watcher.go

// Copyright 2018-2020 Celer Network
//
// This is a wrapper on top of Go-Ethereum's client API (ethclient)
// for fetching log events via filtered query requests.  It handles
// reliable restarts of the application by persisting into the KVStore
// the index of the last log event acknowledged by the application.
// This guarantees that the application does not miss log events that
// occur while the application is not running.

package watcher

import (
	"container/list"
	"context"
	"errors"
	"fmt"
	"math/big"
	"strings"
	"sync"
	"time"

	"github.com/celer-network/goutils/log"
	ethereum "github.com/ethereum/go-ethereum"
	"github.com/ethereum/go-ethereum/accounts/abi"
	"github.com/ethereum/go-ethereum/common"
	"github.com/ethereum/go-ethereum/core/types"
)

const (
	notBlockIndex = ^int64(0) // 0xFFFFFFFF not-a-block-index value
)

var (
	ErrWatchServiceClosed = errors.New("Watch service closed")
	ErrWatcherTimeout     = errors.New("Watcher timeout")
)

// LogEventID tracks the position of a watch event in the event log.
type LogEventID struct {
	BlockNumber uint64 // Number of the block containing the event
	Index       int64  // Index of the event within the block
}

// WatchService holds the active watchers and their connections
// to the Ethereum client and the KVStore persistence layer that
// provides resumability of the watcher after a restart.
type WatchService struct {
	client        WatchClient       // Ethereum client interface
	dal           WatchDAL          // Data access layer
	polling       uint64            // Polling interval (msec)
	maxBlockDelta uint64            // Maximum number of blocks per query, 0 means unlimited
	quit          chan bool         // Terminate the watch service
	mu            sync.RWMutex      // Guards the fields that follow it.
	blkNum        uint64            // Current on-chain block number
	watches       map[string]*Watch // Map of registered watches
}

// Watch provides an iterator over a stream of event logs that match
// an Ethereum filtering query.  It updates the KVStore to persist the
// position in the stream of the last event log that the application
// has acknowledged receiving.
// To handle chain reorganization (ephemeral forking), watch only
// requests from on-chain event logs that are older than a specified
// number of on-chain blocks.
type Watch struct {
	name          string               // Unique name of registered watch
	service       *WatchService        // Service owning this watch
	ackWait       bool                 // Is it waiting for an ACK from the app?
	lastAck       bool                 // One last ACK is allowed after close
	ackID         LogEventID           // ID of log event pending an ACK
	lastID        *LogEventID          // ID of log event for resuming (or nil)
	blkDelay      uint64               // Log event delay in number of blocks
	fwdDelay      uint64               // Block delay when fast forward
	checkInterval uint64               // Check event every checkInterval * service.polling
	fromBlock     uint64               // Start a fetch from this block number
	query         ethereum.FilterQuery // On-chain event log query
	mu            sync.Mutex           // Guards log queue and "closed" flag
	logQueue      *list.List           // Queue of buffered log events
	logQueueCond  *sync.Cond           // Condvar for waiting on the queue
	closed        bool                 // Is the watch closed?
}

// WatchClient is an interface for the subset of functions of the Go-Ethereum
// client API that the watch service uses.
type WatchClient interface {
	HeaderByNumber(ctx context.Context, number *big.Int) (*types.Header, error)
	FilterLogs(ctx context.Context, q ethereum.FilterQuery) ([]types.Log, error)
}

// WatchDAL is an interface for the watch-specific API of the KVStore
// data access layer.
type WatchDAL interface {
	InsertMonitor(event string, blockNum uint64, blockIdx int64, restart bool) error
	GetMonitorBlock(event string) (uint64, int64, bool, error)
	UpdateMonitorBlock(event string, blockNum uint64, blockIdx int64) error
	UpsertMonitorBlock(event string, blockNum uint64, blockIdx int64, restart bool) error
}

// Create a watch service.
// polling: interval (in seconds) to periodically query eth logs
// maxBlockDelta: maximum number of blocks for each eth log query, 0 means unlimited
func NewWatchService(client WatchClient, dal WatchDAL, polling, maxBlockDelta uint64) *WatchService {
	// Note: the incoming polling interval is in seconds.  Purely for
	// unit testing purposes, the internal "polling" variable is in
	// milliseconds to allow tests to set faster polling intervals.
	return makeWatchService(client, dal, polling*1000, maxBlockDelta)
}

// Helper (for testing) to create a watch service with msec polling interval.
func makeWatchService(client WatchClient, dal WatchDAL, polling, maxBlockDelta uint64) *WatchService {
	if client == nil || dal == nil || polling == 0 {
		return nil
	}

	ws := &WatchService{
		client:        client,
		dal:           dal,
		polling:       polling,
		maxBlockDelta: maxBlockDelta,
		quit:          make(chan bool),
		watches:       make(map[string]*Watch),
	}

	// Synchronously initialize the current head block number.
	ws.updateBlockNumber()

	// Start the common on-chain block number watcher.
	go ws.watchBlkNum()

	return ws
}

// Internal goroutine within the watch service that tracks the on-chain
// block number.
func (ws *WatchService) watchBlkNum() {
	ticker := time.NewTicker(time.Duration(ws.polling) * time.Millisecond)
	defer ticker.Stop()

	for {
		select {
		case <-ws.quit:
			log.Debugln("watchBlkNum: quit")
			return

		case <-ticker.C:
			ws.updateBlockNumber()
		}
	}
}

// Fetch the on-chain block number and update the local value if needed.
func (ws *WatchService) updateBlockNumber() {
	head, err := ws.client.HeaderByNumber(context.Background(), nil)
	if err != nil {
		log.Traceln("cannot fetch on-chain block number:", err)
		return
	}

	blkNum := head.Number.Uint64()
	var topBlkNum uint64
	ws.mu.Lock()
	if blkNum > ws.blkNum {
		ws.blkNum = blkNum
	}
	topBlkNum = ws.blkNum
	ws.mu.Unlock()
	log.Tracef("top block #: %d, on-chain #: %d", topBlkNum, blkNum)
}

// Return the most recent on-chain block number.
func (ws *WatchService) GetBlockNumber() uint64 {
	ws.mu.RLock()
	defer ws.mu.RUnlock()

	return ws.blkNum
}

// Return the most recent on-chain block number in big.Int format.
func (ws *WatchService) GetCurrentBlockNumber() *big.Int {
	return new(big.Int).SetUint64(ws.GetBlockNumber())
}

// Close the watch service.
func (ws *WatchService) Close() {
	ws.mu.Lock()
	defer ws.mu.Unlock()

	if ws.watches != nil {
		close(ws.quit)
		for _, w := range ws.watches {
			w.inner_close()
		}
		ws.watches = nil
	}
}

// Register a named watch.
func (ws *WatchService) register(name string, watch *Watch) error {
	ws.mu.Lock()
	defer ws.mu.Unlock()

	if ws.watches == nil {
		return ErrWatchServiceClosed
	}

	if _, exist := ws.watches[name]; exist {
		return fmt.Errorf("watch name '%s' already in use", name)
	}
	ws.watches[name] = watch
	return nil
}

// Unregister a named watch
func (ws *WatchService) unregister(name string) {
	ws.mu.Lock()
	defer ws.mu.Unlock()

	if ws.watches != nil {
		delete(ws.watches, name)
	}
}

// MakeFilterQuery constructs an Ethereum FilterQuery structure from these
// event and contract parameters: address, raw ABI string, event name, and
// the optional start block number.
func (ws *WatchService) MakeFilterQuery(
	addr common.Address, rawABI string, eventName string, startBlock *big.Int) (ethereum.FilterQuery, error) {

	var q ethereum.FilterQuery

	parsedABI, err := abi.JSON(strings.NewReader((rawABI)))
	if err != nil {
		return q, err
	}

	ev, exist := parsedABI.Events[eventName]
	if !exist {
		return q, fmt.Errorf("unknown event name: %s", eventName)
	}

	q.FromBlock = startBlock
	q.Addresses = []common.Address{addr}
	q.Topics = [][]common.Hash{{ev.ID}}
	return q, nil
}

// Create a watch for the given Ethereum log filtering query.
// The block delay is the number of blocks mined used as a time delay
// for fetching event logs, mitigating the effects of chain reorg.
// The block interval controls the polling frequency of fetch logs
// from on-chain, but measured in block numbers (as a delta).
// If "reset" is enabled, the watcher ignores the previously stored
// position in the subscription which resets the stream to its start.
func (ws *WatchService) NewWatch(
	name string, query ethereum.FilterQuery, blkDelay, fwdDelay, checkInterval uint64, reset bool) (*Watch, error) {

	if name == "" {
		return nil, fmt.Errorf("watch name not specified")
	}

	if ws.maxBlockDelta > 0 && ws.maxBlockDelta <= fwdDelay {
		return nil, fmt.Errorf("maxBlockDelta %d should larger than fwdDelay %d", ws.maxBlockDelta, fwdDelay)
	}

	w := &Watch{
		name:          name,
		service:       ws,
		blkDelay:      blkDelay,
		fwdDelay:      fwdDelay,
		checkInterval: checkInterval,
		query:         query,
		logQueue:      list.New(),
	}

	w.logQueueCond = sync.NewCond(&w.mu) // condvar uses "mu"

	// Register the named watch.
	if err := ws.register(name, w); err != nil {
		return nil, err
	}

	go w.watchLogEvents(reset)

	return w, nil
}

// Internal goroutine that periodically fetches and enqueues the log events
// requested by the watch query.
func (w *Watch) watchLogEvents(reset bool) {
	// Set the ID of the last log event acknowledged by the app (if any).
	// It is used to resume the watch from where the application left off.
	w.setWatchLastID(reset)
	log.Debugf("watchLogEvents: start %s from %d", w.name, w.fromBlock)

	// The polling interval is computed in relation to block polling.
	polling := w.checkInterval * w.service.polling
	ticker := time.NewTicker(time.Duration(polling) * time.Millisecond)
	defer ticker.Stop()

	for !w.isClosed() {
		w.fetchLogEvents()

		select {
		case <-w.service.quit:
			log.Debugln("watchLogEvents: quit:", w.name)
			return

		case <-ticker.C:
			continue
		}
	}

	log.Debugln("watchLogEvents: closed:", w.name)
}

// Set the ID of the last log event acknowledged by the app (if any).
// This is used to resume the watcher from where it left off when the
// application was terminated.  If "reset" is enabled, ignore the value
// in the store and reset it to the one provided by the query.
//
// Note: The design of an ack'ed ID keeps open a small window between
// the first time a watch is started (from a user-specified block number)
// and the first log it gives to the app and gets its ACK.  During that
// time window, a server crash/restart will not have any persisted ID
// and would appear to watch as brand-new, even though that watch was
// previously started.  To handle this state (a watch previously started
// from a block number but never had the chance to give a log to the app)
// it is encoded into the persisted ID using the special "notBlockIndex"
// value for the block "Index": ID = {BlockNumber, notBlockIndex}.
// With that in place, a restarted watch will resume from that previous
// block number again, giving persistence to that special time window.
func (w *Watch) setWatchLastID(reset bool) {
	w.lastID = nil
	w.fromBlock = 0

	blockNum, blockIdx, found, err := w.service.dal.GetMonitorBlock(w.name)
	if err == nil && found && !reset {
		if blockIdx != notBlockIndex {
			// This is a real ACK ID from the app.
			w.lastID = &LogEventID{
				BlockNumber: blockNum,
				Index:       blockIdx,
			}
		}

		// Ignore the query's "FromBlock" and start from this block.
		w.fromBlock = blockNum
		return
	}

	// No previously persisted resume pointer (or "reset", or some DB error).
	// Remember this starting block number to resume from in case of a crash.
	if w.query.FromBlock != nil {
		w.fromBlock = w.query.FromBlock.Uint64()
	}

	if err != nil {
		log.Warnln("resume pointer overwritten:", w.name, err)
		err = w.service.dal.UpsertMonitorBlock(w.name,
			w.fromBlock, notBlockIndex, false)
	} else if found {
		err = w.service.dal.UpdateMonitorBlock(w.name,
			w.fromBlock, notBlockIndex)
	} else {
		err = w.service.dal.InsertMonitor(w.name,
			w.fromBlock, notBlockIndex, false)
	}
	if err != nil {
		log.Warnln("cannot persist resume pointer:", w.name, err)
	}
}

// Fetch a batch of log events from on-chain and enqueue them.
// This function is called periodically in a ticker polling loop.
func (w *Watch) fetchLogEvents() {
	// Do nothing if the on-chain block number has not moved forward
	// beyond the desired block delay (to protect from on-chain reorg).
	blkNum := w.service.GetBlockNumber()
	if w.fromBlock+w.blkDelay > blkNum {
		log.Tracef("skip log fetching: %s: want %d, delay %d, blk %d", w.name, w.fromBlock, w.blkDelay, blkNum)
		return
	}

	// Fetch server-side filtered log events in the target range of
	// block numbers.  The block delay limit is used to avoid fetching
	// recently mined blocks that may still be undone by a chain reorg.

	// https://infura.io/docs/ethereum/json-rpc/eth-getLogs up to 10s
	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
	defer cancel()

	toBlock := blkNum - w.blkDelay
	if w.service.maxBlockDelta > 0 && toBlock-w.fromBlock > w.service.maxBlockDelta {
		toBlock = w.fromBlock + w.service.maxBlockDelta
	}
	w.query.FromBlock = new(big.Int).SetUint64(w.fromBlock)
	w.query.ToBlock = new(big.Int).SetUint64(toBlock)

	log.Tracef("fetch logs: %s: [%d-%d]", w.name, w.fromBlock, toBlock)

	logs, err := w.service.client.FilterLogs(ctx, w.query)
	if err != nil {
		log.Warnf("cannot fetch logs: %s: [%d-%d]: %s", w.name, w.fromBlock, toBlock, err)
		return
	}

	// Enqueue log events.  In the first fetch after an application resume,
	// lastID is used to skip log events previously handled within a block.
	// This is needed because log fetching is at the block granularity, but
	// processing is at the log event finer granularity.  After that initial
	// resuming block is refetched and partially de-duped, the lastID is no
	// longer needed.
	maxBlock := uint64(0)
	count := 0
	for i := range logs {
		log := &logs[i]
		if log.BlockNumber > maxBlock {
			maxBlock = log.BlockNumber
		}
		if w.lastID == nil {
			w.enqueue(log)
			count++
		} else if greaterThanLastID(log, w.lastID) {
			w.enqueue(log)
			count++
			if log.BlockNumber > w.lastID.BlockNumber {
				w.lastID = nil
			}
		}
	}

	// Update the next block number to start fetching.
	if maxBlock >= w.fromBlock {
		w.fromBlock = maxBlock + 1
		log.Tracef("added %d logs to queue: %s: next from %d", count, w.name, w.fromBlock)
	} else {
		// we didn't find any event between fromBlock and toBlock, so we can fast forward.
		fromBlock := toBlock
		if fromBlock+w.fwdDelay+w.blkDelay >= blkNum {
			// add additional block delay to mitigate consistency issues from query nodes
			fromBlock = fromBlock - w.fwdDelay
		}
		if fromBlock > w.fromBlock {
			w.fromBlock = fromBlock
		}
		// when there is no event in queue and the Watch is not waiting for Ack(), we
		// assert that at this moment, no event is being processed or going to be processed.
		if w.logQueue.Len() == 0 {
			w.mu.Lock()
			if !w.ackWait {
				if err := w.service.dal.UpdateMonitorBlock(w.name, w.fromBlock, notBlockIndex); err != nil {
					log.Warnln("cannot persist resume pointer:", w.name, err)
				}
			}
			w.mu.Unlock()
		}
		log.Tracef("fast forward %s fromBlock to %d", w.name, w.fromBlock)
	}
}

// Return true if the log event ID is strictly greater than the last ID.
func greaterThanLastID(log *types.Log, lastID *LogEventID) bool {
	if log.BlockNumber > lastID.BlockNumber {
		return true
	} else if log.BlockNumber < lastID.BlockNumber {
		return false
	} else if int64(log.Index) > lastID.Index {
		return true
	}
	return false
}

// Enqueue an event log.
func (w *Watch) enqueue(log *types.Log) {
	w.mu.Lock()
	defer w.mu.Unlock()

	oldLen := w.logQueue.Len()
	w.logQueue.PushBack(log)
	if oldLen == 0 {
		w.logQueueCond.Signal()
	}
}

// Dequeue and return the first event log in the queue, or block waiting
// for one to arrive or the watcher to be closed.
func (w *Watch) dequeue() (*types.Log, error) {
	w.mu.Lock()
	defer w.mu.Unlock()

	for !w.closed && w.logQueue.Len() == 0 && !w.ackWait {
		w.logQueueCond.Wait()
	}

	if w.closed {
		return nil, fmt.Errorf("watch name '%s' closed", w.name)
	}
	if w.ackWait {
		return nil, fmt.Errorf("last event log received not yet ACKed")
	}

	elem := w.logQueue.Front()
	w.logQueue.Remove(elem)
	nextLog := elem.Value.(*types.Log)
	w.ackID.BlockNumber = nextLog.BlockNumber
	w.ackID.Index = int64(nextLog.Index)
	w.ackWait = true
	return nextLog, nil
}

// Fetch the next log event.  The function will block until either an
// event log is available, or the watcher is closed.
func (w *Watch) Next() (types.Log, error) {
	var empty types.Log
	if w.isClosed() {
		return empty, fmt.Errorf("watch name '%s' already closed", w.name)
	}
	if w.ackWait {
		return empty, fmt.Errorf("last event log received not yet ACKed")
	}
	nextLog, err := w.dequeue()
	if err != nil {
		return empty, err
	}
	return *nextLog, nil
}

// The app ACKs the complete processing of the last received event log.
// Be lenient in one case: after the watch is closed, allow at most one
// more ACK to be done.  This allows event processing that was completed
// by the application when an asynchronous Close() took place (between
// the Next() and the Ack() calls) to be persisted into storage instead
// of having it be re-done after the application is restarted.
func (w *Watch) Ack() error {
	if w.isClosed() {
		if w.lastAck {
			return fmt.Errorf("watch name '%s' already closed", w.name)
		}
		w.lastAck = true
	}
	if !w.ackWait {
		return fmt.Errorf("last event log received already ACKed")
	}

	if err := w.service.dal.UpdateMonitorBlock(w.name, w.ackID.BlockNumber, w.ackID.Index); err != nil {
		return err
	}
	w.ackWait = false
	return nil
}

func (w *Watch) isClosed() bool {
	w.mu.Lock()
	defer w.mu.Unlock()
	return w.closed
}

func (w *Watch) inner_close() {
	w.mu.Lock()
	defer w.mu.Unlock()

	if !w.closed {
		w.closed = true
		w.logQueueCond.Broadcast()
	}
}

// Close a watch subscription.
func (w *Watch) Close() {
	w.service.unregister(w.name) // remove from watch service
	w.inner_close()
}