blob: e6635fe7006789cfac33fabfec8cb8f7468687c2 [file] [log] [blame]
David Sterbac1d7c512018-04-03 19:23:33 +02001// SPDX-License-Identifier: GPL-2.0
Chris Masonc8b97812008-10-29 14:49:59 -04002/*
3 * Copyright (C) 2008 Oracle. All rights reserved.
Chris Masonc8b97812008-10-29 14:49:59 -04004 */
5
6#include <linux/kernel.h>
7#include <linux/bio.h>
Chris Masonc8b97812008-10-29 14:49:59 -04008#include <linux/file.h>
9#include <linux/fs.h>
10#include <linux/pagemap.h>
Vishal Moola (Oracle)a75b81c2022-08-23 17:40:19 -070011#include <linux/pagevec.h>
Chris Masonc8b97812008-10-29 14:49:59 -040012#include <linux/highmem.h>
Christoph Hellwige41d12f2021-09-20 14:33:13 +020013#include <linux/kthread.h>
Chris Masonc8b97812008-10-29 14:49:59 -040014#include <linux/time.h>
15#include <linux/init.h>
16#include <linux/string.h>
Chris Masonc8b97812008-10-29 14:49:59 -040017#include <linux/backing-dev.h>
Chris Masonc8b97812008-10-29 14:49:59 -040018#include <linux/writeback.h>
Christoph Hellwig4088a472022-09-15 10:41:58 +010019#include <linux/psi.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
David Sterbafe308532017-05-31 17:14:56 +020021#include <linux/sched/mm.h>
Timofey Titovets19562432017-10-08 16:11:59 +030022#include <linux/log2.h>
Johannes Thumshirnd5178572019-06-03 16:58:57 +020023#include <crypto/hash.h>
David Sterba602cbe92019-08-21 18:48:25 +020024#include "misc.h"
Chris Masonc8b97812008-10-29 14:49:59 -040025#include "ctree.h"
26#include "disk-io.h"
27#include "transaction.h"
28#include "btrfs_inode.h"
29#include "volumes.h"
30#include "ordered-data.h"
Chris Masonc8b97812008-10-29 14:49:59 -040031#include "compression.h"
32#include "extent_io.h"
33#include "extent_map.h"
Qu Wenruo6a404912021-09-27 15:21:47 +080034#include "subpage.h"
Johannes Thumshirn764c7c92021-05-19 00:40:28 +090035#include "zoned.h"
Chris Masonc8b97812008-10-29 14:49:59 -040036
David Sterbae128f9c2017-10-31 17:24:26 +010037static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
38
39const char* btrfs_compress_type2str(enum btrfs_compression_type type)
40{
41 switch (type) {
42 case BTRFS_COMPRESS_ZLIB:
43 case BTRFS_COMPRESS_LZO:
44 case BTRFS_COMPRESS_ZSTD:
45 case BTRFS_COMPRESS_NONE:
46 return btrfs_compress_types[type];
Chengguang Xuce96b7f2019-10-10 15:59:57 +080047 default:
48 break;
David Sterbae128f9c2017-10-31 17:24:26 +010049 }
50
51 return NULL;
52}
53
Johannes Thumshirnaa53e3b2019-06-06 12:07:15 +020054bool btrfs_compress_is_valid_type(const char *str, size_t len)
55{
56 int i;
57
58 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
59 size_t comp_len = strlen(btrfs_compress_types[i]);
60
61 if (len < comp_len)
62 continue;
63
64 if (!strncmp(btrfs_compress_types[i], str, comp_len))
65 return true;
66 }
67 return false;
68}
69
David Sterba1e4eb742019-10-02 00:06:15 +020070static int compression_compress_pages(int type, struct list_head *ws,
71 struct address_space *mapping, u64 start, struct page **pages,
72 unsigned long *out_pages, unsigned long *total_in,
73 unsigned long *total_out)
74{
75 switch (type) {
76 case BTRFS_COMPRESS_ZLIB:
77 return zlib_compress_pages(ws, mapping, start, pages,
78 out_pages, total_in, total_out);
79 case BTRFS_COMPRESS_LZO:
80 return lzo_compress_pages(ws, mapping, start, pages,
81 out_pages, total_in, total_out);
82 case BTRFS_COMPRESS_ZSTD:
83 return zstd_compress_pages(ws, mapping, start, pages,
84 out_pages, total_in, total_out);
85 case BTRFS_COMPRESS_NONE:
86 default:
87 /*
Qu Wenruo1d8ba9e2020-08-04 15:25:47 +080088 * This can happen when compression races with remount setting
89 * it to 'no compress', while caller doesn't call
90 * inode_need_compress() to check if we really need to
91 * compress.
92 *
93 * Not a big deal, just need to inform caller that we
94 * haven't allocated any pages yet.
David Sterba1e4eb742019-10-02 00:06:15 +020095 */
Qu Wenruo1d8ba9e2020-08-04 15:25:47 +080096 *out_pages = 0;
David Sterba1e4eb742019-10-02 00:06:15 +020097 return -E2BIG;
98 }
99}
100
Su Yue4a9e8032021-12-27 18:18:39 +0800101static int compression_decompress_bio(struct list_head *ws,
102 struct compressed_bio *cb)
David Sterba1e4eb742019-10-02 00:06:15 +0200103{
Su Yue4a9e8032021-12-27 18:18:39 +0800104 switch (cb->compress_type) {
David Sterba1e4eb742019-10-02 00:06:15 +0200105 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
106 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
107 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
108 case BTRFS_COMPRESS_NONE:
109 default:
110 /*
111 * This can't happen, the type is validated several times
112 * before we get here.
113 */
114 BUG();
115 }
116}
117
118static int compression_decompress(int type, struct list_head *ws,
119 unsigned char *data_in, struct page *dest_page,
120 unsigned long start_byte, size_t srclen, size_t destlen)
121{
122 switch (type) {
123 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
124 start_byte, srclen, destlen);
125 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
126 start_byte, srclen, destlen);
127 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
128 start_byte, srclen, destlen);
129 case BTRFS_COMPRESS_NONE:
130 default:
131 /*
132 * This can't happen, the type is validated several times
133 * before we get here.
134 */
135 BUG();
136 }
137}
138
Anand Jain8140dc32017-05-26 15:44:58 +0800139static int btrfs_decompress_bio(struct compressed_bio *cb);
Eric Sandeen48a3b632013-04-25 20:41:01 +0000140
Josef Bacike14bfdb2022-02-18 10:03:25 -0500141static void finish_compressed_bio_read(struct compressed_bio *cb)
Qu Wenruo86ccbb42021-09-27 15:21:50 +0800142{
143 unsigned int index;
144 struct page *page;
145
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200146 if (cb->status == BLK_STS_OK)
147 cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
148
Qu Wenruo86ccbb42021-09-27 15:21:50 +0800149 /* Release the compressed pages */
150 for (index = 0; index < cb->nr_pages; index++) {
151 page = cb->compressed_pages[index];
152 page->mapping = NULL;
153 put_page(page);
154 }
155
156 /* Do io completion on the original bio */
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200157 btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status);
Qu Wenruo86ccbb42021-09-27 15:21:50 +0800158
159 /* Finally free the cb struct */
160 kfree(cb->compressed_pages);
161 kfree(cb);
162}
163
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200164/*
165 * Verify the checksums and kick off repair if needed on the uncompressed data
166 * before decompressing it into the original bio and freeing the uncompressed
167 * pages.
Chris Masonc8b97812008-10-29 14:49:59 -0400168 */
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200169static void end_compressed_bio_read(struct btrfs_bio *bbio)
Chris Masonc8b97812008-10-29 14:49:59 -0400170{
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200171 struct compressed_bio *cb = bbio->private;
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200172 struct inode *inode = cb->inode;
173 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
174 struct btrfs_inode *bi = BTRFS_I(inode);
175 bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) &&
176 !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state);
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200177 blk_status_t status = bbio->bio.bi_status;
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200178 struct bvec_iter iter;
179 struct bio_vec bv;
180 u32 offset;
Chris Masonc8b97812008-10-29 14:49:59 -0400181
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200182 btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) {
183 u64 start = bbio->file_offset + offset;
Chris Masonc8b97812008-10-29 14:49:59 -0400184
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200185 if (!status &&
186 (!csum || !btrfs_check_data_csum(inode, bbio, offset,
187 bv.bv_page, bv.bv_offset))) {
Josef Bacik0d0a7622022-09-09 17:53:14 -0400188 btrfs_clean_io_failure(bi, start, bv.bv_page,
189 bv.bv_offset);
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200190 } else {
191 int ret;
Liu Bocf1167d2017-09-20 17:50:18 -0600192
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200193 refcount_inc(&cb->pending_ios);
194 ret = btrfs_repair_one_sector(inode, bbio, offset,
195 bv.bv_page, bv.bv_offset,
196 btrfs_submit_data_read_bio);
197 if (ret) {
198 refcount_dec(&cb->pending_ios);
199 status = errno_to_blk_status(ret);
200 }
201 }
202 }
Liu Boe6311f22017-09-20 17:50:19 -0600203
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200204 if (status)
205 cb->status = status;
Chris Masond20f7042008-12-08 16:58:54 -0500206
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200207 if (refcount_dec_and_test(&cb->pending_ios))
208 finish_compressed_bio_read(cb);
209 btrfs_bio_free_csum(bbio);
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200210 bio_put(&bbio->bio);
Chris Masonc8b97812008-10-29 14:49:59 -0400211}
212
213/*
214 * Clear the writeback bits on all of the file
215 * pages for a compressed write
216 */
Filipe Manana7bdcefc12014-10-07 01:48:26 +0100217static noinline void end_compressed_writeback(struct inode *inode,
218 const struct compressed_bio *cb)
Chris Masonc8b97812008-10-29 14:49:59 -0400219{
Qu Wenruo741ec652021-09-27 15:22:01 +0800220 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300221 unsigned long index = cb->start >> PAGE_SHIFT;
222 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
Vishal Moola (Oracle)a75b81c2022-08-23 17:40:19 -0700223 struct folio_batch fbatch;
Josef Bacik606f82e2022-02-18 10:03:26 -0500224 const int errno = blk_status_to_errno(cb->status);
Chris Masonc8b97812008-10-29 14:49:59 -0400225 int i;
226 int ret;
227
Josef Bacik606f82e2022-02-18 10:03:26 -0500228 if (errno)
229 mapping_set_error(inode->i_mapping, errno);
Filipe Manana7bdcefc12014-10-07 01:48:26 +0100230
Vishal Moola (Oracle)a75b81c2022-08-23 17:40:19 -0700231 folio_batch_init(&fbatch);
232 while (index <= end_index) {
233 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
234 &fbatch);
235
236 if (ret == 0)
237 return;
238
Chris Masonc8b97812008-10-29 14:49:59 -0400239 for (i = 0; i < ret; i++) {
Vishal Moola (Oracle)a75b81c2022-08-23 17:40:19 -0700240 struct folio *folio = fbatch.folios[i];
241
Josef Bacik606f82e2022-02-18 10:03:26 -0500242 if (errno)
Vishal Moola (Oracle)a75b81c2022-08-23 17:40:19 -0700243 folio_set_error(folio);
244 btrfs_page_clamp_clear_writeback(fs_info, &folio->page,
Qu Wenruo741ec652021-09-27 15:22:01 +0800245 cb->start, cb->len);
Chris Masonc8b97812008-10-29 14:49:59 -0400246 }
Vishal Moola (Oracle)a75b81c2022-08-23 17:40:19 -0700247 folio_batch_release(&fbatch);
Chris Masonc8b97812008-10-29 14:49:59 -0400248 }
249 /* the inode may be gone now */
Chris Masonc8b97812008-10-29 14:49:59 -0400250}
251
Qu Wenruo6853c642021-09-27 15:21:51 +0800252static void finish_compressed_bio_write(struct compressed_bio *cb)
Chris Masonc8b97812008-10-29 14:49:59 -0400253{
Qu Wenruo6853c642021-09-27 15:21:51 +0800254 struct inode *inode = cb->inode;
Anand Jain1d08ce582021-05-29 17:48:33 +0800255 unsigned int index;
Chris Masonc8b97812008-10-29 14:49:59 -0400256
Qu Wenruo6853c642021-09-27 15:21:51 +0800257 /*
258 * Ok, we're the last bio for this extent, step one is to call back
259 * into the FS and do all the end_io operations.
Chris Masonc8b97812008-10-29 14:49:59 -0400260 */
Qu Wenruo38a39ac72021-04-08 20:32:27 +0800261 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
Nikolay Borisovc6297322018-11-08 10:18:08 +0200262 cb->start, cb->start + cb->len - 1,
Josef Bacik606f82e2022-02-18 10:03:26 -0500263 cb->status == BLK_STS_OK);
Chris Masonc8b97812008-10-29 14:49:59 -0400264
Omar Sandoval7c0c7262019-08-13 16:00:02 -0700265 if (cb->writeback)
266 end_compressed_writeback(inode, cb);
Qu Wenruo6853c642021-09-27 15:21:51 +0800267 /* Note, our inode could be gone now */
Chris Masonc8b97812008-10-29 14:49:59 -0400268
269 /*
Qu Wenruo6853c642021-09-27 15:21:51 +0800270 * Release the compressed pages, these came from alloc_page and
Chris Masonc8b97812008-10-29 14:49:59 -0400271 * are not attached to the inode at all
272 */
Chris Masonc8b97812008-10-29 14:49:59 -0400273 for (index = 0; index < cb->nr_pages; index++) {
Qu Wenruo6853c642021-09-27 15:21:51 +0800274 struct page *page = cb->compressed_pages[index];
275
Chris Masonc8b97812008-10-29 14:49:59 -0400276 page->mapping = NULL;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300277 put_page(page);
Chris Masonc8b97812008-10-29 14:49:59 -0400278 }
279
Qu Wenruo6853c642021-09-27 15:21:51 +0800280 /* Finally free the cb struct */
Chris Masonc8b97812008-10-29 14:49:59 -0400281 kfree(cb->compressed_pages);
282 kfree(cb);
Qu Wenruo6853c642021-09-27 15:21:51 +0800283}
284
Christoph Hellwigfed8a722022-05-26 09:36:38 +0200285static void btrfs_finish_compressed_write_work(struct work_struct *work)
286{
287 struct compressed_bio *cb =
288 container_of(work, struct compressed_bio, write_end_work);
289
290 finish_compressed_bio_write(cb);
291}
292
Qu Wenruo6853c642021-09-27 15:21:51 +0800293/*
294 * Do the cleanup once all the compressed pages hit the disk. This will clear
295 * writeback on the file pages and free the compressed pages.
296 *
297 * This also calls the writeback end hooks for the file pages so that metadata
298 * and checksums can be updated in the file.
299 */
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200300static void end_compressed_bio_write(struct btrfs_bio *bbio)
Qu Wenruo6853c642021-09-27 15:21:51 +0800301{
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200302 struct compressed_bio *cb = bbio->private;
Qu Wenruo6853c642021-09-27 15:21:51 +0800303
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200304 if (bbio->bio.bi_status)
305 cb->status = bbio->bio.bi_status;
Qu Wenruo6853c642021-09-27 15:21:51 +0800306
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200307 if (refcount_dec_and_test(&cb->pending_ios)) {
Christoph Hellwigfed8a722022-05-26 09:36:38 +0200308 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
Qu Wenruo6853c642021-09-27 15:21:51 +0800309
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200310 btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio);
Christoph Hellwigfed8a722022-05-26 09:36:38 +0200311 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
312 }
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200313 bio_put(&bbio->bio);
Chris Masonc8b97812008-10-29 14:49:59 -0400314}
315
316/*
Qu Wenruof472c28f2021-09-27 15:21:54 +0800317 * Allocate a compressed_bio, which will be used to read/write on-disk
318 * (aka, compressed) * data.
319 *
320 * @cb: The compressed_bio structure, which records all the needed
321 * information to bind the compressed data to the uncompressed
322 * page cache.
323 * @disk_byten: The logical bytenr where the compressed data will be read
324 * from or written to.
325 * @endio_func: The endio function to call after the IO for compressed data
326 * is finished.
327 * @next_stripe_start: Return value of logical bytenr of where next stripe starts.
328 * Let the caller know to only fill the bio up to the stripe
329 * boundary.
Qu Wenruo22c306f2021-09-27 15:21:53 +0800330 */
Qu Wenruof472c28f2021-09-27 15:21:54 +0800331
332
Qu Wenruo22c306f2021-09-27 15:21:53 +0800333static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr,
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200334 blk_opf_t opf,
335 btrfs_bio_end_io_t endio_func,
Qu Wenruof472c28f2021-09-27 15:21:54 +0800336 u64 *next_stripe_start)
Qu Wenruo22c306f2021-09-27 15:21:53 +0800337{
Qu Wenruof472c28f2021-09-27 15:21:54 +0800338 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
339 struct btrfs_io_geometry geom;
340 struct extent_map *em;
Qu Wenruo22c306f2021-09-27 15:21:53 +0800341 struct bio *bio;
Qu Wenruof472c28f2021-09-27 15:21:54 +0800342 int ret;
Qu Wenruo22c306f2021-09-27 15:21:53 +0800343
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200344 bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, endio_func, cb);
Qu Wenruo22c306f2021-09-27 15:21:53 +0800345 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
Qu Wenruo22c306f2021-09-27 15:21:53 +0800346
Qu Wenruof472c28f2021-09-27 15:21:54 +0800347 em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize);
348 if (IS_ERR(em)) {
349 bio_put(bio);
350 return ERR_CAST(em);
Qu Wenruo22c306f2021-09-27 15:21:53 +0800351 }
Qu Wenruof472c28f2021-09-27 15:21:54 +0800352
353 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
354 bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev);
355
356 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom);
357 free_extent_map(em);
358 if (ret < 0) {
359 bio_put(bio);
360 return ERR_PTR(ret);
361 }
362 *next_stripe_start = disk_bytenr + geom.len;
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200363 refcount_inc(&cb->pending_ios);
Qu Wenruo22c306f2021-09-27 15:21:53 +0800364 return bio;
365}
366
Chris Masonc8b97812008-10-29 14:49:59 -0400367/*
368 * worker function to build and submit bios for previously compressed pages.
369 * The corresponding pages in the inode should be marked for writeback
370 * and the compressed pages should have a reference on them for dropping
371 * when the IO is complete.
372 *
373 * This also checksums the file bytes and gets things ready for
374 * the end io hooks.
375 */
Nikolay Borisovc7ee1812020-06-03 08:55:16 +0300376blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
Anand Jain65b53552021-05-29 17:48:35 +0800377 unsigned int len, u64 disk_start,
378 unsigned int compressed_len,
Chris Masonc8b97812008-10-29 14:49:59 -0400379 struct page **compressed_pages,
Anand Jain65b53552021-05-29 17:48:35 +0800380 unsigned int nr_pages,
Bart Van Asschebf9486d2022-07-14 11:07:16 -0700381 blk_opf_t write_flags,
Omar Sandoval7c0c7262019-08-13 16:00:02 -0700382 struct cgroup_subsys_state *blkcg_css,
383 bool writeback)
Chris Masonc8b97812008-10-29 14:49:59 -0400384{
Nikolay Borisovc7ee1812020-06-03 08:55:16 +0300385 struct btrfs_fs_info *fs_info = inode->root->fs_info;
Chris Masonc8b97812008-10-29 14:49:59 -0400386 struct bio *bio = NULL;
Chris Masonc8b97812008-10-29 14:49:59 -0400387 struct compressed_bio *cb;
Qu Wenruo91507242021-09-27 15:21:55 +0800388 u64 cur_disk_bytenr = disk_start;
Qu Wenruof472c28f2021-09-27 15:21:54 +0800389 u64 next_stripe_start;
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200390 blk_status_t ret = BLK_STS_OK;
Nikolay Borisovc7ee1812020-06-03 08:55:16 +0300391 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
Johannes Thumshirn764c7c92021-05-19 00:40:28 +0900392 const bool use_append = btrfs_use_zone_append(inode, disk_start);
Bart Van Asschebf9486d2022-07-14 11:07:16 -0700393 const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
Chris Masonc8b97812008-10-29 14:49:59 -0400394
Qu Wenruobbbff012021-09-27 15:22:00 +0800395 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
396 IS_ALIGNED(len, fs_info->sectorsize));
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200397 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
Yoshinori Sanodac97e52011-02-15 12:01:42 +0000398 if (!cb)
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200399 return BLK_STS_RESOURCE;
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200400 refcount_set(&cb->pending_ios, 1);
Josef Bacik606f82e2022-02-18 10:03:26 -0500401 cb->status = BLK_STS_OK;
Nikolay Borisovc7ee1812020-06-03 08:55:16 +0300402 cb->inode = &inode->vfs_inode;
Chris Masonc8b97812008-10-29 14:49:59 -0400403 cb->start = start;
404 cb->len = len;
405 cb->compressed_pages = compressed_pages;
406 cb->compressed_len = compressed_len;
Omar Sandoval7c0c7262019-08-13 16:00:02 -0700407 cb->writeback = writeback;
Christoph Hellwigfed8a722022-05-26 09:36:38 +0200408 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
Chris Masonc8b97812008-10-29 14:49:59 -0400409 cb->nr_pages = nr_pages;
410
Dennis Zhouacee08a2022-03-31 14:58:28 -0700411 if (blkcg_css)
412 kthread_associate_blkcg(blkcg_css);
413
Qu Wenruo91507242021-09-27 15:21:55 +0800414 while (cur_disk_bytenr < disk_start + compressed_len) {
415 u64 offset = cur_disk_bytenr - disk_start;
416 unsigned int index = offset >> PAGE_SHIFT;
417 unsigned int real_size;
418 unsigned int added;
419 struct page *page = compressed_pages[index];
420 bool submit = false;
Chris Masonec39f762019-07-10 12:28:17 -0700421
Qu Wenruo91507242021-09-27 15:21:55 +0800422 /* Allocate new bio if submitted or not yet allocated */
423 if (!bio) {
424 bio = alloc_compressed_bio(cb, cur_disk_bytenr,
425 bio_op | write_flags, end_compressed_bio_write,
426 &next_stripe_start);
427 if (IS_ERR(bio)) {
428 ret = errno_to_blk_status(PTR_ERR(bio));
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200429 break;
Qu Wenruo91507242021-09-27 15:21:55 +0800430 }
Dennis Zhouacee08a2022-03-31 14:58:28 -0700431 if (blkcg_css)
432 bio->bi_opf |= REQ_CGROUP_PUNT;
Johannes Thumshirn764c7c92021-05-19 00:40:28 +0900433 }
Qu Wenruo91507242021-09-27 15:21:55 +0800434 /*
435 * We should never reach next_stripe_start start as we will
436 * submit comp_bio when reach the boundary immediately.
437 */
438 ASSERT(cur_disk_bytenr != next_stripe_start);
Chris Masonc8b97812008-10-29 14:49:59 -0400439
Qu Wenruo4c80a972021-05-25 13:52:43 +0800440 /*
Qu Wenruo91507242021-09-27 15:21:55 +0800441 * We have various limits on the real read size:
442 * - stripe boundary
443 * - page boundary
444 * - compressed length boundary
Qu Wenruo4c80a972021-05-25 13:52:43 +0800445 */
Qu Wenruo91507242021-09-27 15:21:55 +0800446 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr);
447 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
448 real_size = min_t(u64, real_size, compressed_len - offset);
449 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
Johannes Thumshirn764c7c92021-05-19 00:40:28 +0900450
Qu Wenruo91507242021-09-27 15:21:55 +0800451 if (use_append)
452 added = bio_add_zone_append_page(bio, page, real_size,
453 offset_in_page(offset));
454 else
455 added = bio_add_page(bio, page, real_size,
456 offset_in_page(offset));
457 /* Reached zoned boundary */
458 if (added == 0)
459 submit = true;
Chris Masonc8b97812008-10-29 14:49:59 -0400460
Qu Wenruo91507242021-09-27 15:21:55 +0800461 cur_disk_bytenr += added;
462 /* Reached stripe boundary */
463 if (cur_disk_bytenr == next_stripe_start)
464 submit = true;
465
466 /* Finished the range */
467 if (cur_disk_bytenr == disk_start + compressed_len)
468 submit = true;
469
470 if (submit) {
Li Zefane55179b2011-07-14 03:16:47 +0000471 if (!skip_sum) {
Omar Sandovale331f6b2019-11-06 15:38:43 -0800472 ret = btrfs_csum_one_bio(inode, bio, start, true);
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200473 if (ret) {
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200474 btrfs_bio_end_io(btrfs_bio(bio), ret);
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200475 break;
476 }
Li Zefane55179b2011-07-14 03:16:47 +0000477 }
Chris Masond20f7042008-12-08 16:58:54 -0500478
Christoph Hellwigfed8a722022-05-26 09:36:38 +0200479 ASSERT(bio->bi_iter.bi_size);
Christoph Hellwig1a722d82022-06-17 12:04:07 +0200480 btrfs_submit_bio(fs_info, bio, 0);
Qu Wenruo91507242021-09-27 15:21:55 +0800481 bio = NULL;
Chris Masonc8b97812008-10-29 14:49:59 -0400482 }
Chris Mason771ed682008-11-06 22:02:51 -0500483 cond_resched();
Chris Masonc8b97812008-10-29 14:49:59 -0400484 }
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200485
Dennis Zhou46bcff2b2019-12-11 15:20:15 -0800486 if (blkcg_css)
487 kthread_associate_blkcg(NULL);
488
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200489 if (refcount_dec_and_test(&cb->pending_ios))
490 finish_compressed_bio_write(cb);
Qu Wenruo6853c642021-09-27 15:21:51 +0800491 return ret;
Chris Masonc8b97812008-10-29 14:49:59 -0400492}
493
Christoph Hellwig2a4d0c92016-11-25 09:07:51 +0100494static u64 bio_end_offset(struct bio *bio)
495{
Ming Leic45a8f22017-12-18 20:22:05 +0800496 struct bio_vec *last = bio_last_bvec_all(bio);
Christoph Hellwig2a4d0c92016-11-25 09:07:51 +0100497
498 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
499}
500
Qu Wenruo6a404912021-09-27 15:21:47 +0800501/*
502 * Add extra pages in the same compressed file extent so that we don't need to
503 * re-read the same extent again and again.
504 *
505 * NOTE: this won't work well for subpage, as for subpage read, we lock the
506 * full page then submit bio for each compressed/regular extents.
507 *
508 * This means, if we have several sectors in the same page points to the same
509 * on-disk compressed data, we will re-read the same extent many times and
510 * this function can only help for the next page.
511 */
Chris Mason771ed682008-11-06 22:02:51 -0500512static noinline int add_ra_bio_pages(struct inode *inode,
513 u64 compressed_end,
Christoph Hellwig4088a472022-09-15 10:41:58 +0100514 struct compressed_bio *cb,
Johannes Weiner82e60d02022-11-03 17:34:31 -0400515 int *memstall, unsigned long *pflags)
Chris Mason771ed682008-11-06 22:02:51 -0500516{
Qu Wenruo6a404912021-09-27 15:21:47 +0800517 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
Chris Mason771ed682008-11-06 22:02:51 -0500518 unsigned long end_index;
Qu Wenruo6a404912021-09-27 15:21:47 +0800519 u64 cur = bio_end_offset(cb->orig_bio);
Chris Mason771ed682008-11-06 22:02:51 -0500520 u64 isize = i_size_read(inode);
521 int ret;
522 struct page *page;
Chris Mason771ed682008-11-06 22:02:51 -0500523 struct extent_map *em;
524 struct address_space *mapping = inode->i_mapping;
Chris Mason771ed682008-11-06 22:02:51 -0500525 struct extent_map_tree *em_tree;
526 struct extent_io_tree *tree;
Qu Wenruo6a404912021-09-27 15:21:47 +0800527 int sectors_missed = 0;
Chris Mason771ed682008-11-06 22:02:51 -0500528
Chris Mason771ed682008-11-06 22:02:51 -0500529 em_tree = &BTRFS_I(inode)->extent_tree;
530 tree = &BTRFS_I(inode)->io_tree;
531
532 if (isize == 0)
533 return 0;
534
Qu Wenruoca62e852021-07-26 14:34:52 +0800535 /*
536 * For current subpage support, we only support 64K page size,
537 * which means maximum compressed extent size (128K) is just 2x page
538 * size.
539 * This makes readahead less effective, so here disable readahead for
540 * subpage for now, until full compressed write is supported.
541 */
542 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
543 return 0;
544
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300545 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
Chris Mason771ed682008-11-06 22:02:51 -0500546
Qu Wenruo6a404912021-09-27 15:21:47 +0800547 while (cur < compressed_end) {
548 u64 page_end;
549 u64 pg_index = cur >> PAGE_SHIFT;
550 u32 add_size;
Chris Mason771ed682008-11-06 22:02:51 -0500551
David Sterba306e16c2011-04-19 14:29:38 +0200552 if (pg_index > end_index)
Chris Mason771ed682008-11-06 22:02:51 -0500553 break;
554
Matthew Wilcox0a943c62017-12-04 10:37:22 -0500555 page = xa_load(&mapping->i_pages, pg_index);
Matthew Wilcox3159f942017-11-03 13:30:42 -0400556 if (page && !xa_is_value(page)) {
Qu Wenruo6a404912021-09-27 15:21:47 +0800557 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
558 fs_info->sectorsize_bits;
559
560 /* Beyond threshold, no need to continue */
561 if (sectors_missed > 4)
Chris Mason771ed682008-11-06 22:02:51 -0500562 break;
Qu Wenruo6a404912021-09-27 15:21:47 +0800563
564 /*
565 * Jump to next page start as we already have page for
566 * current offset.
567 */
568 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
569 continue;
Chris Mason771ed682008-11-06 22:02:51 -0500570 }
571
Michal Hockoc62d2552015-11-06 16:28:49 -0800572 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
573 ~__GFP_FS));
Chris Mason771ed682008-11-06 22:02:51 -0500574 if (!page)
575 break;
576
Michal Hockoc62d2552015-11-06 16:28:49 -0800577 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300578 put_page(page);
Qu Wenruo6a404912021-09-27 15:21:47 +0800579 /* There is already a page, skip to page end */
580 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
581 continue;
Chris Mason771ed682008-11-06 22:02:51 -0500582 }
583
Johannes Weiner82e60d02022-11-03 17:34:31 -0400584 if (!*memstall && PageWorkingset(page)) {
Christoph Hellwig4088a472022-09-15 10:41:58 +0100585 psi_memstall_enter(pflags);
Johannes Weiner82e60d02022-11-03 17:34:31 -0400586 *memstall = 1;
587 }
Christoph Hellwig4088a472022-09-15 10:41:58 +0100588
Qu Wenruo32443de2021-01-26 16:34:00 +0800589 ret = set_page_extent_mapped(page);
590 if (ret < 0) {
591 unlock_page(page);
592 put_page(page);
593 break;
594 }
595
Qu Wenruo6a404912021-09-27 15:21:47 +0800596 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
Josef Bacik570eb972022-09-09 17:53:43 -0400597 lock_extent(tree, cur, page_end, NULL);
Chris Mason890871b2009-09-02 16:24:52 -0400598 read_lock(&em_tree->lock);
Qu Wenruo6a404912021-09-27 15:21:47 +0800599 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
Chris Mason890871b2009-09-02 16:24:52 -0400600 read_unlock(&em_tree->lock);
Chris Mason771ed682008-11-06 22:02:51 -0500601
Qu Wenruo6a404912021-09-27 15:21:47 +0800602 /*
603 * At this point, we have a locked page in the page cache for
604 * these bytes in the file. But, we have to make sure they map
605 * to this compressed extent on disk.
606 */
607 if (!em || cur < em->start ||
608 (cur + fs_info->sectorsize > extent_map_end(em)) ||
Kent Overstreet4f024f32013-10-11 15:44:27 -0700609 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
Chris Mason771ed682008-11-06 22:02:51 -0500610 free_extent_map(em);
Josef Bacik570eb972022-09-09 17:53:43 -0400611 unlock_extent(tree, cur, page_end, NULL);
Chris Mason771ed682008-11-06 22:02:51 -0500612 unlock_page(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300613 put_page(page);
Chris Mason771ed682008-11-06 22:02:51 -0500614 break;
615 }
616 free_extent_map(em);
617
618 if (page->index == end_index) {
Johannes Thumshirn70730172018-12-05 15:23:03 +0100619 size_t zero_offset = offset_in_page(isize);
Chris Mason771ed682008-11-06 22:02:51 -0500620
621 if (zero_offset) {
622 int zeros;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300623 zeros = PAGE_SIZE - zero_offset;
Ira Weinyd048b9c2021-05-04 18:40:07 -0700624 memzero_page(page, zero_offset, zeros);
Chris Mason771ed682008-11-06 22:02:51 -0500625 }
626 }
627
Qu Wenruo6a404912021-09-27 15:21:47 +0800628 add_size = min(em->start + em->len, page_end + 1) - cur;
629 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
630 if (ret != add_size) {
Josef Bacik570eb972022-09-09 17:53:43 -0400631 unlock_extent(tree, cur, page_end, NULL);
Chris Mason771ed682008-11-06 22:02:51 -0500632 unlock_page(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300633 put_page(page);
Chris Mason771ed682008-11-06 22:02:51 -0500634 break;
635 }
Qu Wenruo6a404912021-09-27 15:21:47 +0800636 /*
637 * If it's subpage, we also need to increase its
638 * subpage::readers number, as at endio we will decrease
639 * subpage::readers and to unlock the page.
640 */
641 if (fs_info->sectorsize < PAGE_SIZE)
642 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
643 put_page(page);
644 cur += add_size;
Chris Mason771ed682008-11-06 22:02:51 -0500645 }
Chris Mason771ed682008-11-06 22:02:51 -0500646 return 0;
647}
648
Chris Masonc8b97812008-10-29 14:49:59 -0400649/*
650 * for a compressed read, the bio we get passed has all the inode pages
651 * in it. We don't actually do IO on those pages but allocate new ones
652 * to hold the compressed pages on disk.
653 *
Kent Overstreet4f024f32013-10-11 15:44:27 -0700654 * bio->bi_iter.bi_sector points to the compressed extent on disk
Chris Masonc8b97812008-10-29 14:49:59 -0400655 * bio->bi_io_vec points to all of the inode pages
Chris Masonc8b97812008-10-29 14:49:59 -0400656 *
657 * After the compressed pages are read, we copy the bytes into the
658 * bio we were passed and then call the bio end_io calls
659 */
Christoph Hellwigcb4411d2022-04-15 16:33:27 +0200660void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
Goldwyn Rodrigues1d8fa2e22022-04-26 08:47:34 -0500661 int mirror_num)
Chris Masonc8b97812008-10-29 14:49:59 -0400662{
Jeff Mahoney0b246af2016-06-22 18:54:23 -0400663 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
Chris Masonc8b97812008-10-29 14:49:59 -0400664 struct extent_map_tree *em_tree;
665 struct compressed_bio *cb;
Anand Jain356b4a22021-05-29 17:48:34 +0800666 unsigned int compressed_len;
Qu Wenruof472c28f2021-09-27 15:21:54 +0800667 struct bio *comp_bio = NULL;
668 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
669 u64 cur_disk_byte = disk_bytenr;
670 u64 next_stripe_start;
Qu Wenruo557023e2021-07-05 10:00:56 +0800671 u64 file_offset;
Chris Masone04ca622008-11-10 11:44:58 -0500672 u64 em_len;
673 u64 em_start;
Chris Masonc8b97812008-10-29 14:49:59 -0400674 struct extent_map *em;
Johannes Weiner82e60d02022-11-03 17:34:31 -0400675 unsigned long pflags;
676 int memstall = 0;
Josef Bacikf9f15de2022-02-18 10:03:27 -0500677 blk_status_t ret;
Sweet Tea Dorminydd137dd2022-03-30 16:11:22 -0400678 int ret2;
679 int i;
Chris Masonc8b97812008-10-29 14:49:59 -0400680
Chris Masonc8b97812008-10-29 14:49:59 -0400681 em_tree = &BTRFS_I(inode)->extent_tree;
682
Qu Wenruo557023e2021-07-05 10:00:56 +0800683 file_offset = bio_first_bvec_all(bio)->bv_offset +
684 page_offset(bio_first_page_all(bio));
685
Chris Masonc8b97812008-10-29 14:49:59 -0400686 /* we need the actual starting offset of this extent in the file */
Chris Mason890871b2009-09-02 16:24:52 -0400687 read_lock(&em_tree->lock);
Qu Wenruo557023e2021-07-05 10:00:56 +0800688 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
Chris Mason890871b2009-09-02 16:24:52 -0400689 read_unlock(&em_tree->lock);
Josef Bacikf9f15de2022-02-18 10:03:27 -0500690 if (!em) {
691 ret = BLK_STS_IOERR;
692 goto out;
693 }
Chris Masonc8b97812008-10-29 14:49:59 -0400694
Qu Wenruo557023e2021-07-05 10:00:56 +0800695 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
Chris Masond20f7042008-12-08 16:58:54 -0500696 compressed_len = em->block_len;
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200697 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
Josef Bacikf9f15de2022-02-18 10:03:27 -0500698 if (!cb) {
699 ret = BLK_STS_RESOURCE;
liubo6b82ce82011-01-26 06:21:39 +0000700 goto out;
Josef Bacikf9f15de2022-02-18 10:03:27 -0500701 }
liubo6b82ce82011-01-26 06:21:39 +0000702
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200703 refcount_set(&cb->pending_ios, 1);
Josef Bacik606f82e2022-02-18 10:03:26 -0500704 cb->status = BLK_STS_OK;
Chris Masonc8b97812008-10-29 14:49:59 -0400705 cb->inode = inode;
706
Yan Zhengff5b7ee2008-11-10 07:34:43 -0500707 cb->start = em->orig_start;
Chris Masone04ca622008-11-10 11:44:58 -0500708 em_len = em->len;
709 em_start = em->start;
Chris Masond20f7042008-12-08 16:58:54 -0500710
Christoph Hellwig81381052016-11-25 09:07:50 +0100711 cb->len = bio->bi_iter.bi_size;
Chris Masonc8b97812008-10-29 14:49:59 -0400712 cb->compressed_len = compressed_len;
Goldwyn Rodrigues1d8fa2e22022-04-26 08:47:34 -0500713 cb->compress_type = em->compress_type;
Chris Masonc8b97812008-10-29 14:49:59 -0400714 cb->orig_bio = bio;
715
Goldwyn Rodrigues1d8fa2e22022-04-26 08:47:34 -0500716 free_extent_map(em);
717 em = NULL;
718
Sweet Tea Dorminydd137dd2022-03-30 16:11:22 -0400719 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
720 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
Josef Bacikf9f15de2022-02-18 10:03:27 -0500721 if (!cb->compressed_pages) {
722 ret = BLK_STS_RESOURCE;
Sweet Tea Dorminydd137dd2022-03-30 16:11:22 -0400723 goto fail;
Josef Bacikf9f15de2022-02-18 10:03:27 -0500724 }
liubo6b82ce82011-01-26 06:21:39 +0000725
Sweet Tea Dorminydd137dd2022-03-30 16:11:22 -0400726 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
727 if (ret2) {
728 ret = BLK_STS_RESOURCE;
729 goto fail;
Chris Masonc8b97812008-10-29 14:49:59 -0400730 }
Chris Masonc8b97812008-10-29 14:49:59 -0400731
Johannes Weiner82e60d02022-11-03 17:34:31 -0400732 add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags);
Chris Mason771ed682008-11-06 22:02:51 -0500733
Chris Mason771ed682008-11-06 22:02:51 -0500734 /* include any pages we added in add_ra-bio_pages */
Christoph Hellwig81381052016-11-25 09:07:50 +0100735 cb->len = bio->bi_iter.bi_size;
Chris Mason771ed682008-11-06 22:02:51 -0500736
Qu Wenruof472c28f2021-09-27 15:21:54 +0800737 while (cur_disk_byte < disk_bytenr + compressed_len) {
738 u64 offset = cur_disk_byte - disk_bytenr;
739 unsigned int index = offset >> PAGE_SHIFT;
740 unsigned int real_size;
741 unsigned int added;
742 struct page *page = cb->compressed_pages[index];
743 bool submit = false;
Chris Masonc8b97812008-10-29 14:49:59 -0400744
Qu Wenruof472c28f2021-09-27 15:21:54 +0800745 /* Allocate new bio if submitted or not yet allocated */
746 if (!comp_bio) {
747 comp_bio = alloc_compressed_bio(cb, cur_disk_byte,
748 REQ_OP_READ, end_compressed_bio_read,
749 &next_stripe_start);
750 if (IS_ERR(comp_bio)) {
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200751 cb->status = errno_to_blk_status(PTR_ERR(comp_bio));
752 break;
Qu Wenruof472c28f2021-09-27 15:21:54 +0800753 }
754 }
Qu Wenruobe6a1362021-02-04 15:03:23 +0800755 /*
Qu Wenruof472c28f2021-09-27 15:21:54 +0800756 * We should never reach next_stripe_start start as we will
757 * submit comp_bio when reach the boundary immediately.
Qu Wenruobe6a1362021-02-04 15:03:23 +0800758 */
Qu Wenruof472c28f2021-09-27 15:21:54 +0800759 ASSERT(cur_disk_byte != next_stripe_start);
760 /*
761 * We have various limit on the real read size:
762 * - stripe boundary
763 * - page boundary
764 * - compressed length boundary
765 */
766 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte);
767 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
768 real_size = min_t(u64, real_size, compressed_len - offset);
769 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
Qu Wenruobe6a1362021-02-04 15:03:23 +0800770
Qu Wenruof472c28f2021-09-27 15:21:54 +0800771 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
772 /*
773 * Maximum compressed extent is smaller than bio size limit,
774 * thus bio_add_page() should always success.
775 */
776 ASSERT(added == real_size);
777 cur_disk_byte += added;
Chris Masond20f7042008-12-08 16:58:54 -0500778
Qu Wenruof472c28f2021-09-27 15:21:54 +0800779 /* Reached stripe boundary, need to submit */
780 if (cur_disk_byte == next_stripe_start)
781 submit = true;
Chris Masonc8b97812008-10-29 14:49:59 -0400782
Qu Wenruof472c28f2021-09-27 15:21:54 +0800783 /* Has finished the range, need to submit */
784 if (cur_disk_byte == disk_bytenr + compressed_len)
785 submit = true;
786
787 if (submit) {
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200788 /* Save the original iter for read repair */
789 if (bio_op(comp_bio) == REQ_OP_READ)
790 btrfs_bio(comp_bio)->iter = comp_bio->bi_iter;
Johannes Thumshirn10fe6ca2019-05-22 10:19:02 +0200791
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200792 /*
793 * Save the initial offset of this chunk, as there
794 * is no direct correlation between compressed pages and
795 * the original file offset. The field is only used for
796 * priting error messages.
797 */
798 btrfs_bio(comp_bio)->file_offset = file_offset;
Johannes Thumshirn10fe6ca2019-05-22 10:19:02 +0200799
Christoph Hellwig81bd9322022-07-07 07:33:30 +0200800 ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL);
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200801 if (ret) {
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200802 btrfs_bio_end_io(btrfs_bio(comp_bio), ret);
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200803 break;
804 }
Chris Masond20f7042008-12-08 16:58:54 -0500805
Christoph Hellwigfed8a722022-05-26 09:36:38 +0200806 ASSERT(comp_bio->bi_iter.bi_size);
Christoph Hellwig1a722d82022-06-17 12:04:07 +0200807 btrfs_submit_bio(fs_info, comp_bio, mirror_num);
Qu Wenruof472c28f2021-09-27 15:21:54 +0800808 comp_bio = NULL;
Chris Masonc8b97812008-10-29 14:49:59 -0400809 }
Chris Masonc8b97812008-10-29 14:49:59 -0400810 }
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200811
Johannes Weiner82e60d02022-11-03 17:34:31 -0400812 if (memstall)
Christoph Hellwig4088a472022-09-15 10:41:58 +0100813 psi_memstall_leave(&pflags);
814
Christoph Hellwig524bcd12022-07-07 07:33:27 +0200815 if (refcount_dec_and_test(&cb->pending_ios))
816 finish_compressed_bio_read(cb);
Christoph Hellwigcb4411d2022-04-15 16:33:27 +0200817 return;
liubo6b82ce82011-01-26 06:21:39 +0000818
Sweet Tea Dorminydd137dd2022-03-30 16:11:22 -0400819fail:
820 if (cb->compressed_pages) {
821 for (i = 0; i < cb->nr_pages; i++) {
822 if (cb->compressed_pages[i])
823 __free_page(cb->compressed_pages[i]);
824 }
Josef Bacik15e3004a2012-10-05 13:39:50 -0400825 }
liubo6b82ce82011-01-26 06:21:39 +0000826
827 kfree(cb->compressed_pages);
liubo6b82ce82011-01-26 06:21:39 +0000828 kfree(cb);
829out:
830 free_extent_map(em);
Christoph Hellwig917f32a2022-08-06 10:03:26 +0200831 btrfs_bio_end_io(btrfs_bio(bio), ret);
Christoph Hellwigcb4411d2022-04-15 16:33:27 +0200832 return;
Chris Masonc8b97812008-10-29 14:49:59 -0400833}
Li Zefan261507a02010-12-17 14:21:50 +0800834
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300835/*
836 * Heuristic uses systematic sampling to collect data from the input data
837 * range, the logic can be tuned by the following constants:
838 *
839 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
840 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
841 */
842#define SAMPLING_READ_SIZE (16)
843#define SAMPLING_INTERVAL (256)
844
845/*
846 * For statistical analysis of the input data we consider bytes that form a
847 * Galois Field of 256 objects. Each object has an attribute count, ie. how
848 * many times the object appeared in the sample.
849 */
850#define BUCKET_SIZE (256)
851
852/*
853 * The size of the sample is based on a statistical sampling rule of thumb.
854 * The common way is to perform sampling tests as long as the number of
855 * elements in each cell is at least 5.
856 *
857 * Instead of 5, we choose 32 to obtain more accurate results.
858 * If the data contain the maximum number of symbols, which is 256, we obtain a
859 * sample size bound by 8192.
860 *
861 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
862 * from up to 512 locations.
863 */
864#define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
865 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
866
867struct bucket_item {
868 u32 count;
869};
Timofey Titovets4e439a02017-09-28 17:33:36 +0300870
871struct heuristic_ws {
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300872 /* Partial copy of input data */
873 u8 *sample;
Timofey Titovetsa440d482017-09-28 17:33:38 +0300874 u32 sample_size;
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300875 /* Buckets store counters for each byte value */
876 struct bucket_item *bucket;
Timofey Titovets440c8402017-12-04 00:30:33 +0300877 /* Sorting buffer */
878 struct bucket_item *bucket_b;
Timofey Titovets4e439a02017-09-28 17:33:36 +0300879 struct list_head list;
880};
881
Dennis Zhou92ee55302019-02-04 15:20:03 -0500882static struct workspace_manager heuristic_wsm;
883
Timofey Titovets4e439a02017-09-28 17:33:36 +0300884static void free_heuristic_ws(struct list_head *ws)
885{
886 struct heuristic_ws *workspace;
887
888 workspace = list_entry(ws, struct heuristic_ws, list);
889
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300890 kvfree(workspace->sample);
891 kfree(workspace->bucket);
Timofey Titovets440c8402017-12-04 00:30:33 +0300892 kfree(workspace->bucket_b);
Timofey Titovets4e439a02017-09-28 17:33:36 +0300893 kfree(workspace);
894}
895
Dennis Zhou7bf49942019-02-04 15:20:04 -0500896static struct list_head *alloc_heuristic_ws(unsigned int level)
Timofey Titovets4e439a02017-09-28 17:33:36 +0300897{
898 struct heuristic_ws *ws;
899
900 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
901 if (!ws)
902 return ERR_PTR(-ENOMEM);
903
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300904 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
905 if (!ws->sample)
906 goto fail;
Timofey Titovets4e439a02017-09-28 17:33:36 +0300907
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300908 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
909 if (!ws->bucket)
910 goto fail;
911
Timofey Titovets440c8402017-12-04 00:30:33 +0300912 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
913 if (!ws->bucket_b)
914 goto fail;
915
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300916 INIT_LIST_HEAD(&ws->list);
Timofey Titovets4e439a02017-09-28 17:33:36 +0300917 return &ws->list;
Timofey Titovets17b5a6c2017-09-28 17:33:37 +0300918fail:
919 free_heuristic_ws(&ws->list);
920 return ERR_PTR(-ENOMEM);
Timofey Titovets4e439a02017-09-28 17:33:36 +0300921}
922
Dennis Zhouca4ac362019-02-04 15:19:59 -0500923const struct btrfs_compress_op btrfs_heuristic_compress = {
David Sterbabe9510452019-10-02 00:53:31 +0200924 .workspace_manager = &heuristic_wsm,
Dennis Zhouca4ac362019-02-04 15:19:59 -0500925};
926
David Sterbae8c9f182015-01-02 18:23:10 +0100927static const struct btrfs_compress_op * const btrfs_compress_op[] = {
Dennis Zhouca4ac362019-02-04 15:19:59 -0500928 /* The heuristic is represented as compression type 0 */
929 &btrfs_heuristic_compress,
Li Zefan261507a02010-12-17 14:21:50 +0800930 &btrfs_zlib_compress,
Li Zefana6fa6fa2010-10-25 15:12:26 +0800931 &btrfs_lzo_compress,
Nick Terrell5c1aab12017-08-09 19:39:02 -0700932 &btrfs_zstd_compress,
Li Zefan261507a02010-12-17 14:21:50 +0800933};
934
David Sterbac778df12019-10-04 02:47:39 +0200935static struct list_head *alloc_workspace(int type, unsigned int level)
936{
937 switch (type) {
938 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
939 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
940 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
941 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
942 default:
943 /*
944 * This can't happen, the type is validated several times
945 * before we get here.
946 */
947 BUG();
948 }
949}
950
David Sterba1e002352019-10-04 02:57:22 +0200951static void free_workspace(int type, struct list_head *ws)
952{
953 switch (type) {
954 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
955 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
956 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
957 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
958 default:
959 /*
960 * This can't happen, the type is validated several times
961 * before we get here.
962 */
963 BUG();
964 }
965}
966
David Sterbad5517032019-10-02 01:08:03 +0200967static void btrfs_init_workspace_manager(int type)
Li Zefan261507a02010-12-17 14:21:50 +0800968{
David Sterba0cf25212019-10-04 03:09:55 +0200969 struct workspace_manager *wsm;
Timofey Titovets4e439a02017-09-28 17:33:36 +0300970 struct list_head *workspace;
Li Zefan261507a02010-12-17 14:21:50 +0800971
David Sterba0cf25212019-10-04 03:09:55 +0200972 wsm = btrfs_compress_op[type]->workspace_manager;
Dennis Zhou92ee55302019-02-04 15:20:03 -0500973 INIT_LIST_HEAD(&wsm->idle_ws);
974 spin_lock_init(&wsm->ws_lock);
975 atomic_set(&wsm->total_ws, 0);
976 init_waitqueue_head(&wsm->ws_wait);
David Sterbaf77dd0d2016-04-27 02:55:15 +0200977
Dennis Zhou1666eda2019-02-04 15:20:01 -0500978 /*
979 * Preallocate one workspace for each compression type so we can
980 * guarantee forward progress in the worst case
981 */
David Sterbac778df12019-10-04 02:47:39 +0200982 workspace = alloc_workspace(type, 0);
Dennis Zhou1666eda2019-02-04 15:20:01 -0500983 if (IS_ERR(workspace)) {
984 pr_warn(
985 "BTRFS: cannot preallocate compression workspace, will try later\n");
986 } else {
Dennis Zhou92ee55302019-02-04 15:20:03 -0500987 atomic_set(&wsm->total_ws, 1);
988 wsm->free_ws = 1;
989 list_add(workspace, &wsm->idle_ws);
Dennis Zhou1666eda2019-02-04 15:20:01 -0500990 }
991}
992
David Sterba25103072019-10-02 01:08:03 +0200993static void btrfs_cleanup_workspace_manager(int type)
Dennis Zhou1666eda2019-02-04 15:20:01 -0500994{
David Sterba2dba7142019-10-04 01:40:58 +0200995 struct workspace_manager *wsman;
Dennis Zhou1666eda2019-02-04 15:20:01 -0500996 struct list_head *ws;
997
David Sterba2dba7142019-10-04 01:40:58 +0200998 wsman = btrfs_compress_op[type]->workspace_manager;
Dennis Zhou1666eda2019-02-04 15:20:01 -0500999 while (!list_empty(&wsman->idle_ws)) {
1000 ws = wsman->idle_ws.next;
1001 list_del(ws);
David Sterba1e002352019-10-04 02:57:22 +02001002 free_workspace(type, ws);
Dennis Zhou1666eda2019-02-04 15:20:01 -05001003 atomic_dec(&wsman->total_ws);
Li Zefan261507a02010-12-17 14:21:50 +08001004 }
Li Zefan261507a02010-12-17 14:21:50 +08001005}
1006
1007/*
David Sterbae721e492016-04-27 02:41:17 +02001008 * This finds an available workspace or allocates a new one.
1009 * If it's not possible to allocate a new one, waits until there's one.
1010 * Preallocation makes a forward progress guarantees and we do not return
1011 * errors.
Li Zefan261507a02010-12-17 14:21:50 +08001012 */
David Sterba5907a9b2019-10-04 02:50:28 +02001013struct list_head *btrfs_get_workspace(int type, unsigned int level)
Li Zefan261507a02010-12-17 14:21:50 +08001014{
David Sterba5907a9b2019-10-04 02:50:28 +02001015 struct workspace_manager *wsm;
Li Zefan261507a02010-12-17 14:21:50 +08001016 struct list_head *workspace;
1017 int cpus = num_online_cpus();
David Sterbafe308532017-05-31 17:14:56 +02001018 unsigned nofs_flag;
Timofey Titovets4e439a02017-09-28 17:33:36 +03001019 struct list_head *idle_ws;
1020 spinlock_t *ws_lock;
1021 atomic_t *total_ws;
1022 wait_queue_head_t *ws_wait;
1023 int *free_ws;
Li Zefan261507a02010-12-17 14:21:50 +08001024
David Sterba5907a9b2019-10-04 02:50:28 +02001025 wsm = btrfs_compress_op[type]->workspace_manager;
Dennis Zhou92ee55302019-02-04 15:20:03 -05001026 idle_ws = &wsm->idle_ws;
1027 ws_lock = &wsm->ws_lock;
1028 total_ws = &wsm->total_ws;
1029 ws_wait = &wsm->ws_wait;
1030 free_ws = &wsm->free_ws;
Timofey Titovets4e439a02017-09-28 17:33:36 +03001031
Li Zefan261507a02010-12-17 14:21:50 +08001032again:
Byongho Leed9187642015-10-14 14:05:24 +09001033 spin_lock(ws_lock);
1034 if (!list_empty(idle_ws)) {
1035 workspace = idle_ws->next;
Li Zefan261507a02010-12-17 14:21:50 +08001036 list_del(workspace);
David Sterba6ac10a62016-04-27 02:15:15 +02001037 (*free_ws)--;
Byongho Leed9187642015-10-14 14:05:24 +09001038 spin_unlock(ws_lock);
Li Zefan261507a02010-12-17 14:21:50 +08001039 return workspace;
1040
1041 }
David Sterba6ac10a62016-04-27 02:15:15 +02001042 if (atomic_read(total_ws) > cpus) {
Li Zefan261507a02010-12-17 14:21:50 +08001043 DEFINE_WAIT(wait);
1044
Byongho Leed9187642015-10-14 14:05:24 +09001045 spin_unlock(ws_lock);
1046 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
David Sterba6ac10a62016-04-27 02:15:15 +02001047 if (atomic_read(total_ws) > cpus && !*free_ws)
Li Zefan261507a02010-12-17 14:21:50 +08001048 schedule();
Byongho Leed9187642015-10-14 14:05:24 +09001049 finish_wait(ws_wait, &wait);
Li Zefan261507a02010-12-17 14:21:50 +08001050 goto again;
1051 }
David Sterba6ac10a62016-04-27 02:15:15 +02001052 atomic_inc(total_ws);
Byongho Leed9187642015-10-14 14:05:24 +09001053 spin_unlock(ws_lock);
Li Zefan261507a02010-12-17 14:21:50 +08001054
David Sterbafe308532017-05-31 17:14:56 +02001055 /*
1056 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1057 * to turn it off here because we might get called from the restricted
1058 * context of btrfs_compress_bio/btrfs_compress_pages
1059 */
1060 nofs_flag = memalloc_nofs_save();
David Sterbac778df12019-10-04 02:47:39 +02001061 workspace = alloc_workspace(type, level);
David Sterbafe308532017-05-31 17:14:56 +02001062 memalloc_nofs_restore(nofs_flag);
1063
Li Zefan261507a02010-12-17 14:21:50 +08001064 if (IS_ERR(workspace)) {
David Sterba6ac10a62016-04-27 02:15:15 +02001065 atomic_dec(total_ws);
Byongho Leed9187642015-10-14 14:05:24 +09001066 wake_up(ws_wait);
David Sterbae721e492016-04-27 02:41:17 +02001067
1068 /*
1069 * Do not return the error but go back to waiting. There's a
1070 * workspace preallocated for each type and the compression
1071 * time is bounded so we get to a workspace eventually. This
1072 * makes our caller's life easier.
David Sterba523567162016-04-27 03:07:39 +02001073 *
1074 * To prevent silent and low-probability deadlocks (when the
1075 * initial preallocation fails), check if there are any
1076 * workspaces at all.
David Sterbae721e492016-04-27 02:41:17 +02001077 */
David Sterba523567162016-04-27 03:07:39 +02001078 if (atomic_read(total_ws) == 0) {
1079 static DEFINE_RATELIMIT_STATE(_rs,
1080 /* once per minute */ 60 * HZ,
1081 /* no burst */ 1);
1082
1083 if (__ratelimit(&_rs)) {
Jeff Mahoneyab8d0fc2016-09-20 10:05:02 -04001084 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
David Sterba523567162016-04-27 03:07:39 +02001085 }
1086 }
David Sterbae721e492016-04-27 02:41:17 +02001087 goto again;
Li Zefan261507a02010-12-17 14:21:50 +08001088 }
1089 return workspace;
1090}
1091
Dennis Zhou7bf49942019-02-04 15:20:04 -05001092static struct list_head *get_workspace(int type, int level)
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001093{
David Sterba6a0d1272019-10-04 02:36:16 +02001094 switch (type) {
David Sterba5907a9b2019-10-04 02:50:28 +02001095 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
David Sterba6a0d1272019-10-04 02:36:16 +02001096 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
David Sterba5907a9b2019-10-04 02:50:28 +02001097 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
David Sterba6a0d1272019-10-04 02:36:16 +02001098 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1099 default:
1100 /*
1101 * This can't happen, the type is validated several times
1102 * before we get here.
1103 */
1104 BUG();
1105 }
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001106}
1107
Li Zefan261507a02010-12-17 14:21:50 +08001108/*
1109 * put a workspace struct back on the list or free it if we have enough
1110 * idle ones sitting around
1111 */
David Sterbaa3bbd2a2019-10-04 02:50:28 +02001112void btrfs_put_workspace(int type, struct list_head *ws)
Li Zefan261507a02010-12-17 14:21:50 +08001113{
David Sterbaa3bbd2a2019-10-04 02:50:28 +02001114 struct workspace_manager *wsm;
Timofey Titovets4e439a02017-09-28 17:33:36 +03001115 struct list_head *idle_ws;
1116 spinlock_t *ws_lock;
1117 atomic_t *total_ws;
1118 wait_queue_head_t *ws_wait;
1119 int *free_ws;
1120
David Sterbaa3bbd2a2019-10-04 02:50:28 +02001121 wsm = btrfs_compress_op[type]->workspace_manager;
Dennis Zhou92ee55302019-02-04 15:20:03 -05001122 idle_ws = &wsm->idle_ws;
1123 ws_lock = &wsm->ws_lock;
1124 total_ws = &wsm->total_ws;
1125 ws_wait = &wsm->ws_wait;
1126 free_ws = &wsm->free_ws;
Li Zefan261507a02010-12-17 14:21:50 +08001127
Byongho Leed9187642015-10-14 14:05:24 +09001128 spin_lock(ws_lock);
Nick Terrell26b28dc2017-06-29 10:57:26 -07001129 if (*free_ws <= num_online_cpus()) {
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001130 list_add(ws, idle_ws);
David Sterba6ac10a62016-04-27 02:15:15 +02001131 (*free_ws)++;
Byongho Leed9187642015-10-14 14:05:24 +09001132 spin_unlock(ws_lock);
Li Zefan261507a02010-12-17 14:21:50 +08001133 goto wake;
1134 }
Byongho Leed9187642015-10-14 14:05:24 +09001135 spin_unlock(ws_lock);
Li Zefan261507a02010-12-17 14:21:50 +08001136
David Sterba1e002352019-10-04 02:57:22 +02001137 free_workspace(type, ws);
David Sterba6ac10a62016-04-27 02:15:15 +02001138 atomic_dec(total_ws);
Li Zefan261507a02010-12-17 14:21:50 +08001139wake:
David Sterba093258e62018-02-26 16:15:17 +01001140 cond_wake_up(ws_wait);
Li Zefan261507a02010-12-17 14:21:50 +08001141}
1142
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001143static void put_workspace(int type, struct list_head *ws)
1144{
David Sterbabd3a5282019-10-04 02:42:03 +02001145 switch (type) {
David Sterbaa3bbd2a2019-10-04 02:50:28 +02001146 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1147 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1148 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
David Sterbabd3a5282019-10-04 02:42:03 +02001149 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1150 default:
1151 /*
1152 * This can't happen, the type is validated several times
1153 * before we get here.
1154 */
1155 BUG();
1156 }
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001157}
1158
Li Zefan261507a02010-12-17 14:21:50 +08001159/*
Anand Jainadbab642020-05-11 22:37:51 -07001160 * Adjust @level according to the limits of the compression algorithm or
1161 * fallback to default
1162 */
1163static unsigned int btrfs_compress_set_level(int type, unsigned level)
1164{
1165 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1166
1167 if (level == 0)
1168 level = ops->default_level;
1169 else
1170 level = min(level, ops->max_level);
1171
1172 return level;
1173}
1174
1175/*
David Sterba38c31462017-02-14 19:04:07 +01001176 * Given an address space and start and length, compress the bytes into @pages
1177 * that are allocated on demand.
Li Zefan261507a02010-12-17 14:21:50 +08001178 *
David Sterbaf51d2b52017-09-15 17:36:57 +02001179 * @type_level is encoded algorithm and level, where level 0 means whatever
1180 * default the algorithm chooses and is opaque here;
1181 * - compression algo are 0-3
1182 * - the level are bits 4-7
1183 *
David Sterba4d3a8002017-02-14 19:04:07 +01001184 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1185 * and returns number of actually allocated pages
Li Zefan261507a02010-12-17 14:21:50 +08001186 *
David Sterba38c31462017-02-14 19:04:07 +01001187 * @total_in is used to return the number of bytes actually read. It
1188 * may be smaller than the input length if we had to exit early because we
Li Zefan261507a02010-12-17 14:21:50 +08001189 * ran out of room in the pages array or because we cross the
1190 * max_out threshold.
1191 *
David Sterba38c31462017-02-14 19:04:07 +01001192 * @total_out is an in/out parameter, must be set to the input length and will
1193 * be also used to return the total number of compressed bytes
Li Zefan261507a02010-12-17 14:21:50 +08001194 */
David Sterbaf51d2b52017-09-15 17:36:57 +02001195int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
David Sterba38c31462017-02-14 19:04:07 +01001196 u64 start, struct page **pages,
Li Zefan261507a02010-12-17 14:21:50 +08001197 unsigned long *out_pages,
1198 unsigned long *total_in,
David Sterbae5d74902017-02-14 19:45:05 +01001199 unsigned long *total_out)
Li Zefan261507a02010-12-17 14:21:50 +08001200{
Dennis Zhou19727082019-02-04 15:19:57 -05001201 int type = btrfs_compress_type(type_level);
Dennis Zhou7bf49942019-02-04 15:20:04 -05001202 int level = btrfs_compress_level(type_level);
Li Zefan261507a02010-12-17 14:21:50 +08001203 struct list_head *workspace;
1204 int ret;
1205
David Sterbab0c1fe12019-08-09 16:49:06 +02001206 level = btrfs_compress_set_level(type, level);
Dennis Zhou7bf49942019-02-04 15:20:04 -05001207 workspace = get_workspace(type, level);
David Sterba1e4eb742019-10-02 00:06:15 +02001208 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1209 out_pages, total_in, total_out);
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001210 put_workspace(type, workspace);
Li Zefan261507a02010-12-17 14:21:50 +08001211 return ret;
1212}
1213
Anand Jain8140dc32017-05-26 15:44:58 +08001214static int btrfs_decompress_bio(struct compressed_bio *cb)
Li Zefan261507a02010-12-17 14:21:50 +08001215{
1216 struct list_head *workspace;
1217 int ret;
Anand Jain8140dc32017-05-26 15:44:58 +08001218 int type = cb->compress_type;
Li Zefan261507a02010-12-17 14:21:50 +08001219
Dennis Zhou7bf49942019-02-04 15:20:04 -05001220 workspace = get_workspace(type, 0);
Su Yue4a9e8032021-12-27 18:18:39 +08001221 ret = compression_decompress_bio(workspace, cb);
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001222 put_workspace(type, workspace);
Anand Jaine1ddce72017-05-26 15:44:59 +08001223
Li Zefan261507a02010-12-17 14:21:50 +08001224 return ret;
1225}
1226
1227/*
1228 * a less complex decompression routine. Our compressed data fits in a
1229 * single page, and we want to read a single page out of it.
1230 * start_byte tells us the offset into the compressed data we're interested in
1231 */
1232int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1233 unsigned long start_byte, size_t srclen, size_t destlen)
1234{
1235 struct list_head *workspace;
1236 int ret;
1237
Dennis Zhou7bf49942019-02-04 15:20:04 -05001238 workspace = get_workspace(type, 0);
David Sterba1e4eb742019-10-02 00:06:15 +02001239 ret = compression_decompress(type, workspace, data_in, dest_page,
1240 start_byte, srclen, destlen);
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001241 put_workspace(type, workspace);
Dennis Zhou7bf49942019-02-04 15:20:04 -05001242
Li Zefan261507a02010-12-17 14:21:50 +08001243 return ret;
1244}
1245
Dennis Zhou1666eda2019-02-04 15:20:01 -05001246void __init btrfs_init_compress(void)
1247{
David Sterbad5517032019-10-02 01:08:03 +02001248 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1249 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1250 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1251 zstd_init_workspace_manager();
Dennis Zhou1666eda2019-02-04 15:20:01 -05001252}
1253
David Sterbae67c7182018-02-19 17:24:18 +01001254void __cold btrfs_exit_compress(void)
Li Zefan261507a02010-12-17 14:21:50 +08001255{
David Sterba25103072019-10-02 01:08:03 +02001256 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1257 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1258 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1259 zstd_cleanup_workspace_manager();
Li Zefan261507a02010-12-17 14:21:50 +08001260}
Li Zefan3a39c182010-11-08 15:22:19 +08001261
1262/*
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001263 * Copy decompressed data from working buffer to pages.
Li Zefan3a39c182010-11-08 15:22:19 +08001264 *
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001265 * @buf: The decompressed data buffer
1266 * @buf_len: The decompressed data length
1267 * @decompressed: Number of bytes that are already decompressed inside the
1268 * compressed extent
1269 * @cb: The compressed extent descriptor
1270 * @orig_bio: The original bio that the caller wants to read for
Li Zefan3a39c182010-11-08 15:22:19 +08001271 *
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001272 * An easier to understand graph is like below:
1273 *
1274 * |<- orig_bio ->| |<- orig_bio->|
1275 * |<------- full decompressed extent ----->|
1276 * |<----------- @cb range ---->|
1277 * | |<-- @buf_len -->|
1278 * |<--- @decompressed --->|
1279 *
1280 * Note that, @cb can be a subpage of the full decompressed extent, but
1281 * @cb->start always has the same as the orig_file_offset value of the full
1282 * decompressed extent.
1283 *
1284 * When reading compressed extent, we have to read the full compressed extent,
1285 * while @orig_bio may only want part of the range.
1286 * Thus this function will ensure only data covered by @orig_bio will be copied
1287 * to.
1288 *
1289 * Return 0 if we have copied all needed contents for @orig_bio.
1290 * Return >0 if we need continue decompress.
Li Zefan3a39c182010-11-08 15:22:19 +08001291 */
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001292int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1293 struct compressed_bio *cb, u32 decompressed)
Li Zefan3a39c182010-11-08 15:22:19 +08001294{
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001295 struct bio *orig_bio = cb->orig_bio;
1296 /* Offset inside the full decompressed extent */
1297 u32 cur_offset;
Li Zefan3a39c182010-11-08 15:22:19 +08001298
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001299 cur_offset = decompressed;
1300 /* The main loop to do the copy */
1301 while (cur_offset < decompressed + buf_len) {
1302 struct bio_vec bvec;
1303 size_t copy_len;
1304 u32 copy_start;
1305 /* Offset inside the full decompressed extent */
1306 u32 bvec_offset;
Li Zefan3a39c182010-11-08 15:22:19 +08001307
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001308 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1309 /*
1310 * cb->start may underflow, but subtracting that value can still
1311 * give us correct offset inside the full decompressed extent.
1312 */
1313 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
Li Zefan3a39c182010-11-08 15:22:19 +08001314
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001315 /* Haven't reached the bvec range, exit */
1316 if (decompressed + buf_len <= bvec_offset)
1317 return 1;
Li Zefan3a39c182010-11-08 15:22:19 +08001318
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001319 copy_start = max(cur_offset, bvec_offset);
1320 copy_len = min(bvec_offset + bvec.bv_len,
1321 decompressed + buf_len) - copy_start;
1322 ASSERT(copy_len);
Li Zefan3a39c182010-11-08 15:22:19 +08001323
Christoph Hellwig974b1ad2016-11-25 09:07:46 +01001324 /*
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001325 * Extra range check to ensure we didn't go beyond
1326 * @buf + @buf_len.
Christoph Hellwig974b1ad2016-11-25 09:07:46 +01001327 */
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001328 ASSERT(copy_start - decompressed < buf_len);
1329 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1330 buf + copy_start - decompressed, copy_len);
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001331 cur_offset += copy_len;
Li Zefan3a39c182010-11-08 15:22:19 +08001332
Qu Wenruo1c3dc172021-07-05 10:00:58 +08001333 bio_advance(orig_bio, copy_len);
1334 /* Finished the bio */
1335 if (!orig_bio->bi_iter.bi_size)
1336 return 0;
Li Zefan3a39c182010-11-08 15:22:19 +08001337 }
Li Zefan3a39c182010-11-08 15:22:19 +08001338 return 1;
1339}
Timofey Titovetsc2fcdcd2017-07-17 16:52:58 +03001340
Timofey Titovets19562432017-10-08 16:11:59 +03001341/*
1342 * Shannon Entropy calculation
1343 *
Andrea Gelmini52042d82018-11-28 12:05:13 +01001344 * Pure byte distribution analysis fails to determine compressibility of data.
Timofey Titovets19562432017-10-08 16:11:59 +03001345 * Try calculating entropy to estimate the average minimum number of bits
1346 * needed to encode the sampled data.
1347 *
1348 * For convenience, return the percentage of needed bits, instead of amount of
1349 * bits directly.
1350 *
1351 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1352 * and can be compressible with high probability
1353 *
1354 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1355 *
1356 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1357 */
1358#define ENTROPY_LVL_ACEPTABLE (65)
1359#define ENTROPY_LVL_HIGH (80)
1360
1361/*
1362 * For increasead precision in shannon_entropy calculation,
1363 * let's do pow(n, M) to save more digits after comma:
1364 *
1365 * - maximum int bit length is 64
1366 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1367 * - 13 * 4 = 52 < 64 -> M = 4
1368 *
1369 * So use pow(n, 4).
1370 */
1371static inline u32 ilog2_w(u64 n)
1372{
1373 return ilog2(n * n * n * n);
1374}
1375
1376static u32 shannon_entropy(struct heuristic_ws *ws)
1377{
1378 const u32 entropy_max = 8 * ilog2_w(2);
1379 u32 entropy_sum = 0;
1380 u32 p, p_base, sz_base;
1381 u32 i;
1382
1383 sz_base = ilog2_w(ws->sample_size);
1384 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1385 p = ws->bucket[i].count;
1386 p_base = ilog2_w(p);
1387 entropy_sum += p * (sz_base - p_base);
1388 }
1389
1390 entropy_sum /= ws->sample_size;
1391 return entropy_sum * 100 / entropy_max;
1392}
1393
Timofey Titovets440c8402017-12-04 00:30:33 +03001394#define RADIX_BASE 4U
1395#define COUNTERS_SIZE (1U << RADIX_BASE)
Timofey Titovets858177d2017-09-28 17:33:41 +03001396
Timofey Titovets440c8402017-12-04 00:30:33 +03001397static u8 get4bits(u64 num, int shift) {
1398 u8 low4bits;
1399
1400 num >>= shift;
1401 /* Reverse order */
1402 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1403 return low4bits;
1404}
1405
Timofey Titovets440c8402017-12-04 00:30:33 +03001406/*
1407 * Use 4 bits as radix base
Andrea Gelmini52042d82018-11-28 12:05:13 +01001408 * Use 16 u32 counters for calculating new position in buf array
Timofey Titovets440c8402017-12-04 00:30:33 +03001409 *
1410 * @array - array that will be sorted
1411 * @array_buf - buffer array to store sorting results
1412 * must be equal in size to @array
1413 * @num - array size
Timofey Titovets440c8402017-12-04 00:30:33 +03001414 */
David Sterba23ae8c62017-12-12 20:35:02 +01001415static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
David Sterba36243c92017-12-12 20:35:02 +01001416 int num)
Timofey Titovets440c8402017-12-04 00:30:33 +03001417{
1418 u64 max_num;
1419 u64 buf_num;
1420 u32 counters[COUNTERS_SIZE];
1421 u32 new_addr;
1422 u32 addr;
1423 int bitlen;
1424 int shift;
1425 int i;
1426
1427 /*
1428 * Try avoid useless loop iterations for small numbers stored in big
1429 * counters. Example: 48 33 4 ... in 64bit array
1430 */
David Sterba23ae8c62017-12-12 20:35:02 +01001431 max_num = array[0].count;
Timofey Titovets440c8402017-12-04 00:30:33 +03001432 for (i = 1; i < num; i++) {
David Sterba23ae8c62017-12-12 20:35:02 +01001433 buf_num = array[i].count;
Timofey Titovets440c8402017-12-04 00:30:33 +03001434 if (buf_num > max_num)
1435 max_num = buf_num;
1436 }
1437
1438 buf_num = ilog2(max_num);
1439 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1440
1441 shift = 0;
1442 while (shift < bitlen) {
1443 memset(counters, 0, sizeof(counters));
1444
1445 for (i = 0; i < num; i++) {
David Sterba23ae8c62017-12-12 20:35:02 +01001446 buf_num = array[i].count;
Timofey Titovets440c8402017-12-04 00:30:33 +03001447 addr = get4bits(buf_num, shift);
1448 counters[addr]++;
1449 }
1450
1451 for (i = 1; i < COUNTERS_SIZE; i++)
1452 counters[i] += counters[i - 1];
1453
1454 for (i = num - 1; i >= 0; i--) {
David Sterba23ae8c62017-12-12 20:35:02 +01001455 buf_num = array[i].count;
Timofey Titovets440c8402017-12-04 00:30:33 +03001456 addr = get4bits(buf_num, shift);
1457 counters[addr]--;
1458 new_addr = counters[addr];
David Sterba7add17b2017-12-12 20:35:02 +01001459 array_buf[new_addr] = array[i];
Timofey Titovets440c8402017-12-04 00:30:33 +03001460 }
1461
1462 shift += RADIX_BASE;
1463
1464 /*
1465 * Normal radix expects to move data from a temporary array, to
1466 * the main one. But that requires some CPU time. Avoid that
1467 * by doing another sort iteration to original array instead of
1468 * memcpy()
1469 */
1470 memset(counters, 0, sizeof(counters));
1471
1472 for (i = 0; i < num; i ++) {
David Sterba23ae8c62017-12-12 20:35:02 +01001473 buf_num = array_buf[i].count;
Timofey Titovets440c8402017-12-04 00:30:33 +03001474 addr = get4bits(buf_num, shift);
1475 counters[addr]++;
1476 }
1477
1478 for (i = 1; i < COUNTERS_SIZE; i++)
1479 counters[i] += counters[i - 1];
1480
1481 for (i = num - 1; i >= 0; i--) {
David Sterba23ae8c62017-12-12 20:35:02 +01001482 buf_num = array_buf[i].count;
Timofey Titovets440c8402017-12-04 00:30:33 +03001483 addr = get4bits(buf_num, shift);
1484 counters[addr]--;
1485 new_addr = counters[addr];
David Sterba7add17b2017-12-12 20:35:02 +01001486 array[new_addr] = array_buf[i];
Timofey Titovets440c8402017-12-04 00:30:33 +03001487 }
1488
1489 shift += RADIX_BASE;
1490 }
Timofey Titovets858177d2017-09-28 17:33:41 +03001491}
1492
1493/*
1494 * Size of the core byte set - how many bytes cover 90% of the sample
1495 *
1496 * There are several types of structured binary data that use nearly all byte
1497 * values. The distribution can be uniform and counts in all buckets will be
1498 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1499 *
1500 * Other possibility is normal (Gaussian) distribution, where the data could
1501 * be potentially compressible, but we have to take a few more steps to decide
1502 * how much.
1503 *
1504 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1505 * compression algo can easy fix that
1506 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1507 * probability is not compressible
1508 */
1509#define BYTE_CORE_SET_LOW (64)
1510#define BYTE_CORE_SET_HIGH (200)
1511
1512static int byte_core_set_size(struct heuristic_ws *ws)
1513{
1514 u32 i;
1515 u32 coreset_sum = 0;
1516 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1517 struct bucket_item *bucket = ws->bucket;
1518
1519 /* Sort in reverse order */
David Sterba36243c92017-12-12 20:35:02 +01001520 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
Timofey Titovets858177d2017-09-28 17:33:41 +03001521
1522 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1523 coreset_sum += bucket[i].count;
1524
1525 if (coreset_sum > core_set_threshold)
1526 return i;
1527
1528 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1529 coreset_sum += bucket[i].count;
1530 if (coreset_sum > core_set_threshold)
1531 break;
1532 }
1533
1534 return i;
1535}
1536
Timofey Titovetsa288e922017-09-28 17:33:40 +03001537/*
1538 * Count byte values in buckets.
1539 * This heuristic can detect textual data (configs, xml, json, html, etc).
1540 * Because in most text-like data byte set is restricted to limited number of
1541 * possible characters, and that restriction in most cases makes data easy to
1542 * compress.
1543 *
1544 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1545 * less - compressible
1546 * more - need additional analysis
1547 */
1548#define BYTE_SET_THRESHOLD (64)
1549
1550static u32 byte_set_size(const struct heuristic_ws *ws)
1551{
1552 u32 i;
1553 u32 byte_set_size = 0;
1554
1555 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1556 if (ws->bucket[i].count > 0)
1557 byte_set_size++;
1558 }
1559
1560 /*
1561 * Continue collecting count of byte values in buckets. If the byte
1562 * set size is bigger then the threshold, it's pointless to continue,
1563 * the detection technique would fail for this type of data.
1564 */
1565 for (; i < BUCKET_SIZE; i++) {
1566 if (ws->bucket[i].count > 0) {
1567 byte_set_size++;
1568 if (byte_set_size > BYTE_SET_THRESHOLD)
1569 return byte_set_size;
1570 }
1571 }
1572
1573 return byte_set_size;
1574}
1575
Timofey Titovets1fe4f6f2017-09-28 17:33:39 +03001576static bool sample_repeated_patterns(struct heuristic_ws *ws)
1577{
1578 const u32 half_of_sample = ws->sample_size / 2;
1579 const u8 *data = ws->sample;
1580
1581 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1582}
1583
Timofey Titovetsa440d482017-09-28 17:33:38 +03001584static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1585 struct heuristic_ws *ws)
1586{
1587 struct page *page;
1588 u64 index, index_end;
1589 u32 i, curr_sample_pos;
1590 u8 *in_data;
1591
1592 /*
1593 * Compression handles the input data by chunks of 128KiB
1594 * (defined by BTRFS_MAX_UNCOMPRESSED)
1595 *
1596 * We do the same for the heuristic and loop over the whole range.
1597 *
1598 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1599 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1600 */
1601 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1602 end = start + BTRFS_MAX_UNCOMPRESSED;
1603
1604 index = start >> PAGE_SHIFT;
1605 index_end = end >> PAGE_SHIFT;
1606
1607 /* Don't miss unaligned end */
1608 if (!IS_ALIGNED(end, PAGE_SIZE))
1609 index_end++;
1610
1611 curr_sample_pos = 0;
1612 while (index < index_end) {
1613 page = find_get_page(inode->i_mapping, index);
Ira Weiny58c1a352021-02-16 18:48:23 -08001614 in_data = kmap_local_page(page);
Timofey Titovetsa440d482017-09-28 17:33:38 +03001615 /* Handle case where the start is not aligned to PAGE_SIZE */
1616 i = start % PAGE_SIZE;
1617 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1618 /* Don't sample any garbage from the last page */
1619 if (start > end - SAMPLING_READ_SIZE)
1620 break;
1621 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1622 SAMPLING_READ_SIZE);
1623 i += SAMPLING_INTERVAL;
1624 start += SAMPLING_INTERVAL;
1625 curr_sample_pos += SAMPLING_READ_SIZE;
1626 }
Ira Weiny58c1a352021-02-16 18:48:23 -08001627 kunmap_local(in_data);
Timofey Titovetsa440d482017-09-28 17:33:38 +03001628 put_page(page);
1629
1630 index++;
1631 }
1632
1633 ws->sample_size = curr_sample_pos;
1634}
1635
Timofey Titovetsc2fcdcd2017-07-17 16:52:58 +03001636/*
1637 * Compression heuristic.
1638 *
1639 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1640 * quickly (compared to direct compression) detect data characteristics
1641 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1642 * data.
1643 *
1644 * The following types of analysis can be performed:
1645 * - detect mostly zero data
1646 * - detect data with low "byte set" size (text, etc)
1647 * - detect data with low/high "core byte" set
1648 *
1649 * Return non-zero if the compression should be done, 0 otherwise.
1650 */
1651int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1652{
Dennis Zhou7bf49942019-02-04 15:20:04 -05001653 struct list_head *ws_list = get_workspace(0, 0);
Timofey Titovets4e439a02017-09-28 17:33:36 +03001654 struct heuristic_ws *ws;
Timofey Titovetsa440d482017-09-28 17:33:38 +03001655 u32 i;
1656 u8 byte;
Timofey Titovets19562432017-10-08 16:11:59 +03001657 int ret = 0;
Timofey Titovetsc2fcdcd2017-07-17 16:52:58 +03001658
Timofey Titovets4e439a02017-09-28 17:33:36 +03001659 ws = list_entry(ws_list, struct heuristic_ws, list);
1660
Timofey Titovetsa440d482017-09-28 17:33:38 +03001661 heuristic_collect_sample(inode, start, end, ws);
1662
Timofey Titovets1fe4f6f2017-09-28 17:33:39 +03001663 if (sample_repeated_patterns(ws)) {
1664 ret = 1;
1665 goto out;
1666 }
1667
Timofey Titovetsa440d482017-09-28 17:33:38 +03001668 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1669
1670 for (i = 0; i < ws->sample_size; i++) {
1671 byte = ws->sample[i];
1672 ws->bucket[byte].count++;
Timofey Titovetsc2fcdcd2017-07-17 16:52:58 +03001673 }
1674
Timofey Titovetsa288e922017-09-28 17:33:40 +03001675 i = byte_set_size(ws);
1676 if (i < BYTE_SET_THRESHOLD) {
1677 ret = 2;
1678 goto out;
1679 }
1680
Timofey Titovets858177d2017-09-28 17:33:41 +03001681 i = byte_core_set_size(ws);
1682 if (i <= BYTE_CORE_SET_LOW) {
1683 ret = 3;
1684 goto out;
1685 }
1686
1687 if (i >= BYTE_CORE_SET_HIGH) {
1688 ret = 0;
1689 goto out;
1690 }
1691
Timofey Titovets19562432017-10-08 16:11:59 +03001692 i = shannon_entropy(ws);
1693 if (i <= ENTROPY_LVL_ACEPTABLE) {
1694 ret = 4;
1695 goto out;
1696 }
1697
1698 /*
1699 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1700 * needed to give green light to compression.
1701 *
1702 * For now just assume that compression at that level is not worth the
1703 * resources because:
1704 *
1705 * 1. it is possible to defrag the data later
1706 *
1707 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1708 * values, every bucket has counter at level ~54. The heuristic would
1709 * be confused. This can happen when data have some internal repeated
1710 * patterns like "abbacbbc...". This can be detected by analyzing
1711 * pairs of bytes, which is too costly.
1712 */
1713 if (i < ENTROPY_LVL_HIGH) {
1714 ret = 5;
1715 goto out;
1716 } else {
1717 ret = 0;
1718 goto out;
1719 }
1720
Timofey Titovets1fe4f6f2017-09-28 17:33:39 +03001721out:
Dennis Zhou929f4ba2019-02-04 15:20:02 -05001722 put_workspace(0, ws_list);
Timofey Titovetsc2fcdcd2017-07-17 16:52:58 +03001723 return ret;
1724}
David Sterbaf51d2b52017-09-15 17:36:57 +02001725
Dennis Zhoud0ab62c2019-02-04 15:20:05 -05001726/*
1727 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1728 * level, unrecognized string will set the default level
1729 */
1730unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
David Sterbaf51d2b52017-09-15 17:36:57 +02001731{
Dennis Zhoud0ab62c2019-02-04 15:20:05 -05001732 unsigned int level = 0;
1733 int ret;
1734
1735 if (!type)
David Sterbaf51d2b52017-09-15 17:36:57 +02001736 return 0;
1737
Dennis Zhoud0ab62c2019-02-04 15:20:05 -05001738 if (str[0] == ':') {
1739 ret = kstrtouint(str + 1, 10, &level);
1740 if (ret)
1741 level = 0;
1742 }
David Sterbaf51d2b52017-09-15 17:36:57 +02001743
David Sterbab0c1fe12019-08-09 16:49:06 +02001744 level = btrfs_compress_set_level(type, level);
1745
1746 return level;
1747}