| Message ID | 20240329225835.400662-23-michael.roth@amd.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Add AMD Secure Nested Paging (SEV-SNP) Hypervisor Support | expand |
On 3/29/24 23:58, Michael Roth wrote: > + /* > + * If an unaligned PFN corresponds to a 2M region assigned as a > + * large page in he RMP table, PSMASH the region into individual > + * 4K RMP entries before attempting to convert a 4K sub-page. > + */ > + if (!use_2m_update && rmp_level > PG_LEVEL_4K) { > + rc = snp_rmptable_psmash(pfn); > + if (rc) > + pr_err_ratelimited("SEV: Failed to PSMASH RMP entry for PFN 0x%llx error %d\n", > + pfn, rc); > + } Ignoring the PSMASH failure is pretty scary... At this point .free_folio cannot fail, should the psmash part of this patch be done in kvm_gmem_invalidate_begin() before kvm_mmu_unmap_gfn_range()? Also, can you get PSMASH_FAIL_INUSE and if so what's the best way to address it? Should fallocate() return -EBUSY? Thanks, Paolo
On Sat, Mar 30, 2024 at 10:31:47PM +0100, Paolo Bonzini wrote: > On 3/29/24 23:58, Michael Roth wrote: > > + /* > > + * If an unaligned PFN corresponds to a 2M region assigned as a > > + * large page in he RMP table, PSMASH the region into individual > > + * 4K RMP entries before attempting to convert a 4K sub-page. > > + */ > > + if (!use_2m_update && rmp_level > PG_LEVEL_4K) { > > + rc = snp_rmptable_psmash(pfn); > > + if (rc) > > + pr_err_ratelimited("SEV: Failed to PSMASH RMP entry for PFN 0x%llx error %d\n", > > + pfn, rc); > > + } > > Ignoring the PSMASH failure is pretty scary... At this point .free_folio > cannot fail, should the psmash part of this patch be done in > kvm_gmem_invalidate_begin() before kvm_mmu_unmap_gfn_range()? > > Also, can you get PSMASH_FAIL_INUSE and if so what's the best way to address > it? Should fallocate() return -EBUSY? FAIL_INUSE shouldn't occur since at this point the pages have been unmapped from NPT and only the task doing the cleanup should be attempting to access/PSMASH this particular 2M HPA range at this point. However, since FAIL_INUSE is transient, there isn't a good reason why we shouldn't retry until it clears itself up rather than risk hosing the system if some unexpected case ever did pop up, so I've updated snp_rmptable_psmash() to handle that case automatically and simplify the handling in sev_handle_rmp_fault() as well. (in the case of #NPF RMP faults there is actually potential for PSMASH errors other than FAIL_INUSE due to races with other vCPU threads which can interleave and put the RMP entry in an unexpected state, so there's additional handling/reporting to deal with those cases, but here they are not expected and will trigger WARN_*ONCE()'s now) I used this hacked up version of Sean's original patch to re-enable 2MB hugepage support in gmem for the purposes of re-testing this: https://github.com/mdroth/linux/commit/15aa4f81811485997953130fc184e829ba4399d2 -Mike > > Thanks, > > Paolo > >
diff --git a/arch/x86/kvm/Kconfig b/arch/x86/kvm/Kconfig index 286b40d0b07c..32a5c37cbf88 100644 --- a/arch/x86/kvm/Kconfig +++ b/arch/x86/kvm/Kconfig @@ -125,6 +125,7 @@ config KVM_AMD_SEV select ARCH_HAS_CC_PLATFORM select KVM_GENERIC_PRIVATE_MEM select HAVE_KVM_GMEM_PREPARE + select HAVE_KVM_GMEM_INVALIDATE help Provides support for launching Encrypted VMs (SEV) and Encrypted VMs with Encrypted State (SEV-ES) on AMD processors. diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c index e1f8be1df219..87d621d013a4 100644 --- a/arch/x86/kvm/svm/sev.c +++ b/arch/x86/kvm/svm/sev.c @@ -4380,3 +4380,66 @@ int sev_gmem_prepare(struct kvm *kvm, kvm_pfn_t pfn, gfn_t gfn, int max_order) return 0; } + +void sev_gmem_invalidate(kvm_pfn_t start, kvm_pfn_t end) +{ + kvm_pfn_t pfn; + + pr_debug("%s: PFN start 0x%llx PFN end 0x%llx\n", __func__, start, end); + + for (pfn = start; pfn < end;) { + bool use_2m_update = false; + int rc, rmp_level; + bool assigned; + + rc = snp_lookup_rmpentry(pfn, &assigned, &rmp_level); + if (rc) { + pr_debug_ratelimited("SEV: Failed to retrieve RMP entry for PFN 0x%llx error %d\n", + pfn, rc); + goto next_pfn; + } + + if (!assigned) + goto next_pfn; + + use_2m_update = IS_ALIGNED(pfn, PTRS_PER_PMD) && + end >= (pfn + PTRS_PER_PMD) && + rmp_level > PG_LEVEL_4K; + + /* + * If an unaligned PFN corresponds to a 2M region assigned as a + * large page in he RMP table, PSMASH the region into individual + * 4K RMP entries before attempting to convert a 4K sub-page. + */ + if (!use_2m_update && rmp_level > PG_LEVEL_4K) { + rc = snp_rmptable_psmash(pfn); + if (rc) + pr_err_ratelimited("SEV: Failed to PSMASH RMP entry for PFN 0x%llx error %d\n", + pfn, rc); + } + + rc = rmp_make_shared(pfn, use_2m_update ? PG_LEVEL_2M : PG_LEVEL_4K); + if (WARN_ON_ONCE(rc)) { + pr_err_ratelimited("SEV: Failed to update RMP entry for PFN 0x%llx error %d\n", + pfn, rc); + goto next_pfn; + } + + /* + * SEV-ES avoids host/guest cache coherency issues through + * WBINVD hooks issued via MMU notifiers during run-time, and + * KVM's VM destroy path at shutdown. Those MMU notifier events + * don't cover gmem since there is no requirement to map pages + * to a HVA in order to use them for a running guest. While the + * shutdown path would still likely cover things for SNP guests, + * userspace may also free gmem pages during run-time via + * hole-punching operations on the guest_memfd, so flush the + * cache entries for these pages before free'ing them back to + * the host. + */ + clflush_cache_range(__va(pfn_to_hpa(pfn)), + use_2m_update ? PMD_SIZE : PAGE_SIZE); +next_pfn: + pfn += use_2m_update ? PTRS_PER_PMD : 1; + } +} diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c index c099154e326a..b456906f2670 100644 --- a/arch/x86/kvm/svm/svm.c +++ b/arch/x86/kvm/svm/svm.c @@ -5080,6 +5080,7 @@ static struct kvm_x86_ops svm_x86_ops __initdata = { .alloc_apic_backing_page = svm_alloc_apic_backing_page, .gmem_prepare = sev_gmem_prepare, + .gmem_invalidate = sev_gmem_invalidate, }; /* diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h index 53618cfc2b89..3f1f6d3d3ade 100644 --- a/arch/x86/kvm/svm/svm.h +++ b/arch/x86/kvm/svm/svm.h @@ -731,6 +731,7 @@ void sev_handle_rmp_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u64 error_code); void sev_vcpu_unblocking(struct kvm_vcpu *vcpu); void sev_snp_init_protected_guest_state(struct kvm_vcpu *vcpu); int sev_gmem_prepare(struct kvm *kvm, kvm_pfn_t pfn, gfn_t gfn, int max_order); +void sev_gmem_invalidate(kvm_pfn_t start, kvm_pfn_t end); #else static inline struct page *snp_safe_alloc_page(struct kvm_vcpu *vcpu) { return alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO); @@ -751,6 +752,7 @@ static inline int sev_gmem_prepare(struct kvm *kvm, kvm_pfn_t pfn, gfn_t gfn, in { return 0; } +static inline void sev_gmem_invalidate(kvm_pfn_t start, kvm_pfn_t end) {} #endif
Implement a platform hook to do the work of restoring the direct map entries of gmem-managed pages and transitioning the corresponding RMP table entries back to the default shared/hypervisor-owned state. Signed-off-by: Michael Roth <michael.roth@amd.com> --- arch/x86/kvm/Kconfig | 1 + arch/x86/kvm/svm/sev.c | 63 ++++++++++++++++++++++++++++++++++++++++++ arch/x86/kvm/svm/svm.c | 1 + arch/x86/kvm/svm/svm.h | 2 ++ 4 files changed, 67 insertions(+)