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 <crypto/hash.h>
  24#include "misc.h"
  25#include "ctree.h"
  26#include "fs.h"
  27#include "disk-io.h"
  28#include "transaction.h"
  29#include "btrfs_inode.h"
  30#include "bio.h"
  31#include "ordered-data.h"
  32#include "compression.h"
  33#include "extent_io.h"
  34#include "extent_map.h"
  35#include "subpage.h"
  36#include "zoned.h"
  37#include "file-item.h"
  38#include "super.h"
  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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  57bool btrfs_compress_is_valid_type(const char *str, size_t len)
  58{
  59	int i;
  60
  61	for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
  62		size_t comp_len = strlen(btrfs_compress_types[i]);
  63
  64		if (len < comp_len)
  65			continue;
  66
  67		if (!strncmp(btrfs_compress_types[i], str, comp_len))
  68			return true;
  69	}
  70	return false;
  71}
  72
  73static int compression_compress_pages(int type, struct list_head *ws,
  74               struct address_space *mapping, u64 start, struct page **pages,
  75               unsigned long *out_pages, unsigned long *total_in,
  76               unsigned long *total_out)
  77{
  78	switch (type) {
  79	case BTRFS_COMPRESS_ZLIB:
  80		return zlib_compress_pages(ws, mapping, start, pages,
  81				out_pages, total_in, total_out);
  82	case BTRFS_COMPRESS_LZO:
  83		return lzo_compress_pages(ws, mapping, start, pages,
  84				out_pages, total_in, total_out);
  85	case BTRFS_COMPRESS_ZSTD:
  86		return zstd_compress_pages(ws, mapping, start, pages,
  87				out_pages, total_in, total_out);
  88	case BTRFS_COMPRESS_NONE:
  89	default:
  90		/*
  91		 * This can happen when compression races with remount setting
  92		 * it to 'no compress', while caller doesn't call
  93		 * inode_need_compress() to check if we really need to
  94		 * compress.
  95		 *
  96		 * Not a big deal, just need to inform caller that we
  97		 * haven't allocated any pages yet.
  98		 */
  99		*out_pages = 0;
 100		return -E2BIG;
 101	}
 102}
 103
 104static int compression_decompress_bio(struct list_head *ws,
 105				      struct compressed_bio *cb)
 106{
 107	switch (cb->compress_type) {
 108	case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
 109	case BTRFS_COMPRESS_LZO:  return lzo_decompress_bio(ws, cb);
 110	case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
 111	case BTRFS_COMPRESS_NONE:
 112	default:
 113		/*
 114		 * This can't happen, the type is validated several times
 115		 * before we get here.
 116		 */
 117		BUG();
 118	}
 119}
 120
 121static int compression_decompress(int type, struct list_head *ws,
 122               const u8 *data_in, struct page *dest_page,
 123               unsigned long start_byte, size_t srclen, size_t destlen)
 124{
 125	switch (type) {
 126	case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
 127						start_byte, srclen, destlen);
 128	case BTRFS_COMPRESS_LZO:  return lzo_decompress(ws, data_in, dest_page,
 129						start_byte, srclen, destlen);
 130	case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
 131						start_byte, srclen, destlen);
 132	case BTRFS_COMPRESS_NONE:
 133	default:
 134		/*
 135		 * This can't happen, the type is validated several times
 136		 * before we get here.
 137		 */
 138		BUG();
 139	}
 140}
 141
 
 
 
 
 
 
 
 142static int btrfs_decompress_bio(struct compressed_bio *cb);
 143
 144static void finish_compressed_bio_read(struct compressed_bio *cb)
 
 
 
 
 
 
 
 
 
 
 
 145{
 146	unsigned int index;
 147	struct page *page;
 148
 149	if (cb->status == BLK_STS_OK)
 150		cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
 
 
 
 
 
 
 
 151
 152	/* Release the compressed pages */
 153	for (index = 0; index < cb->nr_pages; index++) {
 154		page = cb->compressed_pages[index];
 155		page->mapping = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 156		put_page(page);
 157	}
 158
 159	/* Do io completion on the original bio */
 160	btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status);
 161
 162	/* Finally free the cb struct */
 163	kfree(cb->compressed_pages);
 164	kfree(cb);
 165}
 166
 167/*
 168 * Verify the checksums and kick off repair if needed on the uncompressed data
 169 * before decompressing it into the original bio and freeing the uncompressed
 170 * pages.
 171 */
 172static void end_compressed_bio_read(struct btrfs_bio *bbio)
 173{
 174	struct compressed_bio *cb = bbio->private;
 175	struct inode *inode = cb->inode;
 176	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 177	struct btrfs_inode *bi = BTRFS_I(inode);
 178	bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) &&
 179		    !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state);
 180	blk_status_t status = bbio->bio.bi_status;
 181	struct bvec_iter iter;
 182	struct bio_vec bv;
 183	u32 offset;
 184
 185	btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) {
 186		u64 start = bbio->file_offset + offset;
 187
 188		if (!status &&
 189		    (!csum || !btrfs_check_data_csum(bi, bbio, offset,
 190						     bv.bv_page, bv.bv_offset))) {
 191			btrfs_clean_io_failure(bi, start, bv.bv_page,
 192					       bv.bv_offset);
 193		} else {
 194			int ret;
 195
 196			refcount_inc(&cb->pending_ios);
 197			ret = btrfs_repair_one_sector(BTRFS_I(inode), bbio, offset,
 198						      bv.bv_page, bv.bv_offset,
 199						      true);
 200			if (ret) {
 201				refcount_dec(&cb->pending_ios);
 202				status = errno_to_blk_status(ret);
 203			}
 204		}
 205	}
 
 206
 207	if (status)
 208		cb->status = status;
 209
 210	if (refcount_dec_and_test(&cb->pending_ios))
 211		finish_compressed_bio_read(cb);
 212	btrfs_bio_free_csum(bbio);
 
 
 
 
 
 
 
 
 
 
 
 213	bio_put(&bbio->bio);
 214}
 215
 216/*
 217 * Clear the writeback bits on all of the file
 218 * pages for a compressed write
 219 */
 220static noinline void end_compressed_writeback(struct inode *inode,
 221					      const struct compressed_bio *cb)
 222{
 223	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 
 224	unsigned long index = cb->start >> PAGE_SHIFT;
 225	unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
 226	struct folio_batch fbatch;
 227	const int errno = blk_status_to_errno(cb->status);
 228	int i;
 229	int ret;
 230
 231	if (errno)
 232		mapping_set_error(inode->i_mapping, errno);
 233
 234	folio_batch_init(&fbatch);
 235	while (index <= end_index) {
 236		ret = filemap_get_folios(inode->i_mapping, &index, end_index,
 237				&fbatch);
 238
 239		if (ret == 0)
 240			return;
 241
 242		for (i = 0; i < ret; i++) {
 243			struct folio *folio = fbatch.folios[i];
 244
 245			if (errno)
 246				folio_set_error(folio);
 247			btrfs_page_clamp_clear_writeback(fs_info, &folio->page,
 248							 cb->start, cb->len);
 249		}
 250		folio_batch_release(&fbatch);
 251	}
 252	/* the inode may be gone now */
 253}
 254
 255static void finish_compressed_bio_write(struct compressed_bio *cb)
 256{
 257	struct inode *inode = cb->inode;
 258	unsigned int index;
 259
 260	/*
 261	 * Ok, we're the last bio for this extent, step one is to call back
 262	 * into the FS and do all the end_io operations.
 263	 */
 264	btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
 265			cb->start, cb->start + cb->len - 1,
 266			cb->status == BLK_STS_OK);
 267
 268	if (cb->writeback)
 269		end_compressed_writeback(inode, cb);
 270	/* Note, our inode could be gone now */
 271
 272	/*
 273	 * Release the compressed pages, these came from alloc_page and
 274	 * are not attached to the inode at all
 275	 */
 276	for (index = 0; index < cb->nr_pages; index++) {
 277		struct page *page = cb->compressed_pages[index];
 278
 279		page->mapping = NULL;
 280		put_page(page);
 281	}
 282
 283	/* Finally free the cb struct */
 284	kfree(cb->compressed_pages);
 285	kfree(cb);
 286}
 287
 288static void btrfs_finish_compressed_write_work(struct work_struct *work)
 289{
 290	struct compressed_bio *cb =
 291		container_of(work, struct compressed_bio, write_end_work);
 292
 293	finish_compressed_bio_write(cb);
 294}
 295
 296/*
 297 * Do the cleanup once all the compressed pages hit the disk.  This will clear
 298 * writeback on the file pages and free the compressed pages.
 299 *
 300 * This also calls the writeback end hooks for the file pages so that metadata
 301 * and checksums can be updated in the file.
 302 */
 303static void end_compressed_bio_write(struct btrfs_bio *bbio)
 304{
 305	struct compressed_bio *cb = bbio->private;
 306
 307	if (bbio->bio.bi_status)
 308		cb->status = bbio->bio.bi_status;
 309
 310	if (refcount_dec_and_test(&cb->pending_ios)) {
 311		struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
 312
 313		btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio);
 314		queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
 315	}
 316	bio_put(&bbio->bio);
 317}
 318
 319/*
 320 * Allocate a compressed_bio, which will be used to read/write on-disk
 321 * (aka, compressed) * data.
 322 *
 323 * @cb:                 The compressed_bio structure, which records all the needed
 324 *                      information to bind the compressed data to the uncompressed
 325 *                      page cache.
 326 * @disk_byten:         The logical bytenr where the compressed data will be read
 327 *                      from or written to.
 328 * @endio_func:         The endio function to call after the IO for compressed data
 329 *                      is finished.
 330 * @next_stripe_start:  Return value of logical bytenr of where next stripe starts.
 331 *                      Let the caller know to only fill the bio up to the stripe
 332 *                      boundary.
 333 */
 334
 335
 336static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr,
 337					blk_opf_t opf,
 338					btrfs_bio_end_io_t endio_func,
 339					u64 *next_stripe_start)
 340{
 341	struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
 342	struct btrfs_io_geometry geom;
 343	struct extent_map *em;
 344	struct bio *bio;
 345	int ret;
 346
 347	bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, endio_func, cb);
 348	bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
 
 349
 350	em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize);
 351	if (IS_ERR(em)) {
 352		bio_put(bio);
 353		return ERR_CAST(em);
 
 354	}
 355
 356	if (bio_op(bio) == REQ_OP_ZONE_APPEND)
 357		bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev);
 358
 359	ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom);
 360	free_extent_map(em);
 361	if (ret < 0) {
 362		bio_put(bio);
 363		return ERR_PTR(ret);
 364	}
 365	*next_stripe_start = disk_bytenr + geom.len;
 366	refcount_inc(&cb->pending_ios);
 367	return bio;
 368}
 369
 370/*
 371 * worker function to build and submit bios for previously compressed pages.
 372 * The corresponding pages in the inode should be marked for writeback
 373 * and the compressed pages should have a reference on them for dropping
 374 * when the IO is complete.
 375 *
 376 * This also checksums the file bytes and gets things ready for
 377 * the end io hooks.
 378 */
 379blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
 380				 unsigned int len, u64 disk_start,
 381				 unsigned int compressed_len,
 382				 struct page **compressed_pages,
 383				 unsigned int nr_pages,
 384				 blk_opf_t write_flags,
 385				 struct cgroup_subsys_state *blkcg_css,
 386				 bool writeback)
 387{
 
 388	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 389	struct bio *bio = NULL;
 390	struct compressed_bio *cb;
 391	u64 cur_disk_bytenr = disk_start;
 392	u64 next_stripe_start;
 393	blk_status_t ret = BLK_STS_OK;
 394	int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
 395	const bool use_append = btrfs_use_zone_append(inode, disk_start);
 396	const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
 397
 398	ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
 399	       IS_ALIGNED(len, fs_info->sectorsize));
 400	cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
 401	if (!cb)
 402		return BLK_STS_RESOURCE;
 403	refcount_set(&cb->pending_ios, 1);
 404	cb->status = BLK_STS_OK;
 405	cb->inode = &inode->vfs_inode;
 406	cb->start = start;
 407	cb->len = len;
 408	cb->compressed_pages = compressed_pages;
 409	cb->compressed_len = compressed_len;
 410	cb->writeback = writeback;
 411	INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
 412	cb->nr_pages = nr_pages;
 413
 414	if (blkcg_css)
 415		kthread_associate_blkcg(blkcg_css);
 416
 417	while (cur_disk_bytenr < disk_start + compressed_len) {
 418		u64 offset = cur_disk_bytenr - disk_start;
 419		unsigned int index = offset >> PAGE_SHIFT;
 420		unsigned int real_size;
 421		unsigned int added;
 422		struct page *page = compressed_pages[index];
 423		bool submit = false;
 424
 425		/* Allocate new bio if submitted or not yet allocated */
 426		if (!bio) {
 427			bio = alloc_compressed_bio(cb, cur_disk_bytenr,
 428				bio_op | write_flags, end_compressed_bio_write,
 429				&next_stripe_start);
 430			if (IS_ERR(bio)) {
 431				ret = errno_to_blk_status(PTR_ERR(bio));
 432				break;
 433			}
 434			if (blkcg_css)
 435				bio->bi_opf |= REQ_CGROUP_PUNT;
 436		}
 437		/*
 438		 * We should never reach next_stripe_start start as we will
 439		 * submit comp_bio when reach the boundary immediately.
 440		 */
 441		ASSERT(cur_disk_bytenr != next_stripe_start);
 442
 443		/*
 444		 * We have various limits on the real read size:
 445		 * - stripe boundary
 446		 * - page boundary
 447		 * - compressed length boundary
 448		 */
 449		real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr);
 450		real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
 451		real_size = min_t(u64, real_size, compressed_len - offset);
 452		ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
 453
 454		if (use_append)
 455			added = bio_add_zone_append_page(bio, page, real_size,
 456					offset_in_page(offset));
 457		else
 458			added = bio_add_page(bio, page, real_size,
 459					offset_in_page(offset));
 460		/* Reached zoned boundary */
 461		if (added == 0)
 462			submit = true;
 463
 464		cur_disk_bytenr += added;
 465		/* Reached stripe boundary */
 466		if (cur_disk_bytenr == next_stripe_start)
 467			submit = true;
 468
 469		/* Finished the range */
 470		if (cur_disk_bytenr == disk_start + compressed_len)
 471			submit = true;
 472
 473		if (submit) {
 474			if (!skip_sum) {
 475				ret = btrfs_csum_one_bio(inode, bio, start, true);
 476				if (ret) {
 477					btrfs_bio_end_io(btrfs_bio(bio), ret);
 478					break;
 479				}
 480			}
 481
 482			ASSERT(bio->bi_iter.bi_size);
 483			btrfs_submit_bio(fs_info, bio, 0);
 484			bio = NULL;
 485		}
 486		cond_resched();
 487	}
 488
 489	if (blkcg_css)
 490		kthread_associate_blkcg(NULL);
 491
 492	if (refcount_dec_and_test(&cb->pending_ios))
 493		finish_compressed_bio_write(cb);
 494	return ret;
 495}
 496
 497static u64 bio_end_offset(struct bio *bio)
 498{
 499	struct bio_vec *last = bio_last_bvec_all(bio);
 500
 501	return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
 502}
 503
 504/*
 505 * Add extra pages in the same compressed file extent so that we don't need to
 506 * re-read the same extent again and again.
 507 *
 508 * NOTE: this won't work well for subpage, as for subpage read, we lock the
 509 * full page then submit bio for each compressed/regular extents.
 510 *
 511 * This means, if we have several sectors in the same page points to the same
 512 * on-disk compressed data, we will re-read the same extent many times and
 513 * this function can only help for the next page.
 514 */
 515static noinline int add_ra_bio_pages(struct inode *inode,
 516				     u64 compressed_end,
 517				     struct compressed_bio *cb,
 518				     int *memstall, unsigned long *pflags)
 519{
 520	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 521	unsigned long end_index;
 522	u64 cur = bio_end_offset(cb->orig_bio);
 
 523	u64 isize = i_size_read(inode);
 524	int ret;
 525	struct page *page;
 526	struct extent_map *em;
 527	struct address_space *mapping = inode->i_mapping;
 528	struct extent_map_tree *em_tree;
 529	struct extent_io_tree *tree;
 530	int sectors_missed = 0;
 531
 532	em_tree = &BTRFS_I(inode)->extent_tree;
 533	tree = &BTRFS_I(inode)->io_tree;
 534
 535	if (isize == 0)
 536		return 0;
 537
 538	/*
 539	 * For current subpage support, we only support 64K page size,
 540	 * which means maximum compressed extent size (128K) is just 2x page
 541	 * size.
 542	 * This makes readahead less effective, so here disable readahead for
 543	 * subpage for now, until full compressed write is supported.
 544	 */
 545	if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
 546		return 0;
 547
 548	end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 549
 550	while (cur < compressed_end) {
 551		u64 page_end;
 552		u64 pg_index = cur >> PAGE_SHIFT;
 553		u32 add_size;
 554
 555		if (pg_index > end_index)
 556			break;
 557
 558		page = xa_load(&mapping->i_pages, pg_index);
 559		if (page && !xa_is_value(page)) {
 560			sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
 
 
 
 
 561					  fs_info->sectorsize_bits;
 562
 563			/* Beyond threshold, no need to continue */
 564			if (sectors_missed > 4)
 565				break;
 566
 567			/*
 568			 * Jump to next page start as we already have page for
 569			 * current offset.
 570			 */
 571			cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
 572			continue;
 573		}
 574
 575		page = __page_cache_alloc(mapping_gfp_constraint(mapping,
 576								 ~__GFP_FS));
 577		if (!page)
 578			break;
 579
 580		if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
 581			put_page(page);
 582			/* There is already a page, skip to page end */
 583			cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
 
 584			continue;
 585		}
 586
 587		if (!*memstall && PageWorkingset(page)) {
 588			psi_memstall_enter(pflags);
 589			*memstall = 1;
 590		}
 591
 592		ret = set_page_extent_mapped(page);
 593		if (ret < 0) {
 594			unlock_page(page);
 595			put_page(page);
 596			break;
 597		}
 598
 599		page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
 600		lock_extent(tree, cur, page_end, NULL);
 601		read_lock(&em_tree->lock);
 602		em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
 603		read_unlock(&em_tree->lock);
 604
 605		/*
 606		 * At this point, we have a locked page in the page cache for
 607		 * these bytes in the file.  But, we have to make sure they map
 608		 * to this compressed extent on disk.
 609		 */
 610		if (!em || cur < em->start ||
 611		    (cur + fs_info->sectorsize > extent_map_end(em)) ||
 612		    (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
 
 613			free_extent_map(em);
 614			unlock_extent(tree, cur, page_end, NULL);
 615			unlock_page(page);
 616			put_page(page);
 617			break;
 618		}
 
 619		free_extent_map(em);
 
 620
 621		if (page->index == end_index) {
 622			size_t zero_offset = offset_in_page(isize);
 623
 624			if (zero_offset) {
 625				int zeros;
 626				zeros = PAGE_SIZE - zero_offset;
 627				memzero_page(page, zero_offset, zeros);
 628			}
 629		}
 630
 631		add_size = min(em->start + em->len, page_end + 1) - cur;
 632		ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
 633		if (ret != add_size) {
 634			unlock_extent(tree, cur, page_end, NULL);
 635			unlock_page(page);
 636			put_page(page);
 637			break;
 638		}
 639		/*
 640		 * If it's subpage, we also need to increase its
 641		 * subpage::readers number, as at endio we will decrease
 642		 * subpage::readers and to unlock the page.
 643		 */
 644		if (fs_info->sectorsize < PAGE_SIZE)
 645			btrfs_subpage_start_reader(fs_info, page, cur, add_size);
 646		put_page(page);
 647		cur += add_size;
 648	}
 649	return 0;
 650}
 651
 652/*
 653 * for a compressed read, the bio we get passed has all the inode pages
 654 * in it.  We don't actually do IO on those pages but allocate new ones
 655 * to hold the compressed pages on disk.
 656 *
 657 * bio->bi_iter.bi_sector points to the compressed extent on disk
 658 * bio->bi_io_vec points to all of the inode pages
 659 *
 660 * After the compressed pages are read, we copy the bytes into the
 661 * bio we were passed and then call the bio end_io calls
 662 */
 663void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 664				  int mirror_num)
 665{
 666	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 667	struct extent_map_tree *em_tree;
 
 668	struct compressed_bio *cb;
 669	unsigned int compressed_len;
 670	struct bio *comp_bio = NULL;
 671	const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
 672	u64 cur_disk_byte = disk_bytenr;
 673	u64 next_stripe_start;
 674	u64 file_offset;
 675	u64 em_len;
 676	u64 em_start;
 677	struct extent_map *em;
 678	unsigned long pflags;
 679	int memstall = 0;
 680	blk_status_t ret;
 681	int ret2;
 682	int i;
 683
 684	em_tree = &BTRFS_I(inode)->extent_tree;
 685
 686	file_offset = bio_first_bvec_all(bio)->bv_offset +
 687		      page_offset(bio_first_page_all(bio));
 688
 689	/* we need the actual starting offset of this extent in the file */
 690	read_lock(&em_tree->lock);
 691	em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
 692	read_unlock(&em_tree->lock);
 693	if (!em) {
 694		ret = BLK_STS_IOERR;
 695		goto out;
 696	}
 697
 698	ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
 699	compressed_len = em->block_len;
 700	cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
 701	if (!cb) {
 702		ret = BLK_STS_RESOURCE;
 703		goto out;
 704	}
 705
 706	refcount_set(&cb->pending_ios, 1);
 707	cb->status = BLK_STS_OK;
 708	cb->inode = inode;
 709
 710	cb->start = em->orig_start;
 711	em_len = em->len;
 712	em_start = em->start;
 713
 714	cb->len = bio->bi_iter.bi_size;
 715	cb->compressed_len = compressed_len;
 716	cb->compress_type = em->compress_type;
 717	cb->orig_bio = bio;
 718
 719	free_extent_map(em);
 720	em = NULL;
 721
 722	cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
 723	cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
 724	if (!cb->compressed_pages) {
 725		ret = BLK_STS_RESOURCE;
 726		goto fail;
 727	}
 728
 729	ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
 730	if (ret2) {
 731		ret = BLK_STS_RESOURCE;
 732		goto fail;
 733	}
 734
 735	add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags);
 
 736
 737	/* include any pages we added in add_ra-bio_pages */
 738	cb->len = bio->bi_iter.bi_size;
 739
 740	while (cur_disk_byte < disk_bytenr + compressed_len) {
 741		u64 offset = cur_disk_byte - disk_bytenr;
 742		unsigned int index = offset >> PAGE_SHIFT;
 743		unsigned int real_size;
 744		unsigned int added;
 745		struct page *page = cb->compressed_pages[index];
 746		bool submit = false;
 747
 748		/* Allocate new bio if submitted or not yet allocated */
 749		if (!comp_bio) {
 750			comp_bio = alloc_compressed_bio(cb, cur_disk_byte,
 751					REQ_OP_READ, end_compressed_bio_read,
 752					&next_stripe_start);
 753			if (IS_ERR(comp_bio)) {
 754				cb->status = errno_to_blk_status(PTR_ERR(comp_bio));
 755				break;
 756			}
 757		}
 758		/*
 759		 * We should never reach next_stripe_start start as we will
 760		 * submit comp_bio when reach the boundary immediately.
 761		 */
 762		ASSERT(cur_disk_byte != next_stripe_start);
 763		/*
 764		 * We have various limit on the real read size:
 765		 * - stripe boundary
 766		 * - page boundary
 767		 * - compressed length boundary
 768		 */
 769		real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte);
 770		real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
 771		real_size = min_t(u64, real_size, compressed_len - offset);
 772		ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
 773
 774		added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
 775		/*
 776		 * Maximum compressed extent is smaller than bio size limit,
 777		 * thus bio_add_page() should always success.
 778		 */
 779		ASSERT(added == real_size);
 780		cur_disk_byte += added;
 781
 782		/* Reached stripe boundary, need to submit */
 783		if (cur_disk_byte == next_stripe_start)
 784			submit = true;
 785
 786		/* Has finished the range, need to submit */
 787		if (cur_disk_byte == disk_bytenr + compressed_len)
 788			submit = true;
 789
 790		if (submit) {
 791			/* Save the original iter for read repair */
 792			if (bio_op(comp_bio) == REQ_OP_READ)
 793				btrfs_bio(comp_bio)->iter = comp_bio->bi_iter;
 794
 795			/*
 796			 * Save the initial offset of this chunk, as there
 797			 * is no direct correlation between compressed pages and
 798			 * the original file offset.  The field is only used for
 799			 * priting error messages.
 800			 */
 801			btrfs_bio(comp_bio)->file_offset = file_offset;
 802
 803			ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL);
 804			if (ret) {
 805				btrfs_bio_end_io(btrfs_bio(comp_bio), ret);
 806				break;
 807			}
 808
 809			ASSERT(comp_bio->bi_iter.bi_size);
 810			btrfs_submit_bio(fs_info, comp_bio, mirror_num);
 811			comp_bio = NULL;
 812		}
 813	}
 814
 815	if (memstall)
 816		psi_memstall_leave(&pflags);
 817
 818	if (refcount_dec_and_test(&cb->pending_ios))
 819		finish_compressed_bio_read(cb);
 820	return;
 821
 822fail:
 823	if (cb->compressed_pages) {
 824		for (i = 0; i < cb->nr_pages; i++) {
 825			if (cb->compressed_pages[i])
 826				__free_page(cb->compressed_pages[i]);
 827		}
 828	}
 829
 830	kfree(cb->compressed_pages);
 831	kfree(cb);
 832out:
 833	free_extent_map(em);
 834	btrfs_bio_end_io(btrfs_bio(bio), ret);
 835	return;
 836}
 837
 838/*
 839 * Heuristic uses systematic sampling to collect data from the input data
 840 * range, the logic can be tuned by the following constants:
 841 *
 842 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
 843 * @SAMPLING_INTERVAL  - range from which the sampled data can be collected
 844 */
 845#define SAMPLING_READ_SIZE	(16)
 846#define SAMPLING_INTERVAL	(256)
 847
 848/*
 849 * For statistical analysis of the input data we consider bytes that form a
 850 * Galois Field of 256 objects. Each object has an attribute count, ie. how
 851 * many times the object appeared in the sample.
 852 */
 853#define BUCKET_SIZE		(256)
 854
 855/*
 856 * The size of the sample is based on a statistical sampling rule of thumb.
 857 * The common way is to perform sampling tests as long as the number of
 858 * elements in each cell is at least 5.
 859 *
 860 * Instead of 5, we choose 32 to obtain more accurate results.
 861 * If the data contain the maximum number of symbols, which is 256, we obtain a
 862 * sample size bound by 8192.
 863 *
 864 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
 865 * from up to 512 locations.
 866 */
 867#define MAX_SAMPLE_SIZE		(BTRFS_MAX_UNCOMPRESSED *		\
 868				 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
 869
 870struct bucket_item {
 871	u32 count;
 872};
 873
 874struct heuristic_ws {
 875	/* Partial copy of input data */
 876	u8 *sample;
 877	u32 sample_size;
 878	/* Buckets store counters for each byte value */
 879	struct bucket_item *bucket;
 880	/* Sorting buffer */
 881	struct bucket_item *bucket_b;
 882	struct list_head list;
 883};
 884
 885static struct workspace_manager heuristic_wsm;
 886
 887static void free_heuristic_ws(struct list_head *ws)
 888{
 889	struct heuristic_ws *workspace;
 890
 891	workspace = list_entry(ws, struct heuristic_ws, list);
 892
 893	kvfree(workspace->sample);
 894	kfree(workspace->bucket);
 895	kfree(workspace->bucket_b);
 896	kfree(workspace);
 897}
 898
 899static struct list_head *alloc_heuristic_ws(unsigned int level)
 900{
 901	struct heuristic_ws *ws;
 902
 903	ws = kzalloc(sizeof(*ws), GFP_KERNEL);
 904	if (!ws)
 905		return ERR_PTR(-ENOMEM);
 906
 907	ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
 908	if (!ws->sample)
 909		goto fail;
 910
 911	ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
 912	if (!ws->bucket)
 913		goto fail;
 914
 915	ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
 916	if (!ws->bucket_b)
 917		goto fail;
 918
 919	INIT_LIST_HEAD(&ws->list);
 920	return &ws->list;
 921fail:
 922	free_heuristic_ws(&ws->list);
 923	return ERR_PTR(-ENOMEM);
 924}
 925
 926const struct btrfs_compress_op btrfs_heuristic_compress = {
 927	.workspace_manager = &heuristic_wsm,
 928};
 929
 930static const struct btrfs_compress_op * const btrfs_compress_op[] = {
 931	/* The heuristic is represented as compression type 0 */
 932	&btrfs_heuristic_compress,
 933	&btrfs_zlib_compress,
 934	&btrfs_lzo_compress,
 935	&btrfs_zstd_compress,
 936};
 937
 938static struct list_head *alloc_workspace(int type, unsigned int level)
 939{
 940	switch (type) {
 941	case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
 942	case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
 943	case BTRFS_COMPRESS_LZO:  return lzo_alloc_workspace(level);
 944	case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
 945	default:
 946		/*
 947		 * This can't happen, the type is validated several times
 948		 * before we get here.
 949		 */
 950		BUG();
 951	}
 952}
 953
 954static void free_workspace(int type, struct list_head *ws)
 955{
 956	switch (type) {
 957	case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
 958	case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
 959	case BTRFS_COMPRESS_LZO:  return lzo_free_workspace(ws);
 960	case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
 961	default:
 962		/*
 963		 * This can't happen, the type is validated several times
 964		 * before we get here.
 965		 */
 966		BUG();
 967	}
 968}
 969
 970static void btrfs_init_workspace_manager(int type)
 971{
 972	struct workspace_manager *wsm;
 973	struct list_head *workspace;
 974
 975	wsm = btrfs_compress_op[type]->workspace_manager;
 976	INIT_LIST_HEAD(&wsm->idle_ws);
 977	spin_lock_init(&wsm->ws_lock);
 978	atomic_set(&wsm->total_ws, 0);
 979	init_waitqueue_head(&wsm->ws_wait);
 980
 981	/*
 982	 * Preallocate one workspace for each compression type so we can
 983	 * guarantee forward progress in the worst case
 984	 */
 985	workspace = alloc_workspace(type, 0);
 986	if (IS_ERR(workspace)) {
 987		pr_warn(
 988	"BTRFS: cannot preallocate compression workspace, will try later\n");
 989	} else {
 990		atomic_set(&wsm->total_ws, 1);
 991		wsm->free_ws = 1;
 992		list_add(workspace, &wsm->idle_ws);
 993	}
 994}
 995
 996static void btrfs_cleanup_workspace_manager(int type)
 997{
 998	struct workspace_manager *wsman;
 999	struct list_head *ws;
1000
1001	wsman = btrfs_compress_op[type]->workspace_manager;
1002	while (!list_empty(&wsman->idle_ws)) {
1003		ws = wsman->idle_ws.next;
1004		list_del(ws);
1005		free_workspace(type, ws);
1006		atomic_dec(&wsman->total_ws);
1007	}
1008}
1009
1010/*
1011 * This finds an available workspace or allocates a new one.
1012 * If it's not possible to allocate a new one, waits until there's one.
1013 * Preallocation makes a forward progress guarantees and we do not return
1014 * errors.
1015 */
1016struct list_head *btrfs_get_workspace(int type, unsigned int level)
1017{
1018	struct workspace_manager *wsm;
1019	struct list_head *workspace;
1020	int cpus = num_online_cpus();
1021	unsigned nofs_flag;
1022	struct list_head *idle_ws;
1023	spinlock_t *ws_lock;
1024	atomic_t *total_ws;
1025	wait_queue_head_t *ws_wait;
1026	int *free_ws;
1027
1028	wsm = btrfs_compress_op[type]->workspace_manager;
1029	idle_ws	 = &wsm->idle_ws;
1030	ws_lock	 = &wsm->ws_lock;
1031	total_ws = &wsm->total_ws;
1032	ws_wait	 = &wsm->ws_wait;
1033	free_ws	 = &wsm->free_ws;
1034
1035again:
1036	spin_lock(ws_lock);
1037	if (!list_empty(idle_ws)) {
1038		workspace = idle_ws->next;
1039		list_del(workspace);
1040		(*free_ws)--;
1041		spin_unlock(ws_lock);
1042		return workspace;
1043
1044	}
1045	if (atomic_read(total_ws) > cpus) {
1046		DEFINE_WAIT(wait);
1047
1048		spin_unlock(ws_lock);
1049		prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1050		if (atomic_read(total_ws) > cpus && !*free_ws)
1051			schedule();
1052		finish_wait(ws_wait, &wait);
1053		goto again;
1054	}
1055	atomic_inc(total_ws);
1056	spin_unlock(ws_lock);
1057
1058	/*
1059	 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1060	 * to turn it off here because we might get called from the restricted
1061	 * context of btrfs_compress_bio/btrfs_compress_pages
1062	 */
1063	nofs_flag = memalloc_nofs_save();
1064	workspace = alloc_workspace(type, level);
1065	memalloc_nofs_restore(nofs_flag);
1066
1067	if (IS_ERR(workspace)) {
1068		atomic_dec(total_ws);
1069		wake_up(ws_wait);
1070
1071		/*
1072		 * Do not return the error but go back to waiting. There's a
1073		 * workspace preallocated for each type and the compression
1074		 * time is bounded so we get to a workspace eventually. This
1075		 * makes our caller's life easier.
1076		 *
1077		 * To prevent silent and low-probability deadlocks (when the
1078		 * initial preallocation fails), check if there are any
1079		 * workspaces at all.
1080		 */
1081		if (atomic_read(total_ws) == 0) {
1082			static DEFINE_RATELIMIT_STATE(_rs,
1083					/* once per minute */ 60 * HZ,
1084					/* no burst */ 1);
1085
1086			if (__ratelimit(&_rs)) {
1087				pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1088			}
1089		}
1090		goto again;
1091	}
1092	return workspace;
1093}
1094
1095static struct list_head *get_workspace(int type, int level)
1096{
1097	switch (type) {
1098	case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1099	case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1100	case BTRFS_COMPRESS_LZO:  return btrfs_get_workspace(type, level);
1101	case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1102	default:
1103		/*
1104		 * This can't happen, the type is validated several times
1105		 * before we get here.
1106		 */
1107		BUG();
1108	}
1109}
1110
1111/*
1112 * put a workspace struct back on the list or free it if we have enough
1113 * idle ones sitting around
1114 */
1115void btrfs_put_workspace(int type, struct list_head *ws)
1116{
1117	struct workspace_manager *wsm;
1118	struct list_head *idle_ws;
1119	spinlock_t *ws_lock;
1120	atomic_t *total_ws;
1121	wait_queue_head_t *ws_wait;
1122	int *free_ws;
1123
1124	wsm = btrfs_compress_op[type]->workspace_manager;
1125	idle_ws	 = &wsm->idle_ws;
1126	ws_lock	 = &wsm->ws_lock;
1127	total_ws = &wsm->total_ws;
1128	ws_wait	 = &wsm->ws_wait;
1129	free_ws	 = &wsm->free_ws;
1130
1131	spin_lock(ws_lock);
1132	if (*free_ws <= num_online_cpus()) {
1133		list_add(ws, idle_ws);
1134		(*free_ws)++;
1135		spin_unlock(ws_lock);
1136		goto wake;
1137	}
1138	spin_unlock(ws_lock);
1139
1140	free_workspace(type, ws);
1141	atomic_dec(total_ws);
1142wake:
1143	cond_wake_up(ws_wait);
1144}
1145
1146static void put_workspace(int type, struct list_head *ws)
1147{
1148	switch (type) {
1149	case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1150	case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1151	case BTRFS_COMPRESS_LZO:  return btrfs_put_workspace(type, ws);
1152	case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1153	default:
1154		/*
1155		 * This can't happen, the type is validated several times
1156		 * before we get here.
1157		 */
1158		BUG();
1159	}
1160}
1161
1162/*
1163 * Adjust @level according to the limits of the compression algorithm or
1164 * fallback to default
1165 */
1166static unsigned int btrfs_compress_set_level(int type, unsigned level)
1167{
1168	const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1169
1170	if (level == 0)
1171		level = ops->default_level;
1172	else
1173		level = min(level, ops->max_level);
1174
1175	return level;
1176}
1177
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1178/*
1179 * Given an address space and start and length, compress the bytes into @pages
1180 * that are allocated on demand.
1181 *
1182 * @type_level is encoded algorithm and level, where level 0 means whatever
1183 * default the algorithm chooses and is opaque here;
1184 * - compression algo are 0-3
1185 * - the level are bits 4-7
1186 *
1187 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1188 * and returns number of actually allocated pages
1189 *
1190 * @total_in is used to return the number of bytes actually read.  It
1191 * may be smaller than the input length if we had to exit early because we
1192 * ran out of room in the pages array or because we cross the
1193 * max_out threshold.
1194 *
1195 * @total_out is an in/out parameter, must be set to the input length and will
1196 * be also used to return the total number of compressed bytes
1197 */
1198int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1199			 u64 start, struct page **pages,
1200			 unsigned long *out_pages,
1201			 unsigned long *total_in,
1202			 unsigned long *total_out)
1203{
1204	int type = btrfs_compress_type(type_level);
1205	int level = btrfs_compress_level(type_level);
 
1206	struct list_head *workspace;
1207	int ret;
1208
1209	level = btrfs_compress_set_level(type, level);
1210	workspace = get_workspace(type, level);
1211	ret = compression_compress_pages(type, workspace, mapping, start, pages,
1212					 out_pages, total_in, total_out);
 
 
1213	put_workspace(type, workspace);
1214	return ret;
1215}
1216
1217static int btrfs_decompress_bio(struct compressed_bio *cb)
1218{
1219	struct list_head *workspace;
1220	int ret;
1221	int type = cb->compress_type;
1222
1223	workspace = get_workspace(type, 0);
1224	ret = compression_decompress_bio(workspace, cb);
1225	put_workspace(type, workspace);
1226
 
 
1227	return ret;
1228}
1229
1230/*
1231 * a less complex decompression routine.  Our compressed data fits in a
1232 * single page, and we want to read a single page out of it.
1233 * start_byte tells us the offset into the compressed data we're interested in
1234 */
1235int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1236		     unsigned long start_byte, size_t srclen, size_t destlen)
1237{
 
1238	struct list_head *workspace;
 
1239	int ret;
1240
 
 
 
 
 
 
 
1241	workspace = get_workspace(type, 0);
1242	ret = compression_decompress(type, workspace, data_in, dest_page,
1243				     start_byte, srclen, destlen);
1244	put_workspace(type, workspace);
1245
1246	return ret;
1247}
1248
1249int __init btrfs_init_compress(void)
1250{
 
 
 
 
 
 
 
 
 
1251	btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1252	btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1253	btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1254	zstd_init_workspace_manager();
 
 
 
 
 
 
 
 
 
 
 
 
1255	return 0;
1256}
1257
1258void __cold btrfs_exit_compress(void)
1259{
 
 
 
 
1260	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1261	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1262	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1263	zstd_cleanup_workspace_manager();
 
1264}
1265
1266/*
1267 * Copy decompressed data from working buffer to pages.
1268 *
1269 * @buf:		The decompressed data buffer
1270 * @buf_len:		The decompressed data length
1271 * @decompressed:	Number of bytes that are already decompressed inside the
1272 * 			compressed extent
1273 * @cb:			The compressed extent descriptor
1274 * @orig_bio:		The original bio that the caller wants to read for
1275 *
1276 * An easier to understand graph is like below:
1277 *
1278 * 		|<- orig_bio ->|     |<- orig_bio->|
1279 * 	|<-------      full decompressed extent      ----->|
1280 * 	|<-----------    @cb range   ---->|
1281 * 	|			|<-- @buf_len -->|
1282 * 	|<--- @decompressed --->|
1283 *
1284 * Note that, @cb can be a subpage of the full decompressed extent, but
1285 * @cb->start always has the same as the orig_file_offset value of the full
1286 * decompressed extent.
1287 *
1288 * When reading compressed extent, we have to read the full compressed extent,
1289 * while @orig_bio may only want part of the range.
1290 * Thus this function will ensure only data covered by @orig_bio will be copied
1291 * to.
1292 *
1293 * Return 0 if we have copied all needed contents for @orig_bio.
1294 * Return >0 if we need continue decompress.
1295 */
1296int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1297			      struct compressed_bio *cb, u32 decompressed)
1298{
1299	struct bio *orig_bio = cb->orig_bio;
1300	/* Offset inside the full decompressed extent */
1301	u32 cur_offset;
1302
1303	cur_offset = decompressed;
1304	/* The main loop to do the copy */
1305	while (cur_offset < decompressed + buf_len) {
1306		struct bio_vec bvec;
1307		size_t copy_len;
1308		u32 copy_start;
1309		/* Offset inside the full decompressed extent */
1310		u32 bvec_offset;
1311
1312		bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1313		/*
1314		 * cb->start may underflow, but subtracting that value can still
1315		 * give us correct offset inside the full decompressed extent.
1316		 */
1317		bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1318
1319		/* Haven't reached the bvec range, exit */
1320		if (decompressed + buf_len <= bvec_offset)
1321			return 1;
1322
1323		copy_start = max(cur_offset, bvec_offset);
1324		copy_len = min(bvec_offset + bvec.bv_len,
1325			       decompressed + buf_len) - copy_start;
1326		ASSERT(copy_len);
1327
1328		/*
1329		 * Extra range check to ensure we didn't go beyond
1330		 * @buf + @buf_len.
1331		 */
1332		ASSERT(copy_start - decompressed < buf_len);
1333		memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1334			       buf + copy_start - decompressed, copy_len);
1335		cur_offset += copy_len;
1336
1337		bio_advance(orig_bio, copy_len);
1338		/* Finished the bio */
1339		if (!orig_bio->bi_iter.bi_size)
1340			return 0;
1341	}
1342	return 1;
1343}
1344
1345/*
1346 * Shannon Entropy calculation
1347 *
1348 * Pure byte distribution analysis fails to determine compressibility of data.
1349 * Try calculating entropy to estimate the average minimum number of bits
1350 * needed to encode the sampled data.
1351 *
1352 * For convenience, return the percentage of needed bits, instead of amount of
1353 * bits directly.
1354 *
1355 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1356 *			    and can be compressible with high probability
1357 *
1358 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1359 *
1360 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1361 */
1362#define ENTROPY_LVL_ACEPTABLE		(65)
1363#define ENTROPY_LVL_HIGH		(80)
1364
1365/*
1366 * For increasead precision in shannon_entropy calculation,
1367 * let's do pow(n, M) to save more digits after comma:
1368 *
1369 * - maximum int bit length is 64
1370 * - ilog2(MAX_SAMPLE_SIZE)	-> 13
1371 * - 13 * 4 = 52 < 64		-> M = 4
1372 *
1373 * So use pow(n, 4).
1374 */
1375static inline u32 ilog2_w(u64 n)
1376{
1377	return ilog2(n * n * n * n);
1378}
1379
1380static u32 shannon_entropy(struct heuristic_ws *ws)
1381{
1382	const u32 entropy_max = 8 * ilog2_w(2);
1383	u32 entropy_sum = 0;
1384	u32 p, p_base, sz_base;
1385	u32 i;
1386
1387	sz_base = ilog2_w(ws->sample_size);
1388	for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1389		p = ws->bucket[i].count;
1390		p_base = ilog2_w(p);
1391		entropy_sum += p * (sz_base - p_base);
1392	}
1393
1394	entropy_sum /= ws->sample_size;
1395	return entropy_sum * 100 / entropy_max;
1396}
1397
1398#define RADIX_BASE		4U
1399#define COUNTERS_SIZE		(1U << RADIX_BASE)
1400
1401static u8 get4bits(u64 num, int shift) {
1402	u8 low4bits;
1403
1404	num >>= shift;
1405	/* Reverse order */
1406	low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1407	return low4bits;
1408}
1409
1410/*
1411 * Use 4 bits as radix base
1412 * Use 16 u32 counters for calculating new position in buf array
1413 *
1414 * @array     - array that will be sorted
1415 * @array_buf - buffer array to store sorting results
1416 *              must be equal in size to @array
1417 * @num       - array size
1418 */
1419static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1420		       int num)
1421{
1422	u64 max_num;
1423	u64 buf_num;
1424	u32 counters[COUNTERS_SIZE];
1425	u32 new_addr;
1426	u32 addr;
1427	int bitlen;
1428	int shift;
1429	int i;
1430
1431	/*
1432	 * Try avoid useless loop iterations for small numbers stored in big
1433	 * counters.  Example: 48 33 4 ... in 64bit array
1434	 */
1435	max_num = array[0].count;
1436	for (i = 1; i < num; i++) {
1437		buf_num = array[i].count;
1438		if (buf_num > max_num)
1439			max_num = buf_num;
1440	}
1441
1442	buf_num = ilog2(max_num);
1443	bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1444
1445	shift = 0;
1446	while (shift < bitlen) {
1447		memset(counters, 0, sizeof(counters));
1448
1449		for (i = 0; i < num; i++) {
1450			buf_num = array[i].count;
1451			addr = get4bits(buf_num, shift);
1452			counters[addr]++;
1453		}
1454
1455		for (i = 1; i < COUNTERS_SIZE; i++)
1456			counters[i] += counters[i - 1];
1457
1458		for (i = num - 1; i >= 0; i--) {
1459			buf_num = array[i].count;
1460			addr = get4bits(buf_num, shift);
1461			counters[addr]--;
1462			new_addr = counters[addr];
1463			array_buf[new_addr] = array[i];
1464		}
1465
1466		shift += RADIX_BASE;
1467
1468		/*
1469		 * Normal radix expects to move data from a temporary array, to
1470		 * the main one.  But that requires some CPU time. Avoid that
1471		 * by doing another sort iteration to original array instead of
1472		 * memcpy()
1473		 */
1474		memset(counters, 0, sizeof(counters));
1475
1476		for (i = 0; i < num; i ++) {
1477			buf_num = array_buf[i].count;
1478			addr = get4bits(buf_num, shift);
1479			counters[addr]++;
1480		}
1481
1482		for (i = 1; i < COUNTERS_SIZE; i++)
1483			counters[i] += counters[i - 1];
1484
1485		for (i = num - 1; i >= 0; i--) {
1486			buf_num = array_buf[i].count;
1487			addr = get4bits(buf_num, shift);
1488			counters[addr]--;
1489			new_addr = counters[addr];
1490			array[new_addr] = array_buf[i];
1491		}
1492
1493		shift += RADIX_BASE;
1494	}
1495}
1496
1497/*
1498 * Size of the core byte set - how many bytes cover 90% of the sample
1499 *
1500 * There are several types of structured binary data that use nearly all byte
1501 * values. The distribution can be uniform and counts in all buckets will be
1502 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1503 *
1504 * Other possibility is normal (Gaussian) distribution, where the data could
1505 * be potentially compressible, but we have to take a few more steps to decide
1506 * how much.
1507 *
1508 * @BYTE_CORE_SET_LOW  - main part of byte values repeated frequently,
1509 *                       compression algo can easy fix that
1510 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1511 *                       probability is not compressible
1512 */
1513#define BYTE_CORE_SET_LOW		(64)
1514#define BYTE_CORE_SET_HIGH		(200)
1515
1516static int byte_core_set_size(struct heuristic_ws *ws)
1517{
1518	u32 i;
1519	u32 coreset_sum = 0;
1520	const u32 core_set_threshold = ws->sample_size * 90 / 100;
1521	struct bucket_item *bucket = ws->bucket;
1522
1523	/* Sort in reverse order */
1524	radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1525
1526	for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1527		coreset_sum += bucket[i].count;
1528
1529	if (coreset_sum > core_set_threshold)
1530		return i;
1531
1532	for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1533		coreset_sum += bucket[i].count;
1534		if (coreset_sum > core_set_threshold)
1535			break;
1536	}
1537
1538	return i;
1539}
1540
1541/*
1542 * Count byte values in buckets.
1543 * This heuristic can detect textual data (configs, xml, json, html, etc).
1544 * Because in most text-like data byte set is restricted to limited number of
1545 * possible characters, and that restriction in most cases makes data easy to
1546 * compress.
1547 *
1548 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1549 *	less - compressible
1550 *	more - need additional analysis
1551 */
1552#define BYTE_SET_THRESHOLD		(64)
1553
1554static u32 byte_set_size(const struct heuristic_ws *ws)
1555{
1556	u32 i;
1557	u32 byte_set_size = 0;
1558
1559	for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1560		if (ws->bucket[i].count > 0)
1561			byte_set_size++;
1562	}
1563
1564	/*
1565	 * Continue collecting count of byte values in buckets.  If the byte
1566	 * set size is bigger then the threshold, it's pointless to continue,
1567	 * the detection technique would fail for this type of data.
1568	 */
1569	for (; i < BUCKET_SIZE; i++) {
1570		if (ws->bucket[i].count > 0) {
1571			byte_set_size++;
1572			if (byte_set_size > BYTE_SET_THRESHOLD)
1573				return byte_set_size;
1574		}
1575	}
1576
1577	return byte_set_size;
1578}
1579
1580static bool sample_repeated_patterns(struct heuristic_ws *ws)
1581{
1582	const u32 half_of_sample = ws->sample_size / 2;
1583	const u8 *data = ws->sample;
1584
1585	return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1586}
1587
1588static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1589				     struct heuristic_ws *ws)
1590{
1591	struct page *page;
1592	u64 index, index_end;
1593	u32 i, curr_sample_pos;
1594	u8 *in_data;
1595
1596	/*
1597	 * Compression handles the input data by chunks of 128KiB
1598	 * (defined by BTRFS_MAX_UNCOMPRESSED)
1599	 *
1600	 * We do the same for the heuristic and loop over the whole range.
1601	 *
1602	 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1603	 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1604	 */
1605	if (end - start > BTRFS_MAX_UNCOMPRESSED)
1606		end = start + BTRFS_MAX_UNCOMPRESSED;
1607
1608	index = start >> PAGE_SHIFT;
1609	index_end = end >> PAGE_SHIFT;
1610
1611	/* Don't miss unaligned end */
1612	if (!IS_ALIGNED(end, PAGE_SIZE))
1613		index_end++;
1614
1615	curr_sample_pos = 0;
1616	while (index < index_end) {
1617		page = find_get_page(inode->i_mapping, index);
1618		in_data = kmap_local_page(page);
1619		/* Handle case where the start is not aligned to PAGE_SIZE */
1620		i = start % PAGE_SIZE;
1621		while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1622			/* Don't sample any garbage from the last page */
1623			if (start > end - SAMPLING_READ_SIZE)
1624				break;
1625			memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1626					SAMPLING_READ_SIZE);
1627			i += SAMPLING_INTERVAL;
1628			start += SAMPLING_INTERVAL;
1629			curr_sample_pos += SAMPLING_READ_SIZE;
1630		}
1631		kunmap_local(in_data);
1632		put_page(page);
1633
1634		index++;
1635	}
1636
1637	ws->sample_size = curr_sample_pos;
1638}
1639
1640/*
1641 * Compression heuristic.
1642 *
1643 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1644 * quickly (compared to direct compression) detect data characteristics
1645 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1646 * data.
1647 *
1648 * The following types of analysis can be performed:
1649 * - detect mostly zero data
1650 * - detect data with low "byte set" size (text, etc)
1651 * - detect data with low/high "core byte" set
1652 *
1653 * Return non-zero if the compression should be done, 0 otherwise.
1654 */
1655int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1656{
1657	struct list_head *ws_list = get_workspace(0, 0);
1658	struct heuristic_ws *ws;
1659	u32 i;
1660	u8 byte;
1661	int ret = 0;
1662
1663	ws = list_entry(ws_list, struct heuristic_ws, list);
1664
1665	heuristic_collect_sample(inode, start, end, ws);
1666
1667	if (sample_repeated_patterns(ws)) {
1668		ret = 1;
1669		goto out;
1670	}
1671
1672	memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1673
1674	for (i = 0; i < ws->sample_size; i++) {
1675		byte = ws->sample[i];
1676		ws->bucket[byte].count++;
1677	}
1678
1679	i = byte_set_size(ws);
1680	if (i < BYTE_SET_THRESHOLD) {
1681		ret = 2;
1682		goto out;
1683	}
1684
1685	i = byte_core_set_size(ws);
1686	if (i <= BYTE_CORE_SET_LOW) {
1687		ret = 3;
1688		goto out;
1689	}
1690
1691	if (i >= BYTE_CORE_SET_HIGH) {
1692		ret = 0;
1693		goto out;
1694	}
1695
1696	i = shannon_entropy(ws);
1697	if (i <= ENTROPY_LVL_ACEPTABLE) {
1698		ret = 4;
1699		goto out;
1700	}
1701
1702	/*
1703	 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1704	 * needed to give green light to compression.
1705	 *
1706	 * For now just assume that compression at that level is not worth the
1707	 * resources because:
1708	 *
1709	 * 1. it is possible to defrag the data later
1710	 *
1711	 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1712	 * values, every bucket has counter at level ~54. The heuristic would
1713	 * be confused. This can happen when data have some internal repeated
1714	 * patterns like "abbacbbc...". This can be detected by analyzing
1715	 * pairs of bytes, which is too costly.
1716	 */
1717	if (i < ENTROPY_LVL_HIGH) {
1718		ret = 5;
1719		goto out;
1720	} else {
1721		ret = 0;
1722		goto out;
1723	}
1724
1725out:
1726	put_workspace(0, ws_list);
1727	return ret;
1728}
1729
1730/*
1731 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1732 * level, unrecognized string will set the default level
1733 */
1734unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1735{
1736	unsigned int level = 0;
1737	int ret;
1738
1739	if (!type)
1740		return 0;
1741
1742	if (str[0] == ':') {
1743		ret = kstrtouint(str + 1, 10, &level);
1744		if (ret)
1745			level = 0;
1746	}
1747
1748	level = btrfs_compress_set_level(type, level);
1749
1750	return level;
1751}