flatfs.go 23 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"
22

Jakub Sztandera's avatar
Jakub Sztandera committed
23
	logging "github.com/ipfs/go-log"
Tommi Virtanen's avatar
Tommi Virtanen committed
24 25
)

26 27
var log = logging.Logger("flatfs")

Tommi Virtanen's avatar
Tommi Virtanen committed
28
const (
29 30
	extension                  = ".data"
	diskUsageCheckpointPercent = 1.0
31
	diskUsageCheckpointTimeout = 2.0 * time.Second
Tommi Virtanen's avatar
Tommi Virtanen committed
32 33
)

34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
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
)

58 59 60
type initAccuracy string

const (
61
	unknownA  initAccuracy = "unknown"
62 63 64 65 66 67
	exactA    initAccuracy = "initial-exact"
	approxA   initAccuracy = "initial-approximate"
	timedoutA initAccuracy = "initial-timed-out"
)

func combineAccuracy(a, b initAccuracy) initAccuracy {
68 69 70
	if a == unknownA || b == unknownA {
		return unknownA
	}
71 72 73 74 75 76 77 78 79
	if a == timedoutA || b == timedoutA {
		return timedoutA
	}
	if a == approxA || b == approxA {
		return approxA
	}
	if a == exactA && b == exactA {
		return exactA
	}
80 81 82 83 84 85 86
	if a == "" {
		return b
	}
	if b == "" {
		return a
	}
	return unknownA
87 88
}

89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
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)
)

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
105 106
type Datastore struct {
	path string
107

108 109
	shardStr string
	getDir   ShardFunc
Jeromy's avatar
Jeromy committed
110 111 112

	// sychronize all writes and directory changes for added safety
	sync bool
113

114 115 116 117 118 119
	// atmoic operations should always be used with diskUsage
	diskUsage int64

	// these values should only be used during internalization or
	// inside the checkpoint loop
	dirty       bool
120
	storedValue diskUsageValue
121

Kevin Atkinson's avatar
Kevin Atkinson committed
122 123
	checkpointCh chan struct{}
	done         chan struct{}
124 125 126 127

	// opMap handles concurrent write operations (put/delete)
	// to the same key
	opMap *opMap
Tommi Virtanen's avatar
Tommi Virtanen committed
128 129
}

130
type diskUsageValue struct {
131 132
	DiskUsage int64        `json:"diskUsage"`
	Accuracy  initAccuracy `json:"accuracy"`
133 134
}

135 136
type ShardFunc func(string) string

137
type opT int
138

139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
// 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()
}
187

188
func Create(path string, fun *ShardIdV1) error {
189

kpcyrd's avatar
kpcyrd committed
190
	err := os.Mkdir(path, 0755)
191 192
	if err != nil && !os.IsExist(err) {
		return err
Tommi Virtanen's avatar
Tommi Virtanen committed
193 194
	}

195 196
	dsFun, err := ReadShardFunc(path)
	switch err {
Kevin Atkinson's avatar
Kevin Atkinson committed
197
	case ErrShardingFileMissing:
198 199 200 201 202
		isEmpty, err := DirIsEmpty(path)
		if err != nil {
			return err
		}
		if !isEmpty {
203
			return fmt.Errorf("directory missing %s file: %s", SHARDING_FN, path)
204 205 206
		}

		err = WriteShardFunc(path, fun)
207 208 209
		if err != nil {
			return err
		}
210 211
		err = WriteReadme(path, fun)
		return err
212
	case nil:
213
		if fun.String() != dsFun.String() {
214
			return fmt.Errorf("specified shard func '%s' does not match repo shard func '%s'",
215
				fun.String(), dsFun.String())
216
		}
Kevin Atkinson's avatar
Kevin Atkinson committed
217
		return ErrDatastoreExists
218
	default:
219
		return err
220
	}
221 222
}

223
func Open(path string, syncFiles bool) (*Datastore, error) {
224 225
	_, err := os.Stat(path)
	if os.IsNotExist(err) {
Kevin Atkinson's avatar
Kevin Atkinson committed
226
		return nil, ErrDatastoreDoesNotExist
227 228 229 230
	} else if err != nil {
		return nil, err
	}

231
	shardId, err := ReadShardFunc(path)
232
	if err != nil {
233
		return nil, err
234 235
	}

236
	fs := &Datastore{
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
		path:      path,
		shardStr:  shardId.String(),
		getDir:    shardId.Func(),
		sync:      syncFiles,
		diskUsage: 0,
		opMap:     new(opMap),
	}

	// 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
254
	}
255

Kevin Atkinson's avatar
Kevin Atkinson committed
256 257 258
	fs.checkpointCh = make(chan struct{}, 1)
	fs.done = make(chan struct{})
	go fs.checkpointLoop()
259 260 261
	return fs, nil
}

262
// convenience method
263
func CreateOrOpen(path string, fun *ShardIdV1, sync bool) (*Datastore, error) {
264
	err := Create(path, fun)
Kevin Atkinson's avatar
Kevin Atkinson committed
265
	if err != nil && err != ErrDatastoreExists {
266 267 268 269 270
		return nil, err
	}
	return Open(path, sync)
}

271 272
func (fs *Datastore) ShardStr() string {
	return fs.shardStr
273 274
}

Tommi Virtanen's avatar
Tommi Virtanen committed
275
func (fs *Datastore) encode(key datastore.Key) (dir, file string) {
276
	noslash := key.String()[1:]
Jeromy's avatar
Jeromy committed
277 278
	dir = filepath.Join(fs.path, fs.getDir(noslash))
	file = filepath.Join(dir, noslash+extension)
Tommi Virtanen's avatar
Tommi Virtanen committed
279 280 281
	return dir, file
}

282
func (fs *Datastore) decode(file string) (key datastore.Key, ok bool) {
Jeromy's avatar
Jeromy committed
283
	if filepath.Ext(file) != extension {
284 285 286
		return datastore.Key{}, false
	}
	name := file[:len(file)-len(extension)]
Jeromy's avatar
Jeromy committed
287
	return datastore.NewKey(name), true
288 289
}

290 291
func (fs *Datastore) makeDir(dir string) error {
	if err := fs.makeDirNoSync(dir); err != nil {
Jeromy's avatar
Jeromy committed
292
		return err
293 294 295 296 297 298
	}

	// 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
299 300 301 302
	if fs.sync {
		if err := syncDir(fs.path); err != nil {
			return err
		}
303 304 305 306
	}
	return nil
}

307
func (fs *Datastore) makeDirNoSync(dir string) error {
kpcyrd's avatar
kpcyrd committed
308
	if err := os.Mkdir(dir, 0755); err != nil {
Jeromy's avatar
Jeromy committed
309 310 311 312 313
		// EEXIST is safe to ignore here, that just means the prefix
		// directory already existed.
		if !os.IsExist(err) {
			return err
		}
314
		return nil
Jeromy's avatar
Jeromy committed
315
	}
316 317 318

	// Track DiskUsage of this NEW folder
	fs.updateDiskUsage(dir, true)
Jeromy's avatar
Jeromy committed
319 320 321
	return nil
}

322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
// 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.
337
	err = os.Rename(tmpPath, path)
338 339 340 341
	fs.updateDiskUsage(path, true)
	return err
}

Jeromy's avatar
Jeromy committed
342
var putMaxRetries = 6
343

344 345 346 347 348 349 350 351 352 353
// 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.
Tommi Virtanen's avatar
Tommi Virtanen committed
354 355 356 357 358 359
func (fs *Datastore) Put(key datastore.Key, value interface{}) error {
	val, ok := value.([]byte)
	if !ok {
		return datastore.ErrInvalidType
	}

360
	var err error
Jeromy's avatar
Jeromy committed
361
	for i := 1; i <= putMaxRetries; i++ {
362 363 364 365 366
		err = fs.doWriteOp(&op{
			typ: opPut,
			key: key,
			v:   val,
		})
367
		if err == nil {
Jeromy's avatar
Jeromy committed
368
			break
369 370 371
		}

		if !strings.Contains(err.Error(), "too many open files") {
Jeromy's avatar
Jeromy committed
372
			break
373 374
		}

Or Rikon's avatar
Or Rikon committed
375
		log.Errorf("too many open files, retrying in %dms", 100*i)
Jeromy's avatar
Jeromy committed
376
		time.Sleep(time.Millisecond * 100 * time.Duration(i))
377 378 379 380
	}
	return err
}

381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417
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")
	}
}

// doWrite optmizes out write operations (put/delete) to the same
// key by queueing them and suceeding all queued
// operations if one of them does. In such case,
// we assume that the first suceeding operation
// 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
}

418
func (fs *Datastore) doPut(key datastore.Key, val []byte) error {
419

Tommi Virtanen's avatar
Tommi Virtanen committed
420
	dir, path := fs.encode(key)
421
	if err := fs.makeDir(dir); err != nil {
422
		return err
Tommi Virtanen's avatar
Tommi Virtanen committed
423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444
	}

	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
445
	if fs.sync {
446
		if err := syncFile(tmp); err != nil {
Jeromy's avatar
Jeromy committed
447 448
			return err
		}
449
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
450 451 452 453 454
	if err := tmp.Close(); err != nil {
		return err
	}
	closed = true

455
	err = fs.renameAndUpdateDiskUsage(tmp.Name(), path)
Tommi Virtanen's avatar
Tommi Virtanen committed
456 457 458 459 460
	if err != nil {
		return err
	}
	removed = true

Jeromy's avatar
Jeromy committed
461 462 463 464
	if fs.sync {
		if err := syncDir(dir); err != nil {
			return err
		}
465
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
466 467 468
	return nil
}

Jeromy's avatar
Jeromy committed
469 470
func (fs *Datastore) putMany(data map[datastore.Key]interface{}) error {
	var dirsToSync []string
471
	files := make(map[*os.File]*op)
Jeromy's avatar
Jeromy committed
472 473 474 475 476 477 478

	for key, value := range data {
		val, ok := value.([]byte)
		if !ok {
			return datastore.ErrInvalidType
		}
		dir, path := fs.encode(key)
479
		if err := fs.makeDirNoSync(dir); err != nil {
Jeromy's avatar
Jeromy committed
480 481 482 483 484 485 486 487 488 489 490 491 492
			return err
		}
		dirsToSync = append(dirsToSync, dir)

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

		if _, err := tmp.Write(val); err != nil {
			return err
		}

493 494 495 496 497 498
		files[tmp] = &op{
			typ:  opRename,
			path: path,
			tmp:  tmp.Name(),
			key:  key,
		}
Jeromy's avatar
Jeromy committed
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
	}

	ops := make(map[*os.File]int)

	defer func() {
		for fi, _ := range files {
			val, _ := ops[fi]
			switch val {
			case 0:
				_ = fi.Close()
				fallthrough
			case 1:
				_ = os.Remove(fi.Name())
			}
		}
	}()

	// Now we sync everything
	// sync and close files
	for fi, _ := range files {
Jeromy's avatar
Jeromy committed
519
		if fs.sync {
520
			if err := syncFile(fi); err != nil {
Jeromy's avatar
Jeromy committed
521 522
				return err
			}
Jeromy's avatar
Jeromy committed
523 524 525 526 527 528 529 530 531 532 533
		}

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

		// signify closed
		ops[fi] = 1
	}

	// move files to their proper places
534 535
	for fi, op := range files {
		fs.doWriteOp(op)
Jeromy's avatar
Jeromy committed
536 537 538 539 540
		// signify removed
		ops[fi] = 2
	}

	// now sync the dirs for those files
Jeromy's avatar
Jeromy committed
541 542 543 544 545
	if fs.sync {
		for _, dir := range dirsToSync {
			if err := syncDir(dir); err != nil {
				return err
			}
Jeromy's avatar
Jeromy committed
546 547
		}

Jeromy's avatar
Jeromy committed
548 549 550 551
		// sync top flatfs dir
		if err := syncDir(fs.path); err != nil {
			return err
		}
Jeromy's avatar
Jeromy committed
552 553 554 555 556
	}

	return nil
}

Tommi Virtanen's avatar
Tommi Virtanen committed
557 558 559 560 561 562 563 564 565 566 567 568 569 570
func (fs *Datastore) Get(key datastore.Key) (value interface{}, err error) {
	_, 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
571 572 573 574 575 576 577 578 579
	_, 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
580 581
}

582 583 584
// 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
585
func (fs *Datastore) Delete(key datastore.Key) error {
586 587 588 589 590 591 592 593 594 595
	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
596
	_, path := fs.encode(key)
597 598 599

	fSize := fileSize(path)

Tommi Virtanen's avatar
Tommi Virtanen committed
600 601
	switch err := os.Remove(path); {
	case err == nil:
602 603
		atomic.AddInt64(&fs.diskUsage, -fSize)
		fs.checkpointDiskUsage()
Tommi Virtanen's avatar
Tommi Virtanen committed
604 605 606 607 608 609
		return nil
	case os.IsNotExist(err):
		return datastore.ErrNotFound
	default:
		return err
	}
Tommi Virtanen's avatar
Tommi Virtanen committed
610 611 612
}

func (fs *Datastore) Query(q query.Query) (query.Results, error) {
613 614 615 616 617 618 619 620 621 622 623
	if (q.Prefix != "" && q.Prefix != "/") ||
		len(q.Filters) > 0 ||
		len(q.Orders) > 0 ||
		q.Limit > 0 ||
		q.Offset > 0 ||
		!q.KeysOnly {
		// 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")
	}

624
	reschan := make(chan query.Result, query.KeysOnlyBufSize)
Jeromy's avatar
Jeromy committed
625 626
	go func() {
		defer close(reschan)
627
		err := fs.walkTopLevel(fs.path, reschan)
Jeromy's avatar
Jeromy committed
628
		if err != nil {
629
			reschan <- query.Result{Error: errors.New("walk failed: " + err.Error())}
630
		}
Jeromy's avatar
Jeromy committed
631 632
	}()
	return query.ResultsWithChan(q, reschan), nil
Tommi Virtanen's avatar
Tommi Virtanen committed
633 634
}

635 636 637 638 639
func (fs *Datastore) walkTopLevel(path string, reschan chan query.Result) error {
	dir, err := os.Open(path)
	if err != nil {
		return err
	}
Kevin Atkinson's avatar
Kevin Atkinson committed
640
	defer dir.Close()
641 642 643 644 645
	names, err := dir.Readdirnames(-1)
	if err != nil {
		return err
	}
	for _, dir := range names {
Kevin Atkinson's avatar
Kevin Atkinson committed
646 647 648 649 650

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

651 652 653 654
		err = fs.walk(filepath.Join(path, dir), reschan)
		if err != nil {
			return err
		}
Kevin Atkinson's avatar
Kevin Atkinson committed
655

656 657 658 659
	}
	return nil
}

660 661 662
// folderSize estimates the diskUsage of a folder by reading
// up to DiskUsageFilesAverage entries in it and assumming any
// other files will have an avereage size.
663
func folderSize(path string, deadline time.Time) (int64, initAccuracy, error) {
664 665 666 667
	var du int64

	folder, err := os.Open(path)
	if err != nil {
668
		return 0, "", err
669 670 671 672 673
	}
	defer folder.Close()

	stat, err := folder.Stat()
	if err != nil {
674
		return 0, "", err
675 676 677 678
	}

	files, err := folder.Readdirnames(-1)
	if err != nil {
679
		return 0, "", err
680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696
	}

	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]
	}

697
	accuracy := exactA
698
	for {
699 700 701
		// Do not process any files after deadline is over
		if time.Now().After(deadline) {
			accuracy = timedoutA
702 703 704
			break
		}

705 706 707 708
		if i >= totalFiles || filesProcessed >= maxFiles {
			if filesProcessed >= maxFiles {
				accuracy = approxA
			}
709 710 711 712 713 714 715 716
			break
		}

		// Stat the file
		fname := files[i]
		subpath := filepath.Join(path, fname)
		st, err := os.Stat(subpath)
		if err != nil {
717
			return 0, "", err
718 719 720 721
		}

		// Find folder size recursively
		if st.IsDir() {
722
			du2, acc, err := folderSize(filepath.Join(subpath), deadline)
723
			if err != nil {
724
				return 0, "", err
725
			}
726
			accuracy = combineAccuracy(acc, accuracy)
727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748
			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)
749
	return du, accuracy, nil
750 751 752 753
}

// calculateDiskUsage tries to read the DiskUsageFile for a cached
// diskUsage value, otherwise walks the datastore files.
754
// it is only safe to call in Open()
755 756 757
func (fs *Datastore) calculateDiskUsage() error {
	// Try to obtain a previously stored value from disk
	if persDu := fs.readDiskUsageFile(); persDu > 0 {
758
		fs.diskUsage = persDu
759 760 761 762 763
		return nil
	}

	fmt.Printf("Calculating datastore size. This might take %s at most and will happen only once\n", DiskUsageCalcTimeout.String())
	deadline := time.Now().Add(DiskUsageCalcTimeout)
764
	du, accuracy, err := folderSize(fs.path, deadline)
765 766 767
	if err != nil {
		return err
	}
768
	if accuracy == timedoutA {
769 770 771 772 773 774 775
		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")
	}

776
	fs.storedValue.Accuracy = accuracy
777
	fs.diskUsage = du
778
	fs.writeDiskUsageFile(du, true)
779

780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800
	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 {
801 802
		atomic.AddInt64(&fs.diskUsage, fsize)
		fs.checkpointDiskUsage()
803 804 805
	}
}

806
func (fs *Datastore) checkpointDiskUsage() {
807
	select {
Kevin Atkinson's avatar
Kevin Atkinson committed
808
	case fs.checkpointCh <- struct{}{}:
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
809
		// msg sent
810 811
	default:
		// checkpoint request already pending
812 813 814
	}
}

815
func (fs *Datastore) checkpointLoop() {
816 817
	timerActive := true
	timer := time.NewTimer(0)
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
818
	defer timer.Stop()
819
	for {
820 821 822
		select {
		case _, more := <-fs.checkpointCh:
			du := atomic.LoadInt64(&fs.diskUsage)
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
823
			fs.dirty = true
824
			if !more { // shutting down
825
				fs.writeDiskUsageFile(du, true)
826
				if fs.dirty {
Kevin Atkinson's avatar
Tweaks.  
Kevin Atkinson committed
827
					log.Errorf("could not store final value of disk usage to file, future estimates may be inaccurate")
828
				}
Kevin Atkinson's avatar
Kevin Atkinson committed
829
				fs.done <- struct{}{}
830 831 832 833 834
				return
			}
			// If the difference between the checkpointed disk usage and
			// current one is larger than than `diskUsageCheckpointPercent`
			// of the checkpointed: store it.
835
			newDu := float64(du)
836
			lastCheckpointDu := float64(fs.storedValue.DiskUsage)
837
			diff := math.Abs(newDu - lastCheckpointDu)
838 839
			if lastCheckpointDu*diskUsageCheckpointPercent < diff*100.0 {
				fs.writeDiskUsageFile(du, false)
840
			}
841 842 843 844
			// 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
845
				timerActive = true
846 847 848 849 850
			}
		case <-timer.C:
			timerActive = false
			if fs.dirty {
				du := atomic.LoadInt64(&fs.diskUsage)
851
				fs.writeDiskUsageFile(du, false)
852
			}
853
		}
854
	}
855 856
}

857
func (fs *Datastore) writeDiskUsageFile(du int64, doSync bool) {
858 859
	tmp, err := ioutil.TempFile(fs.path, "du-")
	if err != nil {
860
		log.Warningf("cound not write disk usage: %v", err)
861 862
		return
	}
863

864 865
	toWrite := fs.storedValue
	toWrite.DiskUsage = du
866
	encoder := json.NewEncoder(tmp)
867
	if err := encoder.Encode(&toWrite); err != nil {
868
		log.Warningf("cound not write disk usage: %v", err)
869 870
		return
	}
871

872 873 874 875 876 877 878
	if doSync {
		if err := tmp.Sync(); err != nil {
			log.Warningf("cound not sync %s: %v", DiskUsageFile, err)
			return
		}
	}

879
	if err := tmp.Close(); err != nil {
880
		log.Warningf("cound not write disk usage: %v", err)
881 882 883
		return
	}

884
	if err := os.Rename(tmp.Name(), filepath.Join(fs.path, DiskUsageFile)); err != nil {
885
		log.Warningf("cound not write disk usage: %v", err)
886
		return
887
	}
888

889
	fs.storedValue = toWrite
890
	fs.dirty = false
891 892
}

893
// readDiskUsageFile is only safe to call in Open()
894 895 896 897 898 899
func (fs *Datastore) readDiskUsageFile() int64 {
	fpath := filepath.Join(fs.path, DiskUsageFile)
	duB, err := ioutil.ReadFile(fpath)
	if err != nil {
		return 0
	}
900
	err = json.Unmarshal(duB, &fs.storedValue)
901 902 903
	if err != nil {
		return 0
	}
904
	return fs.storedValue.DiskUsage
905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924
}

// 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
}

925 926 927 928
// 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 {
929
	return string(fs.storedValue.Accuracy)
930 931
}

932 933 934
func (fs *Datastore) walk(path string, reschan chan query.Result) error {
	dir, err := os.Open(path)
	if err != nil {
935 936 937 938
		if os.IsNotExist(err) {
			// not an error if the file disappeared
			return nil
		}
939 940
		return err
	}
Kevin Atkinson's avatar
Kevin Atkinson committed
941
	defer dir.Close()
942 943 944 945 946 947 948 949 950 951

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

952 953 954 955 956 957 958 959 960 961 962 963
	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 {
964
			log.Warningf("failed to decode flatfs entry: %s", fn)
965 966 967 968 969 970 971 972 973 974 975 976
			continue
		}

		reschan <- query.Result{
			Entry: query.Entry{
				Key: key.String(),
			},
		}
	}
	return nil
}

977 978 979 980 981 982 983 984 985
// Deactivate closes background maintenance threads, most write
// operations will fail but readonly operations will continue to
// function
func (fs *Datastore) deactivate() error {
	if fs.checkpointCh != nil {
		close(fs.checkpointCh)
		<-fs.done
		fs.checkpointCh = nil
	}
Jeromy's avatar
Jeromy committed
986 987 988
	return nil
}

989 990 991 992
func (fs *Datastore) Close() error {
	return fs.deactivate()
}

Jeromy's avatar
Jeromy committed
993
type flatfsBatch struct {
Jeromy's avatar
Jeromy committed
994 995 996 997 998 999
	puts    map[datastore.Key]interface{}
	deletes map[datastore.Key]struct{}

	ds *Datastore
}

Jeromy's avatar
Jeromy committed
1000
func (fs *Datastore) Batch() (datastore.Batch, error) {
Jeromy's avatar
Jeromy committed
1001
	return &flatfsBatch{
Jeromy's avatar
Jeromy committed
1002 1003 1004
		puts:    make(map[datastore.Key]interface{}),
		deletes: make(map[datastore.Key]struct{}),
		ds:      fs,
Jeromy's avatar
Jeromy committed
1005
	}, nil
Jeromy's avatar
Jeromy committed
1006 1007
}

Jeromy's avatar
Jeromy committed
1008
func (bt *flatfsBatch) Put(key datastore.Key, val interface{}) error {
Jeromy's avatar
Jeromy committed
1009 1010 1011 1012
	bt.puts[key] = val
	return nil
}

Jeromy's avatar
Jeromy committed
1013
func (bt *flatfsBatch) Delete(key datastore.Key) error {
Jeromy's avatar
Jeromy committed
1014 1015 1016 1017
	bt.deletes[key] = struct{}{}
	return nil
}

Jeromy's avatar
Jeromy committed
1018
func (bt *flatfsBatch) Commit() error {
Jeromy's avatar
Jeromy committed
1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031
	if err := bt.ds.putMany(bt.puts); err != nil {
		return err
	}

	for k, _ := range bt.deletes {
		if err := bt.ds.Delete(k); err != nil {
			return err
		}
	}

	return nil
}

Tommi Virtanen's avatar
Tommi Virtanen committed
1032 1033 1034
var _ datastore.ThreadSafeDatastore = (*Datastore)(nil)

func (*Datastore) IsThreadSafe() {}