path: root/tools/hi-q.c

///////////////
// lightweight connection from qmail to filters e.g. spamassassin
// (hi-q filter, get it?)

// TODO:  Panic stop should signal all children.
// TODO:  Possibly:  Wait for all kids in parallel?
//      That's because they might finish out of order.

#include <unistd.h>
#include <stdlib.h>             /* for exit(), getenv() */
#include <stdio.h>
#include <errno.h>
#include <sys/types.h>          /* for fork(), wait() */
#include <sys/stat.h>
#include <sys/wait.h>

using namespace std;
#include <iostream>
#include <fstream>
#include <sstream>
#include <string>
#include <vector>
#include <sstream>

//  error exit codes, as stated in qmail.c
const int ex_spam = 21;
const int ex_syserr = 71;
const int ex_comerr = 74;

#define bufsize 16384

void panic(const int sts) {
  // FIXME: stop other children
  exit(sts);
}

void slurp(const int inch, const int ouch){
  char buf[bufsize];
  ssize_t todo;
  for (;;) {
    ssize_t got = read(inch, buf, bufsize);
    if (got == 0) {             // EoF
      break;
    }
    if (got < 0) {
      fprintf(stderr, "hi-q: input error: ");
      perror(0);
      panic(ex_comerr);
    }

    todo = got;
    while (todo) {
      ssize_t sent = write(ouch, buf, todo);
      if (sent < 0 && errno != EINTR) {
        fprintf(stderr, "hi-q: output error on fd%d : ", ouch);
        perror(0);
        panic(ex_comerr);
      }
      todo -= sent;
    }
  }
}


void probe_fd(){
  int ii;
  struct stat buf;
  for (ii = 0; ii < 16; ii++) {
    int rslt = fstat(ii, &buf);
    fprintf(stderr, "fd %2d status %2d", ii, rslt);
    if (rslt==0)
      fprintf(stderr, " : %d", (int)buf.st_dev);
    fprintf(stderr, "\n");
  }
  fprintf(stderr, "============\n");
}


void blurb(const int ii, const pid_t* kidpid) {
  int kidstatus;
  /*pid_t somekid = */ waitpid(kidpid[ii], &kidstatus, WUNTRACED);
  if (WIFEXITED(kidstatus))
    fprintf(stderr, "kid #%d (%d) exited with status %d\n",
              ii, kidpid[ii], WEXITSTATUS(kidstatus));
  if (WIFSIGNALED(kidstatus))
    fprintf(stderr, "kid #%d (%d) killed by signal %d\n",
              ii, kidpid[ii], WTERMSIG(kidstatus));

}

// We are fussy about the argument types because we want
// this to compile cleanly under g++ as well as gcc,
// and each is strict about different things, such that
// one or the other will complain unless everything is
// done just right.

// This is the way execve really behaves:
// the characters are held constant
// and the (char*) pointers are held constant:
int Execve(char const * fn,
           char const * const * argv,
           char const * const * env) {
// coerce the arg types to match the unwise declaration in unistd.h :
  return execve(fn, (char*const*) argv, (char*const*) env);
}

void usage() {
  cerr << "Usage:\n"
"  hi-q filter.conf\n"
"or\n"
"  HI_Q_CONF=filter.conf hi-q\n";
}

////////////////////////////////////////
// we have data coming in on fd 0.
// and envelope / control information coming in on fd 1.

void dump(const string var){
  char* str = getenv(var.c_str());
  if (str) cerr << "hi-q: " << var
                << " is set to '" << str << "'" << endl;
  else cerr << "hi-q: " << var << " is not set." << endl;
}

int xclose(int arg){
  cerr << "closing " << arg << endl;
  return close(arg);
}

extern char** environ;

int main(int argc, char** argv) {
  int verbose(1);
  int kidstatus;
  pid_t somekid;

  int rslt;
  int loose_end = 0;

#ifdef SpareStuff
  char* slurp2_args[] = {"/home/jsd/hack/slurp2", 0};
  char* echo_args[] = {"/bin/echo", "hi there", 0};
  char* wc_args[] = {"/usr/bin/wc", 0};
  char* cat_args[] = {"/bin/cat", 0};
  char* spama_args[] = {"/usr/local/bin/spamassassin", "-e", 0};
  char* spamc_args[] = {"/usr/local/bin/spamc", "-Z", "7", 0};
  char* qq_args[] = {"/var/qmail/bin/qmail-queue", 0};



  const char** joblist[] = {
    cat_args,
    slurp2_args,
    0           // required: zero terminates the list
  };

#endif

  typedef vector<string> VS;
  vector<VS> filter;
  string conf_var = "HI_Q_CONF";
  char* auth = getenv("QMAIL_AUTHORIZED");
  if (auth && *auth) conf_var = "HI_Q_AUCONF";

  char* conf_name;
  if (argc == 1) {
    conf_name = getenv(conf_var.c_str());
    if (!conf_name) {
      usage();
      exit(1);
    }
  }

  if (argc >= 2) {
    conf_name = argv[1];
  }

  if (argc >= 3) {
    if (auth && *auth) conf_name = argv[2];
  }

  if (argc > 3) {
    usage();
    exit(1);
  }

  ifstream conf;
  conf.open(conf_name);
  if (! conf.good()) {
    cerr << "hi-q: could not open filter.conf file '"
        << conf_name << "'" << endl;
    exit(1);
  }
  for (;;) {
    string line;
    if (!getline(conf, line).good()) break;
    istringstream parse(line);
    vector<string> job;
    while (parse.good()){
      string token;
      parse >> token;
      if (parse.fail()) break;
      if (token[0] == '#') break;
      job.push_back(token);
    }
    if (job.size()) filter.push_back(job);
  }
  unsigned int nkids = filter.size();
  if (0 && verbose) for (unsigned int ii = 0; ii < nkids; ii++) {
    cerr << "hi-q filter[" << ii << "] :; ";
    for (VS::const_iterator token = filter[ii].begin();
        token != filter[ii].end(); token++){
      cerr << *token << " ";
    }
    cerr << endl;
  }

  vector<pid_t> kidpid(nkids);  // indexed by kid number

  const int rEnd(0);  // end of a pipe for reading
  const int wEnd(1);  // end of a pipe for writing
  int sync[2];
  int resync[2];
  if (pipe(sync) != 0) cerr << "sync pipe failed" << endl;
  if (pipe(resync) != 0) cerr << "resync pipe failed" << endl;

// At this point, there are some loop invariants;
// (a) fd0 is open (standard input) and has the email msg,
//     ready for the next child to read, and
// (b) fd1 is open (nonstandard input) and has envelope information.
//     We need it to be open, so that pipe()
//     doesn't choose it.  That allows N-1 of the kids
//     to close it and dup() something useful onto it.

  for (unsigned int ii=0; ii < nkids; ii++){  /* loop starting all kids */
    int datapipe[2];
    int kid_end;

//xx fprintf(stderr, "Top of loop %d loose: %d\n", ii, loose_end);

    if (loose_end) {
      close(0);
      dup2(loose_end, 0);
      close(loose_end);
    }

// Create a pipe, which will be used to connect
// this child's fd1 to the next child's fd0 ...
// except for the last kid, which reads both fd0 and fd1,
// while writing nothing.

    rslt = pipe(datapipe);
    if (rslt < 0) {
      fprintf(stderr, "hi-q: could not create datapipe: ");
      perror(0);
      panic(ex_syserr);
    }

//xx fprintf(stderr, "pipe: %d %d\n", datapipe[0], datapipe[1]);

// For N-1 kids, the loose end feeds forward.
//    It will be written by this kid and read by the next kid.
// For the last kid, the loose end connects to hi-q.
//    It will be written by hi-q and read by the last kid.

    int lastkid = (ii == nkids-1);
#define flip(a,b) (lastkid ? b : a)
    loose_end = datapipe[flip(rEnd, wEnd)];
    kid_end   = datapipe[flip(wEnd, rEnd)];

    kidpid[ii] = fork();
    if (!kidpid[ii]) {  /*** child code ***/

      pid_t kidgroup(0);    // process group for all kids is
                            // equal to pid of kid#0
      if (ii) kidgroup = kidpid[0];
      if (setpgid(0, kidgroup) != 0) {
        cerr << "*** kid " << ii
           << " setpgid failed! " << errno << " ... ";
        perror(0);
      } else {
        // cerr << "*** kid " << ii << " setpgid OK" << endl;
      }

// ... everybody else has to wait for us to get this far ...
// ... so that the new process group will be valid ...
// Write-a-byte synchronization is released when the *first* guy writes.
      if (ii == 0) {
        int junk(1);
        write(sync[wEnd], &junk, 1);
        //cerr << "sync sent" << endl;
      }
#if 0
      cerr << "kid [" << ii << "] " << getpid()
        << "  kidpid[0]: " << kidpid[0]
        << "  pgid: " << getpgid(0)
        << " starts" << endl;
#endif

      close(resync[wEnd]);             // send resync

// ... now we must wait for everybody else, because ...
// ... if we do the exec(), the new process group becomes invalid ...
// Close synchronization is released when the *last* guy closes.
      if (ii==0) {
        int junk;
        //cerr << "about to read resync" << endl;
        ssize_t rslt = read(resync[rEnd], &junk, 1);
        if (rslt < 0 ) {
          cerr << "bad sync ... " << rslt << endl;
        } else {
          // cerr << "back from read resync, good: " << rslt << endl;
        }
      }

// Now that we are through creating pipes, we don't
// need to continue blocking fd1:
      close(1);

      close(loose_end);          // the reading end is none of this kid's business
                                // except last kid:  writing end

      rslt = dup2(kid_end, 1);   // the writing end is stdout for this kid
                                // except last kid:  nonstandard input
      if (rslt < 0) {
        fprintf(stderr, "hi-q: kid %d: dup2(%d,1) failed: ", ii, kid_end);
        perror(0);
        exit(ex_syserr);
      }

      close(kid_end);           // use fd1 instead now
      // OK, at this point this kid is set up to read fd0 and write fd1
      // (except last kid reads fd1 as well as fd0).
////  probe_fd();

      int ntok = filter[ii].size();
      const char* prog[1+ntok];
      for (int jj = 0; jj < ntok; jj++){
        prog[jj] = filter[ii][jj].c_str();
      }
      prog[ntok] = 0;
      close(resync[rEnd]);
      close(sync[rEnd]);
      close(sync[wEnd]);

      stringstream convert;
      convert << getpgid(0);
      const string grouper("HI_Q_GROUP=" + convert.str());
      if (putenv((char*)grouper.c_str()) != 0) {
        cerr << "putenv failed" << endl;
        exit(1);
      }
      rslt = Execve(prog[0], prog, environ);
      fprintf(stderr, "hi-q: failed to exec '%s': ", prog[0]);
      perror(0);
      exit(ex_syserr);
    }

    /*** parent code ***/
    if (kidpid[ii] < 0) {
      fprintf(stderr, "hi-q: failure to fork kid#%d: ", ii);
      perror(0);
      panic(ex_syserr);
    }
    close(kid_end);

// Let kid #0 run a little ways:
    if (ii==0) {
      int junk;
      //cerr << "about to read sync" << endl;
      ssize_t rslt = read(sync[rEnd], &junk, 1);
      if (rslt != 1) {
        cerr << "bad sync ... 1 != " << rslt << endl;
      } else {
        //cerr << "back from read sync, good: " << rslt << endl;
      }
    }

#if 0
    cerr << "apparent kid #" << ii
            << " (" << kidpid[ii] << ") "
            << endl;
#endif
  } /* end loop starting all kids */

// here with the whole pipeline of kids launched

  close(resync[wEnd]);  // important, so that block gets released
  close(resync[rEnd]);  // less important, just housecleaning

  close(sync[wEnd]);    // more housecleaning
  close(sync[rEnd]);

  close(0);             // Housecleaning: the reading end of stdin was
                        // delegated to the first child, 
                        // so we don't need it.

  if (verbose) for (unsigned int ii = 0; ii < nkids; ii++) {
    cerr << "hi-q filter[" << ii << "] "
    << kidpid[ii]
    << " :; ";
    for (VS::const_iterator token = filter[ii].begin();
        token != filter[ii].end(); token++){
      cerr << *token << " ";
    }
    cerr << endl;
  }

  for (unsigned int ii=0; ii<nkids-1; ii++){        /* loop over N-1 kids */
                                                /* _not_ including last kid */
#ifdef testing
    blurb(ii, kidpid);
#else
    somekid = waitpid(kidpid[ii], &kidstatus, WUNTRACED);
    if (somekid) {}             // avoid silly compiler warning
    if (WIFEXITED(kidstatus)) {
      int sts = WEXITSTATUS(kidstatus);
      if (sts == 1) {
        cerr << "hi-q says: kid " << ii
             << " i.e. '" << filter[ii][0] << "' reports spam."
             << endl;
        panic(ex_spam);
      }
      if (sts != 0) {
        cerr << "hi-q says: kid " << ii
             << " i.e. '" << filter[ii][0] << "'"
             << " suffered an error: " << sts
             << endl;
        panic(ex_syserr);
      }
      /* otherwise kidstatus==0 and we fall through */
    }
    else panic(ex_syserr);      // any kill, not a normal exit
#endif

  }

//xx fprintf(stderr, "slurping %d %d\n", 1, loose_end);

// All filters agree this is not spam.
// Now it is safe to transfer the envelope information:
  slurp(1, loose_end);
  close(1);
  close(loose_end);

// now that the envelope information has been transfered,
// wait for the last kid in the usual way
  {
    int ii = nkids-1;

#ifdef moretesting
fprintf(stderr, "About to wait for kid #%d (%d)\n",
                ii, kidpid[ii]);
    blurb(nkids-1, kidpid);
#endif
    somekid = waitpid(kidpid[ii], &kidstatus, WUNTRACED);
    if (WIFEXITED(kidstatus)) {
      int sts = WEXITSTATUS(kidstatus);
      cerr << "hi-q ends with status: " << sts << endl;
      return sts;
    }
  }

#ifdef testing
  sleep(1);
#endif
// the last kid did not exit but was killed:
  return ex_syserr;

}