Linux: don't resume new LWPs until we've pulled all events out of the kernel

Since the starvation avoidance series
(https://sourceware.org/ml/gdb-patches/2014-12/msg00631.html), both
GDB and GDBserver pull all events out of ptrace before deciding which
event to process.

There's one problem with that though.  Because we resume new threads
immediately when we see a PTRACE_EVENT_CLONE event, if the program
constantly spawns threads fast enough, new threads can spawn threads
faster we can pull events out of the kernel, and thus we'd get stuck
in an infinite loop, never returning any event to the core to process.
I occasionally see this happen with the
attach-many-short-lived-threads.exp test against gdbserver.

The fix is to delay resuming new threads until we've pulled out all
events out of the kernel.

On native, we already have the resume_stopped_resumed_lwps function
that knows to resume LWPs that are stopped with no event to report to
the core.  So the patch just adds another use.  GDBserver didn't have
the equivalent yet, so the patch adds one.

Tested on x86_64 Fedora 20, native and gdbserver (remote and
extended-remote).

gdb/gdbserver/ChangeLog:
2015-02-04  Pedro Alves  <palves@redhat.com>

	* linux-low.c (handle_extended_wait): Don't resume LWPs here.
	(resume_stopped_resumed_lwps): New function.
	(linux_wait_for_event_filtered): Use it.

gdb/ChangeLog:
2015-02-04  Pedro Alves  <palves@redhat.com>

	* linux-nat.c (handle_extended_wait): Don't resume LWPs here.
	(wait_lwp): Don't call wait_lwp if linux_handle_extended_wait
	returns true.
	(resume_stopped_resumed_lwps): Don't check whether the thread is
	marked as executing.
	(linux_nat_wait_1): Use resume_stopped_resumed_lwps.

1 files changed, 36 insertions, 23 deletions

diff --git a/gdb/gdbserver/linux-low.c b/gdb/gdbserver/linux-low.c
index 5e37dd5..1682679 100644
--- a/gdb/gdbserver/linux-low.c
+++ b/gdb/gdbserver/linux-low.c
@@ -427,30 +427,12 @@ handle_extended_wait (struct lwp_info *event_child, int wstat)
       /* Normally we will get the pending SIGSTOP.  But in some cases
 	 we might get another signal delivered to the group first.
 	 If we do get another signal, be sure not to lose it.  */
-      if (WSTOPSIG (status) == SIGSTOP)
-	{
-	  if (stopping_threads == NOT_STOPPING_THREADS)
-	    linux_resume_one_lwp (new_lwp, 0, 0, NULL);
-	}
-      else
+      if (WSTOPSIG (status) != SIGSTOP)
 	{
 	  new_lwp->stop_expected = 1;
-
-	  if (stopping_threads != NOT_STOPPING_THREADS)
-	    {
-	      new_lwp->status_pending_p = 1;
-	      new_lwp->status_pending = status;
-	    }
-	  else
-	    /* Pass the signal on.  This is what GDB does - except
-	       shouldn't we really report it instead?  */
-	    linux_resume_one_lwp (new_lwp, 0, WSTOPSIG (status), NULL);
+	  new_lwp->status_pending_p = 1;
+	  new_lwp->status_pending = status;
 	}
-
-      /* Always resume the current thread.  If we are stopping
-	 threads, it will have a pending SIGSTOP; we may as well
-	 collect it now.  */
-      linux_resume_one_lwp (event_child, event_child->stepping, 0, NULL);
     }
 }
 
@@ -1955,6 +1937,32 @@ linux_low_filter_event (int lwpid, int wstat)
   return child;
 }
 
+/* Resume LWPs that are currently stopped without any pending status
+   to report, but are resumed from the core's perspective.  */
+
+static void
+resume_stopped_resumed_lwps (struct inferior_list_entry *entry)
+{
+  struct thread_info *thread = (struct thread_info *) entry;
+  struct lwp_info *lp = get_thread_lwp (thread);
+
+  if (lp->stopped
+      && !lp->status_pending_p
+      && thread->last_resume_kind != resume_stop
+      && thread->last_status.kind == TARGET_WAITKIND_IGNORE)
+    {
+      int step = thread->last_resume_kind == resume_step;
+
+      if (debug_threads)
+	debug_printf ("RSRL: resuming stopped-resumed LWP %s at %s: step=%d\n",
+		      target_pid_to_str (ptid_of (thread)),
+		      paddress (lp->stop_pc),
+		      step);
+
+      linux_resume_one_lwp (lp, step, GDB_SIGNAL_0, NULL);
+    }
+}
+
 /* Wait for an event from child(ren) WAIT_PTID, and return any that
    match FILTER_PTID (leaving others pending).  The PTIDs can be:
    minus_one_ptid, to specify any child; a pid PTID, specifying all
@@ -2088,8 +2096,13 @@ linux_wait_for_event_filtered (ptid_t wait_ptid, ptid_t filter_ptid,
 	  continue;
 	}
 
-      /* Now that we've pulled all events out of the kernel, check if
-	 there's any LWP with a status to report to the core.  */
+      /* Now that we've pulled all events out of the kernel, resume
+	 LWPs that don't have an interesting event to report.  */
+      if (stopping_threads == NOT_STOPPING_THREADS)
+	for_each_inferior (&all_threads, resume_stopped_resumed_lwps);
+
+      /* ... and find an LWP with a status to report to the core, if
+	 any.  */
       event_thread = (struct thread_info *)
 	find_inferior (&all_threads, status_pending_p_callback, &filter_ptid);
       if (event_thread != NULL)