From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from linux.microsoft.com (linux.microsoft.com [13.77.154.182]) by smtp.subspace.kernel.org (Postfix) with ESMTP id 240CF468C0B; Wed, 5 Aug 2026 11:04:12 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=none smtp.client-ip=13.77.154.182 ARC-Seal:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1785927854; cv=none; b=GOp60DOYu4Im4ZJTcZxj9QGhThdhVwqhkzBUczgZ+gLwHWq/K0dfXHLIX+4dRM3Z2FzNKgHia89OwzsRVRXF0XOulpWkk9SFmZFXBMZYaRttotFjJl1pEiphHIrU/XeUMkJbOQUeSWecEQ0l6knd5YJHeb3eaHVrAdSRLpMXVIk= ARC-Message-Signature:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1785927854; c=relaxed/simple; bh=hp3inTy6vwa9u/UMgIvXdXMwZDjFfkMf6mB3v768YOU=; h=From:To:Cc:Subject:Date:Message-ID:In-Reply-To:References: MIME-Version; b=tlt+AELtqV6XH62ZfjapG/PY+0UpER9/NRAmUMNOlOpzORfWLR7fo9mUHSK7ytKDkypoQQ4PCUf1GDlfMCTPziR/vL5+ag3x3g/vmiKzFYNWZhAFG9bx3+xPZ39vSdtl9IWcVPRLut+ue7/qe4A8Bn0qW/og4c2TL+dhrhgkR4s= ARC-Authentication-Results:i=1; smtp.subspace.kernel.org; dmarc=pass (p=none dis=none) header.from=linux.microsoft.com; spf=pass smtp.mailfrom=linux.microsoft.com; dkim=pass (1024-bit key) header.d=linux.microsoft.com header.i=@linux.microsoft.com header.b=bOqp6a9e; arc=none smtp.client-ip=13.77.154.182 Authentication-Results: smtp.subspace.kernel.org; dmarc=pass (p=none dis=none) header.from=linux.microsoft.com Authentication-Results: smtp.subspace.kernel.org; spf=pass smtp.mailfrom=linux.microsoft.com Authentication-Results: smtp.subspace.kernel.org; dkim=pass (1024-bit key) header.d=linux.microsoft.com header.i=@linux.microsoft.com header.b="bOqp6a9e" Received: from fedora.hsd1.wa.comcast.net (unknown [52.148.140.42]) by linux.microsoft.com (Postfix) with ESMTPSA id C624520B7169; Wed, 5 Aug 2026 04:03:50 -0700 (PDT) DKIM-Filter: OpenDKIM Filter v2.11.0 linux.microsoft.com C624520B7169 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linux.microsoft.com; s=default; t=1785927830; bh=RQx/+tS3Gguj0oQsRoaGqW2DUL7VawR8cXjvWahtnwE=; h=From:To:Cc:Subject:Date:In-Reply-To:References:From; b=bOqp6a9eBubODjXnHNNo2eaStXBqrL+wwVpv5ToXbO2JjG20yOvPCdt1l6fsPtnf6 vg/4UWdFR4XKfOK99EjjFyNgZRwjzPIu1jpo9fVFLEn7WueYm4YdEiP9CQE27sw0ZZ T4BqH/1kdeh/cGHoF3h7OVTljuzFN+Golll62xcg= From: Sriram Nambakam To: kvm@vger.kernel.org Cc: linux-kernel@vger.kernel.org Subject: [RFC PATCH v1 34/42] x86/realmode: skip the sub-1M trampoline for the VBS secure plane Date: Wed, 5 Aug 2026 04:03:16 -0700 Message-ID: <20260805110324.25067-35-snambakam@linux.microsoft.com> X-Mailer: git-send-email 2.55.0 In-Reply-To: <20260805110324.25067-1-snambakam@linux.microsoft.com> References: <20260805110324.25067-1-snambakam@linux.microsoft.com> Precedence: bulk X-Mailing-List: kvm@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 Content-Transfer-Encoding: 8bit The VBS secure plane (plane >0) boots from a single high-memory region carved out of the normal plane's address space and therefore has no sub-1M RAM for the x86 real-mode AP trampoline. reserve_real_mode() followed by init_real_mode() then panics with "Real mode trampoline was not allocated". The secure plane is uniprocessor, enters directly in long mode and never uses the trampoline, so point x86_platform.realmode_reserve/realmode_init at x86_init_noop for it. This mirrors how the Hyper-V VTL (hv_vtl.c) and Xen PV ports disable the trampoline. Gated at compile time on CONFIG_VBS_SECURE_MONITOR (only the secure-plane kernel sets it) and at runtime on the "secure_monitor" early param (the normal plane never passes it), so plane 0 is unaffected. --- arch/x86/include/asm/kvm_host.h | 17 +++++-- arch/x86/kvm/mmu/mmu.c | 36 ++++++++++++++ arch/x86/kvm/mmu/spte.h | 12 +++-- arch/x86/kvm/x86.c | 85 +++++++++++++++++++++++++++++---- arch/x86/realmode/init.c | 26 ++++++++++ include/linux/kvm_host.h | 35 ++++++++++++++ include/uapi/linux/kvm.h | 1 + virt/kvm/kvm_main.c | 6 +++ 8 files changed, 202 insertions(+), 16 deletions(-) diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h index b7d478dcc1a5..bbccb9d3d801 100644 --- a/arch/x86/include/asm/kvm_host.h +++ b/arch/x86/include/asm/kvm_host.h @@ -380,15 +380,24 @@ union kvm_mmu_page_role { */ unsigned cr4_smep:1; - unsigned:3; + /* + * Plane (privilege level) that owns this shadow page. VM + * planes share memslots but must have independent page + * tables so that a higher-privilege plane can restrict a + * lower plane's access (e.g. deny reads of secure-plane + * memory). Tagging the role keeps each plane's roots and + * SPTEs separate. Always 0 when CONFIG_VM_PLANES is off. + */ + unsigned plane:4; /* * This is left at the top of the word so that * kvm_memslots_for_spte_role can extract it with a - * simple shift. While there is room, give it a whole - * byte so it is also faster to load it from memory. + * simple shift. smm is only ever used as a boolean, so it + * is reduced to 7 bits (from a full byte) to make room for + * cr4_smep and the VM-planes plane tag above. */ - unsigned smm:8; + unsigned smm:7; }; }; diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c index 6e41c5df72ed..960e212c5ee3 100644 --- a/arch/x86/kvm/mmu/mmu.c +++ b/arch/x86/kvm/mmu/mmu.c @@ -4708,6 +4708,34 @@ static int kvm_mmu_faultin_pfn(struct kvm_vcpu *vcpu, return -EFAULT; } + /* + * A higher-privilege plane may forbid this plane from accessing a gfn + * (e.g. to hide secure-plane memory from the normal plane). Two cases + * cannot be represented as a present SPTE and must be denied outright, + * exiting to userspace with a memory fault rather than (re)building an + * SPTE the access will immediately re-fault on: + * + * - NO_READ: there is no present-but-unreadable EPT entry, so leave the + * gfn unmapped for this plane. + * + * - NO_WRITE on a write fault: make_spte() strips ACC_WRITE_MASK and + * builds a read-only SPTE, so a guest write would re-fault forever + * (an unresolvable EPT write-violation livelock). Deny it instead so + * the violation is visible and can be mediated (e.g. HEKI text_poke + * is routed through the secure plane rather than written directly). + */ + { + unsigned long plane_attrs = + kvm_plane_access_attributes(vcpu->plane, fault->gfn); + + if ((plane_attrs & KVM_MEMORY_ATTRIBUTE_NO_READ) || + (fault->write && + (plane_attrs & KVM_MEMORY_ATTRIBUTE_NO_WRITE))) { + kvm_mmu_prepare_memory_fault_exit(vcpu, fault); + return -EFAULT; + } + } + if (unlikely(!slot)) return kvm_handle_noslot_fault(vcpu, fault, access); @@ -5884,6 +5912,14 @@ kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu, role.direct = true; role.has_4_byte_gpte = false; + /* + * Give each VM plane its own TDP root. Planes share memslots but + * need independent page tables so a higher-privilege plane can + * restrict a lower plane's access to a GFN. plane_level is 0 (and + * thus a no-op) on non-plane VMs and when CONFIG_VM_PLANES is off. + */ + role.plane = vcpu->plane_level; + /* All TDP pages are supervisor-executable */ role.access = ACC_ALL; if (role.cr4_smep && shadow_user_mask) diff --git a/arch/x86/kvm/mmu/spte.h b/arch/x86/kvm/mmu/spte.h index 144f7c5a1040..ed03bcdbf82d 100644 --- a/arch/x86/kvm/mmu/spte.h +++ b/arch/x86/kvm/mmu/spte.h @@ -580,9 +580,13 @@ void __init kvm_mmu_spte_module_init(void); void kvm_mmu_reset_all_pte_masks(void); /* - * Apply memory protection attributes to pte_access. - * If memory attributes have NO_WRITE or NO_EXEC set for a GFN, - * strip the corresponding access bits before building the SPTE. + * Apply cross-plane access restrictions to pte_access when building an SPTE + * for the faulting plane. A higher-privilege plane may downgrade a lower + * plane's access to a GFN via its per-plane access_attr_array. NO_WRITE and + * NO_EXEC are enforced here by stripping the corresponding access bits. + * NO_READ cannot be expressed as a present-but-unreadable SPTE on EPT, so it + * is enforced earlier in the fault handler (kvm_mmu_faultin_pfn) by refusing + * to map the page. */ #ifdef CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES static inline unsigned int kvm_plane_filter_pte_access(struct kvm_vcpu *vcpu, @@ -591,7 +595,7 @@ static inline unsigned int kvm_plane_filter_pte_access(struct kvm_vcpu *vcpu, { unsigned long attrs; - attrs = kvm_get_memory_attributes(vcpu->kvm, gfn); + attrs = kvm_plane_access_attributes(vcpu->plane, gfn); if (attrs & KVM_MEMORY_ATTRIBUTE_NO_WRITE) pte_access &= ~ACC_WRITE_MASK; if (attrs & KVM_MEMORY_ATTRIBUTE_NO_EXEC) diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c index 3c73ab1dcfe8..c8c37d569023 100644 --- a/arch/x86/kvm/x86.c +++ b/arch/x86/kvm/x86.c @@ -10494,6 +10494,61 @@ static int complete_hypercall_exit(struct kvm_vcpu *vcpu) return kvm_skip_emulated_instruction(vcpu); } +#if defined(CONFIG_VM_PLANES) && defined(CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES) +/* + * Apply cross-plane access restrictions requested by a higher-privilege plane. + * Stores @attrs (NO_READ/NO_WRITE/NO_EXEC) for [@start, @end) in @plane's + * access_attr_array and zaps the range so any pages already mapped in @plane's + * EPT re-fault and pick up the restriction. @attrs == 0 clears the + * restriction for the range. + */ +static int kvm_plane_set_access_attrs(struct kvm *kvm, struct kvm_plane *plane, + gfn_t start, gfn_t end, unsigned long attrs) +{ + void *entry = attrs ? xa_mk_value(attrs) : NULL; + gfn_t gfn; + int r = 0; + + mutex_lock(&kvm->slots_lock); + + /* + * Reserve slots up front so the store loop below cannot fail partway + * through and leave a gap (a still-readable page) in the protected + * range. Clearing a restriction (entry == NULL) never allocates. + */ + if (entry) { + for (gfn = start; gfn < end; gfn++) { + r = xa_reserve(&plane->access_attr_array, gfn, + GFP_KERNEL_ACCOUNT); + if (r) + goto out_unlock; + + cond_resched(); + } + } + + for (gfn = start; gfn < end; gfn++) { + r = xa_err(xa_store(&plane->access_attr_array, gfn, entry, + GFP_KERNEL_ACCOUNT)); + if (KVM_BUG_ON(r, kvm)) + goto out_unlock; + + cond_resched(); + } + + /* + * Re-fault the affected gfns in the plane's EPT so the new restriction + * takes effect on existing mappings. Zapping all roots is harmless; + * other planes simply rebuild identical entries on next access. + */ + kvm_zap_gfn_range(kvm, start, end); + +out_unlock: + mutex_unlock(&kvm->slots_lock); + return r; +} +#endif /* CONFIG_VM_PLANES && CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES */ + int ____kvm_emulate_hypercall(struct kvm_vcpu *vcpu, int cpl, int (*complete_hypercall)(struct kvm_vcpu *)) { @@ -10683,16 +10738,21 @@ int ____kvm_emulate_hypercall(struct kvm_vcpu *vcpu, int cpl, case KVM_HC_VBS_SET_MEM_ATTRS: #if defined(CONFIG_VM_PLANES) && defined(CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES) /* - * The secure plane (plane >0) enforces EPT permissions on the - * normal plane's memory. It cannot issue the host + * The secure plane (plane >0) enforces EPT permissions on a + * lower plane's memory. It cannot issue the host * KVM_SET_MEMORY_ATTRIBUTES ioctl, so it asks KVM to do it via - * this hypercall. Only a higher-privilege plane may call it. + * this hypercall. Only a higher-privilege plane may call it; + * the restriction is applied to the plane directly below the + * caller. * * a0 = guest-physical address (page aligned) * a1 = region size in bytes (page aligned) - * a2 = access bits to retain for lower planes: - * bit0 read (implicit), bit1 write, bit2 exec - * (matches VBS_MEM_READ/WRITE/EXEC) + * a2 = access bits to retain for the lower plane: + * bit0 read, bit1 write, bit2 exec + * (matches VBS_MEM_READ/WRITE/EXEC). A cleared bit adds + * the corresponding NO_READ/NO_WRITE/NO_EXEC restriction; + * a2 = 0 hides the range entirely (e.g. secure-plane + * memory that the normal plane must not read). */ if (vcpu->plane_level == 0) { ret = -KVM_EPERM; @@ -10704,17 +10764,26 @@ int ____kvm_emulate_hypercall(struct kvm_vcpu *vcpu, int cpl, ret = -KVM_EINVAL; goto out; } else { + struct kvm_plane *target; unsigned long attrs = 0; gfn_t start = a0 >> PAGE_SHIFT; gfn_t end = (a0 + a1) >> PAGE_SHIFT; + target = vcpu->kvm->planes[vcpu->plane_level - 1]; + if (!target) { + ret = -KVM_EINVAL; + goto out; + } + + if (!(a2 & BIT(0))) + attrs |= KVM_MEMORY_ATTRIBUTE_NO_READ; if (!(a2 & BIT(1))) attrs |= KVM_MEMORY_ATTRIBUTE_NO_WRITE; if (!(a2 & BIT(2))) attrs |= KVM_MEMORY_ATTRIBUTE_NO_EXEC; - if (kvm_vm_set_mem_attributes(vcpu->kvm, start, end, - attrs)) + if (kvm_plane_set_access_attrs(vcpu->kvm, target, start, + end, attrs)) ret = -KVM_EINVAL; else ret = 0; diff --git a/arch/x86/realmode/init.c b/arch/x86/realmode/init.c index 694d80a5c68e..01855a913b10 100644 --- a/arch/x86/realmode/init.c +++ b/arch/x86/realmode/init.c @@ -11,6 +11,7 @@ #include #include #include +#include struct real_mode_header *real_mode_header; u32 *trampoline_cr4_features; @@ -44,6 +45,31 @@ void load_trampoline_pgtable(void) __flush_tlb_all(); } +#ifdef CONFIG_VBS_SECURE_MONITOR +/* + * A KVM VM-planes secure plane (plane > 0) is entered directly in 64-bit long + * mode and boots from a single carved-out high-memory region that contains no + * RAM below 1 MiB. It runs uniprocessor with no firmware, ACPI sleep, or + * hibernation, so the 16-bit real-mode trampoline can neither be allocated + * (there is no sub-1M memory) nor is it ever used (no AP bringup or wakeup). + * + * Disable the real-mode setup the same way Hyper-V VTL and Xen PV do, by + * pointing the x86_platform real-mode hooks at the no-op handler. This is + * installed from an early_param so it takes effect before setup_arch() calls + * x86_platform.realmode_reserve(). Triggered by the "secure_monitor" + * command-line option, the same switch that activates the in-kernel + * secure-plane monitor. + */ +static int __init secure_plane_no_real_mode(char *arg) +{ + x86_platform.realmode_reserve = x86_init_noop; + x86_platform.realmode_init = x86_init_noop; + pr_info("realmode: secure plane: skipping sub-1M trampoline\n"); + return 0; +} +early_param("secure_monitor", secure_plane_no_real_mode); +#endif /* CONFIG_VBS_SECURE_MONITOR */ + void __init reserve_real_mode(void) { phys_addr_t mem, limit = x86_init.resources.realmode_limit; diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h index f14d78fd8cd3..05c9edd4a73d 100644 --- a/include/linux/kvm_host.h +++ b/include/linux/kvm_host.h @@ -895,6 +895,18 @@ struct kvm_plane { /* Per-Plane VCPU array */ struct xarray vcpu_array; +#ifdef CONFIG_VM_PLANES + /* + * Cross-plane access restrictions imposed on THIS plane by a + * higher-privilege plane. Each entry holds NO_READ/NO_WRITE/NO_EXEC + * bits for a gfn and is enforced when building this plane's SPTEs + * (planes have independent EPT roots). Distinct from + * kvm->mem_attr_array, which holds VM-wide PRIVATE/CoCo attributes. + * Protected by kvm->slots_lock for writes, RCU for reads. + */ + struct xarray access_attr_array; +#endif + struct kvm_arch_plane arch; }; @@ -2739,6 +2751,29 @@ static inline bool kvm_mem_is_private(struct kvm *kvm, gfn_t gfn) } #endif /* CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES */ +#ifdef CONFIG_VM_PLANES +/* + * Cross-plane access restrictions: a higher-privilege plane downgrades a + * lower plane's access (NO_READ/NO_WRITE/NO_EXEC) to a gfn by storing bits in + * that lower plane's access_attr_array. Enforced when building the lower + * plane's SPTEs (planes have independent EPT roots). Returns 0 when no + * restriction applies. + */ +static inline unsigned long kvm_plane_access_attributes(struct kvm_plane *plane, + gfn_t gfn) +{ + if (!plane) + return 0; + return xa_to_value(xa_load(&plane->access_attr_array, gfn)); +} +#else +static inline unsigned long kvm_plane_access_attributes(struct kvm_plane *plane, + gfn_t gfn) +{ + return 0; +} +#endif /* CONFIG_VM_PLANES */ + #ifdef CONFIG_KVM_GUEST_MEMFD int kvm_gmem_get_pfn(struct kvm *kvm, struct kvm_memory_slot *slot, gfn_t gfn, kvm_pfn_t *pfn, struct page **page, diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h index 348628c7b17e..3118b31d13f6 100644 --- a/include/uapi/linux/kvm.h +++ b/include/uapi/linux/kvm.h @@ -1688,6 +1688,7 @@ struct kvm_memory_attributes { #define KVM_MEMORY_ATTRIBUTE_PRIVATE (1ULL << 3) #define KVM_MEMORY_ATTRIBUTE_NO_WRITE (1ULL << 4) #define KVM_MEMORY_ATTRIBUTE_NO_EXEC (1ULL << 5) +#define KVM_MEMORY_ATTRIBUTE_NO_READ (1ULL << 6) #define KVM_CREATE_GUEST_MEMFD _IOWR(KVMIO, 0xd4, struct kvm_create_guest_memfd) #define GUEST_MEMFD_FLAG_MMAP (1ULL << 0) diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c index 553c282500fd..3a1a09f26340 100644 --- a/virt/kvm/kvm_main.c +++ b/virt/kvm/kvm_main.c @@ -1236,6 +1236,9 @@ static struct kvm_plane *kvm_create_plane(struct kvm *kvm, unsigned plane_level) plane->level = plane_level; xa_init(&plane->vcpu_array); +#ifdef CONFIG_VM_PLANES + xa_init(&plane->access_attr_array); +#endif if (kvm_arch_plane_init(kvm, plane, plane_level)) goto out_free_plane; @@ -1254,6 +1257,9 @@ static struct kvm_plane *kvm_create_plane(struct kvm *kvm, unsigned plane_level) static void kvm_destroy_one_plane(struct kvm_plane *plane) { kvm_arch_plane_destroy(plane); +#ifdef CONFIG_VM_PLANES + xa_destroy(&plane->access_attr_array); +#endif kvm_free_plane(plane); } -- 2.55.0