@ -78,6 +78,8 @@ void WriteUnpreparedTxn::Initialize(const TransactionOptions& txn_options) {
} }
unprep_seqs_ . clear ( ) ; unprep_seqs_ . clear ( ) ;
flushed_save_points_ . reset ( nullptr ) ;
unflushed_save_points_ . reset ( nullptr ) ;
recovered_txn_ = false ; recovered_txn_ = false ;
largest_validated_seq_ = 0 ; largest_validated_seq_ = 0 ;
} }
@ -236,6 +238,20 @@ Status WriteUnpreparedTxn::MaybeFlushWriteBatchToDB() {
} }
Status WriteUnpreparedTxn : : FlushWriteBatchToDB ( bool prepared ) { Status WriteUnpreparedTxn : : FlushWriteBatchToDB ( bool prepared ) {
// If the current write batch contains savepoints, then some special handling
// is required so that RollbackToSavepoint can work.
//
// RollbackToSavepoint is not supported after Prepare() is called, so only do
// this for unprepared batches.
if ( ! prepared & & unflushed_save_points_ ! = nullptr & &
! unflushed_save_points_ - > empty ( ) ) {
return FlushWriteBatchWithSavePointToDB ( ) ;
}
return FlushWriteBatchToDBInternal ( prepared ) ;
}
Status WriteUnpreparedTxn : : FlushWriteBatchToDBInternal ( bool prepared ) {
if ( name_ . empty ( ) ) { if ( name_ . empty ( ) ) {
return Status : : InvalidArgument ( " Cannot write to DB without SetName. " ) ; return Status : : InvalidArgument ( " Cannot write to DB without SetName. " ) ;
} }
@ -285,13 +301,118 @@ Status WriteUnpreparedTxn::FlushWriteBatchToDB(bool prepared) {
// Reset transaction state.
// Reset transaction state.
if ( ! prepared ) { if ( ! prepared ) {
prepare_batch_cnt_ = 0 ; prepare_batch_cnt_ = 0 ;
write_batch_ . Clear ( ) ; const bool kClear = true ;
WriteBatchInterna l: : InsertNoop ( wr ite_batch_ . Get WriteBatch ( ) ) ; TransactionBaseImp l: : InitWriteBatch ( kClear ) ;
} }
return s ; return s ;
} }
Status WriteUnpreparedTxn : : FlushWriteBatchWithSavePointToDB ( ) {
assert ( unflushed_save_points_ ! = nullptr & &
unflushed_save_points_ - > size ( ) > 0 ) ;
assert ( save_points_ ! = nullptr & & save_points_ - > size ( ) > 0 ) ;
assert ( save_points_ - > size ( ) > = unflushed_save_points_ - > size ( ) ) ;
// Handler class for creating an unprepared batch from a savepoint.
struct SavePointBatchHandler : public WriteBatch : : Handler {
WriteBatchWithIndex * wb_ ;
const std : : map < uint32_t , ColumnFamilyHandle * > & handles_ ;
SavePointBatchHandler (
WriteBatchWithIndex * wb ,
const std : : map < uint32_t , ColumnFamilyHandle * > & handles )
: wb_ ( wb ) , handles_ ( handles ) { }
Status PutCF ( uint32_t cf , const Slice & key , const Slice & value ) override {
return wb_ - > Put ( handles_ . at ( cf ) , key , value ) ;
}
Status DeleteCF ( uint32_t cf , const Slice & key ) override {
return wb_ - > Delete ( handles_ . at ( cf ) , key ) ;
}
Status SingleDeleteCF ( uint32_t cf , const Slice & key ) override {
return wb_ - > SingleDelete ( handles_ . at ( cf ) , key ) ;
}
Status MergeCF ( uint32_t cf , const Slice & key , const Slice & value ) override {
return wb_ - > Merge ( handles_ . at ( cf ) , key , value ) ;
}
// The only expected 2PC marker is the initial Noop marker.
Status MarkNoop ( bool empty_batch ) override {
return empty_batch ? Status : : OK ( ) : Status : : InvalidArgument ( ) ;
}
Status MarkBeginPrepare ( bool ) override { return Status : : InvalidArgument ( ) ; }
Status MarkEndPrepare ( const Slice & ) override {
return Status : : InvalidArgument ( ) ;
}
Status MarkCommit ( const Slice & ) override {
return Status : : InvalidArgument ( ) ;
}
Status MarkRollback ( const Slice & ) override {
return Status : : InvalidArgument ( ) ;
}
} ;
// The comparator of the default cf is passed in, similar to the
// initialization of TransactionBaseImpl::write_batch_. This comparator is
// only used if the write batch encounters an invalid cf id, and falls back to
// this comparator.
WriteBatchWithIndex wb ( wpt_db_ - > DefaultColumnFamily ( ) - > GetComparator ( ) , 0 ,
true , 0 ) ;
// Swap with write_batch_ so that wb contains the complete write batch. The
// actual write batch that will be flushed to DB will be built in
// write_batch_, and will be read by FlushWriteBatchToDBInternal.
std : : swap ( wb , write_batch_ ) ;
TransactionBaseImpl : : InitWriteBatch ( ) ;
size_t prev_boundary = WriteBatchInternal : : kHeader ;
const bool kPrepared = true ;
for ( size_t i = 0 ; i < unflushed_save_points_ - > size ( ) ; i + + ) {
SavePointBatchHandler sp_handler ( & write_batch_ ,
* wupt_db_ - > GetCFHandleMap ( ) . get ( ) ) ;
size_t curr_boundary = ( * unflushed_save_points_ ) [ i ] ;
// Construct the partial write batch up to the savepoint.
//
// Theoretically, a memcpy between the write batches should be sufficient
// since the rewriting into the batch should produce the exact same byte
// representation. Rebuilding the WriteBatchWithIndex index is still
// necessary though, and would imply doing two passes over the batch though.
Status s = WriteBatchInternal : : Iterate ( wb . GetWriteBatch ( ) , & sp_handler ,
prev_boundary , curr_boundary ) ;
if ( ! s . ok ( ) ) {
return s ;
}
// Flush the write batch.
s = FlushWriteBatchToDBInternal ( ! kPrepared ) ;
if ( ! s . ok ( ) ) {
return s ;
}
if ( flushed_save_points_ = = nullptr ) {
flushed_save_points_ . reset (
new autovector < WriteUnpreparedTxn : : SavePoint > ( ) ) ;
}
flushed_save_points_ - > emplace_back (
unprep_seqs_ , new ManagedSnapshot ( db_impl_ , wupt_db_ - > GetSnapshot ( ) ) ) ;
prev_boundary = curr_boundary ;
const bool kClear = true ;
TransactionBaseImpl : : InitWriteBatch ( kClear ) ;
}
unflushed_save_points_ - > clear ( ) ;
return Status : : OK ( ) ;
}
Status WriteUnpreparedTxn : : PrepareInternal ( ) { Status WriteUnpreparedTxn : : PrepareInternal ( ) {
const bool kPrepared = true ; const bool kPrepared = true ;
return FlushWriteBatchToDB ( kPrepared ) ; return FlushWriteBatchToDB ( kPrepared ) ;
@ -379,6 +500,8 @@ Status WriteUnpreparedTxn::CommitInternal() {
wpt_db_ - > RemovePrepared ( commit_batch_seq , commit_batch_cnt ) ; wpt_db_ - > RemovePrepared ( commit_batch_seq , commit_batch_cnt ) ;
} }
unprep_seqs_ . clear ( ) ; unprep_seqs_ . clear ( ) ;
flushed_save_points_ . reset ( nullptr ) ;
unflushed_save_points_ . reset ( nullptr ) ;
return s ; return s ;
} // else do the 2nd write to publish seq
} // else do the 2nd write to publish seq
@ -410,6 +533,8 @@ Status WriteUnpreparedTxn::CommitInternal() {
wpt_db_ - > RemovePrepared ( seq . first , seq . second ) ; wpt_db_ - > RemovePrepared ( seq . first , seq . second ) ;
} }
unprep_seqs_ . clear ( ) ; unprep_seqs_ . clear ( ) ;
flushed_save_points_ . reset ( nullptr ) ;
unflushed_save_points_ . reset ( nullptr ) ;
return s ; return s ;
} }
@ -488,6 +613,8 @@ Status WriteUnpreparedTxn::RollbackInternal() {
wpt_db_ - > RemovePrepared ( seq . first , seq . second ) ; wpt_db_ - > RemovePrepared ( seq . first , seq . second ) ;
} }
unprep_seqs_ . clear ( ) ; unprep_seqs_ . clear ( ) ;
flushed_save_points_ . reset ( nullptr ) ;
unflushed_save_points_ . reset ( nullptr ) ;
return s ; return s ;
} // else do the 2nd write for commit
} // else do the 2nd write for commit
uint64_t & prepare_seq = seq_used ; uint64_t & prepare_seq = seq_used ;
@ -514,6 +641,8 @@ Status WriteUnpreparedTxn::RollbackInternal() {
} }
unprep_seqs_ . clear ( ) ; unprep_seqs_ . clear ( ) ;
flushed_save_points_ . reset ( nullptr ) ;
unflushed_save_points_ . reset ( nullptr ) ;
return s ; return s ;
} }
@ -524,6 +653,131 @@ void WriteUnpreparedTxn::Clear() {
TransactionBaseImpl : : Clear ( ) ; TransactionBaseImpl : : Clear ( ) ;
} }
void WriteUnpreparedTxn : : SetSavePoint ( ) {
assert ( ( unflushed_save_points_ ? unflushed_save_points_ - > size ( ) : 0 ) +
( flushed_save_points_ ? flushed_save_points_ - > size ( ) : 0 ) = =
( save_points_ ? save_points_ - > size ( ) : 0 ) ) ;
PessimisticTransaction : : SetSavePoint ( ) ;
if ( unflushed_save_points_ = = nullptr ) {
unflushed_save_points_ . reset ( new autovector < size_t > ( ) ) ;
}
unflushed_save_points_ - > push_back ( write_batch_ . GetDataSize ( ) ) ;
}
Status WriteUnpreparedTxn : : RollbackToSavePoint ( ) {
assert ( ( unflushed_save_points_ ? unflushed_save_points_ - > size ( ) : 0 ) +
( flushed_save_points_ ? flushed_save_points_ - > size ( ) : 0 ) = =
( save_points_ ? save_points_ - > size ( ) : 0 ) ) ;
if ( unflushed_save_points_ ! = nullptr & & unflushed_save_points_ - > size ( ) > 0 ) {
Status s = PessimisticTransaction : : RollbackToSavePoint ( ) ;
assert ( ! s . IsNotFound ( ) ) ;
unflushed_save_points_ - > pop_back ( ) ;
return s ;
}
if ( flushed_save_points_ ! = nullptr & & ! flushed_save_points_ - > empty ( ) ) {
return RollbackToSavePointInternal ( ) ;
}
return Status : : NotFound ( ) ;
}
Status WriteUnpreparedTxn : : RollbackToSavePointInternal ( ) {
Status s ;
const bool kClear = true ;
TransactionBaseImpl : : InitWriteBatch ( kClear ) ;
assert ( flushed_save_points_ - > size ( ) > 0 ) ;
WriteUnpreparedTxn : : SavePoint & top = flushed_save_points_ - > back ( ) ;
assert ( top . unprep_seqs_ . size ( ) > 0 ) ;
assert ( save_points_ ! = nullptr & & save_points_ - > size ( ) > 0 ) ;
const TransactionKeyMap & tracked_keys = save_points_ - > top ( ) . new_keys_ ;
// TODO(lth): Reduce duplicate code with RollbackInternal logic.
ReadOptions roptions ;
roptions . snapshot = top . snapshot_ - > snapshot ( ) ;
SequenceNumber min_uncommitted =
static_cast_with_check < const SnapshotImpl , const Snapshot > (
roptions . snapshot )
- > min_uncommitted_ ;
SequenceNumber snap_seq = roptions . snapshot - > GetSequenceNumber ( ) ;
WriteUnpreparedTxnReadCallback callback ( wupt_db_ , snap_seq , min_uncommitted ,
top . unprep_seqs_ ) ;
const auto & cf_map = * wupt_db_ - > GetCFHandleMap ( ) ;
for ( const auto & cfkey : tracked_keys ) {
const auto cfid = cfkey . first ;
const auto & keys = cfkey . second ;
for ( const auto & pair : keys ) {
const auto & key = pair . first ;
const auto & cf_handle = cf_map . at ( cfid ) ;
PinnableSlice pinnable_val ;
bool not_used ;
s = db_impl_ - > GetImpl ( roptions , cf_handle , key , & pinnable_val , & not_used ,
& callback ) ;
if ( s . ok ( ) ) {
s = write_batch_ . Put ( cf_handle , key , pinnable_val ) ;
assert ( s . ok ( ) ) ;
} else if ( s . IsNotFound ( ) ) {
s = write_batch_ . Delete ( cf_handle , key ) ;
assert ( s . ok ( ) ) ;
} else {
return s ;
}
}
}
const bool kPrepared = true ;
s = FlushWriteBatchToDBInternal ( ! kPrepared ) ;
assert ( s . ok ( ) ) ;
if ( ! s . ok ( ) ) {
return s ;
}
// PessimisticTransaction::RollbackToSavePoint will call also call
// RollbackToSavepoint on write_batch_. However, write_batch_ is empty and has
// no savepoints because this savepoint has already been flushed. Work around
// this by setting a fake savepoint.
write_batch_ . SetSavePoint ( ) ;
s = PessimisticTransaction : : RollbackToSavePoint ( ) ;
assert ( s . ok ( ) ) ;
if ( ! s . ok ( ) ) {
return s ;
}
flushed_save_points_ - > pop_back ( ) ;
return s ;
}
Status WriteUnpreparedTxn : : PopSavePoint ( ) {
assert ( ( unflushed_save_points_ ? unflushed_save_points_ - > size ( ) : 0 ) +
( flushed_save_points_ ? flushed_save_points_ - > size ( ) : 0 ) = =
( save_points_ ? save_points_ - > size ( ) : 0 ) ) ;
if ( unflushed_save_points_ ! = nullptr & & unflushed_save_points_ - > size ( ) > 0 ) {
Status s = PessimisticTransaction : : PopSavePoint ( ) ;
assert ( ! s . IsNotFound ( ) ) ;
unflushed_save_points_ - > pop_back ( ) ;
return s ;
}
if ( flushed_save_points_ ! = nullptr & & ! flushed_save_points_ - > empty ( ) ) {
// PessimisticTransaction::PopSavePoint will call also call PopSavePoint on
// write_batch_. However, write_batch_ is empty and has no savepoints
// because this savepoint has already been flushed. Work around this by
// setting a fake savepoint.
write_batch_ . SetSavePoint ( ) ;
Status s = PessimisticTransaction : : PopSavePoint ( ) ;
assert ( ! s . IsNotFound ( ) ) ;
flushed_save_points_ - > pop_back ( ) ;
return s ;
}
return Status : : NotFound ( ) ;
}
void WriteUnpreparedTxn : : MultiGet ( const ReadOptions & options , void WriteUnpreparedTxn : : MultiGet ( const ReadOptions & options ,
ColumnFamilyHandle * column_family , ColumnFamilyHandle * column_family ,
const size_t num_keys , const Slice * keys , const size_t num_keys , const Slice * keys ,
Manuel Ung
5 years ago
committed by