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