#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;
-#ifdef DEBUG
-static unsigned int io_loop_level;
-static struct io_conn *free_later;
-static void io_loop_enter(void)
-{
- io_loop_level++;
-}
-static void io_loop_exit(void)
-{
- io_loop_level--;
- if (io_loop_level == 0) {
- /* Delayed free. */
- while (free_later) {
- struct io_conn *c = free_later;
- free_later = c->finish_arg;
- free(c);
- }
- }
-}
-static void free_conn(struct io_conn *conn)
-{
- /* Only free on final exit: chain via finish. */
- if (io_loop_level > 1) {
- struct io_conn *c;
- for (c = free_later; c; c = c->finish_arg)
- assert(c != conn);
- conn->finish_arg = free_later;
- free_later = conn;
- } else
- free(conn);
-}
-#else
-static void io_loop_enter(void)
-{
-}
-static void io_loop_exit(void)
+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;
}
-static void free_conn(struct io_conn *conn)
+
+int (*io_poll_override(int (*poll)(struct pollfd *fds, nfds_t nfds, int timeout)))(struct pollfd *, nfds_t, int)
{
- free(conn);
+ int (*old)(struct pollfd *fds, nfds_t nfds, int timeout) = pollfn;
+ pollfn = poll;
+ return old;
}
-#endif
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++;
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;
}
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;
+ for (size_t 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;
- /* This can happen with debugging and delayed free... */
- if (conn->fd.backend_info == -1)
+ if (plan->status != IO_ALWAYS)
return;
- pfd = &pollfds[conn->fd.backend_info];
+ pos = find_always(plan);
+ assert(pos >= 0);
- if (pfd->events)
- num_waiting--;
+ /* Move last one down if we made a hole */
+ if (pos != num_always-1)
+ always[pos] = always[num_always-1];
+ num_always--;
+}
+
+bool backend_new_always(struct io_plan *plan)
+{
+ assert(find_always(plan) == -1);
- 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 (!max_always) {
+ assert(num_always == 0);
+ always = tal_arr(NULL, struct io_plan *, 8);
+ if (!always)
+ return false;
+ max_always = 8;
}
+
+ 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);
+ }
}
-void backend_del_conn(struct io_conn *conn)
+static void destroy_conn(struct io_conn *conn, bool close_fd)
{
+ 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 = conn->plan.u.close.saved_errno;
+ 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;
- conn->fd.backend_info = -1;
- } else
- del_fd(&conn->fd);
- num_closing--;
- free_conn(conn);
}
-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);
+}
+
+bool add_conn(struct io_conn *c)
+{
+ if (!add_fd(&c->fd, 0))
+ return false;
+ tal_add_destructor(c, destroy_conn_close_fd);
+ return true;
}
-static void set_plan(struct io_conn *conn, struct io_plan plan)
+void cleanup_conn_without_close(struct io_conn *conn)
{
- conn->plan = plan;
- backend_plan_changed(conn);
+ tal_del_destructor(conn, destroy_conn_close_fd);
+ destroy_conn(conn, false);
}
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 bool finish_conns(struct io_conn **ready)
+/* Return pointer to exclusive flag for this plan. */
+static bool *exclusive(struct io_plan *plan)
{
- unsigned int i;
+ struct io_conn *conn;
- for (i = 0; !io_loop_return && i < num_fds; i++) {
- struct io_conn *c, *duplex;
+ conn = container_of(plan, struct io_conn, plan[plan->dir]);
+ return &conn->fd.exclusive[plan->dir];
+}
- if (!num_closing)
- break;
+/* For simplicity, we do one always at a time */
+static bool handle_always(void)
+{
+ /* Backwards is simple easier to remove entries */
+ for (int 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) {
- if (doing_debug_on(c) && ready) {
- *ready = c;
- return true;
- }
- backend_del_conn(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;
+ if (num_exclusive == 0)
+ return;
+
+ for (size_t 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 *do_io_loop(struct io_conn **ready)
+static void restore_pollfds(void)
{
- void *ret;
+ if (num_exclusive == 0)
+ return;
- io_loop_enter();
+ for (size_t i = 0; i < num_fds; i++) {
+ struct pollfd *pfd = &pollfds[fds[i]->backend_info];
- while (!io_loop_return) {
- int i, r, timeout = INT_MAX;
- struct timespec now;
- bool some_timeouts = false;
+ if (fds[i]->listener) {
+ pfd->events = POLLIN;
+ pfd->fd = fds[i]->fd;
+ } else {
+ struct io_conn *conn = (void *)fds[i];
+ setup_pfd(conn, pfd);
+ }
+ }
+}
- if (timeouts.base) {
- struct timespec first;
- struct list_head expired;
- struct io_timeout *t;
+/* This is the main loop. */
+void *io_loop(struct timers *timers, struct timer **expired)
+{
+ void *ret;
- now = time_now();
+ /* if timers is NULL, expired must be. If not, not. */
+ assert(!timers == !expired);
- /* 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_current(conn);
- set_plan(conn, t->next(conn, t->next_arg));
- some_timeouts = true;
- }
+ /* Make sure this is NULL if we exit for some other reason. */
+ if (expired)
+ *expired = NULL;
- /* 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;
- }
- }
+ while (!io_loop_return) {
+ int i, r, ms_timeout = -1;
- if (num_closing) {
- /* If this finishes a debugging con, return now. */
- if (finish_conns(ready))
- return NULL;
+ if (handle_always()) {
/* Could have started/finished more. */
continue;
}
- /* debug can recurse on io_loop; anything can change. */
- if (doing_debug() && some_timeouts)
- 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) {
- if (doing_debug_on(c->duplex)
- && ready) {
- *ready = c->duplex;
- return NULL;
- }
- io_ready(c->duplex);
- events &= ~mask;
- /* debug can recurse;
- * anything can change. */
- if (doing_debug())
- break;
- if (!(events&(POLLIN|POLLOUT)))
- continue;
- }
- }
- if (doing_debug_on(c) && ready) {
- *ready = c;
- return NULL;
- }
- io_ready(c);
- /* debug can recurse; anything can change. */
- if (doing_debug())
- break;
+ io_ready(c, events);
} else if (events & (POLLHUP|POLLNVAL|POLLERR)) {
r--;
- set_current(c);
errno = EBADF;
- set_plan(c, io_close());
- if (c->duplex) {
- set_current(c->duplex);
- set_plan(c->duplex, io_close());
- }
+ io_close(c);
}
}
}
- while (num_closing && !io_loop_return) {
- if (finish_conns(ready))
- return NULL;
- }
-
ret = io_loop_return;
io_loop_return = NULL;
- io_loop_exit();
return ret;
}
-
-void *io_loop(void)
-{
- return do_io_loop(NULL);
-}