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".
41 * Note that since find_first_bit returns int, at the current granularity of
42 * the bitmap (4KB per byte), this implementation "only" supports up to
43 * 1<<(32+12) == 16 TB...
48 * Access to the *bm_pages is protected by bm_lock.
49 * It is safe to read the other members within the lock.
51 * drbd_bm_set_bits is called from bio_endio callbacks,
52 * We may be called with irq already disabled,
53 * so we need spin_lock_irqsave().
54 * And we need the kmap_atomic.
57 struct page **bm_pages;
59 /* WARNING unsigned long bm_*:
60 * 32bit number of bit offset is just enough for 512 MB bitmap.
61 * it will blow up if we make the bitmap bigger...
62 * not that it makes much sense to have a bitmap that large,
63 * rather change the granularity to 16k or 64k or something.
64 * (that implies other problems, however...)
66 unsigned long bm_set; /* nr of set bits; THINK maybe atomic_t? */
67 unsigned long bm_bits;
69 size_t bm_number_of_pages;
70 sector_t bm_dev_capacity;
71 struct mutex bm_change; /* serializes resize operations */
74 wait_queue_head_t bm_io_wait;
76 unsigned long bm_flags;
78 /* debugging aid, in case we are still racy somewhere */
80 struct task_struct *bm_task;
83 /* definition of bits in bm_flags */
85 #define BM_MD_IO_ERROR 1
86 #define BM_P_VMALLOCED 2
88 static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
89 unsigned long e, int val, const enum km_type km);
91 static int bm_is_locked(struct drbd_bitmap *b)
93 return test_bit(BM_LOCKED, &b->bm_flags);
96 #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__)
97 static void __bm_print_lock_info(struct drbd_conf *mdev, const char *func)
99 struct drbd_bitmap *b = mdev->bitmap;
100 if (!__ratelimit(&drbd_ratelimit_state))
102 dev_err(DEV, "FIXME %s in %s, bitmap locked for '%s' by %s\n",
103 current == mdev->receiver.task ? "receiver" :
104 current == mdev->asender.task ? "asender" :
105 current == mdev->worker.task ? "worker" : current->comm,
106 func, b->bm_why ?: "?",
107 b->bm_task == mdev->receiver.task ? "receiver" :
108 b->bm_task == mdev->asender.task ? "asender" :
109 b->bm_task == mdev->worker.task ? "worker" : "?");
112 void drbd_bm_lock(struct drbd_conf *mdev, char *why)
114 struct drbd_bitmap *b = mdev->bitmap;
118 dev_err(DEV, "FIXME no bitmap in drbd_bm_lock!?\n");
122 trylock_failed = !mutex_trylock(&b->bm_change);
124 if (trylock_failed) {
125 dev_warn(DEV, "%s going to '%s' but bitmap already locked for '%s' by %s\n",
126 current == mdev->receiver.task ? "receiver" :
127 current == mdev->asender.task ? "asender" :
128 current == mdev->worker.task ? "worker" : current->comm,
129 why, b->bm_why ?: "?",
130 b->bm_task == mdev->receiver.task ? "receiver" :
131 b->bm_task == mdev->asender.task ? "asender" :
132 b->bm_task == mdev->worker.task ? "worker" : "?");
133 mutex_lock(&b->bm_change);
135 if (__test_and_set_bit(BM_LOCKED, &b->bm_flags))
136 dev_err(DEV, "FIXME bitmap already locked in bm_lock\n");
139 b->bm_task = current;
142 void drbd_bm_unlock(struct drbd_conf *mdev)
144 struct drbd_bitmap *b = mdev->bitmap;
146 dev_err(DEV, "FIXME no bitmap in drbd_bm_unlock!?\n");
150 if (!__test_and_clear_bit(BM_LOCKED, &mdev->bitmap->bm_flags))
151 dev_err(DEV, "FIXME bitmap not locked in bm_unlock\n");
155 mutex_unlock(&b->bm_change);
158 static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr)
160 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */
161 unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3);
162 BUG_ON(page_nr >= b->bm_number_of_pages);
166 /* word offset to long pointer */
167 static unsigned long *__bm_map_paddr(struct drbd_bitmap *b, unsigned long offset, const enum km_type km)
170 unsigned long page_nr;
172 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */
173 page_nr = offset >> (PAGE_SHIFT - LN2_BPL + 3);
174 BUG_ON(page_nr >= b->bm_number_of_pages);
175 page = b->bm_pages[page_nr];
177 return (unsigned long *) kmap_atomic(page, km);
180 static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx, const enum km_type km)
182 struct page *page = b->bm_pages[idx];
183 return (unsigned long *) kmap_atomic(page, km);
186 static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx)
188 return __bm_map_pidx(b, idx, KM_IRQ1);
191 static unsigned long * bm_map_paddr(struct drbd_bitmap *b, unsigned long offset)
193 return __bm_map_paddr(b, offset, KM_IRQ1);
196 static void __bm_unmap(unsigned long *p_addr, const enum km_type km)
198 kunmap_atomic(p_addr, km);
201 static void bm_unmap(unsigned long *p_addr)
203 return __bm_unmap(p_addr, KM_IRQ1);
206 /* long word offset of _bitmap_ sector */
207 #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
208 /* word offset from start of bitmap to word number _in_page_
209 * modulo longs per page
210 #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long))
211 hm, well, Philipp thinks gcc might not optimze the % into & (... - 1)
214 #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1))
216 /* Long words per page */
217 #define LWPP (PAGE_SIZE/sizeof(long))
220 * actually most functions herein should take a struct drbd_bitmap*, not a
221 * struct drbd_conf*, but for the debug macros I like to have the mdev around
222 * to be able to report device specific.
225 static void bm_free_pages(struct page **pages, unsigned long number)
231 for (i = 0; i < number; i++) {
233 printk(KERN_ALERT "drbd: bm_free_pages tried to free "
234 "a NULL pointer; i=%lu n=%lu\n",
238 __free_page(pages[i]);
243 static void bm_vk_free(void *ptr, int v)
252 * "have" and "want" are NUMBER OF PAGES.
254 static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want)
256 struct page **old_pages = b->bm_pages;
257 struct page **new_pages, *page;
258 unsigned int i, bytes, vmalloced = 0;
259 unsigned long have = b->bm_number_of_pages;
261 BUG_ON(have == 0 && old_pages != NULL);
262 BUG_ON(have != 0 && old_pages == NULL);
267 /* Trying kmalloc first, falling back to vmalloc.
268 * GFP_KERNEL is ok, as this is done when a lower level disk is
269 * "attached" to the drbd. Context is receiver thread or cqueue
270 * thread. As we have no disk yet, we are not in the IO path,
271 * not even the IO path of the peer. */
272 bytes = sizeof(struct page *)*want;
273 new_pages = kmalloc(bytes, GFP_KERNEL);
275 new_pages = vmalloc(bytes);
281 memset(new_pages, 0, bytes);
283 for (i = 0; i < have; i++)
284 new_pages[i] = old_pages[i];
285 for (; i < want; i++) {
286 page = alloc_page(GFP_HIGHUSER);
288 bm_free_pages(new_pages + have, i - have);
289 bm_vk_free(new_pages, vmalloced);
295 for (i = 0; i < want; i++)
296 new_pages[i] = old_pages[i];
297 /* NOT HERE, we are outside the spinlock!
298 bm_free_pages(old_pages + want, have - want);
303 set_bit(BM_P_VMALLOCED, &b->bm_flags);
305 clear_bit(BM_P_VMALLOCED, &b->bm_flags);
311 * called on driver init only. TODO call when a device is created.
312 * allocates the drbd_bitmap, and stores it in mdev->bitmap.
314 int drbd_bm_init(struct drbd_conf *mdev)
316 struct drbd_bitmap *b = mdev->bitmap;
318 b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL);
321 spin_lock_init(&b->bm_lock);
322 mutex_init(&b->bm_change);
323 init_waitqueue_head(&b->bm_io_wait);
330 sector_t drbd_bm_capacity(struct drbd_conf *mdev)
332 ERR_IF(!mdev->bitmap) return 0;
333 return mdev->bitmap->bm_dev_capacity;
336 /* called on driver unload. TODO: call when a device is destroyed.
338 void drbd_bm_cleanup(struct drbd_conf *mdev)
340 ERR_IF (!mdev->bitmap) return;
341 bm_free_pages(mdev->bitmap->bm_pages, mdev->bitmap->bm_number_of_pages);
342 bm_vk_free(mdev->bitmap->bm_pages, test_bit(BM_P_VMALLOCED, &mdev->bitmap->bm_flags));
348 * since (b->bm_bits % BITS_PER_LONG) != 0,
349 * this masks out the remaining bits.
350 * Returns the number of bits cleared.
352 #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3))
353 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1)
354 #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1)
355 static int bm_clear_surplus(struct drbd_bitmap *b)
358 unsigned long *p_addr, *bm;
362 /* number of bits modulo bits per page */
363 tmp = (b->bm_bits & BITS_PER_PAGE_MASK);
364 /* mask the used bits of the word containing the last bit */
365 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1;
366 /* bitmap is always stored little endian,
367 * on disk and in core memory alike */
368 mask = cpu_to_lel(mask);
370 /* because of the "extra long to catch oob access" we allocate in
371 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
372 * containing the last _relevant_ bitmap word */
373 p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, b->bm_bits - 1));
374 bm = p_addr + (tmp/BITS_PER_LONG);
376 /* If mask != 0, we are not exactly aligned, so bm now points
377 * to the long containing the last bit.
378 * If mask == 0, bm already points to the word immediately
379 * after the last (long word aligned) bit. */
380 cleared = hweight_long(*bm & ~mask);
385 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) {
386 /* on a 32bit arch, we may need to zero out
387 * a padding long to align with a 64bit remote */
388 cleared += hweight_long(*bm);
395 static void bm_set_surplus(struct drbd_bitmap *b)
398 unsigned long *p_addr, *bm;
401 /* number of bits modulo bits per page */
402 tmp = (b->bm_bits & BITS_PER_PAGE_MASK);
403 /* mask the used bits of the word containing the last bit */
404 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1;
405 /* bitmap is always stored little endian,
406 * on disk and in core memory alike */
407 mask = cpu_to_lel(mask);
409 /* because of the "extra long to catch oob access" we allocate in
410 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
411 * containing the last _relevant_ bitmap word */
412 p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, b->bm_bits - 1));
413 bm = p_addr + (tmp/BITS_PER_LONG);
415 /* If mask != 0, we are not exactly aligned, so bm now points
416 * to the long containing the last bit.
417 * If mask == 0, bm already points to the word immediately
418 * after the last (long word aligned) bit. */
423 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) {
424 /* on a 32bit arch, we may need to zero out
425 * a padding long to align with a 64bit remote */
431 static unsigned long bm_count_bits(struct drbd_bitmap *b)
433 unsigned long *p_addr, *bm, offset = 0;
434 unsigned long bits = 0;
435 unsigned long i, do_now;
438 /* due to 64bit alignment, the last long on a 32bit arch
439 * may be not used at all. The last used long will likely
440 * be only partially used, always. Don't count those bits,
441 * but mask them out. */
442 words = (b->bm_bits + BITS_PER_LONG - 1) >> LN2_BPL;
444 while (offset < words) {
445 i = do_now = min_t(size_t, words-offset, LWPP);
446 p_addr = __bm_map_paddr(b, offset, KM_USER0);
447 bm = p_addr + MLPP(offset);
449 bits += hweight_long(*bm++);
452 if (offset == words) {
453 /* last word may only be partially used,
454 * see also bm_clear_surplus. */
455 i = (1UL << (b->bm_bits & (BITS_PER_LONG-1))) -1;
457 bits -= hweight_long(p_addr[do_now-1] & ~i);
458 p_addr[do_now-1] &= i;
460 /* 32bit arch, may have an unused padding long */
461 if (words != b->bm_words)
464 __bm_unmap(p_addr, KM_USER0);
471 /* offset and len in long words.*/
472 static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len)
474 unsigned long *p_addr, *bm;
477 #define BM_SECTORS_PER_BIT (BM_BLOCK_SIZE/512)
481 if (end > b->bm_words) {
482 printk(KERN_ALERT "drbd: bm_memset end > bm_words\n");
486 while (offset < end) {
487 do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset;
488 p_addr = bm_map_paddr(b, offset);
489 bm = p_addr + MLPP(offset);
490 if (bm+do_now > p_addr + LWPP) {
491 printk(KERN_ALERT "drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n",
492 p_addr, bm, (int)do_now);
493 break; /* breaks to after catch_oob_access_end() only! */
495 memset(bm, c, do_now * sizeof(long));
502 * make sure the bitmap has enough room for the attached storage,
503 * if necessary, resize.
504 * called whenever we may have changed the device size.
505 * returns -ENOMEM if we could not allocate enough memory, 0 on success.
506 * In case this is actually a resize, we copy the old bitmap into the new one.
507 * Otherwise, the bitmap is initialized to all bits set.
509 int drbd_bm_resize(struct drbd_conf *mdev, sector_t capacity, int set_new_bits)
511 struct drbd_bitmap *b = mdev->bitmap;
512 unsigned long bits, words, owords, obits, *p_addr, *bm;
513 unsigned long want, have, onpages; /* number of pages */
514 struct page **npages, **opages = NULL;
515 int err = 0, growing;
516 int opages_vmalloced;
518 ERR_IF(!b) return -ENOMEM;
520 drbd_bm_lock(mdev, "resize");
522 dev_info(DEV, "drbd_bm_resize called with capacity == %llu\n",
523 (unsigned long long)capacity);
525 if (capacity == b->bm_dev_capacity)
528 opages_vmalloced = test_bit(BM_P_VMALLOCED, &b->bm_flags);
531 spin_lock_irq(&b->bm_lock);
532 opages = b->bm_pages;
533 onpages = b->bm_number_of_pages;
534 owords = b->bm_words;
536 b->bm_number_of_pages =
540 b->bm_dev_capacity = 0;
541 spin_unlock_irq(&b->bm_lock);
542 bm_free_pages(opages, onpages);
543 bm_vk_free(opages, opages_vmalloced);
546 bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT));
549 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL;
550 a 32bit host could present the wrong number of words
553 words = ALIGN(bits, 64) >> LN2_BPL;
555 if (get_ldev(mdev)) {
556 D_ASSERT((u64)bits <= (((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12));
560 /* one extra long to catch off by one errors */
561 want = ALIGN((words+1)*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT;
562 have = b->bm_number_of_pages;
564 D_ASSERT(b->bm_pages != NULL);
565 npages = b->bm_pages;
567 if (drbd_insert_fault(mdev, DRBD_FAULT_BM_ALLOC))
570 npages = bm_realloc_pages(b, want);
578 spin_lock_irq(&b->bm_lock);
579 opages = b->bm_pages;
580 owords = b->bm_words;
583 growing = bits > obits;
584 if (opages && growing && set_new_bits)
587 b->bm_pages = npages;
588 b->bm_number_of_pages = want;
591 b->bm_dev_capacity = capacity;
595 bm_memset(b, owords, 0xff, words-owords);
596 b->bm_set += bits - obits;
598 bm_memset(b, owords, 0x00, words-owords);
603 /* implicit: (opages != NULL) && (opages != npages) */
604 bm_free_pages(opages + want, have - want);
607 p_addr = bm_map_paddr(b, words);
608 bm = p_addr + MLPP(words);
612 (void)bm_clear_surplus(b);
614 spin_unlock_irq(&b->bm_lock);
615 if (opages != npages)
616 bm_vk_free(opages, opages_vmalloced);
618 b->bm_set = bm_count_bits(b);
619 dev_info(DEV, "resync bitmap: bits=%lu words=%lu\n", bits, words);
622 drbd_bm_unlock(mdev);
627 * if not protected by other means, return value may be out of date when
628 * leaving this function...
629 * we still need to lock it, since it is important that this returns
630 * bm_set == 0 precisely.
632 * maybe bm_set should be atomic_t ?
634 unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev)
636 struct drbd_bitmap *b = mdev->bitmap;
641 ERR_IF(!b->bm_pages) return 0;
643 spin_lock_irqsave(&b->bm_lock, flags);
645 spin_unlock_irqrestore(&b->bm_lock, flags);
650 unsigned long drbd_bm_total_weight(struct drbd_conf *mdev)
653 /* if I don't have a disk, I don't know about out-of-sync status */
654 if (!get_ldev_if_state(mdev, D_NEGOTIATING))
656 s = _drbd_bm_total_weight(mdev);
661 size_t drbd_bm_words(struct drbd_conf *mdev)
663 struct drbd_bitmap *b = mdev->bitmap;
665 ERR_IF(!b->bm_pages) return 0;
670 unsigned long drbd_bm_bits(struct drbd_conf *mdev)
672 struct drbd_bitmap *b = mdev->bitmap;
678 /* merge number words from buffer into the bitmap starting at offset.
679 * buffer[i] is expected to be little endian unsigned long.
680 * bitmap must be locked by drbd_bm_lock.
681 * currently only used from receive_bitmap.
683 void drbd_bm_merge_lel(struct drbd_conf *mdev, size_t offset, size_t number,
684 unsigned long *buffer)
686 struct drbd_bitmap *b = mdev->bitmap;
687 unsigned long *p_addr, *bm;
688 unsigned long word, bits;
691 end = offset + number;
694 ERR_IF(!b->bm_pages) return;
697 WARN_ON(offset >= b->bm_words);
698 WARN_ON(end > b->bm_words);
700 spin_lock_irq(&b->bm_lock);
701 while (offset < end) {
702 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset;
703 p_addr = bm_map_paddr(b, offset);
704 bm = p_addr + MLPP(offset);
707 bits = hweight_long(*bm);
708 word = *bm | *buffer++;
710 b->bm_set += hweight_long(word) - bits;
714 /* with 32bit <-> 64bit cross-platform connect
715 * this is only correct for current usage,
716 * where we _know_ that we are 64 bit aligned,
717 * and know that this function is used in this way, too...
719 if (end == b->bm_words)
720 b->bm_set -= bm_clear_surplus(b);
722 spin_unlock_irq(&b->bm_lock);
725 /* copy number words from the bitmap starting at offset into the buffer.
726 * buffer[i] will be little endian unsigned long.
728 void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset, size_t number,
729 unsigned long *buffer)
731 struct drbd_bitmap *b = mdev->bitmap;
732 unsigned long *p_addr, *bm;
735 end = offset + number;
738 ERR_IF(!b->bm_pages) return;
740 spin_lock_irq(&b->bm_lock);
741 if ((offset >= b->bm_words) ||
742 (end > b->bm_words) ||
744 dev_err(DEV, "offset=%lu number=%lu bm_words=%lu\n",
745 (unsigned long) offset,
746 (unsigned long) number,
747 (unsigned long) b->bm_words);
749 while (offset < end) {
750 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset;
751 p_addr = bm_map_paddr(b, offset);
752 bm = p_addr + MLPP(offset);
759 spin_unlock_irq(&b->bm_lock);
762 /* set all bits in the bitmap */
763 void drbd_bm_set_all(struct drbd_conf *mdev)
765 struct drbd_bitmap *b = mdev->bitmap;
767 ERR_IF(!b->bm_pages) return;
769 spin_lock_irq(&b->bm_lock);
770 bm_memset(b, 0, 0xff, b->bm_words);
771 (void)bm_clear_surplus(b);
772 b->bm_set = b->bm_bits;
773 spin_unlock_irq(&b->bm_lock);
776 /* clear all bits in the bitmap */
777 void drbd_bm_clear_all(struct drbd_conf *mdev)
779 struct drbd_bitmap *b = mdev->bitmap;
781 ERR_IF(!b->bm_pages) return;
783 spin_lock_irq(&b->bm_lock);
784 bm_memset(b, 0, 0, b->bm_words);
786 spin_unlock_irq(&b->bm_lock);
789 static void bm_async_io_complete(struct bio *bio, int error)
791 struct drbd_bitmap *b = bio->bi_private;
792 int uptodate = bio_flagged(bio, BIO_UPTODATE);
795 /* strange behavior of some lower level drivers...
796 * fail the request by clearing the uptodate flag,
797 * but do not return any error?!
798 * do we want to WARN() on this? */
799 if (!error && !uptodate)
804 * for now, set all bits, and flag MD_IO_ERROR */
805 __set_bit(BM_MD_IO_ERROR, &b->bm_flags);
807 if (atomic_dec_and_test(&b->bm_async_io))
808 wake_up(&b->bm_io_wait);
813 static void bm_page_io_async(struct drbd_conf *mdev, struct drbd_bitmap *b, int page_nr, int rw) __must_hold(local)
815 /* we are process context. we always get a bio */
816 struct bio *bio = bio_alloc(GFP_KERNEL, 1);
818 sector_t on_disk_sector =
819 mdev->ldev->md.md_offset + mdev->ldev->md.bm_offset;
820 on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9);
822 /* this might happen with very small
823 * flexible external meta data device */
824 len = min_t(unsigned int, PAGE_SIZE,
825 (drbd_md_last_sector(mdev->ldev) - on_disk_sector + 1)<<9);
827 bio->bi_bdev = mdev->ldev->md_bdev;
828 bio->bi_sector = on_disk_sector;
829 bio_add_page(bio, b->bm_pages[page_nr], len, 0);
831 bio->bi_end_io = bm_async_io_complete;
833 if (drbd_insert_fault(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) {
835 bio_endio(bio, -EIO);
842 * bm_rw: read/write the whole bitmap from/to its on disk location.
844 static int bm_rw(struct drbd_conf *mdev, int rw) __must_hold(local)
846 struct drbd_bitmap *b = mdev->bitmap;
847 /* sector_t sector; */
848 int bm_words, num_pages, i;
853 WARN_ON(!bm_is_locked(b));
855 /* no spinlock here, the drbd_bm_lock should be enough! */
857 bm_words = drbd_bm_words(mdev);
858 num_pages = (bm_words*sizeof(long) + PAGE_SIZE-1) >> PAGE_SHIFT;
861 atomic_set(&b->bm_async_io, num_pages);
862 __clear_bit(BM_MD_IO_ERROR, &b->bm_flags);
864 /* let the layers below us try to merge these bios... */
865 for (i = 0; i < num_pages; i++)
866 bm_page_io_async(mdev, b, i, rw);
868 wait_event(b->bm_io_wait, atomic_read(&b->bm_async_io) == 0);
870 if (test_bit(BM_MD_IO_ERROR, &b->bm_flags)) {
871 dev_alert(DEV, "we had at least one MD IO ERROR during bitmap IO\n");
872 drbd_chk_io_error(mdev, 1, true);
879 } else /* rw == READ */ {
880 b->bm_set = bm_count_bits(b);
881 dev_info(DEV, "recounting of set bits took additional %lu jiffies\n",
886 dev_info(DEV, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n",
887 ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now);
893 * drbd_bm_read() - Read the whole bitmap from its on disk location.
894 * @mdev: DRBD device.
896 int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local)
898 return bm_rw(mdev, READ);
902 * drbd_bm_write() - Write the whole bitmap to its on disk location.
903 * @mdev: DRBD device.
905 int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local)
907 return bm_rw(mdev, WRITE);
911 * drbd_bm_write_sect: Writes a 512 (MD_SECTOR_SIZE) byte piece of the bitmap
912 * @mdev: DRBD device.
913 * @enr: Extent number in the resync lru (happens to be sector offset)
915 * The BM_EXT_SIZE is on purpose exactly the amount of the bitmap covered
916 * by a single sector write. Therefore enr == sector offset from the
917 * start of the bitmap.
919 int drbd_bm_write_sect(struct drbd_conf *mdev, unsigned long enr) __must_hold(local)
921 sector_t on_disk_sector = enr + mdev->ldev->md.md_offset
922 + mdev->ldev->md.bm_offset;
923 int bm_words, num_words, offset;
926 mutex_lock(&mdev->md_io_mutex);
927 bm_words = drbd_bm_words(mdev);
928 offset = S2W(enr); /* word offset into bitmap */
929 num_words = min(S2W(1), bm_words - offset);
930 if (num_words < S2W(1))
931 memset(page_address(mdev->md_io_page), 0, MD_SECTOR_SIZE);
932 drbd_bm_get_lel(mdev, offset, num_words,
933 page_address(mdev->md_io_page));
934 if (!drbd_md_sync_page_io(mdev, mdev->ldev, on_disk_sector, WRITE)) {
937 dev_err(DEV, "IO ERROR writing bitmap sector %lu "
938 "(meta-disk sector %llus)\n",
939 enr, (unsigned long long)on_disk_sector);
940 drbd_chk_io_error(mdev, 1, true);
941 for (i = 0; i < AL_EXT_PER_BM_SECT; i++)
942 drbd_bm_ALe_set_all(mdev, enr*AL_EXT_PER_BM_SECT+i);
945 mutex_unlock(&mdev->md_io_mutex);
950 * find_first_bit returns int, we return unsigned long.
951 * should not make much difference anyways, but ...
953 * this returns a bit number, NOT a sector!
955 #define BPP_MASK ((1UL << (PAGE_SHIFT+3)) - 1)
956 static unsigned long __bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo,
957 const int find_zero_bit, const enum km_type km)
959 struct drbd_bitmap *b = mdev->bitmap;
960 unsigned long i = -1UL;
961 unsigned long *p_addr;
962 unsigned long bit_offset; /* bit offset of the mapped page. */
964 if (bm_fo > b->bm_bits) {
965 dev_err(DEV, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits);
967 while (bm_fo < b->bm_bits) {
968 unsigned long offset;
969 bit_offset = bm_fo & ~BPP_MASK; /* bit offset of the page */
970 offset = bit_offset >> LN2_BPL; /* word offset of the page */
971 p_addr = __bm_map_paddr(b, offset, km);
974 i = generic_find_next_zero_le_bit(p_addr, PAGE_SIZE*8, bm_fo & BPP_MASK);
976 i = generic_find_next_le_bit(p_addr, PAGE_SIZE*8, bm_fo & BPP_MASK);
978 __bm_unmap(p_addr, km);
979 if (i < PAGE_SIZE*8) {
985 bm_fo = bit_offset + PAGE_SIZE*8;
993 static unsigned long bm_find_next(struct drbd_conf *mdev,
994 unsigned long bm_fo, const int find_zero_bit)
996 struct drbd_bitmap *b = mdev->bitmap;
997 unsigned long i = -1UL;
1000 ERR_IF(!b->bm_pages) return i;
1002 spin_lock_irq(&b->bm_lock);
1003 if (bm_is_locked(b))
1004 bm_print_lock_info(mdev);
1006 i = __bm_find_next(mdev, bm_fo, find_zero_bit, KM_IRQ1);
1008 spin_unlock_irq(&b->bm_lock);
1012 unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo)
1014 return bm_find_next(mdev, bm_fo, 0);
1018 /* not yet needed for anything. */
1019 unsigned long drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo)
1021 return bm_find_next(mdev, bm_fo, 1);
1025 /* does not spin_lock_irqsave.
1026 * you must take drbd_bm_lock() first */
1027 unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo)
1029 /* WARN_ON(!bm_is_locked(mdev)); */
1030 return __bm_find_next(mdev, bm_fo, 0, KM_USER1);
1033 unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo)
1035 /* WARN_ON(!bm_is_locked(mdev)); */
1036 return __bm_find_next(mdev, bm_fo, 1, KM_USER1);
1039 /* returns number of bits actually changed.
1040 * for val != 0, we change 0 -> 1, return code positive
1041 * for val == 0, we change 1 -> 0, return code negative
1042 * wants bitnr, not sector.
1043 * expected to be called for only a few bits (e - s about BITS_PER_LONG).
1044 * Must hold bitmap lock already. */
1045 static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
1046 unsigned long e, int val, const enum km_type km)
1048 struct drbd_bitmap *b = mdev->bitmap;
1049 unsigned long *p_addr = NULL;
1050 unsigned long bitnr;
1051 unsigned long last_page_nr = -1UL;
1054 if (e >= b->bm_bits) {
1055 dev_err(DEV, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n",
1057 e = b->bm_bits ? b->bm_bits -1 : 0;
1059 for (bitnr = s; bitnr <= e; bitnr++) {
1060 unsigned long offset = bitnr>>LN2_BPL;
1061 unsigned long page_nr = offset >> (PAGE_SHIFT - LN2_BPL + 3);
1062 if (page_nr != last_page_nr) {
1064 __bm_unmap(p_addr, km);
1065 p_addr = __bm_map_paddr(b, offset, km);
1066 last_page_nr = page_nr;
1069 c += (0 == generic___test_and_set_le_bit(bitnr & BPP_MASK, p_addr));
1071 c -= (0 != generic___test_and_clear_le_bit(bitnr & BPP_MASK, p_addr));
1074 __bm_unmap(p_addr, km);
1079 /* returns number of bits actually changed.
1080 * for val != 0, we change 0 -> 1, return code positive
1081 * for val == 0, we change 1 -> 0, return code negative
1082 * wants bitnr, not sector */
1083 static int bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
1084 const unsigned long e, int val)
1086 unsigned long flags;
1087 struct drbd_bitmap *b = mdev->bitmap;
1090 ERR_IF(!b) return 1;
1091 ERR_IF(!b->bm_pages) return 0;
1093 spin_lock_irqsave(&b->bm_lock, flags);
1094 if (bm_is_locked(b))
1095 bm_print_lock_info(mdev);
1097 c = __bm_change_bits_to(mdev, s, e, val, KM_IRQ1);
1099 spin_unlock_irqrestore(&b->bm_lock, flags);
1103 /* returns number of bits changed 0 -> 1 */
1104 int drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1106 return bm_change_bits_to(mdev, s, e, 1);
1109 /* returns number of bits changed 1 -> 0 */
1110 int drbd_bm_clear_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1112 return -bm_change_bits_to(mdev, s, e, 0);
1115 /* sets all bits in full words,
1116 * from first_word up to, but not including, last_word */
1117 static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b,
1118 int page_nr, int first_word, int last_word)
1122 unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr], KM_USER0);
1123 for (i = first_word; i < last_word; i++) {
1124 bits = hweight_long(paddr[i]);
1126 b->bm_set += BITS_PER_LONG - bits;
1128 kunmap_atomic(paddr, KM_USER0);
1131 /* Same thing as drbd_bm_set_bits, but without taking the spin_lock_irqsave.
1132 * You must first drbd_bm_lock().
1133 * Can be called to set the whole bitmap in one go.
1134 * Sets bits from s to e _inclusive_. */
1135 void _drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1137 /* First set_bit from the first bit (s)
1138 * up to the next long boundary (sl),
1139 * then assign full words up to the last long boundary (el),
1140 * then set_bit up to and including the last bit (e).
1142 * Do not use memset, because we must account for changes,
1143 * so we need to loop over the words with hweight() anyways.
1145 unsigned long sl = ALIGN(s,BITS_PER_LONG);
1146 unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1);
1153 if (e - s <= 3*BITS_PER_LONG) {
1154 /* don't bother; el and sl may even be wrong. */
1155 __bm_change_bits_to(mdev, s, e, 1, KM_USER0);
1159 /* difference is large enough that we can trust sl and el */
1161 /* bits filling the current long */
1163 __bm_change_bits_to(mdev, s, sl-1, 1, KM_USER0);
1165 first_page = sl >> (3 + PAGE_SHIFT);
1166 last_page = el >> (3 + PAGE_SHIFT);
1168 /* MLPP: modulo longs per page */
1169 /* LWPP: long words per page */
1170 first_word = MLPP(sl >> LN2_BPL);
1173 /* first and full pages, unless first page == last page */
1174 for (page_nr = first_page; page_nr < last_page; page_nr++) {
1175 bm_set_full_words_within_one_page(mdev->bitmap, page_nr, first_word, last_word);
1180 /* last page (respectively only page, for first page == last page) */
1181 last_word = MLPP(el >> LN2_BPL);
1182 bm_set_full_words_within_one_page(mdev->bitmap, last_page, first_word, last_word);
1184 /* possibly trailing bits.
1185 * example: (e & 63) == 63, el will be e+1.
1186 * if that even was the very last bit,
1187 * it would trigger an assert in __bm_change_bits_to()
1190 __bm_change_bits_to(mdev, el, e, 1, KM_USER0);
1193 /* returns bit state
1194 * wants bitnr, NOT sector.
1195 * inherently racy... area needs to be locked by means of {al,rs}_lru
1198 * -1 ... first out of bounds access, stop testing for bits!
1200 int drbd_bm_test_bit(struct drbd_conf *mdev, const unsigned long bitnr)
1202 unsigned long flags;
1203 struct drbd_bitmap *b = mdev->bitmap;
1204 unsigned long *p_addr;
1207 ERR_IF(!b) return 0;
1208 ERR_IF(!b->bm_pages) return 0;
1210 spin_lock_irqsave(&b->bm_lock, flags);
1211 if (bm_is_locked(b))
1212 bm_print_lock_info(mdev);
1213 if (bitnr < b->bm_bits) {
1214 unsigned long offset = bitnr>>LN2_BPL;
1215 p_addr = bm_map_paddr(b, offset);
1216 i = generic_test_le_bit(bitnr & BPP_MASK, p_addr) ? 1 : 0;
1218 } else if (bitnr == b->bm_bits) {
1220 } else { /* (bitnr > b->bm_bits) */
1221 dev_err(DEV, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits);
1225 spin_unlock_irqrestore(&b->bm_lock, flags);
1229 /* returns number of bits set in the range [s, e] */
1230 int drbd_bm_count_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1232 unsigned long flags;
1233 struct drbd_bitmap *b = mdev->bitmap;
1234 unsigned long *p_addr = NULL, page_nr = -1;
1235 unsigned long bitnr;
1239 /* If this is called without a bitmap, that is a bug. But just to be
1240 * robust in case we screwed up elsewhere, in that case pretend there
1241 * was one dirty bit in the requested area, so we won't try to do a
1242 * local read there (no bitmap probably implies no disk) */
1243 ERR_IF(!b) return 1;
1244 ERR_IF(!b->bm_pages) return 1;
1246 spin_lock_irqsave(&b->bm_lock, flags);
1247 if (bm_is_locked(b))
1248 bm_print_lock_info(mdev);
1249 for (bitnr = s; bitnr <= e; bitnr++) {
1250 w = bitnr >> LN2_BPL;
1251 if (page_nr != w >> (PAGE_SHIFT - LN2_BPL + 3)) {
1252 page_nr = w >> (PAGE_SHIFT - LN2_BPL + 3);
1255 p_addr = bm_map_paddr(b, w);
1257 ERR_IF (bitnr >= b->bm_bits) {
1258 dev_err(DEV, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits);
1260 c += (0 != generic_test_le_bit(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr));
1265 spin_unlock_irqrestore(&b->bm_lock, flags);
1270 /* inherently racy...
1271 * return value may be already out-of-date when this function returns.
1272 * but the general usage is that this is only use during a cstate when bits are
1273 * only cleared, not set, and typically only care for the case when the return
1274 * value is zero, or we already "locked" this "bitmap extent" by other means.
1276 * enr is bm-extent number, since we chose to name one sector (512 bytes)
1277 * worth of the bitmap a "bitmap extent".
1280 * I think since we use it like a reference count, we should use the real
1281 * reference count of some bitmap extent element from some lru instead...
1284 int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr)
1286 struct drbd_bitmap *b = mdev->bitmap;
1288 unsigned long flags;
1289 unsigned long *p_addr, *bm;
1291 ERR_IF(!b) return 0;
1292 ERR_IF(!b->bm_pages) return 0;
1294 spin_lock_irqsave(&b->bm_lock, flags);
1295 if (bm_is_locked(b))
1296 bm_print_lock_info(mdev);
1299 e = min((size_t)S2W(enr+1), b->bm_words);
1301 if (s < b->bm_words) {
1303 p_addr = bm_map_paddr(b, s);
1304 bm = p_addr + MLPP(s);
1306 count += hweight_long(*bm++);
1309 dev_err(DEV, "start offset (%d) too large in drbd_bm_e_weight\n", s);
1311 spin_unlock_irqrestore(&b->bm_lock, flags);
1315 /* set all bits covered by the AL-extent al_enr */
1316 unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev, unsigned long al_enr)
1318 struct drbd_bitmap *b = mdev->bitmap;
1319 unsigned long *p_addr, *bm;
1320 unsigned long weight;
1321 int count, s, e, i, do_now;
1322 ERR_IF(!b) return 0;
1323 ERR_IF(!b->bm_pages) return 0;
1325 spin_lock_irq(&b->bm_lock);
1326 if (bm_is_locked(b))
1327 bm_print_lock_info(mdev);
1330 s = al_enr * BM_WORDS_PER_AL_EXT;
1331 e = min_t(size_t, s + BM_WORDS_PER_AL_EXT, b->bm_words);
1332 /* assert that s and e are on the same page */
1333 D_ASSERT((e-1) >> (PAGE_SHIFT - LN2_BPL + 3)
1334 == s >> (PAGE_SHIFT - LN2_BPL + 3));
1336 if (s < b->bm_words) {
1338 p_addr = bm_map_paddr(b, s);
1339 bm = p_addr + MLPP(s);
1341 count += hweight_long(*bm);
1346 b->bm_set += do_now*BITS_PER_LONG - count;
1347 if (e == b->bm_words)
1348 b->bm_set -= bm_clear_surplus(b);
1350 dev_err(DEV, "start offset (%d) too large in drbd_bm_ALe_set_all\n", s);
1352 weight = b->bm_set - weight;
1353 spin_unlock_irq(&b->bm_lock);