2 * SPDX-License-Identifier: MIT
4 * Copyright © 2016 Intel Corporation
7 #include <linux/prime_numbers.h>
9 #include "gt/intel_engine_pm.h"
10 #include "gt/intel_gpu_commands.h"
11 #include "gt/intel_gt.h"
12 #include "gt/intel_gt_pm.h"
13 #include "gem/i915_gem_region.h"
14 #include "huge_gem_object.h"
15 #include "i915_selftest.h"
16 #include "selftests/i915_random.h"
17 #include "selftests/igt_flush_test.h"
18 #include "selftests/igt_mmap.h"
29 static u64 swizzle_bit(unsigned int bit, u64 offset)
31 return (offset & BIT_ULL(bit)) >> (bit - 6);
34 static u64 tiled_offset(const struct tile *tile, u64 v)
38 if (tile->tiling == I915_TILING_NONE)
41 y = div64_u64_rem(v, tile->stride, &x);
42 v = div64_u64_rem(y, tile->height, &y) * tile->stride * tile->height;
44 if (tile->tiling == I915_TILING_X) {
46 v += div64_u64_rem(x, tile->width, &x) << tile->size;
48 } else if (tile->width == 128) {
49 const unsigned int ytile_span = 16;
50 const unsigned int ytile_height = 512;
53 v += div64_u64_rem(x, ytile_span, &x) * ytile_height;
56 const unsigned int ytile_span = 32;
57 const unsigned int ytile_height = 256;
60 v += div64_u64_rem(x, ytile_span, &x) * ytile_height;
64 switch (tile->swizzle) {
65 case I915_BIT_6_SWIZZLE_9:
66 v ^= swizzle_bit(9, v);
68 case I915_BIT_6_SWIZZLE_9_10:
69 v ^= swizzle_bit(9, v) ^ swizzle_bit(10, v);
71 case I915_BIT_6_SWIZZLE_9_11:
72 v ^= swizzle_bit(9, v) ^ swizzle_bit(11, v);
74 case I915_BIT_6_SWIZZLE_9_10_11:
75 v ^= swizzle_bit(9, v) ^ swizzle_bit(10, v) ^ swizzle_bit(11, v);
82 static int check_partial_mapping(struct drm_i915_gem_object *obj,
83 const struct tile *tile,
84 struct rnd_state *prng)
86 const unsigned long npages = obj->base.size / PAGE_SIZE;
87 struct i915_ggtt_view view;
97 err = i915_gem_object_set_tiling(obj, tile->tiling, tile->stride);
99 pr_err("Failed to set tiling mode=%u, stride=%u, err=%d\n",
100 tile->tiling, tile->stride, err);
104 GEM_BUG_ON(i915_gem_object_get_tiling(obj) != tile->tiling);
105 GEM_BUG_ON(i915_gem_object_get_stride(obj) != tile->stride);
107 i915_gem_object_lock(obj, NULL);
108 err = i915_gem_object_set_to_gtt_domain(obj, true);
109 i915_gem_object_unlock(obj);
111 pr_err("Failed to flush to GTT write domain; err=%d\n", err);
115 page = i915_prandom_u32_max_state(npages, prng);
116 view = compute_partial_view(obj, page, MIN_CHUNK_PAGES);
118 vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, PIN_MAPPABLE);
120 pr_err("Failed to pin partial view: offset=%lu; err=%d\n",
121 page, (int)PTR_ERR(vma));
125 n = page - view.partial.offset;
126 GEM_BUG_ON(n >= view.partial.size);
128 io = i915_vma_pin_iomap(vma);
131 pr_err("Failed to iomap partial view: offset=%lu; err=%d\n",
132 page, (int)PTR_ERR(io));
137 iowrite32(page, io + n * PAGE_SIZE / sizeof(*io));
138 i915_vma_unpin_iomap(vma);
140 offset = tiled_offset(tile, page << PAGE_SHIFT);
141 if (offset >= obj->base.size)
144 intel_gt_flush_ggtt_writes(&to_i915(obj->base.dev)->gt);
146 p = i915_gem_object_get_page(obj, offset >> PAGE_SHIFT);
147 cpu = kmap(p) + offset_in_page(offset);
148 drm_clflush_virt_range(cpu, sizeof(*cpu));
149 if (*cpu != (u32)page) {
150 pr_err("Partial view for %lu [%u] (offset=%llu, size=%u [%llu, row size %u], fence=%d, tiling=%d, stride=%d) misalignment, expected write to page (%llu + %u [0x%llx]) of 0x%x, found 0x%x\n",
154 vma->size >> PAGE_SHIFT,
155 tile->tiling ? tile_row_pages(obj) : 0,
156 vma->fence ? vma->fence->id : -1, tile->tiling, tile->stride,
157 offset >> PAGE_SHIFT,
158 (unsigned int)offset_in_page(offset),
164 drm_clflush_virt_range(cpu, sizeof(*cpu));
172 static int check_partial_mappings(struct drm_i915_gem_object *obj,
173 const struct tile *tile,
174 unsigned long end_time)
176 const unsigned int nreal = obj->scratch / PAGE_SIZE;
177 const unsigned long npages = obj->base.size / PAGE_SIZE;
178 struct i915_vma *vma;
182 err = i915_gem_object_set_tiling(obj, tile->tiling, tile->stride);
184 pr_err("Failed to set tiling mode=%u, stride=%u, err=%d\n",
185 tile->tiling, tile->stride, err);
189 GEM_BUG_ON(i915_gem_object_get_tiling(obj) != tile->tiling);
190 GEM_BUG_ON(i915_gem_object_get_stride(obj) != tile->stride);
192 i915_gem_object_lock(obj, NULL);
193 err = i915_gem_object_set_to_gtt_domain(obj, true);
194 i915_gem_object_unlock(obj);
196 pr_err("Failed to flush to GTT write domain; err=%d\n", err);
200 for_each_prime_number_from(page, 1, npages) {
201 struct i915_ggtt_view view =
202 compute_partial_view(obj, page, MIN_CHUNK_PAGES);
209 GEM_BUG_ON(view.partial.size > nreal);
212 vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, PIN_MAPPABLE);
214 pr_err("Failed to pin partial view: offset=%lu; err=%d\n",
215 page, (int)PTR_ERR(vma));
219 n = page - view.partial.offset;
220 GEM_BUG_ON(n >= view.partial.size);
222 io = i915_vma_pin_iomap(vma);
225 pr_err("Failed to iomap partial view: offset=%lu; err=%d\n",
226 page, (int)PTR_ERR(io));
230 iowrite32(page, io + n * PAGE_SIZE / sizeof(*io));
231 i915_vma_unpin_iomap(vma);
233 offset = tiled_offset(tile, page << PAGE_SHIFT);
234 if (offset >= obj->base.size)
237 intel_gt_flush_ggtt_writes(&to_i915(obj->base.dev)->gt);
239 p = i915_gem_object_get_page(obj, offset >> PAGE_SHIFT);
240 cpu = kmap(p) + offset_in_page(offset);
241 drm_clflush_virt_range(cpu, sizeof(*cpu));
242 if (*cpu != (u32)page) {
243 pr_err("Partial view for %lu [%u] (offset=%llu, size=%u [%llu, row size %u], fence=%d, tiling=%d, stride=%d) misalignment, expected write to page (%llu + %u [0x%llx]) of 0x%x, found 0x%x\n",
247 vma->size >> PAGE_SHIFT,
248 tile->tiling ? tile_row_pages(obj) : 0,
249 vma->fence ? vma->fence->id : -1, tile->tiling, tile->stride,
250 offset >> PAGE_SHIFT,
251 (unsigned int)offset_in_page(offset),
257 drm_clflush_virt_range(cpu, sizeof(*cpu));
264 if (igt_timeout(end_time,
265 "%s: timed out after tiling=%d stride=%d\n",
266 __func__, tile->tiling, tile->stride))
274 setup_tile_size(struct tile *tile, struct drm_i915_private *i915)
276 if (GRAPHICS_VER(i915) <= 2) {
280 } else if (tile->tiling == I915_TILING_Y &&
281 HAS_128_BYTE_Y_TILING(i915)) {
291 if (GRAPHICS_VER(i915) < 4)
292 return 8192 / tile->width;
293 else if (GRAPHICS_VER(i915) < 7)
294 return 128 * I965_FENCE_MAX_PITCH_VAL / tile->width;
296 return 128 * GEN7_FENCE_MAX_PITCH_VAL / tile->width;
299 static int igt_partial_tiling(void *arg)
301 const unsigned int nreal = 1 << 12; /* largest tile row x2 */
302 struct drm_i915_private *i915 = arg;
303 struct drm_i915_gem_object *obj;
304 intel_wakeref_t wakeref;
308 if (!i915_ggtt_has_aperture(&i915->ggtt))
311 /* We want to check the page mapping and fencing of a large object
312 * mmapped through the GTT. The object we create is larger than can
313 * possibly be mmaped as a whole, and so we must use partial GGTT vma.
314 * We then check that a write through each partial GGTT vma ends up
315 * in the right set of pages within the object, and with the expected
316 * tiling, which we verify by manual swizzling.
319 obj = huge_gem_object(i915,
321 (1 + next_prime_number(i915->ggtt.vm.total >> PAGE_SHIFT)) << PAGE_SHIFT);
325 err = i915_gem_object_pin_pages_unlocked(obj);
327 pr_err("Failed to allocate %u pages (%lu total), err=%d\n",
328 nreal, obj->base.size / PAGE_SIZE, err);
332 wakeref = intel_runtime_pm_get(&i915->runtime_pm);
342 tile.swizzle = I915_BIT_6_SWIZZLE_NONE;
343 tile.tiling = I915_TILING_NONE;
345 err = check_partial_mappings(obj, &tile, end);
346 if (err && err != -EINTR)
350 for (tiling = I915_TILING_X; tiling <= I915_TILING_Y; tiling++) {
352 unsigned int max_pitch;
356 if (i915->quirks & QUIRK_PIN_SWIZZLED_PAGES)
358 * The swizzling pattern is actually unknown as it
359 * varies based on physical address of each page.
360 * See i915_gem_detect_bit_6_swizzle().
364 tile.tiling = tiling;
367 tile.swizzle = i915->ggtt.bit_6_swizzle_x;
370 tile.swizzle = i915->ggtt.bit_6_swizzle_y;
374 GEM_BUG_ON(tile.swizzle == I915_BIT_6_SWIZZLE_UNKNOWN);
375 if (tile.swizzle == I915_BIT_6_SWIZZLE_9_17 ||
376 tile.swizzle == I915_BIT_6_SWIZZLE_9_10_17)
379 max_pitch = setup_tile_size(&tile, i915);
381 for (pitch = max_pitch; pitch; pitch >>= 1) {
382 tile.stride = tile.width * pitch;
383 err = check_partial_mappings(obj, &tile, end);
389 if (pitch > 2 && GRAPHICS_VER(i915) >= 4) {
390 tile.stride = tile.width * (pitch - 1);
391 err = check_partial_mappings(obj, &tile, end);
398 if (pitch < max_pitch && GRAPHICS_VER(i915) >= 4) {
399 tile.stride = tile.width * (pitch + 1);
400 err = check_partial_mappings(obj, &tile, end);
408 if (GRAPHICS_VER(i915) >= 4) {
409 for_each_prime_number(pitch, max_pitch) {
410 tile.stride = tile.width * pitch;
411 err = check_partial_mappings(obj, &tile, end);
423 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
424 i915_gem_object_unpin_pages(obj);
426 i915_gem_object_put(obj);
430 static int igt_smoke_tiling(void *arg)
432 const unsigned int nreal = 1 << 12; /* largest tile row x2 */
433 struct drm_i915_private *i915 = arg;
434 struct drm_i915_gem_object *obj;
435 intel_wakeref_t wakeref;
436 I915_RND_STATE(prng);
441 if (!i915_ggtt_has_aperture(&i915->ggtt))
445 * igt_partial_tiling() does an exhastive check of partial tiling
446 * chunking, but will undoubtably run out of time. Here, we do a
447 * randomised search and hope over many runs of 1s with different
448 * seeds we will do a thorough check.
450 * Remember to look at the st_seed if we see a flip-flop in BAT!
453 if (i915->quirks & QUIRK_PIN_SWIZZLED_PAGES)
456 obj = huge_gem_object(i915,
458 (1 + next_prime_number(i915->ggtt.vm.total >> PAGE_SHIFT)) << PAGE_SHIFT);
462 err = i915_gem_object_pin_pages_unlocked(obj);
464 pr_err("Failed to allocate %u pages (%lu total), err=%d\n",
465 nreal, obj->base.size / PAGE_SIZE, err);
469 wakeref = intel_runtime_pm_get(&i915->runtime_pm);
476 i915_prandom_u32_max_state(I915_TILING_Y + 1, &prng);
477 switch (tile.tiling) {
478 case I915_TILING_NONE:
483 tile.swizzle = I915_BIT_6_SWIZZLE_NONE;
487 tile.swizzle = i915->ggtt.bit_6_swizzle_x;
490 tile.swizzle = i915->ggtt.bit_6_swizzle_y;
494 if (tile.swizzle == I915_BIT_6_SWIZZLE_9_17 ||
495 tile.swizzle == I915_BIT_6_SWIZZLE_9_10_17)
498 if (tile.tiling != I915_TILING_NONE) {
499 unsigned int max_pitch = setup_tile_size(&tile, i915);
502 i915_prandom_u32_max_state(max_pitch, &prng);
503 tile.stride = (1 + tile.stride) * tile.width;
504 if (GRAPHICS_VER(i915) < 4)
505 tile.stride = rounddown_pow_of_two(tile.stride);
508 err = check_partial_mapping(obj, &tile, &prng);
513 } while (!__igt_timeout(end, NULL));
515 pr_info("%s: Completed %lu trials\n", __func__, count);
517 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
518 i915_gem_object_unpin_pages(obj);
520 i915_gem_object_put(obj);
524 static int make_obj_busy(struct drm_i915_gem_object *obj)
526 struct drm_i915_private *i915 = to_i915(obj->base.dev);
527 struct intel_engine_cs *engine;
529 for_each_uabi_engine(engine, i915) {
530 struct i915_request *rq;
531 struct i915_vma *vma;
532 struct i915_gem_ww_ctx ww;
535 vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
539 i915_gem_ww_ctx_init(&ww, false);
541 err = i915_gem_object_lock(obj, &ww);
543 err = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_USER);
547 rq = intel_engine_create_kernel_request(engine);
553 err = i915_request_await_object(rq, vma->obj, true);
555 err = i915_vma_move_to_active(vma, rq,
558 i915_request_add(rq);
562 if (err == -EDEADLK) {
563 err = i915_gem_ww_ctx_backoff(&ww);
567 i915_gem_ww_ctx_fini(&ww);
572 i915_gem_object_put(obj); /* leave it only alive via its active ref */
576 static bool assert_mmap_offset(struct drm_i915_private *i915,
580 struct drm_i915_gem_object *obj;
581 struct i915_mmap_offset *mmo;
583 obj = i915_gem_object_create_internal(i915, size);
587 mmo = mmap_offset_attach(obj, I915_MMAP_OFFSET_GTT, NULL);
588 i915_gem_object_put(obj);
590 return PTR_ERR_OR_ZERO(mmo) == expected;
593 static void disable_retire_worker(struct drm_i915_private *i915)
595 i915_gem_driver_unregister__shrinker(i915);
596 intel_gt_pm_get(&i915->gt);
597 cancel_delayed_work_sync(&i915->gt.requests.retire_work);
600 static void restore_retire_worker(struct drm_i915_private *i915)
602 igt_flush_test(i915);
603 intel_gt_pm_put(&i915->gt);
604 i915_gem_driver_register__shrinker(i915);
607 static void mmap_offset_lock(struct drm_i915_private *i915)
608 __acquires(&i915->drm.vma_offset_manager->vm_lock)
610 write_lock(&i915->drm.vma_offset_manager->vm_lock);
613 static void mmap_offset_unlock(struct drm_i915_private *i915)
614 __releases(&i915->drm.vma_offset_manager->vm_lock)
616 write_unlock(&i915->drm.vma_offset_manager->vm_lock);
619 static int igt_mmap_offset_exhaustion(void *arg)
621 struct drm_i915_private *i915 = arg;
622 struct drm_mm *mm = &i915->drm.vma_offset_manager->vm_addr_space_mm;
623 struct drm_i915_gem_object *obj;
624 struct drm_mm_node *hole, *next;
625 struct i915_mmap_offset *mmo;
628 /* Disable background reaper */
629 disable_retire_worker(i915);
630 GEM_BUG_ON(!i915->gt.awake);
631 intel_gt_retire_requests(&i915->gt);
632 i915_gem_drain_freed_objects(i915);
634 /* Trim the device mmap space to only a page */
635 mmap_offset_lock(i915);
636 loop = 1; /* PAGE_SIZE units */
637 list_for_each_entry_safe(hole, next, &mm->hole_stack, hole_stack) {
638 struct drm_mm_node *resv;
640 resv = kzalloc(sizeof(*resv), GFP_NOWAIT);
646 resv->start = drm_mm_hole_node_start(hole) + loop;
647 resv->size = hole->hole_size - loop;
656 pr_debug("Reserving hole [%llx + %llx]\n",
657 resv->start, resv->size);
659 err = drm_mm_reserve_node(mm, resv);
661 pr_err("Failed to trim VMA manager, err=%d\n", err);
666 GEM_BUG_ON(!list_is_singular(&mm->hole_stack));
667 mmap_offset_unlock(i915);
670 if (!assert_mmap_offset(i915, PAGE_SIZE, 0)) {
671 pr_err("Unable to insert object into single page hole\n");
677 if (!assert_mmap_offset(i915, 2 * PAGE_SIZE, -ENOSPC)) {
678 pr_err("Unexpectedly succeeded in inserting too large object into single page hole\n");
683 /* Fill the hole, further allocation attempts should then fail */
684 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
690 mmo = mmap_offset_attach(obj, I915_MMAP_OFFSET_GTT, NULL);
692 pr_err("Unable to insert object into reclaimed hole\n");
697 if (!assert_mmap_offset(i915, PAGE_SIZE, -ENOSPC)) {
698 pr_err("Unexpectedly succeeded in inserting object into no holes!\n");
703 i915_gem_object_put(obj);
705 /* Now fill with busy dead objects that we expect to reap */
706 for (loop = 0; loop < 3; loop++) {
707 if (intel_gt_is_wedged(&i915->gt))
710 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
716 err = make_obj_busy(obj);
718 pr_err("[loop %d] Failed to busy the object\n", loop);
724 mmap_offset_lock(i915);
726 drm_mm_for_each_node_safe(hole, next, mm) {
727 if (hole->color != -1ul)
730 drm_mm_remove_node(hole);
733 mmap_offset_unlock(i915);
734 restore_retire_worker(i915);
737 i915_gem_object_put(obj);
741 static int gtt_set(struct drm_i915_gem_object *obj)
743 struct i915_vma *vma;
747 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, PIN_MAPPABLE);
751 intel_gt_pm_get(vma->vm->gt);
752 map = i915_vma_pin_iomap(vma);
759 memset_io(map, POISON_INUSE, obj->base.size);
760 i915_vma_unpin_iomap(vma);
763 intel_gt_pm_put(vma->vm->gt);
767 static int gtt_check(struct drm_i915_gem_object *obj)
769 struct i915_vma *vma;
773 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, PIN_MAPPABLE);
777 intel_gt_pm_get(vma->vm->gt);
778 map = i915_vma_pin_iomap(vma);
785 if (memchr_inv((void __force *)map, POISON_FREE, obj->base.size)) {
786 pr_err("%s: Write via mmap did not land in backing store (GTT)\n",
787 obj->mm.region->name);
790 i915_vma_unpin_iomap(vma);
793 intel_gt_pm_put(vma->vm->gt);
797 static int wc_set(struct drm_i915_gem_object *obj)
801 vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
803 return PTR_ERR(vaddr);
805 memset(vaddr, POISON_INUSE, obj->base.size);
806 i915_gem_object_flush_map(obj);
807 i915_gem_object_unpin_map(obj);
812 static int wc_check(struct drm_i915_gem_object *obj)
817 vaddr = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WC);
819 return PTR_ERR(vaddr);
821 if (memchr_inv(vaddr, POISON_FREE, obj->base.size)) {
822 pr_err("%s: Write via mmap did not land in backing store (WC)\n",
823 obj->mm.region->name);
826 i915_gem_object_unpin_map(obj);
831 static bool can_mmap(struct drm_i915_gem_object *obj, enum i915_mmap_type type)
833 if (type == I915_MMAP_TYPE_GTT &&
834 !i915_ggtt_has_aperture(&to_i915(obj->base.dev)->ggtt))
837 if (type != I915_MMAP_TYPE_GTT &&
838 !i915_gem_object_has_struct_page(obj) &&
839 !i915_gem_object_type_has(obj, I915_GEM_OBJECT_HAS_IOMEM))
845 static void object_set_placements(struct drm_i915_gem_object *obj,
846 struct intel_memory_region **placements,
847 unsigned int n_placements)
849 GEM_BUG_ON(!n_placements);
851 if (n_placements == 1) {
852 struct drm_i915_private *i915 = to_i915(obj->base.dev);
853 struct intel_memory_region *mr = placements[0];
855 obj->mm.placements = &i915->mm.regions[mr->id];
856 obj->mm.n_placements = 1;
858 obj->mm.placements = placements;
859 obj->mm.n_placements = n_placements;
863 #define expand32(x) (((x) << 0) | ((x) << 8) | ((x) << 16) | ((x) << 24))
864 static int __igt_mmap(struct drm_i915_private *i915,
865 struct drm_i915_gem_object *obj,
866 enum i915_mmap_type type)
868 struct i915_mmap_offset *mmo;
869 struct vm_area_struct *area;
873 if (!can_mmap(obj, type))
882 mmo = mmap_offset_attach(obj, type, NULL);
886 addr = igt_mmap_node(i915, &mmo->vma_node, 0, PROT_WRITE, MAP_SHARED);
887 if (IS_ERR_VALUE(addr))
890 pr_debug("igt_mmap(%s, %d) @ %lx\n", obj->mm.region->name, type, addr);
892 area = vma_lookup(current->mm, addr);
894 pr_err("%s: Did not create a vm_area_struct for the mmap\n",
895 obj->mm.region->name);
900 if (area->vm_private_data != mmo) {
901 pr_err("%s: vm_area_struct did not point back to our mmap_offset object!\n",
902 obj->mm.region->name);
907 for (i = 0; i < obj->base.size / sizeof(u32); i++) {
908 u32 __user *ux = u64_to_user_ptr((u64)(addr + i * sizeof(*ux)));
911 if (get_user(x, ux)) {
912 pr_err("%s: Unable to read from mmap, offset:%zd\n",
913 obj->mm.region->name, i * sizeof(x));
918 if (x != expand32(POISON_INUSE)) {
919 pr_err("%s: Read incorrect value from mmap, offset:%zd, found:%x, expected:%x\n",
920 obj->mm.region->name,
921 i * sizeof(x), x, expand32(POISON_INUSE));
926 x = expand32(POISON_FREE);
927 if (put_user(x, ux)) {
928 pr_err("%s: Unable to write to mmap, offset:%zd\n",
929 obj->mm.region->name, i * sizeof(x));
935 if (type == I915_MMAP_TYPE_GTT)
936 intel_gt_flush_ggtt_writes(&i915->gt);
940 err = gtt_check(obj);
942 vm_munmap(addr, obj->base.size);
946 static int igt_mmap(void *arg)
948 struct drm_i915_private *i915 = arg;
949 struct intel_memory_region *mr;
950 enum intel_region_id id;
952 for_each_memory_region(mr, i915, id) {
953 unsigned long sizes[] = {
960 for (i = 0; i < ARRAY_SIZE(sizes); i++) {
961 struct drm_i915_gem_object *obj;
964 obj = i915_gem_object_create_region(mr, sizes[i], 0);
965 if (obj == ERR_PTR(-ENODEV))
971 object_set_placements(obj, &mr, 1);
973 err = __igt_mmap(i915, obj, I915_MMAP_TYPE_GTT);
975 err = __igt_mmap(i915, obj, I915_MMAP_TYPE_WC);
977 i915_gem_object_put(obj);
986 static const char *repr_mmap_type(enum i915_mmap_type type)
989 case I915_MMAP_TYPE_GTT: return "gtt";
990 case I915_MMAP_TYPE_WB: return "wb";
991 case I915_MMAP_TYPE_WC: return "wc";
992 case I915_MMAP_TYPE_UC: return "uc";
993 default: return "unknown";
997 static bool can_access(const struct drm_i915_gem_object *obj)
999 return i915_gem_object_has_struct_page(obj) ||
1000 i915_gem_object_type_has(obj, I915_GEM_OBJECT_HAS_IOMEM);
1003 static int __igt_mmap_access(struct drm_i915_private *i915,
1004 struct drm_i915_gem_object *obj,
1005 enum i915_mmap_type type)
1007 struct i915_mmap_offset *mmo;
1008 unsigned long __user *ptr;
1014 memset(&A, 0xAA, sizeof(A));
1015 memset(&B, 0xBB, sizeof(B));
1017 if (!can_mmap(obj, type) || !can_access(obj))
1020 mmo = mmap_offset_attach(obj, type, NULL);
1022 return PTR_ERR(mmo);
1024 addr = igt_mmap_node(i915, &mmo->vma_node, 0, PROT_WRITE, MAP_SHARED);
1025 if (IS_ERR_VALUE(addr))
1027 ptr = (unsigned long __user *)addr;
1029 err = __put_user(A, ptr);
1031 pr_err("%s(%s): failed to write into user mmap\n",
1032 obj->mm.region->name, repr_mmap_type(type));
1036 intel_gt_flush_ggtt_writes(&i915->gt);
1038 err = access_process_vm(current, addr, &x, sizeof(x), 0);
1039 if (err != sizeof(x)) {
1040 pr_err("%s(%s): access_process_vm() read failed\n",
1041 obj->mm.region->name, repr_mmap_type(type));
1045 err = access_process_vm(current, addr, &B, sizeof(B), FOLL_WRITE);
1046 if (err != sizeof(B)) {
1047 pr_err("%s(%s): access_process_vm() write failed\n",
1048 obj->mm.region->name, repr_mmap_type(type));
1052 intel_gt_flush_ggtt_writes(&i915->gt);
1054 err = __get_user(y, ptr);
1056 pr_err("%s(%s): failed to read from user mmap\n",
1057 obj->mm.region->name, repr_mmap_type(type));
1061 if (x != A || y != B) {
1062 pr_err("%s(%s): failed to read/write values, found (%lx, %lx)\n",
1063 obj->mm.region->name, repr_mmap_type(type),
1070 vm_munmap(addr, obj->base.size);
1074 static int igt_mmap_access(void *arg)
1076 struct drm_i915_private *i915 = arg;
1077 struct intel_memory_region *mr;
1078 enum intel_region_id id;
1080 for_each_memory_region(mr, i915, id) {
1081 struct drm_i915_gem_object *obj;
1084 obj = i915_gem_object_create_region(mr, PAGE_SIZE, 0);
1085 if (obj == ERR_PTR(-ENODEV))
1089 return PTR_ERR(obj);
1091 object_set_placements(obj, &mr, 1);
1093 err = __igt_mmap_access(i915, obj, I915_MMAP_TYPE_GTT);
1095 err = __igt_mmap_access(i915, obj, I915_MMAP_TYPE_WB);
1097 err = __igt_mmap_access(i915, obj, I915_MMAP_TYPE_WC);
1099 err = __igt_mmap_access(i915, obj, I915_MMAP_TYPE_UC);
1101 i915_gem_object_put(obj);
1109 static int __igt_mmap_gpu(struct drm_i915_private *i915,
1110 struct drm_i915_gem_object *obj,
1111 enum i915_mmap_type type)
1113 struct intel_engine_cs *engine;
1114 struct i915_mmap_offset *mmo;
1121 * Verify that the mmap access into the backing store aligns with
1122 * that of the GPU, i.e. that mmap is indeed writing into the same
1123 * page as being read by the GPU.
1126 if (!can_mmap(obj, type))
1135 mmo = mmap_offset_attach(obj, type, NULL);
1137 return PTR_ERR(mmo);
1139 addr = igt_mmap_node(i915, &mmo->vma_node, 0, PROT_WRITE, MAP_SHARED);
1140 if (IS_ERR_VALUE(addr))
1143 ux = u64_to_user_ptr((u64)addr);
1144 bbe = MI_BATCH_BUFFER_END;
1145 if (put_user(bbe, ux)) {
1146 pr_err("%s: Unable to write to mmap\n", obj->mm.region->name);
1151 if (type == I915_MMAP_TYPE_GTT)
1152 intel_gt_flush_ggtt_writes(&i915->gt);
1154 for_each_uabi_engine(engine, i915) {
1155 struct i915_request *rq;
1156 struct i915_vma *vma;
1157 struct i915_gem_ww_ctx ww;
1159 vma = i915_vma_instance(obj, engine->kernel_context->vm, NULL);
1165 i915_gem_ww_ctx_init(&ww, false);
1167 err = i915_gem_object_lock(obj, &ww);
1169 err = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_USER);
1173 rq = i915_request_create(engine->kernel_context);
1179 err = i915_request_await_object(rq, vma->obj, false);
1181 err = i915_vma_move_to_active(vma, rq, 0);
1183 err = engine->emit_bb_start(rq, vma->node.start, 0, 0);
1184 i915_request_get(rq);
1185 i915_request_add(rq);
1187 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
1188 struct drm_printer p =
1189 drm_info_printer(engine->i915->drm.dev);
1191 pr_err("%s(%s, %s): Failed to execute batch\n",
1192 __func__, engine->name, obj->mm.region->name);
1193 intel_engine_dump(engine, &p,
1194 "%s\n", engine->name);
1196 intel_gt_set_wedged(engine->gt);
1199 i915_request_put(rq);
1202 i915_vma_unpin(vma);
1204 if (err == -EDEADLK) {
1205 err = i915_gem_ww_ctx_backoff(&ww);
1209 i915_gem_ww_ctx_fini(&ww);
1215 vm_munmap(addr, obj->base.size);
1219 static int igt_mmap_gpu(void *arg)
1221 struct drm_i915_private *i915 = arg;
1222 struct intel_memory_region *mr;
1223 enum intel_region_id id;
1225 for_each_memory_region(mr, i915, id) {
1226 struct drm_i915_gem_object *obj;
1229 obj = i915_gem_object_create_region(mr, PAGE_SIZE, 0);
1230 if (obj == ERR_PTR(-ENODEV))
1234 return PTR_ERR(obj);
1236 object_set_placements(obj, &mr, 1);
1238 err = __igt_mmap_gpu(i915, obj, I915_MMAP_TYPE_GTT);
1240 err = __igt_mmap_gpu(i915, obj, I915_MMAP_TYPE_WC);
1242 i915_gem_object_put(obj);
1250 static int check_present_pte(pte_t *pte, unsigned long addr, void *data)
1252 if (!pte_present(*pte) || pte_none(*pte)) {
1253 pr_err("missing PTE:%lx\n",
1254 (addr - (unsigned long)data) >> PAGE_SHIFT);
1261 static int check_absent_pte(pte_t *pte, unsigned long addr, void *data)
1263 if (pte_present(*pte) && !pte_none(*pte)) {
1264 pr_err("present PTE:%lx; expected to be revoked\n",
1265 (addr - (unsigned long)data) >> PAGE_SHIFT);
1272 static int check_present(unsigned long addr, unsigned long len)
1274 return apply_to_page_range(current->mm, addr, len,
1275 check_present_pte, (void *)addr);
1278 static int check_absent(unsigned long addr, unsigned long len)
1280 return apply_to_page_range(current->mm, addr, len,
1281 check_absent_pte, (void *)addr);
1284 static int prefault_range(u64 start, u64 len)
1286 const char __user *addr, *end;
1287 char __maybe_unused c;
1290 addr = u64_to_user_ptr(start);
1293 for (; addr < end; addr += PAGE_SIZE) {
1294 err = __get_user(c, addr);
1299 return __get_user(c, end - 1);
1302 static int __igt_mmap_revoke(struct drm_i915_private *i915,
1303 struct drm_i915_gem_object *obj,
1304 enum i915_mmap_type type)
1306 struct i915_mmap_offset *mmo;
1310 if (!can_mmap(obj, type))
1313 mmo = mmap_offset_attach(obj, type, NULL);
1315 return PTR_ERR(mmo);
1317 addr = igt_mmap_node(i915, &mmo->vma_node, 0, PROT_WRITE, MAP_SHARED);
1318 if (IS_ERR_VALUE(addr))
1321 err = prefault_range(addr, obj->base.size);
1325 err = check_present(addr, obj->base.size);
1327 pr_err("%s: was not present\n", obj->mm.region->name);
1332 * After unbinding the object from the GGTT, its address may be reused
1333 * for other objects. Ergo we have to revoke the previous mmap PTE
1334 * access as it no longer points to the same object.
1336 err = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE);
1338 pr_err("Failed to unbind object!\n");
1342 if (type != I915_MMAP_TYPE_GTT) {
1343 i915_gem_object_lock(obj, NULL);
1344 __i915_gem_object_put_pages(obj);
1345 i915_gem_object_unlock(obj);
1346 if (i915_gem_object_has_pages(obj)) {
1347 pr_err("Failed to put-pages object!\n");
1353 err = check_absent(addr, obj->base.size);
1355 pr_err("%s: was not absent\n", obj->mm.region->name);
1360 vm_munmap(addr, obj->base.size);
1364 static int igt_mmap_revoke(void *arg)
1366 struct drm_i915_private *i915 = arg;
1367 struct intel_memory_region *mr;
1368 enum intel_region_id id;
1370 for_each_memory_region(mr, i915, id) {
1371 struct drm_i915_gem_object *obj;
1374 obj = i915_gem_object_create_region(mr, PAGE_SIZE, 0);
1375 if (obj == ERR_PTR(-ENODEV))
1379 return PTR_ERR(obj);
1381 object_set_placements(obj, &mr, 1);
1383 err = __igt_mmap_revoke(i915, obj, I915_MMAP_TYPE_GTT);
1385 err = __igt_mmap_revoke(i915, obj, I915_MMAP_TYPE_WC);
1387 i915_gem_object_put(obj);
1395 int i915_gem_mman_live_selftests(struct drm_i915_private *i915)
1397 static const struct i915_subtest tests[] = {
1398 SUBTEST(igt_partial_tiling),
1399 SUBTEST(igt_smoke_tiling),
1400 SUBTEST(igt_mmap_offset_exhaustion),
1402 SUBTEST(igt_mmap_access),
1403 SUBTEST(igt_mmap_revoke),
1404 SUBTEST(igt_mmap_gpu),
1407 return i915_subtests(tests, i915);