weston-9 crash on page_flip_handler when system suspend/resume

background

drm-backend.so crashes in the following code

static void
atomic_flip_handler(int fd, unsigned int frame, unsigned int sec,
		    unsigned int usec, unsigned int crtc_id, void *data)
{
	......
	assert(output->page_flip_pending); //fails
	......
}

page_flip_pending value is false causing crash.The log is as follows

audit[576]: ANOM_ABEND auid=4294967295 uid=0 gid=0 ses=4294967295 subj=kernel pid=576 comm="weston" exe="/usr/bin/weston" sig=6 res=1
weston[576]: weston: ../weston-9.0.0/libweston/backend-drm/kms.c:1409: void atomic_flip_handler(int, unsigned int, unsigned int, unsigned int, unsigned int, void *): Assertion `output->atomic_complete_pending' failed.

reason

• The system will trigger session active/deactive when suspend/resume

static void
session_notify(struct wl_listener *listener, void *data)
{
	......
	if (compositor->session_active) {
		weston_log("activating session\n");
		......
		b->state_invalid = true;
		......
	} else {
		weston_log("deactivating session\n");
		......

• When the scheduling timer exits the schedule state, it will still submit an null atomic that does not include crtc and plane requests to drm.

static int
output_repaint_timer_handler(void *data)
{
	......
	wl_list_for_each(output, &compositor->output_list, link) {
		ret = weston_output_maybe_repaint(output, &now, repaint_data); // ret = 0, no repaint
		if (ret)
			break;
	}

	if (ret == 0) {
		if (compositor->backend->repaint_flush)
			ret = compositor->backend->repaint_flush(compositor,
							 repaint_data);         //flush repaint , drmModeAtomicCommit no crtc/plane request
	......
	output_repaint_timer_arm(compositor);

	return 0;
}

• If the session signal event comes earlier than the scheduled timer event. The following will happen:

  1.b->state_invalid = true;

  2.output_repaint_timer_handler commit null atomic;

  3.flags |= DRM_MODE_ATOMIC_ALLOW_MODESET;

  4.page flip event signal;

  5.drm-backend.so assert fail and crash;

static int
drm_pending_state_apply_atomic(struct drm_pending_state *pending_state,
			       enum drm_state_apply_mode mode)
{
	......
	if (b->state_invalid) {                                            // session active happened before
                ......
		flags |= DRM_MODE_ATOMIC_ALLOW_MODESET; 
	}

	wl_list_for_each(output_state, &pending_state->output_list, link) {  // output_list is null, no repaint
		......
		ret |= drm_output_apply_state_atomic(output_state, req, &flags);
	}
        ......
	ret = drmModeAtomicCommit(b->drm.fd, req, flags, b);                //this commit will trigger a page flip event
	......
	wl_list_for_each_safe(output_state, tmp, &pending_state->output_list,
			      link)                                         // output_list is null， output->page_flip_pending will not be set
		drm_output_assign_state(output_state, mode);
        ......
}

how to fix

weston12 has fixed this problem, if weston_output_maybe_repaint return no request, will not trigger repaint_flush.

Because this is just an occasional fault, the current analysis conclusion is only inferred based on the code. Is there any problem with this analysis?

Does weston9 has a patch for this problem? Or I refer to the logic of weston12 to modify output_repaint_timer_handler, will only modifying the logic of output_repaint_timer_handler introduce other problems?