blob: 3030201c69d89e7407230bc8fb2d48dd1b2fbdf3 [file] [log] [blame]
Philipp Reisnerb411b362009-09-25 16:07:19 -07001/*
2 drbd_bitmap.c
3
4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
5
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>.
9
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)
13 any later version.
14
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.
19
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.
23 */
24
25#include <linux/bitops.h>
26#include <linux/vmalloc.h>
27#include <linux/string.h>
28#include <linux/drbd.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090029#include <linux/slab.h>
Philipp Reisnerb411b362009-09-25 16:07:19 -070030#include <asm/kmap_types.h>
Stephen Rothwellf0ff1352011-03-17 15:02:51 +010031
Philipp Reisnerb411b362009-09-25 16:07:19 -070032#include "drbd_int.h"
33
Lars Ellenberg95a0f102010-12-15 08:59:09 +010034
Philipp Reisnerb411b362009-09-25 16:07:19 -070035/* OPAQUE outside this file!
36 * interface defined in drbd_int.h
37
38 * convention:
39 * function name drbd_bm_... => used elsewhere, "public".
40 * function name bm_... => internal to implementation, "private".
Lars Ellenberg4b0715f2010-12-14 15:13:04 +010041 */
Philipp Reisnerb411b362009-09-25 16:07:19 -070042
Lars Ellenberg4b0715f2010-12-14 15:13:04 +010043
44/*
45 * LIMITATIONS:
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.
61 * 1 << (35 - 12)
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.
65 *
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.
69 *
70
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
75 * seems excessive.
76 *
Bart Van Assche24c48302011-05-21 18:32:29 +020077 * We plan to reduce the amount of in-core bitmap pages by paging them in
Lars Ellenberg4b0715f2010-12-14 15:13:04 +010078 * 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.
Philipp Reisnerb411b362009-09-25 16:07:19 -070081 */
82
83/*
84 * NOTE
85 * Access to the *bm_pages is protected by bm_lock.
86 * It is safe to read the other members within the lock.
87 *
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.
92 */
93struct drbd_bitmap {
94 struct page **bm_pages;
95 spinlock_t bm_lock;
Lars Ellenberg4b0715f2010-12-14 15:13:04 +010096
97 /* see LIMITATIONS: above */
98
Philipp Reisnerb411b362009-09-25 16:07:19 -070099 unsigned long bm_set; /* nr of set bits; THINK maybe atomic_t? */
100 unsigned long bm_bits;
101 size_t bm_words;
102 size_t bm_number_of_pages;
103 sector_t bm_dev_capacity;
Thomas Gleixner8a03ae22010-01-29 20:39:07 +0000104 struct mutex bm_change; /* serializes resize operations */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700105
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100106 wait_queue_head_t bm_io_wait; /* used to serialize IO of single pages */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700107
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100108 enum bm_flag bm_flags;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700109
110 /* debugging aid, in case we are still racy somewhere */
111 char *bm_why;
112 struct task_struct *bm_task;
113};
114
Philipp Reisnerb411b362009-09-25 16:07:19 -0700115#define bm_print_lock_info(m) __bm_print_lock_info(m, __func__)
116static void __bm_print_lock_info(struct drbd_conf *mdev, const char *func)
117{
118 struct drbd_bitmap *b = mdev->bitmap;
119 if (!__ratelimit(&drbd_ratelimit_state))
120 return;
121 dev_err(DEV, "FIXME %s in %s, bitmap locked for '%s' by %s\n",
122 current == mdev->receiver.task ? "receiver" :
123 current == mdev->asender.task ? "asender" :
124 current == mdev->worker.task ? "worker" : current->comm,
125 func, b->bm_why ?: "?",
126 b->bm_task == mdev->receiver.task ? "receiver" :
127 b->bm_task == mdev->asender.task ? "asender" :
128 b->bm_task == mdev->worker.task ? "worker" : "?");
129}
130
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100131void drbd_bm_lock(struct drbd_conf *mdev, char *why, enum bm_flag flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700132{
133 struct drbd_bitmap *b = mdev->bitmap;
134 int trylock_failed;
135
136 if (!b) {
137 dev_err(DEV, "FIXME no bitmap in drbd_bm_lock!?\n");
138 return;
139 }
140
Thomas Gleixner8a03ae22010-01-29 20:39:07 +0000141 trylock_failed = !mutex_trylock(&b->bm_change);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700142
143 if (trylock_failed) {
144 dev_warn(DEV, "%s going to '%s' but bitmap already locked for '%s' by %s\n",
145 current == mdev->receiver.task ? "receiver" :
146 current == mdev->asender.task ? "asender" :
147 current == mdev->worker.task ? "worker" : current->comm,
148 why, b->bm_why ?: "?",
149 b->bm_task == mdev->receiver.task ? "receiver" :
150 b->bm_task == mdev->asender.task ? "asender" :
151 b->bm_task == mdev->worker.task ? "worker" : "?");
Thomas Gleixner8a03ae22010-01-29 20:39:07 +0000152 mutex_lock(&b->bm_change);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700153 }
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100154 if (BM_LOCKED_MASK & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700155 dev_err(DEV, "FIXME bitmap already locked in bm_lock\n");
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100156 b->bm_flags |= flags & BM_LOCKED_MASK;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700157
158 b->bm_why = why;
159 b->bm_task = current;
160}
161
162void drbd_bm_unlock(struct drbd_conf *mdev)
163{
164 struct drbd_bitmap *b = mdev->bitmap;
165 if (!b) {
166 dev_err(DEV, "FIXME no bitmap in drbd_bm_unlock!?\n");
167 return;
168 }
169
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100170 if (!(BM_LOCKED_MASK & mdev->bitmap->bm_flags))
Philipp Reisnerb411b362009-09-25 16:07:19 -0700171 dev_err(DEV, "FIXME bitmap not locked in bm_unlock\n");
172
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100173 b->bm_flags &= ~BM_LOCKED_MASK;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700174 b->bm_why = NULL;
175 b->bm_task = NULL;
Thomas Gleixner8a03ae22010-01-29 20:39:07 +0000176 mutex_unlock(&b->bm_change);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700177}
178
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100179/* we store some "meta" info about our pages in page->private */
180/* at a granularity of 4k storage per bitmap bit:
181 * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks
182 * 1<<38 bits,
183 * 1<<23 4k bitmap pages.
184 * Use 24 bits as page index, covers 2 peta byte storage
185 * at a granularity of 4k per bit.
186 * Used to report the failed page idx on io error from the endio handlers.
187 */
188#define BM_PAGE_IDX_MASK ((1UL<<24)-1)
189/* this page is currently read in, or written back */
190#define BM_PAGE_IO_LOCK 31
191/* if there has been an IO error for this page */
192#define BM_PAGE_IO_ERROR 30
193/* this is to be able to intelligently skip disk IO,
194 * set if bits have been set since last IO. */
195#define BM_PAGE_NEED_WRITEOUT 29
196/* to mark for lazy writeout once syncer cleared all clearable bits,
197 * we if bits have been cleared since last IO. */
198#define BM_PAGE_LAZY_WRITEOUT 28
199
Bart Van Assche24c48302011-05-21 18:32:29 +0200200/* store_page_idx uses non-atomic assignment. It is only used directly after
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100201 * allocating the page. All other bm_set_page_* and bm_clear_page_* need to
202 * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap
203 * changes) may happen from various contexts, and wait_on_bit/wake_up_bit
204 * requires it all to be atomic as well. */
205static void bm_store_page_idx(struct page *page, unsigned long idx)
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100206{
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100207 BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK));
208 page_private(page) |= idx;
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100209}
210
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100211static unsigned long bm_page_to_idx(struct page *page)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700212{
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100213 return page_private(page) & BM_PAGE_IDX_MASK;
214}
Philipp Reisnerb411b362009-09-25 16:07:19 -0700215
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100216/* As is very unlikely that the same page is under IO from more than one
217 * context, we can get away with a bit per page and one wait queue per bitmap.
218 */
219static void bm_page_lock_io(struct drbd_conf *mdev, int page_nr)
220{
221 struct drbd_bitmap *b = mdev->bitmap;
222 void *addr = &page_private(b->bm_pages[page_nr]);
223 wait_event(b->bm_io_wait, !test_and_set_bit(BM_PAGE_IO_LOCK, addr));
224}
225
226static void bm_page_unlock_io(struct drbd_conf *mdev, int page_nr)
227{
228 struct drbd_bitmap *b = mdev->bitmap;
229 void *addr = &page_private(b->bm_pages[page_nr]);
230 clear_bit(BM_PAGE_IO_LOCK, addr);
231 smp_mb__after_clear_bit();
232 wake_up(&mdev->bitmap->bm_io_wait);
233}
234
235/* set _before_ submit_io, so it may be reset due to being changed
236 * while this page is in flight... will get submitted later again */
237static void bm_set_page_unchanged(struct page *page)
238{
239 /* use cmpxchg? */
240 clear_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page));
241 clear_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page));
242}
243
244static void bm_set_page_need_writeout(struct page *page)
245{
246 set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page));
247}
248
249static int bm_test_page_unchanged(struct page *page)
250{
251 volatile const unsigned long *addr = &page_private(page);
252 return (*addr & ((1UL<<BM_PAGE_NEED_WRITEOUT)|(1UL<<BM_PAGE_LAZY_WRITEOUT))) == 0;
253}
254
255static void bm_set_page_io_err(struct page *page)
256{
257 set_bit(BM_PAGE_IO_ERROR, &page_private(page));
258}
259
260static void bm_clear_page_io_err(struct page *page)
261{
262 clear_bit(BM_PAGE_IO_ERROR, &page_private(page));
263}
264
265static void bm_set_page_lazy_writeout(struct page *page)
266{
267 set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page));
268}
269
270static int bm_test_page_lazy_writeout(struct page *page)
271{
272 return test_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page));
273}
274
275/* on a 32bit box, this would allow for exactly (2<<38) bits. */
276static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr)
277{
Philipp Reisnerb411b362009-09-25 16:07:19 -0700278 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100279 unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700280 BUG_ON(page_nr >= b->bm_number_of_pages);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100281 return page_nr;
282}
Philipp Reisnerb411b362009-09-25 16:07:19 -0700283
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100284static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr)
285{
286 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */
287 unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3);
288 BUG_ON(page_nr >= b->bm_number_of_pages);
289 return page_nr;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700290}
291
Cong Wang589973a2011-11-25 23:51:58 +0800292static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx)
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100293{
294 struct page *page = b->bm_pages[idx];
Cong Wangcfd80052011-11-25 23:14:18 +0800295 return (unsigned long *) kmap_atomic(page);
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100296}
297
298static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx)
299{
Cong Wangcfd80052011-11-25 23:14:18 +0800300 return __bm_map_pidx(b, idx);
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100301}
302
Cong Wangcfd80052011-11-25 23:14:18 +0800303static void __bm_unmap(unsigned long *p_addr)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700304{
Cong Wangcfd80052011-11-25 23:14:18 +0800305 kunmap_atomic(p_addr);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700306};
307
308static void bm_unmap(unsigned long *p_addr)
309{
Cong Wangcfd80052011-11-25 23:14:18 +0800310 return __bm_unmap(p_addr);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700311}
312
313/* long word offset of _bitmap_ sector */
314#define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
315/* word offset from start of bitmap to word number _in_page_
316 * modulo longs per page
317#define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long))
Bart Van Assche24c48302011-05-21 18:32:29 +0200318 hm, well, Philipp thinks gcc might not optimize the % into & (... - 1)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700319 so do it explicitly:
320 */
321#define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1))
322
323/* Long words per page */
324#define LWPP (PAGE_SIZE/sizeof(long))
325
326/*
327 * actually most functions herein should take a struct drbd_bitmap*, not a
328 * struct drbd_conf*, but for the debug macros I like to have the mdev around
329 * to be able to report device specific.
330 */
331
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100332
Philipp Reisnerb411b362009-09-25 16:07:19 -0700333static void bm_free_pages(struct page **pages, unsigned long number)
334{
335 unsigned long i;
336 if (!pages)
337 return;
338
339 for (i = 0; i < number; i++) {
340 if (!pages[i]) {
341 printk(KERN_ALERT "drbd: bm_free_pages tried to free "
342 "a NULL pointer; i=%lu n=%lu\n",
343 i, number);
344 continue;
345 }
346 __free_page(pages[i]);
347 pages[i] = NULL;
348 }
349}
350
351static void bm_vk_free(void *ptr, int v)
352{
353 if (v)
354 vfree(ptr);
355 else
356 kfree(ptr);
357}
358
359/*
360 * "have" and "want" are NUMBER OF PAGES.
361 */
362static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want)
363{
364 struct page **old_pages = b->bm_pages;
365 struct page **new_pages, *page;
366 unsigned int i, bytes, vmalloced = 0;
367 unsigned long have = b->bm_number_of_pages;
368
369 BUG_ON(have == 0 && old_pages != NULL);
370 BUG_ON(have != 0 && old_pages == NULL);
371
372 if (have == want)
373 return old_pages;
374
375 /* Trying kmalloc first, falling back to vmalloc.
376 * GFP_KERNEL is ok, as this is done when a lower level disk is
377 * "attached" to the drbd. Context is receiver thread or cqueue
378 * thread. As we have no disk yet, we are not in the IO path,
379 * not even the IO path of the peer. */
380 bytes = sizeof(struct page *)*want;
Joe Perches57f32242011-05-28 10:36:24 -0700381 new_pages = kzalloc(bytes, GFP_KERNEL);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700382 if (!new_pages) {
Joe Perches57f32242011-05-28 10:36:24 -0700383 new_pages = vzalloc(bytes);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700384 if (!new_pages)
385 return NULL;
386 vmalloced = 1;
387 }
388
Philipp Reisnerb411b362009-09-25 16:07:19 -0700389 if (want >= have) {
390 for (i = 0; i < have; i++)
391 new_pages[i] = old_pages[i];
392 for (; i < want; i++) {
393 page = alloc_page(GFP_HIGHUSER);
394 if (!page) {
395 bm_free_pages(new_pages + have, i - have);
396 bm_vk_free(new_pages, vmalloced);
397 return NULL;
398 }
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100399 /* we want to know which page it is
400 * from the endio handlers */
401 bm_store_page_idx(page, i);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700402 new_pages[i] = page;
403 }
404 } else {
405 for (i = 0; i < want; i++)
406 new_pages[i] = old_pages[i];
407 /* NOT HERE, we are outside the spinlock!
408 bm_free_pages(old_pages + want, have - want);
409 */
410 }
411
412 if (vmalloced)
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100413 b->bm_flags |= BM_P_VMALLOCED;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700414 else
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100415 b->bm_flags &= ~BM_P_VMALLOCED;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700416
417 return new_pages;
418}
419
420/*
421 * called on driver init only. TODO call when a device is created.
422 * allocates the drbd_bitmap, and stores it in mdev->bitmap.
423 */
424int drbd_bm_init(struct drbd_conf *mdev)
425{
426 struct drbd_bitmap *b = mdev->bitmap;
427 WARN_ON(b != NULL);
428 b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL);
429 if (!b)
430 return -ENOMEM;
431 spin_lock_init(&b->bm_lock);
Thomas Gleixner8a03ae22010-01-29 20:39:07 +0000432 mutex_init(&b->bm_change);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700433 init_waitqueue_head(&b->bm_io_wait);
434
435 mdev->bitmap = b;
436
437 return 0;
438}
439
440sector_t drbd_bm_capacity(struct drbd_conf *mdev)
441{
442 ERR_IF(!mdev->bitmap) return 0;
443 return mdev->bitmap->bm_dev_capacity;
444}
445
446/* called on driver unload. TODO: call when a device is destroyed.
447 */
448void drbd_bm_cleanup(struct drbd_conf *mdev)
449{
450 ERR_IF (!mdev->bitmap) return;
451 bm_free_pages(mdev->bitmap->bm_pages, mdev->bitmap->bm_number_of_pages);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100452 bm_vk_free(mdev->bitmap->bm_pages, (BM_P_VMALLOCED & mdev->bitmap->bm_flags));
Philipp Reisnerb411b362009-09-25 16:07:19 -0700453 kfree(mdev->bitmap);
454 mdev->bitmap = NULL;
455}
456
457/*
458 * since (b->bm_bits % BITS_PER_LONG) != 0,
459 * this masks out the remaining bits.
460 * Returns the number of bits cleared.
461 */
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100462#define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3))
463#define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1)
464#define BITS_PER_LONG_MASK (BITS_PER_LONG - 1)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700465static int bm_clear_surplus(struct drbd_bitmap *b)
466{
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100467 unsigned long mask;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700468 unsigned long *p_addr, *bm;
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100469 int tmp;
470 int cleared = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700471
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100472 /* number of bits modulo bits per page */
473 tmp = (b->bm_bits & BITS_PER_PAGE_MASK);
474 /* mask the used bits of the word containing the last bit */
475 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1;
476 /* bitmap is always stored little endian,
477 * on disk and in core memory alike */
478 mask = cpu_to_lel(mask);
479
Lars Ellenberg6850c442010-12-16 00:32:38 +0100480 p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1);
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100481 bm = p_addr + (tmp/BITS_PER_LONG);
482 if (mask) {
483 /* If mask != 0, we are not exactly aligned, so bm now points
484 * to the long containing the last bit.
485 * If mask == 0, bm already points to the word immediately
486 * after the last (long word aligned) bit. */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700487 cleared = hweight_long(*bm & ~mask);
488 *bm &= mask;
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100489 bm++;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700490 }
491
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100492 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) {
493 /* on a 32bit arch, we may need to zero out
494 * a padding long to align with a 64bit remote */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700495 cleared += hweight_long(*bm);
496 *bm = 0;
497 }
498 bm_unmap(p_addr);
499 return cleared;
500}
501
502static void bm_set_surplus(struct drbd_bitmap *b)
503{
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100504 unsigned long mask;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700505 unsigned long *p_addr, *bm;
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100506 int tmp;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700507
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100508 /* number of bits modulo bits per page */
509 tmp = (b->bm_bits & BITS_PER_PAGE_MASK);
510 /* mask the used bits of the word containing the last bit */
511 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1;
512 /* bitmap is always stored little endian,
513 * on disk and in core memory alike */
514 mask = cpu_to_lel(mask);
515
Lars Ellenberg6850c442010-12-16 00:32:38 +0100516 p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1);
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100517 bm = p_addr + (tmp/BITS_PER_LONG);
518 if (mask) {
519 /* If mask != 0, we are not exactly aligned, so bm now points
520 * to the long containing the last bit.
521 * If mask == 0, bm already points to the word immediately
522 * after the last (long word aligned) bit. */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700523 *bm |= ~mask;
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100524 bm++;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700525 }
526
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100527 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) {
528 /* on a 32bit arch, we may need to zero out
529 * a padding long to align with a 64bit remote */
530 *bm = ~0UL;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700531 }
532 bm_unmap(p_addr);
533}
534
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100535/* you better not modify the bitmap while this is running,
536 * or its results will be stale */
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100537static unsigned long bm_count_bits(struct drbd_bitmap *b)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700538{
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100539 unsigned long *p_addr;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700540 unsigned long bits = 0;
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100541 unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1;
Lars Ellenberg6850c442010-12-16 00:32:38 +0100542 int idx, i, last_word;
Lars Ellenberg7777a8b2010-12-15 23:21:39 +0100543
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100544 /* all but last page */
Lars Ellenberg6850c442010-12-16 00:32:38 +0100545 for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) {
Cong Wangcfd80052011-11-25 23:14:18 +0800546 p_addr = __bm_map_pidx(b, idx);
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100547 for (i = 0; i < LWPP; i++)
548 bits += hweight_long(p_addr[i]);
Cong Wangcfd80052011-11-25 23:14:18 +0800549 __bm_unmap(p_addr);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700550 cond_resched();
551 }
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100552 /* last (or only) page */
553 last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL;
Cong Wang589973a2011-11-25 23:51:58 +0800554 p_addr = __bm_map_pidx(b, idx);
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100555 for (i = 0; i < last_word; i++)
556 bits += hweight_long(p_addr[i]);
557 p_addr[last_word] &= cpu_to_lel(mask);
558 bits += hweight_long(p_addr[last_word]);
559 /* 32bit arch, may have an unused padding long */
560 if (BITS_PER_LONG == 32 && (last_word & 1) == 0)
561 p_addr[last_word+1] = 0;
Cong Wang589973a2011-11-25 23:51:58 +0800562 __bm_unmap(p_addr);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700563 return bits;
564}
565
Philipp Reisnerb411b362009-09-25 16:07:19 -0700566/* offset and len in long words.*/
567static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len)
568{
569 unsigned long *p_addr, *bm;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100570 unsigned int idx;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700571 size_t do_now, end;
572
Philipp Reisnerb411b362009-09-25 16:07:19 -0700573 end = offset + len;
574
575 if (end > b->bm_words) {
576 printk(KERN_ALERT "drbd: bm_memset end > bm_words\n");
577 return;
578 }
579
580 while (offset < end) {
581 do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100582 idx = bm_word_to_page_idx(b, offset);
583 p_addr = bm_map_pidx(b, idx);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700584 bm = p_addr + MLPP(offset);
585 if (bm+do_now > p_addr + LWPP) {
586 printk(KERN_ALERT "drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n",
587 p_addr, bm, (int)do_now);
Lars Ellenberg84e7c0f2010-12-16 00:37:57 +0100588 } else
589 memset(bm, c, do_now * sizeof(long));
Philipp Reisnerb411b362009-09-25 16:07:19 -0700590 bm_unmap(p_addr);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100591 bm_set_page_need_writeout(b->bm_pages[idx]);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700592 offset += do_now;
593 }
594}
595
596/*
597 * make sure the bitmap has enough room for the attached storage,
598 * if necessary, resize.
599 * called whenever we may have changed the device size.
600 * returns -ENOMEM if we could not allocate enough memory, 0 on success.
601 * In case this is actually a resize, we copy the old bitmap into the new one.
602 * Otherwise, the bitmap is initialized to all bits set.
603 */
Philipp Reisner02d9a942010-03-24 16:23:03 +0100604int drbd_bm_resize(struct drbd_conf *mdev, sector_t capacity, int set_new_bits)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700605{
606 struct drbd_bitmap *b = mdev->bitmap;
Lars Ellenberg6850c442010-12-16 00:32:38 +0100607 unsigned long bits, words, owords, obits;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700608 unsigned long want, have, onpages; /* number of pages */
609 struct page **npages, **opages = NULL;
610 int err = 0, growing;
611 int opages_vmalloced;
612
613 ERR_IF(!b) return -ENOMEM;
614
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100615 drbd_bm_lock(mdev, "resize", BM_LOCKED_MASK);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700616
617 dev_info(DEV, "drbd_bm_resize called with capacity == %llu\n",
618 (unsigned long long)capacity);
619
620 if (capacity == b->bm_dev_capacity)
621 goto out;
622
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +0100623 opages_vmalloced = (BM_P_VMALLOCED & b->bm_flags);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700624
625 if (capacity == 0) {
626 spin_lock_irq(&b->bm_lock);
627 opages = b->bm_pages;
628 onpages = b->bm_number_of_pages;
629 owords = b->bm_words;
630 b->bm_pages = NULL;
631 b->bm_number_of_pages =
632 b->bm_set =
633 b->bm_bits =
634 b->bm_words =
635 b->bm_dev_capacity = 0;
636 spin_unlock_irq(&b->bm_lock);
637 bm_free_pages(opages, onpages);
638 bm_vk_free(opages, opages_vmalloced);
639 goto out;
640 }
641 bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT));
642
643 /* if we would use
644 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL;
645 a 32bit host could present the wrong number of words
646 to a 64bit host.
647 */
648 words = ALIGN(bits, 64) >> LN2_BPL;
649
650 if (get_ldev(mdev)) {
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100651 u64 bits_on_disk = ((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700652 put_ldev(mdev);
Lars Ellenberg4b0715f2010-12-14 15:13:04 +0100653 if (bits > bits_on_disk) {
654 dev_info(DEV, "bits = %lu\n", bits);
655 dev_info(DEV, "bits_on_disk = %llu\n", bits_on_disk);
656 err = -ENOSPC;
657 goto out;
658 }
Philipp Reisnerb411b362009-09-25 16:07:19 -0700659 }
660
Lars Ellenberg6850c442010-12-16 00:32:38 +0100661 want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700662 have = b->bm_number_of_pages;
663 if (want == have) {
664 D_ASSERT(b->bm_pages != NULL);
665 npages = b->bm_pages;
666 } else {
Andreas Gruenbacher0cf9d272010-12-07 10:43:29 +0100667 if (drbd_insert_fault(mdev, DRBD_FAULT_BM_ALLOC))
Philipp Reisnerb411b362009-09-25 16:07:19 -0700668 npages = NULL;
669 else
670 npages = bm_realloc_pages(b, want);
671 }
672
673 if (!npages) {
674 err = -ENOMEM;
675 goto out;
676 }
677
678 spin_lock_irq(&b->bm_lock);
679 opages = b->bm_pages;
680 owords = b->bm_words;
681 obits = b->bm_bits;
682
683 growing = bits > obits;
Philipp Reisner52236712010-04-28 14:46:57 +0200684 if (opages && growing && set_new_bits)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700685 bm_set_surplus(b);
686
687 b->bm_pages = npages;
688 b->bm_number_of_pages = want;
689 b->bm_bits = bits;
690 b->bm_words = words;
691 b->bm_dev_capacity = capacity;
692
693 if (growing) {
Philipp Reisner02d9a942010-03-24 16:23:03 +0100694 if (set_new_bits) {
695 bm_memset(b, owords, 0xff, words-owords);
696 b->bm_set += bits - obits;
697 } else
698 bm_memset(b, owords, 0x00, words-owords);
699
Philipp Reisnerb411b362009-09-25 16:07:19 -0700700 }
701
702 if (want < have) {
703 /* implicit: (opages != NULL) && (opages != npages) */
704 bm_free_pages(opages + want, have - want);
705 }
706
Philipp Reisnerb411b362009-09-25 16:07:19 -0700707 (void)bm_clear_surplus(b);
708
709 spin_unlock_irq(&b->bm_lock);
710 if (opages != npages)
711 bm_vk_free(opages, opages_vmalloced);
712 if (!growing)
713 b->bm_set = bm_count_bits(b);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100714 dev_info(DEV, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700715
716 out:
717 drbd_bm_unlock(mdev);
718 return err;
719}
720
721/* inherently racy:
722 * if not protected by other means, return value may be out of date when
723 * leaving this function...
724 * we still need to lock it, since it is important that this returns
725 * bm_set == 0 precisely.
726 *
727 * maybe bm_set should be atomic_t ?
728 */
Philipp Reisner07782862010-08-31 12:00:50 +0200729unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700730{
731 struct drbd_bitmap *b = mdev->bitmap;
732 unsigned long s;
733 unsigned long flags;
734
735 ERR_IF(!b) return 0;
736 ERR_IF(!b->bm_pages) return 0;
737
738 spin_lock_irqsave(&b->bm_lock, flags);
739 s = b->bm_set;
740 spin_unlock_irqrestore(&b->bm_lock, flags);
741
742 return s;
743}
744
745unsigned long drbd_bm_total_weight(struct drbd_conf *mdev)
746{
747 unsigned long s;
748 /* if I don't have a disk, I don't know about out-of-sync status */
749 if (!get_ldev_if_state(mdev, D_NEGOTIATING))
750 return 0;
751 s = _drbd_bm_total_weight(mdev);
752 put_ldev(mdev);
753 return s;
754}
755
756size_t drbd_bm_words(struct drbd_conf *mdev)
757{
758 struct drbd_bitmap *b = mdev->bitmap;
759 ERR_IF(!b) return 0;
760 ERR_IF(!b->bm_pages) return 0;
761
762 return b->bm_words;
763}
764
765unsigned long drbd_bm_bits(struct drbd_conf *mdev)
766{
767 struct drbd_bitmap *b = mdev->bitmap;
768 ERR_IF(!b) return 0;
769
770 return b->bm_bits;
771}
772
773/* merge number words from buffer into the bitmap starting at offset.
774 * buffer[i] is expected to be little endian unsigned long.
775 * bitmap must be locked by drbd_bm_lock.
776 * currently only used from receive_bitmap.
777 */
778void drbd_bm_merge_lel(struct drbd_conf *mdev, size_t offset, size_t number,
779 unsigned long *buffer)
780{
781 struct drbd_bitmap *b = mdev->bitmap;
782 unsigned long *p_addr, *bm;
783 unsigned long word, bits;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100784 unsigned int idx;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700785 size_t end, do_now;
786
787 end = offset + number;
788
789 ERR_IF(!b) return;
790 ERR_IF(!b->bm_pages) return;
791 if (number == 0)
792 return;
793 WARN_ON(offset >= b->bm_words);
794 WARN_ON(end > b->bm_words);
795
796 spin_lock_irq(&b->bm_lock);
797 while (offset < end) {
798 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100799 idx = bm_word_to_page_idx(b, offset);
800 p_addr = bm_map_pidx(b, idx);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700801 bm = p_addr + MLPP(offset);
802 offset += do_now;
803 while (do_now--) {
804 bits = hweight_long(*bm);
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100805 word = *bm | *buffer++;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700806 *bm++ = word;
807 b->bm_set += hweight_long(word) - bits;
808 }
809 bm_unmap(p_addr);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100810 bm_set_page_need_writeout(b->bm_pages[idx]);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700811 }
812 /* with 32bit <-> 64bit cross-platform connect
813 * this is only correct for current usage,
814 * where we _know_ that we are 64 bit aligned,
815 * and know that this function is used in this way, too...
816 */
817 if (end == b->bm_words)
818 b->bm_set -= bm_clear_surplus(b);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700819 spin_unlock_irq(&b->bm_lock);
820}
821
822/* copy number words from the bitmap starting at offset into the buffer.
823 * buffer[i] will be little endian unsigned long.
824 */
825void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset, size_t number,
826 unsigned long *buffer)
827{
828 struct drbd_bitmap *b = mdev->bitmap;
829 unsigned long *p_addr, *bm;
830 size_t end, do_now;
831
832 end = offset + number;
833
834 ERR_IF(!b) return;
835 ERR_IF(!b->bm_pages) return;
836
837 spin_lock_irq(&b->bm_lock);
838 if ((offset >= b->bm_words) ||
839 (end > b->bm_words) ||
840 (number <= 0))
841 dev_err(DEV, "offset=%lu number=%lu bm_words=%lu\n",
842 (unsigned long) offset,
843 (unsigned long) number,
844 (unsigned long) b->bm_words);
845 else {
846 while (offset < end) {
847 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100848 p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset));
Philipp Reisnerb411b362009-09-25 16:07:19 -0700849 bm = p_addr + MLPP(offset);
850 offset += do_now;
851 while (do_now--)
Lars Ellenberg95a0f102010-12-15 08:59:09 +0100852 *buffer++ = *bm++;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700853 bm_unmap(p_addr);
854 }
855 }
856 spin_unlock_irq(&b->bm_lock);
857}
858
859/* set all bits in the bitmap */
860void drbd_bm_set_all(struct drbd_conf *mdev)
861{
862 struct drbd_bitmap *b = mdev->bitmap;
863 ERR_IF(!b) return;
864 ERR_IF(!b->bm_pages) return;
865
866 spin_lock_irq(&b->bm_lock);
867 bm_memset(b, 0, 0xff, b->bm_words);
868 (void)bm_clear_surplus(b);
869 b->bm_set = b->bm_bits;
870 spin_unlock_irq(&b->bm_lock);
871}
872
873/* clear all bits in the bitmap */
874void drbd_bm_clear_all(struct drbd_conf *mdev)
875{
876 struct drbd_bitmap *b = mdev->bitmap;
877 ERR_IF(!b) return;
878 ERR_IF(!b->bm_pages) return;
879
880 spin_lock_irq(&b->bm_lock);
881 bm_memset(b, 0, 0, b->bm_words);
882 b->bm_set = 0;
883 spin_unlock_irq(&b->bm_lock);
884}
885
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100886struct bm_aio_ctx {
887 struct drbd_conf *mdev;
888 atomic_t in_flight;
Lars Ellenberg725a97e2010-12-19 11:29:55 +0100889 struct completion done;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100890 unsigned flags;
891#define BM_AIO_COPY_PAGES 1
892 int error;
893};
894
895/* bv_page may be a copy, or may be the original */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700896static void bm_async_io_complete(struct bio *bio, int error)
897{
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100898 struct bm_aio_ctx *ctx = bio->bi_private;
899 struct drbd_conf *mdev = ctx->mdev;
900 struct drbd_bitmap *b = mdev->bitmap;
901 unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700902 int uptodate = bio_flagged(bio, BIO_UPTODATE);
903
904
905 /* strange behavior of some lower level drivers...
906 * fail the request by clearing the uptodate flag,
907 * but do not return any error?!
908 * do we want to WARN() on this? */
909 if (!error && !uptodate)
910 error = -EIO;
911
Lars Ellenberg7648cdf2010-12-17 23:58:41 +0100912 if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 &&
913 !bm_test_page_unchanged(b->bm_pages[idx]))
914 dev_warn(DEV, "bitmap page idx %u changed during IO!\n", idx);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100915
Philipp Reisnerb411b362009-09-25 16:07:19 -0700916 if (error) {
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100917 /* ctx error will hold the completed-last non-zero error code,
918 * in case error codes differ. */
919 ctx->error = error;
920 bm_set_page_io_err(b->bm_pages[idx]);
921 /* Not identical to on disk version of it.
922 * Is BM_PAGE_IO_ERROR enough? */
923 if (__ratelimit(&drbd_ratelimit_state))
924 dev_err(DEV, "IO ERROR %d on bitmap page idx %u\n",
925 error, idx);
926 } else {
927 bm_clear_page_io_err(b->bm_pages[idx]);
928 dynamic_dev_dbg(DEV, "bitmap page idx %u completed\n", idx);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700929 }
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100930
931 bm_page_unlock_io(mdev, idx);
932
933 /* FIXME give back to page pool */
934 if (ctx->flags & BM_AIO_COPY_PAGES)
935 put_page(bio->bi_io_vec[0].bv_page);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700936
937 bio_put(bio);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100938
939 if (atomic_dec_and_test(&ctx->in_flight))
Lars Ellenberg725a97e2010-12-19 11:29:55 +0100940 complete(&ctx->done);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700941}
942
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100943static void bm_page_io_async(struct bm_aio_ctx *ctx, int page_nr, int rw) __must_hold(local)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700944{
945 /* we are process context. we always get a bio */
946 struct bio *bio = bio_alloc(GFP_KERNEL, 1);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100947 struct drbd_conf *mdev = ctx->mdev;
948 struct drbd_bitmap *b = mdev->bitmap;
949 struct page *page;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700950 unsigned int len;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100951
Philipp Reisnerb411b362009-09-25 16:07:19 -0700952 sector_t on_disk_sector =
953 mdev->ldev->md.md_offset + mdev->ldev->md.bm_offset;
954 on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9);
955
956 /* this might happen with very small
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100957 * flexible external meta data device,
958 * or with PAGE_SIZE > 4k */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700959 len = min_t(unsigned int, PAGE_SIZE,
960 (drbd_md_last_sector(mdev->ldev) - on_disk_sector + 1)<<9);
961
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100962 /* serialize IO on this page */
963 bm_page_lock_io(mdev, page_nr);
964 /* before memcpy and submit,
965 * so it can be redirtied any time */
966 bm_set_page_unchanged(b->bm_pages[page_nr]);
967
968 if (ctx->flags & BM_AIO_COPY_PAGES) {
969 /* FIXME alloc_page is good enough for now, but actually needs
970 * to use pre-allocated page pool */
971 void *src, *dest;
972 page = alloc_page(__GFP_HIGHMEM|__GFP_WAIT);
Cong Wang589973a2011-11-25 23:51:58 +0800973 dest = kmap_atomic(page);
974 src = kmap_atomic(b->bm_pages[page_nr]);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100975 memcpy(dest, src, PAGE_SIZE);
Cong Wang589973a2011-11-25 23:51:58 +0800976 kunmap_atomic(src);
977 kunmap_atomic(dest);
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100978 bm_store_page_idx(page, page_nr);
979 } else
980 page = b->bm_pages[page_nr];
981
Philipp Reisnerb411b362009-09-25 16:07:19 -0700982 bio->bi_bdev = mdev->ldev->md_bdev;
983 bio->bi_sector = on_disk_sector;
Lars Ellenberg19f843a2010-12-15 08:59:11 +0100984 bio_add_page(bio, page, len, 0);
985 bio->bi_private = ctx;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700986 bio->bi_end_io = bm_async_io_complete;
987
Andreas Gruenbacher0cf9d272010-12-07 10:43:29 +0100988 if (drbd_insert_fault(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) {
Philipp Reisnerb411b362009-09-25 16:07:19 -0700989 bio->bi_rw |= rw;
990 bio_endio(bio, -EIO);
991 } else {
992 submit_bio(rw, bio);
Lars Ellenberg5a8b4242011-06-14 14:18:23 +0200993 /* this should not count as user activity and cause the
994 * resync to throttle -- see drbd_rs_should_slow_down(). */
995 atomic_add(len >> 9, &mdev->rs_sect_ev);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700996 }
997}
998
Philipp Reisnerb411b362009-09-25 16:07:19 -0700999/*
1000 * bm_rw: read/write the whole bitmap from/to its on disk location.
1001 */
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001002static int bm_rw(struct drbd_conf *mdev, int rw, unsigned lazy_writeout_upper_idx) __must_hold(local)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001003{
Lars Ellenberg725a97e2010-12-19 11:29:55 +01001004 struct bm_aio_ctx ctx = {
1005 .mdev = mdev,
1006 .in_flight = ATOMIC_INIT(1),
1007 .done = COMPLETION_INITIALIZER_ONSTACK(ctx.done),
1008 .flags = lazy_writeout_upper_idx ? BM_AIO_COPY_PAGES : 0,
1009 };
Philipp Reisnerb411b362009-09-25 16:07:19 -07001010 struct drbd_bitmap *b = mdev->bitmap;
Lars Ellenberg6850c442010-12-16 00:32:38 +01001011 int num_pages, i, count = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001012 unsigned long now;
1013 char ppb[10];
1014 int err = 0;
1015
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001016 /*
1017 * We are protected against bitmap disappearing/resizing by holding an
1018 * ldev reference (caller must have called get_ldev()).
1019 * For read/write, we are protected against changes to the bitmap by
1020 * the bitmap lock (see drbd_bitmap_io).
1021 * For lazy writeout, we don't care for ongoing changes to the bitmap,
1022 * as we submit copies of pages anyways.
1023 */
1024 if (!ctx.flags)
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001025 WARN_ON(!(BM_LOCKED_MASK & b->bm_flags));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001026
Lars Ellenberg6850c442010-12-16 00:32:38 +01001027 num_pages = b->bm_number_of_pages;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001028
Philipp Reisnerb411b362009-09-25 16:07:19 -07001029 now = jiffies;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001030
1031 /* let the layers below us try to merge these bios... */
Lars Ellenberg6850c442010-12-16 00:32:38 +01001032 for (i = 0; i < num_pages; i++) {
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001033 /* ignore completely unchanged pages */
1034 if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx)
1035 break;
1036 if (rw & WRITE) {
1037 if (bm_test_page_unchanged(b->bm_pages[i])) {
1038 dynamic_dev_dbg(DEV, "skipped bm write for idx %u\n", i);
1039 continue;
1040 }
1041 /* during lazy writeout,
1042 * ignore those pages not marked for lazy writeout. */
1043 if (lazy_writeout_upper_idx &&
1044 !bm_test_page_lazy_writeout(b->bm_pages[i])) {
1045 dynamic_dev_dbg(DEV, "skipped bm lazy write for idx %u\n", i);
1046 continue;
1047 }
1048 }
1049 atomic_inc(&ctx.in_flight);
1050 bm_page_io_async(&ctx, i, rw);
1051 ++count;
1052 cond_resched();
1053 }
Philipp Reisnerb411b362009-09-25 16:07:19 -07001054
Lars Ellenberg725a97e2010-12-19 11:29:55 +01001055 /*
1056 * We initialize ctx.in_flight to one to make sure bm_async_io_complete
1057 * will not complete() early, and decrement / test it here. If there
1058 * are still some bios in flight, we need to wait for them here.
1059 */
1060 if (!atomic_dec_and_test(&ctx.in_flight))
1061 wait_for_completion(&ctx.done);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001062 dev_info(DEV, "bitmap %s of %u pages took %lu jiffies\n",
1063 rw == WRITE ? "WRITE" : "READ",
1064 count, jiffies - now);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001065
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001066 if (ctx.error) {
Philipp Reisnerb411b362009-09-25 16:07:19 -07001067 dev_alert(DEV, "we had at least one MD IO ERROR during bitmap IO\n");
Andreas Gruenbacher81e84652010-12-09 15:03:57 +01001068 drbd_chk_io_error(mdev, 1, true);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001069 err = -EIO; /* ctx.error ? */
Philipp Reisnerb411b362009-09-25 16:07:19 -07001070 }
1071
1072 now = jiffies;
1073 if (rw == WRITE) {
Philipp Reisnerb411b362009-09-25 16:07:19 -07001074 drbd_md_flush(mdev);
1075 } else /* rw == READ */ {
Lars Ellenberg95a0f102010-12-15 08:59:09 +01001076 b->bm_set = bm_count_bits(b);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001077 dev_info(DEV, "recounting of set bits took additional %lu jiffies\n",
1078 jiffies - now);
1079 }
1080 now = b->bm_set;
1081
1082 dev_info(DEV, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n",
1083 ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now);
1084
1085 return err;
1086}
1087
1088/**
1089 * drbd_bm_read() - Read the whole bitmap from its on disk location.
1090 * @mdev: DRBD device.
1091 */
1092int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local)
1093{
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001094 return bm_rw(mdev, READ, 0);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001095}
1096
1097/**
1098 * drbd_bm_write() - Write the whole bitmap to its on disk location.
1099 * @mdev: DRBD device.
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001100 *
1101 * Will only write pages that have changed since last IO.
Philipp Reisnerb411b362009-09-25 16:07:19 -07001102 */
1103int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local)
1104{
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001105 return bm_rw(mdev, WRITE, 0);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001106}
1107
1108/**
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001109 * drbd_bm_lazy_write_out() - Write bitmap pages 0 to @upper_idx-1, if they have changed.
Philipp Reisnerb411b362009-09-25 16:07:19 -07001110 * @mdev: DRBD device.
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001111 * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages
Philipp Reisnerb411b362009-09-25 16:07:19 -07001112 */
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001113int drbd_bm_write_lazy(struct drbd_conf *mdev, unsigned upper_idx) __must_hold(local)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001114{
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001115 return bm_rw(mdev, WRITE, upper_idx);
1116}
Philipp Reisnerb411b362009-09-25 16:07:19 -07001117
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001118
1119/**
1120 * drbd_bm_write_page: Writes a PAGE_SIZE aligned piece of bitmap
1121 * @mdev: DRBD device.
1122 * @idx: bitmap page index
1123 *
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001124 * We don't want to special case on logical_block_size of the backend device,
1125 * so we submit PAGE_SIZE aligned pieces.
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001126 * Note that on "most" systems, PAGE_SIZE is 4k.
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001127 *
1128 * In case this becomes an issue on systems with larger PAGE_SIZE,
1129 * we may want to change this again to write 4k aligned 4k pieces.
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001130 */
1131int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local)
1132{
Lars Ellenberg725a97e2010-12-19 11:29:55 +01001133 struct bm_aio_ctx ctx = {
1134 .mdev = mdev,
1135 .in_flight = ATOMIC_INIT(1),
1136 .done = COMPLETION_INITIALIZER_ONSTACK(ctx.done),
1137 .flags = BM_AIO_COPY_PAGES,
1138 };
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001139
1140 if (bm_test_page_unchanged(mdev->bitmap->bm_pages[idx])) {
Lars Ellenberg7648cdf2010-12-17 23:58:41 +01001141 dynamic_dev_dbg(DEV, "skipped bm page write for idx %u\n", idx);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001142 return 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001143 }
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001144
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001145 bm_page_io_async(&ctx, idx, WRITE_SYNC);
Lars Ellenberg725a97e2010-12-19 11:29:55 +01001146 wait_for_completion(&ctx.done);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001147
1148 if (ctx.error)
1149 drbd_chk_io_error(mdev, 1, true);
1150 /* that should force detach, so the in memory bitmap will be
1151 * gone in a moment as well. */
1152
Philipp Reisnerb411b362009-09-25 16:07:19 -07001153 mdev->bm_writ_cnt++;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001154 return ctx.error;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001155}
1156
1157/* NOTE
1158 * find_first_bit returns int, we return unsigned long.
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001159 * For this to work on 32bit arch with bitnumbers > (1<<32),
1160 * we'd need to return u64, and get a whole lot of other places
1161 * fixed where we still use unsigned long.
Philipp Reisnerb411b362009-09-25 16:07:19 -07001162 *
1163 * this returns a bit number, NOT a sector!
1164 */
Philipp Reisnerb411b362009-09-25 16:07:19 -07001165static unsigned long __bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo,
Cong Wangcfd80052011-11-25 23:14:18 +08001166 const int find_zero_bit)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001167{
1168 struct drbd_bitmap *b = mdev->bitmap;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001169 unsigned long *p_addr;
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001170 unsigned long bit_offset;
1171 unsigned i;
1172
Philipp Reisnerb411b362009-09-25 16:07:19 -07001173
1174 if (bm_fo > b->bm_bits) {
1175 dev_err(DEV, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits);
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001176 bm_fo = DRBD_END_OF_BITMAP;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001177 } else {
1178 while (bm_fo < b->bm_bits) {
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001179 /* bit offset of the first bit in the page */
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001180 bit_offset = bm_fo & ~BITS_PER_PAGE_MASK;
Cong Wangcfd80052011-11-25 23:14:18 +08001181 p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001182
1183 if (find_zero_bit)
Linus Torvalds7e599e62011-03-28 07:42:58 -07001184 i = find_next_zero_bit_le(p_addr,
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001185 PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001186 else
Linus Torvalds7e599e62011-03-28 07:42:58 -07001187 i = find_next_bit_le(p_addr,
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001188 PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001189
Cong Wangcfd80052011-11-25 23:14:18 +08001190 __bm_unmap(p_addr);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001191 if (i < PAGE_SIZE*8) {
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001192 bm_fo = bit_offset + i;
1193 if (bm_fo >= b->bm_bits)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001194 break;
1195 goto found;
1196 }
1197 bm_fo = bit_offset + PAGE_SIZE*8;
1198 }
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001199 bm_fo = DRBD_END_OF_BITMAP;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001200 }
1201 found:
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001202 return bm_fo;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001203}
1204
1205static unsigned long bm_find_next(struct drbd_conf *mdev,
1206 unsigned long bm_fo, const int find_zero_bit)
1207{
1208 struct drbd_bitmap *b = mdev->bitmap;
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001209 unsigned long i = DRBD_END_OF_BITMAP;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001210
1211 ERR_IF(!b) return i;
1212 ERR_IF(!b->bm_pages) return i;
1213
1214 spin_lock_irq(&b->bm_lock);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001215 if (BM_DONT_TEST & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001216 bm_print_lock_info(mdev);
1217
Cong Wangcfd80052011-11-25 23:14:18 +08001218 i = __bm_find_next(mdev, bm_fo, find_zero_bit);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001219
1220 spin_unlock_irq(&b->bm_lock);
1221 return i;
1222}
1223
1224unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo)
1225{
1226 return bm_find_next(mdev, bm_fo, 0);
1227}
1228
1229#if 0
1230/* not yet needed for anything. */
1231unsigned long drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo)
1232{
1233 return bm_find_next(mdev, bm_fo, 1);
1234}
1235#endif
1236
1237/* does not spin_lock_irqsave.
1238 * you must take drbd_bm_lock() first */
1239unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo)
1240{
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001241 /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */
Cong Wangcfd80052011-11-25 23:14:18 +08001242 return __bm_find_next(mdev, bm_fo, 0);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001243}
1244
1245unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo)
1246{
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001247 /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */
Cong Wangcfd80052011-11-25 23:14:18 +08001248 return __bm_find_next(mdev, bm_fo, 1);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001249}
1250
1251/* returns number of bits actually changed.
1252 * for val != 0, we change 0 -> 1, return code positive
1253 * for val == 0, we change 1 -> 0, return code negative
1254 * wants bitnr, not sector.
1255 * expected to be called for only a few bits (e - s about BITS_PER_LONG).
1256 * Must hold bitmap lock already. */
Philipp Reisnerb4ee79d2010-04-01 09:57:40 +02001257static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
Lars Ellenberg829c6082011-06-03 21:18:13 +02001258 unsigned long e, int val)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001259{
1260 struct drbd_bitmap *b = mdev->bitmap;
1261 unsigned long *p_addr = NULL;
1262 unsigned long bitnr;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001263 unsigned int last_page_nr = -1U;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001264 int c = 0;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001265 int changed_total = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001266
1267 if (e >= b->bm_bits) {
1268 dev_err(DEV, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n",
1269 s, e, b->bm_bits);
1270 e = b->bm_bits ? b->bm_bits -1 : 0;
1271 }
1272 for (bitnr = s; bitnr <= e; bitnr++) {
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001273 unsigned int page_nr = bm_bit_to_page_idx(b, bitnr);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001274 if (page_nr != last_page_nr) {
1275 if (p_addr)
Cong Wangcfd80052011-11-25 23:14:18 +08001276 __bm_unmap(p_addr);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001277 if (c < 0)
1278 bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]);
1279 else if (c > 0)
1280 bm_set_page_need_writeout(b->bm_pages[last_page_nr]);
1281 changed_total += c;
1282 c = 0;
Cong Wangcfd80052011-11-25 23:14:18 +08001283 p_addr = __bm_map_pidx(b, page_nr);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001284 last_page_nr = page_nr;
1285 }
1286 if (val)
Linus Torvalds7e599e62011-03-28 07:42:58 -07001287 c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001288 else
Linus Torvalds7e599e62011-03-28 07:42:58 -07001289 c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001290 }
1291 if (p_addr)
Cong Wangcfd80052011-11-25 23:14:18 +08001292 __bm_unmap(p_addr);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001293 if (c < 0)
1294 bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]);
1295 else if (c > 0)
1296 bm_set_page_need_writeout(b->bm_pages[last_page_nr]);
1297 changed_total += c;
1298 b->bm_set += changed_total;
1299 return changed_total;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001300}
1301
1302/* returns number of bits actually changed.
1303 * for val != 0, we change 0 -> 1, return code positive
1304 * for val == 0, we change 1 -> 0, return code negative
1305 * wants bitnr, not sector */
Philipp Reisnerb4ee79d2010-04-01 09:57:40 +02001306static int bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
Philipp Reisnerb411b362009-09-25 16:07:19 -07001307 const unsigned long e, int val)
1308{
1309 unsigned long flags;
1310 struct drbd_bitmap *b = mdev->bitmap;
1311 int c = 0;
1312
1313 ERR_IF(!b) return 1;
1314 ERR_IF(!b->bm_pages) return 0;
1315
1316 spin_lock_irqsave(&b->bm_lock, flags);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001317 if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001318 bm_print_lock_info(mdev);
1319
Lars Ellenberg829c6082011-06-03 21:18:13 +02001320 c = __bm_change_bits_to(mdev, s, e, val);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001321
1322 spin_unlock_irqrestore(&b->bm_lock, flags);
1323 return c;
1324}
1325
1326/* returns number of bits changed 0 -> 1 */
1327int drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1328{
1329 return bm_change_bits_to(mdev, s, e, 1);
1330}
1331
1332/* returns number of bits changed 1 -> 0 */
1333int drbd_bm_clear_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1334{
1335 return -bm_change_bits_to(mdev, s, e, 0);
1336}
1337
1338/* sets all bits in full words,
1339 * from first_word up to, but not including, last_word */
1340static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b,
1341 int page_nr, int first_word, int last_word)
1342{
1343 int i;
1344 int bits;
Cong Wangcfd80052011-11-25 23:14:18 +08001345 unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr]);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001346 for (i = first_word; i < last_word; i++) {
1347 bits = hweight_long(paddr[i]);
1348 paddr[i] = ~0UL;
1349 b->bm_set += BITS_PER_LONG - bits;
1350 }
Cong Wangcfd80052011-11-25 23:14:18 +08001351 kunmap_atomic(paddr);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001352}
1353
Lars Ellenberg829c6082011-06-03 21:18:13 +02001354/* Same thing as drbd_bm_set_bits,
1355 * but more efficient for a large bit range.
Philipp Reisnerb411b362009-09-25 16:07:19 -07001356 * You must first drbd_bm_lock().
1357 * Can be called to set the whole bitmap in one go.
1358 * Sets bits from s to e _inclusive_. */
1359void _drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1360{
1361 /* First set_bit from the first bit (s)
1362 * up to the next long boundary (sl),
1363 * then assign full words up to the last long boundary (el),
1364 * then set_bit up to and including the last bit (e).
1365 *
1366 * Do not use memset, because we must account for changes,
1367 * so we need to loop over the words with hweight() anyways.
1368 */
Lars Ellenberg829c6082011-06-03 21:18:13 +02001369 struct drbd_bitmap *b = mdev->bitmap;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001370 unsigned long sl = ALIGN(s,BITS_PER_LONG);
1371 unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1);
1372 int first_page;
1373 int last_page;
1374 int page_nr;
1375 int first_word;
1376 int last_word;
1377
1378 if (e - s <= 3*BITS_PER_LONG) {
1379 /* don't bother; el and sl may even be wrong. */
Lars Ellenberg829c6082011-06-03 21:18:13 +02001380 spin_lock_irq(&b->bm_lock);
1381 __bm_change_bits_to(mdev, s, e, 1);
1382 spin_unlock_irq(&b->bm_lock);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001383 return;
1384 }
1385
1386 /* difference is large enough that we can trust sl and el */
1387
Lars Ellenberg829c6082011-06-03 21:18:13 +02001388 spin_lock_irq(&b->bm_lock);
1389
Philipp Reisnerb411b362009-09-25 16:07:19 -07001390 /* bits filling the current long */
1391 if (sl)
Lars Ellenberg829c6082011-06-03 21:18:13 +02001392 __bm_change_bits_to(mdev, s, sl-1, 1);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001393
1394 first_page = sl >> (3 + PAGE_SHIFT);
1395 last_page = el >> (3 + PAGE_SHIFT);
1396
1397 /* MLPP: modulo longs per page */
1398 /* LWPP: long words per page */
1399 first_word = MLPP(sl >> LN2_BPL);
1400 last_word = LWPP;
1401
1402 /* first and full pages, unless first page == last page */
1403 for (page_nr = first_page; page_nr < last_page; page_nr++) {
1404 bm_set_full_words_within_one_page(mdev->bitmap, page_nr, first_word, last_word);
Lars Ellenberg8ccee202011-06-06 11:31:42 +02001405 spin_unlock_irq(&b->bm_lock);
1406 cond_resched();
Philipp Reisnerb411b362009-09-25 16:07:19 -07001407 first_word = 0;
Lars Ellenberg8ccee202011-06-06 11:31:42 +02001408 spin_lock_irq(&b->bm_lock);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001409 }
1410
1411 /* last page (respectively only page, for first page == last page) */
1412 last_word = MLPP(el >> LN2_BPL);
1413 bm_set_full_words_within_one_page(mdev->bitmap, last_page, first_word, last_word);
1414
1415 /* possibly trailing bits.
1416 * example: (e & 63) == 63, el will be e+1.
1417 * if that even was the very last bit,
1418 * it would trigger an assert in __bm_change_bits_to()
1419 */
1420 if (el <= e)
Lars Ellenberg829c6082011-06-03 21:18:13 +02001421 __bm_change_bits_to(mdev, el, e, 1);
1422 spin_unlock_irq(&b->bm_lock);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001423}
1424
1425/* returns bit state
1426 * wants bitnr, NOT sector.
1427 * inherently racy... area needs to be locked by means of {al,rs}_lru
1428 * 1 ... bit set
1429 * 0 ... bit not set
1430 * -1 ... first out of bounds access, stop testing for bits!
1431 */
1432int drbd_bm_test_bit(struct drbd_conf *mdev, const unsigned long bitnr)
1433{
1434 unsigned long flags;
1435 struct drbd_bitmap *b = mdev->bitmap;
1436 unsigned long *p_addr;
1437 int i;
1438
1439 ERR_IF(!b) return 0;
1440 ERR_IF(!b->bm_pages) return 0;
1441
1442 spin_lock_irqsave(&b->bm_lock, flags);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001443 if (BM_DONT_TEST & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001444 bm_print_lock_info(mdev);
1445 if (bitnr < b->bm_bits) {
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001446 p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr));
Linus Torvalds7e599e62011-03-28 07:42:58 -07001447 i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001448 bm_unmap(p_addr);
1449 } else if (bitnr == b->bm_bits) {
1450 i = -1;
1451 } else { /* (bitnr > b->bm_bits) */
1452 dev_err(DEV, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits);
1453 i = 0;
1454 }
1455
1456 spin_unlock_irqrestore(&b->bm_lock, flags);
1457 return i;
1458}
1459
1460/* returns number of bits set in the range [s, e] */
1461int drbd_bm_count_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e)
1462{
1463 unsigned long flags;
1464 struct drbd_bitmap *b = mdev->bitmap;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001465 unsigned long *p_addr = NULL;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001466 unsigned long bitnr;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001467 unsigned int page_nr = -1U;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001468 int c = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001469
1470 /* If this is called without a bitmap, that is a bug. But just to be
1471 * robust in case we screwed up elsewhere, in that case pretend there
1472 * was one dirty bit in the requested area, so we won't try to do a
1473 * local read there (no bitmap probably implies no disk) */
1474 ERR_IF(!b) return 1;
1475 ERR_IF(!b->bm_pages) return 1;
1476
1477 spin_lock_irqsave(&b->bm_lock, flags);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001478 if (BM_DONT_TEST & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001479 bm_print_lock_info(mdev);
1480 for (bitnr = s; bitnr <= e; bitnr++) {
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001481 unsigned int idx = bm_bit_to_page_idx(b, bitnr);
1482 if (page_nr != idx) {
1483 page_nr = idx;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001484 if (p_addr)
1485 bm_unmap(p_addr);
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001486 p_addr = bm_map_pidx(b, idx);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001487 }
1488 ERR_IF (bitnr >= b->bm_bits) {
1489 dev_err(DEV, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits);
1490 } else {
Linus Torvalds7e599e62011-03-28 07:42:58 -07001491 c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001492 }
1493 }
1494 if (p_addr)
1495 bm_unmap(p_addr);
1496 spin_unlock_irqrestore(&b->bm_lock, flags);
1497 return c;
1498}
1499
1500
1501/* inherently racy...
1502 * return value may be already out-of-date when this function returns.
1503 * but the general usage is that this is only use during a cstate when bits are
1504 * only cleared, not set, and typically only care for the case when the return
1505 * value is zero, or we already "locked" this "bitmap extent" by other means.
1506 *
1507 * enr is bm-extent number, since we chose to name one sector (512 bytes)
1508 * worth of the bitmap a "bitmap extent".
1509 *
1510 * TODO
1511 * I think since we use it like a reference count, we should use the real
1512 * reference count of some bitmap extent element from some lru instead...
1513 *
1514 */
1515int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr)
1516{
1517 struct drbd_bitmap *b = mdev->bitmap;
1518 int count, s, e;
1519 unsigned long flags;
1520 unsigned long *p_addr, *bm;
1521
1522 ERR_IF(!b) return 0;
1523 ERR_IF(!b->bm_pages) return 0;
1524
1525 spin_lock_irqsave(&b->bm_lock, flags);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001526 if (BM_DONT_TEST & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001527 bm_print_lock_info(mdev);
1528
1529 s = S2W(enr);
1530 e = min((size_t)S2W(enr+1), b->bm_words);
1531 count = 0;
1532 if (s < b->bm_words) {
1533 int n = e-s;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001534 p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001535 bm = p_addr + MLPP(s);
1536 while (n--)
1537 count += hweight_long(*bm++);
1538 bm_unmap(p_addr);
1539 } else {
1540 dev_err(DEV, "start offset (%d) too large in drbd_bm_e_weight\n", s);
1541 }
1542 spin_unlock_irqrestore(&b->bm_lock, flags);
1543 return count;
1544}
1545
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001546/* Set all bits covered by the AL-extent al_enr.
1547 * Returns number of bits changed. */
Philipp Reisnerb411b362009-09-25 16:07:19 -07001548unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev, unsigned long al_enr)
1549{
1550 struct drbd_bitmap *b = mdev->bitmap;
1551 unsigned long *p_addr, *bm;
1552 unsigned long weight;
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001553 unsigned long s, e;
1554 int count, i, do_now;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001555 ERR_IF(!b) return 0;
1556 ERR_IF(!b->bm_pages) return 0;
1557
1558 spin_lock_irq(&b->bm_lock);
Lars Ellenberg20ceb2b2011-01-21 10:56:44 +01001559 if (BM_DONT_SET & b->bm_flags)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001560 bm_print_lock_info(mdev);
1561 weight = b->bm_set;
1562
1563 s = al_enr * BM_WORDS_PER_AL_EXT;
1564 e = min_t(size_t, s + BM_WORDS_PER_AL_EXT, b->bm_words);
1565 /* assert that s and e are on the same page */
1566 D_ASSERT((e-1) >> (PAGE_SHIFT - LN2_BPL + 3)
1567 == s >> (PAGE_SHIFT - LN2_BPL + 3));
1568 count = 0;
1569 if (s < b->bm_words) {
1570 i = do_now = e-s;
Lars Ellenberg19f843a2010-12-15 08:59:11 +01001571 p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s));
Philipp Reisnerb411b362009-09-25 16:07:19 -07001572 bm = p_addr + MLPP(s);
1573 while (i--) {
1574 count += hweight_long(*bm);
1575 *bm = -1UL;
1576 bm++;
1577 }
1578 bm_unmap(p_addr);
1579 b->bm_set += do_now*BITS_PER_LONG - count;
1580 if (e == b->bm_words)
1581 b->bm_set -= bm_clear_surplus(b);
1582 } else {
Lars Ellenberg4b0715f2010-12-14 15:13:04 +01001583 dev_err(DEV, "start offset (%lu) too large in drbd_bm_ALe_set_all\n", s);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001584 }
1585 weight = b->bm_set - weight;
1586 spin_unlock_irq(&b->bm_lock);
1587 return weight;
1588}