Exemple #1
0
  protected void _handleMergeRequest(
      Address sender, MergeId merge_id, Collection<? extends Address> mbrs) throws Exception {
    boolean success = matchMergeId(merge_id) || setMergeId(null, merge_id);
    if (!success)
      throw new Exception(
          "merge " + this.merge_id + " is already in progress, received merge-id=" + merge_id);

    /* Clears the view handler queue and discards all JOIN/LEAVE/MERGE requests until after the MERGE  */
    // gms.getViewHandler().suspend();
    if (log.isTraceEnabled())
      log.trace(
          gms.local_addr
              + ": got merge request from "
              + sender
              + ", merge_id="
              + merge_id
              + ", mbrs="
              + mbrs);

    // merge the membership of the current view with mbrs
    List<Address> members = new LinkedList<Address>();
    if (mbrs
        != null) { // didn't use a set because we didn't want to change the membership order at this
      // time (although not incorrect)
      for (Address mbr : mbrs) {
        if (!members.contains(mbr)) members.add(mbr);
      }
    }

    ViewId tmp_vid = gms.getViewId();
    if (tmp_vid != null) tmp_vid = tmp_vid.copy();
    if (tmp_vid == null) throw new Exception("view ID is null; cannot return merge response");

    View view = new View(tmp_vid, new ArrayList<Address>(members));

    // [JGRP-524] - FLUSH and merge: flush doesn't wrap entire merge process
    // [JGRP-770] - Concurrent startup of many channels doesn't stabilize
    // [JGRP-700] - FLUSH: flushing should span merge

    /*if flush is in stack, let this coordinator flush its cluster island */
    if (gms.flushProtocolInStack && !gms.startFlush(view)) throw new Exception("flush failed");

    // we still need to fetch digests from all members, and not just return our own digest
    // (https://issues.jboss.org/browse/JGRP-948)
    Digest digest = fetchDigestsFromAllMembersInSubPartition(members, merge_id);
    if (digest == null || digest.size() == 0)
      throw new Exception(
          "failed fetching digests from subpartition members; dropping merge response");

    sendMergeResponse(sender, view, digest, merge_id);
  }
Exemple #2
0
    /**
     * Merge all MergeData. All MergeData elements should be disjunct (both views and digests).
     * However, this method is prepared to resolve duplicate entries (for the same member).
     * Resolution strategy for views is to merge only 1 of the duplicate members. Resolution
     * strategy for digests is to take the higher seqnos for duplicate digests.
     *
     * <p>After merging all members into a Membership and subsequent sorting, the first member of
     * the sorted membership will be the new coordinator. This method has a lock on merge_rsps.
     *
     * @param merge_rsps A list of MergeData items. Elements with merge_rejected=true were removed
     *     before. Is guaranteed not to be null and to contain at least 1 member.
     */
    private MergeData consolidateMergeData(Vector<MergeData> merge_rsps) {
      long logical_time = 0; // for new_vid
      Membership new_mbrs = new Membership();
      Vector<View> subgroups =
          new Vector<View>(11); // contains a list of Views, each View is a subgroup

      for (MergeData tmp_data : merge_rsps) {
        View tmp_view = tmp_data.getView();
        if (tmp_view != null) {
          ViewId tmp_vid = tmp_view.getVid();
          if (tmp_vid != null) {
            // compute the new view id (max of all vids +1)
            logical_time = Math.max(logical_time, tmp_vid.getId());
          }
          // merge all membership lists into one (prevent duplicates)
          new_mbrs.add(tmp_view.getMembers());
          subgroups.addElement((View) tmp_view.clone());
        }
      }

      // the new coordinator is the first member of the consolidated & sorted membership list
      new_mbrs.sort();
      Address new_coord = new_mbrs.size() > 0 ? new_mbrs.elementAt(0) : null;
      if (new_coord == null) {
        if (log.isErrorEnabled()) log.error("new_coord == null");
        return null;
      }
      // should be the highest view ID seen up to now plus 1
      ViewId new_vid = new ViewId(new_coord, logical_time + 1);

      // determine the new view
      MergeView new_view = new MergeView(new_vid, new_mbrs.getMembers(), subgroups);

      // determine the new digest
      Digest new_digest = consolidateDigests(merge_rsps, new_mbrs.size());
      if (new_digest == null) {
        if (log.isErrorEnabled())
          log.error("Merge leader " + gms.local_addr + ": could not consolidate digest for merge");
        return null;
      }
      if (log.isDebugEnabled())
        log.debug(
            "Merge leader "
                + gms.local_addr
                + ": consolidated view="
                + new_view
                + "\nconsolidated digest="
                + new_digest);
      return new MergeData(gms.local_addr, new_view, new_digest);
    }
Exemple #3
0
    /**
     * Merge all MergeData. All MergeData elements should be disjunct (both views and digests).
     * However, this method is prepared to resolve duplicate entries (for the same member).
     * Resolution strategy for views is to merge only 1 of the duplicate members. Resolution
     * strategy for digests is to take the higher seqnos for duplicate digests.
     *
     * <p>After merging all members into a Membership and subsequent sorting, the first member of
     * the sorted membership will be the new coordinator. This method has a lock on merge_rsps.
     *
     * @param merge_rsps A list of MergeData items. Elements with merge_rejected=true were removed
     *     before. Is guaranteed not to be null and to contain at least 1 member.
     */
    private MergeData consolidateMergeData(List<MergeData> merge_rsps) {
      long logical_time = 0; // for new_vid
      List<View> subgroups =
          new ArrayList<View>(11); // contains a list of Views, each View is a subgroup
      Collection<Collection<Address>> sub_mbrships = new ArrayList<Collection<Address>>();

      for (MergeData tmp_data : merge_rsps) {
        View tmp_view = tmp_data.getView();
        if (tmp_view != null) {
          ViewId tmp_vid = tmp_view.getVid();
          if (tmp_vid != null) {
            // compute the new view id (max of all vids +1)
            logical_time = Math.max(logical_time, tmp_vid.getId());
          }
          // merge all membership lists into one (prevent duplicates)
          sub_mbrships.add(new ArrayList<Address>(tmp_view.getMembers()));
          subgroups.add(tmp_view.copy());
        }
      }

      // determine the new digest
      Digest new_digest = consolidateDigests(merge_rsps, merge_rsps.size());
      if (new_digest == null) return null;

      // remove all members from the new member list that are not in the digest
      Collection<Address> digest_mbrs = new_digest.getMembers();
      for (Collection<Address> coll : sub_mbrships) coll.retainAll(digest_mbrs);

      List<Address> merged_mbrs = gms.computeNewMembership(sub_mbrships);

      // the new coordinator is the first member of the consolidated & sorted membership list
      Address new_coord = merged_mbrs.isEmpty() ? null : merged_mbrs.get(0);
      if (new_coord == null) return null;

      // should be the highest view ID seen up to now plus 1
      ViewId new_vid = new ViewId(new_coord, logical_time + 1);

      // determine the new view
      MergeView new_view = new MergeView(new_vid, merged_mbrs, subgroups);

      if (log.isTraceEnabled())
        log.trace(
            gms.local_addr
                + ": consolidated view="
                + new_view
                + "\nconsolidated digest="
                + new_digest);
      return new MergeData(gms.local_addr, new_view, new_digest);
    }
Exemple #4
0
  /**
   * Computes the next view. Returns a copy that has <code>old_mbrs</code> and <code>suspected_mbrs
   * </code> removed and <code>new_mbrs</code> added.
   */
  public View getNextView(
      Collection<Address> new_mbrs,
      Collection<Address> old_mbrs,
      Collection<Address> suspected_mbrs) {
    synchronized (members) {
      ViewId view_id = view != null ? view.getViewId() : null;
      if (view_id == null) {
        log.error("view_id is null");
        return null; // this should *never* happen !
      }
      long vid = Math.max(view_id.getId(), ltime) + 1;
      ltime = vid;
      Membership tmp_mbrs = tmp_members.copy(); // always operate on the temporary membership
      tmp_mbrs.remove(suspected_mbrs);
      tmp_mbrs.remove(old_mbrs);
      tmp_mbrs.add(new_mbrs);
      List<Address> mbrs = tmp_mbrs.getMembers();
      Address new_coord = local_addr;
      if (!mbrs.isEmpty()) new_coord = mbrs.get(0);
      View v = new View(new_coord, vid, mbrs);

      // Update membership (see DESIGN for explanation):
      tmp_members.set(mbrs);

      // Update joining list (see DESIGN for explanation)
      if (new_mbrs != null) {
        for (Address tmp_mbr : new_mbrs) {
          if (!joining.contains(tmp_mbr)) joining.add(tmp_mbr);
        }
      }

      // Update leaving list (see DESIGN for explanations)
      if (old_mbrs != null) {
        for (Address addr : old_mbrs) {
          if (!leaving.contains(addr)) leaving.add(addr);
        }
      }
      if (suspected_mbrs != null) {
        for (Address addr : suspected_mbrs) {
          if (!leaving.contains(addr)) leaving.add(addr);
        }
      }
      return v;
    }
  }
Exemple #5
0
  /**
   * An event was received from the layer below. Usually the current layer will want to examine the
   * event type and - depending on its type - perform some computation (e.g. removing headers from a
   * MSG event type, or updating the internal membership list when receiving a VIEW_CHANGE event).
   * Finally the event is either a) discarded, or b) an event is sent down the stack using <code>
   * PassDown</code> or c) the event (or another event) is sent up the stack using <code>PassUp
   * </code>.
   *
   * <p>For the PING protocol, the Up operation does the following things. 1. If the event is a
   * Event.MSG then PING will inspect the message header. If the header is null, PING simply passes
   * up the event If the header is PingHeader.GET_MBRS_REQ then the PING protocol will PassDown a
   * PingRequest message If the header is PingHeader.GET_MBRS_RSP we will add the message to the
   * initial members vector and wake up any waiting threads. 2. If the event is Event.SET_LOCAL_ADDR
   * we will simple set the local address of this protocol 3. For all other messages we simple pass
   * it up to the protocol above
   *
   * @param evt - the event that has been sent from the layer below
   */
  @SuppressWarnings("unchecked")
  public Object up(Event evt) {

    switch (evt.getType()) {
      case Event.MSG:
        Message msg = (Message) evt.getArg();
        PingHeader hdr = (PingHeader) msg.getHeader(this.id);
        if (hdr == null) return up_prot.up(evt);

        PingData data = hdr.data;
        Address logical_addr = data != null ? data.getAddress() : null;

        if (is_leaving) return null; // prevents merging back a leaving member
        // (https://issues.jboss.org/browse/JGRP-1336)

        switch (hdr.type) {
          case PingHeader.GET_MBRS_REQ: // return Rsp(local_addr, coord)
            if (group_addr == null || hdr.cluster_name == null) {
              if (log.isWarnEnabled())
                log.warn(
                    "group_addr ("
                        + group_addr
                        + ") or cluster_name of header ("
                        + hdr.cluster_name
                        + ") is null; passing up discovery request from "
                        + msg.getSrc()
                        + ", but this should not"
                        + " be the case");
            } else {
              if (!group_addr.equals(hdr.cluster_name)) {
                if (log.isWarnEnabled())
                  log.warn(
                      "discarding discovery request for cluster '"
                          + hdr.cluster_name
                          + "' from "
                          + msg.getSrc()
                          + "; our cluster name is '"
                          + group_addr
                          + "'. "
                          + "Please separate your clusters cleanly.");
                return null;
              }
            }

            // add physical address and logical name of the discovery sender (if available) to the
            // cache
            if (data != null) {
              if (logical_addr == null) logical_addr = msg.getSrc();
              Collection<PhysicalAddress> physical_addrs = data.getPhysicalAddrs();
              PhysicalAddress physical_addr =
                  physical_addrs != null && !physical_addrs.isEmpty()
                      ? physical_addrs.iterator().next()
                      : null;
              if (logical_addr != null && physical_addr != null)
                down(
                    new Event(
                        Event.SET_PHYSICAL_ADDRESS,
                        new Tuple<Address, PhysicalAddress>(logical_addr, physical_addr)));
              if (logical_addr != null && data.getLogicalName() != null)
                UUID.add(logical_addr, data.getLogicalName());
              discoveryRequestReceived(msg.getSrc(), data.getLogicalName(), physical_addrs);

              synchronized (ping_responses) {
                for (Responses response : ping_responses) {
                  response.addResponse(data, false);
                }
              }
            }

            if (hdr.view_id != null) {
              // If the discovery request is merge-triggered, and the ViewId shipped with it
              // is the same as ours, we don't respond (JGRP-1315).
              ViewId my_view_id = view != null ? view.getViewId() : null;
              if (my_view_id != null && my_view_id.equals(hdr.view_id)) return null;

              boolean send_discovery_rsp =
                  force_sending_discovery_rsps
                      || is_coord
                      || current_coord == null
                      || current_coord.equals(msg.getSrc());
              if (!send_discovery_rsp) {
                if (log.isTraceEnabled())
                  log.trace(
                      local_addr
                          + ": suppressing merge response as I'm not a coordinator and the "
                          + "discovery request was not sent by a coordinator");
                return null;
              }
            }

            if (isMergeRunning()) {
              if (log.isTraceEnabled())
                log.trace(
                    local_addr + ": suppressing merge response as a merge is already in progress");
              return null;
            }

            List<PhysicalAddress> physical_addrs =
                hdr.view_id != null
                    ? null
                    : Arrays.asList(
                        (PhysicalAddress) down(new Event(Event.GET_PHYSICAL_ADDRESS, local_addr)));
            sendDiscoveryResponse(
                local_addr,
                physical_addrs,
                is_server,
                hdr.view_id != null,
                UUID.get(local_addr),
                msg.getSrc());
            return null;

          case PingHeader.GET_MBRS_RSP: // add response to vector and notify waiting thread
            // add physical address (if available) to transport's cache
            if (data != null) {
              Address response_sender = msg.getSrc();
              if (logical_addr == null) logical_addr = msg.getSrc();
              Collection<PhysicalAddress> addrs = data.getPhysicalAddrs();
              PhysicalAddress physical_addr =
                  addrs != null && !addrs.isEmpty() ? addrs.iterator().next() : null;
              if (logical_addr != null && physical_addr != null)
                down(
                    new Event(
                        Event.SET_PHYSICAL_ADDRESS,
                        new Tuple<Address, PhysicalAddress>(logical_addr, physical_addr)));
              if (logical_addr != null && data.getLogicalName() != null)
                UUID.add(logical_addr, data.getLogicalName());

              if (log.isTraceEnabled())
                log.trace(
                    local_addr + ": received GET_MBRS_RSP from " + response_sender + ": " + data);
              boolean overwrite = logical_addr != null && logical_addr.equals(response_sender);
              synchronized (ping_responses) {
                for (Responses response : ping_responses) {
                  response.addResponse(data, overwrite);
                }
              }
            }
            return null;

          default:
            if (log.isWarnEnabled())
              log.warn("got PING header with unknown type (" + hdr.type + ')');
            return null;
        }

      case Event.GET_PHYSICAL_ADDRESS:
        try {
          sendDiscoveryRequest(group_addr, null, null);
        } catch (InterruptedIOException ie) {
          if (log.isWarnEnabled()) {
            log.warn("Discovery request for cluster " + group_addr + " interrupted");
          }
          Thread.currentThread().interrupt();
        } catch (Exception ex) {
          if (log.isErrorEnabled()) log.error("failed sending discovery request", ex);
        }
        return null;

      case Event.FIND_INITIAL_MBRS: // sent by transport
        return findInitialMembers(null);
    }

    return up_prot.up(evt);
  }
Exemple #6
0
  /**
   * Sets the new view and sends a VIEW_CHANGE event up and down the stack. If the view is a
   * MergeView (subclass of View), then digest will be non-null and has to be set before installing
   * the view.
   */
  public void installView(View new_view, Digest digest) {
    ViewId vid = new_view.getVid();
    List<Address> mbrs = new_view.getMembers();
    ltime =
        Math.max(
            vid.getId(),
            ltime); // compute the logical time, regardless of whether the view is accepted

    // Discards view with id lower than or equal to our own. Will be installed without check if it
    // is the first view
    if (view != null) {
      ViewId view_id = view.getViewId();
      int rc = vid.compareToIDs(view_id);
      if (rc <= 0) {
        if (log.isWarnEnabled()
            && rc < 0
            && log_view_warnings) { // only scream if view is smaller, silently discard same views
          log.warn(
              local_addr
                  + ": received view < current view;"
                  + " discarding it (current vid: "
                  + view_id
                  + ", new vid: "
                  + vid
                  + ')');
        }
        return;
      }
    }

    /* Check for self-inclusion: if I'm not part of the new membership, I just discard it.
    This ensures that messages sent in view V1 are only received by members of V1 */
    if (!mbrs.contains(local_addr)) {
      if (log.isWarnEnabled() && log_view_warnings)
        log.warn(local_addr + ": not member of view " + new_view.getViewId() + "; discarding it");
      return;
    }

    if (digest != null) {
      if (new_view instanceof MergeView) mergeDigest(digest);
      else setDigest(digest);
    }

    if (log.isDebugEnabled()) log.debug(local_addr + ": installing view " + new_view);

    Event view_event;
    synchronized (members) {
      view = new View(new_view.getVid(), new_view.getMembers());
      view_event = new Event(Event.VIEW_CHANGE, new_view);

      // Set the membership. Take into account joining members
      if (!mbrs.isEmpty()) {
        members.set(mbrs);
        tmp_members.set(members);
        joining.removeAll(mbrs); // remove all members in mbrs from joining
        // remove all elements from 'leaving' that are not in 'mbrs'
        leaving.retainAll(mbrs);

        tmp_members.add(joining); // add members that haven't yet shown up in the membership
        tmp_members.remove(
            leaving); // remove members that haven't yet been removed from the membership

        // add to prev_members
        for (Address addr : mbrs) {
          if (!prev_members.contains(addr)) prev_members.add(addr);
        }
      }

      Address coord = determineCoordinator();
      if (coord != null && coord.equals(local_addr) && !haveCoordinatorRole()) {
        becomeCoordinator();
      } else {
        if (haveCoordinatorRole() && !local_addr.equals(coord)) {
          becomeParticipant();
          merge_ack_collector.reset(null); // we don't need this one anymore
        }
      }
    }

    // - Changed order of passing view up and down (http://jira.jboss.com/jira/browse/JGRP-347)
    // - Changed it back (bela Sept 4 2007): http://jira.jboss.com/jira/browse/JGRP-564
    // - Moved sending up view_event out of the synchronized block (bela Nov 2011)
    down_prot.down(view_event); // needed e.g. by failure detector or UDP
    up_prot.up(view_event);

    List<Address> tmp_mbrs = new_view.getMembers();
    ack_collector.retainAll(tmp_mbrs);
    merge_ack_collector.retainAll(tmp_mbrs);

    if (new_view instanceof MergeView) merger.forceCancelMerge();

    if (stats) {
      num_views++;
      prev_views.add(new Tuple<View, Long>(new_view, System.currentTimeMillis()));
    }
  }