summaryrefslogtreecommitdiff
path: root/kernel/debug
diff options
context:
space:
mode:
authorNarayana Murty N <nnmlinux@linux.ibm.com>2026-11-06 13:37:43 +0530
committerMadhavan Srinivasan <maddy@linux.ibm.com>2026-11-21 13:16:10 +0530
commitbd6d2f87c2239d10bf5f37ab52b3968565528417 (patch)
tree03d72117af0f87ca528e57db32136a2d9a9dc2ce /kernel/debug
parent0e3d5c0408262aef599df888e9bde88d0337c1dc (diff)
KVM: PPC: Fix TOCTOU kernel stack disclosure via kvm_run->mmio.len
struct kvm_run is mmap'd and writable by host userspace at any time. kvmppc_complete_mmio_load() re-reads run->mmio.len from the shared page on the KVM_RUN completion path. A malicious userspace can: 1. Trigger a guest MMIO load; the kernel sets a valid run->mmio.len (1, 2, 4, or 8) and exits with mmio_needed=1. 2. Rewrite run->mmio.len to an invalid value (0, 3, 5, 6, or 7) before the next KVM_RUN ioctl. 3. On re-entry, kvm_arch_vcpu_ioctl_run() calls kvmppc_complete_mmio_load() unconditionally. 4. No switch case matches the tampered length, so 'gpr' remains uninitialised on the kernel stack. 5. The uninitialised value is written into a guest GPR and is subsequently readable by the attacker via KVM_GET_REGS or KVM_GET_ONE_REG, disclosing kernel stack contents (CWE-457/200). The same issue affects every site that advances paddr_accessed or passes the length to the VSX/VMX repeat-load/store helpers, since all of them re-read run->mmio.len from the shared mapping. Fix by introducing vcpu->arch.mmio_len, a kernel-private u8 field that is written once in __kvmppc_handle_load() and kvmppc_handle_store() alongside the other vcpu->arch.mmio_* fields and never touched by userspace. All consumption sites are updated to use vcpu->arch.mmio_len: - kvmppc_complete_mmio_load(): reads vcpu->arch.mmio_len; validates it is exactly one of {1,2,4,8} (EMULATE_FAIL otherwise); initialises gpr = 0 before the switch so no uninitialised value can escape even if future code changes introduce a new path. Return type changed from void to int (EMULATE_DONE / EMULATE_FAIL) so callers can propagate failures. - kvm_arch_vcpu_ioctl_run(): checks the return value and exits with KVM_EXIT_INTERNAL_ERROR / KVM_INTERNAL_ERROR_EMULATION on failure, matching the VSX/VMX error paths already in place. run->internal.ndata is initialised to 0 so the union field does not carry stale mmio data. Clear mmio_needed before completing an in-kernel MMIO load or a userspace MMIO load on KVM_RUN re-entry, so a completion failure cannot leave stale pending-MMIO state. - kvmppc_handle_vsx_load(), kvmppc_handle_vsx_store(), kvmppc_handle_vmx_load(), kvmppc_handle_vmx_store(): paddr_accessed increments use vcpu->arch.mmio_len. - kvmppc_emulate_mmio_vsx_loadstore() and kvmppc_emulate_mmio_vmx_loadstore() use vcpu->arch.mmio_len for the paddr_accessed increment and for the length passed to the load/store helpers. Fixes: bbf45ba57eae ("KVM: ppc: PowerPC 440 KVM implementation") Tested-by: R Nageswara Sastry <rnsastry@linux.ibm.com> Signed-off-by: Narayana Murty N <nnmlinux@linux.ibm.com> Signed-off-by: Gautam Menghani <gautam@linux.ibm.com> Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Diffstat (limited to 'kernel/debug')
0 files changed, 0 insertions, 0 deletions