1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2016 Intel Corporation
4 */
5
6 #include <linux/kthread.h>
7
8 #include "gem/i915_gem_context.h"
9 #include "gem/i915_gem_internal.h"
10
11 #include "i915_gem_evict.h"
12 #include "intel_gt.h"
13 #include "intel_engine_heartbeat.h"
14 #include "intel_engine_pm.h"
15 #include "selftest_engine_heartbeat.h"
16
17 #include "i915_selftest.h"
18 #include "selftests/i915_random.h"
19 #include "selftests/igt_flush_test.h"
20 #include "selftests/igt_reset.h"
21 #include "selftests/igt_atomic.h"
22 #include "selftests/igt_spinner.h"
23 #include "selftests/intel_scheduler_helpers.h"
24
25 #include "selftests/mock_drm.h"
26
27 #include "gem/selftests/mock_context.h"
28 #include "gem/selftests/igt_gem_utils.h"
29
30 #define IGT_IDLE_TIMEOUT 50 /* ms; time to wait after flushing between tests */
31
32 struct hang {
33 struct intel_gt *gt;
34 struct drm_i915_gem_object *hws;
35 struct drm_i915_gem_object *obj;
36 struct i915_gem_context *ctx;
37 u32 *seqno;
38 u32 *batch;
39 };
40
hang_init(struct hang * h,struct intel_gt * gt)41 static int hang_init(struct hang *h, struct intel_gt *gt)
42 {
43 void *vaddr;
44 int err;
45
46 memset(h, 0, sizeof(*h));
47 h->gt = gt;
48
49 h->ctx = kernel_context(gt->i915, NULL);
50 if (IS_ERR(h->ctx))
51 return PTR_ERR(h->ctx);
52
53 GEM_BUG_ON(i915_gem_context_is_bannable(h->ctx));
54
55 h->hws = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
56 if (IS_ERR(h->hws)) {
57 err = PTR_ERR(h->hws);
58 goto err_ctx;
59 }
60
61 h->obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
62 if (IS_ERR(h->obj)) {
63 err = PTR_ERR(h->obj);
64 goto err_hws;
65 }
66
67 i915_gem_object_set_cache_coherency(h->hws, I915_CACHE_LLC);
68 vaddr = i915_gem_object_pin_map_unlocked(h->hws, I915_MAP_WB);
69 if (IS_ERR(vaddr)) {
70 err = PTR_ERR(vaddr);
71 goto err_obj;
72 }
73 h->seqno = memset(vaddr, 0xff, PAGE_SIZE);
74
75 vaddr = i915_gem_object_pin_map_unlocked(h->obj,
76 intel_gt_coherent_map_type(gt, h->obj, false));
77 if (IS_ERR(vaddr)) {
78 err = PTR_ERR(vaddr);
79 goto err_unpin_hws;
80 }
81 h->batch = vaddr;
82
83 return 0;
84
85 err_unpin_hws:
86 i915_gem_object_unpin_map(h->hws);
87 err_obj:
88 i915_gem_object_put(h->obj);
89 err_hws:
90 i915_gem_object_put(h->hws);
91 err_ctx:
92 kernel_context_close(h->ctx);
93 return err;
94 }
95
hws_address(const struct i915_vma * hws,const struct i915_request * rq)96 static u64 hws_address(const struct i915_vma *hws,
97 const struct i915_request *rq)
98 {
99 return i915_vma_offset(hws) +
100 offset_in_page(sizeof(u32) * rq->fence.context);
101 }
102
103 static struct i915_request *
hang_create_request(struct hang * h,struct intel_engine_cs * engine)104 hang_create_request(struct hang *h, struct intel_engine_cs *engine)
105 {
106 struct intel_gt *gt = h->gt;
107 struct i915_address_space *vm = i915_gem_context_get_eb_vm(h->ctx);
108 struct drm_i915_gem_object *obj;
109 struct i915_request *rq = NULL;
110 struct i915_vma *hws, *vma;
111 unsigned int flags;
112 void *vaddr;
113 u32 *batch;
114 int err;
115
116 obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
117 if (IS_ERR(obj)) {
118 i915_vm_put(vm);
119 return ERR_CAST(obj);
120 }
121
122 vaddr = i915_gem_object_pin_map_unlocked(obj, intel_gt_coherent_map_type(gt, obj, false));
123 if (IS_ERR(vaddr)) {
124 i915_gem_object_put(obj);
125 i915_vm_put(vm);
126 return ERR_CAST(vaddr);
127 }
128
129 i915_gem_object_unpin_map(h->obj);
130 i915_gem_object_put(h->obj);
131
132 h->obj = obj;
133 h->batch = vaddr;
134
135 vma = i915_vma_instance(h->obj, vm, NULL);
136 if (IS_ERR(vma)) {
137 i915_vm_put(vm);
138 return ERR_CAST(vma);
139 }
140
141 hws = i915_vma_instance(h->hws, vm, NULL);
142 if (IS_ERR(hws)) {
143 i915_vm_put(vm);
144 return ERR_CAST(hws);
145 }
146
147 err = i915_vma_pin(vma, 0, 0, PIN_USER);
148 if (err) {
149 i915_vm_put(vm);
150 return ERR_PTR(err);
151 }
152
153 err = i915_vma_pin(hws, 0, 0, PIN_USER);
154 if (err)
155 goto unpin_vma;
156
157 rq = igt_request_alloc(h->ctx, engine);
158 if (IS_ERR(rq)) {
159 err = PTR_ERR(rq);
160 goto unpin_hws;
161 }
162
163 err = igt_vma_move_to_active_unlocked(vma, rq, 0);
164 if (err)
165 goto cancel_rq;
166
167 err = igt_vma_move_to_active_unlocked(hws, rq, 0);
168 if (err)
169 goto cancel_rq;
170
171 batch = h->batch;
172 if (GRAPHICS_VER(gt->i915) >= 8) {
173 *batch++ = MI_STORE_DWORD_IMM_GEN4;
174 *batch++ = lower_32_bits(hws_address(hws, rq));
175 *batch++ = upper_32_bits(hws_address(hws, rq));
176 *batch++ = rq->fence.seqno;
177 *batch++ = MI_NOOP;
178
179 memset(batch, 0, 1024);
180 batch += 1024 / sizeof(*batch);
181
182 *batch++ = MI_NOOP;
183 *batch++ = MI_BATCH_BUFFER_START | 1 << 8 | 1;
184 *batch++ = lower_32_bits(i915_vma_offset(vma));
185 *batch++ = upper_32_bits(i915_vma_offset(vma));
186 } else if (GRAPHICS_VER(gt->i915) >= 6) {
187 *batch++ = MI_STORE_DWORD_IMM_GEN4;
188 *batch++ = 0;
189 *batch++ = lower_32_bits(hws_address(hws, rq));
190 *batch++ = rq->fence.seqno;
191 *batch++ = MI_NOOP;
192
193 memset(batch, 0, 1024);
194 batch += 1024 / sizeof(*batch);
195
196 *batch++ = MI_NOOP;
197 *batch++ = MI_BATCH_BUFFER_START | 1 << 8;
198 *batch++ = lower_32_bits(i915_vma_offset(vma));
199 } else if (GRAPHICS_VER(gt->i915) >= 4) {
200 *batch++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
201 *batch++ = 0;
202 *batch++ = lower_32_bits(hws_address(hws, rq));
203 *batch++ = rq->fence.seqno;
204 *batch++ = MI_NOOP;
205
206 memset(batch, 0, 1024);
207 batch += 1024 / sizeof(*batch);
208
209 *batch++ = MI_NOOP;
210 *batch++ = MI_BATCH_BUFFER_START | 2 << 6;
211 *batch++ = lower_32_bits(i915_vma_offset(vma));
212 } else {
213 *batch++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
214 *batch++ = lower_32_bits(hws_address(hws, rq));
215 *batch++ = rq->fence.seqno;
216 *batch++ = MI_NOOP;
217
218 memset(batch, 0, 1024);
219 batch += 1024 / sizeof(*batch);
220
221 *batch++ = MI_NOOP;
222 *batch++ = MI_BATCH_BUFFER_START | 2 << 6;
223 *batch++ = lower_32_bits(i915_vma_offset(vma));
224 }
225 *batch++ = MI_BATCH_BUFFER_END; /* not reached */
226 intel_gt_chipset_flush(engine->gt);
227
228 if (rq->engine->emit_init_breadcrumb) {
229 err = rq->engine->emit_init_breadcrumb(rq);
230 if (err)
231 goto cancel_rq;
232 }
233
234 flags = 0;
235 if (GRAPHICS_VER(gt->i915) <= 5)
236 flags |= I915_DISPATCH_SECURE;
237
238 err = rq->engine->emit_bb_start(rq, i915_vma_offset(vma), PAGE_SIZE, flags);
239
240 cancel_rq:
241 if (err) {
242 i915_request_set_error_once(rq, err);
243 i915_request_add(rq);
244 }
245 unpin_hws:
246 i915_vma_unpin(hws);
247 unpin_vma:
248 i915_vma_unpin(vma);
249 i915_vm_put(vm);
250 return err ? ERR_PTR(err) : rq;
251 }
252
hws_seqno(const struct hang * h,const struct i915_request * rq)253 static u32 hws_seqno(const struct hang *h, const struct i915_request *rq)
254 {
255 return READ_ONCE(h->seqno[rq->fence.context % (PAGE_SIZE/sizeof(u32))]);
256 }
257
hang_fini(struct hang * h)258 static void hang_fini(struct hang *h)
259 {
260 *h->batch = MI_BATCH_BUFFER_END;
261 intel_gt_chipset_flush(h->gt);
262
263 i915_gem_object_unpin_map(h->obj);
264 i915_gem_object_put(h->obj);
265
266 i915_gem_object_unpin_map(h->hws);
267 i915_gem_object_put(h->hws);
268
269 kernel_context_close(h->ctx);
270
271 igt_flush_test(h->gt->i915);
272 }
273
wait_until_running(struct hang * h,struct i915_request * rq)274 static bool wait_until_running(struct hang *h, struct i915_request *rq)
275 {
276 return !(wait_for_us(i915_seqno_passed(hws_seqno(h, rq),
277 rq->fence.seqno),
278 10) &&
279 wait_for(i915_seqno_passed(hws_seqno(h, rq),
280 rq->fence.seqno),
281 1000));
282 }
283
igt_hang_sanitycheck(void * arg)284 static int igt_hang_sanitycheck(void *arg)
285 {
286 struct intel_gt *gt = arg;
287 struct i915_request *rq;
288 struct intel_engine_cs *engine;
289 enum intel_engine_id id;
290 struct hang h;
291 int err;
292
293 /* Basic check that we can execute our hanging batch */
294
295 err = hang_init(&h, gt);
296 if (err)
297 return err;
298
299 for_each_engine(engine, gt, id) {
300 struct intel_wedge_me w;
301 long timeout;
302
303 if (!intel_engine_can_store_dword(engine))
304 continue;
305
306 rq = hang_create_request(&h, engine);
307 if (IS_ERR(rq)) {
308 err = PTR_ERR(rq);
309 pr_err("Failed to create request for %s, err=%d\n",
310 engine->name, err);
311 goto fini;
312 }
313
314 i915_request_get(rq);
315
316 *h.batch = MI_BATCH_BUFFER_END;
317 intel_gt_chipset_flush(engine->gt);
318
319 i915_request_add(rq);
320
321 timeout = 0;
322 intel_wedge_on_timeout(&w, gt, HZ / 5 /* 200ms */)
323 timeout = i915_request_wait(rq, 0,
324 MAX_SCHEDULE_TIMEOUT);
325 if (intel_gt_is_wedged(gt))
326 timeout = -EIO;
327
328 i915_request_put(rq);
329
330 if (timeout < 0) {
331 err = timeout;
332 pr_err("Wait for request failed on %s, err=%d\n",
333 engine->name, err);
334 goto fini;
335 }
336 }
337
338 fini:
339 hang_fini(&h);
340 return err;
341 }
342
wait_for_idle(struct intel_engine_cs * engine)343 static bool wait_for_idle(struct intel_engine_cs *engine)
344 {
345 return wait_for(intel_engine_is_idle(engine), IGT_IDLE_TIMEOUT) == 0;
346 }
347
igt_reset_nop(void * arg)348 static int igt_reset_nop(void *arg)
349 {
350 struct intel_gt *gt = arg;
351 struct i915_gpu_error *global = >->i915->gpu_error;
352 struct intel_engine_cs *engine;
353 unsigned int reset_count, count;
354 enum intel_engine_id id;
355 IGT_TIMEOUT(end_time);
356 int err = 0;
357
358 /* Check that we can reset during non-user portions of requests */
359
360 reset_count = i915_reset_count(global);
361 count = 0;
362 do {
363 for_each_engine(engine, gt, id) {
364 struct intel_context *ce;
365 int i;
366
367 ce = intel_context_create(engine);
368 if (IS_ERR(ce)) {
369 err = PTR_ERR(ce);
370 pr_err("[%s] Create context failed: %d!\n", engine->name, err);
371 break;
372 }
373
374 for (i = 0; i < 16; i++) {
375 struct i915_request *rq;
376
377 rq = intel_context_create_request(ce);
378 if (IS_ERR(rq)) {
379 err = PTR_ERR(rq);
380 pr_err("[%s] Create request failed: %d!\n",
381 engine->name, err);
382 break;
383 }
384
385 i915_request_add(rq);
386 }
387
388 intel_context_put(ce);
389 }
390
391 igt_global_reset_lock(gt);
392 intel_gt_reset(gt, ALL_ENGINES, NULL);
393 igt_global_reset_unlock(gt);
394
395 if (intel_gt_is_wedged(gt)) {
396 pr_err("[%s] GT is wedged!\n", engine->name);
397 err = -EIO;
398 break;
399 }
400
401 if (i915_reset_count(global) != reset_count + ++count) {
402 pr_err("[%s] Reset not recorded: %d vs %d + %d!\n",
403 engine->name, i915_reset_count(global), reset_count, count);
404 err = -EINVAL;
405 break;
406 }
407
408 err = igt_flush_test(gt->i915);
409 if (err) {
410 pr_err("[%s] Flush failed: %d!\n", engine->name, err);
411 break;
412 }
413 } while (time_before(jiffies, end_time));
414 pr_info("%s: %d resets\n", __func__, count);
415
416 if (igt_flush_test(gt->i915)) {
417 pr_err("Post flush failed: %d!\n", err);
418 err = -EIO;
419 }
420
421 return err;
422 }
423
igt_reset_nop_engine(void * arg)424 static int igt_reset_nop_engine(void *arg)
425 {
426 struct intel_gt *gt = arg;
427 struct i915_gpu_error *global = >->i915->gpu_error;
428 struct intel_engine_cs *engine;
429 enum intel_engine_id id;
430
431 /* Check that we can engine-reset during non-user portions */
432
433 if (!intel_has_reset_engine(gt))
434 return 0;
435
436 for_each_engine(engine, gt, id) {
437 unsigned int reset_count, reset_engine_count, count;
438 struct intel_context *ce;
439 IGT_TIMEOUT(end_time);
440 int err;
441
442 if (intel_engine_uses_guc(engine)) {
443 /* Engine level resets are triggered by GuC when a hang
444 * is detected. They can't be triggered by the KMD any
445 * more. Thus a nop batch cannot be used as a reset test
446 */
447 continue;
448 }
449
450 ce = intel_context_create(engine);
451 if (IS_ERR(ce)) {
452 pr_err("[%s] Create context failed: %pe!\n", engine->name, ce);
453 return PTR_ERR(ce);
454 }
455
456 reset_count = i915_reset_count(global);
457 reset_engine_count = i915_reset_engine_count(global, engine);
458 count = 0;
459
460 st_engine_heartbeat_disable(engine);
461 GEM_BUG_ON(test_and_set_bit(I915_RESET_ENGINE + id,
462 >->reset.flags));
463 do {
464 int i;
465
466 if (!wait_for_idle(engine)) {
467 pr_err("%s failed to idle before reset\n",
468 engine->name);
469 err = -EIO;
470 break;
471 }
472
473 for (i = 0; i < 16; i++) {
474 struct i915_request *rq;
475
476 rq = intel_context_create_request(ce);
477 if (IS_ERR(rq)) {
478 struct drm_printer p =
479 drm_info_printer(gt->i915->drm.dev);
480 intel_engine_dump(engine, &p,
481 "%s(%s): failed to submit request\n",
482 __func__,
483 engine->name);
484
485 GEM_TRACE("%s(%s): failed to submit request\n",
486 __func__,
487 engine->name);
488 GEM_TRACE_DUMP();
489
490 intel_gt_set_wedged(gt);
491
492 err = PTR_ERR(rq);
493 break;
494 }
495
496 i915_request_add(rq);
497 }
498 err = intel_engine_reset(engine, NULL);
499 if (err) {
500 pr_err("intel_engine_reset(%s) failed, err:%d\n",
501 engine->name, err);
502 break;
503 }
504
505 if (i915_reset_count(global) != reset_count) {
506 pr_err("Full GPU reset recorded! (engine reset expected)\n");
507 err = -EINVAL;
508 break;
509 }
510
511 if (i915_reset_engine_count(global, engine) !=
512 reset_engine_count + ++count) {
513 pr_err("%s engine reset not recorded!\n",
514 engine->name);
515 err = -EINVAL;
516 break;
517 }
518 } while (time_before(jiffies, end_time));
519 clear_and_wake_up_bit(I915_RESET_ENGINE + id, >->reset.flags);
520 st_engine_heartbeat_enable(engine);
521
522 pr_info("%s(%s): %d resets\n", __func__, engine->name, count);
523
524 intel_context_put(ce);
525 if (igt_flush_test(gt->i915))
526 err = -EIO;
527 if (err)
528 return err;
529 }
530
531 return 0;
532 }
533
force_reset_timeout(struct intel_engine_cs * engine)534 static void force_reset_timeout(struct intel_engine_cs *engine)
535 {
536 engine->reset_timeout.probability = 999;
537 atomic_set(&engine->reset_timeout.times, -1);
538 }
539
cancel_reset_timeout(struct intel_engine_cs * engine)540 static void cancel_reset_timeout(struct intel_engine_cs *engine)
541 {
542 memset(&engine->reset_timeout, 0, sizeof(engine->reset_timeout));
543 }
544
igt_reset_fail_engine(void * arg)545 static int igt_reset_fail_engine(void *arg)
546 {
547 struct intel_gt *gt = arg;
548 struct intel_engine_cs *engine;
549 enum intel_engine_id id;
550
551 /* Check that we can recover from engine-reset failues */
552
553 if (!intel_has_reset_engine(gt))
554 return 0;
555
556 for_each_engine(engine, gt, id) {
557 unsigned int count;
558 struct intel_context *ce;
559 IGT_TIMEOUT(end_time);
560 int err;
561
562 /* Can't manually break the reset if i915 doesn't perform it */
563 if (intel_engine_uses_guc(engine))
564 continue;
565
566 ce = intel_context_create(engine);
567 if (IS_ERR(ce)) {
568 pr_err("[%s] Create context failed: %pe!\n", engine->name, ce);
569 return PTR_ERR(ce);
570 }
571
572 st_engine_heartbeat_disable(engine);
573 GEM_BUG_ON(test_and_set_bit(I915_RESET_ENGINE + id,
574 >->reset.flags));
575
576 force_reset_timeout(engine);
577 err = intel_engine_reset(engine, NULL);
578 cancel_reset_timeout(engine);
579 if (err == 0) /* timeouts only generated on gen8+ */
580 goto skip;
581
582 count = 0;
583 do {
584 struct i915_request *last = NULL;
585 int i;
586
587 if (!wait_for_idle(engine)) {
588 pr_err("%s failed to idle before reset\n",
589 engine->name);
590 err = -EIO;
591 break;
592 }
593
594 for (i = 0; i < count % 15; i++) {
595 struct i915_request *rq;
596
597 rq = intel_context_create_request(ce);
598 if (IS_ERR(rq)) {
599 struct drm_printer p =
600 drm_info_printer(gt->i915->drm.dev);
601 intel_engine_dump(engine, &p,
602 "%s(%s): failed to submit request\n",
603 __func__,
604 engine->name);
605
606 GEM_TRACE("%s(%s): failed to submit request\n",
607 __func__,
608 engine->name);
609 GEM_TRACE_DUMP();
610
611 intel_gt_set_wedged(gt);
612 if (last)
613 i915_request_put(last);
614
615 err = PTR_ERR(rq);
616 goto out;
617 }
618
619 if (last)
620 i915_request_put(last);
621 last = i915_request_get(rq);
622 i915_request_add(rq);
623 }
624
625 if (count & 1) {
626 err = intel_engine_reset(engine, NULL);
627 if (err) {
628 GEM_TRACE_ERR("intel_engine_reset(%s) failed, err:%d\n",
629 engine->name, err);
630 GEM_TRACE_DUMP();
631 i915_request_put(last);
632 break;
633 }
634 } else {
635 force_reset_timeout(engine);
636 err = intel_engine_reset(engine, NULL);
637 cancel_reset_timeout(engine);
638 if (err != -ETIMEDOUT) {
639 pr_err("intel_engine_reset(%s) did not fail, err:%d\n",
640 engine->name, err);
641 i915_request_put(last);
642 break;
643 }
644 }
645
646 err = 0;
647 if (last) {
648 if (i915_request_wait(last, 0, HZ / 2) < 0) {
649 struct drm_printer p =
650 drm_info_printer(gt->i915->drm.dev);
651
652 intel_engine_dump(engine, &p,
653 "%s(%s): failed to complete request\n",
654 __func__,
655 engine->name);
656
657 GEM_TRACE("%s(%s): failed to complete request\n",
658 __func__,
659 engine->name);
660 GEM_TRACE_DUMP();
661
662 err = -EIO;
663 }
664 i915_request_put(last);
665 }
666 count++;
667 } while (err == 0 && time_before(jiffies, end_time));
668 out:
669 pr_info("%s(%s): %d resets\n", __func__, engine->name, count);
670 skip:
671 clear_and_wake_up_bit(I915_RESET_ENGINE + id, >->reset.flags);
672 st_engine_heartbeat_enable(engine);
673 intel_context_put(ce);
674
675 if (igt_flush_test(gt->i915))
676 err = -EIO;
677 if (err)
678 return err;
679 }
680
681 return 0;
682 }
683
__igt_reset_engine(struct intel_gt * gt,bool active)684 static int __igt_reset_engine(struct intel_gt *gt, bool active)
685 {
686 struct i915_gpu_error *global = >->i915->gpu_error;
687 struct intel_engine_cs *engine;
688 enum intel_engine_id id;
689 struct hang h;
690 int err = 0;
691
692 /* Check that we can issue an engine reset on an idle engine (no-op) */
693
694 if (!intel_has_reset_engine(gt))
695 return 0;
696
697 if (active) {
698 err = hang_init(&h, gt);
699 if (err)
700 return err;
701 }
702
703 for_each_engine(engine, gt, id) {
704 unsigned int reset_count, reset_engine_count;
705 unsigned long count;
706 bool using_guc = intel_engine_uses_guc(engine);
707 IGT_TIMEOUT(end_time);
708
709 if (using_guc && !active)
710 continue;
711
712 if (active && !intel_engine_can_store_dword(engine))
713 continue;
714
715 if (!wait_for_idle(engine)) {
716 pr_err("%s failed to idle before reset\n",
717 engine->name);
718 err = -EIO;
719 break;
720 }
721
722 reset_count = i915_reset_count(global);
723 reset_engine_count = i915_reset_engine_count(global, engine);
724
725 st_engine_heartbeat_disable(engine);
726 GEM_BUG_ON(test_and_set_bit(I915_RESET_ENGINE + id,
727 >->reset.flags));
728 count = 0;
729 do {
730 struct i915_request *rq = NULL;
731 struct intel_selftest_saved_policy saved;
732 int err2;
733
734 err = intel_selftest_modify_policy(engine, &saved,
735 SELFTEST_SCHEDULER_MODIFY_FAST_RESET);
736 if (err) {
737 pr_err("[%s] Modify policy failed: %d!\n", engine->name, err);
738 break;
739 }
740
741 if (active) {
742 rq = hang_create_request(&h, engine);
743 if (IS_ERR(rq)) {
744 err = PTR_ERR(rq);
745 pr_err("[%s] Create hang request failed: %d!\n",
746 engine->name, err);
747 goto restore;
748 }
749
750 i915_request_get(rq);
751 i915_request_add(rq);
752
753 if (!wait_until_running(&h, rq)) {
754 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
755
756 pr_err("%s: Failed to start request %llx, at %x\n",
757 __func__, rq->fence.seqno, hws_seqno(&h, rq));
758 intel_engine_dump(engine, &p,
759 "%s\n", engine->name);
760
761 i915_request_put(rq);
762 err = -EIO;
763 goto restore;
764 }
765 }
766
767 if (!using_guc) {
768 err = intel_engine_reset(engine, NULL);
769 if (err) {
770 pr_err("intel_engine_reset(%s) failed, err:%d\n",
771 engine->name, err);
772 goto skip;
773 }
774 }
775
776 if (rq) {
777 /* Ensure the reset happens and kills the engine */
778 err = intel_selftest_wait_for_rq(rq);
779 if (err)
780 pr_err("[%s] Wait for request %lld:%lld [0x%04X] failed: %d!\n",
781 engine->name, rq->fence.context,
782 rq->fence.seqno, rq->context->guc_id.id, err);
783 }
784
785 skip:
786 if (rq)
787 i915_request_put(rq);
788
789 if (i915_reset_count(global) != reset_count) {
790 pr_err("Full GPU reset recorded! (engine reset expected)\n");
791 err = -EINVAL;
792 goto restore;
793 }
794
795 /* GuC based resets are not logged per engine */
796 if (!using_guc) {
797 if (i915_reset_engine_count(global, engine) !=
798 ++reset_engine_count) {
799 pr_err("%s engine reset not recorded!\n",
800 engine->name);
801 err = -EINVAL;
802 goto restore;
803 }
804 }
805
806 count++;
807
808 restore:
809 err2 = intel_selftest_restore_policy(engine, &saved);
810 if (err2)
811 pr_err("[%s] Restore policy failed: %d!\n", engine->name, err);
812 if (err == 0)
813 err = err2;
814 if (err)
815 break;
816 } while (time_before(jiffies, end_time));
817 clear_and_wake_up_bit(I915_RESET_ENGINE + id, >->reset.flags);
818 st_engine_heartbeat_enable(engine);
819 pr_info("%s: Completed %lu %s resets\n",
820 engine->name, count, active ? "active" : "idle");
821
822 if (err)
823 break;
824
825 err = igt_flush_test(gt->i915);
826 if (err) {
827 pr_err("[%s] Flush failed: %d!\n", engine->name, err);
828 break;
829 }
830 }
831
832 if (intel_gt_is_wedged(gt)) {
833 pr_err("GT is wedged!\n");
834 err = -EIO;
835 }
836
837 if (active)
838 hang_fini(&h);
839
840 return err;
841 }
842
igt_reset_idle_engine(void * arg)843 static int igt_reset_idle_engine(void *arg)
844 {
845 return __igt_reset_engine(arg, false);
846 }
847
igt_reset_active_engine(void * arg)848 static int igt_reset_active_engine(void *arg)
849 {
850 return __igt_reset_engine(arg, true);
851 }
852
853 struct active_engine {
854 struct kthread_worker *worker;
855 struct kthread_work work;
856 struct intel_engine_cs *engine;
857 unsigned long resets;
858 unsigned int flags;
859 bool stop;
860 int result;
861 };
862
863 #define TEST_ACTIVE BIT(0)
864 #define TEST_OTHERS BIT(1)
865 #define TEST_SELF BIT(2)
866 #define TEST_PRIORITY BIT(3)
867
active_request_put(struct i915_request * rq)868 static int active_request_put(struct i915_request *rq)
869 {
870 int err = 0;
871
872 if (!rq)
873 return 0;
874
875 if (i915_request_wait(rq, 0, 10 * HZ) < 0) {
876 GEM_TRACE("%s timed out waiting for completion of fence %llx:%lld\n",
877 rq->engine->name,
878 rq->fence.context,
879 rq->fence.seqno);
880 GEM_TRACE_DUMP();
881
882 intel_gt_set_wedged(rq->engine->gt);
883 err = -EIO;
884 }
885
886 i915_request_put(rq);
887
888 return err;
889 }
890
active_engine(struct kthread_work * work)891 static void active_engine(struct kthread_work *work)
892 {
893 I915_RND_STATE(prng);
894 struct active_engine *arg = container_of(work, typeof(*arg), work);
895 struct intel_engine_cs *engine = arg->engine;
896 struct i915_request *rq[8] = {};
897 struct intel_context *ce[ARRAY_SIZE(rq)];
898 unsigned long count;
899 int err = 0;
900
901 for (count = 0; count < ARRAY_SIZE(ce); count++) {
902 ce[count] = intel_context_create(engine);
903 if (IS_ERR(ce[count])) {
904 arg->result = PTR_ERR(ce[count]);
905 pr_err("[%s] Create context #%ld failed: %d!\n",
906 engine->name, count, arg->result);
907 while (--count)
908 intel_context_put(ce[count]);
909 return;
910 }
911 }
912
913 count = 0;
914 while (!READ_ONCE(arg->stop)) {
915 unsigned int idx = count++ & (ARRAY_SIZE(rq) - 1);
916 struct i915_request *old = rq[idx];
917 struct i915_request *new;
918
919 new = intel_context_create_request(ce[idx]);
920 if (IS_ERR(new)) {
921 err = PTR_ERR(new);
922 pr_err("[%s] Create request #%d failed: %d!\n", engine->name, idx, err);
923 break;
924 }
925
926 rq[idx] = i915_request_get(new);
927 i915_request_add(new);
928
929 if (engine->sched_engine->schedule && arg->flags & TEST_PRIORITY) {
930 struct i915_sched_attr attr = {
931 .priority =
932 i915_prandom_u32_max_state(512, &prng),
933 };
934 engine->sched_engine->schedule(rq[idx], &attr);
935 }
936
937 err = active_request_put(old);
938 if (err) {
939 pr_err("[%s] Request put failed: %d!\n", engine->name, err);
940 break;
941 }
942
943 cond_resched();
944 }
945
946 for (count = 0; count < ARRAY_SIZE(rq); count++) {
947 int err__ = active_request_put(rq[count]);
948
949 if (err)
950 pr_err("[%s] Request put #%ld failed: %d!\n", engine->name, count, err);
951
952 /* Keep the first error */
953 if (!err)
954 err = err__;
955
956 intel_context_put(ce[count]);
957 }
958
959 arg->result = err;
960 }
961
__igt_reset_engines(struct intel_gt * gt,const char * test_name,unsigned int flags)962 static int __igt_reset_engines(struct intel_gt *gt,
963 const char *test_name,
964 unsigned int flags)
965 {
966 struct i915_gpu_error *global = >->i915->gpu_error;
967 struct intel_engine_cs *engine, *other;
968 struct active_engine *threads;
969 enum intel_engine_id id, tmp;
970 struct hang h;
971 int err = 0;
972
973 /* Check that issuing a reset on one engine does not interfere
974 * with any other engine.
975 */
976
977 if (!intel_has_reset_engine(gt))
978 return 0;
979
980 if (flags & TEST_ACTIVE) {
981 err = hang_init(&h, gt);
982 if (err)
983 return err;
984
985 if (flags & TEST_PRIORITY)
986 h.ctx->sched.priority = 1024;
987 }
988
989 threads = kmalloc_array(I915_NUM_ENGINES, sizeof(*threads), GFP_KERNEL);
990 if (!threads)
991 return -ENOMEM;
992
993 for_each_engine(engine, gt, id) {
994 unsigned long device = i915_reset_count(global);
995 unsigned long count = 0, reported;
996 bool using_guc = intel_engine_uses_guc(engine);
997 IGT_TIMEOUT(end_time);
998
999 if (flags & TEST_ACTIVE) {
1000 if (!intel_engine_can_store_dword(engine))
1001 continue;
1002 } else if (using_guc)
1003 continue;
1004
1005 if (!wait_for_idle(engine)) {
1006 pr_err("i915_reset_engine(%s:%s): failed to idle before reset\n",
1007 engine->name, test_name);
1008 err = -EIO;
1009 break;
1010 }
1011
1012 memset(threads, 0, sizeof(*threads) * I915_NUM_ENGINES);
1013 for_each_engine(other, gt, tmp) {
1014 struct kthread_worker *worker;
1015
1016 threads[tmp].resets =
1017 i915_reset_engine_count(global, other);
1018
1019 if (other == engine && !(flags & TEST_SELF))
1020 continue;
1021
1022 if (other != engine && !(flags & TEST_OTHERS))
1023 continue;
1024
1025 threads[tmp].engine = other;
1026 threads[tmp].flags = flags;
1027
1028 worker = kthread_create_worker(0, "igt/%s",
1029 other->name);
1030 if (IS_ERR(worker)) {
1031 err = PTR_ERR(worker);
1032 pr_err("[%s] Worker create failed: %d!\n",
1033 engine->name, err);
1034 goto unwind;
1035 }
1036
1037 threads[tmp].worker = worker;
1038
1039 kthread_init_work(&threads[tmp].work, active_engine);
1040 kthread_queue_work(threads[tmp].worker,
1041 &threads[tmp].work);
1042 }
1043
1044 st_engine_heartbeat_disable_no_pm(engine);
1045 GEM_BUG_ON(test_and_set_bit(I915_RESET_ENGINE + id,
1046 >->reset.flags));
1047 do {
1048 struct i915_request *rq = NULL;
1049 struct intel_selftest_saved_policy saved;
1050 int err2;
1051
1052 err = intel_selftest_modify_policy(engine, &saved,
1053 SELFTEST_SCHEDULER_MODIFY_FAST_RESET);
1054 if (err) {
1055 pr_err("[%s] Modify policy failed: %d!\n", engine->name, err);
1056 break;
1057 }
1058
1059 if (flags & TEST_ACTIVE) {
1060 rq = hang_create_request(&h, engine);
1061 if (IS_ERR(rq)) {
1062 err = PTR_ERR(rq);
1063 pr_err("[%s] Create hang request failed: %d!\n",
1064 engine->name, err);
1065 goto restore;
1066 }
1067
1068 i915_request_get(rq);
1069 i915_request_add(rq);
1070
1071 if (!wait_until_running(&h, rq)) {
1072 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1073
1074 pr_err("%s: Failed to start request %llx, at %x\n",
1075 __func__, rq->fence.seqno, hws_seqno(&h, rq));
1076 intel_engine_dump(engine, &p,
1077 "%s\n", engine->name);
1078
1079 i915_request_put(rq);
1080 err = -EIO;
1081 goto restore;
1082 }
1083 } else {
1084 intel_engine_pm_get(engine);
1085 }
1086
1087 if (!using_guc) {
1088 err = intel_engine_reset(engine, NULL);
1089 if (err) {
1090 pr_err("i915_reset_engine(%s:%s): failed, err=%d\n",
1091 engine->name, test_name, err);
1092 goto restore;
1093 }
1094 }
1095
1096 if (rq) {
1097 /* Ensure the reset happens and kills the engine */
1098 err = intel_selftest_wait_for_rq(rq);
1099 if (err)
1100 pr_err("[%s] Wait for request %lld:%lld [0x%04X] failed: %d!\n",
1101 engine->name, rq->fence.context,
1102 rq->fence.seqno, rq->context->guc_id.id, err);
1103 }
1104
1105 count++;
1106
1107 if (rq) {
1108 if (rq->fence.error != -EIO) {
1109 pr_err("i915_reset_engine(%s:%s): failed to reset request %lld:%lld [0x%04X]\n",
1110 engine->name, test_name,
1111 rq->fence.context,
1112 rq->fence.seqno, rq->context->guc_id.id);
1113 i915_request_put(rq);
1114
1115 GEM_TRACE_DUMP();
1116 intel_gt_set_wedged(gt);
1117 err = -EIO;
1118 goto restore;
1119 }
1120
1121 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
1122 struct drm_printer p =
1123 drm_info_printer(gt->i915->drm.dev);
1124
1125 pr_err("i915_reset_engine(%s:%s):"
1126 " failed to complete request %llx:%lld after reset\n",
1127 engine->name, test_name,
1128 rq->fence.context,
1129 rq->fence.seqno);
1130 intel_engine_dump(engine, &p,
1131 "%s\n", engine->name);
1132 i915_request_put(rq);
1133
1134 GEM_TRACE_DUMP();
1135 intel_gt_set_wedged(gt);
1136 err = -EIO;
1137 goto restore;
1138 }
1139
1140 i915_request_put(rq);
1141 }
1142
1143 if (!(flags & TEST_ACTIVE))
1144 intel_engine_pm_put(engine);
1145
1146 if (!(flags & TEST_SELF) && !wait_for_idle(engine)) {
1147 struct drm_printer p =
1148 drm_info_printer(gt->i915->drm.dev);
1149
1150 pr_err("i915_reset_engine(%s:%s):"
1151 " failed to idle after reset\n",
1152 engine->name, test_name);
1153 intel_engine_dump(engine, &p,
1154 "%s\n", engine->name);
1155
1156 err = -EIO;
1157 goto restore;
1158 }
1159
1160 restore:
1161 err2 = intel_selftest_restore_policy(engine, &saved);
1162 if (err2)
1163 pr_err("[%s] Restore policy failed: %d!\n", engine->name, err2);
1164 if (err == 0)
1165 err = err2;
1166 if (err)
1167 break;
1168 } while (time_before(jiffies, end_time));
1169 clear_and_wake_up_bit(I915_RESET_ENGINE + id, >->reset.flags);
1170 st_engine_heartbeat_enable_no_pm(engine);
1171
1172 pr_info("i915_reset_engine(%s:%s): %lu resets\n",
1173 engine->name, test_name, count);
1174
1175 /* GuC based resets are not logged per engine */
1176 if (!using_guc) {
1177 reported = i915_reset_engine_count(global, engine);
1178 reported -= threads[engine->id].resets;
1179 if (reported != count) {
1180 pr_err("i915_reset_engine(%s:%s): reset %lu times, but reported %lu\n",
1181 engine->name, test_name, count, reported);
1182 if (!err)
1183 err = -EINVAL;
1184 }
1185 }
1186
1187 unwind:
1188 for_each_engine(other, gt, tmp) {
1189 int ret;
1190
1191 if (!threads[tmp].worker)
1192 continue;
1193
1194 WRITE_ONCE(threads[tmp].stop, true);
1195 kthread_flush_work(&threads[tmp].work);
1196 ret = READ_ONCE(threads[tmp].result);
1197 if (ret) {
1198 pr_err("kthread for other engine %s failed, err=%d\n",
1199 other->name, ret);
1200 if (!err)
1201 err = ret;
1202 }
1203
1204 kthread_destroy_worker(threads[tmp].worker);
1205
1206 /* GuC based resets are not logged per engine */
1207 if (!using_guc) {
1208 if (other->uabi_class != engine->uabi_class &&
1209 threads[tmp].resets !=
1210 i915_reset_engine_count(global, other)) {
1211 pr_err("Innocent engine %s was reset (count=%ld)\n",
1212 other->name,
1213 i915_reset_engine_count(global, other) -
1214 threads[tmp].resets);
1215 if (!err)
1216 err = -EINVAL;
1217 }
1218 }
1219 }
1220
1221 if (device != i915_reset_count(global)) {
1222 pr_err("Global reset (count=%ld)!\n",
1223 i915_reset_count(global) - device);
1224 if (!err)
1225 err = -EINVAL;
1226 }
1227
1228 if (err)
1229 break;
1230
1231 err = igt_flush_test(gt->i915);
1232 if (err) {
1233 pr_err("[%s] Flush failed: %d!\n", engine->name, err);
1234 break;
1235 }
1236 }
1237 kfree(threads);
1238
1239 if (intel_gt_is_wedged(gt))
1240 err = -EIO;
1241
1242 if (flags & TEST_ACTIVE)
1243 hang_fini(&h);
1244
1245 return err;
1246 }
1247
igt_reset_engines(void * arg)1248 static int igt_reset_engines(void *arg)
1249 {
1250 static const struct {
1251 const char *name;
1252 unsigned int flags;
1253 } phases[] = {
1254 { "idle", 0 },
1255 { "active", TEST_ACTIVE },
1256 { "others-idle", TEST_OTHERS },
1257 { "others-active", TEST_OTHERS | TEST_ACTIVE },
1258 {
1259 "others-priority",
1260 TEST_OTHERS | TEST_ACTIVE | TEST_PRIORITY
1261 },
1262 {
1263 "self-priority",
1264 TEST_ACTIVE | TEST_PRIORITY | TEST_SELF,
1265 },
1266 { }
1267 };
1268 struct intel_gt *gt = arg;
1269 typeof(*phases) *p;
1270 int err;
1271
1272 for (p = phases; p->name; p++) {
1273 if (p->flags & TEST_PRIORITY) {
1274 if (!(gt->i915->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY))
1275 continue;
1276 }
1277
1278 err = __igt_reset_engines(arg, p->name, p->flags);
1279 if (err)
1280 return err;
1281 }
1282
1283 return 0;
1284 }
1285
fake_hangcheck(struct intel_gt * gt,intel_engine_mask_t mask)1286 static u32 fake_hangcheck(struct intel_gt *gt, intel_engine_mask_t mask)
1287 {
1288 u32 count = i915_reset_count(>->i915->gpu_error);
1289
1290 intel_gt_reset(gt, mask, NULL);
1291
1292 return count;
1293 }
1294
igt_reset_wait(void * arg)1295 static int igt_reset_wait(void *arg)
1296 {
1297 struct intel_gt *gt = arg;
1298 struct i915_gpu_error *global = >->i915->gpu_error;
1299 struct intel_engine_cs *engine;
1300 struct i915_request *rq;
1301 unsigned int reset_count;
1302 struct hang h;
1303 long timeout;
1304 int err;
1305
1306 engine = intel_selftest_find_any_engine(gt);
1307
1308 if (!engine || !intel_engine_can_store_dword(engine))
1309 return 0;
1310
1311 /* Check that we detect a stuck waiter and issue a reset */
1312
1313 igt_global_reset_lock(gt);
1314
1315 err = hang_init(&h, gt);
1316 if (err) {
1317 pr_err("[%s] Hang init failed: %d!\n", engine->name, err);
1318 goto unlock;
1319 }
1320
1321 rq = hang_create_request(&h, engine);
1322 if (IS_ERR(rq)) {
1323 err = PTR_ERR(rq);
1324 pr_err("[%s] Create hang request failed: %d!\n", engine->name, err);
1325 goto fini;
1326 }
1327
1328 i915_request_get(rq);
1329 i915_request_add(rq);
1330
1331 if (!wait_until_running(&h, rq)) {
1332 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1333
1334 pr_err("%s: Failed to start request %llx, at %x\n",
1335 __func__, rq->fence.seqno, hws_seqno(&h, rq));
1336 intel_engine_dump(rq->engine, &p, "%s\n", rq->engine->name);
1337
1338 intel_gt_set_wedged(gt);
1339
1340 err = -EIO;
1341 goto out_rq;
1342 }
1343
1344 reset_count = fake_hangcheck(gt, ALL_ENGINES);
1345
1346 timeout = i915_request_wait(rq, 0, 10);
1347 if (timeout < 0) {
1348 pr_err("i915_request_wait failed on a stuck request: err=%ld\n",
1349 timeout);
1350 err = timeout;
1351 goto out_rq;
1352 }
1353
1354 if (i915_reset_count(global) == reset_count) {
1355 pr_err("No GPU reset recorded!\n");
1356 err = -EINVAL;
1357 goto out_rq;
1358 }
1359
1360 out_rq:
1361 i915_request_put(rq);
1362 fini:
1363 hang_fini(&h);
1364 unlock:
1365 igt_global_reset_unlock(gt);
1366
1367 if (intel_gt_is_wedged(gt))
1368 return -EIO;
1369
1370 return err;
1371 }
1372
1373 struct evict_vma {
1374 struct completion completion;
1375 struct i915_vma *vma;
1376 };
1377
evict_vma(void * data)1378 static int evict_vma(void *data)
1379 {
1380 struct evict_vma *arg = data;
1381 struct i915_address_space *vm = arg->vma->vm;
1382 struct drm_mm_node evict = arg->vma->node;
1383 int err;
1384
1385 complete(&arg->completion);
1386
1387 mutex_lock(&vm->mutex);
1388 err = i915_gem_evict_for_node(vm, NULL, &evict, 0);
1389 mutex_unlock(&vm->mutex);
1390
1391 return err;
1392 }
1393
evict_fence(void * data)1394 static int evict_fence(void *data)
1395 {
1396 struct evict_vma *arg = data;
1397 int err;
1398
1399 complete(&arg->completion);
1400
1401 /* Mark the fence register as dirty to force the mmio update. */
1402 err = i915_gem_object_set_tiling(arg->vma->obj, I915_TILING_Y, 512);
1403 if (err) {
1404 pr_err("Invalid Y-tiling settings; err:%d\n", err);
1405 return err;
1406 }
1407
1408 err = i915_vma_pin(arg->vma, 0, 0, PIN_GLOBAL | PIN_MAPPABLE);
1409 if (err) {
1410 pr_err("Unable to pin vma for Y-tiled fence; err:%d\n", err);
1411 return err;
1412 }
1413
1414 err = i915_vma_pin_fence(arg->vma);
1415 i915_vma_unpin(arg->vma);
1416 if (err) {
1417 pr_err("Unable to pin Y-tiled fence; err:%d\n", err);
1418 return err;
1419 }
1420
1421 i915_vma_unpin_fence(arg->vma);
1422
1423 return 0;
1424 }
1425
__igt_reset_evict_vma(struct intel_gt * gt,struct i915_address_space * vm,int (* fn)(void *),unsigned int flags)1426 static int __igt_reset_evict_vma(struct intel_gt *gt,
1427 struct i915_address_space *vm,
1428 int (*fn)(void *),
1429 unsigned int flags)
1430 {
1431 struct intel_engine_cs *engine;
1432 struct drm_i915_gem_object *obj;
1433 struct task_struct *tsk = NULL;
1434 struct i915_request *rq;
1435 struct evict_vma arg;
1436 struct hang h;
1437 unsigned int pin_flags;
1438 int err;
1439
1440 if (!gt->ggtt->num_fences && flags & EXEC_OBJECT_NEEDS_FENCE)
1441 return 0;
1442
1443 engine = intel_selftest_find_any_engine(gt);
1444
1445 if (!engine || !intel_engine_can_store_dword(engine))
1446 return 0;
1447
1448 /* Check that we can recover an unbind stuck on a hanging request */
1449
1450 err = hang_init(&h, gt);
1451 if (err) {
1452 pr_err("[%s] Hang init failed: %d!\n", engine->name, err);
1453 return err;
1454 }
1455
1456 obj = i915_gem_object_create_internal(gt->i915, SZ_1M);
1457 if (IS_ERR(obj)) {
1458 err = PTR_ERR(obj);
1459 pr_err("[%s] Create object failed: %d!\n", engine->name, err);
1460 goto fini;
1461 }
1462
1463 if (flags & EXEC_OBJECT_NEEDS_FENCE) {
1464 err = i915_gem_object_set_tiling(obj, I915_TILING_X, 512);
1465 if (err) {
1466 pr_err("Invalid X-tiling settings; err:%d\n", err);
1467 goto out_obj;
1468 }
1469 }
1470
1471 arg.vma = i915_vma_instance(obj, vm, NULL);
1472 if (IS_ERR(arg.vma)) {
1473 err = PTR_ERR(arg.vma);
1474 pr_err("[%s] VMA instance failed: %d!\n", engine->name, err);
1475 goto out_obj;
1476 }
1477
1478 rq = hang_create_request(&h, engine);
1479 if (IS_ERR(rq)) {
1480 err = PTR_ERR(rq);
1481 pr_err("[%s] Create hang request failed: %d!\n", engine->name, err);
1482 goto out_obj;
1483 }
1484
1485 pin_flags = i915_vma_is_ggtt(arg.vma) ? PIN_GLOBAL : PIN_USER;
1486
1487 if (flags & EXEC_OBJECT_NEEDS_FENCE)
1488 pin_flags |= PIN_MAPPABLE;
1489
1490 err = i915_vma_pin(arg.vma, 0, 0, pin_flags);
1491 if (err) {
1492 i915_request_add(rq);
1493 pr_err("[%s] VMA pin failed: %d!\n", engine->name, err);
1494 goto out_obj;
1495 }
1496
1497 if (flags & EXEC_OBJECT_NEEDS_FENCE) {
1498 err = i915_vma_pin_fence(arg.vma);
1499 if (err) {
1500 pr_err("Unable to pin X-tiled fence; err:%d\n", err);
1501 i915_vma_unpin(arg.vma);
1502 i915_request_add(rq);
1503 goto out_obj;
1504 }
1505 }
1506
1507 err = igt_vma_move_to_active_unlocked(arg.vma, rq, flags);
1508 if (err)
1509 pr_err("[%s] Move to active failed: %d!\n", engine->name, err);
1510
1511 if (flags & EXEC_OBJECT_NEEDS_FENCE)
1512 i915_vma_unpin_fence(arg.vma);
1513 i915_vma_unpin(arg.vma);
1514
1515 i915_request_get(rq);
1516 i915_request_add(rq);
1517 if (err)
1518 goto out_rq;
1519
1520 if (!wait_until_running(&h, rq)) {
1521 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1522
1523 pr_err("%s: Failed to start request %llx, at %x\n",
1524 __func__, rq->fence.seqno, hws_seqno(&h, rq));
1525 intel_engine_dump(rq->engine, &p, "%s\n", rq->engine->name);
1526
1527 intel_gt_set_wedged(gt);
1528 goto out_reset;
1529 }
1530
1531 init_completion(&arg.completion);
1532
1533 tsk = kthread_run(fn, &arg, "igt/evict_vma");
1534 if (IS_ERR(tsk)) {
1535 err = PTR_ERR(tsk);
1536 pr_err("[%s] Thread spawn failed: %d!\n", engine->name, err);
1537 tsk = NULL;
1538 goto out_reset;
1539 }
1540 get_task_struct(tsk);
1541
1542 wait_for_completion(&arg.completion);
1543
1544 if (wait_for(!list_empty(&rq->fence.cb_list), 10)) {
1545 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1546
1547 pr_err("igt/evict_vma kthread did not wait\n");
1548 intel_engine_dump(rq->engine, &p, "%s\n", rq->engine->name);
1549
1550 intel_gt_set_wedged(gt);
1551 goto out_reset;
1552 }
1553
1554 out_reset:
1555 igt_global_reset_lock(gt);
1556 fake_hangcheck(gt, rq->engine->mask);
1557 igt_global_reset_unlock(gt);
1558
1559 if (tsk) {
1560 struct intel_wedge_me w;
1561
1562 /* The reset, even indirectly, should take less than 10ms. */
1563 intel_wedge_on_timeout(&w, gt, HZ / 10 /* 100ms */)
1564 err = kthread_stop(tsk);
1565
1566 put_task_struct(tsk);
1567 }
1568
1569 out_rq:
1570 i915_request_put(rq);
1571 out_obj:
1572 i915_gem_object_put(obj);
1573 fini:
1574 hang_fini(&h);
1575 if (intel_gt_is_wedged(gt))
1576 return -EIO;
1577
1578 return err;
1579 }
1580
igt_reset_evict_ggtt(void * arg)1581 static int igt_reset_evict_ggtt(void *arg)
1582 {
1583 struct intel_gt *gt = arg;
1584
1585 return __igt_reset_evict_vma(gt, >->ggtt->vm,
1586 evict_vma, EXEC_OBJECT_WRITE);
1587 }
1588
igt_reset_evict_ppgtt(void * arg)1589 static int igt_reset_evict_ppgtt(void *arg)
1590 {
1591 struct intel_gt *gt = arg;
1592 struct i915_ppgtt *ppgtt;
1593 int err;
1594
1595 /* aliasing == global gtt locking, covered above */
1596 if (INTEL_PPGTT(gt->i915) < INTEL_PPGTT_FULL)
1597 return 0;
1598
1599 ppgtt = i915_ppgtt_create(gt, 0);
1600 if (IS_ERR(ppgtt))
1601 return PTR_ERR(ppgtt);
1602
1603 err = __igt_reset_evict_vma(gt, &ppgtt->vm,
1604 evict_vma, EXEC_OBJECT_WRITE);
1605 i915_vm_put(&ppgtt->vm);
1606
1607 return err;
1608 }
1609
igt_reset_evict_fence(void * arg)1610 static int igt_reset_evict_fence(void *arg)
1611 {
1612 struct intel_gt *gt = arg;
1613
1614 return __igt_reset_evict_vma(gt, >->ggtt->vm,
1615 evict_fence, EXEC_OBJECT_NEEDS_FENCE);
1616 }
1617
wait_for_others(struct intel_gt * gt,struct intel_engine_cs * exclude)1618 static int wait_for_others(struct intel_gt *gt,
1619 struct intel_engine_cs *exclude)
1620 {
1621 struct intel_engine_cs *engine;
1622 enum intel_engine_id id;
1623
1624 for_each_engine(engine, gt, id) {
1625 if (engine == exclude)
1626 continue;
1627
1628 if (!wait_for_idle(engine))
1629 return -EIO;
1630 }
1631
1632 return 0;
1633 }
1634
igt_reset_queue(void * arg)1635 static int igt_reset_queue(void *arg)
1636 {
1637 struct intel_gt *gt = arg;
1638 struct i915_gpu_error *global = >->i915->gpu_error;
1639 struct intel_engine_cs *engine;
1640 enum intel_engine_id id;
1641 struct hang h;
1642 int err;
1643
1644 /* Check that we replay pending requests following a hang */
1645
1646 igt_global_reset_lock(gt);
1647
1648 err = hang_init(&h, gt);
1649 if (err)
1650 goto unlock;
1651
1652 for_each_engine(engine, gt, id) {
1653 struct intel_selftest_saved_policy saved;
1654 struct i915_request *prev;
1655 IGT_TIMEOUT(end_time);
1656 unsigned int count;
1657 bool using_guc = intel_engine_uses_guc(engine);
1658
1659 if (!intel_engine_can_store_dword(engine))
1660 continue;
1661
1662 if (using_guc) {
1663 err = intel_selftest_modify_policy(engine, &saved,
1664 SELFTEST_SCHEDULER_MODIFY_NO_HANGCHECK);
1665 if (err) {
1666 pr_err("[%s] Modify policy failed: %d!\n", engine->name, err);
1667 goto fini;
1668 }
1669 }
1670
1671 prev = hang_create_request(&h, engine);
1672 if (IS_ERR(prev)) {
1673 err = PTR_ERR(prev);
1674 pr_err("[%s] Create 'prev' hang request failed: %d!\n", engine->name, err);
1675 goto restore;
1676 }
1677
1678 i915_request_get(prev);
1679 i915_request_add(prev);
1680
1681 count = 0;
1682 do {
1683 struct i915_request *rq;
1684 unsigned int reset_count;
1685
1686 rq = hang_create_request(&h, engine);
1687 if (IS_ERR(rq)) {
1688 err = PTR_ERR(rq);
1689 pr_err("[%s] Create hang request failed: %d!\n", engine->name, err);
1690 goto restore;
1691 }
1692
1693 i915_request_get(rq);
1694 i915_request_add(rq);
1695
1696 /*
1697 * XXX We don't handle resetting the kernel context
1698 * very well. If we trigger a device reset twice in
1699 * quick succession while the kernel context is
1700 * executing, we may end up skipping the breadcrumb.
1701 * This is really only a problem for the selftest as
1702 * normally there is a large interlude between resets
1703 * (hangcheck), or we focus on resetting just one
1704 * engine and so avoid repeatedly resetting innocents.
1705 */
1706 err = wait_for_others(gt, engine);
1707 if (err) {
1708 pr_err("%s(%s): Failed to idle other inactive engines after device reset\n",
1709 __func__, engine->name);
1710 i915_request_put(rq);
1711 i915_request_put(prev);
1712
1713 GEM_TRACE_DUMP();
1714 intel_gt_set_wedged(gt);
1715 goto restore;
1716 }
1717
1718 if (!wait_until_running(&h, prev)) {
1719 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1720
1721 pr_err("%s(%s): Failed to start request %llx, at %x\n",
1722 __func__, engine->name,
1723 prev->fence.seqno, hws_seqno(&h, prev));
1724 intel_engine_dump(engine, &p,
1725 "%s\n", engine->name);
1726
1727 i915_request_put(rq);
1728 i915_request_put(prev);
1729
1730 intel_gt_set_wedged(gt);
1731
1732 err = -EIO;
1733 goto restore;
1734 }
1735
1736 reset_count = fake_hangcheck(gt, BIT(id));
1737
1738 if (prev->fence.error != -EIO) {
1739 pr_err("GPU reset not recorded on hanging request [fence.error=%d]!\n",
1740 prev->fence.error);
1741 i915_request_put(rq);
1742 i915_request_put(prev);
1743 err = -EINVAL;
1744 goto restore;
1745 }
1746
1747 if (rq->fence.error) {
1748 pr_err("Fence error status not zero [%d] after unrelated reset\n",
1749 rq->fence.error);
1750 i915_request_put(rq);
1751 i915_request_put(prev);
1752 err = -EINVAL;
1753 goto restore;
1754 }
1755
1756 if (i915_reset_count(global) == reset_count) {
1757 pr_err("No GPU reset recorded!\n");
1758 i915_request_put(rq);
1759 i915_request_put(prev);
1760 err = -EINVAL;
1761 goto restore;
1762 }
1763
1764 i915_request_put(prev);
1765 prev = rq;
1766 count++;
1767 } while (time_before(jiffies, end_time));
1768 pr_info("%s: Completed %d queued resets\n",
1769 engine->name, count);
1770
1771 *h.batch = MI_BATCH_BUFFER_END;
1772 intel_gt_chipset_flush(engine->gt);
1773
1774 i915_request_put(prev);
1775
1776 restore:
1777 if (using_guc) {
1778 int err2 = intel_selftest_restore_policy(engine, &saved);
1779
1780 if (err2)
1781 pr_err("%s:%d> [%s] Restore policy failed: %d!\n",
1782 __func__, __LINE__, engine->name, err2);
1783 if (err == 0)
1784 err = err2;
1785 }
1786 if (err)
1787 goto fini;
1788
1789 err = igt_flush_test(gt->i915);
1790 if (err) {
1791 pr_err("[%s] Flush failed: %d!\n", engine->name, err);
1792 break;
1793 }
1794 }
1795
1796 fini:
1797 hang_fini(&h);
1798 unlock:
1799 igt_global_reset_unlock(gt);
1800
1801 if (intel_gt_is_wedged(gt))
1802 return -EIO;
1803
1804 return err;
1805 }
1806
igt_handle_error(void * arg)1807 static int igt_handle_error(void *arg)
1808 {
1809 struct intel_gt *gt = arg;
1810 struct i915_gpu_error *global = >->i915->gpu_error;
1811 struct intel_engine_cs *engine;
1812 struct hang h;
1813 struct i915_request *rq;
1814 struct i915_gpu_coredump *error;
1815 int err;
1816
1817 engine = intel_selftest_find_any_engine(gt);
1818
1819 /* Check that we can issue a global GPU and engine reset */
1820
1821 if (!intel_has_reset_engine(gt))
1822 return 0;
1823
1824 if (!engine || !intel_engine_can_store_dword(engine))
1825 return 0;
1826
1827 err = hang_init(&h, gt);
1828 if (err) {
1829 pr_err("[%s] Hang init failed: %d!\n", engine->name, err);
1830 return err;
1831 }
1832
1833 rq = hang_create_request(&h, engine);
1834 if (IS_ERR(rq)) {
1835 err = PTR_ERR(rq);
1836 pr_err("[%s] Create hang request failed: %d!\n", engine->name, err);
1837 goto err_fini;
1838 }
1839
1840 i915_request_get(rq);
1841 i915_request_add(rq);
1842
1843 if (!wait_until_running(&h, rq)) {
1844 struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1845
1846 pr_err("%s: Failed to start request %llx, at %x\n",
1847 __func__, rq->fence.seqno, hws_seqno(&h, rq));
1848 intel_engine_dump(rq->engine, &p, "%s\n", rq->engine->name);
1849
1850 intel_gt_set_wedged(gt);
1851
1852 err = -EIO;
1853 goto err_request;
1854 }
1855
1856 /* Temporarily disable error capture */
1857 error = xchg(&global->first_error, (void *)-1);
1858
1859 intel_gt_handle_error(gt, engine->mask, 0, NULL);
1860
1861 xchg(&global->first_error, error);
1862
1863 if (rq->fence.error != -EIO) {
1864 pr_err("Guilty request not identified!\n");
1865 err = -EINVAL;
1866 goto err_request;
1867 }
1868
1869 err_request:
1870 i915_request_put(rq);
1871 err_fini:
1872 hang_fini(&h);
1873 return err;
1874 }
1875
__igt_atomic_reset_engine(struct intel_engine_cs * engine,const struct igt_atomic_section * p,const char * mode)1876 static int __igt_atomic_reset_engine(struct intel_engine_cs *engine,
1877 const struct igt_atomic_section *p,
1878 const char *mode)
1879 {
1880 struct tasklet_struct * const t = &engine->sched_engine->tasklet;
1881 int err;
1882
1883 GEM_TRACE("i915_reset_engine(%s:%s) under %s\n",
1884 engine->name, mode, p->name);
1885
1886 if (t->func)
1887 tasklet_disable(t);
1888 if (strcmp(p->name, "softirq"))
1889 local_bh_disable();
1890 p->critical_section_begin();
1891
1892 err = __intel_engine_reset_bh(engine, NULL);
1893
1894 p->critical_section_end();
1895 if (strcmp(p->name, "softirq"))
1896 local_bh_enable();
1897 if (t->func) {
1898 tasklet_enable(t);
1899 tasklet_hi_schedule(t);
1900 }
1901
1902 if (err)
1903 pr_err("i915_reset_engine(%s:%s) failed under %s\n",
1904 engine->name, mode, p->name);
1905
1906 return err;
1907 }
1908
igt_atomic_reset_engine(struct intel_engine_cs * engine,const struct igt_atomic_section * p)1909 static int igt_atomic_reset_engine(struct intel_engine_cs *engine,
1910 const struct igt_atomic_section *p)
1911 {
1912 struct i915_request *rq;
1913 struct hang h;
1914 int err;
1915
1916 err = __igt_atomic_reset_engine(engine, p, "idle");
1917 if (err)
1918 return err;
1919
1920 err = hang_init(&h, engine->gt);
1921 if (err) {
1922 pr_err("[%s] Hang init failed: %d!\n", engine->name, err);
1923 return err;
1924 }
1925
1926 rq = hang_create_request(&h, engine);
1927 if (IS_ERR(rq)) {
1928 err = PTR_ERR(rq);
1929 pr_err("[%s] Create hang request failed: %d!\n", engine->name, err);
1930 goto out;
1931 }
1932
1933 i915_request_get(rq);
1934 i915_request_add(rq);
1935
1936 if (wait_until_running(&h, rq)) {
1937 err = __igt_atomic_reset_engine(engine, p, "active");
1938 } else {
1939 pr_err("%s(%s): Failed to start request %llx, at %x\n",
1940 __func__, engine->name,
1941 rq->fence.seqno, hws_seqno(&h, rq));
1942 intel_gt_set_wedged(engine->gt);
1943 err = -EIO;
1944 }
1945
1946 if (err == 0) {
1947 struct intel_wedge_me w;
1948
1949 intel_wedge_on_timeout(&w, engine->gt, HZ / 20 /* 50ms */)
1950 i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
1951 if (intel_gt_is_wedged(engine->gt))
1952 err = -EIO;
1953 }
1954
1955 i915_request_put(rq);
1956 out:
1957 hang_fini(&h);
1958 return err;
1959 }
1960
igt_reset_engines_atomic(void * arg)1961 static int igt_reset_engines_atomic(void *arg)
1962 {
1963 struct intel_gt *gt = arg;
1964 const typeof(*igt_atomic_phases) *p;
1965 int err = 0;
1966
1967 /* Check that the engines resets are usable from atomic context */
1968
1969 if (!intel_has_reset_engine(gt))
1970 return 0;
1971
1972 if (intel_uc_uses_guc_submission(>->uc))
1973 return 0;
1974
1975 igt_global_reset_lock(gt);
1976
1977 /* Flush any requests before we get started and check basics */
1978 if (!igt_force_reset(gt))
1979 goto unlock;
1980
1981 for (p = igt_atomic_phases; p->name; p++) {
1982 struct intel_engine_cs *engine;
1983 enum intel_engine_id id;
1984
1985 for_each_engine(engine, gt, id) {
1986 err = igt_atomic_reset_engine(engine, p);
1987 if (err)
1988 goto out;
1989 }
1990 }
1991
1992 out:
1993 /* As we poke around the guts, do a full reset before continuing. */
1994 igt_force_reset(gt);
1995 unlock:
1996 igt_global_reset_unlock(gt);
1997
1998 return err;
1999 }
2000
intel_hangcheck_live_selftests(struct drm_i915_private * i915)2001 int intel_hangcheck_live_selftests(struct drm_i915_private *i915)
2002 {
2003 static const struct i915_subtest tests[] = {
2004 SUBTEST(igt_hang_sanitycheck),
2005 SUBTEST(igt_reset_nop),
2006 SUBTEST(igt_reset_nop_engine),
2007 SUBTEST(igt_reset_idle_engine),
2008 SUBTEST(igt_reset_active_engine),
2009 SUBTEST(igt_reset_fail_engine),
2010 SUBTEST(igt_reset_engines),
2011 SUBTEST(igt_reset_engines_atomic),
2012 SUBTEST(igt_reset_queue),
2013 SUBTEST(igt_reset_wait),
2014 SUBTEST(igt_reset_evict_ggtt),
2015 SUBTEST(igt_reset_evict_ppgtt),
2016 SUBTEST(igt_reset_evict_fence),
2017 SUBTEST(igt_handle_error),
2018 };
2019 struct intel_gt *gt = to_gt(i915);
2020 intel_wakeref_t wakeref;
2021 int err;
2022
2023 if (!intel_has_gpu_reset(gt))
2024 return 0;
2025
2026 if (intel_gt_is_wedged(gt))
2027 return -EIO; /* we're long past hope of a successful reset */
2028
2029 wakeref = intel_runtime_pm_get(gt->uncore->rpm);
2030
2031 err = intel_gt_live_subtests(tests, gt);
2032
2033 intel_runtime_pm_put(gt->uncore->rpm, wakeref);
2034
2035 return err;
2036 }
2037