command_result.go

// Copyright 2018 The Cockroach Authors.
//
// Use of this software is governed by the Business Source License
// included in the file licenses/BSL.txt.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0, included in the file
// licenses/APL.txt.

package pgwire

import (
	"context"
	"fmt"
	"time"

	"github.com/cockroachdb/cockroach/pkg/col/coldata"
	"github.com/cockroachdb/cockroach/pkg/jobs"
	"github.com/cockroachdb/cockroach/pkg/server/telemetry"
	"github.com/cockroachdb/cockroach/pkg/sql"
	"github.com/cockroachdb/cockroach/pkg/sql/catalog/colinfo"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgnotice"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgwirebase"
	"github.com/cockroachdb/cockroach/pkg/sql/sem/tree"
	"github.com/cockroachdb/cockroach/pkg/sql/sessiondatapb"
	"github.com/cockroachdb/cockroach/pkg/sql/sqltelemetry"
	"github.com/cockroachdb/cockroach/pkg/sql/types"
	"github.com/cockroachdb/errors"
	"github.com/lib/pq/oid"
)

type completionMsgType int

const (
	_ completionMsgType = iota
	commandComplete
	bindComplete
	closeComplete
	parseComplete
	emptyQueryResponse
	readyForQuery
	flush
	// Some commands, like Describe, don't need a completion message.
	noCompletionMsg
)

// commandResult is an implementation of sql.CommandResult that streams a
// commands results over a pgwire network connection.
type commandResult struct {
	// conn is the parent connection of this commandResult.
	conn *conn
	// conv and location indicate the conversion settings for SQL values.
	conv     sessiondatapb.DataConversionConfig
	location *time.Location
	// pos identifies the position of the command within the connection.
	pos sql.CmdPos

	// buffer contains items that are sent before the connection is closed.
	buffer struct {
		notices            []pgnotice.Notice
		paramStatusUpdates []paramStatusUpdate
	}

	err error
	// errExpected, if set, enforces that an error had been set when Close is
	// called.
	errExpected bool

	typ completionMsgType
	// If typ == commandComplete, this is the tag to be written in the
	// CommandComplete message.
	cmdCompleteTag string

	stmtType tree.StatementReturnType
	descOpt  sql.RowDescOpt

	// rowsAffected doesn't reflect the number of changed rows for bulk job
	// (IMPORT, BACKUP and RESTORE). For these jobs, see the usages of
	// log/logutil.LogJobCompletion().
	rowsAffected int

	// formatCodes describes the encoding of each column of result rows. It is nil
	// for statements not returning rows (or for results for commands other than
	// executing statements). It can also be nil for queries returning rows,
	// meaning that all columns will be encoded in the text format (this is the
	// case for queries executed through the simple protocol). Otherwise, it needs
	// to have an entry for every column.
	formatCodes []pgwirebase.FormatCode

	// types is a map from result column index to its type T, similar to formatCodes
	// (except types must always be set).
	types []*types.T

	// bufferingDisabled is conditionally set during planning of certain
	// statements.
	bufferingDisabled bool

	// released is set when the command result has been released so that its
	// memory can be reused. It is also used to assert against use-after-free
	// errors.
	released bool

	// bulkJobInfo stores id for bulk jobs (IMPORT, BACKUP, RESTORE),
	// It's written in commandResult.AddRow() and only if the query is
	// IMPORT, BACKUP or RESTORE.
	bulkJobId uint64
}

// paramStatusUpdate is a status update to send to the client when a parameter is
// updated.
type paramStatusUpdate struct {
	// param is the parameter name.
	param string
	// val is the new value of the given parameter.
	val string
}

var _ sql.CommandResult = &commandResult{}

// Close is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) Close(ctx context.Context, t sql.TransactionStatusIndicator) {
	r.assertNotReleased()
	defer r.release()
	if r.errExpected && r.err == nil {
		panic("expected err to be set on result by Close, but wasn't")
	}

	r.conn.writerState.fi.registerCmd(r.pos)
	if r.err != nil {
		r.conn.bufferErr(ctx, r.err)
		// Sync is the only client message that results in ReadyForQuery, and it
		// must *always* result in ReadyForQuery, even if there are errors during
		// Sync.
		if r.typ != readyForQuery {
			return
		}
	}

	for _, notice := range r.buffer.notices {
		if err := r.conn.bufferNotice(ctx, notice); err != nil {
			panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err when sending notice"))
		}
	}

	// Send a completion message, specific to the type of result.
	switch r.typ {
	case commandComplete:
		tag := cookTag(
			r.cmdCompleteTag, r.conn.writerState.tagBuf[:0], r.stmtType, r.rowsAffected,
		)
		r.conn.bufferCommandComplete(tag)
	case parseComplete:
		r.conn.bufferParseComplete()
	case bindComplete:
		r.conn.bufferBindComplete()
	case closeComplete:
		r.conn.bufferCloseComplete()
	case readyForQuery:
		r.conn.bufferReadyForQuery(byte(t))
		// The error is saved on conn.err.
		_ /* err */ = r.conn.Flush(r.pos)
		r.conn.maybeReallocate()
	case emptyQueryResponse:
		r.conn.bufferEmptyQueryResponse()
	case flush:
		// The error is saved on conn.err.
		_ /* err */ = r.conn.Flush(r.pos)
		r.conn.maybeReallocate()
	case noCompletionMsg:
		// nothing to do
	default:
		panic(errors.AssertionFailedf("unknown type: %v", r.typ))
	}

	for _, paramStatusUpdate := range r.buffer.paramStatusUpdates {
		if err := r.conn.bufferParamStatus(
			paramStatusUpdate.param,
			paramStatusUpdate.val,
		); err != nil {
			panic(
				errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err when sending parameter status update"),
			)
		}
	}
}

// Discard is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) Discard() {
	r.assertNotReleased()
	defer r.release()
}

// Err is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) Err() error {
	r.assertNotReleased()
	return r.err
}

// SetError is part of the sql.RestrictedCommandResult interface.
//
// We're not going to write any bytes to the buffer in order to support future
// SetError() calls. The error will only be serialized at Close() time.
func (r *commandResult) SetError(err error) {
	r.assertNotReleased()
	r.err = err
}

// beforeAdd should be called before rows are buffered.
func (r *commandResult) beforeAdd() error {
	r.assertNotReleased()
	if r.err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(r.err, "can't call AddRow after having set error"))
	}
	r.conn.writerState.fi.registerCmd(r.pos)
	if err := r.conn.GetErr(); err != nil {
		return err
	}
	if r.err != nil {
		panic("can't send row after error")
	}
	return nil
}

// JobIdColIdx is based on jobs.BulkJobExecutionResultHeader and
// jobs.DetachedJobExecutionResultHeader.
var JobIdColIdx int

// AddRow is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) AddRow(ctx context.Context, row tree.Datums) error {
	if err := r.beforeAdd(); err != nil {
		return err
	}
	switch r.cmdCompleteTag {
	case tree.ImportTag, tree.RestoreTag, tree.BackupTag:
		r.bulkJobId = uint64(*row[JobIdColIdx].(*tree.DInt))
	default:
		r.rowsAffected++
	}
	return r.conn.bufferRow(ctx, row, r)
}

// AddBatch is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) AddBatch(ctx context.Context, batch coldata.Batch) error {
	if err := r.beforeAdd(); err != nil {
		return err
	}
	switch r.cmdCompleteTag {
	case tree.ImportTag, tree.RestoreTag, tree.BackupTag:
		r.bulkJobId = uint64(batch.ColVec(JobIdColIdx).Int64()[0])
	default:
		r.rowsAffected += batch.Length()
	}
	return r.conn.bufferBatch(ctx, batch, r)
}

// SupportsAddBatch is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) SupportsAddBatch() bool {
	return true
}

// DisableBuffering is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) DisableBuffering() {
	r.assertNotReleased()
	r.bufferingDisabled = true
}

// BufferParamStatusUpdate is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) BufferParamStatusUpdate(param string, val string) {
	r.buffer.paramStatusUpdates = append(
		r.buffer.paramStatusUpdates,
		paramStatusUpdate{param: param, val: val},
	)
}

// BufferNotice is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) BufferNotice(notice pgnotice.Notice) {
	r.buffer.notices = append(r.buffer.notices, notice)
}

// SetColumns is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) SetColumns(ctx context.Context, cols colinfo.ResultColumns) {
	r.assertNotReleased()
	r.conn.writerState.fi.registerCmd(r.pos)
	if r.descOpt == sql.NeedRowDesc {
		_ /* err */ = r.conn.writeRowDescription(ctx, cols, r.formatCodes, &r.conn.writerState.buf)
	}
	r.types = make([]*types.T, len(cols))
	for i, col := range cols {
		r.types[i] = col.Typ
	}
}

// SetInferredTypes is part of the sql.DescribeResult interface.
func (r *commandResult) SetInferredTypes(types []oid.Oid) {
	r.assertNotReleased()
	r.conn.writerState.fi.registerCmd(r.pos)
	r.conn.bufferParamDesc(types)
}

// SetNoDataRowDescription is part of the sql.DescribeResult interface.
func (r *commandResult) SetNoDataRowDescription() {
	r.assertNotReleased()
	r.conn.writerState.fi.registerCmd(r.pos)
	r.conn.bufferNoDataMsg()
}

// SetPrepStmtOutput is part of the sql.DescribeResult interface.
func (r *commandResult) SetPrepStmtOutput(ctx context.Context, cols colinfo.ResultColumns) {
	r.assertNotReleased()
	r.conn.writerState.fi.registerCmd(r.pos)
	_ /* err */ = r.conn.writeRowDescription(ctx, cols, nil /* formatCodes */, &r.conn.writerState.buf)
}

// SetPortalOutput is part of the sql.DescribeResult interface.
func (r *commandResult) SetPortalOutput(
	ctx context.Context, cols colinfo.ResultColumns, formatCodes []pgwirebase.FormatCode,
) {
	r.assertNotReleased()
	r.conn.writerState.fi.registerCmd(r.pos)
	_ /* err */ = r.conn.writeRowDescription(ctx, cols, formatCodes, &r.conn.writerState.buf)
}

// IncrementRowsAffected is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) IncrementRowsAffected(ctx context.Context, n int) {
	r.assertNotReleased()
	r.rowsAffected += n
}

// RowsAffected is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) RowsAffected() int {
	r.assertNotReleased()
	return r.rowsAffected
}

// ResetStmtType is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) ResetStmtType(stmt tree.Statement) {
	r.assertNotReleased()
	r.stmtType = stmt.StatementReturnType()
	r.cmdCompleteTag = stmt.StatementTag()
}

// GetEntryFromExtraInfo is part of the sql.RestrictedCommandResult interface.
func (r *commandResult) GetBulkJobId() uint64 {
	return r.bulkJobId
}

// release frees the commandResult and allows its memory to be reused.
func (r *commandResult) release() {
	*r = commandResult{released: true}
}

// assertNotReleased asserts that the commandResult is not being used after
// being freed by one of the methods in the CommandResultClose interface. The
// assertion can have false negatives, where it fails to detect a use-after-free
// condition, but will never result in a false positive.
func (r *commandResult) assertNotReleased() {
	if r.released {
		panic("commandResult used after being released")
	}
}

func (c *conn) allocCommandResult() *commandResult {
	r := &c.res
	if r.released {
		r.released = false
	} else {
		// In practice, each conn only ever uses a single commandResult at a
		// time, so we could make this panic. However, doing so would entail
		// complicating the ClientComm interface, which doesn't seem worth it.
		r = new(commandResult)
	}
	return r
}

func (c *conn) newCommandResult(
	descOpt sql.RowDescOpt,
	pos sql.CmdPos,
	stmt tree.Statement,
	formatCodes []pgwirebase.FormatCode,
	conv sessiondatapb.DataConversionConfig,
	location *time.Location,
	limit int,
	portalName string,
	implicitTxn bool,
	portalPausability sql.PortalPausablity,
) sql.CommandResult {
	r := c.allocCommandResult()
	*r = commandResult{
		conn:           c,
		conv:           conv,
		location:       location,
		pos:            pos,
		typ:            commandComplete,
		cmdCompleteTag: stmt.StatementTag(),
		stmtType:       stmt.StatementReturnType(),
		descOpt:        descOpt,
		formatCodes:    formatCodes,
	}
	if limit == 0 {
		return r
	}
	telemetry.Inc(sqltelemetry.PortalWithLimitRequestCounter)
	return &limitedCommandResult{
		limit:            limit,
		portalName:       portalName,
		implicitTxn:      implicitTxn,
		commandResult:    r,
		portalPausablity: portalPausability,
	}
}

func (c *conn) newMiscResult(pos sql.CmdPos, typ completionMsgType) *commandResult {
	r := c.allocCommandResult()
	*r = commandResult{
		conn: c,
		pos:  pos,
		typ:  typ,
	}
	return r
}

// limitedCommandResult is a commandResult that has a limit, after which calls
// to AddRow will block until the associated client connection asks for more
// rows. It essentially implements the "execute portal with limit" part of the
// Postgres protocol.
type limitedCommandResult struct {
	*commandResult
	portalName  string
	implicitTxn bool

	seenTuples int
	// If set, an error will be sent to the client if more rows are produced than
	// this limit.
	limit            int
	reachedLimit     bool
	portalPausablity sql.PortalPausablity
}

// AddRow is part of the sql.RestrictedCommandResult interface.
func (r *limitedCommandResult) AddRow(ctx context.Context, row tree.Datums) error {
	if err := r.commandResult.AddRow(ctx, row); err != nil {
		return err
	}
	r.seenTuples++

	if r.seenTuples == r.limit {
		// If we've seen up to the limit of rows, send a "portal suspended" message
		// and wait for another exec portal message.
		r.conn.bufferPortalSuspended()
		if err := r.conn.Flush(r.pos); err != nil {
			return err
		}
		if r.portalPausablity == sql.PausablePortal {
			r.reachedLimit = true
			return sql.ErrPortalLimitHasBeenReached
		} else {
			// TODO(janexing): we keep using the logic from before we added
			// multiple-active-portals support to avoid bring too many bugs. Eventually
			// we should remove them and use the "return the control to connExecutor"
			// logic for all portals.
			r.seenTuples = 0
			return r.moreResultsNeeded(ctx)
		}
	}
	return nil
}

// SupportsAddBatch is part of the sql.RestrictedCommandResult interface.
// TODO(yuzefovich): implement limiting behavior for AddBatch.
func (r *limitedCommandResult) SupportsAddBatch() bool {
	return false
}

// moreResultsNeeded is a restricted connection handler that waits for more
// requests for rows from the active portal, during the "execute portal" flow
// when a limit has been specified.
func (r *limitedCommandResult) moreResultsNeeded(ctx context.Context) error {
	errBasedOnPausability := func(pausablity sql.PortalPausablity) error {
		switch pausablity {
		case sql.PortalPausabilityDisabled:
			return sql.ErrLimitedResultNotSupported
		case sql.NotPausablePortalForUnsupportedStmt:
			return sql.ErrStmtNotSupportedForPausablePortal
		default:
			return errors.AssertionFailedf("unsupported pausability type for a portal")
		}
	}

	// Keep track of the previous CmdPos so we can rewind if needed.
	prevPos := r.conn.stmtBuf.AdvanceOne()
	for {
		cmd, curPos, err := r.conn.stmtBuf.CurCmd()
		if err != nil {
			return err
		}
		switch c := cmd.(type) {
		case sql.DeletePreparedStmt:
			// The client wants to close a portal or statement. We support the case
			// where it is exactly this portal. We are in effect peeking to see if the
			// next message is a delete portal.
			if c.Type != pgwirebase.PreparePortal || c.Name != r.portalName {
				telemetry.Inc(sqltelemetry.InterleavedPortalRequestCounter)
				return errors.WithDetail(errBasedOnPausability(r.portalPausablity),
					"cannot close a portal while a different one is open")
			}
			return r.rewindAndClosePortal(ctx, prevPos)
		case sql.ExecPortal:
			// The happy case: the client wants more rows from the portal.
			if c.Name != r.portalName {
				telemetry.Inc(sqltelemetry.InterleavedPortalRequestCounter)
				return errors.WithDetail(errBasedOnPausability(r.portalPausablity),
					"cannot execute a portal while a different one is open")
			}
			r.limit = c.Limit
			// In order to get the correct command tag, we need to reset the seen rows.
			r.rowsAffected = 0
			return nil
		case sql.Sync:
			if r.implicitTxn {
				// Implicit transactions should treat a Sync as an auto-commit. This
				// needs to be handled in conn_executor.
				return r.rewindAndClosePortal(ctx, prevPos)
			}
			// The client wants to see a ready for query message
			// back. Send it then run the for loop again.
			r.conn.stmtBuf.AdvanceOne()
			// Trim old statements to reclaim memory. We need to perform this clean up
			// here as the conn_executor cleanup is not executed because of the
			// limitedCommandResult side state machine.
			r.conn.stmtBuf.Ltrim(ctx, prevPos)
			// We can hard code InTxnBlock here because implicit transactions are
			// handled above.
			r.conn.bufferReadyForQuery(byte(sql.InTxnBlock))
			if err := r.conn.Flush(r.pos); err != nil {
				return err
			}
		case sql.Flush:
			// Flush has no client response, so just advance the position and flush
			// any existing results.
			r.conn.stmtBuf.AdvanceOne()
			if err := r.conn.Flush(r.pos); err != nil {
				return err
			}
		default:
			// If the portal is immediately followed by a COMMIT, we can proceed and
			// let the portal be destroyed at the end of the transaction.
			if isCommit, err := r.isCommit(); err != nil {
				return err
			} else if isCommit {
				return r.rewindAndClosePortal(ctx, prevPos)
			}
			// We got some other message, but we only support executing to completion.
			telemetry.Inc(sqltelemetry.InterleavedPortalRequestCounter)
			return errors.WithDetail(errBasedOnPausability(r.portalPausablity),
				fmt.Sprintf("cannot perform operation %T while a different portal is open", c))
		}
		prevPos = curPos
	}
}

// isCommit checks if the statement buffer has a COMMIT at the current
// position. It may either be (1) a COMMIT in the simple protocol, or (2) a
// Parse/Bind/Execute sequence for a COMMIT query.
func (r *limitedCommandResult) isCommit() (bool, error) {
	cmd, _, err := r.conn.stmtBuf.CurCmd()
	if err != nil {
		return false, err
	}
	// Case 1: Check if cmd is a simple COMMIT statement.
	if execStmt, ok := cmd.(sql.ExecStmt); ok {
		if _, isCommit := execStmt.AST.(*tree.CommitTransaction); isCommit {
			return true, nil
		}
	}

	commitStmtName := ""
	commitPortalName := ""
	// Case 2a: Check if cmd is a prepared COMMIT statement.
	if prepareStmt, ok := cmd.(sql.PrepareStmt); ok {
		if _, isCommit := prepareStmt.AST.(*tree.CommitTransaction); isCommit {
			commitStmtName = prepareStmt.Name
		} else {
			return false, nil
		}
	} else {
		return false, nil
	}

	r.conn.stmtBuf.AdvanceOne()
	cmd, _, err = r.conn.stmtBuf.CurCmd()
	if err != nil {
		return false, err
	}
	// Case 2b: The next cmd must be a bind command.
	if bindStmt, ok := cmd.(sql.BindStmt); ok {
		// This bind command must be for the COMMIT statement that we just saw.
		if bindStmt.PreparedStatementName == commitStmtName {
			commitPortalName = bindStmt.PortalName
		} else {
			return false, nil
		}
	} else {
		return false, nil
	}

	r.conn.stmtBuf.AdvanceOne()
	cmd, _, err = r.conn.stmtBuf.CurCmd()
	if err != nil {
		return false, err
	}
	// Case 2c: The next cmd must be an exec portal command.
	if execPortal, ok := cmd.(sql.ExecPortal); ok {
		// This exec command must be for the portal that was just bound.
		if execPortal.Name == commitPortalName {
			return true, nil
		}
	}
	return false, nil
}

// rewindAndClosePortal closes the portal in the same way implicit transactions
// do, but also rewinds the stmtBuf to still point to the portal close so that
// the state machine can do its part of the cleanup.
func (r *limitedCommandResult) rewindAndClosePortal(
	ctx context.Context, rewindTo sql.CmdPos,
) error {
	// Don't send an CommandComplete for the portal; it got suspended.
	r.typ = noCompletionMsg
	// Rewind to before the delete so the AdvanceOne in connExecutor.execCmd ends
	// up back on it.
	r.conn.stmtBuf.Rewind(ctx, rewindTo)
	return sql.ErrLimitedResultClosed
}

func (r *limitedCommandResult) Close(ctx context.Context, t sql.TransactionStatusIndicator) {
	if r.reachedLimit {
		r.commandResult.typ = noCompletionMsg
	}
	r.commandResult.Close(ctx, t)
}

// Err is part of the sql.RestrictedCommandResult interface.
func (r *limitedCommandResult) Err() error {
	return r.err
}

// SetError is part of the sql.RestrictedCommandResult interface.
func (r *limitedCommandResult) SetError(err error) {
	r.err = err
}

// Get the column index for job id based on the result header defined in
// jobs.BulkJobExecutionResultHeader and jobs.DetachedJobExecutionResultHeader.
func init() {
	jobIdIdxInBulkJobExecutionResultHeader := -1
	jobIdIdxInDetachedJobExecutionResultHeader := -1
	for i, col := range jobs.BulkJobExecutionResultHeader {
		if col.Name == "job_id" {
			jobIdIdxInBulkJobExecutionResultHeader = i
			break
		}
	}
	if jobIdIdxInBulkJobExecutionResultHeader == -1 {
		panic("cannot find the job id column in BulkJobExecutionResultHeader")
	}

	for i, col := range jobs.DetachedJobExecutionResultHeader {
		if col.Name == "job_id" {
			if i != jobIdIdxInBulkJobExecutionResultHeader {
				panic("column index of job_id in DetachedJobExecutionResultHeader and" +
					" BulkJobExecutionResultHeader should be the same")
			} else {
				jobIdIdxInDetachedJobExecutionResultHeader = i
				break
			}
		}
	}
	if jobIdIdxInDetachedJobExecutionResultHeader == -1 {
		panic("cannot find the job id column in DetachedJobExecutionResultHeader")
	}
	JobIdColIdx = jobIdIdxInBulkJobExecutionResultHeader
}