1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2019 HUAWEI, Inc. 4 * http://www.huawei.com/ 5 * Created by Gao Xiang <gaoxiang25@huawei.com> 6 */ 7 #include "compress.h" 8 #include <linux/module.h> 9 #include <linux/lz4.h> 10 11 #ifndef LZ4_DISTANCE_MAX /* history window size */ 12 #define LZ4_DISTANCE_MAX 65535 /* set to maximum value by default */ 13 #endif 14 15 #define LZ4_MAX_DISTANCE_PAGES (DIV_ROUND_UP(LZ4_DISTANCE_MAX, PAGE_SIZE) + 1) 16 #ifndef LZ4_DECOMPRESS_INPLACE_MARGIN 17 #define LZ4_DECOMPRESS_INPLACE_MARGIN(srcsize) (((srcsize) >> 8) + 32) 18 #endif 19 20 struct z_erofs_decompressor { 21 /* 22 * if destpages have sparsed pages, fill them with bounce pages. 23 * it also check whether destpages indicate continuous physical memory. 24 */ 25 int (*prepare_destpages)(struct z_erofs_decompress_req *rq, 26 struct list_head *pagepool); 27 int (*decompress)(struct z_erofs_decompress_req *rq, u8 *out); 28 char *name; 29 }; 30 31 static int z_erofs_lz4_prepare_destpages(struct z_erofs_decompress_req *rq, 32 struct list_head *pagepool) 33 { 34 const unsigned int nr = 35 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT; 36 struct page *availables[LZ4_MAX_DISTANCE_PAGES] = { NULL }; 37 unsigned long bounced[DIV_ROUND_UP(LZ4_MAX_DISTANCE_PAGES, 38 BITS_PER_LONG)] = { 0 }; 39 void *kaddr = NULL; 40 unsigned int i, j, top; 41 42 top = 0; 43 for (i = j = 0; i < nr; ++i, ++j) { 44 struct page *const page = rq->out[i]; 45 struct page *victim; 46 47 if (j >= LZ4_MAX_DISTANCE_PAGES) 48 j = 0; 49 50 /* 'valid' bounced can only be tested after a complete round */ 51 if (test_bit(j, bounced)) { 52 DBG_BUGON(i < LZ4_MAX_DISTANCE_PAGES); 53 DBG_BUGON(top >= LZ4_MAX_DISTANCE_PAGES); 54 availables[top++] = rq->out[i - LZ4_MAX_DISTANCE_PAGES]; 55 } 56 57 if (page) { 58 __clear_bit(j, bounced); 59 if (kaddr) { 60 if (kaddr + PAGE_SIZE == page_address(page)) 61 kaddr += PAGE_SIZE; 62 else 63 kaddr = NULL; 64 } else if (!i) { 65 kaddr = page_address(page); 66 } 67 continue; 68 } 69 kaddr = NULL; 70 __set_bit(j, bounced); 71 72 if (top) { 73 victim = availables[--top]; 74 get_page(victim); 75 } else { 76 victim = erofs_allocpage(pagepool, GFP_KERNEL); 77 if (!victim) 78 return -ENOMEM; 79 victim->mapping = Z_EROFS_MAPPING_STAGING; 80 } 81 rq->out[i] = victim; 82 } 83 return kaddr ? 1 : 0; 84 } 85 86 static void *generic_copy_inplace_data(struct z_erofs_decompress_req *rq, 87 u8 *src, unsigned int pageofs_in) 88 { 89 /* 90 * if in-place decompression is ongoing, those decompressed 91 * pages should be copied in order to avoid being overlapped. 92 */ 93 struct page **in = rq->in; 94 u8 *const tmp = erofs_get_pcpubuf(0); 95 u8 *tmpp = tmp; 96 unsigned int inlen = rq->inputsize - pageofs_in; 97 unsigned int count = min_t(uint, inlen, PAGE_SIZE - pageofs_in); 98 99 while (tmpp < tmp + inlen) { 100 if (!src) 101 src = kmap_atomic(*in); 102 memcpy(tmpp, src + pageofs_in, count); 103 kunmap_atomic(src); 104 src = NULL; 105 tmpp += count; 106 pageofs_in = 0; 107 count = PAGE_SIZE; 108 ++in; 109 } 110 return tmp; 111 } 112 113 static int z_erofs_lz4_decompress(struct z_erofs_decompress_req *rq, u8 *out) 114 { 115 unsigned int inputmargin, inlen; 116 u8 *src; 117 bool copied, support_0padding; 118 int ret; 119 120 if (rq->inputsize > PAGE_SIZE) 121 return -EOPNOTSUPP; 122 123 src = kmap_atomic(*rq->in); 124 inputmargin = 0; 125 support_0padding = false; 126 127 /* decompression inplace is only safe when 0padding is enabled */ 128 if (EROFS_SB(rq->sb)->feature_incompat & 129 EROFS_FEATURE_INCOMPAT_LZ4_0PADDING) { 130 support_0padding = true; 131 132 while (!src[inputmargin & ~PAGE_MASK]) 133 if (!(++inputmargin & ~PAGE_MASK)) 134 break; 135 136 if (inputmargin >= rq->inputsize) { 137 kunmap_atomic(src); 138 return -EIO; 139 } 140 } 141 142 copied = false; 143 inlen = rq->inputsize - inputmargin; 144 if (rq->inplace_io) { 145 const uint oend = (rq->pageofs_out + 146 rq->outputsize) & ~PAGE_MASK; 147 const uint nr = PAGE_ALIGN(rq->pageofs_out + 148 rq->outputsize) >> PAGE_SHIFT; 149 150 if (rq->partial_decoding || !support_0padding || 151 rq->out[nr - 1] != rq->in[0] || 152 rq->inputsize - oend < 153 LZ4_DECOMPRESS_INPLACE_MARGIN(inlen)) { 154 src = generic_copy_inplace_data(rq, src, inputmargin); 155 inputmargin = 0; 156 copied = true; 157 } 158 } 159 160 /* legacy format could compress extra data in a pcluster. */ 161 if (rq->partial_decoding || !support_0padding) 162 ret = LZ4_decompress_safe_partial(src + inputmargin, out, 163 inlen, rq->outputsize, 164 rq->outputsize); 165 else 166 ret = LZ4_decompress_safe(src + inputmargin, out, 167 inlen, rq->outputsize); 168 169 if (ret != rq->outputsize) { 170 erofs_err(rq->sb, "failed to decompress %d in[%u, %u] out[%u]", 171 ret, inlen, inputmargin, rq->outputsize); 172 173 WARN_ON(1); 174 print_hex_dump(KERN_DEBUG, "[ in]: ", DUMP_PREFIX_OFFSET, 175 16, 1, src + inputmargin, inlen, true); 176 print_hex_dump(KERN_DEBUG, "[out]: ", DUMP_PREFIX_OFFSET, 177 16, 1, out, rq->outputsize, true); 178 179 if (ret >= 0) 180 memset(out + ret, 0, rq->outputsize - ret); 181 ret = -EIO; 182 } 183 184 if (copied) 185 erofs_put_pcpubuf(src); 186 else 187 kunmap_atomic(src); 188 return ret; 189 } 190 191 static struct z_erofs_decompressor decompressors[] = { 192 [Z_EROFS_COMPRESSION_SHIFTED] = { 193 .name = "shifted" 194 }, 195 [Z_EROFS_COMPRESSION_LZ4] = { 196 .prepare_destpages = z_erofs_lz4_prepare_destpages, 197 .decompress = z_erofs_lz4_decompress, 198 .name = "lz4" 199 }, 200 }; 201 202 static void copy_from_pcpubuf(struct page **out, const char *dst, 203 unsigned short pageofs_out, 204 unsigned int outputsize) 205 { 206 const char *end = dst + outputsize; 207 const unsigned int righthalf = PAGE_SIZE - pageofs_out; 208 const char *cur = dst - pageofs_out; 209 210 while (cur < end) { 211 struct page *const page = *out++; 212 213 if (page) { 214 char *buf = kmap_atomic(page); 215 216 if (cur >= dst) { 217 memcpy(buf, cur, min_t(uint, PAGE_SIZE, 218 end - cur)); 219 } else { 220 memcpy(buf + pageofs_out, cur + pageofs_out, 221 min_t(uint, righthalf, end - cur)); 222 } 223 kunmap_atomic(buf); 224 } 225 cur += PAGE_SIZE; 226 } 227 } 228 229 static int z_erofs_decompress_generic(struct z_erofs_decompress_req *rq, 230 struct list_head *pagepool) 231 { 232 const unsigned int nrpages_out = 233 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT; 234 const struct z_erofs_decompressor *alg = decompressors + rq->alg; 235 unsigned int dst_maptype; 236 void *dst; 237 int ret, i; 238 239 if (nrpages_out == 1 && !rq->inplace_io) { 240 DBG_BUGON(!*rq->out); 241 dst = kmap_atomic(*rq->out); 242 dst_maptype = 0; 243 goto dstmap_out; 244 } 245 246 /* 247 * For the case of small output size (especially much less 248 * than PAGE_SIZE), memcpy the decompressed data rather than 249 * compressed data is preferred. 250 */ 251 if (rq->outputsize <= PAGE_SIZE * 7 / 8) { 252 dst = erofs_get_pcpubuf(0); 253 if (IS_ERR(dst)) 254 return PTR_ERR(dst); 255 256 rq->inplace_io = false; 257 ret = alg->decompress(rq, dst); 258 if (!ret) 259 copy_from_pcpubuf(rq->out, dst, rq->pageofs_out, 260 rq->outputsize); 261 262 erofs_put_pcpubuf(dst); 263 return ret; 264 } 265 266 ret = alg->prepare_destpages(rq, pagepool); 267 if (ret < 0) { 268 return ret; 269 } else if (ret) { 270 dst = page_address(*rq->out); 271 dst_maptype = 1; 272 goto dstmap_out; 273 } 274 275 i = 0; 276 while (1) { 277 dst = vm_map_ram(rq->out, nrpages_out, -1, PAGE_KERNEL); 278 279 /* retry two more times (totally 3 times) */ 280 if (dst || ++i >= 3) 281 break; 282 vm_unmap_aliases(); 283 } 284 285 if (!dst) 286 return -ENOMEM; 287 288 dst_maptype = 2; 289 290 dstmap_out: 291 ret = alg->decompress(rq, dst + rq->pageofs_out); 292 293 if (!dst_maptype) 294 kunmap_atomic(dst); 295 else if (dst_maptype == 2) 296 vm_unmap_ram(dst, nrpages_out); 297 return ret; 298 } 299 300 static int z_erofs_shifted_transform(const struct z_erofs_decompress_req *rq, 301 struct list_head *pagepool) 302 { 303 const unsigned int nrpages_out = 304 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT; 305 const unsigned int righthalf = PAGE_SIZE - rq->pageofs_out; 306 unsigned char *src, *dst; 307 308 if (nrpages_out > 2) { 309 DBG_BUGON(1); 310 return -EIO; 311 } 312 313 if (rq->out[0] == *rq->in) { 314 DBG_BUGON(nrpages_out != 1); 315 return 0; 316 } 317 318 src = kmap_atomic(*rq->in); 319 if (rq->out[0]) { 320 dst = kmap_atomic(rq->out[0]); 321 memcpy(dst + rq->pageofs_out, src, righthalf); 322 kunmap_atomic(dst); 323 } 324 325 if (nrpages_out == 2) { 326 DBG_BUGON(!rq->out[1]); 327 if (rq->out[1] == *rq->in) { 328 memmove(src, src + righthalf, rq->pageofs_out); 329 } else { 330 dst = kmap_atomic(rq->out[1]); 331 memcpy(dst, src + righthalf, rq->pageofs_out); 332 kunmap_atomic(dst); 333 } 334 } 335 kunmap_atomic(src); 336 return 0; 337 } 338 339 int z_erofs_decompress(struct z_erofs_decompress_req *rq, 340 struct list_head *pagepool) 341 { 342 if (rq->alg == Z_EROFS_COMPRESSION_SHIFTED) 343 return z_erofs_shifted_transform(rq, pagepool); 344 return z_erofs_decompress_generic(rq, pagepool); 345 } 346 347