xref: /linux/drivers/block/drbd/drbd_vli.h (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 /*
2 -*- linux-c -*-
3    drbd_receiver.c
4    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
5 
6    Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
7    Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8    Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
9 
10    drbd is free software; you can redistribute it and/or modify
11    it under the terms of the GNU General Public License as published by
12    the Free Software Foundation; either version 2, or (at your option)
13    any later version.
14 
15    drbd is distributed in the hope that it will be useful,
16    but WITHOUT ANY WARRANTY; without even the implied warranty of
17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
18    GNU General Public License for more details.
19 
20    You should have received a copy of the GNU General Public License
21    along with drbd; see the file COPYING.  If not, write to
22    the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
23  */
24 
25 #ifndef _DRBD_VLI_H
26 #define _DRBD_VLI_H
27 
28 /*
29  * At a granularity of 4KiB storage represented per bit,
30  * and stroage sizes of several TiB,
31  * and possibly small-bandwidth replication,
32  * the bitmap transfer time can take much too long,
33  * if transmitted in plain text.
34  *
35  * We try to reduce the transferred bitmap information
36  * by encoding runlengths of bit polarity.
37  *
38  * We never actually need to encode a "zero" (runlengths are positive).
39  * But then we have to store the value of the first bit.
40  * The first bit of information thus shall encode if the first runlength
41  * gives the number of set or unset bits.
42  *
43  * We assume that large areas are either completely set or unset,
44  * which gives good compression with any runlength method,
45  * even when encoding the runlength as fixed size 32bit/64bit integers.
46  *
47  * Still, there may be areas where the polarity flips every few bits,
48  * and encoding the runlength sequence of those areas with fix size
49  * integers would be much worse than plaintext.
50  *
51  * We want to encode small runlength values with minimum code length,
52  * while still being able to encode a Huge run of all zeros.
53  *
54  * Thus we need a Variable Length Integer encoding, VLI.
55  *
56  * For some cases, we produce more code bits than plaintext input.
57  * We need to send incompressible chunks as plaintext, skip over them
58  * and then see if the next chunk compresses better.
59  *
60  * We don't care too much about "excellent" compression ratio for large
61  * runlengths (all set/all clear): whether we achieve a factor of 100
62  * or 1000 is not that much of an issue.
63  * We do not want to waste too much on short runlengths in the "noisy"
64  * parts of the bitmap, though.
65  *
66  * There are endless variants of VLI, we experimented with:
67  *  * simple byte-based
68  *  * various bit based with different code word length.
69  *
70  * To avoid yet an other configuration parameter (choice of bitmap compression
71  * algorithm) which was difficult to explain and tune, we just chose the one
72  * variant that turned out best in all test cases.
73  * Based on real world usage patterns, with device sizes ranging from a few GiB
74  * to several TiB, file server/mailserver/webserver/mysql/postgress,
75  * mostly idle to really busy, the all time winner (though sometimes only
76  * marginally better) is:
77  */
78 
79 /*
80  * encoding is "visualised" as
81  * __little endian__ bitstream, least significant bit first (left most)
82  *
83  * this particular encoding is chosen so that the prefix code
84  * starts as unary encoding the level, then modified so that
85  * 10 levels can be described in 8bit, with minimal overhead
86  * for the smaller levels.
87  *
88  * Number of data bits follow fibonacci sequence, with the exception of the
89  * last level (+1 data bit, so it makes 64bit total).  The only worse code when
90  * encoding bit polarity runlength is 1 plain bits => 2 code bits.
91 prefix    data bits                                    max val  Nº data bits
92 0 x                                                         0x2            1
93 10 x                                                        0x4            1
94 110 xx                                                      0x8            2
95 1110 xxx                                                   0x10            3
96 11110 xxx xx                                               0x30            5
97 111110 xx xxxxxx                                          0x130            8
98 11111100  xxxxxxxx xxxxx                                 0x2130           13
99 11111110  xxxxxxxx xxxxxxxx xxxxx                      0x202130           21
100 11111101  xxxxxxxx xxxxxxxx xxxxxxxx  xxxxxxxx xx   0x400202130           34
101 11111111  xxxxxxxx xxxxxxxx xxxxxxxx  xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx 56
102  * maximum encodable value: 0x100000400202130 == 2**56 + some */
103 
104 /* compression "table":
105  transmitted   x                                0.29
106  as plaintext x                                  ........................
107              x                                   ........................
108             x                                    ........................
109            x    0.59                         0.21........................
110           x      ........................................................
111          x       .. c ...................................................
112         x    0.44.. o ...................................................
113        x .......... d ...................................................
114       x  .......... e ...................................................
115      X.............   ...................................................
116     x.............. b ...................................................
117 2.0x............... i ...................................................
118  #X................ t ...................................................
119  #................. s ...........................  plain bits  ..........
120 -+-----------------------------------------------------------------------
121  1             16              32                              64
122 */
123 
124 /* LEVEL: (total bits, prefix bits, prefix value),
125  * sorted ascending by number of total bits.
126  * The rest of the code table is calculated at compiletime from this. */
127 
128 /* fibonacci data 1, 1, ... */
129 #define VLI_L_1_1() do { \
130 	LEVEL( 2, 1, 0x00); \
131 	LEVEL( 3, 2, 0x01); \
132 	LEVEL( 5, 3, 0x03); \
133 	LEVEL( 7, 4, 0x07); \
134 	LEVEL(10, 5, 0x0f); \
135 	LEVEL(14, 6, 0x1f); \
136 	LEVEL(21, 8, 0x3f); \
137 	LEVEL(29, 8, 0x7f); \
138 	LEVEL(42, 8, 0xbf); \
139 	LEVEL(64, 8, 0xff); \
140 	} while (0)
141 
142 /* finds a suitable level to decode the least significant part of in.
143  * returns number of bits consumed.
144  *
145  * BUG() for bad input, as that would mean a buggy code table. */
146 static inline int vli_decode_bits(u64 *out, const u64 in)
147 {
148 	u64 adj = 1;
149 
150 #define LEVEL(t,b,v)					\
151 	do {						\
152 		if ((in & ((1 << b) -1)) == v) {	\
153 			*out = ((in & ((~0ULL) >> (64-t))) >> b) + adj;	\
154 			return t;			\
155 		}					\
156 		adj += 1ULL << (t - b);			\
157 	} while (0)
158 
159 	VLI_L_1_1();
160 
161 	/* NOT REACHED, if VLI_LEVELS code table is defined properly */
162 	BUG();
163 #undef LEVEL
164 }
165 
166 /* return number of code bits needed,
167  * or negative error number */
168 static inline int __vli_encode_bits(u64 *out, const u64 in)
169 {
170 	u64 max = 0;
171 	u64 adj = 1;
172 
173 	if (in == 0)
174 		return -EINVAL;
175 
176 #define LEVEL(t,b,v) do {		\
177 		max += 1ULL << (t - b);	\
178 		if (in <= max) {	\
179 			if (out)	\
180 				*out = ((in - adj) << b) | v;	\
181 			return t;	\
182 		}			\
183 		adj = max + 1;		\
184 	} while (0)
185 
186 	VLI_L_1_1();
187 
188 	return -EOVERFLOW;
189 #undef LEVEL
190 }
191 
192 #undef VLI_L_1_1
193 
194 /* code from here down is independend of actually used bit code */
195 
196 /*
197  * Code length is determined by some unique (e.g. unary) prefix.
198  * This encodes arbitrary bit length, not whole bytes: we have a bit-stream,
199  * not a byte stream.
200  */
201 
202 /* for the bitstream, we need a cursor */
203 struct bitstream_cursor {
204 	/* the current byte */
205 	u8 *b;
206 	/* the current bit within *b, nomalized: 0..7 */
207 	unsigned int bit;
208 };
209 
210 /* initialize cursor to point to first bit of stream */
211 static inline void bitstream_cursor_reset(struct bitstream_cursor *cur, void *s)
212 {
213 	cur->b = s;
214 	cur->bit = 0;
215 }
216 
217 /* advance cursor by that many bits; maximum expected input value: 64,
218  * but depending on VLI implementation, it may be more. */
219 static inline void bitstream_cursor_advance(struct bitstream_cursor *cur, unsigned int bits)
220 {
221 	bits += cur->bit;
222 	cur->b = cur->b + (bits >> 3);
223 	cur->bit = bits & 7;
224 }
225 
226 /* the bitstream itself knows its length */
227 struct bitstream {
228 	struct bitstream_cursor cur;
229 	unsigned char *buf;
230 	size_t buf_len;		/* in bytes */
231 
232 	/* for input stream:
233 	 * number of trailing 0 bits for padding
234 	 * total number of valid bits in stream: buf_len * 8 - pad_bits */
235 	unsigned int pad_bits;
236 };
237 
238 static inline void bitstream_init(struct bitstream *bs, void *s, size_t len, unsigned int pad_bits)
239 {
240 	bs->buf = s;
241 	bs->buf_len = len;
242 	bs->pad_bits = pad_bits;
243 	bitstream_cursor_reset(&bs->cur, bs->buf);
244 }
245 
246 static inline void bitstream_rewind(struct bitstream *bs)
247 {
248 	bitstream_cursor_reset(&bs->cur, bs->buf);
249 	memset(bs->buf, 0, bs->buf_len);
250 }
251 
252 /* Put (at most 64) least significant bits of val into bitstream, and advance cursor.
253  * Ignores "pad_bits".
254  * Returns zero if bits == 0 (nothing to do).
255  * Returns number of bits used if successful.
256  *
257  * If there is not enough room left in bitstream,
258  * leaves bitstream unchanged and returns -ENOBUFS.
259  */
260 static inline int bitstream_put_bits(struct bitstream *bs, u64 val, const unsigned int bits)
261 {
262 	unsigned char *b = bs->cur.b;
263 	unsigned int tmp;
264 
265 	if (bits == 0)
266 		return 0;
267 
268 	if ((bs->cur.b + ((bs->cur.bit + bits -1) >> 3)) - bs->buf >= bs->buf_len)
269 		return -ENOBUFS;
270 
271 	/* paranoia: strip off hi bits; they should not be set anyways. */
272 	if (bits < 64)
273 		val &= ~0ULL >> (64 - bits);
274 
275 	*b++ |= (val & 0xff) << bs->cur.bit;
276 
277 	for (tmp = 8 - bs->cur.bit; tmp < bits; tmp += 8)
278 		*b++ |= (val >> tmp) & 0xff;
279 
280 	bitstream_cursor_advance(&bs->cur, bits);
281 	return bits;
282 }
283 
284 /* Fetch (at most 64) bits from bitstream into *out, and advance cursor.
285  *
286  * If more than 64 bits are requested, returns -EINVAL and leave *out unchanged.
287  *
288  * If there are less than the requested number of valid bits left in the
289  * bitstream, still fetches all available bits.
290  *
291  * Returns number of actually fetched bits.
292  */
293 static inline int bitstream_get_bits(struct bitstream *bs, u64 *out, int bits)
294 {
295 	u64 val;
296 	unsigned int n;
297 
298 	if (bits > 64)
299 		return -EINVAL;
300 
301 	if (bs->cur.b + ((bs->cur.bit + bs->pad_bits + bits -1) >> 3) - bs->buf >= bs->buf_len)
302 		bits = ((bs->buf_len - (bs->cur.b - bs->buf)) << 3)
303 			- bs->cur.bit - bs->pad_bits;
304 
305 	if (bits == 0) {
306 		*out = 0;
307 		return 0;
308 	}
309 
310 	/* get the high bits */
311 	val = 0;
312 	n = (bs->cur.bit + bits + 7) >> 3;
313 	/* n may be at most 9, if cur.bit + bits > 64 */
314 	/* which means this copies at most 8 byte */
315 	if (n) {
316 		memcpy(&val, bs->cur.b+1, n - 1);
317 		val = le64_to_cpu(val) << (8 - bs->cur.bit);
318 	}
319 
320 	/* we still need the low bits */
321 	val |= bs->cur.b[0] >> bs->cur.bit;
322 
323 	/* and mask out bits we don't want */
324 	val &= ~0ULL >> (64 - bits);
325 
326 	bitstream_cursor_advance(&bs->cur, bits);
327 	*out = val;
328 
329 	return bits;
330 }
331 
332 /* encodes @in as vli into @bs;
333 
334  * return values
335  *  > 0: number of bits successfully stored in bitstream
336  * -ENOBUFS @bs is full
337  * -EINVAL input zero (invalid)
338  * -EOVERFLOW input too large for this vli code (invalid)
339  */
340 static inline int vli_encode_bits(struct bitstream *bs, u64 in)
341 {
342 	u64 code = code;
343 	int bits = __vli_encode_bits(&code, in);
344 
345 	if (bits <= 0)
346 		return bits;
347 
348 	return bitstream_put_bits(bs, code, bits);
349 }
350 
351 #endif
352