Add support for arm cca

[MatiasVara]

This PR aims at adding support to build realm guests. First commit adds support for create_guest_memfd() and set_user_memory_region2(). To do this, the memory_init() is modified by adding a boolean parameter. This is required when building a confidential guest for arm cca and probably also required by other cases. The second commit imports the virtee/cca crate and adds the steps to build a cca guest. The following items should be completed before merge the PR:

• [x] use populate() only for kernel and initialize() for non-kernel area
• [x] handle the case in which the guest switch from shared -> private
• [x] correctly calculation of max_ipa

This has been testing using the v7 series for Linux as a guest and v6 series for KVM on FVP model.

Feedback is welcome.

[jakecorrenti]

@MatiasVara is this ready for review, or should I wait?

[tylerfanelli]

@jakecorrenti I'm addressing the KVM guest_memfd changes (which @MatiasVara also added in this PR) to my latest SEV-SNP patches. This will probably require a rebase after that. Once that happens, it will likely be ready for a review.

[MatiasVara]

@jakecorrenti I'm addressing the KVM guest_memfd changes (which @MatiasVara also added in this PR) to my latest SEV-SNP patches. This will probably require a rebase after that. Once that happens, it will likely be ready for a review.

I think the PR is ready to review. The current issue is that it does not work for the latest series for KVM (v6). I am investigating the issue.

[MatiasVara]

I just updated the PR with the fix for v6 series. The fix is only to set the KVM_ARM_VCPU_REC feature during the initialization of VCPUs otherwise vcpu_finalize would fail.

[MatiasVara]

@jakecorrenti @tylerfanelli @slp Feel free to give a quick look. I am pretty sure there many things to fix.

[MatiasVara]

After reviewing the Jake comments I came to the conclusion that this PR should be split otherwise is hard to review:

0. Add Use create_guest_memfd() and set_user_memory_region2() only when tee is enabled.
1. Add the handling of KVM_MEMORY_FAULT only for tee, which includes the extra thread for handling shared to private and viceverse.
2. Tee should be replaced with tee and x86-64 since all tee is for x86-64.
3. Add CCA support. This shall be added with cca and tee enabled.
4. Eventually some code will be shared for tee in x86-64 and aarch64 but I can't see it yet.

@jakecorrenti @tylerfanelli @slp WDYT?

[jakecorrenti]

After reviewing the Jake comments I came to the conclusion that this PR should be split otherwise is hard to review:

0. Add `Use create_guest_memfd() and set_user_memory_region2()` only when `tee` is enabled.

1. Add the handling of `KVM_MEMORY_FAULT` only for `tee`, which includes the extra thread for handling shared to private and viceverse.

2. `Tee` should be replaced with `tee and x86-64` since all `tee` is for `x86-64`.

I don't know if this is idiomatic by any means, and it depends on whether you can have target arch-specific features, but could we have something like tee-x86 and tee? tee-x86 for the x86 specific functionality, and then tee for the shared bits? Would that be cleaner than just writing #[cfg(all(feature = "tee", target_arch = "x86_64"))]? Again, I don't know if that's "proper" Rust or even possible.

3. Add CCA support. This shall be added with `cca and tee` enabled.

4. Eventually some code will be shared for `tee` in `x86-64` and `aarch64` but I can't see it yet.

@jakecorrenti @tylerfanelli @slp WDYT?

Other than my question above, that all sounds good to me.

[tylerfanelli]

After reviewing the Jake comments I came to the conclusion that this PR should be split otherwise is hard to review:

0. Add Use create_guest_memfd() and set_user_memory_region2() only when tee is enabled.

I don't think set_user_memory_region2 should be tee only, as I believe that on newer kernels, set_user_memory_region is deprecated.

1. Add the handling of KVM_MEMORY_FAULT only for tee, which includes the extra thread for handling shared to private and viceverse.

I'm addressing that in my SEV-SNP PR currently.

2. Tee should be replaced with tee and x86-64 since all tee is for x86-64.

Makes sense to me.

3. Add CCA support. This shall be added with cca and tee enabled.

Do you mean aarch64 and tee?

[MatiasVara]

After reviewing the Jake comments I came to the conclusion that this PR should be split otherwise is hard to review:

0. Add `Use create_guest_memfd() and set_user_memory_region2()` only when `tee` is enabled.

1. Add the handling of `KVM_MEMORY_FAULT` only for `tee`, which includes the extra thread for handling shared to private and viceverse.

2. `Tee` should be replaced with `tee and x86-64` since all `tee` is for `x86-64`.

I don't know if this is idiomatic by any means, and it depends on whether you can have target arch-specific features, but could we have something like tee-x86 and tee? tee-x86 for the x86 specific functionality, and then tee for the shared bits? Would that be cleaner than just writing #[cfg(all(feature = "tee", target_arch = "x86_64"))]? Again, I don't know if that's "proper" Rust or even possible.

I do not know either. I though that tee and x86 would have code that tdx and snv share, if that exists.

3. Add CCA support. This shall be added with `cca and tee` enabled.

4. Eventually some code will be shared for `tee` in `x86-64` and `aarch64` but I can't see it yet.

@jakecorrenti @tylerfanelli @slp WDYT?

Other than my question above, that all sounds good to me.

[MatiasVara]

After reviewing the Jake comments I came to the conclusion that this PR should be split otherwise is hard to review:

0. Add Use create_guest_memfd() and set_user_memory_region2() only when tee is enabled.

I don't think set_user_memory_region2 should be tee only, as I believe that on newer kernels, set_user_memory_region is deprecated.

That's OK. I could just leave set_user_memory_region2 for the future.

1. Add the handling of KVM_MEMORY_FAULT only for tee, which includes the extra thread for handling shared to private and viceverse.

I'm addressing that in my SEV-SNP PR currently.

The first two items in my list are for all tee but they were in my PR for almost a year already. In any case, I need to split those chunks from my PR.

2. Tee should be replaced with tee and x86-64 since all tee is for x86-64.

Makes sense to me.

3. Add CCA support. This shall be added with cca and tee enabled.

Do you mean aarch64 and tee?

Yes for the moment.

[MatiasVara]

After reviewing the Jake comments I came to the conclusion that this PR should be split otherwise is hard to review:

0. Add Use create_guest_memfd() and set_user_memory_region2() only when tee is enabled.
1. Add the handling of KVM_MEMORY_FAULT only for tee, which includes the extra thread for handling shared to private and viceverse.
2. Tee should be replaced with tee and x86-64 since all tee is for x86-64.
3. Add CCA support. This shall be added with cca and tee enabled.
4. Eventually some code will be shared for tee in x86-64 and aarch64 but I can't see it yet.

@jakecorrenti @tylerfanelli @slp WDYT?

I just split the PR into a few commits. It is the same code though. The first three commits are independent of cca and should be shared by all tees. Also [Enable tee only in x86-64 shall for all tees since it aims at separating the code for tee and x86 and tee and aach64.

[MatiasVara]

I open two PR to update kvm-bindings (see https://github.com/virtee/kvm-bindings/pull/2) and kvm-ioctls (see https://github.com/virtee/kvm-ioctls/pull/4) due to changes upstream (v8).

[MatiasVara]

I've Just rebased the PR and tried it on v8 series.

[MatiasVara]

I just rebased and tested on v8. Also, I am now relying on the rust-vmm/kvm repo.

[MatiasVara]

I think the following issues shall be addressed before merging:

• [x] support for blk
• [x] Check the EXIT_FAULT
• [ ] Check FDT hardcoded
• [x] Try v9 series