How does this compare to CHERI?
replies(2):
CHERI-Morello uses 129-bit capability objects to tag operations, has a parallel capability stack, capability pointers, and requires microarchitectural support for a tag storage memory. Basically with CHERI-Morello, your memory operations also need to provide a pointer to a capability object stored in the capability store. Everything that touches memory points to your capability, which tells the processor _what_ you can do with memory and the bounds of the memory you can touch. The capability store is literally a separate bus and memory that isn't accessible by programs, so there are no secrets: even if you leak the pointer to a capability, it doesn't matter, because it's not in a place that "user code" can ever touch. This is fine in theory, but it's incredibly expensive in practice.
MIE is a much simpler notion that seems to use N-bit (maybe 4?) tags to protect heap allocations, and uses the SPTM to protect tag space from kernel compromise. If it's exactly as in the article: heap allocations get a tag. Any load/store operation to the heap needs to provide the tag that was used for their allocation in the pointer. The tag store used by the kernel allocator is protected by SPTM so you can't just dump the tags.
If you combine MIE, SPTM, and PAC, you get close-ish to CHERI, but with independent building blocks. It's less robust, but also a less granular system with less overhead.
MIE is both probabilistic (N-bits of entropy) and protected by a slightly weaker hardware protection (SPTM, which to my understanding is a bus firewall, vs. a separate bus). It also only protects heap allocations, although existing mitigations protect the stack and execution flow.
Going off of the VERY limited information in the post, my naive read is that the biggest vulnerability here will be tag collision. If you try enough times with enough heap spray, or can groom the heap repeatedly, you can probably collide a tag with however many bits of entropy are present in the system. But, because the model is synchronous, you will bus fault every time before that, unlike MTE, so you'll get caught, which is a big problem for nation-state attackers.
MTE is 4 bits with 16 byte granularity. There's usually at least 1 tag reserved so there are 15 random tags. It's possible to dynamically exclude tags to have extra deterministic guarantees. GrapheneOS excludes the previous random tag and adjacent random tags so there are 3 dynamically excluded tags which were themselves random.
Linux kernel MTE integration for internal usage is not very security focused and has to be replaced with a security-focused implementation integrated with pKVM at some point. Google's recently launched Advanced Protection feature currently doesn't use kernel MTE.