Loading...
Note: File does not exist in v3.1.
1/*
2 * Copyright © 2016 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 */
24
25#include <linux/list_sort.h>
26#include <linux/prime_numbers.h>
27
28#include "gem/i915_gem_context.h"
29#include "gem/selftests/mock_context.h"
30#include "gt/intel_context.h"
31#include "gt/intel_gpu_commands.h"
32
33#include "i915_random.h"
34#include "i915_selftest.h"
35
36#include "mock_drm.h"
37#include "mock_gem_device.h"
38#include "mock_gtt.h"
39#include "igt_flush_test.h"
40
41static void cleanup_freed_objects(struct drm_i915_private *i915)
42{
43 i915_gem_drain_freed_objects(i915);
44}
45
46static void fake_free_pages(struct drm_i915_gem_object *obj,
47 struct sg_table *pages)
48{
49 sg_free_table(pages);
50 kfree(pages);
51}
52
53static int fake_get_pages(struct drm_i915_gem_object *obj)
54{
55#define GFP (GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY)
56#define PFN_BIAS 0x1000
57 struct sg_table *pages;
58 struct scatterlist *sg;
59 unsigned int sg_page_sizes;
60 typeof(obj->base.size) rem;
61
62 pages = kmalloc(sizeof(*pages), GFP);
63 if (!pages)
64 return -ENOMEM;
65
66 rem = round_up(obj->base.size, BIT(31)) >> 31;
67 if (sg_alloc_table(pages, rem, GFP)) {
68 kfree(pages);
69 return -ENOMEM;
70 }
71
72 sg_page_sizes = 0;
73 rem = obj->base.size;
74 for (sg = pages->sgl; sg; sg = sg_next(sg)) {
75 unsigned long len = min_t(typeof(rem), rem, BIT(31));
76
77 GEM_BUG_ON(!len);
78 sg_set_page(sg, pfn_to_page(PFN_BIAS), len, 0);
79 sg_dma_address(sg) = page_to_phys(sg_page(sg));
80 sg_dma_len(sg) = len;
81 sg_page_sizes |= len;
82
83 rem -= len;
84 }
85 GEM_BUG_ON(rem);
86
87 __i915_gem_object_set_pages(obj, pages, sg_page_sizes);
88
89 return 0;
90#undef GFP
91}
92
93static void fake_put_pages(struct drm_i915_gem_object *obj,
94 struct sg_table *pages)
95{
96 fake_free_pages(obj, pages);
97 obj->mm.dirty = false;
98}
99
100static const struct drm_i915_gem_object_ops fake_ops = {
101 .name = "fake-gem",
102 .flags = I915_GEM_OBJECT_IS_SHRINKABLE,
103 .get_pages = fake_get_pages,
104 .put_pages = fake_put_pages,
105};
106
107static struct drm_i915_gem_object *
108fake_dma_object(struct drm_i915_private *i915, u64 size)
109{
110 static struct lock_class_key lock_class;
111 struct drm_i915_gem_object *obj;
112
113 GEM_BUG_ON(!size);
114 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
115
116 if (overflows_type(size, obj->base.size))
117 return ERR_PTR(-E2BIG);
118
119 obj = i915_gem_object_alloc();
120 if (!obj)
121 goto err;
122
123 drm_gem_private_object_init(&i915->drm, &obj->base, size);
124 i915_gem_object_init(obj, &fake_ops, &lock_class, 0);
125
126 i915_gem_object_set_volatile(obj);
127
128 obj->write_domain = I915_GEM_DOMAIN_CPU;
129 obj->read_domains = I915_GEM_DOMAIN_CPU;
130 obj->cache_level = I915_CACHE_NONE;
131
132 /* Preallocate the "backing storage" */
133 if (i915_gem_object_pin_pages_unlocked(obj))
134 goto err_obj;
135
136 i915_gem_object_unpin_pages(obj);
137 return obj;
138
139err_obj:
140 i915_gem_object_put(obj);
141err:
142 return ERR_PTR(-ENOMEM);
143}
144
145static int igt_ppgtt_alloc(void *arg)
146{
147 struct drm_i915_private *dev_priv = arg;
148 struct i915_ppgtt *ppgtt;
149 struct i915_gem_ww_ctx ww;
150 u64 size, last, limit;
151 int err = 0;
152
153 /* Allocate a ppggt and try to fill the entire range */
154
155 if (!HAS_PPGTT(dev_priv))
156 return 0;
157
158 ppgtt = i915_ppgtt_create(&dev_priv->gt);
159 if (IS_ERR(ppgtt))
160 return PTR_ERR(ppgtt);
161
162 if (!ppgtt->vm.allocate_va_range)
163 goto err_ppgtt_cleanup;
164
165 /*
166 * While we only allocate the page tables here and so we could
167 * address a much larger GTT than we could actually fit into
168 * RAM, a practical limit is the amount of physical pages in the system.
169 * This should ensure that we do not run into the oomkiller during
170 * the test and take down the machine wilfully.
171 */
172 limit = totalram_pages() << PAGE_SHIFT;
173 limit = min(ppgtt->vm.total, limit);
174
175 i915_gem_ww_ctx_init(&ww, false);
176retry:
177 err = i915_vm_lock_objects(&ppgtt->vm, &ww);
178 if (err)
179 goto err_ppgtt_cleanup;
180
181 /* Check we can allocate the entire range */
182 for (size = 4096; size <= limit; size <<= 2) {
183 struct i915_vm_pt_stash stash = {};
184
185 err = i915_vm_alloc_pt_stash(&ppgtt->vm, &stash, size);
186 if (err)
187 goto err_ppgtt_cleanup;
188
189 err = i915_vm_map_pt_stash(&ppgtt->vm, &stash);
190 if (err) {
191 i915_vm_free_pt_stash(&ppgtt->vm, &stash);
192 goto err_ppgtt_cleanup;
193 }
194
195 ppgtt->vm.allocate_va_range(&ppgtt->vm, &stash, 0, size);
196 cond_resched();
197
198 ppgtt->vm.clear_range(&ppgtt->vm, 0, size);
199
200 i915_vm_free_pt_stash(&ppgtt->vm, &stash);
201 }
202
203 /* Check we can incrementally allocate the entire range */
204 for (last = 0, size = 4096; size <= limit; last = size, size <<= 2) {
205 struct i915_vm_pt_stash stash = {};
206
207 err = i915_vm_alloc_pt_stash(&ppgtt->vm, &stash, size - last);
208 if (err)
209 goto err_ppgtt_cleanup;
210
211 err = i915_vm_map_pt_stash(&ppgtt->vm, &stash);
212 if (err) {
213 i915_vm_free_pt_stash(&ppgtt->vm, &stash);
214 goto err_ppgtt_cleanup;
215 }
216
217 ppgtt->vm.allocate_va_range(&ppgtt->vm, &stash,
218 last, size - last);
219 cond_resched();
220
221 i915_vm_free_pt_stash(&ppgtt->vm, &stash);
222 }
223
224err_ppgtt_cleanup:
225 if (err == -EDEADLK) {
226 err = i915_gem_ww_ctx_backoff(&ww);
227 if (!err)
228 goto retry;
229 }
230 i915_gem_ww_ctx_fini(&ww);
231
232 i915_vm_put(&ppgtt->vm);
233 return err;
234}
235
236static int lowlevel_hole(struct i915_address_space *vm,
237 u64 hole_start, u64 hole_end,
238 unsigned long end_time)
239{
240 I915_RND_STATE(seed_prng);
241 struct i915_vma *mock_vma;
242 unsigned int size;
243
244 mock_vma = kzalloc(sizeof(*mock_vma), GFP_KERNEL);
245 if (!mock_vma)
246 return -ENOMEM;
247
248 /* Keep creating larger objects until one cannot fit into the hole */
249 for (size = 12; (hole_end - hole_start) >> size; size++) {
250 I915_RND_SUBSTATE(prng, seed_prng);
251 struct drm_i915_gem_object *obj;
252 unsigned int *order, count, n;
253 u64 hole_size;
254
255 hole_size = (hole_end - hole_start) >> size;
256 if (hole_size > KMALLOC_MAX_SIZE / sizeof(u32))
257 hole_size = KMALLOC_MAX_SIZE / sizeof(u32);
258 count = hole_size >> 1;
259 if (!count) {
260 pr_debug("%s: hole is too small [%llx - %llx] >> %d: %lld\n",
261 __func__, hole_start, hole_end, size, hole_size);
262 break;
263 }
264
265 do {
266 order = i915_random_order(count, &prng);
267 if (order)
268 break;
269 } while (count >>= 1);
270 if (!count) {
271 kfree(mock_vma);
272 return -ENOMEM;
273 }
274 GEM_BUG_ON(!order);
275
276 GEM_BUG_ON(count * BIT_ULL(size) > vm->total);
277 GEM_BUG_ON(hole_start + count * BIT_ULL(size) > hole_end);
278
279 /* Ignore allocation failures (i.e. don't report them as
280 * a test failure) as we are purposefully allocating very
281 * large objects without checking that we have sufficient
282 * memory. We expect to hit -ENOMEM.
283 */
284
285 obj = fake_dma_object(vm->i915, BIT_ULL(size));
286 if (IS_ERR(obj)) {
287 kfree(order);
288 break;
289 }
290
291 GEM_BUG_ON(obj->base.size != BIT_ULL(size));
292
293 if (i915_gem_object_pin_pages_unlocked(obj)) {
294 i915_gem_object_put(obj);
295 kfree(order);
296 break;
297 }
298
299 for (n = 0; n < count; n++) {
300 u64 addr = hole_start + order[n] * BIT_ULL(size);
301 intel_wakeref_t wakeref;
302
303 GEM_BUG_ON(addr + BIT_ULL(size) > vm->total);
304
305 if (igt_timeout(end_time,
306 "%s timed out before %d/%d\n",
307 __func__, n, count)) {
308 hole_end = hole_start; /* quit */
309 break;
310 }
311
312 if (vm->allocate_va_range) {
313 struct i915_vm_pt_stash stash = {};
314 struct i915_gem_ww_ctx ww;
315 int err;
316
317 i915_gem_ww_ctx_init(&ww, false);
318retry:
319 err = i915_vm_lock_objects(vm, &ww);
320 if (err)
321 goto alloc_vm_end;
322
323 err = -ENOMEM;
324 if (i915_vm_alloc_pt_stash(vm, &stash,
325 BIT_ULL(size)))
326 goto alloc_vm_end;
327
328 err = i915_vm_map_pt_stash(vm, &stash);
329 if (!err)
330 vm->allocate_va_range(vm, &stash,
331 addr, BIT_ULL(size));
332 i915_vm_free_pt_stash(vm, &stash);
333alloc_vm_end:
334 if (err == -EDEADLK) {
335 err = i915_gem_ww_ctx_backoff(&ww);
336 if (!err)
337 goto retry;
338 }
339 i915_gem_ww_ctx_fini(&ww);
340
341 if (err)
342 break;
343 }
344
345 mock_vma->pages = obj->mm.pages;
346 mock_vma->node.size = BIT_ULL(size);
347 mock_vma->node.start = addr;
348
349 with_intel_runtime_pm(vm->gt->uncore->rpm, wakeref)
350 vm->insert_entries(vm, mock_vma,
351 I915_CACHE_NONE, 0);
352 }
353 count = n;
354
355 i915_random_reorder(order, count, &prng);
356 for (n = 0; n < count; n++) {
357 u64 addr = hole_start + order[n] * BIT_ULL(size);
358 intel_wakeref_t wakeref;
359
360 GEM_BUG_ON(addr + BIT_ULL(size) > vm->total);
361 with_intel_runtime_pm(vm->gt->uncore->rpm, wakeref)
362 vm->clear_range(vm, addr, BIT_ULL(size));
363 }
364
365 i915_gem_object_unpin_pages(obj);
366 i915_gem_object_put(obj);
367
368 kfree(order);
369
370 cleanup_freed_objects(vm->i915);
371 }
372
373 kfree(mock_vma);
374 return 0;
375}
376
377static void close_object_list(struct list_head *objects,
378 struct i915_address_space *vm)
379{
380 struct drm_i915_gem_object *obj, *on;
381 int ignored;
382
383 list_for_each_entry_safe(obj, on, objects, st_link) {
384 struct i915_vma *vma;
385
386 vma = i915_vma_instance(obj, vm, NULL);
387 if (!IS_ERR(vma))
388 ignored = i915_vma_unbind(vma);
389
390 list_del(&obj->st_link);
391 i915_gem_object_put(obj);
392 }
393}
394
395static int fill_hole(struct i915_address_space *vm,
396 u64 hole_start, u64 hole_end,
397 unsigned long end_time)
398{
399 const u64 hole_size = hole_end - hole_start;
400 struct drm_i915_gem_object *obj;
401 const unsigned long max_pages =
402 min_t(u64, ULONG_MAX - 1, hole_size/2 >> PAGE_SHIFT);
403 const unsigned long max_step = max(int_sqrt(max_pages), 2UL);
404 unsigned long npages, prime, flags;
405 struct i915_vma *vma;
406 LIST_HEAD(objects);
407 int err;
408
409 /* Try binding many VMA working inwards from either edge */
410
411 flags = PIN_OFFSET_FIXED | PIN_USER;
412 if (i915_is_ggtt(vm))
413 flags |= PIN_GLOBAL;
414
415 for_each_prime_number_from(prime, 2, max_step) {
416 for (npages = 1; npages <= max_pages; npages *= prime) {
417 const u64 full_size = npages << PAGE_SHIFT;
418 const struct {
419 const char *name;
420 u64 offset;
421 int step;
422 } phases[] = {
423 { "top-down", hole_end, -1, },
424 { "bottom-up", hole_start, 1, },
425 { }
426 }, *p;
427
428 obj = fake_dma_object(vm->i915, full_size);
429 if (IS_ERR(obj))
430 break;
431
432 list_add(&obj->st_link, &objects);
433
434 /* Align differing sized objects against the edges, and
435 * check we don't walk off into the void when binding
436 * them into the GTT.
437 */
438 for (p = phases; p->name; p++) {
439 u64 offset;
440
441 offset = p->offset;
442 list_for_each_entry(obj, &objects, st_link) {
443 vma = i915_vma_instance(obj, vm, NULL);
444 if (IS_ERR(vma))
445 continue;
446
447 if (p->step < 0) {
448 if (offset < hole_start + obj->base.size)
449 break;
450 offset -= obj->base.size;
451 }
452
453 err = i915_vma_pin(vma, 0, 0, offset | flags);
454 if (err) {
455 pr_err("%s(%s) pin (forward) failed with err=%d on size=%lu pages (prime=%lu), offset=%llx\n",
456 __func__, p->name, err, npages, prime, offset);
457 goto err;
458 }
459
460 if (!drm_mm_node_allocated(&vma->node) ||
461 i915_vma_misplaced(vma, 0, 0, offset | flags)) {
462 pr_err("%s(%s) (forward) insert failed: vma.node=%llx + %llx [allocated? %d], expected offset %llx\n",
463 __func__, p->name, vma->node.start, vma->node.size, drm_mm_node_allocated(&vma->node),
464 offset);
465 err = -EINVAL;
466 goto err;
467 }
468
469 i915_vma_unpin(vma);
470
471 if (p->step > 0) {
472 if (offset + obj->base.size > hole_end)
473 break;
474 offset += obj->base.size;
475 }
476 }
477
478 offset = p->offset;
479 list_for_each_entry(obj, &objects, st_link) {
480 vma = i915_vma_instance(obj, vm, NULL);
481 if (IS_ERR(vma))
482 continue;
483
484 if (p->step < 0) {
485 if (offset < hole_start + obj->base.size)
486 break;
487 offset -= obj->base.size;
488 }
489
490 if (!drm_mm_node_allocated(&vma->node) ||
491 i915_vma_misplaced(vma, 0, 0, offset | flags)) {
492 pr_err("%s(%s) (forward) moved vma.node=%llx + %llx, expected offset %llx\n",
493 __func__, p->name, vma->node.start, vma->node.size,
494 offset);
495 err = -EINVAL;
496 goto err;
497 }
498
499 err = i915_vma_unbind(vma);
500 if (err) {
501 pr_err("%s(%s) (forward) unbind of vma.node=%llx + %llx failed with err=%d\n",
502 __func__, p->name, vma->node.start, vma->node.size,
503 err);
504 goto err;
505 }
506
507 if (p->step > 0) {
508 if (offset + obj->base.size > hole_end)
509 break;
510 offset += obj->base.size;
511 }
512 }
513
514 offset = p->offset;
515 list_for_each_entry_reverse(obj, &objects, st_link) {
516 vma = i915_vma_instance(obj, vm, NULL);
517 if (IS_ERR(vma))
518 continue;
519
520 if (p->step < 0) {
521 if (offset < hole_start + obj->base.size)
522 break;
523 offset -= obj->base.size;
524 }
525
526 err = i915_vma_pin(vma, 0, 0, offset | flags);
527 if (err) {
528 pr_err("%s(%s) pin (backward) failed with err=%d on size=%lu pages (prime=%lu), offset=%llx\n",
529 __func__, p->name, err, npages, prime, offset);
530 goto err;
531 }
532
533 if (!drm_mm_node_allocated(&vma->node) ||
534 i915_vma_misplaced(vma, 0, 0, offset | flags)) {
535 pr_err("%s(%s) (backward) insert failed: vma.node=%llx + %llx [allocated? %d], expected offset %llx\n",
536 __func__, p->name, vma->node.start, vma->node.size, drm_mm_node_allocated(&vma->node),
537 offset);
538 err = -EINVAL;
539 goto err;
540 }
541
542 i915_vma_unpin(vma);
543
544 if (p->step > 0) {
545 if (offset + obj->base.size > hole_end)
546 break;
547 offset += obj->base.size;
548 }
549 }
550
551 offset = p->offset;
552 list_for_each_entry_reverse(obj, &objects, st_link) {
553 vma = i915_vma_instance(obj, vm, NULL);
554 if (IS_ERR(vma))
555 continue;
556
557 if (p->step < 0) {
558 if (offset < hole_start + obj->base.size)
559 break;
560 offset -= obj->base.size;
561 }
562
563 if (!drm_mm_node_allocated(&vma->node) ||
564 i915_vma_misplaced(vma, 0, 0, offset | flags)) {
565 pr_err("%s(%s) (backward) moved vma.node=%llx + %llx [allocated? %d], expected offset %llx\n",
566 __func__, p->name, vma->node.start, vma->node.size, drm_mm_node_allocated(&vma->node),
567 offset);
568 err = -EINVAL;
569 goto err;
570 }
571
572 err = i915_vma_unbind(vma);
573 if (err) {
574 pr_err("%s(%s) (backward) unbind of vma.node=%llx + %llx failed with err=%d\n",
575 __func__, p->name, vma->node.start, vma->node.size,
576 err);
577 goto err;
578 }
579
580 if (p->step > 0) {
581 if (offset + obj->base.size > hole_end)
582 break;
583 offset += obj->base.size;
584 }
585 }
586 }
587
588 if (igt_timeout(end_time, "%s timed out (npages=%lu, prime=%lu)\n",
589 __func__, npages, prime)) {
590 err = -EINTR;
591 goto err;
592 }
593 }
594
595 close_object_list(&objects, vm);
596 cleanup_freed_objects(vm->i915);
597 }
598
599 return 0;
600
601err:
602 close_object_list(&objects, vm);
603 return err;
604}
605
606static int walk_hole(struct i915_address_space *vm,
607 u64 hole_start, u64 hole_end,
608 unsigned long end_time)
609{
610 const u64 hole_size = hole_end - hole_start;
611 const unsigned long max_pages =
612 min_t(u64, ULONG_MAX - 1, hole_size >> PAGE_SHIFT);
613 unsigned long flags;
614 u64 size;
615
616 /* Try binding a single VMA in different positions within the hole */
617
618 flags = PIN_OFFSET_FIXED | PIN_USER;
619 if (i915_is_ggtt(vm))
620 flags |= PIN_GLOBAL;
621
622 for_each_prime_number_from(size, 1, max_pages) {
623 struct drm_i915_gem_object *obj;
624 struct i915_vma *vma;
625 u64 addr;
626 int err = 0;
627
628 obj = fake_dma_object(vm->i915, size << PAGE_SHIFT);
629 if (IS_ERR(obj))
630 break;
631
632 vma = i915_vma_instance(obj, vm, NULL);
633 if (IS_ERR(vma)) {
634 err = PTR_ERR(vma);
635 goto err_put;
636 }
637
638 for (addr = hole_start;
639 addr + obj->base.size < hole_end;
640 addr += obj->base.size) {
641 err = i915_vma_pin(vma, 0, 0, addr | flags);
642 if (err) {
643 pr_err("%s bind failed at %llx + %llx [hole %llx- %llx] with err=%d\n",
644 __func__, addr, vma->size,
645 hole_start, hole_end, err);
646 goto err_put;
647 }
648 i915_vma_unpin(vma);
649
650 if (!drm_mm_node_allocated(&vma->node) ||
651 i915_vma_misplaced(vma, 0, 0, addr | flags)) {
652 pr_err("%s incorrect at %llx + %llx\n",
653 __func__, addr, vma->size);
654 err = -EINVAL;
655 goto err_put;
656 }
657
658 err = i915_vma_unbind(vma);
659 if (err) {
660 pr_err("%s unbind failed at %llx + %llx with err=%d\n",
661 __func__, addr, vma->size, err);
662 goto err_put;
663 }
664
665 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
666
667 if (igt_timeout(end_time,
668 "%s timed out at %llx\n",
669 __func__, addr)) {
670 err = -EINTR;
671 goto err_put;
672 }
673 }
674
675err_put:
676 i915_gem_object_put(obj);
677 if (err)
678 return err;
679
680 cleanup_freed_objects(vm->i915);
681 }
682
683 return 0;
684}
685
686static int pot_hole(struct i915_address_space *vm,
687 u64 hole_start, u64 hole_end,
688 unsigned long end_time)
689{
690 struct drm_i915_gem_object *obj;
691 struct i915_vma *vma;
692 unsigned long flags;
693 unsigned int pot;
694 int err = 0;
695
696 flags = PIN_OFFSET_FIXED | PIN_USER;
697 if (i915_is_ggtt(vm))
698 flags |= PIN_GLOBAL;
699
700 obj = i915_gem_object_create_internal(vm->i915, 2 * I915_GTT_PAGE_SIZE);
701 if (IS_ERR(obj))
702 return PTR_ERR(obj);
703
704 vma = i915_vma_instance(obj, vm, NULL);
705 if (IS_ERR(vma)) {
706 err = PTR_ERR(vma);
707 goto err_obj;
708 }
709
710 /* Insert a pair of pages across every pot boundary within the hole */
711 for (pot = fls64(hole_end - 1) - 1;
712 pot > ilog2(2 * I915_GTT_PAGE_SIZE);
713 pot--) {
714 u64 step = BIT_ULL(pot);
715 u64 addr;
716
717 for (addr = round_up(hole_start + I915_GTT_PAGE_SIZE, step) - I915_GTT_PAGE_SIZE;
718 addr <= round_down(hole_end - 2*I915_GTT_PAGE_SIZE, step) - I915_GTT_PAGE_SIZE;
719 addr += step) {
720 err = i915_vma_pin(vma, 0, 0, addr | flags);
721 if (err) {
722 pr_err("%s failed to pin object at %llx in hole [%llx - %llx], with err=%d\n",
723 __func__,
724 addr,
725 hole_start, hole_end,
726 err);
727 goto err_obj;
728 }
729
730 if (!drm_mm_node_allocated(&vma->node) ||
731 i915_vma_misplaced(vma, 0, 0, addr | flags)) {
732 pr_err("%s incorrect at %llx + %llx\n",
733 __func__, addr, vma->size);
734 i915_vma_unpin(vma);
735 err = i915_vma_unbind(vma);
736 err = -EINVAL;
737 goto err_obj;
738 }
739
740 i915_vma_unpin(vma);
741 err = i915_vma_unbind(vma);
742 GEM_BUG_ON(err);
743 }
744
745 if (igt_timeout(end_time,
746 "%s timed out after %d/%d\n",
747 __func__, pot, fls64(hole_end - 1) - 1)) {
748 err = -EINTR;
749 goto err_obj;
750 }
751 }
752
753err_obj:
754 i915_gem_object_put(obj);
755 return err;
756}
757
758static int drunk_hole(struct i915_address_space *vm,
759 u64 hole_start, u64 hole_end,
760 unsigned long end_time)
761{
762 I915_RND_STATE(prng);
763 unsigned int size;
764 unsigned long flags;
765
766 flags = PIN_OFFSET_FIXED | PIN_USER;
767 if (i915_is_ggtt(vm))
768 flags |= PIN_GLOBAL;
769
770 /* Keep creating larger objects until one cannot fit into the hole */
771 for (size = 12; (hole_end - hole_start) >> size; size++) {
772 struct drm_i915_gem_object *obj;
773 unsigned int *order, count, n;
774 struct i915_vma *vma;
775 u64 hole_size;
776 int err = -ENODEV;
777
778 hole_size = (hole_end - hole_start) >> size;
779 if (hole_size > KMALLOC_MAX_SIZE / sizeof(u32))
780 hole_size = KMALLOC_MAX_SIZE / sizeof(u32);
781 count = hole_size >> 1;
782 if (!count) {
783 pr_debug("%s: hole is too small [%llx - %llx] >> %d: %lld\n",
784 __func__, hole_start, hole_end, size, hole_size);
785 break;
786 }
787
788 do {
789 order = i915_random_order(count, &prng);
790 if (order)
791 break;
792 } while (count >>= 1);
793 if (!count)
794 return -ENOMEM;
795 GEM_BUG_ON(!order);
796
797 /* Ignore allocation failures (i.e. don't report them as
798 * a test failure) as we are purposefully allocating very
799 * large objects without checking that we have sufficient
800 * memory. We expect to hit -ENOMEM.
801 */
802
803 obj = fake_dma_object(vm->i915, BIT_ULL(size));
804 if (IS_ERR(obj)) {
805 kfree(order);
806 break;
807 }
808
809 vma = i915_vma_instance(obj, vm, NULL);
810 if (IS_ERR(vma)) {
811 err = PTR_ERR(vma);
812 goto err_obj;
813 }
814
815 GEM_BUG_ON(vma->size != BIT_ULL(size));
816
817 for (n = 0; n < count; n++) {
818 u64 addr = hole_start + order[n] * BIT_ULL(size);
819
820 err = i915_vma_pin(vma, 0, 0, addr | flags);
821 if (err) {
822 pr_err("%s failed to pin object at %llx + %llx in hole [%llx - %llx], with err=%d\n",
823 __func__,
824 addr, BIT_ULL(size),
825 hole_start, hole_end,
826 err);
827 goto err_obj;
828 }
829
830 if (!drm_mm_node_allocated(&vma->node) ||
831 i915_vma_misplaced(vma, 0, 0, addr | flags)) {
832 pr_err("%s incorrect at %llx + %llx\n",
833 __func__, addr, BIT_ULL(size));
834 i915_vma_unpin(vma);
835 err = i915_vma_unbind(vma);
836 err = -EINVAL;
837 goto err_obj;
838 }
839
840 i915_vma_unpin(vma);
841 err = i915_vma_unbind(vma);
842 GEM_BUG_ON(err);
843
844 if (igt_timeout(end_time,
845 "%s timed out after %d/%d\n",
846 __func__, n, count)) {
847 err = -EINTR;
848 goto err_obj;
849 }
850 }
851
852err_obj:
853 i915_gem_object_put(obj);
854 kfree(order);
855 if (err)
856 return err;
857
858 cleanup_freed_objects(vm->i915);
859 }
860
861 return 0;
862}
863
864static int __shrink_hole(struct i915_address_space *vm,
865 u64 hole_start, u64 hole_end,
866 unsigned long end_time)
867{
868 struct drm_i915_gem_object *obj;
869 unsigned long flags = PIN_OFFSET_FIXED | PIN_USER;
870 unsigned int order = 12;
871 LIST_HEAD(objects);
872 int err = 0;
873 u64 addr;
874
875 /* Keep creating larger objects until one cannot fit into the hole */
876 for (addr = hole_start; addr < hole_end; ) {
877 struct i915_vma *vma;
878 u64 size = BIT_ULL(order++);
879
880 size = min(size, hole_end - addr);
881 obj = fake_dma_object(vm->i915, size);
882 if (IS_ERR(obj)) {
883 err = PTR_ERR(obj);
884 break;
885 }
886
887 list_add(&obj->st_link, &objects);
888
889 vma = i915_vma_instance(obj, vm, NULL);
890 if (IS_ERR(vma)) {
891 err = PTR_ERR(vma);
892 break;
893 }
894
895 GEM_BUG_ON(vma->size != size);
896
897 err = i915_vma_pin(vma, 0, 0, addr | flags);
898 if (err) {
899 pr_err("%s failed to pin object at %llx + %llx in hole [%llx - %llx], with err=%d\n",
900 __func__, addr, size, hole_start, hole_end, err);
901 break;
902 }
903
904 if (!drm_mm_node_allocated(&vma->node) ||
905 i915_vma_misplaced(vma, 0, 0, addr | flags)) {
906 pr_err("%s incorrect at %llx + %llx\n",
907 __func__, addr, size);
908 i915_vma_unpin(vma);
909 err = i915_vma_unbind(vma);
910 err = -EINVAL;
911 break;
912 }
913
914 i915_vma_unpin(vma);
915 addr += size;
916
917 /*
918 * Since we are injecting allocation faults at random intervals,
919 * wait for this allocation to complete before we change the
920 * faultinjection.
921 */
922 err = i915_vma_sync(vma);
923 if (err)
924 break;
925
926 if (igt_timeout(end_time,
927 "%s timed out at ofset %llx [%llx - %llx]\n",
928 __func__, addr, hole_start, hole_end)) {
929 err = -EINTR;
930 break;
931 }
932 }
933
934 close_object_list(&objects, vm);
935 cleanup_freed_objects(vm->i915);
936 return err;
937}
938
939static int shrink_hole(struct i915_address_space *vm,
940 u64 hole_start, u64 hole_end,
941 unsigned long end_time)
942{
943 unsigned long prime;
944 int err;
945
946 vm->fault_attr.probability = 999;
947 atomic_set(&vm->fault_attr.times, -1);
948
949 for_each_prime_number_from(prime, 0, ULONG_MAX - 1) {
950 vm->fault_attr.interval = prime;
951 err = __shrink_hole(vm, hole_start, hole_end, end_time);
952 if (err)
953 break;
954 }
955
956 memset(&vm->fault_attr, 0, sizeof(vm->fault_attr));
957
958 return err;
959}
960
961static int shrink_boom(struct i915_address_space *vm,
962 u64 hole_start, u64 hole_end,
963 unsigned long end_time)
964{
965 unsigned int sizes[] = { SZ_2M, SZ_1G };
966 struct drm_i915_gem_object *purge;
967 struct drm_i915_gem_object *explode;
968 int err;
969 int i;
970
971 /*
972 * Catch the case which shrink_hole seems to miss. The setup here
973 * requires invoking the shrinker as we do the alloc_pt/alloc_pd, while
974 * ensuring that all vma assiocated with the respective pd/pdp are
975 * unpinned at the time.
976 */
977
978 for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
979 unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
980 unsigned int size = sizes[i];
981 struct i915_vma *vma;
982
983 purge = fake_dma_object(vm->i915, size);
984 if (IS_ERR(purge))
985 return PTR_ERR(purge);
986
987 vma = i915_vma_instance(purge, vm, NULL);
988 if (IS_ERR(vma)) {
989 err = PTR_ERR(vma);
990 goto err_purge;
991 }
992
993 err = i915_vma_pin(vma, 0, 0, flags);
994 if (err)
995 goto err_purge;
996
997 /* Should now be ripe for purging */
998 i915_vma_unpin(vma);
999
1000 explode = fake_dma_object(vm->i915, size);
1001 if (IS_ERR(explode)) {
1002 err = PTR_ERR(explode);
1003 goto err_purge;
1004 }
1005
1006 vm->fault_attr.probability = 100;
1007 vm->fault_attr.interval = 1;
1008 atomic_set(&vm->fault_attr.times, -1);
1009
1010 vma = i915_vma_instance(explode, vm, NULL);
1011 if (IS_ERR(vma)) {
1012 err = PTR_ERR(vma);
1013 goto err_explode;
1014 }
1015
1016 err = i915_vma_pin(vma, 0, 0, flags | size);
1017 if (err)
1018 goto err_explode;
1019
1020 i915_vma_unpin(vma);
1021
1022 i915_gem_object_put(purge);
1023 i915_gem_object_put(explode);
1024
1025 memset(&vm->fault_attr, 0, sizeof(vm->fault_attr));
1026 cleanup_freed_objects(vm->i915);
1027 }
1028
1029 return 0;
1030
1031err_explode:
1032 i915_gem_object_put(explode);
1033err_purge:
1034 i915_gem_object_put(purge);
1035 memset(&vm->fault_attr, 0, sizeof(vm->fault_attr));
1036 return err;
1037}
1038
1039static int exercise_ppgtt(struct drm_i915_private *dev_priv,
1040 int (*func)(struct i915_address_space *vm,
1041 u64 hole_start, u64 hole_end,
1042 unsigned long end_time))
1043{
1044 struct i915_ppgtt *ppgtt;
1045 IGT_TIMEOUT(end_time);
1046 struct file *file;
1047 int err;
1048
1049 if (!HAS_FULL_PPGTT(dev_priv))
1050 return 0;
1051
1052 file = mock_file(dev_priv);
1053 if (IS_ERR(file))
1054 return PTR_ERR(file);
1055
1056 ppgtt = i915_ppgtt_create(&dev_priv->gt);
1057 if (IS_ERR(ppgtt)) {
1058 err = PTR_ERR(ppgtt);
1059 goto out_free;
1060 }
1061 GEM_BUG_ON(offset_in_page(ppgtt->vm.total));
1062 GEM_BUG_ON(!atomic_read(&ppgtt->vm.open));
1063
1064 err = func(&ppgtt->vm, 0, ppgtt->vm.total, end_time);
1065
1066 i915_vm_put(&ppgtt->vm);
1067
1068out_free:
1069 fput(file);
1070 return err;
1071}
1072
1073static int igt_ppgtt_fill(void *arg)
1074{
1075 return exercise_ppgtt(arg, fill_hole);
1076}
1077
1078static int igt_ppgtt_walk(void *arg)
1079{
1080 return exercise_ppgtt(arg, walk_hole);
1081}
1082
1083static int igt_ppgtt_pot(void *arg)
1084{
1085 return exercise_ppgtt(arg, pot_hole);
1086}
1087
1088static int igt_ppgtt_drunk(void *arg)
1089{
1090 return exercise_ppgtt(arg, drunk_hole);
1091}
1092
1093static int igt_ppgtt_lowlevel(void *arg)
1094{
1095 return exercise_ppgtt(arg, lowlevel_hole);
1096}
1097
1098static int igt_ppgtt_shrink(void *arg)
1099{
1100 return exercise_ppgtt(arg, shrink_hole);
1101}
1102
1103static int igt_ppgtt_shrink_boom(void *arg)
1104{
1105 return exercise_ppgtt(arg, shrink_boom);
1106}
1107
1108static int sort_holes(void *priv, const struct list_head *A,
1109 const struct list_head *B)
1110{
1111 struct drm_mm_node *a = list_entry(A, typeof(*a), hole_stack);
1112 struct drm_mm_node *b = list_entry(B, typeof(*b), hole_stack);
1113
1114 if (a->start < b->start)
1115 return -1;
1116 else
1117 return 1;
1118}
1119
1120static int exercise_ggtt(struct drm_i915_private *i915,
1121 int (*func)(struct i915_address_space *vm,
1122 u64 hole_start, u64 hole_end,
1123 unsigned long end_time))
1124{
1125 struct i915_ggtt *ggtt = &i915->ggtt;
1126 u64 hole_start, hole_end, last = 0;
1127 struct drm_mm_node *node;
1128 IGT_TIMEOUT(end_time);
1129 int err = 0;
1130
1131restart:
1132 list_sort(NULL, &ggtt->vm.mm.hole_stack, sort_holes);
1133 drm_mm_for_each_hole(node, &ggtt->vm.mm, hole_start, hole_end) {
1134 if (hole_start < last)
1135 continue;
1136
1137 if (ggtt->vm.mm.color_adjust)
1138 ggtt->vm.mm.color_adjust(node, 0,
1139 &hole_start, &hole_end);
1140 if (hole_start >= hole_end)
1141 continue;
1142
1143 err = func(&ggtt->vm, hole_start, hole_end, end_time);
1144 if (err)
1145 break;
1146
1147 /* As we have manipulated the drm_mm, the list may be corrupt */
1148 last = hole_end;
1149 goto restart;
1150 }
1151
1152 return err;
1153}
1154
1155static int igt_ggtt_fill(void *arg)
1156{
1157 return exercise_ggtt(arg, fill_hole);
1158}
1159
1160static int igt_ggtt_walk(void *arg)
1161{
1162 return exercise_ggtt(arg, walk_hole);
1163}
1164
1165static int igt_ggtt_pot(void *arg)
1166{
1167 return exercise_ggtt(arg, pot_hole);
1168}
1169
1170static int igt_ggtt_drunk(void *arg)
1171{
1172 return exercise_ggtt(arg, drunk_hole);
1173}
1174
1175static int igt_ggtt_lowlevel(void *arg)
1176{
1177 return exercise_ggtt(arg, lowlevel_hole);
1178}
1179
1180static int igt_ggtt_page(void *arg)
1181{
1182 const unsigned int count = PAGE_SIZE/sizeof(u32);
1183 I915_RND_STATE(prng);
1184 struct drm_i915_private *i915 = arg;
1185 struct i915_ggtt *ggtt = &i915->ggtt;
1186 struct drm_i915_gem_object *obj;
1187 intel_wakeref_t wakeref;
1188 struct drm_mm_node tmp;
1189 unsigned int *order, n;
1190 int err;
1191
1192 if (!i915_ggtt_has_aperture(ggtt))
1193 return 0;
1194
1195 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
1196 if (IS_ERR(obj))
1197 return PTR_ERR(obj);
1198
1199 err = i915_gem_object_pin_pages_unlocked(obj);
1200 if (err)
1201 goto out_free;
1202
1203 memset(&tmp, 0, sizeof(tmp));
1204 mutex_lock(&ggtt->vm.mutex);
1205 err = drm_mm_insert_node_in_range(&ggtt->vm.mm, &tmp,
1206 count * PAGE_SIZE, 0,
1207 I915_COLOR_UNEVICTABLE,
1208 0, ggtt->mappable_end,
1209 DRM_MM_INSERT_LOW);
1210 mutex_unlock(&ggtt->vm.mutex);
1211 if (err)
1212 goto out_unpin;
1213
1214 wakeref = intel_runtime_pm_get(&i915->runtime_pm);
1215
1216 for (n = 0; n < count; n++) {
1217 u64 offset = tmp.start + n * PAGE_SIZE;
1218
1219 ggtt->vm.insert_page(&ggtt->vm,
1220 i915_gem_object_get_dma_address(obj, 0),
1221 offset, I915_CACHE_NONE, 0);
1222 }
1223
1224 order = i915_random_order(count, &prng);
1225 if (!order) {
1226 err = -ENOMEM;
1227 goto out_remove;
1228 }
1229
1230 for (n = 0; n < count; n++) {
1231 u64 offset = tmp.start + order[n] * PAGE_SIZE;
1232 u32 __iomem *vaddr;
1233
1234 vaddr = io_mapping_map_atomic_wc(&ggtt->iomap, offset);
1235 iowrite32(n, vaddr + n);
1236 io_mapping_unmap_atomic(vaddr);
1237 }
1238 intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1239
1240 i915_random_reorder(order, count, &prng);
1241 for (n = 0; n < count; n++) {
1242 u64 offset = tmp.start + order[n] * PAGE_SIZE;
1243 u32 __iomem *vaddr;
1244 u32 val;
1245
1246 vaddr = io_mapping_map_atomic_wc(&ggtt->iomap, offset);
1247 val = ioread32(vaddr + n);
1248 io_mapping_unmap_atomic(vaddr);
1249
1250 if (val != n) {
1251 pr_err("insert page failed: found %d, expected %d\n",
1252 val, n);
1253 err = -EINVAL;
1254 break;
1255 }
1256 }
1257
1258 kfree(order);
1259out_remove:
1260 ggtt->vm.clear_range(&ggtt->vm, tmp.start, tmp.size);
1261 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
1262 mutex_lock(&ggtt->vm.mutex);
1263 drm_mm_remove_node(&tmp);
1264 mutex_unlock(&ggtt->vm.mutex);
1265out_unpin:
1266 i915_gem_object_unpin_pages(obj);
1267out_free:
1268 i915_gem_object_put(obj);
1269 return err;
1270}
1271
1272static void track_vma_bind(struct i915_vma *vma)
1273{
1274 struct drm_i915_gem_object *obj = vma->obj;
1275
1276 __i915_gem_object_pin_pages(obj);
1277
1278 GEM_BUG_ON(vma->pages);
1279 atomic_set(&vma->pages_count, I915_VMA_PAGES_ACTIVE);
1280 __i915_gem_object_pin_pages(obj);
1281 vma->pages = obj->mm.pages;
1282
1283 mutex_lock(&vma->vm->mutex);
1284 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
1285 mutex_unlock(&vma->vm->mutex);
1286}
1287
1288static int exercise_mock(struct drm_i915_private *i915,
1289 int (*func)(struct i915_address_space *vm,
1290 u64 hole_start, u64 hole_end,
1291 unsigned long end_time))
1292{
1293 const u64 limit = totalram_pages() << PAGE_SHIFT;
1294 struct i915_address_space *vm;
1295 struct i915_gem_context *ctx;
1296 IGT_TIMEOUT(end_time);
1297 int err;
1298
1299 ctx = mock_context(i915, "mock");
1300 if (!ctx)
1301 return -ENOMEM;
1302
1303 vm = i915_gem_context_get_vm_rcu(ctx);
1304 err = func(vm, 0, min(vm->total, limit), end_time);
1305 i915_vm_put(vm);
1306
1307 mock_context_close(ctx);
1308 return err;
1309}
1310
1311static int igt_mock_fill(void *arg)
1312{
1313 struct i915_ggtt *ggtt = arg;
1314
1315 return exercise_mock(ggtt->vm.i915, fill_hole);
1316}
1317
1318static int igt_mock_walk(void *arg)
1319{
1320 struct i915_ggtt *ggtt = arg;
1321
1322 return exercise_mock(ggtt->vm.i915, walk_hole);
1323}
1324
1325static int igt_mock_pot(void *arg)
1326{
1327 struct i915_ggtt *ggtt = arg;
1328
1329 return exercise_mock(ggtt->vm.i915, pot_hole);
1330}
1331
1332static int igt_mock_drunk(void *arg)
1333{
1334 struct i915_ggtt *ggtt = arg;
1335
1336 return exercise_mock(ggtt->vm.i915, drunk_hole);
1337}
1338
1339static int igt_gtt_reserve(void *arg)
1340{
1341 struct i915_ggtt *ggtt = arg;
1342 struct drm_i915_gem_object *obj, *on;
1343 I915_RND_STATE(prng);
1344 LIST_HEAD(objects);
1345 u64 total;
1346 int err = -ENODEV;
1347
1348 /* i915_gem_gtt_reserve() tries to reserve the precise range
1349 * for the node, and evicts if it has to. So our test checks that
1350 * it can give us the requsted space and prevent overlaps.
1351 */
1352
1353 /* Start by filling the GGTT */
1354 for (total = 0;
1355 total + 2 * I915_GTT_PAGE_SIZE <= ggtt->vm.total;
1356 total += 2 * I915_GTT_PAGE_SIZE) {
1357 struct i915_vma *vma;
1358
1359 obj = i915_gem_object_create_internal(ggtt->vm.i915,
1360 2 * PAGE_SIZE);
1361 if (IS_ERR(obj)) {
1362 err = PTR_ERR(obj);
1363 goto out;
1364 }
1365
1366 err = i915_gem_object_pin_pages_unlocked(obj);
1367 if (err) {
1368 i915_gem_object_put(obj);
1369 goto out;
1370 }
1371
1372 list_add(&obj->st_link, &objects);
1373
1374 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1375 if (IS_ERR(vma)) {
1376 err = PTR_ERR(vma);
1377 goto out;
1378 }
1379
1380 mutex_lock(&ggtt->vm.mutex);
1381 err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node,
1382 obj->base.size,
1383 total,
1384 obj->cache_level,
1385 0);
1386 mutex_unlock(&ggtt->vm.mutex);
1387 if (err) {
1388 pr_err("i915_gem_gtt_reserve (pass 1) failed at %llu/%llu with err=%d\n",
1389 total, ggtt->vm.total, err);
1390 goto out;
1391 }
1392 track_vma_bind(vma);
1393
1394 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1395 if (vma->node.start != total ||
1396 vma->node.size != 2*I915_GTT_PAGE_SIZE) {
1397 pr_err("i915_gem_gtt_reserve (pass 1) placement failed, found (%llx + %llx), expected (%llx + %llx)\n",
1398 vma->node.start, vma->node.size,
1399 total, 2*I915_GTT_PAGE_SIZE);
1400 err = -EINVAL;
1401 goto out;
1402 }
1403 }
1404
1405 /* Now we start forcing evictions */
1406 for (total = I915_GTT_PAGE_SIZE;
1407 total + 2 * I915_GTT_PAGE_SIZE <= ggtt->vm.total;
1408 total += 2 * I915_GTT_PAGE_SIZE) {
1409 struct i915_vma *vma;
1410
1411 obj = i915_gem_object_create_internal(ggtt->vm.i915,
1412 2 * PAGE_SIZE);
1413 if (IS_ERR(obj)) {
1414 err = PTR_ERR(obj);
1415 goto out;
1416 }
1417
1418 err = i915_gem_object_pin_pages_unlocked(obj);
1419 if (err) {
1420 i915_gem_object_put(obj);
1421 goto out;
1422 }
1423
1424 list_add(&obj->st_link, &objects);
1425
1426 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1427 if (IS_ERR(vma)) {
1428 err = PTR_ERR(vma);
1429 goto out;
1430 }
1431
1432 mutex_lock(&ggtt->vm.mutex);
1433 err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node,
1434 obj->base.size,
1435 total,
1436 obj->cache_level,
1437 0);
1438 mutex_unlock(&ggtt->vm.mutex);
1439 if (err) {
1440 pr_err("i915_gem_gtt_reserve (pass 2) failed at %llu/%llu with err=%d\n",
1441 total, ggtt->vm.total, err);
1442 goto out;
1443 }
1444 track_vma_bind(vma);
1445
1446 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1447 if (vma->node.start != total ||
1448 vma->node.size != 2*I915_GTT_PAGE_SIZE) {
1449 pr_err("i915_gem_gtt_reserve (pass 2) placement failed, found (%llx + %llx), expected (%llx + %llx)\n",
1450 vma->node.start, vma->node.size,
1451 total, 2*I915_GTT_PAGE_SIZE);
1452 err = -EINVAL;
1453 goto out;
1454 }
1455 }
1456
1457 /* And then try at random */
1458 list_for_each_entry_safe(obj, on, &objects, st_link) {
1459 struct i915_vma *vma;
1460 u64 offset;
1461
1462 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1463 if (IS_ERR(vma)) {
1464 err = PTR_ERR(vma);
1465 goto out;
1466 }
1467
1468 err = i915_vma_unbind(vma);
1469 if (err) {
1470 pr_err("i915_vma_unbind failed with err=%d!\n", err);
1471 goto out;
1472 }
1473
1474 offset = igt_random_offset(&prng,
1475 0, ggtt->vm.total,
1476 2 * I915_GTT_PAGE_SIZE,
1477 I915_GTT_MIN_ALIGNMENT);
1478
1479 mutex_lock(&ggtt->vm.mutex);
1480 err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node,
1481 obj->base.size,
1482 offset,
1483 obj->cache_level,
1484 0);
1485 mutex_unlock(&ggtt->vm.mutex);
1486 if (err) {
1487 pr_err("i915_gem_gtt_reserve (pass 3) failed at %llu/%llu with err=%d\n",
1488 total, ggtt->vm.total, err);
1489 goto out;
1490 }
1491 track_vma_bind(vma);
1492
1493 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1494 if (vma->node.start != offset ||
1495 vma->node.size != 2*I915_GTT_PAGE_SIZE) {
1496 pr_err("i915_gem_gtt_reserve (pass 3) placement failed, found (%llx + %llx), expected (%llx + %llx)\n",
1497 vma->node.start, vma->node.size,
1498 offset, 2*I915_GTT_PAGE_SIZE);
1499 err = -EINVAL;
1500 goto out;
1501 }
1502 }
1503
1504out:
1505 list_for_each_entry_safe(obj, on, &objects, st_link) {
1506 i915_gem_object_unpin_pages(obj);
1507 i915_gem_object_put(obj);
1508 }
1509 return err;
1510}
1511
1512static int igt_gtt_insert(void *arg)
1513{
1514 struct i915_ggtt *ggtt = arg;
1515 struct drm_i915_gem_object *obj, *on;
1516 struct drm_mm_node tmp = {};
1517 const struct invalid_insert {
1518 u64 size;
1519 u64 alignment;
1520 u64 start, end;
1521 } invalid_insert[] = {
1522 {
1523 ggtt->vm.total + I915_GTT_PAGE_SIZE, 0,
1524 0, ggtt->vm.total,
1525 },
1526 {
1527 2*I915_GTT_PAGE_SIZE, 0,
1528 0, I915_GTT_PAGE_SIZE,
1529 },
1530 {
1531 -(u64)I915_GTT_PAGE_SIZE, 0,
1532 0, 4*I915_GTT_PAGE_SIZE,
1533 },
1534 {
1535 -(u64)2*I915_GTT_PAGE_SIZE, 2*I915_GTT_PAGE_SIZE,
1536 0, 4*I915_GTT_PAGE_SIZE,
1537 },
1538 {
1539 I915_GTT_PAGE_SIZE, I915_GTT_MIN_ALIGNMENT << 1,
1540 I915_GTT_MIN_ALIGNMENT, I915_GTT_MIN_ALIGNMENT << 1,
1541 },
1542 {}
1543 }, *ii;
1544 LIST_HEAD(objects);
1545 u64 total;
1546 int err = -ENODEV;
1547
1548 /* i915_gem_gtt_insert() tries to allocate some free space in the GTT
1549 * to the node, evicting if required.
1550 */
1551
1552 /* Check a couple of obviously invalid requests */
1553 for (ii = invalid_insert; ii->size; ii++) {
1554 mutex_lock(&ggtt->vm.mutex);
1555 err = i915_gem_gtt_insert(&ggtt->vm, &tmp,
1556 ii->size, ii->alignment,
1557 I915_COLOR_UNEVICTABLE,
1558 ii->start, ii->end,
1559 0);
1560 mutex_unlock(&ggtt->vm.mutex);
1561 if (err != -ENOSPC) {
1562 pr_err("Invalid i915_gem_gtt_insert(.size=%llx, .alignment=%llx, .start=%llx, .end=%llx) succeeded (err=%d)\n",
1563 ii->size, ii->alignment, ii->start, ii->end,
1564 err);
1565 return -EINVAL;
1566 }
1567 }
1568
1569 /* Start by filling the GGTT */
1570 for (total = 0;
1571 total + I915_GTT_PAGE_SIZE <= ggtt->vm.total;
1572 total += I915_GTT_PAGE_SIZE) {
1573 struct i915_vma *vma;
1574
1575 obj = i915_gem_object_create_internal(ggtt->vm.i915,
1576 I915_GTT_PAGE_SIZE);
1577 if (IS_ERR(obj)) {
1578 err = PTR_ERR(obj);
1579 goto out;
1580 }
1581
1582 err = i915_gem_object_pin_pages_unlocked(obj);
1583 if (err) {
1584 i915_gem_object_put(obj);
1585 goto out;
1586 }
1587
1588 list_add(&obj->st_link, &objects);
1589
1590 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1591 if (IS_ERR(vma)) {
1592 err = PTR_ERR(vma);
1593 goto out;
1594 }
1595
1596 mutex_lock(&ggtt->vm.mutex);
1597 err = i915_gem_gtt_insert(&ggtt->vm, &vma->node,
1598 obj->base.size, 0, obj->cache_level,
1599 0, ggtt->vm.total,
1600 0);
1601 mutex_unlock(&ggtt->vm.mutex);
1602 if (err == -ENOSPC) {
1603 /* maxed out the GGTT space */
1604 i915_gem_object_put(obj);
1605 break;
1606 }
1607 if (err) {
1608 pr_err("i915_gem_gtt_insert (pass 1) failed at %llu/%llu with err=%d\n",
1609 total, ggtt->vm.total, err);
1610 goto out;
1611 }
1612 track_vma_bind(vma);
1613 __i915_vma_pin(vma);
1614
1615 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1616 }
1617
1618 list_for_each_entry(obj, &objects, st_link) {
1619 struct i915_vma *vma;
1620
1621 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1622 if (IS_ERR(vma)) {
1623 err = PTR_ERR(vma);
1624 goto out;
1625 }
1626
1627 if (!drm_mm_node_allocated(&vma->node)) {
1628 pr_err("VMA was unexpectedly evicted!\n");
1629 err = -EINVAL;
1630 goto out;
1631 }
1632
1633 __i915_vma_unpin(vma);
1634 }
1635
1636 /* If we then reinsert, we should find the same hole */
1637 list_for_each_entry_safe(obj, on, &objects, st_link) {
1638 struct i915_vma *vma;
1639 u64 offset;
1640
1641 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1642 if (IS_ERR(vma)) {
1643 err = PTR_ERR(vma);
1644 goto out;
1645 }
1646
1647 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1648 offset = vma->node.start;
1649
1650 err = i915_vma_unbind(vma);
1651 if (err) {
1652 pr_err("i915_vma_unbind failed with err=%d!\n", err);
1653 goto out;
1654 }
1655
1656 mutex_lock(&ggtt->vm.mutex);
1657 err = i915_gem_gtt_insert(&ggtt->vm, &vma->node,
1658 obj->base.size, 0, obj->cache_level,
1659 0, ggtt->vm.total,
1660 0);
1661 mutex_unlock(&ggtt->vm.mutex);
1662 if (err) {
1663 pr_err("i915_gem_gtt_insert (pass 2) failed at %llu/%llu with err=%d\n",
1664 total, ggtt->vm.total, err);
1665 goto out;
1666 }
1667 track_vma_bind(vma);
1668
1669 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1670 if (vma->node.start != offset) {
1671 pr_err("i915_gem_gtt_insert did not return node to its previous location (the only hole), expected address %llx, found %llx\n",
1672 offset, vma->node.start);
1673 err = -EINVAL;
1674 goto out;
1675 }
1676 }
1677
1678 /* And then force evictions */
1679 for (total = 0;
1680 total + 2 * I915_GTT_PAGE_SIZE <= ggtt->vm.total;
1681 total += 2 * I915_GTT_PAGE_SIZE) {
1682 struct i915_vma *vma;
1683
1684 obj = i915_gem_object_create_internal(ggtt->vm.i915,
1685 2 * I915_GTT_PAGE_SIZE);
1686 if (IS_ERR(obj)) {
1687 err = PTR_ERR(obj);
1688 goto out;
1689 }
1690
1691 err = i915_gem_object_pin_pages_unlocked(obj);
1692 if (err) {
1693 i915_gem_object_put(obj);
1694 goto out;
1695 }
1696
1697 list_add(&obj->st_link, &objects);
1698
1699 vma = i915_vma_instance(obj, &ggtt->vm, NULL);
1700 if (IS_ERR(vma)) {
1701 err = PTR_ERR(vma);
1702 goto out;
1703 }
1704
1705 mutex_lock(&ggtt->vm.mutex);
1706 err = i915_gem_gtt_insert(&ggtt->vm, &vma->node,
1707 obj->base.size, 0, obj->cache_level,
1708 0, ggtt->vm.total,
1709 0);
1710 mutex_unlock(&ggtt->vm.mutex);
1711 if (err) {
1712 pr_err("i915_gem_gtt_insert (pass 3) failed at %llu/%llu with err=%d\n",
1713 total, ggtt->vm.total, err);
1714 goto out;
1715 }
1716 track_vma_bind(vma);
1717
1718 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1719 }
1720
1721out:
1722 list_for_each_entry_safe(obj, on, &objects, st_link) {
1723 i915_gem_object_unpin_pages(obj);
1724 i915_gem_object_put(obj);
1725 }
1726 return err;
1727}
1728
1729int i915_gem_gtt_mock_selftests(void)
1730{
1731 static const struct i915_subtest tests[] = {
1732 SUBTEST(igt_mock_drunk),
1733 SUBTEST(igt_mock_walk),
1734 SUBTEST(igt_mock_pot),
1735 SUBTEST(igt_mock_fill),
1736 SUBTEST(igt_gtt_reserve),
1737 SUBTEST(igt_gtt_insert),
1738 };
1739 struct drm_i915_private *i915;
1740 struct i915_ggtt *ggtt;
1741 int err;
1742
1743 i915 = mock_gem_device();
1744 if (!i915)
1745 return -ENOMEM;
1746
1747 ggtt = kmalloc(sizeof(*ggtt), GFP_KERNEL);
1748 if (!ggtt) {
1749 err = -ENOMEM;
1750 goto out_put;
1751 }
1752 mock_init_ggtt(i915, ggtt);
1753
1754 err = i915_subtests(tests, ggtt);
1755
1756 mock_device_flush(i915);
1757 i915_gem_drain_freed_objects(i915);
1758 mock_fini_ggtt(ggtt);
1759 kfree(ggtt);
1760out_put:
1761 mock_destroy_device(i915);
1762 return err;
1763}
1764
1765static int context_sync(struct intel_context *ce)
1766{
1767 struct i915_request *rq;
1768 long timeout;
1769
1770 rq = intel_context_create_request(ce);
1771 if (IS_ERR(rq))
1772 return PTR_ERR(rq);
1773
1774 i915_request_get(rq);
1775 i915_request_add(rq);
1776
1777 timeout = i915_request_wait(rq, 0, HZ / 5);
1778 i915_request_put(rq);
1779
1780 return timeout < 0 ? -EIO : 0;
1781}
1782
1783static struct i915_request *
1784submit_batch(struct intel_context *ce, u64 addr)
1785{
1786 struct i915_request *rq;
1787 int err;
1788
1789 rq = intel_context_create_request(ce);
1790 if (IS_ERR(rq))
1791 return rq;
1792
1793 err = 0;
1794 if (rq->engine->emit_init_breadcrumb) /* detect a hang */
1795 err = rq->engine->emit_init_breadcrumb(rq);
1796 if (err == 0)
1797 err = rq->engine->emit_bb_start(rq, addr, 0, 0);
1798
1799 if (err == 0)
1800 i915_request_get(rq);
1801 i915_request_add(rq);
1802
1803 return err ? ERR_PTR(err) : rq;
1804}
1805
1806static u32 *spinner(u32 *batch, int i)
1807{
1808 return batch + i * 64 / sizeof(*batch) + 4;
1809}
1810
1811static void end_spin(u32 *batch, int i)
1812{
1813 *spinner(batch, i) = MI_BATCH_BUFFER_END;
1814 wmb();
1815}
1816
1817static int igt_cs_tlb(void *arg)
1818{
1819 const unsigned int count = PAGE_SIZE / 64;
1820 const unsigned int chunk_size = count * PAGE_SIZE;
1821 struct drm_i915_private *i915 = arg;
1822 struct drm_i915_gem_object *bbe, *act, *out;
1823 struct i915_gem_engines_iter it;
1824 struct i915_address_space *vm;
1825 struct i915_gem_context *ctx;
1826 struct intel_context *ce;
1827 struct i915_vma *vma;
1828 I915_RND_STATE(prng);
1829 struct file *file;
1830 unsigned int i;
1831 u32 *result;
1832 u32 *batch;
1833 int err = 0;
1834
1835 /*
1836 * Our mission here is to fool the hardware to execute something
1837 * from scratch as it has not seen the batch move (due to missing
1838 * the TLB invalidate).
1839 */
1840
1841 file = mock_file(i915);
1842 if (IS_ERR(file))
1843 return PTR_ERR(file);
1844
1845 ctx = live_context(i915, file);
1846 if (IS_ERR(ctx)) {
1847 err = PTR_ERR(ctx);
1848 goto out_unlock;
1849 }
1850
1851 vm = i915_gem_context_get_vm_rcu(ctx);
1852 if (i915_is_ggtt(vm))
1853 goto out_vm;
1854
1855 /* Create two pages; dummy we prefill the TLB, and intended */
1856 bbe = i915_gem_object_create_internal(i915, PAGE_SIZE);
1857 if (IS_ERR(bbe)) {
1858 err = PTR_ERR(bbe);
1859 goto out_vm;
1860 }
1861
1862 batch = i915_gem_object_pin_map_unlocked(bbe, I915_MAP_WC);
1863 if (IS_ERR(batch)) {
1864 err = PTR_ERR(batch);
1865 goto out_put_bbe;
1866 }
1867 memset32(batch, MI_BATCH_BUFFER_END, PAGE_SIZE / sizeof(u32));
1868 i915_gem_object_flush_map(bbe);
1869 i915_gem_object_unpin_map(bbe);
1870
1871 act = i915_gem_object_create_internal(i915, PAGE_SIZE);
1872 if (IS_ERR(act)) {
1873 err = PTR_ERR(act);
1874 goto out_put_bbe;
1875 }
1876
1877 /* Track the execution of each request by writing into different slot */
1878 batch = i915_gem_object_pin_map_unlocked(act, I915_MAP_WC);
1879 if (IS_ERR(batch)) {
1880 err = PTR_ERR(batch);
1881 goto out_put_act;
1882 }
1883 for (i = 0; i < count; i++) {
1884 u32 *cs = batch + i * 64 / sizeof(*cs);
1885 u64 addr = (vm->total - PAGE_SIZE) + i * sizeof(u32);
1886
1887 GEM_BUG_ON(GRAPHICS_VER(i915) < 6);
1888 cs[0] = MI_STORE_DWORD_IMM_GEN4;
1889 if (GRAPHICS_VER(i915) >= 8) {
1890 cs[1] = lower_32_bits(addr);
1891 cs[2] = upper_32_bits(addr);
1892 cs[3] = i;
1893 cs[4] = MI_NOOP;
1894 cs[5] = MI_BATCH_BUFFER_START_GEN8;
1895 } else {
1896 cs[1] = 0;
1897 cs[2] = lower_32_bits(addr);
1898 cs[3] = i;
1899 cs[4] = MI_NOOP;
1900 cs[5] = MI_BATCH_BUFFER_START;
1901 }
1902 }
1903
1904 out = i915_gem_object_create_internal(i915, PAGE_SIZE);
1905 if (IS_ERR(out)) {
1906 err = PTR_ERR(out);
1907 goto out_put_batch;
1908 }
1909 i915_gem_object_set_cache_coherency(out, I915_CACHING_CACHED);
1910
1911 vma = i915_vma_instance(out, vm, NULL);
1912 if (IS_ERR(vma)) {
1913 err = PTR_ERR(vma);
1914 goto out_put_out;
1915 }
1916
1917 err = i915_vma_pin(vma, 0, 0,
1918 PIN_USER |
1919 PIN_OFFSET_FIXED |
1920 (vm->total - PAGE_SIZE));
1921 if (err)
1922 goto out_put_out;
1923 GEM_BUG_ON(vma->node.start != vm->total - PAGE_SIZE);
1924
1925 result = i915_gem_object_pin_map_unlocked(out, I915_MAP_WB);
1926 if (IS_ERR(result)) {
1927 err = PTR_ERR(result);
1928 goto out_put_out;
1929 }
1930
1931 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
1932 IGT_TIMEOUT(end_time);
1933 unsigned long pass = 0;
1934
1935 if (!intel_engine_can_store_dword(ce->engine))
1936 continue;
1937
1938 while (!__igt_timeout(end_time, NULL)) {
1939 struct i915_vm_pt_stash stash = {};
1940 struct i915_request *rq;
1941 struct i915_gem_ww_ctx ww;
1942 u64 offset;
1943
1944 offset = igt_random_offset(&prng,
1945 0, vm->total - PAGE_SIZE,
1946 chunk_size, PAGE_SIZE);
1947
1948 memset32(result, STACK_MAGIC, PAGE_SIZE / sizeof(u32));
1949
1950 vma = i915_vma_instance(bbe, vm, NULL);
1951 if (IS_ERR(vma)) {
1952 err = PTR_ERR(vma);
1953 goto end;
1954 }
1955
1956 err = vma->ops->set_pages(vma);
1957 if (err)
1958 goto end;
1959
1960 i915_gem_ww_ctx_init(&ww, false);
1961retry:
1962 err = i915_vm_lock_objects(vm, &ww);
1963 if (err)
1964 goto end_ww;
1965
1966 err = i915_vm_alloc_pt_stash(vm, &stash, chunk_size);
1967 if (err)
1968 goto end_ww;
1969
1970 err = i915_vm_map_pt_stash(vm, &stash);
1971 if (!err)
1972 vm->allocate_va_range(vm, &stash, offset, chunk_size);
1973 i915_vm_free_pt_stash(vm, &stash);
1974end_ww:
1975 if (err == -EDEADLK) {
1976 err = i915_gem_ww_ctx_backoff(&ww);
1977 if (!err)
1978 goto retry;
1979 }
1980 i915_gem_ww_ctx_fini(&ww);
1981 if (err)
1982 goto end;
1983
1984 /* Prime the TLB with the dummy pages */
1985 for (i = 0; i < count; i++) {
1986 vma->node.start = offset + i * PAGE_SIZE;
1987 vm->insert_entries(vm, vma, I915_CACHE_NONE, 0);
1988
1989 rq = submit_batch(ce, vma->node.start);
1990 if (IS_ERR(rq)) {
1991 err = PTR_ERR(rq);
1992 goto end;
1993 }
1994 i915_request_put(rq);
1995 }
1996
1997 vma->ops->clear_pages(vma);
1998
1999 err = context_sync(ce);
2000 if (err) {
2001 pr_err("%s: dummy setup timed out\n",
2002 ce->engine->name);
2003 goto end;
2004 }
2005
2006 vma = i915_vma_instance(act, vm, NULL);
2007 if (IS_ERR(vma)) {
2008 err = PTR_ERR(vma);
2009 goto end;
2010 }
2011
2012 err = vma->ops->set_pages(vma);
2013 if (err)
2014 goto end;
2015
2016 /* Replace the TLB with target batches */
2017 for (i = 0; i < count; i++) {
2018 struct i915_request *rq;
2019 u32 *cs = batch + i * 64 / sizeof(*cs);
2020 u64 addr;
2021
2022 vma->node.start = offset + i * PAGE_SIZE;
2023 vm->insert_entries(vm, vma, I915_CACHE_NONE, 0);
2024
2025 addr = vma->node.start + i * 64;
2026 cs[4] = MI_NOOP;
2027 cs[6] = lower_32_bits(addr);
2028 cs[7] = upper_32_bits(addr);
2029 wmb();
2030
2031 rq = submit_batch(ce, addr);
2032 if (IS_ERR(rq)) {
2033 err = PTR_ERR(rq);
2034 goto end;
2035 }
2036
2037 /* Wait until the context chain has started */
2038 if (i == 0) {
2039 while (READ_ONCE(result[i]) &&
2040 !i915_request_completed(rq))
2041 cond_resched();
2042 } else {
2043 end_spin(batch, i - 1);
2044 }
2045
2046 i915_request_put(rq);
2047 }
2048 end_spin(batch, count - 1);
2049
2050 vma->ops->clear_pages(vma);
2051
2052 err = context_sync(ce);
2053 if (err) {
2054 pr_err("%s: writes timed out\n",
2055 ce->engine->name);
2056 goto end;
2057 }
2058
2059 for (i = 0; i < count; i++) {
2060 if (result[i] != i) {
2061 pr_err("%s: Write lost on pass %lu, at offset %llx, index %d, found %x, expected %x\n",
2062 ce->engine->name, pass,
2063 offset, i, result[i], i);
2064 err = -EINVAL;
2065 goto end;
2066 }
2067 }
2068
2069 vm->clear_range(vm, offset, chunk_size);
2070 pass++;
2071 }
2072 }
2073end:
2074 if (igt_flush_test(i915))
2075 err = -EIO;
2076 i915_gem_context_unlock_engines(ctx);
2077 i915_gem_object_unpin_map(out);
2078out_put_out:
2079 i915_gem_object_put(out);
2080out_put_batch:
2081 i915_gem_object_unpin_map(act);
2082out_put_act:
2083 i915_gem_object_put(act);
2084out_put_bbe:
2085 i915_gem_object_put(bbe);
2086out_vm:
2087 i915_vm_put(vm);
2088out_unlock:
2089 fput(file);
2090 return err;
2091}
2092
2093int i915_gem_gtt_live_selftests(struct drm_i915_private *i915)
2094{
2095 static const struct i915_subtest tests[] = {
2096 SUBTEST(igt_ppgtt_alloc),
2097 SUBTEST(igt_ppgtt_lowlevel),
2098 SUBTEST(igt_ppgtt_drunk),
2099 SUBTEST(igt_ppgtt_walk),
2100 SUBTEST(igt_ppgtt_pot),
2101 SUBTEST(igt_ppgtt_fill),
2102 SUBTEST(igt_ppgtt_shrink),
2103 SUBTEST(igt_ppgtt_shrink_boom),
2104 SUBTEST(igt_ggtt_lowlevel),
2105 SUBTEST(igt_ggtt_drunk),
2106 SUBTEST(igt_ggtt_walk),
2107 SUBTEST(igt_ggtt_pot),
2108 SUBTEST(igt_ggtt_fill),
2109 SUBTEST(igt_ggtt_page),
2110 SUBTEST(igt_cs_tlb),
2111 };
2112
2113 GEM_BUG_ON(offset_in_page(i915->ggtt.vm.total));
2114
2115 return i915_subtests(tests, i915);
2116}