operation_manager.go

// Copyright 2016 Google LLC. All Rights Reserved.
//
// 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 log holds the code that is specific to Trillian logs core operation,
// particularly the code for sequencing.
package log

import (
	"context"
	"fmt"
	"reflect"
	"sort"
	"strconv"
	"sync"
	"time"

	"github.com/google/trillian/extension"
	"github.com/google/trillian/monitoring"
	"github.com/google/trillian/storage"
	"github.com/google/trillian/util/clock"
	"github.com/google/trillian/util/election"
	"golang.org/x/sync/semaphore"
	"k8s.io/klog/v2"
)

var (
	// DefaultTimeout is the default timeout on a single log operation run.
	DefaultTimeout = 60 * time.Second

	once              sync.Once
	knownLogs         monitoring.Gauge
	resignations      monitoring.Counter
	isMaster          monitoring.Gauge
	signingRuns       monitoring.Counter
	failedSigningRuns monitoring.Counter
	entriesAdded      monitoring.Counter
	batchesAdded      monitoring.Counter
)

func createMetrics(mf monitoring.MetricFactory) {
	if mf == nil {
		mf = monitoring.InertMetricFactory{}
	}
	knownLogs = mf.NewGauge("known_logs", "Set to 1 for known logs (whether this instance is master or not)", logIDLabel)
	resignations = mf.NewCounter("master_resignations", "Number of mastership resignations", logIDLabel)
	isMaster = mf.NewGauge("is_master", "Whether this instance is master (0/1)", logIDLabel)
	signingRuns = mf.NewCounter("signing_runs", "Number of times a signing run has succeeded", logIDLabel)
	failedSigningRuns = mf.NewCounter("failed_signing_runs", "Number of times a signing run has failed", logIDLabel)
	// entriesAdded is the total number of entries that have been added to the
	// log during the lifetime of a signer. This allows an operator to determine
	// that the queue is empty for a particular log; if signing runs are succeeding
	// but nothing is being processed then this counter will stop increasing.
	entriesAdded = mf.NewCounter("entries_added", "Number of entries added to the log", logIDLabel)
	// batchesAdded is the number of times a signing run caused entries to be
	// integrated into the log. The value batchesAdded / signingRuns is an
	// indication of how often the signer runs but does no work. The value of
	// entriesAdded / batchesAdded is average batch size. These can be used for
	// tuning sequencing or evaluating performance.
	batchesAdded = mf.NewCounter("batches_added", "Number of times a non zero number of entries was added", logIDLabel)
}

// Operation defines a task that operates on a log. Examples are scheduling, signing,
// consistency checking or cleanup.
type Operation interface {
	// ExecutePass performs a single pass of processing on a single log.  It returns
	// a count of items processed (for logging) and an error.
	ExecutePass(ctx context.Context, logID int64, info *OperationInfo) (int, error)
}

// OperationInfo bundles up information needed for running a set of Operations.
type OperationInfo struct {
	// Registry provides access to Trillian storage.
	Registry extension.Registry

	// The following parameters are passed to individual Operations.

	// BatchSize is the batch size to be passed to tasks run by this manager.
	BatchSize int
	// TimeSource should be used by the Operation to allow mocking for tests.
	TimeSource clock.TimeSource

	// The following parameters govern the overall scheduling of Operations
	// by a OperationManager.

	// Election-related configuration. Copied for each log.
	ElectionConfig election.RunnerConfig

	// RunInterval is the time between starting batches of processing.  If a
	// batch takes longer than this interval to complete, the next batch
	// will start immediately.
	RunInterval time.Duration
	// NumWorkers is the number of worker goroutines to run in parallel.
	NumWorkers int
	// Timeout sets an optional timeout on each operation run.
	// If unset, default to the value of DefaultTimeout.
	Timeout time.Duration
}

// OperationManager controls scheduling activities for logs.
type OperationManager struct {
	info OperationInfo

	// logOperation is the task that gets run for active logs.
	logOperation Operation

	// runnerWG groups all goroutines with election Runners.
	runnerWG sync.WaitGroup
	// runnerCancels contains cancel function for each logID election Runner.
	runnerCancels map[string]context.CancelFunc
	// pendingResignations delivers resignation requests from election Runners.
	pendingResignations chan election.Resignation

	tracker *election.MasterTracker

	// Cache of logID => name. Names are assumed not to change during runtime.
	logNames map[int64]string
	// A recent list of active logs that this instance is master for.
	lastHeld []int64
	// idsMutex guards logNames and lastHeld fields.
	idsMutex sync.Mutex
}

// NewOperationManager creates a new OperationManager instance.
func NewOperationManager(info OperationInfo, logOperation Operation) *OperationManager {
	once.Do(func() {
		createMetrics(info.Registry.MetricFactory)
	})
	if info.Timeout == 0 {
		info.Timeout = DefaultTimeout
	}
	tracker := election.NewMasterTracker(nil, func(id string, v bool) {
		val := 0.0
		if v {
			val = 1.0
		}
		isMaster.Set(val, id)
	})
	return &OperationManager{
		info:                info,
		logOperation:        logOperation,
		runnerCancels:       make(map[string]context.CancelFunc),
		pendingResignations: make(chan election.Resignation, 100),
		tracker:             tracker,
		logNames:            make(map[int64]string),
	}
}

// logName maps a logID to a human-readable name, caching results along the way.
// The human-readable name may non-unique so should only be used for diagnostics.
func (o *OperationManager) logName(ctx context.Context, logID int64) string {
	o.idsMutex.Lock()
	defer o.idsMutex.Unlock()
	if name, ok := o.logNames[logID]; ok {
		return name
	}

	tree, err := storage.GetTree(ctx, o.info.Registry.AdminStorage, logID)
	if err != nil {
		klog.Errorf("%v: failed to get log info: %v", logID, err)
		return "<err>"
	}

	name := tree.DisplayName
	if name == "" {
		name = fmt.Sprintf("<log-%d>", logID)
	}
	o.logNames[logID] = name
	return o.logNames[logID]
}

func (o *OperationManager) heldInfo(ctx context.Context, logIDs []int64) string {
	names := make([]string, 0, len(logIDs))
	for _, logID := range logIDs {
		names = append(names, o.logName(ctx, logID))
	}
	sort.Strings(names)

	result := "master for:"
	for _, name := range names {
		result += " " + name
	}
	return result
}

// masterFor returns the list of log IDs among allIDs that this instance is
// master for. Note that the instance may hold mastership for logs that are not
// listed in allIDs, but such logs are skipped.
func (o *OperationManager) masterFor(ctx context.Context, allIDs []int64) ([]int64, error) {
	if o.info.Registry.ElectionFactory == nil {
		return allIDs, nil
	}
	allStringIDs := make([]string, 0, len(allIDs))
	for _, id := range allIDs {
		s := strconv.FormatInt(id, 10)
		allStringIDs = append(allStringIDs, s)
	}

	// Synchronize the set of log IDs with those we are tracking mastership for.
	for _, logID := range allStringIDs {
		knownLogs.Set(1, logID)
		if o.runnerCancels[logID] == nil {
			o.tracker.Set(logID, false) // Initialise tracking for this ID.
			o.runnerCancels[logID] = o.runElectionWithRestarts(ctx, logID)
		}
	}

	held := o.tracker.Held()
	heldIDs := make([]int64, 0, len(allIDs))
	sort.Strings(allStringIDs)
	for _, s := range held {
		// Skip the log if it is not present in allIDs.
		if i := sort.SearchStrings(allStringIDs, s); i >= len(allStringIDs) || allStringIDs[i] != s {
			continue
		}
		id, err := strconv.ParseInt(s, 10, 64)
		if err != nil {
			return nil, fmt.Errorf("failed to parse logID %v as int64", s)
		}
		heldIDs = append(heldIDs, id)
	}

	return heldIDs, nil
}

// runElectionWithRestarts runs the election/resignation loop for the given log
// indefinitely, until the returned CancelFunc is invoked. Any failure during
// the loop leads to a restart of the loop with a few seconds delay.
//
// TODO(pavelkalinnikov): Restart the whole log operation rather than just the
// election, and have a metric for restarts.
func (o *OperationManager) runElectionWithRestarts(ctx context.Context, logID string) context.CancelFunc {
	klog.Infof("create master election goroutine for %v", logID)
	cctx, cancel := context.WithCancel(ctx)
	run := func(ctx context.Context) {
		e, err := o.info.Registry.ElectionFactory.NewElection(ctx, logID)
		if err != nil {
			klog.Errorf("failed to create election for %v: %v", logID, err)
			return
		}
		// Warning: NewRunner can attempt to modify the config. Make a separate
		// copy of the config for each log, to avoid data races.
		config := o.info.ElectionConfig
		// TODO(pavelkalinnikov): Passing the cancel function is not needed here.
		r := election.NewRunner(logID, &config, o.tracker, cancel, e)
		r.Run(ctx, o.pendingResignations)
	}
	o.runnerWG.Add(1)
	go func(ctx context.Context) {
		defer o.runnerWG.Done()
		// Continue only while the context is active.
		for ctx.Err() == nil {
			run(ctx)
			// Sleep before restarts, to not spam the log with errors.
			// TODO(pavelkalinnikov): Make the interval configurable.
			const pause = time.Duration(5 * time.Second)
			if err := clock.SleepSource(ctx, pause, o.info.TimeSource); err != nil {
				break // The context has been canceled during the sleep.
			}
		}
	}(cctx)
	return cancel
}

// updateHeldIDs updates the process status with the number/list of logs that
// the instance holds mastership for.
func (o *OperationManager) updateHeldIDs(ctx context.Context, logIDs, activeIDs []int64) {
	heldInfo := o.heldInfo(ctx, logIDs)
	msg := fmt.Sprintf("Acting as master for %d / %d active logs: %s", len(logIDs), len(activeIDs), heldInfo)
	o.idsMutex.Lock()
	defer o.idsMutex.Unlock()
	if !reflect.DeepEqual(logIDs, o.lastHeld) {
		o.lastHeld = make([]int64, len(logIDs))
		copy(o.lastHeld, logIDs)
		klog.Info(msg)
		if o.info.Registry.SetProcessStatus != nil {
			o.info.Registry.SetProcessStatus(heldInfo)
		}
	} else {
		klog.V(1).Info(msg)
	}
}

func (o *OperationManager) getLogsAndExecutePass(ctx context.Context) error {
	runCtx, cancel := context.WithTimeout(ctx, o.info.Timeout)
	defer cancel()

	activeIDs, err := o.info.Registry.LogStorage.GetActiveLogIDs(ctx)
	if err != nil {
		return fmt.Errorf("failed to list active log IDs: %v", err)
	}
	// Find the logs we are master for, skipping those logs that are not active,
	// e.g. deleted or FROZEN ones.
	// TODO(pavelkalinnikov): Resign mastership for the inactive logs.
	logIDs, err := o.masterFor(ctx, activeIDs)
	if err != nil {
		return fmt.Errorf("failed to determine log IDs we're master for: %v", err)
	}
	o.updateHeldIDs(ctx, logIDs, activeIDs)

	executePassForAll(runCtx, &o.info, o.logOperation, logIDs)
	return nil
}

// OperationSingle performs a single pass of the manager.
//
// TODO(pavelkalinnikov): Deprecate this because it doesn't clean up any state,
// and is used only for testing.
func (o *OperationManager) OperationSingle(ctx context.Context) {
	if err := o.getLogsAndExecutePass(ctx); err != nil {
		klog.Errorf("failed to perform operation: %v", err)
	}
}

// OperationLoop starts the manager working. It continues until told to exit.
// TODO(Martin2112): No mechanism for error reporting etc., this is OK for v1 but needs work
func (o *OperationManager) OperationLoop(ctx context.Context) {
	klog.Infof("Log operation manager starting")

	// Outer loop, runs until terminated.
	for {
		if err := o.operateOnce(ctx); err != nil {
			klog.Infof("Log operation manager shutting down")
			break
		}
	}

	// Terminate all the election Runners.
	for logID, cancel := range o.runnerCancels {
		if cancel != nil {
			klog.V(1).Infof("cancel election runner for %s", logID)
			cancel()
		}
	}

	// Drain any remaining resignations which might have triggered.
	close(o.pendingResignations)
	for r := range o.pendingResignations {
		resignations.Inc(r.ID)
		r.Execute(ctx)
	}

	klog.Infof("wait for termination of election runners...")
	o.runnerWG.Wait()
	klog.Infof("wait for termination of election runners...done")
}

// operateOnce runs a single round of operation for each of the active logs
// that this instance is master for. Returns an error only if the context is
// canceled, i.e. the operation is being shut down.
func (o *OperationManager) operateOnce(ctx context.Context) error {
	// TODO(alcutter): want a child context with deadline here?
	start := o.info.TimeSource.Now()
	if err := o.getLogsAndExecutePass(ctx); err != nil {
		// Suppress the error if ctx is done (ctx.Err != nil) as we're exiting.
		if ctx.Err() != nil {
			klog.Errorf("failed to execute operation on logs: %v", err)
		}
	}
	klog.V(1).Infof("Log operation manager pass complete")

	// Process any pending resignations while there's no activity.
	doneResigning := false
	for !doneResigning {
		select {
		case r := <-o.pendingResignations:
			resignations.Inc(r.ID)
			r.Execute(ctx)
		default:
			doneResigning = true
		}
	}

	// See if it's time to quit.
	select {
	case <-ctx.Done():
		return ctx.Err()
	default:
	}

	// Wait for the configured time before going for another pass.
	duration := o.info.TimeSource.Now().Sub(start)
	wait := o.info.RunInterval - duration
	if wait > 0 {
		klog.V(1).Infof("Processing started at %v for %v; wait %v before next run", start, duration, wait)
		if err := clock.SleepContext(ctx, wait); err != nil {
			return err
		}
	} else {
		klog.V(1).Infof("Processing started at %v for %v; start next run immediately", start, duration)
	}
	return nil
}

// executePassForAll runs ExecutePass of the given operation for each of the
// passed-in logs, allowing up to a configurable number of parallel operations.
func executePassForAll(ctx context.Context, info *OperationInfo, op Operation, logIDs []int64) {
	startBatch := info.TimeSource.Now()

	numWorkers := info.NumWorkers
	if numWorkers <= 0 {
		klog.Warning("Running executor with NumWorkers <= 0, assuming 1")
		numWorkers = 1
	}
	klog.V(1).Infof("Running executor with %d worker(s)", numWorkers)

	sem := semaphore.NewWeighted(int64(numWorkers))
	var wg sync.WaitGroup
	for _, logID := range logIDs {
		if err := sem.Acquire(ctx, 1); err != nil {
			break // Terminate because the context is canceled.
		}
		wg.Add(1)
		go func(logID int64) {
			defer wg.Done()
			defer sem.Release(1)
			if err := executePass(ctx, info, op, logID); err != nil {
				klog.Errorf("ExecutePass(%v) failed: %v", logID, err)
			}
		}(logID)
	}

	// Wait for the workers to consume all of the logIDs.
	wg.Wait()
	d := clock.SecondsSince(info.TimeSource, startBatch)
	klog.V(1).Infof("Group run completed in %.2f seconds", d)
}

// executePass runs ExecutePass of the given operation for the passed-in log.
func executePass(ctx context.Context, info *OperationInfo, op Operation, logID int64) error {
	label := strconv.FormatInt(logID, 10)
	start := info.TimeSource.Now()
	count, err := op.ExecutePass(ctx, logID, info)
	if err != nil {
		failedSigningRuns.Inc(label)
		return err
	}

	// This indicates signing activity is proceeding on the logID.
	signingRuns.Inc(label)
	if count > 0 {
		d := clock.SecondsSince(info.TimeSource, start)
		klog.Infof("%v: processed %d items in %.2f seconds (%.2f qps)", logID, count, d, float64(count)/d)
		entriesAdded.Add(float64(count), label)
		batchesAdded.Inc(label)
	} else {
		klog.V(1).Infof("%v: no items to process", logID)
	}
	return nil
}