1package bolt 2 3import ( 4 "fmt" 5 "io" 6 "os" 7 "sort" 8 "strings" 9 "time" 10 "unsafe" 11) 12 13// txid represents the internal transaction identifier. 14type txid uint64 15 16// Tx represents a read-only or read/write transaction on the database. 17// Read-only transactions can be used for retrieving values for keys and creating cursors. 18// Read/write transactions can create and remove buckets and create and remove keys. 19// 20// IMPORTANT: You must commit or rollback transactions when you are done with 21// them. Pages can not be reclaimed by the writer until no more transactions 22// are using them. A long running read transaction can cause the database to 23// quickly grow. 24type Tx struct { 25 writable bool 26 managed bool 27 db *DB 28 meta *meta 29 root Bucket 30 pages map[pgid]*page 31 stats TxStats 32 commitHandlers []func() 33 34 // WriteFlag specifies the flag for write-related methods like WriteTo(). 35 // Tx opens the database file with the specified flag to copy the data. 36 // 37 // By default, the flag is unset, which works well for mostly in-memory 38 // workloads. For databases that are much larger than available RAM, 39 // set the flag to syscall.O_DIRECT to avoid trashing the page cache. 40 WriteFlag int 41} 42 43// init initializes the transaction. 44func (tx *Tx) init(db *DB) { 45 tx.db = db 46 tx.pages = nil 47 48 // Copy the meta page since it can be changed by the writer. 49 tx.meta = &meta{} 50 db.meta().copy(tx.meta) 51 52 // Copy over the root bucket. 53 tx.root = newBucket(tx) 54 tx.root.bucket = &bucket{} 55 *tx.root.bucket = tx.meta.root 56 57 // Increment the transaction id and add a page cache for writable transactions. 58 if tx.writable { 59 tx.pages = make(map[pgid]*page) 60 tx.meta.txid += txid(1) 61 } 62} 63 64// ID returns the transaction id. 65func (tx *Tx) ID() int { 66 return int(tx.meta.txid) 67} 68 69// DB returns a reference to the database that created the transaction. 70func (tx *Tx) DB() *DB { 71 return tx.db 72} 73 74// Size returns current database size in bytes as seen by this transaction. 75func (tx *Tx) Size() int64 { 76 return int64(tx.meta.pgid) * int64(tx.db.pageSize) 77} 78 79// Writable returns whether the transaction can perform write operations. 80func (tx *Tx) Writable() bool { 81 return tx.writable 82} 83 84// Cursor creates a cursor associated with the root bucket. 85// All items in the cursor will return a nil value because all root bucket keys point to buckets. 86// The cursor is only valid as long as the transaction is open. 87// Do not use a cursor after the transaction is closed. 88func (tx *Tx) Cursor() *Cursor { 89 return tx.root.Cursor() 90} 91 92// Stats retrieves a copy of the current transaction statistics. 93func (tx *Tx) Stats() TxStats { 94 return tx.stats 95} 96 97// Bucket retrieves a bucket by name. 98// Returns nil if the bucket does not exist. 99// The bucket instance is only valid for the lifetime of the transaction. 100func (tx *Tx) Bucket(name []byte) *Bucket { 101 return tx.root.Bucket(name) 102} 103 104// CreateBucket creates a new bucket. 105// Returns an error if the bucket already exists, if the bucket name is blank, or if the bucket name is too long. 106// The bucket instance is only valid for the lifetime of the transaction. 107func (tx *Tx) CreateBucket(name []byte) (*Bucket, error) { 108 return tx.root.CreateBucket(name) 109} 110 111// CreateBucketIfNotExists creates a new bucket if it doesn't already exist. 112// Returns an error if the bucket name is blank, or if the bucket name is too long. 113// The bucket instance is only valid for the lifetime of the transaction. 114func (tx *Tx) CreateBucketIfNotExists(name []byte) (*Bucket, error) { 115 return tx.root.CreateBucketIfNotExists(name) 116} 117 118// DeleteBucket deletes a bucket. 119// Returns an error if the bucket cannot be found or if the key represents a non-bucket value. 120func (tx *Tx) DeleteBucket(name []byte) error { 121 return tx.root.DeleteBucket(name) 122} 123 124// ForEach executes a function for each bucket in the root. 125// If the provided function returns an error then the iteration is stopped and 126// the error is returned to the caller. 127func (tx *Tx) ForEach(fn func(name []byte, b *Bucket) error) error { 128 return tx.root.ForEach(func(k, v []byte) error { 129 return fn(k, tx.root.Bucket(k)) 130 }) 131} 132 133// OnCommit adds a handler function to be executed after the transaction successfully commits. 134func (tx *Tx) OnCommit(fn func()) { 135 tx.commitHandlers = append(tx.commitHandlers, fn) 136} 137 138// Commit writes all changes to disk and updates the meta page. 139// Returns an error if a disk write error occurs, or if Commit is 140// called on a read-only transaction. 141func (tx *Tx) Commit() error { 142 _assert(!tx.managed, "managed tx commit not allowed") 143 if tx.db == nil { 144 return ErrTxClosed 145 } else if !tx.writable { 146 return ErrTxNotWritable 147 } 148 149 // TODO(benbjohnson): Use vectorized I/O to write out dirty pages. 150 151 // Rebalance nodes which have had deletions. 152 var startTime = time.Now() 153 tx.root.rebalance() 154 if tx.stats.Rebalance > 0 { 155 tx.stats.RebalanceTime += time.Since(startTime) 156 } 157 158 // spill data onto dirty pages. 159 startTime = time.Now() 160 if err := tx.root.spill(); err != nil { 161 tx.rollback() 162 return err 163 } 164 tx.stats.SpillTime += time.Since(startTime) 165 166 // Free the old root bucket. 167 tx.meta.root.root = tx.root.root 168 169 // Free the old freelist because commit writes out a fresh freelist. 170 if tx.meta.freelist != pgidNoFreelist { 171 tx.db.freelist.free(tx.meta.txid, tx.db.page(tx.meta.freelist)) 172 } 173 174 if !tx.db.NoFreelistSync { 175 err := tx.commitFreelist() 176 if err != nil { 177 return err 178 } 179 } else { 180 tx.meta.freelist = pgidNoFreelist 181 } 182 183 // Write dirty pages to disk. 184 startTime = time.Now() 185 if err := tx.write(); err != nil { 186 tx.rollback() 187 return err 188 } 189 190 // If strict mode is enabled then perform a consistency check. 191 // Only the first consistency error is reported in the panic. 192 if tx.db.StrictMode { 193 ch := tx.Check() 194 var errs []string 195 for { 196 err, ok := <-ch 197 if !ok { 198 break 199 } 200 errs = append(errs, err.Error()) 201 } 202 if len(errs) > 0 { 203 panic("check fail: " + strings.Join(errs, "\n")) 204 } 205 } 206 207 // Write meta to disk. 208 if err := tx.writeMeta(); err != nil { 209 tx.rollback() 210 return err 211 } 212 tx.stats.WriteTime += time.Since(startTime) 213 214 // Finalize the transaction. 215 tx.close() 216 217 // Execute commit handlers now that the locks have been removed. 218 for _, fn := range tx.commitHandlers { 219 fn() 220 } 221 222 return nil 223} 224 225func (tx *Tx) commitFreelist() error { 226 // Allocate new pages for the new free list. This will overestimate 227 // the size of the freelist but not underestimate the size (which would be bad). 228 opgid := tx.meta.pgid 229 p, err := tx.allocate((tx.db.freelist.size() / tx.db.pageSize) + 1) 230 if err != nil { 231 tx.rollback() 232 return err 233 } 234 if err := tx.db.freelist.write(p); err != nil { 235 tx.rollback() 236 return err 237 } 238 tx.meta.freelist = p.id 239 // If the high water mark has moved up then attempt to grow the database. 240 if tx.meta.pgid > opgid { 241 if err := tx.db.grow(int(tx.meta.pgid+1) * tx.db.pageSize); err != nil { 242 tx.rollback() 243 return err 244 } 245 } 246 247 return nil 248} 249 250// Rollback closes the transaction and ignores all previous updates. Read-only 251// transactions must be rolled back and not committed. 252func (tx *Tx) Rollback() error { 253 _assert(!tx.managed, "managed tx rollback not allowed") 254 if tx.db == nil { 255 return ErrTxClosed 256 } 257 tx.rollback() 258 return nil 259} 260 261func (tx *Tx) rollback() { 262 if tx.db == nil { 263 return 264 } 265 if tx.writable { 266 tx.db.freelist.rollback(tx.meta.txid) 267 tx.db.freelist.reload(tx.db.page(tx.db.meta().freelist)) 268 } 269 tx.close() 270} 271 272func (tx *Tx) close() { 273 if tx.db == nil { 274 return 275 } 276 if tx.writable { 277 // Grab freelist stats. 278 var freelistFreeN = tx.db.freelist.free_count() 279 var freelistPendingN = tx.db.freelist.pending_count() 280 var freelistAlloc = tx.db.freelist.size() 281 282 // Remove transaction ref & writer lock. 283 tx.db.rwtx = nil 284 tx.db.rwlock.Unlock() 285 286 // Merge statistics. 287 tx.db.statlock.Lock() 288 tx.db.stats.FreePageN = freelistFreeN 289 tx.db.stats.PendingPageN = freelistPendingN 290 tx.db.stats.FreeAlloc = (freelistFreeN + freelistPendingN) * tx.db.pageSize 291 tx.db.stats.FreelistInuse = freelistAlloc 292 tx.db.stats.TxStats.add(&tx.stats) 293 tx.db.statlock.Unlock() 294 } else { 295 tx.db.removeTx(tx) 296 } 297 298 // Clear all references. 299 tx.db = nil 300 tx.meta = nil 301 tx.root = Bucket{tx: tx} 302 tx.pages = nil 303} 304 305// Copy writes the entire database to a writer. 306// This function exists for backwards compatibility. Use WriteTo() instead. 307func (tx *Tx) Copy(w io.Writer) error { 308 _, err := tx.WriteTo(w) 309 return err 310} 311 312// WriteTo writes the entire database to a writer. 313// If err == nil then exactly tx.Size() bytes will be written into the writer. 314func (tx *Tx) WriteTo(w io.Writer) (n int64, err error) { 315 // Attempt to open reader with WriteFlag 316 f, err := os.OpenFile(tx.db.path, os.O_RDONLY|tx.WriteFlag, 0) 317 if err != nil { 318 return 0, err 319 } 320 defer func() { 321 if cerr := f.Close(); err == nil { 322 err = cerr 323 } 324 }() 325 326 // Generate a meta page. We use the same page data for both meta pages. 327 buf := make([]byte, tx.db.pageSize) 328 page := (*page)(unsafe.Pointer(&buf[0])) 329 page.flags = metaPageFlag 330 *page.meta() = *tx.meta 331 332 // Write meta 0. 333 page.id = 0 334 page.meta().checksum = page.meta().sum64() 335 nn, err := w.Write(buf) 336 n += int64(nn) 337 if err != nil { 338 return n, fmt.Errorf("meta 0 copy: %s", err) 339 } 340 341 // Write meta 1 with a lower transaction id. 342 page.id = 1 343 page.meta().txid -= 1 344 page.meta().checksum = page.meta().sum64() 345 nn, err = w.Write(buf) 346 n += int64(nn) 347 if err != nil { 348 return n, fmt.Errorf("meta 1 copy: %s", err) 349 } 350 351 // Move past the meta pages in the file. 352 if _, err := f.Seek(int64(tx.db.pageSize*2), io.SeekStart); err != nil { 353 return n, fmt.Errorf("seek: %s", err) 354 } 355 356 // Copy data pages. 357 wn, err := io.CopyN(w, f, tx.Size()-int64(tx.db.pageSize*2)) 358 n += wn 359 if err != nil { 360 return n, err 361 } 362 363 return n, nil 364} 365 366// CopyFile copies the entire database to file at the given path. 367// A reader transaction is maintained during the copy so it is safe to continue 368// using the database while a copy is in progress. 369func (tx *Tx) CopyFile(path string, mode os.FileMode) error { 370 f, err := os.OpenFile(path, os.O_RDWR|os.O_CREATE|os.O_TRUNC, mode) 371 if err != nil { 372 return err 373 } 374 375 err = tx.Copy(f) 376 if err != nil { 377 _ = f.Close() 378 return err 379 } 380 return f.Close() 381} 382 383// Check performs several consistency checks on the database for this transaction. 384// An error is returned if any inconsistency is found. 385// 386// It can be safely run concurrently on a writable transaction. However, this 387// incurs a high cost for large databases and databases with a lot of subbuckets 388// because of caching. This overhead can be removed if running on a read-only 389// transaction, however, it is not safe to execute other writer transactions at 390// the same time. 391func (tx *Tx) Check() <-chan error { 392 ch := make(chan error) 393 go tx.check(ch) 394 return ch 395} 396 397func (tx *Tx) check(ch chan error) { 398 // Force loading free list if opened in ReadOnly mode. 399 tx.db.loadFreelist() 400 401 // Check if any pages are double freed. 402 freed := make(map[pgid]bool) 403 all := make([]pgid, tx.db.freelist.count()) 404 tx.db.freelist.copyall(all) 405 for _, id := range all { 406 if freed[id] { 407 ch <- fmt.Errorf("page %d: already freed", id) 408 } 409 freed[id] = true 410 } 411 412 // Track every reachable page. 413 reachable := make(map[pgid]*page) 414 reachable[0] = tx.page(0) // meta0 415 reachable[1] = tx.page(1) // meta1 416 if tx.meta.freelist != pgidNoFreelist { 417 for i := uint32(0); i <= tx.page(tx.meta.freelist).overflow; i++ { 418 reachable[tx.meta.freelist+pgid(i)] = tx.page(tx.meta.freelist) 419 } 420 } 421 422 // Recursively check buckets. 423 tx.checkBucket(&tx.root, reachable, freed, ch) 424 425 // Ensure all pages below high water mark are either reachable or freed. 426 for i := pgid(0); i < tx.meta.pgid; i++ { 427 _, isReachable := reachable[i] 428 if !isReachable && !freed[i] { 429 ch <- fmt.Errorf("page %d: unreachable unfreed", int(i)) 430 } 431 } 432 433 // Close the channel to signal completion. 434 close(ch) 435} 436 437func (tx *Tx) checkBucket(b *Bucket, reachable map[pgid]*page, freed map[pgid]bool, ch chan error) { 438 // Ignore inline buckets. 439 if b.root == 0 { 440 return 441 } 442 443 // Check every page used by this bucket. 444 b.tx.forEachPage(b.root, 0, func(p *page, _ int) { 445 if p.id > tx.meta.pgid { 446 ch <- fmt.Errorf("page %d: out of bounds: %d", int(p.id), int(b.tx.meta.pgid)) 447 } 448 449 // Ensure each page is only referenced once. 450 for i := pgid(0); i <= pgid(p.overflow); i++ { 451 var id = p.id + i 452 if _, ok := reachable[id]; ok { 453 ch <- fmt.Errorf("page %d: multiple references", int(id)) 454 } 455 reachable[id] = p 456 } 457 458 // We should only encounter un-freed leaf and branch pages. 459 if freed[p.id] { 460 ch <- fmt.Errorf("page %d: reachable freed", int(p.id)) 461 } else if (p.flags&branchPageFlag) == 0 && (p.flags&leafPageFlag) == 0 { 462 ch <- fmt.Errorf("page %d: invalid type: %s", int(p.id), p.typ()) 463 } 464 }) 465 466 // Check each bucket within this bucket. 467 _ = b.ForEach(func(k, v []byte) error { 468 if child := b.Bucket(k); child != nil { 469 tx.checkBucket(child, reachable, freed, ch) 470 } 471 return nil 472 }) 473} 474 475// allocate returns a contiguous block of memory starting at a given page. 476func (tx *Tx) allocate(count int) (*page, error) { 477 p, err := tx.db.allocate(tx.meta.txid, count) 478 if err != nil { 479 return nil, err 480 } 481 482 // Save to our page cache. 483 tx.pages[p.id] = p 484 485 // Update statistics. 486 tx.stats.PageCount++ 487 tx.stats.PageAlloc += count * tx.db.pageSize 488 489 return p, nil 490} 491 492// write writes any dirty pages to disk. 493func (tx *Tx) write() error { 494 // Sort pages by id. 495 pages := make(pages, 0, len(tx.pages)) 496 for _, p := range tx.pages { 497 pages = append(pages, p) 498 } 499 // Clear out page cache early. 500 tx.pages = make(map[pgid]*page) 501 sort.Sort(pages) 502 503 // Write pages to disk in order. 504 for _, p := range pages { 505 size := (int(p.overflow) + 1) * tx.db.pageSize 506 offset := int64(p.id) * int64(tx.db.pageSize) 507 508 // Write out page in "max allocation" sized chunks. 509 ptr := (*[maxAllocSize]byte)(unsafe.Pointer(p)) 510 for { 511 // Limit our write to our max allocation size. 512 sz := size 513 if sz > maxAllocSize-1 { 514 sz = maxAllocSize - 1 515 } 516 517 // Write chunk to disk. 518 buf := ptr[:sz] 519 if _, err := tx.db.ops.writeAt(buf, offset); err != nil { 520 return err 521 } 522 523 // Update statistics. 524 tx.stats.Write++ 525 526 // Exit inner for loop if we've written all the chunks. 527 size -= sz 528 if size == 0 { 529 break 530 } 531 532 // Otherwise move offset forward and move pointer to next chunk. 533 offset += int64(sz) 534 ptr = (*[maxAllocSize]byte)(unsafe.Pointer(&ptr[sz])) 535 } 536 } 537 538 // Ignore file sync if flag is set on DB. 539 if !tx.db.NoSync || IgnoreNoSync { 540 if err := fdatasync(tx.db); err != nil { 541 return err 542 } 543 } 544 545 // Put small pages back to page pool. 546 for _, p := range pages { 547 // Ignore page sizes over 1 page. 548 // These are allocated using make() instead of the page pool. 549 if int(p.overflow) != 0 { 550 continue 551 } 552 553 buf := (*[maxAllocSize]byte)(unsafe.Pointer(p))[:tx.db.pageSize] 554 555 // See https://go.googlesource.com/go/+/f03c9202c43e0abb130669852082117ca50aa9b1 556 for i := range buf { 557 buf[i] = 0 558 } 559 tx.db.pagePool.Put(buf) 560 } 561 562 return nil 563} 564 565// writeMeta writes the meta to the disk. 566func (tx *Tx) writeMeta() error { 567 // Create a temporary buffer for the meta page. 568 buf := make([]byte, tx.db.pageSize) 569 p := tx.db.pageInBuffer(buf, 0) 570 tx.meta.write(p) 571 572 // Write the meta page to file. 573 if _, err := tx.db.ops.writeAt(buf, int64(p.id)*int64(tx.db.pageSize)); err != nil { 574 return err 575 } 576 if !tx.db.NoSync || IgnoreNoSync { 577 if err := fdatasync(tx.db); err != nil { 578 return err 579 } 580 } 581 582 // Update statistics. 583 tx.stats.Write++ 584 585 return nil 586} 587 588// page returns a reference to the page with a given id. 589// If page has been written to then a temporary buffered page is returned. 590func (tx *Tx) page(id pgid) *page { 591 // Check the dirty pages first. 592 if tx.pages != nil { 593 if p, ok := tx.pages[id]; ok { 594 return p 595 } 596 } 597 598 // Otherwise return directly from the mmap. 599 return tx.db.page(id) 600} 601 602// forEachPage iterates over every page within a given page and executes a function. 603func (tx *Tx) forEachPage(pgid pgid, depth int, fn func(*page, int)) { 604 p := tx.page(pgid) 605 606 // Execute function. 607 fn(p, depth) 608 609 // Recursively loop over children. 610 if (p.flags & branchPageFlag) != 0 { 611 for i := 0; i < int(p.count); i++ { 612 elem := p.branchPageElement(uint16(i)) 613 tx.forEachPage(elem.pgid, depth+1, fn) 614 } 615 } 616} 617 618// Page returns page information for a given page number. 619// This is only safe for concurrent use when used by a writable transaction. 620func (tx *Tx) Page(id int) (*PageInfo, error) { 621 if tx.db == nil { 622 return nil, ErrTxClosed 623 } else if pgid(id) >= tx.meta.pgid { 624 return nil, nil 625 } 626 627 // Build the page info. 628 p := tx.db.page(pgid(id)) 629 info := &PageInfo{ 630 ID: id, 631 Count: int(p.count), 632 OverflowCount: int(p.overflow), 633 } 634 635 // Determine the type (or if it's free). 636 if tx.db.freelist.freed(pgid(id)) { 637 info.Type = "free" 638 } else { 639 info.Type = p.typ() 640 } 641 642 return info, nil 643} 644 645// TxStats represents statistics about the actions performed by the transaction. 646type TxStats struct { 647 // Page statistics. 648 PageCount int // number of page allocations 649 PageAlloc int // total bytes allocated 650 651 // Cursor statistics. 652 CursorCount int // number of cursors created 653 654 // Node statistics 655 NodeCount int // number of node allocations 656 NodeDeref int // number of node dereferences 657 658 // Rebalance statistics. 659 Rebalance int // number of node rebalances 660 RebalanceTime time.Duration // total time spent rebalancing 661 662 // Split/Spill statistics. 663 Split int // number of nodes split 664 Spill int // number of nodes spilled 665 SpillTime time.Duration // total time spent spilling 666 667 // Write statistics. 668 Write int // number of writes performed 669 WriteTime time.Duration // total time spent writing to disk 670} 671 672func (s *TxStats) add(other *TxStats) { 673 s.PageCount += other.PageCount 674 s.PageAlloc += other.PageAlloc 675 s.CursorCount += other.CursorCount 676 s.NodeCount += other.NodeCount 677 s.NodeDeref += other.NodeDeref 678 s.Rebalance += other.Rebalance 679 s.RebalanceTime += other.RebalanceTime 680 s.Split += other.Split 681 s.Spill += other.Spill 682 s.SpillTime += other.SpillTime 683 s.Write += other.Write 684 s.WriteTime += other.WriteTime 685} 686 687// Sub calculates and returns the difference between two sets of transaction stats. 688// This is useful when obtaining stats at two different points and time and 689// you need the performance counters that occurred within that time span. 690func (s *TxStats) Sub(other *TxStats) TxStats { 691 var diff TxStats 692 diff.PageCount = s.PageCount - other.PageCount 693 diff.PageAlloc = s.PageAlloc - other.PageAlloc 694 diff.CursorCount = s.CursorCount - other.CursorCount 695 diff.NodeCount = s.NodeCount - other.NodeCount 696 diff.NodeDeref = s.NodeDeref - other.NodeDeref 697 diff.Rebalance = s.Rebalance - other.Rebalance 698 diff.RebalanceTime = s.RebalanceTime - other.RebalanceTime 699 diff.Split = s.Split - other.Split 700 diff.Spill = s.Spill - other.Spill 701 diff.SpillTime = s.SpillTime - other.SpillTime 702 diff.Write = s.Write - other.Write 703 diff.WriteTime = s.WriteTime - other.WriteTime 704 return diff 705} 706