[PATCH 15/19] drm/i915/selftests: basic huge page tests

[Matthew Auld <matthew.auld@intel.com>]

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>
---
 drivers/gpu/drm/i915/i915_gem.c                    |   1 +
 drivers/gpu/drm/i915/selftests/huge_pages.c        | 604 +++++++++++++++++++++
 .../gpu/drm/i915/selftests/i915_live_selftests.h   |   1 +
 3 files changed, 606 insertions(+)
 create mode 100644 drivers/gpu/drm/i915/selftests/huge_pages.c

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index c4a27b4f419c..6cced283e9de 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -5413,6 +5413,7 @@ i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
 #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
diff --git a/drivers/gpu/drm/i915/selftests/huge_pages.c b/drivers/gpu/drm/i915/selftests/huge_pages.c
new file mode 100644
index 000000000000..5e0437045804
--- /dev/null
+++ b/drivers/gpu/drm/i915/selftests/huge_pages.c
@@ -0,0 +1,604 @@
+/*
+ * 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 unsigned int page_sizes[] = {
+	I915_GTT_PAGE_SIZE_1G,
+	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,
+					  unsigned int rem)
+{
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(page_sizes); ++i) {
+		unsigned int page_size = page_sizes[i];
+
+		if (HAS_PAGE_SIZE(i915, page_size) && rem >= page_size)
+			return page_size;
+	}
+
+	GEM_BUG_ON(1);
+}
+
+static struct sg_table *
+fake_get_huge_pages(struct drm_i915_gem_object *obj,
+		    unsigned int *sg_mask)
+{
+#define GFP (GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY)
+	struct drm_i915_private *i915 = to_i915(obj->base.dev);
+	const unsigned long max = obj->base.size >> PAGE_SHIFT;
+	struct sg_table *st;
+	struct scatterlist *sg;
+	typeof(obj->base.size) rem;
+
+	st = kmalloc(sizeof(*st), GFP);
+	if (!st)
+		return ERR_PTR(-ENOMEM);
+
+	if (sg_alloc_table(st, max, GFP)) {
+		kfree(st);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	/*
+	 * Use the optimal page sized chunks to fill in the sg table from large
+	 * to small.
+	 */
+	rem = obj->base.size;
+	sg = st->sgl;
+	st->nents = 0;
+	do {
+		unsigned int page_size = get_largest_page_size(i915, rem);
+		unsigned int len = page_size * (rem / 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;
+
+	return st;
+#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 struct drm_i915_gem_object *
+fake_huge_paged_object(struct drm_i915_private *i915, u64 size)
+{
+	struct drm_i915_gem_object *obj;
+
+	GEM_BUG_ON(!size);
+	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
+
+	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, &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_ppgtt_huge_fill(void *arg)
+{
+	struct i915_hw_ppgtt *ppgtt = arg;
+	struct drm_i915_private *i915 = ppgtt->base.i915;
+	struct drm_i915_gem_object *obj;
+	unsigned long max_pages = ppgtt->base.total >> PAGE_SHIFT;
+	unsigned long page_num;
+	IGT_TIMEOUT(end_time);
+	int err;
+
+	for_each_prime_number_from(page_num, 1, max_pages) {
+		unsigned int expected_gtt = 0;
+		struct i915_vma *vma;
+		typeof(obj->base.size) size;
+		int i;
+
+		obj = fake_huge_paged_object(i915, page_num << PAGE_SHIFT);
+		if (IS_ERR(obj))
+			return PTR_ERR(obj);
+
+		GEM_BUG_ON(obj->base.size != page_num << PAGE_SHIFT);
+
+		err = i915_gem_object_pin_pages(obj);
+		if (err)
+			goto out_put;
+
+		GEM_BUG_ON(!obj->mm.page_sizes.sg);
+
+		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, PIN_USER);
+		if (err) {
+			i915_vma_close(vma);
+			goto out_unpin;
+		}
+
+		GEM_BUG_ON(obj->mm.page_sizes.gtt);
+		GEM_BUG_ON(!vma->page_sizes.sg);
+		GEM_BUG_ON(!vma->page_sizes.phys);
+
+		size = obj->base.size;
+
+		/* 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_SIZE(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;
+
+		if (i915_vm_is_48bit(&ppgtt->base)) {
+			GEM_BUG_ON(vma->page_sizes.gtt != expected_gtt);
+
+			if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
+				GEM_BUG_ON(!IS_ALIGNED(vma->node.start,
+						       I915_GTT_PAGE_SIZE_2M));
+				GEM_BUG_ON(!IS_ALIGNED(vma->node.size,
+						       I915_GTT_PAGE_SIZE_2M));
+			}
+		} else {
+			GEM_BUG_ON(vma->page_sizes.gtt != I915_GTT_PAGE_SIZE_4K);
+			GEM_BUG_ON(vma->node.size != page_num << PAGE_SHIFT);
+		}
+
+		i915_vma_unpin(vma);
+		i915_vma_close(vma);
+
+		i915_gem_object_unpin_pages(obj);
+		i915_gem_object_put(obj);
+
+		if (igt_timeout(end_time,
+				"%s timed out at size %lx\n",
+				__func__, obj->base.size))
+			break;
+	}
+
+	return 0;
+
+out_unpin:
+	i915_gem_object_unpin_pages(obj);
+out_put:
+	i915_gem_object_put(obj);
+
+	return err;
+}
+
+
+static int igt_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_size_mask;
+	struct drm_i915_gem_object *obj;
+	int err;
+	int bit;
+
+	/* 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, BITS_PER_LONG) {
+		IGT_TIMEOUT(end_time);
+		unsigned int page_size = BIT(bit);
+		unsigned int flags = PIN_USER | PIN_OFFSET_FIXED;
+		struct i915_vma *vma;
+		unsigned int offset;
+		unsigned int size =
+			round_up(page_size, I915_GTT_PAGE_SIZE_2M) << 1;
+
+		obj = fake_huge_paged_object(i915, size);
+		if (IS_ERR(obj))
+			return PTR_ERR(obj);
+
+		GEM_BUG_ON(obj->base.size != size);
+
+		err = i915_gem_object_pin_pages(obj);
+		if (err)
+			goto out_put;
+
+		GEM_BUG_ON(!(obj->mm.page_sizes.sg & page_size));
+
+		/* 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;
+		}
+
+		GEM_BUG_ON(vma->page_sizes.gtt != page_size);
+
+		i915_vma_unpin(vma);
+		err = i915_vma_unbind(vma);
+		if (err)
+			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_pin(vma, 0, 0, flags | offset);
+			if (err) {
+				i915_vma_close(vma);
+				goto out_unpin;
+			}
+
+			GEM_BUG_ON(vma->page_sizes.gtt != I915_GTT_PAGE_SIZE_4K);
+
+			i915_vma_unpin(vma);
+			err = i915_vma_unbind(vma);
+			if (err)
+				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 int igt_ppgtt_64K(void *arg)
+{
+	struct i915_hw_ppgtt *ppgtt = arg;
+	struct drm_i915_private *i915 = ppgtt->base.i915;
+	struct drm_i915_gem_object *obj;
+	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_2M - SZ_4K,
+			.gtt = I915_GTT_PAGE_SIZE_4K,
+			.offset = 0,
+		},
+		{
+			.size = SZ_2M + SZ_64K,
+			.gtt = I915_GTT_PAGE_SIZE_64K,
+			.offset = 0,
+		},
+		{
+			.size = SZ_2M + SZ_4K,
+			.gtt = I915_GTT_PAGE_SIZE_64K | I915_GTT_PAGE_SIZE_4K,
+			.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,
+		},
+	};
+	int i;
+	int err;
+
+	if (!HAS_PAGE_SIZE(i915, I915_GTT_PAGE_SIZE_64K))
+		return 0;
+
+	GEM_BUG_ON(!i915_vm_is_48bit(&ppgtt->base));
+
+	/* 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.
+	 */
+
+	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;
+		struct i915_vma *vma;
+		int flags = PIN_USER;
+
+		obj = fake_huge_paged_object(i915, size);
+		if (IS_ERR(obj))
+			return PTR_ERR(obj);
+
+		err = i915_gem_object_pin_pages(obj);
+		if (err)
+			goto out_put;
+
+		GEM_BUG_ON(!obj->mm.page_sizes.sg);
+
+		/* 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_unpin;
+		}
+
+		if (offset)
+			flags |= PIN_OFFSET_FIXED | offset;
+
+		err = i915_vma_pin(vma, 0, 0, flags);
+		if (err) {
+			i915_vma_close(vma);
+			goto out_unpin;
+		}
+
+		GEM_BUG_ON(obj->mm.page_sizes.gtt);
+		GEM_BUG_ON(!vma->page_sizes.sg);
+		GEM_BUG_ON(!vma->page_sizes.phys);
+
+		GEM_BUG_ON(vma->page_sizes.gtt != expected_gtt);
+
+		if (!offset && vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
+			GEM_BUG_ON(!IS_ALIGNED(vma->node.start,
+					       I915_GTT_PAGE_SIZE_2M));
+			GEM_BUG_ON(!IS_ALIGNED(vma->node.size,
+					       I915_GTT_PAGE_SIZE_2M));
+		}
+
+		i915_vma_unpin(vma);
+		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 int igt_ppgtt_gemfs_huge(void *arg)
+{
+	struct i915_hw_ppgtt *ppgtt = arg;
+	struct drm_i915_private *i915 = ppgtt->base.i915;
+	struct drm_i915_gem_object *obj;
+	unsigned int object_sizes[] = {
+		I915_GTT_PAGE_SIZE_2M,
+		I915_GTT_PAGE_SIZE_2M + I915_GTT_PAGE_SIZE_4K,
+	};
+	int err;
+	int i;
+
+	if (!HAS_PAGE_SIZE(i915, I915_GTT_PAGE_SIZE_2M) ||
+	    !IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) ||
+	    !has_transparent_hugepage())
+		return 0;
+
+	/* Sanity check THP through gemfs */
+
+	for (i = 0; i < ARRAY_SIZE(object_sizes); ++i) {
+		unsigned int size = object_sizes[i];
+		struct i915_vma *vma;
+
+		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;
+
+		GEM_BUG_ON(!obj->mm.page_sizes.sg);
+
+		if (obj->mm.page_sizes.sg < size) {
+			pr_err("Failed to allocate huge pages\n");
+			err = -EINVAL;
+			goto out_unpin;
+		}
+
+		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, PIN_USER);
+		if (err) {
+			i915_vma_close(vma);
+			goto out_unpin;
+		}
+
+		/* TODO: maybe we should try writing to the obj from the gpu
+		 * pov, since we require various things to be enabled/disabled
+		 * for huge-pages to work correctly. Also if we force huge-pages
+		 * for gemfs we should be able to exercise 64K pages, since this
+		 * should always give us a 64K aligned memory region.
+		 */
+
+		GEM_BUG_ON(obj->mm.page_sizes.gtt);
+		GEM_BUG_ON(!vma->page_sizes.sg);
+		GEM_BUG_ON(!vma->page_sizes.phys);
+
+		if (i915_vm_is_48bit(&ppgtt->base)) {
+			GEM_BUG_ON(vma->page_sizes.gtt != size);
+			GEM_BUG_ON(!IS_ALIGNED(vma->node.start, I915_GTT_PAGE_SIZE_2M));
+			GEM_BUG_ON(!IS_ALIGNED(vma->node.size, I915_GTT_PAGE_SIZE_2M));
+		} else {
+			GEM_BUG_ON(vma->page_sizes.gtt != I915_GTT_PAGE_SIZE_4K);
+			GEM_BUG_ON(vma->node.size != size);
+		}
+
+		i915_vma_unpin(vma);
+		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;
+}
+
+int i915_gem_huge_page_live_selftests(struct drm_i915_private *dev_priv)
+{
+	static const struct i915_subtest tests[] = {
+		SUBTEST(igt_ppgtt_huge_fill),
+		SUBTEST(igt_ppgtt_misaligned_dma),
+		SUBTEST(igt_ppgtt_64K),
+		SUBTEST(igt_ppgtt_gemfs_huge),
+	};
+	struct i915_hw_ppgtt *ppgtt;
+	struct drm_file *file;
+	int ret;
+
+	file = mock_file(dev_priv);
+	if (IS_ERR(file))
+		return PTR_ERR(file);
+
+	mutex_lock(&dev_priv->drm.struct_mutex);
+	ppgtt = i915_ppgtt_create(dev_priv, file->driver_priv, "mock");
+	if (IS_ERR(ppgtt)) {
+		ret = PTR_ERR(ppgtt);
+		goto out_unlock;
+	}
+
+	ret = i915_subtests(tests, ppgtt);
+
+	i915_ppgtt_close(&ppgtt->base);
+	i915_ppgtt_put(ppgtt);
+
+out_unlock:
+	mutex_unlock(&dev_priv->drm.struct_mutex);
+
+	mock_file_free(dev_priv, file);
+
+	return ret;
+}
diff --git a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h
index 18b174d855ca..6b49cbb49535 100644
--- a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h
+++ b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h
@@ -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(huge, i915_gem_huge_page_live_selftests)
 selftest(contexts, i915_gem_context_live_selftests)
 selftest(hangcheck, intel_hangcheck_live_selftests)
-- 
2.9.4

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx