1 // SPDX-License-Identifier: GPL-2.0
10 #include <linux/zlib.h>
11 #include <linux/zstd.h>
25 static struct bbuf __bounce_alloc(struct bch_fs *c, unsigned size, int rw)
29 BUG_ON(size > c->opts.encoded_extent_max);
31 b = kmalloc(size, GFP_NOIO|__GFP_NOWARN);
33 return (struct bbuf) { .b = b, .type = BB_KMALLOC, .rw = rw };
35 b = mempool_alloc(&c->compression_bounce[rw], GFP_NOIO);
37 return (struct bbuf) { .b = b, .type = BB_MEMPOOL, .rw = rw };
42 static bool bio_phys_contig(struct bio *bio, struct bvec_iter start)
45 struct bvec_iter iter;
46 void *expected_start = NULL;
48 __bio_for_each_bvec(bv, bio, iter, start) {
50 expected_start != page_address(bv.bv_page) + bv.bv_offset)
53 expected_start = page_address(bv.bv_page) +
54 bv.bv_offset + bv.bv_len;
60 static struct bbuf __bio_map_or_bounce(struct bch_fs *c, struct bio *bio,
61 struct bvec_iter start, int rw)
65 struct bvec_iter iter;
66 unsigned nr_pages = 0;
67 struct page *stack_pages[16];
68 struct page **pages = NULL;
71 BUG_ON(start.bi_size > c->opts.encoded_extent_max);
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
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 &&
87 if (bv.bv_len < iter.bi_size &&
88 bv.bv_offset + bv.bv_len < PAGE_SIZE)
94 BUG_ON(DIV_ROUND_UP(start.bi_size, PAGE_SIZE) > nr_pages);
96 pages = nr_pages > ARRAY_SIZE(stack_pages)
97 ? kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOIO)
103 __bio_for_each_segment(bv, bio, iter, start)
104 pages[nr_pages++] = bv.bv_page;
106 data = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
107 if (pages != stack_pages)
111 return (struct bbuf) {
112 .b = data + bio_iter_offset(bio, start),
113 .type = BB_VMAP, .rw = rw
116 ret = __bounce_alloc(c, start.bi_size, rw);
119 memcpy_from_bio(ret.b, bio, start);
124 static struct bbuf bio_map_or_bounce(struct bch_fs *c, struct bio *bio, int rw)
126 return __bio_map_or_bounce(c, bio, bio->bi_iter, rw);
129 static void bio_unmap_or_unbounce(struct bch_fs *c, struct bbuf buf)
135 vunmap((void *) ((unsigned long) buf.b & PAGE_MASK));
141 mempool_free(buf.b, &c->compression_bounce[buf.rw]);
146 static inline void zlib_set_workspace(z_stream *strm, void *workspace)
149 strm->workspace = workspace;
153 static int __bio_uncompress(struct bch_fs *c, struct bio *src,
154 void *dst_data, struct bch_extent_crc_unpacked crc)
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;
162 src_data = bio_map_or_bounce(c, src, READ);
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);
172 case BCH_COMPRESSION_TYPE_gzip: {
174 .next_in = src_data.b,
176 .next_out = dst_data,
177 .avail_out = dst_len,
180 workspace = mempool_alloc(&c->decompress_workspace, GFP_NOIO);
182 zlib_set_workspace(&strm, workspace);
183 zlib_inflateInit2(&strm, -MAX_WBITS);
184 ret = zlib_inflate(&strm, Z_FINISH);
186 mempool_free(workspace, &c->decompress_workspace);
188 if (ret != Z_STREAM_END)
192 case BCH_COMPRESSION_TYPE_zstd: {
194 size_t real_src_len = le32_to_cpup(src_data.b);
196 if (real_src_len > src_len - 4)
199 workspace = mempool_alloc(&c->decompress_workspace, GFP_NOIO);
200 ctx = zstd_init_dctx(workspace, zstd_dctx_workspace_bound());
202 src_len = le32_to_cpup(src_data.b);
204 ret = zstd_decompress_dctx(ctx,
206 src_data.b + 4, real_src_len);
208 mempool_free(workspace, &c->decompress_workspace);
219 bio_unmap_or_unbounce(c, src_data);
226 int bch2_bio_uncompress_inplace(struct bch_fs *c, struct bio *bio,
227 struct bch_extent_crc_unpacked *crc)
229 struct bbuf data = { NULL };
230 size_t dst_len = crc->uncompressed_size << 9;
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);
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");
242 data = __bounce_alloc(c, dst_len, WRITE);
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);
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
254 bio->bi_iter.bi_size = crc->live_size << 9;
256 memcpy_to_bio(bio, bio->bi_iter, data.b + (crc->offset << 9));
259 crc->compression_type = 0;
260 crc->compressed_size = crc->live_size;
261 crc->uncompressed_size = crc->live_size;
263 crc->csum = (struct bch_csum) { 0, 0 };
265 bio_unmap_or_unbounce(c, data);
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)
273 struct bbuf dst_data = { NULL };
274 size_t dst_len = crc.uncompressed_size << 9;
277 if (crc.uncompressed_size << 9 > c->opts.encoded_extent_max ||
278 crc.compressed_size << 9 > c->opts.encoded_extent_max)
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);
285 ret = __bio_uncompress(c, src, dst_data.b, crc);
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));
293 bio_unmap_or_unbounce(c, dst_data);
297 static int attempt_compress(struct bch_fs *c,
299 void *dst, size_t dst_len,
300 void *src, size_t src_len,
301 enum bch_compression_type compression_type)
303 switch (compression_type) {
304 case BCH_COMPRESSION_TYPE_lz4: {
306 int ret = LZ4_compress_destSize(
316 case BCH_COMPRESSION_TYPE_gzip: {
321 .avail_out = dst_len,
324 zlib_set_workspace(&strm, workspace);
325 zlib_deflateInit2(&strm, Z_DEFAULT_COMPRESSION,
326 Z_DEFLATED, -MAX_WBITS, DEF_MEM_LEVEL,
329 if (zlib_deflate(&strm, Z_FINISH) != Z_STREAM_END)
332 if (zlib_deflateEnd(&strm) != Z_OK)
335 return strm.total_out;
337 case BCH_COMPRESSION_TYPE_zstd: {
338 ZSTD_CCtx *ctx = zstd_init_cctx(workspace,
339 zstd_cctx_workspace_bound(&c->zstd_params.cParams));
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
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
352 size_t len = zstd_compress_cctx(ctx,
353 dst + 4, dst_len - 4 - 7,
356 if (zstd_is_error(len))
359 *((__le32 *) dst) = cpu_to_le32(len);
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)
372 struct bbuf src_data = { NULL }, dst_data = { NULL };
377 BUG_ON(compression_type >= BCH_COMPRESSION_TYPE_NR);
378 BUG_ON(!mempool_initialized(&c->compress_workspace[compression_type]));
380 /* If it's only one block, don't bother trying to compress: */
381 if (src->bi_iter.bi_size <= c->opts.block_size)
384 dst_data = bio_map_or_bounce(c, dst, WRITE);
385 src_data = bio_map_or_bounce(c, src, READ);
387 workspace = mempool_alloc(&c->compress_workspace[compression_type], GFP_NOIO);
389 *src_len = src->bi_iter.bi_size;
390 *dst_len = dst->bi_iter.bi_size;
393 * XXX: this algorithm sucks when the compression code doesn't tell us
394 * how much would fit, like LZ4 does:
397 if (*src_len <= block_bytes(c)) {
402 ret = attempt_compress(c, workspace,
403 dst_data.b, *dst_len,
404 src_data.b, *src_len,
412 /* Didn't fit: should we retry with a smaller amount? */
413 if (*src_len <= *dst_len) {
419 * If ret is negative, it's a hint as to how much data would fit
421 BUG_ON(-ret >= *src_len);
426 *src_len -= (*src_len - *dst_len) / 2;
427 *src_len = round_down(*src_len, block_bytes(c));
430 mempool_free(workspace, &c->compress_workspace[compression_type]);
435 /* Didn't get smaller: */
436 if (round_up(*dst_len, block_bytes(c)) >= *src_len)
439 pad = round_up(*dst_len, block_bytes(c)) - *dst_len;
441 memset(dst_data.b + *dst_len, 0, pad);
444 if (dst_data.type != BB_NONE &&
445 dst_data.type != BB_VMAP)
446 memcpy_to_bio(dst, dst->bi_iter, dst_data.b);
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));
453 bio_unmap_or_unbounce(c, src_data);
454 bio_unmap_or_unbounce(c, dst_data);
455 return compression_type;
457 compression_type = BCH_COMPRESSION_TYPE_incompressible;
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)
466 unsigned orig_dst = dst->bi_iter.bi_size;
467 unsigned orig_src = src->bi_iter.bi_size;
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);
475 if (compression_type == BCH_COMPRESSION_TYPE_lz4_old)
476 compression_type = BCH_COMPRESSION_TYPE_lz4;
479 __bio_compress(c, dst, dst_len, src, src_len, compression_type);
481 dst->bi_iter.bi_size = orig_dst;
482 src->bi_iter.bi_size = orig_src;
483 return compression_type;
486 static int __bch2_fs_compress_init(struct bch_fs *, u64);
488 #define BCH_FEATURE_none 0
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()
496 #undef BCH_FEATURE_none
498 static int __bch2_check_set_has_compressed_data(struct bch_fs *c, u64 f)
502 if ((c->sb.features & f) == f)
505 mutex_lock(&c->sb_lock);
507 if ((c->sb.features & f) == f) {
508 mutex_unlock(&c->sb_lock);
512 ret = __bch2_fs_compress_init(c, c->sb.features|f);
514 mutex_unlock(&c->sb_lock);
518 c->disk_sb.sb->features[0] |= cpu_to_le64(f);
520 mutex_unlock(&c->sb_lock);
525 int bch2_check_set_has_compressed_data(struct bch_fs *c,
526 unsigned compression_type)
528 BUG_ON(compression_type >= ARRAY_SIZE(bch2_compression_opt_to_feature));
530 return compression_type
531 ? __bch2_check_set_has_compressed_data(c,
532 1ULL << bch2_compression_opt_to_feature[compression_type])
536 void bch2_fs_compress_exit(struct bch_fs *c)
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]);
547 static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
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);
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(¶ms.cParams),
564 zstd_dctx_workspace_bound() },
568 pr_verbose_init(c->opts, "");
570 c->zstd_params = params;
572 for (i = compression_types;
573 i < compression_types + ARRAY_SIZE(compression_types);
575 if (features & (1 << i->feature))
576 goto have_compressed;
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);
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);
595 for (i = compression_types;
596 i < compression_types + ARRAY_SIZE(compression_types);
598 decompress_workspace_size =
599 max(decompress_workspace_size, i->decompress_workspace);
601 if (!(features & (1 << i->feature)))
604 if (i->decompress_workspace)
605 decompress_workspace_needed = true;
607 if (mempool_initialized(&c->compress_workspace[i->type]))
610 ret = mempool_init_kvpmalloc_pool(
611 &c->compress_workspace[i->type],
612 1, i->compress_workspace);
617 if (!mempool_initialized(&c->decompress_workspace)) {
618 ret = mempool_init_kvpmalloc_pool(
619 &c->decompress_workspace,
620 1, decompress_workspace_size);
625 pr_verbose_init(c->opts, "ret %i", ret);
629 int bch2_fs_compress_init(struct bch_fs *c)
631 u64 f = c->sb.features;
633 if (c->opts.compression)
634 f |= 1ULL << bch2_compression_opt_to_feature[c->opts.compression];
636 if (c->opts.background_compression)
637 f |= 1ULL << bch2_compression_opt_to_feature[c->opts.background_compression];
639 return __bch2_fs_compress_init(c, f);