1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2018 Intel Corporation
4  */
5 
6 #include "gem/i915_gem_pm.h"
7 #include "gt/intel_engine_user.h"
8 #include "gt/intel_gt.h"
9 #include "i915_selftest.h"
10 #include "intel_reset.h"
11 
12 #include "selftests/igt_flush_test.h"
13 #include "selftests/igt_reset.h"
14 #include "selftests/igt_spinner.h"
15 #include "selftests/intel_scheduler_helpers.h"
16 #include "selftests/mock_drm.h"
17 
18 #include "gem/selftests/igt_gem_utils.h"
19 #include "gem/selftests/mock_context.h"
20 
21 static const struct wo_register {
22 	enum intel_platform platform;
23 	u32 reg;
24 } wo_registers[] = {
25 	{ INTEL_GEMINILAKE, 0x731c }
26 };
27 
28 struct wa_lists {
29 	struct i915_wa_list gt_wa_list;
30 	struct {
31 		struct i915_wa_list wa_list;
32 		struct i915_wa_list ctx_wa_list;
33 	} engine[I915_NUM_ENGINES];
34 };
35 
36 static int request_add_sync(struct i915_request *rq, int err)
37 {
38 	i915_request_get(rq);
39 	i915_request_add(rq);
40 	if (i915_request_wait(rq, 0, HZ / 5) < 0)
41 		err = -EIO;
42 	i915_request_put(rq);
43 
44 	return err;
45 }
46 
47 static int request_add_spin(struct i915_request *rq, struct igt_spinner *spin)
48 {
49 	int err = 0;
50 
51 	i915_request_get(rq);
52 	i915_request_add(rq);
53 	if (spin && !igt_wait_for_spinner(spin, rq))
54 		err = -ETIMEDOUT;
55 	i915_request_put(rq);
56 
57 	return err;
58 }
59 
60 static void
61 reference_lists_init(struct intel_gt *gt, struct wa_lists *lists)
62 {
63 	struct intel_engine_cs *engine;
64 	enum intel_engine_id id;
65 
66 	memset(lists, 0, sizeof(*lists));
67 
68 	wa_init_start(&lists->gt_wa_list, "GT_REF", "global");
69 	gt_init_workarounds(gt, &lists->gt_wa_list);
70 	wa_init_finish(&lists->gt_wa_list);
71 
72 	for_each_engine(engine, gt, id) {
73 		struct i915_wa_list *wal = &lists->engine[id].wa_list;
74 
75 		wa_init_start(wal, "REF", engine->name);
76 		engine_init_workarounds(engine, wal);
77 		wa_init_finish(wal);
78 
79 		__intel_engine_init_ctx_wa(engine,
80 					   &lists->engine[id].ctx_wa_list,
81 					   "CTX_REF");
82 	}
83 }
84 
85 static void
86 reference_lists_fini(struct intel_gt *gt, struct wa_lists *lists)
87 {
88 	struct intel_engine_cs *engine;
89 	enum intel_engine_id id;
90 
91 	for_each_engine(engine, gt, id)
92 		intel_wa_list_free(&lists->engine[id].wa_list);
93 
94 	intel_wa_list_free(&lists->gt_wa_list);
95 }
96 
97 static struct drm_i915_gem_object *
98 read_nonprivs(struct intel_context *ce)
99 {
100 	struct intel_engine_cs *engine = ce->engine;
101 	const u32 base = engine->mmio_base;
102 	struct drm_i915_gem_object *result;
103 	struct i915_request *rq;
104 	struct i915_vma *vma;
105 	u32 srm, *cs;
106 	int err;
107 	int i;
108 
109 	result = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
110 	if (IS_ERR(result))
111 		return result;
112 
113 	i915_gem_object_set_cache_coherency(result, I915_CACHE_LLC);
114 
115 	cs = i915_gem_object_pin_map_unlocked(result, I915_MAP_WB);
116 	if (IS_ERR(cs)) {
117 		err = PTR_ERR(cs);
118 		goto err_obj;
119 	}
120 	memset(cs, 0xc5, PAGE_SIZE);
121 	i915_gem_object_flush_map(result);
122 	i915_gem_object_unpin_map(result);
123 
124 	vma = i915_vma_instance(result, &engine->gt->ggtt->vm, NULL);
125 	if (IS_ERR(vma)) {
126 		err = PTR_ERR(vma);
127 		goto err_obj;
128 	}
129 
130 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
131 	if (err)
132 		goto err_obj;
133 
134 	rq = intel_context_create_request(ce);
135 	if (IS_ERR(rq)) {
136 		err = PTR_ERR(rq);
137 		goto err_pin;
138 	}
139 
140 	i915_vma_lock(vma);
141 	err = i915_request_await_object(rq, vma->obj, true);
142 	if (err == 0)
143 		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
144 	i915_vma_unlock(vma);
145 	if (err)
146 		goto err_req;
147 
148 	srm = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
149 	if (GRAPHICS_VER(engine->i915) >= 8)
150 		srm++;
151 
152 	cs = intel_ring_begin(rq, 4 * RING_MAX_NONPRIV_SLOTS);
153 	if (IS_ERR(cs)) {
154 		err = PTR_ERR(cs);
155 		goto err_req;
156 	}
157 
158 	for (i = 0; i < RING_MAX_NONPRIV_SLOTS; i++) {
159 		*cs++ = srm;
160 		*cs++ = i915_mmio_reg_offset(RING_FORCE_TO_NONPRIV(base, i));
161 		*cs++ = i915_ggtt_offset(vma) + sizeof(u32) * i;
162 		*cs++ = 0;
163 	}
164 	intel_ring_advance(rq, cs);
165 
166 	i915_request_add(rq);
167 	i915_vma_unpin(vma);
168 
169 	return result;
170 
171 err_req:
172 	i915_request_add(rq);
173 err_pin:
174 	i915_vma_unpin(vma);
175 err_obj:
176 	i915_gem_object_put(result);
177 	return ERR_PTR(err);
178 }
179 
180 static u32
181 get_whitelist_reg(const struct intel_engine_cs *engine, unsigned int i)
182 {
183 	i915_reg_t reg = i < engine->whitelist.count ?
184 			 engine->whitelist.list[i].reg :
185 			 RING_NOPID(engine->mmio_base);
186 
187 	return i915_mmio_reg_offset(reg);
188 }
189 
190 static void
191 print_results(const struct intel_engine_cs *engine, const u32 *results)
192 {
193 	unsigned int i;
194 
195 	for (i = 0; i < RING_MAX_NONPRIV_SLOTS; i++) {
196 		u32 expected = get_whitelist_reg(engine, i);
197 		u32 actual = results[i];
198 
199 		pr_info("RING_NONPRIV[%d]: expected 0x%08x, found 0x%08x\n",
200 			i, expected, actual);
201 	}
202 }
203 
204 static int check_whitelist(struct intel_context *ce)
205 {
206 	struct intel_engine_cs *engine = ce->engine;
207 	struct drm_i915_gem_object *results;
208 	struct intel_wedge_me wedge;
209 	u32 *vaddr;
210 	int err;
211 	int i;
212 
213 	results = read_nonprivs(ce);
214 	if (IS_ERR(results))
215 		return PTR_ERR(results);
216 
217 	err = 0;
218 	i915_gem_object_lock(results, NULL);
219 	intel_wedge_on_timeout(&wedge, engine->gt, HZ / 5) /* safety net! */
220 		err = i915_gem_object_set_to_cpu_domain(results, false);
221 
222 	if (intel_gt_is_wedged(engine->gt))
223 		err = -EIO;
224 	if (err)
225 		goto out_put;
226 
227 	vaddr = i915_gem_object_pin_map(results, I915_MAP_WB);
228 	if (IS_ERR(vaddr)) {
229 		err = PTR_ERR(vaddr);
230 		goto out_put;
231 	}
232 
233 	for (i = 0; i < RING_MAX_NONPRIV_SLOTS; i++) {
234 		u32 expected = get_whitelist_reg(engine, i);
235 		u32 actual = vaddr[i];
236 
237 		if (expected != actual) {
238 			print_results(engine, vaddr);
239 			pr_err("Invalid RING_NONPRIV[%d], expected 0x%08x, found 0x%08x\n",
240 			       i, expected, actual);
241 
242 			err = -EINVAL;
243 			break;
244 		}
245 	}
246 
247 	i915_gem_object_unpin_map(results);
248 out_put:
249 	i915_gem_object_unlock(results);
250 	i915_gem_object_put(results);
251 	return err;
252 }
253 
254 static int do_device_reset(struct intel_engine_cs *engine)
255 {
256 	intel_gt_reset(engine->gt, engine->mask, "live_workarounds");
257 	return 0;
258 }
259 
260 static int do_engine_reset(struct intel_engine_cs *engine)
261 {
262 	return intel_engine_reset(engine, "live_workarounds");
263 }
264 
265 static int do_guc_reset(struct intel_engine_cs *engine)
266 {
267 	/* Currently a no-op as the reset is handled by GuC */
268 	return 0;
269 }
270 
271 static int
272 switch_to_scratch_context(struct intel_engine_cs *engine,
273 			  struct igt_spinner *spin,
274 			  struct i915_request **rq)
275 {
276 	struct intel_context *ce;
277 	int err = 0;
278 
279 	ce = intel_context_create(engine);
280 	if (IS_ERR(ce))
281 		return PTR_ERR(ce);
282 
283 	*rq = igt_spinner_create_request(spin, ce, MI_NOOP);
284 	intel_context_put(ce);
285 
286 	if (IS_ERR(*rq)) {
287 		spin = NULL;
288 		err = PTR_ERR(*rq);
289 		goto err;
290 	}
291 
292 	err = request_add_spin(*rq, spin);
293 err:
294 	if (err && spin)
295 		igt_spinner_end(spin);
296 
297 	return err;
298 }
299 
300 static int check_whitelist_across_reset(struct intel_engine_cs *engine,
301 					int (*reset)(struct intel_engine_cs *),
302 					const char *name)
303 {
304 	struct intel_context *ce, *tmp;
305 	struct igt_spinner spin;
306 	struct i915_request *rq;
307 	intel_wakeref_t wakeref;
308 	int err;
309 
310 	pr_info("Checking %d whitelisted registers on %s (RING_NONPRIV) [%s]\n",
311 		engine->whitelist.count, engine->name, name);
312 
313 	ce = intel_context_create(engine);
314 	if (IS_ERR(ce))
315 		return PTR_ERR(ce);
316 
317 	err = igt_spinner_init(&spin, engine->gt);
318 	if (err)
319 		goto out_ctx;
320 
321 	err = check_whitelist(ce);
322 	if (err) {
323 		pr_err("Invalid whitelist *before* %s reset!\n", name);
324 		goto out_spin;
325 	}
326 
327 	err = switch_to_scratch_context(engine, &spin, &rq);
328 	if (err)
329 		goto out_spin;
330 
331 	/* Ensure the spinner hasn't aborted */
332 	if (i915_request_completed(rq)) {
333 		pr_err("%s spinner failed to start\n", name);
334 		err = -ETIMEDOUT;
335 		goto out_spin;
336 	}
337 
338 	with_intel_runtime_pm(engine->uncore->rpm, wakeref)
339 		err = reset(engine);
340 
341 	/* Ensure the reset happens and kills the engine */
342 	if (err == 0)
343 		err = intel_selftest_wait_for_rq(rq);
344 
345 	igt_spinner_end(&spin);
346 
347 	if (err) {
348 		pr_err("%s reset failed\n", name);
349 		goto out_spin;
350 	}
351 
352 	err = check_whitelist(ce);
353 	if (err) {
354 		pr_err("Whitelist not preserved in context across %s reset!\n",
355 		       name);
356 		goto out_spin;
357 	}
358 
359 	tmp = intel_context_create(engine);
360 	if (IS_ERR(tmp)) {
361 		err = PTR_ERR(tmp);
362 		goto out_spin;
363 	}
364 	intel_context_put(ce);
365 	ce = tmp;
366 
367 	err = check_whitelist(ce);
368 	if (err) {
369 		pr_err("Invalid whitelist *after* %s reset in fresh context!\n",
370 		       name);
371 		goto out_spin;
372 	}
373 
374 out_spin:
375 	igt_spinner_fini(&spin);
376 out_ctx:
377 	intel_context_put(ce);
378 	return err;
379 }
380 
381 static struct i915_vma *create_batch(struct i915_address_space *vm)
382 {
383 	struct drm_i915_gem_object *obj;
384 	struct i915_vma *vma;
385 	int err;
386 
387 	obj = i915_gem_object_create_internal(vm->i915, 16 * PAGE_SIZE);
388 	if (IS_ERR(obj))
389 		return ERR_CAST(obj);
390 
391 	vma = i915_vma_instance(obj, vm, NULL);
392 	if (IS_ERR(vma)) {
393 		err = PTR_ERR(vma);
394 		goto err_obj;
395 	}
396 
397 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
398 	if (err)
399 		goto err_obj;
400 
401 	return vma;
402 
403 err_obj:
404 	i915_gem_object_put(obj);
405 	return ERR_PTR(err);
406 }
407 
408 static u32 reg_write(u32 old, u32 new, u32 rsvd)
409 {
410 	if (rsvd == 0x0000ffff) {
411 		old &= ~(new >> 16);
412 		old |= new & (new >> 16);
413 	} else {
414 		old &= ~rsvd;
415 		old |= new & rsvd;
416 	}
417 
418 	return old;
419 }
420 
421 static bool wo_register(struct intel_engine_cs *engine, u32 reg)
422 {
423 	enum intel_platform platform = INTEL_INFO(engine->i915)->platform;
424 	int i;
425 
426 	if ((reg & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
427 	     RING_FORCE_TO_NONPRIV_ACCESS_WR)
428 		return true;
429 
430 	for (i = 0; i < ARRAY_SIZE(wo_registers); i++) {
431 		if (wo_registers[i].platform == platform &&
432 		    wo_registers[i].reg == reg)
433 			return true;
434 	}
435 
436 	return false;
437 }
438 
439 static bool timestamp(const struct intel_engine_cs *engine, u32 reg)
440 {
441 	reg = (reg - engine->mmio_base) & ~RING_FORCE_TO_NONPRIV_ACCESS_MASK;
442 	switch (reg) {
443 	case 0x358:
444 	case 0x35c:
445 	case 0x3a8:
446 		return true;
447 
448 	default:
449 		return false;
450 	}
451 }
452 
453 static bool ro_register(u32 reg)
454 {
455 	if ((reg & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
456 	     RING_FORCE_TO_NONPRIV_ACCESS_RD)
457 		return true;
458 
459 	return false;
460 }
461 
462 static int whitelist_writable_count(struct intel_engine_cs *engine)
463 {
464 	int count = engine->whitelist.count;
465 	int i;
466 
467 	for (i = 0; i < engine->whitelist.count; i++) {
468 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
469 
470 		if (ro_register(reg))
471 			count--;
472 	}
473 
474 	return count;
475 }
476 
477 static int check_dirty_whitelist(struct intel_context *ce)
478 {
479 	const u32 values[] = {
480 		0x00000000,
481 		0x01010101,
482 		0x10100101,
483 		0x03030303,
484 		0x30300303,
485 		0x05050505,
486 		0x50500505,
487 		0x0f0f0f0f,
488 		0xf00ff00f,
489 		0x10101010,
490 		0xf0f01010,
491 		0x30303030,
492 		0xa0a03030,
493 		0x50505050,
494 		0xc0c05050,
495 		0xf0f0f0f0,
496 		0x11111111,
497 		0x33333333,
498 		0x55555555,
499 		0x0000ffff,
500 		0x00ff00ff,
501 		0xff0000ff,
502 		0xffff00ff,
503 		0xffffffff,
504 	};
505 	struct intel_engine_cs *engine = ce->engine;
506 	struct i915_vma *scratch;
507 	struct i915_vma *batch;
508 	int err = 0, i, v, sz;
509 	u32 *cs, *results;
510 
511 	sz = (2 * ARRAY_SIZE(values) + 1) * sizeof(u32);
512 	scratch = __vm_create_scratch_for_read_pinned(ce->vm, sz);
513 	if (IS_ERR(scratch))
514 		return PTR_ERR(scratch);
515 
516 	batch = create_batch(ce->vm);
517 	if (IS_ERR(batch)) {
518 		err = PTR_ERR(batch);
519 		goto out_scratch;
520 	}
521 
522 	for (i = 0; i < engine->whitelist.count; i++) {
523 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
524 		struct i915_gem_ww_ctx ww;
525 		u64 addr = scratch->node.start;
526 		struct i915_request *rq;
527 		u32 srm, lrm, rsvd;
528 		u32 expect;
529 		int idx;
530 		bool ro_reg;
531 
532 		if (wo_register(engine, reg))
533 			continue;
534 
535 		if (timestamp(engine, reg))
536 			continue; /* timestamps are expected to autoincrement */
537 
538 		ro_reg = ro_register(reg);
539 
540 		i915_gem_ww_ctx_init(&ww, false);
541 retry:
542 		cs = NULL;
543 		err = i915_gem_object_lock(scratch->obj, &ww);
544 		if (!err)
545 			err = i915_gem_object_lock(batch->obj, &ww);
546 		if (!err)
547 			err = intel_context_pin_ww(ce, &ww);
548 		if (err)
549 			goto out;
550 
551 		cs = i915_gem_object_pin_map(batch->obj, I915_MAP_WC);
552 		if (IS_ERR(cs)) {
553 			err = PTR_ERR(cs);
554 			goto out_ctx;
555 		}
556 
557 		results = i915_gem_object_pin_map(scratch->obj, I915_MAP_WB);
558 		if (IS_ERR(results)) {
559 			err = PTR_ERR(results);
560 			goto out_unmap_batch;
561 		}
562 
563 		/* Clear non priv flags */
564 		reg &= RING_FORCE_TO_NONPRIV_ADDRESS_MASK;
565 
566 		srm = MI_STORE_REGISTER_MEM;
567 		lrm = MI_LOAD_REGISTER_MEM;
568 		if (GRAPHICS_VER(engine->i915) >= 8)
569 			lrm++, srm++;
570 
571 		pr_debug("%s: Writing garbage to %x\n",
572 			 engine->name, reg);
573 
574 		/* SRM original */
575 		*cs++ = srm;
576 		*cs++ = reg;
577 		*cs++ = lower_32_bits(addr);
578 		*cs++ = upper_32_bits(addr);
579 
580 		idx = 1;
581 		for (v = 0; v < ARRAY_SIZE(values); v++) {
582 			/* LRI garbage */
583 			*cs++ = MI_LOAD_REGISTER_IMM(1);
584 			*cs++ = reg;
585 			*cs++ = values[v];
586 
587 			/* SRM result */
588 			*cs++ = srm;
589 			*cs++ = reg;
590 			*cs++ = lower_32_bits(addr + sizeof(u32) * idx);
591 			*cs++ = upper_32_bits(addr + sizeof(u32) * idx);
592 			idx++;
593 		}
594 		for (v = 0; v < ARRAY_SIZE(values); v++) {
595 			/* LRI garbage */
596 			*cs++ = MI_LOAD_REGISTER_IMM(1);
597 			*cs++ = reg;
598 			*cs++ = ~values[v];
599 
600 			/* SRM result */
601 			*cs++ = srm;
602 			*cs++ = reg;
603 			*cs++ = lower_32_bits(addr + sizeof(u32) * idx);
604 			*cs++ = upper_32_bits(addr + sizeof(u32) * idx);
605 			idx++;
606 		}
607 		GEM_BUG_ON(idx * sizeof(u32) > scratch->size);
608 
609 		/* LRM original -- don't leave garbage in the context! */
610 		*cs++ = lrm;
611 		*cs++ = reg;
612 		*cs++ = lower_32_bits(addr);
613 		*cs++ = upper_32_bits(addr);
614 
615 		*cs++ = MI_BATCH_BUFFER_END;
616 
617 		i915_gem_object_flush_map(batch->obj);
618 		i915_gem_object_unpin_map(batch->obj);
619 		intel_gt_chipset_flush(engine->gt);
620 		cs = NULL;
621 
622 		rq = i915_request_create(ce);
623 		if (IS_ERR(rq)) {
624 			err = PTR_ERR(rq);
625 			goto out_unmap_scratch;
626 		}
627 
628 		if (engine->emit_init_breadcrumb) { /* Be nice if we hang */
629 			err = engine->emit_init_breadcrumb(rq);
630 			if (err)
631 				goto err_request;
632 		}
633 
634 		err = i915_request_await_object(rq, batch->obj, false);
635 		if (err == 0)
636 			err = i915_vma_move_to_active(batch, rq, 0);
637 		if (err)
638 			goto err_request;
639 
640 		err = i915_request_await_object(rq, scratch->obj, true);
641 		if (err == 0)
642 			err = i915_vma_move_to_active(scratch, rq,
643 						      EXEC_OBJECT_WRITE);
644 		if (err)
645 			goto err_request;
646 
647 		err = engine->emit_bb_start(rq,
648 					    batch->node.start, PAGE_SIZE,
649 					    0);
650 		if (err)
651 			goto err_request;
652 
653 err_request:
654 		err = request_add_sync(rq, err);
655 		if (err) {
656 			pr_err("%s: Futzing %x timedout; cancelling test\n",
657 			       engine->name, reg);
658 			intel_gt_set_wedged(engine->gt);
659 			goto out_unmap_scratch;
660 		}
661 
662 		GEM_BUG_ON(values[ARRAY_SIZE(values) - 1] != 0xffffffff);
663 		if (!ro_reg) {
664 			/* detect write masking */
665 			rsvd = results[ARRAY_SIZE(values)];
666 			if (!rsvd) {
667 				pr_err("%s: Unable to write to whitelisted register %x\n",
668 				       engine->name, reg);
669 				err = -EINVAL;
670 				goto out_unmap_scratch;
671 			}
672 		} else {
673 			rsvd = 0;
674 		}
675 
676 		expect = results[0];
677 		idx = 1;
678 		for (v = 0; v < ARRAY_SIZE(values); v++) {
679 			if (ro_reg)
680 				expect = results[0];
681 			else
682 				expect = reg_write(expect, values[v], rsvd);
683 
684 			if (results[idx] != expect)
685 				err++;
686 			idx++;
687 		}
688 		for (v = 0; v < ARRAY_SIZE(values); v++) {
689 			if (ro_reg)
690 				expect = results[0];
691 			else
692 				expect = reg_write(expect, ~values[v], rsvd);
693 
694 			if (results[idx] != expect)
695 				err++;
696 			idx++;
697 		}
698 		if (err) {
699 			pr_err("%s: %d mismatch between values written to whitelisted register [%x], and values read back!\n",
700 			       engine->name, err, reg);
701 
702 			if (ro_reg)
703 				pr_info("%s: Whitelisted read-only register: %x, original value %08x\n",
704 					engine->name, reg, results[0]);
705 			else
706 				pr_info("%s: Whitelisted register: %x, original value %08x, rsvd %08x\n",
707 					engine->name, reg, results[0], rsvd);
708 
709 			expect = results[0];
710 			idx = 1;
711 			for (v = 0; v < ARRAY_SIZE(values); v++) {
712 				u32 w = values[v];
713 
714 				if (ro_reg)
715 					expect = results[0];
716 				else
717 					expect = reg_write(expect, w, rsvd);
718 				pr_info("Wrote %08x, read %08x, expect %08x\n",
719 					w, results[idx], expect);
720 				idx++;
721 			}
722 			for (v = 0; v < ARRAY_SIZE(values); v++) {
723 				u32 w = ~values[v];
724 
725 				if (ro_reg)
726 					expect = results[0];
727 				else
728 					expect = reg_write(expect, w, rsvd);
729 				pr_info("Wrote %08x, read %08x, expect %08x\n",
730 					w, results[idx], expect);
731 				idx++;
732 			}
733 
734 			err = -EINVAL;
735 		}
736 out_unmap_scratch:
737 		i915_gem_object_unpin_map(scratch->obj);
738 out_unmap_batch:
739 		if (cs)
740 			i915_gem_object_unpin_map(batch->obj);
741 out_ctx:
742 		intel_context_unpin(ce);
743 out:
744 		if (err == -EDEADLK) {
745 			err = i915_gem_ww_ctx_backoff(&ww);
746 			if (!err)
747 				goto retry;
748 		}
749 		i915_gem_ww_ctx_fini(&ww);
750 		if (err)
751 			break;
752 	}
753 
754 	if (igt_flush_test(engine->i915))
755 		err = -EIO;
756 
757 	i915_vma_unpin_and_release(&batch, 0);
758 out_scratch:
759 	i915_vma_unpin_and_release(&scratch, 0);
760 	return err;
761 }
762 
763 static int live_dirty_whitelist(void *arg)
764 {
765 	struct intel_gt *gt = arg;
766 	struct intel_engine_cs *engine;
767 	enum intel_engine_id id;
768 
769 	/* Can the user write to the whitelisted registers? */
770 
771 	if (GRAPHICS_VER(gt->i915) < 7) /* minimum requirement for LRI, SRM, LRM */
772 		return 0;
773 
774 	for_each_engine(engine, gt, id) {
775 		struct intel_context *ce;
776 		int err;
777 
778 		if (engine->whitelist.count == 0)
779 			continue;
780 
781 		ce = intel_context_create(engine);
782 		if (IS_ERR(ce))
783 			return PTR_ERR(ce);
784 
785 		err = check_dirty_whitelist(ce);
786 		intel_context_put(ce);
787 		if (err)
788 			return err;
789 	}
790 
791 	return 0;
792 }
793 
794 static int live_reset_whitelist(void *arg)
795 {
796 	struct intel_gt *gt = arg;
797 	struct intel_engine_cs *engine;
798 	enum intel_engine_id id;
799 	int err = 0;
800 
801 	/* If we reset the gpu, we should not lose the RING_NONPRIV */
802 	igt_global_reset_lock(gt);
803 
804 	for_each_engine(engine, gt, id) {
805 		if (engine->whitelist.count == 0)
806 			continue;
807 
808 		if (intel_has_reset_engine(gt)) {
809 			if (intel_engine_uses_guc(engine)) {
810 				struct intel_selftest_saved_policy saved;
811 				int err2;
812 
813 				err = intel_selftest_modify_policy(engine, &saved,
814 								   SELFTEST_SCHEDULER_MODIFY_FAST_RESET);
815 				if (err)
816 					goto out;
817 
818 				err = check_whitelist_across_reset(engine,
819 								   do_guc_reset,
820 								   "guc");
821 
822 				err2 = intel_selftest_restore_policy(engine, &saved);
823 				if (err == 0)
824 					err = err2;
825 			} else {
826 				err = check_whitelist_across_reset(engine,
827 								   do_engine_reset,
828 								   "engine");
829 			}
830 
831 			if (err)
832 				goto out;
833 		}
834 
835 		if (intel_has_gpu_reset(gt)) {
836 			err = check_whitelist_across_reset(engine,
837 							   do_device_reset,
838 							   "device");
839 			if (err)
840 				goto out;
841 		}
842 	}
843 
844 out:
845 	igt_global_reset_unlock(gt);
846 	return err;
847 }
848 
849 static int read_whitelisted_registers(struct intel_context *ce,
850 				      struct i915_vma *results)
851 {
852 	struct intel_engine_cs *engine = ce->engine;
853 	struct i915_request *rq;
854 	int i, err = 0;
855 	u32 srm, *cs;
856 
857 	rq = intel_context_create_request(ce);
858 	if (IS_ERR(rq))
859 		return PTR_ERR(rq);
860 
861 	i915_vma_lock(results);
862 	err = i915_request_await_object(rq, results->obj, true);
863 	if (err == 0)
864 		err = i915_vma_move_to_active(results, rq, EXEC_OBJECT_WRITE);
865 	i915_vma_unlock(results);
866 	if (err)
867 		goto err_req;
868 
869 	srm = MI_STORE_REGISTER_MEM;
870 	if (GRAPHICS_VER(engine->i915) >= 8)
871 		srm++;
872 
873 	cs = intel_ring_begin(rq, 4 * engine->whitelist.count);
874 	if (IS_ERR(cs)) {
875 		err = PTR_ERR(cs);
876 		goto err_req;
877 	}
878 
879 	for (i = 0; i < engine->whitelist.count; i++) {
880 		u64 offset = results->node.start + sizeof(u32) * i;
881 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
882 
883 		/* Clear non priv flags */
884 		reg &= RING_FORCE_TO_NONPRIV_ADDRESS_MASK;
885 
886 		*cs++ = srm;
887 		*cs++ = reg;
888 		*cs++ = lower_32_bits(offset);
889 		*cs++ = upper_32_bits(offset);
890 	}
891 	intel_ring_advance(rq, cs);
892 
893 err_req:
894 	return request_add_sync(rq, err);
895 }
896 
897 static int scrub_whitelisted_registers(struct intel_context *ce)
898 {
899 	struct intel_engine_cs *engine = ce->engine;
900 	struct i915_request *rq;
901 	struct i915_vma *batch;
902 	int i, err = 0;
903 	u32 *cs;
904 
905 	batch = create_batch(ce->vm);
906 	if (IS_ERR(batch))
907 		return PTR_ERR(batch);
908 
909 	cs = i915_gem_object_pin_map_unlocked(batch->obj, I915_MAP_WC);
910 	if (IS_ERR(cs)) {
911 		err = PTR_ERR(cs);
912 		goto err_batch;
913 	}
914 
915 	*cs++ = MI_LOAD_REGISTER_IMM(whitelist_writable_count(engine));
916 	for (i = 0; i < engine->whitelist.count; i++) {
917 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
918 
919 		if (ro_register(reg))
920 			continue;
921 
922 		/* Clear non priv flags */
923 		reg &= RING_FORCE_TO_NONPRIV_ADDRESS_MASK;
924 
925 		*cs++ = reg;
926 		*cs++ = 0xffffffff;
927 	}
928 	*cs++ = MI_BATCH_BUFFER_END;
929 
930 	i915_gem_object_flush_map(batch->obj);
931 	intel_gt_chipset_flush(engine->gt);
932 
933 	rq = intel_context_create_request(ce);
934 	if (IS_ERR(rq)) {
935 		err = PTR_ERR(rq);
936 		goto err_unpin;
937 	}
938 
939 	if (engine->emit_init_breadcrumb) { /* Be nice if we hang */
940 		err = engine->emit_init_breadcrumb(rq);
941 		if (err)
942 			goto err_request;
943 	}
944 
945 	i915_vma_lock(batch);
946 	err = i915_request_await_object(rq, batch->obj, false);
947 	if (err == 0)
948 		err = i915_vma_move_to_active(batch, rq, 0);
949 	i915_vma_unlock(batch);
950 	if (err)
951 		goto err_request;
952 
953 	/* Perform the writes from an unprivileged "user" batch */
954 	err = engine->emit_bb_start(rq, batch->node.start, 0, 0);
955 
956 err_request:
957 	err = request_add_sync(rq, err);
958 
959 err_unpin:
960 	i915_gem_object_unpin_map(batch->obj);
961 err_batch:
962 	i915_vma_unpin_and_release(&batch, 0);
963 	return err;
964 }
965 
966 struct regmask {
967 	i915_reg_t reg;
968 	u8 graphics_ver;
969 };
970 
971 static bool find_reg(struct drm_i915_private *i915,
972 		     i915_reg_t reg,
973 		     const struct regmask *tbl,
974 		     unsigned long count)
975 {
976 	u32 offset = i915_mmio_reg_offset(reg);
977 
978 	while (count--) {
979 		if (GRAPHICS_VER(i915) == tbl->graphics_ver &&
980 		    i915_mmio_reg_offset(tbl->reg) == offset)
981 			return true;
982 		tbl++;
983 	}
984 
985 	return false;
986 }
987 
988 static bool pardon_reg(struct drm_i915_private *i915, i915_reg_t reg)
989 {
990 	/* Alas, we must pardon some whitelists. Mistakes already made */
991 	static const struct regmask pardon[] = {
992 		{ GEN9_CTX_PREEMPT_REG, 9 },
993 		{ GEN8_L3SQCREG4, 9 },
994 	};
995 
996 	return find_reg(i915, reg, pardon, ARRAY_SIZE(pardon));
997 }
998 
999 static bool result_eq(struct intel_engine_cs *engine,
1000 		      u32 a, u32 b, i915_reg_t reg)
1001 {
1002 	if (a != b && !pardon_reg(engine->i915, reg)) {
1003 		pr_err("Whitelisted register 0x%4x not context saved: A=%08x, B=%08x\n",
1004 		       i915_mmio_reg_offset(reg), a, b);
1005 		return false;
1006 	}
1007 
1008 	return true;
1009 }
1010 
1011 static bool writeonly_reg(struct drm_i915_private *i915, i915_reg_t reg)
1012 {
1013 	/* Some registers do not seem to behave and our writes unreadable */
1014 	static const struct regmask wo[] = {
1015 		{ GEN9_SLICE_COMMON_ECO_CHICKEN1, 9 },
1016 	};
1017 
1018 	return find_reg(i915, reg, wo, ARRAY_SIZE(wo));
1019 }
1020 
1021 static bool result_neq(struct intel_engine_cs *engine,
1022 		       u32 a, u32 b, i915_reg_t reg)
1023 {
1024 	if (a == b && !writeonly_reg(engine->i915, reg)) {
1025 		pr_err("Whitelist register 0x%4x:%08x was unwritable\n",
1026 		       i915_mmio_reg_offset(reg), a);
1027 		return false;
1028 	}
1029 
1030 	return true;
1031 }
1032 
1033 static int
1034 check_whitelisted_registers(struct intel_engine_cs *engine,
1035 			    struct i915_vma *A,
1036 			    struct i915_vma *B,
1037 			    bool (*fn)(struct intel_engine_cs *engine,
1038 				       u32 a, u32 b,
1039 				       i915_reg_t reg))
1040 {
1041 	u32 *a, *b;
1042 	int i, err;
1043 
1044 	a = i915_gem_object_pin_map_unlocked(A->obj, I915_MAP_WB);
1045 	if (IS_ERR(a))
1046 		return PTR_ERR(a);
1047 
1048 	b = i915_gem_object_pin_map_unlocked(B->obj, I915_MAP_WB);
1049 	if (IS_ERR(b)) {
1050 		err = PTR_ERR(b);
1051 		goto err_a;
1052 	}
1053 
1054 	err = 0;
1055 	for (i = 0; i < engine->whitelist.count; i++) {
1056 		const struct i915_wa *wa = &engine->whitelist.list[i];
1057 
1058 		if (i915_mmio_reg_offset(wa->reg) &
1059 		    RING_FORCE_TO_NONPRIV_ACCESS_RD)
1060 			continue;
1061 
1062 		if (!fn(engine, a[i], b[i], wa->reg))
1063 			err = -EINVAL;
1064 	}
1065 
1066 	i915_gem_object_unpin_map(B->obj);
1067 err_a:
1068 	i915_gem_object_unpin_map(A->obj);
1069 	return err;
1070 }
1071 
1072 static int live_isolated_whitelist(void *arg)
1073 {
1074 	struct intel_gt *gt = arg;
1075 	struct {
1076 		struct i915_vma *scratch[2];
1077 	} client[2] = {};
1078 	struct intel_engine_cs *engine;
1079 	enum intel_engine_id id;
1080 	int i, err = 0;
1081 
1082 	/*
1083 	 * Check that a write into a whitelist register works, but
1084 	 * invisible to a second context.
1085 	 */
1086 
1087 	if (!intel_engines_has_context_isolation(gt->i915))
1088 		return 0;
1089 
1090 	for (i = 0; i < ARRAY_SIZE(client); i++) {
1091 		client[i].scratch[0] =
1092 			__vm_create_scratch_for_read_pinned(gt->vm, 4096);
1093 		if (IS_ERR(client[i].scratch[0])) {
1094 			err = PTR_ERR(client[i].scratch[0]);
1095 			goto err;
1096 		}
1097 
1098 		client[i].scratch[1] =
1099 			__vm_create_scratch_for_read_pinned(gt->vm, 4096);
1100 		if (IS_ERR(client[i].scratch[1])) {
1101 			err = PTR_ERR(client[i].scratch[1]);
1102 			i915_vma_unpin_and_release(&client[i].scratch[0], 0);
1103 			goto err;
1104 		}
1105 	}
1106 
1107 	for_each_engine(engine, gt, id) {
1108 		struct intel_context *ce[2];
1109 
1110 		if (!engine->kernel_context->vm)
1111 			continue;
1112 
1113 		if (!whitelist_writable_count(engine))
1114 			continue;
1115 
1116 		ce[0] = intel_context_create(engine);
1117 		if (IS_ERR(ce[0])) {
1118 			err = PTR_ERR(ce[0]);
1119 			break;
1120 		}
1121 		ce[1] = intel_context_create(engine);
1122 		if (IS_ERR(ce[1])) {
1123 			err = PTR_ERR(ce[1]);
1124 			intel_context_put(ce[0]);
1125 			break;
1126 		}
1127 
1128 		/* Read default values */
1129 		err = read_whitelisted_registers(ce[0], client[0].scratch[0]);
1130 		if (err)
1131 			goto err_ce;
1132 
1133 		/* Try to overwrite registers (should only affect ctx0) */
1134 		err = scrub_whitelisted_registers(ce[0]);
1135 		if (err)
1136 			goto err_ce;
1137 
1138 		/* Read values from ctx1, we expect these to be defaults */
1139 		err = read_whitelisted_registers(ce[1], client[1].scratch[0]);
1140 		if (err)
1141 			goto err_ce;
1142 
1143 		/* Verify that both reads return the same default values */
1144 		err = check_whitelisted_registers(engine,
1145 						  client[0].scratch[0],
1146 						  client[1].scratch[0],
1147 						  result_eq);
1148 		if (err)
1149 			goto err_ce;
1150 
1151 		/* Read back the updated values in ctx0 */
1152 		err = read_whitelisted_registers(ce[0], client[0].scratch[1]);
1153 		if (err)
1154 			goto err_ce;
1155 
1156 		/* User should be granted privilege to overwhite regs */
1157 		err = check_whitelisted_registers(engine,
1158 						  client[0].scratch[0],
1159 						  client[0].scratch[1],
1160 						  result_neq);
1161 err_ce:
1162 		intel_context_put(ce[1]);
1163 		intel_context_put(ce[0]);
1164 		if (err)
1165 			break;
1166 	}
1167 
1168 err:
1169 	for (i = 0; i < ARRAY_SIZE(client); i++) {
1170 		i915_vma_unpin_and_release(&client[i].scratch[1], 0);
1171 		i915_vma_unpin_and_release(&client[i].scratch[0], 0);
1172 	}
1173 
1174 	if (igt_flush_test(gt->i915))
1175 		err = -EIO;
1176 
1177 	return err;
1178 }
1179 
1180 static bool
1181 verify_wa_lists(struct intel_gt *gt, struct wa_lists *lists,
1182 		const char *str)
1183 {
1184 	struct intel_engine_cs *engine;
1185 	enum intel_engine_id id;
1186 	bool ok = true;
1187 
1188 	ok &= wa_list_verify(gt, &lists->gt_wa_list, str);
1189 
1190 	for_each_engine(engine, gt, id) {
1191 		struct intel_context *ce;
1192 
1193 		ce = intel_context_create(engine);
1194 		if (IS_ERR(ce))
1195 			return false;
1196 
1197 		ok &= engine_wa_list_verify(ce,
1198 					    &lists->engine[id].wa_list,
1199 					    str) == 0;
1200 
1201 		ok &= engine_wa_list_verify(ce,
1202 					    &lists->engine[id].ctx_wa_list,
1203 					    str) == 0;
1204 
1205 		intel_context_put(ce);
1206 	}
1207 
1208 	return ok;
1209 }
1210 
1211 static int
1212 live_gpu_reset_workarounds(void *arg)
1213 {
1214 	struct intel_gt *gt = arg;
1215 	intel_wakeref_t wakeref;
1216 	struct wa_lists *lists;
1217 	bool ok;
1218 
1219 	if (!intel_has_gpu_reset(gt))
1220 		return 0;
1221 
1222 	lists = kzalloc(sizeof(*lists), GFP_KERNEL);
1223 	if (!lists)
1224 		return -ENOMEM;
1225 
1226 	pr_info("Verifying after GPU reset...\n");
1227 
1228 	igt_global_reset_lock(gt);
1229 	wakeref = intel_runtime_pm_get(gt->uncore->rpm);
1230 
1231 	reference_lists_init(gt, lists);
1232 
1233 	ok = verify_wa_lists(gt, lists, "before reset");
1234 	if (!ok)
1235 		goto out;
1236 
1237 	intel_gt_reset(gt, ALL_ENGINES, "live_workarounds");
1238 
1239 	ok = verify_wa_lists(gt, lists, "after reset");
1240 
1241 out:
1242 	reference_lists_fini(gt, lists);
1243 	intel_runtime_pm_put(gt->uncore->rpm, wakeref);
1244 	igt_global_reset_unlock(gt);
1245 	kfree(lists);
1246 
1247 	return ok ? 0 : -ESRCH;
1248 }
1249 
1250 static int
1251 live_engine_reset_workarounds(void *arg)
1252 {
1253 	struct intel_gt *gt = arg;
1254 	struct intel_engine_cs *engine;
1255 	enum intel_engine_id id;
1256 	struct intel_context *ce;
1257 	struct igt_spinner spin;
1258 	struct i915_request *rq;
1259 	intel_wakeref_t wakeref;
1260 	struct wa_lists *lists;
1261 	int ret = 0;
1262 
1263 	if (!intel_has_reset_engine(gt))
1264 		return 0;
1265 
1266 	lists = kzalloc(sizeof(*lists), GFP_KERNEL);
1267 	if (!lists)
1268 		return -ENOMEM;
1269 
1270 	igt_global_reset_lock(gt);
1271 	wakeref = intel_runtime_pm_get(gt->uncore->rpm);
1272 
1273 	reference_lists_init(gt, lists);
1274 
1275 	for_each_engine(engine, gt, id) {
1276 		struct intel_selftest_saved_policy saved;
1277 		bool using_guc = intel_engine_uses_guc(engine);
1278 		bool ok;
1279 		int ret2;
1280 
1281 		pr_info("Verifying after %s reset...\n", engine->name);
1282 		ret = intel_selftest_modify_policy(engine, &saved,
1283 						   SELFTEST_SCHEDULER_MODIFY_FAST_RESET);
1284 		if (ret)
1285 			break;
1286 
1287 		ce = intel_context_create(engine);
1288 		if (IS_ERR(ce)) {
1289 			ret = PTR_ERR(ce);
1290 			goto restore;
1291 		}
1292 
1293 		if (!using_guc) {
1294 			ok = verify_wa_lists(gt, lists, "before reset");
1295 			if (!ok) {
1296 				ret = -ESRCH;
1297 				goto err;
1298 			}
1299 
1300 			ret = intel_engine_reset(engine, "live_workarounds:idle");
1301 			if (ret) {
1302 				pr_err("%s: Reset failed while idle\n", engine->name);
1303 				goto err;
1304 			}
1305 
1306 			ok = verify_wa_lists(gt, lists, "after idle reset");
1307 			if (!ok) {
1308 				ret = -ESRCH;
1309 				goto err;
1310 			}
1311 		}
1312 
1313 		ret = igt_spinner_init(&spin, engine->gt);
1314 		if (ret)
1315 			goto err;
1316 
1317 		rq = igt_spinner_create_request(&spin, ce, MI_NOOP);
1318 		if (IS_ERR(rq)) {
1319 			ret = PTR_ERR(rq);
1320 			igt_spinner_fini(&spin);
1321 			goto err;
1322 		}
1323 
1324 		ret = request_add_spin(rq, &spin);
1325 		if (ret) {
1326 			pr_err("%s: Spinner failed to start\n", engine->name);
1327 			igt_spinner_fini(&spin);
1328 			goto err;
1329 		}
1330 
1331 		/* Ensure the spinner hasn't aborted */
1332 		if (i915_request_completed(rq)) {
1333 			ret = -ETIMEDOUT;
1334 			goto skip;
1335 		}
1336 
1337 		if (!using_guc) {
1338 			ret = intel_engine_reset(engine, "live_workarounds:active");
1339 			if (ret) {
1340 				pr_err("%s: Reset failed on an active spinner\n",
1341 				       engine->name);
1342 				igt_spinner_fini(&spin);
1343 				goto err;
1344 			}
1345 		}
1346 
1347 		/* Ensure the reset happens and kills the engine */
1348 		if (ret == 0)
1349 			ret = intel_selftest_wait_for_rq(rq);
1350 
1351 skip:
1352 		igt_spinner_end(&spin);
1353 		igt_spinner_fini(&spin);
1354 
1355 		ok = verify_wa_lists(gt, lists, "after busy reset");
1356 		if (!ok)
1357 			ret = -ESRCH;
1358 
1359 err:
1360 		intel_context_put(ce);
1361 
1362 restore:
1363 		ret2 = intel_selftest_restore_policy(engine, &saved);
1364 		if (ret == 0)
1365 			ret = ret2;
1366 		if (ret)
1367 			break;
1368 	}
1369 
1370 	reference_lists_fini(gt, lists);
1371 	intel_runtime_pm_put(gt->uncore->rpm, wakeref);
1372 	igt_global_reset_unlock(gt);
1373 	kfree(lists);
1374 
1375 	igt_flush_test(gt->i915);
1376 
1377 	return ret;
1378 }
1379 
1380 int intel_workarounds_live_selftests(struct drm_i915_private *i915)
1381 {
1382 	static const struct i915_subtest tests[] = {
1383 		SUBTEST(live_dirty_whitelist),
1384 		SUBTEST(live_reset_whitelist),
1385 		SUBTEST(live_isolated_whitelist),
1386 		SUBTEST(live_gpu_reset_workarounds),
1387 		SUBTEST(live_engine_reset_workarounds),
1388 	};
1389 
1390 	if (intel_gt_is_wedged(to_gt(i915)))
1391 		return 0;
1392 
1393 	return intel_gt_live_subtests(tests, to_gt(i915));
1394 }
1395