| Bug ID | 98030 |
|---|---|
| Summary | Stuttering video playback in totem after update to 1.19-rc1 |
| Product | xorg |
| Version | unspecified |
| Hardware | Other |
| OS | All |
| Status | NEW |
| Severity | normal |
| Priority | medium |
| Component | Driver/nouveau |
| Assignee | nouveau@lists.freedesktop.org |
| Reporter | piotrdrag@gmail.com |
| QA Contact | xorg-team@lists.x.org |
After updating to xorg-x11-server-1.19.0-0.1.20160929.fc25 and xorg-x11-drv-nouveau-1.0.13-1.fc25 (from <https://bodhi.fedoraproject.org/updates/FEDORA-2016-a5c3ebe67a>)= video playback in totem stutters every few seconds to the point of being unwatcha= ble. It doesn't happen with video playback (e.g. YouTube) in Firefox, and system generally works fine. Downgrading the update "fixes" the problem. I'm using X11 with nouveau on a GT216M (GeForce GT 240M), no hybrid cards: Providers: number : 1 Provider 0: id: 0x64 cap: 0x7, Source Output, Sink Output, Source Offload crtcs: 2 outputs: 3 associated providers: 0 name:nouveau Hans de Goede in <https://bugzilla.redhat.com/show_bug.cgi?id=3D1381014> k= indly helped to discover that xf86-video-nouveau-1.0.13 does not show this problem with xserver-1.18.4, but does show it with xserver-1.19-rc1, and that using= the modesetting driver with xserver-1.19-rc1 makes the problem go away.
Looks like some issue with DRI2. If you enable DRI3 support in= your xorg.conf, should work fine (for DRI3 clients). [I realize this is not a fix, but rath= er a workaround.]
Apparently the issue also exists with modesetting on top of no= uveau - i.e. if you run e.g. LIBGL_DRI3_DISABLE=3D1 glxgears, you'll end up with the same l= ag. Trying to collect more data to see if this is nouveau-specific or if it's an issue across drivers.
Ok, i think this is related: On 1.19-rc1 (current git master) = under DRI2, windowed OpenGL apps like glxgears get stuck unless i press keys on the keyboard or move the mouse to create input events. KDE Plasma 5, being Open= GL driven afaik, has the same problem. If otoh i run applications which use fullscreen kms page-flipped windows, e= .g., Gnome shell, or regular fullscreen GL apps, they work fine under DRI2. If i= use DRI3/Present everything works fine. This both with a server built to use the new input-threads and also built without input-threads. This happens at lea= st on nouveau-ddx and ati/amdgpu-ddx. I just retested glxgears, my own app and totem under gdb, and without me providing mouse/keyboard input, they all get stuck in the DRI2GetBuffersWithFormat request to the X-Server, waiting for a reply. That gets called when Mesa needs new renderbuffers for a new frame after a swapbuffers request, e.g., when glxgears or totem calls glClear(). So far so bad. So something gets stuck in the servers dispatch loop, but "external" input events from the kernel (evdev, kms-pageflip comp= letion events?) gets it unstuck?
| What | Removed | Added |
|---|---|---|
| CC | ajax@nwnk.net, keithp@keithp.com |
I bisected this to commit 8f1edf4bd3a1f050ce9eeb5eac45dd1a8f7a6d5e Author: Keith Packard <keithp&#= 64;keithp.com> Date: Thu May 19 13:59:54 2016 -0700 dix: Use list for ready clients This converts the dispatch loop into using a list of ready clients instead of an array. This changes the WaitForSomething API so that it notifies DIX when a client becomes ready to read, instead of returning the set of ready clients.
(In reply to Mario Kleiner from comment #3) > If otoh i run applications which use fullscreen = kms page-flipped windows, > e.g., Gnome shell, or regular fullscreen GL apps, they work fine under= DRI2. I can confirm that for mutter, but not for e.g. glxgears in fullscreen. May= be compositors just get unblocked by something else, similar to how input unbl= ocks all affected DRI2 clients.
The change in 8f1edf4bd3a1f050ce9eeb5eac45dd1a8f7a6d5e is that=
we no longer
wakeup and flush writers when space in the write queue becomes available -
instead before we poll() we flush the writers iff NewOutputPending since the
last poll(), i.e.
diff --git a/os/WaitFor.c b/os/WaitFor.c
index 8164c30..5b0c342 100644
--- a/os/WaitFor.c
+++ b/os/WaitFor.c
@@ -207,8 +207,7 @@ WaitForSomething(Bool are_ready)
(1000000 / MILLI_PER_SECOND);
wt =3D &waittime;
}
- if (NewOutputPending)
- FlushAllOutput();
+ FlushAllOutput();
/* keep this check close to select() call to minimize race */
if (dispatchException)
i =3D -1;
(In reply to Chris Wilson from comment #6) > - if (NewOutputPending) > - FlushAllOutput(); > + FlushAllOutput(); This doesn't fix the problem for me.
(In reply to Michel D=C3=A4nzer from comment #7) > (In reply to Chris Wilson from comment #6) > > - if (NewOutputPending) > > - FlushAllOutput(); > > + FlushAllOutput(); >=20 > This doesn't fix the problem for me. Yeah, because the first thing FlushAllOutput() does is check NewOutputPendi= ng. Bleh. :| I'm still pretty certain it is the change in considering when to f= lush output that causes the event/replies to become stuck.
| What | Removed | Added |
|---|---|---|
| Component | Driver/nouveau | Server/General |
| Assignee | nouveau@lists.freedesktop.org | xorg-team@lists.x.org |