[inn-innduct.git] / backends / innduct.c

/*
 * Newsfeeds file entries should look like this:
 *     host.name.of.site[/exclude,exclude,...]\
 *             :pattern,pattern...[/distribution,distribution...]\
 *             :Tf,Wnm
 *             :
 * or
 *     sitename[/exclude,exclude,...]\
 *             :pattern,pattern...[/distribution,distribution...]\
 *             :Tf,Wnm
 *             :host.name.of.site
 *
 * Four files full of
 *    token messageid
 * or might be blanked out
 *    <spc><spc><spc><spc>....
 *
 * F site.name                 main feed file
 *                                opened/created, then written, by innd
 *                                read by duct
 *                                unlinked by duct
 *                                tokens blanked out by duct when processed
 *   site.name_lock            lock preventing multiple ducts
 *                                to hold lock must open,F_SETLK[W]
 *                                  and then stat to check that locked file
 *                                  still has name site.name_lock
 *                                holder of this lock is "duct"
 *                                (only) lockholder may remove the lockfile
 * D site.name_flushing        temporary feed file during flush (or crash)
 *                                hardlink created by duct
 *                                unlinked by duct
 *   site.name_defer           431'd articles, still being written,
 *                                created, written, used by duct
 *
 *   site.name_backlog.<date>.<inum>
 *                             431'd articles, ready for innxmit or duct
 *                                created (link/mv) by duct
 *   site.name_backlog<anything-else>  (where <anything-else> does not
 *                                      contain '#' or '~') eg
 *   site.name_backlog.manual
 *                             anything the sysadmin likes (eg, feed files
 *                             from old feeds to be merged into this one)
 *                                created (link/mv) by admin
 *                                may be symlinks (in which case links
 *                                may be written through, but only links
 *                                will be removed.
 *
 *                             It is safe to remove backlog files manually,
 *                             if it's desired to throw away the backlog.
 *
 * Backlog files are also processed by innduct.  We find the oldest
 * backlog file which is at least a certain amount old, and feed it
 * back into our processing.  When every article in it has been read
 * and processed, we unlink it and look for another backlog file.
 *
 * If we don't have a backlog file that we're reading, we close the
 * defer file that we're writing and make it into a backlog file at
 * the first convenient opportunity.


   OVERALL STATES:

                                                                START
                                                                  |
     ,-->--.                                                 check F, D
     |     |                                                      |
     |     |                                                      |
     |     |  <----------------<---------------------------------'|
     |     |                                       F exists       |
     |     |                                       D ENOENT       |
     |     |  duct opens F                                        |
     |     V                                                      |
     |  Normal                                                    |
     |   F: innd writing, duct reading                            |
     |   D: ENOENT                                                |
     |     |                                                      |
     |     |  duct decides time to flush                          |
     |     |  duct makes hardlink                                 |
     |     |                                                      |
     |     V                            <------------------------'|
     |  Hardlinked                                  F==D          |
     |   F == D: innd writing, duct reading         both exist    |
     ^     |                                                      |
     |     |  duct unlinks F                                      |
     |     |                        <-----------<-------------<--'|
     |     |                           open D         F ENOENT    |
     |     |                           if exists                  |
     |     |                                                      |
     |     V                        <---------------------.       |
     |  Moved                                             |       |
     |   F: ENOENT                                        |       |
     |   D: innd writing, duct reading; or ENOENT         |       |
     |     |                                              |       |
     |     |  duct requests flush of feed                 |       |
     |     |   (others can too, harmlessly)               |       |
     |     V                                              |       |
     |  Flushing                                          |       |
     |   F: ENOENT                                        |       |
     |   D: innd flushing, duct; or ENOENT                |       |
     |     |                                              |       |
     |     |   inndcomm flush fails                       |       |
     |     |`-------------------------->------------------'       |
     |     |                                                      |
     |     |   inndcomm reports no such site                      |
     |     |`---------------------------------------------------- | -.
     |     |                                                      |  |
     |     |  innd finishes writing D, creates F                  |  |
     |     |  inndcomm reports flush successful                   |  |
     |     |                                                      |  |
     |     V                                                      |  |
     |  Separated                                <----------------'  |
     |   F: innd writing                            F!=D             /
     |   D: duct reading; or ENOENT                  both exist     /
     |     |                                                       /
     |     |  duct gets to the end of D                           /
     |     |  duct opens F too                                   /
     |     V                                                    /
     |  Finishing                                              /
     |   F: innd writing, duct reading                        |
     |   D: duct finishing                                    V
     |     |                                            Dropping
     |     |  duct finishes processing D                 F: ENOENT
     |     V  duct unlinks D                             D: duct reading
     |     |                                                  |
     `--<--'                                                  | duct finishes
                                                              |  processing D
                                                              | duct unlinks D
                                                              | duct exits
                                                              V
                                                        Dropped
                                                         F: ENOENT
                                                         D: ENOENT
                                                         duct not running

   "duct reading" means innduct is reading the file but also
   overwriting processed tokens.

 *
 *
 */


/*----- general definitions, probably best not changed -----*/

#define PERIOD_SECONDS 60

#define CONNCHILD_ESTATUS_STREAM   4
#define CONNCHILD_ESTATUS_NOSTREAM 5

#define INNDCOMMCHILD_ESTATUS_FAIL     6
#define INNDCOMMCHILD_ESTATUS_NONESUCH 7


/*----- configuration options -----*/

static char *sitename, *feedfile;
static int max_connections=10, max_queue_per_conn=200;
static int connection_setup_timeout=200, port=119, try_stream=1;
static int inndcomm_flush_timeout=100, quiet_if_locked=0;
static const char *remote_host;
static int reconnect_delay_periods, flushfail_retry_periods, open_wait_periods;
static int backlog_retry_minperiods, backlog_spontaneous_rescan_periods;
static const char *inndconffile;

static double accept_proportion;
static double nocheck_thresh_pct= 95.0;
static double nocheck_thresh; /* computed in main from _pct */
static double nocheck_decay_articles= 100; /* converted to _decay */
static double nocheck_decay; /* computed in main from _articles */
static int nocheck, nocheck_reported;


/*----- doubly linked lists -----*/

#define ISNODE(T)    T *next, *back;
#define LIST(T)      struct { T *head, *tail, *tailpred; int count; }

#define NODE(n) ((struct node*)&(n)->head)

#define LIST_ADDHEAD(l,n)                                               \
 (list_addhead((struct list*)&(l), NODE((n))), (void)(l).count++)
#define LIST_ADDTAIL(l,n)                                               \
 (list_addtail((struct list*)&(l), NODE((n))), (void)(l).count++)

#define LIST_REMHEAD(l)                                                   \
 ((l).count ? ((l).count--, (void*)list_remhead((struct list*)&(l))) : 0)
#define LIST_REMTAIL(l)                                                   \
 ((l).count ? ((l).count--, (void*)list_remtail((struct list*)&(l))) : 0)
#define LIST_REMOVE(l,n)                        \
 (list_remove(NODE((n))), (void)(l).count--)
#define LIST_INSERT(l,n,pred) \
 (list_insert((struct list*)&(l), NODE((n)), NODE((pred))), (void)(l).count++)


/*----- statistics -----*/

#define RESULT_COUNTS                           \
  RC(offered)                                   \
  RC(sent)                                      \
  RC(unwanted)                                  \
  RC(accepted)                                  \
  RC(rejected)                                  \
  RC(deferred)

typedef enum {
#define RC_INDEX(x) RCI_##x
  RESULT_COUNTS
  RCI_max
} ResultCountIndex;

typedef struct {
  int articles[2 /* checked */][RCI_max];
} Counts;


/*----- transmission buffers -----*/

#define CONNIOVS 128

typedef enum {
  xk_Malloc, xk_Const, xk_Artdata;
} XmitKind;

typedef struct {
  XmitKind kind;
  union {
    char *malloc_tofree;
    ARTHANDLE *sm_art;
  } info;
} XmitDetails;


/*----- core operational data structure types -----*/

struct Article {
  int midlen;
  int checked, sentbody;
  InputFile *ipf;
  TOKEN token;
  off_t offset;
  int blanklen;
  char messageid[1];
};

typedef struct InputFile {
  /* This is an instance of struct oop_readable */
  struct oop_readable readable; /* first */
  oop_readable_call *readable_callback;
  void *readable_callback_user;

  int fd;
  struct Filemon_Perfile *filemon;

  oop_read *rd;
  long inprogress; /* no. of articles read but not processed */
  off_t offset;

  Counts counts;
  char path[];
} InputFile;

#define SMS_LIST(X)                             \
  X(WAITING)                                    \
  X(NORMAL)                                     \
  X(FLUSHING)                                   \
  X(FLUSHFAIL)                                  \
  X(SEPARATED)                                  \
  X(DROPPING)

typedef enum {
#define SMS_DEF_ENUM(s) sm_##s,
  SMS_LIST(SMS_DEF_ENUM)
} StateMachineState;

static const char *sms_names[]= {
#define SMS_DEF_NAME(s) #s ,
  SMS_LIST(SMS_DEF_NAME)
  0
};

struct Conn {
  ISNODE(Conn);
  int fd, max_queue, stream;
  LIST(Article) queue; /* not yet told peer, or CHECK said send it */
  LIST(Article) sent; /* offered/transmitted - in xmit or waiting reply */
  struct iovec xmit[CONNIOVS];
  XmitDetails xmitd[CONNIOVS];
  int xmitu;
};


/*----- operational variables -----*/

static int nconns;
static LIST(Conn) idle, working, full;
static LIST(Article) *queue;

static char *path_lock, *path_flushing, *path_defer;

#define SMS(newstate, periods, why) \
   (statemc_setstate(sm_##newstate,(periods),#newstate,(why)))

static StateMachineState sms;
static FILE *defer;
static InputFile *main_input_file, *flushing_input_file, *backlog_input_file;
static int sm_period_counter;


/*----- function predeclarations -----*/

static void conn_check_work(Conn *conn);

static int filemon_init(void);
static void filemon_setfile(int mainfeed_fd, const char *mainfeed_path);
static void filemon_callback(void);

static void statemc_setstate(StateMachineState newsms, int periods,
                             const char *forlog, const char *why);

/*========== utility functions etc. ==========*/

static void perhaps_close(int *fd) { if (*fd) { close(*fd); fd=0; } }

static pid_t xfork(const char *what) {
  pid_t child;

  child= fork();
  if (child==-1) sysdie("cannot fork for %s",what);
  if (!child) postfork(what);
  return child;
}

static void on_fd_read_except(int fd, oop_call_fd callback) {
  loop->on_fd(loop, fd, OOP_READ,      callback, 0);
  loop->on_fd(loop, fd, OOP_EXCEPTION, callback, 0);
}
static void cancel_fd_read_except(int fd) {
  loop->cancel_fd(loop, fd, OOP_READ);
  loop->cancel_fd(loop, fd, OOP_EXCEPTION);
}

static void report_child_status(const char *what, int status) {
  if (WIFEXITED(status)) {
    int es= WEXITSTATUS(status);
    if (es)
      warn("%s: child died with error exit status %d",es);
  } else if (WIFSIGNALED(status)) {
    int sig= WTERMSIG(status);
    const char *sigstr= strsignal(sig);
    const char *coredump= WCOREDUMP(status) ? " (core dumped)" : "";
    if (sigstr)
      warn("%s: child died due to fatal signal %s%s", what, sigstr, coredump);
    else
      warn("%s: child died due to unknown fatal signal %d%s",
           what, sig, coredump);
  } else {
    warn("%s: child died with unknown wait status %d", status);
  }
}

static int xwaitpid(pid_t *pid, const char *what) {
  int status;

  r= kill(*pid, SIGKILL);
  if (r) sysdie("cannot kill %s child", what);

  pid_t got= waitpid(*pid, &status, WNOHANG);
  if (got==-1) sysdie("cannot reap %s child", what);

  *pid= 0;

  return status;
}

static void xunlink(const char *path, const char *what) {
  int r= unlink(path);
  if (r) sysdie("can't unlink %s %s", path, what);
}

static void check_isreg(const struct stat *stab, const char *path,
                        const char *what) {
  if (!S_ISREG(stab->st_mode))
    die("%s %s not a plain file (mode 0%lo)",
        what, path, (unsigned long)stab->st_mode);
}

static void xfstat(int fd, struct stat *stab_r, const char *what) {
  int r= fstab(path, stab);
  if (r) sysdie("could not fstat %s %s", what, path);
}

static void xfstat_isreg(int fd, struct stat *stab_r, const char *what) {
  xfstat(fd, stab_r, what);
  check_isreg(stab, path, what);
}

static void xlstat_isreg(const char *path, struct stat *stab,
                         int *enoent_r /* 0 means ENOENT is fatal */
                         const char *what) {
  int r= lstat(path, stab);
  if (r) {
    if (errno==ENOENT && enoent_r) { *enoent_r=1; return; }
    sysdie("could not lstat %s %s", what, path);
  }
  if (enoent_r) *enoent_r= 0;
  check_isreg(stab, path, what);
}

static int samefile(const struct stat *a, const struct stat *b) {
  assert(S_ISREG(a->st_mode));
  assert(S_ISREG(b->st_mode));
  return (a->st_ino == b->st_ino &&
          a->st_dev == b->st_dev);
}

/*========== logging ==========*/

static void logcore(int sysloglevel, const char *fmt, ...)
     __attribute__((printf,2,3))
static void logcore(int sysloglevel, const char *fmt, ...) {
  va_list al;
  va_start(al,fmt);
  if (become_daemon) {
    vsyslog(sysloglevel,fmt,al);
  } else {
    vfprintf(stderr,fmt,al);
    putc('\n',stderr);
  }
}

static void logv(int sysloglevel, const char *pfx, int errnoval,
                 int exitstatus, const char *fmt, va_list al)
     __attribute__((printf,4,0))
static void logv(int sysloglevel, const char *pfx, int errnoval,
                 int exitstatus, const char *fmt, va_list al) {
  char msgbuf[256];
  vsnprintf(msgbuf,sizeof(msgbuf), fmt,al);
  msgbuf[sizeof(msgbuf)-1]= 0;

  if (sysloglevel >= LOG_ERR && (errnoval==EACCES || errnoval==EPERM))
    sysloglevel= LOG_ERR; /* run by wrong user, probably */

  logcore(sysloglevel, "<%s>%s: %s%s%s",
         sitename, pfx, msgbuf,
         errnoval>=0 ? ": " : "",
         errnoval>=0 ? strerror(errnoval) : "");
}

#define logwrap(fn, pfx, sysloglevel, errno, estatus)   \
  static void fn(const char *fmt, ...)                  \
      __attribute__((printf,1,2));                      \
  static void fn(const char *fmt, ...) {                \
    va_list al;                                         \
    va_start(al,fmt);                                   \
    logv(sysloglevel, pfx, errno, estatus, fmt, al);    \
  }

logwrap(sysdie,   " critical", LOG_CRIT,   errno, 16);
logwrap(die,      " critical", LOG_CRIT,   -1,    16);

logwrap(sysfatal, " fatal",    LOG_ERR,    errno, 12);
logwrap(fatal,    " fatal",    LOG_ERR,    -1,    12);

logwrap(syswarn,  " warning",  LOG_WARN,   errno, 0);
logwrap(warn,     " warning",  LOG_WARN,   -1,    0);

logwrap(notice,   "",          LOG_NOTICE, -1,    0);
logwrap(info,     " info",     LOG_INFO,   -1,    0);
logwrap(debug,    " debug",    LOG_DEBUG,  -1,    0);


/*========== making new connections ==========*/

static int connecting_sockets[2]= {-1,-1};
static pid_t connecting_child;

static void connect_attempt_discard(void) {
  if (connecting_sockets[0])
    cancel_fd(connecting_sockets[0]);

  perhaps_close(&connecting_sockets[0]);
  perhaps_close(&connecting_sockets[1]);

  if (connecting_child) {
    int status= xwaitpid(&connecting_child, "connect");

    if (!(WIFEXITED(status) ||
          (WIFSIGNALED(status) && WTERMSIG(status) == SIGKILL)))
      report_child_status("connect", status);
  }
}

#define PREP_DECL_MSG_CMSG(msg)                 \
  struct msghdr msg;                            \
  memset(&msg,0,sizeof(msg));                   \
  char msg##cbuf[CMSG_SPACE(sizeof(fd))];       \
  msg.msg_control= msg##cbuf;                   \
  msg.msg_controllen= sizeof(msg##cbuf);

static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
  Conn *conn= 0;

  conn= xmalloc(sizeof(*conn));
  memset(conn,0,sizeof(*conn));

  DECL_MSG_CMSG(msg);
  struct cmsghdr *h= 0;
  ssize_t rs= recvmsg(fd, &msg, MSG_DONTWAIT);
  if (rs >= 0) h= CMSG_FIRSTHDR(&msg);
  if (!h) {
    int status;
    pid_t got= waitpid(connecting_child, &status, WNOHANG);
    if (got != -1) {
      assert(got==connecting_child);
      connecting_child= 0;
      if (WIFEXITED(status) &&
          (WEXITSTATUS(status) != 0
           WEXITSTATUS(status) != CONNCHILD_ESTATUS_STREAM &&
           WEXITSTATUS(status) != CONNCHILD_ESTATUS_NOSTREAM)) {
        /* child already reported the problem */
      } else if (WIFSIGNALED(status) && WTERMSIG(status) == SIGALARM) {
        warn("connect: connection attempt timed out");
      } else if (!WIFEXITED(status)) {
        report_child_status("connect", status);
        /* that's probably the root cause then */
      }
    } else {
      /* child is still running apparently, report the socket problem */
      if (rs < 0)
        syswarn("connect: read from child socket failed");
      else if (e == OOP_EXCEPTIONN)
        warn("connect: unexpected exception on child socket");
      else
        warn("connect: unexpected EOF on child socket");
    }
    goto x;
  }

#define CHK(field, val)                                                   \
  if (h->cmsg_##field != val) {                                           \
    die("connect: child sent cmsg with cmsg_" #field "=%d, expected %d"); \
    goto x;                                                               \
  }
  CHK(level, SOL_SOCKET);
  CHK(type,  SCM_RIGHTS);
  CHK(len,   CMSG_LEN(sizeof(conn-b>fd)));
#undef CHK

  if (CMSG_NXTHDR,&msg,h) { die("connect: child sent many cmsgs"); goto x; }

  memcpy(&conn->fd, CMSG_DATA(h), sizeof(conn->fd));

  pid_t got= waitpid(connecting_child, &status, 0);
  if (got==-1) sysdie("connect: real wait for child");
  assert(got == connecting_child);
  connecting_child= 0;

  if (!WIFEXITED(status)) { report_child_status("connect",status); goto x; }
  int es= WEXITSTATUS(status);
  switch (es) {
  case CONNCHILD_ESTATUS_STREAM:    conn->stream= 1;   break;
  case CONNCHILD_ESTATUS_NOSTREAM:  conn->stream= 0;   break;
  default:
    fatal("connect: child gave unexpected exit status %d", es);
  }

  set nonblocking;

  /* Phew! */
  LIST_ADDHEAD(idle, conn);
  notice(CN "connected %s", conn->fd, conn->stream ? "streaming" : "plain");
  connect_attempt_discard();
  check_master_queue();
  return 0;

 x:
  if (conn) {
    perhaps_close(&conn->fd);
    free(conn);
  }
  connect_attempt_discard();
}

static void connect_start() {
  assert(!connecting_sockets[0]);
  assert(!connecting_sockets[1]);
  assert(!connecting_child);

  notice("starting connection attempt");

  r= socketpair(AF_UNIX, SOCK_STREAM, 0, connecting_sockets);
  if (r) { syswarn("connect: cannot create socketpair for child"); goto x; }

  connecting_child= xfork("connection");

  if (!connecting_child) {
    FILE *cn_from, *cn_to;
    char buf[NNTP_STRLEN+100];
    int exitstatus= CONNCHILD_ESTATUS_NOSTREAM;

    r= close(connecting_sockets[0]);
    if (r) sysdie("connect: close parent socket in child");

    alarm(connection_setup_timeout);
    if (NNTPconnect(remote_host, port, &cn_from, &cn_to, buf) < 0) {
      if (buf[0]) {
        sanitise_inplace(buf);
        fatal("connect: rejected: %s", buf);
      } else {
        sysfatal("connect: connection attempt failed");
      }
    }
    if (NNTPsendpassword(remote_host, cn_from, cn_to) < 0)
      sysfatal("connect: authentication failed");
    if (try_stream) {
      if (fputs("MODE STREAM\r\n", cn_to) ||
          fflush(cn_to))
        sysfatal("connect: could not send MODE STREAM");
      buf[sizeof(buf)-1]= 0;
      if (!fgets(buf, sizeof(buf)-1, cn_from)) {
        if (ferror(cn_from))
          sysfatal("connect: could not read response to MODE STREAM");
        else
          fatal("connect: connection close in response to MODE STREAM");
      }
      int l= strlen(buf);
      assert(l>=1);
      if (buf[-1]!='\n') {
        sanitise_inplace(buf);
        fatal("connect: response to MODE STREAM is too long: %.100s...",
            remote_host, buf);
      }
      l--;  if (l>0 && buf[1-]=='\r') l--;
      buf[l]= 0;
      char *ep;
      int rcode= strtoul(buf,&ep,10);
      if (ep != buf[3]) {
        sanitise_inplace(buf);
        fatal("connect: bad response to MODE STREAM: %.50s", buf);
      }
      switch (rcode) {
      case 203:
        exitstatus= CONNCHILD_ESTATUS_STREAM;
        break;
      case 480:
      case 500:
        break;
      default:
        sanitise_inplace(buf);
        warn("connect: unexpected response to MODE STREAM: %.50s", buf);
        exitstatus= 2;
        break;
      }
    }
    int fd= fileno(cn_from);

    PREP_DECL_MSG_CMSG(msg);
    struct cmsghdr *cmsg= CMSG_FIRSTHDR(&msg);
    cmsg->cmsg_level= SOL_SOCKET;
    cmsg->cmsg_type=  SCM_RIGHTS;
    cmsg->cmsg_len=   CMSG_LEN(sizeof(fd));
    memcpy(CMSG_DATA(cmsg), &fd, sizeof(fd));

    msg.msg_controllen= cmsg->cmsg_len;
    r= sendmsg(connecting_sockets[1], &msg, 0);
    if (r) sysdie("sendmsg failed for new connection");

    _exit(exitstatus);
  }

  r= close(connecting_sockets[1]);  connecting_sockets[1]= 0;
  if (r) sysdie("connect: close child socket in parent");

  on_fd_read_except(connecting_sockets[0], connchild_event);
  return OOP_CONTINUE;

 x:
  connect_attempt_discard();
}


/*========== overall control of article flow ==========*/

static void check_master_queue(void) {
  if (!queue.count)
    return;

  Conn *last_assigned=0;
  for (;;) {
    if (working.head) {
      conn_assign_one_article(&working, &last_assigned);
    } else if (idle.head) {
      conn_assign_one_article(&idle, &last_assigned);
    } else if (nconns < maxconns && queue.count >= max_queue_per_conn &&
               !connecting_child && !connect_delay) {
      connect_delay= reconnect_delay_periods;
      connect_start();
    } else {
      break;
    }
  }
  conn_check_work(last_assigned);
}

static void conn_assign_one_article(LIST(Conn) *connlist,
                                    Conn **last_assigned) {
  Conn *conn= connlist->head;

  LIST_REMOVE(*connlist, conn);
  Article *art= LIST_REMHEAD(queue);
  LIST_ADDTAIL(conn->queue, art);
  LIST_ADD(*conn_determine_right_list(conn), conn);

  /* This slightly odd arrangement is so that we call conn_check_work
   * once after filling the queue for a new connection in
   * check_master_queue, rather than for each article. */
  if (conn != *last_assigned && *last_assigned)
    conn_check_work(*last_assigned);
  *last_assigned= conn;
}

static int conn_total_queued_articles(Conn *conn) {
  return conn->sent.count + conn->queue.count;
}

static LIST(Conn) *conn_determine_right_list(Conn *conn) {
  int inqueue= conn_total_queued_articles(conn);
  assert(inqueue <= max_queue);
  if (inqueue == 0) return &idle;
  if (inqueue == conn->max_queue) return &full;
  return &working;
}

static void *conn_writeable(oop_source *l, int fd, int ev, void *u) {
  check_conn_work(u);
  return OOP_CONTINUE;
}

static void conn_check_work(Conn *conn)  {
  void *rp= 0;
  for (;;) {
    conn_make_some_xmits(conn);
    if (!conn->xmitu) {
      loop->cancel_fd(loop, conn->fd, OOP_WRITE);
      return;
    }

    void *rp= conn_write_some_xmits(conn);
    if (rp==OOP_CONTINUE) {
      loop->on_fd(loop, conn->fd, OOP_WRITE, conn_writeable, conn);
      return;
    } else if (rp==OOP_HALT) {
      return;
    } else if (!rp) {
      /* transmitted everything */
    } else {
      abort();
    }
  }
}


/*========== article transmission ==========*/

static XmitDetails *xmit_core(Conn *conn, const char *data, int len,
                  XmitKind kind) { /* caller must then fill in details */
  struct iovec *v= &conn->xmit[conn->xmitu];
  XmitDetails *d= &conn->xmitd[conn->xmitu++];
  v->iov_base= data;
  v->iov_len= len;
  d->kind= kind;
  return d;
}

static void xmit_noalloc(Conn *conn, const char *data, int len) {
  xmit_core(conn,data,len, xk_Const);
}
#define XMIT_LITERAL(lit) (xmit_noalloc(conn, (lit), sizeof(lit)-1))

static void xmit_artbody(Conn *conn, ARTHANDLE *ah /* consumed */) {
  XmitDetails *d= xmit_core(conn, ah->data, ah->len, sk_Artdata);
  d->info.sm_art= ah;
}

static void xmit_free(XmitDetails *d) {
  switch (d->kind) {
  case xk_Malloc:  free(d->info.malloc_tofree);   break;
  case xk_Artdata: SMfreearticle(d->info.sm_art); break;
  case xk_Const:                                  break;
  default: abort();
  }
}

static void *conn_write_some_xmits(Conn *conn) {
  /* return values:
   *      0:            nothing more to write, no need to call us again
   *      OOP_CONTINUE: more to write but fd not writeable
   *      OOP_HALT:     disaster, have destroyed conn
   */
  for (;;) {
    int count= conn->xmitu;
    if (!count) return 0;

    if (count > IOV_MAX) count= IOV_MAX;
    ssize_t rs= writev(conn->fd, conn->xmit, count);
    if (rs < 0) {
      if (errno == EAGAIN) return OOP_CONTINUE;
      connfail(conn, "write failed: %s", strerror(errno));
      return OOP_HALT;
    }
    assert(rs > 0);

    for (done=0; rs && done<xmitu; done++) {
      struct iovec *vp= &conn->xmit[done];
      XmitDetails *dp= &conn->xmitd[done];
      if (rs > vp->iov_len) {
        rs -= vp->iov_len;
        xmit_free(dp);
      } else {
        vp->iov_base += rs;
        vp->iov_len -= rs;
      }
    }
    int newu= conn->xmitu - done;
    memmove(conn->xmit,  conn->xmit  + done, newu * sizeof(*conn->xmit));
    memmove(conn->xmitd, conn->xmitd + done, newu * sizeof(*conn->xmitd));
    conn->xmitu= newu;
  }
}

static void conn_make_some_xmits(Conn *conn) {
  for (;;) {
    if (conn->xmitu+5 > CONNIOVS)
      break;

    Article *art= LIST_REMHEAD(queue);
    if (!art) break;

    if (art->checked || (conn->stream && nocheck)) {
      /* actually send it */

      ARTHANDLE *artdata= SMretrieve();

      if (conn->stream) {
        if (artdata) {
          XMIT_LITERAL("TAKETHIS ");
          xmit_noalloc(conn, art->mid, art->midlen);
          XMIT_LITERAL("\r\n");
          xmit_artbody(conn, artdata);
        }
      } else {
        /* we got 235 from IHAVE */
        if (artdata) {
          xmit_artbody(conn, artdata);
        } else {
          XMIT_LITERAL(".\r\n");
        }
      }

      art->sent= 1;
      LIST_ADDTAIL(conn->sent, art);

      art->ipf->counts[art->checked].sent++;

    } else {
      /* check it */

      if (conn->stream)
        XMIT_LITERAL("IHAVE ");
      else
        XMIT_LITERAL("CHECK ");
      xmit_noalloc(art->mid, art->midlen);
      XMIT_LITERAL("\r\n");

      LIST_ADDTAIL(conn->sent, art);
      art->ipf->counts[art->checked].offered++;
    }
  }
}


/*========== handling responses from peer ==========*/

static const oop_rd_style peer_rd_style= {
  OOP_RD_DELIM_STRIP, '\n',
  OOP_RD_NUL_FORBID,
  OOP_RD_SHORTREC_FORBID
};

static void *peer_rd_err(oop_source *lp, oop_read *oread, oop_event ev,
                         const char *errmsg, int errnoval,
                         const char *data, size_t recsz, void *conn_v) {
  Conn *conn= conn_v;
  connfail(conn, "error receiving from peer: %s", errmsg);
  return OOP_CONTINUE;
}

static Article *article_reply_check(Connection *conn, const char *response,
                                    int code_indicates_streaming,
                                    const char *sanitised_response) {
  Article *art= LIST_REMHEAD(conn->sent);

  if (!art) {
    connfail(conn,
             "peer gave unexpected response when no commands outstanding: %s",
             sanitised_response);
    return 0;
  }

  if (code_indicates_streaming) {
    assert(!memchr(response, 0, 4)); /* ensured by peer_rd_ok */
    if (!conn->stream) {
      connfail("peer gave streaming response code "
               " to IHAVE or subsequent body: %s", sanitised_response);
      return 0;
    }
    const char *got_mid= response+4;
    int got_midlen= strcspn(got_mid, " \n\r");
    if (got_midlen<3 || got_mid[0]!='<' || got_mid[got_midlen-1]!='>') {
      connfail("peer gave streaming response with syntactically invalid"
               " messageid: %s", sanitised_response);
      return 0;
    }
    if (got_midlen != art->midlen ||
        memcmp(got_mid, art->messageid, got_midlen)) {
      connfail("peer gave streaming response code to wrong article -"
               " probable synchronisation problem; we offered: %s;"
               " peer said: %s",
               art->messageid, sanitised_response);
      return 0;
    }
  } else {
    if (conn->stream) {
      connfail("peer gave non-streaming response code to CHECK/TAKETHIS: %s",
               sanitised_response);
      return 0;
    }
  }

  return art;
}

static void update_nocheck(int accepted) {
  accept_proportion *= accept_decay;
  accept_proportion += accepted;
  int new_nocheck= accept_proportion >= nocheck_thresh;
  if (new_nocheck && !nocheck_reported) {
    notice("entering nocheck mode for the first time");
    nocheck_reported= 1;
  } else if (new_nocheck != nockech) {
    debug("nocheck mode %s", new_nocheck ? "start" : "stop");
  }
  nocheck= new_nocheck;
}

static void article_done(Connection *conn, Article *art, int whichcount) {
  *count++;
  art->ipf->counts.articles[art->checked][whichcount]++;
  if (whichcount == RC_accepted) update_nocheck(1);
  else if (whichcount == RC_unwanted) update_nocheck(0);

  InputFile *ipf= art->ipf;
  while (art->blanklen) {
    static const char spaces[]=
      "                                                                "
      "                                                                "
      "                                                                "
      "                                                                ";
    int w= art->blanklen;  if (w >= sizeof(spaces)) w= sizeof(spaces)-1;
    int r= pwrite(ipf->fd, spaces, w, art->offset);
    if (r==-1) {
      if (errno==EINTR) continue;
      sysdie("failed to blank entry for %s (length %d at offset %lu) in %s",
             art->messageid, art->blanklen, art->offset, ipf->path);
    }
    assert(r>=0 && r<=w);
    art->blanklen -= w;
    art->offset += w;
  }

  ipf->inprogress--;
  assert(ipf->inprogress >= 0);

  if (!ipf->inprogress)
    loop->on_time(loop, OOP_TIME_NOW, statemc_check_input_done, ipf);

  free(art);
}

static void *peer_rd_ok(oop_source *lp, oop_read *oread, oop_event ev,
                        const char *errmsg, int errnoval,
                        const char *data, size_t recsz, void *conn_v) {
  Conn *conn= conn_v;

  if (ev == OOP_RD_EOF) {
    connfail(conn, "unexpected EOF from peer");
    return OOP_CONTINUE;
  }
  assert(ev == OOP_RD_OK);

  char *ep;
  unsigned long code= strtoul(data, &ep, 10);
  if (ep != data+3 || *ep != ' ' || data[0]=='0') {
    char sanibuf[100];
    const char *p= data;
    char *q= sanibuf;
    *q++= '`';
    for (;;) {
      if (q > sanibuf+sizeof(sanibuf)-8) { strcpy(q,"..."); break; }
      int c= *p++;
      if (!c) { *q++= '\''; break; }
      if (c>=' ' && c<=126 && c!='\\') { *q++= c; continue; }
      sprintf(q,"\\x%02x",c);
      q += 4;
    }
    connfail(conn, "badly formatted response from peer: %s", sanibuf);
    return OOP_CONTINUE;
  }

  if (conn->quitting) {
    if (code!=205) {
      connfail(conn, "peer gave failure response to QUIT: %s", sani);
      return OOP_CONTINUE;
    }
    conn close ok;
    return;
  }

  Article *art;

#define GET_ARTICLE                                                         \
  art= article_reply_check(conn, data, code_streaming, sani);               \
  if (art) ; else return OOP_CONTINUE /* reply_check has failed the conn */

#define ARTICLE_DEALTWITH(streaming,how)                        \
  code_streaming= (streaming)                                   \
  GET_ARTICLE;                                                  \
  article_done(conn, art, RC_##how);  break;

#define PEERBADMSG(m) connfail(conn, m ": %s", sani);  return OOP_CONTINUE

  int code_streaming= 0;

  switch (code) {

  case 400: PEERBADMSG("peer stopped accepting articles");
  case 503: PEERBADMSG("peer timed us out");
  default:  PEERBADMSG("peer sent unexpected message");

  case 435: ARTICLE_DEALTWITH(0,unwanted); /* IHAVE says they have it */
  case 438: ARTICLE_DEALTWITH(1,unwanted); /* CHECK/TAKETHIS: they have it */

  case 235: ARTICLE_DEALTWITH(0,accepted); /* IHAVE says thanks */
  case 239: ARTICLE_DEALTWITH(1,accepted); /* TAKETHIS says thanks */

  case 437: ARTICLE_DEALTWITH(0,rejected); /* IHAVE says rejected */
  case 439: ARTICLE_DEALTWITH(1,rejected); /* TAKETHIS says rejected */

  case 238: /* CHECK says send it */
    code_streaming= 1;
  case 335: /* IHAVE says send it */
    GET_ARTICLE;
    count_checkedwanted++;
    LIST_ADDTAIL(conn->queue);
    if (art->checked) {
      connfail("peer gave %d response to article body: %s",code, sani);
      return OOP_CONTINUE;
    }
    art->checked= 1;
    break;

  case 431: /* CHECK or TAKETHIS says try later */
    code_streaming= 1;
  case 436: /* IHAVE says try later */
    GET_ARTICLE;
    open_defer();
    if (fprintf(defer, "%s %s\n", TokenToText(art->token), art->messageid) <0
        || fflush(defer))
      sysfatal("write to defer file %s",path_defer);
    article_done(conn, art, RC_deferred);
    break;

  }

  check_check_work(conn);
  return OOP_CONTINUE;
}


/*========== monitoring of input files ==========*/

static void feedfile_eof(InputFile *ipf) {
  assert(ipf != main_input_file); /* promised by tailing_try_read */

  inputfile_tailing_stop(ipf);

  if (ipf == backlog_input_file) {
    assert(ipf->fd >= 0);
    if (close(ipf->fd)) sysdie("could not close backlog file %s", ipf->path);
    ipf->fd= -1;
    return;
  }

  assert(ipf == flushing_input_file);

  inputfile_tailing_stop(ipf);
  assert(ipf->fd >= 0);
  if (close(ipf->fd)) sysdie("could not close input file %s", ipf->path);
  ipf->fd= -1;

  assert(sms==sm_SEPARATED || sms==sm_DROPPING);

  if (main_input_file)
    inputfile_tailing_start(main_input_file);
}

static InputFile *open_input_file(const char *path) {
  int fd= open(path, O_RDONLY);
  if (fd<0) {
    if (errno==ENOENT) return 0;
    sysfatal("unable to open input file %s", path);
  }

  InputFile *ipf= xmalloc(sizeof(*ipf) + strlen(path) + 1);
  memset(ipf,0,sizeof(*ipf));

  ipf->readable.on_readable= tailing_on_readable;
  ipf->readable.on_cancel=   tailing_on_cancel;
  ipf->readable.try_read=    tailing_try_read;

  ipf->fd= fd;
  strcpy(ipf->path, path);

  return ipf;
}

static void close_input_file(InputFile *ipf) {
  assert(!ipf->readable_callback); /* must have had ->on_cancel */
  assert(!ipf->filemon); /* must have had inputfile_tailing_stop */
  assert(!ipf->rd); /* must have had inputfile_tailing_stop */
  assert(!ipf->inprogress); /* no dangling pointers pointing here */

  if (ipf->fd >= 0)
    if (close(ipf->fd)) sysdie("could not close input file %s", ipf->path);

  free(ipf);
}


/*---------- dealing with articles read in the input file ----------*/

typedef void *feedfile_got_article(oop_source *lp, oop_read *rd,
                                   oop_rd_event ev, const char *errmsg,
                                   int errnoval,
                                   const char *data, size_t recsz,
                                   void *ipf_v) {
  InputFile *ipf= ipf_v;
  Article *art;
  char tokentextbuf[sizeof(TOKEN)*2+3];

  if (!data) { feedfile_eof(ipf); return OOP_CONTINUE; }

  if (data[0] && data[0]!=' ') {
    char *space= strchr(data,' ');
    int tokenlen= space-data;
    int midlen= (int)recsz-tokenlen-1;
    if (midlen < 0) goto bad_data;

    if (tokenlen != sizeof(TOKEN)*2+2) goto bad_data;
    memcpy(tokentextbuf, data, tokenlen);
    tokentextbuf[tokenlen]= 0;
    if (!IsToken(tokentextbuf)) goto bad_data;

    art= xmalloc(sizeof(*art) - 1 + midlen + 1);
    art->offset= ipf->offset;
    art->blanklen= recsz;
    art->midlen= midlen;
    art->checked= art->sentbody= 0;
    art->ipf= ipf;  ipf->inprogress++;
    art->token= TextToToken(tokentextbuf);
    strcpy(art->messageid, space+1);
    LIST_ADDTAIL(queue, art);
  }
  ipf->offset += recsz + 1;

  if (sms==sm_NORMAL && ipf==main_input_file &&
      (ipf->offset >= flush_threshold || !until_spontaneous_flush) {

    notice("starting flush (%lu >= %lu)",
           (unsigned long)ipf->offset, (unsigned long)flush_threshold);

    int r= link(feedfile, duct_path);
    if (r) sysdie("link feedfile %s to flushing file %s", feedfile, dut_path);
    /* => Hardlinked */

    xunlink(feedfile, "old feedfile link");
    /* => Moved */

    spawn_inndcomm_flush(); /* => Flushing, sets sms to sm_FLUSHING */
  }

  check_master_queue();
}


/*========== tailing input file ==========*/

static void filemon_start(InputFile *ipf) {
  assert(!ipf->filemon);

  ipf->filemon= xmalloc(sizeof(*ipf->filemon));
  memset(ipf->filemon, 0, sizeof(*ipf->filemon));
  filemon_method_startfile(ipf, ipf->filemon);
}

static void filemon_stop(InputFile *ipf) {
  if (!ipf->filemon) return;
  filemon_method_stopfile(ipf, ipf->filemon);
  free(ipf->filemon);
  ipf->filemon= 0;
}

static void filemon_callback(InputFile *ipf) {
  ipf->readable_callback(ipf->readable_callback_user);
}

static void *tailing_rable_call_time(oop_source *loop, struct timeval tv,
                                     void *user) {
  InputFile *ipf= user;
  return ipf->readable_callback(ipf->readable_callback_user);
}

static void on_cancel(struct oop_readable *rable) {
  InputFile *ipf= (void*)rable;

  if (ipf->filemon) filemon_stopfile(ipf);
  loop->cancel_time(loop, OOP_TIME_NOW, tailing_rable_call_time, ipf);
  ipf->readable_callback= 0;
}

static int tailing_on_readable(struct oop_readable *rable,
                                oop_readable_call *cb, void *user) {
  InputFile *ipf= (void*)rable;

  tailing_on_cancel(rable);
  ipf->readable_callback= cb;
  ipf->readable_callback_user= user;
  filemon_startfile(ipf);

  loop->on_time(loop, OOP_TIME_NOW, tailing_rable_call_time, ipf);
  return 0;
}

static ssize_t tailing_try_read(struct oop_readable *rable, void *buffer,
                                size_t length) {
  InputFile *ipf= (void*)rable;
  for (;;) {
    ssize_t r= read(ipf->fd, buffer, length);
    if (r==-1) {
      if (errno==EINTR) continue;
      return r;
    }
    if (!r) {
      if (ipf==main_input_file) {
        errno=EAGAIN;
        return -1;
      } else if (ipf==flushing_input_file) {
        assert(ipf->fd>=0);
        assert(sms==sm_SEPARATED || sms==sm_DROPPING);
      } else if (ipf==backlog_input_file) {
        assert(ipf->fd>=0);
      } else {
        abort();
      }
    }
    return r;
  }
}

/*---------- filemon implemented with inotify ----------*/

#if defined(HAVE_INOTIFY) && !defined(HAVE_FILEMON)
#define HAVE_FILEMON

#include <linux/inotify.h>

static int filemon_inotify_fd;
static int filemon_inotify_wdmax;
static InputFile **filemon_inotify_wd2ipf;

typedef struct Filemon_Perfile {
  int wd;
} Filemon_Inotify_Perfile;

static void filemon_method_startfile(InputFile *ipf, Filemon_Perfile *pf) {
  int wd= inotify_add_watch(filemon_inotify_fd, ipf->path, IN_MODIFY);
  if (wd < 0) sysdie("inotify_add_watch %s", ipf->path);

  if (wd >= filemon_inotify_wdmax) {
    int newmax= wd+2;
    filemon_inotify_wd= xrealloc(filemon_inotify_wd2ipf,
                                 sizeof(*filemon_inotify_wd2ipf) * newmax);
    memset(filemon_inotify_wd2ipf + filemon_inotify_wdmax, 0,
           sizeof(*filemon_inotify_wd2ipf) * (newmax - filemon_inotify_wdmax));
    filemon_inotify_wdmax= newmax;
  }

  assert(!filemon_inotify_wd2ipf[wd]);
  filemon_inotify_wd2ipf[wd]= ipf;

  debug("filemon inotify startfile %p wd=%d wdmax=%d",
        ipf, wd, filemon_inotify_wdmax);

  pf->wd= wd;
}

static void filemon_method_stopfile(InputFile *ipf, Filemon_Perfile *pf) {
  int wd= pf->wd;
  debug("filemon inotify stopfile %p wd=%d", ipf, wd);
  int r= inotify_rm_watch(filemon_inotify_fd, filemon_inotify_wd);
  if (r) sysdie("inotify_rm_watch");
  filemon_inotify_wd2ipf[wd]= 0;
}

static void *filemon_inotify_readable(oop_source *lp, int fd,
                                      oop_event e, void *u) {
  struct inotify_event iev;
  for (;;) {
    int r= read(filemon_inotify_fd, &iev, sizeof(iev));
    if (r==-1) {
      if (errno==EAGAIN) break;
      sysdie("read from inotify master");
    } else if (r==sizeof(iev)) {
      assert(iev.wd >= 0 && iev.wd < filemon_inotify_wdmax);
    } else {
      die("inotify read %d bytes wanted struct of %d", r, (int)sizeof(iev));
    }
    InputFile *ipf= filemon_inotify_wd2ipf[iev.wd];
    debug("filemon inotify readable read %p wd=%p", iev.wd, ipf);
    filemon_callback(ipf);
  }
  return OOP_CONTINUE;
}

static int filemon_method_init(void) {
  filemon_inotify_fd= inotify_init();
  if (filemon_inotify_fd<0) {
    syswarn("could not initialise inotify: inotify_init failed");
    return 0;
  }
  set nonblock;
  loop->on_fd(loop, filemon_inotify_fd, OOP_READ, filemon_inotify_readable);

  debug("filemon inotify init filemon_inotify_fd=%d", filemon_inotify_fd);
  return 1;
}

#endif /* HAVE_INOTIFY && !HAVE_FILEMON *//

/*---------- filemon dummy implementation ----------*/

#if !defined(HAVE_FILEMON)

typedef struct Filemon_Perfile { int dummy; } Filemon_Dummy_Perfile;

static int filemon_method_init(void) { return 0; }
static void filemon_method_startfile(InputFile *ipf, Filemon_Perfile *pf) { }
static void filemon_method_stopfile(InputFile *ipf, Filemon_Perfile *pf) { }

#endif /* !HAVE_FILEMON */

/*---------- interface to start and stop an input file ----------*/

static const oop_rd_style feedfile_rdstyle= {
  OOP_RD_DELIM_STRIP, '\n',
  OOP_RD_NUL_FORBID,
  OOP_RD_SHORTREC_EOF,
};

static void inputfile_tailing_start(InputFile *ipf) {
  assert(!ipf->fd);
  ipf->readable->on_readable= tailing_on_readable;
  ipf->readable->on_cancel=   tailing_on_cancel;
  ipf->readable->try_read=    tailing_try_read;
  ipf->readable->delete_tidy= 0; /* we never call oop_rd_delete_{tidy,kill} */
  ipf->readable->delete_kill= 0;

  ipf->readable_callback= 0;
  ipf->readable_callback_user= 0;

  ipf->rd= oop_rd_new(loop, &ipf->readable, 0,0);
  assert(ipf->fd);

  int r= oop_rd_read(ipf->rd, &feedfile_rdstyle, MAX_LINE_FEEDFILE,
                     feedfile_got_article,ipf, feedfile_problem,ipf);
  if (r) sysdie("unable start reading feedfile %s",ipf->path);
}

static void inputfile_tailing_stop(InputFile *ipf) {
  assert(ipf->fd);
  oop_rd_cancel(ipf->rd);
  oop_rd_delete(ipf->rd);
  ipf->rd= 0;
  assert(!ipf->filemon); /* we shouldn't be monitoring it now */
}


/*========== interaction with innd - state machine ==========*/

/* See official state diagram at top of file.  We implement
 * this as follows:

            .=======.
            ||START||
            `======='
                |
                | open F
                |
                |    F ENOENT
                |`---------------------------------------------------.
      F OPEN OK |                                                    |
                |`---------------- - - -                             |
       D ENOENT |       D EXISTS   see OVERALL STATES diagram        |
                |                  for full startup logic            |
     ,--------->|                                                    |
     |          V                                                    |
     |     ============                                       try to |
     |      NORMAL                                            open D |
     |     [Normal]                                                  |
     |      main F tail                                              |
     |     ============                                              V
     |          |                                                    |
     |          | F IS SO BIG WE SHOULD FLUSH, OR TIMEOUT            |
     ^          | hardlink F to D                                    |
     |     [Hardlinked]                                              |
     |          | unlink F                                           |
     |          | our handle onto F is now onto D                    |
     |     [Moved]                                                   |
     |          |                                                    |
     |          |<-------------------<---------------------<---------+
     |          |                                                    |
     |          | spawn inndcomm flush                               |
     |          V                                                    |
     |     ================                                          |
     |      FLUSHING                                                 |
     |     [Flushing]                                                |
     |     main D tail/none                                          |
     |     ================                                          |
     |          |                                                    |
     |          |   INNDCOMM FLUSH FAILS                             ^
     |          |`----------------------->----------.                |
     |          |                                   |                |
     |          |   NO SUCH SITE                    V                |
     ^          |`--------------->----.          ================    |
     |          |                      \          FLUSHFAIL          |
     |          |                       \         [Moved]            |
     |          | FLUSH OK               \       main D tail/none    |
     |          | open F                  \      ================    |
     |          |                          \        |                |
     |          |                           \       | TIME TO RETRY  |
     |<--------'|                   ,---<---'\      `----------------'
     |  D NONE  |                   | D NONE  \
     |          |                   |          \
     |          |                   |           |
     |          |                   V           |
     |          V                   |           V
     |     =============            |        ============
     |      SEPARATED/              |         DROPPING/
     |      flsh->fd>=0             |         flsh->fd>=0
     |     [Separated]              |        [Dropping]
     |      main F idle             |         main none
     |      old D tail              |         old D tail
     |     =============            |        ============
     |          |                   |          |
     ^          | EOF ON D          |          | EOF ON D
     |          V                   |          V
     |     ===============          |        ===============
     |      SEPARATED/              |         DROPPING/
     |      flsh->fd==-1            V         flsh->fd==-1
     |     [Finishing]              |        [Dropping]
     |      main F tail             `.        main none
     |      old D closed              `.      old D closed
     |     ===============              `.   ===============
     |          |                         `.     |
     |          | ALL D PROCESSED           `.   | ALL D PROCESSED
     |          V install defer as backlog    `. V install defer as backlog
     ^          | close D                       \| close D
     |          | unlink D                       | unlink D
     |          |                                | unlink lock
     |          |                                | exit
     `----------'                                V
                                             ==========
                                              (ESRCH)
                                             [Droppped]
                                             ==========
 */

static void statemc_init(void) {
  struct stat stab, stabf;
  int noent;

  path_lock=        xasprintf("%s_lock",      feedfile);
  path_flushing=    xasprintf("%s_flushing",  feedfile);
  path_defer=       xasprintf("%s_defer",     feedfile);
  globpat_backlog=  xasprintf("%s_backlog*",  feedfile);

  for (;;) {
    int lockfd= open(path_lock, O_CREAT|O_RDWR, 0600);
    if (lockfd<0) sysfatal("open lockfile %s", path_lock);

    struct flock fl;
    memset(&fl,0,sizeof(fl));
    fl.l_type= F_WRLCK;
    fl.l_whence= SEEK_SET;
    r= fcntl(lockfd, F_SETLK, &fl);
    if (r==-1) {
      if (errno==EACCES || errno==EAGAIN) {
        if (quiet_if_locked) exit(0);
        fatal("another duct holds the lockfile");
      }
      sysdie("fcntl F_SETLK lockfile %s", path_lock);
    }

    xfstat_isreg(lockfd, &stabf, "lockfile");
    xlstat_isreg(path_lock, &stab, &noent, "lockfile");

    if (!noent && samefile(&stab, &stabf))
      break;

    if (close(lockfd))
      sysdie("could not close stale lockfile %s", path_lock);
  }
  debug("startup: locked");

  search_backlog_file();

  xlstat_isreg(path_defer, &stab, &noent, "defer file");
  if (noent) {
    debug("startup: ductdefer ENOENT");
  } else {
    debug("startup: ductdefer nlink=%ld", (long)stab.st_nlink);
    switch (stab.st_nlink==1) {
    case 1:
      open_defer(); /* so that we will later close it and rename it */
      break;
    case 2:
      xunlink(path_defer, "stale defer file link"
              " (presumably hardlink to backlog file)");
      break;
    default:
      die("defer file %s has unexpected link count %d",
          path_defer, stab.st_nlink);
    }
  }

  InputFile *file_d= open_input_file(path_flushing);

  if (file_d) {
    struct stat stab_f, stab_d;

    xlstat_isreg(feedfile, &stab_f, &noent, "feed file");
    if (noent) {
      debug("startup: D exists, F ENOENT => Moved");
      goto found_moved;
    }

    debug("startup: F and D both exist");

    xfstat_isreg(file_d->fd, &stab_d, "flushing file");

    if (samefile(&stab_d, &stab_f)) {
      debug("startup: F==D => Hardlinked");
      xunlink(path_flushing, "feed file (during startup)");
    found_moved:
      debug(" => Moved");
      startup_set_input_file(file_d);
      spawn_inndcomm_flush(); /* => Flushing, sets sms to sm_FLUSHING */
    } else {
      debug("F!=D => Separated");
      SMS(SEPARATED, 0, "found both old and current feed files");
      startup_set_input_file(file_d);
    }
  } else {
    debug("startup: D ENOENT => Nothing");
    SMS(WAITING, open_wait_periods, "no feed file currently exists");
  }
}

static void statemc_poll(void) {
  if (sms==sm_WAITING) { statemc_waiting_poll(); return; }

  if (!sm_period_counter) return;
  sm_period_counter--;
  assert(sm_period_counter>=0);

  if (sm_period_counter) return;
  switch (sms) {
  case sm_WAITING:
    fatal("timed out waiting for innd to create feed file %s", feedfile);
  case sm_FLUSHFAIL:
    spawn_inndcomm_flush(void);
    break;
  default:
    abort();
  }
}

static void statemc_waiting_poll(void) {
  InputFile *file_f= open_input_file(feedfile);
  if (!file_f) return;
  startup_set_input_file(file_d);
  SMS(NORMAL, 0, "found and opened feed file");
}

static void startup_set_input_file(InputFile *f) {
  assert(!main_input_file);
  main_input_file= f;
  until_spontaneous_flush= spontaneous_flush_periods;
  inputfile_tailing_start(f);
}

static void *statemc_check_input_done(oop_source *lp,
                                      struct timeval now, void *ipf_v) {
  InputFile *ipf= ipf_v;
  struct stat stab;

  if (ipf->inprogress) return; /* new article in the meantime */
  if (ipf->fd >= 0); return; /* not had EOF */

  if (ipf == backlog_input_file) {
    notice_processed(ipf,"backlog file",ipf->path);
    close_input_file(ipf);
    if (unlink(ipf->path)) {
      if (errno != ENOENT)
        sysdie("could not unlink processed backlog file %s", ipf->path);
      warn("backlog file %s vanished while we were reading it"
           " so we couldn't remove it (but it's done now, anyway)",
           ipf->path);
    }
    backlog_input_file= 0;
    search_backlog_file();
    return;
  }

  assert(ipf == flushing_input_file);
  assert(sms==sm_SEPARATED || sms==sm_DROPPING);

  notice_processed(ipf,"feed file",0);

  close_defer();

  xunlink(path_flushing, "old flushing file");

  if (sms==sm_DROPPING) {
    if (search_backlog_file()) {
      debug("feed dropped but still backlogs to process");
      return;
    }
    notice("feed dropped and our work is complete");
    xunlink(path_lock, "lockfile for old feed");
    exit(0);
  }

  open_defer();

  close_input_file(flushing_input_file);
  flushing_input_file= 0;

  notice("flush complete");
  SMS(NORMAL, 0, "flush complete");
}

static void statemc_setstate(StateMachineState newsms, int periods,
                             const char *forlog, const char *why) {
  sms= newsms;
  sm_period_counter= periods;
  if (periods) {
    info("%s[%d] %s",periods,forlog,why);
  } else {
    info("%s %s",forlog,why);
  }
}

/*---------- defer and backlog files ----------*/

static void open_defer(void) {
  struct stat stab;

  if (defer) return;

  defer= fopen(path_defer, "a+");
  if (!defer) sysfatal("could not open defer file %s", path_defer);

  /* truncate away any half-written records */

  xfstat_isreg(fileno(defer), &stab, "newly opened defer file");

  if (stab.st_size > LONG_MAX)
    die("defer file %s size is far too large", path_defer);

  if (!stab.st_size)
    return;

  long orgsize= stab.st_size;
  long truncto= stab.st_size;
  for (;;) {
    if (!truncto) break; /* was only (if anything) one half-truncated record */
    if (fseek(defer, truncto-1, SEEK_SET) < 0)
      sysdie("seek in defer file %s while truncating partial", path_defer);

    r= getc(defer);
    if (r==EOF) {
      if (ferror(defer))
        sysdie("failed read from defer file %s", path_defer);
      else
        die("defer file %s shrank while we were checking it!", path_defer);
    }
    if (r=='\n') break;
    truncto--;
  }

  if (stab.st_size != truncto) {
    warn("truncating half-record at end of defer file %s -"
         " shrinking by %ld bytes from %ld to %ld",
         path_defer, orgsize - truncto, orgsize, truncto);

    if (fflush(defer))
      sysfatal("could not flush defer file %s", path_defer);
    if (ftruncate(fileno(defer), truncto))
      sysdie("could not truncate defer file %s", path_defer);

  } else {
    info("continuing existing defer file %s (%ld bytes)",
         path_defer, orgsize);
  }
  if (fseek(defer, truncto, SEEK_SET))
    sysdie("could not seek to new end of defer file %s", path_defer);
}

static void close_defer(void) {
  if (!defer)
    return;

  xfstat(fileno(defer), &stab, "defer file");

  if (fclose(defer)) sysfatal("could not close defer file %s", path_defer);
  defer= 0;

  char *backlog= xasprintf("%s_backlog_%lu.%lu", feedfile,
                           (unsigned long)now.tv_sec,
                           (unsigned long)stab.st_ino);
  if (link(path_defer, path_backlog))
    sysfatal("could not install defer file %s as backlog file %s",
           path_defer, backlog);
  if (unlink(path_defer))
    sysdie("could not unlink old defer link %s to backlog file %s",
           path_defer, backlog);

  if (until_backlog_nextscan < 0 ||
      until_backlog_nextscan > backlog_retry_minperiods + 1)
    until_backlog_nextscan= backlog_retry_minperiods + 1;
}

static void poll_backlog_file(void) {
  if (until_backlog_nextscan < 0) return;
  if (until_backlog_nextscan-- > 0) return;
  search_backlog_file();
}

static int search_backlog_file(void) {
  /* returns non-0 iff there are any backlog files */

  glob_t gl;
  int r;
  struct stat stab;
  const char *oldest_path=0;
  time_t oldest_mtime, now;

  if (backlog_input_file) return 3;

 try_again:

  r= glob(globpat_backlog, GLOB_ERR|GLOB_MARK|GLOB_NOSORT, 0, &gl);

  switch (r) {
  case GLOB_ABORTED:
    sysdie("failed to expand backlog pattern %s", globpat_backlog);
  case GLOB_NOSPACE:
    die("out of memory expanding backlog pattern %s", globpat_backlog);
  case 0:
    for (i=0; i<gl.gl_pathc; i++) {
      const char *path= gl.gl_pathv[i];

      if (strchr(path,'#') || strchr(path,'~')) {
        debug("backlog file search skipping %s", path);
        continue;
      }
      r= stat(path, &stab);
      if (r) {
        syswarn("failed to stat backlog file %s", path);
        continue;
      }
      if (!S_ISREG(stab.st_mode)) {
        warn("backlog file %s is not a plain file (or link to one)", path);
        continue;
      }
      if (!oldest_path || stab.st_mtime < oldest_mtime) {
        oldest_path= path;
        oldest_mtime= stab.st_mtime;
      }
    }
  case GLOB_NOMATCH: /* fall through */
    break;
  default:
    sysdie("glob expansion of backlog pattern %s gave unexpected"
           " nonzero (error?) return value %d", globpat_backlog, r);
  }

  globfree(&gl);

  if (!oldest_path) {
    debug("backlog scan: none");
    until_backlog_nextscan= backlog_spontaneous_rescan_periods;
    return 0;
  }

  now= time();  if (now==-1) sysdie("time(2) failed");
  double age= difftime(now, oldest_mtime);
  long age_deficiency= (backlog_retry_minperiods * PERIOD_SECONDS) - age;

  if (age_deficiency <= 0) {
    debug("backlog scan: found age=%f deficiency=%ld oldest=%s",
          age, age_deficiency, oldest_path);

    backlog_input_file= open_input_file(oldest_path);
    if (!backlog_input_file) {
      warn("backlog file %s vanished as we opened it", backlog_input_file);
      goto try_again;
    }
    inputfile_tailing_start(backlog_input_file);
    until_backlog_nextscan= -1;
    return 1;
  }

  until_backlog_nextscan= age_deficiency / PERIOD_SECONDS;

  if (backlog_spontaneous_rescan_periods >= 0 &&
      until_backlog_nextscan > backlog_spontaneous_rescan_periods)
    until_backlog_nextscan= backlog_spontaneous_rescan_periods;

  debug("backlog scan: young age=%f deficiency=%ld nextscan=%d oldest=%s",
        age, age_deficiency, until_backlog_nextscan, oldest_path);
  return 2;
}

/*========== flushing the feed ==========*/

static pid_t inndcomm_child;

static void *inndcomm_event(oop_source *lp, int fd, oop_event e, void *u) {
  assert(inndcomm_child);
  int status= xwaitpid(&inndcomm_child, "inndcomm");
  loop->cancel_fd(fd);
  close(fd);

  assert(!flushing_input_file);

  if (WIFEXITED(status)) {
    switch (WEXITSTATUS(status)) {

    case INNDCOMMCHILD_ESTATUS_FAIL:
      goto failed;

    case INNDCOMMCHILD_ESTATUS_NONESUCH:
      warn("feed has been dropped by innd, finishing up");
      flushing_input_file= main_input_file;
      main_input_file= 0;
      SMS(DROPPING, 0, "dropped by innd");
      return OOP_CONTINUE;

    case 0:
      flushing_input_file= main_input_file;
      main_input_file= open_input_file(feedfile);
      if (!main_input_file)
        die("flush succeeded but feedfile %s does not exist!", feedfile);
      until_spontaneous_flush= spontaneous_flush_periods;
      SMS(SEPARATED, 0, "feed file missing");
      return OOP_CONTINUE;

    default:
      goto unexpected_exitstatus;

    }
  } else if (WIFSIGNALED(status) && WTERMSIG(status) == SIGALRM) {
    warn("flush timed out trying to talk to innd");
    goto failed;
  } else {
  unexpected_exitstatus:
    report_child_status("inndcomm child", status);
  }

 failed:
  SMS(FLUSHFAIL, flushfail_retry_periods, "flush failed, will retry");
}

static void inndcommfail(const char *what) {
  syswarn("error communicating with innd: %s failed: %s", what, ICCfailure);
  exit(INNDCOMMCHILD_ESTATUS_FAIL);
}

void spawn_inndcomm_flush(void) {
  int pipefds[2];

  assert(sms==sm_NORMAL || sms==sm_FLUSHFAIL);
  assert(!inndcomm_child);

  if (pipe(pipefds)) sysdie("create pipe for inndcomm child sentinel");

  inndcomm_child= xfork();

  if (!inndcomm_child) {
    static char flushargv[2]= { sitename, 0 };
    char *reply;

    close(pipefds[0]);

    alarm(inndcomm_flush_timeout);
    r= ICCopen();                         if (r)   inndcommfail("connect");
    r= ICCcommand('f',flushargv,&reply);  if (r<0) inndcommfail("transmit");
    if (!r) exit(0); /* yay! */

    if (!strcmp(reply, "1 No such site")) exit(INNDCOMMCHILD_ESTATUS_NONESUCH);
    syswarn("innd ctlinnd flush failed: innd said %s", reply);
    exit(INNDCOMMCHILD_ESTATUS_FAIL);
  }

  close(pipefds[1]);
  int sentinel_fd= pipefds[0];
  on_fd_read_except(sentinel_fd, inndcomm_event);

  SMS(FLUSHING, 0, "flush is in progress");
}

/*========== main program ==========*/

static void postfork_inputfile(InputFile *ipf) {
  if (!ipf) return;
  assert(ipf->fd >= 0);
  close(ipf->fd);
  ipf->fd= -1;
}

static void postfork_conns(Connection *conn) {
  while (conn) {
    close(conn->fd);
    conn= conn->next;
  }
}

static void postfork_stdio(FILE *f) {
  /* we have no stdio streams that are buffered long-term */
  if (f) fclose(f);
}

static void postfork(const char *what) {
  if (signal(SIGPIPE, SIG_DFL) == SIG_ERR)
    sysdie("%s child: failed to reset SIGPIPE");

  postfork_inputfile(main_input_file);
  postfork_inputfile(flushing_input_file);
  postfork_conns(idle.head);
  postfork_conns(working.head);
  postfork_conns(full.head);
  postfork_stdio(defer);
}

#define EVERY(what, interval, body)                                          \
  static const struct timeval what##_timeout = { 5, 0 };                     \
  static void what##_schedule(void);                                         \
  static void *what##_timedout(oop_source *lp, struct timeval tv, void *u) { \
    { body }                                                                 \
    what##_schedule();                                                       \
  }                                                                          \
  static void what##_schedule(void) {                                        \
    loop->on_time(loop, what##_timeout, what##_timedout, 0);                 \
  }

EVERY(filepoll, {5,0}, {
  if (main_input_file && main_input_file->readable_callback)
    filemon_callback(main_input_file);
});

#define DEBUGF_IPF(wh) " " #wh "=%p/%s:ip=%ld,off=%ld,fd=%d%s"  \
#define DEBUG_IPF(sh)                                           \
  wh##_input_file, debug_ipf_path(wh##_input_file),             \
  wh##_input_file->inprogress, (long)wh##_input_file->offset,   \
  wh##_input_file->fd, wh##_input_file->rd ? "+" : ""
static const char *debug_ipf_path(InputFile *ipf) {
  char *slash= strrchr(ipf->path,'/');
  return slash ? slash+1 : ipf->path;
}

EVERY(period, {PERIOD_SECONDS,0}, {
  debug("PERIOD"
        " sms=%s queue=%d sm_period_counter=%d"
        " connect_delay=%d until_spontaneous_flush=%d"
        " input_files" DEBUGF_IPF(main) DEBUGF_IPF(old) DEBUGF_FMT(flushing)
        " conns idle=%d working=%d full=%d"
        " children connecting=%ld inndcomm_child"
        ,
        sms_names[sms], queue.count, sm_period_counter,
        connect_delay, until_spontaneous_flush,
        DEBUG_IPF(main), DEBUG_IPF(flushing), DEBUG_IPF(flushing),
        idle.count, working.count, full.count,
        (long)connecting_child, (long)inndcomm_child
        );
  if (connect_delay) connect_delay--;
  if (until_spontaneous_flush) until_spontaneous_flush--;
  poll_backlog_file();
  if (!backlog_input_file) close_defer(); /* want to start on a new backlog */
  statemc_poll();
  check_master_queue();
});


/*========== option parsing ==========*/

enum OptFlags {
  of_seconds= 001000u;
  of_boolean= 002000u;
};

typedef struct Option Option;
typedef void OptionParser(const Option*, const char *val);

struct Option {
  int short;
  const char *long;
  void *store;
  OptionParser *fn;
  int noarg;
};

void op_integer(const Option *o, const char *val) {
  char *ep;
  errno= 0;
  unsigned long ul= strtoul(val,&ep,10);
  if (*ep || ep==val || errno || ul>INT_MAX)
    badusage("bad integer value for %s",o->long);
  int *store= o->store;
  *store= ul;
}

void op_double(const Option *o, const char *val) {
  int *store= o->store;
  char *ep;
  errno= 0;
  *store= strtod(val, &ep);
  if (*ep || ep==val || errno)
    badusage("bad floating point value for %s",o->long);
}

void op_string(const Option *o, const char *val) {
  char **store= o->store;
  free(*store);
  *store= val;
}

void op_seconds(const Option *o, const char *val) {
  int *store= o->store;
  char *ep;

  double v= strtod(val,&ep);
  if (ep==val) badusage("bad time/duration value for %s",o->long);

  if (!*ep || !strcmp(ep,"s")) unit= 1;
  else if (!strcmp(ep,"m")) unit= 60;
  else if (!strcmp(ep,"h")) unit= 3600;
  else if (!strcmp(ep,"d")) unit= 86400;
  else badusage("bad units %s for time/duration value for %s",ep,o->long);

  v *= unit;
  v= ceil(v);
  if (v > INT_MAX) badusage("time/duration value for %s out of range",o->long);
  *store= v;
}

void op_periods_rndup(const Option *o, const char *val) {
  int *store= o->store;
  op_seconds(o,val);
  *store += PERIOD_SECONDS-1;
  *store /= PERIOD_SECONDS;
}

void op_periods_booltrue(const Option *o, const char *val) {
  int *store= o->store;
  *store= 1;
}
void op_periods_boolfalse(const Option *o, const char *val) {
  int *store= o->store;
  *store= 0;
}

static const Option options[]= {
{ 0, "max-connections",       &max_connections           op_integer          },
{ 0, "streaming",             &try_stream,               op_booltrue,  1     },
{ 0, "no-streaming",          &try_stream,               op_boolfalse, 1     },
{'h',"host",                  &remote_host,              op_string           },
{'P',"port",                  &port                      op_integer          },
{ 0, "inndconf",              &inndconffile,             op_string           },
{'f',"feedfile",              &feedfile,                 op_string           },
{'q',"quiet-multiple",        &quiet_if_locked,          op_booltrue,  1     },
{ 0, "no-quiet-multiple",     &quiet_if_locked,          op_boolfalse, 1     },
{'d',"daemon",                &become_daemon,            op_booltrue,  1     },
{ 0, "no-daemon",             &become_daemon,            op_boolfalse, 1     },

{ 0, "no-check-proportion",   &nocheck_thresh_pct,       op_double           },
{ 0, "no-check-filter",       &nocheck_decay_articles,   op_double           },

{ 0, "max-queue-size",        &max_queue_per_conn        op_integer          },
{ 0, "reconnect-interval",    &reconnect_delay_periods,  op_periods_rndup    },
{ 0, "flush-retry-interval",  &flushfail_retry_periods,  op_periods_rndup    },
{ 0, "feedfile-open-timeout", &open_wait_periods,        op_periods_rndup    },
{ 0, "connection-timeout",    &connection_timeout,       op_seconds          },
{ 0, "inndcomm-timeout",      &inndcomm_flush_timeout,   op_seconds          },
};

int main(int argc, char **argv) {
  const char *arg;

  for (;;) {
    arg= *++argv;
    if (!arg) break;
    if (*arg != '-') break;
    if (!strcmp(arg,"--")) { arg= *++argv; break; }
    int a;
    while ((a= *++arg)) {
      const Option *o;
      if (a=='-') {
        arg++;
        char *equals= strchr(arg,'=');
        int len= equals ? (equals - arg) : strlen(arg);
        for (o=options; o->long; o++)
          if (strlen(o->long) == len && !memcmp(o->long,arg,len))
            goto found_long;
        badusage("unknown long option --%s",arg);
      found_long:
        if (o->noarg) {
          if (equals) badusage("option --%s does not take a value",o->long);
          arg= 0;
        } else if (equals) {
          arg= equals+1;
        } else {
          arg= *++argv;
          if (!arg) badusage("option --%s needs a value",o->long);
        }
        o->fn(o, arg);
        break; /* eaten the whole argument now */
      }
      for (o=options; o->long; o++)
        if (a == o->short)
          goto found_short;
      badusage("unknown short option -%c",a);
    found_short:
      if (o->noarg) {
        o->fn(o,0);
      } else {
        if (!*++arg) {
          arg= *++argv;
          if (!arg) badusage("option -%c needs a value",o->short);
        }
        o->fn(o,arg);
        break; /* eaten the whole argument now */
      }
    }
  }

  if (!arg) badusage("need site name argument");
  sitename= arg;

  if ((arg= *++argv)) {
    if (remote_host) badusage("must not specify -h and also host as argument");
    remote_host= arg;
  }

  if (*++argv) badusage("too many non-option arguments");

  if (nocheck_thresh_pct < 0 || nocheck_thresh_pct > 100)
    badusage("nocheck threshold percentage must be between 0..100");
  nocheck_thresh= nocheck_thresh_pct * 0.01;

  if (nocheck_decay_articles < 0.1)
    badusage("nocheck decay articles must be at least 0.1");
  nocheck_decay= 1 - 1/nocheck_decay_articles;

  innconf_read(inndconffile);

  if (!feedfile)
    feedfile= xasprintf("%s/%s",pathoutgoing,sitename);
  else if (!feedfile[0])
    badusage("feed filename must be nonempty");
  else if (feedfile[strlen(feedfile)-1]=='/')
    feedfile= xasprintf("%s%s",feedfile,sitename);

  const char *feedfile_forbidden= "?*[~#";
  int c;
  while ((c= *feedfile_forbidden++))
    if (strchr(feedfile, c))
      badusage("feed filename may not contain metacharacter %c",c);

  if (signal(SIGPIPE, SIG_IGN) == SIG_ERR)
    sysdie("could not ignore SIGPIPE");

  if (become_daemon) {
    for (i=3; i<255; i++)
      /* do this now before we open syslog, etc. */
      close(i);
    openlog("innduct",LOG_NDELAY|LOG_PID,LOG_NEWS);

    int null= open("/dev/null",O_RDWR);
    if (null<0) sysdie("failed to open /dev/null");
    dup2(null,0);
    dup2(null,1);
    dup2(null,2);
    close(null);

    pid_t child1= xfork("daemonise first fork");
    if (child1) _exit(0);

    pid_t sid= setsid();
    if (sid != child1) sysdie("setsid failed");

    pid_t child2= xfork("daemonise second fork");
    if (child2) _exit(0);
  }

  notice("starting");

  if (!filemon_init()) {
    warn("no file monitoring available, polling");
    filepoll_schedule();
  }

  period_schedule();

  statemc_init();

  loop->execute.
}