[PATCH v2] Fix logic in exec_file_locate_attach

[Luis Machado]

On 02/22/2016 10:51 AM, Gary Benson wrote:
> Luis Machado wrote:
>> On 02/22/2016 07:40 AM, Gary Benson wrote:
>>> Luis Machado wrote:
>>>> On 02/19/2016 09:21 AM, Gary Benson wrote:
>>>>> This is an updated version of the patch I posted yesterday.
>>>>> It fails silently rather than throwing if the executable is
>>>>> not in the sysroot, which both fixes the sysroot-escape issue
>>>>> and results in a better GDB session for the user.
>>>>>
>>>>> Built and regtested on RHEL 6.6 x86_64.
>>>>>
>>>>> Luis, I think this patch will fix your connection drop without
>>>>> any further changes.  Could you test it please?
>>>>
>>>> Unfortunately it doesn't completely solve the problem i saw, as
>>>> exec_file_find will still potentially throw errors and will
>>>> disrupt the connection attempt or stop execution of a custom
>>>> sequence of commands (as Pedro noted) when "attach" is part of
>>>> the sequence.
>>>>
>>>> define foo
>>>> attach <pid>
>>>>>>>> execution stops here if an error is thrown
>>>> info threads
>>>> info registers
>>>> end
>>>>
>>>> It still looks like a TRY/CATCH block is needed around at least
>>>> exec_file_find.
>>>
>>> What is throwing in exec_file_find?  I'm just seeing lots of calls
>>> to gdb_open_cloexec and openp, and I don't think either of those
>>> should throw except for assertion failures or running out of
>>> memory.
>>
>> Not sure why i had exec_file_find in my mind. I meant to say
>> exec_file_attach still throws errors, when openp fails and
>> scratch_chan < 0. Sorry.
>
> You shouldn't get that now, the "if (full_exec_path == NULL) return"
> should have caught it.  Are you still seeing thrown errors with your
> setup?
>

Yes. With your patch applied, i still see a case where we error out. 
Suppose we have a test binary gdb/test, then:

- chmod -r gdb/test
- Fire up gdbserver with a test binary: ./gdb/gdbserver/gdbserver :2345 
gdb/test
- Fire up GDB: ./gdb/gdb -ex "set sysroot" -ex "tar rem :2345"

You will see something similar to the following:

Sending packet: $qXfer:exec-file:read:3486:0,fff#5f...Packet received: 
l/proc/13446/exe
/proc/13446/exe: Permission denied.
(gdb) i r
The program has no registers now.
(gdb)

This was the testcase suggested by Pedro and it proved to be a good one.

>> There is a symbol_file_add_main call right after calling
>> exec_file_attach in exec_file_locate_attach, but i didn't see any
>> errors being thrown from that one.
>
> You could probably race it (e.g. by deleting the file between the
> calls) but generally symbol_file_add_main won't fail because
> exec_file_attach would have failed if the file was missing or
> inaccessible.

My idea was to guard both exec_file_attach and symbol_file_add_main. We 
can't have anything in that function throwing an error that won't be 
caught, otherwise the above connection attempt will fail.

Luis