xref: /linux/drivers/gpu/drm/i915/gt/selftest_execlists.c (revision 9f2c9170934eace462499ba0bfe042cc72900173)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2018 Intel Corporation
4  */
5 
6 #include <linux/prime_numbers.h>
7 
8 #include "gem/i915_gem_internal.h"
9 #include "gem/i915_gem_pm.h"
10 #include "gt/intel_engine_heartbeat.h"
11 #include "gt/intel_reset.h"
12 #include "gt/selftest_engine_heartbeat.h"
13 
14 #include "i915_selftest.h"
15 #include "selftests/i915_random.h"
16 #include "selftests/igt_flush_test.h"
17 #include "selftests/igt_live_test.h"
18 #include "selftests/igt_spinner.h"
19 #include "selftests/lib_sw_fence.h"
20 
21 #include "gem/selftests/igt_gem_utils.h"
22 #include "gem/selftests/mock_context.h"
23 
24 #define CS_GPR(engine, n) ((engine)->mmio_base + 0x600 + (n) * 4)
25 #define NUM_GPR 16
26 #define NUM_GPR_DW (NUM_GPR * 2) /* each GPR is 2 dwords */
27 
28 static bool is_active(struct i915_request *rq)
29 {
30 	if (i915_request_is_active(rq))
31 		return true;
32 
33 	if (i915_request_on_hold(rq))
34 		return true;
35 
36 	if (i915_request_has_initial_breadcrumb(rq) && i915_request_started(rq))
37 		return true;
38 
39 	return false;
40 }
41 
42 static int wait_for_submit(struct intel_engine_cs *engine,
43 			   struct i915_request *rq,
44 			   unsigned long timeout)
45 {
46 	/* Ignore our own attempts to suppress excess tasklets */
47 	tasklet_hi_schedule(&engine->sched_engine->tasklet);
48 
49 	timeout += jiffies;
50 	do {
51 		bool done = time_after(jiffies, timeout);
52 
53 		if (i915_request_completed(rq)) /* that was quick! */
54 			return 0;
55 
56 		/* Wait until the HW has acknowleged the submission (or err) */
57 		intel_engine_flush_submission(engine);
58 		if (!READ_ONCE(engine->execlists.pending[0]) && is_active(rq))
59 			return 0;
60 
61 		if (done)
62 			return -ETIME;
63 
64 		cond_resched();
65 	} while (1);
66 }
67 
68 static int wait_for_reset(struct intel_engine_cs *engine,
69 			  struct i915_request *rq,
70 			  unsigned long timeout)
71 {
72 	timeout += jiffies;
73 
74 	do {
75 		cond_resched();
76 		intel_engine_flush_submission(engine);
77 
78 		if (READ_ONCE(engine->execlists.pending[0]))
79 			continue;
80 
81 		if (i915_request_completed(rq))
82 			break;
83 
84 		if (READ_ONCE(rq->fence.error))
85 			break;
86 	} while (time_before(jiffies, timeout));
87 
88 	if (rq->fence.error != -EIO) {
89 		pr_err("%s: hanging request %llx:%lld not reset\n",
90 		       engine->name,
91 		       rq->fence.context,
92 		       rq->fence.seqno);
93 		return -EINVAL;
94 	}
95 
96 	/* Give the request a jiffie to complete after flushing the worker */
97 	if (i915_request_wait(rq, 0,
98 			      max(0l, (long)(timeout - jiffies)) + 1) < 0) {
99 		pr_err("%s: hanging request %llx:%lld did not complete\n",
100 		       engine->name,
101 		       rq->fence.context,
102 		       rq->fence.seqno);
103 		return -ETIME;
104 	}
105 
106 	return 0;
107 }
108 
109 static int live_sanitycheck(void *arg)
110 {
111 	struct intel_gt *gt = arg;
112 	struct intel_engine_cs *engine;
113 	enum intel_engine_id id;
114 	struct igt_spinner spin;
115 	int err = 0;
116 
117 	if (!HAS_LOGICAL_RING_CONTEXTS(gt->i915))
118 		return 0;
119 
120 	if (igt_spinner_init(&spin, gt))
121 		return -ENOMEM;
122 
123 	for_each_engine(engine, gt, id) {
124 		struct intel_context *ce;
125 		struct i915_request *rq;
126 
127 		ce = intel_context_create(engine);
128 		if (IS_ERR(ce)) {
129 			err = PTR_ERR(ce);
130 			break;
131 		}
132 
133 		rq = igt_spinner_create_request(&spin, ce, MI_NOOP);
134 		if (IS_ERR(rq)) {
135 			err = PTR_ERR(rq);
136 			goto out_ctx;
137 		}
138 
139 		i915_request_add(rq);
140 		if (!igt_wait_for_spinner(&spin, rq)) {
141 			GEM_TRACE("spinner failed to start\n");
142 			GEM_TRACE_DUMP();
143 			intel_gt_set_wedged(gt);
144 			err = -EIO;
145 			goto out_ctx;
146 		}
147 
148 		igt_spinner_end(&spin);
149 		if (igt_flush_test(gt->i915)) {
150 			err = -EIO;
151 			goto out_ctx;
152 		}
153 
154 out_ctx:
155 		intel_context_put(ce);
156 		if (err)
157 			break;
158 	}
159 
160 	igt_spinner_fini(&spin);
161 	return err;
162 }
163 
164 static int live_unlite_restore(struct intel_gt *gt, int prio)
165 {
166 	struct intel_engine_cs *engine;
167 	enum intel_engine_id id;
168 	struct igt_spinner spin;
169 	int err = -ENOMEM;
170 
171 	/*
172 	 * Check that we can correctly context switch between 2 instances
173 	 * on the same engine from the same parent context.
174 	 */
175 
176 	if (igt_spinner_init(&spin, gt))
177 		return err;
178 
179 	err = 0;
180 	for_each_engine(engine, gt, id) {
181 		struct intel_context *ce[2] = {};
182 		struct i915_request *rq[2];
183 		struct igt_live_test t;
184 		int n;
185 
186 		if (prio && !intel_engine_has_preemption(engine))
187 			continue;
188 
189 		if (!intel_engine_can_store_dword(engine))
190 			continue;
191 
192 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
193 			err = -EIO;
194 			break;
195 		}
196 		st_engine_heartbeat_disable(engine);
197 
198 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
199 			struct intel_context *tmp;
200 
201 			tmp = intel_context_create(engine);
202 			if (IS_ERR(tmp)) {
203 				err = PTR_ERR(tmp);
204 				goto err_ce;
205 			}
206 
207 			err = intel_context_pin(tmp);
208 			if (err) {
209 				intel_context_put(tmp);
210 				goto err_ce;
211 			}
212 
213 			/*
214 			 * Setup the pair of contexts such that if we
215 			 * lite-restore using the RING_TAIL from ce[1] it
216 			 * will execute garbage from ce[0]->ring.
217 			 */
218 			memset(tmp->ring->vaddr,
219 			       POISON_INUSE, /* IPEHR: 0x5a5a5a5a [hung!] */
220 			       tmp->ring->vma->size);
221 
222 			ce[n] = tmp;
223 		}
224 		GEM_BUG_ON(!ce[1]->ring->size);
225 		intel_ring_reset(ce[1]->ring, ce[1]->ring->size / 2);
226 		lrc_update_regs(ce[1], engine, ce[1]->ring->head);
227 
228 		rq[0] = igt_spinner_create_request(&spin, ce[0], MI_ARB_CHECK);
229 		if (IS_ERR(rq[0])) {
230 			err = PTR_ERR(rq[0]);
231 			goto err_ce;
232 		}
233 
234 		i915_request_get(rq[0]);
235 		i915_request_add(rq[0]);
236 		GEM_BUG_ON(rq[0]->postfix > ce[1]->ring->emit);
237 
238 		if (!igt_wait_for_spinner(&spin, rq[0])) {
239 			i915_request_put(rq[0]);
240 			goto err_ce;
241 		}
242 
243 		rq[1] = i915_request_create(ce[1]);
244 		if (IS_ERR(rq[1])) {
245 			err = PTR_ERR(rq[1]);
246 			i915_request_put(rq[0]);
247 			goto err_ce;
248 		}
249 
250 		if (!prio) {
251 			/*
252 			 * Ensure we do the switch to ce[1] on completion.
253 			 *
254 			 * rq[0] is already submitted, so this should reduce
255 			 * to a no-op (a wait on a request on the same engine
256 			 * uses the submit fence, not the completion fence),
257 			 * but it will install a dependency on rq[1] for rq[0]
258 			 * that will prevent the pair being reordered by
259 			 * timeslicing.
260 			 */
261 			i915_request_await_dma_fence(rq[1], &rq[0]->fence);
262 		}
263 
264 		i915_request_get(rq[1]);
265 		i915_request_add(rq[1]);
266 		GEM_BUG_ON(rq[1]->postfix <= rq[0]->postfix);
267 		i915_request_put(rq[0]);
268 
269 		if (prio) {
270 			struct i915_sched_attr attr = {
271 				.priority = prio,
272 			};
273 
274 			/* Alternatively preempt the spinner with ce[1] */
275 			engine->sched_engine->schedule(rq[1], &attr);
276 		}
277 
278 		/* And switch back to ce[0] for good measure */
279 		rq[0] = i915_request_create(ce[0]);
280 		if (IS_ERR(rq[0])) {
281 			err = PTR_ERR(rq[0]);
282 			i915_request_put(rq[1]);
283 			goto err_ce;
284 		}
285 
286 		i915_request_await_dma_fence(rq[0], &rq[1]->fence);
287 		i915_request_get(rq[0]);
288 		i915_request_add(rq[0]);
289 		GEM_BUG_ON(rq[0]->postfix > rq[1]->postfix);
290 		i915_request_put(rq[1]);
291 		i915_request_put(rq[0]);
292 
293 err_ce:
294 		intel_engine_flush_submission(engine);
295 		igt_spinner_end(&spin);
296 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
297 			if (IS_ERR_OR_NULL(ce[n]))
298 				break;
299 
300 			intel_context_unpin(ce[n]);
301 			intel_context_put(ce[n]);
302 		}
303 
304 		st_engine_heartbeat_enable(engine);
305 		if (igt_live_test_end(&t))
306 			err = -EIO;
307 		if (err)
308 			break;
309 	}
310 
311 	igt_spinner_fini(&spin);
312 	return err;
313 }
314 
315 static int live_unlite_switch(void *arg)
316 {
317 	return live_unlite_restore(arg, 0);
318 }
319 
320 static int live_unlite_preempt(void *arg)
321 {
322 	return live_unlite_restore(arg, I915_PRIORITY_MAX);
323 }
324 
325 static int live_unlite_ring(void *arg)
326 {
327 	struct intel_gt *gt = arg;
328 	struct intel_engine_cs *engine;
329 	struct igt_spinner spin;
330 	enum intel_engine_id id;
331 	int err = 0;
332 
333 	/*
334 	 * Setup a preemption event that will cause almost the entire ring
335 	 * to be unwound, potentially fooling our intel_ring_direction()
336 	 * into emitting a forward lite-restore instead of the rollback.
337 	 */
338 
339 	if (igt_spinner_init(&spin, gt))
340 		return -ENOMEM;
341 
342 	for_each_engine(engine, gt, id) {
343 		struct intel_context *ce[2] = {};
344 		struct i915_request *rq;
345 		struct igt_live_test t;
346 		int n;
347 
348 		if (!intel_engine_has_preemption(engine))
349 			continue;
350 
351 		if (!intel_engine_can_store_dword(engine))
352 			continue;
353 
354 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
355 			err = -EIO;
356 			break;
357 		}
358 		st_engine_heartbeat_disable(engine);
359 
360 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
361 			struct intel_context *tmp;
362 
363 			tmp = intel_context_create(engine);
364 			if (IS_ERR(tmp)) {
365 				err = PTR_ERR(tmp);
366 				goto err_ce;
367 			}
368 
369 			err = intel_context_pin(tmp);
370 			if (err) {
371 				intel_context_put(tmp);
372 				goto err_ce;
373 			}
374 
375 			memset32(tmp->ring->vaddr,
376 				 0xdeadbeef, /* trigger a hang if executed */
377 				 tmp->ring->vma->size / sizeof(u32));
378 
379 			ce[n] = tmp;
380 		}
381 
382 		/* Create max prio spinner, followed by N low prio nops */
383 		rq = igt_spinner_create_request(&spin, ce[0], MI_ARB_CHECK);
384 		if (IS_ERR(rq)) {
385 			err = PTR_ERR(rq);
386 			goto err_ce;
387 		}
388 
389 		i915_request_get(rq);
390 		rq->sched.attr.priority = I915_PRIORITY_BARRIER;
391 		i915_request_add(rq);
392 
393 		if (!igt_wait_for_spinner(&spin, rq)) {
394 			intel_gt_set_wedged(gt);
395 			i915_request_put(rq);
396 			err = -ETIME;
397 			goto err_ce;
398 		}
399 
400 		/* Fill the ring, until we will cause a wrap */
401 		n = 0;
402 		while (intel_ring_direction(ce[0]->ring,
403 					    rq->wa_tail,
404 					    ce[0]->ring->tail) <= 0) {
405 			struct i915_request *tmp;
406 
407 			tmp = intel_context_create_request(ce[0]);
408 			if (IS_ERR(tmp)) {
409 				err = PTR_ERR(tmp);
410 				i915_request_put(rq);
411 				goto err_ce;
412 			}
413 
414 			i915_request_add(tmp);
415 			intel_engine_flush_submission(engine);
416 			n++;
417 		}
418 		intel_engine_flush_submission(engine);
419 		pr_debug("%s: Filled ring with %d nop tails {size:%x, tail:%x, emit:%x, rq.tail:%x}\n",
420 			 engine->name, n,
421 			 ce[0]->ring->size,
422 			 ce[0]->ring->tail,
423 			 ce[0]->ring->emit,
424 			 rq->tail);
425 		GEM_BUG_ON(intel_ring_direction(ce[0]->ring,
426 						rq->tail,
427 						ce[0]->ring->tail) <= 0);
428 		i915_request_put(rq);
429 
430 		/* Create a second ring to preempt the first ring after rq[0] */
431 		rq = intel_context_create_request(ce[1]);
432 		if (IS_ERR(rq)) {
433 			err = PTR_ERR(rq);
434 			goto err_ce;
435 		}
436 
437 		rq->sched.attr.priority = I915_PRIORITY_BARRIER;
438 		i915_request_get(rq);
439 		i915_request_add(rq);
440 
441 		err = wait_for_submit(engine, rq, HZ / 2);
442 		i915_request_put(rq);
443 		if (err) {
444 			pr_err("%s: preemption request was not submitted\n",
445 			       engine->name);
446 			err = -ETIME;
447 		}
448 
449 		pr_debug("%s: ring[0]:{ tail:%x, emit:%x }, ring[1]:{ tail:%x, emit:%x }\n",
450 			 engine->name,
451 			 ce[0]->ring->tail, ce[0]->ring->emit,
452 			 ce[1]->ring->tail, ce[1]->ring->emit);
453 
454 err_ce:
455 		intel_engine_flush_submission(engine);
456 		igt_spinner_end(&spin);
457 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
458 			if (IS_ERR_OR_NULL(ce[n]))
459 				break;
460 
461 			intel_context_unpin(ce[n]);
462 			intel_context_put(ce[n]);
463 		}
464 		st_engine_heartbeat_enable(engine);
465 		if (igt_live_test_end(&t))
466 			err = -EIO;
467 		if (err)
468 			break;
469 	}
470 
471 	igt_spinner_fini(&spin);
472 	return err;
473 }
474 
475 static int live_pin_rewind(void *arg)
476 {
477 	struct intel_gt *gt = arg;
478 	struct intel_engine_cs *engine;
479 	enum intel_engine_id id;
480 	int err = 0;
481 
482 	/*
483 	 * We have to be careful not to trust intel_ring too much, for example
484 	 * ring->head is updated upon retire which is out of sync with pinning
485 	 * the context. Thus we cannot use ring->head to set CTX_RING_HEAD,
486 	 * or else we risk writing an older, stale value.
487 	 *
488 	 * To simulate this, let's apply a bit of deliberate sabotague.
489 	 */
490 
491 	for_each_engine(engine, gt, id) {
492 		struct intel_context *ce;
493 		struct i915_request *rq;
494 		struct intel_ring *ring;
495 		struct igt_live_test t;
496 
497 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
498 			err = -EIO;
499 			break;
500 		}
501 
502 		ce = intel_context_create(engine);
503 		if (IS_ERR(ce)) {
504 			err = PTR_ERR(ce);
505 			break;
506 		}
507 
508 		err = intel_context_pin(ce);
509 		if (err) {
510 			intel_context_put(ce);
511 			break;
512 		}
513 
514 		/* Keep the context awake while we play games */
515 		err = i915_active_acquire(&ce->active);
516 		if (err) {
517 			intel_context_unpin(ce);
518 			intel_context_put(ce);
519 			break;
520 		}
521 		ring = ce->ring;
522 
523 		/* Poison the ring, and offset the next request from HEAD */
524 		memset32(ring->vaddr, STACK_MAGIC, ring->size / sizeof(u32));
525 		ring->emit = ring->size / 2;
526 		ring->tail = ring->emit;
527 		GEM_BUG_ON(ring->head);
528 
529 		intel_context_unpin(ce);
530 
531 		/* Submit a simple nop request */
532 		GEM_BUG_ON(intel_context_is_pinned(ce));
533 		rq = intel_context_create_request(ce);
534 		i915_active_release(&ce->active); /* e.g. async retire */
535 		intel_context_put(ce);
536 		if (IS_ERR(rq)) {
537 			err = PTR_ERR(rq);
538 			break;
539 		}
540 		GEM_BUG_ON(!rq->head);
541 		i915_request_add(rq);
542 
543 		/* Expect not to hang! */
544 		if (igt_live_test_end(&t)) {
545 			err = -EIO;
546 			break;
547 		}
548 	}
549 
550 	return err;
551 }
552 
553 static int engine_lock_reset_tasklet(struct intel_engine_cs *engine)
554 {
555 	tasklet_disable(&engine->sched_engine->tasklet);
556 	local_bh_disable();
557 
558 	if (test_and_set_bit(I915_RESET_ENGINE + engine->id,
559 			     &engine->gt->reset.flags)) {
560 		local_bh_enable();
561 		tasklet_enable(&engine->sched_engine->tasklet);
562 
563 		intel_gt_set_wedged(engine->gt);
564 		return -EBUSY;
565 	}
566 
567 	return 0;
568 }
569 
570 static void engine_unlock_reset_tasklet(struct intel_engine_cs *engine)
571 {
572 	clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
573 			      &engine->gt->reset.flags);
574 
575 	local_bh_enable();
576 	tasklet_enable(&engine->sched_engine->tasklet);
577 }
578 
579 static int live_hold_reset(void *arg)
580 {
581 	struct intel_gt *gt = arg;
582 	struct intel_engine_cs *engine;
583 	enum intel_engine_id id;
584 	struct igt_spinner spin;
585 	int err = 0;
586 
587 	/*
588 	 * In order to support offline error capture for fast preempt reset,
589 	 * we need to decouple the guilty request and ensure that it and its
590 	 * descendents are not executed while the capture is in progress.
591 	 */
592 
593 	if (!intel_has_reset_engine(gt))
594 		return 0;
595 
596 	if (igt_spinner_init(&spin, gt))
597 		return -ENOMEM;
598 
599 	for_each_engine(engine, gt, id) {
600 		struct intel_context *ce;
601 		struct i915_request *rq;
602 
603 		ce = intel_context_create(engine);
604 		if (IS_ERR(ce)) {
605 			err = PTR_ERR(ce);
606 			break;
607 		}
608 
609 		st_engine_heartbeat_disable(engine);
610 
611 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
612 		if (IS_ERR(rq)) {
613 			err = PTR_ERR(rq);
614 			goto out;
615 		}
616 		i915_request_add(rq);
617 
618 		if (!igt_wait_for_spinner(&spin, rq)) {
619 			intel_gt_set_wedged(gt);
620 			err = -ETIME;
621 			goto out;
622 		}
623 
624 		/* We have our request executing, now remove it and reset */
625 
626 		err = engine_lock_reset_tasklet(engine);
627 		if (err)
628 			goto out;
629 
630 		engine->sched_engine->tasklet.callback(&engine->sched_engine->tasklet);
631 		GEM_BUG_ON(execlists_active(&engine->execlists) != rq);
632 
633 		i915_request_get(rq);
634 		execlists_hold(engine, rq);
635 		GEM_BUG_ON(!i915_request_on_hold(rq));
636 
637 		__intel_engine_reset_bh(engine, NULL);
638 		GEM_BUG_ON(rq->fence.error != -EIO);
639 
640 		engine_unlock_reset_tasklet(engine);
641 
642 		/* Check that we do not resubmit the held request */
643 		if (!i915_request_wait(rq, 0, HZ / 5)) {
644 			pr_err("%s: on hold request completed!\n",
645 			       engine->name);
646 			i915_request_put(rq);
647 			err = -EIO;
648 			goto out;
649 		}
650 		GEM_BUG_ON(!i915_request_on_hold(rq));
651 
652 		/* But is resubmitted on release */
653 		execlists_unhold(engine, rq);
654 		if (i915_request_wait(rq, 0, HZ / 5) < 0) {
655 			pr_err("%s: held request did not complete!\n",
656 			       engine->name);
657 			intel_gt_set_wedged(gt);
658 			err = -ETIME;
659 		}
660 		i915_request_put(rq);
661 
662 out:
663 		st_engine_heartbeat_enable(engine);
664 		intel_context_put(ce);
665 		if (err)
666 			break;
667 	}
668 
669 	igt_spinner_fini(&spin);
670 	return err;
671 }
672 
673 static const char *error_repr(int err)
674 {
675 	return err ? "bad" : "good";
676 }
677 
678 static int live_error_interrupt(void *arg)
679 {
680 	static const struct error_phase {
681 		enum { GOOD = 0, BAD = -EIO } error[2];
682 	} phases[] = {
683 		{ { BAD,  GOOD } },
684 		{ { BAD,  BAD  } },
685 		{ { BAD,  GOOD } },
686 		{ { GOOD, GOOD } }, /* sentinel */
687 	};
688 	struct intel_gt *gt = arg;
689 	struct intel_engine_cs *engine;
690 	enum intel_engine_id id;
691 
692 	/*
693 	 * We hook up the CS_MASTER_ERROR_INTERRUPT to have forewarning
694 	 * of invalid commands in user batches that will cause a GPU hang.
695 	 * This is a faster mechanism than using hangcheck/heartbeats, but
696 	 * only detects problems the HW knows about -- it will not warn when
697 	 * we kill the HW!
698 	 *
699 	 * To verify our detection and reset, we throw some invalid commands
700 	 * at the HW and wait for the interrupt.
701 	 */
702 
703 	if (!intel_has_reset_engine(gt))
704 		return 0;
705 
706 	for_each_engine(engine, gt, id) {
707 		const struct error_phase *p;
708 		int err = 0;
709 
710 		st_engine_heartbeat_disable(engine);
711 
712 		for (p = phases; p->error[0] != GOOD; p++) {
713 			struct i915_request *client[ARRAY_SIZE(phases->error)];
714 			u32 *cs;
715 			int i;
716 
717 			memset(client, 0, sizeof(*client));
718 			for (i = 0; i < ARRAY_SIZE(client); i++) {
719 				struct intel_context *ce;
720 				struct i915_request *rq;
721 
722 				ce = intel_context_create(engine);
723 				if (IS_ERR(ce)) {
724 					err = PTR_ERR(ce);
725 					goto out;
726 				}
727 
728 				rq = intel_context_create_request(ce);
729 				intel_context_put(ce);
730 				if (IS_ERR(rq)) {
731 					err = PTR_ERR(rq);
732 					goto out;
733 				}
734 
735 				if (rq->engine->emit_init_breadcrumb) {
736 					err = rq->engine->emit_init_breadcrumb(rq);
737 					if (err) {
738 						i915_request_add(rq);
739 						goto out;
740 					}
741 				}
742 
743 				cs = intel_ring_begin(rq, 2);
744 				if (IS_ERR(cs)) {
745 					i915_request_add(rq);
746 					err = PTR_ERR(cs);
747 					goto out;
748 				}
749 
750 				if (p->error[i]) {
751 					*cs++ = 0xdeadbeef;
752 					*cs++ = 0xdeadbeef;
753 				} else {
754 					*cs++ = MI_NOOP;
755 					*cs++ = MI_NOOP;
756 				}
757 
758 				client[i] = i915_request_get(rq);
759 				i915_request_add(rq);
760 			}
761 
762 			err = wait_for_submit(engine, client[0], HZ / 2);
763 			if (err) {
764 				pr_err("%s: first request did not start within time!\n",
765 				       engine->name);
766 				err = -ETIME;
767 				goto out;
768 			}
769 
770 			for (i = 0; i < ARRAY_SIZE(client); i++) {
771 				if (i915_request_wait(client[i], 0, HZ / 5) < 0)
772 					pr_debug("%s: %s request incomplete!\n",
773 						 engine->name,
774 						 error_repr(p->error[i]));
775 
776 				if (!i915_request_started(client[i])) {
777 					pr_err("%s: %s request not started!\n",
778 					       engine->name,
779 					       error_repr(p->error[i]));
780 					err = -ETIME;
781 					goto out;
782 				}
783 
784 				/* Kick the tasklet to process the error */
785 				intel_engine_flush_submission(engine);
786 				if (client[i]->fence.error != p->error[i]) {
787 					pr_err("%s: %s request (%s) with wrong error code: %d\n",
788 					       engine->name,
789 					       error_repr(p->error[i]),
790 					       i915_request_completed(client[i]) ? "completed" : "running",
791 					       client[i]->fence.error);
792 					err = -EINVAL;
793 					goto out;
794 				}
795 			}
796 
797 out:
798 			for (i = 0; i < ARRAY_SIZE(client); i++)
799 				if (client[i])
800 					i915_request_put(client[i]);
801 			if (err) {
802 				pr_err("%s: failed at phase[%zd] { %d, %d }\n",
803 				       engine->name, p - phases,
804 				       p->error[0], p->error[1]);
805 				break;
806 			}
807 		}
808 
809 		st_engine_heartbeat_enable(engine);
810 		if (err) {
811 			intel_gt_set_wedged(gt);
812 			return err;
813 		}
814 	}
815 
816 	return 0;
817 }
818 
819 static int
820 emit_semaphore_chain(struct i915_request *rq, struct i915_vma *vma, int idx)
821 {
822 	u32 *cs;
823 
824 	cs = intel_ring_begin(rq, 10);
825 	if (IS_ERR(cs))
826 		return PTR_ERR(cs);
827 
828 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
829 
830 	*cs++ = MI_SEMAPHORE_WAIT |
831 		MI_SEMAPHORE_GLOBAL_GTT |
832 		MI_SEMAPHORE_POLL |
833 		MI_SEMAPHORE_SAD_NEQ_SDD;
834 	*cs++ = 0;
835 	*cs++ = i915_ggtt_offset(vma) + 4 * idx;
836 	*cs++ = 0;
837 
838 	if (idx > 0) {
839 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
840 		*cs++ = i915_ggtt_offset(vma) + 4 * (idx - 1);
841 		*cs++ = 0;
842 		*cs++ = 1;
843 	} else {
844 		*cs++ = MI_NOOP;
845 		*cs++ = MI_NOOP;
846 		*cs++ = MI_NOOP;
847 		*cs++ = MI_NOOP;
848 	}
849 
850 	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
851 
852 	intel_ring_advance(rq, cs);
853 	return 0;
854 }
855 
856 static struct i915_request *
857 semaphore_queue(struct intel_engine_cs *engine, struct i915_vma *vma, int idx)
858 {
859 	struct intel_context *ce;
860 	struct i915_request *rq;
861 	int err;
862 
863 	ce = intel_context_create(engine);
864 	if (IS_ERR(ce))
865 		return ERR_CAST(ce);
866 
867 	rq = intel_context_create_request(ce);
868 	if (IS_ERR(rq))
869 		goto out_ce;
870 
871 	err = 0;
872 	if (rq->engine->emit_init_breadcrumb)
873 		err = rq->engine->emit_init_breadcrumb(rq);
874 	if (err == 0)
875 		err = emit_semaphore_chain(rq, vma, idx);
876 	if (err == 0)
877 		i915_request_get(rq);
878 	i915_request_add(rq);
879 	if (err)
880 		rq = ERR_PTR(err);
881 
882 out_ce:
883 	intel_context_put(ce);
884 	return rq;
885 }
886 
887 static int
888 release_queue(struct intel_engine_cs *engine,
889 	      struct i915_vma *vma,
890 	      int idx, int prio)
891 {
892 	struct i915_sched_attr attr = {
893 		.priority = prio,
894 	};
895 	struct i915_request *rq;
896 	u32 *cs;
897 
898 	rq = intel_engine_create_kernel_request(engine);
899 	if (IS_ERR(rq))
900 		return PTR_ERR(rq);
901 
902 	cs = intel_ring_begin(rq, 4);
903 	if (IS_ERR(cs)) {
904 		i915_request_add(rq);
905 		return PTR_ERR(cs);
906 	}
907 
908 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
909 	*cs++ = i915_ggtt_offset(vma) + 4 * (idx - 1);
910 	*cs++ = 0;
911 	*cs++ = 1;
912 
913 	intel_ring_advance(rq, cs);
914 
915 	i915_request_get(rq);
916 	i915_request_add(rq);
917 
918 	local_bh_disable();
919 	engine->sched_engine->schedule(rq, &attr);
920 	local_bh_enable(); /* kick tasklet */
921 
922 	i915_request_put(rq);
923 
924 	return 0;
925 }
926 
927 static int
928 slice_semaphore_queue(struct intel_engine_cs *outer,
929 		      struct i915_vma *vma,
930 		      int count)
931 {
932 	struct intel_engine_cs *engine;
933 	struct i915_request *head;
934 	enum intel_engine_id id;
935 	int err, i, n = 0;
936 
937 	head = semaphore_queue(outer, vma, n++);
938 	if (IS_ERR(head))
939 		return PTR_ERR(head);
940 
941 	for_each_engine(engine, outer->gt, id) {
942 		if (!intel_engine_has_preemption(engine))
943 			continue;
944 
945 		for (i = 0; i < count; i++) {
946 			struct i915_request *rq;
947 
948 			rq = semaphore_queue(engine, vma, n++);
949 			if (IS_ERR(rq)) {
950 				err = PTR_ERR(rq);
951 				goto out;
952 			}
953 
954 			i915_request_put(rq);
955 		}
956 	}
957 
958 	err = release_queue(outer, vma, n, I915_PRIORITY_BARRIER);
959 	if (err)
960 		goto out;
961 
962 	if (i915_request_wait(head, 0,
963 			      2 * outer->gt->info.num_engines * (count + 2) * (count + 3)) < 0) {
964 		pr_err("%s: Failed to slice along semaphore chain of length (%d, %d)!\n",
965 		       outer->name, count, n);
966 		GEM_TRACE_DUMP();
967 		intel_gt_set_wedged(outer->gt);
968 		err = -EIO;
969 	}
970 
971 out:
972 	i915_request_put(head);
973 	return err;
974 }
975 
976 static int live_timeslice_preempt(void *arg)
977 {
978 	struct intel_gt *gt = arg;
979 	struct drm_i915_gem_object *obj;
980 	struct intel_engine_cs *engine;
981 	enum intel_engine_id id;
982 	struct i915_vma *vma;
983 	void *vaddr;
984 	int err = 0;
985 
986 	/*
987 	 * If a request takes too long, we would like to give other users
988 	 * a fair go on the GPU. In particular, users may create batches
989 	 * that wait upon external input, where that input may even be
990 	 * supplied by another GPU job. To avoid blocking forever, we
991 	 * need to preempt the current task and replace it with another
992 	 * ready task.
993 	 */
994 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
995 		return 0;
996 
997 	obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
998 	if (IS_ERR(obj))
999 		return PTR_ERR(obj);
1000 
1001 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
1002 	if (IS_ERR(vma)) {
1003 		err = PTR_ERR(vma);
1004 		goto err_obj;
1005 	}
1006 
1007 	vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1008 	if (IS_ERR(vaddr)) {
1009 		err = PTR_ERR(vaddr);
1010 		goto err_obj;
1011 	}
1012 
1013 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1014 	if (err)
1015 		goto err_map;
1016 
1017 	err = i915_vma_sync(vma);
1018 	if (err)
1019 		goto err_pin;
1020 
1021 	for_each_engine(engine, gt, id) {
1022 		if (!intel_engine_has_preemption(engine))
1023 			continue;
1024 
1025 		memset(vaddr, 0, PAGE_SIZE);
1026 
1027 		st_engine_heartbeat_disable(engine);
1028 		err = slice_semaphore_queue(engine, vma, 5);
1029 		st_engine_heartbeat_enable(engine);
1030 		if (err)
1031 			goto err_pin;
1032 
1033 		if (igt_flush_test(gt->i915)) {
1034 			err = -EIO;
1035 			goto err_pin;
1036 		}
1037 	}
1038 
1039 err_pin:
1040 	i915_vma_unpin(vma);
1041 err_map:
1042 	i915_gem_object_unpin_map(obj);
1043 err_obj:
1044 	i915_gem_object_put(obj);
1045 	return err;
1046 }
1047 
1048 static struct i915_request *
1049 create_rewinder(struct intel_context *ce,
1050 		struct i915_request *wait,
1051 		void *slot, int idx)
1052 {
1053 	const u32 offset =
1054 		i915_ggtt_offset(ce->engine->status_page.vma) +
1055 		offset_in_page(slot);
1056 	struct i915_request *rq;
1057 	u32 *cs;
1058 	int err;
1059 
1060 	rq = intel_context_create_request(ce);
1061 	if (IS_ERR(rq))
1062 		return rq;
1063 
1064 	if (wait) {
1065 		err = i915_request_await_dma_fence(rq, &wait->fence);
1066 		if (err)
1067 			goto err;
1068 	}
1069 
1070 	cs = intel_ring_begin(rq, 14);
1071 	if (IS_ERR(cs)) {
1072 		err = PTR_ERR(cs);
1073 		goto err;
1074 	}
1075 
1076 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1077 	*cs++ = MI_NOOP;
1078 
1079 	*cs++ = MI_SEMAPHORE_WAIT |
1080 		MI_SEMAPHORE_GLOBAL_GTT |
1081 		MI_SEMAPHORE_POLL |
1082 		MI_SEMAPHORE_SAD_GTE_SDD;
1083 	*cs++ = idx;
1084 	*cs++ = offset;
1085 	*cs++ = 0;
1086 
1087 	*cs++ = MI_STORE_REGISTER_MEM_GEN8 | MI_USE_GGTT;
1088 	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(rq->engine->mmio_base));
1089 	*cs++ = offset + idx * sizeof(u32);
1090 	*cs++ = 0;
1091 
1092 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1093 	*cs++ = offset;
1094 	*cs++ = 0;
1095 	*cs++ = idx + 1;
1096 
1097 	intel_ring_advance(rq, cs);
1098 
1099 	err = 0;
1100 err:
1101 	i915_request_get(rq);
1102 	i915_request_add(rq);
1103 	if (err) {
1104 		i915_request_put(rq);
1105 		return ERR_PTR(err);
1106 	}
1107 
1108 	return rq;
1109 }
1110 
1111 static int live_timeslice_rewind(void *arg)
1112 {
1113 	struct intel_gt *gt = arg;
1114 	struct intel_engine_cs *engine;
1115 	enum intel_engine_id id;
1116 
1117 	/*
1118 	 * The usual presumption on timeslice expiration is that we replace
1119 	 * the active context with another. However, given a chain of
1120 	 * dependencies we may end up with replacing the context with itself,
1121 	 * but only a few of those requests, forcing us to rewind the
1122 	 * RING_TAIL of the original request.
1123 	 */
1124 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
1125 		return 0;
1126 
1127 	for_each_engine(engine, gt, id) {
1128 		enum { A1, A2, B1 };
1129 		enum { X = 1, Z, Y };
1130 		struct i915_request *rq[3] = {};
1131 		struct intel_context *ce;
1132 		unsigned long timeslice;
1133 		int i, err = 0;
1134 		u32 *slot;
1135 
1136 		if (!intel_engine_has_timeslices(engine))
1137 			continue;
1138 
1139 		/*
1140 		 * A:rq1 -- semaphore wait, timestamp X
1141 		 * A:rq2 -- write timestamp Y
1142 		 *
1143 		 * B:rq1 [await A:rq1] -- write timestamp Z
1144 		 *
1145 		 * Force timeslice, release semaphore.
1146 		 *
1147 		 * Expect execution/evaluation order XZY
1148 		 */
1149 
1150 		st_engine_heartbeat_disable(engine);
1151 		timeslice = xchg(&engine->props.timeslice_duration_ms, 1);
1152 
1153 		slot = memset32(engine->status_page.addr + 1000, 0, 4);
1154 
1155 		ce = intel_context_create(engine);
1156 		if (IS_ERR(ce)) {
1157 			err = PTR_ERR(ce);
1158 			goto err;
1159 		}
1160 
1161 		rq[A1] = create_rewinder(ce, NULL, slot, X);
1162 		if (IS_ERR(rq[A1])) {
1163 			intel_context_put(ce);
1164 			goto err;
1165 		}
1166 
1167 		rq[A2] = create_rewinder(ce, NULL, slot, Y);
1168 		intel_context_put(ce);
1169 		if (IS_ERR(rq[A2]))
1170 			goto err;
1171 
1172 		err = wait_for_submit(engine, rq[A2], HZ / 2);
1173 		if (err) {
1174 			pr_err("%s: failed to submit first context\n",
1175 			       engine->name);
1176 			goto err;
1177 		}
1178 
1179 		ce = intel_context_create(engine);
1180 		if (IS_ERR(ce)) {
1181 			err = PTR_ERR(ce);
1182 			goto err;
1183 		}
1184 
1185 		rq[B1] = create_rewinder(ce, rq[A1], slot, Z);
1186 		intel_context_put(ce);
1187 		if (IS_ERR(rq[2]))
1188 			goto err;
1189 
1190 		err = wait_for_submit(engine, rq[B1], HZ / 2);
1191 		if (err) {
1192 			pr_err("%s: failed to submit second context\n",
1193 			       engine->name);
1194 			goto err;
1195 		}
1196 
1197 		/* ELSP[] = { { A:rq1, A:rq2 }, { B:rq1 } } */
1198 		ENGINE_TRACE(engine, "forcing tasklet for rewind\n");
1199 		while (i915_request_is_active(rq[A2])) { /* semaphore yield! */
1200 			/* Wait for the timeslice to kick in */
1201 			del_timer(&engine->execlists.timer);
1202 			tasklet_hi_schedule(&engine->sched_engine->tasklet);
1203 			intel_engine_flush_submission(engine);
1204 		}
1205 		/* -> ELSP[] = { { A:rq1 }, { B:rq1 } } */
1206 		GEM_BUG_ON(!i915_request_is_active(rq[A1]));
1207 		GEM_BUG_ON(!i915_request_is_active(rq[B1]));
1208 		GEM_BUG_ON(i915_request_is_active(rq[A2]));
1209 
1210 		/* Release the hounds! */
1211 		slot[0] = 1;
1212 		wmb(); /* "pairs" with GPU; paranoid kick of internal CPU$ */
1213 
1214 		for (i = 1; i <= 3; i++) {
1215 			unsigned long timeout = jiffies + HZ / 2;
1216 
1217 			while (!READ_ONCE(slot[i]) &&
1218 			       time_before(jiffies, timeout))
1219 				;
1220 
1221 			if (!time_before(jiffies, timeout)) {
1222 				pr_err("%s: rq[%d] timed out\n",
1223 				       engine->name, i - 1);
1224 				err = -ETIME;
1225 				goto err;
1226 			}
1227 
1228 			pr_debug("%s: slot[%d]:%x\n", engine->name, i, slot[i]);
1229 		}
1230 
1231 		/* XZY: XZ < XY */
1232 		if (slot[Z] - slot[X] >= slot[Y] - slot[X]) {
1233 			pr_err("%s: timeslicing did not run context B [%u] before A [%u]!\n",
1234 			       engine->name,
1235 			       slot[Z] - slot[X],
1236 			       slot[Y] - slot[X]);
1237 			err = -EINVAL;
1238 		}
1239 
1240 err:
1241 		memset32(&slot[0], -1, 4);
1242 		wmb();
1243 
1244 		engine->props.timeslice_duration_ms = timeslice;
1245 		st_engine_heartbeat_enable(engine);
1246 		for (i = 0; i < 3; i++)
1247 			i915_request_put(rq[i]);
1248 		if (igt_flush_test(gt->i915))
1249 			err = -EIO;
1250 		if (err)
1251 			return err;
1252 	}
1253 
1254 	return 0;
1255 }
1256 
1257 static struct i915_request *nop_request(struct intel_engine_cs *engine)
1258 {
1259 	struct i915_request *rq;
1260 
1261 	rq = intel_engine_create_kernel_request(engine);
1262 	if (IS_ERR(rq))
1263 		return rq;
1264 
1265 	i915_request_get(rq);
1266 	i915_request_add(rq);
1267 
1268 	return rq;
1269 }
1270 
1271 static long slice_timeout(struct intel_engine_cs *engine)
1272 {
1273 	long timeout;
1274 
1275 	/* Enough time for a timeslice to kick in, and kick out */
1276 	timeout = 2 * msecs_to_jiffies_timeout(timeslice(engine));
1277 
1278 	/* Enough time for the nop request to complete */
1279 	timeout += HZ / 5;
1280 
1281 	return timeout + 1;
1282 }
1283 
1284 static int live_timeslice_queue(void *arg)
1285 {
1286 	struct intel_gt *gt = arg;
1287 	struct drm_i915_gem_object *obj;
1288 	struct intel_engine_cs *engine;
1289 	enum intel_engine_id id;
1290 	struct i915_vma *vma;
1291 	void *vaddr;
1292 	int err = 0;
1293 
1294 	/*
1295 	 * Make sure that even if ELSP[0] and ELSP[1] are filled with
1296 	 * timeslicing between them disabled, we *do* enable timeslicing
1297 	 * if the queue demands it. (Normally, we do not submit if
1298 	 * ELSP[1] is already occupied, so must rely on timeslicing to
1299 	 * eject ELSP[0] in favour of the queue.)
1300 	 */
1301 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
1302 		return 0;
1303 
1304 	obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
1305 	if (IS_ERR(obj))
1306 		return PTR_ERR(obj);
1307 
1308 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
1309 	if (IS_ERR(vma)) {
1310 		err = PTR_ERR(vma);
1311 		goto err_obj;
1312 	}
1313 
1314 	vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1315 	if (IS_ERR(vaddr)) {
1316 		err = PTR_ERR(vaddr);
1317 		goto err_obj;
1318 	}
1319 
1320 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1321 	if (err)
1322 		goto err_map;
1323 
1324 	err = i915_vma_sync(vma);
1325 	if (err)
1326 		goto err_pin;
1327 
1328 	for_each_engine(engine, gt, id) {
1329 		struct i915_sched_attr attr = { .priority = I915_PRIORITY_MAX };
1330 		struct i915_request *rq, *nop;
1331 
1332 		if (!intel_engine_has_preemption(engine))
1333 			continue;
1334 
1335 		st_engine_heartbeat_disable(engine);
1336 		memset(vaddr, 0, PAGE_SIZE);
1337 
1338 		/* ELSP[0]: semaphore wait */
1339 		rq = semaphore_queue(engine, vma, 0);
1340 		if (IS_ERR(rq)) {
1341 			err = PTR_ERR(rq);
1342 			goto err_heartbeat;
1343 		}
1344 		engine->sched_engine->schedule(rq, &attr);
1345 		err = wait_for_submit(engine, rq, HZ / 2);
1346 		if (err) {
1347 			pr_err("%s: Timed out trying to submit semaphores\n",
1348 			       engine->name);
1349 			goto err_rq;
1350 		}
1351 
1352 		/* ELSP[1]: nop request */
1353 		nop = nop_request(engine);
1354 		if (IS_ERR(nop)) {
1355 			err = PTR_ERR(nop);
1356 			goto err_rq;
1357 		}
1358 		err = wait_for_submit(engine, nop, HZ / 2);
1359 		i915_request_put(nop);
1360 		if (err) {
1361 			pr_err("%s: Timed out trying to submit nop\n",
1362 			       engine->name);
1363 			goto err_rq;
1364 		}
1365 
1366 		GEM_BUG_ON(i915_request_completed(rq));
1367 		GEM_BUG_ON(execlists_active(&engine->execlists) != rq);
1368 
1369 		/* Queue: semaphore signal, matching priority as semaphore */
1370 		err = release_queue(engine, vma, 1, effective_prio(rq));
1371 		if (err)
1372 			goto err_rq;
1373 
1374 		/* Wait until we ack the release_queue and start timeslicing */
1375 		do {
1376 			cond_resched();
1377 			intel_engine_flush_submission(engine);
1378 		} while (READ_ONCE(engine->execlists.pending[0]));
1379 
1380 		/* Timeslice every jiffy, so within 2 we should signal */
1381 		if (i915_request_wait(rq, 0, slice_timeout(engine)) < 0) {
1382 			struct drm_printer p =
1383 				drm_info_printer(gt->i915->drm.dev);
1384 
1385 			pr_err("%s: Failed to timeslice into queue\n",
1386 			       engine->name);
1387 			intel_engine_dump(engine, &p,
1388 					  "%s\n", engine->name);
1389 
1390 			memset(vaddr, 0xff, PAGE_SIZE);
1391 			err = -EIO;
1392 		}
1393 err_rq:
1394 		i915_request_put(rq);
1395 err_heartbeat:
1396 		st_engine_heartbeat_enable(engine);
1397 		if (err)
1398 			break;
1399 	}
1400 
1401 err_pin:
1402 	i915_vma_unpin(vma);
1403 err_map:
1404 	i915_gem_object_unpin_map(obj);
1405 err_obj:
1406 	i915_gem_object_put(obj);
1407 	return err;
1408 }
1409 
1410 static int live_timeslice_nopreempt(void *arg)
1411 {
1412 	struct intel_gt *gt = arg;
1413 	struct intel_engine_cs *engine;
1414 	enum intel_engine_id id;
1415 	struct igt_spinner spin;
1416 	int err = 0;
1417 
1418 	/*
1419 	 * We should not timeslice into a request that is marked with
1420 	 * I915_REQUEST_NOPREEMPT.
1421 	 */
1422 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
1423 		return 0;
1424 
1425 	if (igt_spinner_init(&spin, gt))
1426 		return -ENOMEM;
1427 
1428 	for_each_engine(engine, gt, id) {
1429 		struct intel_context *ce;
1430 		struct i915_request *rq;
1431 		unsigned long timeslice;
1432 
1433 		if (!intel_engine_has_preemption(engine))
1434 			continue;
1435 
1436 		ce = intel_context_create(engine);
1437 		if (IS_ERR(ce)) {
1438 			err = PTR_ERR(ce);
1439 			break;
1440 		}
1441 
1442 		st_engine_heartbeat_disable(engine);
1443 		timeslice = xchg(&engine->props.timeslice_duration_ms, 1);
1444 
1445 		/* Create an unpreemptible spinner */
1446 
1447 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
1448 		intel_context_put(ce);
1449 		if (IS_ERR(rq)) {
1450 			err = PTR_ERR(rq);
1451 			goto out_heartbeat;
1452 		}
1453 
1454 		i915_request_get(rq);
1455 		i915_request_add(rq);
1456 
1457 		if (!igt_wait_for_spinner(&spin, rq)) {
1458 			i915_request_put(rq);
1459 			err = -ETIME;
1460 			goto out_spin;
1461 		}
1462 
1463 		set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags);
1464 		i915_request_put(rq);
1465 
1466 		/* Followed by a maximum priority barrier (heartbeat) */
1467 
1468 		ce = intel_context_create(engine);
1469 		if (IS_ERR(ce)) {
1470 			err = PTR_ERR(ce);
1471 			goto out_spin;
1472 		}
1473 
1474 		rq = intel_context_create_request(ce);
1475 		intel_context_put(ce);
1476 		if (IS_ERR(rq)) {
1477 			err = PTR_ERR(rq);
1478 			goto out_spin;
1479 		}
1480 
1481 		rq->sched.attr.priority = I915_PRIORITY_BARRIER;
1482 		i915_request_get(rq);
1483 		i915_request_add(rq);
1484 
1485 		/*
1486 		 * Wait until the barrier is in ELSP, and we know timeslicing
1487 		 * will have been activated.
1488 		 */
1489 		if (wait_for_submit(engine, rq, HZ / 2)) {
1490 			i915_request_put(rq);
1491 			err = -ETIME;
1492 			goto out_spin;
1493 		}
1494 
1495 		/*
1496 		 * Since the ELSP[0] request is unpreemptible, it should not
1497 		 * allow the maximum priority barrier through. Wait long
1498 		 * enough to see if it is timesliced in by mistake.
1499 		 */
1500 		if (i915_request_wait(rq, 0, slice_timeout(engine)) >= 0) {
1501 			pr_err("%s: I915_PRIORITY_BARRIER request completed, bypassing no-preempt request\n",
1502 			       engine->name);
1503 			err = -EINVAL;
1504 		}
1505 		i915_request_put(rq);
1506 
1507 out_spin:
1508 		igt_spinner_end(&spin);
1509 out_heartbeat:
1510 		xchg(&engine->props.timeslice_duration_ms, timeslice);
1511 		st_engine_heartbeat_enable(engine);
1512 		if (err)
1513 			break;
1514 
1515 		if (igt_flush_test(gt->i915)) {
1516 			err = -EIO;
1517 			break;
1518 		}
1519 	}
1520 
1521 	igt_spinner_fini(&spin);
1522 	return err;
1523 }
1524 
1525 static int live_busywait_preempt(void *arg)
1526 {
1527 	struct intel_gt *gt = arg;
1528 	struct i915_gem_context *ctx_hi, *ctx_lo;
1529 	struct intel_engine_cs *engine;
1530 	struct drm_i915_gem_object *obj;
1531 	struct i915_vma *vma;
1532 	enum intel_engine_id id;
1533 	int err = -ENOMEM;
1534 	u32 *map;
1535 
1536 	/*
1537 	 * Verify that even without HAS_LOGICAL_RING_PREEMPTION, we can
1538 	 * preempt the busywaits used to synchronise between rings.
1539 	 */
1540 
1541 	ctx_hi = kernel_context(gt->i915, NULL);
1542 	if (!ctx_hi)
1543 		return -ENOMEM;
1544 	ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
1545 
1546 	ctx_lo = kernel_context(gt->i915, NULL);
1547 	if (!ctx_lo)
1548 		goto err_ctx_hi;
1549 	ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
1550 
1551 	obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
1552 	if (IS_ERR(obj)) {
1553 		err = PTR_ERR(obj);
1554 		goto err_ctx_lo;
1555 	}
1556 
1557 	map = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1558 	if (IS_ERR(map)) {
1559 		err = PTR_ERR(map);
1560 		goto err_obj;
1561 	}
1562 
1563 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
1564 	if (IS_ERR(vma)) {
1565 		err = PTR_ERR(vma);
1566 		goto err_map;
1567 	}
1568 
1569 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1570 	if (err)
1571 		goto err_map;
1572 
1573 	err = i915_vma_sync(vma);
1574 	if (err)
1575 		goto err_vma;
1576 
1577 	for_each_engine(engine, gt, id) {
1578 		struct i915_request *lo, *hi;
1579 		struct igt_live_test t;
1580 		u32 *cs;
1581 
1582 		if (!intel_engine_has_preemption(engine))
1583 			continue;
1584 
1585 		if (!intel_engine_can_store_dword(engine))
1586 			continue;
1587 
1588 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
1589 			err = -EIO;
1590 			goto err_vma;
1591 		}
1592 
1593 		/*
1594 		 * We create two requests. The low priority request
1595 		 * busywaits on a semaphore (inside the ringbuffer where
1596 		 * is should be preemptible) and the high priority requests
1597 		 * uses a MI_STORE_DWORD_IMM to update the semaphore value
1598 		 * allowing the first request to complete. If preemption
1599 		 * fails, we hang instead.
1600 		 */
1601 
1602 		lo = igt_request_alloc(ctx_lo, engine);
1603 		if (IS_ERR(lo)) {
1604 			err = PTR_ERR(lo);
1605 			goto err_vma;
1606 		}
1607 
1608 		cs = intel_ring_begin(lo, 8);
1609 		if (IS_ERR(cs)) {
1610 			err = PTR_ERR(cs);
1611 			i915_request_add(lo);
1612 			goto err_vma;
1613 		}
1614 
1615 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1616 		*cs++ = i915_ggtt_offset(vma);
1617 		*cs++ = 0;
1618 		*cs++ = 1;
1619 
1620 		/* XXX Do we need a flush + invalidate here? */
1621 
1622 		*cs++ = MI_SEMAPHORE_WAIT |
1623 			MI_SEMAPHORE_GLOBAL_GTT |
1624 			MI_SEMAPHORE_POLL |
1625 			MI_SEMAPHORE_SAD_EQ_SDD;
1626 		*cs++ = 0;
1627 		*cs++ = i915_ggtt_offset(vma);
1628 		*cs++ = 0;
1629 
1630 		intel_ring_advance(lo, cs);
1631 
1632 		i915_request_get(lo);
1633 		i915_request_add(lo);
1634 
1635 		if (wait_for(READ_ONCE(*map), 10)) {
1636 			i915_request_put(lo);
1637 			err = -ETIMEDOUT;
1638 			goto err_vma;
1639 		}
1640 
1641 		/* Low priority request should be busywaiting now */
1642 		if (i915_request_wait(lo, 0, 1) != -ETIME) {
1643 			i915_request_put(lo);
1644 			pr_err("%s: Busywaiting request did not!\n",
1645 			       engine->name);
1646 			err = -EIO;
1647 			goto err_vma;
1648 		}
1649 
1650 		hi = igt_request_alloc(ctx_hi, engine);
1651 		if (IS_ERR(hi)) {
1652 			err = PTR_ERR(hi);
1653 			i915_request_put(lo);
1654 			goto err_vma;
1655 		}
1656 
1657 		cs = intel_ring_begin(hi, 4);
1658 		if (IS_ERR(cs)) {
1659 			err = PTR_ERR(cs);
1660 			i915_request_add(hi);
1661 			i915_request_put(lo);
1662 			goto err_vma;
1663 		}
1664 
1665 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1666 		*cs++ = i915_ggtt_offset(vma);
1667 		*cs++ = 0;
1668 		*cs++ = 0;
1669 
1670 		intel_ring_advance(hi, cs);
1671 		i915_request_add(hi);
1672 
1673 		if (i915_request_wait(lo, 0, HZ / 5) < 0) {
1674 			struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1675 
1676 			pr_err("%s: Failed to preempt semaphore busywait!\n",
1677 			       engine->name);
1678 
1679 			intel_engine_dump(engine, &p, "%s\n", engine->name);
1680 			GEM_TRACE_DUMP();
1681 
1682 			i915_request_put(lo);
1683 			intel_gt_set_wedged(gt);
1684 			err = -EIO;
1685 			goto err_vma;
1686 		}
1687 		GEM_BUG_ON(READ_ONCE(*map));
1688 		i915_request_put(lo);
1689 
1690 		if (igt_live_test_end(&t)) {
1691 			err = -EIO;
1692 			goto err_vma;
1693 		}
1694 	}
1695 
1696 	err = 0;
1697 err_vma:
1698 	i915_vma_unpin(vma);
1699 err_map:
1700 	i915_gem_object_unpin_map(obj);
1701 err_obj:
1702 	i915_gem_object_put(obj);
1703 err_ctx_lo:
1704 	kernel_context_close(ctx_lo);
1705 err_ctx_hi:
1706 	kernel_context_close(ctx_hi);
1707 	return err;
1708 }
1709 
1710 static struct i915_request *
1711 spinner_create_request(struct igt_spinner *spin,
1712 		       struct i915_gem_context *ctx,
1713 		       struct intel_engine_cs *engine,
1714 		       u32 arb)
1715 {
1716 	struct intel_context *ce;
1717 	struct i915_request *rq;
1718 
1719 	ce = i915_gem_context_get_engine(ctx, engine->legacy_idx);
1720 	if (IS_ERR(ce))
1721 		return ERR_CAST(ce);
1722 
1723 	rq = igt_spinner_create_request(spin, ce, arb);
1724 	intel_context_put(ce);
1725 	return rq;
1726 }
1727 
1728 static int live_preempt(void *arg)
1729 {
1730 	struct intel_gt *gt = arg;
1731 	struct i915_gem_context *ctx_hi, *ctx_lo;
1732 	struct igt_spinner spin_hi, spin_lo;
1733 	struct intel_engine_cs *engine;
1734 	enum intel_engine_id id;
1735 	int err = -ENOMEM;
1736 
1737 	ctx_hi = kernel_context(gt->i915, NULL);
1738 	if (!ctx_hi)
1739 		return -ENOMEM;
1740 	ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
1741 
1742 	ctx_lo = kernel_context(gt->i915, NULL);
1743 	if (!ctx_lo)
1744 		goto err_ctx_hi;
1745 	ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
1746 
1747 	if (igt_spinner_init(&spin_hi, gt))
1748 		goto err_ctx_lo;
1749 
1750 	if (igt_spinner_init(&spin_lo, gt))
1751 		goto err_spin_hi;
1752 
1753 	for_each_engine(engine, gt, id) {
1754 		struct igt_live_test t;
1755 		struct i915_request *rq;
1756 
1757 		if (!intel_engine_has_preemption(engine))
1758 			continue;
1759 
1760 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
1761 			err = -EIO;
1762 			goto err_spin_lo;
1763 		}
1764 
1765 		rq = spinner_create_request(&spin_lo, ctx_lo, engine,
1766 					    MI_ARB_CHECK);
1767 		if (IS_ERR(rq)) {
1768 			err = PTR_ERR(rq);
1769 			goto err_spin_lo;
1770 		}
1771 
1772 		i915_request_add(rq);
1773 		if (!igt_wait_for_spinner(&spin_lo, rq)) {
1774 			GEM_TRACE("lo spinner failed to start\n");
1775 			GEM_TRACE_DUMP();
1776 			intel_gt_set_wedged(gt);
1777 			err = -EIO;
1778 			goto err_spin_lo;
1779 		}
1780 
1781 		rq = spinner_create_request(&spin_hi, ctx_hi, engine,
1782 					    MI_ARB_CHECK);
1783 		if (IS_ERR(rq)) {
1784 			igt_spinner_end(&spin_lo);
1785 			err = PTR_ERR(rq);
1786 			goto err_spin_lo;
1787 		}
1788 
1789 		i915_request_add(rq);
1790 		if (!igt_wait_for_spinner(&spin_hi, rq)) {
1791 			GEM_TRACE("hi spinner failed to start\n");
1792 			GEM_TRACE_DUMP();
1793 			intel_gt_set_wedged(gt);
1794 			err = -EIO;
1795 			goto err_spin_lo;
1796 		}
1797 
1798 		igt_spinner_end(&spin_hi);
1799 		igt_spinner_end(&spin_lo);
1800 
1801 		if (igt_live_test_end(&t)) {
1802 			err = -EIO;
1803 			goto err_spin_lo;
1804 		}
1805 	}
1806 
1807 	err = 0;
1808 err_spin_lo:
1809 	igt_spinner_fini(&spin_lo);
1810 err_spin_hi:
1811 	igt_spinner_fini(&spin_hi);
1812 err_ctx_lo:
1813 	kernel_context_close(ctx_lo);
1814 err_ctx_hi:
1815 	kernel_context_close(ctx_hi);
1816 	return err;
1817 }
1818 
1819 static int live_late_preempt(void *arg)
1820 {
1821 	struct intel_gt *gt = arg;
1822 	struct i915_gem_context *ctx_hi, *ctx_lo;
1823 	struct igt_spinner spin_hi, spin_lo;
1824 	struct intel_engine_cs *engine;
1825 	struct i915_sched_attr attr = {};
1826 	enum intel_engine_id id;
1827 	int err = -ENOMEM;
1828 
1829 	ctx_hi = kernel_context(gt->i915, NULL);
1830 	if (!ctx_hi)
1831 		return -ENOMEM;
1832 
1833 	ctx_lo = kernel_context(gt->i915, NULL);
1834 	if (!ctx_lo)
1835 		goto err_ctx_hi;
1836 
1837 	if (igt_spinner_init(&spin_hi, gt))
1838 		goto err_ctx_lo;
1839 
1840 	if (igt_spinner_init(&spin_lo, gt))
1841 		goto err_spin_hi;
1842 
1843 	/* Make sure ctx_lo stays before ctx_hi until we trigger preemption. */
1844 	ctx_lo->sched.priority = 1;
1845 
1846 	for_each_engine(engine, gt, id) {
1847 		struct igt_live_test t;
1848 		struct i915_request *rq;
1849 
1850 		if (!intel_engine_has_preemption(engine))
1851 			continue;
1852 
1853 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
1854 			err = -EIO;
1855 			goto err_spin_lo;
1856 		}
1857 
1858 		rq = spinner_create_request(&spin_lo, ctx_lo, engine,
1859 					    MI_ARB_CHECK);
1860 		if (IS_ERR(rq)) {
1861 			err = PTR_ERR(rq);
1862 			goto err_spin_lo;
1863 		}
1864 
1865 		i915_request_add(rq);
1866 		if (!igt_wait_for_spinner(&spin_lo, rq)) {
1867 			pr_err("First context failed to start\n");
1868 			goto err_wedged;
1869 		}
1870 
1871 		rq = spinner_create_request(&spin_hi, ctx_hi, engine,
1872 					    MI_NOOP);
1873 		if (IS_ERR(rq)) {
1874 			igt_spinner_end(&spin_lo);
1875 			err = PTR_ERR(rq);
1876 			goto err_spin_lo;
1877 		}
1878 
1879 		i915_request_add(rq);
1880 		if (igt_wait_for_spinner(&spin_hi, rq)) {
1881 			pr_err("Second context overtook first?\n");
1882 			goto err_wedged;
1883 		}
1884 
1885 		attr.priority = I915_PRIORITY_MAX;
1886 		engine->sched_engine->schedule(rq, &attr);
1887 
1888 		if (!igt_wait_for_spinner(&spin_hi, rq)) {
1889 			pr_err("High priority context failed to preempt the low priority context\n");
1890 			GEM_TRACE_DUMP();
1891 			goto err_wedged;
1892 		}
1893 
1894 		igt_spinner_end(&spin_hi);
1895 		igt_spinner_end(&spin_lo);
1896 
1897 		if (igt_live_test_end(&t)) {
1898 			err = -EIO;
1899 			goto err_spin_lo;
1900 		}
1901 	}
1902 
1903 	err = 0;
1904 err_spin_lo:
1905 	igt_spinner_fini(&spin_lo);
1906 err_spin_hi:
1907 	igt_spinner_fini(&spin_hi);
1908 err_ctx_lo:
1909 	kernel_context_close(ctx_lo);
1910 err_ctx_hi:
1911 	kernel_context_close(ctx_hi);
1912 	return err;
1913 
1914 err_wedged:
1915 	igt_spinner_end(&spin_hi);
1916 	igt_spinner_end(&spin_lo);
1917 	intel_gt_set_wedged(gt);
1918 	err = -EIO;
1919 	goto err_spin_lo;
1920 }
1921 
1922 struct preempt_client {
1923 	struct igt_spinner spin;
1924 	struct i915_gem_context *ctx;
1925 };
1926 
1927 static int preempt_client_init(struct intel_gt *gt, struct preempt_client *c)
1928 {
1929 	c->ctx = kernel_context(gt->i915, NULL);
1930 	if (!c->ctx)
1931 		return -ENOMEM;
1932 
1933 	if (igt_spinner_init(&c->spin, gt))
1934 		goto err_ctx;
1935 
1936 	return 0;
1937 
1938 err_ctx:
1939 	kernel_context_close(c->ctx);
1940 	return -ENOMEM;
1941 }
1942 
1943 static void preempt_client_fini(struct preempt_client *c)
1944 {
1945 	igt_spinner_fini(&c->spin);
1946 	kernel_context_close(c->ctx);
1947 }
1948 
1949 static int live_nopreempt(void *arg)
1950 {
1951 	struct intel_gt *gt = arg;
1952 	struct intel_engine_cs *engine;
1953 	struct preempt_client a, b;
1954 	enum intel_engine_id id;
1955 	int err = -ENOMEM;
1956 
1957 	/*
1958 	 * Verify that we can disable preemption for an individual request
1959 	 * that may be being observed and not want to be interrupted.
1960 	 */
1961 
1962 	if (preempt_client_init(gt, &a))
1963 		return -ENOMEM;
1964 	if (preempt_client_init(gt, &b))
1965 		goto err_client_a;
1966 	b.ctx->sched.priority = I915_PRIORITY_MAX;
1967 
1968 	for_each_engine(engine, gt, id) {
1969 		struct i915_request *rq_a, *rq_b;
1970 
1971 		if (!intel_engine_has_preemption(engine))
1972 			continue;
1973 
1974 		engine->execlists.preempt_hang.count = 0;
1975 
1976 		rq_a = spinner_create_request(&a.spin,
1977 					      a.ctx, engine,
1978 					      MI_ARB_CHECK);
1979 		if (IS_ERR(rq_a)) {
1980 			err = PTR_ERR(rq_a);
1981 			goto err_client_b;
1982 		}
1983 
1984 		/* Low priority client, but unpreemptable! */
1985 		__set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq_a->fence.flags);
1986 
1987 		i915_request_add(rq_a);
1988 		if (!igt_wait_for_spinner(&a.spin, rq_a)) {
1989 			pr_err("First client failed to start\n");
1990 			goto err_wedged;
1991 		}
1992 
1993 		rq_b = spinner_create_request(&b.spin,
1994 					      b.ctx, engine,
1995 					      MI_ARB_CHECK);
1996 		if (IS_ERR(rq_b)) {
1997 			err = PTR_ERR(rq_b);
1998 			goto err_client_b;
1999 		}
2000 
2001 		i915_request_add(rq_b);
2002 
2003 		/* B is much more important than A! (But A is unpreemptable.) */
2004 		GEM_BUG_ON(rq_prio(rq_b) <= rq_prio(rq_a));
2005 
2006 		/* Wait long enough for preemption and timeslicing */
2007 		if (igt_wait_for_spinner(&b.spin, rq_b)) {
2008 			pr_err("Second client started too early!\n");
2009 			goto err_wedged;
2010 		}
2011 
2012 		igt_spinner_end(&a.spin);
2013 
2014 		if (!igt_wait_for_spinner(&b.spin, rq_b)) {
2015 			pr_err("Second client failed to start\n");
2016 			goto err_wedged;
2017 		}
2018 
2019 		igt_spinner_end(&b.spin);
2020 
2021 		if (engine->execlists.preempt_hang.count) {
2022 			pr_err("Preemption recorded x%d; should have been suppressed!\n",
2023 			       engine->execlists.preempt_hang.count);
2024 			err = -EINVAL;
2025 			goto err_wedged;
2026 		}
2027 
2028 		if (igt_flush_test(gt->i915))
2029 			goto err_wedged;
2030 	}
2031 
2032 	err = 0;
2033 err_client_b:
2034 	preempt_client_fini(&b);
2035 err_client_a:
2036 	preempt_client_fini(&a);
2037 	return err;
2038 
2039 err_wedged:
2040 	igt_spinner_end(&b.spin);
2041 	igt_spinner_end(&a.spin);
2042 	intel_gt_set_wedged(gt);
2043 	err = -EIO;
2044 	goto err_client_b;
2045 }
2046 
2047 struct live_preempt_cancel {
2048 	struct intel_engine_cs *engine;
2049 	struct preempt_client a, b;
2050 };
2051 
2052 static int __cancel_active0(struct live_preempt_cancel *arg)
2053 {
2054 	struct i915_request *rq;
2055 	struct igt_live_test t;
2056 	int err;
2057 
2058 	/* Preempt cancel of ELSP0 */
2059 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2060 	if (igt_live_test_begin(&t, arg->engine->i915,
2061 				__func__, arg->engine->name))
2062 		return -EIO;
2063 
2064 	rq = spinner_create_request(&arg->a.spin,
2065 				    arg->a.ctx, arg->engine,
2066 				    MI_ARB_CHECK);
2067 	if (IS_ERR(rq))
2068 		return PTR_ERR(rq);
2069 
2070 	clear_bit(CONTEXT_BANNED, &rq->context->flags);
2071 	i915_request_get(rq);
2072 	i915_request_add(rq);
2073 	if (!igt_wait_for_spinner(&arg->a.spin, rq)) {
2074 		err = -EIO;
2075 		goto out;
2076 	}
2077 
2078 	intel_context_ban(rq->context, rq);
2079 	err = intel_engine_pulse(arg->engine);
2080 	if (err)
2081 		goto out;
2082 
2083 	err = wait_for_reset(arg->engine, rq, HZ / 2);
2084 	if (err) {
2085 		pr_err("Cancelled inflight0 request did not reset\n");
2086 		goto out;
2087 	}
2088 
2089 out:
2090 	i915_request_put(rq);
2091 	if (igt_live_test_end(&t))
2092 		err = -EIO;
2093 	return err;
2094 }
2095 
2096 static int __cancel_active1(struct live_preempt_cancel *arg)
2097 {
2098 	struct i915_request *rq[2] = {};
2099 	struct igt_live_test t;
2100 	int err;
2101 
2102 	/* Preempt cancel of ELSP1 */
2103 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2104 	if (igt_live_test_begin(&t, arg->engine->i915,
2105 				__func__, arg->engine->name))
2106 		return -EIO;
2107 
2108 	rq[0] = spinner_create_request(&arg->a.spin,
2109 				       arg->a.ctx, arg->engine,
2110 				       MI_NOOP); /* no preemption */
2111 	if (IS_ERR(rq[0]))
2112 		return PTR_ERR(rq[0]);
2113 
2114 	clear_bit(CONTEXT_BANNED, &rq[0]->context->flags);
2115 	i915_request_get(rq[0]);
2116 	i915_request_add(rq[0]);
2117 	if (!igt_wait_for_spinner(&arg->a.spin, rq[0])) {
2118 		err = -EIO;
2119 		goto out;
2120 	}
2121 
2122 	rq[1] = spinner_create_request(&arg->b.spin,
2123 				       arg->b.ctx, arg->engine,
2124 				       MI_ARB_CHECK);
2125 	if (IS_ERR(rq[1])) {
2126 		err = PTR_ERR(rq[1]);
2127 		goto out;
2128 	}
2129 
2130 	clear_bit(CONTEXT_BANNED, &rq[1]->context->flags);
2131 	i915_request_get(rq[1]);
2132 	err = i915_request_await_dma_fence(rq[1], &rq[0]->fence);
2133 	i915_request_add(rq[1]);
2134 	if (err)
2135 		goto out;
2136 
2137 	intel_context_ban(rq[1]->context, rq[1]);
2138 	err = intel_engine_pulse(arg->engine);
2139 	if (err)
2140 		goto out;
2141 
2142 	igt_spinner_end(&arg->a.spin);
2143 	err = wait_for_reset(arg->engine, rq[1], HZ / 2);
2144 	if (err)
2145 		goto out;
2146 
2147 	if (rq[0]->fence.error != 0) {
2148 		pr_err("Normal inflight0 request did not complete\n");
2149 		err = -EINVAL;
2150 		goto out;
2151 	}
2152 
2153 	if (rq[1]->fence.error != -EIO) {
2154 		pr_err("Cancelled inflight1 request did not report -EIO\n");
2155 		err = -EINVAL;
2156 		goto out;
2157 	}
2158 
2159 out:
2160 	i915_request_put(rq[1]);
2161 	i915_request_put(rq[0]);
2162 	if (igt_live_test_end(&t))
2163 		err = -EIO;
2164 	return err;
2165 }
2166 
2167 static int __cancel_queued(struct live_preempt_cancel *arg)
2168 {
2169 	struct i915_request *rq[3] = {};
2170 	struct igt_live_test t;
2171 	int err;
2172 
2173 	/* Full ELSP and one in the wings */
2174 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2175 	if (igt_live_test_begin(&t, arg->engine->i915,
2176 				__func__, arg->engine->name))
2177 		return -EIO;
2178 
2179 	rq[0] = spinner_create_request(&arg->a.spin,
2180 				       arg->a.ctx, arg->engine,
2181 				       MI_ARB_CHECK);
2182 	if (IS_ERR(rq[0]))
2183 		return PTR_ERR(rq[0]);
2184 
2185 	clear_bit(CONTEXT_BANNED, &rq[0]->context->flags);
2186 	i915_request_get(rq[0]);
2187 	i915_request_add(rq[0]);
2188 	if (!igt_wait_for_spinner(&arg->a.spin, rq[0])) {
2189 		err = -EIO;
2190 		goto out;
2191 	}
2192 
2193 	rq[1] = igt_request_alloc(arg->b.ctx, arg->engine);
2194 	if (IS_ERR(rq[1])) {
2195 		err = PTR_ERR(rq[1]);
2196 		goto out;
2197 	}
2198 
2199 	clear_bit(CONTEXT_BANNED, &rq[1]->context->flags);
2200 	i915_request_get(rq[1]);
2201 	err = i915_request_await_dma_fence(rq[1], &rq[0]->fence);
2202 	i915_request_add(rq[1]);
2203 	if (err)
2204 		goto out;
2205 
2206 	rq[2] = spinner_create_request(&arg->b.spin,
2207 				       arg->a.ctx, arg->engine,
2208 				       MI_ARB_CHECK);
2209 	if (IS_ERR(rq[2])) {
2210 		err = PTR_ERR(rq[2]);
2211 		goto out;
2212 	}
2213 
2214 	i915_request_get(rq[2]);
2215 	err = i915_request_await_dma_fence(rq[2], &rq[1]->fence);
2216 	i915_request_add(rq[2]);
2217 	if (err)
2218 		goto out;
2219 
2220 	intel_context_ban(rq[2]->context, rq[2]);
2221 	err = intel_engine_pulse(arg->engine);
2222 	if (err)
2223 		goto out;
2224 
2225 	err = wait_for_reset(arg->engine, rq[2], HZ / 2);
2226 	if (err)
2227 		goto out;
2228 
2229 	if (rq[0]->fence.error != -EIO) {
2230 		pr_err("Cancelled inflight0 request did not report -EIO\n");
2231 		err = -EINVAL;
2232 		goto out;
2233 	}
2234 
2235 	/*
2236 	 * The behavior between having semaphores and not is different. With
2237 	 * semaphores the subsequent request is on the hardware and not cancelled
2238 	 * while without the request is held in the driver and cancelled.
2239 	 */
2240 	if (intel_engine_has_semaphores(rq[1]->engine) &&
2241 	    rq[1]->fence.error != 0) {
2242 		pr_err("Normal inflight1 request did not complete\n");
2243 		err = -EINVAL;
2244 		goto out;
2245 	}
2246 
2247 	if (rq[2]->fence.error != -EIO) {
2248 		pr_err("Cancelled queued request did not report -EIO\n");
2249 		err = -EINVAL;
2250 		goto out;
2251 	}
2252 
2253 out:
2254 	i915_request_put(rq[2]);
2255 	i915_request_put(rq[1]);
2256 	i915_request_put(rq[0]);
2257 	if (igt_live_test_end(&t))
2258 		err = -EIO;
2259 	return err;
2260 }
2261 
2262 static int __cancel_hostile(struct live_preempt_cancel *arg)
2263 {
2264 	struct i915_request *rq;
2265 	int err;
2266 
2267 	/* Preempt cancel non-preemptible spinner in ELSP0 */
2268 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
2269 		return 0;
2270 
2271 	if (!intel_has_reset_engine(arg->engine->gt))
2272 		return 0;
2273 
2274 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2275 	rq = spinner_create_request(&arg->a.spin,
2276 				    arg->a.ctx, arg->engine,
2277 				    MI_NOOP); /* preemption disabled */
2278 	if (IS_ERR(rq))
2279 		return PTR_ERR(rq);
2280 
2281 	clear_bit(CONTEXT_BANNED, &rq->context->flags);
2282 	i915_request_get(rq);
2283 	i915_request_add(rq);
2284 	if (!igt_wait_for_spinner(&arg->a.spin, rq)) {
2285 		err = -EIO;
2286 		goto out;
2287 	}
2288 
2289 	intel_context_ban(rq->context, rq);
2290 	err = intel_engine_pulse(arg->engine); /* force reset */
2291 	if (err)
2292 		goto out;
2293 
2294 	err = wait_for_reset(arg->engine, rq, HZ / 2);
2295 	if (err) {
2296 		pr_err("Cancelled inflight0 request did not reset\n");
2297 		goto out;
2298 	}
2299 
2300 out:
2301 	i915_request_put(rq);
2302 	if (igt_flush_test(arg->engine->i915))
2303 		err = -EIO;
2304 	return err;
2305 }
2306 
2307 static void force_reset_timeout(struct intel_engine_cs *engine)
2308 {
2309 	engine->reset_timeout.probability = 999;
2310 	atomic_set(&engine->reset_timeout.times, -1);
2311 }
2312 
2313 static void cancel_reset_timeout(struct intel_engine_cs *engine)
2314 {
2315 	memset(&engine->reset_timeout, 0, sizeof(engine->reset_timeout));
2316 }
2317 
2318 static int __cancel_fail(struct live_preempt_cancel *arg)
2319 {
2320 	struct intel_engine_cs *engine = arg->engine;
2321 	struct i915_request *rq;
2322 	int err;
2323 
2324 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
2325 		return 0;
2326 
2327 	if (!intel_has_reset_engine(engine->gt))
2328 		return 0;
2329 
2330 	GEM_TRACE("%s(%s)\n", __func__, engine->name);
2331 	rq = spinner_create_request(&arg->a.spin,
2332 				    arg->a.ctx, engine,
2333 				    MI_NOOP); /* preemption disabled */
2334 	if (IS_ERR(rq))
2335 		return PTR_ERR(rq);
2336 
2337 	clear_bit(CONTEXT_BANNED, &rq->context->flags);
2338 	i915_request_get(rq);
2339 	i915_request_add(rq);
2340 	if (!igt_wait_for_spinner(&arg->a.spin, rq)) {
2341 		err = -EIO;
2342 		goto out;
2343 	}
2344 
2345 	intel_context_set_banned(rq->context);
2346 
2347 	err = intel_engine_pulse(engine);
2348 	if (err)
2349 		goto out;
2350 
2351 	force_reset_timeout(engine);
2352 
2353 	/* force preempt reset [failure] */
2354 	while (!engine->execlists.pending[0])
2355 		intel_engine_flush_submission(engine);
2356 	del_timer_sync(&engine->execlists.preempt);
2357 	intel_engine_flush_submission(engine);
2358 
2359 	cancel_reset_timeout(engine);
2360 
2361 	/* after failure, require heartbeats to reset device */
2362 	intel_engine_set_heartbeat(engine, 1);
2363 	err = wait_for_reset(engine, rq, HZ / 2);
2364 	intel_engine_set_heartbeat(engine,
2365 				   engine->defaults.heartbeat_interval_ms);
2366 	if (err) {
2367 		pr_err("Cancelled inflight0 request did not reset\n");
2368 		goto out;
2369 	}
2370 
2371 out:
2372 	i915_request_put(rq);
2373 	if (igt_flush_test(engine->i915))
2374 		err = -EIO;
2375 	return err;
2376 }
2377 
2378 static int live_preempt_cancel(void *arg)
2379 {
2380 	struct intel_gt *gt = arg;
2381 	struct live_preempt_cancel data;
2382 	enum intel_engine_id id;
2383 	int err = -ENOMEM;
2384 
2385 	/*
2386 	 * To cancel an inflight context, we need to first remove it from the
2387 	 * GPU. That sounds like preemption! Plus a little bit of bookkeeping.
2388 	 */
2389 
2390 	if (preempt_client_init(gt, &data.a))
2391 		return -ENOMEM;
2392 	if (preempt_client_init(gt, &data.b))
2393 		goto err_client_a;
2394 
2395 	for_each_engine(data.engine, gt, id) {
2396 		if (!intel_engine_has_preemption(data.engine))
2397 			continue;
2398 
2399 		err = __cancel_active0(&data);
2400 		if (err)
2401 			goto err_wedged;
2402 
2403 		err = __cancel_active1(&data);
2404 		if (err)
2405 			goto err_wedged;
2406 
2407 		err = __cancel_queued(&data);
2408 		if (err)
2409 			goto err_wedged;
2410 
2411 		err = __cancel_hostile(&data);
2412 		if (err)
2413 			goto err_wedged;
2414 
2415 		err = __cancel_fail(&data);
2416 		if (err)
2417 			goto err_wedged;
2418 	}
2419 
2420 	err = 0;
2421 err_client_b:
2422 	preempt_client_fini(&data.b);
2423 err_client_a:
2424 	preempt_client_fini(&data.a);
2425 	return err;
2426 
2427 err_wedged:
2428 	GEM_TRACE_DUMP();
2429 	igt_spinner_end(&data.b.spin);
2430 	igt_spinner_end(&data.a.spin);
2431 	intel_gt_set_wedged(gt);
2432 	goto err_client_b;
2433 }
2434 
2435 static int live_suppress_self_preempt(void *arg)
2436 {
2437 	struct i915_sched_attr attr = { .priority = I915_PRIORITY_MAX };
2438 	struct intel_gt *gt = arg;
2439 	struct intel_engine_cs *engine;
2440 	struct preempt_client a, b;
2441 	enum intel_engine_id id;
2442 	int err = -ENOMEM;
2443 
2444 	/*
2445 	 * Verify that if a preemption request does not cause a change in
2446 	 * the current execution order, the preempt-to-idle injection is
2447 	 * skipped and that we do not accidentally apply it after the CS
2448 	 * completion event.
2449 	 */
2450 
2451 	if (intel_uc_uses_guc_submission(&gt->uc))
2452 		return 0; /* presume black blox */
2453 
2454 	if (intel_vgpu_active(gt->i915))
2455 		return 0; /* GVT forces single port & request submission */
2456 
2457 	if (preempt_client_init(gt, &a))
2458 		return -ENOMEM;
2459 	if (preempt_client_init(gt, &b))
2460 		goto err_client_a;
2461 
2462 	for_each_engine(engine, gt, id) {
2463 		struct i915_request *rq_a, *rq_b;
2464 		int depth;
2465 
2466 		if (!intel_engine_has_preemption(engine))
2467 			continue;
2468 
2469 		if (igt_flush_test(gt->i915))
2470 			goto err_wedged;
2471 
2472 		st_engine_heartbeat_disable(engine);
2473 		engine->execlists.preempt_hang.count = 0;
2474 
2475 		rq_a = spinner_create_request(&a.spin,
2476 					      a.ctx, engine,
2477 					      MI_NOOP);
2478 		if (IS_ERR(rq_a)) {
2479 			err = PTR_ERR(rq_a);
2480 			st_engine_heartbeat_enable(engine);
2481 			goto err_client_b;
2482 		}
2483 
2484 		i915_request_add(rq_a);
2485 		if (!igt_wait_for_spinner(&a.spin, rq_a)) {
2486 			pr_err("First client failed to start\n");
2487 			st_engine_heartbeat_enable(engine);
2488 			goto err_wedged;
2489 		}
2490 
2491 		/* Keep postponing the timer to avoid premature slicing */
2492 		mod_timer(&engine->execlists.timer, jiffies + HZ);
2493 		for (depth = 0; depth < 8; depth++) {
2494 			rq_b = spinner_create_request(&b.spin,
2495 						      b.ctx, engine,
2496 						      MI_NOOP);
2497 			if (IS_ERR(rq_b)) {
2498 				err = PTR_ERR(rq_b);
2499 				st_engine_heartbeat_enable(engine);
2500 				goto err_client_b;
2501 			}
2502 			i915_request_add(rq_b);
2503 
2504 			GEM_BUG_ON(i915_request_completed(rq_a));
2505 			engine->sched_engine->schedule(rq_a, &attr);
2506 			igt_spinner_end(&a.spin);
2507 
2508 			if (!igt_wait_for_spinner(&b.spin, rq_b)) {
2509 				pr_err("Second client failed to start\n");
2510 				st_engine_heartbeat_enable(engine);
2511 				goto err_wedged;
2512 			}
2513 
2514 			swap(a, b);
2515 			rq_a = rq_b;
2516 		}
2517 		igt_spinner_end(&a.spin);
2518 
2519 		if (engine->execlists.preempt_hang.count) {
2520 			pr_err("Preemption on %s recorded x%d, depth %d; should have been suppressed!\n",
2521 			       engine->name,
2522 			       engine->execlists.preempt_hang.count,
2523 			       depth);
2524 			st_engine_heartbeat_enable(engine);
2525 			err = -EINVAL;
2526 			goto err_client_b;
2527 		}
2528 
2529 		st_engine_heartbeat_enable(engine);
2530 		if (igt_flush_test(gt->i915))
2531 			goto err_wedged;
2532 	}
2533 
2534 	err = 0;
2535 err_client_b:
2536 	preempt_client_fini(&b);
2537 err_client_a:
2538 	preempt_client_fini(&a);
2539 	return err;
2540 
2541 err_wedged:
2542 	igt_spinner_end(&b.spin);
2543 	igt_spinner_end(&a.spin);
2544 	intel_gt_set_wedged(gt);
2545 	err = -EIO;
2546 	goto err_client_b;
2547 }
2548 
2549 static int live_chain_preempt(void *arg)
2550 {
2551 	struct intel_gt *gt = arg;
2552 	struct intel_engine_cs *engine;
2553 	struct preempt_client hi, lo;
2554 	enum intel_engine_id id;
2555 	int err = -ENOMEM;
2556 
2557 	/*
2558 	 * Build a chain AB...BA between two contexts (A, B) and request
2559 	 * preemption of the last request. It should then complete before
2560 	 * the previously submitted spinner in B.
2561 	 */
2562 
2563 	if (preempt_client_init(gt, &hi))
2564 		return -ENOMEM;
2565 
2566 	if (preempt_client_init(gt, &lo))
2567 		goto err_client_hi;
2568 
2569 	for_each_engine(engine, gt, id) {
2570 		struct i915_sched_attr attr = { .priority = I915_PRIORITY_MAX };
2571 		struct igt_live_test t;
2572 		struct i915_request *rq;
2573 		int ring_size, count, i;
2574 
2575 		if (!intel_engine_has_preemption(engine))
2576 			continue;
2577 
2578 		rq = spinner_create_request(&lo.spin,
2579 					    lo.ctx, engine,
2580 					    MI_ARB_CHECK);
2581 		if (IS_ERR(rq))
2582 			goto err_wedged;
2583 
2584 		i915_request_get(rq);
2585 		i915_request_add(rq);
2586 
2587 		ring_size = rq->wa_tail - rq->head;
2588 		if (ring_size < 0)
2589 			ring_size += rq->ring->size;
2590 		ring_size = rq->ring->size / ring_size;
2591 		pr_debug("%s(%s): Using maximum of %d requests\n",
2592 			 __func__, engine->name, ring_size);
2593 
2594 		igt_spinner_end(&lo.spin);
2595 		if (i915_request_wait(rq, 0, HZ / 2) < 0) {
2596 			pr_err("Timed out waiting to flush %s\n", engine->name);
2597 			i915_request_put(rq);
2598 			goto err_wedged;
2599 		}
2600 		i915_request_put(rq);
2601 
2602 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
2603 			err = -EIO;
2604 			goto err_wedged;
2605 		}
2606 
2607 		for_each_prime_number_from(count, 1, ring_size) {
2608 			rq = spinner_create_request(&hi.spin,
2609 						    hi.ctx, engine,
2610 						    MI_ARB_CHECK);
2611 			if (IS_ERR(rq))
2612 				goto err_wedged;
2613 			i915_request_add(rq);
2614 			if (!igt_wait_for_spinner(&hi.spin, rq))
2615 				goto err_wedged;
2616 
2617 			rq = spinner_create_request(&lo.spin,
2618 						    lo.ctx, engine,
2619 						    MI_ARB_CHECK);
2620 			if (IS_ERR(rq))
2621 				goto err_wedged;
2622 			i915_request_add(rq);
2623 
2624 			for (i = 0; i < count; i++) {
2625 				rq = igt_request_alloc(lo.ctx, engine);
2626 				if (IS_ERR(rq))
2627 					goto err_wedged;
2628 				i915_request_add(rq);
2629 			}
2630 
2631 			rq = igt_request_alloc(hi.ctx, engine);
2632 			if (IS_ERR(rq))
2633 				goto err_wedged;
2634 
2635 			i915_request_get(rq);
2636 			i915_request_add(rq);
2637 			engine->sched_engine->schedule(rq, &attr);
2638 
2639 			igt_spinner_end(&hi.spin);
2640 			if (i915_request_wait(rq, 0, HZ / 5) < 0) {
2641 				struct drm_printer p =
2642 					drm_info_printer(gt->i915->drm.dev);
2643 
2644 				pr_err("Failed to preempt over chain of %d\n",
2645 				       count);
2646 				intel_engine_dump(engine, &p,
2647 						  "%s\n", engine->name);
2648 				i915_request_put(rq);
2649 				goto err_wedged;
2650 			}
2651 			igt_spinner_end(&lo.spin);
2652 			i915_request_put(rq);
2653 
2654 			rq = igt_request_alloc(lo.ctx, engine);
2655 			if (IS_ERR(rq))
2656 				goto err_wedged;
2657 
2658 			i915_request_get(rq);
2659 			i915_request_add(rq);
2660 
2661 			if (i915_request_wait(rq, 0, HZ / 5) < 0) {
2662 				struct drm_printer p =
2663 					drm_info_printer(gt->i915->drm.dev);
2664 
2665 				pr_err("Failed to flush low priority chain of %d requests\n",
2666 				       count);
2667 				intel_engine_dump(engine, &p,
2668 						  "%s\n", engine->name);
2669 
2670 				i915_request_put(rq);
2671 				goto err_wedged;
2672 			}
2673 			i915_request_put(rq);
2674 		}
2675 
2676 		if (igt_live_test_end(&t)) {
2677 			err = -EIO;
2678 			goto err_wedged;
2679 		}
2680 	}
2681 
2682 	err = 0;
2683 err_client_lo:
2684 	preempt_client_fini(&lo);
2685 err_client_hi:
2686 	preempt_client_fini(&hi);
2687 	return err;
2688 
2689 err_wedged:
2690 	igt_spinner_end(&hi.spin);
2691 	igt_spinner_end(&lo.spin);
2692 	intel_gt_set_wedged(gt);
2693 	err = -EIO;
2694 	goto err_client_lo;
2695 }
2696 
2697 static int create_gang(struct intel_engine_cs *engine,
2698 		       struct i915_request **prev)
2699 {
2700 	struct drm_i915_gem_object *obj;
2701 	struct intel_context *ce;
2702 	struct i915_request *rq;
2703 	struct i915_vma *vma;
2704 	u32 *cs;
2705 	int err;
2706 
2707 	ce = intel_context_create(engine);
2708 	if (IS_ERR(ce))
2709 		return PTR_ERR(ce);
2710 
2711 	obj = i915_gem_object_create_internal(engine->i915, 4096);
2712 	if (IS_ERR(obj)) {
2713 		err = PTR_ERR(obj);
2714 		goto err_ce;
2715 	}
2716 
2717 	vma = i915_vma_instance(obj, ce->vm, NULL);
2718 	if (IS_ERR(vma)) {
2719 		err = PTR_ERR(vma);
2720 		goto err_obj;
2721 	}
2722 
2723 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
2724 	if (err)
2725 		goto err_obj;
2726 
2727 	cs = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
2728 	if (IS_ERR(cs)) {
2729 		err = PTR_ERR(cs);
2730 		goto err_obj;
2731 	}
2732 
2733 	/* Semaphore target: spin until zero */
2734 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
2735 
2736 	*cs++ = MI_SEMAPHORE_WAIT |
2737 		MI_SEMAPHORE_POLL |
2738 		MI_SEMAPHORE_SAD_EQ_SDD;
2739 	*cs++ = 0;
2740 	*cs++ = lower_32_bits(vma->node.start);
2741 	*cs++ = upper_32_bits(vma->node.start);
2742 
2743 	if (*prev) {
2744 		u64 offset = (*prev)->batch->node.start;
2745 
2746 		/* Terminate the spinner in the next lower priority batch. */
2747 		*cs++ = MI_STORE_DWORD_IMM_GEN4;
2748 		*cs++ = lower_32_bits(offset);
2749 		*cs++ = upper_32_bits(offset);
2750 		*cs++ = 0;
2751 	}
2752 
2753 	*cs++ = MI_BATCH_BUFFER_END;
2754 	i915_gem_object_flush_map(obj);
2755 	i915_gem_object_unpin_map(obj);
2756 
2757 	rq = intel_context_create_request(ce);
2758 	if (IS_ERR(rq)) {
2759 		err = PTR_ERR(rq);
2760 		goto err_obj;
2761 	}
2762 
2763 	rq->batch = i915_vma_get(vma);
2764 	i915_request_get(rq);
2765 
2766 	i915_vma_lock(vma);
2767 	err = i915_vma_move_to_active(vma, rq, 0);
2768 	if (!err)
2769 		err = rq->engine->emit_bb_start(rq,
2770 						vma->node.start,
2771 						PAGE_SIZE, 0);
2772 	i915_vma_unlock(vma);
2773 	i915_request_add(rq);
2774 	if (err)
2775 		goto err_rq;
2776 
2777 	i915_gem_object_put(obj);
2778 	intel_context_put(ce);
2779 
2780 	rq->mock.link.next = &(*prev)->mock.link;
2781 	*prev = rq;
2782 	return 0;
2783 
2784 err_rq:
2785 	i915_vma_put(rq->batch);
2786 	i915_request_put(rq);
2787 err_obj:
2788 	i915_gem_object_put(obj);
2789 err_ce:
2790 	intel_context_put(ce);
2791 	return err;
2792 }
2793 
2794 static int __live_preempt_ring(struct intel_engine_cs *engine,
2795 			       struct igt_spinner *spin,
2796 			       int queue_sz, int ring_sz)
2797 {
2798 	struct intel_context *ce[2] = {};
2799 	struct i915_request *rq;
2800 	struct igt_live_test t;
2801 	int err = 0;
2802 	int n;
2803 
2804 	if (igt_live_test_begin(&t, engine->i915, __func__, engine->name))
2805 		return -EIO;
2806 
2807 	for (n = 0; n < ARRAY_SIZE(ce); n++) {
2808 		struct intel_context *tmp;
2809 
2810 		tmp = intel_context_create(engine);
2811 		if (IS_ERR(tmp)) {
2812 			err = PTR_ERR(tmp);
2813 			goto err_ce;
2814 		}
2815 
2816 		tmp->ring_size = ring_sz;
2817 
2818 		err = intel_context_pin(tmp);
2819 		if (err) {
2820 			intel_context_put(tmp);
2821 			goto err_ce;
2822 		}
2823 
2824 		memset32(tmp->ring->vaddr,
2825 			 0xdeadbeef, /* trigger a hang if executed */
2826 			 tmp->ring->vma->size / sizeof(u32));
2827 
2828 		ce[n] = tmp;
2829 	}
2830 
2831 	rq = igt_spinner_create_request(spin, ce[0], MI_ARB_CHECK);
2832 	if (IS_ERR(rq)) {
2833 		err = PTR_ERR(rq);
2834 		goto err_ce;
2835 	}
2836 
2837 	i915_request_get(rq);
2838 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
2839 	i915_request_add(rq);
2840 
2841 	if (!igt_wait_for_spinner(spin, rq)) {
2842 		intel_gt_set_wedged(engine->gt);
2843 		i915_request_put(rq);
2844 		err = -ETIME;
2845 		goto err_ce;
2846 	}
2847 
2848 	/* Fill the ring, until we will cause a wrap */
2849 	n = 0;
2850 	while (ce[0]->ring->tail - rq->wa_tail <= queue_sz) {
2851 		struct i915_request *tmp;
2852 
2853 		tmp = intel_context_create_request(ce[0]);
2854 		if (IS_ERR(tmp)) {
2855 			err = PTR_ERR(tmp);
2856 			i915_request_put(rq);
2857 			goto err_ce;
2858 		}
2859 
2860 		i915_request_add(tmp);
2861 		intel_engine_flush_submission(engine);
2862 		n++;
2863 	}
2864 	intel_engine_flush_submission(engine);
2865 	pr_debug("%s: Filled %d with %d nop tails {size:%x, tail:%x, emit:%x, rq.tail:%x}\n",
2866 		 engine->name, queue_sz, n,
2867 		 ce[0]->ring->size,
2868 		 ce[0]->ring->tail,
2869 		 ce[0]->ring->emit,
2870 		 rq->tail);
2871 	i915_request_put(rq);
2872 
2873 	/* Create a second request to preempt the first ring */
2874 	rq = intel_context_create_request(ce[1]);
2875 	if (IS_ERR(rq)) {
2876 		err = PTR_ERR(rq);
2877 		goto err_ce;
2878 	}
2879 
2880 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
2881 	i915_request_get(rq);
2882 	i915_request_add(rq);
2883 
2884 	err = wait_for_submit(engine, rq, HZ / 2);
2885 	i915_request_put(rq);
2886 	if (err) {
2887 		pr_err("%s: preemption request was not submitted\n",
2888 		       engine->name);
2889 		err = -ETIME;
2890 	}
2891 
2892 	pr_debug("%s: ring[0]:{ tail:%x, emit:%x }, ring[1]:{ tail:%x, emit:%x }\n",
2893 		 engine->name,
2894 		 ce[0]->ring->tail, ce[0]->ring->emit,
2895 		 ce[1]->ring->tail, ce[1]->ring->emit);
2896 
2897 err_ce:
2898 	intel_engine_flush_submission(engine);
2899 	igt_spinner_end(spin);
2900 	for (n = 0; n < ARRAY_SIZE(ce); n++) {
2901 		if (IS_ERR_OR_NULL(ce[n]))
2902 			break;
2903 
2904 		intel_context_unpin(ce[n]);
2905 		intel_context_put(ce[n]);
2906 	}
2907 	if (igt_live_test_end(&t))
2908 		err = -EIO;
2909 	return err;
2910 }
2911 
2912 static int live_preempt_ring(void *arg)
2913 {
2914 	struct intel_gt *gt = arg;
2915 	struct intel_engine_cs *engine;
2916 	struct igt_spinner spin;
2917 	enum intel_engine_id id;
2918 	int err = 0;
2919 
2920 	/*
2921 	 * Check that we rollback large chunks of a ring in order to do a
2922 	 * preemption event. Similar to live_unlite_ring, but looking at
2923 	 * ring size rather than the impact of intel_ring_direction().
2924 	 */
2925 
2926 	if (igt_spinner_init(&spin, gt))
2927 		return -ENOMEM;
2928 
2929 	for_each_engine(engine, gt, id) {
2930 		int n;
2931 
2932 		if (!intel_engine_has_preemption(engine))
2933 			continue;
2934 
2935 		if (!intel_engine_can_store_dword(engine))
2936 			continue;
2937 
2938 		st_engine_heartbeat_disable(engine);
2939 
2940 		for (n = 0; n <= 3; n++) {
2941 			err = __live_preempt_ring(engine, &spin,
2942 						  n * SZ_4K / 4, SZ_4K);
2943 			if (err)
2944 				break;
2945 		}
2946 
2947 		st_engine_heartbeat_enable(engine);
2948 		if (err)
2949 			break;
2950 	}
2951 
2952 	igt_spinner_fini(&spin);
2953 	return err;
2954 }
2955 
2956 static int live_preempt_gang(void *arg)
2957 {
2958 	struct intel_gt *gt = arg;
2959 	struct intel_engine_cs *engine;
2960 	enum intel_engine_id id;
2961 
2962 	/*
2963 	 * Build as long a chain of preempters as we can, with each
2964 	 * request higher priority than the last. Once we are ready, we release
2965 	 * the last batch which then precolates down the chain, each releasing
2966 	 * the next oldest in turn. The intent is to simply push as hard as we
2967 	 * can with the number of preemptions, trying to exceed narrow HW
2968 	 * limits. At a minimum, we insist that we can sort all the user
2969 	 * high priority levels into execution order.
2970 	 */
2971 
2972 	for_each_engine(engine, gt, id) {
2973 		struct i915_request *rq = NULL;
2974 		struct igt_live_test t;
2975 		IGT_TIMEOUT(end_time);
2976 		int prio = 0;
2977 		int err = 0;
2978 		u32 *cs;
2979 
2980 		if (!intel_engine_has_preemption(engine))
2981 			continue;
2982 
2983 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name))
2984 			return -EIO;
2985 
2986 		do {
2987 			struct i915_sched_attr attr = { .priority = prio++ };
2988 
2989 			err = create_gang(engine, &rq);
2990 			if (err)
2991 				break;
2992 
2993 			/* Submit each spinner at increasing priority */
2994 			engine->sched_engine->schedule(rq, &attr);
2995 		} while (prio <= I915_PRIORITY_MAX &&
2996 			 !__igt_timeout(end_time, NULL));
2997 		pr_debug("%s: Preempt chain of %d requests\n",
2998 			 engine->name, prio);
2999 
3000 		/*
3001 		 * Such that the last spinner is the highest priority and
3002 		 * should execute first. When that spinner completes,
3003 		 * it will terminate the next lowest spinner until there
3004 		 * are no more spinners and the gang is complete.
3005 		 */
3006 		cs = i915_gem_object_pin_map_unlocked(rq->batch->obj, I915_MAP_WC);
3007 		if (!IS_ERR(cs)) {
3008 			*cs = 0;
3009 			i915_gem_object_unpin_map(rq->batch->obj);
3010 		} else {
3011 			err = PTR_ERR(cs);
3012 			intel_gt_set_wedged(gt);
3013 		}
3014 
3015 		while (rq) { /* wait for each rq from highest to lowest prio */
3016 			struct i915_request *n = list_next_entry(rq, mock.link);
3017 
3018 			if (err == 0 && i915_request_wait(rq, 0, HZ / 5) < 0) {
3019 				struct drm_printer p =
3020 					drm_info_printer(engine->i915->drm.dev);
3021 
3022 				pr_err("Failed to flush chain of %d requests, at %d\n",
3023 				       prio, rq_prio(rq));
3024 				intel_engine_dump(engine, &p,
3025 						  "%s\n", engine->name);
3026 
3027 				err = -ETIME;
3028 			}
3029 
3030 			i915_vma_put(rq->batch);
3031 			i915_request_put(rq);
3032 			rq = n;
3033 		}
3034 
3035 		if (igt_live_test_end(&t))
3036 			err = -EIO;
3037 		if (err)
3038 			return err;
3039 	}
3040 
3041 	return 0;
3042 }
3043 
3044 static struct i915_vma *
3045 create_gpr_user(struct intel_engine_cs *engine,
3046 		struct i915_vma *result,
3047 		unsigned int offset)
3048 {
3049 	struct drm_i915_gem_object *obj;
3050 	struct i915_vma *vma;
3051 	u32 *cs;
3052 	int err;
3053 	int i;
3054 
3055 	obj = i915_gem_object_create_internal(engine->i915, 4096);
3056 	if (IS_ERR(obj))
3057 		return ERR_CAST(obj);
3058 
3059 	vma = i915_vma_instance(obj, result->vm, NULL);
3060 	if (IS_ERR(vma)) {
3061 		i915_gem_object_put(obj);
3062 		return vma;
3063 	}
3064 
3065 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
3066 	if (err) {
3067 		i915_vma_put(vma);
3068 		return ERR_PTR(err);
3069 	}
3070 
3071 	cs = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
3072 	if (IS_ERR(cs)) {
3073 		i915_vma_put(vma);
3074 		return ERR_CAST(cs);
3075 	}
3076 
3077 	/* All GPR are clear for new contexts. We use GPR(0) as a constant */
3078 	*cs++ = MI_LOAD_REGISTER_IMM(1);
3079 	*cs++ = CS_GPR(engine, 0);
3080 	*cs++ = 1;
3081 
3082 	for (i = 1; i < NUM_GPR; i++) {
3083 		u64 addr;
3084 
3085 		/*
3086 		 * Perform: GPR[i]++
3087 		 *
3088 		 * As we read and write into the context saved GPR[i], if
3089 		 * we restart this batch buffer from an earlier point, we
3090 		 * will repeat the increment and store a value > 1.
3091 		 */
3092 		*cs++ = MI_MATH(4);
3093 		*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(i));
3094 		*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(0));
3095 		*cs++ = MI_MATH_ADD;
3096 		*cs++ = MI_MATH_STORE(MI_MATH_REG(i), MI_MATH_REG_ACCU);
3097 
3098 		addr = result->node.start + offset + i * sizeof(*cs);
3099 		*cs++ = MI_STORE_REGISTER_MEM_GEN8;
3100 		*cs++ = CS_GPR(engine, 2 * i);
3101 		*cs++ = lower_32_bits(addr);
3102 		*cs++ = upper_32_bits(addr);
3103 
3104 		*cs++ = MI_SEMAPHORE_WAIT |
3105 			MI_SEMAPHORE_POLL |
3106 			MI_SEMAPHORE_SAD_GTE_SDD;
3107 		*cs++ = i;
3108 		*cs++ = lower_32_bits(result->node.start);
3109 		*cs++ = upper_32_bits(result->node.start);
3110 	}
3111 
3112 	*cs++ = MI_BATCH_BUFFER_END;
3113 	i915_gem_object_flush_map(obj);
3114 	i915_gem_object_unpin_map(obj);
3115 
3116 	return vma;
3117 }
3118 
3119 static struct i915_vma *create_global(struct intel_gt *gt, size_t sz)
3120 {
3121 	struct drm_i915_gem_object *obj;
3122 	struct i915_vma *vma;
3123 	int err;
3124 
3125 	obj = i915_gem_object_create_internal(gt->i915, sz);
3126 	if (IS_ERR(obj))
3127 		return ERR_CAST(obj);
3128 
3129 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
3130 	if (IS_ERR(vma)) {
3131 		i915_gem_object_put(obj);
3132 		return vma;
3133 	}
3134 
3135 	err = i915_ggtt_pin(vma, NULL, 0, 0);
3136 	if (err) {
3137 		i915_vma_put(vma);
3138 		return ERR_PTR(err);
3139 	}
3140 
3141 	return vma;
3142 }
3143 
3144 static struct i915_request *
3145 create_gpr_client(struct intel_engine_cs *engine,
3146 		  struct i915_vma *global,
3147 		  unsigned int offset)
3148 {
3149 	struct i915_vma *batch, *vma;
3150 	struct intel_context *ce;
3151 	struct i915_request *rq;
3152 	int err;
3153 
3154 	ce = intel_context_create(engine);
3155 	if (IS_ERR(ce))
3156 		return ERR_CAST(ce);
3157 
3158 	vma = i915_vma_instance(global->obj, ce->vm, NULL);
3159 	if (IS_ERR(vma)) {
3160 		err = PTR_ERR(vma);
3161 		goto out_ce;
3162 	}
3163 
3164 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
3165 	if (err)
3166 		goto out_ce;
3167 
3168 	batch = create_gpr_user(engine, vma, offset);
3169 	if (IS_ERR(batch)) {
3170 		err = PTR_ERR(batch);
3171 		goto out_vma;
3172 	}
3173 
3174 	rq = intel_context_create_request(ce);
3175 	if (IS_ERR(rq)) {
3176 		err = PTR_ERR(rq);
3177 		goto out_batch;
3178 	}
3179 
3180 	i915_vma_lock(vma);
3181 	err = i915_vma_move_to_active(vma, rq, 0);
3182 	i915_vma_unlock(vma);
3183 
3184 	i915_vma_lock(batch);
3185 	if (!err)
3186 		err = i915_vma_move_to_active(batch, rq, 0);
3187 	if (!err)
3188 		err = rq->engine->emit_bb_start(rq,
3189 						batch->node.start,
3190 						PAGE_SIZE, 0);
3191 	i915_vma_unlock(batch);
3192 	i915_vma_unpin(batch);
3193 
3194 	if (!err)
3195 		i915_request_get(rq);
3196 	i915_request_add(rq);
3197 
3198 out_batch:
3199 	i915_vma_put(batch);
3200 out_vma:
3201 	i915_vma_unpin(vma);
3202 out_ce:
3203 	intel_context_put(ce);
3204 	return err ? ERR_PTR(err) : rq;
3205 }
3206 
3207 static int preempt_user(struct intel_engine_cs *engine,
3208 			struct i915_vma *global,
3209 			int id)
3210 {
3211 	struct i915_sched_attr attr = {
3212 		.priority = I915_PRIORITY_MAX
3213 	};
3214 	struct i915_request *rq;
3215 	int err = 0;
3216 	u32 *cs;
3217 
3218 	rq = intel_engine_create_kernel_request(engine);
3219 	if (IS_ERR(rq))
3220 		return PTR_ERR(rq);
3221 
3222 	cs = intel_ring_begin(rq, 4);
3223 	if (IS_ERR(cs)) {
3224 		i915_request_add(rq);
3225 		return PTR_ERR(cs);
3226 	}
3227 
3228 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
3229 	*cs++ = i915_ggtt_offset(global);
3230 	*cs++ = 0;
3231 	*cs++ = id;
3232 
3233 	intel_ring_advance(rq, cs);
3234 
3235 	i915_request_get(rq);
3236 	i915_request_add(rq);
3237 
3238 	engine->sched_engine->schedule(rq, &attr);
3239 
3240 	if (i915_request_wait(rq, 0, HZ / 2) < 0)
3241 		err = -ETIME;
3242 	i915_request_put(rq);
3243 
3244 	return err;
3245 }
3246 
3247 static int live_preempt_user(void *arg)
3248 {
3249 	struct intel_gt *gt = arg;
3250 	struct intel_engine_cs *engine;
3251 	struct i915_vma *global;
3252 	enum intel_engine_id id;
3253 	u32 *result;
3254 	int err = 0;
3255 
3256 	/*
3257 	 * In our other tests, we look at preemption in carefully
3258 	 * controlled conditions in the ringbuffer. Since most of the
3259 	 * time is spent in user batches, most of our preemptions naturally
3260 	 * occur there. We want to verify that when we preempt inside a batch
3261 	 * we continue on from the current instruction and do not roll back
3262 	 * to the start, or another earlier arbitration point.
3263 	 *
3264 	 * To verify this, we create a batch which is a mixture of
3265 	 * MI_MATH (gpr++) MI_SRM (gpr) and preemption points. Then with
3266 	 * a few preempting contexts thrown into the mix, we look for any
3267 	 * repeated instructions (which show up as incorrect values).
3268 	 */
3269 
3270 	global = create_global(gt, 4096);
3271 	if (IS_ERR(global))
3272 		return PTR_ERR(global);
3273 
3274 	result = i915_gem_object_pin_map_unlocked(global->obj, I915_MAP_WC);
3275 	if (IS_ERR(result)) {
3276 		i915_vma_unpin_and_release(&global, 0);
3277 		return PTR_ERR(result);
3278 	}
3279 
3280 	for_each_engine(engine, gt, id) {
3281 		struct i915_request *client[3] = {};
3282 		struct igt_live_test t;
3283 		int i;
3284 
3285 		if (!intel_engine_has_preemption(engine))
3286 			continue;
3287 
3288 		if (GRAPHICS_VER(gt->i915) == 8 && engine->class != RENDER_CLASS)
3289 			continue; /* we need per-context GPR */
3290 
3291 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
3292 			err = -EIO;
3293 			break;
3294 		}
3295 
3296 		memset(result, 0, 4096);
3297 
3298 		for (i = 0; i < ARRAY_SIZE(client); i++) {
3299 			struct i915_request *rq;
3300 
3301 			rq = create_gpr_client(engine, global,
3302 					       NUM_GPR * i * sizeof(u32));
3303 			if (IS_ERR(rq)) {
3304 				err = PTR_ERR(rq);
3305 				goto end_test;
3306 			}
3307 
3308 			client[i] = rq;
3309 		}
3310 
3311 		/* Continuously preempt the set of 3 running contexts */
3312 		for (i = 1; i <= NUM_GPR; i++) {
3313 			err = preempt_user(engine, global, i);
3314 			if (err)
3315 				goto end_test;
3316 		}
3317 
3318 		if (READ_ONCE(result[0]) != NUM_GPR) {
3319 			pr_err("%s: Failed to release semaphore\n",
3320 			       engine->name);
3321 			err = -EIO;
3322 			goto end_test;
3323 		}
3324 
3325 		for (i = 0; i < ARRAY_SIZE(client); i++) {
3326 			int gpr;
3327 
3328 			if (i915_request_wait(client[i], 0, HZ / 2) < 0) {
3329 				err = -ETIME;
3330 				goto end_test;
3331 			}
3332 
3333 			for (gpr = 1; gpr < NUM_GPR; gpr++) {
3334 				if (result[NUM_GPR * i + gpr] != 1) {
3335 					pr_err("%s: Invalid result, client %d, gpr %d, result: %d\n",
3336 					       engine->name,
3337 					       i, gpr, result[NUM_GPR * i + gpr]);
3338 					err = -EINVAL;
3339 					goto end_test;
3340 				}
3341 			}
3342 		}
3343 
3344 end_test:
3345 		for (i = 0; i < ARRAY_SIZE(client); i++) {
3346 			if (!client[i])
3347 				break;
3348 
3349 			i915_request_put(client[i]);
3350 		}
3351 
3352 		/* Flush the semaphores on error */
3353 		smp_store_mb(result[0], -1);
3354 		if (igt_live_test_end(&t))
3355 			err = -EIO;
3356 		if (err)
3357 			break;
3358 	}
3359 
3360 	i915_vma_unpin_and_release(&global, I915_VMA_RELEASE_MAP);
3361 	return err;
3362 }
3363 
3364 static int live_preempt_timeout(void *arg)
3365 {
3366 	struct intel_gt *gt = arg;
3367 	struct i915_gem_context *ctx_hi, *ctx_lo;
3368 	struct igt_spinner spin_lo;
3369 	struct intel_engine_cs *engine;
3370 	enum intel_engine_id id;
3371 	int err = -ENOMEM;
3372 
3373 	/*
3374 	 * Check that we force preemption to occur by cancelling the previous
3375 	 * context if it refuses to yield the GPU.
3376 	 */
3377 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
3378 		return 0;
3379 
3380 	if (!intel_has_reset_engine(gt))
3381 		return 0;
3382 
3383 	ctx_hi = kernel_context(gt->i915, NULL);
3384 	if (!ctx_hi)
3385 		return -ENOMEM;
3386 	ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
3387 
3388 	ctx_lo = kernel_context(gt->i915, NULL);
3389 	if (!ctx_lo)
3390 		goto err_ctx_hi;
3391 	ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
3392 
3393 	if (igt_spinner_init(&spin_lo, gt))
3394 		goto err_ctx_lo;
3395 
3396 	for_each_engine(engine, gt, id) {
3397 		unsigned long saved_timeout;
3398 		struct i915_request *rq;
3399 
3400 		if (!intel_engine_has_preemption(engine))
3401 			continue;
3402 
3403 		rq = spinner_create_request(&spin_lo, ctx_lo, engine,
3404 					    MI_NOOP); /* preemption disabled */
3405 		if (IS_ERR(rq)) {
3406 			err = PTR_ERR(rq);
3407 			goto err_spin_lo;
3408 		}
3409 
3410 		i915_request_add(rq);
3411 		if (!igt_wait_for_spinner(&spin_lo, rq)) {
3412 			intel_gt_set_wedged(gt);
3413 			err = -EIO;
3414 			goto err_spin_lo;
3415 		}
3416 
3417 		rq = igt_request_alloc(ctx_hi, engine);
3418 		if (IS_ERR(rq)) {
3419 			igt_spinner_end(&spin_lo);
3420 			err = PTR_ERR(rq);
3421 			goto err_spin_lo;
3422 		}
3423 
3424 		/* Flush the previous CS ack before changing timeouts */
3425 		while (READ_ONCE(engine->execlists.pending[0]))
3426 			cpu_relax();
3427 
3428 		saved_timeout = engine->props.preempt_timeout_ms;
3429 		engine->props.preempt_timeout_ms = 1; /* in ms, -> 1 jiffie */
3430 
3431 		i915_request_get(rq);
3432 		i915_request_add(rq);
3433 
3434 		intel_engine_flush_submission(engine);
3435 		engine->props.preempt_timeout_ms = saved_timeout;
3436 
3437 		if (i915_request_wait(rq, 0, HZ / 10) < 0) {
3438 			intel_gt_set_wedged(gt);
3439 			i915_request_put(rq);
3440 			err = -ETIME;
3441 			goto err_spin_lo;
3442 		}
3443 
3444 		igt_spinner_end(&spin_lo);
3445 		i915_request_put(rq);
3446 	}
3447 
3448 	err = 0;
3449 err_spin_lo:
3450 	igt_spinner_fini(&spin_lo);
3451 err_ctx_lo:
3452 	kernel_context_close(ctx_lo);
3453 err_ctx_hi:
3454 	kernel_context_close(ctx_hi);
3455 	return err;
3456 }
3457 
3458 static int random_range(struct rnd_state *rnd, int min, int max)
3459 {
3460 	return i915_prandom_u32_max_state(max - min, rnd) + min;
3461 }
3462 
3463 static int random_priority(struct rnd_state *rnd)
3464 {
3465 	return random_range(rnd, I915_PRIORITY_MIN, I915_PRIORITY_MAX);
3466 }
3467 
3468 struct preempt_smoke {
3469 	struct intel_gt *gt;
3470 	struct kthread_work work;
3471 	struct i915_gem_context **contexts;
3472 	struct intel_engine_cs *engine;
3473 	struct drm_i915_gem_object *batch;
3474 	unsigned int ncontext;
3475 	struct rnd_state prng;
3476 	unsigned long count;
3477 	int result;
3478 };
3479 
3480 static struct i915_gem_context *smoke_context(struct preempt_smoke *smoke)
3481 {
3482 	return smoke->contexts[i915_prandom_u32_max_state(smoke->ncontext,
3483 							  &smoke->prng)];
3484 }
3485 
3486 static int smoke_submit(struct preempt_smoke *smoke,
3487 			struct i915_gem_context *ctx, int prio,
3488 			struct drm_i915_gem_object *batch)
3489 {
3490 	struct i915_request *rq;
3491 	struct i915_vma *vma = NULL;
3492 	int err = 0;
3493 
3494 	if (batch) {
3495 		struct i915_address_space *vm;
3496 
3497 		vm = i915_gem_context_get_eb_vm(ctx);
3498 		vma = i915_vma_instance(batch, vm, NULL);
3499 		i915_vm_put(vm);
3500 		if (IS_ERR(vma))
3501 			return PTR_ERR(vma);
3502 
3503 		err = i915_vma_pin(vma, 0, 0, PIN_USER);
3504 		if (err)
3505 			return err;
3506 	}
3507 
3508 	ctx->sched.priority = prio;
3509 
3510 	rq = igt_request_alloc(ctx, smoke->engine);
3511 	if (IS_ERR(rq)) {
3512 		err = PTR_ERR(rq);
3513 		goto unpin;
3514 	}
3515 
3516 	if (vma) {
3517 		i915_vma_lock(vma);
3518 		err = i915_vma_move_to_active(vma, rq, 0);
3519 		if (!err)
3520 			err = rq->engine->emit_bb_start(rq,
3521 							vma->node.start,
3522 							PAGE_SIZE, 0);
3523 		i915_vma_unlock(vma);
3524 	}
3525 
3526 	i915_request_add(rq);
3527 
3528 unpin:
3529 	if (vma)
3530 		i915_vma_unpin(vma);
3531 
3532 	return err;
3533 }
3534 
3535 static void smoke_crescendo_work(struct kthread_work *work)
3536 {
3537 	struct preempt_smoke *smoke = container_of(work, typeof(*smoke), work);
3538 	IGT_TIMEOUT(end_time);
3539 	unsigned long count;
3540 
3541 	count = 0;
3542 	do {
3543 		struct i915_gem_context *ctx = smoke_context(smoke);
3544 
3545 		smoke->result = smoke_submit(smoke, ctx,
3546 					     count % I915_PRIORITY_MAX,
3547 					     smoke->batch);
3548 
3549 		count++;
3550 	} while (!smoke->result && count < smoke->ncontext &&
3551 		 !__igt_timeout(end_time, NULL));
3552 
3553 	smoke->count = count;
3554 }
3555 
3556 static int smoke_crescendo(struct preempt_smoke *smoke, unsigned int flags)
3557 #define BATCH BIT(0)
3558 {
3559 	struct kthread_worker *worker[I915_NUM_ENGINES] = {};
3560 	struct preempt_smoke *arg;
3561 	struct intel_engine_cs *engine;
3562 	enum intel_engine_id id;
3563 	unsigned long count;
3564 	int err = 0;
3565 
3566 	arg = kmalloc_array(I915_NUM_ENGINES, sizeof(*arg), GFP_KERNEL);
3567 	if (!arg)
3568 		return -ENOMEM;
3569 
3570 	memset(arg, 0, I915_NUM_ENGINES * sizeof(*arg));
3571 
3572 	for_each_engine(engine, smoke->gt, id) {
3573 		arg[id] = *smoke;
3574 		arg[id].engine = engine;
3575 		if (!(flags & BATCH))
3576 			arg[id].batch = NULL;
3577 		arg[id].count = 0;
3578 
3579 		worker[id] = kthread_create_worker(0, "igt/smoke:%d", id);
3580 		if (IS_ERR(worker[id])) {
3581 			err = PTR_ERR(worker[id]);
3582 			break;
3583 		}
3584 
3585 		kthread_init_work(&arg[id].work, smoke_crescendo_work);
3586 		kthread_queue_work(worker[id], &arg[id].work);
3587 	}
3588 
3589 	count = 0;
3590 	for_each_engine(engine, smoke->gt, id) {
3591 		if (IS_ERR_OR_NULL(worker[id]))
3592 			continue;
3593 
3594 		kthread_flush_work(&arg[id].work);
3595 		if (arg[id].result && !err)
3596 			err = arg[id].result;
3597 
3598 		count += arg[id].count;
3599 
3600 		kthread_destroy_worker(worker[id]);
3601 	}
3602 
3603 	pr_info("Submitted %lu crescendo:%x requests across %d engines and %d contexts\n",
3604 		count, flags, smoke->gt->info.num_engines, smoke->ncontext);
3605 
3606 	kfree(arg);
3607 	return 0;
3608 }
3609 
3610 static int smoke_random(struct preempt_smoke *smoke, unsigned int flags)
3611 {
3612 	enum intel_engine_id id;
3613 	IGT_TIMEOUT(end_time);
3614 	unsigned long count;
3615 
3616 	count = 0;
3617 	do {
3618 		for_each_engine(smoke->engine, smoke->gt, id) {
3619 			struct i915_gem_context *ctx = smoke_context(smoke);
3620 			int err;
3621 
3622 			err = smoke_submit(smoke,
3623 					   ctx, random_priority(&smoke->prng),
3624 					   flags & BATCH ? smoke->batch : NULL);
3625 			if (err)
3626 				return err;
3627 
3628 			count++;
3629 		}
3630 	} while (count < smoke->ncontext && !__igt_timeout(end_time, NULL));
3631 
3632 	pr_info("Submitted %lu random:%x requests across %d engines and %d contexts\n",
3633 		count, flags, smoke->gt->info.num_engines, smoke->ncontext);
3634 	return 0;
3635 }
3636 
3637 static int live_preempt_smoke(void *arg)
3638 {
3639 	struct preempt_smoke smoke = {
3640 		.gt = arg,
3641 		.prng = I915_RND_STATE_INITIALIZER(i915_selftest.random_seed),
3642 		.ncontext = 256,
3643 	};
3644 	const unsigned int phase[] = { 0, BATCH };
3645 	struct igt_live_test t;
3646 	int err = -ENOMEM;
3647 	u32 *cs;
3648 	int n;
3649 
3650 	smoke.contexts = kmalloc_array(smoke.ncontext,
3651 				       sizeof(*smoke.contexts),
3652 				       GFP_KERNEL);
3653 	if (!smoke.contexts)
3654 		return -ENOMEM;
3655 
3656 	smoke.batch =
3657 		i915_gem_object_create_internal(smoke.gt->i915, PAGE_SIZE);
3658 	if (IS_ERR(smoke.batch)) {
3659 		err = PTR_ERR(smoke.batch);
3660 		goto err_free;
3661 	}
3662 
3663 	cs = i915_gem_object_pin_map_unlocked(smoke.batch, I915_MAP_WB);
3664 	if (IS_ERR(cs)) {
3665 		err = PTR_ERR(cs);
3666 		goto err_batch;
3667 	}
3668 	for (n = 0; n < PAGE_SIZE / sizeof(*cs) - 1; n++)
3669 		cs[n] = MI_ARB_CHECK;
3670 	cs[n] = MI_BATCH_BUFFER_END;
3671 	i915_gem_object_flush_map(smoke.batch);
3672 	i915_gem_object_unpin_map(smoke.batch);
3673 
3674 	if (igt_live_test_begin(&t, smoke.gt->i915, __func__, "all")) {
3675 		err = -EIO;
3676 		goto err_batch;
3677 	}
3678 
3679 	for (n = 0; n < smoke.ncontext; n++) {
3680 		smoke.contexts[n] = kernel_context(smoke.gt->i915, NULL);
3681 		if (!smoke.contexts[n])
3682 			goto err_ctx;
3683 	}
3684 
3685 	for (n = 0; n < ARRAY_SIZE(phase); n++) {
3686 		err = smoke_crescendo(&smoke, phase[n]);
3687 		if (err)
3688 			goto err_ctx;
3689 
3690 		err = smoke_random(&smoke, phase[n]);
3691 		if (err)
3692 			goto err_ctx;
3693 	}
3694 
3695 err_ctx:
3696 	if (igt_live_test_end(&t))
3697 		err = -EIO;
3698 
3699 	for (n = 0; n < smoke.ncontext; n++) {
3700 		if (!smoke.contexts[n])
3701 			break;
3702 		kernel_context_close(smoke.contexts[n]);
3703 	}
3704 
3705 err_batch:
3706 	i915_gem_object_put(smoke.batch);
3707 err_free:
3708 	kfree(smoke.contexts);
3709 
3710 	return err;
3711 }
3712 
3713 static int nop_virtual_engine(struct intel_gt *gt,
3714 			      struct intel_engine_cs **siblings,
3715 			      unsigned int nsibling,
3716 			      unsigned int nctx,
3717 			      unsigned int flags)
3718 #define CHAIN BIT(0)
3719 {
3720 	IGT_TIMEOUT(end_time);
3721 	struct i915_request *request[16] = {};
3722 	struct intel_context *ve[16];
3723 	unsigned long n, prime, nc;
3724 	struct igt_live_test t;
3725 	ktime_t times[2] = {};
3726 	int err;
3727 
3728 	GEM_BUG_ON(!nctx || nctx > ARRAY_SIZE(ve));
3729 
3730 	for (n = 0; n < nctx; n++) {
3731 		ve[n] = intel_engine_create_virtual(siblings, nsibling, 0);
3732 		if (IS_ERR(ve[n])) {
3733 			err = PTR_ERR(ve[n]);
3734 			nctx = n;
3735 			goto out;
3736 		}
3737 
3738 		err = intel_context_pin(ve[n]);
3739 		if (err) {
3740 			intel_context_put(ve[n]);
3741 			nctx = n;
3742 			goto out;
3743 		}
3744 	}
3745 
3746 	err = igt_live_test_begin(&t, gt->i915, __func__, ve[0]->engine->name);
3747 	if (err)
3748 		goto out;
3749 
3750 	for_each_prime_number_from(prime, 1, 8192) {
3751 		times[1] = ktime_get_raw();
3752 
3753 		if (flags & CHAIN) {
3754 			for (nc = 0; nc < nctx; nc++) {
3755 				for (n = 0; n < prime; n++) {
3756 					struct i915_request *rq;
3757 
3758 					rq = i915_request_create(ve[nc]);
3759 					if (IS_ERR(rq)) {
3760 						err = PTR_ERR(rq);
3761 						goto out;
3762 					}
3763 
3764 					if (request[nc])
3765 						i915_request_put(request[nc]);
3766 					request[nc] = i915_request_get(rq);
3767 					i915_request_add(rq);
3768 				}
3769 			}
3770 		} else {
3771 			for (n = 0; n < prime; n++) {
3772 				for (nc = 0; nc < nctx; nc++) {
3773 					struct i915_request *rq;
3774 
3775 					rq = i915_request_create(ve[nc]);
3776 					if (IS_ERR(rq)) {
3777 						err = PTR_ERR(rq);
3778 						goto out;
3779 					}
3780 
3781 					if (request[nc])
3782 						i915_request_put(request[nc]);
3783 					request[nc] = i915_request_get(rq);
3784 					i915_request_add(rq);
3785 				}
3786 			}
3787 		}
3788 
3789 		for (nc = 0; nc < nctx; nc++) {
3790 			if (i915_request_wait(request[nc], 0, HZ / 10) < 0) {
3791 				pr_err("%s(%s): wait for %llx:%lld timed out\n",
3792 				       __func__, ve[0]->engine->name,
3793 				       request[nc]->fence.context,
3794 				       request[nc]->fence.seqno);
3795 
3796 				GEM_TRACE("%s(%s) failed at request %llx:%lld\n",
3797 					  __func__, ve[0]->engine->name,
3798 					  request[nc]->fence.context,
3799 					  request[nc]->fence.seqno);
3800 				GEM_TRACE_DUMP();
3801 				intel_gt_set_wedged(gt);
3802 				break;
3803 			}
3804 		}
3805 
3806 		times[1] = ktime_sub(ktime_get_raw(), times[1]);
3807 		if (prime == 1)
3808 			times[0] = times[1];
3809 
3810 		for (nc = 0; nc < nctx; nc++) {
3811 			i915_request_put(request[nc]);
3812 			request[nc] = NULL;
3813 		}
3814 
3815 		if (__igt_timeout(end_time, NULL))
3816 			break;
3817 	}
3818 
3819 	err = igt_live_test_end(&t);
3820 	if (err)
3821 		goto out;
3822 
3823 	pr_info("Requestx%d latencies on %s: 1 = %lluns, %lu = %lluns\n",
3824 		nctx, ve[0]->engine->name, ktime_to_ns(times[0]),
3825 		prime, div64_u64(ktime_to_ns(times[1]), prime));
3826 
3827 out:
3828 	if (igt_flush_test(gt->i915))
3829 		err = -EIO;
3830 
3831 	for (nc = 0; nc < nctx; nc++) {
3832 		i915_request_put(request[nc]);
3833 		intel_context_unpin(ve[nc]);
3834 		intel_context_put(ve[nc]);
3835 	}
3836 	return err;
3837 }
3838 
3839 static unsigned int
3840 __select_siblings(struct intel_gt *gt,
3841 		  unsigned int class,
3842 		  struct intel_engine_cs **siblings,
3843 		  bool (*filter)(const struct intel_engine_cs *))
3844 {
3845 	unsigned int n = 0;
3846 	unsigned int inst;
3847 
3848 	for (inst = 0; inst <= MAX_ENGINE_INSTANCE; inst++) {
3849 		if (!gt->engine_class[class][inst])
3850 			continue;
3851 
3852 		if (filter && !filter(gt->engine_class[class][inst]))
3853 			continue;
3854 
3855 		siblings[n++] = gt->engine_class[class][inst];
3856 	}
3857 
3858 	return n;
3859 }
3860 
3861 static unsigned int
3862 select_siblings(struct intel_gt *gt,
3863 		unsigned int class,
3864 		struct intel_engine_cs **siblings)
3865 {
3866 	return __select_siblings(gt, class, siblings, NULL);
3867 }
3868 
3869 static int live_virtual_engine(void *arg)
3870 {
3871 	struct intel_gt *gt = arg;
3872 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
3873 	struct intel_engine_cs *engine;
3874 	enum intel_engine_id id;
3875 	unsigned int class;
3876 	int err;
3877 
3878 	if (intel_uc_uses_guc_submission(&gt->uc))
3879 		return 0;
3880 
3881 	for_each_engine(engine, gt, id) {
3882 		err = nop_virtual_engine(gt, &engine, 1, 1, 0);
3883 		if (err) {
3884 			pr_err("Failed to wrap engine %s: err=%d\n",
3885 			       engine->name, err);
3886 			return err;
3887 		}
3888 	}
3889 
3890 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
3891 		int nsibling, n;
3892 
3893 		nsibling = select_siblings(gt, class, siblings);
3894 		if (nsibling < 2)
3895 			continue;
3896 
3897 		for (n = 1; n <= nsibling + 1; n++) {
3898 			err = nop_virtual_engine(gt, siblings, nsibling,
3899 						 n, 0);
3900 			if (err)
3901 				return err;
3902 		}
3903 
3904 		err = nop_virtual_engine(gt, siblings, nsibling, n, CHAIN);
3905 		if (err)
3906 			return err;
3907 	}
3908 
3909 	return 0;
3910 }
3911 
3912 static int mask_virtual_engine(struct intel_gt *gt,
3913 			       struct intel_engine_cs **siblings,
3914 			       unsigned int nsibling)
3915 {
3916 	struct i915_request *request[MAX_ENGINE_INSTANCE + 1];
3917 	struct intel_context *ve;
3918 	struct igt_live_test t;
3919 	unsigned int n;
3920 	int err;
3921 
3922 	/*
3923 	 * Check that by setting the execution mask on a request, we can
3924 	 * restrict it to our desired engine within the virtual engine.
3925 	 */
3926 
3927 	ve = intel_engine_create_virtual(siblings, nsibling, 0);
3928 	if (IS_ERR(ve)) {
3929 		err = PTR_ERR(ve);
3930 		goto out_close;
3931 	}
3932 
3933 	err = intel_context_pin(ve);
3934 	if (err)
3935 		goto out_put;
3936 
3937 	err = igt_live_test_begin(&t, gt->i915, __func__, ve->engine->name);
3938 	if (err)
3939 		goto out_unpin;
3940 
3941 	for (n = 0; n < nsibling; n++) {
3942 		request[n] = i915_request_create(ve);
3943 		if (IS_ERR(request[n])) {
3944 			err = PTR_ERR(request[n]);
3945 			nsibling = n;
3946 			goto out;
3947 		}
3948 
3949 		/* Reverse order as it's more likely to be unnatural */
3950 		request[n]->execution_mask = siblings[nsibling - n - 1]->mask;
3951 
3952 		i915_request_get(request[n]);
3953 		i915_request_add(request[n]);
3954 	}
3955 
3956 	for (n = 0; n < nsibling; n++) {
3957 		if (i915_request_wait(request[n], 0, HZ / 10) < 0) {
3958 			pr_err("%s(%s): wait for %llx:%lld timed out\n",
3959 			       __func__, ve->engine->name,
3960 			       request[n]->fence.context,
3961 			       request[n]->fence.seqno);
3962 
3963 			GEM_TRACE("%s(%s) failed at request %llx:%lld\n",
3964 				  __func__, ve->engine->name,
3965 				  request[n]->fence.context,
3966 				  request[n]->fence.seqno);
3967 			GEM_TRACE_DUMP();
3968 			intel_gt_set_wedged(gt);
3969 			err = -EIO;
3970 			goto out;
3971 		}
3972 
3973 		if (request[n]->engine != siblings[nsibling - n - 1]) {
3974 			pr_err("Executed on wrong sibling '%s', expected '%s'\n",
3975 			       request[n]->engine->name,
3976 			       siblings[nsibling - n - 1]->name);
3977 			err = -EINVAL;
3978 			goto out;
3979 		}
3980 	}
3981 
3982 	err = igt_live_test_end(&t);
3983 out:
3984 	if (igt_flush_test(gt->i915))
3985 		err = -EIO;
3986 
3987 	for (n = 0; n < nsibling; n++)
3988 		i915_request_put(request[n]);
3989 
3990 out_unpin:
3991 	intel_context_unpin(ve);
3992 out_put:
3993 	intel_context_put(ve);
3994 out_close:
3995 	return err;
3996 }
3997 
3998 static int live_virtual_mask(void *arg)
3999 {
4000 	struct intel_gt *gt = arg;
4001 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4002 	unsigned int class;
4003 	int err;
4004 
4005 	if (intel_uc_uses_guc_submission(&gt->uc))
4006 		return 0;
4007 
4008 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4009 		unsigned int nsibling;
4010 
4011 		nsibling = select_siblings(gt, class, siblings);
4012 		if (nsibling < 2)
4013 			continue;
4014 
4015 		err = mask_virtual_engine(gt, siblings, nsibling);
4016 		if (err)
4017 			return err;
4018 	}
4019 
4020 	return 0;
4021 }
4022 
4023 static int slicein_virtual_engine(struct intel_gt *gt,
4024 				  struct intel_engine_cs **siblings,
4025 				  unsigned int nsibling)
4026 {
4027 	const long timeout = slice_timeout(siblings[0]);
4028 	struct intel_context *ce;
4029 	struct i915_request *rq;
4030 	struct igt_spinner spin;
4031 	unsigned int n;
4032 	int err = 0;
4033 
4034 	/*
4035 	 * Virtual requests must take part in timeslicing on the target engines.
4036 	 */
4037 
4038 	if (igt_spinner_init(&spin, gt))
4039 		return -ENOMEM;
4040 
4041 	for (n = 0; n < nsibling; n++) {
4042 		ce = intel_context_create(siblings[n]);
4043 		if (IS_ERR(ce)) {
4044 			err = PTR_ERR(ce);
4045 			goto out;
4046 		}
4047 
4048 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
4049 		intel_context_put(ce);
4050 		if (IS_ERR(rq)) {
4051 			err = PTR_ERR(rq);
4052 			goto out;
4053 		}
4054 
4055 		i915_request_add(rq);
4056 	}
4057 
4058 	ce = intel_engine_create_virtual(siblings, nsibling, 0);
4059 	if (IS_ERR(ce)) {
4060 		err = PTR_ERR(ce);
4061 		goto out;
4062 	}
4063 
4064 	rq = intel_context_create_request(ce);
4065 	intel_context_put(ce);
4066 	if (IS_ERR(rq)) {
4067 		err = PTR_ERR(rq);
4068 		goto out;
4069 	}
4070 
4071 	i915_request_get(rq);
4072 	i915_request_add(rq);
4073 	if (i915_request_wait(rq, 0, timeout) < 0) {
4074 		GEM_TRACE_ERR("%s(%s) failed to slice in virtual request\n",
4075 			      __func__, rq->engine->name);
4076 		GEM_TRACE_DUMP();
4077 		intel_gt_set_wedged(gt);
4078 		err = -EIO;
4079 	}
4080 	i915_request_put(rq);
4081 
4082 out:
4083 	igt_spinner_end(&spin);
4084 	if (igt_flush_test(gt->i915))
4085 		err = -EIO;
4086 	igt_spinner_fini(&spin);
4087 	return err;
4088 }
4089 
4090 static int sliceout_virtual_engine(struct intel_gt *gt,
4091 				   struct intel_engine_cs **siblings,
4092 				   unsigned int nsibling)
4093 {
4094 	const long timeout = slice_timeout(siblings[0]);
4095 	struct intel_context *ce;
4096 	struct i915_request *rq;
4097 	struct igt_spinner spin;
4098 	unsigned int n;
4099 	int err = 0;
4100 
4101 	/*
4102 	 * Virtual requests must allow others a fair timeslice.
4103 	 */
4104 
4105 	if (igt_spinner_init(&spin, gt))
4106 		return -ENOMEM;
4107 
4108 	/* XXX We do not handle oversubscription and fairness with normal rq */
4109 	for (n = 0; n < nsibling; n++) {
4110 		ce = intel_engine_create_virtual(siblings, nsibling, 0);
4111 		if (IS_ERR(ce)) {
4112 			err = PTR_ERR(ce);
4113 			goto out;
4114 		}
4115 
4116 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
4117 		intel_context_put(ce);
4118 		if (IS_ERR(rq)) {
4119 			err = PTR_ERR(rq);
4120 			goto out;
4121 		}
4122 
4123 		i915_request_add(rq);
4124 	}
4125 
4126 	for (n = 0; !err && n < nsibling; n++) {
4127 		ce = intel_context_create(siblings[n]);
4128 		if (IS_ERR(ce)) {
4129 			err = PTR_ERR(ce);
4130 			goto out;
4131 		}
4132 
4133 		rq = intel_context_create_request(ce);
4134 		intel_context_put(ce);
4135 		if (IS_ERR(rq)) {
4136 			err = PTR_ERR(rq);
4137 			goto out;
4138 		}
4139 
4140 		i915_request_get(rq);
4141 		i915_request_add(rq);
4142 		if (i915_request_wait(rq, 0, timeout) < 0) {
4143 			GEM_TRACE_ERR("%s(%s) failed to slice out virtual request\n",
4144 				      __func__, siblings[n]->name);
4145 			GEM_TRACE_DUMP();
4146 			intel_gt_set_wedged(gt);
4147 			err = -EIO;
4148 		}
4149 		i915_request_put(rq);
4150 	}
4151 
4152 out:
4153 	igt_spinner_end(&spin);
4154 	if (igt_flush_test(gt->i915))
4155 		err = -EIO;
4156 	igt_spinner_fini(&spin);
4157 	return err;
4158 }
4159 
4160 static int live_virtual_slice(void *arg)
4161 {
4162 	struct intel_gt *gt = arg;
4163 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4164 	unsigned int class;
4165 	int err;
4166 
4167 	if (intel_uc_uses_guc_submission(&gt->uc))
4168 		return 0;
4169 
4170 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4171 		unsigned int nsibling;
4172 
4173 		nsibling = __select_siblings(gt, class, siblings,
4174 					     intel_engine_has_timeslices);
4175 		if (nsibling < 2)
4176 			continue;
4177 
4178 		err = slicein_virtual_engine(gt, siblings, nsibling);
4179 		if (err)
4180 			return err;
4181 
4182 		err = sliceout_virtual_engine(gt, siblings, nsibling);
4183 		if (err)
4184 			return err;
4185 	}
4186 
4187 	return 0;
4188 }
4189 
4190 static int preserved_virtual_engine(struct intel_gt *gt,
4191 				    struct intel_engine_cs **siblings,
4192 				    unsigned int nsibling)
4193 {
4194 	struct i915_request *last = NULL;
4195 	struct intel_context *ve;
4196 	struct i915_vma *scratch;
4197 	struct igt_live_test t;
4198 	unsigned int n;
4199 	int err = 0;
4200 	u32 *cs;
4201 
4202 	scratch =
4203 		__vm_create_scratch_for_read_pinned(&siblings[0]->gt->ggtt->vm,
4204 						    PAGE_SIZE);
4205 	if (IS_ERR(scratch))
4206 		return PTR_ERR(scratch);
4207 
4208 	err = i915_vma_sync(scratch);
4209 	if (err)
4210 		goto out_scratch;
4211 
4212 	ve = intel_engine_create_virtual(siblings, nsibling, 0);
4213 	if (IS_ERR(ve)) {
4214 		err = PTR_ERR(ve);
4215 		goto out_scratch;
4216 	}
4217 
4218 	err = intel_context_pin(ve);
4219 	if (err)
4220 		goto out_put;
4221 
4222 	err = igt_live_test_begin(&t, gt->i915, __func__, ve->engine->name);
4223 	if (err)
4224 		goto out_unpin;
4225 
4226 	for (n = 0; n < NUM_GPR_DW; n++) {
4227 		struct intel_engine_cs *engine = siblings[n % nsibling];
4228 		struct i915_request *rq;
4229 
4230 		rq = i915_request_create(ve);
4231 		if (IS_ERR(rq)) {
4232 			err = PTR_ERR(rq);
4233 			goto out_end;
4234 		}
4235 
4236 		i915_request_put(last);
4237 		last = i915_request_get(rq);
4238 
4239 		cs = intel_ring_begin(rq, 8);
4240 		if (IS_ERR(cs)) {
4241 			i915_request_add(rq);
4242 			err = PTR_ERR(cs);
4243 			goto out_end;
4244 		}
4245 
4246 		*cs++ = MI_STORE_REGISTER_MEM_GEN8 | MI_USE_GGTT;
4247 		*cs++ = CS_GPR(engine, n);
4248 		*cs++ = i915_ggtt_offset(scratch) + n * sizeof(u32);
4249 		*cs++ = 0;
4250 
4251 		*cs++ = MI_LOAD_REGISTER_IMM(1);
4252 		*cs++ = CS_GPR(engine, (n + 1) % NUM_GPR_DW);
4253 		*cs++ = n + 1;
4254 
4255 		*cs++ = MI_NOOP;
4256 		intel_ring_advance(rq, cs);
4257 
4258 		/* Restrict this request to run on a particular engine */
4259 		rq->execution_mask = engine->mask;
4260 		i915_request_add(rq);
4261 	}
4262 
4263 	if (i915_request_wait(last, 0, HZ / 5) < 0) {
4264 		err = -ETIME;
4265 		goto out_end;
4266 	}
4267 
4268 	cs = i915_gem_object_pin_map_unlocked(scratch->obj, I915_MAP_WB);
4269 	if (IS_ERR(cs)) {
4270 		err = PTR_ERR(cs);
4271 		goto out_end;
4272 	}
4273 
4274 	for (n = 0; n < NUM_GPR_DW; n++) {
4275 		if (cs[n] != n) {
4276 			pr_err("Incorrect value[%d] found for GPR[%d]\n",
4277 			       cs[n], n);
4278 			err = -EINVAL;
4279 			break;
4280 		}
4281 	}
4282 
4283 	i915_gem_object_unpin_map(scratch->obj);
4284 
4285 out_end:
4286 	if (igt_live_test_end(&t))
4287 		err = -EIO;
4288 	i915_request_put(last);
4289 out_unpin:
4290 	intel_context_unpin(ve);
4291 out_put:
4292 	intel_context_put(ve);
4293 out_scratch:
4294 	i915_vma_unpin_and_release(&scratch, 0);
4295 	return err;
4296 }
4297 
4298 static int live_virtual_preserved(void *arg)
4299 {
4300 	struct intel_gt *gt = arg;
4301 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4302 	unsigned int class;
4303 
4304 	/*
4305 	 * Check that the context image retains non-privileged (user) registers
4306 	 * from one engine to the next. For this we check that the CS_GPR
4307 	 * are preserved.
4308 	 */
4309 
4310 	if (intel_uc_uses_guc_submission(&gt->uc))
4311 		return 0;
4312 
4313 	/* As we use CS_GPR we cannot run before they existed on all engines. */
4314 	if (GRAPHICS_VER(gt->i915) < 9)
4315 		return 0;
4316 
4317 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4318 		int nsibling, err;
4319 
4320 		nsibling = select_siblings(gt, class, siblings);
4321 		if (nsibling < 2)
4322 			continue;
4323 
4324 		err = preserved_virtual_engine(gt, siblings, nsibling);
4325 		if (err)
4326 			return err;
4327 	}
4328 
4329 	return 0;
4330 }
4331 
4332 static int reset_virtual_engine(struct intel_gt *gt,
4333 				struct intel_engine_cs **siblings,
4334 				unsigned int nsibling)
4335 {
4336 	struct intel_engine_cs *engine;
4337 	struct intel_context *ve;
4338 	struct igt_spinner spin;
4339 	struct i915_request *rq;
4340 	unsigned int n;
4341 	int err = 0;
4342 
4343 	/*
4344 	 * In order to support offline error capture for fast preempt reset,
4345 	 * we need to decouple the guilty request and ensure that it and its
4346 	 * descendents are not executed while the capture is in progress.
4347 	 */
4348 
4349 	if (igt_spinner_init(&spin, gt))
4350 		return -ENOMEM;
4351 
4352 	ve = intel_engine_create_virtual(siblings, nsibling, 0);
4353 	if (IS_ERR(ve)) {
4354 		err = PTR_ERR(ve);
4355 		goto out_spin;
4356 	}
4357 
4358 	for (n = 0; n < nsibling; n++)
4359 		st_engine_heartbeat_disable(siblings[n]);
4360 
4361 	rq = igt_spinner_create_request(&spin, ve, MI_ARB_CHECK);
4362 	if (IS_ERR(rq)) {
4363 		err = PTR_ERR(rq);
4364 		goto out_heartbeat;
4365 	}
4366 	i915_request_add(rq);
4367 
4368 	if (!igt_wait_for_spinner(&spin, rq)) {
4369 		intel_gt_set_wedged(gt);
4370 		err = -ETIME;
4371 		goto out_heartbeat;
4372 	}
4373 
4374 	engine = rq->engine;
4375 	GEM_BUG_ON(engine == ve->engine);
4376 
4377 	/* Take ownership of the reset and tasklet */
4378 	err = engine_lock_reset_tasklet(engine);
4379 	if (err)
4380 		goto out_heartbeat;
4381 
4382 	engine->sched_engine->tasklet.callback(&engine->sched_engine->tasklet);
4383 	GEM_BUG_ON(execlists_active(&engine->execlists) != rq);
4384 
4385 	/* Fake a preemption event; failed of course */
4386 	spin_lock_irq(&engine->sched_engine->lock);
4387 	__unwind_incomplete_requests(engine);
4388 	spin_unlock_irq(&engine->sched_engine->lock);
4389 	GEM_BUG_ON(rq->engine != engine);
4390 
4391 	/* Reset the engine while keeping our active request on hold */
4392 	execlists_hold(engine, rq);
4393 	GEM_BUG_ON(!i915_request_on_hold(rq));
4394 
4395 	__intel_engine_reset_bh(engine, NULL);
4396 	GEM_BUG_ON(rq->fence.error != -EIO);
4397 
4398 	/* Release our grasp on the engine, letting CS flow again */
4399 	engine_unlock_reset_tasklet(engine);
4400 
4401 	/* Check that we do not resubmit the held request */
4402 	i915_request_get(rq);
4403 	if (!i915_request_wait(rq, 0, HZ / 5)) {
4404 		pr_err("%s: on hold request completed!\n",
4405 		       engine->name);
4406 		intel_gt_set_wedged(gt);
4407 		err = -EIO;
4408 		goto out_rq;
4409 	}
4410 	GEM_BUG_ON(!i915_request_on_hold(rq));
4411 
4412 	/* But is resubmitted on release */
4413 	execlists_unhold(engine, rq);
4414 	if (i915_request_wait(rq, 0, HZ / 5) < 0) {
4415 		pr_err("%s: held request did not complete!\n",
4416 		       engine->name);
4417 		intel_gt_set_wedged(gt);
4418 		err = -ETIME;
4419 	}
4420 
4421 out_rq:
4422 	i915_request_put(rq);
4423 out_heartbeat:
4424 	for (n = 0; n < nsibling; n++)
4425 		st_engine_heartbeat_enable(siblings[n]);
4426 
4427 	intel_context_put(ve);
4428 out_spin:
4429 	igt_spinner_fini(&spin);
4430 	return err;
4431 }
4432 
4433 static int live_virtual_reset(void *arg)
4434 {
4435 	struct intel_gt *gt = arg;
4436 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4437 	unsigned int class;
4438 
4439 	/*
4440 	 * Check that we handle a reset event within a virtual engine.
4441 	 * Only the physical engine is reset, but we have to check the flow
4442 	 * of the virtual requests around the reset, and make sure it is not
4443 	 * forgotten.
4444 	 */
4445 
4446 	if (intel_uc_uses_guc_submission(&gt->uc))
4447 		return 0;
4448 
4449 	if (!intel_has_reset_engine(gt))
4450 		return 0;
4451 
4452 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4453 		int nsibling, err;
4454 
4455 		nsibling = select_siblings(gt, class, siblings);
4456 		if (nsibling < 2)
4457 			continue;
4458 
4459 		err = reset_virtual_engine(gt, siblings, nsibling);
4460 		if (err)
4461 			return err;
4462 	}
4463 
4464 	return 0;
4465 }
4466 
4467 int intel_execlists_live_selftests(struct drm_i915_private *i915)
4468 {
4469 	static const struct i915_subtest tests[] = {
4470 		SUBTEST(live_sanitycheck),
4471 		SUBTEST(live_unlite_switch),
4472 		SUBTEST(live_unlite_preempt),
4473 		SUBTEST(live_unlite_ring),
4474 		SUBTEST(live_pin_rewind),
4475 		SUBTEST(live_hold_reset),
4476 		SUBTEST(live_error_interrupt),
4477 		SUBTEST(live_timeslice_preempt),
4478 		SUBTEST(live_timeslice_rewind),
4479 		SUBTEST(live_timeslice_queue),
4480 		SUBTEST(live_timeslice_nopreempt),
4481 		SUBTEST(live_busywait_preempt),
4482 		SUBTEST(live_preempt),
4483 		SUBTEST(live_late_preempt),
4484 		SUBTEST(live_nopreempt),
4485 		SUBTEST(live_preempt_cancel),
4486 		SUBTEST(live_suppress_self_preempt),
4487 		SUBTEST(live_chain_preempt),
4488 		SUBTEST(live_preempt_ring),
4489 		SUBTEST(live_preempt_gang),
4490 		SUBTEST(live_preempt_timeout),
4491 		SUBTEST(live_preempt_user),
4492 		SUBTEST(live_preempt_smoke),
4493 		SUBTEST(live_virtual_engine),
4494 		SUBTEST(live_virtual_mask),
4495 		SUBTEST(live_virtual_preserved),
4496 		SUBTEST(live_virtual_slice),
4497 		SUBTEST(live_virtual_reset),
4498 	};
4499 
4500 	if (to_gt(i915)->submission_method != INTEL_SUBMISSION_ELSP)
4501 		return 0;
4502 
4503 	if (intel_gt_is_wedged(to_gt(i915)))
4504 		return 0;
4505 
4506 	return intel_gt_live_subtests(tests, to_gt(i915));
4507 }
4508