13:50 <aneesh> maz, with respect to MTE_PERM usecase, one of the goals for MTE_PERM was to allow support for smaller pool of allocation tagged memory. But supporting such a pool will require wider MM changes and there was no consensus w.r.t the usefulness of supporting that usecase. We can also use MTE_PERM for another usecase. ie, supporting usage of MTE in a guest where we don't allow MTE now.

13:51 <aneesh> I am looking at implementing that specific use case such a way that future updates are possible for smaller allocation tagged memory pool.

13:53 <maz> aneesh: what are these use-cases, and why don't they currently work?

13:53 <aneesh> The two i listed in my email were. "libkrun and virtiofs

13:54 <aneesh> They don't support MTE now because we don't support MTE with page cache and some of the memory regions back guest memory are from page cache

13:55 <aneesh> in both the case, these memory regions are special in that user space allocation will not come from these regions and we can be sure there will not be a tag access check against those pfns

14:04 <maz> and what does it mean for that memory that is now tagged for VMs only? what guarantees that there will not be a tagged access outside of the VM? How do you transition a page back from tagged to non-tagged?

14:05 <maz> and crucially, what is the userspace ABI for managing these transition?

14:07 <maz> you need to have clear answers to these points before we can even start reviewing patches.

14:27 <aneesh> sure. I understand the confusion. It is also that i am may be explaining all the details in my cover letter. For the usecase i am trying to address, there will not be a conversion from tagged to non-tagged. For ex: we can look at guest memory made of two vmas , one anon and one pagecache.

14:29 <aneesh> the guest usage of its memory is such that the memory backed by page cache is considered special and will never be allocatable by userspace. Hence there won't be a tag access check on this pfns. In both the examples, these pfns are not part of buddy for any sort of page allocation from that.

14:29 <aneesh> Considering there will be no transition between tagged to non-tagged, there is no user API for that.

14:32 <aneesh> on the host all pfns are considered access tagged. on guest since some of the guest pfns are backed by pagcache, we deny the MTE usage in guest now. (kvm_arch_prepare_memory_regions()) . The changes allow the usage of MTE and if the guest break the rule that no tag access to those "special" pfns, we get a tag access check fault

14:32 <aneesh> the VM get terminated on such event.

14:46 <aneesh> maz, ^

15:40 <maz> aneesh: I really think that lacking the reverse conversion is short-sighted. you *really* want it in.

15:41 <aneesh> maz, what do you mean by reverse conversion?

15:47 <maz> from tagged to non-tagged.

15:49 <aneesh> for supporting an allocation from a special pool and making it useful as a general page allocator requires large MM changes ( possibly a separate zone, or a custom device, identifying lru pages to swap things out etc).

15:50 <aneesh> if the goal is to reduce the memory wastage , there could be other approaches. I am not sure we will have a smaller special pool of allocation tag pages soon.

15:50 <aneesh> considering such a facility doesn't exit, it didn't make sense to look at converting from tagged to non-tagged.