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

[Borislav Petkov]

On Wed, May 13, 2015 at 12:46:30PM +0200, Ingo Molnar wrote:
> 
> * Borislav Petkov <bp@xxxxxxxxx> wrote:
> 
> > On Wed, May 13, 2015 at 12:31:40PM +0200, Ingo Molnar wrote:
> > > So why should an alternatives-CALL, inlined directly into call sites,
> > > cost more kernel space?
> > 
> > Not the alternatives CALL alone but inlining _copy_*_user with all 
> > the preparation glue around it would. Basically what we're doing 
> > currently.
> 
> So I reacted to this comment of yours:
> 
> > > > The disadvantage is that we have CALL after CALL [...]
> 
> Is the CALL after CALL caused by us calling an alternatives patched 
> function? If yes then we probably should not do that: alternatives 
> switching should IMHO happen at the highest possible level.

Right, so I was trying to analyze Linus' suggestion to uninline stuff
and put it in uaccess_64.c. And that does save us some size and
alternatives patch sites but produces the CALL ... CALL thing.

So let me show you what we have now:

ffffffff8102a774:       0f 1f 40 00             nopl   0x0(%rax)
ffffffff8102a778:       ba 58 00 00 00          mov    $0x58,%edx
ffffffff8102a77d:       4c 89 ff                mov    %r15,%rdi
ffffffff8102a780:       49 83 c7 58             add    $0x58,%r15
ffffffff8102a784:       e8 b7 19 2f 00          callq  ffffffff8131c140 <_copy_to_user>

...

ffffffff8131c140 <_copy_to_user>:
ffffffff8131c140:       65 48 8b 04 25 88 a9    mov    %gs:0xa988,%rax
ffffffff8131c147:       00 00 
ffffffff8131c149:       48 2d 00 40 00 00       sub    $0x4000,%rax
ffffffff8131c14f:       48 89 f9                mov    %rdi,%rcx
ffffffff8131c152:       48 01 d1                add    %rdx,%rcx
ffffffff8131c155:       0f 82 bb c5 57 00       jb     ffffffff81898716 <bad_to_user>
ffffffff8131c15b:       48 3b 48 18             cmp    0x18(%rax),%rcx
ffffffff8131c15f:       0f 87 b1 c5 57 00       ja     ffffffff81898716 <bad_to_user>
ffffffff8131c165:       e9 36 00 00 00          jmpq   ffffffff8131c1a0 <copy_user_generic_unrolled>
ffffffff8131c16a:       66 0f 1f 44 00 00       nopw   0x0(%rax,%rax,1)

so we prep args, call _copy_to_user, do checks and then JMP to the
optimal alternative function.

What I'd like to do is (hypothetically copy'pasted together):

ffffffff8102a778:       ba 58 00 00 00          mov    $0x58,%edx
ffffffff8102a77d:       4c 89 ff                mov    %r15,%rdi

        					movq    -16360(%r14), %rax      # _208->addr_limit.seg, tmp347
					        subq    $88, %rax       #, D.37904
        					cmpq    %rax, %r15      # D.37904, ubuf
					        ja      .L493   #,
        					call copy_user_generic_unrolled #

which saves us the first CALL to _copy_to_user and we do the
alternatives <copy_user_generic_unrolled> CALL directly.

This would mean that we will have to inline _copy_*_user() and switch it
to use copy_user_generic() which already does the proper alternatives.

For the price of some minor size increase and more alternatives patch
sites.

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/