pyln: Work around the socket path length on Linux OSs

Some Linux OSs impose a length limit on the path a Unix socket may have. This
is not an issue in `lightningd` since we `chdir()` into that directory before
opening the socket, however in pyln this became a problem for some tests,
since we use absolute paths in the testing framework. It's also a rather
strange quirk to expose to users.

This patch introduces a `UnixSocket` abstraction that attempts to work around
these limitations by aliasing the directory containing the socket into
`/proc/self/fd` and then connecting using that alias.

It was inspired by Open vSwitch code here https://github.com/openvswitch/ovs/blob/master/python/ovs/socket_util.py

Signed-off-by: Christian Decker <@cdecker>

contrib/pyln-client/pyln/client/lightning.py

@@ -1,9 +1,10 @@
+from decimal import Decimal
+from math import floor, log10
 import json
 import logging
+import os
 import socket
 import warnings
-from decimal import Decimal
-from math import floor, log10
 
 
 class RpcError(ValueError):
@@ -159,6 +160,73 @@ class Millisatoshi:
         return Millisatoshi(int(self) + int(other))
 
 
+class UnixSocket(object):
+    """A wrapper for socket.socket that is specialized to unix sockets.
+
+    Some OS implementations impose restrictions on the Unix sockets.
+
+     - On linux OSs the socket path must be shorter than the in-kernel buffer
+       size (somewhere around 100 bytes), thus long paths may end up failing
+       the `socket.connect` call.
+
+    This is a small wrapper that tries to work around these limitations.
+
+    """
+
+    def __init__(self, path):
+        self.path = path
+        self.sock = None
+        self.connect()
+
+    def connect(self):
+        try:
+            self.sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+            return self.sock.connect(self.path)
+        except OSError as e:
+            self.sock.close()
+
+            if (e.args[0] == "AF_UNIX path too long" and os.uname()[0] == "Linux"):
+                # If this is a Linux system we may be able to work around this
+                # issue by opening our directory and using `/proc/self/fd/` to
+                # get a short alias for the socket file.
+                #
+                # This was heavily inspired by the Open vSwitch code see here:
+                # https://github.com/openvswitch/ovs/blob/master/python/ovs/socket_util.py
+
+                dirname = os.path.dirname(self.path)
+                basename = os.path.basename(self.path)
+
+                # Open an fd to our home directory, that we can then find
+                # through `/proc/self/fd` and access the contents.
+                dirfd = os.open(dirname, os.O_DIRECTORY | os.O_RDONLY)
+                short_path = "/proc/self/fd/%d/%s" % (dirfd, basename)
+                self.sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+                return self.sock.connect(short_path)
+            else:
+                # There is no good way to recover from this.
+                raise
+
+    def close(self):
+        if self.sock is not None:
+            self.sock.close()
+        self.sock = None
+
+    def sendall(self, b):
+        if self.sock is None:
+            raise socket.error("not connected")
+
+        self.sock.sendall(b)
+
+    def recv(self, length):
+        if self.sock is None:
+            raise socket.error("not connected")
+
+        return self.sock.recv(length)
+
+    def __del__(self):
+        self.close()
+
+
 class UnixDomainSocketRpc(object):
     def __init__(self, socket_path, executor=None, logger=logging, encoder_cls=json.JSONEncoder, decoder=json.JSONDecoder()):
         self.socket_path = socket_path
@@ -215,8 +283,7 @@ class UnixDomainSocketRpc(object):
             payload = {k: v for k, v in payload.items() if v is not None}
 
         # FIXME: we open a new socket for every readobj call...
-        sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
-        sock.connect(self.socket_path)
+        sock = UnixSocket(self.socket_path)
         self._writeobj(sock, {
             "method": method,
             "params": payload,

tests/test_misc.py

@@ -1,11 +1,11 @@
 from bitcoin.rpc import RawProxy
 from decimal import Decimal
 from fixtures import *  # noqa: F401,F403
-from fixtures import TEST_NETWORK
+from fixtures import LightningNode, TEST_NETWORK
 from flaky import flaky  # noqa: F401
 from lightning import RpcError
 from threading import Event
-from utils import DEVELOPER, TIMEOUT, VALGRIND, sync_blockheight, only_one, wait_for, TailableProc, env
+from pyln.testing.utils import DEVELOPER, TIMEOUT, VALGRIND, sync_blockheight, only_one, wait_for, TailableProc, env
 from ephemeral_port_reserve import reserve
 
 import json
@@ -2002,3 +2002,20 @@ def test_unicode_rpc(node_factory):
     assert(len(invoices) == 1)
     assert(invoices[0]['description'] == desc)
     assert(invoices[0]['label'] == desc)
+
+
+@unittest.skipIf(VALGRIND, "Testing pyln doesn't exercise anything interesting in the c code.")
+def test_unix_socket_path_length(node_factory, bitcoind, directory, executor, db_provider, test_base_dir):
+    lightning_dir = os.path.join(directory, "anode" + "far" * 30 + "away")
+    os.makedirs(lightning_dir)
+    db = db_provider.get_db(lightning_dir, "test_unix_socket_path_length", 1)
+
+    l1 = LightningNode(1, lightning_dir, bitcoind, executor, db=db, port=node_factory.get_next_port())
+
+    # `LightningNode.start()` internally calls `LightningRpc.getinfo()` which
+    # exercises the socket logic, and raises an issue if it fails.
+    l1.start()
+
+    # Let's just call it again to make sure it really works.
+    l1.rpc.listconfigs()
+    l1.stop()