RE: [PATCH v2] riscv: fix incorrect use of __user pointer

[David Laight]

From: Clément Léger
> Sent: 27 November 2023 10:37
> 
> On 27/11/2023 11:35, David Laight wrote:
> > From: Clément Léger
> >> Sent: 27 November 2023 10:24
> >>
> >> On 25/11/2023 16:37, David Laight wrote:
> >>> ...
> >>>> @@ -491,7 +486,7 @@ int handle_misaligned_load(struct pt_regs *regs)
> >>>>
> >>>>  	val.data_u64 = 0;
> >>>>  	for (i = 0; i < len; i++) {
> >>>> -		if (load_u8(regs, (void *)(addr + i), &val.data_bytes[i]))
> >>>> +		if (load_u8(regs, addr + i, &val.data_bytes[i]))
> >>>>  			return -1;
> >>>>  	}
> >>>
> >>> I'd really have thought that you'd want to pull the kernel/user
> >>> check way outside the loop?
> >>
> >> Hi David,
> >>
> >> I hope the compiler is able to extract that 'if' out of the loop since
> >> regs isn't modified in the loop. Nevertheless, that could be more
> >> "clear" if put outside indeed.
> >
> > If has access regs->xxx then the compiler can't do so because it
> > will must assume that the assignment might alias into 'regs'.
> > That is even true for byte writes if 'strict-aliasing' is enabled
> > - which it isn't for linux kernel builds.
> >
> > It might do so if 'regs' were 'const'; it tends to assume that if
> > it can't change  something nothing can - although that isn't true.
> 
> Ok, good to know ! As I said, I'll modify that in a subsequent patch.

Actually the following loops will (probably) generate much better code:
	// Read kernel
	val = 0;
	for (i = 0; i < len; i++)
		val |= addr[i] << (i * 8);
	// write kernel
	for (i = 0; i < len; i++, val >>= 8)
		addr[i] = val;
For user using __get/put_user() as appropriate.
I think there is a 'goto' variant of the user access functions
that probably make the code clearer.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)