4 ms·

LXC unprivileged containers are actually pretty secure by design. The Canonical LXC page does a pretty great job of explaining why: https://linuxcontainers.org/

LXC unprivileged containers are actually pretty secure by design. The Canonical LXC page does a pretty great job of explaining why: https://linuxcontainers.org/lxc/security/ https://linuxcontainers.org/lxc/security/

dgacmu 9y ago
Please (OP, and others allowing access to your machines), be cautious on this front. While LXC unprivileged containers are a start at isolation, LXC containers + GPU passthrough has a much, much larger attack surface (the nvidia binary blob drivers are complex and out of your control). The most common ways of giving GPU access to a container involve installing the CUDA drivers in the host and simply allowing the container to access them. Depending on your threat model, this may be borderline OK, or may be insanely high risk. We've seen driver exploits in the recent past: http://nvidia.custhelp.com/app/answers/detail/a_id/3140/~/security-bulletin%3A-nvidia-unix-graphics-driver-exploit-advisory http://nvidia.custhelp.com/app/answers/detail/a_id/3140/~/se.... that can vector through the GPU to gain access to arbitrary host memory (aka, you're in deep trouble). The only "right" way to give access to your GPUs requires IOMMU support and running in a true guest VM, which Nvidia's consumer cards and drivers are explicitly prevented from doing (because they'd like you to buy the Tesla versions, thanks). This may be a totally acceptable risk on an isolated mining rig, but people should be aware of the heightened risks they're exposing themselves to. Breaking your machine is well within the capabilities of state-sponsored actors on this one, and from time to time (just after vulnerabilities get announced), could be decently within advanced-script-kiddie zone.
- Samin100 9y ago
  This is a very, very valid point. I'm going to mention this on the website & even advise people to run the client on an isolated machine (instead of their daily driver gaming rig). Thank you for pointing this out. Also, one of my friends working on this project is a sophomore in the CS dept at CMU, and given your interest in distributed systems & DL, would it be possible to meet up for a couple of minutes and discuss security a bit more in depth? (if yes, I'll shoot you an email)
  - dgacmu 9y ago
    Happy to. Drop me a note.
    - rememberlenny 9y ago
      ^^^ This is why HN is amazing.

dgacmu 9y ago

Please (OP, and others allowing access to your machines), be cautious on this front. While LXC unprivileged containers are a start at isolation, LXC containers + GPU passthrough has a much, much larger attack surface (the nvidia binary blob drivers are complex and out of your control). The most common ways of giving GPU access to a container involve installing the CUDA drivers in the host and simply allowing the container to access them. Depending on your threat model, this may be borderline OK, or may be insanely high risk. We've seen driver exploits in the recent past: http://nvidia.custhelp.com/app/answers/detail/a_id/3140/~/security-bulletin%3A-nvidia-unix-graphics-driver-exploit-advisory http://nvidia.custhelp.com/app/answers/detail/a_id/3140/~/se.... that can vector through the GPU to gain access to arbitrary host memory (aka, you're in deep trouble). The only "right" way to give access to your GPUs requires IOMMU support and running in a true guest VM, which Nvidia's consumer cards and drivers are explicitly prevented from doing (because they'd like you to buy the Tesla versions, thanks). This may be a totally acceptable risk on an isolated mining rig, but people should be aware of the heightened risks they're exposing themselves to. Breaking your machine is well within the capabilities of state-sponsored actors on this one, and from time to time (just after vulnerabilities get announced), could be decently within advanced-script-kiddie zone.

Samin100 9y ago
This is a very, very valid point. I'm going to mention this on the website & even advise people to run the client on an isolated machine (instead of their daily driver gaming rig). Thank you for pointing this out. Also, one of my friends working on this project is a sophomore in the CS dept at CMU, and given your interest in distributed systems & DL, would it be possible to meet up for a couple of minutes and discuss security a bit more in depth? (if yes, I'll shoot you an email)
- dgacmu 9y ago
  Happy to. Drop me a note.
  - rememberlenny 9y ago
    ^^^ This is why HN is amazing.