xref: /illumos-gate/usr/src/uts/common/os/compress.c (revision 3ba944265c4ae1fcf23ef758537c2e4f4feec16e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 1998 by Sun Microsystems, Inc.
24  * All rights reserved.
25  */
26 
27 /*
28  * NOTE: this file is compiled into the kernel, cprboot, and savecore.
29  * Therefore it must compile in kernel, boot, and userland source context;
30  * so if you ever change this code, avoid references to external symbols.
31  *
32  * This compression algorithm is a derivative of LZRW1, which I'll call
33  * LZJB in the classic LZ* spirit.  All LZ* (Lempel-Ziv) algorithms are
34  * based on the same basic principle: when a "phrase" (sequences of bytes)
35  * is repeated in a data stream, we can save space by storing a reference to
36  * the previous instance of that phrase (a "copy item") rather than storing
37  * the phrase itself (a "literal item").  The compressor remembers phrases
38  * in a simple hash table (the "Lempel history") that maps three-character
39  * sequences (the minimum match) to the addresses where they were last seen.
40  *
41  * A copy item must encode both the length and the location of the matching
42  * phrase so that decompress() can reconstruct the original data stream.
43  * For example, here's how we'd encode "yadda yadda yadda, blah blah blah"
44  * (with "_" replacing spaces for readability):
45  *
46  * Original:
47  *
48  * y a d d a _ y a d d a _ y a d d a , _ b l a h _ b l a h _ b l a h
49  *
50  * Compressed:
51  *
52  * y a d d a _ 6 11 , _ b l a h 5 10
53  *
54  * In the compressed output, the "6 11" simply means "to get the original
55  * data, execute memmove(ptr, ptr - 6, 11)".  Note that in this example,
56  * the match at "6 11" actually extends beyond the current location and
57  * overlaps it.  That's OK; like memmove(), decompress() handles overlap.
58  *
59  * There's still one more thing decompress() needs to know, which is how to
60  * distinguish literal items from copy items.  We encode this information
61  * in an 8-bit bitmap that precedes each 8 items of output; if the Nth bit
62  * is set, then the Nth item is a copy item.  Thus the full encoding for
63  * the example above would be:
64  *
65  * 0x40 y a d d a _ 6 11 , 0x20 _ b l a h 5 10
66  *
67  * Finally, the "6 11" isn't really encoded as the two byte values 6 and 11
68  * in the output stream because, empirically, we get better compression by
69  * dedicating more bits to offset, fewer to match length.  LZJB uses 6 bits
70  * to encode the match length, 10 bits to encode the offset.  Since copy-item
71  * encoding consumes 2 bytes, we don't generate copy items unless the match
72  * length is at least 3; therefore, we can store (length - 3) in the 6-bit
73  * match length field, which extends the maximum match from 63 to 66 bytes.
74  * Thus the 2-byte encoding for a copy item is as follows:
75  *
76  *	byte[0] = ((length - 3) << 2) | (offset >> 8);
77  *	byte[1] = (uint8_t)offset;
78  *
79  * In our example above, an offset of 6 with length 11 would be encoded as:
80  *
81  *	byte[0] = ((11 - 3) << 2) | (6 >> 8) = 0x20
82  *	byte[1] = (uint8_t)6 = 0x6
83  *
84  * Similarly, an offset of 5 with length 10 would be encoded as:
85  *
86  *	byte[0] = ((10 - 3) << 2) | (5 >> 8) = 0x1c
87  *	byte[1] = (uint8_t)5 = 0x5
88  *
89  * Putting it all together, the actual LZJB output for our example is:
90  *
91  * 0x40 y a d d a _ 0x2006 , 0x20 _ b l a h 0x1c05
92  *
93  * The main differences between LZRW1 and LZJB are as follows:
94  *
95  * (1) LZRW1 is sloppy about buffer overruns.  LZJB never reads past the
96  *     end of its input, and never writes past the end of its output.
97  *
98  * (2) LZJB allows a maximum match length of 66 (vs. 18 for LZRW1), with
99  *     the trade-off being a shorter look-behind (1K vs. 4K for LZRW1).
100  *
101  * (3) LZJB records only the low-order 16 bits of pointers in the Lempel
102  *     history (which is all we need since the maximum look-behind is 1K),
103  *     and uses only 256 hash entries (vs. 4096 for LZRW1).  This makes
104  *     the compression hash small enough to allocate on the stack, which
105  *     solves two problems: (1) it saves 64K of kernel/cprboot memory,
106  *     and (2) it makes the code MT-safe without any locking, since we
107  *     don't have multiple threads sharing a common hash table.
108  *
109  * (4) LZJB is faster at both compression and decompression, has a
110  *     better compression ratio, and is somewhat simpler than LZRW1.
111  *
112  * Finally, note that LZJB is non-deterministic: given the same input,
113  * two calls to compress() may produce different output.  This is a
114  * general characteristic of most Lempel-Ziv derivatives because there's
115  * no need to initialize the Lempel history; not doing so saves time.
116  */
117 
118 #include <sys/types.h>
119 #include <sys/param.h>
120 
121 #define	MATCH_BITS	6
122 #define	MATCH_MIN	3
123 #define	MATCH_MAX	((1 << MATCH_BITS) + (MATCH_MIN - 1))
124 #define	OFFSET_MASK	((1 << (16 - MATCH_BITS)) - 1)
125 #define	LEMPEL_SIZE	256
126 
127 size_t
128 compress(void *s_start, void *d_start, size_t s_len)
129 {
130 	uchar_t *src = s_start;
131 	uchar_t *dst = d_start;
132 	uchar_t *cpy, *copymap = NULL;
133 	int copymask = 1 << (NBBY - 1);
134 	int mlen, offset;
135 	uint16_t *hp;
136 	uint16_t lempel[LEMPEL_SIZE];	/* uninitialized; see above */
137 
138 	while (src < (uchar_t *)s_start + s_len) {
139 		if ((copymask <<= 1) == (1 << NBBY)) {
140 			if (dst >= (uchar_t *)d_start + s_len - 1 - 2 * NBBY) {
141 				mlen = s_len;
142 				for (src = s_start, dst = d_start; mlen; mlen--)
143 					*dst++ = *src++;
144 				return (s_len);
145 			}
146 			copymask = 1;
147 			copymap = dst;
148 			*dst++ = 0;
149 		}
150 		if (src > (uchar_t *)s_start + s_len - MATCH_MAX) {
151 			*dst++ = *src++;
152 			continue;
153 		}
154 		hp = &lempel[((src[0] + 13) ^ (src[1] - 13) ^ src[2]) &
155 		    (LEMPEL_SIZE - 1)];
156 		offset = (intptr_t)(src - *hp) & OFFSET_MASK;
157 		*hp = (uint16_t)(uintptr_t)src;
158 		cpy = src - offset;
159 		if (cpy >= (uchar_t *)s_start && cpy != src &&
160 		    src[0] == cpy[0] && src[1] == cpy[1] && src[2] == cpy[2]) {
161 			*copymap |= copymask;
162 			for (mlen = MATCH_MIN; mlen < MATCH_MAX; mlen++)
163 				if (src[mlen] != cpy[mlen])
164 					break;
165 			*dst++ = ((mlen - MATCH_MIN) << (NBBY - MATCH_BITS)) |
166 			    (offset >> NBBY);
167 			*dst++ = (uchar_t)offset;
168 			src += mlen;
169 		} else {
170 			*dst++ = *src++;
171 		}
172 	}
173 	return (dst - (uchar_t *)d_start);
174 }
175 
176 size_t
177 decompress(void *s_start, void *d_start, size_t s_len, size_t d_len)
178 {
179 	uchar_t *src = s_start;
180 	uchar_t *dst = d_start;
181 	uchar_t *s_end = (uchar_t *)s_start + s_len;
182 	uchar_t *d_end = (uchar_t *)d_start + d_len;
183 	uchar_t *cpy, copymap = '\0';
184 	int copymask = 1 << (NBBY - 1);
185 
186 	if (s_len >= d_len) {
187 		size_t d_rem = d_len;
188 		while (d_rem-- != 0)
189 			*dst++ = *src++;
190 		return (d_len);
191 	}
192 
193 	while (src < s_end && dst < d_end) {
194 		if ((copymask <<= 1) == (1 << NBBY)) {
195 			copymask = 1;
196 			copymap = *src++;
197 		}
198 		if (copymap & copymask) {
199 			int mlen = (src[0] >> (NBBY - MATCH_BITS)) + MATCH_MIN;
200 			int offset = ((src[0] << NBBY) | src[1]) & OFFSET_MASK;
201 			src += 2;
202 			if ((cpy = dst - offset) >= (uchar_t *)d_start)
203 				while (--mlen >= 0 && dst < d_end)
204 					*dst++ = *cpy++;
205 			else
206 				/*
207 				 * offset before start of destination buffer
208 				 * indicates corrupt source data
209 				 */
210 				return (dst - (uchar_t *)d_start);
211 		} else {
212 			*dst++ = *src++;
213 		}
214 	}
215 	return (dst - (uchar_t *)d_start);
216 }
217 
218 uint32_t
219 checksum32(void *cp_arg, size_t length)
220 {
221 	uchar_t *cp, *ep;
222 	uint32_t sum = 0;
223 
224 	for (cp = cp_arg, ep = cp + length; cp < ep; cp++)
225 		sum = ((sum >> 1) | (sum << 31)) + *cp;
226 	return (sum);
227 }
228