chiark / gitweb /
options for everything
[inn-innduct.git] / backends / innduct.c
index 610504e8d84c0dc858d424d3ca890f753b7a3cc8..8e599f56026835595780ad3a28725754699fd06f 100644 (file)
 /*
 /*
+ * TODO
+ *  - actually implement badusage
+ *  - options for all options
+ *  - manpage
+ *  - pid, sitename, hostname in lockfile
+ *  - -k kill mode ?
+ */
+
+/*
+ * 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
  * Four files full of
- *    token article
+ *    token messageid
+ * or might be blanked out
+ *    <spc><spc><spc><spc>....
  *
  *
- *   site.name_ductlock        lock taken out by innduct
- * F site.name                 written by innd
- * D site.name_duct            moved aside by innduct, for flush
- *   site.name_deferwork       431'd articles, still being written
- *   site.name_defergo_<inum>  431'd articles, ready for innxmit
- *   site.name_deferlock       lock taken out by innxmit wrapper
+ * 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.
  *
  *
- * OVERALL STATES:
+ * 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 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.
+
+ * ->8- -^L-
  *
  *
- *                                                  START
- *   ,-->--.                                           |
- *   |     |                                         stat D
- *   |     |                                         /   |
- *   |     |                                  ENOENT/    |exists
- *   |     V                            <----------'     |
- *   |  Normal                                         stat F
- *   |   F: innd writing, duct reading                  /|\
- *   |   D: ENOENT                                     / | \
- *   |     |                                          /  |  \
- *   |     |  duct decides time to flush      same   /   |   |
- *   |     |  duct makes hardlink             as D  /    |   |
- *   |     |                                       /     |   |
- *   |     V                            <---------'      |   |
- *   |  Hardlinked                                       |   |
- *   |   F == D: innd writing, duct reading              |   |
- *   ^     |                                             |   |
- *   |     |  duct unlinks F                            /    |
- *   |     V                                  ENOENT   /     |
- *   |  Moved                            <------------'      |
- *   |   F: ENOENT                                           |
- *   |   D: innd writing, duct reading                       |
- *   |     |                                                 |
- *   |     |  duct flushes feed                              |
- *   |     |   (others can too, harmlessly)                  |
- *   |     V                                                 |
- *   |  Separated                          <-----------------'
- *   |   F: innd writing                        different to D
- *   |   D: duct reading
- *   |     |
- *   |     V  duct completes processing of D
- *   |     |  duct unlinks D
- *   |     |
- *   `--<--'
+ * rune for printing diagrams:
+
+perl -ne 'print if m/-8\<-/..m/-\>8-/; print "\f" if m/-\^L-/' backends/innduct.c |a2ps -R -B -ops
+
  *
  */
 
  *
  */
 
-static int max_connections, max_queue_per_conn;
-static int connection_setup_timeout, port, try_stream;
-static const char *remote_host;
+/*============================== PROGRAM ==============================*/
+
+#define _GNU_SOURCE 1
+
+#include "config.h"
+#include "storage.h"
+#include "nntp.h"
+#include "libinn.h"
+#include "inndcomm.h"
+
+#include "inn/list.h"
+#include "inn/innconf.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 <stdarg.h>
+#include <assert.h>
+#include <stdlib.h>
+#include <glob.h>
+#include <time.h>
+#include <math.h>
 
 
-#define ISNODE(T)    T *next, *back;
-#define LIST(T)      struct { T *head, *tail, *tailpred; int count; }
+#include <oop.h>
+#include <oop-read.h>
 
 
-#define NODE(n) ((struct node*)&(n)->head)
+/*----- general definitions, probably best not changed -----*/
 
 
-#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 CONNCHILD_ESTATUS_STREAM   4
+#define CONNCHILD_ESTATUS_NOSTREAM 5
+
+#define INNDCOMMCHILD_ESTATUS_FAIL     6
+#define INNDCOMMCHILD_ESTATUS_NONESUCH 7
+
+#define MAX_LINE_FEEDFILE (NNTP_MSGID_MAXLEN + sizeof(TOKEN)*2 + 10)
+
+#define VA                va_list al;  va_start(al,fmt)
+#define PRINTF(f,a)       __attribute__((__format__(printf,f,a)))
+#define NORET_PRINTF(f,a) __attribute__((__noreturn__,__format__(printf,f,a)))
+
+/*----- doubly linked lists -----*/
+
+#define ISNODE(T)   struct node list_node
+#define DEFLIST(T)                             \
+   typedef struct {                            \
+     union { struct list li; T *for_type; } u; \
+     int count;                                        \
+   } T##List
+
+#define NODE(n) (assert((void*)&(n)->list_node == &(n)), &(n)->list_node)
+
+#define LIST_CHECKCANHAVENODE(l,n) \
+  ((void)((n) == ((l).u.for_type))) /* just for the type check */
+
+#define LIST_ADDSOMEHOW(l,n,list_addsomehow)   \
+ ( LIST_CHECKCANHAVENODE(l,n),                 \
+   list_addsomehow(&(l).u.li, NODE((n))),      \
+   (void)(l).count++                           \
+   )
+
+#define LIST_REMSOMEHOW(l,list_remsomehow)     \
+ ( (typeof((l).u.for_type))                    \
+   ( (l).count                                 \
+     ? ( (l).count--,                          \
+        list_remsomehow(&(l).u.li) )           \
+     : 0                                       \
+     )                                         \
+   )
+
+
+#define LIST_ADDHEAD(l,n) LIST_ADDSOMEHOW((l),(n),list_addhead)
+#define LIST_ADDTAIL(l,n) LIST_ADDSOMEHOW((l),(n),list_addtail)
+#define LIST_REMHEAD(l) LIST_REMSOMEHOW((l),list_remhead)
+#define LIST_REMTAIL(l) LIST_REMSOMEHOW((l),list_remtail)
+
+#define LIST_INIT(l) (list_new(&(l).u.li))
+#define LIST_HEAD(l) ((typeof((l).u.for_type))(list_head((struct list*)&(l))))
+#define LIST_NEXT(n) ((typeof(n))list_succ(NODE((n))))
+#define LIST_BACK(n) ((typeof(n))list_pred(NODE((n))))
 
 
-#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)                       \
 #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++)
+ ( LIST_CHECKCANHAVENODE(l,n),                 \
+   list_remove(NODE((n))),                     \
+   (void)(l).count--                           \
+   )
+
+#define LIST_INSERT(l,n,pred)                                  \
+ ( LIST_CHECKCANHAVENODE(l,n),                                 \
+   LIST_CHECKCANHAVENODE(l,pred),                              \
+   list_insert((struct list*)&(l), NODE((n)), NODE((pred))),   \
+   (void)(l).count++                                           \
+   )
+
+/*----- type predeclarations -----*/
+
+typedef struct Conn Conn;
+typedef struct Article Article;
+typedef struct InputFile InputFile;
+typedef struct XmitDetails XmitDetails;
+typedef struct Filemon_Perfile Filemon_Perfile;
+typedef enum StateMachineState StateMachineState;
+
+DEFLIST(Conn);
+DEFLIST(Article);
+
+/*----- function predeclarations -----*/
+
+static void conn_maybe_write(Conn *conn);
+static void conn_make_some_xmits(Conn *conn);
+static void *conn_write_some_xmits(Conn *conn);
+
+static void xmit_free(XmitDetails *d);
+
+#define SMS(newstate, periods, why) \
+   (statemc_setstate(sm_##newstate,(periods),#newstate,(why)))
+static void statemc_setstate(StateMachineState newsms, int periods,
+                            const char *forlog, const char *why);
+
+static void statemc_start_flush(const char *why); /* Normal => Flushing */
+static void spawn_inndcomm_flush(const char *why); /* Moved => Flushing */
+
+static void check_assign_articles(void);
+static void queue_check_input_done(void);
+
+static void statemc_check_flushing_done(void);
+static void statemc_check_backlog_done(void);
+
+static void postfork(void);
+
+static void open_defer(void);
+static void close_defer(void);
+static void search_backlog_file(void);
+
+static void inputfile_reading_start(InputFile *ipf);
+static void inputfile_reading_stop(InputFile *ipf);
+
+static void filemon_start(InputFile *ipf);
+static void filemon_stop(InputFile *ipf);
+static void filemon_callback(InputFile *ipf);
+
+static void vconnfail(Conn *conn, const char *fmt, va_list al) PRINTF(2,0);
+static void connfail(Conn *conn, const char *fmt, ...)         PRINTF(2,3);
+
+static const oop_rd_style peer_rd_style;
+static oop_rd_call peer_rd_err, peer_rd_ok;
+
+/*----- configuration options -----*/
+
+static const char *sitename, *remote_host;
+static const char *feedfile;
+static int quiet_multiple=0;
+static int become_daemon=1;
+static int try_stream=1;
+static int port=119;
+static const char *inndconffile;
+
+static int max_connections=10;
+static int max_queue_per_conn=200;
+static int target_max_feedfile_size=100000;
+static int period_seconds=60;
+
+static int connection_setup_timeout=200;
+static int inndcomm_flush_timeout=100;
+
+static double nocheck_thresh= 95.0; /* converted from percentage by main */
+static double nocheck_decay= 100; /* conv'd from articles to lambda by main */
+
+/* all these are initialised to seconds, and converted to periods in main */
+static int reconnect_delay_periods=1000;
+static int flushfail_retry_periods=1000;
+static int backlog_retry_minperiods=50;
+static int backlog_spontrescan_periods=300;
+static int spontaneous_flush_periods=100000;
+static int need_activity_periods=1000;
+
+static double max_bad_data_ratio= 1; /* conv'd from percentage by main */
+static int max_bad_data_initial= 30;
+  /* in one corrupt 4096-byte block the number of newlines has
+   * mean 16 and standard deviation 3.99.  30 corresponds to z=+3.5 */
+
+
+/*----- 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) RC_##x,
+  RESULT_COUNTS(RC_INDEX, RC_INDEX)
+  RCI_max
+} ResultCountIndex;
+
+
+/*----- transmission buffers -----*/
+
+#define CONNIOVS 128
+
+typedef enum {
+  xk_Malloc, xk_Const, xk_Artdata
+} XmitKind;
+
+struct XmitDetails {
+  XmitKind kind;
+  union {
+    char *malloc_tofree;
+    ARTHANDLE *sm_art;
+  } info;
+};
+
+
+/*----- core operational data structure types -----*/
+
+struct InputFile {
+  /* This is also an instance of struct oop_readable */
+  struct oop_readable readable; /* first */
+  oop_readable_call *readable_callback;
+  void *readable_callback_user;
+
+  int fd;
+  Filemon_Perfile *filemon;
+
+  oop_read *rd; /* non-0: reading; 0: constructing, or had EOF */
+  long inprogress; /* no. of articles read but not processed */
+  off_t offset;
+  int skippinglong;
+
+  int counts[art_MaxState][RCI_max];
+  int readcount_ok, readcount_blank, readcount_err;
+  char path[];
+};
 
 struct Article {
 
 struct Article {
-  char *mid;
+  ISNODE(Article);
+  ArtState state;
   int midlen;
   int midlen;
-  int nocheck; /* also used when CHECK says yes please */
+  InputFile *ipf;
+  TOKEN token;
+  off_t offset;
+  int blanklen;
+  char messageid[1];
 };
 
 };
 
-#define CONNBUFSZ 16384
+#define SMS_LIST(X)                            \
+  X(NORMAL)                                    \
+  X(FLUSHING)                                  \
+  X(FLUSHFAILED)                               \
+  X(SEPARATED)                                 \
+  X(DROPPING)                                  \
+  X(DROPPED)
 
 
-#define CN "<%d> "
+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
+};
 
 
-typedef struct Conn Conn;
 struct Conn {
   ISNODE(Conn);
 struct Conn {
   ISNODE(Conn);
-  int fd, max_queue, stream;
-  LIST(Article) queue;
-  Article *tosend; /* points into queue */
-  char circ_buf[CONNBUFSZ];
-  unsigned circ_read, circ_write;
+  int fd; /* may be 0, meaning closed (during construction/destruction) */
+  oop_read *rd; /* likewise */
+  int max_queue, stream, quitting;
+  int since_activity; /* periods */
+  ArticleList waiting; /* not yet told peer */
+  ArticleList priority; /* peer says send it now */
+  ArticleList sent; /* offered/transmitted - in xmit or waiting reply */
+  struct iovec xmit[CONNIOVS];
+  XmitDetails xmitd[CONNIOVS];
+  int xmitu;
 };
 
 
 };
 
 
-#define CHILD_ESTATUS_STREAM   4
-#define CHILD_ESTATUS_NOSTREAM 5
+/*----- general operational variables -----*/
 
 
-static int since_connect_attempt;
-static int nconns;
-static LIST(Conn) idle, working, full;
+/* main initialises */
+static oop_source *loop;
+static ConnList conns;
+static ArticleList queue;
+static char *path_lock, *path_flushing, *path_defer, *globpat_backlog;
 
 
-static LIST(Article) *queue;
+/* statemc_init initialises */
+static StateMachineState sms;
+static FILE *defer;
+static InputFile *main_input_file, *flushing_input_file, *backlog_input_file;
+static int sm_period_counter;
 
 
-static void perhaps_close(int *fd) { if (*fd) { close(*fd); fd=0; } }
+/* initialisation to 0 is good */
+static int until_connect, until_backlog_nextscan;
+static double accept_proportion;
+static int nocheck, nocheck_reported;
 
 
+/*========== logging ==========*/
 
 
-/*========== making new connections ==========*/
+static void logcore(int sysloglevel, const char *fmt, ...) PRINTF(2,3);
+static void logcore(int sysloglevel, const char *fmt, ...) {
+  VA;
+  if (become_daemon) {
+    vsyslog(sysloglevel,fmt,al);
+  } else {
+    vfprintf(stderr,fmt,al);
+    putc('\n',stderr);
+  }
+  va_end(al);
+}
 
 
-static int connecting_sockets[2]= {-1,-1};
-static pid_t connecting_child;
+static void logv(int sysloglevel, const char *pfx, int errnoval,
+                const char *fmt, va_list al) PRINTF(5,0);
+static void logv(int sysloglevel, const char *pfx, int errnoval,
+                const char *fmt, va_list al) {
+  char msgbuf[256]; /* NB do not call xvasprintf here or you'll recurse */
+  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 diewrap(fn, pfx, sysloglevel, err, estatus)            \
+  static void fn(const char *fmt, ...) NORET_PRINTF(1,2);      \
+  static void fn(const char *fmt, ...) {                       \
+    VA;                                                                \
+    logv(sysloglevel, pfx, err, fmt, al);                      \
+    exit(estatus);                                             \
+  }
+
+#define logwrap(fn, pfx, sysloglevel, err)             \
+  static void fn(const char *fmt, ...) PRINTF(1,2);    \
+  static void fn(const char *fmt, ...) {               \
+    VA;                                                        \
+    logv(sysloglevel, pfx, err, fmt, al);              \
+    va_end(al);                                                \
+  }
+
+diewrap(sysdie,   " critical", LOG_CRIT,    errno, 16);
+diewrap(die,      " critical", LOG_CRIT,    -1,    16);
+
+diewrap(sysfatal, " fatal",    LOG_ERR,     errno, 12);
+diewrap(fatal,    " fatal",    LOG_ERR,     -1,    12);
+
+logwrap(syswarn,  " warning",  LOG_WARNING, errno);
+logwrap(warn,     " warning",  LOG_WARNING, -1);
+
+logwrap(notice,   "",          LOG_NOTICE,  -1);
+logwrap(info,     " info",     LOG_INFO,    -1);
+logwrap(debug,    " debug",    LOG_DEBUG,   -1);
+
+
+/*========== utility functions etc. ==========*/
+
+static char *xvasprintf(const char *fmt, va_list al) PRINTF(1,0);
+static char *xvasprintf(const char *fmt, va_list al) {
+  char *str;
+  int rc= vasprintf(&str,fmt,al);
+  if (rc<0) sysdie("vasprintf(\"%s\",...) failed", fmt);
+  return str;
+}
+static char *xasprintf(const char *fmt, ...) PRINTF(1,2);
+static char *xasprintf(const char *fmt, ...) {
+  VA;
+  char *str= xvasprintf(fmt,al);
+  va_end(al);
+  return str;
+}
+
+static int close_perhaps(int *fd) {
+  if (!*fd) return 0;
+  int r= close(*fd);
+  *fd=0;
+  return r;
+}
+static void xclose(int fd, const char *what, const char *what2) {
+  int r= close(fd);
+  if (r) sysdie("close %s%s",what,what2?what2:"");
+}
+static void xclose_perhaps(int *fd, const char *what, const char *what2) {
+  if (!*fd) return;
+  xclose(*fd,what,what2);
+  *fd=0;
+}
+
+static pid_t xfork(const char *what) {
+  pid_t child;
+
+  child= fork();
+  if (child==-1) sysdie("cannot fork for %s",what);
+  debug("forked %s %ld", what, (unsigned long)child);
+  if (!child) postfork();
+  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)
 
 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);
+      warn("%s: child died with error exit status %d", what, es);
   } else if (WIFSIGNALED(status)) {
     int sig= WTERMSIG(status);
     const char *sigstr= strsignal(sig);
   } else if (WIFSIGNALED(status)) {
     int sig= WTERMSIG(status);
     const char *sigstr= strsignal(sig);
@@ -126,31 +607,228 @@ static void report_child_status(const char *what, int status) {
       warn("%s: child died due to unknown fatal signal %d%s",
           what, sig, 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);
+    warn("%s: child died with unknown wait status %d", what,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 time_t xtime(void) {
+  time_t now= time(0);
+  if (now==-1) sysdie("time(2) failed");
+  return now;
+}
+
+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= fstat(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);
   }
   }
-  perhaps_close(&connecting_sockets[0]);
-  perhaps_close(&connecting_sockets[1]);
+  if (enoent_r) *enoent_r= 0;
+  check_isreg(stab, path, what);
+}
 
 
-  if (connecting_child) {
-    int status;
-    r= kill(connecting_child, SIGKILL);
-    if (r) sysdie("cannot kill connect child");
+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);
+}
 
 
-    pid_t got= waitpid(connecting_child, &status, WNOHANG);
-    if (got==-1) sysdie("cannot reap connect child");
+static char *sanitise(const char *input) {
+  static char sanibuf[100]; /* returns pointer to this buffer! */
 
 
-    if (!(WIFEXITED(status) ||
-         (WIFSIGNALED(status) && WTERMSIG(status) == SIGKILL))) {
-      report_child_status("connect"
+  const char *p= input;
+  char *q= sanibuf;
+  *q++= '`';
+  for (;;) {
+    if (q > sanibuf+sizeof(sanibuf)-8) { strcpy(q,"'.."); break; }
+    int c= *p++;
+    if (!c) { *q++= '\''; *q=0; break; }
+    if (c>=' ' && c<=126 && c!='\\') { *q++= c; continue; }
+    sprintf(q,"\\x%02x",c);
+    q += 4;
+  }
+  return sanibuf;
+}
+
+static int isewouldblock(int errnoval) {
+  return errnoval==EWOULDBLOCK || errnoval==EAGAIN;
+}
+
+/*========== management of connections ==========*/
+
+static void conn_closefd(Conn *conn, const char *msgprefix) {
+  int r= close_perhaps(&conn->fd);
+  if (r) info("C%d %serror closing socket: %s",
+             conn->fd, msgprefix, strerror(errno));
+}
+
+static void conn_dispose(Conn *conn) {
+  if (!conn) return;
+  if (conn->rd) {
+    oop_rd_cancel(conn->rd);
+    oop_rd_delete_kill(conn->rd);
+    conn->rd= 0;
+  }
+  if (conn->fd) {
+    loop->cancel_fd(loop, conn->fd, OOP_WRITE);
+    loop->cancel_fd(loop, conn->fd, OOP_EXCEPTION);
+  }
+  conn_closefd(conn,"");
+  free(conn);
+  until_connect= reconnect_delay_periods;
+}
+
+static void *conn_exception(oop_source *lp, int fd,
+                           oop_event ev, void *conn_v) {
+  Conn *conn= conn_v;
+  unsigned char ch;
+  assert(fd == conn->fd);
+  assert(ev == OOP_EXCEPTION);
+  int r= read(conn->fd, &ch, 1);
+  if (r<0) connfail(conn,"read failed: %s",strerror(errno));
+  else connfail(conn,"exceptional condition on socket (peer sent urgent"
+               " data? read(,&ch,1)=%d,ch='\\x%02x')",r,ch);
+  return OOP_CONTINUE;
+}  
+
+static void vconnfail(Conn *conn, const char *fmt, va_list al) {
+  int requeue[art_MaxState];
+
+  Article *art;
+  while ((art= LIST_REMHEAD(conn->priority))) LIST_ADDTAIL(queue, art);
+  while ((art= LIST_REMHEAD(conn->waiting))) LIST_ADDTAIL(queue, art);
+  while ((art= LIST_REMHEAD(conn->sent))) {
+    requeue[art->state]++;
+    if (art->state==art_Unsolicited) art->state= art_Unchecked;
+    LIST_ADDTAIL(queue,art);
+  }
+
+  int i;
+  XmitDetails *d;
+  for (i=0, d=conn->xmitd; i<conn->xmitu; i++, d++)
+    xmit_free(d);
+
+  char *m= xvasprintf(fmt,al);
+  warn("C%d connection failed (requeueing " RCI_TRIPLE_FMT_BASE "): %s",
+       conn->fd, RCI_TRIPLE_VALS_BASE(requeue, /*nothing*/), m);
+  free(m);
+
+  LIST_REMOVE(conns,conn);
+  conn_dispose(conn);
+  check_assign_articles();
+}
+
+static void connfail(Conn *conn, const char *fmt, ...) {
+  va_list al;
+  va_start(al,fmt);
+  vconnfail(conn,fmt,al);
+  va_end(al);
+}
+
+static void check_idle_conns(void) {
+  Conn *conn;
+  for (conn=LIST_HEAD(conns); conn; conn=LIST_NEXT(conn))
+    conn->since_activity++;
+ search_again:
+  for (conn=LIST_HEAD(conns); conn; conn=LIST_NEXT(conn)) {
+    if (conn->since_activity <= need_activity_periods) continue;
+
+    /* We need to shut this down */
+    if (conn->quitting)
+      connfail(conn,"timed out waiting for response to QUIT");
+    else if (conn->sent.count)
+      connfail(conn,"timed out waiting for responses");
+    else if (conn->waiting.count || conn->priority.count)
+      connfail(conn,"BUG IN INNDUCT conn has queue but nothing sent");
+    else if (conn->xmitu)
+      connfail(conn,"peer has been sending responses"
+              " before receiving our commands!");
+    else {
+      static const char quitcmd[]= "QUIT\r\n";
+      int todo= sizeof(quitcmd)-1;
+      const char *p= quitcmd;
+      for (;;) {
+       int r= write(conn->fd, p, todo);
+       if (r<0) {
+         if (isewouldblock(errno))
+           connfail(conn, "blocked writing QUIT to idle connection");
+         else
+           connfail(conn, "failed to write QUIT to idle connection: %s",
+                    strerror(errno));
+         break;
+       }
+       assert(r<=todo);
+       todo -= r;
+       if (!todo) {
+         conn->quitting= 1;
+         conn->since_activity= 0;
+         debug("C%d is idle, quitting", conn->fd);
+         break;
+       }
+      }
     }
     }
-    connecting_child= 0;
+    goto search_again;
+  }
+}  
+
+/*---------- making new connections ----------*/
+
+static pid_t connecting_child;
+static int connecting_fdpass_sock;
+
+static void connect_attempt_discard(void) {
+  if (connecting_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", status);
+  }
+  if (connecting_fdpass_sock) {
+    cancel_fd_read_except(connecting_fdpass_sock);
+    xclose_perhaps(&connecting_fdpass_sock, "connecting fdpass socket",0);
   }
 }
 
   }
 }
 
@@ -164,9 +842,12 @@ static void connect_attempt_discard(void) {
 static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
   Conn *conn= 0;
 
 static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
   Conn *conn= 0;
 
-  conn= xcalloc(sizeof(*conn));
-  
-  DECL_MSG_CMSG(msg);
+  assert(fd == connecting_fdpass_sock);
+
+  conn= xmalloc(sizeof(*conn));
+  memset(conn,0,sizeof(*conn));
+
+  PREP_DECL_MSG_CMSG(msg);
   struct cmsghdr *h= 0;
   ssize_t rs= recvmsg(fd, &msg, MSG_DONTWAIT);
   if (rs >= 0) h= CMSG_FIRSTHDR(&msg);
   struct cmsghdr *h= 0;
   ssize_t rs= recvmsg(fd, &msg, MSG_DONTWAIT);
   if (rs >= 0) h= CMSG_FIRSTHDR(&msg);
@@ -176,43 +857,49 @@ static void *connchild_event(oop_source *lp, int fd, oop_event e, void *u) {
     if (got != -1) {
       assert(got==connecting_child);
       connecting_child= 0;
     if (got != -1) {
       assert(got==connecting_child);
       connecting_child= 0;
-      if (WIFEXITED(status) &&
-         (WEXITSTATUS(status) != 0
-          WEXITSTATUS(status) != CHILD_ESTATUS_STREAM &&
-          WEXITSTATUS(status) != CHILD_ESTATUS_NOSTREAM)) {
-       /* child already reported the problem */
-      } else if (WIFSIGNALED(status) && WTERMSIG(status) == SIGALARM) {
+      if (WIFEXITED(status)) {
+       if (WEXITSTATUS(status) != 0 &&
+           WEXITSTATUS(status) != CONNCHILD_ESTATUS_STREAM &&
+           WEXITSTATUS(status) != CONNCHILD_ESTATUS_NOSTREAM)
+         /* child already reported the problem */;
+       else
+         warn("connect: connection child exited code %d but no cmsg",
+              WEXITSTATUS(status));
+      } else if (WIFSIGNALED(status) && WTERMSIG(status) == SIGALRM) {
        warn("connect: connection attempt timed out");
        warn("connect: connection attempt timed out");
-      } else if (!WIFEXITED(status)) {
+      } else {
        report_child_status("connect", 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)
       }
     } 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");
+       syswarn("connect: read from fdpass socket failed");
+      else if (e == OOP_EXCEPTION)
+       warn("connect: unexpected exception on fdpass socket");
+      else if (!rs)
+       warn("connect: unexpected EOF on fdpass socket");
       else
       else
-       warn("connect: unexpected EOF on child socket");
+       fatal("connect: unexpected lack of cmsg from child");
     }
     goto x;
   }
 
     }
     goto x;
   }
 
-#define CHK(field, val)                                                          \
-  if (h->cmsg_##field != val) {                                                  \
-    die("connect: child sent cmsg with cmsg_" #field "=%d, expected %d"); \
-    goto x;                                                              \
+#define CHK(field, val)                                                         \
+  if (h->cmsg_##field != val) {                                                 \
+    die("connect: child sent cmsg with cmsg_" #field "=%d, expected %d", \
+       h->cmsg_##field, val);                                           \
+    goto x;                                                             \
   }
   CHK(level, SOL_SOCKET);
   CHK(type,  SCM_RIGHTS);
   }
   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
 
 #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");
 
   memcpy(&conn->fd, CMSG_DATA(h), sizeof(conn->fd));
 
 
   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);
   pid_t got= waitpid(connecting_child, &status, 0);
   if (got==-1) sysdie("connect: real wait for child");
   assert(got == connecting_child);
@@ -221,100 +908,104 @@ 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) {
   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:
   default:
-    die("connect: child gave unexpected exit status %d", es);
+    fatal("connect: child gave unexpected exit status %d", es);
   }
 
   }
 
-  set nonblocking;
-
   /* Phew! */
   /* Phew! */
-  LIST_ADDHEAD(idle, conn);
-  notice(CN "connected %s", conn->fd, conn->stream ? "streaming" : "plain");
+  LIST_INIT(conn->waiting);
+  LIST_INIT(conn->priority);
+  LIST_INIT(conn->sent);
+  conn->max_queue= conn->stream ? max_queue_per_conn : 1;
+
+  loop->on_fd(loop, conn->fd, OOP_EXCEPTION, conn_exception, conn);
+  conn->rd= oop_rd_new_fd(loop,conn->fd, 0, 0); /* sets nonblocking, too */
+  if (!conn->fd) sysdie("oop_rd_new_fd (fd=%d)",conn->fd);
+  int r= oop_rd_read(conn->rd, &peer_rd_style, NNTP_STRLEN,
+                    &peer_rd_ok, conn,
+                    &peer_rd_err, conn);
+  if (r) sysdie("oop_rd_read for peer (fd=%d)",conn->fd);
+
+  notice("C%d connected %s", conn->fd, conn->stream ? "streaming" : "plain");
+  LIST_ADDHEAD(conns, conn);
+
   connect_attempt_discard();
   connect_attempt_discard();
-  process_queue();
+  check_assign_articles();
   return 0;
 
  x:
   return 0;
 
  x:
-  if (conn) {
-    perhaps_close(&conn->fd);
-    free(conn);
-  }
+  conn_dispose(conn);
   connect_attempt_discard();
   connect_attempt_discard();
+  return OOP_CONTINUE;
+}
+
+static int allow_connect_start(void) {
+  return conns.count < max_connections
+    && !connecting_child
+    && !until_connect;
 }
 
 }
 
-static void connect_start() {
-  assert(!connecting_sockets[0]);
-  assert(!connecting_sockets[1]);
+static void connect_start(void) {
   assert(!connecting_child);
   assert(!connecting_child);
+  assert(!connecting_fdpass_sock);
 
   notice("starting connection attempt");
 
 
   notice("starting connection attempt");
 
-  r= socketpair(AF_UNIX, SOCK_STREAM, 0, connecting_sockets);
-  if (r) { syswarn("connect: cannot create socketpair for child"); goto x; }
+  int socks[2];
+  int r= socketpair(AF_UNIX, SOCK_STREAM, 0, socks);
+  if (r) { syswarn("connect: cannot create socketpair for child"); return; }
 
 
-  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];
 
   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");
+    xclose(socks[0], "(in child) parent's connection fdpass socket",0);
 
     alarm(connection_setup_timeout);
 
     alarm(connection_setup_timeout);
-    if (NNTPconnect(remote_host, port, &cn_from, &cn_to, buf) < 0) {
-      if (buf[0]) {
-       sanitise_inplace(buf);
-       die("connect: rejected: %s", buf);
-      } else {
-       sysdie("connect: connection attempt failed");
-      }
+    if (NNTPconnect((char*)remote_host, port, &cn_from, &cn_to, buf) < 0) {
+      if (buf[0]) fatal("connect: rejected: %s", sanitise(buf));
+      else sysfatal("connect: connection attempt failed");
     }
     }
-    if (NNTPsendpassword(remote_host, cn_from, cn_to) < 0)
-      sysdie("connect: authentication failed");
+    if (NNTPsendpassword((char*)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))
     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))
       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
        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);
       }
       int l= strlen(buf);
       assert(l>=1);
-      if (buf[-1]!='\n') {
-       sanitise_inplace(buf);
-       die("connect: response to MODE STREAM is too long: %.100s...",
-           remote_host, buf);
-      }
-      l--;  if (l>0 && buf[1-]=='\r') l--;
+      if (buf[-1]!='\n')
+       fatal("connect: response to MODE STREAM is too long: %.100s...",
+             sanitise(buf));
+      l--;  if (l>0 && buf[l-1]=='\r') l--;
       buf[l]= 0;
       char *ep;
       int rcode= strtoul(buf,&ep,10);
       buf[l]= 0;
       char *ep;
       int rcode= strtoul(buf,&ep,10);
-      if (ep != buf[3]) {
-       sanitise_inplace(buf);
-       die("connect: bad response to MODE STREAM: %.50s", buf);
-      }
+      if (ep != &buf[3])
+       fatal("connect: bad response to MODE STREAM: %.50s", sanitise(buf));
+
       switch (rcode) {
       case 203:
       switch (rcode) {
       case 203:
-       exitstatus= CHILD_ESTATUS_STREAM;
+       exitstatus= CONNCHILD_ESTATUS_STREAM;
        break;
       case 480:
       case 500:
        break;
       default:
        break;
       case 480:
       case 500:
        break;
       default:
-       sanitise_inplace(buf);
-       warn("connect: unexpected response to MODE STREAM: %.50s", buf);
+       warn("connect: unexpected response to MODE STREAM: %.50s",
+            sanitise(buf));
        exitstatus= 2;
        break;
       }
        exitstatus= 2;
        break;
       }
@@ -329,83 +1020,1839 @@ static void connect_start() {
     memcpy(CMSG_DATA(cmsg), &fd, sizeof(fd));
 
     msg.msg_controllen= cmsg->cmsg_len;
     memcpy(CMSG_DATA(cmsg), &fd, sizeof(fd));
 
     msg.msg_controllen= cmsg->cmsg_len;
-    r= sendmsg(connecting_sockets[1], &msg, 0);
+    r= sendmsg(socks[1], &msg, 0);
     if (r) sysdie("sendmsg failed for new connection");
 
     _exit(exitstatus);
   }
 
     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");
+  xclose(socks[1], "connecting fdpass child's socket",0);
+  connecting_fdpass_sock= socks[0];
+  on_fd_read_except(connecting_fdpass_sock, connchild_event);
+}
 
 
-  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;
+/*---------- assigning articles to conns, and transmitting ----------*/
 
 
- x:
-  connect_attempt_discard();
-}
+static void check_assign_articles(void) {
+  for (;;) {
+    if (!queue.count)
+      break;
 
 
-/*========== overall control of article flow ==========*/
-static void process_queue() {
-  if (!queue.count)
-    return;
+    Conn *walk, *use=0;
+    int spare=0, inqueue=0;
 
 
-  if (working.head) {
-    transmit(working.head);
-  } else if (idle.head) {
-    transmit(idle.head);
-  } else if (nconns < maxconns && queue.count >= max_queue_per_conn &&
-            !connecting_child && !connect_delay) {
-    connect_delay= reconnect_delay_periods;
-    connect_start();
-  }
-} 
-/*========== article transmission ==========*/
+    /* Find a connection to offer this article.  We prefer a busy
+     * connection to an idle one, provided it's not full.  We take the
+     * first (oldest) and since that's stable, it will mean we fill up
+     * connections in order.  That way if we have too many
+     * connections, the spare ones will go away eventually.
+     */
+    for (walk=LIST_HEAD(conns); walk; walk=LIST_NEXT(walk)) {
+      if (walk->quitting) continue;
+      inqueue= walk->sent.count + walk->priority.count
+            + walk->waiting.count;
+      spare= walk->max_queue - inqueue;
+      assert(inqueue <= max_queue_per_conn);
+      assert(spare >= 0);
+      if (inqueue==0) /*idle*/ { if (!use) use= walk; }
+      else if (spare>0) /*working*/ { use= walk; break; }
+    }
+    if (use) {
+      if (!inqueue) use->since_activity= 0; /* reset idle counter */
+      while (spare>0) {
+       Article *art= LIST_REMHEAD(queue);
+       LIST_ADDTAIL(use->waiting, art);
+       spare--;
+      }
+      conn_maybe_write(use);
+    } else if (allow_connect_start()) {
+      until_connect= reconnect_delay_periods;
+      connect_start();
+      break;
+    } else {
+      break;
+    }
+  }
+}
 
 
-static void *conn_writeable() {
-  for (;;) {
-    int circ_used= circ_write - circ_read;
-    if (circ_used < 0) circ_used += CONNBUFSZ;
-    writeable_moredata(conn, CONNBUFSZ-1 - circ_used);
+static void *conn_writeable(oop_source *l, int fd, oop_event ev, void *u) {
+  conn_maybe_write(u);
+  return OOP_CONTINUE;
+}
 
 
-    if (conn->circ_read == conn->circ_write)
-      return OOP_CONTINUE;
+static void conn_maybe_write(Conn *conn)  {
+  for (;;) {
+    conn_make_some_xmits(conn);
+    if (!conn->xmitu) {
+      loop->cancel_fd(loop, conn->fd, OOP_WRITE);
+      return;
+    }
 
 
-    struct iovec iov[2];
-    int niov= 1;
-    iov[0].iov_base= conn->circ_buf + conn->circ_read;
-    if (conn->circ_read > conn->circ_write) { /* wrapped */
-      iov[0].iov_len= CONNBUFSZ - conn->circ_read;
-      iov[1].iov_base= conn->circ_buf;
-      iov[1].iov_len= conn->circ_write;
-      if (niov[1].iov_len) niov= 2;
+    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 {
     } else {
-      iov[0].iov_len= conn->circ_write - conn->circ_read;
+      abort();
     }
     }
-    ssize_t rs= writev(conn->fd, &iov, niov);
+  }
+}
+
+/*========== 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= (char*)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, xk_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 (rs < 0) {
-      if (errno == EAGAIN) return OOP_CONTINUE;
-      syswarn(CN "write failed", conn->fd);
-      conn_failed(conn);
-      return OOP_CONTINUE;
+      if (isewouldblock(errno)) return OOP_CONTINUE;
+      connfail(conn, "write failed: %s", strerror(errno));
+      return OOP_HALT;
     }
     assert(rs > 0);
 
     }
     assert(rs > 0);
 
-    conn->circ_read += rs;
-    if (conn->circ_read > CONNBUFSZ)
-      conn->circ_read -= CONNBUFSZ;
+    int done;
+    for (done=0; rs && done<conn->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= (char*)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(conn->priority);
+    if (!art) art= LIST_REMHEAD(conn->waiting);
+    if (!art) break;
+
+    if (art->state >= art_Wanted || (conn->stream && nocheck)) {
+      /* actually send it */
+
+      ARTHANDLE *artdata= SMretrieve(art->token, RETR_ALL);
+
+      if (conn->stream) {
+       if (artdata) {
+         XMIT_LITERAL("TAKETHIS ");
+         xmit_noalloc(conn, art->messageid, 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->state=
+       art->state == art_Unchecked ? art_Unsolicited :
+       art->state == art_Wanted    ? art_Wanted      :
+       (abort(),-1);
+      art->ipf->counts[art->state][RC_sent]++;
+      LIST_ADDTAIL(conn->sent, art);
+
+    } else {
+      /* check it */
+
+      if (conn->stream)
+       XMIT_LITERAL("IHAVE ");
+      else
+       XMIT_LITERAL("CHECK ");
+      xmit_noalloc(conn, art->messageid, art->midlen);
+      XMIT_LITERAL("\r\n");
+
+      assert(art->state == art_Unchecked);
+      art->ipf->counts[art->state][RC_sent]++;
+      LIST_ADDTAIL(conn->sent, art);
+    }
   }
 }
 
 
   }
 }
 
 
-static void transmit(Conn *conn) {
-  assert(conn->queue.count < max_queue);
-  
+/*========== handling responses from peer ==========*/
 
 
-main {
-  ignore sigpipe;
+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_rd_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(Conn *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_HEAD(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(conn, "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(conn, "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(conn, "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(conn, "peer gave non-streaming response code to"
+              " CHECK/TAKETHIS: %s", sanitised_response);
+      return 0;
+    }
+  }
+
+  if (must_have_sent>0 && art->state < art_Wanted) {
+    connfail(conn, "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(conn, "peer says please sent the article but we just did: %s",
+            sanitised_response);
+    return 0;
+  }
+
+  Article *art_again= LIST_REMHEAD(conn->sent);
+  assert(art_again == art);
+  return art;
+}
+
+static void update_nocheck(int accepted) {
+  accept_proportion *= nocheck_decay;
+  accept_proportion += accepted * (1.0 - nocheck_decay);
+  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(Conn *conn, Article *art, int whichcount) {
+  art->ipf->counts[art->state][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,
+            (unsigned long)art->offset, ipf->path);
+    }
+    assert(r>=0 && r<=w);
+    art->blanklen -= w;
+    art->offset += w;
+  }
+
+  ipf->inprogress--;
+  assert(ipf->inprogress >= 0);
+  free(art);
+
+  if (!ipf->inprogress && ipf != main_input_file)
+    queue_check_input_done();
+}
+
+static void *peer_rd_ok(oop_source *lp, oop_read *oread, oop_rd_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 *sani= sanitise(data);
+
+  char *ep;
+  unsigned long code= strtoul(data, &ep, 10);
+  if (ep != data+3 || *ep != ' ' || data[0]=='0') {
+    connfail(conn, "badly formatted response from peer: %s", sani);
+    return OOP_CONTINUE;
+  }
+
+  if (conn->quitting) {
+    if (code!=205 && code!=503) {
+      connfail(conn, "peer gave unexpected response to QUIT: %s", sani);
+    } else {
+      notice("C%d idle connection closed", conn->fd);
+      assert(!conn->waiting.count);
+      assert(!conn->priority.count);
+      assert(!conn->sent.count);
+      assert(!conn->xmitu);
+      LIST_REMOVE(conns,conn);
+      conn_dispose(conn);
+    }
+    return OOP_CONTINUE;
+  }
+
+  conn->since_activity= 0;
+  Article *art;
+
+#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,musthavesent,how)          \
+  code_streaming= (streaming);                                 \
+  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 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(0,1,accepted); /* IHAVE says thanks */
+  case 239: ARTICLE_DEALTWITH(1,1,accepted); /* TAKETHIS says thanks */
+
+  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(-1);
+    assert(art->state == art_Unchecked);
+    art->ipf->counts[art->state][RC_accepted]++;
+    art->state= art_Wanted;
+    LIST_ADDTAIL(conn->priority, art);
+    break;
+
+  case 431: /* CHECK or TAKETHIS says try later */
+    code_streaming= 1;
+  case 436: /* IHAVE says try later */
+    GET_ARTICLE(0);
+    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;
+
+  }
+
+  conn_maybe_write(conn);
+  check_assign_articles();
+  return OOP_CONTINUE;
+}
+
+
+/*========== monitoring of input files ==========*/
+
+static void feedfile_eof(InputFile *ipf) {
+  assert(ipf != main_input_file); /* promised by tailing_try_read */
+  inputfile_reading_stop(ipf);
+
+  if (ipf == flushing_input_file) {
+    assert(sms==sm_SEPARATED || sms==sm_DROPPING);
+    if (main_input_file) inputfile_reading_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;
+    sysfatal("unable to open input file %s", path);
+  }
+  assert(fd>0);
+
+  InputFile *ipf= xmalloc(sizeof(*ipf) + strlen(path) + 1);
+  memset(ipf,0,sizeof(*ipf));
+
+  ipf->fd= fd;
+  strcpy(ipf->path, path);
+
+  return ipf;
+}
+
+static void close_input_file(InputFile *ipf) { /* does not free */
+  assert(!ipf->readable_callback); /* must have had ->on_cancel */
+  assert(!ipf->filemon); /* must have had inputfile_reading_stop */
+  assert(!ipf->rd); /* must have had inputfile_reading_stop */
+  assert(!ipf->inprogress); /* no dangling pointers pointing here */
+  xclose_perhaps(&ipf->fd, "input file ", ipf->path);
+}
+
+
+/*---------- dealing with articles read in the input file ----------*/
+
+static void *feedfile_got_bad_data(InputFile *ipf, off_t offset,
+                                  const char *data, const char *how) {
+  warn("corrupted file: %s, offset %lu: %s: %s",
+       ipf->path, (unsigned long)offset, how, sanitise(data));
+  ipf->readcount_err++;
+  if (ipf->readcount_err > max_bad_data_initial +
+      (ipf->readcount_ok+ipf->readcount_blank) / max_bad_data_ratio)
+    die("too much garbage in input file!  (%d errs, %d ok, %d blank)",
+       ipf->readcount_err, ipf->readcount_ok, ipf->readcount_blank);
+  return OOP_CONTINUE;
+}
+
+static void *feedfile_read_err(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;
+  assert(ev == OOP_RD_SYSTEM);
+  errno= errnoval;
+  sysdie("error reading input file: %s, offset %lu",
+        ipf->path, (unsigned long)ipf->offset);
+}
+
+static 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; }
+
+  off_t old_offset= ipf->offset;
+  ipf->offset += recsz + 1;
+
+#define X_BAD_DATA(m) return feedfile_got_bad_data(ipf,old_offset,data,m);
+
+  if (ev==OOP_RD_PARTREC)
+    feedfile_got_bad_data(ipf,old_offset,data,"missing final newline");
+    /* but process it anyway */
+
+  if (ipf->skippinglong) {
+    if (ev==OOP_RD_OK) ipf->skippinglong= 0; /* fine now */
+    return OOP_CONTINUE;
+  }
+  if (ev==OOP_RD_LONG) {
+    ipf->skippinglong= 1;
+    X_BAD_DATA("overly long line");
+  }
+
+  if (memchr(data,'\0',recsz)) X_BAD_DATA("nul byte");
+  if (!recsz) X_BAD_DATA("empty line");
+
+  if (data[0]==' ') {
+    if (strspn(data," ") != recsz) X_BAD_DATA("line partially blanked");
+    ipf->readcount_blank++;
+    return OOP_CONTINUE;
+  }
+
+  char *space= strchr(data,' ');
+  int tokenlen= space-data;
+  int midlen= (int)recsz-tokenlen-1;
+  if (midlen <= 2) X_BAD_DATA("no room for messageid");
+  if (space[1]!='<' || space[midlen]!='>') X_BAD_DATA("invalid messageid");
+
+  if (tokenlen != sizeof(TOKEN)*2+2) X_BAD_DATA("token wrong length");
+  memcpy(tokentextbuf, data, tokenlen);
+  tokentextbuf[tokenlen]= 0;
+  if (!IsToken(tokentextbuf)) X_BAD_DATA("token wrong syntax");
+
+  ipf->readcount_ok++;
+
+  art= xmalloc(sizeof(*art) - 1 + midlen + 1);
+  art->state= art_Unchecked;
+  art->midlen= midlen;
+  art->ipf= ipf;  ipf->inprogress++;
+  art->token= TextToToken(tokentextbuf);
+  art->offset= ipf->offset;
+  art->blanklen= recsz;
+  strcpy(art->messageid, space+1);
+  LIST_ADDTAIL(queue, art);
+
+  if (sms==sm_NORMAL && ipf==main_input_file &&
+      ipf->offset >= target_max_feedfile_size)
+    statemc_start_flush("feed file size");
+
+  check_assign_articles();
+  return OOP_CONTINUE;
+}
+
+/*========== tailing input file ==========*/
+
+static void *tailing_rable_call_time(oop_source *loop, struct timeval tv,
+                                    void *user) {
+  InputFile *ipf= user;
+  return ipf->readable_callback(loop, &ipf->readable,
+                               ipf->readable_callback_user);
+}
+
+static void tailing_on_cancel(struct oop_readable *rable) {
+  InputFile *ipf= (void*)rable;
+
+  if (ipf->filemon) filemon_stop(ipf);
+  loop->cancel_time(loop, OOP_TIME_NOW, tailing_rable_call_time, ipf);
+  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;
+
+  tailing_on_cancel(rable);
+  ipf->readable_callback= cb;
+  ipf->readable_callback_user= user;
+  filemon_start(ipf);
+
+  tailing_queue_readable(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->rd);
+       assert(sms==sm_SEPARATED || sms==sm_DROPPING);
+      } else if (ipf==backlog_input_file) {
+       assert(ipf->rd);
+      } 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;
+
+struct Filemon_Perfile {
+  int wd;
+};
+
+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 (isewouldblock(errno)) 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)
+
+struct Filemon_Perfile { int dummy; };
+
+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 */
+
+/*---------- filemon generic interface ----------*/
+
+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(loop, &ipf->readable, ipf->readable_callback_user);
+}
+
+/*---------- interface to start and stop an input file ----------*/
+
+static const oop_rd_style feedfile_rdstyle= {
+  OOP_RD_DELIM_STRIP, '\n',
+  OOP_RD_NUL_PERMIT,
+  OOP_RD_SHORTREC_LONG,
+};
+
+static void inputfile_reading_start(InputFile *ipf) {
+  assert(!ipf->rd);
+  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->rd);
+
+  int r= oop_rd_read(ipf->rd, &feedfile_rdstyle, MAX_LINE_FEEDFILE,
+                    feedfile_got_article,ipf, feedfile_read_err, ipf);
+  if (r) sysdie("unable start reading feedfile %s",ipf->path);
+}
+
+static void inputfile_reading_stop(InputFile *ipf) {
+  assert(ipf->rd);
+  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:
+ * -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[-ABSENT]                                        |
+     |     [Flushing]                                                |
+     |     main D tail/none                                          |
+     |     ==================                                        |
+     |          |                                                    |
+     |          |   INNDCOMM FLUSH FAILS                             ^
+     |          |`----------------------->----------.                |
+     |          |                                   |                |
+     |          |   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->rd!=0       |     |              flsh->rd!=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->rd==0       |     V              flsh->rd==0
+     |     [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 void startup_set_input_file(InputFile *f) {
+  assert(!main_input_file);
+  main_input_file= f;
+  inputfile_reading_start(f);
+}
+
+static void statemc_init(void) {
+  struct stat stab, stabf;
+
+  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;
+    int r= fcntl(lockfd, F_SETLK, &fl);
+    if (r==-1) {
+      if (errno==EACCES || isewouldblock(errno)) {
+       if (quiet_multiple) exit(0);
+       fatal("another duct holds the lockfile");
+      }
+      sysdie("fcntl F_SETLK lockfile %s", path_lock);
+    }
+
+    xfstat_isreg(lockfd, &stabf, path_lock, "lockfile");
+    int lock_noent;
+    xlstat_isreg(path_lock, &stab, &lock_noent, "lockfile");
+
+    if (!lock_noent && samefile(&stab, &stabf))
+      break;
+
+    xclose(lockfd, "stale lockfile ", 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, path_flushing,"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");
+      InputFile *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_start_flush(const char *why) { /* Normal => Flushing */
+  assert(sms == sm_NORMAL);
+
+  debug("starting flush (%s) (%lu >?= %lu) (%d)",
+       why,
+       (unsigned long)(main_input_file ? main_input_file->offset : 0),
+       (unsigned long)target_max_feedfile_size,
+       sm_period_counter);
+
+  int r= link(feedfile, path_flushing);
+  if (r) sysdie("link feedfile %s to flushing file %s",
+               feedfile, path_flushing);
+  /* => Hardlinked */
+
+  xunlink(feedfile, "old feedfile link");
+  /* => Moved */
+
+  spawn_inndcomm_flush(why); /* => Flushing FLUSHING */
+}
+
+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 int inputfile_is_done(InputFile *ipf) {
+  if (!ipf) return 0;
+  if (ipf->inprogress) return 0; /* new article in the meantime */
+  if (ipf->rd) 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, [RC_##x])
+
+#define CNT(art,rc) (ipf->counts[art_##art][RC_##rc])
+
+  info("processed %s%s read=%d(+%dbl,+%derr)"
+       " offered=%d(ch%d,nc%d) accepted=%d(ch%d+nc%d)"
+       RESULT_COUNTS(RCI_NOTHING, RCI_TRIPLE_FMT)
+       ,
+       what, spec,
+       ipf->readcount_ok, ipf->readcount_blank, ipf->readcount_err,
+       CNT(Unchecked,sent) + CNT(Unsolicited,sent)
+       , CNT(Unchecked,sent), CNT(Unsolicited,sent),
+       CNT(Wanted,accepted) + CNT(Unsolicited,accepted)
+       , CNT(Wanted,accepted), CNT(Unsolicited,accepted)
+       RESULT_COUNTS(RCI_NOTHING,  RCI_TRIPLE_VALS)
+       );
+
+#undef CNT
+}
+
+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:
+  case sm_FLUSHFAILED:
+    if (!main_input_file) xtra= "-ABSENT";
+    break;
+  case sm_SEPARATED:
+  case sm_DROPPING:
+    xtra= flushing_input_file->rd ? "-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;
+
+  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, path_defer, "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);
+
+    int 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;
+
+  struct stat stab;
+  xfstat_isreg(fileno(defer), &stab, path_defer, "defer file");
+
+  if (fclose(defer)) sysfatal("could not close defer file %s", path_defer);
+  defer= 0;
+
+  time_t now= xtime();
+
+  char *backlog= xasprintf("%s_backlog_%lu.%lu", feedfile,
+                          (unsigned long)now,
+                          (unsigned long)stab.st_ino);
+  if (link(path_defer, 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);
+
+  free(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, i;
+  struct stat stab;
+  const char *oldest_path=0;
+  time_t oldest_mtime=0, now;
+
+  if (backlog_input_file) return;
+
+ 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);
+  }
+
+  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_spontrescan_periods;
+    goto xfree;
+  }
+
+  now= xtime();
+  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", oldest_path);
+      globfree(&gl);
+      goto try_again;
+    }
+    inputfile_reading_start(backlog_input_file);
+    until_backlog_nextscan= -1;
+    goto xfree;
+  }
+
+  until_backlog_nextscan= age_deficiency / period_seconds;
+
+  if (backlog_spontrescan_periods >= 0 &&
+      until_backlog_nextscan > backlog_spontrescan_periods)
+    until_backlog_nextscan= backlog_spontrescan_periods;
+
+  debug("backlog scan: young age=%f deficiency=%ld nextscan=%d oldest=%s",
+       age, age_deficiency, until_backlog_nextscan, oldest_path);
+
+ xfree:
+  globfree(&gl);
+  return;
+}
+
+/*========== flushing the feed ==========*/
+
+static pid_t inndcomm_child;
+static int inndcomm_sentinel_fd;
+
+static void *inndcomm_event(oop_source *lp, int fd, oop_event e, void *u) {
+  assert(inndcomm_child);
+  assert(fd == inndcomm_sentinel_fd);
+  int status= xwaitpid(&inndcomm_child, "inndcomm");
+  cancel_fd_read_except(fd);
+  xclose_perhaps(&fd, "inndcomm sentinel pipe",0);
+
+  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;
+
+    }
+  } 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");
+  return OOP_CONTINUE;
+}
+
+static void inndcommfail(const char *what) {
+  syswarn("error communicating with innd: %s failed: %s", what, ICCfailure);
+  exit(INNDCOMMCHILD_ESTATUS_FAIL);
+}
+
+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);
+  assert(!inndcomm_sentinel_fd);
+
+  if (pipe(pipefds)) sysdie("create pipe for inndcomm child sentinel");
+
+  inndcomm_child= xfork("inndcomm child");
+
+  if (!inndcomm_child) {
+    const char *flushargv[2]= { sitename, 0 };
+    char *reply;
+    int r;
+
+    xclose(pipefds[0], "(in child) inndcomm sentinel parent's end",0);
+    /* parent spots the autoclose of pipefds[1] when we die or exit */
+
+    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);
+  }
+
+  xclose(pipefds[1], "inndcomm sentinel child's end",0);
+  inndcomm_sentinel_fd= pipefds[0];
+  assert(inndcomm_sentinel_fd);
+  on_fd_read_except(inndcomm_sentinel_fd, inndcomm_event);
+
+  SMS(FLUSHING, 0, why);
+}
+
+/*========== main program ==========*/
+
+static void postfork_inputfile(InputFile *ipf) {
+  if (!ipf) return;
+  xclose(ipf->fd, "(in child) input file ", ipf->path);
+}
+
+static void postfork_stdio(FILE *f, const char *what, const char *what2) {
+  /* we have no stdio streams that are buffered long-term */
+  if (!f) return;
+  if (fclose(f)) sysdie("(in child) close %s%s", what, what2?what2:0);
+}
+
+static void postfork(void) {
+  if (signal(SIGPIPE, SIG_DFL) == SIG_ERR)
+    sysdie("(in child) failed to reset SIGPIPE");
+
+  postfork_inputfile(main_input_file);
+  postfork_inputfile(flushing_input_file);
+
+  Conn *conn;
+  for (conn=LIST_HEAD(conns); conn; conn=LIST_NEXT(conn))
+    conn_closefd(conn,"(in child) ");
+
+  postfork_stdio(defer, "defer file ", path_defer);
+}
+
+#define EVERY(what, interval_sec, interval_usec, body)                      \
+  static struct timeval what##_timeout = { interval_sec, interval_usec };    \
+  static void what##_schedule(void);                                        \
+  static void *what##_timedout(oop_source *lp, struct timeval tv, void *u) { \
+    body;                                                                   \
+    what##_schedule();                                                      \
+    return OOP_CONTINUE;                                                    \
+  }                                                                         \
+  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);
+}));
+
+static char *debug_report_ipf(InputFile *ipf) {
+  if (!ipf) return xasprintf("-");
+
+  const char *slash= strrchr(ipf->path,'/');
+  const char *path= slash ? slash+1 : ipf->path;
+
+  return xasprintf("%p/%s:ip=%ld,off=%ld,fd=%d%s",
+                  ipf, path,
+                  ipf->inprogress, (long)ipf->offset,
+                  ipf->fd, ipf->rd ? "+" : "");
+}
+
+EVERY(period, -1,0, ({
+  char *dipf_main=     debug_report_ipf(main_input_file);
+  char *dipf_flushing= debug_report_ipf(flushing_input_file);
+  char *dipf_backlog=  debug_report_ipf(backlog_input_file);
+
+  debug("PERIOD"
+       " sms=%s[%d] conns=%d queue=%d until_connect=%d"
+       " input_files main:%s old:%s flushing:%s"
+       " children connecting=%ld inndcomm_child=%ld"
+       ,
+       sms_names[sms], sm_period_counter,
+         conns.count, queue.count, until_connect,
+       dipf_main, dipf_flushing, dipf_backlog,
+       (long)connecting_child, (long)inndcomm_child
+       );
+
+  free(dipf_main);
+  free(dipf_flushing);
+  free(dipf_backlog);
+
+  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_assign_articles();
+  check_idle_conns();
+}));
+
+
+/*========== option parsing ==========*/
+
+static void vbadusage(const char *fmt, va_list al) NORET_PRINTF(1,0);
+static void vbadusage(const char *fmt, va_list al) {
+  char *m= xvasprintf(fmt,al);
+  fprintf(stderr, "bad usage: %s\n"
+         "say --help for help, or read the manpage\n",
+         m);
+  if (become_daemon)
+    syslog(LOG_CRIT,"innduct: invoked with bad usage: %s",m);
+  exit(8);
+}
+
+/*---------- generic option parser ----------*/
+
+static void badusage(const char *fmt, ...) NORET_PRINTF(1,2);
+static void badusage(const char *fmt, ...) {
+  va_list al;
+  va_start(al,fmt);
+  vbadusage(fmt,al);
+}
+
+enum OptFlags {
+  of_seconds= 001000u,
+  of_boolean= 002000u,
+};
+
+typedef struct Option Option;
+typedef void OptionParser(const Option*, const char *val);
+
+struct Option {
+  int shrt;
+  const char *lng, *formarg;
+  void *store;
+  OptionParser *fn;
+  int intval;
+};
+
+static void parse_options(const Option *options, char ***argvp) {
+  /* on return *argvp is first non-option arg; argc is not updated */
+
+  for (;;) {
+    const char *arg= *++(*argvp);
+    if (!arg) break;
+    if (*arg != '-') break;
+    if (!strcmp(arg,"--")) { arg= *++(*argvp); 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->shrt || o->lng; o++)
+         if (strlen(o->lng) == len && !memcmp(o->lng,arg,len))
+           goto found_long;
+       badusage("unknown long option --%s",arg);
+      found_long:
+       if (!o->formarg) {
+         if (equals) badusage("option --%s does not take a value",o->lng);
+         arg= 0;
+       } else if (equals) {
+         arg= equals+1;
+       } else {
+         arg= *++(*argvp);
+         if (!arg) badusage("option --%s needs a value for %s",
+                            o->lng, o->formarg);
+       }
+       o->fn(o, arg);
+       break; /* eaten the whole argument now */
+      }
+      for (o=options; o->shrt || o->lng; o++)
+       if (a == o->shrt)
+         goto found_short;
+      badusage("unknown short option -%c",a);
+    found_short:
+      if (!o->formarg) {
+       o->fn(o,0);
+      } else {
+       if (!*++arg) {
+         arg= *++(*argvp);
+         if (!arg) badusage("option -%c needs a value for %s",
+                            o->shrt, o->formarg);
+       }
+       o->fn(o,arg);
+       break; /* eaten the whole argument now */
+      }
+    }
+  }
+}
+
+#define DELIMPERHAPS(delim,str)  (str) ? (delim) : "", (str) ? (str) : ""
+
+static void print_options(const Option *options, FILE *f) {
+  const Option *o;
+  for (o=options; o->shrt || o->lng; o++) {
+    char shrt[2] = { o->shrt, 0 };
+    char *optspec= xasprintf("%s%s%s%s%s",
+                            o->shrt ? "-" : "", shrt,
+                            o->shrt && o->lng ? "|" : "",
+                            DELIMPERHAPS("--", o->lng));
+    fprintf(f, "  %s%s%s\n", optspec, DELIMPERHAPS(" ", o->formarg));
+    free(optspec);
+  }
+}
+
+/*---------- specific option types ----------*/
+
+static 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->lng);
+  int *store= o->store;
+  *store= ul;
+}
+
+static 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->lng);
+}
+
+static void op_string(const Option *o, const char *val) {
+  const char **store= o->store;
+  *store= val;
+}
+
+static void op_seconds(const Option *o, const char *val) {
+  int *store= o->store;
+  char *ep;
+  int unit;
+
+  double v= strtod(val,&ep);
+  if (ep==val) badusage("bad time/duration value for %s",o->lng);
+
+  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->lng);
+
+  v *= unit;
+  v= ceil(v);
+  if (v > INT_MAX) badusage("time/duration value for %s out of range",o->lng);
+  *store= v;
+}
+
+static void op_setint(const Option *o, const char *val) {
+  int *store= o->store;
+  *store= o->intval;
+}
+
+/*---------- specific options ----------*/
+
+static void help(const Option *o, const char *val);
+
+static const Option innduct_options[]= {
+{'f',"feedfile",         "F",     &feedfile,                 op_string      },
+{'q',"quiet-multiple",   0,       &quiet_multiple,           op_setint, 1   },
+{0,"no-daemon",          0,       &become_daemon,            op_setint, 0   },
+{0,"no-streaming",       0,       &try_stream,               op_setint, 0   },
+{0,"inndconf",           "F",     &inndconffile,             op_string      },
+{'P',"port",             "PORT",  &port,                     op_integer     },
+{0,"help",               0,       0,                         help           },
+
+{0,"max-connections",    "N",     &max_connections,          op_integer     },
+{0,"max-queue-per-conn", "N",     &max_queue_per_conn,       op_integer     },
+{0,"feedfile-flush-size","BYTES", &target_max_feedfile_size, op_integer     },
+{0,"period-interval",    "TIME",  &period_seconds,           op_seconds     },
+
+{0,"connection-timeout", "TIME",  &connection_setup_timeout, op_seconds     },
+{0,"stuck-flush-timeout","TIME",  &inndcomm_flush_timeout,   op_seconds     },
+
+{0,"no-check-proportion",   "PERCENT",   &nocheck_thresh,       op_double   },
+{0,"no-check-response-time","ARTICLES",  &nocheck_decay,        op_double   },
+
+{0,"reconnect-interval",     "PERIOD", &reconnect_delay_periods,  op_seconds },
+{0,"flush-retry-interval",   "PERIOD", &flushfail_retry_periods,  op_seconds },
+{0,"earliest-deferred-retry","PERIOD", &backlog_retry_minperiods, op_seconds },
+{0,"backlog-rescan-interval","PERIOD",&backlog_spontrescan_periods,op_seconds},
+{0,"max-flush-interval",     "PERIOD", &spontaneous_flush_periods,op_seconds },
+{0,"idle-timeout",           "PERIOD", &need_activity_periods,    op_seconds },
+
+{0,"max-bad-input-data-ratio","PERCENT", &max_bad_data_ratio,   op_double    },
+{0,"max-bad-input-data-init", "PERCENT", &max_bad_data_initial, op_integer   },
+
+{0,0}
+};
+
+static void printusage(FILE *f) {
+  fputs("usage: innduct [options] site [fqdn]\n"
+       "available options are:\n", f);
+  print_options(innduct_options, f);
+}
+
+static void help(const Option *o, const char *val) {
+  printusage(stdout);
+  if (ferror(stdout) || fflush(stdout)) {
+    perror("innduct: writing help");
+    exit(12);
+  }
+  exit(0);
+}
+
+static void convert_to_periods_rndup(int *store) {
+  *store += period_seconds-1;
+  *store /= period_seconds;
+}
+
+int main(int argc, char **argv) {
+  if (!argv[1]) {
+    printusage(stderr);
+    exit(12);
+  }
+
+  parse_options(innduct_options, &argv);
+
+  /* arguments */
+
+  sitename= *argv++;
+  if (!sitename) badusage("need site name argument");
+  remote_host= *argv++;
+  if (*argv) badusage("too many non-option arguments");
+
+  /* defaults */
+
+  if (!remote_host) remote_host= sitename;
+
+  if (nocheck_thresh < 0 || nocheck_thresh > 100)
+    badusage("nocheck threshold percentage must be between 0..100");
+  nocheck_thresh *= 0.01;
+
+  if (nocheck_decay < 0.1)
+    badusage("nocheck decay articles must be at least 0.1");
+  nocheck_decay= 1 - 1.0/nocheck_decay;
+
+  convert_to_periods_rndup(&reconnect_delay_periods);
+  convert_to_periods_rndup(&flushfail_retry_periods);
+  convert_to_periods_rndup(&backlog_retry_minperiods);
+  convert_to_periods_rndup(&backlog_spontrescan_periods);
+  convert_to_periods_rndup(&spontaneous_flush_periods);
+  convert_to_periods_rndup(&need_activity_periods);
+
+  if (max_bad_data_ratio < 0 || max_bad_data_ratio > 100)
+    badusage("bad input data ratio must be between 0..100");
+  max_bad_data_ratio *= 0.01;
+
+  if (!feedfile) {
+    innconf_read(inndconffile);
+    feedfile= xasprintf("%s/%s",innconf->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);
+
+  /* set things up */
+
+  oop_source_sys *sysloop= oop_sys_new();
+  if (!sysloop) sysdie("could not create liboop event loop");
+  loop= (oop_source*)sysloop;
+
+  if (signal(SIGPIPE, SIG_IGN) == SIG_ERR)
+    sysdie("could not ignore SIGPIPE");
+
+  LIST_INIT(conns);
+  LIST_INIT(queue);
+
+  if (become_daemon) {
+    int i;
+    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);
+    xclose(null, "/dev/null original fd",0);
+
+    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_method_init()) {
+    warn("no file monitoring available, polling");
+    filepoll_schedule();
+  }
+
+  period_timeout.tv_sec= period_seconds;
+  period_schedule();
+
+  statemc_init();
+
+  /* let's go */
+
+  void *r= oop_sys_run(sysloop);
+  assert(r == OOP_ERROR);
+  sysdie("event loop failed");
+}