1// Copyright 2011 The Snappy-Go Authors. All rights reserved.
2// Use of this source code is governed by a BSD-style
3// license that can be found in the LICENSE file.
4
5package snappy
6
7import (
8	"encoding/binary"
9	"errors"
10	"io"
11)
12
13// Encode returns the encoded form of src. The returned slice may be a sub-
14// slice of dst if dst was large enough to hold the entire encoded block.
15// Otherwise, a newly allocated slice will be returned.
16//
17// The dst and src must not overlap. It is valid to pass a nil dst.
18//
19// Encode handles the Snappy block format, not the Snappy stream format.
20func Encode(dst, src []byte) []byte {
21	if n := MaxEncodedLen(len(src)); n < 0 {
22		panic(ErrTooLarge)
23	} else if len(dst) < n {
24		dst = make([]byte, n)
25	}
26
27	// The block starts with the varint-encoded length of the decompressed bytes.
28	d := binary.PutUvarint(dst, uint64(len(src)))
29
30	for len(src) > 0 {
31		p := src
32		src = nil
33		if len(p) > maxBlockSize {
34			p, src = p[:maxBlockSize], p[maxBlockSize:]
35		}
36		if len(p) < minNonLiteralBlockSize {
37			d += emitLiteral(dst[d:], p)
38		} else {
39			d += encodeBlock(dst[d:], p)
40		}
41	}
42	return dst[:d]
43}
44
45// inputMargin is the minimum number of extra input bytes to keep, inside
46// encodeBlock's inner loop. On some architectures, this margin lets us
47// implement a fast path for emitLiteral, where the copy of short (<= 16 byte)
48// literals can be implemented as a single load to and store from a 16-byte
49// register. That literal's actual length can be as short as 1 byte, so this
50// can copy up to 15 bytes too much, but that's OK as subsequent iterations of
51// the encoding loop will fix up the copy overrun, and this inputMargin ensures
52// that we don't overrun the dst and src buffers.
53const inputMargin = 16 - 1
54
55// minNonLiteralBlockSize is the minimum size of the input to encodeBlock that
56// could be encoded with a copy tag. This is the minimum with respect to the
57// algorithm used by encodeBlock, not a minimum enforced by the file format.
58//
59// The encoded output must start with at least a 1 byte literal, as there are
60// no previous bytes to copy. A minimal (1 byte) copy after that, generated
61// from an emitCopy call in encodeBlock's main loop, would require at least
62// another inputMargin bytes, for the reason above: we want any emitLiteral
63// calls inside encodeBlock's main loop to use the fast path if possible, which
64// requires being able to overrun by inputMargin bytes. Thus,
65// minNonLiteralBlockSize equals 1 + 1 + inputMargin.
66//
67// The C++ code doesn't use this exact threshold, but it could, as discussed at
68// https://groups.google.com/d/topic/snappy-compression/oGbhsdIJSJ8/discussion
69// The difference between Go (2+inputMargin) and C++ (inputMargin) is purely an
70// optimization. It should not affect the encoded form. This is tested by
71// TestSameEncodingAsCppShortCopies.
72const minNonLiteralBlockSize = 1 + 1 + inputMargin
73
74// MaxEncodedLen returns the maximum length of a snappy block, given its
75// uncompressed length.
76//
77// It will return a negative value if srcLen is too large to encode.
78func MaxEncodedLen(srcLen int) int {
79	n := uint64(srcLen)
80	if n > 0xffffffff {
81		return -1
82	}
83	// Compressed data can be defined as:
84	//    compressed := item* literal*
85	//    item       := literal* copy
86	//
87	// The trailing literal sequence has a space blowup of at most 62/60
88	// since a literal of length 60 needs one tag byte + one extra byte
89	// for length information.
90	//
91	// Item blowup is trickier to measure. Suppose the "copy" op copies
92	// 4 bytes of data. Because of a special check in the encoding code,
93	// we produce a 4-byte copy only if the offset is < 65536. Therefore
94	// the copy op takes 3 bytes to encode, and this type of item leads
95	// to at most the 62/60 blowup for representing literals.
96	//
97	// Suppose the "copy" op copies 5 bytes of data. If the offset is big
98	// enough, it will take 5 bytes to encode the copy op. Therefore the
99	// worst case here is a one-byte literal followed by a five-byte copy.
100	// That is, 6 bytes of input turn into 7 bytes of "compressed" data.
101	//
102	// This last factor dominates the blowup, so the final estimate is:
103	n = 32 + n + n/6
104	if n > 0xffffffff {
105		return -1
106	}
107	return int(n)
108}
109
110var errClosed = errors.New("snappy: Writer is closed")
111
112// NewWriter returns a new Writer that compresses to w.
113//
114// The Writer returned does not buffer writes. There is no need to Flush or
115// Close such a Writer.
116//
117// Deprecated: the Writer returned is not suitable for many small writes, only
118// for few large writes. Use NewBufferedWriter instead, which is efficient
119// regardless of the frequency and shape of the writes, and remember to Close
120// that Writer when done.
121func NewWriter(w io.Writer) *Writer {
122	return &Writer{
123		w:    w,
124		obuf: make([]byte, obufLen),
125	}
126}
127
128// NewBufferedWriter returns a new Writer that compresses to w, using the
129// framing format described at
130// https://github.com/google/snappy/blob/master/framing_format.txt
131//
132// The Writer returned buffers writes. Users must call Close to guarantee all
133// data has been forwarded to the underlying io.Writer. They may also call
134// Flush zero or more times before calling Close.
135func NewBufferedWriter(w io.Writer) *Writer {
136	return &Writer{
137		w:    w,
138		ibuf: make([]byte, 0, maxBlockSize),
139		obuf: make([]byte, obufLen),
140	}
141}
142
143// Writer is an io.Writer that can write Snappy-compressed bytes.
144//
145// Writer handles the Snappy stream format, not the Snappy block format.
146type Writer struct {
147	w   io.Writer
148	err error
149
150	// ibuf is a buffer for the incoming (uncompressed) bytes.
151	//
152	// Its use is optional. For backwards compatibility, Writers created by the
153	// NewWriter function have ibuf == nil, do not buffer incoming bytes, and
154	// therefore do not need to be Flush'ed or Close'd.
155	ibuf []byte
156
157	// obuf is a buffer for the outgoing (compressed) bytes.
158	obuf []byte
159
160	// wroteStreamHeader is whether we have written the stream header.
161	wroteStreamHeader bool
162}
163
164// Reset discards the writer's state and switches the Snappy writer to write to
165// w. This permits reusing a Writer rather than allocating a new one.
166func (w *Writer) Reset(writer io.Writer) {
167	w.w = writer
168	w.err = nil
169	if w.ibuf != nil {
170		w.ibuf = w.ibuf[:0]
171	}
172	w.wroteStreamHeader = false
173}
174
175// Write satisfies the io.Writer interface.
176func (w *Writer) Write(p []byte) (nRet int, errRet error) {
177	if w.ibuf == nil {
178		// Do not buffer incoming bytes. This does not perform or compress well
179		// if the caller of Writer.Write writes many small slices. This
180		// behavior is therefore deprecated, but still supported for backwards
181		// compatibility with code that doesn't explicitly Flush or Close.
182		return w.write(p)
183	}
184
185	// The remainder of this method is based on bufio.Writer.Write from the
186	// standard library.
187
188	for len(p) > (cap(w.ibuf)-len(w.ibuf)) && w.err == nil {
189		var n int
190		if len(w.ibuf) == 0 {
191			// Large write, empty buffer.
192			// Write directly from p to avoid copy.
193			n, _ = w.write(p)
194		} else {
195			n = copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
196			w.ibuf = w.ibuf[:len(w.ibuf)+n]
197			w.Flush()
198		}
199		nRet += n
200		p = p[n:]
201	}
202	if w.err != nil {
203		return nRet, w.err
204	}
205	n := copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
206	w.ibuf = w.ibuf[:len(w.ibuf)+n]
207	nRet += n
208	return nRet, nil
209}
210
211func (w *Writer) write(p []byte) (nRet int, errRet error) {
212	if w.err != nil {
213		return 0, w.err
214	}
215	for len(p) > 0 {
216		obufStart := len(magicChunk)
217		if !w.wroteStreamHeader {
218			w.wroteStreamHeader = true
219			copy(w.obuf, magicChunk)
220			obufStart = 0
221		}
222
223		var uncompressed []byte
224		if len(p) > maxBlockSize {
225			uncompressed, p = p[:maxBlockSize], p[maxBlockSize:]
226		} else {
227			uncompressed, p = p, nil
228		}
229		checksum := crc(uncompressed)
230
231		// Compress the buffer, discarding the result if the improvement
232		// isn't at least 12.5%.
233		compressed := Encode(w.obuf[obufHeaderLen:], uncompressed)
234		chunkType := uint8(chunkTypeCompressedData)
235		chunkLen := 4 + len(compressed)
236		obufEnd := obufHeaderLen + len(compressed)
237		if len(compressed) >= len(uncompressed)-len(uncompressed)/8 {
238			chunkType = chunkTypeUncompressedData
239			chunkLen = 4 + len(uncompressed)
240			obufEnd = obufHeaderLen
241		}
242
243		// Fill in the per-chunk header that comes before the body.
244		w.obuf[len(magicChunk)+0] = chunkType
245		w.obuf[len(magicChunk)+1] = uint8(chunkLen >> 0)
246		w.obuf[len(magicChunk)+2] = uint8(chunkLen >> 8)
247		w.obuf[len(magicChunk)+3] = uint8(chunkLen >> 16)
248		w.obuf[len(magicChunk)+4] = uint8(checksum >> 0)
249		w.obuf[len(magicChunk)+5] = uint8(checksum >> 8)
250		w.obuf[len(magicChunk)+6] = uint8(checksum >> 16)
251		w.obuf[len(magicChunk)+7] = uint8(checksum >> 24)
252
253		if _, err := w.w.Write(w.obuf[obufStart:obufEnd]); err != nil {
254			w.err = err
255			return nRet, err
256		}
257		if chunkType == chunkTypeUncompressedData {
258			if _, err := w.w.Write(uncompressed); err != nil {
259				w.err = err
260				return nRet, err
261			}
262		}
263		nRet += len(uncompressed)
264	}
265	return nRet, nil
266}
267
268// Flush flushes the Writer to its underlying io.Writer.
269func (w *Writer) Flush() error {
270	if w.err != nil {
271		return w.err
272	}
273	if len(w.ibuf) == 0 {
274		return nil
275	}
276	w.write(w.ibuf)
277	w.ibuf = w.ibuf[:0]
278	return w.err
279}
280
281// Close calls Flush and then closes the Writer.
282func (w *Writer) Close() error {
283	w.Flush()
284	ret := w.err
285	if w.err == nil {
286		w.err = errClosed
287	}
288	return ret
289}
290