This results in unexpected behaviour if the socket has been `dup()`d, as O_NONBLOCK is shared.
Close is always 'blocking' anyway
See https://github.com/wahern/cqueues/issues/13 for an example use case
In this example:
>Client send: MDTM test.txt
>Server response: 213 20120824120909
Because FTP server do not open new channel (2XX response)
and LuaSocket try open new channel we get timeout.
```lua
local ftp = require "socket.ftp"
local ltn12 = require "ltn12"
local url = require("socket.url")
local URL = "ftp://USER:TEST@127.0.0.1";
local CMD = 'MDTM test.txt';
-- get timeout
ftp.get{
url = URL;
command = CMD;
sink = ltn12.sink.table{};
}
-- or we can use ftp.command
ftp.command{
url = URL;
command = URL,
check = function(...)
local status, data = ...
return true
end;
}
```
inet_pton was copying the entire sockaddr_in struct,
rather than just the sin_addr field...
I am a bit unsure about the UDP fix, because it may affect
TCP as well. On UDP sockets, when a sendto fails, the next
receive/receivefrom fails with CONNRESET. I changed
sock_recv/sock_recvfrom in wsocket.c to skip the CONNRESET
from the recv/recvfrom, hoping that if the socket is TCP,
sock_waitfd will get the CONNRESET again. The tests pass,
but this should be tested more thoroughly.
Previous implementation was not making sure the socket
had the same family as the addr returned by getaddrinfo.
So instead of "connection refused", we could get "invalid
argument", which was our fault.
There seems to be a curious difference between MacOS and
Linux and I am not sure if this is documented. When you
break a "connection" on Mac OS, you only eliminate the peer
association, but the local address remains bound. On Linux,
breaking a "connection" eliminates the binding to the local
address. Have you guys ever come accross this?
Another irritating difference is that connect() returns the
error EAFNOSUPPORT on Mac OS. I am going to ignore all
errors when the reason for calling connect() is simply to
break the "connection".
Bug was caught by user moteus.
Code was checking if arguments was nil after using
luaL_Buffer code, which messes with the stack.
Simple to fix, though.
This avoid socket.lua duplicating the iteration over the results
of getaddrinfo(). Some problems with the C implementation not
initializing sockets or the luasocket family have also been fixed,
and error reporting made more robust.
This wrapper takes a domain name or an IP as first argument
and a service name or port as second argument.
Either argument may be nil.
It returns a list of names (always only one in the IP case) and a
service name.
- Added IPv6 support to getsockname
- Simplified getpeername implementation
- Added family to return of getsockname and getpeername
and added modification to the manual to describe
Documented headers.lua
Update copyright date everywhere
Remove RCSID from files
Move version back to 2.1 rather than 2.1.1
Fixed url package to support ipv6 hosts
Changed "domain" to "family" in tcp and udp structures
Implemented getfamily methods
Update Lua and Luasocket version in samples and in documentation
Documented ipv5_v6only default option being set
Documented tcp6 and udp6
Documented dns.getaddrinfo
Documented zero-sized datagram change?
Documented getoption
Conflicts in options.c were just due to independent small functions
being close to each other.
unix.c in mwild was broken, it wasn't using LUASOCKET_API.
serial.c needed luaL_reg renamed, and to use LUASOCKET_API.
makefile didn't respect standard DESTDIR and prefix makefile
variables, and didn't allow LUAV variable to select lua version to build
against.
I've tested the top-level install-both target builds and installs
against both lua5.1 and lua5.2, but not done further testing.
Conflicts:
README
config
gem/ltn012.tex
makefile
src/makefile
src/options.c
src/options.h
src/tcp.c
src/usocket.c
Looks like a historical bug. Its err argument is an error number, but
if it isn't using a custom error message for it, it just calls
strerror() with the errno global, effectively ignoring its argument
and returning a semi-random string.
This dependency was spuriously added, maybe for debug reasons,
as confirmed to me by Diego Nehab by mail.
Some systems based in Lua (e.g. Ginga) prohibit the use of
io module for security reasons, so this dependency makes
mime unusable; even worse this makes other modules, based
on mime, unusable too (e.g. html).
Connect timeouts are implemented by waiting on the new socket
descriptor. When select() is used for this, it imposes an arbitrary
limit on the number of connections that can be made, usually 1024-3.
Using poll() removes this limit on the number of simultaneous TCP
connections can be made using luasocket. The previous default
implementation using select() is available by defining SOCKET_SELECT.
Note that using socket.select() always uses select(), so it isn't
possible to wait on an arbitrary number of connections at once.
A zero-length send is invalid with TCP, but well defined with UDP.
udp:send"" was returning (nil,"refused"), indicating that it failed when
the packet was actually sent. The test script reproduces the bug, and
includes a tcpdump of the zero length packet being sent.
