classification
Title: socket.getservbyname(), socket.getservbyport(), socket.getprotobyname() are not threadsafe
Type: behavior Stage: needs patch
Components: Extension Modules Versions: Python 3.7, Python 3.6, Python 3.5, Python 2.7
process
Status: open Resolution:
Dependencies: Superseder:
Assigned To: Nosy List: dwfreed, njs, pitrou, vstinner
Priority: normal Keywords:

Created on 2017-05-26 09:32 by dwfreed, last changed 2017-05-27 07:09 by njs.

Messages (3)
msg294540 - (view) Author: Doug Freed (dwfreed) * Date: 2017-05-26 09:32
On at least Linux (and probably most other UNIXes, except OS X), the C functions getservbyname(), getservbyport(), and getprotobyname() are not threadsafe.  CPython's wrappers around these functions in the socket module do nothing to cover up this fact.  Simple reproduction script for getservbyname (others similar):

```
import threading
import socket

def getservbyname_loop(service, port):
        while True:
                result = socket.getservbyname(service)
                if result != port:
                        raise RuntimeError("thread-safety broken, got %d, expected %d" % (result, port))

thread1 = threading.Thread(target=getservbyname_loop, args=("ssh", 22))
thread2 = threading.Thread(target=getservbyname_loop, args=("smtp", 25))
thread1.start()
thread2.start()
```

One of the threads will throw the RuntimeError, saying it got the port number the other thread should have gotten.

Naive fix: a lock (eg, just use the netdb_lock already created in the module)

Proper fix: use the libc's reentrant variant if available, and fall back to locking if not (see gethostbyname_ex() implementation for example).

I'd be happy to work on this, but as I don't have access to anything other than Linux and OS X at the moment, it would be helpful if platform maintainers could chime in on what if any reentrant variants of these functions exist on their platforms so we can have a more proper fix.
msg294545 - (view) Author: Antoine Pitrou (pitrou) * (Python committer) Date: 2017-05-26 10:11
The configure.ac script should check for the availability of gethostbyname_r and friends, like it already does for other functions.
msg294564 - (view) Author: Doug Freed (dwfreed) * Date: 2017-05-26 18:16
It already checks for gethostbyname_r, but the comments in socketmodule.c mention that configure seems to get it wrong.  Those comments are probably old, though, so perhaps that can be revisited as well.
History
Date User Action Args
2017-05-27 07:09:19njssetnosy: + njs
2017-05-26 18:16:23dwfreedsetmessages: + msg294564
2017-05-26 10:11:12pitrousetversions: - Python 3.3, Python 3.4
nosy: + vstinner, pitrou

messages: + msg294545

stage: needs patch
2017-05-26 09:32:15dwfreedcreate