The address translation logic in get_physical_address() will currently truncate physical addresses to 32 bits unless long mode is enabled. This is incorrect when using physical address extensions (PAE) outside of long mode, with the result that a 32-bit operating system using PAE to access memory above 4G will experience undefined behaviour. Instead, truncation must be applied only to non-paging mode, because all paths that go through page table accesses already produce a correctly-masked address.
Furthermore, when inspecting the code I noticed that the A20 mask is applied incorrectly when NPT is active. The mask should not be applied to the addresses that are looked up in the NPT, only to the physical addresses. Obviously no hypervisor is going to leave A20 masking on, but the code actually becomes simpler so let's do it. Patches 1 and 2 fix cases in which the addresses must be masked, or overflow is otherwise invalid, for MMU_PHYS_IDX accesses. Patch 3 fixes the bug, by limiting the masking to the case of CR0.PG=0. Patches 4 and 5 further clean up the MMU functions to centralize application of the A20 mask and fix bugs in the process. Untested except for running the SVM tests from kvm-unit-tests (which is better than nothing, still). Supersedes: <0102018c8d11471f-9a6d73eb-0c34-4f61-8d37-5a4418f9e0d7-000...@eu-west-1.amazonses.com> Paolo Bonzini (5): target/i386: mask high bits of CR3 in 32-bit mode target/i386: check validity of VMCB addresses target/i386: Fix physical address truncation target/i386: remove unnecessary/wrong application of the A20 mask target/i386: leave the A20 bit set in the final NPT walk target/i386/tcg/sysemu/excp_helper.c | 44 ++++++++++++---------------- target/i386/tcg/sysemu/misc_helper.c | 3 ++ target/i386/tcg/sysemu/svm_helper.c | 27 +++++++++++++---- 3 files changed, 43 insertions(+), 31 deletions(-) -- 2.43.0