RE: [PATCH V7] amd64-mpx: initialize bnd register before performing inferior calls.

["Tedeschi, Walfred" <walfred dot tedeschi at intel dot com>]

-----Original Message-----
From: gdb-patches-owner@sourceware.org [mailto:gdb-patches-owner@sourceware.org] On Behalf Of Pedro Alves
Sent: Thursday, February 16, 2017 2:52 PM
To: Tedeschi, Walfred <walfred.tedeschi@intel.com>; qiyaoltc@gmail.com; brobecker@adacore.com
Cc: gdb-patches@sourceware.org
Subject: Re: [PATCH V7] amd64-mpx: initialize bnd register before performing inferior calls.

On 02/16/2017 01:49 PM, Tedeschi, Walfred wrote:

>> Correct?
> Yes.
> But actual behavior at 7 has an issue!

7. is:

>>> 7. expected - control should be back to 1, i.e. on stop mode.
>>> 7. actual behavior - application finishes with the signal

But the rest of your email doesn't talk about this at all.
I'm confused....

FRED:
Issue is that I forgot to say in here that segv is not presented when we set the BND register in the prolog.
(The manual set is meant) 
I intended to force a boundary violation. But did not work.

> 
> When we set the BND registers from gdb itself (applying the patch) it 
> looks like changing the values of BND again while in the prolog have 
> no effect.
> Lets go to the reproducer:
> 
> The inferior call i want to do is "upper (x, a, b, c, d, 100)".
> it has the following relevant prolog:
> 
>   0x0000000000400a0b <+1>:    mov    %rsp,%rbp
>    0x0000000000400a0e <+4>:    sub    $0x18,%rsp
>    0x0000000000400a12 <+8>:    mov    %rdi,-0x18(%rbp)
>    0x0000000000400a16 <+12>:    mov    %rsi,-0x20(%rbp)
>    0x0000000000400a1a <+16>:    mov    %rdx,-0x28(%rbp)
>    0x0000000000400a1e <+20>:    mov    %rcx,-0x30(%rbp)
>    0x0000000000400a22 <+24>:    mov    %r8,-0x38(%rbp)
>    0x0000000000400a26 <+28>:    mov    %r9d,-0x3c(%rbp)
>    0x0000000000400a2a <+32>:    bndmov %bnd0,-0x50(%rbp)
>    0x0000000000400a2f <+37>:    bndmov %bnd1,-0x60(%rbp)
>    0x0000000000400a34 <+42>:    bndmov %bnd2,-0x70(%rbp)
>    0x0000000000400a39 <+47>:    bndmov %bnd3,-0x80(%rbp)
> 
> I can stop at the first instruction of "upper" by issuing b (void*)&upper.

FYI, the usual way to do that is with "b *upper".


> In order to verify the effective change in the BND i have printed 
> bnd0..bnd3. Register values were same as entered with the GDB command.

printed how?  and printed when exactly?

FRED:
So, stopped at 0x0000000000400a0b set BND0 and did a step instruction and printed BND0
Commands were: 
P $bnd0={0xFFFFF..., 0} 
Si
P $bnd0 (output was the same as input)

For the value I set I should have got an segv. But I did not. 
> Other way is to do instruction stepping till " bndmov %bnd3,-0x80(%rbp)"
> and examine the memory at the indicated places.

Memory?  I thought you'd examine the registers?  What indicated places, BTW?

FRED:
Yes, tring to discover what happened to the set I stepped instructions till 0x0000000000400a2f
This tells me that the BND0 should be at  0x50(%rbp) but value was not the one I added by hand but the automatically set (with the patch)
Interestingly without the patch, and repeating the same procedure, the value is the one I have set by hand while stopped at the prolog.

 
> 
> Surprise! In the gdb i have applied the patch though changing the
> BND0..BND3 values at 0x0400a0b value present on memory was still set 
> to the init state.

_memory_ set to the init state?

FRED:
X /4w ($rbp -0x50)

When stopped at 0x0000000000400a2f we should see the value set of $BND0 what we see is the automatic set value and not the value I set when at the prolog.

Pedro:
Can you please explain what you're seeing in a bit more detail?
You're leaving out details I'm finding myself needing to guess, and I'd probably guess wrong.

But still, if I have to guess, I'd think that the problem with stopping at function entry and poking the bnd registers _before_ the prologue runs, would be that whatever bnd register value you patch in, would be overridden by the bndmov instructions in the prologue.  I.e., you need to single-step past those bndmov instructions, and patch the bnd registers _then_, otherwise the bndmovs undo your patching.


But this comment:

> In the version without applying the patch it i could see the value i 
> entered while stopped at the first instruction.

... seems to contradict that.  So I'm double confused.

Still, I don't see what does this have to do with point 7.

Thanks,
Pedro Alves


Fred:
Thanks a lot again! I hope I could get the message trough now. With memory registers set and not set the setup was a bit complicated in case it is still not clear, I will send the full log.

Regards,
/Fred
Intel Deutschland GmbH
Registered Address: Am Campeon 10-12, 85579 Neubiberg, Germany
Tel: +49 89 99 8853-0, www.intel.de
Managing Directors: Christin Eisenschmid, Christian Lamprechter
Chairperson of the Supervisory Board: Nicole Lau
Registered Office: Munich
Commercial Register: Amtsgericht Muenchen HRB 186928