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>
34 /* OPAQUE outside this file!
35 * interface defined in drbd_int.h
38 * function name drbd_bm_... => used elsewhere, "public".
39 * function name bm_... => internal to implementation, "private".
45 * We want to support >= peta byte of backend storage, while for now still using
46 * a granularity of one bit per 4KiB of storage.
47 * 1 << 50 bytes backend storage (1 PiB)
48 * 1 << (50 - 12) bits needed
49 * 38 --> we need u64 to index and count bits
50 * 1 << (38 - 3) bitmap bytes needed
51 * 35 --> we still need u64 to index and count bytes
52 * (that's 32 GiB of bitmap for 1 PiB storage)
53 * 1 << (35 - 2) 32bit longs needed
54 * 33 --> we'd even need u64 to index and count 32bit long words.
55 * 1 << (35 - 3) 64bit longs needed
56 * 32 --> we could get away with a 32bit unsigned int to index and count
57 * 64bit long words, but I rather stay with unsigned long for now.
58 * We probably should neither count nor point to bytes or long words
59 * directly, but either by bitnumber, or by page index and offset.
61 * 22 --> we need that much 4KiB pages of bitmap.
62 * 1 << (22 + 3) --> on a 64bit arch,
63 * we need 32 MiB to store the array of page pointers.
65 * Because I'm lazy, and because the resulting patch was too large, too ugly
66 * and still incomplete, on 32bit we still "only" support 16 TiB (minus some),
67 * (1 << 32) bits * 4k storage.
70 * bitmap storage and IO:
71 * Bitmap is stored little endian on disk, and is kept little endian in
72 * core memory. Currently we still hold the full bitmap in core as long
73 * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage
76 * We plan to reduce the amount of in-core bitmap pages by pageing them in
77 * and out against their on-disk location as necessary, but need to make
78 * sure we don't cause too much meta data IO, and must not deadlock in
79 * tight memory situations. This needs some more work.
84 * Access to the *bm_pages is protected by bm_lock.
85 * It is safe to read the other members within the lock.
87 * drbd_bm_set_bits is called from bio_endio callbacks,
88 * We may be called with irq already disabled,
89 * so we need spin_lock_irqsave().
90 * And we need the kmap_atomic.
93 struct page
**bm_pages
;
96 /* see LIMITATIONS: above */
98 unsigned long bm_set
; /* nr of set bits; THINK maybe atomic_t? */
99 unsigned long bm_bits
;
101 size_t bm_number_of_pages
;
102 sector_t bm_dev_capacity
;
103 struct mutex bm_change
; /* serializes resize operations */
105 wait_queue_head_t bm_io_wait
; /* used to serialize IO of single pages */
107 unsigned long bm_flags
;
109 /* debugging aid, in case we are still racy somewhere */
111 struct task_struct
*bm_task
;
114 /* definition of bits in bm_flags */
116 // #define BM_MD_IO_ERROR 1 unused now.
117 #define BM_P_VMALLOCED 2
119 static int __bm_change_bits_to(struct drbd_conf
*mdev
, const unsigned long s
,
120 unsigned long e
, int val
, const enum km_type km
);
122 static int bm_is_locked(struct drbd_bitmap
*b
)
124 return test_bit(BM_LOCKED
, &b
->bm_flags
);
127 #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__)
128 static void __bm_print_lock_info(struct drbd_conf
*mdev
, const char *func
)
130 struct drbd_bitmap
*b
= mdev
->bitmap
;
131 if (!__ratelimit(&drbd_ratelimit_state
))
133 dev_err(DEV
, "FIXME %s in %s, bitmap locked for '%s' by %s\n",
134 current
== mdev
->receiver
.task
? "receiver" :
135 current
== mdev
->asender
.task
? "asender" :
136 current
== mdev
->worker
.task
? "worker" : current
->comm
,
137 func
, b
->bm_why
?: "?",
138 b
->bm_task
== mdev
->receiver
.task
? "receiver" :
139 b
->bm_task
== mdev
->asender
.task
? "asender" :
140 b
->bm_task
== mdev
->worker
.task
? "worker" : "?");
143 void drbd_bm_lock(struct drbd_conf
*mdev
, char *why
)
145 struct drbd_bitmap
*b
= mdev
->bitmap
;
149 dev_err(DEV
, "FIXME no bitmap in drbd_bm_lock!?\n");
153 trylock_failed
= !mutex_trylock(&b
->bm_change
);
155 if (trylock_failed
) {
156 dev_warn(DEV
, "%s going to '%s' but bitmap already locked for '%s' by %s\n",
157 current
== mdev
->receiver
.task
? "receiver" :
158 current
== mdev
->asender
.task
? "asender" :
159 current
== mdev
->worker
.task
? "worker" : current
->comm
,
160 why
, b
->bm_why
?: "?",
161 b
->bm_task
== mdev
->receiver
.task
? "receiver" :
162 b
->bm_task
== mdev
->asender
.task
? "asender" :
163 b
->bm_task
== mdev
->worker
.task
? "worker" : "?");
164 mutex_lock(&b
->bm_change
);
166 if (__test_and_set_bit(BM_LOCKED
, &b
->bm_flags
))
167 dev_err(DEV
, "FIXME bitmap already locked in bm_lock\n");
170 b
->bm_task
= current
;
173 void drbd_bm_unlock(struct drbd_conf
*mdev
)
175 struct drbd_bitmap
*b
= mdev
->bitmap
;
177 dev_err(DEV
, "FIXME no bitmap in drbd_bm_unlock!?\n");
181 if (!__test_and_clear_bit(BM_LOCKED
, &mdev
->bitmap
->bm_flags
))
182 dev_err(DEV
, "FIXME bitmap not locked in bm_unlock\n");
186 mutex_unlock(&b
->bm_change
);
189 /* we store some "meta" info about our pages in page->private */
190 /* at a granularity of 4k storage per bitmap bit:
191 * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks
193 * 1<<23 4k bitmap pages.
194 * Use 24 bits as page index, covers 2 peta byte storage
195 * at a granularity of 4k per bit.
196 * Used to report the failed page idx on io error from the endio handlers.
198 #define BM_PAGE_IDX_MASK ((1UL<<24)-1)
199 /* this page is currently read in, or written back */
200 #define BM_PAGE_IO_LOCK 31
201 /* if there has been an IO error for this page */
202 #define BM_PAGE_IO_ERROR 30
203 /* this is to be able to intelligently skip disk IO,
204 * set if bits have been set since last IO. */
205 #define BM_PAGE_NEED_WRITEOUT 29
206 /* to mark for lazy writeout once syncer cleared all clearable bits,
207 * we if bits have been cleared since last IO. */
208 #define BM_PAGE_LAZY_WRITEOUT 28
210 /* store_page_idx uses non-atomic assingment. It is only used directly after
211 * allocating the page. All other bm_set_page_* and bm_clear_page_* need to
212 * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap
213 * changes) may happen from various contexts, and wait_on_bit/wake_up_bit
214 * requires it all to be atomic as well. */
215 static void bm_store_page_idx(struct page
*page
, unsigned long idx
)
217 BUG_ON(0 != (idx
& ~BM_PAGE_IDX_MASK
));
218 page_private(page
) |= idx
;
221 static unsigned long bm_page_to_idx(struct page
*page
)
223 return page_private(page
) & BM_PAGE_IDX_MASK
;
226 /* As is very unlikely that the same page is under IO from more than one
227 * context, we can get away with a bit per page and one wait queue per bitmap.
229 static void bm_page_lock_io(struct drbd_conf
*mdev
, int page_nr
)
231 struct drbd_bitmap
*b
= mdev
->bitmap
;
232 void *addr
= &page_private(b
->bm_pages
[page_nr
]);
233 wait_event(b
->bm_io_wait
, !test_and_set_bit(BM_PAGE_IO_LOCK
, addr
));
236 static void bm_page_unlock_io(struct drbd_conf
*mdev
, int page_nr
)
238 struct drbd_bitmap
*b
= mdev
->bitmap
;
239 void *addr
= &page_private(b
->bm_pages
[page_nr
]);
240 clear_bit(BM_PAGE_IO_LOCK
, addr
);
241 smp_mb__after_clear_bit();
242 wake_up(&mdev
->bitmap
->bm_io_wait
);
245 /* set _before_ submit_io, so it may be reset due to being changed
246 * while this page is in flight... will get submitted later again */
247 static void bm_set_page_unchanged(struct page
*page
)
250 clear_bit(BM_PAGE_NEED_WRITEOUT
, &page_private(page
));
251 clear_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
254 static void bm_set_page_need_writeout(struct page
*page
)
256 set_bit(BM_PAGE_NEED_WRITEOUT
, &page_private(page
));
259 static int bm_test_page_unchanged(struct page
*page
)
261 volatile const unsigned long *addr
= &page_private(page
);
262 return (*addr
& ((1UL<<BM_PAGE_NEED_WRITEOUT
)|(1UL<<BM_PAGE_LAZY_WRITEOUT
))) == 0;
265 static void bm_set_page_io_err(struct page
*page
)
267 set_bit(BM_PAGE_IO_ERROR
, &page_private(page
));
270 static void bm_clear_page_io_err(struct page
*page
)
272 clear_bit(BM_PAGE_IO_ERROR
, &page_private(page
));
275 static void bm_set_page_lazy_writeout(struct page
*page
)
277 set_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
280 static int bm_test_page_lazy_writeout(struct page
*page
)
282 return test_bit(BM_PAGE_LAZY_WRITEOUT
, &page_private(page
));
285 /* on a 32bit box, this would allow for exactly (2<<38) bits. */
286 static unsigned int bm_word_to_page_idx(struct drbd_bitmap
*b
, unsigned long long_nr
)
288 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */
289 unsigned int page_nr
= long_nr
>> (PAGE_SHIFT
- LN2_BPL
+ 3);
290 BUG_ON(page_nr
>= b
->bm_number_of_pages
);
294 static unsigned int bm_bit_to_page_idx(struct drbd_bitmap
*b
, u64 bitnr
)
296 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */
297 unsigned int page_nr
= bitnr
>> (PAGE_SHIFT
+ 3);
298 BUG_ON(page_nr
>= b
->bm_number_of_pages
);
302 static unsigned long *__bm_map_pidx(struct drbd_bitmap
*b
, unsigned int idx
, const enum km_type km
)
304 struct page
*page
= b
->bm_pages
[idx
];
305 return (unsigned long *) kmap_atomic(page
, km
);
308 static unsigned long *bm_map_pidx(struct drbd_bitmap
*b
, unsigned int idx
)
310 return __bm_map_pidx(b
, idx
, KM_IRQ1
);
313 static void __bm_unmap(unsigned long *p_addr
, const enum km_type km
)
315 kunmap_atomic(p_addr
, km
);
318 static void bm_unmap(unsigned long *p_addr
)
320 return __bm_unmap(p_addr
, KM_IRQ1
);
323 /* long word offset of _bitmap_ sector */
324 #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
325 /* word offset from start of bitmap to word number _in_page_
326 * modulo longs per page
327 #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long))
328 hm, well, Philipp thinks gcc might not optimze the % into & (... - 1)
331 #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1))
333 /* Long words per page */
334 #define LWPP (PAGE_SIZE/sizeof(long))
337 * actually most functions herein should take a struct drbd_bitmap*, not a
338 * struct drbd_conf*, but for the debug macros I like to have the mdev around
339 * to be able to report device specific.
343 static void bm_free_pages(struct page
**pages
, unsigned long number
)
349 for (i
= 0; i
< number
; i
++) {
351 printk(KERN_ALERT
"drbd: bm_free_pages tried to free "
352 "a NULL pointer; i=%lu n=%lu\n",
356 __free_page(pages
[i
]);
361 static void bm_vk_free(void *ptr
, int v
)
370 * "have" and "want" are NUMBER OF PAGES.
372 static struct page
**bm_realloc_pages(struct drbd_bitmap
*b
, unsigned long want
)
374 struct page
**old_pages
= b
->bm_pages
;
375 struct page
**new_pages
, *page
;
376 unsigned int i
, bytes
, vmalloced
= 0;
377 unsigned long have
= b
->bm_number_of_pages
;
379 BUG_ON(have
== 0 && old_pages
!= NULL
);
380 BUG_ON(have
!= 0 && old_pages
== NULL
);
385 /* Trying kmalloc first, falling back to vmalloc.
386 * GFP_KERNEL is ok, as this is done when a lower level disk is
387 * "attached" to the drbd. Context is receiver thread or cqueue
388 * thread. As we have no disk yet, we are not in the IO path,
389 * not even the IO path of the peer. */
390 bytes
= sizeof(struct page
*)*want
;
391 new_pages
= kmalloc(bytes
, GFP_KERNEL
);
393 new_pages
= vmalloc(bytes
);
399 memset(new_pages
, 0, bytes
);
401 for (i
= 0; i
< have
; i
++)
402 new_pages
[i
] = old_pages
[i
];
403 for (; i
< want
; i
++) {
404 page
= alloc_page(GFP_HIGHUSER
);
406 bm_free_pages(new_pages
+ have
, i
- have
);
407 bm_vk_free(new_pages
, vmalloced
);
410 /* we want to know which page it is
411 * from the endio handlers */
412 bm_store_page_idx(page
, i
);
416 for (i
= 0; i
< want
; i
++)
417 new_pages
[i
] = old_pages
[i
];
418 /* NOT HERE, we are outside the spinlock!
419 bm_free_pages(old_pages + want, have - want);
424 set_bit(BM_P_VMALLOCED
, &b
->bm_flags
);
426 clear_bit(BM_P_VMALLOCED
, &b
->bm_flags
);
432 * called on driver init only. TODO call when a device is created.
433 * allocates the drbd_bitmap, and stores it in mdev->bitmap.
435 int drbd_bm_init(struct drbd_conf
*mdev
)
437 struct drbd_bitmap
*b
= mdev
->bitmap
;
439 b
= kzalloc(sizeof(struct drbd_bitmap
), GFP_KERNEL
);
442 spin_lock_init(&b
->bm_lock
);
443 mutex_init(&b
->bm_change
);
444 init_waitqueue_head(&b
->bm_io_wait
);
451 sector_t
drbd_bm_capacity(struct drbd_conf
*mdev
)
453 ERR_IF(!mdev
->bitmap
) return 0;
454 return mdev
->bitmap
->bm_dev_capacity
;
457 /* called on driver unload. TODO: call when a device is destroyed.
459 void drbd_bm_cleanup(struct drbd_conf
*mdev
)
461 ERR_IF (!mdev
->bitmap
) return;
462 bm_free_pages(mdev
->bitmap
->bm_pages
, mdev
->bitmap
->bm_number_of_pages
);
463 bm_vk_free(mdev
->bitmap
->bm_pages
, test_bit(BM_P_VMALLOCED
, &mdev
->bitmap
->bm_flags
));
469 * since (b->bm_bits % BITS_PER_LONG) != 0,
470 * this masks out the remaining bits.
471 * Returns the number of bits cleared.
473 #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3))
474 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1)
475 #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1)
476 static int bm_clear_surplus(struct drbd_bitmap
*b
)
479 unsigned long *p_addr
, *bm
;
483 /* number of bits modulo bits per page */
484 tmp
= (b
->bm_bits
& BITS_PER_PAGE_MASK
);
485 /* mask the used bits of the word containing the last bit */
486 mask
= (1UL << (tmp
& BITS_PER_LONG_MASK
)) -1;
487 /* bitmap is always stored little endian,
488 * on disk and in core memory alike */
489 mask
= cpu_to_lel(mask
);
491 /* because of the "extra long to catch oob access" we allocate in
492 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
493 * containing the last _relevant_ bitmap word */
494 p_addr
= bm_map_pidx(b
, bm_bit_to_page_idx(b
, b
->bm_bits
- 1));
495 bm
= p_addr
+ (tmp
/BITS_PER_LONG
);
497 /* If mask != 0, we are not exactly aligned, so bm now points
498 * to the long containing the last bit.
499 * If mask == 0, bm already points to the word immediately
500 * after the last (long word aligned) bit. */
501 cleared
= hweight_long(*bm
& ~mask
);
506 if (BITS_PER_LONG
== 32 && ((bm
- p_addr
) & 1) == 1) {
507 /* on a 32bit arch, we may need to zero out
508 * a padding long to align with a 64bit remote */
509 cleared
+= hweight_long(*bm
);
516 static void bm_set_surplus(struct drbd_bitmap
*b
)
519 unsigned long *p_addr
, *bm
;
522 /* number of bits modulo bits per page */
523 tmp
= (b
->bm_bits
& BITS_PER_PAGE_MASK
);
524 /* mask the used bits of the word containing the last bit */
525 mask
= (1UL << (tmp
& BITS_PER_LONG_MASK
)) -1;
526 /* bitmap is always stored little endian,
527 * on disk and in core memory alike */
528 mask
= cpu_to_lel(mask
);
530 /* because of the "extra long to catch oob access" we allocate in
531 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
532 * containing the last _relevant_ bitmap word */
533 p_addr
= bm_map_pidx(b
, bm_bit_to_page_idx(b
, b
->bm_bits
- 1));
534 bm
= p_addr
+ (tmp
/BITS_PER_LONG
);
536 /* If mask != 0, we are not exactly aligned, so bm now points
537 * to the long containing the last bit.
538 * If mask == 0, bm already points to the word immediately
539 * after the last (long word aligned) bit. */
544 if (BITS_PER_LONG
== 32 && ((bm
- p_addr
) & 1) == 1) {
545 /* on a 32bit arch, we may need to zero out
546 * a padding long to align with a 64bit remote */
552 /* you better not modify the bitmap while this is running,
553 * or its results will be stale */
554 static unsigned long bm_count_bits(struct drbd_bitmap
*b
)
556 unsigned long *p_addr
;
557 unsigned long bits
= 0;
558 unsigned long mask
= (1UL << (b
->bm_bits
& BITS_PER_LONG_MASK
)) -1;
559 int idx
, last_page
, i
, last_word
;
561 /* because of the "extra long to catch oob access" we allocate in
562 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
563 * containing the last _relevant_ bitmap word */
564 last_page
= bm_bit_to_page_idx(b
, b
->bm_bits
-1);
566 /* all but last page */
567 for (idx
= 0; idx
< last_page
; idx
++) {
568 p_addr
= __bm_map_pidx(b
, idx
, KM_USER0
);
569 for (i
= 0; i
< LWPP
; i
++)
570 bits
+= hweight_long(p_addr
[i
]);
571 __bm_unmap(p_addr
, KM_USER0
);
574 /* last (or only) page */
575 last_word
= ((b
->bm_bits
- 1) & BITS_PER_PAGE_MASK
) >> LN2_BPL
;
576 p_addr
= __bm_map_pidx(b
, idx
, KM_USER0
);
577 for (i
= 0; i
< last_word
; i
++)
578 bits
+= hweight_long(p_addr
[i
]);
579 p_addr
[last_word
] &= cpu_to_lel(mask
);
580 bits
+= hweight_long(p_addr
[last_word
]);
581 /* 32bit arch, may have an unused padding long */
582 if (BITS_PER_LONG
== 32 && (last_word
& 1) == 0)
583 p_addr
[last_word
+1] = 0;
584 __bm_unmap(p_addr
, KM_USER0
);
588 /* offset and len in long words.*/
589 static void bm_memset(struct drbd_bitmap
*b
, size_t offset
, int c
, size_t len
)
591 unsigned long *p_addr
, *bm
;
597 if (end
> b
->bm_words
) {
598 printk(KERN_ALERT
"drbd: bm_memset end > bm_words\n");
602 while (offset
< end
) {
603 do_now
= min_t(size_t, ALIGN(offset
+ 1, LWPP
), end
) - offset
;
604 idx
= bm_word_to_page_idx(b
, offset
);
605 p_addr
= bm_map_pidx(b
, idx
);
606 bm
= p_addr
+ MLPP(offset
);
607 if (bm
+do_now
> p_addr
+ LWPP
) {
608 printk(KERN_ALERT
"drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n",
609 p_addr
, bm
, (int)do_now
);
610 break; /* breaks to after catch_oob_access_end() only! */
612 memset(bm
, c
, do_now
* sizeof(long));
614 bm_set_page_need_writeout(b
->bm_pages
[idx
]);
620 * make sure the bitmap has enough room for the attached storage,
621 * if necessary, resize.
622 * called whenever we may have changed the device size.
623 * returns -ENOMEM if we could not allocate enough memory, 0 on success.
624 * In case this is actually a resize, we copy the old bitmap into the new one.
625 * Otherwise, the bitmap is initialized to all bits set.
627 int drbd_bm_resize(struct drbd_conf
*mdev
, sector_t capacity
, int set_new_bits
)
629 struct drbd_bitmap
*b
= mdev
->bitmap
;
630 unsigned long bits
, words
, owords
, obits
, *p_addr
, *bm
;
631 unsigned long want
, have
, onpages
; /* number of pages */
632 struct page
**npages
, **opages
= NULL
;
633 int err
= 0, growing
;
634 int opages_vmalloced
;
636 ERR_IF(!b
) return -ENOMEM
;
638 drbd_bm_lock(mdev
, "resize");
640 dev_info(DEV
, "drbd_bm_resize called with capacity == %llu\n",
641 (unsigned long long)capacity
);
643 if (capacity
== b
->bm_dev_capacity
)
646 opages_vmalloced
= test_bit(BM_P_VMALLOCED
, &b
->bm_flags
);
649 spin_lock_irq(&b
->bm_lock
);
650 opages
= b
->bm_pages
;
651 onpages
= b
->bm_number_of_pages
;
652 owords
= b
->bm_words
;
654 b
->bm_number_of_pages
=
658 b
->bm_dev_capacity
= 0;
659 spin_unlock_irq(&b
->bm_lock
);
660 bm_free_pages(opages
, onpages
);
661 bm_vk_free(opages
, opages_vmalloced
);
664 bits
= BM_SECT_TO_BIT(ALIGN(capacity
, BM_SECT_PER_BIT
));
667 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL;
668 a 32bit host could present the wrong number of words
671 words
= ALIGN(bits
, 64) >> LN2_BPL
;
673 if (get_ldev(mdev
)) {
674 u64 bits_on_disk
= ((u64
)mdev
->ldev
->md
.md_size_sect
-MD_BM_OFFSET
) << 12;
676 if (bits
> bits_on_disk
) {
677 dev_info(DEV
, "bits = %lu\n", bits
);
678 dev_info(DEV
, "bits_on_disk = %llu\n", bits_on_disk
);
684 /* one extra long to catch off by one errors */
685 want
= ALIGN((words
+1)*sizeof(long), PAGE_SIZE
) >> PAGE_SHIFT
;
686 have
= b
->bm_number_of_pages
;
688 D_ASSERT(b
->bm_pages
!= NULL
);
689 npages
= b
->bm_pages
;
691 if (drbd_insert_fault(mdev
, DRBD_FAULT_BM_ALLOC
))
694 npages
= bm_realloc_pages(b
, want
);
702 spin_lock_irq(&b
->bm_lock
);
703 opages
= b
->bm_pages
;
704 owords
= b
->bm_words
;
707 growing
= bits
> obits
;
708 if (opages
&& growing
&& set_new_bits
)
711 b
->bm_pages
= npages
;
712 b
->bm_number_of_pages
= want
;
715 b
->bm_dev_capacity
= capacity
;
719 bm_memset(b
, owords
, 0xff, words
-owords
);
720 b
->bm_set
+= bits
- obits
;
722 bm_memset(b
, owords
, 0x00, words
-owords
);
727 /* implicit: (opages != NULL) && (opages != npages) */
728 bm_free_pages(opages
+ want
, have
- want
);
731 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, words
));
732 bm
= p_addr
+ MLPP(words
);
736 (void)bm_clear_surplus(b
);
738 spin_unlock_irq(&b
->bm_lock
);
739 if (opages
!= npages
)
740 bm_vk_free(opages
, opages_vmalloced
);
742 b
->bm_set
= bm_count_bits(b
);
743 dev_info(DEV
, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits
, words
, want
);
746 drbd_bm_unlock(mdev
);
751 * if not protected by other means, return value may be out of date when
752 * leaving this function...
753 * we still need to lock it, since it is important that this returns
754 * bm_set == 0 precisely.
756 * maybe bm_set should be atomic_t ?
758 unsigned long _drbd_bm_total_weight(struct drbd_conf
*mdev
)
760 struct drbd_bitmap
*b
= mdev
->bitmap
;
765 ERR_IF(!b
->bm_pages
) return 0;
767 spin_lock_irqsave(&b
->bm_lock
, flags
);
769 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
774 unsigned long drbd_bm_total_weight(struct drbd_conf
*mdev
)
777 /* if I don't have a disk, I don't know about out-of-sync status */
778 if (!get_ldev_if_state(mdev
, D_NEGOTIATING
))
780 s
= _drbd_bm_total_weight(mdev
);
785 size_t drbd_bm_words(struct drbd_conf
*mdev
)
787 struct drbd_bitmap
*b
= mdev
->bitmap
;
789 ERR_IF(!b
->bm_pages
) return 0;
794 unsigned long drbd_bm_bits(struct drbd_conf
*mdev
)
796 struct drbd_bitmap
*b
= mdev
->bitmap
;
802 /* merge number words from buffer into the bitmap starting at offset.
803 * buffer[i] is expected to be little endian unsigned long.
804 * bitmap must be locked by drbd_bm_lock.
805 * currently only used from receive_bitmap.
807 void drbd_bm_merge_lel(struct drbd_conf
*mdev
, size_t offset
, size_t number
,
808 unsigned long *buffer
)
810 struct drbd_bitmap
*b
= mdev
->bitmap
;
811 unsigned long *p_addr
, *bm
;
812 unsigned long word
, bits
;
816 end
= offset
+ number
;
819 ERR_IF(!b
->bm_pages
) return;
822 WARN_ON(offset
>= b
->bm_words
);
823 WARN_ON(end
> b
->bm_words
);
825 spin_lock_irq(&b
->bm_lock
);
826 while (offset
< end
) {
827 do_now
= min_t(size_t, ALIGN(offset
+1, LWPP
), end
) - offset
;
828 idx
= bm_word_to_page_idx(b
, offset
);
829 p_addr
= bm_map_pidx(b
, idx
);
830 bm
= p_addr
+ MLPP(offset
);
833 bits
= hweight_long(*bm
);
834 word
= *bm
| *buffer
++;
836 b
->bm_set
+= hweight_long(word
) - bits
;
839 bm_set_page_need_writeout(b
->bm_pages
[idx
]);
841 /* with 32bit <-> 64bit cross-platform connect
842 * this is only correct for current usage,
843 * where we _know_ that we are 64 bit aligned,
844 * and know that this function is used in this way, too...
846 if (end
== b
->bm_words
)
847 b
->bm_set
-= bm_clear_surplus(b
);
849 spin_unlock_irq(&b
->bm_lock
);
852 /* copy number words from the bitmap starting at offset into the buffer.
853 * buffer[i] will be little endian unsigned long.
855 void drbd_bm_get_lel(struct drbd_conf
*mdev
, size_t offset
, size_t number
,
856 unsigned long *buffer
)
858 struct drbd_bitmap
*b
= mdev
->bitmap
;
859 unsigned long *p_addr
, *bm
;
862 end
= offset
+ number
;
865 ERR_IF(!b
->bm_pages
) return;
867 spin_lock_irq(&b
->bm_lock
);
868 if ((offset
>= b
->bm_words
) ||
869 (end
> b
->bm_words
) ||
871 dev_err(DEV
, "offset=%lu number=%lu bm_words=%lu\n",
872 (unsigned long) offset
,
873 (unsigned long) number
,
874 (unsigned long) b
->bm_words
);
876 while (offset
< end
) {
877 do_now
= min_t(size_t, ALIGN(offset
+1, LWPP
), end
) - offset
;
878 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, offset
));
879 bm
= p_addr
+ MLPP(offset
);
886 spin_unlock_irq(&b
->bm_lock
);
889 /* set all bits in the bitmap */
890 void drbd_bm_set_all(struct drbd_conf
*mdev
)
892 struct drbd_bitmap
*b
= mdev
->bitmap
;
894 ERR_IF(!b
->bm_pages
) return;
896 spin_lock_irq(&b
->bm_lock
);
897 bm_memset(b
, 0, 0xff, b
->bm_words
);
898 (void)bm_clear_surplus(b
);
899 b
->bm_set
= b
->bm_bits
;
900 spin_unlock_irq(&b
->bm_lock
);
903 /* clear all bits in the bitmap */
904 void drbd_bm_clear_all(struct drbd_conf
*mdev
)
906 struct drbd_bitmap
*b
= mdev
->bitmap
;
908 ERR_IF(!b
->bm_pages
) return;
910 spin_lock_irq(&b
->bm_lock
);
911 bm_memset(b
, 0, 0, b
->bm_words
);
913 spin_unlock_irq(&b
->bm_lock
);
917 struct drbd_conf
*mdev
;
919 wait_queue_head_t io_wait
;
921 #define BM_AIO_COPY_PAGES 1
925 /* bv_page may be a copy, or may be the original */
926 static void bm_async_io_complete(struct bio
*bio
, int error
)
928 struct bm_aio_ctx
*ctx
= bio
->bi_private
;
929 struct drbd_conf
*mdev
= ctx
->mdev
;
930 struct drbd_bitmap
*b
= mdev
->bitmap
;
931 unsigned int idx
= bm_page_to_idx(bio
->bi_io_vec
[0].bv_page
);
932 int uptodate
= bio_flagged(bio
, BIO_UPTODATE
);
935 /* strange behavior of some lower level drivers...
936 * fail the request by clearing the uptodate flag,
937 * but do not return any error?!
938 * do we want to WARN() on this? */
939 if (!error
&& !uptodate
)
942 if (!bm_test_page_unchanged(b
->bm_pages
[idx
]))
943 dev_info(DEV
, "bitmap page idx %u changed during IO!\n", idx
);
946 /* ctx error will hold the completed-last non-zero error code,
947 * in case error codes differ. */
949 bm_set_page_io_err(b
->bm_pages
[idx
]);
950 /* Not identical to on disk version of it.
951 * Is BM_PAGE_IO_ERROR enough? */
952 if (__ratelimit(&drbd_ratelimit_state
))
953 dev_err(DEV
, "IO ERROR %d on bitmap page idx %u\n",
956 bm_clear_page_io_err(b
->bm_pages
[idx
]);
957 dynamic_dev_dbg(DEV
, "bitmap page idx %u completed\n", idx
);
960 bm_page_unlock_io(mdev
, idx
);
962 /* FIXME give back to page pool */
963 if (ctx
->flags
& BM_AIO_COPY_PAGES
)
964 put_page(bio
->bi_io_vec
[0].bv_page
);
968 if (atomic_dec_and_test(&ctx
->in_flight
))
969 wake_up(&ctx
->io_wait
);
972 static void bm_page_io_async(struct bm_aio_ctx
*ctx
, int page_nr
, int rw
) __must_hold(local
)
974 /* we are process context. we always get a bio */
975 struct bio
*bio
= bio_alloc(GFP_KERNEL
, 1);
976 struct drbd_conf
*mdev
= ctx
->mdev
;
977 struct drbd_bitmap
*b
= mdev
->bitmap
;
981 sector_t on_disk_sector
=
982 mdev
->ldev
->md
.md_offset
+ mdev
->ldev
->md
.bm_offset
;
983 on_disk_sector
+= ((sector_t
)page_nr
) << (PAGE_SHIFT
-9);
985 /* this might happen with very small
986 * flexible external meta data device,
987 * or with PAGE_SIZE > 4k */
988 len
= min_t(unsigned int, PAGE_SIZE
,
989 (drbd_md_last_sector(mdev
->ldev
) - on_disk_sector
+ 1)<<9);
991 /* serialize IO on this page */
992 bm_page_lock_io(mdev
, page_nr
);
993 /* before memcpy and submit,
994 * so it can be redirtied any time */
995 bm_set_page_unchanged(b
->bm_pages
[page_nr
]);
997 if (ctx
->flags
& BM_AIO_COPY_PAGES
) {
998 /* FIXME alloc_page is good enough for now, but actually needs
999 * to use pre-allocated page pool */
1001 page
= alloc_page(__GFP_HIGHMEM
|__GFP_WAIT
);
1002 dest
= kmap_atomic(page
, KM_USER0
);
1003 src
= kmap_atomic(b
->bm_pages
[page_nr
], KM_USER1
);
1004 memcpy(dest
, src
, PAGE_SIZE
);
1005 kunmap_atomic(src
, KM_USER1
);
1006 kunmap_atomic(dest
, KM_USER0
);
1007 bm_store_page_idx(page
, page_nr
);
1009 page
= b
->bm_pages
[page_nr
];
1011 bio
->bi_bdev
= mdev
->ldev
->md_bdev
;
1012 bio
->bi_sector
= on_disk_sector
;
1013 bio_add_page(bio
, page
, len
, 0);
1014 bio
->bi_private
= ctx
;
1015 bio
->bi_end_io
= bm_async_io_complete
;
1017 if (drbd_insert_fault(mdev
, (rw
& WRITE
) ? DRBD_FAULT_MD_WR
: DRBD_FAULT_MD_RD
)) {
1019 bio_endio(bio
, -EIO
);
1021 submit_bio(rw
, bio
);
1026 * bm_rw: read/write the whole bitmap from/to its on disk location.
1028 static int bm_rw(struct drbd_conf
*mdev
, int rw
, unsigned lazy_writeout_upper_idx
) __must_hold(local
)
1030 struct bm_aio_ctx ctx
=
1031 { .flags
= lazy_writeout_upper_idx
? BM_AIO_COPY_PAGES
: 0 };
1032 struct drbd_bitmap
*b
= mdev
->bitmap
;
1033 int last_page
, i
, count
= 0;
1039 * We are protected against bitmap disappearing/resizing by holding an
1040 * ldev reference (caller must have called get_ldev()).
1041 * For read/write, we are protected against changes to the bitmap by
1042 * the bitmap lock (see drbd_bitmap_io).
1043 * For lazy writeout, we don't care for ongoing changes to the bitmap,
1044 * as we submit copies of pages anyways.
1047 WARN_ON(!bm_is_locked(b
));
1049 /* because of the "extra long to catch oob access" we allocate in
1050 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
1051 * containing the last _relevant_ bitmap word */
1052 last_page
= bm_word_to_page_idx(b
, b
->bm_words
- 1);
1056 atomic_set(&ctx
.in_flight
, 1); /* one extra ref */
1057 init_waitqueue_head(&ctx
.io_wait
);
1060 /* let the layers below us try to merge these bios... */
1061 for (i
= 0; i
<= last_page
; i
++) {
1062 /* ignore completely unchanged pages */
1063 if (lazy_writeout_upper_idx
&& i
== lazy_writeout_upper_idx
)
1066 if (bm_test_page_unchanged(b
->bm_pages
[i
])) {
1067 dynamic_dev_dbg(DEV
, "skipped bm write for idx %u\n", i
);
1070 /* during lazy writeout,
1071 * ignore those pages not marked for lazy writeout. */
1072 if (lazy_writeout_upper_idx
&&
1073 !bm_test_page_lazy_writeout(b
->bm_pages
[i
])) {
1074 dynamic_dev_dbg(DEV
, "skipped bm lazy write for idx %u\n", i
);
1078 atomic_inc(&ctx
.in_flight
);
1079 bm_page_io_async(&ctx
, i
, rw
);
1084 atomic_dec(&ctx
.in_flight
); /* drop the extra ref */
1085 wait_event(ctx
.io_wait
, atomic_read(&ctx
.in_flight
) == 0);
1086 dev_info(DEV
, "bitmap %s of %u pages took %lu jiffies\n",
1087 rw
== WRITE
? "WRITE" : "READ",
1088 count
, jiffies
- now
);
1091 dev_alert(DEV
, "we had at least one MD IO ERROR during bitmap IO\n");
1092 drbd_chk_io_error(mdev
, 1, true);
1093 err
= -EIO
; /* ctx.error ? */
1098 drbd_md_flush(mdev
);
1099 } else /* rw == READ */ {
1100 b
->bm_set
= bm_count_bits(b
);
1101 dev_info(DEV
, "recounting of set bits took additional %lu jiffies\n",
1106 dev_info(DEV
, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n",
1107 ppsize(ppb
, now
<< (BM_BLOCK_SHIFT
-10)), now
);
1113 * drbd_bm_read() - Read the whole bitmap from its on disk location.
1114 * @mdev: DRBD device.
1116 int drbd_bm_read(struct drbd_conf
*mdev
) __must_hold(local
)
1118 return bm_rw(mdev
, READ
, 0);
1122 * drbd_bm_write() - Write the whole bitmap to its on disk location.
1123 * @mdev: DRBD device.
1125 * Will only write pages that have changed since last IO.
1127 int drbd_bm_write(struct drbd_conf
*mdev
) __must_hold(local
)
1129 return bm_rw(mdev
, WRITE
, 0);
1133 * drbd_bm_lazy_write_out() - Write bitmap pages 0 to @upper_idx-1, if they have changed.
1134 * @mdev: DRBD device.
1135 * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages
1137 int drbd_bm_write_lazy(struct drbd_conf
*mdev
, unsigned upper_idx
) __must_hold(local
)
1139 return bm_rw(mdev
, WRITE
, upper_idx
);
1144 * drbd_bm_write_page: Writes a PAGE_SIZE aligned piece of bitmap
1145 * @mdev: DRBD device.
1146 * @idx: bitmap page index
1148 * We don't want to special case on logical_block_size of the backend device,
1149 * so we submit PAGE_SIZE aligned pieces.
1150 * Note that on "most" systems, PAGE_SIZE is 4k.
1152 * In case this becomes an issue on systems with larger PAGE_SIZE,
1153 * we may want to change this again to write 4k aligned 4k pieces.
1155 int drbd_bm_write_page(struct drbd_conf
*mdev
, unsigned int idx
) __must_hold(local
)
1157 struct bm_aio_ctx ctx
= { .flags
= BM_AIO_COPY_PAGES
, };
1159 if (bm_test_page_unchanged(mdev
->bitmap
->bm_pages
[idx
])) {
1160 dev_info(DEV
, "skipped bm page write for idx %u\n", idx
);
1165 atomic_set(&ctx
.in_flight
, 1);
1166 init_waitqueue_head(&ctx
.io_wait
);
1168 bm_page_io_async(&ctx
, idx
, WRITE_SYNC
);
1169 wait_event(ctx
.io_wait
, atomic_read(&ctx
.in_flight
) == 0);
1172 drbd_chk_io_error(mdev
, 1, true);
1173 /* that should force detach, so the in memory bitmap will be
1174 * gone in a moment as well. */
1176 mdev
->bm_writ_cnt
++;
1181 * find_first_bit returns int, we return unsigned long.
1182 * For this to work on 32bit arch with bitnumbers > (1<<32),
1183 * we'd need to return u64, and get a whole lot of other places
1184 * fixed where we still use unsigned long.
1186 * this returns a bit number, NOT a sector!
1188 static unsigned long __bm_find_next(struct drbd_conf
*mdev
, unsigned long bm_fo
,
1189 const int find_zero_bit
, const enum km_type km
)
1191 struct drbd_bitmap
*b
= mdev
->bitmap
;
1192 unsigned long *p_addr
;
1193 unsigned long bit_offset
;
1197 if (bm_fo
> b
->bm_bits
) {
1198 dev_err(DEV
, "bm_fo=%lu bm_bits=%lu\n", bm_fo
, b
->bm_bits
);
1199 bm_fo
= DRBD_END_OF_BITMAP
;
1201 while (bm_fo
< b
->bm_bits
) {
1202 /* bit offset of the first bit in the page */
1203 bit_offset
= bm_fo
& ~BITS_PER_PAGE_MASK
;
1204 p_addr
= __bm_map_pidx(b
, bm_bit_to_page_idx(b
, bm_fo
), km
);
1207 i
= generic_find_next_zero_le_bit(p_addr
,
1208 PAGE_SIZE
*8, bm_fo
& BITS_PER_PAGE_MASK
);
1210 i
= generic_find_next_le_bit(p_addr
,
1211 PAGE_SIZE
*8, bm_fo
& BITS_PER_PAGE_MASK
);
1213 __bm_unmap(p_addr
, km
);
1214 if (i
< PAGE_SIZE
*8) {
1215 bm_fo
= bit_offset
+ i
;
1216 if (bm_fo
>= b
->bm_bits
)
1220 bm_fo
= bit_offset
+ PAGE_SIZE
*8;
1222 bm_fo
= DRBD_END_OF_BITMAP
;
1228 static unsigned long bm_find_next(struct drbd_conf
*mdev
,
1229 unsigned long bm_fo
, const int find_zero_bit
)
1231 struct drbd_bitmap
*b
= mdev
->bitmap
;
1232 unsigned long i
= DRBD_END_OF_BITMAP
;
1234 ERR_IF(!b
) return i
;
1235 ERR_IF(!b
->bm_pages
) return i
;
1237 spin_lock_irq(&b
->bm_lock
);
1238 if (bm_is_locked(b
))
1239 bm_print_lock_info(mdev
);
1241 i
= __bm_find_next(mdev
, bm_fo
, find_zero_bit
, KM_IRQ1
);
1243 spin_unlock_irq(&b
->bm_lock
);
1247 unsigned long drbd_bm_find_next(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1249 return bm_find_next(mdev
, bm_fo
, 0);
1253 /* not yet needed for anything. */
1254 unsigned long drbd_bm_find_next_zero(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1256 return bm_find_next(mdev
, bm_fo
, 1);
1260 /* does not spin_lock_irqsave.
1261 * you must take drbd_bm_lock() first */
1262 unsigned long _drbd_bm_find_next(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1264 /* WARN_ON(!bm_is_locked(mdev)); */
1265 return __bm_find_next(mdev
, bm_fo
, 0, KM_USER1
);
1268 unsigned long _drbd_bm_find_next_zero(struct drbd_conf
*mdev
, unsigned long bm_fo
)
1270 /* WARN_ON(!bm_is_locked(mdev)); */
1271 return __bm_find_next(mdev
, bm_fo
, 1, KM_USER1
);
1274 /* returns number of bits actually changed.
1275 * for val != 0, we change 0 -> 1, return code positive
1276 * for val == 0, we change 1 -> 0, return code negative
1277 * wants bitnr, not sector.
1278 * expected to be called for only a few bits (e - s about BITS_PER_LONG).
1279 * Must hold bitmap lock already. */
1280 static int __bm_change_bits_to(struct drbd_conf
*mdev
, const unsigned long s
,
1281 unsigned long e
, int val
, const enum km_type km
)
1283 struct drbd_bitmap
*b
= mdev
->bitmap
;
1284 unsigned long *p_addr
= NULL
;
1285 unsigned long bitnr
;
1286 unsigned int last_page_nr
= -1U;
1288 int changed_total
= 0;
1290 if (e
>= b
->bm_bits
) {
1291 dev_err(DEV
, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n",
1293 e
= b
->bm_bits
? b
->bm_bits
-1 : 0;
1295 for (bitnr
= s
; bitnr
<= e
; bitnr
++) {
1296 unsigned int page_nr
= bm_bit_to_page_idx(b
, bitnr
);
1297 if (page_nr
!= last_page_nr
) {
1299 __bm_unmap(p_addr
, km
);
1301 bm_set_page_lazy_writeout(b
->bm_pages
[last_page_nr
]);
1303 bm_set_page_need_writeout(b
->bm_pages
[last_page_nr
]);
1306 p_addr
= __bm_map_pidx(b
, page_nr
, km
);
1307 last_page_nr
= page_nr
;
1310 c
+= (0 == generic___test_and_set_le_bit(bitnr
& BITS_PER_PAGE_MASK
, p_addr
));
1312 c
-= (0 != generic___test_and_clear_le_bit(bitnr
& BITS_PER_PAGE_MASK
, p_addr
));
1315 __bm_unmap(p_addr
, km
);
1317 bm_set_page_lazy_writeout(b
->bm_pages
[last_page_nr
]);
1319 bm_set_page_need_writeout(b
->bm_pages
[last_page_nr
]);
1321 b
->bm_set
+= changed_total
;
1322 return changed_total
;
1325 /* returns number of bits actually changed.
1326 * for val != 0, we change 0 -> 1, return code positive
1327 * for val == 0, we change 1 -> 0, return code negative
1328 * wants bitnr, not sector */
1329 static int bm_change_bits_to(struct drbd_conf
*mdev
, const unsigned long s
,
1330 const unsigned long e
, int val
)
1332 unsigned long flags
;
1333 struct drbd_bitmap
*b
= mdev
->bitmap
;
1336 ERR_IF(!b
) return 1;
1337 ERR_IF(!b
->bm_pages
) return 0;
1339 spin_lock_irqsave(&b
->bm_lock
, flags
);
1340 if (bm_is_locked(b
))
1341 bm_print_lock_info(mdev
);
1343 c
= __bm_change_bits_to(mdev
, s
, e
, val
, KM_IRQ1
);
1345 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1349 /* returns number of bits changed 0 -> 1 */
1350 int drbd_bm_set_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1352 return bm_change_bits_to(mdev
, s
, e
, 1);
1355 /* returns number of bits changed 1 -> 0 */
1356 int drbd_bm_clear_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1358 return -bm_change_bits_to(mdev
, s
, e
, 0);
1361 /* sets all bits in full words,
1362 * from first_word up to, but not including, last_word */
1363 static inline void bm_set_full_words_within_one_page(struct drbd_bitmap
*b
,
1364 int page_nr
, int first_word
, int last_word
)
1368 unsigned long *paddr
= kmap_atomic(b
->bm_pages
[page_nr
], KM_USER0
);
1369 for (i
= first_word
; i
< last_word
; i
++) {
1370 bits
= hweight_long(paddr
[i
]);
1372 b
->bm_set
+= BITS_PER_LONG
- bits
;
1374 kunmap_atomic(paddr
, KM_USER0
);
1377 /* Same thing as drbd_bm_set_bits, but without taking the spin_lock_irqsave.
1378 * You must first drbd_bm_lock().
1379 * Can be called to set the whole bitmap in one go.
1380 * Sets bits from s to e _inclusive_. */
1381 void _drbd_bm_set_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1383 /* First set_bit from the first bit (s)
1384 * up to the next long boundary (sl),
1385 * then assign full words up to the last long boundary (el),
1386 * then set_bit up to and including the last bit (e).
1388 * Do not use memset, because we must account for changes,
1389 * so we need to loop over the words with hweight() anyways.
1391 unsigned long sl
= ALIGN(s
,BITS_PER_LONG
);
1392 unsigned long el
= (e
+1) & ~((unsigned long)BITS_PER_LONG
-1);
1399 if (e
- s
<= 3*BITS_PER_LONG
) {
1400 /* don't bother; el and sl may even be wrong. */
1401 __bm_change_bits_to(mdev
, s
, e
, 1, KM_USER0
);
1405 /* difference is large enough that we can trust sl and el */
1407 /* bits filling the current long */
1409 __bm_change_bits_to(mdev
, s
, sl
-1, 1, KM_USER0
);
1411 first_page
= sl
>> (3 + PAGE_SHIFT
);
1412 last_page
= el
>> (3 + PAGE_SHIFT
);
1414 /* MLPP: modulo longs per page */
1415 /* LWPP: long words per page */
1416 first_word
= MLPP(sl
>> LN2_BPL
);
1419 /* first and full pages, unless first page == last page */
1420 for (page_nr
= first_page
; page_nr
< last_page
; page_nr
++) {
1421 bm_set_full_words_within_one_page(mdev
->bitmap
, page_nr
, first_word
, last_word
);
1426 /* last page (respectively only page, for first page == last page) */
1427 last_word
= MLPP(el
>> LN2_BPL
);
1428 bm_set_full_words_within_one_page(mdev
->bitmap
, last_page
, first_word
, last_word
);
1430 /* possibly trailing bits.
1431 * example: (e & 63) == 63, el will be e+1.
1432 * if that even was the very last bit,
1433 * it would trigger an assert in __bm_change_bits_to()
1436 __bm_change_bits_to(mdev
, el
, e
, 1, KM_USER0
);
1439 /* returns bit state
1440 * wants bitnr, NOT sector.
1441 * inherently racy... area needs to be locked by means of {al,rs}_lru
1444 * -1 ... first out of bounds access, stop testing for bits!
1446 int drbd_bm_test_bit(struct drbd_conf
*mdev
, const unsigned long bitnr
)
1448 unsigned long flags
;
1449 struct drbd_bitmap
*b
= mdev
->bitmap
;
1450 unsigned long *p_addr
;
1453 ERR_IF(!b
) return 0;
1454 ERR_IF(!b
->bm_pages
) return 0;
1456 spin_lock_irqsave(&b
->bm_lock
, flags
);
1457 if (bm_is_locked(b
))
1458 bm_print_lock_info(mdev
);
1459 if (bitnr
< b
->bm_bits
) {
1460 p_addr
= bm_map_pidx(b
, bm_bit_to_page_idx(b
, bitnr
));
1461 i
= generic_test_le_bit(bitnr
& BITS_PER_PAGE_MASK
, p_addr
) ? 1 : 0;
1463 } else if (bitnr
== b
->bm_bits
) {
1465 } else { /* (bitnr > b->bm_bits) */
1466 dev_err(DEV
, "bitnr=%lu > bm_bits=%lu\n", bitnr
, b
->bm_bits
);
1470 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1474 /* returns number of bits set in the range [s, e] */
1475 int drbd_bm_count_bits(struct drbd_conf
*mdev
, const unsigned long s
, const unsigned long e
)
1477 unsigned long flags
;
1478 struct drbd_bitmap
*b
= mdev
->bitmap
;
1479 unsigned long *p_addr
= NULL
;
1480 unsigned long bitnr
;
1481 unsigned int page_nr
= -1U;
1484 /* If this is called without a bitmap, that is a bug. But just to be
1485 * robust in case we screwed up elsewhere, in that case pretend there
1486 * was one dirty bit in the requested area, so we won't try to do a
1487 * local read there (no bitmap probably implies no disk) */
1488 ERR_IF(!b
) return 1;
1489 ERR_IF(!b
->bm_pages
) return 1;
1491 spin_lock_irqsave(&b
->bm_lock
, flags
);
1492 if (bm_is_locked(b
))
1493 bm_print_lock_info(mdev
);
1494 for (bitnr
= s
; bitnr
<= e
; bitnr
++) {
1495 unsigned int idx
= bm_bit_to_page_idx(b
, bitnr
);
1496 if (page_nr
!= idx
) {
1500 p_addr
= bm_map_pidx(b
, idx
);
1502 ERR_IF (bitnr
>= b
->bm_bits
) {
1503 dev_err(DEV
, "bitnr=%lu bm_bits=%lu\n", bitnr
, b
->bm_bits
);
1505 c
+= (0 != generic_test_le_bit(bitnr
- (page_nr
<< (PAGE_SHIFT
+3)), p_addr
));
1510 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1515 /* inherently racy...
1516 * return value may be already out-of-date when this function returns.
1517 * but the general usage is that this is only use during a cstate when bits are
1518 * only cleared, not set, and typically only care for the case when the return
1519 * value is zero, or we already "locked" this "bitmap extent" by other means.
1521 * enr is bm-extent number, since we chose to name one sector (512 bytes)
1522 * worth of the bitmap a "bitmap extent".
1525 * I think since we use it like a reference count, we should use the real
1526 * reference count of some bitmap extent element from some lru instead...
1529 int drbd_bm_e_weight(struct drbd_conf
*mdev
, unsigned long enr
)
1531 struct drbd_bitmap
*b
= mdev
->bitmap
;
1533 unsigned long flags
;
1534 unsigned long *p_addr
, *bm
;
1536 ERR_IF(!b
) return 0;
1537 ERR_IF(!b
->bm_pages
) return 0;
1539 spin_lock_irqsave(&b
->bm_lock
, flags
);
1540 if (bm_is_locked(b
))
1541 bm_print_lock_info(mdev
);
1544 e
= min((size_t)S2W(enr
+1), b
->bm_words
);
1546 if (s
< b
->bm_words
) {
1548 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, s
));
1549 bm
= p_addr
+ MLPP(s
);
1551 count
+= hweight_long(*bm
++);
1554 dev_err(DEV
, "start offset (%d) too large in drbd_bm_e_weight\n", s
);
1556 spin_unlock_irqrestore(&b
->bm_lock
, flags
);
1560 /* Set all bits covered by the AL-extent al_enr.
1561 * Returns number of bits changed. */
1562 unsigned long drbd_bm_ALe_set_all(struct drbd_conf
*mdev
, unsigned long al_enr
)
1564 struct drbd_bitmap
*b
= mdev
->bitmap
;
1565 unsigned long *p_addr
, *bm
;
1566 unsigned long weight
;
1568 int count
, i
, do_now
;
1569 ERR_IF(!b
) return 0;
1570 ERR_IF(!b
->bm_pages
) return 0;
1572 spin_lock_irq(&b
->bm_lock
);
1573 if (bm_is_locked(b
))
1574 bm_print_lock_info(mdev
);
1577 s
= al_enr
* BM_WORDS_PER_AL_EXT
;
1578 e
= min_t(size_t, s
+ BM_WORDS_PER_AL_EXT
, b
->bm_words
);
1579 /* assert that s and e are on the same page */
1580 D_ASSERT((e
-1) >> (PAGE_SHIFT
- LN2_BPL
+ 3)
1581 == s
>> (PAGE_SHIFT
- LN2_BPL
+ 3));
1583 if (s
< b
->bm_words
) {
1585 p_addr
= bm_map_pidx(b
, bm_word_to_page_idx(b
, s
));
1586 bm
= p_addr
+ MLPP(s
);
1588 count
+= hweight_long(*bm
);
1593 b
->bm_set
+= do_now
*BITS_PER_LONG
- count
;
1594 if (e
== b
->bm_words
)
1595 b
->bm_set
-= bm_clear_surplus(b
);
1597 dev_err(DEV
, "start offset (%lu) too large in drbd_bm_ALe_set_all\n", s
);
1599 weight
= b
->bm_set
- weight
;
1600 spin_unlock_irq(&b
->bm_lock
);