Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/slab.h>
4#include "messages.h"
5#include "ctree.h"
6#include "subpage.h"
7#include "btrfs_inode.h"
8
9/*
10 * Subpage (sectorsize < PAGE_SIZE) support overview:
11 *
12 * Limitations:
13 *
14 * - Only support 64K page size for now
15 * This is to make metadata handling easier, as 64K page would ensure
16 * all nodesize would fit inside one page, thus we don't need to handle
17 * cases where a tree block crosses several pages.
18 *
19 * - Only metadata read-write for now
20 * The data read-write part is in development.
21 *
22 * - Metadata can't cross 64K page boundary
23 * btrfs-progs and kernel have done that for a while, thus only ancient
24 * filesystems could have such problem. For such case, do a graceful
25 * rejection.
26 *
27 * Special behavior:
28 *
29 * - Metadata
30 * Metadata read is fully supported.
31 * Meaning when reading one tree block will only trigger the read for the
32 * needed range, other unrelated range in the same page will not be touched.
33 *
34 * Metadata write support is partial.
35 * The writeback is still for the full page, but we will only submit
36 * the dirty extent buffers in the page.
37 *
38 * This means, if we have a metadata page like this:
39 *
40 * Page offset
41 * 0 16K 32K 48K 64K
42 * |/////////| |///////////|
43 * \- Tree block A \- Tree block B
44 *
45 * Even if we just want to writeback tree block A, we will also writeback
46 * tree block B if it's also dirty.
47 *
48 * This may cause extra metadata writeback which results more COW.
49 *
50 * Implementation:
51 *
52 * - Common
53 * Both metadata and data will use a new structure, btrfs_subpage, to
54 * record the status of each sector inside a page. This provides the extra
55 * granularity needed.
56 *
57 * - Metadata
58 * Since we have multiple tree blocks inside one page, we can't rely on page
59 * locking anymore, or we will have greatly reduced concurrency or even
60 * deadlocks (hold one tree lock while trying to lock another tree lock in
61 * the same page).
62 *
63 * Thus for metadata locking, subpage support relies on io_tree locking only.
64 * This means a slightly higher tree locking latency.
65 */
66
67#if PAGE_SIZE > SZ_4K
68bool btrfs_is_subpage(const struct btrfs_fs_info *fs_info, struct address_space *mapping)
69{
70 if (fs_info->sectorsize >= PAGE_SIZE)
71 return false;
72
73 /*
74 * Only data pages (either through DIO or compression) can have no
75 * mapping. And if page->mapping->host is data inode, it's subpage.
76 * As we have ruled our sectorsize >= PAGE_SIZE case already.
77 */
78 if (!mapping || !mapping->host || is_data_inode(BTRFS_I(mapping->host)))
79 return true;
80
81 /*
82 * Now the only remaining case is metadata, which we only go subpage
83 * routine if nodesize < PAGE_SIZE.
84 */
85 if (fs_info->nodesize < PAGE_SIZE)
86 return true;
87 return false;
88}
89#endif
90
91int btrfs_attach_subpage(const struct btrfs_fs_info *fs_info,
92 struct folio *folio, enum btrfs_subpage_type type)
93{
94 struct btrfs_subpage *subpage;
95
96 /*
97 * We have cases like a dummy extent buffer page, which is not mapped
98 * and doesn't need to be locked.
99 */
100 if (folio->mapping)
101 ASSERT(folio_test_locked(folio));
102
103 /* Either not subpage, or the folio already has private attached. */
104 if (!btrfs_is_subpage(fs_info, folio->mapping) || folio_test_private(folio))
105 return 0;
106
107 subpage = btrfs_alloc_subpage(fs_info, type);
108 if (IS_ERR(subpage))
109 return PTR_ERR(subpage);
110
111 folio_attach_private(folio, subpage);
112 return 0;
113}
114
115void btrfs_detach_subpage(const struct btrfs_fs_info *fs_info, struct folio *folio)
116{
117 struct btrfs_subpage *subpage;
118
119 /* Either not subpage, or the folio already has private attached. */
120 if (!btrfs_is_subpage(fs_info, folio->mapping) || !folio_test_private(folio))
121 return;
122
123 subpage = folio_detach_private(folio);
124 ASSERT(subpage);
125 btrfs_free_subpage(subpage);
126}
127
128struct btrfs_subpage *btrfs_alloc_subpage(const struct btrfs_fs_info *fs_info,
129 enum btrfs_subpage_type type)
130{
131 struct btrfs_subpage *ret;
132 unsigned int real_size;
133
134 ASSERT(fs_info->sectorsize < PAGE_SIZE);
135
136 real_size = struct_size(ret, bitmaps,
137 BITS_TO_LONGS(btrfs_bitmap_nr_max * fs_info->sectors_per_page));
138 ret = kzalloc(real_size, GFP_NOFS);
139 if (!ret)
140 return ERR_PTR(-ENOMEM);
141
142 spin_lock_init(&ret->lock);
143 if (type == BTRFS_SUBPAGE_METADATA)
144 atomic_set(&ret->eb_refs, 0);
145 else
146 atomic_set(&ret->nr_locked, 0);
147 return ret;
148}
149
150void btrfs_free_subpage(struct btrfs_subpage *subpage)
151{
152 kfree(subpage);
153}
154
155/*
156 * Increase the eb_refs of current subpage.
157 *
158 * This is important for eb allocation, to prevent race with last eb freeing
159 * of the same page.
160 * With the eb_refs increased before the eb inserted into radix tree,
161 * detach_extent_buffer_page() won't detach the folio private while we're still
162 * allocating the extent buffer.
163 */
164void btrfs_folio_inc_eb_refs(const struct btrfs_fs_info *fs_info, struct folio *folio)
165{
166 struct btrfs_subpage *subpage;
167
168 if (!btrfs_is_subpage(fs_info, folio->mapping))
169 return;
170
171 ASSERT(folio_test_private(folio) && folio->mapping);
172 lockdep_assert_held(&folio->mapping->i_private_lock);
173
174 subpage = folio_get_private(folio);
175 atomic_inc(&subpage->eb_refs);
176}
177
178void btrfs_folio_dec_eb_refs(const struct btrfs_fs_info *fs_info, struct folio *folio)
179{
180 struct btrfs_subpage *subpage;
181
182 if (!btrfs_is_subpage(fs_info, folio->mapping))
183 return;
184
185 ASSERT(folio_test_private(folio) && folio->mapping);
186 lockdep_assert_held(&folio->mapping->i_private_lock);
187
188 subpage = folio_get_private(folio);
189 ASSERT(atomic_read(&subpage->eb_refs));
190 atomic_dec(&subpage->eb_refs);
191}
192
193static void btrfs_subpage_assert(const struct btrfs_fs_info *fs_info,
194 struct folio *folio, u64 start, u32 len)
195{
196 /* For subpage support, the folio must be single page. */
197 ASSERT(folio_order(folio) == 0);
198
199 /* Basic checks */
200 ASSERT(folio_test_private(folio) && folio_get_private(folio));
201 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
202 IS_ALIGNED(len, fs_info->sectorsize));
203 /*
204 * The range check only works for mapped page, we can still have
205 * unmapped page like dummy extent buffer pages.
206 */
207 if (folio->mapping)
208 ASSERT(folio_pos(folio) <= start &&
209 start + len <= folio_pos(folio) + PAGE_SIZE);
210}
211
212#define subpage_calc_start_bit(fs_info, folio, name, start, len) \
213({ \
214 unsigned int __start_bit; \
215 \
216 btrfs_subpage_assert(fs_info, folio, start, len); \
217 __start_bit = offset_in_page(start) >> fs_info->sectorsize_bits; \
218 __start_bit += fs_info->sectors_per_page * btrfs_bitmap_nr_##name; \
219 __start_bit; \
220})
221
222static void btrfs_subpage_clamp_range(struct folio *folio, u64 *start, u32 *len)
223{
224 u64 orig_start = *start;
225 u32 orig_len = *len;
226
227 *start = max_t(u64, folio_pos(folio), orig_start);
228 /*
229 * For certain call sites like btrfs_drop_pages(), we may have pages
230 * beyond the target range. In that case, just set @len to 0, subpage
231 * helpers can handle @len == 0 without any problem.
232 */
233 if (folio_pos(folio) >= orig_start + orig_len)
234 *len = 0;
235 else
236 *len = min_t(u64, folio_pos(folio) + PAGE_SIZE,
237 orig_start + orig_len) - *start;
238}
239
240static bool btrfs_subpage_end_and_test_lock(const struct btrfs_fs_info *fs_info,
241 struct folio *folio, u64 start, u32 len)
242{
243 struct btrfs_subpage *subpage = folio_get_private(folio);
244 const int start_bit = subpage_calc_start_bit(fs_info, folio, locked, start, len);
245 const int nbits = (len >> fs_info->sectorsize_bits);
246 unsigned long flags;
247 unsigned int cleared = 0;
248 int bit = start_bit;
249 bool last;
250
251 btrfs_subpage_assert(fs_info, folio, start, len);
252
253 spin_lock_irqsave(&subpage->lock, flags);
254 /*
255 * We have call sites passing @lock_page into
256 * extent_clear_unlock_delalloc() for compression path.
257 *
258 * This @locked_page is locked by plain lock_page(), thus its
259 * subpage::locked is 0. Handle them in a special way.
260 */
261 if (atomic_read(&subpage->nr_locked) == 0) {
262 spin_unlock_irqrestore(&subpage->lock, flags);
263 return true;
264 }
265
266 for_each_set_bit_from(bit, subpage->bitmaps, start_bit + nbits) {
267 clear_bit(bit, subpage->bitmaps);
268 cleared++;
269 }
270 ASSERT(atomic_read(&subpage->nr_locked) >= cleared);
271 last = atomic_sub_and_test(cleared, &subpage->nr_locked);
272 spin_unlock_irqrestore(&subpage->lock, flags);
273 return last;
274}
275
276/*
277 * Handle different locked folios:
278 *
279 * - Non-subpage folio
280 * Just unlock it.
281 *
282 * - folio locked but without any subpage locked
283 * This happens either before writepage_delalloc() or the delalloc range is
284 * already handled by previous folio.
285 * We can simple unlock it.
286 *
287 * - folio locked with subpage range locked.
288 * We go through the locked sectors inside the range and clear their locked
289 * bitmap, reduce the writer lock number, and unlock the page if that's
290 * the last locked range.
291 */
292void btrfs_folio_end_lock(const struct btrfs_fs_info *fs_info,
293 struct folio *folio, u64 start, u32 len)
294{
295 struct btrfs_subpage *subpage = folio_get_private(folio);
296
297 ASSERT(folio_test_locked(folio));
298
299 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, folio->mapping)) {
300 folio_unlock(folio);
301 return;
302 }
303
304 /*
305 * For subpage case, there are two types of locked page. With or
306 * without locked number.
307 *
308 * Since we own the page lock, no one else could touch subpage::locked
309 * and we are safe to do several atomic operations without spinlock.
310 */
311 if (atomic_read(&subpage->nr_locked) == 0) {
312 /* No subpage lock, locked by plain lock_page(). */
313 folio_unlock(folio);
314 return;
315 }
316
317 btrfs_subpage_clamp_range(folio, &start, &len);
318 if (btrfs_subpage_end_and_test_lock(fs_info, folio, start, len))
319 folio_unlock(folio);
320}
321
322void btrfs_folio_end_lock_bitmap(const struct btrfs_fs_info *fs_info,
323 struct folio *folio, unsigned long bitmap)
324{
325 struct btrfs_subpage *subpage = folio_get_private(folio);
326 const int start_bit = fs_info->sectors_per_page * btrfs_bitmap_nr_locked;
327 unsigned long flags;
328 bool last = false;
329 int cleared = 0;
330 int bit;
331
332 if (!btrfs_is_subpage(fs_info, folio->mapping)) {
333 folio_unlock(folio);
334 return;
335 }
336
337 if (atomic_read(&subpage->nr_locked) == 0) {
338 /* No subpage lock, locked by plain lock_page(). */
339 folio_unlock(folio);
340 return;
341 }
342
343 spin_lock_irqsave(&subpage->lock, flags);
344 for_each_set_bit(bit, &bitmap, fs_info->sectors_per_page) {
345 if (test_and_clear_bit(bit + start_bit, subpage->bitmaps))
346 cleared++;
347 }
348 ASSERT(atomic_read(&subpage->nr_locked) >= cleared);
349 last = atomic_sub_and_test(cleared, &subpage->nr_locked);
350 spin_unlock_irqrestore(&subpage->lock, flags);
351 if (last)
352 folio_unlock(folio);
353}
354
355#define subpage_test_bitmap_all_set(fs_info, subpage, name) \
356 bitmap_test_range_all_set(subpage->bitmaps, \
357 fs_info->sectors_per_page * btrfs_bitmap_nr_##name, \
358 fs_info->sectors_per_page)
359
360#define subpage_test_bitmap_all_zero(fs_info, subpage, name) \
361 bitmap_test_range_all_zero(subpage->bitmaps, \
362 fs_info->sectors_per_page * btrfs_bitmap_nr_##name, \
363 fs_info->sectors_per_page)
364
365void btrfs_subpage_set_uptodate(const struct btrfs_fs_info *fs_info,
366 struct folio *folio, u64 start, u32 len)
367{
368 struct btrfs_subpage *subpage = folio_get_private(folio);
369 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
370 uptodate, start, len);
371 unsigned long flags;
372
373 spin_lock_irqsave(&subpage->lock, flags);
374 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
375 if (subpage_test_bitmap_all_set(fs_info, subpage, uptodate))
376 folio_mark_uptodate(folio);
377 spin_unlock_irqrestore(&subpage->lock, flags);
378}
379
380void btrfs_subpage_clear_uptodate(const struct btrfs_fs_info *fs_info,
381 struct folio *folio, u64 start, u32 len)
382{
383 struct btrfs_subpage *subpage = folio_get_private(folio);
384 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
385 uptodate, start, len);
386 unsigned long flags;
387
388 spin_lock_irqsave(&subpage->lock, flags);
389 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
390 folio_clear_uptodate(folio);
391 spin_unlock_irqrestore(&subpage->lock, flags);
392}
393
394void btrfs_subpage_set_dirty(const struct btrfs_fs_info *fs_info,
395 struct folio *folio, u64 start, u32 len)
396{
397 struct btrfs_subpage *subpage = folio_get_private(folio);
398 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
399 dirty, start, len);
400 unsigned long flags;
401
402 spin_lock_irqsave(&subpage->lock, flags);
403 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
404 spin_unlock_irqrestore(&subpage->lock, flags);
405 folio_mark_dirty(folio);
406}
407
408/*
409 * Extra clear_and_test function for subpage dirty bitmap.
410 *
411 * Return true if we're the last bits in the dirty_bitmap and clear the
412 * dirty_bitmap.
413 * Return false otherwise.
414 *
415 * NOTE: Callers should manually clear page dirty for true case, as we have
416 * extra handling for tree blocks.
417 */
418bool btrfs_subpage_clear_and_test_dirty(const struct btrfs_fs_info *fs_info,
419 struct folio *folio, u64 start, u32 len)
420{
421 struct btrfs_subpage *subpage = folio_get_private(folio);
422 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
423 dirty, start, len);
424 unsigned long flags;
425 bool last = false;
426
427 spin_lock_irqsave(&subpage->lock, flags);
428 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
429 if (subpage_test_bitmap_all_zero(fs_info, subpage, dirty))
430 last = true;
431 spin_unlock_irqrestore(&subpage->lock, flags);
432 return last;
433}
434
435void btrfs_subpage_clear_dirty(const struct btrfs_fs_info *fs_info,
436 struct folio *folio, u64 start, u32 len)
437{
438 bool last;
439
440 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, start, len);
441 if (last)
442 folio_clear_dirty_for_io(folio);
443}
444
445void btrfs_subpage_set_writeback(const struct btrfs_fs_info *fs_info,
446 struct folio *folio, u64 start, u32 len)
447{
448 struct btrfs_subpage *subpage = folio_get_private(folio);
449 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
450 writeback, start, len);
451 unsigned long flags;
452
453 spin_lock_irqsave(&subpage->lock, flags);
454 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
455 if (!folio_test_writeback(folio))
456 folio_start_writeback(folio);
457 spin_unlock_irqrestore(&subpage->lock, flags);
458}
459
460void btrfs_subpage_clear_writeback(const struct btrfs_fs_info *fs_info,
461 struct folio *folio, u64 start, u32 len)
462{
463 struct btrfs_subpage *subpage = folio_get_private(folio);
464 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
465 writeback, start, len);
466 unsigned long flags;
467
468 spin_lock_irqsave(&subpage->lock, flags);
469 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
470 if (subpage_test_bitmap_all_zero(fs_info, subpage, writeback)) {
471 ASSERT(folio_test_writeback(folio));
472 folio_end_writeback(folio);
473 }
474 spin_unlock_irqrestore(&subpage->lock, flags);
475}
476
477void btrfs_subpage_set_ordered(const struct btrfs_fs_info *fs_info,
478 struct folio *folio, u64 start, u32 len)
479{
480 struct btrfs_subpage *subpage = folio_get_private(folio);
481 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
482 ordered, start, len);
483 unsigned long flags;
484
485 spin_lock_irqsave(&subpage->lock, flags);
486 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
487 folio_set_ordered(folio);
488 spin_unlock_irqrestore(&subpage->lock, flags);
489}
490
491void btrfs_subpage_clear_ordered(const struct btrfs_fs_info *fs_info,
492 struct folio *folio, u64 start, u32 len)
493{
494 struct btrfs_subpage *subpage = folio_get_private(folio);
495 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
496 ordered, start, len);
497 unsigned long flags;
498
499 spin_lock_irqsave(&subpage->lock, flags);
500 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
501 if (subpage_test_bitmap_all_zero(fs_info, subpage, ordered))
502 folio_clear_ordered(folio);
503 spin_unlock_irqrestore(&subpage->lock, flags);
504}
505
506void btrfs_subpage_set_checked(const struct btrfs_fs_info *fs_info,
507 struct folio *folio, u64 start, u32 len)
508{
509 struct btrfs_subpage *subpage = folio_get_private(folio);
510 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
511 checked, start, len);
512 unsigned long flags;
513
514 spin_lock_irqsave(&subpage->lock, flags);
515 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
516 if (subpage_test_bitmap_all_set(fs_info, subpage, checked))
517 folio_set_checked(folio);
518 spin_unlock_irqrestore(&subpage->lock, flags);
519}
520
521void btrfs_subpage_clear_checked(const struct btrfs_fs_info *fs_info,
522 struct folio *folio, u64 start, u32 len)
523{
524 struct btrfs_subpage *subpage = folio_get_private(folio);
525 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
526 checked, start, len);
527 unsigned long flags;
528
529 spin_lock_irqsave(&subpage->lock, flags);
530 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
531 folio_clear_checked(folio);
532 spin_unlock_irqrestore(&subpage->lock, flags);
533}
534
535/*
536 * Unlike set/clear which is dependent on each page status, for test all bits
537 * are tested in the same way.
538 */
539#define IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(name) \
540bool btrfs_subpage_test_##name(const struct btrfs_fs_info *fs_info, \
541 struct folio *folio, u64 start, u32 len) \
542{ \
543 struct btrfs_subpage *subpage = folio_get_private(folio); \
544 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio, \
545 name, start, len); \
546 unsigned long flags; \
547 bool ret; \
548 \
549 spin_lock_irqsave(&subpage->lock, flags); \
550 ret = bitmap_test_range_all_set(subpage->bitmaps, start_bit, \
551 len >> fs_info->sectorsize_bits); \
552 spin_unlock_irqrestore(&subpage->lock, flags); \
553 return ret; \
554}
555IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(uptodate);
556IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(dirty);
557IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(writeback);
558IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(ordered);
559IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(checked);
560
561/*
562 * Note that, in selftests (extent-io-tests), we can have empty fs_info passed
563 * in. We only test sectorsize == PAGE_SIZE cases so far, thus we can fall
564 * back to regular sectorsize branch.
565 */
566#define IMPLEMENT_BTRFS_PAGE_OPS(name, folio_set_func, \
567 folio_clear_func, folio_test_func) \
568void btrfs_folio_set_##name(const struct btrfs_fs_info *fs_info, \
569 struct folio *folio, u64 start, u32 len) \
570{ \
571 if (unlikely(!fs_info) || \
572 !btrfs_is_subpage(fs_info, folio->mapping)) { \
573 folio_set_func(folio); \
574 return; \
575 } \
576 btrfs_subpage_set_##name(fs_info, folio, start, len); \
577} \
578void btrfs_folio_clear_##name(const struct btrfs_fs_info *fs_info, \
579 struct folio *folio, u64 start, u32 len) \
580{ \
581 if (unlikely(!fs_info) || \
582 !btrfs_is_subpage(fs_info, folio->mapping)) { \
583 folio_clear_func(folio); \
584 return; \
585 } \
586 btrfs_subpage_clear_##name(fs_info, folio, start, len); \
587} \
588bool btrfs_folio_test_##name(const struct btrfs_fs_info *fs_info, \
589 struct folio *folio, u64 start, u32 len) \
590{ \
591 if (unlikely(!fs_info) || \
592 !btrfs_is_subpage(fs_info, folio->mapping)) \
593 return folio_test_func(folio); \
594 return btrfs_subpage_test_##name(fs_info, folio, start, len); \
595} \
596void btrfs_folio_clamp_set_##name(const struct btrfs_fs_info *fs_info, \
597 struct folio *folio, u64 start, u32 len) \
598{ \
599 if (unlikely(!fs_info) || \
600 !btrfs_is_subpage(fs_info, folio->mapping)) { \
601 folio_set_func(folio); \
602 return; \
603 } \
604 btrfs_subpage_clamp_range(folio, &start, &len); \
605 btrfs_subpage_set_##name(fs_info, folio, start, len); \
606} \
607void btrfs_folio_clamp_clear_##name(const struct btrfs_fs_info *fs_info, \
608 struct folio *folio, u64 start, u32 len) \
609{ \
610 if (unlikely(!fs_info) || \
611 !btrfs_is_subpage(fs_info, folio->mapping)) { \
612 folio_clear_func(folio); \
613 return; \
614 } \
615 btrfs_subpage_clamp_range(folio, &start, &len); \
616 btrfs_subpage_clear_##name(fs_info, folio, start, len); \
617} \
618bool btrfs_folio_clamp_test_##name(const struct btrfs_fs_info *fs_info, \
619 struct folio *folio, u64 start, u32 len) \
620{ \
621 if (unlikely(!fs_info) || \
622 !btrfs_is_subpage(fs_info, folio->mapping)) \
623 return folio_test_func(folio); \
624 btrfs_subpage_clamp_range(folio, &start, &len); \
625 return btrfs_subpage_test_##name(fs_info, folio, start, len); \
626}
627IMPLEMENT_BTRFS_PAGE_OPS(uptodate, folio_mark_uptodate, folio_clear_uptodate,
628 folio_test_uptodate);
629IMPLEMENT_BTRFS_PAGE_OPS(dirty, folio_mark_dirty, folio_clear_dirty_for_io,
630 folio_test_dirty);
631IMPLEMENT_BTRFS_PAGE_OPS(writeback, folio_start_writeback, folio_end_writeback,
632 folio_test_writeback);
633IMPLEMENT_BTRFS_PAGE_OPS(ordered, folio_set_ordered, folio_clear_ordered,
634 folio_test_ordered);
635IMPLEMENT_BTRFS_PAGE_OPS(checked, folio_set_checked, folio_clear_checked,
636 folio_test_checked);
637
638/*
639 * Make sure not only the page dirty bit is cleared, but also subpage dirty bit
640 * is cleared.
641 */
642void btrfs_folio_assert_not_dirty(const struct btrfs_fs_info *fs_info,
643 struct folio *folio, u64 start, u32 len)
644{
645 struct btrfs_subpage *subpage;
646 unsigned int start_bit;
647 unsigned int nbits;
648 unsigned long flags;
649
650 if (!IS_ENABLED(CONFIG_BTRFS_ASSERT))
651 return;
652
653 if (!btrfs_is_subpage(fs_info, folio->mapping)) {
654 ASSERT(!folio_test_dirty(folio));
655 return;
656 }
657
658 start_bit = subpage_calc_start_bit(fs_info, folio, dirty, start, len);
659 nbits = len >> fs_info->sectorsize_bits;
660 subpage = folio_get_private(folio);
661 ASSERT(subpage);
662 spin_lock_irqsave(&subpage->lock, flags);
663 ASSERT(bitmap_test_range_all_zero(subpage->bitmaps, start_bit, nbits));
664 spin_unlock_irqrestore(&subpage->lock, flags);
665}
666
667/*
668 * This is for folio already locked by plain lock_page()/folio_lock(), which
669 * doesn't have any subpage awareness.
670 *
671 * This populates the involved subpage ranges so that subpage helpers can
672 * properly unlock them.
673 */
674void btrfs_folio_set_lock(const struct btrfs_fs_info *fs_info,
675 struct folio *folio, u64 start, u32 len)
676{
677 struct btrfs_subpage *subpage;
678 unsigned long flags;
679 unsigned int start_bit;
680 unsigned int nbits;
681 int ret;
682
683 ASSERT(folio_test_locked(folio));
684 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, folio->mapping))
685 return;
686
687 subpage = folio_get_private(folio);
688 start_bit = subpage_calc_start_bit(fs_info, folio, locked, start, len);
689 nbits = len >> fs_info->sectorsize_bits;
690 spin_lock_irqsave(&subpage->lock, flags);
691 /* Target range should not yet be locked. */
692 ASSERT(bitmap_test_range_all_zero(subpage->bitmaps, start_bit, nbits));
693 bitmap_set(subpage->bitmaps, start_bit, nbits);
694 ret = atomic_add_return(nbits, &subpage->nr_locked);
695 ASSERT(ret <= fs_info->sectors_per_page);
696 spin_unlock_irqrestore(&subpage->lock, flags);
697}
698
699#define GET_SUBPAGE_BITMAP(subpage, fs_info, name, dst) \
700{ \
701 const int sectors_per_page = fs_info->sectors_per_page; \
702 \
703 ASSERT(sectors_per_page < BITS_PER_LONG); \
704 *dst = bitmap_read(subpage->bitmaps, \
705 sectors_per_page * btrfs_bitmap_nr_##name, \
706 sectors_per_page); \
707}
708
709void __cold btrfs_subpage_dump_bitmap(const struct btrfs_fs_info *fs_info,
710 struct folio *folio, u64 start, u32 len)
711{
712 struct btrfs_subpage *subpage;
713 const u32 sectors_per_page = fs_info->sectors_per_page;
714 unsigned long uptodate_bitmap;
715 unsigned long dirty_bitmap;
716 unsigned long writeback_bitmap;
717 unsigned long ordered_bitmap;
718 unsigned long checked_bitmap;
719 unsigned long locked_bitmap;
720 unsigned long flags;
721
722 ASSERT(folio_test_private(folio) && folio_get_private(folio));
723 ASSERT(sectors_per_page > 1);
724 subpage = folio_get_private(folio);
725
726 spin_lock_irqsave(&subpage->lock, flags);
727 GET_SUBPAGE_BITMAP(subpage, fs_info, uptodate, &uptodate_bitmap);
728 GET_SUBPAGE_BITMAP(subpage, fs_info, dirty, &dirty_bitmap);
729 GET_SUBPAGE_BITMAP(subpage, fs_info, writeback, &writeback_bitmap);
730 GET_SUBPAGE_BITMAP(subpage, fs_info, ordered, &ordered_bitmap);
731 GET_SUBPAGE_BITMAP(subpage, fs_info, checked, &checked_bitmap);
732 GET_SUBPAGE_BITMAP(subpage, fs_info, locked, &locked_bitmap);
733 spin_unlock_irqrestore(&subpage->lock, flags);
734
735 dump_page(folio_page(folio, 0), "btrfs subpage dump");
736 btrfs_warn(fs_info,
737"start=%llu len=%u page=%llu, bitmaps uptodate=%*pbl dirty=%*pbl locked=%*pbl writeback=%*pbl ordered=%*pbl checked=%*pbl",
738 start, len, folio_pos(folio),
739 sectors_per_page, &uptodate_bitmap,
740 sectors_per_page, &dirty_bitmap,
741 sectors_per_page, &locked_bitmap,
742 sectors_per_page, &writeback_bitmap,
743 sectors_per_page, &ordered_bitmap,
744 sectors_per_page, &checked_bitmap);
745}
746
747void btrfs_get_subpage_dirty_bitmap(struct btrfs_fs_info *fs_info,
748 struct folio *folio,
749 unsigned long *ret_bitmap)
750{
751 struct btrfs_subpage *subpage;
752 unsigned long flags;
753
754 ASSERT(folio_test_private(folio) && folio_get_private(folio));
755 ASSERT(fs_info->sectors_per_page > 1);
756 subpage = folio_get_private(folio);
757
758 spin_lock_irqsave(&subpage->lock, flags);
759 GET_SUBPAGE_BITMAP(subpage, fs_info, dirty, ret_bitmap);
760 spin_unlock_irqrestore(&subpage->lock, flags);
761}
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/slab.h>
4#include "messages.h"
5#include "ctree.h"
6#include "subpage.h"
7#include "btrfs_inode.h"
8
9/*
10 * Subpage (sectorsize < PAGE_SIZE) support overview:
11 *
12 * Limitations:
13 *
14 * - Only support 64K page size for now
15 * This is to make metadata handling easier, as 64K page would ensure
16 * all nodesize would fit inside one page, thus we don't need to handle
17 * cases where a tree block crosses several pages.
18 *
19 * - Only metadata read-write for now
20 * The data read-write part is in development.
21 *
22 * - Metadata can't cross 64K page boundary
23 * btrfs-progs and kernel have done that for a while, thus only ancient
24 * filesystems could have such problem. For such case, do a graceful
25 * rejection.
26 *
27 * Special behavior:
28 *
29 * - Metadata
30 * Metadata read is fully supported.
31 * Meaning when reading one tree block will only trigger the read for the
32 * needed range, other unrelated range in the same page will not be touched.
33 *
34 * Metadata write support is partial.
35 * The writeback is still for the full page, but we will only submit
36 * the dirty extent buffers in the page.
37 *
38 * This means, if we have a metadata page like this:
39 *
40 * Page offset
41 * 0 16K 32K 48K 64K
42 * |/////////| |///////////|
43 * \- Tree block A \- Tree block B
44 *
45 * Even if we just want to writeback tree block A, we will also writeback
46 * tree block B if it's also dirty.
47 *
48 * This may cause extra metadata writeback which results more COW.
49 *
50 * Implementation:
51 *
52 * - Common
53 * Both metadata and data will use a new structure, btrfs_subpage, to
54 * record the status of each sector inside a page. This provides the extra
55 * granularity needed.
56 *
57 * - Metadata
58 * Since we have multiple tree blocks inside one page, we can't rely on page
59 * locking anymore, or we will have greatly reduced concurrency or even
60 * deadlocks (hold one tree lock while trying to lock another tree lock in
61 * the same page).
62 *
63 * Thus for metadata locking, subpage support relies on io_tree locking only.
64 * This means a slightly higher tree locking latency.
65 */
66
67bool btrfs_is_subpage(const struct btrfs_fs_info *fs_info, struct address_space *mapping)
68{
69 if (fs_info->sectorsize >= PAGE_SIZE)
70 return false;
71
72 /*
73 * Only data pages (either through DIO or compression) can have no
74 * mapping. And if page->mapping->host is data inode, it's subpage.
75 * As we have ruled our sectorsize >= PAGE_SIZE case already.
76 */
77 if (!mapping || !mapping->host || is_data_inode(mapping->host))
78 return true;
79
80 /*
81 * Now the only remaining case is metadata, which we only go subpage
82 * routine if nodesize < PAGE_SIZE.
83 */
84 if (fs_info->nodesize < PAGE_SIZE)
85 return true;
86 return false;
87}
88
89void btrfs_init_subpage_info(struct btrfs_subpage_info *subpage_info, u32 sectorsize)
90{
91 unsigned int cur = 0;
92 unsigned int nr_bits;
93
94 ASSERT(IS_ALIGNED(PAGE_SIZE, sectorsize));
95
96 nr_bits = PAGE_SIZE / sectorsize;
97 subpage_info->bitmap_nr_bits = nr_bits;
98
99 subpage_info->uptodate_offset = cur;
100 cur += nr_bits;
101
102 subpage_info->dirty_offset = cur;
103 cur += nr_bits;
104
105 subpage_info->writeback_offset = cur;
106 cur += nr_bits;
107
108 subpage_info->ordered_offset = cur;
109 cur += nr_bits;
110
111 subpage_info->checked_offset = cur;
112 cur += nr_bits;
113
114 subpage_info->locked_offset = cur;
115 cur += nr_bits;
116
117 subpage_info->total_nr_bits = cur;
118}
119
120int btrfs_attach_subpage(const struct btrfs_fs_info *fs_info,
121 struct folio *folio, enum btrfs_subpage_type type)
122{
123 struct btrfs_subpage *subpage;
124
125 /*
126 * We have cases like a dummy extent buffer page, which is not mapped
127 * and doesn't need to be locked.
128 */
129 if (folio->mapping)
130 ASSERT(folio_test_locked(folio));
131
132 /* Either not subpage, or the folio already has private attached. */
133 if (!btrfs_is_subpage(fs_info, folio->mapping) || folio_test_private(folio))
134 return 0;
135
136 subpage = btrfs_alloc_subpage(fs_info, type);
137 if (IS_ERR(subpage))
138 return PTR_ERR(subpage);
139
140 folio_attach_private(folio, subpage);
141 return 0;
142}
143
144void btrfs_detach_subpage(const struct btrfs_fs_info *fs_info, struct folio *folio)
145{
146 struct btrfs_subpage *subpage;
147
148 /* Either not subpage, or the folio already has private attached. */
149 if (!btrfs_is_subpage(fs_info, folio->mapping) || !folio_test_private(folio))
150 return;
151
152 subpage = folio_detach_private(folio);
153 ASSERT(subpage);
154 btrfs_free_subpage(subpage);
155}
156
157struct btrfs_subpage *btrfs_alloc_subpage(const struct btrfs_fs_info *fs_info,
158 enum btrfs_subpage_type type)
159{
160 struct btrfs_subpage *ret;
161 unsigned int real_size;
162
163 ASSERT(fs_info->sectorsize < PAGE_SIZE);
164
165 real_size = struct_size(ret, bitmaps,
166 BITS_TO_LONGS(fs_info->subpage_info->total_nr_bits));
167 ret = kzalloc(real_size, GFP_NOFS);
168 if (!ret)
169 return ERR_PTR(-ENOMEM);
170
171 spin_lock_init(&ret->lock);
172 if (type == BTRFS_SUBPAGE_METADATA) {
173 atomic_set(&ret->eb_refs, 0);
174 } else {
175 atomic_set(&ret->readers, 0);
176 atomic_set(&ret->writers, 0);
177 }
178 return ret;
179}
180
181void btrfs_free_subpage(struct btrfs_subpage *subpage)
182{
183 kfree(subpage);
184}
185
186/*
187 * Increase the eb_refs of current subpage.
188 *
189 * This is important for eb allocation, to prevent race with last eb freeing
190 * of the same page.
191 * With the eb_refs increased before the eb inserted into radix tree,
192 * detach_extent_buffer_page() won't detach the folio private while we're still
193 * allocating the extent buffer.
194 */
195void btrfs_folio_inc_eb_refs(const struct btrfs_fs_info *fs_info, struct folio *folio)
196{
197 struct btrfs_subpage *subpage;
198
199 if (!btrfs_is_subpage(fs_info, folio->mapping))
200 return;
201
202 ASSERT(folio_test_private(folio) && folio->mapping);
203 lockdep_assert_held(&folio->mapping->i_private_lock);
204
205 subpage = folio_get_private(folio);
206 atomic_inc(&subpage->eb_refs);
207}
208
209void btrfs_folio_dec_eb_refs(const struct btrfs_fs_info *fs_info, struct folio *folio)
210{
211 struct btrfs_subpage *subpage;
212
213 if (!btrfs_is_subpage(fs_info, folio->mapping))
214 return;
215
216 ASSERT(folio_test_private(folio) && folio->mapping);
217 lockdep_assert_held(&folio->mapping->i_private_lock);
218
219 subpage = folio_get_private(folio);
220 ASSERT(atomic_read(&subpage->eb_refs));
221 atomic_dec(&subpage->eb_refs);
222}
223
224static void btrfs_subpage_assert(const struct btrfs_fs_info *fs_info,
225 struct folio *folio, u64 start, u32 len)
226{
227 /* For subpage support, the folio must be single page. */
228 ASSERT(folio_order(folio) == 0);
229
230 /* Basic checks */
231 ASSERT(folio_test_private(folio) && folio_get_private(folio));
232 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
233 IS_ALIGNED(len, fs_info->sectorsize));
234 /*
235 * The range check only works for mapped page, we can still have
236 * unmapped page like dummy extent buffer pages.
237 */
238 if (folio->mapping)
239 ASSERT(folio_pos(folio) <= start &&
240 start + len <= folio_pos(folio) + PAGE_SIZE);
241}
242
243#define subpage_calc_start_bit(fs_info, folio, name, start, len) \
244({ \
245 unsigned int start_bit; \
246 \
247 btrfs_subpage_assert(fs_info, folio, start, len); \
248 start_bit = offset_in_page(start) >> fs_info->sectorsize_bits; \
249 start_bit += fs_info->subpage_info->name##_offset; \
250 start_bit; \
251})
252
253void btrfs_subpage_start_reader(const struct btrfs_fs_info *fs_info,
254 struct folio *folio, u64 start, u32 len)
255{
256 struct btrfs_subpage *subpage = folio_get_private(folio);
257 const int start_bit = subpage_calc_start_bit(fs_info, folio, locked, start, len);
258 const int nbits = len >> fs_info->sectorsize_bits;
259 unsigned long flags;
260
261
262 btrfs_subpage_assert(fs_info, folio, start, len);
263
264 spin_lock_irqsave(&subpage->lock, flags);
265 /*
266 * Even though it's just for reading the page, no one should have
267 * locked the subpage range.
268 */
269 ASSERT(bitmap_test_range_all_zero(subpage->bitmaps, start_bit, nbits));
270 bitmap_set(subpage->bitmaps, start_bit, nbits);
271 atomic_add(nbits, &subpage->readers);
272 spin_unlock_irqrestore(&subpage->lock, flags);
273}
274
275void btrfs_subpage_end_reader(const struct btrfs_fs_info *fs_info,
276 struct folio *folio, u64 start, u32 len)
277{
278 struct btrfs_subpage *subpage = folio_get_private(folio);
279 const int start_bit = subpage_calc_start_bit(fs_info, folio, locked, start, len);
280 const int nbits = len >> fs_info->sectorsize_bits;
281 unsigned long flags;
282 bool is_data;
283 bool last;
284
285 btrfs_subpage_assert(fs_info, folio, start, len);
286 is_data = is_data_inode(folio->mapping->host);
287
288 spin_lock_irqsave(&subpage->lock, flags);
289
290 /* The range should have already been locked. */
291 ASSERT(bitmap_test_range_all_set(subpage->bitmaps, start_bit, nbits));
292 ASSERT(atomic_read(&subpage->readers) >= nbits);
293
294 bitmap_clear(subpage->bitmaps, start_bit, nbits);
295 last = atomic_sub_and_test(nbits, &subpage->readers);
296
297 /*
298 * For data we need to unlock the page if the last read has finished.
299 *
300 * And please don't replace @last with atomic_sub_and_test() call
301 * inside if () condition.
302 * As we want the atomic_sub_and_test() to be always executed.
303 */
304 if (is_data && last)
305 folio_unlock(folio);
306 spin_unlock_irqrestore(&subpage->lock, flags);
307}
308
309static void btrfs_subpage_clamp_range(struct folio *folio, u64 *start, u32 *len)
310{
311 u64 orig_start = *start;
312 u32 orig_len = *len;
313
314 *start = max_t(u64, folio_pos(folio), orig_start);
315 /*
316 * For certain call sites like btrfs_drop_pages(), we may have pages
317 * beyond the target range. In that case, just set @len to 0, subpage
318 * helpers can handle @len == 0 without any problem.
319 */
320 if (folio_pos(folio) >= orig_start + orig_len)
321 *len = 0;
322 else
323 *len = min_t(u64, folio_pos(folio) + PAGE_SIZE,
324 orig_start + orig_len) - *start;
325}
326
327static void btrfs_subpage_start_writer(const struct btrfs_fs_info *fs_info,
328 struct folio *folio, u64 start, u32 len)
329{
330 struct btrfs_subpage *subpage = folio_get_private(folio);
331 const int start_bit = subpage_calc_start_bit(fs_info, folio, locked, start, len);
332 const int nbits = (len >> fs_info->sectorsize_bits);
333 unsigned long flags;
334 int ret;
335
336 btrfs_subpage_assert(fs_info, folio, start, len);
337
338 spin_lock_irqsave(&subpage->lock, flags);
339 ASSERT(atomic_read(&subpage->readers) == 0);
340 ASSERT(bitmap_test_range_all_zero(subpage->bitmaps, start_bit, nbits));
341 bitmap_set(subpage->bitmaps, start_bit, nbits);
342 ret = atomic_add_return(nbits, &subpage->writers);
343 ASSERT(ret == nbits);
344 spin_unlock_irqrestore(&subpage->lock, flags);
345}
346
347static bool btrfs_subpage_end_and_test_writer(const struct btrfs_fs_info *fs_info,
348 struct folio *folio, u64 start, u32 len)
349{
350 struct btrfs_subpage *subpage = folio_get_private(folio);
351 const int start_bit = subpage_calc_start_bit(fs_info, folio, locked, start, len);
352 const int nbits = (len >> fs_info->sectorsize_bits);
353 unsigned long flags;
354 bool last;
355
356 btrfs_subpage_assert(fs_info, folio, start, len);
357
358 spin_lock_irqsave(&subpage->lock, flags);
359 /*
360 * We have call sites passing @lock_page into
361 * extent_clear_unlock_delalloc() for compression path.
362 *
363 * This @locked_page is locked by plain lock_page(), thus its
364 * subpage::writers is 0. Handle them in a special way.
365 */
366 if (atomic_read(&subpage->writers) == 0) {
367 spin_unlock_irqrestore(&subpage->lock, flags);
368 return true;
369 }
370
371 ASSERT(atomic_read(&subpage->writers) >= nbits);
372 /* The target range should have been locked. */
373 ASSERT(bitmap_test_range_all_set(subpage->bitmaps, start_bit, nbits));
374 bitmap_clear(subpage->bitmaps, start_bit, nbits);
375 last = atomic_sub_and_test(nbits, &subpage->writers);
376 spin_unlock_irqrestore(&subpage->lock, flags);
377 return last;
378}
379
380/*
381 * Lock a folio for delalloc page writeback.
382 *
383 * Return -EAGAIN if the page is not properly initialized.
384 * Return 0 with the page locked, and writer counter updated.
385 *
386 * Even with 0 returned, the page still need extra check to make sure
387 * it's really the correct page, as the caller is using
388 * filemap_get_folios_contig(), which can race with page invalidating.
389 */
390int btrfs_folio_start_writer_lock(const struct btrfs_fs_info *fs_info,
391 struct folio *folio, u64 start, u32 len)
392{
393 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, folio->mapping)) {
394 folio_lock(folio);
395 return 0;
396 }
397 folio_lock(folio);
398 if (!folio_test_private(folio) || !folio_get_private(folio)) {
399 folio_unlock(folio);
400 return -EAGAIN;
401 }
402 btrfs_subpage_clamp_range(folio, &start, &len);
403 btrfs_subpage_start_writer(fs_info, folio, start, len);
404 return 0;
405}
406
407void btrfs_folio_end_writer_lock(const struct btrfs_fs_info *fs_info,
408 struct folio *folio, u64 start, u32 len)
409{
410 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, folio->mapping)) {
411 folio_unlock(folio);
412 return;
413 }
414 btrfs_subpage_clamp_range(folio, &start, &len);
415 if (btrfs_subpage_end_and_test_writer(fs_info, folio, start, len))
416 folio_unlock(folio);
417}
418
419#define subpage_test_bitmap_all_set(fs_info, subpage, name) \
420 bitmap_test_range_all_set(subpage->bitmaps, \
421 fs_info->subpage_info->name##_offset, \
422 fs_info->subpage_info->bitmap_nr_bits)
423
424#define subpage_test_bitmap_all_zero(fs_info, subpage, name) \
425 bitmap_test_range_all_zero(subpage->bitmaps, \
426 fs_info->subpage_info->name##_offset, \
427 fs_info->subpage_info->bitmap_nr_bits)
428
429void btrfs_subpage_set_uptodate(const struct btrfs_fs_info *fs_info,
430 struct folio *folio, u64 start, u32 len)
431{
432 struct btrfs_subpage *subpage = folio_get_private(folio);
433 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
434 uptodate, start, len);
435 unsigned long flags;
436
437 spin_lock_irqsave(&subpage->lock, flags);
438 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
439 if (subpage_test_bitmap_all_set(fs_info, subpage, uptodate))
440 folio_mark_uptodate(folio);
441 spin_unlock_irqrestore(&subpage->lock, flags);
442}
443
444void btrfs_subpage_clear_uptodate(const struct btrfs_fs_info *fs_info,
445 struct folio *folio, u64 start, u32 len)
446{
447 struct btrfs_subpage *subpage = folio_get_private(folio);
448 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
449 uptodate, start, len);
450 unsigned long flags;
451
452 spin_lock_irqsave(&subpage->lock, flags);
453 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
454 folio_clear_uptodate(folio);
455 spin_unlock_irqrestore(&subpage->lock, flags);
456}
457
458void btrfs_subpage_set_dirty(const struct btrfs_fs_info *fs_info,
459 struct folio *folio, u64 start, u32 len)
460{
461 struct btrfs_subpage *subpage = folio_get_private(folio);
462 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
463 dirty, start, len);
464 unsigned long flags;
465
466 spin_lock_irqsave(&subpage->lock, flags);
467 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
468 spin_unlock_irqrestore(&subpage->lock, flags);
469 folio_mark_dirty(folio);
470}
471
472/*
473 * Extra clear_and_test function for subpage dirty bitmap.
474 *
475 * Return true if we're the last bits in the dirty_bitmap and clear the
476 * dirty_bitmap.
477 * Return false otherwise.
478 *
479 * NOTE: Callers should manually clear page dirty for true case, as we have
480 * extra handling for tree blocks.
481 */
482bool btrfs_subpage_clear_and_test_dirty(const struct btrfs_fs_info *fs_info,
483 struct folio *folio, u64 start, u32 len)
484{
485 struct btrfs_subpage *subpage = folio_get_private(folio);
486 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
487 dirty, start, len);
488 unsigned long flags;
489 bool last = false;
490
491 spin_lock_irqsave(&subpage->lock, flags);
492 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
493 if (subpage_test_bitmap_all_zero(fs_info, subpage, dirty))
494 last = true;
495 spin_unlock_irqrestore(&subpage->lock, flags);
496 return last;
497}
498
499void btrfs_subpage_clear_dirty(const struct btrfs_fs_info *fs_info,
500 struct folio *folio, u64 start, u32 len)
501{
502 bool last;
503
504 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, start, len);
505 if (last)
506 folio_clear_dirty_for_io(folio);
507}
508
509void btrfs_subpage_set_writeback(const struct btrfs_fs_info *fs_info,
510 struct folio *folio, u64 start, u32 len)
511{
512 struct btrfs_subpage *subpage = folio_get_private(folio);
513 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
514 writeback, start, len);
515 unsigned long flags;
516
517 spin_lock_irqsave(&subpage->lock, flags);
518 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
519 if (!folio_test_writeback(folio))
520 folio_start_writeback(folio);
521 spin_unlock_irqrestore(&subpage->lock, flags);
522}
523
524void btrfs_subpage_clear_writeback(const struct btrfs_fs_info *fs_info,
525 struct folio *folio, u64 start, u32 len)
526{
527 struct btrfs_subpage *subpage = folio_get_private(folio);
528 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
529 writeback, start, len);
530 unsigned long flags;
531
532 spin_lock_irqsave(&subpage->lock, flags);
533 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
534 if (subpage_test_bitmap_all_zero(fs_info, subpage, writeback)) {
535 ASSERT(folio_test_writeback(folio));
536 folio_end_writeback(folio);
537 }
538 spin_unlock_irqrestore(&subpage->lock, flags);
539}
540
541void btrfs_subpage_set_ordered(const struct btrfs_fs_info *fs_info,
542 struct folio *folio, u64 start, u32 len)
543{
544 struct btrfs_subpage *subpage = folio_get_private(folio);
545 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
546 ordered, start, len);
547 unsigned long flags;
548
549 spin_lock_irqsave(&subpage->lock, flags);
550 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
551 folio_set_ordered(folio);
552 spin_unlock_irqrestore(&subpage->lock, flags);
553}
554
555void btrfs_subpage_clear_ordered(const struct btrfs_fs_info *fs_info,
556 struct folio *folio, u64 start, u32 len)
557{
558 struct btrfs_subpage *subpage = folio_get_private(folio);
559 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
560 ordered, start, len);
561 unsigned long flags;
562
563 spin_lock_irqsave(&subpage->lock, flags);
564 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
565 if (subpage_test_bitmap_all_zero(fs_info, subpage, ordered))
566 folio_clear_ordered(folio);
567 spin_unlock_irqrestore(&subpage->lock, flags);
568}
569
570void btrfs_subpage_set_checked(const struct btrfs_fs_info *fs_info,
571 struct folio *folio, u64 start, u32 len)
572{
573 struct btrfs_subpage *subpage = folio_get_private(folio);
574 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
575 checked, start, len);
576 unsigned long flags;
577
578 spin_lock_irqsave(&subpage->lock, flags);
579 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
580 if (subpage_test_bitmap_all_set(fs_info, subpage, checked))
581 folio_set_checked(folio);
582 spin_unlock_irqrestore(&subpage->lock, flags);
583}
584
585void btrfs_subpage_clear_checked(const struct btrfs_fs_info *fs_info,
586 struct folio *folio, u64 start, u32 len)
587{
588 struct btrfs_subpage *subpage = folio_get_private(folio);
589 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio,
590 checked, start, len);
591 unsigned long flags;
592
593 spin_lock_irqsave(&subpage->lock, flags);
594 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
595 folio_clear_checked(folio);
596 spin_unlock_irqrestore(&subpage->lock, flags);
597}
598
599/*
600 * Unlike set/clear which is dependent on each page status, for test all bits
601 * are tested in the same way.
602 */
603#define IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(name) \
604bool btrfs_subpage_test_##name(const struct btrfs_fs_info *fs_info, \
605 struct folio *folio, u64 start, u32 len) \
606{ \
607 struct btrfs_subpage *subpage = folio_get_private(folio); \
608 unsigned int start_bit = subpage_calc_start_bit(fs_info, folio, \
609 name, start, len); \
610 unsigned long flags; \
611 bool ret; \
612 \
613 spin_lock_irqsave(&subpage->lock, flags); \
614 ret = bitmap_test_range_all_set(subpage->bitmaps, start_bit, \
615 len >> fs_info->sectorsize_bits); \
616 spin_unlock_irqrestore(&subpage->lock, flags); \
617 return ret; \
618}
619IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(uptodate);
620IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(dirty);
621IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(writeback);
622IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(ordered);
623IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(checked);
624
625/*
626 * Note that, in selftests (extent-io-tests), we can have empty fs_info passed
627 * in. We only test sectorsize == PAGE_SIZE cases so far, thus we can fall
628 * back to regular sectorsize branch.
629 */
630#define IMPLEMENT_BTRFS_PAGE_OPS(name, folio_set_func, \
631 folio_clear_func, folio_test_func) \
632void btrfs_folio_set_##name(const struct btrfs_fs_info *fs_info, \
633 struct folio *folio, u64 start, u32 len) \
634{ \
635 if (unlikely(!fs_info) || \
636 !btrfs_is_subpage(fs_info, folio->mapping)) { \
637 folio_set_func(folio); \
638 return; \
639 } \
640 btrfs_subpage_set_##name(fs_info, folio, start, len); \
641} \
642void btrfs_folio_clear_##name(const struct btrfs_fs_info *fs_info, \
643 struct folio *folio, u64 start, u32 len) \
644{ \
645 if (unlikely(!fs_info) || \
646 !btrfs_is_subpage(fs_info, folio->mapping)) { \
647 folio_clear_func(folio); \
648 return; \
649 } \
650 btrfs_subpage_clear_##name(fs_info, folio, start, len); \
651} \
652bool btrfs_folio_test_##name(const struct btrfs_fs_info *fs_info, \
653 struct folio *folio, u64 start, u32 len) \
654{ \
655 if (unlikely(!fs_info) || \
656 !btrfs_is_subpage(fs_info, folio->mapping)) \
657 return folio_test_func(folio); \
658 return btrfs_subpage_test_##name(fs_info, folio, start, len); \
659} \
660void btrfs_folio_clamp_set_##name(const struct btrfs_fs_info *fs_info, \
661 struct folio *folio, u64 start, u32 len) \
662{ \
663 if (unlikely(!fs_info) || \
664 !btrfs_is_subpage(fs_info, folio->mapping)) { \
665 folio_set_func(folio); \
666 return; \
667 } \
668 btrfs_subpage_clamp_range(folio, &start, &len); \
669 btrfs_subpage_set_##name(fs_info, folio, start, len); \
670} \
671void btrfs_folio_clamp_clear_##name(const struct btrfs_fs_info *fs_info, \
672 struct folio *folio, u64 start, u32 len) \
673{ \
674 if (unlikely(!fs_info) || \
675 !btrfs_is_subpage(fs_info, folio->mapping)) { \
676 folio_clear_func(folio); \
677 return; \
678 } \
679 btrfs_subpage_clamp_range(folio, &start, &len); \
680 btrfs_subpage_clear_##name(fs_info, folio, start, len); \
681} \
682bool btrfs_folio_clamp_test_##name(const struct btrfs_fs_info *fs_info, \
683 struct folio *folio, u64 start, u32 len) \
684{ \
685 if (unlikely(!fs_info) || \
686 !btrfs_is_subpage(fs_info, folio->mapping)) \
687 return folio_test_func(folio); \
688 btrfs_subpage_clamp_range(folio, &start, &len); \
689 return btrfs_subpage_test_##name(fs_info, folio, start, len); \
690}
691IMPLEMENT_BTRFS_PAGE_OPS(uptodate, folio_mark_uptodate, folio_clear_uptodate,
692 folio_test_uptodate);
693IMPLEMENT_BTRFS_PAGE_OPS(dirty, folio_mark_dirty, folio_clear_dirty_for_io,
694 folio_test_dirty);
695IMPLEMENT_BTRFS_PAGE_OPS(writeback, folio_start_writeback, folio_end_writeback,
696 folio_test_writeback);
697IMPLEMENT_BTRFS_PAGE_OPS(ordered, folio_set_ordered, folio_clear_ordered,
698 folio_test_ordered);
699IMPLEMENT_BTRFS_PAGE_OPS(checked, folio_set_checked, folio_clear_checked,
700 folio_test_checked);
701
702/*
703 * Make sure not only the page dirty bit is cleared, but also subpage dirty bit
704 * is cleared.
705 */
706void btrfs_folio_assert_not_dirty(const struct btrfs_fs_info *fs_info, struct folio *folio)
707{
708 struct btrfs_subpage *subpage = folio_get_private(folio);
709
710 if (!IS_ENABLED(CONFIG_BTRFS_ASSERT))
711 return;
712
713 ASSERT(!folio_test_dirty(folio));
714 if (!btrfs_is_subpage(fs_info, folio->mapping))
715 return;
716
717 ASSERT(folio_test_private(folio) && folio_get_private(folio));
718 ASSERT(subpage_test_bitmap_all_zero(fs_info, subpage, dirty));
719}
720
721/*
722 * Handle different locked pages with different page sizes:
723 *
724 * - Page locked by plain lock_page()
725 * It should not have any subpage::writers count.
726 * Can be unlocked by unlock_page().
727 * This is the most common locked page for __extent_writepage() called
728 * inside extent_write_cache_pages().
729 * Rarer cases include the @locked_page from extent_write_locked_range().
730 *
731 * - Page locked by lock_delalloc_pages()
732 * There is only one caller, all pages except @locked_page for
733 * extent_write_locked_range().
734 * In this case, we have to call subpage helper to handle the case.
735 */
736void btrfs_folio_unlock_writer(struct btrfs_fs_info *fs_info,
737 struct folio *folio, u64 start, u32 len)
738{
739 struct btrfs_subpage *subpage;
740
741 ASSERT(folio_test_locked(folio));
742 /* For non-subpage case, we just unlock the page */
743 if (!btrfs_is_subpage(fs_info, folio->mapping)) {
744 folio_unlock(folio);
745 return;
746 }
747
748 ASSERT(folio_test_private(folio) && folio_get_private(folio));
749 subpage = folio_get_private(folio);
750
751 /*
752 * For subpage case, there are two types of locked page. With or
753 * without writers number.
754 *
755 * Since we own the page lock, no one else could touch subpage::writers
756 * and we are safe to do several atomic operations without spinlock.
757 */
758 if (atomic_read(&subpage->writers) == 0) {
759 /* No writers, locked by plain lock_page() */
760 folio_unlock(folio);
761 return;
762 }
763
764 /* Have writers, use proper subpage helper to end it */
765 btrfs_folio_end_writer_lock(fs_info, folio, start, len);
766}
767
768#define GET_SUBPAGE_BITMAP(subpage, subpage_info, name, dst) \
769 bitmap_cut(dst, subpage->bitmaps, 0, \
770 subpage_info->name##_offset, subpage_info->bitmap_nr_bits)
771
772void __cold btrfs_subpage_dump_bitmap(const struct btrfs_fs_info *fs_info,
773 struct folio *folio, u64 start, u32 len)
774{
775 struct btrfs_subpage_info *subpage_info = fs_info->subpage_info;
776 struct btrfs_subpage *subpage;
777 unsigned long uptodate_bitmap;
778 unsigned long error_bitmap;
779 unsigned long dirty_bitmap;
780 unsigned long writeback_bitmap;
781 unsigned long ordered_bitmap;
782 unsigned long checked_bitmap;
783 unsigned long flags;
784
785 ASSERT(folio_test_private(folio) && folio_get_private(folio));
786 ASSERT(subpage_info);
787 subpage = folio_get_private(folio);
788
789 spin_lock_irqsave(&subpage->lock, flags);
790 GET_SUBPAGE_BITMAP(subpage, subpage_info, uptodate, &uptodate_bitmap);
791 GET_SUBPAGE_BITMAP(subpage, subpage_info, dirty, &dirty_bitmap);
792 GET_SUBPAGE_BITMAP(subpage, subpage_info, writeback, &writeback_bitmap);
793 GET_SUBPAGE_BITMAP(subpage, subpage_info, ordered, &ordered_bitmap);
794 GET_SUBPAGE_BITMAP(subpage, subpage_info, checked, &checked_bitmap);
795 GET_SUBPAGE_BITMAP(subpage, subpage_info, locked, &checked_bitmap);
796 spin_unlock_irqrestore(&subpage->lock, flags);
797
798 dump_page(folio_page(folio, 0), "btrfs subpage dump");
799 btrfs_warn(fs_info,
800"start=%llu len=%u page=%llu, bitmaps uptodate=%*pbl error=%*pbl dirty=%*pbl writeback=%*pbl ordered=%*pbl checked=%*pbl",
801 start, len, folio_pos(folio),
802 subpage_info->bitmap_nr_bits, &uptodate_bitmap,
803 subpage_info->bitmap_nr_bits, &error_bitmap,
804 subpage_info->bitmap_nr_bits, &dirty_bitmap,
805 subpage_info->bitmap_nr_bits, &writeback_bitmap,
806 subpage_info->bitmap_nr_bits, &ordered_bitmap,
807 subpage_info->bitmap_nr_bits, &checked_bitmap);
808}