Re: A question about PROT_NONE on ARM and ARM26

[Jamie Lokier <jamie@shareable.org>]

Russell King wrote:
> > Here's the potential improvement to current 32-bit ARM.  It's
> > 4 instructions instead of 8 and one less load, in the common case:
> > 
> > __get_user_4:
> > 	cmp	r0, #TASK_SIZE-4
> > 4:	ldrlet	r1, [r0]
> > 	movle	r0, #0
> > 	movle	pc, lr
> > 	bic	r1, sp, #0x1f00
> > 	bic	r1, r1, #0x00ff
> > 	ldr	r1, [r1, #TI_ADDR_LIMIT]
> > 	sub	r1, r1, #4
> > 	cmp	r0, r1
> > 14:	ldrlet	r1, [r0]
> > 	movle	r0, #0
> > 	movle	pc, lr
> > 	b	__get_user_bad
> 
> Ok, this could work, but there's one gotcha - TASK_SIZE-4 doesn't fit
> in an 8-bit rotated constants, so we need 2 extra instructions:
> 
> __get_user_4:
> 	mov	r1, #TASK_SIZE
> 	sub	r1, r1, #4
> 	cmp	r0, r1
> 4:	ldrlet	r1, [r0]
> 	movle	r0, #0
> 	movle	pc, lr
> 	...

One more possibility:

	cmp	r0, #(TASK_SIZE - (1<<24))

I.e. just compare against the largest constant that can be
represented.  For accesses to the last part of userspace, it's a
penalty of 4 instructions -- but it might work out to be a net gain.

Actually, since the shortest path is only three instructions in the
fast case, not counting control flow, it might be good to inline those
3 in uaccess.h, and change the "bl" to a conditonal "blhi" there.

> > Finally, I think I see a bug in current ARM.  Shouldn't this use
> > ldrlet instead of ldrlst?  Think about accesses to addresses
> > TASK_SIZE-4 and 0xfffffffc.
> 
> LS = unsigned less than or same.  LE = signed less than or equal.  You
> need the unsigned compare because addresses are unsigned.

Ah.  I was guessing the mnemonic.

That's because of the way "ge" is used on ARM26 in places, which
therefore look buggy or subtly clever:

        ldr     r1, [r1, #TI_ADDR_LIMIT]
        sub     r1, r1, #4
        cmp     r0, r1
        bge     __get_user_bad
        cmp     r0, #0x02000000
4:      ldrlst  r1, [r0]
        ldrge   r1, [r0]

"ge" is a signed comparison, and unsigned is needed here, unless I
missed something subtle.  So "bge" and "ldrge" should be "bhi" and "ldrhi".

Thanks,
-- Jamie