python subprocess deadlock in demonized server

问题

I'm trying to set up a remote backup server for dar, along these lines. I'd really like to do all the piping with python if possible, but I've asked a separate question about that.

Using netcat in subprocess.Popen(cmd, shell=True), I succeeded in making a differential backup, as in the examples on the dar site. The only two problems are:

I don't know how to assign port numbers dynamically this way
If I execute the server in the background, it fails. Why?

Update: This doesn't seem to be related to netcat; it hangs even without netcat in the mix.

Here's my code:

from socket import socket, AF_INET, SOCK_STREAM
import os, sys
import SocketServer
import subprocess

class DarHandler(SocketServer.BaseRequestHandler):
    def handle(self):
        print('entering handler')
        data = self.request.recv(1024).strip()
        print('got: ' + data)
        if data == 'xform':
            cmd1 = 'nc -dl 41201 | dar_slave archives/remotehost | nc -l 41202'
            print(cmd1)
            cmd2 = 'nc -dl 41200 | dar_xform -s 10k - archives/diffbackup'
            print(cmd2)
            proc1 = subprocess.Popen(cmd1, shell=True)
            proc2 = subprocess.Popen(cmd2, shell=True)
            print('sending port number')
            self.request.send('41200')
            print('waiting')
            result = str(proc1.wait())
            print('nc-dar_slave-nc returned ' + result)
            result = str(proc2.wait())
            print('nc-dar_xform returned ' + result)
        else:
            result = 'bad request'
        self.request.send(result)
        print('send result, exiting handler')

myaddress = ('localhost', 18010)
def server():
    server = SocketServer.TCPServer(myaddress, DarHandler)
    print('listening')
    server.serve_forever()

def client():
    sock = socket(AF_INET, SOCK_STREAM)
    print('connecting')
    sock.connect(('localhost', 18010))
    print('connected, sending request')
    sock.send('xform')
    print('waiting for response')
    port = sock.recv(1024)
    print('got: ' + port)
    try:
        os.unlink('toslave')
    except:
        pass
    os.mkfifo('toslave')
    cmd1 = 'nc -w3 localhost 41201 < toslave'
    cmd2 = 'nc -w3 localhost 41202 | dar -B config/test.dcf -A - -o toslave -c - | nc -w3 localhost ' + port
    print(cmd2)
    proc1 = subprocess.Popen(cmd1, shell=True)
    proc2 = subprocess.Popen(cmd2, shell=True)
    print('waiting')
    result2 = proc2.wait()
    result1 = proc1.wait()
    print('nc<fifo returned: ' + str(result1))
    print('nc-dar-nc returned: ' + str(result2))
    result = sock.recv(1024)
    print('received: ' + result)
    sock.close()
    print('socket closed, exiting')

if __name__ == "__main__":
    if sys.argv[1].startswith('serv'):
        server()
    else:
        client()

Here's what happens on the server:

$ python clientserver.py serve &
[1] 4651
$ listening
entering handler
got: xform
nc -dl 41201 | dar_slave archives/remotehost | nc -l 41202
nc -dl 41200 | dar_xform -s 10k - archives/diffbackup
sending port number
waiting

[1]+  Stopped                 python clientserver.py serve

Here's what happens on the client:

$ python clientserver.py client
connecting
connected, sending request
waiting for response
got: 41200
nc -w3 localhost 41202 | dar -B config/test.dcf -A - -o toslave -c - | nc -w3 localhost 41200
waiting
FATAL error, aborting operation
Corrupted data read on pipe
nc<fifo returned: 1
nc-dar-nc returned: 1

The client also hangs, and I have to kill it with a keyboard interrupt.

回答1:

I'd cut my losses and start over. This solution attempt is very complicated and kludgy. There are many ready-made solutions in the area:

10 outstanding Linux backup utilities

Fwbackups sounds good if you want to take the easy route, rsync+ssh for the hard core.

回答2:

Use Popen.communicate() instead of Popen.wait().

The python documentation for wait() states:

Warning: This will deadlock if the child process generates enough output to a stdout or stderr pipe such that it blocks waiting for the OS pipe buffer to accept more data. Use communicate() to avoid that.
Dar and its related executables should get a -Q if they aren't running interactively.
When syncronizing multiple processes, make sure to call communicate() on the 'weakest link' first: dar_slave before dar_xform and dar before cat. This was done correctly in the question, but it's worth noting.
Clean up shared resources. The client process is holding open a socket from which dar_xform is still reading. Attempting to send/recv data on the initial socket after dar and friends are finished without closing that socket will therefore cause a deadlock.

Here is a working example which doesn't use shell=True or netcat. An advantage of this is I can have the secondary ports assigned dynamically and therefore could conceivably serve multiple backup clients simultaneously.

来源：https://stackoverflow.com/questions/7922477/python-subprocess-deadlock-in-demonized-server

标签

python

deadlock

subprocess