4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 Copyright (C) 2004-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 2004-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2004-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 drbd is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
15 drbd is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with drbd; see the file COPYING. If not, write to
22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
25 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
27 #include <linux/bitops.h>
28 #include <linux/vmalloc.h>
29 #include <linux/string.h>
30 #include <linux/drbd.h>
31 #include <linux/slab.h>
32 #include <asm/kmap_types.h>
37 /* OPAQUE outside this file!
38 * interface defined in drbd_int.h
41 * function name drbd_bm_... => used elsewhere, "public".
42 * function name bm_... => internal to implementation, "private".
48 * We want to support >= peta byte of backend storage, while for now still using
49 * a granularity of one bit per 4KiB of storage.
50 * 1 << 50 bytes backend storage (1 PiB)
51 * 1 << (50 - 12) bits needed
52 * 38 --> we need u64 to index and count bits
53 * 1 << (38 - 3) bitmap bytes needed
54 * 35 --> we still need u64 to index and count bytes
55 * (that's 32 GiB of bitmap for 1 PiB storage)
56 * 1 << (35 - 2) 32bit longs needed
57 * 33 --> we'd even need u64 to index and count 32bit long words.
58 * 1 << (35 - 3) 64bit longs needed
59 * 32 --> we could get away with a 32bit unsigned int to index and count
60 * 64bit long words, but I rather stay with unsigned long for now.
61 * We probably should neither count nor point to bytes or long words
62 * directly, but either by bitnumber, or by page index and offset.
64 * 22 --> we need that much 4KiB pages of bitmap.
65 * 1 << (22 + 3) --> on a 64bit arch,
66 * we need 32 MiB to store the array of page pointers.
68 * Because I'm lazy, and because the resulting patch was too large, too ugly
69 * and still incomplete, on 32bit we still "only" support 16 TiB (minus some),
70 * (1 << 32) bits * 4k storage.
73 * bitmap storage and IO:
74 * Bitmap is stored little endian on disk, and is kept little endian in
75 * core memory. Currently we still hold the full bitmap in core as long
76 * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage
79 * We plan to reduce the amount of in-core bitmap pages by paging them in
80 * and out against their on-disk location as necessary, but need to make
81 * sure we don't cause too much meta data IO, and must not deadlock in
82 * tight memory situations. This needs some more work.
87 * Access to the *bm_pages is protected by bm_lock.
88 * It is safe to read the other members within the lock.
90 * drbd_bm_set_bits is called from bio_endio callbacks,
91 * We may be called with irq already disabled,
92 * so we need spin_lock_irqsave().
93 * And we need the kmap_atomic.
96 struct page
**bm_pages
;
99 /* see LIMITATIONS: above */
101 unsigned long bm_set
; /* nr of set bits; THINK maybe atomic_t? */
102 unsigned long bm_bits
;
104 size_t bm_number_of_pages
;
105 sector_t bm_dev_capacity
;
106 struct mutex bm_change
; /* serializes resize operations */
108 wait_queue_head_t bm_io_wait
; /* used to serialize IO of single pages */
110 enum bm_flag bm_flags
;
112 /* debugging aid, in case we are still racy somewhere */
114 struct task_struct
*bm_task
;
117 #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__)
118 static void __bm_print_lock_info(struct drbd_device
*device
, const char *func
)
120 struct drbd_bitmap
*b
= device
->bitmap
;
121 if (!__ratelimit(&drbd_ratelimit_state
))
123 drbd_err(device
, "FIXME %s[%d] in %s, bitmap locked for '%s' by %s[%d]\n",
124 current
->comm
, task_pid_nr(current
),
125 func
, b
->bm_why
?: "?",
126 b
->bm_task
->comm
, task_pid_nr(b
->bm_task
));
129 void drbd_bm_lock(struct drbd_device
*device
, char *why
, enum bm_flag flags
)
131 struct drbd_bitmap
*b
= device
->bitmap
;
135 drbd_err(device
, "FIXME no bitmap in drbd_bm_lock!?\n");
139 trylock_failed
= !mutex_trylock(&b
->bm_change
);
141 if (trylock_failed
) {
142 drbd_warn(device
, "%s[%d] going to '%s' but bitmap already locked for '%s' by %s[%d]\n",
143 current
->comm
, task_pid_nr(current
),
144 why
, b
->bm_why
?: "?",
145 b
->bm_task
->comm
, task_pid_nr(b
->bm_task
));
146 mutex_lock(&b
->bm_change
);
148 if (BM_LOCKED_MASK
& b
->bm_flags
)
149 drbd_err(device
, "FIXME bitmap already locked in bm_lock\n");
150 b
->bm_flags
|= flags
& BM_LOCKED_MASK
;
153 b
->bm_task
= current
;
156 void drbd_bm_unlock(struct drbd_device
*device
)
158 struct drbd_bitmap
*b
= device
->bitmap
;
160 drbd_err(device
, "FIXME no bitmap in drbd_bm_unlock!?\n");
164 if (!(BM_LOCKED_MASK
& device
->bitmap
->bm_flags
))
165 drbd_err(device
, "FIXME bitmap not locked in bm_unlock\n");
167 b
->bm_flags
&= ~BM_LOCKED_MASK
;
170 mutex_unlock(&b
->bm_change
);
173 /* we store some "meta" info about our pages in page->private */
174 /* at a granularity of 4k storage per bitmap bit:
175 * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks
177 * 1<<23 4k bitmap pages.
178 * Use 24 bits as page index, covers 2 peta byte storage
179 * at a granularity of 4k per bit.
180 * Used to report the failed page idx on io error from the endio handlers.
182 #define BM_PAGE_IDX_MASK ((1UL<<24)-1)
183 /* this page is currently read in, or written back */
184 #define BM_PAGE_IO_LOCK 31
185 /* if there has been an IO error for this page */
186 #define BM_PAGE_IO_ERROR 30
187 /* this is to be able to intelligently skip disk IO,
188 * set if bits have been set since last IO. */
189 #define BM_PAGE_NEED_WRITEOUT 29
190 /* to mark for lazy writeout once syncer cleared all clearable bits,
191 * we if bits have been cleared since last IO. */
192 #define BM_PAGE_LAZY_WRITEOUT 28
193 /* pages marked with this "HINT" will be considered for writeout
194 * on activity log transactions */
195 #define BM_PAGE_HINT_WRITEOUT 27
197 /* store_page_idx uses non-atomic assignment. It is only used directly after
198 * allocating the page. All other bm_set_page_* and bm_clear_page_* need to
199 * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap
200 * changes) may happen from various contexts, and wait_on_bit/wake_up_bit
201 * requires it all to be atomic as well. */
202 static void bm_store_page_idx(struct page
*page
, unsigned long idx
)
204 BUG_ON(0 != (idx
& ~BM_PAGE_IDX_MASK
));
205 set_page_private(page
, idx
);
208 static unsigned long bm_page_to_idx(struct page
*page
)
210 return page_private(page
) & BM_PAGE_IDX_MASK
;
213 /* As is very unlikely that the same page is under IO from more than one
214 * context, we can get away with a bit per page and one wait queue per bitmap.
216 static void bm_page_lock_io(struct drbd_device
*device
, int page_nr
)
218 struct drbd_bitmap
*b
= device
->bitmap
;
219 void *addr
= &page_private(b
->bm_pages
[page_nr
]);
220 wait_event(b
->bm_io_wait
, !test_and_set_bit(BM_PAGE_IO_LOCK
, addr
));
223 static void bm_page_unlock_io(struct drbd_device
*device
, int page_nr
)
225 struct drbd_bitmap
*b
= device
->bitmap
;
226 void *addr
= &page_private(b
->bm_pages
[page_nr
]);
227 clear_bit_unlock(BM_PAGE_IO_LOCK
, addr
);
228 wake_up(&device
->bitmap
->bm_io_wait
);
231 /* set _before_ submit_io, so it may be reset due to being changed
232 * while this page is in flight... will get submitted later again */
233 static void bm_set_page_unchanged(struct page
*page
)
236 clear_bit(BM_PAGE_NEED_WRITEOUT
, &page_private(page
));
237 clear_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
240 static void bm_set_page_need_writeout(struct page
*page
)
242 set_bit(BM_PAGE_NEED_WRITEOUT
, &page_private(page
));
246 * drbd_bm_mark_for_writeout() - mark a page with a "hint" to be considered for writeout
247 * @device: DRBD device.
248 * @page_nr: the bitmap page to mark with the "hint" flag
250 * From within an activity log transaction, we mark a few pages with these
251 * hints, then call drbd_bm_write_hinted(), which will only write out changed
252 * pages which are flagged with this mark.
254 void drbd_bm_mark_for_writeout(struct drbd_device
*device
, int page_nr
)
257 if (page_nr
>= device
->bitmap
->bm_number_of_pages
) {
258 drbd_warn(device
, "BAD: page_nr: %u, number_of_pages: %u\n",
259 page_nr
, (int)device
->bitmap
->bm_number_of_pages
);
262 page
= device
->bitmap
->bm_pages
[page_nr
];
263 set_bit(BM_PAGE_HINT_WRITEOUT
, &page_private(page
));
266 static int bm_test_page_unchanged(struct page
*page
)
268 volatile const unsigned long *addr
= &page_private(page
);
269 return (*addr
& ((1UL<<BM_PAGE_NEED_WRITEOUT
)|(1UL<<BM_PAGE_LAZY_WRITEOUT
))) == 0;
272 static void bm_set_page_io_err(struct page
*page
)
274 set_bit(BM_PAGE_IO_ERROR
, &page_private(page
));
277 static void bm_clear_page_io_err(struct page
*page
)
279 clear_bit(BM_PAGE_IO_ERROR
, &page_private(page
));
282 static void bm_set_page_lazy_writeout(struct page
*page
)
284 set_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
287 static int bm_test_page_lazy_writeout(struct page
*page
)
289 return test_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
292 /* on a 32bit box, this would allow for exactly (2<<38) bits. */
293 static unsigned int bm_word_to_page_idx(struct drbd_bitmap
*b
, unsigned long long_nr
)
295 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */
296 unsigned int page_nr
= long_nr
>> (PAGE_SHIFT
- LN2_BPL
+ 3);
297 BUG_ON(page_nr
>= b
->bm_number_of_pages
);
301 static unsigned int bm_bit_to_page_idx(struct drbd_bitmap
*b
, u64 bitnr
)
303 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */
304 unsigned int page_nr
= bitnr
>> (PAGE_SHIFT
+ 3);
305 BUG_ON(page_nr
>= b
->bm_number_of_pages
);
309 static unsigned long *__bm_map_pidx(struct drbd_bitmap
*b
, unsigned int idx
)
311 struct page
*page
= b
->bm_pages
[idx
];
312 return (unsigned long *) kmap_atomic(page
);
315 static unsigned long *bm_map_pidx(struct drbd_bitmap
*b
, unsigned int idx
)
317 return __bm_map_pidx(b
, idx
);
320 static void __bm_unmap(unsigned long *p_addr
)
322 kunmap_atomic(p_addr
);
325 static void bm_unmap(unsigned long *p_addr
)
327 return __bm_unmap(p_addr
);
330 /* long word offset of _bitmap_ sector */
331 #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
332 /* word offset from start of bitmap to word number _in_page_
333 * modulo longs per page
334 #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long))
335 hm, well, Philipp thinks gcc might not optimize the % into & (... - 1)
338 #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1))
340 /* Long words per page */
341 #define LWPP (PAGE_SIZE/sizeof(long))
344 * actually most functions herein should take a struct drbd_bitmap*, not a
345 * struct drbd_device*, but for the debug macros I like to have the device around
346 * to be able to report device specific.
350 static void bm_free_pages(struct page
**pages
, unsigned long number
)
356 for (i
= 0; i
< number
; i
++) {
358 pr_alert("bm_free_pages tried to free a NULL pointer; i=%lu n=%lu\n",
362 __free_page(pages
[i
]);
367 static void bm_vk_free(void *ptr
, int v
)
376 * "have" and "want" are NUMBER OF PAGES.
378 static struct page
**bm_realloc_pages(struct drbd_bitmap
*b
, unsigned long want
)
380 struct page
**old_pages
= b
->bm_pages
;
381 struct page
**new_pages
, *page
;
382 unsigned int i
, bytes
, vmalloced
= 0;
383 unsigned long have
= b
->bm_number_of_pages
;
385 BUG_ON(have
== 0 && old_pages
!= NULL
);
386 BUG_ON(have
!= 0 && old_pages
== NULL
);
391 /* Trying kmalloc first, falling back to vmalloc.
392 * GFP_NOIO, as this is called while drbd IO is "suspended",
393 * and during resize or attach on diskless Primary,
394 * we must not block on IO to ourselves.
395 * Context is receiver thread or dmsetup. */
396 bytes
= sizeof(struct page
*)*want
;
397 new_pages
= kzalloc(bytes
, GFP_NOIO
| __GFP_NOWARN
);
399 new_pages
= __vmalloc(bytes
,
400 GFP_NOIO
| __GFP_HIGHMEM
| __GFP_ZERO
,
408 for (i
= 0; i
< have
; i
++)
409 new_pages
[i
] = old_pages
[i
];
410 for (; i
< want
; i
++) {
411 page
= alloc_page(GFP_NOIO
| __GFP_HIGHMEM
);
413 bm_free_pages(new_pages
+ have
, i
- have
);
414 bm_vk_free(new_pages
, vmalloced
);
417 /* we want to know which page it is
418 * from the endio handlers */
419 bm_store_page_idx(page
, i
);
423 for (i
= 0; i
< want
; i
++)
424 new_pages
[i
] = old_pages
[i
];
425 /* NOT HERE, we are outside the spinlock!
426 bm_free_pages(old_pages + want, have - want);
431 b
->bm_flags
|= BM_P_VMALLOCED
;
433 b
->bm_flags
&= ~BM_P_VMALLOCED
;
439 * called on driver init only. TODO call when a device is created.
440 * allocates the drbd_bitmap, and stores it in device->bitmap.
442 int drbd_bm_init(struct drbd_device
*device
)
444 struct drbd_bitmap
*b
= device
->bitmap
;
446 b
= kzalloc(sizeof(struct drbd_bitmap
), GFP_KERNEL
);
449 spin_lock_init(&b
->bm_lock
);
450 mutex_init(&b
->bm_change
);
451 init_waitqueue_head(&b
->bm_io_wait
);
458 sector_t
drbd_bm_capacity(struct drbd_device
*device
)
460 if (!expect(device
->bitmap
))
462 return device
->bitmap
->bm_dev_capacity
;
465 /* called on driver unload. TODO: call when a device is destroyed.
467 void drbd_bm_cleanup(struct drbd_device
*device
)
469 if (!expect(device
->bitmap
))
471 bm_free_pages(device
->bitmap
->bm_pages
, device
->bitmap
->bm_number_of_pages
);
472 bm_vk_free(device
->bitmap
->bm_pages
, (BM_P_VMALLOCED
& device
->bitmap
->bm_flags
));
473 kfree(device
->bitmap
);
474 device
->bitmap
= NULL
;
478 * since (b->bm_bits % BITS_PER_LONG) != 0,
479 * this masks out the remaining bits.
480 * Returns the number of bits cleared.
482 #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3))
483 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1)
484 #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1)
485 static int bm_clear_surplus(struct drbd_bitmap
*b
)
488 unsigned long *p_addr
, *bm
;
492 /* number of bits modulo bits per page */
493 tmp
= (b
->bm_bits
& BITS_PER_PAGE_MASK
);
494 /* mask the used bits of the word containing the last bit */
495 mask
= (1UL << (tmp
& BITS_PER_LONG_MASK
)) -1;
496 /* bitmap is always stored little endian,
497 * on disk and in core memory alike */
498 mask
= cpu_to_lel(mask
);
500 p_addr
= bm_map_pidx(b
, b
->bm_number_of_pages
- 1);
501 bm
= p_addr
+ (tmp
/BITS_PER_LONG
);
503 /* If mask != 0, we are not exactly aligned, so bm now points
504 * to the long containing the last bit.
505 * If mask == 0, bm already points to the word immediately
506 * after the last (long word aligned) bit. */
507 cleared
= hweight_long(*bm
& ~mask
);
512 if (BITS_PER_LONG
== 32 && ((bm
- p_addr
) & 1) == 1) {
513 /* on a 32bit arch, we may need to zero out
514 * a padding long to align with a 64bit remote */
515 cleared
+= hweight_long(*bm
);
522 static void bm_set_surplus(struct drbd_bitmap
*b
)
525 unsigned long *p_addr
, *bm
;
528 /* number of bits modulo bits per page */
529 tmp
= (b
->bm_bits
& BITS_PER_PAGE_MASK
);
530 /* mask the used bits of the word containing the last bit */
531 mask
= (1UL << (tmp
& BITS_PER_LONG_MASK
)) -1;
532 /* bitmap is always stored little endian,
533 * on disk and in core memory alike */
534 mask
= cpu_to_lel(mask
);
536 p_addr
= bm_map_pidx(b
, b
->bm_number_of_pages
- 1);
537 bm
= p_addr
+ (tmp
/BITS_PER_LONG
);
539 /* If mask != 0, we are not exactly aligned, so bm now points
540 * to the long containing the last bit.
541 * If mask == 0, bm already points to the word immediately
542 * after the last (long word aligned) bit. */
547 if (BITS_PER_LONG
== 32 && ((bm
- p_addr
) & 1) == 1) {
548 /* on a 32bit arch, we may need to zero out
549 * a padding long to align with a 64bit remote */
555 /* you better not modify the bitmap while this is running,
556 * or its results will be stale */
557 static unsigned long bm_count_bits(struct drbd_bitmap
*b
)
559 unsigned long *p_addr
;
560 unsigned long bits
= 0;
561 unsigned long mask
= (1UL << (b
->bm_bits
& BITS_PER_LONG_MASK
)) -1;
562 int idx
, i
, last_word
;
564 /* all but last page */
565 for (idx
= 0; idx
< b
->bm_number_of_pages
- 1; idx
++) {
566 p_addr
= __bm_map_pidx(b
, idx
);
567 for (i
= 0; i
< LWPP
; i
++)
568 bits
+= hweight_long(p_addr
[i
]);
572 /* last (or only) page */
573 last_word
= ((b
->bm_bits
- 1) & BITS_PER_PAGE_MASK
) >> LN2_BPL
;
574 p_addr
= __bm_map_pidx(b
, idx
);
575 for (i
= 0; i
< last_word
; i
++)
576 bits
+= hweight_long(p_addr
[i
]);
577 p_addr
[last_word
] &= cpu_to_lel(mask
);
578 bits
+= hweight_long(p_addr
[last_word
]);
579 /* 32bit arch, may have an unused padding long */
580 if (BITS_PER_LONG
== 32 && (last_word
& 1) == 0)
581 p_addr
[last_word
+1] = 0;
586 /* offset and len in long words.*/
587 static void bm_memset(struct drbd_bitmap
*b
, size_t offset
, int c
, size_t len
)
589 unsigned long *p_addr
, *bm
;
595 if (end
> b
->bm_words
) {
596 pr_alert("bm_memset end > bm_words\n");
600 while (offset
< end
) {
601 do_now
= min_t(size_t, ALIGN(offset
+ 1, LWPP
), end
) - offset
;
602 idx
= bm_word_to_page_idx(b
, offset
);
603 p_addr
= bm_map_pidx(b
, idx
);
604 bm
= p_addr
+ MLPP(offset
);
605 if (bm
+do_now
> p_addr
+ LWPP
) {
606 pr_alert("BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n",
607 p_addr
, bm
, (int)do_now
);
609 memset(bm
, c
, do_now
* sizeof(long));
611 bm_set_page_need_writeout(b
->bm_pages
[idx
]);
616 /* For the layout, see comment above drbd_md_set_sector_offsets(). */
617 static u64
drbd_md_on_disk_bits(struct drbd_backing_dev
*ldev
)
620 if (ldev
->md
.al_offset
== 8)
621 bitmap_sectors
= ldev
->md
.md_size_sect
- ldev
->md
.bm_offset
;
623 bitmap_sectors
= ldev
->md
.al_offset
- ldev
->md
.bm_offset
;
624 return bitmap_sectors
<< (9 + 3);
628 * make sure the bitmap has enough room for the attached storage,
629 * if necessary, resize.
630 * called whenever we may have changed the device size.
631 * returns -ENOMEM if we could not allocate enough memory, 0 on success.
632 * In case this is actually a resize, we copy the old bitmap into the new one.
633 * Otherwise, the bitmap is initialized to all bits set.
635 int drbd_bm_resize(struct drbd_device
*device
, sector_t capacity
, int set_new_bits
)
637 struct drbd_bitmap
*b
= device
->bitmap
;
638 unsigned long bits
, words
, owords
, obits
;
639 unsigned long want
, have
, onpages
; /* number of pages */
640 struct page
**npages
, **opages
= NULL
;
641 int err
= 0, growing
;
642 int opages_vmalloced
;
647 drbd_bm_lock(device
, "resize", BM_LOCKED_MASK
);
649 drbd_info(device
, "drbd_bm_resize called with capacity == %llu\n",
650 (unsigned long long)capacity
);
652 if (capacity
== b
->bm_dev_capacity
)
655 opages_vmalloced
= (BM_P_VMALLOCED
& b
->bm_flags
);
658 spin_lock_irq(&b
->bm_lock
);
659 opages
= b
->bm_pages
;
660 onpages
= b
->bm_number_of_pages
;
661 owords
= b
->bm_words
;
663 b
->bm_number_of_pages
=
667 b
->bm_dev_capacity
= 0;
668 spin_unlock_irq(&b
->bm_lock
);
669 bm_free_pages(opages
, onpages
);
670 bm_vk_free(opages
, opages_vmalloced
);
673 bits
= BM_SECT_TO_BIT(ALIGN(capacity
, BM_SECT_PER_BIT
));
676 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL;
677 a 32bit host could present the wrong number of words
680 words
= ALIGN(bits
, 64) >> LN2_BPL
;
682 if (get_ldev(device
)) {
683 u64 bits_on_disk
= drbd_md_on_disk_bits(device
->ldev
);
685 if (bits
> bits_on_disk
) {
686 drbd_info(device
, "bits = %lu\n", bits
);
687 drbd_info(device
, "bits_on_disk = %llu\n", bits_on_disk
);
693 want
= ALIGN(words
*sizeof(long), PAGE_SIZE
) >> PAGE_SHIFT
;
694 have
= b
->bm_number_of_pages
;
696 D_ASSERT(device
, b
->bm_pages
!= NULL
);
697 npages
= b
->bm_pages
;
699 if (drbd_insert_fault(device
, DRBD_FAULT_BM_ALLOC
))
702 npages
= bm_realloc_pages(b
, want
);
710 spin_lock_irq(&b
->bm_lock
);
711 opages
= b
->bm_pages
;
712 owords
= b
->bm_words
;
715 growing
= bits
> obits
;
716 if (opages
&& growing
&& set_new_bits
)
719 b
->bm_pages
= npages
;
720 b
->bm_number_of_pages
= want
;
723 b
->bm_dev_capacity
= capacity
;
727 bm_memset(b
, owords
, 0xff, words
-owords
);
728 b
->bm_set
+= bits
- obits
;
730 bm_memset(b
, owords
, 0x00, words
-owords
);
735 /* implicit: (opages != NULL) && (opages != npages) */
736 bm_free_pages(opages
+ want
, have
- want
);
739 (void)bm_clear_surplus(b
);
741 spin_unlock_irq(&b
->bm_lock
);
742 if (opages
!= npages
)
743 bm_vk_free(opages
, opages_vmalloced
);
745 b
->bm_set
= bm_count_bits(b
);
746 drbd_info(device
, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits
, words
, want
);
749 drbd_bm_unlock(device
);
754 * if not protected by other means, return value may be out of date when
755 * leaving this function...
756 * we still need to lock it, since it is important that this returns
757 * bm_set == 0 precisely.
759 * maybe bm_set should be atomic_t ?
761 unsigned long _drbd_bm_total_weight(struct drbd_device
*device
)
763 struct drbd_bitmap
*b
= device
->bitmap
;
769 if (!expect(b
->bm_pages
))
772 spin_lock_irqsave(&b
->bm_lock
, flags
);
774 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
779 unsigned long drbd_bm_total_weight(struct drbd_device
*device
)
782 /* if I don't have a disk, I don't know about out-of-sync status */
783 if (!get_ldev_if_state(device
, D_NEGOTIATING
))
785 s
= _drbd_bm_total_weight(device
);
790 size_t drbd_bm_words(struct drbd_device
*device
)
792 struct drbd_bitmap
*b
= device
->bitmap
;
795 if (!expect(b
->bm_pages
))
801 unsigned long drbd_bm_bits(struct drbd_device
*device
)
803 struct drbd_bitmap
*b
= device
->bitmap
;
810 /* merge number words from buffer into the bitmap starting at offset.
811 * buffer[i] is expected to be little endian unsigned long.
812 * bitmap must be locked by drbd_bm_lock.
813 * currently only used from receive_bitmap.
815 void drbd_bm_merge_lel(struct drbd_device
*device
, size_t offset
, size_t number
,
816 unsigned long *buffer
)
818 struct drbd_bitmap
*b
= device
->bitmap
;
819 unsigned long *p_addr
, *bm
;
820 unsigned long word
, bits
;
824 end
= offset
+ number
;
828 if (!expect(b
->bm_pages
))
832 WARN_ON(offset
>= b
->bm_words
);
833 WARN_ON(end
> b
->bm_words
);
835 spin_lock_irq(&b
->bm_lock
);
836 while (offset
< end
) {
837 do_now
= min_t(size_t, ALIGN(offset
+1, LWPP
), end
) - offset
;
838 idx
= bm_word_to_page_idx(b
, offset
);
839 p_addr
= bm_map_pidx(b
, idx
);
840 bm
= p_addr
+ MLPP(offset
);
843 bits
= hweight_long(*bm
);
844 word
= *bm
| *buffer
++;
846 b
->bm_set
+= hweight_long(word
) - bits
;
849 bm_set_page_need_writeout(b
->bm_pages
[idx
]);
851 /* with 32bit <-> 64bit cross-platform connect
852 * this is only correct for current usage,
853 * where we _know_ that we are 64 bit aligned,
854 * and know that this function is used in this way, too...
856 if (end
== b
->bm_words
)
857 b
->bm_set
-= bm_clear_surplus(b
);
858 spin_unlock_irq(&b
->bm_lock
);
861 /* copy number words from the bitmap starting at offset into the buffer.
862 * buffer[i] will be little endian unsigned long.
864 void drbd_bm_get_lel(struct drbd_device
*device
, size_t offset
, size_t number
,
865 unsigned long *buffer
)
867 struct drbd_bitmap
*b
= device
->bitmap
;
868 unsigned long *p_addr
, *bm
;
871 end
= offset
+ number
;
875 if (!expect(b
->bm_pages
))
878 spin_lock_irq(&b
->bm_lock
);
879 if ((offset
>= b
->bm_words
) ||
880 (end
> b
->bm_words
) ||
882 drbd_err(device
, "offset=%lu number=%lu bm_words=%lu\n",
883 (unsigned long) offset
,
884 (unsigned long) number
,
885 (unsigned long) b
->bm_words
);
887 while (offset
< end
) {
888 do_now
= min_t(size_t, ALIGN(offset
+1, LWPP
), end
) - offset
;
889 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, offset
));
890 bm
= p_addr
+ MLPP(offset
);
897 spin_unlock_irq(&b
->bm_lock
);
900 /* set all bits in the bitmap */
901 void drbd_bm_set_all(struct drbd_device
*device
)
903 struct drbd_bitmap
*b
= device
->bitmap
;
906 if (!expect(b
->bm_pages
))
909 spin_lock_irq(&b
->bm_lock
);
910 bm_memset(b
, 0, 0xff, b
->bm_words
);
911 (void)bm_clear_surplus(b
);
912 b
->bm_set
= b
->bm_bits
;
913 spin_unlock_irq(&b
->bm_lock
);
916 /* clear all bits in the bitmap */
917 void drbd_bm_clear_all(struct drbd_device
*device
)
919 struct drbd_bitmap
*b
= device
->bitmap
;
922 if (!expect(b
->bm_pages
))
925 spin_lock_irq(&b
->bm_lock
);
926 bm_memset(b
, 0, 0, b
->bm_words
);
928 spin_unlock_irq(&b
->bm_lock
);
931 static void drbd_bm_aio_ctx_destroy(struct kref
*kref
)
933 struct drbd_bm_aio_ctx
*ctx
= container_of(kref
, struct drbd_bm_aio_ctx
, kref
);
936 spin_lock_irqsave(&ctx
->device
->resource
->req_lock
, flags
);
937 list_del(&ctx
->list
);
938 spin_unlock_irqrestore(&ctx
->device
->resource
->req_lock
, flags
);
939 put_ldev(ctx
->device
);
943 /* bv_page may be a copy, or may be the original */
944 static void drbd_bm_endio(struct bio
*bio
, int error
)
946 struct drbd_bm_aio_ctx
*ctx
= bio
->bi_private
;
947 struct drbd_device
*device
= ctx
->device
;
948 struct drbd_bitmap
*b
= device
->bitmap
;
949 unsigned int idx
= bm_page_to_idx(bio
->bi_io_vec
[0].bv_page
);
950 int uptodate
= bio_flagged(bio
, BIO_UPTODATE
);
953 /* strange behavior of some lower level drivers...
954 * fail the request by clearing the uptodate flag,
955 * but do not return any error?!
956 * do we want to WARN() on this? */
957 if (!error
&& !uptodate
)
960 if ((ctx
->flags
& BM_AIO_COPY_PAGES
) == 0 &&
961 !bm_test_page_unchanged(b
->bm_pages
[idx
]))
962 drbd_warn(device
, "bitmap page idx %u changed during IO!\n", idx
);
965 /* ctx error will hold the completed-last non-zero error code,
966 * in case error codes differ. */
968 bm_set_page_io_err(b
->bm_pages
[idx
]);
969 /* Not identical to on disk version of it.
970 * Is BM_PAGE_IO_ERROR enough? */
971 if (__ratelimit(&drbd_ratelimit_state
))
972 drbd_err(device
, "IO ERROR %d on bitmap page idx %u\n",
975 bm_clear_page_io_err(b
->bm_pages
[idx
]);
976 dynamic_drbd_dbg(device
, "bitmap page idx %u completed\n", idx
);
979 bm_page_unlock_io(device
, idx
);
981 if (ctx
->flags
& BM_AIO_COPY_PAGES
)
982 mempool_free(bio
->bi_io_vec
[0].bv_page
, drbd_md_io_page_pool
);
986 if (atomic_dec_and_test(&ctx
->in_flight
)) {
988 wake_up(&device
->misc_wait
);
989 kref_put(&ctx
->kref
, &drbd_bm_aio_ctx_destroy
);
993 static void bm_page_io_async(struct drbd_bm_aio_ctx
*ctx
, int page_nr
) __must_hold(local
)
995 struct bio
*bio
= bio_alloc_drbd(GFP_NOIO
);
996 struct drbd_device
*device
= ctx
->device
;
997 struct drbd_bitmap
*b
= device
->bitmap
;
1000 unsigned int rw
= (ctx
->flags
& BM_AIO_READ
) ? READ
: WRITE
;
1002 sector_t on_disk_sector
=
1003 device
->ldev
->md
.md_offset
+ device
->ldev
->md
.bm_offset
;
1004 on_disk_sector
+= ((sector_t
)page_nr
) << (PAGE_SHIFT
-9);
1006 /* this might happen with very small
1007 * flexible external meta data device,
1008 * or with PAGE_SIZE > 4k */
1009 len
= min_t(unsigned int, PAGE_SIZE
,
1010 (drbd_md_last_sector(device
->ldev
) - on_disk_sector
+ 1)<<9);
1012 /* serialize IO on this page */
1013 bm_page_lock_io(device
, page_nr
);
1014 /* before memcpy and submit,
1015 * so it can be redirtied any time */
1016 bm_set_page_unchanged(b
->bm_pages
[page_nr
]);
1018 if (ctx
->flags
& BM_AIO_COPY_PAGES
) {
1019 page
= mempool_alloc(drbd_md_io_page_pool
, __GFP_HIGHMEM
|__GFP_WAIT
);
1020 copy_highpage(page
, b
->bm_pages
[page_nr
]);
1021 bm_store_page_idx(page
, page_nr
);
1023 page
= b
->bm_pages
[page_nr
];
1024 bio
->bi_bdev
= device
->ldev
->md_bdev
;
1025 bio
->bi_iter
.bi_sector
= on_disk_sector
;
1026 /* bio_add_page of a single page to an empty bio will always succeed,
1027 * according to api. Do we want to assert that? */
1028 bio_add_page(bio
, page
, len
, 0);
1029 bio
->bi_private
= ctx
;
1030 bio
->bi_end_io
= drbd_bm_endio
;
1032 if (drbd_insert_fault(device
, (rw
& WRITE
) ? DRBD_FAULT_MD_WR
: DRBD_FAULT_MD_RD
)) {
1034 bio_endio(bio
, -EIO
);
1036 submit_bio(rw
, bio
);
1037 /* this should not count as user activity and cause the
1038 * resync to throttle -- see drbd_rs_should_slow_down(). */
1039 atomic_add(len
>> 9, &device
->rs_sect_ev
);
1044 * bm_rw: read/write the whole bitmap from/to its on disk location.
1046 static int bm_rw(struct drbd_device
*device
, const unsigned int flags
, unsigned lazy_writeout_upper_idx
) __must_hold(local
)
1048 struct drbd_bm_aio_ctx
*ctx
;
1049 struct drbd_bitmap
*b
= device
->bitmap
;
1050 int num_pages
, i
, count
= 0;
1056 * We are protected against bitmap disappearing/resizing by holding an
1057 * ldev reference (caller must have called get_ldev()).
1058 * For read/write, we are protected against changes to the bitmap by
1059 * the bitmap lock (see drbd_bitmap_io).
1060 * For lazy writeout, we don't care for ongoing changes to the bitmap,
1061 * as we submit copies of pages anyways.
1064 ctx
= kmalloc(sizeof(struct drbd_bm_aio_ctx
), GFP_NOIO
);
1068 *ctx
= (struct drbd_bm_aio_ctx
) {
1070 .start_jif
= jiffies
,
1071 .in_flight
= ATOMIC_INIT(1),
1075 .kref
= { ATOMIC_INIT(2) },
1078 if (!get_ldev_if_state(device
, D_ATTACHING
)) { /* put is in drbd_bm_aio_ctx_destroy() */
1079 drbd_err(device
, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n");
1083 /* Here D_ATTACHING is sufficient since drbd_bm_read() is called only from
1084 drbd_adm_attach(), after device->ldev was assigned. */
1086 if (0 == (ctx
->flags
& ~BM_AIO_READ
))
1087 WARN_ON(!(BM_LOCKED_MASK
& b
->bm_flags
));
1089 spin_lock_irq(&device
->resource
->req_lock
);
1090 list_add_tail(&ctx
->list
, &device
->pending_bitmap_io
);
1091 spin_unlock_irq(&device
->resource
->req_lock
);
1093 num_pages
= b
->bm_number_of_pages
;
1097 /* let the layers below us try to merge these bios... */
1098 for (i
= 0; i
< num_pages
; i
++) {
1099 /* ignore completely unchanged pages */
1100 if (lazy_writeout_upper_idx
&& i
== lazy_writeout_upper_idx
)
1102 if (!(flags
& BM_AIO_READ
)) {
1103 if ((flags
& BM_AIO_WRITE_HINTED
) &&
1104 !test_and_clear_bit(BM_PAGE_HINT_WRITEOUT
,
1105 &page_private(b
->bm_pages
[i
])))
1108 if (!(flags
& BM_AIO_WRITE_ALL_PAGES
) &&
1109 bm_test_page_unchanged(b
->bm_pages
[i
])) {
1110 dynamic_drbd_dbg(device
, "skipped bm write for idx %u\n", i
);
1113 /* during lazy writeout,
1114 * ignore those pages not marked for lazy writeout. */
1115 if (lazy_writeout_upper_idx
&&
1116 !bm_test_page_lazy_writeout(b
->bm_pages
[i
])) {
1117 dynamic_drbd_dbg(device
, "skipped bm lazy write for idx %u\n", i
);
1121 atomic_inc(&ctx
->in_flight
);
1122 bm_page_io_async(ctx
, i
);
1128 * We initialize ctx->in_flight to one to make sure drbd_bm_endio
1129 * will not set ctx->done early, and decrement / test it here. If there
1130 * are still some bios in flight, we need to wait for them here.
1131 * If all IO is done already (or nothing had been submitted), there is
1132 * no need to wait. Still, we need to put the kref associated with the
1133 * "in_flight reached zero, all done" event.
1135 if (!atomic_dec_and_test(&ctx
->in_flight
))
1136 wait_until_done_or_force_detached(device
, device
->ldev
, &ctx
->done
);
1138 kref_put(&ctx
->kref
, &drbd_bm_aio_ctx_destroy
);
1140 /* summary for global bitmap IO */
1142 drbd_info(device
, "bitmap %s of %u pages took %lu jiffies\n",
1143 (flags
& BM_AIO_READ
) ? "READ" : "WRITE",
1144 count
, jiffies
- now
);
1147 drbd_alert(device
, "we had at least one MD IO ERROR during bitmap IO\n");
1148 drbd_chk_io_error(device
, 1, DRBD_META_IO_ERROR
);
1149 err
= -EIO
; /* ctx->error ? */
1152 if (atomic_read(&ctx
->in_flight
))
1153 err
= -EIO
; /* Disk timeout/force-detach during IO... */
1156 if (flags
& BM_AIO_READ
) {
1157 b
->bm_set
= bm_count_bits(b
);
1158 drbd_info(device
, "recounting of set bits took additional %lu jiffies\n",
1163 if ((flags
& ~BM_AIO_READ
) == 0)
1164 drbd_info(device
, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n",
1165 ppsize(ppb
, now
<< (BM_BLOCK_SHIFT
-10)), now
);
1167 kref_put(&ctx
->kref
, &drbd_bm_aio_ctx_destroy
);
1172 * drbd_bm_read() - Read the whole bitmap from its on disk location.
1173 * @device: DRBD device.
1175 int drbd_bm_read(struct drbd_device
*device
) __must_hold(local
)
1177 return bm_rw(device
, BM_AIO_READ
, 0);
1181 * drbd_bm_write() - Write the whole bitmap to its on disk location.
1182 * @device: DRBD device.
1184 * Will only write pages that have changed since last IO.
1186 int drbd_bm_write(struct drbd_device
*device
) __must_hold(local
)
1188 return bm_rw(device
, 0, 0);
1192 * drbd_bm_write_all() - Write the whole bitmap to its on disk location.
1193 * @device: DRBD device.
1195 * Will write all pages.
1197 int drbd_bm_write_all(struct drbd_device
*device
) __must_hold(local
)
1199 return bm_rw(device
, BM_AIO_WRITE_ALL_PAGES
, 0);
1203 * drbd_bm_write_lazy() - Write bitmap pages 0 to @upper_idx-1, if they have changed.
1204 * @device: DRBD device.
1205 * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages
1207 int drbd_bm_write_lazy(struct drbd_device
*device
, unsigned upper_idx
) __must_hold(local
)
1209 return bm_rw(device
, BM_AIO_COPY_PAGES
, upper_idx
);
1213 * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location.
1214 * @device: DRBD device.
1216 * Will only write pages that have changed since last IO.
1217 * In contrast to drbd_bm_write(), this will copy the bitmap pages
1218 * to temporary writeout pages. It is intended to trigger a full write-out
1219 * while still allowing the bitmap to change, for example if a resync or online
1220 * verify is aborted due to a failed peer disk, while local IO continues, or
1221 * pending resync acks are still being processed.
1223 int drbd_bm_write_copy_pages(struct drbd_device
*device
) __must_hold(local
)
1225 return bm_rw(device
, BM_AIO_COPY_PAGES
, 0);
1229 * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed.
1230 * @device: DRBD device.
1232 int drbd_bm_write_hinted(struct drbd_device
*device
) __must_hold(local
)
1234 return bm_rw(device
, BM_AIO_WRITE_HINTED
| BM_AIO_COPY_PAGES
, 0);
1238 * find_first_bit returns int, we return unsigned long.
1239 * For this to work on 32bit arch with bitnumbers > (1<<32),
1240 * we'd need to return u64, and get a whole lot of other places
1241 * fixed where we still use unsigned long.
1243 * this returns a bit number, NOT a sector!
1245 static unsigned long __bm_find_next(struct drbd_device
*device
, unsigned long bm_fo
,
1246 const int find_zero_bit
)
1248 struct drbd_bitmap
*b
= device
->bitmap
;
1249 unsigned long *p_addr
;
1250 unsigned long bit_offset
;
1254 if (bm_fo
> b
->bm_bits
) {
1255 drbd_err(device
, "bm_fo=%lu bm_bits=%lu\n", bm_fo
, b
->bm_bits
);
1256 bm_fo
= DRBD_END_OF_BITMAP
;
1258 while (bm_fo
< b
->bm_bits
) {
1259 /* bit offset of the first bit in the page */
1260 bit_offset
= bm_fo
& ~BITS_PER_PAGE_MASK
;
1261 p_addr
= __bm_map_pidx(b
, bm_bit_to_page_idx(b
, bm_fo
));
1264 i
= find_next_zero_bit_le(p_addr
,
1265 PAGE_SIZE
*8, bm_fo
& BITS_PER_PAGE_MASK
);
1267 i
= find_next_bit_le(p_addr
,
1268 PAGE_SIZE
*8, bm_fo
& BITS_PER_PAGE_MASK
);
1271 if (i
< PAGE_SIZE
*8) {
1272 bm_fo
= bit_offset
+ i
;
1273 if (bm_fo
>= b
->bm_bits
)
1277 bm_fo
= bit_offset
+ PAGE_SIZE
*8;
1279 bm_fo
= DRBD_END_OF_BITMAP
;
1285 static unsigned long bm_find_next(struct drbd_device
*device
,
1286 unsigned long bm_fo
, const int find_zero_bit
)
1288 struct drbd_bitmap
*b
= device
->bitmap
;
1289 unsigned long i
= DRBD_END_OF_BITMAP
;
1293 if (!expect(b
->bm_pages
))
1296 spin_lock_irq(&b
->bm_lock
);
1297 if (BM_DONT_TEST
& b
->bm_flags
)
1298 bm_print_lock_info(device
);
1300 i
= __bm_find_next(device
, bm_fo
, find_zero_bit
);
1302 spin_unlock_irq(&b
->bm_lock
);
1306 unsigned long drbd_bm_find_next(struct drbd_device
*device
, unsigned long bm_fo
)
1308 return bm_find_next(device
, bm_fo
, 0);
1312 /* not yet needed for anything. */
1313 unsigned long drbd_bm_find_next_zero(struct drbd_device
*device
, unsigned long bm_fo
)
1315 return bm_find_next(device
, bm_fo
, 1);
1319 /* does not spin_lock_irqsave.
1320 * you must take drbd_bm_lock() first */
1321 unsigned long _drbd_bm_find_next(struct drbd_device
*device
, unsigned long bm_fo
)
1323 /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */
1324 return __bm_find_next(device
, bm_fo
, 0);
1327 unsigned long _drbd_bm_find_next_zero(struct drbd_device
*device
, unsigned long bm_fo
)
1329 /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */
1330 return __bm_find_next(device
, bm_fo
, 1);
1333 /* returns number of bits actually changed.
1334 * for val != 0, we change 0 -> 1, return code positive
1335 * for val == 0, we change 1 -> 0, return code negative
1336 * wants bitnr, not sector.
1337 * expected to be called for only a few bits (e - s about BITS_PER_LONG).
1338 * Must hold bitmap lock already. */
1339 static int __bm_change_bits_to(struct drbd_device
*device
, const unsigned long s
,
1340 unsigned long e
, int val
)
1342 struct drbd_bitmap
*b
= device
->bitmap
;
1343 unsigned long *p_addr
= NULL
;
1344 unsigned long bitnr
;
1345 unsigned int last_page_nr
= -1U;
1347 int changed_total
= 0;
1349 if (e
>= b
->bm_bits
) {
1350 drbd_err(device
, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n",
1352 e
= b
->bm_bits
? b
->bm_bits
-1 : 0;
1354 for (bitnr
= s
; bitnr
<= e
; bitnr
++) {
1355 unsigned int page_nr
= bm_bit_to_page_idx(b
, bitnr
);
1356 if (page_nr
!= last_page_nr
) {
1360 bm_set_page_lazy_writeout(b
->bm_pages
[last_page_nr
]);
1362 bm_set_page_need_writeout(b
->bm_pages
[last_page_nr
]);
1365 p_addr
= __bm_map_pidx(b
, page_nr
);
1366 last_page_nr
= page_nr
;
1369 c
+= (0 == __test_and_set_bit_le(bitnr
& BITS_PER_PAGE_MASK
, p_addr
));
1371 c
-= (0 != __test_and_clear_bit_le(bitnr
& BITS_PER_PAGE_MASK
, p_addr
));
1376 bm_set_page_lazy_writeout(b
->bm_pages
[last_page_nr
]);
1378 bm_set_page_need_writeout(b
->bm_pages
[last_page_nr
]);
1380 b
->bm_set
+= changed_total
;
1381 return changed_total
;
1384 /* returns number of bits actually changed.
1385 * for val != 0, we change 0 -> 1, return code positive
1386 * for val == 0, we change 1 -> 0, return code negative
1387 * wants bitnr, not sector */
1388 static int bm_change_bits_to(struct drbd_device
*device
, const unsigned long s
,
1389 const unsigned long e
, int val
)
1391 unsigned long flags
;
1392 struct drbd_bitmap
*b
= device
->bitmap
;
1397 if (!expect(b
->bm_pages
))
1400 spin_lock_irqsave(&b
->bm_lock
, flags
);
1401 if ((val
? BM_DONT_SET
: BM_DONT_CLEAR
) & b
->bm_flags
)
1402 bm_print_lock_info(device
);
1404 c
= __bm_change_bits_to(device
, s
, e
, val
);
1406 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1410 /* returns number of bits changed 0 -> 1 */
1411 int drbd_bm_set_bits(struct drbd_device
*device
, const unsigned long s
, const unsigned long e
)
1413 return bm_change_bits_to(device
, s
, e
, 1);
1416 /* returns number of bits changed 1 -> 0 */
1417 int drbd_bm_clear_bits(struct drbd_device
*device
, const unsigned long s
, const unsigned long e
)
1419 return -bm_change_bits_to(device
, s
, e
, 0);
1422 /* sets all bits in full words,
1423 * from first_word up to, but not including, last_word */
1424 static inline void bm_set_full_words_within_one_page(struct drbd_bitmap
*b
,
1425 int page_nr
, int first_word
, int last_word
)
1430 unsigned long *paddr
= kmap_atomic(b
->bm_pages
[page_nr
]);
1431 for (i
= first_word
; i
< last_word
; i
++) {
1432 bits
= hweight_long(paddr
[i
]);
1434 changed
+= BITS_PER_LONG
- bits
;
1436 kunmap_atomic(paddr
);
1438 /* We only need lazy writeout, the information is still in the
1439 * remote bitmap as well, and is reconstructed during the next
1440 * bitmap exchange, if lost locally due to a crash. */
1441 bm_set_page_lazy_writeout(b
->bm_pages
[page_nr
]);
1442 b
->bm_set
+= changed
;
1446 /* Same thing as drbd_bm_set_bits,
1447 * but more efficient for a large bit range.
1448 * You must first drbd_bm_lock().
1449 * Can be called to set the whole bitmap in one go.
1450 * Sets bits from s to e _inclusive_. */
1451 void _drbd_bm_set_bits(struct drbd_device
*device
, const unsigned long s
, const unsigned long e
)
1453 /* First set_bit from the first bit (s)
1454 * up to the next long boundary (sl),
1455 * then assign full words up to the last long boundary (el),
1456 * then set_bit up to and including the last bit (e).
1458 * Do not use memset, because we must account for changes,
1459 * so we need to loop over the words with hweight() anyways.
1461 struct drbd_bitmap
*b
= device
->bitmap
;
1462 unsigned long sl
= ALIGN(s
,BITS_PER_LONG
);
1463 unsigned long el
= (e
+1) & ~((unsigned long)BITS_PER_LONG
-1);
1470 if (e
- s
<= 3*BITS_PER_LONG
) {
1471 /* don't bother; el and sl may even be wrong. */
1472 spin_lock_irq(&b
->bm_lock
);
1473 __bm_change_bits_to(device
, s
, e
, 1);
1474 spin_unlock_irq(&b
->bm_lock
);
1478 /* difference is large enough that we can trust sl and el */
1480 spin_lock_irq(&b
->bm_lock
);
1482 /* bits filling the current long */
1484 __bm_change_bits_to(device
, s
, sl
-1, 1);
1486 first_page
= sl
>> (3 + PAGE_SHIFT
);
1487 last_page
= el
>> (3 + PAGE_SHIFT
);
1489 /* MLPP: modulo longs per page */
1490 /* LWPP: long words per page */
1491 first_word
= MLPP(sl
>> LN2_BPL
);
1494 /* first and full pages, unless first page == last page */
1495 for (page_nr
= first_page
; page_nr
< last_page
; page_nr
++) {
1496 bm_set_full_words_within_one_page(device
->bitmap
, page_nr
, first_word
, last_word
);
1497 spin_unlock_irq(&b
->bm_lock
);
1500 spin_lock_irq(&b
->bm_lock
);
1502 /* last page (respectively only page, for first page == last page) */
1503 last_word
= MLPP(el
>> LN2_BPL
);
1505 /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples).
1506 * ==> e = 32767, el = 32768, last_page = 2,
1507 * and now last_word = 0.
1508 * We do not want to touch last_page in this case,
1509 * as we did not allocate it, it is not present in bitmap->bm_pages.
1512 bm_set_full_words_within_one_page(device
->bitmap
, last_page
, first_word
, last_word
);
1514 /* possibly trailing bits.
1515 * example: (e & 63) == 63, el will be e+1.
1516 * if that even was the very last bit,
1517 * it would trigger an assert in __bm_change_bits_to()
1520 __bm_change_bits_to(device
, el
, e
, 1);
1521 spin_unlock_irq(&b
->bm_lock
);
1524 /* returns bit state
1525 * wants bitnr, NOT sector.
1526 * inherently racy... area needs to be locked by means of {al,rs}_lru
1529 * -1 ... first out of bounds access, stop testing for bits!
1531 int drbd_bm_test_bit(struct drbd_device
*device
, const unsigned long bitnr
)
1533 unsigned long flags
;
1534 struct drbd_bitmap
*b
= device
->bitmap
;
1535 unsigned long *p_addr
;
1540 if (!expect(b
->bm_pages
))
1543 spin_lock_irqsave(&b
->bm_lock
, flags
);
1544 if (BM_DONT_TEST
& b
->bm_flags
)
1545 bm_print_lock_info(device
);
1546 if (bitnr
< b
->bm_bits
) {
1547 p_addr
= bm_map_pidx(b
, bm_bit_to_page_idx(b
, bitnr
));
1548 i
= test_bit_le(bitnr
& BITS_PER_PAGE_MASK
, p_addr
) ? 1 : 0;
1550 } else if (bitnr
== b
->bm_bits
) {
1552 } else { /* (bitnr > b->bm_bits) */
1553 drbd_err(device
, "bitnr=%lu > bm_bits=%lu\n", bitnr
, b
->bm_bits
);
1557 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1561 /* returns number of bits set in the range [s, e] */
1562 int drbd_bm_count_bits(struct drbd_device
*device
, const unsigned long s
, const unsigned long e
)
1564 unsigned long flags
;
1565 struct drbd_bitmap
*b
= device
->bitmap
;
1566 unsigned long *p_addr
= NULL
;
1567 unsigned long bitnr
;
1568 unsigned int page_nr
= -1U;
1571 /* If this is called without a bitmap, that is a bug. But just to be
1572 * robust in case we screwed up elsewhere, in that case pretend there
1573 * was one dirty bit in the requested area, so we won't try to do a
1574 * local read there (no bitmap probably implies no disk) */
1577 if (!expect(b
->bm_pages
))
1580 spin_lock_irqsave(&b
->bm_lock
, flags
);
1581 if (BM_DONT_TEST
& b
->bm_flags
)
1582 bm_print_lock_info(device
);
1583 for (bitnr
= s
; bitnr
<= e
; bitnr
++) {
1584 unsigned int idx
= bm_bit_to_page_idx(b
, bitnr
);
1585 if (page_nr
!= idx
) {
1589 p_addr
= bm_map_pidx(b
, idx
);
1591 if (expect(bitnr
< b
->bm_bits
))
1592 c
+= (0 != test_bit_le(bitnr
- (page_nr
<< (PAGE_SHIFT
+3)), p_addr
));
1594 drbd_err(device
, "bitnr=%lu bm_bits=%lu\n", bitnr
, b
->bm_bits
);
1598 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1603 /* inherently racy...
1604 * return value may be already out-of-date when this function returns.
1605 * but the general usage is that this is only use during a cstate when bits are
1606 * only cleared, not set, and typically only care for the case when the return
1607 * value is zero, or we already "locked" this "bitmap extent" by other means.
1609 * enr is bm-extent number, since we chose to name one sector (512 bytes)
1610 * worth of the bitmap a "bitmap extent".
1613 * I think since we use it like a reference count, we should use the real
1614 * reference count of some bitmap extent element from some lru instead...
1617 int drbd_bm_e_weight(struct drbd_device
*device
, unsigned long enr
)
1619 struct drbd_bitmap
*b
= device
->bitmap
;
1621 unsigned long flags
;
1622 unsigned long *p_addr
, *bm
;
1626 if (!expect(b
->bm_pages
))
1629 spin_lock_irqsave(&b
->bm_lock
, flags
);
1630 if (BM_DONT_TEST
& b
->bm_flags
)
1631 bm_print_lock_info(device
);
1634 e
= min((size_t)S2W(enr
+1), b
->bm_words
);
1636 if (s
< b
->bm_words
) {
1638 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, s
));
1639 bm
= p_addr
+ MLPP(s
);
1641 count
+= hweight_long(*bm
++);
1644 drbd_err(device
, "start offset (%d) too large in drbd_bm_e_weight\n", s
);
1646 spin_unlock_irqrestore(&b
->bm_lock
, flags
);