Re: [PATCH 31/40] target_ops/C++: Base FreeBSD target

[Pedro Alves <palves at redhat dot com>]

On 04/18/2018 01:37 AM, John Baldwin wrote:
> On Tuesday, April 17, 2018 06:07:37 PM Pedro Alves wrote:
>> On 04/17/2018 05:05 PM, John Baldwin wrote:
>>> On Saturday, April 14, 2018 08:09:44 PM Pedro Alves wrote:
>>>> The
>>>>
>>>>   $architecture x NetBSD/OpenBSD/FreeBSD
>>>>
>>>> support matrix complicates things a bit.  There's common BSD target
>>>> code, and there's common architecture-specific code shared between the
>>>> different BSDs.  Current, all that is stiched together to form a final
>>>> target, via the i386bsd_target, x86bsd_target, fbsd_nat_add_target
>>>> functions etc.
>>>>
>>>> Introduces a fbsd_nat_target base/prototype target.  To be used in
>>>> following patches.
>>>
>>> I will do some tests of FreeBSD/amd64 first and let you know what I find.
>>
>> Thank you!
> 
> I've pushed a target_ops-cxx branch to github.com/bsdjhb/gdb.git that has
> some small fixups (compile fixes).  

Thank you!

> I've built the amd64, i386, arm, and
> aarch64 FreeBSD native targets so far.  Simple testing of the the amd64
> and i386 binaries seems to work, but I encountered a new test failure
> in the testsuite for FreeBSD/amd64 that is a bit odd.  In particular,
> I get a core dump running 'info set' when it tries to display the
> current setting of whether ASLR is disabled.  Looking at the core of gdb:

Ah, I wasn't seeing this because the Linux target implements
the supports_disable_randomization method.  If I hack that away, I
can reproduce the crash.

> 
> Program terminated with signal SIGSEGV, Segmentation fault.
> #0  0x0000000000d3a7d4 in target_ops::supports_disable_randomization (
>     this=0x28d4c68 <the_amd64_fbsd_nat_target>)
>     at ../../gdb/target-delegates.c:2732
> 2732      return this->beneath->supports_disable_randomization ();
> 

... 

> 
> From the stack trace we can see that it already bounced down to the dummy
> target which calls find_default_supports_disable_randomization.  That
> finds the native "run" target and invokes its method without pushing
> it onto the stack.  I think before if a native target didn't support ASLR
> at all it just didn't set the function pointer and no harm was done.
> Now the function pointer is effectively always set but to something that
> assumes 'beneath' is valid.  I'm not quite sure how you want to fix this.
> The simple solution is to change the default method to return false if
> beneath is NULL, but I'm not quite sure that fits in with the design this
> branch is aiming for.

Thanks for the analysis, that helps.

There's a small set of target methods that must always be
implemented by native targets, to avoid falling down to the
target beneath (since the target may not be pushed when the
method is called).  See target.c:complete_target_initialization
in current master.  Since all native targets inherit from inf-child
(just like in current master), in the C++ version, we can override
it there.  Like below.  This fixes it for me.

diff --git c/gdb/inf-child.h w/gdb/inf-child.h
index d3f8b71589..956cee2a0a 100644
--- c/gdb/inf-child.h
+++ w/gdb/inf-child.h
@@ -72,6 +72,7 @@ public:
      target that can run.  */
   bool can_async_p ()  override { return false; }
   bool supports_non_stop ()  override { return false; }
+  bool supports_disable_randomization () override { return false; }
 
   char *pid_to_exec_file (int pid) override;

Thanks,
Pedro Alves