flatfs.go 24.4 KB
Newer Older
Tommi Virtanen's avatar
Tommi Virtanen committed
1 2 3 4 5 6
// Package flatfs is a Datastore implementation that stores all
// objects in a two-level directory structure in the local file
// system, regardless of the hierarchy of the keys.
package flatfs

import (
7
	"encoding/json"
Tommi Virtanen's avatar
Tommi Virtanen committed
8
	"errors"
9
	"fmt"
Tommi Virtanen's avatar
Tommi Virtanen committed
10
	"io/ioutil"
11 12
	"math"
	"math/rand"
Tommi Virtanen's avatar
Tommi Virtanen committed
13
	"os"
Jeromy's avatar
Jeromy committed
14
	"path/filepath"
Tommi Virtanen's avatar
Tommi Virtanen committed
15
	"strings"
16 17
	"sync"
	"sync/atomic"
18
	"time"
Tommi Virtanen's avatar
Tommi Virtanen committed
19

Jeromy's avatar
Jeromy committed
20 21
	"github.com/ipfs/go-datastore"
	"github.com/ipfs/go-datastore/query"
Steven Allen's avatar
Steven Allen committed
22
	"github.com/jbenet/goprocess"
23

Jakub Sztandera's avatar
Jakub Sztandera committed
24
	logging "github.com/ipfs/go-log"
Tommi Virtanen's avatar
Tommi Virtanen committed
25 26
)

27 28
var log = logging.Logger("flatfs")

Tommi Virtanen's avatar
Tommi Virtanen committed
29
const (
30
	extension                  = ".data"
31
	diskUsageMessageTimeout    = 5 * time.Second
32
	diskUsageCheckpointPercent = 1.0
33
	diskUsageCheckpointTimeout = 2 * time.Second
Tommi Virtanen's avatar
Tommi Virtanen committed
34 35
)

36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
var (
	// DiskUsageFile is the name of the file to cache the size of the
	// datastore in disk
	DiskUsageFile = "diskUsage.cache"
	// DiskUsageFilesAverage is the maximum number of files per folder
	// to stat in order to calculate the size of the datastore.
	// The size of the rest of the files in a folder will be assumed
	// to be the average of the values obtained. This includes
	// regular files and directories.
	DiskUsageFilesAverage = 2000
	// DiskUsageCalcTimeout is the maximum time to spend
	// calculating the DiskUsage upon a start when no
	// DiskUsageFile is present.
	// If this period did not suffice to read the size of the datastore,
	// the remaining sizes will be stimated.
	DiskUsageCalcTimeout = 5 * time.Minute
)

const (
	opPut = iota
	opDelete
	opRename
)

60 61 62
type initAccuracy string

const (
63
	unknownA  initAccuracy = "unknown"
64 65 66 67 68 69
	exactA    initAccuracy = "initial-exact"
	approxA   initAccuracy = "initial-approximate"
	timedoutA initAccuracy = "initial-timed-out"
)

func combineAccuracy(a, b initAccuracy) initAccuracy {
70 71 72
	if a == unknownA || b == unknownA {
		return unknownA
	}
73 74 75 76 77 78 79 80 81
	if a == timedoutA || b == timedoutA {
		return timedoutA
	}
	if a == approxA || b == approxA {
		return approxA
	}
	if a == exactA && b == exactA {
		return exactA
	}
82 83 84 85 86 87 88
	if a == "" {
		return b
	}
	if b == "" {
		return a
	}
	return unknownA
89 90
}

91 92 93 94 95 96
var _ datastore.Datastore = (*Datastore)(nil)

var (
	ErrDatastoreExists       = errors.New("datastore already exists")
	ErrDatastoreDoesNotExist = errors.New("datastore directory does not exist")
	ErrShardingFileMissing   = fmt.Errorf("%s file not found in datastore", SHARDING_FN)
Steven Allen's avatar
Steven Allen committed
97
	ErrClosed                = errors.New("datastore closed")
98 99 100 101 102 103 104 105 106 107
)

func init() {
	rand.Seed(time.Now().UTC().UnixNano())
}

// Datastore implements the go-datastore Interface.
// Note this datastore cannot guarantee order of concurrent
// write operations to the same key. See the explanation in
// Put().
Tommi Virtanen's avatar
Tommi Virtanen committed
108
type Datastore struct {
109 110 111 112 113
	// atmoic operations should always be used with diskUsage.
	// Must be first in struct to ensure correct alignment
	// (see https://golang.org/pkg/sync/atomic/#pkg-note-BUG)
	diskUsage int64

Tommi Virtanen's avatar
Tommi Virtanen committed
114
	path string
115

116 117
	shardStr string
	getDir   ShardFunc
Jeromy's avatar
Jeromy committed
118 119 120

	// sychronize all writes and directory changes for added safety
	sync bool
121

122 123 124
	// these values should only be used during internalization or
	// inside the checkpoint loop
	dirty       bool
125
	storedValue diskUsageValue
126

Steven Allen's avatar
Steven Allen committed
127
	// Used to trigger a checkpoint.
Kevin Atkinson's avatar
Kevin Atkinson committed
128 129
	checkpointCh chan struct{}
	done         chan struct{}
130

Steven Allen's avatar
Steven Allen committed
131 132 133
	shutdownLock sync.RWMutex
	shutdown     bool

134 135 136
	// opMap handles concurrent write operations (put/delete)
	// to the same key
	opMap *opMap
Tommi Virtanen's avatar
Tommi Virtanen committed
137 138
}

139
type diskUsageValue struct {
140 141
	DiskUsage int64        `json:"diskUsage"`
	Accuracy  initAccuracy `json:"accuracy"`
142 143
}

144 145
type ShardFunc func(string) string

146
type opT int
147

148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
// op wraps useful arguments of write operations
type op struct {
	typ  opT           // operation type
	key  datastore.Key // datastore key. Mandatory.
	tmp  string        // temp file path
	path string        // file path
	v    []byte        // value
}

type opMap struct {
	ops sync.Map
}

type opResult struct {
	mu      sync.RWMutex
	success bool

	opMap *opMap
	name  string
}

// Returns nil if there's nothing to do.
func (m *opMap) Begin(name string) *opResult {
	for {
		myOp := &opResult{opMap: m, name: name}
		myOp.mu.Lock()
		opIface, loaded := m.ops.LoadOrStore(name, myOp)
		if !loaded { // no one else doing ops with this key
			return myOp
		}

		op := opIface.(*opResult)
		// someone else doing ops with this key, wait for
		// the result
		op.mu.RLock()
		if op.success {
			return nil
		}

		// if we are here, we will retry the operation
	}
}

func (o *opResult) Finish(ok bool) {
	o.success = ok
	o.opMap.ops.Delete(o.name)
	o.mu.Unlock()
}
196

197
func Create(path string, fun *ShardIdV1) error {
198

kpcyrd's avatar
kpcyrd committed
199
	err := os.Mkdir(path, 0755)
200 201
	if err != nil && !os.IsExist(err) {
		return err
Tommi Virtanen's avatar
Tommi Virtanen committed
202 203
	}

204 205
	dsFun, err := ReadShardFunc(path)
	switch err {
Kevin Atkinson's avatar
Kevin Atkinson committed
206
	case ErrShardingFileMissing:
207 208 209 210 211
		isEmpty, err := DirIsEmpty(path)
		if err != nil {
			return err
		}
		if !isEmpty {
212
			return fmt.Errorf("directory missing %s file: %s", SHARDING_FN, path)
213 214 215
		}

		err = WriteShardFunc(path, fun)
216 217 218
		if err != nil {
			return err
		}
219 220
		err = WriteReadme(path, fun)
		return err
221
	case nil:
222
		if fun.String() != dsFun.String() {
223
			return fmt.Errorf("specified shard func '%s' does not match repo shard func '%s'",
224
				fun.String(), dsFun.String())
225
		}
Kevin Atkinson's avatar
Kevin Atkinson committed
226
		return ErrDatastoreExists
227
	default:
228
		return err
229
	}
230 231
}

232
func Open(path string, syncFiles bool) (*Datastore, error) {
233 234
	_, err := os.Stat(path)
	if os.IsNotExist(err) {
Kevin Atkinson's avatar
Kevin Atkinson committed
235
		return nil, ErrDatastoreDoesNotExist
236 237 238 239
	} else if err != nil {
		return nil, err
	}

240
	shardId, err := ReadShardFunc(path)
241
	if err != nil {
242
		return nil, err
243 244
	}

245
	fs := &Datastore{
Steven Allen's avatar
Steven Allen committed
246 247 248 249 250 251 252 253
		path:         path,
		shardStr:     shardId.String(),
		getDir:       shardId.Func(),
		sync:         syncFiles,
		checkpointCh: make(chan struct{}, 1),
		done:         make(chan struct{}),
		diskUsage:    0,
		opMap:        new(opMap),
254 255 256 257 258 259 260 261 262 263 264
	}

	// This sets diskUsage to the correct value
	// It might be slow, but allowing it to happen
	// while the datastore is usable might
	// cause diskUsage to not be accurate.
	err = fs.calculateDiskUsage()
	if err != nil {
		// Cannot stat() all
		// elements in the datastore.
		return nil, err
265
	}
266

Kevin Atkinson's avatar
Kevin Atkinson committed
267
	go fs.checkpointLoop()
268 269 270
	return fs, nil
}

271
// convenience method
272
func CreateOrOpen(path string, fun *ShardIdV1, sync bool) (*Datastore, error) {
273
	err := Create(path, fun)
Kevin Atkinson's avatar
Kevin Atkinson committed
274
	if err != nil && err != ErrDatastoreExists {
275 276 277 278 279
		return nil, err
	}
	return Open(path, sync)
}

280 281
func (fs *Datastore) ShardStr() string {
	return fs.shardStr
282 283
}

Tommi Virtanen's avatar
Tommi Virtanen committed
284
func (fs *Datastore) encode(key datastore.Key) (dir, file string) {
285
	noslash := key.String()[1:]
Jeromy's avatar
Jeromy committed
286 287
	dir = filepath.Join(fs.path, fs.getDir(noslash))
	file = filepath.Join(dir, noslash+extension)
Tommi Virtanen's avatar
Tommi Virtanen committed
288 289 290
	return dir, file
}

291
func (fs *Datastore) decode(file string) (key datastore.Key, ok bool) {
Jeromy's avatar
Jeromy committed
292
	if filepath.Ext(file) != extension {
293 294 295
		return datastore.Key{}, false
	}
	name := file[:len(file)-len(extension)]
Jeromy's avatar
Jeromy committed
296
	return datastore.NewKey(name), true
297 298
}

299 300
func (fs *Datastore) makeDir(dir string) error {
	if err := fs.makeDirNoSync(dir); err != nil {
Jeromy's avatar
Jeromy committed
301
		return err
302 303 304 305 306 307
	}

	// In theory, if we create a new prefix dir and add a file to
	// it, the creation of the prefix dir itself might not be
	// durable yet. Sync the root dir after a successful mkdir of
	// a prefix dir, just to be paranoid.
Jeromy's avatar
Jeromy committed
308 309 310 311
	if fs.sync {
		if err := syncDir(fs.path); err != nil {
			return err
		}
312 313 314 315
	}
	return nil
}

316
func (fs *Datastore) makeDirNoSync(dir string) error {
kpcyrd's avatar
kpcyrd committed
317
	if err := os.Mkdir(dir, 0755); err != nil {
Jeromy's avatar
Jeromy committed
318 319 320 321 322
		// EEXIST is safe to ignore here, that just means the prefix
		// directory already existed.
		if !os.IsExist(err) {
			return err
		}
323
		return nil
Jeromy's avatar
Jeromy committed
324
	}
325 326 327

	// Track DiskUsage of this NEW folder
	fs.updateDiskUsage(dir, true)
Jeromy's avatar
Jeromy committed
328 329 330
	return nil
}

331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
// This function always runs under an opLock. Therefore, only one thread is
// touching the affected files.
func (fs *Datastore) renameAndUpdateDiskUsage(tmpPath, path string) error {
	fi, err := os.Stat(path)

	// Destination exists, we need to discount it from diskUsage
	if fs != nil && err == nil {
		atomic.AddInt64(&fs.diskUsage, -fi.Size())
	} else if !os.IsNotExist(err) {
		return err
	}

	// Rename and add new file's diskUsage. If the rename fails,
	// it will either a) Re-add the size of an existing file, which
	// was sustracted before b) Add 0 if there is no existing file.
346
	err = os.Rename(tmpPath, path)
347 348 349 350
	fs.updateDiskUsage(path, true)
	return err
}

Jeromy's avatar
Jeromy committed
351
var putMaxRetries = 6
352

353 354 355 356 357 358 359 360 361 362
// Put stores a key/value in the datastore.
//
// Note, that we do not guarantee order of write operations (Put or Delete)
// to the same key in this datastore.
//
// For example. i.e. in the case of two concurrent Put, we only guarantee
// that one of them will come through, but cannot assure which one even if
// one arrived slightly later than the other. In the case of a
// concurrent Put and a Delete operation, we cannot guarantee which one
// will win.
363
func (fs *Datastore) Put(key datastore.Key, value []byte) error {
Steven Allen's avatar
Steven Allen committed
364 365 366 367 368 369
	fs.shutdownLock.RLock()
	defer fs.shutdownLock.RUnlock()
	if fs.shutdown {
		return ErrClosed
	}

370
	var err error
Jeromy's avatar
Jeromy committed
371
	for i := 1; i <= putMaxRetries; i++ {
372 373 374
		err = fs.doWriteOp(&op{
			typ: opPut,
			key: key,
375
			v:   value,
376
		})
377
		if err == nil {
Jeromy's avatar
Jeromy committed
378
			break
379 380 381
		}

		if !strings.Contains(err.Error(), "too many open files") {
Jeromy's avatar
Jeromy committed
382
			break
383 384
		}

Or Rikon's avatar
Or Rikon committed
385
		log.Errorf("too many open files, retrying in %dms", 100*i)
Jeromy's avatar
Jeromy committed
386
		time.Sleep(time.Millisecond * 100 * time.Duration(i))
387 388 389 390
	}
	return err
}

391 392 393 394 395 396 397 398 399 400 401 402 403
func (fs *Datastore) doOp(oper *op) error {
	switch oper.typ {
	case opPut:
		return fs.doPut(oper.key, oper.v)
	case opDelete:
		return fs.doDelete(oper.key)
	case opRename:
		return fs.renameAndUpdateDiskUsage(oper.tmp, oper.path)
	default:
		panic("bad operation, this is a bug")
	}
}

404 405
// doWrite optimizes out write operations (put/delete) to the same
// key by queueing them and succeeding all queued
406
// operations if one of them does. In such case,
407
// we assume that the first succeeding operation
408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
// on that key was the last one to happen after
// all successful others.
func (fs *Datastore) doWriteOp(oper *op) error {
	keyStr := oper.key.String()

	opRes := fs.opMap.Begin(keyStr)
	if opRes == nil { // nothing to do, a concurrent op succeeded
		return nil
	}

	// Do the operation
	err := fs.doOp(oper)

	// Finish it. If no error, it will signal other operations
	// waiting on this result to succeed. Otherwise, they will
	// retry.
	opRes.Finish(err == nil)
	return err
}

428
func (fs *Datastore) doPut(key datastore.Key, val []byte) error {
429

Tommi Virtanen's avatar
Tommi Virtanen committed
430
	dir, path := fs.encode(key)
431
	if err := fs.makeDir(dir); err != nil {
432
		return err
Tommi Virtanen's avatar
Tommi Virtanen committed
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454
	}

	tmp, err := ioutil.TempFile(dir, "put-")
	if err != nil {
		return err
	}
	closed := false
	removed := false
	defer func() {
		if !closed {
			// silence errcheck
			_ = tmp.Close()
		}
		if !removed {
			// silence errcheck
			_ = os.Remove(tmp.Name())
		}
	}()

	if _, err := tmp.Write(val); err != nil {
		return err
	}
Jeromy's avatar
Jeromy committed
455
	if fs.sync {
456
		if err := syncFile(tmp); err != nil {
Jeromy's avatar
Jeromy committed
457 458
			return err
		}
459
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
460 461 462 463 464
	if err := tmp.Close(); err != nil {
		return err
	}
	closed = true

465
	err = fs.renameAndUpdateDiskUsage(tmp.Name(), path)
Tommi Virtanen's avatar
Tommi Virtanen committed
466 467 468 469 470
	if err != nil {
		return err
	}
	removed = true

Jeromy's avatar
Jeromy committed
471 472 473 474
	if fs.sync {
		if err := syncDir(dir); err != nil {
			return err
		}
475
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
476 477 478
	return nil
}

479
func (fs *Datastore) putMany(data map[datastore.Key][]byte) error {
Steven Allen's avatar
Steven Allen committed
480 481 482 483 484 485
	fs.shutdownLock.RLock()
	defer fs.shutdownLock.RUnlock()
	if fs.shutdown {
		return ErrClosed
	}

Jeromy's avatar
Jeromy committed
486 487
	var dirsToSync []string

488 489 490 491 492
	files := make(map[*os.File]*op, len(data))
	ops := make(map[*os.File]int, len(data))

	defer func() {
		for fi := range files {
Steven Allen's avatar
Steven Allen committed
493
			val := ops[fi]
494 495 496 497 498 499 500
			switch val {
			case 0:
				_ = fi.Close()
				fallthrough
			case 1:
				_ = os.Remove(fi.Name())
			}
Jeromy's avatar
Jeromy committed
501
		}
502 503 504
	}()

	for key, value := range data {
Jeromy's avatar
Jeromy committed
505
		dir, path := fs.encode(key)
506
		if err := fs.makeDirNoSync(dir); err != nil {
Jeromy's avatar
Jeromy committed
507 508 509 510 511 512 513 514 515
			return err
		}
		dirsToSync = append(dirsToSync, dir)

		tmp, err := ioutil.TempFile(dir, "put-")
		if err != nil {
			return err
		}

516
		if _, err := tmp.Write(value); err != nil {
Jeromy's avatar
Jeromy committed
517 518 519
			return err
		}

520 521 522 523 524 525
		files[tmp] = &op{
			typ:  opRename,
			path: path,
			tmp:  tmp.Name(),
			key:  key,
		}
Jeromy's avatar
Jeromy committed
526 527 528 529
	}

	// Now we sync everything
	// sync and close files
530
	for fi := range files {
Jeromy's avatar
Jeromy committed
531
		if fs.sync {
532
			if err := syncFile(fi); err != nil {
Jeromy's avatar
Jeromy committed
533 534
				return err
			}
Jeromy's avatar
Jeromy committed
535 536 537 538 539 540 541 542 543 544 545
		}

		if err := fi.Close(); err != nil {
			return err
		}

		// signify closed
		ops[fi] = 1
	}

	// move files to their proper places
546
	for fi, op := range files {
547 548 549 550
		err := fs.doWriteOp(op)
		if err != nil {
			return err
		}
Jeromy's avatar
Jeromy committed
551 552 553 554 555
		// signify removed
		ops[fi] = 2
	}

	// now sync the dirs for those files
Jeromy's avatar
Jeromy committed
556 557 558 559 560
	if fs.sync {
		for _, dir := range dirsToSync {
			if err := syncDir(dir); err != nil {
				return err
			}
Jeromy's avatar
Jeromy committed
561 562
		}

Jeromy's avatar
Jeromy committed
563 564 565 566
		// sync top flatfs dir
		if err := syncDir(fs.path); err != nil {
			return err
		}
Jeromy's avatar
Jeromy committed
567 568 569 570 571
	}

	return nil
}

572
func (fs *Datastore) Get(key datastore.Key) (value []byte, err error) {
Tommi Virtanen's avatar
Tommi Virtanen committed
573 574 575 576 577 578 579 580 581 582 583 584 585
	_, path := fs.encode(key)
	data, err := ioutil.ReadFile(path)
	if err != nil {
		if os.IsNotExist(err) {
			return nil, datastore.ErrNotFound
		}
		// no specific error to return, so just pass it through
		return nil, err
	}
	return data, nil
}

func (fs *Datastore) Has(key datastore.Key) (exists bool, err error) {
Tommi Virtanen's avatar
Tommi Virtanen committed
586 587 588 589 590 591 592 593 594
	_, path := fs.encode(key)
	switch _, err := os.Stat(path); {
	case err == nil:
		return true, nil
	case os.IsNotExist(err):
		return false, nil
	default:
		return false, err
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
595 596
}

Steven Allen's avatar
Steven Allen committed
597 598 599 600 601 602 603 604 605 606 607 608
func (fs *Datastore) GetSize(key datastore.Key) (size int, err error) {
	_, path := fs.encode(key)
	switch s, err := os.Stat(path); {
	case err == nil:
		return int(s.Size()), nil
	case os.IsNotExist(err):
		return -1, datastore.ErrNotFound
	default:
		return -1, err
	}
}

609 610 611
// Delete removes a key/value from the Datastore. Please read
// the Put() explanation about the handling of concurrent write
// operations to the same key.
Tommi Virtanen's avatar
Tommi Virtanen committed
612
func (fs *Datastore) Delete(key datastore.Key) error {
Steven Allen's avatar
Steven Allen committed
613 614 615 616 617 618
	fs.shutdownLock.RLock()
	defer fs.shutdownLock.RUnlock()
	if fs.shutdown {
		return ErrClosed
	}

619 620 621 622 623 624 625 626 627 628
	return fs.doWriteOp(&op{
		typ: opDelete,
		key: key,
		v:   nil,
	})
}

// This function always runs within an opLock for the given
// key, and not concurrently.
func (fs *Datastore) doDelete(key datastore.Key) error {
Tommi Virtanen's avatar
Tommi Virtanen committed
629
	_, path := fs.encode(key)
630 631 632

	fSize := fileSize(path)

Tommi Virtanen's avatar
Tommi Virtanen committed
633 634
	switch err := os.Remove(path); {
	case err == nil:
635 636
		atomic.AddInt64(&fs.diskUsage, -fSize)
		fs.checkpointDiskUsage()
Tommi Virtanen's avatar
Tommi Virtanen committed
637 638
		return nil
	case os.IsNotExist(err):
Steven Allen's avatar
Steven Allen committed
639
		return nil
Tommi Virtanen's avatar
Tommi Virtanen committed
640 641 642
	default:
		return err
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
643 644 645
}

func (fs *Datastore) Query(q query.Query) (query.Results, error) {
646 647 648 649 650
	if (q.Prefix != "" && q.Prefix != "/") ||
		len(q.Filters) > 0 ||
		len(q.Orders) > 0 ||
		q.Limit > 0 ||
		q.Offset > 0 ||
651 652 653
		!q.KeysOnly ||
		q.ReturnExpirations ||
		q.ReturnsSizes {
654 655 656 657 658
		// TODO this is overly simplistic, but the only caller is
		// `ipfs refs local` for now, and this gets us moving.
		return nil, errors.New("flatfs only supports listing all keys in random order")
	}

Steven Allen's avatar
Steven Allen committed
659 660 661 662 663 664 665
	// Replicates the logic in ResultsWithChan but actually respects calls
	// to `Close`.
	b := query.NewResultBuilder(q)
	b.Process.Go(func(p goprocess.Process) {
		err := fs.walkTopLevel(fs.path, b)
		if err == nil {
			return
666
		}
Steven Allen's avatar
Steven Allen committed
667 668 669 670 671
		select {
		case b.Output <- query.Result{Error: errors.New("walk failed: " + err.Error())}:
		case <-p.Closing():
		}
	})
Steven Allen's avatar
Steven Allen committed
672
	go b.Process.CloseAfterChildren() //nolint
Steven Allen's avatar
Steven Allen committed
673 674

	return b.Results(), nil
Tommi Virtanen's avatar
Tommi Virtanen committed
675 676
}

Steven Allen's avatar
Steven Allen committed
677
func (fs *Datastore) walkTopLevel(path string, result *query.ResultBuilder) error {
678 679 680 681
	dir, err := os.Open(path)
	if err != nil {
		return err
	}
Kevin Atkinson's avatar
Kevin Atkinson committed
682
	defer dir.Close()
683 684 685 686 687
	names, err := dir.Readdirnames(-1)
	if err != nil {
		return err
	}
	for _, dir := range names {
Kevin Atkinson's avatar
Kevin Atkinson committed
688 689 690 691
		if len(dir) == 0 || dir[0] == '.' {
			continue
		}

Steven Allen's avatar
Steven Allen committed
692
		err = fs.walk(filepath.Join(path, dir), result)
693 694 695
		if err != nil {
			return err
		}
Kevin Atkinson's avatar
Kevin Atkinson committed
696

Steven Allen's avatar
Steven Allen committed
697 698 699 700 701 702
		// Are we closing?
		select {
		case <-result.Process.Closing():
			return nil
		default:
		}
703 704 705 706
	}
	return nil
}

707
// folderSize estimates the diskUsage of a folder by reading
708 709
// up to DiskUsageFilesAverage entries in it and assuming any
// other files will have an average size.
710
func folderSize(path string, deadline time.Time) (int64, initAccuracy, error) {
711 712 713 714
	var du int64

	folder, err := os.Open(path)
	if err != nil {
715
		return 0, "", err
716 717 718 719 720
	}
	defer folder.Close()

	stat, err := folder.Stat()
	if err != nil {
721
		return 0, "", err
722 723 724 725
	}

	files, err := folder.Readdirnames(-1)
	if err != nil {
726
		return 0, "", err
727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743
	}

	totalFiles := len(files)
	i := 0
	filesProcessed := 0
	maxFiles := DiskUsageFilesAverage
	if maxFiles <= 0 {
		maxFiles = totalFiles
	}

	// randomize file order
	// https://stackoverflow.com/a/42776696
	for i := len(files) - 1; i > 0; i-- {
		j := rand.Intn(i + 1)
		files[i], files[j] = files[j], files[i]
	}

744
	accuracy := exactA
745
	for {
746 747 748
		// Do not process any files after deadline is over
		if time.Now().After(deadline) {
			accuracy = timedoutA
749 750 751
			break
		}

752 753 754 755
		if i >= totalFiles || filesProcessed >= maxFiles {
			if filesProcessed >= maxFiles {
				accuracy = approxA
			}
756 757 758 759 760 761 762 763
			break
		}

		// Stat the file
		fname := files[i]
		subpath := filepath.Join(path, fname)
		st, err := os.Stat(subpath)
		if err != nil {
764
			return 0, "", err
765 766 767 768
		}

		// Find folder size recursively
		if st.IsDir() {
769
			du2, acc, err := folderSize(filepath.Join(subpath), deadline)
770
			if err != nil {
771
				return 0, "", err
772
			}
773
			accuracy = combineAccuracy(acc, accuracy)
774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795
			du += du2
			filesProcessed++
		} else { // in any other case, add the file size
			du += st.Size()
			filesProcessed++
		}

		i++
	}

	nonProcessed := totalFiles - filesProcessed

	// Avg is total size in this folder up to now / total files processed
	// it includes folders ant not folders
	avg := 0.0
	if filesProcessed > 0 {
		avg = float64(du) / float64(filesProcessed)
	}
	duEstimation := int64(avg * float64(nonProcessed))
	du += duEstimation
	du += stat.Size()
	//fmt.Println(path, "total:", totalFiles, "totalStat:", i, "totalFile:", filesProcessed, "left:", nonProcessed, "avg:", int(avg), "est:", int(duEstimation), "du:", du)
796
	return du, accuracy, nil
797 798 799 800
}

// calculateDiskUsage tries to read the DiskUsageFile for a cached
// diskUsage value, otherwise walks the datastore files.
801
// it is only safe to call in Open()
802 803 804
func (fs *Datastore) calculateDiskUsage() error {
	// Try to obtain a previously stored value from disk
	if persDu := fs.readDiskUsageFile(); persDu > 0 {
805
		fs.diskUsage = persDu
806 807 808
		return nil
	}

809 810 811 812 813 814 815
	msgDone := make(chan struct{}, 1) // prevent race condition
	msgTimer := time.AfterFunc(diskUsageMessageTimeout, func() {
		fmt.Printf("Calculating datastore size. This might take %s at most and will happen only once\n",
			DiskUsageCalcTimeout.String())
		msgDone <- struct{}{}
	})
	defer msgTimer.Stop()
816
	deadline := time.Now().Add(DiskUsageCalcTimeout)
817
	du, accuracy, err := folderSize(fs.path, deadline)
818 819 820
	if err != nil {
		return err
	}
821 822 823
	if !msgTimer.Stop() {
		<-msgDone
	}
824
	if accuracy == timedoutA {
825 826 827 828 829 830 831
		fmt.Println("WARN: It took to long to calculate the datastore size")
		fmt.Printf("WARN: The total size (%d) is an estimation. You can fix errors by\n", du)
		fmt.Printf("WARN: replacing the %s file with the right disk usage in bytes and\n",
			filepath.Join(fs.path, DiskUsageFile))
		fmt.Println("WARN: re-opening the datastore")
	}

832
	fs.storedValue.Accuracy = accuracy
833
	fs.diskUsage = du
834
	fs.writeDiskUsageFile(du, true)
835

836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856
	return nil
}

func fileSize(path string) int64 {
	fi, err := os.Stat(path)
	if err != nil {
		return 0
	}
	return fi.Size()
}

// updateDiskUsage reads the size of path and atomically
// increases or decreases the diskUsage variable.
// setting add to false will subtract from disk usage.
func (fs *Datastore) updateDiskUsage(path string, add bool) {
	fsize := fileSize(path)
	if !add {
		fsize = -fsize
	}

	if fsize != 0 {
857 858
		atomic.AddInt64(&fs.diskUsage, fsize)
		fs.checkpointDiskUsage()
859 860 861
	}
}

862
func (fs *Datastore) checkpointDiskUsage() {
863
	select {
Kevin Atkinson's avatar
Kevin Atkinson committed
864
	case fs.checkpointCh <- struct{}{}:
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
865
		// msg sent
866 867
	default:
		// checkpoint request already pending
868 869 870
	}
}

871
func (fs *Datastore) checkpointLoop() {
Steven Allen's avatar
Steven Allen committed
872 873
	defer close(fs.done)

874 875
	timerActive := true
	timer := time.NewTimer(0)
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
876
	defer timer.Stop()
877
	for {
878 879 880
		select {
		case _, more := <-fs.checkpointCh:
			du := atomic.LoadInt64(&fs.diskUsage)
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
881
			fs.dirty = true
882
			if !more { // shutting down
883
				fs.writeDiskUsageFile(du, true)
884
				if fs.dirty {
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
885
					log.Errorf("could not store final value of disk usage to file, future estimates may be inaccurate")
886
				}
887 888 889 890 891
				return
			}
			// If the difference between the checkpointed disk usage and
			// current one is larger than than `diskUsageCheckpointPercent`
			// of the checkpointed: store it.
892
			newDu := float64(du)
893
			lastCheckpointDu := float64(fs.storedValue.DiskUsage)
894
			diff := math.Abs(newDu - lastCheckpointDu)
895 896
			if lastCheckpointDu*diskUsageCheckpointPercent < diff*100.0 {
				fs.writeDiskUsageFile(du, false)
897
			}
898 899 900 901
			// Otherwise insure the value will be written to disk after
			// `diskUsageCheckpointTimeout`
			if fs.dirty && !timerActive {
				timer.Reset(diskUsageCheckpointTimeout)
Kevin Atkinson's avatar
Kevin Atkinson committed
902
				timerActive = true
903 904 905 906 907
			}
		case <-timer.C:
			timerActive = false
			if fs.dirty {
				du := atomic.LoadInt64(&fs.diskUsage)
908
				fs.writeDiskUsageFile(du, false)
909
			}
910
		}
911
	}
912 913
}

914
func (fs *Datastore) writeDiskUsageFile(du int64, doSync bool) {
915
	tmp, err := ioutil.TempFile(fs.path, "du-")
916 917 918 919 920
	if err != nil {
		log.Warningf("cound not write disk usage: %v", err)
		return
	}

921 922 923 924 925 926 927
	removed := false
	defer func() {
		if !removed {
			// silence errcheck
			_ = os.Remove(tmp.Name())
		}
	}()
928

929 930
	toWrite := fs.storedValue
	toWrite.DiskUsage = du
931
	encoder := json.NewEncoder(tmp)
932
	if err := encoder.Encode(&toWrite); err != nil {
933
		log.Warningf("cound not write disk usage: %v", err)
934 935
		return
	}
936

937 938 939 940 941 942 943
	if doSync {
		if err := tmp.Sync(); err != nil {
			log.Warningf("cound not sync %s: %v", DiskUsageFile, err)
			return
		}
	}

944
	if err := tmp.Close(); err != nil {
945
		log.Warningf("cound not write disk usage: %v", err)
946 947 948
		return
	}

949
	if err := os.Rename(tmp.Name(), filepath.Join(fs.path, DiskUsageFile)); err != nil {
950
		log.Warningf("cound not write disk usage: %v", err)
951
		return
952
	}
953
	removed = true
954

955
	fs.storedValue = toWrite
956
	fs.dirty = false
957 958
}

959
// readDiskUsageFile is only safe to call in Open()
960 961 962 963 964 965
func (fs *Datastore) readDiskUsageFile() int64 {
	fpath := filepath.Join(fs.path, DiskUsageFile)
	duB, err := ioutil.ReadFile(fpath)
	if err != nil {
		return 0
	}
966
	err = json.Unmarshal(duB, &fs.storedValue)
967 968 969
	if err != nil {
		return 0
	}
970
	return fs.storedValue.DiskUsage
971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990
}

// DiskUsage implements the PersistentDatastore interface
// and returns the current disk usage in bytes used by
// this datastore.
//
// The size is approximative and may slightly differ from
// the real disk values.
func (fs *Datastore) DiskUsage() (uint64, error) {
	// it may differ from real disk values if
	// the filesystem has allocated for blocks
	// for a directory because it has many files in it
	// we don't account for "resized" directories.
	// In a large datastore, the differences should be
	// are negligible though.

	du := atomic.LoadInt64(&fs.diskUsage)
	return uint64(du), nil
}

991 992 993 994
// Accuracy returns a string representing the accuracy of the
// DiskUsage() result, the value returned is implementation defined
// and for informational purposes only
func (fs *Datastore) Accuracy() string {
995
	return string(fs.storedValue.Accuracy)
996 997
}

Steven Allen's avatar
Steven Allen committed
998
func (fs *Datastore) walk(path string, result *query.ResultBuilder) error {
999 1000
	dir, err := os.Open(path)
	if err != nil {
1001 1002 1003 1004
		if os.IsNotExist(err) {
			// not an error if the file disappeared
			return nil
		}
1005 1006
		return err
	}
Kevin Atkinson's avatar
Kevin Atkinson committed
1007
	defer dir.Close()
1008 1009 1010 1011 1012 1013 1014 1015 1016 1017

	// ignore non-directories
	fileInfo, err := dir.Stat()
	if err != nil {
		return err
	}
	if !fileInfo.IsDir() {
		return nil
	}

1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029
	names, err := dir.Readdirnames(-1)
	if err != nil {
		return err
	}
	for _, fn := range names {

		if len(fn) == 0 || fn[0] == '.' {
			continue
		}

		key, ok := fs.decode(fn)
		if !ok {
1030
			log.Warningf("failed to decode flatfs entry: %s", fn)
1031 1032 1033
			continue
		}

Steven Allen's avatar
Steven Allen committed
1034 1035
		select {
		case result.Output <- query.Result{
1036 1037 1038
			Entry: query.Entry{
				Key: key.String(),
			},
Steven Allen's avatar
Steven Allen committed
1039 1040 1041
		}:
		case <-result.Process.Closing():
			return nil
1042 1043 1044 1045 1046
		}
	}
	return nil
}

1047 1048 1049
// Deactivate closes background maintenance threads, most write
// operations will fail but readonly operations will continue to
// function
Steven Allen's avatar
Steven Allen committed
1050
func (fs *Datastore) deactivate() {
Steven Allen's avatar
Steven Allen committed
1051 1052 1053
	fs.shutdownLock.Lock()
	defer fs.shutdownLock.Unlock()
	if fs.shutdown {
Steven Allen's avatar
Steven Allen committed
1054
		return
1055
	}
Steven Allen's avatar
Steven Allen committed
1056 1057 1058
	fs.shutdown = true
	close(fs.checkpointCh)
	<-fs.done
Jeromy's avatar
Jeromy committed
1059 1060
}

1061
func (fs *Datastore) Close() error {
Steven Allen's avatar
Steven Allen committed
1062 1063
	fs.deactivate()
	return nil
1064 1065
}

Jeromy's avatar
Jeromy committed
1066
type flatfsBatch struct {
1067
	puts    map[datastore.Key][]byte
Jeromy's avatar
Jeromy committed
1068 1069 1070 1071 1072
	deletes map[datastore.Key]struct{}

	ds *Datastore
}

Jeromy's avatar
Jeromy committed
1073
func (fs *Datastore) Batch() (datastore.Batch, error) {
Jeromy's avatar
Jeromy committed
1074
	return &flatfsBatch{
1075
		puts:    make(map[datastore.Key][]byte),
Jeromy's avatar
Jeromy committed
1076 1077
		deletes: make(map[datastore.Key]struct{}),
		ds:      fs,
Jeromy's avatar
Jeromy committed
1078
	}, nil
Jeromy's avatar
Jeromy committed
1079 1080
}

1081
func (bt *flatfsBatch) Put(key datastore.Key, val []byte) error {
Jeromy's avatar
Jeromy committed
1082 1083 1084 1085
	bt.puts[key] = val
	return nil
}

Jeromy's avatar
Jeromy committed
1086
func (bt *flatfsBatch) Delete(key datastore.Key) error {
Jeromy's avatar
Jeromy committed
1087 1088 1089 1090
	bt.deletes[key] = struct{}{}
	return nil
}

Jeromy's avatar
Jeromy committed
1091
func (bt *flatfsBatch) Commit() error {
Jeromy's avatar
Jeromy committed
1092 1093 1094 1095
	if err := bt.ds.putMany(bt.puts); err != nil {
		return err
	}

Steven Allen's avatar
Steven Allen committed
1096
	for k := range bt.deletes {
Jeromy's avatar
Jeromy committed
1097 1098 1099 1100 1101 1102 1103
		if err := bt.ds.Delete(k); err != nil {
			return err
		}
	}

	return nil
}