Commit | Line | Data |
---|---|---|
b411b363 PR |
1 | /* |
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> | |
5a0e3ad6 | 29 | #include <linux/slab.h> |
b411b363 | 30 | #include <asm/kmap_types.h> |
f0ff1357 | 31 | |
b411b363 PR |
32 | #include "drbd_int.h" |
33 | ||
95a0f10c | 34 | |
b411b363 PR |
35 | /* 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". | |
4b0715f0 LE |
41 | */ |
42 | ||
b411b363 | 43 | |
4b0715f0 LE |
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 | * | |
24c4830c | 77 | * We plan to reduce the amount of in-core bitmap pages by paging them in |
4b0715f0 LE |
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. | |
b411b363 PR |
81 | */ |
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 | */ | |
93 | struct drbd_bitmap { | |
94 | struct page **bm_pages; | |
95 | spinlock_t bm_lock; | |
4b0715f0 LE |
96 | |
97 | /* see LIMITATIONS: above */ | |
98 | ||
b411b363 PR |
99 | 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; | |
8a03ae2a | 104 | struct mutex bm_change; /* serializes resize operations */ |
b411b363 | 105 | |
19f843aa | 106 | wait_queue_head_t bm_io_wait; /* used to serialize IO of single pages */ |
b411b363 | 107 | |
20ceb2b2 | 108 | enum bm_flag bm_flags; |
b411b363 PR |
109 | |
110 | /* debugging aid, in case we are still racy somewhere */ | |
111 | char *bm_why; | |
112 | struct task_struct *bm_task; | |
113 | }; | |
114 | ||
b411b363 PR |
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) | |
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 | ||
20ceb2b2 | 131 | void drbd_bm_lock(struct drbd_conf *mdev, char *why, enum bm_flag flags) |
b411b363 PR |
132 | { |
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 | ||
8a03ae2a | 141 | trylock_failed = !mutex_trylock(&b->bm_change); |
b411b363 PR |
142 | |
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" : "?"); | |
8a03ae2a | 152 | mutex_lock(&b->bm_change); |
b411b363 | 153 | } |
20ceb2b2 | 154 | if (BM_LOCKED_MASK & b->bm_flags) |
b411b363 | 155 | dev_err(DEV, "FIXME bitmap already locked in bm_lock\n"); |
20ceb2b2 | 156 | b->bm_flags |= flags & BM_LOCKED_MASK; |
b411b363 PR |
157 | |
158 | b->bm_why = why; | |
159 | b->bm_task = current; | |
160 | } | |
161 | ||
162 | void 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 | ||
20ceb2b2 | 170 | if (!(BM_LOCKED_MASK & mdev->bitmap->bm_flags)) |
b411b363 PR |
171 | dev_err(DEV, "FIXME bitmap not locked in bm_unlock\n"); |
172 | ||
20ceb2b2 | 173 | b->bm_flags &= ~BM_LOCKED_MASK; |
b411b363 PR |
174 | b->bm_why = NULL; |
175 | b->bm_task = NULL; | |
8a03ae2a | 176 | mutex_unlock(&b->bm_change); |
b411b363 PR |
177 | } |
178 | ||
19f843aa LE |
179 | /* we store some "meta" info about our pages in page->private */ |
180 | /* at a granularity of 4k storage per bitmap bit: | |
181 | * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks | |
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 | ||
24c4830c | 200 | /* store_page_idx uses non-atomic assignment. It is only used directly after |
19f843aa LE |
201 | * allocating the page. All other bm_set_page_* and bm_clear_page_* need to |
202 | * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap | |
203 | * changes) may happen from various contexts, and wait_on_bit/wake_up_bit | |
204 | * requires it all to be atomic as well. */ | |
205 | static void bm_store_page_idx(struct page *page, unsigned long idx) | |
95a0f10c | 206 | { |
19f843aa | 207 | BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK)); |
0c7db279 | 208 | set_page_private(page, idx); |
95a0f10c LE |
209 | } |
210 | ||
19f843aa | 211 | static unsigned long bm_page_to_idx(struct page *page) |
b411b363 | 212 | { |
19f843aa LE |
213 | return page_private(page) & BM_PAGE_IDX_MASK; |
214 | } | |
215 | ||
216 | /* As is very unlikely that the same page is under IO from more than one | |
217 | * context, we can get away with a bit per page and one wait queue per bitmap. | |
218 | */ | |
219 | static 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 | ||
226 | static 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 */ | |
237 | static 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 | ||
244 | static void bm_set_page_need_writeout(struct page *page) | |
245 | { | |
246 | set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
247 | } | |
248 | ||
249 | static 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 | } | |
b411b363 | 254 | |
19f843aa LE |
255 | static void bm_set_page_io_err(struct page *page) |
256 | { | |
257 | set_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
258 | } | |
259 | ||
260 | static void bm_clear_page_io_err(struct page *page) | |
261 | { | |
262 | clear_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
263 | } | |
264 | ||
265 | static void bm_set_page_lazy_writeout(struct page *page) | |
266 | { | |
267 | set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
268 | } | |
269 | ||
270 | static 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. */ | |
276 | static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr) | |
277 | { | |
b411b363 | 278 | /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */ |
19f843aa | 279 | unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3); |
b411b363 | 280 | BUG_ON(page_nr >= b->bm_number_of_pages); |
19f843aa LE |
281 | return page_nr; |
282 | } | |
b411b363 | 283 | |
19f843aa LE |
284 | static 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; | |
b411b363 PR |
290 | } |
291 | ||
589973a7 | 292 | static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) |
95a0f10c LE |
293 | { |
294 | struct page *page = b->bm_pages[idx]; | |
cfd8005c | 295 | return (unsigned long *) kmap_atomic(page); |
95a0f10c LE |
296 | } |
297 | ||
298 | static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) | |
299 | { | |
cfd8005c | 300 | return __bm_map_pidx(b, idx); |
95a0f10c LE |
301 | } |
302 | ||
cfd8005c | 303 | static void __bm_unmap(unsigned long *p_addr) |
b411b363 | 304 | { |
cfd8005c | 305 | kunmap_atomic(p_addr); |
b411b363 PR |
306 | }; |
307 | ||
308 | static void bm_unmap(unsigned long *p_addr) | |
309 | { | |
cfd8005c | 310 | return __bm_unmap(p_addr); |
b411b363 PR |
311 | } |
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)) | |
24c4830c | 318 | hm, well, Philipp thinks gcc might not optimize the % into & (... - 1) |
b411b363 PR |
319 | 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 | ||
19f843aa | 332 | |
b411b363 PR |
333 | static 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 | ||
351 | static 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 | */ | |
362 | static 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; | |
57f3224c | 381 | new_pages = kzalloc(bytes, GFP_KERNEL); |
b411b363 | 382 | if (!new_pages) { |
57f3224c | 383 | new_pages = vzalloc(bytes); |
b411b363 PR |
384 | if (!new_pages) |
385 | return NULL; | |
386 | vmalloced = 1; | |
387 | } | |
388 | ||
b411b363 PR |
389 | 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 | } | |
19f843aa LE |
399 | /* we want to know which page it is |
400 | * from the endio handlers */ | |
401 | bm_store_page_idx(page, i); | |
b411b363 PR |
402 | 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) | |
20ceb2b2 | 413 | b->bm_flags |= BM_P_VMALLOCED; |
b411b363 | 414 | else |
20ceb2b2 | 415 | b->bm_flags &= ~BM_P_VMALLOCED; |
b411b363 PR |
416 | |
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 | */ | |
424 | int 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); | |
8a03ae2a | 432 | mutex_init(&b->bm_change); |
b411b363 PR |
433 | init_waitqueue_head(&b->bm_io_wait); |
434 | ||
435 | mdev->bitmap = b; | |
436 | ||
437 | return 0; | |
438 | } | |
439 | ||
440 | sector_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 | */ | |
448 | void 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); | |
20ceb2b2 | 452 | bm_vk_free(mdev->bitmap->bm_pages, (BM_P_VMALLOCED & mdev->bitmap->bm_flags)); |
b411b363 PR |
453 | 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 | */ | |
95a0f10c LE |
462 | #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3)) |
463 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1) | |
464 | #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1) | |
b411b363 PR |
465 | static int bm_clear_surplus(struct drbd_bitmap *b) |
466 | { | |
95a0f10c | 467 | unsigned long mask; |
b411b363 | 468 | unsigned long *p_addr, *bm; |
95a0f10c LE |
469 | int tmp; |
470 | int cleared = 0; | |
b411b363 | 471 | |
95a0f10c LE |
472 | /* number of bits modulo bits per page */ |
473 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
474 | /* mask the used bits of the word containing the last bit */ | |
475 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
476 | /* bitmap is always stored little endian, | |
477 | * on disk and in core memory alike */ | |
478 | mask = cpu_to_lel(mask); | |
479 | ||
6850c442 | 480 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
481 | 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. */ | |
b411b363 PR |
487 | cleared = hweight_long(*bm & ~mask); |
488 | *bm &= mask; | |
95a0f10c | 489 | bm++; |
b411b363 PR |
490 | } |
491 | ||
95a0f10c LE |
492 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
493 | /* on a 32bit arch, we may need to zero out | |
494 | * a padding long to align with a 64bit remote */ | |
b411b363 PR |
495 | cleared += hweight_long(*bm); |
496 | *bm = 0; | |
497 | } | |
498 | bm_unmap(p_addr); | |
499 | return cleared; | |
500 | } | |
501 | ||
502 | static void bm_set_surplus(struct drbd_bitmap *b) | |
503 | { | |
95a0f10c | 504 | unsigned long mask; |
b411b363 | 505 | unsigned long *p_addr, *bm; |
95a0f10c LE |
506 | int tmp; |
507 | ||
508 | /* number of bits modulo bits per page */ | |
509 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
510 | /* mask the used bits of the word containing the last bit */ | |
511 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
512 | /* bitmap is always stored little endian, | |
513 | * on disk and in core memory alike */ | |
514 | mask = cpu_to_lel(mask); | |
515 | ||
6850c442 | 516 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
517 | 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. */ | |
b411b363 | 523 | *bm |= ~mask; |
95a0f10c | 524 | bm++; |
b411b363 PR |
525 | } |
526 | ||
95a0f10c LE |
527 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
528 | /* on a 32bit arch, we may need to zero out | |
529 | * a padding long to align with a 64bit remote */ | |
530 | *bm = ~0UL; | |
b411b363 PR |
531 | } |
532 | bm_unmap(p_addr); | |
533 | } | |
534 | ||
4b0715f0 LE |
535 | /* you better not modify the bitmap while this is running, |
536 | * or its results will be stale */ | |
95a0f10c | 537 | static unsigned long bm_count_bits(struct drbd_bitmap *b) |
b411b363 | 538 | { |
4b0715f0 | 539 | unsigned long *p_addr; |
b411b363 | 540 | unsigned long bits = 0; |
4b0715f0 | 541 | unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1; |
6850c442 | 542 | int idx, i, last_word; |
4b0715f0 LE |
543 | |
544 | /* all but last page */ | |
6850c442 | 545 | for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) { |
cfd8005c | 546 | p_addr = __bm_map_pidx(b, idx); |
4b0715f0 LE |
547 | for (i = 0; i < LWPP; i++) |
548 | bits += hweight_long(p_addr[i]); | |
cfd8005c | 549 | __bm_unmap(p_addr); |
b411b363 PR |
550 | cond_resched(); |
551 | } | |
4b0715f0 LE |
552 | /* last (or only) page */ |
553 | last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL; | |
589973a7 | 554 | p_addr = __bm_map_pidx(b, idx); |
4b0715f0 LE |
555 | for (i = 0; i < last_word; i++) |
556 | bits += hweight_long(p_addr[i]); | |
557 | p_addr[last_word] &= cpu_to_lel(mask); | |
558 | bits += hweight_long(p_addr[last_word]); | |
559 | /* 32bit arch, may have an unused padding long */ | |
560 | if (BITS_PER_LONG == 32 && (last_word & 1) == 0) | |
561 | p_addr[last_word+1] = 0; | |
589973a7 | 562 | __bm_unmap(p_addr); |
b411b363 PR |
563 | return bits; |
564 | } | |
565 | ||
b411b363 PR |
566 | /* offset and len in long words.*/ |
567 | static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len) | |
568 | { | |
569 | unsigned long *p_addr, *bm; | |
19f843aa | 570 | unsigned int idx; |
b411b363 PR |
571 | size_t do_now, end; |
572 | ||
b411b363 PR |
573 | 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; | |
19f843aa LE |
582 | idx = bm_word_to_page_idx(b, offset); |
583 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
584 | bm = p_addr + MLPP(offset); |
585 | if (bm+do_now > p_addr + LWPP) { | |
586 | printk(KERN_ALERT "drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n", | |
587 | p_addr, bm, (int)do_now); | |
84e7c0f7 LE |
588 | } else |
589 | memset(bm, c, do_now * sizeof(long)); | |
b411b363 | 590 | bm_unmap(p_addr); |
19f843aa | 591 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
592 | 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 | */ | |
02d9a94b | 604 | int drbd_bm_resize(struct drbd_conf *mdev, sector_t capacity, int set_new_bits) |
b411b363 PR |
605 | { |
606 | struct drbd_bitmap *b = mdev->bitmap; | |
6850c442 | 607 | unsigned long bits, words, owords, obits; |
b411b363 PR |
608 | 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 | ||
20ceb2b2 | 615 | drbd_bm_lock(mdev, "resize", BM_LOCKED_MASK); |
b411b363 PR |
616 | |
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 | ||
20ceb2b2 | 623 | opages_vmalloced = (BM_P_VMALLOCED & b->bm_flags); |
b411b363 PR |
624 | |
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)) { | |
4b0715f0 | 651 | u64 bits_on_disk = ((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12; |
b411b363 | 652 | put_ldev(mdev); |
4b0715f0 LE |
653 | if (bits > bits_on_disk) { |
654 | dev_info(DEV, "bits = %lu\n", bits); | |
655 | dev_info(DEV, "bits_on_disk = %llu\n", bits_on_disk); | |
656 | err = -ENOSPC; | |
657 | goto out; | |
658 | } | |
b411b363 PR |
659 | } |
660 | ||
6850c442 | 661 | want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT; |
b411b363 PR |
662 | have = b->bm_number_of_pages; |
663 | if (want == have) { | |
664 | D_ASSERT(b->bm_pages != NULL); | |
665 | npages = b->bm_pages; | |
666 | } else { | |
0cf9d27e | 667 | if (drbd_insert_fault(mdev, DRBD_FAULT_BM_ALLOC)) |
b411b363 PR |
668 | 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; | |
5223671b | 684 | if (opages && growing && set_new_bits) |
b411b363 PR |
685 | 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) { | |
02d9a94b PR |
694 | 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 | ||
b411b363 PR |
700 | } |
701 | ||
702 | if (want < have) { | |
703 | /* implicit: (opages != NULL) && (opages != npages) */ | |
704 | bm_free_pages(opages + want, have - want); | |
705 | } | |
706 | ||
b411b363 PR |
707 | (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); | |
19f843aa | 714 | dev_info(DEV, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want); |
b411b363 PR |
715 | |
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 | */ | |
0778286a | 729 | unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev) |
b411b363 PR |
730 | { |
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 | ||
745 | unsigned 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 | ||
756 | size_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 | ||
765 | unsigned 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 | */ | |
778 | void 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; | |
19f843aa | 784 | unsigned int idx; |
b411b363 PR |
785 | 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; | |
19f843aa LE |
799 | idx = bm_word_to_page_idx(b, offset); |
800 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
801 | bm = p_addr + MLPP(offset); |
802 | offset += do_now; | |
803 | while (do_now--) { | |
804 | bits = hweight_long(*bm); | |
95a0f10c | 805 | word = *bm | *buffer++; |
b411b363 PR |
806 | *bm++ = word; |
807 | b->bm_set += hweight_long(word) - bits; | |
808 | } | |
809 | bm_unmap(p_addr); | |
19f843aa | 810 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
811 | } |
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); | |
b411b363 PR |
819 | 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 | */ | |
825 | void 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; | |
19f843aa | 848 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset)); |
b411b363 PR |
849 | bm = p_addr + MLPP(offset); |
850 | offset += do_now; | |
851 | while (do_now--) | |
95a0f10c | 852 | *buffer++ = *bm++; |
b411b363 PR |
853 | bm_unmap(p_addr); |
854 | } | |
855 | } | |
856 | spin_unlock_irq(&b->bm_lock); | |
857 | } | |
858 | ||
859 | /* set all bits in the bitmap */ | |
860 | void 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 */ | |
874 | void 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 | ||
19f843aa LE |
886 | struct bm_aio_ctx { |
887 | struct drbd_conf *mdev; | |
888 | atomic_t in_flight; | |
9e58c4da | 889 | unsigned int done; |
19f843aa LE |
890 | unsigned flags; |
891 | #define BM_AIO_COPY_PAGES 1 | |
d1aa4d04 | 892 | #define BM_WRITE_ALL_PAGES 2 |
19f843aa | 893 | int error; |
d1f3779b | 894 | struct kref kref; |
19f843aa LE |
895 | }; |
896 | ||
d1f3779b PR |
897 | static void bm_aio_ctx_destroy(struct kref *kref) |
898 | { | |
899 | struct bm_aio_ctx *ctx = container_of(kref, struct bm_aio_ctx, kref); | |
900 | ||
9e58c4da | 901 | put_ldev(ctx->mdev); |
d1f3779b PR |
902 | kfree(ctx); |
903 | } | |
904 | ||
19f843aa | 905 | /* bv_page may be a copy, or may be the original */ |
b411b363 PR |
906 | static void bm_async_io_complete(struct bio *bio, int error) |
907 | { | |
19f843aa LE |
908 | struct bm_aio_ctx *ctx = bio->bi_private; |
909 | struct drbd_conf *mdev = ctx->mdev; | |
910 | struct drbd_bitmap *b = mdev->bitmap; | |
911 | unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page); | |
b411b363 PR |
912 | int uptodate = bio_flagged(bio, BIO_UPTODATE); |
913 | ||
914 | ||
915 | /* strange behavior of some lower level drivers... | |
916 | * fail the request by clearing the uptodate flag, | |
917 | * but do not return any error?! | |
918 | * do we want to WARN() on this? */ | |
919 | if (!error && !uptodate) | |
920 | error = -EIO; | |
921 | ||
7648cdfe LE |
922 | if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 && |
923 | !bm_test_page_unchanged(b->bm_pages[idx])) | |
924 | dev_warn(DEV, "bitmap page idx %u changed during IO!\n", idx); | |
19f843aa | 925 | |
b411b363 | 926 | if (error) { |
19f843aa LE |
927 | /* ctx error will hold the completed-last non-zero error code, |
928 | * in case error codes differ. */ | |
929 | ctx->error = error; | |
930 | bm_set_page_io_err(b->bm_pages[idx]); | |
931 | /* Not identical to on disk version of it. | |
932 | * Is BM_PAGE_IO_ERROR enough? */ | |
933 | if (__ratelimit(&drbd_ratelimit_state)) | |
934 | dev_err(DEV, "IO ERROR %d on bitmap page idx %u\n", | |
935 | error, idx); | |
936 | } else { | |
937 | bm_clear_page_io_err(b->bm_pages[idx]); | |
938 | dynamic_dev_dbg(DEV, "bitmap page idx %u completed\n", idx); | |
b411b363 | 939 | } |
19f843aa LE |
940 | |
941 | bm_page_unlock_io(mdev, idx); | |
942 | ||
19f843aa | 943 | if (ctx->flags & BM_AIO_COPY_PAGES) |
4d95a10f | 944 | mempool_free(bio->bi_io_vec[0].bv_page, drbd_md_io_page_pool); |
b411b363 PR |
945 | |
946 | bio_put(bio); | |
19f843aa | 947 | |
d1f3779b | 948 | if (atomic_dec_and_test(&ctx->in_flight)) { |
9e58c4da PR |
949 | ctx->done = 1; |
950 | wake_up(&mdev->misc_wait); | |
d1f3779b PR |
951 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); |
952 | } | |
b411b363 PR |
953 | } |
954 | ||
19f843aa | 955 | static void bm_page_io_async(struct bm_aio_ctx *ctx, int page_nr, int rw) __must_hold(local) |
b411b363 | 956 | { |
9476f39d | 957 | struct bio *bio = bio_alloc_drbd(GFP_NOIO); |
19f843aa LE |
958 | struct drbd_conf *mdev = ctx->mdev; |
959 | struct drbd_bitmap *b = mdev->bitmap; | |
960 | struct page *page; | |
b411b363 | 961 | unsigned int len; |
19f843aa | 962 | |
b411b363 PR |
963 | sector_t on_disk_sector = |
964 | mdev->ldev->md.md_offset + mdev->ldev->md.bm_offset; | |
965 | on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9); | |
966 | ||
967 | /* this might happen with very small | |
19f843aa LE |
968 | * flexible external meta data device, |
969 | * or with PAGE_SIZE > 4k */ | |
b411b363 PR |
970 | len = min_t(unsigned int, PAGE_SIZE, |
971 | (drbd_md_last_sector(mdev->ldev) - on_disk_sector + 1)<<9); | |
972 | ||
19f843aa LE |
973 | /* serialize IO on this page */ |
974 | bm_page_lock_io(mdev, page_nr); | |
975 | /* before memcpy and submit, | |
976 | * so it can be redirtied any time */ | |
977 | bm_set_page_unchanged(b->bm_pages[page_nr]); | |
978 | ||
979 | if (ctx->flags & BM_AIO_COPY_PAGES) { | |
19f843aa | 980 | void *src, *dest; |
4d95a10f | 981 | page = mempool_alloc(drbd_md_io_page_pool, __GFP_HIGHMEM|__GFP_WAIT); |
589973a7 CW |
982 | dest = kmap_atomic(page); |
983 | src = kmap_atomic(b->bm_pages[page_nr]); | |
19f843aa | 984 | memcpy(dest, src, PAGE_SIZE); |
589973a7 CW |
985 | kunmap_atomic(src); |
986 | kunmap_atomic(dest); | |
19f843aa LE |
987 | bm_store_page_idx(page, page_nr); |
988 | } else | |
989 | page = b->bm_pages[page_nr]; | |
990 | ||
b411b363 PR |
991 | bio->bi_bdev = mdev->ldev->md_bdev; |
992 | bio->bi_sector = on_disk_sector; | |
4d95a10f LE |
993 | /* bio_add_page of a single page to an empty bio will always succeed, |
994 | * according to api. Do we want to assert that? */ | |
19f843aa LE |
995 | bio_add_page(bio, page, len, 0); |
996 | bio->bi_private = ctx; | |
b411b363 PR |
997 | bio->bi_end_io = bm_async_io_complete; |
998 | ||
0cf9d27e | 999 | if (drbd_insert_fault(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) { |
b411b363 PR |
1000 | bio->bi_rw |= rw; |
1001 | bio_endio(bio, -EIO); | |
1002 | } else { | |
1003 | submit_bio(rw, bio); | |
5a8b4242 LE |
1004 | /* this should not count as user activity and cause the |
1005 | * resync to throttle -- see drbd_rs_should_slow_down(). */ | |
1006 | atomic_add(len >> 9, &mdev->rs_sect_ev); | |
b411b363 PR |
1007 | } |
1008 | } | |
1009 | ||
b411b363 PR |
1010 | /* |
1011 | * bm_rw: read/write the whole bitmap from/to its on disk location. | |
1012 | */ | |
0e8488ad | 1013 | static int bm_rw(struct drbd_conf *mdev, int rw, unsigned flags, unsigned lazy_writeout_upper_idx) __must_hold(local) |
b411b363 | 1014 | { |
d1f3779b | 1015 | struct bm_aio_ctx *ctx; |
b411b363 | 1016 | struct drbd_bitmap *b = mdev->bitmap; |
6850c442 | 1017 | int num_pages, i, count = 0; |
b411b363 PR |
1018 | unsigned long now; |
1019 | char ppb[10]; | |
1020 | int err = 0; | |
1021 | ||
19f843aa LE |
1022 | /* |
1023 | * We are protected against bitmap disappearing/resizing by holding an | |
1024 | * ldev reference (caller must have called get_ldev()). | |
1025 | * For read/write, we are protected against changes to the bitmap by | |
1026 | * the bitmap lock (see drbd_bitmap_io). | |
1027 | * For lazy writeout, we don't care for ongoing changes to the bitmap, | |
1028 | * as we submit copies of pages anyways. | |
1029 | */ | |
d1f3779b | 1030 | |
22f46ce2 | 1031 | ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO); |
d1f3779b PR |
1032 | if (!ctx) |
1033 | return -ENOMEM; | |
1034 | ||
1035 | *ctx = (struct bm_aio_ctx) { | |
1036 | .mdev = mdev, | |
1037 | .in_flight = ATOMIC_INIT(1), | |
9e58c4da | 1038 | .done = 0, |
0e8488ad | 1039 | .flags = flags, |
d1f3779b PR |
1040 | .error = 0, |
1041 | .kref = { ATOMIC_INIT(2) }, | |
1042 | }; | |
1043 | ||
9e58c4da PR |
1044 | if (!get_ldev_if_state(mdev, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */ |
1045 | dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n"); | |
1046 | kfree(ctx); | |
1047 | return -ENODEV; | |
1048 | } | |
1049 | ||
d1f3779b | 1050 | if (!ctx->flags) |
20ceb2b2 | 1051 | WARN_ON(!(BM_LOCKED_MASK & b->bm_flags)); |
b411b363 | 1052 | |
6850c442 | 1053 | num_pages = b->bm_number_of_pages; |
b411b363 | 1054 | |
b411b363 | 1055 | now = jiffies; |
b411b363 PR |
1056 | |
1057 | /* let the layers below us try to merge these bios... */ | |
6850c442 | 1058 | for (i = 0; i < num_pages; i++) { |
19f843aa LE |
1059 | /* ignore completely unchanged pages */ |
1060 | if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx) | |
1061 | break; | |
1062 | if (rw & WRITE) { | |
d1aa4d04 PR |
1063 | if (!(flags & BM_WRITE_ALL_PAGES) && |
1064 | bm_test_page_unchanged(b->bm_pages[i])) { | |
19f843aa LE |
1065 | dynamic_dev_dbg(DEV, "skipped bm write for idx %u\n", i); |
1066 | continue; | |
1067 | } | |
1068 | /* during lazy writeout, | |
1069 | * ignore those pages not marked for lazy writeout. */ | |
1070 | if (lazy_writeout_upper_idx && | |
1071 | !bm_test_page_lazy_writeout(b->bm_pages[i])) { | |
1072 | dynamic_dev_dbg(DEV, "skipped bm lazy write for idx %u\n", i); | |
1073 | continue; | |
1074 | } | |
1075 | } | |
d1f3779b PR |
1076 | atomic_inc(&ctx->in_flight); |
1077 | bm_page_io_async(ctx, i, rw); | |
19f843aa LE |
1078 | ++count; |
1079 | cond_resched(); | |
1080 | } | |
b411b363 | 1081 | |
725a97e4 | 1082 | /* |
d1f3779b | 1083 | * We initialize ctx->in_flight to one to make sure bm_async_io_complete |
9e58c4da | 1084 | * will not set ctx->done early, and decrement / test it here. If there |
725a97e4 | 1085 | * are still some bios in flight, we need to wait for them here. |
9e58c4da PR |
1086 | * If all IO is done already (or nothing had been submitted), there is |
1087 | * no need to wait. Still, we need to put the kref associated with the | |
1088 | * "in_flight reached zero, all done" event. | |
725a97e4 | 1089 | */ |
d1f3779b | 1090 | if (!atomic_dec_and_test(&ctx->in_flight)) |
7caacb69 | 1091 | wait_until_done_or_disk_failure(mdev, mdev->ldev, &ctx->done); |
9e58c4da PR |
1092 | else |
1093 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); | |
d1f3779b | 1094 | |
19f843aa LE |
1095 | dev_info(DEV, "bitmap %s of %u pages took %lu jiffies\n", |
1096 | rw == WRITE ? "WRITE" : "READ", | |
1097 | count, jiffies - now); | |
b411b363 | 1098 | |
d1f3779b | 1099 | if (ctx->error) { |
b411b363 | 1100 | dev_alert(DEV, "we had at least one MD IO ERROR during bitmap IO\n"); |
383606e0 | 1101 | drbd_chk_io_error(mdev, 1, DRBD_META_IO_ERROR); |
d1f3779b | 1102 | err = -EIO; /* ctx->error ? */ |
b411b363 PR |
1103 | } |
1104 | ||
9e58c4da PR |
1105 | if (atomic_read(&ctx->in_flight)) |
1106 | err = -EIO; /* Disk failed during IO... */ | |
1107 | ||
b411b363 PR |
1108 | now = jiffies; |
1109 | if (rw == WRITE) { | |
b411b363 PR |
1110 | drbd_md_flush(mdev); |
1111 | } else /* rw == READ */ { | |
95a0f10c | 1112 | b->bm_set = bm_count_bits(b); |
b411b363 PR |
1113 | dev_info(DEV, "recounting of set bits took additional %lu jiffies\n", |
1114 | jiffies - now); | |
1115 | } | |
1116 | now = b->bm_set; | |
1117 | ||
1118 | dev_info(DEV, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n", | |
1119 | ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now); | |
1120 | ||
d1f3779b | 1121 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); |
b411b363 PR |
1122 | return err; |
1123 | } | |
1124 | ||
1125 | /** | |
1126 | * drbd_bm_read() - Read the whole bitmap from its on disk location. | |
1127 | * @mdev: DRBD device. | |
1128 | */ | |
1129 | int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local) | |
1130 | { | |
0e8488ad | 1131 | return bm_rw(mdev, READ, 0, 0); |
b411b363 PR |
1132 | } |
1133 | ||
1134 | /** | |
1135 | * drbd_bm_write() - Write the whole bitmap to its on disk location. | |
1136 | * @mdev: DRBD device. | |
19f843aa LE |
1137 | * |
1138 | * Will only write pages that have changed since last IO. | |
b411b363 PR |
1139 | */ |
1140 | int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local) | |
1141 | { | |
0e8488ad | 1142 | return bm_rw(mdev, WRITE, 0, 0); |
b411b363 PR |
1143 | } |
1144 | ||
d1aa4d04 PR |
1145 | /** |
1146 | * drbd_bm_write_all() - Write the whole bitmap to its on disk location. | |
1147 | * @mdev: DRBD device. | |
1148 | * | |
1149 | * Will write all pages. | |
1150 | */ | |
1151 | int drbd_bm_write_all(struct drbd_conf *mdev) __must_hold(local) | |
1152 | { | |
1153 | return bm_rw(mdev, WRITE, BM_WRITE_ALL_PAGES, 0); | |
1154 | } | |
1155 | ||
b411b363 | 1156 | /** |
19f843aa | 1157 | * drbd_bm_lazy_write_out() - Write bitmap pages 0 to @upper_idx-1, if they have changed. |
b411b363 | 1158 | * @mdev: DRBD device. |
19f843aa LE |
1159 | * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages |
1160 | */ | |
1161 | int drbd_bm_write_lazy(struct drbd_conf *mdev, unsigned upper_idx) __must_hold(local) | |
1162 | { | |
0e8488ad LE |
1163 | return bm_rw(mdev, WRITE, BM_AIO_COPY_PAGES, upper_idx); |
1164 | } | |
1165 | ||
1166 | /** | |
1167 | * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location. | |
1168 | * @mdev: DRBD device. | |
1169 | * | |
1170 | * Will only write pages that have changed since last IO. | |
1171 | * In contrast to drbd_bm_write(), this will copy the bitmap pages | |
1172 | * to temporary writeout pages. It is intended to trigger a full write-out | |
1173 | * while still allowing the bitmap to change, for example if a resync or online | |
1174 | * verify is aborted due to a failed peer disk, while local IO continues, or | |
1175 | * pending resync acks are still being processed. | |
1176 | */ | |
1177 | int drbd_bm_write_copy_pages(struct drbd_conf *mdev) __must_hold(local) | |
1178 | { | |
1179 | return bm_rw(mdev, WRITE, BM_AIO_COPY_PAGES, 0); | |
19f843aa LE |
1180 | } |
1181 | ||
1182 | ||
1183 | /** | |
1184 | * drbd_bm_write_page: Writes a PAGE_SIZE aligned piece of bitmap | |
1185 | * @mdev: DRBD device. | |
1186 | * @idx: bitmap page index | |
b411b363 | 1187 | * |
4b0715f0 LE |
1188 | * We don't want to special case on logical_block_size of the backend device, |
1189 | * so we submit PAGE_SIZE aligned pieces. | |
19f843aa | 1190 | * Note that on "most" systems, PAGE_SIZE is 4k. |
4b0715f0 LE |
1191 | * |
1192 | * In case this becomes an issue on systems with larger PAGE_SIZE, | |
1193 | * we may want to change this again to write 4k aligned 4k pieces. | |
b411b363 | 1194 | */ |
19f843aa | 1195 | int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local) |
b411b363 | 1196 | { |
d1f3779b PR |
1197 | struct bm_aio_ctx *ctx; |
1198 | int err; | |
b411b363 | 1199 | |
19f843aa | 1200 | if (bm_test_page_unchanged(mdev->bitmap->bm_pages[idx])) { |
7648cdfe | 1201 | dynamic_dev_dbg(DEV, "skipped bm page write for idx %u\n", idx); |
19f843aa | 1202 | return 0; |
b411b363 | 1203 | } |
19f843aa | 1204 | |
22f46ce2 | 1205 | ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO); |
d1f3779b PR |
1206 | if (!ctx) |
1207 | return -ENOMEM; | |
1208 | ||
1209 | *ctx = (struct bm_aio_ctx) { | |
1210 | .mdev = mdev, | |
1211 | .in_flight = ATOMIC_INIT(1), | |
9e58c4da | 1212 | .done = 0, |
d1f3779b PR |
1213 | .flags = BM_AIO_COPY_PAGES, |
1214 | .error = 0, | |
1215 | .kref = { ATOMIC_INIT(2) }, | |
1216 | }; | |
1217 | ||
9e58c4da PR |
1218 | if (!get_ldev_if_state(mdev, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */ |
1219 | dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in drbd_bm_write_page()\n"); | |
1220 | kfree(ctx); | |
1221 | return -ENODEV; | |
1222 | } | |
1223 | ||
d1f3779b | 1224 | bm_page_io_async(ctx, idx, WRITE_SYNC); |
7caacb69 | 1225 | wait_until_done_or_disk_failure(mdev, mdev->ldev, &ctx->done); |
19f843aa | 1226 | |
d1f3779b | 1227 | if (ctx->error) |
383606e0 | 1228 | drbd_chk_io_error(mdev, 1, DRBD_META_IO_ERROR); |
19f843aa LE |
1229 | /* that should force detach, so the in memory bitmap will be |
1230 | * gone in a moment as well. */ | |
1231 | ||
b411b363 | 1232 | mdev->bm_writ_cnt++; |
9e58c4da | 1233 | err = atomic_read(&ctx->in_flight) ? -EIO : ctx->error; |
d1f3779b | 1234 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); |
d1f3779b | 1235 | return err; |
b411b363 PR |
1236 | } |
1237 | ||
1238 | /* NOTE | |
1239 | * find_first_bit returns int, we return unsigned long. | |
4b0715f0 LE |
1240 | * For this to work on 32bit arch with bitnumbers > (1<<32), |
1241 | * we'd need to return u64, and get a whole lot of other places | |
1242 | * fixed where we still use unsigned long. | |
b411b363 PR |
1243 | * |
1244 | * this returns a bit number, NOT a sector! | |
1245 | */ | |
b411b363 | 1246 | static unsigned long __bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo, |
cfd8005c | 1247 | const int find_zero_bit) |
b411b363 PR |
1248 | { |
1249 | struct drbd_bitmap *b = mdev->bitmap; | |
b411b363 | 1250 | unsigned long *p_addr; |
4b0715f0 LE |
1251 | unsigned long bit_offset; |
1252 | unsigned i; | |
1253 | ||
b411b363 PR |
1254 | |
1255 | if (bm_fo > b->bm_bits) { | |
1256 | dev_err(DEV, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits); | |
4b0715f0 | 1257 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1258 | } else { |
1259 | while (bm_fo < b->bm_bits) { | |
19f843aa | 1260 | /* bit offset of the first bit in the page */ |
4b0715f0 | 1261 | bit_offset = bm_fo & ~BITS_PER_PAGE_MASK; |
cfd8005c | 1262 | p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo)); |
b411b363 PR |
1263 | |
1264 | if (find_zero_bit) | |
7e599e6e | 1265 | i = find_next_zero_bit_le(p_addr, |
4b0715f0 | 1266 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1267 | else |
7e599e6e | 1268 | i = find_next_bit_le(p_addr, |
4b0715f0 | 1269 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1270 | |
cfd8005c | 1271 | __bm_unmap(p_addr); |
b411b363 | 1272 | if (i < PAGE_SIZE*8) { |
4b0715f0 LE |
1273 | bm_fo = bit_offset + i; |
1274 | if (bm_fo >= b->bm_bits) | |
b411b363 PR |
1275 | break; |
1276 | goto found; | |
1277 | } | |
1278 | bm_fo = bit_offset + PAGE_SIZE*8; | |
1279 | } | |
4b0715f0 | 1280 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1281 | } |
1282 | found: | |
4b0715f0 | 1283 | return bm_fo; |
b411b363 PR |
1284 | } |
1285 | ||
1286 | static unsigned long bm_find_next(struct drbd_conf *mdev, | |
1287 | unsigned long bm_fo, const int find_zero_bit) | |
1288 | { | |
1289 | struct drbd_bitmap *b = mdev->bitmap; | |
4b0715f0 | 1290 | unsigned long i = DRBD_END_OF_BITMAP; |
b411b363 PR |
1291 | |
1292 | ERR_IF(!b) return i; | |
1293 | ERR_IF(!b->bm_pages) return i; | |
1294 | ||
1295 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1296 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1297 | bm_print_lock_info(mdev); |
1298 | ||
cfd8005c | 1299 | i = __bm_find_next(mdev, bm_fo, find_zero_bit); |
b411b363 PR |
1300 | |
1301 | spin_unlock_irq(&b->bm_lock); | |
1302 | return i; | |
1303 | } | |
1304 | ||
1305 | unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo) | |
1306 | { | |
1307 | return bm_find_next(mdev, bm_fo, 0); | |
1308 | } | |
1309 | ||
1310 | #if 0 | |
1311 | /* not yet needed for anything. */ | |
1312 | unsigned long drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo) | |
1313 | { | |
1314 | return bm_find_next(mdev, bm_fo, 1); | |
1315 | } | |
1316 | #endif | |
1317 | ||
1318 | /* does not spin_lock_irqsave. | |
1319 | * you must take drbd_bm_lock() first */ | |
1320 | unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo) | |
1321 | { | |
20ceb2b2 | 1322 | /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */ |
cfd8005c | 1323 | return __bm_find_next(mdev, bm_fo, 0); |
b411b363 PR |
1324 | } |
1325 | ||
1326 | unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo) | |
1327 | { | |
20ceb2b2 | 1328 | /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */ |
cfd8005c | 1329 | return __bm_find_next(mdev, bm_fo, 1); |
b411b363 PR |
1330 | } |
1331 | ||
1332 | /* returns number of bits actually changed. | |
1333 | * for val != 0, we change 0 -> 1, return code positive | |
1334 | * for val == 0, we change 1 -> 0, return code negative | |
1335 | * wants bitnr, not sector. | |
1336 | * expected to be called for only a few bits (e - s about BITS_PER_LONG). | |
1337 | * Must hold bitmap lock already. */ | |
b4ee79da | 1338 | static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s, |
829c6087 | 1339 | unsigned long e, int val) |
b411b363 PR |
1340 | { |
1341 | struct drbd_bitmap *b = mdev->bitmap; | |
1342 | unsigned long *p_addr = NULL; | |
1343 | unsigned long bitnr; | |
19f843aa | 1344 | unsigned int last_page_nr = -1U; |
b411b363 | 1345 | int c = 0; |
19f843aa | 1346 | int changed_total = 0; |
b411b363 PR |
1347 | |
1348 | if (e >= b->bm_bits) { | |
1349 | dev_err(DEV, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n", | |
1350 | s, e, b->bm_bits); | |
1351 | e = b->bm_bits ? b->bm_bits -1 : 0; | |
1352 | } | |
1353 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa | 1354 | unsigned int page_nr = bm_bit_to_page_idx(b, bitnr); |
b411b363 PR |
1355 | if (page_nr != last_page_nr) { |
1356 | if (p_addr) | |
cfd8005c | 1357 | __bm_unmap(p_addr); |
19f843aa LE |
1358 | if (c < 0) |
1359 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1360 | else if (c > 0) | |
1361 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1362 | changed_total += c; | |
1363 | c = 0; | |
cfd8005c | 1364 | p_addr = __bm_map_pidx(b, page_nr); |
b411b363 PR |
1365 | last_page_nr = page_nr; |
1366 | } | |
1367 | if (val) | |
7e599e6e | 1368 | c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 | 1369 | else |
7e599e6e | 1370 | c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 PR |
1371 | } |
1372 | if (p_addr) | |
cfd8005c | 1373 | __bm_unmap(p_addr); |
19f843aa LE |
1374 | if (c < 0) |
1375 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1376 | else if (c > 0) | |
1377 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1378 | changed_total += c; | |
1379 | b->bm_set += changed_total; | |
1380 | return changed_total; | |
b411b363 PR |
1381 | } |
1382 | ||
1383 | /* returns number of bits actually changed. | |
1384 | * for val != 0, we change 0 -> 1, return code positive | |
1385 | * for val == 0, we change 1 -> 0, return code negative | |
1386 | * wants bitnr, not sector */ | |
b4ee79da | 1387 | static int bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s, |
b411b363 PR |
1388 | const unsigned long e, int val) |
1389 | { | |
1390 | unsigned long flags; | |
1391 | struct drbd_bitmap *b = mdev->bitmap; | |
1392 | int c = 0; | |
1393 | ||
1394 | ERR_IF(!b) return 1; | |
1395 | ERR_IF(!b->bm_pages) return 0; | |
1396 | ||
1397 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1398 | if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags) |
b411b363 PR |
1399 | bm_print_lock_info(mdev); |
1400 | ||
829c6087 | 1401 | c = __bm_change_bits_to(mdev, s, e, val); |
b411b363 PR |
1402 | |
1403 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1404 | return c; | |
1405 | } | |
1406 | ||
1407 | /* returns number of bits changed 0 -> 1 */ | |
1408 | int drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1409 | { | |
1410 | return bm_change_bits_to(mdev, s, e, 1); | |
1411 | } | |
1412 | ||
1413 | /* returns number of bits changed 1 -> 0 */ | |
1414 | int drbd_bm_clear_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1415 | { | |
1416 | return -bm_change_bits_to(mdev, s, e, 0); | |
1417 | } | |
1418 | ||
1419 | /* sets all bits in full words, | |
1420 | * from first_word up to, but not including, last_word */ | |
1421 | static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b, | |
1422 | int page_nr, int first_word, int last_word) | |
1423 | { | |
1424 | int i; | |
1425 | int bits; | |
cfd8005c | 1426 | unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr]); |
b411b363 PR |
1427 | for (i = first_word; i < last_word; i++) { |
1428 | bits = hweight_long(paddr[i]); | |
1429 | paddr[i] = ~0UL; | |
1430 | b->bm_set += BITS_PER_LONG - bits; | |
1431 | } | |
cfd8005c | 1432 | kunmap_atomic(paddr); |
b411b363 PR |
1433 | } |
1434 | ||
829c6087 LE |
1435 | /* Same thing as drbd_bm_set_bits, |
1436 | * but more efficient for a large bit range. | |
b411b363 PR |
1437 | * You must first drbd_bm_lock(). |
1438 | * Can be called to set the whole bitmap in one go. | |
1439 | * Sets bits from s to e _inclusive_. */ | |
1440 | void _drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1441 | { | |
1442 | /* First set_bit from the first bit (s) | |
1443 | * up to the next long boundary (sl), | |
1444 | * then assign full words up to the last long boundary (el), | |
1445 | * then set_bit up to and including the last bit (e). | |
1446 | * | |
1447 | * Do not use memset, because we must account for changes, | |
1448 | * so we need to loop over the words with hweight() anyways. | |
1449 | */ | |
829c6087 | 1450 | struct drbd_bitmap *b = mdev->bitmap; |
b411b363 PR |
1451 | unsigned long sl = ALIGN(s,BITS_PER_LONG); |
1452 | unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1); | |
1453 | int first_page; | |
1454 | int last_page; | |
1455 | int page_nr; | |
1456 | int first_word; | |
1457 | int last_word; | |
1458 | ||
1459 | if (e - s <= 3*BITS_PER_LONG) { | |
1460 | /* don't bother; el and sl may even be wrong. */ | |
829c6087 LE |
1461 | spin_lock_irq(&b->bm_lock); |
1462 | __bm_change_bits_to(mdev, s, e, 1); | |
1463 | spin_unlock_irq(&b->bm_lock); | |
b411b363 PR |
1464 | return; |
1465 | } | |
1466 | ||
1467 | /* difference is large enough that we can trust sl and el */ | |
1468 | ||
829c6087 LE |
1469 | spin_lock_irq(&b->bm_lock); |
1470 | ||
b411b363 PR |
1471 | /* bits filling the current long */ |
1472 | if (sl) | |
829c6087 | 1473 | __bm_change_bits_to(mdev, s, sl-1, 1); |
b411b363 PR |
1474 | |
1475 | first_page = sl >> (3 + PAGE_SHIFT); | |
1476 | last_page = el >> (3 + PAGE_SHIFT); | |
1477 | ||
1478 | /* MLPP: modulo longs per page */ | |
1479 | /* LWPP: long words per page */ | |
1480 | first_word = MLPP(sl >> LN2_BPL); | |
1481 | last_word = LWPP; | |
1482 | ||
1483 | /* first and full pages, unless first page == last page */ | |
1484 | for (page_nr = first_page; page_nr < last_page; page_nr++) { | |
1485 | bm_set_full_words_within_one_page(mdev->bitmap, page_nr, first_word, last_word); | |
8ccee20e LE |
1486 | spin_unlock_irq(&b->bm_lock); |
1487 | cond_resched(); | |
b411b363 | 1488 | first_word = 0; |
8ccee20e | 1489 | spin_lock_irq(&b->bm_lock); |
b411b363 | 1490 | } |
b411b363 PR |
1491 | /* last page (respectively only page, for first page == last page) */ |
1492 | last_word = MLPP(el >> LN2_BPL); | |
4eccc579 LE |
1493 | |
1494 | /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples). | |
1495 | * ==> e = 32767, el = 32768, last_page = 2, | |
1496 | * and now last_word = 0. | |
1497 | * We do not want to touch last_page in this case, | |
1498 | * as we did not allocate it, it is not present in bitmap->bm_pages. | |
1499 | */ | |
1500 | if (last_word) | |
1501 | bm_set_full_words_within_one_page(mdev->bitmap, last_page, first_word, last_word); | |
b411b363 PR |
1502 | |
1503 | /* possibly trailing bits. | |
1504 | * example: (e & 63) == 63, el will be e+1. | |
1505 | * if that even was the very last bit, | |
1506 | * it would trigger an assert in __bm_change_bits_to() | |
1507 | */ | |
1508 | if (el <= e) | |
829c6087 LE |
1509 | __bm_change_bits_to(mdev, el, e, 1); |
1510 | spin_unlock_irq(&b->bm_lock); | |
b411b363 PR |
1511 | } |
1512 | ||
1513 | /* returns bit state | |
1514 | * wants bitnr, NOT sector. | |
1515 | * inherently racy... area needs to be locked by means of {al,rs}_lru | |
1516 | * 1 ... bit set | |
1517 | * 0 ... bit not set | |
1518 | * -1 ... first out of bounds access, stop testing for bits! | |
1519 | */ | |
1520 | int drbd_bm_test_bit(struct drbd_conf *mdev, const unsigned long bitnr) | |
1521 | { | |
1522 | unsigned long flags; | |
1523 | struct drbd_bitmap *b = mdev->bitmap; | |
1524 | unsigned long *p_addr; | |
1525 | int i; | |
1526 | ||
1527 | ERR_IF(!b) return 0; | |
1528 | ERR_IF(!b->bm_pages) return 0; | |
1529 | ||
1530 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1531 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1532 | bm_print_lock_info(mdev); |
1533 | if (bitnr < b->bm_bits) { | |
19f843aa | 1534 | p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr)); |
7e599e6e | 1535 | i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0; |
b411b363 PR |
1536 | bm_unmap(p_addr); |
1537 | } else if (bitnr == b->bm_bits) { | |
1538 | i = -1; | |
1539 | } else { /* (bitnr > b->bm_bits) */ | |
1540 | dev_err(DEV, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits); | |
1541 | i = 0; | |
1542 | } | |
1543 | ||
1544 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1545 | return i; | |
1546 | } | |
1547 | ||
1548 | /* returns number of bits set in the range [s, e] */ | |
1549 | int drbd_bm_count_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1550 | { | |
1551 | unsigned long flags; | |
1552 | struct drbd_bitmap *b = mdev->bitmap; | |
19f843aa | 1553 | unsigned long *p_addr = NULL; |
b411b363 | 1554 | unsigned long bitnr; |
19f843aa | 1555 | unsigned int page_nr = -1U; |
b411b363 | 1556 | int c = 0; |
b411b363 PR |
1557 | |
1558 | /* If this is called without a bitmap, that is a bug. But just to be | |
1559 | * robust in case we screwed up elsewhere, in that case pretend there | |
1560 | * was one dirty bit in the requested area, so we won't try to do a | |
1561 | * local read there (no bitmap probably implies no disk) */ | |
1562 | ERR_IF(!b) return 1; | |
1563 | ERR_IF(!b->bm_pages) return 1; | |
1564 | ||
1565 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1566 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1567 | bm_print_lock_info(mdev); |
1568 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa LE |
1569 | unsigned int idx = bm_bit_to_page_idx(b, bitnr); |
1570 | if (page_nr != idx) { | |
1571 | page_nr = idx; | |
b411b363 PR |
1572 | if (p_addr) |
1573 | bm_unmap(p_addr); | |
19f843aa | 1574 | p_addr = bm_map_pidx(b, idx); |
b411b363 PR |
1575 | } |
1576 | ERR_IF (bitnr >= b->bm_bits) { | |
1577 | dev_err(DEV, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits); | |
1578 | } else { | |
7e599e6e | 1579 | c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr)); |
b411b363 PR |
1580 | } |
1581 | } | |
1582 | if (p_addr) | |
1583 | bm_unmap(p_addr); | |
1584 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1585 | return c; | |
1586 | } | |
1587 | ||
1588 | ||
1589 | /* inherently racy... | |
1590 | * return value may be already out-of-date when this function returns. | |
1591 | * but the general usage is that this is only use during a cstate when bits are | |
1592 | * only cleared, not set, and typically only care for the case when the return | |
1593 | * value is zero, or we already "locked" this "bitmap extent" by other means. | |
1594 | * | |
1595 | * enr is bm-extent number, since we chose to name one sector (512 bytes) | |
1596 | * worth of the bitmap a "bitmap extent". | |
1597 | * | |
1598 | * TODO | |
1599 | * I think since we use it like a reference count, we should use the real | |
1600 | * reference count of some bitmap extent element from some lru instead... | |
1601 | * | |
1602 | */ | |
1603 | int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr) | |
1604 | { | |
1605 | struct drbd_bitmap *b = mdev->bitmap; | |
1606 | int count, s, e; | |
1607 | unsigned long flags; | |
1608 | unsigned long *p_addr, *bm; | |
1609 | ||
1610 | ERR_IF(!b) return 0; | |
1611 | ERR_IF(!b->bm_pages) return 0; | |
1612 | ||
1613 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1614 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1615 | bm_print_lock_info(mdev); |
1616 | ||
1617 | s = S2W(enr); | |
1618 | e = min((size_t)S2W(enr+1), b->bm_words); | |
1619 | count = 0; | |
1620 | if (s < b->bm_words) { | |
1621 | int n = e-s; | |
19f843aa | 1622 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); |
b411b363 PR |
1623 | bm = p_addr + MLPP(s); |
1624 | while (n--) | |
1625 | count += hweight_long(*bm++); | |
1626 | bm_unmap(p_addr); | |
1627 | } else { | |
1628 | dev_err(DEV, "start offset (%d) too large in drbd_bm_e_weight\n", s); | |
1629 | } | |
1630 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1631 | return count; | |
1632 | } | |
1633 | ||
4b0715f0 LE |
1634 | /* Set all bits covered by the AL-extent al_enr. |
1635 | * Returns number of bits changed. */ | |
b411b363 PR |
1636 | unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev, unsigned long al_enr) |
1637 | { | |
1638 | struct drbd_bitmap *b = mdev->bitmap; | |
1639 | unsigned long *p_addr, *bm; | |
1640 | unsigned long weight; | |
4b0715f0 LE |
1641 | unsigned long s, e; |
1642 | int count, i, do_now; | |
b411b363 PR |
1643 | ERR_IF(!b) return 0; |
1644 | ERR_IF(!b->bm_pages) return 0; | |
1645 | ||
1646 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1647 | if (BM_DONT_SET & b->bm_flags) |
b411b363 PR |
1648 | bm_print_lock_info(mdev); |
1649 | weight = b->bm_set; | |
1650 | ||
1651 | s = al_enr * BM_WORDS_PER_AL_EXT; | |
1652 | e = min_t(size_t, s + BM_WORDS_PER_AL_EXT, b->bm_words); | |
1653 | /* assert that s and e are on the same page */ | |
1654 | D_ASSERT((e-1) >> (PAGE_SHIFT - LN2_BPL + 3) | |
1655 | == s >> (PAGE_SHIFT - LN2_BPL + 3)); | |
1656 | count = 0; | |
1657 | if (s < b->bm_words) { | |
1658 | i = do_now = e-s; | |
19f843aa | 1659 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); |
b411b363 PR |
1660 | bm = p_addr + MLPP(s); |
1661 | while (i--) { | |
1662 | count += hweight_long(*bm); | |
1663 | *bm = -1UL; | |
1664 | bm++; | |
1665 | } | |
1666 | bm_unmap(p_addr); | |
1667 | b->bm_set += do_now*BITS_PER_LONG - count; | |
1668 | if (e == b->bm_words) | |
1669 | b->bm_set -= bm_clear_surplus(b); | |
1670 | } else { | |
4b0715f0 | 1671 | dev_err(DEV, "start offset (%lu) too large in drbd_bm_ALe_set_all\n", s); |
b411b363 PR |
1672 | } |
1673 | weight = b->bm_set - weight; | |
1674 | spin_unlock_irq(&b->bm_lock); | |
1675 | return weight; | |
1676 | } |