Re: [PATCH v1] Intel(R) MPX - Bound violation handling.

[Pedro Alves <palves@redhat.com>]

Hi Walfred,

(Please don't top post.)

On 12/15/2015 10:58 AM, Tedeschi, Walfred wrote:
> From: Pedro Alves [mailto:palves@redhat.com] 
> On 12/14/2015 05:43 PM, Tedeschi, Walfred wrote:

>>> >> +  set_running (user_visible_resume_ptid (1), 0);
>> > 
>> > This is the part that _really_ concerns me, not necessary because I think it's wrong (although, it is a big red flag for me), but because I don't understand why it's needed, and how it will affect things.
>> > (From Joel)
>>> >> +  si_code = parse_and_eval_long ("$_siginfo.si_code\n");
>> > 
>> > During the debugging time I understood that inferior was stopped. Gdb is that was in the process to determine in which state the inferior was.
>> > In this sense I set the flag at this point to allow for the evaluation.
> Where is the error thrown that required brute-forcing set_running away?
> Can we try to find some other way to handle this?  E.g., use something a bit lower level than parse_and_eval_long that bypasses the error?  E.g., start from lookup_internalvar and then use type/value manipulation routines?
>
> It comes from the infrun.c (validate_siginfo_access) .
> The requirement is not running is not fulfilled. Also in the case that we execute a lookup_interval and ask for value_contents we trigger the same code.
>
> What would be the suggestion here:
> Additional function to be used internally in infrun or add a flag.

I gave this some thought, and ended up filling 2 PRs and a proposal forward:

  [Bug breakpoints/19388] Access $_siginfo in breakpoint (catch signal) condition
  https://sourceware.org/bugzilla/show_bug.cgi?id=19388

  [Bug gdb/19389] GDB sometimes mistakenly allows accessing registers of running threads
  https://sourceware.org/bugzilla/show_bug.cgi?id=19389

I think that to move forward we should change validate_siginfo_access to
check is_executing instead of is_running for now.  I think it'll
fix your case (please give it a try).  It's the same check that we do to
prevent accessing registers from a running thread.  $_siginfo is
conceptually really no different from registers -- we could think of it as
just another register.

I sent a patch that does that here, along with testcase that justifies
it on its own, independently of MPX:

  [PATCH] Fix PR19388: Can't access $_siginfo in breakpoint (catch signal) condition
  https://sourceware.org/ml/gdb-patches/2015-12/msg00443.html

By making this change, $_siginfo will become exposed to PR19389 too,
just like registers, and we should definitely fix it, but that seems
like a lesser evil than not being able to get at the info at all.

Thanks,
Pedro Alves