704 err = engine->emit_bb_start(rq, vma->node.start, vma->node.size, flags);
1708 i915_vma_unpin(vma);
1710 i915_request_add(rq);
1712 i915_gem_object_lock(obj, NULL);
1713 err = i915_gem_object_set_to_cpu_domain(obj, false);
1714 i915_gem_object_unlock(obj);
1718 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
1724 *value = cmd[result / sizeof(*cmd)];
1725 i915_gem_object_unpin_map(obj);
1729 i915_request_set_error_once(rq, err);
1730 i915_request_add(rq);
1732 i915_vma_unpin(vma);
1736 i915_gem_object_put(obj);
1740 static int check_scratch_page(struct i915_gem_context *ctx, u32 *out)
1742 struct i915_address_space *vm;
1751 page = __px_page(vm->scratch[0]);
1753 pr_err("No scratch page!\n");
1759 pr_err("No (mappable) scratch page!\n");
1763 memcpy(out, vaddr, sizeof(*out));
1764 if (memchr_inv(vaddr, *out, PAGE_SIZE)) {
1765 pr_err("Inconsistent initial state of scratch page!\n");
1773 static int igt_vm_isolation(void *arg)
1775 struct drm_i915_private *i915 = arg;
1776 struct i915_gem_context *ctx_a, *ctx_b;
1777 unsigned long num_engines, count;
1778 struct intel_engine_cs *engine;
1779 struct igt_live_test t;
1780 I915_RND_STATE(prng);
1786 if (INTEL_GEN(i915) < 7)
1790 * The simple goal here is that a write into one context is not
1791 * observed in a second (separate page tables and scratch).
1794 file = mock_file(i915);
1796 return PTR_ERR(file);
1798 err = igt_live_test_begin(&t, i915, __func__, "");
1802 ctx_a = live_context(i915, file);
1803 if (IS_ERR(ctx_a)) {
1804 err = PTR_ERR(ctx_a);
1808 ctx_b = live_context(i915, file);
1809 if (IS_ERR(ctx_b)) {
1810 err = PTR_ERR(ctx_b);
1814 /* We can only test vm isolation, if the vm are distinct */
1815 if (ctx_vm(ctx_a) == ctx_vm(ctx_b))
1818 /* Read the initial state of the scratch page */
1819 err = check_scratch_page(ctx_a, &expected);
1823 err = check_scratch_page(ctx_b, &expected);
1827 vm_total = ctx_vm(ctx_a)->total;
1828 GEM_BUG_ON(ctx_vm(ctx_b)->total != vm_total);
1832 for_each_uabi_engine(engine, i915) {
1833 IGT_TIMEOUT(end_time);
1834 unsigned long this = 0;
1836 if (!intel_engine_can_store_dword(engine))
1839 /* Not all engines have their own GPR! */
1840 if (INTEL_GEN(i915) < 8 && engine->class != RENDER_CLASS)
1843 while (!__igt_timeout(end_time, NULL)) {
1844 u32 value = 0xc5c5c5c5;
1847 /* Leave enough space at offset 0 for the batch */
1848 offset = igt_random_offset(&prng,
1849 I915_GTT_PAGE_SIZE, vm_total,
1850 sizeof(u32), alignof_dword);
1852 err = write_to_scratch(ctx_a, engine,
1853 offset, 0xdeadbeef);
1855 err = read_from_scratch(ctx_b, engine,
1860 if (value != expected) {
1861 pr_err("%s: Read %08x from scratch (offset 0x%08x_%08x), after %lu reads!\n",
1862 engine->name, value,
1863 upper_32_bits(offset),
1864 lower_32_bits(offset),
1875 pr_info("Checked %lu scratch offsets across %lu engines\n",
1876 count, num_engines);
1879 if (igt_live_test_end(&t))
1885 static bool skip_unused_engines(struct intel_context *ce, void *data)
1890 static void mock_barrier_task(void *data)
1892 unsigned int *counter = data;
1897 static int mock_context_barrier(void *arg)
1900 #define pr_fmt(x) "context_barrier_task():" # x
1901 struct drm_i915_private *i915 = arg;
1902 struct i915_gem_context *ctx;
1903 struct i915_request *rq;
1904 unsigned int counter;
1908 * The context barrier provides us with a callback after it emits
1909 * a request; useful for retiring old state after loading new.
1912 ctx = mock_context(i915, "mock");
1917 err = context_barrier_task(ctx, 0, NULL, NULL, NULL,
1918 mock_barrier_task, &counter);
1920 pr_err("Failed at line %d, err=%d\n", __LINE__, err);
1924 pr_err("Did not retire immediately with 0 engines\n");
1930 err = context_barrier_task(ctx, ALL_ENGINES, skip_unused_engines,
1931 NULL, NULL, mock_barrier_task, &counter);
1933 pr_err("Failed at line %d, err=%d\n", __LINE__, err);
1937 pr_err("Did not retire immediately for all unused engines\n");
1942 rq = igt_request_alloc(ctx, i915->gt.engine[RCS0]);
1944 pr_err("Request allocation failed!\n");
1947 i915_request_add(rq);
1950 context_barrier_inject_fault = BIT(RCS0);
1951 err = context_barrier_task(ctx, ALL_ENGINES, NULL, NULL, NULL,
1952 mock_barrier_task, &counter);
1953 context_barrier_inject_fault = 0;
1957 pr_err("Did not hit fault injection!\n");
1959 pr_err("Invoked callback on error!\n");
1966 err = context_barrier_task(ctx, ALL_ENGINES, skip_unused_engines,
1967 NULL, NULL, mock_barrier_task, &counter);
1969 pr_err("Failed at line %d, err=%d\n", __LINE__, err);
1972 mock_device_flush(i915);
1974 pr_err("Did not retire on each active engines\n");
1980 mock_context_close(ctx);
1986 int i915_gem_context_mock_selftests(void)
1988 static const struct i915_subtest tests[] = {
1989 SUBTEST(mock_context_barrier),
1991 struct drm_i915_private *i915;
1994 i915 = mock_gem_device();
1998 err = i915_subtests(tests, i915);
2000 mock_destroy_device(i915);
2004 int i915_gem_context_live_selftests(struct drm_i915_private *i915)
2006 static const struct i915_subtest tests[] = {
2007 SUBTEST(live_nop_switch),
2008 SUBTEST(live_parallel_switch),
2009 SUBTEST(igt_ctx_exec),
2010 SUBTEST(igt_ctx_readonly),
2011 SUBTEST(igt_ctx_sseu),
2012 SUBTEST(igt_shared_ctx_exec),
2013 SUBTEST(igt_vm_isolation),
2016 if (intel_gt_is_wedged(&i915->gt))
2019 return i915_live_subtests(tests, i915);