Re: [PATCH v3 2/2] x86/tdx: Fix zero-extension for 32-bit port I/O

[David Laight <david.laight.linux@gmail.com>]

On Thu, 28 May 2026 11:14:38 +0100
Kiryl Shutsemau <kas@kernel.org> wrote:

> On Wed, May 27, 2026 at 10:45:28AM -0700, Dave Hansen wrote:
> > On 5/27/26 05:05, Kiryl Shutsemau (Meta) wrote:
> > ...  
> > > -	/* Update part of the register affected by the emulated instruction */
> > > -	regs->ax &= ~mask;
> > > +	/*
> > > +	 * IN writes the result into a sub-register of RAX. Only the
> > > +	 * 32-bit form zero-extends; the smaller forms leave the upper
> > > +	 * bits untouched:
> > > +	 *
> > > +	 *   insn  dest  size  bits written     bits preserved
> > > +	 *   inb   AL    1     RAX[ 7: 0]       RAX[63: 8]
> > > +	 *   inw   AX    2     RAX[15: 0]       RAX[63:16]
> > > +	 *   inl   EAX   4     RAX[63: 0]       (none, zero-extended)
> > > +	 *
> > > +	 * 'mask' only covers the low 'size' bytes, which is exactly the
> > > +	 * range affected for size 1 and 2. For size 4 the write also
> > > +	 * clears RAX[63:32], so widen the clear-mask.
> > > +	 */
> > > +	if (size == 4)
> > > +		regs->ax = 0;
> > > +	else
> > > +		regs->ax &= ~mask;
> > > +  
> > 
> > Is there any way we could do this with fewer comments and more code?
> > 
> > I mean, there's only three cases. Why have;
> > 
> > 	u64 mask = GENMASK(BITS_PER_BYTE * size - 1, 0);
> > 
> > When there are only 3 possible cases:
> > 
> > 	1 => 0xf
> > 	2 => 0xff
> > 	4 => 0xffff
> > 
> > and one of those cases needs a special case on top of it.
> > 
> > Maybe something like this?
> > 
> > 	/* Clear out part of RAX so part of args.r11 can be OR'd in: */
> > 	switch (size) {
> > 	case 1:
> > 		/* inb consumes lower 8 bits of r11: */
> > 		regs->ax &= ~GENMASK_ULL(7, 0);
> > 		args.r11 &=  GENMASK_ULL(7, 0);
> > 		break;
> > 	case 2:
> > 		/* inw consumes lower 16 bits of r11: */
> > 		regs->ax &= ~GENMASK_ULL(15, 0);
> > 		args.r11 &=  GENMASK_ULL(15, 0);
> > 		break;
> > 	case 4:
> > 		/* inl is weird and zeros the whole register: */
> > 		regs->ax &= ~GENMASK_ULL(63, 0);
> > 		/* But only consumes 32-bits from r11: */
> > 		args.r11 &=  GENMASK_ULL(31, 0);
> > 		break;
> > 	default:
> > 		/* Probable TDX module bug. Illegal in[bwl] size: */
> > 		WARN_ON_ONCE(1);
> > 		success = 0;
> > 	}
> > 
> > 	if (success)
> > 		regs->ax |= args.r11;
> > 
> > It might need a temporary variable for args.r11, but you get the point.
> > That's basically the data from the comment but written as code.  
> 
> I hate how verbose it is. All these GENMASK_ULL() make it hard to
> follow.
> 
> What about the patch below. Inspired by kvm's assign_register().
> 
> diff --git a/arch/x86/coco/tdx/tdx.c b/arch/x86/coco/tdx/tdx.c
> index 65119362f9a2..460b9fbabf14 100644
> --- a/arch/x86/coco/tdx/tdx.c
> +++ b/arch/x86/coco/tdx/tdx.c
> @@ -693,8 +693,8 @@ static bool handle_in(struct pt_regs *regs, int size, int port)
>  		.r13 = PORT_READ,
>  		.r14 = port,
>  	};
> -	u64 mask = GENMASK(BITS_PER_BYTE * size - 1, 0);
>  	bool success;
> +	u32 val;
>  
>  	/*
>  	 * Emulate the I/O read via hypercall. More info about ABI can be found
> @@ -703,10 +703,33 @@ static bool handle_in(struct pt_regs *regs, int size, int port)
>  	 */
>  	success = !__tdx_hypercall(&args);
>  
> -	/* Update part of the register affected by the emulated instruction */
> -	regs->ax &= ~mask;
>  	if (success)
> -		regs->ax |= args.r11 & mask;
> +		val = args.r11;
> +	else
> +		val = 0;
> +
> +	/*
> +	 * IN writes the result into a sub-register of RAX.
> +	 *
> +	 * Only the 32-bit form zero-extends; the smaller forms leave
> +	 * the upper bits untouched.
> +	 */
> +	switch (size) {
> +	case 1:
> +		*(u8 *)&regs->ax = (u8)val;
> +		break;
> +	case 2:
> +		*(u16 *)&regs->ax = (u16)val;
> +		break;
> +	case 4:
> +		/* zero-extended */
> +		regs->ax = val;
> +		break;
> +	default:
> +		/* Probable TDX module bug. Illegal in[bwl] size. */
> +		WARN_ON_ONCE(1);
> +		break;
> +	}

Just write it as normal arithmetic code:

	/* IN writes the result into a sub-register of RAX. */
	switch (size) {
	case 1:
		regs->ax = (regs->ax & ~0xfful) | (val & 0xff);
		break;
	case 2:
		regs->ax = (regs->ax & ~0xfffful) | (val & 0xffff);
		break;
	case 4:
	default:
		/* 32bit 'INB' will zero the high bits. */
		regs->ax = val
		break;
	}

Succinct, obvious and readable.

-- David


>  
>  	return success;
>  }