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
67bool btrfs_is_subpage(const struct btrfs_fs_info *fs_info, struct page *page)
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 (!page->mapping || !page->mapping->host ||
78 is_data_inode(page->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
90void btrfs_init_subpage_info(struct btrfs_subpage_info *subpage_info, u32 sectorsize)
91{
92 unsigned int cur = 0;
93 unsigned int nr_bits;
94
95 ASSERT(IS_ALIGNED(PAGE_SIZE, sectorsize));
96
97 nr_bits = PAGE_SIZE / sectorsize;
98 subpage_info->bitmap_nr_bits = nr_bits;
99
100 subpage_info->uptodate_offset = cur;
101 cur += nr_bits;
102
103 subpage_info->error_offset = cur;
104 cur += nr_bits;
105
106 subpage_info->dirty_offset = cur;
107 cur += nr_bits;
108
109 subpage_info->writeback_offset = cur;
110 cur += nr_bits;
111
112 subpage_info->ordered_offset = cur;
113 cur += nr_bits;
114
115 subpage_info->checked_offset = cur;
116 cur += nr_bits;
117
118 subpage_info->total_nr_bits = cur;
119}
120
121int btrfs_attach_subpage(const struct btrfs_fs_info *fs_info,
122 struct page *page, enum btrfs_subpage_type type)
123{
124 struct btrfs_subpage *subpage;
125
126 /*
127 * We have cases like a dummy extent buffer page, which is not mapped
128 * and doesn't need to be locked.
129 */
130 if (page->mapping)
131 ASSERT(PageLocked(page));
132
133 /* Either not subpage, or the page already has private attached */
134 if (!btrfs_is_subpage(fs_info, page) || PagePrivate(page))
135 return 0;
136
137 subpage = btrfs_alloc_subpage(fs_info, type);
138 if (IS_ERR(subpage))
139 return PTR_ERR(subpage);
140
141 attach_page_private(page, subpage);
142 return 0;
143}
144
145void btrfs_detach_subpage(const struct btrfs_fs_info *fs_info,
146 struct page *page)
147{
148 struct btrfs_subpage *subpage;
149
150 /* Either not subpage, or already detached */
151 if (!btrfs_is_subpage(fs_info, page) || !PagePrivate(page))
152 return;
153
154 subpage = detach_page_private(page);
155 ASSERT(subpage);
156 btrfs_free_subpage(subpage);
157}
158
159struct btrfs_subpage *btrfs_alloc_subpage(const struct btrfs_fs_info *fs_info,
160 enum btrfs_subpage_type type)
161{
162 struct btrfs_subpage *ret;
163 unsigned int real_size;
164
165 ASSERT(fs_info->sectorsize < PAGE_SIZE);
166
167 real_size = struct_size(ret, bitmaps,
168 BITS_TO_LONGS(fs_info->subpage_info->total_nr_bits));
169 ret = kzalloc(real_size, GFP_NOFS);
170 if (!ret)
171 return ERR_PTR(-ENOMEM);
172
173 spin_lock_init(&ret->lock);
174 if (type == BTRFS_SUBPAGE_METADATA) {
175 atomic_set(&ret->eb_refs, 0);
176 } else {
177 atomic_set(&ret->readers, 0);
178 atomic_set(&ret->writers, 0);
179 }
180 return ret;
181}
182
183void btrfs_free_subpage(struct btrfs_subpage *subpage)
184{
185 kfree(subpage);
186}
187
188/*
189 * Increase the eb_refs of current subpage.
190 *
191 * This is important for eb allocation, to prevent race with last eb freeing
192 * of the same page.
193 * With the eb_refs increased before the eb inserted into radix tree,
194 * detach_extent_buffer_page() won't detach the page private while we're still
195 * allocating the extent buffer.
196 */
197void btrfs_page_inc_eb_refs(const struct btrfs_fs_info *fs_info,
198 struct page *page)
199{
200 struct btrfs_subpage *subpage;
201
202 if (!btrfs_is_subpage(fs_info, page))
203 return;
204
205 ASSERT(PagePrivate(page) && page->mapping);
206 lockdep_assert_held(&page->mapping->private_lock);
207
208 subpage = (struct btrfs_subpage *)page->private;
209 atomic_inc(&subpage->eb_refs);
210}
211
212void btrfs_page_dec_eb_refs(const struct btrfs_fs_info *fs_info,
213 struct page *page)
214{
215 struct btrfs_subpage *subpage;
216
217 if (!btrfs_is_subpage(fs_info, page))
218 return;
219
220 ASSERT(PagePrivate(page) && page->mapping);
221 lockdep_assert_held(&page->mapping->private_lock);
222
223 subpage = (struct btrfs_subpage *)page->private;
224 ASSERT(atomic_read(&subpage->eb_refs));
225 atomic_dec(&subpage->eb_refs);
226}
227
228static void btrfs_subpage_assert(const struct btrfs_fs_info *fs_info,
229 struct page *page, u64 start, u32 len)
230{
231 /* Basic checks */
232 ASSERT(PagePrivate(page) && page->private);
233 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
234 IS_ALIGNED(len, fs_info->sectorsize));
235 /*
236 * The range check only works for mapped page, we can still have
237 * unmapped page like dummy extent buffer pages.
238 */
239 if (page->mapping)
240 ASSERT(page_offset(page) <= start &&
241 start + len <= page_offset(page) + PAGE_SIZE);
242}
243
244void btrfs_subpage_start_reader(const struct btrfs_fs_info *fs_info,
245 struct page *page, u64 start, u32 len)
246{
247 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
248 const int nbits = len >> fs_info->sectorsize_bits;
249
250 btrfs_subpage_assert(fs_info, page, start, len);
251
252 atomic_add(nbits, &subpage->readers);
253}
254
255void btrfs_subpage_end_reader(const struct btrfs_fs_info *fs_info,
256 struct page *page, u64 start, u32 len)
257{
258 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
259 const int nbits = len >> fs_info->sectorsize_bits;
260 bool is_data;
261 bool last;
262
263 btrfs_subpage_assert(fs_info, page, start, len);
264 is_data = is_data_inode(page->mapping->host);
265 ASSERT(atomic_read(&subpage->readers) >= nbits);
266 last = atomic_sub_and_test(nbits, &subpage->readers);
267
268 /*
269 * For data we need to unlock the page if the last read has finished.
270 *
271 * And please don't replace @last with atomic_sub_and_test() call
272 * inside if () condition.
273 * As we want the atomic_sub_and_test() to be always executed.
274 */
275 if (is_data && last)
276 unlock_page(page);
277}
278
279static void btrfs_subpage_clamp_range(struct page *page, u64 *start, u32 *len)
280{
281 u64 orig_start = *start;
282 u32 orig_len = *len;
283
284 *start = max_t(u64, page_offset(page), orig_start);
285 /*
286 * For certain call sites like btrfs_drop_pages(), we may have pages
287 * beyond the target range. In that case, just set @len to 0, subpage
288 * helpers can handle @len == 0 without any problem.
289 */
290 if (page_offset(page) >= orig_start + orig_len)
291 *len = 0;
292 else
293 *len = min_t(u64, page_offset(page) + PAGE_SIZE,
294 orig_start + orig_len) - *start;
295}
296
297void btrfs_subpage_start_writer(const struct btrfs_fs_info *fs_info,
298 struct page *page, u64 start, u32 len)
299{
300 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
301 const int nbits = (len >> fs_info->sectorsize_bits);
302 int ret;
303
304 btrfs_subpage_assert(fs_info, page, start, len);
305
306 ASSERT(atomic_read(&subpage->readers) == 0);
307 ret = atomic_add_return(nbits, &subpage->writers);
308 ASSERT(ret == nbits);
309}
310
311bool btrfs_subpage_end_and_test_writer(const struct btrfs_fs_info *fs_info,
312 struct page *page, u64 start, u32 len)
313{
314 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
315 const int nbits = (len >> fs_info->sectorsize_bits);
316
317 btrfs_subpage_assert(fs_info, page, start, len);
318
319 /*
320 * We have call sites passing @lock_page into
321 * extent_clear_unlock_delalloc() for compression path.
322 *
323 * This @locked_page is locked by plain lock_page(), thus its
324 * subpage::writers is 0. Handle them in a special way.
325 */
326 if (atomic_read(&subpage->writers) == 0)
327 return true;
328
329 ASSERT(atomic_read(&subpage->writers) >= nbits);
330 return atomic_sub_and_test(nbits, &subpage->writers);
331}
332
333/*
334 * Lock a page for delalloc page writeback.
335 *
336 * Return -EAGAIN if the page is not properly initialized.
337 * Return 0 with the page locked, and writer counter updated.
338 *
339 * Even with 0 returned, the page still need extra check to make sure
340 * it's really the correct page, as the caller is using
341 * filemap_get_folios_contig(), which can race with page invalidating.
342 */
343int btrfs_page_start_writer_lock(const struct btrfs_fs_info *fs_info,
344 struct page *page, u64 start, u32 len)
345{
346 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) {
347 lock_page(page);
348 return 0;
349 }
350 lock_page(page);
351 if (!PagePrivate(page) || !page->private) {
352 unlock_page(page);
353 return -EAGAIN;
354 }
355 btrfs_subpage_clamp_range(page, &start, &len);
356 btrfs_subpage_start_writer(fs_info, page, start, len);
357 return 0;
358}
359
360void btrfs_page_end_writer_lock(const struct btrfs_fs_info *fs_info,
361 struct page *page, u64 start, u32 len)
362{
363 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page))
364 return unlock_page(page);
365 btrfs_subpage_clamp_range(page, &start, &len);
366 if (btrfs_subpage_end_and_test_writer(fs_info, page, start, len))
367 unlock_page(page);
368}
369
370static bool bitmap_test_range_all_set(unsigned long *addr, unsigned int start,
371 unsigned int nbits)
372{
373 unsigned int found_zero;
374
375 found_zero = find_next_zero_bit(addr, start + nbits, start);
376 if (found_zero == start + nbits)
377 return true;
378 return false;
379}
380
381static bool bitmap_test_range_all_zero(unsigned long *addr, unsigned int start,
382 unsigned int nbits)
383{
384 unsigned int found_set;
385
386 found_set = find_next_bit(addr, start + nbits, start);
387 if (found_set == start + nbits)
388 return true;
389 return false;
390}
391
392#define subpage_calc_start_bit(fs_info, page, name, start, len) \
393({ \
394 unsigned int start_bit; \
395 \
396 btrfs_subpage_assert(fs_info, page, start, len); \
397 start_bit = offset_in_page(start) >> fs_info->sectorsize_bits; \
398 start_bit += fs_info->subpage_info->name##_offset; \
399 start_bit; \
400})
401
402#define subpage_test_bitmap_all_set(fs_info, subpage, name) \
403 bitmap_test_range_all_set(subpage->bitmaps, \
404 fs_info->subpage_info->name##_offset, \
405 fs_info->subpage_info->bitmap_nr_bits)
406
407#define subpage_test_bitmap_all_zero(fs_info, subpage, name) \
408 bitmap_test_range_all_zero(subpage->bitmaps, \
409 fs_info->subpage_info->name##_offset, \
410 fs_info->subpage_info->bitmap_nr_bits)
411
412void btrfs_subpage_set_uptodate(const struct btrfs_fs_info *fs_info,
413 struct page *page, u64 start, u32 len)
414{
415 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
416 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
417 uptodate, start, len);
418 unsigned long flags;
419
420 spin_lock_irqsave(&subpage->lock, flags);
421 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
422 if (subpage_test_bitmap_all_set(fs_info, subpage, uptodate))
423 SetPageUptodate(page);
424 spin_unlock_irqrestore(&subpage->lock, flags);
425}
426
427void btrfs_subpage_clear_uptodate(const struct btrfs_fs_info *fs_info,
428 struct page *page, u64 start, u32 len)
429{
430 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
431 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
432 uptodate, start, len);
433 unsigned long flags;
434
435 spin_lock_irqsave(&subpage->lock, flags);
436 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
437 ClearPageUptodate(page);
438 spin_unlock_irqrestore(&subpage->lock, flags);
439}
440
441void btrfs_subpage_set_error(const struct btrfs_fs_info *fs_info,
442 struct page *page, u64 start, u32 len)
443{
444 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
445 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
446 error, start, len);
447 unsigned long flags;
448
449 spin_lock_irqsave(&subpage->lock, flags);
450 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
451 SetPageError(page);
452 spin_unlock_irqrestore(&subpage->lock, flags);
453}
454
455void btrfs_subpage_clear_error(const struct btrfs_fs_info *fs_info,
456 struct page *page, u64 start, u32 len)
457{
458 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
459 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
460 error, start, len);
461 unsigned long flags;
462
463 spin_lock_irqsave(&subpage->lock, flags);
464 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
465 if (subpage_test_bitmap_all_zero(fs_info, subpage, error))
466 ClearPageError(page);
467 spin_unlock_irqrestore(&subpage->lock, flags);
468}
469
470void btrfs_subpage_set_dirty(const struct btrfs_fs_info *fs_info,
471 struct page *page, u64 start, u32 len)
472{
473 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
474 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
475 dirty, start, len);
476 unsigned long flags;
477
478 spin_lock_irqsave(&subpage->lock, flags);
479 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
480 spin_unlock_irqrestore(&subpage->lock, flags);
481 set_page_dirty(page);
482}
483
484/*
485 * Extra clear_and_test function for subpage dirty bitmap.
486 *
487 * Return true if we're the last bits in the dirty_bitmap and clear the
488 * dirty_bitmap.
489 * Return false otherwise.
490 *
491 * NOTE: Callers should manually clear page dirty for true case, as we have
492 * extra handling for tree blocks.
493 */
494bool btrfs_subpage_clear_and_test_dirty(const struct btrfs_fs_info *fs_info,
495 struct page *page, u64 start, u32 len)
496{
497 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
498 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
499 dirty, start, len);
500 unsigned long flags;
501 bool last = false;
502
503 spin_lock_irqsave(&subpage->lock, flags);
504 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
505 if (subpage_test_bitmap_all_zero(fs_info, subpage, dirty))
506 last = true;
507 spin_unlock_irqrestore(&subpage->lock, flags);
508 return last;
509}
510
511void btrfs_subpage_clear_dirty(const struct btrfs_fs_info *fs_info,
512 struct page *page, u64 start, u32 len)
513{
514 bool last;
515
516 last = btrfs_subpage_clear_and_test_dirty(fs_info, page, start, len);
517 if (last)
518 clear_page_dirty_for_io(page);
519}
520
521void btrfs_subpage_set_writeback(const struct btrfs_fs_info *fs_info,
522 struct page *page, u64 start, u32 len)
523{
524 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
525 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
526 writeback, start, len);
527 unsigned long flags;
528
529 spin_lock_irqsave(&subpage->lock, flags);
530 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
531 set_page_writeback(page);
532 spin_unlock_irqrestore(&subpage->lock, flags);
533}
534
535void btrfs_subpage_clear_writeback(const struct btrfs_fs_info *fs_info,
536 struct page *page, u64 start, u32 len)
537{
538 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
539 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
540 writeback, start, len);
541 unsigned long flags;
542
543 spin_lock_irqsave(&subpage->lock, flags);
544 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
545 if (subpage_test_bitmap_all_zero(fs_info, subpage, writeback)) {
546 ASSERT(PageWriteback(page));
547 end_page_writeback(page);
548 }
549 spin_unlock_irqrestore(&subpage->lock, flags);
550}
551
552void btrfs_subpage_set_ordered(const struct btrfs_fs_info *fs_info,
553 struct page *page, u64 start, u32 len)
554{
555 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
556 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
557 ordered, start, len);
558 unsigned long flags;
559
560 spin_lock_irqsave(&subpage->lock, flags);
561 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
562 SetPageOrdered(page);
563 spin_unlock_irqrestore(&subpage->lock, flags);
564}
565
566void btrfs_subpage_clear_ordered(const struct btrfs_fs_info *fs_info,
567 struct page *page, u64 start, u32 len)
568{
569 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
570 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
571 ordered, start, len);
572 unsigned long flags;
573
574 spin_lock_irqsave(&subpage->lock, flags);
575 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
576 if (subpage_test_bitmap_all_zero(fs_info, subpage, ordered))
577 ClearPageOrdered(page);
578 spin_unlock_irqrestore(&subpage->lock, flags);
579}
580
581void btrfs_subpage_set_checked(const struct btrfs_fs_info *fs_info,
582 struct page *page, u64 start, u32 len)
583{
584 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
585 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
586 checked, start, len);
587 unsigned long flags;
588
589 spin_lock_irqsave(&subpage->lock, flags);
590 bitmap_set(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
591 if (subpage_test_bitmap_all_set(fs_info, subpage, checked))
592 SetPageChecked(page);
593 spin_unlock_irqrestore(&subpage->lock, flags);
594}
595
596void btrfs_subpage_clear_checked(const struct btrfs_fs_info *fs_info,
597 struct page *page, u64 start, u32 len)
598{
599 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
600 unsigned int start_bit = subpage_calc_start_bit(fs_info, page,
601 checked, start, len);
602 unsigned long flags;
603
604 spin_lock_irqsave(&subpage->lock, flags);
605 bitmap_clear(subpage->bitmaps, start_bit, len >> fs_info->sectorsize_bits);
606 ClearPageChecked(page);
607 spin_unlock_irqrestore(&subpage->lock, flags);
608}
609
610/*
611 * Unlike set/clear which is dependent on each page status, for test all bits
612 * are tested in the same way.
613 */
614#define IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(name) \
615bool btrfs_subpage_test_##name(const struct btrfs_fs_info *fs_info, \
616 struct page *page, u64 start, u32 len) \
617{ \
618 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private; \
619 unsigned int start_bit = subpage_calc_start_bit(fs_info, page, \
620 name, start, len); \
621 unsigned long flags; \
622 bool ret; \
623 \
624 spin_lock_irqsave(&subpage->lock, flags); \
625 ret = bitmap_test_range_all_set(subpage->bitmaps, start_bit, \
626 len >> fs_info->sectorsize_bits); \
627 spin_unlock_irqrestore(&subpage->lock, flags); \
628 return ret; \
629}
630IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(uptodate);
631IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(error);
632IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(dirty);
633IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(writeback);
634IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(ordered);
635IMPLEMENT_BTRFS_SUBPAGE_TEST_OP(checked);
636
637/*
638 * Note that, in selftests (extent-io-tests), we can have empty fs_info passed
639 * in. We only test sectorsize == PAGE_SIZE cases so far, thus we can fall
640 * back to regular sectorsize branch.
641 */
642#define IMPLEMENT_BTRFS_PAGE_OPS(name, set_page_func, clear_page_func, \
643 test_page_func) \
644void btrfs_page_set_##name(const struct btrfs_fs_info *fs_info, \
645 struct page *page, u64 start, u32 len) \
646{ \
647 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) { \
648 set_page_func(page); \
649 return; \
650 } \
651 btrfs_subpage_set_##name(fs_info, page, start, len); \
652} \
653void btrfs_page_clear_##name(const struct btrfs_fs_info *fs_info, \
654 struct page *page, u64 start, u32 len) \
655{ \
656 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) { \
657 clear_page_func(page); \
658 return; \
659 } \
660 btrfs_subpage_clear_##name(fs_info, page, start, len); \
661} \
662bool btrfs_page_test_##name(const struct btrfs_fs_info *fs_info, \
663 struct page *page, u64 start, u32 len) \
664{ \
665 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) \
666 return test_page_func(page); \
667 return btrfs_subpage_test_##name(fs_info, page, start, len); \
668} \
669void btrfs_page_clamp_set_##name(const struct btrfs_fs_info *fs_info, \
670 struct page *page, u64 start, u32 len) \
671{ \
672 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) { \
673 set_page_func(page); \
674 return; \
675 } \
676 btrfs_subpage_clamp_range(page, &start, &len); \
677 btrfs_subpage_set_##name(fs_info, page, start, len); \
678} \
679void btrfs_page_clamp_clear_##name(const struct btrfs_fs_info *fs_info, \
680 struct page *page, u64 start, u32 len) \
681{ \
682 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) { \
683 clear_page_func(page); \
684 return; \
685 } \
686 btrfs_subpage_clamp_range(page, &start, &len); \
687 btrfs_subpage_clear_##name(fs_info, page, start, len); \
688} \
689bool btrfs_page_clamp_test_##name(const struct btrfs_fs_info *fs_info, \
690 struct page *page, u64 start, u32 len) \
691{ \
692 if (unlikely(!fs_info) || !btrfs_is_subpage(fs_info, page)) \
693 return test_page_func(page); \
694 btrfs_subpage_clamp_range(page, &start, &len); \
695 return btrfs_subpage_test_##name(fs_info, page, start, len); \
696}
697IMPLEMENT_BTRFS_PAGE_OPS(uptodate, SetPageUptodate, ClearPageUptodate,
698 PageUptodate);
699IMPLEMENT_BTRFS_PAGE_OPS(error, SetPageError, ClearPageError, PageError);
700IMPLEMENT_BTRFS_PAGE_OPS(dirty, set_page_dirty, clear_page_dirty_for_io,
701 PageDirty);
702IMPLEMENT_BTRFS_PAGE_OPS(writeback, set_page_writeback, end_page_writeback,
703 PageWriteback);
704IMPLEMENT_BTRFS_PAGE_OPS(ordered, SetPageOrdered, ClearPageOrdered,
705 PageOrdered);
706IMPLEMENT_BTRFS_PAGE_OPS(checked, SetPageChecked, ClearPageChecked, PageChecked);
707
708/*
709 * Make sure not only the page dirty bit is cleared, but also subpage dirty bit
710 * is cleared.
711 */
712void btrfs_page_assert_not_dirty(const struct btrfs_fs_info *fs_info,
713 struct page *page)
714{
715 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
716
717 if (!IS_ENABLED(CONFIG_BTRFS_ASSERT))
718 return;
719
720 ASSERT(!PageDirty(page));
721 if (!btrfs_is_subpage(fs_info, page))
722 return;
723
724 ASSERT(PagePrivate(page) && page->private);
725 ASSERT(subpage_test_bitmap_all_zero(fs_info, subpage, dirty));
726}
727
728/*
729 * Handle different locked pages with different page sizes:
730 *
731 * - Page locked by plain lock_page()
732 * It should not have any subpage::writers count.
733 * Can be unlocked by unlock_page().
734 * This is the most common locked page for __extent_writepage() called
735 * inside extent_write_cache_pages().
736 * Rarer cases include the @locked_page from extent_write_locked_range().
737 *
738 * - Page locked by lock_delalloc_pages()
739 * There is only one caller, all pages except @locked_page for
740 * extent_write_locked_range().
741 * In this case, we have to call subpage helper to handle the case.
742 */
743void btrfs_page_unlock_writer(struct btrfs_fs_info *fs_info, struct page *page,
744 u64 start, u32 len)
745{
746 struct btrfs_subpage *subpage;
747
748 ASSERT(PageLocked(page));
749 /* For non-subpage case, we just unlock the page */
750 if (!btrfs_is_subpage(fs_info, page))
751 return unlock_page(page);
752
753 ASSERT(PagePrivate(page) && page->private);
754 subpage = (struct btrfs_subpage *)page->private;
755
756 /*
757 * For subpage case, there are two types of locked page. With or
758 * without writers number.
759 *
760 * Since we own the page lock, no one else could touch subpage::writers
761 * and we are safe to do several atomic operations without spinlock.
762 */
763 if (atomic_read(&subpage->writers) == 0)
764 /* No writers, locked by plain lock_page() */
765 return unlock_page(page);
766
767 /* Have writers, use proper subpage helper to end it */
768 btrfs_page_end_writer_lock(fs_info, page, start, len);
769}
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}