// newWatcher creates a new watcher.
func newWatcher(c *config.Config, clients *dep.ClientSet, once bool) (*watch.Watcher, error) {
log.Printf("[INFO] (runner) creating watcher")
w, err := watch.NewWatcher(&watch.NewWatcherInput{
Clients: clients,
MaxStale: config.TimeDurationVal(c.MaxStale),
Once: once,
RenewVault: config.StringPresent(c.Vault.Token) && config.BoolVal(c.Vault.RenewToken),
RetryFuncConsul: watch.RetryFunc(c.Consul.Retry.RetryFunc()),
// TODO: Add a sane default retry - right now this only affects "local"
// dependencies like reading a file from disk.
RetryFuncDefault: nil,
RetryFuncVault: watch.RetryFunc(c.Vault.Retry.RetryFunc()),
})
if err != nil {
return nil, errors.Wrap(err, "runner")
}
return w, nil
}
// Run iterates over each template in this Runner and conditionally executes
// the template rendering and command execution.
//
// The template is rendered atomicly. If and only if the template render
// completes successfully, the optional commands will be executed, if given.
// Please note that all templates are rendered **and then** any commands are
// executed.
func (r *Runner) Run() error {
log.Printf("[INFO] (runner) initiating run")
var wouldRenderAny, renderedAny bool
var commands []*config.TemplateConfig
depsMap := make(map[string]dep.Dependency)
for _, tmpl := range r.templates {
log.Printf("[DEBUG] (runner) checking template %s", tmpl.ID())
// Create the event
event := &RenderEvent{
Template: tmpl,
TemplateConfigs: r.templateConfigsFor(tmpl),
}
// Check if we are currently the leader instance
isLeader := true
if r.dedup != nil {
isLeader = r.dedup.IsLeader(tmpl)
}
// If we are in once mode and this template was already rendered, move
// onto the next one. We do not want to re-render the template if we are
// in once mode, and we certainly do not want to re-run any commands.
if r.once {
r.renderEventsLock.RLock()
_, rendered := r.renderEvents[tmpl.ID()]
r.renderEventsLock.RUnlock()
if rendered {
log.Printf("[DEBUG] (runner) once mode and already rendered")
continue
}
}
// Attempt to render the template, returning any missing dependencies and
// the rendered contents. If there are any missing dependencies, the
// contents cannot be rendered or trusted!
result, err := tmpl.Execute(&template.ExecuteInput{
Brain: r.brain,
Env: r.childEnv(),
})
if err != nil {
return errors.Wrap(err, tmpl.Source())
}
// Grab the list of used and missing dependencies.
missing, used := result.Missing, result.Used
// Add the dependency to the list of dependencies for this runner.
for _, d := range used.List() {
// If we've taken over leadership for a template, we may have data
// that is cached, but not have the watcher. We must treat this as
// missing so that we create the watcher and re-run the template.
if isLeader && !r.watcher.Watching(d) {
missing.Add(d)
}
if _, ok := depsMap[d.String()]; !ok {
depsMap[d.String()] = d
}
}
// Diff any missing dependencies the template reported with dependencies
// the watcher is watching.
unwatched := new(dep.Set)
for _, d := range missing.List() {
if !r.watcher.Watching(d) {
unwatched.Add(d)
}
}
// If there are unwatched dependencies, start the watcher and move onto the
// next one.
if l := unwatched.Len(); l > 0 {
log.Printf("[DEBUG] (runner) was not watching %d dependencies", l)
for _, d := range unwatched.List() {
// If we are deduplicating, we must still handle non-sharable
// dependencies, since those will be ignored.
if isLeader || !d.CanShare() {
r.watcher.Add(d)
}
}
continue
}
// If the template is missing data for some dependencies then we are not
// ready to render and need to move on to the next one.
if l := missing.Len(); l > 0 {
log.Printf("[DEBUG] (runner) missing data for %d dependencies", l)
continue
}
// Trigger an update of the de-duplicaiton manager
if r.dedup != nil && isLeader {
if err := r.dedup.UpdateDeps(tmpl, used.List()); err != nil {
log.Printf("[ERR] (runner) failed to update dependency data for de-duplication: %v", err)
}
}
// Update event information with dependencies.
event.MissingDeps = missing
event.UnwatchedDeps = unwatched
event.UsedDeps = used
// If quiescence is activated, start/update the timers and loop back around.
// We do not want to render the templates yet.
if q, ok := r.quiescenceMap[tmpl.ID()]; ok {
q.tick()
continue
}
// For each template configuration that is tied to this template, attempt to
// render it to disk and accumulate commands for later use.
for _, templateConfig := range r.templateConfigsFor(tmpl) {
log.Printf("[DEBUG] (runner) rendering %s", templateConfig.Display())
// Render the template, taking dry mode into account
result, err := Render(&RenderInput{
Backup: config.BoolVal(templateConfig.Backup),
Contents: result.Output,
Dry: r.dry,
DryStream: r.outStream,
Path: config.StringVal(templateConfig.Destination),
Perms: config.FileModeVal(templateConfig.Perms),
})
if err != nil {
return errors.Wrap(err, "error rendering "+templateConfig.Display())
}
// If we would have rendered this template (but we did not because the
// contents were the same or something), we should consider this template
// rendered even though the contents on disk have not been updated. We
// will not fire commands unless the template was _actually_ rendered to
// disk though.
if result.WouldRender {
// This event would have rendered
event.WouldRender = true
// Record that at least one template would have been rendered.
wouldRenderAny = true
}
// If we _actually_ rendered the template to disk, we want to run the
// appropriate commands.
if result.DidRender {
log.Printf("[INFO] (runner) rendered %s", templateConfig.Display())
// This event did render
event.DidRender = true
// Record that at least one template was rendered.
renderedAny = true
if !r.dry {
// If the template was rendered (changed) and we are not in dry-run mode,
// aggregate commands, ignoring previously known commands
//
// Future-self Q&A: Why not use a map for the commands instead of an
// array with an expensive lookup option? Well I'm glad you asked that
// future-self! One of the API promises is that commands are executed
// in the order in which they are provided in the TemplateConfig
// definitions. If we inserted commands into a map, we would lose that
// relative ordering and people would be unhappy.
// if config.StringPresent(ctemplate.Command)
if c := config.StringVal(templateConfig.Exec.Command); c != "" {
existing := findCommand(templateConfig, commands)
if existing != nil {
log.Printf("[DEBUG] (runner) skipping command %q from %s (already appended from %s)",
c, templateConfig.Display(), existing.Display())
} else {
log.Printf("[DEBUG] (runner) appending command %q from %s",
c, templateConfig.Display())
commands = append(commands, templateConfig)
}
}
}
}
}
// Send updated render event
r.renderEventsLock.Lock()
event.UpdatedAt = time.Now().UTC()
r.renderEvents[tmpl.ID()] = event
r.renderEventsLock.Unlock()
}
// Check if we need to deliver any rendered signals
if wouldRenderAny || renderedAny {
// Send the signal that a template got rendered
select {
case r.renderedCh <- struct{}{}:
default:
}
}
// Perform the diff and update the known dependencies.
r.diffAndUpdateDeps(depsMap)
// Execute each command in sequence, collecting any errors that occur - this
// ensures all commands execute at least once.
var errs []error
for _, t := range commands {
command := config.StringVal(t.Exec.Command)
log.Printf("[INFO] (runner) executing command %q from %s", command, t.Display())
env := t.Exec.Env.Copy()
env.Custom = append(r.childEnv(), env.Custom...)
if _, err := spawnChild(&spawnChildInput{
Stdin: r.inStream,
Stdout: r.outStream,
Stderr: r.errStream,
Command: command,
Env: env.Env(),
Timeout: config.TimeDurationVal(t.Exec.Timeout),
ReloadSignal: config.SignalVal(t.Exec.ReloadSignal),
KillSignal: config.SignalVal(t.Exec.KillSignal),
KillTimeout: config.TimeDurationVal(t.Exec.KillTimeout),
Splay: config.TimeDurationVal(t.Exec.Splay),
}); err != nil {
s := fmt.Sprintf("failed to execute command %q from %s", command, t.Display())
errs = append(errs, errors.Wrap(err, s))
}
}
// If we got this far and have a child process, we need to send the reload
// signal to the child process.
if renderedAny && r.child != nil {
r.childLock.RLock()
if err := r.child.Reload(); err != nil {
errs = append(errs, err)
}
r.childLock.RUnlock()
}
// If any errors were returned, convert them to an ErrorList for human
// readability.
if len(errs) != 0 {
var result *multierror.Error
for _, err := range errs {
result = multierror.Append(result, err)
}
return result.ErrorOrNil()
}
return nil
}
// Start begins the polling for this runner. Any errors that occur will cause
// this function to push an item onto the runner's error channel and the halt
// execution. This function is blocking and should be called as a goroutine.
func (r *Runner) Start() {
log.Printf("[INFO] (runner) starting")
// Create the pid before doing anything.
if err := r.storePid(); err != nil {
r.ErrCh <- err
return
}
// Start the de-duplication manager
var dedupCh <-chan struct{}
if r.dedup != nil {
if err := r.dedup.Start(); err != nil {
r.ErrCh <- err
return
}
dedupCh = r.dedup.UpdateCh()
}
// Setup the child process exit channel
var childExitCh <-chan int
// Fire an initial run to parse all the templates and setup the first-pass
// dependencies. This also forces any templates that have no dependencies to
// be rendered immediately (since they are already renderable).
log.Printf("[DEBUG] (runner) running initial templates")
if err := r.Run(); err != nil {
r.ErrCh <- err
return
}
for {
// Enable quiescence for all templates if we have specified wait
// intervals.
NEXT_Q:
for _, t := range r.templates {
if _, ok := r.quiescenceMap[t.ID()]; ok {
continue NEXT_Q
}
for _, c := range r.templateConfigsFor(t) {
if *c.Wait.Enabled {
log.Printf("[DEBUG] (runner) enabling template-specific quiescence for %q", t.ID())
r.quiescenceMap[t.ID()] = newQuiescence(
r.quiescenceCh, *c.Wait.Min, *c.Wait.Max, t)
continue NEXT_Q
}
}
if *r.config.Wait.Enabled {
log.Printf("[DEBUG] (runner) enabling global quiescence for %q", t.ID())
r.quiescenceMap[t.ID()] = newQuiescence(
r.quiescenceCh, *r.config.Wait.Min, *r.config.Wait.Max, t)
continue NEXT_Q
}
}
// Warn the user if they are watching too many dependencies.
if r.watcher.Size() > saneViewLimit {
log.Printf("[WARN] (runner) watching %d dependencies - watching this "+
"many dependencies could DDoS your consul cluster", r.watcher.Size())
} else {
log.Printf("[DEBUG] (runner) watching %d dependencies", r.watcher.Size())
}
if r.allTemplatesRendered() {
// If an exec command was given and a command is not currently running,
// spawn the child process for supervision.
if config.StringPresent(r.config.Exec.Command) {
// Lock the child because we are about to check if it exists.
r.childLock.Lock()
if r.child == nil {
env := r.config.Exec.Env.Copy()
env.Custom = append(r.childEnv(), env.Custom...)
child, err := spawnChild(&spawnChildInput{
Stdin: r.inStream,
Stdout: r.outStream,
Stderr: r.errStream,
Command: config.StringVal(r.config.Exec.Command),
Env: env.Env(),
ReloadSignal: config.SignalVal(r.config.Exec.ReloadSignal),
KillSignal: config.SignalVal(r.config.Exec.KillSignal),
KillTimeout: config.TimeDurationVal(r.config.Exec.KillTimeout),
Splay: config.TimeDurationVal(r.config.Exec.Splay),
})
if err != nil {
r.ErrCh <- err
r.childLock.Unlock()
return
}
r.child = child
}
// Unlock the child, we are done now.
r.childLock.Unlock()
// It's possible that we didn't start a process, in which case no
// channel is returned. If we did get a new exitCh, that means a child
// was spawned, so we need to watch a new exitCh. It is also possible
// that during a run, the child process was restarted, which means a
// new exit channel should be used.
nexitCh := r.child.ExitCh()
if nexitCh != nil {
childExitCh = nexitCh
}
}
// If we are running in once mode and all our templates are rendered,
// then we should exit here.
if r.once {
log.Printf("[INFO] (runner) once mode and all templates rendered")
if r.child != nil {
r.stopDedup()
r.stopWatcher()
log.Printf("[INFO] (runner) waiting for child process to exit")
select {
case c := <-childExitCh:
log.Printf("[INFO] (runner) child process died")
r.ErrCh <- NewErrChildDied(c)
return
case <-r.DoneCh:
}
}
r.Stop()
return
}
}
OUTER:
select {
case view := <-r.watcher.DataCh():
// Receive this update
r.Receive(view.Dependency(), view.Data())
// Drain all dependency data. Given a large number of dependencies, it is
// feasible that we have data for more than one of them. Instead of
// wasting CPU cycles rendering templates when we have more dependencies
// waiting to be added to the brain, we drain the entire buffered channel
// on the watcher and then reports when it is done receiving new data
// which the parent select listens for.
//
// Please see https://github.com/hashicorp/consul-template/issues/168 for
// more information about this optimization and the entire backstory.
for {
select {
case view := <-r.watcher.DataCh():
r.Receive(view.Dependency(), view.Data())
default:
break OUTER
}
}
case <-dedupCh:
// We may get triggered by the de-duplication manager for either a change
// in leadership (acquired or lost lock), or an update of data for a template
// that we are watching.
log.Printf("[INFO] (runner) watcher triggered by de-duplication manager")
break OUTER
case err := <-r.watcher.ErrCh():
// Push the error back up the stack
log.Printf("[ERR] (runner) watcher reported error: %s", err)
r.ErrCh <- err
return
case tmpl := <-r.quiescenceCh:
// Remove the quiescence for this template from the map. This will force
// the upcoming Run call to actually evaluate and render the template.
log.Printf("[DEBUG] (runner) received template %q from quiescence", tmpl.ID())
delete(r.quiescenceMap, tmpl.ID())
case c := <-childExitCh:
log.Printf("[INFO] (runner) child process died")
r.ErrCh <- NewErrChildDied(c)
return
case <-r.DoneCh:
log.Printf("[INFO] (runner) received finish")
return
}
// If we got this far, that means we got new data or one of the timers
// fired, so attempt to re-render.
if err := r.Run(); err != nil {
r.ErrCh <- err
return
}
}
}