You should take a look at the shell injection document before this one.
A lot of the time, our codebase uses shell=True because it’s convenient. The shell provides the ability to pipe things around without buffering them in memory, and allows a malicious user to chain additional commands after a legitimate command is run.
Here is a simple function that uses curl to grab a page from a website, and pipe it directly to the wordcount program to tell us how many lines there are in the HTML source code.
def count_lines(website): return subprocess.check_output('curl %s | wc -l' % website, shell=True) #>>> count_lines('www.google.com') #'7\n'
(That output is correct, by the way - the google html source does have 7 lines.)
The function is insecure because it uses shell=True, which allows shell injection. A user to who instructs your code to fetch the website ; rm -rf / can do terrible things to what used to be your machine.
If we convert the function to use shell=False, it doesn’t work.
def count_lines(website): args = ['curl', website, '|', 'wc', '-l'] return subprocess.check_output(args, shell=False) # >>> count_lines('www.google.com') # curl: (6) Could not resolve host: | # curl: (6) Could not resolve host: wc # Traceback (most recent call last): # File "<stdin>", line 3, in count_lines # File "/usr/lib/python2.7/subprocess.py", line 573, in check_output # raise CalledProcessError(retcode, cmd, output=output) # subprocess.CalledProcessError: Command # '['curl', 'www.google.com', '|', 'wc', '-l']' returned non-zero exit status 6
The pipe doesn’t mean anything special when shell=False, and so curl tries to download the website called ‘|’. This does not fix the issue, rather it causes it to be more broken than before.
If we can’t rely on pipes if we have shell=False, how should we do this?
def count_lines(website): args = ['curl', website] args2 = ['wc', '-l'] process_curl = subprocess.Popen(args, stdout=subprocess.PIPE, shell=False) process_wc = subprocess.Popen(args2, stdin=process_curl.stdout, stdout=subprocess.PIPE, shell=False) # Allow process_curl to receive a SIGPIPE if process_wc exits. process_curl.stdout.close() return process_wc.communicate() # >>> count_lines('www.google.com') # '7\n'
Rather than calling a single shell process that runs each of our programs, we run them separately and connect stdout from curl to stdin for wc. We specify stdout=subprocess.PIPE, which tells subprocess to send that output to the respective file handler.
Treat pipes like file descriptors (you can actually use FDs if you want) they may block on reading and writing if nothing is connected to the other end. That’s why we use communicate(), which reads until EOF on the output and then waits for the process to terminate. You should generally avoid reading and writing to pipes directly unless you really know what you’re doing - it’s easy to work yourself into a situation that can deadlock.
Note that communicate() buffers the result in memory - if that’s not what you want, use a file descriptor for stdout to pipe that output into a file.