brintos

brintos / linux-shallow public Read only

0
0
Text · 42.5 KiB · 90aef26 Raw
1611 lines · c
1// SPDX-License-Identifier: GPL-2.02/*3 * Copyright (C) 2008 Oracle.  All rights reserved.4 */5 6#include <linux/kernel.h>7#include <linux/bio.h>8#include <linux/file.h>9#include <linux/fs.h>10#include <linux/pagemap.h>11#include <linux/pagevec.h>12#include <linux/highmem.h>13#include <linux/kthread.h>14#include <linux/time.h>15#include <linux/init.h>16#include <linux/string.h>17#include <linux/backing-dev.h>18#include <linux/writeback.h>19#include <linux/psi.h>20#include <linux/slab.h>21#include <linux/sched/mm.h>22#include <linux/log2.h>23#include <linux/shrinker.h>24#include <crypto/hash.h>25#include "misc.h"26#include "ctree.h"27#include "fs.h"28#include "btrfs_inode.h"29#include "bio.h"30#include "ordered-data.h"31#include "compression.h"32#include "extent_io.h"33#include "extent_map.h"34#include "subpage.h"35#include "messages.h"36#include "super.h"37 38static struct bio_set btrfs_compressed_bioset;39 40static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };41 42const char* btrfs_compress_type2str(enum btrfs_compression_type type)43{44	switch (type) {45	case BTRFS_COMPRESS_ZLIB:46	case BTRFS_COMPRESS_LZO:47	case BTRFS_COMPRESS_ZSTD:48	case BTRFS_COMPRESS_NONE:49		return btrfs_compress_types[type];50	default:51		break;52	}53 54	return NULL;55}56 57static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio)58{59	return container_of(bbio, struct compressed_bio, bbio);60}61 62static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode,63						   u64 start, blk_opf_t op,64						   btrfs_bio_end_io_t end_io)65{66	struct btrfs_bio *bbio;67 68	bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op,69					  GFP_NOFS, &btrfs_compressed_bioset));70	btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);71	bbio->inode = inode;72	bbio->file_offset = start;73	return to_compressed_bio(bbio);74}75 76bool btrfs_compress_is_valid_type(const char *str, size_t len)77{78	int i;79 80	for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {81		size_t comp_len = strlen(btrfs_compress_types[i]);82 83		if (len < comp_len)84			continue;85 86		if (!strncmp(btrfs_compress_types[i], str, comp_len))87			return true;88	}89	return false;90}91 92static int compression_compress_pages(int type, struct list_head *ws,93				      struct address_space *mapping, u64 start,94				      struct folio **folios, unsigned long *out_folios,95				      unsigned long *total_in, unsigned long *total_out)96{97	switch (type) {98	case BTRFS_COMPRESS_ZLIB:99		return zlib_compress_folios(ws, mapping, start, folios,100					    out_folios, total_in, total_out);101	case BTRFS_COMPRESS_LZO:102		return lzo_compress_folios(ws, mapping, start, folios,103					   out_folios, total_in, total_out);104	case BTRFS_COMPRESS_ZSTD:105		return zstd_compress_folios(ws, mapping, start, folios,106					    out_folios, total_in, total_out);107	case BTRFS_COMPRESS_NONE:108	default:109		/*110		 * This can happen when compression races with remount setting111		 * it to 'no compress', while caller doesn't call112		 * inode_need_compress() to check if we really need to113		 * compress.114		 *115		 * Not a big deal, just need to inform caller that we116		 * haven't allocated any pages yet.117		 */118		*out_folios = 0;119		return -E2BIG;120	}121}122 123static int compression_decompress_bio(struct list_head *ws,124				      struct compressed_bio *cb)125{126	switch (cb->compress_type) {127	case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);128	case BTRFS_COMPRESS_LZO:  return lzo_decompress_bio(ws, cb);129	case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);130	case BTRFS_COMPRESS_NONE:131	default:132		/*133		 * This can't happen, the type is validated several times134		 * before we get here.135		 */136		BUG();137	}138}139 140static int compression_decompress(int type, struct list_head *ws,141		const u8 *data_in, struct folio *dest_folio,142		unsigned long dest_pgoff, size_t srclen, size_t destlen)143{144	switch (type) {145	case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_folio,146						dest_pgoff, srclen, destlen);147	case BTRFS_COMPRESS_LZO:  return lzo_decompress(ws, data_in, dest_folio,148						dest_pgoff, srclen, destlen);149	case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_folio,150						dest_pgoff, srclen, destlen);151	case BTRFS_COMPRESS_NONE:152	default:153		/*154		 * This can't happen, the type is validated several times155		 * before we get here.156		 */157		BUG();158	}159}160 161static void btrfs_free_compressed_folios(struct compressed_bio *cb)162{163	for (unsigned int i = 0; i < cb->nr_folios; i++)164		btrfs_free_compr_folio(cb->compressed_folios[i]);165	kfree(cb->compressed_folios);166}167 168static int btrfs_decompress_bio(struct compressed_bio *cb);169 170/*171 * Global cache of last unused pages for compression/decompression.172 */173static struct btrfs_compr_pool {174	struct shrinker *shrinker;175	spinlock_t lock;176	struct list_head list;177	int count;178	int thresh;179} compr_pool;180 181static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)182{183	int ret;184 185	/*186	 * We must not read the values more than once if 'ret' gets expanded in187	 * the return statement so we don't accidentally return a negative188	 * number, even if the first condition finds it positive.189	 */190	ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);191 192	return ret > 0 ? ret : 0;193}194 195static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)196{197	struct list_head remove;198	struct list_head *tmp, *next;199	int freed;200 201	if (compr_pool.count == 0)202		return SHRINK_STOP;203 204	INIT_LIST_HEAD(&remove);205 206	/* For now, just simply drain the whole list. */207	spin_lock(&compr_pool.lock);208	list_splice_init(&compr_pool.list, &remove);209	freed = compr_pool.count;210	compr_pool.count = 0;211	spin_unlock(&compr_pool.lock);212 213	list_for_each_safe(tmp, next, &remove) {214		struct page *page = list_entry(tmp, struct page, lru);215 216		ASSERT(page_ref_count(page) == 1);217		put_page(page);218	}219 220	return freed;221}222 223/*224 * Common wrappers for page allocation from compression wrappers225 */226struct folio *btrfs_alloc_compr_folio(void)227{228	struct folio *folio = NULL;229 230	spin_lock(&compr_pool.lock);231	if (compr_pool.count > 0) {232		folio = list_first_entry(&compr_pool.list, struct folio, lru);233		list_del_init(&folio->lru);234		compr_pool.count--;235	}236	spin_unlock(&compr_pool.lock);237 238	if (folio)239		return folio;240 241	return folio_alloc(GFP_NOFS, 0);242}243 244void btrfs_free_compr_folio(struct folio *folio)245{246	bool do_free = false;247 248	spin_lock(&compr_pool.lock);249	if (compr_pool.count > compr_pool.thresh) {250		do_free = true;251	} else {252		list_add(&folio->lru, &compr_pool.list);253		compr_pool.count++;254	}255	spin_unlock(&compr_pool.lock);256 257	if (!do_free)258		return;259 260	ASSERT(folio_ref_count(folio) == 1);261	folio_put(folio);262}263 264static void end_bbio_compressed_read(struct btrfs_bio *bbio)265{266	struct compressed_bio *cb = to_compressed_bio(bbio);267	blk_status_t status = bbio->bio.bi_status;268 269	if (!status)270		status = errno_to_blk_status(btrfs_decompress_bio(cb));271 272	btrfs_free_compressed_folios(cb);273	btrfs_bio_end_io(cb->orig_bbio, status);274	bio_put(&bbio->bio);275}276 277/*278 * Clear the writeback bits on all of the file279 * pages for a compressed write280 */281static noinline void end_compressed_writeback(const struct compressed_bio *cb)282{283	struct inode *inode = &cb->bbio.inode->vfs_inode;284	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);285	unsigned long index = cb->start >> PAGE_SHIFT;286	unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;287	struct folio_batch fbatch;288	const int error = blk_status_to_errno(cb->bbio.bio.bi_status);289	int i;290	int ret;291 292	if (error)293		mapping_set_error(inode->i_mapping, error);294 295	folio_batch_init(&fbatch);296	while (index <= end_index) {297		ret = filemap_get_folios(inode->i_mapping, &index, end_index,298				&fbatch);299 300		if (ret == 0)301			return;302 303		for (i = 0; i < ret; i++) {304			struct folio *folio = fbatch.folios[i];305 306			btrfs_folio_clamp_clear_writeback(fs_info, folio,307							  cb->start, cb->len);308		}309		folio_batch_release(&fbatch);310	}311	/* the inode may be gone now */312}313 314static void btrfs_finish_compressed_write_work(struct work_struct *work)315{316	struct compressed_bio *cb =317		container_of(work, struct compressed_bio, write_end_work);318 319	btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,320				    cb->bbio.bio.bi_status == BLK_STS_OK);321 322	if (cb->writeback)323		end_compressed_writeback(cb);324	/* Note, our inode could be gone now */325 326	btrfs_free_compressed_folios(cb);327	bio_put(&cb->bbio.bio);328}329 330/*331 * Do the cleanup once all the compressed pages hit the disk.  This will clear332 * writeback on the file pages and free the compressed pages.333 *334 * This also calls the writeback end hooks for the file pages so that metadata335 * and checksums can be updated in the file.336 */337static void end_bbio_compressed_write(struct btrfs_bio *bbio)338{339	struct compressed_bio *cb = to_compressed_bio(bbio);340	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;341 342	queue_work(fs_info->compressed_write_workers, &cb->write_end_work);343}344 345static void btrfs_add_compressed_bio_folios(struct compressed_bio *cb)346{347	struct bio *bio = &cb->bbio.bio;348	u32 offset = 0;349 350	while (offset < cb->compressed_len) {351		int ret;352		u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);353 354		/* Maximum compressed extent is smaller than bio size limit. */355		ret = bio_add_folio(bio, cb->compressed_folios[offset >> PAGE_SHIFT],356				    len, 0);357		ASSERT(ret);358		offset += len;359	}360}361 362/*363 * worker function to build and submit bios for previously compressed pages.364 * The corresponding pages in the inode should be marked for writeback365 * and the compressed pages should have a reference on them for dropping366 * when the IO is complete.367 *368 * This also checksums the file bytes and gets things ready for369 * the end io hooks.370 */371void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,372				   struct folio **compressed_folios,373				   unsigned int nr_folios,374				   blk_opf_t write_flags,375				   bool writeback)376{377	struct btrfs_inode *inode = ordered->inode;378	struct btrfs_fs_info *fs_info = inode->root->fs_info;379	struct compressed_bio *cb;380 381	ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));382	ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));383 384	cb = alloc_compressed_bio(inode, ordered->file_offset,385				  REQ_OP_WRITE | write_flags,386				  end_bbio_compressed_write);387	cb->start = ordered->file_offset;388	cb->len = ordered->num_bytes;389	cb->compressed_folios = compressed_folios;390	cb->compressed_len = ordered->disk_num_bytes;391	cb->writeback = writeback;392	INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);393	cb->nr_folios = nr_folios;394	cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;395	cb->bbio.ordered = ordered;396	btrfs_add_compressed_bio_folios(cb);397 398	btrfs_submit_bbio(&cb->bbio, 0);399}400 401/*402 * Add extra pages in the same compressed file extent so that we don't need to403 * re-read the same extent again and again.404 *405 * NOTE: this won't work well for subpage, as for subpage read, we lock the406 * full page then submit bio for each compressed/regular extents.407 *408 * This means, if we have several sectors in the same page points to the same409 * on-disk compressed data, we will re-read the same extent many times and410 * this function can only help for the next page.411 */412static noinline int add_ra_bio_pages(struct inode *inode,413				     u64 compressed_end,414				     struct compressed_bio *cb,415				     int *memstall, unsigned long *pflags)416{417	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);418	unsigned long end_index;419	struct bio *orig_bio = &cb->orig_bbio->bio;420	u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;421	u64 isize = i_size_read(inode);422	int ret;423	struct folio *folio;424	struct extent_map *em;425	struct address_space *mapping = inode->i_mapping;426	struct extent_map_tree *em_tree;427	struct extent_io_tree *tree;428	int sectors_missed = 0;429 430	em_tree = &BTRFS_I(inode)->extent_tree;431	tree = &BTRFS_I(inode)->io_tree;432 433	if (isize == 0)434		return 0;435 436	/*437	 * For current subpage support, we only support 64K page size,438	 * which means maximum compressed extent size (128K) is just 2x page439	 * size.440	 * This makes readahead less effective, so here disable readahead for441	 * subpage for now, until full compressed write is supported.442	 */443	if (fs_info->sectorsize < PAGE_SIZE)444		return 0;445 446	end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;447 448	while (cur < compressed_end) {449		u64 page_end;450		u64 pg_index = cur >> PAGE_SHIFT;451		u32 add_size;452 453		if (pg_index > end_index)454			break;455 456		folio = __filemap_get_folio(mapping, pg_index, 0, 0);457		if (!IS_ERR(folio)) {458			u64 folio_sz = folio_size(folio);459			u64 offset = offset_in_folio(folio, cur);460 461			folio_put(folio);462			sectors_missed += (folio_sz - offset) >>463					  fs_info->sectorsize_bits;464 465			/* Beyond threshold, no need to continue */466			if (sectors_missed > 4)467				break;468 469			/*470			 * Jump to next page start as we already have page for471			 * current offset.472			 */473			cur += (folio_sz - offset);474			continue;475		}476 477		folio = filemap_alloc_folio(mapping_gfp_constraint(mapping,478								   ~__GFP_FS), 0);479		if (!folio)480			break;481 482		if (filemap_add_folio(mapping, folio, pg_index, GFP_NOFS)) {483			/* There is already a page, skip to page end */484			cur += folio_size(folio);485			folio_put(folio);486			continue;487		}488 489		if (!*memstall && folio_test_workingset(folio)) {490			psi_memstall_enter(pflags);491			*memstall = 1;492		}493 494		ret = set_folio_extent_mapped(folio);495		if (ret < 0) {496			folio_unlock(folio);497			folio_put(folio);498			break;499		}500 501		page_end = (pg_index << PAGE_SHIFT) + folio_size(folio) - 1;502		lock_extent(tree, cur, page_end, NULL);503		read_lock(&em_tree->lock);504		em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);505		read_unlock(&em_tree->lock);506 507		/*508		 * At this point, we have a locked page in the page cache for509		 * these bytes in the file.  But, we have to make sure they map510		 * to this compressed extent on disk.511		 */512		if (!em || cur < em->start ||513		    (cur + fs_info->sectorsize > extent_map_end(em)) ||514		    (extent_map_block_start(em) >> SECTOR_SHIFT) !=515		    orig_bio->bi_iter.bi_sector) {516			free_extent_map(em);517			unlock_extent(tree, cur, page_end, NULL);518			folio_unlock(folio);519			folio_put(folio);520			break;521		}522		add_size = min(em->start + em->len, page_end + 1) - cur;523		free_extent_map(em);524		unlock_extent(tree, cur, page_end, NULL);525 526		if (folio->index == end_index) {527			size_t zero_offset = offset_in_folio(folio, isize);528 529			if (zero_offset) {530				int zeros;531				zeros = folio_size(folio) - zero_offset;532				folio_zero_range(folio, zero_offset, zeros);533			}534		}535 536		if (!bio_add_folio(orig_bio, folio, add_size,537				   offset_in_folio(folio, cur))) {538			folio_unlock(folio);539			folio_put(folio);540			break;541		}542		/*543		 * If it's subpage, we also need to increase its544		 * subpage::readers number, as at endio we will decrease545		 * subpage::readers and to unlock the page.546		 */547		if (fs_info->sectorsize < PAGE_SIZE)548			btrfs_subpage_start_reader(fs_info, folio, cur,549						   add_size);550		folio_put(folio);551		cur += add_size;552	}553	return 0;554}555 556/*557 * for a compressed read, the bio we get passed has all the inode pages558 * in it.  We don't actually do IO on those pages but allocate new ones559 * to hold the compressed pages on disk.560 *561 * bio->bi_iter.bi_sector points to the compressed extent on disk562 * bio->bi_io_vec points to all of the inode pages563 *564 * After the compressed pages are read, we copy the bytes into the565 * bio we were passed and then call the bio end_io calls566 */567void btrfs_submit_compressed_read(struct btrfs_bio *bbio)568{569	struct btrfs_inode *inode = bbio->inode;570	struct btrfs_fs_info *fs_info = inode->root->fs_info;571	struct extent_map_tree *em_tree = &inode->extent_tree;572	struct compressed_bio *cb;573	unsigned int compressed_len;574	u64 file_offset = bbio->file_offset;575	u64 em_len;576	u64 em_start;577	struct extent_map *em;578	unsigned long pflags;579	int memstall = 0;580	blk_status_t ret;581	int ret2;582 583	/* we need the actual starting offset of this extent in the file */584	read_lock(&em_tree->lock);585	em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);586	read_unlock(&em_tree->lock);587	if (!em) {588		ret = BLK_STS_IOERR;589		goto out;590	}591 592	ASSERT(extent_map_is_compressed(em));593	compressed_len = em->disk_num_bytes;594 595	cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,596				  end_bbio_compressed_read);597 598	cb->start = em->start - em->offset;599	em_len = em->len;600	em_start = em->start;601 602	cb->len = bbio->bio.bi_iter.bi_size;603	cb->compressed_len = compressed_len;604	cb->compress_type = extent_map_compression(em);605	cb->orig_bbio = bbio;606 607	free_extent_map(em);608 609	cb->nr_folios = DIV_ROUND_UP(compressed_len, PAGE_SIZE);610	cb->compressed_folios = kcalloc(cb->nr_folios, sizeof(struct page *), GFP_NOFS);611	if (!cb->compressed_folios) {612		ret = BLK_STS_RESOURCE;613		goto out_free_bio;614	}615 616	ret2 = btrfs_alloc_folio_array(cb->nr_folios, cb->compressed_folios);617	if (ret2) {618		ret = BLK_STS_RESOURCE;619		goto out_free_compressed_pages;620	}621 622	add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,623			 &pflags);624 625	/* include any pages we added in add_ra-bio_pages */626	cb->len = bbio->bio.bi_iter.bi_size;627	cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;628	btrfs_add_compressed_bio_folios(cb);629 630	if (memstall)631		psi_memstall_leave(&pflags);632 633	btrfs_submit_bbio(&cb->bbio, 0);634	return;635 636out_free_compressed_pages:637	kfree(cb->compressed_folios);638out_free_bio:639	bio_put(&cb->bbio.bio);640out:641	btrfs_bio_end_io(bbio, ret);642}643 644/*645 * Heuristic uses systematic sampling to collect data from the input data646 * range, the logic can be tuned by the following constants:647 *648 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample649 * @SAMPLING_INTERVAL  - range from which the sampled data can be collected650 */651#define SAMPLING_READ_SIZE	(16)652#define SAMPLING_INTERVAL	(256)653 654/*655 * For statistical analysis of the input data we consider bytes that form a656 * Galois Field of 256 objects. Each object has an attribute count, ie. how657 * many times the object appeared in the sample.658 */659#define BUCKET_SIZE		(256)660 661/*662 * The size of the sample is based on a statistical sampling rule of thumb.663 * The common way is to perform sampling tests as long as the number of664 * elements in each cell is at least 5.665 *666 * Instead of 5, we choose 32 to obtain more accurate results.667 * If the data contain the maximum number of symbols, which is 256, we obtain a668 * sample size bound by 8192.669 *670 * For a sample of at most 8KB of data per data range: 16 consecutive bytes671 * from up to 512 locations.672 */673#define MAX_SAMPLE_SIZE		(BTRFS_MAX_UNCOMPRESSED *		\674				 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)675 676struct bucket_item {677	u32 count;678};679 680struct heuristic_ws {681	/* Partial copy of input data */682	u8 *sample;683	u32 sample_size;684	/* Buckets store counters for each byte value */685	struct bucket_item *bucket;686	/* Sorting buffer */687	struct bucket_item *bucket_b;688	struct list_head list;689};690 691static struct workspace_manager heuristic_wsm;692 693static void free_heuristic_ws(struct list_head *ws)694{695	struct heuristic_ws *workspace;696 697	workspace = list_entry(ws, struct heuristic_ws, list);698 699	kvfree(workspace->sample);700	kfree(workspace->bucket);701	kfree(workspace->bucket_b);702	kfree(workspace);703}704 705static struct list_head *alloc_heuristic_ws(unsigned int level)706{707	struct heuristic_ws *ws;708 709	ws = kzalloc(sizeof(*ws), GFP_KERNEL);710	if (!ws)711		return ERR_PTR(-ENOMEM);712 713	ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);714	if (!ws->sample)715		goto fail;716 717	ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);718	if (!ws->bucket)719		goto fail;720 721	ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);722	if (!ws->bucket_b)723		goto fail;724 725	INIT_LIST_HEAD(&ws->list);726	return &ws->list;727fail:728	free_heuristic_ws(&ws->list);729	return ERR_PTR(-ENOMEM);730}731 732const struct btrfs_compress_op btrfs_heuristic_compress = {733	.workspace_manager = &heuristic_wsm,734};735 736static const struct btrfs_compress_op * const btrfs_compress_op[] = {737	/* The heuristic is represented as compression type 0 */738	&btrfs_heuristic_compress,739	&btrfs_zlib_compress,740	&btrfs_lzo_compress,741	&btrfs_zstd_compress,742};743 744static struct list_head *alloc_workspace(int type, unsigned int level)745{746	switch (type) {747	case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);748	case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);749	case BTRFS_COMPRESS_LZO:  return lzo_alloc_workspace(level);750	case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);751	default:752		/*753		 * This can't happen, the type is validated several times754		 * before we get here.755		 */756		BUG();757	}758}759 760static void free_workspace(int type, struct list_head *ws)761{762	switch (type) {763	case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);764	case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);765	case BTRFS_COMPRESS_LZO:  return lzo_free_workspace(ws);766	case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);767	default:768		/*769		 * This can't happen, the type is validated several times770		 * before we get here.771		 */772		BUG();773	}774}775 776static void btrfs_init_workspace_manager(int type)777{778	struct workspace_manager *wsm;779	struct list_head *workspace;780 781	wsm = btrfs_compress_op[type]->workspace_manager;782	INIT_LIST_HEAD(&wsm->idle_ws);783	spin_lock_init(&wsm->ws_lock);784	atomic_set(&wsm->total_ws, 0);785	init_waitqueue_head(&wsm->ws_wait);786 787	/*788	 * Preallocate one workspace for each compression type so we can789	 * guarantee forward progress in the worst case790	 */791	workspace = alloc_workspace(type, 0);792	if (IS_ERR(workspace)) {793		pr_warn(794	"BTRFS: cannot preallocate compression workspace, will try later\n");795	} else {796		atomic_set(&wsm->total_ws, 1);797		wsm->free_ws = 1;798		list_add(workspace, &wsm->idle_ws);799	}800}801 802static void btrfs_cleanup_workspace_manager(int type)803{804	struct workspace_manager *wsman;805	struct list_head *ws;806 807	wsman = btrfs_compress_op[type]->workspace_manager;808	while (!list_empty(&wsman->idle_ws)) {809		ws = wsman->idle_ws.next;810		list_del(ws);811		free_workspace(type, ws);812		atomic_dec(&wsman->total_ws);813	}814}815 816/*817 * This finds an available workspace or allocates a new one.818 * If it's not possible to allocate a new one, waits until there's one.819 * Preallocation makes a forward progress guarantees and we do not return820 * errors.821 */822struct list_head *btrfs_get_workspace(int type, unsigned int level)823{824	struct workspace_manager *wsm;825	struct list_head *workspace;826	int cpus = num_online_cpus();827	unsigned nofs_flag;828	struct list_head *idle_ws;829	spinlock_t *ws_lock;830	atomic_t *total_ws;831	wait_queue_head_t *ws_wait;832	int *free_ws;833 834	wsm = btrfs_compress_op[type]->workspace_manager;835	idle_ws	 = &wsm->idle_ws;836	ws_lock	 = &wsm->ws_lock;837	total_ws = &wsm->total_ws;838	ws_wait	 = &wsm->ws_wait;839	free_ws	 = &wsm->free_ws;840 841again:842	spin_lock(ws_lock);843	if (!list_empty(idle_ws)) {844		workspace = idle_ws->next;845		list_del(workspace);846		(*free_ws)--;847		spin_unlock(ws_lock);848		return workspace;849 850	}851	if (atomic_read(total_ws) > cpus) {852		DEFINE_WAIT(wait);853 854		spin_unlock(ws_lock);855		prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);856		if (atomic_read(total_ws) > cpus && !*free_ws)857			schedule();858		finish_wait(ws_wait, &wait);859		goto again;860	}861	atomic_inc(total_ws);862	spin_unlock(ws_lock);863 864	/*865	 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have866	 * to turn it off here because we might get called from the restricted867	 * context of btrfs_compress_bio/btrfs_compress_pages868	 */869	nofs_flag = memalloc_nofs_save();870	workspace = alloc_workspace(type, level);871	memalloc_nofs_restore(nofs_flag);872 873	if (IS_ERR(workspace)) {874		atomic_dec(total_ws);875		wake_up(ws_wait);876 877		/*878		 * Do not return the error but go back to waiting. There's a879		 * workspace preallocated for each type and the compression880		 * time is bounded so we get to a workspace eventually. This881		 * makes our caller's life easier.882		 *883		 * To prevent silent and low-probability deadlocks (when the884		 * initial preallocation fails), check if there are any885		 * workspaces at all.886		 */887		if (atomic_read(total_ws) == 0) {888			static DEFINE_RATELIMIT_STATE(_rs,889					/* once per minute */ 60 * HZ,890					/* no burst */ 1);891 892			if (__ratelimit(&_rs)) {893				pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");894			}895		}896		goto again;897	}898	return workspace;899}900 901static struct list_head *get_workspace(int type, int level)902{903	switch (type) {904	case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);905	case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);906	case BTRFS_COMPRESS_LZO:  return btrfs_get_workspace(type, level);907	case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);908	default:909		/*910		 * This can't happen, the type is validated several times911		 * before we get here.912		 */913		BUG();914	}915}916 917/*918 * put a workspace struct back on the list or free it if we have enough919 * idle ones sitting around920 */921void btrfs_put_workspace(int type, struct list_head *ws)922{923	struct workspace_manager *wsm;924	struct list_head *idle_ws;925	spinlock_t *ws_lock;926	atomic_t *total_ws;927	wait_queue_head_t *ws_wait;928	int *free_ws;929 930	wsm = btrfs_compress_op[type]->workspace_manager;931	idle_ws	 = &wsm->idle_ws;932	ws_lock	 = &wsm->ws_lock;933	total_ws = &wsm->total_ws;934	ws_wait	 = &wsm->ws_wait;935	free_ws	 = &wsm->free_ws;936 937	spin_lock(ws_lock);938	if (*free_ws <= num_online_cpus()) {939		list_add(ws, idle_ws);940		(*free_ws)++;941		spin_unlock(ws_lock);942		goto wake;943	}944	spin_unlock(ws_lock);945 946	free_workspace(type, ws);947	atomic_dec(total_ws);948wake:949	cond_wake_up(ws_wait);950}951 952static void put_workspace(int type, struct list_head *ws)953{954	switch (type) {955	case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);956	case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);957	case BTRFS_COMPRESS_LZO:  return btrfs_put_workspace(type, ws);958	case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);959	default:960		/*961		 * This can't happen, the type is validated several times962		 * before we get here.963		 */964		BUG();965	}966}967 968/*969 * Adjust @level according to the limits of the compression algorithm or970 * fallback to default971 */972static unsigned int btrfs_compress_set_level(int type, unsigned level)973{974	const struct btrfs_compress_op *ops = btrfs_compress_op[type];975 976	if (level == 0)977		level = ops->default_level;978	else979		level = min(level, ops->max_level);980 981	return level;982}983 984/* Wrapper around find_get_page(), with extra error message. */985int btrfs_compress_filemap_get_folio(struct address_space *mapping, u64 start,986				     struct folio **in_folio_ret)987{988	struct folio *in_folio;989 990	/*991	 * The compressed write path should have the folio locked already, thus992	 * we only need to grab one reference.993	 */994	in_folio = filemap_get_folio(mapping, start >> PAGE_SHIFT);995	if (IS_ERR(in_folio)) {996		struct btrfs_inode *inode = BTRFS_I(mapping->host);997 998		btrfs_crit(inode->root->fs_info,999		"failed to get page cache, root %lld ino %llu file offset %llu",1000			   btrfs_root_id(inode->root), btrfs_ino(inode), start);1001		return -ENOENT;1002	}1003	*in_folio_ret = in_folio;1004	return 0;1005}1006 1007/*1008 * Given an address space and start and length, compress the bytes into @pages1009 * that are allocated on demand.1010 *1011 * @type_level is encoded algorithm and level, where level 0 means whatever1012 * default the algorithm chooses and is opaque here;1013 * - compression algo are 0-31014 * - the level are bits 4-71015 *1016 * @out_pages is an in/out parameter, holds maximum number of pages to allocate1017 * and returns number of actually allocated pages1018 *1019 * @total_in is used to return the number of bytes actually read.  It1020 * may be smaller than the input length if we had to exit early because we1021 * ran out of room in the pages array or because we cross the1022 * max_out threshold.1023 *1024 * @total_out is an in/out parameter, must be set to the input length and will1025 * be also used to return the total number of compressed bytes1026 */1027int btrfs_compress_folios(unsigned int type_level, struct address_space *mapping,1028			 u64 start, struct folio **folios, unsigned long *out_folios,1029			 unsigned long *total_in, unsigned long *total_out)1030{1031	int type = btrfs_compress_type(type_level);1032	int level = btrfs_compress_level(type_level);1033	struct list_head *workspace;1034	int ret;1035 1036	level = btrfs_compress_set_level(type, level);1037	workspace = get_workspace(type, level);1038	ret = compression_compress_pages(type, workspace, mapping, start, folios,1039					 out_folios, total_in, total_out);1040	put_workspace(type, workspace);1041	return ret;1042}1043 1044static int btrfs_decompress_bio(struct compressed_bio *cb)1045{1046	struct list_head *workspace;1047	int ret;1048	int type = cb->compress_type;1049 1050	workspace = get_workspace(type, 0);1051	ret = compression_decompress_bio(workspace, cb);1052	put_workspace(type, workspace);1053 1054	if (!ret)1055		zero_fill_bio(&cb->orig_bbio->bio);1056	return ret;1057}1058 1059/*1060 * a less complex decompression routine.  Our compressed data fits in a1061 * single page, and we want to read a single page out of it.1062 * start_byte tells us the offset into the compressed data we're interested in1063 */1064int btrfs_decompress(int type, const u8 *data_in, struct folio *dest_folio,1065		     unsigned long dest_pgoff, size_t srclen, size_t destlen)1066{1067	struct btrfs_fs_info *fs_info = folio_to_fs_info(dest_folio);1068	struct list_head *workspace;1069	const u32 sectorsize = fs_info->sectorsize;1070	int ret;1071 1072	/*1073	 * The full destination page range should not exceed the page size.1074	 * And the @destlen should not exceed sectorsize, as this is only called for1075	 * inline file extents, which should not exceed sectorsize.1076	 */1077	ASSERT(dest_pgoff + destlen <= PAGE_SIZE && destlen <= sectorsize);1078 1079	workspace = get_workspace(type, 0);1080	ret = compression_decompress(type, workspace, data_in, dest_folio,1081				     dest_pgoff, srclen, destlen);1082	put_workspace(type, workspace);1083 1084	return ret;1085}1086 1087int __init btrfs_init_compress(void)1088{1089	if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,1090			offsetof(struct compressed_bio, bbio.bio),1091			BIOSET_NEED_BVECS))1092		return -ENOMEM;1093 1094	compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");1095	if (!compr_pool.shrinker)1096		return -ENOMEM;1097 1098	btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);1099	btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);1100	btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);1101	zstd_init_workspace_manager();1102 1103	spin_lock_init(&compr_pool.lock);1104	INIT_LIST_HEAD(&compr_pool.list);1105	compr_pool.count = 0;1106	/* 128K / 4K = 32, for 8 threads is 256 pages. */1107	compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;1108	compr_pool.shrinker->count_objects = btrfs_compr_pool_count;1109	compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;1110	compr_pool.shrinker->batch = 32;1111	compr_pool.shrinker->seeks = DEFAULT_SEEKS;1112	shrinker_register(compr_pool.shrinker);1113 1114	return 0;1115}1116 1117void __cold btrfs_exit_compress(void)1118{1119	/* For now scan drains all pages and does not touch the parameters. */1120	btrfs_compr_pool_scan(NULL, NULL);1121	shrinker_free(compr_pool.shrinker);1122 1123	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);1124	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);1125	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);1126	zstd_cleanup_workspace_manager();1127	bioset_exit(&btrfs_compressed_bioset);1128}1129 1130/*1131 * Copy decompressed data from working buffer to pages.1132 *1133 * @buf:		The decompressed data buffer1134 * @buf_len:		The decompressed data length1135 * @decompressed:	Number of bytes that are already decompressed inside the1136 * 			compressed extent1137 * @cb:			The compressed extent descriptor1138 * @orig_bio:		The original bio that the caller wants to read for1139 *1140 * An easier to understand graph is like below:1141 *1142 * 		|<- orig_bio ->|     |<- orig_bio->|1143 * 	|<-------      full decompressed extent      ----->|1144 * 	|<-----------    @cb range   ---->|1145 * 	|			|<-- @buf_len -->|1146 * 	|<--- @decompressed --->|1147 *1148 * Note that, @cb can be a subpage of the full decompressed extent, but1149 * @cb->start always has the same as the orig_file_offset value of the full1150 * decompressed extent.1151 *1152 * When reading compressed extent, we have to read the full compressed extent,1153 * while @orig_bio may only want part of the range.1154 * Thus this function will ensure only data covered by @orig_bio will be copied1155 * to.1156 *1157 * Return 0 if we have copied all needed contents for @orig_bio.1158 * Return >0 if we need continue decompress.1159 */1160int btrfs_decompress_buf2page(const char *buf, u32 buf_len,1161			      struct compressed_bio *cb, u32 decompressed)1162{1163	struct bio *orig_bio = &cb->orig_bbio->bio;1164	/* Offset inside the full decompressed extent */1165	u32 cur_offset;1166 1167	cur_offset = decompressed;1168	/* The main loop to do the copy */1169	while (cur_offset < decompressed + buf_len) {1170		struct bio_vec bvec;1171		size_t copy_len;1172		u32 copy_start;1173		/* Offset inside the full decompressed extent */1174		u32 bvec_offset;1175 1176		bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);1177		/*1178		 * cb->start may underflow, but subtracting that value can still1179		 * give us correct offset inside the full decompressed extent.1180		 */1181		bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;1182 1183		/* Haven't reached the bvec range, exit */1184		if (decompressed + buf_len <= bvec_offset)1185			return 1;1186 1187		copy_start = max(cur_offset, bvec_offset);1188		copy_len = min(bvec_offset + bvec.bv_len,1189			       decompressed + buf_len) - copy_start;1190		ASSERT(copy_len);1191 1192		/*1193		 * Extra range check to ensure we didn't go beyond1194		 * @buf + @buf_len.1195		 */1196		ASSERT(copy_start - decompressed < buf_len);1197		memcpy_to_page(bvec.bv_page, bvec.bv_offset,1198			       buf + copy_start - decompressed, copy_len);1199		cur_offset += copy_len;1200 1201		bio_advance(orig_bio, copy_len);1202		/* Finished the bio */1203		if (!orig_bio->bi_iter.bi_size)1204			return 0;1205	}1206	return 1;1207}1208 1209/*1210 * Shannon Entropy calculation1211 *1212 * Pure byte distribution analysis fails to determine compressibility of data.1213 * Try calculating entropy to estimate the average minimum number of bits1214 * needed to encode the sampled data.1215 *1216 * For convenience, return the percentage of needed bits, instead of amount of1217 * bits directly.1218 *1219 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy1220 *			    and can be compressible with high probability1221 *1222 * @ENTROPY_LVL_HIGH - data are not compressible with high probability1223 *1224 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.1225 */1226#define ENTROPY_LVL_ACEPTABLE		(65)1227#define ENTROPY_LVL_HIGH		(80)1228 1229/*1230 * For increasead precision in shannon_entropy calculation,1231 * let's do pow(n, M) to save more digits after comma:1232 *1233 * - maximum int bit length is 641234 * - ilog2(MAX_SAMPLE_SIZE)	-> 131235 * - 13 * 4 = 52 < 64		-> M = 41236 *1237 * So use pow(n, 4).1238 */1239static inline u32 ilog2_w(u64 n)1240{1241	return ilog2(n * n * n * n);1242}1243 1244static u32 shannon_entropy(struct heuristic_ws *ws)1245{1246	const u32 entropy_max = 8 * ilog2_w(2);1247	u32 entropy_sum = 0;1248	u32 p, p_base, sz_base;1249	u32 i;1250 1251	sz_base = ilog2_w(ws->sample_size);1252	for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {1253		p = ws->bucket[i].count;1254		p_base = ilog2_w(p);1255		entropy_sum += p * (sz_base - p_base);1256	}1257 1258	entropy_sum /= ws->sample_size;1259	return entropy_sum * 100 / entropy_max;1260}1261 1262#define RADIX_BASE		4U1263#define COUNTERS_SIZE		(1U << RADIX_BASE)1264 1265static u8 get4bits(u64 num, int shift) {1266	u8 low4bits;1267 1268	num >>= shift;1269	/* Reverse order */1270	low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);1271	return low4bits;1272}1273 1274/*1275 * Use 4 bits as radix base1276 * Use 16 u32 counters for calculating new position in buf array1277 *1278 * @array     - array that will be sorted1279 * @array_buf - buffer array to store sorting results1280 *              must be equal in size to @array1281 * @num       - array size1282 */1283static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,1284		       int num)1285{1286	u64 max_num;1287	u64 buf_num;1288	u32 counters[COUNTERS_SIZE];1289	u32 new_addr;1290	u32 addr;1291	int bitlen;1292	int shift;1293	int i;1294 1295	/*1296	 * Try avoid useless loop iterations for small numbers stored in big1297	 * counters.  Example: 48 33 4 ... in 64bit array1298	 */1299	max_num = array[0].count;1300	for (i = 1; i < num; i++) {1301		buf_num = array[i].count;1302		if (buf_num > max_num)1303			max_num = buf_num;1304	}1305 1306	buf_num = ilog2(max_num);1307	bitlen = ALIGN(buf_num, RADIX_BASE * 2);1308 1309	shift = 0;1310	while (shift < bitlen) {1311		memset(counters, 0, sizeof(counters));1312 1313		for (i = 0; i < num; i++) {1314			buf_num = array[i].count;1315			addr = get4bits(buf_num, shift);1316			counters[addr]++;1317		}1318 1319		for (i = 1; i < COUNTERS_SIZE; i++)1320			counters[i] += counters[i - 1];1321 1322		for (i = num - 1; i >= 0; i--) {1323			buf_num = array[i].count;1324			addr = get4bits(buf_num, shift);1325			counters[addr]--;1326			new_addr = counters[addr];1327			array_buf[new_addr] = array[i];1328		}1329 1330		shift += RADIX_BASE;1331 1332		/*1333		 * Normal radix expects to move data from a temporary array, to1334		 * the main one.  But that requires some CPU time. Avoid that1335		 * by doing another sort iteration to original array instead of1336		 * memcpy()1337		 */1338		memset(counters, 0, sizeof(counters));1339 1340		for (i = 0; i < num; i ++) {1341			buf_num = array_buf[i].count;1342			addr = get4bits(buf_num, shift);1343			counters[addr]++;1344		}1345 1346		for (i = 1; i < COUNTERS_SIZE; i++)1347			counters[i] += counters[i - 1];1348 1349		for (i = num - 1; i >= 0; i--) {1350			buf_num = array_buf[i].count;1351			addr = get4bits(buf_num, shift);1352			counters[addr]--;1353			new_addr = counters[addr];1354			array[new_addr] = array_buf[i];1355		}1356 1357		shift += RADIX_BASE;1358	}1359}1360 1361/*1362 * Size of the core byte set - how many bytes cover 90% of the sample1363 *1364 * There are several types of structured binary data that use nearly all byte1365 * values. The distribution can be uniform and counts in all buckets will be1366 * nearly the same (eg. encrypted data). Unlikely to be compressible.1367 *1368 * Other possibility is normal (Gaussian) distribution, where the data could1369 * be potentially compressible, but we have to take a few more steps to decide1370 * how much.1371 *1372 * @BYTE_CORE_SET_LOW  - main part of byte values repeated frequently,1373 *                       compression algo can easy fix that1374 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high1375 *                       probability is not compressible1376 */1377#define BYTE_CORE_SET_LOW		(64)1378#define BYTE_CORE_SET_HIGH		(200)1379 1380static int byte_core_set_size(struct heuristic_ws *ws)1381{1382	u32 i;1383	u32 coreset_sum = 0;1384	const u32 core_set_threshold = ws->sample_size * 90 / 100;1385	struct bucket_item *bucket = ws->bucket;1386 1387	/* Sort in reverse order */1388	radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);1389 1390	for (i = 0; i < BYTE_CORE_SET_LOW; i++)1391		coreset_sum += bucket[i].count;1392 1393	if (coreset_sum > core_set_threshold)1394		return i;1395 1396	for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {1397		coreset_sum += bucket[i].count;1398		if (coreset_sum > core_set_threshold)1399			break;1400	}1401 1402	return i;1403}1404 1405/*1406 * Count byte values in buckets.1407 * This heuristic can detect textual data (configs, xml, json, html, etc).1408 * Because in most text-like data byte set is restricted to limited number of1409 * possible characters, and that restriction in most cases makes data easy to1410 * compress.1411 *1412 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:1413 *	less - compressible1414 *	more - need additional analysis1415 */1416#define BYTE_SET_THRESHOLD		(64)1417 1418static u32 byte_set_size(const struct heuristic_ws *ws)1419{1420	u32 i;1421	u32 byte_set_size = 0;1422 1423	for (i = 0; i < BYTE_SET_THRESHOLD; i++) {1424		if (ws->bucket[i].count > 0)1425			byte_set_size++;1426	}1427 1428	/*1429	 * Continue collecting count of byte values in buckets.  If the byte1430	 * set size is bigger then the threshold, it's pointless to continue,1431	 * the detection technique would fail for this type of data.1432	 */1433	for (; i < BUCKET_SIZE; i++) {1434		if (ws->bucket[i].count > 0) {1435			byte_set_size++;1436			if (byte_set_size > BYTE_SET_THRESHOLD)1437				return byte_set_size;1438		}1439	}1440 1441	return byte_set_size;1442}1443 1444static bool sample_repeated_patterns(struct heuristic_ws *ws)1445{1446	const u32 half_of_sample = ws->sample_size / 2;1447	const u8 *data = ws->sample;1448 1449	return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;1450}1451 1452static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,1453				     struct heuristic_ws *ws)1454{1455	struct page *page;1456	u64 index, index_end;1457	u32 i, curr_sample_pos;1458	u8 *in_data;1459 1460	/*1461	 * Compression handles the input data by chunks of 128KiB1462	 * (defined by BTRFS_MAX_UNCOMPRESSED)1463	 *1464	 * We do the same for the heuristic and loop over the whole range.1465	 *1466	 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will1467	 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.1468	 */1469	if (end - start > BTRFS_MAX_UNCOMPRESSED)1470		end = start + BTRFS_MAX_UNCOMPRESSED;1471 1472	index = start >> PAGE_SHIFT;1473	index_end = end >> PAGE_SHIFT;1474 1475	/* Don't miss unaligned end */1476	if (!PAGE_ALIGNED(end))1477		index_end++;1478 1479	curr_sample_pos = 0;1480	while (index < index_end) {1481		page = find_get_page(inode->i_mapping, index);1482		in_data = kmap_local_page(page);1483		/* Handle case where the start is not aligned to PAGE_SIZE */1484		i = start % PAGE_SIZE;1485		while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {1486			/* Don't sample any garbage from the last page */1487			if (start > end - SAMPLING_READ_SIZE)1488				break;1489			memcpy(&ws->sample[curr_sample_pos], &in_data[i],1490					SAMPLING_READ_SIZE);1491			i += SAMPLING_INTERVAL;1492			start += SAMPLING_INTERVAL;1493			curr_sample_pos += SAMPLING_READ_SIZE;1494		}1495		kunmap_local(in_data);1496		put_page(page);1497 1498		index++;1499	}1500 1501	ws->sample_size = curr_sample_pos;1502}1503 1504/*1505 * Compression heuristic.1506 *1507 * The following types of analysis can be performed:1508 * - detect mostly zero data1509 * - detect data with low "byte set" size (text, etc)1510 * - detect data with low/high "core byte" set1511 *1512 * Return non-zero if the compression should be done, 0 otherwise.1513 */1514int btrfs_compress_heuristic(struct btrfs_inode *inode, u64 start, u64 end)1515{1516	struct list_head *ws_list = get_workspace(0, 0);1517	struct heuristic_ws *ws;1518	u32 i;1519	u8 byte;1520	int ret = 0;1521 1522	ws = list_entry(ws_list, struct heuristic_ws, list);1523 1524	heuristic_collect_sample(&inode->vfs_inode, start, end, ws);1525 1526	if (sample_repeated_patterns(ws)) {1527		ret = 1;1528		goto out;1529	}1530 1531	memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);1532 1533	for (i = 0; i < ws->sample_size; i++) {1534		byte = ws->sample[i];1535		ws->bucket[byte].count++;1536	}1537 1538	i = byte_set_size(ws);1539	if (i < BYTE_SET_THRESHOLD) {1540		ret = 2;1541		goto out;1542	}1543 1544	i = byte_core_set_size(ws);1545	if (i <= BYTE_CORE_SET_LOW) {1546		ret = 3;1547		goto out;1548	}1549 1550	if (i >= BYTE_CORE_SET_HIGH) {1551		ret = 0;1552		goto out;1553	}1554 1555	i = shannon_entropy(ws);1556	if (i <= ENTROPY_LVL_ACEPTABLE) {1557		ret = 4;1558		goto out;1559	}1560 1561	/*1562	 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be1563	 * needed to give green light to compression.1564	 *1565	 * For now just assume that compression at that level is not worth the1566	 * resources because:1567	 *1568	 * 1. it is possible to defrag the data later1569	 *1570	 * 2. the data would turn out to be hardly compressible, eg. 150 byte1571	 * values, every bucket has counter at level ~54. The heuristic would1572	 * be confused. This can happen when data have some internal repeated1573	 * patterns like "abbacbbc...". This can be detected by analyzing1574	 * pairs of bytes, which is too costly.1575	 */1576	if (i < ENTROPY_LVL_HIGH) {1577		ret = 5;1578		goto out;1579	} else {1580		ret = 0;1581		goto out;1582	}1583 1584out:1585	put_workspace(0, ws_list);1586	return ret;1587}1588 1589/*1590 * Convert the compression suffix (eg. after "zlib" starting with ":") to1591 * level, unrecognized string will set the default level1592 */1593unsigned int btrfs_compress_str2level(unsigned int type, const char *str)1594{1595	unsigned int level = 0;1596	int ret;1597 1598	if (!type)1599		return 0;1600 1601	if (str[0] == ':') {1602		ret = kstrtouint(str + 1, 10, &level);1603		if (ret)1604			level = 0;1605	}1606 1607	level = btrfs_compress_set_level(type, level);1608 1609	return level;1610}1611