/** * @param cctx Context. * @param id Partition ID. */ @SuppressWarnings("ExternalizableWithoutPublicNoArgConstructor") GridDhtLocalPartition(GridCacheContext cctx, int id) { assert cctx != null; this.id = id; this.cctx = cctx; log = U.logger(cctx.kernalContext(), logRef, this); rent = new GridFutureAdapter<Object>() { @Override public String toString() { return "PartitionRentFuture [part=" + GridDhtLocalPartition.this + ", map=" + map + ']'; } }; map = new ConcurrentHashMap8<>(cctx.config().getStartSize() / cctx.affinity().partitions()); int delQueueSize = CU.isSystemCache(cctx.name()) ? 100 : Math.max(MAX_DELETE_QUEUE_SIZE / cctx.affinity().partitions(), 20); rmvQueue = new GridCircularBuffer<>(U.ceilPow2(delQueueSize)); }
/** {@inheritDoc} */ @Override void onCancelAtStop() { super.onCancelAtStop(); for (GridCacheQueryFutureAdapter fut : futs.values()) try { fut.cancel(); } catch (IgniteCheckedException e) { U.error(log, "Failed to cancel running query future: " + fut, e); } U.interrupt(threads.values()); }
/** Clears swap entries for evicted partition. */ private void clearSwap() { assert state() == EVICTED; assert !GridQueryProcessor.isEnabled(cctx.config()) : "Indexing needs to have unswapped values."; try { GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> it = cctx.swap().iterator(id); boolean isLocStore = cctx.store().isLocal(); if (it != null) { // We can safely remove these values because no entries will be created for evicted // partition. while (it.hasNext()) { Map.Entry<byte[], GridCacheSwapEntry> entry = it.next(); byte[] keyBytes = entry.getKey(); KeyCacheObject key = cctx.toCacheKeyObject(keyBytes); cctx.swap().remove(key); if (isLocStore) cctx.store().remove(null, key.value(cctx.cacheObjectContext(), false)); } } } catch (IgniteCheckedException e) { U.error(log, "Failed to clear swap for evicted partition: " + this, e); } }
/** @return Nodes to execute on. */ private Collection<ClusterNode> nodes() { CacheMode cacheMode = cctx.config().getCacheMode(); switch (cacheMode) { case LOCAL: if (prj != null) U.warn( log, "Ignoring query projection because it's executed over LOCAL cache " + "(only local node will be queried): " + this); return Collections.singletonList(cctx.localNode()); case REPLICATED: if (prj != null || partition() != null) return nodes(cctx, prj, partition()); return cctx.affinityNode() ? Collections.singletonList(cctx.localNode()) : Collections.singletonList(F.rand(nodes(cctx, null, partition()))); case PARTITIONED: return nodes(cctx, prj, partition()); default: throw new IllegalStateException("Unknown cache distribution mode: " + cacheMode); } }
/** * Processes cache query response. * * @param sndId Sender node id. * @param res Query response. */ @SuppressWarnings("unchecked") private void processQueryResponse(UUID sndId, GridCacheQueryResponse res) { if (log.isDebugEnabled()) log.debug("Received query response: " + res); GridCacheQueryFutureAdapter fut = getQueryFuture(res.requestId()); if (fut != null) if (res.fields()) ((GridCacheDistributedFieldsQueryFuture) fut) .onPage( sndId, res.metadata(), (Collection<Map<String, Object>>) ((Collection) res.data()), res.error(), res.isFinished()); else fut.onPage(sndId, res.data(), res.error(), res.isFinished()); else if (!cancelled.contains(res.requestId())) U.warn( log, "Received response for finished or unknown query [rmtNodeId=" + sndId + ", res=" + res + ']'); }
/** {@inheritDoc} */ @Override public String toString() { return S.toString( GridDhtLocalPartition.class, this, "state", state(), "reservations", reservations(), "empty", map.isEmpty(), "createTime", U.format(createTime), "mapPubSize", mapPubSize); }
/** * Processes cache query request. * * @param sndId Sender node id. * @param req Query request. */ @SuppressWarnings("unchecked") @Override void processQueryRequest(UUID sndId, GridCacheQueryRequest req) { if (req.cancel()) { cancelIds.add(new CancelMessageId(req.id(), sndId)); if (req.fields()) removeFieldsQueryResult(sndId, req.id()); else removeQueryResult(sndId, req.id()); } else { if (!cancelIds.contains(new CancelMessageId(req.id(), sndId))) { if (!F.eq(req.cacheName(), cctx.name())) { GridCacheQueryResponse res = new GridCacheQueryResponse( cctx.cacheId(), req.id(), new IgniteCheckedException( "Received request for incorrect cache [expected=" + cctx.name() + ", actual=" + req.cacheName())); sendQueryResponse(sndId, res, 0); } else { threads.put(req.id(), Thread.currentThread()); try { GridCacheQueryInfo info = distributedQueryInfo(sndId, req); if (info == null) return; if (req.fields()) runFieldsQuery(info); else runQuery(info); } catch (Throwable e) { U.error(log(), "Failed to run query.", e); sendQueryResponse( sndId, new GridCacheQueryResponse(cctx.cacheId(), req.id(), e.getCause()), 0); if (e instanceof Error) throw (Error) e; } finally { threads.remove(req.id()); } } } } }
/** Clears values for this partition. */ private void clearAll() { GridCacheVersion clearVer = cctx.versions().next(); boolean swap = cctx.isSwapOrOffheapEnabled(); boolean rec = cctx.events().isRecordable(EVT_CACHE_REBALANCE_OBJECT_UNLOADED); Iterator<GridDhtCacheEntry> it = map.values().iterator(); GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> swapIt = null; if (swap && GridQueryProcessor.isEnabled(cctx.config())) { // Indexing needs to unswap cache values. Iterator<GridDhtCacheEntry> unswapIt = null; try { swapIt = cctx.swap().iterator(id); unswapIt = unswapIterator(swapIt); } catch (Exception e) { U.error(log, "Failed to clear swap for evicted partition: " + this, e); } if (unswapIt != null) it = F.concat(it, unswapIt); } try { while (it.hasNext()) { GridDhtCacheEntry cached = it.next(); try { if (cached.clearInternal(clearVer, swap)) { map.remove(cached.key(), cached); if (!cached.isInternal()) { mapPubSize.decrement(); if (rec) cctx.events() .addEvent( cached.partition(), cached.key(), cctx.localNodeId(), (IgniteUuid) null, null, EVT_CACHE_REBALANCE_OBJECT_UNLOADED, null, false, cached.rawGet(), cached.hasValue(), null, null, null); } } } catch (IgniteCheckedException e) { U.error(log, "Failed to clear cache entry for evicted partition: " + cached, e); } } } finally { U.close(swapIt, log); } }
/** Key partition. */ public class GridDhtLocalPartition implements Comparable<GridDhtLocalPartition>, GridReservable { /** Maximum size for delete queue. */ public static final int MAX_DELETE_QUEUE_SIZE = Integer.getInteger(IGNITE_ATOMIC_CACHE_DELETE_HISTORY_SIZE, 200_000); /** Static logger to avoid re-creation. */ private static final AtomicReference<IgniteLogger> logRef = new AtomicReference<>(); /** Logger. */ private static volatile IgniteLogger log; /** Partition ID. */ private final int id; /** State. */ @GridToStringExclude private final AtomicStampedReference<GridDhtPartitionState> state = new AtomicStampedReference<>(MOVING, 0); /** Rent future. */ @GridToStringExclude private final GridFutureAdapter<?> rent; /** Entries map. */ private final ConcurrentMap<KeyCacheObject, GridDhtCacheEntry> map; /** Context. */ private final GridCacheContext cctx; /** Create time. */ @GridToStringExclude private final long createTime = U.currentTimeMillis(); /** Eviction history. */ private volatile Map<KeyCacheObject, GridCacheVersion> evictHist = new HashMap<>(); /** Lock. */ private final ReentrantLock lock = new ReentrantLock(); /** Public size counter. */ private final LongAdder8 mapPubSize = new LongAdder8(); /** Remove queue. */ private final GridCircularBuffer<T2<KeyCacheObject, GridCacheVersion>> rmvQueue; /** Group reservations. */ private final CopyOnWriteArrayList<GridDhtPartitionsReservation> reservations = new CopyOnWriteArrayList<>(); /** * @param cctx Context. * @param id Partition ID. */ @SuppressWarnings("ExternalizableWithoutPublicNoArgConstructor") GridDhtLocalPartition(GridCacheContext cctx, int id) { assert cctx != null; this.id = id; this.cctx = cctx; log = U.logger(cctx.kernalContext(), logRef, this); rent = new GridFutureAdapter<Object>() { @Override public String toString() { return "PartitionRentFuture [part=" + GridDhtLocalPartition.this + ", map=" + map + ']'; } }; map = new ConcurrentHashMap8<>(cctx.config().getStartSize() / cctx.affinity().partitions()); int delQueueSize = CU.isSystemCache(cctx.name()) ? 100 : Math.max(MAX_DELETE_QUEUE_SIZE / cctx.affinity().partitions(), 20); rmvQueue = new GridCircularBuffer<>(U.ceilPow2(delQueueSize)); } /** * Adds group reservation to this partition. * * @param r Reservation. * @return {@code false} If such reservation already added. */ public boolean addReservation(GridDhtPartitionsReservation r) { assert state.getReference() != EVICTED : "we can reserve only active partitions"; assert state.getStamp() != 0 : "partition must be already reserved before adding group reservation"; return reservations.addIfAbsent(r); } /** @param r Reservation. */ public void removeReservation(GridDhtPartitionsReservation r) { if (!reservations.remove(r)) throw new IllegalStateException("Reservation was already removed."); } /** @return Partition ID. */ public int id() { return id; } /** @return Create time. */ long createTime() { return createTime; } /** @return Partition state. */ public GridDhtPartitionState state() { return state.getReference(); } /** @return Reservations. */ public int reservations() { return state.getStamp(); } /** @return Keys belonging to partition. */ public Set<KeyCacheObject> keySet() { return map.keySet(); } /** @return Entries belonging to partition. */ public Collection<GridDhtCacheEntry> entries() { return map.values(); } /** @return {@code True} if partition is empty. */ public boolean isEmpty() { return map.isEmpty(); } /** @return Number of entries in this partition (constant-time method). */ public int size() { return map.size(); } /** Increments public size of the map. */ public void incrementPublicSize() { mapPubSize.increment(); } /** Decrements public size of the map. */ public void decrementPublicSize() { mapPubSize.decrement(); } /** @return Number of public (non-internal) entries in this partition. */ public int publicSize() { return mapPubSize.intValue(); } /** @return If partition is moving or owning or renting. */ public boolean valid() { GridDhtPartitionState state = state(); return state == MOVING || state == OWNING || state == RENTING; } /** @param entry Entry to add. */ void onAdded(GridDhtCacheEntry entry) { GridDhtPartitionState state = state(); if (state == EVICTED) throw new GridDhtInvalidPartitionException( id, "Adding entry to invalid partition [part=" + id + ']'); map.put(entry.key(), entry); if (!entry.isInternal()) mapPubSize.increment(); } /** @param entry Entry to remove. */ @SuppressWarnings("SynchronizationOnLocalVariableOrMethodParameter") void onRemoved(GridDhtCacheEntry entry) { assert entry.obsolete(); // Make sure to remove exactly this entry. synchronized (entry) { map.remove(entry.key(), entry); if (!entry.isInternal() && !entry.deleted()) mapPubSize.decrement(); } // Attempt to evict. tryEvict(true); } /** * @param key Removed key. * @param ver Removed version. * @throws IgniteCheckedException If failed. */ public void onDeferredDelete(KeyCacheObject key, GridCacheVersion ver) throws IgniteCheckedException { try { T2<KeyCacheObject, GridCacheVersion> evicted = rmvQueue.add(new T2<>(key, ver)); if (evicted != null) cctx.dht().removeVersionedEntry(evicted.get1(), evicted.get2()); } catch (InterruptedException e) { Thread.currentThread().interrupt(); throw new IgniteInterruptedCheckedException(e); } } /** Locks partition. */ @SuppressWarnings({"LockAcquiredButNotSafelyReleased"}) public void lock() { lock.lock(); } /** Unlocks partition. */ public void unlock() { lock.unlock(); } /** * @param key Key. * @param ver Version. */ public void onEntryEvicted(KeyCacheObject key, GridCacheVersion ver) { assert key != null; assert ver != null; assert lock.isHeldByCurrentThread(); // Only one thread can enter this method at a time. if (state() != MOVING) return; Map<KeyCacheObject, GridCacheVersion> evictHist0 = evictHist; if (evictHist0 != null) { GridCacheVersion ver0 = evictHist0.get(key); if (ver0 == null || ver0.isLess(ver)) { GridCacheVersion ver1 = evictHist0.put(key, ver); assert ver1 == ver0; } } } /** * Cache preloader should call this method within partition lock. * * @param key Key. * @param ver Version. * @return {@code True} if preloading is permitted. */ public boolean preloadingPermitted(KeyCacheObject key, GridCacheVersion ver) { assert key != null; assert ver != null; assert lock.isHeldByCurrentThread(); // Only one thread can enter this method at a time. if (state() != MOVING) return false; Map<KeyCacheObject, GridCacheVersion> evictHist0 = evictHist; if (evictHist0 != null) { GridCacheVersion ver0 = evictHist0.get(key); // Permit preloading if version in history // is missing or less than passed in. return ver0 == null || ver0.isLess(ver); } return false; } /** * Reserves a partition so it won't be cleared. * * @return {@code True} if reserved. */ @Override public boolean reserve() { while (true) { int reservations = state.getStamp(); GridDhtPartitionState s = state.getReference(); if (s == EVICTED) return false; if (state.compareAndSet(s, s, reservations, reservations + 1)) return true; } } /** Releases previously reserved partition. */ @Override public void release() { while (true) { int reservations = state.getStamp(); if (reservations == 0) return; GridDhtPartitionState s = state.getReference(); assert s != EVICTED; // Decrement reservations. if (state.compareAndSet(s, s, reservations, --reservations)) { tryEvict(true); break; } } } /** @return {@code True} if transitioned to OWNING state. */ boolean own() { while (true) { int reservations = state.getStamp(); GridDhtPartitionState s = state.getReference(); if (s == RENTING || s == EVICTED) return false; if (s == OWNING) return true; assert s == MOVING; if (state.compareAndSet(MOVING, OWNING, reservations, reservations)) { if (log.isDebugEnabled()) log.debug("Owned partition: " + this); // No need to keep history any more. evictHist = null; return true; } } } /** * @param updateSeq Update sequence. * @return Future to signal that this node is no longer an owner or backup. */ IgniteInternalFuture<?> rent(boolean updateSeq) { while (true) { int reservations = state.getStamp(); GridDhtPartitionState s = state.getReference(); if (s == RENTING || s == EVICTED) return rent; if (state.compareAndSet(s, RENTING, reservations, reservations)) { if (log.isDebugEnabled()) log.debug("Moved partition to RENTING state: " + this); // Evict asynchronously, as the 'rent' method may be called // from within write locks on local partition. tryEvictAsync(updateSeq); break; } } return rent; } /** * @param updateSeq Update sequence. * @return Future for evict attempt. */ IgniteInternalFuture<Boolean> tryEvictAsync(boolean updateSeq) { if (map.isEmpty() && !GridQueryProcessor.isEnabled(cctx.config()) && state.compareAndSet(RENTING, EVICTED, 0, 0)) { if (log.isDebugEnabled()) log.debug("Evicted partition: " + this); clearSwap(); if (cctx.isDrEnabled()) cctx.dr().partitionEvicted(id); cctx.dataStructures().onPartitionEvicted(id); rent.onDone(); ((GridDhtPreloader) cctx.preloader()).onPartitionEvicted(this, updateSeq); clearDeferredDeletes(); return new GridFinishedFuture<>(true); } return cctx.closures() .callLocalSafe( new GPC<Boolean>() { @Override public Boolean call() { return tryEvict(true); } }, /*system pool*/ true); } /** @return {@code true} If there is a group reservation. */ private boolean groupReserved() { for (GridDhtPartitionsReservation reservation : reservations) { if (!reservation.invalidate()) return true; // Failed to invalidate reservation -> we are reserved. } return false; } /** * @param updateSeq Update sequence. * @return {@code True} if entry has been transitioned to state EVICTED. */ boolean tryEvict(boolean updateSeq) { if (state.getReference() != RENTING || state.getStamp() != 0 || groupReserved()) return false; // Attempt to evict partition entries from cache. clearAll(); if (map.isEmpty() && state.compareAndSet(RENTING, EVICTED, 0, 0)) { if (log.isDebugEnabled()) log.debug("Evicted partition: " + this); if (!GridQueryProcessor.isEnabled(cctx.config())) clearSwap(); if (cctx.isDrEnabled()) cctx.dr().partitionEvicted(id); cctx.dataStructures().onPartitionEvicted(id); rent.onDone(); ((GridDhtPreloader) cctx.preloader()).onPartitionEvicted(this, updateSeq); clearDeferredDeletes(); return true; } return false; } /** Clears swap entries for evicted partition. */ private void clearSwap() { assert state() == EVICTED; assert !GridQueryProcessor.isEnabled(cctx.config()) : "Indexing needs to have unswapped values."; try { GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> it = cctx.swap().iterator(id); boolean isLocStore = cctx.store().isLocal(); if (it != null) { // We can safely remove these values because no entries will be created for evicted // partition. while (it.hasNext()) { Map.Entry<byte[], GridCacheSwapEntry> entry = it.next(); byte[] keyBytes = entry.getKey(); KeyCacheObject key = cctx.toCacheKeyObject(keyBytes); cctx.swap().remove(key); if (isLocStore) cctx.store().remove(null, key.value(cctx.cacheObjectContext(), false)); } } } catch (IgniteCheckedException e) { U.error(log, "Failed to clear swap for evicted partition: " + this, e); } } /** */ void onUnlock() { tryEvict(true); } /** * @param topVer Topology version. * @return {@code True} if local node is primary for this partition. */ public boolean primary(AffinityTopologyVersion topVer) { return cctx.affinity().primary(cctx.localNode(), id, topVer); } /** Clears values for this partition. */ private void clearAll() { GridCacheVersion clearVer = cctx.versions().next(); boolean swap = cctx.isSwapOrOffheapEnabled(); boolean rec = cctx.events().isRecordable(EVT_CACHE_REBALANCE_OBJECT_UNLOADED); Iterator<GridDhtCacheEntry> it = map.values().iterator(); GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> swapIt = null; if (swap && GridQueryProcessor.isEnabled(cctx.config())) { // Indexing needs to unswap cache values. Iterator<GridDhtCacheEntry> unswapIt = null; try { swapIt = cctx.swap().iterator(id); unswapIt = unswapIterator(swapIt); } catch (Exception e) { U.error(log, "Failed to clear swap for evicted partition: " + this, e); } if (unswapIt != null) it = F.concat(it, unswapIt); } try { while (it.hasNext()) { GridDhtCacheEntry cached = it.next(); try { if (cached.clearInternal(clearVer, swap)) { map.remove(cached.key(), cached); if (!cached.isInternal()) { mapPubSize.decrement(); if (rec) cctx.events() .addEvent( cached.partition(), cached.key(), cctx.localNodeId(), (IgniteUuid) null, null, EVT_CACHE_REBALANCE_OBJECT_UNLOADED, null, false, cached.rawGet(), cached.hasValue(), null, null, null); } } } catch (IgniteCheckedException e) { U.error(log, "Failed to clear cache entry for evicted partition: " + cached, e); } } } finally { U.close(swapIt, log); } } /** * @param it Swap iterator. * @return Unswapping iterator over swapped entries. */ private Iterator<GridDhtCacheEntry> unswapIterator( final GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> it) { if (it == null) return null; return new Iterator<GridDhtCacheEntry>() { /** */ GridDhtCacheEntry lastEntry; @Override public boolean hasNext() { return it.hasNext(); } @Override public GridDhtCacheEntry next() { Map.Entry<byte[], GridCacheSwapEntry> entry = it.next(); byte[] keyBytes = entry.getKey(); try { KeyCacheObject key = cctx.toCacheKeyObject(keyBytes); lastEntry = (GridDhtCacheEntry) cctx.cache().entryEx(key, false); lastEntry.unswap(true); return lastEntry; } catch (IgniteCheckedException e) { throw new CacheException(e); } } @Override public void remove() { map.remove(lastEntry.key(), lastEntry); } }; } /** */ private void clearDeferredDeletes() { rmvQueue.forEach( new CI1<T2<KeyCacheObject, GridCacheVersion>>() { @Override public void apply(T2<KeyCacheObject, GridCacheVersion> t) { cctx.dht().removeVersionedEntry(t.get1(), t.get2()); } }); } /** {@inheritDoc} */ @Override public int hashCode() { return id; } /** {@inheritDoc} */ @SuppressWarnings({"OverlyStrongTypeCast"}) @Override public boolean equals(Object obj) { return obj instanceof GridDhtLocalPartition && (obj == this || ((GridDhtLocalPartition) obj).id() == id); } /** {@inheritDoc} */ @Override public int compareTo(@NotNull GridDhtLocalPartition part) { if (part == null) return 1; return Integer.compare(id, part.id()); } /** {@inheritDoc} */ @Override public String toString() { return S.toString( GridDhtLocalPartition.class, this, "state", state(), "reservations", reservations(), "empty", map.isEmpty(), "createTime", U.format(createTime), "mapPubSize", mapPubSize); } }
/** * Sends cache query response. * * @param nodeId Node to send response. * @param res Cache query response. * @param timeout Message timeout. * @return {@code true} if response was sent, {@code false} otherwise. */ private boolean sendQueryResponse(UUID nodeId, GridCacheQueryResponse res, long timeout) { ClusterNode node = cctx.node(nodeId); if (node == null) return false; int attempt = 1; IgniteCheckedException err = null; while (!Thread.currentThread().isInterrupted()) { try { if (log.isDebugEnabled()) log.debug("Send query response: " + res); Object topic = topic(nodeId, res.requestId()); cctx.io() .sendOrderedMessage( node, topic, res, cctx.ioPolicy(), timeout > 0 ? timeout : Long.MAX_VALUE); return true; } catch (ClusterTopologyCheckedException ignored) { if (log.isDebugEnabled()) log.debug( "Failed to send query response since node left grid [nodeId=" + nodeId + ", res=" + res + "]"); return false; } catch (IgniteCheckedException e) { if (err == null) err = e; if (Thread.currentThread().isInterrupted()) break; if (attempt < RESEND_ATTEMPTS) { if (log.isDebugEnabled()) log.debug( "Failed to send queries response (will try again) [nodeId=" + nodeId + ", res=" + res + ", attempt=" + attempt + ", err=" + e + "]"); if (!Thread.currentThread().isInterrupted()) try { U.sleep(RESEND_FREQ); } catch (IgniteInterruptedCheckedException e1) { U.error( log, "Waiting for queries response resending was interrupted (response will not be sent) " + "[nodeId=" + nodeId + ", response=" + res + "]", e1); return false; } } else { U.error( log, "Failed to sender cache response [nodeId=" + nodeId + ", response=" + res + "]", err); return false; } } attempt++; } return false; }