tree 00246960f9cc7b704f2cefd420fc8b337c57d02e
parent 7ed689ad61de0cbfe4e5a6f18f097776128202e4
author Pedro Alves <palves@redhat.com> 1402481071 +0100
committer Pedro Alves <palves@redhat.com> 1402481071 +0100

PR remote/17028: GDB+GDBserver hangs on Windows

Since target-async was turned on by default, debugging on Windows
using GDB+GDBserver sometimes hangs while waiting for a RSP reply.

The problem is a race in the gdb_select machinery.

This is what we see for a faulty next on the GDB side:

    (gdb) n
    infrun: clear_proceed_status_thread (Thread 4424)
    infrun: proceed (addr=0xffffffff, signal=GDB_SIGNAL_DEFAULT, step=1)
    (...)
    infrun: resume (step=1, signal=GDB_SIGNAL_0), ...
    Sending packet: $vCont;s:1148;c#5e...
    *hang*

At this point, attaching a debugger to the hanging GDB confirms that
it is blocked, waiting for a socket event:

    #6  0x757841d8 in WaitForMultipleObjects ()
       from C:\Windows\syswow64\kernel32.dll
    #7  0x004708e7 in gdb_select (n=469, readfds=0x88ca50 <gdb_notifier+784>,
        writefds=0x88cb54 <gdb_notifier+1044>,
        exceptfds=0x88cc58 <gdb_notifier+1304>, timeout=0x0)
        at /[...]/gdb/mingw-hdep.c:172
    #8  0x00527926 in gdb_wait_for_event (block=1)
        at /[...]/gdb/event-loop.c:831
    #9  0x00526ff1 in gdb_do_one_event ()
        at /[...]/gdb/event-loop.c:403

However, on the GDBserver side, we see that GDBserver already sent a
T05 packet reply:

    gdbserver: kernel event EXCEPTION_DEBUG_EVENT for pid=4968 tid=1148
    EXCEPTION_SINGLE_STEP
    Child Stopped with signal = 5
    Writing resume reply for LWP 4968.4424:1
    DEBUG: write_prim ($T0505:c8fe2800;04:a0fe2800;08:38164000;thread:1148;#f0)
           -> 55

To recap, on Windows, 'select' only works with sockets, so we have a
wrapper, gdb_select, that uses the GDB serial abstraction to handle
sockets, consoles, pipes, and serial ports.  Each serial descriptor
has a thread associated (we call those the select threads), and those
threads communicate with the main thread by means of standard Windows
events.

It basically goes like this: gdb_select first loops through all fds of
interest, calling their wait_handle hooks, which returns an event that
WaitForMultipleObjects can wait on.  gdb_select then blocks in
WaitForMultipleObjects with all those event handles.  The wait_handle
hook is responsible for arranging for the returned event to become set
once data is available.  This is done by setting the descriptor's
helper thread running, which itself knows how to wait for data from
the type of handle it manages (sockets, pipes, consoles, files, etc.).
Once data arrives, the select thread sets the corresponding event
which unblocks WaitForMultipleObjects within gdb_select.  However, the
wait_handle hook can also apply an optimization: if data is already
pending, then there's no need to set the thread running, and the
descriptors event can be set immediately.  It's around this latter
aspect that lies the bug/race.

Adding some ad hoc debug logs to ser-mingw.c and mingw-hdep.c, we see
the following sequence of events, right after sending
"$vCont;s:1148;c#5e".  Thread 1 is the main thread, and thread 2 is
the socket's helper/select thread.  gdb_select was only passed one
descriptor to wait on, the remote target's socket.
net_windows_select_thread is the entry point of the select threads for
sockets.

 #1 - thread 1: gdb_select: enter
 #2 - thread 2: net_windows_select_thread: WaitForMultipleObjects blocking

gdb_select walked over the wait_handle hooks, and woke up the socket's
helper thread.  The helper thread is now blocked waiting for socket
events.

 #3 - thread 1: gdb_select: WaitForMultipleObjects polling (timeout=0ms)
 #4 - thread 1: gdb_select: WaitForMultipleObjects returned 102 (WAIT_TIMEOUT)

There was no pending data available yet, and gdb_select was passed
timeout==0ms, and so WaitForMultipleObjects times out immediately.

 #5 - thread 2: net_windows_select_thread: WaitForMultipleObjects returned 1

Just afterwards, socket data arrives, and thread 2 wakes up.  Thread 2
calls WSAEnumNetworkEvents, which clears state->sock_event, and marks
the serial's read_event event, telling the main thread that data is
available.

 #6 - thread 1: gdb_select: call serial_done_wait_handle on each serial

gdb_select stops all the helper/select threads.

 #7 - thread 1: gdb_select: return 0 (WAIT_TIMEOUT)

gdb_select in the main thread returns to the caller.

Note that at this point, data is pending on the socket, the serial's
read_event is set, but the socket's sock_event event is not set, until
_further_ data arrives.

Now GDB does its thing and goes back to the event loop.  That calls
gdb_select, but with timeout==INFINITE.

Again, gdb_select calls the socket serial's wait_handle hook.  It
first clears its events, starting from a clean slate:

  ResetEvent (state->base.read_event);
  ResetEvent (state->base.except_event);
  ResetEvent (state->base.stop_select);

That cleared read_event, which was previously set in #5 above.  And
then it checks for pending events, in the sock_event event:

  /* Check any pending events.  This both avoids starting the thread
     unnecessarily, and handles stray FD_READ events (see below).  */
  if (WaitForSingleObject (state->sock_event, 0) == WAIT_OBJECT_0)
    {

That also fails because state->sock_event was cleared in #5 too...

So the wait_handle hook erroneously decides that it needs to start the
helper thread to wait for input:

 #8 - thread 2: net_windows_select_thread: WaitForMultipleObjects blocking
 #9 - thread 1: gdb_select: WaitForMultipleObjects blocking (INFINITE)

But, GDBserver already sent all it had to send, so both threads waits
forever...

At first I thought that net_windows_wait_handle shouldn't be resetting
state->base.read_event or state->base.except_event, but looking
deeper, the pipe and console wait_handle hooks reset all events too.
It actually makes sense that way -- consuming an event from different
threads is bad practice, and, we should always be able to query
pending state without looking at the state->sock_event from within
net_windows_wait_handle.  The end result is much simpler, and makes
net_windows_select_thread look a lot like console_select_thread,
actually.

gdb/
2014-06-11  Pedro Alves  <palves@redhat.com>

	PR remote/17028
	* ser-mingw.c (net_windows_socket_check_pending): New function.
	(net_windows_select_thread): Ignore spurious wakeups.  Use
	net_windows_socket_check_pending.
	(net_windows_wait_handle): Check for pending events with
	ioctlsocket, through net_windows_socket_check_pending, instead of
	checking the socket's event.