adb: fix adb client running out of sockets on Windows
Background
==========
On Windows, if you run "adb shell exit" in a loop in two windows,
eventually the adb client will be unable to connect to the adb server. I
think connect() is returning WSAEADDRINUSE: "Only one usage of each
socket address (protocol/network address/port) is normally permitted.
(10048)". The Windows System Event Log may also show Event 4227, Tcpip.
Netstat output is filled with:
# for the adb server
TCP 127.0.0.1:5037 127.0.0.1:65523 TIME_WAIT
# for the adb client
TCP 127.0.0.1:65523 127.0.0.1:5037 TIME_WAIT
The error probably means that the client is running out of free
address:port pairs.
The first netstat line is unavoidable, but the second line exists
because the adb client is not waiting for orderly/graceful shutdown of
the socket, and that is apparently required on Windows to get rid of the
second line. For more info, see
https://github.com/CompareAndSwap/SocketCloseTest .
This is exacerbated by the fact that "adb shell exit" makes 4 socket
connections to the adb server: 1) host:version, 2) host:features, 3)
host:version (again), 4) shell:exit. Also exacerbating is the fact that
the adb protocol is length-prefixed so the client typically does not
have to 'read() until zero' which effectively waits for orderly/graceful
shutdown.
The Fix
=======
Introduce a function, ReadOrderlyShutdown(), that should be called in
the adb client to wait for the server to close its socket, before
closing the client socket.
I reviewed all code where the adb client makes a connection to the adb
server and added ReadOrderlyShutdown() when it made sense. I wasn't able
to add it to the following:
* interactive_shell: this doesn't matter because this is interactive and
thus can't be run fast enough to use up ports.
* adb sideload: I couldn't get enough test coverage and I don't think
this is being called frequently enough to be a problem.
* send_shell_command, backup, adb_connect_command, adb shell, adb
exec-out, install_multiple_app, adb_send_emulator_command: These
already wait for server socket shutdown since they already call
recv() until zero.
* restore, adb exec-in: protocol design can't have the server close
first.
* adb start-server: no fd is actually returned
* create_local_service_socket, local_connect_arbitrary_ports,
connect_device: probably called rarely enough not to be a problem.
Also in this change
===================
* Clarify comments in when adb_shutdown() is called before exit().
* add some missing adb_close() in adb sideload.
* Fixup error handling and comments in adb_send_emulator_command().
* Make SyncConnection::SendQuit return a success boolean.
* Add unittest for adb emu kill command. This gets code coverage over
this very careful piece of code.
Change-Id: Iad0b1336f5b74186af2cd35f7ea827d0fa77a17c
Signed-off-by: Spencer Low <CompareAndSwap@gmail.com>
diff --git a/adb/adb_io.cpp b/adb/adb_io.cpp
index 7788996..a37fbc0 100644
--- a/adb/adb_io.cpp
+++ b/adb/adb_io.cpp
@@ -137,3 +137,43 @@
return WriteFdExactly(fd, str);
}
+
+bool ReadOrderlyShutdown(int fd) {
+ char buf[16];
+
+ // Only call this function if you're sure that the peer does
+ // orderly/graceful shutdown of the socket, closing the socket so that
+ // adb_read() will return 0. If the peer keeps the socket open, adb_read()
+ // will never return.
+ int result = adb_read(fd, buf, sizeof(buf));
+ if (result == -1) {
+ // If errno is EAGAIN, that means this function was called on a
+ // nonblocking socket and it would have blocked (which would be bad
+ // because we'd probably block the main thread where nonblocking IO is
+ // done). Don't do that. If you have a nonblocking socket, use the
+ // fdevent APIs to get called on FDE_READ, and then call this function
+ // if you really need to, but it shouldn't be needed for server sockets.
+ CHECK_NE(errno, EAGAIN);
+
+ // Note that on Windows, orderly shutdown sometimes causes
+ // recv() == SOCKET_ERROR && WSAGetLastError() == WSAECONNRESET. That
+ // can be ignored.
+ return false;
+ } else if (result == 0) {
+ // Peer has performed an orderly/graceful shutdown.
+ return true;
+ } else {
+ // Unexpectedly received data. This is essentially a protocol error
+ // because you should not call this function unless you expect no more
+ // data. We don't repeatedly call adb_read() until we get zero because
+ // we don't know how long that would take, but we do know that the
+ // caller wants to close the socket soon.
+ VLOG(RWX) << "ReadOrderlyShutdown(" << fd << ") unexpectedly read "
+ << dump_hex(buf, result);
+ // Shutdown the socket to prevent the caller from reading or writing to
+ // it which doesn't make sense if we just read and discarded some data.
+ adb_shutdown(fd);
+ errno = EINVAL;
+ return false;
+ }
+}