func (p *planner) getTableNames(dbDesc *DatabaseDescriptor) (parser.QualifiedNames, *roachpb.Error) {
prefix := MakeNameMetadataKey(dbDesc.ID, "")
sr, pErr := p.txn.Scan(prefix, prefix.PrefixEnd(), 0)
if pErr != nil {
return nil, pErr
}
var qualifiedNames parser.QualifiedNames
for _, row := range sr {
_, tableName, err := encoding.DecodeStringAscending(
bytes.TrimPrefix(row.Key, prefix), nil)
if err != nil {
return nil, roachpb.NewError(err)
}
qname := &parser.QualifiedName{
Base: parser.Name(dbDesc.Name),
Indirect: parser.Indirection{parser.NameIndirection(tableName)},
}
if err := qname.NormalizeTableName(""); err != nil {
return nil, roachpb.NewError(err)
}
qualifiedNames = append(qualifiedNames, qname)
}
return qualifiedNames, nil
}
// sendRPC sends one or more RPCs to replicas from the supplied roachpb.Replica
// slice. First, replicas which have gossiped addresses are corralled (and
// rearranged depending on proximity and whether the request needs to go to a
// leader) and then sent via Send, with requirement that one RPC to a server
// must succeed. Returns an RPC error if the request could not be sent. Note
// that the reply may contain a higher level error and must be checked in
// addition to the RPC error.
func (ds *DistSender) sendRPC(ctx context.Context, rangeID roachpb.RangeID, replicas ReplicaSlice,
order orderingPolicy, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if len(replicas) == 0 {
return nil, roachpb.NewError(noNodeAddrsAvailError{})
}
// TODO(pmattis): This needs to be tested. If it isn't set we'll
// still route the request appropriately by key, but won't receive
// RangeNotFoundErrors.
ba.RangeID = rangeID
// Set RPC opts with stipulation that one of N RPCs must succeed.
rpcOpts := SendOptions{
Ordering: order,
SendNextTimeout: defaultSendNextTimeout,
Timeout: base.NetworkTimeout,
Context: ctx,
}
tracing.AnnotateTrace()
defer tracing.AnnotateTrace()
reply, err := ds.rpcSend(rpcOpts, replicas, ba, ds.rpcContext)
if err != nil {
return nil, roachpb.NewError(err)
}
return reply, nil
}
// TestEndWriteRestartReadOnlyTransaction verifies that if
// a transaction writes, then restarts and turns read-only,
// an explicit EndTransaction call is still sent if retry-
// able didn't, regardless of whether there is an error
// or not.
func TestEndWriteRestartReadOnlyTransaction(t *testing.T) {
defer leaktest.AfterTest(t)
for _, success := range []bool{true, false} {
expCalls := []roachpb.Method{roachpb.BeginTransaction, roachpb.Put, roachpb.EndTransaction}
var calls []roachpb.Method
db := newDB(newTestSender(func(ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
calls = append(calls, ba.Methods()...)
return ba.CreateReply(), nil
}, nil))
ok := false
if pErr := db.Txn(func(txn *Txn) *roachpb.Error {
if !ok {
if pErr := txn.Put("consider", "phlebas"); pErr != nil {
t.Fatal(pErr)
}
ok = true
return roachpb.NewError(&roachpb.TransactionRetryError{}) // immediate txn retry
}
if !success {
return roachpb.NewError(errors.New("aborting on purpose"))
}
return nil
}); pErr == nil != success {
t.Errorf("expected error: %t, got error: %v", !success, pErr)
}
if !reflect.DeepEqual(expCalls, calls) {
t.Fatalf("expected %v, got %v", expCalls, calls)
}
}
}
// Send forwards the call to the single store. This is a poor man's
// version of kv.TxnCoordSender, but it serves the purposes of
// supporting tests in this package. Transactions are not supported.
// Since kv/ depends on storage/, we can't get access to a
// TxnCoordSender from here.
// TODO(tschottdorf): {kv->storage}.LocalSender
func (db *testSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if et, ok := ba.GetArg(roachpb.EndTransaction); ok {
return nil, roachpb.NewError(util.Errorf("%s method not supported", et.Method()))
}
// Lookup range and direct request.
key, endKey := keys.Range(ba)
rng := db.store.LookupReplica(key, endKey)
if rng == nil {
return nil, roachpb.NewError(roachpb.NewRangeKeyMismatchError(key, endKey, nil))
}
ba.RangeID = rng.Desc().RangeID
replica := rng.GetReplica()
if replica == nil {
return nil, roachpb.NewError(util.Errorf("own replica missing in range"))
}
ba.Replica = *replica
br, pErr := db.store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(db.store, br))
}
if pErr != nil {
return nil, pErr
}
return br, nil
}
// getCachedDatabaseDesc looks up the database descriptor given its name in the
// descriptor cache.
func (p *planner) getCachedDatabaseDesc(name string) (*DatabaseDescriptor, *roachpb.Error) {
if name == systemDB.Name {
return &systemDB, nil
}
nameKey := databaseKey{name}
nameVal := p.systemConfig.GetValue(nameKey.Key())
if nameVal == nil {
return nil, roachpb.NewUErrorf("database %q does not exist in system cache", name)
}
id, err := nameVal.GetInt()
if err != nil {
return nil, roachpb.NewError(err)
}
descKey := MakeDescMetadataKey(ID(id))
descVal := p.systemConfig.GetValue(descKey)
if descVal == nil {
return nil, roachpb.NewUErrorf("database %q has name entry, but no descriptor in system cache", name)
}
desc := &Descriptor{}
if err := descVal.GetProto(desc); err != nil {
return nil, roachpb.NewError(err)
}
database := desc.GetDatabase()
if database == nil {
return nil, roachpb.NewErrorf("%q is not a database", name)
}
return database, roachpb.NewError(database.Validate())
}
// getDescriptorFromTargetList examines a TargetList and fetches the
// appropriate descriptor.
// TODO(marc): support multiple targets.
func (p *planner) getDescriptorFromTargetList(targets parser.TargetList) (descriptorProto, *roachpb.Error) {
if targets.Databases != nil {
if len(targets.Databases) == 0 {
return nil, roachpb.NewError(errNoDatabase)
} else if len(targets.Databases) != 1 {
return nil, roachpb.NewErrorf("TODO(marc): multiple targets not implemented")
}
descriptor, err := p.getDatabaseDesc(targets.Databases[0])
if err != nil {
return nil, err
}
return descriptor, nil
}
if len(targets.Tables) == 0 {
return nil, roachpb.NewError(errNoTable)
} else if len(targets.Tables) != 1 {
return nil, roachpb.NewErrorf("TODO(marc): multiple targets not implemented")
}
descriptor, err := p.getTableDesc(targets.Tables[0])
if err != nil {
return nil, err
}
return descriptor, nil
}
// Prepare returns the result types of the given statement. Args may be a
// partially populated val args map. Prepare will populate the missing val
// args. The column result types are returned (or nil if there are no results).
func (e *Executor) Prepare(query string, session *Session, args parser.MapArgs) (
[]ResultColumn, *roachpb.Error) {
stmt, err := parser.ParseOne(query, parser.Syntax(session.Syntax))
if err != nil {
return nil, roachpb.NewError(err)
}
session.planner.resetForBatch(e)
session.planner.evalCtx.Args = args
session.planner.evalCtx.PrepareOnly = true
// TODO(andrei): does the prepare phase really need a Txn?
txn := client.NewTxn(*e.ctx.DB)
txn.Proto.Isolation = session.DefaultIsolationLevel
session.planner.setTxn(txn)
defer session.planner.setTxn(nil)
plan, pErr := session.planner.prepare(stmt)
if pErr != nil {
return nil, pErr
}
if plan == nil {
return nil, nil
}
cols := plan.Columns()
for _, c := range cols {
if err := checkResultDatum(c.Typ); err != nil {
return nil, roachpb.NewError(err)
}
}
return cols, nil
}
func (p *planner) SetTimeZone(n *parser.SetTimeZone) (planNode, *roachpb.Error) {
d, err := n.Value.Eval(p.evalCtx)
if err != nil {
return nil, roachpb.NewError(err)
}
var offset int64
switch v := d.(type) {
case parser.DString:
location := string(v)
if location == "DEFAULT" || location == "LOCAL" {
location = "UTC"
}
if _, err := time.LoadLocation(location); err != nil {
return nil, roachpb.NewError(err)
}
p.session.Timezone = &Session_Location{Location: location}
case parser.DInterval:
offset = int64(v.Duration / time.Second)
case parser.DInt:
offset = int64(v) * 60 * 60
case parser.DFloat:
offset = int64(float64(v) * 60.0 * 60.0)
default:
return nil, roachpb.NewUErrorf("bad time zone value: %v", n.Value)
}
if offset != 0 {
p.session.Timezone = &Session_Offset{Offset: offset}
}
p.evalCtx.GetLocation = p.session.getLocation
return &valuesNode{}, nil
}
// execStmtInAbortedTxn executes a statement in a txn that's in state
// Aborted or RestartWait.
// Everything but COMMIT/ROLLBACK/RESTART causes errors.
func (e *Executor) execStmtInAbortedTxn(
stmt parser.Statement, txnState *txnState) (Result, *roachpb.Error) {
if txnState.State != Aborted && txnState.State != RestartWait {
panic("execStmtInAbortedTxn called outside of an aborted txn")
}
switch stmt.(type) {
case *parser.CommitTransaction, *parser.RollbackTransaction:
if txnState.State == RestartWait {
if pErr := txnState.txn.Rollback(); pErr != nil {
log.Errorf("failure rolling back transaction: %s", pErr)
}
}
// Reset the state to allow new transactions to start.
// Note: postgres replies to COMMIT of failed txn with "ROLLBACK" too.
result := Result{PGTag: (*parser.RollbackTransaction)(nil).StatementTag()}
txnState.resetStateAndTxn(NoTxn)
return result, nil
case *parser.RestartTransaction:
if txnState.State == RestartWait {
// Reset the state. Txn is Open again.
txnState.State = Open
// TODO(andrei/cdo): add a counter for user-directed retries.
return Result{}, nil
}
pErr := roachpb.NewError(&roachpb.SqlTransactionAbortedError{
CustomMsg: "RETRY INTENT has not been used or a non-retriable error was encountered."})
return Result{PErr: pErr}, pErr
default:
pErr := roachpb.NewError(&roachpb.SqlTransactionAbortedError{})
return Result{PErr: pErr}, pErr
}
}
// getTableID retrieves the table ID for the specified table. It uses the
// descriptor cache to perform lookups, falling back to the KV store when
// necessary.
func (p *planner) getTableID(qname *parser.QualifiedName) (ID, *roachpb.Error) {
if err := qname.NormalizeTableName(p.session.Database); err != nil {
return 0, roachpb.NewError(err)
}
dbID, pErr := p.getDatabaseID(qname.Database())
if pErr != nil {
return 0, pErr
}
// Lookup the ID of the table in the cache. The use of the cache might cause
// the usage of a recently renamed table, but that's a race that could occur
// anyways.
nameKey := tableKey{dbID, qname.Table()}
key := nameKey.Key()
if nameVal := p.systemConfig.GetValue(key); nameVal != nil {
id, err := nameVal.GetInt()
return ID(id), roachpb.NewError(err)
}
gr, pErr := p.txn.Get(key)
if pErr != nil {
return 0, pErr
}
if !gr.Exists() {
return 0, roachpb.NewErrorf("table %q does not exist", nameKey.Name())
}
return ID(gr.ValueInt()), nil
}
func (n *scanNode) initWhere(where *parser.Where) *roachpb.Error {
if where == nil {
return nil
}
n.filter, n.pErr = n.resolveQNames(where.Expr)
if n.pErr == nil {
var whereType parser.Datum
var err error
whereType, err = n.filter.TypeCheck(n.planner.evalCtx.Args)
n.pErr = roachpb.NewError(err)
if n.pErr == nil {
if !(whereType == parser.DummyBool || whereType == parser.DNull) {
n.pErr = roachpb.NewUErrorf("argument of WHERE must be type %s, not type %s", parser.DummyBool.Type(), whereType.Type())
}
}
}
if n.pErr == nil {
// Normalize the expression (this will also evaluate any branches that are
// constant).
var err error
n.filter, err = n.planner.parser.NormalizeExpr(n.planner.evalCtx, n.filter)
n.pErr = roachpb.NewError(err)
}
if n.pErr == nil {
n.filter, n.pErr = n.planner.expandSubqueries(n.filter, 1)
}
return n.pErr
}
// RenameDatabase renames the database.
// Privileges: "root" user.
// Notes: postgres requires superuser, db owner, or "CREATEDB".
// mysql >= 5.1.23 does not allow database renames.
func (p *planner) RenameDatabase(n *parser.RenameDatabase) (planNode, *roachpb.Error) {
if n.Name == "" || n.NewName == "" {
return nil, roachpb.NewError(errEmptyDatabaseName)
}
if p.user != security.RootUser {
return nil, roachpb.NewUErrorf("only %s is allowed to rename databases", security.RootUser)
}
dbDesc, pErr := p.getDatabaseDesc(string(n.Name))
if pErr != nil {
return nil, pErr
}
if n.Name == n.NewName {
// Noop.
return &valuesNode{}, nil
}
// Now update the nameMetadataKey and the descriptor.
descKey := MakeDescMetadataKey(dbDesc.GetID())
dbDesc.SetName(string(n.NewName))
if err := dbDesc.Validate(); err != nil {
return nil, roachpb.NewError(err)
}
newKey := databaseKey{string(n.NewName)}.Key()
oldKey := databaseKey{string(n.Name)}.Key()
descID := dbDesc.GetID()
descDesc := wrapDescriptor(dbDesc)
b := client.Batch{}
b.CPut(newKey, descID, nil)
b.Put(descKey, descDesc)
b.Del(oldKey)
if pErr := p.txn.Run(&b); pErr != nil {
if _, ok := pErr.GoError().(*roachpb.ConditionFailedError); ok {
return nil, roachpb.NewUErrorf("the new database name %q already exists", string(n.NewName))
}
return nil, pErr
}
p.testingVerifyMetadata = func(systemConfig config.SystemConfig) error {
if err := expectDescriptorID(systemConfig, newKey, descID); err != nil {
return err
}
if err := expectDescriptor(systemConfig, descKey, descDesc); err != nil {
return err
}
return expectDeleted(systemConfig, oldKey)
}
return &valuesNode{}, nil
}
// initScan initializes but does not perform the key-value scan.
func (n *scanNode) initScan() bool {
// Initialize our key/values.
if len(n.spans) == 0 {
// If no spans were specified retrieve all of the keys that start with our
// index key prefix.
start := roachpb.Key(MakeIndexKeyPrefix(n.desc.ID, n.index.ID))
n.spans = append(n.spans, span{
start: start,
end: start.PrefixEnd(),
})
}
if n.valTypes == nil {
// Prepare our index key vals slice.
var err error
n.valTypes, err = makeKeyVals(&n.desc, n.columnIDs)
n.pErr = roachpb.NewError(err)
if n.pErr != nil {
return false
}
n.vals = make([]parser.Datum, len(n.valTypes))
if n.isSecondaryIndex && n.index.Unique {
// Unique secondary indexes have a value that is the primary index
// key. Prepare implicitVals for use in decoding this value.
// Primary indexes only contain ascendingly-encoded values. If this
// ever changes, we'll probably have to figure out the directions here too.
var err error
n.implicitValTypes, err = makeKeyVals(&n.desc, n.index.ImplicitColumnIDs)
n.pErr = roachpb.NewError(err)
if n.pErr != nil {
return false
}
n.implicitVals = make([]parser.Datum, len(n.implicitValTypes))
}
}
// If we have a limit hint, we limit the first batch size. Subsequent batches use the normal
// size, to avoid making things too slow (e.g. in case we have a very restrictive filter and
// actually have to retrieve a lot of rows).
firstBatchLimit := n.limitHint
if firstBatchLimit != 0 {
// For a secondary index, we have one key per row.
if !n.isSecondaryIndex {
// We have a sentinel key per row plus at most one key per non-PK column. Of course, we
// may have other keys due to a schema change, but this is only a hint.
firstBatchLimit *= int64(1 + len(n.visibleCols) - len(n.index.ColumnIDs))
}
// We need an extra key to make sure we form the last row.
firstBatchLimit++
}
n.fetcher = makeKVFetcher(n.txn, n.spans, n.reverse, firstBatchLimit)
n.scanInitialized = true
return true
}
// RenameIndex renames the index.
// Privileges: CREATE on table.
// notes: postgres requires CREATE on the table.
// mysql requires ALTER, CREATE, INSERT on the table.
func (p *planner) RenameIndex(n *parser.RenameIndex) (planNode, *roachpb.Error) {
newIdxName := string(n.NewName)
if newIdxName == "" {
return nil, roachpb.NewError(errEmptyIndexName)
}
if err := n.Name.NormalizeTableName(p.session.Database); err != nil {
return nil, roachpb.NewError(err)
}
tableDesc, pErr := p.getTableDesc(n.Name)
if pErr != nil {
return nil, pErr
}
idxName := n.Name.Index()
status, i, pErr := tableDesc.FindIndexByName(idxName)
if pErr != nil {
if n.IfExists {
// Noop.
return &valuesNode{}, nil
}
// Index does not exist, but we want it to: error out.
return nil, pErr
}
if pErr := p.checkPrivilege(tableDesc, privilege.CREATE); pErr != nil {
return nil, pErr
}
if equalName(idxName, newIdxName) {
// Noop.
return &valuesNode{}, nil
}
if _, _, err := tableDesc.FindIndexByName(newIdxName); err == nil {
return nil, roachpb.NewUErrorf("index name %q already exists", n.NewName)
}
if status == DescriptorActive {
tableDesc.Indexes[i].Name = newIdxName
} else {
tableDesc.Mutations[i].GetIndex().Name = newIdxName
}
tableDesc.UpVersion = true
descKey := MakeDescMetadataKey(tableDesc.GetID())
if err := tableDesc.Validate(); err != nil {
return nil, roachpb.NewError(err)
}
if pErr := p.txn.Put(descKey, wrapDescriptor(tableDesc)); pErr != nil {
return nil, pErr
}
p.notifySchemaChange(tableDesc.ID, invalidMutationID)
return &valuesNode{}, nil
}
func (p *planner) query(sql string, args ...interface{}) (planNode, *roachpb.Error) {
stmt, err := parser.ParseOneTraditional(sql)
if err != nil {
return nil, roachpb.NewError(err)
}
if err := parser.FillArgs(stmt, golangParameters(args)); err != nil {
return nil, roachpb.NewError(err)
}
return p.makePlan(stmt, false)
}
// Initializes a scanNode with a tableName. Returns the table or index name that can be used for
// fully-qualified columns if an alias is not specified.
func (n *scanNode) initTable(p *planner, tableName *parser.QualifiedName) (string, *roachpb.Error) {
if n.desc, n.pErr = p.getTableLease(tableName); n.pErr != nil {
return "", n.pErr
}
if err := p.checkPrivilege(&n.desc, privilege.SELECT); err != nil {
return "", roachpb.NewError(err)
}
alias := n.desc.Name
indexName := tableName.Index()
if indexName != "" && !equalName(n.desc.PrimaryIndex.Name, indexName) {
for i := range n.desc.Indexes {
if equalName(n.desc.Indexes[i].Name, indexName) {
// Remove all but the matching index from the descriptor.
n.desc.Indexes = n.desc.Indexes[i : i+1]
n.index = &n.desc.Indexes[0]
break
}
}
if n.index == nil {
n.pErr = roachpb.NewUErrorf("index \"%s\" not found", indexName)
return "", n.pErr
}
// Use the index name instead of the table name for fully-qualified columns in the
// expression.
alias = n.index.Name
// Strip out any columns from the table that are not present in the
// index.
visibleCols := make([]ColumnDescriptor, 0, len(n.index.ColumnIDs)+len(n.index.ImplicitColumnIDs))
for _, colID := range n.index.ColumnIDs {
col, err := n.desc.FindColumnByID(colID)
n.pErr = roachpb.NewError(err)
if n.pErr != nil {
return "", n.pErr
}
visibleCols = append(visibleCols, *col)
}
for _, colID := range n.index.ImplicitColumnIDs {
col, err := n.desc.FindColumnByID(colID)
n.pErr = roachpb.NewError(err)
if n.pErr != nil {
return "", n.pErr
}
visibleCols = append(visibleCols, *col)
}
n.isSecondaryIndex = true
n.initVisibleCols(visibleCols, len(n.index.ImplicitColumnIDs))
} else {
n.initDescDefaults()
}
return alias, nil
}
func (s *selectNode) addRender(target parser.SelectExpr) *roachpb.Error {
// outputName will be empty if the target is not aliased.
outputName := string(target.As)
if isStar, cols, exprs, err := checkRenderStar(target, &s.table, s.qvals); err != nil {
s.pErr = roachpb.NewError(err)
return s.pErr
} else if isStar {
s.columns = append(s.columns, cols...)
s.render = append(s.render, exprs...)
return nil
}
// When generating an output column name it should exactly match the original
// expression, so determine the output column name before we perform any
// manipulations to the expression.
outputName = getRenderColName(target)
// Resolve qualified names. This has the side-effect of normalizing any
// qualified name found.
var resolved parser.Expr
var err error
if resolved, err = s.resolveQNames(target.Expr); err != nil {
s.pErr = roachpb.NewError(err)
return s.pErr
}
if resolved, s.pErr = s.planner.expandSubqueries(resolved, 1); s.pErr != nil {
return s.pErr
}
var typ parser.Datum
typ, err = resolved.TypeCheck(s.planner.evalCtx.Args)
s.pErr = roachpb.NewError(err)
if s.pErr != nil {
return s.pErr
}
var normalized parser.Expr
normalized, err = s.planner.parser.NormalizeExpr(s.planner.evalCtx, resolved)
s.pErr = roachpb.NewError(err)
if s.pErr != nil {
return s.pErr
}
s.render = append(s.render, normalized)
if target.As == "" {
switch t := target.Expr.(type) {
case *parser.QualifiedName:
// If the expression is a qualified name, use the column name, not the
// full qualification as the column name to return.
outputName = t.Column()
}
}
s.columns = append(s.columns, ResultColumn{Name: outputName, Typ: typ})
return nil
}
// CreateTable creates a table.
// Privileges: CREATE on database.
// Notes: postgres/mysql require CREATE on database.
func (p *planner) CreateTable(n *parser.CreateTable) (planNode, *roachpb.Error) {
if err := n.Table.NormalizeTableName(p.session.Database); err != nil {
return nil, roachpb.NewError(err)
}
dbDesc, pErr := p.getDatabaseDesc(n.Table.Database())
if pErr != nil {
return nil, pErr
}
if err := p.checkPrivilege(dbDesc, privilege.CREATE); err != nil {
return nil, roachpb.NewError(err)
}
desc, pErr := makeTableDesc(n, dbDesc.ID)
if pErr != nil {
return nil, pErr
}
// Inherit permissions from the database descriptor.
desc.Privileges = dbDesc.GetPrivileges()
if len(desc.PrimaryIndex.ColumnNames) == 0 {
// Ensure a Primary Key exists.
s := "experimental_unique_int()"
col := ColumnDescriptor{
Name: "rowid",
Type: ColumnType{
Kind: ColumnType_INT,
},
DefaultExpr: &s,
Hidden: true,
Nullable: false,
}
desc.AddColumn(col)
idx := IndexDescriptor{
Unique: true,
ColumnNames: []string{col.Name},
ColumnDirections: []IndexDescriptor_Direction{IndexDescriptor_ASC},
}
if err := desc.AddIndex(idx, true); err != nil {
return nil, roachpb.NewError(err)
}
}
if pErr := desc.AllocateIDs(); pErr != nil {
return nil, pErr
}
if pErr := p.createDescriptor(tableKey{dbDesc.ID, n.Table.Table()}, &desc, n.IfNotExists); pErr != nil {
return nil, pErr
}
return &emptyNode{}, nil
}
// execStmtInOpenTxn executes one statement in the context
// of the planner's transaction (which is assumed to exist).
// It handles statements that affect the transaction state (BEGIN, COMMIT)
// and delegates everything else to `execStmt`.
// It binds placeholders.
//
// The current transaction might be committed/rolled back when this returns.
//
// Args:
// abortedMode: if set, we're in a transaction that has encountered errors, so we
// must reject the statement unless it's a COMMIT/ROLLBACK.
// implicitTxn: set if the current transaction was implicitly
// created by the system (i.e. the client sent the statement outside of
// a transaction).
// COMMIT/ROLLBACK statements are rejected if set. Also, the transaction
// might be auto-committed in this function.
// firstInTxn: set for the first statement in a transaction. Used
// so that nested BEGIN statements are caught.
// stmtTimestamp: Used as the statement_timestamp().
//
// Returns:
// - a Result
// - an error, if any. In case of error, the result returned also reflects this error.
func (e *Executor) execStmtInOpenTxn(
stmt parser.Statement, planMaker *planner,
implicitTxn bool,
firstInTxn bool,
stmtTimestamp parser.DTimestamp,
txnState *txnState) (Result, *roachpb.Error) {
if txnState.state() != openTransaction {
panic("execStmtInOpenTxn called outside of an open txn")
}
if planMaker.txn == nil {
panic("execStmtInOpenTxn called with the a txn not set on the planner")
}
planMaker.evalCtx.StmtTimestamp = stmtTimestamp
// TODO(cdo): Figure out how to not double count on retries.
e.updateStmtCounts(stmt)
switch stmt.(type) {
case *parser.BeginTransaction:
if !firstInTxn {
txnState.aborted = true
pErr := roachpb.NewError(errTransactionInProgress)
return Result{PErr: pErr}, pErr
}
case *parser.CommitTransaction, *parser.RollbackTransaction, *parser.SetTransaction:
if implicitTxn {
txnState.aborted = true
pErr := roachpb.NewError(errNoTransactionInProgress)
return Result{PErr: pErr}, pErr
}
}
// Bind all the placeholder variables in the stmt to actual values.
stmt, err := parser.FillArgs(stmt, &planMaker.params)
if err != nil {
txnState.aborted = true
pErr := roachpb.NewError(err)
return Result{PErr: pErr}, pErr
}
result, pErr := e.execStmt(stmt, planMaker, timeutil.Now(),
implicitTxn /* autoCommit */)
txnDone := planMaker.txn == nil
if pErr != nil {
result = Result{PErr: pErr}
txnState.aborted = true
}
if txnDone {
txnState.aborted = false
txnState.txn = nil
}
return result, pErr
}
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
// If we aren't given a Replica, then a little bending over
// backwards here. This case applies exclusively to unittests.
if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
rs, err := keys.Range(ba)
if err != nil {
return nil, roachpb.NewError(err)
}
rangeID, repl, err := ls.lookupReplica(rs.Key, rs.EndKey)
if err != nil {
return nil, roachpb.NewError(err)
}
ba.RangeID = rangeID
ba.Replica = *repl
}
ctx = log.Add(ctx,
log.RangeID, ba.RangeID)
store, err := ls.GetStore(ba.Replica.StoreID)
if err != nil {
return nil, roachpb.NewError(err)
}
if ba.Txn != nil {
// For calls that read data within a txn, we keep track of timestamps
// observed from the various participating nodes' HLC clocks. If we have
// a timestamp on file for this Node which is smaller than MaxTimestamp,
// we can lower MaxTimestamp accordingly. If MaxTimestamp drops below
// OrigTimestamp, we effectively can't see uncertainty restarts any
// more.
// Note that it's not an issue if MaxTimestamp propagates back out to
// the client via a returned Transaction update - when updating a Txn
// from another, the larger MaxTimestamp wins.
if maxTS, ok := ba.Txn.GetObservedTimestamp(ba.Replica.NodeID); ok && maxTS.Less(ba.Txn.MaxTimestamp) {
// Copy-on-write to protect others we might be sharing the Txn with.
shallowTxn := *ba.Txn
// The uncertainty window is [OrigTimestamp, maxTS), so if that window
// is empty, there won't be any uncertainty restarts.
if !ba.Txn.OrigTimestamp.Less(maxTS) {
log.Trace(ctx, "read has no clock uncertainty")
}
shallowTxn.MaxTimestamp.Backward(maxTS)
ba.Txn = &shallowTxn
}
}
br, pErr := store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(store, br))
}
return br, pErr
}
// DropIndex drops an index.
// Privileges: CREATE on table.
// Notes: postgres allows only the index owner to DROP an index.
// mysql requires the INDEX privilege on the table.
func (p *planner) DropIndex(n *parser.DropIndex) (planNode, *roachpb.Error) {
for _, indexQualifiedName := range n.Names {
if err := indexQualifiedName.NormalizeTableName(p.session.Database); err != nil {
return nil, roachpb.NewError(err)
}
tableDesc, pErr := p.getTableDesc(indexQualifiedName)
if pErr != nil {
return nil, pErr
}
if err := p.checkPrivilege(&tableDesc, privilege.CREATE); err != nil {
return nil, roachpb.NewError(err)
}
idxName := indexQualifiedName.Index()
status, i, err := tableDesc.FindIndexByName(idxName)
if err != nil {
if n.IfExists {
// Noop.
return &emptyNode{}, nil
}
// Index does not exist, but we want it to: error out.
return nil, roachpb.NewError(err)
}
switch status {
case DescriptorActive:
tableDesc.addIndexMutation(tableDesc.Indexes[i], DescriptorMutation_DROP)
tableDesc.Indexes = append(tableDesc.Indexes[:i], tableDesc.Indexes[i+1:]...)
case DescriptorIncomplete:
switch tableDesc.Mutations[i].Direction {
case DescriptorMutation_ADD:
return nil, roachpb.NewUErrorf("index %q in the middle of being added, try again later", idxName)
case DescriptorMutation_DROP:
return &emptyNode{}, nil
}
}
tableDesc.UpVersion = true
mutationID := tableDesc.NextMutationID
tableDesc.NextMutationID++
if err := tableDesc.Validate(); err != nil {
return nil, roachpb.NewError(err)
}
if pErr := p.txn.Put(MakeDescMetadataKey(tableDesc.GetID()), wrapDescriptor(&tableDesc)); pErr != nil {
return nil, pErr
}
p.notifySchemaChange(tableDesc.ID, mutationID)
}
return &emptyNode{}, nil
}
// Values constructs a valuesNode from a VALUES expression.
func (p *planner) Values(n *parser.Values) (planNode, *roachpb.Error) {
v := &valuesNode{
rows: make([]parser.DTuple, 0, len(n.Tuples)),
}
for num, tupleOrig := range n.Tuples {
if num == 0 {
v.columns = make([]ResultColumn, 0, len(tupleOrig.Exprs))
} else if a, e := len(tupleOrig.Exprs), len(v.columns); a != e {
return nil, roachpb.NewUErrorf("VALUES lists must all be the same length, %d for %d", a, e)
}
// We must not modify the Values node in-place (or the changes will leak if we retry this
// transaction).
tuple := parser.Tuple{Exprs: parser.Exprs(append([]parser.Expr(nil), tupleOrig.Exprs...))}
for i := range tuple.Exprs {
var pErr *roachpb.Error
tuple.Exprs[i], pErr = p.expandSubqueries(tuple.Exprs[i], 1)
if pErr != nil {
return nil, pErr
}
var err error
typ, err := tuple.Exprs[i].TypeCheck(p.evalCtx.Args)
if err != nil {
return nil, roachpb.NewError(err)
}
tuple.Exprs[i], err = p.parser.NormalizeExpr(p.evalCtx, tuple.Exprs[i])
if err != nil {
return nil, roachpb.NewError(err)
}
if num == 0 {
v.columns = append(v.columns, ResultColumn{Name: "column" + strconv.Itoa(i+1), Typ: typ})
} else if v.columns[i].Typ == parser.DNull {
v.columns[i].Typ = typ
} else if typ != parser.DNull && !typ.TypeEqual(v.columns[i].Typ) {
return nil, roachpb.NewUErrorf("VALUES list type mismatch, %s for %s", typ.Type(), v.columns[i].Typ.Type())
}
}
data, err := tuple.Eval(p.evalCtx)
if err != nil {
return nil, roachpb.NewError(err)
}
vals, ok := data.(parser.DTuple)
if !ok {
return nil, roachpb.NewUErrorf("expected a tuple, but found %T", data)
}
v.rows = append(v.rows, vals)
}
return v, nil
}
// CreateIndex creates an index.
// Privileges: CREATE on table.
// notes: postgres requires CREATE on the table.
// mysql requires INDEX on the table.
func (p *planner) CreateIndex(n *parser.CreateIndex) (planNode, *roachpb.Error) {
tableDesc, pErr := p.getTableDesc(n.Table)
if pErr != nil {
return nil, pErr
}
status, i, err := tableDesc.FindIndexByName(string(n.Name))
if err == nil {
if status == DescriptorIncomplete {
switch tableDesc.Mutations[i].Direction {
case DescriptorMutation_DROP:
return nil, roachpb.NewUErrorf("index %q being dropped, try again later", string(n.Name))
case DescriptorMutation_ADD:
// Noop, will fail in AllocateIDs below.
}
}
if n.IfNotExists {
// Noop.
return &emptyNode{}, nil
}
}
if err := p.checkPrivilege(&tableDesc, privilege.CREATE); err != nil {
return nil, roachpb.NewError(err)
}
indexDesc := IndexDescriptor{
Name: string(n.Name),
Unique: n.Unique,
StoreColumnNames: n.Storing,
}
if err := indexDesc.fillColumns(n.Columns); err != nil {
return nil, roachpb.NewError(err)
}
tableDesc.addIndexMutation(indexDesc, DescriptorMutation_ADD)
tableDesc.UpVersion = true
mutationID := tableDesc.NextMutationID
tableDesc.NextMutationID++
if err := tableDesc.AllocateIDs(); err != nil {
return nil, roachpb.NewError(err)
}
if pErr := p.txn.Put(MakeDescMetadataKey(tableDesc.GetID()), wrapDescriptor(&tableDesc)); pErr != nil {
return nil, pErr
}
p.notifySchemaChange(tableDesc.ID, mutationID)
return &emptyNode{}, nil
}
// limit constructs a limitNode based on the LIMIT and OFFSET clauses.
func (p *planner) limit(n *parser.Select, plan planNode) (planNode, *roachpb.Error) {
if n.Limit == nil {
return plan, nil
}
var count, offset int64
data := []struct {
name string
src parser.Expr
dst *int64
defaultVal int64
}{
{"LIMIT", n.Limit.Count, &count, math.MaxInt64},
{"OFFSET", n.Limit.Offset, &offset, 0},
}
for _, datum := range data {
if datum.src == nil {
*datum.dst = datum.defaultVal
} else {
if parser.ContainsVars(datum.src) {
return nil, roachpb.NewUErrorf("argument of %s must not contain variables", datum.name)
}
normalized, err := p.parser.NormalizeExpr(p.evalCtx, datum.src)
if err != nil {
return nil, roachpb.NewError(err)
}
dstDatum, err := normalized.Eval(p.evalCtx)
if err != nil {
return nil, roachpb.NewError(err)
}
if dstDatum == parser.DNull {
*datum.dst = datum.defaultVal
continue
}
if dstDInt, ok := dstDatum.(parser.DInt); ok {
*datum.dst = int64(dstDInt)
continue
}
return nil, roachpb.NewUErrorf("argument of %s must be type %s, not type %s", datum.name, parser.DummyInt.Type(), dstDatum.Type())
}
}
return &limitNode{planNode: plan, count: count, offset: offset}, nil
}
// send runs the specified calls synchronously in a single batch and
// returns any errors. If the transaction is read-only or has already
// been successfully committed or aborted, a potential trailing
// EndTransaction call is silently dropped, allowing the caller to
// always commit or clean-up explicitly even when that may not be
// required (or even erroneous).
func (txn *Txn) send(reqs ...roachpb.Request) (*roachpb.BatchResponse, *roachpb.Error) {
if txn.Proto.Status != roachpb.PENDING {
return nil, roachpb.NewError(util.Errorf("attempting to use %s transaction", txn.Proto.Status))
}
lastIndex := len(reqs) - 1
if lastIndex < 0 {
return &roachpb.BatchResponse{}, nil
}
lastReq := reqs[lastIndex]
// haveTxnWrite tracks intention to write. This is in contrast to
// txn.Proto.Writing, which is set by the coordinator when the first
// intent has been created, and which lives for the life of the
// transaction.
haveTxnWrite := roachpb.IsTransactionWrite(lastReq)
for _, args := range reqs[:lastIndex] {
if _, ok := args.(*roachpb.EndTransactionRequest); ok {
return nil, roachpb.NewError(util.Errorf("%s sent as non-terminal call", args.Method()))
}
if !haveTxnWrite {
haveTxnWrite = roachpb.IsTransactionWrite(args)
}
}
endTxnRequest, haveEndTxn := lastReq.(*roachpb.EndTransactionRequest)
needEndTxn := txn.Proto.Writing || haveTxnWrite
elideEndTxn := haveEndTxn && !needEndTxn
if elideEndTxn {
reqs = reqs[:lastIndex]
}
br, pErr := txn.db.send(reqs...)
if elideEndTxn && pErr == nil {
// This normally happens on the server and sent back in response
// headers, but this transaction was optimized away. The caller may
// still inspect the transaction struct, so we manually update it
// here to emulate a true transaction.
if endTxnRequest.Commit {
txn.Proto.Status = roachpb.COMMITTED
} else {
txn.Proto.Status = roachpb.ABORTED
}
}
return br, pErr
}
// Values constructs a valuesNode from a VALUES expression.
func (p *planner) Values(n parser.Values) (planNode, *roachpb.Error) {
v := &valuesNode{
rows: make([]parser.DTuple, 0, len(n)),
}
for num, tuple := range n {
if num == 0 {
v.columns = make([]resultColumn, 0, len(tuple))
} else if a, e := len(tuple), len(v.columns); a != e {
return nil, roachpb.NewUErrorf("VALUES lists must all be the same length, %d for %d", a, e)
}
for i := range tuple {
var pErr *roachpb.Error
tuple[i], pErr = p.expandSubqueries(tuple[i], 1)
if pErr != nil {
return nil, pErr
}
var err error
typ, err := tuple[i].TypeCheck(p.evalCtx.Args)
if err != nil {
return nil, roachpb.NewError(err)
}
tuple[i], err = p.parser.NormalizeExpr(p.evalCtx, tuple[i])
if err != nil {
return nil, roachpb.NewError(err)
}
if num == 0 {
v.columns = append(v.columns, resultColumn{name: "column" + strconv.Itoa(i+1), typ: typ})
} else if v.columns[i].typ == parser.DNull {
v.columns[i].typ = typ
} else if typ != parser.DNull && typ != v.columns[i].typ {
return nil, roachpb.NewUErrorf("VALUES list type mismatch, %s for %s", typ.Type(), v.columns[i].typ.Type())
}
}
data, err := tuple.Eval(p.evalCtx)
if err != nil {
return nil, roachpb.NewError(err)
}
vals, ok := data.(parser.DTuple)
if !ok {
return nil, roachpb.NewUErrorf("expected a tuple, but found %T", data)
}
v.rows = append(v.rows, vals)
}
return v, nil
}
// unmarshalColumnValue decodes the value from a key-value pair using the type
// expected by the column. An error is returned if the value's type does not
// match the column's type.
func unmarshalColumnValue(kind ColumnType_Kind, value *roachpb.Value) (parser.Datum, *roachpb.Error) {
if value == nil {
return parser.DNull, nil
}
switch kind {
case ColumnType_BOOL:
v, err := value.GetInt()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DBool(v != 0), nil
case ColumnType_INT:
v, err := value.GetInt()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DInt(v), nil
case ColumnType_FLOAT:
v, err := value.GetFloat()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DFloat(v), nil
case ColumnType_STRING:
v, err := value.GetBytes()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DString(v), nil
case ColumnType_BYTES:
v, err := value.GetBytes()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DBytes(v), nil
case ColumnType_DATE:
v, err := value.GetInt()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DDate(v), nil
case ColumnType_TIMESTAMP:
v, err := value.GetTime()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DTimestamp{Time: v}, nil
case ColumnType_INTERVAL:
v, err := value.GetInt()
if err != nil {
return nil, roachpb.NewError(err)
}
return parser.DInterval{Duration: time.Duration(v)}, nil
default:
return nil, roachpb.NewErrorf("unsupported column type: %s", kind)
}
}
// CreateDatabase creates a database.
// Privileges: security.RootUser user.
// Notes: postgres requires superuser or "CREATEDB".
// mysql uses the mysqladmin command.
func (p *planner) CreateDatabase(n *parser.CreateDatabase) (planNode, *roachpb.Error) {
if n.Name == "" {
return nil, roachpb.NewError(errEmptyDatabaseName)
}
if p.session.User != security.RootUser {
return nil, roachpb.NewUErrorf("only %s is allowed to create databases", security.RootUser)
}
desc := makeDatabaseDesc(n)
created, err := p.createDescriptor(databaseKey{string(n.Name)}, &desc, n.IfNotExists)
if err != nil {
return nil, err
}
if created {
// Log Create Database event.
if pErr := MakeEventLogger(p.leaseMgr).InsertEventRecord(p.txn,
EventLogCreateDatabase,
int32(desc.ID),
int32(p.evalCtx.NodeID),
struct {
DatabaseName string
Statement string
User string
}{n.Name.String(), n.String(), p.session.User},
); pErr != nil {
return nil, pErr
}
}
return &emptyNode{}, nil
}
// lookupReplica looks up replica by key [range]. Lookups are done
// by consulting each store in turn via Store.LookupRange(key).
// Returns RangeID and replica on success; RangeKeyMismatch error
// if not found.
// This is only for testing usage; performance doesn't matter.
func (ls *Stores) lookupReplica(start, end roachpb.RKey) (rangeID roachpb.RangeID, replica *roachpb.ReplicaDescriptor, pErr *roachpb.Error) {
ls.mu.RLock()
defer ls.mu.RUnlock()
var rng *Replica
for _, store := range ls.storeMap {
rng = store.LookupReplica(start, end)
if rng == nil {
if tmpRng := store.LookupReplica(start, nil); tmpRng != nil {
log.Warningf(fmt.Sprintf("range not contained in one range: [%s,%s), but have [%s,%s)", start, end, tmpRng.Desc().StartKey, tmpRng.Desc().EndKey))
}
continue
}
if replica == nil {
rangeID = rng.RangeID
replica = rng.GetReplica()
continue
}
// Should never happen outside of tests.
return 0, nil, roachpb.NewErrorf(
"range %+v exists on additional store: %+v", rng, store)
}
if replica == nil {
pErr = roachpb.NewError(roachpb.NewRangeKeyMismatchError(start.AsRawKey(), end.AsRawKey(), nil))
}
return rangeID, replica, pErr
}
// GetProto retrieves the value for a key and decodes the result as a proto
// message.
//
// key can be either a byte slice or a string.
func (txn *Txn) GetProto(key interface{}, msg proto.Message) *roachpb.Error {
r, pErr := txn.Get(key)
if pErr != nil {
return pErr
}
return roachpb.NewError(r.ValueProto(msg))
}