1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2008 Oracle. All rights reserved.
4 *
5 * Based on jffs2 zlib code:
6 * Copyright © 2001-2007 Red Hat, Inc.
7 * Created by David Woodhouse <dwmw2@infradead.org>
8 */
9
10 #include <linux/kernel.h>
11 #include <linux/slab.h>
12 #include <linux/zlib.h>
13 #include <linux/zutil.h>
14 #include <linux/mm.h>
15 #include <linux/init.h>
16 #include <linux/err.h>
17 #include <linux/sched.h>
18 #include <linux/pagemap.h>
19 #include <linux/bio.h>
20 #include <linux/refcount.h>
21 #include "btrfs_inode.h"
22 #include "compression.h"
23 #include "fs.h"
24 #include "subpage.h"
25
26 /* workspace buffer size for s390 zlib hardware support */
27 #define ZLIB_DFLTCC_BUF_SIZE (4 * PAGE_SIZE)
28
29 struct workspace {
30 z_stream strm;
31 char *buf;
32 unsigned int buf_size;
33 struct list_head list;
34 int level;
35 };
36
zlib_get_workspace(struct btrfs_fs_info * fs_info,unsigned int level)37 struct list_head *zlib_get_workspace(struct btrfs_fs_info *fs_info, unsigned int level)
38 {
39 struct list_head *ws = btrfs_get_workspace(fs_info, BTRFS_COMPRESS_ZLIB, level);
40 struct workspace *workspace = list_entry(ws, struct workspace, list);
41
42 workspace->level = level;
43
44 return ws;
45 }
46
zlib_free_workspace(struct list_head * ws)47 void zlib_free_workspace(struct list_head *ws)
48 {
49 struct workspace *workspace = list_entry(ws, struct workspace, list);
50
51 kvfree(workspace->strm.workspace);
52 kfree(workspace->buf);
53 kfree(workspace);
54 }
55
56 /*
57 * For s390 hardware acceleration, the buffer size should be at least
58 * ZLIB_DFLTCC_BUF_SIZE to achieve the best performance.
59 *
60 * But if bs > ps we can have large enough folios that meet the s390 hardware
61 * handling.
62 */
need_special_buffer(struct btrfs_fs_info * fs_info)63 static bool need_special_buffer(struct btrfs_fs_info *fs_info)
64 {
65 if (!zlib_deflate_dfltcc_enabled())
66 return false;
67 if (btrfs_min_folio_size(fs_info) >= ZLIB_DFLTCC_BUF_SIZE)
68 return false;
69 return true;
70 }
71
zlib_alloc_workspace(struct btrfs_fs_info * fs_info,unsigned int level)72 struct list_head *zlib_alloc_workspace(struct btrfs_fs_info *fs_info, unsigned int level)
73 {
74 struct workspace *workspace;
75 int workspacesize;
76
77 workspace = kzalloc_obj(*workspace);
78 if (!workspace)
79 return ERR_PTR(-ENOMEM);
80
81 workspacesize = max(zlib_deflate_workspacesize(MAX_WBITS, MAX_MEM_LEVEL),
82 zlib_inflate_workspacesize());
83 workspace->strm.workspace = kvzalloc(workspacesize, GFP_KERNEL | __GFP_NOWARN);
84 workspace->level = level;
85 workspace->buf = NULL;
86 if (need_special_buffer(fs_info)) {
87 workspace->buf = kmalloc(ZLIB_DFLTCC_BUF_SIZE,
88 __GFP_NOMEMALLOC | __GFP_NORETRY |
89 __GFP_NOWARN | GFP_NOIO);
90 workspace->buf_size = ZLIB_DFLTCC_BUF_SIZE;
91 }
92 if (!workspace->buf) {
93 workspace->buf = kmalloc(fs_info->sectorsize, GFP_KERNEL);
94 workspace->buf_size = fs_info->sectorsize;
95 }
96 if (!workspace->strm.workspace || !workspace->buf)
97 goto fail;
98
99 INIT_LIST_HEAD(&workspace->list);
100
101 return &workspace->list;
102 fail:
103 zlib_free_workspace(&workspace->list);
104 return ERR_PTR(-ENOMEM);
105 }
106
107 /*
108 * Helper for S390x with hardware zlib compression support.
109 *
110 * That hardware acceleration requires a buffer size larger than a single page
111 * to get ideal performance, thus we need to do the memory copy rather than
112 * use the page cache directly as input buffer.
113 */
copy_data_into_buffer(struct address_space * mapping,struct workspace * workspace,u64 filepos,unsigned long length)114 static int copy_data_into_buffer(struct address_space *mapping,
115 struct workspace *workspace, u64 filepos,
116 unsigned long length)
117 {
118 u64 cur = filepos;
119
120 /* It's only for hardware accelerated zlib code. */
121 ASSERT(zlib_deflate_dfltcc_enabled());
122
123 while (cur < filepos + length) {
124 struct folio *folio;
125 void *data_in;
126 unsigned int offset;
127 unsigned long copy_length;
128 int ret;
129
130 ret = btrfs_compress_filemap_get_folio(mapping, cur, &folio);
131 if (ret < 0)
132 return ret;
133
134 offset = offset_in_folio(folio, cur);
135 copy_length = min(folio_size(folio) - offset,
136 filepos + length - cur);
137
138 data_in = kmap_local_folio(folio, offset);
139 memcpy(workspace->buf + cur - filepos, data_in, copy_length);
140 kunmap_local(data_in);
141 folio_put(folio);
142 cur += copy_length;
143 }
144 return 0;
145 }
146
zlib_compress_bio(struct list_head * ws,struct compressed_bio * cb)147 int zlib_compress_bio(struct list_head *ws, struct compressed_bio *cb)
148 {
149 struct btrfs_inode *inode = cb->bbio.inode;
150 struct btrfs_fs_info *fs_info = inode->root->fs_info;
151 struct workspace *workspace = list_entry(ws, struct workspace, list);
152 struct address_space *mapping = inode->vfs_inode.i_mapping;
153 struct bio *bio = &cb->bbio.bio;
154 u64 start = cb->start;
155 u32 len = cb->len;
156 const u32 min_folio_size = btrfs_min_folio_size(fs_info);
157 int ret;
158 char *data_in = NULL;
159 struct folio *in_folio = NULL;
160 struct folio *out_folio = NULL;
161 const u64 orig_end = start + len;
162
163 ret = zlib_deflateInit(&workspace->strm, workspace->level);
164 if (unlikely(ret != Z_OK)) {
165 btrfs_err(fs_info,
166 "zlib compression init failed, error %d root %llu inode %llu offset %llu",
167 ret, btrfs_root_id(inode->root), btrfs_ino(inode), start);
168 ret = -EIO;
169 goto out;
170 }
171
172 workspace->strm.total_in = 0;
173 workspace->strm.total_out = 0;
174
175 out_folio = btrfs_alloc_compr_folio(fs_info, GFP_NOFS);
176 if (out_folio == NULL) {
177 ret = -ENOMEM;
178 goto out;
179 }
180
181 workspace->strm.next_in = workspace->buf;
182 workspace->strm.avail_in = 0;
183 workspace->strm.next_out = folio_address(out_folio);
184 workspace->strm.avail_out = min_folio_size;
185
186 while (workspace->strm.total_in < len) {
187 /*
188 * Get next input pages and copy the contents to the workspace
189 * buffer if required.
190 */
191 if (workspace->strm.avail_in == 0) {
192 unsigned long bytes_left = len - workspace->strm.total_in;
193 unsigned int copy_length = min(bytes_left, workspace->buf_size);
194
195 /*
196 * For s390 hardware accelerated zlib, and our folio is smaller
197 * than the copy_length, we need to fill the buffer so that
198 * we can take full advantage of hardware acceleration.
199 */
200 if (need_special_buffer(fs_info)) {
201 ret = copy_data_into_buffer(mapping, workspace,
202 start, copy_length);
203 if (ret < 0)
204 goto out;
205 start += copy_length;
206 workspace->strm.next_in = workspace->buf;
207 workspace->strm.avail_in = copy_length;
208 } else {
209 unsigned int cur_len;
210
211 if (data_in) {
212 kunmap_local(data_in);
213 folio_put(in_folio);
214 data_in = NULL;
215 }
216 ret = btrfs_compress_filemap_get_folio(mapping,
217 start, &in_folio);
218 if (ret < 0)
219 goto out;
220 cur_len = btrfs_calc_input_length(in_folio, orig_end, start);
221 data_in = kmap_local_folio(in_folio,
222 offset_in_folio(in_folio, start));
223 start += cur_len;
224 workspace->strm.next_in = data_in;
225 workspace->strm.avail_in = cur_len;
226 }
227 }
228
229 ret = zlib_deflate(&workspace->strm, Z_SYNC_FLUSH);
230 if (unlikely(ret != Z_OK)) {
231 btrfs_warn(fs_info,
232 "zlib compression failed, error %d root %llu inode %llu offset %llu",
233 ret, btrfs_root_id(inode->root), btrfs_ino(inode),
234 start);
235 zlib_deflateEnd(&workspace->strm);
236 ret = -EIO;
237 goto out;
238 }
239
240 /* We're making it bigger, give up. */
241 if (workspace->strm.total_in > fs_info->sectorsize * 2 &&
242 workspace->strm.total_in < workspace->strm.total_out) {
243 ret = -E2BIG;
244 goto out;
245 }
246 if (workspace->strm.total_out >= len) {
247 ret = -E2BIG;
248 goto out;
249 }
250 /* Queue the full folio and allocate a new one. */
251 if (workspace->strm.avail_out == 0) {
252 if (!bio_add_folio(bio, out_folio, folio_size(out_folio), 0)) {
253 ret = -E2BIG;
254 goto out;
255 }
256
257 out_folio = btrfs_alloc_compr_folio(fs_info, GFP_NOFS);
258 if (out_folio == NULL) {
259 ret = -ENOMEM;
260 goto out;
261 }
262 workspace->strm.avail_out = min_folio_size;
263 workspace->strm.next_out = folio_address(out_folio);
264 }
265 /* We're all done. */
266 if (workspace->strm.total_in >= len)
267 break;
268 }
269
270 workspace->strm.avail_in = 0;
271
272 /*
273 * Call deflate with Z_FINISH flush parameter providing more output
274 * space but no more input data, until it returns with Z_STREAM_END.
275 */
276 while (ret != Z_STREAM_END) {
277 ret = zlib_deflate(&workspace->strm, Z_FINISH);
278 if (ret == Z_STREAM_END)
279 break;
280 if (unlikely(ret != Z_OK && ret != Z_BUF_ERROR)) {
281 zlib_deflateEnd(&workspace->strm);
282 ret = -EIO;
283 goto out;
284 } else if (workspace->strm.avail_out == 0) {
285 if (workspace->strm.total_out >= len) {
286 ret = -E2BIG;
287 goto out;
288 }
289 if (!bio_add_folio(bio, out_folio, folio_size(out_folio), 0)) {
290 ret = -E2BIG;
291 goto out;
292 }
293 /* Get another folio for the stream end. */
294 out_folio = btrfs_alloc_compr_folio(fs_info, GFP_NOFS);
295 if (out_folio == NULL) {
296 ret = -ENOMEM;
297 goto out;
298 }
299 workspace->strm.avail_out = min_folio_size;
300 workspace->strm.next_out = folio_address(out_folio);
301 }
302 }
303 /* Queue the remaining part of the folio. */
304 if (workspace->strm.total_out > bio->bi_iter.bi_size) {
305 const u32 cur_len = workspace->strm.total_out - bio->bi_iter.bi_size;
306
307 ASSERT(cur_len <= folio_size(out_folio));
308
309 if (!bio_add_folio(bio, out_folio, cur_len, 0)) {
310 ret = -E2BIG;
311 goto out;
312 }
313 } else {
314 /* The last folio hasn't' been utilized. */
315 btrfs_free_compr_folio(out_folio);
316 }
317 out_folio = NULL;
318 ASSERT(bio->bi_iter.bi_size == workspace->strm.total_out);
319 zlib_deflateEnd(&workspace->strm);
320
321 if (workspace->strm.total_out >= workspace->strm.total_in) {
322 ret = -E2BIG;
323 goto out;
324 }
325
326 ret = 0;
327 out:
328 if (out_folio)
329 btrfs_free_compr_folio(out_folio);
330 if (data_in) {
331 kunmap_local(data_in);
332 folio_put(in_folio);
333 }
334
335 return ret;
336 }
337
zlib_decompress_bio(struct list_head * ws,struct compressed_bio * cb)338 int zlib_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
339 {
340 struct btrfs_fs_info *fs_info = cb_to_fs_info(cb);
341 struct workspace *workspace = list_entry(ws, struct workspace, list);
342 struct folio_iter fi;
343 const u32 min_folio_size = btrfs_min_folio_size(fs_info);
344 int ret = 0, ret2;
345 int wbits = MAX_WBITS;
346 char *data_in;
347 size_t total_out = 0;
348 const size_t srclen = bio_get_size(&cb->bbio.bio);
349 unsigned long buf_start;
350
351 bio_first_folio(&fi, &cb->bbio.bio, 0);
352
353 /* We must have at least one folio here, that has the correct size. */
354 if (unlikely(!fi.folio))
355 return -EINVAL;
356 ASSERT(folio_size(fi.folio) == min_folio_size);
357
358 data_in = kmap_local_folio(fi.folio, 0);
359 workspace->strm.next_in = data_in;
360 workspace->strm.avail_in = min_t(size_t, srclen, min_folio_size);
361 workspace->strm.total_in = 0;
362
363 workspace->strm.total_out = 0;
364 workspace->strm.next_out = workspace->buf;
365 workspace->strm.avail_out = workspace->buf_size;
366
367 /* If it's deflate, and it's got no preset dictionary, then
368 we can tell zlib to skip the adler32 check. */
369 if (srclen > 2 && !(data_in[1] & PRESET_DICT) &&
370 ((data_in[0] & 0x0f) == Z_DEFLATED) &&
371 !(((data_in[0]<<8) + data_in[1]) % 31)) {
372
373 wbits = -((data_in[0] >> 4) + 8);
374 workspace->strm.next_in += 2;
375 workspace->strm.avail_in -= 2;
376 }
377
378 ret = zlib_inflateInit2(&workspace->strm, wbits);
379 if (unlikely(ret != Z_OK)) {
380 struct btrfs_inode *inode = cb->bbio.inode;
381
382 kunmap_local(data_in);
383 btrfs_err(inode->root->fs_info,
384 "zlib decompression init failed, error %d root %llu inode %llu offset %llu",
385 ret, btrfs_root_id(inode->root), btrfs_ino(inode), cb->start);
386 return -EIO;
387 }
388 while (workspace->strm.total_in < srclen) {
389 ret = zlib_inflate(&workspace->strm, Z_NO_FLUSH);
390 if (ret != Z_OK && ret != Z_STREAM_END)
391 break;
392
393 buf_start = total_out;
394 total_out = workspace->strm.total_out;
395
396 /* we didn't make progress in this inflate call, we're done */
397 if (buf_start == total_out)
398 break;
399
400 ret2 = btrfs_decompress_buf2page(workspace->buf,
401 total_out - buf_start, cb, buf_start);
402 if (ret2 == 0) {
403 ret = 0;
404 goto done;
405 }
406
407 workspace->strm.next_out = workspace->buf;
408 workspace->strm.avail_out = workspace->buf_size;
409
410 if (workspace->strm.avail_in == 0) {
411 unsigned long tmp;
412 kunmap_local(data_in);
413 bio_next_folio(&fi, &cb->bbio.bio);
414 if (!fi.folio) {
415 data_in = NULL;
416 break;
417 }
418 ASSERT(folio_size(fi.folio) == min_folio_size);
419 data_in = kmap_local_folio(fi.folio, 0);
420 workspace->strm.next_in = data_in;
421 tmp = srclen - workspace->strm.total_in;
422 workspace->strm.avail_in = min(tmp, min_folio_size);
423 }
424 }
425 if (unlikely(ret != Z_STREAM_END)) {
426 btrfs_err(cb->bbio.inode->root->fs_info,
427 "zlib decompression failed, error %d root %llu inode %llu offset %llu",
428 ret, btrfs_root_id(cb->bbio.inode->root),
429 btrfs_ino(cb->bbio.inode), cb->start);
430 ret = -EIO;
431 } else {
432 ret = 0;
433 }
434 done:
435 zlib_inflateEnd(&workspace->strm);
436 if (data_in)
437 kunmap_local(data_in);
438 return ret;
439 }
440
zlib_decompress(struct list_head * ws,const u8 * data_in,struct folio * dest_folio,unsigned long dest_pgoff,size_t srclen,size_t destlen)441 int zlib_decompress(struct list_head *ws, const u8 *data_in,
442 struct folio *dest_folio, unsigned long dest_pgoff, size_t srclen,
443 size_t destlen)
444 {
445 struct workspace *workspace = list_entry(ws, struct workspace, list);
446 int ret = 0;
447 int wbits = MAX_WBITS;
448 unsigned long to_copy;
449
450 workspace->strm.next_in = data_in;
451 workspace->strm.avail_in = srclen;
452 workspace->strm.total_in = 0;
453
454 workspace->strm.next_out = workspace->buf;
455 workspace->strm.avail_out = workspace->buf_size;
456 workspace->strm.total_out = 0;
457 /* If it's deflate, and it's got no preset dictionary, then
458 we can tell zlib to skip the adler32 check. */
459 if (srclen > 2 && !(data_in[1] & PRESET_DICT) &&
460 ((data_in[0] & 0x0f) == Z_DEFLATED) &&
461 !(((data_in[0]<<8) + data_in[1]) % 31)) {
462
463 wbits = -((data_in[0] >> 4) + 8);
464 workspace->strm.next_in += 2;
465 workspace->strm.avail_in -= 2;
466 }
467
468 ret = zlib_inflateInit2(&workspace->strm, wbits);
469 if (unlikely(ret != Z_OK)) {
470 struct btrfs_inode *inode = folio_to_inode(dest_folio);
471
472 btrfs_err(inode->root->fs_info,
473 "zlib decompression init failed, error %d root %llu inode %llu offset %llu",
474 ret, btrfs_root_id(inode->root), btrfs_ino(inode),
475 folio_pos(dest_folio));
476 return -EIO;
477 }
478
479 /*
480 * Everything (in/out buf) should be at most one sector, there should
481 * be no need to switch any input/output buffer.
482 */
483 ret = zlib_inflate(&workspace->strm, Z_FINISH);
484 to_copy = min(workspace->strm.total_out, destlen);
485 if (ret != Z_STREAM_END)
486 goto out;
487
488 memcpy_to_folio(dest_folio, dest_pgoff, workspace->buf, to_copy);
489
490 out:
491 if (unlikely(to_copy != destlen)) {
492 struct btrfs_inode *inode = folio_to_inode(dest_folio);
493
494 btrfs_err(inode->root->fs_info,
495 "zlib decompression failed, error %d root %llu inode %llu offset %llu decompressed %lu expected %zu",
496 ret, btrfs_root_id(inode->root), btrfs_ino(inode),
497 folio_pos(dest_folio), to_copy, destlen);
498 ret = -EIO;
499 } else {
500 ret = 0;
501 }
502
503 zlib_inflateEnd(&workspace->strm);
504
505 if (unlikely(to_copy < destlen))
506 folio_zero_range(dest_folio, dest_pgoff + to_copy, destlen - to_copy);
507 return ret;
508 }
509
510 const struct btrfs_compress_levels btrfs_zlib_compress = {
511 .min_level = 1,
512 .max_level = 9,
513 .default_level = BTRFS_ZLIB_DEFAULT_LEVEL,
514 };
515