compression.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   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, start, end_io, NULL);
  71	return to_compressed_bio(bbio);
  72}
  73
  74bool btrfs_compress_is_valid_type(const char *str, size_t len)
  75{
  76	int i;
  77
  78	for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
  79		size_t comp_len = strlen(btrfs_compress_types[i]);
  80
  81		if (len < comp_len)
  82			continue;
  83
  84		if (!strncmp(btrfs_compress_types[i], str, comp_len))
  85			return true;
  86	}
  87	return false;
  88}
  89
  90static int compression_compress_pages(int type, struct list_head *ws,
  91				      struct btrfs_inode *inode, u64 start,
  92				      struct folio **folios, unsigned long *out_folios,
  93				      unsigned long *total_in, unsigned long *total_out)
  94{
  95	switch (type) {
  96	case BTRFS_COMPRESS_ZLIB:
  97		return zlib_compress_folios(ws, inode, start, folios,
  98					    out_folios, total_in, total_out);
  99	case BTRFS_COMPRESS_LZO:
 100		return lzo_compress_folios(ws, inode, start, folios,
 101					   out_folios, total_in, total_out);
 102	case BTRFS_COMPRESS_ZSTD:
 103		return zstd_compress_folios(ws, inode, start, folios,
 104					    out_folios, total_in, total_out);
 105	case BTRFS_COMPRESS_NONE:
 106	default:
 107		/*
 108		 * This can happen when compression races with remount setting
 109		 * it to 'no compress', while caller doesn't call
 110		 * inode_need_compress() to check if we really need to
 111		 * compress.
 112		 *
 113		 * Not a big deal, just need to inform caller that we
 114		 * haven't allocated any pages yet.
 115		 */
 116		*out_folios = 0;
 117		return -E2BIG;
 118	}
 119}
 120
 121static int compression_decompress_bio(struct list_head *ws,
 122				      struct compressed_bio *cb)
 123{
 124	switch (cb->compress_type) {
 125	case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
 126	case BTRFS_COMPRESS_LZO:  return lzo_decompress_bio(ws, cb);
 127	case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
 128	case BTRFS_COMPRESS_NONE:
 129	default:
 130		/*
 131		 * This can't happen, the type is validated several times
 132		 * before we get here.
 133		 */
 134		BUG();
 135	}
 136}
 137
 138static int compression_decompress(int type, struct list_head *ws,
 139		const u8 *data_in, struct folio *dest_folio,
 140		unsigned long dest_pgoff, size_t srclen, size_t destlen)
 141{
 142	switch (type) {
 143	case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_folio,
 144						dest_pgoff, srclen, destlen);
 145	case BTRFS_COMPRESS_LZO:  return lzo_decompress(ws, data_in, dest_folio,
 146						dest_pgoff, srclen, destlen);
 147	case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_folio,
 148						dest_pgoff, srclen, destlen);
 149	case BTRFS_COMPRESS_NONE:
 150	default:
 151		/*
 152		 * This can't happen, the type is validated several times
 153		 * before we get here.
 154		 */
 155		BUG();
 156	}
 157}
 158
 159static void btrfs_free_compressed_folios(struct compressed_bio *cb)
 160{
 161	for (unsigned int i = 0; i < cb->nr_folios; i++)
 162		btrfs_free_compr_folio(cb->compressed_folios[i]);
 163	kfree(cb->compressed_folios);
 164}
 165
 166static int btrfs_decompress_bio(struct compressed_bio *cb);
 167
 168/*
 169 * Global cache of last unused pages for compression/decompression.
 170 */
 171static struct btrfs_compr_pool {
 172	struct shrinker *shrinker;
 173	spinlock_t lock;
 174	struct list_head list;
 175	int count;
 176	int thresh;
 177} compr_pool;
 178
 179static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)
 180{
 181	int ret;
 182
 183	/*
 184	 * We must not read the values more than once if 'ret' gets expanded in
 185	 * the return statement so we don't accidentally return a negative
 186	 * number, even if the first condition finds it positive.
 187	 */
 188	ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);
 189
 190	return ret > 0 ? ret : 0;
 191}
 192
 193static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)
 194{
 195	LIST_HEAD(remove);
 196	struct list_head *tmp, *next;
 197	int freed;
 198
 199	if (compr_pool.count == 0)
 200		return SHRINK_STOP;
 201
 202	/* For now, just simply drain the whole list. */
 203	spin_lock(&compr_pool.lock);
 204	list_splice_init(&compr_pool.list, &remove);
 205	freed = compr_pool.count;
 206	compr_pool.count = 0;
 207	spin_unlock(&compr_pool.lock);
 208
 209	list_for_each_safe(tmp, next, &remove) {
 210		struct page *page = list_entry(tmp, struct page, lru);
 211
 212		ASSERT(page_ref_count(page) == 1);
 213		put_page(page);
 214	}
 215
 216	return freed;
 217}
 218
 219/*
 220 * Common wrappers for page allocation from compression wrappers
 221 */
 222struct folio *btrfs_alloc_compr_folio(struct btrfs_fs_info *fs_info)
 223{
 224	struct folio *folio = NULL;
 225
 226	/* For bs > ps cases, no cached folio pool for now. */
 227	if (fs_info->block_min_order)
 228		goto alloc;
 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
 241alloc:
 242	return folio_alloc(GFP_NOFS, fs_info->block_min_order);
 243}
 244
 245void btrfs_free_compr_folio(struct folio *folio)
 246{
 247	bool do_free = false;
 248
 249	/* The folio is from bs > ps fs, no cached pool for now. */
 250	if (folio_order(folio))
 251		goto free;
 252
 253	spin_lock(&compr_pool.lock);
 254	if (compr_pool.count > compr_pool.thresh) {
 255		do_free = true;
 256	} else {
 257		list_add(&folio->lru, &compr_pool.list);
 258		compr_pool.count++;
 259	}
 260	spin_unlock(&compr_pool.lock);
 261
 262	if (!do_free)
 263		return;
 264
 265free:
 266	ASSERT(folio_ref_count(folio) == 1);
 267	folio_put(folio);
 268}
 269
 270static void end_bbio_compressed_read(struct btrfs_bio *bbio)
 271{
 272	struct compressed_bio *cb = to_compressed_bio(bbio);
 273	blk_status_t status = bbio->bio.bi_status;
 274
 275	if (!status)
 276		status = errno_to_blk_status(btrfs_decompress_bio(cb));
 277
 278	btrfs_free_compressed_folios(cb);
 279	btrfs_bio_end_io(cb->orig_bbio, status);
 280	bio_put(&bbio->bio);
 281}
 282
 283/*
 284 * Clear the writeback bits on all of the file
 285 * pages for a compressed write
 286 */
 287static noinline void end_compressed_writeback(const struct compressed_bio *cb)
 288{
 289	struct inode *inode = &cb->bbio.inode->vfs_inode;
 290	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
 291	pgoff_t index = cb->start >> PAGE_SHIFT;
 292	const pgoff_t end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
 293	struct folio_batch fbatch;
 294	int i;
 295	int ret;
 296
 297	ret = blk_status_to_errno(cb->bbio.bio.bi_status);
 298	if (ret)
 299		mapping_set_error(inode->i_mapping, ret);
 300
 301	folio_batch_init(&fbatch);
 302	while (index <= end_index) {
 303		ret = filemap_get_folios(inode->i_mapping, &index, end_index,
 304				&fbatch);
 305
 306		if (ret == 0)
 307			return;
 308
 309		for (i = 0; i < ret; i++) {
 310			struct folio *folio = fbatch.folios[i];
 311
 312			btrfs_folio_clamp_clear_writeback(fs_info, folio,
 313							  cb->start, cb->len);
 314		}
 315		folio_batch_release(&fbatch);
 316	}
 317	/* the inode may be gone now */
 318}
 319
 320/*
 321 * Do the cleanup once all the compressed pages hit the disk.  This will clear
 322 * writeback on the file pages and free the compressed pages.
 323 *
 324 * This also calls the writeback end hooks for the file pages so that metadata
 325 * and checksums can be updated in the file.
 326 */
 327static void end_bbio_compressed_write(struct btrfs_bio *bbio)
 328{
 329	struct compressed_bio *cb = to_compressed_bio(bbio);
 330
 331	btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
 332				    cb->bbio.bio.bi_status == BLK_STS_OK);
 333
 334	if (cb->writeback)
 335		end_compressed_writeback(cb);
 336	/* Note, our inode could be gone now. */
 337	btrfs_free_compressed_folios(cb);
 338	bio_put(&cb->bbio.bio);
 339}
 340
 341static void btrfs_add_compressed_bio_folios(struct compressed_bio *cb)
 342{
 343	struct bio *bio = &cb->bbio.bio;
 344	u32 offset = 0;
 345	unsigned int findex = 0;
 346
 347	while (offset < cb->compressed_len) {
 348		struct folio *folio = cb->compressed_folios[findex];
 349		u32 len = min_t(u32, cb->compressed_len - offset, folio_size(folio));
 350		int ret;
 351
 352		/* Maximum compressed extent is smaller than bio size limit. */
 353		ret = bio_add_folio(bio, folio, len, 0);
 354		ASSERT(ret);
 355		offset += len;
 356		findex++;
 357	}
 358}
 359
 360/*
 361 * worker function to build and submit bios for previously compressed pages.
 362 * The corresponding pages in the inode should be marked for writeback
 363 * and the compressed pages should have a reference on them for dropping
 364 * when the IO is complete.
 365 *
 366 * This also checksums the file bytes and gets things ready for
 367 * the end io hooks.
 368 */
 369void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
 370				   struct folio **compressed_folios,
 371				   unsigned int nr_folios,
 372				   blk_opf_t write_flags,
 373				   bool writeback)
 374{
 375	struct btrfs_inode *inode = ordered->inode;
 376	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 377	struct compressed_bio *cb;
 378
 379	ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
 380	ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
 381
 382	cb = alloc_compressed_bio(inode, ordered->file_offset,
 383				  REQ_OP_WRITE | write_flags,
 384				  end_bbio_compressed_write);
 385	cb->start = ordered->file_offset;
 386	cb->len = ordered->num_bytes;
 387	cb->compressed_folios = compressed_folios;
 388	cb->compressed_len = ordered->disk_num_bytes;
 389	cb->writeback = writeback;
 390	cb->nr_folios = nr_folios;
 391	cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
 392	cb->bbio.ordered = ordered;
 393	btrfs_add_compressed_bio_folios(cb);
 394
 395	btrfs_submit_bbio(&cb->bbio, 0);
 396}
 397
 398/*
 399 * Add extra pages in the same compressed file extent so that we don't need to
 400 * re-read the same extent again and again.
 401 *
 402 * NOTE: this won't work well for subpage, as for subpage read, we lock the
 403 * full page then submit bio for each compressed/regular extents.
 404 *
 405 * This means, if we have several sectors in the same page points to the same
 406 * on-disk compressed data, we will re-read the same extent many times and
 407 * this function can only help for the next page.
 408 */
 409static noinline int add_ra_bio_pages(struct inode *inode,
 410				     u64 compressed_end,
 411				     struct compressed_bio *cb,
 412				     int *memstall, unsigned long *pflags)
 413{
 414	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
 415	pgoff_t end_index;
 416	struct bio *orig_bio = &cb->orig_bbio->bio;
 417	u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
 418	u64 isize = i_size_read(inode);
 419	int ret;
 420	struct folio *folio;
 421	struct extent_map *em;
 422	struct address_space *mapping = inode->i_mapping;
 423	struct extent_map_tree *em_tree;
 424	struct extent_io_tree *tree;
 425	int sectors_missed = 0;
 426
 427	em_tree = &BTRFS_I(inode)->extent_tree;
 428	tree = &BTRFS_I(inode)->io_tree;
 429
 430	if (isize == 0)
 431		return 0;
 432
 433	/*
 434	 * For current subpage support, we only support 64K page size,
 435	 * which means maximum compressed extent size (128K) is just 2x page
 436	 * size.
 437	 * This makes readahead less effective, so here disable readahead for
 438	 * subpage for now, until full compressed write is supported.
 439	 */
 440	if (fs_info->sectorsize < PAGE_SIZE)
 441		return 0;
 442
 443	/* For bs > ps cases, we don't support readahead for compressed folios for now. */
 444	if (fs_info->block_min_order)
 445		return 0;
 446
 447	end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 448
 449	while (cur < compressed_end) {
 450		pgoff_t page_end;
 451		pgoff_t pg_index = cur >> PAGE_SHIFT;
 452		u32 add_size;
 453
 454		if (pg_index > end_index)
 455			break;
 456
 457		folio = filemap_get_folio(mapping, pg_index);
 458		if (!IS_ERR(folio)) {
 459			u64 folio_sz = folio_size(folio);
 460			u64 offset = offset_in_folio(folio, cur);
 461
 462			folio_put(folio);
 463			sectors_missed += (folio_sz - offset) >>
 464					  fs_info->sectorsize_bits;
 465
 466			/* Beyond threshold, no need to continue */
 467			if (sectors_missed > 4)
 468				break;
 469
 470			/*
 471			 * Jump to next page start as we already have page for
 472			 * current offset.
 473			 */
 474			cur += (folio_sz - offset);
 475			continue;
 476		}
 477
 478		folio = filemap_alloc_folio(mapping_gfp_constraint(mapping, ~__GFP_FS),
 479					    0, NULL);
 480		if (!folio)
 481			break;
 482
 483		if (filemap_add_folio(mapping, folio, pg_index, GFP_NOFS)) {
 484			/* There is already a page, skip to page end */
 485			cur += folio_size(folio);
 486			folio_put(folio);
 487			continue;
 488		}
 489
 490		if (!*memstall && folio_test_workingset(folio)) {
 491			psi_memstall_enter(pflags);
 492			*memstall = 1;
 493		}
 494
 495		ret = set_folio_extent_mapped(folio);
 496		if (ret < 0) {
 497			folio_unlock(folio);
 498			folio_put(folio);
 499			break;
 500		}
 501
 502		page_end = (pg_index << PAGE_SHIFT) + folio_size(folio) - 1;
 503		btrfs_lock_extent(tree, cur, page_end, NULL);
 504		read_lock(&em_tree->lock);
 505		em = btrfs_lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
 506		read_unlock(&em_tree->lock);
 507
 508		/*
 509		 * At this point, we have a locked page in the page cache for
 510		 * these bytes in the file.  But, we have to make sure they map
 511		 * to this compressed extent on disk.
 512		 */
 513		if (!em || cur < em->start ||
 514		    (cur + fs_info->sectorsize > btrfs_extent_map_end(em)) ||
 515		    (btrfs_extent_map_block_start(em) >> SECTOR_SHIFT) !=
 516		    orig_bio->bi_iter.bi_sector) {
 517			btrfs_free_extent_map(em);
 518			btrfs_unlock_extent(tree, cur, page_end, NULL);
 519			folio_unlock(folio);
 520			folio_put(folio);
 521			break;
 522		}
 523		add_size = min(em->start + em->len, page_end + 1) - cur;
 524		btrfs_free_extent_map(em);
 525		btrfs_unlock_extent(tree, cur, page_end, NULL);
 526
 527		if (folio_contains(folio, end_index)) {
 528			size_t zero_offset = offset_in_folio(folio, isize);
 529
 530			if (zero_offset) {
 531				int zeros;
 532				zeros = folio_size(folio) - zero_offset;
 533				folio_zero_range(folio, zero_offset, zeros);
 534			}
 535		}
 536
 537		if (!bio_add_folio(orig_bio, folio, add_size,
 538				   offset_in_folio(folio, cur))) {
 539			folio_unlock(folio);
 540			folio_put(folio);
 541			break;
 542		}
 543		/*
 544		 * If it's subpage, we also need to increase its
 545		 * subpage::readers number, as at endio we will decrease
 546		 * subpage::readers and to unlock the page.
 547		 */
 548		if (fs_info->sectorsize < PAGE_SIZE)
 549			btrfs_folio_set_lock(fs_info, folio, cur, 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 pages
 558 * in it.  We don't actually do IO on those pages but allocate new ones
 559 * to hold the compressed pages on disk.
 560 *
 561 * bio->bi_iter.bi_sector points to the compressed extent on disk
 562 * bio->bi_io_vec points to all of the inode pages
 563 *
 564 * After the compressed pages are read, we copy the bytes into the
 565 * bio we were passed and then call the bio end_io calls
 566 */
 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 status;
 581	int ret;
 582
 583	/* we need the actual starting offset of this extent in the file */
 584	read_lock(&em_tree->lock);
 585	em = btrfs_lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
 586	read_unlock(&em_tree->lock);
 587	if (!em) {
 588		status = BLK_STS_IOERR;
 589		goto out;
 590	}
 591
 592	ASSERT(btrfs_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 = btrfs_extent_map_compression(em);
 605	cb->orig_bbio = bbio;
 606	cb->bbio.csum_search_commit_root = bbio->csum_search_commit_root;
 607
 608	btrfs_free_extent_map(em);
 609
 610	cb->nr_folios = DIV_ROUND_UP(compressed_len, btrfs_min_folio_size(fs_info));
 611	cb->compressed_folios = kcalloc(cb->nr_folios, sizeof(struct folio *), GFP_NOFS);
 612	if (!cb->compressed_folios) {
 613		status = BLK_STS_RESOURCE;
 614		goto out_free_bio;
 615	}
 616
 617	ret = btrfs_alloc_folio_array(cb->nr_folios, fs_info->block_min_order,
 618				      cb->compressed_folios);
 619	if (ret) {
 620		status = BLK_STS_RESOURCE;
 621		goto out_free_compressed_pages;
 622	}
 623
 624	add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
 625			 &pflags);
 626
 627	/* include any pages we added in add_ra-bio_pages */
 628	cb->len = bbio->bio.bi_iter.bi_size;
 629	cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
 630	btrfs_add_compressed_bio_folios(cb);
 631
 632	if (memstall)
 633		psi_memstall_leave(&pflags);
 634
 635	btrfs_submit_bbio(&cb->bbio, 0);
 636	return;
 637
 638out_free_compressed_pages:
 639	kfree(cb->compressed_folios);
 640out_free_bio:
 641	bio_put(&cb->bbio.bio);
 642out:
 643	btrfs_bio_end_io(bbio, status);
 644}
 645
 646/*
 647 * Heuristic uses systematic sampling to collect data from the input data
 648 * range, the logic can be tuned by the following constants:
 649 *
 650 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
 651 * @SAMPLING_INTERVAL  - range from which the sampled data can be collected
 652 */
 653#define SAMPLING_READ_SIZE	(16)
 654#define SAMPLING_INTERVAL	(256)
 655
 656/*
 657 * For statistical analysis of the input data we consider bytes that form a
 658 * Galois Field of 256 objects. Each object has an attribute count, ie. how
 659 * many times the object appeared in the sample.
 660 */
 661#define BUCKET_SIZE		(256)
 662
 663/*
 664 * The size of the sample is based on a statistical sampling rule of thumb.
 665 * The common way is to perform sampling tests as long as the number of
 666 * elements in each cell is at least 5.
 667 *
 668 * Instead of 5, we choose 32 to obtain more accurate results.
 669 * If the data contain the maximum number of symbols, which is 256, we obtain a
 670 * sample size bound by 8192.
 671 *
 672 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
 673 * from up to 512 locations.
 674 */
 675#define MAX_SAMPLE_SIZE		(BTRFS_MAX_UNCOMPRESSED *		\
 676				 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
 677
 678struct bucket_item {
 679	u32 count;
 680};
 681
 682struct heuristic_ws {
 683	/* Partial copy of input data */
 684	u8 *sample;
 685	u32 sample_size;
 686	/* Buckets store counters for each byte value */
 687	struct bucket_item *bucket;
 688	/* Sorting buffer */
 689	struct bucket_item *bucket_b;
 690	struct list_head list;
 691};
 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(struct btrfs_fs_info *fs_info)
 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_levels btrfs_heuristic_compress = { 0 };
 733
 734static const struct btrfs_compress_levels * const btrfs_compress_levels[] = {
 735	/* The heuristic is represented as compression type 0 */
 736	&btrfs_heuristic_compress,
 737	&btrfs_zlib_compress,
 738	&btrfs_lzo_compress,
 739	&btrfs_zstd_compress,
 740};
 741
 742static struct list_head *alloc_workspace(struct btrfs_fs_info *fs_info, int type, int level)
 743{
 744	switch (type) {
 745	case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(fs_info);
 746	case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(fs_info, level);
 747	case BTRFS_COMPRESS_LZO:  return lzo_alloc_workspace(fs_info);
 748	case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(fs_info, level);
 749	default:
 750		/*
 751		 * This can't happen, the type is validated several times
 752		 * before we get here.
 753		 */
 754		BUG();
 755	}
 756}
 757
 758static void free_workspace(int type, struct list_head *ws)
 759{
 760	switch (type) {
 761	case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
 762	case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
 763	case BTRFS_COMPRESS_LZO:  return lzo_free_workspace(ws);
 764	case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
 765	default:
 766		/*
 767		 * This can't happen, the type is validated several times
 768		 * before we get here.
 769		 */
 770		BUG();
 771	}
 772}
 773
 774static int alloc_workspace_manager(struct btrfs_fs_info *fs_info,
 775				   enum btrfs_compression_type type)
 776{
 777	struct workspace_manager *gwsm;
 778	struct list_head *workspace;
 779
 780	ASSERT(fs_info->compr_wsm[type] == NULL);
 781	gwsm = kzalloc(sizeof(*gwsm), GFP_KERNEL);
 782	if (!gwsm)
 783		return -ENOMEM;
 784
 785	INIT_LIST_HEAD(&gwsm->idle_ws);
 786	spin_lock_init(&gwsm->ws_lock);
 787	atomic_set(&gwsm->total_ws, 0);
 788	init_waitqueue_head(&gwsm->ws_wait);
 789	fs_info->compr_wsm[type] = gwsm;
 790
 791	/*
 792	 * Preallocate one workspace for each compression type so we can
 793	 * guarantee forward progress in the worst case
 794	 */
 795	workspace = alloc_workspace(fs_info, type, 0);
 796	if (IS_ERR(workspace)) {
 797		btrfs_warn(fs_info,
 798	"cannot preallocate compression workspace for %s, will try later",
 799			   btrfs_compress_type2str(type));
 800	} else {
 801		atomic_set(&gwsm->total_ws, 1);
 802		gwsm->free_ws = 1;
 803		list_add(workspace, &gwsm->idle_ws);
 804	}
 805	return 0;
 806}
 807
 808static void free_workspace_manager(struct btrfs_fs_info *fs_info,
 809				   enum btrfs_compression_type type)
 810{
 811	struct list_head *ws;
 812	struct workspace_manager *gwsm = fs_info->compr_wsm[type];
 813
 814	/* ZSTD uses its own workspace manager, should enter here. */
 815	ASSERT(type != BTRFS_COMPRESS_ZSTD && type < BTRFS_NR_COMPRESS_TYPES);
 816	if (!gwsm)
 817		return;
 818	fs_info->compr_wsm[type] = NULL;
 819	while (!list_empty(&gwsm->idle_ws)) {
 820		ws = gwsm->idle_ws.next;
 821		list_del(ws);
 822		free_workspace(type, ws);
 823		atomic_dec(&gwsm->total_ws);
 824	}
 825	kfree(gwsm);
 826}
 827
 828/*
 829 * This finds an available workspace or allocates a new one.
 830 * If it's not possible to allocate a new one, waits until there's one.
 831 * Preallocation makes a forward progress guarantees and we do not return
 832 * errors.
 833 */
 834struct list_head *btrfs_get_workspace(struct btrfs_fs_info *fs_info, int type, int level)
 835{
 836	struct workspace_manager *wsm = fs_info->compr_wsm[type];
 837	struct list_head *workspace;
 838	int cpus = num_online_cpus();
 839	unsigned nofs_flag;
 840	struct list_head *idle_ws;
 841	spinlock_t *ws_lock;
 842	atomic_t *total_ws;
 843	wait_queue_head_t *ws_wait;
 844	int *free_ws;
 845
 846	ASSERT(wsm);
 847	idle_ws	 = &wsm->idle_ws;
 848	ws_lock	 = &wsm->ws_lock;
 849	total_ws = &wsm->total_ws;
 850	ws_wait	 = &wsm->ws_wait;
 851	free_ws	 = &wsm->free_ws;
 852
 853again:
 854	spin_lock(ws_lock);
 855	if (!list_empty(idle_ws)) {
 856		workspace = idle_ws->next;
 857		list_del(workspace);
 858		(*free_ws)--;
 859		spin_unlock(ws_lock);
 860		return workspace;
 861
 862	}
 863	if (atomic_read(total_ws) > cpus) {
 864		DEFINE_WAIT(wait);
 865
 866		spin_unlock(ws_lock);
 867		prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
 868		if (atomic_read(total_ws) > cpus && !*free_ws)
 869			schedule();
 870		finish_wait(ws_wait, &wait);
 871		goto again;
 872	}
 873	atomic_inc(total_ws);
 874	spin_unlock(ws_lock);
 875
 876	/*
 877	 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
 878	 * to turn it off here because we might get called from the restricted
 879	 * context of btrfs_compress_bio/btrfs_compress_pages
 880	 */
 881	nofs_flag = memalloc_nofs_save();
 882	workspace = alloc_workspace(fs_info, type, level);
 883	memalloc_nofs_restore(nofs_flag);
 884
 885	if (IS_ERR(workspace)) {
 886		atomic_dec(total_ws);
 887		wake_up(ws_wait);
 888
 889		/*
 890		 * Do not return the error but go back to waiting. There's a
 891		 * workspace preallocated for each type and the compression
 892		 * time is bounded so we get to a workspace eventually. This
 893		 * makes our caller's life easier.
 894		 *
 895		 * To prevent silent and low-probability deadlocks (when the
 896		 * initial preallocation fails), check if there are any
 897		 * workspaces at all.
 898		 */
 899		if (atomic_read(total_ws) == 0) {
 900			static DEFINE_RATELIMIT_STATE(_rs,
 901					/* once per minute */ 60 * HZ,
 902					/* no burst */ 1);
 903
 904			if (__ratelimit(&_rs))
 905				btrfs_warn(fs_info,
 906				"no compression workspaces, low memory, retrying");
 907		}
 908		goto again;
 909	}
 910	return workspace;
 911}
 912
 913static struct list_head *get_workspace(struct btrfs_fs_info *fs_info, int type, int level)
 914{
 915	switch (type) {
 916	case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(fs_info, type, level);
 917	case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(fs_info, level);
 918	case BTRFS_COMPRESS_LZO:  return btrfs_get_workspace(fs_info, type, level);
 919	case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(fs_info, level);
 920	default:
 921		/*
 922		 * This can't happen, the type is validated several times
 923		 * before we get here.
 924		 */
 925		BUG();
 926	}
 927}
 928
 929/*
 930 * put a workspace struct back on the list or free it if we have enough
 931 * idle ones sitting around
 932 */
 933void btrfs_put_workspace(struct btrfs_fs_info *fs_info, int type, struct list_head *ws)
 934{
 935	struct workspace_manager *gwsm = fs_info->compr_wsm[type];
 936	struct list_head *idle_ws;
 937	spinlock_t *ws_lock;
 938	atomic_t *total_ws;
 939	wait_queue_head_t *ws_wait;
 940	int *free_ws;
 941
 942	ASSERT(gwsm);
 943	idle_ws	 = &gwsm->idle_ws;
 944	ws_lock	 = &gwsm->ws_lock;
 945	total_ws = &gwsm->total_ws;
 946	ws_wait	 = &gwsm->ws_wait;
 947	free_ws	 = &gwsm->free_ws;
 948
 949	spin_lock(ws_lock);
 950	if (*free_ws <= num_online_cpus()) {
 951		list_add(ws, idle_ws);
 952		(*free_ws)++;
 953		spin_unlock(ws_lock);
 954		goto wake;
 955	}
 956	spin_unlock(ws_lock);
 957
 958	free_workspace(type, ws);
 959	atomic_dec(total_ws);
 960wake:
 961	cond_wake_up(ws_wait);
 962}
 963
 964static void put_workspace(struct btrfs_fs_info *fs_info, int type, struct list_head *ws)
 965{
 966	switch (type) {
 967	case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(fs_info, type, ws);
 968	case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(fs_info, type, ws);
 969	case BTRFS_COMPRESS_LZO:  return btrfs_put_workspace(fs_info, type, ws);
 970	case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(fs_info, ws);
 971	default:
 972		/*
 973		 * This can't happen, the type is validated several times
 974		 * before we get here.
 975		 */
 976		BUG();
 977	}
 978}
 979
 980/*
 981 * Adjust @level according to the limits of the compression algorithm or
 982 * fallback to default
 983 */
 984static int btrfs_compress_set_level(unsigned int type, int level)
 985{
 986	const struct btrfs_compress_levels *levels = btrfs_compress_levels[type];
 987
 988	if (level == 0)
 989		level = levels->default_level;
 990	else
 991		level = clamp(level, levels->min_level, levels->max_level);
 992
 993	return level;
 994}
 995
 996/*
 997 * Check whether the @level is within the valid range for the given type.
 998 */
 999bool btrfs_compress_level_valid(unsigned int type, int level)
1000{
1001	const struct btrfs_compress_levels *levels = btrfs_compress_levels[type];
1002
1003	return levels->min_level <= level && level <= levels->max_level;
1004}
1005
1006/* Wrapper around find_get_page(), with extra error message. */
1007int btrfs_compress_filemap_get_folio(struct address_space *mapping, u64 start,
1008				     struct folio **in_folio_ret)
1009{
1010	struct folio *in_folio;
1011
1012	/*
1013	 * The compressed write path should have the folio locked already, thus
1014	 * we only need to grab one reference.
1015	 */
1016	in_folio = filemap_get_folio(mapping, start >> PAGE_SHIFT);
1017	if (IS_ERR(in_folio)) {
1018		struct btrfs_inode *inode = BTRFS_I(mapping->host);
1019
1020		btrfs_crit(inode->root->fs_info,
1021		"failed to get page cache, root %lld ino %llu file offset %llu",
1022			   btrfs_root_id(inode->root), btrfs_ino(inode), start);
1023		return -ENOENT;
1024	}
1025	*in_folio_ret = in_folio;
1026	return 0;
1027}
1028
1029/*
1030 * Given an address space and start and length, compress the bytes into @pages
1031 * that are allocated on demand.
1032 *
1033 * @type_level is encoded algorithm and level, where level 0 means whatever
1034 * default the algorithm chooses and is opaque here;
1035 * - compression algo are 0-3
1036 * - the level are bits 4-7
1037 *
1038 * @out_folios is an in/out parameter, holds maximum number of folios to allocate
1039 * and returns number of actually allocated folios
1040 *
1041 * @total_in is used to return the number of bytes actually read.  It
1042 * may be smaller than the input length if we had to exit early because we
1043 * ran out of room in the folios array or because we cross the
1044 * max_out threshold.
1045 *
1046 * @total_out is an in/out parameter, must be set to the input length and will
1047 * be also used to return the total number of compressed bytes
1048 */
1049int btrfs_compress_folios(unsigned int type, int level, struct btrfs_inode *inode,
1050			 u64 start, struct folio **folios, unsigned long *out_folios,
1051			 unsigned long *total_in, unsigned long *total_out)
1052{
1053	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1054	const unsigned long orig_len = *total_out;
1055	struct list_head *workspace;
1056	int ret;
1057
1058	level = btrfs_compress_set_level(type, level);
1059	workspace = get_workspace(fs_info, type, level);
1060	ret = compression_compress_pages(type, workspace, inode, start, folios,
1061					 out_folios, total_in, total_out);
1062	/* The total read-in bytes should be no larger than the input. */
1063	ASSERT(*total_in <= orig_len);
1064	put_workspace(fs_info, type, workspace);
1065	return ret;
1066}
1067
1068static int btrfs_decompress_bio(struct compressed_bio *cb)
1069{
1070	struct btrfs_fs_info *fs_info = cb_to_fs_info(cb);
1071	struct list_head *workspace;
1072	int ret;
1073	int type = cb->compress_type;
1074
1075	workspace = get_workspace(fs_info, type, 0);
1076	ret = compression_decompress_bio(workspace, cb);
1077	put_workspace(fs_info, type, workspace);
1078
1079	if (!ret)
1080		zero_fill_bio(&cb->orig_bbio->bio);
1081	return ret;
1082}
1083
1084/*
1085 * a less complex decompression routine.  Our compressed data fits in a
1086 * single page, and we want to read a single page out of it.
1087 * dest_pgoff tells us the offset into the destination folio where we write the
1088 * decompressed data.
1089 */
1090int btrfs_decompress(int type, const u8 *data_in, struct folio *dest_folio,
1091		     unsigned long dest_pgoff, size_t srclen, size_t destlen)
1092{
1093	struct btrfs_fs_info *fs_info = folio_to_fs_info(dest_folio);
1094	struct list_head *workspace;
1095	const u32 sectorsize = fs_info->sectorsize;
1096	int ret;
1097
1098	/*
1099	 * The full destination folio range should not exceed the folio size.
1100	 * And the @destlen should not exceed sectorsize, as this is only called for
1101	 * inline file extents, which should not exceed sectorsize.
1102	 */
1103	ASSERT(dest_pgoff + destlen <= folio_size(dest_folio) && destlen <= sectorsize);
1104
1105	workspace = get_workspace(fs_info, type, 0);
1106	ret = compression_decompress(type, workspace, data_in, dest_folio,
1107				     dest_pgoff, srclen, destlen);
1108	put_workspace(fs_info, type, workspace);
1109
1110	return ret;
1111}
1112
1113int btrfs_alloc_compress_wsm(struct btrfs_fs_info *fs_info)
1114{
1115	int ret;
1116
1117	ret = alloc_workspace_manager(fs_info, BTRFS_COMPRESS_NONE);
1118	if (ret < 0)
1119		goto error;
1120	ret = alloc_workspace_manager(fs_info, BTRFS_COMPRESS_ZLIB);
1121	if (ret < 0)
1122		goto error;
1123	ret = alloc_workspace_manager(fs_info, BTRFS_COMPRESS_LZO);
1124	if (ret < 0)
1125		goto error;
1126	ret = zstd_alloc_workspace_manager(fs_info);
1127	if (ret < 0)
1128		goto error;
1129	return 0;
1130error:
1131	btrfs_free_compress_wsm(fs_info);
1132	return ret;
1133}
1134
1135void btrfs_free_compress_wsm(struct btrfs_fs_info *fs_info)
1136{
1137	free_workspace_manager(fs_info, BTRFS_COMPRESS_NONE);
1138	free_workspace_manager(fs_info, BTRFS_COMPRESS_ZLIB);
1139	free_workspace_manager(fs_info, BTRFS_COMPRESS_LZO);
1140	zstd_free_workspace_manager(fs_info);
1141}
1142
1143int __init btrfs_init_compress(void)
1144{
1145	if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1146			offsetof(struct compressed_bio, bbio.bio),
1147			BIOSET_NEED_BVECS))
1148		return -ENOMEM;
1149
1150	compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1151	if (!compr_pool.shrinker)
1152		return -ENOMEM;
1153
1154	spin_lock_init(&compr_pool.lock);
1155	INIT_LIST_HEAD(&compr_pool.list);
1156	compr_pool.count = 0;
1157	/* 128K / 4K = 32, for 8 threads is 256 pages. */
1158	compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1159	compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1160	compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1161	compr_pool.shrinker->batch = 32;
1162	compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1163	shrinker_register(compr_pool.shrinker);
1164
1165	return 0;
1166}
1167
1168void __cold btrfs_exit_compress(void)
1169{
1170	/* For now scan drains all pages and does not touch the parameters. */
1171	btrfs_compr_pool_scan(NULL, NULL);
1172	shrinker_free(compr_pool.shrinker);
1173
1174	bioset_exit(&btrfs_compressed_bioset);
1175}
1176
1177/*
1178 * The bvec is a single page bvec from a bio that contains folios from a filemap.
1179 *
1180 * Since the folio may be a large one, and if the bv_page is not a head page of
1181 * a large folio, then page->index is unreliable.
1182 *
1183 * Thus we need this helper to grab the proper file offset.
1184 */
1185static u64 file_offset_from_bvec(const struct bio_vec *bvec)
1186{
1187	const struct page *page = bvec->bv_page;
1188	const struct folio *folio = page_folio(page);
1189
1190	return (page_pgoff(folio, page) << PAGE_SHIFT) + bvec->bv_offset;
1191}
1192
1193/*
1194 * Copy decompressed data from working buffer to pages.
1195 *
1196 * @buf:		The decompressed data buffer
1197 * @buf_len:		The decompressed data length
1198 * @decompressed:	Number of bytes that are already decompressed inside the
1199 * 			compressed extent
1200 * @cb:			The compressed extent descriptor
1201 * @orig_bio:		The original bio that the caller wants to read for
1202 *
1203 * An easier to understand graph is like below:
1204 *
1205 * 		|<- orig_bio ->|     |<- orig_bio->|
1206 * 	|<-------      full decompressed extent      ----->|
1207 * 	|<-----------    @cb range   ---->|
1208 * 	|			|<-- @buf_len -->|
1209 * 	|<--- @decompressed --->|
1210 *
1211 * Note that, @cb can be a subpage of the full decompressed extent, but
1212 * @cb->start always has the same as the orig_file_offset value of the full
1213 * decompressed extent.
1214 *
1215 * When reading compressed extent, we have to read the full compressed extent,
1216 * while @orig_bio may only want part of the range.
1217 * Thus this function will ensure only data covered by @orig_bio will be copied
1218 * to.
1219 *
1220 * Return 0 if we have copied all needed contents for @orig_bio.
1221 * Return >0 if we need continue decompress.
1222 */
1223int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1224			      struct compressed_bio *cb, u32 decompressed)
1225{
1226	struct bio *orig_bio = &cb->orig_bbio->bio;
1227	/* Offset inside the full decompressed extent */
1228	u32 cur_offset;
1229
1230	cur_offset = decompressed;
1231	/* The main loop to do the copy */
1232	while (cur_offset < decompressed + buf_len) {
1233		struct bio_vec bvec;
1234		size_t copy_len;
1235		u32 copy_start;
1236		/* Offset inside the full decompressed extent */
1237		u32 bvec_offset;
1238		void *kaddr;
1239
1240		bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1241		/*
1242		 * cb->start may underflow, but subtracting that value can still
1243		 * give us correct offset inside the full decompressed extent.
1244		 */
1245		bvec_offset = file_offset_from_bvec(&bvec) - cb->start;
1246
1247		/* Haven't reached the bvec range, exit */
1248		if (decompressed + buf_len <= bvec_offset)
1249			return 1;
1250
1251		copy_start = max(cur_offset, bvec_offset);
1252		copy_len = min(bvec_offset + bvec.bv_len,
1253			       decompressed + buf_len) - copy_start;
1254		ASSERT(copy_len);
1255
1256		/*
1257		 * Extra range check to ensure we didn't go beyond
1258		 * @buf + @buf_len.
1259		 */
1260		ASSERT(copy_start - decompressed < buf_len);
1261
1262		kaddr = bvec_kmap_local(&bvec);
1263		memcpy(kaddr, buf + copy_start - decompressed, copy_len);
1264		kunmap_local(kaddr);
1265
1266		cur_offset += copy_len;
1267		bio_advance(orig_bio, copy_len);
1268		/* Finished the bio */
1269		if (!orig_bio->bi_iter.bi_size)
1270			return 0;
1271	}
1272	return 1;
1273}
1274
1275/*
1276 * Shannon Entropy calculation
1277 *
1278 * Pure byte distribution analysis fails to determine compressibility of data.
1279 * Try calculating entropy to estimate the average minimum number of bits
1280 * needed to encode the sampled data.
1281 *
1282 * For convenience, return the percentage of needed bits, instead of amount of
1283 * bits directly.
1284 *
1285 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1286 *			    and can be compressible with high probability
1287 *
1288 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1289 *
1290 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1291 */
1292#define ENTROPY_LVL_ACEPTABLE		(65)
1293#define ENTROPY_LVL_HIGH		(80)
1294
1295/*
1296 * For increased precision in shannon_entropy calculation,
1297 * let's do pow(n, M) to save more digits after comma:
1298 *
1299 * - maximum int bit length is 64
1300 * - ilog2(MAX_SAMPLE_SIZE)	-> 13
1301 * - 13 * 4 = 52 < 64		-> M = 4
1302 *
1303 * So use pow(n, 4).
1304 */
1305static inline u32 ilog2_w(u64 n)
1306{
1307	return ilog2(n * n * n * n);
1308}
1309
1310static u32 shannon_entropy(struct heuristic_ws *ws)
1311{
1312	const u32 entropy_max = 8 * ilog2_w(2);
1313	u32 entropy_sum = 0;
1314	u32 p, p_base, sz_base;
1315	u32 i;
1316
1317	sz_base = ilog2_w(ws->sample_size);
1318	for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1319		p = ws->bucket[i].count;
1320		p_base = ilog2_w(p);
1321		entropy_sum += p * (sz_base - p_base);
1322	}
1323
1324	entropy_sum /= ws->sample_size;
1325	return entropy_sum * 100 / entropy_max;
1326}
1327
1328#define RADIX_BASE		4U
1329#define COUNTERS_SIZE		(1U << RADIX_BASE)
1330
1331static u8 get4bits(u64 num, int shift) {
1332	u8 low4bits;
1333
1334	num >>= shift;
1335	/* Reverse order */
1336	low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1337	return low4bits;
1338}
1339
1340/*
1341 * Use 4 bits as radix base
1342 * Use 16 u32 counters for calculating new position in buf array
1343 *
1344 * @array     - array that will be sorted
1345 * @array_buf - buffer array to store sorting results
1346 *              must be equal in size to @array
1347 * @num       - array size
1348 */
1349static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1350		       int num)
1351{
1352	u64 max_num;
1353	u64 buf_num;
1354	u32 counters[COUNTERS_SIZE];
1355	u32 new_addr;
1356	u32 addr;
1357	int bitlen;
1358	int shift;
1359	int i;
1360
1361	/*
1362	 * Try avoid useless loop iterations for small numbers stored in big
1363	 * counters.  Example: 48 33 4 ... in 64bit array
1364	 */
1365	max_num = array[0].count;
1366	for (i = 1; i < num; i++) {
1367		buf_num = array[i].count;
1368		if (buf_num > max_num)
1369			max_num = buf_num;
1370	}
1371
1372	buf_num = ilog2(max_num);
1373	bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1374
1375	shift = 0;
1376	while (shift < bitlen) {
1377		memset(counters, 0, sizeof(counters));
1378
1379		for (i = 0; i < num; i++) {
1380			buf_num = array[i].count;
1381			addr = get4bits(buf_num, shift);
1382			counters[addr]++;
1383		}
1384
1385		for (i = 1; i < COUNTERS_SIZE; i++)
1386			counters[i] += counters[i - 1];
1387
1388		for (i = num - 1; i >= 0; i--) {
1389			buf_num = array[i].count;
1390			addr = get4bits(buf_num, shift);
1391			counters[addr]--;
1392			new_addr = counters[addr];
1393			array_buf[new_addr] = array[i];
1394		}
1395
1396		shift += RADIX_BASE;
1397
1398		/*
1399		 * Normal radix expects to move data from a temporary array, to
1400		 * the main one.  But that requires some CPU time. Avoid that
1401		 * by doing another sort iteration to original array instead of
1402		 * memcpy()
1403		 */
1404		memset(counters, 0, sizeof(counters));
1405
1406		for (i = 0; i < num; i ++) {
1407			buf_num = array_buf[i].count;
1408			addr = get4bits(buf_num, shift);
1409			counters[addr]++;
1410		}
1411
1412		for (i = 1; i < COUNTERS_SIZE; i++)
1413			counters[i] += counters[i - 1];
1414
1415		for (i = num - 1; i >= 0; i--) {
1416			buf_num = array_buf[i].count;
1417			addr = get4bits(buf_num, shift);
1418			counters[addr]--;
1419			new_addr = counters[addr];
1420			array[new_addr] = array_buf[i];
1421		}
1422
1423		shift += RADIX_BASE;
1424	}
1425}
1426
1427/*
1428 * Size of the core byte set - how many bytes cover 90% of the sample
1429 *
1430 * There are several types of structured binary data that use nearly all byte
1431 * values. The distribution can be uniform and counts in all buckets will be
1432 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1433 *
1434 * Other possibility is normal (Gaussian) distribution, where the data could
1435 * be potentially compressible, but we have to take a few more steps to decide
1436 * how much.
1437 *
1438 * @BYTE_CORE_SET_LOW  - main part of byte values repeated frequently,
1439 *                       compression algo can easy fix that
1440 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1441 *                       probability is not compressible
1442 */
1443#define BYTE_CORE_SET_LOW		(64)
1444#define BYTE_CORE_SET_HIGH		(200)
1445
1446static int byte_core_set_size(struct heuristic_ws *ws)
1447{
1448	u32 i;
1449	u32 coreset_sum = 0;
1450	const u32 core_set_threshold = ws->sample_size * 90 / 100;
1451	struct bucket_item *bucket = ws->bucket;
1452
1453	/* Sort in reverse order */
1454	radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1455
1456	for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1457		coreset_sum += bucket[i].count;
1458
1459	if (coreset_sum > core_set_threshold)
1460		return i;
1461
1462	for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1463		coreset_sum += bucket[i].count;
1464		if (coreset_sum > core_set_threshold)
1465			break;
1466	}
1467
1468	return i;
1469}
1470
1471/*
1472 * Count byte values in buckets.
1473 * This heuristic can detect textual data (configs, xml, json, html, etc).
1474 * Because in most text-like data byte set is restricted to limited number of
1475 * possible characters, and that restriction in most cases makes data easy to
1476 * compress.
1477 *
1478 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1479 *	less - compressible
1480 *	more - need additional analysis
1481 */
1482#define BYTE_SET_THRESHOLD		(64)
1483
1484static u32 byte_set_size(const struct heuristic_ws *ws)
1485{
1486	u32 i;
1487	u32 byte_set_size = 0;
1488
1489	for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1490		if (ws->bucket[i].count > 0)
1491			byte_set_size++;
1492	}
1493
1494	/*
1495	 * Continue collecting count of byte values in buckets.  If the byte
1496	 * set size is bigger then the threshold, it's pointless to continue,
1497	 * the detection technique would fail for this type of data.
1498	 */
1499	for (; i < BUCKET_SIZE; i++) {
1500		if (ws->bucket[i].count > 0) {
1501			byte_set_size++;
1502			if (byte_set_size > BYTE_SET_THRESHOLD)
1503				return byte_set_size;
1504		}
1505	}
1506
1507	return byte_set_size;
1508}
1509
1510static bool sample_repeated_patterns(struct heuristic_ws *ws)
1511{
1512	const u32 half_of_sample = ws->sample_size / 2;
1513	const u8 *data = ws->sample;
1514
1515	return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1516}
1517
1518static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1519				     struct heuristic_ws *ws)
1520{
1521	struct page *page;
1522	pgoff_t index, index_end;
1523	u32 i, curr_sample_pos;
1524	u8 *in_data;
1525
1526	/*
1527	 * Compression handles the input data by chunks of 128KiB
1528	 * (defined by BTRFS_MAX_UNCOMPRESSED)
1529	 *
1530	 * We do the same for the heuristic and loop over the whole range.
1531	 *
1532	 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1533	 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1534	 */
1535	if (end - start > BTRFS_MAX_UNCOMPRESSED)
1536		end = start + BTRFS_MAX_UNCOMPRESSED;
1537
1538	index = start >> PAGE_SHIFT;
1539	index_end = end >> PAGE_SHIFT;
1540
1541	/* Don't miss unaligned end */
1542	if (!PAGE_ALIGNED(end))
1543		index_end++;
1544
1545	curr_sample_pos = 0;
1546	while (index < index_end) {
1547		page = find_get_page(inode->i_mapping, index);
1548		in_data = kmap_local_page(page);
1549		/* Handle case where the start is not aligned to PAGE_SIZE */
1550		i = start % PAGE_SIZE;
1551		while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1552			/* Don't sample any garbage from the last page */
1553			if (start > end - SAMPLING_READ_SIZE)
1554				break;
1555			memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1556					SAMPLING_READ_SIZE);
1557			i += SAMPLING_INTERVAL;
1558			start += SAMPLING_INTERVAL;
1559			curr_sample_pos += SAMPLING_READ_SIZE;
1560		}
1561		kunmap_local(in_data);
1562		put_page(page);
1563
1564		index++;
1565	}
1566
1567	ws->sample_size = curr_sample_pos;
1568}
1569
1570/*
1571 * Compression heuristic.
1572 *
1573 * The following types of analysis can be performed:
1574 * - detect mostly zero data
1575 * - detect data with low "byte set" size (text, etc)
1576 * - detect data with low/high "core byte" set
1577 *
1578 * Return non-zero if the compression should be done, 0 otherwise.
1579 */
1580int btrfs_compress_heuristic(struct btrfs_inode *inode, u64 start, u64 end)
1581{
1582	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1583	struct list_head *ws_list = get_workspace(fs_info, 0, 0);
1584	struct heuristic_ws *ws;
1585	u32 i;
1586	u8 byte;
1587	int ret = 0;
1588
1589	ws = list_entry(ws_list, struct heuristic_ws, list);
1590
1591	heuristic_collect_sample(&inode->vfs_inode, start, end, ws);
1592
1593	if (sample_repeated_patterns(ws)) {
1594		ret = 1;
1595		goto out;
1596	}
1597
1598	memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1599
1600	for (i = 0; i < ws->sample_size; i++) {
1601		byte = ws->sample[i];
1602		ws->bucket[byte].count++;
1603	}
1604
1605	i = byte_set_size(ws);
1606	if (i < BYTE_SET_THRESHOLD) {
1607		ret = 2;
1608		goto out;
1609	}
1610
1611	i = byte_core_set_size(ws);
1612	if (i <= BYTE_CORE_SET_LOW) {
1613		ret = 3;
1614		goto out;
1615	}
1616
1617	if (i >= BYTE_CORE_SET_HIGH) {
1618		ret = 0;
1619		goto out;
1620	}
1621
1622	i = shannon_entropy(ws);
1623	if (i <= ENTROPY_LVL_ACEPTABLE) {
1624		ret = 4;
1625		goto out;
1626	}
1627
1628	/*
1629	 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1630	 * needed to give green light to compression.
1631	 *
1632	 * For now just assume that compression at that level is not worth the
1633	 * resources because:
1634	 *
1635	 * 1. it is possible to defrag the data later
1636	 *
1637	 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1638	 * values, every bucket has counter at level ~54. The heuristic would
1639	 * be confused. This can happen when data have some internal repeated
1640	 * patterns like "abbacbbc...". This can be detected by analyzing
1641	 * pairs of bytes, which is too costly.
1642	 */
1643	if (i < ENTROPY_LVL_HIGH) {
1644		ret = 5;
1645		goto out;
1646	} else {
1647		ret = 0;
1648		goto out;
1649	}
1650
1651out:
1652	put_workspace(fs_info, 0, ws_list);
1653	return ret;
1654}
1655
1656/*
1657 * Convert the compression suffix (eg. after "zlib" starting with ":") to level.
1658 *
1659 * If the resulting level exceeds the algo's supported levels, it will be clamped.
1660 *
1661 * Return <0 if no valid string can be found.
1662 * Return 0 if everything is fine.
1663 */
1664int btrfs_compress_str2level(unsigned int type, const char *str, int *level_ret)
1665{
1666	int level = 0;
1667	int ret;
1668
1669	if (!type) {
1670		*level_ret = btrfs_compress_set_level(type, level);
1671		return 0;
1672	}
1673
1674	if (str[0] == ':') {
1675		ret = kstrtoint(str + 1, 10, &level);
1676		if (ret)
1677			return ret;
1678	}
1679
1680	*level_ret = btrfs_compress_set_level(type, level);
1681	return 0;
1682}