0

In virtualization,is sensitive instructions an issue because it probably gets executed in non-privileged mode without getting trapped or raising an interrupt?

In hardware-assisted virtualization, is the Intel-VT, AMD Pacifica solving the problem by actually executing the sensitive instruction or it is just creating a interrupt trap?

Is Hypervisor catching the trap from Intel-VT/AMD Pacifica and then choosing to execute the sensitive instruction or not?

James Dsouza
  • 23
  • 2

2 Answers2

2

In virtualization,is sensitive instructions an issue because it probably gets executed in non-privileged mode without getting trapped or raising an interrupt?

For virtualization; sensitive instructions would be a major problem if they aren't trapped.

In hardware-assisted virtualization, is the Intel-VT, AMD Pacifica solving the problem by actually executing the sensitive instruction or it is just creating a interrupt trap? Is Hypervisor catching the trap from Intel-VT/AMD Pacifica and then choosing to execute the sensitive instruction or not?

For hardware-assisted virtualization; the CPU sees the sensitive instruction and switches to the hyper-visor (e.g. "VMEXIT") instead of executing the sensitive instruction; and the hyper-visor emulates the instruction (it doesn't ignore the instruction or ask the CPU to execute the instruction).

The emulation of an instruction may be very simple (e.g. clearing a bit in a variable somewhere in response to a cli instruction) and may be very complicated (e.g. using sockets on the host to emulate the behavior of a real NE2000 network card to emulate an out instruction in the guest).

Brendan
  • 35,656
  • 2
  • 39
  • 66
1

First paragraph: Yes, sensitive instructions that do not trap are an issue for x86 virtualisation.

Extensions like SVM generally trap for all sensitive instructions AFAIK. The hypervisor's handling of the instructions depends on what the instructions are. It can be either ignored (skipped), handled by some sort of emulation, or directly be executed by the hypervisor.

However, in (non-AMD64) V86M's VME (Virtual Mode Extensions), specific software interrupts (which are generally sensitive) can be handled without trapping.

ecm
  • 2,583
  • 4
  • 21
  • 29
  • I'm pretty sure some common semi-sensitive instructions like `pushf` or `popf` are just virtualized without trapping. It's also configurable for `rdtsc` and `cpuid`, I think, and there's HW support for scaling and offsetting the RDTSC time (in a guest VM only, not in general.) – Peter Cordes Jul 24 '19 at 21:01
  • I would if I was sure enough about any details :P – Peter Cordes Jul 24 '19 at 21:05