Skip to contentSkip to conversation

KVM: x86/mmu: Don't synthesize triple fault on bad root

PATCH In review Sean Christopherson 1 participant 7 messages Jul 29, 2023 – Aug 18, 2023

Subsystems (3)
lore
Share & tools
lore

Sean Christopherson

author maintainer
#0 lore raw

v2 cover KVM: x86/mmu: Don't synthesize triple fault on bad root

Rework the handling of !visible guest root gfns to wait until the guest actually tries to access memory before synthesizing a fault. KVM currently just immediately synthesizes triple fault, which causes problems for nVMX and nSVM as immediately injecting a fault causes KVM to try and forward the fault to L1 (as a VM-Exit) before completing nested VM-Enter, e.g. if L1 runs L2 with a "bad" nested TDP root.

To get around the conundrum of not wanting to shadow garbage, load a dummy root, backed by the zero page, into CR3/EPTP/nCR3, and then inject an appropriate page fault when the guest (likely) hits a !PRESENT fault.

Note, KVM's behavior is still not strictly correct with respect to x86 architecture, the primary goal is purely to prevent triggering KVM's WARN at will. No real world guest intentionally loads CR3 (or EPTP or nCR3) with a GPA that points at MMIO and expects it to work (and KVM has a long and storied history of punting on emulated MMIO corner cases).

I didn't Cc any of this for stable because syzkaller is really the only thing that I expect to care, and the whole dummy root thing isn't exactly risk free. If someone _really_ wants to squash the WARN in LTS kernels, the way to do that would be to exempt triple fault shutdown VM-Exits from the sanity checks in nVMX and nSVM, i.e. sweep the problem under the rug.

I have a KUT test for this that'll I'll post next week (I said that about v1 and then forgot).

v2:
- Finish writing the changelog for patch 3. [Yu]
- Use KVM_REQ_MMU_FREE_OBSOLETE_ROOTS instead of directly unloading all roots. [Yu]

v1: https://lore.kernel.org/all/[email protected]

Sean Christopherson (5):
KVM: x86/mmu: Add helper to convert root hpa to shadow page
KVM: x86/mmu: Harden new PGD against roots without shadow pages
KVM: x86/mmu: Harden TDP MMU iteration against root w/o shadow page
KVM: x86/mmu: Disallow guest from using !visible slots for page tables
KVM: x86/mmu: Use dummy root, backed by zero page, for !visible guest
roots

 arch/x86/kvm/mmu/mmu.c          | 94 ++++++++++++++++++---------------
 arch/x86/kvm/mmu/mmu_internal.h | 10 ++++
 arch/x86/kvm/mmu/paging_tmpl.h  | 18 ++++++-
 arch/x86/kvm/mmu/spte.h         | 12 +++++
 arch/x86/kvm/mmu/tdp_iter.c     | 11 ++--
 arch/x86/kvm/mmu/tdp_mmu.c      |  2 +-
 6 files changed, 98 insertions(+), 49 deletions(-)

base-commit: fdf0eaf11452d72945af31804e2a1048ee1b574c

Signature · 1 line
2.41.0.487.g6d72f3e995-goog
author maintainer
#1 lore raw

PATCH v2 1/5 KVM: x86/mmu: Add helper to convert root hpa to shadow page

Add a dedicated helper for converting a root hpa to a shadow page in anticipation of using a "dummy" root to handle the scenario where KVM needs to load a valid shadow root (from hardware's perspective), but the guest doesn't have a visible root to shadow. Similar to PAE roots, the dummy root won't have an associated kvm_mmu_page and will need special handling when finding a shadow page given a root.

Opportunistically retrieve the root shadow page in kvm_mmu_sync_roots() *after* verifying the root is unsync (the dummy root can never be unsync).

---
 arch/x86/kvm/mmu/mmu.c     | 28 +++++++++++++---------------
 arch/x86/kvm/mmu/spte.h    |  9 +++++++++
 arch/x86/kvm/mmu/tdp_mmu.c |  2 +-
 3 files changed, 23 insertions(+), 16 deletions(-)
arch/x86/kvm/mmu/mmu.c +13 -15
index ec169f5c7dce..1eadfcde30be 100644
@@ -3574,11 +3574,7 @@ static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
 	if (!VALID_PAGE(*root_hpa))
 		return;
 
-	/*
-	 * The "root" may be a special root, e.g. a PAE entry, treat it as a
-	 * SPTE to ensure any non-PA bits are dropped.
-	 */
-	sp = spte_to_child_sp(*root_hpa);
+	sp = root_to_sp(*root_hpa);
 	if (WARN_ON(!sp))
 		return;
 
@@ -3624,7 +3620,7 @@ void kvm_mmu_free_roots(struct kvm *kvm, struct kvm_mmu *mmu,
 					   &invalid_list);
 
 	if (free_active_root) {
-		if (to_shadow_page(mmu->root.hpa)) {
+		if (root_to_sp(mmu->root.hpa)) {
 			mmu_free_root_page(kvm, &mmu->root.hpa, &invalid_list);
 		} else if (mmu->pae_root) {
 			for (i = 0; i < 4; ++i) {
@@ -3648,6 +3644,7 @@ EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
 void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu)
 {
 	unsigned long roots_to_free = 0;
+	struct kvm_mmu_page *sp;
 	hpa_t root_hpa;
 	int i;
 
@@ -3662,8 +3659,8 @@ void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu)
 		if (!VALID_PAGE(root_hpa))
 			continue;
 
-		if (!to_shadow_page(root_hpa) ||
-			to_shadow_page(root_hpa)->role.guest_mode)
+		sp = root_to_sp(root_hpa);
+		if (!sp || sp->role.guest_mode)
 			roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
 	}
 
@@ -4018,7 +4015,7 @@ static bool is_unsync_root(hpa_t root)
 	 * requirement isn't satisfied.
 	 */
 	smp_rmb();
-	sp = to_shadow_page(root);
+	sp = root_to_sp(root);
 
 	/*
 	 * PAE roots (somewhat arbitrarily) aren't backed by shadow pages, the
@@ -4048,11 +4045,12 @@ void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
 
 	if (vcpu->arch.mmu->cpu_role.base.level >= PT64_ROOT_4LEVEL) {
 		hpa_t root = vcpu->arch.mmu->root.hpa;
-		sp = to_shadow_page(root);
 
 		if (!is_unsync_root(root))
 			return;
 
+		sp = root_to_sp(root);
+
 		write_lock(&vcpu->kvm->mmu_lock);
 		mmu_sync_children(vcpu, sp, true);
 		write_unlock(&vcpu->kvm->mmu_lock);
@@ -4382,7 +4380,7 @@ static int kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
 static bool is_page_fault_stale(struct kvm_vcpu *vcpu,
 				struct kvm_page_fault *fault)
 {
-	struct kvm_mmu_page *sp = to_shadow_page(vcpu->arch.mmu->root.hpa);
+	struct kvm_mmu_page *sp = root_to_sp(vcpu->arch.mmu->root.hpa);
 
 	/* Special roots, e.g. pae_root, are not backed by shadow pages. */
 	if (sp && is_obsolete_sp(vcpu->kvm, sp))
@@ -4564,7 +4562,7 @@ static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd,
 {
 	return (role.direct || pgd == root->pgd) &&
 	       VALID_PAGE(root->hpa) &&
-	       role.word == to_shadow_page(root->hpa)->role.word;
+	       role.word == root_to_sp(root->hpa)->role.word;
 }
 
 /*
@@ -4638,7 +4636,7 @@ static bool fast_pgd_switch(struct kvm *kvm, struct kvm_mmu *mmu,
 	 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
 	 * later if necessary.
 	 */
-	if (VALID_PAGE(mmu->root.hpa) && !to_shadow_page(mmu->root.hpa))
+	if (VALID_PAGE(mmu->root.hpa) && !root_to_sp(mmu->root.hpa))
 		kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT);
 
 	if (VALID_PAGE(mmu->root.hpa))
@@ -4686,7 +4684,7 @@ void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd)
 	 */
 	if (!new_role.direct)
 		__clear_sp_write_flooding_count(
-				to_shadow_page(vcpu->arch.mmu->root.hpa));
+				root_to_sp(vcpu->arch.mmu->root.hpa));
 }
 EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd);
 
@@ -5555,7 +5553,7 @@ static bool is_obsolete_root(struct kvm *kvm, hpa_t root_hpa)
 	 *  (c) KVM doesn't track previous roots for PAE paging, and the guest
 	 *      is unlikely to zap an in-use PGD.
 	 */
-	sp = to_shadow_page(root_hpa);
+	sp = root_to_sp(root_hpa);
 	return !sp || is_obsolete_sp(kvm, sp);
 }
 
arch/x86/kvm/mmu/spte.h +9 -0
index 1279db2eab44..9f8e8cda89e8 100644
@@ -236,6 +236,15 @@ static inline struct kvm_mmu_page *sptep_to_sp(u64 *sptep)
 	return to_shadow_page(__pa(sptep));
 }
 
+static inline struct kvm_mmu_page *root_to_sp(hpa_t root)
+{
+	/*
+	 * The "root" may be a special root, e.g. a PAE entry, treat it as a
+	 * SPTE to ensure any non-PA bits are dropped.
+	 */
+	return spte_to_child_sp(root);
+}
+
 static inline bool is_mmio_spte(u64 spte)
 {
 	return (spte & shadow_mmio_mask) == shadow_mmio_value &&
arch/x86/kvm/mmu/tdp_mmu.c +1 -1
index 512163d52194..046ac2589611 100644
@@ -689,7 +689,7 @@ static inline void tdp_mmu_iter_set_spte(struct kvm *kvm, struct tdp_iter *iter,
 		else
 
 #define tdp_mmu_for_each_pte(_iter, _mmu, _start, _end)		\
-	for_each_tdp_pte(_iter, to_shadow_page(_mmu->root.hpa), _start, _end)
+	for_each_tdp_pte(_iter, root_to_sp(_mmu->root.hpa), _start, _end)
 
 /*
  * Yield if the MMU lock is contended or this thread needs to return control
Signature · 1 line
2.41.0.487.g6d72f3e995-goog
author maintainer
#2 lore raw

PATCH v2 2/5 KVM: x86/mmu: Harden new PGD against roots without shadow pages

In reply to Sean Christopherson

Harden kvm_mmu_new_pgd() against NULL pointer dereference bugs by sanity checking that the target root has an associated shadow page prior to dereferencing said shadow page. The code in question is guaranteed to only see roots with shadow pages as fast_pgd_switch() explicitly frees the current root if it doesn't have a shadow page, i.e. is a PAE root, and that in turn prevents valid roots from being cached, but that's all very subtle.

---
 arch/x86/kvm/mmu/mmu.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)
arch/x86/kvm/mmu/mmu.c +19 -6
index 1eadfcde30be..dd8cc46551b2 100644
@@ -4560,9 +4560,19 @@ static void nonpaging_init_context(struct kvm_mmu *context)
 static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd,
 				  union kvm_mmu_page_role role)
 {
-	return (role.direct || pgd == root->pgd) &&
-	       VALID_PAGE(root->hpa) &&
-	       role.word == root_to_sp(root->hpa)->role.word;
+	struct kvm_mmu_page *sp;
+
+	if (!VALID_PAGE(root->hpa))
+		return false;
+
+	if (!role.direct && pgd != root->pgd)
+		return false;
+
+	sp = root_to_sp(root->hpa);
+	if (WARN_ON_ONCE(!sp))
+		return false;
+
+	return role.word == sp->role.word;
 }
 
 /*
@@ -4682,9 +4692,12 @@ void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd)
 	 * If this is a direct root page, it doesn't have a write flooding
 	 * count. Otherwise, clear the write flooding count.
 	 */
-	if (!new_role.direct)
-		__clear_sp_write_flooding_count(
-				root_to_sp(vcpu->arch.mmu->root.hpa));
+	if (!new_role.direct) {
+		struct kvm_mmu_page *sp = root_to_sp(vcpu->arch.mmu->root.hpa);
+
+		if (!WARN_ON_ONCE(!sp))
+			__clear_sp_write_flooding_count(sp);
+	}
 }
 EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd);
 
Signature · 1 line
2.41.0.487.g6d72f3e995-goog
author maintainer
#3 lore raw

PATCH v2 3/5 KVM: x86/mmu: Harden TDP MMU iteration against root w/o shadow page

In reply to Sean Christopherson

Explicitly check that tdp_iter_start() is handed a valid shadow page to harden KVM against bugs, e.g. if KVM calls into the TDP MMU with an invalid or shadow MMU root (which would be a fatal KVM bug), the shadow page pointer will be NULL.

Opportunistically stop the TDP MMU iteration instead of continuing on with garbage if the incoming root is bogus. Attempting to walk a garbage root is more likely to caused major problems than doing nothing.

---
 arch/x86/kvm/mmu/tdp_iter.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
arch/x86/kvm/mmu/tdp_iter.c +6 -5
index d2eb0d4f8710..bd30ebfb2f2c 100644
@@ -39,13 +39,14 @@ void tdp_iter_restart(struct tdp_iter *iter)
 void tdp_iter_start(struct tdp_iter *iter, struct kvm_mmu_page *root,
 		    int min_level, gfn_t next_last_level_gfn)
 {
-	int root_level = root->role.level;
-
-	WARN_ON(root_level < 1);
-	WARN_ON(root_level > PT64_ROOT_MAX_LEVEL);
+	if (WARN_ON_ONCE(!root || (root->role.level < 1) ||
+			 (root->role.level > PT64_ROOT_MAX_LEVEL))) {
+		iter->valid = false;
+		return;
+	}
 
 	iter->next_last_level_gfn = next_last_level_gfn;
-	iter->root_level = root_level;
+	iter->root_level = root->role.level;
 	iter->min_level = min_level;
 	iter->pt_path[iter->root_level - 1] = (tdp_ptep_t)root->spt;
 	iter->as_id = kvm_mmu_page_as_id(root);
Signature · 1 line
2.41.0.487.g6d72f3e995-goog
author maintainer
#4 lore raw

PATCH v2 4/5 KVM: x86/mmu: Disallow guest from using !visible slots for page tables

In reply to Sean Christopherson

Explicitly inject a page fault if guest attempts to use a !visible gfn as a page table. kvm_vcpu_gfn_to_hva_prot() will naturally handle the case where there is no memslot, but doesn't catch the scenario where the gfn points at a KVM-internal memslot.

Letting the guest backdoor its way into accessing KVM-internal memslots isn't dangerous on its own, e.g. at worst the guest can crash itself, but disallowing the behavior will simplify fixing how KVM handles !visible guest root gfns (immediately synthesizing a triple fault when loading the root is architecturally wrong).

---
 arch/x86/kvm/mmu/paging_tmpl.h | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
arch/x86/kvm/mmu/paging_tmpl.h +6 -1
index 0662e0278e70..122bfc0124d3 100644
@@ -351,6 +351,7 @@ static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 	++walker->level;
 
 	do {
+		struct kvm_memory_slot *slot;
 		unsigned long host_addr;
 
 		pt_access = pte_access;
@@ -381,7 +382,11 @@ static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 		if (unlikely(real_gpa == INVALID_GPA))
 			return 0;
 
-		host_addr = kvm_vcpu_gfn_to_hva_prot(vcpu, gpa_to_gfn(real_gpa),
+		slot = kvm_vcpu_gfn_to_memslot(vcpu, gpa_to_gfn(real_gpa));
+		if (!kvm_is_visible_memslot(slot))
+			goto error;
+
+		host_addr = gfn_to_hva_memslot_prot(slot, gpa_to_gfn(real_gpa),
 					    &walker->pte_writable[walker->level - 1]);
 		if (unlikely(kvm_is_error_hva(host_addr)))
 			goto error;
Signature · 1 line
2.41.0.487.g6d72f3e995-goog
author maintainer
#5 lore raw

PATCH v2 5/5 KVM: x86/mmu: Use dummy root, backed by zero page, for !visible guest roots

In reply to Sean Christopherson

When attempting to allocate a shadow root for a !visible guest root gfn, e.g. that resides in MMIO space, load a dummy root that is backed by the zero page instead of immediately synthesizing a triple fault shutdown (using the zero page ensures any attempt to translate memory will generate a !PRESENT fault and thus VM-Exit).

Unless the vCPU is racing with memslot activity, KVM will inject a page fault due to not finding a visible slot in FNAME(walk_addr_generic), i.e. the end result is mostly same, but critically KVM will inject a fault only *after* KVM runs the vCPU with the bogus root.

Waiting to inject a fault until after running the vCPU fixes a bug where KVM would bail from nested VM-Enter if L1 tried to run L2 with TDP enabled and a !visible root. Even though a bad root will *probably* lead to shutdown, (a) it's not guaranteed and (b) the CPU won't read the underlying memory until after VM-Enter succeeds. E.g. if L1 runs L2 with a VMX preemption timer value of '0', then architecturally the preemption timer VM-Exit is guaranteed to occur before the CPU executes any instruction, i.e. before the CPU needs to translate a GPA to a HPA (so long as there are no injected events with higher priority than the preemption timer).

If KVM manages to get to FNAME(fetch) with a dummy root, e.g. because userspace created a memslot between installing the dummy root and handling the page fault, simply unload the MMU to allocate a new root and retry the instruction. Use KVM_REQ_MMU_FREE_OBSOLETE_ROOTS to drop the root, as invoking kvm_mmu_free_roots() while holding mmu_lock would deadlock, and conceptually the dummy root has indeeed become obsolete. The only difference versus existing usage of KVM_REQ_MMU_FREE_OBSOLETE_ROOTS is that the root has become obsolete due to memslot *creation*, not memslot deletion or movement.

Cc Yu Zhang <[email protected]>

---
 arch/x86/kvm/mmu/mmu.c          | 47 ++++++++++++++++-----------------
 arch/x86/kvm/mmu/mmu_internal.h | 10 +++++++
 arch/x86/kvm/mmu/paging_tmpl.h  | 11 ++++++++
 arch/x86/kvm/mmu/spte.h         |  3 +++
 4 files changed, 47 insertions(+), 24 deletions(-)
arch/x86/kvm/mmu/mmu.c +23 -24
index dd8cc46551b2..565988c81b57 100644
@@ -3620,7 +3620,9 @@ void kvm_mmu_free_roots(struct kvm *kvm, struct kvm_mmu *mmu,
 					   &invalid_list);
 
 	if (free_active_root) {
-		if (root_to_sp(mmu->root.hpa)) {
+		if (kvm_mmu_is_dummy_root(mmu->root.hpa)) {
+			/* Nothing to cleanup for dummy roots. */
+		} else if (root_to_sp(mmu->root.hpa)) {
 			mmu_free_root_page(kvm, &mmu->root.hpa, &invalid_list);
 		} else if (mmu->pae_root) {
 			for (i = 0; i < 4; ++i) {
@@ -3668,19 +3670,6 @@ void kvm_mmu_free_guest_mode_roots(struct kvm *kvm, struct kvm_mmu *mmu)
 }
 EXPORT_SYMBOL_GPL(kvm_mmu_free_guest_mode_roots);
 
-
-static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
-{
-	int ret = 0;
-
-	if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) {
-		kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
-		ret = 1;
-	}
-
-	return ret;
-}
-
 static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, int quadrant,
 			    u8 level)
 {
@@ -3818,8 +3807,10 @@ static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
 	root_pgd = kvm_mmu_get_guest_pgd(vcpu, mmu);
 	root_gfn = root_pgd >> PAGE_SHIFT;
 
-	if (mmu_check_root(vcpu, root_gfn))
-		return 1;
+	if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) {
+		mmu->root.hpa = kvm_mmu_get_dummy_root();
+		return 0;
+	}
 
 	/*
 	 * On SVM, reading PDPTRs might access guest memory, which might fault
@@ -3831,8 +3822,8 @@ static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
 			if (!(pdptrs[i] & PT_PRESENT_MASK))
 				continue;
 
-			if (mmu_check_root(vcpu, pdptrs[i] >> PAGE_SHIFT))
-				return 1;
+			if (!kvm_vcpu_is_visible_gfn(vcpu, pdptrs[i] >> PAGE_SHIFT))
+				pdptrs[i] = 0;
 		}
 	}
 
@@ -3999,7 +3990,7 @@ static bool is_unsync_root(hpa_t root)
 {
 	struct kvm_mmu_page *sp;
 
-	if (!VALID_PAGE(root))
+	if (!VALID_PAGE(root) || kvm_mmu_is_dummy_root(root))
 		return false;
 
 	/*
@@ -4405,6 +4396,10 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault
 {
 	int r;
 
+	/* Dummy roots are used only for shadowing bad guest roots. */
+	if (WARN_ON_ONCE(kvm_mmu_is_dummy_root(vcpu->arch.mmu->root.hpa)))
+		return RET_PF_RETRY;
+
 	if (page_fault_handle_page_track(vcpu, fault))
 		return RET_PF_EMULATE;
 
@@ -4642,9 +4637,8 @@ static bool fast_pgd_switch(struct kvm *kvm, struct kvm_mmu *mmu,
 			    gpa_t new_pgd, union kvm_mmu_page_role new_role)
 {
 	/*
-	 * For now, limit the caching to 64-bit hosts+VMs in order to avoid
-	 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
-	 * later if necessary.
+	 * Limit reuse to 64-bit hosts+VMs without "special" roots in order to
+	 * avoid having to deal with PDPTEs and other complexities.
 	 */
 	if (VALID_PAGE(mmu->root.hpa) && !root_to_sp(mmu->root.hpa))
 		kvm_mmu_free_roots(kvm, mmu, KVM_MMU_ROOT_CURRENT);
@@ -5557,14 +5551,19 @@ static bool is_obsolete_root(struct kvm *kvm, hpa_t root_hpa)
 
 	/*
 	 * When freeing obsolete roots, treat roots as obsolete if they don't
-	 * have an associated shadow page.  This does mean KVM will get false
+	 * have an associated shadow page, as it's impossible to determine if
+	 * such roots are fresh or stale.  This does mean KVM will get false
 	 * positives and free roots that don't strictly need to be freed, but
 	 * such false positives are relatively rare:
 	 *
-	 *  (a) only PAE paging and nested NPT has roots without shadow pages
+	 *  (a) only PAE paging and nested NPT have roots without shadow pages
+	 *      (or any shadow paging flavor with a dummy root, see note below)
 	 *  (b) remote reloads due to a memslot update obsoletes _all_ roots
 	 *  (c) KVM doesn't track previous roots for PAE paging, and the guest
 	 *      is unlikely to zap an in-use PGD.
+	 *
+	 * Note!  Dummy roots are unique in that they are obsoleted by memslot
+	 * _creation_!  See also FNAME(fetch).
 	 */
 	sp = root_to_sp(root_hpa);
 	return !sp || is_obsolete_sp(kvm, sp);
arch/x86/kvm/mmu/mmu_internal.h +10 -0
index d39af5639ce9..3ca986450393 100644
@@ -44,6 +44,16 @@ extern bool dbg;
 #define INVALID_PAE_ROOT	0
 #define IS_VALID_PAE_ROOT(x)	(!!(x))
 
+static inline hpa_t kvm_mmu_get_dummy_root(void)
+{
+	return my_zero_pfn(0) << PAGE_SHIFT;
+}
+
+static inline bool kvm_mmu_is_dummy_root(hpa_t shadow_page)
+{
+	return is_zero_pfn(shadow_page >> PAGE_SHIFT);
+}
+
 typedef u64 __rcu *tdp_ptep_t;
 
 struct kvm_mmu_page {
arch/x86/kvm/mmu/paging_tmpl.h +11 -0
index 122bfc0124d3..caf7623e5f91 100644
@@ -646,6 +646,17 @@ static int FNAME(fetch)(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
 	if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root.hpa)))
 		goto out_gpte_changed;
 
+	/*
+	 * Load a new root and retry the faulting instruction in the extremely
+	 * unlikely scenario that the guest root gfn became visible between
+	 * loading a dummy root and handling the resulting page fault, e.g. if
+	 * userspace create a memslot in the interim.
+	 */
+	if (unlikely(kvm_mmu_is_dummy_root(vcpu->arch.mmu->root.hpa))) {
+		kvm_make_request(KVM_REQ_MMU_FREE_OBSOLETE_ROOTS, vcpu);
+		goto out_gpte_changed;
+	}
+
 	for_each_shadow_entry(vcpu, fault->addr, it) {
 		gfn_t table_gfn;
 
arch/x86/kvm/mmu/spte.h +3 -0
index 9f8e8cda89e8..ac8ad12f9698 100644
@@ -238,6 +238,9 @@ static inline struct kvm_mmu_page *sptep_to_sp(u64 *sptep)
 
 static inline struct kvm_mmu_page *root_to_sp(hpa_t root)
 {
+	if (kvm_mmu_is_dummy_root(root))
+		return NULL;
+
 	/*
 	 * The "root" may be a special root, e.g. a PAE entry, treat it as a
 	 * SPTE to ensure any non-PA bits are dropped.
Signature · 1 line
2.41.0.487.g6d72f3e995-goog
author maintainer
#6 lore raw
In reply to Sean Christopherson
Quoted 8 lines · Sean Christopherson

Rework the handling of !visible guest root gfns to wait until the guest actually tries to access memory before synthesizing a fault. KVM currently just immediately synthesizes triple fault, which causes problems for nVMX and nSVM as immediately injecting a fault causes KVM to try and forward the fault to L1 (as a VM-Exit) before completing nested VM-Enter, e.g. if L1 runs L2 with a "bad" nested TDP root.

[...]

With some trepidation that I'm overlooking something, applied to kvm-x86 mmu. Patches 1-4 are worthwhile on their own, so even if the actual fix is wildly broken somehow and needs to be reverted, it'll just be that one commit that gets nuked.

[1/5] KVM: x86/mmu: Add helper to convert root hpa to shadow page

      https://github.com/kvm-x86/linux/commit/732f57612d5c
[2/5] KVM: x86/mmu: Harden new PGD against roots without shadow pages
      https://github.com/kvm-x86/linux/commit/9e3f832edfca
[3/5] KVM: x86/mmu: Harden TDP MMU iteration against root w/o shadow page
      https://github.com/kvm-x86/linux/commit/004c297c327f
[4/5] KVM: x86/mmu: Disallow guest from using !visible slots for page tables
      https://github.com/kvm-x86/linux/commit/81d4621b7d9f
[5/5] KVM: x86/mmu: Use dummy root, backed by zero page, for !visible guest roots
      https://github.com/kvm-x86/linux/commit/a328a359d99b

--
https://github.com/kvm-x86/linux/tree/next
https://github.com/kvm-x86/linux/tree/fixes