open-nomad/client/rpc.go

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

465 lines
13 KiB
Go
Raw Normal View History

2018-01-11 19:24:57 +00:00
package client
import (
"errors"
2018-01-11 21:23:57 +00:00
"io"
2018-01-11 19:24:57 +00:00
"net"
"net/rpc"
"strings"
2018-01-09 23:26:53 +00:00
"time"
2018-01-11 19:24:57 +00:00
2018-01-11 21:23:57 +00:00
metrics "github.com/armon/go-metrics"
"github.com/hashicorp/go-msgpack/codec"
2018-02-01 01:35:21 +00:00
"github.com/hashicorp/nomad/client/servers"
"github.com/hashicorp/nomad/helper"
inmem "github.com/hashicorp/nomad/helper/codec"
2018-01-12 21:58:44 +00:00
"github.com/hashicorp/nomad/helper/pool"
2018-01-09 23:26:53 +00:00
"github.com/hashicorp/nomad/nomad/structs"
2018-01-11 19:24:57 +00:00
)
// rpcEndpoints holds the RPC endpoints
type rpcEndpoints struct {
ClientStats *ClientStats
CSI *CSI
FileSystem *FileSystem
Allocations *Allocations
Agent *Agent
2018-01-11 19:24:57 +00:00
}
// ClientRPC is used to make a local, client only RPC call
func (c *Client) ClientRPC(method string, args interface{}, reply interface{}) error {
codec := &inmem.InmemCodec{
2018-01-11 19:24:57 +00:00
Method: method,
Args: args,
Reply: reply,
}
if err := c.rpcServer.ServeRequest(codec); err != nil {
return err
}
return codec.Err
}
2018-01-31 21:49:21 +00:00
// StreamingRpcHandler is used to make a local, client only streaming RPC
2018-01-21 01:19:55 +00:00
// call.
2018-01-31 21:49:21 +00:00
func (c *Client) StreamingRpcHandler(method string) (structs.StreamingRpcHandler, error) {
2018-01-21 01:19:55 +00:00
return c.streamingRpcs.GetHandler(method)
}
2018-01-11 19:24:57 +00:00
// RPC is used to forward an RPC call to a nomad server, or fail if no servers.
func (c *Client) RPC(method string, args interface{}, reply interface{}) error {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
conf := c.GetConfig()
2018-01-11 19:24:57 +00:00
// Invoke the RPCHandler if it exists
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if conf.RPCHandler != nil {
return conf.RPCHandler.RPC(method, args, reply)
2018-01-11 19:24:57 +00:00
}
// We will try to automatically retry requests that fail due to things like server unavailability
// but instead of retrying forever, lets have a solid upper-bound
deadline := time.Now()
// A reasonable amount of time for leader election. Note when servers forward() our RPC requests
// to the leader they may also allow for an RPCHoldTimeout while waiting for leader election.
// That's OK, we won't double up because we are using it here not as a sleep but
// as a hint to give up
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
deadline = deadline.Add(conf.RPCHoldTimeout)
// If its a blocking query, allow the time specified by the request
if info, ok := args.(structs.RPCInfo); ok {
deadline = deadline.Add(info.TimeToBlock())
}
2018-01-09 23:26:53 +00:00
TRY:
var rpcErr error
2018-01-09 23:26:53 +00:00
server := c.servers.FindServer()
if server == nil {
rpcErr = noServersErr
} else {
// Make the request.
rpcErr = c.connPool.RPC(c.Region(), server.Addr, method, args, reply)
if rpcErr == nil {
c.fireRpcRetryWatcher()
return nil
}
2018-01-11 19:24:57 +00:00
// If shutting down, exit without logging the error
select {
case <-c.shutdownCh:
return nil
default:
}
// Move off to another server, and see if we can retry.
c.rpcLogger.Error("error performing RPC to server", "error", rpcErr, "rpc", method, "server", server.Addr)
c.servers.NotifyFailedServer(server)
2018-01-11 19:24:57 +00:00
if !canRetry(args, rpcErr) {
c.rpcLogger.Error("error performing RPC to server which is not safe to automatically retry", "error", rpcErr, "rpc", method, "server", server.Addr)
return rpcErr
}
}
if time.Now().After(deadline) {
// Blocking queries are tricky. jitters and rpcholdtimes in multiple places can result in our server call taking longer than we wanted it to. For example:
// a block time of 5s may easily turn into the server blocking for 10s since it applies its own RPCHoldTime. If the server dies at t=7s we still want to retry
// so before we give up on blocking queries make one last attempt for an immediate answer
if info, ok := args.(structs.RPCInfo); ok && info.TimeToBlock() > 0 {
info.SetTimeToBlock(0)
return c.RPC(method, args, reply)
}
c.rpcLogger.Error("error performing RPC to server, deadline exceeded, cannot retry", "error", rpcErr, "rpc", method)
2018-01-09 23:26:53 +00:00
return rpcErr
}
// Wait to avoid thundering herd
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
timer, cancel := helper.NewSafeTimer(helper.RandomStagger(conf.RPCHoldTimeout / structs.JitterFraction))
defer cancel()
select {
case <-timer.C:
// If we are going to retry a blocking query we need to update the time to block so it finishes by our deadline.
if info, ok := args.(structs.RPCInfo); ok && info.TimeToBlock() > 0 {
2020-12-09 19:05:18 +00:00
newBlockTime := time.Until(deadline)
// We can get below 0 here on slow computers because we slept for jitter so at least try to get an immediate response
if newBlockTime < 0 {
newBlockTime = 0
}
info.SetTimeToBlock(newBlockTime)
return c.RPC(method, args, reply)
}
goto TRY
case <-c.shutdownCh:
2018-01-09 23:26:53 +00:00
}
return rpcErr
}
// canRetry returns true if the given situation is safe for a retry.
func canRetry(args interface{}, err error) bool {
2018-01-09 23:26:53 +00:00
// No leader errors are always safe to retry since no state could have
// been changed.
if structs.IsErrNoLeader(err) {
return true
}
// Reads are safe to retry for stream errors, such as if a server was
// being shut down.
info, ok := args.(structs.RPCInfo)
if ok && info.IsRead() && helper.IsErrEOF(err) {
2018-01-09 23:26:53 +00:00
return true
}
return false
2018-01-11 19:24:57 +00:00
}
2018-02-01 01:35:21 +00:00
// RemoteStreamingRpcHandler is used to make a streaming RPC call to a remote
// server.
func (c *Client) RemoteStreamingRpcHandler(method string) (structs.StreamingRpcHandler, error) {
server := c.servers.FindServer()
if server == nil {
return nil, noServersErr
}
conn, err := c.streamingRpcConn(server, method)
if err != nil {
// Move off to another server
c.rpcLogger.Error("error performing RPC to server", "error", err, "rpc", method, "server", server.Addr)
2018-02-01 01:35:21 +00:00
c.servers.NotifyFailedServer(server)
return nil, err
}
return bridgedStreamingRpcHandler(conn), nil
}
// bridgedStreamingRpcHandler creates a bridged streaming RPC handler by copying
// data between the two sides.
func bridgedStreamingRpcHandler(sideA io.ReadWriteCloser) structs.StreamingRpcHandler {
return func(sideB io.ReadWriteCloser) {
defer sideA.Close()
defer sideB.Close()
structs.Bridge(sideA, sideB)
}
}
// streamingRpcConn is used to retrieve a connection to a server to conduct a
// streaming RPC.
func (c *Client) streamingRpcConn(server *servers.Server, method string) (net.Conn, error) {
// Dial the server
conn, err := net.DialTimeout("tcp", server.Addr.String(), 10*time.Second)
if err != nil {
return nil, err
}
// Cast to TCPConn
if tcp, ok := conn.(*net.TCPConn); ok {
tcp.SetKeepAlive(true)
tcp.SetNoDelay(true)
}
// Check if TLS is enabled
2018-02-15 23:22:57 +00:00
c.tlsWrapLock.RLock()
tlsWrap := c.tlsWrap
c.tlsWrapLock.RUnlock()
if tlsWrap != nil {
// Switch the connection into TLS mode
if _, err := conn.Write([]byte{byte(pool.RpcTLS)}); err != nil {
conn.Close()
return nil, err
}
// Wrap the connection in a TLS client
tlsConn, err := tlsWrap(c.Region(), conn)
if err != nil {
conn.Close()
return nil, err
}
conn = tlsConn
}
2018-02-01 01:35:21 +00:00
// Write the multiplex byte to set the mode
if _, err := conn.Write([]byte{byte(pool.RpcStreaming)}); err != nil {
conn.Close()
return nil, err
}
// Send the header
encoder := codec.NewEncoder(conn, structs.MsgpackHandle)
decoder := codec.NewDecoder(conn, structs.MsgpackHandle)
2018-02-01 01:35:21 +00:00
header := structs.StreamingRpcHeader{
Method: method,
}
if err := encoder.Encode(header); err != nil {
conn.Close()
return nil, err
}
// Wait for the acknowledgement
var ack structs.StreamingRpcAck
if err := decoder.Decode(&ack); err != nil {
conn.Close()
return nil, err
}
if ack.Error != "" {
conn.Close()
return nil, errors.New(ack.Error)
}
2018-02-01 01:35:21 +00:00
return conn, nil
}
2018-01-11 19:24:57 +00:00
// setupClientRpc is used to setup the Client's RPC endpoints
func (c *Client) setupClientRpc(rpcs map[string]interface{}) {
2018-01-11 19:24:57 +00:00
// Create the RPC Server
c.rpcServer = rpc.NewServer()
// Initialize the RPC handlers
if rpcs != nil {
// override RPCs
for name, rpc := range rpcs {
c.rpcServer.RegisterName(name, rpc)
}
} else {
c.endpoints.ClientStats = &ClientStats{c}
c.endpoints.CSI = &CSI{c}
c.endpoints.FileSystem = NewFileSystemEndpoint(c)
c.endpoints.Allocations = NewAllocationsEndpoint(c)
c.endpoints.Agent = NewAgentEndpoint(c)
c.setupClientRpcServer(c.rpcServer)
}
2018-01-11 21:23:57 +00:00
go c.rpcConnListener()
}
// setupClientRpcServer is used to populate a client RPC server with endpoints.
func (c *Client) setupClientRpcServer(server *rpc.Server) {
// Register the endpoints
server.Register(c.endpoints.ClientStats)
server.Register(c.endpoints.CSI)
server.Register(c.endpoints.FileSystem)
server.Register(c.endpoints.Allocations)
server.Register(c.endpoints.Agent)
}
2018-01-11 21:23:57 +00:00
// rpcConnListener is a long lived function that listens for new connections
// being made on the connection pool and starts an RPC listener for each
// connection.
func (c *Client) rpcConnListener() {
// Make a channel for new connections.
conns := make(chan *pool.Conn, 4)
2018-01-11 21:23:57 +00:00
c.connPool.SetConnListener(conns)
for {
select {
case <-c.shutdownCh:
return
case conn, ok := <-conns:
2018-01-11 21:23:57 +00:00
if !ok {
continue
}
go c.listenConn(conn)
2018-01-11 21:23:57 +00:00
}
}
}
// listenConn is used to listen for connections being made from the server on
// pre-existing connection. This should be called in a goroutine.
func (c *Client) listenConn(conn *pool.Conn) {
2018-01-11 21:23:57 +00:00
for {
stream, err := conn.AcceptStream()
2018-01-11 21:23:57 +00:00
if err != nil {
if conn.IsClosed() {
2018-01-11 21:23:57 +00:00
return
}
2019-01-09 14:22:47 +00:00
c.rpcLogger.Error("failed to accept RPC conn", "error", err)
2018-01-11 21:23:57 +00:00
continue
}
go c.handleConn(stream)
2018-01-11 21:23:57 +00:00
metrics.IncrCounter([]string{"client", "rpc", "accept_conn"}, 1)
}
}
// handleConn is used to determine if this is a RPC or Streaming RPC connection and
// invoke the correct handler
2018-01-11 21:23:57 +00:00
func (c *Client) handleConn(conn net.Conn) {
// Read a single byte
buf := make([]byte, 1)
if _, err := conn.Read(buf); err != nil {
if err != io.EOF {
c.rpcLogger.Error("error reading byte", "error", err)
}
conn.Close()
return
}
// Switch on the byte
switch pool.RPCType(buf[0]) {
case pool.RpcNomad:
c.handleNomadConn(conn)
case pool.RpcStreaming:
c.handleStreamingConn(conn)
default:
c.rpcLogger.Error("unrecognized RPC byte", "byte", buf[0])
conn.Close()
return
}
}
// handleNomadConn is used to handle a single Nomad RPC connection.
func (c *Client) handleNomadConn(conn net.Conn) {
2018-01-11 21:23:57 +00:00
defer conn.Close()
2018-01-12 21:58:44 +00:00
rpcCodec := pool.NewServerCodec(conn)
2018-01-11 21:23:57 +00:00
for {
select {
case <-c.shutdownCh:
return
default:
}
if err := c.rpcServer.ServeRequest(rpcCodec); err != nil {
if err != io.EOF && !strings.Contains(err.Error(), "closed") {
c.rpcLogger.Error("error performing RPC", "error", err, "addr", conn.RemoteAddr())
2018-01-11 21:23:57 +00:00
metrics.IncrCounter([]string{"client", "rpc", "request_error"}, 1)
}
return
}
metrics.IncrCounter([]string{"client", "rpc", "request"}, 1)
}
2018-01-11 19:24:57 +00:00
}
// handleStreamingConn is used to handle a single Streaming Nomad RPC connection.
func (c *Client) handleStreamingConn(conn net.Conn) {
defer conn.Close()
// Decode the header
var header structs.StreamingRpcHeader
decoder := codec.NewDecoder(conn, structs.MsgpackHandle)
if err := decoder.Decode(&header); err != nil {
if err != io.EOF && !strings.Contains(err.Error(), "closed") {
c.rpcLogger.Error("error performing streaming RPC", "error", err, "addr", conn.RemoteAddr())
metrics.IncrCounter([]string{"client", "streaming_rpc", "request_error"}, 1)
}
return
}
ack := structs.StreamingRpcAck{}
handler, err := c.streamingRpcs.GetHandler(header.Method)
if err != nil {
c.rpcLogger.Error("streaming RPC error", "addr", conn.RemoteAddr(), "error", err)
metrics.IncrCounter([]string{"client", "streaming_rpc", "request_error"}, 1)
ack.Error = err.Error()
}
// Send the acknowledgement
encoder := codec.NewEncoder(conn, structs.MsgpackHandle)
if err := encoder.Encode(ack); err != nil {
conn.Close()
return
}
if ack.Error != "" {
return
}
// Invoke the handler
metrics.IncrCounter([]string{"client", "streaming_rpc", "request"}, 1)
handler(conn)
}
2018-01-11 19:24:57 +00:00
// resolveServer given a sever's address as a string, return it's resolved
// net.Addr or an error.
func resolveServer(s string) (net.Addr, error) {
const defaultClientPort = "4647" // default client RPC port
host, port, err := net.SplitHostPort(s)
if err != nil {
if strings.Contains(err.Error(), "missing port") {
host = s
port = defaultClientPort
} else {
return nil, err
}
}
return net.ResolveTCPAddr("tcp", net.JoinHostPort(host, port))
}
2018-01-09 23:26:53 +00:00
// Ping is used to ping a particular server and returns whether it is healthy or
// a potential error.
2018-01-25 02:00:21 +00:00
func (c *Client) Ping(srv net.Addr) error {
2018-01-09 23:26:53 +00:00
var reply struct{}
err := c.connPool.RPC(c.Region(), srv, "Status.Ping", struct{}{}, &reply)
2018-01-25 02:00:21 +00:00
return err
2018-01-09 23:26:53 +00:00
}
2018-04-04 01:05:28 +00:00
// rpcRetryWatcher returns a channel that will be closed if an event happens
// such that we expect the next RPC to be successful.
func (c *Client) rpcRetryWatcher() <-chan struct{} {
c.rpcRetryLock.Lock()
defer c.rpcRetryLock.Unlock()
if c.rpcRetryCh == nil {
c.rpcRetryCh = make(chan struct{})
}
return c.rpcRetryCh
}
// fireRpcRetryWatcher causes any RPC retryloops to retry their RPCs because we
// believe the will be successful.
func (c *Client) fireRpcRetryWatcher() {
c.rpcRetryLock.Lock()
defer c.rpcRetryLock.Unlock()
if c.rpcRetryCh != nil {
close(c.rpcRetryCh)
c.rpcRetryCh = nil
}
}