Re: [RFC 00/10] Process-local memory allocations for hiding KVM secrets

[Liran Alon]

> On 12 Jun 2019, at 21:25, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> 
> On Wed, Jun 12, 2019 at 07:08:24PM +0200, Marius Hillenbrand wrote:
>> The Linux kernel has a global address space that is the same for any
>> kernel code. This address space becomes a liability in a world with
>> processor information leak vulnerabilities, such as L1TF. With the right
>> cache load gadget, an attacker-controlled hyperthread pair can leak
>> arbitrary data via L1TF. Disabling hyperthreading is one recommended
>> mitigation, but it comes with a large performance hit for a wide range
>> of workloads.
>> 
>> An alternative mitigation is to not make certain data in the kernel
>> globally visible, but only when the kernel executes in the context of
>> the process where this data belongs to.
>> 
>> This patch series proposes to introduce a region for what we call
>> process-local memory into the kernel's virtual address space. Page
>> tables and mappings in that region will be exclusive to one address
>> space, instead of implicitly shared between all kernel address spaces.
>> Any data placed in that region will be out of reach of cache load
>> gadgets that execute in different address spaces. To implement
>> process-local memory, we introduce a new interface kmalloc_proclocal() /
>> kfree_proclocal() that allocates and maps pages exclusively into the
>> current kernel address space. As a first use case, we move architectural
>> state of guest CPUs in KVM out of reach of other kernel address spaces.
> 
> Can you briefly describe what types of attacks this is intended to
> mitigate?  E.g. guest-guest, userspace-guest, etc...  I don't want to
> make comments based on my potentially bad assumptions.

I think I can assist in the explanation.

Consider the following scenario:
1) Hyperthread A in CPU core runs in guest and triggers a VMExit which is handled by host kernel.
While hyperthread A runs VMExit handler, it populates CPU core cache / internal-resources (e.g. MDS buffers)
with some sensitive data it have speculatively/architecturally access.
2) During hyperthread A running on host kernel, hyperthread B on same CPU core runs in guest and use
some CPU speculative execution vulnerability to leak the sensitive host data populated by hyperthread A
in CPU core cache / internal-resources.

Current CPU microcode mitigations (L1D/MDS flush) only handle the case of a single hyperthread and donât
provide a mechanism to mitigate this hyperthreading attack scenario.

Assuming there is some guest triggerable speculative load gadget in some VMExit path,
it can be used to force any data that is mapped into kernel address space to be loaded into CPU resource that is subject to leak.
Therefore, there were multiple attempts to reduce sensitive information from being mapped into the kernel address space
that is accessible by this VMExit path.

One attempt was XPFO which attempts to remove from kernel direct-map any page that is currently used only by userspace.
Unfortunately, XPFO currently exhibits multiple performance issues that *currently* makes it impractical as far as I know.

Another attempt is this patch-series which attempts to remove from one vCPU thread host kernel address space,
the state of vCPUs of other guests. Which is very specific but I personally have additional ideas on how this patch series can be further used.
For example, vhost-net needs to kmap entire guest memory into kernel-space to write ingress packets data into guest memory.
Thus, vCPU thread kernel address space now maps entire other guest memory which can be leaked using the technique described above.
Therefore, it should be useful to also move this kmap() to happen on process-local kernel virtual address region.

One could argue however that there is still a much bigger issue because of kernel direct-map that maps all physical pages that kernel
manage (i.e. have struct page) in kernel virtual address space. And all of those pages can theoretically be leaked.
However, this could be handled by complementary techniques such as booting host kernel with âmem=Xâ and mapping guest memory
by directly mmap relevant portion of /dev/mem.
Which is probably what AWS does given these upstream KVM patches they have contributed:
bd53cb35a3e9 X86/KVM: Handle PFNs outside of kernel reach when touching GPTEs
e45adf665a53 KVM: Introduce a new guest mapping API
0c55671f84ff kvm, x86: Properly check whether a pfn is an MMIO or not

Also note that when using such âmem=Xâ technique, you can also avoid performance penalties introduced by CPU microcode mitigations.
E.g. You can avoid doing L1D flush on VMEntry if VMExit handler run only in kernel and didnât context-switch as you assume kernel address
space donât map any host sensitive data.

Itâs also worth mentioning that another alternative that I have attempted to this âmem=Xâ technique
was to create an isolated address space that is only used when running KVM VMExit handlers.
For more information, refer to:
https://lkml.org/lkml/2019/5/13/515
(See some of my comments on that thread)

This is my 2cents on this at least.

-Liran