Refactored xevents callback deletion

Some changes in the xevents code for deleting callback events,
required for the lifespan QoS implementation. With these changes,
there is no longer a possibility of the callback still being
executing or getting invoked after delete_xevent_callback has
returned.

Signed-off-by: Dennis Potman <dennis.potman@adlinktech.com>

src/core/ddsi/include/dds/ddsi/q_time.h

@@ -46,6 +46,9 @@ typedef struct {
   int64_t v;
 } nn_etime_t;
 
+#define NN_MTIME_NEVER ((nn_mtime_t) { T_NEVER })
+#define NN_WCTIME_NEVER ((nn_wctime_t) { T_NEVER })
+#define NN_ETIME_NEVER ((nn_etime_t) { T_NEVER })
 #define NN_WCTIME_INVALID ((nn_wctime_t) { INT64_MIN })
 
 int valid_ddsi_timestamp (ddsi_time_t t);

src/core/ddsi/include/dds/ddsi/q_xevent.h

@@ -12,6 +12,9 @@
 #ifndef NN_XEVENT_H
 #define NN_XEVENT_H
 
+#include "dds/ddsrt/retcode.h"
+#include "dds/ddsi/ddsi_guid.h"
+
 #if defined (__cplusplus)
 extern "C" {
 #endif
@@ -30,6 +33,7 @@ struct xevent;
 struct xeventq;
 struct proxy_writer;
 struct proxy_reader;
+struct nn_xmsg;
 
 struct xeventq *xeventq_new
 (
@@ -55,6 +59,7 @@ DDS_EXPORT int qxev_msg_rexmit_wrlock_held (struct xeventq *evq, struct nn_xmsg
 
 /* All of the following lock EVQ for the duration of the operation */
 DDS_EXPORT void delete_xevent (struct xevent *ev);
+DDS_EXPORT void delete_xevent_callback (struct xevent *ev);
 DDS_EXPORT int resched_xevent_if_earlier (struct xevent *ev, nn_mtime_t tsched);
 
 DDS_EXPORT struct xevent *qxev_heartbeat (struct xeventq *evq, nn_mtime_t tsched, const ddsi_guid_t *wr_guid);

src/core/ddsi/src/q_ddsi_discovery.c

@@ -1245,11 +1245,10 @@ static void handle_SEDP_alive (const struct receiver_state *rst, seqno_t seq, nn
     GVLOGDISC (" known%s", vendor_is_cloud (vendorid) ? "-DS" : "");
     if (vendor_is_cloud (vendorid) && pp->implicitly_created && memcmp(&pp->privileged_pp_guid.prefix, src_guid_prefix, sizeof(pp->privileged_pp_guid.prefix)) != 0)
     {
-      nn_etime_t never = { T_NEVER };
       GVLOGDISC (" "PGUIDFMT" attach-to-DS "PGUIDFMT, PGUID(pp->e.guid), PGUIDPREFIX(*src_guid_prefix), pp->privileged_pp_guid.entityid.u);
       ddsrt_mutex_lock (&pp->e.lock);
       pp->privileged_pp_guid.prefix = *src_guid_prefix;
-      lease_set_expiry(pp->lease, never);
+      lease_set_expiry(pp->lease, NN_ETIME_NEVER);
       ddsrt_mutex_unlock (&pp->e.lock);
     }
     GVLOGDISC ("\n");

src/core/ddsi/src/q_init.c

@@ -761,7 +761,7 @@ static void wait_for_receive_threads (struct q_globals *gv)
   }
   if (trigev)
   {
-    delete_xevent (trigev);
+    delete_xevent_callback (trigev);
   }
 }
 

src/core/ddsi/src/q_xevent.c

@@ -94,6 +94,7 @@ struct xevent
     struct {
       void (*cb) (struct xevent *ev, void *arg, nn_mtime_t tnow);
       void *arg;
+      bool executing;
     } callback;
   } u;
 };
@@ -232,7 +233,7 @@ static void add_to_non_timed_xmit_list (struct xeventq *evq, struct xevent_nt *e
   if (ev->kind == XEVK_MSG_REXMIT)
     remember_msg (evq, ev);
 
-  ddsrt_cond_signal (&evq->cond);
+  ddsrt_cond_broadcast (&evq->cond);
 }
 
 static struct xevent_nt *getnext_from_non_timed_xmit_list  (struct xeventq *evq)
@@ -316,6 +317,7 @@ void delete_xevent (struct xevent *ev)
 {
   struct xeventq *evq = ev->evq;
   ddsrt_mutex_lock (&evq->lock);
+  assert (ev->kind != XEVK_CALLBACK || ev->u.callback.executing);
   /* Can delete it only once, no matter how we implement it internally */
   assert (ev->tsched.v != TSCHED_DELETE);
   assert (TSCHED_DELETE < ev->tsched.v);
@@ -331,27 +333,44 @@ void delete_xevent (struct xevent *ev)
   }
   /* TSCHED_DELETE is absolute minimum time, so chances are we need to
      wake up the thread.  The superfluous signal is harmless. */
-  ddsrt_cond_signal (&evq->cond);
+  ddsrt_cond_broadcast (&evq->cond);
   ddsrt_mutex_unlock (&evq->lock);
 }
 
+void delete_xevent_callback (struct xevent *ev)
+{
+  struct xeventq *evq = ev->evq;
+  assert (ev->kind == XEVK_CALLBACK);
+  ddsrt_mutex_lock (&evq->lock);
+  if (ev->tsched.v != T_NEVER)
+  {
+    assert (ev->tsched.v != TSCHED_DELETE);
+    ddsrt_fibheap_delete (&evq_xevents_fhdef, &evq->xevents, ev);
+    ev->tsched.v = TSCHED_DELETE;
+  }
+  while (ev->u.callback.executing)
+    ddsrt_cond_wait (&evq->cond, &evq->lock);
+  ddsrt_mutex_unlock (&evq->lock);
+  free_xevent (evq, ev);
+}
+
 int resched_xevent_if_earlier (struct xevent *ev, nn_mtime_t tsched)
 {
   struct xeventq *evq = ev->evq;
   int is_resched;
+  if (tsched.v == T_NEVER)
+    return 0;
   ddsrt_mutex_lock (&evq->lock);
-  assert (tsched.v != TSCHED_DELETE);
   /* If you want to delete it, you to say so by calling the right
      function. Don't want to reschedule an event marked for deletion,
      but with TSCHED_DELETE = MIN_INT64, tsched >= ev->tsched is
      guaranteed to be false. */
-  assert (tsched.v > TSCHED_DELETE);
+  assert (tsched.v != TSCHED_DELETE);
   if (tsched.v >= ev->tsched.v)
     is_resched = 0;
   else
   {
     nn_mtime_t tbefore = earliest_in_xeventq (evq);
-    assert (tsched.v != T_NEVER);
     if (ev->tsched.v != T_NEVER)
     {
       ev->tsched = tsched;
@@ -364,7 +383,7 @@ int resched_xevent_if_earlier (struct xevent *ev, nn_mtime_t tsched)
     }
     is_resched = 1;
     if (tsched.v < tbefore.v)
-      ddsrt_cond_signal (&evq->cond);
+      ddsrt_cond_broadcast (&evq->cond);
   }
   ddsrt_mutex_unlock (&evq->lock);
   return is_resched;
@@ -406,13 +425,7 @@ static nn_mtime_t earliest_in_xeventq (struct xeventq *evq)
 {
   struct xevent *min;
   ASSERT_MUTEX_HELD (&evq->lock);
-  if ((min = ddsrt_fibheap_min (&evq_xevents_fhdef, &evq->xevents)) != NULL)
-    return min->tsched;
-  else
-  {
-    nn_mtime_t r = { T_NEVER };
-    return r;
-  }
+  return ((min = ddsrt_fibheap_min (&evq_xevents_fhdef, &evq->xevents)) != NULL) ? min->tsched : NN_MTIME_NEVER;
 }
 
 static void qxev_insert (struct xevent *ev)
@@ -426,7 +439,7 @@ static void qxev_insert (struct xevent *ev)
     nn_mtime_t tbefore = earliest_in_xeventq (evq);
     ddsrt_fibheap_insert (&evq_xevents_fhdef, &evq->xevents, ev);
     if (ev->tsched.v < tbefore.v)
-      ddsrt_cond_signal (&evq->cond);
+      ddsrt_cond_broadcast (&evq->cond);
   }
 }
 
@@ -502,7 +515,7 @@ void xeventq_stop (struct xeventq *evq)
   assert (evq->ts != NULL);
   ddsrt_mutex_lock (&evq->lock);
   evq->terminate = 1;
-  ddsrt_cond_signal (&evq->cond);
+  ddsrt_cond_broadcast (&evq->cond);
   ddsrt_mutex_unlock (&evq->lock);
   join_thread (evq->ts);
   evq->ts = NULL;
@@ -513,22 +526,7 @@ void xeventq_free (struct xeventq *evq)
   struct xevent *ev;
   assert (evq->ts == NULL);
   while ((ev = ddsrt_fibheap_extract_min (&evq_xevents_fhdef, &evq->xevents)) != NULL)
-  {
-    if (ev->tsched.v == TSCHED_DELETE || ev->kind != XEVK_CALLBACK)
     free_xevent (evq, ev);
-    else
-    {
-      ev->tsched.v = T_NEVER;
-      ev->u.callback.cb (ev, ev->u.callback.arg, ev->tsched);
-      if (ev->tsched.v != TSCHED_DELETE)
-      {
-        union { void *v; void (*f) (struct xevent *ev, void *arg, nn_mtime_t tnow); } fp;
-        fp.f = ev->u.callback.cb;
-        DDS_CWARNING (&evq->gv->logconfig, "xeventq_free: callback %p did not schedule deletion as required, deleting event anyway\n", fp.v);
-        delete_xevent (ev);
-      }
-    }
-  }
 
   {
     struct nn_xpack *xp = nn_xpack_new (evq->tev_conn, evq->auxiliary_bandwidth_limit, false);
@@ -1137,6 +1135,22 @@ static void handle_xevk_delete_writer (UNUSED_ARG (struct nn_xpack *xp), struct
 
 static void handle_individual_xevent (struct thread_state1 * const ts1, struct xevent *xev, struct nn_xpack *xp, nn_mtime_t tnow)
 {
+  struct xeventq *xevq = xev->evq;
+  /* We relinquish the lock while processing the event, but require it
+     held for administrative work. */
+  ASSERT_MUTEX_HELD (&xevq->lock);
+  if (xev->kind == XEVK_CALLBACK)
+  {
+    xev->u.callback.executing = true;
+    ddsrt_mutex_unlock (&xevq->lock);
+    xev->u.callback.cb (xev, xev->u.callback.arg, tnow);
+    ddsrt_mutex_lock (&xevq->lock);
+    xev->u.callback.executing = false;
+    ddsrt_cond_broadcast (&xevq->cond);
+  }
+  else
+  {
+    ddsrt_mutex_unlock (&xevq->lock);
     switch (xev->kind)
     {
       case XEVK_HEARTBEAT:
@@ -1155,9 +1169,12 @@ static void handle_individual_xevent (struct thread_state1 * const ts1, struct x
         handle_xevk_delete_writer (xp, xev, tnow);
         break;
       case XEVK_CALLBACK:
-      xev->u.callback.cb (xev, xev->u.callback.arg, tnow);
+        assert (0);
         break;
     }
+    ddsrt_mutex_lock (&xevq->lock);
+  }
+  ASSERT_MUTEX_HELD (&xevq->lock);
 }
 
 static void handle_individual_xevent_nt (struct xevent_nt *xev, struct nn_xpack *xp)
@@ -1181,20 +1198,8 @@ static void handle_timed_xevent (struct thread_state1 * const ts1, struct xevent
 {
    /* This function handles the individual xevent irrespective of
       whether it is a "timed" or "non-timed" xevent */
-  struct xeventq *xevq = xev->evq;
-
-  /* We relinquish the lock while processing the event, but require it
-     held for administrative work. */
-  ASSERT_MUTEX_HELD (&xevq->lock);
-
-  assert (xev->evq == xevq);
   assert (xev->tsched.v != TSCHED_DELETE);
-
-  ddsrt_mutex_unlock (&xevq->lock);
   handle_individual_xevent (ts1, xev, xp, tnow /* monotonic */);
-  ddsrt_mutex_lock (&xevq->lock);
-
-  ASSERT_MUTEX_HELD (&xevq->lock);
 }
 
 static void handle_nontimed_xevent (struct xevent_nt *xev, struct nn_xpack *xp)
@@ -1517,6 +1522,7 @@ struct xevent *qxev_callback (struct xeventq *evq, nn_mtime_t tsched, void (*cb)
   ev = qxev_common (evq, tsched, XEVK_CALLBACK);
   ev->u.callback.cb = cb;
   ev->u.callback.arg = arg;
+  ev->u.callback.executing = false;
   qxev_insert (ev);
   ddsrt_mutex_unlock (&evq->lock);
   return ev;