uvloop does not support original unix udp socketpair

[ZhuYuJin]

import asyncio
import os, socket
import uvloop
asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())

try:
 os.unlink("./tmp")
except:
 pass
try:
 os.unlink("./tmp1")
except:
 pass

socket_pair = socket.socketpair(socket.AF_UNIX, socket.SOCK_DGRAM, 0)

class WorkerConnection(asyncio.Protocol):
 def __init__(self):
 print("init connection in worker")
 self.aio_connection = None

 def connection_made(self, transport):
 self.aio_connection = transport

 def datagram_received(self, data: bytes, addr):
 	print(data)

class DispatcherConnection(asyncio.Protocol):
 def __init__(self):
 print("init connection in dispatcher")
 self.aio_connection = None

 def connection_made(self, transport):
 self.aio_connection = transport

 def datagram_received(self, data: bytes, addr):
 	print(data)

loop = uvloop.new_event_loop()
asyncio.set_event_loop(loop)

coros = loop.create_datagram_endpoint(WorkerConnection, sock=socket_pair[0])
(transport1, protocol1) = loop.run_until_complete(coros)

coros = loop.create_datagram_endpoint(DispatcherConnection, sock=socket_pair[1])
(transport2, protocol2) = loop.run_until_complete(coros)

transport1.sendto(b'transport1')
transport2.sendto(b'transport2')

loop.run_forever()

I initialize a unix udp socketpair, and try to use uvloop to manage both sockets.
The code above has the following err:

However, when I bind address to both sockets, the err disappears.

socket_pair[0].bind("./tmp")
socket_pair[1].bind("./tmp1")

It shows uvloop does not support the original unix udp socketpair.

[fantix]

This seems to be a low-level behavior:

#include <unistd.h>
#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <sys/errno.h>
#include <sys/ioctl.h>


void do_send(int fd) {
 int32_t w_buf = 32;
 write(fd, &w_buf, sizeof(w_buf));
}

void do_recv(int fd) {
 // copied from uv__udp_recvmsg
 struct sockaddr_storage peer;
 struct msghdr h;
 ssize_t nread;
 struct iovec buf;
 int r;
 int nonblocking = 1;
 buf.iov_base = malloc((32) * sizeof(char));
 buf.iov_len = 32;

 memset(&h, 0, sizeof(h));
 memset(&peer, 0, sizeof(peer));
 h.msg_name = &peer;
 h.msg_namelen = sizeof(peer);
 h.msg_iov = (void*) &buf;
 h.msg_iovlen = 1;

 // copied from uv__nonblock_ioctl
 do
 r = ioctl(fd, FIONBIO, &nonblocking);
 while (r == -1 && errno == EINTR);
 if (r) {
 printf("ioctl errno: %d %s\n", r, strerror(errno));
 }

 nread = recvmsg(fd, &h, 0);
 if (nread == -1 || errno != 0) {
 printf("%ld %s\n", nread, strerror(errno));
 } else {
 printf("nread=%ld sa_family=%d msg_namelen=%d\n", nread, ((struct sockaddr *)&peer)->sa_family, h.msg_namelen);
 }
}

int main() {
 int socket_vector[2];

 // use either SOCK_DGRAM or SOCK_STREAM
 if(0 != socketpair(AF_UNIX, SOCK_DGRAM, 0, socket_vector)) {
 perror("cannot create socket pair");
 return 2;
 }

 do_send(socket_vector[0]);
 do_recv(socket_vector[1]);
}

OS	socketpair type	nonblocking	peer.sa_family	msg_namelen
Linux	SOCK_DGRAM	0	AF_UNSPEC	0
Linux	SOCK_DGRAM	1	AF_UNSPEC	0
Linux	SOCK_STREAM	0	AF_UNSPEC	0
Linux	SOCK_STREAM	1	AF_UNSPEC	0
macOS	SOCK_DGRAM	0	AF_UNIX	16
macOS	SOCK_DGRAM	1	AF_UNIX	16
macOS	SOCK_STREAM	0	AF_UNSPEC	0
macOS	SOCK_STREAM	1	AF_UNSPEC	0

And ((struct sockaddr_un *) &peer)->sun_path is always an empty string.

Related code:

• https://github.com/libuv/libuv/blob/0ed9692a80b2f310ad52060a8cb400f2a963d09c/src/unix/udp.c#L285-L301

• uvloop/uvloop/handles/udp.pyx

  Line 335 in 2870219

  pyaddr = __convert_sockaddr_to_pyaddr(addr)

• uvloop/uvloop/dns.pyx

  Lines 64 to 68 in 2870219

  	elif addr.sa_family == uv.AF_UNIX:
  	addr_un = <system.sockaddr_un*>addr
  	return system.MakeUnixSockPyAddr(addr_un)
  	raise RuntimeError("cannot convert sockaddr into Python object")

@ZhuYuJin 's code fails because __convert_sockaddr_to_pyaddr doesn't support AF_UNSPEC. @1st1 shall we patch __convert_sockaddr_to_pyaddr so that it returns an empty string if the family is AF_UNSPEC like this (Solution A)?

diff --git a/uvloop/dns.pyx b/uvloop/dns.pyx
index a252a24..843b5d5 100644
--- a/uvloop/dns.pyx
+++ b/uvloop/dns.pyx
@@ -65,6 +67,9 @@ cdef __convert_sockaddr_to_pyaddr(const system.sockaddr* addr):
 addr_un = <system.sockaddr_un*>addr
 return system.MakeUnixSockPyAddr(addr_un)

+ elif addr.sa_family == uv.AF_UNSPEC:
+ return ""
+
 raise RuntimeError("cannot convert sockaddr into Python object")

Alternatively, we could patch udp.pyx to handle it like this (Solution B):

diff --git a/uvloop/handles/udp.pyx b/uvloop/handles/udp.pyx
index 675407d..5a0f225 100644
--- a/uvloop/handles/udp.pyx
+++ b/uvloop/handles/udp.pyx
@@ -330,6 +330,8 @@ cdef void __uv_udp_on_receive(uv.uv_udp_t* handle,

 if addr is NULL:
 pyaddr = None
+ elif addr.sa_family == uv.AF_UNSPEC:
+ pyaddr = ""
 else:
 try:
 pyaddr = __convert_sockaddr_to_pyaddr(addr)

Or less efficiently but support potential future changes to __convert_sockaddr_to_pyaddr (Solution C):

diff --git a/uvloop/handles/udp.pyx b/uvloop/handles/udp.pyx
index 675407d..16aced6 100644
--- a/uvloop/handles/udp.pyx
+++ b/uvloop/handles/udp.pyx
@@ -334,8 +334,11 @@ cdef void __uv_udp_on_receive(uv.uv_udp_t* handle,
 try:
 pyaddr = __convert_sockaddr_to_pyaddr(addr)
 except BaseException as exc:
- udp._error(exc, False)
- return
+ if addr.sa_family == uv.AF_UNSPEC:
+ pyaddr = ""
+ else:
+ udp._error(exc, False)
+ return

 if nread < 0:
 exc = convert_error(nread)

[ZhuYuJin]

@fantix It should solve the problem.
However, I specify a unix socket pair. Why does it become a AF_UNSPEC?
This code performs well in Mac OS.

[fantix]

Yeah, I think it is from Linux kernel - I tried to track it down but it's a lot of code to read, so we might want to work it around in uvloop first.

[1st1]

C or B would work. How does CPython deal with this btw?

[fantix]

That's a really good point - let me check.

[fantix]

Looks like it'll just return None:

https://github.com/python/cpython/blob/87502ddd710eb1f030b8ff5a60b05becea3f474f/Modules/socketmodule.c#L1335-L1341

static PyObject *
makesockaddr(SOCKET_T sockfd, struct sockaddr *addr, size_t addrlen, int proto)
{
 if (addrlen == 0) {
 /* No address -- may be recvfrom() from known socket */
 Py_RETURN_NONE;
 }

Guess solution D would be to update __convert_sockaddr_to_pyaddr to include this logic.

[1st1]

Guess solution D would be to update __convert_sockaddr_to_pyaddr to include this logic.

Yep!

[fantix]

Bad news: libuv didn't provide the addrlen (the msg_namelen below as a value-result argument) for callbacks (I'll file an issue in libuv):

https://github.com/libuv/libuv/blob/16910d23a34b8b8c313d24824fd85383aa8a8d40/src/unix/udp.c#L180-L200

static void uv__udp_recvmsg(uv_udp_t* handle) {
 struct sockaddr_storage peer;
 struct msghdr h;
 ...
 memset(&h, 0, sizeof(h));
 memset(&peer, 0, sizeof(peer));
 h.msg_name = &peer;
 h.msg_namelen = sizeof(peer);
 ...
 handle->recv_cb(handle, nread, &buf, (const struct sockaddr*) &peer, flags);

And I don't think it is a reliable way to check if the first byte is \0 as an equivalent to checking addrlen == 0 in __convert_sockaddr_to_pyaddr, because we don't know for sure what addr->sa_family is.

In this case, I would go for solution B if no objections.

---

By the way, @ZhuYuJin 's code doesn't work with vanilla asyncio - it uses sendto if peername is not set:

https://github.com/python/cpython/blob/515fce4fc4bb0d2db97b17df275cf90640017f56/Lib/asyncio/selector_events.py#L1043-L1046

 def sendto(self, data, addr=None):
 ...
 if self._extra['peername']:
 self._sock.send(data)
 else:
 self._sock.sendto(data, addr)

But the peername is always empty for a socketpair:

import socket

s1, s2 = socket.socketpair(socket.AF_UNIX, socket.SOCK_DGRAM, 0)
print("peername:", repr(s1.getpeername()))
s1.sendto(b'xxx', None)
s1.sendto(b'xxx', b'')

peername: ''
Traceback (most recent call last):
 File "t304.py", line 5, in <module>
 s1.sendto(b'xxx', None)
TypeError: a bytes-like object is required, not 'NoneType'

Traceback (most recent call last):
 File "t304.py", line 5, in <module>
 s1.sendto(b'xxx', b'')
OSError: [Errno 56] Socket is already connected

Should we consider this a bug in CPython asyncio and fix it there too?

[1st1]

Yeah, let's to B.

Bad news: libuv didn't provide the addrlen (the msg_namelen below as a value-result argument) for callbacks (I'll file an issue in libuv):

I think I also encountered this limitation myself one time, so resolving this at the libuv level would be great. IIRC I also had problems with dealing with abstract unix sockets in libuv, and maybe it was somehow related to this, I don't remember now.