Comment 1:

Owner changed to @bradfitz.

Status changed to Accepted.

Comment 2:

I think I'd rather catch the exact error that Windows generates.
Could you change that line to log.Printf("error %T %#v", err, err) and let me know what
the windows error type is?

Comment 3 by hectorchu:

I don't have the source checked out at the moment, but I believe I can make a pretty
good guess of the type.
package main
import (
    "log"
    "net"
    "os"
)
func main() {
    err := &net.OpError{"WSARecv", "tcp", &net.TCPAddr{net.ParseIP("192.168.137.130"), 55092}, os.Errno(64)}
    log.Printf("err=%v", err)
    log.Printf("error %T %#v", err, err)
}
Outputs:
2011/07/11 23:20:34 err=WSARecv tcp 192.168.137.130:55092: The specified network name is
no longer available.
2011/07/11 23:20:34 error *net.OpError &net.OpError{Op:"WSARecv", Net:"tcp",
Addr:(*net.TCPAddr)(0x10b727f0), Error:64}

Comment 4:

errno=64 would be
#define ERROR_NETNAME_DELETED 64L
you could add the const to syscall/ztypes_windows.go if it helps you any.

Comment 5 by hectorchu:

I think the problem is broader than Windows.  I'm getting a similar error under Linux:
connection reset by peer (probably ECONNRESET).  Here's the website I'm connecting to,
in a Go program:
package main
import (
    "bytes"
    "http"
    "io"
    "log"
)
func main() {
    r, err := http.Get("https://www.mbtrading.com/secure/getquoteserverxml.asp")
    if err != nil {
        return
    }
    var buf bytes.Buffer
    io.Copy(&buf, r.Body)
    r.Body.Close()
    log.Printf("%s\n", buf.Bytes())
    select {}
}
The disconnect happens after 2 minutes.

Comment 6:

I agree.
This particular server closes connection in such way that:
err == os.EOF || err == os.EINVAL
in http/transport.go wouldn't catch that case.
On Windows I receive ERROR_NETNAME_DELETED while reading from connection.
On Linux - ECONNRESET.
We need to handle these errors.

Comment 7:

Alex, I can fix this if there's an easy way to reproduce.  Is it just the server closing
the connection?  You want to write (at least part of) a unit test that I can finish +
fix the transport?
I guess I have a Windows machine now, so I can experiment.

Comment 8:

Brad,
I don't know how to reproduce the problem reliably and on all different platforms. But
here I can demonstrate it on Windows. Build this program:
package main
import (
       "fmt"
       "io"
       "net"
       "log"
       "os"
)
func serveClient(fd io.ReadWriteCloser) {
       defer fd.Close()
       var buf [1024]byte
       for {
               // both parties are blocked reading
               n, err := fd.Read(buf[0:])
               if err != nil || n == 0 {
                       fmt.Printf("disconneted: %q\n", err)
                       break
               }
       }
}
func server() {
       const addr = ":0"
       fmt.Printf("running server ...\n")
       l, err := net.Listen("tcp", addr)
       if err != nil {
               log.Fatalf("Listen(%q) failed: %v", addr, err)
       }
       defer l.Close()
       fmt.Printf("listening on %s ...\n", l.Addr().String())
       for {
               fd, err := l.Accept()
               if err != nil {
                       break
               }
               go serveClient(fd)
       }
}
func client(addr string) {
       fmt.Printf("running client ...\n")
       fd, err := net.Dial("tcp", addr)
       if err != nil {
               log.Fatalf("Dial(%q) failed: %v", addr, err)
       }
       defer fd.Close()
       // both parties are blocked reading
       var b [1024]byte
       n, err := fd.Read(b[:])
       if err != nil {
               log.Fatalf("fd.Read() = %d, %v", n, err)
       }
}
func main() {
       if len(os.Args) == 1 {
               server()
       } else {
               client(os.Args[1])
       }
}
and then run it on Windows:
C:\>a.exe
running server ...
listening on 0.0.0.0:1713 ...
then run the same program, but in a new window:
C:\>a.exe localhost:1713
running client ...
Go back to original (server) window and hit Ctrl+C to kill the program. You should see
client program exit with:
2011/07/22 19:33:16 fd.Read() = 0, WSARecv tcp 127.0.0.1:1713: The specified network
name is no longer available.
If you run client program on Linux, the message is:
2011/07/22 19:52:22 fd.Read() = 0, read tcp 10.0.0.13:1713: connection reset by peer
Alex

Comment 9 by hectorchu:

> Is it just the server closing the connection?
Yep, the server is doing a dirty disconnect.  Should be pretty straightforward to write
a Go http server that does something similar.

Comment 10:

Status changed to HelpWanted.

Comment 12 by hectorchu:

The hard close is likely to be caused by the peer setting SO_LINGER l_onoff to 1 (true)
and l_linger to 0 (zero time) (see
http://groups.google.com/group/comp.unix.programmer/msg/a21de3c46091ed9c).  I had
trouble writing a test case that reproduced this on Windows, but at least the post
indicates that ECONNRESET isn't to be completely unexpected.  Anyway this code will
probably change as we're not supposed to use syscall-type errors in general use.

Comment 13:

I thought I had a fix for this at one point.  Maybe it's on my Windows machine at home,
stranded without a working codereview.py.

Comment 14:

This issue was closed by revision c4b9c8b.

Status changed to Fixed.