On Sun, Feb 28, 2021 at 10:10 AM Uwe Sauter <uwe.sauter.de@xxxxxxxxx> wrote:
Am 27.02.21 um 22:26 schrieb Ilia Mirkin:
On Sat, Feb 27, 2021 at 7:28 AM Uwe Sauter <uwe.sauter.de@xxxxxxxxx> wrote:
I can also report that the modesetting ddx that comes with xorg-server 1.20.10-3 (Arch Linux package) shows this kind of
cursor-cut-into-horizontal-stripes behavior. Changing to xf86-video-nouveau 1.0.17-1 solves this issue. But nouveau has
issues with Mate 1.24 (as discussed earlier this month).
My hardware:
# lspci -s 3:0.0 -v
03:00.0 VGA compatible controller: NVIDIA Corporation GK208B [GeForce GT 710] (rev a1) (prog-if 00 [VGA controller])
Subsystem: ASUSTeK Computer Inc. GT710-4H-SL-2GD5
Flags: bus master, fast devsel, latency 0, IRQ 36, IOMMU group 12
Memory at fb000000 (32-bit, non-prefetchable) [size=16M]
Memory at fff0000000 (64-bit, prefetchable) [size=128M]
Memory at fff8000000 (64-bit, prefetchable) [size=32M]
I/O ports at f000 [size=128]
Expansion ROM at fc000000 [disabled] [size=512K]
Capabilities: [60] Power Management version 3
Capabilities: [68] MSI: Enable+ Count=1/1 Maskable- 64bit+
Capabilities: [78] Express Legacy Endpoint, MSI 00
Capabilities: [100] Virtual Channel
Capabilities: [128] Power Budgeting <?>
Capabilities: [600] Vendor Specific Information: ID=0001 Rev=1 Len=024 <?>
Capabilities: [900] Secondary PCI Express
Kernel driver in use: nouveau
Kernel modules: nouveau
If I can help in any way please let me know.
Thanks, that's good info. Simon - you originally said that everything
looked good on your GK208, so a retest would be super.
I just double-checked on a GP108 (with an older kernel, but same idea
should apply), and it seems like 256x256 cursors are fine there.
However the display logic has gone through some ideally no-op updates
since that kernel version, but there could very easily be issues.
Can you try Alex's patch to modetest and confirm that you see issues
with modetest? If so, can you (and maybe Alex as well) try an older
kernel (I'm on 5.6) and see if modetest behaves well there. [The patch
in question was to expose 256x256 as the 'preferred' size, but support
for the larger cursors has been around for a while.] Alex - if you
have time, same question to you.
You can find the patch here:
https://lists.x.org/archives/nouveau/2021-February/037992.html
I had to install a parallel Arch Linux to my existing production system in order to keep it clean from all the
development stuff.
System summary (most recent):
AMD Ryzen 3 3100
Gigabyte B550M S2H with BIOS F13c
Asus GT710-4H-SL-2GD5 (GK208B [GeForce GT 710] (rev a1)) using nouveau kernel module
32GB DDR4-3200MHz ECC
libdrm 2.4.104-1
linux 5.11.2.arch1-1
mesa 20.3.4-3
xf86-video-nouveau 1.0.17-1
I built libdrm 2.4.104.r16.ga9bb32cf in order to get modetest.
With unmodified kernel / modetest (cw=64, ch=64) I call:
$ ./modetest -c |grep '^[0-9]\|preferred'
85 86 connected HDMI-A-1 530x300 40 86
#0 1920x1080 60.00 1920 2008 2052 2200 1080 1084 1089 1125 148500 flags: phsync, pvsync; type: preferred, driver
87 0 disconnected HDMI-A-2 0x0 0 88
89 0 disconnected HDMI-A-3 0x0 0 90
91 0 disconnected HDMI-A-4 0x0 0 92
./modetest -s 85:1920x1080 -C
trying to open device 'i915'...failed
trying to open device 'amdgpu'...failed
trying to open device 'radeon'...failed
trying to open device 'nouveau'...done
setting mode 1920x1080-60.00Hz on connectors 85, crtc 50
starting cursor
^C
This shows several things:
* There is a moving gray, half transparent square bouncing around. I believe that this is
the mentioned cursor.
* The cursor movement happens at various speeds, sometimes staying half a second on the
same position to then move quite fast to another, then slowing down.
* The cursor is flickering.
* When (forcefully) ending the test the screen is not properly reset, leaving the
previous content in a state similar to the phenomenon with the mouse cursor that stated
this discussion. On my FullHD display the console output is sliced horizontally and
offset with about 1/5th of the screen width.
This also happens on my other machine with a Xeon E3-1245 v3 with integrated graphics on a ASRock C226 WS, using the
i915 kernel module and same software versions as above.
Applying Alex' patch with (cw=128, ch=128) shows a cursor that contains the same test pattern as is shown in the
background. The behavior is as jumpy and flickery as it was with size 64.
When killing the test the last position of the cursor still shows the test pattern while the background is again sliced
and shuffled horizontally.
Setting the size to 256 shows an even larger cursor. It shows the same jumpy and flickery behavior as the other two. The
cursor itself also shows a horizontal sliced in the lower half. After killing the test the cursor's last position still
shows the test pattern while the background is sliced.
This testing was all conducted with packages from the Arch Linux distribution (but for modetest).
Questions:
1) Is this jumpy and flickery behavior expected or should the cursor move smoothly?
Good question. It's definitely jumpy/flickery for me too. I haven't
looked at why, but I wouldn't worry about it. I suspect it has to do
with the mechanics of how it causes the cursor to be moved.
2) With unmodified modetest, what should the cursor look like? Without further inspection
of the code I suspect that the change from UTIL_PATTERN_PLAIN to UTIL_PATTERN_SMPTE
changed the cursor's appearance.
The PLAIN pattern is just gray, which isn't necessarily a great test
to see visual corruption.
4) How long is modetest expected to run? On the first run I let it test for over 10min
before deciding to kill it.
Until you hit enter (or escape or maybe any key, I forget).
5) Is modetest expected to reset the display to the state it was before? Why doesn't it
do that when being killed?
No. You can switch vt's back and forth to restore. It's just a test
application. It's unfortunately not an entirely trivial thing to do.
6) Where do you expect this bug to come from? Kernel nouveau driver, modesetting ddx,
nouveau ddx?
modetest interacts with the kernel directly. The bug is most likely in
the hardware, and we should just not use the 256x256 size even though
allegedly the hw supports it. But perhaps the kernel screws something
up.
7) Any proposal what kernel to test next?
If you could test modetest with 256x256 cursor on a pre-5.9 kernel and
ensure that you see the same corruption in the cursor image, that'd
confirm that we didn't just screw something up in the macro rework
which landed in 5.9, vs a hw issue.
