David Sterba | c1d7c51 | 2018-04-03 19:23:33 +0200 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 2 | /* |
| 3 | * Copyright (C) 2008 Oracle. All rights reserved. |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 4 | */ |
| 5 | |
| 6 | #include <linux/kernel.h> |
| 7 | #include <linux/bio.h> |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 8 | #include <linux/file.h> |
| 9 | #include <linux/fs.h> |
| 10 | #include <linux/pagemap.h> |
Vishal Moola (Oracle) | a75b81c | 2022-08-23 17:40:19 -0700 | [diff] [blame] | 11 | #include <linux/pagevec.h> |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 12 | #include <linux/highmem.h> |
Christoph Hellwig | e41d12f | 2021-09-20 14:33:13 +0200 | [diff] [blame] | 13 | #include <linux/kthread.h> |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 14 | #include <linux/time.h> |
| 15 | #include <linux/init.h> |
| 16 | #include <linux/string.h> |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 17 | #include <linux/backing-dev.h> |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 18 | #include <linux/writeback.h> |
Christoph Hellwig | 4088a47 | 2022-09-15 10:41:58 +0100 | [diff] [blame] | 19 | #include <linux/psi.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 20 | #include <linux/slab.h> |
David Sterba | fe30853 | 2017-05-31 17:14:56 +0200 | [diff] [blame] | 21 | #include <linux/sched/mm.h> |
Timofey Titovets | 1956243 | 2017-10-08 16:11:59 +0300 | [diff] [blame] | 22 | #include <linux/log2.h> |
Johannes Thumshirn | d517857 | 2019-06-03 16:58:57 +0200 | [diff] [blame] | 23 | #include <crypto/hash.h> |
David Sterba | 602cbe9 | 2019-08-21 18:48:25 +0200 | [diff] [blame] | 24 | #include "misc.h" |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 25 | #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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 31 | #include "compression.h" |
| 32 | #include "extent_io.h" |
| 33 | #include "extent_map.h" |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 34 | #include "subpage.h" |
Johannes Thumshirn | 764c7c9 | 2021-05-19 00:40:28 +0900 | [diff] [blame] | 35 | #include "zoned.h" |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 36 | |
David Sterba | e128f9c | 2017-10-31 17:24:26 +0100 | [diff] [blame] | 37 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
| 38 | |
| 39 | const 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 Xu | ce96b7f | 2019-10-10 15:59:57 +0800 | [diff] [blame] | 47 | default: |
| 48 | break; |
David Sterba | e128f9c | 2017-10-31 17:24:26 +0100 | [diff] [blame] | 49 | } |
| 50 | |
| 51 | return NULL; |
| 52 | } |
| 53 | |
Johannes Thumshirn | aa53e3b | 2019-06-06 12:07:15 +0200 | [diff] [blame] | 54 | bool 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 Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 70 | static 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 Wenruo | 1d8ba9e | 2020-08-04 15:25:47 +0800 | [diff] [blame] | 88 | * 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 Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 95 | */ |
Qu Wenruo | 1d8ba9e | 2020-08-04 15:25:47 +0800 | [diff] [blame] | 96 | *out_pages = 0; |
David Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 97 | return -E2BIG; |
| 98 | } |
| 99 | } |
| 100 | |
Su Yue | 4a9e803 | 2021-12-27 18:18:39 +0800 | [diff] [blame] | 101 | static int compression_decompress_bio(struct list_head *ws, |
| 102 | struct compressed_bio *cb) |
David Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 103 | { |
Su Yue | 4a9e803 | 2021-12-27 18:18:39 +0800 | [diff] [blame] | 104 | switch (cb->compress_type) { |
David Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 105 | 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 | |
| 118 | static 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 Jain | 8140dc3 | 2017-05-26 15:44:58 +0800 | [diff] [blame] | 139 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
Eric Sandeen | 48a3b63 | 2013-04-25 20:41:01 +0000 | [diff] [blame] | 140 | |
Josef Bacik | e14bfdb | 2022-02-18 10:03:25 -0500 | [diff] [blame] | 141 | static void finish_compressed_bio_read(struct compressed_bio *cb) |
Qu Wenruo | 86ccbb4 | 2021-09-27 15:21:50 +0800 | [diff] [blame] | 142 | { |
| 143 | unsigned int index; |
| 144 | struct page *page; |
| 145 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 146 | if (cb->status == BLK_STS_OK) |
| 147 | cb->status = errno_to_blk_status(btrfs_decompress_bio(cb)); |
| 148 | |
Qu Wenruo | 86ccbb4 | 2021-09-27 15:21:50 +0800 | [diff] [blame] | 149 | /* 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 Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 157 | btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status); |
Qu Wenruo | 86ccbb4 | 2021-09-27 15:21:50 +0800 | [diff] [blame] | 158 | |
| 159 | /* Finally free the cb struct */ |
| 160 | kfree(cb->compressed_pages); |
| 161 | kfree(cb); |
| 162 | } |
| 163 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 164 | /* |
| 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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 168 | */ |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 169 | static void end_compressed_bio_read(struct btrfs_bio *bbio) |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 170 | { |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 171 | struct compressed_bio *cb = bbio->private; |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 172 | 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 Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 177 | blk_status_t status = bbio->bio.bi_status; |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 178 | struct bvec_iter iter; |
| 179 | struct bio_vec bv; |
| 180 | u32 offset; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 181 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 182 | btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) { |
| 183 | u64 start = bbio->file_offset + offset; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 184 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 185 | if (!status && |
| 186 | (!csum || !btrfs_check_data_csum(inode, bbio, offset, |
| 187 | bv.bv_page, bv.bv_offset))) { |
Josef Bacik | 0d0a762 | 2022-09-09 17:53:14 -0400 | [diff] [blame] | 188 | btrfs_clean_io_failure(bi, start, bv.bv_page, |
| 189 | bv.bv_offset); |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 190 | } else { |
| 191 | int ret; |
Liu Bo | cf1167d | 2017-09-20 17:50:18 -0600 | [diff] [blame] | 192 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 193 | 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 Bo | e6311f2 | 2017-09-20 17:50:19 -0600 | [diff] [blame] | 203 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 204 | if (status) |
| 205 | cb->status = status; |
Chris Mason | d20f704 | 2008-12-08 16:58:54 -0500 | [diff] [blame] | 206 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 207 | if (refcount_dec_and_test(&cb->pending_ios)) |
| 208 | finish_compressed_bio_read(cb); |
| 209 | btrfs_bio_free_csum(bbio); |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 210 | bio_put(&bbio->bio); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 211 | } |
| 212 | |
| 213 | /* |
| 214 | * Clear the writeback bits on all of the file |
| 215 | * pages for a compressed write |
| 216 | */ |
Filipe Manana | 7bdcefc1 | 2014-10-07 01:48:26 +0100 | [diff] [blame] | 217 | static noinline void end_compressed_writeback(struct inode *inode, |
| 218 | const struct compressed_bio *cb) |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 219 | { |
Qu Wenruo | 741ec65 | 2021-09-27 15:22:01 +0800 | [diff] [blame] | 220 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 221 | unsigned long index = cb->start >> PAGE_SHIFT; |
| 222 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; |
Vishal Moola (Oracle) | a75b81c | 2022-08-23 17:40:19 -0700 | [diff] [blame] | 223 | struct folio_batch fbatch; |
Josef Bacik | 606f82e | 2022-02-18 10:03:26 -0500 | [diff] [blame] | 224 | const int errno = blk_status_to_errno(cb->status); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 225 | int i; |
| 226 | int ret; |
| 227 | |
Josef Bacik | 606f82e | 2022-02-18 10:03:26 -0500 | [diff] [blame] | 228 | if (errno) |
| 229 | mapping_set_error(inode->i_mapping, errno); |
Filipe Manana | 7bdcefc1 | 2014-10-07 01:48:26 +0100 | [diff] [blame] | 230 | |
Vishal Moola (Oracle) | a75b81c | 2022-08-23 17:40:19 -0700 | [diff] [blame] | 231 | 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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 239 | for (i = 0; i < ret; i++) { |
Vishal Moola (Oracle) | a75b81c | 2022-08-23 17:40:19 -0700 | [diff] [blame] | 240 | struct folio *folio = fbatch.folios[i]; |
| 241 | |
Josef Bacik | 606f82e | 2022-02-18 10:03:26 -0500 | [diff] [blame] | 242 | if (errno) |
Vishal Moola (Oracle) | a75b81c | 2022-08-23 17:40:19 -0700 | [diff] [blame] | 243 | folio_set_error(folio); |
| 244 | btrfs_page_clamp_clear_writeback(fs_info, &folio->page, |
Qu Wenruo | 741ec65 | 2021-09-27 15:22:01 +0800 | [diff] [blame] | 245 | cb->start, cb->len); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 246 | } |
Vishal Moola (Oracle) | a75b81c | 2022-08-23 17:40:19 -0700 | [diff] [blame] | 247 | folio_batch_release(&fbatch); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 248 | } |
| 249 | /* the inode may be gone now */ |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 250 | } |
| 251 | |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 252 | static void finish_compressed_bio_write(struct compressed_bio *cb) |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 253 | { |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 254 | struct inode *inode = cb->inode; |
Anand Jain | 1d08ce58 | 2021-05-29 17:48:33 +0800 | [diff] [blame] | 255 | unsigned int index; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 256 | |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 257 | /* |
| 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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 260 | */ |
Qu Wenruo | 38a39ac7 | 2021-04-08 20:32:27 +0800 | [diff] [blame] | 261 | btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL, |
Nikolay Borisov | c629732 | 2018-11-08 10:18:08 +0200 | [diff] [blame] | 262 | cb->start, cb->start + cb->len - 1, |
Josef Bacik | 606f82e | 2022-02-18 10:03:26 -0500 | [diff] [blame] | 263 | cb->status == BLK_STS_OK); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 264 | |
Omar Sandoval | 7c0c726 | 2019-08-13 16:00:02 -0700 | [diff] [blame] | 265 | if (cb->writeback) |
| 266 | end_compressed_writeback(inode, cb); |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 267 | /* Note, our inode could be gone now */ |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 268 | |
| 269 | /* |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 270 | * Release the compressed pages, these came from alloc_page and |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 271 | * are not attached to the inode at all |
| 272 | */ |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 273 | for (index = 0; index < cb->nr_pages; index++) { |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 274 | struct page *page = cb->compressed_pages[index]; |
| 275 | |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 276 | page->mapping = NULL; |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 277 | put_page(page); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 278 | } |
| 279 | |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 280 | /* Finally free the cb struct */ |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 281 | kfree(cb->compressed_pages); |
| 282 | kfree(cb); |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 283 | } |
| 284 | |
Christoph Hellwig | fed8a72 | 2022-05-26 09:36:38 +0200 | [diff] [blame] | 285 | static 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 Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 293 | /* |
| 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 Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 300 | static void end_compressed_bio_write(struct btrfs_bio *bbio) |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 301 | { |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 302 | struct compressed_bio *cb = bbio->private; |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 303 | |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 304 | if (bbio->bio.bi_status) |
| 305 | cb->status = bbio->bio.bi_status; |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 306 | |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 307 | if (refcount_dec_and_test(&cb->pending_ios)) { |
Christoph Hellwig | fed8a72 | 2022-05-26 09:36:38 +0200 | [diff] [blame] | 308 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 309 | |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 310 | btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio); |
Christoph Hellwig | fed8a72 | 2022-05-26 09:36:38 +0200 | [diff] [blame] | 311 | queue_work(fs_info->compressed_write_workers, &cb->write_end_work); |
| 312 | } |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 313 | bio_put(&bbio->bio); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 314 | } |
| 315 | |
| 316 | /* |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 317 | * 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 Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 330 | */ |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 331 | |
| 332 | |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 333 | static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr, |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 334 | blk_opf_t opf, |
| 335 | btrfs_bio_end_io_t endio_func, |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 336 | u64 *next_stripe_start) |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 337 | { |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 338 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); |
| 339 | struct btrfs_io_geometry geom; |
| 340 | struct extent_map *em; |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 341 | struct bio *bio; |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 342 | int ret; |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 343 | |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 344 | bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, endio_func, cb); |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 345 | bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT; |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 346 | |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 347 | 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 Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 351 | } |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 352 | |
| 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 Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 363 | refcount_inc(&cb->pending_ios); |
Qu Wenruo | 22c306f | 2021-09-27 15:21:53 +0800 | [diff] [blame] | 364 | return bio; |
| 365 | } |
| 366 | |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 367 | /* |
| 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 Borisov | c7ee181 | 2020-06-03 08:55:16 +0300 | [diff] [blame] | 376 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
Anand Jain | 65b5355 | 2021-05-29 17:48:35 +0800 | [diff] [blame] | 377 | unsigned int len, u64 disk_start, |
| 378 | unsigned int compressed_len, |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 379 | struct page **compressed_pages, |
Anand Jain | 65b5355 | 2021-05-29 17:48:35 +0800 | [diff] [blame] | 380 | unsigned int nr_pages, |
Bart Van Assche | bf9486d | 2022-07-14 11:07:16 -0700 | [diff] [blame] | 381 | blk_opf_t write_flags, |
Omar Sandoval | 7c0c726 | 2019-08-13 16:00:02 -0700 | [diff] [blame] | 382 | struct cgroup_subsys_state *blkcg_css, |
| 383 | bool writeback) |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 384 | { |
Nikolay Borisov | c7ee181 | 2020-06-03 08:55:16 +0300 | [diff] [blame] | 385 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 386 | struct bio *bio = NULL; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 387 | struct compressed_bio *cb; |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 388 | u64 cur_disk_bytenr = disk_start; |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 389 | u64 next_stripe_start; |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 390 | blk_status_t ret = BLK_STS_OK; |
Nikolay Borisov | c7ee181 | 2020-06-03 08:55:16 +0300 | [diff] [blame] | 391 | int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; |
Johannes Thumshirn | 764c7c9 | 2021-05-19 00:40:28 +0900 | [diff] [blame] | 392 | const bool use_append = btrfs_use_zone_append(inode, disk_start); |
Bart Van Assche | bf9486d | 2022-07-14 11:07:16 -0700 | [diff] [blame] | 393 | const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 394 | |
Qu Wenruo | bbbff01 | 2021-09-27 15:22:00 +0800 | [diff] [blame] | 395 | ASSERT(IS_ALIGNED(start, fs_info->sectorsize) && |
| 396 | IS_ALIGNED(len, fs_info->sectorsize)); |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 397 | cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS); |
Yoshinori Sano | dac97e5 | 2011-02-15 12:01:42 +0000 | [diff] [blame] | 398 | if (!cb) |
Christoph Hellwig | 4e4cbee | 2017-06-03 09:38:06 +0200 | [diff] [blame] | 399 | return BLK_STS_RESOURCE; |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 400 | refcount_set(&cb->pending_ios, 1); |
Josef Bacik | 606f82e | 2022-02-18 10:03:26 -0500 | [diff] [blame] | 401 | cb->status = BLK_STS_OK; |
Nikolay Borisov | c7ee181 | 2020-06-03 08:55:16 +0300 | [diff] [blame] | 402 | cb->inode = &inode->vfs_inode; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 403 | cb->start = start; |
| 404 | cb->len = len; |
| 405 | cb->compressed_pages = compressed_pages; |
| 406 | cb->compressed_len = compressed_len; |
Omar Sandoval | 7c0c726 | 2019-08-13 16:00:02 -0700 | [diff] [blame] | 407 | cb->writeback = writeback; |
Christoph Hellwig | fed8a72 | 2022-05-26 09:36:38 +0200 | [diff] [blame] | 408 | INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 409 | cb->nr_pages = nr_pages; |
| 410 | |
Dennis Zhou | acee08a | 2022-03-31 14:58:28 -0700 | [diff] [blame] | 411 | if (blkcg_css) |
| 412 | kthread_associate_blkcg(blkcg_css); |
| 413 | |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 414 | 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 Mason | ec39f76 | 2019-07-10 12:28:17 -0700 | [diff] [blame] | 421 | |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 422 | /* 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 Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 429 | break; |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 430 | } |
Dennis Zhou | acee08a | 2022-03-31 14:58:28 -0700 | [diff] [blame] | 431 | if (blkcg_css) |
| 432 | bio->bi_opf |= REQ_CGROUP_PUNT; |
Johannes Thumshirn | 764c7c9 | 2021-05-19 00:40:28 +0900 | [diff] [blame] | 433 | } |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 434 | /* |
| 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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 439 | |
Qu Wenruo | 4c80a97 | 2021-05-25 13:52:43 +0800 | [diff] [blame] | 440 | /* |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 441 | * We have various limits on the real read size: |
| 442 | * - stripe boundary |
| 443 | * - page boundary |
| 444 | * - compressed length boundary |
Qu Wenruo | 4c80a97 | 2021-05-25 13:52:43 +0800 | [diff] [blame] | 445 | */ |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 446 | 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 Thumshirn | 764c7c9 | 2021-05-19 00:40:28 +0900 | [diff] [blame] | 450 | |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 451 | 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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 460 | |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 461 | 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 Zefan | e55179b | 2011-07-14 03:16:47 +0000 | [diff] [blame] | 471 | if (!skip_sum) { |
Omar Sandoval | e331f6b | 2019-11-06 15:38:43 -0800 | [diff] [blame] | 472 | ret = btrfs_csum_one_bio(inode, bio, start, true); |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 473 | if (ret) { |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 474 | btrfs_bio_end_io(btrfs_bio(bio), ret); |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 475 | break; |
| 476 | } |
Li Zefan | e55179b | 2011-07-14 03:16:47 +0000 | [diff] [blame] | 477 | } |
Chris Mason | d20f704 | 2008-12-08 16:58:54 -0500 | [diff] [blame] | 478 | |
Christoph Hellwig | fed8a72 | 2022-05-26 09:36:38 +0200 | [diff] [blame] | 479 | ASSERT(bio->bi_iter.bi_size); |
Christoph Hellwig | 1a722d8 | 2022-06-17 12:04:07 +0200 | [diff] [blame] | 480 | btrfs_submit_bio(fs_info, bio, 0); |
Qu Wenruo | 9150724 | 2021-09-27 15:21:55 +0800 | [diff] [blame] | 481 | bio = NULL; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 482 | } |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 483 | cond_resched(); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 484 | } |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 485 | |
Dennis Zhou | 46bcff2b | 2019-12-11 15:20:15 -0800 | [diff] [blame] | 486 | if (blkcg_css) |
| 487 | kthread_associate_blkcg(NULL); |
| 488 | |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 489 | if (refcount_dec_and_test(&cb->pending_ios)) |
| 490 | finish_compressed_bio_write(cb); |
Qu Wenruo | 6853c64 | 2021-09-27 15:21:51 +0800 | [diff] [blame] | 491 | return ret; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 492 | } |
| 493 | |
Christoph Hellwig | 2a4d0c9 | 2016-11-25 09:07:51 +0100 | [diff] [blame] | 494 | static u64 bio_end_offset(struct bio *bio) |
| 495 | { |
Ming Lei | c45a8f2 | 2017-12-18 20:22:05 +0800 | [diff] [blame] | 496 | struct bio_vec *last = bio_last_bvec_all(bio); |
Christoph Hellwig | 2a4d0c9 | 2016-11-25 09:07:51 +0100 | [diff] [blame] | 497 | |
| 498 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; |
| 499 | } |
| 500 | |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 501 | /* |
| 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 Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 512 | static noinline int add_ra_bio_pages(struct inode *inode, |
| 513 | u64 compressed_end, |
Christoph Hellwig | 4088a47 | 2022-09-15 10:41:58 +0100 | [diff] [blame] | 514 | struct compressed_bio *cb, |
Johannes Weiner | 82e60d0 | 2022-11-03 17:34:31 -0400 | [diff] [blame] | 515 | int *memstall, unsigned long *pflags) |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 516 | { |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 517 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 518 | unsigned long end_index; |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 519 | u64 cur = bio_end_offset(cb->orig_bio); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 520 | u64 isize = i_size_read(inode); |
| 521 | int ret; |
| 522 | struct page *page; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 523 | struct extent_map *em; |
| 524 | struct address_space *mapping = inode->i_mapping; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 525 | struct extent_map_tree *em_tree; |
| 526 | struct extent_io_tree *tree; |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 527 | int sectors_missed = 0; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 528 | |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 529 | em_tree = &BTRFS_I(inode)->extent_tree; |
| 530 | tree = &BTRFS_I(inode)->io_tree; |
| 531 | |
| 532 | if (isize == 0) |
| 533 | return 0; |
| 534 | |
Qu Wenruo | ca62e85 | 2021-07-26 14:34:52 +0800 | [diff] [blame] | 535 | /* |
| 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. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 545 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 546 | |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 547 | while (cur < compressed_end) { |
| 548 | u64 page_end; |
| 549 | u64 pg_index = cur >> PAGE_SHIFT; |
| 550 | u32 add_size; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 551 | |
David Sterba | 306e16c | 2011-04-19 14:29:38 +0200 | [diff] [blame] | 552 | if (pg_index > end_index) |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 553 | break; |
| 554 | |
Matthew Wilcox | 0a943c6 | 2017-12-04 10:37:22 -0500 | [diff] [blame] | 555 | page = xa_load(&mapping->i_pages, pg_index); |
Matthew Wilcox | 3159f94 | 2017-11-03 13:30:42 -0400 | [diff] [blame] | 556 | if (page && !xa_is_value(page)) { |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 557 | 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 Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 562 | break; |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 563 | |
| 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 Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 570 | } |
| 571 | |
Michal Hocko | c62d255 | 2015-11-06 16:28:49 -0800 | [diff] [blame] | 572 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
| 573 | ~__GFP_FS)); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 574 | if (!page) |
| 575 | break; |
| 576 | |
Michal Hocko | c62d255 | 2015-11-06 16:28:49 -0800 | [diff] [blame] | 577 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 578 | put_page(page); |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 579 | /* There is already a page, skip to page end */ |
| 580 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; |
| 581 | continue; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 582 | } |
| 583 | |
Johannes Weiner | 82e60d0 | 2022-11-03 17:34:31 -0400 | [diff] [blame] | 584 | if (!*memstall && PageWorkingset(page)) { |
Christoph Hellwig | 4088a47 | 2022-09-15 10:41:58 +0100 | [diff] [blame] | 585 | psi_memstall_enter(pflags); |
Johannes Weiner | 82e60d0 | 2022-11-03 17:34:31 -0400 | [diff] [blame] | 586 | *memstall = 1; |
| 587 | } |
Christoph Hellwig | 4088a47 | 2022-09-15 10:41:58 +0100 | [diff] [blame] | 588 | |
Qu Wenruo | 32443de | 2021-01-26 16:34:00 +0800 | [diff] [blame] | 589 | ret = set_page_extent_mapped(page); |
| 590 | if (ret < 0) { |
| 591 | unlock_page(page); |
| 592 | put_page(page); |
| 593 | break; |
| 594 | } |
| 595 | |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 596 | page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; |
Josef Bacik | 570eb97 | 2022-09-09 17:53:43 -0400 | [diff] [blame] | 597 | lock_extent(tree, cur, page_end, NULL); |
Chris Mason | 890871b | 2009-09-02 16:24:52 -0400 | [diff] [blame] | 598 | read_lock(&em_tree->lock); |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 599 | em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); |
Chris Mason | 890871b | 2009-09-02 16:24:52 -0400 | [diff] [blame] | 600 | read_unlock(&em_tree->lock); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 601 | |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 602 | /* |
| 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 Overstreet | 4f024f3 | 2013-10-11 15:44:27 -0700 | [diff] [blame] | 609 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 610 | free_extent_map(em); |
Josef Bacik | 570eb97 | 2022-09-09 17:53:43 -0400 | [diff] [blame] | 611 | unlock_extent(tree, cur, page_end, NULL); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 612 | unlock_page(page); |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 613 | put_page(page); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 614 | break; |
| 615 | } |
| 616 | free_extent_map(em); |
| 617 | |
| 618 | if (page->index == end_index) { |
Johannes Thumshirn | 7073017 | 2018-12-05 15:23:03 +0100 | [diff] [blame] | 619 | size_t zero_offset = offset_in_page(isize); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 620 | |
| 621 | if (zero_offset) { |
| 622 | int zeros; |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 623 | zeros = PAGE_SIZE - zero_offset; |
Ira Weiny | d048b9c | 2021-05-04 18:40:07 -0700 | [diff] [blame] | 624 | memzero_page(page, zero_offset, zeros); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 625 | } |
| 626 | } |
| 627 | |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 628 | 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 Bacik | 570eb97 | 2022-09-09 17:53:43 -0400 | [diff] [blame] | 631 | unlock_extent(tree, cur, page_end, NULL); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 632 | unlock_page(page); |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 633 | put_page(page); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 634 | break; |
| 635 | } |
Qu Wenruo | 6a40491 | 2021-09-27 15:21:47 +0800 | [diff] [blame] | 636 | /* |
| 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 Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 645 | } |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 646 | return 0; |
| 647 | } |
| 648 | |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 649 | /* |
| 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 Overstreet | 4f024f3 | 2013-10-11 15:44:27 -0700 | [diff] [blame] | 654 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 655 | * bio->bi_io_vec points to all of the inode pages |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 656 | * |
| 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 Hellwig | cb4411d | 2022-04-15 16:33:27 +0200 | [diff] [blame] | 660 | void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
Goldwyn Rodrigues | 1d8fa2e2 | 2022-04-26 08:47:34 -0500 | [diff] [blame] | 661 | int mirror_num) |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 662 | { |
Jeff Mahoney | 0b246af | 2016-06-22 18:54:23 -0400 | [diff] [blame] | 663 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 664 | struct extent_map_tree *em_tree; |
| 665 | struct compressed_bio *cb; |
Anand Jain | 356b4a2 | 2021-05-29 17:48:34 +0800 | [diff] [blame] | 666 | unsigned int compressed_len; |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 667 | 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 Wenruo | 557023e | 2021-07-05 10:00:56 +0800 | [diff] [blame] | 671 | u64 file_offset; |
Chris Mason | e04ca62 | 2008-11-10 11:44:58 -0500 | [diff] [blame] | 672 | u64 em_len; |
| 673 | u64 em_start; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 674 | struct extent_map *em; |
Johannes Weiner | 82e60d0 | 2022-11-03 17:34:31 -0400 | [diff] [blame] | 675 | unsigned long pflags; |
| 676 | int memstall = 0; |
Josef Bacik | f9f15de | 2022-02-18 10:03:27 -0500 | [diff] [blame] | 677 | blk_status_t ret; |
Sweet Tea Dorminy | dd137dd | 2022-03-30 16:11:22 -0400 | [diff] [blame] | 678 | int ret2; |
| 679 | int i; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 680 | |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 681 | em_tree = &BTRFS_I(inode)->extent_tree; |
| 682 | |
Qu Wenruo | 557023e | 2021-07-05 10:00:56 +0800 | [diff] [blame] | 683 | file_offset = bio_first_bvec_all(bio)->bv_offset + |
| 684 | page_offset(bio_first_page_all(bio)); |
| 685 | |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 686 | /* we need the actual starting offset of this extent in the file */ |
Chris Mason | 890871b | 2009-09-02 16:24:52 -0400 | [diff] [blame] | 687 | read_lock(&em_tree->lock); |
Qu Wenruo | 557023e | 2021-07-05 10:00:56 +0800 | [diff] [blame] | 688 | em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); |
Chris Mason | 890871b | 2009-09-02 16:24:52 -0400 | [diff] [blame] | 689 | read_unlock(&em_tree->lock); |
Josef Bacik | f9f15de | 2022-02-18 10:03:27 -0500 | [diff] [blame] | 690 | if (!em) { |
| 691 | ret = BLK_STS_IOERR; |
| 692 | goto out; |
| 693 | } |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 694 | |
Qu Wenruo | 557023e | 2021-07-05 10:00:56 +0800 | [diff] [blame] | 695 | ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); |
Chris Mason | d20f704 | 2008-12-08 16:58:54 -0500 | [diff] [blame] | 696 | compressed_len = em->block_len; |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 697 | cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS); |
Josef Bacik | f9f15de | 2022-02-18 10:03:27 -0500 | [diff] [blame] | 698 | if (!cb) { |
| 699 | ret = BLK_STS_RESOURCE; |
liubo | 6b82ce8 | 2011-01-26 06:21:39 +0000 | [diff] [blame] | 700 | goto out; |
Josef Bacik | f9f15de | 2022-02-18 10:03:27 -0500 | [diff] [blame] | 701 | } |
liubo | 6b82ce8 | 2011-01-26 06:21:39 +0000 | [diff] [blame] | 702 | |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 703 | refcount_set(&cb->pending_ios, 1); |
Josef Bacik | 606f82e | 2022-02-18 10:03:26 -0500 | [diff] [blame] | 704 | cb->status = BLK_STS_OK; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 705 | cb->inode = inode; |
| 706 | |
Yan Zheng | ff5b7ee | 2008-11-10 07:34:43 -0500 | [diff] [blame] | 707 | cb->start = em->orig_start; |
Chris Mason | e04ca62 | 2008-11-10 11:44:58 -0500 | [diff] [blame] | 708 | em_len = em->len; |
| 709 | em_start = em->start; |
Chris Mason | d20f704 | 2008-12-08 16:58:54 -0500 | [diff] [blame] | 710 | |
Christoph Hellwig | 8138105 | 2016-11-25 09:07:50 +0100 | [diff] [blame] | 711 | cb->len = bio->bi_iter.bi_size; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 712 | cb->compressed_len = compressed_len; |
Goldwyn Rodrigues | 1d8fa2e2 | 2022-04-26 08:47:34 -0500 | [diff] [blame] | 713 | cb->compress_type = em->compress_type; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 714 | cb->orig_bio = bio; |
| 715 | |
Goldwyn Rodrigues | 1d8fa2e2 | 2022-04-26 08:47:34 -0500 | [diff] [blame] | 716 | free_extent_map(em); |
| 717 | em = NULL; |
| 718 | |
Sweet Tea Dorminy | dd137dd | 2022-03-30 16:11:22 -0400 | [diff] [blame] | 719 | cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
| 720 | cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS); |
Josef Bacik | f9f15de | 2022-02-18 10:03:27 -0500 | [diff] [blame] | 721 | if (!cb->compressed_pages) { |
| 722 | ret = BLK_STS_RESOURCE; |
Sweet Tea Dorminy | dd137dd | 2022-03-30 16:11:22 -0400 | [diff] [blame] | 723 | goto fail; |
Josef Bacik | f9f15de | 2022-02-18 10:03:27 -0500 | [diff] [blame] | 724 | } |
liubo | 6b82ce8 | 2011-01-26 06:21:39 +0000 | [diff] [blame] | 725 | |
Sweet Tea Dorminy | dd137dd | 2022-03-30 16:11:22 -0400 | [diff] [blame] | 726 | ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages); |
| 727 | if (ret2) { |
| 728 | ret = BLK_STS_RESOURCE; |
| 729 | goto fail; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 730 | } |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 731 | |
Johannes Weiner | 82e60d0 | 2022-11-03 17:34:31 -0400 | [diff] [blame] | 732 | add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags); |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 733 | |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 734 | /* include any pages we added in add_ra-bio_pages */ |
Christoph Hellwig | 8138105 | 2016-11-25 09:07:50 +0100 | [diff] [blame] | 735 | cb->len = bio->bi_iter.bi_size; |
Chris Mason | 771ed68 | 2008-11-06 22:02:51 -0500 | [diff] [blame] | 736 | |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 737 | 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 Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 744 | |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 745 | /* 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 Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 751 | cb->status = errno_to_blk_status(PTR_ERR(comp_bio)); |
| 752 | break; |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 753 | } |
| 754 | } |
Qu Wenruo | be6a136 | 2021-02-04 15:03:23 +0800 | [diff] [blame] | 755 | /* |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 756 | * We should never reach next_stripe_start start as we will |
| 757 | * submit comp_bio when reach the boundary immediately. |
Qu Wenruo | be6a136 | 2021-02-04 15:03:23 +0800 | [diff] [blame] | 758 | */ |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 759 | 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 Wenruo | be6a136 | 2021-02-04 15:03:23 +0800 | [diff] [blame] | 770 | |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 771 | 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 Mason | d20f704 | 2008-12-08 16:58:54 -0500 | [diff] [blame] | 778 | |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 779 | /* Reached stripe boundary, need to submit */ |
| 780 | if (cur_disk_byte == next_stripe_start) |
| 781 | submit = true; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 782 | |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 783 | /* Has finished the range, need to submit */ |
| 784 | if (cur_disk_byte == disk_bytenr + compressed_len) |
| 785 | submit = true; |
| 786 | |
| 787 | if (submit) { |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 788 | /* 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 Thumshirn | 10fe6ca | 2019-05-22 10:19:02 +0200 | [diff] [blame] | 791 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 792 | /* |
| 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 Thumshirn | 10fe6ca | 2019-05-22 10:19:02 +0200 | [diff] [blame] | 799 | |
Christoph Hellwig | 81bd932 | 2022-07-07 07:33:30 +0200 | [diff] [blame] | 800 | ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL); |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 801 | if (ret) { |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 802 | btrfs_bio_end_io(btrfs_bio(comp_bio), ret); |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 803 | break; |
| 804 | } |
Chris Mason | d20f704 | 2008-12-08 16:58:54 -0500 | [diff] [blame] | 805 | |
Christoph Hellwig | fed8a72 | 2022-05-26 09:36:38 +0200 | [diff] [blame] | 806 | ASSERT(comp_bio->bi_iter.bi_size); |
Christoph Hellwig | 1a722d8 | 2022-06-17 12:04:07 +0200 | [diff] [blame] | 807 | btrfs_submit_bio(fs_info, comp_bio, mirror_num); |
Qu Wenruo | f472c28f | 2021-09-27 15:21:54 +0800 | [diff] [blame] | 808 | comp_bio = NULL; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 809 | } |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 810 | } |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 811 | |
Johannes Weiner | 82e60d0 | 2022-11-03 17:34:31 -0400 | [diff] [blame] | 812 | if (memstall) |
Christoph Hellwig | 4088a47 | 2022-09-15 10:41:58 +0100 | [diff] [blame] | 813 | psi_memstall_leave(&pflags); |
| 814 | |
Christoph Hellwig | 524bcd1 | 2022-07-07 07:33:27 +0200 | [diff] [blame] | 815 | if (refcount_dec_and_test(&cb->pending_ios)) |
| 816 | finish_compressed_bio_read(cb); |
Christoph Hellwig | cb4411d | 2022-04-15 16:33:27 +0200 | [diff] [blame] | 817 | return; |
liubo | 6b82ce8 | 2011-01-26 06:21:39 +0000 | [diff] [blame] | 818 | |
Sweet Tea Dorminy | dd137dd | 2022-03-30 16:11:22 -0400 | [diff] [blame] | 819 | fail: |
| 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 Bacik | 15e3004a | 2012-10-05 13:39:50 -0400 | [diff] [blame] | 825 | } |
liubo | 6b82ce8 | 2011-01-26 06:21:39 +0000 | [diff] [blame] | 826 | |
| 827 | kfree(cb->compressed_pages); |
liubo | 6b82ce8 | 2011-01-26 06:21:39 +0000 | [diff] [blame] | 828 | kfree(cb); |
| 829 | out: |
| 830 | free_extent_map(em); |
Christoph Hellwig | 917f32a | 2022-08-06 10:03:26 +0200 | [diff] [blame] | 831 | btrfs_bio_end_io(btrfs_bio(bio), ret); |
Christoph Hellwig | cb4411d | 2022-04-15 16:33:27 +0200 | [diff] [blame] | 832 | return; |
Chris Mason | c8b9781 | 2008-10-29 14:49:59 -0400 | [diff] [blame] | 833 | } |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 834 | |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 835 | /* |
| 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 | |
| 867 | struct bucket_item { |
| 868 | u32 count; |
| 869 | }; |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 870 | |
| 871 | struct heuristic_ws { |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 872 | /* Partial copy of input data */ |
| 873 | u8 *sample; |
Timofey Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 874 | u32 sample_size; |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 875 | /* Buckets store counters for each byte value */ |
| 876 | struct bucket_item *bucket; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 877 | /* Sorting buffer */ |
| 878 | struct bucket_item *bucket_b; |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 879 | struct list_head list; |
| 880 | }; |
| 881 | |
Dennis Zhou | 92ee5530 | 2019-02-04 15:20:03 -0500 | [diff] [blame] | 882 | static struct workspace_manager heuristic_wsm; |
| 883 | |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 884 | static 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 Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 890 | kvfree(workspace->sample); |
| 891 | kfree(workspace->bucket); |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 892 | kfree(workspace->bucket_b); |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 893 | kfree(workspace); |
| 894 | } |
| 895 | |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 896 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 897 | { |
| 898 | struct heuristic_ws *ws; |
| 899 | |
| 900 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); |
| 901 | if (!ws) |
| 902 | return ERR_PTR(-ENOMEM); |
| 903 | |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 904 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
| 905 | if (!ws->sample) |
| 906 | goto fail; |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 907 | |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 908 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); |
| 909 | if (!ws->bucket) |
| 910 | goto fail; |
| 911 | |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 912 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
| 913 | if (!ws->bucket_b) |
| 914 | goto fail; |
| 915 | |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 916 | INIT_LIST_HEAD(&ws->list); |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 917 | return &ws->list; |
Timofey Titovets | 17b5a6c | 2017-09-28 17:33:37 +0300 | [diff] [blame] | 918 | fail: |
| 919 | free_heuristic_ws(&ws->list); |
| 920 | return ERR_PTR(-ENOMEM); |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 921 | } |
| 922 | |
Dennis Zhou | ca4ac36 | 2019-02-04 15:19:59 -0500 | [diff] [blame] | 923 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
David Sterba | be951045 | 2019-10-02 00:53:31 +0200 | [diff] [blame] | 924 | .workspace_manager = &heuristic_wsm, |
Dennis Zhou | ca4ac36 | 2019-02-04 15:19:59 -0500 | [diff] [blame] | 925 | }; |
| 926 | |
David Sterba | e8c9f18 | 2015-01-02 18:23:10 +0100 | [diff] [blame] | 927 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
Dennis Zhou | ca4ac36 | 2019-02-04 15:19:59 -0500 | [diff] [blame] | 928 | /* The heuristic is represented as compression type 0 */ |
| 929 | &btrfs_heuristic_compress, |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 930 | &btrfs_zlib_compress, |
Li Zefan | a6fa6fa | 2010-10-25 15:12:26 +0800 | [diff] [blame] | 931 | &btrfs_lzo_compress, |
Nick Terrell | 5c1aab1 | 2017-08-09 19:39:02 -0700 | [diff] [blame] | 932 | &btrfs_zstd_compress, |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 933 | }; |
| 934 | |
David Sterba | c778df1 | 2019-10-04 02:47:39 +0200 | [diff] [blame] | 935 | static 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 Sterba | 1e00235 | 2019-10-04 02:57:22 +0200 | [diff] [blame] | 951 | static 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 Sterba | d551703 | 2019-10-02 01:08:03 +0200 | [diff] [blame] | 967 | static void btrfs_init_workspace_manager(int type) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 968 | { |
David Sterba | 0cf2521 | 2019-10-04 03:09:55 +0200 | [diff] [blame] | 969 | struct workspace_manager *wsm; |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 970 | struct list_head *workspace; |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 971 | |
David Sterba | 0cf2521 | 2019-10-04 03:09:55 +0200 | [diff] [blame] | 972 | wsm = btrfs_compress_op[type]->workspace_manager; |
Dennis Zhou | 92ee5530 | 2019-02-04 15:20:03 -0500 | [diff] [blame] | 973 | 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 Sterba | f77dd0d | 2016-04-27 02:55:15 +0200 | [diff] [blame] | 977 | |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 978 | /* |
| 979 | * Preallocate one workspace for each compression type so we can |
| 980 | * guarantee forward progress in the worst case |
| 981 | */ |
David Sterba | c778df1 | 2019-10-04 02:47:39 +0200 | [diff] [blame] | 982 | workspace = alloc_workspace(type, 0); |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 983 | if (IS_ERR(workspace)) { |
| 984 | pr_warn( |
| 985 | "BTRFS: cannot preallocate compression workspace, will try later\n"); |
| 986 | } else { |
Dennis Zhou | 92ee5530 | 2019-02-04 15:20:03 -0500 | [diff] [blame] | 987 | atomic_set(&wsm->total_ws, 1); |
| 988 | wsm->free_ws = 1; |
| 989 | list_add(workspace, &wsm->idle_ws); |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 990 | } |
| 991 | } |
| 992 | |
David Sterba | 2510307 | 2019-10-02 01:08:03 +0200 | [diff] [blame] | 993 | static void btrfs_cleanup_workspace_manager(int type) |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 994 | { |
David Sterba | 2dba714 | 2019-10-04 01:40:58 +0200 | [diff] [blame] | 995 | struct workspace_manager *wsman; |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 996 | struct list_head *ws; |
| 997 | |
David Sterba | 2dba714 | 2019-10-04 01:40:58 +0200 | [diff] [blame] | 998 | wsman = btrfs_compress_op[type]->workspace_manager; |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 999 | while (!list_empty(&wsman->idle_ws)) { |
| 1000 | ws = wsman->idle_ws.next; |
| 1001 | list_del(ws); |
David Sterba | 1e00235 | 2019-10-04 02:57:22 +0200 | [diff] [blame] | 1002 | free_workspace(type, ws); |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 1003 | atomic_dec(&wsman->total_ws); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1004 | } |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1005 | } |
| 1006 | |
| 1007 | /* |
David Sterba | e721e49 | 2016-04-27 02:41:17 +0200 | [diff] [blame] | 1008 | * 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 Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1012 | */ |
David Sterba | 5907a9b | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1013 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1014 | { |
David Sterba | 5907a9b | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1015 | struct workspace_manager *wsm; |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1016 | struct list_head *workspace; |
| 1017 | int cpus = num_online_cpus(); |
David Sterba | fe30853 | 2017-05-31 17:14:56 +0200 | [diff] [blame] | 1018 | unsigned nofs_flag; |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 1019 | 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 Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1024 | |
David Sterba | 5907a9b | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1025 | wsm = btrfs_compress_op[type]->workspace_manager; |
Dennis Zhou | 92ee5530 | 2019-02-04 15:20:03 -0500 | [diff] [blame] | 1026 | 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 Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 1031 | |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1032 | again: |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1033 | spin_lock(ws_lock); |
| 1034 | if (!list_empty(idle_ws)) { |
| 1035 | workspace = idle_ws->next; |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1036 | list_del(workspace); |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1037 | (*free_ws)--; |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1038 | spin_unlock(ws_lock); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1039 | return workspace; |
| 1040 | |
| 1041 | } |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1042 | if (atomic_read(total_ws) > cpus) { |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1043 | DEFINE_WAIT(wait); |
| 1044 | |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1045 | spin_unlock(ws_lock); |
| 1046 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1047 | if (atomic_read(total_ws) > cpus && !*free_ws) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1048 | schedule(); |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1049 | finish_wait(ws_wait, &wait); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1050 | goto again; |
| 1051 | } |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1052 | atomic_inc(total_ws); |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1053 | spin_unlock(ws_lock); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1054 | |
David Sterba | fe30853 | 2017-05-31 17:14:56 +0200 | [diff] [blame] | 1055 | /* |
| 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 Sterba | c778df1 | 2019-10-04 02:47:39 +0200 | [diff] [blame] | 1061 | workspace = alloc_workspace(type, level); |
David Sterba | fe30853 | 2017-05-31 17:14:56 +0200 | [diff] [blame] | 1062 | memalloc_nofs_restore(nofs_flag); |
| 1063 | |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1064 | if (IS_ERR(workspace)) { |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1065 | atomic_dec(total_ws); |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1066 | wake_up(ws_wait); |
David Sterba | e721e49 | 2016-04-27 02:41:17 +0200 | [diff] [blame] | 1067 | |
| 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 Sterba | 52356716 | 2016-04-27 03:07:39 +0200 | [diff] [blame] | 1073 | * |
| 1074 | * To prevent silent and low-probability deadlocks (when the |
| 1075 | * initial preallocation fails), check if there are any |
| 1076 | * workspaces at all. |
David Sterba | e721e49 | 2016-04-27 02:41:17 +0200 | [diff] [blame] | 1077 | */ |
David Sterba | 52356716 | 2016-04-27 03:07:39 +0200 | [diff] [blame] | 1078 | 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 Mahoney | ab8d0fc | 2016-09-20 10:05:02 -0400 | [diff] [blame] | 1084 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
David Sterba | 52356716 | 2016-04-27 03:07:39 +0200 | [diff] [blame] | 1085 | } |
| 1086 | } |
David Sterba | e721e49 | 2016-04-27 02:41:17 +0200 | [diff] [blame] | 1087 | goto again; |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1088 | } |
| 1089 | return workspace; |
| 1090 | } |
| 1091 | |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1092 | static struct list_head *get_workspace(int type, int level) |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1093 | { |
David Sterba | 6a0d127 | 2019-10-04 02:36:16 +0200 | [diff] [blame] | 1094 | switch (type) { |
David Sterba | 5907a9b | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1095 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
David Sterba | 6a0d127 | 2019-10-04 02:36:16 +0200 | [diff] [blame] | 1096 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
David Sterba | 5907a9b | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1097 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
David Sterba | 6a0d127 | 2019-10-04 02:36:16 +0200 | [diff] [blame] | 1098 | 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 Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1106 | } |
| 1107 | |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1108 | /* |
| 1109 | * put a workspace struct back on the list or free it if we have enough |
| 1110 | * idle ones sitting around |
| 1111 | */ |
David Sterba | a3bbd2a | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1112 | void btrfs_put_workspace(int type, struct list_head *ws) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1113 | { |
David Sterba | a3bbd2a | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1114 | struct workspace_manager *wsm; |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 1115 | 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 Sterba | a3bbd2a | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1121 | wsm = btrfs_compress_op[type]->workspace_manager; |
Dennis Zhou | 92ee5530 | 2019-02-04 15:20:03 -0500 | [diff] [blame] | 1122 | 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 Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1127 | |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1128 | spin_lock(ws_lock); |
Nick Terrell | 26b28dc | 2017-06-29 10:57:26 -0700 | [diff] [blame] | 1129 | if (*free_ws <= num_online_cpus()) { |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1130 | list_add(ws, idle_ws); |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1131 | (*free_ws)++; |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1132 | spin_unlock(ws_lock); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1133 | goto wake; |
| 1134 | } |
Byongho Lee | d918764 | 2015-10-14 14:05:24 +0900 | [diff] [blame] | 1135 | spin_unlock(ws_lock); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1136 | |
David Sterba | 1e00235 | 2019-10-04 02:57:22 +0200 | [diff] [blame] | 1137 | free_workspace(type, ws); |
David Sterba | 6ac10a6 | 2016-04-27 02:15:15 +0200 | [diff] [blame] | 1138 | atomic_dec(total_ws); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1139 | wake: |
David Sterba | 093258e6 | 2018-02-26 16:15:17 +0100 | [diff] [blame] | 1140 | cond_wake_up(ws_wait); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1141 | } |
| 1142 | |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1143 | static void put_workspace(int type, struct list_head *ws) |
| 1144 | { |
David Sterba | bd3a528 | 2019-10-04 02:42:03 +0200 | [diff] [blame] | 1145 | switch (type) { |
David Sterba | a3bbd2a | 2019-10-04 02:50:28 +0200 | [diff] [blame] | 1146 | 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 Sterba | bd3a528 | 2019-10-04 02:42:03 +0200 | [diff] [blame] | 1149 | 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 Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1157 | } |
| 1158 | |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1159 | /* |
Anand Jain | adbab64 | 2020-05-11 22:37:51 -0700 | [diff] [blame] | 1160 | * Adjust @level according to the limits of the compression algorithm or |
| 1161 | * fallback to default |
| 1162 | */ |
| 1163 | static 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 Sterba | 38c3146 | 2017-02-14 19:04:07 +0100 | [diff] [blame] | 1176 | * Given an address space and start and length, compress the bytes into @pages |
| 1177 | * that are allocated on demand. |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1178 | * |
David Sterba | f51d2b5 | 2017-09-15 17:36:57 +0200 | [diff] [blame] | 1179 | * @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 Sterba | 4d3a800 | 2017-02-14 19:04:07 +0100 | [diff] [blame] | 1184 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
| 1185 | * and returns number of actually allocated pages |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1186 | * |
David Sterba | 38c3146 | 2017-02-14 19:04:07 +0100 | [diff] [blame] | 1187 | * @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 Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1189 | * ran out of room in the pages array or because we cross the |
| 1190 | * max_out threshold. |
| 1191 | * |
David Sterba | 38c3146 | 2017-02-14 19:04:07 +0100 | [diff] [blame] | 1192 | * @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 Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1194 | */ |
David Sterba | f51d2b5 | 2017-09-15 17:36:57 +0200 | [diff] [blame] | 1195 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
David Sterba | 38c3146 | 2017-02-14 19:04:07 +0100 | [diff] [blame] | 1196 | u64 start, struct page **pages, |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1197 | unsigned long *out_pages, |
| 1198 | unsigned long *total_in, |
David Sterba | e5d7490 | 2017-02-14 19:45:05 +0100 | [diff] [blame] | 1199 | unsigned long *total_out) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1200 | { |
Dennis Zhou | 1972708 | 2019-02-04 15:19:57 -0500 | [diff] [blame] | 1201 | int type = btrfs_compress_type(type_level); |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1202 | int level = btrfs_compress_level(type_level); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1203 | struct list_head *workspace; |
| 1204 | int ret; |
| 1205 | |
David Sterba | b0c1fe1 | 2019-08-09 16:49:06 +0200 | [diff] [blame] | 1206 | level = btrfs_compress_set_level(type, level); |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1207 | workspace = get_workspace(type, level); |
David Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 1208 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
| 1209 | out_pages, total_in, total_out); |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1210 | put_workspace(type, workspace); |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1211 | return ret; |
| 1212 | } |
| 1213 | |
Anand Jain | 8140dc3 | 2017-05-26 15:44:58 +0800 | [diff] [blame] | 1214 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1215 | { |
| 1216 | struct list_head *workspace; |
| 1217 | int ret; |
Anand Jain | 8140dc3 | 2017-05-26 15:44:58 +0800 | [diff] [blame] | 1218 | int type = cb->compress_type; |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1219 | |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1220 | workspace = get_workspace(type, 0); |
Su Yue | 4a9e803 | 2021-12-27 18:18:39 +0800 | [diff] [blame] | 1221 | ret = compression_decompress_bio(workspace, cb); |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1222 | put_workspace(type, workspace); |
Anand Jain | e1ddce7 | 2017-05-26 15:44:59 +0800 | [diff] [blame] | 1223 | |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1224 | 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 | */ |
| 1232 | int 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 Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1238 | workspace = get_workspace(type, 0); |
David Sterba | 1e4eb74 | 2019-10-02 00:06:15 +0200 | [diff] [blame] | 1239 | ret = compression_decompress(type, workspace, data_in, dest_page, |
| 1240 | start_byte, srclen, destlen); |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1241 | put_workspace(type, workspace); |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1242 | |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1243 | return ret; |
| 1244 | } |
| 1245 | |
Dennis Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 1246 | void __init btrfs_init_compress(void) |
| 1247 | { |
David Sterba | d551703 | 2019-10-02 01:08:03 +0200 | [diff] [blame] | 1248 | 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 Zhou | 1666eda | 2019-02-04 15:20:01 -0500 | [diff] [blame] | 1252 | } |
| 1253 | |
David Sterba | e67c718 | 2018-02-19 17:24:18 +0100 | [diff] [blame] | 1254 | void __cold btrfs_exit_compress(void) |
Li Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1255 | { |
David Sterba | 2510307 | 2019-10-02 01:08:03 +0200 | [diff] [blame] | 1256 | 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 Zefan | 261507a0 | 2010-12-17 14:21:50 +0800 | [diff] [blame] | 1260 | } |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1261 | |
| 1262 | /* |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1263 | * Copy decompressed data from working buffer to pages. |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1264 | * |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1265 | * @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 Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1271 | * |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1272 | * 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 Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1291 | */ |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1292 | int btrfs_decompress_buf2page(const char *buf, u32 buf_len, |
| 1293 | struct compressed_bio *cb, u32 decompressed) |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1294 | { |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1295 | struct bio *orig_bio = cb->orig_bio; |
| 1296 | /* Offset inside the full decompressed extent */ |
| 1297 | u32 cur_offset; |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1298 | |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1299 | 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 Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1307 | |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1308 | 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 Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1314 | |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1315 | /* Haven't reached the bvec range, exit */ |
| 1316 | if (decompressed + buf_len <= bvec_offset) |
| 1317 | return 1; |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1318 | |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1319 | 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 Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1323 | |
Christoph Hellwig | 974b1ad | 2016-11-25 09:07:46 +0100 | [diff] [blame] | 1324 | /* |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1325 | * Extra range check to ensure we didn't go beyond |
| 1326 | * @buf + @buf_len. |
Christoph Hellwig | 974b1ad | 2016-11-25 09:07:46 +0100 | [diff] [blame] | 1327 | */ |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1328 | ASSERT(copy_start - decompressed < buf_len); |
| 1329 | memcpy_to_page(bvec.bv_page, bvec.bv_offset, |
| 1330 | buf + copy_start - decompressed, copy_len); |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1331 | cur_offset += copy_len; |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1332 | |
Qu Wenruo | 1c3dc17 | 2021-07-05 10:00:58 +0800 | [diff] [blame] | 1333 | bio_advance(orig_bio, copy_len); |
| 1334 | /* Finished the bio */ |
| 1335 | if (!orig_bio->bi_iter.bi_size) |
| 1336 | return 0; |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1337 | } |
Li Zefan | 3a39c18 | 2010-11-08 15:22:19 +0800 | [diff] [blame] | 1338 | return 1; |
| 1339 | } |
Timofey Titovets | c2fcdcd | 2017-07-17 16:52:58 +0300 | [diff] [blame] | 1340 | |
Timofey Titovets | 1956243 | 2017-10-08 16:11:59 +0300 | [diff] [blame] | 1341 | /* |
| 1342 | * Shannon Entropy calculation |
| 1343 | * |
Andrea Gelmini | 52042d8 | 2018-11-28 12:05:13 +0100 | [diff] [blame] | 1344 | * Pure byte distribution analysis fails to determine compressibility of data. |
Timofey Titovets | 1956243 | 2017-10-08 16:11:59 +0300 | [diff] [blame] | 1345 | * 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 | */ |
| 1371 | static inline u32 ilog2_w(u64 n) |
| 1372 | { |
| 1373 | return ilog2(n * n * n * n); |
| 1374 | } |
| 1375 | |
| 1376 | static 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 Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1394 | #define RADIX_BASE 4U |
| 1395 | #define COUNTERS_SIZE (1U << RADIX_BASE) |
Timofey Titovets | 858177d | 2017-09-28 17:33:41 +0300 | [diff] [blame] | 1396 | |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1397 | static 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 Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1406 | /* |
| 1407 | * Use 4 bits as radix base |
Andrea Gelmini | 52042d8 | 2018-11-28 12:05:13 +0100 | [diff] [blame] | 1408 | * Use 16 u32 counters for calculating new position in buf array |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1409 | * |
| 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 Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1414 | */ |
David Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1415 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
David Sterba | 36243c9 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1416 | int num) |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1417 | { |
| 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 Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1431 | max_num = array[0].count; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1432 | for (i = 1; i < num; i++) { |
David Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1433 | buf_num = array[i].count; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1434 | 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 Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1446 | buf_num = array[i].count; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1447 | 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 Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1455 | buf_num = array[i].count; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1456 | addr = get4bits(buf_num, shift); |
| 1457 | counters[addr]--; |
| 1458 | new_addr = counters[addr]; |
David Sterba | 7add17b | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1459 | array_buf[new_addr] = array[i]; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1460 | } |
| 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 Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1473 | buf_num = array_buf[i].count; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1474 | 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 Sterba | 23ae8c6 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1482 | buf_num = array_buf[i].count; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1483 | addr = get4bits(buf_num, shift); |
| 1484 | counters[addr]--; |
| 1485 | new_addr = counters[addr]; |
David Sterba | 7add17b | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1486 | array[new_addr] = array_buf[i]; |
Timofey Titovets | 440c840 | 2017-12-04 00:30:33 +0300 | [diff] [blame] | 1487 | } |
| 1488 | |
| 1489 | shift += RADIX_BASE; |
| 1490 | } |
Timofey Titovets | 858177d | 2017-09-28 17:33:41 +0300 | [diff] [blame] | 1491 | } |
| 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 | |
| 1512 | static 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 Sterba | 36243c9 | 2017-12-12 20:35:02 +0100 | [diff] [blame] | 1520 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
Timofey Titovets | 858177d | 2017-09-28 17:33:41 +0300 | [diff] [blame] | 1521 | |
| 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 Titovets | a288e92 | 2017-09-28 17:33:40 +0300 | [diff] [blame] | 1537 | /* |
| 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 | |
| 1550 | static 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 Titovets | 1fe4f6f | 2017-09-28 17:33:39 +0300 | [diff] [blame] | 1576 | static 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 Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 1584 | static 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 Weiny | 58c1a35 | 2021-02-16 18:48:23 -0800 | [diff] [blame] | 1614 | in_data = kmap_local_page(page); |
Timofey Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 1615 | /* 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 Weiny | 58c1a35 | 2021-02-16 18:48:23 -0800 | [diff] [blame] | 1627 | kunmap_local(in_data); |
Timofey Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 1628 | put_page(page); |
| 1629 | |
| 1630 | index++; |
| 1631 | } |
| 1632 | |
| 1633 | ws->sample_size = curr_sample_pos; |
| 1634 | } |
| 1635 | |
Timofey Titovets | c2fcdcd | 2017-07-17 16:52:58 +0300 | [diff] [blame] | 1636 | /* |
| 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 | */ |
| 1651 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) |
| 1652 | { |
Dennis Zhou | 7bf4994 | 2019-02-04 15:20:04 -0500 | [diff] [blame] | 1653 | struct list_head *ws_list = get_workspace(0, 0); |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 1654 | struct heuristic_ws *ws; |
Timofey Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 1655 | u32 i; |
| 1656 | u8 byte; |
Timofey Titovets | 1956243 | 2017-10-08 16:11:59 +0300 | [diff] [blame] | 1657 | int ret = 0; |
Timofey Titovets | c2fcdcd | 2017-07-17 16:52:58 +0300 | [diff] [blame] | 1658 | |
Timofey Titovets | 4e439a0 | 2017-09-28 17:33:36 +0300 | [diff] [blame] | 1659 | ws = list_entry(ws_list, struct heuristic_ws, list); |
| 1660 | |
Timofey Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 1661 | heuristic_collect_sample(inode, start, end, ws); |
| 1662 | |
Timofey Titovets | 1fe4f6f | 2017-09-28 17:33:39 +0300 | [diff] [blame] | 1663 | if (sample_repeated_patterns(ws)) { |
| 1664 | ret = 1; |
| 1665 | goto out; |
| 1666 | } |
| 1667 | |
Timofey Titovets | a440d48 | 2017-09-28 17:33:38 +0300 | [diff] [blame] | 1668 | 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 Titovets | c2fcdcd | 2017-07-17 16:52:58 +0300 | [diff] [blame] | 1673 | } |
| 1674 | |
Timofey Titovets | a288e92 | 2017-09-28 17:33:40 +0300 | [diff] [blame] | 1675 | i = byte_set_size(ws); |
| 1676 | if (i < BYTE_SET_THRESHOLD) { |
| 1677 | ret = 2; |
| 1678 | goto out; |
| 1679 | } |
| 1680 | |
Timofey Titovets | 858177d | 2017-09-28 17:33:41 +0300 | [diff] [blame] | 1681 | 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 Titovets | 1956243 | 2017-10-08 16:11:59 +0300 | [diff] [blame] | 1692 | 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 Titovets | 1fe4f6f | 2017-09-28 17:33:39 +0300 | [diff] [blame] | 1721 | out: |
Dennis Zhou | 929f4ba | 2019-02-04 15:20:02 -0500 | [diff] [blame] | 1722 | put_workspace(0, ws_list); |
Timofey Titovets | c2fcdcd | 2017-07-17 16:52:58 +0300 | [diff] [blame] | 1723 | return ret; |
| 1724 | } |
David Sterba | f51d2b5 | 2017-09-15 17:36:57 +0200 | [diff] [blame] | 1725 | |
Dennis Zhou | d0ab62c | 2019-02-04 15:20:05 -0500 | [diff] [blame] | 1726 | /* |
| 1727 | * Convert the compression suffix (eg. after "zlib" starting with ":") to |
| 1728 | * level, unrecognized string will set the default level |
| 1729 | */ |
| 1730 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) |
David Sterba | f51d2b5 | 2017-09-15 17:36:57 +0200 | [diff] [blame] | 1731 | { |
Dennis Zhou | d0ab62c | 2019-02-04 15:20:05 -0500 | [diff] [blame] | 1732 | unsigned int level = 0; |
| 1733 | int ret; |
| 1734 | |
| 1735 | if (!type) |
David Sterba | f51d2b5 | 2017-09-15 17:36:57 +0200 | [diff] [blame] | 1736 | return 0; |
| 1737 | |
Dennis Zhou | d0ab62c | 2019-02-04 15:20:05 -0500 | [diff] [blame] | 1738 | if (str[0] == ':') { |
| 1739 | ret = kstrtouint(str + 1, 10, &level); |
| 1740 | if (ret) |
| 1741 | level = 0; |
| 1742 | } |
David Sterba | f51d2b5 | 2017-09-15 17:36:57 +0200 | [diff] [blame] | 1743 | |
David Sterba | b0c1fe1 | 2019-08-09 16:49:06 +0200 | [diff] [blame] | 1744 | level = btrfs_compress_set_level(type, level); |
| 1745 | |
| 1746 | return level; |
| 1747 | } |