Hi all,
So in the course of beating the pageflip paths senseless to test the new locking scheme I've hunted down some bugs on g33. Sometimes we seemingly time out waiting for the event - apparently the pageflip irq is lost somewhere. I've tried to dig in, but lost myself completely in the rather fickle vblank mess in drm_irq.c.
To not write off the last few days completely, I've figured I'll submit the two sane patches for inclusion - the first should fix fbcon doing stupid things, the second is purely paranoia (I hope at least). Imho longterm we need to seriously look at the current vblank stuff - since it operates on pipe ides instead of modeset crtc objects it's all rather fiddly and easily falls appart. But keeping both ums and kms paths working is a challenge ...
Comments highly welcome.
Yours, Daniel
Daniel Vetter (2): drm/fb_helper: check whether fbcon is bound drm/i915: wake up all pageflip waiters
drivers/gpu/drm/drm_fb_helper.c | 39 +++++++++++++++++++++++++--------- drivers/gpu/drm/i915/intel_display.c | 2 +- 2 files changed, 30 insertions(+), 11 deletions(-)
We need to make sure that the fbcon is still bound when touching the hw, since otherwise we might corrupt the modeset state of kms clients. X mostly works around that with VT switching and setting the VT into raw mode, which disables most fbcon events.
Raw kms test programs though don't do that dance, and in the future we might want to aim to abolish CONFIG_VT anyway. So improve preventive measures a bit. To do so, extract the existing logic for handling hotplug events (which X can't block with the current set of tricks) and reuse it for the fbdev blanking helper.
Long-term we really need to either scrap this all and only have a OOPS console, or come up with a saner model for device ownership sharing between fbdev/fbcon and kms userspace.
And for those wondering whether this check can't be added to fb_set_par e.g. to prevent undue panning - drm initialization locking is too screwed up for that to work out, I'm sorry :(
Signed-off-by: Daniel Vetter daniel.vetter@ffwll.ch --- drivers/gpu/drm/drm_fb_helper.c | 39 +++++++++++++++++++++++++++++---------- 1 file changed, 29 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c index be0f2d6..0c6e25e 100644 --- a/drivers/gpu/drm/drm_fb_helper.c +++ b/drivers/gpu/drm/drm_fb_helper.c @@ -305,6 +305,24 @@ void drm_fb_helper_restore(void) } EXPORT_SYMBOL(drm_fb_helper_restore);
+static bool drm_fb_helper_is_bound(struct drm_fb_helper *fb_helper) +{ + struct drm_device *dev = fb_helper->dev; + struct drm_crtc *crtc; + int bound = 0, crtcs_bound = 0; + + list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { + if (crtc->fb) + crtcs_bound++; + if (crtc->fb == fb_helper->fb) + bound++; + } + + if (bound < crtcs_bound) + return false; + return true; +} + #ifdef CONFIG_MAGIC_SYSRQ static void drm_fb_helper_restore_work_fn(struct work_struct *ignored) { @@ -338,6 +356,11 @@ static void drm_fb_helper_dpms(struct fb_info *info, int dpms_mode) * For each CRTC in this fb, turn the connectors on/off. */ drm_modeset_lock_all(dev); + if (!drm_fb_helper_is_bound(fb_helper)) { + drm_modeset_unlock_all(dev); + return; + } + for (i = 0; i < fb_helper->crtc_count; i++) { crtc = fb_helper->crtc_info[i].mode_set.crtc;
@@ -702,6 +725,11 @@ int drm_fb_helper_pan_display(struct fb_var_screeninfo *var, int i;
drm_modeset_lock_all(dev); + if (!drm_fb_helper_is_bound(fb_helper)) { + drm_modeset_unlock_all(dev); + return -EBUSY; + } + for (i = 0; i < fb_helper->crtc_count; i++) { crtc = fb_helper->crtc_info[i].mode_set.crtc;
@@ -1369,21 +1397,12 @@ int drm_fb_helper_hotplug_event(struct drm_fb_helper *fb_helper) struct drm_device *dev = fb_helper->dev; int count = 0; u32 max_width, max_height, bpp_sel; - int bound = 0, crtcs_bound = 0; - struct drm_crtc *crtc;
if (!fb_helper->fb) return 0;
drm_modeset_lock_all(dev); - list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { - if (crtc->fb) - crtcs_bound++; - if (crtc->fb == fb_helper->fb) - bound++; - } - - if (bound < crtcs_bound) { + if (!drm_fb_helper_is_bound(fb_helper)) { fb_helper->delayed_hotplug = true; drm_modeset_unlock_all(dev); return 0;
Otherwise it seems like we can get stuck with concurrent waiters. Right now this /shouldn't/ be a problem, since all pending pageflip waiters are serialized by the one mode_config.mutex, so there's at most on waiter. But better paranoid than sorry, since this is tricky code.
Signed-off-by: Daniel Vetter daniel.vetter@ffwll.ch --- drivers/gpu/drm/i915/intel_display.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 6056972..f49fd2a 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -6788,7 +6788,7 @@ static void do_intel_finish_page_flip(struct drm_device *dev,
obj = work->old_fb_obj;
- wake_up(&dev_priv->pending_flip_queue); + wake_up_all(&dev_priv->pending_flip_queue);
queue_work(dev_priv->wq, &work->work);
On Mon, 17 Dec 2012 16:59:00 +0100, Daniel Vetter daniel.vetter@ffwll.ch wrote:
Otherwise it seems like we can get stuck with concurrent waiters. Right now this /shouldn't/ be a problem, since all pending pageflip waiters are serialized by the one mode_config.mutex, so there's at most on waiter. But better paranoid than sorry, since this is tricky code.
Hmm.
BUG_ON(waitqueue_active()) before joining the waitqueue? -Chris
dri-devel@lists.freedesktop.org