Re: [Bug 216867] New: KVM instruction emulation breaks LOCK instruction atomicity when CMPXCHG fails

[Sean Christopherson <seanjc@google.com>]

On Fri, Dec 30, 2022, bugzilla-daemon@kernel.org wrote:
 
> My code performs the following experiment repeatedly on 3 CPUs:
> 
> * Initially, "ptr" at address 0xb8000 (VGA memory mapped I/O) is set to 0
> * CPU 0 writes 0x12345678 to ptr, then increases counter "count0".
> * In an infinite loop, CPU 1 tries exchanges ptr with register EAX (contains 0)
> using the XCHG instruction. If CPU 1 sees 0x12345678, it increases counter
> "count1".
> * CPU 2's behavior is similar to CPU 1, except it increases counter "count2"
> when it sees 0x12345678.
> 
> Ideally, after each experiment there should always be count1 + count2 = count0.
> However, in KVM, there may be count1 + count2 > count0. This because CPU 0
> writes 0x12345678 to ptr once, but CPU 1 and CPU 2 both get 0x12345678 in XCHG.
> Note that XCHG instruction always implements the locking protocol.
> 
> There is also a deadlock after running the experiment a few times. However I am
> not trying to explain it for now.

Is the suspect deadlock in userspace, the guest, or in the host kernel?

> Guessed cause:
> 
> I guess that KVM emulates the XCHG instruction that accesses 0xb8000. The call
> stack should be:
> 
> ...
>  x86_emulate_instruction (arch/x86/kvm/x86.c)
>   x86_emulate_insn (arch/x86/kvm/emulate.c)
>    writeback (arch/x86/kvm/emulate.c)
>     segmented_cmpxchg (arch/x86/kvm/emulate.c)
>      emulator_cmpxchg_emulated (arch/x86/kvm/x86.c, ->cmpxchg_emulated)
>       emulator_try_cmpxchg_user (arch/x86/kvm/x86.c)
>        ...
>         CMPXCHG instruction
> 
> Suppose CPU 2 wants to write 0 to ptr using writeback(), and expecting ptr to
> already contain 0x13245678. However, CPU 1 changes the content of ptr to 0. So
> * The CMPXCHG instruction fails (clears ZF).
> * emulator_try_cmpxchg_user returns 1.
> * emulator_cmpxchg_emulated() returns X86EMUL_CMPXCHG_FAILED.
> * segmented_cmpxchg() returns X86EMUL_CMPXCHG_FAILED.
> * writeback() returns X86EMUL_CMPXCHG_FAILED.
> * x86_emulate_insn() returns EMULATION_OK.
> 
> Thus, I think the root cause of this bug is that x86_emulate_insn() ignores the
> X86EMUL_CMPXCHG_FAILED error. The correct behavior should be retrying the
> emulation using the updated value (similar to load-linked/store-conditional).

KVM does retry the emulation, albeit in a very roundabout and non-robust way.
On X86EMUL_CMPXCHG_FAILED, x86_emulate_insn() skips the EIP update and doesn't
writeback GPRs.  x86_emulate_instruction() is flawed and emulates single-step, but
the "eip" written should be the original RIP, i.e. shouldn't advance past the
instructions being emulated.  The single-step mess should be fixed, but I doubt
that's the root cause here.

Is there a memslot for 0xb8000?  I assume not since KVM is emulating (have you
actually verified that, e.g. with tracepoints?).  KVM's ABI doesn't support
atomic MMIO operations, i.e. if there's no memslot, KVM will effectively drop
the LOCK semantics.  If that's indeed what's happening, you should see

  kvm: emulating exchange as write

in the host dmesg (just once though).