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
close_objects(struct intel_memory_region * mem,struct list_head * objects)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
igt_mock_fill(void * arg)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 *
igt_object_create(struct intel_memory_region * mem,struct list_head * objects,u64 size,unsigned int flags)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
igt_object_release(struct drm_i915_gem_object * obj)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
is_contiguous(struct drm_i915_gem_object * obj)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
igt_mock_reserve(void * arg)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
igt_mock_contiguous(void * arg)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
igt_mock_splintered_region(void * arg)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
igt_mock_max_segment(void * arg)452 static int igt_mock_max_segment(void *arg)
453 {
454 struct intel_memory_region *mem = arg;
455 struct drm_i915_private *i915 = mem->i915;
456 struct i915_ttm_buddy_resource *res;
457 struct drm_i915_gem_object *obj;
458 struct drm_buddy_block *block;
459 struct drm_buddy *mm;
460 struct list_head *blocks;
461 struct scatterlist *sg;
462 I915_RND_STATE(prng);
463 LIST_HEAD(objects);
464 unsigned int max_segment;
465 unsigned int ps;
466 u64 size;
467 int err = 0;
468
469 /*
470 * While we may create very large contiguous blocks, we may need
471 * to break those down for consumption elsewhere. In particular,
472 * dma-mapping with scatterlist elements have an implicit limit of
473 * UINT_MAX on each element.
474 */
475
476 size = SZ_8G;
477 ps = PAGE_SIZE;
478 if (i915_prandom_u64_state(&prng) & 1)
479 ps = SZ_64K; /* For something like DG2 */
480
481 max_segment = round_down(UINT_MAX, ps);
482
483 mem = mock_region_create(i915, 0, size, ps, 0, 0);
484 if (IS_ERR(mem))
485 return PTR_ERR(mem);
486
487 obj = igt_object_create(mem, &objects, size, 0);
488 if (IS_ERR(obj)) {
489 err = PTR_ERR(obj);
490 goto out_put;
491 }
492
493 res = to_ttm_buddy_resource(obj->mm.res);
494 blocks = &res->blocks;
495 mm = res->mm;
496 size = 0;
497 list_for_each_entry(block, blocks, link) {
498 if (drm_buddy_block_size(mm, block) > size)
499 size = drm_buddy_block_size(mm, block);
500 }
501 if (size < max_segment) {
502 pr_err("%s: Failed to create a huge contiguous block [> %u], largest block %lld\n",
503 __func__, max_segment, size);
504 err = -EINVAL;
505 goto out_close;
506 }
507
508 for (sg = obj->mm.pages->sgl; sg; sg = sg_next(sg)) {
509 dma_addr_t daddr = sg_dma_address(sg);
510
511 if (sg->length > max_segment) {
512 pr_err("%s: Created an oversized scatterlist entry, %u > %u\n",
513 __func__, sg->length, max_segment);
514 err = -EINVAL;
515 goto out_close;
516 }
517
518 if (!IS_ALIGNED(daddr, ps)) {
519 pr_err("%s: Created an unaligned scatterlist entry, addr=%pa, ps=%u\n",
520 __func__, &daddr, ps);
521 err = -EINVAL;
522 goto out_close;
523 }
524 }
525
526 out_close:
527 close_objects(mem, &objects);
528 out_put:
529 intel_memory_region_destroy(mem);
530 return err;
531 }
532
igt_object_mappable_total(struct drm_i915_gem_object * obj)533 static u64 igt_object_mappable_total(struct drm_i915_gem_object *obj)
534 {
535 struct intel_memory_region *mr = obj->mm.region;
536 struct i915_ttm_buddy_resource *bman_res =
537 to_ttm_buddy_resource(obj->mm.res);
538 struct drm_buddy *mm = bman_res->mm;
539 struct drm_buddy_block *block;
540 u64 total;
541
542 total = 0;
543 list_for_each_entry(block, &bman_res->blocks, link) {
544 u64 start = drm_buddy_block_offset(block);
545 u64 end = start + drm_buddy_block_size(mm, block);
546
547 if (start < mr->io_size)
548 total += min_t(u64, end, mr->io_size) - start;
549 }
550
551 return total;
552 }
553
igt_mock_io_size(void * arg)554 static int igt_mock_io_size(void *arg)
555 {
556 struct intel_memory_region *mr = arg;
557 struct drm_i915_private *i915 = mr->i915;
558 struct drm_i915_gem_object *obj;
559 u64 mappable_theft_total;
560 u64 io_size;
561 u64 total;
562 u64 ps;
563 u64 rem;
564 u64 size;
565 I915_RND_STATE(prng);
566 LIST_HEAD(objects);
567 int err = 0;
568
569 ps = SZ_4K;
570 if (i915_prandom_u64_state(&prng) & 1)
571 ps = SZ_64K; /* For something like DG2 */
572
573 div64_u64_rem(i915_prandom_u64_state(&prng), SZ_8G, &total);
574 total = round_down(total, ps);
575 total = max_t(u64, total, SZ_1G);
576
577 div64_u64_rem(i915_prandom_u64_state(&prng), total - ps, &io_size);
578 io_size = round_down(io_size, ps);
579 io_size = max_t(u64, io_size, SZ_256M); /* 256M seems to be the common lower limit */
580
581 pr_info("%s with ps=%llx, io_size=%llx, total=%llx\n",
582 __func__, ps, io_size, total);
583
584 mr = mock_region_create(i915, 0, total, ps, 0, io_size);
585 if (IS_ERR(mr)) {
586 err = PTR_ERR(mr);
587 goto out_err;
588 }
589
590 mappable_theft_total = 0;
591 rem = total - io_size;
592 do {
593 div64_u64_rem(i915_prandom_u64_state(&prng), rem, &size);
594 size = round_down(size, ps);
595 size = max(size, ps);
596
597 obj = igt_object_create(mr, &objects, size,
598 I915_BO_ALLOC_GPU_ONLY);
599 if (IS_ERR(obj)) {
600 pr_err("%s TOPDOWN failed with rem=%llx, size=%llx\n",
601 __func__, rem, size);
602 err = PTR_ERR(obj);
603 goto out_close;
604 }
605
606 mappable_theft_total += igt_object_mappable_total(obj);
607 rem -= size;
608 } while (rem);
609
610 pr_info("%s mappable theft=(%lluMiB/%lluMiB), total=%lluMiB\n",
611 __func__,
612 (u64)mappable_theft_total >> 20,
613 (u64)io_size >> 20,
614 (u64)total >> 20);
615
616 /*
617 * Even if we allocate all of the non-mappable portion, we should still
618 * be able to dip into the mappable portion.
619 */
620 obj = igt_object_create(mr, &objects, io_size,
621 I915_BO_ALLOC_GPU_ONLY);
622 if (IS_ERR(obj)) {
623 pr_err("%s allocation unexpectedly failed\n", __func__);
624 err = PTR_ERR(obj);
625 goto out_close;
626 }
627
628 close_objects(mr, &objects);
629
630 rem = io_size;
631 do {
632 div64_u64_rem(i915_prandom_u64_state(&prng), rem, &size);
633 size = round_down(size, ps);
634 size = max(size, ps);
635
636 obj = igt_object_create(mr, &objects, size, 0);
637 if (IS_ERR(obj)) {
638 pr_err("%s MAPPABLE failed with rem=%llx, size=%llx\n",
639 __func__, rem, size);
640 err = PTR_ERR(obj);
641 goto out_close;
642 }
643
644 if (igt_object_mappable_total(obj) != size) {
645 pr_err("%s allocation is not mappable(size=%llx)\n",
646 __func__, size);
647 err = -EINVAL;
648 goto out_close;
649 }
650 rem -= size;
651 } while (rem);
652
653 /*
654 * We assume CPU access is required by default, which should result in a
655 * failure here, even though the non-mappable portion is free.
656 */
657 obj = igt_object_create(mr, &objects, ps, 0);
658 if (!IS_ERR(obj)) {
659 pr_err("%s allocation unexpectedly succeeded\n", __func__);
660 err = -EINVAL;
661 goto out_close;
662 }
663
664 out_close:
665 close_objects(mr, &objects);
666 intel_memory_region_destroy(mr);
667 out_err:
668 if (err == -ENOMEM)
669 err = 0;
670
671 return err;
672 }
673
igt_gpu_write_dw(struct intel_context * ce,struct i915_vma * vma,u32 dword,u32 value)674 static int igt_gpu_write_dw(struct intel_context *ce,
675 struct i915_vma *vma,
676 u32 dword,
677 u32 value)
678 {
679 return igt_gpu_fill_dw(ce, vma, dword * sizeof(u32),
680 vma->size >> PAGE_SHIFT, value);
681 }
682
igt_cpu_check(struct drm_i915_gem_object * obj,u32 dword,u32 val)683 static int igt_cpu_check(struct drm_i915_gem_object *obj, u32 dword, u32 val)
684 {
685 unsigned long n = obj->base.size >> PAGE_SHIFT;
686 u32 *ptr;
687 int err;
688
689 err = i915_gem_object_wait(obj, 0, MAX_SCHEDULE_TIMEOUT);
690 if (err)
691 return err;
692
693 ptr = i915_gem_object_pin_map(obj, I915_MAP_WC);
694 if (IS_ERR(ptr))
695 return PTR_ERR(ptr);
696
697 ptr += dword;
698 while (n--) {
699 if (*ptr != val) {
700 pr_err("base[%u]=%08x, val=%08x\n",
701 dword, *ptr, val);
702 err = -EINVAL;
703 break;
704 }
705
706 ptr += PAGE_SIZE / sizeof(*ptr);
707 }
708
709 i915_gem_object_unpin_map(obj);
710 return err;
711 }
712
igt_gpu_write(struct i915_gem_context * ctx,struct drm_i915_gem_object * obj)713 static int igt_gpu_write(struct i915_gem_context *ctx,
714 struct drm_i915_gem_object *obj)
715 {
716 struct i915_gem_engines *engines;
717 struct i915_gem_engines_iter it;
718 struct i915_address_space *vm;
719 struct intel_context *ce;
720 I915_RND_STATE(prng);
721 IGT_TIMEOUT(end_time);
722 unsigned int count;
723 struct i915_vma *vma;
724 int *order;
725 int i, n;
726 int err = 0;
727
728 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
729
730 n = 0;
731 count = 0;
732 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
733 count++;
734 if (!intel_engine_can_store_dword(ce->engine))
735 continue;
736
737 vm = ce->vm;
738 n++;
739 }
740 i915_gem_context_unlock_engines(ctx);
741 if (!n)
742 return 0;
743
744 order = i915_random_order(count * count, &prng);
745 if (!order)
746 return -ENOMEM;
747
748 vma = i915_vma_instance(obj, vm, NULL);
749 if (IS_ERR(vma)) {
750 err = PTR_ERR(vma);
751 goto out_free;
752 }
753
754 err = i915_vma_pin(vma, 0, 0, PIN_USER);
755 if (err)
756 goto out_free;
757
758 i = 0;
759 engines = i915_gem_context_lock_engines(ctx);
760 do {
761 u32 rng = prandom_u32_state(&prng);
762 u32 dword = offset_in_page(rng) / 4;
763
764 ce = engines->engines[order[i] % engines->num_engines];
765 i = (i + 1) % (count * count);
766 if (!ce || !intel_engine_can_store_dword(ce->engine))
767 continue;
768
769 err = igt_gpu_write_dw(ce, vma, dword, rng);
770 if (err)
771 break;
772
773 i915_gem_object_lock(obj, NULL);
774 err = igt_cpu_check(obj, dword, rng);
775 i915_gem_object_unlock(obj);
776 if (err)
777 break;
778 } while (!__igt_timeout(end_time, NULL));
779 i915_gem_context_unlock_engines(ctx);
780
781 out_free:
782 kfree(order);
783
784 if (err == -ENOMEM)
785 err = 0;
786
787 return err;
788 }
789
igt_lmem_create(void * arg)790 static int igt_lmem_create(void *arg)
791 {
792 struct drm_i915_private *i915 = arg;
793 struct drm_i915_gem_object *obj;
794 int err = 0;
795
796 obj = i915_gem_object_create_lmem(i915, PAGE_SIZE, 0);
797 if (IS_ERR(obj))
798 return PTR_ERR(obj);
799
800 err = i915_gem_object_pin_pages_unlocked(obj);
801 if (err)
802 goto out_put;
803
804 i915_gem_object_unpin_pages(obj);
805 out_put:
806 i915_gem_object_put(obj);
807
808 return err;
809 }
810
igt_lmem_create_with_ps(void * arg)811 static int igt_lmem_create_with_ps(void *arg)
812 {
813 struct drm_i915_private *i915 = arg;
814 int err = 0;
815 u32 ps;
816
817 for (ps = PAGE_SIZE; ps <= SZ_1G; ps <<= 1) {
818 struct drm_i915_gem_object *obj;
819 dma_addr_t daddr;
820
821 obj = __i915_gem_object_create_lmem_with_ps(i915, ps, ps, 0);
822 if (IS_ERR(obj)) {
823 err = PTR_ERR(obj);
824 if (err == -ENXIO || err == -E2BIG) {
825 pr_info("%s not enough lmem for ps(%u) err=%d\n",
826 __func__, ps, err);
827 err = 0;
828 }
829
830 break;
831 }
832
833 if (obj->base.size != ps) {
834 pr_err("%s size(%zu) != ps(%u)\n",
835 __func__, obj->base.size, ps);
836 err = -EINVAL;
837 goto out_put;
838 }
839
840 i915_gem_object_lock(obj, NULL);
841 err = i915_gem_object_pin_pages(obj);
842 if (err) {
843 if (err == -ENXIO || err == -E2BIG || err == -ENOMEM) {
844 pr_info("%s not enough lmem for ps(%u) err=%d\n",
845 __func__, ps, err);
846 err = 0;
847 }
848 goto out_put;
849 }
850
851 daddr = i915_gem_object_get_dma_address(obj, 0);
852 if (!IS_ALIGNED(daddr, ps)) {
853 pr_err("%s daddr(%pa) not aligned with ps(%u)\n",
854 __func__, &daddr, ps);
855 err = -EINVAL;
856 goto out_unpin;
857 }
858
859 out_unpin:
860 i915_gem_object_unpin_pages(obj);
861 __i915_gem_object_put_pages(obj);
862 out_put:
863 i915_gem_object_unlock(obj);
864 i915_gem_object_put(obj);
865
866 if (err)
867 break;
868 }
869
870 return err;
871 }
872
igt_lmem_create_cleared_cpu(void * arg)873 static int igt_lmem_create_cleared_cpu(void *arg)
874 {
875 struct drm_i915_private *i915 = arg;
876 I915_RND_STATE(prng);
877 IGT_TIMEOUT(end_time);
878 u32 size, i;
879 int err;
880
881 i915_gem_drain_freed_objects(i915);
882
883 size = max_t(u32, PAGE_SIZE, i915_prandom_u32_max_state(SZ_32M, &prng));
884 size = round_up(size, PAGE_SIZE);
885 i = 0;
886
887 do {
888 struct drm_i915_gem_object *obj;
889 unsigned int flags;
890 u32 dword, val;
891 void *vaddr;
892
893 /*
894 * Alternate between cleared and uncleared allocations, while
895 * also dirtying the pages each time to check that the pages are
896 * always cleared if requested, since we should get some overlap
897 * of the underlying pages, if not all, since we are the only
898 * user.
899 */
900
901 flags = I915_BO_ALLOC_CPU_CLEAR;
902 if (i & 1)
903 flags = 0;
904
905 obj = i915_gem_object_create_lmem(i915, size, flags);
906 if (IS_ERR(obj))
907 return PTR_ERR(obj);
908
909 i915_gem_object_lock(obj, NULL);
910 err = i915_gem_object_pin_pages(obj);
911 if (err)
912 goto out_put;
913
914 dword = i915_prandom_u32_max_state(PAGE_SIZE / sizeof(u32),
915 &prng);
916
917 if (flags & I915_BO_ALLOC_CPU_CLEAR) {
918 err = igt_cpu_check(obj, dword, 0);
919 if (err) {
920 pr_err("%s failed with size=%u, flags=%u\n",
921 __func__, size, flags);
922 goto out_unpin;
923 }
924 }
925
926 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WC);
927 if (IS_ERR(vaddr)) {
928 err = PTR_ERR(vaddr);
929 goto out_unpin;
930 }
931
932 val = prandom_u32_state(&prng);
933
934 memset32(vaddr, val, obj->base.size / sizeof(u32));
935
936 i915_gem_object_flush_map(obj);
937 i915_gem_object_unpin_map(obj);
938 out_unpin:
939 i915_gem_object_unpin_pages(obj);
940 __i915_gem_object_put_pages(obj);
941 out_put:
942 i915_gem_object_unlock(obj);
943 i915_gem_object_put(obj);
944
945 if (err)
946 break;
947 ++i;
948 } while (!__igt_timeout(end_time, NULL));
949
950 pr_info("%s completed (%u) iterations\n", __func__, i);
951
952 return err;
953 }
954
igt_lmem_write_gpu(void * arg)955 static int igt_lmem_write_gpu(void *arg)
956 {
957 struct drm_i915_private *i915 = arg;
958 struct drm_i915_gem_object *obj;
959 struct i915_gem_context *ctx;
960 struct file *file;
961 I915_RND_STATE(prng);
962 u32 sz;
963 int err;
964
965 file = mock_file(i915);
966 if (IS_ERR(file))
967 return PTR_ERR(file);
968
969 ctx = live_context(i915, file);
970 if (IS_ERR(ctx)) {
971 err = PTR_ERR(ctx);
972 goto out_file;
973 }
974
975 sz = round_up(prandom_u32_state(&prng) % SZ_32M, PAGE_SIZE);
976
977 obj = i915_gem_object_create_lmem(i915, sz, 0);
978 if (IS_ERR(obj)) {
979 err = PTR_ERR(obj);
980 goto out_file;
981 }
982
983 err = i915_gem_object_pin_pages_unlocked(obj);
984 if (err)
985 goto out_put;
986
987 err = igt_gpu_write(ctx, obj);
988 if (err)
989 pr_err("igt_gpu_write failed(%d)\n", err);
990
991 i915_gem_object_unpin_pages(obj);
992 out_put:
993 i915_gem_object_put(obj);
994 out_file:
995 fput(file);
996 return err;
997 }
998
999 static struct intel_engine_cs *
random_engine_class(struct drm_i915_private * i915,unsigned int class,struct rnd_state * prng)1000 random_engine_class(struct drm_i915_private *i915,
1001 unsigned int class,
1002 struct rnd_state *prng)
1003 {
1004 struct intel_engine_cs *engine;
1005 unsigned int count;
1006
1007 count = 0;
1008 for (engine = intel_engine_lookup_user(i915, class, 0);
1009 engine && engine->uabi_class == class;
1010 engine = rb_entry_safe(rb_next(&engine->uabi_node),
1011 typeof(*engine), uabi_node))
1012 count++;
1013
1014 count = i915_prandom_u32_max_state(count, prng);
1015 return intel_engine_lookup_user(i915, class, count);
1016 }
1017
igt_lmem_write_cpu(void * arg)1018 static int igt_lmem_write_cpu(void *arg)
1019 {
1020 struct drm_i915_private *i915 = arg;
1021 struct drm_i915_gem_object *obj;
1022 I915_RND_STATE(prng);
1023 IGT_TIMEOUT(end_time);
1024 u32 bytes[] = {
1025 0, /* rng placeholder */
1026 sizeof(u32),
1027 sizeof(u64),
1028 64, /* cl */
1029 PAGE_SIZE,
1030 PAGE_SIZE - sizeof(u32),
1031 PAGE_SIZE - sizeof(u64),
1032 PAGE_SIZE - 64,
1033 };
1034 struct intel_engine_cs *engine;
1035 struct i915_request *rq;
1036 u32 *vaddr;
1037 u32 sz;
1038 u32 i;
1039 int *order;
1040 int count;
1041 int err;
1042
1043 engine = random_engine_class(i915, I915_ENGINE_CLASS_COPY, &prng);
1044 if (!engine)
1045 return 0;
1046
1047 pr_info("%s: using %s\n", __func__, engine->name);
1048
1049 sz = round_up(prandom_u32_state(&prng) % SZ_32M, PAGE_SIZE);
1050 sz = max_t(u32, 2 * PAGE_SIZE, sz);
1051
1052 obj = i915_gem_object_create_lmem(i915, sz, I915_BO_ALLOC_CONTIGUOUS);
1053 if (IS_ERR(obj))
1054 return PTR_ERR(obj);
1055
1056 vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
1057 if (IS_ERR(vaddr)) {
1058 err = PTR_ERR(vaddr);
1059 goto out_put;
1060 }
1061
1062 i915_gem_object_lock(obj, NULL);
1063
1064 err = dma_resv_reserve_fences(obj->base.resv, 1);
1065 if (err) {
1066 i915_gem_object_unlock(obj);
1067 goto out_put;
1068 }
1069
1070 /* Put the pages into a known state -- from the gpu for added fun */
1071 intel_engine_pm_get(engine);
1072 err = intel_context_migrate_clear(engine->gt->migrate.context, NULL,
1073 obj->mm.pages->sgl,
1074 i915_gem_get_pat_index(i915,
1075 I915_CACHE_NONE),
1076 true, 0xdeadbeaf, &rq);
1077 if (rq) {
1078 dma_resv_add_fence(obj->base.resv, &rq->fence,
1079 DMA_RESV_USAGE_WRITE);
1080 i915_request_put(rq);
1081 }
1082
1083 intel_engine_pm_put(engine);
1084 if (!err)
1085 err = i915_gem_object_set_to_wc_domain(obj, true);
1086 i915_gem_object_unlock(obj);
1087 if (err)
1088 goto out_unpin;
1089
1090 count = ARRAY_SIZE(bytes);
1091 order = i915_random_order(count * count, &prng);
1092 if (!order) {
1093 err = -ENOMEM;
1094 goto out_unpin;
1095 }
1096
1097 /* A random multiple of u32, picked between [64, PAGE_SIZE - 64] */
1098 bytes[0] = igt_random_offset(&prng, 64, PAGE_SIZE - 64, 0, sizeof(u32));
1099 GEM_BUG_ON(!IS_ALIGNED(bytes[0], sizeof(u32)));
1100
1101 i = 0;
1102 do {
1103 u32 offset;
1104 u32 align;
1105 u32 dword;
1106 u32 size;
1107 u32 val;
1108
1109 size = bytes[order[i] % count];
1110 i = (i + 1) % (count * count);
1111
1112 align = bytes[order[i] % count];
1113 i = (i + 1) % (count * count);
1114
1115 align = max_t(u32, sizeof(u32), rounddown_pow_of_two(align));
1116
1117 offset = igt_random_offset(&prng, 0, obj->base.size,
1118 size, align);
1119
1120 val = prandom_u32_state(&prng);
1121 memset32(vaddr + offset / sizeof(u32), val ^ 0xdeadbeaf,
1122 size / sizeof(u32));
1123
1124 /*
1125 * Sample random dw -- don't waste precious time reading every
1126 * single dw.
1127 */
1128 dword = igt_random_offset(&prng, offset,
1129 offset + size,
1130 sizeof(u32), sizeof(u32));
1131 dword /= sizeof(u32);
1132 if (vaddr[dword] != (val ^ 0xdeadbeaf)) {
1133 pr_err("%s vaddr[%u]=%u, val=%u, size=%u, align=%u, offset=%u\n",
1134 __func__, dword, vaddr[dword], val ^ 0xdeadbeaf,
1135 size, align, offset);
1136 err = -EINVAL;
1137 break;
1138 }
1139 } while (!__igt_timeout(end_time, NULL));
1140
1141 out_unpin:
1142 i915_gem_object_unpin_map(obj);
1143 out_put:
1144 i915_gem_object_put(obj);
1145
1146 return err;
1147 }
1148
repr_type(u32 type)1149 static const char *repr_type(u32 type)
1150 {
1151 switch (type) {
1152 case I915_MAP_WB:
1153 return "WB";
1154 case I915_MAP_WC:
1155 return "WC";
1156 }
1157
1158 return "";
1159 }
1160
1161 static struct drm_i915_gem_object *
create_region_for_mapping(struct intel_memory_region * mr,u64 size,u32 type,void ** out_addr)1162 create_region_for_mapping(struct intel_memory_region *mr, u64 size, u32 type,
1163 void **out_addr)
1164 {
1165 struct drm_i915_gem_object *obj;
1166 void *addr;
1167
1168 obj = i915_gem_object_create_region(mr, size, 0, 0);
1169 if (IS_ERR(obj)) {
1170 if (PTR_ERR(obj) == -ENOSPC) /* Stolen memory */
1171 return ERR_PTR(-ENODEV);
1172 return obj;
1173 }
1174
1175 addr = i915_gem_object_pin_map_unlocked(obj, type);
1176 if (IS_ERR(addr)) {
1177 i915_gem_object_put(obj);
1178 if (PTR_ERR(addr) == -ENXIO)
1179 return ERR_PTR(-ENODEV);
1180 return addr;
1181 }
1182
1183 *out_addr = addr;
1184 return obj;
1185 }
1186
wrap_ktime_compare(const void * A,const void * B)1187 static int wrap_ktime_compare(const void *A, const void *B)
1188 {
1189 const ktime_t *a = A, *b = B;
1190
1191 return ktime_compare(*a, *b);
1192 }
1193
igt_memcpy_long(void * dst,const void * src,size_t size)1194 static void igt_memcpy_long(void *dst, const void *src, size_t size)
1195 {
1196 unsigned long *tmp = dst;
1197 const unsigned long *s = src;
1198
1199 size = size / sizeof(unsigned long);
1200 while (size--)
1201 *tmp++ = *s++;
1202 }
1203
igt_memcpy(void * dst,const void * src,size_t size)1204 static inline void igt_memcpy(void *dst, const void *src, size_t size)
1205 {
1206 memcpy(dst, src, size);
1207 }
1208
igt_memcpy_from_wc(void * dst,const void * src,size_t size)1209 static inline void igt_memcpy_from_wc(void *dst, const void *src, size_t size)
1210 {
1211 i915_memcpy_from_wc(dst, src, size);
1212 }
1213
_perf_memcpy(struct intel_memory_region * src_mr,struct intel_memory_region * dst_mr,u64 size,u32 src_type,u32 dst_type)1214 static int _perf_memcpy(struct intel_memory_region *src_mr,
1215 struct intel_memory_region *dst_mr,
1216 u64 size, u32 src_type, u32 dst_type)
1217 {
1218 struct drm_i915_private *i915 = src_mr->i915;
1219 const struct {
1220 const char *name;
1221 void (*copy)(void *dst, const void *src, size_t size);
1222 bool skip;
1223 } tests[] = {
1224 {
1225 "memcpy",
1226 igt_memcpy,
1227 },
1228 {
1229 "memcpy_long",
1230 igt_memcpy_long,
1231 },
1232 {
1233 "memcpy_from_wc",
1234 igt_memcpy_from_wc,
1235 !i915_has_memcpy_from_wc(),
1236 },
1237 };
1238 struct drm_i915_gem_object *src, *dst;
1239 void *src_addr, *dst_addr;
1240 int ret = 0;
1241 int i;
1242
1243 src = create_region_for_mapping(src_mr, size, src_type, &src_addr);
1244 if (IS_ERR(src)) {
1245 ret = PTR_ERR(src);
1246 goto out;
1247 }
1248
1249 dst = create_region_for_mapping(dst_mr, size, dst_type, &dst_addr);
1250 if (IS_ERR(dst)) {
1251 ret = PTR_ERR(dst);
1252 goto out_unpin_src;
1253 }
1254
1255 for (i = 0; i < ARRAY_SIZE(tests); ++i) {
1256 ktime_t t[5];
1257 int pass;
1258
1259 if (tests[i].skip)
1260 continue;
1261
1262 for (pass = 0; pass < ARRAY_SIZE(t); pass++) {
1263 ktime_t t0, t1;
1264
1265 t0 = ktime_get();
1266
1267 tests[i].copy(dst_addr, src_addr, size);
1268
1269 t1 = ktime_get();
1270 t[pass] = ktime_sub(t1, t0);
1271 }
1272
1273 sort(t, ARRAY_SIZE(t), sizeof(*t), wrap_ktime_compare, NULL);
1274 if (t[0] <= 0) {
1275 /* ignore the impossible to protect our sanity */
1276 pr_debug("Skipping %s src(%s, %s) -> dst(%s, %s) %14s %4lluKiB copy, unstable measurement [%lld, %lld]\n",
1277 __func__,
1278 src_mr->name, repr_type(src_type),
1279 dst_mr->name, repr_type(dst_type),
1280 tests[i].name, size >> 10,
1281 t[0], t[4]);
1282 continue;
1283 }
1284
1285 pr_info("%s src(%s, %s) -> dst(%s, %s) %14s %4llu KiB copy: %5lld MiB/s\n",
1286 __func__,
1287 src_mr->name, repr_type(src_type),
1288 dst_mr->name, repr_type(dst_type),
1289 tests[i].name, size >> 10,
1290 div64_u64(mul_u32_u32(4 * size,
1291 1000 * 1000 * 1000),
1292 t[1] + 2 * t[2] + t[3]) >> 20);
1293
1294 cond_resched();
1295 }
1296
1297 i915_gem_object_unpin_map(dst);
1298 i915_gem_object_put(dst);
1299 out_unpin_src:
1300 i915_gem_object_unpin_map(src);
1301 i915_gem_object_put(src);
1302
1303 i915_gem_drain_freed_objects(i915);
1304 out:
1305 if (ret == -ENODEV)
1306 ret = 0;
1307
1308 return ret;
1309 }
1310
perf_memcpy(void * arg)1311 static int perf_memcpy(void *arg)
1312 {
1313 struct drm_i915_private *i915 = arg;
1314 static const u32 types[] = {
1315 I915_MAP_WB,
1316 I915_MAP_WC,
1317 };
1318 static const u32 sizes[] = {
1319 SZ_4K,
1320 SZ_64K,
1321 SZ_4M,
1322 };
1323 struct intel_memory_region *src_mr, *dst_mr;
1324 int src_id, dst_id;
1325 int i, j, k;
1326 int ret;
1327
1328 for_each_memory_region(src_mr, i915, src_id) {
1329 for_each_memory_region(dst_mr, i915, dst_id) {
1330 for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
1331 for (j = 0; j < ARRAY_SIZE(types); ++j) {
1332 for (k = 0; k < ARRAY_SIZE(types); ++k) {
1333 ret = _perf_memcpy(src_mr,
1334 dst_mr,
1335 sizes[i],
1336 types[j],
1337 types[k]);
1338 if (ret)
1339 return ret;
1340 }
1341 }
1342 }
1343 }
1344 }
1345
1346 return 0;
1347 }
1348
intel_memory_region_mock_selftests(void)1349 int intel_memory_region_mock_selftests(void)
1350 {
1351 static const struct i915_subtest tests[] = {
1352 SUBTEST(igt_mock_reserve),
1353 SUBTEST(igt_mock_fill),
1354 SUBTEST(igt_mock_contiguous),
1355 SUBTEST(igt_mock_splintered_region),
1356 SUBTEST(igt_mock_max_segment),
1357 SUBTEST(igt_mock_io_size),
1358 };
1359 struct intel_memory_region *mem;
1360 struct drm_i915_private *i915;
1361 int err;
1362
1363 i915 = mock_gem_device();
1364 if (!i915)
1365 return -ENOMEM;
1366
1367 mem = mock_region_create(i915, 0, SZ_2G, I915_GTT_PAGE_SIZE_4K, 0, 0);
1368 if (IS_ERR(mem)) {
1369 pr_err("failed to create memory region\n");
1370 err = PTR_ERR(mem);
1371 goto out_unref;
1372 }
1373
1374 err = i915_subtests(tests, mem);
1375
1376 intel_memory_region_destroy(mem);
1377 out_unref:
1378 mock_destroy_device(i915);
1379 return err;
1380 }
1381
intel_memory_region_live_selftests(struct drm_i915_private * i915)1382 int intel_memory_region_live_selftests(struct drm_i915_private *i915)
1383 {
1384 static const struct i915_subtest tests[] = {
1385 SUBTEST(igt_lmem_create),
1386 SUBTEST(igt_lmem_create_with_ps),
1387 SUBTEST(igt_lmem_create_cleared_cpu),
1388 SUBTEST(igt_lmem_write_cpu),
1389 SUBTEST(igt_lmem_write_gpu),
1390 };
1391
1392 if (!HAS_LMEM(i915)) {
1393 pr_info("device lacks LMEM support, skipping\n");
1394 return 0;
1395 }
1396
1397 if (intel_gt_is_wedged(to_gt(i915)))
1398 return 0;
1399
1400 return i915_live_subtests(tests, i915);
1401 }
1402
intel_memory_region_perf_selftests(struct drm_i915_private * i915)1403 int intel_memory_region_perf_selftests(struct drm_i915_private *i915)
1404 {
1405 static const struct i915_subtest tests[] = {
1406 SUBTEST(perf_memcpy),
1407 };
1408
1409 if (intel_gt_is_wedged(to_gt(i915)))
1410 return 0;
1411
1412 return i915_live_subtests(tests, i915);
1413 }
1414