drm/i915/region: fix max size calculation
authorMatthew Auld <matthew.auld@intel.com>
Wed, 21 Oct 2020 10:36:06 +0000 (11:36 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 21 Oct 2020 18:21:04 +0000 (19:21 +0100)
We are incorrectly limiting the max allocation size as per the mm
max_order, which is effectively the largest power-of-two that we can fit
in the region size. However, it's normal to setup the region or
allocator with a non-power-of-two size(for example 3G), which we should
already handle correctly, except it seems for the early too-big-check.

v2: make sure we also exercise the I915_BO_ALLOC_CONTIGUOUS path, which
is quite different, since for that we are actually limited by the
largest power-of-two that we can fit within the region size. (Chris)

Fixes: b908be543e44 ("drm/i915: support creating LMEM objects")
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: CQ Tang <cq.tang@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20201021103606.241395-1-matthew.auld@intel.com
drivers/gpu/drm/i915/intel_memory_region.c
drivers/gpu/drm/i915/selftests/intel_memory_region.c
drivers/gpu/drm/i915/selftests/mock_region.c

index 6b5e9d8..180e107 100644 (file)
@@ -87,7 +87,7 @@ __intel_memory_region_get_pages_buddy(struct intel_memory_region *mem,
                min_order = ilog2(size) - ilog2(mem->mm.chunk_size);
        }
 
-       if (size > BIT(mem->mm.max_order) * mem->mm.chunk_size)
+       if (size > mem->mm.size)
                return -E2BIG;
 
        n_pages = size >> ilog2(mem->mm.chunk_size);
index 93a38a3..4829c47 100644 (file)
@@ -261,6 +261,82 @@ err_close_objects:
        return err;
 }
 
+static int igt_mock_splintered_region(void *arg)
+{
+       struct intel_memory_region *mem = arg;
+       struct drm_i915_private *i915 = mem->i915;
+       struct drm_i915_gem_object *obj;
+       unsigned int expected_order;
+       LIST_HEAD(objects);
+       u64 size;
+       int err = 0;
+
+       /*
+        * Sanity check we can still allocate everything even if the
+        * mm.max_order != mm.size. i.e our starting address space size is not a
+        * power-of-two.
+        */
+
+       size = (SZ_4G - 1) & PAGE_MASK;
+       mem = mock_region_create(i915, 0, size, PAGE_SIZE, 0);
+       if (IS_ERR(mem))
+               return PTR_ERR(mem);
+
+       if (mem->mm.size != size) {
+               pr_err("%s size mismatch(%llu != %llu)\n",
+                      __func__, mem->mm.size, size);
+               err = -EINVAL;
+               goto out_put;
+       }
+
+       expected_order = get_order(rounddown_pow_of_two(size));
+       if (mem->mm.max_order != expected_order) {
+               pr_err("%s order mismatch(%u != %u)\n",
+                      __func__, mem->mm.max_order, expected_order);
+               err = -EINVAL;
+               goto out_put;
+       }
+
+       obj = igt_object_create(mem, &objects, size, 0);
+       if (IS_ERR(obj)) {
+               err = PTR_ERR(obj);
+               goto out_close;
+       }
+
+       close_objects(mem, &objects);
+
+       /*
+        * While we should be able allocate everything without any flag
+        * restrictions, if we consider I915_BO_ALLOC_CONTIGUOUS then we are
+        * actually limited to the largest power-of-two for the region size i.e
+        * max_order, due to the inner workings of the buddy allocator. So make
+        * sure that does indeed hold true.
+        */
+
+       obj = igt_object_create(mem, &objects, size, I915_BO_ALLOC_CONTIGUOUS);
+       if (!IS_ERR(obj)) {
+               pr_err("%s too large contiguous allocation was not rejected\n",
+                      __func__);
+               err = -EINVAL;
+               goto out_close;
+       }
+
+       obj = igt_object_create(mem, &objects, rounddown_pow_of_two(size),
+                               I915_BO_ALLOC_CONTIGUOUS);
+       if (IS_ERR(obj)) {
+               pr_err("%s largest possible contiguous allocation failed\n",
+                      __func__);
+               err = PTR_ERR(obj);
+               goto out_close;
+       }
+
+out_close:
+       close_objects(mem, &objects);
+out_put:
+       intel_memory_region_put(mem);
+       return err;
+}
+
 static int igt_gpu_write_dw(struct intel_context *ce,
                            struct i915_vma *vma,
                            u32 dword,
@@ -771,6 +847,7 @@ int intel_memory_region_mock_selftests(void)
        static const struct i915_subtest tests[] = {
                SUBTEST(igt_mock_fill),
                SUBTEST(igt_mock_contiguous),
+               SUBTEST(igt_mock_splintered_region),
        };
        struct intel_memory_region *mem;
        struct drm_i915_private *i915;
index 09660f5..979d96f 100644 (file)
@@ -24,7 +24,7 @@ mock_object_create(struct intel_memory_region *mem,
        struct drm_i915_private *i915 = mem->i915;
        struct drm_i915_gem_object *obj;
 
-       if (size > BIT(mem->mm.max_order) * mem->mm.chunk_size)
+       if (size > mem->mm.size)
                return ERR_PTR(-E2BIG);
 
        obj = i915_gem_object_alloc();