brintos

brintos / linux-shallow public Read only

0
0
Text · 17.6 KiB · 1410365 Raw
729 lines · c
1// SPDX-License-Identifier: GPL-2.02#include "bcachefs.h"3#include "checksum.h"4#include "compress.h"5#include "extents.h"6#include "super-io.h"7 8#include <linux/lz4.h>9#include <linux/zlib.h>10#include <linux/zstd.h>11 12/* Bounce buffer: */13struct bbuf {14	void		*b;15	enum {16		BB_NONE,17		BB_VMAP,18		BB_KMALLOC,19		BB_MEMPOOL,20	}		type;21	int		rw;22};23 24static struct bbuf __bounce_alloc(struct bch_fs *c, unsigned size, int rw)25{26	void *b;27 28	BUG_ON(size > c->opts.encoded_extent_max);29 30	b = kmalloc(size, GFP_NOFS|__GFP_NOWARN);31	if (b)32		return (struct bbuf) { .b = b, .type = BB_KMALLOC, .rw = rw };33 34	b = mempool_alloc(&c->compression_bounce[rw], GFP_NOFS);35	if (b)36		return (struct bbuf) { .b = b, .type = BB_MEMPOOL, .rw = rw };37 38	BUG();39}40 41static bool bio_phys_contig(struct bio *bio, struct bvec_iter start)42{43	struct bio_vec bv;44	struct bvec_iter iter;45	void *expected_start = NULL;46 47	__bio_for_each_bvec(bv, bio, iter, start) {48		if (expected_start &&49		    expected_start != page_address(bv.bv_page) + bv.bv_offset)50			return false;51 52		expected_start = page_address(bv.bv_page) +53			bv.bv_offset + bv.bv_len;54	}55 56	return true;57}58 59static struct bbuf __bio_map_or_bounce(struct bch_fs *c, struct bio *bio,60				       struct bvec_iter start, int rw)61{62	struct bbuf ret;63	struct bio_vec bv;64	struct bvec_iter iter;65	unsigned nr_pages = 0;66	struct page *stack_pages[16];67	struct page **pages = NULL;68	void *data;69 70	BUG_ON(start.bi_size > c->opts.encoded_extent_max);71 72	if (!PageHighMem(bio_iter_page(bio, start)) &&73	    bio_phys_contig(bio, start))74		return (struct bbuf) {75			.b = page_address(bio_iter_page(bio, start)) +76				bio_iter_offset(bio, start),77			.type = BB_NONE, .rw = rw78		};79 80	/* check if we can map the pages contiguously: */81	__bio_for_each_segment(bv, bio, iter, start) {82		if (iter.bi_size != start.bi_size &&83		    bv.bv_offset)84			goto bounce;85 86		if (bv.bv_len < iter.bi_size &&87		    bv.bv_offset + bv.bv_len < PAGE_SIZE)88			goto bounce;89 90		nr_pages++;91	}92 93	BUG_ON(DIV_ROUND_UP(start.bi_size, PAGE_SIZE) > nr_pages);94 95	pages = nr_pages > ARRAY_SIZE(stack_pages)96		? kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS)97		: stack_pages;98	if (!pages)99		goto bounce;100 101	nr_pages = 0;102	__bio_for_each_segment(bv, bio, iter, start)103		pages[nr_pages++] = bv.bv_page;104 105	data = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);106	if (pages != stack_pages)107		kfree(pages);108 109	if (data)110		return (struct bbuf) {111			.b = data + bio_iter_offset(bio, start),112			.type = BB_VMAP, .rw = rw113		};114bounce:115	ret = __bounce_alloc(c, start.bi_size, rw);116 117	if (rw == READ)118		memcpy_from_bio(ret.b, bio, start);119 120	return ret;121}122 123static struct bbuf bio_map_or_bounce(struct bch_fs *c, struct bio *bio, int rw)124{125	return __bio_map_or_bounce(c, bio, bio->bi_iter, rw);126}127 128static void bio_unmap_or_unbounce(struct bch_fs *c, struct bbuf buf)129{130	switch (buf.type) {131	case BB_NONE:132		break;133	case BB_VMAP:134		vunmap((void *) ((unsigned long) buf.b & PAGE_MASK));135		break;136	case BB_KMALLOC:137		kfree(buf.b);138		break;139	case BB_MEMPOOL:140		mempool_free(buf.b, &c->compression_bounce[buf.rw]);141		break;142	}143}144 145static inline void zlib_set_workspace(z_stream *strm, void *workspace)146{147#ifdef __KERNEL__148	strm->workspace = workspace;149#endif150}151 152static int __bio_uncompress(struct bch_fs *c, struct bio *src,153			    void *dst_data, struct bch_extent_crc_unpacked crc)154{155	struct bbuf src_data = { NULL };156	size_t src_len = src->bi_iter.bi_size;157	size_t dst_len = crc.uncompressed_size << 9;158	void *workspace;159	int ret;160 161	src_data = bio_map_or_bounce(c, src, READ);162 163	switch (crc.compression_type) {164	case BCH_COMPRESSION_TYPE_lz4_old:165	case BCH_COMPRESSION_TYPE_lz4:166		ret = LZ4_decompress_safe_partial(src_data.b, dst_data,167						  src_len, dst_len, dst_len);168		if (ret != dst_len)169			goto err;170		break;171	case BCH_COMPRESSION_TYPE_gzip: {172		z_stream strm = {173			.next_in	= src_data.b,174			.avail_in	= src_len,175			.next_out	= dst_data,176			.avail_out	= dst_len,177		};178 179		workspace = mempool_alloc(&c->decompress_workspace, GFP_NOFS);180 181		zlib_set_workspace(&strm, workspace);182		zlib_inflateInit2(&strm, -MAX_WBITS);183		ret = zlib_inflate(&strm, Z_FINISH);184 185		mempool_free(workspace, &c->decompress_workspace);186 187		if (ret != Z_STREAM_END)188			goto err;189		break;190	}191	case BCH_COMPRESSION_TYPE_zstd: {192		ZSTD_DCtx *ctx;193		size_t real_src_len = le32_to_cpup(src_data.b);194 195		if (real_src_len > src_len - 4)196			goto err;197 198		workspace = mempool_alloc(&c->decompress_workspace, GFP_NOFS);199		ctx = zstd_init_dctx(workspace, zstd_dctx_workspace_bound());200 201		ret = zstd_decompress_dctx(ctx,202				dst_data,	dst_len,203				src_data.b + 4, real_src_len);204 205		mempool_free(workspace, &c->decompress_workspace);206 207		if (ret != dst_len)208			goto err;209		break;210	}211	default:212		BUG();213	}214	ret = 0;215out:216	bio_unmap_or_unbounce(c, src_data);217	return ret;218err:219	ret = -EIO;220	goto out;221}222 223int bch2_bio_uncompress_inplace(struct bch_fs *c, struct bio *bio,224				struct bch_extent_crc_unpacked *crc)225{226	struct bbuf data = { NULL };227	size_t dst_len = crc->uncompressed_size << 9;228 229	/* bio must own its pages: */230	BUG_ON(!bio->bi_vcnt);231	BUG_ON(DIV_ROUND_UP(crc->live_size, PAGE_SECTORS) > bio->bi_max_vecs);232 233	if (crc->uncompressed_size << 9	> c->opts.encoded_extent_max ||234	    crc->compressed_size << 9	> c->opts.encoded_extent_max) {235		bch_err(c, "error rewriting existing data: extent too big");236		return -EIO;237	}238 239	data = __bounce_alloc(c, dst_len, WRITE);240 241	if (__bio_uncompress(c, bio, data.b, *crc)) {242		if (!c->opts.no_data_io)243			bch_err(c, "error rewriting existing data: decompression error");244		bio_unmap_or_unbounce(c, data);245		return -EIO;246	}247 248	/*249	 * XXX: don't have a good way to assert that the bio was allocated with250	 * enough space, we depend on bch2_move_extent doing the right thing251	 */252	bio->bi_iter.bi_size = crc->live_size << 9;253 254	memcpy_to_bio(bio, bio->bi_iter, data.b + (crc->offset << 9));255 256	crc->csum_type		= 0;257	crc->compression_type	= 0;258	crc->compressed_size	= crc->live_size;259	crc->uncompressed_size	= crc->live_size;260	crc->offset		= 0;261	crc->csum		= (struct bch_csum) { 0, 0 };262 263	bio_unmap_or_unbounce(c, data);264	return 0;265}266 267int bch2_bio_uncompress(struct bch_fs *c, struct bio *src,268		       struct bio *dst, struct bvec_iter dst_iter,269		       struct bch_extent_crc_unpacked crc)270{271	struct bbuf dst_data = { NULL };272	size_t dst_len = crc.uncompressed_size << 9;273	int ret;274 275	if (crc.uncompressed_size << 9	> c->opts.encoded_extent_max ||276	    crc.compressed_size << 9	> c->opts.encoded_extent_max)277		return -EIO;278 279	dst_data = dst_len == dst_iter.bi_size280		? __bio_map_or_bounce(c, dst, dst_iter, WRITE)281		: __bounce_alloc(c, dst_len, WRITE);282 283	ret = __bio_uncompress(c, src, dst_data.b, crc);284	if (ret)285		goto err;286 287	if (dst_data.type != BB_NONE &&288	    dst_data.type != BB_VMAP)289		memcpy_to_bio(dst, dst_iter, dst_data.b + (crc.offset << 9));290err:291	bio_unmap_or_unbounce(c, dst_data);292	return ret;293}294 295static int attempt_compress(struct bch_fs *c,296			    void *workspace,297			    void *dst, size_t dst_len,298			    void *src, size_t src_len,299			    struct bch_compression_opt compression)300{301	enum bch_compression_type compression_type =302		__bch2_compression_opt_to_type[compression.type];303 304	switch (compression_type) {305	case BCH_COMPRESSION_TYPE_lz4:306		if (compression.level < LZ4HC_MIN_CLEVEL) {307			int len = src_len;308			int ret = LZ4_compress_destSize(309					src,		dst,310					&len,		dst_len,311					workspace);312			if (len < src_len)313				return -len;314 315			return ret;316		} else {317			int ret = LZ4_compress_HC(318					src,		dst,319					src_len,	dst_len,320					compression.level,321					workspace);322 323			return ret ?: -1;324		}325	case BCH_COMPRESSION_TYPE_gzip: {326		z_stream strm = {327			.next_in	= src,328			.avail_in	= src_len,329			.next_out	= dst,330			.avail_out	= dst_len,331		};332 333		zlib_set_workspace(&strm, workspace);334		zlib_deflateInit2(&strm,335				  compression.level336				  ? clamp_t(unsigned, compression.level,337					    Z_BEST_SPEED, Z_BEST_COMPRESSION)338				  : Z_DEFAULT_COMPRESSION,339				  Z_DEFLATED, -MAX_WBITS, DEF_MEM_LEVEL,340				  Z_DEFAULT_STRATEGY);341 342		if (zlib_deflate(&strm, Z_FINISH) != Z_STREAM_END)343			return 0;344 345		if (zlib_deflateEnd(&strm) != Z_OK)346			return 0;347 348		return strm.total_out;349	}350	case BCH_COMPRESSION_TYPE_zstd: {351		/*352		 * rescale:353		 * zstd max compression level is 22, our max level is 15354		 */355		unsigned level = min((compression.level * 3) / 2, zstd_max_clevel());356		ZSTD_parameters params = zstd_get_params(level, c->opts.encoded_extent_max);357		ZSTD_CCtx *ctx = zstd_init_cctx(workspace, c->zstd_workspace_size);358 359		/*360		 * ZSTD requires that when we decompress we pass in the exact361		 * compressed size - rounding it up to the nearest sector362		 * doesn't work, so we use the first 4 bytes of the buffer for363		 * that.364		 *365		 * Additionally, the ZSTD code seems to have a bug where it will366		 * write just past the end of the buffer - so subtract a fudge367		 * factor (7 bytes) from the dst buffer size to account for368		 * that.369		 */370		size_t len = zstd_compress_cctx(ctx,371				dst + 4,	dst_len - 4 - 7,372				src,		src_len,373				&params);374		if (zstd_is_error(len))375			return 0;376 377		*((__le32 *) dst) = cpu_to_le32(len);378		return len + 4;379	}380	default:381		BUG();382	}383}384 385static unsigned __bio_compress(struct bch_fs *c,386			       struct bio *dst, size_t *dst_len,387			       struct bio *src, size_t *src_len,388			       struct bch_compression_opt compression)389{390	struct bbuf src_data = { NULL }, dst_data = { NULL };391	void *workspace;392	enum bch_compression_type compression_type =393		__bch2_compression_opt_to_type[compression.type];394	unsigned pad;395	int ret = 0;396 397	BUG_ON(compression_type >= BCH_COMPRESSION_TYPE_NR);398	BUG_ON(!mempool_initialized(&c->compress_workspace[compression_type]));399 400	/* If it's only one block, don't bother trying to compress: */401	if (src->bi_iter.bi_size <= c->opts.block_size)402		return BCH_COMPRESSION_TYPE_incompressible;403 404	dst_data = bio_map_or_bounce(c, dst, WRITE);405	src_data = bio_map_or_bounce(c, src, READ);406 407	workspace = mempool_alloc(&c->compress_workspace[compression_type], GFP_NOFS);408 409	*src_len = src->bi_iter.bi_size;410	*dst_len = dst->bi_iter.bi_size;411 412	/*413	 * XXX: this algorithm sucks when the compression code doesn't tell us414	 * how much would fit, like LZ4 does:415	 */416	while (1) {417		if (*src_len <= block_bytes(c)) {418			ret = -1;419			break;420		}421 422		ret = attempt_compress(c, workspace,423				       dst_data.b,	*dst_len,424				       src_data.b,	*src_len,425				       compression);426		if (ret > 0) {427			*dst_len = ret;428			ret = 0;429			break;430		}431 432		/* Didn't fit: should we retry with a smaller amount?  */433		if (*src_len <= *dst_len) {434			ret = -1;435			break;436		}437 438		/*439		 * If ret is negative, it's a hint as to how much data would fit440		 */441		BUG_ON(-ret >= *src_len);442 443		if (ret < 0)444			*src_len = -ret;445		else446			*src_len -= (*src_len - *dst_len) / 2;447		*src_len = round_down(*src_len, block_bytes(c));448	}449 450	mempool_free(workspace, &c->compress_workspace[compression_type]);451 452	if (ret)453		goto err;454 455	/* Didn't get smaller: */456	if (round_up(*dst_len, block_bytes(c)) >= *src_len)457		goto err;458 459	pad = round_up(*dst_len, block_bytes(c)) - *dst_len;460 461	memset(dst_data.b + *dst_len, 0, pad);462	*dst_len += pad;463 464	if (dst_data.type != BB_NONE &&465	    dst_data.type != BB_VMAP)466		memcpy_to_bio(dst, dst->bi_iter, dst_data.b);467 468	BUG_ON(!*dst_len || *dst_len > dst->bi_iter.bi_size);469	BUG_ON(!*src_len || *src_len > src->bi_iter.bi_size);470	BUG_ON(*dst_len & (block_bytes(c) - 1));471	BUG_ON(*src_len & (block_bytes(c) - 1));472	ret = compression_type;473out:474	bio_unmap_or_unbounce(c, src_data);475	bio_unmap_or_unbounce(c, dst_data);476	return ret;477err:478	ret = BCH_COMPRESSION_TYPE_incompressible;479	goto out;480}481 482unsigned bch2_bio_compress(struct bch_fs *c,483			   struct bio *dst, size_t *dst_len,484			   struct bio *src, size_t *src_len,485			   unsigned compression_opt)486{487	unsigned orig_dst = dst->bi_iter.bi_size;488	unsigned orig_src = src->bi_iter.bi_size;489	unsigned compression_type;490 491	/* Don't consume more than BCH_ENCODED_EXTENT_MAX from @src: */492	src->bi_iter.bi_size = min_t(unsigned, src->bi_iter.bi_size,493				     c->opts.encoded_extent_max);494	/* Don't generate a bigger output than input: */495	dst->bi_iter.bi_size = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);496 497	compression_type =498		__bio_compress(c, dst, dst_len, src, src_len,499			       bch2_compression_decode(compression_opt));500 501	dst->bi_iter.bi_size = orig_dst;502	src->bi_iter.bi_size = orig_src;503	return compression_type;504}505 506static int __bch2_fs_compress_init(struct bch_fs *, u64);507 508#define BCH_FEATURE_none	0509 510static const unsigned bch2_compression_opt_to_feature[] = {511#define x(t, n) [BCH_COMPRESSION_OPT_##t] = BCH_FEATURE_##t,512	BCH_COMPRESSION_OPTS()513#undef x514};515 516#undef BCH_FEATURE_none517 518static int __bch2_check_set_has_compressed_data(struct bch_fs *c, u64 f)519{520	int ret = 0;521 522	if ((c->sb.features & f) == f)523		return 0;524 525	mutex_lock(&c->sb_lock);526 527	if ((c->sb.features & f) == f) {528		mutex_unlock(&c->sb_lock);529		return 0;530	}531 532	ret = __bch2_fs_compress_init(c, c->sb.features|f);533	if (ret) {534		mutex_unlock(&c->sb_lock);535		return ret;536	}537 538	c->disk_sb.sb->features[0] |= cpu_to_le64(f);539	bch2_write_super(c);540	mutex_unlock(&c->sb_lock);541 542	return 0;543}544 545int bch2_check_set_has_compressed_data(struct bch_fs *c,546				       unsigned compression_opt)547{548	unsigned compression_type = bch2_compression_decode(compression_opt).type;549 550	BUG_ON(compression_type >= ARRAY_SIZE(bch2_compression_opt_to_feature));551 552	return compression_type553		? __bch2_check_set_has_compressed_data(c,554				1ULL << bch2_compression_opt_to_feature[compression_type])555		: 0;556}557 558void bch2_fs_compress_exit(struct bch_fs *c)559{560	unsigned i;561 562	mempool_exit(&c->decompress_workspace);563	for (i = 0; i < ARRAY_SIZE(c->compress_workspace); i++)564		mempool_exit(&c->compress_workspace[i]);565	mempool_exit(&c->compression_bounce[WRITE]);566	mempool_exit(&c->compression_bounce[READ]);567}568 569static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)570{571	size_t decompress_workspace_size = 0;572	ZSTD_parameters params = zstd_get_params(zstd_max_clevel(),573						 c->opts.encoded_extent_max);574 575	c->zstd_workspace_size = zstd_cctx_workspace_bound(&params.cParams);576 577	struct {578		unsigned			feature;579		enum bch_compression_type	type;580		size_t				compress_workspace;581		size_t				decompress_workspace;582	} compression_types[] = {583		{ BCH_FEATURE_lz4, BCH_COMPRESSION_TYPE_lz4,584			max_t(size_t, LZ4_MEM_COMPRESS, LZ4HC_MEM_COMPRESS),585			0 },586		{ BCH_FEATURE_gzip, BCH_COMPRESSION_TYPE_gzip,587			zlib_deflate_workspacesize(MAX_WBITS, DEF_MEM_LEVEL),588			zlib_inflate_workspacesize(), },589		{ BCH_FEATURE_zstd, BCH_COMPRESSION_TYPE_zstd,590			c->zstd_workspace_size,591			zstd_dctx_workspace_bound() },592	}, *i;593	bool have_compressed = false;594 595	for (i = compression_types;596	     i < compression_types + ARRAY_SIZE(compression_types);597	     i++)598		have_compressed |= (features & (1 << i->feature)) != 0;599 600	if (!have_compressed)601		return 0;602 603	if (!mempool_initialized(&c->compression_bounce[READ]) &&604	    mempool_init_kvmalloc_pool(&c->compression_bounce[READ],605				       1, c->opts.encoded_extent_max))606		return -BCH_ERR_ENOMEM_compression_bounce_read_init;607 608	if (!mempool_initialized(&c->compression_bounce[WRITE]) &&609	    mempool_init_kvmalloc_pool(&c->compression_bounce[WRITE],610				       1, c->opts.encoded_extent_max))611		return -BCH_ERR_ENOMEM_compression_bounce_write_init;612 613	for (i = compression_types;614	     i < compression_types + ARRAY_SIZE(compression_types);615	     i++) {616		decompress_workspace_size =617			max(decompress_workspace_size, i->decompress_workspace);618 619		if (!(features & (1 << i->feature)))620			continue;621 622		if (mempool_initialized(&c->compress_workspace[i->type]))623			continue;624 625		if (mempool_init_kvmalloc_pool(626				&c->compress_workspace[i->type],627				1, i->compress_workspace))628			return -BCH_ERR_ENOMEM_compression_workspace_init;629	}630 631	if (!mempool_initialized(&c->decompress_workspace) &&632	    mempool_init_kvmalloc_pool(&c->decompress_workspace,633				       1, decompress_workspace_size))634		return -BCH_ERR_ENOMEM_decompression_workspace_init;635 636	return 0;637}638 639static u64 compression_opt_to_feature(unsigned v)640{641	unsigned type = bch2_compression_decode(v).type;642 643	return BIT_ULL(bch2_compression_opt_to_feature[type]);644}645 646int bch2_fs_compress_init(struct bch_fs *c)647{648	u64 f = c->sb.features;649 650	f |= compression_opt_to_feature(c->opts.compression);651	f |= compression_opt_to_feature(c->opts.background_compression);652 653	return __bch2_fs_compress_init(c, f);654}655 656int bch2_opt_compression_parse(struct bch_fs *c, const char *_val, u64 *res,657			       struct printbuf *err)658{659	char *val = kstrdup(_val, GFP_KERNEL);660	char *p = val, *type_str, *level_str;661	struct bch_compression_opt opt = { 0 };662	int ret;663 664	if (!val)665		return -ENOMEM;666 667	type_str = strsep(&p, ":");668	level_str = p;669 670	ret = match_string(bch2_compression_opts, -1, type_str);671	if (ret < 0 && err)672		prt_str(err, "invalid compression type");673	if (ret < 0)674		goto err;675 676	opt.type = ret;677 678	if (level_str) {679		unsigned level;680 681		ret = kstrtouint(level_str, 10, &level);682		if (!ret && !opt.type && level)683			ret = -EINVAL;684		if (!ret && level > 15)685			ret = -EINVAL;686		if (ret < 0 && err)687			prt_str(err, "invalid compression level");688		if (ret < 0)689			goto err;690 691		opt.level = level;692	}693 694	*res = bch2_compression_encode(opt);695err:696	kfree(val);697	return ret;698}699 700void bch2_compression_opt_to_text(struct printbuf *out, u64 v)701{702	struct bch_compression_opt opt = bch2_compression_decode(v);703 704	if (opt.type < BCH_COMPRESSION_OPT_NR)705		prt_str(out, bch2_compression_opts[opt.type]);706	else707		prt_printf(out, "(unknown compression opt %u)", opt.type);708	if (opt.level)709		prt_printf(out, ":%u", opt.level);710}711 712void bch2_opt_compression_to_text(struct printbuf *out,713				  struct bch_fs *c,714				  struct bch_sb *sb,715				  u64 v)716{717	return bch2_compression_opt_to_text(out, v);718}719 720int bch2_opt_compression_validate(u64 v, struct printbuf *err)721{722	if (!bch2_compression_opt_valid(v)) {723		prt_printf(err, "invalid compression opt %llu", v);724		return -BCH_ERR_invalid_sb_opt_compression;725	}726 727	return 0;728}729