Driver fails to validate 3440x1440@160Hz on Linux, works on Windows

[crimist]

Posting here as suggested by Andy.

@crimist: Yes, I think your issue is unrelated to the DSC + multiple-heads-driving-one-monitor-behind-your-back functionality of this Issue.

It would probably be best to start a new thread on the NVIDIA Linux forums. I would suggest adding the “ModeDebug” X configuration option to your X config file (nvidia-xconfig --mode-debug, or add it manually), then restart X11, and look at the X log. The log should list why each mode was considered valid or invalid. That should tell us why the driver is invalidating the 3440x1440@160Hz mode.

On Linux, X logs show that the 3440x1440@160Hz mode fails to validate. However, the mode works perfectly fine on Windows. I have had this issue since I started daily driving Linux. At the time the NVIDIA latest driver was 515.57.

GPU	1080 Ti
Kernel	6.3.8
Driver	535.54.03

Log excerpt below. Full logs available here.

[ 72444.150] (II) NVIDIA(GPU-0):   Validating Mode "3440x1440_144":
[ 72444.150] (II) NVIDIA(GPU-0):     Mode Source: EDID
[ 72444.150] (II) NVIDIA(GPU-0):     3440 x 1440 @ 144 Hz
[ 72444.150] (II) NVIDIA(GPU-0):       Pixel Clock      : 799.75 MHz
[ 72444.150] (II) NVIDIA(GPU-0):       HRes, HSyncStart : 3440, 3488
[ 72444.150] (II) NVIDIA(GPU-0):       HSyncEnd, HTotal : 3520, 3680
[ 72444.150] (II) NVIDIA(GPU-0):       VRes, VSyncStart : 1440, 1443
[ 72444.150] (II) NVIDIA(GPU-0):       VSyncEnd, VTotal : 1453, 1510
[ 72444.150] (II) NVIDIA(GPU-0):       Sync Polarity    : +H -V
[ 72444.150] (II) NVIDIA(GPU-0):     The EDID for LG Electronics LG ULTRAGEAR (DP-4)
[ 72444.150] (II) NVIDIA(GPU-0):     contradicts itself: mode "3440x1440" is specified in the
[ 72444.150] (II) NVIDIA(GPU-0):     EDID; however, the EDID's valid HorizSync range (250.000
[ 72444.150] (II) NVIDIA(GPU-0):     kHz) would exclude this mode's HorizSync (217.323 kHz);
[ 72444.150] (II) NVIDIA(GPU-0):     ignoring HorizSync check for mode "3440x1440".
[ 72444.150] (II) NVIDIA(GPU-0):     DualHead Mode: No
[ 72444.150] (II) NVIDIA(GPU-0):     Viewport                 3440x1440+0+0
[ 72444.150] (II) NVIDIA(GPU-0):       Horizontal Taps        1
[ 72444.150] (II) NVIDIA(GPU-0):       Vertical Taps          1
[ 72444.150] (II) NVIDIA(GPU-0):     Mode "3440x1440_144" is valid.
[ 72444.150] (II) NVIDIA(GPU-0):
[ 72444.150] (WW) NVIDIA(GPU-0):   Validating Mode "3440x1440_160":
[ 72444.150] (WW) NVIDIA(GPU-0):     Mode Source: EDID
[ 72444.150] (WW) NVIDIA(GPU-0):     3440 x 1440 @ 160 Hz
[ 72444.150] (WW) NVIDIA(GPU-0):       Pixel Clock      : 889.09 MHz
[ 72444.150] (WW) NVIDIA(GPU-0):       HRes, HSyncStart : 3440, 3584
[ 72444.150] (WW) NVIDIA(GPU-0):       HSyncEnd, HTotal : 3616, 3680
[ 72444.150] (WW) NVIDIA(GPU-0):       VRes, VSyncStart : 1440, 1443
[ 72444.150] (WW) NVIDIA(GPU-0):       VSyncEnd, VTotal : 1453, 1510
[ 72444.150] (WW) NVIDIA(GPU-0):       Sync Polarity    : +H -V
[ 72444.150] (WW) NVIDIA(GPU-0):     The EDID for LG Electronics LG ULTRAGEAR (DP-4)
[ 72444.150] (WW) NVIDIA(GPU-0):     contradicts itself: mode "3440x1440" is specified in the
[ 72444.151] (WW) NVIDIA(GPU-0):     EDID; however, the EDID's valid HorizSync range (250.000
[ 72444.151] (WW) NVIDIA(GPU-0):     kHz) would exclude this mode's HorizSync (241.601 kHz);
[ 72444.151] (WW) NVIDIA(GPU-0):     ignoring HorizSync check for mode "3440x1440".
[ 72444.151] (WW) NVIDIA(GPU-0):     ViewPort 3440x1440 exceeds hardware capabilities.
[ 72444.151] (WW) NVIDIA(GPU-0):     Mode is rejected: GPU extended capability check failed.
[ 72444.151] (WW) NVIDIA(GPU-0):     GPU extended capability check failed.
[ 72444.151] (WW) NVIDIA(GPU-0):     Mode "3440x1440_160" is invalid.

So far I have tried setting ModeValidation=NoExtendedGpuCapabilitiesCheck but this just causes X to fail to start.

[crimist]

Seems like I am not the only person with such issues. Other Arch users have the same issue with 3840x2160@160Hz here and here.

[crimist]

Adding my nvidia-bug-report.log.gz.

[zetten]

For what it’s worth, I suspect this is a regression in the latest driver. I’m trying to drive mode 3840x2160_144 from a 1080, and was successfully doing so with ModeValidation=NoExtendedGpuCapabilitiesCheck.

With the 530.41.03 driver all was fine, but since upgrading to 535.54.03 that option simply causes X to fail, regardless of the configured mode - even modes well within hardware capabilities report:

[   540.393] (WW) NVIDIA(GPU-0):     GPU extended capability check failed.
[   540.393] (WW) NVIDIA(GPU-0):     Mode "1920x1080_60" is invalid.

Removing NoExtendedGpuCapabilitiesCheck lets X start up, but the best mode it can reach is 3840x2160_60.

[crimist]

Confirming this is a new regression on 535.54.03 - thanks for the suggestion.

After downgrading to 530.41.03 via nvidia-dkms ModeValidation=NoExtendedGpuCapabilitiesCheck no longer crashes and allows me to set my monitor to 160Hz with no issues thus far.

Seems like there’s 2 issues going on:

1. The Linux drivers fail to validate high bandwidth modes despite the cards supporting them. ModeValidation=NoExtendedGpuCapabilitiesCheck can be used to bypass the check and “fix” the issue.

2. A regression in 535.54.03 causes ModeValidation=NoExtendedGpuCapabilitiesCheck to fail to validate any modes, therefore crashing the X server.

[JimKib]

I am seeing this with my GTX 970. 525.116.04 could do 2560x1440@144Hz, but 535.54.03 can only do 120Hz. I didn’t use a driver between those versions, so have no experience with NoExtendedGpuCapabilitiesCheck, but I did try playing with the new MaxOneHardwareHead, with no luck.

[crimist]

@amrits I believe we’ve found a clear and reproducible regression. Would it be possible to confirm with a developer and file an internal bug?

[birdie]

Try filing a bug report here:

Issues · NVIDIA/open-gpu-kernel-modules

[crimist]

Unfortunately I cannot use nvidia-open as I have a Pascal card. If someone with a Turing or later card can confirm this affects nvidia-open then I’d be happy to put a report in.

[crimist]

Seems like another users report finally got Nvidia’s attention on this issue: Linux 535.xx does not honor ModeValidation, making headless Fullscreen RANDR usage with ConnectedMonitor impossible

[crimist]

Fixed in NVIDIA driver 545.29.02.