1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2017 Intel Corporation
5  */
6 
7 #include <linux/prime_numbers.h>
8 
9 #include "i915_selftest.h"
10 
11 #include "gem/i915_gem_region.h"
12 #include "gem/i915_gem_lmem.h"
13 #include "gem/i915_gem_pm.h"
14 
15 #include "gt/intel_gt.h"
16 
17 #include "igt_gem_utils.h"
18 #include "mock_context.h"
19 
20 #include "selftests/mock_drm.h"
21 #include "selftests/mock_gem_device.h"
22 #include "selftests/mock_region.h"
23 #include "selftests/i915_random.h"
24 
25 static const unsigned int page_sizes[] = {
26 	I915_GTT_PAGE_SIZE_2M,
27 	I915_GTT_PAGE_SIZE_64K,
28 	I915_GTT_PAGE_SIZE_4K,
29 };
30 
31 static unsigned int get_largest_page_size(struct drm_i915_private *i915,
32 					  u64 rem)
33 {
34 	int i;
35 
36 	for (i = 0; i < ARRAY_SIZE(page_sizes); ++i) {
37 		unsigned int page_size = page_sizes[i];
38 
39 		if (HAS_PAGE_SIZES(i915, page_size) && rem >= page_size)
40 			return page_size;
41 	}
42 
43 	return 0;
44 }
45 
46 static void huge_pages_free_pages(struct sg_table *st)
47 {
48 	struct scatterlist *sg;
49 
50 	for (sg = st->sgl; sg; sg = __sg_next(sg)) {
51 		if (sg_page(sg))
52 			__free_pages(sg_page(sg), get_order(sg->length));
53 	}
54 
55 	sg_free_table(st);
56 	kfree(st);
57 }
58 
59 static int get_huge_pages(struct drm_i915_gem_object *obj)
60 {
61 #define GFP (GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY)
62 	unsigned int page_mask = obj->mm.page_mask;
63 	struct sg_table *st;
64 	struct scatterlist *sg;
65 	unsigned int sg_page_sizes;
66 	u64 rem;
67 
68 	st = kmalloc(sizeof(*st), GFP);
69 	if (!st)
70 		return -ENOMEM;
71 
72 	if (sg_alloc_table(st, obj->base.size >> PAGE_SHIFT, GFP)) {
73 		kfree(st);
74 		return -ENOMEM;
75 	}
76 
77 	rem = obj->base.size;
78 	sg = st->sgl;
79 	st->nents = 0;
80 	sg_page_sizes = 0;
81 
82 	/*
83 	 * Our goal here is simple, we want to greedily fill the object from
84 	 * largest to smallest page-size, while ensuring that we use *every*
85 	 * page-size as per the given page-mask.
86 	 */
87 	do {
88 		unsigned int bit = ilog2(page_mask);
89 		unsigned int page_size = BIT(bit);
90 		int order = get_order(page_size);
91 
92 		do {
93 			struct page *page;
94 
95 			GEM_BUG_ON(order >= MAX_ORDER);
96 			page = alloc_pages(GFP | __GFP_ZERO, order);
97 			if (!page)
98 				goto err;
99 
100 			sg_set_page(sg, page, page_size, 0);
101 			sg_page_sizes |= page_size;
102 			st->nents++;
103 
104 			rem -= page_size;
105 			if (!rem) {
106 				sg_mark_end(sg);
107 				break;
108 			}
109 
110 			sg = __sg_next(sg);
111 		} while ((rem - ((page_size-1) & page_mask)) >= page_size);
112 
113 		page_mask &= (page_size-1);
114 	} while (page_mask);
115 
116 	if (i915_gem_gtt_prepare_pages(obj, st))
117 		goto err;
118 
119 	GEM_BUG_ON(sg_page_sizes != obj->mm.page_mask);
120 	__i915_gem_object_set_pages(obj, st, sg_page_sizes);
121 
122 	return 0;
123 
124 err:
125 	sg_set_page(sg, NULL, 0, 0);
126 	sg_mark_end(sg);
127 	huge_pages_free_pages(st);
128 
129 	return -ENOMEM;
130 }
131 
132 static void put_huge_pages(struct drm_i915_gem_object *obj,
133 			   struct sg_table *pages)
134 {
135 	i915_gem_gtt_finish_pages(obj, pages);
136 	huge_pages_free_pages(pages);
137 
138 	obj->mm.dirty = false;
139 }
140 
141 static const struct drm_i915_gem_object_ops huge_page_ops = {
142 	.flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
143 		 I915_GEM_OBJECT_IS_SHRINKABLE,
144 	.get_pages = get_huge_pages,
145 	.put_pages = put_huge_pages,
146 };
147 
148 static struct drm_i915_gem_object *
149 huge_pages_object(struct drm_i915_private *i915,
150 		  u64 size,
151 		  unsigned int page_mask)
152 {
153 	static struct lock_class_key lock_class;
154 	struct drm_i915_gem_object *obj;
155 
156 	GEM_BUG_ON(!size);
157 	GEM_BUG_ON(!IS_ALIGNED(size, BIT(__ffs(page_mask))));
158 
159 	if (size >> PAGE_SHIFT > INT_MAX)
160 		return ERR_PTR(-E2BIG);
161 
162 	if (overflows_type(size, obj->base.size))
163 		return ERR_PTR(-E2BIG);
164 
165 	obj = i915_gem_object_alloc();
166 	if (!obj)
167 		return ERR_PTR(-ENOMEM);
168 
169 	drm_gem_private_object_init(&i915->drm, &obj->base, size);
170 	i915_gem_object_init(obj, &huge_page_ops, &lock_class);
171 
172 	i915_gem_object_set_volatile(obj);
173 
174 	obj->write_domain = I915_GEM_DOMAIN_CPU;
175 	obj->read_domains = I915_GEM_DOMAIN_CPU;
176 	obj->cache_level = I915_CACHE_NONE;
177 
178 	obj->mm.page_mask = page_mask;
179 
180 	return obj;
181 }
182 
183 static int fake_get_huge_pages(struct drm_i915_gem_object *obj)
184 {
185 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
186 	const u64 max_len = rounddown_pow_of_two(UINT_MAX);
187 	struct sg_table *st;
188 	struct scatterlist *sg;
189 	unsigned int sg_page_sizes;
190 	u64 rem;
191 
192 	st = kmalloc(sizeof(*st), GFP);
193 	if (!st)
194 		return -ENOMEM;
195 
196 	if (sg_alloc_table(st, obj->base.size >> PAGE_SHIFT, GFP)) {
197 		kfree(st);
198 		return -ENOMEM;
199 	}
200 
201 	/* Use optimal page sized chunks to fill in the sg table */
202 	rem = obj->base.size;
203 	sg = st->sgl;
204 	st->nents = 0;
205 	sg_page_sizes = 0;
206 	do {
207 		unsigned int page_size = get_largest_page_size(i915, rem);
208 		unsigned int len = min(page_size * div_u64(rem, page_size),
209 				       max_len);
210 
211 		GEM_BUG_ON(!page_size);
212 
213 		sg->offset = 0;
214 		sg->length = len;
215 		sg_dma_len(sg) = len;
216 		sg_dma_address(sg) = page_size;
217 
218 		sg_page_sizes |= len;
219 
220 		st->nents++;
221 
222 		rem -= len;
223 		if (!rem) {
224 			sg_mark_end(sg);
225 			break;
226 		}
227 
228 		sg = sg_next(sg);
229 	} while (1);
230 
231 	i915_sg_trim(st);
232 
233 	__i915_gem_object_set_pages(obj, st, sg_page_sizes);
234 
235 	return 0;
236 }
237 
238 static int fake_get_huge_pages_single(struct drm_i915_gem_object *obj)
239 {
240 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
241 	struct sg_table *st;
242 	struct scatterlist *sg;
243 	unsigned int page_size;
244 
245 	st = kmalloc(sizeof(*st), GFP);
246 	if (!st)
247 		return -ENOMEM;
248 
249 	if (sg_alloc_table(st, 1, GFP)) {
250 		kfree(st);
251 		return -ENOMEM;
252 	}
253 
254 	sg = st->sgl;
255 	st->nents = 1;
256 
257 	page_size = get_largest_page_size(i915, obj->base.size);
258 	GEM_BUG_ON(!page_size);
259 
260 	sg->offset = 0;
261 	sg->length = obj->base.size;
262 	sg_dma_len(sg) = obj->base.size;
263 	sg_dma_address(sg) = page_size;
264 
265 	__i915_gem_object_set_pages(obj, st, sg->length);
266 
267 	return 0;
268 #undef GFP
269 }
270 
271 static void fake_free_huge_pages(struct drm_i915_gem_object *obj,
272 				 struct sg_table *pages)
273 {
274 	sg_free_table(pages);
275 	kfree(pages);
276 }
277 
278 static void fake_put_huge_pages(struct drm_i915_gem_object *obj,
279 				struct sg_table *pages)
280 {
281 	fake_free_huge_pages(obj, pages);
282 	obj->mm.dirty = false;
283 }
284 
285 static const struct drm_i915_gem_object_ops fake_ops = {
286 	.flags = I915_GEM_OBJECT_IS_SHRINKABLE,
287 	.get_pages = fake_get_huge_pages,
288 	.put_pages = fake_put_huge_pages,
289 };
290 
291 static const struct drm_i915_gem_object_ops fake_ops_single = {
292 	.flags = I915_GEM_OBJECT_IS_SHRINKABLE,
293 	.get_pages = fake_get_huge_pages_single,
294 	.put_pages = fake_put_huge_pages,
295 };
296 
297 static struct drm_i915_gem_object *
298 fake_huge_pages_object(struct drm_i915_private *i915, u64 size, bool single)
299 {
300 	static struct lock_class_key lock_class;
301 	struct drm_i915_gem_object *obj;
302 
303 	GEM_BUG_ON(!size);
304 	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
305 
306 	if (size >> PAGE_SHIFT > UINT_MAX)
307 		return ERR_PTR(-E2BIG);
308 
309 	if (overflows_type(size, obj->base.size))
310 		return ERR_PTR(-E2BIG);
311 
312 	obj = i915_gem_object_alloc();
313 	if (!obj)
314 		return ERR_PTR(-ENOMEM);
315 
316 	drm_gem_private_object_init(&i915->drm, &obj->base, size);
317 
318 	if (single)
319 		i915_gem_object_init(obj, &fake_ops_single, &lock_class);
320 	else
321 		i915_gem_object_init(obj, &fake_ops, &lock_class);
322 
323 	i915_gem_object_set_volatile(obj);
324 
325 	obj->write_domain = I915_GEM_DOMAIN_CPU;
326 	obj->read_domains = I915_GEM_DOMAIN_CPU;
327 	obj->cache_level = I915_CACHE_NONE;
328 
329 	return obj;
330 }
331 
332 static int igt_check_page_sizes(struct i915_vma *vma)
333 {
334 	struct drm_i915_private *i915 = vma->vm->i915;
335 	unsigned int supported = INTEL_INFO(i915)->page_sizes;
336 	struct drm_i915_gem_object *obj = vma->obj;
337 	int err;
338 
339 	/* We have to wait for the async bind to complete before our asserts */
340 	err = i915_vma_sync(vma);
341 	if (err)
342 		return err;
343 
344 	if (!HAS_PAGE_SIZES(i915, vma->page_sizes.sg)) {
345 		pr_err("unsupported page_sizes.sg=%u, supported=%u\n",
346 		       vma->page_sizes.sg & ~supported, supported);
347 		err = -EINVAL;
348 	}
349 
350 	if (!HAS_PAGE_SIZES(i915, vma->page_sizes.gtt)) {
351 		pr_err("unsupported page_sizes.gtt=%u, supported=%u\n",
352 		       vma->page_sizes.gtt & ~supported, supported);
353 		err = -EINVAL;
354 	}
355 
356 	if (vma->page_sizes.phys != obj->mm.page_sizes.phys) {
357 		pr_err("vma->page_sizes.phys(%u) != obj->mm.page_sizes.phys(%u)\n",
358 		       vma->page_sizes.phys, obj->mm.page_sizes.phys);
359 		err = -EINVAL;
360 	}
361 
362 	if (vma->page_sizes.sg != obj->mm.page_sizes.sg) {
363 		pr_err("vma->page_sizes.sg(%u) != obj->mm.page_sizes.sg(%u)\n",
364 		       vma->page_sizes.sg, obj->mm.page_sizes.sg);
365 		err = -EINVAL;
366 	}
367 
368 	if (obj->mm.page_sizes.gtt) {
369 		pr_err("obj->page_sizes.gtt(%u) should never be set\n",
370 		       obj->mm.page_sizes.gtt);
371 		err = -EINVAL;
372 	}
373 
374 	return err;
375 }
376 
377 static int igt_mock_exhaust_device_supported_pages(void *arg)
378 {
379 	struct i915_ppgtt *ppgtt = arg;
380 	struct drm_i915_private *i915 = ppgtt->vm.i915;
381 	unsigned int saved_mask = INTEL_INFO(i915)->page_sizes;
382 	struct drm_i915_gem_object *obj;
383 	struct i915_vma *vma;
384 	int i, j, single;
385 	int err;
386 
387 	/*
388 	 * Sanity check creating objects with every valid page support
389 	 * combination for our mock device.
390 	 */
391 
392 	for (i = 1; i < BIT(ARRAY_SIZE(page_sizes)); i++) {
393 		unsigned int combination = 0;
394 
395 		for (j = 0; j < ARRAY_SIZE(page_sizes); j++) {
396 			if (i & BIT(j))
397 				combination |= page_sizes[j];
398 		}
399 
400 		mkwrite_device_info(i915)->page_sizes = combination;
401 
402 		for (single = 0; single <= 1; ++single) {
403 			obj = fake_huge_pages_object(i915, combination, !!single);
404 			if (IS_ERR(obj)) {
405 				err = PTR_ERR(obj);
406 				goto out_device;
407 			}
408 
409 			if (obj->base.size != combination) {
410 				pr_err("obj->base.size=%zu, expected=%u\n",
411 				       obj->base.size, combination);
412 				err = -EINVAL;
413 				goto out_put;
414 			}
415 
416 			vma = i915_vma_instance(obj, &ppgtt->vm, NULL);
417 			if (IS_ERR(vma)) {
418 				err = PTR_ERR(vma);
419 				goto out_put;
420 			}
421 
422 			err = i915_vma_pin(vma, 0, 0, PIN_USER);
423 			if (err)
424 				goto out_close;
425 
426 			err = igt_check_page_sizes(vma);
427 
428 			if (vma->page_sizes.sg != combination) {
429 				pr_err("page_sizes.sg=%u, expected=%u\n",
430 				       vma->page_sizes.sg, combination);
431 				err = -EINVAL;
432 			}
433 
434 			i915_vma_unpin(vma);
435 			i915_vma_close(vma);
436 
437 			i915_gem_object_put(obj);
438 
439 			if (err)
440 				goto out_device;
441 		}
442 	}
443 
444 	goto out_device;
445 
446 out_close:
447 	i915_vma_close(vma);
448 out_put:
449 	i915_gem_object_put(obj);
450 out_device:
451 	mkwrite_device_info(i915)->page_sizes = saved_mask;
452 
453 	return err;
454 }
455 
456 static int igt_mock_memory_region_huge_pages(void *arg)
457 {
458 	const unsigned int flags[] = { 0, I915_BO_ALLOC_CONTIGUOUS };
459 	struct i915_ppgtt *ppgtt = arg;
460 	struct drm_i915_private *i915 = ppgtt->vm.i915;
461 	unsigned long supported = INTEL_INFO(i915)->page_sizes;
462 	struct intel_memory_region *mem;
463 	struct drm_i915_gem_object *obj;
464 	struct i915_vma *vma;
465 	int bit;
466 	int err = 0;
467 
468 	mem = mock_region_create(i915, 0, SZ_2G, I915_GTT_PAGE_SIZE_4K, 0);
469 	if (IS_ERR(mem)) {
470 		pr_err("%s failed to create memory region\n", __func__);
471 		return PTR_ERR(mem);
472 	}
473 
474 	for_each_set_bit(bit, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
475 		unsigned int page_size = BIT(bit);
476 		resource_size_t phys;
477 		int i;
478 
479 		for (i = 0; i < ARRAY_SIZE(flags); ++i) {
480 			obj = i915_gem_object_create_region(mem, page_size,
481 							    flags[i]);
482 			if (IS_ERR(obj)) {
483 				err = PTR_ERR(obj);
484 				goto out_region;
485 			}
486 
487 			vma = i915_vma_instance(obj, &ppgtt->vm, NULL);
488 			if (IS_ERR(vma)) {
489 				err = PTR_ERR(vma);
490 				goto out_put;
491 			}
492 
493 			err = i915_vma_pin(vma, 0, 0, PIN_USER);
494 			if (err)
495 				goto out_close;
496 
497 			err = igt_check_page_sizes(vma);
498 			if (err)
499 				goto out_unpin;
500 
501 			phys = i915_gem_object_get_dma_address(obj, 0);
502 			if (!IS_ALIGNED(phys, page_size)) {
503 				pr_err("%s addr misaligned(%pa) page_size=%u\n",
504 				       __func__, &phys, page_size);
505 				err = -EINVAL;
506 				goto out_unpin;
507 			}
508 
509 			if (vma->page_sizes.gtt != page_size) {
510 				pr_err("%s page_sizes.gtt=%u, expected=%u\n",
511 				       __func__, vma->page_sizes.gtt,
512 				       page_size);
513 				err = -EINVAL;
514 				goto out_unpin;
515 			}
516 
517 			i915_vma_unpin(vma);
518 			i915_vma_close(vma);
519 
520 			__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
521 			i915_gem_object_put(obj);
522 		}
523 	}
524 
525 	goto out_region;
526 
527 out_unpin:
528 	i915_vma_unpin(vma);
529 out_close:
530 	i915_vma_close(vma);
531 out_put:
532 	i915_gem_object_put(obj);
533 out_region:
534 	intel_memory_region_put(mem);
535 	return err;
536 }
537 
538 static int igt_mock_ppgtt_misaligned_dma(void *arg)
539 {
540 	struct i915_ppgtt *ppgtt = arg;
541 	struct drm_i915_private *i915 = ppgtt->vm.i915;
542 	unsigned long supported = INTEL_INFO(i915)->page_sizes;
543 	struct drm_i915_gem_object *obj;
544 	int bit;
545 	int err;
546 
547 	/*
548 	 * Sanity check dma misalignment for huge pages -- the dma addresses we
549 	 * insert into the paging structures need to always respect the page
550 	 * size alignment.
551 	 */
552 
553 	bit = ilog2(I915_GTT_PAGE_SIZE_64K);
554 
555 	for_each_set_bit_from(bit, &supported,
556 			      ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
557 		IGT_TIMEOUT(end_time);
558 		unsigned int page_size = BIT(bit);
559 		unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
560 		unsigned int offset;
561 		unsigned int size =
562 			round_up(page_size, I915_GTT_PAGE_SIZE_2M) << 1;
563 		struct i915_vma *vma;
564 
565 		obj = fake_huge_pages_object(i915, size, true);
566 		if (IS_ERR(obj))
567 			return PTR_ERR(obj);
568 
569 		if (obj->base.size != size) {
570 			pr_err("obj->base.size=%zu, expected=%u\n",
571 			       obj->base.size, size);
572 			err = -EINVAL;
573 			goto out_put;
574 		}
575 
576 		err = i915_gem_object_pin_pages(obj);
577 		if (err)
578 			goto out_put;
579 
580 		/* Force the page size for this object */
581 		obj->mm.page_sizes.sg = page_size;
582 
583 		vma = i915_vma_instance(obj, &ppgtt->vm, NULL);
584 		if (IS_ERR(vma)) {
585 			err = PTR_ERR(vma);
586 			goto out_unpin;
587 		}
588 
589 		err = i915_vma_pin(vma, 0, 0, flags);
590 		if (err) {
591 			i915_vma_close(vma);
592 			goto out_unpin;
593 		}
594 
595 
596 		err = igt_check_page_sizes(vma);
597 
598 		if (vma->page_sizes.gtt != page_size) {
599 			pr_err("page_sizes.gtt=%u, expected %u\n",
600 			       vma->page_sizes.gtt, page_size);
601 			err = -EINVAL;
602 		}
603 
604 		i915_vma_unpin(vma);
605 
606 		if (err) {
607 			i915_vma_close(vma);
608 			goto out_unpin;
609 		}
610 
611 		/*
612 		 * Try all the other valid offsets until the next
613 		 * boundary -- should always fall back to using 4K
614 		 * pages.
615 		 */
616 		for (offset = 4096; offset < page_size; offset += 4096) {
617 			err = i915_vma_unbind(vma);
618 			if (err) {
619 				i915_vma_close(vma);
620 				goto out_unpin;
621 			}
622 
623 			err = i915_vma_pin(vma, 0, 0, flags | offset);
624 			if (err) {
625 				i915_vma_close(vma);
626 				goto out_unpin;
627 			}
628 
629 			err = igt_check_page_sizes(vma);
630 
631 			if (vma->page_sizes.gtt != I915_GTT_PAGE_SIZE_4K) {
632 				pr_err("page_sizes.gtt=%u, expected %llu\n",
633 				       vma->page_sizes.gtt, I915_GTT_PAGE_SIZE_4K);
634 				err = -EINVAL;
635 			}
636 
637 			i915_vma_unpin(vma);
638 
639 			if (err) {
640 				i915_vma_close(vma);
641 				goto out_unpin;
642 			}
643 
644 			if (igt_timeout(end_time,
645 					"%s timed out at offset %x with page-size %x\n",
646 					__func__, offset, page_size))
647 				break;
648 		}
649 
650 		i915_vma_close(vma);
651 
652 		i915_gem_object_unpin_pages(obj);
653 		__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
654 		i915_gem_object_put(obj);
655 	}
656 
657 	return 0;
658 
659 out_unpin:
660 	i915_gem_object_unpin_pages(obj);
661 out_put:
662 	i915_gem_object_put(obj);
663 
664 	return err;
665 }
666 
667 static void close_object_list(struct list_head *objects,
668 			      struct i915_ppgtt *ppgtt)
669 {
670 	struct drm_i915_gem_object *obj, *on;
671 
672 	list_for_each_entry_safe(obj, on, objects, st_link) {
673 		struct i915_vma *vma;
674 
675 		vma = i915_vma_instance(obj, &ppgtt->vm, NULL);
676 		if (!IS_ERR(vma))
677 			i915_vma_close(vma);
678 
679 		list_del(&obj->st_link);
680 		i915_gem_object_unpin_pages(obj);
681 		__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
682 		i915_gem_object_put(obj);
683 	}
684 }
685 
686 static int igt_mock_ppgtt_huge_fill(void *arg)
687 {
688 	struct i915_ppgtt *ppgtt = arg;
689 	struct drm_i915_private *i915 = ppgtt->vm.i915;
690 	unsigned long max_pages = ppgtt->vm.total >> PAGE_SHIFT;
691 	unsigned long page_num;
692 	bool single = false;
693 	LIST_HEAD(objects);
694 	IGT_TIMEOUT(end_time);
695 	int err = -ENODEV;
696 
697 	for_each_prime_number_from(page_num, 1, max_pages) {
698 		struct drm_i915_gem_object *obj;
699 		u64 size = page_num << PAGE_SHIFT;
700 		struct i915_vma *vma;
701 		unsigned int expected_gtt = 0;
702 		int i;
703 
704 		obj = fake_huge_pages_object(i915, size, single);
705 		if (IS_ERR(obj)) {
706 			err = PTR_ERR(obj);
707 			break;
708 		}
709 
710 		if (obj->base.size != size) {
711 			pr_err("obj->base.size=%zd, expected=%llu\n",
712 			       obj->base.size, size);
713 			i915_gem_object_put(obj);
714 			err = -EINVAL;
715 			break;
716 		}
717 
718 		err = i915_gem_object_pin_pages(obj);
719 		if (err) {
720 			i915_gem_object_put(obj);
721 			break;
722 		}
723 
724 		list_add(&obj->st_link, &objects);
725 
726 		vma = i915_vma_instance(obj, &ppgtt->vm, NULL);
727 		if (IS_ERR(vma)) {
728 			err = PTR_ERR(vma);
729 			break;
730 		}
731 
732 		err = i915_vma_pin(vma, 0, 0, PIN_USER);
733 		if (err)
734 			break;
735 
736 		err = igt_check_page_sizes(vma);
737 		if (err) {
738 			i915_vma_unpin(vma);
739 			break;
740 		}
741 
742 		/*
743 		 * Figure out the expected gtt page size knowing that we go from
744 		 * largest to smallest page size sg chunks, and that we align to
745 		 * the largest page size.
746 		 */
747 		for (i = 0; i < ARRAY_SIZE(page_sizes); ++i) {
748 			unsigned int page_size = page_sizes[i];
749 
750 			if (HAS_PAGE_SIZES(i915, page_size) &&
751 			    size >= page_size) {
752 				expected_gtt |= page_size;
753 				size &= page_size-1;
754 			}
755 		}
756 
757 		GEM_BUG_ON(!expected_gtt);
758 		GEM_BUG_ON(size);
759 
760 		if (expected_gtt & I915_GTT_PAGE_SIZE_4K)
761 			expected_gtt &= ~I915_GTT_PAGE_SIZE_64K;
762 
763 		i915_vma_unpin(vma);
764 
765 		if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
766 			if (!IS_ALIGNED(vma->node.start,
767 					I915_GTT_PAGE_SIZE_2M)) {
768 				pr_err("node.start(%llx) not aligned to 2M\n",
769 				       vma->node.start);
770 				err = -EINVAL;
771 				break;
772 			}
773 
774 			if (!IS_ALIGNED(vma->node.size,
775 					I915_GTT_PAGE_SIZE_2M)) {
776 				pr_err("node.size(%llx) not aligned to 2M\n",
777 				       vma->node.size);
778 				err = -EINVAL;
779 				break;
780 			}
781 		}
782 
783 		if (vma->page_sizes.gtt != expected_gtt) {
784 			pr_err("gtt=%u, expected=%u, size=%zd, single=%s\n",
785 			       vma->page_sizes.gtt, expected_gtt,
786 			       obj->base.size, yesno(!!single));
787 			err = -EINVAL;
788 			break;
789 		}
790 
791 		if (igt_timeout(end_time,
792 				"%s timed out at size %zd\n",
793 				__func__, obj->base.size))
794 			break;
795 
796 		single = !single;
797 	}
798 
799 	close_object_list(&objects, ppgtt);
800 
801 	if (err == -ENOMEM || err == -ENOSPC)
802 		err = 0;
803 
804 	return err;
805 }
806 
807 static int igt_mock_ppgtt_64K(void *arg)
808 {
809 	struct i915_ppgtt *ppgtt = arg;
810 	struct drm_i915_private *i915 = ppgtt->vm.i915;
811 	struct drm_i915_gem_object *obj;
812 	const struct object_info {
813 		unsigned int size;
814 		unsigned int gtt;
815 		unsigned int offset;
816 	} objects[] = {
817 		/* Cases with forced padding/alignment */
818 		{
819 			.size = SZ_64K,
820 			.gtt = I915_GTT_PAGE_SIZE_64K,
821 			.offset = 0,
822 		},
823 		{
824 			.size = SZ_64K + SZ_4K,
825 			.gtt = I915_GTT_PAGE_SIZE_4K,
826 			.offset = 0,
827 		},
828 		{
829 			.size = SZ_64K - SZ_4K,
830 			.gtt = I915_GTT_PAGE_SIZE_4K,
831 			.offset = 0,
832 		},
833 		{
834 			.size = SZ_2M,
835 			.gtt = I915_GTT_PAGE_SIZE_64K,
836 			.offset = 0,
837 		},
838 		{
839 			.size = SZ_2M - SZ_4K,
840 			.gtt = I915_GTT_PAGE_SIZE_4K,
841 			.offset = 0,
842 		},
843 		{
844 			.size = SZ_2M + SZ_4K,
845 			.gtt = I915_GTT_PAGE_SIZE_64K | I915_GTT_PAGE_SIZE_4K,
846 			.offset = 0,
847 		},
848 		{
849 			.size = SZ_2M + SZ_64K,
850 			.gtt = I915_GTT_PAGE_SIZE_64K,
851 			.offset = 0,
852 		},
853 		{
854 			.size = SZ_2M - SZ_64K,
855 			.gtt = I915_GTT_PAGE_SIZE_64K,
856 			.offset = 0,
857 		},
858 		/* Try without any forced padding/alignment */
859 		{
860 			.size = SZ_64K,
861 			.offset = SZ_2M,
862 			.gtt = I915_GTT_PAGE_SIZE_4K,
863 		},
864 		{
865 			.size = SZ_128K,
866 			.offset = SZ_2M - SZ_64K,
867 			.gtt = I915_GTT_PAGE_SIZE_4K,
868 		},
869 	};
870 	struct i915_vma *vma;
871 	int i, single;
872 	int err;
873 
874 	/*
875 	 * Sanity check some of the trickiness with 64K pages -- either we can
876 	 * safely mark the whole page-table(2M block) as 64K, or we have to
877 	 * always fallback to 4K.
878 	 */
879 
880 	if (!HAS_PAGE_SIZES(i915, I915_GTT_PAGE_SIZE_64K))
881 		return 0;
882 
883 	for (i = 0; i < ARRAY_SIZE(objects); ++i) {
884 		unsigned int size = objects[i].size;
885 		unsigned int expected_gtt = objects[i].gtt;
886 		unsigned int offset = objects[i].offset;
887 		unsigned int flags = PIN_USER;
888 
889 		for (single = 0; single <= 1; single++) {
890 			obj = fake_huge_pages_object(i915, size, !!single);
891 			if (IS_ERR(obj))
892 				return PTR_ERR(obj);
893 
894 			err = i915_gem_object_pin_pages(obj);
895 			if (err)
896 				goto out_object_put;
897 
898 			/*
899 			 * Disable 2M pages -- We only want to use 64K/4K pages
900 			 * for this test.
901 			 */
902 			obj->mm.page_sizes.sg &= ~I915_GTT_PAGE_SIZE_2M;
903 
904 			vma = i915_vma_instance(obj, &ppgtt->vm, NULL);
905 			if (IS_ERR(vma)) {
906 				err = PTR_ERR(vma);
907 				goto out_object_unpin;
908 			}
909 
910 			if (offset)
911 				flags |= PIN_OFFSET_FIXED | offset;
912 
913 			err = i915_vma_pin(vma, 0, 0, flags);
914 			if (err)
915 				goto out_vma_close;
916 
917 			err = igt_check_page_sizes(vma);
918 			if (err)
919 				goto out_vma_unpin;
920 
921 			if (!offset && vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
922 				if (!IS_ALIGNED(vma->node.start,
923 						I915_GTT_PAGE_SIZE_2M)) {
924 					pr_err("node.start(%llx) not aligned to 2M\n",
925 					       vma->node.start);
926 					err = -EINVAL;
927 					goto out_vma_unpin;
928 				}
929 
930 				if (!IS_ALIGNED(vma->node.size,
931 						I915_GTT_PAGE_SIZE_2M)) {
932 					pr_err("node.size(%llx) not aligned to 2M\n",
933 					       vma->node.size);
934 					err = -EINVAL;
935 					goto out_vma_unpin;
936 				}
937 			}
938 
939 			if (vma->page_sizes.gtt != expected_gtt) {
940 				pr_err("gtt=%u, expected=%u, i=%d, single=%s\n",
941 				       vma->page_sizes.gtt, expected_gtt, i,
942 				       yesno(!!single));
943 				err = -EINVAL;
944 				goto out_vma_unpin;
945 			}
946 
947 			i915_vma_unpin(vma);
948 			i915_vma_close(vma);
949 
950 			i915_gem_object_unpin_pages(obj);
951 			__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
952 			i915_gem_object_put(obj);
953 		}
954 	}
955 
956 	return 0;
957 
958 out_vma_unpin:
959 	i915_vma_unpin(vma);
960 out_vma_close:
961 	i915_vma_close(vma);
962 out_object_unpin:
963 	i915_gem_object_unpin_pages(obj);
964 out_object_put:
965 	i915_gem_object_put(obj);
966 
967 	return err;
968 }
969 
970 static int gpu_write(struct intel_context *ce,
971 		     struct i915_vma *vma,
972 		     u32 dw,
973 		     u32 val)
974 {
975 	int err;
976 
977 	i915_gem_object_lock(vma->obj);
978 	err = i915_gem_object_set_to_gtt_domain(vma->obj, true);
979 	i915_gem_object_unlock(vma->obj);
980 	if (err)
981 		return err;
982 
983 	return igt_gpu_fill_dw(ce, vma, dw * sizeof(u32),
984 			       vma->size >> PAGE_SHIFT, val);
985 }
986 
987 static int
988 __cpu_check_shmem(struct drm_i915_gem_object *obj, u32 dword, u32 val)
989 {
990 	unsigned int needs_flush;
991 	unsigned long n;
992 	int err;
993 
994 	err = i915_gem_object_prepare_read(obj, &needs_flush);
995 	if (err)
996 		return err;
997 
998 	for (n = 0; n < obj->base.size >> PAGE_SHIFT; ++n) {
999 		u32 *ptr = kmap_atomic(i915_gem_object_get_page(obj, n));
1000 
1001 		if (needs_flush & CLFLUSH_BEFORE)
1002 			drm_clflush_virt_range(ptr, PAGE_SIZE);
1003 
1004 		if (ptr[dword] != val) {
1005 			pr_err("n=%lu ptr[%u]=%u, val=%u\n",
1006 			       n, dword, ptr[dword], val);
1007 			kunmap_atomic(ptr);
1008 			err = -EINVAL;
1009 			break;
1010 		}
1011 
1012 		kunmap_atomic(ptr);
1013 	}
1014 
1015 	i915_gem_object_finish_access(obj);
1016 
1017 	return err;
1018 }
1019 
1020 static int __cpu_check_lmem(struct drm_i915_gem_object *obj, u32 dword, u32 val)
1021 {
1022 	unsigned long n;
1023 	int err;
1024 
1025 	i915_gem_object_lock(obj);
1026 	err = i915_gem_object_set_to_wc_domain(obj, false);
1027 	i915_gem_object_unlock(obj);
1028 	if (err)
1029 		return err;
1030 
1031 	err = i915_gem_object_pin_pages(obj);
1032 	if (err)
1033 		return err;
1034 
1035 	for (n = 0; n < obj->base.size >> PAGE_SHIFT; ++n) {
1036 		u32 __iomem *base;
1037 		u32 read_val;
1038 
1039 		base = i915_gem_object_lmem_io_map_page_atomic(obj, n);
1040 
1041 		read_val = ioread32(base + dword);
1042 		io_mapping_unmap_atomic(base);
1043 		if (read_val != val) {
1044 			pr_err("n=%lu base[%u]=%u, val=%u\n",
1045 			       n, dword, read_val, val);
1046 			err = -EINVAL;
1047 			break;
1048 		}
1049 	}
1050 
1051 	i915_gem_object_unpin_pages(obj);
1052 	return err;
1053 }
1054 
1055 static int cpu_check(struct drm_i915_gem_object *obj, u32 dword, u32 val)
1056 {
1057 	if (i915_gem_object_has_struct_page(obj))
1058 		return __cpu_check_shmem(obj, dword, val);
1059 	else if (i915_gem_object_is_lmem(obj))
1060 		return __cpu_check_lmem(obj, dword, val);
1061 
1062 	return -ENODEV;
1063 }
1064 
1065 static int __igt_write_huge(struct intel_context *ce,
1066 			    struct drm_i915_gem_object *obj,
1067 			    u64 size, u64 offset,
1068 			    u32 dword, u32 val)
1069 {
1070 	unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
1071 	struct i915_vma *vma;
1072 	int err;
1073 
1074 	vma = i915_vma_instance(obj, ce->vm, NULL);
1075 	if (IS_ERR(vma))
1076 		return PTR_ERR(vma);
1077 
1078 	err = i915_vma_unbind(vma);
1079 	if (err)
1080 		goto out_vma_close;
1081 
1082 	err = i915_vma_pin(vma, size, 0, flags | offset);
1083 	if (err) {
1084 		/*
1085 		 * The ggtt may have some pages reserved so
1086 		 * refrain from erroring out.
1087 		 */
1088 		if (err == -ENOSPC && i915_is_ggtt(ce->vm))
1089 			err = 0;
1090 
1091 		goto out_vma_close;
1092 	}
1093 
1094 	err = igt_check_page_sizes(vma);
1095 	if (err)
1096 		goto out_vma_unpin;
1097 
1098 	err = gpu_write(ce, vma, dword, val);
1099 	if (err) {
1100 		pr_err("gpu-write failed at offset=%llx\n", offset);
1101 		goto out_vma_unpin;
1102 	}
1103 
1104 	err = cpu_check(obj, dword, val);
1105 	if (err) {
1106 		pr_err("cpu-check failed at offset=%llx\n", offset);
1107 		goto out_vma_unpin;
1108 	}
1109 
1110 out_vma_unpin:
1111 	i915_vma_unpin(vma);
1112 out_vma_close:
1113 	i915_vma_destroy(vma);
1114 
1115 	return err;
1116 }
1117 
1118 static int igt_write_huge(struct i915_gem_context *ctx,
1119 			  struct drm_i915_gem_object *obj)
1120 {
1121 	struct i915_gem_engines *engines;
1122 	struct i915_gem_engines_iter it;
1123 	struct intel_context *ce;
1124 	I915_RND_STATE(prng);
1125 	IGT_TIMEOUT(end_time);
1126 	unsigned int max_page_size;
1127 	unsigned int count;
1128 	u64 max;
1129 	u64 num;
1130 	u64 size;
1131 	int *order;
1132 	int i, n;
1133 	int err = 0;
1134 
1135 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
1136 
1137 	size = obj->base.size;
1138 	if (obj->mm.page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
1139 		size = round_up(size, I915_GTT_PAGE_SIZE_2M);
1140 
1141 	n = 0;
1142 	count = 0;
1143 	max = U64_MAX;
1144 	for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
1145 		count++;
1146 		if (!intel_engine_can_store_dword(ce->engine))
1147 			continue;
1148 
1149 		max = min(max, ce->vm->total);
1150 		n++;
1151 	}
1152 	i915_gem_context_unlock_engines(ctx);
1153 	if (!n)
1154 		return 0;
1155 
1156 	/*
1157 	 * To keep things interesting when alternating between engines in our
1158 	 * randomized order, lets also make feeding to the same engine a few
1159 	 * times in succession a possibility by enlarging the permutation array.
1160 	 */
1161 	order = i915_random_order(count * count, &prng);
1162 	if (!order)
1163 		return -ENOMEM;
1164 
1165 	max_page_size = rounddown_pow_of_two(obj->mm.page_sizes.sg);
1166 	max = div_u64(max - size, max_page_size);
1167 
1168 	/*
1169 	 * Try various offsets in an ascending/descending fashion until we
1170 	 * timeout -- we want to avoid issues hidden by effectively always using
1171 	 * offset = 0.
1172 	 */
1173 	i = 0;
1174 	engines = i915_gem_context_lock_engines(ctx);
1175 	for_each_prime_number_from(num, 0, max) {
1176 		u64 offset_low = num * max_page_size;
1177 		u64 offset_high = (max - num) * max_page_size;
1178 		u32 dword = offset_in_page(num) / 4;
1179 		struct intel_context *ce;
1180 
1181 		ce = engines->engines[order[i] % engines->num_engines];
1182 		i = (i + 1) % (count * count);
1183 		if (!ce || !intel_engine_can_store_dword(ce->engine))
1184 			continue;
1185 
1186 		/*
1187 		 * In order to utilize 64K pages we need to both pad the vma
1188 		 * size and ensure the vma offset is at the start of the pt
1189 		 * boundary, however to improve coverage we opt for testing both
1190 		 * aligned and unaligned offsets.
1191 		 */
1192 		if (obj->mm.page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
1193 			offset_low = round_down(offset_low,
1194 						I915_GTT_PAGE_SIZE_2M);
1195 
1196 		err = __igt_write_huge(ce, obj, size, offset_low,
1197 				       dword, num + 1);
1198 		if (err)
1199 			break;
1200 
1201 		err = __igt_write_huge(ce, obj, size, offset_high,
1202 				       dword, num + 1);
1203 		if (err)
1204 			break;
1205 
1206 		if (igt_timeout(end_time,
1207 				"%s timed out on %s, offset_low=%llx offset_high=%llx, max_page_size=%x\n",
1208 				__func__, ce->engine->name, offset_low, offset_high,
1209 				max_page_size))
1210 			break;
1211 	}
1212 	i915_gem_context_unlock_engines(ctx);
1213 
1214 	kfree(order);
1215 
1216 	return err;
1217 }
1218 
1219 static int igt_ppgtt_exhaust_huge(void *arg)
1220 {
1221 	struct i915_gem_context *ctx = arg;
1222 	struct drm_i915_private *i915 = ctx->i915;
1223 	unsigned long supported = INTEL_INFO(i915)->page_sizes;
1224 	static unsigned int pages[ARRAY_SIZE(page_sizes)];
1225 	struct drm_i915_gem_object *obj;
1226 	unsigned int size_mask;
1227 	unsigned int page_mask;
1228 	int n, i;
1229 	int err = -ENODEV;
1230 
1231 	if (supported == I915_GTT_PAGE_SIZE_4K)
1232 		return 0;
1233 
1234 	/*
1235 	 * Sanity check creating objects with a varying mix of page sizes --
1236 	 * ensuring that our writes lands in the right place.
1237 	 */
1238 
1239 	n = 0;
1240 	for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1)
1241 		pages[n++] = BIT(i);
1242 
1243 	for (size_mask = 2; size_mask < BIT(n); size_mask++) {
1244 		unsigned int size = 0;
1245 
1246 		for (i = 0; i < n; i++) {
1247 			if (size_mask & BIT(i))
1248 				size |= pages[i];
1249 		}
1250 
1251 		/*
1252 		 * For our page mask we want to enumerate all the page-size
1253 		 * combinations which will fit into our chosen object size.
1254 		 */
1255 		for (page_mask = 2; page_mask <= size_mask; page_mask++) {
1256 			unsigned int page_sizes = 0;
1257 
1258 			for (i = 0; i < n; i++) {
1259 				if (page_mask & BIT(i))
1260 					page_sizes |= pages[i];
1261 			}
1262 
1263 			/*
1264 			 * Ensure that we can actually fill the given object
1265 			 * with our chosen page mask.
1266 			 */
1267 			if (!IS_ALIGNED(size, BIT(__ffs(page_sizes))))
1268 				continue;
1269 
1270 			obj = huge_pages_object(i915, size, page_sizes);
1271 			if (IS_ERR(obj)) {
1272 				err = PTR_ERR(obj);
1273 				goto out_device;
1274 			}
1275 
1276 			err = i915_gem_object_pin_pages(obj);
1277 			if (err) {
1278 				i915_gem_object_put(obj);
1279 
1280 				if (err == -ENOMEM) {
1281 					pr_info("unable to get pages, size=%u, pages=%u\n",
1282 						size, page_sizes);
1283 					err = 0;
1284 					break;
1285 				}
1286 
1287 				pr_err("pin_pages failed, size=%u, pages=%u\n",
1288 				       size_mask, page_mask);
1289 
1290 				goto out_device;
1291 			}
1292 
1293 			/* Force the page-size for the gtt insertion */
1294 			obj->mm.page_sizes.sg = page_sizes;
1295 
1296 			err = igt_write_huge(ctx, obj);
1297 			if (err) {
1298 				pr_err("exhaust write-huge failed with size=%u\n",
1299 				       size);
1300 				goto out_unpin;
1301 			}
1302 
1303 			i915_gem_object_unpin_pages(obj);
1304 			__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
1305 			i915_gem_object_put(obj);
1306 		}
1307 	}
1308 
1309 	goto out_device;
1310 
1311 out_unpin:
1312 	i915_gem_object_unpin_pages(obj);
1313 	i915_gem_object_put(obj);
1314 out_device:
1315 	mkwrite_device_info(i915)->page_sizes = supported;
1316 
1317 	return err;
1318 }
1319 
1320 typedef struct drm_i915_gem_object *
1321 (*igt_create_fn)(struct drm_i915_private *i915, u32 size, u32 flags);
1322 
1323 static inline bool igt_can_allocate_thp(struct drm_i915_private *i915)
1324 {
1325 	return i915->mm.gemfs && has_transparent_hugepage();
1326 }
1327 
1328 static struct drm_i915_gem_object *
1329 igt_create_shmem(struct drm_i915_private *i915, u32 size, u32 flags)
1330 {
1331 	if (!igt_can_allocate_thp(i915)) {
1332 		pr_info("%s missing THP support, skipping\n", __func__);
1333 		return ERR_PTR(-ENODEV);
1334 	}
1335 
1336 	return i915_gem_object_create_shmem(i915, size);
1337 }
1338 
1339 static struct drm_i915_gem_object *
1340 igt_create_internal(struct drm_i915_private *i915, u32 size, u32 flags)
1341 {
1342 	return i915_gem_object_create_internal(i915, size);
1343 }
1344 
1345 static struct drm_i915_gem_object *
1346 igt_create_system(struct drm_i915_private *i915, u32 size, u32 flags)
1347 {
1348 	return huge_pages_object(i915, size, size);
1349 }
1350 
1351 static struct drm_i915_gem_object *
1352 igt_create_local(struct drm_i915_private *i915, u32 size, u32 flags)
1353 {
1354 	return i915_gem_object_create_lmem(i915, size, flags);
1355 }
1356 
1357 static u32 igt_random_size(struct rnd_state *prng,
1358 			   u32 min_page_size,
1359 			   u32 max_page_size)
1360 {
1361 	u64 mask;
1362 	u32 size;
1363 
1364 	GEM_BUG_ON(!is_power_of_2(min_page_size));
1365 	GEM_BUG_ON(!is_power_of_2(max_page_size));
1366 	GEM_BUG_ON(min_page_size < PAGE_SIZE);
1367 	GEM_BUG_ON(min_page_size > max_page_size);
1368 
1369 	mask = ((max_page_size << 1ULL) - 1) & PAGE_MASK;
1370 	size = prandom_u32_state(prng) & mask;
1371 	if (size < min_page_size)
1372 		size |= min_page_size;
1373 
1374 	return size;
1375 }
1376 
1377 static int igt_ppgtt_smoke_huge(void *arg)
1378 {
1379 	struct i915_gem_context *ctx = arg;
1380 	struct drm_i915_private *i915 = ctx->i915;
1381 	struct drm_i915_gem_object *obj;
1382 	I915_RND_STATE(prng);
1383 	struct {
1384 		igt_create_fn fn;
1385 		u32 min;
1386 		u32 max;
1387 	} backends[] = {
1388 		{ igt_create_internal, SZ_64K, SZ_2M,  },
1389 		{ igt_create_shmem,    SZ_64K, SZ_32M, },
1390 		{ igt_create_local,    SZ_64K, SZ_1G,  },
1391 	};
1392 	int err;
1393 	int i;
1394 
1395 	/*
1396 	 * Sanity check that the HW uses huge pages correctly through our
1397 	 * various backends -- ensure that our writes land in the right place.
1398 	 */
1399 
1400 	for (i = 0; i < ARRAY_SIZE(backends); ++i) {
1401 		u32 min = backends[i].min;
1402 		u32 max = backends[i].max;
1403 		u32 size = max;
1404 try_again:
1405 		size = igt_random_size(&prng, min, rounddown_pow_of_two(size));
1406 
1407 		obj = backends[i].fn(i915, size, 0);
1408 		if (IS_ERR(obj)) {
1409 			err = PTR_ERR(obj);
1410 			if (err == -E2BIG) {
1411 				size >>= 1;
1412 				goto try_again;
1413 			} else if (err == -ENODEV) {
1414 				err = 0;
1415 				continue;
1416 			}
1417 
1418 			return err;
1419 		}
1420 
1421 		err = i915_gem_object_pin_pages(obj);
1422 		if (err) {
1423 			if (err == -ENXIO) {
1424 				i915_gem_object_put(obj);
1425 				size >>= 1;
1426 				goto try_again;
1427 			}
1428 			goto out_put;
1429 		}
1430 
1431 		if (obj->mm.page_sizes.phys < min) {
1432 			pr_info("%s unable to allocate huge-page(s) with size=%u, i=%d\n",
1433 				__func__, size, i);
1434 			err = -ENOMEM;
1435 			goto out_unpin;
1436 		}
1437 
1438 		err = igt_write_huge(ctx, obj);
1439 		if (err) {
1440 			pr_err("%s write-huge failed with size=%u, i=%d\n",
1441 			       __func__, size, i);
1442 		}
1443 out_unpin:
1444 		i915_gem_object_unpin_pages(obj);
1445 		__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
1446 out_put:
1447 		i915_gem_object_put(obj);
1448 
1449 		if (err == -ENOMEM || err == -ENXIO)
1450 			err = 0;
1451 
1452 		if (err)
1453 			break;
1454 
1455 		cond_resched();
1456 	}
1457 
1458 	return err;
1459 }
1460 
1461 static int igt_ppgtt_sanity_check(void *arg)
1462 {
1463 	struct i915_gem_context *ctx = arg;
1464 	struct drm_i915_private *i915 = ctx->i915;
1465 	unsigned int supported = INTEL_INFO(i915)->page_sizes;
1466 	struct {
1467 		igt_create_fn fn;
1468 		unsigned int flags;
1469 	} backends[] = {
1470 		{ igt_create_system, 0,                        },
1471 		{ igt_create_local,  I915_BO_ALLOC_CONTIGUOUS, },
1472 	};
1473 	struct {
1474 		u32 size;
1475 		u32 pages;
1476 	} combos[] = {
1477 		{ SZ_64K,		SZ_64K		},
1478 		{ SZ_2M,		SZ_2M		},
1479 		{ SZ_2M,		SZ_64K		},
1480 		{ SZ_2M - SZ_64K,	SZ_64K		},
1481 		{ SZ_2M - SZ_4K,	SZ_64K | SZ_4K	},
1482 		{ SZ_2M + SZ_4K,	SZ_64K | SZ_4K	},
1483 		{ SZ_2M + SZ_4K,	SZ_2M  | SZ_4K	},
1484 		{ SZ_2M + SZ_64K,	SZ_2M  | SZ_64K },
1485 	};
1486 	int i, j;
1487 	int err;
1488 
1489 	if (supported == I915_GTT_PAGE_SIZE_4K)
1490 		return 0;
1491 
1492 	/*
1493 	 * Sanity check that the HW behaves with a limited set of combinations.
1494 	 * We already have a bunch of randomised testing, which should give us
1495 	 * a decent amount of variation between runs, however we should keep
1496 	 * this to limit the chances of introducing a temporary regression, by
1497 	 * testing the most obvious cases that might make something blow up.
1498 	 */
1499 
1500 	for (i = 0; i < ARRAY_SIZE(backends); ++i) {
1501 		for (j = 0; j < ARRAY_SIZE(combos); ++j) {
1502 			struct drm_i915_gem_object *obj;
1503 			u32 size = combos[j].size;
1504 			u32 pages = combos[j].pages;
1505 
1506 			obj = backends[i].fn(i915, size, backends[i].flags);
1507 			if (IS_ERR(obj)) {
1508 				err = PTR_ERR(obj);
1509 				if (err == -ENODEV) {
1510 					pr_info("Device lacks local memory, skipping\n");
1511 					err = 0;
1512 					break;
1513 				}
1514 
1515 				return err;
1516 			}
1517 
1518 			err = i915_gem_object_pin_pages(obj);
1519 			if (err) {
1520 				i915_gem_object_put(obj);
1521 				goto out;
1522 			}
1523 
1524 			GEM_BUG_ON(pages > obj->base.size);
1525 			pages = pages & supported;
1526 
1527 			if (pages)
1528 				obj->mm.page_sizes.sg = pages;
1529 
1530 			err = igt_write_huge(ctx, obj);
1531 
1532 			i915_gem_object_unpin_pages(obj);
1533 			__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
1534 			i915_gem_object_put(obj);
1535 
1536 			if (err) {
1537 				pr_err("%s write-huge failed with size=%u pages=%u i=%d, j=%d\n",
1538 				       __func__, size, pages, i, j);
1539 				goto out;
1540 			}
1541 		}
1542 
1543 		cond_resched();
1544 	}
1545 
1546 out:
1547 	if (err == -ENOMEM)
1548 		err = 0;
1549 
1550 	return err;
1551 }
1552 
1553 static int igt_ppgtt_pin_update(void *arg)
1554 {
1555 	struct i915_gem_context *ctx = arg;
1556 	struct drm_i915_private *dev_priv = ctx->i915;
1557 	unsigned long supported = INTEL_INFO(dev_priv)->page_sizes;
1558 	struct drm_i915_gem_object *obj;
1559 	struct i915_gem_engines_iter it;
1560 	struct i915_address_space *vm;
1561 	struct intel_context *ce;
1562 	struct i915_vma *vma;
1563 	unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
1564 	unsigned int n;
1565 	int first, last;
1566 	int err = 0;
1567 
1568 	/*
1569 	 * Make sure there's no funny business when doing a PIN_UPDATE -- in the
1570 	 * past we had a subtle issue with being able to incorrectly do multiple
1571 	 * alloc va ranges on the same object when doing a PIN_UPDATE, which
1572 	 * resulted in some pretty nasty bugs, though only when using
1573 	 * huge-gtt-pages.
1574 	 */
1575 
1576 	vm = i915_gem_context_get_vm_rcu(ctx);
1577 	if (!i915_vm_is_4lvl(vm)) {
1578 		pr_info("48b PPGTT not supported, skipping\n");
1579 		goto out_vm;
1580 	}
1581 
1582 	first = ilog2(I915_GTT_PAGE_SIZE_64K);
1583 	last = ilog2(I915_GTT_PAGE_SIZE_2M);
1584 
1585 	for_each_set_bit_from(first, &supported, last + 1) {
1586 		unsigned int page_size = BIT(first);
1587 
1588 		obj = i915_gem_object_create_internal(dev_priv, page_size);
1589 		if (IS_ERR(obj))
1590 			return PTR_ERR(obj);
1591 
1592 		vma = i915_vma_instance(obj, vm, NULL);
1593 		if (IS_ERR(vma)) {
1594 			err = PTR_ERR(vma);
1595 			goto out_put;
1596 		}
1597 
1598 		err = i915_vma_pin(vma, SZ_2M, 0, flags);
1599 		if (err)
1600 			goto out_close;
1601 
1602 		if (vma->page_sizes.sg < page_size) {
1603 			pr_info("Unable to allocate page-size %x, finishing test early\n",
1604 				page_size);
1605 			goto out_unpin;
1606 		}
1607 
1608 		err = igt_check_page_sizes(vma);
1609 		if (err)
1610 			goto out_unpin;
1611 
1612 		if (vma->page_sizes.gtt != page_size) {
1613 			dma_addr_t addr = i915_gem_object_get_dma_address(obj, 0);
1614 
1615 			/*
1616 			 * The only valid reason for this to ever fail would be
1617 			 * if the dma-mapper screwed us over when we did the
1618 			 * dma_map_sg(), since it has the final say over the dma
1619 			 * address.
1620 			 */
1621 			if (IS_ALIGNED(addr, page_size)) {
1622 				pr_err("page_sizes.gtt=%u, expected=%u\n",
1623 				       vma->page_sizes.gtt, page_size);
1624 				err = -EINVAL;
1625 			} else {
1626 				pr_info("dma address misaligned, finishing test early\n");
1627 			}
1628 
1629 			goto out_unpin;
1630 		}
1631 
1632 		err = i915_vma_bind(vma, I915_CACHE_NONE, PIN_UPDATE, NULL);
1633 		if (err)
1634 			goto out_unpin;
1635 
1636 		i915_vma_unpin(vma);
1637 		i915_vma_close(vma);
1638 
1639 		i915_gem_object_put(obj);
1640 	}
1641 
1642 	obj = i915_gem_object_create_internal(dev_priv, PAGE_SIZE);
1643 	if (IS_ERR(obj))
1644 		return PTR_ERR(obj);
1645 
1646 	vma = i915_vma_instance(obj, vm, NULL);
1647 	if (IS_ERR(vma)) {
1648 		err = PTR_ERR(vma);
1649 		goto out_put;
1650 	}
1651 
1652 	err = i915_vma_pin(vma, 0, 0, flags);
1653 	if (err)
1654 		goto out_close;
1655 
1656 	/*
1657 	 * Make sure we don't end up with something like where the pde is still
1658 	 * pointing to the 2M page, and the pt we just filled-in is dangling --
1659 	 * we can check this by writing to the first page where it would then
1660 	 * land in the now stale 2M page.
1661 	 */
1662 
1663 	n = 0;
1664 	for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
1665 		if (!intel_engine_can_store_dword(ce->engine))
1666 			continue;
1667 
1668 		err = gpu_write(ce, vma, n++, 0xdeadbeaf);
1669 		if (err)
1670 			break;
1671 	}
1672 	i915_gem_context_unlock_engines(ctx);
1673 	if (err)
1674 		goto out_unpin;
1675 
1676 	while (n--) {
1677 		err = cpu_check(obj, n, 0xdeadbeaf);
1678 		if (err)
1679 			goto out_unpin;
1680 	}
1681 
1682 out_unpin:
1683 	i915_vma_unpin(vma);
1684 out_close:
1685 	i915_vma_close(vma);
1686 out_put:
1687 	i915_gem_object_put(obj);
1688 out_vm:
1689 	i915_vm_put(vm);
1690 
1691 	return err;
1692 }
1693 
1694 static int igt_tmpfs_fallback(void *arg)
1695 {
1696 	struct i915_gem_context *ctx = arg;
1697 	struct drm_i915_private *i915 = ctx->i915;
1698 	struct vfsmount *gemfs = i915->mm.gemfs;
1699 	struct i915_address_space *vm = i915_gem_context_get_vm_rcu(ctx);
1700 	struct drm_i915_gem_object *obj;
1701 	struct i915_vma *vma;
1702 	u32 *vaddr;
1703 	int err = 0;
1704 
1705 	/*
1706 	 * Make sure that we don't burst into a ball of flames upon falling back
1707 	 * to tmpfs, which we rely on if on the off-chance we encouter a failure
1708 	 * when setting up gemfs.
1709 	 */
1710 
1711 	i915->mm.gemfs = NULL;
1712 
1713 	obj = i915_gem_object_create_shmem(i915, PAGE_SIZE);
1714 	if (IS_ERR(obj)) {
1715 		err = PTR_ERR(obj);
1716 		goto out_restore;
1717 	}
1718 
1719 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1720 	if (IS_ERR(vaddr)) {
1721 		err = PTR_ERR(vaddr);
1722 		goto out_put;
1723 	}
1724 	*vaddr = 0xdeadbeaf;
1725 
1726 	__i915_gem_object_flush_map(obj, 0, 64);
1727 	i915_gem_object_unpin_map(obj);
1728 
1729 	vma = i915_vma_instance(obj, vm, NULL);
1730 	if (IS_ERR(vma)) {
1731 		err = PTR_ERR(vma);
1732 		goto out_put;
1733 	}
1734 
1735 	err = i915_vma_pin(vma, 0, 0, PIN_USER);
1736 	if (err)
1737 		goto out_close;
1738 
1739 	err = igt_check_page_sizes(vma);
1740 
1741 	i915_vma_unpin(vma);
1742 out_close:
1743 	i915_vma_close(vma);
1744 out_put:
1745 	i915_gem_object_put(obj);
1746 out_restore:
1747 	i915->mm.gemfs = gemfs;
1748 
1749 	i915_vm_put(vm);
1750 	return err;
1751 }
1752 
1753 static int igt_shrink_thp(void *arg)
1754 {
1755 	struct i915_gem_context *ctx = arg;
1756 	struct drm_i915_private *i915 = ctx->i915;
1757 	struct i915_address_space *vm = i915_gem_context_get_vm_rcu(ctx);
1758 	struct drm_i915_gem_object *obj;
1759 	struct i915_gem_engines_iter it;
1760 	struct intel_context *ce;
1761 	struct i915_vma *vma;
1762 	unsigned int flags = PIN_USER;
1763 	unsigned int n;
1764 	int err = 0;
1765 
1766 	/*
1767 	 * Sanity check shrinking huge-paged object -- make sure nothing blows
1768 	 * up.
1769 	 */
1770 
1771 	if (!igt_can_allocate_thp(i915)) {
1772 		pr_info("missing THP support, skipping\n");
1773 		goto out_vm;
1774 	}
1775 
1776 	obj = i915_gem_object_create_shmem(i915, SZ_2M);
1777 	if (IS_ERR(obj)) {
1778 		err = PTR_ERR(obj);
1779 		goto out_vm;
1780 	}
1781 
1782 	vma = i915_vma_instance(obj, vm, NULL);
1783 	if (IS_ERR(vma)) {
1784 		err = PTR_ERR(vma);
1785 		goto out_put;
1786 	}
1787 
1788 	err = i915_vma_pin(vma, 0, 0, flags);
1789 	if (err)
1790 		goto out_close;
1791 
1792 	if (obj->mm.page_sizes.phys < I915_GTT_PAGE_SIZE_2M) {
1793 		pr_info("failed to allocate THP, finishing test early\n");
1794 		goto out_unpin;
1795 	}
1796 
1797 	err = igt_check_page_sizes(vma);
1798 	if (err)
1799 		goto out_unpin;
1800 
1801 	n = 0;
1802 
1803 	for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
1804 		if (!intel_engine_can_store_dword(ce->engine))
1805 			continue;
1806 
1807 		err = gpu_write(ce, vma, n++, 0xdeadbeaf);
1808 		if (err)
1809 			break;
1810 	}
1811 	i915_gem_context_unlock_engines(ctx);
1812 	i915_vma_unpin(vma);
1813 	if (err)
1814 		goto out_close;
1815 
1816 	/*
1817 	 * Now that the pages are *unpinned* shrink-all should invoke
1818 	 * shmem to truncate our pages.
1819 	 */
1820 	i915_gem_shrink_all(i915);
1821 	if (i915_gem_object_has_pages(obj)) {
1822 		pr_err("shrink-all didn't truncate the pages\n");
1823 		err = -EINVAL;
1824 		goto out_close;
1825 	}
1826 
1827 	if (obj->mm.page_sizes.sg || obj->mm.page_sizes.phys) {
1828 		pr_err("residual page-size bits left\n");
1829 		err = -EINVAL;
1830 		goto out_close;
1831 	}
1832 
1833 	err = i915_vma_pin(vma, 0, 0, flags);
1834 	if (err)
1835 		goto out_close;
1836 
1837 	while (n--) {
1838 		err = cpu_check(obj, n, 0xdeadbeaf);
1839 		if (err)
1840 			break;
1841 	}
1842 
1843 out_unpin:
1844 	i915_vma_unpin(vma);
1845 out_close:
1846 	i915_vma_close(vma);
1847 out_put:
1848 	i915_gem_object_put(obj);
1849 out_vm:
1850 	i915_vm_put(vm);
1851 
1852 	return err;
1853 }
1854 
1855 int i915_gem_huge_page_mock_selftests(void)
1856 {
1857 	static const struct i915_subtest tests[] = {
1858 		SUBTEST(igt_mock_exhaust_device_supported_pages),
1859 		SUBTEST(igt_mock_memory_region_huge_pages),
1860 		SUBTEST(igt_mock_ppgtt_misaligned_dma),
1861 		SUBTEST(igt_mock_ppgtt_huge_fill),
1862 		SUBTEST(igt_mock_ppgtt_64K),
1863 	};
1864 	struct drm_i915_private *dev_priv;
1865 	struct i915_ppgtt *ppgtt;
1866 	int err;
1867 
1868 	dev_priv = mock_gem_device();
1869 	if (!dev_priv)
1870 		return -ENOMEM;
1871 
1872 	/* Pretend to be a device which supports the 48b PPGTT */
1873 	mkwrite_device_info(dev_priv)->ppgtt_type = INTEL_PPGTT_FULL;
1874 	mkwrite_device_info(dev_priv)->ppgtt_size = 48;
1875 
1876 	ppgtt = i915_ppgtt_create(dev_priv);
1877 	if (IS_ERR(ppgtt)) {
1878 		err = PTR_ERR(ppgtt);
1879 		goto out_unlock;
1880 	}
1881 
1882 	if (!i915_vm_is_4lvl(&ppgtt->vm)) {
1883 		pr_err("failed to create 48b PPGTT\n");
1884 		err = -EINVAL;
1885 		goto out_close;
1886 	}
1887 
1888 	/* If we were ever hit this then it's time to mock the 64K scratch */
1889 	if (!i915_vm_has_scratch_64K(&ppgtt->vm)) {
1890 		pr_err("PPGTT missing 64K scratch page\n");
1891 		err = -EINVAL;
1892 		goto out_close;
1893 	}
1894 
1895 	err = i915_subtests(tests, ppgtt);
1896 
1897 out_close:
1898 	i915_vm_put(&ppgtt->vm);
1899 
1900 out_unlock:
1901 	drm_dev_put(&dev_priv->drm);
1902 	return err;
1903 }
1904 
1905 int i915_gem_huge_page_live_selftests(struct drm_i915_private *i915)
1906 {
1907 	static const struct i915_subtest tests[] = {
1908 		SUBTEST(igt_shrink_thp),
1909 		SUBTEST(igt_ppgtt_pin_update),
1910 		SUBTEST(igt_tmpfs_fallback),
1911 		SUBTEST(igt_ppgtt_exhaust_huge),
1912 		SUBTEST(igt_ppgtt_smoke_huge),
1913 		SUBTEST(igt_ppgtt_sanity_check),
1914 	};
1915 	struct drm_file *file;
1916 	struct i915_gem_context *ctx;
1917 	struct i915_address_space *vm;
1918 	int err;
1919 
1920 	if (!HAS_PPGTT(i915)) {
1921 		pr_info("PPGTT not supported, skipping live-selftests\n");
1922 		return 0;
1923 	}
1924 
1925 	if (intel_gt_is_wedged(&i915->gt))
1926 		return 0;
1927 
1928 	file = mock_file(i915);
1929 	if (IS_ERR(file))
1930 		return PTR_ERR(file);
1931 
1932 	ctx = live_context(i915, file);
1933 	if (IS_ERR(ctx)) {
1934 		err = PTR_ERR(ctx);
1935 		goto out_file;
1936 	}
1937 
1938 	mutex_lock(&ctx->mutex);
1939 	vm = i915_gem_context_vm(ctx);
1940 	if (vm)
1941 		WRITE_ONCE(vm->scrub_64K, true);
1942 	mutex_unlock(&ctx->mutex);
1943 
1944 	err = i915_subtests(tests, ctx);
1945 
1946 out_file:
1947 	mock_file_free(i915, file);
1948 	return err;
1949 }
1950