]> git.ozlabs.org Git - ccan/blobdiff - ccan/io/poll.c
io: fix another leak path for always array.
[ccan] / ccan / io / poll.c
index 5982ef1535e363723ac4904729fa824e1d718855..17f9384587d92e39b37e199b0f05159bb899d112 100644 (file)
@@ -7,36 +7,68 @@
 #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_next = 0, num_finished = 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].fd = fd->fd;
        pollfds[num_fds].events = events;
+       /* In case it's idle. */
+       if (!events)
+               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++;
+
        return true;
 }
 
@@ -46,6 +78,8 @@ static void del_fd(struct fd *fd)
 
        assert(n != -1);
        assert(n < num_fds);
+       if (pollfds[n].events)
+               num_waiting--;
        if (n != num_fds - 1) {
                /* Move last one over us. */
                pollfds[n] = pollfds[num_fds-1];
@@ -54,242 +88,371 @@ 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;
-       num_waiting++;
+       tal_add_destructor(l, destroy_listener);
        return true;
 }
 
-bool add_conn(struct io_conn *c)
+static int find_always(const struct io_plan *plan)
 {
-       if (!add_fd(&c->fd, 0))
-               return false;
-       num_next++;
-       return true;
+       for (size_t i = 0; i < num_always; i++)
+               if (always[i] == plan)
+                       return i;
+       return -1;
 }
 
-bool add_duplex(struct io_conn *c)
+static void remove_from_always(const struct io_plan *plan)
 {
-       c->fd.backend_info = c->duplex->fd.backend_info;
-       num_next++;
-       return true;
+       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;
+       }
 }
 
-static void del_conn(struct io_conn *conn)
+bool backend_new_always(struct io_plan *plan)
 {
-       if (conn->finish)
-               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);
-       if (conn->state == FINISHED)
-               num_finished--;
-       else if (conn->state == NEXT)
-               num_next--;
+       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 del_listener(struct io_listener *l)
+static void setup_pfd(struct io_conn *conn, struct pollfd *pfd)
 {
-       del_fd(&l->fd);
+       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) {
+               pfd->fd = conn->fd.fd;
+       } else {
+               pfd->fd = -conn->fd.fd - 1;
+       }
 }
 
-void backend_set_state(struct io_conn *conn, struct io_plan *plan)
+void backend_new_plan(struct io_conn *conn)
 {
-       enum io_state state = from_ioplan(plan);
        struct pollfd *pfd = &pollfds[conn->fd.backend_info];
 
        if (pfd->events)
                num_waiting--;
 
-       pfd->events = conn->pollflag;
-       if (conn->duplex) {
-               int mask = conn->duplex->pollflag;
-               /* You can't *both* read/write. */
-               assert(!mask || pfd->events != mask);
-               pfd->events |= mask;
-       }
+       setup_pfd(conn, pfd);
+
        if (pfd->events)
                num_waiting++;
+}
+
+void backend_wake(const void *wait)
+{
+       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 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 = saved_errno;
+               conn->finish(conn, conn->finish_arg);
+       }
+}
+
+static void destroy_conn_close_fd(struct io_conn *conn)
+{
+       destroy_conn(conn, true);
+}
 
-       if (state == NEXT)
-               num_next++;
-       else if (state == FINISHED)
-               num_finished++;
+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;
+}
 
-       conn->state = state;
+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)
 {
-       struct io_conn *c;
        int fd = accept(l->fd.fd, NULL, NULL);
 
        /* FIXME: What to do here? */
        if (fd < 0)
                return;
-       c = io_new_conn(fd, l->next, l->finish, l->conn_arg);
-       if (!c) {
-               close(fd);
-               return;
-       }
+
+       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_and_next(bool finished_only)
+/* 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_finished) {
-                       if (finished_only || num_next == 0)
-                               break;
-               }
-               if (fds[i]->listener)
+/* 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 (num_exclusive && !*exclusive(plan))
                        continue;
-               c = (void *)fds[i];
-               for (duplex = c->duplex; c; c = duplex, duplex = NULL) {
-                       if (c->state == FINISHED) {
-                               del_conn(c);
-                               free(c);
-                               i--;
-                       } else if (!finished_only && c->state == NEXT) {
-                               backend_set_state(c, c->next(c, c->next_arg));
-                               num_next--;
-                       }
-               }
+               /* 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;
 }
 
-static void ready(struct io_conn *c)
+bool backend_set_exclusive(struct io_plan *plan, bool excl)
 {
-       backend_set_state(c, do_ready(c));
+       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_add_timeout(struct io_conn *conn, struct timespec duration)
+/* FIXME: We could do this once at set_exclusive time, and catch everywhere
+ * else that we manipulate events. */
+static void exclude_pollfds(void)
 {
-       if (!timeouts.base)
-               timers_init(&timeouts, time_now());
-       timer_add(&timeouts, &conn->timeout->timer,
-                 time_add(time_now(), duration));
-       conn->timeout->conn = conn;
+       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;
+       }
 }
 
-void backend_del_timeout(struct io_conn *conn)
+static void restore_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]->listener) {
+                       pfd->events = POLLIN;
+                       pfd->fd = fds[i]->fd;
+               } else {
+                       struct io_conn *conn = (void *)fds[i];
+                       setup_pfd(conn, pfd);
+               }
+       }
 }
 
 /* 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, timeout = INT_MAX;
-               struct timespec now;
-
-               if (timeouts.base) {
-                       struct timespec first;
-                       struct list_head expired;
-                       struct io_timeout *t;
+       /* if timers is NULL, expired must be.  If not, not. */
+       assert(!timers == !expired);
 
-                       now = time_now();
+       /* Make sure this is NULL if we exit for some other reason. */
+       if (expired)
+               *expired = NULL;
 
-                       /* 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;
-                               backend_set_state(conn, t->next(conn, t->next_arg));
-                       }
-
-                       /* 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_finished || num_next) {
-                       finish_and_next(false);
+               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->pollflag;
-                                       if (events & mask) {
-                                               ready(c->duplex);
-                                               events &= ~mask;
-                                               if (!(events&(POLLIN|POLLOUT)))
-                                                       continue;
-                                       }
-                               }
-                               ready(c);
-                       } else if (events & POLLHUP) {
+                               io_ready(c, events);
+                       } else if (events & (POLLHUP|POLLNVAL|POLLERR)) {
                                r--;
-                               backend_set_state(c, io_close(c, NULL));
-                               if (c->duplex)
-                                       backend_set_state(c->duplex,
-                                                         io_close(c->duplex,
-                                                                  NULL));
+                               errno = EBADF;
+                               io_close(c);
                        }
                }
        }
 
-       while (num_finished)
-               finish_and_next(true);
-
        ret = io_loop_return;
        io_loop_return = NULL;
+
        return ret;
 }