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 #include <linux/bitops.h>
26 #include <linux/vmalloc.h>
27 #include <linux/string.h>
28 #include <linux/drbd.h>
29 #include <linux/slab.h>
30 #include <asm/kmap_types.h>
35 /* OPAQUE outside this file!
36 * interface defined in drbd_int.h
39 * function name drbd_bm_... => used elsewhere, "public".
40 * function name bm_... => internal to implementation, "private".
46 * We want to support >= peta byte of backend storage, while for now still using
47 * a granularity of one bit per 4KiB of storage.
48 * 1 << 50 bytes backend storage (1 PiB)
49 * 1 << (50 - 12) bits needed
50 * 38 --> we need u64 to index and count bits
51 * 1 << (38 - 3) bitmap bytes needed
52 * 35 --> we still need u64 to index and count bytes
53 * (that's 32 GiB of bitmap for 1 PiB storage)
54 * 1 << (35 - 2) 32bit longs needed
55 * 33 --> we'd even need u64 to index and count 32bit long words.
56 * 1 << (35 - 3) 64bit longs needed
57 * 32 --> we could get away with a 32bit unsigned int to index and count
58 * 64bit long words, but I rather stay with unsigned long for now.
59 * We probably should neither count nor point to bytes or long words
60 * directly, but either by bitnumber, or by page index and offset.
62 * 22 --> we need that much 4KiB pages of bitmap.
63 * 1 << (22 + 3) --> on a 64bit arch,
64 * we need 32 MiB to store the array of page pointers.
66 * Because I'm lazy, and because the resulting patch was too large, too ugly
67 * and still incomplete, on 32bit we still "only" support 16 TiB (minus some),
68 * (1 << 32) bits * 4k storage.
71 * bitmap storage and IO:
72 * Bitmap is stored little endian on disk, and is kept little endian in
73 * core memory. Currently we still hold the full bitmap in core as long
74 * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage
77 * We plan to reduce the amount of in-core bitmap pages by paging them in
78 * and out against their on-disk location as necessary, but need to make
79 * sure we don't cause too much meta data IO, and must not deadlock in
80 * tight memory situations. This needs some more work.
85 * Access to the *bm_pages is protected by bm_lock.
86 * It is safe to read the other members within the lock.
88 * drbd_bm_set_bits is called from bio_endio callbacks,
89 * We may be called with irq already disabled,
90 * so we need spin_lock_irqsave().
91 * And we need the kmap_atomic.
94 struct page
**bm_pages
;
97 /* see LIMITATIONS: above */
99 unsigned long bm_set
; /* nr of set bits; THINK maybe atomic_t? */
100 unsigned long bm_bits
;
102 size_t bm_number_of_pages
;
103 sector_t bm_dev_capacity
;
104 struct mutex bm_change
; /* serializes resize operations */
106 wait_queue_head_t bm_io_wait
; /* used to serialize IO of single pages */
108 enum bm_flag bm_flags
;
110 /* debugging aid, in case we are still racy somewhere */
112 struct task_struct
*bm_task
;
115 #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__)
116 static void __bm_print_lock_info(struct drbd_conf
*mdev
, const char *func
)
118 struct drbd_bitmap
*b
= mdev
->bitmap
;
119 if (!__ratelimit(&drbd_ratelimit_state
))
121 dev_err(DEV
, "FIXME %s in %s, bitmap locked for '%s' by %s\n",
122 current
== mdev
->receiver
.task
? "receiver" :
123 current
== mdev
->asender
.task
? "asender" :
124 current
== mdev
->worker
.task
? "worker" : current
->comm
,
125 func
, b
->bm_why
?: "?",
126 b
->bm_task
== mdev
->receiver
.task
? "receiver" :
127 b
->bm_task
== mdev
->asender
.task
? "asender" :
128 b
->bm_task
== mdev
->worker
.task
? "worker" : "?");
131 void drbd_bm_lock(struct drbd_conf
*mdev
, char *why
, enum bm_flag flags
)
133 struct drbd_bitmap
*b
= mdev
->bitmap
;
137 dev_err(DEV
, "FIXME no bitmap in drbd_bm_lock!?\n");
141 trylock_failed
= !mutex_trylock(&b
->bm_change
);
143 if (trylock_failed
) {
144 dev_warn(DEV
, "%s going to '%s' but bitmap already locked for '%s' by %s\n",
145 current
== mdev
->receiver
.task
? "receiver" :
146 current
== mdev
->asender
.task
? "asender" :
147 current
== mdev
->worker
.task
? "worker" : current
->comm
,
148 why
, b
->bm_why
?: "?",
149 b
->bm_task
== mdev
->receiver
.task
? "receiver" :
150 b
->bm_task
== mdev
->asender
.task
? "asender" :
151 b
->bm_task
== mdev
->worker
.task
? "worker" : "?");
152 mutex_lock(&b
->bm_change
);
154 if (BM_LOCKED_MASK
& b
->bm_flags
)
155 dev_err(DEV
, "FIXME bitmap already locked in bm_lock\n");
156 b
->bm_flags
|= flags
& BM_LOCKED_MASK
;
159 b
->bm_task
= current
;
162 void drbd_bm_unlock(struct drbd_conf
*mdev
)
164 struct drbd_bitmap
*b
= mdev
->bitmap
;
166 dev_err(DEV
, "FIXME no bitmap in drbd_bm_unlock!?\n");
170 if (!(BM_LOCKED_MASK
& mdev
->bitmap
->bm_flags
))
171 dev_err(DEV
, "FIXME bitmap not locked in bm_unlock\n");
173 b
->bm_flags
&= ~BM_LOCKED_MASK
;
176 mutex_unlock(&b
->bm_change
);
179 /* we store some "meta" info about our pages in page->private */
180 /* at a granularity of 4k storage per bitmap bit:
181 * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks
183 * 1<<23 4k bitmap pages.
184 * Use 24 bits as page index, covers 2 peta byte storage
185 * at a granularity of 4k per bit.
186 * Used to report the failed page idx on io error from the endio handlers.
188 #define BM_PAGE_IDX_MASK ((1UL<<24)-1)
189 /* this page is currently read in, or written back */
190 #define BM_PAGE_IO_LOCK 31
191 /* if there has been an IO error for this page */
192 #define BM_PAGE_IO_ERROR 30
193 /* this is to be able to intelligently skip disk IO,
194 * set if bits have been set since last IO. */
195 #define BM_PAGE_NEED_WRITEOUT 29
196 /* to mark for lazy writeout once syncer cleared all clearable bits,
197 * we if bits have been cleared since last IO. */
198 #define BM_PAGE_LAZY_WRITEOUT 28
200 /* store_page_idx uses non-atomic assignment. It is only used directly after
201 * allocating the page. All other bm_set_page_* and bm_clear_page_* need to
202 * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap
203 * changes) may happen from various contexts, and wait_on_bit/wake_up_bit
204 * requires it all to be atomic as well. */
205 static void bm_store_page_idx(struct page
*page
, unsigned long idx
)
207 BUG_ON(0 != (idx
& ~BM_PAGE_IDX_MASK
));
208 set_page_private(page
, idx
);
211 static unsigned long bm_page_to_idx(struct page
*page
)
213 return page_private(page
) & BM_PAGE_IDX_MASK
;
216 /* As is very unlikely that the same page is under IO from more than one
217 * context, we can get away with a bit per page and one wait queue per bitmap.
219 static void bm_page_lock_io(struct drbd_conf
*mdev
, int page_nr
)
221 struct drbd_bitmap
*b
= mdev
->bitmap
;
222 void *addr
= &page_private(b
->bm_pages
[page_nr
]);
223 wait_event(b
->bm_io_wait
, !test_and_set_bit(BM_PAGE_IO_LOCK
, addr
));
226 static void bm_page_unlock_io(struct drbd_conf
*mdev
, int page_nr
)
228 struct drbd_bitmap
*b
= mdev
->bitmap
;
229 void *addr
= &page_private(b
->bm_pages
[page_nr
]);
230 clear_bit(BM_PAGE_IO_LOCK
, addr
);
231 smp_mb__after_clear_bit();
232 wake_up(&mdev
->bitmap
->bm_io_wait
);
235 /* set _before_ submit_io, so it may be reset due to being changed
236 * while this page is in flight... will get submitted later again */
237 static void bm_set_page_unchanged(struct page
*page
)
240 clear_bit(BM_PAGE_NEED_WRITEOUT
, &page_private(page
));
241 clear_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
244 static void bm_set_page_need_writeout(struct page
*page
)
246 set_bit(BM_PAGE_NEED_WRITEOUT
, &page_private(page
));
249 static int bm_test_page_unchanged(struct page
*page
)
251 volatile const unsigned long *addr
= &page_private(page
);
252 return (*addr
& ((1UL<<BM_PAGE_NEED_WRITEOUT
)|(1UL<<BM_PAGE_LAZY_WRITEOUT
))) == 0;
255 static void bm_set_page_io_err(struct page
*page
)
257 set_bit(BM_PAGE_IO_ERROR
, &page_private(page
));
260 static void bm_clear_page_io_err(struct page
*page
)
262 clear_bit(BM_PAGE_IO_ERROR
, &page_private(page
));
265 static void bm_set_page_lazy_writeout(struct page
*page
)
267 set_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
270 static int bm_test_page_lazy_writeout(struct page
*page
)
272 return test_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
275 /* on a 32bit box, this would allow for exactly (2<<38) bits. */
276 static unsigned int bm_word_to_page_idx(struct drbd_bitmap
*b
, unsigned long long_nr
)
278 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */
279 unsigned int page_nr
= long_nr
>> (PAGE_SHIFT
- LN2_BPL
+ 3);
280 BUG_ON(page_nr
>= b
->bm_number_of_pages
);
284 static unsigned int bm_bit_to_page_idx(struct drbd_bitmap
*b
, u64 bitnr
)
286 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */
287 unsigned int page_nr
= bitnr
>> (PAGE_SHIFT
+ 3);
288 BUG_ON(page_nr
>= b
->bm_number_of_pages
);
292 static unsigned long *__bm_map_pidx(struct drbd_bitmap
*b
, unsigned int idx
)
294 struct page
*page
= b
->bm_pages
[idx
];
295 return (unsigned long *) kmap_atomic(page
);
298 static unsigned long *bm_map_pidx(struct drbd_bitmap
*b
, unsigned int idx
)
300 return __bm_map_pidx(b
, idx
);
303 static void __bm_unmap(unsigned long *p_addr
)
305 kunmap_atomic(p_addr
);
308 static void bm_unmap(unsigned long *p_addr
)
310 return __bm_unmap(p_addr
);
313 /* long word offset of _bitmap_ sector */
314 #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
315 /* word offset from start of bitmap to word number _in_page_
316 * modulo longs per page
317 #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long))
318 hm, well, Philipp thinks gcc might not optimize the % into & (... - 1)
321 #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1))
323 /* Long words per page */
324 #define LWPP (PAGE_SIZE/sizeof(long))
327 * actually most functions herein should take a struct drbd_bitmap*, not a
328 * struct drbd_conf*, but for the debug macros I like to have the mdev around
329 * to be able to report device specific.
333 static void bm_free_pages(struct page
**pages
, unsigned long number
)
339 for (i
= 0; i
< number
; i
++) {
341 printk(KERN_ALERT
"drbd: bm_free_pages tried to free "
342 "a NULL pointer; i=%lu n=%lu\n",
346 __free_page(pages
[i
]);
351 static void bm_vk_free(void *ptr
, int v
)
360 * "have" and "want" are NUMBER OF PAGES.
362 static struct page
**bm_realloc_pages(struct drbd_bitmap
*b
, unsigned long want
)
364 struct page
**old_pages
= b
->bm_pages
;
365 struct page
**new_pages
, *page
;
366 unsigned int i
, bytes
, vmalloced
= 0;
367 unsigned long have
= b
->bm_number_of_pages
;
369 BUG_ON(have
== 0 && old_pages
!= NULL
);
370 BUG_ON(have
!= 0 && old_pages
== NULL
);
375 /* Trying kmalloc first, falling back to vmalloc.
376 * GFP_KERNEL is ok, as this is done when a lower level disk is
377 * "attached" to the drbd. Context is receiver thread or cqueue
378 * thread. As we have no disk yet, we are not in the IO path,
379 * not even the IO path of the peer. */
380 bytes
= sizeof(struct page
*)*want
;
381 new_pages
= kzalloc(bytes
, GFP_KERNEL
);
383 new_pages
= vzalloc(bytes
);
390 for (i
= 0; i
< have
; i
++)
391 new_pages
[i
] = old_pages
[i
];
392 for (; i
< want
; i
++) {
393 page
= alloc_page(GFP_HIGHUSER
);
395 bm_free_pages(new_pages
+ have
, i
- have
);
396 bm_vk_free(new_pages
, vmalloced
);
399 /* we want to know which page it is
400 * from the endio handlers */
401 bm_store_page_idx(page
, i
);
405 for (i
= 0; i
< want
; i
++)
406 new_pages
[i
] = old_pages
[i
];
407 /* NOT HERE, we are outside the spinlock!
408 bm_free_pages(old_pages + want, have - want);
413 b
->bm_flags
|= BM_P_VMALLOCED
;
415 b
->bm_flags
&= ~BM_P_VMALLOCED
;
421 * called on driver init only. TODO call when a device is created.
422 * allocates the drbd_bitmap, and stores it in mdev->bitmap.
424 int drbd_bm_init(struct drbd_conf
*mdev
)
426 struct drbd_bitmap
*b
= mdev
->bitmap
;
428 b
= kzalloc(sizeof(struct drbd_bitmap
), GFP_KERNEL
);
431 spin_lock_init(&b
->bm_lock
);
432 mutex_init(&b
->bm_change
);
433 init_waitqueue_head(&b
->bm_io_wait
);
440 sector_t
drbd_bm_capacity(struct drbd_conf
*mdev
)
442 ERR_IF(!mdev
->bitmap
) return 0;
443 return mdev
->bitmap
->bm_dev_capacity
;
446 /* called on driver unload. TODO: call when a device is destroyed.
448 void drbd_bm_cleanup(struct drbd_conf
*mdev
)
450 ERR_IF (!mdev
->bitmap
) return;
451 bm_free_pages(mdev
->bitmap
->bm_pages
, mdev
->bitmap
->bm_number_of_pages
);
452 bm_vk_free(mdev
->bitmap
->bm_pages
, (BM_P_VMALLOCED
& mdev
->bitmap
->bm_flags
));
458 * since (b->bm_bits % BITS_PER_LONG) != 0,
459 * this masks out the remaining bits.
460 * Returns the number of bits cleared.
462 #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3))
463 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1)
464 #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1)
465 static int bm_clear_surplus(struct drbd_bitmap
*b
)
468 unsigned long *p_addr
, *bm
;
472 /* number of bits modulo bits per page */
473 tmp
= (b
->bm_bits
& BITS_PER_PAGE_MASK
);
474 /* mask the used bits of the word containing the last bit */
475 mask
= (1UL << (tmp
& BITS_PER_LONG_MASK
)) -1;
476 /* bitmap is always stored little endian,
477 * on disk and in core memory alike */
478 mask
= cpu_to_lel(mask
);
480 p_addr
= bm_map_pidx(b
, b
->bm_number_of_pages
- 1);
481 bm
= p_addr
+ (tmp
/BITS_PER_LONG
);
483 /* If mask != 0, we are not exactly aligned, so bm now points
484 * to the long containing the last bit.
485 * If mask == 0, bm already points to the word immediately
486 * after the last (long word aligned) bit. */
487 cleared
= hweight_long(*bm
& ~mask
);
492 if (BITS_PER_LONG
== 32 && ((bm
- p_addr
) & 1) == 1) {
493 /* on a 32bit arch, we may need to zero out
494 * a padding long to align with a 64bit remote */
495 cleared
+= hweight_long(*bm
);
502 static void bm_set_surplus(struct drbd_bitmap
*b
)
505 unsigned long *p_addr
, *bm
;
508 /* number of bits modulo bits per page */
509 tmp
= (b
->bm_bits
& BITS_PER_PAGE_MASK
);
510 /* mask the used bits of the word containing the last bit */
511 mask
= (1UL << (tmp
& BITS_PER_LONG_MASK
)) -1;
512 /* bitmap is always stored little endian,
513 * on disk and in core memory alike */
514 mask
= cpu_to_lel(mask
);
516 p_addr
= bm_map_pidx(b
, b
->bm_number_of_pages
- 1);
517 bm
= p_addr
+ (tmp
/BITS_PER_LONG
);
519 /* If mask != 0, we are not exactly aligned, so bm now points
520 * to the long containing the last bit.
521 * If mask == 0, bm already points to the word immediately
522 * after the last (long word aligned) bit. */
527 if (BITS_PER_LONG
== 32 && ((bm
- p_addr
) & 1) == 1) {
528 /* on a 32bit arch, we may need to zero out
529 * a padding long to align with a 64bit remote */
535 /* you better not modify the bitmap while this is running,
536 * or its results will be stale */
537 static unsigned long bm_count_bits(struct drbd_bitmap
*b
)
539 unsigned long *p_addr
;
540 unsigned long bits
= 0;
541 unsigned long mask
= (1UL << (b
->bm_bits
& BITS_PER_LONG_MASK
)) -1;
542 int idx
, i
, last_word
;
544 /* all but last page */
545 for (idx
= 0; idx
< b
->bm_number_of_pages
- 1; idx
++) {
546 p_addr
= __bm_map_pidx(b
, idx
);
547 for (i
= 0; i
< LWPP
; i
++)
548 bits
+= hweight_long(p_addr
[i
]);
552 /* last (or only) page */
553 last_word
= ((b
->bm_bits
- 1) & BITS_PER_PAGE_MASK
) >> LN2_BPL
;
554 p_addr
= __bm_map_pidx(b
, idx
);
555 for (i
= 0; i
< last_word
; i
++)
556 bits
+= hweight_long(p_addr
[i
]);
557 p_addr
[last_word
] &= cpu_to_lel(mask
);
558 bits
+= hweight_long(p_addr
[last_word
]);
559 /* 32bit arch, may have an unused padding long */
560 if (BITS_PER_LONG
== 32 && (last_word
& 1) == 0)
561 p_addr
[last_word
+1] = 0;
566 /* offset and len in long words.*/
567 static void bm_memset(struct drbd_bitmap
*b
, size_t offset
, int c
, size_t len
)
569 unsigned long *p_addr
, *bm
;
575 if (end
> b
->bm_words
) {
576 printk(KERN_ALERT
"drbd: bm_memset end > bm_words\n");
580 while (offset
< end
) {
581 do_now
= min_t(size_t, ALIGN(offset
+ 1, LWPP
), end
) - offset
;
582 idx
= bm_word_to_page_idx(b
, offset
);
583 p_addr
= bm_map_pidx(b
, idx
);
584 bm
= p_addr
+ MLPP(offset
);
585 if (bm
+do_now
> p_addr
+ LWPP
) {
586 printk(KERN_ALERT
"drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n",
587 p_addr
, bm
, (int)do_now
);
589 memset(bm
, c
, do_now
* sizeof(long));
591 bm_set_page_need_writeout(b
->bm_pages
[idx
]);
597 * make sure the bitmap has enough room for the attached storage,
598 * if necessary, resize.
599 * called whenever we may have changed the device size.
600 * returns -ENOMEM if we could not allocate enough memory, 0 on success.
601 * In case this is actually a resize, we copy the old bitmap into the new one.
602 * Otherwise, the bitmap is initialized to all bits set.
604 int drbd_bm_resize(struct drbd_conf
*mdev
, sector_t capacity
, int set_new_bits
)
606 struct drbd_bitmap
*b
= mdev
->bitmap
;
607 unsigned long bits
, words
, owords
, obits
;
608 unsigned long want
, have
, onpages
; /* number of pages */
609 struct page
**npages
, **opages
= NULL
;
610 int err
= 0, growing
;
611 int opages_vmalloced
;
613 ERR_IF(!b
) return -ENOMEM
;
615 drbd_bm_lock(mdev
, "resize", BM_LOCKED_MASK
);
617 dev_info(DEV
, "drbd_bm_resize called with capacity == %llu\n",
618 (unsigned long long)capacity
);
620 if (capacity
== b
->bm_dev_capacity
)
623 opages_vmalloced
= (BM_P_VMALLOCED
& b
->bm_flags
);
626 spin_lock_irq(&b
->bm_lock
);
627 opages
= b
->bm_pages
;
628 onpages
= b
->bm_number_of_pages
;
629 owords
= b
->bm_words
;
631 b
->bm_number_of_pages
=
635 b
->bm_dev_capacity
= 0;
636 spin_unlock_irq(&b
->bm_lock
);
637 bm_free_pages(opages
, onpages
);
638 bm_vk_free(opages
, opages_vmalloced
);
641 bits
= BM_SECT_TO_BIT(ALIGN(capacity
, BM_SECT_PER_BIT
));
644 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL;
645 a 32bit host could present the wrong number of words
648 words
= ALIGN(bits
, 64) >> LN2_BPL
;
650 if (get_ldev(mdev
)) {
651 u64 bits_on_disk
= ((u64
)mdev
->ldev
->md
.md_size_sect
-MD_BM_OFFSET
) << 12;
653 if (bits
> bits_on_disk
) {
654 dev_info(DEV
, "bits = %lu\n", bits
);
655 dev_info(DEV
, "bits_on_disk = %llu\n", bits_on_disk
);
661 want
= ALIGN(words
*sizeof(long), PAGE_SIZE
) >> PAGE_SHIFT
;
662 have
= b
->bm_number_of_pages
;
664 D_ASSERT(b
->bm_pages
!= NULL
);
665 npages
= b
->bm_pages
;
667 if (drbd_insert_fault(mdev
, DRBD_FAULT_BM_ALLOC
))
670 npages
= bm_realloc_pages(b
, want
);
678 spin_lock_irq(&b
->bm_lock
);
679 opages
= b
->bm_pages
;
680 owords
= b
->bm_words
;
683 growing
= bits
> obits
;
684 if (opages
&& growing
&& set_new_bits
)
687 b
->bm_pages
= npages
;
688 b
->bm_number_of_pages
= want
;
691 b
->bm_dev_capacity
= capacity
;
695 bm_memset(b
, owords
, 0xff, words
-owords
);
696 b
->bm_set
+= bits
- obits
;
698 bm_memset(b
, owords
, 0x00, words
-owords
);
703 /* implicit: (opages != NULL) && (opages != npages) */
704 bm_free_pages(opages
+ want
, have
- want
);
707 (void)bm_clear_surplus(b
);
709 spin_unlock_irq(&b
->bm_lock
);
710 if (opages
!= npages
)
711 bm_vk_free(opages
, opages_vmalloced
);
713 b
->bm_set
= bm_count_bits(b
);
714 dev_info(DEV
, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits
, words
, want
);
717 drbd_bm_unlock(mdev
);
722 * if not protected by other means, return value may be out of date when
723 * leaving this function...
724 * we still need to lock it, since it is important that this returns
725 * bm_set == 0 precisely.
727 * maybe bm_set should be atomic_t ?
729 unsigned long _drbd_bm_total_weight(struct drbd_conf
*mdev
)
731 struct drbd_bitmap
*b
= mdev
->bitmap
;
736 ERR_IF(!b
->bm_pages
) return 0;
738 spin_lock_irqsave(&b
->bm_lock
, flags
);
740 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
745 unsigned long drbd_bm_total_weight(struct drbd_conf
*mdev
)
748 /* if I don't have a disk, I don't know about out-of-sync status */
749 if (!get_ldev_if_state(mdev
, D_NEGOTIATING
))
751 s
= _drbd_bm_total_weight(mdev
);
756 size_t drbd_bm_words(struct drbd_conf
*mdev
)
758 struct drbd_bitmap
*b
= mdev
->bitmap
;
760 ERR_IF(!b
->bm_pages
) return 0;
765 unsigned long drbd_bm_bits(struct drbd_conf
*mdev
)
767 struct drbd_bitmap
*b
= mdev
->bitmap
;
773 /* merge number words from buffer into the bitmap starting at offset.
774 * buffer[i] is expected to be little endian unsigned long.
775 * bitmap must be locked by drbd_bm_lock.
776 * currently only used from receive_bitmap.
778 void drbd_bm_merge_lel(struct drbd_conf
*mdev
, size_t offset
, size_t number
,
779 unsigned long *buffer
)
781 struct drbd_bitmap
*b
= mdev
->bitmap
;
782 unsigned long *p_addr
, *bm
;
783 unsigned long word
, bits
;
787 end
= offset
+ number
;
790 ERR_IF(!b
->bm_pages
) return;
793 WARN_ON(offset
>= b
->bm_words
);
794 WARN_ON(end
> b
->bm_words
);
796 spin_lock_irq(&b
->bm_lock
);
797 while (offset
< end
) {
798 do_now
= min_t(size_t, ALIGN(offset
+1, LWPP
), end
) - offset
;
799 idx
= bm_word_to_page_idx(b
, offset
);
800 p_addr
= bm_map_pidx(b
, idx
);
801 bm
= p_addr
+ MLPP(offset
);
804 bits
= hweight_long(*bm
);
805 word
= *bm
| *buffer
++;
807 b
->bm_set
+= hweight_long(word
) - bits
;
810 bm_set_page_need_writeout(b
->bm_pages
[idx
]);
812 /* with 32bit <-> 64bit cross-platform connect
813 * this is only correct for current usage,
814 * where we _know_ that we are 64 bit aligned,
815 * and know that this function is used in this way, too...
817 if (end
== b
->bm_words
)
818 b
->bm_set
-= bm_clear_surplus(b
);
819 spin_unlock_irq(&b
->bm_lock
);
822 /* copy number words from the bitmap starting at offset into the buffer.
823 * buffer[i] will be little endian unsigned long.
825 void drbd_bm_get_lel(struct drbd_conf
*mdev
, size_t offset
, size_t number
,
826 unsigned long *buffer
)
828 struct drbd_bitmap
*b
= mdev
->bitmap
;
829 unsigned long *p_addr
, *bm
;
832 end
= offset
+ number
;
835 ERR_IF(!b
->bm_pages
) return;
837 spin_lock_irq(&b
->bm_lock
);
838 if ((offset
>= b
->bm_words
) ||
839 (end
> b
->bm_words
) ||
841 dev_err(DEV
, "offset=%lu number=%lu bm_words=%lu\n",
842 (unsigned long) offset
,
843 (unsigned long) number
,
844 (unsigned long) b
->bm_words
);
846 while (offset
< end
) {
847 do_now
= min_t(size_t, ALIGN(offset
+1, LWPP
), end
) - offset
;
848 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, offset
));
849 bm
= p_addr
+ MLPP(offset
);
856 spin_unlock_irq(&b
->bm_lock
);
859 /* set all bits in the bitmap */
860 void drbd_bm_set_all(struct drbd_conf
*mdev
)
862 struct drbd_bitmap
*b
= mdev
->bitmap
;
864 ERR_IF(!b
->bm_pages
) return;
866 spin_lock_irq(&b
->bm_lock
);
867 bm_memset(b
, 0, 0xff, b
->bm_words
);
868 (void)bm_clear_surplus(b
);
869 b
->bm_set
= b
->bm_bits
;
870 spin_unlock_irq(&b
->bm_lock
);
873 /* clear all bits in the bitmap */
874 void drbd_bm_clear_all(struct drbd_conf
*mdev
)
876 struct drbd_bitmap
*b
= mdev
->bitmap
;
878 ERR_IF(!b
->bm_pages
) return;
880 spin_lock_irq(&b
->bm_lock
);
881 bm_memset(b
, 0, 0, b
->bm_words
);
883 spin_unlock_irq(&b
->bm_lock
);
887 struct drbd_conf
*mdev
;
891 #define BM_AIO_COPY_PAGES 1
896 static void bm_aio_ctx_destroy(struct kref
*kref
)
898 struct bm_aio_ctx
*ctx
= container_of(kref
, struct bm_aio_ctx
, kref
);
904 /* bv_page may be a copy, or may be the original */
905 static void bm_async_io_complete(struct bio
*bio
, int error
)
907 struct bm_aio_ctx
*ctx
= bio
->bi_private
;
908 struct drbd_conf
*mdev
= ctx
->mdev
;
909 struct drbd_bitmap
*b
= mdev
->bitmap
;
910 unsigned int idx
= bm_page_to_idx(bio
->bi_io_vec
[0].bv_page
);
911 int uptodate
= bio_flagged(bio
, BIO_UPTODATE
);
914 /* strange behavior of some lower level drivers...
915 * fail the request by clearing the uptodate flag,
916 * but do not return any error?!
917 * do we want to WARN() on this? */
918 if (!error
&& !uptodate
)
921 if ((ctx
->flags
& BM_AIO_COPY_PAGES
) == 0 &&
922 !bm_test_page_unchanged(b
->bm_pages
[idx
]))
923 dev_warn(DEV
, "bitmap page idx %u changed during IO!\n", idx
);
926 /* ctx error will hold the completed-last non-zero error code,
927 * in case error codes differ. */
929 bm_set_page_io_err(b
->bm_pages
[idx
]);
930 /* Not identical to on disk version of it.
931 * Is BM_PAGE_IO_ERROR enough? */
932 if (__ratelimit(&drbd_ratelimit_state
))
933 dev_err(DEV
, "IO ERROR %d on bitmap page idx %u\n",
936 bm_clear_page_io_err(b
->bm_pages
[idx
]);
937 dynamic_dev_dbg(DEV
, "bitmap page idx %u completed\n", idx
);
940 bm_page_unlock_io(mdev
, idx
);
942 if (ctx
->flags
& BM_AIO_COPY_PAGES
)
943 mempool_free(bio
->bi_io_vec
[0].bv_page
, drbd_md_io_page_pool
);
947 if (atomic_dec_and_test(&ctx
->in_flight
)) {
949 wake_up(&mdev
->misc_wait
);
950 kref_put(&ctx
->kref
, &bm_aio_ctx_destroy
);
954 static void bm_page_io_async(struct bm_aio_ctx
*ctx
, int page_nr
, int rw
) __must_hold(local
)
956 struct bio
*bio
= bio_alloc_drbd(GFP_NOIO
);
957 struct drbd_conf
*mdev
= ctx
->mdev
;
958 struct drbd_bitmap
*b
= mdev
->bitmap
;
962 sector_t on_disk_sector
=
963 mdev
->ldev
->md
.md_offset
+ mdev
->ldev
->md
.bm_offset
;
964 on_disk_sector
+= ((sector_t
)page_nr
) << (PAGE_SHIFT
-9);
966 /* this might happen with very small
967 * flexible external meta data device,
968 * or with PAGE_SIZE > 4k */
969 len
= min_t(unsigned int, PAGE_SIZE
,
970 (drbd_md_last_sector(mdev
->ldev
) - on_disk_sector
+ 1)<<9);
972 /* serialize IO on this page */
973 bm_page_lock_io(mdev
, page_nr
);
974 /* before memcpy and submit,
975 * so it can be redirtied any time */
976 bm_set_page_unchanged(b
->bm_pages
[page_nr
]);
978 if (ctx
->flags
& BM_AIO_COPY_PAGES
) {
980 page
= mempool_alloc(drbd_md_io_page_pool
, __GFP_HIGHMEM
|__GFP_WAIT
);
981 dest
= kmap_atomic(page
);
982 src
= kmap_atomic(b
->bm_pages
[page_nr
]);
983 memcpy(dest
, src
, PAGE_SIZE
);
986 bm_store_page_idx(page
, page_nr
);
988 page
= b
->bm_pages
[page_nr
];
990 bio
->bi_bdev
= mdev
->ldev
->md_bdev
;
991 bio
->bi_sector
= on_disk_sector
;
992 /* bio_add_page of a single page to an empty bio will always succeed,
993 * according to api. Do we want to assert that? */
994 bio_add_page(bio
, page
, len
, 0);
995 bio
->bi_private
= ctx
;
996 bio
->bi_end_io
= bm_async_io_complete
;
998 if (drbd_insert_fault(mdev
, (rw
& WRITE
) ? DRBD_FAULT_MD_WR
: DRBD_FAULT_MD_RD
)) {
1000 bio_endio(bio
, -EIO
);
1002 submit_bio(rw
, bio
);
1003 /* this should not count as user activity and cause the
1004 * resync to throttle -- see drbd_rs_should_slow_down(). */
1005 atomic_add(len
>> 9, &mdev
->rs_sect_ev
);
1010 * bm_rw: read/write the whole bitmap from/to its on disk location.
1012 static int bm_rw(struct drbd_conf
*mdev
, int rw
, unsigned flags
, unsigned lazy_writeout_upper_idx
) __must_hold(local
)
1014 struct bm_aio_ctx
*ctx
;
1015 struct drbd_bitmap
*b
= mdev
->bitmap
;
1016 int num_pages
, i
, count
= 0;
1022 * We are protected against bitmap disappearing/resizing by holding an
1023 * ldev reference (caller must have called get_ldev()).
1024 * For read/write, we are protected against changes to the bitmap by
1025 * the bitmap lock (see drbd_bitmap_io).
1026 * For lazy writeout, we don't care for ongoing changes to the bitmap,
1027 * as we submit copies of pages anyways.
1030 ctx
= kmalloc(sizeof(struct bm_aio_ctx
), GFP_NOIO
);
1034 *ctx
= (struct bm_aio_ctx
) {
1036 .in_flight
= ATOMIC_INIT(1),
1040 .kref
= { ATOMIC_INIT(2) },
1043 if (!get_ldev_if_state(mdev
, D_ATTACHING
)) { /* put is in bm_aio_ctx_destroy() */
1044 dev_err(DEV
, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n");
1050 WARN_ON(!(BM_LOCKED_MASK
& b
->bm_flags
));
1052 num_pages
= b
->bm_number_of_pages
;
1056 /* let the layers below us try to merge these bios... */
1057 for (i
= 0; i
< num_pages
; i
++) {
1058 /* ignore completely unchanged pages */
1059 if (lazy_writeout_upper_idx
&& i
== lazy_writeout_upper_idx
)
1062 if (bm_test_page_unchanged(b
->bm_pages
[i
])) {
1063 dynamic_dev_dbg(DEV
, "skipped bm write for idx %u\n", i
);
1066 /* during lazy writeout,
1067 * ignore those pages not marked for lazy writeout. */
1068 if (lazy_writeout_upper_idx
&&
1069 !bm_test_page_lazy_writeout(b
->bm_pages
[i
])) {
1070 dynamic_dev_dbg(DEV
, "skipped bm lazy write for idx %u\n", i
);
1074 atomic_inc(&ctx
->in_flight
);
1075 bm_page_io_async(ctx
, i
, rw
);
1081 * We initialize ctx->in_flight to one to make sure bm_async_io_complete
1082 * will not set ctx->done early, and decrement / test it here. If there
1083 * are still some bios in flight, we need to wait for them here.
1084 * If all IO is done already (or nothing had been submitted), there is
1085 * no need to wait. Still, we need to put the kref associated with the
1086 * "in_flight reached zero, all done" event.
1088 if (!atomic_dec_and_test(&ctx
->in_flight
))
1089 wait_until_done_or_disk_failure(mdev
, mdev
->ldev
, &ctx
->done
);
1091 kref_put(&ctx
->kref
, &bm_aio_ctx_destroy
);
1093 dev_info(DEV
, "bitmap %s of %u pages took %lu jiffies\n",
1094 rw
== WRITE
? "WRITE" : "READ",
1095 count
, jiffies
- now
);
1098 dev_alert(DEV
, "we had at least one MD IO ERROR during bitmap IO\n");
1099 drbd_chk_io_error(mdev
, 1, true);
1100 err
= -EIO
; /* ctx->error ? */
1103 if (atomic_read(&ctx
->in_flight
))
1104 err
= -EIO
; /* Disk failed during IO... */
1108 drbd_md_flush(mdev
);
1109 } else /* rw == READ */ {
1110 b
->bm_set
= bm_count_bits(b
);
1111 dev_info(DEV
, "recounting of set bits took additional %lu jiffies\n",
1116 dev_info(DEV
, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n",
1117 ppsize(ppb
, now
<< (BM_BLOCK_SHIFT
-10)), now
);
1119 kref_put(&ctx
->kref
, &bm_aio_ctx_destroy
);
1124 * drbd_bm_read() - Read the whole bitmap from its on disk location.
1125 * @mdev: DRBD device.
1127 int drbd_bm_read(struct drbd_conf
*mdev
) __must_hold(local
)
1129 return bm_rw(mdev
, READ
, 0, 0);
1133 * drbd_bm_write() - Write the whole bitmap to its on disk location.
1134 * @mdev: DRBD device.
1136 * Will only write pages that have changed since last IO.
1138 int drbd_bm_write(struct drbd_conf
*mdev
) __must_hold(local
)
1140 return bm_rw(mdev
, WRITE
, 0, 0);
1144 * drbd_bm_lazy_write_out() - Write bitmap pages 0 to @upper_idx-1, if they have changed.
1145 * @mdev: DRBD device.
1146 * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages
1148 int drbd_bm_write_lazy(struct drbd_conf
*mdev
, unsigned upper_idx
) __must_hold(local
)
1150 return bm_rw(mdev
, WRITE
, BM_AIO_COPY_PAGES
, upper_idx
);
1154 * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location.
1155 * @mdev: DRBD device.
1157 * Will only write pages that have changed since last IO.
1158 * In contrast to drbd_bm_write(), this will copy the bitmap pages
1159 * to temporary writeout pages. It is intended to trigger a full write-out
1160 * while still allowing the bitmap to change, for example if a resync or online
1161 * verify is aborted due to a failed peer disk, while local IO continues, or
1162 * pending resync acks are still being processed.
1164 int drbd_bm_write_copy_pages(struct drbd_conf
*mdev
) __must_hold(local
)
1166 return bm_rw(mdev
, WRITE
, BM_AIO_COPY_PAGES
, 0);
1171 * drbd_bm_write_page: Writes a PAGE_SIZE aligned piece of bitmap
1172 * @mdev: DRBD device.
1173 * @idx: bitmap page index
1175 * We don't want to special case on logical_block_size of the backend device,
1176 * so we submit PAGE_SIZE aligned pieces.
1177 * Note that on "most" systems, PAGE_SIZE is 4k.
1179 * In case this becomes an issue on systems with larger PAGE_SIZE,
1180 * we may want to change this again to write 4k aligned 4k pieces.
1182 int drbd_bm_write_page(struct drbd_conf
*mdev
, unsigned int idx
) __must_hold(local
)
1184 struct bm_aio_ctx
*ctx
;
1187 if (bm_test_page_unchanged(mdev
->bitmap
->bm_pages
[idx
])) {
1188 dynamic_dev_dbg(DEV
, "skipped bm page write for idx %u\n", idx
);
1192 ctx
= kmalloc(sizeof(struct bm_aio_ctx
), GFP_NOIO
);
1196 *ctx
= (struct bm_aio_ctx
) {
1198 .in_flight
= ATOMIC_INIT(1),
1200 .flags
= BM_AIO_COPY_PAGES
,
1202 .kref
= { ATOMIC_INIT(2) },
1205 if (!get_ldev_if_state(mdev
, D_ATTACHING
)) { /* put is in bm_aio_ctx_destroy() */
1206 dev_err(DEV
, "ASSERT FAILED: get_ldev_if_state() == 1 in drbd_bm_write_page()\n");
1211 bm_page_io_async(ctx
, idx
, WRITE_SYNC
);
1212 wait_until_done_or_disk_failure(mdev
, mdev
->ldev
, &ctx
->done
);
1215 drbd_chk_io_error(mdev
, 1, true);
1216 /* that should force detach, so the in memory bitmap will be
1217 * gone in a moment as well. */
1219 mdev
->bm_writ_cnt
++;
1220 err
= atomic_read(&ctx
->in_flight
) ? -EIO
: ctx
->error
;
1221 kref_put(&ctx
->kref
, &bm_aio_ctx_destroy
);
1226 * find_first_bit returns int, we return unsigned long.
1227 * For this to work on 32bit arch with bitnumbers > (1<<32),
1228 * we'd need to return u64, and get a whole lot of other places
1229 * fixed where we still use unsigned long.
1231 * this returns a bit number, NOT a sector!
1233 static unsigned long __bm_find_next(struct drbd_conf
*mdev
, unsigned long bm_fo
,
1234 const int find_zero_bit
)
1236 struct drbd_bitmap
*b
= mdev
->bitmap
;
1237 unsigned long *p_addr
;
1238 unsigned long bit_offset
;
1242 if (bm_fo
> b
->bm_bits
) {
1243 dev_err(DEV
, "bm_fo=%lu bm_bits=%lu\n", bm_fo
, b
->bm_bits
);
1244 bm_fo
= DRBD_END_OF_BITMAP
;
1246 while (bm_fo
< b
->bm_bits
) {
1247 /* bit offset of the first bit in the page */
1248 bit_offset
= bm_fo
& ~BITS_PER_PAGE_MASK
;
1249 p_addr
= __bm_map_pidx(b
, bm_bit_to_page_idx(b
, bm_fo
));
1252 i
= find_next_zero_bit_le(p_addr
,
1253 PAGE_SIZE
*8, bm_fo
& BITS_PER_PAGE_MASK
);
1255 i
= find_next_bit_le(p_addr
,
1256 PAGE_SIZE
*8, bm_fo
& BITS_PER_PAGE_MASK
);
1259 if (i
< PAGE_SIZE
*8) {
1260 bm_fo
= bit_offset
+ i
;
1261 if (bm_fo
>= b
->bm_bits
)
1265 bm_fo
= bit_offset
+ PAGE_SIZE
*8;
1267 bm_fo
= DRBD_END_OF_BITMAP
;
1273 static unsigned long bm_find_next(struct drbd_conf
*mdev
,
1274 unsigned long bm_fo
, const int find_zero_bit
)
1276 struct drbd_bitmap
*b
= mdev
->bitmap
;
1277 unsigned long i
= DRBD_END_OF_BITMAP
;
1279 ERR_IF(!b
) return i
;
1280 ERR_IF(!b
->bm_pages
) return i
;
1282 spin_lock_irq(&b
->bm_lock
);
1283 if (BM_DONT_TEST
& b
->bm_flags
)
1284 bm_print_lock_info(mdev
);
1286 i
= __bm_find_next(mdev
, bm_fo
, find_zero_bit
);
1288 spin_unlock_irq(&b
->bm_lock
);
1292 unsigned long drbd_bm_find_next(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1294 return bm_find_next(mdev
, bm_fo
, 0);
1298 /* not yet needed for anything. */
1299 unsigned long drbd_bm_find_next_zero(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1301 return bm_find_next(mdev
, bm_fo
, 1);
1305 /* does not spin_lock_irqsave.
1306 * you must take drbd_bm_lock() first */
1307 unsigned long _drbd_bm_find_next(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1309 /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */
1310 return __bm_find_next(mdev
, bm_fo
, 0);
1313 unsigned long _drbd_bm_find_next_zero(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1315 /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */
1316 return __bm_find_next(mdev
, bm_fo
, 1);
1319 /* returns number of bits actually changed.
1320 * for val != 0, we change 0 -> 1, return code positive
1321 * for val == 0, we change 1 -> 0, return code negative
1322 * wants bitnr, not sector.
1323 * expected to be called for only a few bits (e - s about BITS_PER_LONG).
1324 * Must hold bitmap lock already. */
1325 static int __bm_change_bits_to(struct drbd_conf
*mdev
, const unsigned long s
,
1326 unsigned long e
, int val
)
1328 struct drbd_bitmap
*b
= mdev
->bitmap
;
1329 unsigned long *p_addr
= NULL
;
1330 unsigned long bitnr
;
1331 unsigned int last_page_nr
= -1U;
1333 int changed_total
= 0;
1335 if (e
>= b
->bm_bits
) {
1336 dev_err(DEV
, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n",
1338 e
= b
->bm_bits
? b
->bm_bits
-1 : 0;
1340 for (bitnr
= s
; bitnr
<= e
; bitnr
++) {
1341 unsigned int page_nr
= bm_bit_to_page_idx(b
, bitnr
);
1342 if (page_nr
!= last_page_nr
) {
1346 bm_set_page_lazy_writeout(b
->bm_pages
[last_page_nr
]);
1348 bm_set_page_need_writeout(b
->bm_pages
[last_page_nr
]);
1351 p_addr
= __bm_map_pidx(b
, page_nr
);
1352 last_page_nr
= page_nr
;
1355 c
+= (0 == __test_and_set_bit_le(bitnr
& BITS_PER_PAGE_MASK
, p_addr
));
1357 c
-= (0 != __test_and_clear_bit_le(bitnr
& BITS_PER_PAGE_MASK
, p_addr
));
1362 bm_set_page_lazy_writeout(b
->bm_pages
[last_page_nr
]);
1364 bm_set_page_need_writeout(b
->bm_pages
[last_page_nr
]);
1366 b
->bm_set
+= changed_total
;
1367 return changed_total
;
1370 /* returns number of bits actually changed.
1371 * for val != 0, we change 0 -> 1, return code positive
1372 * for val == 0, we change 1 -> 0, return code negative
1373 * wants bitnr, not sector */
1374 static int bm_change_bits_to(struct drbd_conf
*mdev
, const unsigned long s
,
1375 const unsigned long e
, int val
)
1377 unsigned long flags
;
1378 struct drbd_bitmap
*b
= mdev
->bitmap
;
1381 ERR_IF(!b
) return 1;
1382 ERR_IF(!b
->bm_pages
) return 0;
1384 spin_lock_irqsave(&b
->bm_lock
, flags
);
1385 if ((val
? BM_DONT_SET
: BM_DONT_CLEAR
) & b
->bm_flags
)
1386 bm_print_lock_info(mdev
);
1388 c
= __bm_change_bits_to(mdev
, s
, e
, val
);
1390 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1394 /* returns number of bits changed 0 -> 1 */
1395 int drbd_bm_set_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1397 return bm_change_bits_to(mdev
, s
, e
, 1);
1400 /* returns number of bits changed 1 -> 0 */
1401 int drbd_bm_clear_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1403 return -bm_change_bits_to(mdev
, s
, e
, 0);
1406 /* sets all bits in full words,
1407 * from first_word up to, but not including, last_word */
1408 static inline void bm_set_full_words_within_one_page(struct drbd_bitmap
*b
,
1409 int page_nr
, int first_word
, int last_word
)
1413 unsigned long *paddr
= kmap_atomic(b
->bm_pages
[page_nr
]);
1414 for (i
= first_word
; i
< last_word
; i
++) {
1415 bits
= hweight_long(paddr
[i
]);
1417 b
->bm_set
+= BITS_PER_LONG
- bits
;
1419 kunmap_atomic(paddr
);
1422 /* Same thing as drbd_bm_set_bits,
1423 * but more efficient for a large bit range.
1424 * You must first drbd_bm_lock().
1425 * Can be called to set the whole bitmap in one go.
1426 * Sets bits from s to e _inclusive_. */
1427 void _drbd_bm_set_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1429 /* First set_bit from the first bit (s)
1430 * up to the next long boundary (sl),
1431 * then assign full words up to the last long boundary (el),
1432 * then set_bit up to and including the last bit (e).
1434 * Do not use memset, because we must account for changes,
1435 * so we need to loop over the words with hweight() anyways.
1437 struct drbd_bitmap
*b
= mdev
->bitmap
;
1438 unsigned long sl
= ALIGN(s
,BITS_PER_LONG
);
1439 unsigned long el
= (e
+1) & ~((unsigned long)BITS_PER_LONG
-1);
1446 if (e
- s
<= 3*BITS_PER_LONG
) {
1447 /* don't bother; el and sl may even be wrong. */
1448 spin_lock_irq(&b
->bm_lock
);
1449 __bm_change_bits_to(mdev
, s
, e
, 1);
1450 spin_unlock_irq(&b
->bm_lock
);
1454 /* difference is large enough that we can trust sl and el */
1456 spin_lock_irq(&b
->bm_lock
);
1458 /* bits filling the current long */
1460 __bm_change_bits_to(mdev
, s
, sl
-1, 1);
1462 first_page
= sl
>> (3 + PAGE_SHIFT
);
1463 last_page
= el
>> (3 + PAGE_SHIFT
);
1465 /* MLPP: modulo longs per page */
1466 /* LWPP: long words per page */
1467 first_word
= MLPP(sl
>> LN2_BPL
);
1470 /* first and full pages, unless first page == last page */
1471 for (page_nr
= first_page
; page_nr
< last_page
; page_nr
++) {
1472 bm_set_full_words_within_one_page(mdev
->bitmap
, page_nr
, first_word
, last_word
);
1473 spin_unlock_irq(&b
->bm_lock
);
1476 spin_lock_irq(&b
->bm_lock
);
1478 /* last page (respectively only page, for first page == last page) */
1479 last_word
= MLPP(el
>> LN2_BPL
);
1481 /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples).
1482 * ==> e = 32767, el = 32768, last_page = 2,
1483 * and now last_word = 0.
1484 * We do not want to touch last_page in this case,
1485 * as we did not allocate it, it is not present in bitmap->bm_pages.
1488 bm_set_full_words_within_one_page(mdev
->bitmap
, last_page
, first_word
, last_word
);
1490 /* possibly trailing bits.
1491 * example: (e & 63) == 63, el will be e+1.
1492 * if that even was the very last bit,
1493 * it would trigger an assert in __bm_change_bits_to()
1496 __bm_change_bits_to(mdev
, el
, e
, 1);
1497 spin_unlock_irq(&b
->bm_lock
);
1500 /* returns bit state
1501 * wants bitnr, NOT sector.
1502 * inherently racy... area needs to be locked by means of {al,rs}_lru
1505 * -1 ... first out of bounds access, stop testing for bits!
1507 int drbd_bm_test_bit(struct drbd_conf
*mdev
, const unsigned long bitnr
)
1509 unsigned long flags
;
1510 struct drbd_bitmap
*b
= mdev
->bitmap
;
1511 unsigned long *p_addr
;
1514 ERR_IF(!b
) return 0;
1515 ERR_IF(!b
->bm_pages
) return 0;
1517 spin_lock_irqsave(&b
->bm_lock
, flags
);
1518 if (BM_DONT_TEST
& b
->bm_flags
)
1519 bm_print_lock_info(mdev
);
1520 if (bitnr
< b
->bm_bits
) {
1521 p_addr
= bm_map_pidx(b
, bm_bit_to_page_idx(b
, bitnr
));
1522 i
= test_bit_le(bitnr
& BITS_PER_PAGE_MASK
, p_addr
) ? 1 : 0;
1524 } else if (bitnr
== b
->bm_bits
) {
1526 } else { /* (bitnr > b->bm_bits) */
1527 dev_err(DEV
, "bitnr=%lu > bm_bits=%lu\n", bitnr
, b
->bm_bits
);
1531 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1535 /* returns number of bits set in the range [s, e] */
1536 int drbd_bm_count_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1538 unsigned long flags
;
1539 struct drbd_bitmap
*b
= mdev
->bitmap
;
1540 unsigned long *p_addr
= NULL
;
1541 unsigned long bitnr
;
1542 unsigned int page_nr
= -1U;
1545 /* If this is called without a bitmap, that is a bug. But just to be
1546 * robust in case we screwed up elsewhere, in that case pretend there
1547 * was one dirty bit in the requested area, so we won't try to do a
1548 * local read there (no bitmap probably implies no disk) */
1549 ERR_IF(!b
) return 1;
1550 ERR_IF(!b
->bm_pages
) return 1;
1552 spin_lock_irqsave(&b
->bm_lock
, flags
);
1553 if (BM_DONT_TEST
& b
->bm_flags
)
1554 bm_print_lock_info(mdev
);
1555 for (bitnr
= s
; bitnr
<= e
; bitnr
++) {
1556 unsigned int idx
= bm_bit_to_page_idx(b
, bitnr
);
1557 if (page_nr
!= idx
) {
1561 p_addr
= bm_map_pidx(b
, idx
);
1563 ERR_IF (bitnr
>= b
->bm_bits
) {
1564 dev_err(DEV
, "bitnr=%lu bm_bits=%lu\n", bitnr
, b
->bm_bits
);
1566 c
+= (0 != test_bit_le(bitnr
- (page_nr
<< (PAGE_SHIFT
+3)), p_addr
));
1571 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1576 /* inherently racy...
1577 * return value may be already out-of-date when this function returns.
1578 * but the general usage is that this is only use during a cstate when bits are
1579 * only cleared, not set, and typically only care for the case when the return
1580 * value is zero, or we already "locked" this "bitmap extent" by other means.
1582 * enr is bm-extent number, since we chose to name one sector (512 bytes)
1583 * worth of the bitmap a "bitmap extent".
1586 * I think since we use it like a reference count, we should use the real
1587 * reference count of some bitmap extent element from some lru instead...
1590 int drbd_bm_e_weight(struct drbd_conf
*mdev
, unsigned long enr
)
1592 struct drbd_bitmap
*b
= mdev
->bitmap
;
1594 unsigned long flags
;
1595 unsigned long *p_addr
, *bm
;
1597 ERR_IF(!b
) return 0;
1598 ERR_IF(!b
->bm_pages
) return 0;
1600 spin_lock_irqsave(&b
->bm_lock
, flags
);
1601 if (BM_DONT_TEST
& b
->bm_flags
)
1602 bm_print_lock_info(mdev
);
1605 e
= min((size_t)S2W(enr
+1), b
->bm_words
);
1607 if (s
< b
->bm_words
) {
1609 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, s
));
1610 bm
= p_addr
+ MLPP(s
);
1612 count
+= hweight_long(*bm
++);
1615 dev_err(DEV
, "start offset (%d) too large in drbd_bm_e_weight\n", s
);
1617 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1621 /* Set all bits covered by the AL-extent al_enr.
1622 * Returns number of bits changed. */
1623 unsigned long drbd_bm_ALe_set_all(struct drbd_conf
*mdev
, unsigned long al_enr
)
1625 struct drbd_bitmap
*b
= mdev
->bitmap
;
1626 unsigned long *p_addr
, *bm
;
1627 unsigned long weight
;
1629 int count
, i
, do_now
;
1630 ERR_IF(!b
) return 0;
1631 ERR_IF(!b
->bm_pages
) return 0;
1633 spin_lock_irq(&b
->bm_lock
);
1634 if (BM_DONT_SET
& b
->bm_flags
)
1635 bm_print_lock_info(mdev
);
1638 s
= al_enr
* BM_WORDS_PER_AL_EXT
;
1639 e
= min_t(size_t, s
+ BM_WORDS_PER_AL_EXT
, b
->bm_words
);
1640 /* assert that s and e are on the same page */
1641 D_ASSERT((e
-1) >> (PAGE_SHIFT
- LN2_BPL
+ 3)
1642 == s
>> (PAGE_SHIFT
- LN2_BPL
+ 3));
1644 if (s
< b
->bm_words
) {
1646 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, s
));
1647 bm
= p_addr
+ MLPP(s
);
1649 count
+= hweight_long(*bm
);
1654 b
->bm_set
+= do_now
*BITS_PER_LONG
- count
;
1655 if (e
== b
->bm_words
)
1656 b
->bm_set
-= bm_clear_surplus(b
);
1658 dev_err(DEV
, "start offset (%lu) too large in drbd_bm_ALe_set_all\n", s
);
1660 weight
= b
->bm_set
- weight
;
1661 spin_unlock_irq(&b
->bm_lock
);