1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2019 HUAWEI, Inc. 4 * https://www.huawei.com/ 5 */ 6 #include "compress.h" 7 #include <linux/module.h> 8 #include <linux/lz4.h> 9 10 #ifndef LZ4_DISTANCE_MAX /* history window size */ 11 #define LZ4_DISTANCE_MAX 65535 /* set to maximum value by default */ 12 #endif 13 14 #define LZ4_MAX_DISTANCE_PAGES (DIV_ROUND_UP(LZ4_DISTANCE_MAX, PAGE_SIZE) + 1) 15 #ifndef LZ4_DECOMPRESS_INPLACE_MARGIN 16 #define LZ4_DECOMPRESS_INPLACE_MARGIN(srcsize) (((srcsize) >> 8) + 32) 17 #endif 18 19 struct z_erofs_lz4_decompress_ctx { 20 struct z_erofs_decompress_req *rq; 21 /* # of encoded, decoded pages */ 22 unsigned int inpages, outpages; 23 /* decoded block total length (used for in-place decompression) */ 24 unsigned int oend; 25 }; 26 27 int z_erofs_load_lz4_config(struct super_block *sb, 28 struct erofs_super_block *dsb, 29 struct z_erofs_lz4_cfgs *lz4, int size) 30 { 31 struct erofs_sb_info *sbi = EROFS_SB(sb); 32 u16 distance; 33 34 if (lz4) { 35 if (size < sizeof(struct z_erofs_lz4_cfgs)) { 36 erofs_err(sb, "invalid lz4 cfgs, size=%u", size); 37 return -EINVAL; 38 } 39 distance = le16_to_cpu(lz4->max_distance); 40 41 sbi->lz4.max_pclusterblks = le16_to_cpu(lz4->max_pclusterblks); 42 if (!sbi->lz4.max_pclusterblks) { 43 sbi->lz4.max_pclusterblks = 1; /* reserved case */ 44 } else if (sbi->lz4.max_pclusterblks > 45 Z_EROFS_PCLUSTER_MAX_SIZE / EROFS_BLKSIZ) { 46 erofs_err(sb, "too large lz4 pclusterblks %u", 47 sbi->lz4.max_pclusterblks); 48 return -EINVAL; 49 } else if (sbi->lz4.max_pclusterblks >= 2) { 50 erofs_info(sb, "EXPERIMENTAL big pcluster feature in use. Use at your own risk!"); 51 } 52 } else { 53 distance = le16_to_cpu(dsb->u1.lz4_max_distance); 54 sbi->lz4.max_pclusterblks = 1; 55 } 56 57 sbi->lz4.max_distance_pages = distance ? 58 DIV_ROUND_UP(distance, PAGE_SIZE) + 1 : 59 LZ4_MAX_DISTANCE_PAGES; 60 return erofs_pcpubuf_growsize(sbi->lz4.max_pclusterblks); 61 } 62 63 /* 64 * Fill all gaps with bounce pages if it's a sparse page list. Also check if 65 * all physical pages are consecutive, which can be seen for moderate CR. 66 */ 67 static int z_erofs_lz4_prepare_dstpages(struct z_erofs_lz4_decompress_ctx *ctx, 68 struct page **pagepool) 69 { 70 struct z_erofs_decompress_req *rq = ctx->rq; 71 struct page *availables[LZ4_MAX_DISTANCE_PAGES] = { NULL }; 72 unsigned long bounced[DIV_ROUND_UP(LZ4_MAX_DISTANCE_PAGES, 73 BITS_PER_LONG)] = { 0 }; 74 unsigned int lz4_max_distance_pages = 75 EROFS_SB(rq->sb)->lz4.max_distance_pages; 76 void *kaddr = NULL; 77 unsigned int i, j, top; 78 79 top = 0; 80 for (i = j = 0; i < ctx->outpages; ++i, ++j) { 81 struct page *const page = rq->out[i]; 82 struct page *victim; 83 84 if (j >= lz4_max_distance_pages) 85 j = 0; 86 87 /* 'valid' bounced can only be tested after a complete round */ 88 if (test_bit(j, bounced)) { 89 DBG_BUGON(i < lz4_max_distance_pages); 90 DBG_BUGON(top >= lz4_max_distance_pages); 91 availables[top++] = rq->out[i - lz4_max_distance_pages]; 92 } 93 94 if (page) { 95 __clear_bit(j, bounced); 96 if (kaddr) { 97 if (kaddr + PAGE_SIZE == page_address(page)) 98 kaddr += PAGE_SIZE; 99 else 100 kaddr = NULL; 101 } else if (!i) { 102 kaddr = page_address(page); 103 } 104 continue; 105 } 106 kaddr = NULL; 107 __set_bit(j, bounced); 108 109 if (top) { 110 victim = availables[--top]; 111 get_page(victim); 112 } else { 113 victim = erofs_allocpage(pagepool, 114 GFP_KERNEL | __GFP_NOFAIL); 115 set_page_private(victim, Z_EROFS_SHORTLIVED_PAGE); 116 } 117 rq->out[i] = victim; 118 } 119 return kaddr ? 1 : 0; 120 } 121 122 static void *z_erofs_lz4_handle_overlap(struct z_erofs_lz4_decompress_ctx *ctx, 123 void *inpage, unsigned int *inputmargin, int *maptype, 124 bool may_inplace) 125 { 126 struct z_erofs_decompress_req *rq = ctx->rq; 127 unsigned int omargin, total, i, j; 128 struct page **in; 129 void *src, *tmp; 130 131 if (rq->inplace_io) { 132 omargin = PAGE_ALIGN(ctx->oend) - ctx->oend; 133 if (rq->partial_decoding || !may_inplace || 134 omargin < LZ4_DECOMPRESS_INPLACE_MARGIN(rq->inputsize)) 135 goto docopy; 136 137 for (i = 0; i < ctx->inpages; ++i) { 138 DBG_BUGON(rq->in[i] == NULL); 139 for (j = 0; j < ctx->outpages - ctx->inpages + i; ++j) 140 if (rq->out[j] == rq->in[i]) 141 goto docopy; 142 } 143 } 144 145 if (ctx->inpages <= 1) { 146 *maptype = 0; 147 return inpage; 148 } 149 kunmap_atomic(inpage); 150 might_sleep(); 151 src = erofs_vm_map_ram(rq->in, ctx->inpages); 152 if (!src) 153 return ERR_PTR(-ENOMEM); 154 *maptype = 1; 155 return src; 156 157 docopy: 158 /* Or copy compressed data which can be overlapped to per-CPU buffer */ 159 in = rq->in; 160 src = erofs_get_pcpubuf(ctx->inpages); 161 if (!src) { 162 DBG_BUGON(1); 163 kunmap_atomic(inpage); 164 return ERR_PTR(-EFAULT); 165 } 166 167 tmp = src; 168 total = rq->inputsize; 169 while (total) { 170 unsigned int page_copycnt = 171 min_t(unsigned int, total, PAGE_SIZE - *inputmargin); 172 173 if (!inpage) 174 inpage = kmap_atomic(*in); 175 memcpy(tmp, inpage + *inputmargin, page_copycnt); 176 kunmap_atomic(inpage); 177 inpage = NULL; 178 tmp += page_copycnt; 179 total -= page_copycnt; 180 ++in; 181 *inputmargin = 0; 182 } 183 *maptype = 2; 184 return src; 185 } 186 187 /* 188 * Get the exact inputsize with zero_padding feature. 189 * - For LZ4, it should work if zero_padding feature is on (5.3+); 190 * - For MicroLZMA, it'd be enabled all the time. 191 */ 192 int z_erofs_fixup_insize(struct z_erofs_decompress_req *rq, const char *padbuf, 193 unsigned int padbufsize) 194 { 195 const char *padend; 196 197 padend = memchr_inv(padbuf, 0, padbufsize); 198 if (!padend) 199 return -EFSCORRUPTED; 200 rq->inputsize -= padend - padbuf; 201 rq->pageofs_in += padend - padbuf; 202 return 0; 203 } 204 205 static int z_erofs_lz4_decompress_mem(struct z_erofs_lz4_decompress_ctx *ctx, 206 u8 *out) 207 { 208 struct z_erofs_decompress_req *rq = ctx->rq; 209 bool support_0padding = false, may_inplace = false; 210 unsigned int inputmargin; 211 u8 *headpage, *src; 212 int ret, maptype; 213 214 DBG_BUGON(*rq->in == NULL); 215 headpage = kmap_atomic(*rq->in); 216 217 /* LZ4 decompression inplace is only safe if zero_padding is enabled */ 218 if (erofs_sb_has_zero_padding(EROFS_SB(rq->sb))) { 219 support_0padding = true; 220 ret = z_erofs_fixup_insize(rq, headpage + rq->pageofs_in, 221 min_t(unsigned int, rq->inputsize, 222 EROFS_BLKSIZ - rq->pageofs_in)); 223 if (ret) { 224 kunmap_atomic(headpage); 225 return ret; 226 } 227 may_inplace = !((rq->pageofs_in + rq->inputsize) & 228 (EROFS_BLKSIZ - 1)); 229 } 230 231 inputmargin = rq->pageofs_in; 232 src = z_erofs_lz4_handle_overlap(ctx, headpage, &inputmargin, 233 &maptype, may_inplace); 234 if (IS_ERR(src)) 235 return PTR_ERR(src); 236 237 /* legacy format could compress extra data in a pcluster. */ 238 if (rq->partial_decoding || !support_0padding) 239 ret = LZ4_decompress_safe_partial(src + inputmargin, out, 240 rq->inputsize, rq->outputsize, rq->outputsize); 241 else 242 ret = LZ4_decompress_safe(src + inputmargin, out, 243 rq->inputsize, rq->outputsize); 244 245 if (ret != rq->outputsize) { 246 erofs_err(rq->sb, "failed to decompress %d in[%u, %u] out[%u]", 247 ret, rq->inputsize, inputmargin, rq->outputsize); 248 249 print_hex_dump(KERN_DEBUG, "[ in]: ", DUMP_PREFIX_OFFSET, 250 16, 1, src + inputmargin, rq->inputsize, true); 251 print_hex_dump(KERN_DEBUG, "[out]: ", DUMP_PREFIX_OFFSET, 252 16, 1, out, rq->outputsize, true); 253 254 if (ret >= 0) 255 memset(out + ret, 0, rq->outputsize - ret); 256 ret = -EIO; 257 } else { 258 ret = 0; 259 } 260 261 if (maptype == 0) { 262 kunmap_atomic(headpage); 263 } else if (maptype == 1) { 264 vm_unmap_ram(src, ctx->inpages); 265 } else if (maptype == 2) { 266 erofs_put_pcpubuf(src); 267 } else { 268 DBG_BUGON(1); 269 return -EFAULT; 270 } 271 return ret; 272 } 273 274 static int z_erofs_lz4_decompress(struct z_erofs_decompress_req *rq, 275 struct page **pagepool) 276 { 277 struct z_erofs_lz4_decompress_ctx ctx; 278 unsigned int dst_maptype; 279 void *dst; 280 int ret; 281 282 ctx.rq = rq; 283 ctx.oend = rq->pageofs_out + rq->outputsize; 284 ctx.outpages = PAGE_ALIGN(ctx.oend) >> PAGE_SHIFT; 285 ctx.inpages = PAGE_ALIGN(rq->inputsize) >> PAGE_SHIFT; 286 287 /* one optimized fast path only for non bigpcluster cases yet */ 288 if (ctx.inpages == 1 && ctx.outpages == 1 && !rq->inplace_io) { 289 DBG_BUGON(!*rq->out); 290 dst = kmap_atomic(*rq->out); 291 dst_maptype = 0; 292 goto dstmap_out; 293 } 294 295 /* general decoding path which can be used for all cases */ 296 ret = z_erofs_lz4_prepare_dstpages(&ctx, pagepool); 297 if (ret < 0) { 298 return ret; 299 } else if (ret > 0) { 300 dst = page_address(*rq->out); 301 dst_maptype = 1; 302 } else { 303 dst = erofs_vm_map_ram(rq->out, ctx.outpages); 304 if (!dst) 305 return -ENOMEM; 306 dst_maptype = 2; 307 } 308 309 dstmap_out: 310 ret = z_erofs_lz4_decompress_mem(&ctx, dst + rq->pageofs_out); 311 if (!dst_maptype) 312 kunmap_atomic(dst); 313 else if (dst_maptype == 2) 314 vm_unmap_ram(dst, ctx.outpages); 315 return ret; 316 } 317 318 static int z_erofs_shifted_transform(struct z_erofs_decompress_req *rq, 319 struct page **pagepool) 320 { 321 const unsigned int nrpages_out = 322 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT; 323 const unsigned int righthalf = min_t(unsigned int, rq->outputsize, 324 PAGE_SIZE - rq->pageofs_out); 325 unsigned char *src, *dst; 326 327 if (nrpages_out > 2) { 328 DBG_BUGON(1); 329 return -EIO; 330 } 331 332 if (rq->out[0] == *rq->in) { 333 DBG_BUGON(nrpages_out != 1); 334 return 0; 335 } 336 337 src = kmap_atomic(*rq->in) + rq->pageofs_in; 338 if (rq->out[0]) { 339 dst = kmap_atomic(rq->out[0]); 340 memcpy(dst + rq->pageofs_out, src, righthalf); 341 kunmap_atomic(dst); 342 } 343 344 if (nrpages_out == 2) { 345 DBG_BUGON(!rq->out[1]); 346 if (rq->out[1] == *rq->in) { 347 memmove(src, src + righthalf, rq->pageofs_out); 348 } else { 349 dst = kmap_atomic(rq->out[1]); 350 memcpy(dst, src + righthalf, rq->pageofs_out); 351 kunmap_atomic(dst); 352 } 353 } 354 kunmap_atomic(src); 355 return 0; 356 } 357 358 static struct z_erofs_decompressor decompressors[] = { 359 [Z_EROFS_COMPRESSION_SHIFTED] = { 360 .decompress = z_erofs_shifted_transform, 361 .name = "shifted" 362 }, 363 [Z_EROFS_COMPRESSION_LZ4] = { 364 .decompress = z_erofs_lz4_decompress, 365 .name = "lz4" 366 }, 367 #ifdef CONFIG_EROFS_FS_ZIP_LZMA 368 [Z_EROFS_COMPRESSION_LZMA] = { 369 .decompress = z_erofs_lzma_decompress, 370 .name = "lzma" 371 }, 372 #endif 373 }; 374 375 int z_erofs_decompress(struct z_erofs_decompress_req *rq, 376 struct page **pagepool) 377 { 378 return decompressors[rq->alg].decompress(rq, pagepool); 379 } 380