mirror of
https://github.com/ceph/ceph-csi.git
synced 2024-12-23 05:20:19 +00:00
f080b9e0c9
Signed-off-by: Niels de Vos <ndevos@ibm.com>
503 lines
17 KiB
Go
503 lines
17 KiB
Go
/*
|
|
Copyright 2023 The Kubernetes Authors.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
*/
|
|
|
|
package remotecommand
|
|
|
|
import (
|
|
"context"
|
|
"errors"
|
|
"fmt"
|
|
"io"
|
|
"net"
|
|
"net/http"
|
|
"sync"
|
|
"time"
|
|
|
|
gwebsocket "github.com/gorilla/websocket"
|
|
|
|
v1 "k8s.io/api/core/v1"
|
|
"k8s.io/apimachinery/pkg/util/httpstream"
|
|
"k8s.io/apimachinery/pkg/util/remotecommand"
|
|
restclient "k8s.io/client-go/rest"
|
|
"k8s.io/client-go/transport/websocket"
|
|
"k8s.io/klog/v2"
|
|
)
|
|
|
|
// writeDeadline defines the time that a write to the websocket connection
|
|
// must complete by, otherwise an i/o timeout occurs. The writeDeadline
|
|
// has nothing to do with a response from the other websocket connection
|
|
// endpoint; only that the message was successfully processed by the
|
|
// local websocket connection. The typical write deadline within the websocket
|
|
// library is one second.
|
|
const writeDeadline = 2 * time.Second
|
|
|
|
var (
|
|
_ Executor = &wsStreamExecutor{}
|
|
_ streamCreator = &wsStreamCreator{}
|
|
_ httpstream.Stream = &stream{}
|
|
|
|
streamType2streamID = map[string]byte{
|
|
v1.StreamTypeStdin: remotecommand.StreamStdIn,
|
|
v1.StreamTypeStdout: remotecommand.StreamStdOut,
|
|
v1.StreamTypeStderr: remotecommand.StreamStdErr,
|
|
v1.StreamTypeError: remotecommand.StreamErr,
|
|
v1.StreamTypeResize: remotecommand.StreamResize,
|
|
}
|
|
)
|
|
|
|
const (
|
|
// pingPeriod defines how often a heartbeat "ping" message is sent.
|
|
pingPeriod = 5 * time.Second
|
|
// pingReadDeadline defines the time waiting for a response heartbeat
|
|
// "pong" message before a timeout error occurs for websocket reading.
|
|
// This duration must always be greater than the "pingPeriod". By defining
|
|
// this deadline in terms of the ping period, we are essentially saying
|
|
// we can drop "X-1" (e.g. 3-1=2) pings before firing the timeout.
|
|
pingReadDeadline = (pingPeriod * 3) + (1 * time.Second)
|
|
)
|
|
|
|
// wsStreamExecutor handles transporting standard shell streams over an httpstream connection.
|
|
type wsStreamExecutor struct {
|
|
transport http.RoundTripper
|
|
upgrader websocket.ConnectionHolder
|
|
method string
|
|
url string
|
|
// requested protocols in priority order (e.g. v5.channel.k8s.io before v4.channel.k8s.io).
|
|
protocols []string
|
|
// selected protocol from the handshake process; could be empty string if handshake fails.
|
|
negotiated string
|
|
// period defines how often a "ping" heartbeat message is sent to the other endpoint.
|
|
heartbeatPeriod time.Duration
|
|
// deadline defines the amount of time before "pong" response must be received.
|
|
heartbeatDeadline time.Duration
|
|
}
|
|
|
|
func NewWebSocketExecutor(config *restclient.Config, method, url string) (Executor, error) {
|
|
// Only supports V5 protocol for correct version skew functionality.
|
|
// Previous api servers will proxy upgrade requests to legacy websocket
|
|
// servers on container runtimes which support V1-V4. These legacy
|
|
// websocket servers will not handle the new CLOSE signal.
|
|
return NewWebSocketExecutorForProtocols(config, method, url, remotecommand.StreamProtocolV5Name)
|
|
}
|
|
|
|
// NewWebSocketExecutorForProtocols allows to execute commands via a WebSocket connection.
|
|
func NewWebSocketExecutorForProtocols(config *restclient.Config, method, url string, protocols ...string) (Executor, error) {
|
|
transport, upgrader, err := websocket.RoundTripperFor(config)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("error creating websocket transports: %v", err)
|
|
}
|
|
return &wsStreamExecutor{
|
|
transport: transport,
|
|
upgrader: upgrader,
|
|
method: method,
|
|
url: url,
|
|
protocols: protocols,
|
|
heartbeatPeriod: pingPeriod,
|
|
heartbeatDeadline: pingReadDeadline,
|
|
}, nil
|
|
}
|
|
|
|
// Deprecated: use StreamWithContext instead to avoid possible resource leaks.
|
|
// See https://github.com/kubernetes/kubernetes/pull/103177 for details.
|
|
func (e *wsStreamExecutor) Stream(options StreamOptions) error {
|
|
return e.StreamWithContext(context.Background(), options)
|
|
}
|
|
|
|
// StreamWithContext upgrades an HTTPRequest to a WebSocket connection, and starts the various
|
|
// goroutines to implement the necessary streams over the connection. The "options" parameter
|
|
// defines which streams are requested. Returns an error if one occurred. This method is NOT
|
|
// safe to run concurrently with the same executor (because of the state stored in the upgrader).
|
|
func (e *wsStreamExecutor) StreamWithContext(ctx context.Context, options StreamOptions) error {
|
|
req, err := http.NewRequestWithContext(ctx, e.method, e.url, nil)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
conn, err := websocket.Negotiate(e.transport, e.upgrader, req, e.protocols...)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
if conn == nil {
|
|
panic(fmt.Errorf("websocket connection is nil"))
|
|
}
|
|
defer conn.Close()
|
|
e.negotiated = conn.Subprotocol()
|
|
klog.V(4).Infof("The subprotocol is %s", e.negotiated)
|
|
|
|
var streamer streamProtocolHandler
|
|
switch e.negotiated {
|
|
case remotecommand.StreamProtocolV5Name:
|
|
streamer = newStreamProtocolV5(options)
|
|
case remotecommand.StreamProtocolV4Name:
|
|
streamer = newStreamProtocolV4(options)
|
|
case remotecommand.StreamProtocolV3Name:
|
|
streamer = newStreamProtocolV3(options)
|
|
case remotecommand.StreamProtocolV2Name:
|
|
streamer = newStreamProtocolV2(options)
|
|
case "":
|
|
klog.V(4).Infof("The server did not negotiate a streaming protocol version. Falling back to %s", remotecommand.StreamProtocolV1Name)
|
|
fallthrough
|
|
case remotecommand.StreamProtocolV1Name:
|
|
streamer = newStreamProtocolV1(options)
|
|
}
|
|
|
|
panicChan := make(chan any, 1)
|
|
errorChan := make(chan error, 1)
|
|
go func() {
|
|
defer func() {
|
|
if p := recover(); p != nil {
|
|
panicChan <- p
|
|
}
|
|
}()
|
|
creator := newWSStreamCreator(conn)
|
|
go creator.readDemuxLoop(
|
|
e.upgrader.DataBufferSize(),
|
|
e.heartbeatPeriod,
|
|
e.heartbeatDeadline,
|
|
)
|
|
errorChan <- streamer.stream(creator)
|
|
}()
|
|
|
|
select {
|
|
case p := <-panicChan:
|
|
panic(p)
|
|
case err := <-errorChan:
|
|
return err
|
|
case <-ctx.Done():
|
|
return ctx.Err()
|
|
}
|
|
}
|
|
|
|
type wsStreamCreator struct {
|
|
conn *gwebsocket.Conn
|
|
// Protects writing to websocket connection; reading is lock-free
|
|
connWriteLock sync.Mutex
|
|
// map of stream id to stream; multiple streams read/write the connection
|
|
streams map[byte]*stream
|
|
streamsMu sync.Mutex
|
|
}
|
|
|
|
func newWSStreamCreator(conn *gwebsocket.Conn) *wsStreamCreator {
|
|
return &wsStreamCreator{
|
|
conn: conn,
|
|
streams: map[byte]*stream{},
|
|
}
|
|
}
|
|
|
|
func (c *wsStreamCreator) getStream(id byte) *stream {
|
|
c.streamsMu.Lock()
|
|
defer c.streamsMu.Unlock()
|
|
return c.streams[id]
|
|
}
|
|
|
|
func (c *wsStreamCreator) setStream(id byte, s *stream) {
|
|
c.streamsMu.Lock()
|
|
defer c.streamsMu.Unlock()
|
|
c.streams[id] = s
|
|
}
|
|
|
|
// CreateStream uses id from passed headers to create a stream over "c.conn" connection.
|
|
// Returns a Stream structure or nil and an error if one occurred.
|
|
func (c *wsStreamCreator) CreateStream(headers http.Header) (httpstream.Stream, error) {
|
|
streamType := headers.Get(v1.StreamType)
|
|
id, ok := streamType2streamID[streamType]
|
|
if !ok {
|
|
return nil, fmt.Errorf("unknown stream type: %s", streamType)
|
|
}
|
|
if s := c.getStream(id); s != nil {
|
|
return nil, fmt.Errorf("duplicate stream for type %s", streamType)
|
|
}
|
|
reader, writer := io.Pipe()
|
|
s := &stream{
|
|
headers: headers,
|
|
readPipe: reader,
|
|
writePipe: writer,
|
|
conn: c.conn,
|
|
connWriteLock: &c.connWriteLock,
|
|
id: id,
|
|
}
|
|
c.setStream(id, s)
|
|
return s, nil
|
|
}
|
|
|
|
// readDemuxLoop is the lock-free reading processor for this endpoint of the websocket
|
|
// connection. This loop reads the connection, and demultiplexes the data
|
|
// into one of the individual stream pipes (by checking the stream id). This
|
|
// loop can *not* be run concurrently, because there can only be one websocket
|
|
// connection reader at a time (a read mutex would provide no benefit).
|
|
func (c *wsStreamCreator) readDemuxLoop(bufferSize int, period time.Duration, deadline time.Duration) {
|
|
// Initialize and start the ping/pong heartbeat.
|
|
h := newHeartbeat(c.conn, period, deadline)
|
|
// Set initial timeout for websocket connection reading.
|
|
if err := c.conn.SetReadDeadline(time.Now().Add(deadline)); err != nil {
|
|
klog.Errorf("Websocket initial setting read deadline failed %v", err)
|
|
return
|
|
}
|
|
go h.start()
|
|
// Buffer size must correspond to the same size allocated
|
|
// for the read buffer during websocket client creation. A
|
|
// difference can cause incomplete connection reads.
|
|
readBuffer := make([]byte, bufferSize)
|
|
for {
|
|
// NextReader() only returns data messages (BinaryMessage or Text
|
|
// Message). Even though this call will never return control frames
|
|
// such as ping, pong, or close, this call is necessary for these
|
|
// message types to be processed. There can only be one reader
|
|
// at a time, so this reader loop must *not* be run concurrently;
|
|
// there is no lock for reading. Calling "NextReader()" before the
|
|
// current reader has been processed will close the current reader.
|
|
// If the heartbeat read deadline times out, this "NextReader()" will
|
|
// return an i/o error, and error handling will clean up.
|
|
messageType, r, err := c.conn.NextReader()
|
|
if err != nil {
|
|
websocketErr, ok := err.(*gwebsocket.CloseError)
|
|
if ok && websocketErr.Code == gwebsocket.CloseNormalClosure {
|
|
err = nil // readers will get io.EOF as it's a normal closure
|
|
} else {
|
|
err = fmt.Errorf("next reader: %w", err)
|
|
}
|
|
c.closeAllStreamReaders(err)
|
|
return
|
|
}
|
|
// All remote command protocols send/receive only binary data messages.
|
|
if messageType != gwebsocket.BinaryMessage {
|
|
c.closeAllStreamReaders(fmt.Errorf("unexpected message type: %d", messageType))
|
|
return
|
|
}
|
|
// It's ok to read just a single byte because the underlying library wraps the actual
|
|
// connection with a buffered reader anyway.
|
|
_, err = io.ReadFull(r, readBuffer[:1])
|
|
if err != nil {
|
|
c.closeAllStreamReaders(fmt.Errorf("read stream id: %w", err))
|
|
return
|
|
}
|
|
streamID := readBuffer[0]
|
|
s := c.getStream(streamID)
|
|
if s == nil {
|
|
klog.Errorf("Unknown stream id %d, discarding message", streamID)
|
|
continue
|
|
}
|
|
for {
|
|
nr, errRead := r.Read(readBuffer)
|
|
if nr > 0 {
|
|
// Write the data to the stream's pipe. This can block.
|
|
_, errWrite := s.writePipe.Write(readBuffer[:nr])
|
|
if errWrite != nil {
|
|
// Pipe must have been closed by the stream user.
|
|
// Nothing to do, discard the message.
|
|
break
|
|
}
|
|
}
|
|
if errRead != nil {
|
|
if errRead == io.EOF {
|
|
break
|
|
}
|
|
c.closeAllStreamReaders(fmt.Errorf("read message: %w", err))
|
|
return
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// closeAllStreamReaders closes readers in all streams.
|
|
// This unblocks all stream.Read() calls.
|
|
func (c *wsStreamCreator) closeAllStreamReaders(err error) {
|
|
c.streamsMu.Lock()
|
|
defer c.streamsMu.Unlock()
|
|
for _, s := range c.streams {
|
|
// Closing writePipe unblocks all readPipe.Read() callers and prevents any future writes.
|
|
_ = s.writePipe.CloseWithError(err)
|
|
}
|
|
}
|
|
|
|
type stream struct {
|
|
headers http.Header
|
|
readPipe *io.PipeReader
|
|
writePipe *io.PipeWriter
|
|
// conn is used for writing directly into the connection.
|
|
// Is nil after Close() / Reset() to prevent future writes.
|
|
conn *gwebsocket.Conn
|
|
// connWriteLock protects conn against concurrent write operations. There must be a single writer and a single reader only.
|
|
// The mutex is shared across all streams because the underlying connection is shared.
|
|
connWriteLock *sync.Mutex
|
|
id byte
|
|
}
|
|
|
|
func (s *stream) Read(p []byte) (n int, err error) {
|
|
return s.readPipe.Read(p)
|
|
}
|
|
|
|
// Write writes directly to the underlying WebSocket connection.
|
|
func (s *stream) Write(p []byte) (n int, err error) {
|
|
klog.V(4).Infof("Write() on stream %d", s.id)
|
|
defer klog.V(4).Infof("Write() done on stream %d", s.id)
|
|
s.connWriteLock.Lock()
|
|
defer s.connWriteLock.Unlock()
|
|
if s.conn == nil {
|
|
return 0, fmt.Errorf("write on closed stream %d", s.id)
|
|
}
|
|
err = s.conn.SetWriteDeadline(time.Now().Add(writeDeadline))
|
|
if err != nil {
|
|
klog.V(7).Infof("Websocket setting write deadline failed %v", err)
|
|
return 0, err
|
|
}
|
|
// Message writer buffers the message data, so we don't need to do that ourselves.
|
|
// Just write id and the data as two separate writes to avoid allocating an intermediate buffer.
|
|
w, err := s.conn.NextWriter(gwebsocket.BinaryMessage)
|
|
if err != nil {
|
|
return 0, err
|
|
}
|
|
defer func() {
|
|
if w != nil {
|
|
w.Close()
|
|
}
|
|
}()
|
|
_, err = w.Write([]byte{s.id})
|
|
if err != nil {
|
|
return 0, err
|
|
}
|
|
n, err = w.Write(p)
|
|
if err != nil {
|
|
return n, err
|
|
}
|
|
err = w.Close()
|
|
w = nil
|
|
return n, err
|
|
}
|
|
|
|
// Close half-closes the stream, indicating this side is finished with the stream.
|
|
func (s *stream) Close() error {
|
|
klog.V(4).Infof("Close() on stream %d", s.id)
|
|
defer klog.V(4).Infof("Close() done on stream %d", s.id)
|
|
s.connWriteLock.Lock()
|
|
defer s.connWriteLock.Unlock()
|
|
if s.conn == nil {
|
|
return fmt.Errorf("Close() on already closed stream %d", s.id)
|
|
}
|
|
// Communicate the CLOSE stream signal to the other websocket endpoint.
|
|
err := s.conn.WriteMessage(gwebsocket.BinaryMessage, []byte{remotecommand.StreamClose, s.id})
|
|
s.conn = nil
|
|
return err
|
|
}
|
|
|
|
func (s *stream) Reset() error {
|
|
klog.V(4).Infof("Reset() on stream %d", s.id)
|
|
defer klog.V(4).Infof("Reset() done on stream %d", s.id)
|
|
s.Close()
|
|
return s.writePipe.Close()
|
|
}
|
|
|
|
func (s *stream) Headers() http.Header {
|
|
return s.headers
|
|
}
|
|
|
|
func (s *stream) Identifier() uint32 {
|
|
return uint32(s.id)
|
|
}
|
|
|
|
// heartbeat encasulates data necessary for the websocket ping/pong heartbeat. This
|
|
// heartbeat works by setting a read deadline on the websocket connection, then
|
|
// pushing this deadline into the future for every successful heartbeat. If the
|
|
// heartbeat "pong" fails to respond within the deadline, then the "NextReader()" call
|
|
// inside the "readDemuxLoop" will return an i/o error prompting a connection close
|
|
// and cleanup.
|
|
type heartbeat struct {
|
|
conn *gwebsocket.Conn
|
|
// period defines how often a "ping" heartbeat message is sent to the other endpoint
|
|
period time.Duration
|
|
// closing the "closer" channel will clean up the heartbeat timers
|
|
closer chan struct{}
|
|
// optional data to send with "ping" message
|
|
message []byte
|
|
// optionally received data message with "pong" message, same as sent with ping
|
|
pongMessage []byte
|
|
}
|
|
|
|
// newHeartbeat creates heartbeat structure encapsulating fields necessary to
|
|
// run the websocket connection ping/pong mechanism and sets up handlers on
|
|
// the websocket connection.
|
|
func newHeartbeat(conn *gwebsocket.Conn, period time.Duration, deadline time.Duration) *heartbeat {
|
|
h := &heartbeat{
|
|
conn: conn,
|
|
period: period,
|
|
closer: make(chan struct{}),
|
|
}
|
|
// Set up handler for receiving returned "pong" message from other endpoint
|
|
// by pushing the read deadline into the future. The "msg" received could
|
|
// be empty.
|
|
h.conn.SetPongHandler(func(msg string) error {
|
|
// Push the read deadline into the future.
|
|
klog.V(8).Infof("Pong message received (%s)--resetting read deadline", msg)
|
|
err := h.conn.SetReadDeadline(time.Now().Add(deadline))
|
|
if err != nil {
|
|
klog.Errorf("Websocket setting read deadline failed %v", err)
|
|
return err
|
|
}
|
|
if len(msg) > 0 {
|
|
h.pongMessage = []byte(msg)
|
|
}
|
|
return nil
|
|
})
|
|
// Set up handler to cleanup timers when this endpoint receives "Close" message.
|
|
closeHandler := h.conn.CloseHandler()
|
|
h.conn.SetCloseHandler(func(code int, text string) error {
|
|
close(h.closer)
|
|
return closeHandler(code, text)
|
|
})
|
|
return h
|
|
}
|
|
|
|
// setMessage is optional data sent with "ping" heartbeat. According to the websocket RFC
|
|
// this data sent with "ping" message should be returned in "pong" message.
|
|
func (h *heartbeat) setMessage(msg string) {
|
|
h.message = []byte(msg)
|
|
}
|
|
|
|
// start the heartbeat by setting up necesssary handlers and looping by sending "ping"
|
|
// message every "period" until the "closer" channel is closed.
|
|
func (h *heartbeat) start() {
|
|
// Loop to continually send "ping" message through websocket connection every "period".
|
|
t := time.NewTicker(h.period)
|
|
defer t.Stop()
|
|
for {
|
|
select {
|
|
case <-h.closer:
|
|
klog.V(8).Infof("closed channel--returning")
|
|
return
|
|
case <-t.C:
|
|
// "WriteControl" does not need to be protected by a mutex. According to
|
|
// gorilla/websockets library docs: "The Close and WriteControl methods can
|
|
// be called concurrently with all other methods."
|
|
if err := h.conn.WriteControl(gwebsocket.PingMessage, h.message, time.Now().Add(writeDeadline)); err == nil {
|
|
klog.V(8).Infof("Websocket Ping succeeeded")
|
|
} else {
|
|
klog.Errorf("Websocket Ping failed: %v", err)
|
|
if errors.Is(err, gwebsocket.ErrCloseSent) {
|
|
// we continue because c.conn.CloseChan will manage closing the connection already
|
|
continue
|
|
} else if e, ok := err.(net.Error); ok && e.Timeout() {
|
|
// Continue, in case this is a transient failure.
|
|
// c.conn.CloseChan above will tell us when the connection is
|
|
// actually closed.
|
|
// If Temporary function hadn't been deprecated, we would have used it.
|
|
// But most of temporary errors are timeout errors anyway.
|
|
continue
|
|
}
|
|
return
|
|
}
|
|
}
|
|
}
|
|
}
|