Re: [PATCH] PR threads/20743: Don't attempt to suspend or resume exited threads.

[John Baldwin <jhb@freebsd.org>]

On Tuesday, December 27, 2016 05:43:29 PM Vasil Dimov wrote:
> On Fri, Dec 23, 2016 at 15:43:19 -0600, Luis Machado wrote:
> > On 12/23/2016 03:28 PM, John Baldwin wrote:
> > > When resuming a native FreeBSD process, ignore exited threads when
> > > suspending/resuming individual threads prior to continuing the process.
> > >
> > > gdb/ChangeLog:
> > >
> > > 	PR threads/20743
> > > 	* fbsd-nat.c (resume_one_thread_cb): Ignore exited threads.
> > > 	(resume_all_threads_cb): Likewise.
> > > 	(fbsd_resume): Assert resuming thread has not exited.
> > > ---
> > >  gdb/ChangeLog  | 7 +++++++
> > >  gdb/fbsd-nat.c | 7 +++++++
> > >  2 files changed, 14 insertions(+)
> > >
> > > diff --git a/gdb/ChangeLog b/gdb/ChangeLog
> > > index db6e913..4fb3732 100644
> > > --- a/gdb/ChangeLog
> > > +++ b/gdb/ChangeLog
> > > @@ -1,3 +1,10 @@
> > > +2016-12-23  John Baldwin  <jhb@FreeBSD.org>
> > > +
> > > +	PR threads/20743
> > > +	* fbsd-nat.c (resume_one_thread_cb): Ignore exited threads.
> > > +	(resume_all_threads_cb): Likewise.
> > > +	(fbsd_resume): Assert resuming thread has not exited.
> > > +
> > >  2016-12-22  Doug Evans  <xdje42@gmail.com>
> > >
> > >  	* infrun.c (set_step_over_info): Add comment.
> > > diff --git a/gdb/fbsd-nat.c b/gdb/fbsd-nat.c
> > > index ade62f1..7cd08c6 100644
> > > --- a/gdb/fbsd-nat.c
> > > +++ b/gdb/fbsd-nat.c
> > > @@ -662,6 +662,9 @@ resume_one_thread_cb (struct thread_info *tp, void *data)
> > >    if (ptid_get_pid (tp->ptid) != ptid_get_pid (*ptid))
> > >      return 0;
> > >
> > > +  if (is_exited (tp->ptid))
> > > +    return 0;
> > > +		
> > >    if (ptid_get_lwp (tp->ptid) == ptid_get_lwp (*ptid))
> > >      request = PT_RESUME;
> > >    else
> > > @@ -680,6 +683,9 @@ resume_all_threads_cb (struct thread_info *tp, void *data)
> > >    if (!ptid_match (tp->ptid, *filter))
> > >      return 0;
> > >
> > > +  if (is_exited (tp->ptid))
> > > +    return 0;
> > > +		
> > >    if (ptrace (PT_RESUME, ptid_get_lwp (tp->ptid), NULL, 0) == -1)
> > >      perror_with_name (("ptrace"));
> > >    return 0;
> > > @@ -711,6 +717,7 @@ fbsd_resume (struct target_ops *ops,
> > >    if (ptid_lwp_p (ptid))
> > >      {
> > >        /* If ptid is a specific LWP, suspend all other LWPs in the process.  */
> > > +      gdb_assert (!is_exited (ptid));
> > 
> > If we're asserting on this (since supposedly it shouldn't happen), do we 
> > need to check for is_exited on the two functions above?
> > 
> > Also, is there a reason why we're not detecting a thread that has 
> > exited? Aren't all threads stopped at this point (for all-stop mode at 
> > least)?
> [...]
> 
> Hello,
> 
> I just nailed this down after it has been annoying me for some time,
> fixed it with a similar patch as the one submitted by John, and came
> here to report it.
> 
> The reason that we are "not detecting" an exited thread (at least in the
> scenario I got is), gdb/thread.c:
> 
>   --- cut ---
>   static void
>   delete_thread_1 (ptid_t ptid, int silent)
>   {
>   ...
>     /* If this is the current thread, or there's code out there that
>        relies on it existing (refcount > 0) we can't delete yet.  Mark
>        it as exited, and notify it.  */
>     if (tp->refcount > 0
>         || ptid_equal (tp->ptid, inferior_ptid))
>       {
>   ...
>          /* Will be really deleted some other time.  */
>          printf_unfiltered ("========== Will be really deleted some other time %u\n", ptid);
>          return;
>        }
>   ...
>   if (tpprev)
>     tpprev->next = tp->next;
>   else
>     thread_list = tp->next;
>   --- cut ---
> 
> In my scenario tp->refcount is 0, but
> "ptid_equal (tp->ptid, inferior_ptid)" is true, so the thread's entry is
> not removed from the global "threads_list".
> 
> The gdb output (with "set debug fbsd-lwp" enabled):
> 
>   --- cut ---
>   FLWP: adding thread for LWP 102009
>   [New LWP 102009 of process 40304]
>   FLWP: fbsd_resume for ptid (-1, 0, 0)
>   FLWP: fbsd_resume for ptid (40304, 102009, 0)
>   FLWP: fbsd_resume for ptid (-1, 0, 0)
>   FLWP: fbsd_resume for ptid (40304, 102009, 0)
>   FLWP: fbsd_resume for ptid (-1, 0, 0)
>   FLWP: deleting thread for LWP 102009
>   [LWP 102009 of process 40304 exited]
>   ...
>   ptrace: No such process.
>   --- cut ---
> 
> Hope this helps.

In particular, the sequence of events is this:

- an LWP (T1) reports a "normal" event (in the test case it is hitting a
  breakpoint).  This is reported to the core and sets the current thread
  (and thus inferior_ptid) to T1.
- the same LWP (T1) then exits and a thread exit event is reported via
  ptrace() to the native target.  The native target calls delete_thread,
  but the thread is not removed, just marked EXITING since it ==
  inferior_ptid as Vasil noted.  The native target just
  continues the process explicitly via ptrace() without reporting any
  event to the core aside from the call to delete_thread().
- some other LWP (T2) reports an event (in the test case it is a
  breakpoint).
- the user continues which invokes fbsd_resume() which wants to resume
  all threads.  Here iterate_over_threads() in fbsd_resume() will
  encounters the exited thread for T1 since nothing has called
  thread_update_list() (which would invoke delete_exited_threads() from
  fbsd_update_thread_list()).  Since the thread is exited, trying to
  manipulate it via ptrace() results in an error.

I have tried changing fbsd_wait() to return a TARGET_WAITKIND_SPURIOUS
instead of explicitly continuing the process, but that doesn't help, and it
means that the ptid being returned is still T1 in that case.

I'm not sure if I should explicitly be calling delete_exited_threads() in
fbsd_resume() before calling iterate_threads()?  Alternatively, fbsd_resume()
could use ALL_NONEXITED_THREADS() instead of iterate_threads() (it isn't
clear to me which of these is preferred since both are in use).

I added the assertion for my own sanity.  I suspect gdb should never try to
invoke target_resume() with a ptid of an exited thread, but if for some
reason it did the effect on FreeBSD would be a hang since we would suspend
all the other threads and when the process was continued via PT_CONTINUE it
would have nothing to do and would never return from wait().  I'd rather have
gdb fail an assertion in that case rather than hang.

-- 
John Baldwin