Re: [PATCH] Remove lwp -> pid conversion in linux_nat_xfer_partial

[Pedro Alves <palves at redhat dot com>]

On 03/22/2017 12:42 AM, Simon Marchi wrote:

>>> There is also linux_proc_xfer_partial and linux_proc_xfer_spu, which
>>> both only use the pid field of inferior_ptid and ignore lwp.  However,
>>> since they use "/proc/<pid>", using the id of any thread in the process
>>> will give the same result (AFAIK).
>>
>> It's generally better to use the lwp id:
>>
>> - some files under /proc/<pid>/ may not work if the <pid> thread is
>>   running, just like ptrace requires a stopped thread.  The current
>>   thread's lwp id is more likely to be in the necessary state (stopped).
>>
>> - if the leader exits, and goes zombie, then several files under
>>   "/proc/<pid>" won't work, though using "/proc/<pid>/task/<tid>" would.
>>   (try poking at leader-exit.exp a bit.)
>>   The latter path form is also generally better for being robust in
>>   the case TID exits and is reused in another process, much like
>>   tkill vs tgkill.
> 
> I thought that the process exited whenever the main thread exits, that's
> not the case?  I guess not if there's a test for it...

Nope.  If you call "exit", then yes.  The kernel kills the whole thread
group in response to that system call.  If the leader does
pthread_exit, then no, the thread group stays around until all children
exit too.  The kernel won't report the main thread's exit status (i.e.,
we can't reap that zombie, and we'd hang if we tried a blocking waitpid)
until all the children are reaped first.  That's why we have
linux-nat.c:check_zombie_leaders (and the equivalent in gdbserver).

>> So if possible to switch those spots too, I'd recommend/prefer it.
> 
> Ok, I'll just replace ptid_get_pid with get_ptrace_pid* in this patch

Since this is linux-specific code, you should be able to use
ptid_get_lwp directly.

> and look at using /proc/<pid>/task/<tid> after.  When doing the latter,
> do I still have to consider cases where ptid is a single-process/thread
> ptid (lwp == 0)?  From my experience, there's always a lwp on Linux, but
> perhaps there are some setups I don't know about with which it can happen?

Right, on Linux there's always an lwp.  Before NPTL, the 
/proc/<pid>/task/<tid> path didn't exist at all, but we no longer
support LinuxThreads.

Thanks,
Pedro Alves