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 drbd_task_to_thread_name(mdev->tconn, current),
123 func, b->bm_why ?: "?",
124 drbd_task_to_thread_name(mdev->tconn, b->bm_task));
127 void drbd_bm_lock(struct drbd_conf *mdev, char *why, enum bm_flag flags)
129 struct drbd_bitmap *b = mdev->bitmap;
133 dev_err(DEV, "FIXME no bitmap in drbd_bm_lock!?\n");
137 trylock_failed = !mutex_trylock(&b->bm_change);
139 if (trylock_failed) {
140 dev_warn(DEV, "%s going to '%s' but bitmap already locked for '%s' by %s\n",
141 drbd_task_to_thread_name(mdev->tconn, current),
142 why, b->bm_why ?: "?",
143 drbd_task_to_thread_name(mdev->tconn, b->bm_task));
144 mutex_lock(&b->bm_change);
146 if (BM_LOCKED_MASK & b->bm_flags)
147 dev_err(DEV, "FIXME bitmap already locked in bm_lock\n");
148 b->bm_flags |= flags & BM_LOCKED_MASK;
151 b->bm_task = current;
154 void drbd_bm_unlock(struct drbd_conf *mdev)
156 struct drbd_bitmap *b = mdev->bitmap;
158 dev_err(DEV, "FIXME no bitmap in drbd_bm_unlock!?\n");
162 if (!(BM_LOCKED_MASK & mdev->bitmap->bm_flags))
163 dev_err(DEV, "FIXME bitmap not locked in bm_unlock\n");
165 b->bm_flags &= ~BM_LOCKED_MASK;
168 mutex_unlock(&b->bm_change);
171 /* we store some "meta" info about our pages in page->private */
172 /* at a granularity of 4k storage per bitmap bit:
173 * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks
175 * 1<<23 4k bitmap pages.
176 * Use 24 bits as page index, covers 2 peta byte storage
177 * at a granularity of 4k per bit.
178 * Used to report the failed page idx on io error from the endio handlers.
180 #define BM_PAGE_IDX_MASK ((1UL<<24)-1)
181 /* this page is currently read in, or written back */
182 #define BM_PAGE_IO_LOCK 31
183 /* if there has been an IO error for this page */
184 #define BM_PAGE_IO_ERROR 30
185 /* this is to be able to intelligently skip disk IO,
186 * set if bits have been set since last IO. */
187 #define BM_PAGE_NEED_WRITEOUT 29
188 /* to mark for lazy writeout once syncer cleared all clearable bits,
189 * we if bits have been cleared since last IO. */
190 #define BM_PAGE_LAZY_WRITEOUT 28
191 /* pages marked with this "HINT" will be considered for writeout
192 * on activity log transactions */
193 #define BM_PAGE_HINT_WRITEOUT 27
195 /* store_page_idx uses non-atomic assignment. It is only used directly after
196 * allocating the page. All other bm_set_page_* and bm_clear_page_* need to
197 * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap
198 * changes) may happen from various contexts, and wait_on_bit/wake_up_bit
199 * requires it all to be atomic as well. */
200 static void bm_store_page_idx(struct page *page, unsigned long idx)
202 BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK));
203 set_page_private(page, idx);
206 static unsigned long bm_page_to_idx(struct page *page)
208 return page_private(page) & BM_PAGE_IDX_MASK;
211 /* As is very unlikely that the same page is under IO from more than one
212 * context, we can get away with a bit per page and one wait queue per bitmap.
214 static void bm_page_lock_io(struct drbd_conf *mdev, int page_nr)
216 struct drbd_bitmap *b = mdev->bitmap;
217 void *addr = &page_private(b->bm_pages[page_nr]);
218 wait_event(b->bm_io_wait, !test_and_set_bit(BM_PAGE_IO_LOCK, addr));
221 static void bm_page_unlock_io(struct drbd_conf *mdev, int page_nr)
223 struct drbd_bitmap *b = mdev->bitmap;
224 void *addr = &page_private(b->bm_pages[page_nr]);
225 clear_bit_unlock(BM_PAGE_IO_LOCK, addr);
226 wake_up(&mdev->bitmap->bm_io_wait);
229 /* set _before_ submit_io, so it may be reset due to being changed
230 * while this page is in flight... will get submitted later again */
231 static void bm_set_page_unchanged(struct page *page)
234 clear_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page));
235 clear_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page));
238 static void bm_set_page_need_writeout(struct page *page)
240 set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page));
244 * drbd_bm_mark_for_writeout() - mark a page with a "hint" to be considered for writeout
245 * @mdev: DRBD device.
246 * @page_nr: the bitmap page to mark with the "hint" flag
248 * From within an activity log transaction, we mark a few pages with these
249 * hints, then call drbd_bm_write_hinted(), which will only write out changed
250 * pages which are flagged with this mark.
252 void drbd_bm_mark_for_writeout(struct drbd_conf *mdev, int page_nr)
255 if (page_nr >= mdev->bitmap->bm_number_of_pages) {
256 dev_warn(DEV, "BAD: page_nr: %u, number_of_pages: %u\n",
257 page_nr, (int)mdev->bitmap->bm_number_of_pages);
260 page = mdev->bitmap->bm_pages[page_nr];
261 set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page));
264 static int bm_test_page_unchanged(struct page *page)
266 volatile const unsigned long *addr = &page_private(page);
267 return (*addr & ((1UL<<BM_PAGE_NEED_WRITEOUT)|(1UL<<BM_PAGE_LAZY_WRITEOUT))) == 0;
270 static void bm_set_page_io_err(struct page *page)
272 set_bit(BM_PAGE_IO_ERROR, &page_private(page));
275 static void bm_clear_page_io_err(struct page *page)
277 clear_bit(BM_PAGE_IO_ERROR, &page_private(page));
280 static void bm_set_page_lazy_writeout(struct page *page)
282 set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page));
285 static int bm_test_page_lazy_writeout(struct page *page)
287 return test_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page));
290 /* on a 32bit box, this would allow for exactly (2<<38) bits. */
291 static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr)
293 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */
294 unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3);
295 BUG_ON(page_nr >= b->bm_number_of_pages);
299 static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr)
301 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */
302 unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3);
303 BUG_ON(page_nr >= b->bm_number_of_pages);
307 static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx, const enum km_type km)
309 struct page *page = b->bm_pages[idx];
310 return (unsigned long *) kmap_atomic(page, km);
313 static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx)
315 return __bm_map_pidx(b, idx, KM_IRQ1);
318 static void __bm_unmap(unsigned long *p_addr, const enum km_type km)
320 kunmap_atomic(p_addr, km);
323 static void bm_unmap(unsigned long *p_addr)
325 return __bm_unmap(p_addr, KM_IRQ1);
328 /* long word offset of _bitmap_ sector */
329 #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
330 /* word offset from start of bitmap to word number _in_page_
331 * modulo longs per page
332 #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long))
333 hm, well, Philipp thinks gcc might not optimize the % into & (... - 1)
336 #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1))
338 /* Long words per page */
339 #define LWPP (PAGE_SIZE/sizeof(long))
342 * actually most functions herein should take a struct drbd_bitmap*, not a
343 * struct drbd_conf*, but for the debug macros I like to have the mdev around
344 * to be able to report device specific.
348 static void bm_free_pages(struct page **pages, unsigned long number)
354 for (i = 0; i < number; i++) {
356 printk(KERN_ALERT "drbd: bm_free_pages tried to free "
357 "a NULL pointer; i=%lu n=%lu\n",
361 __free_page(pages[i]);
366 static void bm_vk_free(void *ptr, int v)
375 * "have" and "want" are NUMBER OF PAGES.
377 static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want)
379 struct page **old_pages = b->bm_pages;
380 struct page **new_pages, *page;
381 unsigned int i, bytes, vmalloced = 0;
382 unsigned long have = b->bm_number_of_pages;
384 BUG_ON(have == 0 && old_pages != NULL);
385 BUG_ON(have != 0 && old_pages == NULL);
390 /* Trying kmalloc first, falling back to vmalloc.
391 * GFP_NOIO, as this is called while drbd IO is "suspended",
392 * and during resize or attach on diskless Primary,
393 * we must not block on IO to ourselves.
394 * Context is receiver thread or dmsetup. */
395 bytes = sizeof(struct page *)*want;
396 new_pages = kmalloc(bytes, GFP_NOIO);
398 new_pages = __vmalloc(bytes,
399 GFP_NOIO | __GFP_HIGHMEM,
406 memset(new_pages, 0, bytes);
408 for (i = 0; i < have; i++)
409 new_pages[i] = old_pages[i];
410 for (; i < want; i++) {
411 page = alloc_page(GFP_NOIO | __GFP_HIGHMEM);
413 bm_free_pages(new_pages + have, i - have);
414 bm_vk_free(new_pages, vmalloced);
417 /* we want to know which page it is
418 * from the endio handlers */
419 bm_store_page_idx(page, i);
423 for (i = 0; i < want; i++)
424 new_pages[i] = old_pages[i];
425 /* NOT HERE, we are outside the spinlock!
426 bm_free_pages(old_pages + want, have - want);
431 b->bm_flags |= BM_P_VMALLOCED;
433 b->bm_flags &= ~BM_P_VMALLOCED;
439 * called on driver init only. TODO call when a device is created.
440 * allocates the drbd_bitmap, and stores it in mdev->bitmap.
442 int drbd_bm_init(struct drbd_conf *mdev)
444 struct drbd_bitmap *b = mdev->bitmap;
446 b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL);
449 spin_lock_init(&b->bm_lock);
450 mutex_init(&b->bm_change);
451 init_waitqueue_head(&b->bm_io_wait);
458 sector_t drbd_bm_capacity(struct drbd_conf *mdev)
460 if (!expect(mdev->bitmap))
462 return mdev->bitmap->bm_dev_capacity;
465 /* called on driver unload. TODO: call when a device is destroyed.
467 void drbd_bm_cleanup(struct drbd_conf *mdev)
469 if (!expect(mdev->bitmap))
471 bm_free_pages(mdev->bitmap->bm_pages, mdev->bitmap->bm_number_of_pages);
472 bm_vk_free(mdev->bitmap->bm_pages, (BM_P_VMALLOCED & mdev->bitmap->bm_flags));
478 * since (b->bm_bits % BITS_PER_LONG) != 0,
479 * this masks out the remaining bits.
480 * Returns the number of bits cleared.
482 #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3))
483 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1)
484 #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1)
485 static int bm_clear_surplus(struct drbd_bitmap *b)
488 unsigned long *p_addr, *bm;
492 /* number of bits modulo bits per page */
493 tmp = (b->bm_bits & BITS_PER_PAGE_MASK);
494 /* mask the used bits of the word containing the last bit */
495 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1;
496 /* bitmap is always stored little endian,
497 * on disk and in core memory alike */
498 mask = cpu_to_lel(mask);
500 p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1);
501 bm = p_addr + (tmp/BITS_PER_LONG);
503 /* If mask != 0, we are not exactly aligned, so bm now points
504 * to the long containing the last bit.
505 * If mask == 0, bm already points to the word immediately
506 * after the last (long word aligned) bit. */
507 cleared = hweight_long(*bm & ~mask);
512 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) {
513 /* on a 32bit arch, we may need to zero out
514 * a padding long to align with a 64bit remote */
515 cleared += hweight_long(*bm);
522 static void bm_set_surplus(struct drbd_bitmap *b)
525 unsigned long *p_addr, *bm;
528 /* number of bits modulo bits per page */
529 tmp = (b->bm_bits & BITS_PER_PAGE_MASK);
530 /* mask the used bits of the word containing the last bit */
531 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1;
532 /* bitmap is always stored little endian,
533 * on disk and in core memory alike */
534 mask = cpu_to_lel(mask);
536 p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1);
537 bm = p_addr + (tmp/BITS_PER_LONG);
539 /* If mask != 0, we are not exactly aligned, so bm now points
540 * to the long containing the last bit.
541 * If mask == 0, bm already points to the word immediately
542 * after the last (long word aligned) bit. */
547 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) {
548 /* on a 32bit arch, we may need to zero out
549 * a padding long to align with a 64bit remote */
555 /* you better not modify the bitmap while this is running,
556 * or its results will be stale */
557 static unsigned long bm_count_bits(struct drbd_bitmap *b)
559 unsigned long *p_addr;
560 unsigned long bits = 0;
561 unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1;
562 int idx, i, last_word;
564 /* all but last page */
565 for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) {
566 p_addr = __bm_map_pidx(b, idx, KM_USER0);
567 for (i = 0; i < LWPP; i++)
568 bits += hweight_long(p_addr[i]);
569 __bm_unmap(p_addr, KM_USER0);
572 /* last (or only) page */
573 last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL;
574 p_addr = __bm_map_pidx(b, idx, KM_USER0);
575 for (i = 0; i < last_word; i++)
576 bits += hweight_long(p_addr[i]);
577 p_addr[last_word] &= cpu_to_lel(mask);
578 bits += hweight_long(p_addr[last_word]);
579 /* 32bit arch, may have an unused padding long */
580 if (BITS_PER_LONG == 32 && (last_word & 1) == 0)
581 p_addr[last_word+1] = 0;
582 __bm_unmap(p_addr, KM_USER0);
586 /* offset and len in long words.*/
587 static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len)
589 unsigned long *p_addr, *bm;
595 if (end > b->bm_words) {
596 printk(KERN_ALERT "drbd: bm_memset end > bm_words\n");
600 while (offset < end) {
601 do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset;
602 idx = bm_word_to_page_idx(b, offset);
603 p_addr = bm_map_pidx(b, idx);
604 bm = p_addr + MLPP(offset);
605 if (bm+do_now > p_addr + LWPP) {
606 printk(KERN_ALERT "drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n",
607 p_addr, bm, (int)do_now);
609 memset(bm, c, do_now * sizeof(long));
611 bm_set_page_need_writeout(b->bm_pages[idx]);
617 * make sure the bitmap has enough room for the attached storage,
618 * if necessary, resize.
619 * called whenever we may have changed the device size.
620 * returns -ENOMEM if we could not allocate enough memory, 0 on success.
621 * In case this is actually a resize, we copy the old bitmap into the new one.
622 * Otherwise, the bitmap is initialized to all bits set.
624 int drbd_bm_resize(struct drbd_conf *mdev, sector_t capacity, int set_new_bits)
626 struct drbd_bitmap *b = mdev->bitmap;
627 unsigned long bits, words, owords, obits;
628 unsigned long want, have, onpages; /* number of pages */
629 struct page **npages, **opages = NULL;
630 int err = 0, growing;
631 int opages_vmalloced;
636 drbd_bm_lock(mdev, "resize", BM_LOCKED_MASK);
638 dev_info(DEV, "drbd_bm_resize called with capacity == %llu\n",
639 (unsigned long long)capacity);
641 if (capacity == b->bm_dev_capacity)
644 opages_vmalloced = (BM_P_VMALLOCED & b->bm_flags);
647 spin_lock_irq(&b->bm_lock);
648 opages = b->bm_pages;
649 onpages = b->bm_number_of_pages;
650 owords = b->bm_words;
652 b->bm_number_of_pages =
656 b->bm_dev_capacity = 0;
657 spin_unlock_irq(&b->bm_lock);
658 bm_free_pages(opages, onpages);
659 bm_vk_free(opages, opages_vmalloced);
662 bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT));
665 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL;
666 a 32bit host could present the wrong number of words
669 words = ALIGN(bits, 64) >> LN2_BPL;
671 if (get_ldev(mdev)) {
672 u64 bits_on_disk = ((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12;
674 if (bits > bits_on_disk) {
675 dev_info(DEV, "bits = %lu\n", bits);
676 dev_info(DEV, "bits_on_disk = %llu\n", bits_on_disk);
682 want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT;
683 have = b->bm_number_of_pages;
685 D_ASSERT(b->bm_pages != NULL);
686 npages = b->bm_pages;
688 if (drbd_insert_fault(mdev, DRBD_FAULT_BM_ALLOC))
691 npages = bm_realloc_pages(b, want);
699 spin_lock_irq(&b->bm_lock);
700 opages = b->bm_pages;
701 owords = b->bm_words;
704 growing = bits > obits;
705 if (opages && growing && set_new_bits)
708 b->bm_pages = npages;
709 b->bm_number_of_pages = want;
712 b->bm_dev_capacity = capacity;
716 bm_memset(b, owords, 0xff, words-owords);
717 b->bm_set += bits - obits;
719 bm_memset(b, owords, 0x00, words-owords);
724 /* implicit: (opages != NULL) && (opages != npages) */
725 bm_free_pages(opages + want, have - want);
728 (void)bm_clear_surplus(b);
730 spin_unlock_irq(&b->bm_lock);
731 if (opages != npages)
732 bm_vk_free(opages, opages_vmalloced);
734 b->bm_set = bm_count_bits(b);
735 dev_info(DEV, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want);
738 drbd_bm_unlock(mdev);
743 * if not protected by other means, return value may be out of date when
744 * leaving this function...
745 * we still need to lock it, since it is important that this returns
746 * bm_set == 0 precisely.
748 * maybe bm_set should be atomic_t ?
750 unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev)
752 struct drbd_bitmap *b = mdev->bitmap;
758 if (!expect(b->bm_pages))
761 spin_lock_irqsave(&b->bm_lock, flags);
763 spin_unlock_irqrestore(&b->bm_lock, flags);
768 unsigned long drbd_bm_total_weight(struct drbd_conf *mdev)
771 /* if I don't have a disk, I don't know about out-of-sync status */
772 if (!get_ldev_if_state(mdev, D_NEGOTIATING))
774 s = _drbd_bm_total_weight(mdev);
779 size_t drbd_bm_words(struct drbd_conf *mdev)
781 struct drbd_bitmap *b = mdev->bitmap;
784 if (!expect(b->bm_pages))
790 unsigned long drbd_bm_bits(struct drbd_conf *mdev)
792 struct drbd_bitmap *b = mdev->bitmap;
799 /* merge number words from buffer into the bitmap starting at offset.
800 * buffer[i] is expected to be little endian unsigned long.
801 * bitmap must be locked by drbd_bm_lock.
802 * currently only used from receive_bitmap.
804 void drbd_bm_merge_lel(struct drbd_conf *mdev, size_t offset, size_t number,
805 unsigned long *buffer)
807 struct drbd_bitmap *b = mdev->bitmap;
808 unsigned long *p_addr, *bm;
809 unsigned long word, bits;
813 end = offset + number;
817 if (!expect(b->bm_pages))
821 WARN_ON(offset >= b->bm_words);
822 WARN_ON(end > b->bm_words);
824 spin_lock_irq(&b->bm_lock);
825 while (offset < end) {
826 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset;
827 idx = bm_word_to_page_idx(b, offset);
828 p_addr = bm_map_pidx(b, idx);
829 bm = p_addr + MLPP(offset);
832 bits = hweight_long(*bm);
833 word = *bm | *buffer++;
835 b->bm_set += hweight_long(word) - bits;
838 bm_set_page_need_writeout(b->bm_pages[idx]);
840 /* with 32bit <-> 64bit cross-platform connect
841 * this is only correct for current usage,
842 * where we _know_ that we are 64 bit aligned,
843 * and know that this function is used in this way, too...
845 if (end == b->bm_words)
846 b->bm_set -= bm_clear_surplus(b);
847 spin_unlock_irq(&b->bm_lock);
850 /* copy number words from the bitmap starting at offset into the buffer.
851 * buffer[i] will be little endian unsigned long.
853 void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset, size_t number,
854 unsigned long *buffer)
856 struct drbd_bitmap *b = mdev->bitmap;
857 unsigned long *p_addr, *bm;
860 end = offset + number;
864 if (!expect(b->bm_pages))
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;
895 if (!expect(b->bm_pages))
898 spin_lock_irq(&b->bm_lock);
899 bm_memset(b, 0, 0xff, b->bm_words);
900 (void)bm_clear_surplus(b);
901 b->bm_set = b->bm_bits;
902 spin_unlock_irq(&b->bm_lock);
905 /* clear all bits in the bitmap */
906 void drbd_bm_clear_all(struct drbd_conf *mdev)
908 struct drbd_bitmap *b = mdev->bitmap;
911 if (!expect(b->bm_pages))
914 spin_lock_irq(&b->bm_lock);
915 bm_memset(b, 0, 0, b->bm_words);
917 spin_unlock_irq(&b->bm_lock);
921 struct drbd_conf *mdev;
925 #define BM_AIO_COPY_PAGES 1
926 #define BM_AIO_WRITE_HINTED 2
927 #define BM_WRITE_ALL_PAGES 4
932 static void bm_aio_ctx_destroy(struct kref *kref)
934 struct bm_aio_ctx *ctx = container_of(kref, struct bm_aio_ctx, kref);
940 /* bv_page may be a copy, or may be the original */
941 static void bm_async_io_complete(struct bio *bio, int error)
943 struct bm_aio_ctx *ctx = bio->bi_private;
944 struct drbd_conf *mdev = ctx->mdev;
945 struct drbd_bitmap *b = mdev->bitmap;
946 unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page);
947 int uptodate = bio_flagged(bio, BIO_UPTODATE);
950 /* strange behavior of some lower level drivers...
951 * fail the request by clearing the uptodate flag,
952 * but do not return any error?!
953 * do we want to WARN() on this? */
954 if (!error && !uptodate)
957 if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 &&
958 !bm_test_page_unchanged(b->bm_pages[idx]))
959 dev_warn(DEV, "bitmap page idx %u changed during IO!\n", idx);
962 /* ctx error will hold the completed-last non-zero error code,
963 * in case error codes differ. */
965 bm_set_page_io_err(b->bm_pages[idx]);
966 /* Not identical to on disk version of it.
967 * Is BM_PAGE_IO_ERROR enough? */
968 if (__ratelimit(&drbd_ratelimit_state))
969 dev_err(DEV, "IO ERROR %d on bitmap page idx %u\n",
972 bm_clear_page_io_err(b->bm_pages[idx]);
973 dynamic_dev_dbg(DEV, "bitmap page idx %u completed\n", idx);
976 bm_page_unlock_io(mdev, idx);
978 if (ctx->flags & BM_AIO_COPY_PAGES)
979 mempool_free(bio->bi_io_vec[0].bv_page, drbd_md_io_page_pool);
983 if (atomic_dec_and_test(&ctx->in_flight)) {
985 wake_up(&mdev->misc_wait);
986 kref_put(&ctx->kref, &bm_aio_ctx_destroy);
990 static void bm_page_io_async(struct bm_aio_ctx *ctx, int page_nr, int rw) __must_hold(local)
992 struct bio *bio = bio_alloc_drbd(GFP_NOIO);
993 struct drbd_conf *mdev = ctx->mdev;
994 struct drbd_bitmap *b = mdev->bitmap;
998 sector_t on_disk_sector =
999 mdev->ldev->md.md_offset + mdev->ldev->md.bm_offset;
1000 on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9);
1002 /* this might happen with very small
1003 * flexible external meta data device,
1004 * or with PAGE_SIZE > 4k */
1005 len = min_t(unsigned int, PAGE_SIZE,
1006 (drbd_md_last_sector(mdev->ldev) - on_disk_sector + 1)<<9);
1008 /* serialize IO on this page */
1009 bm_page_lock_io(mdev, page_nr);
1010 /* before memcpy and submit,
1011 * so it can be redirtied any time */
1012 bm_set_page_unchanged(b->bm_pages[page_nr]);
1014 if (ctx->flags & BM_AIO_COPY_PAGES) {
1016 page = mempool_alloc(drbd_md_io_page_pool, __GFP_HIGHMEM|__GFP_WAIT);
1017 dest = kmap_atomic(page, KM_USER0);
1018 src = kmap_atomic(b->bm_pages[page_nr], KM_USER1);
1019 memcpy(dest, src, PAGE_SIZE);
1020 kunmap_atomic(src, KM_USER1);
1021 kunmap_atomic(dest, KM_USER0);
1022 bm_store_page_idx(page, page_nr);
1024 page = b->bm_pages[page_nr];
1026 bio->bi_bdev = mdev->ldev->md_bdev;
1027 bio->bi_sector = on_disk_sector;
1028 /* bio_add_page of a single page to an empty bio will always succeed,
1029 * according to api. Do we want to assert that? */
1030 bio_add_page(bio, page, len, 0);
1031 bio->bi_private = ctx;
1032 bio->bi_end_io = bm_async_io_complete;
1034 if (drbd_insert_fault(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) {
1036 bio_endio(bio, -EIO);
1038 submit_bio(rw, bio);
1039 /* this should not count as user activity and cause the
1040 * resync to throttle -- see drbd_rs_should_slow_down(). */
1041 atomic_add(len >> 9, &mdev->rs_sect_ev);
1046 * bm_rw: read/write the whole bitmap from/to its on disk location.
1048 static int bm_rw(struct drbd_conf *mdev, int rw, unsigned flags, unsigned lazy_writeout_upper_idx) __must_hold(local)
1050 struct bm_aio_ctx *ctx;
1051 struct drbd_bitmap *b = mdev->bitmap;
1052 int num_pages, i, count = 0;
1058 * We are protected against bitmap disappearing/resizing by holding an
1059 * ldev reference (caller must have called get_ldev()).
1060 * For read/write, we are protected against changes to the bitmap by
1061 * the bitmap lock (see drbd_bitmap_io).
1062 * For lazy writeout, we don't care for ongoing changes to the bitmap,
1063 * as we submit copies of pages anyways.
1066 ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO);
1070 *ctx = (struct bm_aio_ctx) {
1072 .in_flight = ATOMIC_INIT(1),
1076 .kref = { ATOMIC_INIT(2) },
1079 if (!get_ldev_if_state(mdev, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */
1080 dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n");
1086 WARN_ON(!(BM_LOCKED_MASK & b->bm_flags));
1088 num_pages = b->bm_number_of_pages;
1092 /* let the layers below us try to merge these bios... */
1093 for (i = 0; i < num_pages; i++) {
1094 /* ignore completely unchanged pages */
1095 if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx)
1098 if ((flags & BM_AIO_WRITE_HINTED) &&
1099 !test_and_clear_bit(BM_PAGE_HINT_WRITEOUT,
1100 &page_private(b->bm_pages[i])))
1103 if (!(flags & BM_WRITE_ALL_PAGES) &&
1104 bm_test_page_unchanged(b->bm_pages[i])) {
1105 dynamic_dev_dbg(DEV, "skipped bm write for idx %u\n", i);
1108 /* during lazy writeout,
1109 * ignore those pages not marked for lazy writeout. */
1110 if (lazy_writeout_upper_idx &&
1111 !bm_test_page_lazy_writeout(b->bm_pages[i])) {
1112 dynamic_dev_dbg(DEV, "skipped bm lazy write for idx %u\n", i);
1116 atomic_inc(&ctx->in_flight);
1117 bm_page_io_async(ctx, i, rw);
1123 * We initialize ctx->in_flight to one to make sure bm_async_io_complete
1124 * will not set ctx->done early, and decrement / test it here. If there
1125 * are still some bios in flight, we need to wait for them here.
1126 * If all IO is done already (or nothing had been submitted), there is
1127 * no need to wait. Still, we need to put the kref associated with the
1128 * "in_flight reached zero, all done" event.
1130 if (!atomic_dec_and_test(&ctx->in_flight))
1131 wait_until_done_or_disk_failure(mdev, mdev->ldev, &ctx->done);
1133 kref_put(&ctx->kref, &bm_aio_ctx_destroy);
1135 /* summary for global bitmap IO */
1137 dev_info(DEV, "bitmap %s of %u pages took %lu jiffies\n",
1138 rw == WRITE ? "WRITE" : "READ",
1139 count, jiffies - now);
1142 dev_alert(DEV, "we had at least one MD IO ERROR during bitmap IO\n");
1143 drbd_chk_io_error(mdev, 1, DRBD_META_IO_ERROR);
1144 err = -EIO; /* ctx->error ? */
1147 if (atomic_read(&ctx->in_flight))
1148 err = -EIO; /* Disk failed during IO... */
1152 drbd_md_flush(mdev);
1153 } else /* rw == READ */ {
1154 b->bm_set = bm_count_bits(b);
1155 dev_info(DEV, "recounting of set bits took additional %lu jiffies\n",
1161 dev_info(DEV, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n",
1162 ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now);
1164 kref_put(&ctx->kref, &bm_aio_ctx_destroy);
1169 * drbd_bm_read() - Read the whole bitmap from its on disk location.
1170 * @mdev: DRBD device.
1172 int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local)
1174 return bm_rw(mdev, READ, 0, 0);
1178 * drbd_bm_write() - Write the whole bitmap to its on disk location.
1179 * @mdev: DRBD device.
1181 * Will only write pages that have changed since last IO.
1183 int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local)
1185 return bm_rw(mdev, WRITE, 0, 0);
1189 * drbd_bm_write_all() - Write the whole bitmap to its on disk location.
1190 * @mdev: DRBD device.
1192 * Will write all pages.
1194 int drbd_bm_write_all(struct drbd_conf *mdev) __must_hold(local)
1196 return bm_rw(mdev, WRITE, BM_WRITE_ALL_PAGES, 0);
1200 * drbd_bm_lazy_write_out() - Write bitmap pages 0 to @upper_idx-1, if they have changed.
1201 * @mdev: DRBD device.
1202 * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages
1204 int drbd_bm_write_lazy(struct drbd_conf *mdev, unsigned upper_idx) __must_hold(local)
1206 return bm_rw(mdev, WRITE, BM_AIO_COPY_PAGES, upper_idx);
1210 * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location.
1211 * @mdev: DRBD device.
1213 * Will only write pages that have changed since last IO.
1214 * In contrast to drbd_bm_write(), this will copy the bitmap pages
1215 * to temporary writeout pages. It is intended to trigger a full write-out
1216 * while still allowing the bitmap to change, for example if a resync or online
1217 * verify is aborted due to a failed peer disk, while local IO continues, or
1218 * pending resync acks are still being processed.
1220 int drbd_bm_write_copy_pages(struct drbd_conf *mdev) __must_hold(local)
1222 return bm_rw(mdev, WRITE, BM_AIO_COPY_PAGES, 0);
1226 * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed.
1227 * @mdev: DRBD device.
1229 int drbd_bm_write_hinted(struct drbd_conf *mdev) __must_hold(local)
1231 return bm_rw(mdev, WRITE, BM_AIO_WRITE_HINTED | BM_AIO_COPY_PAGES, 0);
1235 * drbd_bm_write_page() - Writes a PAGE_SIZE aligned piece of bitmap
1236 * @mdev: DRBD device.
1237 * @idx: bitmap page index
1239 * We don't want to special case on logical_block_size of the backend device,
1240 * so we submit PAGE_SIZE aligned pieces.
1241 * Note that on "most" systems, PAGE_SIZE is 4k.
1243 * In case this becomes an issue on systems with larger PAGE_SIZE,
1244 * we may want to change this again to write 4k aligned 4k pieces.
1246 int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local)
1248 struct bm_aio_ctx *ctx;
1251 if (bm_test_page_unchanged(mdev->bitmap->bm_pages[idx])) {
1252 dynamic_dev_dbg(DEV, "skipped bm page write for idx %u\n", idx);
1256 ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO);
1260 *ctx = (struct bm_aio_ctx) {
1262 .in_flight = ATOMIC_INIT(1),
1264 .flags = BM_AIO_COPY_PAGES,
1266 .kref = { ATOMIC_INIT(2) },
1269 if (!get_ldev_if_state(mdev, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */
1270 dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in drbd_bm_write_page()\n");
1275 bm_page_io_async(ctx, idx, WRITE_SYNC);
1276 wait_until_done_or_disk_failure(mdev, mdev->ldev, &ctx->done);
1279 drbd_chk_io_error(mdev, 1, DRBD_META_IO_ERROR);
1280 /* that should force detach, so the in memory bitmap will be
1281 * gone in a moment as well. */
1283 mdev->bm_writ_cnt++;
1284 err = atomic_read(&ctx->in_flight) ? -EIO : ctx->error;
1285 kref_put(&ctx->kref, &bm_aio_ctx_destroy);
1290 * find_first_bit returns int, we return unsigned long.
1291 * For this to work on 32bit arch with bitnumbers > (1<<32),
1292 * we'd need to return u64, and get a whole lot of other places
1293 * fixed where we still use unsigned long.
1295 * this returns a bit number, NOT a sector!
1297 static unsigned long __bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo,
1298 const int find_zero_bit, const enum km_type km)
1300 struct drbd_bitmap *b = mdev->bitmap;
1301 unsigned long *p_addr;
1302 unsigned long bit_offset;
1306 if (bm_fo > b->bm_bits) {
1307 dev_err(DEV, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits);
1308 bm_fo = DRBD_END_OF_BITMAP;
1310 while (bm_fo < b->bm_bits) {
1311 /* bit offset of the first bit in the page */
1312 bit_offset = bm_fo & ~BITS_PER_PAGE_MASK;
1313 p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo), km);
1316 i = find_next_zero_bit_le(p_addr,
1317 PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK);
1319 i = find_next_bit_le(p_addr,
1320 PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK);
1322 __bm_unmap(p_addr, km);
1323 if (i < PAGE_SIZE*8) {
1324 bm_fo = bit_offset + i;
1325 if (bm_fo >= b->bm_bits)
1329 bm_fo = bit_offset + PAGE_SIZE*8;
1331 bm_fo = DRBD_END_OF_BITMAP;
1337 static unsigned long bm_find_next(struct drbd_conf *mdev,
1338 unsigned long bm_fo, const int find_zero_bit)
1340 struct drbd_bitmap *b = mdev->bitmap;
1341 unsigned long i = DRBD_END_OF_BITMAP;
1345 if (!expect(b->bm_pages))
1348 spin_lock_irq(&b->bm_lock);
1349 if (BM_DONT_TEST & b->bm_flags)
1350 bm_print_lock_info(mdev);
1352 i = __bm_find_next(mdev, bm_fo, find_zero_bit, KM_IRQ1);
1354 spin_unlock_irq(&b->bm_lock);
1358 unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo)
1360 return bm_find_next(mdev, bm_fo, 0);
1364 /* not yet needed for anything. */
1365 unsigned long drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo)
1367 return bm_find_next(mdev, bm_fo, 1);
1371 /* does not spin_lock_irqsave.
1372 * you must take drbd_bm_lock() first */
1373 unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo)
1375 /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */
1376 return __bm_find_next(mdev, bm_fo, 0, KM_USER1);
1379 unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo)
1381 /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */
1382 return __bm_find_next(mdev, bm_fo, 1, KM_USER1);
1385 /* returns number of bits actually changed.
1386 * for val != 0, we change 0 -> 1, return code positive
1387 * for val == 0, we change 1 -> 0, return code negative
1388 * wants bitnr, not sector.
1389 * expected to be called for only a few bits (e - s about BITS_PER_LONG).
1390 * Must hold bitmap lock already. */
1391 static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
1392 unsigned long e, int val)
1394 struct drbd_bitmap *b = mdev->bitmap;
1395 unsigned long *p_addr = NULL;
1396 unsigned long bitnr;
1397 unsigned int last_page_nr = -1U;
1399 int changed_total = 0;
1401 if (e >= b->bm_bits) {
1402 dev_err(DEV, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n",
1404 e = b->bm_bits ? b->bm_bits -1 : 0;
1406 for (bitnr = s; bitnr <= e; bitnr++) {
1407 unsigned int page_nr = bm_bit_to_page_idx(b, bitnr);
1408 if (page_nr != last_page_nr) {
1410 __bm_unmap(p_addr, KM_IRQ1);
1412 bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]);
1414 bm_set_page_need_writeout(b->bm_pages[last_page_nr]);
1417 p_addr = __bm_map_pidx(b, page_nr, KM_IRQ1);
1418 last_page_nr = page_nr;
1421 c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr));
1423 c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr));
1426 __bm_unmap(p_addr, KM_IRQ1);
1428 bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]);
1430 bm_set_page_need_writeout(b->bm_pages[last_page_nr]);
1432 b->bm_set += changed_total;
1433 return changed_total;
1436 /* returns number of bits actually changed.
1437 * for val != 0, we change 0 -> 1, return code positive
1438 * for val == 0, we change 1 -> 0, return code negative
1439 * wants bitnr, not sector */
1440 static int bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
1441 const unsigned long e, int val)
1443 unsigned long flags;
1444 struct drbd_bitmap *b = mdev->bitmap;
1449 if (!expect(b->bm_pages))
1452 spin_lock_irqsave(&b->bm_lock, flags);
1453 if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags)
1454 bm_print_lock_info(mdev);
1456 c = __bm_change_bits_to(mdev, s, e, val);
1458 spin_unlock_irqrestore(&b->bm_lock, flags);
1462 /* returns number of bits changed 0 -> 1 */
1463 int drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1465 return bm_change_bits_to(mdev, s, e, 1);
1468 /* returns number of bits changed 1 -> 0 */
1469 int drbd_bm_clear_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1471 return -bm_change_bits_to(mdev, s, e, 0);
1474 /* sets all bits in full words,
1475 * from first_word up to, but not including, last_word */
1476 static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b,
1477 int page_nr, int first_word, int last_word)
1482 unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr], KM_IRQ1);
1483 for (i = first_word; i < last_word; i++) {
1484 bits = hweight_long(paddr[i]);
1486 changed += BITS_PER_LONG - bits;
1488 kunmap_atomic(paddr, KM_IRQ1);
1490 /* We only need lazy writeout, the information is still in the
1491 * remote bitmap as well, and is reconstructed during the next
1492 * bitmap exchange, if lost locally due to a crash. */
1493 bm_set_page_lazy_writeout(b->bm_pages[page_nr]);
1494 b->bm_set += changed;
1498 /* Same thing as drbd_bm_set_bits,
1499 * but more efficient for a large bit range.
1500 * You must first drbd_bm_lock().
1501 * Can be called to set the whole bitmap in one go.
1502 * Sets bits from s to e _inclusive_. */
1503 void _drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1505 /* First set_bit from the first bit (s)
1506 * up to the next long boundary (sl),
1507 * then assign full words up to the last long boundary (el),
1508 * then set_bit up to and including the last bit (e).
1510 * Do not use memset, because we must account for changes,
1511 * so we need to loop over the words with hweight() anyways.
1513 struct drbd_bitmap *b = mdev->bitmap;
1514 unsigned long sl = ALIGN(s,BITS_PER_LONG);
1515 unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1);
1522 if (e - s <= 3*BITS_PER_LONG) {
1523 /* don't bother; el and sl may even be wrong. */
1524 spin_lock_irq(&b->bm_lock);
1525 __bm_change_bits_to(mdev, s, e, 1);
1526 spin_unlock_irq(&b->bm_lock);
1530 /* difference is large enough that we can trust sl and el */
1532 spin_lock_irq(&b->bm_lock);
1534 /* bits filling the current long */
1536 __bm_change_bits_to(mdev, s, sl-1, 1);
1538 first_page = sl >> (3 + PAGE_SHIFT);
1539 last_page = el >> (3 + PAGE_SHIFT);
1541 /* MLPP: modulo longs per page */
1542 /* LWPP: long words per page */
1543 first_word = MLPP(sl >> LN2_BPL);
1546 /* first and full pages, unless first page == last page */
1547 for (page_nr = first_page; page_nr < last_page; page_nr++) {
1548 bm_set_full_words_within_one_page(mdev->bitmap, page_nr, first_word, last_word);
1549 spin_unlock_irq(&b->bm_lock);
1552 spin_lock_irq(&b->bm_lock);
1554 /* last page (respectively only page, for first page == last page) */
1555 last_word = MLPP(el >> LN2_BPL);
1557 /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples).
1558 * ==> e = 32767, el = 32768, last_page = 2,
1559 * and now last_word = 0.
1560 * We do not want to touch last_page in this case,
1561 * as we did not allocate it, it is not present in bitmap->bm_pages.
1564 bm_set_full_words_within_one_page(mdev->bitmap, last_page, first_word, last_word);
1566 /* possibly trailing bits.
1567 * example: (e & 63) == 63, el will be e+1.
1568 * if that even was the very last bit,
1569 * it would trigger an assert in __bm_change_bits_to()
1572 __bm_change_bits_to(mdev, el, e, 1);
1573 spin_unlock_irq(&b->bm_lock);
1576 /* returns bit state
1577 * wants bitnr, NOT sector.
1578 * inherently racy... area needs to be locked by means of {al,rs}_lru
1581 * -1 ... first out of bounds access, stop testing for bits!
1583 int drbd_bm_test_bit(struct drbd_conf *mdev, const unsigned long bitnr)
1585 unsigned long flags;
1586 struct drbd_bitmap *b = mdev->bitmap;
1587 unsigned long *p_addr;
1592 if (!expect(b->bm_pages))
1595 spin_lock_irqsave(&b->bm_lock, flags);
1596 if (BM_DONT_TEST & b->bm_flags)
1597 bm_print_lock_info(mdev);
1598 if (bitnr < b->bm_bits) {
1599 p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr));
1600 i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0;
1602 } else if (bitnr == b->bm_bits) {
1604 } else { /* (bitnr > b->bm_bits) */
1605 dev_err(DEV, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits);
1609 spin_unlock_irqrestore(&b->bm_lock, flags);
1613 /* returns number of bits set in the range [s, e] */
1614 int drbd_bm_count_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1616 unsigned long flags;
1617 struct drbd_bitmap *b = mdev->bitmap;
1618 unsigned long *p_addr = NULL;
1619 unsigned long bitnr;
1620 unsigned int page_nr = -1U;
1623 /* If this is called without a bitmap, that is a bug. But just to be
1624 * robust in case we screwed up elsewhere, in that case pretend there
1625 * was one dirty bit in the requested area, so we won't try to do a
1626 * local read there (no bitmap probably implies no disk) */
1629 if (!expect(b->bm_pages))
1632 spin_lock_irqsave(&b->bm_lock, flags);
1633 if (BM_DONT_TEST & b->bm_flags)
1634 bm_print_lock_info(mdev);
1635 for (bitnr = s; bitnr <= e; bitnr++) {
1636 unsigned int idx = bm_bit_to_page_idx(b, bitnr);
1637 if (page_nr != idx) {
1641 p_addr = bm_map_pidx(b, idx);
1643 if (expect(bitnr < b->bm_bits))
1644 c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr));
1646 dev_err(DEV, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits);
1650 spin_unlock_irqrestore(&b->bm_lock, flags);
1655 /* inherently racy...
1656 * return value may be already out-of-date when this function returns.
1657 * but the general usage is that this is only use during a cstate when bits are
1658 * only cleared, not set, and typically only care for the case when the return
1659 * value is zero, or we already "locked" this "bitmap extent" by other means.
1661 * enr is bm-extent number, since we chose to name one sector (512 bytes)
1662 * worth of the bitmap a "bitmap extent".
1665 * I think since we use it like a reference count, we should use the real
1666 * reference count of some bitmap extent element from some lru instead...
1669 int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr)
1671 struct drbd_bitmap *b = mdev->bitmap;
1673 unsigned long flags;
1674 unsigned long *p_addr, *bm;
1678 if (!expect(b->bm_pages))
1681 spin_lock_irqsave(&b->bm_lock, flags);
1682 if (BM_DONT_TEST & b->bm_flags)
1683 bm_print_lock_info(mdev);
1686 e = min((size_t)S2W(enr+1), b->bm_words);
1688 if (s < b->bm_words) {
1690 p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s));
1691 bm = p_addr + MLPP(s);
1693 count += hweight_long(*bm++);
1696 dev_err(DEV, "start offset (%d) too large in drbd_bm_e_weight\n", s);
1698 spin_unlock_irqrestore(&b->bm_lock, flags);