Nvidia GPU devices in systemd will be ignored after 1.1.3

[panli889]

I found a PR https://github.com/opencontainers/runc/pull/3504/files in 1.13, in which it will ignore the devices can't be os.Stat. However, Nvidia GPU is that kind of device, though I do not know why.

After the version, If we use the command like this: 'docker run --device /dev/nvidia0:/dev/nvidia0', the DeviceAllow list systemd received from runc will be different from the list runc operates. So if we run systemctl daemon-reload, the device will not be able to access anymore.

So I'm wondering is it possible to revert the PR, or is there any other solutions?

[AkihiroSuda]

@cyphar PTAL

[panli889]

@cyphar Hi, any progress?

[kolyshkin]

@panli889 can you please check and let us know if the fix in PR #3620 also fixes your issue?

[panli889]

@panli889 can you please check and let us know if the fix in PR #3620 also fixes your issue?

Thanks for the reply! I've test it with the code, it reports the same issue.

For the normal nvidia GPU device, the location is at /dev/nvidia0, and in the common.go now it turns to /dev/char/195:0. If I run it with the os.Stat, it still reports no such file or directory.

[kolyshkin]

This should be fixed in runc-1.1.7 (by PR #3845), provided systemd >= v240 is used. For older systemd versions, there's nothing we can do.

@panli889 can you please check that this fixes the issue?

[panli889]

This should be fixed in runc-1.1.7 (by PR #3845), provided systemd >= v240 is used. For older systemd versions, there's nothing we can do.

@panli889 can you please check that this fixes the issue?

Hi @kolyshkin, Thanks for your efforts! It's ok from my vision, our systemd version is >= 240!