fs/erofs/decompressor.c

   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_decompressor {
  20         /*
  21          * if destpages have sparsed pages, fill them with bounce pages.
  22          * it also check whether destpages indicate continuous physical memory.
  23          */
  24         int (*prepare_destpages)(struct z_erofs_decompress_req *rq,
  25                                  struct list_head *pagepool);
  26         int (*decompress)(struct z_erofs_decompress_req *rq, u8 *out);
  27         char *name;
  28 };
  29
  30 int z_erofs_load_lz4_config(struct super_block *sb,
  31                             struct erofs_super_block *dsb,
  32                             struct z_erofs_lz4_cfgs *lz4, int size)
  33 {
  34         struct erofs_sb_info *sbi = EROFS_SB(sb);
  35         u16 distance;
  36
  37         if (lz4) {
  38                 if (size < sizeof(struct z_erofs_lz4_cfgs)) {
  39                         erofs_err(sb, "invalid lz4 cfgs, size=%u", size);
  40                         return -EINVAL;
  41                 }
  42                 distance = le16_to_cpu(lz4->max_distance);
  43
  44                 sbi->lz4.max_pclusterblks = le16_to_cpu(lz4->max_pclusterblks);
  45                 if (!sbi->lz4.max_pclusterblks) {
  46                         sbi->lz4.max_pclusterblks = 1;  /* reserved case */
  47                 } else if (sbi->lz4.max_pclusterblks >
  48                            Z_EROFS_PCLUSTER_MAX_SIZE / EROFS_BLKSIZ) {
  49                         erofs_err(sb, "too large lz4 pclusterblks %u",
  50                                   sbi->lz4.max_pclusterblks);
  51                         return -EINVAL;
  52                 } else if (sbi->lz4.max_pclusterblks >= 2) {
  53                         erofs_info(sb, "EXPERIMENTAL big pcluster feature in use. Use at your own risk!");
  54                 }
  55         } else {
  56                 distance = le16_to_cpu(dsb->u1.lz4_max_distance);
  57                 sbi->lz4.max_pclusterblks = 1;
  58         }
  59
  60         sbi->lz4.max_distance_pages = distance ?
  61                                         DIV_ROUND_UP(distance, PAGE_SIZE) + 1 :
  62                                         LZ4_MAX_DISTANCE_PAGES;
  63         return erofs_pcpubuf_growsize(sbi->lz4.max_pclusterblks);
  64 }
  65
  66 static int z_erofs_lz4_prepare_destpages(struct z_erofs_decompress_req *rq,
  67                                          struct list_head *pagepool)
  68 {
  69         const unsigned int nr =
  70                 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
  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 < nr; ++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 (!PageHighMem(page)) {
  97                                 if (!i) {
  98                                         kaddr = page_address(page);
  99                                         continue;
 100                                 }
 101                                 if (kaddr &&
 102                                     kaddr + PAGE_SIZE == page_address(page)) {
 103                                         kaddr += PAGE_SIZE;
 104                                         continue;
 105                                 }
 106                         }
 107                         kaddr = NULL;
 108                         continue;
 109                 }
 110                 kaddr = NULL;
 111                 __set_bit(j, bounced);
 112
 113                 if (top) {
 114                         victim = availables[--top];
 115                         get_page(victim);
 116                 } else {
 117                         victim = erofs_allocpage(pagepool,
 118                                                  GFP_KERNEL | __GFP_NOFAIL);
 119                         set_page_private(victim, Z_EROFS_SHORTLIVED_PAGE);
 120                 }
 121                 rq->out[i] = victim;
 122         }
 123         return kaddr ? 1 : 0;
 124 }
 125
 126 static void *z_erofs_handle_inplace_io(struct z_erofs_decompress_req *rq,
 127                         void *inpage, unsigned int *inputmargin, int *maptype,
 128                         bool support_0padding)
 129 {
 130         unsigned int nrpages_in, nrpages_out;
 131         unsigned int ofull, oend, inputsize, total, i, j;
 132         struct page **in;
 133         void *src, *tmp;
 134
 135         inputsize = rq->inputsize;
 136         nrpages_in = PAGE_ALIGN(inputsize) >> PAGE_SHIFT;
 137         oend = rq->pageofs_out + rq->outputsize;
 138         ofull = PAGE_ALIGN(oend);
 139         nrpages_out = ofull >> PAGE_SHIFT;
 140
 141         if (rq->inplace_io) {
 142                 if (rq->partial_decoding || !support_0padding ||
 143                     ofull - oend < LZ4_DECOMPRESS_INPLACE_MARGIN(inputsize))
 144                         goto docopy;
 145
 146                 for (i = 0; i < nrpages_in; ++i) {
 147                         DBG_BUGON(rq->in[i] == NULL);
 148                         for (j = 0; j < nrpages_out - nrpages_in + i; ++j)
 149                                 if (rq->out[j] == rq->in[i])
 150                                         goto docopy;
 151                 }
 152         }
 153
 154         if (nrpages_in <= 1) {
 155                 *maptype = 0;
 156                 return inpage;
 157         }
 158         kunmap_atomic(inpage);
 159         might_sleep();
 160         src = erofs_vm_map_ram(rq->in, nrpages_in);
 161         if (!src)
 162                 return ERR_PTR(-ENOMEM);
 163         *maptype = 1;
 164         return src;
 165
 166 docopy:
 167         /* Or copy compressed data which can be overlapped to per-CPU buffer */
 168         in = rq->in;
 169         src = erofs_get_pcpubuf(nrpages_in);
 170         if (!src) {
 171                 DBG_BUGON(1);
 172                 kunmap_atomic(inpage);
 173                 return ERR_PTR(-EFAULT);
 174         }
 175
 176         tmp = src;
 177         total = rq->inputsize;
 178         while (total) {
 179                 unsigned int page_copycnt =
 180                         min_t(unsigned int, total, PAGE_SIZE - *inputmargin);
 181
 182                 if (!inpage)
 183                         inpage = kmap_atomic(*in);
 184                 memcpy(tmp, inpage + *inputmargin, page_copycnt);
 185                 kunmap_atomic(inpage);
 186                 inpage = NULL;
 187                 tmp += page_copycnt;
 188                 total -= page_copycnt;
 189                 ++in;
 190                 *inputmargin = 0;
 191         }
 192         *maptype = 2;
 193         return src;
 194 }
 195
 196 static int z_erofs_lz4_decompress(struct z_erofs_decompress_req *rq, u8 *out)
 197 {
 198         unsigned int inputmargin;
 199         u8 *headpage, *src;
 200         bool support_0padding;
 201         int ret, maptype;
 202
 203         DBG_BUGON(*rq->in == NULL);
 204         headpage = kmap_atomic(*rq->in);
 205         inputmargin = 0;
 206         support_0padding = false;
 207
 208         /* decompression inplace is only safe when 0padding is enabled */
 209         if (erofs_sb_has_lz4_0padding(EROFS_SB(rq->sb))) {
 210                 support_0padding = true;
 211
 212                 while (!headpage[inputmargin & ~PAGE_MASK])
 213                         if (!(++inputmargin & ~PAGE_MASK))
 214                                 break;
 215
 216                 if (inputmargin >= rq->inputsize) {
 217                         kunmap_atomic(headpage);
 218                         return -EIO;
 219                 }
 220         }
 221
 222         rq->inputsize -= inputmargin;
 223         src = z_erofs_handle_inplace_io(rq, headpage, &inputmargin, &maptype,
 224                                         support_0padding);
 225         if (IS_ERR(src))
 226                 return PTR_ERR(src);
 227
 228         /* legacy format could compress extra data in a pcluster. */
 229         if (rq->partial_decoding || !support_0padding)
 230                 ret = LZ4_decompress_safe_partial(src + inputmargin, out,
 231                                 rq->inputsize, rq->outputsize, rq->outputsize);
 232         else
 233                 ret = LZ4_decompress_safe(src + inputmargin, out,
 234                                           rq->inputsize, rq->outputsize);
 235
 236         if (ret != rq->outputsize) {
 237                 erofs_err(rq->sb, "failed to decompress %d in[%u, %u] out[%u]",
 238                           ret, rq->inputsize, inputmargin, rq->outputsize);
 239
 240                 print_hex_dump(KERN_DEBUG, "[ in]: ", DUMP_PREFIX_OFFSET,
 241                                16, 1, src + inputmargin, rq->inputsize, true);
 242                 print_hex_dump(KERN_DEBUG, "[out]: ", DUMP_PREFIX_OFFSET,
 243                                16, 1, out, rq->outputsize, true);
 244
 245                 if (ret >= 0)
 246                         memset(out + ret, 0, rq->outputsize - ret);
 247                 ret = -EIO;
 248         }
 249
 250         if (maptype == 0) {
 251                 kunmap_atomic(src);
 252         } else if (maptype == 1) {
 253                 vm_unmap_ram(src, PAGE_ALIGN(rq->inputsize) >> PAGE_SHIFT);
 254         } else if (maptype == 2) {
 255                 erofs_put_pcpubuf(src);
 256         } else {
 257                 DBG_BUGON(1);
 258                 return -EFAULT;
 259         }
 260         return ret;
 261 }
 262
 263 static struct z_erofs_decompressor decompressors[] = {
 264         [Z_EROFS_COMPRESSION_SHIFTED] = {
 265                 .name = "shifted"
 266         },
 267         [Z_EROFS_COMPRESSION_LZ4] = {
 268                 .prepare_destpages = z_erofs_lz4_prepare_destpages,
 269                 .decompress = z_erofs_lz4_decompress,
 270                 .name = "lz4"
 271         },
 272 };
 273
 274 static void copy_from_pcpubuf(struct page **out, const char *dst,
 275                               unsigned short pageofs_out,
 276                               unsigned int outputsize)
 277 {
 278         const char *end = dst + outputsize;
 279         const unsigned int righthalf = PAGE_SIZE - pageofs_out;
 280         const char *cur = dst - pageofs_out;
 281
 282         while (cur < end) {
 283                 struct page *const page = *out++;
 284
 285                 if (page) {
 286                         char *buf = kmap_atomic(page);
 287
 288                         if (cur >= dst) {
 289                                 memcpy(buf, cur, min_t(uint, PAGE_SIZE,
 290                                                        end - cur));
 291                         } else {
 292                                 memcpy(buf + pageofs_out, cur + pageofs_out,
 293                                        min_t(uint, righthalf, end - cur));
 294                         }
 295                         kunmap_atomic(buf);
 296                 }
 297                 cur += PAGE_SIZE;
 298         }
 299 }
 300
 301 static int z_erofs_decompress_generic(struct z_erofs_decompress_req *rq,
 302                                       struct list_head *pagepool)
 303 {
 304         const unsigned int nrpages_out =
 305                 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
 306         const struct z_erofs_decompressor *alg = decompressors + rq->alg;
 307         unsigned int dst_maptype;
 308         void *dst;
 309         int ret;
 310
 311         /* two optimized fast paths only for non bigpcluster cases yet */
 312         if (rq->inputsize <= PAGE_SIZE) {
 313                 if (nrpages_out == 1 && !rq->inplace_io) {
 314                         DBG_BUGON(!*rq->out);
 315                         dst = kmap_atomic(*rq->out);
 316                         dst_maptype = 0;
 317                         goto dstmap_out;
 318                 }
 319
 320                 /*
 321                  * For the case of small output size (especially much less
 322                  * than PAGE_SIZE), memcpy the decompressed data rather than
 323                  * compressed data is preferred.
 324                  */
 325                 if (rq->outputsize <= PAGE_SIZE * 7 / 8) {
 326                         dst = erofs_get_pcpubuf(1);
 327                         if (IS_ERR(dst))
 328                                 return PTR_ERR(dst);
 329
 330                         rq->inplace_io = false;
 331                         ret = alg->decompress(rq, dst);
 332                         if (!ret)
 333                                 copy_from_pcpubuf(rq->out, dst, rq->pageofs_out,
 334                                                   rq->outputsize);
 335
 336                         erofs_put_pcpubuf(dst);
 337                         return ret;
 338                 }
 339         }
 340
 341         /* general decoding path which can be used for all cases */
 342         ret = alg->prepare_destpages(rq, pagepool);
 343         if (ret < 0)
 344                 return ret;
 345         if (ret) {
 346                 dst = page_address(*rq->out);
 347                 dst_maptype = 1;
 348                 goto dstmap_out;
 349         }
 350
 351         dst = erofs_vm_map_ram(rq->out, nrpages_out);
 352         if (!dst)
 353                 return -ENOMEM;
 354         dst_maptype = 2;
 355
 356 dstmap_out:
 357         ret = alg->decompress(rq, dst + rq->pageofs_out);
 358
 359         if (!dst_maptype)
 360                 kunmap_atomic(dst);
 361         else if (dst_maptype == 2)
 362                 vm_unmap_ram(dst, nrpages_out);
 363         return ret;
 364 }
 365
 366 static int z_erofs_shifted_transform(const struct z_erofs_decompress_req *rq,
 367                                      struct list_head *pagepool)
 368 {
 369         const unsigned int nrpages_out =
 370                 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
 371         const unsigned int righthalf = PAGE_SIZE - rq->pageofs_out;
 372         unsigned char *src, *dst;
 373
 374         if (nrpages_out > 2) {
 375                 DBG_BUGON(1);
 376                 return -EIO;
 377         }
 378
 379         if (rq->out[0] == *rq->in) {
 380                 DBG_BUGON(nrpages_out != 1);
 381                 return 0;
 382         }
 383
 384         src = kmap_atomic(*rq->in);
 385         if (rq->out[0]) {
 386                 dst = kmap_atomic(rq->out[0]);
 387                 memcpy(dst + rq->pageofs_out, src, righthalf);
 388                 kunmap_atomic(dst);
 389         }
 390
 391         if (nrpages_out == 2) {
 392                 DBG_BUGON(!rq->out[1]);
 393                 if (rq->out[1] == *rq->in) {
 394                         memmove(src, src + righthalf, rq->pageofs_out);
 395                 } else {
 396                         dst = kmap_atomic(rq->out[1]);
 397                         memcpy(dst, src + righthalf, rq->pageofs_out);
 398                         kunmap_atomic(dst);
 399                 }
 400         }
 401         kunmap_atomic(src);
 402         return 0;
 403 }
 404
 405 int z_erofs_decompress(struct z_erofs_decompress_req *rq,
 406                        struct list_head *pagepool)
 407 {
 408         if (rq->alg == Z_EROFS_COMPRESSION_SHIFTED)
 409                 return z_erofs_shifted_transform(rq, pagepool);
 410         return z_erofs_decompress_generic(rq, pagepool);
 411 }