Bug 15439

Created attachment 15812 [details]
Digicam of screen issue, notice the bars to the left and right of the google earth pixmap.

Black bars are only present when tiling is enabled

Kevin, are you able to git-bisect to find the culprit commit?

So you saw this with intel driver git master branch right? Could you try xf86-video-intel-2.3-branch instead? It can tell us if this is caused by recent i965 render improvement patches or keith's transform change.

Bug is present in xf86-video-intel-2.3-branch

Attaching Xorg.0.log as confirmation that I have the right driver

Created attachment 15836 [details]
Xorg.0.log with 2.3 branch of the driver

Tried driver 2.2.1 and it has the same issue with Tiling enabled

Tried to go back to driver 2.1 and it won't compile on my system due to errors in I810AccelInit as it is missing members in _I810Rec

Also a note on the master branch with tiling enabled. If I leave glxgears running for about 30 seconds. The machine will hardlock. If I disable Tiling, it will run without locking.

As of Apr 11, 2008
drm - git
mesa - git
xf86-drv-intel - git

RPM of xorg from fedora rawhide
xorg-x11-server 1.4.99.901-18.20080401.fc9

Kevin, I'll just ignore your glxgears problem in this track.

I've also done like your tests with intel driver 2.3 branch and 2.2.1, both googleearth caused screen artifacts with xserver-1.5 and pixman master.

I also tried intel 2.3 branch with xserver-1.4, mesa-7.0, and I _can't_ see the problem. So this seems a xserver or pixman regression, bisect xserver exa/ changes might be able to show what's wrong.

I did a couple of tests with my setup

I switched the AccelMethod to XAA and the lines did disappear with Tiling and without Tiling, however the display in XAA mode was corrupted in other ways so it was not usable for me. 

I also tried to compile mesa 7.1 but I was unable to do so. So I am still running mesa master.

So it looks like the bug is in EXA mode + Tiling. At this point I have no idea how to bisect it anymore. I'm not sure I can bisect the change that broke it since the error occurs in all of the components I can compile. 

I am willing to test and try and recreate things but other than that, I think it will have to be debugged by someone with a little more experience in this area.

I will continue to test and report back what I find.

Kevin, I tried with today's master branches, drm, mesa, xserver, xf86-video-intel (master and 2.3-branch), it seems if I loaded ttm-enabled drm kernel modules, googleearth can work just fine (no screen corrupts and fonts are right in googleearth).

If I use vanilla 2.6.24 kernel's drm modules, it just shows the corrupt. And as I tested before, mesa-7.0 also works fine.

Could you try that too? Looks this's a mesa master problem in ttm fallback classic mode.

I tested with todays code using git master of drm, mesa and xf86-drv-intel and I believe I am running in TTM mode (is there a way to verify) and I am still getting the issue. 

Again this is with tiling on, if I turn tiling off the problem goes away.

(In reply to comment #11)
> I tested with todays code using git master of drm, mesa and xf86-drv-intel and
> I believe I am running in TTM mode (is there a way to verify) and I am still
> getting the issue. 

Do you load drm modules from drm source linux-core/? (make i915.o; insmod drm.ko; insmod i915.ko) dmesg shows what's version of i915 module you've loaded, and export LIBGL_DEBUG=verbose might have more info. I don't know other environ you might set, but mesa master i965 dri driver will say if it fails to detect ttm and fallback to classic.

> 
> Again this is with tiling on, if I turn tiling off the problem goes away.
> 

Even if disable tiling, I still can see fonts in googleearth window are broken (city names have missing glyphs,etc.), although mine is T61 with GM965.

Yes, I am running the full drm stack from git master. I do compile the main code and the linux_core code and install it. 

From mesa I do not get any messages that state that I am falling back to classic mode (I have seen that message before, but I don't get it now). So I am pretty sure I am running all the right parts.

on my chip (g35) I do not see any font issues in google earth with tiling _disabled_. I even looked at the Hong Kong airport and verified that non-latin characters were working ok.


$ export LIBGL_DEBUG=verbose
$ googleearth 
libGL: XF86DRIGetClientDriverName: 1.9.0 i965 (screen 0)
libGL: OpenDriver: trying /usr/lib/dri/i965_dri.so
drmOpenDevice: node name is /dev/dri/card0
drmOpenDevice: open result is 8, (OK)
drmOpenByBusid: Searching for BusID pci:0000:00:02.0
drmOpenDevice: node name is /dev/dri/card0
drmOpenDevice: open result is 8, (OK)
drmOpenByBusid: drmOpenMinor returns 8
drmOpenByBusid: drmGetBusid reports pci:0000:00:02.0
libGL error: 
Can't open configuration file /etc/drirc: No such file or directory.
libGL error: 
Can't open configuration file /home/kdekorte/.drirc: No such file or directory.
wx=0.312690, wy=0.328990, rx=0.639990, ry=0.330000


I'll attach the output of glxinfo as well (Tiling off)

Created attachment 15905 [details]
output of glxinfo with tiling off

Probably not real useful, but should be able to show you that I am using the right stuff to test.

I think I found another clue with googleearth with tiling on. I found that if I zoomed in slowly from the globe to the continent. If the texture didn't change the screen corruption didn't change. However if I found a spot where the textures changed to to LOD changes. IE high lod to a lower one and vis versa that is when the corruption came up. So I believe it is a problem when changing textures on a surface. But that is just my guess to the problem.

Also, I tried getting xorg server from git master and the thing just exploded on me. So I had to revert back to my distros Xserver.

Just another data point, but I finally got xserver upgraded to git master and even with that a tiling on I get screen corruption as seen in the attached jpeg. So still a problem..

from git

drm
mesa
xserver
xf86_drv_intel

So I think I'm at the same level as you now.

Attaching Xorg.0.log for confirmation

Created attachment 15910 [details]
Xorg.0.log when using git master of xserver

Just for grins, I upgraded pixman to git master as well and that did not seem to change anything.

Also, I tried with AccelMethod XAA and No offscreen pixmaps set and noticed something... I don't get the blackbars outside if the window like I do with XAA but while the texture is being refreshed, I do see screen corruption like I do with EXA just for a split second until the texture is fully loaded. What it looks like is the same corruption you get outside the window, but constrained to the OpenGL surface... so you see the black bars, but after the texture is fully loaded and the surface is refreshed the bars go away.

I've been thinking about this problem and I'm wondering if in tiling mode if there is call that clears the memory for the texture being uploaded. I'm thinking that since the black bars are being seen that rather than memory being cleared for each tile or tile band, it is being cleared for the entire texture and the tiles and the the memory are not quite the same or the X offset for each row of the tile array for the clear is not reset prior to the clear. 

The problem I see it looks like this

Say the screen is 30x30 and the texture is at 5,5 and is 10,10

--------------------------------
|
|
|
|     *XXXXXXXXX
|                XXXXXXXXXX
|                          XXXXX
|XXXXXXXXXXXXXXX

etc...


The black bars are the X's in the diagram so it looks to me that either when the memory is cleared the X offset is not reset properly or a block of memory is just cleared once for the entire size of the texture and it is clearing the wrong blocks. The titles are being placed in the right spot since when the texture is used the texture looks correct.

I think that the difference between XAA and EXA mode could be that XAA has better clipping of where bits are written and so the problem is not in either EXA or XAA but in the tile upload process.

Feel free to tell me I'm totally wrong on this. This is just a guess based on observation of the garbage on the screen.

Did some more testing and switched the batch buffer branch in DRI2 mode and everything looks great there with Tiling mode on.

OpenGL code in some cases (etracer) is slightly fewer FPS in batch buffer mode over non-batchbuffer in tiling mode, but that fact that it works without artifacts in a positive in my opionion.

Once batchbuffer is merged into master, we can retest and if things look good we can close this.

ok one more note. Using intel-batchbuffer branch I switched from DRI2 mode to DRI mode just by changing the setting in the xorg.conf file. After that the black lines came back. So that appears to point the problem in the direction of the dri code. 

I tried on one G35 here, master branches seem work fine on it like on GM965.

batch-buffer driver issue should be on seperate track though.

I have git/master of drm, mesa, xserver, and pixman. What else could I need master off?

Also odd that DRI1 shows the error and DRI2 makes it go away. Can I ask what hardware you are testing with the G35 chip?

Sorry for delay, it's an ASUS board with G35 and HDMI output, also on my T61 with GM965. 

Kevin, could you try current drm master? Keith has pushed several fixes in his googleearth testing.

assign back to gordon to hand this to somebody else.

Well I tried to compile the entire X stack. drm, mesa, xserver and the intel driver.

drm and mesa compiled ok, but the xserver won't compile right now. And the intel driver gives an awful mess for text right now. (it appears to have been broken for a couple of days, but not sure if it is related to xserver or not).

Anyway, with updated drm and mesa, I try to run anything 3d and the machine crashes pretty much right away.

Ok, I got this to build with all the latest stuff...

drm, mesa, xserver, pixman all from git master

xf86-drv-intel is from the batchbuffer branch (git master has problems with text display and screen corruption)

With those items in place, running with EXA and Tiling and DRI (version 1) I am able to play mplayer -vo gl video and not get the horizontal stripes. 

I tried running with googleearth, and the machine hangs.

Also I experienced other hangs with the system using gtkperf and other display tests. 

Using the batchbuffer branch with DRI2, googleearth runs ok, 15500 is still outstanding.

Using git master versions of drm, mesa and xf86-video-intel (4/21/08) , 
and setting tiling disabled , this problem still exist , but if I change 
xf86-video-intel to 2.3 branch , this case can run fine . 

reassigning to Nanhai, as Zhenyu believes it's 3d issue.

I updated the xserver to the master tip , this bug can't be reproduced . 

Is this bug gone?

it 's ok in the latest commit of master.

verified.