Repository navigation
Network 2.8 missing conversions for Socket family and type. #427
Description
Activity
I don't have motivation to improve
2.8by myself. But if you send a PR, I would love to merge it and @eborden will release a new version.☝️ agreed. 2.x is in end of life. PRs are accepted for bug fixes, but we are focusing on 3.x.
@vdukhovni it is also good to note that stackage nightly is on 3.x, which is a good indicator that support will spread across the ecosystem.
The reason I reported an issue with 2.8, is that I am looking to contribute a PR for
warpthat fixes races in its acceptor loop shutdown logic. And I don't know which version ofnetworkmight be used by someone building an updated version ofwarp. Can I assume it will be a 3.x build?And in fact, though I've not yet opened a similar issue for 3.x, some of the same sort of usability gaps exist in 3.x. The various sum types in Network don't necessarily have exposed to/from int instances, which then makes some get/set option calls less than useful.
@kazu-yamamoto asked about what to do for 4.x. My take is that there are some deep structural issues with the API. More structures need to become opaque, and more getter (and as appropriate setter) interfaces provided. Pattern synonyms need to be used to generalize some types away from a fixed list of hoices.
Warp supports both
networkv2 and v3 currently. But personally, I assumenetworkv3 for Warp.If you want to bring big changes to
network, v4 is perfectly suitable. For instance, we should change fix #426 probably resulting in breaking backward compatibility.@vdukhovni I believe that we are ready to accept this proposal. Would you send a PR?
@vdukhovni If you wan to include this new feature to v3.1.2.0, please send us a PR as soon as possible. You can add this new feature in any new versions.
OK, I'll try to get that done shortly.
- added 6 commits that reference this issue
on May 21, 2020
In Network 2.8, it is difficult to create a new socket of the same shape as an existing socket due to missing interfaces. In C, I can do this via:
But
Network.Socketdoes not export a way to recover the address family of a socket other than case matching the constructor, and no way to map the integer returned from(getSocketOption sock Type)back to aSocketTypethat can be used to create a new socket, andSO_PROTOCOLis missing entirely, so one can simply hope thatdefaultProtocolis good enough.As systems evolve and new values for some of these fields are added, the "sum type" representation of these fields creates constant compatibility issues. These could in some cases be much better handled via
PatternSynonymsthan fixed sum types. Something like the below suitably refined (e.g. Perhaps SocketFamily can be 8 bits on BSD and 16 bits on Linux, with some Haskell type defined accordingly to match the platform'ssa_family_t){-# LANGUAGE CPP #-} {-# LANGUAGE PatternSynonyms #-} {-# LANGUAGE GeneralizedNewtypeDeriving #-} #include <sys/param.h> #include <sys/socket.h> module Network.Socket.Types ( AddressFamily ( AF_INET , AF_INET6 , AF_UNIX ) ) where import Data.Word newtype AddressFamily = AddressFamily { packAddressFamily :: Word16 } deriving (Eq) #ifdef AF_INET pattern AF_INET :: AddressFamily pattern AF_INET = AddressFamily (#const AF_INET) #else pattern AF_INET = AddressFamily 0 #endif #ifdef AF_INET6 pattern AF_INET6 :: AddressFamily pattern AF_INET6 = AddressFamily (#const AF_INET6) #else pattern AF_INET6 = AddressFamily 0 #endif #ifdef AF_UNIX pattern AF_UNIX :: AddressFamily pattern AF_UNIX = AddressFamily (#const AF_UNIX) #else pattern AF_UNIX = AddressFamily 0 #endif instance Show AddressFamily where show AF_INET = "inet" show AF_INET6 = "inet6" show AF_UNIX = "unix" show (AddressFamily n) = "<family" ++ show n ++ ">"and so on for various similar cases.