[1/3] PR remote/19496, internal err forking-threads-plus-bkpt

Message ID 1453942111-1215-2-git-send-email-donb@codesourcery.com
State New, archived

Commit Message

Don Breazeal Jan. 28, 2016, 12:48 a.m. UTC
  This patch fixes an internal error that occurs in

/blah/binutils-gdb/gdb/target.c:2723: internal-error: Can't determine the
current address space of thread Thread 3170.3170

In default_thread_address_space, find_inferior_ptid couldn't find 3170.3170
because it had been overwritten in inferior_appeared, called as follows:


The cause of the problem was the following sequence of events:

* GDB knows only about the main thread

* the first fork event is reported to GDB, saved as pending_event

* qXfer:threads_read gets the threads from the remote.
  remove_new_fork_children id's the fork child from the pending event
  and removes it from the list reported to GDB.  All the rest of the
  threads, including the fork parent, are added to the GDB thread list.

* GDB stops all the threads.  All the stop events are pushed onto the
  stop reply queue behind the pending fork event.

* remote_wait_ns calls queued_stop_reply and process_stop_reply to
  remove the fork event from the front of the stop reply queue and save
  event information in the thread_info structure for the fork parent
  thread.  Unfortunately, none of the information saved in this way is
  the fork-specific information, so the actual fork event info is lost.

* A subsequent qXfer:threads:read packet gets the thread list including
  the fork parent and fork child.  remove_new_fork_children checks the
  thread list to see if there is a fork parent, doesn't find one, checks
  the stop reply queue for a pending fork event, doesn't find one, and
  allows the fork child thread to be reported to GDB before the fork
  event has been handled.  remote_update_thread_list calls
  remote_notice_new_thread and overwrites the current (main) thread in
  inferior_appeared.  GDB has now lost all knowledge of the main thread,
  and an internal error results.

The fix was to make sure that when the stop reply was removed from the
stop reply queuei, all of the necessary fork event information was stored
in the parent thread structure.  In process_stop_reply we call a new
function, update_thread_if_fork_parent, to store the pending_follow
information from the fork stop reply in the fork parent thread.

Tested on x86_64 and Nios II Linux.  No regressions, but more failures,
which are addressed in subsequent patches in this patchset.


2016-01-27  Don Breazeal  <donb@codesourcery.com>

	PR remote/19496
	* remote.c (update_thread_if_fork_parent): New function.
	(process_stop_reply): Call update_thread_if_fork_parent.

 gdb/remote.c | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)


diff --git a/gdb/remote.c b/gdb/remote.c
index b0303f6..f072ce4 100644
--- a/gdb/remote.c
+++ b/gdb/remote.c
@@ -6799,6 +6799,25 @@  remote_notif_get_pending_events (struct notif_client *nc)
+/* Check if the specified stop reply is for a fork event.  If it is,
+   update the corresponding thread to contain the pending follow
+   information required to identify it as the fork parent.  */
+static void
+update_thread_if_fork_parent (struct stop_reply *stop_reply)
+  ptid_t ptid;
+  ptid = stop_reply->ptid;
+  if (stop_reply->ws.kind == TARGET_WAITKIND_FORKED
+      || stop_reply->ws.kind == TARGET_WAITKIND_VFORKED)
+    {
+      struct thread_info *tp = find_thread_ptid (ptid);
+      tp->pending_follow = stop_reply->ws;
+    }
 /* Called when it is decided that STOP_REPLY holds the info of the
    event that is to be returned to the core.  This function always
    destroys STOP_REPLY.  */
@@ -6844,8 +6863,11 @@  process_stop_reply (struct stop_reply *stop_reply,
       remote_thr->core = stop_reply->core;
       remote_thr->stop_reason = stop_reply->stop_reason;
       remote_thr->watch_data_address = stop_reply->watch_data_address;
-    }
+      /* Make sure we record any pending fork events.  */
+      update_thread_if_fork_parent (stop_reply);
+    }
   stop_reply_xfree (stop_reply);
   return ptid;