3

On my dev system (Kubuntu 20.04) I just installed CUDA 11.3. After checking the installation, I found some older CUDA packages installed, which are remainders from experiments some time ago. That CUDA 11.2 installation was incomplete, so I needed to remove the packages one-by-one (not via the "cuda" metapackage). I was able remove all of them except one: cuda-cudart-11-2.

When running sudo apt remove cuda-cudart-11-2 or sudo apt --purge remove cuda-cudart-11-2, apt names two new packages that need to be installed before removing cuda-cudart-11-2:

  • cuda-toolkit-11-config-common
  • cuda-toolkit-config-common

My first question: How can removing a package require installing new packages? I have never seen this before.

I confirmed the installation of those packages as proposed by apt and got errors:

me@mypc:~$ sudo apt remove cuda-cudart-11-2
Reading package lists... Done
Building dependency tree       
Reading state information... Done
The following additional packages will be installed:
  cuda-toolkit-11-config-common cuda-toolkit-config-common
The following packages will be REMOVED:
  cuda-cudart-11-2
The following NEW packages will be installed:
  cuda-toolkit-11-config-common cuda-toolkit-config-common
0 upgraded, 2 newly installed, 1 to remove and 0 not upgraded.
Need to get 0 B/32.2 kB of archives.
After this operation, 514 kB disk space will be freed.
Do you want to continue? [Y/n] y
(Reading database ... 380817 files and directories currently installed.)
Preparing to unpack .../cuda-toolkit-11-config-common_11.3.109-1_all.deb ...
Unpacking cuda-toolkit-11-config-common (11.3.109-1) ...
dpkg: error processing archive /var/cache/apt/archives/cuda-toolkit-11-config-common_11.3.109-1_all.deb (--unpack):
 trying to overwrite '/etc/ld.so.conf.d/989_cuda-11.conf', which is also in package cuda-cudart-11-2 11.2.152-1
Preparing to unpack .../cuda-toolkit-config-common_11.3.109-1_all.deb ...
Unpacking cuda-toolkit-config-common (11.3.109-1) ...
dpkg: error processing archive /var/cache/apt/archives/cuda-toolkit-config-common_11.3.109-1_all.deb (--unpack):
 trying to overwrite '/etc/ld.so.conf.d/000_cuda.conf', which is also in package cuda-cudart-11-2 11.2.152-1
Errors were encountered while processing:
 /var/cache/apt/archives/cuda-toolkit-11-config-common_11.3.109-1_all.deb
 /var/cache/apt/archives/cuda-toolkit-config-common_11.3.109-1_all.deb
E: Sub-process /usr/bin/dpkg returned an error code (1)

As far as I understand, this says that cuda-toolkit-11-config-common can not be installed because the config file /etc/ld.so.conf.d/989_cuda-11.conf can not be overwritten because it is in use by cuda-cudart-11-2 (which can not be removed until cuda-toolkit-11-config-common is installed). This sounds like a kind of deadlock to me.

My seconds question: If this is actually a deadlock, how can I solve it?

Silicomancer
  • 583
  • 2
  • 10
  • 33
  • Just do the process and try to remove cuda-toolkit-11-config-common after. Maybe it will resolve the dependency problem and you'll be able to remove CUDA. If it doesn't work then you are in a deadlock :/ – Jeryosh Jun 14 '21 at 19:30
  • @Jerare I can't install cuda-toolkit-11-config-common because it wants to overwrite a configuration file that is in use. Is my English unclear on this? – Silicomancer Jun 14 '21 at 19:56
  • 1
    Could you try the solution in https://askubuntu.com/q/56761/26246 .About you question, it is possible that apt was already having a stuck packages, and they will show up whenever you launch apt. Also if you have used aptitude or synaptic, if there are packages that are manually installed and depend on that package. aptitude and synaptic may look for alternatives to not break dependencies. – user.dz Jun 15 '21 at 20:58
  • 1
    @user.dz

    Never used aptitude or synaptic.

    If those are stuck packages, apt should offer them anytime, even when installing something totally unrelated? I tried to install some random application package and apt did not mention the cuda-toolkit*-config-common packages.

     – Silicomancer Jun 15 '21 at 23:19
  • 1
    @user.dz The linked solution kind of worked. I was able to force-install the proposed packages and could remove the old package immediately after. The config files have not been update but I guess they are still valid. – Silicomancer Jun 16 '21 at 21:07
  • Still my first question is not answered – Silicomancer Jun 16 '21 at 21:07
  • @Silicomancer they are dependencies of cuda as per https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/Packages , it just happen the installation hadn't complete in first place. When 1st time did you notice that APT has issue, have you ever run apt for another reason like update after installing cuda? – user.dz Jun 16 '21 at 21:24
  • @user.dz I'm pretty sure I didn't do anyhing else after installing cuda package. I installed it, then noticed the version I got was newer than expected (11.3 instead of 11.2), then I listed all packages containing "cuda", found old 11.2 packages and removed them (well, most of them). I didn't notice an error during cuda installation. Is there some log in linux that contains package installation runs? – Silicomancer Jun 16 '21 at 21:46
  • Yes, /var/log/dpkg.log* – user.dz Jun 16 '21 at 21:49
  • Had a look in the log. I don't get it. I don't see anything that looks like a failure, even for the failed removales nor for the failed installs. – Silicomancer Jun 18 '21 at 20:46

0 Answers0