Garbage collection needs to go into the kernel. Providing fast, fine-grained control over virtual memory mappings gets you a read barrier on any system with virtual memory, which lets you build real-time GCs. Azul's patches to Linux show how this can be done:
They're using the VT-x for the virtualized page tables and still run the GC in userspace, but if you move the GC into the kernel you won't even need virtualized page tables.
http://lambda-the-ultimate.org/node/4165
They're using the VT-x for the virtualized page tables and still run the GC in userspace, but if you move the GC into the kernel you won't even need virtualized page tables.