drm/i915/selftests: huge page tests
authorMatthew Auld <matthew.auld@intel.com>
Fri, 6 Oct 2017 22:18:29 +0000 (23:18 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Sat, 7 Oct 2017 09:12:00 +0000 (10:12 +0100)
v2: mock test page support configurations and add MI_STORE_DWORD test

v3: run all mockable huge page tests on all platforms via the mock_device

v4: add pin_update regression test
    various improvements suggested by Chris

v5: fix issues reported by kbuild
    test single sg spanning multiple page sizes
    don't explode when running the live-tests through the appgtt

v6: lots of improvements from Chris

v7: run on each engine for igt_write_huge
    add simple tmpfs fallback test

v8: size_t is bad
    don't break the i386 build

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171006145041.21673-18-matthew.auld@intel.com
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171006221833.32439-17-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_object.h
drivers/gpu/drm/i915/selftests/huge_pages.c [new file with mode: 0644]
drivers/gpu/drm/i915/selftests/i915_live_selftests.h
drivers/gpu/drm/i915/selftests/i915_mock_selftests.h

index 34398696824c65bd84df0dc7008159426035239c..f8c3ac1c8c671e0c74ff5dc74e89a09f26aa78dc 100644 (file)
@@ -5412,6 +5412,7 @@ err_unlock:
 #include "selftests/scatterlist.c"
 #include "selftests/mock_gem_device.c"
 #include "selftests/huge_gem_object.c"
+#include "selftests/huge_pages.c"
 #include "selftests/i915_gem_object.c"
 #include "selftests/i915_gem_coherency.c"
 #endif
index e4e6dd93889d07de1c45622edfced77084f70bd5..956c911c2cbffe120e0ae05904fc37d4d22f4fd7 100644 (file)
@@ -196,6 +196,8 @@ struct drm_i915_gem_object {
                        unsigned int gtt;
                } page_sizes;
 
+               I915_SELFTEST_DECLARE(unsigned int page_mask);
+
                struct i915_gem_object_page_iter {
                        struct scatterlist *sg_pos;
                        unsigned int sg_idx; /* in pages, but 32bit eek! */
diff --git a/drivers/gpu/drm/i915/selftests/huge_pages.c b/drivers/gpu/drm/i915/selftests/huge_pages.c
new file mode 100644 (file)
index 0000000..b849588
--- /dev/null
@@ -0,0 +1,1715 @@
+/*
+ * Copyright © 2017 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ */
+
+#include "../i915_selftest.h"
+
+#include <linux/prime_numbers.h>
+
+#include "mock_drm.h"
+
+static const unsigned int page_sizes[] = {
+       I915_GTT_PAGE_SIZE_2M,
+       I915_GTT_PAGE_SIZE_64K,
+       I915_GTT_PAGE_SIZE_4K,
+};
+
+static unsigned int get_largest_page_size(struct drm_i915_private *i915,
+                                         u64 rem)
+{
+       int i;
+
+       for (i = 0; i < ARRAY_SIZE(page_sizes); ++i) {
+               unsigned int page_size = page_sizes[i];
+
+               if (HAS_PAGE_SIZES(i915, page_size) && rem >= page_size)
+                       return page_size;
+       }
+
+       return 0;
+}
+
+static void huge_pages_free_pages(struct sg_table *st)
+{
+       struct scatterlist *sg;
+
+       for (sg = st->sgl; sg; sg = __sg_next(sg)) {
+               if (sg_page(sg))
+                       __free_pages(sg_page(sg), get_order(sg->length));
+       }
+
+       sg_free_table(st);
+       kfree(st);
+}
+
+static int get_huge_pages(struct drm_i915_gem_object *obj)
+{
+#define GFP (GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY)
+       unsigned int page_mask = obj->mm.page_mask;
+       struct sg_table *st;
+       struct scatterlist *sg;
+       unsigned int sg_mask;
+       u64 rem;
+
+       st = kmalloc(sizeof(*st), GFP);
+       if (!st)
+               return -ENOMEM;
+
+       if (sg_alloc_table(st, obj->base.size >> PAGE_SHIFT, GFP)) {
+               kfree(st);
+               return -ENOMEM;
+       }
+
+       rem = obj->base.size;
+       sg = st->sgl;
+       st->nents = 0;
+       sg_mask = 0;
+
+       /*
+        * Our goal here is simple, we want to greedily fill the object from
+        * largest to smallest page-size, while ensuring that we use *every*
+        * page-size as per the given page-mask.
+        */
+       do {
+               unsigned int bit = ilog2(page_mask);
+               unsigned int page_size = BIT(bit);
+               int order = get_order(page_size);
+
+               do {
+                       struct page *page;
+
+                       GEM_BUG_ON(order >= MAX_ORDER);
+                       page = alloc_pages(GFP | __GFP_ZERO, order);
+                       if (!page)
+                               goto err;
+
+                       sg_set_page(sg, page, page_size, 0);
+                       sg_mask |= page_size;
+                       st->nents++;
+
+                       rem -= page_size;
+                       if (!rem) {
+                               sg_mark_end(sg);
+                               break;
+                       }
+
+                       sg = __sg_next(sg);
+               } while ((rem - ((page_size-1) & page_mask)) >= page_size);
+
+               page_mask &= (page_size-1);
+       } while (page_mask);
+
+       if (i915_gem_gtt_prepare_pages(obj, st))
+               goto err;
+
+       obj->mm.madv = I915_MADV_DONTNEED;
+
+       GEM_BUG_ON(sg_mask != obj->mm.page_mask);
+       __i915_gem_object_set_pages(obj, st, sg_mask);
+
+       return 0;
+
+err:
+       sg_set_page(sg, NULL, 0, 0);
+       sg_mark_end(sg);
+       huge_pages_free_pages(st);
+
+       return -ENOMEM;
+}
+
+static void put_huge_pages(struct drm_i915_gem_object *obj,
+                          struct sg_table *pages)
+{
+       i915_gem_gtt_finish_pages(obj, pages);
+       huge_pages_free_pages(pages);
+
+       obj->mm.dirty = false;
+       obj->mm.madv = I915_MADV_WILLNEED;
+}
+
+static const struct drm_i915_gem_object_ops huge_page_ops = {
+       .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
+                I915_GEM_OBJECT_IS_SHRINKABLE,
+       .get_pages = get_huge_pages,
+       .put_pages = put_huge_pages,
+};
+
+static struct drm_i915_gem_object *
+huge_pages_object(struct drm_i915_private *i915,
+                 u64 size,
+                 unsigned int page_mask)
+{
+       struct drm_i915_gem_object *obj;
+
+       GEM_BUG_ON(!size);
+       GEM_BUG_ON(!IS_ALIGNED(size, BIT(__ffs(page_mask))));
+
+       if (size >> PAGE_SHIFT > INT_MAX)
+               return ERR_PTR(-E2BIG);
+
+       if (overflows_type(size, obj->base.size))
+               return ERR_PTR(-E2BIG);
+
+       obj = i915_gem_object_alloc(i915);
+       if (!obj)
+               return ERR_PTR(-ENOMEM);
+
+       drm_gem_private_object_init(&i915->drm, &obj->base, size);
+       i915_gem_object_init(obj, &huge_page_ops);
+
+       obj->base.write_domain = I915_GEM_DOMAIN_CPU;
+       obj->base.read_domains = I915_GEM_DOMAIN_CPU;
+       obj->cache_level = I915_CACHE_NONE;
+
+       obj->mm.page_mask = page_mask;
+
+       return obj;
+}
+
+static int fake_get_huge_pages(struct drm_i915_gem_object *obj)
+{
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       const u64 max_len = rounddown_pow_of_two(UINT_MAX);
+       struct sg_table *st;
+       struct scatterlist *sg;
+       unsigned int sg_mask;
+       u64 rem;
+
+       st = kmalloc(sizeof(*st), GFP);
+       if (!st)
+               return -ENOMEM;
+
+       if (sg_alloc_table(st, obj->base.size >> PAGE_SHIFT, GFP)) {
+               kfree(st);
+               return -ENOMEM;
+       }
+
+       /* Use optimal page sized chunks to fill in the sg table */
+       rem = obj->base.size;
+       sg = st->sgl;
+       st->nents = 0;
+       sg_mask = 0;
+       do {
+               unsigned int page_size = get_largest_page_size(i915, rem);
+               unsigned int len = min(page_size * div_u64(rem, page_size),
+                                      max_len);
+
+               GEM_BUG_ON(!page_size);
+
+               sg->offset = 0;
+               sg->length = len;
+               sg_dma_len(sg) = len;
+               sg_dma_address(sg) = page_size;
+
+               sg_mask |= len;
+
+               st->nents++;
+
+               rem -= len;
+               if (!rem) {
+                       sg_mark_end(sg);
+                       break;
+               }
+
+               sg = sg_next(sg);
+       } while (1);
+
+       obj->mm.madv = I915_MADV_DONTNEED;
+
+       __i915_gem_object_set_pages(obj, st, sg_mask);
+
+       return 0;
+}
+
+static int fake_get_huge_pages_single(struct drm_i915_gem_object *obj)
+{
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       struct sg_table *st;
+       struct scatterlist *sg;
+       unsigned int page_size;
+
+       st = kmalloc(sizeof(*st), GFP);
+       if (!st)
+               return -ENOMEM;
+
+       if (sg_alloc_table(st, 1, GFP)) {
+               kfree(st);
+               return -ENOMEM;
+       }
+
+       sg = st->sgl;
+       st->nents = 1;
+
+       page_size = get_largest_page_size(i915, obj->base.size);
+       GEM_BUG_ON(!page_size);
+
+       sg->offset = 0;
+       sg->length = obj->base.size;
+       sg_dma_len(sg) = obj->base.size;
+       sg_dma_address(sg) = page_size;
+
+       obj->mm.madv = I915_MADV_DONTNEED;
+
+       __i915_gem_object_set_pages(obj, st, sg->length);
+
+       return 0;
+#undef GFP
+}
+
+static void fake_free_huge_pages(struct drm_i915_gem_object *obj,
+                                struct sg_table *pages)
+{
+       sg_free_table(pages);
+       kfree(pages);
+}
+
+static void fake_put_huge_pages(struct drm_i915_gem_object *obj,
+                               struct sg_table *pages)
+{
+       fake_free_huge_pages(obj, pages);
+       obj->mm.dirty = false;
+       obj->mm.madv = I915_MADV_WILLNEED;
+}
+
+static const struct drm_i915_gem_object_ops fake_ops = {
+       .flags = I915_GEM_OBJECT_IS_SHRINKABLE,
+       .get_pages = fake_get_huge_pages,
+       .put_pages = fake_put_huge_pages,
+};
+
+static const struct drm_i915_gem_object_ops fake_ops_single = {
+       .flags = I915_GEM_OBJECT_IS_SHRINKABLE,
+       .get_pages = fake_get_huge_pages_single,
+       .put_pages = fake_put_huge_pages,
+};
+
+static struct drm_i915_gem_object *
+fake_huge_pages_object(struct drm_i915_private *i915, u64 size, bool single)
+{
+       struct drm_i915_gem_object *obj;
+
+       GEM_BUG_ON(!size);
+       GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
+
+       if (size >> PAGE_SHIFT > UINT_MAX)
+               return ERR_PTR(-E2BIG);
+
+       if (overflows_type(size, obj->base.size))
+               return ERR_PTR(-E2BIG);
+
+       obj = i915_gem_object_alloc(i915);
+       if (!obj)
+               return ERR_PTR(-ENOMEM);
+
+       drm_gem_private_object_init(&i915->drm, &obj->base, size);
+
+       if (single)
+               i915_gem_object_init(obj, &fake_ops_single);
+       else
+               i915_gem_object_init(obj, &fake_ops);
+
+       obj->base.write_domain = I915_GEM_DOMAIN_CPU;
+       obj->base.read_domains = I915_GEM_DOMAIN_CPU;
+       obj->cache_level = I915_CACHE_NONE;
+
+       return obj;
+}
+
+static int igt_check_page_sizes(struct i915_vma *vma)
+{
+       struct drm_i915_private *i915 = to_i915(vma->obj->base.dev);
+       unsigned int supported = INTEL_INFO(i915)->page_sizes;
+       struct drm_i915_gem_object *obj = vma->obj;
+       int err = 0;
+
+       if (!HAS_PAGE_SIZES(i915, vma->page_sizes.sg)) {
+               pr_err("unsupported page_sizes.sg=%u, supported=%u\n",
+                      vma->page_sizes.sg & ~supported, supported);
+               err = -EINVAL;
+       }
+
+       if (!HAS_PAGE_SIZES(i915, vma->page_sizes.gtt)) {
+               pr_err("unsupported page_sizes.gtt=%u, supported=%u\n",
+                      vma->page_sizes.gtt & ~supported, supported);
+               err = -EINVAL;
+       }
+
+       if (vma->page_sizes.phys != obj->mm.page_sizes.phys) {
+               pr_err("vma->page_sizes.phys(%u) != obj->mm.page_sizes.phys(%u)\n",
+                      vma->page_sizes.phys, obj->mm.page_sizes.phys);
+               err = -EINVAL;
+       }
+
+       if (vma->page_sizes.sg != obj->mm.page_sizes.sg) {
+               pr_err("vma->page_sizes.sg(%u) != obj->mm.page_sizes.sg(%u)\n",
+                      vma->page_sizes.sg, obj->mm.page_sizes.sg);
+               err = -EINVAL;
+       }
+
+       if (obj->mm.page_sizes.gtt) {
+               pr_err("obj->page_sizes.gtt(%u) should never be set\n",
+                      obj->mm.page_sizes.gtt);
+               err = -EINVAL;
+       }
+
+       return err;
+}
+
+static int igt_mock_exhaust_device_supported_pages(void *arg)
+{
+       struct i915_hw_ppgtt *ppgtt = arg;
+       struct drm_i915_private *i915 = ppgtt->base.i915;
+       unsigned int saved_mask = INTEL_INFO(i915)->page_sizes;
+       struct drm_i915_gem_object *obj;
+       struct i915_vma *vma;
+       int i, j, single;
+       int err;
+
+       /*
+        * Sanity check creating objects with every valid page support
+        * combination for our mock device.
+        */
+
+       for (i = 1; i < BIT(ARRAY_SIZE(page_sizes)); i++) {
+               unsigned int combination = 0;
+
+               for (j = 0; j < ARRAY_SIZE(page_sizes); j++) {
+                       if (i & BIT(j))
+                               combination |= page_sizes[j];
+               }
+
+               mkwrite_device_info(i915)->page_sizes = combination;
+
+               for (single = 0; single <= 1; ++single) {
+                       obj = fake_huge_pages_object(i915, combination, !!single);
+                       if (IS_ERR(obj)) {
+                               err = PTR_ERR(obj);
+                               goto out_device;
+                       }
+
+                       if (obj->base.size != combination) {
+                               pr_err("obj->base.size=%zu, expected=%u\n",
+                                      obj->base.size, combination);
+                               err = -EINVAL;
+                               goto out_put;
+                       }
+
+                       vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+                       if (IS_ERR(vma)) {
+                               err = PTR_ERR(vma);
+                               goto out_put;
+                       }
+
+                       err = i915_vma_pin(vma, 0, 0, PIN_USER);
+                       if (err)
+                               goto out_close;
+
+                       err = igt_check_page_sizes(vma);
+
+                       if (vma->page_sizes.sg != combination) {
+                               pr_err("page_sizes.sg=%u, expected=%u\n",
+                                      vma->page_sizes.sg, combination);
+                               err = -EINVAL;
+                       }
+
+                       i915_vma_unpin(vma);
+                       i915_vma_close(vma);
+
+                       i915_gem_object_put(obj);
+
+                       if (err)
+                               goto out_device;
+               }
+       }
+
+       goto out_device;
+
+out_close:
+       i915_vma_close(vma);
+out_put:
+       i915_gem_object_put(obj);
+out_device:
+       mkwrite_device_info(i915)->page_sizes = saved_mask;
+
+       return err;
+}
+
+static int igt_mock_ppgtt_misaligned_dma(void *arg)
+{
+       struct i915_hw_ppgtt *ppgtt = arg;
+       struct drm_i915_private *i915 = ppgtt->base.i915;
+       unsigned long supported = INTEL_INFO(i915)->page_sizes;
+       struct drm_i915_gem_object *obj;
+       int bit;
+       int err;
+
+       /*
+        * Sanity check dma misalignment for huge pages -- the dma addresses we
+        * insert into the paging structures need to always respect the page
+        * size alignment.
+        */
+
+       bit = ilog2(I915_GTT_PAGE_SIZE_64K);
+
+       for_each_set_bit_from(bit, &supported,
+                             ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
+               IGT_TIMEOUT(end_time);
+               unsigned int page_size = BIT(bit);
+               unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
+               unsigned int offset;
+               unsigned int size =
+                       round_up(page_size, I915_GTT_PAGE_SIZE_2M) << 1;
+               struct i915_vma *vma;
+
+               obj = fake_huge_pages_object(i915, size, true);
+               if (IS_ERR(obj))
+                       return PTR_ERR(obj);
+
+               if (obj->base.size != size) {
+                       pr_err("obj->base.size=%zu, expected=%u\n",
+                              obj->base.size, size);
+                       err = -EINVAL;
+                       goto out_put;
+               }
+
+               err = i915_gem_object_pin_pages(obj);
+               if (err)
+                       goto out_put;
+
+               /* Force the page size for this object */
+               obj->mm.page_sizes.sg = page_size;
+
+               vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+               if (IS_ERR(vma)) {
+                       err = PTR_ERR(vma);
+                       goto out_unpin;
+               }
+
+               err = i915_vma_pin(vma, 0, 0, flags);
+               if (err) {
+                       i915_vma_close(vma);
+                       goto out_unpin;
+               }
+
+
+               err = igt_check_page_sizes(vma);
+
+               if (vma->page_sizes.gtt != page_size) {
+                       pr_err("page_sizes.gtt=%u, expected %u\n",
+                              vma->page_sizes.gtt, page_size);
+                       err = -EINVAL;
+               }
+
+               i915_vma_unpin(vma);
+
+               if (err) {
+                       i915_vma_close(vma);
+                       goto out_unpin;
+               }
+
+               /*
+                * Try all the other valid offsets until the next
+                * boundary -- should always fall back to using 4K
+                * pages.
+                */
+               for (offset = 4096; offset < page_size; offset += 4096) {
+                       err = i915_vma_unbind(vma);
+                       if (err) {
+                               i915_vma_close(vma);
+                               goto out_unpin;
+                       }
+
+                       err = i915_vma_pin(vma, 0, 0, flags | offset);
+                       if (err) {
+                               i915_vma_close(vma);
+                               goto out_unpin;
+                       }
+
+                       err = igt_check_page_sizes(vma);
+
+                       if (vma->page_sizes.gtt != I915_GTT_PAGE_SIZE_4K) {
+                               pr_err("page_sizes.gtt=%u, expected %lu\n",
+                                      vma->page_sizes.gtt, I915_GTT_PAGE_SIZE_4K);
+                               err = -EINVAL;
+                       }
+
+                       i915_vma_unpin(vma);
+
+                       if (err) {
+                               i915_vma_close(vma);
+                               goto out_unpin;
+                       }
+
+                       if (igt_timeout(end_time,
+                                       "%s timed out at offset %x with page-size %x\n",
+                                       __func__, offset, page_size))
+                               break;
+               }
+
+               i915_vma_close(vma);
+
+               i915_gem_object_unpin_pages(obj);
+               i915_gem_object_put(obj);
+       }
+
+       return 0;
+
+out_unpin:
+       i915_gem_object_unpin_pages(obj);
+out_put:
+       i915_gem_object_put(obj);
+
+       return err;
+}
+
+static void close_object_list(struct list_head *objects,
+                             struct i915_hw_ppgtt *ppgtt)
+{
+       struct drm_i915_gem_object *obj, *on;
+
+       list_for_each_entry_safe(obj, on, objects, st_link) {
+               struct i915_vma *vma;
+
+               vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+               if (!IS_ERR(vma))
+                       i915_vma_close(vma);
+
+               list_del(&obj->st_link);
+               i915_gem_object_unpin_pages(obj);
+               i915_gem_object_put(obj);
+       }
+}
+
+static int igt_mock_ppgtt_huge_fill(void *arg)
+{
+       struct i915_hw_ppgtt *ppgtt = arg;
+       struct drm_i915_private *i915 = ppgtt->base.i915;
+       unsigned long max_pages = ppgtt->base.total >> PAGE_SHIFT;
+       unsigned long page_num;
+       bool single = false;
+       LIST_HEAD(objects);
+       IGT_TIMEOUT(end_time);
+       int err;
+
+       for_each_prime_number_from(page_num, 1, max_pages) {
+               struct drm_i915_gem_object *obj;
+               u64 size = page_num << PAGE_SHIFT;
+               struct i915_vma *vma;
+               unsigned int expected_gtt = 0;
+               int i;
+
+               obj = fake_huge_pages_object(i915, size, single);
+               if (IS_ERR(obj)) {
+                       err = PTR_ERR(obj);
+                       break;
+               }
+
+               if (obj->base.size != size) {
+                       pr_err("obj->base.size=%zd, expected=%llu\n",
+                              obj->base.size, size);
+                       i915_gem_object_put(obj);
+                       err = -EINVAL;
+                       break;
+               }
+
+               err = i915_gem_object_pin_pages(obj);
+               if (err) {
+                       i915_gem_object_put(obj);
+                       break;
+               }
+
+               list_add(&obj->st_link, &objects);
+
+               vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+               if (IS_ERR(vma)) {
+                       err = PTR_ERR(vma);
+                       break;
+               }
+
+               err = i915_vma_pin(vma, 0, 0, PIN_USER);
+               if (err)
+                       break;
+
+               err = igt_check_page_sizes(vma);
+               if (err) {
+                       i915_vma_unpin(vma);
+                       break;
+               }
+
+               /*
+                * Figure out the expected gtt page size knowing that we go from
+                * largest to smallest page size sg chunks, and that we align to
+                * the largest page size.
+                */
+               for (i = 0; i < ARRAY_SIZE(page_sizes); ++i) {
+                       unsigned int page_size = page_sizes[i];
+
+                       if (HAS_PAGE_SIZES(i915, page_size) &&
+                           size >= page_size) {
+                               expected_gtt |= page_size;
+                               size &= page_size-1;
+                       }
+               }
+
+               GEM_BUG_ON(!expected_gtt);
+               GEM_BUG_ON(size);
+
+               if (expected_gtt & I915_GTT_PAGE_SIZE_4K)
+                       expected_gtt &= ~I915_GTT_PAGE_SIZE_64K;
+
+               i915_vma_unpin(vma);
+
+               if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
+                       if (!IS_ALIGNED(vma->node.start,
+                                       I915_GTT_PAGE_SIZE_2M)) {
+                               pr_err("node.start(%llx) not aligned to 2M\n",
+                                      vma->node.start);
+                               err = -EINVAL;
+                               break;
+                       }
+
+                       if (!IS_ALIGNED(vma->node.size,
+                                       I915_GTT_PAGE_SIZE_2M)) {
+                               pr_err("node.size(%llx) not aligned to 2M\n",
+                                      vma->node.size);
+                               err = -EINVAL;
+                               break;
+                       }
+               }
+
+               if (vma->page_sizes.gtt != expected_gtt) {
+                       pr_err("gtt=%u, expected=%u, size=%zd, single=%s\n",
+                              vma->page_sizes.gtt, expected_gtt,
+                              obj->base.size, yesno(!!single));
+                       err = -EINVAL;
+                       break;
+               }
+
+               if (igt_timeout(end_time,
+                               "%s timed out at size %zd\n",
+                               __func__, obj->base.size))
+                       break;
+
+               single = !single;
+       }
+
+       close_object_list(&objects, ppgtt);
+
+       if (err == -ENOMEM || err == -ENOSPC)
+               err = 0;
+
+       return err;
+}
+
+static int igt_mock_ppgtt_64K(void *arg)
+{
+       struct i915_hw_ppgtt *ppgtt = arg;
+       struct drm_i915_private *i915 = ppgtt->base.i915;
+       struct drm_i915_gem_object *obj;
+       const struct object_info {
+               unsigned int size;
+               unsigned int gtt;
+               unsigned int offset;
+       } objects[] = {
+               /* Cases with forced padding/alignment */
+               {
+                       .size = SZ_64K,
+                       .gtt = I915_GTT_PAGE_SIZE_64K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_64K + SZ_4K,
+                       .gtt = I915_GTT_PAGE_SIZE_4K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_64K - SZ_4K,
+                       .gtt = I915_GTT_PAGE_SIZE_4K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_2M,
+                       .gtt = I915_GTT_PAGE_SIZE_64K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_2M - SZ_4K,
+                       .gtt = I915_GTT_PAGE_SIZE_4K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_2M + SZ_4K,
+                       .gtt = I915_GTT_PAGE_SIZE_64K | I915_GTT_PAGE_SIZE_4K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_2M + SZ_64K,
+                       .gtt = I915_GTT_PAGE_SIZE_64K,
+                       .offset = 0,
+               },
+               {
+                       .size = SZ_2M - SZ_64K,
+                       .gtt = I915_GTT_PAGE_SIZE_64K,
+                       .offset = 0,
+               },
+               /* Try without any forced padding/alignment */
+               {
+                       .size = SZ_64K,
+                       .offset = SZ_2M,
+                       .gtt = I915_GTT_PAGE_SIZE_4K,
+               },
+               {
+                       .size = SZ_128K,
+                       .offset = SZ_2M - SZ_64K,
+                       .gtt = I915_GTT_PAGE_SIZE_4K,
+               },
+       };
+       struct i915_vma *vma;
+       int i, single;
+       int err;
+
+       /*
+        * Sanity check some of the trickiness with 64K pages -- either we can
+        * safely mark the whole page-table(2M block) as 64K, or we have to
+        * always fallback to 4K.
+        */
+
+       if (!HAS_PAGE_SIZES(i915, I915_GTT_PAGE_SIZE_64K))
+               return 0;
+
+       for (i = 0; i < ARRAY_SIZE(objects); ++i) {
+               unsigned int size = objects[i].size;
+               unsigned int expected_gtt = objects[i].gtt;
+               unsigned int offset = objects[i].offset;
+               unsigned int flags = PIN_USER;
+
+               for (single = 0; single <= 1; single++) {
+                       obj = fake_huge_pages_object(i915, size, !!single);
+                       if (IS_ERR(obj))
+                               return PTR_ERR(obj);
+
+                       err = i915_gem_object_pin_pages(obj);
+                       if (err)
+                               goto out_object_put;
+
+                       /*
+                        * Disable 2M pages -- We only want to use 64K/4K pages
+                        * for this test.
+                        */
+                       obj->mm.page_sizes.sg &= ~I915_GTT_PAGE_SIZE_2M;
+
+                       vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+                       if (IS_ERR(vma)) {
+                               err = PTR_ERR(vma);
+                               goto out_object_unpin;
+                       }
+
+                       if (offset)
+                               flags |= PIN_OFFSET_FIXED | offset;
+
+                       err = i915_vma_pin(vma, 0, 0, flags);
+                       if (err)
+                               goto out_vma_close;
+
+                       err = igt_check_page_sizes(vma);
+                       if (err)
+                               goto out_vma_unpin;
+
+                       if (!offset && vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
+                               if (!IS_ALIGNED(vma->node.start,
+                                               I915_GTT_PAGE_SIZE_2M)) {
+                                       pr_err("node.start(%llx) not aligned to 2M\n",
+                                              vma->node.start);
+                                       err = -EINVAL;
+                                       goto out_vma_unpin;
+                               }
+
+                               if (!IS_ALIGNED(vma->node.size,
+                                               I915_GTT_PAGE_SIZE_2M)) {
+                                       pr_err("node.size(%llx) not aligned to 2M\n",
+                                              vma->node.size);
+                                       err = -EINVAL;
+                                       goto out_vma_unpin;
+                               }
+                       }
+
+                       if (vma->page_sizes.gtt != expected_gtt) {
+                               pr_err("gtt=%u, expected=%u, i=%d, single=%s\n",
+                                      vma->page_sizes.gtt, expected_gtt, i,
+                                      yesno(!!single));
+                               err = -EINVAL;
+                               goto out_vma_unpin;
+                       }
+
+                       i915_vma_unpin(vma);
+                       i915_vma_close(vma);
+
+                       i915_gem_object_unpin_pages(obj);
+                       i915_gem_object_put(obj);
+               }
+       }
+
+       return 0;
+
+out_vma_unpin:
+       i915_vma_unpin(vma);
+out_vma_close:
+       i915_vma_close(vma);
+out_object_unpin:
+       i915_gem_object_unpin_pages(obj);
+out_object_put:
+       i915_gem_object_put(obj);
+
+       return err;
+}
+
+static struct i915_vma *
+gpu_write_dw(struct i915_vma *vma, u64 offset, u32 val)
+{
+       struct drm_i915_private *i915 = to_i915(vma->obj->base.dev);
+       const int gen = INTEL_GEN(vma->vm->i915);
+       unsigned int count = vma->size >> PAGE_SHIFT;
+       struct drm_i915_gem_object *obj;
+       struct i915_vma *batch;
+       unsigned int size;
+       u32 *cmd;
+       int n;
+       int err;
+
+       size = (1 + 4 * count) * sizeof(u32);
+       size = round_up(size, PAGE_SIZE);
+       obj = i915_gem_object_create_internal(i915, size);
+       if (IS_ERR(obj))
+               return ERR_CAST(obj);
+
+       cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
+       if (IS_ERR(cmd)) {
+               err = PTR_ERR(cmd);
+               goto err;
+       }
+
+       offset += vma->node.start;
+
+       for (n = 0; n < count; n++) {
+               if (gen >= 8) {
+                       *cmd++ = MI_STORE_DWORD_IMM_GEN4;
+                       *cmd++ = lower_32_bits(offset);
+                       *cmd++ = upper_32_bits(offset);
+                       *cmd++ = val;
+               } else if (gen >= 4) {
+                       *cmd++ = MI_STORE_DWORD_IMM_GEN4 |
+                               (gen < 6 ? 1 << 22 : 0);
+                       *cmd++ = 0;
+                       *cmd++ = offset;
+                       *cmd++ = val;
+               } else {
+                       *cmd++ = MI_STORE_DWORD_IMM | 1 << 22;
+                       *cmd++ = offset;
+                       *cmd++ = val;
+               }
+
+               offset += PAGE_SIZE;
+       }
+
+       *cmd = MI_BATCH_BUFFER_END;
+
+       i915_gem_object_unpin_map(obj);
+
+       err = i915_gem_object_set_to_gtt_domain(obj, false);
+       if (err)
+               goto err;
+
+       batch = i915_vma_instance(obj, vma->vm, NULL);
+       if (IS_ERR(batch)) {
+               err = PTR_ERR(batch);
+               goto err;
+       }
+
+       err = i915_vma_pin(batch, 0, 0, PIN_USER);
+       if (err)
+               goto err;
+
+       return batch;
+
+err:
+       i915_gem_object_put(obj);
+
+       return ERR_PTR(err);
+}
+
+static int gpu_write(struct i915_vma *vma,
+                    struct i915_gem_context *ctx,
+                    struct intel_engine_cs *engine,
+                    u32 dword,
+                    u32 value)
+{
+       struct drm_i915_gem_request *rq;
+       struct i915_vma *batch;
+       int flags = 0;
+       int err;
+
+       GEM_BUG_ON(!intel_engine_can_store_dword(engine));
+
+       err = i915_gem_object_set_to_gtt_domain(vma->obj, true);
+       if (err)
+               return err;
+
+       rq = i915_gem_request_alloc(engine, ctx);
+       if (IS_ERR(rq))
+               return PTR_ERR(rq);
+
+       batch = gpu_write_dw(vma, dword * sizeof(u32), value);
+       if (IS_ERR(batch)) {
+               err = PTR_ERR(batch);
+               goto err_request;
+       }
+
+       i915_vma_move_to_active(batch, rq, 0);
+       i915_gem_object_set_active_reference(batch->obj);
+       i915_vma_unpin(batch);
+       i915_vma_close(batch);
+
+       err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
+       if (err)
+               goto err_request;
+
+       err = i915_switch_context(rq);
+       if (err)
+               goto err_request;
+
+       err = rq->engine->emit_bb_start(rq,
+                                       batch->node.start, batch->node.size,
+                                       flags);
+       if (err)
+               goto err_request;
+
+       i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
+
+       reservation_object_lock(vma->resv, NULL);
+       reservation_object_add_excl_fence(vma->resv, &rq->fence);
+       reservation_object_unlock(vma->resv);
+
+err_request:
+       __i915_add_request(rq, err == 0);
+
+       return err;
+}
+
+static int cpu_check(struct drm_i915_gem_object *obj, u32 dword, u32 val)
+{
+       unsigned int needs_flush;
+       unsigned long n;
+       int err;
+
+       err = i915_gem_obj_prepare_shmem_read(obj, &needs_flush);
+       if (err)
+               return err;
+
+       for (n = 0; n < obj->base.size >> PAGE_SHIFT; ++n) {
+               u32 *ptr = kmap_atomic(i915_gem_object_get_page(obj, n));
+
+               if (needs_flush & CLFLUSH_BEFORE)
+                       drm_clflush_virt_range(ptr, PAGE_SIZE);
+
+               if (ptr[dword] != val) {
+                       pr_err("n=%lu ptr[%u]=%u, val=%u\n",
+                              n, dword, ptr[dword], val);
+                       kunmap_atomic(ptr);
+                       err = -EINVAL;
+                       break;
+               }
+
+               kunmap_atomic(ptr);
+       }
+
+       i915_gem_obj_finish_shmem_access(obj);
+
+       return err;
+}
+
+static int igt_write_huge(struct drm_i915_gem_object *obj)
+{
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       struct i915_gem_context *ctx = i915->kernel_context;
+       struct i915_address_space *vm = ctx->ppgtt ? &ctx->ppgtt->base : &i915->ggtt.base;
+       struct intel_engine_cs *engine;
+       struct i915_vma *vma;
+       unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
+       unsigned int max_page_size;
+       unsigned int id;
+       u64 max;
+       u64 num;
+       u64 size;
+       int err = 0;
+
+       GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
+
+       size = obj->base.size;
+       if (obj->mm.page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
+               size = round_up(size, I915_GTT_PAGE_SIZE_2M);
+
+       max_page_size = rounddown_pow_of_two(obj->mm.page_sizes.sg);
+       max = div_u64((vm->total - size), max_page_size);
+
+       vma = i915_vma_instance(obj, vm, NULL);
+       if (IS_ERR(vma))
+               return PTR_ERR(vma);
+
+       for_each_engine(engine, i915, id) {
+               IGT_TIMEOUT(end_time);
+
+               if (!intel_engine_can_store_dword(engine)) {
+                       pr_info("store-dword-imm not supported on engine=%u\n",
+                               id);
+                       continue;
+               }
+
+               /*
+                * Try various offsets until we timeout -- we want to avoid
+                * issues hidden by effectively always using offset = 0.
+                */
+               for_each_prime_number_from(num, 0, max) {
+                       u64 offset = num * max_page_size;
+                       u32 dword;
+
+                       err = i915_vma_unbind(vma);
+                       if (err)
+                               goto out_vma_close;
+
+                       err = i915_vma_pin(vma, size, max_page_size, flags | offset);
+                       if (err) {
+                               /*
+                                * The ggtt may have some pages reserved so
+                                * refrain from erroring out.
+                                */
+                               if (err == -ENOSPC && i915_is_ggtt(vm)) {
+                                       err = 0;
+                                       continue;
+                               }
+
+                               goto out_vma_close;
+                       }
+
+                       err = igt_check_page_sizes(vma);
+                       if (err)
+                               goto out_vma_unpin;
+
+                       dword = offset_in_page(num) / 4;
+
+                       err = gpu_write(vma, ctx, engine, dword, num + 1);
+                       if (err) {
+                               pr_err("gpu-write failed at offset=%llx", offset);
+                               goto out_vma_unpin;
+                       }
+
+                       err = cpu_check(obj, dword, num + 1);
+                       if (err) {
+                               pr_err("cpu-check failed at offset=%llx", offset);
+                               goto out_vma_unpin;
+                       }
+
+                       i915_vma_unpin(vma);
+
+                       if (num > 0 &&
+                           igt_timeout(end_time,
+                                       "%s timed out on engine=%u at offset=%llx, max_page_size=%x\n",
+                                       __func__, id, offset, max_page_size))
+                               break;
+               }
+       }
+
+out_vma_unpin:
+       if (i915_vma_is_pinned(vma))
+               i915_vma_unpin(vma);
+out_vma_close:
+       i915_vma_close(vma);
+
+       return err;
+}
+
+static int igt_ppgtt_exhaust_huge(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       unsigned long supported = INTEL_INFO(i915)->page_sizes;
+       static unsigned int pages[ARRAY_SIZE(page_sizes)];
+       struct drm_i915_gem_object *obj;
+       unsigned int size_mask;
+       unsigned int page_mask;
+       int n, i;
+       int err;
+
+       /*
+        * Sanity check creating objects with a varying mix of page sizes --
+        * ensuring that our writes lands in the right place.
+        */
+
+       n = 0;
+       for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1)
+               pages[n++] = BIT(i);
+
+       for (size_mask = 2; size_mask < BIT(n); size_mask++) {
+               unsigned int size = 0;
+
+               for (i = 0; i < n; i++) {
+                       if (size_mask & BIT(i))
+                               size |= pages[i];
+               }
+
+               /*
+                * For our page mask we want to enumerate all the page-size
+                * combinations which will fit into our chosen object size.
+                */
+               for (page_mask = 2; page_mask <= size_mask; page_mask++) {
+                       unsigned int page_sizes = 0;
+
+                       for (i = 0; i < n; i++) {
+                               if (page_mask & BIT(i))
+                                       page_sizes |= pages[i];
+                       }
+
+                       /*
+                        * Ensure that we can actually fill the given object
+                        * with our chosen page mask.
+                        */
+                       if (!IS_ALIGNED(size, BIT(__ffs(page_sizes))))
+                               continue;
+
+                       obj = huge_pages_object(i915, size, page_sizes);
+                       if (IS_ERR(obj)) {
+                               err = PTR_ERR(obj);
+                               goto out_device;
+                       }
+
+                       err = i915_gem_object_pin_pages(obj);
+                       if (err) {
+                               i915_gem_object_put(obj);
+
+                               if (err == -ENOMEM) {
+                                       pr_info("unable to get pages, size=%u, pages=%u\n",
+                                               size, page_sizes);
+                                       err = 0;
+                                       break;
+                               }
+
+                               pr_err("pin_pages failed, size=%u, pages=%u\n",
+                                      size_mask, page_mask);
+
+                               goto out_device;
+                       }
+
+                       /* Force the page-size for the gtt insertion */
+                       obj->mm.page_sizes.sg = page_sizes;
+
+                       err = igt_write_huge(obj);
+                       if (err) {
+                               pr_err("exhaust write-huge failed with size=%u\n",
+                                      size);
+                               goto out_unpin;
+                       }
+
+                       i915_gem_object_unpin_pages(obj);
+                       i915_gem_object_put(obj);
+               }
+       }
+
+       goto out_device;
+
+out_unpin:
+       i915_gem_object_unpin_pages(obj);
+       i915_gem_object_put(obj);
+out_device:
+       mkwrite_device_info(i915)->page_sizes = supported;
+
+       return err;
+}
+
+static int igt_ppgtt_internal_huge(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       struct drm_i915_gem_object *obj;
+       static const unsigned int sizes[] = {
+               SZ_64K,
+               SZ_128K,
+               SZ_256K,
+               SZ_512K,
+               SZ_1M,
+               SZ_2M,
+       };
+       int i;
+       int err;
+
+       /*
+        * Sanity check that the HW uses huge pages correctly through internal
+        * -- ensure that our writes land in the right place.
+        */
+
+       for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
+               unsigned int size = sizes[i];
+
+               obj = i915_gem_object_create_internal(i915, size);
+               if (IS_ERR(obj))
+                       return PTR_ERR(obj);
+
+               err = i915_gem_object_pin_pages(obj);
+               if (err)
+                       goto out_put;
+
+               if (obj->mm.page_sizes.phys < I915_GTT_PAGE_SIZE_64K) {
+                       pr_info("internal unable to allocate huge-page(s) with size=%u\n",
+                               size);
+                       goto out_unpin;
+               }
+
+               err = igt_write_huge(obj);
+               if (err) {
+                       pr_err("internal write-huge failed with size=%u\n",
+                              size);
+                       goto out_unpin;
+               }
+
+               i915_gem_object_unpin_pages(obj);
+               i915_gem_object_put(obj);
+       }
+
+       return 0;
+
+out_unpin:
+       i915_gem_object_unpin_pages(obj);
+out_put:
+       i915_gem_object_put(obj);
+
+       return err;
+}
+
+static inline bool igt_can_allocate_thp(struct drm_i915_private *i915)
+{
+       return i915->mm.gemfs && has_transparent_hugepage();
+}
+
+static int igt_ppgtt_gemfs_huge(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       struct drm_i915_gem_object *obj;
+       static const unsigned int sizes[] = {
+               SZ_2M,
+               SZ_4M,
+               SZ_8M,
+               SZ_16M,
+               SZ_32M,
+       };
+       int i;
+       int err;
+
+       /*
+        * Sanity check that the HW uses huge pages correctly through gemfs --
+        * ensure that our writes land in the right place.
+        */
+
+       if (!igt_can_allocate_thp(i915)) {
+               pr_info("missing THP support, skipping\n");
+               return 0;
+       }
+
+       for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
+               unsigned int size = sizes[i];
+
+               obj = i915_gem_object_create(i915, size);
+               if (IS_ERR(obj))
+                       return PTR_ERR(obj);
+
+               err = i915_gem_object_pin_pages(obj);
+               if (err)
+                       goto out_put;
+
+               if (obj->mm.page_sizes.phys < I915_GTT_PAGE_SIZE_2M) {
+                       pr_info("finishing test early, gemfs unable to allocate huge-page(s) with size=%u\n",
+                               size);
+                       goto out_unpin;
+               }
+
+               err = igt_write_huge(obj);
+               if (err) {
+                       pr_err("gemfs write-huge failed with size=%u\n",
+                              size);
+                       goto out_unpin;
+               }
+
+               i915_gem_object_unpin_pages(obj);
+               i915_gem_object_put(obj);
+       }
+
+       return 0;
+
+out_unpin:
+       i915_gem_object_unpin_pages(obj);
+out_put:
+       i915_gem_object_put(obj);
+
+       return err;
+}
+
+static int igt_ppgtt_pin_update(void *arg)
+{
+       struct drm_i915_private *dev_priv = arg;
+       unsigned long supported = INTEL_INFO(dev_priv)->page_sizes;
+       struct i915_hw_ppgtt *ppgtt = dev_priv->kernel_context->ppgtt;
+       struct drm_i915_gem_object *obj;
+       struct i915_vma *vma;
+       unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
+       int first, last;
+       int err;
+
+       /*
+        * Make sure there's no funny business when doing a PIN_UPDATE -- in the
+        * past we had a subtle issue with being able to incorrectly do multiple
+        * alloc va ranges on the same object when doing a PIN_UPDATE, which
+        * resulted in some pretty nasty bugs, though only when using
+        * huge-gtt-pages.
+        */
+
+       if (!USES_FULL_48BIT_PPGTT(dev_priv)) {
+               pr_info("48b PPGTT not supported, skipping\n");
+               return 0;
+       }
+
+       first = ilog2(I915_GTT_PAGE_SIZE_64K);
+       last = ilog2(I915_GTT_PAGE_SIZE_2M);
+
+       for_each_set_bit_from(first, &supported, last + 1) {
+               unsigned int page_size = BIT(first);
+
+               obj = i915_gem_object_create_internal(dev_priv, page_size);
+               if (IS_ERR(obj))
+                       return PTR_ERR(obj);
+
+               vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+               if (IS_ERR(vma)) {
+                       err = PTR_ERR(vma);
+                       goto out_put;
+               }
+
+               err = i915_vma_pin(vma, SZ_2M, 0, flags);
+               if (err)
+                       goto out_close;
+
+               if (vma->page_sizes.sg < page_size) {
+                       pr_info("Unable to allocate page-size %x, finishing test early\n",
+                               page_size);
+                       goto out_unpin;
+               }
+
+               err = igt_check_page_sizes(vma);
+               if (err)
+                       goto out_unpin;
+
+               if (vma->page_sizes.gtt != page_size) {
+                       dma_addr_t addr = i915_gem_object_get_dma_address(obj, 0);
+
+                       /*
+                        * The only valid reason for this to ever fail would be
+                        * if the dma-mapper screwed us over when we did the
+                        * dma_map_sg(), since it has the final say over the dma
+                        * address.
+                        */
+                       if (IS_ALIGNED(addr, page_size)) {
+                               pr_err("page_sizes.gtt=%u, expected=%u\n",
+                                      vma->page_sizes.gtt, page_size);
+                               err = -EINVAL;
+                       } else {
+                               pr_info("dma address misaligned, finishing test early\n");
+                       }
+
+                       goto out_unpin;
+               }
+
+               err = i915_vma_bind(vma, I915_CACHE_NONE, PIN_UPDATE);
+               if (err)
+                       goto out_unpin;
+
+               i915_vma_unpin(vma);
+               i915_vma_close(vma);
+
+               i915_gem_object_put(obj);
+       }
+
+       obj = i915_gem_object_create_internal(dev_priv, PAGE_SIZE);
+       if (IS_ERR(obj))
+               return PTR_ERR(obj);
+
+       vma = i915_vma_instance(obj, &ppgtt->base, NULL);
+       if (IS_ERR(vma)) {
+               err = PTR_ERR(vma);
+               goto out_put;
+       }
+
+       err = i915_vma_pin(vma, 0, 0, flags);
+       if (err)
+               goto out_close;
+
+       /*
+        * Make sure we don't end up with something like where the pde is still
+        * pointing to the 2M page, and the pt we just filled-in is dangling --
+        * we can check this by writing to the first page where it would then
+        * land in the now stale 2M page.
+        */
+
+       err = gpu_write(vma, dev_priv->kernel_context, dev_priv->engine[RCS],
+                       0, 0xdeadbeaf);
+       if (err)
+               goto out_unpin;
+
+       err = cpu_check(obj, 0, 0xdeadbeaf);
+
+out_unpin:
+       i915_vma_unpin(vma);
+out_close:
+       i915_vma_close(vma);
+out_put:
+       i915_gem_object_put(obj);
+
+       return err;
+}
+
+static int igt_tmpfs_fallback(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       struct vfsmount *gemfs = i915->mm.gemfs;
+       struct i915_gem_context *ctx = i915->kernel_context;
+       struct i915_address_space *vm = ctx->ppgtt ? &ctx->ppgtt->base : &i915->ggtt.base;
+       struct drm_i915_gem_object *obj;
+       struct i915_vma *vma;
+       u32 *vaddr;
+       int err = 0;
+
+       /*
+        * Make sure that we don't burst into a ball of flames upon falling back
+        * to tmpfs, which we rely on if on the off-chance we encouter a failure
+        * when setting up gemfs.
+        */
+
+       i915->mm.gemfs = NULL;
+
+       obj = i915_gem_object_create(i915, PAGE_SIZE);
+       if (IS_ERR(obj)) {
+               err = PTR_ERR(obj);
+               goto out_restore;
+       }
+
+       vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
+       if (IS_ERR(vaddr)) {
+               err = PTR_ERR(vaddr);
+               goto out_put;
+       }
+       *vaddr = 0xdeadbeaf;
+
+       i915_gem_object_unpin_map(obj);
+
+       vma = i915_vma_instance(obj, vm, NULL);
+       if (IS_ERR(vma)) {
+               err = PTR_ERR(vma);
+               goto out_put;
+       }
+
+       err = i915_vma_pin(vma, 0, 0, PIN_USER);
+       if (err)
+               goto out_close;
+
+       err = igt_check_page_sizes(vma);
+
+       i915_vma_unpin(vma);
+out_close:
+       i915_vma_close(vma);
+out_put:
+       i915_gem_object_put(obj);
+out_restore:
+       i915->mm.gemfs = gemfs;
+
+       return err;
+}
+
+static int igt_shrink_thp(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       struct i915_gem_context *ctx = i915->kernel_context;
+       struct i915_address_space *vm = ctx->ppgtt ? &ctx->ppgtt->base : &i915->ggtt.base;
+       struct drm_i915_gem_object *obj;
+       struct i915_vma *vma;
+       unsigned int flags = PIN_USER;
+       int err;
+
+       /*
+        * Sanity check shrinking huge-paged object -- make sure nothing blows
+        * up.
+        */
+
+       if (!igt_can_allocate_thp(i915)) {
+               pr_info("missing THP support, skipping\n");
+               return 0;
+       }
+
+       obj = i915_gem_object_create(i915, SZ_2M);
+       if (IS_ERR(obj))
+               return PTR_ERR(obj);
+
+       vma = i915_vma_instance(obj, vm, NULL);
+       if (IS_ERR(vma)) {
+               err = PTR_ERR(vma);
+               goto out_put;
+       }
+
+       err = i915_vma_pin(vma, 0, 0, flags);
+       if (err)
+               goto out_close;
+
+       if (obj->mm.page_sizes.phys < I915_GTT_PAGE_SIZE_2M) {
+               pr_info("failed to allocate THP, finishing test early\n");
+               goto out_unpin;
+       }
+
+       err = igt_check_page_sizes(vma);
+       if (err)
+               goto out_unpin;
+
+       err = gpu_write(vma, i915->kernel_context, i915->engine[RCS], 0,
+                       0xdeadbeaf);
+       if (err)
+               goto out_unpin;
+
+       i915_vma_unpin(vma);
+
+       /*
+        * Now that the pages are *unpinned* shrink-all should invoke
+        * shmem to truncate our pages.
+        */
+       i915_gem_shrink_all(i915);
+       if (!IS_ERR_OR_NULL(obj->mm.pages)) {
+               pr_err("shrink-all didn't truncate the pages\n");
+               err = -EINVAL;
+               goto out_close;
+       }
+
+       if (obj->mm.page_sizes.sg || obj->mm.page_sizes.phys) {
+               pr_err("residual page-size bits left\n");
+               err = -EINVAL;
+               goto out_close;
+       }
+
+       err = i915_vma_pin(vma, 0, 0, flags);
+       if (err)
+               goto out_close;
+
+       err = cpu_check(obj, 0, 0xdeadbeaf);
+
+out_unpin:
+       i915_vma_unpin(vma);
+out_close:
+       i915_vma_close(vma);
+out_put:
+       i915_gem_object_put(obj);
+
+       return err;
+}
+
+int i915_gem_huge_page_mock_selftests(void)
+{
+       static const struct i915_subtest tests[] = {
+               SUBTEST(igt_mock_exhaust_device_supported_pages),
+               SUBTEST(igt_mock_ppgtt_misaligned_dma),
+               SUBTEST(igt_mock_ppgtt_huge_fill),
+               SUBTEST(igt_mock_ppgtt_64K),
+       };
+       int saved_ppgtt = i915_modparams.enable_ppgtt;
+       struct drm_i915_private *dev_priv;
+       struct pci_dev *pdev;
+       struct i915_hw_ppgtt *ppgtt;
+       int err;
+
+       dev_priv = mock_gem_device();
+       if (!dev_priv)
+               return -ENOMEM;
+
+       /* Pretend to be a device which supports the 48b PPGTT */
+       i915_modparams.enable_ppgtt = 3;
+
+       pdev = dev_priv->drm.pdev;
+       dma_coerce_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(39));
+
+       mutex_lock(&dev_priv->drm.struct_mutex);
+       ppgtt = i915_ppgtt_create(dev_priv, ERR_PTR(-ENODEV), "mock");
+       if (IS_ERR(ppgtt)) {
+               err = PTR_ERR(ppgtt);
+               goto out_unlock;
+       }
+
+       if (!i915_vm_is_48bit(&ppgtt->base)) {
+               pr_err("failed to create 48b PPGTT\n");
+               err = -EINVAL;
+               goto out_close;
+       }
+
+       /* If we were ever hit this then it's time to mock the 64K scratch */
+       if (!i915_vm_has_scratch_64K(&ppgtt->base)) {
+               pr_err("PPGTT missing 64K scratch page\n");
+               err = -EINVAL;
+               goto out_close;
+       }
+
+       err = i915_subtests(tests, ppgtt);
+
+out_close:
+       i915_ppgtt_close(&ppgtt->base);
+       i915_ppgtt_put(ppgtt);
+
+out_unlock:
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+
+       i915_modparams.enable_ppgtt = saved_ppgtt;
+
+       drm_dev_unref(&dev_priv->drm);
+
+       return err;
+}
+
+int i915_gem_huge_page_live_selftests(struct drm_i915_private *dev_priv)
+{
+       static const struct i915_subtest tests[] = {
+               SUBTEST(igt_shrink_thp),
+               SUBTEST(igt_ppgtt_pin_update),
+               SUBTEST(igt_tmpfs_fallback),
+               SUBTEST(igt_ppgtt_exhaust_huge),
+               SUBTEST(igt_ppgtt_gemfs_huge),
+               SUBTEST(igt_ppgtt_internal_huge),
+       };
+       int err;
+
+       if (!USES_PPGTT(dev_priv)) {
+               pr_info("PPGTT not supported, skipping live-selftests\n");
+               return 0;
+       }
+
+       mutex_lock(&dev_priv->drm.struct_mutex);
+       err = i915_subtests(tests, dev_priv);
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+
+       return err;
+}
index 18b174d855ca7079956951368a8284804275d44b..64acd7eccc5c67213f436a5d0f2c53caa9a64002 100644 (file)
@@ -15,5 +15,6 @@ selftest(objects, i915_gem_object_live_selftests)
 selftest(dmabuf, i915_gem_dmabuf_live_selftests)
 selftest(coherency, i915_gem_coherency_live_selftests)
 selftest(gtt, i915_gem_gtt_live_selftests)
+selftest(hugepages, i915_gem_huge_page_live_selftests)
 selftest(contexts, i915_gem_context_live_selftests)
 selftest(hangcheck, intel_hangcheck_live_selftests)
index fc74687501ba923c97f5afe14a13deba7ee6c893..9961b44f76ed4e3166761137bda500fa076f79cf 100644 (file)
@@ -21,3 +21,4 @@ selftest(dmabuf, i915_gem_dmabuf_mock_selftests)
 selftest(vma, i915_vma_mock_selftests)
 selftest(evict, i915_gem_evict_mock_selftests)
 selftest(gtt, i915_gem_gtt_mock_selftests)
+selftest(hugepages, i915_gem_huge_page_mock_selftests)