github.com/sberex/go-sberex@v1.8.2-0.20181113200658-ed96ac38f7d7/swarm/storage/chunker.go (about) 1 // This file is part of the go-sberex library. The go-sberex library is 2 // free software: you can redistribute it and/or modify it under the terms 3 // of the GNU Lesser General Public License as published by the Free 4 // Software Foundation, either version 3 of the License, or (at your option) 5 // any later version. 6 // 7 // The go-sberex library is distributed in the hope that it will be useful, 8 // but WITHOUT ANY WARRANTY; without even the implied warranty of 9 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser 10 // General Public License <http://www.gnu.org/licenses/> for more details. 11 12 package storage 13 14 import ( 15 "encoding/binary" 16 "errors" 17 "fmt" 18 "io" 19 "sync" 20 "time" 21 22 "github.com/Sberex/go-sberex/metrics" 23 ) 24 25 /* 26 The distributed storage implemented in this package requires fix sized chunks of content. 27 28 Chunker is the interface to a component that is responsible for disassembling and assembling larger data. 29 30 TreeChunker implements a Chunker based on a tree structure defined as follows: 31 32 1 each node in the tree including the root and other branching nodes are stored as a chunk. 33 34 2 branching nodes encode data contents that includes the size of the dataslice covered by its entire subtree under the node as well as the hash keys of all its children : 35 data_{i} := size(subtree_{i}) || key_{j} || key_{j+1} .... || key_{j+n-1} 36 37 3 Leaf nodes encode an actual subslice of the input data. 38 39 4 if data size is not more than maximum chunksize, the data is stored in a single chunk 40 key = hash(int64(size) + data) 41 42 5 if data size is more than chunksize*branches^l, but no more than chunksize* 43 branches^(l+1), the data vector is split into slices of chunksize* 44 branches^l length (except the last one). 45 key = hash(int64(size) + key(slice0) + key(slice1) + ...) 46 47 The underlying hash function is configurable 48 */ 49 50 /* 51 Tree chunker is a concrete implementation of data chunking. 52 This chunker works in a simple way, it builds a tree out of the document so that each node either represents a chunk of real data or a chunk of data representing an branching non-leaf node of the tree. In particular each such non-leaf chunk will represent is a concatenation of the hash of its respective children. This scheme simultaneously guarantees data integrity as well as self addressing. Abstract nodes are transparent since their represented size component is strictly greater than their maximum data size, since they encode a subtree. 53 54 If all is well it is possible to implement this by simply composing readers so that no extra allocation or buffering is necessary for the data splitting and joining. This means that in principle there can be direct IO between : memory, file system, network socket (bzz peers storage request is read from the socket). In practice there may be need for several stages of internal buffering. 55 The hashing itself does use extra copies and allocation though, since it does need it. 56 */ 57 58 var ( 59 errAppendOppNotSuported = errors.New("Append operation not supported") 60 errOperationTimedOut = errors.New("operation timed out") 61 ) 62 63 //metrics variables 64 var ( 65 newChunkCounter = metrics.NewRegisteredCounter("storage.chunks.new", nil) 66 ) 67 68 type TreeChunker struct { 69 branches int64 70 hashFunc SwarmHasher 71 // calculated 72 hashSize int64 // self.hashFunc.New().Size() 73 chunkSize int64 // hashSize* branches 74 workerCount int64 // the number of worker routines used 75 workerLock sync.RWMutex // lock for the worker count 76 } 77 78 func NewTreeChunker(params *ChunkerParams) (self *TreeChunker) { 79 self = &TreeChunker{} 80 self.hashFunc = MakeHashFunc(params.Hash) 81 self.branches = params.Branches 82 self.hashSize = int64(self.hashFunc().Size()) 83 self.chunkSize = self.hashSize * self.branches 84 self.workerCount = 0 85 86 return 87 } 88 89 // func (self *TreeChunker) KeySize() int64 { 90 // return self.hashSize 91 // } 92 93 // String() for pretty printing 94 func (self *Chunk) String() string { 95 return fmt.Sprintf("Key: %v TreeSize: %v Chunksize: %v", self.Key.Log(), self.Size, len(self.SData)) 96 } 97 98 type hashJob struct { 99 key Key 100 chunk []byte 101 size int64 102 parentWg *sync.WaitGroup 103 } 104 105 func (self *TreeChunker) incrementWorkerCount() { 106 self.workerLock.Lock() 107 defer self.workerLock.Unlock() 108 self.workerCount += 1 109 } 110 111 func (self *TreeChunker) getWorkerCount() int64 { 112 self.workerLock.RLock() 113 defer self.workerLock.RUnlock() 114 return self.workerCount 115 } 116 117 func (self *TreeChunker) decrementWorkerCount() { 118 self.workerLock.Lock() 119 defer self.workerLock.Unlock() 120 self.workerCount -= 1 121 } 122 123 func (self *TreeChunker) Split(data io.Reader, size int64, chunkC chan *Chunk, swg, wwg *sync.WaitGroup) (Key, error) { 124 if self.chunkSize <= 0 { 125 panic("chunker must be initialised") 126 } 127 128 jobC := make(chan *hashJob, 2*ChunkProcessors) 129 wg := &sync.WaitGroup{} 130 errC := make(chan error) 131 quitC := make(chan bool) 132 133 // wwg = workers waitgroup keeps track of hashworkers spawned by this split call 134 if wwg != nil { 135 wwg.Add(1) 136 } 137 138 self.incrementWorkerCount() 139 go self.hashWorker(jobC, chunkC, errC, quitC, swg, wwg) 140 141 depth := 0 142 treeSize := self.chunkSize 143 144 // takes lowest depth such that chunksize*HashCount^(depth+1) > size 145 // power series, will find the order of magnitude of the data size in base hashCount or numbers of levels of branching in the resulting tree. 146 for ; treeSize < size; treeSize *= self.branches { 147 depth++ 148 } 149 150 key := make([]byte, self.hashFunc().Size()) 151 // this waitgroup member is released after the root hash is calculated 152 wg.Add(1) 153 //launch actual recursive function passing the waitgroups 154 go self.split(depth, treeSize/self.branches, key, data, size, jobC, chunkC, errC, quitC, wg, swg, wwg) 155 156 // closes internal error channel if all subprocesses in the workgroup finished 157 go func() { 158 // waiting for all threads to finish 159 wg.Wait() 160 // if storage waitgroup is non-nil, we wait for storage to finish too 161 if swg != nil { 162 swg.Wait() 163 } 164 close(errC) 165 }() 166 167 defer close(quitC) 168 select { 169 case err := <-errC: 170 if err != nil { 171 return nil, err 172 } 173 case <-time.NewTimer(splitTimeout).C: 174 return nil, errOperationTimedOut 175 } 176 177 return key, nil 178 } 179 180 func (self *TreeChunker) split(depth int, treeSize int64, key Key, data io.Reader, size int64, jobC chan *hashJob, chunkC chan *Chunk, errC chan error, quitC chan bool, parentWg, swg, wwg *sync.WaitGroup) { 181 182 // 183 184 for depth > 0 && size < treeSize { 185 treeSize /= self.branches 186 depth-- 187 } 188 189 if depth == 0 { 190 // leaf nodes -> content chunks 191 chunkData := make([]byte, size+8) 192 binary.LittleEndian.PutUint64(chunkData[0:8], uint64(size)) 193 var readBytes int64 194 for readBytes < size { 195 n, err := data.Read(chunkData[8+readBytes:]) 196 readBytes += int64(n) 197 if err != nil && !(err == io.EOF && readBytes == size) { 198 errC <- err 199 return 200 } 201 } 202 select { 203 case jobC <- &hashJob{key, chunkData, size, parentWg}: 204 case <-quitC: 205 } 206 return 207 } 208 // dept > 0 209 // intermediate chunk containing child nodes hashes 210 branchCnt := (size + treeSize - 1) / treeSize 211 212 var chunk = make([]byte, branchCnt*self.hashSize+8) 213 var pos, i int64 214 215 binary.LittleEndian.PutUint64(chunk[0:8], uint64(size)) 216 217 childrenWg := &sync.WaitGroup{} 218 var secSize int64 219 for i < branchCnt { 220 // the last item can have shorter data 221 if size-pos < treeSize { 222 secSize = size - pos 223 } else { 224 secSize = treeSize 225 } 226 // the hash of that data 227 subTreeKey := chunk[8+i*self.hashSize : 8+(i+1)*self.hashSize] 228 229 childrenWg.Add(1) 230 self.split(depth-1, treeSize/self.branches, subTreeKey, data, secSize, jobC, chunkC, errC, quitC, childrenWg, swg, wwg) 231 232 i++ 233 pos += treeSize 234 } 235 // wait for all the children to complete calculating their hashes and copying them onto sections of the chunk 236 // parentWg.Add(1) 237 // go func() { 238 childrenWg.Wait() 239 240 worker := self.getWorkerCount() 241 if int64(len(jobC)) > worker && worker < ChunkProcessors { 242 if wwg != nil { 243 wwg.Add(1) 244 } 245 self.incrementWorkerCount() 246 go self.hashWorker(jobC, chunkC, errC, quitC, swg, wwg) 247 248 } 249 select { 250 case jobC <- &hashJob{key, chunk, size, parentWg}: 251 case <-quitC: 252 } 253 } 254 255 func (self *TreeChunker) hashWorker(jobC chan *hashJob, chunkC chan *Chunk, errC chan error, quitC chan bool, swg, wwg *sync.WaitGroup) { 256 defer self.decrementWorkerCount() 257 258 hasher := self.hashFunc() 259 if wwg != nil { 260 defer wwg.Done() 261 } 262 for { 263 select { 264 265 case job, ok := <-jobC: 266 if !ok { 267 return 268 } 269 // now we got the hashes in the chunk, then hash the chunks 270 self.hashChunk(hasher, job, chunkC, swg) 271 case <-quitC: 272 return 273 } 274 } 275 } 276 277 // The treeChunkers own Hash hashes together 278 // - the size (of the subtree encoded in the Chunk) 279 // - the Chunk, ie. the contents read from the input reader 280 func (self *TreeChunker) hashChunk(hasher SwarmHash, job *hashJob, chunkC chan *Chunk, swg *sync.WaitGroup) { 281 hasher.ResetWithLength(job.chunk[:8]) // 8 bytes of length 282 hasher.Write(job.chunk[8:]) // minus 8 []byte length 283 h := hasher.Sum(nil) 284 285 newChunk := &Chunk{ 286 Key: h, 287 SData: job.chunk, 288 Size: job.size, 289 wg: swg, 290 } 291 292 // report hash of this chunk one level up (keys corresponds to the proper subslice of the parent chunk) 293 copy(job.key, h) 294 // send off new chunk to storage 295 if chunkC != nil { 296 if swg != nil { 297 swg.Add(1) 298 } 299 } 300 job.parentWg.Done() 301 302 if chunkC != nil { 303 //NOTE: this increases the chunk count even if the local node already has this chunk; 304 //on file upload the node will increase this counter even if the same file has already been uploaded 305 //So it should be evaluated whether it is worth keeping this counter 306 //and/or actually better track when the chunk is Put to the local database 307 //(which may question the need for disambiguation when a completely new chunk has been created 308 //and/or a chunk is being put to the local DB; for chunk tracking it may be worth distinguishing 309 newChunkCounter.Inc(1) 310 chunkC <- newChunk 311 } 312 } 313 314 func (self *TreeChunker) Append(key Key, data io.Reader, chunkC chan *Chunk, swg, wwg *sync.WaitGroup) (Key, error) { 315 return nil, errAppendOppNotSuported 316 } 317 318 // LazyChunkReader implements LazySectionReader 319 type LazyChunkReader struct { 320 key Key // root key 321 chunkC chan *Chunk // chunk channel to send retrieve requests on 322 chunk *Chunk // size of the entire subtree 323 off int64 // offset 324 chunkSize int64 // inherit from chunker 325 branches int64 // inherit from chunker 326 hashSize int64 // inherit from chunker 327 } 328 329 // implements the Joiner interface 330 func (self *TreeChunker) Join(key Key, chunkC chan *Chunk) LazySectionReader { 331 return &LazyChunkReader{ 332 key: key, 333 chunkC: chunkC, 334 chunkSize: self.chunkSize, 335 branches: self.branches, 336 hashSize: self.hashSize, 337 } 338 } 339 340 // Size is meant to be called on the LazySectionReader 341 func (self *LazyChunkReader) Size(quitC chan bool) (n int64, err error) { 342 if self.chunk != nil { 343 return self.chunk.Size, nil 344 } 345 chunk := retrieve(self.key, self.chunkC, quitC) 346 if chunk == nil { 347 select { 348 case <-quitC: 349 return 0, errors.New("aborted") 350 default: 351 return 0, fmt.Errorf("root chunk not found for %v", self.key.Hex()) 352 } 353 } 354 self.chunk = chunk 355 return chunk.Size, nil 356 } 357 358 // read at can be called numerous times 359 // concurrent reads are allowed 360 // Size() needs to be called synchronously on the LazyChunkReader first 361 func (self *LazyChunkReader) ReadAt(b []byte, off int64) (read int, err error) { 362 // this is correct, a swarm doc cannot be zero length, so no EOF is expected 363 if len(b) == 0 { 364 return 0, nil 365 } 366 quitC := make(chan bool) 367 size, err := self.Size(quitC) 368 if err != nil { 369 return 0, err 370 } 371 372 errC := make(chan error) 373 374 // } 375 var treeSize int64 376 var depth int 377 // calculate depth and max treeSize 378 treeSize = self.chunkSize 379 for ; treeSize < size; treeSize *= self.branches { 380 depth++ 381 } 382 wg := sync.WaitGroup{} 383 wg.Add(1) 384 go self.join(b, off, off+int64(len(b)), depth, treeSize/self.branches, self.chunk, &wg, errC, quitC) 385 go func() { 386 wg.Wait() 387 close(errC) 388 }() 389 390 err = <-errC 391 if err != nil { 392 close(quitC) 393 394 return 0, err 395 } 396 if off+int64(len(b)) >= size { 397 return len(b), io.EOF 398 } 399 return len(b), nil 400 } 401 402 func (self *LazyChunkReader) join(b []byte, off int64, eoff int64, depth int, treeSize int64, chunk *Chunk, parentWg *sync.WaitGroup, errC chan error, quitC chan bool) { 403 defer parentWg.Done() 404 // return NewDPA(&LocalStore{}) 405 406 // chunk.Size = int64(binary.LittleEndian.Uint64(chunk.SData[0:8])) 407 408 // find appropriate block level 409 for chunk.Size < treeSize && depth > 0 { 410 treeSize /= self.branches 411 depth-- 412 } 413 414 // leaf chunk found 415 if depth == 0 { 416 extra := 8 + eoff - int64(len(chunk.SData)) 417 if extra > 0 { 418 eoff -= extra 419 } 420 copy(b, chunk.SData[8+off:8+eoff]) 421 return // simply give back the chunks reader for content chunks 422 } 423 424 // subtree 425 start := off / treeSize 426 end := (eoff + treeSize - 1) / treeSize 427 428 wg := &sync.WaitGroup{} 429 defer wg.Wait() 430 431 for i := start; i < end; i++ { 432 soff := i * treeSize 433 roff := soff 434 seoff := soff + treeSize 435 436 if soff < off { 437 soff = off 438 } 439 if seoff > eoff { 440 seoff = eoff 441 } 442 if depth > 1 { 443 wg.Wait() 444 } 445 wg.Add(1) 446 go func(j int64) { 447 childKey := chunk.SData[8+j*self.hashSize : 8+(j+1)*self.hashSize] 448 chunk := retrieve(childKey, self.chunkC, quitC) 449 if chunk == nil { 450 select { 451 case errC <- fmt.Errorf("chunk %v-%v not found", off, off+treeSize): 452 case <-quitC: 453 } 454 return 455 } 456 if soff < off { 457 soff = off 458 } 459 self.join(b[soff-off:seoff-off], soff-roff, seoff-roff, depth-1, treeSize/self.branches, chunk, wg, errC, quitC) 460 }(i) 461 } //for 462 } 463 464 // the helper method submits chunks for a key to a oueue (DPA) and 465 // block until they time out or arrive 466 // abort if quitC is readable 467 func retrieve(key Key, chunkC chan *Chunk, quitC chan bool) *Chunk { 468 chunk := &Chunk{ 469 Key: key, 470 C: make(chan bool), // close channel to signal data delivery 471 } 472 // submit chunk for retrieval 473 select { 474 case chunkC <- chunk: // submit retrieval request, someone should be listening on the other side (or we will time out globally) 475 case <-quitC: 476 return nil 477 } 478 // waiting for the chunk retrieval 479 select { // chunk.Size = int64(binary.LittleEndian.Uint64(chunk.SData[0:8])) 480 481 case <-quitC: 482 // this is how we control process leakage (quitC is closed once join is finished (after timeout)) 483 return nil 484 case <-chunk.C: // bells are ringing, data have been delivered 485 } 486 if len(chunk.SData) == 0 { 487 return nil // chunk.Size = int64(binary.LittleEndian.Uint64(chunk.SData[0:8])) 488 489 } 490 return chunk 491 } 492 493 // Read keeps a cursor so cannot be called simulateously, see ReadAt 494 func (self *LazyChunkReader) Read(b []byte) (read int, err error) { 495 read, err = self.ReadAt(b, self.off) 496 497 self.off += int64(read) 498 return 499 } 500 501 // completely analogous to standard SectionReader implementation 502 var errWhence = errors.New("Seek: invalid whence") 503 var errOffset = errors.New("Seek: invalid offset") 504 505 func (s *LazyChunkReader) Seek(offset int64, whence int) (int64, error) { 506 switch whence { 507 default: 508 return 0, errWhence 509 case 0: 510 offset += 0 511 case 1: 512 offset += s.off 513 case 2: 514 if s.chunk == nil { //seek from the end requires rootchunk for size. call Size first 515 _, err := s.Size(nil) 516 if err != nil { 517 return 0, fmt.Errorf("can't get size: %v", err) 518 } 519 } 520 offset += s.chunk.Size 521 } 522 523 if offset < 0 { 524 return 0, errOffset 525 } 526 s.off = offset 527 return offset, nil 528 }