Skip to contentSkip to conversation

igt/kms: Make fence waiting explicit.

PATCH In review Maarten Lankhorst 2 participants 8 messages Jul 6, 2017 – Jul 10, 2017 Library code

lore
Share & tools
lore
#0 lore raw

cover igt/kms: Make fence waiting explicit.

I wanted to make kms_atomic_transition pass, but the nonblocking modeset fencing tests were bogus.

This series changes the semantics for fencing slightly. It only keeps the out fences in pipe_obj->out_fence_fd, it's up to the test to decide what to do with it, which is probably just waiting on the fd or checking if it completed.

Maarten Lankhorst (5):
tests/kms_atomic_transition: Add test for plane completion ordering. lib/kms: Handle fence interaction correctly WRT TEST_ONLY. tests/kms_atomic_transition: Only request fence on enabled pipes tests/kms_atomic_transition: Do not clear in-fences after atomic
commit.
igt/kms: Do not wait for fence completion during commit

 lib/igt_kms.c                 | 41 ++++++++----------
 tests/kms_atomic_transition.c | 99 +++++++++++++++++++++++++++++++++----------
 2 files changed, 95 insertions(+), 45 deletions(-)
Signature · 6 lines
2.11.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH 1/5 tests/kms_atomic_transition: Add test for plane completion ordering.

When a plane gets disabled, that commit has to complete before an atomic commit on that disabled plane only. Else for the old commit, new_plane_state may have been freed.

---
 tests/kms_atomic_transition.c | 31 +++++++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)
tests/kms_atomic_transition.c +31 -0
index ba5cd4d6e289..2b56eef66b7b 100644
@@ -30,6 +30,7 @@
 #include <stdio.h>
 #include <string.h>
 #include <time.h>
+#include <poll.h>
 
 #ifndef DRM_CAP_CURSOR_WIDTH
 #define DRM_CAP_CURSOR_WIDTH 0x8
@@ -152,6 +153,7 @@ static drmEventContext drm_events = {
 
 enum transition_type {
 	TRANSITION_PLANES,
+	TRANSITION_AFTER_FREE,
 	TRANSITION_MODESET,
 	TRANSITION_MODESET_DISABLE,
 };
@@ -480,6 +482,31 @@ run_transition_test(igt_display_t *display, enum pipe pipe, igt_output_t *output
 			igt_skip("Cannot run tests without proper size sprite planes\n");
 	}
 
+	igt_display_commit2(display, COMMIT_ATOMIC);
+
+	if (type == TRANSITION_AFTER_FREE) {
+		struct pollfd pfd = { display->drm_fd, POLLIN };
+
+		wm_setup_plane(display, pipe, 0, parms);
+
+		atomic_commit(display, pipe, flags, (void *)(unsigned long)0, fencing);
+
+		for_each_plane_on_pipe(display, pipe, plane)
+			plane->fb_changed = true;
+
+		igt_display_commit2(display, COMMIT_ATOMIC);
+
+		/*
+		 * Previous atomic commit should have completed
+		 * before this plane-only atomic commit.
+		 */
+		igt_assert_eq(poll(&pfd, 1, 0), 1);
+
+		drmHandleEvent(display->drm_fd, &drm_events);
+
+		goto cleanup;
+	}
+
 	for (i = 0; i < iter_max; i++) {
 		igt_output_set_pipe(output, pipe);
 
@@ -854,6 +881,10 @@ igt_main
 		for_each_pipe_with_valid_output(&display, pipe, output)
 			run_transition_test(&display, pipe, output, TRANSITION_PLANES, true, true);
 
+	igt_subtest("plane-use-after-nonblocking-unbind")
+		for_each_pipe_with_valid_output(&display, pipe, output)
+			run_transition_test(&display, pipe, output, TRANSITION_AFTER_FREE, true, false);
+
 	igt_subtest("plane-all-modeset-transition")
 		for_each_pipe_with_valid_output(&display, pipe, output)
 			run_transition_test(&display, pipe, output, TRANSITION_MODESET, false, false);
Signature · 6 lines
2.11.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

PATCH 2/5 lib/kms: Handle fence interaction correctly WRT TEST_ONLY.

In reply to Maarten Lankhorst

All other atomic properties are reset in display_commit_changed, which is the right place because TEST_ONLY commits don't need to be reset.

---
 lib/igt_kms.c | 38 ++++++++++++++++----------------------
 1 file changed, 16 insertions(+), 22 deletions(-)
lib/igt_kms.c +16 -22
index faf9df2fcedf..8bf56faf41e9 100644
@@ -2505,7 +2505,6 @@ static void igt_atomic_prepare_crtc_commit(igt_pipe_t *pipe_obj, drmModeAtomicRe
 	pipe_obj->out_fence_fd = -1;
 	if (pipe_obj->out_fence_requested)
 	{
-		pipe_obj->out_fence_requested = false;
 		igt_atomic_populate_crtc_req(req, pipe_obj, IGT_CRTC_OUT_FENCE_PTR,
 		    (uint64_t)(uintptr_t) &pipe_obj->out_fence_fd);
 	}
@@ -2586,27 +2585,6 @@ static int igt_atomic_commit(igt_display_t *display, uint32_t flags, void *user_
 	}
 
 	ret = drmModeAtomicCommit(display->drm_fd, req, flags, user_data);
-	if (!ret) {
-
-		for_each_pipe(display, pipe) {
-			igt_pipe_t *pipe_obj = &display->pipes[pipe];
-			igt_plane_t *plane;
-
-			/* reset fence_fd to prevent it from being set for the next commit */
-			for_each_plane_on_pipe(display, pipe, plane) {
-				igt_plane_set_fence_fd(plane, -1);
-			}
-
-			if (pipe_obj->out_fence_fd == -1)
-				continue;
-
-			igt_assert(pipe_obj->out_fence_fd >= 0);
-			ret = sync_fence_wait(pipe_obj->out_fence_fd, 1000);
-			igt_assert(ret == 0);
-			close(pipe_obj->out_fence_fd);
-			pipe_obj->out_fence_fd = -1;
-		}
-	}
 
 	drmModeAtomicFree(req);
 	return ret;
@@ -2629,6 +2607,18 @@ display_commit_changed(igt_display_t *display, enum igt_commit_style s)
 		if (s != COMMIT_UNIVERSAL)
 			pipe_obj->mode_changed = false;
 
+		if (s == COMMIT_ATOMIC) {
+			pipe_obj->out_fence_requested = false;
+
+			if (pipe_obj->out_fence_fd == -1)
+				continue;
+
+			igt_assert(pipe_obj->out_fence_fd >= 0);
+			igt_assert_eq(sync_fence_wait(pipe_obj->out_fence_fd, 1000), 0);
+			close(pipe_obj->out_fence_fd);
+			pipe_obj->out_fence_fd = -1;
+		}
+
 		for_each_plane_on_pipe(display, pipe, plane) {
 			plane->fb_changed = false;
 			plane->position_changed = false;
@@ -2638,6 +2628,10 @@ display_commit_changed(igt_display_t *display, enum igt_commit_style s)
 			    !(plane->type == DRM_PLANE_TYPE_PRIMARY ||
 			      plane->type == DRM_PLANE_TYPE_CURSOR))
 				plane->rotation_changed = false;
+
+			if (s == COMMIT_ATOMIC)
+				/* reset fence_fd to prevent it from being set for the next commit */
+				igt_plane_set_fence_fd(plane, -1);
 		}
 	}
 
Signature · 6 lines
2.11.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw

PATCH 3/5 tests/kms_atomic_transition: Only request fence on enabled pipes

In reply to Maarten Lankhorst

Trying to set a fence on disabled pipes will be rejected by the kernel:

[ 1275.865029] [drm:drm_atomic_check_only [drm]] [CRTC:39:pipe B] requesting event but off

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=99911

---
 tests/kms_atomic_transition.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
tests/kms_atomic_transition.c +3 -3
index 2b56eef66b7b..a2470c0c3571 100644
@@ -717,11 +717,11 @@ static void run_modeset_tests(igt_display_t *display, int howmany, bool nonblock
 			igt_plane_set_fb(plane, &fbs[1]);
 			igt_fb_set_size(&fbs[1], plane, mode->hdisplay, mode->vdisplay);
 			igt_plane_set_size(plane, mode->hdisplay, mode->vdisplay);
+
+			if (fencing)
+				igt_pipe_request_out_fence(&display->pipes[i]);
 		} else
 			igt_plane_set_fb(plane, NULL);
-
-		if(fencing)
-			igt_pipe_request_out_fence(&display->pipes[i]);
 	}
 
 	/*
Signature · 6 lines
2.11.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw

PATCH 4/5 tests/kms_atomic_transition: Do not clear in-fences after atomic commit.

In reply to Maarten Lankhorst

This is already done in igt_display_commit* functions.

---
 tests/kms_atomic_transition.c | 11 -----------
 1 file changed, 11 deletions(-)
tests/kms_atomic_transition.c +0 -11
index a2470c0c3571..6e2b25da75db 100644
@@ -367,14 +367,6 @@ static void configure_fencing(igt_display_t *display, enum pipe pipe)
 	}
 }
 
-static void clear_fencing(igt_display_t *display, enum pipe pipe)
-{
-	igt_plane_t *plane;
-
-	for_each_plane_on_pipe(display, pipe, plane)
-		igt_plane_set_fence_fd(plane, -1);
-}
-
 static void atomic_commit(igt_display_t *display, enum pipe pipe, unsigned int flags, void *data, bool fencing)
 {
 	if (fencing) {
@@ -383,9 +375,6 @@ static void atomic_commit(igt_display_t *display, enum pipe pipe, unsigned int f
 	}
 
 	igt_display_commit_atomic(display, flags, data);
-
-	if (fencing)
-		clear_fencing(display, pipe);
 }
 
 /*
Signature · 6 lines
2.11.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw

PATCH 5/5 igt/kms: Do not wait for fence completion during commit

In reply to Maarten Lankhorst

This will make the IGT tests that use fences more useful, since they can perform the waiting themselves when required.

To celebrate, also add plane-use-after-nonblocking-unbind-fencing, the fence version of plane-use-after-nonblocking-unbind.

---
 lib/igt_kms.c                 | 19 ++++++------
 tests/kms_atomic_transition.c | 71 ++++++++++++++++++++++++++++++++-----------
 2 files changed, 63 insertions(+), 27 deletions(-)
lib/igt_kms.c +10 -9
index 8bf56faf41e9..6390229f1546 100644
@@ -1698,6 +1698,7 @@ void igt_display_init(igt_display_t *display, int drm_fd)
 		pipe->plane_cursor = -1;
 		pipe->plane_primary = -1;
 		pipe->planes = NULL;
+		pipe->out_fence_fd = -1;
 
 		get_crtc_property(display->drm_fd, pipe->crtc_id,
 				    "background_color",
@@ -1913,6 +1914,9 @@ static void igt_pipe_fini(igt_pipe_t *pipe)
 
 	free(pipe->planes);
 	pipe->planes = NULL;
+
+	if (pipe->out_fence_fd != -1)
+		close(pipe->out_fence_fd);
 }
 
 static void igt_output_fini(igt_output_t *output)
@@ -2502,7 +2506,11 @@ static void igt_atomic_prepare_crtc_commit(igt_pipe_t *pipe_obj, drmModeAtomicRe
 		igt_atomic_populate_crtc_req(req, pipe_obj, IGT_CRTC_ACTIVE, !!output);
 	}
 
-	pipe_obj->out_fence_fd = -1;
+	if (pipe_obj->out_fence_fd != -1) {
+		close(pipe_obj->out_fence_fd);
+		pipe_obj->out_fence_fd = -1;
+	}
+
 	if (pipe_obj->out_fence_requested)
 	{
 		igt_atomic_populate_crtc_req(req, pipe_obj, IGT_CRTC_OUT_FENCE_PTR,
@@ -2607,16 +2615,9 @@ display_commit_changed(igt_display_t *display, enum igt_commit_style s)
 		if (s != COMMIT_UNIVERSAL)
 			pipe_obj->mode_changed = false;
 
-		if (s == COMMIT_ATOMIC) {
+		if (s == COMMIT_ATOMIC && pipe_obj->out_fence_requested) {
 			pipe_obj->out_fence_requested = false;
-
-			if (pipe_obj->out_fence_fd == -1)
-				continue;
-
 			igt_assert(pipe_obj->out_fence_fd >= 0);
-			igt_assert_eq(sync_fence_wait(pipe_obj->out_fence_fd, 1000), 0);
-			close(pipe_obj->out_fence_fd);
-			pipe_obj->out_fence_fd = -1;
 		}
 
 		for_each_plane_on_pipe(display, pipe, plane) {
tests/kms_atomic_transition.c +53 -18

142 lines not on this page · raw message

Signature · 6 lines
2.11.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Acked-by on the v1 series
#6 lore raw
In reply to Maarten Lankhorst

Hi Maarten,

2017-07-06 Maarten Lankhorst <[email protected]>:

Quoted 15 lines

I wanted to make kms_atomic_transition pass, but the nonblocking modeset fencing tests were bogus.

This series changes the semantics for fencing slightly. It only keeps the out fences in pipe_obj->out_fence_fd, it's up to the test to decide what to do with it, which is probably just waiting on the fd or checking if it completed.

Maarten Lankhorst (5):
tests/kms_atomic_transition: Add test for plane completion ordering. lib/kms: Handle fence interaction correctly WRT TEST_ONLY. tests/kms_atomic_transition: Only request fence on enabled pipes tests/kms_atomic_transition: Do not clear in-fences after atomic
commit.
igt/kms: Do not wait for fence completion during commit

Quoted 3 lines
 lib/igt_kms.c                 | 41 ++++++++----------
 tests/kms_atomic_transition.c | 99 +++++++++++++++++++++++++++++++++----------
 2 files changed, 95 insertions(+), 45 deletions(-)

This make fencing a lot cleaner.

For the whole series:

Gustavo

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#7 lore raw

Op 07-07-17 om 02:15 schreef Gustavo Padovan:

Quoted 22 lines

Hi Maarten,

2017-07-06 Maarten Lankhorst <[email protected]>:

Quoted 15 lines

I wanted to make kms_atomic_transition pass, but the nonblocking modeset fencing tests were bogus.

This series changes the semantics for fencing slightly. It only keeps the out fences in pipe_obj->out_fence_fd, it's up to the test to decide what to do with it, which is probably just waiting on the fd or checking if it completed.

Maarten Lankhorst (5):
tests/kms_atomic_transition: Add test for plane completion ordering. lib/kms: Handle fence interaction correctly WRT TEST_ONLY. tests/kms_atomic_transition: Only request fence on enabled pipes tests/kms_atomic_transition: Do not clear in-fences after atomic
commit.
igt/kms: Do not wait for fence completion during commit

Quoted 3 lines
 lib/igt_kms.c                 | 41 ++++++++----------
 tests/kms_atomic_transition.c | 99 +++++++++++++++++++++++++++++++++----------
 2 files changed, 95 insertions(+), 45 deletions(-)
Quoted 8 lines

This make fencing a lot cleaner.

For the whole series:

Gustavo

Thanks, pushed. :)

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx