Merge pull request #701 from qlyoung/mt-safe-cancel

mt-safe thread_cancel()

lib/thread.c

@@ -328,6 +328,12 @@ thread_timer_update(void *node, int actual_position)
   thread->index = actual_position;
 }
 
+static void
+cancelreq_del (void *cr)
+{
+  XFREE (MTYPE_TMP, cr);
+}
+
 /* Allocate new thread master.  */
 struct thread_master *
 thread_master_create (void)
@@ -351,6 +357,7 @@ thread_master_create (void)
     return NULL;
 
   pthread_mutex_init (&rv->mtx, NULL);
+  pthread_cond_init (&rv->cancel_cond, NULL);
 
   rv->fd_limit = (int)limit.rlim_cur;
   rv->read = XCALLOC (MTYPE_THREAD, sizeof (struct thread *) * rv->fd_limit);
@@ -375,6 +382,9 @@ thread_master_create (void)
   rv->spin = true;
   rv->handle_signals = true;
   rv->owner = pthread_self();
+  rv->cancel_req = list_new ();
+  rv->cancel_req->del = cancelreq_del;
+  rv->canceled = true;
   pipe (rv->io_pipe);
   set_nonblocking (rv->io_pipe[0]);
   set_nonblocking (rv->io_pipe[1]);
@@ -533,6 +543,7 @@ thread_master_free (struct thread_master *m)
   pthread_mutex_destroy (&m->mtx);
   close (m->io_pipe[0]);
   close (m->io_pipe[1]);
+  list_delete (m->cancel_req);
 
   XFREE (MTYPE_THREAD_MASTER, m->handler.pfds);
   XFREE (MTYPE_THREAD_MASTER, m->handler.copy);
@@ -637,7 +648,7 @@ thread_get (struct thread_master *m, u_char type,
 
 static int
 fd_poll (struct thread_master *m, struct pollfd *pfds, nfds_t pfdsize,
-         nfds_t count, struct timeval *timer_wait)
+         nfds_t count, const struct timeval *timer_wait)
 {
   /* If timer_wait is null here, that means poll() should block indefinitely,
    * unless the thread_master has overriden it by setting ->selectpoll_timeout.
@@ -864,151 +875,288 @@ funcname_thread_add_event (struct thread_master *m,
   return thread;
 }
 
-static void
-thread_cancel_read_or_write (struct thread *thread, short int state)
-{
-  for (nfds_t i = 0; i < thread->master->handler.pfdcount; ++i)
-    if (thread->master->handler.pfds[i].fd == thread->u.fd)
-      {
-        thread->master->handler.pfds[i].events &= ~(state);
+/* Thread cancellation ------------------------------------------------------ */
 
-        /* remove thread fds from pfd list */
-        if (thread->master->handler.pfds[i].events == 0)
-          {
-            memmove(thread->master->handler.pfds+i,
-                    thread->master->handler.pfds+i+1,
-                    (thread->master->handler.pfdsize-i-1) * sizeof(struct pollfd));
-            thread->master->handler.pfdcount--;
-            return;
-          }
-      }
+/**
+ * NOT's out the .events field of pollfd corresponding to the given file
+ * descriptor. The event to be NOT'd is passed in the 'state' parameter.
+ *
+ * This needs to happen for both copies of pollfd's. See 'thread_fetch'
+ * implementation for details.
+ *
+ * @param master
+ * @param fd
+ * @param state the event to cancel. One or more (OR'd together) of the
+ * following:
+ *   - POLLIN
+ *   - POLLOUT
+ */
+static void
+thread_cancel_rw (struct thread_master *master, int fd, short state)
+{
+  /* Cancel POLLHUP too just in case some bozo set it */
+  state |= POLLHUP;
+
+  /* find the index of corresponding pollfd */
+  nfds_t i;
+
+  for (i = 0; i < master->handler.pfdcount; i++)
+    if (master->handler.pfds[i].fd == fd)
+      break;
+
+  /* NOT out event. */
+  master->handler.pfds[i].events &= ~(state);
+
+  /* If all events are canceled, delete / resize the pollfd array. */
+  if (master->handler.pfds[i].events == 0)
+    {
+      memmove(master->handler.pfds + i, master->handler.pfds + i + 1,
+              (master->handler.pfdcount - i - 1) * sizeof (struct pollfd));
+      master->handler.pfdcount--;
+    }
+
+  /* If we have the same pollfd in the copy, perform the same operations,
+   * otherwise return. */
+  if (i >= master->handler.copycount)
+    return;
+
+  master->handler.copy[i].events &= ~(state);
+
+  if (master->handler.copy[i].events == 0)
+    {
+      memmove(master->handler.copy + i, master->handler.copy + i + 1,
+              (master->handler.copycount - i - 1) * sizeof (struct pollfd));
+      master->handler.copycount--;
+    }
 }
 
 /**
- * Cancel thread from scheduler.
+ * Process cancellation requests.
  *
- * This function is *NOT* MT-safe. DO NOT call it from any other pthread except
- * the one which owns thread->master. You will crash.
+ * This may only be run from the pthread which owns the thread_master.
+ *
+ * @param master the thread master to process
+ * @REQUIRE master->mtx
  */
-void
-thread_cancel (struct thread *thread)
+static void
+do_thread_cancel (struct thread_master *master)
 {
   struct thread_list *list = NULL;
   struct pqueue *queue = NULL;
   struct thread **thread_array = NULL;
-
-  pthread_mutex_lock (&thread->mtx);
-  pthread_mutex_lock (&thread->master->mtx);
-
-  assert (pthread_self() == thread->master->owner);
-
-  switch (thread->type)
-    {
-    case THREAD_READ:
-      thread_cancel_read_or_write (thread, POLLIN | POLLHUP);
-      thread_array = thread->master->read;
-      break;
-    case THREAD_WRITE:
-      thread_cancel_read_or_write (thread, POLLOUT | POLLHUP);
-      thread_array = thread->master->write;
-      break;
-    case THREAD_TIMER:
-      queue = thread->master->timer;
-      break;
-    case THREAD_EVENT:
-      list = &thread->master->event;
-      break;
-    case THREAD_READY:
-      list = &thread->master->ready;
-      break;
-    default:
-      goto done;
-      break;
-    }
-
-  if (queue)
-    {
-      assert(thread->index >= 0);
-      pqueue_remove (thread, queue);
-    }
-  else if (list)
-    {
-      thread_list_delete (list, thread);
-    }
-  else if (thread_array)
-    {
-      thread_array[thread->u.fd] = NULL;
-    }
-  else
-    {
-      assert(!"Thread should be either in queue or list or array!");
-    }
-
-  if (thread->ref)
-    *thread->ref = NULL;
-
-  thread_add_unuse (thread->master, thread);
-
-done:
-  pthread_mutex_unlock (&thread->master->mtx);
-  pthread_mutex_unlock (&thread->mtx);
-}
-
-/* Delete all events which has argument value arg. */
-unsigned int
-thread_cancel_event (struct thread_master *m, void *arg)
-{
-  unsigned int ret = 0;
   struct thread *thread;
-  struct thread *t;
 
-  pthread_mutex_lock (&m->mtx);
-  {
-    thread = m->event.head;
-    while (thread)
-      {
-        t = thread;
-        pthread_mutex_lock (&t->mtx);
+  struct cancel_req *cr;
+  struct listnode *ln;
+  for (ALL_LIST_ELEMENTS_RO (master->cancel_req, ln, cr))
+    {
+      /* If this is an event object cancellation, linear search through event
+       * list deleting any events which have the specified argument. We also
+       * need to check every thread in the ready queue. */
+      if (cr->eventobj)
         {
-          thread = t->next;
+          struct thread *t;
+          thread = master->event.head;
 
-          if (t->arg == arg)
+          while (thread)
             {
-              ret++;
-              thread_list_delete (&m->event, t);
-              if (t->ref)
-                *t->ref = NULL;
-              thread_add_unuse (m, t);
-            }
-        }
-        pthread_mutex_unlock (&t->mtx);
-      }
+              t = thread;
+              thread = t->next;
 
-    /* thread can be on the ready list too */
-    thread = m->ready.head;
-    while (thread)
-      {
-        t = thread;
-        pthread_mutex_lock (&t->mtx);
+              if (t->arg == cr->eventobj)
+                {
+                  thread_list_delete (&master->event, t);
+                  if (t->ref)
+                    *t->ref = NULL;
+                  thread_add_unuse (master, t);
+                }
+            }
+
+          thread = master->ready.head;
+          while (thread)
+            {
+              t = thread;
+              thread = t->next;
+
+              if (t->arg == cr->eventobj)
+                {
+                  thread_list_delete (&master->ready, t);
+                  if (t->ref)
+                    *t->ref = NULL;
+                  thread_add_unuse (master, t);
+                }
+            }
+          continue;
+        }
+
+      /* The pointer varies depending on whether the cancellation request was
+       * made asynchronously or not. If it was, we need to check whether the
+       * thread even exists anymore before cancelling it. */
+      thread = (cr->thread) ? cr->thread : *cr->threadref;
+
+      if (!thread)
+        continue;
+
+      /* Determine the appropriate queue to cancel the thread from */
+      switch (thread->type)
         {
-          thread = t->next;
-
-          if (t->arg == arg)
-            {
-              ret++;
-              thread_list_delete (&m->ready, t);
-              if (t->ref)
-                *t->ref = NULL;
-              thread_add_unuse (m, t);
-            }
+        case THREAD_READ:
+          thread_cancel_rw (master, thread->u.fd, POLLIN);
+          thread_array = master->read;
+          break;
+        case THREAD_WRITE:
+          thread_cancel_rw (master, thread->u.fd, POLLOUT);
+          thread_array = master->write;
+          break;
+        case THREAD_TIMER:
+          queue = master->timer;
+          break;
+        case THREAD_EVENT:
+          list = &master->event;
+          break;
+        case THREAD_READY:
+          list = &master->ready;
+          break;
+        default:
+          continue;
+          break;
         }
-        pthread_mutex_unlock (&t->mtx);
-      }
-  }
-  pthread_mutex_unlock (&m->mtx);
-  return ret;
+
+      if (queue)
+        {
+          assert(thread->index >= 0);
+          pqueue_remove (thread, queue);
+        }
+      else if (list)
+        {
+          thread_list_delete (list, thread);
+        }
+      else if (thread_array)
+        {
+          thread_array[thread->u.fd] = NULL;
+        }
+      else
+        {
+          assert(!"Thread should be either in queue or list or array!");
+        }
+
+      if (thread->ref)
+        *thread->ref = NULL;
+
+      thread_add_unuse (thread->master, thread);
+    }
+
+  /* Delete and free all cancellation requests */
+  list_delete_all_node (master->cancel_req);
+
+  /* Wake up any threads which may be blocked in thread_cancel_async() */
+  master->canceled = true;
+  pthread_cond_broadcast (&master->cancel_cond);
 }
 
+/**
+ * Cancel any events which have the specified argument.
+ *
+ * MT-Unsafe
+ *
+ * @param m the thread_master to cancel from
+ * @param arg the argument passed when creating the event
+ */
+void
+thread_cancel_event (struct thread_master *master, void *arg)
+{
+  assert (master->owner == pthread_self());
+
+  pthread_mutex_lock (&master->mtx);
+  {
+    struct cancel_req *cr = XCALLOC (MTYPE_TMP, sizeof (struct cancel_req));
+    cr->eventobj = arg;
+    listnode_add (master->cancel_req, cr);
+    do_thread_cancel(master);
+  }
+  pthread_mutex_unlock (&master->mtx);
+}
+
+/**
+ * Cancel a specific task.
+ *
+ * MT-Unsafe
+ *
+ * @param thread task to cancel
+ */
+void
+thread_cancel (struct thread *thread)
+{
+  assert (thread->master->owner == pthread_self());
+
+  pthread_mutex_lock (&thread->master->mtx);
+  {
+    struct cancel_req *cr = XCALLOC (MTYPE_TMP, sizeof (struct cancel_req));
+    cr->thread = thread;
+    listnode_add (thread->master->cancel_req, cr);
+    do_thread_cancel (thread->master);
+  }
+  pthread_mutex_unlock (&thread->master->mtx);
+}
+
+/**
+ * Asynchronous cancellation.
+ *
+ * Called with either a struct thread ** or void * to an event argument,
+ * this function posts the correct cancellation request and blocks until it is
+ * serviced.
+ *
+ * If the thread is currently running, execution blocks until it completes.
+ *
+ * The last two parameters are mutually exclusive, i.e. if you pass one the
+ * other must be NULL.
+ *
+ * When the cancellation procedure executes on the target thread_master, the
+ * thread * provided is checked for nullity. If it is null, the thread is
+ * assumed to no longer exist and the cancellation request is a no-op. Thus
+ * users of this API must pass a back-reference when scheduling the original
+ * task.
+ *
+ * MT-Safe
+ *
+ * @param master the thread master with the relevant event / task
+ * @param thread pointer to thread to cancel
+ * @param eventobj the event
+ */
+void
+thread_cancel_async (struct thread_master *master, struct thread **thread,
+                     void *eventobj)
+{
+  assert (!(thread && eventobj) && (thread || eventobj));
+  assert (master->owner != pthread_self());
+
+  pthread_mutex_lock (&master->mtx);
+  {
+    master->canceled = false;
+
+    if (thread)
+      {
+        struct cancel_req *cr =
+          XCALLOC (MTYPE_TMP, sizeof (struct cancel_req));
+        cr->threadref = thread;
+        listnode_add (master->cancel_req, cr);
+      }
+    else if (eventobj)
+      {
+        struct cancel_req *cr =
+          XCALLOC (MTYPE_TMP, sizeof (struct cancel_req));
+        cr->eventobj = eventobj;
+        listnode_add (master->cancel_req, cr);
+      }
+    AWAKEN (master);
+
+    while (!master->canceled)
+      pthread_cond_wait (&master->cancel_cond, &master->mtx);
+  }
+  pthread_mutex_unlock (&master->mtx);
+}
+/* ------------------------------------------------------------------------- */
+
 static struct timeval *
 thread_timer_wait (struct pqueue *queue, struct timeval *timer_val)
 {
@@ -1053,13 +1201,22 @@ thread_process_io_helper (struct thread_master *m, struct thread *thread,
   return 1;
 }
 
+/**
+ * Process I/O events.
+ *
+ * Walks through file descriptor array looking for those pollfds whose .revents
+ * field has something interesting. Deletes any invalid file descriptors.
+ *
+ * @param m the thread master
+ * @param num the number of active file descriptors (return value of poll())
+ */
 static void
-thread_process_io (struct thread_master *m, struct pollfd *pfds,
-        unsigned int num, unsigned int count)
+thread_process_io (struct thread_master *m, unsigned int num)
 {
   unsigned int ready = 0;
+  struct pollfd *pfds = m->handler.copy;
 
-  for (nfds_t i = 0; i < count && ready < num ; ++i)
+  for (nfds_t i = 0; i < m->handler.copycount && ready < num ; ++i)
     {
       /* no event for current fd? immediately continue */
       if (pfds[i].revents == 0)
@@ -1088,8 +1245,9 @@ thread_process_io (struct thread_master *m, struct pollfd *pfds,
           m->handler.pfdcount--;
 
           memmove (pfds + i, pfds + i + 1,
-                   (count - i - 1) * sizeof(struct pollfd));
-          count--;
+                   (m->handler.copycount - i - 1) * sizeof(struct pollfd));
+          m->handler.copycount--;
+
           i--;
         }
     }
@@ -1139,98 +1297,108 @@ thread_process (struct thread_list *list)
 struct thread *
 thread_fetch (struct thread_master *m, struct thread *fetch)
 {
-  struct thread *thread;
+  struct thread *thread = NULL;
   struct timeval now;
-  struct timeval timer_val = { .tv_sec = 0, .tv_usec = 0 };
-  struct timeval *timer_wait = &timer_val;
+  struct timeval zerotime = { 0, 0 };
+  struct timeval tv;
+  struct timeval *tw;
 
-  do
-    {
-      int num = 0;
+  int num = 0;
 
-      /* Signals pre-empt everything */
-      if (m->handle_signals)
-        quagga_sigevent_process ();
-       
-      pthread_mutex_lock (&m->mtx);
-      /* Drain the ready queue of already scheduled jobs, before scheduling
-       * more.
-       */
-      if ((thread = thread_trim_head (&m->ready)) != NULL)
-        {
-          fetch = thread_run (m, thread, fetch);
-          if (fetch->ref)
-            *fetch->ref = NULL;
-          pthread_mutex_unlock (&m->mtx);
-          return fetch;
-        }
-      
-      /* To be fair to all kinds of threads, and avoid starvation, we
-       * need to be careful to consider all thread types for scheduling
-       * in each quanta. I.e. we should not return early from here on.
-       */
-       
-      /* Normal event are the next highest priority.  */
-      thread_process (&m->event);
-      
-      /* Calculate select wait timer if nothing else to do */
-      if (m->ready.count == 0)
-        {
-          timer_wait = thread_timer_wait (m->timer, &timer_val);
-        }
+  do {
+    /* Handle signals if any */
+    if (m->handle_signals)
+      quagga_sigevent_process ();
 
-      if (timer_wait && timer_wait->tv_sec < 0)
-        {
-          timerclear(&timer_val);
-          timer_wait = &timer_val;
-        }
+    pthread_mutex_lock (&m->mtx);
 
-      unsigned int count = m->handler.pfdcount + m->handler.pfdcountsnmp;
-      memcpy (m->handler.copy, m->handler.pfds, count * sizeof (struct pollfd));
+    /* Process any pending cancellation requests */
+    do_thread_cancel (m);
 
-      pthread_mutex_unlock (&m->mtx);
-      {
-        num = fd_poll (m, m->handler.copy, m->handler.pfdsize, count, timer_wait);
+    /* Post events to ready queue. This must come before the following block
+     * since events should occur immediately */
+    thread_process (&m->event);
+
+    /* If there are no tasks on the ready queue, we will poll() until a timer
+     * expires or we receive I/O, whichever comes first. The strategy for doing
+     * this is:
+     *
+     * - If there are events pending, set the poll() timeout to zero
+     * - If there are no events pending, but there are timers pending, set the
+     *   timeout to the smallest remaining time on any timer
+     * - If there are neither timers nor events pending, but there are file
+     *   descriptors pending, block indefinitely in poll()
+     * - If nothing is pending, it's time for the application to die
+     *
+     * In every case except the last, we need to hit poll() at least once per
+     * loop to avoid starvation by events */
+
+    if (m->ready.count == 0)
+      tw = thread_timer_wait (m->timer, &tv);
+
+    if (m->ready.count != 0 || (tw && !timercmp (tw, &zerotime, >)))
+      tw = &zerotime;
+
+    if (!tw && m->handler.pfdcount == 0)
+      { /* die */
+        pthread_mutex_unlock (&m->mtx);
+        fetch = NULL;
+        break;
       }
-      pthread_mutex_lock (&m->mtx);
 
-      /* Signals should get quick treatment */
-      if (num < 0)
-        {
-          if (errno == EINTR)
-            {
-              pthread_mutex_unlock (&m->mtx);
-              continue; /* signal received - process it */
-            }
-          zlog_warn ("poll() error: %s", safe_strerror (errno));
-          pthread_mutex_unlock (&m->mtx);
-          return NULL;
-        }
+    /* Copy pollfd array + # active pollfds in it. Not necessary to copy
+     * the array size as this is fixed. */
+    m->handler.copycount = m->handler.pfdcount;
+    memcpy (m->handler.copy, m->handler.pfds,
+            m->handler.copycount * sizeof (struct pollfd));
 
-      /* Check foreground timers.  Historically, they have had higher
-       * priority than I/O threads, so let's push them onto the ready
-       * list in front of the I/O threads. */
-      monotime(&now);
-      thread_process_timers (m->timer, &now);
-      
-      /* Got IO, process it */
-      if (num > 0)
-        thread_process_io (m, m->handler.copy, num, count);
+    pthread_mutex_unlock (&m->mtx);
+    {
+      num = fd_poll (m, m->handler.copy, m->handler.pfdsize,
+                     m->handler.copycount, tw);
+    }
+    pthread_mutex_lock (&m->mtx);
 
-      if ((thread = thread_trim_head (&m->ready)) != NULL)
-        {
-          fetch = thread_run (m, thread, fetch);
-          if (fetch->ref)
-            *fetch->ref = NULL;
-          pthread_mutex_unlock (&m->mtx);
-          return fetch;
-        }
+    /* Handle any errors received in poll() */
+    if (num < 0)
+      {
+        if (errno == EINTR)
+          {
+            pthread_mutex_unlock (&m->mtx);
+            continue; /* loop around to signal handler */
+          }
 
-      pthread_mutex_unlock (&m->mtx);
+        /* else die */
+        zlog_warn ("poll() error: %s", safe_strerror (errno));
+        pthread_mutex_unlock (&m->mtx);
+        fetch = NULL;
+        break;
+      }
 
-    } while (m->spin);
+    /* Since we could have received more cancellation requests during poll(), process those */
+    do_thread_cancel (m);
 
-  return NULL;
+    /* Post timers to ready queue. */
+    monotime(&now);
+    thread_process_timers (m->timer, &now);
+
+    /* Post I/O to ready queue. */
+    if (num > 0)
+      thread_process_io (m, num);
+
+    /* If we have a ready task, break the loop and return it to the caller */
+    if ((thread = thread_trim_head (&m->ready)))
+      {
+        fetch = thread_run (m, thread, fetch);
+        if (fetch->ref)
+          *fetch->ref = NULL;
+      }
+
+    pthread_mutex_unlock (&m->mtx);
+
+  } while (!thread && m->spin);
+
+  return fetch;
 }
 
 unsigned long

lib/thread.h

@@ -47,16 +47,25 @@ struct pqueue;
 
 struct fd_handler
 {
-  /* number of pfd stored in pfds */
-  nfds_t pfdcount;
-  /* number of pfd stored in pfds + number of snmp pfd */
-  nfds_t pfdcountsnmp;
-  /* number of pfd that fit in the allocated space of pfds */
+  /* number of pfd that fit in the allocated space of pfds. This is a constant
+   * and is the same for both pfds and copy. */
   nfds_t pfdsize;
+
   /* file descriptors to monitor for i/o */
   struct pollfd *pfds;
+  /* number of pollfds stored in pfds */
+  nfds_t pfdcount;
+
   /* chunk used for temp copy of pollfds */
   struct pollfd *copy;
+  /* number of pollfds stored in copy */
+  nfds_t copycount;
+};
+
+struct cancel_req {
+  struct thread *thread;
+  void *eventobj;
+  struct thread **threadref;
 };
 
 /* Master of the theads. */
@@ -68,6 +77,9 @@ struct thread_master
   struct thread_list event;
   struct thread_list ready;
   struct thread_list unuse;
+  struct list *cancel_req;
+  bool canceled;
+  pthread_cond_t cancel_cond;
   int io_pipe[2];
   int fd_limit;
   struct fd_handler handler;
@@ -189,7 +201,8 @@ extern void funcname_thread_execute (struct thread_master *,
 #undef debugargdef
 
 extern void thread_cancel (struct thread *);
-extern unsigned int thread_cancel_event (struct thread_master *, void *);
+extern void thread_cancel_async (struct thread_master *, struct thread **, void *);
+extern void thread_cancel_event (struct thread_master *, void *);
 extern struct thread *thread_fetch (struct thread_master *, struct thread *);
 extern void thread_call (struct thread *);
 extern unsigned long thread_timer_remain_second (struct thread *);