A patch for HEAD

Gwynne Raskind wrote:

I'm proposing the inclusion of this patch in HEAD (which I'll port it to
if I get a thumbs-up here), and possibly 5.3.2. Criticism and angry
flames welcome. Constructive critcism and good-natured comments will be
ignored ;) (just kidding... or am I?).

I'm fine with this in HEAD.

-Andrei

The idea is great. In fact this was in my todo list for php 5.3..
Please give me a few more days to review the patch.

Nuno

P.S.: you can add on more point to your list: you get to know the PID of the
exec'ed process instead of the PID of the shell.

----- Original Message -----

I've just finished making this patch for my own use (diffed against 5.3
CVS):

http://darkrainfall.org/php-5.3-shellbypass.patch

In short, what it does is make proc_open()'s shell_bypass option
available to UNIX systems. This is accomplished by allowing the "command"
parameter to proc_open() to be an array of arguments to pass to
execve. I've included a few tests to check the functionality.

(A few more tests could be devised to, for example, check that the
correct warning is issued if you pass an array without bypass_shell set,
or a string with it set, etc.)

The exact behavior of the argument array is:

1. The array must contain at least one element, at index 0.
2. The element at index 0 is always the exact command path passed to
   execve (after being filtered through any safe_mode restrictions, as
   with the normal behavior of proc_open()).
3. Any other elements form the argv array passed to execve. By
   convention the first of these arguments (argv[0] in the child process) is
   the same as the command path, however my patch does NOT enforce or assume
   this; it simply calls execv[e]($argument_array[0],
   array_slice($argument_array, 1)).

This patch currently provides the only useful way to fork a process
without running a shell (pcntl_fork() + pcntl_exec() are useless since
there's no pcntl_dup2() to control the descriptors of the child).

Why would you want to avoid the shell?

• Efficiency. The shell is an extra, often unnecessary process, which
  must parse the commandline given to it into individual arguments
  according to all its various rules. Not to mention the overhead of
  setting up another entire process just to run a third process.

• Resource control. The shell is an extra process. If you don't need it,
  and your system is tight on process space, best to avoid it.

• Sanity. Correctly quoting arguments to a shell command ranges from
  mildly annoying (escapeshellarg() in simple cases) to nightmarish (manual
  parsing of a string in some edge cases). Passing arguments directly
  completely bypasses this, quite possibly saving you quite a bit of string
  parsing time if you were doing something like "$shell_args = implode(' ',
  array_map('escapeshellarg', $raw_args));".

• Oddly enough, security. Since there's no shell, it's more difficult to
  subvert the child process to do other things than the coder intended
  (unless of course, said coder executes a shell this way).

This patch does nothing on Windows, since the option was already
implemented there. It also does nothing on Netware, since from what I
could see in the code, Netware doesn't have a shell in the first place.

I'm proposing the inclusion of this patch in HEAD (which I'll port it to
if I get a thumbs-up here), and possibly 5.3.2. Criticism and angry
flames welcome. Constructive critcism and good-natured comments will be
ignored ;) (just kidding... or am I?).

-- Gwynne

Hi,

So the patch looks generally good. Here are some minor comments about it:

• I believe _php_array_to_argv() doesn't need TSRMLS_DC. If that's the
  case, please remove it.
• in _php_array_to_argv() you modify the input array destructively (when
  calling convert_to_string_ex). You should not modify the input.
• in _php_array_to_argv(), HASH_OF will never return NULL (because we
  already know that it's an array).
• apparently the check 'if (cnt < 1) {' will never be true, as the array
  will always have one element. please verify that this special case does what
  you want (maybe change to 'cnt <= 1')
• argvs with length==0 are perfectly valid, and so you shouldn't skip them.
• the api is a little inconsistent, as you use the idx 0 to retrieve the
  command name, but then you use the array order to retrieve the rest of the
  elements (and thus if the idx 0 doesn't appear in the beginning your code
  will fail). I would just stick with the array order. e.g.
  array(1=>'/usr/bin/echo', 0=>'foo') or similar should print 'foo'.
• in exit_fail you can remove the check for bypass_shell, as
  _php_free_argv() will check the argument against NULL.
• the line 'proc->argv = (bypass_shell ? child_argv : NULL);' can be
  simplified to 'proc->argv = child_argv;' since child_argv is already
  initialized to NULL

I think that's it :) It's only minor things, I guess.
As soon as you fix these things, please go ahead and commit the patch, or
mail it back to the ML in case you need me to do it.

Nuno

----- Original Message -----
From: "Gwynne Raskind" gwynne@darkrainfall.org
To: "PHP internals" internals@lists.php.net
Sent: Friday, June 26, 2009 10:20 PM
Subject: [PHP-DEV] A patch for HEAD

I've just finished making this patch for my own use (diffed against 5.3
CVS):

http://darkrainfall.org/php-5.3-shellbypass.patch

In short, what it does is make proc_open()'s shell_bypass option
available to UNIX systems. This is accomplished by allowing the "command"
parameter to proc_open() to be an array of arguments to pass to
execve. I've included a few tests to check the functionality.

(A few more tests could be devised to, for example, check that the
correct warning is issued if you pass an array without bypass_shell set,
or a string with it set, etc.)

The exact behavior of the argument array is:

1. The array must contain at least one element, at index 0.
2. The element at index 0 is always the exact command path passed to
   execve (after being filtered through any safe_mode restrictions, as
   with the normal behavior of proc_open()).
3. Any other elements form the argv array passed to execve. By
   convention the first of these arguments (argv[0] in the child process) is
   the same as the command path, however my patch does NOT enforce or assume
   this; it simply calls execv[e]($argument_array[0],
   array_slice($argument_array, 1)).

This patch currently provides the only useful way to fork a process
without running a shell (pcntl_fork() + pcntl_exec() are useless since
there's no pcntl_dup2() to control the descriptors of the child).

Why would you want to avoid the shell?

• Efficiency. The shell is an extra, often unnecessary process, which
  must parse the commandline given to it into individual arguments
  according to all its various rules. Not to mention the overhead of
  setting up another entire process just to run a third process.

• Resource control. The shell is an extra process. If you don't need it,
  and your system is tight on process space, best to avoid it.

• Sanity. Correctly quoting arguments to a shell command ranges from
  mildly annoying (escapeshellarg() in simple cases) to nightmarish (manual
  parsing of a string in some edge cases). Passing arguments directly
  completely bypasses this, quite possibly saving you quite a bit of string
  parsing time if you were doing something like "$shell_args = implode(' ',
  array_map('escapeshellarg', $raw_args));".

• Oddly enough, security. Since there's no shell, it's more difficult to
  subvert the child process to do other things than the coder intended
  (unless of course, said coder executes a shell this way).

This patch does nothing on Windows, since the option was already
implemented there. It also does nothing on Netware, since from what I
could see in the code, Netware doesn't have a shell in the first place.

I'm proposing the inclusion of this patch in HEAD (which I'll port it to
if I get a thumbs-up here), and possibly 5.3.2. Criticism and angry
flames welcome. Constructive critcism and good-natured comments will be
ignored ;) (just kidding... or am I?).

-- Gwynne