Re: [PATCH v2 0/2] MTE support for KVM guest

From: Steven Price
Date: Wed Sep 09 2020 - 12:52:52 EST

On 09/09/2020 16:25, Andrew Jones wrote:
On Fri, Sep 04, 2020 at 05:00:16PM +0100, Steven Price wrote:
Arm's Memory Tagging Extension (MTE) adds 4 bits of tag data to every 16
bytes of memory in the system. This along with stashing a tag within the
high bit of virtual addresses allows runtime checking of memory

These patches add support to KVM to enable MTE within a guest. They are
based on Catalin's v9 MTE user-space support series[1].

I'd welcome feedback on the proposed user-kernel ABI. Specifically this
series currently:

0. Feature probing

Probably a KVM cap, rather than requiring userspace to attempt VCPU
features one at a time with a scratch VCPU.

Ah, good point - thanks for pointing that out.

1. Requires the VMM to enable MTE per-VCPU.

I suppose. We're collecting many features that are enabling CPU features,
so they map nicely to VCPU features, yet they're effectively VM features
due to a shared resource such as an irq or memory.

Yeah this is a little weird I'll admit. The architectural feature is described per-CPU (well "processing element"), but it makes little sense to have it only on some CPUs and has effects on the rest of the memory system. Given that it's theoretically possible to build e.g. a big.LITTLE setup with only some CPUs support MTE it seemed more future-proof to design the API to allow it even though I hope no-one will use it.

2. Automatically promotes (normal host) memory given to the guest to be
tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
tags are cleared if the memory wasn't previously MTE enabled.

Shouldn't this be up to the guest? Or, is this required in order for the
guest to use tagging at all. Something like making the guest IPAs memtag
capable, but if the guest doesn't enable tagging then there is no guest
impact? In any case, shouldn't userspace be the one that adds PROT_MTE
to the memory regions it wants the guest to be able to use tagging with,
rather than KVM adding the attribute page by page?

I think I've probably explained this badly.

The guest can choose how to populate the stage 1 mapping - so can choose which parts of memory are accessed tagged or not. However, the hypervisor cannot restrict this in stage 2 (except by e.g. making the memory uncached but that's obviously not great - however devices forward to the guest can be handled like this).

Because the hypervisor cannot restrict the guest's access to the tags, the hypervisor must assume that all memory given to the guest could have the tags accessed. So it must (a) clear any stale data from the tags, and (b) ensure that the tags are preserved (e.g. when swapping pages out).

Because of the above the current series automatically sets PG_mte_tagged on the pages. Note that this doesn't change the mappings that the VMM has (a non-PROT_MTE mapping will still not have access to the tags).

It's a shame that the stage-2 can't usefully restrict tag access, but this matches the architectural expectation: that if MTE is supported then all standard memory will be MTE-enabled.

3. Doesn't provide any new methods for the VMM to access the tags on

(2) and (3) are particularly interesting from the aspect of VM migration.
The guest is able to store/retrieve data in the tags (presumably for the
purpose of tag checking, but architecturally it could be used as just
storage). This means that when migrating a guest the data needs to be
transferred (or saved/restored).

MTE tags are controlled by the same permission model as normal pages
(i.e. a read-only page has read-only tags), so the normal methods of
detecting guest changes to pages can be used. But this would also
require the tags within a page to be migrated at the same time as the
data (since the access control for tags is the same as the normal data
within a page).

(3) may be problematic and I'd welcome input from those familiar with
VMMs. User space cannot access tags unless the memory is mapped with the
PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
for the user space process (assuming the VMM enables tag checking for
the process) and since the tags in memory are controlled by the guest
it's unlikely the VMM would have an appropriately tagged pointer for its
access. This means the VMM would either need to maintain two mappings of
memory (one to access tags, the other to access data) or disable tag
checking during the accesses to data.

If userspace needs to write to guest memory then it should be due to
a device DMA or other specific hardware emulation. Those accesses can
be done with tag checking disabled.

Yes, the question is can the VMM (sensibly) wrap the accesses with a disable/renable tag checking for the process sequence. The alternative at the moment is to maintain a separate (untagged) mapping for the purpose which might present it's own problems.

If it's not practical to either disable tag checking in the VMM or
maintain multiple mappings then the alternatives I'm aware of are:

* Provide a KVM-specific method to extract the tags from guest memory.
This might also have benefits in terms of providing an easy way to
read bulk tag data from guest memory (since the LDGM instruction
isn't available at EL0).

Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
the tags for all addresses of each dirty page.

Certainly possible, although it seems to conflate two operations: "get list of dirty pages", "get tags from page". It would also require a lot of return space (size of slot/32).

* Provide support for user space setting the TCMA0 or TCMA1 bits in
TCR_EL1. These would allow the VMM to generate pointers which are not
tag checked.

So this is necessary to allow the VMM to keep tag checking enabled for
itself, plus map guest memory as PROT_MTE, and write to that memory when

This is certainly one option. The architecture provides two "magic" values (all-0s and all-1s) which can be configured using TCMAx to be treated differently. The VMM could therefore construct pointers to otherwise tagged memory which would be treated as untagged.

However, Catalin's user space series doesn't at the moment expose this functionality.