backport of commit c2dc1c5

plus MinJob() from nomad/mock/job.go

Author: gulducat

.changelog/17104.txt

@@ -0,0 +1,3 @@
+```release-note:bug
+client: clean up resources upon failure to restore task during client restart
+```

client/allocrunner/alloc_runner.go

@@ -8,6 +8,7 @@ import (
 
 	log "github.com/hashicorp/go-hclog"
 	multierror "github.com/hashicorp/go-multierror"
+
 	"github.com/hashicorp/nomad/client/allocdir"
 	"github.com/hashicorp/nomad/client/allocrunner/interfaces"
 	"github.com/hashicorp/nomad/client/allocrunner/state"
@@ -344,17 +345,15 @@ func (ar *allocRunner) Run() {
 			ar.logger.Error("prerun failed", "error", err)
 
 			for _, tr := range ar.tasks {
-				tr.MarkFailedDead(fmt.Sprintf("failed to setup alloc: %v", err))
+				// emit event and mark task to be cleaned up during runTasks()
+				tr.MarkFailedKill(fmt.Sprintf("failed to setup alloc: %v", err))
 			}
-
-			goto POST
 		}
 	}
 
 	// Run the runners (blocks until they exit)
 	ar.runTasks()
 
-POST:
 	if ar.isShuttingDown() {
 		return
 	}

client/allocrunner/alloc_runner_hooks.go

@@ -5,6 +5,7 @@ import (
 	"time"
 
 	multierror "github.com/hashicorp/go-multierror"
+
 	"github.com/hashicorp/nomad/client/allocrunner/interfaces"
 	clientconfig "github.com/hashicorp/nomad/client/config"
 	"github.com/hashicorp/nomad/client/taskenv"
@@ -135,6 +136,9 @@ func (ar *allocRunner) initRunnerHooks(config *clientconfig.Config) error {
 		newCSIHook(alloc, hookLogger, ar.csiManager, ar.rpcClient, ar, ar.hookResources, ar.clientConfig.Node.SecretID),
 		newChecksHook(hookLogger, alloc, ar.checkStore, ar),
 	}
+	if config.ExtraAllocHooks != nil {
+		ar.runnerHooks = append(ar.runnerHooks, config.ExtraAllocHooks...)
+	}
 
 	return nil
 }

client/allocrunner/fail_hook.go

@@ -0,0 +1,118 @@
+// Copyright (c) HashiCorp, Inc.
+// SPDX-License-Identifier: MPL-2.0
+
+// FailHook is designed to fail for testing purposes,
+// so should never be included in a release.
+//go:build !release
+
+package allocrunner
+
+import (
+	"errors"
+	"fmt"
+	"os"
+
+	"github.com/hashicorp/go-hclog"
+	"github.com/hashicorp/hcl/v2/hclsimple"
+
+	"github.com/hashicorp/nomad/client/allocrunner/interfaces"
+)
+
+var ErrFailHookError = errors.New("failed successfully")
+
+func NewFailHook(l hclog.Logger, name string) *FailHook {
+	return &FailHook{
+		name:   name,
+		logger: l.Named(name),
+	}
+}
+
+type FailHook struct {
+	name   string
+	logger hclog.Logger
+	Fail   struct {
+		Prerun         bool `hcl:"prerun,optional"`
+		PreKill        bool `hcl:"prekill,optional"`
+		Postrun        bool `hcl:"postrun,optional"`
+		Destroy        bool `hcl:"destroy,optional"`
+		Update         bool `hcl:"update,optional"`
+		PreTaskRestart bool `hcl:"pretaskrestart,optional"`
+		Shutdown       bool `hcl:"shutdown,optional"`
+	}
+}
+
+func (h *FailHook) Name() string {
+	return h.name
+}
+
+func (h *FailHook) LoadConfig(path string) *FailHook {
+	if _, err := os.Stat(path); os.IsNotExist(err) {
+		h.logger.Error("couldn't load config", "error", err)
+		return h
+	}
+	if err := hclsimple.DecodeFile(path, nil, &h.Fail); err != nil {
+		h.logger.Error("error parsing config", "path", path, "error", err)
+	}
+	return h
+}
+
+var _ interfaces.RunnerPrerunHook = &FailHook{}
+
+func (h *FailHook) Prerun() error {
+	if h.Fail.Prerun {
+		return fmt.Errorf("prerun %w", ErrFailHookError)
+	}
+	return nil
+}
+
+var _ interfaces.RunnerPreKillHook = &FailHook{}
+
+func (h *FailHook) PreKill() {
+	if h.Fail.PreKill {
+		h.logger.Error("prekill", "error", ErrFailHookError)
+	}
+}
+
+var _ interfaces.RunnerPostrunHook = &FailHook{}
+
+func (h *FailHook) Postrun() error {
+	if h.Fail.Postrun {
+		return fmt.Errorf("postrun %w", ErrFailHookError)
+	}
+	return nil
+}
+
+var _ interfaces.RunnerDestroyHook = &FailHook{}
+
+func (h *FailHook) Destroy() error {
+	if h.Fail.Destroy {
+		return fmt.Errorf("destroy %w", ErrFailHookError)
+	}
+	return nil
+}
+
+var _ interfaces.RunnerUpdateHook = &FailHook{}
+
+func (h *FailHook) Update(request *interfaces.RunnerUpdateRequest) error {
+	if h.Fail.Update {
+		return fmt.Errorf("update %w", ErrFailHookError)
+	}
+	return nil
+}
+
+var _ interfaces.RunnerTaskRestartHook = &FailHook{}
+
+func (h *FailHook) PreTaskRestart() error {
+	if h.Fail.PreTaskRestart {
+		return fmt.Errorf("destroy %w", ErrFailHookError)
+	}
+	return nil
+}
+
+var _ interfaces.ShutdownHook = &FailHook{}
+
+func (h *FailHook) Shutdown() {
+	if h.Fail.Shutdown {
+		h.logger.Error("shutdown", "error", ErrFailHookError)
+	}
+}

client/allocrunner/taskrunner/task_runner.go

@@ -8,13 +8,13 @@ import (
 	"sync"
 	"time"
 
-	"github.com/hashicorp/nomad/client/lib/cgutil"
 	"golang.org/x/exp/slices"
 
 	metrics "github.com/armon/go-metrics"
 	log "github.com/hashicorp/go-hclog"
 	multierror "github.com/hashicorp/go-multierror"
 	"github.com/hashicorp/hcl/v2/hcldec"
+
 	"github.com/hashicorp/nomad/client/allocdir"
 	"github.com/hashicorp/nomad/client/allocrunner/interfaces"
 	"github.com/hashicorp/nomad/client/allocrunner/taskrunner/restarts"
@@ -24,6 +24,7 @@ import (
 	"github.com/hashicorp/nomad/client/devicemanager"
 	"github.com/hashicorp/nomad/client/dynamicplugins"
 	cinterfaces "github.com/hashicorp/nomad/client/interfaces"
+	"github.com/hashicorp/nomad/client/lib/cgutil"
 	"github.com/hashicorp/nomad/client/pluginmanager/csimanager"
 	"github.com/hashicorp/nomad/client/pluginmanager/drivermanager"
 	"github.com/hashicorp/nomad/client/serviceregistration"
@@ -492,30 +493,20 @@ func (tr *TaskRunner) initLabels() {
 	}
 }
 
-// MarkFailedDead marks a task as failed and not to run. Aimed to be invoked
-// when alloc runner prestart hooks failed. Should never be called with Run().
-func (tr *TaskRunner) MarkFailedDead(reason string) {
-	defer close(tr.waitCh)
-
-	tr.stateLock.Lock()
-	if err := tr.stateDB.PutTaskRunnerLocalState(tr.allocID, tr.taskName, tr.localState); err != nil {
-		//TODO Nomad will be unable to restore this task; try to kill
-		//     it now and fail? In general we prefer to leave running
-		//     tasks running even if the agent encounters an error.
-		tr.logger.Warn("error persisting local failed task state; may be unable to restore after a Nomad restart",
-			"error", err)
-	}
-	tr.stateLock.Unlock()
-
+// MarkFailedKill marks a task as failed and should be killed.
+// It should be invoked when alloc runner prestart hooks fail.
+// Afterwards, Run() will perform any necessary cleanup.
+func (tr *TaskRunner) MarkFailedKill(reason string) {
+	// Emit an event that fails the task and gives reasons for humans.
 	event := structs.NewTaskEvent(structs.TaskSetupFailure).
+		SetKillReason(structs.TaskRestoreFailed).
 		SetDisplayMessage(reason).
 		SetFailsTask()
-	tr.UpdateState(structs.TaskStateDead, event)
+	tr.EmitEvent(event)
 
-	// Run the stop hooks in case task was a restored task that failed prestart
-	if err := tr.stop(); err != nil {
-		tr.logger.Error("stop failed while marking task dead", "error", err)
-	}
+	// Cancel kill context, so later when allocRunner runs tr.Run(),
+	// we'll follow the usual kill path and do all the appropriate cleanup steps.
+	tr.killCtxCancel()
 }
 
 // Run the TaskRunner. Starts the user's task or reattaches to a restored task.

client/allocrunner/taskrunner/task_runner_test.go

@@ -13,6 +13,12 @@ import (
 	"time"
 
 	"github.com/golang/snappy"
+	"github.com/kr/pretty"
+	"github.com/shoenig/test"
+	"github.com/shoenig/test/must"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
 	"github.com/hashicorp/nomad/ci"
 	"github.com/hashicorp/nomad/client/allocdir"
 	"github.com/hashicorp/nomad/client/allocrunner/interfaces"
@@ -38,9 +44,6 @@ import (
 	"github.com/hashicorp/nomad/plugins/device"
 	"github.com/hashicorp/nomad/plugins/drivers"
 	"github.com/hashicorp/nomad/testutil"
-	"github.com/kr/pretty"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
 )
 
 type MockTaskStateUpdater struct {
@@ -658,6 +661,61 @@ func TestTaskRunner_Restore_System(t *testing.T) {
 	})
 }
 
+// TestTaskRunner_MarkFailedKill asserts that MarkFailedKill marks the task as failed
+// and cancels the killCtx so a subsequent Run() will do any necessary task cleanup.
+func TestTaskRunner_MarkFailedKill(t *testing.T) {
+	ci.Parallel(t)
+
+	// set up some taskrunner
+	alloc := mock.MinAlloc()
+	task := alloc.Job.TaskGroups[0].Tasks[0]
+	conf, cleanup := testTaskRunnerConfig(t, alloc, task.Name)
+	t.Cleanup(cleanup)
+	tr, err := NewTaskRunner(conf)
+	must.NoError(t, err)
+
+	// side quest: set this lifecycle coordination channel,
+	// so early in tr MAIN, it doesn't randomly follow that route.
+	// test config creates this already closed, but not so in real life.
+	startCh := make(chan struct{})
+	t.Cleanup(func() { close(startCh) })
+	tr.startConditionMetCh = startCh
+
+	// function under test: should mark the task as failed and cancel kill context
+	reason := "because i said so"
+	tr.MarkFailedKill(reason)
+
+	// explicitly check kill context.
+	select {
+	case <-tr.killCtx.Done():
+	default:
+		t.Fatal("kill context should be done")
+	}
+
+	// Run() should now follow the kill path.
+	go tr.Run()
+
+	select { // it should finish up very quickly
+	case <-tr.WaitCh():
+	case <-time.After(time.Second):
+		t.Error("task not killed (or not as fast as expected)")
+	}
+
+	// check state for expected values and events
+	state := tr.TaskState()
+
+	// this gets set directly by MarkFailedKill()
+	test.True(t, state.Failed, test.Sprint("task should have failed"))
+	// this is set in Run()
+	test.Eq(t, structs.TaskStateDead, state.State, test.Sprint("task should be dead"))
+	// reason "because i said so" should be a task event message
+	foundMessages := make(map[string]bool)
+	for _, event := range state.Events {
+		foundMessages[event.DisplayMessage] = true
+	}
+	test.True(t, foundMessages[reason], test.Sprintf("expected '%s' in events: %#v", reason, foundMessages))
+}
+
 // TestTaskRunner_TaskEnv_Interpolated asserts driver configurations are
 // interpolated.
 func TestTaskRunner_TaskEnv_Interpolated(t *testing.T) {