Arch Linux

WeaponMan

Member

Registered: 2018-02-13

Posts: 3

Re: Terrible performance regression with Nvidia 390.25 driver

I don't know if I am affected.
I have swapped HD 6750 to GT 710 + Intel HD on last friday. (Needed to fit GPU to small case)
Both cards have 1GB of VRAM.
Currently running 2 monitors on GT 710 and one on Intel HD.
All 3 monitors have resolution of 1920x1200 at 59.95Hz.
GT 710 is running as main card and Intel HD is Prime offload. (xorg.conf is empty file and xorg.conf.d has only one config with keyboard setup stuff).
I have observed tearing in chrome when scrolling and when I open new tab with some webpage chrome just freeze for a second.
Sometimes second screen on GT 710 start flashing black for second.

On HD 6750 was everything fine...

For specs:
Asus B85M-G
Intel i5-4460
32GB RAM
Kingston SSD 120GB
Nvidia GT 710 1GB VRAM (Vendor EVGA)
DE: GNOME Shell on Xorg

I have running system at home with GTX 1080 and I am little bit scared to update the driver.

EDIT: typo

Last edited by WeaponMan (2018-02-13 16:31:44)

Tom B

Member

Registered: 2014-01-15

Posts: 187

Website

Re: Terrible performance regression with Nvidia 390.25 driver

@WeaponMan. Tearing is a separate issue, check the nvidia troubleshooting page on the wiki. However, the Chrome scrolling issue sounds like the one people here are having if you're on 390.25.

WeaponMan

Member

Registered: 2018-02-13

Posts: 3

Re: Terrible performance regression with Nvidia 390.25 driver

Just updated my system at home and I happy to say it is Unaffected.

Specs:
ASUS PRIME X370-Pro
AMD Ryzen 7 1700
16GB RAM
Nvidia GTX 1080 (Vendor EVGA)
1 monitor 2560x1440 at 59.95Hz on DisplayPort
DE: GNOME Shell
CMDLINE: nvidia-drm.modeset=1

On the other system I downgraded the driver but system is remote (at work) so tomorrow I will find out if it helps or not.

Last edited by WeaponMan (2018-02-13 17:27:01)

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

@kokoko3k blispx is stuck here with https://git.kernel.org/pub/scm/linux/ke … 15.3#n1864 with modesetting enabled
so the issue there appears to be the previous flip has not occurred but another frame is ready and it times out waiting on the old flip.
Perhaps something similar is happening in the none modesetting path for all the others affected.

kokoko3k

Member

Registered: 2008-11-14

Posts: 2,421

Re: Terrible performance regression with Nvidia 390.25 driver

loqs, the first post refers to a specific problem with images, so i think textures and memory allocation could be part of the issue, because flipping should occour even with no images at all, right?

Seriously slow image rendering and stuttering while scrolling in chromium, with the website window freezing for up to 3 seconds when loading a larger image (noticeable on larger wikipedia articles which would scroll smoothly with 387, now every time an image comes into view there is bad stutter).

---

Help me to improve ssh-rdp !
Retroarch User? Try my koko-aio shader !

blispx

Member

Registered: 2017-11-29

Posts: 53

Re: Terrible performance regression with Nvidia 390.25 driver

I will check on HDMI later if there is a difference between DisplayPort

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

@kokoko3k that seems perfectly logical to me but somehow on the modesetting path flipping is stalling or at least that is my interpretation of the dmesg output.
@blispx could you try 390.25 with modesetting under linux-lts the patch for 4.15 patch is versioned so it will not have any effect on 4.14 I want to see if with the older 4.14 modesettng code if the issue still occurs.
Edit:
Ideally some use of kernel performance analysis (perf / bpf ?) would be able to detect a call in 390.25 taking much more time but that is outside my expertise.

Last edited by loqs (2018-02-13 19:58:02)

Roken

Member

From: South Wales, UK

Registered: 2012-01-16

Posts: 1,281

Re: Terrible performance regression with Nvidia 390.25 driver

This is just one of a sudden uplift in NV 390. Would it not be worth merging them all?

---

Ryzen 5900X 12 core/24 thread - RTX 3090 FE 24 Gb, Asus Prime B450 Plus, 32Gb Corsair DDR4, Cooler Master N300 chassis, 5 HD (1 NvME PCI, 4SSD) + 1 x optical.
Linux user #545703

/ is the root of all problems.

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

Roken which of the other nvidia 390.25 threads has the "Failed to initialize the GLX module" as the error message?

ewaller

Administrator

From: Pasadena, CA

Registered: 2009-07-13

Posts: 20,208

Re: Terrible performance regression with Nvidia 390.25 driver

This is just one of a sudden uplift in NV 390. Would it not be worth merging them all?

Done.  Let us know if there are others that should be merged as well.

---

Nothing is too wonderful to be true, if it be consistent with the laws of nature -- Michael Faraday
Sometimes it is the people no one can imagine anything of who do the things no one can imagine. -- Alan Turing
---
How to Ask Questions the Smart Way

Roken

Member

From: South Wales, UK

Registered: 2012-01-16

Posts: 1,281

Re: Terrible performance regression with Nvidia 390.25 driver

@ loqs:I meant generally there are many issues with NV 390 suddenly cropping up. As someone has mentioned, the drivers should have probably not been released in their current state. Not the Arch maintainer's problem, but NVidia's.

Last edited by Roken (2018-02-13 21:20:21)

---

Ryzen 5900X 12 core/24 thread - RTX 3090 FE 24 Gb, Asus Prime B450 Plus, 32Gb Corsair DDR4, Cooler Master N300 chassis, 5 HD (1 NvME PCI, 4SSD) + 1 x optical.
Linux user #545703

/ is the root of all problems.

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

I wonder how merging that particular thread will be beneficial though as it appeared to be a distinct issue.  The other performance issues covered in this thread need a functioning nvidia libgl.
Edit:
Perhaps comparing the output between 387.34 and 390.25 of

# perf top -U

Last edited by loqs (2018-02-13 23:14:53)

WeaponMan

Member

Registered: 2018-02-13

Posts: 3

Re: Terrible performance regression with Nvidia 390.25 driver

Yep downgrade helped on my system at work so I am affected.
I suppect is something with VRAM swapping becauce card was at 800 - 960 MB on 390.25 then start flickering on second screen with freeze for a second.
On 387.34-23 is almost same but the swapping is not noticable (no freeze, no screen flickering).

Tom B

Member

Registered: 2014-01-15

Posts: 187

Website

Re: Terrible performance regression with Nvidia 390.25 driver

I just installed nvidia-390.25-10, unfortunately, no difference.

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

I just installed nvidia-390.25-10, unfortunately, no difference.

nvidia-390.25-10 was just a rebuild against 4.15.3

Guiluge

Member

Registered: 2016-04-12

Posts: 9

Re: Terrible performance regression with Nvidia 390.25 driver

Performance regression also confirmed on my setup, with nvidia 390.25-10.
Graphic Card : GeForce GTX 980
Desktop : Plasma 5

Until nvidia fixes (again) this annoying bug, I've switched to nouveau driver : smooth and stable for now !

Batou

Member

Registered: 2017-01-03

Posts: 259

Re: Terrible performance regression with Nvidia 390.25 driver

Just tried 4.15.3 from the repo:

Linux batou 4.15.3-1-ARCH #1 SMP PREEMPT Mon Feb 12 23:01:17 UTC 2018 x86_64 GNU/Linux

and:

nvidia 390.25-10
nvidia-lts 1:390.25-2
nvidia-settings 390.25-1
nvidia-utils 390.25-2
opencl-nvidia 390.25-2

My Xorg will not start Cinnamon and all I get is a black screen and a mouse pointer. This is consistent with the behavior I saw when I tried 4.15.1.

Only when I removed 'force (full) composition pipeline from the Xorg conf file would me allow to start cinnamon:

Option         "metamodes" "nvidia-auto-select +0+0 {ForceCompositionPipeline=On}" 
#this won't work either:
Option         "metamodes" "nvidia-auto-select +0+0 {ForceCompositionPipeline=On, ForceFullCompositionPipeline=On}"

I also removed these lines since they go hand-in-hand with composition pipeline:

Option         "TripleBuffer" "on"
Option         "AllowIndirectGLXProtocol" "off"
 

When I had these enabled, I saw this in dmesg:

[    0.497872] nvidia: loading out-of-tree module taints kernel.
[    0.497879] nvidia: module license 'NVIDIA' taints kernel.
[    0.497879] Disabling lock debugging due to kernel taint
[    0.508074] nvidia-nvlink: Nvlink Core is being initialized, major device number 244
[    0.508295] nvidia 0000:01:00.0: vgaarb: changed VGA decodes: olddecodes=io+mem,decodes=none:owns=io+mem
[    0.508354] NVRM: loading NVIDIA UNIX x86_64 Kernel Module  390.25  Wed Jan 24 20:02:43 PST 2018 (using threaded interrupts)
[    0.509938] nvidia-modeset: Loading NVIDIA Kernel Mode Setting Driver for UNIX platforms  390.25  Wed Jan 24 19:29:37 PST 2018
[    0.519754] nvidia-uvm: Loaded the UVM driver in 8 mode, major device number 243
[    0.520315] Linux agpgart interface v0.103
[    0.531802] [drm] [nvidia-drm] [GPU ID 0x00000100] Loading driver
[    0.901965] resource sanity check: requesting [mem 0x000c0000-0x000fffff], which spans more than PCI Bus 0000:00 [mem 0x000d0000-0x000d3fff window]
[    0.902104] caller _nv001170rm+0xe3/0x1d0 [nvidia] mapping multiple BARs
[    0.954484] resource sanity check: requesting [mem 0x000e0000-0x000fffff], which spans more than PCI Bus 0000:00 [mem 0x000e0000-0x000e3fff window]
[    0.954604] caller _nv029921rm+0x57/0x90 [nvidia] mapping multiple BARs
[    1.249522] nvidia-modeset: Allocated GPU:0 (GPU-85aca018-59c5-d11a-30b6-55a83246689c) @ PCI:0000:01:00.0
[    1.274716] [drm] Supports vblank timestamp caching Rev 2 (21.10.2013).
[    1.274717] [drm] No driver support for vblank timestamp query. 

This BARs stuff I have no idea what it is.

I've also had to disable mdm and start X manually.
Edit: I've replaced mdm with LightDM and Xorg is starting fine now.
I also haven't noticed many screentears.

Last edited by Batou (2018-02-14 13:22:46)

---

Please vote for all the AUR packages you're using. You can mass-vote for all of them by doing: "pacman -Qqm | xargs aurvote -v" (make sure to run "aurvote --configure"  first)

blispx

Member

Registered: 2017-11-29

Posts: 53

Re: Terrible performance regression with Nvidia 390.25 driver

\
@blispx could you try 390.25 with modesetting under linux-lts the patch for 4.15 patch is versioned so it will not have any effect on 4.14 I want to see if with the older 4.14 modesettng code if the issue still occurs.

I have built nvidia 390.25-10 without 4.15-FS57305.patch and KMS works

Coming back to 4.14 I do not have it now

Switching from DP to HDMI did not give anything

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

I have built nvidia 390.25-10 without 4.15-FS57305.patch and KMS works

When using nvidia 390.25-10 without 4.15-FS57305.patch what is the output of

$ lsmod | grep drm

Batou

Member

Registered: 2017-01-03

Posts: 259

Re: Terrible performance regression with Nvidia 390.25 driver

When I open a large image in Chromium, my whole machine now slows down. I have mpd playing music in the background and it starts stuttering too. I think this is related to the fact that I have Chromium set up to use the GPU.

What's causing this? Do kernel devs even know this is an issue with Nvidia? Is there a kernel.org bug for this that we can check? I find it strange that it's 4.15.3 already and it hasn't been fixed in the past 3 releases so far.

I'm probably going to roll back to 4.14 again if it becomes unbearable but I'm gonna stick with 4.15 for at least a day.

---

Please vote for all the AUR packages you're using. You can mass-vote for all of them by doing: "pacman -Qqm | xargs aurvote -v" (make sure to run "aurvote --configure"  first)

deafeningsylence

Member

Registered: 2016-09-23

Posts: 52

Re: Terrible performance regression with Nvidia 390.25 driver

Just wanted to add my two cents of evidence, I am affected by very laggy video or rather buffering every few seconds, already excluded download or hardware performance as a cause.

I have a DP and a HDMI screen, the latter runs on 60hz and the former on 144hz, both show the same problem, neither is affected from screen freezing or stuttery scrolling though. Only during Netflix streaming some tearing appears.

Most noticeably is this problem on Twitch, much less to not present on YouTube.

I also checked streaming twitch via streamlink into different players to exclude browsers from the equation. Here, both mpv and vlc player show the same stuttering.

So much for my evidence.

loqs

Member

Registered: 2014-03-06

Posts: 18,055

Re: Terrible performance regression with Nvidia 390.25 driver

@Batou upstream would be the nvidia linux forum.  The nvidia modules are not supplied and hence not supported by kernel.org.

Batou

Member

Registered: 2017-01-03

Posts: 259

Re: Terrible performance regression with Nvidia 390.25 driver

@Ioqs thanks. I checked out Nvidia's forum and it has a bunch of threads about 390.25... including this one, which also links to this thread: https://devtalk.nvidia.com/default/topi … on-390-25/ Unfortunately, no comments from Nvidia employees.
I guess we're stuck with an old kernel (or terrible perf with new one) until Nvidia fixes this. I doubt we can do much more ourselves. This is all inside of the binary blob and only Nvidia can fix it. Complaining to Nvidia might help.

---

Please vote for all the AUR packages you're using. You can mass-vote for all of them by doing: "pacman -Qqm | xargs aurvote -v" (make sure to run "aurvote --configure"  first)

Batou

Member

Registered: 2017-01-03

Posts: 259

Re: Terrible performance regression with Nvidia 390.25 driver

@Ioqs is your patch from this post: https://bugs.archlinux.org/task/57305?getfile=16124 applied to 4.15.3 from the repo or do we still have to compile it manually?

---

Please vote for all the AUR packages you're using. You can mass-vote for all of them by doing: "pacman -Qqm | xargs aurvote -v" (make sure to run "aurvote --configure"  first)

deafeningsylence

Member

Registered: 2016-09-23

Posts: 52

Re: Terrible performance regression with Nvidia 390.25 driver

Quick update, changing to nouveau drivers did not solve the issue, it is the same lag/buffering every 2 seconds present as before for me. I checked via lspci -v if the driver is in use and it is.