2018-04-25 21:21:03 +00:00
|
|
|
package proxy
|
|
|
|
|
|
|
|
import (
|
2018-04-25 23:54:00 +00:00
|
|
|
"fmt"
|
|
|
|
"log"
|
|
|
|
"os"
|
2018-04-25 21:21:03 +00:00
|
|
|
"os/exec"
|
2018-05-02 18:02:58 +00:00
|
|
|
"reflect"
|
2018-05-03 20:56:42 +00:00
|
|
|
"strconv"
|
2018-04-25 23:54:00 +00:00
|
|
|
"sync"
|
2018-04-26 00:39:32 +00:00
|
|
|
"time"
|
2018-05-03 20:56:42 +00:00
|
|
|
|
|
|
|
"github.com/hashicorp/consul/lib/file"
|
2018-05-03 22:46:00 +00:00
|
|
|
"github.com/mitchellh/mapstructure"
|
2018-04-26 00:39:32 +00:00
|
|
|
)
|
|
|
|
|
|
|
|
// Constants related to restart timers with the daemon mode proxies. At some
|
|
|
|
// point we will probably want to expose these knobs to an end user, but
|
|
|
|
// reasonable defaults are chosen.
|
|
|
|
const (
|
|
|
|
DaemonRestartHealthy = 10 * time.Second // time before considering healthy
|
|
|
|
DaemonRestartBackoffMin = 3 // 3 attempts before backing off
|
|
|
|
DaemonRestartMaxWait = 1 * time.Minute // maximum backoff wait time
|
2018-04-25 21:21:03 +00:00
|
|
|
)
|
|
|
|
|
|
|
|
// Daemon is a long-running proxy process. It is expected to keep running
|
|
|
|
// and to use blocking queries to detect changes in configuration, certs,
|
|
|
|
// and more.
|
|
|
|
//
|
|
|
|
// Consul will ensure that if the daemon crashes, that it is restarted.
|
|
|
|
type Daemon struct {
|
|
|
|
// Command is the command to execute to start this daemon. This must
|
|
|
|
// be a Cmd that isn't yet started.
|
|
|
|
Command *exec.Cmd
|
|
|
|
|
2018-05-06 15:41:25 +00:00
|
|
|
// ProxyId is the ID of the proxy service. This is required for API
|
|
|
|
// requests (along with the token) and is passed via env var.
|
|
|
|
ProxyId string
|
|
|
|
|
2018-04-25 21:21:03 +00:00
|
|
|
// ProxyToken is the special local-only ACL token that allows a proxy
|
|
|
|
// to communicate to the Connect-specific endpoints.
|
|
|
|
ProxyToken string
|
2018-04-25 23:54:00 +00:00
|
|
|
|
|
|
|
// Logger is where logs will be sent around the management of this
|
|
|
|
// daemon. The actual logs for the daemon itself will be sent to
|
|
|
|
// a file.
|
|
|
|
Logger *log.Logger
|
|
|
|
|
2018-05-03 20:56:42 +00:00
|
|
|
// PidPath is the path where a pid file will be created storing the
|
|
|
|
// pid of the active process. If this is empty then a pid-file won't
|
|
|
|
// be created. Under erroneous conditions, the pid file may not be
|
|
|
|
// created but the error will be logged to the Logger.
|
|
|
|
PidPath string
|
|
|
|
|
2018-05-02 19:25:00 +00:00
|
|
|
// For tests, they can set this to change the default duration to wait
|
|
|
|
// for a graceful quit.
|
|
|
|
gracefulWait time.Duration
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
// process is the started process
|
2018-05-02 19:25:00 +00:00
|
|
|
lock sync.Mutex
|
|
|
|
stopped bool
|
|
|
|
stopCh chan struct{}
|
|
|
|
exitedCh chan struct{}
|
|
|
|
process *os.Process
|
2018-04-25 21:21:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Start starts the daemon and keeps it running.
|
2018-04-25 23:54:00 +00:00
|
|
|
//
|
|
|
|
// This function returns after the process is successfully started.
|
2018-04-25 21:21:03 +00:00
|
|
|
func (p *Daemon) Start() error {
|
2018-04-25 23:54:00 +00:00
|
|
|
p.lock.Lock()
|
|
|
|
defer p.lock.Unlock()
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// A stopped proxy cannot be restarted
|
|
|
|
if p.stopped {
|
|
|
|
return fmt.Errorf("stopped")
|
2018-04-25 23:54:00 +00:00
|
|
|
}
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// If we're already running, that is okay
|
|
|
|
if p.process != nil {
|
|
|
|
return nil
|
2018-04-25 23:54:00 +00:00
|
|
|
}
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// Setup our stop channel
|
2018-04-25 23:54:00 +00:00
|
|
|
stopCh := make(chan struct{})
|
2018-05-02 19:25:00 +00:00
|
|
|
exitedCh := make(chan struct{})
|
2018-04-25 23:54:00 +00:00
|
|
|
p.stopCh = stopCh
|
2018-05-02 19:25:00 +00:00
|
|
|
p.exitedCh = exitedCh
|
2018-04-25 23:54:00 +00:00
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// Start the loop.
|
2018-05-02 19:25:00 +00:00
|
|
|
go p.keepAlive(stopCh, exitedCh)
|
2018-04-25 23:54:00 +00:00
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// keepAlive starts and keeps the configured process alive until it
|
|
|
|
// is stopped via Stop.
|
2018-05-02 19:25:00 +00:00
|
|
|
func (p *Daemon) keepAlive(stopCh <-chan struct{}, exitedCh chan<- struct{}) {
|
|
|
|
defer close(exitedCh)
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
p.lock.Lock()
|
|
|
|
process := p.process
|
|
|
|
p.lock.Unlock()
|
|
|
|
|
2018-04-26 00:39:32 +00:00
|
|
|
// attemptsDeadline is the time at which we consider the daemon to have
|
|
|
|
// been alive long enough that we can reset the attempt counter.
|
|
|
|
//
|
|
|
|
// attempts keeps track of the number of restart attempts we've had and
|
|
|
|
// is used to calculate the wait time using an exponential backoff.
|
|
|
|
var attemptsDeadline time.Time
|
|
|
|
var attempts uint
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
for {
|
|
|
|
if process == nil {
|
2018-04-26 00:39:32 +00:00
|
|
|
// If we're passed the attempt deadline then reset the attempts
|
|
|
|
if !attemptsDeadline.IsZero() && time.Now().After(attemptsDeadline) {
|
|
|
|
attempts = 0
|
|
|
|
}
|
|
|
|
attemptsDeadline = time.Now().Add(DaemonRestartHealthy)
|
|
|
|
attempts++
|
|
|
|
|
|
|
|
// Calculate the exponential backoff and wait if we have to
|
|
|
|
if attempts > DaemonRestartBackoffMin {
|
|
|
|
waitTime := (1 << (attempts - DaemonRestartBackoffMin)) * time.Second
|
|
|
|
if waitTime > DaemonRestartMaxWait {
|
|
|
|
waitTime = DaemonRestartMaxWait
|
|
|
|
}
|
|
|
|
|
|
|
|
if waitTime > 0 {
|
|
|
|
p.Logger.Printf(
|
|
|
|
"[WARN] agent/proxy: waiting %s before restarting daemon",
|
|
|
|
waitTime)
|
2018-04-27 17:44:16 +00:00
|
|
|
|
|
|
|
timer := time.NewTimer(waitTime)
|
|
|
|
select {
|
|
|
|
case <-timer.C:
|
|
|
|
// Timer is up, good!
|
|
|
|
|
|
|
|
case <-stopCh:
|
|
|
|
// During our backoff wait, we've been signalled to
|
|
|
|
// quit, so just quit.
|
|
|
|
timer.Stop()
|
|
|
|
return
|
|
|
|
}
|
2018-04-26 00:39:32 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
p.lock.Lock()
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// If we gracefully stopped then don't restart.
|
|
|
|
if p.stopped {
|
2018-04-25 23:54:00 +00:00
|
|
|
p.lock.Unlock()
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// Process isn't started currently. We're restarting. Start it
|
|
|
|
// and save the process if we have it.
|
2018-04-25 23:54:00 +00:00
|
|
|
var err error
|
|
|
|
process, err = p.start()
|
2018-04-27 17:44:16 +00:00
|
|
|
if err == nil {
|
|
|
|
p.process = process
|
|
|
|
}
|
|
|
|
p.lock.Unlock()
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
if err != nil {
|
|
|
|
p.Logger.Printf("[ERR] agent/proxy: error restarting daemon: %s", err)
|
2018-04-27 17:44:16 +00:00
|
|
|
continue
|
2018-04-25 23:54:00 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2018-05-04 01:25:11 +00:00
|
|
|
// Wait for the process to exit. Note that if we restored this proxy
|
|
|
|
// then Wait will always fail because we likely aren't the parent
|
|
|
|
// process. Therefore, we do an extra sanity check after to use other
|
|
|
|
// syscalls to verify the process is truly dead.
|
2018-05-02 18:38:18 +00:00
|
|
|
ps, err := process.Wait()
|
2018-05-04 01:25:11 +00:00
|
|
|
if _, err := findProcess(process.Pid); err == nil {
|
|
|
|
select {
|
|
|
|
case <-time.After(1 * time.Second):
|
|
|
|
// We want a busy loop, but not too busy. 1 second between
|
|
|
|
// detecting a process death seems reasonable.
|
|
|
|
|
|
|
|
case <-stopCh:
|
|
|
|
// If we receive a stop request we want to exit immediately.
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
process = nil
|
2018-05-02 18:38:18 +00:00
|
|
|
if err != nil {
|
|
|
|
p.Logger.Printf("[INFO] agent/proxy: daemon exited with error: %s", err)
|
2018-05-02 18:47:57 +00:00
|
|
|
} else if status, ok := exitStatus(ps); ok {
|
|
|
|
p.Logger.Printf("[INFO] agent/proxy: daemon exited with exit code: %d", status)
|
2018-05-02 18:38:18 +00:00
|
|
|
}
|
2018-04-25 23:54:00 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// start starts and returns the process. This will create a copy of the
|
|
|
|
// configured *exec.Command with the modifications documented on Daemon
|
|
|
|
// such as setting the proxy token environmental variable.
|
|
|
|
func (p *Daemon) start() (*os.Process, error) {
|
|
|
|
cmd := *p.Command
|
|
|
|
|
|
|
|
// Add the proxy token to the environment. We first copy the env because
|
|
|
|
// it is a slice and therefore the "copy" above will only copy the slice
|
|
|
|
// reference. We allocate an exactly sized slice.
|
|
|
|
cmd.Env = make([]string, len(p.Command.Env), len(p.Command.Env)+1)
|
|
|
|
copy(cmd.Env, p.Command.Env)
|
2018-05-06 15:41:25 +00:00
|
|
|
cmd.Env = append(cmd.Env,
|
|
|
|
fmt.Sprintf("%s=%s", EnvProxyId, p.ProxyId),
|
|
|
|
fmt.Sprintf("%s=%s", EnvProxyToken, p.ProxyToken))
|
2018-04-25 23:54:00 +00:00
|
|
|
|
2018-05-02 18:47:57 +00:00
|
|
|
// Args must always contain a 0 entry which is usually the executed binary.
|
|
|
|
// To be safe and a bit more robust we default this, but only to prevent
|
|
|
|
// a panic below.
|
|
|
|
if len(cmd.Args) == 0 {
|
|
|
|
cmd.Args = []string{cmd.Path}
|
|
|
|
}
|
|
|
|
|
2018-04-25 23:54:00 +00:00
|
|
|
// Start it
|
2018-05-02 18:47:57 +00:00
|
|
|
p.Logger.Printf("[DEBUG] agent/proxy: starting proxy: %q %#v", cmd.Path, cmd.Args[1:])
|
2018-05-03 20:56:42 +00:00
|
|
|
if err := cmd.Start(); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
|
|
|
// Write the pid file. This might error and that's okay.
|
|
|
|
if p.PidPath != "" {
|
|
|
|
pid := strconv.FormatInt(int64(cmd.Process.Pid), 10)
|
|
|
|
if err := file.WriteAtomic(p.PidPath, []byte(pid)); err != nil {
|
|
|
|
p.Logger.Printf(
|
|
|
|
"[DEBUG] agent/proxy: error writing pid file %q: %s",
|
|
|
|
p.PidPath, err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return cmd.Process, nil
|
2018-04-25 23:54:00 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Stop stops the daemon.
|
|
|
|
//
|
|
|
|
// This will attempt a graceful stop (SIGINT) before force killing the
|
|
|
|
// process (SIGKILL). In either case, the process won't be automatically
|
|
|
|
// restarted unless Start is called again.
|
|
|
|
//
|
|
|
|
// This is safe to call multiple times. If the daemon is already stopped,
|
|
|
|
// then this returns no error.
|
|
|
|
func (p *Daemon) Stop() error {
|
|
|
|
p.lock.Lock()
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// If we're already stopped or never started, then no problem.
|
|
|
|
if p.stopped || p.process == nil {
|
|
|
|
// In the case we never even started, calling Stop makes it so
|
|
|
|
// that we can't ever start in the future, either, so mark this.
|
|
|
|
p.stopped = true
|
2018-05-02 19:25:00 +00:00
|
|
|
p.lock.Unlock()
|
2018-04-25 23:54:00 +00:00
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2018-04-27 17:44:16 +00:00
|
|
|
// Note that we've stopped
|
|
|
|
p.stopped = true
|
|
|
|
close(p.stopCh)
|
2018-05-02 19:25:00 +00:00
|
|
|
process := p.process
|
|
|
|
p.lock.Unlock()
|
2018-04-25 23:54:00 +00:00
|
|
|
|
2018-05-02 19:25:00 +00:00
|
|
|
gracefulWait := p.gracefulWait
|
|
|
|
if gracefulWait == 0 {
|
|
|
|
gracefulWait = 5 * time.Second
|
|
|
|
}
|
|
|
|
|
2018-05-04 16:44:59 +00:00
|
|
|
// Defer removing the pid file. Even under error conditions we
|
|
|
|
// delete the pid file since Stop means that the manager is no
|
|
|
|
// longer managing this proxy and therefore nothing else will ever
|
|
|
|
// clean it up.
|
|
|
|
if p.PidPath != "" {
|
|
|
|
defer func() {
|
|
|
|
if err := os.Remove(p.PidPath); err != nil && !os.IsNotExist(err) {
|
|
|
|
p.Logger.Printf(
|
|
|
|
"[DEBUG] agent/proxy: error removing pid file %q: %s",
|
|
|
|
p.PidPath, err)
|
|
|
|
}
|
|
|
|
}()
|
|
|
|
}
|
|
|
|
|
2018-05-02 19:25:00 +00:00
|
|
|
// First, try a graceful stop
|
|
|
|
err := process.Signal(os.Interrupt)
|
|
|
|
if err == nil {
|
|
|
|
select {
|
|
|
|
case <-p.exitedCh:
|
|
|
|
// Success!
|
|
|
|
return nil
|
|
|
|
|
|
|
|
case <-time.After(gracefulWait):
|
|
|
|
// Interrupt didn't work
|
|
|
|
}
|
|
|
|
}
|
2018-04-25 23:54:00 +00:00
|
|
|
|
2018-05-02 19:25:00 +00:00
|
|
|
// Graceful didn't work, forcibly kill
|
|
|
|
return process.Kill()
|
2018-04-25 21:21:03 +00:00
|
|
|
}
|
2018-05-02 18:02:58 +00:00
|
|
|
|
2018-05-03 22:46:00 +00:00
|
|
|
// stopKeepAlive is like Stop but keeps the process running. This is
|
|
|
|
// used only for tests.
|
|
|
|
func (p *Daemon) stopKeepAlive() error {
|
|
|
|
p.lock.Lock()
|
2018-05-04 15:09:44 +00:00
|
|
|
defer p.lock.Unlock()
|
2018-05-03 22:46:00 +00:00
|
|
|
|
|
|
|
// If we're already stopped or never started, then no problem.
|
|
|
|
if p.stopped || p.process == nil {
|
|
|
|
p.stopped = true
|
|
|
|
p.lock.Unlock()
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// Note that we've stopped
|
|
|
|
p.stopped = true
|
|
|
|
close(p.stopCh)
|
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2018-05-02 18:02:58 +00:00
|
|
|
// Equal implements Proxy to check for equality.
|
|
|
|
func (p *Daemon) Equal(raw Proxy) bool {
|
|
|
|
p2, ok := raw.(*Daemon)
|
|
|
|
if !ok {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
|
|
|
// We compare equality on a subset of the command configuration
|
|
|
|
return p.ProxyToken == p2.ProxyToken &&
|
|
|
|
p.Command.Path == p2.Command.Path &&
|
|
|
|
p.Command.Dir == p2.Command.Dir &&
|
|
|
|
reflect.DeepEqual(p.Command.Args, p2.Command.Args) &&
|
|
|
|
reflect.DeepEqual(p.Command.Env, p2.Command.Env)
|
|
|
|
}
|
2018-05-03 22:46:00 +00:00
|
|
|
|
|
|
|
// MarshalSnapshot implements Proxy
|
|
|
|
func (p *Daemon) MarshalSnapshot() map[string]interface{} {
|
|
|
|
p.lock.Lock()
|
|
|
|
defer p.lock.Unlock()
|
|
|
|
|
|
|
|
// If we're stopped or have no process, then nothing to snapshot.
|
|
|
|
if p.stopped || p.process == nil {
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
return map[string]interface{}{
|
|
|
|
"Pid": p.process.Pid,
|
|
|
|
"CommandPath": p.Command.Path,
|
|
|
|
"CommandArgs": p.Command.Args,
|
|
|
|
"CommandDir": p.Command.Dir,
|
|
|
|
"CommandEnv": p.Command.Env,
|
|
|
|
"ProxyToken": p.ProxyToken,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// UnmarshalSnapshot implements Proxy
|
|
|
|
func (p *Daemon) UnmarshalSnapshot(m map[string]interface{}) error {
|
|
|
|
var s daemonSnapshot
|
|
|
|
if err := mapstructure.Decode(m, &s); err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
|
|
|
|
p.lock.Lock()
|
|
|
|
defer p.lock.Unlock()
|
|
|
|
|
|
|
|
// Set the basic fields
|
|
|
|
p.ProxyToken = s.ProxyToken
|
|
|
|
p.Command = &exec.Cmd{
|
|
|
|
Path: s.CommandPath,
|
|
|
|
Args: s.CommandArgs,
|
|
|
|
Dir: s.CommandDir,
|
|
|
|
Env: s.CommandEnv,
|
|
|
|
}
|
|
|
|
|
2018-05-03 22:55:49 +00:00
|
|
|
// FindProcess on many systems returns no error even if the process
|
|
|
|
// is now dead. We perform an extra check that the process is alive.
|
2018-05-04 01:25:11 +00:00
|
|
|
proc, err := findProcess(s.Pid)
|
|
|
|
if err != nil {
|
2018-05-03 22:55:49 +00:00
|
|
|
return err
|
|
|
|
}
|
2018-05-03 22:46:00 +00:00
|
|
|
|
|
|
|
// "Start it"
|
|
|
|
stopCh := make(chan struct{})
|
|
|
|
exitedCh := make(chan struct{})
|
|
|
|
p.stopCh = stopCh
|
|
|
|
p.exitedCh = exitedCh
|
|
|
|
p.process = proc
|
|
|
|
go p.keepAlive(stopCh, exitedCh)
|
|
|
|
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// daemonSnapshot is the structure of the marshalled data for snapshotting.
|
2018-05-06 15:41:25 +00:00
|
|
|
//
|
|
|
|
// Note we don't have to store the ProxyId because this is stored directly
|
|
|
|
// within the manager snapshot and is restored automatically.
|
2018-05-03 22:46:00 +00:00
|
|
|
type daemonSnapshot struct {
|
|
|
|
// Pid of the process. This is the only value actually required to
|
|
|
|
// regain mangement control. The remainder values are for Equal.
|
|
|
|
Pid int
|
|
|
|
|
|
|
|
// Command information
|
|
|
|
CommandPath string
|
|
|
|
CommandArgs []string
|
|
|
|
CommandDir string
|
|
|
|
CommandEnv []string
|
|
|
|
|
|
|
|
// NOTE(mitchellh): longer term there are discussions/plans to only
|
|
|
|
// store the hash of the token but for now we need the full token in
|
|
|
|
// case the process dies and has to be restarted.
|
|
|
|
ProxyToken string
|
|
|
|
}
|