Memory layout and vulnerability to side-channel attacks

[alveko]

Q1: Is the memory layout of the guest same as if it runs directly on the host kernel? Specifically, if the underlying host kernel maps all physical memory to virtual address space of each process (the way which made Meltdown possible), will it preserve such mapping in the gVisor container?

Q2: Assuming the underlying host kernel and CPU is vulnerable side-channel attacks (same as or similar to Meltdown/Spectre), would one guest in one gVisor container be able to attack another guest in another gVisor container on the same host? and why so?

[amscanne]

Sorry, somehow I think this issue went unnoticed!

Q1: For the KVM platform, only host user space spaces are available to the guest, so Meltdown would not apply. For ptrace, yes, you still have the upper half (but you also now have KPTI).

Q2: I'm sure it's theoretically possible. Not sure how easy or practical. Applies the same way to different VMs on the same host. (If you can probe cache timing, you can exploit Spectre.) Heck, apparently some clever folks have successfully exploited Spectre by probing cache timings with network packets.

Feel free to reopen for follow-up.