Race condition in KVM gmap invalidation / shadow paging (memory safety and isolation risk)

Note: A reliable runnable PoC for this kernel-level race is not readily practical in a short snippet due to the need to orchestrate kernel-internal gmap/shadow-pagetable races within a KVM guest/host context. Below is a high-level PoC outline and steps that could guide a reproduce attempt in a controlled lab with a patched kernel containing this fix. PoC outline (high level, not a runnable exploit): 1) Environment: a Linux host with virtualization support, running a guest via KVM on an affected kernel prior to this patch. Patch the host kernel with the commit addressed here and boot. 2) Prepare a guest workload that repeatedly exercises gmap shadow paging (shadow CRSTE/PTE paths) and can trigger partial gmap invalidations while shadow updates are in-flight. 3) Create two kernel threads in the host that operate on the same gmap: - Thread A: perform partial gmap invalidation on a shadow gmap (simulating an invalidation race). - Thread B: concurrently perform shadow-page updates (reading/writing PTE/CRSTE) that rely on sg->parent state. 4) Observe outcomes: - Before the fix: races could lead to inconsistent checks around sg->parent vs sg->invalidated, potentially causing EAGAIN paths to be skipped or misused pointers, in rare cases enabling memory safety violations or guest memory isolation leaks. - After the fix: sg->invalidated is checked first in critical paths, and invalidated transitions help prevent use-after-free-like states; the race is mitigated. 5) Validation signals: - Look for crashes, use-after-free warnings, or unexpected EAGAIN returns that correlate with concurrent invalidation/shadow events. - Optional: instrument the kernel with ftrace/dtrace to confirm the invalidated flag gating in the hot paths (e.g., _do_shadow_pte, _do_shadow_crste, _gaccess_do_shadow). Note: This PoC outline assumes access to a patched kernel and the ability to orchestrate kernel-internal race conditions. A small, safe PoC in user space is not feasible due to the need to touch kernel KVM shadow/gmap internals.

diff --git a/arch/s390/kvm/gaccess.c b/arch/s390/kvm/gaccess.c index 53a8550e7102e8..290e03a13a9562 100644 --- a/arch/s390/kvm/gaccess.c +++ b/arch/s390/kvm/gaccess.c @@ -1449,7 +1449,7 @@ static int _do_shadow_pte(struct gmap *sg, gpa_t raddr, union pte *ptep_h, union pgste_set_unlock(ptep_h, pgste); if (rc) return rc; - if (!sg->parent) + if (sg->invalidated) return -EAGAIN; newpte = _pte(f->pfn, 0, !p, 0); @@ -1479,7 +1479,7 @@ static int _do_shadow_crste(struct gmap *sg, gpa_t raddr, union crste *host, uni do { /* _gmap_crstep_xchg_atomic() could have unshadowed this shadow gmap */ - if (!sg->parent) + if (sg->invalidated) return -EAGAIN; oldcrste = READ_ONCE(*host); newcrste = _crste_fc1(f->pfn, oldcrste.h.tt, f->writable, !p); @@ -1492,7 +1492,7 @@ static int _do_shadow_crste(struct gmap *sg, gpa_t raddr, union crste *host, uni if (!newcrste.h.p && !f->writable) return -EOPNOTSUPP; } while (!_gmap_crstep_xchg_atomic(sg->parent, host, oldcrste, newcrste, f->gfn, false)); - if (!sg->parent) + if (sg->invalidated) return -EAGAIN; newcrste = _crste_fc1(f->pfn, oldcrste.h.tt, 0, !p); @@ -1545,7 +1545,7 @@ static int _gaccess_do_shadow(struct kvm_s390_mmu_cache *mc, struct gmap *sg, entries[i].pfn, i + 1, entries[i].writable); if (rc) return rc; - if (!sg->parent) + if (sg->invalidated) return -EAGAIN; } @@ -1601,6 +1601,7 @@ static inline int _gaccess_shadow_fault(struct kvm_vcpu *vcpu, struct gmap *sg, scoped_guard(spinlock, &parent->children_lock) { if (READ_ONCE(sg->parent) != parent) return -EAGAIN; + sg->invalidated = false; rc = _gaccess_do_shadow(vcpu->arch.mc, sg, saddr, walk); } if (rc == -ENOMEM) diff --git a/arch/s390/kvm/gmap.c b/arch/s390/kvm/gmap.c index 645c32c767d24b..0111d31e038656 100644 --- a/arch/s390/kvm/gmap.c +++ b/arch/s390/kvm/gmap.c @@ -181,6 +181,7 @@ void gmap_remove_child(struct gmap *child) list_del(&child->list); child->parent = NULL; + child->invalidated = true; } /** @@ -1069,6 +1070,7 @@ static void gmap_unshadow_level(struct gmap *sg, gfn_t r_gfn, int level) if (level > TABLE_TYPE_PAGE_TABLE) align = 1UL << (11 * level + _SEGMENT_SHIFT); kvm_s390_vsie_gmap_notifier(sg, ALIGN_DOWN(gaddr, align), ALIGN(gaddr + 1, align)); + sg->invalidated = true; if (dat_entry_walk(NULL, r_gfn, sg->asce, 0, level, &crstep, &ptep)) return; if (ptep) { @@ -1174,6 +1176,7 @@ static inline int __gmap_protect_asce_top_level(struct kvm_s390_mmu_cache *mc, s scoped_guard(spinlock, &parent->children_lock) { if (READ_ONCE(sg->parent) != parent) return -EAGAIN; + sg->invalidated = false; for (i = 0; i < CRST_TABLE_PAGES; i++) { if (!context->f[i].valid) continue; diff --git a/arch/s390/kvm/gmap.h b/arch/s390/kvm/gmap.h index 579399ef54803d..31ea13fda142bc 100644 --- a/arch/s390/kvm/gmap.h +++ b/arch/s390/kvm/gmap.h @@ -60,6 +60,7 @@ enum gmap_flags { struct gmap { unsigned long flags; unsigned char edat_level; + bool invalidated; struct kvm *kvm; union asce asce; struct list_head list; diff --git a/arch/x86/include/uapi/asm/kvm.h b/arch/x86/include/uapi/asm/kvm.h index 0d4538fa6c31ab..5f2b30d0405c87 100644 --- a/arch/x86/include/uapi/asm/kvm.h +++ b/arch/x86/include/uapi/asm/kvm.h @@ -197,13 +197,13 @@ struct kvm_msrs { __u32 nmsrs; /* number of msrs in entries */ __u32 pad; - struct kvm_msr_entry entries[]; + __DECLARE_FLEX_ARRAY(struct kvm_msr_entry, entries); }; /* for KVM_GET_MSR_INDEX_LIST */ struct kvm_msr_list { __u32 nmsrs; /* number of msrs in entries */ - __u32 indices[]; + __DECLARE_FLEX_ARRAY(__u32, indices); }; /* Maximum size of any access bitmap in bytes */ @@ -245,7 +245,7 @@ struct kvm_cpuid_entry { struct kvm_cpuid { __u32 nent; __u32 padding; - struct kvm_cpuid_entry entries[]; + __DECLARE_FLEX_ARRAY(struct kvm_cpuid_entry, entries); }; struct kvm_cpuid_entry2 { @@ -267,7 +267,7 @@ struct kvm_cpuid_entry2 { struct kvm_cpuid2 { __u32 nent; __u32 padding; - struct kvm_cpuid_entry2 entries[]; + __DECLARE_FLEX_ARRAY(struct kvm_cpuid_entry2, entries); }; /* for KVM_GET_PIT and KVM_SET_PIT */ @@ -398,7 +398,7 @@ struct kvm_xsave { * the contents of CPUID leaf 0xD on the host. */ __u32 region[1024]; - __u32 extra[]; + __DECLARE_FLEX_ARRAY(__u32, extra); }; #define KVM_MAX_XCRS 16 @@ -566,7 +566,7 @@ struct kvm_pmu_event_filter { __u32 fixed_counter_bitmap; __u32 flags; __u32 pad[4]; - __u64 events[]; + __DECLARE_FLEX_ARRAY(__u64, events); }; #define KVM_PMU_EVENT_ALLOW 0 diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h index 80364d4dbebb0c..3f0d8d3c3dafd0 100644 --- a/include/uapi/linux/kvm.h +++ b/include/uapi/linux/kvm.h @@ -11,6 +11,7 @@ #include <linux/const.h> #include <linux/types.h> #include <linux/compiler.h> +#include <linux/stddef.h> #include <linux/ioctl.h> #include <asm/kvm.h> @@ -542,7 +543,7 @@ struct kvm_coalesced_mmio { struct kvm_coalesced_mmio_ring { __u32 first, last; - struct kvm_coalesced_mmio coalesced_mmio[]; + __DECLARE_FLEX_ARRAY(struct kvm_coalesced_mmio, coalesced_mmio); }; #define KVM_COALESCED_MMIO_MAX \ @@ -592,7 +593,7 @@ struct kvm_clear_dirty_log { /* for KVM_SET_SIGNAL_MASK */ struct kvm_signal_mask { __u32 len; - __u8 sigset[]; + __DECLARE_FLEX_ARRAY(__u8, sigset); }; /* for KVM_TPR_ACCESS_REPORTING */ @@ -1051,7 +1052,7 @@ struct kvm_irq_routing_entry { struct kvm_irq_routing { __u32 nr; __u32 flags; - struct kvm_irq_routing_entry entries[]; + __DECLARE_FLEX_ARRAY(struct kvm_irq_routing_entry, entries); }; #define KVM_IRQFD_FLAG_DEASSIGN (1 << 0) @@ -1142,7 +1143,7 @@ struct kvm_dirty_tlb { struct kvm_reg_list { __u64 n; /* number of regs */ - __u64 reg[]; + __DECLARE_FLEX_ARRAY(__u64, reg); }; struct kvm_one_reg { @@ -1608,7 +1609,7 @@ struct kvm_stats_desc { #ifdef __KERNEL__ char name[KVM_STATS_NAME_SIZE]; #else - char name[]; + __DECLARE_FLEX_ARRAY(char, name); #endif };