Re: [PATCH] KVM: X86: Fix warning caused by stale emulation context

[Sean Christopherson <seanjc@google.com>]

On Sun, May 23, 2021, Wanpeng Li wrote:
> From: Wanpeng Li <wanpengli@tencent.com>
> 
> Reported by syzkaller:
> 
>   WARNING: CPU: 7 PID: 10526 at /home/kernel/ssd/linux/arch/x86/kvm//x86.c:7621 x86_emulate_instruction+0x41b/0x510 [kvm]
>   RIP: 0010:x86_emulate_instruction+0x41b/0x510 [kvm]
>   Call Trace:
>    kvm_mmu_page_fault+0x126/0x8f0 [kvm]
>    vmx_handle_exit+0x11e/0x680 [kvm_intel]
>    vcpu_enter_guest+0xd95/0x1b40 [kvm]
>    kvm_arch_vcpu_ioctl_run+0x377/0x6a0 [kvm]
>    kvm_vcpu_ioctl+0x389/0x630 [kvm]
>    __x64_sys_ioctl+0x8e/0xd0
>    do_syscall_64+0x3c/0xb0
>    entry_SYSCALL_64_after_hwframe+0x44/0xae
> 
> Commit 4a1e10d5b5d8c (KVM: x86: handle hardware breakpoints during emulation())
> adds hardware breakpoints check before emulation the instruction and parts of 
> emulation context initialization, actually we don't have EMULTYPE_NO_DECODE flag 
> here and the emulation context will not be reused. Commit c8848cee74ff (KVM: x86: 
> set ctxt->have_exception in x86_decode_insn()) triggers the warning because it 
> catches the stale emulation context has #UD, however, it is not during instruction 
> decoding which should result in EMULATION_FAILED. This patch fixes it by moving 
> the second part emulation context initialization before hardware breakpoints check.
> 
> syzkaller source: https://syzkaller.appspot.com/x/repro.c?x=134683fdd00000
> 
> Reported-by: syzbot+71271244f206d17f6441@syzkaller.appspotmail.com
> Fixes: 4a1e10d5b5d8 (KVM: x86: handle hardware breakpoints during emulation)
> Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
> ---
>  arch/x86/kvm/x86.c | 14 +++++++-------
>  1 file changed, 7 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index bbc4e04..eca69f9 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -7552,6 +7552,13 @@ int x86_decode_emulated_instruction(struct kvm_vcpu *vcpu, int emulation_type,
>  
>  	init_emulate_ctxt(vcpu);
>  
> +	ctxt->interruptibility = 0;
> +	ctxt->have_exception = false;
> +	ctxt->exception.vector = -1;
> +	ctxt->perm_ok = false;

What about moving this block all the way into init_emulate_ctxt()?

> +	ctxt->ud = emulation_type & EMULTYPE_TRAP_UD;

This can be left where it is since ctxt->ud is consumed only by x86_decode_insn().
I don't have a strong preference as it really only matters for the backport.  For
upstream, we can kill it off in a follow-up patch by passing emulation_type to
x86_decode_insn() and dropping ctxt->ud altogether.  Tracking that info in ctxt
for literally one call is silly.

diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 8a0ccdb56076..b62944046d7d 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -5322,7 +5322,8 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len)

        ctxt->execute = opcode.u.execute;

-       if (unlikely(ctxt->ud) && likely(!(ctxt->d & EmulateOnUD)))
+       if (unlikely(emulation_type & EMULTYPE_TRAP_UD) &&
+           likely(!(ctxt->d & EmulateOnUD)))
                return EMULATION_FAILED;

        if (unlikely(ctxt->d &
diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
index f016838faedd..2ad32600a8e3 100644
--- a/arch/x86/kvm/kvm_emulate.h
+++ b/arch/x86/kvm/kvm_emulate.h
@@ -314,7 +314,6 @@ struct x86_emulate_ctxt {
        int interruptibility;

        bool perm_ok; /* do not check permissions if true */
-       bool ud;        /* inject an #UD if host doesn't support insn */
        bool tf;        /* TF value before instruction (after for syscall/sysret) */

        bool have_exception;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index a224601d89e2..48b49c24c086 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -7552,8 +7552,6 @@ int x86_decode_emulated_instruction(struct kvm_vcpu *vcpu, int emulation_type,

        init_emulate_ctxt(vcpu);

-       ctxt->ud = emulation_type & EMULTYPE_TRAP_UD;
-
        /*
         * We will reenter on the same instruction since we do not set
         * complete_userspace_io. This does not handle watchpoints yet,
@@ -7563,7 +7561,7 @@ int x86_decode_emulated_instruction(struct kvm_vcpu *vcpu, int emulation_type,
            kvm_vcpu_check_breakpoint(vcpu, &r))
                return r;

-       r = x86_decode_insn(ctxt, insn, insn_len);
+       r = x86_decode_insn(ctxt, insn, insn_len, emulation_type);

        trace_kvm_emulate_insn_start(vcpu);
        ++vcpu->stat.insn_emulation;

> +
>  	/*
>  	 * We will reenter on the same instruction since we do not set
>  	 * complete_userspace_io. This does not handle watchpoints yet,
> @@ -7561,13 +7568,6 @@ int x86_decode_emulated_instruction(struct kvm_vcpu *vcpu, int emulation_type,
>  	    kvm_vcpu_check_breakpoint(vcpu, &r))
>  		return r;
>  
> -	ctxt->interruptibility = 0;
> -	ctxt->have_exception = false;
> -	ctxt->exception.vector = -1;
> -	ctxt->perm_ok = false;
> -
> -	ctxt->ud = emulation_type & EMULTYPE_TRAP_UD;
> -
>  	r = x86_decode_insn(ctxt, insn, insn_len);
>  
>  	trace_kvm_emulate_insn_start(vcpu);
> -- 
> 2.7.4
>