32-bit to 64-bit Conversion Woes with strtoul()

Proper error checking when using the strtol() family of functions is notoriously difficult.  See this stackoverflow thread explaining it.

Focusing only on strtoul() here, the correct error checking in most cases is:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

unsigned long mystrtoul(char *str, bool *success) { unsigned long val = 0; char *eptr = NULL; errno = 0; *success = true; val = strtoul(str, &eptr, 0); if (eptr == str || *eptr != '\0' || (val == ULONG_MAX && errno == ERANGE) || (val == 0 && errno == EINVAL)) *success = false; return val; }

That is:

• IF your end pointer points to the beginning of your string, you've got a problem
• IF your end pointer doesn't point to the end of your string, you've got a problem
• IF your output is at the max end of the range (a potentially valid value) AND you get errno, you've got a problem
• IF your output is 0 (a potentially valid value) AND you get errno, you've got a problem
• ELSE congratulations, you've converted a string to an integer

Understandably, a lot of people get this error checking wrong when using this function. I've reviewed source code I work on and found mistakes.  But even if you got all of this right, you may still have gotten it wrong.

If you originally wrote this code for a 32-bit system, with the intention of converting and storing a 32-bit number, then what you wrote is correct.  If you then recompile this correct code on a 64-bit system, it becomes incorrect.

That's right with no modifications (and no compiler warnings), your carefully reviewed integer conversion function goes from right to (dangerously) wrong.

For example give as input "4294967296" and assign to an unsigned int:

1 2 3 4 5 6 7 8 9 10 11

int main(void) { bool success = false; unsigned int val = 0; val = mystrtoul("4294967296", &success); printf("%s %u\n", success?"success":"failure", val); return 0; }

Result:

• On 32-bit Linux:  failure 4294967296 (ERANGE)
• On 64-bit Linux:  success 0

The 32-bit compilation catches the integer overflow and errors out.  The 64-bit version does not and rolls back around to 0.  So your output is wrong and the usual possible errors and security holes with integer overflows apply:

• malloc(4294967296) allocate 0 bytes and a buffer overflow is created
• uid_t 4294967296 become 0 and a user is root

And strings passed in to strtoul() are very often user controlled input.  That's why you need to convert them from an on-the-wire, CLI, or on-disk string format to an in-memory integer.

This isn't strictly a strtoul() problem.  This is a LP64 data model problem.  Any assumption about the 32-bit size of a long needs to be checked when converting from 32-bit to 64-bit runtime.  This is just a particularly tricky example of the trade offs made by LP64.

Using shquote() to Combat Command Injection

I've been incredibly impressed with FreeBSD's shquote(3) function as a singular way to prevent command injection in a server process.  I'm currently working on a RESTful API (aren't we all), backed in part by a server daemon written in C.  The largest threat to any networked API is user input.  A server developer must assume that all input is malicious until proven valid.  There are many types of injection attacks against web services; I've been focused on Unix command injection today.

Let's say you're building a service which takes a directory path as input and creates this directory path on the server's file system.  In C we can just call:

 void mkdir_recursive(char* path) {
      char* cmd;
      asprintf(&cmd, "mkdir -p %s", path);
      system(cmd);
 }

Now wait!  Using system() is bad!  It's prone to command injection attacks and there's usually always a programmable API to do the same thing in a safer, cleaner way.

I agree, but I am interested in exploring command injection and mkdir -p is so convenient for recursive directory creation.  There's no easy libc alternative and writing it myself would take 10x as much code.  Yes, I'm complaining about 3 lines vs 30, but Larry Wall says laziness is one of a programmers greatest virtues.

So normally to protect this simple code from command injection we'd validate the input for shell meta-characters.  We must check for ; & ( ) ` | > < $.  These could allow an attacker to append an arbitrary command onto our simple mkdir, read or write files, and access environment variables.  For instance we could receive a path which would make the above command look like this:

mkdir -p a/new/dir ; cat /etc/passwd > /usr/local/apache/htdocs/passwd

Which unfiltered would cause the server's local user list to now be published in a publicly accessible directory.  To further cause issues, in my particular use case, all of the above characters are valid for Unix directory paths.  So simply removing them from the user's input would be incorrect behavior.

In comes shquote().  It protects against all of these command injections and allows these characters to be used in path names, because it puts everything in strong quotes.  With a refactor of our above code snippet:

 void mkdir_recursive(char* path) {
      char *cmd;
      int len = shquote(path, NULL, 0) + 1;
      char *safe_path = malloc(len);
      shquote(path, &safe_path, len);
      asprintf(&cmd, "mkdir -p %s", path);
      system(cmd);
 }

The execute command now looks like this:

mkdir -p 'a/new/dir ; cat /etc/passwd > /usr/local/apache/htdocs/passwd'

Which will safely create 10 new directories, instead of calling 2 commands.

# while [ "`ls`" ]; do ls -F && cd "`ls`"; done
a/
new/
dir ; cat /
etc/
passwd > /
usr/
local/
apache/
htdocs/
passwd/

Surely we can still trick shquote() with embedded \ or ' characters right?  Let's try adding intermediate quotes to break up the two commands:

a/new/dir' ; 'cat /etc/passwd > /usr/local/apache/htdocs/passwd

becomes

mkdir -p 'a/new/dir'\'' ; '\''cat /etc/passwd > /usr/local/apache/htdocs/passwd'

# while [ "`ls`" ]; do ls -F && cd "`ls`"; done
a/
new/
dir' ; 'cat /
etc/
...

That didn't give us an exploit.  So let's try escaping those intermediate quotes:

a/new/dir\' ; \'cat /etc/passwd > /usr/local/apache/htdocs/passwd

becomes

mkdir -p 'a/new/dir\'\'' ; \'\''cat /etc/passwd > /usr/local/apache/htdocs/passwd'

# while [ "`ls`" ]; do ls -F && cd "`ls`"; done
a/
new/
dir\' ; \'cat /
etc/
...

My simple function handles them both correctly with successive strong quote wrappings.  It still isn't protected from other attacks like path traversal trickery using embedded .. but shquote() goes a long way in santizing input.  I couldn't think of a way to break it.  Please tell me if you can.