Skip to contentSkip to conversation

tests/gem_wait_render_timeout: make sure the GPU is idle before exiting

PATCH In review Imre Deak 2 participants 6 messages May 27, 2013 – May 28, 2013 Library code

lore
Share & tools
lore

Imre Deak

author
#0 lore raw

PATCH 1/1 tests/gem_wait_render_timeout: make sure the GPU is idle before exiting

Leaving the GPU running after we exit can mess up timing dependent tests we run afterwards.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64270

---
 tests/gem_wait_render_timeout.c | 5 +++++
 1 file changed, 5 insertions(+)
tests/gem_wait_render_timeout.c +5 -0
index 14482e3..9ec4e49 100644
@@ -216,6 +216,11 @@ int main(int argc, char **argv)
 	assert(gem_bo_wait_timeout(fd, dst2->handle, &timeout) == -ETIME);
 	assert(timeout == 0);
 
+	/* Make sure we exit only after the GPU is idle. Apply some fuzz
+	 * to the timeout since the calibrated iteration count can be
+	 * twice the amount the GPU can execute in a second. */
+	timeout = 5 * ENOUGH_WORK_IN_SECONDS * NSEC_PER_SEC;
+	assert(gem_bo_wait_timeout(fd, dst2->handle, &timeout) == 0);
 
 	if (do_signals)
 		drmtest_stop_signal_helper();
Signature · 1 line
1.8.1.2
#1 lore raw
Quoted 6 lines · Imre Deak

Leaving the GPU running after we exit can mess up timing dependent tests we run afterwards.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64270

I've flailed around in this area since forever, see gem_quiescent_gpu in drmtest.c. I wonder whether we shouldn't just install an atexit handler with this and maybe for full paranoia run it in drm_open_any, too?

One ugly bug with that function is that it doesn't quiescent all rings on gen6+ though. So I guess that needs fixing.
-Daniel

Quoted 27 lines
---
 tests/gem_wait_render_timeout.c | 5 +++++
 1 file changed, 5 insertions(+)
tests/gem_wait_render_timeout.c +5 -0
index 14482e3..9ec4e49 100644
@@ -216,6 +216,11 @@ int main(int argc, char **argv)
 	assert(gem_bo_wait_timeout(fd, dst2->handle, &timeout) == -ETIME);
 	assert(timeout == 0);
 
+	/* Make sure we exit only after the GPU is idle. Apply some fuzz
+	 * to the timeout since the calibrated iteration count can be
+	 * twice the amount the GPU can execute in a second. */
+	timeout = 5 * ENOUGH_WORK_IN_SECONDS * NSEC_PER_SEC;
+	assert(gem_bo_wait_timeout(fd, dst2->handle, &timeout) == 0);
 
 	if (do_signals)
 		drmtest_stop_signal_helper();
Signature · 6 lines
1.8.1.2

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
author
#2 lore raw

PATCH v2 1/2 lib: make sure all rings are idle in gpu_quiescent_gpu()

In reply to Imre Deak
---
 lib/drmtest.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
lib/drmtest.c +16 -0
index 16f5be1..2d37fb6 100644
@@ -109,6 +109,7 @@ int gem_available_fences(int fd)
 }
 
 
+#define LOCAL_I915_EXEC_VEBOX	(4 << 0)
 /* Ensure the gpu is idle by launching a nop execbuf and stalling for it. */
 void gem_quiescent_gpu(int fd)
 {
@@ -143,6 +144,21 @@ void gem_quiescent_gpu(int fd)
 
 	do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
 
+	if (gem_has_blt(fd)) {
+		execbuf.flags = I915_EXEC_BLT;
+		do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
+	}
+
+	if (gem_has_bsd(fd)) {
+		execbuf.flags = I915_EXEC_BSD;
+		do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
+	}
+
+	if (gem_has_vebox(fd)) {
+		execbuf.flags = LOCAL_I915_EXEC_VEBOX;
+		do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
+	}
+
 	gem_sync(fd, handle);
 }
 
Signature · 1 line
1.8.1.2
author
#3 lore raw

PATCH v2 2/2 tests/lib: make sure the GPU is idle at test start and exit

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64270

v2:
- Make sure also that the GPU is idle at start and error exit of any test using drm_open_any(). (Daniel)

---
 lib/drmtest.c | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)
lib/drmtest.c +24 -1
index 2d37fb6..198104d 100644
@@ -228,7 +228,7 @@ int drm_get_card(int master)
 }
 
 /** Open the first DRM device we can find, searching up to 16 device nodes */
-int drm_open_any(void)
+static int __drm_open_any(void)
 {
 	char *name;
 	int ret, fd;
@@ -248,6 +248,29 @@ int drm_open_any(void)
 	return fd;
 }
 
+static void quiescent_gpu_at_exit(int sig)
+{
+	int fd;
+
+	fd = __drm_open_any();
+	if (fd >= 0)
+		close(fd);
+}
+
+int drm_open_any(void)
+{
+	static int open_count;
+	int fd = __drm_open_any();
+
+	if (fd < 0 || __sync_fetch_and_add(&open_count, 1))
+		return fd;
+
+	gem_quiescent_gpu(fd);
+	drmtest_install_exit_handler(quiescent_gpu_at_exit);
+
+	return fd;
+}
+
 /**
  * Open the first DRM device we can find where we end up being the master.
  */
Signature · 1 line
1.8.1.2
author
#4 lore raw

PATCH v3 2/2 tests/lib: make sure the GPU is idle at test start and exit

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64270

v2:
- Make sure also that the GPU is idle at start and error exit of any test using drm_open_any(). (Daniel)
v3:
- actually call gem_quiescent_gpu() at exit

---
 lib/drmtest.c | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)
lib/drmtest.c +26 -1
index 2d37fb6..d17dbb0 100644
@@ -228,7 +228,7 @@ int drm_get_card(int master)
 }
 
 /** Open the first DRM device we can find, searching up to 16 device nodes */
-int drm_open_any(void)
+static int __drm_open_any(void)
 {
 	char *name;
 	int ret, fd;
@@ -248,6 +248,31 @@ int drm_open_any(void)
 	return fd;
 }
 
+static void quiescent_gpu_at_exit(int sig)
+{
+	int fd;
+
+	fd = __drm_open_any();
+	if (fd >= 0) {
+		gem_quiescent_gpu(fd);
+		close(fd);
+	}
+}
+
+int drm_open_any(void)
+{
+	static int open_count;
+	int fd = __drm_open_any();
+
+	if (fd < 0 || __sync_fetch_and_add(&open_count, 1))
+		return fd;
+
+	gem_quiescent_gpu(fd);
+	drmtest_install_exit_handler(quiescent_gpu_at_exit);
+
+	return fd;
+}
+
 /**
  * Open the first DRM device we can find where we end up being the master.
  */
Signature · 1 line
1.8.1.2
#5 lore raw

Re: [i-g-t PATCH v3 2/2] tests/lib: make sure the GPU is idle at test start and exit

Quoted 9 lines · Imre Deak

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64270

v2:
- Make sure also that the GPU is idle at start and error exit of any test using drm_open_any(). (Daniel)
v3:
- actually call gem_quiescent_gpu() at exit

Both merged to igt, thanks.
-Daniel

Quoted 56 lines
---
 lib/drmtest.c | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)
lib/drmtest.c +26 -1
index 2d37fb6..d17dbb0 100644
@@ -228,7 +228,7 @@ int drm_get_card(int master)
 }
 
 /** Open the first DRM device we can find, searching up to 16 device nodes */
-int drm_open_any(void)
+static int __drm_open_any(void)
 {
 	char *name;
 	int ret, fd;
@@ -248,6 +248,31 @@ int drm_open_any(void)
 	return fd;
 }
 
+static void quiescent_gpu_at_exit(int sig)
+{
+	int fd;
+
+	fd = __drm_open_any();
+	if (fd >= 0) {
+		gem_quiescent_gpu(fd);
+		close(fd);
+	}
+}
+
+int drm_open_any(void)
+{
+	static int open_count;
+	int fd = __drm_open_any();
+
+	if (fd < 0 || __sync_fetch_and_add(&open_count, 1))
+		return fd;
+
+	gem_quiescent_gpu(fd);
+	drmtest_install_exit_handler(quiescent_gpu_at_exit);
+
+	return fd;
+}
+
 /**
  * Open the first DRM device we can find where we end up being the master.
  */
Signature · 6 lines
1.8.1.2

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch