git-lfs/tq/transfer_queue.go

package tq

import (
	"os"
	"sort"
	"sync"

	"github.com/git-lfs/git-lfs/errors"
	"github.com/git-lfs/git-lfs/lfsapi"
	"github.com/git-lfs/git-lfs/tools"
	"github.com/rubyist/tracerx"
)

const (
	defaultBatchSize = 100
)

type retryCounter struct {
	MaxRetries int `git:"lfs.transfer.maxretries"`

	// cmu guards count
	cmu sync.Mutex
	// count maps OIDs to number of retry attempts
	count map[string]int
}

// newRetryCounter instantiates a new *retryCounter. It parses the gitconfig
// value: `lfs.transfer.maxretries`, and falls back to defaultMaxRetries if none
// was provided.
//
// If it encountered an error in Unmarshaling the *config.Configuration, it will
// be returned, otherwise nil.
func newRetryCounter() *retryCounter {
	return &retryCounter{
		MaxRetries: defaultMaxRetries,
		count:      make(map[string]int),
	}
}

// Increment increments the number of retries for a given OID. It is safe to
// call across multiple goroutines.
func (r *retryCounter) Increment(oid string) {
	r.cmu.Lock()
	defer r.cmu.Unlock()

	r.count[oid]++
}

// CountFor returns the current number of retries for a given OID. It is safe to
// call across multiple goroutines.
func (r *retryCounter) CountFor(oid string) int {
	r.cmu.Lock()
	defer r.cmu.Unlock()

	return r.count[oid]
}

// CanRetry returns the current number of retries, and whether or not it exceeds
// the maximum number of retries (see: retryCounter.MaxRetries).
func (r *retryCounter) CanRetry(oid string) (int, bool) {
	count := r.CountFor(oid)
	return count, count < r.MaxRetries
}

// batch implements the sort.Interface interface and enables sorting on a slice
// of `*Transfer`s by object size.
//
// This interface is implemented here so that the largest objects can be
// processed first. Since adding a new batch is unable to occur until the
// current batch has finished processing, this enables us to reduce the risk of
// a single worker getting tied up on a large item at the end of a batch while
// all other workers are sitting idle.
type batch []*objectTuple

// Concat concatenates two batches together, returning a single, clamped batch as
// "left", and the remainder of elements as "right". If the union of the
// receiver and "other" has cardinality less than "size", "right" will be
// returned as nil.
func (b batch) Concat(other batch, size int) (left, right batch) {
	u := batch(append(b, other...))
	if len(u) <= size {
		return u, nil
	}
	return u[:size], u[size:]
}

func (b batch) ToTransfers() []*Transfer {
	transfers := make([]*Transfer, 0, len(b))
	for _, t := range b {
		transfers = append(transfers, &Transfer{Oid: t.Oid, Size: t.Size})
	}
	return transfers
}

func (b batch) Len() int           { return len(b) }
func (b batch) Less(i, j int) bool { return b[i].Size < b[j].Size }
func (b batch) Swap(i, j int)      { b[i], b[j] = b[j], b[i] }

// TransferQueue organises the wider process of uploading and downloading,
// including calling the API, passing the actual transfer request to transfer
// adapters, and dealing with progress, errors and retries.
type TransferQueue struct {
	direction         Direction
	client            *tqClient
	remote            string
	adapter           Adapter
	adapterInProgress bool
	adapterInitMutex  sync.Mutex
	dryRun            bool
	cb                tools.CopyCallback
	meter             *Meter
	errors            []error
	transfers         map[string]*objects
	batchSize         int
	bufferDepth       int
	incoming          chan *objectTuple // Channel for processing incoming items
	errorc            chan error        // Channel for processing errors
	watchers          []chan *Transfer
	trMutex           *sync.Mutex
	collectorWait     sync.WaitGroup
	errorwait         sync.WaitGroup
	// wait is used to keep track of pending transfers. It is incremented
	// once per unique OID on Add(), and is decremented when that transfer
	// is marked as completed or failed, but not retried.
	wait     sync.WaitGroup
	manifest *Manifest
	rc       *retryCounter
}

// objects holds a set of objects.
type objects struct {
	completed bool
	objects   []*objectTuple
}

// All returns all *objectTuple's contained in the *objects set.
func (s *objects) All() []*objectTuple {
	return s.objects
}

// Append returns a new *objects with the given *objectTuple(s) appended to the
// end of the known objects.
func (s *objects) Append(os ...*objectTuple) *objects {
	return &objects{
		completed: s.completed,
		objects:   append(s.objects, os...),
	}
}

// First returns the first *objectTuple in the chain of objects.
func (s *objects) First() *objectTuple {
	if len(s.objects) == 0 {
		return nil
	}
	return s.objects[0]
}

type objectTuple struct {
	Name, Path, Oid string
	Size            int64
}

func (o *objectTuple) ToTransfer() *Transfer {
	return &Transfer{
		Name: o.Name,
		Path: o.Path,
		Oid:  o.Oid,
		Size: o.Size,
	}
}

type Option func(*TransferQueue)

func DryRun(dryRun bool) Option {
	return func(tq *TransferQueue) {
		tq.dryRun = dryRun
	}
}

func WithProgress(m *Meter) Option {
	return func(tq *TransferQueue) {
		tq.meter = m
	}
}

func WithProgressCallback(cb tools.CopyCallback) Option {
	return func(tq *TransferQueue) {
		tq.cb = cb
	}
}

func WithBatchSize(size int) Option {
	return func(tq *TransferQueue) { tq.batchSize = size }
}

func WithBufferDepth(depth int) Option {
	return func(tq *TransferQueue) { tq.bufferDepth = depth }
}

// NewTransferQueue builds a TransferQueue, direction and underlying mechanism determined by adapter
func NewTransferQueue(dir Direction, manifest *Manifest, remote string, options ...Option) *TransferQueue {
	q := &TransferQueue{
		direction: dir,
		client:    &tqClient{Client: manifest.APIClient()},
		remote:    remote,
		errorc:    make(chan error),
		transfers: make(map[string]*objects),
		trMutex:   &sync.Mutex{},
		manifest:  manifest,
		rc:        newRetryCounter(),
	}

	for _, opt := range options {
		opt(q)
	}

	q.rc.MaxRetries = q.manifest.maxRetries
	q.client.MaxRetries = q.manifest.maxRetries

	if q.batchSize <= 0 {
		q.batchSize = defaultBatchSize
	}
	if q.bufferDepth <= 0 {
		q.bufferDepth = q.batchSize
	}
	if q.meter != nil {
		q.meter.Direction = q.direction
	}

	q.incoming = make(chan *objectTuple, q.bufferDepth)
	q.collectorWait.Add(1)
	q.errorwait.Add(1)
	q.run()

	return q
}

// Add adds a *Transfer to the transfer queue. It only increments the amount
// of waiting the TransferQueue has to do if the *Transfer "t" is new.
//
// If another transfer(s) with the same OID has been added to the *TransferQueue
// already, the given transfer will not be enqueued, but will be sent to any
// channel created by Watch() once the oldest transfer has completed.
//
// Only one file will be transferred to/from the Path element of the first
// transfer.
func (q *TransferQueue) Add(name, path, oid string, size int64) {
	t := &objectTuple{
		Name: name,
		Path: path,
		Oid:  oid,
		Size: size,
	}

	if objs := q.remember(t); len(objs.objects) > 1 {
		if objs.completed {
			// If there is already a completed transfer chain for
			// this OID, then this object is already "done", and can
			// be sent through as completed to the watchers.
			for _, w := range q.watchers {
				w <- t.ToTransfer()
			}
		}

		// If the chain is not done, there is no reason to enqueue this
		// transfer into 'q.incoming'.
		tracerx.Printf("already transferring %q, skipping duplicate", t.Oid)
		return
	}

	q.incoming <- t
}

// remember remembers the *Transfer "t" if the *TransferQueue doesn't already
// know about a Transfer with the same OID.
//
// It returns if the value is new or not.
func (q *TransferQueue) remember(t *objectTuple) objects {
	q.trMutex.Lock()
	defer q.trMutex.Unlock()

	if _, ok := q.transfers[t.Oid]; !ok {
		q.wait.Add(1)
		q.transfers[t.Oid] = &objects{
			objects: []*objectTuple{t},
		}

		return *q.transfers[t.Oid]
	}

	q.transfers[t.Oid] = q.transfers[t.Oid].Append(t)

	return *q.transfers[t.Oid]
}

// collectBatches collects batches in a loop, prioritizing failed items from the
// previous before adding new items. The process works as follows:
//
//   1. Create a new batch, of size `q.batchSize`, and containing no items
//   2. While the batch contains less items than `q.batchSize` AND the channel
//      is open, read one item from the `q.incoming` channel.
//      a. If the read was a channel close, go to step 4.
//      b. If the read was a transferable item, go to step 3.
//   3. Append the item to the batch.
//   4. Sort the batch by descending object size, make a batch API call, send
//      the items to the `*adapterBase`.
//   5. In a separate goroutine, process the worker results, incrementing and
//      appending retries if possible. On the main goroutine, accept new items
//      into "pending".
//   6. Concat() the "next" and "pending" batches such that no more items than
//      the maximum allowed per batch are in next, and the rest are in pending.
//   7. If the `q.incoming` channel is open, go to step 2.
//   8. If the next batch is empty AND the `q.incoming` channel is closed,
//      terminate immediately.
//
// collectBatches runs in its own goroutine.
func (q *TransferQueue) collectBatches() {
	defer q.collectorWait.Done()

	var closing bool
	next := q.makeBatch()
	pending := q.makeBatch()

	for {
		for !closing && (len(next) < q.batchSize) {
			t, ok := <-q.incoming
			if !ok {
				closing = true
				break
			}

			next = append(next, t)
		}

		// Before enqueuing the next batch, sort by descending object
		// size.
		sort.Sort(sort.Reverse(next))

		done := make(chan struct{})

		var retries batch

		go func() {
			var err error

			retries, err = q.enqueueAndCollectRetriesFor(next)
			if err != nil {
				q.errorc <- err
			}

			close(done)
		}()

		var collected batch
		collected, closing = q.collectPendingUntil(done)

		// Ensure the next batch is filled with, in order:
		//
		// - retries from the previous batch,
		// - new additions that were enqueued behind retries, &
		// - items collected while the batch was processing.
		next, pending = retries.Concat(append(pending, collected...), q.batchSize)

		if closing && len(next) == 0 {
			// If len(next) == 0, there are no items in "pending",
			// and it is safe to exit.
			break
		}
	}
}

// collectPendingUntil collects items from q.incoming into a "pending" batch
// until the given "done" channel is written to, or is closed.
//
// A "pending" batch is returned, along with whether or not "q.incoming" is
// closed.
func (q *TransferQueue) collectPendingUntil(done <-chan struct{}) (pending batch, closing bool) {
	for {
		select {
		case t, ok := <-q.incoming:
			if !ok {
				closing = true
				<-done
				return
			}

			pending = append(pending, t)
		case <-done:
			return
		}
	}
}

// enqueueAndCollectRetriesFor makes a Batch API call and returns a "next" batch
// containing all of the objects that failed from the previous batch and had
// retries availale to them.
//
// If an error was encountered while making the API request, _all_ of the items
// from the previous batch (that have retries available to them) will be
// returned immediately, along with the error that was encountered.
//
// enqueueAndCollectRetriesFor blocks until the entire Batch "batch" has been
// processed.
func (q *TransferQueue) enqueueAndCollectRetriesFor(batch batch) (batch, error) {
	next := q.makeBatch()
	tracerx.Printf("tq: sending batch of size %d", len(batch))

	q.meter.Pause()
	var bRes *BatchResponse
	if q.manifest.standaloneTransferAgent != "" {
		// Trust the external transfer agent can do everything by itself.
		objects := make([]*Transfer, 0, len(batch))
		for _, t := range batch {
			objects = append(objects, &Transfer{Oid: t.Oid, Size: t.Size, Path: t.Path})
		}
		bRes = &BatchResponse{
			Objects:             objects,
			TransferAdapterName: q.manifest.standaloneTransferAgent,
		}
	} else {
		// Query the Git LFS server for what transfer method to use and
		// details such as URLs, authentication, etc.
		var err error
		bRes, err = Batch(q.manifest, q.direction, q.remote, batch.ToTransfers())
		if err != nil {
			// If there was an error making the batch API call, mark all of
			// the objects for retry, and return them along with the error
			// that was encountered. If any of the objects couldn't be
			// retried, they will be marked as failed.
			for _, t := range batch {
				if q.canRetryObject(t.Oid, err) {
					q.rc.Increment(t.Oid)

					next = append(next, t)
				} else {
					q.wait.Done()
				}
			}

			return next, err
		}
	}

	if len(bRes.Objects) == 0 {
		return next, nil
	}

	q.useAdapter(bRes.TransferAdapterName)
	q.meter.Start()

	toTransfer := make([]*Transfer, 0, len(bRes.Objects))

	for _, o := range bRes.Objects {
		if o.Error != nil {
			q.errorc <- errors.Wrapf(o.Error, "[%v] %v", o.Oid, o.Error.Message)
			q.Skip(o.Size)
			q.wait.Done()

			continue
		}

		q.trMutex.Lock()
		objects, ok := q.transfers[o.Oid]
		q.trMutex.Unlock()
		if !ok {
			// If we couldn't find any associated
			// Transfer object, then we give up on the
			// transfer by telling the progress meter to
			// skip the number of bytes in "o".
			q.errorc <- errors.Errorf("[%v] The server returned an unknown OID.", o.Oid)

			q.Skip(o.Size)
			q.wait.Done()
		} else {
			// Pick t[0], since it will cover all transfers with the
			// same OID.
			tr := newTransfer(o, objects.First().Name, objects.First().Path)

			if a, err := tr.Rel(q.direction.String()); err != nil {
				// XXX(taylor): duplication
				if q.canRetryObject(tr.Oid, err) {
					q.rc.Increment(tr.Oid)
					count := q.rc.CountFor(tr.Oid)

					tracerx.Printf("tq: enqueue retry #%d for %q (size: %d): %s", count, tr.Oid, tr.Size, err)
					next = append(next, objects.First())
				} else {
					q.errorc <- errors.Errorf("[%v] %v", tr.Name, err)

					q.Skip(o.Size)
					q.wait.Done()
				}
			} else if a == nil && q.manifest.standaloneTransferAgent == "" {
				q.Skip(o.Size)
				q.wait.Done()
			} else {
				q.meter.StartTransfer(objects.First().Name)
				toTransfer = append(toTransfer, tr)
			}
		}
	}

	retries := q.addToAdapter(bRes.endpoint, toTransfer)
	for t := range retries {
		q.rc.Increment(t.Oid)
		count := q.rc.CountFor(t.Oid)

		tracerx.Printf("tq: enqueue retry #%d for %q (size: %d)", count, t.Oid, t.Size)

		next = append(next, t)
	}

	return next, nil
}

// makeBatch returns a new, empty batch, with a capacity equal to the maximum
// batch size designated by the `*TransferQueue`.
func (q *TransferQueue) makeBatch() batch { return make(batch, 0, q.batchSize) }

// addToAdapter adds the given "pending" transfers to the transfer adapters and
// returns a channel of Transfers that are to be retried in the next batch.
// After all of the items in the batch have been processed, the channel is
// closed.
//
// addToAdapter returns immediately, and does not block.
func (q *TransferQueue) addToAdapter(e lfsapi.Endpoint, pending []*Transfer) <-chan *objectTuple {
	retries := make(chan *objectTuple, len(pending))

	if err := q.ensureAdapterBegun(e); err != nil {
		close(retries)

		q.errorc <- err
		for _, t := range pending {
			q.Skip(t.Size)
			q.wait.Done()
		}

		return retries
	}

	present, missingResults := q.partitionTransfers(pending)

	go func() {
		defer close(retries)

		var results <-chan TransferResult
		if q.dryRun {
			results = q.makeDryRunResults(present)
		} else {
			results = q.adapter.Add(present...)
		}

		for _, res := range missingResults {
			q.handleTransferResult(res, retries)
		}
		for res := range results {
			q.handleTransferResult(res, retries)
		}
	}()

	return retries
}

func (q *TransferQueue) partitionTransfers(transfers []*Transfer) (present []*Transfer, results []TransferResult) {
	if q.direction != Upload {
		return transfers, nil
	}

	present = make([]*Transfer, 0, len(transfers))
	results = make([]TransferResult, 0, len(transfers))

	for _, t := range transfers {
		var err error

		if t.Size < 0 {
			err = errors.Errorf("Git LFS: object %q has invalid size (got: %d)", t.Oid, t.Size)
		} else {
			fd, serr := os.Stat(t.Path)
			if serr != nil {
				if os.IsNotExist(serr) {
					err = newObjectMissingError(t.Name, t.Oid)
				} else {
					err = serr
				}
			} else if t.Size != fd.Size() {
				err = newCorruptObjectError(t.Name, t.Oid)
			}
		}

		if err != nil {
			results = append(results, TransferResult{
				Transfer: t,
				Error:    err,
			})
		} else {
			present = append(present, t)
		}
	}

	return
}

// makeDryRunResults returns a channel populated immediately with "successful"
// results for all of the given transfers in "ts".
func (q *TransferQueue) makeDryRunResults(ts []*Transfer) <-chan TransferResult {
	results := make(chan TransferResult, len(ts))
	for _, t := range ts {
		results <- TransferResult{t, nil}
	}

	close(results)

	return results
}

// handleTransferResult observes the transfer result, sending it on the retries
// channel if it was able to be retried.
func (q *TransferQueue) handleTransferResult(
	res TransferResult, retries chan<- *objectTuple,
) {
	oid := res.Transfer.Oid

	if res.Error != nil {
		// If there was an error encountered when processing the
		// transfer (res.Transfer), handle the error as is appropriate:

		if q.canRetryObject(oid, res.Error) {
			// If the object can be retried, send it on the retries
			// channel, where it will be read at the call-site and
			// its retry count will be incremented.
			tracerx.Printf("tq: retrying object %s: %s", oid, res.Error)

			q.trMutex.Lock()
			objects, ok := q.transfers[oid]
			q.trMutex.Unlock()

			if ok {
				retries <- objects.First()
			} else {
				q.errorc <- res.Error
			}
		} else {
			// If the error wasn't retriable, OR the object has
			// exceeded its retry budget, it will be NOT be sent to
			// the retry channel, and the error will be reported
			// immediately.
			q.errorc <- res.Error
			q.wait.Done()
		}
	} else {
		q.trMutex.Lock()
		objects := q.transfers[oid]
		objects.completed = true

		// Otherwise, if the transfer was successful, notify all of the
		// watchers, and mark it as finished.
		for _, c := range q.watchers {
			// Send one update for each transfer with the
			// same OID.
			for _, t := range objects.All() {
				c <- &Transfer{
					Name: t.Name,
					Path: t.Path,
					Oid:  t.Oid,
					Size: t.Size,
				}
			}
		}

		q.trMutex.Unlock()

		q.meter.FinishTransfer(res.Transfer.Name)
		q.wait.Done()
	}
}

func (q *TransferQueue) useAdapter(name string) {
	q.adapterInitMutex.Lock()
	defer q.adapterInitMutex.Unlock()

	if q.adapter != nil {
		if q.adapter.Name() == name {
			// re-use, this is the normal path
			return
		}
		// If the adapter we're using isn't the same as the one we've been
		// told to use now, must wait for the current one to finish then switch
		// This will probably never happen but is just in case server starts
		// changing adapter support in between batches
		q.finishAdapter()
	}
	q.adapter = q.manifest.NewAdapterOrDefault(name, q.direction)
}

func (q *TransferQueue) finishAdapter() {
	if q.adapterInProgress {
		q.adapter.End()
		q.adapterInProgress = false
		q.adapter = nil
	}
}

// BatchSize returns the batch size of the receiving *TransferQueue, or, the
// number of transfers to accept before beginning work on them.
func (q *TransferQueue) BatchSize() int {
	return q.batchSize
}

func (q *TransferQueue) Skip(size int64) {
	q.meter.Skip(size)
}

func (q *TransferQueue) ensureAdapterBegun(e lfsapi.Endpoint) error {
	q.adapterInitMutex.Lock()
	defer q.adapterInitMutex.Unlock()

	if q.adapterInProgress {
		return nil
	}

	// Progress callback - receives byte updates
	cb := func(name string, total, read int64, current int) error {
		q.meter.TransferBytes(q.direction.String(), name, read, total, current)
		if q.cb != nil {
			// NOTE: this is the mechanism by which the logpath
			// specified by GIT_LFS_PROGRESS is written to.
			//
			// See: lfs.downloadFile() for more.
			q.cb(total, read, current)
		}
		return nil
	}

	tracerx.Printf("tq: starting transfer adapter %q", q.adapter.Name())
	err := q.adapter.Begin(q.toAdapterCfg(e), cb)
	if err != nil {
		return err
	}
	q.adapterInProgress = true

	return nil
}

func (q *TransferQueue) toAdapterCfg(e lfsapi.Endpoint) AdapterConfig {
	apiClient := q.manifest.APIClient()
	concurrency := q.manifest.ConcurrentTransfers()
	if apiClient.Endpoints.AccessFor(e.Url) == lfsapi.NTLMAccess {
		concurrency = 1
	}

	return &adapterConfig{
		concurrentTransfers: concurrency,
		apiClient:           apiClient,
		remote:              q.remote,
	}
}

// Wait waits for the queue to finish processing all transfers. Once Wait is
// called, Add will no longer add transfers to the queue. Any failed
// transfers will be automatically retried once.
func (q *TransferQueue) Wait() {
	close(q.incoming)

	q.wait.Wait()
	q.collectorWait.Wait()

	q.finishAdapter()
	close(q.errorc)

	for _, watcher := range q.watchers {
		close(watcher)
	}

	q.meter.Flush()
	q.errorwait.Wait()
}

// Watch returns a channel where the queue will write the value of each transfer
// as it completes. If multiple transfers exist with the same OID, they will all
// be recorded here, even though only one actual transfer took place. The
// channel will be closed when the queue finishes processing.
func (q *TransferQueue) Watch() chan *Transfer {
	c := make(chan *Transfer, q.batchSize)
	q.watchers = append(q.watchers, c)
	return c
}

// This goroutine collects errors returned from transfers
func (q *TransferQueue) errorCollector() {
	for err := range q.errorc {
		q.errors = append(q.errors, err)
	}
	q.errorwait.Done()
}

// run begins the transfer queue. It transfers files sequentially or
// concurrently depending on the Config.ConcurrentTransfers() value.
func (q *TransferQueue) run() {
	tracerx.Printf("tq: running as batched queue, batch size of %d", q.batchSize)

	go q.errorCollector()
	go q.collectBatches()
}

// canRetry returns whether or not the given error "err" is retriable.
func (q *TransferQueue) canRetry(err error) bool {
	return errors.IsRetriableError(err)
}

// canRetryObject returns whether the given error is retriable for the object
// given by "oid". If the an OID has met its retry limit, then it will not be
// able to be retried again. If so, canRetryObject returns whether or not that
// given error "err" is retriable.
func (q *TransferQueue) canRetryObject(oid string, err error) bool {
	if count, ok := q.rc.CanRetry(oid); !ok {
		tracerx.Printf("tq: refusing to retry %q, too many retries (%d)", oid, count)
		return false
	}

	return q.canRetry(err)
}

// Errors returns any errors encountered during transfer.
func (q *TransferQueue) Errors() []error {
	return q.errors
}