Golang Network File Server
NFSv3 protocol implementation in pure Golang.
Current Status:
- Minimally tested
- Mounts, read-only and read-write support
Usage
The most interesting demo is currently in example/osview
.
Start the server
go run ./example/osview .
.
The local folder at .
will be the initial view in the mount. mutations to metadata or contents
will be stored purely in memory and not written back to the OS. When run, this
demo will print the port it is listening on.
The mount can be accessed using a command similar to
mount -o port=<n>,mountport=<n> -t nfs localhost:/mount <mountpoint>
(For Mac users)
or
mount -o port=<n>,mountport=<n>,nfsvers=3,noacl,tcp -t nfs localhost:/mount <mountpoint>
(For Linux users)
API
The NFS server runs on a net.Listener
to export a file system to NFS clients.
Usage is structured similarly to many other golang network servers.
package main
import (
"fmt"
"log"
"net"
"github.com/go-git/go-billy/v5/memfs"
nfs "github.com/willscott/go-nfs"
nfshelper "github.com/willscott/go-nfs/helpers"
)
func main() {
listener, err := net.Listen("tcp", ":0")
panicOnErr(err, "starting TCP listener")
fmt.Printf("Server running at %s\n", listener.Addr())
mem := memfs.New()
f, err := mem.Create("hello.txt")
panicOnErr(err, "creating file")
_, err = f.Write([]byte("hello world"))
panicOnErr(err, "writing data")
f.Close()
handler := nfshelper.NewNullAuthHandler(mem)
cacheHelper := nfshelper.NewCachingHandler(handler, 1)
panicOnErr(nfs.Serve(listener, cacheHelper), "serving nfs")
}
func panicOnErr(err error, desc ...interface{}) {
if err == nil {
return
}
log.Println(desc...)
log.Panicln(err)
}
Notes
-
Ports are typically determined through portmap. The need for running portmap (which is the only part that needs a privileged listening port) can be avoided through specific mount options. e.g.
mount -o port=n,mountport=n -t nfs host:/mount /localmount
-
This server currently uses billy to provide a file system abstraction layer. There are some edges of the NFS protocol which do not translate to this abstraction.
- NFS expects access to an
inode
or equivalent unique identifier to reference files in a file system. These are considered opaque identifiers here, which means they will not work as expected in cases of hard linking. - The billy abstraction layer does not extend to exposing
uid
andgid
ownership of files. If ownership is important to your file system, you will need to ensure that theos.FileInfo
meets additional constraints. In particular, theSys()
escape hatch is queried by this library, and if your file system populates asyscall.Stat_t
concrete struct, the ownership specified in that object will be used.
- NFS expects access to an
-
Relevant RFCS: 5531 - RPC protocol, 1813 - NFSv3, 1094 - NFS