[PATCH] KVM: x86/emulator: Bounds check reg nr against reg array size

[PATCH] KVM: x86/emulator: Bounds check reg nr against reg array size

[Kees Cook]

GCC 12 sees that it might be possible for "nr" to be outside the _regs
array. Add explicit bounds checking.

In function 'reg_read',
    inlined from 'reg_rmw' at ../arch/x86/kvm/emulate.c:266:2:
../arch/x86/kvm/emulate.c:254:27: warning: array subscript 32 is above array bounds of 'long unsigned int[17]' [-Warray-bounds]
  254 |         return ctxt->_regs[nr];
      |                ~~~~~~~~~~~^~~~
In file included from ../arch/x86/kvm/emulate.c:23:
../arch/x86/kvm/kvm_emulate.h: In function 'reg_rmw':
../arch/x86/kvm/kvm_emulate.h:366:23: note: while referencing '_regs'
  366 |         unsigned long _regs[NR_VCPU_REGS];
      |                       ^~~~~

Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sean Christopherson <seanjc@google.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Wanpeng Li <wanpengli@tencent.com>
Cc: Jim Mattson <jmattson@google.com>
Cc: x86@kernel.org
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: kvm@vger.kernel.org
Signed-off-by: Kees Cook <keescook@chromium.org>
---
 arch/x86/kvm/emulate.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 89b11e7dca8a..fbcbc012a3ae 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -247,6 +247,8 @@ enum x86_transfer_type {
 
 static ulong reg_read(struct x86_emulate_ctxt *ctxt, unsigned nr)
 {
+	if (WARN_ON(nr >= ARRAY_SIZE(ctxt->_regs)))
+		return 0;
 	if (!(ctxt->regs_valid & (1 << nr))) {
 		ctxt->regs_valid |= 1 << nr;
 		ctxt->_regs[nr] = ctxt->ops->read_gpr(ctxt, nr);
@@ -256,6 +258,8 @@ static ulong reg_read(struct x86_emulate_ctxt *ctxt, unsigned nr)
 
 static ulong *reg_write(struct x86_emulate_ctxt *ctxt, unsigned nr)
 {
+	if (WARN_ON(nr >= ARRAY_SIZE(ctxt->_regs)))
+		return 0;
 	ctxt->regs_valid |= 1 << nr;
 	ctxt->regs_dirty |= 1 << nr;
 	return &ctxt->_regs[nr];
-- 
2.32.0

Re: [PATCH] KVM: x86/emulator: Bounds check reg nr against reg array size

[Sean Christopherson]

On Fri, May 20, 2022, Kees Cook wrote:
> GCC 12 sees that it might be possible for "nr" to be outside the _regs
> array. Add explicit bounds checking.

I think GCC 12 is wrong.

There are four uses of reg_rmw() that don't use hardcoded registers:

   $ git grep reg_rmw | grep -v VCPU_REGS_
   emulate.c:static ulong *reg_rmw(struct x86_emulate_ctxt *ctxt, unsigned nr)
1  emulate.c:	ulong *preg = reg_rmw(ctxt, reg);
2  emulate.c:		p = (unsigned char *)reg_rmw(ctxt, modrm_reg & 3) + 1;
3  emulate.c:		p = reg_rmw(ctxt, modrm_reg);
4  emulate.c:		assign_register(reg_rmw(ctxt, reg), val, ctxt->op_bytes);

#1 has three users, but two of those use hardcoded registers.

  $ git grep register_address_increment | grep -v VCPU_REGS_
  emulate.c:register_address_increment(struct x86_emulate_ctxt *ctxt, int reg, int inc)
  emulate.c:	register_address_increment(ctxt, reg, df * op->bytes);
 
and that last one is string_addr_inc(), which is only called with RDI or RSI.

#2 can't overflow as the register can only be 0-3 (yay AH/BH/CH/DH operands).

#3 is the !highbyte path of decode_register(), and is a bit messy, but modrm_reg
is always sanitized.

   $ git grep -E "decode_register\("
   emulate.c:static void *decode_register(struct x86_emulate_ctxt *ctxt, u8 modrm_reg,
a  emulate.c:      op->addr.reg = decode_register(ctxt, reg, ctxt->d & ByteOp);
b  emulate.c:              op->addr.reg = decode_register(ctxt, ctxt->modrm_rm,
c  emulate.c:                      ctxt->memop.addr.reg = decode_register(ctxt,
                                                                          ctxt->modrm_rm, true);

For (b) and (c), modrm_reg == ctxt->modrm_rm, which is computed in one place and
is bounded to 0-15:

	base_reg = (ctxt->rex_prefix << 3) & 8; /* REX.B */
	ctxt->modrm_rm = base_reg | (ctxt->modrm & 0x07);

For (a), "reg" is either modrm_reg or a register that is encoded in the opcode,
both of which are again bounded to 0-15:

	unsigned reg = ctxt->modrm_reg;

	if (!(ctxt->d & ModRM))
		reg = (ctxt->b & 7) | ((ctxt->rex_prefix & 1) << 3);

and

	ctxt->modrm_reg = ((ctxt->rex_prefix << 1) & 8); /* REX.R */
	ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;

#4 is em_popa() and is just funky hardcoding of popping RAX-RDI, minus RSP.

I did the same exercise for reg_reg() and write_reg(), and the handful of
non-hardcoded use are all bounded in similar ways.

> In function 'reg_read',
>     inlined from 'reg_rmw' at ../arch/x86/kvm/emulate.c:266:2:

Is there more of the "stack" available?  I don't mind the WARN too much, but if
there is a bug lurking I would much rather fix the bug.

> ../arch/x86/kvm/emulate.c:254:27: warning: array subscript 32 is above array bounds of 'long unsigned int[17]' [-Warray-bounds]
>   254 |         return ctxt->_regs[nr];
>       |                ~~~~~~~~~~~^~~~
> In file included from ../arch/x86/kvm/emulate.c:23:
> ../arch/x86/kvm/kvm_emulate.h: In function 'reg_rmw':
> ../arch/x86/kvm/kvm_emulate.h:366:23: note: while referencing '_regs'
>   366 |         unsigned long _regs[NR_VCPU_REGS];
>       |                       ^~~~~

Re: [PATCH] KVM: x86/emulator: Bounds check reg nr against reg array size

[Kees Cook]

On Fri, May 20, 2022 at 05:32:04PM +0000, Sean Christopherson wrote:
> On Fri, May 20, 2022, Kees Cook wrote:
> > GCC 12 sees that it might be possible for "nr" to be outside the _regs
> > array. Add explicit bounds checking.
> 
> I think GCC 12 is wrong.

I think it's more like GCC is extremely conservative about these things,
and assumes the worst when, for whatever reason, it can't track
something.

> There are four uses of reg_rmw() that don't use hardcoded registers:
> 
>    $ git grep reg_rmw | grep -v VCPU_REGS_
>    emulate.c:static ulong *reg_rmw(struct x86_emulate_ctxt *ctxt, unsigned nr)
> 1  emulate.c:	ulong *preg = reg_rmw(ctxt, reg);
> 2  emulate.c:		p = (unsigned char *)reg_rmw(ctxt, modrm_reg & 3) + 1;
> 3  emulate.c:		p = reg_rmw(ctxt, modrm_reg);
> 4  emulate.c:		assign_register(reg_rmw(ctxt, reg), val, ctxt->op_bytes);
> 
> #1 has three users, but two of those use hardcoded registers.
> 
>   $ git grep register_address_increment | grep -v VCPU_REGS_
>   emulate.c:register_address_increment(struct x86_emulate_ctxt *ctxt, int reg, int inc)
>   emulate.c:	register_address_increment(ctxt, reg, df * op->bytes);
>  
> and that last one is string_addr_inc(), which is only called with RDI or RSI.
> 
> #2 can't overflow as the register can only be 0-3 (yay AH/BH/CH/DH operands).
> 
> #3 is the !highbyte path of decode_register(), and is a bit messy, but modrm_reg
> is always sanitized.
> 
>    $ git grep -E "decode_register\("
>    emulate.c:static void *decode_register(struct x86_emulate_ctxt *ctxt, u8 modrm_reg,
> a  emulate.c:      op->addr.reg = decode_register(ctxt, reg, ctxt->d & ByteOp);
> b  emulate.c:              op->addr.reg = decode_register(ctxt, ctxt->modrm_rm,
> c  emulate.c:                      ctxt->memop.addr.reg = decode_register(ctxt,
>                                                                           ctxt->modrm_rm, true);
> 
> For (b) and (c), modrm_reg == ctxt->modrm_rm, which is computed in one place and
> is bounded to 0-15:
> 
> 	base_reg = (ctxt->rex_prefix << 3) & 8; /* REX.B */
> 	ctxt->modrm_rm = base_reg | (ctxt->modrm & 0x07);
> 
> For (a), "reg" is either modrm_reg or a register that is encoded in the opcode,
> both of which are again bounded to 0-15:
> 
> 	unsigned reg = ctxt->modrm_reg;
> 
> 	if (!(ctxt->d & ModRM))
> 		reg = (ctxt->b & 7) | ((ctxt->rex_prefix & 1) << 3);
> 
> and
> 
> 	ctxt->modrm_reg = ((ctxt->rex_prefix << 1) & 8); /* REX.R */
> 	ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;
> 
> #4 is em_popa() and is just funky hardcoding of popping RAX-RDI, minus RSP.
> 
> I did the same exercise for reg_reg() and write_reg(), and the handful of
> non-hardcoded use are all bounded in similar ways.

Thanks for digging into this. I tried to do the same and started to lose
track of things.

> 
> > In function 'reg_read',
> >     inlined from 'reg_rmw' at ../arch/x86/kvm/emulate.c:266:2:
> 
> Is there more of the "stack" available?  I don't mind the WARN too much, but if
> there is a bug lurking I would much rather fix the bug.

Agreed, but I haven't found a way to get more context here. I think I
found a separate place where GCC really does look to have a bug, as it
complains about array usage that is explicitly bounded. :P

-- 
Kees Cook

Re: [PATCH] KVM: x86/emulator: Bounds check reg nr against reg array size

[Sean Christopherson]

On Fri, May 20, 2022, Kees Cook wrote:
> GCC 12 sees that it might be possible for "nr" to be outside the _regs
> array. Add explicit bounds checking.
> 
> In function 'reg_read',
>     inlined from 'reg_rmw' at ../arch/x86/kvm/emulate.c:266:2:
> ../arch/x86/kvm/emulate.c:254:27: warning: array subscript 32 is above array bounds of 'long unsigned int[17]' [-Warray-bounds]
>   254 |         return ctxt->_regs[nr];
>       |                ~~~~~~~~~~~^~~~
> In file included from ../arch/x86/kvm/emulate.c:23:
> ../arch/x86/kvm/kvm_emulate.h: In function 'reg_rmw':
> ../arch/x86/kvm/kvm_emulate.h:366:23: note: while referencing '_regs'
>   366 |         unsigned long _regs[NR_VCPU_REGS];
>       |                       ^~~~~
> 
> Cc: Paolo Bonzini <pbonzini@redhat.com>
> Cc: Sean Christopherson <seanjc@google.com>
> Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
> Cc: Wanpeng Li <wanpengli@tencent.com>
> Cc: Jim Mattson <jmattson@google.com>
> Cc: x86@kernel.org
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: kvm@vger.kernel.org
> Signed-off-by: Kees Cook <keescook@chromium.org>
> ---
>  arch/x86/kvm/emulate.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 89b11e7dca8a..fbcbc012a3ae 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -247,6 +247,8 @@ enum x86_transfer_type {
>  
>  static ulong reg_read(struct x86_emulate_ctxt *ctxt, unsigned nr)
>  {
> +	if (WARN_ON(nr >= ARRAY_SIZE(ctxt->_regs)))
> +		return 0;
>  	if (!(ctxt->regs_valid & (1 << nr))) {
>  		ctxt->regs_valid |= 1 << nr;
>  		ctxt->_regs[nr] = ctxt->ops->read_gpr(ctxt, nr);
> @@ -256,6 +258,8 @@ static ulong reg_read(struct x86_emulate_ctxt *ctxt, unsigned nr)
>  
>  static ulong *reg_write(struct x86_emulate_ctxt *ctxt, unsigned nr)
>  {
> +	if (WARN_ON(nr >= ARRAY_SIZE(ctxt->_regs)))
> +		return 0;

This is wrong, reg_write() confusingly returns a pointer the register to be written,
it doesn't actually do the write.  So if we want to guard against array overflow,
it would be better to cap @nr and continue on, i.e. assume some higher bit was
spuriously set.

The other oddity here is that VCPU_REGS_RIP should never be read, the RIP relative
code reads _eip directly.  I.e. _regs[] should really be VCPU_REGS_R15+1.  And
adding a #define for that would clean up this bit of code in writeback_registers()
that hardcodes 16 (rax - r15) GPRs:

	for_each_set_bit(reg, (ulong *)&ctxt->regs_dirty, 16)
		ctxt->ops->write_gpr(ctxt, reg, ctxt->_regs[reg]);

Lastly, casting regs_dirty to an unsigned long pointer is all kinds of gross, e.g.
if it were moved to the end of struct x86_emulate_ctxt then the above could trigger
an out-of-bounds read.

I'll whip up a small series to clean this code up and add WARNs similar to above.