chiark / gitweb /
wip compile
[inn-innduct.git] / backends / innduct.c
index ba4c6bd..a5bb448 100644 (file)
  * or might be blanked out
  *    <spc><spc><spc><spc>....
  *
- *   site.name_duct.lock       lock preventing multiple ducts
- *                                holder of this lock is "duct"
  * 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
- * D site.name_duct            temporary feed file during flush (or crash)
+ *   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_duct.defer      431'd articles, still being written,
+ *   site.name_defer           431'd articles, still being written,
  *                                created, written, used by duct
- *   site.name_backlog.lock    lock taken out by innxmit wrapper
- *                                holder and its child are "xmit"
- *   site.name_backlog_<date>.<inum>
- *                             431'd articles, ready for innxmit
+ *
+ *   site.name_backlog.<date>.<inum>
+ *                             431'd articles, ready for innxmit or duct
  *                                created (link/mv) by duct
- *                                read by xmit
- *                                unlinked by xmit
- *   site.name_backlog_<letters> eg
- *   site.name_backlog_manual
+ *   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
- *                                read by xmit
- *                                unlinked by xmit
+ *                                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.
+ * -8<-
 
 
    OVERALL STATES:
 
                                                                START
                                                                   |
-                                                             check D, F
-                                                                  |
-                          <--------------------------------------'|
-        Nothing                            F, D both ENOENT       |
-         F: ENOENT                                                       |
-         D: ENOENT                                                       |
-         duct: not not reading anything                           |
-           |                                                     |
-           |`---------------------.                               |
-           |                      | duct times out waiting for F  |
-           V  innd creates F      | duct exits                    |
-           |                      V                               |
-        Noduct                    GO TO Dropped                   |
-         F: innd writing                                         |
-         D: ENOENT                                               |
-         duct: not running or not reading anything                |
-           |                                                     |
-           |                                                     |
-     ,-->--+                   <---------------------------------'|
-     |     |  duct opens F                         F exists       |
-     |     |                                              D ENOENT       |
+     ,-->--.                                                 check F, D
+     |     |                                                      |
+     |     |                                                      |
+     |     |  <----------------<---------------------------------'|
+     |     |                                       F exists       |
+     |     |                                       D ENOENT       |
+     |     |  duct opens F                                        |
      |     V                                                     |
      |  Normal                                                    |
      |   F: innd writing, duct reading                            |
      |   F == D: innd writing, duct reading         both exist    |
      ^     |                                                      |
      |     |  duct unlinks F                                             |
-     |     V                                                     |
-     |  Moved                               <----+------------<--'|
-     |   F: ENOENT                               |  F ENOENT      |
-     |   D: innd writing, duct reading           |  D exists      |
-     |     |                                     |               |
-     |     |  duct requests flush of feed        |               |
-     |     |   (others can too, harmlessly)      |               |
-     |     V                                    |                |
-     |  Flushing                                |                |
-     |   F: ENOENT                              |                |
-     |   D: innd flushing, duct reading                 |                |
-     |     |                                            |                |
-     |     |   inndcomm flush fails                     |                |
-     |     |`-------------------------->---------'               |
-     |     |                                                             |
+     |     |                        <-----------<-------------<--'|
+     |     |                           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                             both exist     /
+     |   D: duct reading; or ENOENT                  both exist     /
      |     |                                                       /
      |     |  duct gets to the end of D                           /
      |     |  duct opens F too                                          /
    "duct reading" means innduct is reading the file but also
    overwriting processed tokens.
 
+ * ->8- -^L-
  *
+ * rune for printing diagrams:
+
+perl -ne 'print if m/-8\<-/..m/-\>8-/; print "\f" if m/-\^L-/' backends/innduct.c |a2ps -R -B -ops
+
  *
  */
 
+/*============================== PROGRAM ==============================*/
+
+#define _GNU_SOURCE
+
+#include "config.h"
+#include "storage.h"
+#include "nntp.h"
+
+#include <sys/uio.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <sys/stat.h>
+#include <sys/socket.h>
+#include <unistd.h>
+#include <string.h>
+#include <signal.h>
+#include <stdio.h>
+#include <errno.h>
+#include <syslog.h>
+#include <fcntl.h>
+
+#include <oop.h>
+#include <oop-read.h>
+
+/*----- general definitions, probably best not changed -----*/
+
 #define PERIOD_SECONDS 60
 
-static char *feedfile;
-static int max_connections, max_queue_per_conn;
-static int connection_setup_timeout, port, try_stream;
-static const char *remote_host;
+#define CONNCHILD_ESTATUS_STREAM   4
+#define CONNCHILD_ESTATUS_NOSTREAM 5
 
-static double accept_proportion;
-static double nocheck_thresh= 0.95;
-static double nocheck_decay= 1-1/100;
-static int nocheck, nocheck_reported;
+#define INNDCOMMCHILD_ESTATUS_FAIL     6
+#define INNDCOMMCHILD_ESTATUS_NONESUCH 7
+
+/*----- doubly linked lists -----*/
 
-#define ISNODE(T)    T *next, *back;
-#define LIST(T)      struct { T *head, *tail, *tailpred; int count; }
+#define ISNODE(T)   struct { T *succ, *pred; } node
+#define DEFLIST(T)  typedef struct { T *head, *tail, *tp; int count; } T##List
 
-#define NODE(n) ((struct node*)&(n)->head)
+#define NODE(n) ((struct node*)&(n)->node)
 
 #define LIST_ADDHEAD(l,n)                                              \
  (list_addhead((struct list*)&(l), NODE((n))), (void)(l).count++)
@@ -165,43 +200,89 @@ static int nocheck, nocheck_reported;
 #define LIST_INSERT(l,n,pred) \
  (list_insert((struct list*)&(l), NODE((n)), NODE((pred))), (void)(l).count++)
 
+/*----- type predeclarations -----*/
+
+typedef struct Conn Conn;
+typedef struct Article Article;
+typedef enum StateMachineState StateMachineState;
+
+DEFLIST(Conn);
+DEFLIST(Article);
+
+/*----- 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 conn_assign_one_article(ConnList *connlist, Conn **last_assigned);
+static void statemc_setstate(StateMachineState newsms, int periods,
+                            const char *forlog, const char *why);
+static void check_master_queue(void);
+
+/*----- configuration options -----*/
+
+static char *sitename, *feedfile;
+static const char *remote_host;
+static int quiet_multiple=0, become_daemon=1;
+
+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;
+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;
+
 
-#define RESULT_COUNTS                          \
-  RC(offered)                                  \
-  RC(sent)                                     \
-  RC(unwanted)                                 \
-  RC(accepted)                                 \
-  RC(rejected)                                 \
-  RC(deferred)
+/*----- statistics -----*/
+
+typedef enum {      /* in queue                 in conn->sent             */
+  art_Unchecked,    /*   not checked, not sent    checking                */
+  art_Wanted,       /*   checked, wanted          sent body as requested  */
+  art_Unsolicited,  /*   -                        sent body without check */
+  art_MaxState
+} ArtState;
+
+#define RESULT_COUNTS(RCS,RCN)                 \
+  RCS(sent)                                    \
+  RCS(accepted)                                        \
+  RCN(unwanted)                                        \
+  RCN(rejected)                                        \
+  RCN(deferred)                                        \
+  RCN(connretry)
+
+#define RCI_TRIPLE_FMT_BASE "%d(id%d+bd%d+nc%d)"
+#define RCI_TRIPLE_VALS_BASE(counts,x)         \
+       , counts[art_Unchecked] x               \
+       + counts[art_Wanted] x                  \
+       + counts[art_Unsolicited] x,            \
+       counts[art_Unchecked] x                 \
+       , counts[art_Wanted] x                  \
+       , counts[art_Unsolicited] x
 
 typedef enum {
-#define RC_INDEX(x) RCI_##x
-  RESULT_COUNTS
+#define RC_INDEX(x) RCI_##x,
+  RESULT_COUNTS(RC_INDEX, RC_INDEX)
   RCI_max
 } ResultCountIndex;
 
-typedef struct {
-  int articles[2 /* checked */][RCI_max];
-} Counts;
 
-struct Article {
-  int midlen;
-  int checked, sentbody;
-  InputFile *ipf;
-  TOKEN token;
-  off_t offset;
-  int blanklen;
-  char messageid[1];
-};
+/*----- transmission buffers -----*/
 
 #define CONNIOVS 128
 
-#define CN "<%d> "
-
-typedef struct Conn Conn;
-
 typedef enum {
-  xk_Malloc, xk_Const, xk_Artdata;
+  xk_Malloc, xk_Const, xk_Artdata
 } XmitKind;
 
 typedef struct {
@@ -212,38 +293,172 @@ typedef struct {
   } info;
 } XmitDetails;
 
+
+/*----- core operational data structure types -----*/
+
+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;
+
+  int counts[art_MaxState][RCI_max];
+  char path[];
+} InputFile;
+
+struct Article {
+  ISNODE(Article);
+  ArtState state;
+  int midlen;
+  InputFile *ipf;
+  TOKEN token;
+  off_t offset;
+  int blanklen;
+  char messageid[1];
+};
+
+#define SMS_LIST(X)                            \
+  X(NORMAL)                                    \
+  X(FLUSHING)                                  \
+  X(FLUSHFAILED)                               \
+  X(SEPARATED)                                 \
+  X(DROPPING)                                  \
+  X(DROPPED)
+
+enum StateMachineState {
+#define SMS_DEF_ENUM(s) sm_##s,
+  SMS_LIST(SMS_DEF_ENUM)
+};
+
+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 */
+  ArticleList queue; /* not yet told peer, or CHECK said send it */
+  ArticleList sent; /* offered/transmitted - in xmit or waiting reply */
   struct iovec xmit[CONNIOVS];
   XmitDetails xmitd[CONNIOVS];
   int xmitu;
 };
 
+
+/*----- operational variables -----*/
+
+static oop_source *loop;
+
+static int nconns, until_connect;
+static ConnList idle, working, full;
+static ArticleList 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;
 
-static int filemon_init(void);
-static void filemon_setfile(int mainfeed_fd, const char *mainfeed_path);
-static void filemon_callback(void);
 
+/*========== logging ==========*/
 
-#define CHILD_ESTATUS_STREAM   4
-#define CHILD_ESTATUS_NOSTREAM 5
+static void logcore(int sysloglevel, const char *fmt, ...)
+     __attribute__((__format__(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);
+  }
+  va_end(al);
+}
+
+static void logv(int sysloglevel, const char *pfx, int errnoval,
+                int exitstatus, const char *fmt, va_list al)
+     __attribute__((__format__(printf,5,0)));
+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, err, estatus)    \
+  static void fn(const char *fmt, ...)                 \
+      __attribute__((__format__(printf,1,2)));         \
+  static void fn(const char *fmt, ...) {               \
+    va_list al;                                                \
+    va_start(al,fmt);                                  \
+    logv(sysloglevel, pfx, err, 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_WARNING, errno, 0);
+logwrap(warn,     " warning",  LOG_WARNING, -1,    0);
+
+logwrap(notice,   "",          LOG_NOTICE,  -1,    0);
+logwrap(info,     " info",     LOG_INFO,    -1,    0);
+logwrap(debug,    " debug",    LOG_DEBUG,   -1,    0);
 
-static int since_connect_attempt;
-static int nconns;
-static LIST(Conn) idle, working, full;
 
-static LIST(Article) *queue;
+/*========== utility functions etc. ==========*/
 
 static void perhaps_close(int *fd) { if (*fd) { close(*fd); fd=0; } }
 
-/*========== making new connections ==========*/
+static void *xmalloc(size_t sz) {
+  if (!sz) return 0;
+  void *r= malloc(sz);
+  if (r) return r;
+  sysdie("malloc (%ld bytes) failed", (unsigned long)sz);
+}
 
-static int connecting_sockets[2]= {-1,-1};
-static pid_t connecting_child;
+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);
+  debug("forked %s %ld", what, (unsigned long)child);
+  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)) {
@@ -264,27 +479,89 @@ static void report_child_status(const char *what, int status) {
   }
 }
 
-static void connect_attempt_discard(void) {
-  if (connecting_sockets[0]) {
-    cancel_fd(loop, connecting_sockets[0], OOP_READ);
-    cancel_fd(loop, connecting_sockets[0], OOP_EXCEPTION);
+static int xwaitpid(pid_t *pid, const char *what) {
+  int status;
+
+  int 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(fd, stab_r);
+  if (r) sysdie("could not fstat %s", what);
+}
+
+static void xfstat_isreg(int fd, struct stat *stab_r,
+                        const char *path, const char *what) {
+  xfstat(fd, stab_r, what);
+  check_isreg(stab_r, 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 void setnonblock(int fd, int nonblocking) {
+  int r= fcntl(fd, F_GETFL);  if (r<0) sysdie("setnonblocking fcntl F_GETFL");
+  if (nonblocking) r |= O_NONBLOCK;
+  else r &= ~O_NONBLOCK;
+  r= fcntl(fd, F_SETFL, r);  if (r<0) sysdie("setnonblocking fcntl F_SETFL");
+}
+
+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);
+}
+
+/*========== 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;
-    r= kill(connecting_child, SIGKILL);
-    if (r) sysdie("cannot kill connect child");
-
-    pid_t got= waitpid(connecting_child, &status, WNOHANG);
-    if (got==-1) sysdie("cannot reap connect child");
+    int r= kill(connecting_child, SIGTERM);
+    if (r) syswarn("failed to kill connecting child");
+    int status= xwaitpid(&connecting_child, "connect");
 
     if (!(WIFEXITED(status) ||
-         (WIFSIGNALED(status) && WTERMSIG(status) == SIGKILL))) {
-      report_child_status("connect"
-    }
-    connecting_child= 0;
+         (WIFSIGNALED(status) && WTERMSIG(status) == SIGKILL)))
+      report_child_status("connect", status);
   }
 }
 
@@ -298,9 +575,10 @@ static void connect_attempt_discard(void) {
 static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
   Conn *conn= 0;
 
-  conn= xcalloc(sizeof(*conn));
+  conn= xmalloc(sizeof(*conn));
+  memset(conn,0,sizeof(*conn));
 
-  DECL_MSG_CMSG(msg);
+  PREP_DECL_MSG_CMSG(msg);
   struct cmsghdr *h= 0;
   ssize_t rs= recvmsg(fd, &msg, MSG_DONTWAIT);
   if (rs >= 0) h= CMSG_FIRSTHDR(&msg);
@@ -311,11 +589,11 @@ static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
       assert(got==connecting_child);
       connecting_child= 0;
       if (WIFEXITED(status) &&
-         (WEXITSTATUS(status) != 0
-          WEXITSTATUS(status) != CHILD_ESTATUS_STREAM &&
-          WEXITSTATUS(status) != CHILD_ESTATUS_NOSTREAM)) {
+         (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) {
+      } else if (WIFSIGNALED(status) && WTERMSIG(status) == SIGALRM) {
        warn("connect: connection attempt timed out");
       } else if (!WIFEXITED(status)) {
        report_child_status("connect", status);
@@ -325,7 +603,7 @@ static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
       /* child is still running apparently, report the socket problem */
       if (rs < 0)
        syswarn("connect: read from child socket failed");
-      else if (e == OOP_EXCEPTIONN)
+      else if (e == OOP_EXCEPTION)
        warn("connect: unexpected exception on child socket");
       else
        warn("connect: unexpected EOF on child socket");
@@ -340,13 +618,14 @@ static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
   }
   CHK(level, SOL_SOCKET);
   CHK(type,  SCM_RIGHTS);
-  CHK(len,   CMSG_LEN(sizeof(conn-b>fd)));
+  CHK(len,   CMSG_LEN(sizeof(conn->fd)));
 #undef CHK
 
-  if (CMSG_NXTHDR,&msg,h) { die("connect: child sent many cmsgs"); goto x; }
+  if (CMSG_NXTHDR(&msg,h)) { die("connect: child sent many cmsgs"); goto x; }
 
   memcpy(&conn->fd, CMSG_DATA(h), sizeof(conn->fd));
 
+  int status;
   pid_t got= waitpid(connecting_child, &status, 0);
   if (got==-1) sysdie("connect: real wait for child");
   assert(got == connecting_child);
@@ -355,17 +634,17 @@ static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
   if (!WIFEXITED(status)) { report_child_status("connect",status); goto x; }
   int es= WEXITSTATUS(status);
   switch (es) {
-  case CHILD_ESTATUS_STREAM:    conn->stream= 1;   break;
-  case CHILD_ESTATUS_NOSTREAM:  conn->stream= 0;   break;
+  case CONNCHILD_ESTATUS_STREAM:    conn->stream= 1;   break;
+  case CONNCHILD_ESTATUS_NOSTREAM:  conn->stream= 0;   break;
   default:
-    die("connect: child gave unexpected exit status %d", es);
+    fatal("connect: child gave unexpected exit status %d", es);
   }
 
-  set nonblocking;
+  setnonblocking(conn->fd, 1);
 
   /* Phew! */
   LIST_ADDHEAD(idle, conn);
-  notice(CN "connected %s", conn->fd, conn->stream ? "streaming" : "plain");
+  notice("#%d connected %s", conn->fd, conn->stream ? "streaming" : "plain");
   connect_attempt_discard();
   check_master_queue();
   return 0;
@@ -378,26 +657,22 @@ static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
   connect_attempt_discard();
 }
 
-static void connect_start() {
+static void connect_start(void) {
   assert(!connecting_sockets[0]);
   assert(!connecting_sockets[1]);
   assert(!connecting_child);
 
   notice("starting connection attempt");
 
-  r= socketpair(AF_UNIX, SOCK_STREAM, 0, connecting_sockets);
+  int r= socketpair(AF_UNIX, SOCK_STREAM, 0, connecting_sockets);
   if (r) { syswarn("connect: cannot create socketpair for child"); goto x; }
 
-  connecting_child= fork();
-  if (connecting_child==-1) { syswarn("connect: cannot fork"); goto x; }
+  connecting_child= xfork("connection");
 
   if (!connecting_child) {
     FILE *cn_from, *cn_to;
     char buf[NNTP_STRLEN+100];
-    int exitstatus= CHILD_ESTATUS_NOSTREAM;
-
-    put sigpipe back;
-    close unwanted fds;
+    int exitstatus= CONNCHILD_ESTATUS_NOSTREAM;
 
     r= close(connecting_sockets[0]);
     if (r) sysdie("connect: close parent socket in child");
@@ -406,42 +681,42 @@ static void connect_start() {
     if (NNTPconnect(remote_host, port, &cn_from, &cn_to, buf) < 0) {
       if (buf[0]) {
        sanitise_inplace(buf);
-       die("connect: rejected: %s", buf);
+       fatal("connect: rejected: %s", buf);
       } else {
-       sysdie("connect: connection attempt failed");
+       sysfatal("connect: connection attempt failed");
       }
     }
     if (NNTPsendpassword(remote_host, cn_from, cn_to) < 0)
-      sysdie("connect: authentication failed");
+      sysfatal("connect: authentication failed");
     if (try_stream) {
       if (fputs("MODE STREAM\r\n", cn_to) ||
          fflush(cn_to))
-       sysdie("connect: could not send MODE STREAM");
+       sysfatal("connect: could not send MODE STREAM");
       buf[sizeof(buf)-1]= 0;
       if (!fgets(buf, sizeof(buf)-1, cn_from)) {
        if (ferror(cn_from))
-         sysdie("connect: could not read response to MODE STREAM");
+         sysfatal("connect: could not read response to MODE STREAM");
        else
-         die("connect: connection close in response to MODE STREAM");
+         fatal("connect: connection close in response to MODE STREAM");
       }
       int l= strlen(buf);
       assert(l>=1);
       if (buf[-1]!='\n') {
        sanitise_inplace(buf);
-       die("connect: response to MODE STREAM is too long: %.100s...",
+       fatal("connect: response to MODE STREAM is too long: %.100s...",
            remote_host, buf);
       }
-      l--;  if (l>0 && buf[1-]=='\r') l--;
+      l--;  if (l>0 && buf[l-1]=='\r') l--;
       buf[l]= 0;
       char *ep;
       int rcode= strtoul(buf,&ep,10);
-      if (ep != buf[3]) {
+      if (ep != &buf[3]) {
        sanitise_inplace(buf);
-       die("connect: bad response to MODE STREAM: %.50s", buf);
+       fatal("connect: bad response to MODE STREAM: %.50s", buf);
       }
       switch (rcode) {
       case 203:
-       exitstatus= CHILD_ESTATUS_STREAM;
+       exitstatus= CONNCHILD_ESTATUS_STREAM;
        break;
       case 480:
       case 500:
@@ -472,21 +747,17 @@ static void connect_start() {
   r= close(connecting_sockets[1]);  connecting_sockets[1]= 0;
   if (r) sysdie("connect: close child socket in parent");
 
-  loop->on_fd(loop, connecting_sockets[0], OOP_READ,      connchild_event, 0);
-  loop->on_fd(loop, connecting_sockets[0], OOP_EXCEPTION, connchild_event, 0);
-  return OOP_CONTINUE;
+  on_fd_read_except(connecting_sockets[0], connchild_event);
+  return;
 
  x:
   connect_attempt_discard();
 }
 
-/*========== overall control of article flow ==========*/
 
-static void conn_check_work(Conn *conn);
+/*========== overall control of article flow ==========*/
 
 static void check_master_queue(void) {
-  try reading current feed file;
-
   if (!queue.count)
     return;
 
@@ -496,9 +767,9 @@ static void check_master_queue(void) {
       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;
+    } else if (nconns < max_connections && queue.count >= max_queue_per_conn &&
+              !connecting_child && !until_connect) {
+      until_connect= reconnect_delay_periods;
       connect_start();
     } else {
       break;
@@ -507,8 +778,7 @@ static void check_master_queue(void) {
   conn_check_work(last_assigned);
 }
 
-static void conn_assign_one_article(LIST(Conn) *connlist,
-                                   Conn **last_assigned) {
+static void conn_assign_one_article(ConnList *connlist, Conn **last_assigned) {
   Conn *conn= connlist->head;
 
   LIST_REMOVE(*connlist, conn);
@@ -528,7 +798,7 @@ static int conn_total_queued_articles(Conn *conn) {
   return conn->sent.count + conn->queue.count;
 }
 
-static LIST(Conn) *conn_determine_right_list(Conn *conn) {
+static ConnList *conn_determine_right_list(Conn *conn) {
   int inqueue= conn_total_queued_articles(conn);
   assert(inqueue <= max_queue);
   if (inqueue == 0) return &idle;
@@ -564,8 +834,47 @@ static void conn_check_work(Conn *conn)  {
   }
 }
 
-/*========== article transmission ==========*/
+static void vconnfail(Conn *conn, const char *fmt, va_list al)
+     __attribute__((printf,2,0));
+
+static void vconnfail(Conn *conn, const char *fmt, va_list al) {
+  int requeue[art_MaxState];
+
+  Article *art;
+  while ((art= LIST_REMHEAD(conn->queue))) LIST_ADDTAIL(queue);
+  while ((art= LIST_REMHEAD(conn->sent))) {
+    counts[art->state]++;
+    if (art->state==art_Unsolicited) art->state= art_Unchecked;
+    LIST_ADDTAIL(queue);
+  }
+
+  int i;
+  XmitDetails *xd;
+  for (i=0, dp=&conn->xmitd; i<conn->xmitu; i++, dp++)
+    xmit_free(dp);
+
+  char *m= xvasprintf(fmt,al);
+  warn("#%d connection failed, requeueing " RCI_TRIPLE_FMT_BASE ": %s",
+       conn->fd, RCI_TRIPLE_FMT_VALS(requeue, /*nothing*/), m);
+  free(m);
+
+  close(conn->fd);
+  free(conn);
 
+  until_connect= reconnect_delay_periods;
+  check_master_queue();
+}
+
+static void connfail(Connection *conn, const char *fmt, ...)
+     __attribute__((printf,2,3));
+static void connfail(Connection *conn, const char *fmt, ...) {
+  va_list al;
+  va_start(al,fmt);
+  vconnfail(fmt,al);
+  va_end(al);
+}
+
+/*========== article transmission ==========*/
 
 static XmitDetails *xmit_core(Conn *conn, const char *data, int len,
                   XmitKind kind) { /* caller must then fill in details */
@@ -610,8 +919,7 @@ static void *conn_write_some_xmits(Conn *conn) {
     ssize_t rs= writev(conn->fd, conn->xmit, count);
     if (rs < 0) {
       if (errno == EAGAIN) return OOP_CONTINUE;
-      syswarn(CN "write failed", conn->fd);
-      conn_failed(conn);
+      connfail(conn, "write failed: %s", strerror(errno));
       return OOP_HALT;
     }
     assert(rs > 0);
@@ -642,7 +950,7 @@ static void conn_make_some_xmits(Conn *conn) {
     Article *art= LIST_REMHEAD(queue);
     if (!art) break;
 
-    if (art->checked || (conn->stream && nocheck)) {
+    if (art->state >= art_Wanted || (conn->stream && nocheck)) {
       /* actually send it */
 
       ARTHANDLE *artdata= SMretrieve();
@@ -663,11 +971,13 @@ static void conn_make_some_xmits(Conn *conn) {
        }
       }
 
-      art->sent= 1;
+      art->state=
+       art->state == art_Unchecked ? art_Unsolicited :
+       art->state == art_Wanted    ? art_Wanted      :
+       abort();
+      art->ipf->counts[art->state].sent++;
       LIST_ADDTAIL(conn->sent, art);
 
-      counts[art->checked].sent++;
-
     } else {
       /* check it */
 
@@ -678,13 +988,15 @@ static void conn_make_some_xmits(Conn *conn) {
       xmit_noalloc(art->mid, art->midlen);
       XMIT_LITERAL("\r\n");
 
+      assert(art->state == art_Unchecked);
+      art->ipf->counts[art->state].sent++;
       LIST_ADDTAIL(conn->sent, art);
-      counts[art->checked].offered++;
     }
   }
 }
 
-/*========== responses from peer ==========*/
+
+/*========== handling responses from peer ==========*/
 
 static const oop_rd_style peer_rd_style= {
   OOP_RD_DELIM_STRIP, '\n',
@@ -692,66 +1004,89 @@ static const oop_rd_style peer_rd_style= {
   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,
+                                   int must_have_sent
+                                       /* 1:yes, -1:no, 0:dontcare */,
                                    const char *sanitised_response) {
-  Article *art= LIST_REMHEAD(conn->sent);
+  Article *art= conn->sent.head;
 
   if (!art) {
-    warn("peer gave unexpected response when no commands outstanding: %s",
-        sanitised_response);
-    goto failed;
+    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) {
-      warn("peer gave streaming response code "
-          " to IHAVE or subsequent body: %s", sanitised_response);
-      goto failed;
+      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]!='>') {
-      warn("peer gave streaming response with syntactically invalid"
-          " messageid: %s", sanitised_response);
-      goto failed;
+      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)) {
-      peer("peer gave streaming response code to wrong article -"
-          " probable synchronisation problem; we offered: %s; peer said: %s",
-          art->messageid, sanitised_response);
-      goto failed;
+      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) {
-      warn("peer gave non-streaming response code to CHECK/TAKETHIS: %s",
-          sanitised_response);
-      goto failed;
+      connfail("peer gave non-streaming response code to CHECK/TAKETHIS: %s",
+              sanitised_response);
+      return 0;
     }
   }
 
-  return art;
+  if (must_have_sent>0 && art->state < art_Wanted) {
+    connfail("peer says article accepted but we had not sent the body: %s",
+            sanitised_response);
+    return 0;
+  }
+  if (must_have_sent<0 && art->state >= art_Wanted) {
+    connfail("peer says please sent the article but we just did: %s",
+            sanitised_response);
+    return 0;
+  }
 
- failed:
-  conn_failed(conn);
-  return 0;
+  Article *art_again= LIST_REMHEAD(conn->sent);
+  assert(art_again == art);
+  return art;
 }
 
 static void update_nocheck(int accepted) {
   accept_proportion *= accept_decay;
   accept_proportion += accepted;
-  nocheck= accept_proportion >= nocheck_thresh;
-  if (nocheck && !nocheck_reported) {
+  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 != nocheck) {
+    debug("nocheck mode %s", new_nocheck ? "start" : "stop");
   }
+  nocheck= new_nocheck;
 }
 
 static void article_done(Connection *conn, Article *art, int whichcount) {
-  *count++;
-  counts.articles[art->checked][whichcount]++;
+  art->ipf->counts[art->state][whichcount]++;
   if (whichcount == RC_accepted) update_nocheck(1);
   else if (whichcount == RC_unwanted) update_nocheck(0);
 
@@ -777,6 +1112,9 @@ static void article_done(Connection *conn, Article *art, int whichcount) {
   ipf->inprogress--;
   assert(ipf->inprogress >= 0);
 
+  if (!ipf->inprogress && ipf != main_input_file)
+    queue_check_input_done();
+
   free(art);
 }
 
@@ -786,9 +1124,8 @@ static void *peer_rd_ok(oop_source *lp, oop_read *oread, oop_event ev,
   Conn *conn= conn_v;
 
   if (ev == OOP_RD_EOF) {
-    warn("unexpected EOF from peer");
-    conn_failed(conn);
-    return;
+    connfail(conn, "unexpected EOF from peer");
+    return OOP_CONTINUE;
   }
   assert(ev == OOP_RD_OK);
 
@@ -807,16 +1144,14 @@ static void *peer_rd_ok(oop_source *lp, oop_read *oread, oop_event ev,
       sprintf(q,"\\x%02x",c);
       q += 4;
     }
-    warn("badly formatted response from peer: %s", sanibuf);
-    conn_failed(conn);
-    return;
+    connfail(conn, "badly formatted response from peer: %s", sanibuf);
+    return OOP_CONTINUE;
   }
 
   if (conn->quitting) {
     if (code!=205) {
-      warn("peer gave failure response to QUIT: %s", sani);
-      conn_failed(conn);
-      return;
+      connfail(conn, "peer gave failure response to QUIT: %s", sani);
+      return OOP_CONTINUE;
     }
     conn close ok;
     return;
@@ -824,127 +1159,100 @@ static void *peer_rd_ok(oop_source *lp, oop_read *oread, oop_event ev,
 
   Article *art;
 
-#define GET_ARTICLE                                                        \
-  art= article_reply_check(conn, data, code_streaming, sani);              \
+#define GET_ARTICLE(musthavesent)                                          \
+  art= article_reply_check(conn, data, musthavesent, code_streaming, sani); \
   if (art) ; else return OOP_CONTINUE /* reply_check has failed the conn */
 
-#define ARTICLE_DEALTWITH(streaming,how)                       \
+#define ARTICLE_DEALTWITH(streaming,musthavesent,how)          \
   code_streaming= (streaming)                                  \
-  GET_ARTICLE;                                                 \
+  GET_ARTICLE(musthavesent);                                   \
   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 435: ARTICLE_DEALTWITH(0,unwanted); /* IHAVE says they have it */
-  case 438: ARTICLE_DEALTWITH(1,unwanted); /* CHECK/TAKETHIS: they have it */
+  case 400: PEERBADMSG("peer stopped accepting articles");
+  case 503: PEERBADMSG("peer timed us out");
+  default:  PEERBADMSG("peer sent unexpected message");
+
+  case 435: ARTICLE_DEALTWITH(0,0,unwanted); /* IHAVE says they have it */
+  case 438: ARTICLE_DEALTWITH(1,0,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 235: ARTICLE_DEALTWITH(0,1,accepted); /* IHAVE says thanks */
+  case 239: ARTICLE_DEALTWITH(1,1,accepted); /* TAKETHIS says thanks */
 
-  case 437: ARTICLE_DEALTWITH(0,rejected); /* IHAVE says rejected */
-  case 439: ARTICLE_DEALTWITH(1,rejected); /* TAKETHIS says rejected */
+  case 437: ARTICLE_DEALTWITH(0,0,rejected); /* IHAVE says rejected */
+  case 439: ARTICLE_DEALTWITH(1,0,rejected); /* TAKETHIS says rejected */
 
   case 238: /* CHECK says send it */
     code_streaming= 1;
   case 335: /* IHAVE says send it */
-    GET_ARTICLE;
-    count_checkedwanted++;
+    GET_ARTICLE(-1);
+    assert(art->state == art_Unchecked);
+    art->ipf->counts[art->state].accepted++;
+    art->state= art_Wanted;
     LIST_ADDTAIL(conn->queue);
-    if (art->checked) {
-      warn("peer gave %d response to article body",code);
-      goto failed;
-    }
-    art->checked= 1;
     break;
 
   case 431: /* CHECK or TAKETHIS says try later */
     code_streaming= 1;
   case 436: /* IHAVE says try later */
-    GET_ARTICLE;
+    GET_ARTICLE(0);
+    open_defer();
     if (fprintf(defer, "%s %s\n", TokenToText(art->token), art->messageid) <0
        || fflush(defer))
-      sysdie("write to defer file %s",path_ductdefer);
+      sysfatal("write to defer file %s",path_defer);
     article_done(conn, art, RC_deferred);
     break;
 
-  case 400: warn("peer has stopped accepting articles: %s", sani); goto failed;
-  case 503: warn("peer timed us out: %s", sani);                   goto failed;
-  default:  warn("peer sent unexpected message: %s", sani);        goto failed;
-
-  failed:
-    conn_failed(conn);
-    return OOP_CONTINUE;;
   }
 
   check_check_work(conn);
   return OOP_CONTINUE;
 }
 
-/*========== monitoring of input file ==========*/
+
+/*========== monitoring of input files ==========*/
 
 static void feedfile_eof(InputFile *ipf) {
   assert(ipf != main_input_file); /* promised by tailing_try_read */
-  assert(ipf == old_input_file);
-  assert(sms == sm_SEPARATED);
-  sms= sm_FINISHING;
-  inputfile_tailing_stop(ipf);
-  inputfile_tailing_start(main_input_file);
-}
-
-static void statmc_finishdone(void) {
-  time_t now;
-  struct stat stab;
-
-  assert(sms == sm_FINISHING);
-
-  r= fstat(fileno(defer), &stab);
-  if (r) sysdie("check defer file %s", path_defer);
-
-  if (fclose(defer)) sysdie("could not close defer file %s", path_defer);
-  defer= 0;
-
-  now= time(0);
-  if (now==-1) sysdie("could not get current time for backlog filename");
-
-  char *backlog= xasprintf("%s_backlog_%lu.%lu", feedfile,
-                          (unsigned long)now,
-                          (unsigned long)stab.st_ino);
-  if (link(path_defer, path_backlog))
-    sysdie("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);
-  open_defer();
-
-  close_input_file(old_input_file);
-  old_input_file= 0;
 
-  if (unlink(path_duct))
-    sysdie("could not unlink old duct file %s", path_duct);
-
-  sms= sm_NORMAL;
+  inputfile_tailing_stop(ipf);
+  assert(ipf->fd >= 0);
+  if (close(ipf->fd)) sysdie("could not close input file %s", ipf->path);
+  ipf->fd= -1;
+
+  if (ipf == flushing_input_file) {
+    assert(sms==sm_SEPARATED || sms==sm_DROPPING);
+    if (main_input_file) inputfile_tailing_start(main_input_file);
+    statemc_check_flushing_done();
+  } else if (ipf == backlog_input_file) {
+    statemc_check_backlog_done();
+  } else {
+    abort(); /* supposed to wait rather than get EOF on 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;
-    sysdie("unable to open input file %s", path);
+    sysfatal("unable to open input file %s", path);
   }
 
-  InputFile *ipf= xmalloc(sizeof(InputFile));
+  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;
-  ipf->path= path;
-  
+  strcpy(ipf->path, path);
+
   return ipf;
 }
 
@@ -954,10 +1262,11 @@ static void close_input_file(InputFile *ipf) {
   assert(!ipf->rd); /* must have had inputfile_tailing_stop */
   assert(!ipf->inprogress); /* no dangling pointers pointing here */
 
-  if (close(ipf->fd)) sysdie("could not close input file %s", ipf->path);
-  free(ipf);
+  if (ipf->fd >= 0)
+    if (close(ipf->fd)) sysdie("could not close input file %s", ipf->path);
 }
 
+
 /*---------- dealing with articles read in the input file ----------*/
 
 typedef void *feedfile_got_article(oop_source *lp, oop_read *rd,
@@ -986,7 +1295,7 @@ typedef void *feedfile_got_article(oop_source *lp, oop_read *rd,
     art->offset= ipf->offset;
     art->blanklen= recsz;
     art->midlen= midlen;
-    art->checked= art->sentbody= 0;
+    art->state= art_Unchecked;
     art->ipf= ipf;  ipf->inprogress++;
     art->token= TextToToken(tokentextbuf);
     strcpy(art->messageid, space+1);
@@ -994,22 +1303,33 @@ typedef void *feedfile_got_article(oop_source *lp, oop_read *rd,
   }
   ipf->offset += recsz + 1;
 
-  if (sms==sm_NORMAL && ipf->offset >= flush_threshold) {
-    int r= link(feedfile, duct_path);
-    if (r) sysdie("link feedfile %s to ductfile %s", feedfile, dut_path);
-    /* => Hardlinked */
+  if (sms==sm_NORMAL && ipf==main_input_file &&
+      ipf->offset >= flush_threshold)
+    statemc_start_flush("feed file size");
+
+  check_master_queue();
+}
+
+static void statemc_start_flush(const char *why) { /* Normal => Flushing */
+  assert(sms == sm_NORMAL);
 
-    r= unlink(feedfile);
-    if (r) sysdie("unlink old feedfile link %s", feedfile);
-    /* => Moved */
+  debug("starting flush (%s) (%lu >= %lu) (%d)",
+       why,
+       (unsigned long)ipf->offset, (unsigned long)flush_threshold,
+       sm_period_counter);
 
-    spawn_inndcomm_flush(); /* => Flushing, sets sms to sm_FLUSHING */
-  }
+  int r= link(feedfile, duct_path);
+  if (r) sysdie("link feedfile %s to flushing file %s", feedfile,
+               path_duct);
+  /* => Hardlinked */
 
-  check_master_queue();
+  xunlink(feedfile, "old feedfile link");
+  /* => Moved */
+
+  spawn_inndcomm_flush(why); /* => Flushing FLUSHING */
 }
 
-/*---------- tailing input file ----------*/
+/*========== tailing input file ==========*/
 
 static void filemon_start(InputFile *ipf) {
   assert(!ipf->filemon);
@@ -1044,6 +1364,12 @@ static void on_cancel(struct oop_readable *rable) {
   ipf->readable_callback= 0;
 }
 
+static void tailing_queue_readable(InputFile *ipf) {
+  /* lifetime of ipf here is OK because destruction will cause
+   * on_cancel which will cancel this callback */
+  loop->on_time(loop, OOP_TIME_NOW, tailing_rable_call_time, ipf);
+}
+
 static int tailing_on_readable(struct oop_readable *rable,
                                oop_readable_call *cb, void *user) {
   InputFile *ipf= (void*)rable;
@@ -1053,7 +1379,7 @@ static int tailing_on_readable(struct oop_readable *rable,
   ipf->readable_callback_user= user;
   filemon_startfile(ipf);
 
-  loop->on_time(loop, OOP_TIME_NOW, tailing_rable_call_time, ipf);
+  tailing_queue_readable(ipf);
   return 0;
 }
 
@@ -1062,8 +1388,23 @@ static ssize_t tailing_try_read(struct oop_readable *rable, void *buffer,
   InputFile *ipf= (void*)rable;
   for (;;) {
     ssize_t r= read(ipf->fd, buffer, length);
-    if (!r && ipf==main_input_file) { errno=EAGAIN; return -1; }
-    if (r==-1 && errno==EINTR) continue;
+    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;
   }
 }
@@ -1099,11 +1440,15 @@ static void filemon_method_startfile(InputFile *ipf, Filemon_Perfile *pf) {
   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;
@@ -1123,6 +1468,7 @@ static void *filemon_inotify_readable(oop_source *lp, int fd,
       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;
@@ -1137,6 +1483,7 @@ static int filemon_method_init(void) {
   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;
 }
 
@@ -1183,139 +1530,358 @@ static void inputfile_tailing_start(InputFile *ipf) {
 
 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 ==========*/
+
+/*========== interaction with innd - state machine ==========*/
 
 /* See official state diagram at top of file.  We implement
  * this as follows:
- *
-          ================
-           WAITING
-          [Nothing/Noduct]
-           poll for F
-          ================
-               |
-               |     TIMEOUT
-               |`--------------------------.
-               |                           | install defer as backlog
-                | OPEN F SUCCEEDS           | exit
-     ,--------->|                           V
-     |          V                         =========
-     |     ========                        (ESRCH)
-     |      NORMAL                        [Dropped]
-     |     [Normal]                       =========
-     |      read F
-     |     ========
-     |          |
-     |          | F IS SO BIG WE SHOULD FLUSH
-     ^          | hardlink F to D
-     |     [Hardlinked]
-     |          | unlink F
-     |          | our handle onto F is now onto D
-     |     [Moved]
-     |          |
-     |          |<---------------------------------------------------.
+ * -8<-
+
+           .=======.
+           ||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[-ABSENT]                                        |
      |     [Flushing]                                                |
-     |      read D                                                   |
-     |     ==========                                                |
+     |     main D tail/none                                          |
+     |     ==================                                        |
      |          |                                                    |
      |          |   INNDCOMM FLUSH FAILS                             ^
-     |          |`----------------------->--------.                  |
-     |          |                                 |                  |
-     |          |   NO SUCH SITE                  V                  |
-     ^          |`----------------.            =========             |
-     |          |                 |            FLUSHFAIL             |
-     |          |                 V            [Moved]               |
-     |          |            ==========        read D                |
-     |          |             DROPPING         =========             |
-     |          |            [Dropping]           |                  |
-     |          |             read D              | TIME TO RETRY    |
-     |          |            ==========           `------------------'
-     |          | FLUSH OK        |
-     |          | open F          | AT EOF OF D AND ALL PROCESSED
-     |          V                 | install defer as backlog
-     |     ===========            | unlink D
-     |      SEPARATED             | exit
-     |     [Separated]            V
-     |      read D            ==========
-     |     ===========         (ESRCH)
-     |          |             [Droppped]
-     |          |             ==========
-     |          V
-     |          | AT EOF OF D
-     ^          |
-     |     ===========
-     |      FINISHING
-     |     [Finishing]
-     |      read F
-     |      write D
-     |     ===========
-     |          |
-     |          | ALL D PROCESSED
-     |          | install defer as backlog
-     |          | start new defer
-     ^          V unlink D
-     |          | close D
-     |          |
-     `----------'
-
- *
+     |          |`----------------------->----------.                |
+     |          |                                   |                |
+     |          |   NO SUCH SITE                    V                |
+     ^          |`--------------->----.         ==================== |
+     |          |                      \        FLUSHFAILED[-ABSENT] |
+     |          |                       \         [Moved]            |
+     |          | FLUSH OK               \       main D tail/none    |
+     |          | open F                  \     ==================== |
+     |          |                          \        |                |
+     |          |                           \       | TIME TO RETRY  |
+     |          |`------->----.     ,---<---'\      `----------------'
+     |          |    D NONE   |     | D NONE  `----.
+     |          V             |     |              V
+     |     =============      V     V                    ============
+     |      SEPARATED-1       |     |              DROPPING-1
+     |      flsh->fd>=0       |     |              flsh->fd>=0
+     |     [Separated]        |     |             [Dropping]
+     |      main F idle       |     |              main none
+     |      old D tail        |     |              old D tail
+     |     =============      |     |             ============
+     |          |             |     | install       |
+     ^          | EOF ON D    |     |  defer        | EOF ON D
+     |          V             |     |               V
+     |     ===============    |     |             ===============
+     |      SEPARATED-2       |     |              DROPPING-2
+     |      flsh->fd==-1      |     V              flsh->fd==-1
+     |     [Finishing]        |     |             [Dropping]
+     |      main F tail              |     `.             main none
+     |      old D closed      |       `.           old D closed
+     |     ===============    V                `.        ===============
+     |          |                        `.          |
+     |          | ALL D PROCESSED           `.        | ALL D PROCESSED
+     |          V install defer as backlog    `.      | install defer
+     ^          | close D                       `.    | close D
+     |          | unlink D                        `.  | unlink D
+     |          |                                  |  |
+     |          |                                  V  V
+     `----------'                               ==============
+                                                 DROPPED
+                                                [Dropped]
+                                                 main none
+                                                 old none
+                                                 some backlog
+                                                ==============
+                                                      |
+                                                      | ALL BACKLOG DONE
+                                                      |
+                                                      | unlink lock
+                                                      | exit
+                                                      V
+                                                  ==========
+                                                  (ESRCH)
+                                                 [Droppped]
+                                                 ==========
+ * ->8-
  */
 
-static char *path_ductlock, *path_duct, *path_ductdefer;
+static void statemc_init(void) {
+  struct stat stab, stabf;
 
-typedef struct {
-  /* This is an instance of struct oop_readable */
-  struct oop_readable readable; /* first */
-  oop_readable_call *readable_callback;
-  void *readable_callback_user;
+  path_lock=        xasprintf("%s_lock",      feedfile);
+  path_flushing=    xasprintf("%s_flushing",  feedfile);
+  path_defer=       xasprintf("%s_defer",     feedfile);
+  globpat_backlog=  xasprintf("%s_backlog*",  feedfile);
 
-  int fd;
-  const char *path; /* ptr copy of path_<foo> or feedfile */
-  struct Filemon_Perfile *filemon;
+  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_multiple) exit(0);
+       fatal("another duct holds the lockfile");
+      }
+      sysdie("fcntl F_SETLK lockfile %s", path_lock);
+    }
 
-  oop_read *rd;
-  long inprogress; /* no. of articles read but not processed */
-  off_t offset;
-} InputFile;
+    xfstat_isreg(lockfd, &stabf, "lockfile");
+    int lock_noent;
+    xlstat_isreg(path_lock, &stab, &lock_noent, "lockfile");
 
-typedef enum {
-  sm_WAITING,
-  sm_NORMAL,
-  sm_FLUSHING,
-  sm_FLUSHFAIL,
-  sm_DROPPING,
-  sm_SEPARATED,
-  sm_FINISHING;
-} StateMachineState;
-
-static InputFile *main_input_file, *old_input_file;
-static StateMachineState sms;
-static int waiting_periods_sofar;
+    if (!lock_noent && samefile(&stab, &stabf))
+      break;
+
+    if (close(lockfd))
+      sysdie("could not close stale lockfile %s", path_lock);
+  }
+  debug("startup: locked");
+
+  search_backlog_file();
+
+  int defer_noent;
+  xlstat_isreg(path_defer, &stab, &defer_noent, "defer file");
+  if (defer_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);
+    }
+  }
+
+  struct stat stab_f, stab_d;
+  int noent_f;
+
+  InputFile *file_d= open_input_file(path_flushing);
+  if (file_d) xfstat_isreg(file_d->fd, &stab_d, "flushing file");
+
+  xlstat_isreg(feedfile, &stab_f, &noent_f, "feedfile");
+
+  if (!noent_f && file_d && samefile(&stab_f, &stab_d)) {
+    debug("startup: F==D => Hardlinked");
+    xunlink(feedfile, "feed file (during startup)"); /* => Moved */
+    noent_f= 1;
+  }
+
+  if (noent_f) {
+    debug("startup: F ENOENT => Moved");
+    if (file_d) startup_set_input_file(file_d);
+    spawn_inndcomm_flush("feedfile missing at startup");
+    /* => Flushing, sms:=FLUSHING */
+  } else {
+    if (file_d) {
+      debug("startup: F!=D => Separated");
+      startup_set_input_file(file_d);
+      SMS(SEPARATED, 0, "found both old and current feed files");
+    } else {
+      debug("startup: F exists, D ENOENT => Normal");
+      FILE *file_f= open_input_file(feedfile);
+      if (!file_f) die("feed file vanished during startup");
+      startup_set_input_file(file_f);
+      SMS(NORMAL, flushfail_retry_periods, "normal startup");
+    }
+  }
+}
+
+static void statemc_period_poll(void) {
+  if (!sm_period_counter) return;
+  sm_period_counter--;
+  assert(sm_period_counter>=0);
+
+  if (sm_period_counter) return;
+  switch (sms) {
+  case sm_NORMAL:
+    statemc_start_flush("periodic"); /* Normal => Flushing; => FLUSHING */
+    break;
+  case sm_FLUSHFAILED:
+    spawn_inndcomm_flush("retry"); /* Moved => Flushing; => FLUSHING */
+    break;
+  default:
+    abort();
+  }
+}
+
+static void startup_set_input_file(InputFile *f) {
+  assert(!main_input_file);
+  main_input_file= f;
+  inputfile_tailing_start(f);
+}
+
+static int inputfile_is_done(InputFile *ipf) {
+  if (!ipf) return 0;
+  if (ipf->inprogress) return 0; /* new article in the meantime */
+  if (ipf->fd >= 0); return 0; /* not had EOF */
+  return 1;
+}
+
+static void notice_processed(InputFile *ipf, const char *what,
+                            const char *spec) {
+#define RCI_NOTHING(x) /* nothing */
+#define RCI_TRIPLE_FMT(x) " " #x "=" RCI_TRIPLE_FMT_BASE
+#define RCI_TRIPLE_VALS(x) RCI_TRIPLE_VALS_BASE(ipf->counts, .x)
+
+  info("processed %s%s offered=%d(ch%d,nc%d) accepted=%d(ch%d+nc%d)"
+       RESULT_COUNTS(RCI_NOTHING, RCI_TRIPLE_FMT)
+       ,
+       what,spec,
+       ipf->counts[art_Unchecked].sent + ipf->counts[art_Unsolicited].sent
+       , ipf->counts[art_Unchecked].sent, ipf->counts[art_Unsolicited].sent,
+       ipf->counts[art_Wanted].accepted + ipf->counts[art_Unsolicited].accepted
+       ,ipf->counts[art_Wanted].accepted,ipf->counts[art_Unsolicited].accepted
+       RESULT_COUNTS(RCI_NOTHING,  RCI_TRIPLE_VALS)
+       );
+}
+
+static void statemc_check_backlog_done(void) {
+  InputFile *ipf= backlog_input_file();
+  if (!inputfile_is_done(ipf)) return;
+
+  const char *slash= strrchr(ipf->path, "/");
+  const char *leaf= slash ? slash+1 : ipf->path;
+  const char *under= strchr(slash, "_");
+  const char *rest= under ? under+1 : leaf;
+  if (!strncmp(rest,"backlog",7)) rest += 7;
+  notice_processed(ipf,"backlog:",rest);
+  
+  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);
+  }
+  free(ipf);
+  backlog_input_file= 0;
+  search_backlog_file();
+  return;
+}
+
+static void statemc_check_flushing_done(void) {
+  InputFile *ipf= flushing_input_file;
+  if (!inputfile_is_done(ipf)) return;
+
+  assert(sms==sm_SEPARATED || sms==sm_DROPPING);
+
+  notice_processed(ipf,"feedfile",0);
+
+  close_defer();
+
+  xunlink(path_flushing, "old flushing file");
+
+  close_input_file(flushing_input_file);
+  free(flushing_input_file);
+  flushing_input_file= 0;
+
+  if (sms==sm_SEPARATED) {
+    notice("flush complete");
+    SMS(NORMAL, 0, "flush complete");
+  } else if (sms==sm_DROPPING) {
+    SMS(DROPPED, 0, "old flush complete");
+    search_backlog_file();
+    notice("feed dropped, but will continue until backlog is finished");
+  }
+}
+
+static void *statemc_check_input_done(oop_source *lp, struct timeval now,
+                                     void *u) {
+  assert(!inputfile_is_done(main_input_file));
+  statemc_check_flushing_done();
+  statemc_check_backlog_done();
+  return OOP_CONTINUE;
+}
+
+static void queue_check_input_done(void) {
+  loop->on_time(loop, OOP_TIME_NOW, statemc_check_input_done, 0);
+}
+
+static void statemc_setstate(StateMachineState newsms, int periods,
+                            const char *forlog, const char *why) {
+  sms= newsms;
+  sm_period_counter= periods;
+
+  const char *xtra= "";
+  switch (sms) {
+  case sm_FLUSHING: sm_FLUSHFAILED:
+    if (!main_input_file) xtra= "-ABSENT";
+    break;
+  case sm_SEPARATED: case sm_DROPPING:
+    xtra= flushing_input_file->fd >= 0 ? "-1" : "-2";
+    break;
+  default:;
+  }
+
+  if (periods) {
+    info("%s%s[%d] %s",forlog,xtra,periods,why);
+  } else {
+    info("%s%s %s",forlog,xtra,why);
+  }
+}
+
+/*---------- defer and backlog files ----------*/
 
 static void open_defer(void) {
   struct stat stab;
 
-  assert(!defer);
-  defer= fopen(path_ductdefer, "a+");
-  if (!defer) sysdie("could not open defer file %s", path_ductdefer);
+  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 */
 
-  r= fstat(fileno(defer), &stab);
-  if (r) sysdie("could not stat newly opened defer file %s", path_ductdefer);
+  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_ductdefer);
+    die("defer file %s size is far too large", path_defer);
 
   if (!stab.st_size)
     return;
@@ -1325,14 +1891,14 @@ static void open_defer(void) {
   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_ductdefer);
+      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_ductdefer);
+       sysdie("failed read from defer file %s", path_defer);
       else
-       die("defer file %s shrank while we were checking it!", path_ductdefer);
+       die("defer file %s shrank while we were checking it!", path_defer);
     }
     if (r=='\n') break;
     truncto--;
@@ -1341,125 +1907,285 @@ static void open_defer(void) {
   if (stab.st_size != truncto) {
     warn("truncating half-record at end of defer file %s -"
         " shrinking by %ld bytes from %ld to %ld",
-        path_ductdefer, orgsize - truncto, orgsize, truncto);
+        path_defer, orgsize - truncto, orgsize, truncto);
 
-    if (fflush(defer)) sysdie("could not flush defer file %s", path_ductdefer);
+    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_ductdefer);
+      sysdie("could not truncate defer file %s", path_defer);
 
   } else {
     info("continuing existing defer file %s (%ld bytes)",
-        path_ductdefer, orgsize);
+        path_defer, orgsize);
   }
   if (fseek(defer, truncto, SEEK_SET))
-    sysdie("could not seek to new end of defer file %s", path_ductdefer);
+    sysdie("could not seek to new end of defer file %s", path_defer);
 }
 
-static void statemc_init(void) {
+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 void 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;
 
-  path_ductlock=  xasprintf("%s_duct.lock",  feedfile);
-  path_duct=      xasprintf("%s_duct",       feedfile);
-  path_ductdefer= xasprintf("%s_duct.defer", feedfile);
+  if (backlog_input_file) return 3;
 
-  if (lstat(path_ductdefer, &stab)) {
-    if (errno!=ENOENT) sysdie("could not check defer file %s", path_defer);
-  } else {
-    if (!S_ISREG(stab.st_mode))
-      die("defer file %s not a plain file (mode 0%lo)",
-         path_defer, (unsigned long)stab.st_mode);
-    switch (stab.st_nlink==1) {
-    case 1: /* ok */ break;
-    case 2:
-      if (unlink(path_defer))
-       sysdie("could not unlink stale defer file link %s (presumably"
-              " hardlink to backlog file)", path_defer);
-      break;
-    default:
-      die("defer file %s has unexpected link count %d",
-         path_defer, stab.st_nlink);
+ 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);
   }
-  open_defer();
-
-  int lockfd= open(path_ductlock, O_CREAT|O_RDWR, 0600);
-  if (lockfd<0) sysdie("open lockfile %s", path_ductlock);
-
-  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)
-      die("another duct holds the lockfile");
-    sysdie("fcntl F_SETLK lockfile %s", path_ductlock);
+
+  globfree(&gl);
+
+  if (!oldest_path) {
+    debug("backlog scan: none");
+
+    if (sms==sm_DROPPED) {
+      notice("feed dropped and our work is complete");
+      xunlink(path_lock, "lockfile for old feed");
+      exit(0);
+    }
+    until_backlog_nextscan= backlog_spontaneous_rescan_periods;
+    return 0;
   }
 
-  InputFile *file_d= open_input_file(path_duct);
+  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 (file_d) {
-    struct stat stab_f, stab_d;
+  if (age_deficiency <= 0) {
+    debug("backlog scan: found age=%f deficiency=%ld oldest=%s",
+         age, age_deficiency, oldest_path);
 
-    r= stat(feedfile, &stab_f);
-    if (r) {
-      if (errno!=ENOENT) sysdie("check feed file %s", feedfile);
-      /* D exists, F ENOENT => Moved */
-      goto found_moved;
+    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;
+  }
 
-    /* F and D both exist */
+  until_backlog_nextscan= age_deficiency / PERIOD_SECONDS;
 
-    r= fstat(file_d->fd, &stab_d);
-    if (r) sysdie("check duct file %s", ductfile);
+  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;
+      tailing_queue_readable(flushing_input_file);
+        /* we probably previously returned EAGAIN from our fake read method
+        * when in fact we were at EOF, so signal another readable event
+        * so we actually see the EOF */
+
+      main_input_file= 0;
+
+      if (flushing_input_file) {
+       SMS(DROPPING, 0, "feed dropped by innd, but must finish last flush");
+      } else {
+       close_defer();
+       SMS(DROPPED, 0, "feed dropped by innd");
+       search_backlog_file();
+      }
+      return OOP_CONTINUE;
+
+    case 0:
+      /* as above */
+      flushing_input_file= main_input_file;
+      tailing_queue_readable(flushing_input_file);
+
+      main_input_file= open_input_file(feedfile);
+      if (!main_input_file)
+       die("flush succeeded but feedfile %s does not exist!", feedfile);
+
+      if (flushing_input_file) {
+       SMS(SEPARATED, spontaneous_flush_periods, "recovery flush complete");
+      } else {
+       close_defer();
+       SMS(NORMAL, spontaneous_flush_periods, "flush complete");
+      }
+      return OOP_CONTINUE;
+
+    default:
+      goto unexpected_exitstatus;
 
-    if (stab_d.st_ino == stab_f.st_ino &&
-       stab_d.st_dev == stab_f.st_dev) {
-      /* F==D => Hardlinked*/
-      r= unlink(path_duct);
-      if (r) sysdie("unlink feed file %s during startup", feedfile);
-    found_moved:
-      /* => Moved */
-      startup_set_input_file(file_d);
-      spawn_inndcomm_flush(); /* => Flushing, sets sms to sm_FLUSHING */
-    } else {
-      /* F!=D => Separated */
-      sms= sm_SEPARATED;
-      startup_set_input_file(file_d);
     }
-  } else { /*!file_d*/
-    sms= sm_WAITING;
-    statemc_waiting_poll();
+  } 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(FLUSHFAILED, flushfail_retry_periods, "flush failed, will retry");
 }
 
-static void statemc_poll(void) {
-  if (sms == sm_WAITING) statemc_waiting_poll();
-  if (sms == sm_FINISHING && !old_input_file->inprogress) statemc_finishdone();
+static void inndcommfail(const char *what) {
+  syswarn("error communicating with innd: %s failed: %s", what, ICCfailure);
+  exit(INNDCOMMCHILD_ESTATUS_FAIL);
 }
 
-static void statemc_waiting_poll(void) {
-  InputFile *file_f= open_input_file(feedfile);
-  if (!file_f) {
-    if (waiting_periods_sofar++ > waiting_timeout_periods)
-      die("timed out waiting for innd to create feed file %s", feedfile);
-    return;
+void spawn_inndcomm_flush(const char *why) { /* Moved => Flushing */
+  int pipefds[2];
+
+  notice("flushing %s",why);
+
+  assert(sms==sm_NORMAL || sms==sm_FLUSHFAILED);
+  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);
   }
-  startup_set_input_file(file_d);
-  sms= sm_NORMAL;
+
+  close(pipefds[1]);
+  int sentinel_fd= pipefds[0];
+  on_fd_read_except(sentinel_fd, inndcomm_event);
+
+  SMS(FLUSHING, 0, why);
 }
 
-static void startup_set_input_file(InputFile *f) {
-  assert(!main_input_file);
-  main_input_file= f;
-  inputfile_tailing_start(f);
+/*========== main program ==========*/
+
+static void postfork_inputfile(InputFile *ipf) {
+  if (!ipf) return;
+  assert(ipf->fd >= 0);
+  close(ipf->fd);
+  ipf->fd= -1;
 }
 
-/*========== flushing the feed ==========*/
+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);
+}
 
-/*========== main program ==========*/
+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 };                    \
@@ -1472,17 +2198,267 @@ static void startup_set_input_file(InputFile *f) {
     loop->on_time(loop, what##_timeout, what##_timedout, 0);                \
   }
 
-EVERY(filepoll, {5,0}, { check_master_queue(); })
+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}, {
-  if (connect_delay) connect_delay--;
-  statemc_poll();
+  debug("PERIOD"
+       " sms=%s[%d] queue=%d until_connect=%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], sm_period_counter, queue.count, until_connect,
+       DEBUG_IPF(main), DEBUG_IPF(flushing), DEBUG_IPF(flushing),
+       idle.count, working.count, full.count,
+       (long)connecting_child, (long)inndcomm_child
+       );
+
+  if (until_connect) until_connect--;
+
+  poll_backlog_file();
+  if (!backlog_input_file) close_defer(); /* want to start on a new backlog */
+  statemc_period_poll();
   check_master_queue();
 });
 
-main {
-  ignore sigpipe;
-  if (!filemon_init())
+
+/*========== 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[]= {
+{'f',"feedfile",              &feedfile,                 op_string           },
+{'q',"quiet-multiple",        &quiet_multiple,           op_booltrue,  1     },
+
+{ 0, "max-connections",       &max_connections           op_integer          },
+{ 0, "max-queue-per-conn",    &max_queue_per_conn        op_integer          },
+
+
+{ 0, "streaming",             &try_stream,               op_booltrue,  1     },
+{ 0, "no-streaming",          &try_stream,               op_boolfalse, 1     },
+{'P',"port",                  &port                      op_integer          },
+{ 0, "inndconf",              &inndconffile,             op_string           },
+{'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, "reconnect-interval",    &reconnect_delay_periods,  op_periods_rndup    },
+{ 0, "flush-retry-interval",  &flushfail_retry_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))
+    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;
+
+  if (!pathoutgoing)
+    pathoutgoing= innconf->pathoutgoing;
+  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);
+
+  loop= oop_sys_new();
+  if (!loop) sysdie("could not create liboop event loop");
+
+  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.
+}