state_store.go

package state

import (
	"fmt"
	"io"
	"log"
	"sync"

	"github.com/hashicorp/go-memdb"
	"github.com/hashicorp/nomad/nomad/structs"
	"github.com/hashicorp/nomad/nomad/watch"
)

// IndexEntry is used with the "index" table
// for managing the latest Raft index affecting a table.
type IndexEntry struct {
	Key   string
	Value uint64
}

// The StateStore is responsible for maintaining all the Nomad
// state. It is manipulated by the FSM which maintains consistency
// through the use of Raft. The goals of the StateStore are to provide
// high concurrency for read operations without blocking writes, and
// to provide write availability in the face of reads. EVERY object
// returned as a result of a read against the state store should be
// considered a constant and NEVER modified in place.
type StateStore struct {
	logger *log.Logger
	db     *memdb.MemDB
	watch  *stateWatch
}

// NewStateStore is used to create a new state store
func NewStateStore(logOutput io.Writer) (*StateStore, error) {
	// Create the MemDB
	db, err := memdb.NewMemDB(stateStoreSchema())
	if err != nil {
		return nil, fmt.Errorf("state store setup failed: %v", err)
	}

	// Create the state store
	s := &StateStore{
		logger: log.New(logOutput, "", log.LstdFlags),
		db:     db,
		watch:  newStateWatch(),
	}
	return s, nil
}

// Snapshot is used to create a point in time snapshot. Because
// we use MemDB, we just need to snapshot the state of the underlying
// database.
func (s *StateStore) Snapshot() (*StateSnapshot, error) {
	snap := &StateSnapshot{
		StateStore: StateStore{
			logger: s.logger,
			db:     s.db.Snapshot(),
			watch:  s.watch,
		},
	}
	return snap, nil
}

// Restore is used to optimize the efficiency of rebuilding
// state by minimizing the number of transactions and checking
// overhead.
func (s *StateStore) Restore() (*StateRestore, error) {
	txn := s.db.Txn(true)
	r := &StateRestore{
		txn:   txn,
		watch: s.watch,
		items: watch.NewItems(),
	}
	return r, nil
}

// Watch subscribes a channel to a set of watch items.
func (s *StateStore) Watch(items watch.Items, notify chan struct{}) {
	s.watch.watch(items, notify)
}

// StopWatch unsubscribes a channel from a set of watch items.
func (s *StateStore) StopWatch(items watch.Items, notify chan struct{}) {
	s.watch.stopWatch(items, notify)
}

// UpsertNode is used to register a node or update a node definition
// This is assumed to be triggered by the client, so we retain the value
// of drain which is set by the scheduler.
func (s *StateStore) UpsertNode(index uint64, node *structs.Node) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "nodes"})
	watcher.Add(watch.Item{Node: node.ID})

	// Check if the node already exists
	existing, err := txn.First("nodes", "id", node.ID)
	if err != nil {
		return fmt.Errorf("node lookup failed: %v", err)
	}

	// Setup the indexes correctly
	if existing != nil {
		exist := existing.(*structs.Node)
		node.CreateIndex = exist.CreateIndex
		node.ModifyIndex = index
		node.Drain = exist.Drain // Retain the drain mode
	} else {
		node.CreateIndex = index
		node.ModifyIndex = index
	}

	// Insert the node
	if err := txn.Insert("nodes", node); err != nil {
		return fmt.Errorf("node insert failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"nodes", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// DeleteNode is used to deregister a node
func (s *StateStore) DeleteNode(index uint64, nodeID string) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "nodes"})
	watcher.Add(watch.Item{Node: nodeID})

	// Lookup the node
	existing, err := txn.First("nodes", "id", nodeID)
	if err != nil {
		return fmt.Errorf("node lookup failed: %v", err)
	}
	if existing == nil {
		return fmt.Errorf("node not found")
	}

	// Delete the node
	if err := txn.Delete("nodes", existing); err != nil {
		return fmt.Errorf("node delete failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"nodes", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// UpdateNodeStatus is used to update the status of a node
func (s *StateStore) UpdateNodeStatus(index uint64, nodeID, status string) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "nodes"})
	watcher.Add(watch.Item{Node: nodeID})

	// Lookup the node
	existing, err := txn.First("nodes", "id", nodeID)
	if err != nil {
		return fmt.Errorf("node lookup failed: %v", err)
	}
	if existing == nil {
		return fmt.Errorf("node not found")
	}

	// Copy the existing node
	existingNode := existing.(*structs.Node)
	copyNode := new(structs.Node)
	*copyNode = *existingNode

	// Update the status in the copy
	copyNode.Status = status
	copyNode.ModifyIndex = index

	// Insert the node
	if err := txn.Insert("nodes", copyNode); err != nil {
		return fmt.Errorf("node update failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"nodes", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// UpdateNodeDrain is used to update the drain of a node
func (s *StateStore) UpdateNodeDrain(index uint64, nodeID string, drain bool) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "nodes"})
	watcher.Add(watch.Item{Node: nodeID})

	// Lookup the node
	existing, err := txn.First("nodes", "id", nodeID)
	if err != nil {
		return fmt.Errorf("node lookup failed: %v", err)
	}
	if existing == nil {
		return fmt.Errorf("node not found")
	}

	// Copy the existing node
	existingNode := existing.(*structs.Node)
	copyNode := new(structs.Node)
	*copyNode = *existingNode

	// Update the drain in the copy
	copyNode.Drain = drain
	copyNode.ModifyIndex = index

	// Insert the node
	if err := txn.Insert("nodes", copyNode); err != nil {
		return fmt.Errorf("node update failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"nodes", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// NodeByID is used to lookup a node by ID
func (s *StateStore) NodeByID(nodeID string) (*structs.Node, error) {
	txn := s.db.Txn(false)

	existing, err := txn.First("nodes", "id", nodeID)
	if err != nil {
		return nil, fmt.Errorf("node lookup failed: %v", err)
	}

	if existing != nil {
		return existing.(*structs.Node), nil
	}
	return nil, nil
}

// Nodes returns an iterator over all the nodes
func (s *StateStore) Nodes() (memdb.ResultIterator, error) {
	txn := s.db.Txn(false)

	// Walk the entire nodes table
	iter, err := txn.Get("nodes", "id")
	if err != nil {
		return nil, err
	}
	return iter, nil
}

// UpsertJob is used to register a job or update a job definition
func (s *StateStore) UpsertJob(index uint64, job *structs.Job) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "jobs"})
	watcher.Add(watch.Item{Job: job.ID})

	// Check if the job already exists
	existing, err := txn.First("jobs", "id", job.ID)
	if err != nil {
		return fmt.Errorf("job lookup failed: %v", err)
	}

	// Setup the indexes correctly
	if existing != nil {
		job.CreateIndex = existing.(*structs.Job).CreateIndex
		job.ModifyIndex = index
	} else {
		job.CreateIndex = index
		job.ModifyIndex = index
	}

	// Insert the job
	if err := txn.Insert("jobs", job); err != nil {
		return fmt.Errorf("job insert failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"jobs", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// DeleteJob is used to deregister a job
func (s *StateStore) DeleteJob(index uint64, jobID string) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "jobs"})
	watcher.Add(watch.Item{Job: jobID})

	// Lookup the node
	existing, err := txn.First("jobs", "id", jobID)
	if err != nil {
		return fmt.Errorf("job lookup failed: %v", err)
	}
	if existing == nil {
		return fmt.Errorf("job not found")
	}

	// Delete the node
	if err := txn.Delete("jobs", existing); err != nil {
		return fmt.Errorf("job delete failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"jobs", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// JobByID is used to lookup a job by its ID
func (s *StateStore) JobByID(id string) (*structs.Job, error) {
	txn := s.db.Txn(false)

	existing, err := txn.First("jobs", "id", id)
	if err != nil {
		return nil, fmt.Errorf("job lookup failed: %v", err)
	}

	if existing != nil {
		return existing.(*structs.Job), nil
	}
	return nil, nil
}

// Jobs returns an iterator over all the jobs
func (s *StateStore) Jobs() (memdb.ResultIterator, error) {
	txn := s.db.Txn(false)

	// Walk the entire jobs table
	iter, err := txn.Get("jobs", "id")
	if err != nil {
		return nil, err
	}
	return iter, nil
}

// JobsByScheduler returns an iterator over all the jobs with the specific
// scheduler type.
func (s *StateStore) JobsByScheduler(schedulerType string) (memdb.ResultIterator, error) {
	txn := s.db.Txn(false)

	// Return an iterator for jobs with the specific type.
	iter, err := txn.Get("jobs", "type", schedulerType)
	if err != nil {
		return nil, err
	}
	return iter, nil
}

// UpsertEvaluation is used to upsert an evaluation
func (s *StateStore) UpsertEvals(index uint64, evals []*structs.Evaluation) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "evals"})

	// Do a nested upsert
	for _, eval := range evals {
		watcher.Add(watch.Item{Eval: eval.ID})
		if err := s.nestedUpsertEval(txn, index, eval); err != nil {
			return err
		}
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// nestedUpsertEvaluation is used to nest an evaluation upsert within a transaction
func (s *StateStore) nestedUpsertEval(txn *memdb.Txn, index uint64, eval *structs.Evaluation) error {
	// Lookup the evaluation
	existing, err := txn.First("evals", "id", eval.ID)
	if err != nil {
		return fmt.Errorf("eval lookup failed: %v", err)
	}

	// Update the indexes
	if existing != nil {
		eval.CreateIndex = existing.(*structs.Evaluation).CreateIndex
		eval.ModifyIndex = index
	} else {
		eval.CreateIndex = index
		eval.ModifyIndex = index
	}

	// Insert the eval
	if err := txn.Insert("evals", eval); err != nil {
		return fmt.Errorf("eval insert failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"evals", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}
	return nil
}

// DeleteEval is used to delete an evaluation
func (s *StateStore) DeleteEval(index uint64, evals []string, allocs []string) error {
	txn := s.db.Txn(true)
	defer txn.Abort()
	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "evals"})
	watcher.Add(watch.Item{Table: "allocs"})

	for _, eval := range evals {
		existing, err := txn.First("evals", "id", eval)
		if err != nil {
			return fmt.Errorf("eval lookup failed: %v", err)
		}
		if existing == nil {
			continue
		}
		if err := txn.Delete("evals", existing); err != nil {
			return fmt.Errorf("eval delete failed: %v", err)
		}
		watcher.Add(watch.Item{Eval: eval})
	}

	for _, alloc := range allocs {
		existing, err := txn.First("allocs", "id", alloc)
		if err != nil {
			return fmt.Errorf("alloc lookup failed: %v", err)
		}
		if existing == nil {
			continue
		}
		if err := txn.Delete("allocs", existing); err != nil {
			return fmt.Errorf("alloc delete failed: %v", err)
		}
		realAlloc := existing.(*structs.Allocation)
		watcher.Add(watch.Item{Alloc: realAlloc.ID})
		watcher.Add(watch.Item{AllocEval: realAlloc.EvalID})
		watcher.Add(watch.Item{AllocJob: realAlloc.JobID})
		watcher.Add(watch.Item{AllocNode: realAlloc.NodeID})
	}

	// Update the indexes
	if err := txn.Insert("index", &IndexEntry{"evals", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}
	if err := txn.Insert("index", &IndexEntry{"allocs", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// EvalByID is used to lookup an eval by its ID
func (s *StateStore) EvalByID(id string) (*structs.Evaluation, error) {
	txn := s.db.Txn(false)

	existing, err := txn.First("evals", "id", id)
	if err != nil {
		return nil, fmt.Errorf("eval lookup failed: %v", err)
	}

	if existing != nil {
		return existing.(*structs.Evaluation), nil
	}
	return nil, nil
}

// EvalsByJob returns all the evaluations by job id
func (s *StateStore) EvalsByJob(jobID string) ([]*structs.Evaluation, error) {
	txn := s.db.Txn(false)

	// Get an iterator over the node allocations
	iter, err := txn.Get("evals", "job", jobID)
	if err != nil {
		return nil, err
	}

	var out []*structs.Evaluation
	for {
		raw := iter.Next()
		if raw == nil {
			break
		}
		out = append(out, raw.(*structs.Evaluation))
	}
	return out, nil
}

// Evals returns an iterator over all the evaluations
func (s *StateStore) Evals() (memdb.ResultIterator, error) {
	txn := s.db.Txn(false)

	// Walk the entire table
	iter, err := txn.Get("evals", "id")
	if err != nil {
		return nil, err
	}
	return iter, nil
}

// UpdateAllocFromClient is used to update an allocation based on input
// from a client. While the schedulers are the authority on the allocation for
// most things, some updates are authoritative from the client. Specifically,
// the desired state comes from the schedulers, while the actual state comes
// from clients.
func (s *StateStore) UpdateAllocFromClient(index uint64, alloc *structs.Allocation) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "allocs"})
	watcher.Add(watch.Item{Alloc: alloc.ID})
	watcher.Add(watch.Item{AllocEval: alloc.EvalID})
	watcher.Add(watch.Item{AllocJob: alloc.JobID})
	watcher.Add(watch.Item{AllocNode: alloc.NodeID})

	// Look for existing alloc
	existing, err := txn.First("allocs", "id", alloc.ID)
	if err != nil {
		return fmt.Errorf("alloc lookup failed: %v", err)
	}

	// Nothing to do if this does not exist
	if existing == nil {
		return nil
	}
	exist := existing.(*structs.Allocation)

	// Copy everything from the existing allocation
	copyAlloc := new(structs.Allocation)
	*copyAlloc = *exist

	// Pull in anything the client is the authority on
	copyAlloc.ClientStatus = alloc.ClientStatus
	copyAlloc.ClientDescription = alloc.ClientDescription

	// Update the modify index
	copyAlloc.ModifyIndex = index

	// Update the allocation
	if err := txn.Insert("allocs", copyAlloc); err != nil {
		return fmt.Errorf("alloc insert failed: %v", err)
	}

	// Update the indexes
	if err := txn.Insert("index", &IndexEntry{"allocs", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// UpsertAllocs is used to evict a set of allocations
// and allocate new ones at the same time.
func (s *StateStore) UpsertAllocs(index uint64, allocs []*structs.Allocation) error {
	txn := s.db.Txn(true)
	defer txn.Abort()

	watcher := watch.NewItems()
	watcher.Add(watch.Item{Table: "allocs"})

	// Handle the allocations
	for _, alloc := range allocs {
		existing, err := txn.First("allocs", "id", alloc.ID)
		if err != nil {
			return fmt.Errorf("alloc lookup failed: %v", err)
		}

		if existing == nil {
			alloc.CreateIndex = index
			alloc.ModifyIndex = index
		} else {
			exist := existing.(*structs.Allocation)
			alloc.CreateIndex = exist.CreateIndex
			alloc.ModifyIndex = index
			alloc.ClientStatus = exist.ClientStatus
			alloc.ClientDescription = exist.ClientDescription
		}
		if err := txn.Insert("allocs", alloc); err != nil {
			return fmt.Errorf("alloc insert failed: %v", err)
		}

		watcher.Add(watch.Item{Alloc: alloc.ID})
		watcher.Add(watch.Item{AllocEval: alloc.EvalID})
		watcher.Add(watch.Item{AllocJob: alloc.JobID})
		watcher.Add(watch.Item{AllocNode: alloc.NodeID})
	}

	// Update the indexes
	if err := txn.Insert("index", &IndexEntry{"allocs", index}); err != nil {
		return fmt.Errorf("index update failed: %v", err)
	}

	txn.Defer(func() { s.watch.notify(watcher) })
	txn.Commit()
	return nil
}

// AllocByID is used to lookup an allocation by its ID
func (s *StateStore) AllocByID(id string) (*structs.Allocation, error) {
	txn := s.db.Txn(false)

	existing, err := txn.First("allocs", "id", id)
	if err != nil {
		return nil, fmt.Errorf("alloc lookup failed: %v", err)
	}

	if existing != nil {
		return existing.(*structs.Allocation), nil
	}
	return nil, nil
}

// AllocsByNode returns all the allocations by node
func (s *StateStore) AllocsByNode(node string) ([]*structs.Allocation, error) {
	txn := s.db.Txn(false)

	// Get an iterator over the node allocations
	iter, err := txn.Get("allocs", "node", node)
	if err != nil {
		return nil, err
	}

	var out []*structs.Allocation
	for {
		raw := iter.Next()
		if raw == nil {
			break
		}
		out = append(out, raw.(*structs.Allocation))
	}
	return out, nil
}

// AllocsByJob returns all the allocations by job id
func (s *StateStore) AllocsByJob(jobID string) ([]*structs.Allocation, error) {
	txn := s.db.Txn(false)

	// Get an iterator over the node allocations
	iter, err := txn.Get("allocs", "job", jobID)
	if err != nil {
		return nil, err
	}

	var out []*structs.Allocation
	for {
		raw := iter.Next()
		if raw == nil {
			break
		}
		out = append(out, raw.(*structs.Allocation))
	}
	return out, nil
}

// AllocsByEval returns all the allocations by eval id
func (s *StateStore) AllocsByEval(evalID string) ([]*structs.Allocation, error) {
	txn := s.db.Txn(false)

	// Get an iterator over the eval allocations
	iter, err := txn.Get("allocs", "eval", evalID)
	if err != nil {
		return nil, err
	}

	var out []*structs.Allocation
	for {
		raw := iter.Next()
		if raw == nil {
			break
		}
		out = append(out, raw.(*structs.Allocation))
	}
	return out, nil
}

// Allocs returns an iterator over all the evaluations
func (s *StateStore) Allocs() (memdb.ResultIterator, error) {
	txn := s.db.Txn(false)

	// Walk the entire table
	iter, err := txn.Get("allocs", "id")
	if err != nil {
		return nil, err
	}
	return iter, nil
}

// Index finds the matching index value
func (s *StateStore) Index(name string) (uint64, error) {
	txn := s.db.Txn(false)

	// Lookup the first matching index
	out, err := txn.First("index", "id", name)
	if err != nil {
		return 0, err
	}
	if out == nil {
		return 0, nil
	}
	return out.(*IndexEntry).Value, nil
}

// Indexes returns an iterator over all the indexes
func (s *StateStore) Indexes() (memdb.ResultIterator, error) {
	txn := s.db.Txn(false)

	// Walk the entire nodes table
	iter, err := txn.Get("index", "id")
	if err != nil {
		return nil, err
	}
	return iter, nil
}

// StateSnapshot is used to provide a point-in-time snapshot
type StateSnapshot struct {
	StateStore
}

// StateRestore is used to optimize the performance when
// restoring state by only using a single large transaction
// instead of thousands of sub transactions
type StateRestore struct {
	txn   *memdb.Txn
	watch *stateWatch
	items watch.Items
}

// Abort is used to abort the restore operation
func (s *StateRestore) Abort() {
	s.txn.Abort()
}

// Commit is used to commit the restore operation
func (s *StateRestore) Commit() {
	s.txn.Defer(func() { s.watch.notify(s.items) })
	s.txn.Commit()
}

// NodeRestore is used to restore a node
func (r *StateRestore) NodeRestore(node *structs.Node) error {
	r.items.Add(watch.Item{Table: "nodes"})
	r.items.Add(watch.Item{Node: node.ID})
	if err := r.txn.Insert("nodes", node); err != nil {
		return fmt.Errorf("node insert failed: %v", err)
	}
	return nil
}

// JobRestore is used to restore a job
func (r *StateRestore) JobRestore(job *structs.Job) error {
	r.items.Add(watch.Item{Table: "jobs"})
	r.items.Add(watch.Item{Job: job.ID})
	if err := r.txn.Insert("jobs", job); err != nil {
		return fmt.Errorf("job insert failed: %v", err)
	}
	return nil
}

// EvalRestore is used to restore an evaluation
func (r *StateRestore) EvalRestore(eval *structs.Evaluation) error {
	r.items.Add(watch.Item{Table: "evals"})
	r.items.Add(watch.Item{Eval: eval.ID})
	if err := r.txn.Insert("evals", eval); err != nil {
		return fmt.Errorf("eval insert failed: %v", err)
	}
	return nil
}

// AllocRestore is used to restore an allocation
func (r *StateRestore) AllocRestore(alloc *structs.Allocation) error {
	r.items.Add(watch.Item{Table: "allocs"})
	r.items.Add(watch.Item{Alloc: alloc.ID})
	r.items.Add(watch.Item{AllocEval: alloc.EvalID})
	r.items.Add(watch.Item{AllocJob: alloc.JobID})
	r.items.Add(watch.Item{AllocNode: alloc.NodeID})
	if err := r.txn.Insert("allocs", alloc); err != nil {
		return fmt.Errorf("alloc insert failed: %v", err)
	}
	return nil
}

// IndexRestore is used to restore an index
func (r *StateRestore) IndexRestore(idx *IndexEntry) error {
	if err := r.txn.Insert("index", idx); err != nil {
		return fmt.Errorf("index insert failed: %v", err)
	}
	return nil
}

// stateWatch holds shared state for watching updates. This is
// outside of StateStore so it can be shared with snapshots.
type stateWatch struct {
	items map[watch.Item]*NotifyGroup
	l     sync.Mutex
}

// newStateWatch creates a new stateWatch for change notification.
func newStateWatch() *stateWatch {
	return &stateWatch{
		items: make(map[watch.Item]*NotifyGroup),
	}
}

// watch subscribes a channel to the given watch items.
func (w *stateWatch) watch(items watch.Items, ch chan struct{}) {
	w.l.Lock()
	defer w.l.Unlock()

	for item, _ := range items {
		grp, ok := w.items[item]
		if !ok {
			grp = new(NotifyGroup)
			w.items[item] = grp
		}
		grp.Wait(ch)
	}
}

// stopWatch unsubscribes a channel from the given watch items.
func (w *stateWatch) stopWatch(items watch.Items, ch chan struct{}) {
	w.l.Lock()
	defer w.l.Unlock()

	for item, _ := range items {
		if grp, ok := w.items[item]; ok {
			grp.Clear(ch)
			if grp.Empty() {
				delete(w.items, item)
			}
		}
	}
}

// notify is used to fire notifications on the given watch items.
func (w *stateWatch) notify(items watch.Items) {
	w.l.Lock()
	defer w.l.Unlock()

	for wi, _ := range items {
		if grp, ok := w.items[wi]; ok {
			grp.Notify()
		}
	}
}