How does curl protect a password from appearing in ps output?

Question

I noticed some time ago that usernames and passwords given to curl as command line arguments don't appear in ps output (although of course they may appear in your bash history).

They likewise don't appear in /proc/PID/cmdline.

(The length of the combined username/password argument can be derived, though.)

Demonstration below:

[root@localhost ~]# nc -l 80 &
[1] 3342
[root@localhost ~]# curl -u iamsam:samiam localhost &
[2] 3343
[root@localhost ~]# GET / HTTP/1.1
Authorization: Basic aWFtc2FtOnNhbWlhbQ==
User-Agent: curl/7.19.7 (x86_64-redhat-linux-gnu) libcurl/7.19.7 NSS/3.15.3 zlib/1.2.3 libidn/1.18 libssh2/1.4.2
Host: localhost
Accept: */*



[1]+  Stopped                 nc -l 80
[root@localhost ~]# jobs
[1]+  Stopped                 nc -l 80
[2]-  Running                 curl -u iamsam:samiam localhost &
[root@localhost ~]# ps -ef | grep curl
root      3343  3258  0 22:37 pts/1    00:00:00 curl -u               localhost
root      3347  3258  0 22:38 pts/1    00:00:00 grep curl
[root@localhost ~]# od -xa /proc/3343/cmdline 
0000000    7563    6c72    2d00    0075    2020    2020    2020    2020
          c   u   r   l nul   -   u nul  sp  sp  sp  sp  sp  sp  sp  sp
0000020    2020    2020    0020    6f6c    6163    686c    736f    0074
         sp  sp  sp  sp  sp nul   l   o   c   a   l   h   o   s   t nul
0000040
[root@localhost ~]# 

How is this effect achieved? Is it somewhere in the source code of curl? (I assume it is a curl feature, not a ps feature? Or is it a kernel feature of some sort?)

---

Also: can this be achieved from outside the source code of a binary executable? E.g. by using shell commands, probably combined with root permissions?

In other words could I somehow mask an argument from appearing in /proc or in ps output (same thing, I think) that I passed to some arbitrary shell command? (I would guess the answer to this is "no" but it seems worth including this extra half-a-question.)

Answer 1

When the kernel executes a process, it copies the command line arguments to read-write memory belonging to the process (on the stack, at least on Linux). The process can write to that memory like any other memory. When ps displays the argument, it reads back whatever is stored at that particular address in the process's memory. Most programs keep the original arguments, but it's possible to change them. The POSIX description of ps states that

It is unspecified whether the string represented is a version of the argument list as it was passed to the command when it started, or is a version of the arguments as they may have been modified by the application. Applications cannot depend on being able to modify their argument list and having that modification be reflected in the output of ps.

The reason this is mentioned is that most unix variants do reflect the change, but POSIX implementations on other types of operating systems may not.

This feature is of limited use because the process can't make arbitrary changes. At the very least, the total length of the arguments cannot be increased, because the program can't change the location where ps will fetch the arguments and can't extend the area beyond its original size. The length can effectively be decreased by putting null bytes at the end, because arguments are C-style null-terminated strings (this is indistinguishable from having a bunch of empty arguments at the end).

If you really want to dig, you can look at the source of an open-source implementation. On Linux, the source of ps isn't interesting, all you'll see there is that it reads the command line arguments from the proc filesystem, in /proc/PID/cmdline. The code that generates the content of this file is in the kernel, in proc_pid_cmdline_read in fs/proc/base.c. The part of the process's memory (accessed with access_remote_vm) goes from the address mm->arg_start to mm->arg_end; these addresses are recorded in the kernel when the process starts and can't be changed afterwards.

Some daemons use this ability to reflect their status, e.g. they change their argv[1] to a string like starting or available or exiting. Many unix variants have a setproctitle function to do this. Some programs use this ability to hide confidential data. Note that this is of limited use since the command line arguments are visible while the process starts.

Most high-level languages copy the arguments to string objects and don't give a way to modify the original storage. Here's a C program that demonstrates this ability by changing argv elements directly.

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
int main(int argc, char *argv[])
{
    int i;
    system("ps -p $PPID -o args=");
    for (i = 0; i < argc; i++)
    {
        memset(argv[i], '0' + (i % 10), strlen(argv[i]));
    }
    system("ps -p $PPID -o args=");
    return 0;
}

Sample output:

./a.out hello world
0000000 11111 22222

You can see argv modification in the curl source code. Curl defines a function cleanarg in src/tool_paramhlp.c which is used to change an argument to all spaces using memset. In src/tool_getparam.c this function is used a few times, e.g. by redacting the user password. Since the function is called from the parameter parsing, it happens early in a curl invocation, but dumping the command line before this happens will still show any passwords.

Since the arguments are stored in the process's own memory, they cannot be changed from the outside except by using a debugger.

Answer 2

The other answers answer the question well in a general manner. To specifically answer "How is this effect achieved? Is it somewhere in the source code of curl?":

In the argument parsing section of the curl source code, the -u option is handled as follows:

    case 'u':
      /* user:password  */
      GetStr(&config->userpwd, nextarg);
      cleanarg(nextarg);
      break;

And the cleanarg() function is defined as follows:

void cleanarg(char *str)
{
#ifdef HAVE_WRITABLE_ARGV
  /* now that GetStr has copied the contents of nextarg, wipe the next
   * argument out so that the username:password isn't displayed in the
   * system process list */
  if(str) {
    size_t len = strlen(str);
    memset(str, ' ', len);
  }
#else
  (void)str;
#endif
}

So we can explicitly see that the username:password argument in argv is overwritten with spaces, as described by the other answers.

Answer 3

A process can not only read its parameters but write them, too.

I am not a developer so I am not familiar with this stuff but it may be possible from the outside with an approach similar to the changing of environment parameters:

https://stackoverflow.com/questions/205064/is-there-a-way-to-change-another-processs-environment-variables