conn.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 (
	"bufio"
	"bytes"
	"context"
	"fmt"
	"io"
	"net"
	"strconv"
	"sync"
	"sync/atomic"
	"time"

	"github.com/cockroachdb/cockroach/pkg/col/coldata"
	"github.com/cockroachdb/cockroach/pkg/server/telemetry"
	"github.com/cockroachdb/cockroach/pkg/settings"
	"github.com/cockroachdb/cockroach/pkg/sql"
	"github.com/cockroachdb/cockroach/pkg/sql/catalog/colinfo"
	"github.com/cockroachdb/cockroach/pkg/sql/parser"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgcode"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgerror"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgnotice"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgwirebase"
	"github.com/cockroachdb/cockroach/pkg/sql/pgwire/pgwirecancel"
	"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/cockroach/pkg/util/envutil"
	"github.com/cockroachdb/cockroach/pkg/util/log"
	"github.com/cockroachdb/cockroach/pkg/util/log/eventpb"
	"github.com/cockroachdb/cockroach/pkg/util/mon"
	"github.com/cockroachdb/cockroach/pkg/util/netutil"
	"github.com/cockroachdb/cockroach/pkg/util/ring"
	"github.com/cockroachdb/cockroach/pkg/util/timeutil"
	"github.com/cockroachdb/cockroach/pkg/util/tracing"
	"github.com/cockroachdb/errors"
	"github.com/cockroachdb/logtags"
	"github.com/cockroachdb/redact"
	"github.com/lib/pq/oid"
	"go.opentelemetry.io/otel/attribute"
)

// conn implements a pgwire network connection (version 3 of the protocol,
// implemented by Postgres v7.4 and later). conn.serve() reads protocol
// messages, transforms them into commands that it pushes onto a StmtBuf (where
// they'll be picked up and executed by the connExecutor).
// The connExecutor produces results for the commands, which are delivered to
// the client through the sql.ClientComm interface, implemented by this conn
// (code is in command_result.go).
type conn struct {
	conn net.Conn

	sessionArgs sql.SessionArgs
	metrics     *ServerMetrics

	// startTime is the time when the connection attempt was first received
	// by the server.
	startTime time.Time

	// rd is a buffered reader consuming conn. All reads from conn go through
	// this.
	rd bufio.Reader

	// parser is used to avoid allocating a parser each time.
	parser parser.Parser

	// stmtBuf is populated with commands queued for execution by this conn.
	stmtBuf sql.StmtBuf

	// res is used to avoid allocations in the conn's ClientComm implementation.
	res commandResult

	// err is an error, accessed atomically. It represents any error encountered
	// while accessing the underlying network connection. This can read via
	// GetErr() by anybody. If it is found to be != nil, the conn is no longer to
	// be used.
	err atomic.Value

	// writerState groups together all aspects of the write-side state of the
	// connection.
	writerState struct {
		fi flushInfo
		// buf contains command results (rows, etc.) until they're flushed to the
		// network connection.
		buf    bytes.Buffer
		tagBuf [64]byte
	}

	readBuf    pgwirebase.ReadBuffer
	msgBuilder writeBuffer

	// vecsScratch is a scratch space used by bufferBatch.
	vecsScratch coldata.TypedVecs

	sv *settings.Values

	// alwaysLogAuthActivity is used force-enables logging of authn events.
	alwaysLogAuthActivity bool

	// afterReadMsgTestingKnob is called after reading every message.
	afterReadMsgTestingKnob func(context.Context) error
}

// serveConn creates a conn that will serve the netConn. It returns once the
// network connection is closed.
//
// Internally, a connExecutor will be created to execute commands. Commands read
// from the network are buffered in a stmtBuf which is consumed by the
// connExecutor. The connExecutor produces results which are buffered and
// sometimes synchronously flushed to the network.
//
// The reader goroutine (this one) outlives the connExecutor's goroutine (the
// "processor goroutine").
// However, they can both signal each other to stop. Here's how the different
// cases work:
// 1) The reader receives a ClientMsgTerminate protocol packet: the reader
// closes the stmtBuf and also cancels the command processing context. These
// actions will prompt the command processor to finish.
// 2) The reader gets a read error from the network connection: like above, the
// reader closes the command processor.
// 3) The reader's context is canceled (happens when the server is draining but
// the connection was busy and hasn't quit yet): the reader notices the canceled
// context and, like above, closes the processor.
// 4) The processor encounters an error. This error can come from various fatal
// conditions encountered internally by the processor, or from a network
// communication error encountered while flushing results to the network.
// The processor will cancel the reader's context and terminate.
// Note that query processing errors are different; they don't cause the
// termination of the connection.
//
// Draining notes:
//
// The reader notices that the server is draining by polling the IsDraining
// closure passed to serveImpl. At that point, the reader delegates the
// responsibility of closing the connection to the statement processor: it will
// push a DrainRequest to the stmtBuf which signals the processor to quit ASAP.
// The processor will quit immediately upon seeing that command if it's not
// currently in a transaction. If it is in a transaction, it will wait until the
// first time a Sync command is processed outside of a transaction - the logic
// being that we want to stop when we're both outside transactions and outside
// batches.
func (s *Server) serveConn(
	ctx context.Context,
	netConn net.Conn,
	sArgs sql.SessionArgs,
	reserved mon.BoundAccount,
	connStart time.Time,
	authOpt authOptions,
) {
	if log.V(2) {
		log.Infof(ctx, "new connection with options: %+v", sArgs)
	}

	c := newConn(netConn, sArgs, &s.metrics, connStart, &s.execCfg.Settings.SV)
	c.alwaysLogAuthActivity = alwaysLogAuthActivity || atomic.LoadInt32(&s.testingAuthLogEnabled) > 0
	if s.execCfg.PGWireTestingKnobs != nil {
		c.afterReadMsgTestingKnob = s.execCfg.PGWireTestingKnobs.AfterReadMsgTestingKnob
	}

	// Do the reading of commands from the network.
	c.serveImpl(ctx, s.IsDraining, s.SQLServer, reserved, authOpt)
}

// alwaysLogAuthActivity makes it possible to unconditionally enable
// authentication logging when cluster settings do not work reliably,
// e.g. in multi-tenant setups in v20.2. This override mechanism
// can be removed after all of CC is moved to use v21.1 or a version
// which supports cluster settings.
var alwaysLogAuthActivity = envutil.EnvOrDefaultBool("COCKROACH_ALWAYS_LOG_AUTHN_EVENTS", false)

func newConn(
	netConn net.Conn,
	sArgs sql.SessionArgs,
	metrics *ServerMetrics,
	connStart time.Time,
	sv *settings.Values,
) *conn {
	c := &conn{
		conn:        netConn,
		sessionArgs: sArgs,
		metrics:     metrics,
		startTime:   connStart,
		rd:          *bufio.NewReader(netConn),
		sv:          sv,
		readBuf:     pgwirebase.MakeReadBuffer(pgwirebase.ReadBufferOptionWithClusterSettings(sv)),
	}
	c.stmtBuf.Init()
	c.res.released = true
	c.writerState.fi.buf = &c.writerState.buf
	c.writerState.fi.lastFlushed = -1
	c.msgBuilder.init(metrics.BytesOutCount)

	return c
}

func (c *conn) setErr(err error) {
	c.err.Store(err)
}

func (c *conn) GetErr() error {
	err := c.err.Load()
	if err != nil {
		return err.(error)
	}
	return nil
}

func (c *conn) sendError(ctx context.Context, execCfg *sql.ExecutorConfig, err error) error {
	// We could, but do not, report server-side network errors while
	// trying to send the client error. This is because clients that
	// receive error payload are highly correlated with clients
	// disconnecting abruptly.
	_ /* err */ = writeErr(ctx, &execCfg.Settings.SV, err, &c.msgBuilder, c.conn)
	return err
}

func (c *conn) checkMaxConnections(ctx context.Context, sqlServer *sql.Server) error {
	if c.sessionArgs.IsSuperuser {
		// This user is a super user and is therefore not affected by connection limits.
		sqlServer.IncrementConnectionCount()
		return nil
	}

	maxNumConnectionsValue := maxNumConnections.Get(&sqlServer.GetExecutorConfig().Settings.SV)
	if maxNumConnectionsValue < 0 {
		// Unlimited connections are allowed.
		sqlServer.IncrementConnectionCount()
		return nil
	}
	if !sqlServer.IncrementConnectionCountIfLessThan(maxNumConnectionsValue) {
		return c.sendError(ctx, sqlServer.GetExecutorConfig(), errors.WithHintf(
			pgerror.New(pgcode.TooManyConnections, "sorry, too many clients already"),
			"the maximum number of allowed connections is %d and can be modified using the %s config key",
			maxNumConnectionsValue,
			maxNumConnections.Key(),
		))
	}
	return nil
}

func (c *conn) authLogEnabled() bool {
	return c.alwaysLogAuthActivity || logSessionAuth.Get(c.sv)
}

// maxRepeatedErrorCount is the number of times an error can be received
// while reading from the network connection before the server decides to give
// up and abort the connection.
const maxRepeatedErrorCount = 1 << 15

// serveImpl continuously reads from the network connection and pushes execution
// instructions into a sql.StmtBuf, from where they'll be processed by a command
// "processor" goroutine (a connExecutor).
// The method returns when the pgwire termination message is received, when
// network communication fails, when the server is draining or when ctx is
// canceled (which also happens when draining (but not from the get-go), and
// when the processor encounters a fatal error).
//
// serveImpl always closes the network connection before returning.
//
// sqlServer is used to create the command processor. As a special facility for
// tests, sqlServer can be nil, in which case the command processor and the
// write-side of the connection will not be created.
func (c *conn) serveImpl(
	ctx context.Context,
	draining func() bool,
	sqlServer *sql.Server,
	reserved mon.BoundAccount,
	authOpt authOptions,
) {
	defer func() { _ = c.conn.Close() }()

	if c.sessionArgs.User.IsRootUser() || c.sessionArgs.User.IsNodeUser() {
		ctx = logtags.AddTag(ctx, "user", redact.Safe(c.sessionArgs.User))
	} else {
		ctx = logtags.AddTag(ctx, "user", c.sessionArgs.User)
	}
	tracing.SpanFromContext(ctx).SetTag("user", attribute.StringValue(c.sessionArgs.User.Normalized()))

	inTestWithoutSQL := sqlServer == nil
	if !inTestWithoutSQL {
		sessionStart := timeutil.Now()
		defer func() {
			if c.authLogEnabled() {
				endTime := timeutil.Now()
				ev := &eventpb.ClientSessionEnd{
					CommonEventDetails:      eventpb.CommonEventDetails{Timestamp: endTime.UnixNano()},
					CommonConnectionDetails: authOpt.connDetails,
					Duration:                endTime.Sub(sessionStart).Nanoseconds(),
				}
				log.StructuredEvent(ctx, ev)
			}
		}()
	}

	// NOTE: We're going to write a few messages to the connection in this method,
	// for the handshake. After that, all writes are done async, in the
	// startWriter() goroutine.

	ctx, cancelConn := context.WithCancel(ctx)
	defer cancelConn() // This calms the linter that wants these callbacks to always be called.

	var sentDrainSignal bool
	// The net.Conn is switched to a conn that exits if the ctx is canceled.
	c.conn = NewReadTimeoutConn(c.conn, func() error {
		// If the context was canceled, it's time to stop reading. Either a
		// higher-level server or the command processor have canceled us.
		if ctx.Err() != nil {
			return ctx.Err()
		}
		// If the server is draining, we'll let the processor know by pushing a
		// DrainRequest. This will make the processor quit whenever it finds a good
		// time.
		if !sentDrainSignal && draining() {
			_ /* err */ = c.stmtBuf.Push(ctx, sql.DrainRequest{})
			sentDrainSignal = true
		}
		return nil
	})
	c.rd = *bufio.NewReader(c.conn)

	// the authPipe below logs authentication messages iff its auth
	// logger is non-nil. We define this here.
	logAuthn := !inTestWithoutSQL && c.authLogEnabled()

	// We'll build an authPipe to communicate with the authentication process.
	authPipe := newAuthPipe(c, logAuthn, authOpt, c.sessionArgs.User)
	var authenticator authenticatorIO = authPipe

	// procCh is the channel on which we'll receive the termination signal from
	// the command processor.
	var procCh <-chan error

	// We need a value for the unqualified int size here, but it is controlled
	// by a session variable, and this layer doesn't have access to the session
	// data. The callback below is called whenever default_int_size changes.
	// It happens in a different goroutine, so it has to be changed atomically.
	var atomicUnqualifiedIntSize = new(int32)
	onDefaultIntSizeChange := func(newSize int32) {
		atomic.StoreInt32(atomicUnqualifiedIntSize, newSize)
	}

	if sqlServer != nil {
		// Spawn the command processing goroutine, which also handles connection
		// authentication). It will notify us when it's done through procCh, and
		// we'll also interact with the authentication process through ac.
		var ac AuthConn = authPipe
		procCh = c.processCommandsAsync(
			ctx,
			authOpt,
			ac,
			sqlServer,
			reserved,
			cancelConn,
			onDefaultIntSizeChange,
		)
	} else {
		// sqlServer == nil means we are in a local test. In this case
		// we only need the minimum to make pgx happy.
		var err error
		for param, value := range testingStatusReportParams {
			err = c.sendParamStatus(param, value)
			if err != nil {
				break
			}
		}
		if err != nil {
			reserved.Close(ctx)
			return
		}
		var ac AuthConn = authPipe
		// Simulate auth succeeding.
		ac.AuthOK(ctx)
		dummyCh := make(chan error)
		close(dummyCh)
		procCh = dummyCh

		if err := c.sendReadyForQuery(0 /* queryCancelKey */); err != nil {
			reserved.Close(ctx)
			return
		}
	}

	var terminateSeen bool
	var authDone, ignoreUntilSync bool
	var repeatedErrorCount int
	for {
		breakLoop, isSimpleQuery, err := func() (bool, bool, error) {
			typ, n, err := c.readBuf.ReadTypedMsg(&c.rd)
			c.metrics.BytesInCount.Inc(int64(n))
			if err == nil && c.afterReadMsgTestingKnob != nil {
				err = c.afterReadMsgTestingKnob(ctx)
			}
			isSimpleQuery := typ == pgwirebase.ClientMsgSimpleQuery
			if err != nil {
				if pgwirebase.IsMessageTooBigError(err) {
					log.VInfof(ctx, 1, "pgwire: found big error message; attempting to slurp bytes and return error: %s", err)

					// Slurp the remaining bytes.
					slurpN, slurpErr := c.readBuf.SlurpBytes(&c.rd, pgwirebase.GetMessageTooBigSize(err))
					c.metrics.BytesInCount.Inc(int64(slurpN))
					if slurpErr != nil {
						return false, isSimpleQuery, errors.Wrap(slurpErr, "pgwire: error slurping remaining bytes")
					}
				}

				// Write out the error over pgwire.
				if err := c.stmtBuf.Push(ctx, sql.SendError{Err: err}); err != nil {
					return false, isSimpleQuery, errors.New("pgwire: error writing too big error message to the client")
				}

				// If this is a simple query, we have to send the sync message back as
				// well.
				if isSimpleQuery {
					if err := c.stmtBuf.Push(ctx, sql.Sync{}); err != nil {
						return false, isSimpleQuery, errors.New("pgwire: error writing sync to the client whilst message is too big")
					}
				}

				// We need to continue processing here for pgwire clients to be able to
				// successfully read the error message off pgwire.
				//
				// If break here, we terminate the connection. The client will instead see that
				// we terminated the connection prematurely (as opposed to seeing a ClientMsgTerminate
				// packet) and instead return a broken pipe or io.EOF error message.
				return false, isSimpleQuery, errors.Wrap(err, "pgwire: error reading input")
			}
			timeReceived := timeutil.Now()
			log.VEventf(ctx, 2, "pgwire: processing %s", typ)

			if ignoreUntilSync {
				if typ != pgwirebase.ClientMsgSync {
					log.VInfof(ctx, 1, "pgwire: skipping non-sync message after encountering error")
					return false, isSimpleQuery, nil
				}
				ignoreUntilSync = false
			}

			if !authDone {
				if typ == pgwirebase.ClientMsgPassword {
					var pwd []byte
					if pwd, err = c.readBuf.GetBytes(n - 4); err != nil {
						return false, isSimpleQuery, err
					}
					// Pass the data to the authenticator. This hopefully causes it to finish
					// authentication in the background and give us an intSizer when we loop
					// around.
					if err = authenticator.sendPwdData(pwd); err != nil {
						return false, isSimpleQuery, err
					}
					return false, isSimpleQuery, nil
				}
				// Wait for the auth result.
				if err = authenticator.authResult(); err != nil {
					// The error has already been sent to the client.
					return true, isSimpleQuery, nil //nolint:returnerrcheck
				}
				authDone = true
			}

			switch typ {
			case pgwirebase.ClientMsgPassword:
				// This messages are only acceptable during the auth phase, handled above.
				err = pgwirebase.NewProtocolViolationErrorf("unexpected authentication data")
				return true, isSimpleQuery, writeErr(
					ctx, &sqlServer.GetExecutorConfig().Settings.SV, err,
					&c.msgBuilder, &c.writerState.buf)
			case pgwirebase.ClientMsgSimpleQuery:
				if err = c.handleSimpleQuery(
					ctx, &c.readBuf, timeReceived, parser.NakedIntTypeFromDefaultIntSize(atomic.LoadInt32(atomicUnqualifiedIntSize)),
				); err != nil {
					return false, isSimpleQuery, err
				}
				return false, isSimpleQuery, c.stmtBuf.Push(ctx, sql.Sync{})

			case pgwirebase.ClientMsgExecute:
				// To support the 1PC txn fast path, we peek at the next command to
				// see if it is a Sync. This is because in the extended protocol, an
				// implicit transaction cannot commit until the Sync is seen. If there's
				// an error while peeking (for example, there are no bytes in the
				// buffer), the error is ignored since it will be handled on the next
				// loop iteration.
				followedBySync := false
				if nextMsgType, err := c.rd.Peek(1); err == nil &&
					pgwirebase.ClientMessageType(nextMsgType[0]) == pgwirebase.ClientMsgSync {
					followedBySync = true
				}
				return false, isSimpleQuery, c.handleExecute(ctx, &c.readBuf, timeReceived, followedBySync)

			case pgwirebase.ClientMsgParse:
				return false, isSimpleQuery, c.handleParse(ctx, &c.readBuf, parser.NakedIntTypeFromDefaultIntSize(atomic.LoadInt32(atomicUnqualifiedIntSize)))

			case pgwirebase.ClientMsgDescribe:
				return false, isSimpleQuery, c.handleDescribe(ctx, &c.readBuf)

			case pgwirebase.ClientMsgBind:
				return false, isSimpleQuery, c.handleBind(ctx, &c.readBuf)

			case pgwirebase.ClientMsgClose:
				return false, isSimpleQuery, c.handleClose(ctx, &c.readBuf)

			case pgwirebase.ClientMsgTerminate:
				terminateSeen = true
				return true, isSimpleQuery, nil

			case pgwirebase.ClientMsgSync:
				// We're starting a batch here. If the client continues using the extended
				// protocol and encounters an error, everything until the next sync
				// message has to be skipped. See:
				// https://www.postgresql.org/docs/current/10/protocol-flow.html#PROTOCOL-FLOW-EXT-QUERY

				return false, isSimpleQuery, c.stmtBuf.Push(ctx, sql.Sync{})

			case pgwirebase.ClientMsgFlush:
				return false, isSimpleQuery, c.handleFlush(ctx)

			case pgwirebase.ClientMsgCopyData, pgwirebase.ClientMsgCopyDone, pgwirebase.ClientMsgCopyFail:
				// We're supposed to ignore these messages, per the protocol spec. This
				// state will happen when an error occurs on the server-side during a copy
				// operation: the server will send an error and a ready message back to
				// the client, and must then ignore further copy messages. See:
				// https://github.com/postgres/postgres/blob/6e1dd2773eb60a6ab87b27b8d9391b756e904ac3/src/backend/tcop/postgres.c#L4295
				return false, isSimpleQuery, nil
			default:
				return false, isSimpleQuery, c.stmtBuf.Push(
					ctx,
					sql.SendError{Err: pgwirebase.NewUnrecognizedMsgTypeErr(typ)})
			}
		}()
		if err != nil {
			log.VEventf(ctx, 1, "pgwire: error processing message: %s", err)
			if !isSimpleQuery {
				// In the extended protocol, after seeing an error, we ignore all
				// messages until receiving a sync.
				ignoreUntilSync = true
			}
			repeatedErrorCount++
			// If we can't read data because of any one of the following conditions,
			// then we should break:
			// 1. the connection was closed.
			// 2. the context was canceled (e.g. during authentication).
			// 3. we reached an arbitrary threshold of repeated errors.
			if netutil.IsClosedConnection(err) ||
				errors.Is(err, context.Canceled) ||
				repeatedErrorCount > maxRepeatedErrorCount {
				break
			}
		} else {
			repeatedErrorCount = 0
		}
		if breakLoop {
			break
		}
	}

	// We're done reading data from the client, so make the communication
	// goroutine stop. Depending on what that goroutine is currently doing (or
	// blocked on), we cancel and close all the possible channels to make sure we
	// tickle it in the right way.

	// Signal command processing to stop. It might be the case that the processor
	// canceled our context and that's how we got here; in that case, this will
	// be a no-op.
	c.stmtBuf.Close()
	// Cancel the processor's context.
	cancelConn()
	// In case the authenticator is blocked on waiting for data from the client,
	// tell it that there's no more data coming. This is a no-op if authentication
	// was completed already.
	authenticator.noMorePwdData()

	// Wait for the processor goroutine to finish, if it hasn't already. We're
	// ignoring the error we get from it, as we have no use for it. It might be a
	// connection error, or a context cancelation error case this goroutine is the
	// one that triggered the execution to stop.
	<-procCh

	if terminateSeen {
		return
	}
	// If we're draining, let the client know by piling on an AdminShutdownError
	// and flushing the buffer.
	if draining() {
		// TODO(andrei): I think sending this extra error to the client if we also
		// sent another error for the last query (like a context canceled) is a bad
		// idea; see #22630. I think we should find a way to return the
		// AdminShutdown error as the only result of the query.
		log.Ops.Info(ctx, "closing existing connection while server is draining")
		_ /* err */ = writeErr(ctx, &sqlServer.GetExecutorConfig().Settings.SV,
			newAdminShutdownErr(ErrDrainingExistingConn), &c.msgBuilder, &c.writerState.buf)
		_ /* n */, _ /* err */ = c.writerState.buf.WriteTo(c.conn)
	}
}

// processCommandsAsync spawns a goroutine that authenticates the connection and
// then processes commands from c.stmtBuf.
//
// It returns a channel that will be signaled when this goroutine is done.
// Whatever error is returned on that channel has already been written to the
// client connection, if applicable.
//
// If authentication fails, this goroutine finishes and, as always, cancelConn
// is called.
//
// Args:
// ac: An interface used by the authentication process to receive password data
//   and to ultimately declare the authentication successful.
// reserved: Reserved memory. This method takes ownership.
// cancelConn: A function to be called when this goroutine exits. Its goal is to
//   cancel the connection's context, thus stopping the connection's goroutine.
//   The returned channel is also closed before this goroutine dies, but the
//   connection's goroutine is not expected to be reading from that channel
//   (instead, it's expected to always be monitoring the network connection).
func (c *conn) processCommandsAsync(
	ctx context.Context,
	authOpt authOptions,
	ac AuthConn,
	sqlServer *sql.Server,
	reserved mon.BoundAccount,
	cancelConn func(),
	onDefaultIntSizeChange func(newSize int32),
) <-chan error {
	// reservedOwned is true while we own reserved, false when we pass ownership
	// away.
	reservedOwned := true
	retCh := make(chan error, 1)
	go func() {
		var retErr error
		var connHandler sql.ConnectionHandler
		var authOK bool
		var connCloseAuthHandler func()
		defer func() {
			// Release resources, if we still own them.
			if reservedOwned {
				reserved.Close(ctx)
			}
			// Notify the connection's goroutine that we're terminating. The
			// connection might know already, as it might have triggered this
			// goroutine's finish, but it also might be us that we're triggering the
			// connection's death. This context cancelation serves to interrupt a
			// network read on the connection's goroutine.
			cancelConn()

			pgwireKnobs := sqlServer.GetExecutorConfig().PGWireTestingKnobs
			if pgwireKnobs != nil && pgwireKnobs.CatchPanics {
				if r := recover(); r != nil {
					// Catch the panic and return it to the client as an error.
					if err, ok := r.(error); ok {
						// Mask the cause but keep the details.
						retErr = errors.Handled(err)
					} else {
						retErr = errors.Newf("%+v", r)
					}
					retErr = pgerror.WithCandidateCode(retErr, pgcode.CrashShutdown)
					// Add a prefix. This also adds a stack trace.
					retErr = errors.Wrap(retErr, "caught fatal error")
					_ = writeErr(
						ctx, &sqlServer.GetExecutorConfig().Settings.SV, retErr,
						&c.msgBuilder, &c.writerState.buf)
					_ /* n */, _ /* err */ = c.writerState.buf.WriteTo(c.conn)
					c.stmtBuf.Close()
					// Send a ready for query to make sure the client can react.
					// TODO(andrei, jordan): Why are we sending this exactly?
					c.bufferReadyForQuery('I')
				}
			}
			if !authOK {
				ac.AuthFail(retErr)
			}
			if connCloseAuthHandler != nil {
				connCloseAuthHandler()
			}
			// Inform the connection goroutine of success or failure.
			retCh <- retErr
		}()

		// Authenticate the connection.
		if connCloseAuthHandler, retErr = c.handleAuthentication(
			ctx, ac, authOpt, sqlServer.GetExecutorConfig(),
		); retErr != nil {
			// Auth failed or some other error.
			return
		}

		if retErr = c.checkMaxConnections(ctx, sqlServer); retErr != nil {
			return
		}
		defer sqlServer.DecrementConnectionCount()

		if retErr = c.authOKMessage(); retErr != nil {
			return
		}

		// Inform the client of the default session settings.
		connHandler, retErr = c.sendInitialConnData(ctx, sqlServer, onDefaultIntSizeChange)
		if retErr != nil {
			return
		}
		// Signal the connection was established to the authenticator.
		ac.AuthOK(ctx)
		ac.LogAuthOK(ctx)

		// We count the connection establish latency until we are ready to
		// serve a SQL query. It includes the time it takes to authenticate and
		// send the initial ReadyForQuery message.
		duration := timeutil.Since(c.startTime).Nanoseconds()
		c.metrics.ConnLatency.RecordValue(duration)

		// Mark the authentication as succeeded in case a panic
		// is thrown below and we need to report to the client
		// using the defer above.
		authOK = true

		// Now actually process commands.
		reservedOwned = false // We're about to pass ownership away.
		retErr = sqlServer.ServeConn(ctx, connHandler, reserved, cancelConn)
	}()
	return retCh
}

func (c *conn) sendParamStatus(param, value string) error {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgParameterStatus)
	c.msgBuilder.writeTerminatedString(param)
	c.msgBuilder.writeTerminatedString(value)
	return c.msgBuilder.finishMsg(c.conn)
}

func (c *conn) bufferParamStatus(param, value string) error {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgParameterStatus)
	c.msgBuilder.writeTerminatedString(param)
	c.msgBuilder.writeTerminatedString(value)
	return c.msgBuilder.finishMsg(&c.writerState.buf)
}

func (c *conn) bufferNotice(ctx context.Context, noticeErr pgnotice.Notice) error {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgNoticeResponse)
	return writeErrFields(ctx, c.sv, noticeErr, &c.msgBuilder, &c.writerState.buf)
}

func (c *conn) sendInitialConnData(
	ctx context.Context, sqlServer *sql.Server, onDefaultIntSizeChange func(newSize int32),
) (sql.ConnectionHandler, error) {
	connHandler, err := sqlServer.SetupConn(
		ctx,
		c.sessionArgs,
		&c.stmtBuf,
		c,
		c.metrics.SQLMemMetrics,
		onDefaultIntSizeChange,
	)
	if err != nil {
		_ /* err */ = writeErr(
			ctx, &sqlServer.GetExecutorConfig().Settings.SV, err, &c.msgBuilder, c.conn)
		return sql.ConnectionHandler{}, err
	}

	// Send the initial "status parameters" to the client.  This
	// overlaps partially with session variables. The client wants to
	// see the values that result from the combination of server-side
	// defaults with client-provided values.
	// For details see: https://www.postgresql.org/docs/10/static/libpq-status.html
	for _, param := range statusReportParams {
		param := param
		value := connHandler.GetParamStatus(ctx, param)
		if err := c.sendParamStatus(param, value); err != nil {
			return sql.ConnectionHandler{}, err
		}
	}
	// The two following status parameters have no equivalent session
	// variable.
	if err := c.sendParamStatus("session_authorization", c.sessionArgs.User.Normalized()); err != nil {
		return sql.ConnectionHandler{}, err
	}

	if err := c.sendReadyForQuery(connHandler.GetQueryCancelKey()); err != nil {
		return sql.ConnectionHandler{}, err
	}
	return connHandler, nil
}

// sendReadyForQuery sends the final messages of the connection handshake.
// This includes a BackendKeyData message and a ServerMsgReady
// message indicating that there is no active transaction.
func (c *conn) sendReadyForQuery(queryCancelKey pgwirecancel.BackendKeyData) error {
	// Send our BackendKeyData to the client, so they can cancel the connection.
	c.msgBuilder.initMsg(pgwirebase.ServerMsgBackendKeyData)
	c.msgBuilder.putInt64(int64(queryCancelKey))
	if err := c.msgBuilder.finishMsg(c.conn); err != nil {
		return err
	}

	// An initial ServerMsgReady message is part of the handshake.
	c.msgBuilder.initMsg(pgwirebase.ServerMsgReady)
	c.msgBuilder.writeByte(byte(sql.IdleTxnBlock))
	if err := c.msgBuilder.finishMsg(c.conn); err != nil {
		return err
	}
	return nil
}

// An error is returned iff the statement buffer has been closed. In that case,
// the connection should be considered toast.
func (c *conn) handleSimpleQuery(
	ctx context.Context,
	buf *pgwirebase.ReadBuffer,
	timeReceived time.Time,
	unqualifiedIntSize *types.T,
) error {
	query, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}

	startParse := timeutil.Now()
	stmts, err := c.parser.ParseWithInt(query, unqualifiedIntSize)
	if err != nil {
		log.SqlExec.Errorf(ctx, "failed to parse simple query: %s", query)
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	endParse := timeutil.Now()

	if len(stmts) == 0 {
		return c.stmtBuf.Push(
			ctx, sql.ExecStmt{
				Statement:    parser.Statement{},
				TimeReceived: timeReceived,
				ParseStart:   startParse,
				ParseEnd:     endParse,
			})
	}

	for i := range stmts {
		// The CopyFrom statement is special. We need to detect it so we can hand
		// control of the connection, through the stmtBuf, to a copyMachine, and
		// block this network routine until control is passed back.
		if cp, ok := stmts[i].AST.(*tree.CopyFrom); ok {
			if len(stmts) != 1 {
				// NOTE(andrei): I don't know if Postgres supports receiving a COPY
				// together with other statements in the "simple" protocol, but I'd
				// rather not worry about it since execution of COPY is special - it
				// takes control over the connection.
				return c.stmtBuf.Push(
					ctx,
					sql.SendError{
						Err: pgwirebase.NewProtocolViolationErrorf(
							"COPY together with other statements in a query string is not supported"),
					})
			}
			copyDone := sync.WaitGroup{}
			copyDone.Add(1)
			if err := c.stmtBuf.Push(ctx, sql.CopyIn{Conn: c, Stmt: cp, CopyDone: &copyDone}); err != nil {
				return err
			}
			copyDone.Wait()
			return nil
		}

		if err := c.stmtBuf.Push(
			ctx,
			sql.ExecStmt{
				Statement:    stmts[i],
				TimeReceived: timeReceived,
				ParseStart:   startParse,
				ParseEnd:     endParse,
				LastInBatch:  i == len(stmts)-1,
			}); err != nil {
			return err
		}
	}
	return nil
}

// An error is returned iff the statement buffer has been closed. In that case,
// the connection should be considered toast.
func (c *conn) handleParse(
	ctx context.Context, buf *pgwirebase.ReadBuffer, nakedIntSize *types.T,
) error {
	telemetry.Inc(sqltelemetry.ParseRequestCounter)
	name, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	query, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	// The client may provide type information for (some of) the placeholders.
	numQArgTypes, err := buf.GetUint16()
	if err != nil {
		return err
	}
	inTypeHints := make([]oid.Oid, numQArgTypes)
	for i := range inTypeHints {
		typ, err := buf.GetUint32()
		if err != nil {
			return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
		}
		inTypeHints[i] = oid.Oid(typ)
	}

	startParse := timeutil.Now()
	stmts, err := c.parser.ParseWithInt(query, nakedIntSize)
	if err != nil {
		log.SqlExec.Errorf(ctx, "failed to parse: %s", query)
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	if len(stmts) > 1 {
		err := pgerror.WrongNumberOfPreparedStatements(len(stmts))
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	var stmt parser.Statement
	if len(stmts) == 1 {
		stmt = stmts[0]
	}
	// len(stmts) == 0 results in a nil (empty) statement.

	if len(inTypeHints) > stmt.NumPlaceholders {
		err := pgwirebase.NewProtocolViolationErrorf(
			"received too many type hints: %d vs %d placeholders in query",
			len(inTypeHints), stmt.NumPlaceholders,
		)
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}

	var sqlTypeHints tree.PlaceholderTypes
	if len(inTypeHints) > 0 {
		// Prepare the mapping of SQL placeholder names to types. Pre-populate it with
		// the type hints received from the client, if any.
		sqlTypeHints = make(tree.PlaceholderTypes, stmt.NumPlaceholders)
		for i, t := range inTypeHints {
			if t == 0 {
				continue
			}
			// If the OID is user defined or unknown, then write nil into the type
			// hints and let the consumer of the PrepareStmt resolve the types.
			if t == oid.T_unknown || types.IsOIDUserDefinedType(t) {
				sqlTypeHints[i] = nil
				continue
			}
			v, ok := types.OidToType[t]
			if !ok {
				err := pgwirebase.NewProtocolViolationErrorf("unknown oid type: %v", t)
				return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
			}
			sqlTypeHints[i] = v
		}
	}

	endParse := timeutil.Now()

	if _, ok := stmt.AST.(*tree.CopyFrom); ok {
		// We don't support COPY in extended protocol because it'd be complicated:
		// it wouldn't be the preparing, but the execution that would need to
		// execute the copyMachine.
		// Be aware that the copyMachine assumes it always runs in the simple
		// protocol, so if we ever support this, many parts of the copyMachine
		// would need to be changed.
		// Also note that COPY FROM in extended mode seems to be quite broken in
		// Postgres too:
		// https://www.postgresql.org/message-id/flat/CAMsr%2BYGvp2wRx9pPSxaKFdaObxX8DzWse%2BOkWk2xpXSvT0rq-g%40mail.gmail.com#CAMsr+YGvp2wRx9pPSxaKFdaObxX8DzWse+OkWk2xpXSvT0rq-g@mail.gmail.com
		return c.stmtBuf.Push(ctx, sql.SendError{Err: fmt.Errorf("CopyFrom not supported in extended protocol mode")})
	}

	return c.stmtBuf.Push(
		ctx,
		sql.PrepareStmt{
			Name:         name,
			Statement:    stmt,
			TypeHints:    sqlTypeHints,
			RawTypeHints: inTypeHints,
			ParseStart:   startParse,
			ParseEnd:     endParse,
		})
}

// An error is returned iff the statement buffer has been closed. In that case,
// the connection should be considered toast.
func (c *conn) handleDescribe(ctx context.Context, buf *pgwirebase.ReadBuffer) error {
	telemetry.Inc(sqltelemetry.DescribeRequestCounter)
	typ, err := buf.GetPrepareType()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	name, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	return c.stmtBuf.Push(
		ctx,
		sql.DescribeStmt{
			Name: name,
			Type: typ,
		})
}

// An error is returned iff the statement buffer has been closed. In that case,
// the connection should be considered toast.
func (c *conn) handleClose(ctx context.Context, buf *pgwirebase.ReadBuffer) error {
	telemetry.Inc(sqltelemetry.CloseRequestCounter)
	typ, err := buf.GetPrepareType()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	name, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	return c.stmtBuf.Push(
		ctx,
		sql.DeletePreparedStmt{
			Name: name,
			Type: typ,
		})
}

// If no format codes are provided then all arguments/result-columns use
// the default format, text.
var formatCodesAllText = []pgwirebase.FormatCode{pgwirebase.FormatText}

// handleBind queues instructions for creating a portal from a prepared
// statement.
// An error is returned iff the statement buffer has been closed. In that case,
// the connection should be considered toast.
func (c *conn) handleBind(ctx context.Context, buf *pgwirebase.ReadBuffer) error {
	telemetry.Inc(sqltelemetry.BindRequestCounter)
	portalName, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	statementName, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}

	// From the docs on number of argument format codes to bind:
	// This can be zero to indicate that there are no arguments or that the
	// arguments all use the default format (text); or one, in which case the
	// specified format code is applied to all arguments; or it can equal the
	// actual number of arguments.
	// http://www.postgresql.org/docs/current/static/protocol-message-formats.html
	numQArgFormatCodes, err := buf.GetUint16()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	var qArgFormatCodes []pgwirebase.FormatCode
	switch numQArgFormatCodes {
	case 0:
		// No format codes means all arguments are passed as text.
		qArgFormatCodes = formatCodesAllText
	case 1:
		// `1` means read one code and apply it to every argument.
		ch, err := buf.GetUint16()
		if err != nil {
			return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
		}
		code := pgwirebase.FormatCode(ch)
		if code == pgwirebase.FormatText {
			qArgFormatCodes = formatCodesAllText
		} else {
			qArgFormatCodes = []pgwirebase.FormatCode{code}
		}
	default:
		qArgFormatCodes = make([]pgwirebase.FormatCode, numQArgFormatCodes)
		// Read one format code for each argument and apply it to that argument.
		for i := range qArgFormatCodes {
			ch, err := buf.GetUint16()
			if err != nil {
				return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
			}
			qArgFormatCodes[i] = pgwirebase.FormatCode(ch)
		}
	}

	numValues, err := buf.GetUint16()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	qargs := make([][]byte, numValues)
	for i := 0; i < int(numValues); i++ {
		plen, err := buf.GetUint32()
		if err != nil {
			return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
		}
		if int32(plen) == -1 {
			// The argument is a NULL value.
			qargs[i] = nil
			continue
		}
		b, err := buf.GetBytes(int(plen))
		if err != nil {
			return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
		}
		qargs[i] = b
	}

	// From the docs on number of result-column format codes to bind:
	// This can be zero to indicate that there are no result columns or that
	// the result columns should all use the default format (text); or one, in
	// which case the specified format code is applied to all result columns
	// (if any); or it can equal the actual number of result columns of the
	// query.
	// http://www.postgresql.org/docs/current/static/protocol-message-formats.html
	numColumnFormatCodes, err := buf.GetUint16()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	var columnFormatCodes []pgwirebase.FormatCode
	switch numColumnFormatCodes {
	case 0:
		// All columns will use the text format.
		columnFormatCodes = formatCodesAllText
	case 1:
		// All columns will use the one specified format.
		ch, err := buf.GetUint16()
		if err != nil {
			return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
		}
		code := pgwirebase.FormatCode(ch)
		if code == pgwirebase.FormatText {
			columnFormatCodes = formatCodesAllText
		} else {
			columnFormatCodes = []pgwirebase.FormatCode{code}
		}
	default:
		columnFormatCodes = make([]pgwirebase.FormatCode, numColumnFormatCodes)
		// Read one format code for each column and apply it to that column.
		for i := range columnFormatCodes {
			ch, err := buf.GetUint16()
			if err != nil {
				return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
			}
			columnFormatCodes[i] = pgwirebase.FormatCode(ch)
		}
	}
	return c.stmtBuf.Push(
		ctx,
		sql.BindStmt{
			PreparedStatementName: statementName,
			PortalName:            portalName,
			Args:                  qargs,
			ArgFormatCodes:        qArgFormatCodes,
			OutFormats:            columnFormatCodes,
		})
}

// An error is returned iff the statement buffer has been closed. In that case,
// the connection should be considered toast.
func (c *conn) handleExecute(
	ctx context.Context, buf *pgwirebase.ReadBuffer, timeReceived time.Time, followedBySync bool,
) error {
	telemetry.Inc(sqltelemetry.ExecuteRequestCounter)
	portalName, err := buf.GetString()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	limit, err := buf.GetUint32()
	if err != nil {
		return c.stmtBuf.Push(ctx, sql.SendError{Err: err})
	}
	return c.stmtBuf.Push(ctx, sql.ExecPortal{
		Name:           portalName,
		TimeReceived:   timeReceived,
		Limit:          int(limit),
		FollowedBySync: followedBySync,
	})
}

func (c *conn) handleFlush(ctx context.Context) error {
	telemetry.Inc(sqltelemetry.FlushRequestCounter)
	return c.stmtBuf.Push(ctx, sql.Flush{})
}

// BeginCopyIn is part of the pgwirebase.Conn interface.
func (c *conn) BeginCopyIn(
	ctx context.Context, columns []colinfo.ResultColumn, format pgwirebase.FormatCode,
) error {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgCopyInResponse)
	c.msgBuilder.writeByte(byte(format))
	c.msgBuilder.putInt16(int16(len(columns)))
	for range columns {
		c.msgBuilder.putInt16(int16(format))
	}
	return c.msgBuilder.finishMsg(c.conn)
}

// SendCommandComplete is part of the pgwirebase.Conn interface.
func (c *conn) SendCommandComplete(tag []byte) error {
	c.bufferCommandComplete(tag)
	return nil
}

// Rd is part of the pgwirebase.Conn interface.
func (c *conn) Rd() pgwirebase.BufferedReader {
	return &pgwireReader{conn: c}
}

// flushInfo encapsulates information about what results have been flushed to
// the network.
type flushInfo struct {
	// buf is a reference to writerState.buf.
	buf *bytes.Buffer
	// lastFlushed indicates the highest command for which results have been
	// flushed. The command may have further results in the buffer that haven't
	// been flushed.
	lastFlushed sql.CmdPos
	// cmdStarts maintains the state about where the results for the respective
	// positions begin. We utilize the invariant that positions are
	// monotonically increasing sequences.
	cmdStarts cmdIdxBuffer
}

type cmdIdx struct {
	pos sql.CmdPos
	idx int
}

var cmdIdxPool = sync.Pool{
	New: func() interface{} {
		return &cmdIdx{}
	},
}

func (c *cmdIdx) release() {
	*c = cmdIdx{}
	cmdIdxPool.Put(c)
}

type cmdIdxBuffer struct {
	// We intentionally do not just embed ring.Buffer in order to restrict the
	// methods that can be called on cmdIdxBuffer.
	buffer ring.Buffer
}

func (b *cmdIdxBuffer) empty() bool {
	return b.buffer.Len() == 0
}

func (b *cmdIdxBuffer) addLast(pos sql.CmdPos, idx int) {
	cmdIdx := cmdIdxPool.Get().(*cmdIdx)
	cmdIdx.pos = pos
	cmdIdx.idx = idx
	b.buffer.AddLast(cmdIdx)
}

// removeLast removes the last cmdIdx from the buffer and will panic if the
// buffer is empty.
func (b *cmdIdxBuffer) removeLast() {
	b.getLast().release()
	b.buffer.RemoveLast()
}

// getLast returns the last cmdIdx in the buffer and will panic if the buffer is
// empty.
func (b *cmdIdxBuffer) getLast() *cmdIdx {
	return b.buffer.GetLast().(*cmdIdx)
}

func (b *cmdIdxBuffer) clear() {
	for !b.empty() {
		b.removeLast()
	}
}

// registerCmd updates cmdStarts buffer when the first result for a new command
// is received.
func (fi *flushInfo) registerCmd(pos sql.CmdPos) {
	if !fi.cmdStarts.empty() && fi.cmdStarts.getLast().pos >= pos {
		// Not a new command, nothing to do.
		return
	}
	fi.cmdStarts.addLast(pos, fi.buf.Len())
}

func cookTag(
	tagStr string, buf []byte, stmtType tree.StatementReturnType, rowsAffected int,
) []byte {
	if tagStr == "INSERT" {
		// From the postgres docs (49.5. Message Formats):
		// `INSERT oid rows`... oid is the object ID of the inserted row if
		//	rows is 1 and the target table has OIDs; otherwise oid is 0.
		tagStr = "INSERT 0"
	}
	tag := append(buf, tagStr...)

	switch stmtType {
	case tree.RowsAffected:
		tag = append(tag, ' ')
		tag = strconv.AppendInt(tag, int64(rowsAffected), 10)

	case tree.Rows:
		tag = append(tag, ' ')
		tag = strconv.AppendUint(tag, uint64(rowsAffected), 10)

	case tree.Ack, tree.DDL:
		if tagStr == "SELECT" {
			tag = append(tag, ' ')
			tag = strconv.AppendInt(tag, int64(rowsAffected), 10)
		}

	case tree.CopyIn:
		// Nothing to do. The CommandComplete message has been sent elsewhere.
		panic(errors.AssertionFailedf("CopyIn statements should have been handled elsewhere " +
			"and not produce results"))
	default:
		panic(errors.AssertionFailedf("unexpected result type %v", stmtType))
	}

	return tag
}

// bufferRow serializes a row and adds it to the buffer. Depending on the buffer
// size limit, bufferRow may flush the buffered data to the connection.
func (c *conn) bufferRow(ctx context.Context, row tree.Datums, r *commandResult) error {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgDataRow)
	c.msgBuilder.putInt16(int16(len(row)))
	for i, col := range row {
		fmtCode := pgwirebase.FormatText
		if r.formatCodes != nil {
			fmtCode = r.formatCodes[i]
		}
		switch fmtCode {
		case pgwirebase.FormatText:
			c.msgBuilder.writeTextDatum(ctx, col, r.conv, r.location, r.types[i])
		case pgwirebase.FormatBinary:
			c.msgBuilder.writeBinaryDatum(ctx, col, r.location, r.types[i])
		default:
			c.msgBuilder.setError(errors.Errorf("unsupported format code %s", fmtCode))
		}
	}
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
	return c.maybeFlush(r.pos, r.bufferingDisabled)
}

// bufferBatch serializes a batch and adds all the rows from it to the buffer.
// It is a noop for zero-length batch. Depending on the buffer size limit,
// bufferBatch may flush the buffered data to the connection.
func (c *conn) bufferBatch(ctx context.Context, batch coldata.Batch, r *commandResult) error {
	sel := batch.Selection()
	n := batch.Length()
	if n > 0 {
		c.vecsScratch.SetBatch(batch)
		// Make sure that c doesn't hold on to the memory of the batch.
		defer c.vecsScratch.Reset()
		width := int16(len(c.vecsScratch.Vecs))
		for i := 0; i < n; i++ {
			rowIdx := i
			if sel != nil {
				rowIdx = sel[rowIdx]
			}
			c.msgBuilder.initMsg(pgwirebase.ServerMsgDataRow)
			c.msgBuilder.putInt16(width)
			for vecIdx := 0; vecIdx < len(c.vecsScratch.Vecs); vecIdx++ {
				fmtCode := pgwirebase.FormatText
				if r.formatCodes != nil {
					fmtCode = r.formatCodes[vecIdx]
				}
				switch fmtCode {
				case pgwirebase.FormatText:
					c.msgBuilder.writeTextColumnarElement(ctx, &c.vecsScratch, vecIdx, rowIdx, r.conv, r.location)
				case pgwirebase.FormatBinary:
					c.msgBuilder.writeBinaryColumnarElement(ctx, &c.vecsScratch, vecIdx, rowIdx, r.location)
				default:
					c.msgBuilder.setError(errors.Errorf("unsupported format code %s", fmtCode))
				}
			}
			if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
				panic(fmt.Sprintf("unexpected err from buffer: %s", err))
			}
			if err := c.maybeFlush(r.pos, r.bufferingDisabled); err != nil {
				return err
			}
		}
	}
	return nil
}

func (c *conn) bufferReadyForQuery(txnStatus byte) {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgReady)
	c.msgBuilder.writeByte(txnStatus)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferParseComplete() {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgParseComplete)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferBindComplete() {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgBindComplete)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferCloseComplete() {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgCloseComplete)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferCommandComplete(tag []byte) {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgCommandComplete)
	c.msgBuilder.write(tag)
	c.msgBuilder.nullTerminate()
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferPortalSuspended() {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgPortalSuspended)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferErr(ctx context.Context, err error) {
	if err := writeErr(ctx, c.sv,
		err, &c.msgBuilder, &c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferEmptyQueryResponse() {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgEmptyQuery)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func writeErr(
	ctx context.Context, sv *settings.Values, err error, msgBuilder *writeBuffer, w io.Writer,
) error {
	// Record telemetry for the error.
	sqltelemetry.RecordError(ctx, err, sv)
	msgBuilder.initMsg(pgwirebase.ServerMsgErrorResponse)
	return writeErrFields(ctx, sv, err, msgBuilder, w)
}

func writeErrFields(
	ctx context.Context, sv *settings.Values, err error, msgBuilder *writeBuffer, w io.Writer,
) error {
	pgErr := pgerror.Flatten(err)

	msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldSeverity)
	msgBuilder.writeTerminatedString(pgErr.Severity)

	msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldSeverityNonLocalized)
	msgBuilder.writeTerminatedString(pgErr.Severity)

	msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldSQLState)
	msgBuilder.writeTerminatedString(pgErr.Code)

	if pgErr.Detail != "" {
		msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldDetail)
		msgBuilder.writeTerminatedString(pgErr.Detail)
	}

	if pgErr.Hint != "" {
		msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldHint)
		msgBuilder.writeTerminatedString(pgErr.Hint)
	}

	if pgErr.ConstraintName != "" {
		msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldConstraintName)
		msgBuilder.writeTerminatedString(pgErr.ConstraintName)
	}

	if pgErr.Source != nil {
		errCtx := pgErr.Source
		if errCtx.File != "" {
			msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldSrcFile)
			msgBuilder.writeTerminatedString(errCtx.File)
		}

		if errCtx.Line > 0 {
			msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldSrcLine)
			msgBuilder.writeTerminatedString(strconv.Itoa(int(errCtx.Line)))
		}

		if errCtx.Function != "" {
			msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldSrcFunction)
			msgBuilder.writeTerminatedString(errCtx.Function)
		}
	}

	msgBuilder.putErrFieldMsg(pgwirebase.ServerErrFieldMsgPrimary)
	msgBuilder.writeTerminatedString(pgErr.Message)

	msgBuilder.nullTerminate()
	return msgBuilder.finishMsg(w)
}

func (c *conn) bufferParamDesc(types []oid.Oid) {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgParameterDescription)
	c.msgBuilder.putInt16(int16(len(types)))
	for _, t := range types {
		c.msgBuilder.putInt32(int32(t))
	}
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

func (c *conn) bufferNoDataMsg() {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgNoData)
	if err := c.msgBuilder.finishMsg(&c.writerState.buf); err != nil {
		panic(errors.NewAssertionErrorWithWrappedErrf(err, "unexpected err from buffer"))
	}
}

// writeRowDescription writes a row description to the given writer.
//
// formatCodes specifies the format for each column. It can be nil, in which
// case all columns will use FormatText.
//
// If an error is returned, it has also been saved on c.err.
func (c *conn) writeRowDescription(
	ctx context.Context,
	columns []colinfo.ResultColumn,
	formatCodes []pgwirebase.FormatCode,
	w io.Writer,
) error {
	c.msgBuilder.initMsg(pgwirebase.ServerMsgRowDescription)
	c.msgBuilder.putInt16(int16(len(columns)))
	for i, column := range columns {
		if log.V(2) {
			log.Infof(ctx, "pgwire: writing column %s of type: %s", column.Name, column.Typ)
		}
		c.msgBuilder.writeTerminatedString(column.Name)
		typ := pgTypeForParserType(column.Typ)
		c.msgBuilder.putInt32(int32(column.TableID))        // Table OID (optional).
		c.msgBuilder.putInt16(int16(column.PGAttributeNum)) // Column attribute ID (optional).
		c.msgBuilder.putInt32(int32(typ.oid))
		c.msgBuilder.putInt16(int16(typ.size))
		c.msgBuilder.putInt32(column.GetTypeModifier()) // Type modifier
		if formatCodes == nil {
			c.msgBuilder.putInt16(int16(pgwirebase.FormatText))
		} else {
			c.msgBuilder.putInt16(int16(formatCodes[i]))
		}
	}
	if err := c.msgBuilder.finishMsg(w); err != nil {
		c.setErr(err)
		return err
	}
	return nil
}

// Flush is part of the ClientComm interface.
//
// In case conn.err is set, this is a no-op - the previous err is returned.
func (c *conn) Flush(pos sql.CmdPos) error {
	// Check that there were no previous network errors. If there were, we'd
	// probably also fail the write below, but this check is here to make
	// absolutely sure that we don't send some results after we previously had
	// failed to send others.
	if err := c.GetErr(); err != nil {
		return err
	}

	c.writerState.fi.lastFlushed = pos
	// Make sure that the entire cmdStarts buffer is drained.
	c.writerState.fi.cmdStarts.clear()

	_ /* n */, err := c.writerState.buf.WriteTo(c.conn)
	if err != nil {
		c.setErr(err)
		return err
	}
	return nil
}

// maybeFlush flushes the buffer to the network connection if it exceeded
// sessionArgs.ConnResultsBufferSize or if buffering is disabled.
func (c *conn) maybeFlush(pos sql.CmdPos, bufferingDisabled bool) error {
	if !bufferingDisabled && int64(c.writerState.buf.Len()) <= c.sessionArgs.ConnResultsBufferSize {
		return nil
	}
	return c.Flush(pos)
}

// LockCommunication is part of the ClientComm interface.
//
// The current implementation of conn writes results to the network
// synchronously, as they are produced (modulo buffering). Therefore, there's
// nothing to "lock" - communication is naturally blocked as the command
// processor won't write any more results.
func (c *conn) LockCommunication() sql.ClientLock {
	return (*clientConnLock)(&c.writerState.fi)
}

// clientConnLock is the connection's implementation of sql.ClientLock. It lets
// the sql module lock the flushing of results and find out what has already
// been flushed.
type clientConnLock flushInfo

var _ sql.ClientLock = &clientConnLock{}

// Close is part of the sql.ClientLock interface.
func (cl *clientConnLock) Close() {
	// Nothing to do. See LockCommunication note.
}

// ClientPos is part of the sql.ClientLock interface.
func (cl *clientConnLock) ClientPos() sql.CmdPos {
	return cl.lastFlushed
}

// RTrim is part of the sql.ClientLock interface.
func (cl *clientConnLock) RTrim(ctx context.Context, pos sql.CmdPos) {
	if pos <= cl.lastFlushed {
		panic(errors.AssertionFailedf("asked to trim to pos: %d, below the last flush: %d", pos, cl.lastFlushed))
	}
	// If we don't have a start index for pos yet, it must be that no results
	// for it yet have been produced yet.
	truncateIdx := cl.buf.Len()
	// Update cmdStarts buffer: delete commands that were trimmed from the back
	// of the cmdStarts buffer.
	for !cl.cmdStarts.empty() {
		cmdStart := cl.cmdStarts.getLast()
		if cmdStart.pos < pos {
			break
		}
		truncateIdx = cmdStart.idx
		cl.cmdStarts.removeLast()
	}
	cl.buf.Truncate(truncateIdx)
}

// CreateStatementResult is part of the sql.ClientComm interface.
func (c *conn) CreateStatementResult(
	stmt tree.Statement,
	descOpt sql.RowDescOpt,
	pos sql.CmdPos,
	formatCodes []pgwirebase.FormatCode,
	conv sessiondatapb.DataConversionConfig,
	location *time.Location,
	limit int,
	portalName string,
	implicitTxn bool,
) sql.CommandResult {
	return c.newCommandResult(descOpt, pos, stmt, formatCodes, conv, location, limit, portalName, implicitTxn)
}

// CreateSyncResult is part of the sql.ClientComm interface.
func (c *conn) CreateSyncResult(pos sql.CmdPos) sql.SyncResult {
	return c.newMiscResult(pos, readyForQuery)
}

// CreateFlushResult is part of the sql.ClientComm interface.
func (c *conn) CreateFlushResult(pos sql.CmdPos) sql.FlushResult {
	return c.newMiscResult(pos, flush)
}

// CreateDrainResult is part of the sql.ClientComm interface.
func (c *conn) CreateDrainResult(pos sql.CmdPos) sql.DrainResult {
	return c.newMiscResult(pos, noCompletionMsg)
}

// CreateBindResult is part of the sql.ClientComm interface.
func (c *conn) CreateBindResult(pos sql.CmdPos) sql.BindResult {
	return c.newMiscResult(pos, bindComplete)
}

// CreatePrepareResult is part of the sql.ClientComm interface.
func (c *conn) CreatePrepareResult(pos sql.CmdPos) sql.ParseResult {
	return c.newMiscResult(pos, parseComplete)
}

// CreateDescribeResult is part of the sql.ClientComm interface.
func (c *conn) CreateDescribeResult(pos sql.CmdPos) sql.DescribeResult {
	return c.newMiscResult(pos, noCompletionMsg)
}

// CreateEmptyQueryResult is part of the sql.ClientComm interface.
func (c *conn) CreateEmptyQueryResult(pos sql.CmdPos) sql.EmptyQueryResult {
	return c.newMiscResult(pos, emptyQueryResponse)
}

// CreateDeleteResult is part of the sql.ClientComm interface.
func (c *conn) CreateDeleteResult(pos sql.CmdPos) sql.DeleteResult {
	return c.newMiscResult(pos, closeComplete)
}

// CreateErrorResult is part of the sql.ClientComm interface.
func (c *conn) CreateErrorResult(pos sql.CmdPos) sql.ErrorResult {
	res := c.newMiscResult(pos, noCompletionMsg)
	res.errExpected = true
	return res
}

// CreateCopyInResult is part of the sql.ClientComm interface.
func (c *conn) CreateCopyInResult(pos sql.CmdPos) sql.CopyInResult {
	return c.newMiscResult(pos, noCompletionMsg)
}

// pgwireReader is an io.Reader that wraps a conn, maintaining its metrics as
// it is consumed.
type pgwireReader struct {
	conn *conn
}

// pgwireReader implements the pgwirebase.BufferedReader interface.
var _ pgwirebase.BufferedReader = &pgwireReader{}

// Read is part of the pgwirebase.BufferedReader interface.
func (r *pgwireReader) Read(p []byte) (int, error) {
	n, err := r.conn.rd.Read(p)
	r.conn.metrics.BytesInCount.Inc(int64(n))
	return n, err
}

// ReadString is part of the pgwirebase.BufferedReader interface.
func (r *pgwireReader) ReadString(delim byte) (string, error) {
	s, err := r.conn.rd.ReadString(delim)
	r.conn.metrics.BytesInCount.Inc(int64(len(s)))
	return s, err
}

// ReadByte is part of the pgwirebase.BufferedReader interface.
func (r *pgwireReader) ReadByte() (byte, error) {
	b, err := r.conn.rd.ReadByte()
	if err == nil {
		r.conn.metrics.BytesInCount.Inc(1)
	}
	return b, err
}

// statusReportParams is a list of session variables that are also
// reported as server run-time parameters in the pgwire connection
// initialization.
//
// The standard PostgreSQL status vars are listed here:
// https://www.postgresql.org/docs/10/static/libpq-status.html
var statusReportParams = []string{
	"server_version",
	"server_encoding",
	"client_encoding",
	"application_name",
	// Note: session_authorization is handled specially in serveImpl().
	"DateStyle",
	"IntervalStyle",
	"is_superuser",
	"TimeZone",
	"integer_datetimes",
	"standard_conforming_strings",
	"crdb_version", // CockroachDB extension.
}

// testingStatusReportParams is the minimum set of status parameters
// needed to make pgx tests in the local package happy.
var testingStatusReportParams = map[string]string{
	"client_encoding":             "UTF8",
	"standard_conforming_strings": "on",
}

// readTimeoutConn overloads net.Conn.Read by periodically calling
// checkExitConds() and aborting the read if an error is returned.
type readTimeoutConn struct {
	net.Conn
	// checkExitConds is called periodically by Read(). If it returns an error,
	// the Read() returns that error. Future calls to Read() are allowed, in which
	// case checkExitConds() will be called again.
	checkExitConds func() error
}

// NewReadTimeoutConn wraps the given connection with a readTimeoutConn.
func NewReadTimeoutConn(c net.Conn, checkExitConds func() error) net.Conn {
	return &readTimeoutConn{
		Conn:           c,
		checkExitConds: checkExitConds,
	}
}

func (c *readTimeoutConn) Read(b []byte) (int, error) {
	// readTimeout is the amount of time ReadTimeoutConn should wait on a
	// read before checking for exit conditions. The tradeoff is between the
	// time it takes to react to session context cancellation and the overhead
	// of waking up and checking for exit conditions.
	const readTimeout = 1 * time.Second

	// Remove the read deadline when returning from this function to avoid
	// unexpected behavior.
	defer func() { _ = c.SetReadDeadline(time.Time{}) }()
	for {
		if err := c.checkExitConds(); err != nil {
			return 0, err
		}
		if err := c.SetReadDeadline(timeutil.Now().Add(readTimeout)); err != nil {
			return 0, err
		}
		n, err := c.Conn.Read(b)
		if err != nil {
			// Continue if the error is due to timing out.
			if ne := (net.Error)(nil); errors.As(err, &ne) && ne.Timeout() {
				continue
			}
		}
		return n, err
	}
}