Scott Baker | ed4efab | 2020-01-13 19:12:25 -0800 | [diff] [blame] | 1 | // Copyright 2019+ Klaus Post. All rights reserved. |
| 2 | // License information can be found in the LICENSE file. |
| 3 | // Based on work by Yann Collet, released under BSD License. |
| 4 | |
| 5 | package zstd |
| 6 | |
| 7 | import ( |
| 8 | "bytes" |
| 9 | "encoding/hex" |
| 10 | "errors" |
| 11 | "hash" |
| 12 | "io" |
| 13 | "sync" |
| 14 | |
| 15 | "github.com/klauspost/compress/zstd/internal/xxhash" |
| 16 | ) |
| 17 | |
| 18 | type frameDec struct { |
| 19 | o decoderOptions |
| 20 | crc hash.Hash64 |
| 21 | frameDone sync.WaitGroup |
| 22 | offset int64 |
| 23 | |
| 24 | WindowSize uint64 |
| 25 | DictionaryID uint32 |
| 26 | FrameContentSize uint64 |
| 27 | HasCheckSum bool |
| 28 | SingleSegment bool |
| 29 | |
| 30 | // maxWindowSize is the maximum windows size to support. |
| 31 | // should never be bigger than max-int. |
| 32 | maxWindowSize uint64 |
| 33 | |
| 34 | // In order queue of blocks being decoded. |
| 35 | decoding chan *blockDec |
| 36 | |
| 37 | // Frame history passed between blocks |
| 38 | history history |
| 39 | |
| 40 | rawInput byteBuffer |
| 41 | |
| 42 | // Byte buffer that can be reused for small input blocks. |
| 43 | bBuf byteBuf |
| 44 | |
| 45 | // asyncRunning indicates whether the async routine processes input on 'decoding'. |
| 46 | asyncRunning bool |
| 47 | asyncRunningMu sync.Mutex |
| 48 | } |
| 49 | |
| 50 | const ( |
| 51 | // The minimum Window_Size is 1 KB. |
| 52 | MinWindowSize = 1 << 10 |
| 53 | MaxWindowSize = 1 << 30 |
| 54 | ) |
| 55 | |
| 56 | var ( |
| 57 | frameMagic = []byte{0x28, 0xb5, 0x2f, 0xfd} |
| 58 | skippableFrameMagic = []byte{0x2a, 0x4d, 0x18} |
| 59 | ) |
| 60 | |
| 61 | func newFrameDec(o decoderOptions) *frameDec { |
| 62 | d := frameDec{ |
| 63 | o: o, |
| 64 | maxWindowSize: MaxWindowSize, |
| 65 | } |
| 66 | if d.maxWindowSize > o.maxDecodedSize { |
| 67 | d.maxWindowSize = o.maxDecodedSize |
| 68 | } |
| 69 | return &d |
| 70 | } |
| 71 | |
| 72 | // reset will read the frame header and prepare for block decoding. |
| 73 | // If nothing can be read from the input, io.EOF will be returned. |
| 74 | // Any other error indicated that the stream contained data, but |
| 75 | // there was a problem. |
| 76 | func (d *frameDec) reset(br byteBuffer) error { |
| 77 | d.HasCheckSum = false |
| 78 | d.WindowSize = 0 |
| 79 | var b []byte |
| 80 | for { |
| 81 | b = br.readSmall(4) |
| 82 | if b == nil { |
| 83 | return io.EOF |
| 84 | } |
| 85 | if !bytes.Equal(b[1:4], skippableFrameMagic) || b[0]&0xf0 != 0x50 { |
| 86 | if debug { |
| 87 | println("Not skippable", hex.EncodeToString(b), hex.EncodeToString(skippableFrameMagic)) |
| 88 | } |
| 89 | // Break if not skippable frame. |
| 90 | break |
| 91 | } |
| 92 | // Read size to skip |
| 93 | b = br.readSmall(4) |
| 94 | if b == nil { |
| 95 | println("Reading Frame Size EOF") |
| 96 | return io.ErrUnexpectedEOF |
| 97 | } |
| 98 | n := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24) |
| 99 | println("Skipping frame with", n, "bytes.") |
| 100 | err := br.skipN(int(n)) |
| 101 | if err != nil { |
| 102 | if debug { |
| 103 | println("Reading discarded frame", err) |
| 104 | } |
| 105 | return err |
| 106 | } |
| 107 | } |
| 108 | if !bytes.Equal(b, frameMagic) { |
| 109 | println("Got magic numbers: ", b, "want:", frameMagic) |
| 110 | return ErrMagicMismatch |
| 111 | } |
| 112 | |
| 113 | // Read Frame_Header_Descriptor |
| 114 | fhd, err := br.readByte() |
| 115 | if err != nil { |
| 116 | println("Reading Frame_Header_Descriptor", err) |
| 117 | return err |
| 118 | } |
| 119 | d.SingleSegment = fhd&(1<<5) != 0 |
| 120 | |
| 121 | if fhd&(1<<3) != 0 { |
| 122 | return errors.New("Reserved bit set on frame header") |
| 123 | } |
| 124 | |
| 125 | // Read Window_Descriptor |
| 126 | // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#window_descriptor |
| 127 | d.WindowSize = 0 |
| 128 | if !d.SingleSegment { |
| 129 | wd, err := br.readByte() |
| 130 | if err != nil { |
| 131 | println("Reading Window_Descriptor", err) |
| 132 | return err |
| 133 | } |
| 134 | printf("raw: %x, mantissa: %d, exponent: %d\n", wd, wd&7, wd>>3) |
| 135 | windowLog := 10 + (wd >> 3) |
| 136 | windowBase := uint64(1) << windowLog |
| 137 | windowAdd := (windowBase / 8) * uint64(wd&0x7) |
| 138 | d.WindowSize = windowBase + windowAdd |
| 139 | } |
| 140 | |
| 141 | // Read Dictionary_ID |
| 142 | // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary_id |
| 143 | d.DictionaryID = 0 |
| 144 | if size := fhd & 3; size != 0 { |
| 145 | if size == 3 { |
| 146 | size = 4 |
| 147 | } |
| 148 | b = br.readSmall(int(size)) |
| 149 | if b == nil { |
| 150 | if debug { |
| 151 | println("Reading Dictionary_ID", io.ErrUnexpectedEOF) |
| 152 | } |
| 153 | return io.ErrUnexpectedEOF |
| 154 | } |
| 155 | switch size { |
| 156 | case 1: |
| 157 | d.DictionaryID = uint32(b[0]) |
| 158 | case 2: |
| 159 | d.DictionaryID = uint32(b[0]) | (uint32(b[1]) << 8) |
| 160 | case 4: |
| 161 | d.DictionaryID = uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24) |
| 162 | } |
| 163 | if debug { |
| 164 | println("Dict size", size, "ID:", d.DictionaryID) |
| 165 | } |
| 166 | if d.DictionaryID != 0 { |
| 167 | return ErrUnknownDictionary |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | // Read Frame_Content_Size |
| 172 | // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frame_content_size |
| 173 | var fcsSize int |
| 174 | v := fhd >> 6 |
| 175 | switch v { |
| 176 | case 0: |
| 177 | if d.SingleSegment { |
| 178 | fcsSize = 1 |
| 179 | } |
| 180 | default: |
| 181 | fcsSize = 1 << v |
| 182 | } |
| 183 | d.FrameContentSize = 0 |
| 184 | if fcsSize > 0 { |
| 185 | b := br.readSmall(fcsSize) |
| 186 | if b == nil { |
| 187 | println("Reading Frame content", io.ErrUnexpectedEOF) |
| 188 | return io.ErrUnexpectedEOF |
| 189 | } |
| 190 | switch fcsSize { |
| 191 | case 1: |
| 192 | d.FrameContentSize = uint64(b[0]) |
| 193 | case 2: |
| 194 | // When FCS_Field_Size is 2, the offset of 256 is added. |
| 195 | d.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) + 256 |
| 196 | case 4: |
| 197 | d.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) | (uint64(b[2]) << 16) | (uint64(b[3]) << 24) |
| 198 | case 8: |
| 199 | d1 := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24) |
| 200 | d2 := uint32(b[4]) | (uint32(b[5]) << 8) | (uint32(b[6]) << 16) | (uint32(b[7]) << 24) |
| 201 | d.FrameContentSize = uint64(d1) | (uint64(d2) << 32) |
| 202 | } |
| 203 | if debug { |
| 204 | println("field size bits:", v, "fcsSize:", fcsSize, "FrameContentSize:", d.FrameContentSize, hex.EncodeToString(b[:fcsSize]), "singleseg:", d.SingleSegment, "window:", d.WindowSize) |
| 205 | } |
| 206 | } |
| 207 | // Move this to shared. |
| 208 | d.HasCheckSum = fhd&(1<<2) != 0 |
| 209 | if d.HasCheckSum { |
| 210 | if d.crc == nil { |
| 211 | d.crc = xxhash.New() |
| 212 | } |
| 213 | d.crc.Reset() |
| 214 | } |
| 215 | |
| 216 | if d.WindowSize == 0 && d.SingleSegment { |
| 217 | // We may not need window in this case. |
| 218 | d.WindowSize = d.FrameContentSize |
| 219 | if d.WindowSize < MinWindowSize { |
| 220 | d.WindowSize = MinWindowSize |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | if d.WindowSize > d.maxWindowSize { |
| 225 | printf("window size %d > max %d\n", d.WindowSize, d.maxWindowSize) |
| 226 | return ErrWindowSizeExceeded |
| 227 | } |
| 228 | // The minimum Window_Size is 1 KB. |
| 229 | if d.WindowSize < MinWindowSize { |
| 230 | println("got window size: ", d.WindowSize) |
| 231 | return ErrWindowSizeTooSmall |
| 232 | } |
| 233 | d.history.windowSize = int(d.WindowSize) |
| 234 | d.history.maxSize = d.history.windowSize + maxBlockSize |
| 235 | // history contains input - maybe we do something |
| 236 | d.rawInput = br |
| 237 | return nil |
| 238 | } |
| 239 | |
| 240 | // next will start decoding the next block from stream. |
| 241 | func (d *frameDec) next(block *blockDec) error { |
| 242 | if debug { |
| 243 | printf("decoding new block %p:%p", block, block.data) |
| 244 | } |
| 245 | err := block.reset(d.rawInput, d.WindowSize) |
| 246 | if err != nil { |
| 247 | println("block error:", err) |
| 248 | // Signal the frame decoder we have a problem. |
| 249 | d.sendErr(block, err) |
| 250 | return err |
| 251 | } |
| 252 | block.input <- struct{}{} |
| 253 | if debug { |
| 254 | println("next block:", block) |
| 255 | } |
| 256 | d.asyncRunningMu.Lock() |
| 257 | defer d.asyncRunningMu.Unlock() |
| 258 | if !d.asyncRunning { |
| 259 | return nil |
| 260 | } |
| 261 | if block.Last { |
| 262 | // We indicate the frame is done by sending io.EOF |
| 263 | d.decoding <- block |
| 264 | return io.EOF |
| 265 | } |
| 266 | d.decoding <- block |
| 267 | return nil |
| 268 | } |
| 269 | |
| 270 | // sendEOF will queue an error block on the frame. |
| 271 | // This will cause the frame decoder to return when it encounters the block. |
| 272 | // Returns true if the decoder was added. |
| 273 | func (d *frameDec) sendErr(block *blockDec, err error) bool { |
| 274 | d.asyncRunningMu.Lock() |
| 275 | defer d.asyncRunningMu.Unlock() |
| 276 | if !d.asyncRunning { |
| 277 | return false |
| 278 | } |
| 279 | |
| 280 | println("sending error", err.Error()) |
| 281 | block.sendErr(err) |
| 282 | d.decoding <- block |
| 283 | return true |
| 284 | } |
| 285 | |
| 286 | // checkCRC will check the checksum if the frame has one. |
| 287 | // Will return ErrCRCMismatch if crc check failed, otherwise nil. |
| 288 | func (d *frameDec) checkCRC() error { |
| 289 | if !d.HasCheckSum { |
| 290 | return nil |
| 291 | } |
| 292 | var tmp [4]byte |
| 293 | got := d.crc.Sum64() |
| 294 | // Flip to match file order. |
| 295 | tmp[0] = byte(got >> 0) |
| 296 | tmp[1] = byte(got >> 8) |
| 297 | tmp[2] = byte(got >> 16) |
| 298 | tmp[3] = byte(got >> 24) |
| 299 | |
| 300 | // We can overwrite upper tmp now |
| 301 | want := d.rawInput.readSmall(4) |
| 302 | if want == nil { |
| 303 | println("CRC missing?") |
| 304 | return io.ErrUnexpectedEOF |
| 305 | } |
| 306 | |
| 307 | if !bytes.Equal(tmp[:], want) { |
| 308 | if debug { |
| 309 | println("CRC Check Failed:", tmp[:], "!=", want) |
| 310 | } |
| 311 | return ErrCRCMismatch |
| 312 | } |
| 313 | if debug { |
| 314 | println("CRC ok", tmp[:]) |
| 315 | } |
| 316 | return nil |
| 317 | } |
| 318 | |
| 319 | func (d *frameDec) initAsync() { |
| 320 | if !d.o.lowMem && !d.SingleSegment { |
| 321 | // set max extra size history to 20MB. |
| 322 | d.history.maxSize = d.history.windowSize + maxBlockSize*10 |
| 323 | } |
| 324 | // re-alloc if more than one extra block size. |
| 325 | if d.o.lowMem && cap(d.history.b) > d.history.maxSize+maxBlockSize { |
| 326 | d.history.b = make([]byte, 0, d.history.maxSize) |
| 327 | } |
| 328 | if cap(d.history.b) < d.history.maxSize { |
| 329 | d.history.b = make([]byte, 0, d.history.maxSize) |
| 330 | } |
| 331 | if cap(d.decoding) < d.o.concurrent { |
| 332 | d.decoding = make(chan *blockDec, d.o.concurrent) |
| 333 | } |
| 334 | if debug { |
| 335 | h := d.history |
| 336 | printf("history init. len: %d, cap: %d", len(h.b), cap(h.b)) |
| 337 | } |
| 338 | d.asyncRunningMu.Lock() |
| 339 | d.asyncRunning = true |
| 340 | d.asyncRunningMu.Unlock() |
| 341 | } |
| 342 | |
| 343 | // startDecoder will start decoding blocks and write them to the writer. |
| 344 | // The decoder will stop as soon as an error occurs or at end of frame. |
| 345 | // When the frame has finished decoding the *bufio.Reader |
| 346 | // containing the remaining input will be sent on frameDec.frameDone. |
| 347 | func (d *frameDec) startDecoder(output chan decodeOutput) { |
| 348 | // TODO: Init to dictionary |
| 349 | d.history.reset() |
| 350 | written := int64(0) |
| 351 | |
| 352 | defer func() { |
| 353 | d.asyncRunningMu.Lock() |
| 354 | d.asyncRunning = false |
| 355 | d.asyncRunningMu.Unlock() |
| 356 | |
| 357 | // Drain the currently decoding. |
| 358 | d.history.error = true |
| 359 | flushdone: |
| 360 | for { |
| 361 | select { |
| 362 | case b := <-d.decoding: |
| 363 | b.history <- &d.history |
| 364 | output <- <-b.result |
| 365 | default: |
| 366 | break flushdone |
| 367 | } |
| 368 | } |
| 369 | println("frame decoder done, signalling done") |
| 370 | d.frameDone.Done() |
| 371 | }() |
| 372 | // Get decoder for first block. |
| 373 | block := <-d.decoding |
| 374 | block.history <- &d.history |
| 375 | for { |
| 376 | var next *blockDec |
| 377 | // Get result |
| 378 | r := <-block.result |
| 379 | if r.err != nil { |
| 380 | println("Result contained error", r.err) |
| 381 | output <- r |
| 382 | return |
| 383 | } |
| 384 | if debug { |
| 385 | println("got result, from ", d.offset, "to", d.offset+int64(len(r.b))) |
| 386 | d.offset += int64(len(r.b)) |
| 387 | } |
| 388 | if !block.Last { |
| 389 | // Send history to next block |
| 390 | select { |
| 391 | case next = <-d.decoding: |
| 392 | if debug { |
| 393 | println("Sending ", len(d.history.b), "bytes as history") |
| 394 | } |
| 395 | next.history <- &d.history |
| 396 | default: |
| 397 | // Wait until we have sent the block, so |
| 398 | // other decoders can potentially get the decoder. |
| 399 | next = nil |
| 400 | } |
| 401 | } |
| 402 | |
| 403 | // Add checksum, async to decoding. |
| 404 | if d.HasCheckSum { |
| 405 | n, err := d.crc.Write(r.b) |
| 406 | if err != nil { |
| 407 | r.err = err |
| 408 | if n != len(r.b) { |
| 409 | r.err = io.ErrShortWrite |
| 410 | } |
| 411 | output <- r |
| 412 | return |
| 413 | } |
| 414 | } |
| 415 | written += int64(len(r.b)) |
| 416 | if d.SingleSegment && uint64(written) > d.FrameContentSize { |
| 417 | println("runDecoder: single segment and", uint64(written), ">", d.FrameContentSize) |
| 418 | r.err = ErrFrameSizeExceeded |
| 419 | output <- r |
| 420 | return |
| 421 | } |
| 422 | if block.Last { |
| 423 | r.err = d.checkCRC() |
| 424 | output <- r |
| 425 | return |
| 426 | } |
| 427 | output <- r |
| 428 | if next == nil { |
| 429 | // There was no decoder available, we wait for one now that we have sent to the writer. |
| 430 | if debug { |
| 431 | println("Sending ", len(d.history.b), " bytes as history") |
| 432 | } |
| 433 | next = <-d.decoding |
| 434 | next.history <- &d.history |
| 435 | } |
| 436 | block = next |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | // runDecoder will create a sync decoder that will decode a block of data. |
| 441 | func (d *frameDec) runDecoder(dst []byte, dec *blockDec) ([]byte, error) { |
| 442 | // TODO: Init to dictionary |
| 443 | d.history.reset() |
| 444 | saved := d.history.b |
| 445 | |
| 446 | // We use the history for output to avoid copying it. |
| 447 | d.history.b = dst |
| 448 | // Store input length, so we only check new data. |
| 449 | crcStart := len(dst) |
| 450 | var err error |
| 451 | for { |
| 452 | err = dec.reset(d.rawInput, d.WindowSize) |
| 453 | if err != nil { |
| 454 | break |
| 455 | } |
| 456 | if debug { |
| 457 | println("next block:", dec) |
| 458 | } |
| 459 | err = dec.decodeBuf(&d.history) |
| 460 | if err != nil || dec.Last { |
| 461 | break |
| 462 | } |
| 463 | if uint64(len(d.history.b)) > d.o.maxDecodedSize { |
| 464 | err = ErrDecoderSizeExceeded |
| 465 | break |
| 466 | } |
| 467 | if d.SingleSegment && uint64(len(d.history.b)) > d.o.maxDecodedSize { |
| 468 | println("runDecoder: single segment and", uint64(len(d.history.b)), ">", d.o.maxDecodedSize) |
| 469 | err = ErrFrameSizeExceeded |
| 470 | break |
| 471 | } |
| 472 | } |
| 473 | dst = d.history.b |
| 474 | if err == nil { |
| 475 | if d.HasCheckSum { |
| 476 | var n int |
| 477 | n, err = d.crc.Write(dst[crcStart:]) |
| 478 | if err == nil { |
| 479 | if n != len(dst)-crcStart { |
| 480 | err = io.ErrShortWrite |
| 481 | } else { |
| 482 | err = d.checkCRC() |
| 483 | } |
| 484 | } |
| 485 | } |
| 486 | } |
| 487 | d.history.b = saved |
| 488 | return dst, err |
| 489 | } |