Re: [RFC PATCH] Drop some asm from copy_user_64.S

[Borislav Petkov <bp@alien8.de>]

On Wed, May 13, 2015 at 08:19:55AM +0200, Ingo Molnar wrote:
> Looks nice. Would be useful to do before/after analysis of the
> generated asm with a defconfig and document that in the changelog.

Right, so I'm looking at what we have now:

/* Standard copy_to_user with segment limit checking */
ENTRY(_copy_to_user)
	CFI_STARTPROC
	GET_THREAD_INFO(%rax)
	movq %rdi,%rcx
	addq %rdx,%rcx
	jc bad_to_user
	cmpq TI_addr_limit(%rax),%rcx
	ja bad_to_user

This is adding @to (in %rdi) with size (in %rdx) and then looking at the
carry flag. __chk_range_not_ok() does the same thing, but with a single
operation, AFAICT:

static inline bool __chk_range_not_ok(unsigned long addr, unsigned long size, unsigned long limit)
{
        /*
         * If we have used "sizeof()" for the size,
         * we know it won't overflow the limit (but
         * it might overflow the 'addr', so it's
         * important to subtract the size from the
         * limit, not add it to the address).
         */
        if (__builtin_constant_p(size))
                return addr > limit - size;

and we're avoiding the addr overflow by subtracting size from limit.

So the resulting asm looks like this:

        .file 22 "./arch/x86/include/asm/uaccess.h"
        .loc 22 54 0
        movq    -16360(%r14), %rax      # _208->addr_limit.seg, tmp347		%r14 contains thread_info
        subq    $88, %rax       #, D.37904					88 is the size

        .file 23 "./arch/x86/include/asm/uaccess_64.h"
        .loc 23 165 0
        cmpq    %rax, %r12      # D.37904, ubuf					%r12 contains the user ptr
        ja      .L493   #,
        movq    %r12, %rdi      # ubuf, to					prep args for copy_user...
        movl    $88, %edx       #, len

										alternative starts here
	#APP
	# 36 "./arch/x86/include/asm/uaccess_64.h" 1
	661:
	call copy_user_generic_unrolled	#
	....


so we end up replacing

	MOV
	ADD
	JC
	CMP
	JA
	JMP (alternative)

with

	MOV
	SUB
	CMP
	JA
	MOV
	MOV
	CALL (alternative)

The only problem I see here is that we have to do two MOVs to put args
in proper registers before calling the copy_user* version. But we end
up with a single conditional instead of two. And the MOVs are cheaper.
Also, we gets rid of asm glue, even betterer :-)

> I'd keep any changes to inlining decisions a separate patch and do
> vmlinux before/after size analysis as well, so that we don't mix the
> effects of the various enhancements.

Yap.

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--