gdbserver crash if gdb attaches too fast

With "maint set target-non-stop on", the attach tests occasionally
crash gdbserver.

Basically, gdb attaches with vAttach;PID, and then shortly after reads
the xml target description for that process, to figure out the
process' architecture.  On the gdbserver side, the target description
is only filled in when the first process/thread in the thread group
reports its initial PTRACE_ATTACH SIGSTOP.  So if GDB is fast enough,
it can read the target description _before_ that initial stop, and
then gdbserver dies dereferencing a NULL tdesc pointer.

gdb/gdbserver/ChangeLog:
2015-11-30  Pedro Alves  <palves@redhat.com>

	* linux-low.c (linux_attach): In non-stop mode, wait for one stop
	before returning.

2 files changed, 46 insertions, 12 deletions

diff --git a/gdb/gdbserver/ChangeLog b/gdb/gdbserver/ChangeLog
index 22d729f..3d694ac 100644
--- a/gdb/gdbserver/ChangeLog
+++ b/gdb/gdbserver/ChangeLog
@@ -1,5 +1,10 @@
 2015-11-30  Pedro Alves  <palves@redhat.com>
 
+	* linux-low.c (linux_attach): In non-stop mode, wait for one stop
+	before returning.
+
+2015-11-30  Pedro Alves  <palves@redhat.com>
+
 	* server.c (handle_v_requests): Handle vCtrlC.
 
 2015-11-30  Pedro Alves  <palves@redhat.com>
diff --git a/gdb/gdbserver/linux-low.c b/gdb/gdbserver/linux-low.c
index a70868c..69fda7f 100644
--- a/gdb/gdbserver/linux-low.c
+++ b/gdb/gdbserver/linux-low.c
@@ -1029,12 +1029,16 @@ attach_proc_task_lwp_callback (ptid_t ptid)
   return 0;
 }
 
+static void async_file_mark (void);
+
 /* Attach to PID.  If PID is the tgid, attach to it and all
    of its threads.  */
 
 static int
 linux_attach (unsigned long pid)
 {
+  struct process_info *proc;
+  struct thread_info *initial_thread;
   ptid_t ptid = ptid_build (pid, pid, 0);
   int err;
 
@@ -1045,17 +1049,12 @@ linux_attach (unsigned long pid)
     error ("Cannot attach to process %ld: %s",
 	   pid, linux_ptrace_attach_fail_reason_string (ptid, err));
 
-  linux_add_process (pid, 1);
-
-  if (!non_stop)
-    {
-      struct thread_info *thread;
+  proc = linux_add_process (pid, 1);
 
-     /* Don't ignore the initial SIGSTOP if we just attached to this
-	process.  It will be collected by wait shortly.  */
-      thread = find_thread_ptid (ptid_build (pid, pid, 0));
-      thread->last_resume_kind = resume_stop;
-    }
+  /* Don't ignore the initial SIGSTOP if we just attached to this
+     process.  It will be collected by wait shortly.  */
+  initial_thread = find_thread_ptid (ptid_build (pid, pid, 0));
+  initial_thread->last_resume_kind = resume_stop;
 
   /* We must attach to every LWP.  If /proc is mounted, use that to
      find them now.  On the one hand, the inferior may be using raw
@@ -1067,6 +1066,38 @@ linux_attach (unsigned long pid)
      that once thread_db is loaded, we'll still use it to list threads
      and associate pthread info with each LWP.  */
   linux_proc_attach_tgid_threads (pid, attach_proc_task_lwp_callback);
+
+  /* GDB will shortly read the xml target description for this
+     process, to figure out the process' architecture.  But the target
+     description is only filled in when the first process/thread in
+     the thread group reports its initial PTRACE_ATTACH SIGSTOP.  Do
+     that now, otherwise, if GDB is fast enough, it could read the
+     target description _before_ that initial stop.  */
+  if (non_stop)
+    {
+      struct lwp_info *lwp;
+      int wstat, lwpid;
+      ptid_t pid_ptid = pid_to_ptid (pid);
+
+      lwpid = linux_wait_for_event_filtered (pid_ptid, pid_ptid,
+					     &wstat, __WALL);
+      gdb_assert (lwpid > 0);
+
+      lwp = find_lwp_pid (pid_to_ptid (lwpid));
+
+      if (!WIFSTOPPED (wstat) || WSTOPSIG (wstat) != SIGSTOP)
+	{
+	  lwp->status_pending_p = 1;
+	  lwp->status_pending = wstat;
+	}
+
+      initial_thread->last_resume_kind = resume_continue;
+
+      async_file_mark ();
+
+      gdb_assert (proc->tdesc != NULL);
+    }
+
   return 0;
 }
 
@@ -2868,8 +2899,6 @@ linux_stabilize_threads (void)
     }
 }
 
-static void async_file_mark (void);
-
 /* Convenience function that is called when the kernel reports an
    event that is not passed out to GDB.  */