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 | ||
f88c5d90 LE |
25 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
26 | ||
5fb3bc4d | 27 | #include <linux/bitmap.h> |
b411b363 PR |
28 | #include <linux/vmalloc.h> |
29 | #include <linux/string.h> | |
30 | #include <linux/drbd.h> | |
5a0e3ad6 | 31 | #include <linux/slab.h> |
dbcbdc43 | 32 | #include <linux/highmem.h> |
f0ff1357 | 33 | |
b411b363 PR |
34 | #include "drbd_int.h" |
35 | ||
95a0f10c | 36 | |
b411b363 PR |
37 | /* OPAQUE outside this file! |
38 | * interface defined in drbd_int.h | |
39 | ||
40 | * convention: | |
41 | * function name drbd_bm_... => used elsewhere, "public". | |
42 | * function name bm_... => internal to implementation, "private". | |
4b0715f0 LE |
43 | */ |
44 | ||
b411b363 | 45 | |
4b0715f0 LE |
46 | /* |
47 | * LIMITATIONS: | |
48 | * We want to support >= peta byte of backend storage, while for now still using | |
49 | * a granularity of one bit per 4KiB of storage. | |
50 | * 1 << 50 bytes backend storage (1 PiB) | |
51 | * 1 << (50 - 12) bits needed | |
52 | * 38 --> we need u64 to index and count bits | |
53 | * 1 << (38 - 3) bitmap bytes needed | |
54 | * 35 --> we still need u64 to index and count bytes | |
55 | * (that's 32 GiB of bitmap for 1 PiB storage) | |
56 | * 1 << (35 - 2) 32bit longs needed | |
57 | * 33 --> we'd even need u64 to index and count 32bit long words. | |
58 | * 1 << (35 - 3) 64bit longs needed | |
59 | * 32 --> we could get away with a 32bit unsigned int to index and count | |
60 | * 64bit long words, but I rather stay with unsigned long for now. | |
61 | * We probably should neither count nor point to bytes or long words | |
62 | * directly, but either by bitnumber, or by page index and offset. | |
63 | * 1 << (35 - 12) | |
64 | * 22 --> we need that much 4KiB pages of bitmap. | |
65 | * 1 << (22 + 3) --> on a 64bit arch, | |
66 | * we need 32 MiB to store the array of page pointers. | |
67 | * | |
68 | * Because I'm lazy, and because the resulting patch was too large, too ugly | |
69 | * and still incomplete, on 32bit we still "only" support 16 TiB (minus some), | |
70 | * (1 << 32) bits * 4k storage. | |
71 | * | |
72 | ||
73 | * bitmap storage and IO: | |
74 | * Bitmap is stored little endian on disk, and is kept little endian in | |
75 | * core memory. Currently we still hold the full bitmap in core as long | |
76 | * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage | |
77 | * seems excessive. | |
78 | * | |
24c4830c | 79 | * We plan to reduce the amount of in-core bitmap pages by paging them in |
4b0715f0 LE |
80 | * and out against their on-disk location as necessary, but need to make |
81 | * sure we don't cause too much meta data IO, and must not deadlock in | |
82 | * tight memory situations. This needs some more work. | |
b411b363 PR |
83 | */ |
84 | ||
85 | /* | |
86 | * NOTE | |
87 | * Access to the *bm_pages is protected by bm_lock. | |
88 | * It is safe to read the other members within the lock. | |
89 | * | |
90 | * drbd_bm_set_bits is called from bio_endio callbacks, | |
91 | * We may be called with irq already disabled, | |
92 | * so we need spin_lock_irqsave(). | |
93 | * And we need the kmap_atomic. | |
94 | */ | |
95 | struct drbd_bitmap { | |
96 | struct page **bm_pages; | |
97 | spinlock_t bm_lock; | |
4b0715f0 LE |
98 | |
99 | /* see LIMITATIONS: above */ | |
100 | ||
b411b363 PR |
101 | unsigned long bm_set; /* nr of set bits; THINK maybe atomic_t? */ |
102 | unsigned long bm_bits; | |
103 | size_t bm_words; | |
104 | size_t bm_number_of_pages; | |
105 | sector_t bm_dev_capacity; | |
8a03ae2a | 106 | struct mutex bm_change; /* serializes resize operations */ |
b411b363 | 107 | |
19f843aa | 108 | wait_queue_head_t bm_io_wait; /* used to serialize IO of single pages */ |
b411b363 | 109 | |
20ceb2b2 | 110 | enum bm_flag bm_flags; |
b411b363 PR |
111 | |
112 | /* debugging aid, in case we are still racy somewhere */ | |
113 | char *bm_why; | |
114 | struct task_struct *bm_task; | |
115 | }; | |
116 | ||
b411b363 | 117 | #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__) |
b30ab791 | 118 | static void __bm_print_lock_info(struct drbd_device *device, const char *func) |
b411b363 | 119 | { |
b30ab791 | 120 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
121 | if (!__ratelimit(&drbd_ratelimit_state)) |
122 | return; | |
c60b0251 AG |
123 | drbd_err(device, "FIXME %s[%d] in %s, bitmap locked for '%s' by %s[%d]\n", |
124 | current->comm, task_pid_nr(current), | |
125 | func, b->bm_why ?: "?", | |
126 | b->bm_task->comm, task_pid_nr(b->bm_task)); | |
b411b363 PR |
127 | } |
128 | ||
b30ab791 | 129 | void drbd_bm_lock(struct drbd_device *device, char *why, enum bm_flag flags) |
b411b363 | 130 | { |
b30ab791 | 131 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
132 | int trylock_failed; |
133 | ||
134 | if (!b) { | |
d0180171 | 135 | drbd_err(device, "FIXME no bitmap in drbd_bm_lock!?\n"); |
b411b363 PR |
136 | return; |
137 | } | |
138 | ||
8a03ae2a | 139 | trylock_failed = !mutex_trylock(&b->bm_change); |
b411b363 PR |
140 | |
141 | if (trylock_failed) { | |
c60b0251 AG |
142 | drbd_warn(device, "%s[%d] going to '%s' but bitmap already locked for '%s' by %s[%d]\n", |
143 | current->comm, task_pid_nr(current), | |
144 | why, b->bm_why ?: "?", | |
145 | b->bm_task->comm, task_pid_nr(b->bm_task)); | |
8a03ae2a | 146 | mutex_lock(&b->bm_change); |
b411b363 | 147 | } |
20ceb2b2 | 148 | if (BM_LOCKED_MASK & b->bm_flags) |
d0180171 | 149 | drbd_err(device, "FIXME bitmap already locked in bm_lock\n"); |
20ceb2b2 | 150 | b->bm_flags |= flags & BM_LOCKED_MASK; |
b411b363 PR |
151 | |
152 | b->bm_why = why; | |
153 | b->bm_task = current; | |
154 | } | |
155 | ||
b30ab791 | 156 | void drbd_bm_unlock(struct drbd_device *device) |
b411b363 | 157 | { |
b30ab791 | 158 | struct drbd_bitmap *b = device->bitmap; |
b411b363 | 159 | if (!b) { |
d0180171 | 160 | drbd_err(device, "FIXME no bitmap in drbd_bm_unlock!?\n"); |
b411b363 PR |
161 | return; |
162 | } | |
163 | ||
b30ab791 | 164 | if (!(BM_LOCKED_MASK & device->bitmap->bm_flags)) |
d0180171 | 165 | drbd_err(device, "FIXME bitmap not locked in bm_unlock\n"); |
b411b363 | 166 | |
20ceb2b2 | 167 | b->bm_flags &= ~BM_LOCKED_MASK; |
b411b363 PR |
168 | b->bm_why = NULL; |
169 | b->bm_task = NULL; | |
8a03ae2a | 170 | mutex_unlock(&b->bm_change); |
b411b363 PR |
171 | } |
172 | ||
19f843aa LE |
173 | /* we store some "meta" info about our pages in page->private */ |
174 | /* at a granularity of 4k storage per bitmap bit: | |
175 | * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks | |
176 | * 1<<38 bits, | |
177 | * 1<<23 4k bitmap pages. | |
178 | * Use 24 bits as page index, covers 2 peta byte storage | |
179 | * at a granularity of 4k per bit. | |
180 | * Used to report the failed page idx on io error from the endio handlers. | |
181 | */ | |
182 | #define BM_PAGE_IDX_MASK ((1UL<<24)-1) | |
183 | /* this page is currently read in, or written back */ | |
184 | #define BM_PAGE_IO_LOCK 31 | |
185 | /* if there has been an IO error for this page */ | |
186 | #define BM_PAGE_IO_ERROR 30 | |
187 | /* this is to be able to intelligently skip disk IO, | |
188 | * set if bits have been set since last IO. */ | |
189 | #define BM_PAGE_NEED_WRITEOUT 29 | |
190 | /* to mark for lazy writeout once syncer cleared all clearable bits, | |
191 | * we if bits have been cleared since last IO. */ | |
192 | #define BM_PAGE_LAZY_WRITEOUT 28 | |
45dfffeb LE |
193 | /* pages marked with this "HINT" will be considered for writeout |
194 | * on activity log transactions */ | |
195 | #define BM_PAGE_HINT_WRITEOUT 27 | |
19f843aa | 196 | |
24c4830c | 197 | /* store_page_idx uses non-atomic assignment. It is only used directly after |
19f843aa LE |
198 | * allocating the page. All other bm_set_page_* and bm_clear_page_* need to |
199 | * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap | |
200 | * changes) may happen from various contexts, and wait_on_bit/wake_up_bit | |
201 | * requires it all to be atomic as well. */ | |
202 | static void bm_store_page_idx(struct page *page, unsigned long idx) | |
95a0f10c | 203 | { |
19f843aa | 204 | BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK)); |
0c7db279 | 205 | set_page_private(page, idx); |
95a0f10c LE |
206 | } |
207 | ||
19f843aa | 208 | static unsigned long bm_page_to_idx(struct page *page) |
b411b363 | 209 | { |
19f843aa LE |
210 | return page_private(page) & BM_PAGE_IDX_MASK; |
211 | } | |
212 | ||
213 | /* As is very unlikely that the same page is under IO from more than one | |
214 | * context, we can get away with a bit per page and one wait queue per bitmap. | |
215 | */ | |
b30ab791 | 216 | static void bm_page_lock_io(struct drbd_device *device, int page_nr) |
19f843aa | 217 | { |
b30ab791 | 218 | struct drbd_bitmap *b = device->bitmap; |
19f843aa LE |
219 | void *addr = &page_private(b->bm_pages[page_nr]); |
220 | wait_event(b->bm_io_wait, !test_and_set_bit(BM_PAGE_IO_LOCK, addr)); | |
221 | } | |
222 | ||
b30ab791 | 223 | static void bm_page_unlock_io(struct drbd_device *device, int page_nr) |
19f843aa | 224 | { |
b30ab791 | 225 | struct drbd_bitmap *b = device->bitmap; |
19f843aa | 226 | void *addr = &page_private(b->bm_pages[page_nr]); |
4738fa16 | 227 | clear_bit_unlock(BM_PAGE_IO_LOCK, addr); |
b30ab791 | 228 | wake_up(&device->bitmap->bm_io_wait); |
19f843aa LE |
229 | } |
230 | ||
231 | /* set _before_ submit_io, so it may be reset due to being changed | |
232 | * while this page is in flight... will get submitted later again */ | |
233 | static void bm_set_page_unchanged(struct page *page) | |
234 | { | |
235 | /* use cmpxchg? */ | |
236 | clear_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
237 | clear_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
238 | } | |
239 | ||
240 | static void bm_set_page_need_writeout(struct page *page) | |
241 | { | |
242 | set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
243 | } | |
244 | ||
45dfffeb LE |
245 | /** |
246 | * drbd_bm_mark_for_writeout() - mark a page with a "hint" to be considered for writeout | |
b30ab791 | 247 | * @device: DRBD device. |
45dfffeb LE |
248 | * @page_nr: the bitmap page to mark with the "hint" flag |
249 | * | |
250 | * From within an activity log transaction, we mark a few pages with these | |
251 | * hints, then call drbd_bm_write_hinted(), which will only write out changed | |
252 | * pages which are flagged with this mark. | |
253 | */ | |
b30ab791 | 254 | void drbd_bm_mark_for_writeout(struct drbd_device *device, int page_nr) |
45dfffeb LE |
255 | { |
256 | struct page *page; | |
b30ab791 | 257 | if (page_nr >= device->bitmap->bm_number_of_pages) { |
d0180171 | 258 | drbd_warn(device, "BAD: page_nr: %u, number_of_pages: %u\n", |
b30ab791 | 259 | page_nr, (int)device->bitmap->bm_number_of_pages); |
45dfffeb LE |
260 | return; |
261 | } | |
b30ab791 | 262 | page = device->bitmap->bm_pages[page_nr]; |
45dfffeb LE |
263 | set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page)); |
264 | } | |
265 | ||
19f843aa LE |
266 | static int bm_test_page_unchanged(struct page *page) |
267 | { | |
268 | volatile const unsigned long *addr = &page_private(page); | |
269 | return (*addr & ((1UL<<BM_PAGE_NEED_WRITEOUT)|(1UL<<BM_PAGE_LAZY_WRITEOUT))) == 0; | |
270 | } | |
b411b363 | 271 | |
19f843aa LE |
272 | static void bm_set_page_io_err(struct page *page) |
273 | { | |
274 | set_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
275 | } | |
276 | ||
277 | static void bm_clear_page_io_err(struct page *page) | |
278 | { | |
279 | clear_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
280 | } | |
281 | ||
282 | static void bm_set_page_lazy_writeout(struct page *page) | |
283 | { | |
284 | set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
285 | } | |
286 | ||
287 | static int bm_test_page_lazy_writeout(struct page *page) | |
288 | { | |
289 | return test_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
290 | } | |
291 | ||
292 | /* on a 32bit box, this would allow for exactly (2<<38) bits. */ | |
293 | static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr) | |
294 | { | |
b411b363 | 295 | /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */ |
19f843aa | 296 | unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3); |
b411b363 | 297 | BUG_ON(page_nr >= b->bm_number_of_pages); |
19f843aa LE |
298 | return page_nr; |
299 | } | |
b411b363 | 300 | |
19f843aa LE |
301 | static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr) |
302 | { | |
303 | /* page_nr = (bitnr/8) >> PAGE_SHIFT; */ | |
304 | unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3); | |
305 | BUG_ON(page_nr >= b->bm_number_of_pages); | |
306 | return page_nr; | |
b411b363 PR |
307 | } |
308 | ||
589973a7 | 309 | static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) |
95a0f10c LE |
310 | { |
311 | struct page *page = b->bm_pages[idx]; | |
cfd8005c | 312 | return (unsigned long *) kmap_atomic(page); |
95a0f10c LE |
313 | } |
314 | ||
315 | static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) | |
316 | { | |
cfd8005c | 317 | return __bm_map_pidx(b, idx); |
95a0f10c LE |
318 | } |
319 | ||
cfd8005c | 320 | static void __bm_unmap(unsigned long *p_addr) |
b411b363 | 321 | { |
cfd8005c | 322 | kunmap_atomic(p_addr); |
b411b363 PR |
323 | }; |
324 | ||
325 | static void bm_unmap(unsigned long *p_addr) | |
326 | { | |
cfd8005c | 327 | return __bm_unmap(p_addr); |
b411b363 PR |
328 | } |
329 | ||
330 | /* long word offset of _bitmap_ sector */ | |
331 | #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) | |
332 | /* word offset from start of bitmap to word number _in_page_ | |
333 | * modulo longs per page | |
334 | #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long)) | |
24c4830c | 335 | hm, well, Philipp thinks gcc might not optimize the % into & (... - 1) |
b411b363 PR |
336 | so do it explicitly: |
337 | */ | |
338 | #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1)) | |
339 | ||
340 | /* Long words per page */ | |
341 | #define LWPP (PAGE_SIZE/sizeof(long)) | |
342 | ||
343 | /* | |
344 | * actually most functions herein should take a struct drbd_bitmap*, not a | |
b30ab791 | 345 | * struct drbd_device*, but for the debug macros I like to have the device around |
b411b363 PR |
346 | * to be able to report device specific. |
347 | */ | |
348 | ||
19f843aa | 349 | |
b411b363 PR |
350 | static void bm_free_pages(struct page **pages, unsigned long number) |
351 | { | |
352 | unsigned long i; | |
353 | if (!pages) | |
354 | return; | |
355 | ||
356 | for (i = 0; i < number; i++) { | |
357 | if (!pages[i]) { | |
f88c5d90 LE |
358 | pr_alert("bm_free_pages tried to free a NULL pointer; i=%lu n=%lu\n", |
359 | i, number); | |
b411b363 PR |
360 | continue; |
361 | } | |
362 | __free_page(pages[i]); | |
363 | pages[i] = NULL; | |
364 | } | |
365 | } | |
366 | ||
1d5cfdb0 | 367 | static inline void bm_vk_free(void *ptr) |
b411b363 | 368 | { |
1d5cfdb0 | 369 | kvfree(ptr); |
b411b363 PR |
370 | } |
371 | ||
372 | /* | |
373 | * "have" and "want" are NUMBER OF PAGES. | |
374 | */ | |
375 | static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want) | |
376 | { | |
377 | struct page **old_pages = b->bm_pages; | |
378 | struct page **new_pages, *page; | |
1d5cfdb0 | 379 | unsigned int i, bytes; |
b411b363 PR |
380 | unsigned long have = b->bm_number_of_pages; |
381 | ||
382 | BUG_ON(have == 0 && old_pages != NULL); | |
383 | BUG_ON(have != 0 && old_pages == NULL); | |
384 | ||
385 | if (have == want) | |
386 | return old_pages; | |
387 | ||
388 | /* Trying kmalloc first, falling back to vmalloc. | |
0b143d43 LE |
389 | * GFP_NOIO, as this is called while drbd IO is "suspended", |
390 | * and during resize or attach on diskless Primary, | |
391 | * we must not block on IO to ourselves. | |
bc891c9a | 392 | * Context is receiver thread or dmsetup. */ |
b411b363 | 393 | bytes = sizeof(struct page *)*want; |
8be04b93 | 394 | new_pages = kzalloc(bytes, GFP_NOIO | __GFP_NOWARN); |
b411b363 | 395 | if (!new_pages) { |
0b143d43 LE |
396 | new_pages = __vmalloc(bytes, |
397 | GFP_NOIO | __GFP_HIGHMEM | __GFP_ZERO, | |
398 | PAGE_KERNEL); | |
b411b363 PR |
399 | if (!new_pages) |
400 | return NULL; | |
b411b363 PR |
401 | } |
402 | ||
b411b363 PR |
403 | if (want >= have) { |
404 | for (i = 0; i < have; i++) | |
405 | new_pages[i] = old_pages[i]; | |
406 | for (; i < want; i++) { | |
0b143d43 | 407 | page = alloc_page(GFP_NOIO | __GFP_HIGHMEM); |
b411b363 PR |
408 | if (!page) { |
409 | bm_free_pages(new_pages + have, i - have); | |
1d5cfdb0 | 410 | bm_vk_free(new_pages); |
b411b363 PR |
411 | return NULL; |
412 | } | |
19f843aa LE |
413 | /* we want to know which page it is |
414 | * from the endio handlers */ | |
415 | bm_store_page_idx(page, i); | |
b411b363 PR |
416 | new_pages[i] = page; |
417 | } | |
418 | } else { | |
419 | for (i = 0; i < want; i++) | |
420 | new_pages[i] = old_pages[i]; | |
421 | /* NOT HERE, we are outside the spinlock! | |
422 | bm_free_pages(old_pages + want, have - want); | |
423 | */ | |
424 | } | |
425 | ||
b411b363 PR |
426 | return new_pages; |
427 | } | |
428 | ||
429 | /* | |
430 | * called on driver init only. TODO call when a device is created. | |
b30ab791 | 431 | * allocates the drbd_bitmap, and stores it in device->bitmap. |
b411b363 | 432 | */ |
b30ab791 | 433 | int drbd_bm_init(struct drbd_device *device) |
b411b363 | 434 | { |
b30ab791 | 435 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
436 | WARN_ON(b != NULL); |
437 | b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL); | |
438 | if (!b) | |
439 | return -ENOMEM; | |
440 | spin_lock_init(&b->bm_lock); | |
8a03ae2a | 441 | mutex_init(&b->bm_change); |
b411b363 PR |
442 | init_waitqueue_head(&b->bm_io_wait); |
443 | ||
b30ab791 | 444 | device->bitmap = b; |
b411b363 PR |
445 | |
446 | return 0; | |
447 | } | |
448 | ||
b30ab791 | 449 | sector_t drbd_bm_capacity(struct drbd_device *device) |
b411b363 | 450 | { |
b30ab791 | 451 | if (!expect(device->bitmap)) |
841ce241 | 452 | return 0; |
b30ab791 | 453 | return device->bitmap->bm_dev_capacity; |
b411b363 PR |
454 | } |
455 | ||
456 | /* called on driver unload. TODO: call when a device is destroyed. | |
457 | */ | |
b30ab791 | 458 | void drbd_bm_cleanup(struct drbd_device *device) |
b411b363 | 459 | { |
b30ab791 | 460 | if (!expect(device->bitmap)) |
841ce241 | 461 | return; |
b30ab791 | 462 | bm_free_pages(device->bitmap->bm_pages, device->bitmap->bm_number_of_pages); |
1d5cfdb0 | 463 | bm_vk_free(device->bitmap->bm_pages); |
b30ab791 AG |
464 | kfree(device->bitmap); |
465 | device->bitmap = NULL; | |
b411b363 PR |
466 | } |
467 | ||
468 | /* | |
469 | * since (b->bm_bits % BITS_PER_LONG) != 0, | |
470 | * this masks out the remaining bits. | |
471 | * Returns the number of bits cleared. | |
472 | */ | |
2630628b | 473 | #ifndef BITS_PER_PAGE |
95a0f10c LE |
474 | #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3)) |
475 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1) | |
2630628b LE |
476 | #else |
477 | # if BITS_PER_PAGE != (1UL << (PAGE_SHIFT + 3)) | |
478 | # error "ambiguous BITS_PER_PAGE" | |
479 | # endif | |
480 | #endif | |
95a0f10c | 481 | #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1) |
b411b363 PR |
482 | static int bm_clear_surplus(struct drbd_bitmap *b) |
483 | { | |
95a0f10c | 484 | unsigned long mask; |
b411b363 | 485 | unsigned long *p_addr, *bm; |
95a0f10c LE |
486 | int tmp; |
487 | int cleared = 0; | |
b411b363 | 488 | |
95a0f10c LE |
489 | /* number of bits modulo bits per page */ |
490 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
491 | /* mask the used bits of the word containing the last bit */ | |
492 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
493 | /* bitmap is always stored little endian, | |
494 | * on disk and in core memory alike */ | |
495 | mask = cpu_to_lel(mask); | |
496 | ||
6850c442 | 497 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
498 | bm = p_addr + (tmp/BITS_PER_LONG); |
499 | if (mask) { | |
500 | /* If mask != 0, we are not exactly aligned, so bm now points | |
501 | * to the long containing the last bit. | |
502 | * If mask == 0, bm already points to the word immediately | |
503 | * after the last (long word aligned) bit. */ | |
b411b363 PR |
504 | cleared = hweight_long(*bm & ~mask); |
505 | *bm &= mask; | |
95a0f10c | 506 | bm++; |
b411b363 PR |
507 | } |
508 | ||
95a0f10c LE |
509 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
510 | /* on a 32bit arch, we may need to zero out | |
511 | * a padding long to align with a 64bit remote */ | |
b411b363 PR |
512 | cleared += hweight_long(*bm); |
513 | *bm = 0; | |
514 | } | |
515 | bm_unmap(p_addr); | |
516 | return cleared; | |
517 | } | |
518 | ||
519 | static void bm_set_surplus(struct drbd_bitmap *b) | |
520 | { | |
95a0f10c | 521 | unsigned long mask; |
b411b363 | 522 | unsigned long *p_addr, *bm; |
95a0f10c LE |
523 | int tmp; |
524 | ||
525 | /* number of bits modulo bits per page */ | |
526 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
527 | /* mask the used bits of the word containing the last bit */ | |
528 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
529 | /* bitmap is always stored little endian, | |
530 | * on disk and in core memory alike */ | |
531 | mask = cpu_to_lel(mask); | |
532 | ||
6850c442 | 533 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
534 | bm = p_addr + (tmp/BITS_PER_LONG); |
535 | if (mask) { | |
536 | /* If mask != 0, we are not exactly aligned, so bm now points | |
537 | * to the long containing the last bit. | |
538 | * If mask == 0, bm already points to the word immediately | |
539 | * after the last (long word aligned) bit. */ | |
b411b363 | 540 | *bm |= ~mask; |
95a0f10c | 541 | bm++; |
b411b363 PR |
542 | } |
543 | ||
95a0f10c LE |
544 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
545 | /* on a 32bit arch, we may need to zero out | |
546 | * a padding long to align with a 64bit remote */ | |
547 | *bm = ~0UL; | |
b411b363 PR |
548 | } |
549 | bm_unmap(p_addr); | |
550 | } | |
551 | ||
4b0715f0 LE |
552 | /* you better not modify the bitmap while this is running, |
553 | * or its results will be stale */ | |
95a0f10c | 554 | static unsigned long bm_count_bits(struct drbd_bitmap *b) |
b411b363 | 555 | { |
4b0715f0 | 556 | unsigned long *p_addr; |
b411b363 | 557 | unsigned long bits = 0; |
4b0715f0 | 558 | unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1; |
5fb3bc4d | 559 | int idx, last_word; |
4b0715f0 LE |
560 | |
561 | /* all but last page */ | |
6850c442 | 562 | for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) { |
cfd8005c | 563 | p_addr = __bm_map_pidx(b, idx); |
5fb3bc4d | 564 | bits += bitmap_weight(p_addr, BITS_PER_PAGE); |
cfd8005c | 565 | __bm_unmap(p_addr); |
b411b363 PR |
566 | cond_resched(); |
567 | } | |
4b0715f0 LE |
568 | /* last (or only) page */ |
569 | last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL; | |
589973a7 | 570 | p_addr = __bm_map_pidx(b, idx); |
5fb3bc4d | 571 | bits += bitmap_weight(p_addr, last_word * BITS_PER_LONG); |
4b0715f0 LE |
572 | p_addr[last_word] &= cpu_to_lel(mask); |
573 | bits += hweight_long(p_addr[last_word]); | |
574 | /* 32bit arch, may have an unused padding long */ | |
575 | if (BITS_PER_LONG == 32 && (last_word & 1) == 0) | |
576 | p_addr[last_word+1] = 0; | |
589973a7 | 577 | __bm_unmap(p_addr); |
b411b363 PR |
578 | return bits; |
579 | } | |
580 | ||
b411b363 PR |
581 | /* offset and len in long words.*/ |
582 | static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len) | |
583 | { | |
584 | unsigned long *p_addr, *bm; | |
19f843aa | 585 | unsigned int idx; |
b411b363 PR |
586 | size_t do_now, end; |
587 | ||
b411b363 PR |
588 | end = offset + len; |
589 | ||
590 | if (end > b->bm_words) { | |
f88c5d90 | 591 | pr_alert("bm_memset end > bm_words\n"); |
b411b363 PR |
592 | return; |
593 | } | |
594 | ||
595 | while (offset < end) { | |
596 | do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset; | |
19f843aa LE |
597 | idx = bm_word_to_page_idx(b, offset); |
598 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
599 | bm = p_addr + MLPP(offset); |
600 | if (bm+do_now > p_addr + LWPP) { | |
f88c5d90 | 601 | pr_alert("BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n", |
b411b363 | 602 | p_addr, bm, (int)do_now); |
84e7c0f7 LE |
603 | } else |
604 | memset(bm, c, do_now * sizeof(long)); | |
b411b363 | 605 | bm_unmap(p_addr); |
19f843aa | 606 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
607 | offset += do_now; |
608 | } | |
609 | } | |
610 | ||
ae8bf312 LE |
611 | /* For the layout, see comment above drbd_md_set_sector_offsets(). */ |
612 | static u64 drbd_md_on_disk_bits(struct drbd_backing_dev *ldev) | |
613 | { | |
614 | u64 bitmap_sectors; | |
615 | if (ldev->md.al_offset == 8) | |
616 | bitmap_sectors = ldev->md.md_size_sect - ldev->md.bm_offset; | |
617 | else | |
618 | bitmap_sectors = ldev->md.al_offset - ldev->md.bm_offset; | |
619 | return bitmap_sectors << (9 + 3); | |
620 | } | |
621 | ||
b411b363 PR |
622 | /* |
623 | * make sure the bitmap has enough room for the attached storage, | |
624 | * if necessary, resize. | |
625 | * called whenever we may have changed the device size. | |
626 | * returns -ENOMEM if we could not allocate enough memory, 0 on success. | |
627 | * In case this is actually a resize, we copy the old bitmap into the new one. | |
628 | * Otherwise, the bitmap is initialized to all bits set. | |
629 | */ | |
b30ab791 | 630 | int drbd_bm_resize(struct drbd_device *device, sector_t capacity, int set_new_bits) |
b411b363 | 631 | { |
b30ab791 | 632 | struct drbd_bitmap *b = device->bitmap; |
6850c442 | 633 | unsigned long bits, words, owords, obits; |
b411b363 PR |
634 | unsigned long want, have, onpages; /* number of pages */ |
635 | struct page **npages, **opages = NULL; | |
636 | int err = 0, growing; | |
b411b363 | 637 | |
841ce241 AG |
638 | if (!expect(b)) |
639 | return -ENOMEM; | |
b411b363 | 640 | |
b30ab791 | 641 | drbd_bm_lock(device, "resize", BM_LOCKED_MASK); |
b411b363 | 642 | |
d0180171 | 643 | drbd_info(device, "drbd_bm_resize called with capacity == %llu\n", |
b411b363 PR |
644 | (unsigned long long)capacity); |
645 | ||
646 | if (capacity == b->bm_dev_capacity) | |
647 | goto out; | |
648 | ||
b411b363 PR |
649 | if (capacity == 0) { |
650 | spin_lock_irq(&b->bm_lock); | |
651 | opages = b->bm_pages; | |
652 | onpages = b->bm_number_of_pages; | |
653 | owords = b->bm_words; | |
654 | b->bm_pages = NULL; | |
655 | b->bm_number_of_pages = | |
656 | b->bm_set = | |
657 | b->bm_bits = | |
658 | b->bm_words = | |
659 | b->bm_dev_capacity = 0; | |
660 | spin_unlock_irq(&b->bm_lock); | |
661 | bm_free_pages(opages, onpages); | |
1d5cfdb0 | 662 | bm_vk_free(opages); |
b411b363 PR |
663 | goto out; |
664 | } | |
665 | bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT)); | |
666 | ||
667 | /* if we would use | |
668 | words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL; | |
669 | a 32bit host could present the wrong number of words | |
670 | to a 64bit host. | |
671 | */ | |
672 | words = ALIGN(bits, 64) >> LN2_BPL; | |
673 | ||
b30ab791 AG |
674 | if (get_ldev(device)) { |
675 | u64 bits_on_disk = drbd_md_on_disk_bits(device->ldev); | |
676 | put_ldev(device); | |
4b0715f0 | 677 | if (bits > bits_on_disk) { |
d0180171 AG |
678 | drbd_info(device, "bits = %lu\n", bits); |
679 | drbd_info(device, "bits_on_disk = %llu\n", bits_on_disk); | |
4b0715f0 LE |
680 | err = -ENOSPC; |
681 | goto out; | |
682 | } | |
b411b363 PR |
683 | } |
684 | ||
6850c442 | 685 | want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT; |
b411b363 PR |
686 | have = b->bm_number_of_pages; |
687 | if (want == have) { | |
0b0ba1ef | 688 | D_ASSERT(device, b->bm_pages != NULL); |
b411b363 PR |
689 | npages = b->bm_pages; |
690 | } else { | |
b30ab791 | 691 | if (drbd_insert_fault(device, DRBD_FAULT_BM_ALLOC)) |
b411b363 PR |
692 | npages = NULL; |
693 | else | |
694 | npages = bm_realloc_pages(b, want); | |
695 | } | |
696 | ||
697 | if (!npages) { | |
698 | err = -ENOMEM; | |
699 | goto out; | |
700 | } | |
701 | ||
702 | spin_lock_irq(&b->bm_lock); | |
703 | opages = b->bm_pages; | |
704 | owords = b->bm_words; | |
705 | obits = b->bm_bits; | |
706 | ||
707 | growing = bits > obits; | |
5223671b | 708 | if (opages && growing && set_new_bits) |
b411b363 PR |
709 | bm_set_surplus(b); |
710 | ||
711 | b->bm_pages = npages; | |
712 | b->bm_number_of_pages = want; | |
713 | b->bm_bits = bits; | |
714 | b->bm_words = words; | |
715 | b->bm_dev_capacity = capacity; | |
716 | ||
717 | if (growing) { | |
02d9a94b PR |
718 | if (set_new_bits) { |
719 | bm_memset(b, owords, 0xff, words-owords); | |
720 | b->bm_set += bits - obits; | |
721 | } else | |
722 | bm_memset(b, owords, 0x00, words-owords); | |
723 | ||
b411b363 PR |
724 | } |
725 | ||
726 | if (want < have) { | |
727 | /* implicit: (opages != NULL) && (opages != npages) */ | |
728 | bm_free_pages(opages + want, have - want); | |
729 | } | |
730 | ||
b411b363 PR |
731 | (void)bm_clear_surplus(b); |
732 | ||
733 | spin_unlock_irq(&b->bm_lock); | |
734 | if (opages != npages) | |
1d5cfdb0 | 735 | bm_vk_free(opages); |
b411b363 PR |
736 | if (!growing) |
737 | b->bm_set = bm_count_bits(b); | |
d0180171 | 738 | drbd_info(device, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want); |
b411b363 PR |
739 | |
740 | out: | |
b30ab791 | 741 | drbd_bm_unlock(device); |
b411b363 PR |
742 | return err; |
743 | } | |
744 | ||
745 | /* inherently racy: | |
746 | * if not protected by other means, return value may be out of date when | |
747 | * leaving this function... | |
748 | * we still need to lock it, since it is important that this returns | |
749 | * bm_set == 0 precisely. | |
750 | * | |
751 | * maybe bm_set should be atomic_t ? | |
752 | */ | |
b30ab791 | 753 | unsigned long _drbd_bm_total_weight(struct drbd_device *device) |
b411b363 | 754 | { |
b30ab791 | 755 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
756 | unsigned long s; |
757 | unsigned long flags; | |
758 | ||
841ce241 AG |
759 | if (!expect(b)) |
760 | return 0; | |
761 | if (!expect(b->bm_pages)) | |
762 | return 0; | |
b411b363 PR |
763 | |
764 | spin_lock_irqsave(&b->bm_lock, flags); | |
765 | s = b->bm_set; | |
766 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
767 | ||
768 | return s; | |
769 | } | |
770 | ||
b30ab791 | 771 | unsigned long drbd_bm_total_weight(struct drbd_device *device) |
b411b363 PR |
772 | { |
773 | unsigned long s; | |
774 | /* if I don't have a disk, I don't know about out-of-sync status */ | |
b30ab791 | 775 | if (!get_ldev_if_state(device, D_NEGOTIATING)) |
b411b363 | 776 | return 0; |
b30ab791 AG |
777 | s = _drbd_bm_total_weight(device); |
778 | put_ldev(device); | |
b411b363 PR |
779 | return s; |
780 | } | |
781 | ||
b30ab791 | 782 | size_t drbd_bm_words(struct drbd_device *device) |
b411b363 | 783 | { |
b30ab791 | 784 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
785 | if (!expect(b)) |
786 | return 0; | |
787 | if (!expect(b->bm_pages)) | |
788 | return 0; | |
b411b363 PR |
789 | |
790 | return b->bm_words; | |
791 | } | |
792 | ||
b30ab791 | 793 | unsigned long drbd_bm_bits(struct drbd_device *device) |
b411b363 | 794 | { |
b30ab791 | 795 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
796 | if (!expect(b)) |
797 | return 0; | |
b411b363 PR |
798 | |
799 | return b->bm_bits; | |
800 | } | |
801 | ||
802 | /* merge number words from buffer into the bitmap starting at offset. | |
803 | * buffer[i] is expected to be little endian unsigned long. | |
804 | * bitmap must be locked by drbd_bm_lock. | |
805 | * currently only used from receive_bitmap. | |
806 | */ | |
b30ab791 | 807 | void drbd_bm_merge_lel(struct drbd_device *device, size_t offset, size_t number, |
b411b363 PR |
808 | unsigned long *buffer) |
809 | { | |
b30ab791 | 810 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
811 | unsigned long *p_addr, *bm; |
812 | unsigned long word, bits; | |
19f843aa | 813 | unsigned int idx; |
b411b363 PR |
814 | size_t end, do_now; |
815 | ||
816 | end = offset + number; | |
817 | ||
841ce241 AG |
818 | if (!expect(b)) |
819 | return; | |
820 | if (!expect(b->bm_pages)) | |
821 | return; | |
b411b363 PR |
822 | if (number == 0) |
823 | return; | |
824 | WARN_ON(offset >= b->bm_words); | |
825 | WARN_ON(end > b->bm_words); | |
826 | ||
827 | spin_lock_irq(&b->bm_lock); | |
828 | while (offset < end) { | |
829 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa LE |
830 | idx = bm_word_to_page_idx(b, offset); |
831 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
832 | bm = p_addr + MLPP(offset); |
833 | offset += do_now; | |
834 | while (do_now--) { | |
835 | bits = hweight_long(*bm); | |
95a0f10c | 836 | word = *bm | *buffer++; |
b411b363 PR |
837 | *bm++ = word; |
838 | b->bm_set += hweight_long(word) - bits; | |
839 | } | |
840 | bm_unmap(p_addr); | |
19f843aa | 841 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
842 | } |
843 | /* with 32bit <-> 64bit cross-platform connect | |
844 | * this is only correct for current usage, | |
845 | * where we _know_ that we are 64 bit aligned, | |
846 | * and know that this function is used in this way, too... | |
847 | */ | |
848 | if (end == b->bm_words) | |
849 | b->bm_set -= bm_clear_surplus(b); | |
b411b363 PR |
850 | spin_unlock_irq(&b->bm_lock); |
851 | } | |
852 | ||
853 | /* copy number words from the bitmap starting at offset into the buffer. | |
854 | * buffer[i] will be little endian unsigned long. | |
855 | */ | |
b30ab791 | 856 | void drbd_bm_get_lel(struct drbd_device *device, size_t offset, size_t number, |
b411b363 PR |
857 | unsigned long *buffer) |
858 | { | |
b30ab791 | 859 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
860 | unsigned long *p_addr, *bm; |
861 | size_t end, do_now; | |
862 | ||
863 | end = offset + number; | |
864 | ||
841ce241 AG |
865 | if (!expect(b)) |
866 | return; | |
867 | if (!expect(b->bm_pages)) | |
868 | return; | |
b411b363 PR |
869 | |
870 | spin_lock_irq(&b->bm_lock); | |
871 | if ((offset >= b->bm_words) || | |
872 | (end > b->bm_words) || | |
873 | (number <= 0)) | |
d0180171 | 874 | drbd_err(device, "offset=%lu number=%lu bm_words=%lu\n", |
b411b363 PR |
875 | (unsigned long) offset, |
876 | (unsigned long) number, | |
877 | (unsigned long) b->bm_words); | |
878 | else { | |
879 | while (offset < end) { | |
880 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa | 881 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset)); |
b411b363 PR |
882 | bm = p_addr + MLPP(offset); |
883 | offset += do_now; | |
884 | while (do_now--) | |
95a0f10c | 885 | *buffer++ = *bm++; |
b411b363 PR |
886 | bm_unmap(p_addr); |
887 | } | |
888 | } | |
889 | spin_unlock_irq(&b->bm_lock); | |
890 | } | |
891 | ||
892 | /* set all bits in the bitmap */ | |
b30ab791 | 893 | void drbd_bm_set_all(struct drbd_device *device) |
b411b363 | 894 | { |
b30ab791 | 895 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
896 | if (!expect(b)) |
897 | return; | |
898 | if (!expect(b->bm_pages)) | |
899 | return; | |
b411b363 PR |
900 | |
901 | spin_lock_irq(&b->bm_lock); | |
902 | bm_memset(b, 0, 0xff, b->bm_words); | |
903 | (void)bm_clear_surplus(b); | |
904 | b->bm_set = b->bm_bits; | |
905 | spin_unlock_irq(&b->bm_lock); | |
906 | } | |
907 | ||
908 | /* clear all bits in the bitmap */ | |
b30ab791 | 909 | void drbd_bm_clear_all(struct drbd_device *device) |
b411b363 | 910 | { |
b30ab791 | 911 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
912 | if (!expect(b)) |
913 | return; | |
914 | if (!expect(b->bm_pages)) | |
915 | return; | |
b411b363 PR |
916 | |
917 | spin_lock_irq(&b->bm_lock); | |
918 | bm_memset(b, 0, 0, b->bm_words); | |
919 | b->bm_set = 0; | |
920 | spin_unlock_irq(&b->bm_lock); | |
921 | } | |
922 | ||
4ce49266 | 923 | static void drbd_bm_aio_ctx_destroy(struct kref *kref) |
d1f3779b | 924 | { |
4ce49266 LE |
925 | struct drbd_bm_aio_ctx *ctx = container_of(kref, struct drbd_bm_aio_ctx, kref); |
926 | unsigned long flags; | |
d1f3779b | 927 | |
4ce49266 LE |
928 | spin_lock_irqsave(&ctx->device->resource->req_lock, flags); |
929 | list_del(&ctx->list); | |
930 | spin_unlock_irqrestore(&ctx->device->resource->req_lock, flags); | |
b30ab791 | 931 | put_ldev(ctx->device); |
d1f3779b PR |
932 | kfree(ctx); |
933 | } | |
934 | ||
19f843aa | 935 | /* bv_page may be a copy, or may be the original */ |
4246a0b6 | 936 | static void drbd_bm_endio(struct bio *bio) |
b411b363 | 937 | { |
4ce49266 | 938 | struct drbd_bm_aio_ctx *ctx = bio->bi_private; |
b30ab791 AG |
939 | struct drbd_device *device = ctx->device; |
940 | struct drbd_bitmap *b = device->bitmap; | |
19f843aa | 941 | unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page); |
b411b363 | 942 | |
7648cdfe LE |
943 | if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 && |
944 | !bm_test_page_unchanged(b->bm_pages[idx])) | |
d0180171 | 945 | drbd_warn(device, "bitmap page idx %u changed during IO!\n", idx); |
19f843aa | 946 | |
4246a0b6 | 947 | if (bio->bi_error) { |
19f843aa LE |
948 | /* ctx error will hold the completed-last non-zero error code, |
949 | * in case error codes differ. */ | |
4246a0b6 | 950 | ctx->error = bio->bi_error; |
19f843aa LE |
951 | bm_set_page_io_err(b->bm_pages[idx]); |
952 | /* Not identical to on disk version of it. | |
953 | * Is BM_PAGE_IO_ERROR enough? */ | |
954 | if (__ratelimit(&drbd_ratelimit_state)) | |
d0180171 | 955 | drbd_err(device, "IO ERROR %d on bitmap page idx %u\n", |
4246a0b6 | 956 | bio->bi_error, idx); |
19f843aa LE |
957 | } else { |
958 | bm_clear_page_io_err(b->bm_pages[idx]); | |
d0180171 | 959 | dynamic_drbd_dbg(device, "bitmap page idx %u completed\n", idx); |
b411b363 | 960 | } |
19f843aa | 961 | |
b30ab791 | 962 | bm_page_unlock_io(device, idx); |
19f843aa | 963 | |
19f843aa | 964 | if (ctx->flags & BM_AIO_COPY_PAGES) |
4d95a10f | 965 | mempool_free(bio->bi_io_vec[0].bv_page, drbd_md_io_page_pool); |
b411b363 PR |
966 | |
967 | bio_put(bio); | |
19f843aa | 968 | |
d1f3779b | 969 | if (atomic_dec_and_test(&ctx->in_flight)) { |
9e58c4da | 970 | ctx->done = 1; |
b30ab791 | 971 | wake_up(&device->misc_wait); |
4ce49266 | 972 | kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); |
d1f3779b | 973 | } |
b411b363 PR |
974 | } |
975 | ||
4ce49266 | 976 | static void bm_page_io_async(struct drbd_bm_aio_ctx *ctx, int page_nr) __must_hold(local) |
b411b363 | 977 | { |
9476f39d | 978 | struct bio *bio = bio_alloc_drbd(GFP_NOIO); |
b30ab791 AG |
979 | struct drbd_device *device = ctx->device; |
980 | struct drbd_bitmap *b = device->bitmap; | |
19f843aa | 981 | struct page *page; |
b411b363 | 982 | unsigned int len; |
4ce49266 | 983 | unsigned int rw = (ctx->flags & BM_AIO_READ) ? READ : WRITE; |
19f843aa | 984 | |
b411b363 | 985 | sector_t on_disk_sector = |
b30ab791 | 986 | device->ldev->md.md_offset + device->ldev->md.bm_offset; |
b411b363 PR |
987 | on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9); |
988 | ||
989 | /* this might happen with very small | |
19f843aa LE |
990 | * flexible external meta data device, |
991 | * or with PAGE_SIZE > 4k */ | |
b411b363 | 992 | len = min_t(unsigned int, PAGE_SIZE, |
b30ab791 | 993 | (drbd_md_last_sector(device->ldev) - on_disk_sector + 1)<<9); |
b411b363 | 994 | |
19f843aa | 995 | /* serialize IO on this page */ |
b30ab791 | 996 | bm_page_lock_io(device, page_nr); |
19f843aa LE |
997 | /* before memcpy and submit, |
998 | * so it can be redirtied any time */ | |
999 | bm_set_page_unchanged(b->bm_pages[page_nr]); | |
1000 | ||
1001 | if (ctx->flags & BM_AIO_COPY_PAGES) { | |
71baba4b | 1002 | page = mempool_alloc(drbd_md_io_page_pool, __GFP_HIGHMEM|__GFP_RECLAIM); |
f1d6a328 | 1003 | copy_highpage(page, b->bm_pages[page_nr]); |
19f843aa LE |
1004 | bm_store_page_idx(page, page_nr); |
1005 | } else | |
1006 | page = b->bm_pages[page_nr]; | |
b30ab791 | 1007 | bio->bi_bdev = device->ldev->md_bdev; |
4f024f37 | 1008 | bio->bi_iter.bi_sector = on_disk_sector; |
4d95a10f LE |
1009 | /* bio_add_page of a single page to an empty bio will always succeed, |
1010 | * according to api. Do we want to assert that? */ | |
19f843aa LE |
1011 | bio_add_page(bio, page, len, 0); |
1012 | bio->bi_private = ctx; | |
ed15b795 | 1013 | bio->bi_end_io = drbd_bm_endio; |
b411b363 | 1014 | |
b30ab791 | 1015 | if (drbd_insert_fault(device, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) { |
b411b363 | 1016 | bio->bi_rw |= rw; |
4246a0b6 | 1017 | bio_io_error(bio); |
b411b363 PR |
1018 | } else { |
1019 | submit_bio(rw, bio); | |
5a8b4242 LE |
1020 | /* this should not count as user activity and cause the |
1021 | * resync to throttle -- see drbd_rs_should_slow_down(). */ | |
b30ab791 | 1022 | atomic_add(len >> 9, &device->rs_sect_ev); |
b411b363 PR |
1023 | } |
1024 | } | |
1025 | ||
b411b363 PR |
1026 | /* |
1027 | * bm_rw: read/write the whole bitmap from/to its on disk location. | |
1028 | */ | |
4ce49266 | 1029 | static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned lazy_writeout_upper_idx) __must_hold(local) |
b411b363 | 1030 | { |
4ce49266 | 1031 | struct drbd_bm_aio_ctx *ctx; |
b30ab791 | 1032 | struct drbd_bitmap *b = device->bitmap; |
6850c442 | 1033 | int num_pages, i, count = 0; |
b411b363 PR |
1034 | unsigned long now; |
1035 | char ppb[10]; | |
1036 | int err = 0; | |
1037 | ||
19f843aa LE |
1038 | /* |
1039 | * We are protected against bitmap disappearing/resizing by holding an | |
1040 | * ldev reference (caller must have called get_ldev()). | |
1041 | * For read/write, we are protected against changes to the bitmap by | |
1042 | * the bitmap lock (see drbd_bitmap_io). | |
1043 | * For lazy writeout, we don't care for ongoing changes to the bitmap, | |
1044 | * as we submit copies of pages anyways. | |
1045 | */ | |
d1f3779b | 1046 | |
4ce49266 | 1047 | ctx = kmalloc(sizeof(struct drbd_bm_aio_ctx), GFP_NOIO); |
d1f3779b PR |
1048 | if (!ctx) |
1049 | return -ENOMEM; | |
1050 | ||
4ce49266 | 1051 | *ctx = (struct drbd_bm_aio_ctx) { |
b30ab791 | 1052 | .device = device, |
4ce49266 | 1053 | .start_jif = jiffies, |
d1f3779b | 1054 | .in_flight = ATOMIC_INIT(1), |
9e58c4da | 1055 | .done = 0, |
0e8488ad | 1056 | .flags = flags, |
d1f3779b PR |
1057 | .error = 0, |
1058 | .kref = { ATOMIC_INIT(2) }, | |
1059 | }; | |
1060 | ||
4ce49266 | 1061 | if (!get_ldev_if_state(device, D_ATTACHING)) { /* put is in drbd_bm_aio_ctx_destroy() */ |
d0180171 | 1062 | drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n"); |
9e58c4da PR |
1063 | kfree(ctx); |
1064 | return -ENODEV; | |
1065 | } | |
8fe39aac PR |
1066 | /* Here D_ATTACHING is sufficient since drbd_bm_read() is called only from |
1067 | drbd_adm_attach(), after device->ldev was assigned. */ | |
9e58c4da | 1068 | |
4ce49266 | 1069 | if (0 == (ctx->flags & ~BM_AIO_READ)) |
20ceb2b2 | 1070 | WARN_ON(!(BM_LOCKED_MASK & b->bm_flags)); |
b411b363 | 1071 | |
4ce49266 LE |
1072 | spin_lock_irq(&device->resource->req_lock); |
1073 | list_add_tail(&ctx->list, &device->pending_bitmap_io); | |
1074 | spin_unlock_irq(&device->resource->req_lock); | |
1075 | ||
6850c442 | 1076 | num_pages = b->bm_number_of_pages; |
b411b363 | 1077 | |
b411b363 | 1078 | now = jiffies; |
b411b363 PR |
1079 | |
1080 | /* let the layers below us try to merge these bios... */ | |
6850c442 | 1081 | for (i = 0; i < num_pages; i++) { |
19f843aa LE |
1082 | /* ignore completely unchanged pages */ |
1083 | if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx) | |
1084 | break; | |
4ce49266 | 1085 | if (!(flags & BM_AIO_READ)) { |
45dfffeb LE |
1086 | if ((flags & BM_AIO_WRITE_HINTED) && |
1087 | !test_and_clear_bit(BM_PAGE_HINT_WRITEOUT, | |
1088 | &page_private(b->bm_pages[i]))) | |
1089 | continue; | |
fef45d29 | 1090 | |
4ce49266 | 1091 | if (!(flags & BM_AIO_WRITE_ALL_PAGES) && |
d1aa4d04 | 1092 | bm_test_page_unchanged(b->bm_pages[i])) { |
d0180171 | 1093 | dynamic_drbd_dbg(device, "skipped bm write for idx %u\n", i); |
19f843aa LE |
1094 | continue; |
1095 | } | |
1096 | /* during lazy writeout, | |
1097 | * ignore those pages not marked for lazy writeout. */ | |
1098 | if (lazy_writeout_upper_idx && | |
1099 | !bm_test_page_lazy_writeout(b->bm_pages[i])) { | |
d0180171 | 1100 | dynamic_drbd_dbg(device, "skipped bm lazy write for idx %u\n", i); |
19f843aa LE |
1101 | continue; |
1102 | } | |
1103 | } | |
d1f3779b | 1104 | atomic_inc(&ctx->in_flight); |
4ce49266 | 1105 | bm_page_io_async(ctx, i); |
19f843aa LE |
1106 | ++count; |
1107 | cond_resched(); | |
1108 | } | |
b411b363 | 1109 | |
725a97e4 | 1110 | /* |
ed15b795 | 1111 | * We initialize ctx->in_flight to one to make sure drbd_bm_endio |
9e58c4da | 1112 | * will not set ctx->done early, and decrement / test it here. If there |
725a97e4 | 1113 | * are still some bios in flight, we need to wait for them here. |
9e58c4da PR |
1114 | * If all IO is done already (or nothing had been submitted), there is |
1115 | * no need to wait. Still, we need to put the kref associated with the | |
1116 | * "in_flight reached zero, all done" event. | |
725a97e4 | 1117 | */ |
d1f3779b | 1118 | if (!atomic_dec_and_test(&ctx->in_flight)) |
b30ab791 | 1119 | wait_until_done_or_force_detached(device, device->ldev, &ctx->done); |
9e58c4da | 1120 | else |
4ce49266 | 1121 | kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); |
d1f3779b | 1122 | |
c9d963a4 LE |
1123 | /* summary for global bitmap IO */ |
1124 | if (flags == 0) | |
d0180171 | 1125 | drbd_info(device, "bitmap %s of %u pages took %lu jiffies\n", |
4ce49266 | 1126 | (flags & BM_AIO_READ) ? "READ" : "WRITE", |
cdfda633 | 1127 | count, jiffies - now); |
b411b363 | 1128 | |
d1f3779b | 1129 | if (ctx->error) { |
d0180171 | 1130 | drbd_alert(device, "we had at least one MD IO ERROR during bitmap IO\n"); |
b30ab791 | 1131 | drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR); |
d1f3779b | 1132 | err = -EIO; /* ctx->error ? */ |
b411b363 PR |
1133 | } |
1134 | ||
9e58c4da | 1135 | if (atomic_read(&ctx->in_flight)) |
44edfb0d | 1136 | err = -EIO; /* Disk timeout/force-detach during IO... */ |
9e58c4da | 1137 | |
b411b363 | 1138 | now = jiffies; |
4ce49266 | 1139 | if (flags & BM_AIO_READ) { |
95a0f10c | 1140 | b->bm_set = bm_count_bits(b); |
d0180171 | 1141 | drbd_info(device, "recounting of set bits took additional %lu jiffies\n", |
b411b363 PR |
1142 | jiffies - now); |
1143 | } | |
1144 | now = b->bm_set; | |
1145 | ||
4ce49266 | 1146 | if ((flags & ~BM_AIO_READ) == 0) |
d0180171 | 1147 | drbd_info(device, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n", |
c9d963a4 | 1148 | ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now); |
b411b363 | 1149 | |
4ce49266 | 1150 | kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); |
b411b363 PR |
1151 | return err; |
1152 | } | |
1153 | ||
1154 | /** | |
1155 | * drbd_bm_read() - Read the whole bitmap from its on disk location. | |
b30ab791 | 1156 | * @device: DRBD device. |
b411b363 | 1157 | */ |
b30ab791 | 1158 | int drbd_bm_read(struct drbd_device *device) __must_hold(local) |
b411b363 | 1159 | { |
4ce49266 | 1160 | return bm_rw(device, BM_AIO_READ, 0); |
b411b363 PR |
1161 | } |
1162 | ||
1163 | /** | |
1164 | * drbd_bm_write() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1165 | * @device: DRBD device. |
19f843aa LE |
1166 | * |
1167 | * Will only write pages that have changed since last IO. | |
b411b363 | 1168 | */ |
b30ab791 | 1169 | int drbd_bm_write(struct drbd_device *device) __must_hold(local) |
b411b363 | 1170 | { |
4ce49266 | 1171 | return bm_rw(device, 0, 0); |
b411b363 PR |
1172 | } |
1173 | ||
d1aa4d04 PR |
1174 | /** |
1175 | * drbd_bm_write_all() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1176 | * @device: DRBD device. |
d1aa4d04 PR |
1177 | * |
1178 | * Will write all pages. | |
1179 | */ | |
b30ab791 | 1180 | int drbd_bm_write_all(struct drbd_device *device) __must_hold(local) |
d1aa4d04 | 1181 | { |
4ce49266 | 1182 | return bm_rw(device, BM_AIO_WRITE_ALL_PAGES, 0); |
d1aa4d04 PR |
1183 | } |
1184 | ||
c7a58db4 LE |
1185 | /** |
1186 | * drbd_bm_write_lazy() - Write bitmap pages 0 to @upper_idx-1, if they have changed. | |
1187 | * @device: DRBD device. | |
1188 | * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages | |
1189 | */ | |
1190 | int drbd_bm_write_lazy(struct drbd_device *device, unsigned upper_idx) __must_hold(local) | |
1191 | { | |
1192 | return bm_rw(device, BM_AIO_COPY_PAGES, upper_idx); | |
1193 | } | |
1194 | ||
0e8488ad LE |
1195 | /** |
1196 | * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1197 | * @device: DRBD device. |
0e8488ad LE |
1198 | * |
1199 | * Will only write pages that have changed since last IO. | |
1200 | * In contrast to drbd_bm_write(), this will copy the bitmap pages | |
1201 | * to temporary writeout pages. It is intended to trigger a full write-out | |
1202 | * while still allowing the bitmap to change, for example if a resync or online | |
1203 | * verify is aborted due to a failed peer disk, while local IO continues, or | |
1204 | * pending resync acks are still being processed. | |
1205 | */ | |
b30ab791 | 1206 | int drbd_bm_write_copy_pages(struct drbd_device *device) __must_hold(local) |
0e8488ad | 1207 | { |
4ce49266 | 1208 | return bm_rw(device, BM_AIO_COPY_PAGES, 0); |
19f843aa LE |
1209 | } |
1210 | ||
45dfffeb LE |
1211 | /** |
1212 | * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed. | |
b30ab791 | 1213 | * @device: DRBD device. |
45dfffeb | 1214 | */ |
b30ab791 | 1215 | int drbd_bm_write_hinted(struct drbd_device *device) __must_hold(local) |
45dfffeb | 1216 | { |
4ce49266 | 1217 | return bm_rw(device, BM_AIO_WRITE_HINTED | BM_AIO_COPY_PAGES, 0); |
45dfffeb | 1218 | } |
19f843aa | 1219 | |
b411b363 PR |
1220 | /* NOTE |
1221 | * find_first_bit returns int, we return unsigned long. | |
4b0715f0 LE |
1222 | * For this to work on 32bit arch with bitnumbers > (1<<32), |
1223 | * we'd need to return u64, and get a whole lot of other places | |
1224 | * fixed where we still use unsigned long. | |
b411b363 PR |
1225 | * |
1226 | * this returns a bit number, NOT a sector! | |
1227 | */ | |
b30ab791 | 1228 | static unsigned long __bm_find_next(struct drbd_device *device, unsigned long bm_fo, |
cfd8005c | 1229 | const int find_zero_bit) |
b411b363 | 1230 | { |
b30ab791 | 1231 | struct drbd_bitmap *b = device->bitmap; |
b411b363 | 1232 | unsigned long *p_addr; |
4b0715f0 LE |
1233 | unsigned long bit_offset; |
1234 | unsigned i; | |
1235 | ||
b411b363 PR |
1236 | |
1237 | if (bm_fo > b->bm_bits) { | |
d0180171 | 1238 | drbd_err(device, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits); |
4b0715f0 | 1239 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1240 | } else { |
1241 | while (bm_fo < b->bm_bits) { | |
19f843aa | 1242 | /* bit offset of the first bit in the page */ |
4b0715f0 | 1243 | bit_offset = bm_fo & ~BITS_PER_PAGE_MASK; |
cfd8005c | 1244 | p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo)); |
b411b363 PR |
1245 | |
1246 | if (find_zero_bit) | |
7e599e6e | 1247 | i = find_next_zero_bit_le(p_addr, |
4b0715f0 | 1248 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1249 | else |
7e599e6e | 1250 | i = find_next_bit_le(p_addr, |
4b0715f0 | 1251 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1252 | |
cfd8005c | 1253 | __bm_unmap(p_addr); |
b411b363 | 1254 | if (i < PAGE_SIZE*8) { |
4b0715f0 LE |
1255 | bm_fo = bit_offset + i; |
1256 | if (bm_fo >= b->bm_bits) | |
b411b363 PR |
1257 | break; |
1258 | goto found; | |
1259 | } | |
1260 | bm_fo = bit_offset + PAGE_SIZE*8; | |
1261 | } | |
4b0715f0 | 1262 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1263 | } |
1264 | found: | |
4b0715f0 | 1265 | return bm_fo; |
b411b363 PR |
1266 | } |
1267 | ||
b30ab791 | 1268 | static unsigned long bm_find_next(struct drbd_device *device, |
b411b363 PR |
1269 | unsigned long bm_fo, const int find_zero_bit) |
1270 | { | |
b30ab791 | 1271 | struct drbd_bitmap *b = device->bitmap; |
4b0715f0 | 1272 | unsigned long i = DRBD_END_OF_BITMAP; |
b411b363 | 1273 | |
841ce241 AG |
1274 | if (!expect(b)) |
1275 | return i; | |
1276 | if (!expect(b->bm_pages)) | |
1277 | return i; | |
b411b363 PR |
1278 | |
1279 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1280 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1281 | bm_print_lock_info(device); |
b411b363 | 1282 | |
b30ab791 | 1283 | i = __bm_find_next(device, bm_fo, find_zero_bit); |
b411b363 PR |
1284 | |
1285 | spin_unlock_irq(&b->bm_lock); | |
1286 | return i; | |
1287 | } | |
1288 | ||
b30ab791 | 1289 | unsigned long drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1290 | { |
b30ab791 | 1291 | return bm_find_next(device, bm_fo, 0); |
b411b363 PR |
1292 | } |
1293 | ||
1294 | #if 0 | |
1295 | /* not yet needed for anything. */ | |
b30ab791 | 1296 | unsigned long drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1297 | { |
b30ab791 | 1298 | return bm_find_next(device, bm_fo, 1); |
b411b363 PR |
1299 | } |
1300 | #endif | |
1301 | ||
1302 | /* does not spin_lock_irqsave. | |
1303 | * you must take drbd_bm_lock() first */ | |
b30ab791 | 1304 | unsigned long _drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1305 | { |
b30ab791 AG |
1306 | /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ |
1307 | return __bm_find_next(device, bm_fo, 0); | |
b411b363 PR |
1308 | } |
1309 | ||
b30ab791 | 1310 | unsigned long _drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1311 | { |
b30ab791 AG |
1312 | /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ |
1313 | return __bm_find_next(device, bm_fo, 1); | |
b411b363 PR |
1314 | } |
1315 | ||
1316 | /* returns number of bits actually changed. | |
1317 | * for val != 0, we change 0 -> 1, return code positive | |
1318 | * for val == 0, we change 1 -> 0, return code negative | |
1319 | * wants bitnr, not sector. | |
1320 | * expected to be called for only a few bits (e - s about BITS_PER_LONG). | |
1321 | * Must hold bitmap lock already. */ | |
b30ab791 | 1322 | static int __bm_change_bits_to(struct drbd_device *device, const unsigned long s, |
829c6087 | 1323 | unsigned long e, int val) |
b411b363 | 1324 | { |
b30ab791 | 1325 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1326 | unsigned long *p_addr = NULL; |
1327 | unsigned long bitnr; | |
19f843aa | 1328 | unsigned int last_page_nr = -1U; |
b411b363 | 1329 | int c = 0; |
19f843aa | 1330 | int changed_total = 0; |
b411b363 PR |
1331 | |
1332 | if (e >= b->bm_bits) { | |
d0180171 | 1333 | drbd_err(device, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n", |
b411b363 PR |
1334 | s, e, b->bm_bits); |
1335 | e = b->bm_bits ? b->bm_bits -1 : 0; | |
1336 | } | |
1337 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa | 1338 | unsigned int page_nr = bm_bit_to_page_idx(b, bitnr); |
b411b363 PR |
1339 | if (page_nr != last_page_nr) { |
1340 | if (p_addr) | |
cfd8005c | 1341 | __bm_unmap(p_addr); |
19f843aa LE |
1342 | if (c < 0) |
1343 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1344 | else if (c > 0) | |
1345 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1346 | changed_total += c; | |
1347 | c = 0; | |
cfd8005c | 1348 | p_addr = __bm_map_pidx(b, page_nr); |
b411b363 PR |
1349 | last_page_nr = page_nr; |
1350 | } | |
1351 | if (val) | |
7e599e6e | 1352 | c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 | 1353 | else |
7e599e6e | 1354 | c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 PR |
1355 | } |
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 | b->bm_set += changed_total; | |
1364 | return changed_total; | |
b411b363 PR |
1365 | } |
1366 | ||
1367 | /* returns number of bits actually changed. | |
1368 | * for val != 0, we change 0 -> 1, return code positive | |
1369 | * for val == 0, we change 1 -> 0, return code negative | |
1370 | * wants bitnr, not sector */ | |
b30ab791 | 1371 | static int bm_change_bits_to(struct drbd_device *device, const unsigned long s, |
b411b363 PR |
1372 | const unsigned long e, int val) |
1373 | { | |
1374 | unsigned long flags; | |
b30ab791 | 1375 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1376 | int c = 0; |
1377 | ||
841ce241 AG |
1378 | if (!expect(b)) |
1379 | return 1; | |
1380 | if (!expect(b->bm_pages)) | |
1381 | return 0; | |
b411b363 PR |
1382 | |
1383 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1384 | if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags) |
b30ab791 | 1385 | bm_print_lock_info(device); |
b411b363 | 1386 | |
b30ab791 | 1387 | c = __bm_change_bits_to(device, s, e, val); |
b411b363 PR |
1388 | |
1389 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1390 | return c; | |
1391 | } | |
1392 | ||
1393 | /* returns number of bits changed 0 -> 1 */ | |
b30ab791 | 1394 | int drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 | 1395 | { |
b30ab791 | 1396 | return bm_change_bits_to(device, s, e, 1); |
b411b363 PR |
1397 | } |
1398 | ||
1399 | /* returns number of bits changed 1 -> 0 */ | |
b30ab791 | 1400 | int drbd_bm_clear_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 | 1401 | { |
b30ab791 | 1402 | return -bm_change_bits_to(device, s, e, 0); |
b411b363 PR |
1403 | } |
1404 | ||
1405 | /* sets all bits in full words, | |
1406 | * from first_word up to, but not including, last_word */ | |
1407 | static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b, | |
1408 | int page_nr, int first_word, int last_word) | |
1409 | { | |
1410 | int i; | |
1411 | int bits; | |
22d81140 | 1412 | int changed = 0; |
cfd8005c | 1413 | unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr]); |
5fb3bc4d LE |
1414 | |
1415 | /* I think it is more cache line friendly to hweight_long then set to ~0UL, | |
1416 | * than to first bitmap_weight() all words, then bitmap_fill() all words */ | |
b411b363 PR |
1417 | for (i = first_word; i < last_word; i++) { |
1418 | bits = hweight_long(paddr[i]); | |
1419 | paddr[i] = ~0UL; | |
22d81140 | 1420 | changed += BITS_PER_LONG - bits; |
b411b363 | 1421 | } |
cfd8005c | 1422 | kunmap_atomic(paddr); |
22d81140 LE |
1423 | if (changed) { |
1424 | /* We only need lazy writeout, the information is still in the | |
1425 | * remote bitmap as well, and is reconstructed during the next | |
1426 | * bitmap exchange, if lost locally due to a crash. */ | |
1427 | bm_set_page_lazy_writeout(b->bm_pages[page_nr]); | |
1428 | b->bm_set += changed; | |
1429 | } | |
b411b363 PR |
1430 | } |
1431 | ||
829c6087 LE |
1432 | /* Same thing as drbd_bm_set_bits, |
1433 | * but more efficient for a large bit range. | |
b411b363 PR |
1434 | * You must first drbd_bm_lock(). |
1435 | * Can be called to set the whole bitmap in one go. | |
1436 | * Sets bits from s to e _inclusive_. */ | |
b30ab791 | 1437 | void _drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 PR |
1438 | { |
1439 | /* First set_bit from the first bit (s) | |
1440 | * up to the next long boundary (sl), | |
1441 | * then assign full words up to the last long boundary (el), | |
1442 | * then set_bit up to and including the last bit (e). | |
1443 | * | |
1444 | * Do not use memset, because we must account for changes, | |
1445 | * so we need to loop over the words with hweight() anyways. | |
1446 | */ | |
b30ab791 | 1447 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1448 | unsigned long sl = ALIGN(s,BITS_PER_LONG); |
1449 | unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1); | |
1450 | int first_page; | |
1451 | int last_page; | |
1452 | int page_nr; | |
1453 | int first_word; | |
1454 | int last_word; | |
1455 | ||
1456 | if (e - s <= 3*BITS_PER_LONG) { | |
1457 | /* don't bother; el and sl may even be wrong. */ | |
829c6087 | 1458 | spin_lock_irq(&b->bm_lock); |
b30ab791 | 1459 | __bm_change_bits_to(device, s, e, 1); |
829c6087 | 1460 | spin_unlock_irq(&b->bm_lock); |
b411b363 PR |
1461 | return; |
1462 | } | |
1463 | ||
1464 | /* difference is large enough that we can trust sl and el */ | |
1465 | ||
829c6087 LE |
1466 | spin_lock_irq(&b->bm_lock); |
1467 | ||
b411b363 PR |
1468 | /* bits filling the current long */ |
1469 | if (sl) | |
b30ab791 | 1470 | __bm_change_bits_to(device, s, sl-1, 1); |
b411b363 PR |
1471 | |
1472 | first_page = sl >> (3 + PAGE_SHIFT); | |
1473 | last_page = el >> (3 + PAGE_SHIFT); | |
1474 | ||
1475 | /* MLPP: modulo longs per page */ | |
1476 | /* LWPP: long words per page */ | |
1477 | first_word = MLPP(sl >> LN2_BPL); | |
1478 | last_word = LWPP; | |
1479 | ||
1480 | /* first and full pages, unless first page == last page */ | |
1481 | for (page_nr = first_page; page_nr < last_page; page_nr++) { | |
b30ab791 | 1482 | bm_set_full_words_within_one_page(device->bitmap, page_nr, first_word, last_word); |
8ccee20e LE |
1483 | spin_unlock_irq(&b->bm_lock); |
1484 | cond_resched(); | |
b411b363 | 1485 | first_word = 0; |
8ccee20e | 1486 | spin_lock_irq(&b->bm_lock); |
b411b363 | 1487 | } |
b411b363 PR |
1488 | /* last page (respectively only page, for first page == last page) */ |
1489 | last_word = MLPP(el >> LN2_BPL); | |
4eccc579 LE |
1490 | |
1491 | /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples). | |
1492 | * ==> e = 32767, el = 32768, last_page = 2, | |
1493 | * and now last_word = 0. | |
1494 | * We do not want to touch last_page in this case, | |
1495 | * as we did not allocate it, it is not present in bitmap->bm_pages. | |
1496 | */ | |
1497 | if (last_word) | |
b30ab791 | 1498 | bm_set_full_words_within_one_page(device->bitmap, last_page, first_word, last_word); |
b411b363 PR |
1499 | |
1500 | /* possibly trailing bits. | |
1501 | * example: (e & 63) == 63, el will be e+1. | |
1502 | * if that even was the very last bit, | |
1503 | * it would trigger an assert in __bm_change_bits_to() | |
1504 | */ | |
1505 | if (el <= e) | |
b30ab791 | 1506 | __bm_change_bits_to(device, el, e, 1); |
829c6087 | 1507 | spin_unlock_irq(&b->bm_lock); |
b411b363 PR |
1508 | } |
1509 | ||
1510 | /* returns bit state | |
1511 | * wants bitnr, NOT sector. | |
1512 | * inherently racy... area needs to be locked by means of {al,rs}_lru | |
1513 | * 1 ... bit set | |
1514 | * 0 ... bit not set | |
1515 | * -1 ... first out of bounds access, stop testing for bits! | |
1516 | */ | |
b30ab791 | 1517 | int drbd_bm_test_bit(struct drbd_device *device, const unsigned long bitnr) |
b411b363 PR |
1518 | { |
1519 | unsigned long flags; | |
b30ab791 | 1520 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1521 | unsigned long *p_addr; |
1522 | int i; | |
1523 | ||
841ce241 AG |
1524 | if (!expect(b)) |
1525 | return 0; | |
1526 | if (!expect(b->bm_pages)) | |
1527 | return 0; | |
b411b363 PR |
1528 | |
1529 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1530 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1531 | bm_print_lock_info(device); |
b411b363 | 1532 | if (bitnr < b->bm_bits) { |
19f843aa | 1533 | p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr)); |
7e599e6e | 1534 | i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0; |
b411b363 PR |
1535 | bm_unmap(p_addr); |
1536 | } else if (bitnr == b->bm_bits) { | |
1537 | i = -1; | |
1538 | } else { /* (bitnr > b->bm_bits) */ | |
d0180171 | 1539 | drbd_err(device, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits); |
b411b363 PR |
1540 | i = 0; |
1541 | } | |
1542 | ||
1543 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1544 | return i; | |
1545 | } | |
1546 | ||
1547 | /* returns number of bits set in the range [s, e] */ | |
b30ab791 | 1548 | int drbd_bm_count_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 PR |
1549 | { |
1550 | unsigned long flags; | |
b30ab791 | 1551 | struct drbd_bitmap *b = device->bitmap; |
19f843aa | 1552 | unsigned long *p_addr = NULL; |
b411b363 | 1553 | unsigned long bitnr; |
19f843aa | 1554 | unsigned int page_nr = -1U; |
b411b363 | 1555 | int c = 0; |
b411b363 PR |
1556 | |
1557 | /* If this is called without a bitmap, that is a bug. But just to be | |
1558 | * robust in case we screwed up elsewhere, in that case pretend there | |
1559 | * was one dirty bit in the requested area, so we won't try to do a | |
1560 | * local read there (no bitmap probably implies no disk) */ | |
841ce241 AG |
1561 | if (!expect(b)) |
1562 | return 1; | |
1563 | if (!expect(b->bm_pages)) | |
1564 | return 1; | |
b411b363 PR |
1565 | |
1566 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1567 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1568 | bm_print_lock_info(device); |
b411b363 | 1569 | for (bitnr = s; bitnr <= e; bitnr++) { |
19f843aa LE |
1570 | unsigned int idx = bm_bit_to_page_idx(b, bitnr); |
1571 | if (page_nr != idx) { | |
1572 | page_nr = idx; | |
b411b363 PR |
1573 | if (p_addr) |
1574 | bm_unmap(p_addr); | |
19f843aa | 1575 | p_addr = bm_map_pidx(b, idx); |
b411b363 | 1576 | } |
841ce241 | 1577 | if (expect(bitnr < b->bm_bits)) |
7e599e6e | 1578 | c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr)); |
841ce241 | 1579 | else |
d0180171 | 1580 | drbd_err(device, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits); |
b411b363 PR |
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 | */ | |
b30ab791 | 1603 | int drbd_bm_e_weight(struct drbd_device *device, unsigned long enr) |
b411b363 | 1604 | { |
b30ab791 | 1605 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1606 | int count, s, e; |
1607 | unsigned long flags; | |
1608 | unsigned long *p_addr, *bm; | |
1609 | ||
841ce241 AG |
1610 | if (!expect(b)) |
1611 | return 0; | |
1612 | if (!expect(b->bm_pages)) | |
1613 | return 0; | |
b411b363 PR |
1614 | |
1615 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1616 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1617 | bm_print_lock_info(device); |
b411b363 PR |
1618 | |
1619 | s = S2W(enr); | |
1620 | e = min((size_t)S2W(enr+1), b->bm_words); | |
1621 | count = 0; | |
1622 | if (s < b->bm_words) { | |
1623 | int n = e-s; | |
19f843aa | 1624 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); |
b411b363 | 1625 | bm = p_addr + MLPP(s); |
5fb3bc4d | 1626 | count += bitmap_weight(bm, n * BITS_PER_LONG); |
b411b363 PR |
1627 | bm_unmap(p_addr); |
1628 | } else { | |
d0180171 | 1629 | drbd_err(device, "start offset (%d) too large in drbd_bm_e_weight\n", s); |
b411b363 PR |
1630 | } |
1631 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1632 | return count; | |
1633 | } |