[Intel-gfx] [PATCH i-g-t v2 0/6] Exercise NEEDS_CPU_ACCESS

[Intel-gfx] [PATCH i-g-t v2 0/6] Exercise NEEDS_CPU_ACCESS

[Matthew Auld]

Add some initial tests for the new I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS
flag.

v2: Add missing cover letter. Likely also needed when using 'Test-with:'

-- 
2.34.1

[Intel-gfx] [PATCH i-g-t v2 1/6] lib/i915_drm_local: Add I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS

[Matthew Auld]

For now dump into i915_drm_local.h. Once the uapi on the kernel side is
merged, and is part of drm-next, we can sync the kernel headers and
remove this.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 lib/i915/i915_drm_local.h | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)

diff --git a/lib/i915/i915_drm_local.h b/lib/i915/i915_drm_local.h
index 9e82c968..7b5285f3 100644
--- a/lib/i915/i915_drm_local.h
+++ b/lib/i915/i915_drm_local.h
@@ -21,6 +21,33 @@ extern "C" {
  */
 #define I915_ENGINE_CLASS_COMPUTE 4
 
+/*
+ * Signal to the kernel that the object will need to be accessed via
+ * the CPU.
+ *
+ * Only valid when placing objects in I915_MEMORY_CLASS_DEVICE, and only
+ * strictly required on platforms where only some of the device memory
+ * is directly visible or mappable through the CPU, like on DG2+.
+ *
+ * One of the placements MUST also be I915_MEMORY_CLASS_SYSTEM, to
+ * ensure we can always spill the allocation to system memory, if we
+ * can't place the object in the mappable part of
+ * I915_MEMORY_CLASS_DEVICE.
+ *
+ * Note that buffers that need to be captured with EXEC_OBJECT_CAPTURE,
+ * will need to enable this hint, if the object can also be placed in
+ * I915_MEMORY_CLASS_DEVICE, starting from DG2+. The execbuf call will
+ * throw an error otherwise. This also means that such objects will need
+ * I915_MEMORY_CLASS_SYSTEM set as a possible placement.
+ *
+ * Without this hint, the kernel will assume that non-mappable
+ * I915_MEMORY_CLASS_DEVICE is preferred for this object. Note that the
+ * kernel can still migrate the object to the mappable part, as a last
+ * resort, if userspace ever CPU faults this object, but this might be
+ * expensive, and so ideally should be avoided.
+ */
+#define I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS (1 << 0)
+
 #if defined(__cplusplus)
 }
 #endif
-- 
2.34.1

[Intel-gfx] [PATCH i-g-t v2 2/6] lib/i915: wire up optional flags for gem_create_ext

[Matthew Auld]

For now limit to direct callers.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 lib/i915/gem_create.c          |  9 ++++++---
 lib/i915/gem_create.h          |  5 +++--
 lib/i915/intel_memory_region.c |  2 +-
 tests/i915/gem_create.c        | 24 ++++++++++++------------
 tests/i915/gem_pxp.c           |  2 +-
 5 files changed, 23 insertions(+), 19 deletions(-)

diff --git a/lib/i915/gem_create.c b/lib/i915/gem_create.c
index b2e8d559..d245a32e 100644
--- a/lib/i915/gem_create.c
+++ b/lib/i915/gem_create.c
@@ -48,11 +48,12 @@ uint32_t gem_create(int fd, uint64_t size)
 	return handle;
 }
 
-int __gem_create_ext(int fd, uint64_t *size, uint32_t *handle,
+int __gem_create_ext(int fd, uint64_t *size, uint32_t flags, uint32_t *handle,
 		     struct i915_user_extension *ext)
 {
 	struct drm_i915_gem_create_ext create = {
 		.size = *size,
+		.flags = flags,
 		.extensions = to_user_pointer(ext),
 	};
 	int err = 0;
@@ -73,6 +74,7 @@ int __gem_create_ext(int fd, uint64_t *size, uint32_t *handle,
  * gem_create_ext:
  * @fd: open i915 drm file descriptor
  * @size: desired size of the buffer
+ * @flags: optional flags
  * @ext: optional extensions chain
  *
  * This wraps the GEM_CREATE_EXT ioctl, which allocates a new gem buffer object
@@ -80,11 +82,12 @@ int __gem_create_ext(int fd, uint64_t *size, uint32_t *handle,
  *
  * Returns: The file-private handle of the created buffer object
  */
-uint32_t gem_create_ext(int fd, uint64_t size, struct i915_user_extension *ext)
+uint32_t gem_create_ext(int fd, uint64_t size, uint32_t flags,
+			struct i915_user_extension *ext)
 {
 	uint32_t handle;
 
-	igt_assert_eq(__gem_create_ext(fd, &size, &handle, ext), 0);
+	igt_assert_eq(__gem_create_ext(fd, &size, flags, &handle, ext), 0);
 
 	return handle;
 }
diff --git a/lib/i915/gem_create.h b/lib/i915/gem_create.h
index c2b531b4..02232693 100644
--- a/lib/i915/gem_create.h
+++ b/lib/i915/gem_create.h
@@ -12,8 +12,9 @@
 
 int __gem_create(int fd, uint64_t *size, uint32_t *handle);
 uint32_t gem_create(int fd, uint64_t size);
-int __gem_create_ext(int fd, uint64_t *size, uint32_t *handle,
+int __gem_create_ext(int fd, uint64_t *size, uint32_t flags, uint32_t *handle,
                      struct i915_user_extension *ext);
-uint32_t gem_create_ext(int fd, uint64_t size, struct i915_user_extension *ext);
+uint32_t gem_create_ext(int fd, uint64_t size, uint32_t flags,
+			struct i915_user_extension *ext);
 
 #endif /* GEM_CREATE_H */
diff --git a/lib/i915/intel_memory_region.c b/lib/i915/intel_memory_region.c
index a8759e06..f0c8bc7c 100644
--- a/lib/i915/intel_memory_region.c
+++ b/lib/i915/intel_memory_region.c
@@ -208,7 +208,7 @@ int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size,
 	};
 	int ret;
 
-	ret = __gem_create_ext(fd, size, handle, &ext_regions.base);
+	ret = __gem_create_ext(fd, size, 0, handle, &ext_regions.base);
 
 	/*
 	 * Provide fallback for stable kernels if region passed in the array
diff --git a/tests/i915/gem_create.c b/tests/i915/gem_create.c
index 45804cde..a6c3c9d9 100644
--- a/tests/i915/gem_create.c
+++ b/tests/i915/gem_create.c
@@ -331,38 +331,38 @@ static void create_ext_placement_sanity_check(int fd)
 	 * behaviour.
 	 */
 	size = PAGE_SIZE;
-	igt_assert_eq(__gem_create_ext(fd, &size, &handle, 0), 0);
+	igt_assert_eq(__gem_create_ext(fd, &size, 0, &handle, 0), 0);
 	gem_close(fd, handle);
 
 	/* Try some uncreative invalid combinations */
 	setparam_region.regions = to_user_pointer(&region_smem);
 	setparam_region.num_regions = 0;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	setparam_region.regions = to_user_pointer(&region_smem);
 	setparam_region.num_regions = regions->num_regions + 1;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	setparam_region.regions = to_user_pointer(&region_smem);
 	setparam_region.num_regions = -1;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	setparam_region.regions = to_user_pointer(&region_invalid);
 	setparam_region.num_regions = 1;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	setparam_region.regions = to_user_pointer(&region_invalid);
 	setparam_region.num_regions = 0;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	uregions = calloc(regions->num_regions + 1, sizeof(uint32_t));
@@ -373,7 +373,7 @@ static void create_ext_placement_sanity_check(int fd)
 	setparam_region.regions = to_user_pointer(uregions);
 	setparam_region.num_regions = regions->num_regions + 1;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	if (regions->num_regions > 1)  {
@@ -386,7 +386,7 @@ static void create_ext_placement_sanity_check(int fd)
 			setparam_region.regions = to_user_pointer(dups);
 			setparam_region.num_regions = 2;
 			size = PAGE_SIZE;
-			igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+			igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 							&setparam_region.base), 0);
 		}
 	}
@@ -396,7 +396,7 @@ static void create_ext_placement_sanity_check(int fd)
 	setparam_region.regions = to_user_pointer(uregions);
 	setparam_region.num_regions = regions->num_regions;
 	size = PAGE_SIZE;
-	igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 					&setparam_region.base), 0);
 
 	free(uregions);
@@ -412,7 +412,7 @@ static void create_ext_placement_sanity_check(int fd)
 				to_user_pointer(&setparam_region_next);
 
 		size = PAGE_SIZE;
-		igt_assert_neq(__gem_create_ext(fd, &size, &handle,
+		igt_assert_neq(__gem_create_ext(fd, &size, 0, &handle,
 						&setparam_region.base), 0);
 		setparam_region.base.next_extension = 0;
 	}
@@ -444,7 +444,7 @@ static void create_ext_placement_all(int fd)
 	setparam_region.num_regions = regions->num_regions;
 
 	size = PAGE_SIZE;
-	igt_assert_eq(__gem_create_ext(fd, &size, &handle,
+	igt_assert_eq(__gem_create_ext(fd, &size, 0, &handle,
 				       &setparam_region.base), 0);
 	gem_close(fd, handle);
 	free(uregions);
@@ -473,7 +473,7 @@ static void create_ext_placement_each(int fd)
 		setparam_region.num_regions = 1;
 
 		size = PAGE_SIZE;
-		igt_assert_eq(__gem_create_ext(fd, &size, &handle,
+		igt_assert_eq(__gem_create_ext(fd, &size, 0, &handle,
 					       &setparam_region.base), 0);
 		gem_close(fd, handle);
 	}
diff --git a/tests/i915/gem_pxp.c b/tests/i915/gem_pxp.c
index 5f269bab..65618556 100644
--- a/tests/i915/gem_pxp.c
+++ b/tests/i915/gem_pxp.c
@@ -40,7 +40,7 @@ static int create_bo_ext(int i915, uint32_t size, bool protected_is_true, uint32
 		ext = &protected_ext.base;
 
 	*bo_out = 0;
-	ret = __gem_create_ext(i915, &size64, bo_out, ext);
+	ret = __gem_create_ext(i915, &size64, 0, bo_out, ext);
 
 	return ret;
 }
-- 
2.34.1

[Intel-gfx] [PATCH i-g-t v2 3/6] tests/i915/gem_create: test NEEDS_CPU_ACCESS

[Matthew Auld]

Add some basic tests for this new flag.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 tests/i915/gem_create.c | 334 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 334 insertions(+)

diff --git a/tests/i915/gem_create.c b/tests/i915/gem_create.c
index a6c3c9d9..318e6491 100644
--- a/tests/i915/gem_create.c
+++ b/tests/i915/gem_create.c
@@ -43,6 +43,8 @@
 #include <getopt.h>
 #include <pthread.h>
 #include <stdatomic.h>
+#include <setjmp.h>
+#include <signal.h>
 
 #include "drm.h"
 #include "drmtest.h"
@@ -481,6 +483,327 @@ static void create_ext_placement_each(int fd)
 	free(regions);
 }
 
+static bool supports_needs_cpu_access(int fd)
+{
+	struct drm_i915_gem_memory_class_instance regions[] = {
+		{ I915_MEMORY_CLASS_DEVICE, },
+		{ I915_MEMORY_CLASS_SYSTEM, },
+	};
+	struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+		.base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+		.regions = to_user_pointer(&regions),
+		.num_regions = ARRAY_SIZE(regions),
+	};
+	uint64_t size = PAGE_SIZE;
+	uint32_t handle;
+	int ret;
+
+	ret = __gem_create_ext(fd, &size,
+			       I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+			       &handle, &setparam_region.base);
+	if (!ret) {
+		gem_close(fd, handle);
+		igt_assert(gem_has_lmem(fd)); /* Should be dgpu only */
+	}
+
+	return ret == 0;
+}
+
+static uint32_t batch_create_size(int fd, uint64_t size)
+{
+	const uint32_t bbe = MI_BATCH_BUFFER_END;
+	uint32_t handle;
+
+	handle = gem_create(fd, size);
+	gem_write(fd, handle, 0, &bbe, sizeof(bbe));
+
+	return handle;
+}
+
+static int upload(int fd, uint32_t handle)
+{
+	struct drm_i915_gem_exec_object2 exec[2] = {};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&exec),
+		.buffer_count = 2,
+	};
+
+	/*
+	 * To be reasonably sure that we are not being swindled, let's make sure
+	 * to 'touch' the pages from the GPU first to ensure the object is for
+	 * sure placed in one of requested regions.
+	 */
+	exec[0].handle = handle;
+	exec[1].handle = batch_create_size(fd, PAGE_SIZE);
+
+	return __gem_execbuf(fd, &execbuf);
+}
+
+static int alloc_lmem(int fd, uint32_t *handle,
+		      struct drm_i915_gem_memory_class_instance ci,
+		      uint64_t size, bool cpu_access, bool do_upload)
+{
+	struct drm_i915_gem_memory_class_instance regions[] = {
+		ci, { I915_MEMORY_CLASS_SYSTEM, },
+	};
+	struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+		.base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+		.regions = to_user_pointer(&regions),
+	};
+	uint32_t flags;
+
+	igt_assert_eq(ci.memory_class, I915_MEMORY_CLASS_DEVICE);
+
+	flags = 0;
+	setparam_region.num_regions = 1;
+	if (cpu_access) {
+		flags = I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+		setparam_region.num_regions = 2;
+	}
+
+	*handle = gem_create_ext(fd, size, flags, &setparam_region.base);
+
+	if (do_upload)
+		return upload(fd, *handle);
+
+	return 0;
+}
+
+static void create_ext_cpu_access_sanity_check(int fd)
+{
+	struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+		.base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+	};
+	struct drm_i915_query_memory_regions *regions;
+	uint64_t size = PAGE_SIZE;
+	uint32_t handle;
+	int i;
+
+	/*
+	 * The ABI is that FLAG_NEEDS_CPU_ACCESS can only be applied to LMEM +
+	 * SMEM objects. Make sure the kernel follows that. Let's check if we
+	 * can indeed fault the object.
+	 */
+
+	/* Implicit placement; should fail */
+	igt_assert_eq(__gem_create_ext(fd, &size,
+				       I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+				       &handle, NULL), -EINVAL);
+
+	regions = gem_get_query_memory_regions(fd);
+	igt_assert(regions);
+	igt_assert(regions->num_regions);
+
+	for (i = 0; i < regions->num_regions; i++) {
+		struct drm_i915_gem_memory_class_instance ci_regions[2] = {
+			regions->regions[i].region,
+			{ I915_MEMORY_CLASS_SYSTEM, },
+		};
+		uint32_t *ptr;
+
+		setparam_region.regions = to_user_pointer(ci_regions);
+		setparam_region.num_regions = 1;
+
+		/* Single explicit placement; should fail */
+		igt_assert_eq(__gem_create_ext(fd, &size,
+					       I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+					       &handle, &setparam_region.base),
+			      -EINVAL);
+
+		if (ci_regions[0].memory_class == I915_MEMORY_CLASS_SYSTEM)
+			continue;
+
+		/*
+		 * Now combine with system memory; should pass. We should also
+		 * be able to fault it.
+		 */
+		setparam_region.num_regions = 2;
+		igt_assert_eq(__gem_create_ext(fd, &size,
+					       I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+					       &handle, &setparam_region.base),
+			      0);
+		upload(fd, handle);
+		ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+					     PROT_READ | PROT_WRITE);
+		ptr[0] = 0xdeadbeaf;
+		gem_close(fd, handle);
+
+		/*
+		 * It should also work just fine without the flag, where in the
+		 * worst case we need to migrate it when faulting it.
+		 */
+		igt_assert_eq(__gem_create_ext(fd, &size,
+					       0,
+					       &handle, &setparam_region.base),
+			      0);
+		upload(fd, handle);
+		ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+					     PROT_READ | PROT_WRITE);
+		ptr[0] = 0xdeadbeaf;
+		gem_close(fd, handle);
+	}
+
+	free(regions);
+}
+
+static uint64_t get_visible_size(int device, uint16_t instance)
+{
+	FILE *file;
+	char *line = NULL;
+	size_t line_size;
+	int fd, i;
+
+	fd = igt_debugfs_open(device, "i915_gem_objects", O_RDONLY);
+	file = fdopen(fd, "r");
+	igt_require(getline(&line, &line_size, file) > 0);
+
+	i = 0;
+	while (getline(&line, &line_size, file) > 0) {
+		const char needle[] = "visible_size: ";
+		const char *ptr = strstr(line, needle);
+		long int size_mb;
+		char *end;
+
+		if (!ptr)
+			continue;
+
+		if (i++ != instance)
+			continue;
+
+		size_mb = strtol(ptr + ARRAY_SIZE(needle)-1, &end, 10);
+
+		free(line);
+		fclose(file);
+		close(fd);
+
+		return size_mb << 20ULL;
+	}
+
+	igt_assert(!"reached");
+}
+
+static jmp_buf jmp;
+
+__noreturn static void sigtrap(int sig)
+{
+	siglongjmp(jmp, sig);
+}
+
+static void trap_sigbus(uint32_t *ptr)
+{
+	sighandler_t old_sigbus;
+
+	old_sigbus = signal(SIGBUS, sigtrap);
+	switch (sigsetjmp(jmp, SIGBUS)) {
+	case SIGBUS:
+		break;
+	case 0:
+		*ptr = 0xdeadbeaf;
+	default:
+		igt_assert(!"reached");
+		break;
+	}
+	signal(SIGBUS, old_sigbus);
+}
+
+static void create_ext_cpu_access_big(int fd)
+{
+	struct drm_i915_query_memory_regions *regions;
+	int i;
+
+	/*
+	 * Sanity check that we can still CPU map an overly large object, even
+	 * if it happens to be larger the CPU visible portion of LMEM. Also
+	 * check that an overly large allocation, which can't be spilled into
+	 * system memory does indeed fail.
+	 */
+
+	regions = gem_get_query_memory_regions(fd);
+	igt_assert(regions);
+	igt_assert(regions->num_regions);
+
+	for (i = 0; i < regions->num_regions; i++) {
+		struct drm_i915_memory_region_info qmr = regions->regions[i];
+		struct drm_i915_gem_memory_class_instance ci = qmr.region;
+		uint64_t size, visible_size, lmem_size;
+		uint32_t handle;
+		uint32_t *ptr;
+
+		if (ci.memory_class == I915_MEMORY_CLASS_SYSTEM)
+			continue;
+
+		lmem_size = qmr.probed_size;
+		visible_size = get_visible_size(fd, ci.memory_instance);
+		if (!visible_size)
+			continue;
+
+		if (visible_size <= (0.70 * lmem_size)) {
+			/*
+			 * Too big. We should still be able to allocate it just
+			 * fine, but faulting should result in tears.
+			 */
+			size = visible_size;
+			igt_assert_eq(alloc_lmem(fd, &handle, ci, size, false, true), 0);
+			ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+						     PROT_READ | PROT_WRITE);
+			trap_sigbus(ptr);
+			gem_close(fd, handle);
+
+			/*
+			 * Too big again, but this time we can spill to system
+			 * memory when faulting the object.
+			 */
+			size = visible_size;
+			igt_assert_eq(alloc_lmem(fd, &handle, ci, size, true, true), 0);
+			ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+						     PROT_READ | PROT_WRITE);
+			ptr[0] = 0xdeadbeaf;
+			gem_close(fd, handle);
+
+			/*
+			 * Let's also move the upload to after faulting the
+			 * pages. The current behaviour is that the pages are
+			 * only allocated in device memory when initially
+			 * touched by the GPU. With this in mind we should also
+			 * make sure that the pages are indeed migrated, as
+			 * expected.
+			 */
+			size = visible_size;
+			igt_assert_eq(alloc_lmem(fd, &handle, ci, size, false, false), 0);
+			ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+						     PROT_READ | PROT_WRITE);
+			ptr[0] = 0xdeadbeaf; /* temp system memory */
+			igt_assert_eq(upload(fd, handle), 0);
+			trap_sigbus(ptr); /* non-mappable device memory */
+			gem_close(fd, handle);
+		}
+
+		/*
+		 * Should fit. We likely need to migrate to the mappable portion
+		 * on fault though, if this device has a small BAR, given how
+		 * large the initial allocation is.
+		 */
+		size = visible_size >> 1;
+		igt_assert_eq(alloc_lmem(fd, &handle, ci, size, false, true), 0);
+		ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+					     PROT_READ | PROT_WRITE);
+		ptr[0] = 0xdeadbeaf;
+		gem_close(fd, handle);
+
+		/*
+		 * And then with the CPU_ACCESS flag enabled; should also be no
+		 * surprises here.
+		 */
+		igt_assert_eq(alloc_lmem(fd, &handle, ci, size, true, true), 0);
+		ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+					     PROT_READ | PROT_WRITE);
+		ptr[0] = 0xdeadbeaf;
+		gem_close(fd, handle);
+	}
+
+	free(regions);
+}
+
 igt_main
 {
 	int fd = -1;
@@ -516,4 +839,15 @@ igt_main
 	igt_subtest("create-ext-placement-all")
 		create_ext_placement_all(fd);
 
+	igt_describe("Verify the basic functionally and expected ABI contract around I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS");
+	igt_subtest("create-ext-cpu-access-sanity-check") {
+		igt_require(supports_needs_cpu_access(fd));
+		create_ext_cpu_access_sanity_check(fd);
+	}
+
+	igt_describe("Verify the extreme cases with very large objects and I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS");
+	igt_subtest("create-ext-cpu-access-big") {
+		igt_require(supports_needs_cpu_access(fd));
+		create_ext_cpu_access_big(fd);
+	}
 }
-- 
2.34.1

[Intel-gfx] [PATCH i-g-t v2 4/6] lib/i915: add gem_create_with_cpu_access_in_memory_regions

[Matthew Auld]

Most users shouldn't care about such an interface, but where required,
this should be useful to aid in setting NEEDS_CPU_ACCESS for a given BO.
Underneath we try to smooth over needing to provide an explicit SMEM
region, or if this is SMEM-only, we don't want the kernel to throw an
error.

Put it to use in gem_exec_capture, where a proper hint is now required
by the kernel on DG2+, for objects marked with EXEC_OBJECT_CAPTURE, that
can also be placed in LMEM.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 lib/i915/intel_memory_region.c | 10 +++++---
 lib/i915/intel_memory_region.h | 46 +++++++++++++++++++++++++++++++---
 tests/i915/gem_exec_capture.c  |  6 ++---
 tests/i915/gem_lmem_swapping.c |  2 +-
 4 files changed, 52 insertions(+), 12 deletions(-)

diff --git a/lib/i915/intel_memory_region.c b/lib/i915/intel_memory_region.c
index f0c8bc7c..4893c5ba 100644
--- a/lib/i915/intel_memory_region.c
+++ b/lib/i915/intel_memory_region.c
@@ -197,7 +197,7 @@ bool gem_has_lmem(int fd)
 
 /* A version of gem_create_in_memory_region_list which can be allowed to
    fail so that the object creation can be retried */
-int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size,
+int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size, uint32_t flags,
 				       struct drm_i915_gem_memory_class_instance *mem_regions,
 				       int num_regions)
 {
@@ -208,7 +208,9 @@ int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size,
 	};
 	int ret;
 
-	ret = __gem_create_ext(fd, size, 0, handle, &ext_regions.base);
+	ret = __gem_create_ext(fd, size, flags, handle, &ext_regions.base);
+	if (flags && ret == -EINVAL)
+		ret = __gem_create_ext(fd, size, 0, handle, &ext_regions.base);
 
 	/*
 	 * Provide fallback for stable kernels if region passed in the array
@@ -231,12 +233,12 @@ int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size,
  * @mem_regions: memory regions array (priority list)
  * @num_regions: @mem_regions length
  */
-uint32_t gem_create_in_memory_region_list(int fd, uint64_t size,
+uint32_t gem_create_in_memory_region_list(int fd, uint64_t size, uint32_t flags,
 					  struct drm_i915_gem_memory_class_instance *mem_regions,
 					  int num_regions)
 {
 	uint32_t handle;
-	int ret = __gem_create_in_memory_region_list(fd, &handle, &size,
+	int ret = __gem_create_in_memory_region_list(fd, &handle, &size, flags,
 						     mem_regions, num_regions);
 	igt_assert_eq(ret, 0);
 	return handle;
diff --git a/lib/i915/intel_memory_region.h b/lib/i915/intel_memory_region.h
index 936e7d1c..7cc119ec 100644
--- a/lib/i915/intel_memory_region.h
+++ b/lib/i915/intel_memory_region.h
@@ -21,6 +21,7 @@
  * IN THE SOFTWARE.
  */
 #include "igt_collection.h"
+#include "i915_drm_local.h"
 
 #ifndef INTEL_MEMORY_REGION_H
 #define INTEL_MEMORY_REGION_H
@@ -64,11 +65,11 @@ bool gem_has_lmem(int fd);
 
 struct drm_i915_gem_memory_class_instance;
 
-int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size,
+int __gem_create_in_memory_region_list(int fd, uint32_t *handle, uint64_t *size, uint32_t flags,
 				       struct drm_i915_gem_memory_class_instance *mem_regions,
 				       int num_regions);
 
-uint32_t gem_create_in_memory_region_list(int fd, uint64_t size,
+uint32_t gem_create_in_memory_region_list(int fd, uint64_t size, uint32_t flags,
 					  struct drm_i915_gem_memory_class_instance *mem_regions,
 					  int num_regions);
 
@@ -84,7 +85,7 @@ uint32_t gem_create_in_memory_region_list(int fd, uint64_t size,
 		arr_query__[i__].memory_class = MEMORY_TYPE_FROM_REGION(arr__[i__]);  \
 		arr_query__[i__].memory_instance = MEMORY_INSTANCE_FROM_REGION(arr__[i__]);  \
 	} \
-	__gem_create_in_memory_region_list(fd, handle, size, arr_query__, ARRAY_SIZE(arr_query__)); \
+	__gem_create_in_memory_region_list(fd, handle, size, 0, arr_query__, ARRAY_SIZE(arr_query__)); \
 })
 #define gem_create_in_memory_regions(fd, size, regions...) ({ \
 	unsigned int arr__[] = { regions }; \
@@ -93,7 +94,44 @@ uint32_t gem_create_in_memory_region_list(int fd, uint64_t size,
 		arr_query__[i__].memory_class = MEMORY_TYPE_FROM_REGION(arr__[i__]);  \
 		arr_query__[i__].memory_instance = MEMORY_INSTANCE_FROM_REGION(arr__[i__]);  \
 	} \
-	gem_create_in_memory_region_list(fd, size, arr_query__, ARRAY_SIZE(arr_query__)); \
+	gem_create_in_memory_region_list(fd, size, 0, arr_query__, ARRAY_SIZE(arr_query__)); \
+})
+
+/*
+ * Create an object that requires CPU access. This only becomes interesting on
+ * platforms that have a small BAR for LMEM CPU access. Without this the object
+ * might need to be migrated when CPU faulting the object, or if that is not
+ * possible we hit SIGBUS. Most users should be fine with this. If enabled the
+ * kernel will never allocate this object in the non-CPU visible portion of
+ * LMEM.
+ *
+ * Underneath this just enables the I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS
+ * flag, if we also have an LMEM placement. Also since the kernel requires SMEM
+ * as a potential placement, we automatically attach that as a possible
+ * placement, if not already provided. If this happens to be an SMEM-only
+ * placement then we don't supply the flag, and instead just treat as normal
+ * allocation.
+ */
+#define gem_create_with_cpu_access_in_memory_regions(fd, size, regions...) ({ \
+	unsigned int arr__[] = { regions }; \
+	struct drm_i915_gem_memory_class_instance arr_query__[ARRAY_SIZE(arr__) + 1]; \
+	int i__, arr_query_size__ = ARRAY_SIZE(arr__); \
+	uint32_t ext_flags__ = 0; \
+	bool ext_found_smem__ = false; \
+	for (i__  = 0; i__ < arr_query_size__; ++i__) { \
+		arr_query__[i__].memory_class = MEMORY_TYPE_FROM_REGION(arr__[i__]);  \
+		if (arr_query__[i__].memory_class == I915_MEMORY_CLASS_DEVICE) \
+			ext_flags__ = I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS; \
+		else \
+			ext_found_smem__ = true; \
+		arr_query__[i__].memory_instance = MEMORY_INSTANCE_FROM_REGION(arr__[i__]);  \
+	} \
+	if (ext_flags__ && !ext_found_smem__) { \
+		arr_query__[i__].memory_class = I915_MEMORY_CLASS_SYSTEM; \
+		arr_query__[i__].memory_instance = 0; \
+		arr_query_size__++; \
+	} \
+	gem_create_in_memory_region_list(fd, size, ext_flags__, arr_query__, arr_query_size__); \
 })
 
 struct igt_collection *
diff --git a/tests/i915/gem_exec_capture.c b/tests/i915/gem_exec_capture.c
index 60f8df04..24ba6036 100644
--- a/tests/i915/gem_exec_capture.c
+++ b/tests/i915/gem_exec_capture.c
@@ -268,7 +268,7 @@ static void __capture1(int fd, int dir, uint64_t ahnd, const intel_ctx_t *ctx,
 	saved_engine = configure_hangs(fd, e, ctx->id);
 
 	memset(obj, 0, sizeof(obj));
-	obj[SCRATCH].handle = gem_create_in_memory_regions(fd, 4096, region);
+	obj[SCRATCH].handle = gem_create_with_cpu_access_in_memory_regions(fd, 4096, region);
 	obj[SCRATCH].flags = EXEC_OBJECT_WRITE;
 	obj[CAPTURE].handle = target;
 	obj[CAPTURE].flags = EXEC_OBJECT_CAPTURE;
@@ -387,9 +387,9 @@ static void capture(int fd, int dir, const intel_ctx_t *ctx,
 		    const struct intel_execution_engine2 *e, uint32_t region)
 {
 	uint32_t handle;
-	uint64_t ahnd, obj_size = 4096;
+	uint64_t ahnd, obj_size = 16 * 4096;
 
-	igt_assert_eq(__gem_create_in_memory_regions(fd, &handle, &obj_size, region), 0);
+	handle = gem_create_with_cpu_access_in_memory_regions(fd, obj_size, region);
 	ahnd = get_reloc_ahnd(fd, ctx->id);
 
 	__capture1(fd, dir, ahnd, ctx, e, handle, obj_size, region);
diff --git a/tests/i915/gem_lmem_swapping.c b/tests/i915/gem_lmem_swapping.c
index 39f9e1f5..18b66f09 100644
--- a/tests/i915/gem_lmem_swapping.c
+++ b/tests/i915/gem_lmem_swapping.c
@@ -80,7 +80,7 @@ static uint32_t create_bo(int i915,
 	int ret;
 
 retry:
-	ret = __gem_create_in_memory_region_list(i915, &handle, &size, region, 1);
+	ret = __gem_create_in_memory_region_list(i915, &handle, &size, 0, region, 1);
 	if (do_oom_test && ret == -ENOMEM)
 		goto retry;
 	igt_assert_eq(ret, 0);
-- 
2.34.1

[Intel-gfx] [PATCH i-g-t v2 5/6] i915/tests/capture: add a negative test for NEEDS_CPU_ACCESS

[Matthew Auld]

Sanity check that the kernel does indeed reject LMEM buffers marked with
EXEC_OBJECT_CAPTURE, that are not also marked with NEEDS_CPU_ACCESS.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 tests/i915/gem_exec_capture.c | 69 +++++++++++++++++++++++++++++++++++
 1 file changed, 69 insertions(+)

diff --git a/tests/i915/gem_exec_capture.c b/tests/i915/gem_exec_capture.c
index 24ba6036..09187f62 100644
--- a/tests/i915/gem_exec_capture.c
+++ b/tests/i915/gem_exec_capture.c
@@ -735,6 +735,71 @@ static void userptr(int fd, int dir)
 	gem_engine_properties_restore(fd, &saved_engine);
 }
 
+static bool supports_needs_cpu_access(int fd)
+{
+	struct drm_i915_gem_memory_class_instance regions[] = {
+		{ I915_MEMORY_CLASS_DEVICE, },
+		{ I915_MEMORY_CLASS_SYSTEM, },
+	};
+	struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+		.base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+		.regions = to_user_pointer(&regions),
+		.num_regions = ARRAY_SIZE(regions),
+	};
+	uint64_t size = 4096;
+	uint32_t handle;
+	int ret;
+
+	ret = __gem_create_ext(fd, &size,
+			       I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+			       &handle, &setparam_region.base);
+	if (!ret) {
+		gem_close(fd, handle);
+		igt_assert(gem_has_lmem(fd)); /* Should be dgpu only */
+	}
+
+	return ret == 0;
+}
+
+static void capture_no_cpu_access(int fd)
+{
+	struct drm_i915_gem_exec_object2 exec = {
+		.flags = EXEC_OBJECT_CAPTURE,
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&exec),
+		.buffer_count = 1,
+	};
+	uint64_t size = 4096;
+	uint32_t handle;
+	int ret;
+
+	igt_require(gem_has_lmem(fd));
+	igt_require(supports_needs_cpu_access(fd));
+
+	/*
+	 * Sanity check that execbuf rejects EXEC_OBJECT_CAPTURE marked BO, that
+	 * is not also tagged with I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS, if
+	 * it can be placed in LMEM. This is only relevant for Dg2+.
+	 */
+
+	igt_require(__gem_create_in_memory_regions(fd, &handle, &size,
+						   REGION_LMEM(0)) == 0);
+
+	exec.handle = handle;
+	ret = __gem_execbuf(fd, &execbuf);
+	if (IS_DG1(fd)) /* Should be no impact on existing ABI */
+		igt_assert(ret == 0);
+	else
+		igt_assert(ret == -EINVAL);
+
+	/* SMEM only buffers should work as normal */
+	igt_assert(__gem_create_in_memory_regions(fd, &handle, &size,
+						  REGION_SMEM) == 0);
+	exec.handle = handle;
+	igt_assert(__gem_execbuf(fd, &execbuf) == 0);
+}
+
 static bool has_capture(int fd)
 {
 	drm_i915_getparam_t gp;
@@ -839,6 +904,10 @@ igt_main
 		igt_dynamic_f("%s", (e)->name)
 			prioinv(fd, dir, ctx, e);
 
+	igt_describe("Verify the ABI contract when using EXEC_OBJECT_CAPTURE without I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS");
+	igt_subtest_f("capture-non-cpu-access")
+		capture_no_cpu_access(fd);
+
 	igt_fixture {
 		close(dir);
 		igt_disallow_hang(fd, hang);
-- 
2.34.1

[Intel-gfx] [PATCH i-g-t v2 6/6] lib/i915: request CPU_ACCESS for fb objects

[Matthew Auld]

kms_frontbuffer_tracking@basic falls over if the fb needs to be migrated
from non-mappable device memory, to the mappable part, due to being
temporarily pinned for scanout, when hitting the CPU fault handler,
which just gives us SIGBUS. If the device has a small BAR let's attempt
to use the mappable portion, if possible.

XXX: perhaps the kernel needs to somehow handle this better?

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 lib/ioctl_wrappers.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/lib/ioctl_wrappers.c b/lib/ioctl_wrappers.c
index 09eb3ce7..7713e78b 100644
--- a/lib/ioctl_wrappers.c
+++ b/lib/ioctl_wrappers.c
@@ -635,7 +635,8 @@ uint32_t gem_buffer_create_fb_obj(int fd, uint64_t size)
 	uint32_t handle;
 
 	if (gem_has_lmem(fd))
-		handle = gem_create_in_memory_regions(fd, size, REGION_LMEM(0));
+		handle = gem_create_with_cpu_access_in_memory_regions(fd, size,
+								      REGION_LMEM(0));
 	else
 		handle = gem_create(fd, size);
 
-- 
2.34.1