Is the data in siginfo trustworthy?

Question

I\'ve found that on Linux, by making my own call to the rt_sigqueue syscall, I can put whatever I like in the si_uid and si_pid fields and

Answer 1

That section of the POSIX page you quote also lists what si-code means, and here's the meaning:

SI_QUEUE
    The signal was sent by the sigqueue() function.

That section goes on to say:

If the signal was not generated by one of the functions or events listed above, si_code shall be set either to one of the signal-specific values described in XBD , or to an implementation-defined value that is not equal to any of the values defined above.

Nothing is violated if only the sigqueue() function uses SI_QUEUE. Your scenario involves code other than the sigqueue() function using SI_QUEUE The question is whether POSIX envisions an operating system enforcing that only a specified library function (as opposed to some function which is not a POSIX-defined library function) be permitted to make a system call with certain characteristics. I believe the answer is "no".

EDIT as of 2011-03-26, 14:00 PST:

This edit is in response to R..'s comment from eight hours ago, since the page wouldn't let me leave an adequately voluminous comment:

I think you're basically right. But either a system is POSIX compliant or it is not. If a non-library function does a syscall which results in a non-compliant combination of uid, pid, and 'si_code', then the second statement I quoted makes it clear that the call itself is not compliant. One can interpret this in two ways. One ways is: "If a user breaks this rule, then he makes the system non-compliant." But you're right, I think that's silly. What good is a system when any nonprivileged user can make it noncompliant? The fix, as I see it, is somehow to have the system know that it's not the library 'sigqueue()' making the system call, then the kernel itself should set 'si_code' to something other than 'SI_QUEUE', and leave the uid and pid as you set them. In my opinion, you should raise this with the kernel folks. They may have difficulty, however; I don't know of any secure way for them to detect whether a syscall is made by a particular library function, seeing as how the library functions. almost by definition, are merely convenience wrappers around the syscalls. And that may be the position they take, which I know will be a disappointment.

(voluminous) EDIT as of 2011-03-26, 18:00 PST:

Again because of limitations on comment length.

This is in response to R..'s comment of about an hour ago.

I'm a little new to the syscall subject, so please bear with me.

By "the kernel sysqueue syscall", do you mean the `__NR_rt_sigqueueinfo' call? That's the only one that I found when I did this:

grep -Ri 'NR.*queue' /usr/include

If that's the case, I think I'm not understanding your original point. The kernel will let (non-root) me use SI-QUEUE with a faked pid and uid without error. If I have the sending side coded thus:

#include 
#include 
#include 
#include 
#include 
#include 

int main(int    argc,
         char **argv
        )
{
  long john_silver;

  siginfo_t my_siginfo;

  if(argc!=2)
  {
    fprintf(stderr,"missing pid argument\n");

    exit(1);
  }

  john_silver=strtol(argv[1],NULL,0);

  if(kill(john_silver,SIGUSR1))
  {
    fprintf(stderr,"kill() fail\n");

    exit(1);
  }

  sleep(1);

  my_siginfo.si_signo=SIGUSR1;
  my_siginfo.si_code=SI_QUEUE;
  my_siginfo.si_pid=getpid();
  my_siginfo.si_uid=getuid();
  my_siginfo.si_value.sival_int=41;

  if(syscall(__NR_rt_sigqueueinfo,john_silver,SIGUSR1,&my_siginfo))
  {
    perror("syscall()");

    exit(1);
  }

  sleep(1);

  my_siginfo.si_signo=SIGUSR2;
  my_siginfo.si_code=SI_QUEUE;
  my_siginfo.si_pid=getpid()+1;
  my_siginfo.si_uid=getuid()+1;
  my_siginfo.si_value.sival_int=42;

  if(syscall(__NR_rt_sigqueueinfo,john_silver,SIGUSR2,&my_siginfo))
  {
    perror("syscall()");

    exit(1);
  }

  return 0;

} /* main() */

and the receiving side coded thus:

#include 
#include 
#include 
#include 
#include 
#include 

int signaled_flag=0;

siginfo_t received_information;

void
my_handler(int        signal_number,
           siginfo_t *signal_information,
           void      *we_ignore_this
          )
{
  memmove(&received_information,
          signal_information,
          sizeof(received_information)
         );

  signaled_flag=1;

} /* my_handler() */

/*--------------------------------------------------------------------------*/

int
main(void)
{
  pid_t            myself;

  struct sigaction the_action;

  myself=getpid();

  printf("signal receiver is process %d\n",myself);

  the_action.sa_sigaction=my_handler;
  sigemptyset(&the_action.sa_mask);
  the_action.sa_flags=SA_SIGINFO;

  if(sigaction(SIGUSR1,&the_action,NULL))
  {
    fprintf(stderr,"sigaction(SIGUSR1) fail\n");

    exit(1);
  }

  if(sigaction(SIGUSR2,&the_action,NULL))
  {
    fprintf(stderr,"sigaction(SIGUSR2) fail\n");

    exit(1);
  }

  for(;;)
  {
    while(!signaled_flag)
    {
      sleep(1);
    }

    printf("si_signo: %d\n",received_information.si_signo);
    printf("si_pid  : %d\n",received_information.si_pid  );
    printf("si_uid  : %d\n",received_information.si_uid  );

    if(received_information.si_signo==SIGUSR2)
    {
      break;
    }

    signaled_flag=0;
  }

  return 0;

} /* main() */

I can then run (non-root) the receiving side thus:

wally:~/tmp/20110326$ receive
signal receiver is process 9023
si_signo: 10
si_pid  : 9055
si_uid  : 4000
si_signo: 10
si_pid  : 9055
si_uid  : 4000
si_signo: 12
si_pid  : 9056
si_uid  : 4001
wally:~/tmp/20110326$ 

And see this (non-root) on the send end:

wally:~/tmp/20110326$ send 9023
wally:~/tmp/20110326$ 

As you can see, the third event has spoofed pid and uid. Isn't that what you originally objected to? There's no EINVAL or EPERM in sight. I guess I'm confused.