X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Fio%2Fpoll.c;h=634f83d286a66669c54694daf2fab0de4ce1a416;hp=7f3df7ed301728965cbc2fc203f0bcc46262ec11;hb=HEAD;hpb=cf86c1e30648c9b482b92444c39ea6ea0ff936a8

diff --git a/ccan/io/poll.c b/ccan/io/poll.c
index 7f3df7ed..634f83d2 100644
--- a/ccan/io/poll.c
+++ b/ccan/io/poll.c
@@ -7,39 +7,64 @@
 #include <sys/types.h>
 #include <sys/socket.h>
 #include <limits.h>
+#include <errno.h>
+#include <ccan/time/time.h>
+#include <ccan/timer/timer.h>
 
-static size_t num_fds = 0, max_fds = 0, num_closing = 0, num_waiting = 0;
+static size_t num_fds = 0, max_fds = 0, num_waiting = 0, num_always = 0, max_always = 0, num_exclusive = 0;
 static struct pollfd *pollfds = NULL;
 static struct fd **fds = NULL;
-static struct timers timeouts;
+static struct io_plan **always = NULL;
+static struct timemono (*nowfn)(void) = time_mono;
+static int (*pollfn)(struct pollfd *fds, nfds_t nfds, int timeout) = poll;
+
+struct timemono (*io_time_override(struct timemono (*now)(void)))(void)
+{
+	struct timemono (*old)(void) = nowfn;
+	nowfn = now;
+	return old;
+}
+
+int (*io_poll_override(int (*poll)(struct pollfd *fds, nfds_t nfds, int timeout)))(struct pollfd *, nfds_t, int)
+{
+	int (*old)(struct pollfd *fds, nfds_t nfds, int timeout) = pollfn;
+	pollfn = poll;
+	return old;
+}
 
 static bool add_fd(struct fd *fd, short events)
 {
+	if (!max_fds) {
+		assert(num_fds == 0);
+		pollfds = tal_arr(NULL, struct pollfd, 8);
+		if (!pollfds)
+			return false;
+		fds = tal_arr(pollfds, struct fd *, 8);
+		if (!fds)
+			return false;
+		max_fds = 8;
+	}
+
 	if (num_fds + 1 > max_fds) {
-		struct pollfd *newpollfds;
-		struct fd **newfds;
-		size_t num = max_fds ? max_fds * 2 : 8;
+		size_t num = max_fds * 2;
 
-		newpollfds = realloc(pollfds, sizeof(*newpollfds) * num);
-		if (!newpollfds)
+		if (!tal_resize(&pollfds, num))
 			return false;
-		pollfds = newpollfds;
-		newfds = realloc(fds, sizeof(*newfds) * num);
-		if (!newfds)
+		if (!tal_resize(&fds, num))
 			return false;
-		fds = newfds;
 		max_fds = num;
 	}
 
 	pollfds[num_fds].events = events;
 	/* In case it's idle. */
 	if (!events)
-		pollfds[num_fds].fd = -fd->fd;
+		pollfds[num_fds].fd = -fd->fd - 1;
 	else
 		pollfds[num_fds].fd = fd->fd;
 	pollfds[num_fds].revents = 0; /* In case we're iterating now */
 	fds[num_fds] = fd;
 	fd->backend_info = num_fds;
+	fd->exclusive[0] = fd->exclusive[1] = false;
 	num_fds++;
 	if (events)
 		num_waiting++;
@@ -63,90 +88,182 @@ static void del_fd(struct fd *fd)
 		fds[n]->backend_info = n;
 	} else if (num_fds == 1) {
 		/* Free everything when no more fds. */
-		free(pollfds);
-		free(fds);
-		pollfds = NULL;
+		pollfds = tal_free(pollfds);
 		fds = NULL;
 		max_fds = 0;
+		if (num_always == 0) {
+			always = tal_free(always);
+			max_always = 0;
+		}
 	}
 	num_fds--;
 	fd->backend_info = -1;
-	close(fd->fd);
+
+	if (fd->exclusive[IO_IN])
+		num_exclusive--;
+	if (fd->exclusive[IO_OUT])
+		num_exclusive--;
+}
+
+static void destroy_listener(struct io_listener *l)
+{
+	close(l->fd.fd);
+	del_fd(&l->fd);
 }
 
 bool add_listener(struct io_listener *l)
 {
 	if (!add_fd(&l->fd, POLLIN))
 		return false;
+	tal_add_destructor(l, destroy_listener);
 	return true;
 }
 
-void backend_plan_changed(struct io_conn *conn)
+static int find_always(const struct io_plan *plan)
 {
-	struct pollfd *pfd = &pollfds[conn->fd.backend_info];
+	size_t i = 0;
 
-	if (pfd->events)
-		num_waiting--;
+	for (i = 0; i < num_always; i++)
+		if (always[i] == plan)
+			return i;
+	return -1;
+}
+
+static void remove_from_always(const struct io_plan *plan)
+{
+	int pos;
+
+	if (plan->status != IO_ALWAYS)
+		return;
+
+	pos = find_always(plan);
+	assert(pos >= 0);
+
+	/* Move last one down if we made a hole */
+	if (pos != num_always-1)
+		always[pos] = always[num_always-1];
+	num_always--;
+
+	/* Only free if no fds left either. */
+	if (num_always == 0 && max_fds == 0) {
+		always = tal_free(always);
+		max_always = 0;
+	}
+}
+
+bool backend_new_always(struct io_plan *plan)
+{
+	assert(find_always(plan) == -1);
+
+	if (!max_always) {
+		assert(num_always == 0);
+		always = tal_arr(NULL, struct io_plan *, 8);
+		if (!always)
+			return false;
+		max_always = 8;
+	}
 
-	pfd->events = conn->plan.pollflag;
-	if (conn->duplex) {
-		int mask = conn->duplex->plan.pollflag;
-		/* You can't *both* read/write. */
-		assert(!mask || pfd->events != mask);
-		pfd->events |= mask;
+	if (num_always + 1 > max_always) {
+		size_t num = max_always * 2;
+
+		if (!tal_resize(&always, num))
+			return false;
+		max_always = num;
 	}
+
+	always[num_always++] = plan;
+	return true;
+}
+
+static void setup_pfd(struct io_conn *conn, struct pollfd *pfd)
+{
+	assert(pfd == &pollfds[conn->fd.backend_info]);
+
+	pfd->events = 0;
+	if (conn->plan[IO_IN].status == IO_POLLING_NOTSTARTED
+	    || conn->plan[IO_IN].status == IO_POLLING_STARTED)
+		pfd->events |= POLLIN;
+	if (conn->plan[IO_OUT].status == IO_POLLING_NOTSTARTED
+	    || conn->plan[IO_OUT].status == IO_POLLING_STARTED)
+		pfd->events |= POLLOUT;
+
 	if (pfd->events) {
-		num_waiting++;
 		pfd->fd = conn->fd.fd;
-	} else
-		pfd->fd = -conn->fd.fd;
-
-	if (!conn->plan.next)
-		num_closing++;
+	} else {
+		pfd->fd = -conn->fd.fd - 1;
+	}
 }
 
-bool add_conn(struct io_conn *c)
+void backend_new_plan(struct io_conn *conn)
 {
-	if (!add_fd(&c->fd, c->plan.pollflag))
-		return false;
-	/* Immediate close is allowed. */
-	if (!c->plan.next)
-		num_closing++;
-	return true;
+	struct pollfd *pfd = &pollfds[conn->fd.backend_info];
+
+	if (pfd->events)
+		num_waiting--;
+
+	setup_pfd(conn, pfd);
+
+	if (pfd->events)
+		num_waiting++;
 }
 
-bool add_duplex(struct io_conn *c)
+void backend_wake(const void *wait)
 {
-	c->fd.backend_info = c->duplex->fd.backend_info;
-	backend_plan_changed(c);
-	return true;
+	unsigned int i;
+
+	for (i = 0; i < num_fds; i++) {
+		struct io_conn *c;
+
+		/* Ignore listeners */
+		if (fds[i]->listener)
+			continue;
+
+		c = (void *)fds[i];
+		if (c->plan[IO_IN].status == IO_WAITING
+		    && c->plan[IO_IN].arg.u1.const_vp == wait)
+			io_do_wakeup(c, IO_IN);
+
+		if (c->plan[IO_OUT].status == IO_WAITING
+		    && c->plan[IO_OUT].arg.u1.const_vp == wait)
+			io_do_wakeup(c, IO_OUT);
+	}
 }
 
-static void del_conn(struct io_conn *conn)
+static void destroy_conn(struct io_conn *conn, bool close_fd)
 {
-	if (conn->finish)
+	int saved_errno = errno;
+
+	if (close_fd)
+		close(conn->fd.fd);
+	del_fd(&conn->fd);
+
+	remove_from_always(&conn->plan[IO_IN]);
+	remove_from_always(&conn->plan[IO_OUT]);
+
+	/* errno saved/restored by tal_free itself. */
+	if (conn->finish) {
+		errno = saved_errno;
 		conn->finish(conn, conn->finish_arg);
-	if (timeout_active(conn))
-		backend_del_timeout(conn);
-	free(conn->timeout);
-	if (conn->duplex) {
-		/* In case fds[] pointed to the other one. */
-		fds[conn->fd.backend_info] = &conn->duplex->fd;
-		conn->duplex->duplex = NULL;
-	} else
-		del_fd(&conn->fd);
-	num_closing--;
+	}
 }
 
-void del_listener(struct io_listener *l)
+static void destroy_conn_close_fd(struct io_conn *conn)
 {
-	del_fd(&l->fd);
+	destroy_conn(conn, true);
 }
 
-static void set_plan(struct io_conn *conn, struct io_plan plan)
+bool add_conn(struct io_conn *c)
 {
-	conn->plan = plan;
-	backend_plan_changed(conn);
+	if (!add_fd(&c->fd, 0))
+		return false;
+	tal_add_destructor(c, destroy_conn_close_fd);
+	return true;
+}
+
+void cleanup_conn_without_close(struct io_conn *conn)
+{
+	tal_del_destructor(conn, destroy_conn_close_fd);
+	destroy_conn(conn, false);
 }
 
 static void accept_conn(struct io_listener *l)
@@ -156,136 +273,206 @@ static void accept_conn(struct io_listener *l)
 	/* FIXME: What to do here? */
 	if (fd < 0)
 		return;
-	l->init(fd, l->arg);
+
+	io_new_conn(l->ctx, fd, l->init, l->arg);
 }
 
-/* It's OK to miss some, as long as we make progress. */
-static void finish_conns(void)
+/* Return pointer to exclusive flag for this plan. */
+static bool *exclusive(struct io_plan *plan)
 {
-	unsigned int i;
+	struct io_conn *conn;
+
+	conn = container_of(plan, struct io_conn, plan[plan->dir]);
+	return &conn->fd.exclusive[plan->dir];
+}
 
-	for (i = 0; !io_loop_return && i < num_fds; i++) {
-		struct io_conn *c, *duplex;
+/* For simplicity, we do one always at a time */
+static bool handle_always(void)
+{
+	int i;
 
-		if (!num_closing)
-			break;
+	/* Backwards is simple easier to remove entries */
+	for (i = num_always - 1; i >= 0; i--) {
+		struct io_plan *plan = always[i];
 
-		if (fds[i]->listener)
+		if (num_exclusive && !*exclusive(plan))
 			continue;
-		c = (void *)fds[i];
-		for (duplex = c->duplex; c; c = duplex, duplex = NULL) {
-			if (!c->plan.next) {
-				del_conn(c);
-				free(c);
-				i--;
-			}
-		}
+		/* Remove first: it might re-add */
+		if (i != num_always-1)
+			always[i] = always[num_always-1];
+		num_always--;
+		io_do_always(plan);
+		return true;
 	}
+
+	return false;
 }
 
-void backend_add_timeout(struct io_conn *conn, struct timespec duration)
+bool backend_set_exclusive(struct io_plan *plan, bool excl)
 {
-	if (!timeouts.base)
-		timers_init(&timeouts, time_now());
-	timer_add(&timeouts, &conn->timeout->timer,
-		  time_add(time_now(), duration));
-	conn->timeout->conn = conn;
+	bool *excl_ptr = exclusive(plan);
+
+	if (excl != *excl_ptr) {
+		*excl_ptr = excl;
+		if (!excl)
+			num_exclusive--;
+		else
+			num_exclusive++;
+	}
+
+	return num_exclusive != 0;
 }
 
-void backend_del_timeout(struct io_conn *conn)
+/* FIXME: We could do this once at set_exclusive time, and catch everywhere
+ * else that we manipulate events. */
+static void exclude_pollfds(void)
 {
-	assert(conn->timeout->conn == conn);
-	timer_del(&timeouts, &conn->timeout->timer);
-	conn->timeout->conn = NULL;
+	size_t i;
+
+	if (num_exclusive == 0)
+		return;
+
+	for (i = 0; i < num_fds; i++) {
+		struct pollfd *pfd = &pollfds[fds[i]->backend_info];
+
+		if (!fds[i]->exclusive[IO_IN])
+			pfd->events &= ~POLLIN;
+		if (!fds[i]->exclusive[IO_OUT])
+			pfd->events &= ~POLLOUT;
+
+		/* If we're not listening, we don't want error events
+		 * either. */
+		if (!pfd->events)
+			pfd->fd = -fds[i]->fd - 1;
+	}
 }
 
-/* This is the main loop. */
-void *io_loop(void)
+static void restore_pollfds(void)
 {
-	void *ret;
+	size_t i;
 
-	while (!io_loop_return) {
-		int i, r, timeout = INT_MAX;
-		struct timespec now;
+	if (num_exclusive == 0)
+		return;
+
+	for (i = 0; i < num_fds; i++) {
+		struct pollfd *pfd = &pollfds[fds[i]->backend_info];
 
-		if (timeouts.base) {
-			struct timespec first;
-			struct list_head expired;
-			struct io_timeout *t;
+		if (fds[i]->listener) {
+			pfd->events = POLLIN;
+			pfd->fd = fds[i]->fd;
+		} else {
+			struct io_conn *conn = (void *)fds[i];
+			setup_pfd(conn, pfd);
+		}
+	}
+}
 
-			now = time_now();
+/* This is the main loop. */
+void *io_loop(struct timers *timers, struct timer **expired)
+{
+	void *ret;
 
-			/* Call functions for expired timers. */
-			timers_expire(&timeouts, now, &expired);
-			while ((t = list_pop(&expired, struct io_timeout, timer.list))) {
-				struct io_conn *conn = t->conn;
-				/* Clear, in case timer re-adds */
-				t->conn = NULL;
-				set_plan(conn, t->next(conn, t->next_arg));
-			}
+	/* if timers is NULL, expired must be.  If not, not. */
+	assert(!timers == !expired);
 
-			/* Now figure out how long to wait for the next one. */
-			if (timer_earliest(&timeouts, &first)) {
-				uint64_t f = time_to_msec(time_sub(first, now));
-				if (f < INT_MAX)
-					timeout = f;
-			}
-		}
+	/* Make sure this is NULL if we exit for some other reason. */
+	if (expired)
+		*expired = NULL;
+
+	while (!io_loop_return) {
+		int i, r, ms_timeout = -1;
 
-		if (num_closing) {
-			finish_conns();
+		if (handle_always()) {
 			/* Could have started/finished more. */
 			continue;
 		}
 
+		/* Everything closed? */
 		if (num_fds == 0)
 			break;
 
 		/* You can't tell them all to go to sleep! */
 		assert(num_waiting);
 
-		r = poll(pollfds, num_fds, timeout);
-		if (r < 0)
+		if (timers) {
+			struct timemono now, first;
+
+			now = nowfn();
+
+			/* Call functions for expired timers. */
+			*expired = timers_expire(timers, now);
+			if (*expired)
+				break;
+
+			/* Now figure out how long to wait for the next one. */
+			if (timer_earliest(timers, &first)) {
+				uint64_t next;
+				next = time_to_msec(timemono_between(first, now));
+				if (next < INT_MAX)
+					ms_timeout = next;
+				else
+					ms_timeout = INT_MAX;
+			}
+		}
+
+		/* We do this temporarily, assuming exclusive is unusual */
+		exclude_pollfds();
+		r = pollfn(pollfds, num_fds, ms_timeout);
+		restore_pollfds();
+
+		if (r < 0) {
+			/* Signals shouldn't break us, unless they set
+			 * io_loop_return. */
+			if (errno == EINTR)
+				continue;
 			break;
+		}
 
 		for (i = 0; i < num_fds && !io_loop_return; i++) {
 			struct io_conn *c = (void *)fds[i];
 			int events = pollfds[i].revents;
 
+			/* Clear so we don't get confused if exclusive next time */
+			pollfds[i].revents = 0;
+
 			if (r == 0)
 				break;
 
 			if (fds[i]->listener) {
+				struct io_listener *l = (void *)fds[i];
 				if (events & POLLIN) {
-					accept_conn((void *)c);
+					accept_conn(l);
+					r--;
+				} else if (events & (POLLHUP|POLLNVAL|POLLERR)) {
 					r--;
+					errno = EBADF;
+					io_close_listener(l);
 				}
 			} else if (events & (POLLIN|POLLOUT)) {
 				r--;
-				if (c->duplex) {
-					int mask = c->duplex->plan.pollflag;
-					if (events & mask) {
-						io_ready(c->duplex);
-						events &= ~mask;
-						if (!(events&(POLLIN|POLLOUT)))
-							continue;
-					}
-				}
-				io_ready(c);
-			} else if (events & POLLHUP) {
+				io_ready(c, events);
+			} else if (events & (POLLHUP|POLLNVAL|POLLERR)) {
 				r--;
-				set_plan(c, io_close(c, NULL));
-				if (c->duplex)
-					set_plan(c->duplex,
-						 io_close(c->duplex, NULL));
+				errno = EBADF;
+				io_close(c);
 			}
 		}
 	}
 
-	while (num_closing)
-		finish_conns();
-
 	ret = io_loop_return;
 	io_loop_return = NULL;
+
 	return ret;
 }
+
+const void *io_have_fd(int fd, bool *listener)
+{
+	for (size_t i = 0; i < num_fds; i++) {
+		if (fds[i]->fd != fd)
+			continue;
+		if (listener)
+			*listener = fds[i]->listener;
+		return fds[i];
+	}
+	return NULL;
+}