On 17.03.2020 20:00, Simon Marchi wrote:
> On 2020-03-17 1:45 p.m., Kamil Rytarowski wrote:
>> On 17.03.2020 17:39, Simon Marchi wrote:
>>> On 2020-03-17 12:30 p.m., Kamil Rytarowski wrote:
>>>> NetBSD tracks the PID and LWP pair separately and both values are
>>>> needed and meaningful.
>>>> ---
>>>>  gdb/inf-ptrace.c | 6 +++++-
>>>>  1 file changed, 5 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/gdb/inf-ptrace.c b/gdb/inf-ptrace.c
>>>> index db17a76d946..6a6cb554ba7 100644
>>>> --- a/gdb/inf-ptrace.c
>>>> +++ b/gdb/inf-ptrace.c
>>>> @@ -321,10 +321,14 @@ get_ptrace_pid (ptid_t ptid)
>>>>  {
>>>>    pid_t pid;
>>>>
>>>> +#if !defined(__NetBSD__)
>>>>    /* If we have an LWPID to work with, use it.  Otherwise, we're
>>>> -     dealing with a non-threaded program/target.  */
>>>> +     dealing with a non-threaded program/target.
>>>> +
>>>> +     NetBSD tracks the PID and LWP pair separately. */
>>>>    pid = ptid.lwp ();
>>>>    if (pid == 0)
>>>> +#endif
>>>>      pid = ptid.pid ();
>>>>    return pid;
>>>>  }
>>>> --
>>>> 2.25.0
>>>>
>>>
>>> I think you should just avoid using get_ptrace_pid on NetBSD altogether, since
>>> it is meant for OSes that require passing a single thread identifier to ptrace
>>> (whereas NetBSD requires the (pid, lwp) pair).
>>>
>>> Even with this modification in get_ptrace_pid, you need to change all the ptrace
>>> call sites to pass the lwp on top of it.
>>>
>>> I would suggest to instead #ifdef out get_ptrace_pid entirely on NetBSD, to avoid
>>> using it by mistake, and just replace all ptrace call sites possibly used on BSD
>>> to be
>>>
>>>   ptrace (request, ptid.pid (), addr, ptid.lwp ());
>>>
>>> This matches what I suggested in:
>>>
>>>   https://sourceware.org/pipermail/gdb-patches/2020-March/166735.html
>>>
>>> Simon
>>>
>>
>> Avoiding is possibly nice.. however in the current code it is much more
>> intrusive. We would need to patch now generic and OS/CPU specific code
>> (some of that is also shared with other OSs due to legacy reasons).
>>
>> I think it is much cleaner to return ptid. pid() for NetBSD and reflect
>> the meaning of get_ptrace_pid().
>>
>> If I follow your advice I end up with ifdefs like here:
>>
>> diff --git a/gdb/inf-ptrace.c b/gdb/inf-ptrace.c
>> index b63a1bf88ef..a5d9c1d10ea 100644
>> --- a/gdb/inf-ptrace.c
>> +++ b/gdb/inf-ptrace.c
>> @@ -349,7 +349,11 @@ inf_ptrace_target::resume (ptid_t ptid, int step,
>> enum gdb_signal signal)
>>         single-threaded processes, so simply resume the inferior.  */
>>      pid = inferior_ptid.pid ();
>>    else
>> +#ifdef __NetBSD__
>> +    pid = ptid. pid();
>> +#else
>>      pid = get_ptrace_pid (ptid);
>> +#endif
>>
>>    if (catch_syscall_enabled () > 0)
>>      request = PT_SYSCALL;
>> @@ -533,7 +537,11 @@ inf_ptrace_target::xfer_partial (enum target_object
>> object,
>>  				 const gdb_byte *writebuf,
>>  				 ULONGEST offset, ULONGEST len, ULONGEST *xfered_len)
>>  {
>> +#ifdef __NetBSD__
>> +  pid_t pid = inferior_ptid. pid();
>> +#else
>>    pid_t pid = get_ptrace_pid (inferior_ptid);
>> +#endif
>>
>>    switch (object)
>>      {
> 
> I was thinking more about using a "gdb_ptrace" function in this file, as
> you have added in the other patch.  The only ifdef would be in that function.
> get_ptrace_pid would only be used in the !__NetBSD__ branch of that function.
> 
> inf_ptrace_target::xfer_partial and inf_ptrace_target::resume would just call
> gdb_ptrace, passing the right ptid.
> 
>> Maintaining that will be certainly harder and it will be prone to
>> recurring regressions.
>>
>> If we want to take the route of cleanups and refactoring I think it
>> would be better to rethink the pid,lwp separation in Linux; but that is
>> much beyond the scope of my patches.
> 
> I'm not sure what do mean by "rethink the pid,lwp separation in Linux".
> 

Please disregard. Refactoring (and removal assumptions about specific
threading model) is out of scope of my patches.

>> Last but not least, get_ptrace_pid() would work now for NetBSD literally
>> as specified in the function name now... just extracting pid from ptid,
>> not calculating it from lwp/pid. It's now questionable whether a wrapper
>> function is still needed, but that would be optimized to .pid () in future.
> 
> Yeah, but since NetBSD doesn't need get_ptrace_pid, I'd like if we could avoid
> making get_ptrace_pid more complex and if the ifdefs were concentrated around
> the ptrace calls (ideally, in as few spots possible, thanks to the gdb_ptrace
> functions).
> 
> Simon
> 

I've submitted a patch disabling get_ptrace_pid() for NetBSD and
switching in 2 places to ptid. pid()me file.

Not all ptrace(2) calls accept PID+LWP on NetBSD as here are calls that
work for the whole process such as DETACH, ATTACH, PT_SET_EVENT_MASK etc.

It's easier for now to go for this approach in the proposed patch.