A working implementation that uses a pthread to start the receiver

[rsync/rsync-patches.git] / threaded-receiver.diff

This patch changes the receiving side to have the receiving code use a thread
instead of a forked process.  This extra thread does read from the socket, but
it sends any stdout/stderr messages to the generator (main thread) to output.

** This is very new code. **  Yes, it passes the "make test" testsuite, but
there may still be some problems, especially in some of the untested features.
(For one thing, I haven't yet added code to properly handle any keep-alive
messages that arrive on the receiving side during the --delete-after phase!)

This code just uses pthread.h directly, so configure changes will probably be
needed to make this compatible with more systems.  I have also tested that
this code works fine using the GNU pth library without any code changes if
you configured it with --enable-syscall-soft --enable-pthread (you may need
to twiddle the Makefile options if you didn't install the library, though).

If you try this out, please send some email to wayned@samba.org or the rsync
mailing list with your results, build changes, bug reports, etc.  Thanks!

Be sure to run "make proto" before running "make".

--- orig/Makefile.in	2005-11-07 04:29:00
+++ Makefile.in	2005-12-10 18:35:39
@@ -6,7 +6,7 @@ exec_prefix=@exec_prefix@
 bindir=@bindir@
 mandir=@mandir@
 
-LIBS=@LIBS@
+LIBS=@LIBS@ -lpthread
 CC=@CC@
 CFLAGS=@CFLAGS@
 CPPFLAGS=@CPPFLAGS@
--- orig/cleanup.c	2005-11-10 16:58:36
+++ cleanup.c	2005-12-08 23:17:08
@@ -94,9 +94,6 @@ void _exit_cleanup(int code, const char 
 	}
 	inside_cleanup++;
 
-	signal(SIGUSR1, SIG_IGN);
-	signal(SIGUSR2, SIG_IGN);
-
 	if (verbose > 3) {
 		rprintf(FINFO,"_exit_cleanup(code=%d, file=%s, line=%d): entered\n",
 			code, safe_fname(file), line);
@@ -127,8 +124,6 @@ void _exit_cleanup(int code, const char 
 	io_flush(FULL_FLUSH);
 	if (cleanup_fname)
 		do_unlink(cleanup_fname);
-	if (code)
-		kill_all(SIGUSR1);
 	if (cleanup_pid && cleanup_pid == getpid()) {
 		char *pidf = lp_pid_file();
 		if (pidf && *pidf)
--- orig/generator.c	2005-11-12 20:31:04
+++ generator.c	2005-12-08 23:17:08
@@ -65,7 +65,6 @@ extern OFF_T min_size;
 extern int io_error;
 extern int allowed_lull;
 extern int sock_f_out;
-extern int ignore_timeout;
 extern int protocol_version;
 extern int fuzzy_basis;
 extern int always_checksum;
@@ -94,6 +93,11 @@ extern struct filter_list_struct server_
 
 static int deletion_count = 0; /* used to implement --max-delete */
 
+/* These vars are local copies so that the receiver can use the originals. */
+static int GEN_append_mode;
+static int GEN_make_backups;
+static int GEN_csum_length;
+
 /* For calling delete_file() */
 #define DEL_FORCE_RECURSE	(1<<1) /* recurse even w/o --force */
 #define DEL_TERSE		(1<<3)
@@ -446,8 +450,8 @@ static void sum_sizes_sqroot(struct sum_
 	}
 
 	if (protocol_version < 27) {
-		s2length = csum_length;
-	} else if (csum_length == SUM_LENGTH) {
+		s2length = GEN_csum_length;
+	} else if (GEN_csum_length == SUM_LENGTH) {
 		s2length = SUM_LENGTH;
 	} else {
 		int32 c;
@@ -457,7 +461,7 @@ static void sum_sizes_sqroot(struct sum_
 		for (c = blength; c >>= 1 && b; b--) {}
 		/* add a bit, subtract rollsum, round up. */
 		s2length = (b + 1 - 32 + 7) / 8; /* --optimize in compiler-- */
-		s2length = MAX(s2length, csum_length);
+		s2length = MAX(s2length, GEN_csum_length);
 		s2length = MIN(s2length, SUM_LENGTH);
 	}
 
@@ -491,7 +495,7 @@ static void generate_and_send_sums(int f
 	sum_sizes_sqroot(&sum, len);
 	write_sum_head(f_out, &sum);
 
-	if (append_mode > 0 && f_copy < 0)
+	if (GEN_append_mode > 0 && f_copy < 0)
 		return;
 
 	if (len > 0)
@@ -510,7 +514,7 @@ static void generate_and_send_sums(int f
 
 		if (f_copy >= 0) {
 			full_write(f_copy, map, n1);
-			if (append_mode > 0)
+			if (GEN_append_mode > 0)
 				continue;
 		}
 
@@ -1063,7 +1067,7 @@ static void recv_generator(char *fname, 
 		return;
 	}
 
-	if (append_mode && st.st_size > file->length)
+	if (GEN_append_mode && st.st_size > file->length)
 		return;
 
 	if (!compare_dest && fnamecmp_type <= FNAMECMP_BASIS_DIR_HIGH)
@@ -1124,7 +1128,7 @@ static void recv_generator(char *fname, 
 		goto notify_others;
 	}
 
-	if (inplace && make_backups && fnamecmp_type == FNAMECMP_FNAME) {
+	if (inplace && GEN_make_backups && fnamecmp_type == FNAMECMP_FNAME) {
 		if (!(backupptr = get_backup_name(fname))) {
 			close(fd);
 			return;
@@ -1214,7 +1218,10 @@ void generate_files(int f_out, struct fi
 	int save_ignore_existing = ignore_existing;
 	int save_ignore_non_existing = ignore_non_existing;
 	int save_do_progress = do_progress;
-	int save_make_backups = make_backups;
+	int save_make_backups = GEN_make_backups = make_backups;
+
+	GEN_append_mode = append_mode;
+	GEN_csum_length = csum_length;
 
 	if (protocol_version >= 29) {
 		itemizing = 1;
@@ -1243,7 +1250,7 @@ void generate_files(int f_out, struct fi
 		do_delete_pass(flist);
 	do_progress = 0;
 
-	if (append_mode || whole_file < 0)
+	if (GEN_append_mode || whole_file < 0)
 		whole_file = 0;
 	if (verbose >= 2) {
 		rprintf(FINFO, "delta-transmission %s\n",
@@ -1252,12 +1259,6 @@ void generate_files(int f_out, struct fi
 			: "enabled");
 	}
 
-	/* Since we often fill up the outgoing socket and then just sit around
-	 * waiting for the other 2 processes to do their thing, we don't want
-	 * to exit on a timeout.  If the data stops flowing, the receiver will
-	 * notice that and let us know via the redo pipe (or its closing). */
-	ignore_timeout = 1;
-
 	for (i = 0; i < flist->count; i++) {
 		struct file_struct *file = flist->files[i];
 
@@ -1301,23 +1302,34 @@ void generate_files(int f_out, struct fi
 		delete_in_dir(NULL, NULL, NULL);
 
 	phase++;
-	csum_length = SUM_LENGTH;
+	GEN_csum_length = SUM_LENGTH; /* csum_length is set by the receiver */
 	max_size = min_size = ignore_existing = ignore_non_existing = 0;
 	update_only = always_checksum = size_only = 0;
 	ignore_times = 1;
-	if (append_mode)  /* resend w/o append mode */
-		append_mode = -1; /* ... but only longer files */
-	make_backups = 0; /* avoid a duplicate backup for inplace processing */
+	if (GEN_append_mode)  /* resend w/o append mode */
+		GEN_append_mode = -1; /* ... but only longer files */
+	GEN_make_backups = 0; /* avoid a duplicate backup for inplace processing */
 
 	if (verbose > 2)
 		rprintf(FINFO,"generate_files phase=%d\n",phase);
 
 	write_int(f_out, -1);
+	io_flush(NORMAL_FLUSH);
 
 	/* files can cycle through the system more than once
 	 * to catch initial checksum errors */
-	while ((i = get_redo_num(itemizing, code)) != -1) {
-		struct file_struct *file = flist->files[i];
+	while (1) {
+		struct file_struct *file;
+		if (preserve_hard_links)
+			check_for_finished_hlinks(itemizing, code);
+		if ((i = get_redo_num()) < 0) {
+			if (i == -2)
+				break;
+			io_flush(NORMAL_FLUSH);
+			msleep(20);
+			continue;
+		}
+		file = flist->files[i];
 		if (local_name)
 			strlcpy(fbuf, local_name, sizeof fbuf);
 		else
@@ -1329,27 +1341,43 @@ void generate_files(int f_out, struct fi
 	phase++;
 	ignore_non_existing = save_ignore_non_existing;
 	ignore_existing = save_ignore_existing;
-	make_backups = save_make_backups;
+	GEN_make_backups = save_make_backups;
 
 	if (verbose > 2)
 		rprintf(FINFO,"generate_files phase=%d\n",phase);
 
 	write_int(f_out, -1);
+	io_flush(NORMAL_FLUSH);
+
 	/* Reduce round-trip lag-time for a useless delay-updates phase. */
-	if (protocol_version >= 29 && !delay_updates)
+	if (protocol_version >= 29 && !delay_updates) {
 		write_int(f_out, -1);
+		io_flush(NORMAL_FLUSH);
+	}
 
-	/* Read MSG_DONE for the redo phase (and any prior messages). */
-	get_redo_num(itemizing, code);
+	/* Read end marker for the redo phase (and any prior messages). */
+	while (1) {
+		if (preserve_hard_links)
+			check_for_finished_hlinks(itemizing, code);
+		if (get_redo_num() == -2)
+			break;
+		io_flush(NORMAL_FLUSH);
+		msleep(20);
+	}
 
 	if (protocol_version >= 29) {
 		phase++;
 		if (verbose > 2)
 			rprintf(FINFO, "generate_files phase=%d\n", phase);
-		if (delay_updates)
+		if (delay_updates) {
 			write_int(f_out, -1);
-		/* Read MSG_DONE for delay-updates phase & prior messages. */
-		get_redo_num(itemizing, code);
+			io_flush(NORMAL_FLUSH);
+		}
+		/* Read end marker for delay-updates phase & prior messages. */
+		while (get_redo_num() != -2) {
+			io_flush(NORMAL_FLUSH);
+			msleep(20);
+		}
 	}
 
 	do_progress = save_do_progress;
--- orig/io.c	2005-12-08 21:19:31
+++ io.c	2005-12-10 19:03:08
@@ -47,7 +47,6 @@ extern int allowed_lull;
 extern int am_server;
 extern int am_daemon;
 extern int am_sender;
-extern int am_generator;
 extern int eol_nulls;
 extern int read_batch;
 extern int csum_length;
@@ -60,7 +59,6 @@ extern struct stats stats;
 extern struct file_list *the_file_list;
 
 const char phase_unknown[] = "unknown";
-int ignore_timeout = 0;
 int batch_fd = -1;
 int batch_gen_fd = -1;
 
@@ -84,7 +82,6 @@ const char *io_read_phase = phase_unknow
 int kluge_around_eof = 0;
 
 int msg_fd_in = -1;
-int msg_fd_out = -1;
 int sock_f_in = -1;
 int sock_f_out = -1;
 
@@ -109,27 +106,32 @@ static int select_timeout = SELECT_TIMEO
 static void read_loop(int fd, char *buf, size_t len);
 
 struct flist_ndx_item {
-	struct flist_ndx_item *next;
+	volatile struct flist_ndx_item *next;
 	int ndx;
 };
 
 struct flist_ndx_list {
-	struct flist_ndx_item *head, *tail;
+	volatile struct flist_ndx_item *head, *tail;
+	pthread_mutex_t mutex;
 };
 
-static struct flist_ndx_list redo_list, hlink_list;
+static struct flist_ndx_list redo_list = { NULL, NULL, PTHREAD_MUTEX_INITIALIZER };
+static struct flist_ndx_list hlink_list = { NULL, NULL, PTHREAD_MUTEX_INITIALIZER };
 
 struct msg_list_item {
-	struct msg_list_item *next;
+	volatile struct msg_list_item *next;
+	pthread_mutex_t mutex;
 	char *buf;
 	int len;
+	enum msgcode code;
 };
 
 struct msg_list {
-	struct msg_list_item *head, *tail;
+	volatile struct msg_list_item *head, *tail;
+	pthread_mutex_t mutex;
 };
 
-static struct msg_list msg_list;
+static struct msg_list msg_list = { NULL, NULL, PTHREAD_MUTEX_INITIALIZER };
 
 static void flist_ndx_push(struct flist_ndx_list *lp, int ndx)
 {
@@ -139,27 +141,31 @@ static void flist_ndx_push(struct flist_
 		out_of_memory("flist_ndx_push");
 	item->next = NULL;
 	item->ndx = ndx;
+	pthread_mutex_lock(&redo_list.mutex);
 	if (lp->tail)
 		lp->tail->next = item;
 	else
 		lp->head = item;
 	lp->tail = item;
+	pthread_mutex_unlock(&redo_list.mutex);
 }
 
 static int flist_ndx_pop(struct flist_ndx_list *lp)
 {
-	struct flist_ndx_item *next;
+	struct flist_ndx_item *head, *next;
 	int ndx;
 
 	if (!lp->head)
 		return -1;
 
-	ndx = lp->head->ndx;
-	next = lp->head->next;
-	free(lp->head);
-	lp->head = next;
-	if (!next)
+	pthread_mutex_lock(&hlink_list.mutex);
+	head = (struct flist_ndx_item *)lp->head;
+	next = (struct flist_ndx_item *)head->next;
+	ndx = head->ndx;
+	if (!(lp->head = next))
 		lp->tail = NULL;
+	pthread_mutex_unlock(&hlink_list.mutex);
+	free(head);
 
 	return ndx;
 }
@@ -168,7 +174,7 @@ static void check_timeout(void)
 {
 	time_t t;
 
-	if (!io_timeout || ignore_timeout)
+	if (!io_timeout)
 		return;
 
 	if (!last_io_in) {
@@ -209,45 +215,40 @@ void set_io_timeout(int secs)
 
 /* Setup the fd used to receive MSG_* messages.  Only needed during the
  * early stages of being a local sender (up through the sending of the
- * file list) or when we're the generator (to fetch the messages from
- * the receiver). */
+ * file list). */
 void set_msg_fd_in(int fd)
 {
 	msg_fd_in = fd;
 }
 
-/* Setup the fd used to send our MSG_* messages.  Only needed when
- * we're the receiver (to send our messages to the generator). */
-void set_msg_fd_out(int fd)
-{
-	msg_fd_out = fd;
-	set_nonblocking(msg_fd_out);
-}
-
 /* Add a message to the pending MSG_* list. */
 static void msg_list_add(int code, char *buf, int len)
 {
 	struct msg_list_item *ml;
 
+	assert(am_receiver());
 	if (!(ml = new(struct msg_list_item)))
 		out_of_memory("msg_list_add");
 	ml->next = NULL;
-	if (!(ml->buf = new_array(char, len+4)))
+	/* NOTE: the "+ 1" allows rwrite() to use the buf! */
+	if (!(ml->buf = new_array(char, len + 1)))
 		out_of_memory("msg_list_add");
-	SIVAL(ml->buf, 0, ((code+MPLEX_BASE)<<24) | len);
-	memcpy(ml->buf+4, buf, len);
-	ml->len = len+4;
+	memcpy(ml->buf, buf, len);
+	ml->len = len;
+	ml->code = code;
+
+	pthread_mutex_lock(&msg_list.mutex);
 	if (msg_list.tail)
 		msg_list.tail->next = ml;
 	else
 		msg_list.head = ml;
 	msg_list.tail = ml;
+	pthread_mutex_unlock(&msg_list.mutex);
 }
 
-/* Read a message from the MSG_* fd and handle it.  This is called either
+/* Read a message from the MSG_* fd and handle it.  This is only called
  * during the early stages of being a local sender (up through the sending
- * of the file list) or when we're the generator (to fetch the messages
- * from the receiver). */
+ * of the file list). */
 static void read_msg_fd(void)
 {
 	char buf[2048];
@@ -266,40 +267,6 @@ static void read_msg_fd(void)
 	tag = (tag >> 24) - MPLEX_BASE;
 
 	switch (tag) {
-	case MSG_DONE:
-		if (len != 0 || !am_generator) {
-			rprintf(FERROR, "invalid message %d:%d\n", tag, len);
-			exit_cleanup(RERR_STREAMIO);
-		}
-		flist_ndx_push(&redo_list, -1);
-		break;
-	case MSG_REDO:
-		if (len != 4 || !am_generator) {
-			rprintf(FERROR, "invalid message %d:%d\n", tag, len);
-			exit_cleanup(RERR_STREAMIO);
-		}
-		read_loop(fd, buf, 4);
-		flist_ndx_push(&redo_list, IVAL(buf,0));
-		break;
-	case MSG_DELETED:
-		if (len >= (int)sizeof buf || !am_generator) {
-			rprintf(FERROR, "invalid message %d:%d\n", tag, len);
-			exit_cleanup(RERR_STREAMIO);
-		}
-		read_loop(fd, buf, len);
-		io_multiplex_write(MSG_DELETED, buf, len);
-		break;
-	case MSG_SUCCESS:
-		if (len != 4 || !am_generator) {
-			rprintf(FERROR, "invalid message %d:%d\n", tag, len);
-			exit_cleanup(RERR_STREAMIO);
-		}
-		read_loop(fd, buf, len);
-		if (remove_sent_files)
-			io_multiplex_write(MSG_SUCCESS, buf, len);
-		if (preserve_hard_links)
-			flist_ndx_push(&hlink_list, IVAL(buf,0));
-		break;
 	case MSG_INFO:
 	case MSG_ERROR:
 	case MSG_LOG:
@@ -320,71 +287,72 @@ static void read_msg_fd(void)
 	msg_fd_in = fd;
 }
 
-/* Try to push messages off the list onto the wire.  If we leave with more
+/* Try to pop messages off the list onto the wire.  If we leave with more
  * to do, return 0.  On error, return -1.  If everything flushed, return 1.
- * This is only active in the receiver. */
+ * This is only called by the generator. */
 static int msg_list_flush(int flush_it_all)
 {
-	static int written = 0;
-	struct timeval tv;
-	fd_set fds;
-
-	if (msg_fd_out < 0)
-		return -1;
-
+	assert(am_generator());
+	no_flush++;
 	while (msg_list.head) {
-		struct msg_list_item *ml = msg_list.head;
-		int n = write(msg_fd_out, ml->buf + written, ml->len - written);
-		if (n < 0) {
-			if (errno == EINTR)
-				continue;
-			if (errno != EWOULDBLOCK && errno != EAGAIN)
-				return -1;
-			if (!flush_it_all)
-				return 0;
-			FD_ZERO(&fds);
-			FD_SET(msg_fd_out, &fds);
-			tv.tv_sec = select_timeout;
-			tv.tv_usec = 0;
-			if (!select(msg_fd_out+1, NULL, &fds, NULL, &tv))
-				check_timeout();
-		} else if ((written += n) == ml->len) {
-			free(ml->buf);
-			msg_list.head = ml->next;
-			if (!msg_list.head)
-				msg_list.tail = NULL;
-			free(ml);
-			written = 0;
+		struct msg_list_item *ml = (struct msg_list_item *)msg_list.head;
+		switch (ml->code) {
+		case MSG_INFO:
+		case MSG_ERROR:
+		case MSG_LOG:
+			rwrite(ml->code, ml->buf, ml->len);
+			break;
+		default:
+			io_multiplex_write(ml->code, ml->buf, ml->len);
+			break;
 		}
+		pthread_mutex_lock(&msg_list.mutex);
+		if (!(msg_list.head = ml->next))
+			msg_list.tail = NULL;
+		pthread_mutex_unlock(&msg_list.mutex);
+		free(ml->buf);
+		free(ml);
+		if (!flush_it_all)
+			break;
 	}
+	no_flush--;
+
 	return 1;
 }
 
 void send_msg(enum msgcode code, char *buf, int len)
 {
-	if (msg_fd_out < 0) {
+	if (am_receiver())
+		msg_list_add(code, buf, len);
+	else
 		io_multiplex_write(code, buf, len);
-		return;
-	}
-	msg_list_add(code, buf, len);
-	msg_list_flush(NORMAL_FLUSH);
 }
 
-int get_redo_num(int itemizing, enum logcode code)
+/* This is only used by the receiver. */
+void push_redo_num(int ndx)
 {
-	while (1) {
-		if (hlink_list.head)
-			check_for_finished_hlinks(itemizing, code);
-		if (redo_list.head)
-			break;
-		read_msg_fd();
-	}
+	assert(am_receiver());
+	flist_ndx_push(&redo_list, ndx);
+}
 
+/* This is only used by the generator. */
+int get_redo_num(void)
+{
+	assert(am_generator());
 	return flist_ndx_pop(&redo_list);
 }
 
+/* This is only used by the receiver. */
+void push_hlink_num(int ndx)
+{
+	assert(am_receiver());
+	flist_ndx_push(&hlink_list, ndx);
+}
+
+/* This is only used by the generator. */
 int get_hlink_num(void)
 {
+	assert(am_generator());
 	return flist_ndx_pop(&hlink_list);
 }
 
@@ -465,11 +433,6 @@ static int read_timeout(int fd, char *bu
 		FD_ZERO(&r_fds);
 		FD_ZERO(&w_fds);
 		FD_SET(fd, &r_fds);
-		if (msg_list.head) {
-			FD_SET(msg_fd_out, &w_fds);
-			if (msg_fd_out > maxfd)
-				maxfd = msg_fd_out;
-		}
 		if (io_filesfrom_f_out >= 0) {
 			int new_fd;
 			if (io_filesfrom_buflen == 0) {
@@ -502,9 +465,6 @@ static int read_timeout(int fd, char *bu
 			continue;
 		}
 
-		if (msg_list.head && FD_ISSET(msg_fd_out, &w_fds))
-			msg_list_flush(NORMAL_FLUSH);
-
 		if (io_filesfrom_f_out >= 0) {
 			if (io_filesfrom_buflen) {
 				if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
@@ -832,6 +792,8 @@ static void readfd(int fd, char *buffer,
 	}
 
 	if (fd == write_batch_monitor_in) {
+		if (am_generator())
+		    rprintf(FINFO, "writing %d bytes to batch file from generator\n", total);
 		if ((size_t)write(batch_fd, buffer, total) != total)
 			exit_cleanup(RERR_FILEIO);
 	}
@@ -1091,7 +1053,6 @@ static void writefd_unbuffered(int fd,ch
 			 * to grab any messages they sent before they died. */
 			while (fd == sock_f_out && io_multiplexing_in) {
 				set_io_timeout(30);
-				ignore_timeout = 0;
 				readfd_unbuffered(sock_f_in, io_filesfrom_buf,
 						  sizeof io_filesfrom_buf);
 			}
@@ -1101,7 +1062,7 @@ static void writefd_unbuffered(int fd,ch
 		total += ret;
 
 		if (fd == sock_f_out) {
-			if (io_timeout || am_generator)
+			if (io_timeout || am_generator())
 				last_io_out = time(NULL);
 			sleep_for_bwlimit(ret);
 		}
@@ -1126,7 +1087,7 @@ static void mplex_write(enum msgcode cod
 	 * cause output to occur down the socket.  Setting contiguous_write_len
 	 * prevents the reading of msg_fd_in once we actually start to write
 	 * this sequence of data (though we might read it before the start). */
-	if (am_generator && msg_fd_in >= 0)
+	if (am_generator() && msg_fd_in >= 0)
 		contiguous_write_len = len + 4;
 
 	if (n > sizeof buffer - 4)
@@ -1141,33 +1102,31 @@ static void mplex_write(enum msgcode cod
 	if (len)
 		writefd_unbuffered(sock_f_out, buf, len);
 
-	if (am_generator && msg_fd_in >= 0)
+	if (am_generator() && msg_fd_in >= 0)
 		contiguous_write_len = 0;
 }
 
 
 void io_flush(int flush_it_all)
 {
-	msg_list_flush(flush_it_all);
-
-	if (!iobuf_out_cnt || no_flush)
+	if (no_flush)
 		return;
 
-	if (io_multiplexing_out)
-		mplex_write(MSG_DATA, iobuf_out, iobuf_out_cnt);
-	else
-		writefd_unbuffered(sock_f_out, iobuf_out, iobuf_out_cnt);
-	iobuf_out_cnt = 0;
+	if (iobuf_out_cnt) {
+		if (io_multiplexing_out)
+			mplex_write(MSG_DATA, iobuf_out, iobuf_out_cnt);
+		else
+			writefd_unbuffered(sock_f_out, iobuf_out, iobuf_out_cnt);
+		iobuf_out_cnt = 0;
+	}
+
+	if (am_generator())
+		msg_list_flush(flush_it_all);
 }
 
 
 static void writefd(int fd,char *buf,size_t len)
 {
-	if (fd == msg_fd_out) {
-		rprintf(FERROR, "Internal error: wrong write used in receiver.\n");
-		exit_cleanup(RERR_PROTOCOL);
-	}
-
 	if (fd == sock_f_out)
 		stats.total_written += len;
 
@@ -1387,9 +1346,3 @@ void start_write_batch(int fd)
 	else
 		write_batch_monitor_in = fd;
 }
-
-void stop_write_batch(void)
-{
-	write_batch_monitor_out = -1;
-	write_batch_monitor_in = -1;
-}
--- orig/log.c	2005-10-26 16:49:08
+++ log.c	2005-12-08 23:17:09
@@ -35,7 +35,6 @@ extern int am_sender;
 extern int local_server;
 extern int quiet;
 extern int module_id;
-extern int msg_fd_out;
 extern int protocol_version;
 extern int preserve_times;
 extern int log_format_has_o_or_i;
@@ -67,7 +66,7 @@ struct {
 	{ RERR_IPC        , "error in IPC code" },
 	{ RERR_CRASHED    , "sibling process crashed" },
 	{ RERR_TERMINATED , "sibling process terminated abnormally" },
-	{ RERR_SIGNAL     , "received SIGUSR1 or SIGINT" },
+	{ RERR_SIGNAL     , "received SIGINT, SIGHUP, or SIGTERM" },
 	{ RERR_WAITCHILD  , "waitpid() failed" },
 	{ RERR_MALLOC     , "error allocating core memory buffers" },
 	{ RERR_PARTIAL    , "some files could not be transferred" },
@@ -204,8 +203,8 @@ void rwrite(enum logcode code, char *buf
 
 	buf[len] = 0;
 
-	if (am_server && msg_fd_out >= 0) {
-		/* Pass the message to our sibling. */
+	if (am_receiver()) {
+		/* Pass the message to the generator thread. */
 		send_msg((enum msgcode)code, buf, len);
 		return;
 	}
--- orig/main.c	2005-12-15 07:55:39
+++ main.c	2005-12-10 19:02:56
@@ -30,7 +30,6 @@ extern int list_only;
 extern int am_root;
 extern int am_server;
 extern int am_sender;
-extern int am_generator;
 extern int am_daemon;
 extern int blocking_io;
 extern int remove_sent_files;
@@ -75,9 +74,20 @@ struct pid_status {
 
 static time_t starttime, endtime;
 static int64 total_read, total_written;
+static pthread_t receiver_tid;
 
 static void show_malloc_stats(void);
 
+int am_generator()
+{
+	return receiver_tid != 0 && pthread_self() != receiver_tid;
+}
+
+int am_receiver()
+{
+	return receiver_tid != 0 && pthread_self() == receiver_tid;
+}
+
 /* Works like waitpid(), but if we already harvested the child pid in our
  * sigchld_handler(), we succeed instead of returning an error. */
 pid_t wait_process(pid_t pid, int *status_ptr, int flags)
@@ -154,7 +164,7 @@ static void handle_stats(int f)
 		show_flist_stats();
 	}
 
-	if (am_generator)
+	if (am_generator())
 		return;
 
 	if (am_daemon) {
@@ -558,12 +568,30 @@ static void do_server_sender(int f_in, i
 	exit_cleanup(0);
 }
 
+struct thread_args {
+    struct file_list *flist;
+    char *local_name;
+    int f_in;
+};
+
+static void *start_receiver_thread(void *arg)
+{
+	static int exit_code;
+	struct thread_args *ta = (struct thread_args *)arg;
+
+	recv_files(ta->f_in, ta->flist, ta->local_name);
+	handle_stats(ta->f_in);
+
+	push_redo_num(-2);
+
+	exit_code = log_got_error ? RERR_PARTIAL : 0;
+	return &exit_code;
+}
 
 static int do_recv(int f_in,int f_out,struct file_list *flist,char *local_name)
 {
-	int pid;
-	int exit_code = 0;
-	int error_pipe[2];
+	void *value_ptr;
+	struct thread_args args;
 
 	/* The receiving side mustn't obey this, or an existing symlink that
 	 * points to an identical file won't be replaced by the referent. */
@@ -572,70 +600,16 @@ static int do_recv(int f_in,int f_out,st
 	if (preserve_hard_links)
 		init_hard_links();
 
-	if (fd_pair(error_pipe) < 0) {
-		rsyserr(FERROR, errno, "pipe failed in do_recv");
+	args.f_in = f_in;
+	args.flist = flist;
+	args.local_name = local_name;
+	if (pthread_create(&receiver_tid, NULL, start_receiver_thread, &args) < 0) {
+		rsyserr(FERROR, errno, "pthread_create failed in do_recv");
 		exit_cleanup(RERR_IPC);
 	}
 
-	io_flush(NORMAL_FLUSH);
-
-	if ((pid = do_fork()) == -1) {
-		rsyserr(FERROR, errno, "fork failed in do_recv");
-		exit_cleanup(RERR_IPC);
-	}
-
-	if (pid == 0) {
-		close(error_pipe[0]);
-		if (f_in != f_out)
-			close(f_out);
-
-		/* we can't let two processes write to the socket at one time */
-		close_multiplexing_out();
-
-		/* set place to send errors */
-		set_msg_fd_out(error_pipe[1]);
-
-		recv_files(f_in, flist, local_name);
-		io_flush(FULL_FLUSH);
-		handle_stats(f_in);
-
-		send_msg(MSG_DONE, "", 0);
-		io_flush(FULL_FLUSH);
-
-		/* Handle any keep-alive packets from the post-processing work
-		 * that the generator does. */
-		if (protocol_version >= 29) {
-			kluge_around_eof = -1;
-
-			/* This should only get stopped via a USR2 signal. */
-			while (read_int(f_in) == flist->count
-			    && read_shortint(f_in) == ITEM_IS_NEW) {}
-
-			rprintf(FERROR, "Invalid packet at end of run [%s]\n",
-				who_am_i());
-			exit_cleanup(RERR_PROTOCOL);
-		}
-
-		/* Finally, we go to sleep until our parent kills us with a
-		 * USR2 signal.  We sleep for a short time, as on some OSes
-		 * a signal won't interrupt a sleep! */
-		while (1)
-			msleep(20);
-	}
-
-	am_generator = 1;
-	close_multiplexing_in();
-	if (write_batch && !am_server)
-		stop_write_batch();
-
-	close(error_pipe[1]);
-	if (f_in != f_out)
-		close(f_in);
-
 	io_start_buffering_out();
 
-	set_msg_fd_in(error_pipe[0]);
-
 	generate_files(f_out, flist, local_name);
 
 	handle_stats(-1);
@@ -646,10 +620,13 @@ static int do_recv(int f_in,int f_out,st
 	}
 	io_flush(FULL_FLUSH);
 
-	set_msg_fd_in(-1);
-	kill(pid, SIGUSR2);
-	wait_process_with_flush(pid, &exit_code);
-	return exit_code;
+	pthread_join(receiver_tid, &value_ptr);
+	if (!am_server)
+		output_summary();
+
+	close_all();
+
+	return *(int*)value_ptr;
 }
 
 
@@ -1011,22 +988,6 @@ static int start_client(int argc, char *
 	return ret;
 }
 
-
-static RETSIGTYPE sigusr1_handler(UNUSED(int val))
-{
-	exit_cleanup(RERR_SIGNAL);
-}
-
-static RETSIGTYPE sigusr2_handler(UNUSED(int val))
-{
-	if (!am_server)
-		output_summary();
-	close_all();
-	if (log_got_error)
-		_exit(RERR_PARTIAL);
-	_exit(0);
-}
-
 static RETSIGTYPE sigchld_handler(UNUSED(int val))
 {
 #ifdef WNOHANG
@@ -1109,8 +1070,6 @@ int main(int argc,char *argv[])
 	int orig_argc = argc;
 	char **orig_argv = argv;
 
-	signal(SIGUSR1, sigusr1_handler);
-	signal(SIGUSR2, sigusr2_handler);
 	signal(SIGCHLD, sigchld_handler);
 #ifdef MAINTAINER_MODE
 	signal(SIGSEGV, rsync_panic_handler);
--- orig/match.c	2005-11-10 16:58:36
+++ match.c	2005-12-08 23:17:09
@@ -21,7 +21,7 @@
 
 extern int verbose;
 extern int am_server;
-extern int do_progress;
+extern int recv_progress;
 extern int checksum_seed;
 extern int append_mode;
 
@@ -133,7 +133,7 @@ static void matched(int f, struct sum_st
 	else
 		last_match = offset;
 
-	if (buf && do_progress)
+	if (buf && recv_progress)
 		show_progress(last_match, buf->file_size);
 }
 
@@ -333,7 +333,7 @@ void match_sums(int f, struct sum_struct
 	if (append_mode) {
 		OFF_T j = 0;
 		for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
-			if (buf && do_progress)
+			if (buf && recv_progress)
 				show_progress(last_match, buf->file_size);
 			sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
 				   CHUNK_SIZE);
@@ -341,7 +341,7 @@ void match_sums(int f, struct sum_struct
 		}
 		if (last_match < s->flength) {
 			int32 len = s->flength - last_match;
-			if (buf && do_progress)
+			if (buf && recv_progress)
 				show_progress(last_match, buf->file_size);
 			sum_update(map_ptr(buf, last_match, len), len);
 			last_match = s->flength;
--- orig/options.c	2005-11-15 18:21:22
+++ options.c	2005-12-08 23:17:09
@@ -69,7 +69,6 @@ int def_compress_level = Z_DEFAULT_COMPR
 int am_root = 0;
 int am_server = 0;
 int am_sender = 0;
-int am_generator = 0;
 int am_starting_up = 1;
 int orig_umask = 0;
 int relative_paths = -1;
@@ -89,6 +88,7 @@ int am_daemon = 0;
 int daemon_over_rsh = 0;
 int do_stats = 0;
 int do_progress = 0;
+int recv_progress = 0;
 int keep_partial = 0;
 int safe_symlinks = 0;
 int copy_unsafe_links = 0;
@@ -1232,6 +1232,7 @@ int parse_arguments(int *argc, const cha
 	if ((do_progress || dry_run) && !verbose && !log_before_transfer
 	    && !am_server)
 		verbose = 1;
+	recv_progress = do_progress;
 
 	if (dry_run)
 		do_xfers = 0;
--- orig/pipe.c	2005-10-24 21:04:45
+++ pipe.c	2005-12-08 23:17:09
@@ -55,7 +55,7 @@ pid_t piped_child(char **command, int *f
 		exit_cleanup(RERR_IPC);
 	}
 
-	pid = do_fork();
+	pid = fork();
 	if (pid == -1) {
 		rsyserr(FERROR, errno, "fork");
 		exit_cleanup(RERR_IPC);
@@ -117,7 +117,7 @@ pid_t local_child(int argc, char **argv,
 		exit_cleanup(RERR_IPC);
 	}
 
-	pid = do_fork();
+	pid = fork();
 	if (pid == -1) {
 		rsyserr(FERROR, errno, "fork");
 		exit_cleanup(RERR_IPC);
--- orig/receiver.c	2005-11-10 16:58:36
+++ receiver.c	2005-12-08 23:17:10
@@ -24,7 +24,7 @@ extern int verbose;
 extern int do_xfers;
 extern int am_daemon;
 extern int am_server;
-extern int do_progress;
+extern int recv_progress;
 extern int log_before_transfer;
 extern int log_format_has_i;
 extern int daemon_log_format_has_i;
@@ -220,7 +220,7 @@ static int receive_data(int f_in, char *
 		if (sum.remainder)
 			sum.flength -= sum.blength - sum.remainder;
 		for (j = CHUNK_SIZE; j < sum.flength; j += CHUNK_SIZE) {
-			if (do_progress)
+			if (recv_progress)
 				show_progress(offset, total_size);
 			sum_update(map_ptr(mapbuf, offset, CHUNK_SIZE),
 				   CHUNK_SIZE);
@@ -228,7 +228,7 @@ static int receive_data(int f_in, char *
 		}
 		if (offset < sum.flength) {
 			int32 len = sum.flength - offset;
-			if (do_progress)
+			if (recv_progress)
 				show_progress(offset, total_size);
 			sum_update(map_ptr(mapbuf, offset, len), len);
 			offset = sum.flength;
@@ -241,7 +241,7 @@ static int receive_data(int f_in, char *
 	}
 
 	while ((i = recv_token(f_in, &data)) != 0) {
-		if (do_progress)
+		if (recv_progress)
 			show_progress(offset, total_size);
 
 		if (i > 0) {
@@ -309,7 +309,7 @@ static int receive_data(int f_in, char *
 		ftruncate(fd, offset);
 #endif
 
-	if (do_progress)
+	if (recv_progress)
 		end_progress(total_size);
 
 	if (fd != -1 && offset > 0 && sparse_end(fd) != 0) {
@@ -362,14 +362,13 @@ static void handle_delayed_updates(struc
 					full_fname(fname),
 					safe_fname(partialptr));
 			} else {
-				if (remove_sent_files
-				    || (preserve_hard_links
-				     && file->link_u.links)) {
+				if (remove_sent_files) {
 					SIVAL(numbuf, 0, i);
 					send_msg(MSG_SUCCESS,numbuf,4);
 				}
-				handle_partial_dir(partialptr,
-						   PDIR_DELETE);
+				if (preserve_hard_links && file->link_u.links)
+					push_hlink_num(i);
+				handle_partial_dir(partialptr, PDIR_DELETE);
 			}
 		}
 	}
@@ -419,11 +418,6 @@ int recv_files(int f_in, struct file_lis
 	if (verbose > 2)
 		rprintf(FINFO,"recv_files(%d) starting\n",flist->count);
 
-	if (flist->hlink_pool) {
-		pool_destroy(flist->hlink_pool);
-		flist->hlink_pool = NULL;
-	}
-
 	if (delay_updates)
 		init_delayed_bits(flist->count);
 
@@ -444,7 +438,7 @@ int recv_files(int f_in, struct file_lis
 				rprintf(FINFO, "recv_files phase=%d\n", phase);
 			if (phase == 2 && delay_updates)
 				handle_delayed_updates(flist, local_name);
-			send_msg(MSG_DONE, "", 0);
+			push_redo_num(-2);
 			if (keep_partial && !partial_dir)
 				make_backups = 0; /* prevents double backup */
 			if (append_mode) {
@@ -665,7 +659,7 @@ int recv_files(int f_in, struct file_lis
 		/* log the transfer */
 		if (log_before_transfer)
 			log_item(file, &initial_stats, iflags, NULL);
-		else if (!am_server && verbose && do_progress)
+		else if (!am_server && verbose && recv_progress)
 			rprintf(FINFO, "%s\n", safe_fname(fname));
 
 		/* recv file data */
@@ -705,11 +699,12 @@ int recv_files(int f_in, struct file_lis
 		cleanup_disable();
 
 		if (recv_ok > 0) {
-			if (remove_sent_files
-			    || (preserve_hard_links && file->link_u.links)) {
+			if (remove_sent_files) {
 				SIVAL(numbuf, 0, i);
 				send_msg(MSG_SUCCESS, numbuf, 4);
 			}
+			if (preserve_hard_links && file->link_u.links)
+				push_hlink_num(i);
 		} else if (!recv_ok) {
 			int msgtype = phase || read_batch ? FERROR : FINFO;
 			if (msgtype == FERROR || verbose) {
@@ -732,10 +727,8 @@ int recv_files(int f_in, struct file_lis
 					errstr, safe_fname(fname),
 					keptstr, redostr);
 			}
-			if (!phase) {
-				SIVAL(numbuf, 0, i);
-				send_msg(MSG_REDO, numbuf, 4);
-			}
+			if (!phase)
+				push_redo_num(i);
 		}
 	}
 	make_backups = save_make_backups;
--- orig/rsync.c	2005-07-27 23:31:12
+++ rsync.c	2005-12-08 23:17:10
@@ -30,7 +30,6 @@ extern int omit_dir_times;
 extern int am_root;
 extern int am_server;
 extern int am_sender;
-extern int am_generator;
 extern int am_starting_up;
 extern int preserve_uid;
 extern int preserve_gid;
@@ -210,5 +209,5 @@ const char *who_am_i(void)
 {
 	if (am_starting_up)
 		return am_server ? "server" : "client";
-	return am_sender ? "sender" : am_generator ? "generator" : "receiver";
+	return am_sender ? "sender" : am_generator() ? "generator" : "receiver";
 }
--- orig/rsync.h	2005-11-12 20:31:04
+++ rsync.h	2005-12-10 19:02:58
@@ -166,10 +166,8 @@ enum msgcode {
 	MSG_DATA=0,	/* raw data on the multiplexed stream */
 	MSG_ERROR=FERROR, MSG_INFO=FINFO, /* remote logging */
 	MSG_LOG=FLOG, MSG_FCLIENT=FCLIENT, /* sibling logging */
-	MSG_REDO=9,	/* reprocess indicated flist index */
 	MSG_SUCCESS=100,/* successfully updated indicated flist index */
 	MSG_DELETED=101,/* successfully deleted a file on receiving side */
-	MSG_DONE=86	/* current phase is done */
 };
 
 #include "errcode.h"
@@ -320,6 +318,7 @@ enum msgcode {
 #endif
 
 #include <assert.h>
+#include <pthread.h>
 
 #include "lib/pool_alloc.h"
 
--- orig/util.c	2005-11-12 20:13:05
+++ util.c	2005-12-08 23:17:10
@@ -405,51 +405,6 @@ int robust_rename(char *from, char *to, 
 	return -1;
 }
 
-
-static pid_t all_pids[10];
-static int num_pids;
-
-/** Fork and record the pid of the child. **/
-pid_t do_fork(void)
-{
-	pid_t newpid = fork();
-
-	if (newpid != 0  &&  newpid != -1) {
-		all_pids[num_pids++] = newpid;
-	}
-	return newpid;
-}
-
-/**
- * Kill all children.
- *
- * @todo It would be kind of nice to make sure that they are actually
- * all our children before we kill them, because their pids may have
- * been recycled by some other process.  Perhaps when we wait for a
- * child, we should remove it from this array.  Alternatively we could
- * perhaps use process groups, but I think that would not work on
- * ancient Unix versions that don't support them.
- **/
-void kill_all(int sig)
-{
-	int i;
-
-	for (i = 0; i < num_pids; i++) {
-		/* Let's just be a little careful where we
-		 * point that gun, hey?  See kill(2) for the
-		 * magic caused by negative values. */
-		pid_t p = all_pids[i];
-
-		if (p == getpid())
-			continue;
-		if (p <= 0)
-			continue;
-
-		kill(p, sig);
-	}
-}
-
-
 /** Turn a user name into a uid */
 int name_to_uid(char *name, uid_t *uid)
 {