[PATCH] gpu/tests/gpu_buddy: Add gpu_test_buddy_alloc_range for exact-range allocation

[PATCH] gpu/tests/gpu_buddy: Add gpu_test_buddy_alloc_range for exact-range allocation

[Sanjay Yadav]

Add a new kunit test gpu_test_buddy_alloc_range() that exercises the
__gpu_buddy_alloc_range() exact-range allocation path, triggered when
start + size == end with flags=0.

The test covers:
- Basic exact-range allocation of the full mm
- Exact-range allocation of equal sub-ranges (quarters)
- Minimum chunk-size exact ranges at start, middle, and end offsets
- Non power-of-two mm size with multiple roots, including cross-root
  exact-range allocation
- Randomized exact-range allocations of N contiguous page-aligned
  slices in random order
- Negative: partially allocated range must reject overlapping exact
  alloc
- Negative: checkerboard allocation pattern rejects exact range over
  partially occupied pairs
- Negative: misaligned start, unaligned size, and out-of-bounds end
- Free and re-allocate the same exact range across multiple iterations
- Various power-of-two exact ranges at natural alignment

Cc: Christian König <christian.koenig@amd.com>
Cc: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
Suggested-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Sanjay Yadav <sanjay.kumar.yadav@intel.com>
---
 drivers/gpu/tests/gpu_buddy_test.c | 327 +++++++++++++++++++++++++++++
 1 file changed, 327 insertions(+)

diff --git a/drivers/gpu/tests/gpu_buddy_test.c b/drivers/gpu/tests/gpu_buddy_test.c
index 5429010f34d3..9738fd09518f 100644
--- a/drivers/gpu/tests/gpu_buddy_test.c
+++ b/drivers/gpu/tests/gpu_buddy_test.c
@@ -362,6 +362,332 @@ static void gpu_test_buddy_alloc_range_bias(struct kunit *test)
 	gpu_buddy_fini(&mm);
 }
 
+static void gpu_test_buddy_alloc_range(struct kunit *test)
+{
+	GPU_RND_STATE(prng, random_seed);
+	struct gpu_buddy_block *block;
+	struct gpu_buddy mm;
+	u32 mm_size, total;
+	LIST_HEAD(blocks);
+	LIST_HEAD(tmp);
+	u32 ps = SZ_4K;
+	int ret;
+
+	mm_size = SZ_16M;
+
+	KUNIT_ASSERT_FALSE_MSG(test, gpu_buddy_init(&mm, mm_size, ps),
+			       "buddy_init failed\n");
+
+	/*
+	 * Basic exact-range allocation.
+	 * Allocate the entire mm as one exact range (start + size == end).
+	 * This is the simplest case exercising __gpu_buddy_alloc_range.
+	 */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, mm_size, mm_size, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ_MSG(test, ret, 0,
+			    "exact-range alloc of full mm failed\n");
+
+	total = 0;
+	list_for_each_entry(block, &blocks, link) {
+		u64 offset = gpu_buddy_block_offset(block);
+		u64 bsize = gpu_buddy_block_size(&mm, block);
+
+		KUNIT_EXPECT_TRUE_MSG(test, offset + bsize <= (u64)mm_size,
+				      "block [%llx, %llx) outside mm\n", offset, offset + bsize);
+		total += (u32)bsize;
+	}
+	KUNIT_EXPECT_EQ(test, total, mm_size);
+	KUNIT_EXPECT_EQ(test, mm.avail, 0ULL);
+
+	/* Full mm should be exhausted */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, ps, ps, ps, &tmp, 0);
+	KUNIT_EXPECT_NE_MSG(test, ret, 0, "alloc should fail when mm is full\n");
+
+	gpu_buddy_free_list(&mm, &blocks, 0);
+	KUNIT_EXPECT_EQ(test, mm.avail, (u64)mm_size);
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Exact-range allocation of sub-ranges.
+	 * Split the mm into four equal quarters and allocate each as an exact
+	 * range. Validates splitting and non-overlapping exact allocations.
+	 */
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	{
+		u32 quarter = mm_size / 4;
+		int i;
+
+		for (i = 0; i < 4; i++) {
+			u32 start = i * quarter;
+			u32 end = start + quarter;
+
+			ret = gpu_buddy_alloc_blocks(&mm, start, end, quarter, ps, &blocks, 0);
+			KUNIT_ASSERT_EQ_MSG(test, ret, 0,
+					    "exact-range alloc quarter %d [%x, %x) failed\n",
+					    i, start, end);
+		}
+		KUNIT_EXPECT_EQ(test, mm.avail, 0ULL);
+		gpu_buddy_free_list(&mm, &blocks, 0);
+	}
+
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Minimum chunk-size exact range at various offsets.
+	 * Allocate single-page exact ranges at the start, middle and end.
+	 */
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	ret = gpu_buddy_alloc_blocks(&mm, 0, ps, ps, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	ret = gpu_buddy_alloc_blocks(&mm, mm_size / 2, mm_size / 2 + ps, ps, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	ret = gpu_buddy_alloc_blocks(&mm, mm_size - ps, mm_size, ps, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	total = 0;
+	list_for_each_entry(block, &blocks, link)
+		total += (u32)gpu_buddy_block_size(&mm, block);
+	KUNIT_EXPECT_EQ(test, total, 3 * ps);
+
+	gpu_buddy_free_list(&mm, &blocks, 0);
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Non power-of-two mm size (multiple roots).
+	 * Exact-range allocations that span root boundaries must still work.
+	 */
+	mm_size = SZ_4M + SZ_2M + SZ_1M; /* 7 MiB, three roots */
+
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+	KUNIT_EXPECT_GT(test, mm.n_roots, 1U);
+
+	/* Allocate first 4M root exactly */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, SZ_4M, SZ_4M, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	/* Allocate second root (4M-6M) exactly */
+	ret = gpu_buddy_alloc_blocks(&mm, SZ_4M, SZ_4M + SZ_2M, SZ_2M, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	/* Allocate third root (6M-7M) exactly */
+	ret = gpu_buddy_alloc_blocks(&mm, SZ_4M + SZ_2M, mm_size, SZ_1M, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	KUNIT_EXPECT_EQ(test, mm.avail, 0ULL);
+	gpu_buddy_free_list(&mm, &blocks, 0);
+
+	/* Cross-root exact-range: the entire non-pot mm */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, mm_size, mm_size, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+	KUNIT_EXPECT_EQ(test, mm.avail, 0ULL);
+
+	gpu_buddy_free_list(&mm, &blocks, 0);
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Randomized exact-range allocations.
+	 * Divide the mm into N random-sized, contiguous, page-aligned slices
+	 * and allocate each as an exact range in random order.
+	 */
+	mm_size = SZ_16M;
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	{
+#define N_RAND_RANGES 16
+		u32 ranges[N_RAND_RANGES + 1]; /* boundaries */
+		u32 order_arr[N_RAND_RANGES];
+		u32 remaining = mm_size;
+		int i;
+
+		ranges[0] = 0;
+		for (i = 0; i < N_RAND_RANGES - 1; i++) {
+			u32 max_chunk = remaining - (N_RAND_RANGES - 1 - i) * ps;
+			u32 sz = max(round_up(prandom_u32_state(&prng) % max_chunk, ps), ps);
+
+			ranges[i + 1] = ranges[i] + sz;
+			remaining -= sz;
+		}
+		ranges[N_RAND_RANGES] = mm_size;
+
+		/* Create a random order */
+		for (i = 0; i < N_RAND_RANGES; i++)
+			order_arr[i] = i;
+		for (i = N_RAND_RANGES - 1; i > 0; i--) {
+			u32 j = prandom_u32_state(&prng) % (i + 1);
+			u32 tmp_val = order_arr[i];
+
+			order_arr[i] = order_arr[j];
+			order_arr[j] = tmp_val;
+		}
+
+		for (i = 0; i < N_RAND_RANGES; i++) {
+			u32 idx = order_arr[i];
+			u32 start = ranges[idx];
+			u32 end = ranges[idx + 1];
+			u32 sz = end - start;
+
+			ret = gpu_buddy_alloc_blocks(&mm, start, end, sz, ps, &blocks, 0);
+			KUNIT_ASSERT_EQ_MSG(test, ret, 0,
+					    "random exact-range [%x, %x) sz=%x failed\n",
+					    start, end, sz);
+		}
+
+		KUNIT_EXPECT_EQ(test, mm.avail, 0ULL);
+		gpu_buddy_free_list(&mm, &blocks, 0);
+#undef N_RAND_RANGES
+	}
+
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Negative case - partially allocated range.
+	 * Allocate the first half, then try to exact-range allocate the full
+	 * mm. This must fail because the first half is already occupied.
+	 */
+	mm_size = SZ_16M;
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	ret = gpu_buddy_alloc_blocks(&mm, 0, mm_size / 2, mm_size / 2, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	ret = gpu_buddy_alloc_blocks(&mm, 0, mm_size, mm_size, ps, &tmp, 0);
+	KUNIT_EXPECT_NE_MSG(test, ret, 0,
+			    "exact-range alloc should fail when range is partially used\n");
+
+	/* Also try the already-occupied sub-range directly */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, mm_size / 2, mm_size / 2, ps, &tmp, 0);
+	KUNIT_EXPECT_NE_MSG(test, ret, 0,
+			    "double alloc of same exact range should fail\n");
+
+	/* The free second half should still be allocatable */
+	ret = gpu_buddy_alloc_blocks(&mm, mm_size / 2, mm_size, mm_size / 2, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	KUNIT_EXPECT_EQ(test, mm.avail, 0ULL);
+	gpu_buddy_free_list(&mm, &blocks, 0);
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Negative case - checkerboard partial allocation.
+	 * Allocate every other page-sized chunk in a small mm, then try to
+	 * exact-range allocate a range covering two pages (one allocated, one
+	 * free). This must fail.
+	 */
+	mm_size = SZ_64K;
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	{
+		u32 off;
+
+		for (off = 0; off < mm_size; off += 2 * ps) {
+			ret = gpu_buddy_alloc_blocks(&mm, off, off + ps, ps, ps, &blocks, 0);
+			KUNIT_ASSERT_EQ(test, ret, 0);
+		}
+
+		/* Try exact range over a pair [allocated, free] */
+		ret = gpu_buddy_alloc_blocks(&mm, 0, 2 * ps, 2 * ps, ps, &tmp, 0);
+		KUNIT_EXPECT_NE_MSG(test, ret, 0,
+				    "exact-range over partially allocated pair should fail\n");
+
+		/* The free pages individually should still work */
+		ret = gpu_buddy_alloc_blocks(&mm, ps, 2 * ps, ps, ps, &blocks, 0);
+		KUNIT_ASSERT_EQ(test, ret, 0);
+
+		gpu_buddy_free_list(&mm, &blocks, 0);
+	}
+
+	gpu_buddy_fini(&mm);
+
+	/* Negative case - misaligned start/end/size */
+	mm_size = SZ_16M;
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	/* start not aligned to chunk_size */
+	ret = gpu_buddy_alloc_blocks(&mm, ps / 2, ps / 2 + ps, ps, ps, &tmp, 0);
+	KUNIT_EXPECT_NE(test, ret, 0);
+
+	/* size not aligned */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, ps + 1, ps + 1, ps, &tmp, 0);
+	KUNIT_EXPECT_NE(test, ret, 0);
+
+	/* end exceeds mm size */
+	ret = gpu_buddy_alloc_blocks(&mm, mm_size, mm_size + ps, ps, ps, &tmp, 0);
+	KUNIT_EXPECT_NE(test, ret, 0);
+
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Free and re-allocate the same exact range.
+	 * This exercises merge-on-free followed by exact-range re-split.
+	 */
+	mm_size = SZ_16M;
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	{
+		int i;
+
+		for (i = 0; i < 5; i++) {
+			ret = gpu_buddy_alloc_blocks(&mm, SZ_4M, SZ_4M + SZ_2M,
+						     SZ_2M, ps, &blocks, 0);
+			KUNIT_ASSERT_EQ_MSG(test, ret, 0,
+					    "re-alloc iteration %d failed\n", i);
+
+			total = 0;
+			list_for_each_entry(block, &blocks, link) {
+				u64 offset = gpu_buddy_block_offset(block);
+				u64 bsize = gpu_buddy_block_size(&mm, block);
+
+				KUNIT_EXPECT_GE(test, offset, (u64)SZ_4M);
+				KUNIT_EXPECT_LE(test, offset + bsize, (u64)(SZ_4M + SZ_2M));
+				total += (u32)bsize;
+			}
+			KUNIT_EXPECT_EQ(test, total, SZ_2M);
+
+			gpu_buddy_free_list(&mm, &blocks, 0);
+		}
+
+		KUNIT_EXPECT_EQ(test, mm.avail, (u64)mm_size);
+	}
+
+	gpu_buddy_fini(&mm);
+
+	/*
+	 * Various power-of-two exact ranges within a large mm.
+	 * Allocate non-overlapping power-of-two exact ranges at their natural
+	 * alignment, validating that the allocator handles different orders.
+	 */
+	mm_size = SZ_16M;
+	KUNIT_ASSERT_FALSE(test, gpu_buddy_init(&mm, mm_size, ps));
+
+	/* Allocate 4K at offset 0 */
+	ret = gpu_buddy_alloc_blocks(&mm, 0, SZ_4K, SZ_4K, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	/* Allocate 64K at offset 64K */
+	ret = gpu_buddy_alloc_blocks(&mm, SZ_64K, SZ_64K + SZ_64K, SZ_64K, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	/* Allocate 1M at offset 1M */
+	ret = gpu_buddy_alloc_blocks(&mm, SZ_1M, SZ_1M + SZ_1M, SZ_1M, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	/* Allocate 4M at offset 4M */
+	ret = gpu_buddy_alloc_blocks(&mm, SZ_4M, SZ_4M + SZ_4M, SZ_4M, ps, &blocks, 0);
+	KUNIT_ASSERT_EQ(test, ret, 0);
+
+	total = 0;
+	list_for_each_entry(block, &blocks, link)
+		total += (u32)gpu_buddy_block_size(&mm, block);
+	KUNIT_EXPECT_EQ(test, total, SZ_4K + SZ_64K + SZ_1M + SZ_4M);
+
+	gpu_buddy_free_list(&mm, &blocks, 0);
+	gpu_buddy_fini(&mm);
+}
+
 static void gpu_test_buddy_alloc_clear(struct kunit *test)
 {
 	unsigned long n_pages, total, i = 0;
@@ -909,6 +1235,7 @@ static struct kunit_case gpu_buddy_tests[] = {
 	KUNIT_CASE(gpu_test_buddy_alloc_pathological),
 	KUNIT_CASE(gpu_test_buddy_alloc_contiguous),
 	KUNIT_CASE(gpu_test_buddy_alloc_clear),
+	KUNIT_CASE(gpu_test_buddy_alloc_range),
 	KUNIT_CASE(gpu_test_buddy_alloc_range_bias),
 	KUNIT_CASE_SLOW(gpu_test_buddy_fragmentation_performance),
 	KUNIT_CASE(gpu_test_buddy_alloc_exceeds_max_order),
-- 
2.52.0

Claude review: gpu/tests/gpu_buddy: Add gpu_test_buddy_alloc_range for exact-range allocation

[Claude Code Review Bot]

Overall Series Review

Subject: gpu/tests/gpu_buddy: Add gpu_test_buddy_alloc_range for exact-range allocation
Author: Sanjay Yadav <sanjay.kumar.yadav@intel.com>
Patches: 1
Reviewed: 2026-03-03T13:14:20.435073

---

This is a single patch adding a new KUnit test function `gpu_test_buddy_alloc_range()` that exercises the exact-range allocation path in the GPU buddy allocator (the `start + size == end` codepath that dispatches to `__gpu_buddy_alloc_range()`). The test is well-structured with good scenario coverage including positive, negative, and randomized cases. The code is consistent with existing test style (e.g., `gpu_test_buddy_alloc_range_bias`) and uses the same variable types (`u32` for sizes) and patterns.

The test is generally clean and correct, with only minor observations below.

---
Generated by Claude Code Patch Reviewer

Claude review: gpu/tests/gpu_buddy: Add gpu_test_buddy_alloc_range for exact-range allocation

[Claude Code Review Bot]

Patch Review

**Positive aspects:**

- Comprehensive coverage: basic full-mm, sub-range quarters, min-chunk at multiple offsets, non-power-of-two mm sizes (multi-root), randomized slices, negative overlap cases, checkerboard patterns, misalignment rejection, free/re-alloc cycling, and mixed power-of-two sizes.
- Negative tests correctly trigger the exact-range path (they satisfy `start + size == end`) and verify that the appropriate validation catches them.
- The randomized range generation maintains the invariant that `max_chunk >= ps > 0` (since both `remaining` and `(N_RAND_RANGES - 1 - i) * ps` are always multiples of `ps`), so there is no division-by-zero or underflow risk. I verified this carefully.

**Observations and minor issues:**

1. **Leaked blocks on unexpected negative-test success.** In several negative tests, the allocation result goes into `tmp`, but if the allocation unexpectedly *succeeds* (indicating an allocator bug), those blocks are never freed before `gpu_buddy_fini()`:

```c
ret = gpu_buddy_alloc_blocks(&mm, 0, mm_size, mm_size, ps, &tmp, 0);
KUNIT_EXPECT_NE_MSG(test, ret, 0,
		    "exact-range alloc should fail when range is partially used\n");
```

This is consistent with how `gpu_test_buddy_alloc_range_bias` handles its negative cases, so it's an existing pattern. But it means a failing EXPECT won't trigger the WARN in `gpu_buddy_fini()` about unreleased blocks — consider adding `gpu_buddy_free_list(&mm, &tmp, 0)` after each negative assertion, or at least before `gpu_buddy_fini()`, to make failures easier to debug.

2. **Function length.** At 327 lines, the function is quite long. The sub-test blocks are already scoped with `{ }` for variable locality, which is good, but consider whether splitting into separate `kunit_case` functions (e.g., `gpu_test_buddy_alloc_range_negative`, `gpu_test_buddy_alloc_range_random`) would improve readability and give finer-grained pass/fail reporting from KUnit. This is a stylistic suggestion, not a blocker.

3. **Test ordering in the case array.** The new test is inserted *before* `gpu_test_buddy_alloc_range_bias`:

```c
KUNIT_CASE(gpu_test_buddy_alloc_clear),
KUNIT_CASE(gpu_test_buddy_alloc_range),
KUNIT_CASE(gpu_test_buddy_alloc_range_bias),
```

This is a reasonable grouping. However, the name `gpu_test_buddy_alloc_range` is very similar to `gpu_test_buddy_alloc_range_bias` and could cause confusion since the former tests the "exact-range" path (flags=0, start+size==end) while the latter tests range-biased allocation (GPU_BUDDY_RANGE_ALLOCATION flag). A more descriptive name like `gpu_test_buddy_alloc_range_exact` would clarify intent and better distinguish the two.

4. **Negative misalignment tests.** The three misalignment sub-tests are good but it's worth noting they are all caught by the *general* alignment check at `buddy.c:1117`:

```c
if (!IS_ALIGNED(start | end | size, mm->chunk_size))
    return -EINVAL;
```

rather than the exact-range-specific check at line 1128. The test still validates correct rejection, but a comment clarifying this would help future maintainers understand which validation layer is being exercised. If the intent is to specifically test the exact-range alignment check on line 1128, you'd need start/end/size all aligned to `chunk_size` but start not aligned to `min_block_size` (when `min_block_size > chunk_size`).

5. **Minor style:** The commit subject line prefix `gpu/tests/gpu_buddy:` differs slightly from what you might see as a conventional `drm/buddy:` or `drm/tests:` prefix. Check the existing convention for this test file in the git log.

**Correctness verification of key test logic:**

- Quarter allocation: All four allocations go into the same `blocks` list; `gpu_buddy_free_list` correctly frees them all. ✓
- Checkerboard: Allocates every other page into `blocks`, tries exact-range over a partially-occupied pair, correctly verifies failure, then allocates the free page. The final `gpu_buddy_free_list(&mm, &blocks, 0)` frees both the checkerboard and the single free-page allocation. ✓
- Randomized: `round_up(prandom_u32_state(&prng) % max_chunk, ps)` — since `max_chunk` is always a multiple of `ps`, the rounded-up result never exceeds `max_chunk`, preserving the budget invariant. ✓
- Out-of-bounds negative test: caught by `end > mm->size` check at `buddy.c:1120`. ✓

**Verdict:** The patch is correct and provides valuable test coverage for the exact-range allocation path. The suggestions above (naming, tmp cleanup, possible split) are minor improvements. No blocking issues.

---
Generated by Claude Code Patch Reviewer