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 	flush_scheduled_work();
89 
90 	if (rq->fence.error != -EIO) {
91 		pr_err("%s: hanging request %llx:%lld not reset\n",
92 		       engine->name,
93 		       rq->fence.context,
94 		       rq->fence.seqno);
95 		return -EINVAL;
96 	}
97 
98 	/* Give the request a jiffie to complete after flushing the worker */
99 	if (i915_request_wait(rq, 0,
100 			      max(0l, (long)(timeout - jiffies)) + 1) < 0) {
101 		pr_err("%s: hanging request %llx:%lld did not complete\n",
102 		       engine->name,
103 		       rq->fence.context,
104 		       rq->fence.seqno);
105 		return -ETIME;
106 	}
107 
108 	return 0;
109 }
110 
111 static int live_sanitycheck(void *arg)
112 {
113 	struct intel_gt *gt = arg;
114 	struct intel_engine_cs *engine;
115 	enum intel_engine_id id;
116 	struct igt_spinner spin;
117 	int err = 0;
118 
119 	if (!HAS_LOGICAL_RING_CONTEXTS(gt->i915))
120 		return 0;
121 
122 	if (igt_spinner_init(&spin, gt))
123 		return -ENOMEM;
124 
125 	for_each_engine(engine, gt, id) {
126 		struct intel_context *ce;
127 		struct i915_request *rq;
128 
129 		ce = intel_context_create(engine);
130 		if (IS_ERR(ce)) {
131 			err = PTR_ERR(ce);
132 			break;
133 		}
134 
135 		rq = igt_spinner_create_request(&spin, ce, MI_NOOP);
136 		if (IS_ERR(rq)) {
137 			err = PTR_ERR(rq);
138 			goto out_ctx;
139 		}
140 
141 		i915_request_add(rq);
142 		if (!igt_wait_for_spinner(&spin, rq)) {
143 			GEM_TRACE("spinner failed to start\n");
144 			GEM_TRACE_DUMP();
145 			intel_gt_set_wedged(gt);
146 			err = -EIO;
147 			goto out_ctx;
148 		}
149 
150 		igt_spinner_end(&spin);
151 		if (igt_flush_test(gt->i915)) {
152 			err = -EIO;
153 			goto out_ctx;
154 		}
155 
156 out_ctx:
157 		intel_context_put(ce);
158 		if (err)
159 			break;
160 	}
161 
162 	igt_spinner_fini(&spin);
163 	return err;
164 }
165 
166 static int live_unlite_restore(struct intel_gt *gt, int prio)
167 {
168 	struct intel_engine_cs *engine;
169 	enum intel_engine_id id;
170 	struct igt_spinner spin;
171 	int err = -ENOMEM;
172 
173 	/*
174 	 * Check that we can correctly context switch between 2 instances
175 	 * on the same engine from the same parent context.
176 	 */
177 
178 	if (igt_spinner_init(&spin, gt))
179 		return err;
180 
181 	err = 0;
182 	for_each_engine(engine, gt, id) {
183 		struct intel_context *ce[2] = {};
184 		struct i915_request *rq[2];
185 		struct igt_live_test t;
186 		int n;
187 
188 		if (prio && !intel_engine_has_preemption(engine))
189 			continue;
190 
191 		if (!intel_engine_can_store_dword(engine))
192 			continue;
193 
194 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
195 			err = -EIO;
196 			break;
197 		}
198 		st_engine_heartbeat_disable(engine);
199 
200 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
201 			struct intel_context *tmp;
202 
203 			tmp = intel_context_create(engine);
204 			if (IS_ERR(tmp)) {
205 				err = PTR_ERR(tmp);
206 				goto err_ce;
207 			}
208 
209 			err = intel_context_pin(tmp);
210 			if (err) {
211 				intel_context_put(tmp);
212 				goto err_ce;
213 			}
214 
215 			/*
216 			 * Setup the pair of contexts such that if we
217 			 * lite-restore using the RING_TAIL from ce[1] it
218 			 * will execute garbage from ce[0]->ring.
219 			 */
220 			memset(tmp->ring->vaddr,
221 			       POISON_INUSE, /* IPEHR: 0x5a5a5a5a [hung!] */
222 			       tmp->ring->vma->size);
223 
224 			ce[n] = tmp;
225 		}
226 		GEM_BUG_ON(!ce[1]->ring->size);
227 		intel_ring_reset(ce[1]->ring, ce[1]->ring->size / 2);
228 		lrc_update_regs(ce[1], engine, ce[1]->ring->head);
229 
230 		rq[0] = igt_spinner_create_request(&spin, ce[0], MI_ARB_CHECK);
231 		if (IS_ERR(rq[0])) {
232 			err = PTR_ERR(rq[0]);
233 			goto err_ce;
234 		}
235 
236 		i915_request_get(rq[0]);
237 		i915_request_add(rq[0]);
238 		GEM_BUG_ON(rq[0]->postfix > ce[1]->ring->emit);
239 
240 		if (!igt_wait_for_spinner(&spin, rq[0])) {
241 			i915_request_put(rq[0]);
242 			goto err_ce;
243 		}
244 
245 		rq[1] = i915_request_create(ce[1]);
246 		if (IS_ERR(rq[1])) {
247 			err = PTR_ERR(rq[1]);
248 			i915_request_put(rq[0]);
249 			goto err_ce;
250 		}
251 
252 		if (!prio) {
253 			/*
254 			 * Ensure we do the switch to ce[1] on completion.
255 			 *
256 			 * rq[0] is already submitted, so this should reduce
257 			 * to a no-op (a wait on a request on the same engine
258 			 * uses the submit fence, not the completion fence),
259 			 * but it will install a dependency on rq[1] for rq[0]
260 			 * that will prevent the pair being reordered by
261 			 * timeslicing.
262 			 */
263 			i915_request_await_dma_fence(rq[1], &rq[0]->fence);
264 		}
265 
266 		i915_request_get(rq[1]);
267 		i915_request_add(rq[1]);
268 		GEM_BUG_ON(rq[1]->postfix <= rq[0]->postfix);
269 		i915_request_put(rq[0]);
270 
271 		if (prio) {
272 			struct i915_sched_attr attr = {
273 				.priority = prio,
274 			};
275 
276 			/* Alternatively preempt the spinner with ce[1] */
277 			engine->sched_engine->schedule(rq[1], &attr);
278 		}
279 
280 		/* And switch back to ce[0] for good measure */
281 		rq[0] = i915_request_create(ce[0]);
282 		if (IS_ERR(rq[0])) {
283 			err = PTR_ERR(rq[0]);
284 			i915_request_put(rq[1]);
285 			goto err_ce;
286 		}
287 
288 		i915_request_await_dma_fence(rq[0], &rq[1]->fence);
289 		i915_request_get(rq[0]);
290 		i915_request_add(rq[0]);
291 		GEM_BUG_ON(rq[0]->postfix > rq[1]->postfix);
292 		i915_request_put(rq[1]);
293 		i915_request_put(rq[0]);
294 
295 err_ce:
296 		intel_engine_flush_submission(engine);
297 		igt_spinner_end(&spin);
298 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
299 			if (IS_ERR_OR_NULL(ce[n]))
300 				break;
301 
302 			intel_context_unpin(ce[n]);
303 			intel_context_put(ce[n]);
304 		}
305 
306 		st_engine_heartbeat_enable(engine);
307 		if (igt_live_test_end(&t))
308 			err = -EIO;
309 		if (err)
310 			break;
311 	}
312 
313 	igt_spinner_fini(&spin);
314 	return err;
315 }
316 
317 static int live_unlite_switch(void *arg)
318 {
319 	return live_unlite_restore(arg, 0);
320 }
321 
322 static int live_unlite_preempt(void *arg)
323 {
324 	return live_unlite_restore(arg, I915_PRIORITY_MAX);
325 }
326 
327 static int live_unlite_ring(void *arg)
328 {
329 	struct intel_gt *gt = arg;
330 	struct intel_engine_cs *engine;
331 	struct igt_spinner spin;
332 	enum intel_engine_id id;
333 	int err = 0;
334 
335 	/*
336 	 * Setup a preemption event that will cause almost the entire ring
337 	 * to be unwound, potentially fooling our intel_ring_direction()
338 	 * into emitting a forward lite-restore instead of the rollback.
339 	 */
340 
341 	if (igt_spinner_init(&spin, gt))
342 		return -ENOMEM;
343 
344 	for_each_engine(engine, gt, id) {
345 		struct intel_context *ce[2] = {};
346 		struct i915_request *rq;
347 		struct igt_live_test t;
348 		int n;
349 
350 		if (!intel_engine_has_preemption(engine))
351 			continue;
352 
353 		if (!intel_engine_can_store_dword(engine))
354 			continue;
355 
356 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
357 			err = -EIO;
358 			break;
359 		}
360 		st_engine_heartbeat_disable(engine);
361 
362 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
363 			struct intel_context *tmp;
364 
365 			tmp = intel_context_create(engine);
366 			if (IS_ERR(tmp)) {
367 				err = PTR_ERR(tmp);
368 				goto err_ce;
369 			}
370 
371 			err = intel_context_pin(tmp);
372 			if (err) {
373 				intel_context_put(tmp);
374 				goto err_ce;
375 			}
376 
377 			memset32(tmp->ring->vaddr,
378 				 0xdeadbeef, /* trigger a hang if executed */
379 				 tmp->ring->vma->size / sizeof(u32));
380 
381 			ce[n] = tmp;
382 		}
383 
384 		/* Create max prio spinner, followed by N low prio nops */
385 		rq = igt_spinner_create_request(&spin, ce[0], MI_ARB_CHECK);
386 		if (IS_ERR(rq)) {
387 			err = PTR_ERR(rq);
388 			goto err_ce;
389 		}
390 
391 		i915_request_get(rq);
392 		rq->sched.attr.priority = I915_PRIORITY_BARRIER;
393 		i915_request_add(rq);
394 
395 		if (!igt_wait_for_spinner(&spin, rq)) {
396 			intel_gt_set_wedged(gt);
397 			i915_request_put(rq);
398 			err = -ETIME;
399 			goto err_ce;
400 		}
401 
402 		/* Fill the ring, until we will cause a wrap */
403 		n = 0;
404 		while (intel_ring_direction(ce[0]->ring,
405 					    rq->wa_tail,
406 					    ce[0]->ring->tail) <= 0) {
407 			struct i915_request *tmp;
408 
409 			tmp = intel_context_create_request(ce[0]);
410 			if (IS_ERR(tmp)) {
411 				err = PTR_ERR(tmp);
412 				i915_request_put(rq);
413 				goto err_ce;
414 			}
415 
416 			i915_request_add(tmp);
417 			intel_engine_flush_submission(engine);
418 			n++;
419 		}
420 		intel_engine_flush_submission(engine);
421 		pr_debug("%s: Filled ring with %d nop tails {size:%x, tail:%x, emit:%x, rq.tail:%x}\n",
422 			 engine->name, n,
423 			 ce[0]->ring->size,
424 			 ce[0]->ring->tail,
425 			 ce[0]->ring->emit,
426 			 rq->tail);
427 		GEM_BUG_ON(intel_ring_direction(ce[0]->ring,
428 						rq->tail,
429 						ce[0]->ring->tail) <= 0);
430 		i915_request_put(rq);
431 
432 		/* Create a second ring to preempt the first ring after rq[0] */
433 		rq = intel_context_create_request(ce[1]);
434 		if (IS_ERR(rq)) {
435 			err = PTR_ERR(rq);
436 			goto err_ce;
437 		}
438 
439 		rq->sched.attr.priority = I915_PRIORITY_BARRIER;
440 		i915_request_get(rq);
441 		i915_request_add(rq);
442 
443 		err = wait_for_submit(engine, rq, HZ / 2);
444 		i915_request_put(rq);
445 		if (err) {
446 			pr_err("%s: preemption request was not submitted\n",
447 			       engine->name);
448 			err = -ETIME;
449 		}
450 
451 		pr_debug("%s: ring[0]:{ tail:%x, emit:%x }, ring[1]:{ tail:%x, emit:%x }\n",
452 			 engine->name,
453 			 ce[0]->ring->tail, ce[0]->ring->emit,
454 			 ce[1]->ring->tail, ce[1]->ring->emit);
455 
456 err_ce:
457 		intel_engine_flush_submission(engine);
458 		igt_spinner_end(&spin);
459 		for (n = 0; n < ARRAY_SIZE(ce); n++) {
460 			if (IS_ERR_OR_NULL(ce[n]))
461 				break;
462 
463 			intel_context_unpin(ce[n]);
464 			intel_context_put(ce[n]);
465 		}
466 		st_engine_heartbeat_enable(engine);
467 		if (igt_live_test_end(&t))
468 			err = -EIO;
469 		if (err)
470 			break;
471 	}
472 
473 	igt_spinner_fini(&spin);
474 	return err;
475 }
476 
477 static int live_pin_rewind(void *arg)
478 {
479 	struct intel_gt *gt = arg;
480 	struct intel_engine_cs *engine;
481 	enum intel_engine_id id;
482 	int err = 0;
483 
484 	/*
485 	 * We have to be careful not to trust intel_ring too much, for example
486 	 * ring->head is updated upon retire which is out of sync with pinning
487 	 * the context. Thus we cannot use ring->head to set CTX_RING_HEAD,
488 	 * or else we risk writing an older, stale value.
489 	 *
490 	 * To simulate this, let's apply a bit of deliberate sabotague.
491 	 */
492 
493 	for_each_engine(engine, gt, id) {
494 		struct intel_context *ce;
495 		struct i915_request *rq;
496 		struct intel_ring *ring;
497 		struct igt_live_test t;
498 
499 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
500 			err = -EIO;
501 			break;
502 		}
503 
504 		ce = intel_context_create(engine);
505 		if (IS_ERR(ce)) {
506 			err = PTR_ERR(ce);
507 			break;
508 		}
509 
510 		err = intel_context_pin(ce);
511 		if (err) {
512 			intel_context_put(ce);
513 			break;
514 		}
515 
516 		/* Keep the context awake while we play games */
517 		err = i915_active_acquire(&ce->active);
518 		if (err) {
519 			intel_context_unpin(ce);
520 			intel_context_put(ce);
521 			break;
522 		}
523 		ring = ce->ring;
524 
525 		/* Poison the ring, and offset the next request from HEAD */
526 		memset32(ring->vaddr, STACK_MAGIC, ring->size / sizeof(u32));
527 		ring->emit = ring->size / 2;
528 		ring->tail = ring->emit;
529 		GEM_BUG_ON(ring->head);
530 
531 		intel_context_unpin(ce);
532 
533 		/* Submit a simple nop request */
534 		GEM_BUG_ON(intel_context_is_pinned(ce));
535 		rq = intel_context_create_request(ce);
536 		i915_active_release(&ce->active); /* e.g. async retire */
537 		intel_context_put(ce);
538 		if (IS_ERR(rq)) {
539 			err = PTR_ERR(rq);
540 			break;
541 		}
542 		GEM_BUG_ON(!rq->head);
543 		i915_request_add(rq);
544 
545 		/* Expect not to hang! */
546 		if (igt_live_test_end(&t)) {
547 			err = -EIO;
548 			break;
549 		}
550 	}
551 
552 	return err;
553 }
554 
555 static int engine_lock_reset_tasklet(struct intel_engine_cs *engine)
556 {
557 	tasklet_disable(&engine->sched_engine->tasklet);
558 	local_bh_disable();
559 
560 	if (test_and_set_bit(I915_RESET_ENGINE + engine->id,
561 			     &engine->gt->reset.flags)) {
562 		local_bh_enable();
563 		tasklet_enable(&engine->sched_engine->tasklet);
564 
565 		intel_gt_set_wedged(engine->gt);
566 		return -EBUSY;
567 	}
568 
569 	return 0;
570 }
571 
572 static void engine_unlock_reset_tasklet(struct intel_engine_cs *engine)
573 {
574 	clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
575 			      &engine->gt->reset.flags);
576 
577 	local_bh_enable();
578 	tasklet_enable(&engine->sched_engine->tasklet);
579 }
580 
581 static int live_hold_reset(void *arg)
582 {
583 	struct intel_gt *gt = arg;
584 	struct intel_engine_cs *engine;
585 	enum intel_engine_id id;
586 	struct igt_spinner spin;
587 	int err = 0;
588 
589 	/*
590 	 * In order to support offline error capture for fast preempt reset,
591 	 * we need to decouple the guilty request and ensure that it and its
592 	 * descendents are not executed while the capture is in progress.
593 	 */
594 
595 	if (!intel_has_reset_engine(gt))
596 		return 0;
597 
598 	if (igt_spinner_init(&spin, gt))
599 		return -ENOMEM;
600 
601 	for_each_engine(engine, gt, id) {
602 		struct intel_context *ce;
603 		struct i915_request *rq;
604 
605 		ce = intel_context_create(engine);
606 		if (IS_ERR(ce)) {
607 			err = PTR_ERR(ce);
608 			break;
609 		}
610 
611 		st_engine_heartbeat_disable(engine);
612 
613 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
614 		if (IS_ERR(rq)) {
615 			err = PTR_ERR(rq);
616 			goto out;
617 		}
618 		i915_request_add(rq);
619 
620 		if (!igt_wait_for_spinner(&spin, rq)) {
621 			intel_gt_set_wedged(gt);
622 			err = -ETIME;
623 			goto out;
624 		}
625 
626 		/* We have our request executing, now remove it and reset */
627 
628 		err = engine_lock_reset_tasklet(engine);
629 		if (err)
630 			goto out;
631 
632 		engine->sched_engine->tasklet.callback(&engine->sched_engine->tasklet);
633 		GEM_BUG_ON(execlists_active(&engine->execlists) != rq);
634 
635 		i915_request_get(rq);
636 		execlists_hold(engine, rq);
637 		GEM_BUG_ON(!i915_request_on_hold(rq));
638 
639 		__intel_engine_reset_bh(engine, NULL);
640 		GEM_BUG_ON(rq->fence.error != -EIO);
641 
642 		engine_unlock_reset_tasklet(engine);
643 
644 		/* Check that we do not resubmit the held request */
645 		if (!i915_request_wait(rq, 0, HZ / 5)) {
646 			pr_err("%s: on hold request completed!\n",
647 			       engine->name);
648 			i915_request_put(rq);
649 			err = -EIO;
650 			goto out;
651 		}
652 		GEM_BUG_ON(!i915_request_on_hold(rq));
653 
654 		/* But is resubmitted on release */
655 		execlists_unhold(engine, rq);
656 		if (i915_request_wait(rq, 0, HZ / 5) < 0) {
657 			pr_err("%s: held request did not complete!\n",
658 			       engine->name);
659 			intel_gt_set_wedged(gt);
660 			err = -ETIME;
661 		}
662 		i915_request_put(rq);
663 
664 out:
665 		st_engine_heartbeat_enable(engine);
666 		intel_context_put(ce);
667 		if (err)
668 			break;
669 	}
670 
671 	igt_spinner_fini(&spin);
672 	return err;
673 }
674 
675 static const char *error_repr(int err)
676 {
677 	return err ? "bad" : "good";
678 }
679 
680 static int live_error_interrupt(void *arg)
681 {
682 	static const struct error_phase {
683 		enum { GOOD = 0, BAD = -EIO } error[2];
684 	} phases[] = {
685 		{ { BAD,  GOOD } },
686 		{ { BAD,  BAD  } },
687 		{ { BAD,  GOOD } },
688 		{ { GOOD, GOOD } }, /* sentinel */
689 	};
690 	struct intel_gt *gt = arg;
691 	struct intel_engine_cs *engine;
692 	enum intel_engine_id id;
693 
694 	/*
695 	 * We hook up the CS_MASTER_ERROR_INTERRUPT to have forewarning
696 	 * of invalid commands in user batches that will cause a GPU hang.
697 	 * This is a faster mechanism than using hangcheck/heartbeats, but
698 	 * only detects problems the HW knows about -- it will not warn when
699 	 * we kill the HW!
700 	 *
701 	 * To verify our detection and reset, we throw some invalid commands
702 	 * at the HW and wait for the interrupt.
703 	 */
704 
705 	if (!intel_has_reset_engine(gt))
706 		return 0;
707 
708 	for_each_engine(engine, gt, id) {
709 		const struct error_phase *p;
710 		int err = 0;
711 
712 		st_engine_heartbeat_disable(engine);
713 
714 		for (p = phases; p->error[0] != GOOD; p++) {
715 			struct i915_request *client[ARRAY_SIZE(phases->error)];
716 			u32 *cs;
717 			int i;
718 
719 			memset(client, 0, sizeof(*client));
720 			for (i = 0; i < ARRAY_SIZE(client); i++) {
721 				struct intel_context *ce;
722 				struct i915_request *rq;
723 
724 				ce = intel_context_create(engine);
725 				if (IS_ERR(ce)) {
726 					err = PTR_ERR(ce);
727 					goto out;
728 				}
729 
730 				rq = intel_context_create_request(ce);
731 				intel_context_put(ce);
732 				if (IS_ERR(rq)) {
733 					err = PTR_ERR(rq);
734 					goto out;
735 				}
736 
737 				if (rq->engine->emit_init_breadcrumb) {
738 					err = rq->engine->emit_init_breadcrumb(rq);
739 					if (err) {
740 						i915_request_add(rq);
741 						goto out;
742 					}
743 				}
744 
745 				cs = intel_ring_begin(rq, 2);
746 				if (IS_ERR(cs)) {
747 					i915_request_add(rq);
748 					err = PTR_ERR(cs);
749 					goto out;
750 				}
751 
752 				if (p->error[i]) {
753 					*cs++ = 0xdeadbeef;
754 					*cs++ = 0xdeadbeef;
755 				} else {
756 					*cs++ = MI_NOOP;
757 					*cs++ = MI_NOOP;
758 				}
759 
760 				client[i] = i915_request_get(rq);
761 				i915_request_add(rq);
762 			}
763 
764 			err = wait_for_submit(engine, client[0], HZ / 2);
765 			if (err) {
766 				pr_err("%s: first request did not start within time!\n",
767 				       engine->name);
768 				err = -ETIME;
769 				goto out;
770 			}
771 
772 			for (i = 0; i < ARRAY_SIZE(client); i++) {
773 				if (i915_request_wait(client[i], 0, HZ / 5) < 0)
774 					pr_debug("%s: %s request incomplete!\n",
775 						 engine->name,
776 						 error_repr(p->error[i]));
777 
778 				if (!i915_request_started(client[i])) {
779 					pr_err("%s: %s request not started!\n",
780 					       engine->name,
781 					       error_repr(p->error[i]));
782 					err = -ETIME;
783 					goto out;
784 				}
785 
786 				/* Kick the tasklet to process the error */
787 				intel_engine_flush_submission(engine);
788 				if (client[i]->fence.error != p->error[i]) {
789 					pr_err("%s: %s request (%s) with wrong error code: %d\n",
790 					       engine->name,
791 					       error_repr(p->error[i]),
792 					       i915_request_completed(client[i]) ? "completed" : "running",
793 					       client[i]->fence.error);
794 					err = -EINVAL;
795 					goto out;
796 				}
797 			}
798 
799 out:
800 			for (i = 0; i < ARRAY_SIZE(client); i++)
801 				if (client[i])
802 					i915_request_put(client[i]);
803 			if (err) {
804 				pr_err("%s: failed at phase[%zd] { %d, %d }\n",
805 				       engine->name, p - phases,
806 				       p->error[0], p->error[1]);
807 				break;
808 			}
809 		}
810 
811 		st_engine_heartbeat_enable(engine);
812 		if (err) {
813 			intel_gt_set_wedged(gt);
814 			return err;
815 		}
816 	}
817 
818 	return 0;
819 }
820 
821 static int
822 emit_semaphore_chain(struct i915_request *rq, struct i915_vma *vma, int idx)
823 {
824 	u32 *cs;
825 
826 	cs = intel_ring_begin(rq, 10);
827 	if (IS_ERR(cs))
828 		return PTR_ERR(cs);
829 
830 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
831 
832 	*cs++ = MI_SEMAPHORE_WAIT |
833 		MI_SEMAPHORE_GLOBAL_GTT |
834 		MI_SEMAPHORE_POLL |
835 		MI_SEMAPHORE_SAD_NEQ_SDD;
836 	*cs++ = 0;
837 	*cs++ = i915_ggtt_offset(vma) + 4 * idx;
838 	*cs++ = 0;
839 
840 	if (idx > 0) {
841 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
842 		*cs++ = i915_ggtt_offset(vma) + 4 * (idx - 1);
843 		*cs++ = 0;
844 		*cs++ = 1;
845 	} else {
846 		*cs++ = MI_NOOP;
847 		*cs++ = MI_NOOP;
848 		*cs++ = MI_NOOP;
849 		*cs++ = MI_NOOP;
850 	}
851 
852 	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
853 
854 	intel_ring_advance(rq, cs);
855 	return 0;
856 }
857 
858 static struct i915_request *
859 semaphore_queue(struct intel_engine_cs *engine, struct i915_vma *vma, int idx)
860 {
861 	struct intel_context *ce;
862 	struct i915_request *rq;
863 	int err;
864 
865 	ce = intel_context_create(engine);
866 	if (IS_ERR(ce))
867 		return ERR_CAST(ce);
868 
869 	rq = intel_context_create_request(ce);
870 	if (IS_ERR(rq))
871 		goto out_ce;
872 
873 	err = 0;
874 	if (rq->engine->emit_init_breadcrumb)
875 		err = rq->engine->emit_init_breadcrumb(rq);
876 	if (err == 0)
877 		err = emit_semaphore_chain(rq, vma, idx);
878 	if (err == 0)
879 		i915_request_get(rq);
880 	i915_request_add(rq);
881 	if (err)
882 		rq = ERR_PTR(err);
883 
884 out_ce:
885 	intel_context_put(ce);
886 	return rq;
887 }
888 
889 static int
890 release_queue(struct intel_engine_cs *engine,
891 	      struct i915_vma *vma,
892 	      int idx, int prio)
893 {
894 	struct i915_sched_attr attr = {
895 		.priority = prio,
896 	};
897 	struct i915_request *rq;
898 	u32 *cs;
899 
900 	rq = intel_engine_create_kernel_request(engine);
901 	if (IS_ERR(rq))
902 		return PTR_ERR(rq);
903 
904 	cs = intel_ring_begin(rq, 4);
905 	if (IS_ERR(cs)) {
906 		i915_request_add(rq);
907 		return PTR_ERR(cs);
908 	}
909 
910 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
911 	*cs++ = i915_ggtt_offset(vma) + 4 * (idx - 1);
912 	*cs++ = 0;
913 	*cs++ = 1;
914 
915 	intel_ring_advance(rq, cs);
916 
917 	i915_request_get(rq);
918 	i915_request_add(rq);
919 
920 	local_bh_disable();
921 	engine->sched_engine->schedule(rq, &attr);
922 	local_bh_enable(); /* kick tasklet */
923 
924 	i915_request_put(rq);
925 
926 	return 0;
927 }
928 
929 static int
930 slice_semaphore_queue(struct intel_engine_cs *outer,
931 		      struct i915_vma *vma,
932 		      int count)
933 {
934 	struct intel_engine_cs *engine;
935 	struct i915_request *head;
936 	enum intel_engine_id id;
937 	int err, i, n = 0;
938 
939 	head = semaphore_queue(outer, vma, n++);
940 	if (IS_ERR(head))
941 		return PTR_ERR(head);
942 
943 	for_each_engine(engine, outer->gt, id) {
944 		if (!intel_engine_has_preemption(engine))
945 			continue;
946 
947 		for (i = 0; i < count; i++) {
948 			struct i915_request *rq;
949 
950 			rq = semaphore_queue(engine, vma, n++);
951 			if (IS_ERR(rq)) {
952 				err = PTR_ERR(rq);
953 				goto out;
954 			}
955 
956 			i915_request_put(rq);
957 		}
958 	}
959 
960 	err = release_queue(outer, vma, n, I915_PRIORITY_BARRIER);
961 	if (err)
962 		goto out;
963 
964 	if (i915_request_wait(head, 0,
965 			      2 * outer->gt->info.num_engines * (count + 2) * (count + 3)) < 0) {
966 		pr_err("%s: Failed to slice along semaphore chain of length (%d, %d)!\n",
967 		       outer->name, count, n);
968 		GEM_TRACE_DUMP();
969 		intel_gt_set_wedged(outer->gt);
970 		err = -EIO;
971 	}
972 
973 out:
974 	i915_request_put(head);
975 	return err;
976 }
977 
978 static int live_timeslice_preempt(void *arg)
979 {
980 	struct intel_gt *gt = arg;
981 	struct drm_i915_gem_object *obj;
982 	struct intel_engine_cs *engine;
983 	enum intel_engine_id id;
984 	struct i915_vma *vma;
985 	void *vaddr;
986 	int err = 0;
987 
988 	/*
989 	 * If a request takes too long, we would like to give other users
990 	 * a fair go on the GPU. In particular, users may create batches
991 	 * that wait upon external input, where that input may even be
992 	 * supplied by another GPU job. To avoid blocking forever, we
993 	 * need to preempt the current task and replace it with another
994 	 * ready task.
995 	 */
996 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
997 		return 0;
998 
999 	obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
1000 	if (IS_ERR(obj))
1001 		return PTR_ERR(obj);
1002 
1003 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
1004 	if (IS_ERR(vma)) {
1005 		err = PTR_ERR(vma);
1006 		goto err_obj;
1007 	}
1008 
1009 	vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1010 	if (IS_ERR(vaddr)) {
1011 		err = PTR_ERR(vaddr);
1012 		goto err_obj;
1013 	}
1014 
1015 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1016 	if (err)
1017 		goto err_map;
1018 
1019 	err = i915_vma_sync(vma);
1020 	if (err)
1021 		goto err_pin;
1022 
1023 	for_each_engine(engine, gt, id) {
1024 		if (!intel_engine_has_preemption(engine))
1025 			continue;
1026 
1027 		memset(vaddr, 0, PAGE_SIZE);
1028 
1029 		st_engine_heartbeat_disable(engine);
1030 		err = slice_semaphore_queue(engine, vma, 5);
1031 		st_engine_heartbeat_enable(engine);
1032 		if (err)
1033 			goto err_pin;
1034 
1035 		if (igt_flush_test(gt->i915)) {
1036 			err = -EIO;
1037 			goto err_pin;
1038 		}
1039 	}
1040 
1041 err_pin:
1042 	i915_vma_unpin(vma);
1043 err_map:
1044 	i915_gem_object_unpin_map(obj);
1045 err_obj:
1046 	i915_gem_object_put(obj);
1047 	return err;
1048 }
1049 
1050 static struct i915_request *
1051 create_rewinder(struct intel_context *ce,
1052 		struct i915_request *wait,
1053 		void *slot, int idx)
1054 {
1055 	const u32 offset =
1056 		i915_ggtt_offset(ce->engine->status_page.vma) +
1057 		offset_in_page(slot);
1058 	struct i915_request *rq;
1059 	u32 *cs;
1060 	int err;
1061 
1062 	rq = intel_context_create_request(ce);
1063 	if (IS_ERR(rq))
1064 		return rq;
1065 
1066 	if (wait) {
1067 		err = i915_request_await_dma_fence(rq, &wait->fence);
1068 		if (err)
1069 			goto err;
1070 	}
1071 
1072 	cs = intel_ring_begin(rq, 14);
1073 	if (IS_ERR(cs)) {
1074 		err = PTR_ERR(cs);
1075 		goto err;
1076 	}
1077 
1078 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1079 	*cs++ = MI_NOOP;
1080 
1081 	*cs++ = MI_SEMAPHORE_WAIT |
1082 		MI_SEMAPHORE_GLOBAL_GTT |
1083 		MI_SEMAPHORE_POLL |
1084 		MI_SEMAPHORE_SAD_GTE_SDD;
1085 	*cs++ = idx;
1086 	*cs++ = offset;
1087 	*cs++ = 0;
1088 
1089 	*cs++ = MI_STORE_REGISTER_MEM_GEN8 | MI_USE_GGTT;
1090 	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(rq->engine->mmio_base));
1091 	*cs++ = offset + idx * sizeof(u32);
1092 	*cs++ = 0;
1093 
1094 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1095 	*cs++ = offset;
1096 	*cs++ = 0;
1097 	*cs++ = idx + 1;
1098 
1099 	intel_ring_advance(rq, cs);
1100 
1101 	err = 0;
1102 err:
1103 	i915_request_get(rq);
1104 	i915_request_add(rq);
1105 	if (err) {
1106 		i915_request_put(rq);
1107 		return ERR_PTR(err);
1108 	}
1109 
1110 	return rq;
1111 }
1112 
1113 static int live_timeslice_rewind(void *arg)
1114 {
1115 	struct intel_gt *gt = arg;
1116 	struct intel_engine_cs *engine;
1117 	enum intel_engine_id id;
1118 
1119 	/*
1120 	 * The usual presumption on timeslice expiration is that we replace
1121 	 * the active context with another. However, given a chain of
1122 	 * dependencies we may end up with replacing the context with itself,
1123 	 * but only a few of those requests, forcing us to rewind the
1124 	 * RING_TAIL of the original request.
1125 	 */
1126 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
1127 		return 0;
1128 
1129 	for_each_engine(engine, gt, id) {
1130 		enum { A1, A2, B1 };
1131 		enum { X = 1, Z, Y };
1132 		struct i915_request *rq[3] = {};
1133 		struct intel_context *ce;
1134 		unsigned long timeslice;
1135 		int i, err = 0;
1136 		u32 *slot;
1137 
1138 		if (!intel_engine_has_timeslices(engine))
1139 			continue;
1140 
1141 		/*
1142 		 * A:rq1 -- semaphore wait, timestamp X
1143 		 * A:rq2 -- write timestamp Y
1144 		 *
1145 		 * B:rq1 [await A:rq1] -- write timestamp Z
1146 		 *
1147 		 * Force timeslice, release semaphore.
1148 		 *
1149 		 * Expect execution/evaluation order XZY
1150 		 */
1151 
1152 		st_engine_heartbeat_disable(engine);
1153 		timeslice = xchg(&engine->props.timeslice_duration_ms, 1);
1154 
1155 		slot = memset32(engine->status_page.addr + 1000, 0, 4);
1156 
1157 		ce = intel_context_create(engine);
1158 		if (IS_ERR(ce)) {
1159 			err = PTR_ERR(ce);
1160 			goto err;
1161 		}
1162 
1163 		rq[A1] = create_rewinder(ce, NULL, slot, X);
1164 		if (IS_ERR(rq[A1])) {
1165 			intel_context_put(ce);
1166 			goto err;
1167 		}
1168 
1169 		rq[A2] = create_rewinder(ce, NULL, slot, Y);
1170 		intel_context_put(ce);
1171 		if (IS_ERR(rq[A2]))
1172 			goto err;
1173 
1174 		err = wait_for_submit(engine, rq[A2], HZ / 2);
1175 		if (err) {
1176 			pr_err("%s: failed to submit first context\n",
1177 			       engine->name);
1178 			goto err;
1179 		}
1180 
1181 		ce = intel_context_create(engine);
1182 		if (IS_ERR(ce)) {
1183 			err = PTR_ERR(ce);
1184 			goto err;
1185 		}
1186 
1187 		rq[B1] = create_rewinder(ce, rq[A1], slot, Z);
1188 		intel_context_put(ce);
1189 		if (IS_ERR(rq[2]))
1190 			goto err;
1191 
1192 		err = wait_for_submit(engine, rq[B1], HZ / 2);
1193 		if (err) {
1194 			pr_err("%s: failed to submit second context\n",
1195 			       engine->name);
1196 			goto err;
1197 		}
1198 
1199 		/* ELSP[] = { { A:rq1, A:rq2 }, { B:rq1 } } */
1200 		ENGINE_TRACE(engine, "forcing tasklet for rewind\n");
1201 		while (i915_request_is_active(rq[A2])) { /* semaphore yield! */
1202 			/* Wait for the timeslice to kick in */
1203 			del_timer(&engine->execlists.timer);
1204 			tasklet_hi_schedule(&engine->sched_engine->tasklet);
1205 			intel_engine_flush_submission(engine);
1206 		}
1207 		/* -> ELSP[] = { { A:rq1 }, { B:rq1 } } */
1208 		GEM_BUG_ON(!i915_request_is_active(rq[A1]));
1209 		GEM_BUG_ON(!i915_request_is_active(rq[B1]));
1210 		GEM_BUG_ON(i915_request_is_active(rq[A2]));
1211 
1212 		/* Release the hounds! */
1213 		slot[0] = 1;
1214 		wmb(); /* "pairs" with GPU; paranoid kick of internal CPU$ */
1215 
1216 		for (i = 1; i <= 3; i++) {
1217 			unsigned long timeout = jiffies + HZ / 2;
1218 
1219 			while (!READ_ONCE(slot[i]) &&
1220 			       time_before(jiffies, timeout))
1221 				;
1222 
1223 			if (!time_before(jiffies, timeout)) {
1224 				pr_err("%s: rq[%d] timed out\n",
1225 				       engine->name, i - 1);
1226 				err = -ETIME;
1227 				goto err;
1228 			}
1229 
1230 			pr_debug("%s: slot[%d]:%x\n", engine->name, i, slot[i]);
1231 		}
1232 
1233 		/* XZY: XZ < XY */
1234 		if (slot[Z] - slot[X] >= slot[Y] - slot[X]) {
1235 			pr_err("%s: timeslicing did not run context B [%u] before A [%u]!\n",
1236 			       engine->name,
1237 			       slot[Z] - slot[X],
1238 			       slot[Y] - slot[X]);
1239 			err = -EINVAL;
1240 		}
1241 
1242 err:
1243 		memset32(&slot[0], -1, 4);
1244 		wmb();
1245 
1246 		engine->props.timeslice_duration_ms = timeslice;
1247 		st_engine_heartbeat_enable(engine);
1248 		for (i = 0; i < 3; i++)
1249 			i915_request_put(rq[i]);
1250 		if (igt_flush_test(gt->i915))
1251 			err = -EIO;
1252 		if (err)
1253 			return err;
1254 	}
1255 
1256 	return 0;
1257 }
1258 
1259 static struct i915_request *nop_request(struct intel_engine_cs *engine)
1260 {
1261 	struct i915_request *rq;
1262 
1263 	rq = intel_engine_create_kernel_request(engine);
1264 	if (IS_ERR(rq))
1265 		return rq;
1266 
1267 	i915_request_get(rq);
1268 	i915_request_add(rq);
1269 
1270 	return rq;
1271 }
1272 
1273 static long slice_timeout(struct intel_engine_cs *engine)
1274 {
1275 	long timeout;
1276 
1277 	/* Enough time for a timeslice to kick in, and kick out */
1278 	timeout = 2 * msecs_to_jiffies_timeout(timeslice(engine));
1279 
1280 	/* Enough time for the nop request to complete */
1281 	timeout += HZ / 5;
1282 
1283 	return timeout + 1;
1284 }
1285 
1286 static int live_timeslice_queue(void *arg)
1287 {
1288 	struct intel_gt *gt = arg;
1289 	struct drm_i915_gem_object *obj;
1290 	struct intel_engine_cs *engine;
1291 	enum intel_engine_id id;
1292 	struct i915_vma *vma;
1293 	void *vaddr;
1294 	int err = 0;
1295 
1296 	/*
1297 	 * Make sure that even if ELSP[0] and ELSP[1] are filled with
1298 	 * timeslicing between them disabled, we *do* enable timeslicing
1299 	 * if the queue demands it. (Normally, we do not submit if
1300 	 * ELSP[1] is already occupied, so must rely on timeslicing to
1301 	 * eject ELSP[0] in favour of the queue.)
1302 	 */
1303 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
1304 		return 0;
1305 
1306 	obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
1307 	if (IS_ERR(obj))
1308 		return PTR_ERR(obj);
1309 
1310 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
1311 	if (IS_ERR(vma)) {
1312 		err = PTR_ERR(vma);
1313 		goto err_obj;
1314 	}
1315 
1316 	vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1317 	if (IS_ERR(vaddr)) {
1318 		err = PTR_ERR(vaddr);
1319 		goto err_obj;
1320 	}
1321 
1322 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1323 	if (err)
1324 		goto err_map;
1325 
1326 	err = i915_vma_sync(vma);
1327 	if (err)
1328 		goto err_pin;
1329 
1330 	for_each_engine(engine, gt, id) {
1331 		struct i915_sched_attr attr = { .priority = I915_PRIORITY_MAX };
1332 		struct i915_request *rq, *nop;
1333 
1334 		if (!intel_engine_has_preemption(engine))
1335 			continue;
1336 
1337 		st_engine_heartbeat_disable(engine);
1338 		memset(vaddr, 0, PAGE_SIZE);
1339 
1340 		/* ELSP[0]: semaphore wait */
1341 		rq = semaphore_queue(engine, vma, 0);
1342 		if (IS_ERR(rq)) {
1343 			err = PTR_ERR(rq);
1344 			goto err_heartbeat;
1345 		}
1346 		engine->sched_engine->schedule(rq, &attr);
1347 		err = wait_for_submit(engine, rq, HZ / 2);
1348 		if (err) {
1349 			pr_err("%s: Timed out trying to submit semaphores\n",
1350 			       engine->name);
1351 			goto err_rq;
1352 		}
1353 
1354 		/* ELSP[1]: nop request */
1355 		nop = nop_request(engine);
1356 		if (IS_ERR(nop)) {
1357 			err = PTR_ERR(nop);
1358 			goto err_rq;
1359 		}
1360 		err = wait_for_submit(engine, nop, HZ / 2);
1361 		i915_request_put(nop);
1362 		if (err) {
1363 			pr_err("%s: Timed out trying to submit nop\n",
1364 			       engine->name);
1365 			goto err_rq;
1366 		}
1367 
1368 		GEM_BUG_ON(i915_request_completed(rq));
1369 		GEM_BUG_ON(execlists_active(&engine->execlists) != rq);
1370 
1371 		/* Queue: semaphore signal, matching priority as semaphore */
1372 		err = release_queue(engine, vma, 1, effective_prio(rq));
1373 		if (err)
1374 			goto err_rq;
1375 
1376 		/* Wait until we ack the release_queue and start timeslicing */
1377 		do {
1378 			cond_resched();
1379 			intel_engine_flush_submission(engine);
1380 		} while (READ_ONCE(engine->execlists.pending[0]));
1381 
1382 		/* Timeslice every jiffy, so within 2 we should signal */
1383 		if (i915_request_wait(rq, 0, slice_timeout(engine)) < 0) {
1384 			struct drm_printer p =
1385 				drm_info_printer(gt->i915->drm.dev);
1386 
1387 			pr_err("%s: Failed to timeslice into queue\n",
1388 			       engine->name);
1389 			intel_engine_dump(engine, &p,
1390 					  "%s\n", engine->name);
1391 
1392 			memset(vaddr, 0xff, PAGE_SIZE);
1393 			err = -EIO;
1394 		}
1395 err_rq:
1396 		i915_request_put(rq);
1397 err_heartbeat:
1398 		st_engine_heartbeat_enable(engine);
1399 		if (err)
1400 			break;
1401 	}
1402 
1403 err_pin:
1404 	i915_vma_unpin(vma);
1405 err_map:
1406 	i915_gem_object_unpin_map(obj);
1407 err_obj:
1408 	i915_gem_object_put(obj);
1409 	return err;
1410 }
1411 
1412 static int live_timeslice_nopreempt(void *arg)
1413 {
1414 	struct intel_gt *gt = arg;
1415 	struct intel_engine_cs *engine;
1416 	enum intel_engine_id id;
1417 	struct igt_spinner spin;
1418 	int err = 0;
1419 
1420 	/*
1421 	 * We should not timeslice into a request that is marked with
1422 	 * I915_REQUEST_NOPREEMPT.
1423 	 */
1424 	if (!CONFIG_DRM_I915_TIMESLICE_DURATION)
1425 		return 0;
1426 
1427 	if (igt_spinner_init(&spin, gt))
1428 		return -ENOMEM;
1429 
1430 	for_each_engine(engine, gt, id) {
1431 		struct intel_context *ce;
1432 		struct i915_request *rq;
1433 		unsigned long timeslice;
1434 
1435 		if (!intel_engine_has_preemption(engine))
1436 			continue;
1437 
1438 		ce = intel_context_create(engine);
1439 		if (IS_ERR(ce)) {
1440 			err = PTR_ERR(ce);
1441 			break;
1442 		}
1443 
1444 		st_engine_heartbeat_disable(engine);
1445 		timeslice = xchg(&engine->props.timeslice_duration_ms, 1);
1446 
1447 		/* Create an unpreemptible spinner */
1448 
1449 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
1450 		intel_context_put(ce);
1451 		if (IS_ERR(rq)) {
1452 			err = PTR_ERR(rq);
1453 			goto out_heartbeat;
1454 		}
1455 
1456 		i915_request_get(rq);
1457 		i915_request_add(rq);
1458 
1459 		if (!igt_wait_for_spinner(&spin, rq)) {
1460 			i915_request_put(rq);
1461 			err = -ETIME;
1462 			goto out_spin;
1463 		}
1464 
1465 		set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags);
1466 		i915_request_put(rq);
1467 
1468 		/* Followed by a maximum priority barrier (heartbeat) */
1469 
1470 		ce = intel_context_create(engine);
1471 		if (IS_ERR(ce)) {
1472 			err = PTR_ERR(ce);
1473 			goto out_spin;
1474 		}
1475 
1476 		rq = intel_context_create_request(ce);
1477 		intel_context_put(ce);
1478 		if (IS_ERR(rq)) {
1479 			err = PTR_ERR(rq);
1480 			goto out_spin;
1481 		}
1482 
1483 		rq->sched.attr.priority = I915_PRIORITY_BARRIER;
1484 		i915_request_get(rq);
1485 		i915_request_add(rq);
1486 
1487 		/*
1488 		 * Wait until the barrier is in ELSP, and we know timeslicing
1489 		 * will have been activated.
1490 		 */
1491 		if (wait_for_submit(engine, rq, HZ / 2)) {
1492 			i915_request_put(rq);
1493 			err = -ETIME;
1494 			goto out_spin;
1495 		}
1496 
1497 		/*
1498 		 * Since the ELSP[0] request is unpreemptible, it should not
1499 		 * allow the maximum priority barrier through. Wait long
1500 		 * enough to see if it is timesliced in by mistake.
1501 		 */
1502 		if (i915_request_wait(rq, 0, slice_timeout(engine)) >= 0) {
1503 			pr_err("%s: I915_PRIORITY_BARRIER request completed, bypassing no-preempt request\n",
1504 			       engine->name);
1505 			err = -EINVAL;
1506 		}
1507 		i915_request_put(rq);
1508 
1509 out_spin:
1510 		igt_spinner_end(&spin);
1511 out_heartbeat:
1512 		xchg(&engine->props.timeslice_duration_ms, timeslice);
1513 		st_engine_heartbeat_enable(engine);
1514 		if (err)
1515 			break;
1516 
1517 		if (igt_flush_test(gt->i915)) {
1518 			err = -EIO;
1519 			break;
1520 		}
1521 	}
1522 
1523 	igt_spinner_fini(&spin);
1524 	return err;
1525 }
1526 
1527 static int live_busywait_preempt(void *arg)
1528 {
1529 	struct intel_gt *gt = arg;
1530 	struct i915_gem_context *ctx_hi, *ctx_lo;
1531 	struct intel_engine_cs *engine;
1532 	struct drm_i915_gem_object *obj;
1533 	struct i915_vma *vma;
1534 	enum intel_engine_id id;
1535 	int err = -ENOMEM;
1536 	u32 *map;
1537 
1538 	/*
1539 	 * Verify that even without HAS_LOGICAL_RING_PREEMPTION, we can
1540 	 * preempt the busywaits used to synchronise between rings.
1541 	 */
1542 
1543 	ctx_hi = kernel_context(gt->i915, NULL);
1544 	if (!ctx_hi)
1545 		return -ENOMEM;
1546 	ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
1547 
1548 	ctx_lo = kernel_context(gt->i915, NULL);
1549 	if (!ctx_lo)
1550 		goto err_ctx_hi;
1551 	ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
1552 
1553 	obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
1554 	if (IS_ERR(obj)) {
1555 		err = PTR_ERR(obj);
1556 		goto err_ctx_lo;
1557 	}
1558 
1559 	map = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1560 	if (IS_ERR(map)) {
1561 		err = PTR_ERR(map);
1562 		goto err_obj;
1563 	}
1564 
1565 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
1566 	if (IS_ERR(vma)) {
1567 		err = PTR_ERR(vma);
1568 		goto err_map;
1569 	}
1570 
1571 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
1572 	if (err)
1573 		goto err_map;
1574 
1575 	err = i915_vma_sync(vma);
1576 	if (err)
1577 		goto err_vma;
1578 
1579 	for_each_engine(engine, gt, id) {
1580 		struct i915_request *lo, *hi;
1581 		struct igt_live_test t;
1582 		u32 *cs;
1583 
1584 		if (!intel_engine_has_preemption(engine))
1585 			continue;
1586 
1587 		if (!intel_engine_can_store_dword(engine))
1588 			continue;
1589 
1590 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
1591 			err = -EIO;
1592 			goto err_vma;
1593 		}
1594 
1595 		/*
1596 		 * We create two requests. The low priority request
1597 		 * busywaits on a semaphore (inside the ringbuffer where
1598 		 * is should be preemptible) and the high priority requests
1599 		 * uses a MI_STORE_DWORD_IMM to update the semaphore value
1600 		 * allowing the first request to complete. If preemption
1601 		 * fails, we hang instead.
1602 		 */
1603 
1604 		lo = igt_request_alloc(ctx_lo, engine);
1605 		if (IS_ERR(lo)) {
1606 			err = PTR_ERR(lo);
1607 			goto err_vma;
1608 		}
1609 
1610 		cs = intel_ring_begin(lo, 8);
1611 		if (IS_ERR(cs)) {
1612 			err = PTR_ERR(cs);
1613 			i915_request_add(lo);
1614 			goto err_vma;
1615 		}
1616 
1617 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1618 		*cs++ = i915_ggtt_offset(vma);
1619 		*cs++ = 0;
1620 		*cs++ = 1;
1621 
1622 		/* XXX Do we need a flush + invalidate here? */
1623 
1624 		*cs++ = MI_SEMAPHORE_WAIT |
1625 			MI_SEMAPHORE_GLOBAL_GTT |
1626 			MI_SEMAPHORE_POLL |
1627 			MI_SEMAPHORE_SAD_EQ_SDD;
1628 		*cs++ = 0;
1629 		*cs++ = i915_ggtt_offset(vma);
1630 		*cs++ = 0;
1631 
1632 		intel_ring_advance(lo, cs);
1633 
1634 		i915_request_get(lo);
1635 		i915_request_add(lo);
1636 
1637 		if (wait_for(READ_ONCE(*map), 10)) {
1638 			i915_request_put(lo);
1639 			err = -ETIMEDOUT;
1640 			goto err_vma;
1641 		}
1642 
1643 		/* Low priority request should be busywaiting now */
1644 		if (i915_request_wait(lo, 0, 1) != -ETIME) {
1645 			i915_request_put(lo);
1646 			pr_err("%s: Busywaiting request did not!\n",
1647 			       engine->name);
1648 			err = -EIO;
1649 			goto err_vma;
1650 		}
1651 
1652 		hi = igt_request_alloc(ctx_hi, engine);
1653 		if (IS_ERR(hi)) {
1654 			err = PTR_ERR(hi);
1655 			i915_request_put(lo);
1656 			goto err_vma;
1657 		}
1658 
1659 		cs = intel_ring_begin(hi, 4);
1660 		if (IS_ERR(cs)) {
1661 			err = PTR_ERR(cs);
1662 			i915_request_add(hi);
1663 			i915_request_put(lo);
1664 			goto err_vma;
1665 		}
1666 
1667 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1668 		*cs++ = i915_ggtt_offset(vma);
1669 		*cs++ = 0;
1670 		*cs++ = 0;
1671 
1672 		intel_ring_advance(hi, cs);
1673 		i915_request_add(hi);
1674 
1675 		if (i915_request_wait(lo, 0, HZ / 5) < 0) {
1676 			struct drm_printer p = drm_info_printer(gt->i915->drm.dev);
1677 
1678 			pr_err("%s: Failed to preempt semaphore busywait!\n",
1679 			       engine->name);
1680 
1681 			intel_engine_dump(engine, &p, "%s\n", engine->name);
1682 			GEM_TRACE_DUMP();
1683 
1684 			i915_request_put(lo);
1685 			intel_gt_set_wedged(gt);
1686 			err = -EIO;
1687 			goto err_vma;
1688 		}
1689 		GEM_BUG_ON(READ_ONCE(*map));
1690 		i915_request_put(lo);
1691 
1692 		if (igt_live_test_end(&t)) {
1693 			err = -EIO;
1694 			goto err_vma;
1695 		}
1696 	}
1697 
1698 	err = 0;
1699 err_vma:
1700 	i915_vma_unpin(vma);
1701 err_map:
1702 	i915_gem_object_unpin_map(obj);
1703 err_obj:
1704 	i915_gem_object_put(obj);
1705 err_ctx_lo:
1706 	kernel_context_close(ctx_lo);
1707 err_ctx_hi:
1708 	kernel_context_close(ctx_hi);
1709 	return err;
1710 }
1711 
1712 static struct i915_request *
1713 spinner_create_request(struct igt_spinner *spin,
1714 		       struct i915_gem_context *ctx,
1715 		       struct intel_engine_cs *engine,
1716 		       u32 arb)
1717 {
1718 	struct intel_context *ce;
1719 	struct i915_request *rq;
1720 
1721 	ce = i915_gem_context_get_engine(ctx, engine->legacy_idx);
1722 	if (IS_ERR(ce))
1723 		return ERR_CAST(ce);
1724 
1725 	rq = igt_spinner_create_request(spin, ce, arb);
1726 	intel_context_put(ce);
1727 	return rq;
1728 }
1729 
1730 static int live_preempt(void *arg)
1731 {
1732 	struct intel_gt *gt = arg;
1733 	struct i915_gem_context *ctx_hi, *ctx_lo;
1734 	struct igt_spinner spin_hi, spin_lo;
1735 	struct intel_engine_cs *engine;
1736 	enum intel_engine_id id;
1737 	int err = -ENOMEM;
1738 
1739 	if (igt_spinner_init(&spin_hi, gt))
1740 		return -ENOMEM;
1741 
1742 	if (igt_spinner_init(&spin_lo, gt))
1743 		goto err_spin_hi;
1744 
1745 	ctx_hi = kernel_context(gt->i915, NULL);
1746 	if (!ctx_hi)
1747 		goto err_spin_lo;
1748 	ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
1749 
1750 	ctx_lo = kernel_context(gt->i915, NULL);
1751 	if (!ctx_lo)
1752 		goto err_ctx_hi;
1753 	ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
1754 
1755 	for_each_engine(engine, gt, id) {
1756 		struct igt_live_test t;
1757 		struct i915_request *rq;
1758 
1759 		if (!intel_engine_has_preemption(engine))
1760 			continue;
1761 
1762 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
1763 			err = -EIO;
1764 			goto err_ctx_lo;
1765 		}
1766 
1767 		rq = spinner_create_request(&spin_lo, ctx_lo, engine,
1768 					    MI_ARB_CHECK);
1769 		if (IS_ERR(rq)) {
1770 			err = PTR_ERR(rq);
1771 			goto err_ctx_lo;
1772 		}
1773 
1774 		i915_request_add(rq);
1775 		if (!igt_wait_for_spinner(&spin_lo, rq)) {
1776 			GEM_TRACE("lo spinner failed to start\n");
1777 			GEM_TRACE_DUMP();
1778 			intel_gt_set_wedged(gt);
1779 			err = -EIO;
1780 			goto err_ctx_lo;
1781 		}
1782 
1783 		rq = spinner_create_request(&spin_hi, ctx_hi, engine,
1784 					    MI_ARB_CHECK);
1785 		if (IS_ERR(rq)) {
1786 			igt_spinner_end(&spin_lo);
1787 			err = PTR_ERR(rq);
1788 			goto err_ctx_lo;
1789 		}
1790 
1791 		i915_request_add(rq);
1792 		if (!igt_wait_for_spinner(&spin_hi, rq)) {
1793 			GEM_TRACE("hi spinner failed to start\n");
1794 			GEM_TRACE_DUMP();
1795 			intel_gt_set_wedged(gt);
1796 			err = -EIO;
1797 			goto err_ctx_lo;
1798 		}
1799 
1800 		igt_spinner_end(&spin_hi);
1801 		igt_spinner_end(&spin_lo);
1802 
1803 		if (igt_live_test_end(&t)) {
1804 			err = -EIO;
1805 			goto err_ctx_lo;
1806 		}
1807 	}
1808 
1809 	err = 0;
1810 err_ctx_lo:
1811 	kernel_context_close(ctx_lo);
1812 err_ctx_hi:
1813 	kernel_context_close(ctx_hi);
1814 err_spin_lo:
1815 	igt_spinner_fini(&spin_lo);
1816 err_spin_hi:
1817 	igt_spinner_fini(&spin_hi);
1818 	return err;
1819 }
1820 
1821 static int live_late_preempt(void *arg)
1822 {
1823 	struct intel_gt *gt = arg;
1824 	struct i915_gem_context *ctx_hi, *ctx_lo;
1825 	struct igt_spinner spin_hi, spin_lo;
1826 	struct intel_engine_cs *engine;
1827 	struct i915_sched_attr attr = {};
1828 	enum intel_engine_id id;
1829 	int err = -ENOMEM;
1830 
1831 	if (igt_spinner_init(&spin_hi, gt))
1832 		return -ENOMEM;
1833 
1834 	if (igt_spinner_init(&spin_lo, gt))
1835 		goto err_spin_hi;
1836 
1837 	ctx_hi = kernel_context(gt->i915, NULL);
1838 	if (!ctx_hi)
1839 		goto err_spin_lo;
1840 
1841 	ctx_lo = kernel_context(gt->i915, NULL);
1842 	if (!ctx_lo)
1843 		goto err_ctx_hi;
1844 
1845 	/* Make sure ctx_lo stays before ctx_hi until we trigger preemption. */
1846 	ctx_lo->sched.priority = 1;
1847 
1848 	for_each_engine(engine, gt, id) {
1849 		struct igt_live_test t;
1850 		struct i915_request *rq;
1851 
1852 		if (!intel_engine_has_preemption(engine))
1853 			continue;
1854 
1855 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
1856 			err = -EIO;
1857 			goto err_ctx_lo;
1858 		}
1859 
1860 		rq = spinner_create_request(&spin_lo, ctx_lo, engine,
1861 					    MI_ARB_CHECK);
1862 		if (IS_ERR(rq)) {
1863 			err = PTR_ERR(rq);
1864 			goto err_ctx_lo;
1865 		}
1866 
1867 		i915_request_add(rq);
1868 		if (!igt_wait_for_spinner(&spin_lo, rq)) {
1869 			pr_err("First context failed to start\n");
1870 			goto err_wedged;
1871 		}
1872 
1873 		rq = spinner_create_request(&spin_hi, ctx_hi, engine,
1874 					    MI_NOOP);
1875 		if (IS_ERR(rq)) {
1876 			igt_spinner_end(&spin_lo);
1877 			err = PTR_ERR(rq);
1878 			goto err_ctx_lo;
1879 		}
1880 
1881 		i915_request_add(rq);
1882 		if (igt_wait_for_spinner(&spin_hi, rq)) {
1883 			pr_err("Second context overtook first?\n");
1884 			goto err_wedged;
1885 		}
1886 
1887 		attr.priority = I915_PRIORITY_MAX;
1888 		engine->sched_engine->schedule(rq, &attr);
1889 
1890 		if (!igt_wait_for_spinner(&spin_hi, rq)) {
1891 			pr_err("High priority context failed to preempt the low priority context\n");
1892 			GEM_TRACE_DUMP();
1893 			goto err_wedged;
1894 		}
1895 
1896 		igt_spinner_end(&spin_hi);
1897 		igt_spinner_end(&spin_lo);
1898 
1899 		if (igt_live_test_end(&t)) {
1900 			err = -EIO;
1901 			goto err_ctx_lo;
1902 		}
1903 	}
1904 
1905 	err = 0;
1906 err_ctx_lo:
1907 	kernel_context_close(ctx_lo);
1908 err_ctx_hi:
1909 	kernel_context_close(ctx_hi);
1910 err_spin_lo:
1911 	igt_spinner_fini(&spin_lo);
1912 err_spin_hi:
1913 	igt_spinner_fini(&spin_hi);
1914 	return err;
1915 
1916 err_wedged:
1917 	igt_spinner_end(&spin_hi);
1918 	igt_spinner_end(&spin_lo);
1919 	intel_gt_set_wedged(gt);
1920 	err = -EIO;
1921 	goto err_ctx_lo;
1922 }
1923 
1924 struct preempt_client {
1925 	struct igt_spinner spin;
1926 	struct i915_gem_context *ctx;
1927 };
1928 
1929 static int preempt_client_init(struct intel_gt *gt, struct preempt_client *c)
1930 {
1931 	c->ctx = kernel_context(gt->i915, NULL);
1932 	if (!c->ctx)
1933 		return -ENOMEM;
1934 
1935 	if (igt_spinner_init(&c->spin, gt))
1936 		goto err_ctx;
1937 
1938 	return 0;
1939 
1940 err_ctx:
1941 	kernel_context_close(c->ctx);
1942 	return -ENOMEM;
1943 }
1944 
1945 static void preempt_client_fini(struct preempt_client *c)
1946 {
1947 	igt_spinner_fini(&c->spin);
1948 	kernel_context_close(c->ctx);
1949 }
1950 
1951 static int live_nopreempt(void *arg)
1952 {
1953 	struct intel_gt *gt = arg;
1954 	struct intel_engine_cs *engine;
1955 	struct preempt_client a, b;
1956 	enum intel_engine_id id;
1957 	int err = -ENOMEM;
1958 
1959 	/*
1960 	 * Verify that we can disable preemption for an individual request
1961 	 * that may be being observed and not want to be interrupted.
1962 	 */
1963 
1964 	if (preempt_client_init(gt, &a))
1965 		return -ENOMEM;
1966 	if (preempt_client_init(gt, &b))
1967 		goto err_client_a;
1968 	b.ctx->sched.priority = I915_PRIORITY_MAX;
1969 
1970 	for_each_engine(engine, gt, id) {
1971 		struct i915_request *rq_a, *rq_b;
1972 
1973 		if (!intel_engine_has_preemption(engine))
1974 			continue;
1975 
1976 		engine->execlists.preempt_hang.count = 0;
1977 
1978 		rq_a = spinner_create_request(&a.spin,
1979 					      a.ctx, engine,
1980 					      MI_ARB_CHECK);
1981 		if (IS_ERR(rq_a)) {
1982 			err = PTR_ERR(rq_a);
1983 			goto err_client_b;
1984 		}
1985 
1986 		/* Low priority client, but unpreemptable! */
1987 		__set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq_a->fence.flags);
1988 
1989 		i915_request_add(rq_a);
1990 		if (!igt_wait_for_spinner(&a.spin, rq_a)) {
1991 			pr_err("First client failed to start\n");
1992 			goto err_wedged;
1993 		}
1994 
1995 		rq_b = spinner_create_request(&b.spin,
1996 					      b.ctx, engine,
1997 					      MI_ARB_CHECK);
1998 		if (IS_ERR(rq_b)) {
1999 			err = PTR_ERR(rq_b);
2000 			goto err_client_b;
2001 		}
2002 
2003 		i915_request_add(rq_b);
2004 
2005 		/* B is much more important than A! (But A is unpreemptable.) */
2006 		GEM_BUG_ON(rq_prio(rq_b) <= rq_prio(rq_a));
2007 
2008 		/* Wait long enough for preemption and timeslicing */
2009 		if (igt_wait_for_spinner(&b.spin, rq_b)) {
2010 			pr_err("Second client started too early!\n");
2011 			goto err_wedged;
2012 		}
2013 
2014 		igt_spinner_end(&a.spin);
2015 
2016 		if (!igt_wait_for_spinner(&b.spin, rq_b)) {
2017 			pr_err("Second client failed to start\n");
2018 			goto err_wedged;
2019 		}
2020 
2021 		igt_spinner_end(&b.spin);
2022 
2023 		if (engine->execlists.preempt_hang.count) {
2024 			pr_err("Preemption recorded x%d; should have been suppressed!\n",
2025 			       engine->execlists.preempt_hang.count);
2026 			err = -EINVAL;
2027 			goto err_wedged;
2028 		}
2029 
2030 		if (igt_flush_test(gt->i915))
2031 			goto err_wedged;
2032 	}
2033 
2034 	err = 0;
2035 err_client_b:
2036 	preempt_client_fini(&b);
2037 err_client_a:
2038 	preempt_client_fini(&a);
2039 	return err;
2040 
2041 err_wedged:
2042 	igt_spinner_end(&b.spin);
2043 	igt_spinner_end(&a.spin);
2044 	intel_gt_set_wedged(gt);
2045 	err = -EIO;
2046 	goto err_client_b;
2047 }
2048 
2049 struct live_preempt_cancel {
2050 	struct intel_engine_cs *engine;
2051 	struct preempt_client a, b;
2052 };
2053 
2054 static int __cancel_active0(struct live_preempt_cancel *arg)
2055 {
2056 	struct i915_request *rq;
2057 	struct igt_live_test t;
2058 	int err;
2059 
2060 	/* Preempt cancel of ELSP0 */
2061 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2062 	if (igt_live_test_begin(&t, arg->engine->i915,
2063 				__func__, arg->engine->name))
2064 		return -EIO;
2065 
2066 	rq = spinner_create_request(&arg->a.spin,
2067 				    arg->a.ctx, arg->engine,
2068 				    MI_ARB_CHECK);
2069 	if (IS_ERR(rq))
2070 		return PTR_ERR(rq);
2071 
2072 	clear_bit(CONTEXT_BANNED, &rq->context->flags);
2073 	i915_request_get(rq);
2074 	i915_request_add(rq);
2075 	if (!igt_wait_for_spinner(&arg->a.spin, rq)) {
2076 		err = -EIO;
2077 		goto out;
2078 	}
2079 
2080 	intel_context_set_banned(rq->context);
2081 	err = intel_engine_pulse(arg->engine);
2082 	if (err)
2083 		goto out;
2084 
2085 	err = wait_for_reset(arg->engine, rq, HZ / 2);
2086 	if (err) {
2087 		pr_err("Cancelled inflight0 request did not reset\n");
2088 		goto out;
2089 	}
2090 
2091 out:
2092 	i915_request_put(rq);
2093 	if (igt_live_test_end(&t))
2094 		err = -EIO;
2095 	return err;
2096 }
2097 
2098 static int __cancel_active1(struct live_preempt_cancel *arg)
2099 {
2100 	struct i915_request *rq[2] = {};
2101 	struct igt_live_test t;
2102 	int err;
2103 
2104 	/* Preempt cancel of ELSP1 */
2105 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2106 	if (igt_live_test_begin(&t, arg->engine->i915,
2107 				__func__, arg->engine->name))
2108 		return -EIO;
2109 
2110 	rq[0] = spinner_create_request(&arg->a.spin,
2111 				       arg->a.ctx, arg->engine,
2112 				       MI_NOOP); /* no preemption */
2113 	if (IS_ERR(rq[0]))
2114 		return PTR_ERR(rq[0]);
2115 
2116 	clear_bit(CONTEXT_BANNED, &rq[0]->context->flags);
2117 	i915_request_get(rq[0]);
2118 	i915_request_add(rq[0]);
2119 	if (!igt_wait_for_spinner(&arg->a.spin, rq[0])) {
2120 		err = -EIO;
2121 		goto out;
2122 	}
2123 
2124 	rq[1] = spinner_create_request(&arg->b.spin,
2125 				       arg->b.ctx, arg->engine,
2126 				       MI_ARB_CHECK);
2127 	if (IS_ERR(rq[1])) {
2128 		err = PTR_ERR(rq[1]);
2129 		goto out;
2130 	}
2131 
2132 	clear_bit(CONTEXT_BANNED, &rq[1]->context->flags);
2133 	i915_request_get(rq[1]);
2134 	err = i915_request_await_dma_fence(rq[1], &rq[0]->fence);
2135 	i915_request_add(rq[1]);
2136 	if (err)
2137 		goto out;
2138 
2139 	intel_context_set_banned(rq[1]->context);
2140 	err = intel_engine_pulse(arg->engine);
2141 	if (err)
2142 		goto out;
2143 
2144 	igt_spinner_end(&arg->a.spin);
2145 	err = wait_for_reset(arg->engine, rq[1], HZ / 2);
2146 	if (err)
2147 		goto out;
2148 
2149 	if (rq[0]->fence.error != 0) {
2150 		pr_err("Normal inflight0 request did not complete\n");
2151 		err = -EINVAL;
2152 		goto out;
2153 	}
2154 
2155 	if (rq[1]->fence.error != -EIO) {
2156 		pr_err("Cancelled inflight1 request did not report -EIO\n");
2157 		err = -EINVAL;
2158 		goto out;
2159 	}
2160 
2161 out:
2162 	i915_request_put(rq[1]);
2163 	i915_request_put(rq[0]);
2164 	if (igt_live_test_end(&t))
2165 		err = -EIO;
2166 	return err;
2167 }
2168 
2169 static int __cancel_queued(struct live_preempt_cancel *arg)
2170 {
2171 	struct i915_request *rq[3] = {};
2172 	struct igt_live_test t;
2173 	int err;
2174 
2175 	/* Full ELSP and one in the wings */
2176 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2177 	if (igt_live_test_begin(&t, arg->engine->i915,
2178 				__func__, arg->engine->name))
2179 		return -EIO;
2180 
2181 	rq[0] = spinner_create_request(&arg->a.spin,
2182 				       arg->a.ctx, arg->engine,
2183 				       MI_ARB_CHECK);
2184 	if (IS_ERR(rq[0]))
2185 		return PTR_ERR(rq[0]);
2186 
2187 	clear_bit(CONTEXT_BANNED, &rq[0]->context->flags);
2188 	i915_request_get(rq[0]);
2189 	i915_request_add(rq[0]);
2190 	if (!igt_wait_for_spinner(&arg->a.spin, rq[0])) {
2191 		err = -EIO;
2192 		goto out;
2193 	}
2194 
2195 	rq[1] = igt_request_alloc(arg->b.ctx, arg->engine);
2196 	if (IS_ERR(rq[1])) {
2197 		err = PTR_ERR(rq[1]);
2198 		goto out;
2199 	}
2200 
2201 	clear_bit(CONTEXT_BANNED, &rq[1]->context->flags);
2202 	i915_request_get(rq[1]);
2203 	err = i915_request_await_dma_fence(rq[1], &rq[0]->fence);
2204 	i915_request_add(rq[1]);
2205 	if (err)
2206 		goto out;
2207 
2208 	rq[2] = spinner_create_request(&arg->b.spin,
2209 				       arg->a.ctx, arg->engine,
2210 				       MI_ARB_CHECK);
2211 	if (IS_ERR(rq[2])) {
2212 		err = PTR_ERR(rq[2]);
2213 		goto out;
2214 	}
2215 
2216 	i915_request_get(rq[2]);
2217 	err = i915_request_await_dma_fence(rq[2], &rq[1]->fence);
2218 	i915_request_add(rq[2]);
2219 	if (err)
2220 		goto out;
2221 
2222 	intel_context_set_banned(rq[2]->context);
2223 	err = intel_engine_pulse(arg->engine);
2224 	if (err)
2225 		goto out;
2226 
2227 	err = wait_for_reset(arg->engine, rq[2], HZ / 2);
2228 	if (err)
2229 		goto out;
2230 
2231 	if (rq[0]->fence.error != -EIO) {
2232 		pr_err("Cancelled inflight0 request did not report -EIO\n");
2233 		err = -EINVAL;
2234 		goto out;
2235 	}
2236 
2237 	if (rq[1]->fence.error != 0) {
2238 		pr_err("Normal inflight1 request did not complete\n");
2239 		err = -EINVAL;
2240 		goto out;
2241 	}
2242 
2243 	if (rq[2]->fence.error != -EIO) {
2244 		pr_err("Cancelled queued request did not report -EIO\n");
2245 		err = -EINVAL;
2246 		goto out;
2247 	}
2248 
2249 out:
2250 	i915_request_put(rq[2]);
2251 	i915_request_put(rq[1]);
2252 	i915_request_put(rq[0]);
2253 	if (igt_live_test_end(&t))
2254 		err = -EIO;
2255 	return err;
2256 }
2257 
2258 static int __cancel_hostile(struct live_preempt_cancel *arg)
2259 {
2260 	struct i915_request *rq;
2261 	int err;
2262 
2263 	/* Preempt cancel non-preemptible spinner in ELSP0 */
2264 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
2265 		return 0;
2266 
2267 	if (!intel_has_reset_engine(arg->engine->gt))
2268 		return 0;
2269 
2270 	GEM_TRACE("%s(%s)\n", __func__, arg->engine->name);
2271 	rq = spinner_create_request(&arg->a.spin,
2272 				    arg->a.ctx, arg->engine,
2273 				    MI_NOOP); /* preemption disabled */
2274 	if (IS_ERR(rq))
2275 		return PTR_ERR(rq);
2276 
2277 	clear_bit(CONTEXT_BANNED, &rq->context->flags);
2278 	i915_request_get(rq);
2279 	i915_request_add(rq);
2280 	if (!igt_wait_for_spinner(&arg->a.spin, rq)) {
2281 		err = -EIO;
2282 		goto out;
2283 	}
2284 
2285 	intel_context_set_banned(rq->context);
2286 	err = intel_engine_pulse(arg->engine); /* force reset */
2287 	if (err)
2288 		goto out;
2289 
2290 	err = wait_for_reset(arg->engine, rq, HZ / 2);
2291 	if (err) {
2292 		pr_err("Cancelled inflight0 request did not reset\n");
2293 		goto out;
2294 	}
2295 
2296 out:
2297 	i915_request_put(rq);
2298 	if (igt_flush_test(arg->engine->i915))
2299 		err = -EIO;
2300 	return err;
2301 }
2302 
2303 static void force_reset_timeout(struct intel_engine_cs *engine)
2304 {
2305 	engine->reset_timeout.probability = 999;
2306 	atomic_set(&engine->reset_timeout.times, -1);
2307 }
2308 
2309 static void cancel_reset_timeout(struct intel_engine_cs *engine)
2310 {
2311 	memset(&engine->reset_timeout, 0, sizeof(engine->reset_timeout));
2312 }
2313 
2314 static int __cancel_fail(struct live_preempt_cancel *arg)
2315 {
2316 	struct intel_engine_cs *engine = arg->engine;
2317 	struct i915_request *rq;
2318 	int err;
2319 
2320 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
2321 		return 0;
2322 
2323 	if (!intel_has_reset_engine(engine->gt))
2324 		return 0;
2325 
2326 	GEM_TRACE("%s(%s)\n", __func__, engine->name);
2327 	rq = spinner_create_request(&arg->a.spin,
2328 				    arg->a.ctx, engine,
2329 				    MI_NOOP); /* preemption disabled */
2330 	if (IS_ERR(rq))
2331 		return PTR_ERR(rq);
2332 
2333 	clear_bit(CONTEXT_BANNED, &rq->context->flags);
2334 	i915_request_get(rq);
2335 	i915_request_add(rq);
2336 	if (!igt_wait_for_spinner(&arg->a.spin, rq)) {
2337 		err = -EIO;
2338 		goto out;
2339 	}
2340 
2341 	intel_context_set_banned(rq->context);
2342 
2343 	err = intel_engine_pulse(engine);
2344 	if (err)
2345 		goto out;
2346 
2347 	force_reset_timeout(engine);
2348 
2349 	/* force preempt reset [failure] */
2350 	while (!engine->execlists.pending[0])
2351 		intel_engine_flush_submission(engine);
2352 	del_timer_sync(&engine->execlists.preempt);
2353 	intel_engine_flush_submission(engine);
2354 
2355 	cancel_reset_timeout(engine);
2356 
2357 	/* after failure, require heartbeats to reset device */
2358 	intel_engine_set_heartbeat(engine, 1);
2359 	err = wait_for_reset(engine, rq, HZ / 2);
2360 	intel_engine_set_heartbeat(engine,
2361 				   engine->defaults.heartbeat_interval_ms);
2362 	if (err) {
2363 		pr_err("Cancelled inflight0 request did not reset\n");
2364 		goto out;
2365 	}
2366 
2367 out:
2368 	i915_request_put(rq);
2369 	if (igt_flush_test(engine->i915))
2370 		err = -EIO;
2371 	return err;
2372 }
2373 
2374 static int live_preempt_cancel(void *arg)
2375 {
2376 	struct intel_gt *gt = arg;
2377 	struct live_preempt_cancel data;
2378 	enum intel_engine_id id;
2379 	int err = -ENOMEM;
2380 
2381 	/*
2382 	 * To cancel an inflight context, we need to first remove it from the
2383 	 * GPU. That sounds like preemption! Plus a little bit of bookkeeping.
2384 	 */
2385 
2386 	if (preempt_client_init(gt, &data.a))
2387 		return -ENOMEM;
2388 	if (preempt_client_init(gt, &data.b))
2389 		goto err_client_a;
2390 
2391 	for_each_engine(data.engine, gt, id) {
2392 		if (!intel_engine_has_preemption(data.engine))
2393 			continue;
2394 
2395 		err = __cancel_active0(&data);
2396 		if (err)
2397 			goto err_wedged;
2398 
2399 		err = __cancel_active1(&data);
2400 		if (err)
2401 			goto err_wedged;
2402 
2403 		err = __cancel_queued(&data);
2404 		if (err)
2405 			goto err_wedged;
2406 
2407 		err = __cancel_hostile(&data);
2408 		if (err)
2409 			goto err_wedged;
2410 
2411 		err = __cancel_fail(&data);
2412 		if (err)
2413 			goto err_wedged;
2414 	}
2415 
2416 	err = 0;
2417 err_client_b:
2418 	preempt_client_fini(&data.b);
2419 err_client_a:
2420 	preempt_client_fini(&data.a);
2421 	return err;
2422 
2423 err_wedged:
2424 	GEM_TRACE_DUMP();
2425 	igt_spinner_end(&data.b.spin);
2426 	igt_spinner_end(&data.a.spin);
2427 	intel_gt_set_wedged(gt);
2428 	goto err_client_b;
2429 }
2430 
2431 static int live_suppress_self_preempt(void *arg)
2432 {
2433 	struct i915_sched_attr attr = { .priority = I915_PRIORITY_MAX };
2434 	struct intel_gt *gt = arg;
2435 	struct intel_engine_cs *engine;
2436 	struct preempt_client a, b;
2437 	enum intel_engine_id id;
2438 	int err = -ENOMEM;
2439 
2440 	/*
2441 	 * Verify that if a preemption request does not cause a change in
2442 	 * the current execution order, the preempt-to-idle injection is
2443 	 * skipped and that we do not accidentally apply it after the CS
2444 	 * completion event.
2445 	 */
2446 
2447 	if (intel_uc_uses_guc_submission(&gt->uc))
2448 		return 0; /* presume black blox */
2449 
2450 	if (intel_vgpu_active(gt->i915))
2451 		return 0; /* GVT forces single port & request submission */
2452 
2453 	if (preempt_client_init(gt, &a))
2454 		return -ENOMEM;
2455 	if (preempt_client_init(gt, &b))
2456 		goto err_client_a;
2457 
2458 	for_each_engine(engine, gt, id) {
2459 		struct i915_request *rq_a, *rq_b;
2460 		int depth;
2461 
2462 		if (!intel_engine_has_preemption(engine))
2463 			continue;
2464 
2465 		if (igt_flush_test(gt->i915))
2466 			goto err_wedged;
2467 
2468 		st_engine_heartbeat_disable(engine);
2469 		engine->execlists.preempt_hang.count = 0;
2470 
2471 		rq_a = spinner_create_request(&a.spin,
2472 					      a.ctx, engine,
2473 					      MI_NOOP);
2474 		if (IS_ERR(rq_a)) {
2475 			err = PTR_ERR(rq_a);
2476 			st_engine_heartbeat_enable(engine);
2477 			goto err_client_b;
2478 		}
2479 
2480 		i915_request_add(rq_a);
2481 		if (!igt_wait_for_spinner(&a.spin, rq_a)) {
2482 			pr_err("First client failed to start\n");
2483 			st_engine_heartbeat_enable(engine);
2484 			goto err_wedged;
2485 		}
2486 
2487 		/* Keep postponing the timer to avoid premature slicing */
2488 		mod_timer(&engine->execlists.timer, jiffies + HZ);
2489 		for (depth = 0; depth < 8; depth++) {
2490 			rq_b = spinner_create_request(&b.spin,
2491 						      b.ctx, engine,
2492 						      MI_NOOP);
2493 			if (IS_ERR(rq_b)) {
2494 				err = PTR_ERR(rq_b);
2495 				st_engine_heartbeat_enable(engine);
2496 				goto err_client_b;
2497 			}
2498 			i915_request_add(rq_b);
2499 
2500 			GEM_BUG_ON(i915_request_completed(rq_a));
2501 			engine->sched_engine->schedule(rq_a, &attr);
2502 			igt_spinner_end(&a.spin);
2503 
2504 			if (!igt_wait_for_spinner(&b.spin, rq_b)) {
2505 				pr_err("Second client failed to start\n");
2506 				st_engine_heartbeat_enable(engine);
2507 				goto err_wedged;
2508 			}
2509 
2510 			swap(a, b);
2511 			rq_a = rq_b;
2512 		}
2513 		igt_spinner_end(&a.spin);
2514 
2515 		if (engine->execlists.preempt_hang.count) {
2516 			pr_err("Preemption on %s recorded x%d, depth %d; should have been suppressed!\n",
2517 			       engine->name,
2518 			       engine->execlists.preempt_hang.count,
2519 			       depth);
2520 			st_engine_heartbeat_enable(engine);
2521 			err = -EINVAL;
2522 			goto err_client_b;
2523 		}
2524 
2525 		st_engine_heartbeat_enable(engine);
2526 		if (igt_flush_test(gt->i915))
2527 			goto err_wedged;
2528 	}
2529 
2530 	err = 0;
2531 err_client_b:
2532 	preempt_client_fini(&b);
2533 err_client_a:
2534 	preempt_client_fini(&a);
2535 	return err;
2536 
2537 err_wedged:
2538 	igt_spinner_end(&b.spin);
2539 	igt_spinner_end(&a.spin);
2540 	intel_gt_set_wedged(gt);
2541 	err = -EIO;
2542 	goto err_client_b;
2543 }
2544 
2545 static int live_chain_preempt(void *arg)
2546 {
2547 	struct intel_gt *gt = arg;
2548 	struct intel_engine_cs *engine;
2549 	struct preempt_client hi, lo;
2550 	enum intel_engine_id id;
2551 	int err = -ENOMEM;
2552 
2553 	/*
2554 	 * Build a chain AB...BA between two contexts (A, B) and request
2555 	 * preemption of the last request. It should then complete before
2556 	 * the previously submitted spinner in B.
2557 	 */
2558 
2559 	if (preempt_client_init(gt, &hi))
2560 		return -ENOMEM;
2561 
2562 	if (preempt_client_init(gt, &lo))
2563 		goto err_client_hi;
2564 
2565 	for_each_engine(engine, gt, id) {
2566 		struct i915_sched_attr attr = { .priority = I915_PRIORITY_MAX };
2567 		struct igt_live_test t;
2568 		struct i915_request *rq;
2569 		int ring_size, count, i;
2570 
2571 		if (!intel_engine_has_preemption(engine))
2572 			continue;
2573 
2574 		rq = spinner_create_request(&lo.spin,
2575 					    lo.ctx, engine,
2576 					    MI_ARB_CHECK);
2577 		if (IS_ERR(rq))
2578 			goto err_wedged;
2579 
2580 		i915_request_get(rq);
2581 		i915_request_add(rq);
2582 
2583 		ring_size = rq->wa_tail - rq->head;
2584 		if (ring_size < 0)
2585 			ring_size += rq->ring->size;
2586 		ring_size = rq->ring->size / ring_size;
2587 		pr_debug("%s(%s): Using maximum of %d requests\n",
2588 			 __func__, engine->name, ring_size);
2589 
2590 		igt_spinner_end(&lo.spin);
2591 		if (i915_request_wait(rq, 0, HZ / 2) < 0) {
2592 			pr_err("Timed out waiting to flush %s\n", engine->name);
2593 			i915_request_put(rq);
2594 			goto err_wedged;
2595 		}
2596 		i915_request_put(rq);
2597 
2598 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
2599 			err = -EIO;
2600 			goto err_wedged;
2601 		}
2602 
2603 		for_each_prime_number_from(count, 1, ring_size) {
2604 			rq = spinner_create_request(&hi.spin,
2605 						    hi.ctx, engine,
2606 						    MI_ARB_CHECK);
2607 			if (IS_ERR(rq))
2608 				goto err_wedged;
2609 			i915_request_add(rq);
2610 			if (!igt_wait_for_spinner(&hi.spin, rq))
2611 				goto err_wedged;
2612 
2613 			rq = spinner_create_request(&lo.spin,
2614 						    lo.ctx, engine,
2615 						    MI_ARB_CHECK);
2616 			if (IS_ERR(rq))
2617 				goto err_wedged;
2618 			i915_request_add(rq);
2619 
2620 			for (i = 0; i < count; i++) {
2621 				rq = igt_request_alloc(lo.ctx, engine);
2622 				if (IS_ERR(rq))
2623 					goto err_wedged;
2624 				i915_request_add(rq);
2625 			}
2626 
2627 			rq = igt_request_alloc(hi.ctx, engine);
2628 			if (IS_ERR(rq))
2629 				goto err_wedged;
2630 
2631 			i915_request_get(rq);
2632 			i915_request_add(rq);
2633 			engine->sched_engine->schedule(rq, &attr);
2634 
2635 			igt_spinner_end(&hi.spin);
2636 			if (i915_request_wait(rq, 0, HZ / 5) < 0) {
2637 				struct drm_printer p =
2638 					drm_info_printer(gt->i915->drm.dev);
2639 
2640 				pr_err("Failed to preempt over chain of %d\n",
2641 				       count);
2642 				intel_engine_dump(engine, &p,
2643 						  "%s\n", engine->name);
2644 				i915_request_put(rq);
2645 				goto err_wedged;
2646 			}
2647 			igt_spinner_end(&lo.spin);
2648 			i915_request_put(rq);
2649 
2650 			rq = igt_request_alloc(lo.ctx, engine);
2651 			if (IS_ERR(rq))
2652 				goto err_wedged;
2653 
2654 			i915_request_get(rq);
2655 			i915_request_add(rq);
2656 
2657 			if (i915_request_wait(rq, 0, HZ / 5) < 0) {
2658 				struct drm_printer p =
2659 					drm_info_printer(gt->i915->drm.dev);
2660 
2661 				pr_err("Failed to flush low priority chain of %d requests\n",
2662 				       count);
2663 				intel_engine_dump(engine, &p,
2664 						  "%s\n", engine->name);
2665 
2666 				i915_request_put(rq);
2667 				goto err_wedged;
2668 			}
2669 			i915_request_put(rq);
2670 		}
2671 
2672 		if (igt_live_test_end(&t)) {
2673 			err = -EIO;
2674 			goto err_wedged;
2675 		}
2676 	}
2677 
2678 	err = 0;
2679 err_client_lo:
2680 	preempt_client_fini(&lo);
2681 err_client_hi:
2682 	preempt_client_fini(&hi);
2683 	return err;
2684 
2685 err_wedged:
2686 	igt_spinner_end(&hi.spin);
2687 	igt_spinner_end(&lo.spin);
2688 	intel_gt_set_wedged(gt);
2689 	err = -EIO;
2690 	goto err_client_lo;
2691 }
2692 
2693 static int create_gang(struct intel_engine_cs *engine,
2694 		       struct i915_request **prev)
2695 {
2696 	struct drm_i915_gem_object *obj;
2697 	struct intel_context *ce;
2698 	struct i915_request *rq;
2699 	struct i915_vma *vma;
2700 	u32 *cs;
2701 	int err;
2702 
2703 	ce = intel_context_create(engine);
2704 	if (IS_ERR(ce))
2705 		return PTR_ERR(ce);
2706 
2707 	obj = i915_gem_object_create_internal(engine->i915, 4096);
2708 	if (IS_ERR(obj)) {
2709 		err = PTR_ERR(obj);
2710 		goto err_ce;
2711 	}
2712 
2713 	vma = i915_vma_instance(obj, ce->vm, NULL);
2714 	if (IS_ERR(vma)) {
2715 		err = PTR_ERR(vma);
2716 		goto err_obj;
2717 	}
2718 
2719 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
2720 	if (err)
2721 		goto err_obj;
2722 
2723 	cs = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
2724 	if (IS_ERR(cs)) {
2725 		err = PTR_ERR(cs);
2726 		goto err_obj;
2727 	}
2728 
2729 	/* Semaphore target: spin until zero */
2730 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
2731 
2732 	*cs++ = MI_SEMAPHORE_WAIT |
2733 		MI_SEMAPHORE_POLL |
2734 		MI_SEMAPHORE_SAD_EQ_SDD;
2735 	*cs++ = 0;
2736 	*cs++ = lower_32_bits(vma->node.start);
2737 	*cs++ = upper_32_bits(vma->node.start);
2738 
2739 	if (*prev) {
2740 		u64 offset = (*prev)->batch->node.start;
2741 
2742 		/* Terminate the spinner in the next lower priority batch. */
2743 		*cs++ = MI_STORE_DWORD_IMM_GEN4;
2744 		*cs++ = lower_32_bits(offset);
2745 		*cs++ = upper_32_bits(offset);
2746 		*cs++ = 0;
2747 	}
2748 
2749 	*cs++ = MI_BATCH_BUFFER_END;
2750 	i915_gem_object_flush_map(obj);
2751 	i915_gem_object_unpin_map(obj);
2752 
2753 	rq = intel_context_create_request(ce);
2754 	if (IS_ERR(rq)) {
2755 		err = PTR_ERR(rq);
2756 		goto err_obj;
2757 	}
2758 
2759 	rq->batch = i915_vma_get(vma);
2760 	i915_request_get(rq);
2761 
2762 	i915_vma_lock(vma);
2763 	err = i915_request_await_object(rq, vma->obj, false);
2764 	if (!err)
2765 		err = i915_vma_move_to_active(vma, rq, 0);
2766 	if (!err)
2767 		err = rq->engine->emit_bb_start(rq,
2768 						vma->node.start,
2769 						PAGE_SIZE, 0);
2770 	i915_vma_unlock(vma);
2771 	i915_request_add(rq);
2772 	if (err)
2773 		goto err_rq;
2774 
2775 	i915_gem_object_put(obj);
2776 	intel_context_put(ce);
2777 
2778 	rq->mock.link.next = &(*prev)->mock.link;
2779 	*prev = rq;
2780 	return 0;
2781 
2782 err_rq:
2783 	i915_vma_put(rq->batch);
2784 	i915_request_put(rq);
2785 err_obj:
2786 	i915_gem_object_put(obj);
2787 err_ce:
2788 	intel_context_put(ce);
2789 	return err;
2790 }
2791 
2792 static int __live_preempt_ring(struct intel_engine_cs *engine,
2793 			       struct igt_spinner *spin,
2794 			       int queue_sz, int ring_sz)
2795 {
2796 	struct intel_context *ce[2] = {};
2797 	struct i915_request *rq;
2798 	struct igt_live_test t;
2799 	int err = 0;
2800 	int n;
2801 
2802 	if (igt_live_test_begin(&t, engine->i915, __func__, engine->name))
2803 		return -EIO;
2804 
2805 	for (n = 0; n < ARRAY_SIZE(ce); n++) {
2806 		struct intel_context *tmp;
2807 
2808 		tmp = intel_context_create(engine);
2809 		if (IS_ERR(tmp)) {
2810 			err = PTR_ERR(tmp);
2811 			goto err_ce;
2812 		}
2813 
2814 		tmp->ring_size = ring_sz;
2815 
2816 		err = intel_context_pin(tmp);
2817 		if (err) {
2818 			intel_context_put(tmp);
2819 			goto err_ce;
2820 		}
2821 
2822 		memset32(tmp->ring->vaddr,
2823 			 0xdeadbeef, /* trigger a hang if executed */
2824 			 tmp->ring->vma->size / sizeof(u32));
2825 
2826 		ce[n] = tmp;
2827 	}
2828 
2829 	rq = igt_spinner_create_request(spin, ce[0], MI_ARB_CHECK);
2830 	if (IS_ERR(rq)) {
2831 		err = PTR_ERR(rq);
2832 		goto err_ce;
2833 	}
2834 
2835 	i915_request_get(rq);
2836 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
2837 	i915_request_add(rq);
2838 
2839 	if (!igt_wait_for_spinner(spin, rq)) {
2840 		intel_gt_set_wedged(engine->gt);
2841 		i915_request_put(rq);
2842 		err = -ETIME;
2843 		goto err_ce;
2844 	}
2845 
2846 	/* Fill the ring, until we will cause a wrap */
2847 	n = 0;
2848 	while (ce[0]->ring->tail - rq->wa_tail <= queue_sz) {
2849 		struct i915_request *tmp;
2850 
2851 		tmp = intel_context_create_request(ce[0]);
2852 		if (IS_ERR(tmp)) {
2853 			err = PTR_ERR(tmp);
2854 			i915_request_put(rq);
2855 			goto err_ce;
2856 		}
2857 
2858 		i915_request_add(tmp);
2859 		intel_engine_flush_submission(engine);
2860 		n++;
2861 	}
2862 	intel_engine_flush_submission(engine);
2863 	pr_debug("%s: Filled %d with %d nop tails {size:%x, tail:%x, emit:%x, rq.tail:%x}\n",
2864 		 engine->name, queue_sz, n,
2865 		 ce[0]->ring->size,
2866 		 ce[0]->ring->tail,
2867 		 ce[0]->ring->emit,
2868 		 rq->tail);
2869 	i915_request_put(rq);
2870 
2871 	/* Create a second request to preempt the first ring */
2872 	rq = intel_context_create_request(ce[1]);
2873 	if (IS_ERR(rq)) {
2874 		err = PTR_ERR(rq);
2875 		goto err_ce;
2876 	}
2877 
2878 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
2879 	i915_request_get(rq);
2880 	i915_request_add(rq);
2881 
2882 	err = wait_for_submit(engine, rq, HZ / 2);
2883 	i915_request_put(rq);
2884 	if (err) {
2885 		pr_err("%s: preemption request was not submitted\n",
2886 		       engine->name);
2887 		err = -ETIME;
2888 	}
2889 
2890 	pr_debug("%s: ring[0]:{ tail:%x, emit:%x }, ring[1]:{ tail:%x, emit:%x }\n",
2891 		 engine->name,
2892 		 ce[0]->ring->tail, ce[0]->ring->emit,
2893 		 ce[1]->ring->tail, ce[1]->ring->emit);
2894 
2895 err_ce:
2896 	intel_engine_flush_submission(engine);
2897 	igt_spinner_end(spin);
2898 	for (n = 0; n < ARRAY_SIZE(ce); n++) {
2899 		if (IS_ERR_OR_NULL(ce[n]))
2900 			break;
2901 
2902 		intel_context_unpin(ce[n]);
2903 		intel_context_put(ce[n]);
2904 	}
2905 	if (igt_live_test_end(&t))
2906 		err = -EIO;
2907 	return err;
2908 }
2909 
2910 static int live_preempt_ring(void *arg)
2911 {
2912 	struct intel_gt *gt = arg;
2913 	struct intel_engine_cs *engine;
2914 	struct igt_spinner spin;
2915 	enum intel_engine_id id;
2916 	int err = 0;
2917 
2918 	/*
2919 	 * Check that we rollback large chunks of a ring in order to do a
2920 	 * preemption event. Similar to live_unlite_ring, but looking at
2921 	 * ring size rather than the impact of intel_ring_direction().
2922 	 */
2923 
2924 	if (igt_spinner_init(&spin, gt))
2925 		return -ENOMEM;
2926 
2927 	for_each_engine(engine, gt, id) {
2928 		int n;
2929 
2930 		if (!intel_engine_has_preemption(engine))
2931 			continue;
2932 
2933 		if (!intel_engine_can_store_dword(engine))
2934 			continue;
2935 
2936 		st_engine_heartbeat_disable(engine);
2937 
2938 		for (n = 0; n <= 3; n++) {
2939 			err = __live_preempt_ring(engine, &spin,
2940 						  n * SZ_4K / 4, SZ_4K);
2941 			if (err)
2942 				break;
2943 		}
2944 
2945 		st_engine_heartbeat_enable(engine);
2946 		if (err)
2947 			break;
2948 	}
2949 
2950 	igt_spinner_fini(&spin);
2951 	return err;
2952 }
2953 
2954 static int live_preempt_gang(void *arg)
2955 {
2956 	struct intel_gt *gt = arg;
2957 	struct intel_engine_cs *engine;
2958 	enum intel_engine_id id;
2959 
2960 	/*
2961 	 * Build as long a chain of preempters as we can, with each
2962 	 * request higher priority than the last. Once we are ready, we release
2963 	 * the last batch which then precolates down the chain, each releasing
2964 	 * the next oldest in turn. The intent is to simply push as hard as we
2965 	 * can with the number of preemptions, trying to exceed narrow HW
2966 	 * limits. At a minimum, we insist that we can sort all the user
2967 	 * high priority levels into execution order.
2968 	 */
2969 
2970 	for_each_engine(engine, gt, id) {
2971 		struct i915_request *rq = NULL;
2972 		struct igt_live_test t;
2973 		IGT_TIMEOUT(end_time);
2974 		int prio = 0;
2975 		int err = 0;
2976 		u32 *cs;
2977 
2978 		if (!intel_engine_has_preemption(engine))
2979 			continue;
2980 
2981 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name))
2982 			return -EIO;
2983 
2984 		do {
2985 			struct i915_sched_attr attr = { .priority = prio++ };
2986 
2987 			err = create_gang(engine, &rq);
2988 			if (err)
2989 				break;
2990 
2991 			/* Submit each spinner at increasing priority */
2992 			engine->sched_engine->schedule(rq, &attr);
2993 		} while (prio <= I915_PRIORITY_MAX &&
2994 			 !__igt_timeout(end_time, NULL));
2995 		pr_debug("%s: Preempt chain of %d requests\n",
2996 			 engine->name, prio);
2997 
2998 		/*
2999 		 * Such that the last spinner is the highest priority and
3000 		 * should execute first. When that spinner completes,
3001 		 * it will terminate the next lowest spinner until there
3002 		 * are no more spinners and the gang is complete.
3003 		 */
3004 		cs = i915_gem_object_pin_map_unlocked(rq->batch->obj, I915_MAP_WC);
3005 		if (!IS_ERR(cs)) {
3006 			*cs = 0;
3007 			i915_gem_object_unpin_map(rq->batch->obj);
3008 		} else {
3009 			err = PTR_ERR(cs);
3010 			intel_gt_set_wedged(gt);
3011 		}
3012 
3013 		while (rq) { /* wait for each rq from highest to lowest prio */
3014 			struct i915_request *n = list_next_entry(rq, mock.link);
3015 
3016 			if (err == 0 && i915_request_wait(rq, 0, HZ / 5) < 0) {
3017 				struct drm_printer p =
3018 					drm_info_printer(engine->i915->drm.dev);
3019 
3020 				pr_err("Failed to flush chain of %d requests, at %d\n",
3021 				       prio, rq_prio(rq));
3022 				intel_engine_dump(engine, &p,
3023 						  "%s\n", engine->name);
3024 
3025 				err = -ETIME;
3026 			}
3027 
3028 			i915_vma_put(rq->batch);
3029 			i915_request_put(rq);
3030 			rq = n;
3031 		}
3032 
3033 		if (igt_live_test_end(&t))
3034 			err = -EIO;
3035 		if (err)
3036 			return err;
3037 	}
3038 
3039 	return 0;
3040 }
3041 
3042 static struct i915_vma *
3043 create_gpr_user(struct intel_engine_cs *engine,
3044 		struct i915_vma *result,
3045 		unsigned int offset)
3046 {
3047 	struct drm_i915_gem_object *obj;
3048 	struct i915_vma *vma;
3049 	u32 *cs;
3050 	int err;
3051 	int i;
3052 
3053 	obj = i915_gem_object_create_internal(engine->i915, 4096);
3054 	if (IS_ERR(obj))
3055 		return ERR_CAST(obj);
3056 
3057 	vma = i915_vma_instance(obj, result->vm, NULL);
3058 	if (IS_ERR(vma)) {
3059 		i915_gem_object_put(obj);
3060 		return vma;
3061 	}
3062 
3063 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
3064 	if (err) {
3065 		i915_vma_put(vma);
3066 		return ERR_PTR(err);
3067 	}
3068 
3069 	cs = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
3070 	if (IS_ERR(cs)) {
3071 		i915_vma_put(vma);
3072 		return ERR_CAST(cs);
3073 	}
3074 
3075 	/* All GPR are clear for new contexts. We use GPR(0) as a constant */
3076 	*cs++ = MI_LOAD_REGISTER_IMM(1);
3077 	*cs++ = CS_GPR(engine, 0);
3078 	*cs++ = 1;
3079 
3080 	for (i = 1; i < NUM_GPR; i++) {
3081 		u64 addr;
3082 
3083 		/*
3084 		 * Perform: GPR[i]++
3085 		 *
3086 		 * As we read and write into the context saved GPR[i], if
3087 		 * we restart this batch buffer from an earlier point, we
3088 		 * will repeat the increment and store a value > 1.
3089 		 */
3090 		*cs++ = MI_MATH(4);
3091 		*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(i));
3092 		*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(0));
3093 		*cs++ = MI_MATH_ADD;
3094 		*cs++ = MI_MATH_STORE(MI_MATH_REG(i), MI_MATH_REG_ACCU);
3095 
3096 		addr = result->node.start + offset + i * sizeof(*cs);
3097 		*cs++ = MI_STORE_REGISTER_MEM_GEN8;
3098 		*cs++ = CS_GPR(engine, 2 * i);
3099 		*cs++ = lower_32_bits(addr);
3100 		*cs++ = upper_32_bits(addr);
3101 
3102 		*cs++ = MI_SEMAPHORE_WAIT |
3103 			MI_SEMAPHORE_POLL |
3104 			MI_SEMAPHORE_SAD_GTE_SDD;
3105 		*cs++ = i;
3106 		*cs++ = lower_32_bits(result->node.start);
3107 		*cs++ = upper_32_bits(result->node.start);
3108 	}
3109 
3110 	*cs++ = MI_BATCH_BUFFER_END;
3111 	i915_gem_object_flush_map(obj);
3112 	i915_gem_object_unpin_map(obj);
3113 
3114 	return vma;
3115 }
3116 
3117 static struct i915_vma *create_global(struct intel_gt *gt, size_t sz)
3118 {
3119 	struct drm_i915_gem_object *obj;
3120 	struct i915_vma *vma;
3121 	int err;
3122 
3123 	obj = i915_gem_object_create_internal(gt->i915, sz);
3124 	if (IS_ERR(obj))
3125 		return ERR_CAST(obj);
3126 
3127 	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
3128 	if (IS_ERR(vma)) {
3129 		i915_gem_object_put(obj);
3130 		return vma;
3131 	}
3132 
3133 	err = i915_ggtt_pin(vma, NULL, 0, 0);
3134 	if (err) {
3135 		i915_vma_put(vma);
3136 		return ERR_PTR(err);
3137 	}
3138 
3139 	return vma;
3140 }
3141 
3142 static struct i915_request *
3143 create_gpr_client(struct intel_engine_cs *engine,
3144 		  struct i915_vma *global,
3145 		  unsigned int offset)
3146 {
3147 	struct i915_vma *batch, *vma;
3148 	struct intel_context *ce;
3149 	struct i915_request *rq;
3150 	int err;
3151 
3152 	ce = intel_context_create(engine);
3153 	if (IS_ERR(ce))
3154 		return ERR_CAST(ce);
3155 
3156 	vma = i915_vma_instance(global->obj, ce->vm, NULL);
3157 	if (IS_ERR(vma)) {
3158 		err = PTR_ERR(vma);
3159 		goto out_ce;
3160 	}
3161 
3162 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
3163 	if (err)
3164 		goto out_ce;
3165 
3166 	batch = create_gpr_user(engine, vma, offset);
3167 	if (IS_ERR(batch)) {
3168 		err = PTR_ERR(batch);
3169 		goto out_vma;
3170 	}
3171 
3172 	rq = intel_context_create_request(ce);
3173 	if (IS_ERR(rq)) {
3174 		err = PTR_ERR(rq);
3175 		goto out_batch;
3176 	}
3177 
3178 	i915_vma_lock(vma);
3179 	err = i915_request_await_object(rq, vma->obj, false);
3180 	if (!err)
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_request_await_object(rq, batch->obj, false);
3187 	if (!err)
3188 		err = i915_vma_move_to_active(batch, rq, 0);
3189 	if (!err)
3190 		err = rq->engine->emit_bb_start(rq,
3191 						batch->node.start,
3192 						PAGE_SIZE, 0);
3193 	i915_vma_unlock(batch);
3194 	i915_vma_unpin(batch);
3195 
3196 	if (!err)
3197 		i915_request_get(rq);
3198 	i915_request_add(rq);
3199 
3200 out_batch:
3201 	i915_vma_put(batch);
3202 out_vma:
3203 	i915_vma_unpin(vma);
3204 out_ce:
3205 	intel_context_put(ce);
3206 	return err ? ERR_PTR(err) : rq;
3207 }
3208 
3209 static int preempt_user(struct intel_engine_cs *engine,
3210 			struct i915_vma *global,
3211 			int id)
3212 {
3213 	struct i915_sched_attr attr = {
3214 		.priority = I915_PRIORITY_MAX
3215 	};
3216 	struct i915_request *rq;
3217 	int err = 0;
3218 	u32 *cs;
3219 
3220 	rq = intel_engine_create_kernel_request(engine);
3221 	if (IS_ERR(rq))
3222 		return PTR_ERR(rq);
3223 
3224 	cs = intel_ring_begin(rq, 4);
3225 	if (IS_ERR(cs)) {
3226 		i915_request_add(rq);
3227 		return PTR_ERR(cs);
3228 	}
3229 
3230 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
3231 	*cs++ = i915_ggtt_offset(global);
3232 	*cs++ = 0;
3233 	*cs++ = id;
3234 
3235 	intel_ring_advance(rq, cs);
3236 
3237 	i915_request_get(rq);
3238 	i915_request_add(rq);
3239 
3240 	engine->sched_engine->schedule(rq, &attr);
3241 
3242 	if (i915_request_wait(rq, 0, HZ / 2) < 0)
3243 		err = -ETIME;
3244 	i915_request_put(rq);
3245 
3246 	return err;
3247 }
3248 
3249 static int live_preempt_user(void *arg)
3250 {
3251 	struct intel_gt *gt = arg;
3252 	struct intel_engine_cs *engine;
3253 	struct i915_vma *global;
3254 	enum intel_engine_id id;
3255 	u32 *result;
3256 	int err = 0;
3257 
3258 	/*
3259 	 * In our other tests, we look at preemption in carefully
3260 	 * controlled conditions in the ringbuffer. Since most of the
3261 	 * time is spent in user batches, most of our preemptions naturally
3262 	 * occur there. We want to verify that when we preempt inside a batch
3263 	 * we continue on from the current instruction and do not roll back
3264 	 * to the start, or another earlier arbitration point.
3265 	 *
3266 	 * To verify this, we create a batch which is a mixture of
3267 	 * MI_MATH (gpr++) MI_SRM (gpr) and preemption points. Then with
3268 	 * a few preempting contexts thrown into the mix, we look for any
3269 	 * repeated instructions (which show up as incorrect values).
3270 	 */
3271 
3272 	global = create_global(gt, 4096);
3273 	if (IS_ERR(global))
3274 		return PTR_ERR(global);
3275 
3276 	result = i915_gem_object_pin_map_unlocked(global->obj, I915_MAP_WC);
3277 	if (IS_ERR(result)) {
3278 		i915_vma_unpin_and_release(&global, 0);
3279 		return PTR_ERR(result);
3280 	}
3281 
3282 	for_each_engine(engine, gt, id) {
3283 		struct i915_request *client[3] = {};
3284 		struct igt_live_test t;
3285 		int i;
3286 
3287 		if (!intel_engine_has_preemption(engine))
3288 			continue;
3289 
3290 		if (GRAPHICS_VER(gt->i915) == 8 && engine->class != RENDER_CLASS)
3291 			continue; /* we need per-context GPR */
3292 
3293 		if (igt_live_test_begin(&t, gt->i915, __func__, engine->name)) {
3294 			err = -EIO;
3295 			break;
3296 		}
3297 
3298 		memset(result, 0, 4096);
3299 
3300 		for (i = 0; i < ARRAY_SIZE(client); i++) {
3301 			struct i915_request *rq;
3302 
3303 			rq = create_gpr_client(engine, global,
3304 					       NUM_GPR * i * sizeof(u32));
3305 			if (IS_ERR(rq)) {
3306 				err = PTR_ERR(rq);
3307 				goto end_test;
3308 			}
3309 
3310 			client[i] = rq;
3311 		}
3312 
3313 		/* Continuously preempt the set of 3 running contexts */
3314 		for (i = 1; i <= NUM_GPR; i++) {
3315 			err = preempt_user(engine, global, i);
3316 			if (err)
3317 				goto end_test;
3318 		}
3319 
3320 		if (READ_ONCE(result[0]) != NUM_GPR) {
3321 			pr_err("%s: Failed to release semaphore\n",
3322 			       engine->name);
3323 			err = -EIO;
3324 			goto end_test;
3325 		}
3326 
3327 		for (i = 0; i < ARRAY_SIZE(client); i++) {
3328 			int gpr;
3329 
3330 			if (i915_request_wait(client[i], 0, HZ / 2) < 0) {
3331 				err = -ETIME;
3332 				goto end_test;
3333 			}
3334 
3335 			for (gpr = 1; gpr < NUM_GPR; gpr++) {
3336 				if (result[NUM_GPR * i + gpr] != 1) {
3337 					pr_err("%s: Invalid result, client %d, gpr %d, result: %d\n",
3338 					       engine->name,
3339 					       i, gpr, result[NUM_GPR * i + gpr]);
3340 					err = -EINVAL;
3341 					goto end_test;
3342 				}
3343 			}
3344 		}
3345 
3346 end_test:
3347 		for (i = 0; i < ARRAY_SIZE(client); i++) {
3348 			if (!client[i])
3349 				break;
3350 
3351 			i915_request_put(client[i]);
3352 		}
3353 
3354 		/* Flush the semaphores on error */
3355 		smp_store_mb(result[0], -1);
3356 		if (igt_live_test_end(&t))
3357 			err = -EIO;
3358 		if (err)
3359 			break;
3360 	}
3361 
3362 	i915_vma_unpin_and_release(&global, I915_VMA_RELEASE_MAP);
3363 	return err;
3364 }
3365 
3366 static int live_preempt_timeout(void *arg)
3367 {
3368 	struct intel_gt *gt = arg;
3369 	struct i915_gem_context *ctx_hi, *ctx_lo;
3370 	struct igt_spinner spin_lo;
3371 	struct intel_engine_cs *engine;
3372 	enum intel_engine_id id;
3373 	int err = -ENOMEM;
3374 
3375 	/*
3376 	 * Check that we force preemption to occur by cancelling the previous
3377 	 * context if it refuses to yield the GPU.
3378 	 */
3379 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
3380 		return 0;
3381 
3382 	if (!intel_has_reset_engine(gt))
3383 		return 0;
3384 
3385 	if (igt_spinner_init(&spin_lo, gt))
3386 		return -ENOMEM;
3387 
3388 	ctx_hi = kernel_context(gt->i915, NULL);
3389 	if (!ctx_hi)
3390 		goto err_spin_lo;
3391 	ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
3392 
3393 	ctx_lo = kernel_context(gt->i915, NULL);
3394 	if (!ctx_lo)
3395 		goto err_ctx_hi;
3396 	ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
3397 
3398 	for_each_engine(engine, gt, id) {
3399 		unsigned long saved_timeout;
3400 		struct i915_request *rq;
3401 
3402 		if (!intel_engine_has_preemption(engine))
3403 			continue;
3404 
3405 		rq = spinner_create_request(&spin_lo, ctx_lo, engine,
3406 					    MI_NOOP); /* preemption disabled */
3407 		if (IS_ERR(rq)) {
3408 			err = PTR_ERR(rq);
3409 			goto err_ctx_lo;
3410 		}
3411 
3412 		i915_request_add(rq);
3413 		if (!igt_wait_for_spinner(&spin_lo, rq)) {
3414 			intel_gt_set_wedged(gt);
3415 			err = -EIO;
3416 			goto err_ctx_lo;
3417 		}
3418 
3419 		rq = igt_request_alloc(ctx_hi, engine);
3420 		if (IS_ERR(rq)) {
3421 			igt_spinner_end(&spin_lo);
3422 			err = PTR_ERR(rq);
3423 			goto err_ctx_lo;
3424 		}
3425 
3426 		/* Flush the previous CS ack before changing timeouts */
3427 		while (READ_ONCE(engine->execlists.pending[0]))
3428 			cpu_relax();
3429 
3430 		saved_timeout = engine->props.preempt_timeout_ms;
3431 		engine->props.preempt_timeout_ms = 1; /* in ms, -> 1 jiffie */
3432 
3433 		i915_request_get(rq);
3434 		i915_request_add(rq);
3435 
3436 		intel_engine_flush_submission(engine);
3437 		engine->props.preempt_timeout_ms = saved_timeout;
3438 
3439 		if (i915_request_wait(rq, 0, HZ / 10) < 0) {
3440 			intel_gt_set_wedged(gt);
3441 			i915_request_put(rq);
3442 			err = -ETIME;
3443 			goto err_ctx_lo;
3444 		}
3445 
3446 		igt_spinner_end(&spin_lo);
3447 		i915_request_put(rq);
3448 	}
3449 
3450 	err = 0;
3451 err_ctx_lo:
3452 	kernel_context_close(ctx_lo);
3453 err_ctx_hi:
3454 	kernel_context_close(ctx_hi);
3455 err_spin_lo:
3456 	igt_spinner_fini(&spin_lo);
3457 	return err;
3458 }
3459 
3460 static int random_range(struct rnd_state *rnd, int min, int max)
3461 {
3462 	return i915_prandom_u32_max_state(max - min, rnd) + min;
3463 }
3464 
3465 static int random_priority(struct rnd_state *rnd)
3466 {
3467 	return random_range(rnd, I915_PRIORITY_MIN, I915_PRIORITY_MAX);
3468 }
3469 
3470 struct preempt_smoke {
3471 	struct intel_gt *gt;
3472 	struct i915_gem_context **contexts;
3473 	struct intel_engine_cs *engine;
3474 	struct drm_i915_gem_object *batch;
3475 	unsigned int ncontext;
3476 	struct rnd_state prng;
3477 	unsigned long count;
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_request_await_object(rq, vma->obj, false);
3519 		if (!err)
3520 			err = i915_vma_move_to_active(vma, rq, 0);
3521 		if (!err)
3522 			err = rq->engine->emit_bb_start(rq,
3523 							vma->node.start,
3524 							PAGE_SIZE, 0);
3525 		i915_vma_unlock(vma);
3526 	}
3527 
3528 	i915_request_add(rq);
3529 
3530 unpin:
3531 	if (vma)
3532 		i915_vma_unpin(vma);
3533 
3534 	return err;
3535 }
3536 
3537 static int smoke_crescendo_thread(void *arg)
3538 {
3539 	struct preempt_smoke *smoke = arg;
3540 	IGT_TIMEOUT(end_time);
3541 	unsigned long count;
3542 
3543 	count = 0;
3544 	do {
3545 		struct i915_gem_context *ctx = smoke_context(smoke);
3546 		int err;
3547 
3548 		err = smoke_submit(smoke,
3549 				   ctx, count % I915_PRIORITY_MAX,
3550 				   smoke->batch);
3551 		if (err)
3552 			return err;
3553 
3554 		count++;
3555 	} while (count < smoke->ncontext && !__igt_timeout(end_time, NULL));
3556 
3557 	smoke->count = count;
3558 	return 0;
3559 }
3560 
3561 static int smoke_crescendo(struct preempt_smoke *smoke, unsigned int flags)
3562 #define BATCH BIT(0)
3563 {
3564 	struct task_struct *tsk[I915_NUM_ENGINES] = {};
3565 	struct preempt_smoke *arg;
3566 	struct intel_engine_cs *engine;
3567 	enum intel_engine_id id;
3568 	unsigned long count;
3569 	int err = 0;
3570 
3571 	arg = kmalloc_array(I915_NUM_ENGINES, sizeof(*arg), GFP_KERNEL);
3572 	if (!arg)
3573 		return -ENOMEM;
3574 
3575 	for_each_engine(engine, smoke->gt, id) {
3576 		arg[id] = *smoke;
3577 		arg[id].engine = engine;
3578 		if (!(flags & BATCH))
3579 			arg[id].batch = NULL;
3580 		arg[id].count = 0;
3581 
3582 		tsk[id] = kthread_run(smoke_crescendo_thread, arg,
3583 				      "igt/smoke:%d", id);
3584 		if (IS_ERR(tsk[id])) {
3585 			err = PTR_ERR(tsk[id]);
3586 			break;
3587 		}
3588 		get_task_struct(tsk[id]);
3589 	}
3590 
3591 	yield(); /* start all threads before we kthread_stop() */
3592 
3593 	count = 0;
3594 	for_each_engine(engine, smoke->gt, id) {
3595 		int status;
3596 
3597 		if (IS_ERR_OR_NULL(tsk[id]))
3598 			continue;
3599 
3600 		status = kthread_stop(tsk[id]);
3601 		if (status && !err)
3602 			err = status;
3603 
3604 		count += arg[id].count;
3605 
3606 		put_task_struct(tsk[id]);
3607 	}
3608 
3609 	pr_info("Submitted %lu crescendo:%x requests across %d engines and %d contexts\n",
3610 		count, flags, smoke->gt->info.num_engines, smoke->ncontext);
3611 
3612 	kfree(arg);
3613 	return 0;
3614 }
3615 
3616 static int smoke_random(struct preempt_smoke *smoke, unsigned int flags)
3617 {
3618 	enum intel_engine_id id;
3619 	IGT_TIMEOUT(end_time);
3620 	unsigned long count;
3621 
3622 	count = 0;
3623 	do {
3624 		for_each_engine(smoke->engine, smoke->gt, id) {
3625 			struct i915_gem_context *ctx = smoke_context(smoke);
3626 			int err;
3627 
3628 			err = smoke_submit(smoke,
3629 					   ctx, random_priority(&smoke->prng),
3630 					   flags & BATCH ? smoke->batch : NULL);
3631 			if (err)
3632 				return err;
3633 
3634 			count++;
3635 		}
3636 	} while (count < smoke->ncontext && !__igt_timeout(end_time, NULL));
3637 
3638 	pr_info("Submitted %lu random:%x requests across %d engines and %d contexts\n",
3639 		count, flags, smoke->gt->info.num_engines, smoke->ncontext);
3640 	return 0;
3641 }
3642 
3643 static int live_preempt_smoke(void *arg)
3644 {
3645 	struct preempt_smoke smoke = {
3646 		.gt = arg,
3647 		.prng = I915_RND_STATE_INITIALIZER(i915_selftest.random_seed),
3648 		.ncontext = 256,
3649 	};
3650 	const unsigned int phase[] = { 0, BATCH };
3651 	struct igt_live_test t;
3652 	int err = -ENOMEM;
3653 	u32 *cs;
3654 	int n;
3655 
3656 	smoke.contexts = kmalloc_array(smoke.ncontext,
3657 				       sizeof(*smoke.contexts),
3658 				       GFP_KERNEL);
3659 	if (!smoke.contexts)
3660 		return -ENOMEM;
3661 
3662 	smoke.batch =
3663 		i915_gem_object_create_internal(smoke.gt->i915, PAGE_SIZE);
3664 	if (IS_ERR(smoke.batch)) {
3665 		err = PTR_ERR(smoke.batch);
3666 		goto err_free;
3667 	}
3668 
3669 	cs = i915_gem_object_pin_map_unlocked(smoke.batch, I915_MAP_WB);
3670 	if (IS_ERR(cs)) {
3671 		err = PTR_ERR(cs);
3672 		goto err_batch;
3673 	}
3674 	for (n = 0; n < PAGE_SIZE / sizeof(*cs) - 1; n++)
3675 		cs[n] = MI_ARB_CHECK;
3676 	cs[n] = MI_BATCH_BUFFER_END;
3677 	i915_gem_object_flush_map(smoke.batch);
3678 	i915_gem_object_unpin_map(smoke.batch);
3679 
3680 	if (igt_live_test_begin(&t, smoke.gt->i915, __func__, "all")) {
3681 		err = -EIO;
3682 		goto err_batch;
3683 	}
3684 
3685 	for (n = 0; n < smoke.ncontext; n++) {
3686 		smoke.contexts[n] = kernel_context(smoke.gt->i915, NULL);
3687 		if (!smoke.contexts[n])
3688 			goto err_ctx;
3689 	}
3690 
3691 	for (n = 0; n < ARRAY_SIZE(phase); n++) {
3692 		err = smoke_crescendo(&smoke, phase[n]);
3693 		if (err)
3694 			goto err_ctx;
3695 
3696 		err = smoke_random(&smoke, phase[n]);
3697 		if (err)
3698 			goto err_ctx;
3699 	}
3700 
3701 err_ctx:
3702 	if (igt_live_test_end(&t))
3703 		err = -EIO;
3704 
3705 	for (n = 0; n < smoke.ncontext; n++) {
3706 		if (!smoke.contexts[n])
3707 			break;
3708 		kernel_context_close(smoke.contexts[n]);
3709 	}
3710 
3711 err_batch:
3712 	i915_gem_object_put(smoke.batch);
3713 err_free:
3714 	kfree(smoke.contexts);
3715 
3716 	return err;
3717 }
3718 
3719 static int nop_virtual_engine(struct intel_gt *gt,
3720 			      struct intel_engine_cs **siblings,
3721 			      unsigned int nsibling,
3722 			      unsigned int nctx,
3723 			      unsigned int flags)
3724 #define CHAIN BIT(0)
3725 {
3726 	IGT_TIMEOUT(end_time);
3727 	struct i915_request *request[16] = {};
3728 	struct intel_context *ve[16];
3729 	unsigned long n, prime, nc;
3730 	struct igt_live_test t;
3731 	ktime_t times[2] = {};
3732 	int err;
3733 
3734 	GEM_BUG_ON(!nctx || nctx > ARRAY_SIZE(ve));
3735 
3736 	for (n = 0; n < nctx; n++) {
3737 		ve[n] = intel_engine_create_virtual(siblings, nsibling, 0);
3738 		if (IS_ERR(ve[n])) {
3739 			err = PTR_ERR(ve[n]);
3740 			nctx = n;
3741 			goto out;
3742 		}
3743 
3744 		err = intel_context_pin(ve[n]);
3745 		if (err) {
3746 			intel_context_put(ve[n]);
3747 			nctx = n;
3748 			goto out;
3749 		}
3750 	}
3751 
3752 	err = igt_live_test_begin(&t, gt->i915, __func__, ve[0]->engine->name);
3753 	if (err)
3754 		goto out;
3755 
3756 	for_each_prime_number_from(prime, 1, 8192) {
3757 		times[1] = ktime_get_raw();
3758 
3759 		if (flags & CHAIN) {
3760 			for (nc = 0; nc < nctx; nc++) {
3761 				for (n = 0; n < prime; n++) {
3762 					struct i915_request *rq;
3763 
3764 					rq = i915_request_create(ve[nc]);
3765 					if (IS_ERR(rq)) {
3766 						err = PTR_ERR(rq);
3767 						goto out;
3768 					}
3769 
3770 					if (request[nc])
3771 						i915_request_put(request[nc]);
3772 					request[nc] = i915_request_get(rq);
3773 					i915_request_add(rq);
3774 				}
3775 			}
3776 		} else {
3777 			for (n = 0; n < prime; n++) {
3778 				for (nc = 0; nc < nctx; nc++) {
3779 					struct i915_request *rq;
3780 
3781 					rq = i915_request_create(ve[nc]);
3782 					if (IS_ERR(rq)) {
3783 						err = PTR_ERR(rq);
3784 						goto out;
3785 					}
3786 
3787 					if (request[nc])
3788 						i915_request_put(request[nc]);
3789 					request[nc] = i915_request_get(rq);
3790 					i915_request_add(rq);
3791 				}
3792 			}
3793 		}
3794 
3795 		for (nc = 0; nc < nctx; nc++) {
3796 			if (i915_request_wait(request[nc], 0, HZ / 10) < 0) {
3797 				pr_err("%s(%s): wait for %llx:%lld timed out\n",
3798 				       __func__, ve[0]->engine->name,
3799 				       request[nc]->fence.context,
3800 				       request[nc]->fence.seqno);
3801 
3802 				GEM_TRACE("%s(%s) failed at request %llx:%lld\n",
3803 					  __func__, ve[0]->engine->name,
3804 					  request[nc]->fence.context,
3805 					  request[nc]->fence.seqno);
3806 				GEM_TRACE_DUMP();
3807 				intel_gt_set_wedged(gt);
3808 				break;
3809 			}
3810 		}
3811 
3812 		times[1] = ktime_sub(ktime_get_raw(), times[1]);
3813 		if (prime == 1)
3814 			times[0] = times[1];
3815 
3816 		for (nc = 0; nc < nctx; nc++) {
3817 			i915_request_put(request[nc]);
3818 			request[nc] = NULL;
3819 		}
3820 
3821 		if (__igt_timeout(end_time, NULL))
3822 			break;
3823 	}
3824 
3825 	err = igt_live_test_end(&t);
3826 	if (err)
3827 		goto out;
3828 
3829 	pr_info("Requestx%d latencies on %s: 1 = %lluns, %lu = %lluns\n",
3830 		nctx, ve[0]->engine->name, ktime_to_ns(times[0]),
3831 		prime, div64_u64(ktime_to_ns(times[1]), prime));
3832 
3833 out:
3834 	if (igt_flush_test(gt->i915))
3835 		err = -EIO;
3836 
3837 	for (nc = 0; nc < nctx; nc++) {
3838 		i915_request_put(request[nc]);
3839 		intel_context_unpin(ve[nc]);
3840 		intel_context_put(ve[nc]);
3841 	}
3842 	return err;
3843 }
3844 
3845 static unsigned int
3846 __select_siblings(struct intel_gt *gt,
3847 		  unsigned int class,
3848 		  struct intel_engine_cs **siblings,
3849 		  bool (*filter)(const struct intel_engine_cs *))
3850 {
3851 	unsigned int n = 0;
3852 	unsigned int inst;
3853 
3854 	for (inst = 0; inst <= MAX_ENGINE_INSTANCE; inst++) {
3855 		if (!gt->engine_class[class][inst])
3856 			continue;
3857 
3858 		if (filter && !filter(gt->engine_class[class][inst]))
3859 			continue;
3860 
3861 		siblings[n++] = gt->engine_class[class][inst];
3862 	}
3863 
3864 	return n;
3865 }
3866 
3867 static unsigned int
3868 select_siblings(struct intel_gt *gt,
3869 		unsigned int class,
3870 		struct intel_engine_cs **siblings)
3871 {
3872 	return __select_siblings(gt, class, siblings, NULL);
3873 }
3874 
3875 static int live_virtual_engine(void *arg)
3876 {
3877 	struct intel_gt *gt = arg;
3878 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
3879 	struct intel_engine_cs *engine;
3880 	enum intel_engine_id id;
3881 	unsigned int class;
3882 	int err;
3883 
3884 	if (intel_uc_uses_guc_submission(&gt->uc))
3885 		return 0;
3886 
3887 	for_each_engine(engine, gt, id) {
3888 		err = nop_virtual_engine(gt, &engine, 1, 1, 0);
3889 		if (err) {
3890 			pr_err("Failed to wrap engine %s: err=%d\n",
3891 			       engine->name, err);
3892 			return err;
3893 		}
3894 	}
3895 
3896 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
3897 		int nsibling, n;
3898 
3899 		nsibling = select_siblings(gt, class, siblings);
3900 		if (nsibling < 2)
3901 			continue;
3902 
3903 		for (n = 1; n <= nsibling + 1; n++) {
3904 			err = nop_virtual_engine(gt, siblings, nsibling,
3905 						 n, 0);
3906 			if (err)
3907 				return err;
3908 		}
3909 
3910 		err = nop_virtual_engine(gt, siblings, nsibling, n, CHAIN);
3911 		if (err)
3912 			return err;
3913 	}
3914 
3915 	return 0;
3916 }
3917 
3918 static int mask_virtual_engine(struct intel_gt *gt,
3919 			       struct intel_engine_cs **siblings,
3920 			       unsigned int nsibling)
3921 {
3922 	struct i915_request *request[MAX_ENGINE_INSTANCE + 1];
3923 	struct intel_context *ve;
3924 	struct igt_live_test t;
3925 	unsigned int n;
3926 	int err;
3927 
3928 	/*
3929 	 * Check that by setting the execution mask on a request, we can
3930 	 * restrict it to our desired engine within the virtual engine.
3931 	 */
3932 
3933 	ve = intel_engine_create_virtual(siblings, nsibling, 0);
3934 	if (IS_ERR(ve)) {
3935 		err = PTR_ERR(ve);
3936 		goto out_close;
3937 	}
3938 
3939 	err = intel_context_pin(ve);
3940 	if (err)
3941 		goto out_put;
3942 
3943 	err = igt_live_test_begin(&t, gt->i915, __func__, ve->engine->name);
3944 	if (err)
3945 		goto out_unpin;
3946 
3947 	for (n = 0; n < nsibling; n++) {
3948 		request[n] = i915_request_create(ve);
3949 		if (IS_ERR(request[n])) {
3950 			err = PTR_ERR(request[n]);
3951 			nsibling = n;
3952 			goto out;
3953 		}
3954 
3955 		/* Reverse order as it's more likely to be unnatural */
3956 		request[n]->execution_mask = siblings[nsibling - n - 1]->mask;
3957 
3958 		i915_request_get(request[n]);
3959 		i915_request_add(request[n]);
3960 	}
3961 
3962 	for (n = 0; n < nsibling; n++) {
3963 		if (i915_request_wait(request[n], 0, HZ / 10) < 0) {
3964 			pr_err("%s(%s): wait for %llx:%lld timed out\n",
3965 			       __func__, ve->engine->name,
3966 			       request[n]->fence.context,
3967 			       request[n]->fence.seqno);
3968 
3969 			GEM_TRACE("%s(%s) failed at request %llx:%lld\n",
3970 				  __func__, ve->engine->name,
3971 				  request[n]->fence.context,
3972 				  request[n]->fence.seqno);
3973 			GEM_TRACE_DUMP();
3974 			intel_gt_set_wedged(gt);
3975 			err = -EIO;
3976 			goto out;
3977 		}
3978 
3979 		if (request[n]->engine != siblings[nsibling - n - 1]) {
3980 			pr_err("Executed on wrong sibling '%s', expected '%s'\n",
3981 			       request[n]->engine->name,
3982 			       siblings[nsibling - n - 1]->name);
3983 			err = -EINVAL;
3984 			goto out;
3985 		}
3986 	}
3987 
3988 	err = igt_live_test_end(&t);
3989 out:
3990 	if (igt_flush_test(gt->i915))
3991 		err = -EIO;
3992 
3993 	for (n = 0; n < nsibling; n++)
3994 		i915_request_put(request[n]);
3995 
3996 out_unpin:
3997 	intel_context_unpin(ve);
3998 out_put:
3999 	intel_context_put(ve);
4000 out_close:
4001 	return err;
4002 }
4003 
4004 static int live_virtual_mask(void *arg)
4005 {
4006 	struct intel_gt *gt = arg;
4007 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4008 	unsigned int class;
4009 	int err;
4010 
4011 	if (intel_uc_uses_guc_submission(&gt->uc))
4012 		return 0;
4013 
4014 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4015 		unsigned int nsibling;
4016 
4017 		nsibling = select_siblings(gt, class, siblings);
4018 		if (nsibling < 2)
4019 			continue;
4020 
4021 		err = mask_virtual_engine(gt, siblings, nsibling);
4022 		if (err)
4023 			return err;
4024 	}
4025 
4026 	return 0;
4027 }
4028 
4029 static int slicein_virtual_engine(struct intel_gt *gt,
4030 				  struct intel_engine_cs **siblings,
4031 				  unsigned int nsibling)
4032 {
4033 	const long timeout = slice_timeout(siblings[0]);
4034 	struct intel_context *ce;
4035 	struct i915_request *rq;
4036 	struct igt_spinner spin;
4037 	unsigned int n;
4038 	int err = 0;
4039 
4040 	/*
4041 	 * Virtual requests must take part in timeslicing on the target engines.
4042 	 */
4043 
4044 	if (igt_spinner_init(&spin, gt))
4045 		return -ENOMEM;
4046 
4047 	for (n = 0; n < nsibling; n++) {
4048 		ce = intel_context_create(siblings[n]);
4049 		if (IS_ERR(ce)) {
4050 			err = PTR_ERR(ce);
4051 			goto out;
4052 		}
4053 
4054 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
4055 		intel_context_put(ce);
4056 		if (IS_ERR(rq)) {
4057 			err = PTR_ERR(rq);
4058 			goto out;
4059 		}
4060 
4061 		i915_request_add(rq);
4062 	}
4063 
4064 	ce = intel_engine_create_virtual(siblings, nsibling, 0);
4065 	if (IS_ERR(ce)) {
4066 		err = PTR_ERR(ce);
4067 		goto out;
4068 	}
4069 
4070 	rq = intel_context_create_request(ce);
4071 	intel_context_put(ce);
4072 	if (IS_ERR(rq)) {
4073 		err = PTR_ERR(rq);
4074 		goto out;
4075 	}
4076 
4077 	i915_request_get(rq);
4078 	i915_request_add(rq);
4079 	if (i915_request_wait(rq, 0, timeout) < 0) {
4080 		GEM_TRACE_ERR("%s(%s) failed to slice in virtual request\n",
4081 			      __func__, rq->engine->name);
4082 		GEM_TRACE_DUMP();
4083 		intel_gt_set_wedged(gt);
4084 		err = -EIO;
4085 	}
4086 	i915_request_put(rq);
4087 
4088 out:
4089 	igt_spinner_end(&spin);
4090 	if (igt_flush_test(gt->i915))
4091 		err = -EIO;
4092 	igt_spinner_fini(&spin);
4093 	return err;
4094 }
4095 
4096 static int sliceout_virtual_engine(struct intel_gt *gt,
4097 				   struct intel_engine_cs **siblings,
4098 				   unsigned int nsibling)
4099 {
4100 	const long timeout = slice_timeout(siblings[0]);
4101 	struct intel_context *ce;
4102 	struct i915_request *rq;
4103 	struct igt_spinner spin;
4104 	unsigned int n;
4105 	int err = 0;
4106 
4107 	/*
4108 	 * Virtual requests must allow others a fair timeslice.
4109 	 */
4110 
4111 	if (igt_spinner_init(&spin, gt))
4112 		return -ENOMEM;
4113 
4114 	/* XXX We do not handle oversubscription and fairness with normal rq */
4115 	for (n = 0; n < nsibling; n++) {
4116 		ce = intel_engine_create_virtual(siblings, nsibling, 0);
4117 		if (IS_ERR(ce)) {
4118 			err = PTR_ERR(ce);
4119 			goto out;
4120 		}
4121 
4122 		rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK);
4123 		intel_context_put(ce);
4124 		if (IS_ERR(rq)) {
4125 			err = PTR_ERR(rq);
4126 			goto out;
4127 		}
4128 
4129 		i915_request_add(rq);
4130 	}
4131 
4132 	for (n = 0; !err && n < nsibling; n++) {
4133 		ce = intel_context_create(siblings[n]);
4134 		if (IS_ERR(ce)) {
4135 			err = PTR_ERR(ce);
4136 			goto out;
4137 		}
4138 
4139 		rq = intel_context_create_request(ce);
4140 		intel_context_put(ce);
4141 		if (IS_ERR(rq)) {
4142 			err = PTR_ERR(rq);
4143 			goto out;
4144 		}
4145 
4146 		i915_request_get(rq);
4147 		i915_request_add(rq);
4148 		if (i915_request_wait(rq, 0, timeout) < 0) {
4149 			GEM_TRACE_ERR("%s(%s) failed to slice out virtual request\n",
4150 				      __func__, siblings[n]->name);
4151 			GEM_TRACE_DUMP();
4152 			intel_gt_set_wedged(gt);
4153 			err = -EIO;
4154 		}
4155 		i915_request_put(rq);
4156 	}
4157 
4158 out:
4159 	igt_spinner_end(&spin);
4160 	if (igt_flush_test(gt->i915))
4161 		err = -EIO;
4162 	igt_spinner_fini(&spin);
4163 	return err;
4164 }
4165 
4166 static int live_virtual_slice(void *arg)
4167 {
4168 	struct intel_gt *gt = arg;
4169 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4170 	unsigned int class;
4171 	int err;
4172 
4173 	if (intel_uc_uses_guc_submission(&gt->uc))
4174 		return 0;
4175 
4176 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4177 		unsigned int nsibling;
4178 
4179 		nsibling = __select_siblings(gt, class, siblings,
4180 					     intel_engine_has_timeslices);
4181 		if (nsibling < 2)
4182 			continue;
4183 
4184 		err = slicein_virtual_engine(gt, siblings, nsibling);
4185 		if (err)
4186 			return err;
4187 
4188 		err = sliceout_virtual_engine(gt, siblings, nsibling);
4189 		if (err)
4190 			return err;
4191 	}
4192 
4193 	return 0;
4194 }
4195 
4196 static int preserved_virtual_engine(struct intel_gt *gt,
4197 				    struct intel_engine_cs **siblings,
4198 				    unsigned int nsibling)
4199 {
4200 	struct i915_request *last = NULL;
4201 	struct intel_context *ve;
4202 	struct i915_vma *scratch;
4203 	struct igt_live_test t;
4204 	unsigned int n;
4205 	int err = 0;
4206 	u32 *cs;
4207 
4208 	scratch =
4209 		__vm_create_scratch_for_read_pinned(&siblings[0]->gt->ggtt->vm,
4210 						    PAGE_SIZE);
4211 	if (IS_ERR(scratch))
4212 		return PTR_ERR(scratch);
4213 
4214 	err = i915_vma_sync(scratch);
4215 	if (err)
4216 		goto out_scratch;
4217 
4218 	ve = intel_engine_create_virtual(siblings, nsibling, 0);
4219 	if (IS_ERR(ve)) {
4220 		err = PTR_ERR(ve);
4221 		goto out_scratch;
4222 	}
4223 
4224 	err = intel_context_pin(ve);
4225 	if (err)
4226 		goto out_put;
4227 
4228 	err = igt_live_test_begin(&t, gt->i915, __func__, ve->engine->name);
4229 	if (err)
4230 		goto out_unpin;
4231 
4232 	for (n = 0; n < NUM_GPR_DW; n++) {
4233 		struct intel_engine_cs *engine = siblings[n % nsibling];
4234 		struct i915_request *rq;
4235 
4236 		rq = i915_request_create(ve);
4237 		if (IS_ERR(rq)) {
4238 			err = PTR_ERR(rq);
4239 			goto out_end;
4240 		}
4241 
4242 		i915_request_put(last);
4243 		last = i915_request_get(rq);
4244 
4245 		cs = intel_ring_begin(rq, 8);
4246 		if (IS_ERR(cs)) {
4247 			i915_request_add(rq);
4248 			err = PTR_ERR(cs);
4249 			goto out_end;
4250 		}
4251 
4252 		*cs++ = MI_STORE_REGISTER_MEM_GEN8 | MI_USE_GGTT;
4253 		*cs++ = CS_GPR(engine, n);
4254 		*cs++ = i915_ggtt_offset(scratch) + n * sizeof(u32);
4255 		*cs++ = 0;
4256 
4257 		*cs++ = MI_LOAD_REGISTER_IMM(1);
4258 		*cs++ = CS_GPR(engine, (n + 1) % NUM_GPR_DW);
4259 		*cs++ = n + 1;
4260 
4261 		*cs++ = MI_NOOP;
4262 		intel_ring_advance(rq, cs);
4263 
4264 		/* Restrict this request to run on a particular engine */
4265 		rq->execution_mask = engine->mask;
4266 		i915_request_add(rq);
4267 	}
4268 
4269 	if (i915_request_wait(last, 0, HZ / 5) < 0) {
4270 		err = -ETIME;
4271 		goto out_end;
4272 	}
4273 
4274 	cs = i915_gem_object_pin_map_unlocked(scratch->obj, I915_MAP_WB);
4275 	if (IS_ERR(cs)) {
4276 		err = PTR_ERR(cs);
4277 		goto out_end;
4278 	}
4279 
4280 	for (n = 0; n < NUM_GPR_DW; n++) {
4281 		if (cs[n] != n) {
4282 			pr_err("Incorrect value[%d] found for GPR[%d]\n",
4283 			       cs[n], n);
4284 			err = -EINVAL;
4285 			break;
4286 		}
4287 	}
4288 
4289 	i915_gem_object_unpin_map(scratch->obj);
4290 
4291 out_end:
4292 	if (igt_live_test_end(&t))
4293 		err = -EIO;
4294 	i915_request_put(last);
4295 out_unpin:
4296 	intel_context_unpin(ve);
4297 out_put:
4298 	intel_context_put(ve);
4299 out_scratch:
4300 	i915_vma_unpin_and_release(&scratch, 0);
4301 	return err;
4302 }
4303 
4304 static int live_virtual_preserved(void *arg)
4305 {
4306 	struct intel_gt *gt = arg;
4307 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4308 	unsigned int class;
4309 
4310 	/*
4311 	 * Check that the context image retains non-privileged (user) registers
4312 	 * from one engine to the next. For this we check that the CS_GPR
4313 	 * are preserved.
4314 	 */
4315 
4316 	if (intel_uc_uses_guc_submission(&gt->uc))
4317 		return 0;
4318 
4319 	/* As we use CS_GPR we cannot run before they existed on all engines. */
4320 	if (GRAPHICS_VER(gt->i915) < 9)
4321 		return 0;
4322 
4323 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4324 		int nsibling, err;
4325 
4326 		nsibling = select_siblings(gt, class, siblings);
4327 		if (nsibling < 2)
4328 			continue;
4329 
4330 		err = preserved_virtual_engine(gt, siblings, nsibling);
4331 		if (err)
4332 			return err;
4333 	}
4334 
4335 	return 0;
4336 }
4337 
4338 static int reset_virtual_engine(struct intel_gt *gt,
4339 				struct intel_engine_cs **siblings,
4340 				unsigned int nsibling)
4341 {
4342 	struct intel_engine_cs *engine;
4343 	struct intel_context *ve;
4344 	struct igt_spinner spin;
4345 	struct i915_request *rq;
4346 	unsigned int n;
4347 	int err = 0;
4348 
4349 	/*
4350 	 * In order to support offline error capture for fast preempt reset,
4351 	 * we need to decouple the guilty request and ensure that it and its
4352 	 * descendents are not executed while the capture is in progress.
4353 	 */
4354 
4355 	if (igt_spinner_init(&spin, gt))
4356 		return -ENOMEM;
4357 
4358 	ve = intel_engine_create_virtual(siblings, nsibling, 0);
4359 	if (IS_ERR(ve)) {
4360 		err = PTR_ERR(ve);
4361 		goto out_spin;
4362 	}
4363 
4364 	for (n = 0; n < nsibling; n++)
4365 		st_engine_heartbeat_disable(siblings[n]);
4366 
4367 	rq = igt_spinner_create_request(&spin, ve, MI_ARB_CHECK);
4368 	if (IS_ERR(rq)) {
4369 		err = PTR_ERR(rq);
4370 		goto out_heartbeat;
4371 	}
4372 	i915_request_add(rq);
4373 
4374 	if (!igt_wait_for_spinner(&spin, rq)) {
4375 		intel_gt_set_wedged(gt);
4376 		err = -ETIME;
4377 		goto out_heartbeat;
4378 	}
4379 
4380 	engine = rq->engine;
4381 	GEM_BUG_ON(engine == ve->engine);
4382 
4383 	/* Take ownership of the reset and tasklet */
4384 	err = engine_lock_reset_tasklet(engine);
4385 	if (err)
4386 		goto out_heartbeat;
4387 
4388 	engine->sched_engine->tasklet.callback(&engine->sched_engine->tasklet);
4389 	GEM_BUG_ON(execlists_active(&engine->execlists) != rq);
4390 
4391 	/* Fake a preemption event; failed of course */
4392 	spin_lock_irq(&engine->sched_engine->lock);
4393 	__unwind_incomplete_requests(engine);
4394 	spin_unlock_irq(&engine->sched_engine->lock);
4395 	GEM_BUG_ON(rq->engine != engine);
4396 
4397 	/* Reset the engine while keeping our active request on hold */
4398 	execlists_hold(engine, rq);
4399 	GEM_BUG_ON(!i915_request_on_hold(rq));
4400 
4401 	__intel_engine_reset_bh(engine, NULL);
4402 	GEM_BUG_ON(rq->fence.error != -EIO);
4403 
4404 	/* Release our grasp on the engine, letting CS flow again */
4405 	engine_unlock_reset_tasklet(engine);
4406 
4407 	/* Check that we do not resubmit the held request */
4408 	i915_request_get(rq);
4409 	if (!i915_request_wait(rq, 0, HZ / 5)) {
4410 		pr_err("%s: on hold request completed!\n",
4411 		       engine->name);
4412 		intel_gt_set_wedged(gt);
4413 		err = -EIO;
4414 		goto out_rq;
4415 	}
4416 	GEM_BUG_ON(!i915_request_on_hold(rq));
4417 
4418 	/* But is resubmitted on release */
4419 	execlists_unhold(engine, rq);
4420 	if (i915_request_wait(rq, 0, HZ / 5) < 0) {
4421 		pr_err("%s: held request did not complete!\n",
4422 		       engine->name);
4423 		intel_gt_set_wedged(gt);
4424 		err = -ETIME;
4425 	}
4426 
4427 out_rq:
4428 	i915_request_put(rq);
4429 out_heartbeat:
4430 	for (n = 0; n < nsibling; n++)
4431 		st_engine_heartbeat_enable(siblings[n]);
4432 
4433 	intel_context_put(ve);
4434 out_spin:
4435 	igt_spinner_fini(&spin);
4436 	return err;
4437 }
4438 
4439 static int live_virtual_reset(void *arg)
4440 {
4441 	struct intel_gt *gt = arg;
4442 	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
4443 	unsigned int class;
4444 
4445 	/*
4446 	 * Check that we handle a reset event within a virtual engine.
4447 	 * Only the physical engine is reset, but we have to check the flow
4448 	 * of the virtual requests around the reset, and make sure it is not
4449 	 * forgotten.
4450 	 */
4451 
4452 	if (intel_uc_uses_guc_submission(&gt->uc))
4453 		return 0;
4454 
4455 	if (!intel_has_reset_engine(gt))
4456 		return 0;
4457 
4458 	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
4459 		int nsibling, err;
4460 
4461 		nsibling = select_siblings(gt, class, siblings);
4462 		if (nsibling < 2)
4463 			continue;
4464 
4465 		err = reset_virtual_engine(gt, siblings, nsibling);
4466 		if (err)
4467 			return err;
4468 	}
4469 
4470 	return 0;
4471 }
4472 
4473 int intel_execlists_live_selftests(struct drm_i915_private *i915)
4474 {
4475 	static const struct i915_subtest tests[] = {
4476 		SUBTEST(live_sanitycheck),
4477 		SUBTEST(live_unlite_switch),
4478 		SUBTEST(live_unlite_preempt),
4479 		SUBTEST(live_unlite_ring),
4480 		SUBTEST(live_pin_rewind),
4481 		SUBTEST(live_hold_reset),
4482 		SUBTEST(live_error_interrupt),
4483 		SUBTEST(live_timeslice_preempt),
4484 		SUBTEST(live_timeslice_rewind),
4485 		SUBTEST(live_timeslice_queue),
4486 		SUBTEST(live_timeslice_nopreempt),
4487 		SUBTEST(live_busywait_preempt),
4488 		SUBTEST(live_preempt),
4489 		SUBTEST(live_late_preempt),
4490 		SUBTEST(live_nopreempt),
4491 		SUBTEST(live_preempt_cancel),
4492 		SUBTEST(live_suppress_self_preempt),
4493 		SUBTEST(live_chain_preempt),
4494 		SUBTEST(live_preempt_ring),
4495 		SUBTEST(live_preempt_gang),
4496 		SUBTEST(live_preempt_timeout),
4497 		SUBTEST(live_preempt_user),
4498 		SUBTEST(live_preempt_smoke),
4499 		SUBTEST(live_virtual_engine),
4500 		SUBTEST(live_virtual_mask),
4501 		SUBTEST(live_virtual_preserved),
4502 		SUBTEST(live_virtual_slice),
4503 		SUBTEST(live_virtual_reset),
4504 	};
4505 
4506 	if (to_gt(i915)->submission_method != INTEL_SUBMISSION_ELSP)
4507 		return 0;
4508 
4509 	if (intel_gt_is_wedged(to_gt(i915)))
4510 		return 0;
4511 
4512 	return intel_gt_live_subtests(tests, to_gt(i915));
4513 }
4514