On Fri, 2024-09-06 at 14:10 +1200, Huang, Kai wrote:
>
> > --- a/arch/x86/kvm/vmx/main.c
> > +++ b/arch/x86/kvm/vmx/main.c
> > @@ -36,9 +36,21 @@ static __init int vt_hardware_setup(void)
> > * is KVM may allocate couple of more bytes than needed for
> > * each VM.
> > */
> > - if (enable_tdx)
> > + if (enable_tdx) {
> > vt_x86_ops.vm_size = max_t(unsigned int, vt_x86_ops.vm_size,
> > sizeof(struct kvm_tdx));
> > + /*
> > + * Note, TDX may fail to initialize in a later time in
> > + * vt_init(), in which case it is not necessary to setup
> > + * those callbacks. But making them valid here even
> > + * when TDX fails to init later is fine because those
> > + * callbacks won't be called if the VM isn't TDX guest.
> > + */
> > + vt_x86_ops.link_external_spt = tdx_sept_link_private_spt;
> > + vt_x86_ops.set_external_spte = tdx_sept_set_private_spte;
> > + vt_x86_ops.free_external_spt = tdx_sept_free_private_spt;
> > + vt_x86_ops.remove_external_spte =
> > tdx_sept_remove_private_spte;
>
> Nit: The callbacks in 'struct kvm_x86_ops' have name "external", but
> TDX callbacks have name "private". Should we rename TDX callbacks to
> make them aligned?
"external" is the core MMU naming abstraction. I think you were part of the
discussion to purge special TDX private naming from the core MMU to avoid
confusion with AMD private memory in the last MMU series.
So external page tables ended up being a general concept, and private mem is the
TDX use. In practice of course it will likely only be used for TDX. So I thought
the external<->private connection here was nice to have.
>
> > + }
> >
> > return 0;
> > }
> > diff --git a/arch/x86/kvm/vmx/tdx.c b/arch/x86/kvm/vmx/tdx.c
> > index 6feb3ab96926..b8cd5a629a80 100644
> > --- a/arch/x86/kvm/vmx/tdx.c
> > +++ b/arch/x86/kvm/vmx/tdx.c
> > @@ -447,6 +447,177 @@ void tdx_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t
> > root_hpa, int pgd_level)
> > td_vmcs_write64(to_tdx(vcpu), SHARED_EPT_POINTER, root_hpa);
> > }
> >
> > +static void tdx_unpin(struct kvm *kvm, kvm_pfn_t pfn)
> > +{
> > + struct page *page = pfn_to_page(pfn);
> > +
> > + put_page(page);
> > +}
> > +
> > +static int tdx_mem_page_aug(struct kvm *kvm, gfn_t gfn,
> > + enum pg_level level, kvm_pfn_t pfn)
> > +{
> > + int tdx_level = pg_level_to_tdx_sept_level(level);
> > + struct kvm_tdx *kvm_tdx = to_kvm_tdx(kvm);
> > + hpa_t hpa = pfn_to_hpa(pfn);
> > + gpa_t gpa = gfn_to_gpa(gfn);
> > + u64 entry, level_state;
> > + u64 err;
> > +
> > + err = tdh_mem_page_aug(kvm_tdx, gpa, hpa, &entry, &level_state);
> > + if (unlikely(err == TDX_ERROR_SEPT_BUSY)) {
> > + tdx_unpin(kvm, pfn);
> > + return -EAGAIN;
> > + }
>
> Nit: Here (and other non-fatal error cases) I think we should return
> -EBUSY to make it consistent with non-TDX case? E.g., the non-TDX case has:
>
> if (!try_cmpxchg64(sptep, &iter->old_spte, new_spte))
> return -EBUSY;
>
> And the comment of tdp_mmu_set_spte_atomic() currently says it can only
> return 0 or -EBUSY. It needs to be patched to reflect it can also
> return other non-0 errors like -EIO but those are fatal. In terms of
> non-fatal error I don't think we need another -EAGAIN.
Yes, good point.
>
> /*
> * tdp_mmu_set_spte_atomic - Set a TDP MMU SPTE atomically
>
> [...]
>
> * Return:
> * * 0 - If the SPTE was set.
> * * -EBUSY - If the SPTE cannot be set. In this case this function will
> * have no side-effects other than setting iter->old_spte to
> * the last known value of the spte.
> */
>
> [...]
>
> > +
> > +static int tdx_sept_drop_private_spte(struct kvm *kvm, gfn_t gfn,
> > + enum pg_level level, kvm_pfn_t pfn)
> > +{
> >
> [...]
>
> > +
> > + hpa_with_hkid = set_hkid_to_hpa(hpa, (u16)kvm_tdx->hkid);
> > + do {
> > + /*
> > + * TDX_OPERAND_BUSY can happen on locking PAMT entry.
> > Because
> > + * this page was removed above, other thread shouldn't be
> > + * repeatedly operating on this page. Just retry loop.
> > + */
> > + err = tdh_phymem_page_wbinvd(hpa_with_hkid);
> > + } while (unlikely(err == (TDX_OPERAND_BUSY | TDX_OPERAND_ID_RCX)));
>
> In what case(s) other threads can concurrently lock the PAMT entry,
> leading to the above BUSY?
Good question, lets add this to the seamcall retry research.
>
> [...]
>
> > +
> > +int tdx_sept_remove_private_spte(struct kvm *kvm, gfn_t gfn,
> > + enum pg_level level, kvm_pfn_t pfn)
> > +{
> > + int ret;
> > +
> > + /*
> > + * HKID is released when vm_free() which is after closing gmem_fd
>
> From latest dev branch HKID is freed from vt_vm_destroy(), but not
> vm_free() (which should be tdx_vm_free() btw).
Oh, yes, we should update the comment.
>
> static void vt_vm_destroy(struct kvm *kvm)
> {
> if (is_td(kvm))
> return tdx_mmu_release_hkid(kvm);
>
> vmx_vm_destroy(kvm);
> }
>
> Btw, why not have a tdx_vm_destroy() wrapper? Seems all other vt_xx()s
> have a tdx_xx() but only this one calls tdx_mmu_release_hkid() directly.
No strong reason. It's asymmetric to the other tdx callbacks, but KVM code tends
to be less wrapped and a tdx_vm_destory would be a oneline function. So I think
it fits in other ways.
>
> > + * which causes gmem invalidation to zap all spte.
> > + * Population is only allowed after KVM_TDX_INIT_VM.
> > + */
>
> What does the second sentence ("Population ...") meaning? Why is it
> relevant here?
>
How about:
/*
* HKID is released after all private pages have been removed,
* and set before any might be populated. Warn if zapping is
* attempted when there can't be anything populated in the private
* EPT.
*/
But actually, I wonder if we need to remove the KVM_BUG_ON(). I think if you did
a KVM_PRE_FAULT_MEMORY and then deleted the memslot you could hit it?
