1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include <linux/prime_numbers.h>
7 #include <linux/sort.h>
8 
9 #include <drm/drm_buddy.h>
10 
11 #include "../i915_selftest.h"
12 
13 #include "mock_drm.h"
14 #include "mock_gem_device.h"
15 #include "mock_region.h"
16 
17 #include "gem/i915_gem_context.h"
18 #include "gem/i915_gem_lmem.h"
19 #include "gem/i915_gem_region.h"
20 #include "gem/i915_gem_ttm.h"
21 #include "gem/selftests/igt_gem_utils.h"
22 #include "gem/selftests/mock_context.h"
23 #include "gt/intel_engine_pm.h"
24 #include "gt/intel_engine_user.h"
25 #include "gt/intel_gt.h"
26 #include "gt/intel_migrate.h"
27 #include "i915_memcpy.h"
28 #include "i915_ttm_buddy_manager.h"
29 #include "selftests/igt_flush_test.h"
30 #include "selftests/i915_random.h"
31 
32 static void close_objects(struct intel_memory_region *mem,
33 			  struct list_head *objects)
34 {
35 	struct drm_i915_private *i915 = mem->i915;
36 	struct drm_i915_gem_object *obj, *on;
37 
38 	list_for_each_entry_safe(obj, on, objects, st_link) {
39 		i915_gem_object_lock(obj, NULL);
40 		if (i915_gem_object_has_pinned_pages(obj))
41 			i915_gem_object_unpin_pages(obj);
42 		/* No polluting the memory region between tests */
43 		__i915_gem_object_put_pages(obj);
44 		i915_gem_object_unlock(obj);
45 		list_del(&obj->st_link);
46 		i915_gem_object_put(obj);
47 	}
48 
49 	cond_resched();
50 
51 	i915_gem_drain_freed_objects(i915);
52 }
53 
54 static int igt_mock_fill(void *arg)
55 {
56 	struct intel_memory_region *mem = arg;
57 	resource_size_t total = resource_size(&mem->region);
58 	resource_size_t page_size;
59 	resource_size_t rem;
60 	unsigned long max_pages;
61 	unsigned long page_num;
62 	LIST_HEAD(objects);
63 	int err = 0;
64 
65 	page_size = PAGE_SIZE;
66 	max_pages = div64_u64(total, page_size);
67 	rem = total;
68 
69 	for_each_prime_number_from(page_num, 1, max_pages) {
70 		resource_size_t size = page_num * page_size;
71 		struct drm_i915_gem_object *obj;
72 
73 		obj = i915_gem_object_create_region(mem, size, 0, 0);
74 		if (IS_ERR(obj)) {
75 			err = PTR_ERR(obj);
76 			break;
77 		}
78 
79 		err = i915_gem_object_pin_pages_unlocked(obj);
80 		if (err) {
81 			i915_gem_object_put(obj);
82 			break;
83 		}
84 
85 		list_add(&obj->st_link, &objects);
86 		rem -= size;
87 	}
88 
89 	if (err == -ENOMEM)
90 		err = 0;
91 	if (err == -ENXIO) {
92 		if (page_num * page_size <= rem) {
93 			pr_err("%s failed, space still left in region\n",
94 			       __func__);
95 			err = -EINVAL;
96 		} else {
97 			err = 0;
98 		}
99 	}
100 
101 	close_objects(mem, &objects);
102 
103 	return err;
104 }
105 
106 static struct drm_i915_gem_object *
107 igt_object_create(struct intel_memory_region *mem,
108 		  struct list_head *objects,
109 		  u64 size,
110 		  unsigned int flags)
111 {
112 	struct drm_i915_gem_object *obj;
113 	int err;
114 
115 	obj = i915_gem_object_create_region(mem, size, 0, flags);
116 	if (IS_ERR(obj))
117 		return obj;
118 
119 	err = i915_gem_object_pin_pages_unlocked(obj);
120 	if (err)
121 		goto put;
122 
123 	list_add(&obj->st_link, objects);
124 	return obj;
125 
126 put:
127 	i915_gem_object_put(obj);
128 	return ERR_PTR(err);
129 }
130 
131 static void igt_object_release(struct drm_i915_gem_object *obj)
132 {
133 	i915_gem_object_lock(obj, NULL);
134 	i915_gem_object_unpin_pages(obj);
135 	__i915_gem_object_put_pages(obj);
136 	i915_gem_object_unlock(obj);
137 	list_del(&obj->st_link);
138 	i915_gem_object_put(obj);
139 }
140 
141 static bool is_contiguous(struct drm_i915_gem_object *obj)
142 {
143 	struct scatterlist *sg;
144 	dma_addr_t addr = -1;
145 
146 	for (sg = obj->mm.pages->sgl; sg; sg = sg_next(sg)) {
147 		if (addr != -1 && sg_dma_address(sg) != addr)
148 			return false;
149 
150 		addr = sg_dma_address(sg) + sg_dma_len(sg);
151 	}
152 
153 	return true;
154 }
155 
156 static int igt_mock_reserve(void *arg)
157 {
158 	struct intel_memory_region *mem = arg;
159 	struct drm_i915_private *i915 = mem->i915;
160 	resource_size_t avail = resource_size(&mem->region);
161 	struct drm_i915_gem_object *obj;
162 	const u32 chunk_size = SZ_32M;
163 	u32 i, offset, count, *order;
164 	u64 allocated, cur_avail;
165 	I915_RND_STATE(prng);
166 	LIST_HEAD(objects);
167 	int err = 0;
168 
169 	count = avail / chunk_size;
170 	order = i915_random_order(count, &prng);
171 	if (!order)
172 		return 0;
173 
174 	mem = mock_region_create(i915, 0, SZ_2G, I915_GTT_PAGE_SIZE_4K, 0, 0);
175 	if (IS_ERR(mem)) {
176 		pr_err("failed to create memory region\n");
177 		err = PTR_ERR(mem);
178 		goto out_free_order;
179 	}
180 
181 	/* Reserve a bunch of ranges within the region */
182 	for (i = 0; i < count; ++i) {
183 		u64 start = order[i] * chunk_size;
184 		u64 size = i915_prandom_u32_max_state(chunk_size, &prng);
185 
186 		/* Allow for some really big holes */
187 		if (!size)
188 			continue;
189 
190 		size = round_up(size, PAGE_SIZE);
191 		offset = igt_random_offset(&prng, 0, chunk_size, size,
192 					   PAGE_SIZE);
193 
194 		err = intel_memory_region_reserve(mem, start + offset, size);
195 		if (err) {
196 			pr_err("%s failed to reserve range", __func__);
197 			goto out_close;
198 		}
199 
200 		/* XXX: maybe sanity check the block range here? */
201 		avail -= size;
202 	}
203 
204 	/* Try to see if we can allocate from the remaining space */
205 	allocated = 0;
206 	cur_avail = avail;
207 	do {
208 		u32 size = i915_prandom_u32_max_state(cur_avail, &prng);
209 
210 		size = max_t(u32, round_up(size, PAGE_SIZE), PAGE_SIZE);
211 		obj = igt_object_create(mem, &objects, size, 0);
212 		if (IS_ERR(obj)) {
213 			if (PTR_ERR(obj) == -ENXIO)
214 				break;
215 
216 			err = PTR_ERR(obj);
217 			goto out_close;
218 		}
219 		cur_avail -= size;
220 		allocated += size;
221 	} while (1);
222 
223 	if (allocated != avail) {
224 		pr_err("%s mismatch between allocation and free space", __func__);
225 		err = -EINVAL;
226 	}
227 
228 out_close:
229 	close_objects(mem, &objects);
230 	intel_memory_region_destroy(mem);
231 out_free_order:
232 	kfree(order);
233 	return err;
234 }
235 
236 static int igt_mock_contiguous(void *arg)
237 {
238 	struct intel_memory_region *mem = arg;
239 	struct drm_i915_gem_object *obj;
240 	unsigned long n_objects;
241 	LIST_HEAD(objects);
242 	LIST_HEAD(holes);
243 	I915_RND_STATE(prng);
244 	resource_size_t total;
245 	resource_size_t min;
246 	u64 target;
247 	int err = 0;
248 
249 	total = resource_size(&mem->region);
250 
251 	/* Min size */
252 	obj = igt_object_create(mem, &objects, PAGE_SIZE,
253 				I915_BO_ALLOC_CONTIGUOUS);
254 	if (IS_ERR(obj))
255 		return PTR_ERR(obj);
256 
257 	if (!is_contiguous(obj)) {
258 		pr_err("%s min object spans disjoint sg entries\n", __func__);
259 		err = -EINVAL;
260 		goto err_close_objects;
261 	}
262 
263 	igt_object_release(obj);
264 
265 	/* Max size */
266 	obj = igt_object_create(mem, &objects, total, I915_BO_ALLOC_CONTIGUOUS);
267 	if (IS_ERR(obj))
268 		return PTR_ERR(obj);
269 
270 	if (!is_contiguous(obj)) {
271 		pr_err("%s max object spans disjoint sg entries\n", __func__);
272 		err = -EINVAL;
273 		goto err_close_objects;
274 	}
275 
276 	igt_object_release(obj);
277 
278 	/* Internal fragmentation should not bleed into the object size */
279 	target = i915_prandom_u64_state(&prng);
280 	div64_u64_rem(target, total, &target);
281 	target = round_up(target, PAGE_SIZE);
282 	target = max_t(u64, PAGE_SIZE, target);
283 
284 	obj = igt_object_create(mem, &objects, target,
285 				I915_BO_ALLOC_CONTIGUOUS);
286 	if (IS_ERR(obj))
287 		return PTR_ERR(obj);
288 
289 	if (obj->base.size != target) {
290 		pr_err("%s obj->base.size(%zx) != target(%llx)\n", __func__,
291 		       obj->base.size, target);
292 		err = -EINVAL;
293 		goto err_close_objects;
294 	}
295 
296 	if (!is_contiguous(obj)) {
297 		pr_err("%s object spans disjoint sg entries\n", __func__);
298 		err = -EINVAL;
299 		goto err_close_objects;
300 	}
301 
302 	igt_object_release(obj);
303 
304 	/*
305 	 * Try to fragment the address space, such that half of it is free, but
306 	 * the max contiguous block size is SZ_64K.
307 	 */
308 
309 	target = SZ_64K;
310 	n_objects = div64_u64(total, target);
311 
312 	while (n_objects--) {
313 		struct list_head *list;
314 
315 		if (n_objects % 2)
316 			list = &holes;
317 		else
318 			list = &objects;
319 
320 		obj = igt_object_create(mem, list, target,
321 					I915_BO_ALLOC_CONTIGUOUS);
322 		if (IS_ERR(obj)) {
323 			err = PTR_ERR(obj);
324 			goto err_close_objects;
325 		}
326 	}
327 
328 	close_objects(mem, &holes);
329 
330 	min = target;
331 	target = total >> 1;
332 
333 	/* Make sure we can still allocate all the fragmented space */
334 	obj = igt_object_create(mem, &objects, target, 0);
335 	if (IS_ERR(obj)) {
336 		err = PTR_ERR(obj);
337 		goto err_close_objects;
338 	}
339 
340 	igt_object_release(obj);
341 
342 	/*
343 	 * Even though we have enough free space, we don't have a big enough
344 	 * contiguous block. Make sure that holds true.
345 	 */
346 
347 	do {
348 		bool should_fail = target > min;
349 
350 		obj = igt_object_create(mem, &objects, target,
351 					I915_BO_ALLOC_CONTIGUOUS);
352 		if (should_fail != IS_ERR(obj)) {
353 			pr_err("%s target allocation(%llx) mismatch\n",
354 			       __func__, target);
355 			err = -EINVAL;
356 			goto err_close_objects;
357 		}
358 
359 		target >>= 1;
360 	} while (target >= PAGE_SIZE);
361 
362 err_close_objects:
363 	list_splice_tail(&holes, &objects);
364 	close_objects(mem, &objects);
365 	return err;
366 }
367 
368 static int igt_mock_splintered_region(void *arg)
369 {
370 	struct intel_memory_region *mem = arg;
371 	struct drm_i915_private *i915 = mem->i915;
372 	struct i915_ttm_buddy_resource *res;
373 	struct drm_i915_gem_object *obj;
374 	struct drm_buddy *mm;
375 	unsigned int expected_order;
376 	LIST_HEAD(objects);
377 	u64 size;
378 	int err = 0;
379 
380 	/*
381 	 * Sanity check we can still allocate everything even if the
382 	 * mm.max_order != mm.size. i.e our starting address space size is not a
383 	 * power-of-two.
384 	 */
385 
386 	size = (SZ_4G - 1) & PAGE_MASK;
387 	mem = mock_region_create(i915, 0, size, PAGE_SIZE, 0, 0);
388 	if (IS_ERR(mem))
389 		return PTR_ERR(mem);
390 
391 	obj = igt_object_create(mem, &objects, size, 0);
392 	if (IS_ERR(obj)) {
393 		err = PTR_ERR(obj);
394 		goto out_close;
395 	}
396 
397 	res = to_ttm_buddy_resource(obj->mm.res);
398 	mm = res->mm;
399 	if (mm->size != size) {
400 		pr_err("%s size mismatch(%llu != %llu)\n",
401 		       __func__, mm->size, size);
402 		err = -EINVAL;
403 		goto out_put;
404 	}
405 
406 	expected_order = get_order(rounddown_pow_of_two(size));
407 	if (mm->max_order != expected_order) {
408 		pr_err("%s order mismatch(%u != %u)\n",
409 		       __func__, mm->max_order, expected_order);
410 		err = -EINVAL;
411 		goto out_put;
412 	}
413 
414 	close_objects(mem, &objects);
415 
416 	/*
417 	 * While we should be able allocate everything without any flag
418 	 * restrictions, if we consider I915_BO_ALLOC_CONTIGUOUS then we are
419 	 * actually limited to the largest power-of-two for the region size i.e
420 	 * max_order, due to the inner workings of the buddy allocator. So make
421 	 * sure that does indeed hold true.
422 	 */
423 
424 	obj = igt_object_create(mem, &objects, size, I915_BO_ALLOC_CONTIGUOUS);
425 	if (!IS_ERR(obj)) {
426 		pr_err("%s too large contiguous allocation was not rejected\n",
427 		       __func__);
428 		err = -EINVAL;
429 		goto out_close;
430 	}
431 
432 	obj = igt_object_create(mem, &objects, rounddown_pow_of_two(size),
433 				I915_BO_ALLOC_CONTIGUOUS);
434 	if (IS_ERR(obj)) {
435 		pr_err("%s largest possible contiguous allocation failed\n",
436 		       __func__);
437 		err = PTR_ERR(obj);
438 		goto out_close;
439 	}
440 
441 out_close:
442 	close_objects(mem, &objects);
443 out_put:
444 	intel_memory_region_destroy(mem);
445 	return err;
446 }
447 
448 #ifndef SZ_8G
449 #define SZ_8G BIT_ULL(33)
450 #endif
451 
452 static int igt_mock_max_segment(void *arg)
453 {
454 	const unsigned int max_segment = rounddown(UINT_MAX, PAGE_SIZE);
455 	struct intel_memory_region *mem = arg;
456 	struct drm_i915_private *i915 = mem->i915;
457 	struct i915_ttm_buddy_resource *res;
458 	struct drm_i915_gem_object *obj;
459 	struct drm_buddy_block *block;
460 	struct drm_buddy *mm;
461 	struct list_head *blocks;
462 	struct scatterlist *sg;
463 	LIST_HEAD(objects);
464 	u64 size;
465 	int err = 0;
466 
467 	/*
468 	 * While we may create very large contiguous blocks, we may need
469 	 * to break those down for consumption elsewhere. In particular,
470 	 * dma-mapping with scatterlist elements have an implicit limit of
471 	 * UINT_MAX on each element.
472 	 */
473 
474 	size = SZ_8G;
475 	mem = mock_region_create(i915, 0, size, PAGE_SIZE, 0, 0);
476 	if (IS_ERR(mem))
477 		return PTR_ERR(mem);
478 
479 	obj = igt_object_create(mem, &objects, size, 0);
480 	if (IS_ERR(obj)) {
481 		err = PTR_ERR(obj);
482 		goto out_put;
483 	}
484 
485 	res = to_ttm_buddy_resource(obj->mm.res);
486 	blocks = &res->blocks;
487 	mm = res->mm;
488 	size = 0;
489 	list_for_each_entry(block, blocks, link) {
490 		if (drm_buddy_block_size(mm, block) > size)
491 			size = drm_buddy_block_size(mm, block);
492 	}
493 	if (size < max_segment) {
494 		pr_err("%s: Failed to create a huge contiguous block [> %u], largest block %lld\n",
495 		       __func__, max_segment, size);
496 		err = -EINVAL;
497 		goto out_close;
498 	}
499 
500 	for (sg = obj->mm.pages->sgl; sg; sg = sg_next(sg)) {
501 		if (sg->length > max_segment) {
502 			pr_err("%s: Created an oversized scatterlist entry, %u > %u\n",
503 			       __func__, sg->length, max_segment);
504 			err = -EINVAL;
505 			goto out_close;
506 		}
507 	}
508 
509 out_close:
510 	close_objects(mem, &objects);
511 out_put:
512 	intel_memory_region_destroy(mem);
513 	return err;
514 }
515 
516 static u64 igt_object_mappable_total(struct drm_i915_gem_object *obj)
517 {
518 	struct intel_memory_region *mr = obj->mm.region;
519 	struct i915_ttm_buddy_resource *bman_res =
520 		to_ttm_buddy_resource(obj->mm.res);
521 	struct drm_buddy *mm = bman_res->mm;
522 	struct drm_buddy_block *block;
523 	u64 total;
524 
525 	total = 0;
526 	list_for_each_entry(block, &bman_res->blocks, link) {
527 		u64 start = drm_buddy_block_offset(block);
528 		u64 end = start + drm_buddy_block_size(mm, block);
529 
530 		if (start < mr->io_size)
531 			total += min_t(u64, end, mr->io_size) - start;
532 	}
533 
534 	return total;
535 }
536 
537 static int igt_mock_io_size(void *arg)
538 {
539 	struct intel_memory_region *mr = arg;
540 	struct drm_i915_private *i915 = mr->i915;
541 	struct drm_i915_gem_object *obj;
542 	u64 mappable_theft_total;
543 	u64 io_size;
544 	u64 total;
545 	u64 ps;
546 	u64 rem;
547 	u64 size;
548 	I915_RND_STATE(prng);
549 	LIST_HEAD(objects);
550 	int err = 0;
551 
552 	ps = SZ_4K;
553 	if (i915_prandom_u64_state(&prng) & 1)
554 		ps = SZ_64K; /* For something like DG2 */
555 
556 	div64_u64_rem(i915_prandom_u64_state(&prng), SZ_8G, &total);
557 	total = round_down(total, ps);
558 	total = max_t(u64, total, SZ_1G);
559 
560 	div64_u64_rem(i915_prandom_u64_state(&prng), total - ps, &io_size);
561 	io_size = round_down(io_size, ps);
562 	io_size = max_t(u64, io_size, SZ_256M); /* 256M seems to be the common lower limit */
563 
564 	pr_info("%s with ps=%llx, io_size=%llx, total=%llx\n",
565 		__func__, ps, io_size, total);
566 
567 	mr = mock_region_create(i915, 0, total, ps, 0, io_size);
568 	if (IS_ERR(mr)) {
569 		err = PTR_ERR(mr);
570 		goto out_err;
571 	}
572 
573 	mappable_theft_total = 0;
574 	rem = total - io_size;
575 	do {
576 		div64_u64_rem(i915_prandom_u64_state(&prng), rem, &size);
577 		size = round_down(size, ps);
578 		size = max(size, ps);
579 
580 		obj = igt_object_create(mr, &objects, size,
581 					I915_BO_ALLOC_GPU_ONLY);
582 		if (IS_ERR(obj)) {
583 			pr_err("%s TOPDOWN failed with rem=%llx, size=%llx\n",
584 			       __func__, rem, size);
585 			err = PTR_ERR(obj);
586 			goto out_close;
587 		}
588 
589 		mappable_theft_total += igt_object_mappable_total(obj);
590 		rem -= size;
591 	} while (rem);
592 
593 	pr_info("%s mappable theft=(%lluMiB/%lluMiB), total=%lluMiB\n",
594 		__func__,
595 		(u64)mappable_theft_total >> 20,
596 		(u64)io_size >> 20,
597 		(u64)total >> 20);
598 
599 	/*
600 	 * Even if we allocate all of the non-mappable portion, we should still
601 	 * be able to dip into the mappable portion.
602 	 */
603 	obj = igt_object_create(mr, &objects, io_size,
604 				I915_BO_ALLOC_GPU_ONLY);
605 	if (IS_ERR(obj)) {
606 		pr_err("%s allocation unexpectedly failed\n", __func__);
607 		err = PTR_ERR(obj);
608 		goto out_close;
609 	}
610 
611 	close_objects(mr, &objects);
612 
613 	rem = io_size;
614 	do {
615 		div64_u64_rem(i915_prandom_u64_state(&prng), rem, &size);
616 		size = round_down(size, ps);
617 		size = max(size, ps);
618 
619 		obj = igt_object_create(mr, &objects, size, 0);
620 		if (IS_ERR(obj)) {
621 			pr_err("%s MAPPABLE failed with rem=%llx, size=%llx\n",
622 			       __func__, rem, size);
623 			err = PTR_ERR(obj);
624 			goto out_close;
625 		}
626 
627 		if (igt_object_mappable_total(obj) != size) {
628 			pr_err("%s allocation is not mappable(size=%llx)\n",
629 			       __func__, size);
630 			err = -EINVAL;
631 			goto out_close;
632 		}
633 		rem -= size;
634 	} while (rem);
635 
636 	/*
637 	 * We assume CPU access is required by default, which should result in a
638 	 * failure here, even though the non-mappable portion is free.
639 	 */
640 	obj = igt_object_create(mr, &objects, ps, 0);
641 	if (!IS_ERR(obj)) {
642 		pr_err("%s allocation unexpectedly succeeded\n", __func__);
643 		err = -EINVAL;
644 		goto out_close;
645 	}
646 
647 out_close:
648 	close_objects(mr, &objects);
649 	intel_memory_region_destroy(mr);
650 out_err:
651 	if (err == -ENOMEM)
652 		err = 0;
653 
654 	return err;
655 }
656 
657 static int igt_gpu_write_dw(struct intel_context *ce,
658 			    struct i915_vma *vma,
659 			    u32 dword,
660 			    u32 value)
661 {
662 	return igt_gpu_fill_dw(ce, vma, dword * sizeof(u32),
663 			       vma->size >> PAGE_SHIFT, value);
664 }
665 
666 static int igt_cpu_check(struct drm_i915_gem_object *obj, u32 dword, u32 val)
667 {
668 	unsigned long n = obj->base.size >> PAGE_SHIFT;
669 	u32 *ptr;
670 	int err;
671 
672 	err = i915_gem_object_wait(obj, 0, MAX_SCHEDULE_TIMEOUT);
673 	if (err)
674 		return err;
675 
676 	ptr = i915_gem_object_pin_map(obj, I915_MAP_WC);
677 	if (IS_ERR(ptr))
678 		return PTR_ERR(ptr);
679 
680 	ptr += dword;
681 	while (n--) {
682 		if (*ptr != val) {
683 			pr_err("base[%u]=%08x, val=%08x\n",
684 			       dword, *ptr, val);
685 			err = -EINVAL;
686 			break;
687 		}
688 
689 		ptr += PAGE_SIZE / sizeof(*ptr);
690 	}
691 
692 	i915_gem_object_unpin_map(obj);
693 	return err;
694 }
695 
696 static int igt_gpu_write(struct i915_gem_context *ctx,
697 			 struct drm_i915_gem_object *obj)
698 {
699 	struct i915_gem_engines *engines;
700 	struct i915_gem_engines_iter it;
701 	struct i915_address_space *vm;
702 	struct intel_context *ce;
703 	I915_RND_STATE(prng);
704 	IGT_TIMEOUT(end_time);
705 	unsigned int count;
706 	struct i915_vma *vma;
707 	int *order;
708 	int i, n;
709 	int err = 0;
710 
711 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
712 
713 	n = 0;
714 	count = 0;
715 	for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
716 		count++;
717 		if (!intel_engine_can_store_dword(ce->engine))
718 			continue;
719 
720 		vm = ce->vm;
721 		n++;
722 	}
723 	i915_gem_context_unlock_engines(ctx);
724 	if (!n)
725 		return 0;
726 
727 	order = i915_random_order(count * count, &prng);
728 	if (!order)
729 		return -ENOMEM;
730 
731 	vma = i915_vma_instance(obj, vm, NULL);
732 	if (IS_ERR(vma)) {
733 		err = PTR_ERR(vma);
734 		goto out_free;
735 	}
736 
737 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
738 	if (err)
739 		goto out_free;
740 
741 	i = 0;
742 	engines = i915_gem_context_lock_engines(ctx);
743 	do {
744 		u32 rng = prandom_u32_state(&prng);
745 		u32 dword = offset_in_page(rng) / 4;
746 
747 		ce = engines->engines[order[i] % engines->num_engines];
748 		i = (i + 1) % (count * count);
749 		if (!ce || !intel_engine_can_store_dword(ce->engine))
750 			continue;
751 
752 		err = igt_gpu_write_dw(ce, vma, dword, rng);
753 		if (err)
754 			break;
755 
756 		i915_gem_object_lock(obj, NULL);
757 		err = igt_cpu_check(obj, dword, rng);
758 		i915_gem_object_unlock(obj);
759 		if (err)
760 			break;
761 	} while (!__igt_timeout(end_time, NULL));
762 	i915_gem_context_unlock_engines(ctx);
763 
764 out_free:
765 	kfree(order);
766 
767 	if (err == -ENOMEM)
768 		err = 0;
769 
770 	return err;
771 }
772 
773 static int igt_lmem_create(void *arg)
774 {
775 	struct drm_i915_private *i915 = arg;
776 	struct drm_i915_gem_object *obj;
777 	int err = 0;
778 
779 	obj = i915_gem_object_create_lmem(i915, PAGE_SIZE, 0);
780 	if (IS_ERR(obj))
781 		return PTR_ERR(obj);
782 
783 	err = i915_gem_object_pin_pages_unlocked(obj);
784 	if (err)
785 		goto out_put;
786 
787 	i915_gem_object_unpin_pages(obj);
788 out_put:
789 	i915_gem_object_put(obj);
790 
791 	return err;
792 }
793 
794 static int igt_lmem_create_with_ps(void *arg)
795 {
796 	struct drm_i915_private *i915 = arg;
797 	int err = 0;
798 	u32 ps;
799 
800 	for (ps = PAGE_SIZE; ps <= SZ_1G; ps <<= 1) {
801 		struct drm_i915_gem_object *obj;
802 		dma_addr_t daddr;
803 
804 		obj = __i915_gem_object_create_lmem_with_ps(i915, ps, ps, 0);
805 		if (IS_ERR(obj)) {
806 			err = PTR_ERR(obj);
807 			if (err == -ENXIO || err == -E2BIG) {
808 				pr_info("%s not enough lmem for ps(%u) err=%d\n",
809 					__func__, ps, err);
810 				err = 0;
811 			}
812 
813 			break;
814 		}
815 
816 		if (obj->base.size != ps) {
817 			pr_err("%s size(%zu) != ps(%u)\n",
818 			       __func__, obj->base.size, ps);
819 			err = -EINVAL;
820 			goto out_put;
821 		}
822 
823 		i915_gem_object_lock(obj, NULL);
824 		err = i915_gem_object_pin_pages(obj);
825 		if (err) {
826 			if (err == -ENXIO || err == -E2BIG || err == -ENOMEM) {
827 				pr_info("%s not enough lmem for ps(%u) err=%d\n",
828 					__func__, ps, err);
829 				err = 0;
830 			}
831 			goto out_put;
832 		}
833 
834 		daddr = i915_gem_object_get_dma_address(obj, 0);
835 		if (!IS_ALIGNED(daddr, ps)) {
836 			pr_err("%s daddr(%pa) not aligned with ps(%u)\n",
837 			       __func__, &daddr, ps);
838 			err = -EINVAL;
839 			goto out_unpin;
840 		}
841 
842 out_unpin:
843 		i915_gem_object_unpin_pages(obj);
844 		__i915_gem_object_put_pages(obj);
845 out_put:
846 		i915_gem_object_unlock(obj);
847 		i915_gem_object_put(obj);
848 
849 		if (err)
850 			break;
851 	}
852 
853 	return err;
854 }
855 
856 static int igt_lmem_create_cleared_cpu(void *arg)
857 {
858 	struct drm_i915_private *i915 = arg;
859 	I915_RND_STATE(prng);
860 	IGT_TIMEOUT(end_time);
861 	u32 size, i;
862 	int err;
863 
864 	i915_gem_drain_freed_objects(i915);
865 
866 	size = max_t(u32, PAGE_SIZE, i915_prandom_u32_max_state(SZ_32M, &prng));
867 	size = round_up(size, PAGE_SIZE);
868 	i = 0;
869 
870 	do {
871 		struct drm_i915_gem_object *obj;
872 		unsigned int flags;
873 		u32 dword, val;
874 		void *vaddr;
875 
876 		/*
877 		 * Alternate between cleared and uncleared allocations, while
878 		 * also dirtying the pages each time to check that the pages are
879 		 * always cleared if requested, since we should get some overlap
880 		 * of the underlying pages, if not all, since we are the only
881 		 * user.
882 		 */
883 
884 		flags = I915_BO_ALLOC_CPU_CLEAR;
885 		if (i & 1)
886 			flags = 0;
887 
888 		obj = i915_gem_object_create_lmem(i915, size, flags);
889 		if (IS_ERR(obj))
890 			return PTR_ERR(obj);
891 
892 		i915_gem_object_lock(obj, NULL);
893 		err = i915_gem_object_pin_pages(obj);
894 		if (err)
895 			goto out_put;
896 
897 		dword = i915_prandom_u32_max_state(PAGE_SIZE / sizeof(u32),
898 						   &prng);
899 
900 		if (flags & I915_BO_ALLOC_CPU_CLEAR) {
901 			err = igt_cpu_check(obj, dword, 0);
902 			if (err) {
903 				pr_err("%s failed with size=%u, flags=%u\n",
904 				       __func__, size, flags);
905 				goto out_unpin;
906 			}
907 		}
908 
909 		vaddr = i915_gem_object_pin_map(obj, I915_MAP_WC);
910 		if (IS_ERR(vaddr)) {
911 			err = PTR_ERR(vaddr);
912 			goto out_unpin;
913 		}
914 
915 		val = prandom_u32_state(&prng);
916 
917 		memset32(vaddr, val, obj->base.size / sizeof(u32));
918 
919 		i915_gem_object_flush_map(obj);
920 		i915_gem_object_unpin_map(obj);
921 out_unpin:
922 		i915_gem_object_unpin_pages(obj);
923 		__i915_gem_object_put_pages(obj);
924 out_put:
925 		i915_gem_object_unlock(obj);
926 		i915_gem_object_put(obj);
927 
928 		if (err)
929 			break;
930 		++i;
931 	} while (!__igt_timeout(end_time, NULL));
932 
933 	pr_info("%s completed (%u) iterations\n", __func__, i);
934 
935 	return err;
936 }
937 
938 static int igt_lmem_write_gpu(void *arg)
939 {
940 	struct drm_i915_private *i915 = arg;
941 	struct drm_i915_gem_object *obj;
942 	struct i915_gem_context *ctx;
943 	struct file *file;
944 	I915_RND_STATE(prng);
945 	u32 sz;
946 	int err;
947 
948 	file = mock_file(i915);
949 	if (IS_ERR(file))
950 		return PTR_ERR(file);
951 
952 	ctx = live_context(i915, file);
953 	if (IS_ERR(ctx)) {
954 		err = PTR_ERR(ctx);
955 		goto out_file;
956 	}
957 
958 	sz = round_up(prandom_u32_state(&prng) % SZ_32M, PAGE_SIZE);
959 
960 	obj = i915_gem_object_create_lmem(i915, sz, 0);
961 	if (IS_ERR(obj)) {
962 		err = PTR_ERR(obj);
963 		goto out_file;
964 	}
965 
966 	err = i915_gem_object_pin_pages_unlocked(obj);
967 	if (err)
968 		goto out_put;
969 
970 	err = igt_gpu_write(ctx, obj);
971 	if (err)
972 		pr_err("igt_gpu_write failed(%d)\n", err);
973 
974 	i915_gem_object_unpin_pages(obj);
975 out_put:
976 	i915_gem_object_put(obj);
977 out_file:
978 	fput(file);
979 	return err;
980 }
981 
982 static struct intel_engine_cs *
983 random_engine_class(struct drm_i915_private *i915,
984 		    unsigned int class,
985 		    struct rnd_state *prng)
986 {
987 	struct intel_engine_cs *engine;
988 	unsigned int count;
989 
990 	count = 0;
991 	for (engine = intel_engine_lookup_user(i915, class, 0);
992 	     engine && engine->uabi_class == class;
993 	     engine = rb_entry_safe(rb_next(&engine->uabi_node),
994 				    typeof(*engine), uabi_node))
995 		count++;
996 
997 	count = i915_prandom_u32_max_state(count, prng);
998 	return intel_engine_lookup_user(i915, class, count);
999 }
1000 
1001 static int igt_lmem_write_cpu(void *arg)
1002 {
1003 	struct drm_i915_private *i915 = arg;
1004 	struct drm_i915_gem_object *obj;
1005 	I915_RND_STATE(prng);
1006 	IGT_TIMEOUT(end_time);
1007 	u32 bytes[] = {
1008 		0, /* rng placeholder */
1009 		sizeof(u32),
1010 		sizeof(u64),
1011 		64, /* cl */
1012 		PAGE_SIZE,
1013 		PAGE_SIZE - sizeof(u32),
1014 		PAGE_SIZE - sizeof(u64),
1015 		PAGE_SIZE - 64,
1016 	};
1017 	struct intel_engine_cs *engine;
1018 	struct i915_request *rq;
1019 	u32 *vaddr;
1020 	u32 sz;
1021 	u32 i;
1022 	int *order;
1023 	int count;
1024 	int err;
1025 
1026 	engine = random_engine_class(i915, I915_ENGINE_CLASS_COPY, &prng);
1027 	if (!engine)
1028 		return 0;
1029 
1030 	pr_info("%s: using %s\n", __func__, engine->name);
1031 
1032 	sz = round_up(prandom_u32_state(&prng) % SZ_32M, PAGE_SIZE);
1033 	sz = max_t(u32, 2 * PAGE_SIZE, sz);
1034 
1035 	obj = i915_gem_object_create_lmem(i915, sz, I915_BO_ALLOC_CONTIGUOUS);
1036 	if (IS_ERR(obj))
1037 		return PTR_ERR(obj);
1038 
1039 	vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1040 	if (IS_ERR(vaddr)) {
1041 		err = PTR_ERR(vaddr);
1042 		goto out_put;
1043 	}
1044 
1045 	i915_gem_object_lock(obj, NULL);
1046 
1047 	err = dma_resv_reserve_fences(obj->base.resv, 1);
1048 	if (err) {
1049 		i915_gem_object_unlock(obj);
1050 		goto out_put;
1051 	}
1052 
1053 	/* Put the pages into a known state -- from the gpu for added fun */
1054 	intel_engine_pm_get(engine);
1055 	err = intel_context_migrate_clear(engine->gt->migrate.context, NULL,
1056 					  obj->mm.pages->sgl, I915_CACHE_NONE,
1057 					  true, 0xdeadbeaf, &rq);
1058 	if (rq) {
1059 		dma_resv_add_fence(obj->base.resv, &rq->fence,
1060 				   DMA_RESV_USAGE_WRITE);
1061 		i915_request_put(rq);
1062 	}
1063 
1064 	intel_engine_pm_put(engine);
1065 	if (!err)
1066 		err = i915_gem_object_set_to_wc_domain(obj, true);
1067 	i915_gem_object_unlock(obj);
1068 	if (err)
1069 		goto out_unpin;
1070 
1071 	count = ARRAY_SIZE(bytes);
1072 	order = i915_random_order(count * count, &prng);
1073 	if (!order) {
1074 		err = -ENOMEM;
1075 		goto out_unpin;
1076 	}
1077 
1078 	/* A random multiple of u32, picked between [64, PAGE_SIZE - 64] */
1079 	bytes[0] = igt_random_offset(&prng, 64, PAGE_SIZE - 64, 0, sizeof(u32));
1080 	GEM_BUG_ON(!IS_ALIGNED(bytes[0], sizeof(u32)));
1081 
1082 	i = 0;
1083 	do {
1084 		u32 offset;
1085 		u32 align;
1086 		u32 dword;
1087 		u32 size;
1088 		u32 val;
1089 
1090 		size = bytes[order[i] % count];
1091 		i = (i + 1) % (count * count);
1092 
1093 		align = bytes[order[i] % count];
1094 		i = (i + 1) % (count * count);
1095 
1096 		align = max_t(u32, sizeof(u32), rounddown_pow_of_two(align));
1097 
1098 		offset = igt_random_offset(&prng, 0, obj->base.size,
1099 					   size, align);
1100 
1101 		val = prandom_u32_state(&prng);
1102 		memset32(vaddr + offset / sizeof(u32), val ^ 0xdeadbeaf,
1103 			 size / sizeof(u32));
1104 
1105 		/*
1106 		 * Sample random dw -- don't waste precious time reading every
1107 		 * single dw.
1108 		 */
1109 		dword = igt_random_offset(&prng, offset,
1110 					  offset + size,
1111 					  sizeof(u32), sizeof(u32));
1112 		dword /= sizeof(u32);
1113 		if (vaddr[dword] != (val ^ 0xdeadbeaf)) {
1114 			pr_err("%s vaddr[%u]=%u, val=%u, size=%u, align=%u, offset=%u\n",
1115 			       __func__, dword, vaddr[dword], val ^ 0xdeadbeaf,
1116 			       size, align, offset);
1117 			err = -EINVAL;
1118 			break;
1119 		}
1120 	} while (!__igt_timeout(end_time, NULL));
1121 
1122 out_unpin:
1123 	i915_gem_object_unpin_map(obj);
1124 out_put:
1125 	i915_gem_object_put(obj);
1126 
1127 	return err;
1128 }
1129 
1130 static const char *repr_type(u32 type)
1131 {
1132 	switch (type) {
1133 	case I915_MAP_WB:
1134 		return "WB";
1135 	case I915_MAP_WC:
1136 		return "WC";
1137 	}
1138 
1139 	return "";
1140 }
1141 
1142 static struct drm_i915_gem_object *
1143 create_region_for_mapping(struct intel_memory_region *mr, u64 size, u32 type,
1144 			  void **out_addr)
1145 {
1146 	struct drm_i915_gem_object *obj;
1147 	void *addr;
1148 
1149 	obj = i915_gem_object_create_region(mr, size, 0, 0);
1150 	if (IS_ERR(obj)) {
1151 		if (PTR_ERR(obj) == -ENOSPC) /* Stolen memory */
1152 			return ERR_PTR(-ENODEV);
1153 		return obj;
1154 	}
1155 
1156 	addr = i915_gem_object_pin_map_unlocked(obj, type);
1157 	if (IS_ERR(addr)) {
1158 		i915_gem_object_put(obj);
1159 		if (PTR_ERR(addr) == -ENXIO)
1160 			return ERR_PTR(-ENODEV);
1161 		return addr;
1162 	}
1163 
1164 	*out_addr = addr;
1165 	return obj;
1166 }
1167 
1168 static int wrap_ktime_compare(const void *A, const void *B)
1169 {
1170 	const ktime_t *a = A, *b = B;
1171 
1172 	return ktime_compare(*a, *b);
1173 }
1174 
1175 static void igt_memcpy_long(void *dst, const void *src, size_t size)
1176 {
1177 	unsigned long *tmp = dst;
1178 	const unsigned long *s = src;
1179 
1180 	size = size / sizeof(unsigned long);
1181 	while (size--)
1182 		*tmp++ = *s++;
1183 }
1184 
1185 static inline void igt_memcpy(void *dst, const void *src, size_t size)
1186 {
1187 	memcpy(dst, src, size);
1188 }
1189 
1190 static inline void igt_memcpy_from_wc(void *dst, const void *src, size_t size)
1191 {
1192 	i915_memcpy_from_wc(dst, src, size);
1193 }
1194 
1195 static int _perf_memcpy(struct intel_memory_region *src_mr,
1196 			struct intel_memory_region *dst_mr,
1197 			u64 size, u32 src_type, u32 dst_type)
1198 {
1199 	struct drm_i915_private *i915 = src_mr->i915;
1200 	const struct {
1201 		const char *name;
1202 		void (*copy)(void *dst, const void *src, size_t size);
1203 		bool skip;
1204 	} tests[] = {
1205 		{
1206 			"memcpy",
1207 			igt_memcpy,
1208 		},
1209 		{
1210 			"memcpy_long",
1211 			igt_memcpy_long,
1212 		},
1213 		{
1214 			"memcpy_from_wc",
1215 			igt_memcpy_from_wc,
1216 			!i915_has_memcpy_from_wc(),
1217 		},
1218 	};
1219 	struct drm_i915_gem_object *src, *dst;
1220 	void *src_addr, *dst_addr;
1221 	int ret = 0;
1222 	int i;
1223 
1224 	src = create_region_for_mapping(src_mr, size, src_type, &src_addr);
1225 	if (IS_ERR(src)) {
1226 		ret = PTR_ERR(src);
1227 		goto out;
1228 	}
1229 
1230 	dst = create_region_for_mapping(dst_mr, size, dst_type, &dst_addr);
1231 	if (IS_ERR(dst)) {
1232 		ret = PTR_ERR(dst);
1233 		goto out_unpin_src;
1234 	}
1235 
1236 	for (i = 0; i < ARRAY_SIZE(tests); ++i) {
1237 		ktime_t t[5];
1238 		int pass;
1239 
1240 		if (tests[i].skip)
1241 			continue;
1242 
1243 		for (pass = 0; pass < ARRAY_SIZE(t); pass++) {
1244 			ktime_t t0, t1;
1245 
1246 			t0 = ktime_get();
1247 
1248 			tests[i].copy(dst_addr, src_addr, size);
1249 
1250 			t1 = ktime_get();
1251 			t[pass] = ktime_sub(t1, t0);
1252 		}
1253 
1254 		sort(t, ARRAY_SIZE(t), sizeof(*t), wrap_ktime_compare, NULL);
1255 		if (t[0] <= 0) {
1256 			/* ignore the impossible to protect our sanity */
1257 			pr_debug("Skipping %s src(%s, %s) -> dst(%s, %s) %14s %4lluKiB copy, unstable measurement [%lld, %lld]\n",
1258 				 __func__,
1259 				 src_mr->name, repr_type(src_type),
1260 				 dst_mr->name, repr_type(dst_type),
1261 				 tests[i].name, size >> 10,
1262 				 t[0], t[4]);
1263 			continue;
1264 		}
1265 
1266 		pr_info("%s src(%s, %s) -> dst(%s, %s) %14s %4llu KiB copy: %5lld MiB/s\n",
1267 			__func__,
1268 			src_mr->name, repr_type(src_type),
1269 			dst_mr->name, repr_type(dst_type),
1270 			tests[i].name, size >> 10,
1271 			div64_u64(mul_u32_u32(4 * size,
1272 					      1000 * 1000 * 1000),
1273 				  t[1] + 2 * t[2] + t[3]) >> 20);
1274 
1275 		cond_resched();
1276 	}
1277 
1278 	i915_gem_object_unpin_map(dst);
1279 	i915_gem_object_put(dst);
1280 out_unpin_src:
1281 	i915_gem_object_unpin_map(src);
1282 	i915_gem_object_put(src);
1283 
1284 	i915_gem_drain_freed_objects(i915);
1285 out:
1286 	if (ret == -ENODEV)
1287 		ret = 0;
1288 
1289 	return ret;
1290 }
1291 
1292 static int perf_memcpy(void *arg)
1293 {
1294 	struct drm_i915_private *i915 = arg;
1295 	static const u32 types[] = {
1296 		I915_MAP_WB,
1297 		I915_MAP_WC,
1298 	};
1299 	static const u32 sizes[] = {
1300 		SZ_4K,
1301 		SZ_64K,
1302 		SZ_4M,
1303 	};
1304 	struct intel_memory_region *src_mr, *dst_mr;
1305 	int src_id, dst_id;
1306 	int i, j, k;
1307 	int ret;
1308 
1309 	for_each_memory_region(src_mr, i915, src_id) {
1310 		for_each_memory_region(dst_mr, i915, dst_id) {
1311 			for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
1312 				for (j = 0; j < ARRAY_SIZE(types); ++j) {
1313 					for (k = 0; k < ARRAY_SIZE(types); ++k) {
1314 						ret = _perf_memcpy(src_mr,
1315 								   dst_mr,
1316 								   sizes[i],
1317 								   types[j],
1318 								   types[k]);
1319 						if (ret)
1320 							return ret;
1321 					}
1322 				}
1323 			}
1324 		}
1325 	}
1326 
1327 	return 0;
1328 }
1329 
1330 int intel_memory_region_mock_selftests(void)
1331 {
1332 	static const struct i915_subtest tests[] = {
1333 		SUBTEST(igt_mock_reserve),
1334 		SUBTEST(igt_mock_fill),
1335 		SUBTEST(igt_mock_contiguous),
1336 		SUBTEST(igt_mock_splintered_region),
1337 		SUBTEST(igt_mock_max_segment),
1338 		SUBTEST(igt_mock_io_size),
1339 	};
1340 	struct intel_memory_region *mem;
1341 	struct drm_i915_private *i915;
1342 	int err;
1343 
1344 	i915 = mock_gem_device();
1345 	if (!i915)
1346 		return -ENOMEM;
1347 
1348 	mem = mock_region_create(i915, 0, SZ_2G, I915_GTT_PAGE_SIZE_4K, 0, 0);
1349 	if (IS_ERR(mem)) {
1350 		pr_err("failed to create memory region\n");
1351 		err = PTR_ERR(mem);
1352 		goto out_unref;
1353 	}
1354 
1355 	err = i915_subtests(tests, mem);
1356 
1357 	intel_memory_region_destroy(mem);
1358 out_unref:
1359 	mock_destroy_device(i915);
1360 	return err;
1361 }
1362 
1363 int intel_memory_region_live_selftests(struct drm_i915_private *i915)
1364 {
1365 	static const struct i915_subtest tests[] = {
1366 		SUBTEST(igt_lmem_create),
1367 		SUBTEST(igt_lmem_create_with_ps),
1368 		SUBTEST(igt_lmem_create_cleared_cpu),
1369 		SUBTEST(igt_lmem_write_cpu),
1370 		SUBTEST(igt_lmem_write_gpu),
1371 	};
1372 
1373 	if (!HAS_LMEM(i915)) {
1374 		pr_info("device lacks LMEM support, skipping\n");
1375 		return 0;
1376 	}
1377 
1378 	if (intel_gt_is_wedged(to_gt(i915)))
1379 		return 0;
1380 
1381 	return i915_live_subtests(tests, i915);
1382 }
1383 
1384 int intel_memory_region_perf_selftests(struct drm_i915_private *i915)
1385 {
1386 	static const struct i915_subtest tests[] = {
1387 		SUBTEST(perf_memcpy),
1388 	};
1389 
1390 	if (intel_gt_is_wedged(to_gt(i915)))
1391 		return 0;
1392 
1393 	return i915_live_subtests(tests, i915);
1394 }
1395