Return a special error code if a debugger request is made for a dead …

…target.

src/replayer/dbg_gdb.c

@@ -713,6 +713,24 @@ static int process_packet(struct dbg_context* dbg)
  return ret;
 }
 
+void dbg_notify_no_such_thread(struct dbg_context* dbg,
+ const struct dbg_request* req)
+{
+ assert(!memcmp(&dbg->req, req, sizeof(dbg->req)));
+
+ /* '10' is the errno ECHILD. We use it as a magic code to
+ * notify the user that the thread that was the target of this
+ * request has died, and either gdb didn't notice that, or rr
+ * didn't notify gdb. Either way, the user should restart
+ * their debugging session. */
+ log_err(
+"Targeted thread no longer exists; this is the result of either a gdb or\n"
+"rr bug. Please restart your debugging session and avoid doing whatever\n"
+"triggered this bug.");
+ write_packet(dbg, "E10");
+ consume_request(dbg);
+}
+
 struct dbg_request dbg_get_request(struct dbg_context* dbg)
 {
  /* Can't ask for the next request until you've satisfied the

src/replayer/dbg_gdb.h

@@ -131,6 +131,14 @@ struct dbg_context* dbg_await_client_connection(const char* addr,
  unsigned short port,
  int probe);
 
+/**
+ * Call this when the target of |req| is needed to fulfill the
+ * request, but the target is dead. This situation is a symptom of a
+ * gdb or rr bug.
+ */
+void dbg_notify_no_such_thread(struct dbg_context* dbg,
+ const struct dbg_request* req);
+
 /**
  * Return the current request made by the debugger host, that needs to
  * be satisfied. This function will block until either there's a

src/replayer/replayer.c

@@ -346,8 +346,12 @@ static struct dbg_request process_debugger_requests(struct dbg_context* dbg,
  }
 
  /* These requests require a valid target task. We
- * trust gdb to use the information provided above to
- * only query valid tasks. */
+ * don't trust the debugger to use the information
+ * provided above to only query valid tasks. */
+ if (!target) {
+ dbg_notify_no_such_thread(dbg, &req);
+ continue;
+ }
  switch (req.type) {
  case DREQ_GET_MEM: {
  size_t len;

src/test/dead_thread_target.py

@@ -1,4 +1,5 @@
 from rrutil import *
+import re
 
 send_gdb('b hit_barrier\n')
 expect_gdb('Breakpoint 1')
@@ -16,6 +17,18 @@
 expect_gdb('Switching to thread 2')
 
 send_gdb('c\n')
-expect_gdb('Breakpoint 2, joined_threads')
+# TODO: with the gdb in fedora 19, if a thread dies while it's the
+# resume target, then rr notifies gdb, but gdb doesn't ask for a new
+# thread list. This seems like a gdb bug, because we don't have any
+# other way to notify gdb of thread death, and the same code works
+# just fine in concurrent ubuntu and older versions.
+#
+# So we work around that problem by returning this special error code
+# to the user. Once gdb has made this mistake, the debugging session
+# is "stuck" because won't let any other threads continue. But at
+# least this error code tells the user that they need to restart the
+# session.
+expect_gdb(re.compile(
+ r'Breakpoint 2, joined_threads|Remote failure reply: E10'))
 
 ok()