git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/compress.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "checksum.h"
   4 #include "compress.h"
   5 #include "extents.h"
   6 #include "io.h"
   7 #include "super-io.h"
   8
   9 #include <linux/lz4.h>
  10 #include <linux/zlib.h>
  11 #include <linux/zstd.h>
  12
  13 /* Bounce buffer: */
  14 struct bbuf {
  15         void            *b;
  16         enum {
  17                 BB_NONE,
  18                 BB_VMAP,
  19                 BB_KMALLOC,
  20                 BB_MEMPOOL,
  21         }               type;
  22         int             rw;
  23 };
  24
  25 static struct bbuf __bounce_alloc(struct bch_fs *c, unsigned size, int rw)
  26 {
  27         void *b;
  28
  29         BUG_ON(size > c->opts.encoded_extent_max);
  30
  31         b = kmalloc(size, GFP_NOIO|__GFP_NOWARN);
  32         if (b)
  33                 return (struct bbuf) { .b = b, .type = BB_KMALLOC, .rw = rw };
  34
  35         b = mempool_alloc(&c->compression_bounce[rw], GFP_NOIO);
  36         if (b)
  37                 return (struct bbuf) { .b = b, .type = BB_MEMPOOL, .rw = rw };
  38
  39         BUG();
  40 }
  41
  42 static bool bio_phys_contig(struct bio *bio, struct bvec_iter start)
  43 {
  44         struct bio_vec bv;
  45         struct bvec_iter iter;
  46         void *expected_start = NULL;
  47
  48         __bio_for_each_bvec(bv, bio, iter, start) {
  49                 if (expected_start &&
  50                     expected_start != page_address(bv.bv_page) + bv.bv_offset)
  51                         return false;
  52
  53                 expected_start = page_address(bv.bv_page) +
  54                         bv.bv_offset + bv.bv_len;
  55         }
  56
  57         return true;
  58 }
  59
  60 static struct bbuf __bio_map_or_bounce(struct bch_fs *c, struct bio *bio,
  61                                        struct bvec_iter start, int rw)
  62 {
  63         struct bbuf ret;
  64         struct bio_vec bv;
  65         struct bvec_iter iter;
  66         unsigned nr_pages = 0;
  67         struct page *stack_pages[16];
  68         struct page **pages = NULL;
  69         void *data;
  70
  71         BUG_ON(start.bi_size > c->opts.encoded_extent_max);
  72
  73         if (!PageHighMem(bio_iter_page(bio, start)) &&
  74             bio_phys_contig(bio, start))
  75                 return (struct bbuf) {
  76                         .b = page_address(bio_iter_page(bio, start)) +
  77                                 bio_iter_offset(bio, start),
  78                         .type = BB_NONE, .rw = rw
  79                 };
  80
  81         /* check if we can map the pages contiguously: */
  82         __bio_for_each_segment(bv, bio, iter, start) {
  83                 if (iter.bi_size != start.bi_size &&
  84                     bv.bv_offset)
  85                         goto bounce;
  86
  87                 if (bv.bv_len < iter.bi_size &&
  88                     bv.bv_offset + bv.bv_len < PAGE_SIZE)
  89                         goto bounce;
  90
  91                 nr_pages++;
  92         }
  93
  94         BUG_ON(DIV_ROUND_UP(start.bi_size, PAGE_SIZE) > nr_pages);
  95
  96         pages = nr_pages > ARRAY_SIZE(stack_pages)
  97                 ? kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOIO)
  98                 : stack_pages;
  99         if (!pages)
 100                 goto bounce;
 101
 102         nr_pages = 0;
 103         __bio_for_each_segment(bv, bio, iter, start)
 104                 pages[nr_pages++] = bv.bv_page;
 105
 106         data = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
 107         if (pages != stack_pages)
 108                 kfree(pages);
 109
 110         if (data)
 111                 return (struct bbuf) {
 112                         .b = data + bio_iter_offset(bio, start),
 113                         .type = BB_VMAP, .rw = rw
 114                 };
 115 bounce:
 116         ret = __bounce_alloc(c, start.bi_size, rw);
 117
 118         if (rw == READ)
 119                 memcpy_from_bio(ret.b, bio, start);
 120
 121         return ret;
 122 }
 123
 124 static struct bbuf bio_map_or_bounce(struct bch_fs *c, struct bio *bio, int rw)
 125 {
 126         return __bio_map_or_bounce(c, bio, bio->bi_iter, rw);
 127 }
 128
 129 static void bio_unmap_or_unbounce(struct bch_fs *c, struct bbuf buf)
 130 {
 131         switch (buf.type) {
 132         case BB_NONE:
 133                 break;
 134         case BB_VMAP:
 135                 vunmap((void *) ((unsigned long) buf.b & PAGE_MASK));
 136                 break;
 137         case BB_KMALLOC:
 138                 kfree(buf.b);
 139                 break;
 140         case BB_MEMPOOL:
 141                 mempool_free(buf.b, &c->compression_bounce[buf.rw]);
 142                 break;
 143         }
 144 }
 145
 146 static inline void zlib_set_workspace(z_stream *strm, void *workspace)
 147 {
 148 #ifdef __KERNEL__
 149         strm->workspace = workspace;
 150 #endif
 151 }
 152
 153 static int __bio_uncompress(struct bch_fs *c, struct bio *src,
 154                             void *dst_data, struct bch_extent_crc_unpacked crc)
 155 {
 156         struct bbuf src_data = { NULL };
 157         size_t src_len = src->bi_iter.bi_size;
 158         size_t dst_len = crc.uncompressed_size << 9;
 159         void *workspace;
 160         int ret;
 161
 162         src_data = bio_map_or_bounce(c, src, READ);
 163
 164         switch (crc.compression_type) {
 165         case BCH_COMPRESSION_TYPE_lz4_old:
 166         case BCH_COMPRESSION_TYPE_lz4:
 167                 ret = LZ4_decompress_safe_partial(src_data.b, dst_data,
 168                                                   src_len, dst_len, dst_len);
 169                 if (ret != dst_len)
 170                         goto err;
 171                 break;
 172         case BCH_COMPRESSION_TYPE_gzip: {
 173                 z_stream strm = {
 174                         .next_in        = src_data.b,
 175                         .avail_in       = src_len,
 176                         .next_out       = dst_data,
 177                         .avail_out      = dst_len,
 178                 };
 179
 180                 workspace = mempool_alloc(&c->decompress_workspace, GFP_NOIO);
 181
 182                 zlib_set_workspace(&strm, workspace);
 183                 zlib_inflateInit2(&strm, -MAX_WBITS);
 184                 ret = zlib_inflate(&strm, Z_FINISH);
 185
 186                 mempool_free(workspace, &c->decompress_workspace);
 187
 188                 if (ret != Z_STREAM_END)
 189                         goto err;
 190                 break;
 191         }
 192         case BCH_COMPRESSION_TYPE_zstd: {
 193                 ZSTD_DCtx *ctx;
 194                 size_t real_src_len = le32_to_cpup(src_data.b);
 195
 196                 if (real_src_len > src_len - 4)
 197                         goto err;
 198
 199                 workspace = mempool_alloc(&c->decompress_workspace, GFP_NOIO);
 200                 ctx = zstd_init_dctx(workspace, zstd_dctx_workspace_bound());
 201
 202                 src_len = le32_to_cpup(src_data.b);
 203
 204                 ret = zstd_decompress_dctx(ctx,
 205                                 dst_data,       dst_len,
 206                                 src_data.b + 4, real_src_len);
 207
 208                 mempool_free(workspace, &c->decompress_workspace);
 209
 210                 if (ret != dst_len)
 211                         goto err;
 212                 break;
 213         }
 214         default:
 215                 BUG();
 216         }
 217         ret = 0;
 218 out:
 219         bio_unmap_or_unbounce(c, src_data);
 220         return ret;
 221 err:
 222         ret = -EIO;
 223         goto out;
 224 }
 225
 226 int bch2_bio_uncompress_inplace(struct bch_fs *c, struct bio *bio,
 227                                 struct bch_extent_crc_unpacked *crc)
 228 {
 229         struct bbuf data = { NULL };
 230         size_t dst_len = crc->uncompressed_size << 9;
 231
 232         /* bio must own its pages: */
 233         BUG_ON(!bio->bi_vcnt);
 234         BUG_ON(DIV_ROUND_UP(crc->live_size, PAGE_SECTORS) > bio->bi_max_vecs);
 235
 236         if (crc->uncompressed_size << 9 > c->opts.encoded_extent_max ||
 237             crc->compressed_size << 9   > c->opts.encoded_extent_max) {
 238                 bch_err(c, "error rewriting existing data: extent too big");
 239                 return -EIO;
 240         }
 241
 242         data = __bounce_alloc(c, dst_len, WRITE);
 243
 244         if (__bio_uncompress(c, bio, data.b, *crc)) {
 245                 bch_err(c, "error rewriting existing data: decompression error");
 246                 bio_unmap_or_unbounce(c, data);
 247                 return -EIO;
 248         }
 249
 250         /*
 251          * XXX: don't have a good way to assert that the bio was allocated with
 252          * enough space, we depend on bch2_move_extent doing the right thing
 253          */
 254         bio->bi_iter.bi_size = crc->live_size << 9;
 255
 256         memcpy_to_bio(bio, bio->bi_iter, data.b + (crc->offset << 9));
 257
 258         crc->csum_type          = 0;
 259         crc->compression_type   = 0;
 260         crc->compressed_size    = crc->live_size;
 261         crc->uncompressed_size  = crc->live_size;
 262         crc->offset             = 0;
 263         crc->csum               = (struct bch_csum) { 0, 0 };
 264
 265         bio_unmap_or_unbounce(c, data);
 266         return 0;
 267 }
 268
 269 int bch2_bio_uncompress(struct bch_fs *c, struct bio *src,
 270                        struct bio *dst, struct bvec_iter dst_iter,
 271                        struct bch_extent_crc_unpacked crc)
 272 {
 273         struct bbuf dst_data = { NULL };
 274         size_t dst_len = crc.uncompressed_size << 9;
 275         int ret = -ENOMEM;
 276
 277         if (crc.uncompressed_size << 9  > c->opts.encoded_extent_max ||
 278             crc.compressed_size << 9    > c->opts.encoded_extent_max)
 279                 return -EIO;
 280
 281         dst_data = dst_len == dst_iter.bi_size
 282                 ? __bio_map_or_bounce(c, dst, dst_iter, WRITE)
 283                 : __bounce_alloc(c, dst_len, WRITE);
 284
 285         ret = __bio_uncompress(c, src, dst_data.b, crc);
 286         if (ret)
 287                 goto err;
 288
 289         if (dst_data.type != BB_NONE &&
 290             dst_data.type != BB_VMAP)
 291                 memcpy_to_bio(dst, dst_iter, dst_data.b + (crc.offset << 9));
 292 err:
 293         bio_unmap_or_unbounce(c, dst_data);
 294         return ret;
 295 }
 296
 297 static int attempt_compress(struct bch_fs *c,
 298                             void *workspace,
 299                             void *dst, size_t dst_len,
 300                             void *src, size_t src_len,
 301                             enum bch_compression_type compression_type)
 302 {
 303         switch (compression_type) {
 304         case BCH_COMPRESSION_TYPE_lz4: {
 305                 int len = src_len;
 306                 int ret = LZ4_compress_destSize(
 307                                 src,            dst,
 308                                 &len,           dst_len,
 309                                 workspace);
 310
 311                 if (len < src_len)
 312                         return -len;
 313
 314                 return ret;
 315         }
 316         case BCH_COMPRESSION_TYPE_gzip: {
 317                 z_stream strm = {
 318                         .next_in        = src,
 319                         .avail_in       = src_len,
 320                         .next_out       = dst,
 321                         .avail_out      = dst_len,
 322                 };
 323
 324                 zlib_set_workspace(&strm, workspace);
 325                 zlib_deflateInit2(&strm, Z_DEFAULT_COMPRESSION,
 326                                   Z_DEFLATED, -MAX_WBITS, DEF_MEM_LEVEL,
 327                                   Z_DEFAULT_STRATEGY);
 328
 329                 if (zlib_deflate(&strm, Z_FINISH) != Z_STREAM_END)
 330                         return 0;
 331
 332                 if (zlib_deflateEnd(&strm) != Z_OK)
 333                         return 0;
 334
 335                 return strm.total_out;
 336         }
 337         case BCH_COMPRESSION_TYPE_zstd: {
 338                 ZSTD_CCtx *ctx = zstd_init_cctx(workspace,
 339                         zstd_cctx_workspace_bound(&c->zstd_params.cParams));
 340
 341                 /*
 342                  * ZSTD requires that when we decompress we pass in the exact
 343                  * compressed size - rounding it up to the nearest sector
 344                  * doesn't work, so we use the first 4 bytes of the buffer for
 345                  * that.
 346                  *
 347                  * Additionally, the ZSTD code seems to have a bug where it will
 348                  * write just past the end of the buffer - so subtract a fudge
 349                  * factor (7 bytes) from the dst buffer size to account for
 350                  * that.
 351                  */
 352                 size_t len = zstd_compress_cctx(ctx,
 353                                 dst + 4,        dst_len - 4 - 7,
 354                                 src,            src_len,
 355                                 &c->zstd_params);
 356                 if (zstd_is_error(len))
 357                         return 0;
 358
 359                 *((__le32 *) dst) = cpu_to_le32(len);
 360                 return len + 4;
 361         }
 362         default:
 363                 BUG();
 364         }
 365 }
 366
 367 static unsigned __bio_compress(struct bch_fs *c,
 368                                struct bio *dst, size_t *dst_len,
 369                                struct bio *src, size_t *src_len,
 370                                enum bch_compression_type compression_type)
 371 {
 372         struct bbuf src_data = { NULL }, dst_data = { NULL };
 373         void *workspace;
 374         unsigned pad;
 375         int ret = 0;
 376
 377         BUG_ON(compression_type >= BCH_COMPRESSION_TYPE_NR);
 378         BUG_ON(!mempool_initialized(&c->compress_workspace[compression_type]));
 379
 380         /* If it's only one block, don't bother trying to compress: */
 381         if (src->bi_iter.bi_size <= c->opts.block_size)
 382                 return 0;
 383
 384         dst_data = bio_map_or_bounce(c, dst, WRITE);
 385         src_data = bio_map_or_bounce(c, src, READ);
 386
 387         workspace = mempool_alloc(&c->compress_workspace[compression_type], GFP_NOIO);
 388
 389         *src_len = src->bi_iter.bi_size;
 390         *dst_len = dst->bi_iter.bi_size;
 391
 392         /*
 393          * XXX: this algorithm sucks when the compression code doesn't tell us
 394          * how much would fit, like LZ4 does:
 395          */
 396         while (1) {
 397                 if (*src_len <= block_bytes(c)) {
 398                         ret = -1;
 399                         break;
 400                 }
 401
 402                 ret = attempt_compress(c, workspace,
 403                                        dst_data.b,      *dst_len,
 404                                        src_data.b,      *src_len,
 405                                        compression_type);
 406                 if (ret > 0) {
 407                         *dst_len = ret;
 408                         ret = 0;
 409                         break;
 410                 }
 411
 412                 /* Didn't fit: should we retry with a smaller amount?  */
 413                 if (*src_len <= *dst_len) {
 414                         ret = -1;
 415                         break;
 416                 }
 417
 418                 /*
 419                  * If ret is negative, it's a hint as to how much data would fit
 420                  */
 421                 BUG_ON(-ret >= *src_len);
 422
 423                 if (ret < 0)
 424                         *src_len = -ret;
 425                 else
 426                         *src_len -= (*src_len - *dst_len) / 2;
 427                 *src_len = round_down(*src_len, block_bytes(c));
 428         }
 429
 430         mempool_free(workspace, &c->compress_workspace[compression_type]);
 431
 432         if (ret)
 433                 goto err;
 434
 435         /* Didn't get smaller: */
 436         if (round_up(*dst_len, block_bytes(c)) >= *src_len)
 437                 goto err;
 438
 439         pad = round_up(*dst_len, block_bytes(c)) - *dst_len;
 440
 441         memset(dst_data.b + *dst_len, 0, pad);
 442         *dst_len += pad;
 443
 444         if (dst_data.type != BB_NONE &&
 445             dst_data.type != BB_VMAP)
 446                 memcpy_to_bio(dst, dst->bi_iter, dst_data.b);
 447
 448         BUG_ON(!*dst_len || *dst_len > dst->bi_iter.bi_size);
 449         BUG_ON(!*src_len || *src_len > src->bi_iter.bi_size);
 450         BUG_ON(*dst_len & (block_bytes(c) - 1));
 451         BUG_ON(*src_len & (block_bytes(c) - 1));
 452 out:
 453         bio_unmap_or_unbounce(c, src_data);
 454         bio_unmap_or_unbounce(c, dst_data);
 455         return compression_type;
 456 err:
 457         compression_type = BCH_COMPRESSION_TYPE_incompressible;
 458         goto out;
 459 }
 460
 461 unsigned bch2_bio_compress(struct bch_fs *c,
 462                            struct bio *dst, size_t *dst_len,
 463                            struct bio *src, size_t *src_len,
 464                            unsigned compression_type)
 465 {
 466         unsigned orig_dst = dst->bi_iter.bi_size;
 467         unsigned orig_src = src->bi_iter.bi_size;
 468
 469         /* Don't consume more than BCH_ENCODED_EXTENT_MAX from @src: */
 470         src->bi_iter.bi_size = min_t(unsigned, src->bi_iter.bi_size,
 471                                      c->opts.encoded_extent_max);
 472         /* Don't generate a bigger output than input: */
 473         dst->bi_iter.bi_size = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 474
 475         if (compression_type == BCH_COMPRESSION_TYPE_lz4_old)
 476                 compression_type = BCH_COMPRESSION_TYPE_lz4;
 477
 478         compression_type =
 479                 __bio_compress(c, dst, dst_len, src, src_len, compression_type);
 480
 481         dst->bi_iter.bi_size = orig_dst;
 482         src->bi_iter.bi_size = orig_src;
 483         return compression_type;
 484 }
 485
 486 static int __bch2_fs_compress_init(struct bch_fs *, u64);
 487
 488 #define BCH_FEATURE_none        0
 489
 490 static const unsigned bch2_compression_opt_to_feature[] = {
 491 #define x(t, n) [BCH_COMPRESSION_OPT_##t] = BCH_FEATURE_##t,
 492         BCH_COMPRESSION_OPTS()
 493 #undef x
 494 };
 495
 496 #undef BCH_FEATURE_none
 497
 498 static int __bch2_check_set_has_compressed_data(struct bch_fs *c, u64 f)
 499 {
 500         int ret = 0;
 501
 502         if ((c->sb.features & f) == f)
 503                 return 0;
 504
 505         mutex_lock(&c->sb_lock);
 506
 507         if ((c->sb.features & f) == f) {
 508                 mutex_unlock(&c->sb_lock);
 509                 return 0;
 510         }
 511
 512         ret = __bch2_fs_compress_init(c, c->sb.features|f);
 513         if (ret) {
 514                 mutex_unlock(&c->sb_lock);
 515                 return ret;
 516         }
 517
 518         c->disk_sb.sb->features[0] |= cpu_to_le64(f);
 519         bch2_write_super(c);
 520         mutex_unlock(&c->sb_lock);
 521
 522         return 0;
 523 }
 524
 525 int bch2_check_set_has_compressed_data(struct bch_fs *c,
 526                                        unsigned compression_type)
 527 {
 528         BUG_ON(compression_type >= ARRAY_SIZE(bch2_compression_opt_to_feature));
 529
 530         return compression_type
 531                 ? __bch2_check_set_has_compressed_data(c,
 532                                 1ULL << bch2_compression_opt_to_feature[compression_type])
 533                 : 0;
 534 }
 535
 536 void bch2_fs_compress_exit(struct bch_fs *c)
 537 {
 538         unsigned i;
 539
 540         mempool_exit(&c->decompress_workspace);
 541         for (i = 0; i < ARRAY_SIZE(c->compress_workspace); i++)
 542                 mempool_exit(&c->compress_workspace[i]);
 543         mempool_exit(&c->compression_bounce[WRITE]);
 544         mempool_exit(&c->compression_bounce[READ]);
 545 }
 546
 547 static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
 548 {
 549         size_t decompress_workspace_size = 0;
 550         bool decompress_workspace_needed;
 551         ZSTD_parameters params = zstd_get_params(0, c->opts.encoded_extent_max);
 552         struct {
 553                 unsigned        feature;
 554                 unsigned        type;
 555                 size_t          compress_workspace;
 556                 size_t          decompress_workspace;
 557         } compression_types[] = {
 558                 { BCH_FEATURE_lz4, BCH_COMPRESSION_TYPE_lz4, LZ4_MEM_COMPRESS, 0 },
 559                 { BCH_FEATURE_gzip, BCH_COMPRESSION_TYPE_gzip,
 560                         zlib_deflate_workspacesize(MAX_WBITS, DEF_MEM_LEVEL),
 561                         zlib_inflate_workspacesize(), },
 562                 { BCH_FEATURE_zstd, BCH_COMPRESSION_TYPE_zstd,
 563                         zstd_cctx_workspace_bound(&params.cParams),
 564                         zstd_dctx_workspace_bound() },
 565         }, *i;
 566         int ret = 0;
 567
 568         pr_verbose_init(c->opts, "");
 569
 570         c->zstd_params = params;
 571
 572         for (i = compression_types;
 573              i < compression_types + ARRAY_SIZE(compression_types);
 574              i++)
 575                 if (features & (1 << i->feature))
 576                         goto have_compressed;
 577
 578         goto out;
 579 have_compressed:
 580
 581         if (!mempool_initialized(&c->compression_bounce[READ])) {
 582                 ret = mempool_init_kvpmalloc_pool(&c->compression_bounce[READ],
 583                                                   1, c->opts.encoded_extent_max);
 584                 if (ret)
 585                         goto out;
 586         }
 587
 588         if (!mempool_initialized(&c->compression_bounce[WRITE])) {
 589                 ret = mempool_init_kvpmalloc_pool(&c->compression_bounce[WRITE],
 590                                                   1, c->opts.encoded_extent_max);
 591                 if (ret)
 592                         goto out;
 593         }
 594
 595         for (i = compression_types;
 596              i < compression_types + ARRAY_SIZE(compression_types);
 597              i++) {
 598                 decompress_workspace_size =
 599                         max(decompress_workspace_size, i->decompress_workspace);
 600
 601                 if (!(features & (1 << i->feature)))
 602                         continue;
 603
 604                 if (i->decompress_workspace)
 605                         decompress_workspace_needed = true;
 606
 607                 if (mempool_initialized(&c->compress_workspace[i->type]))
 608                         continue;
 609
 610                 ret = mempool_init_kvpmalloc_pool(
 611                                 &c->compress_workspace[i->type],
 612                                 1, i->compress_workspace);
 613                 if (ret)
 614                         goto out;
 615         }
 616
 617         if (!mempool_initialized(&c->decompress_workspace)) {
 618                 ret = mempool_init_kvpmalloc_pool(
 619                                 &c->decompress_workspace,
 620                                 1, decompress_workspace_size);
 621                 if (ret)
 622                         goto out;
 623         }
 624 out:
 625         pr_verbose_init(c->opts, "ret %i", ret);
 626         return ret;
 627 }
 628
 629 int bch2_fs_compress_init(struct bch_fs *c)
 630 {
 631         u64 f = c->sb.features;
 632
 633         if (c->opts.compression)
 634                 f |= 1ULL << bch2_compression_opt_to_feature[c->opts.compression];
 635
 636         if (c->opts.background_compression)
 637                 f |= 1ULL << bch2_compression_opt_to_feature[c->opts.background_compression];
 638
 639         return __bch2_fs_compress_init(c, f);
 640
 641 }