/**
* 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);
}
public void handleMembershipChange(Collection<Request> requests) {
boolean joinAndStateTransferInitiated = false;
boolean useFlushIfPresent = gms.use_flush_if_present;
Collection<Address> new_mbrs = new LinkedHashSet<>(requests.size());
Collection<Address> suspected_mbrs = new LinkedHashSet<>(requests.size());
Collection<Address> leaving_mbrs = new LinkedHashSet<>(requests.size());
boolean self_leaving = false; // is the coord leaving
for (Request req : requests) {
switch (req.type) {
case Request.JOIN:
new_mbrs.add(req.mbr);
if (req.useFlushIfPresent) useFlushIfPresent = true;
break;
case Request.JOIN_WITH_STATE_TRANSFER:
new_mbrs.add(req.mbr);
joinAndStateTransferInitiated = true;
if (req.useFlushIfPresent) useFlushIfPresent = true;
break;
case Request.LEAVE:
if (req.suspected) suspected_mbrs.add(req.mbr);
else {
leaving_mbrs.add(req.mbr);
if (gms.local_addr != null && gms.local_addr.equals(req.mbr)) self_leaving = true;
}
break;
case Request.SUSPECT:
suspected_mbrs.add(req.mbr);
break;
}
}
new_mbrs.remove(gms.local_addr); // remove myself - cannot join myself (already joined)
if (gms.getViewId() == null) {
// we're probably not the coord anymore (we just left ourselves), let someone else do it
// (client will retry when it doesn't get a response)
log.debug(
"gms.view_id is null, I'm not the coordinator anymore (leaving=%b); "
+ "the new coordinator will handle the leave request",
self_leaving);
return;
}
List<Address> current_members = gms.members.getMembers();
leaving_mbrs.retainAll(
current_members); // remove all elements of leaving_mbrs which are not current members
if (suspected_mbrs.remove(gms.local_addr))
log.warn("I am the coord and I'm being suspected -- will probably leave shortly");
suspected_mbrs.retainAll(
current_members); // remove all elements of suspected_mbrs which are not current members
// for the members that have already joined, return the current digest and membership
for (Iterator<Address> it = new_mbrs.iterator(); it.hasNext(); ) {
Address mbr = it.next();
if (gms.members.contains(mbr)) { // already joined: return current digest and membership
log.trace(
"%s: %s already present; returning existing view %s", gms.local_addr, mbr, gms.view);
Tuple<View, Digest> tuple = gms.getViewAndDigest();
if (tuple != null) gms.sendJoinResponse(new JoinRsp(tuple.getVal1(), tuple.getVal2()), mbr);
else
log.warn(
"%s: did not find a digest matching view %s; dropping JOIN-RSP",
gms.local_addr, gms.view);
it
.remove(); // remove it anyway, even if we didn't find a digest matching the view
// (joiner will retry)
}
}
if (new_mbrs.isEmpty() && leaving_mbrs.isEmpty() && suspected_mbrs.isEmpty()) {
log.trace("%s: found no members to add or remove, will not create new view", gms.local_addr);
return;
}
View new_view = gms.getNextView(new_mbrs, leaving_mbrs, suspected_mbrs);
if (new_view.size() == 0
&& gms.local_addr != null
&& gms.local_addr.equals(new_view.getCreator())) {
if (self_leaving) gms.initState(); // in case connect() is called again
return;
}
log.trace(
"%s: joiners=%s, suspected=%s, leaving=%s, new view: %s",
gms.local_addr, new_mbrs, suspected_mbrs, leaving_mbrs, new_view);
JoinRsp join_rsp = null;
boolean hasJoiningMembers = !new_mbrs.isEmpty();
try {
boolean successfulFlush =
!useFlushIfPresent || !gms.flushProtocolInStack || gms.startFlush(new_view);
if (!successfulFlush && hasJoiningMembers) {
// Don't send a join response if the flush fails
// (http://jira.jboss.org/jira/browse/JGRP-759)
// The joiner should block until the previous FLUSH completed
sendLeaveResponses(leaving_mbrs); // we still have to send potential leave responses
// but let the joining client timeout and send another join request
return;
}
// we cannot garbage collect during joining a new member *if* we're the only member
// Example: {A}, B joins, after returning JoinRsp to B, A garbage collects messages higher
// than those
// in the digest returned to the client, so the client will *not* be able to ask for
// retransmission
// of those messages if he misses them
if (hasJoiningMembers) {
gms.getDownProtocol().down(new Event(Event.SUSPEND_STABLE, MAX_SUSPEND_TIMEOUT));
// create a new digest, which contains the new members, minus left members
MutableDigest join_digest =
new MutableDigest(new_view.getMembersRaw()).set(gms.getDigest());
for (Address member : new_mbrs)
join_digest.set(member, 0, 0); // ... and set the new members. their first seqno will be 1
// If the digest from NAKACK doesn't include all members of the view, we try once more; if
// it is still
// incomplete, we don't send a JoinRsp back to the joiner(s). This shouldn't be a problem as
// they will retry
if (join_digest.allSet() || join_digest.set(gms.getDigest()).allSet())
join_rsp = new JoinRsp(new_view, join_digest);
else
log.warn(
"%s: digest does not match view (missing seqnos for %s); dropping JOIN-RSP",
gms.local_addr, Arrays.toString(join_digest.getNonSetMembers()));
}
sendLeaveResponses(leaving_mbrs); // no-op if no leaving members
// we don't need to send the digest to existing members:
// https://issues.jboss.org/browse/JGRP-1317
gms.castViewChange(new_view, null, new_mbrs);
gms.sendJoinResponses(join_rsp, new_mbrs);
} finally {
if (hasJoiningMembers) gms.getDownProtocol().down(new Event(Event.RESUME_STABLE));
if (!joinAndStateTransferInitiated && useFlushIfPresent) gms.stopFlush();
if (self_leaving) gms.initState(); // in case connect() is called again
}
}