This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

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




On 02/13/2017 09:33 AM, Tedeschi, Walfred wrote:


On 02/08/2017 01:27 PM, Pedro Alves wrote:
On 02/07/2017 08:56 AM, Tedeschi, Walfred wrote:
On 01/31/2017 03:13 PM, Walfred Tedeschi wrote:
This patch initializes the bnd registers before executing the inferior
call.  BND registers can be in arbitrary values at the moment of the
inferior call.  In case the function being called uses as part of the
parameters bnd register, e.g. when passing a pointer as parameter, the
current value of the register will be used.  This can cause boundary
violations that are not due to a real bug or even desired by the user.
In this sense the best to be done is set the bnd registers to allow
access to the whole memory, i.e. initialized state, before pushing the
inferior call.
This explains the reason for clearing better ...
Yes, it was my intention. Do you see value to have the patch in then?
I think I do.
At this point i think it is woth to mention a bit more on the usage model I had in mind.

At the inferior call time all BND registers are set to the INIT state.
That covers 90% of the usage case and keeps the actual behavior of GDB.
In case it is desirable to investigate any bound violation user can add a breakpoint on the prolog of the function and set the BND register there, I understand that this is the normal use case for the debugger as well.
In reality it did not work as expected. When an inferior call is done and there is a breakpoint in the executed GDB issues a message like:

"The program being debugged stopped while in a function called from GDB.
Evaluation of the expression containing the function
(foo) will be abandoned.
When the function is done executing, GDB will silently stop."

Also after that no signal generated within the function call is reported.
Not sure if this a limitation or a bug. In case you think it is a bug i can add some entries in Bugzilla.

On the other hand i see then the following option:
Add the code to INIT the BND regs based on a set/show option.
In case not set user will have to take care of setting and restoring the register himself.

As optimal solution we could also provide 4 debugger variables, or convenience variables to keep the register values chosen by the user.
Use those to set the registers before the inferior call.
 I have the impression that this option is an overkill.

Would you agree with the creation of the new set and show variable to INIT the BND registers before the inferior call?


Thanks again and best regards,
/Fred

For passing local pointers to some function, it might be
that GDB could be able to figure out which bound registers
contains the bound for a given variable, or if spilled, where
to find then, and set up the call to use the right bounds, but
I have no idea of how to retrieve that information.  I suspect
that it's not a mapping we could retrieve from the dwarf?  And
then there's also the case of passing pointers to global
variables, and pointers to memory that gdb malloc's into the
inferior, like for array/string coercion:
ABI defines which BND is used for which parameter and what to do when it is needed to pass more bounds than BND registers available.

(gdb) p strlen ("hello")

this will alloc a block of memory in the inferior for
"hello", by calling malloc in the inferior.  If strlen
is compiled to do BND checks, would we need to setup the
BND registers to the range of the pointer returned by malloc ?
Yes we would need to set the BND in GDB in this scenario.
However we could simply let that attribution to the user.
My point of view is that this could complicate a lot the code to cover very little use.


I've not used BND myself, so I don't have any experience
with it.  But my impression is that disabling BND for
infcalls makes infcalls work again on BND-enabled systems, and
that we could perhaps try to do something smart to re-enable it
in some infcall cases, if there's sufficient benefit.

Now, a question is: could this auto-clearing of BND registers
get in the way of some use cases?  E.g., could a user want to poke
at the BND registers manually before calling a function in the
inferior, in order to debug BND-related code.
If so, that may call for a new option users could tweak
if necessary.
I was also considering this model as my first implementation.
But this way I suppose that it will be even more complicated.
We will need special commands for the architecture, special documentation...

if we set afterwards: Inferior call + Break point + register set.
we should not need any additional set and show for the architecture.
Hover as you pointed out in the other e-mail it is better to increase testing and document it.
Thanks,
Pedro Alves

Thanks a lot for your review!

/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


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]