[PATCH i-g-t v4] tests/intel/xe_prefetch_fault: Add SVM and hit-under-miss validation

[Varun Gupta <varun.gupta@intel.com>]

Add a second batch to validate hit-under-miss behavior: placing the BB
at PREFETCH_ADDR ensures the page is already mapped when the shader
prefetch runs, so the prefetch fault counter should not increment.

Add SVM mode (prefetch-fault-svm subtest) that creates the VM with
CPU_ADDR_MIRROR and uses mmap() at PREFETCH_ADDR to make the page
GPU-accessible before the hit-under-miss batch.

Other changes:
  - Extract PREFETCH_ADDR and USER_FENCE_VALUE into defines
  - Update stat name to 'invalid_prefetch_pagefault_count'
  - Add missing cleanup (intel_buf_destroy, xe_exec_queue_destroy,
    xe_vm_destroy)

v2:
  - Move xe_vm_create() before if(svm) as it is common to both paths (Nishit)
  - Separate SVM and non-SVM specific code into if/else blocks (Nishit)
  - Add BB_OFFSET and BB_OFFSET_SVM defines (Nishit)
  - Add comments wherever needed (Nishit)

v3:
  - Drop const from bb_offset and bb_offset2 (Nishit)
  - Rename prefetch_pos to prefetch_post (Nishit)

v4:
  - Fix build error: misplaced closing bracket.

Reviewed-by: Nishit Sharma <nishit.sharma@intel.com>
Reviewed-by: Priyanka Dandamudi <priyanka.dandamudi@intel.com>
Signed-off-by: Varun Gupta <varun.gupta@intel.com>
---
 tests/intel/xe_prefetch_fault.c | 129 +++++++++++++++++++++++++-------
 1 file changed, 104 insertions(+), 25 deletions(-)

diff --git a/tests/intel/xe_prefetch_fault.c b/tests/intel/xe_prefetch_fault.c
index 4a143374a..66bcdfe50 100644
--- a/tests/intel/xe_prefetch_fault.c
+++ b/tests/intel/xe_prefetch_fault.c
@@ -25,6 +25,10 @@
 #define WALKER_Y_DIM	1
 #define PAGE_SIZE 4096
 #define COLOR_C4 0xC4C4C4C4
+#define USER_FENCE_VALUE	0xdeadbeefdeadbeefull
+#define PREFETCH_ADDR	0x1f000000
+#define BB_OFFSET	0x1b000000
+#define BB_OFFSET_SVM	0x2b000000
 
 struct dim_t {
 	uint32_t x;
@@ -118,7 +122,7 @@ static struct gpgpu_shader *get_prefetch_shader(int fd)
 	static struct gpgpu_shader *shader;
 
 	shader = gpgpu_shader_create(fd);
-	gpgpu_shader__prefetch_fault(shader, xe_canonical_va(fd, 0x1f000000));
+	gpgpu_shader__prefetch_fault(shader, xe_canonical_va(fd, PREFETCH_ADDR));
 	gpgpu_shader__eot(shader);
 
 	return shader;
@@ -126,63 +130,127 @@ static struct gpgpu_shader *get_prefetch_shader(int fd)
 
 /**
  * SUBTEST: prefetch-fault
- * Description: Validate L1/L2 cache prefetch fault.
+ * Description: Validate prefetch fault and hit-under-miss behavior
+ * Run type: FULL
+ *
+ * SUBTEST: prefetch-fault-svm
+ * Description: Validate prefetch fault and hit-under-miss behavior in SVM mode
  * Run type: FULL
  */
-
-static void test_prefetch(int fd, struct drm_xe_engine_class_instance *hwe)
+static void test_prefetch_fault(int fd, struct drm_xe_engine_class_instance *hwe, bool svm)
 {
-	/* faulty address 0x1f000000 should be beyond bb_offset+bb_size. */
-	const uint64_t bb_offset = 0x1b000000;
+	uint64_t bb_offset = BB_OFFSET;
+	/*
+	 * For the hit-under-miss run, place the batch at PREFETCH_ADDR in
+	 * non-SVM mode so the BO bind maps that page before the shader runs.
+	 * In SVM mode PREFETCH_ADDR is reserved for the mmap, so use a
+	 * separate offset that doesn't collide with it.
+	 */
+	uint64_t bb_offset2 = svm ? BB_OFFSET_SVM : PREFETCH_ADDR;
 	const size_t bb_size = 4096;
-	struct dim_t w_dim;
+	static const char *stat = "invalid_prefetch_pagefault_count";
+	struct dim_t w_dim = { .x = WALKER_X_DIM, .y = WALKER_Y_DIM };
 	struct gpgpu_shader *shader;
 	struct intel_bb *ibb;
 	struct intel_buf *buf;
-	uint32_t *ptr;
 	uint32_t exec_queue_id, vm;
-	int prefetch_pre, prefetch_pos;
-	static const char *stat = "prefetch_pagefault_count";
-
-	w_dim.x = WALKER_X_DIM;
-	w_dim.y = WALKER_Y_DIM;
+	void *cpu_data = NULL;
+	int prefetch_pre, prefetch_post;
+	uint32_t *ptr;
 
 	buf = create_buf(fd, w_dim.x, w_dim.y, COLOR_C4);
 
-	prefetch_pre = xe_gt_stats_get_count(fd, hwe->gt_id, stat);
-
 	vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE |
 			  DRM_XE_VM_CREATE_FLAG_FAULT_MODE, 0);
 
+	if (svm) {
+		/*
+		 * Enable SVM: mirror the full VA space so GPU page faults are
+		 * resolved via HMM against the CPU page tables.
+		 */
+		struct xe_device *xe = xe_device_get(fd);
+		uint64_t vm_sync = 0;
+		struct drm_xe_sync sync[1] = {
+			{ .type = DRM_XE_SYNC_TYPE_USER_FENCE, .flags = DRM_XE_SYNC_FLAG_SIGNAL,
+			  .timeline_value = USER_FENCE_VALUE },
+		};
+
+		sync[0].addr = to_user_pointer(&vm_sync);
+		__xe_vm_bind_assert(fd, vm, 0, 0, 0, 0, 0x1ull << xe->va_bits,
+				    DRM_XE_VM_BIND_OP_MAP,
+				    DRM_XE_VM_BIND_FLAG_CPU_ADDR_MIRROR,
+				    sync, 1, 0, 0);
+		xe_wait_ufence(fd, &vm_sync, USER_FENCE_VALUE, 0, NSEC_PER_SEC);
+	}
+
 	exec_queue_id = xe_exec_queue_create(fd, vm, hwe, 0);
 
-	ibb = xe_bb_create_on_offset(fd, exec_queue_id, vm,
-				     bb_offset, bb_size);
+	prefetch_pre = xe_gt_stats_get_count(fd, hwe->gt_id, stat);
+
+	/* First run: PREFETCH_ADDR is unmapped, so each shader lane raises a prefetch fault. */
+	ibb = xe_bb_create_on_offset(fd, exec_queue_id, vm, bb_offset, bb_size);
 	intel_bb_set_lr_mode(ibb, true);
 
 	shader = get_prefetch_shader(fd);
-
 	gpgpu_shader_exec(ibb, buf, w_dim.x, w_dim.y, shader, NULL, 0, 0);
-
 	gpgpu_shader_destroy(shader);
+	intel_bb_sync(ibb);
+	intel_bb_destroy(ibb);
 
+	prefetch_post = xe_gt_stats_get_count(fd, hwe->gt_id, stat);
+	igt_assert_eq(prefetch_post, prefetch_pre + w_dim.x * w_dim.y);
+
+	/*
+	 * Hit-under-miss: ensure the page at PREFETCH_ADDR is already mapped
+	 * before the prefetch shader runs again. The fault is resolved
+	 * successfully so the prefetch counter must not change.
+	 *
+	 * SVM:     mmap at PREFETCH_ADDR creates a CPU page table entry.
+	 *          The kernel resolves the GPU pagefault via HMM.
+	 * Non-SVM: placing the batch buffer at PREFETCH_ADDR causes the
+	 *          BO fault path to map the page.
+	 */
+	if (svm) {
+		cpu_data = mmap((void *)PREFETCH_ADDR, PAGE_SIZE,
+				PROT_READ | PROT_WRITE,
+				MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED,
+				-1, 0);
+		igt_assert(cpu_data == (void *)PREFETCH_ADDR);
+		/* Touch the page to populate the CPU PTE so HMM can resolve the GPU fault. */
+		memset(cpu_data, 0xAB, PAGE_SIZE);
+		prefetch_pre = xe_gt_stats_get_count(fd, hwe->gt_id, stat);
+	} else {
+		prefetch_pre = prefetch_post;
+	}
+	ibb = xe_bb_create_on_offset(fd, exec_queue_id, vm, bb_offset2, bb_size);
+	intel_bb_set_lr_mode(ibb, true);
+
+	shader = get_prefetch_shader(fd);
+	gpgpu_shader_exec(ibb, buf, w_dim.x, w_dim.y, shader, NULL, 0, 0);
+	gpgpu_shader_destroy(shader);
 	intel_bb_sync(ibb);
 
-	ptr = xe_bo_mmap_ext(fd, buf->handle, buf->size, PROT_READ);
+	prefetch_post = xe_gt_stats_get_count(fd, hwe->gt_id, stat);
+	igt_assert_eq(prefetch_post, prefetch_pre);
 
+	/* Verify buffer contents */
+	ptr = xe_bo_mmap_ext(fd, buf->handle, buf->size, PROT_READ);
 	for (int j = 0; j < w_dim.y; j++)
 		for (int i = 0; i < w_dim.x; i++) {
 			igt_assert_f(ptr[j * w_dim.x + i] == COLOR_C4,
 				     "Expected 0x%02x, found 0x%02x at (%d,%d)\n",
 				     COLOR_C4, ptr[j * w_dim.x + i], i, j);
 		}
-	/* Validate prefetch count. */
-	prefetch_pos = xe_gt_stats_get_count(fd, hwe->gt_id, stat);
-	igt_assert_eq(prefetch_pos, prefetch_pre + w_dim.x * w_dim.y);
-
 	munmap(ptr, buf->size);
 
+	/* Cleanup */
+	if (svm && cpu_data)
+		munmap(cpu_data, PAGE_SIZE);
+
 	intel_bb_destroy(ibb);
+	intel_buf_destroy(buf);
+	xe_exec_queue_destroy(fd, exec_queue_id);
+	xe_vm_destroy(fd, vm);
 }
 
 int igt_main()
@@ -201,7 +269,18 @@ int igt_main()
 			    hwe->engine_class == DRM_XE_ENGINE_CLASS_COMPUTE) {
 				igt_dynamic_f("%s%d", xe_engine_class_string(hwe->engine_class),
 					      hwe->engine_instance)
-					test_prefetch(fd, hwe);
+					test_prefetch_fault(fd, hwe, false);
+			}
+		}
+	}
+
+	igt_subtest_with_dynamic("prefetch-fault-svm") {
+		xe_for_each_engine(fd, hwe) {
+			if (hwe->engine_class == DRM_XE_ENGINE_CLASS_RENDER ||
+			    hwe->engine_class == DRM_XE_ENGINE_CLASS_COMPUTE) {
+				igt_dynamic_f("%s%d", xe_engine_class_string(hwe->engine_class),
+					      hwe->engine_instance)
+					test_prefetch_fault(fd, hwe, true);
 			}
 		}
 	}
-- 
2.43.0