s3: Use poll for the event loop
authorVolker Lendecke <vl@samba.org>
Thu, 3 Feb 2011 15:34:07 +0000 (16:34 +0100)
committerVolker Lendecke <vlendec@samba.org>
Mon, 28 Feb 2011 15:40:19 +0000 (16:40 +0100)
source3/include/event.h
source3/lib/events.c

index dc45b768963aa855e53b64ec12fc39195f0666ad..c6689d9e919a33d03ee8ebef3808be54eeddcb22 100644 (file)
@@ -36,3 +36,8 @@ struct timeval *get_timed_events_timeout(struct event_context *event_ctx,
 void dump_event_list(struct event_context *event_ctx);
 struct tevent_context *s3_tevent_context_init(TALLOC_CTX *mem_ctx);
 
+bool event_add_to_poll_args(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
+                           struct pollfd **pfds, int *num_pfds,
+                           int *ptimeout);
+bool run_events_poll(struct tevent_context *ev, int pollrtn,
+                    struct pollfd *pfds, int num_pfds);
index 07a270c1e7481a83733553986400b5a919e43149..f850d23da23f03448ef823535ec0e63f07b0df38 100644 (file)
@@ -21,6 +21,7 @@
 #include "includes.h"
 #include <tevent_internal.h>
 #include "../lib/util/select.h"
+#include "system/select.h"
 
 /*
  * Return if there's something in the queue
@@ -136,6 +137,256 @@ bool run_events(struct tevent_context *ev,
        return false;
 }
 
+struct tevent_poll_private {
+       /*
+        * Index from file descriptor into the pollfd array
+        */
+       int *pollfd_idx;
+
+       /*
+        * Cache for s3_event_loop_once to avoid reallocs
+        */
+       struct pollfd *pfds;
+};
+
+static struct tevent_poll_private *tevent_get_poll_private(
+       struct tevent_context *ev)
+{
+       struct tevent_poll_private *state;
+
+       state = (struct tevent_poll_private *)ev->additional_data;
+       if (state == NULL) {
+               state = TALLOC_ZERO_P(ev, struct tevent_poll_private);
+               ev->additional_data = (void *)state;
+               if (state == NULL) {
+                       DEBUG(10, ("talloc failed\n"));
+               }
+       }
+       return state;
+}
+
+static void count_fds(struct tevent_context *ev,
+                     int *pnum_fds, int *pmax_fd)
+{
+       struct tevent_fd *fde;
+       int num_fds = 0;
+       int max_fd = 0;
+
+       for (fde = ev->fd_events; fde != NULL; fde = fde->next) {
+               if (fde->flags & (EVENT_FD_READ|EVENT_FD_WRITE)) {
+                       num_fds += 1;
+                       if (fde->fd > max_fd) {
+                               max_fd = fde->fd;
+                       }
+               }
+       }
+       *pnum_fds = num_fds;
+       *pmax_fd = max_fd;
+}
+
+bool event_add_to_poll_args(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
+                           struct pollfd **pfds, int *pnum_pfds,
+                           int *ptimeout)
+{
+       struct tevent_poll_private *state;
+       struct tevent_fd *fde;
+       int i, num_fds, max_fd, num_pollfds, idx_len;
+       struct pollfd *fds;
+       struct timeval now, diff;
+       int timeout;
+
+       state = tevent_get_poll_private(ev);
+       if (state == NULL) {
+               return false;
+       }
+       count_fds(ev, &num_fds, &max_fd);
+
+       idx_len = max_fd+1;
+
+       if (talloc_array_length(state->pollfd_idx) < idx_len) {
+               state->pollfd_idx = TALLOC_REALLOC_ARRAY(
+                       state, state->pollfd_idx, int, idx_len);
+               if (state->pollfd_idx == NULL) {
+                       DEBUG(10, ("talloc_realloc failed\n"));
+                       return false;
+               }
+       }
+
+       fds = *pfds;
+       num_pollfds = *pnum_pfds;
+
+       /*
+        * The +1 is for the sys_poll calling convention. It expects
+        * an array 1 longer for the signal pipe
+        */
+
+       if (talloc_array_length(fds) < num_pollfds + num_fds + 1) {
+               fds = TALLOC_REALLOC_ARRAY(mem_ctx, fds, struct pollfd,
+                                          num_pollfds + num_fds + 1);
+               if (fds == NULL) {
+                       DEBUG(10, ("talloc_realloc failed\n"));
+                       return false;
+               }
+       }
+
+       memset(&fds[num_pollfds], 0, sizeof(struct pollfd) * num_fds);
+
+       /*
+        * This needs tuning. We need to cope with multiple fde's for a file
+        * descriptor. The problem is that we need to re-use pollfd_idx across
+        * calls for efficiency. One way would be a direct bitmask that might
+        * be initialized quicker, but our bitmap_init implementation is
+        * pretty heavy-weight as well.
+        */
+       for (i=0; i<idx_len; i++) {
+               state->pollfd_idx[i] = -1;
+       }
+
+       for (fde = ev->fd_events; fde; fde = fde->next) {
+               struct pollfd *pfd;
+
+               if ((fde->flags & (EVENT_FD_READ|EVENT_FD_WRITE)) == 0) {
+                       continue;
+               }
+
+               if (state->pollfd_idx[fde->fd] == -1) {
+                       /*
+                        * We haven't seen this fd yet. Allocate a new pollfd.
+                        */
+                       state->pollfd_idx[fde->fd] = num_pollfds;
+                       pfd = &fds[num_pollfds];
+                       num_pollfds += 1;
+               } else {
+                       /*
+                        * We have already seen this fd. OR in the flags.
+                        */
+                       pfd = &fds[state->pollfd_idx[fde->fd]];
+               }
+
+               pfd->fd = fde->fd;
+
+               if (fde->flags & EVENT_FD_READ) {
+                       pfd->events |= (POLLIN|POLLHUP);
+               }
+               if (fde->flags & EVENT_FD_WRITE) {
+                       pfd->events |= POLLOUT;
+               }
+       }
+       *pfds = fds;
+       *pnum_pfds = num_pollfds;
+
+       if (ev->immediate_events != NULL) {
+               *ptimeout = 0;
+               return true;
+       }
+       if (ev->timer_events == NULL) {
+               *ptimeout = INT_MAX;
+               return true;
+       }
+
+       now = timeval_current();
+       diff = timeval_until(&now, &ev->timer_events->next_event);
+       timeout = timeval_to_msec(diff);
+
+       if (timeout < *ptimeout) {
+               *ptimeout = timeout;
+       }
+
+       return true;
+}
+
+bool run_events_poll(struct tevent_context *ev, int pollrtn,
+                    struct pollfd *pfds, int num_pfds)
+{
+       struct tevent_poll_private *state;
+       int *pollfd_idx;
+       struct tevent_fd *fde;
+       struct timeval now;
+
+       if (ev->signal_events &&
+           tevent_common_check_signal(ev)) {
+               return true;
+       }
+
+       if (ev->immediate_events &&
+           tevent_common_loop_immediate(ev)) {
+               return true;
+       }
+
+       GetTimeOfDay(&now);
+
+       if ((ev->timer_events != NULL)
+           && (timeval_compare(&now, &ev->timer_events->next_event) >= 0)) {
+               /* this older events system did not auto-free timed
+                  events on running them, and had a race condition
+                  where the event could be called twice if the
+                  talloc_free of the te happened after the callback
+                  made a call which invoked the event loop. To avoid
+                  this while still allowing old code which frees the
+                  te, we need to create a temporary context which
+                  will be used to ensure the te is freed. We also
+                  remove the te from the timed event list before we
+                  call the handler, to ensure we can't loop */
+
+               struct tevent_timer *te = ev->timer_events;
+               TALLOC_CTX *tmp_ctx = talloc_new(ev);
+
+               DEBUG(10, ("Running timed event \"%s\" %p\n",
+                          ev->timer_events->handler_name, ev->timer_events));
+
+               DLIST_REMOVE(ev->timer_events, te);
+               talloc_steal(tmp_ctx, te);
+
+               te->handler(ev, te, now, te->private_data);
+
+               talloc_free(tmp_ctx);
+               return true;
+       }
+
+       if (pollrtn <= 0) {
+               /*
+                * No fd ready
+                */
+               return false;
+       }
+
+       state = (struct tevent_poll_private *)ev->additional_data;
+       pollfd_idx = state->pollfd_idx;
+
+       for (fde = ev->fd_events; fde; fde = fde->next) {
+               struct pollfd *pfd;
+               uint16 flags = 0;
+
+               if (pollfd_idx[fde->fd] >= num_pfds) {
+                       DEBUG(1, ("internal error: pollfd_idx[fde->fd] (%d) "
+                                 ">= num_pfds (%d)\n", pollfd_idx[fde->fd],
+                                 num_pfds));
+                       return false;
+               }
+               pfd = &pfds[pollfd_idx[fde->fd]];
+
+               if (pfd->fd != fde->fd) {
+                       DEBUG(1, ("internal error: pfd->fd (%d) "
+                                 "!= fde->fd (%d)\n", pollfd_idx[fde->fd],
+                                  num_pfds));
+                       return false;
+               }
+
+               if (pfd->revents & (POLLIN|POLLHUP|POLLERR)) {
+                       flags |= EVENT_FD_READ;
+               }
+               if (pfd->revents & POLLOUT) {
+                       flags |= EVENT_FD_WRITE;
+               }
+               if (flags & fde->flags) {
+                       DLIST_DEMOTE(ev->fd_events, fde, struct tevent_fd);
+                       fde->handler(ev, fde, flags, fde->private_data);
+                       return true;
+               }
+       }
+
+       return false;
+}
 
 struct timeval *get_timed_events_timeout(struct tevent_context *ev,
                                         struct timeval *to_ret)
@@ -161,35 +412,38 @@ struct timeval *get_timed_events_timeout(struct tevent_context *ev,
 
 static int s3_event_loop_once(struct tevent_context *ev, const char *location)
 {
-       struct timeval to;
-       fd_set r_fds, w_fds;
-       int maxfd = 0;
+       struct tevent_poll_private *state;
+       int timeout;
+       int num_pfds;
        int ret;
 
-       FD_ZERO(&r_fds);
-       FD_ZERO(&w_fds);
+       timeout = INT_MAX;
 
-       to.tv_sec = 9999;       /* Max timeout */
-       to.tv_usec = 0;
+       state = tevent_get_poll_private(ev);
+       if (state == NULL) {
+               errno = ENOMEM;
+               return -1;
+       }
 
-       if (run_events(ev, 0, NULL, NULL)) {
+       if (run_events_poll(ev, 0, NULL, 0)) {
                return 0;
        }
 
-       if (!event_add_to_select_args(ev, &r_fds, &w_fds, &to, &maxfd)) {
+       num_pfds = 0;
+       if (!event_add_to_poll_args(ev, state,
+                                   &state->pfds, &num_pfds, &timeout)) {
                return -1;
        }
 
-       ret = sys_select(maxfd+1, &r_fds, &w_fds, NULL, &to);
-
+       ret = sys_poll(state->pfds, num_pfds, timeout);
        if (ret == -1 && errno != EINTR) {
                tevent_debug(ev, TEVENT_DEBUG_FATAL,
-                            "sys_select() failed: %d:%s\n",
+                            "poll() failed: %d:%s\n",
                             errno, strerror(errno));
                return -1;
        }
 
-       run_events(ev, ret, &r_fds, &w_fds);
+       run_events_poll(ev, ret, state->pfds, num_pfds);
        return 0;
 }