xref: /linux/drivers/gpu/drm/i915/gt/selftest_workarounds.c (revision 9a87ffc99ec8eb8d35eed7c4f816d75f5cc9662e)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2018 Intel Corporation
4  */
5 
6 #include "gem/i915_gem_internal.h"
7 #include "gem/i915_gem_pm.h"
8 #include "gt/intel_engine_user.h"
9 #include "gt/intel_gt.h"
10 #include "i915_selftest.h"
11 #include "intel_reset.h"
12 
13 #include "selftests/igt_flush_test.h"
14 #include "selftests/igt_reset.h"
15 #include "selftests/igt_spinner.h"
16 #include "selftests/intel_scheduler_helpers.h"
17 #include "selftests/mock_drm.h"
18 
19 #include "gem/selftests/igt_gem_utils.h"
20 #include "gem/selftests/mock_context.h"
21 
22 static const struct wo_register {
23 	enum intel_platform platform;
24 	u32 reg;
25 } wo_registers[] = {
26 	{ INTEL_GEMINILAKE, 0x731c }
27 };
28 
29 struct wa_lists {
30 	struct i915_wa_list gt_wa_list;
31 	struct {
32 		struct i915_wa_list wa_list;
33 		struct i915_wa_list ctx_wa_list;
34 	} engine[I915_NUM_ENGINES];
35 };
36 
request_add_sync(struct i915_request * rq,int err)37 static int request_add_sync(struct i915_request *rq, int err)
38 {
39 	i915_request_get(rq);
40 	i915_request_add(rq);
41 	if (i915_request_wait(rq, 0, HZ / 5) < 0)
42 		err = -EIO;
43 	i915_request_put(rq);
44 
45 	return err;
46 }
47 
request_add_spin(struct i915_request * rq,struct igt_spinner * spin)48 static int request_add_spin(struct i915_request *rq, struct igt_spinner *spin)
49 {
50 	int err = 0;
51 
52 	i915_request_get(rq);
53 	i915_request_add(rq);
54 	if (spin && !igt_wait_for_spinner(spin, rq))
55 		err = -ETIMEDOUT;
56 	i915_request_put(rq);
57 
58 	return err;
59 }
60 
61 static void
reference_lists_init(struct intel_gt * gt,struct wa_lists * lists)62 reference_lists_init(struct intel_gt *gt, struct wa_lists *lists)
63 {
64 	struct intel_engine_cs *engine;
65 	enum intel_engine_id id;
66 
67 	memset(lists, 0, sizeof(*lists));
68 
69 	wa_init_start(&lists->gt_wa_list, gt, "GT_REF", "global");
70 	gt_init_workarounds(gt, &lists->gt_wa_list);
71 	wa_init_finish(&lists->gt_wa_list);
72 
73 	for_each_engine(engine, gt, id) {
74 		struct i915_wa_list *wal = &lists->engine[id].wa_list;
75 
76 		wa_init_start(wal, gt, "REF", engine->name);
77 		engine_init_workarounds(engine, wal);
78 		wa_init_finish(wal);
79 
80 		__intel_engine_init_ctx_wa(engine,
81 					   &lists->engine[id].ctx_wa_list,
82 					   "CTX_REF");
83 	}
84 }
85 
86 static void
reference_lists_fini(struct intel_gt * gt,struct wa_lists * lists)87 reference_lists_fini(struct intel_gt *gt, struct wa_lists *lists)
88 {
89 	struct intel_engine_cs *engine;
90 	enum intel_engine_id id;
91 
92 	for_each_engine(engine, gt, id)
93 		intel_wa_list_free(&lists->engine[id].wa_list);
94 
95 	intel_wa_list_free(&lists->gt_wa_list);
96 }
97 
98 static struct drm_i915_gem_object *
read_nonprivs(struct intel_context * ce)99 read_nonprivs(struct intel_context *ce)
100 {
101 	struct intel_engine_cs *engine = ce->engine;
102 	const u32 base = engine->mmio_base;
103 	struct drm_i915_gem_object *result;
104 	struct i915_request *rq;
105 	struct i915_vma *vma;
106 	u32 srm, *cs;
107 	int err;
108 	int i;
109 
110 	result = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
111 	if (IS_ERR(result))
112 		return result;
113 
114 	i915_gem_object_set_cache_coherency(result, I915_CACHE_LLC);
115 
116 	cs = i915_gem_object_pin_map_unlocked(result, I915_MAP_WB);
117 	if (IS_ERR(cs)) {
118 		err = PTR_ERR(cs);
119 		goto err_obj;
120 	}
121 	memset(cs, 0xc5, PAGE_SIZE);
122 	i915_gem_object_flush_map(result);
123 	i915_gem_object_unpin_map(result);
124 
125 	vma = i915_vma_instance(result, &engine->gt->ggtt->vm, NULL);
126 	if (IS_ERR(vma)) {
127 		err = PTR_ERR(vma);
128 		goto err_obj;
129 	}
130 
131 	err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL);
132 	if (err)
133 		goto err_obj;
134 
135 	rq = intel_context_create_request(ce);
136 	if (IS_ERR(rq)) {
137 		err = PTR_ERR(rq);
138 		goto err_pin;
139 	}
140 
141 	err = igt_vma_move_to_active_unlocked(vma, rq, EXEC_OBJECT_WRITE);
142 	if (err)
143 		goto err_req;
144 
145 	srm = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
146 	if (GRAPHICS_VER(engine->i915) >= 8)
147 		srm++;
148 
149 	cs = intel_ring_begin(rq, 4 * RING_MAX_NONPRIV_SLOTS);
150 	if (IS_ERR(cs)) {
151 		err = PTR_ERR(cs);
152 		goto err_req;
153 	}
154 
155 	for (i = 0; i < RING_MAX_NONPRIV_SLOTS; i++) {
156 		*cs++ = srm;
157 		*cs++ = i915_mmio_reg_offset(RING_FORCE_TO_NONPRIV(base, i));
158 		*cs++ = i915_ggtt_offset(vma) + sizeof(u32) * i;
159 		*cs++ = 0;
160 	}
161 	intel_ring_advance(rq, cs);
162 
163 	i915_request_add(rq);
164 	i915_vma_unpin(vma);
165 
166 	return result;
167 
168 err_req:
169 	i915_request_add(rq);
170 err_pin:
171 	i915_vma_unpin(vma);
172 err_obj:
173 	i915_gem_object_put(result);
174 	return ERR_PTR(err);
175 }
176 
177 static u32
get_whitelist_reg(const struct intel_engine_cs * engine,unsigned int i)178 get_whitelist_reg(const struct intel_engine_cs *engine, unsigned int i)
179 {
180 	i915_reg_t reg = i < engine->whitelist.count ?
181 			 engine->whitelist.list[i].reg :
182 			 RING_NOPID(engine->mmio_base);
183 
184 	return i915_mmio_reg_offset(reg);
185 }
186 
187 static void
print_results(const struct intel_engine_cs * engine,const u32 * results)188 print_results(const struct intel_engine_cs *engine, const u32 *results)
189 {
190 	unsigned int i;
191 
192 	for (i = 0; i < RING_MAX_NONPRIV_SLOTS; i++) {
193 		u32 expected = get_whitelist_reg(engine, i);
194 		u32 actual = results[i];
195 
196 		pr_info("RING_NONPRIV[%d]: expected 0x%08x, found 0x%08x\n",
197 			i, expected, actual);
198 	}
199 }
200 
check_whitelist(struct intel_context * ce)201 static int check_whitelist(struct intel_context *ce)
202 {
203 	struct intel_engine_cs *engine = ce->engine;
204 	struct drm_i915_gem_object *results;
205 	struct intel_wedge_me wedge;
206 	u32 *vaddr;
207 	int err;
208 	int i;
209 
210 	results = read_nonprivs(ce);
211 	if (IS_ERR(results))
212 		return PTR_ERR(results);
213 
214 	err = 0;
215 	i915_gem_object_lock(results, NULL);
216 	intel_wedge_on_timeout(&wedge, engine->gt, HZ / 5) /* safety net! */
217 		err = i915_gem_object_set_to_cpu_domain(results, false);
218 
219 	if (intel_gt_is_wedged(engine->gt))
220 		err = -EIO;
221 	if (err)
222 		goto out_put;
223 
224 	vaddr = i915_gem_object_pin_map(results, I915_MAP_WB);
225 	if (IS_ERR(vaddr)) {
226 		err = PTR_ERR(vaddr);
227 		goto out_put;
228 	}
229 
230 	for (i = 0; i < RING_MAX_NONPRIV_SLOTS; i++) {
231 		u32 expected = get_whitelist_reg(engine, i);
232 		u32 actual = vaddr[i];
233 
234 		if (expected != actual) {
235 			print_results(engine, vaddr);
236 			pr_err("Invalid RING_NONPRIV[%d], expected 0x%08x, found 0x%08x\n",
237 			       i, expected, actual);
238 
239 			err = -EINVAL;
240 			break;
241 		}
242 	}
243 
244 	i915_gem_object_unpin_map(results);
245 out_put:
246 	i915_gem_object_unlock(results);
247 	i915_gem_object_put(results);
248 	return err;
249 }
250 
do_device_reset(struct intel_engine_cs * engine)251 static int do_device_reset(struct intel_engine_cs *engine)
252 {
253 	intel_gt_reset(engine->gt, engine->mask, "live_workarounds");
254 	return 0;
255 }
256 
do_engine_reset(struct intel_engine_cs * engine)257 static int do_engine_reset(struct intel_engine_cs *engine)
258 {
259 	return intel_engine_reset(engine, "live_workarounds");
260 }
261 
do_guc_reset(struct intel_engine_cs * engine)262 static int do_guc_reset(struct intel_engine_cs *engine)
263 {
264 	/* Currently a no-op as the reset is handled by GuC */
265 	return 0;
266 }
267 
268 static int
switch_to_scratch_context(struct intel_engine_cs * engine,struct igt_spinner * spin,struct i915_request ** rq)269 switch_to_scratch_context(struct intel_engine_cs *engine,
270 			  struct igt_spinner *spin,
271 			  struct i915_request **rq)
272 {
273 	struct intel_context *ce;
274 	int err = 0;
275 
276 	ce = intel_context_create(engine);
277 	if (IS_ERR(ce))
278 		return PTR_ERR(ce);
279 
280 	*rq = igt_spinner_create_request(spin, ce, MI_NOOP);
281 	intel_context_put(ce);
282 
283 	if (IS_ERR(*rq)) {
284 		spin = NULL;
285 		err = PTR_ERR(*rq);
286 		goto err;
287 	}
288 
289 	err = request_add_spin(*rq, spin);
290 err:
291 	if (err && spin)
292 		igt_spinner_end(spin);
293 
294 	return err;
295 }
296 
check_whitelist_across_reset(struct intel_engine_cs * engine,int (* reset)(struct intel_engine_cs *),const char * name)297 static int check_whitelist_across_reset(struct intel_engine_cs *engine,
298 					int (*reset)(struct intel_engine_cs *),
299 					const char *name)
300 {
301 	struct intel_context *ce, *tmp;
302 	struct igt_spinner spin;
303 	struct i915_request *rq;
304 	intel_wakeref_t wakeref;
305 	int err;
306 
307 	pr_info("Checking %d whitelisted registers on %s (RING_NONPRIV) [%s]\n",
308 		engine->whitelist.count, engine->name, name);
309 
310 	ce = intel_context_create(engine);
311 	if (IS_ERR(ce))
312 		return PTR_ERR(ce);
313 
314 	err = igt_spinner_init(&spin, engine->gt);
315 	if (err)
316 		goto out_ctx;
317 
318 	err = check_whitelist(ce);
319 	if (err) {
320 		pr_err("Invalid whitelist *before* %s reset!\n", name);
321 		goto out_spin;
322 	}
323 
324 	err = switch_to_scratch_context(engine, &spin, &rq);
325 	if (err)
326 		goto out_spin;
327 
328 	/* Ensure the spinner hasn't aborted */
329 	if (i915_request_completed(rq)) {
330 		pr_err("%s spinner failed to start\n", name);
331 		err = -ETIMEDOUT;
332 		goto out_spin;
333 	}
334 
335 	with_intel_runtime_pm(engine->uncore->rpm, wakeref)
336 		err = reset(engine);
337 
338 	/* Ensure the reset happens and kills the engine */
339 	if (err == 0)
340 		err = intel_selftest_wait_for_rq(rq);
341 
342 	igt_spinner_end(&spin);
343 
344 	if (err) {
345 		pr_err("%s reset failed\n", name);
346 		goto out_spin;
347 	}
348 
349 	err = check_whitelist(ce);
350 	if (err) {
351 		pr_err("Whitelist not preserved in context across %s reset!\n",
352 		       name);
353 		goto out_spin;
354 	}
355 
356 	tmp = intel_context_create(engine);
357 	if (IS_ERR(tmp)) {
358 		err = PTR_ERR(tmp);
359 		goto out_spin;
360 	}
361 	intel_context_put(ce);
362 	ce = tmp;
363 
364 	err = check_whitelist(ce);
365 	if (err) {
366 		pr_err("Invalid whitelist *after* %s reset in fresh context!\n",
367 		       name);
368 		goto out_spin;
369 	}
370 
371 out_spin:
372 	igt_spinner_fini(&spin);
373 out_ctx:
374 	intel_context_put(ce);
375 	return err;
376 }
377 
create_batch(struct i915_address_space * vm)378 static struct i915_vma *create_batch(struct i915_address_space *vm)
379 {
380 	struct drm_i915_gem_object *obj;
381 	struct i915_vma *vma;
382 	int err;
383 
384 	obj = i915_gem_object_create_internal(vm->i915, 16 * PAGE_SIZE);
385 	if (IS_ERR(obj))
386 		return ERR_CAST(obj);
387 
388 	vma = i915_vma_instance(obj, vm, NULL);
389 	if (IS_ERR(vma)) {
390 		err = PTR_ERR(vma);
391 		goto err_obj;
392 	}
393 
394 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
395 	if (err)
396 		goto err_obj;
397 
398 	return vma;
399 
400 err_obj:
401 	i915_gem_object_put(obj);
402 	return ERR_PTR(err);
403 }
404 
reg_write(u32 old,u32 new,u32 rsvd)405 static u32 reg_write(u32 old, u32 new, u32 rsvd)
406 {
407 	if (rsvd == 0x0000ffff) {
408 		old &= ~(new >> 16);
409 		old |= new & (new >> 16);
410 	} else {
411 		old &= ~rsvd;
412 		old |= new & rsvd;
413 	}
414 
415 	return old;
416 }
417 
wo_register(struct intel_engine_cs * engine,u32 reg)418 static bool wo_register(struct intel_engine_cs *engine, u32 reg)
419 {
420 	enum intel_platform platform = INTEL_INFO(engine->i915)->platform;
421 	int i;
422 
423 	if ((reg & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
424 	     RING_FORCE_TO_NONPRIV_ACCESS_WR)
425 		return true;
426 
427 	for (i = 0; i < ARRAY_SIZE(wo_registers); i++) {
428 		if (wo_registers[i].platform == platform &&
429 		    wo_registers[i].reg == reg)
430 			return true;
431 	}
432 
433 	return false;
434 }
435 
timestamp(const struct intel_engine_cs * engine,u32 reg)436 static bool timestamp(const struct intel_engine_cs *engine, u32 reg)
437 {
438 	reg = (reg - engine->mmio_base) & ~RING_FORCE_TO_NONPRIV_ACCESS_MASK;
439 	switch (reg) {
440 	case 0x358:
441 	case 0x35c:
442 	case 0x3a8:
443 		return true;
444 
445 	default:
446 		return false;
447 	}
448 }
449 
ro_register(u32 reg)450 static bool ro_register(u32 reg)
451 {
452 	if ((reg & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
453 	     RING_FORCE_TO_NONPRIV_ACCESS_RD)
454 		return true;
455 
456 	return false;
457 }
458 
whitelist_writable_count(struct intel_engine_cs * engine)459 static int whitelist_writable_count(struct intel_engine_cs *engine)
460 {
461 	int count = engine->whitelist.count;
462 	int i;
463 
464 	for (i = 0; i < engine->whitelist.count; i++) {
465 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
466 
467 		if (ro_register(reg))
468 			count--;
469 	}
470 
471 	return count;
472 }
473 
check_dirty_whitelist(struct intel_context * ce)474 static int check_dirty_whitelist(struct intel_context *ce)
475 {
476 	const u32 values[] = {
477 		0x00000000,
478 		0x01010101,
479 		0x10100101,
480 		0x03030303,
481 		0x30300303,
482 		0x05050505,
483 		0x50500505,
484 		0x0f0f0f0f,
485 		0xf00ff00f,
486 		0x10101010,
487 		0xf0f01010,
488 		0x30303030,
489 		0xa0a03030,
490 		0x50505050,
491 		0xc0c05050,
492 		0xf0f0f0f0,
493 		0x11111111,
494 		0x33333333,
495 		0x55555555,
496 		0x0000ffff,
497 		0x00ff00ff,
498 		0xff0000ff,
499 		0xffff00ff,
500 		0xffffffff,
501 	};
502 	struct intel_engine_cs *engine = ce->engine;
503 	struct i915_vma *scratch;
504 	struct i915_vma *batch;
505 	int err = 0, i, v, sz;
506 	u32 *cs, *results;
507 
508 	sz = (2 * ARRAY_SIZE(values) + 1) * sizeof(u32);
509 	scratch = __vm_create_scratch_for_read_pinned(ce->vm, sz);
510 	if (IS_ERR(scratch))
511 		return PTR_ERR(scratch);
512 
513 	batch = create_batch(ce->vm);
514 	if (IS_ERR(batch)) {
515 		err = PTR_ERR(batch);
516 		goto out_scratch;
517 	}
518 
519 	for (i = 0; i < engine->whitelist.count; i++) {
520 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
521 		struct i915_gem_ww_ctx ww;
522 		u64 addr = i915_vma_offset(scratch);
523 		struct i915_request *rq;
524 		u32 srm, lrm, rsvd;
525 		u32 expect;
526 		int idx;
527 		bool ro_reg;
528 
529 		if (wo_register(engine, reg))
530 			continue;
531 
532 		if (timestamp(engine, reg))
533 			continue; /* timestamps are expected to autoincrement */
534 
535 		ro_reg = ro_register(reg);
536 
537 		i915_gem_ww_ctx_init(&ww, false);
538 retry:
539 		cs = NULL;
540 		err = i915_gem_object_lock(scratch->obj, &ww);
541 		if (!err)
542 			err = i915_gem_object_lock(batch->obj, &ww);
543 		if (!err)
544 			err = intel_context_pin_ww(ce, &ww);
545 		if (err)
546 			goto out;
547 
548 		cs = i915_gem_object_pin_map(batch->obj, I915_MAP_WC);
549 		if (IS_ERR(cs)) {
550 			err = PTR_ERR(cs);
551 			goto out_ctx;
552 		}
553 
554 		results = i915_gem_object_pin_map(scratch->obj, I915_MAP_WB);
555 		if (IS_ERR(results)) {
556 			err = PTR_ERR(results);
557 			goto out_unmap_batch;
558 		}
559 
560 		/* Clear non priv flags */
561 		reg &= RING_FORCE_TO_NONPRIV_ADDRESS_MASK;
562 
563 		srm = MI_STORE_REGISTER_MEM;
564 		lrm = MI_LOAD_REGISTER_MEM;
565 		if (GRAPHICS_VER(engine->i915) >= 8)
566 			lrm++, srm++;
567 
568 		pr_debug("%s: Writing garbage to %x\n",
569 			 engine->name, reg);
570 
571 		/* SRM original */
572 		*cs++ = srm;
573 		*cs++ = reg;
574 		*cs++ = lower_32_bits(addr);
575 		*cs++ = upper_32_bits(addr);
576 
577 		idx = 1;
578 		for (v = 0; v < ARRAY_SIZE(values); v++) {
579 			/* LRI garbage */
580 			*cs++ = MI_LOAD_REGISTER_IMM(1);
581 			*cs++ = reg;
582 			*cs++ = values[v];
583 
584 			/* SRM result */
585 			*cs++ = srm;
586 			*cs++ = reg;
587 			*cs++ = lower_32_bits(addr + sizeof(u32) * idx);
588 			*cs++ = upper_32_bits(addr + sizeof(u32) * idx);
589 			idx++;
590 		}
591 		for (v = 0; v < ARRAY_SIZE(values); v++) {
592 			/* LRI garbage */
593 			*cs++ = MI_LOAD_REGISTER_IMM(1);
594 			*cs++ = reg;
595 			*cs++ = ~values[v];
596 
597 			/* SRM result */
598 			*cs++ = srm;
599 			*cs++ = reg;
600 			*cs++ = lower_32_bits(addr + sizeof(u32) * idx);
601 			*cs++ = upper_32_bits(addr + sizeof(u32) * idx);
602 			idx++;
603 		}
604 		GEM_BUG_ON(idx * sizeof(u32) > scratch->size);
605 
606 		/* LRM original -- don't leave garbage in the context! */
607 		*cs++ = lrm;
608 		*cs++ = reg;
609 		*cs++ = lower_32_bits(addr);
610 		*cs++ = upper_32_bits(addr);
611 
612 		*cs++ = MI_BATCH_BUFFER_END;
613 
614 		i915_gem_object_flush_map(batch->obj);
615 		i915_gem_object_unpin_map(batch->obj);
616 		intel_gt_chipset_flush(engine->gt);
617 		cs = NULL;
618 
619 		rq = i915_request_create(ce);
620 		if (IS_ERR(rq)) {
621 			err = PTR_ERR(rq);
622 			goto out_unmap_scratch;
623 		}
624 
625 		if (engine->emit_init_breadcrumb) { /* Be nice if we hang */
626 			err = engine->emit_init_breadcrumb(rq);
627 			if (err)
628 				goto err_request;
629 		}
630 
631 		err = i915_vma_move_to_active(batch, rq, 0);
632 		if (err)
633 			goto err_request;
634 
635 		err = i915_vma_move_to_active(scratch, rq,
636 					      EXEC_OBJECT_WRITE);
637 		if (err)
638 			goto err_request;
639 
640 		err = engine->emit_bb_start(rq,
641 					    i915_vma_offset(batch), PAGE_SIZE,
642 					    0);
643 		if (err)
644 			goto err_request;
645 
646 err_request:
647 		err = request_add_sync(rq, err);
648 		if (err) {
649 			pr_err("%s: Futzing %x timedout; cancelling test\n",
650 			       engine->name, reg);
651 			intel_gt_set_wedged(engine->gt);
652 			goto out_unmap_scratch;
653 		}
654 
655 		GEM_BUG_ON(values[ARRAY_SIZE(values) - 1] != 0xffffffff);
656 		if (!ro_reg) {
657 			/* detect write masking */
658 			rsvd = results[ARRAY_SIZE(values)];
659 			if (!rsvd) {
660 				pr_err("%s: Unable to write to whitelisted register %x\n",
661 				       engine->name, reg);
662 				err = -EINVAL;
663 				goto out_unmap_scratch;
664 			}
665 		} else {
666 			rsvd = 0;
667 		}
668 
669 		expect = results[0];
670 		idx = 1;
671 		for (v = 0; v < ARRAY_SIZE(values); v++) {
672 			if (ro_reg)
673 				expect = results[0];
674 			else
675 				expect = reg_write(expect, values[v], rsvd);
676 
677 			if (results[idx] != expect)
678 				err++;
679 			idx++;
680 		}
681 		for (v = 0; v < ARRAY_SIZE(values); v++) {
682 			if (ro_reg)
683 				expect = results[0];
684 			else
685 				expect = reg_write(expect, ~values[v], rsvd);
686 
687 			if (results[idx] != expect)
688 				err++;
689 			idx++;
690 		}
691 		if (err) {
692 			pr_err("%s: %d mismatch between values written to whitelisted register [%x], and values read back!\n",
693 			       engine->name, err, reg);
694 
695 			if (ro_reg)
696 				pr_info("%s: Whitelisted read-only register: %x, original value %08x\n",
697 					engine->name, reg, results[0]);
698 			else
699 				pr_info("%s: Whitelisted register: %x, original value %08x, rsvd %08x\n",
700 					engine->name, reg, results[0], rsvd);
701 
702 			expect = results[0];
703 			idx = 1;
704 			for (v = 0; v < ARRAY_SIZE(values); v++) {
705 				u32 w = values[v];
706 
707 				if (ro_reg)
708 					expect = results[0];
709 				else
710 					expect = reg_write(expect, w, rsvd);
711 				pr_info("Wrote %08x, read %08x, expect %08x\n",
712 					w, results[idx], expect);
713 				idx++;
714 			}
715 			for (v = 0; v < ARRAY_SIZE(values); v++) {
716 				u32 w = ~values[v];
717 
718 				if (ro_reg)
719 					expect = results[0];
720 				else
721 					expect = reg_write(expect, w, rsvd);
722 				pr_info("Wrote %08x, read %08x, expect %08x\n",
723 					w, results[idx], expect);
724 				idx++;
725 			}
726 
727 			err = -EINVAL;
728 		}
729 out_unmap_scratch:
730 		i915_gem_object_unpin_map(scratch->obj);
731 out_unmap_batch:
732 		if (cs)
733 			i915_gem_object_unpin_map(batch->obj);
734 out_ctx:
735 		intel_context_unpin(ce);
736 out:
737 		if (err == -EDEADLK) {
738 			err = i915_gem_ww_ctx_backoff(&ww);
739 			if (!err)
740 				goto retry;
741 		}
742 		i915_gem_ww_ctx_fini(&ww);
743 		if (err)
744 			break;
745 	}
746 
747 	if (igt_flush_test(engine->i915))
748 		err = -EIO;
749 
750 	i915_vma_unpin_and_release(&batch, 0);
751 out_scratch:
752 	i915_vma_unpin_and_release(&scratch, 0);
753 	return err;
754 }
755 
live_dirty_whitelist(void * arg)756 static int live_dirty_whitelist(void *arg)
757 {
758 	struct intel_gt *gt = arg;
759 	struct intel_engine_cs *engine;
760 	enum intel_engine_id id;
761 
762 	/* Can the user write to the whitelisted registers? */
763 
764 	if (GRAPHICS_VER(gt->i915) < 7) /* minimum requirement for LRI, SRM, LRM */
765 		return 0;
766 
767 	for_each_engine(engine, gt, id) {
768 		struct intel_context *ce;
769 		int err;
770 
771 		if (engine->whitelist.count == 0)
772 			continue;
773 
774 		ce = intel_context_create(engine);
775 		if (IS_ERR(ce))
776 			return PTR_ERR(ce);
777 
778 		err = check_dirty_whitelist(ce);
779 		intel_context_put(ce);
780 		if (err)
781 			return err;
782 	}
783 
784 	return 0;
785 }
786 
live_reset_whitelist(void * arg)787 static int live_reset_whitelist(void *arg)
788 {
789 	struct intel_gt *gt = arg;
790 	struct intel_engine_cs *engine;
791 	enum intel_engine_id id;
792 	int err = 0;
793 
794 	/* If we reset the gpu, we should not lose the RING_NONPRIV */
795 	igt_global_reset_lock(gt);
796 
797 	for_each_engine(engine, gt, id) {
798 		if (engine->whitelist.count == 0)
799 			continue;
800 
801 		if (intel_has_reset_engine(gt)) {
802 			if (intel_engine_uses_guc(engine)) {
803 				struct intel_selftest_saved_policy saved;
804 				int err2;
805 
806 				err = intel_selftest_modify_policy(engine, &saved,
807 								   SELFTEST_SCHEDULER_MODIFY_FAST_RESET);
808 				if (err)
809 					goto out;
810 
811 				err = check_whitelist_across_reset(engine,
812 								   do_guc_reset,
813 								   "guc");
814 
815 				err2 = intel_selftest_restore_policy(engine, &saved);
816 				if (err == 0)
817 					err = err2;
818 			} else {
819 				err = check_whitelist_across_reset(engine,
820 								   do_engine_reset,
821 								   "engine");
822 			}
823 
824 			if (err)
825 				goto out;
826 		}
827 
828 		if (intel_has_gpu_reset(gt)) {
829 			err = check_whitelist_across_reset(engine,
830 							   do_device_reset,
831 							   "device");
832 			if (err)
833 				goto out;
834 		}
835 	}
836 
837 out:
838 	igt_global_reset_unlock(gt);
839 	return err;
840 }
841 
read_whitelisted_registers(struct intel_context * ce,struct i915_vma * results)842 static int read_whitelisted_registers(struct intel_context *ce,
843 				      struct i915_vma *results)
844 {
845 	struct intel_engine_cs *engine = ce->engine;
846 	struct i915_request *rq;
847 	int i, err = 0;
848 	u32 srm, *cs;
849 
850 	rq = intel_context_create_request(ce);
851 	if (IS_ERR(rq))
852 		return PTR_ERR(rq);
853 
854 	err = igt_vma_move_to_active_unlocked(results, rq, EXEC_OBJECT_WRITE);
855 	if (err)
856 		goto err_req;
857 
858 	srm = MI_STORE_REGISTER_MEM;
859 	if (GRAPHICS_VER(engine->i915) >= 8)
860 		srm++;
861 
862 	cs = intel_ring_begin(rq, 4 * engine->whitelist.count);
863 	if (IS_ERR(cs)) {
864 		err = PTR_ERR(cs);
865 		goto err_req;
866 	}
867 
868 	for (i = 0; i < engine->whitelist.count; i++) {
869 		u64 offset = i915_vma_offset(results) + sizeof(u32) * i;
870 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
871 
872 		/* Clear non priv flags */
873 		reg &= RING_FORCE_TO_NONPRIV_ADDRESS_MASK;
874 
875 		*cs++ = srm;
876 		*cs++ = reg;
877 		*cs++ = lower_32_bits(offset);
878 		*cs++ = upper_32_bits(offset);
879 	}
880 	intel_ring_advance(rq, cs);
881 
882 err_req:
883 	return request_add_sync(rq, err);
884 }
885 
scrub_whitelisted_registers(struct intel_context * ce)886 static int scrub_whitelisted_registers(struct intel_context *ce)
887 {
888 	struct intel_engine_cs *engine = ce->engine;
889 	struct i915_request *rq;
890 	struct i915_vma *batch;
891 	int i, err = 0;
892 	u32 *cs;
893 
894 	batch = create_batch(ce->vm);
895 	if (IS_ERR(batch))
896 		return PTR_ERR(batch);
897 
898 	cs = i915_gem_object_pin_map_unlocked(batch->obj, I915_MAP_WC);
899 	if (IS_ERR(cs)) {
900 		err = PTR_ERR(cs);
901 		goto err_batch;
902 	}
903 
904 	*cs++ = MI_LOAD_REGISTER_IMM(whitelist_writable_count(engine));
905 	for (i = 0; i < engine->whitelist.count; i++) {
906 		u32 reg = i915_mmio_reg_offset(engine->whitelist.list[i].reg);
907 
908 		if (ro_register(reg))
909 			continue;
910 
911 		/* Clear non priv flags */
912 		reg &= RING_FORCE_TO_NONPRIV_ADDRESS_MASK;
913 
914 		*cs++ = reg;
915 		*cs++ = 0xffffffff;
916 	}
917 	*cs++ = MI_BATCH_BUFFER_END;
918 
919 	i915_gem_object_flush_map(batch->obj);
920 	intel_gt_chipset_flush(engine->gt);
921 
922 	rq = intel_context_create_request(ce);
923 	if (IS_ERR(rq)) {
924 		err = PTR_ERR(rq);
925 		goto err_unpin;
926 	}
927 
928 	if (engine->emit_init_breadcrumb) { /* Be nice if we hang */
929 		err = engine->emit_init_breadcrumb(rq);
930 		if (err)
931 			goto err_request;
932 	}
933 
934 	err = igt_vma_move_to_active_unlocked(batch, rq, 0);
935 	if (err)
936 		goto err_request;
937 
938 	/* Perform the writes from an unprivileged "user" batch */
939 	err = engine->emit_bb_start(rq, i915_vma_offset(batch), 0, 0);
940 
941 err_request:
942 	err = request_add_sync(rq, err);
943 
944 err_unpin:
945 	i915_gem_object_unpin_map(batch->obj);
946 err_batch:
947 	i915_vma_unpin_and_release(&batch, 0);
948 	return err;
949 }
950 
951 struct regmask {
952 	i915_reg_t reg;
953 	u8 graphics_ver;
954 };
955 
find_reg(struct drm_i915_private * i915,i915_reg_t reg,const struct regmask * tbl,unsigned long count)956 static bool find_reg(struct drm_i915_private *i915,
957 		     i915_reg_t reg,
958 		     const struct regmask *tbl,
959 		     unsigned long count)
960 {
961 	u32 offset = i915_mmio_reg_offset(reg);
962 
963 	while (count--) {
964 		if (GRAPHICS_VER(i915) == tbl->graphics_ver &&
965 		    i915_mmio_reg_offset(tbl->reg) == offset)
966 			return true;
967 		tbl++;
968 	}
969 
970 	return false;
971 }
972 
pardon_reg(struct drm_i915_private * i915,i915_reg_t reg)973 static bool pardon_reg(struct drm_i915_private *i915, i915_reg_t reg)
974 {
975 	/* Alas, we must pardon some whitelists. Mistakes already made */
976 	static const struct regmask pardon[] = {
977 		{ GEN9_CTX_PREEMPT_REG, 9 },
978 		{ _MMIO(0xb118), 9 }, /* GEN8_L3SQCREG4 */
979 	};
980 
981 	return find_reg(i915, reg, pardon, ARRAY_SIZE(pardon));
982 }
983 
result_eq(struct intel_engine_cs * engine,u32 a,u32 b,i915_reg_t reg)984 static bool result_eq(struct intel_engine_cs *engine,
985 		      u32 a, u32 b, i915_reg_t reg)
986 {
987 	if (a != b && !pardon_reg(engine->i915, reg)) {
988 		pr_err("Whitelisted register 0x%4x not context saved: A=%08x, B=%08x\n",
989 		       i915_mmio_reg_offset(reg), a, b);
990 		return false;
991 	}
992 
993 	return true;
994 }
995 
writeonly_reg(struct drm_i915_private * i915,i915_reg_t reg)996 static bool writeonly_reg(struct drm_i915_private *i915, i915_reg_t reg)
997 {
998 	/* Some registers do not seem to behave and our writes unreadable */
999 	static const struct regmask wo[] = {
1000 		{ GEN9_SLICE_COMMON_ECO_CHICKEN1, 9 },
1001 	};
1002 
1003 	return find_reg(i915, reg, wo, ARRAY_SIZE(wo));
1004 }
1005 
result_neq(struct intel_engine_cs * engine,u32 a,u32 b,i915_reg_t reg)1006 static bool result_neq(struct intel_engine_cs *engine,
1007 		       u32 a, u32 b, i915_reg_t reg)
1008 {
1009 	if (a == b && !writeonly_reg(engine->i915, reg)) {
1010 		pr_err("Whitelist register 0x%4x:%08x was unwritable\n",
1011 		       i915_mmio_reg_offset(reg), a);
1012 		return false;
1013 	}
1014 
1015 	return true;
1016 }
1017 
1018 static int
check_whitelisted_registers(struct intel_engine_cs * engine,struct i915_vma * A,struct i915_vma * B,bool (* fn)(struct intel_engine_cs * engine,u32 a,u32 b,i915_reg_t reg))1019 check_whitelisted_registers(struct intel_engine_cs *engine,
1020 			    struct i915_vma *A,
1021 			    struct i915_vma *B,
1022 			    bool (*fn)(struct intel_engine_cs *engine,
1023 				       u32 a, u32 b,
1024 				       i915_reg_t reg))
1025 {
1026 	u32 *a, *b;
1027 	int i, err;
1028 
1029 	a = i915_gem_object_pin_map_unlocked(A->obj, I915_MAP_WB);
1030 	if (IS_ERR(a))
1031 		return PTR_ERR(a);
1032 
1033 	b = i915_gem_object_pin_map_unlocked(B->obj, I915_MAP_WB);
1034 	if (IS_ERR(b)) {
1035 		err = PTR_ERR(b);
1036 		goto err_a;
1037 	}
1038 
1039 	err = 0;
1040 	for (i = 0; i < engine->whitelist.count; i++) {
1041 		const struct i915_wa *wa = &engine->whitelist.list[i];
1042 
1043 		if (i915_mmio_reg_offset(wa->reg) &
1044 		    RING_FORCE_TO_NONPRIV_ACCESS_RD)
1045 			continue;
1046 
1047 		if (!fn(engine, a[i], b[i], wa->reg))
1048 			err = -EINVAL;
1049 	}
1050 
1051 	i915_gem_object_unpin_map(B->obj);
1052 err_a:
1053 	i915_gem_object_unpin_map(A->obj);
1054 	return err;
1055 }
1056 
live_isolated_whitelist(void * arg)1057 static int live_isolated_whitelist(void *arg)
1058 {
1059 	struct intel_gt *gt = arg;
1060 	struct {
1061 		struct i915_vma *scratch[2];
1062 	} client[2] = {};
1063 	struct intel_engine_cs *engine;
1064 	enum intel_engine_id id;
1065 	int i, err = 0;
1066 
1067 	/*
1068 	 * Check that a write into a whitelist register works, but
1069 	 * invisible to a second context.
1070 	 */
1071 
1072 	if (!intel_engines_has_context_isolation(gt->i915))
1073 		return 0;
1074 
1075 	for (i = 0; i < ARRAY_SIZE(client); i++) {
1076 		client[i].scratch[0] =
1077 			__vm_create_scratch_for_read_pinned(gt->vm, 4096);
1078 		if (IS_ERR(client[i].scratch[0])) {
1079 			err = PTR_ERR(client[i].scratch[0]);
1080 			goto err;
1081 		}
1082 
1083 		client[i].scratch[1] =
1084 			__vm_create_scratch_for_read_pinned(gt->vm, 4096);
1085 		if (IS_ERR(client[i].scratch[1])) {
1086 			err = PTR_ERR(client[i].scratch[1]);
1087 			i915_vma_unpin_and_release(&client[i].scratch[0], 0);
1088 			goto err;
1089 		}
1090 	}
1091 
1092 	for_each_engine(engine, gt, id) {
1093 		struct intel_context *ce[2];
1094 
1095 		if (!engine->kernel_context->vm)
1096 			continue;
1097 
1098 		if (!whitelist_writable_count(engine))
1099 			continue;
1100 
1101 		ce[0] = intel_context_create(engine);
1102 		if (IS_ERR(ce[0])) {
1103 			err = PTR_ERR(ce[0]);
1104 			break;
1105 		}
1106 		ce[1] = intel_context_create(engine);
1107 		if (IS_ERR(ce[1])) {
1108 			err = PTR_ERR(ce[1]);
1109 			intel_context_put(ce[0]);
1110 			break;
1111 		}
1112 
1113 		/* Read default values */
1114 		err = read_whitelisted_registers(ce[0], client[0].scratch[0]);
1115 		if (err)
1116 			goto err_ce;
1117 
1118 		/* Try to overwrite registers (should only affect ctx0) */
1119 		err = scrub_whitelisted_registers(ce[0]);
1120 		if (err)
1121 			goto err_ce;
1122 
1123 		/* Read values from ctx1, we expect these to be defaults */
1124 		err = read_whitelisted_registers(ce[1], client[1].scratch[0]);
1125 		if (err)
1126 			goto err_ce;
1127 
1128 		/* Verify that both reads return the same default values */
1129 		err = check_whitelisted_registers(engine,
1130 						  client[0].scratch[0],
1131 						  client[1].scratch[0],
1132 						  result_eq);
1133 		if (err)
1134 			goto err_ce;
1135 
1136 		/* Read back the updated values in ctx0 */
1137 		err = read_whitelisted_registers(ce[0], client[0].scratch[1]);
1138 		if (err)
1139 			goto err_ce;
1140 
1141 		/* User should be granted privilege to overwhite regs */
1142 		err = check_whitelisted_registers(engine,
1143 						  client[0].scratch[0],
1144 						  client[0].scratch[1],
1145 						  result_neq);
1146 err_ce:
1147 		intel_context_put(ce[1]);
1148 		intel_context_put(ce[0]);
1149 		if (err)
1150 			break;
1151 	}
1152 
1153 err:
1154 	for (i = 0; i < ARRAY_SIZE(client); i++) {
1155 		i915_vma_unpin_and_release(&client[i].scratch[1], 0);
1156 		i915_vma_unpin_and_release(&client[i].scratch[0], 0);
1157 	}
1158 
1159 	if (igt_flush_test(gt->i915))
1160 		err = -EIO;
1161 
1162 	return err;
1163 }
1164 
1165 static bool
verify_wa_lists(struct intel_gt * gt,struct wa_lists * lists,const char * str)1166 verify_wa_lists(struct intel_gt *gt, struct wa_lists *lists,
1167 		const char *str)
1168 {
1169 	struct intel_engine_cs *engine;
1170 	enum intel_engine_id id;
1171 	bool ok = true;
1172 
1173 	ok &= wa_list_verify(gt, &lists->gt_wa_list, str);
1174 
1175 	for_each_engine(engine, gt, id) {
1176 		struct intel_context *ce;
1177 
1178 		ce = intel_context_create(engine);
1179 		if (IS_ERR(ce))
1180 			return false;
1181 
1182 		ok &= engine_wa_list_verify(ce,
1183 					    &lists->engine[id].wa_list,
1184 					    str) == 0;
1185 
1186 		ok &= engine_wa_list_verify(ce,
1187 					    &lists->engine[id].ctx_wa_list,
1188 					    str) == 0;
1189 
1190 		intel_context_put(ce);
1191 	}
1192 
1193 	return ok;
1194 }
1195 
1196 static int
live_gpu_reset_workarounds(void * arg)1197 live_gpu_reset_workarounds(void *arg)
1198 {
1199 	struct intel_gt *gt = arg;
1200 	intel_wakeref_t wakeref;
1201 	struct wa_lists *lists;
1202 	bool ok;
1203 
1204 	if (!intel_has_gpu_reset(gt))
1205 		return 0;
1206 
1207 	lists = kzalloc(sizeof(*lists), GFP_KERNEL);
1208 	if (!lists)
1209 		return -ENOMEM;
1210 
1211 	pr_info("Verifying after GPU reset...\n");
1212 
1213 	igt_global_reset_lock(gt);
1214 	wakeref = intel_runtime_pm_get(gt->uncore->rpm);
1215 
1216 	reference_lists_init(gt, lists);
1217 
1218 	ok = verify_wa_lists(gt, lists, "before reset");
1219 	if (!ok)
1220 		goto out;
1221 
1222 	intel_gt_reset(gt, ALL_ENGINES, "live_workarounds");
1223 
1224 	ok = verify_wa_lists(gt, lists, "after reset");
1225 
1226 out:
1227 	reference_lists_fini(gt, lists);
1228 	intel_runtime_pm_put(gt->uncore->rpm, wakeref);
1229 	igt_global_reset_unlock(gt);
1230 	kfree(lists);
1231 
1232 	return ok ? 0 : -ESRCH;
1233 }
1234 
1235 static int
live_engine_reset_workarounds(void * arg)1236 live_engine_reset_workarounds(void *arg)
1237 {
1238 	struct intel_gt *gt = arg;
1239 	struct intel_engine_cs *engine;
1240 	enum intel_engine_id id;
1241 	struct intel_context *ce;
1242 	struct igt_spinner spin;
1243 	struct i915_request *rq;
1244 	intel_wakeref_t wakeref;
1245 	struct wa_lists *lists;
1246 	int ret = 0;
1247 
1248 	if (!intel_has_reset_engine(gt))
1249 		return 0;
1250 
1251 	lists = kzalloc(sizeof(*lists), GFP_KERNEL);
1252 	if (!lists)
1253 		return -ENOMEM;
1254 
1255 	igt_global_reset_lock(gt);
1256 	wakeref = intel_runtime_pm_get(gt->uncore->rpm);
1257 
1258 	reference_lists_init(gt, lists);
1259 
1260 	for_each_engine(engine, gt, id) {
1261 		struct intel_selftest_saved_policy saved;
1262 		bool using_guc = intel_engine_uses_guc(engine);
1263 		bool ok;
1264 		int ret2;
1265 
1266 		pr_info("Verifying after %s reset...\n", engine->name);
1267 		ret = intel_selftest_modify_policy(engine, &saved,
1268 						   SELFTEST_SCHEDULER_MODIFY_FAST_RESET);
1269 		if (ret)
1270 			break;
1271 
1272 		ce = intel_context_create(engine);
1273 		if (IS_ERR(ce)) {
1274 			ret = PTR_ERR(ce);
1275 			goto restore;
1276 		}
1277 
1278 		if (!using_guc) {
1279 			ok = verify_wa_lists(gt, lists, "before reset");
1280 			if (!ok) {
1281 				ret = -ESRCH;
1282 				goto err;
1283 			}
1284 
1285 			ret = intel_engine_reset(engine, "live_workarounds:idle");
1286 			if (ret) {
1287 				pr_err("%s: Reset failed while idle\n", engine->name);
1288 				goto err;
1289 			}
1290 
1291 			ok = verify_wa_lists(gt, lists, "after idle reset");
1292 			if (!ok) {
1293 				ret = -ESRCH;
1294 				goto err;
1295 			}
1296 		}
1297 
1298 		ret = igt_spinner_init(&spin, engine->gt);
1299 		if (ret)
1300 			goto err;
1301 
1302 		rq = igt_spinner_create_request(&spin, ce, MI_NOOP);
1303 		if (IS_ERR(rq)) {
1304 			ret = PTR_ERR(rq);
1305 			igt_spinner_fini(&spin);
1306 			goto err;
1307 		}
1308 
1309 		ret = request_add_spin(rq, &spin);
1310 		if (ret) {
1311 			pr_err("%s: Spinner failed to start\n", engine->name);
1312 			igt_spinner_fini(&spin);
1313 			goto err;
1314 		}
1315 
1316 		/* Ensure the spinner hasn't aborted */
1317 		if (i915_request_completed(rq)) {
1318 			ret = -ETIMEDOUT;
1319 			goto skip;
1320 		}
1321 
1322 		if (!using_guc) {
1323 			ret = intel_engine_reset(engine, "live_workarounds:active");
1324 			if (ret) {
1325 				pr_err("%s: Reset failed on an active spinner\n",
1326 				       engine->name);
1327 				igt_spinner_fini(&spin);
1328 				goto err;
1329 			}
1330 		}
1331 
1332 		/* Ensure the reset happens and kills the engine */
1333 		if (ret == 0)
1334 			ret = intel_selftest_wait_for_rq(rq);
1335 
1336 skip:
1337 		igt_spinner_end(&spin);
1338 		igt_spinner_fini(&spin);
1339 
1340 		ok = verify_wa_lists(gt, lists, "after busy reset");
1341 		if (!ok)
1342 			ret = -ESRCH;
1343 
1344 err:
1345 		intel_context_put(ce);
1346 
1347 restore:
1348 		ret2 = intel_selftest_restore_policy(engine, &saved);
1349 		if (ret == 0)
1350 			ret = ret2;
1351 		if (ret)
1352 			break;
1353 	}
1354 
1355 	reference_lists_fini(gt, lists);
1356 	intel_runtime_pm_put(gt->uncore->rpm, wakeref);
1357 	igt_global_reset_unlock(gt);
1358 	kfree(lists);
1359 
1360 	igt_flush_test(gt->i915);
1361 
1362 	return ret;
1363 }
1364 
intel_workarounds_live_selftests(struct drm_i915_private * i915)1365 int intel_workarounds_live_selftests(struct drm_i915_private *i915)
1366 {
1367 	static const struct i915_subtest tests[] = {
1368 		SUBTEST(live_dirty_whitelist),
1369 		SUBTEST(live_reset_whitelist),
1370 		SUBTEST(live_isolated_whitelist),
1371 		SUBTEST(live_gpu_reset_workarounds),
1372 		SUBTEST(live_engine_reset_workarounds),
1373 	};
1374 
1375 	if (intel_gt_is_wedged(to_gt(i915)))
1376 		return 0;
1377 
1378 	return intel_gt_live_subtests(tests, to_gt(i915));
1379 }
1380