lib/lzo/lzo1x_compress.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  *  LZO1X Compressor from LZO
0004  *
0005  *  Copyright (C) 1996-2012 Markus F.X.J. Oberhumer <markus@oberhumer.com>
0006  *
0007  *  The full LZO package can be found at:
0008  *  http://www.oberhumer.com/opensource/lzo/
0009  *
0010  *  Changed for Linux kernel use by:
0011  *  Nitin Gupta <nitingupta910@gmail.com>
0012  *  Richard Purdie <rpurdie@openedhand.com>
0013  */
0014
0015 #include <linux/module.h>
0016 #include <linux/kernel.h>
0017 #include <asm/unaligned.h>
0018 #include <linux/lzo.h>
0019 #include "lzodefs.h"
0020
0021 static noinline size_t
0022 lzo1x_1_do_compress(const unsigned char *in, size_t in_len,
0023             unsigned char *out, size_t *out_len,
0024             size_t ti, void *wrkmem, signed char *state_offset,
0025             const unsigned char bitstream_version)
0026 {
0027     const unsigned char *ip;
0028     unsigned char *op;
0029     const unsigned char * const in_end = in + in_len;
0030     const unsigned char * const ip_end = in + in_len - 20;
0031     const unsigned char *ii;
0032     lzo_dict_t * const dict = (lzo_dict_t *) wrkmem;
0033
0034     op = out;
0035     ip = in;
0036     ii = ip;
0037     ip += ti < 4 ? 4 - ti : 0;
0038
0039     for (;;) {
0040         const unsigned char *m_pos = NULL;
0041         size_t t, m_len, m_off;
0042         u32 dv;
0043         u32 run_length = 0;
0044 literal:
0045         ip += 1 + ((ip - ii) >> 5);
0046 next:
0047         if (unlikely(ip >= ip_end))
0048             break;
0049         dv = get_unaligned_le32(ip);
0050
0051         if (dv == 0 && bitstream_version) {
0052             const unsigned char *ir = ip + 4;
0053             const unsigned char *limit = min(ip_end, ip + MAX_ZERO_RUN_LENGTH + 1);
0054 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && \
0055     defined(LZO_FAST_64BIT_MEMORY_ACCESS)
0056             u64 dv64;
0057
0058             for (; (ir + 32) <= limit; ir += 32) {
0059                 dv64 = get_unaligned((u64 *)ir);
0060                 dv64 |= get_unaligned((u64 *)ir + 1);
0061                 dv64 |= get_unaligned((u64 *)ir + 2);
0062                 dv64 |= get_unaligned((u64 *)ir + 3);
0063                 if (dv64)
0064                     break;
0065             }
0066             for (; (ir + 8) <= limit; ir += 8) {
0067                 dv64 = get_unaligned((u64 *)ir);
0068                 if (dv64) {
0069 #  if defined(__LITTLE_ENDIAN)
0070                     ir += __builtin_ctzll(dv64) >> 3;
0071 #  elif defined(__BIG_ENDIAN)
0072                     ir += __builtin_clzll(dv64) >> 3;
0073 #  else
0074 #    error "missing endian definition"
0075 #  endif
0076                     break;
0077                 }
0078             }
0079 #else
0080             while ((ir < (const unsigned char *)
0081                     ALIGN((uintptr_t)ir, 4)) &&
0082                     (ir < limit) && (*ir == 0))
0083                 ir++;
0084             if (IS_ALIGNED((uintptr_t)ir, 4)) {
0085                 for (; (ir + 4) <= limit; ir += 4) {
0086                     dv = *((u32 *)ir);
0087                     if (dv) {
0088 #  if defined(__LITTLE_ENDIAN)
0089                         ir += __builtin_ctz(dv) >> 3;
0090 #  elif defined(__BIG_ENDIAN)
0091                         ir += __builtin_clz(dv) >> 3;
0092 #  else
0093 #    error "missing endian definition"
0094 #  endif
0095                         break;
0096                     }
0097                 }
0098             }
0099 #endif
0100             while (likely(ir < limit) && unlikely(*ir == 0))
0101                 ir++;
0102             run_length = ir - ip;
0103             if (run_length > MAX_ZERO_RUN_LENGTH)
0104                 run_length = MAX_ZERO_RUN_LENGTH;
0105         } else {
0106             t = ((dv * 0x1824429d) >> (32 - D_BITS)) & D_MASK;
0107             m_pos = in + dict[t];
0108             dict[t] = (lzo_dict_t) (ip - in);
0109             if (unlikely(dv != get_unaligned_le32(m_pos)))
0110                 goto literal;
0111         }
0112
0113         ii -= ti;
0114         ti = 0;
0115         t = ip - ii;
0116         if (t != 0) {
0117             if (t <= 3) {
0118                 op[*state_offset] |= t;
0119                 COPY4(op, ii);
0120                 op += t;
0121             } else if (t <= 16) {
0122                 *op++ = (t - 3);
0123                 COPY8(op, ii);
0124                 COPY8(op + 8, ii + 8);
0125                 op += t;
0126             } else {
0127                 if (t <= 18) {
0128                     *op++ = (t - 3);
0129                 } else {
0130                     size_t tt = t - 18;
0131                     *op++ = 0;
0132                     while (unlikely(tt > 255)) {
0133                         tt -= 255;
0134                         *op++ = 0;
0135                     }
0136                     *op++ = tt;
0137                 }
0138                 do {
0139                     COPY8(op, ii);
0140                     COPY8(op + 8, ii + 8);
0141                     op += 16;
0142                     ii += 16;
0143                     t -= 16;
0144                 } while (t >= 16);
0145                 if (t > 0) do {
0146                     *op++ = *ii++;
0147                 } while (--t > 0);
0148             }
0149         }
0150
0151         if (unlikely(run_length)) {
0152             ip += run_length;
0153             run_length -= MIN_ZERO_RUN_LENGTH;
0154             put_unaligned_le32((run_length << 21) | 0xfffc18
0155                        | (run_length & 0x7), op);
0156             op += 4;
0157             run_length = 0;
0158             *state_offset = -3;
0159             goto finished_writing_instruction;
0160         }
0161
0162         m_len = 4;
0163         {
0164 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(LZO_USE_CTZ64)
0165         u64 v;
0166         v = get_unaligned((const u64 *) (ip + m_len)) ^
0167             get_unaligned((const u64 *) (m_pos + m_len));
0168         if (unlikely(v == 0)) {
0169             do {
0170                 m_len += 8;
0171                 v = get_unaligned((const u64 *) (ip + m_len)) ^
0172                     get_unaligned((const u64 *) (m_pos + m_len));
0173                 if (unlikely(ip + m_len >= ip_end))
0174                     goto m_len_done;
0175             } while (v == 0);
0176         }
0177 #  if defined(__LITTLE_ENDIAN)
0178         m_len += (unsigned) __builtin_ctzll(v) / 8;
0179 #  elif defined(__BIG_ENDIAN)
0180         m_len += (unsigned) __builtin_clzll(v) / 8;
0181 #  else
0182 #    error "missing endian definition"
0183 #  endif
0184 #elif defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(LZO_USE_CTZ32)
0185         u32 v;
0186         v = get_unaligned((const u32 *) (ip + m_len)) ^
0187             get_unaligned((const u32 *) (m_pos + m_len));
0188         if (unlikely(v == 0)) {
0189             do {
0190                 m_len += 4;
0191                 v = get_unaligned((const u32 *) (ip + m_len)) ^
0192                     get_unaligned((const u32 *) (m_pos + m_len));
0193                 if (v != 0)
0194                     break;
0195                 m_len += 4;
0196                 v = get_unaligned((const u32 *) (ip + m_len)) ^
0197                     get_unaligned((const u32 *) (m_pos + m_len));
0198                 if (unlikely(ip + m_len >= ip_end))
0199                     goto m_len_done;
0200             } while (v == 0);
0201         }
0202 #  if defined(__LITTLE_ENDIAN)
0203         m_len += (unsigned) __builtin_ctz(v) / 8;
0204 #  elif defined(__BIG_ENDIAN)
0205         m_len += (unsigned) __builtin_clz(v) / 8;
0206 #  else
0207 #    error "missing endian definition"
0208 #  endif
0209 #else
0210         if (unlikely(ip[m_len] == m_pos[m_len])) {
0211             do {
0212                 m_len += 1;
0213                 if (ip[m_len] != m_pos[m_len])
0214                     break;
0215                 m_len += 1;
0216                 if (ip[m_len] != m_pos[m_len])
0217                     break;
0218                 m_len += 1;
0219                 if (ip[m_len] != m_pos[m_len])
0220                     break;
0221                 m_len += 1;
0222                 if (ip[m_len] != m_pos[m_len])
0223                     break;
0224                 m_len += 1;
0225                 if (ip[m_len] != m_pos[m_len])
0226                     break;
0227                 m_len += 1;
0228                 if (ip[m_len] != m_pos[m_len])
0229                     break;
0230                 m_len += 1;
0231                 if (ip[m_len] != m_pos[m_len])
0232                     break;
0233                 m_len += 1;
0234                 if (unlikely(ip + m_len >= ip_end))
0235                     goto m_len_done;
0236             } while (ip[m_len] == m_pos[m_len]);
0237         }
0238 #endif
0239         }
0240 m_len_done:
0241
0242         m_off = ip - m_pos;
0243         ip += m_len;
0244         if (m_len <= M2_MAX_LEN && m_off <= M2_MAX_OFFSET) {
0245             m_off -= 1;
0246             *op++ = (((m_len - 1) << 5) | ((m_off & 7) << 2));
0247             *op++ = (m_off >> 3);
0248         } else if (m_off <= M3_MAX_OFFSET) {
0249             m_off -= 1;
0250             if (m_len <= M3_MAX_LEN)
0251                 *op++ = (M3_MARKER | (m_len - 2));
0252             else {
0253                 m_len -= M3_MAX_LEN;
0254                 *op++ = M3_MARKER | 0;
0255                 while (unlikely(m_len > 255)) {
0256                     m_len -= 255;
0257                     *op++ = 0;
0258                 }
0259                 *op++ = (m_len);
0260             }
0261             *op++ = (m_off << 2);
0262             *op++ = (m_off >> 6);
0263         } else {
0264             m_off -= 0x4000;
0265             if (m_len <= M4_MAX_LEN)
0266                 *op++ = (M4_MARKER | ((m_off >> 11) & 8)
0267                         | (m_len - 2));
0268             else {
0269                 if (unlikely(((m_off & 0x403f) == 0x403f)
0270                         && (m_len >= 261)
0271                         && (m_len <= 264))
0272                         && likely(bitstream_version)) {
0273                     // Under lzo-rle, block copies
0274                     // for 261 <= length <= 264 and
0275                     // (distance & 0x80f3) == 0x80f3
0276                     // can result in ambiguous
0277                     // output. Adjust length
0278                     // to 260 to prevent ambiguity.
0279                     ip -= m_len - 260;
0280                     m_len = 260;
0281                 }
0282                 m_len -= M4_MAX_LEN;
0283                 *op++ = (M4_MARKER | ((m_off >> 11) & 8));
0284                 while (unlikely(m_len > 255)) {
0285                     m_len -= 255;
0286                     *op++ = 0;
0287                 }
0288                 *op++ = (m_len);
0289             }
0290             *op++ = (m_off << 2);
0291             *op++ = (m_off >> 6);
0292         }
0293         *state_offset = -2;
0294 finished_writing_instruction:
0295         ii = ip;
0296         goto next;
0297     }
0298     *out_len = op - out;
0299     return in_end - (ii - ti);
0300 }
0301
0302 static int lzogeneric1x_1_compress(const unsigned char *in, size_t in_len,
0303              unsigned char *out, size_t *out_len,
0304              void *wrkmem, const unsigned char bitstream_version)
0305 {
0306     const unsigned char *ip = in;
0307     unsigned char *op = out;
0308     unsigned char *data_start;
0309     size_t l = in_len;
0310     size_t t = 0;
0311     signed char state_offset = -2;
0312     unsigned int m4_max_offset;
0313
0314     // LZO v0 will never write 17 as first byte (except for zero-length
0315     // input), so this is used to version the bitstream
0316     if (bitstream_version > 0) {
0317         *op++ = 17;
0318         *op++ = bitstream_version;
0319         m4_max_offset = M4_MAX_OFFSET_V1;
0320     } else {
0321         m4_max_offset = M4_MAX_OFFSET_V0;
0322     }
0323
0324     data_start = op;
0325
0326     while (l > 20) {
0327         size_t ll = min_t(size_t, l, m4_max_offset + 1);
0328         uintptr_t ll_end = (uintptr_t) ip + ll;
0329         if ((ll_end + ((t + ll) >> 5)) <= ll_end)
0330             break;
0331         BUILD_BUG_ON(D_SIZE * sizeof(lzo_dict_t) > LZO1X_1_MEM_COMPRESS);
0332         memset(wrkmem, 0, D_SIZE * sizeof(lzo_dict_t));
0333         t = lzo1x_1_do_compress(ip, ll, op, out_len, t, wrkmem,
0334                     &state_offset, bitstream_version);
0335         ip += ll;
0336         op += *out_len;
0337         l  -= ll;
0338     }
0339     t += l;
0340
0341     if (t > 0) {
0342         const unsigned char *ii = in + in_len - t;
0343
0344         if (op == data_start && t <= 238) {
0345             *op++ = (17 + t);
0346         } else if (t <= 3) {
0347             op[state_offset] |= t;
0348         } else if (t <= 18) {
0349             *op++ = (t - 3);
0350         } else {
0351             size_t tt = t - 18;
0352             *op++ = 0;
0353             while (tt > 255) {
0354                 tt -= 255;
0355                 *op++ = 0;
0356             }
0357             *op++ = tt;
0358         }
0359         if (t >= 16) do {
0360             COPY8(op, ii);
0361             COPY8(op + 8, ii + 8);
0362             op += 16;
0363             ii += 16;
0364             t -= 16;
0365         } while (t >= 16);
0366         if (t > 0) do {
0367             *op++ = *ii++;
0368         } while (--t > 0);
0369     }
0370
0371     *op++ = M4_MARKER | 1;
0372     *op++ = 0;
0373     *op++ = 0;
0374
0375     *out_len = op - out;
0376     return LZO_E_OK;
0377 }
0378
0379 int lzo1x_1_compress(const unsigned char *in, size_t in_len,
0380              unsigned char *out, size_t *out_len,
0381              void *wrkmem)
0382 {
0383     return lzogeneric1x_1_compress(in, in_len, out, out_len, wrkmem, 0);
0384 }
0385
0386 int lzorle1x_1_compress(const unsigned char *in, size_t in_len,
0387              unsigned char *out, size_t *out_len,
0388              void *wrkmem)
0389 {
0390     return lzogeneric1x_1_compress(in, in_len, out, out_len,
0391                        wrkmem, LZO_VERSION);
0392 }
0393
0394 EXPORT_SYMBOL_GPL(lzo1x_1_compress);
0395 EXPORT_SYMBOL_GPL(lzorle1x_1_compress);
0396
0397 MODULE_LICENSE("GPL");
0398 MODULE_DESCRIPTION("LZO1X-1 Compressor");