]> git.ozlabs.org Git - ccan/blobdiff - ccan/io/poll.c
io: don't leak memory on clean shutdown.
[ccan] / ccan / io / poll.c
index c1a624525e9185811ad9510eef555a4dd80bfb30..a02152e3792ae07569d4a75b1ca423ecce5fb626 100644 (file)
@@ -8,11 +8,29 @@
 #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_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 io_conn *closing = NULL, *always = NULL;
+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)
 {
@@ -40,12 +58,13 @@ static bool add_fd(struct fd *fd, short events)
        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++;
@@ -75,62 +94,113 @@ static void del_fd(struct fd *fd)
        }
        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_new_closing(struct io_conn *conn)
+static int find_always(const struct io_plan *plan)
 {
-       /* Already on always list?  Remove it. */
-       if (conn->list) {
-               struct io_conn **p = &always;
+       for (size_t i = 0; i < num_always; i++)
+               if (always[i] == plan)
+                       return i;
+       return -1;
+}
 
-               while (*p != conn)
-                       p = &(*p)->list;
+static void remove_from_always(const struct io_plan *plan)
+{
+       int pos;
 
-               *p = conn->list;
-       }
+       if (plan->status != IO_ALWAYS)
+               return;
+
+       pos = find_always(plan);
+       assert(pos >= 0);
 
-       conn->list = closing;
-       closing = conn;
+       /* 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) {
+               tal_free(always);
+               max_always = 0;
+       }
 }
 
-void backend_new_always(struct io_conn *conn)
+bool backend_new_always(struct io_plan *plan)
 {
-       /* May already be in always list (other plan), or closing. */
-       if (!conn->list) {
-               conn->list = always;
-               always = conn;
+       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;
+       }
+
+       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;
 }
 
-void backend_new_plan(struct io_conn *conn)
+static void setup_pfd(struct io_conn *conn, struct pollfd *pfd)
 {
-       struct pollfd *pfd = &pollfds[conn->fd.backend_info];
-
-       if (pfd->events)
-               num_waiting--;
+       assert(pfd == &pollfds[conn->fd.backend_info]);
 
        pfd->events = 0;
-       if (conn->plan[IO_IN].status == IO_POLLING)
+       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)
+       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;
+               pfd->fd = -conn->fd.fd - 1;
        }
 }
 
+void backend_new_plan(struct io_conn *conn)
+{
+       struct pollfd *pfd = &pollfds[conn->fd.backend_info];
+
+       if (pfd->events)
+               num_waiting--;
+
+       setup_pfd(conn, pfd);
+
+       if (pfd->events)
+               num_waiting++;
+}
+
 void backend_wake(const void *wait)
 {
        unsigned int i;
@@ -144,34 +214,50 @@ void backend_wake(const void *wait)
 
                c = (void *)fds[i];
                if (c->plan[IO_IN].status == IO_WAITING
-                   && c->plan[IO_IN].u1.const_vp == wait)
-                       io_do_wakeup(c, &c->plan[IO_IN]);
+                   && 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].u1.const_vp == wait)
-                       io_do_wakeup(c, &c->plan[IO_OUT]);
+                   && c->plan[IO_OUT].arg.u1.const_vp == wait)
+                       io_do_wakeup(c, IO_OUT);
        }
 }
 
-bool add_conn(struct io_conn *c)
+static void destroy_conn(struct io_conn *conn, bool close_fd)
 {
-       return add_fd(&c->fd, 0);
-}
+       int saved_errno = errno;
 
-static void del_conn(struct io_conn *conn)
-{
+       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) {
-               /* Saved by io_close */
-               errno = conn->plan[IO_IN].u1.s;
+               errno = saved_errno;
                conn->finish(conn, conn->finish_arg);
        }
-       tal_free(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;
+}
+
+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)
@@ -185,55 +271,104 @@ static void accept_conn(struct io_listener *l)
        io_new_conn(l->ctx, fd, l->init, l->arg);
 }
 
-/* It's OK to miss some, as long as we make progress. */
-static bool close_conns(void)
+/* Return pointer to exclusive flag for this plan. */
+static bool *exclusive(struct io_plan *plan)
 {
-       bool ret = false;
+       struct io_conn *conn;
 
-       while (closing) {
-               struct io_conn *conn = closing;
+       conn = container_of(plan, struct io_conn, plan[plan->dir]);
+       return &conn->fd.exclusive[plan->dir];
+}
 
-               assert(conn->plan[IO_IN].status == IO_CLOSING);
-               assert(conn->plan[IO_OUT].status == IO_CLOSING);
+/* 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];
 
-               closing = closing->list;
-               del_conn(conn);
-               ret = true;
+               if (num_exclusive && !*exclusive(plan))
+                       continue;
+               /* 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 ret;
+
+       return false;
 }
 
-static bool handle_always(void)
+bool backend_set_exclusive(struct io_plan *plan, bool excl)
+{
+       bool *excl_ptr = exclusive(plan);
+
+       if (excl != *excl_ptr) {
+               *excl_ptr = excl;
+               if (!excl)
+                       num_exclusive--;
+               else
+                       num_exclusive++;
+       }
+
+       return num_exclusive != 0;
+}
+
+/* FIXME: We could do this once at set_exclusive time, and catch everywhere
+ * else that we manipulate events. */
+static void exclude_pollfds(void)
 {
-       bool ret = false;
+       if (num_exclusive == 0)
+               return;
 
-       while (always) {
-               struct io_conn *conn = always;
+       for (size_t i = 0; i < num_fds; i++) {
+               struct pollfd *pfd = &pollfds[fds[i]->backend_info];
 
-               assert(conn->plan[IO_IN].status == IO_ALWAYS
-                      || conn->plan[IO_OUT].status == IO_ALWAYS);
+               if (!fds[i]->exclusive[IO_IN])
+                       pfd->events &= ~POLLIN;
+               if (!fds[i]->exclusive[IO_OUT])
+                       pfd->events &= ~POLLOUT;
 
-               /* Remove from list, and mark it so it knows that. */
-               always = always->list;
-               conn->list = NULL;
-               io_do_always(conn);
-               ret = true;
+               /* If we're not listening, we don't want error events
+                * either. */
+               if (!pfd->events)
+                       pfd->fd = -fds[i]->fd - 1;
+       }
+}
+
+static void restore_pollfds(void)
+{
+       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]->listener) {
+                       pfd->events = POLLIN;
+                       pfd->fd = fds[i]->fd;
+               } else {
+                       struct io_conn *conn = (void *)fds[i];
+                       setup_pfd(conn, pfd);
+               }
        }
-       return ret;
 }
 
 /* This is the main loop. */
-void *io_loop(void)
+void *io_loop(struct timers *timers, struct timer **expired)
 {
        void *ret;
 
-       while (!io_loop_return) {
-               int i, r;
+       /* if timers is NULL, expired must be.  If not, not. */
+       assert(!timers == !expired);
 
-               if (close_conns()) {
-                       /* Could have started/finished more. */
-                       continue;
-               }
+       /* 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 (handle_always()) {
                        /* Could have started/finished more. */
@@ -247,21 +382,59 @@ void *io_loop(void)
                /* You can't tell them all to go to sleep! */
                assert(num_waiting);
 
-               r = poll(pollfds, num_fds, -1);
-               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--;
@@ -274,8 +447,6 @@ void *io_loop(void)
                }
        }
 
-       close_conns();
-
        ret = io_loop_return;
        io_loop_return = NULL;