Re: [Intel-gfx] [PATCH v2] Return only active connectors for get_resources ioctl

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Nov 29, 2018 at 8:37 AM Lisovskiy, Stanislav
<stanislav.lisovskiy@xxxxxxxxx> wrote:
>
> On Wed, 2018-11-28 at 22:21 +0100, Daniel Vetter wrote:
> > On Wed, Nov 28, 2018 at 09:51:13PM +0100, Daniel Vetter wrote:
> > > On Wed, Nov 28, 2018 at 03:55:58PM +0200, Stanislav Lisovskiy
> > > wrote:
> > > > Currently kernel might allocate different connector ids
> > > > for the same outputs in case of DP MST, which seems to
> > > > confuse userspace. There are can be different connector
> > > > ids in the list, which could be assigned to the same
> > > > output, while being in different states.
> > > > This results in issues, like external displays staying
> > > > blank after quick unplugging and plugging back(bug #106250).
> > > > Returning only active DP connectors fixes the issue.
> > > >
> > > > v2: Removed caps from the title
> > > >
> > > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=106250
> > > > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy@xxxxxxxxx
> > > > >
> > > > ---
> > > >  drivers/gpu/drm/drm_mode_config.c | 16 +++++++++++-----
> > > >  1 file changed, 11 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/drm_mode_config.c
> > > > b/drivers/gpu/drm/drm_mode_config.c
> > > > index ee80788f2c40..ec5b2b08a45e 100644
> > > > --- a/drivers/gpu/drm/drm_mode_config.c
> > > > +++ b/drivers/gpu/drm/drm_mode_config.c
> > > > @@ -143,6 +143,7 @@ int drm_mode_getresources(struct drm_device
> > > > *dev, void *data,
> > > >   drm_connector_list_iter_begin(dev, &conn_iter);
> > > >   count = 0;
> > > >   connector_id = u64_to_user_ptr(card_res-
> > > > >connector_id_ptr);
> > > > + DRM_DEBUG_KMS("GetResources: writing connectors start");
> > > >   drm_for_each_connector_iter(connector, &conn_iter) {
> > > >           /* only expose writeback connectors if userspace
> > > > understands them */
> > > >           if (!file_priv->writeback_connectors &&
> > > > @@ -150,15 +151,20 @@ int drm_mode_getresources(struct drm_device
> > > > *dev, void *data,
> > > >                   continue;
> > > >
> > > >           if (drm_lease_held(file_priv, connector-
> > > > >base.id)) {
> > > > -                 if (count < card_res->count_connectors
> > > > &&
> > > > -                     put_user(connector->base.id,
> > > > connector_id + count)) {
> > > > -                         drm_connector_list_iter_end(&con
> > > > n_iter);
> > > > -                         return -EFAULT;
> > > > +                 if (connector->connector_type !=
> > > > DRM_MODE_CONNECTOR_DisplayPort ||
> > > > +                         connector->status !=
> > > > connector_status_disconnected) {
> > > > +                         if (count < card_res-
> > > > >count_connectors &&
> > > > +                             put_user(connector->base.id,
> > > > connector_id + count)) {
> > > > +                                 drm_connector_list_iter_
> > > > end(&conn_iter);
> > > > +                                 return -EFAULT;
> > > > +                         }
> > > > +                         DRM_DEBUG_KMS("GetResources:
> > > > connector %s", connector->name);
> > > > +                         count++;
> > >
> > > I tried to read the bug and I have no idea what's going on here.
> > > Userspace
> > > is supposed to shut off outputs that are disconnected, whether
> > > that's DP,
> > > DP MST or something else shouldn't matter. New connectors can
> > > come&go as
> > > they see fit. Also not really something special.
> > >
> > > Why do we need to dynamically hide an output here? Note that this
> > > also
> > > affects normal DP ports, which I have no idea is actually what you
> > > want to
> > > do or not.
>
> This bug is real and easily reproducible with recent drm-tip.
> Unplugging and then quickly plugging back periodically leaves both my
> external displays connected to the docking station blank, there are
> also many duplicate bugs for this, which I simply didn't track. This
> patch at least fixes that annoying thing.
> The userspace seems to get confused when we are returning two different
> connector ids, one in disconnected state and another in connected state
> for the same output. This results in userspace believing that nothing
> had changed and drm_mode_setcrtc call is not required( I have done
> traces confirming that theory).
>
> This could be also fixed in userspace by checking connectors more
> carefully - that fix I've also implemented for Intel DDX and attached
> to the bug, however seems that this happens also for Wayland.

I didn't check wayland (well, weston or any of the others, wayland
doesn't exist as an implementation), but -modesetting and -amdgpu. And
they all have the same issue. It's roughly:
- You unplug
- Kernel sends out uevent
- Userspace is slow with shutting down that output
- You replug
- Kernel creates a new connector (it's always a new connector from the
kernel's pov for MST, which might or might not happen to have the same
id as a previous one).
- Userspace sees the hotplug, and uses the PATH property to check
whether it has seen this mst connector already. This is needed on X
since you can't unplug connectors in xrandr ever, so reusing old ones
is a good idea.
- Userspace is confused.

Your fix essentially goes back to really old MST world where the
kernel yanked the MST output right away, without waiting for
userspace. That's kinda not nice. Except it's only hidden, not even
shut off, so userspace has a hard time shutting it down. It also has
the fundamental problem that if you yank&replug fast enough (fast
enough that userspace can't see the in-between unplugged state), then
nothing happens, and userspace is again confused. Aside: You want
Chris patch to implement your idea.

There's a few ways to fix this in proper fashion I think:
- We wait on the kernel side with registering the new connector until
userspace has caught up. Would be a bit a hack, and I haven't figured
out a good way yet where to put that hack. Probably the least hacky
way would be:
1. In getresources we scan all the previous connector's PATH property,
if it matches we don't (yet) expose that connector. This should only
expose the older connector.
2. In the connector cleanup work we need to fire another uevent, to
make sure userspace notices.

My little patch to order connectors the other way round was kinda
aimed at implementing this the cheap way (I think userspace ignores
all but the last connector it sees), but I guess that didn't work.

- 2nd option, and what I think we should aim for long-term: Kernel
reuses the same connector if it notices that userspace hasn't cleaned
it up yet. This is going to be real tricky to implement from a object
lifetime pov. Since the link won't work anymore we also need to send a
link_status update to inform userspace that it needs to do a modeset
(like in other cases with DP link issues). Lyude's port lifetime
rework should at least get us closer to this I think.

Cheers, Daniel

> >
> > For entertainment and other reasons, testing the below diff would be
> > interesting.
> >
> >
> > diff --git a/drivers/gpu/drm/i915/intel_dp_mst.c
> > b/drivers/gpu/drm/i915/intel_dp_mst.c
> > index 4de247ddf05f..e1b66396c83b 100644
> > --- a/drivers/gpu/drm/i915/intel_dp_mst.c
> > +++ b/drivers/gpu/drm/i915/intel_dp_mst.c
> > @@ -499,6 +499,8 @@ static void
> > intel_dp_register_mst_connector(struct drm_connector *connector)
> >               drm_fb_helper_add_one_connector(&dev_priv->fbdev-
> > >helper,
> >                                               connector);
> >
> > +     list_move(&connector->head, &connector->dev-
> > >mode_config.connector_list);
> > +
> >       drm_connector_register(connector);
> >  }
> >
> --
> Best Regards,
>
> Lisovskiy Stanislav



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch



[Index of Archives]     [Linux Kernel]     [Kernel Development Newbies]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite Hiking]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux