1

Back in December I downloaded and burned an Ubuntu install CD that had kernel 4.13.0-19-generic. I built two minimal servers and they worked fine with nVidia CUDA.RUN package. I recently tried building a 3rd system openssh server using the same CD. I see kernel ...19 go in but after the download from the US repository I end up with kernel ...32 which has a bug documented on the nvidia developer forum that does not allow the kernel to be built. I assume they will have a fix for this eventually but in the meantime I would like to install ...19 or any version that still works with the nVidia cuda run package.

I retried the install from my old CD but there was no option to select the version of the package to download. I see 19 come off the CD but end up with 32.

How do I restrict the install to 19?

Maybe I can revert back to 19 after getting 32 running?

Alternately, I could clone the flash drive from one of the other systems. I did try that with acronis clone tool under windows 10, but it didnt boot. Maybe there is a better cloning tool?

Thank you for looking at this.

[SOLVED WITH HELP FROM MIKEWHATEVER thanks mike!]

sudo apt-get install linux-image-4.13.0-19-generic

(reboot and select 19)

sudo apt-get purge linux-image-4.13.0-32-generic
sudo apt-get purge linux-headers-4.13.0.32
sudo apt-get install linux-headers-$(uname -r)
sudo apt-get install libxi-dev libxmu-dev freeglut3-dev build-essential binutils-gold
sudo sh ./cuda384.run --override

results as follows:

27-Jan-2018 12:24:55 [---] CUDA: NVIDIA GPU 0: GeForce GTX 670 (driver version 384.81, CUDA version 9.0, compute capability 3.0, 1994MB, 1874MB available, 2704 GFLOPS peak)
27-Jan-2018 12:24:55 [---] OpenCL: NVIDIA GPU 0: GeForce GTX 670 (driver version 384.81, device version OpenCL 1.2 CUDA, 1994MB, 1874MB available, 2704 GFLOPS peak)

I thought I had to use update-initramfs -u after the CUDA install. That put 32 into the boot which caused a problem. It was not necessary and forced a re-install to recover.

I went back to 19 because it worked with the nVidia CUDA kit and nVidia has not released a fix yet for having their source headers out of order when being built into the 32 kernel.

[EDIT] Strange - I see where 5 gurus here report this question is "an exact duplicate of an existing question". I went there and looked and that question was how to prevent an upgrade which is nice if you know the upgrade will cause a problem. This question was how to recover from an upgrade. I actually spent some time googling for how to restore a previous version. I found nothing informative until I came here and was told to use "apt-get install linux_image...". Want to ask the gurus here, what good is it to install 32 and get protection from meltdown and specter when there are no known exploits of those defects and the application you are required to run requires a package that does not work under 32?

Zanna
  • 70,465
Joseph Stateson
  • 143
  • The 4.13.0.-32-generic is indeed the latest kernel in Ubuntu 17.10. It has important fixes for Meltdown and Spectre. I don't know what you mean by "not allow the kernel to be built", and why you need to build the kernel. It works just fine with nvidia cards/driver here. Anyway, if you must use an older kernel, try installing it with sudo apt-get install linux-image-4.13.0-19-generic. – mikewhatever Jan 27 '18 at 15:03
  • Executing the cuda.run package rebuilds the kernel. The compile fails because the headers in the nVidia source are out of order for the 32 package. nVidia uBuntu specifically says that gcc must be installed along with build-essential and kernel headers. I will try your suggestion. AFAICT those defects truely be fixed on in the bios or chip replacement and my systems are old and extremely unlikely to be updated by the manufacturer. – Joseph Stateson Jan 27 '18 at 15:12
  • The nVidia drivers work fine. Any type of crypto-currency mining (or co-processor work) requires CUDA, not just an nVidia graphics board. I installed the microsoft fix for those two bugs but when I ran the diagnostics to verify I got the message that I needed a bios fix to correct the problem on my Dell xps730. That is an old system and Dell will never release a bios upgrade for it nor swap out my CPUs. – Joseph Stateson Jan 27 '18 at 15:33
  • Bios updates are out of the question for old system, so there is no point talking about it. The only possible mitigation for old PCs is at the OS level. – mikewhatever Jan 27 '18 at 15:36
  • Your suggestion seems to be working and I am on my way to install CUDA. However, I noticed that I have "-1" vote. What is the significance of that? I have no control over how nVidia creates their CUDA installation package and can only wait for them to fix the problem. I will post the solution details once I get CUDA working. – Joseph Stateson Jan 27 '18 at 16:14
  • I did the following and everything went OK even the compile and installation. sudo apt-get install linux-image-4.13.0-19-generic (reboot and select 19) sudo apt-get install linux-headers-$(uname -r) sudo apt-get install libxi-dev libxmu-dev freeglut3-dev build-essential binutils-gold sudo sh ./cuda384.run --override sudo update-initramfs -u. The last step installed initrd.img-4.13.0.32 instead of 19. I cannot boot and holding down shift does not help. I will have to use CD to recover. Nouveau driver is blacklisted with last step as I understand and that needs to be done – Joseph Stateson Jan 27 '18 at 17:00
  • Worked, I credited you – Joseph Stateson Jan 27 '18 at 18:39
  • Your recent edit seems to indicate a new question (at least to me). The original question looks to be adequately answered by the called duplicate. – Elder Geek Jan 28 '18 at 15:29
  • Clonezilla is an excellent cloning program and linux based. The current kernel only provides meltdown protection. If you do not have an intel CPU the module is not activated, thus the updated kernel isn't necessary. – ravery Jan 28 '18 at 17:46

0 Answers0