[PATCH 4/4] tests/gem_gtt_hog: Fix for BDW

[Rodrigo Vivi <rodrigo.vivi@gmail.com>]

Update XY_COLOR_BLT command for Broadwell.

v2: stash devid and remove ugly double allocation. (by Chris).
v3: fix inverted blt command size and stash fd, devid and intel_gen.
v4: improved len calculation and noop between blt commands. (by Chris).

Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=73724

Cc: Chris Wilson chris@chris-wilson.co.uk>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
---
 tests/gem_gtt_hog.c | 52 +++++++++++++++++++++++++++++++++-------------------
 1 file changed, 33 insertions(+), 19 deletions(-)

diff --git a/tests/gem_gtt_hog.c b/tests/gem_gtt_hog.c
index d22e074..c7206dc 100644
--- a/tests/gem_gtt_hog.c
+++ b/tests/gem_gtt_hog.c
@@ -44,20 +44,26 @@
 
 static const uint32_t canary = 0xdeadbeef;
 
+typedef struct data {
+	int fd;
+	int devid;
+	int intel_gen;
+} data_t;
+
 static double elapsed(const struct timeval *start,
 		      const struct timeval *end)
 {
 	return 1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec);
 }
 
-static void busy(int fd, uint32_t handle, int size, int loops)
+static void busy(data_t *data, uint32_t handle, int size, int loops)
 {
 	struct drm_i915_gem_relocation_entry reloc[20];
 	struct drm_i915_gem_exec_object2 gem_exec[2];
 	struct drm_i915_gem_execbuffer2 execbuf;
 	struct drm_i915_gem_pwrite gem_pwrite;
 	struct drm_i915_gem_create create;
-	uint32_t buf[122], *b;
+	uint32_t buf[170], *b;
 	int i;
 
 	memset(reloc, 0, sizeof(reloc));
@@ -66,7 +72,8 @@ static void busy(int fd, uint32_t handle, int size, int loops)
 
 	b = buf;
 	for (i = 0; i < 20; i++) {
-		*b++ = XY_COLOR_BLT_CMD_NOLEN | 4 |
+		*b++ = XY_COLOR_BLT_CMD_NOLEN |
+			((data->intel_gen >= 8) ? 5 : 4) |
 			COLOR_BLT_WRITE_ALPHA | XY_COLOR_BLT_WRITE_RGB;
 		*b++ = 0xf0 << 16 | 1 << 25 | 1 << 24 | 4096;
 		*b++ = 0;
@@ -76,66 +83,68 @@ static void busy(int fd, uint32_t handle, int size, int loops)
 		reloc[i].read_domains = I915_GEM_DOMAIN_RENDER;
 		reloc[i].write_domain = I915_GEM_DOMAIN_RENDER;
 		*b++ = 0;
+		if (data->intel_gen >= 8)
+			*b++ = 0;
 		*b++ = canary;
 	}
 	*b++ = MI_BATCH_BUFFER_END;
-	*b++ = 0;
+	if ((b - buf) & 1)
+		*b++ = 0;
 
 	gem_exec[0].handle = handle;
 	gem_exec[0].flags = EXEC_OBJECT_NEEDS_FENCE;
 
 	create.handle = 0;
 	create.size = 4096;
-	drmIoctl(fd, DRM_IOCTL_I915_GEM_CREATE, &create);
+	drmIoctl(data->fd, DRM_IOCTL_I915_GEM_CREATE, &create);
 	gem_exec[1].handle = create.handle;
 	gem_exec[1].relocation_count = 20;
 	gem_exec[1].relocs_ptr = (uintptr_t)reloc;
 
 	execbuf.buffers_ptr = (uintptr_t)gem_exec;
 	execbuf.buffer_count = 2;
-	execbuf.batch_len = sizeof(buf);
+	execbuf.batch_len = (b - buf) * sizeof(buf[0]);
 	execbuf.flags = 1 << 11;
-	if (HAS_BLT_RING(intel_get_drm_devid(fd)))
+	if (HAS_BLT_RING(data->devid))
 		execbuf.flags |= I915_EXEC_BLT;
 
 	gem_pwrite.handle = gem_exec[1].handle;
 	gem_pwrite.offset = 0;
-	gem_pwrite.size = sizeof(buf);
+	gem_pwrite.size = execbuf.batch_len;
 	gem_pwrite.data_ptr = (uintptr_t)buf;
-	if (drmIoctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &gem_pwrite) == 0) {
+	if (drmIoctl(data->fd, DRM_IOCTL_I915_GEM_PWRITE, &gem_pwrite) == 0) {
 		while (loops--)
-			drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
+			drmIoctl(data->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
 	}
 
-	drmIoctl(fd, DRM_IOCTL_GEM_CLOSE, &create.handle);
+	drmIoctl(data->fd, DRM_IOCTL_GEM_CLOSE, &create.handle);
 }
 
-static void run(int child)
+static void run(data_t *data, int child)
 {
 	const int size = 4096 * (256 + child * child);
 	const int tiling = child % 2;
 	const int write = child % 2;
-	int fd = drm_open_any();
-	uint32_t handle = gem_create(fd, size);
+	uint32_t handle = gem_create(data->fd, size);
 	uint32_t *ptr;
 	uint32_t x;
 
 	igt_assert(handle);
 
 	if (tiling != I915_TILING_NONE)
-		gem_set_tiling(fd, handle, tiling, 4096);
+		gem_set_tiling(data->fd, handle, tiling, 4096);
 
 	/* load up the unfaulted bo */
-	busy(fd, handle, size, 100);
+	busy(data, handle, size, 100);
 
 	/* Note that we ignore the API and rely on the implict
 	 * set-to-gtt-domain within the fault handler.
 	 */
 	if (write) {
-		ptr = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
+		ptr = gem_mmap(data->fd, handle, size, PROT_READ | PROT_WRITE);
 		ptr[rand() % (size / 4)] = canary;
 	} else
-		ptr = gem_mmap(fd, handle, size, PROT_READ);
+		ptr = gem_mmap(data->fd, handle, size, PROT_READ);
 	x = ptr[rand() % (size / 4)];
 	munmap(ptr, size);
 
@@ -147,6 +156,7 @@ igt_simple_main
 {
 	struct timeval start, end;
 	pid_t children[64];
+	data_t data = {};
 	int n;
 
 	/* check for an intel gpu before goint nuts. */
@@ -155,11 +165,15 @@ igt_simple_main
 
 	igt_skip_on_simulation();
 
+	data.fd = drm_open_any();
+	data.devid = intel_get_drm_devid(data.fd);
+	data.intel_gen = intel_gen(data.devid);
+
 	gettimeofday(&start, NULL);
 	for (n = 0; n < ARRAY_SIZE(children); n++) {
 		switch ((children[n] = fork())) {
 		case -1: igt_assert(0);
-		case 0: run(n); break;
+		case 0: run(&data, n); break;
 		default: break;
 		}
 	}
-- 
1.8.5.3