storage: hold batchTx lock during KV txn

One txn is treated as atomic, and might contain multiple Put/Delete/Range
operations. For now, between these operations, we might call forecCommit
to sync the change to disk, or backend may commit it in background.
Thus the snapshot state might contains an unfinished multiple objects
transaction, which is dangerous if database is restored from the snapshot.

This PR makes KV txn hold batchTx lock during the process and avoids
commit to happen.
This commit is contained in:
Yicheng Qin 2015-09-29 23:26:01 -07:00
parent 533e728b64
commit c97dda766e
2 changed files with 36 additions and 13 deletions

View File

@ -52,6 +52,7 @@ type store struct {
// the main revision of the last compaction // the main revision of the last compaction
compactMainRev int64 compactMainRev int64
tx backend.BatchTx
tmu sync.Mutex // protect the txnID field tmu sync.Mutex // protect the txnID field
txnID int64 // tracks the current txnID to verify txn operations txnID int64 // tracks the current txnID to verify txn operations
@ -122,6 +123,8 @@ func (s *store) DeleteRange(key, end []byte) (n, rev int64) {
func (s *store) TxnBegin() int64 { func (s *store) TxnBegin() int64 {
s.mu.Lock() s.mu.Lock()
s.currentRev.sub = 0 s.currentRev.sub = 0
s.tx = s.b.BatchTx()
s.tx.Lock()
s.tmu.Lock() s.tmu.Lock()
defer s.tmu.Unlock() defer s.tmu.Unlock()
@ -148,6 +151,7 @@ func (s *store) txnEnd(txnID int64) error {
return ErrTxnIDMismatch return ErrTxnIDMismatch
} }
s.tx.Unlock()
if s.currentRev.sub != 0 { if s.currentRev.sub != 0 {
s.currentRev.main += 1 s.currentRev.main += 1
} }
@ -392,14 +396,11 @@ func (s *store) rangeKeys(key, end []byte, limit, rangeRev int64) (kvs []storage
return nil, rev, nil return nil, rev, nil
} }
tx := s.b.BatchTx()
tx.Lock()
defer tx.Unlock()
for _, revpair := range revpairs { for _, revpair := range revpairs {
revbytes := newRevBytes() revbytes := newRevBytes()
revToBytes(revpair, revbytes) revToBytes(revpair, revbytes)
_, vs := tx.UnsafeRange(keyBucketName, revbytes, nil, 0) _, vs := s.tx.UnsafeRange(keyBucketName, revbytes, nil, 0)
if len(vs) != 1 { if len(vs) != 1 {
log.Fatalf("storage: range cannot find rev (%d,%d)", revpair.main, revpair.sub) log.Fatalf("storage: range cannot find rev (%d,%d)", revpair.main, revpair.sub)
} }
@ -446,10 +447,7 @@ func (s *store) put(key, value []byte) {
log.Fatalf("storage: cannot marshal event: %v", err) log.Fatalf("storage: cannot marshal event: %v", err)
} }
tx := s.b.BatchTx() s.tx.UnsafePut(keyBucketName, ibytes, d)
tx.Lock()
defer tx.Unlock()
tx.UnsafePut(keyBucketName, ibytes, d)
s.kvindex.Put(key, revision{main: rev, sub: s.currentRev.sub}) s.kvindex.Put(key, revision{main: rev, sub: s.currentRev.sub})
s.currentRev.sub += 1 s.currentRev.sub += 1
} }
@ -474,10 +472,6 @@ func (s *store) deleteRange(key, end []byte) int64 {
func (s *store) delete(key []byte) { func (s *store) delete(key []byte) {
mainrev := s.currentRev.main + 1 mainrev := s.currentRev.main + 1
tx := s.b.BatchTx()
tx.Lock()
defer tx.Unlock()
ibytes := newRevBytes() ibytes := newRevBytes()
revToBytes(revision{main: mainrev, sub: s.currentRev.sub}, ibytes) revToBytes(revision{main: mainrev, sub: s.currentRev.sub}, ibytes)
@ -493,7 +487,7 @@ func (s *store) delete(key []byte) {
log.Fatalf("storage: cannot marshal event: %v", err) log.Fatalf("storage: cannot marshal event: %v", err)
} }
tx.UnsafePut(keyBucketName, ibytes, d) s.tx.UnsafePut(keyBucketName, ibytes, d)
err = s.kvindex.Tombstone(key, revision{main: mainrev, sub: s.currentRev.sub}) err = s.kvindex.Tombstone(key, revision{main: mainrev, sub: s.currentRev.sub})
if err != nil { if err != nil {
log.Fatalf("storage: cannot tombstone an existing key (%s): %v", string(key), err) log.Fatalf("storage: cannot tombstone an existing key (%s): %v", string(key), err)

View File

@ -103,6 +103,7 @@ func TestStorePut(t *testing.T) {
for i, tt := range tests { for i, tt := range tests {
s, b, index := newFakeStore() s, b, index := newFakeStore()
s.currentRev = tt.rev s.currentRev = tt.rev
s.tx = b.BatchTx()
index.indexGetRespc <- tt.r index.indexGetRespc <- tt.r
s.put([]byte("foo"), []byte("bar")) s.put([]byte("foo"), []byte("bar"))
@ -164,6 +165,7 @@ func TestStoreRange(t *testing.T) {
for i, tt := range tests { for i, tt := range tests {
s, b, index := newFakeStore() s, b, index := newFakeStore()
s.currentRev = currev s.currentRev = currev
s.tx = b.BatchTx()
b.tx.rangeRespc <- tt.r b.tx.rangeRespc <- tt.r
index.indexRangeRespc <- tt.idxr index.indexRangeRespc <- tt.idxr
@ -223,6 +225,7 @@ func TestStoreDeleteRange(t *testing.T) {
for i, tt := range tests { for i, tt := range tests {
s, b, index := newFakeStore() s, b, index := newFakeStore()
s.currentRev = tt.rev s.currentRev = tt.rev
s.tx = b.BatchTx()
index.indexRangeRespc <- tt.r index.indexRangeRespc <- tt.r
n := s.deleteRange([]byte("foo"), []byte("goo")) n := s.deleteRange([]byte("foo"), []byte("goo"))
@ -651,6 +654,32 @@ func TestRestoreContinueUnfinishedCompaction(t *testing.T) {
tx.Unlock() tx.Unlock()
} }
func TestTxnBlockBackendForceCommit(t *testing.T) {
s := newStore(tmpPath)
defer os.Remove(tmpPath)
id := s.TxnBegin()
done := make(chan struct{})
go func() {
s.b.ForceCommit()
done <- struct{}{}
}()
select {
case <-done:
t.Fatalf("failed to block ForceCommit")
case <-time.After(100 * time.Millisecond):
}
s.TxnEnd(id)
select {
case <-done:
case <-time.After(100 * time.Millisecond):
t.Fatalf("failed to execute ForceCommit")
}
}
func BenchmarkStorePut(b *testing.B) { func BenchmarkStorePut(b *testing.B) {
s := newStore(tmpPath) s := newStore(tmpPath)
defer os.Remove(tmpPath) defer os.Remove(tmpPath)