with nvidia config and latest nvidia-325.15-5, X does not start

[jantman]

upgrade to latest nvidia, after reboot get only a black screen. console in (boot with "" kernel option) and Xorg.log.0 shows:

[ 12.441] (EE) NVIDIA(GPU-0): Failed to tear down EVO channel
...
[ 16.440] (EE) NVIDIA(GPU-0): EVO Push buffer channel allocation failed
[ 16.450] (EE) NVIDIA(GPU-0): Failed to allocate EVO core DMA push buffer
[ 16.452] (EE)
Fatal server error:
[ 16.452] (EE) AddScreen/ScreenInit failed for driver 0
...
[ 16.461] (EE) Server terminated with error (1). Closing log file.

[jantman]

I'm already using

    Option         "UseDPLib" "off"

adding

    Option         "ProbeAllGpus" "False"

didn't help.

This may be related to an upstream but that's being discussed a fair bit online, with the latest nvidia 325 not working with 3.10 kernels. Some success has been reported downgrading to a 3.9 kernel and nvidia 319.

[jantman]

Something at least very similar is being tracked on the Arch forums: https://bbs.archlinux.org/viewtopic.php?id=167195

[jantman]

This is an upstream bug in the nvidia driver, specifically with the 3.10 kernel series.

For the time being, my solution (which gave me a pretty quick, working system) was to downgrade to my previous 3.9.8-1 kernel and the matching nvidia driver, which were still cached locally on my system. The specific command (for the versions I was previously running):

cd /var/cache/pacman/pkg
pacman -U linux-3.9.8-1-x86_64.pkg.tar.xz linux-headers-3.9.8-1-x86_64.pkg.tar.xz nvidia-319.32-2-x86_64.pkg.tar.xz nvidia-utils-319.32-1-x86_64.pkg.tar.xz nvidia-libgl-319.32-1-x86_64.pkg.tar.xz virtualbox-4.2.14-1-x86_64.pkg.tar.xz virtualbox-guest-iso-4.2.14-1-any.pkg.tar.xz virtualbox-host-modules-4.2.14-2-x86_64.pkg.tar.xz

[tazjin]

Did you ever find a proper solution to this? I'm still unable to get the nvidia drivers to run using kernel 3.12.9 and I've tried all tricks and hacks found on the internet. I don't really want to go back to an older version of the kernel so I'm currently stuck with the nouveau driver (on which I can't get background brightness control to work)

[jantman]

@tazjin I haven't attempted a kernel upgrade in almost 3 months because of this (last time I tried nouveau I had serious issues trying to hot-plug to my 30" Dell monitor at work, so I gave up on it).

I'd been following a thread on the Nvidia forum: https://devtalk.nvidia.com/default/topic/573252/evo-push-buffer-channel-allocation-failed-is-back-as-usedplib-false-no-more-supported-with-325-15/#3999262 that seemed to be this exact issue. But it looks like the last reply from an nvidia person was 2 months ago asking people to reproduce the problem and submit bug reports... which 3 people seem to have done, and no update from nvidia since then.

All that being said, the last time I tried this was with nvidia 331.17 and kernel 3.11.6, and there's a 331.38 out now, which does seem to include some X-related fixes: http://www.nvidia.com/Download/driverResults.aspx/72250/en-us

[hh]

I tested with 331.38 and the current 3.12.x kernel and still get the same EVO Push buffer channel allocation failed error.

From https://wiki.archlinux.org/index.php/MacBookPro10,x#Graphics

Note** that as of September, 2013 the current nvidia driver (325.15-5)
does not work with the current 3.10 series kernels; X will die with an
error about "Failed to allocate EVO core DMA push buffer" and
leave you with a black screen (but able to SSH in to the machine).
Your best current bet is to use a 3.9-series kernel and the older
319.32-series nvidia driver.

You have the luxury of the downgrade, I'm just getting into Arch and don't see an obvious way to install kernel 3.9.8-1 or nvidia 331.17.

Any thoughts on how to install 3.9 series kernels via arch, or should I try and do it manually.

[jantman]

@hh actually, I'm the one who added that note on the wiki from 9/2013 :)

Well I installed the older kernels by manually installing the package from the local pacman package cache. I'd assume there has to be an archive somewhere that you can download the packages from, but it'll certainly be more difficult without the local cache.

At this point I think the best bet is for all of us to sign up for devtalk.nvidia.com accounts, and post our problems to the above-linked thread, in the hopes that with enough renewed interest, the nvidia folks will actually fix this...

[tazjin]

You could check which Kernel versions are available in the Arch Rollback Machine. What I did was build the kernel by getting the PKGBUILD for the appropriate version from git. You'll also need to rebuild linux-headers and xorg-server the same way. It requires a few more dependencies which Pacman will tell you about when installing stuff.

Also remember to blacklist the packages you changed afterwards so that they don't get updated in a pacman -Syu

[yvesh]

Any updates on this? Just tried the Arch nvidia-beta (334.16) - still not working.

[jantman]

Nope, sorry. I'm going to be doing an arch install on my desktop next week or so, so hopefully I'll revisit this then (once I know I have at least one working Arch system). Unfortunately this laptop is my primary work machine, so it's not too often that I can risk the downtime of trying a full upgrade, seeing if it works, and then rolling back.

The forum thread I've been following (https://devtalk.nvidia.com/default/topic/573252/linux/evo-push-buffer-channel-allocation-failed-is-back-as-usedplib-false-no-more-supported-with-325-15/3/) still appears to just be filled with reports of the same issue and no reply from Nvidia, so I'm assuming it's not fixed...

[hh]

I'm looking at this again and have a bit more time on my hands. Any updates @jantman ?

[krobertson]

@jantman @yvesh @hh was doing a series of updates on my 2012 MBP this weekend and decided to give the latest nvidia drivers a try and found they actually worked. I had been using the 319.32 drivers on the 3.10.10 kernel, and even had to rebuilt an order xorg. Updated today to nvidia 337.12 and linux 3.14.1. I had to rebuild xf86-input-mtrack against the newer xorg, but was surprised when X actually fired up.

I still have the issue where if I exit X and go back to the console, it doesn't show... but I can live with that.

[hh]

@krobertson That's exciting, I have console and nouveau (still haven't gotten X up) but will try your setup with linux 3.14.1 and nvidea 337.12 and a rebuilt xf86-input-mtrack.... how do I rebuild against the newer xorg? git? I had basically given up and started working on a macbook air with 4gig ram and an intel. ;)

[hh]

I'll try and take a look tonight, would be awesome to get this box up. - @hh

[krobertson]

@hh to rebuild mtrack, I just reinstalled it from the AUR

[hh]

I was able to get X up and running! Thanks for the update.

[jantman]

@krobertson @hh Wow, that's absolutely wonderful news, thanks so much!

I actually did a fresh install of Arch on my brand new desktop, and have been planning a very major rewrite of this module to make it a "real" puppet module (on the Forge), and get it to work with my new method which uses r10k to get and deploy modules (https://github.com/jantman/workstation-bootstrap) and split off a lot of the non-MBP-specific stuff into an archlinux-workstation module (https://github.com/jantman/puppet-archlinux-workstation).

Now that it seems like this issue is finally fixed, I can try rebuilding my MBP and getting this to work.

I also have some plans around doing regular automated testing of this by spinning up a vagrant/virtualbox install of the latest Arch and then using beaker to test the Puppet stuff.