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 | 115 | #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__) |
b30ab791 | 116 | static void __bm_print_lock_info(struct drbd_device *device, const char *func) |
b411b363 | 117 | { |
b30ab791 | 118 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
119 | if (!__ratelimit(&drbd_ratelimit_state)) |
120 | return; | |
c60b0251 AG |
121 | drbd_err(device, "FIXME %s[%d] in %s, bitmap locked for '%s' by %s[%d]\n", |
122 | current->comm, task_pid_nr(current), | |
123 | func, b->bm_why ?: "?", | |
124 | b->bm_task->comm, task_pid_nr(b->bm_task)); | |
b411b363 PR |
125 | } |
126 | ||
b30ab791 | 127 | void drbd_bm_lock(struct drbd_device *device, char *why, enum bm_flag flags) |
b411b363 | 128 | { |
b30ab791 | 129 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
130 | int trylock_failed; |
131 | ||
132 | if (!b) { | |
d0180171 | 133 | drbd_err(device, "FIXME no bitmap in drbd_bm_lock!?\n"); |
b411b363 PR |
134 | return; |
135 | } | |
136 | ||
8a03ae2a | 137 | trylock_failed = !mutex_trylock(&b->bm_change); |
b411b363 PR |
138 | |
139 | if (trylock_failed) { | |
c60b0251 AG |
140 | drbd_warn(device, "%s[%d] going to '%s' but bitmap already locked for '%s' by %s[%d]\n", |
141 | current->comm, task_pid_nr(current), | |
142 | why, b->bm_why ?: "?", | |
143 | b->bm_task->comm, task_pid_nr(b->bm_task)); | |
8a03ae2a | 144 | mutex_lock(&b->bm_change); |
b411b363 | 145 | } |
20ceb2b2 | 146 | if (BM_LOCKED_MASK & b->bm_flags) |
d0180171 | 147 | drbd_err(device, "FIXME bitmap already locked in bm_lock\n"); |
20ceb2b2 | 148 | b->bm_flags |= flags & BM_LOCKED_MASK; |
b411b363 PR |
149 | |
150 | b->bm_why = why; | |
151 | b->bm_task = current; | |
152 | } | |
153 | ||
b30ab791 | 154 | void drbd_bm_unlock(struct drbd_device *device) |
b411b363 | 155 | { |
b30ab791 | 156 | struct drbd_bitmap *b = device->bitmap; |
b411b363 | 157 | if (!b) { |
d0180171 | 158 | drbd_err(device, "FIXME no bitmap in drbd_bm_unlock!?\n"); |
b411b363 PR |
159 | return; |
160 | } | |
161 | ||
b30ab791 | 162 | if (!(BM_LOCKED_MASK & device->bitmap->bm_flags)) |
d0180171 | 163 | drbd_err(device, "FIXME bitmap not locked in bm_unlock\n"); |
b411b363 | 164 | |
20ceb2b2 | 165 | b->bm_flags &= ~BM_LOCKED_MASK; |
b411b363 PR |
166 | b->bm_why = NULL; |
167 | b->bm_task = NULL; | |
8a03ae2a | 168 | mutex_unlock(&b->bm_change); |
b411b363 PR |
169 | } |
170 | ||
19f843aa LE |
171 | /* we store some "meta" info about our pages in page->private */ |
172 | /* at a granularity of 4k storage per bitmap bit: | |
173 | * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks | |
174 | * 1<<38 bits, | |
175 | * 1<<23 4k bitmap pages. | |
176 | * Use 24 bits as page index, covers 2 peta byte storage | |
177 | * at a granularity of 4k per bit. | |
178 | * Used to report the failed page idx on io error from the endio handlers. | |
179 | */ | |
180 | #define BM_PAGE_IDX_MASK ((1UL<<24)-1) | |
181 | /* this page is currently read in, or written back */ | |
182 | #define BM_PAGE_IO_LOCK 31 | |
183 | /* if there has been an IO error for this page */ | |
184 | #define BM_PAGE_IO_ERROR 30 | |
185 | /* this is to be able to intelligently skip disk IO, | |
186 | * set if bits have been set since last IO. */ | |
187 | #define BM_PAGE_NEED_WRITEOUT 29 | |
188 | /* to mark for lazy writeout once syncer cleared all clearable bits, | |
189 | * we if bits have been cleared since last IO. */ | |
190 | #define BM_PAGE_LAZY_WRITEOUT 28 | |
45dfffeb LE |
191 | /* pages marked with this "HINT" will be considered for writeout |
192 | * on activity log transactions */ | |
193 | #define BM_PAGE_HINT_WRITEOUT 27 | |
19f843aa | 194 | |
24c4830c | 195 | /* store_page_idx uses non-atomic assignment. It is only used directly after |
19f843aa LE |
196 | * allocating the page. All other bm_set_page_* and bm_clear_page_* need to |
197 | * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap | |
198 | * changes) may happen from various contexts, and wait_on_bit/wake_up_bit | |
199 | * requires it all to be atomic as well. */ | |
200 | static void bm_store_page_idx(struct page *page, unsigned long idx) | |
95a0f10c | 201 | { |
19f843aa | 202 | BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK)); |
0c7db279 | 203 | set_page_private(page, idx); |
95a0f10c LE |
204 | } |
205 | ||
19f843aa | 206 | static unsigned long bm_page_to_idx(struct page *page) |
b411b363 | 207 | { |
19f843aa LE |
208 | return page_private(page) & BM_PAGE_IDX_MASK; |
209 | } | |
210 | ||
211 | /* As is very unlikely that the same page is under IO from more than one | |
212 | * context, we can get away with a bit per page and one wait queue per bitmap. | |
213 | */ | |
b30ab791 | 214 | static void bm_page_lock_io(struct drbd_device *device, int page_nr) |
19f843aa | 215 | { |
b30ab791 | 216 | struct drbd_bitmap *b = device->bitmap; |
19f843aa LE |
217 | void *addr = &page_private(b->bm_pages[page_nr]); |
218 | wait_event(b->bm_io_wait, !test_and_set_bit(BM_PAGE_IO_LOCK, addr)); | |
219 | } | |
220 | ||
b30ab791 | 221 | static void bm_page_unlock_io(struct drbd_device *device, int page_nr) |
19f843aa | 222 | { |
b30ab791 | 223 | struct drbd_bitmap *b = device->bitmap; |
19f843aa | 224 | void *addr = &page_private(b->bm_pages[page_nr]); |
4738fa16 | 225 | clear_bit_unlock(BM_PAGE_IO_LOCK, addr); |
b30ab791 | 226 | wake_up(&device->bitmap->bm_io_wait); |
19f843aa LE |
227 | } |
228 | ||
229 | /* set _before_ submit_io, so it may be reset due to being changed | |
230 | * while this page is in flight... will get submitted later again */ | |
231 | static void bm_set_page_unchanged(struct page *page) | |
232 | { | |
233 | /* use cmpxchg? */ | |
234 | clear_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
235 | clear_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
236 | } | |
237 | ||
238 | static void bm_set_page_need_writeout(struct page *page) | |
239 | { | |
240 | set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
241 | } | |
242 | ||
45dfffeb LE |
243 | /** |
244 | * drbd_bm_mark_for_writeout() - mark a page with a "hint" to be considered for writeout | |
b30ab791 | 245 | * @device: DRBD device. |
45dfffeb LE |
246 | * @page_nr: the bitmap page to mark with the "hint" flag |
247 | * | |
248 | * From within an activity log transaction, we mark a few pages with these | |
249 | * hints, then call drbd_bm_write_hinted(), which will only write out changed | |
250 | * pages which are flagged with this mark. | |
251 | */ | |
b30ab791 | 252 | void drbd_bm_mark_for_writeout(struct drbd_device *device, int page_nr) |
45dfffeb LE |
253 | { |
254 | struct page *page; | |
b30ab791 | 255 | if (page_nr >= device->bitmap->bm_number_of_pages) { |
d0180171 | 256 | drbd_warn(device, "BAD: page_nr: %u, number_of_pages: %u\n", |
b30ab791 | 257 | page_nr, (int)device->bitmap->bm_number_of_pages); |
45dfffeb LE |
258 | return; |
259 | } | |
b30ab791 | 260 | page = device->bitmap->bm_pages[page_nr]; |
45dfffeb LE |
261 | set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page)); |
262 | } | |
263 | ||
19f843aa LE |
264 | static int bm_test_page_unchanged(struct page *page) |
265 | { | |
266 | volatile const unsigned long *addr = &page_private(page); | |
267 | return (*addr & ((1UL<<BM_PAGE_NEED_WRITEOUT)|(1UL<<BM_PAGE_LAZY_WRITEOUT))) == 0; | |
268 | } | |
b411b363 | 269 | |
19f843aa LE |
270 | static void bm_set_page_io_err(struct page *page) |
271 | { | |
272 | set_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
273 | } | |
274 | ||
275 | static void bm_clear_page_io_err(struct page *page) | |
276 | { | |
277 | clear_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
278 | } | |
279 | ||
280 | static void bm_set_page_lazy_writeout(struct page *page) | |
281 | { | |
282 | set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
283 | } | |
284 | ||
285 | static int bm_test_page_lazy_writeout(struct page *page) | |
286 | { | |
287 | return test_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
288 | } | |
289 | ||
290 | /* on a 32bit box, this would allow for exactly (2<<38) bits. */ | |
291 | static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr) | |
292 | { | |
b411b363 | 293 | /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */ |
19f843aa | 294 | unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3); |
b411b363 | 295 | BUG_ON(page_nr >= b->bm_number_of_pages); |
19f843aa LE |
296 | return page_nr; |
297 | } | |
b411b363 | 298 | |
19f843aa LE |
299 | static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr) |
300 | { | |
301 | /* page_nr = (bitnr/8) >> PAGE_SHIFT; */ | |
302 | unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3); | |
303 | BUG_ON(page_nr >= b->bm_number_of_pages); | |
304 | return page_nr; | |
b411b363 PR |
305 | } |
306 | ||
589973a7 | 307 | static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) |
95a0f10c LE |
308 | { |
309 | struct page *page = b->bm_pages[idx]; | |
cfd8005c | 310 | return (unsigned long *) kmap_atomic(page); |
95a0f10c LE |
311 | } |
312 | ||
313 | static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) | |
314 | { | |
cfd8005c | 315 | return __bm_map_pidx(b, idx); |
95a0f10c LE |
316 | } |
317 | ||
cfd8005c | 318 | static void __bm_unmap(unsigned long *p_addr) |
b411b363 | 319 | { |
cfd8005c | 320 | kunmap_atomic(p_addr); |
b411b363 PR |
321 | }; |
322 | ||
323 | static void bm_unmap(unsigned long *p_addr) | |
324 | { | |
cfd8005c | 325 | return __bm_unmap(p_addr); |
b411b363 PR |
326 | } |
327 | ||
328 | /* long word offset of _bitmap_ sector */ | |
329 | #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) | |
330 | /* word offset from start of bitmap to word number _in_page_ | |
331 | * modulo longs per page | |
332 | #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long)) | |
24c4830c | 333 | hm, well, Philipp thinks gcc might not optimize the % into & (... - 1) |
b411b363 PR |
334 | so do it explicitly: |
335 | */ | |
336 | #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1)) | |
337 | ||
338 | /* Long words per page */ | |
339 | #define LWPP (PAGE_SIZE/sizeof(long)) | |
340 | ||
341 | /* | |
342 | * actually most functions herein should take a struct drbd_bitmap*, not a | |
b30ab791 | 343 | * struct drbd_device*, but for the debug macros I like to have the device around |
b411b363 PR |
344 | * to be able to report device specific. |
345 | */ | |
346 | ||
19f843aa | 347 | |
b411b363 PR |
348 | static void bm_free_pages(struct page **pages, unsigned long number) |
349 | { | |
350 | unsigned long i; | |
351 | if (!pages) | |
352 | return; | |
353 | ||
354 | for (i = 0; i < number; i++) { | |
355 | if (!pages[i]) { | |
356 | printk(KERN_ALERT "drbd: bm_free_pages tried to free " | |
357 | "a NULL pointer; i=%lu n=%lu\n", | |
358 | i, number); | |
359 | continue; | |
360 | } | |
361 | __free_page(pages[i]); | |
362 | pages[i] = NULL; | |
363 | } | |
364 | } | |
365 | ||
366 | static void bm_vk_free(void *ptr, int v) | |
367 | { | |
368 | if (v) | |
369 | vfree(ptr); | |
370 | else | |
371 | kfree(ptr); | |
372 | } | |
373 | ||
374 | /* | |
375 | * "have" and "want" are NUMBER OF PAGES. | |
376 | */ | |
377 | static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want) | |
378 | { | |
379 | struct page **old_pages = b->bm_pages; | |
380 | struct page **new_pages, *page; | |
381 | unsigned int i, bytes, vmalloced = 0; | |
382 | unsigned long have = b->bm_number_of_pages; | |
383 | ||
384 | BUG_ON(have == 0 && old_pages != NULL); | |
385 | BUG_ON(have != 0 && old_pages == NULL); | |
386 | ||
387 | if (have == want) | |
388 | return old_pages; | |
389 | ||
390 | /* Trying kmalloc first, falling back to vmalloc. | |
0b143d43 LE |
391 | * GFP_NOIO, as this is called while drbd IO is "suspended", |
392 | * and during resize or attach on diskless Primary, | |
393 | * we must not block on IO to ourselves. | |
bc891c9a | 394 | * Context is receiver thread or dmsetup. */ |
b411b363 | 395 | bytes = sizeof(struct page *)*want; |
8be04b93 | 396 | new_pages = kzalloc(bytes, GFP_NOIO | __GFP_NOWARN); |
b411b363 | 397 | if (!new_pages) { |
0b143d43 LE |
398 | new_pages = __vmalloc(bytes, |
399 | GFP_NOIO | __GFP_HIGHMEM | __GFP_ZERO, | |
400 | PAGE_KERNEL); | |
b411b363 PR |
401 | if (!new_pages) |
402 | return NULL; | |
403 | vmalloced = 1; | |
404 | } | |
405 | ||
b411b363 PR |
406 | if (want >= have) { |
407 | for (i = 0; i < have; i++) | |
408 | new_pages[i] = old_pages[i]; | |
409 | for (; i < want; i++) { | |
0b143d43 | 410 | page = alloc_page(GFP_NOIO | __GFP_HIGHMEM); |
b411b363 PR |
411 | if (!page) { |
412 | bm_free_pages(new_pages + have, i - have); | |
413 | bm_vk_free(new_pages, vmalloced); | |
414 | return NULL; | |
415 | } | |
19f843aa LE |
416 | /* we want to know which page it is |
417 | * from the endio handlers */ | |
418 | bm_store_page_idx(page, i); | |
b411b363 PR |
419 | new_pages[i] = page; |
420 | } | |
421 | } else { | |
422 | for (i = 0; i < want; i++) | |
423 | new_pages[i] = old_pages[i]; | |
424 | /* NOT HERE, we are outside the spinlock! | |
425 | bm_free_pages(old_pages + want, have - want); | |
426 | */ | |
427 | } | |
428 | ||
429 | if (vmalloced) | |
20ceb2b2 | 430 | b->bm_flags |= BM_P_VMALLOCED; |
b411b363 | 431 | else |
20ceb2b2 | 432 | b->bm_flags &= ~BM_P_VMALLOCED; |
b411b363 PR |
433 | |
434 | return new_pages; | |
435 | } | |
436 | ||
437 | /* | |
438 | * called on driver init only. TODO call when a device is created. | |
b30ab791 | 439 | * allocates the drbd_bitmap, and stores it in device->bitmap. |
b411b363 | 440 | */ |
b30ab791 | 441 | int drbd_bm_init(struct drbd_device *device) |
b411b363 | 442 | { |
b30ab791 | 443 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
444 | WARN_ON(b != NULL); |
445 | b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL); | |
446 | if (!b) | |
447 | return -ENOMEM; | |
448 | spin_lock_init(&b->bm_lock); | |
8a03ae2a | 449 | mutex_init(&b->bm_change); |
b411b363 PR |
450 | init_waitqueue_head(&b->bm_io_wait); |
451 | ||
b30ab791 | 452 | device->bitmap = b; |
b411b363 PR |
453 | |
454 | return 0; | |
455 | } | |
456 | ||
b30ab791 | 457 | sector_t drbd_bm_capacity(struct drbd_device *device) |
b411b363 | 458 | { |
b30ab791 | 459 | if (!expect(device->bitmap)) |
841ce241 | 460 | return 0; |
b30ab791 | 461 | return device->bitmap->bm_dev_capacity; |
b411b363 PR |
462 | } |
463 | ||
464 | /* called on driver unload. TODO: call when a device is destroyed. | |
465 | */ | |
b30ab791 | 466 | void drbd_bm_cleanup(struct drbd_device *device) |
b411b363 | 467 | { |
b30ab791 | 468 | if (!expect(device->bitmap)) |
841ce241 | 469 | return; |
b30ab791 AG |
470 | bm_free_pages(device->bitmap->bm_pages, device->bitmap->bm_number_of_pages); |
471 | bm_vk_free(device->bitmap->bm_pages, (BM_P_VMALLOCED & device->bitmap->bm_flags)); | |
472 | kfree(device->bitmap); | |
473 | device->bitmap = NULL; | |
b411b363 PR |
474 | } |
475 | ||
476 | /* | |
477 | * since (b->bm_bits % BITS_PER_LONG) != 0, | |
478 | * this masks out the remaining bits. | |
479 | * Returns the number of bits cleared. | |
480 | */ | |
95a0f10c LE |
481 | #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3)) |
482 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1) | |
483 | #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1) | |
b411b363 PR |
484 | static int bm_clear_surplus(struct drbd_bitmap *b) |
485 | { | |
95a0f10c | 486 | unsigned long mask; |
b411b363 | 487 | unsigned long *p_addr, *bm; |
95a0f10c LE |
488 | int tmp; |
489 | int cleared = 0; | |
b411b363 | 490 | |
95a0f10c LE |
491 | /* number of bits modulo bits per page */ |
492 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
493 | /* mask the used bits of the word containing the last bit */ | |
494 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
495 | /* bitmap is always stored little endian, | |
496 | * on disk and in core memory alike */ | |
497 | mask = cpu_to_lel(mask); | |
498 | ||
6850c442 | 499 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
500 | bm = p_addr + (tmp/BITS_PER_LONG); |
501 | if (mask) { | |
502 | /* If mask != 0, we are not exactly aligned, so bm now points | |
503 | * to the long containing the last bit. | |
504 | * If mask == 0, bm already points to the word immediately | |
505 | * after the last (long word aligned) bit. */ | |
b411b363 PR |
506 | cleared = hweight_long(*bm & ~mask); |
507 | *bm &= mask; | |
95a0f10c | 508 | bm++; |
b411b363 PR |
509 | } |
510 | ||
95a0f10c LE |
511 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
512 | /* on a 32bit arch, we may need to zero out | |
513 | * a padding long to align with a 64bit remote */ | |
b411b363 PR |
514 | cleared += hweight_long(*bm); |
515 | *bm = 0; | |
516 | } | |
517 | bm_unmap(p_addr); | |
518 | return cleared; | |
519 | } | |
520 | ||
521 | static void bm_set_surplus(struct drbd_bitmap *b) | |
522 | { | |
95a0f10c | 523 | unsigned long mask; |
b411b363 | 524 | unsigned long *p_addr, *bm; |
95a0f10c LE |
525 | int tmp; |
526 | ||
527 | /* number of bits modulo bits per page */ | |
528 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
529 | /* mask the used bits of the word containing the last bit */ | |
530 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
531 | /* bitmap is always stored little endian, | |
532 | * on disk and in core memory alike */ | |
533 | mask = cpu_to_lel(mask); | |
534 | ||
6850c442 | 535 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
536 | bm = p_addr + (tmp/BITS_PER_LONG); |
537 | if (mask) { | |
538 | /* If mask != 0, we are not exactly aligned, so bm now points | |
539 | * to the long containing the last bit. | |
540 | * If mask == 0, bm already points to the word immediately | |
541 | * after the last (long word aligned) bit. */ | |
b411b363 | 542 | *bm |= ~mask; |
95a0f10c | 543 | bm++; |
b411b363 PR |
544 | } |
545 | ||
95a0f10c LE |
546 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
547 | /* on a 32bit arch, we may need to zero out | |
548 | * a padding long to align with a 64bit remote */ | |
549 | *bm = ~0UL; | |
b411b363 PR |
550 | } |
551 | bm_unmap(p_addr); | |
552 | } | |
553 | ||
4b0715f0 LE |
554 | /* you better not modify the bitmap while this is running, |
555 | * or its results will be stale */ | |
95a0f10c | 556 | static unsigned long bm_count_bits(struct drbd_bitmap *b) |
b411b363 | 557 | { |
4b0715f0 | 558 | unsigned long *p_addr; |
b411b363 | 559 | unsigned long bits = 0; |
4b0715f0 | 560 | unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1; |
6850c442 | 561 | int idx, i, last_word; |
4b0715f0 LE |
562 | |
563 | /* all but last page */ | |
6850c442 | 564 | for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) { |
cfd8005c | 565 | p_addr = __bm_map_pidx(b, idx); |
4b0715f0 LE |
566 | for (i = 0; i < LWPP; i++) |
567 | bits += hweight_long(p_addr[i]); | |
cfd8005c | 568 | __bm_unmap(p_addr); |
b411b363 PR |
569 | cond_resched(); |
570 | } | |
4b0715f0 LE |
571 | /* last (or only) page */ |
572 | last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL; | |
589973a7 | 573 | p_addr = __bm_map_pidx(b, idx); |
4b0715f0 LE |
574 | for (i = 0; i < last_word; i++) |
575 | bits += hweight_long(p_addr[i]); | |
576 | p_addr[last_word] &= cpu_to_lel(mask); | |
577 | bits += hweight_long(p_addr[last_word]); | |
578 | /* 32bit arch, may have an unused padding long */ | |
579 | if (BITS_PER_LONG == 32 && (last_word & 1) == 0) | |
580 | p_addr[last_word+1] = 0; | |
589973a7 | 581 | __bm_unmap(p_addr); |
b411b363 PR |
582 | return bits; |
583 | } | |
584 | ||
b411b363 PR |
585 | /* offset and len in long words.*/ |
586 | static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len) | |
587 | { | |
588 | unsigned long *p_addr, *bm; | |
19f843aa | 589 | unsigned int idx; |
b411b363 PR |
590 | size_t do_now, end; |
591 | ||
b411b363 PR |
592 | end = offset + len; |
593 | ||
594 | if (end > b->bm_words) { | |
595 | printk(KERN_ALERT "drbd: bm_memset end > bm_words\n"); | |
596 | return; | |
597 | } | |
598 | ||
599 | while (offset < end) { | |
600 | do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset; | |
19f843aa LE |
601 | idx = bm_word_to_page_idx(b, offset); |
602 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
603 | bm = p_addr + MLPP(offset); |
604 | if (bm+do_now > p_addr + LWPP) { | |
605 | printk(KERN_ALERT "drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n", | |
606 | p_addr, bm, (int)do_now); | |
84e7c0f7 LE |
607 | } else |
608 | memset(bm, c, do_now * sizeof(long)); | |
b411b363 | 609 | bm_unmap(p_addr); |
19f843aa | 610 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
611 | offset += do_now; |
612 | } | |
613 | } | |
614 | ||
ae8bf312 LE |
615 | /* For the layout, see comment above drbd_md_set_sector_offsets(). */ |
616 | static u64 drbd_md_on_disk_bits(struct drbd_backing_dev *ldev) | |
617 | { | |
618 | u64 bitmap_sectors; | |
619 | if (ldev->md.al_offset == 8) | |
620 | bitmap_sectors = ldev->md.md_size_sect - ldev->md.bm_offset; | |
621 | else | |
622 | bitmap_sectors = ldev->md.al_offset - ldev->md.bm_offset; | |
623 | return bitmap_sectors << (9 + 3); | |
624 | } | |
625 | ||
b411b363 PR |
626 | /* |
627 | * make sure the bitmap has enough room for the attached storage, | |
628 | * if necessary, resize. | |
629 | * called whenever we may have changed the device size. | |
630 | * returns -ENOMEM if we could not allocate enough memory, 0 on success. | |
631 | * In case this is actually a resize, we copy the old bitmap into the new one. | |
632 | * Otherwise, the bitmap is initialized to all bits set. | |
633 | */ | |
b30ab791 | 634 | int drbd_bm_resize(struct drbd_device *device, sector_t capacity, int set_new_bits) |
b411b363 | 635 | { |
b30ab791 | 636 | struct drbd_bitmap *b = device->bitmap; |
6850c442 | 637 | unsigned long bits, words, owords, obits; |
b411b363 PR |
638 | unsigned long want, have, onpages; /* number of pages */ |
639 | struct page **npages, **opages = NULL; | |
640 | int err = 0, growing; | |
641 | int opages_vmalloced; | |
642 | ||
841ce241 AG |
643 | if (!expect(b)) |
644 | return -ENOMEM; | |
b411b363 | 645 | |
b30ab791 | 646 | drbd_bm_lock(device, "resize", BM_LOCKED_MASK); |
b411b363 | 647 | |
d0180171 | 648 | drbd_info(device, "drbd_bm_resize called with capacity == %llu\n", |
b411b363 PR |
649 | (unsigned long long)capacity); |
650 | ||
651 | if (capacity == b->bm_dev_capacity) | |
652 | goto out; | |
653 | ||
20ceb2b2 | 654 | opages_vmalloced = (BM_P_VMALLOCED & b->bm_flags); |
b411b363 PR |
655 | |
656 | if (capacity == 0) { | |
657 | spin_lock_irq(&b->bm_lock); | |
658 | opages = b->bm_pages; | |
659 | onpages = b->bm_number_of_pages; | |
660 | owords = b->bm_words; | |
661 | b->bm_pages = NULL; | |
662 | b->bm_number_of_pages = | |
663 | b->bm_set = | |
664 | b->bm_bits = | |
665 | b->bm_words = | |
666 | b->bm_dev_capacity = 0; | |
667 | spin_unlock_irq(&b->bm_lock); | |
668 | bm_free_pages(opages, onpages); | |
669 | bm_vk_free(opages, opages_vmalloced); | |
670 | goto out; | |
671 | } | |
672 | bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT)); | |
673 | ||
674 | /* if we would use | |
675 | words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL; | |
676 | a 32bit host could present the wrong number of words | |
677 | to a 64bit host. | |
678 | */ | |
679 | words = ALIGN(bits, 64) >> LN2_BPL; | |
680 | ||
b30ab791 AG |
681 | if (get_ldev(device)) { |
682 | u64 bits_on_disk = drbd_md_on_disk_bits(device->ldev); | |
683 | put_ldev(device); | |
4b0715f0 | 684 | if (bits > bits_on_disk) { |
d0180171 AG |
685 | drbd_info(device, "bits = %lu\n", bits); |
686 | drbd_info(device, "bits_on_disk = %llu\n", bits_on_disk); | |
4b0715f0 LE |
687 | err = -ENOSPC; |
688 | goto out; | |
689 | } | |
b411b363 PR |
690 | } |
691 | ||
6850c442 | 692 | want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT; |
b411b363 PR |
693 | have = b->bm_number_of_pages; |
694 | if (want == have) { | |
0b0ba1ef | 695 | D_ASSERT(device, b->bm_pages != NULL); |
b411b363 PR |
696 | npages = b->bm_pages; |
697 | } else { | |
b30ab791 | 698 | if (drbd_insert_fault(device, DRBD_FAULT_BM_ALLOC)) |
b411b363 PR |
699 | npages = NULL; |
700 | else | |
701 | npages = bm_realloc_pages(b, want); | |
702 | } | |
703 | ||
704 | if (!npages) { | |
705 | err = -ENOMEM; | |
706 | goto out; | |
707 | } | |
708 | ||
709 | spin_lock_irq(&b->bm_lock); | |
710 | opages = b->bm_pages; | |
711 | owords = b->bm_words; | |
712 | obits = b->bm_bits; | |
713 | ||
714 | growing = bits > obits; | |
5223671b | 715 | if (opages && growing && set_new_bits) |
b411b363 PR |
716 | bm_set_surplus(b); |
717 | ||
718 | b->bm_pages = npages; | |
719 | b->bm_number_of_pages = want; | |
720 | b->bm_bits = bits; | |
721 | b->bm_words = words; | |
722 | b->bm_dev_capacity = capacity; | |
723 | ||
724 | if (growing) { | |
02d9a94b PR |
725 | if (set_new_bits) { |
726 | bm_memset(b, owords, 0xff, words-owords); | |
727 | b->bm_set += bits - obits; | |
728 | } else | |
729 | bm_memset(b, owords, 0x00, words-owords); | |
730 | ||
b411b363 PR |
731 | } |
732 | ||
733 | if (want < have) { | |
734 | /* implicit: (opages != NULL) && (opages != npages) */ | |
735 | bm_free_pages(opages + want, have - want); | |
736 | } | |
737 | ||
b411b363 PR |
738 | (void)bm_clear_surplus(b); |
739 | ||
740 | spin_unlock_irq(&b->bm_lock); | |
741 | if (opages != npages) | |
742 | bm_vk_free(opages, opages_vmalloced); | |
743 | if (!growing) | |
744 | b->bm_set = bm_count_bits(b); | |
d0180171 | 745 | drbd_info(device, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want); |
b411b363 PR |
746 | |
747 | out: | |
b30ab791 | 748 | drbd_bm_unlock(device); |
b411b363 PR |
749 | return err; |
750 | } | |
751 | ||
752 | /* inherently racy: | |
753 | * if not protected by other means, return value may be out of date when | |
754 | * leaving this function... | |
755 | * we still need to lock it, since it is important that this returns | |
756 | * bm_set == 0 precisely. | |
757 | * | |
758 | * maybe bm_set should be atomic_t ? | |
759 | */ | |
b30ab791 | 760 | unsigned long _drbd_bm_total_weight(struct drbd_device *device) |
b411b363 | 761 | { |
b30ab791 | 762 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
763 | unsigned long s; |
764 | unsigned long flags; | |
765 | ||
841ce241 AG |
766 | if (!expect(b)) |
767 | return 0; | |
768 | if (!expect(b->bm_pages)) | |
769 | return 0; | |
b411b363 PR |
770 | |
771 | spin_lock_irqsave(&b->bm_lock, flags); | |
772 | s = b->bm_set; | |
773 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
774 | ||
775 | return s; | |
776 | } | |
777 | ||
b30ab791 | 778 | unsigned long drbd_bm_total_weight(struct drbd_device *device) |
b411b363 PR |
779 | { |
780 | unsigned long s; | |
781 | /* if I don't have a disk, I don't know about out-of-sync status */ | |
b30ab791 | 782 | if (!get_ldev_if_state(device, D_NEGOTIATING)) |
b411b363 | 783 | return 0; |
b30ab791 AG |
784 | s = _drbd_bm_total_weight(device); |
785 | put_ldev(device); | |
b411b363 PR |
786 | return s; |
787 | } | |
788 | ||
b30ab791 | 789 | size_t drbd_bm_words(struct drbd_device *device) |
b411b363 | 790 | { |
b30ab791 | 791 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
792 | if (!expect(b)) |
793 | return 0; | |
794 | if (!expect(b->bm_pages)) | |
795 | return 0; | |
b411b363 PR |
796 | |
797 | return b->bm_words; | |
798 | } | |
799 | ||
b30ab791 | 800 | unsigned long drbd_bm_bits(struct drbd_device *device) |
b411b363 | 801 | { |
b30ab791 | 802 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
803 | if (!expect(b)) |
804 | return 0; | |
b411b363 PR |
805 | |
806 | return b->bm_bits; | |
807 | } | |
808 | ||
809 | /* merge number words from buffer into the bitmap starting at offset. | |
810 | * buffer[i] is expected to be little endian unsigned long. | |
811 | * bitmap must be locked by drbd_bm_lock. | |
812 | * currently only used from receive_bitmap. | |
813 | */ | |
b30ab791 | 814 | void drbd_bm_merge_lel(struct drbd_device *device, size_t offset, size_t number, |
b411b363 PR |
815 | unsigned long *buffer) |
816 | { | |
b30ab791 | 817 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
818 | unsigned long *p_addr, *bm; |
819 | unsigned long word, bits; | |
19f843aa | 820 | unsigned int idx; |
b411b363 PR |
821 | size_t end, do_now; |
822 | ||
823 | end = offset + number; | |
824 | ||
841ce241 AG |
825 | if (!expect(b)) |
826 | return; | |
827 | if (!expect(b->bm_pages)) | |
828 | return; | |
b411b363 PR |
829 | if (number == 0) |
830 | return; | |
831 | WARN_ON(offset >= b->bm_words); | |
832 | WARN_ON(end > b->bm_words); | |
833 | ||
834 | spin_lock_irq(&b->bm_lock); | |
835 | while (offset < end) { | |
836 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa LE |
837 | idx = bm_word_to_page_idx(b, offset); |
838 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
839 | bm = p_addr + MLPP(offset); |
840 | offset += do_now; | |
841 | while (do_now--) { | |
842 | bits = hweight_long(*bm); | |
95a0f10c | 843 | word = *bm | *buffer++; |
b411b363 PR |
844 | *bm++ = word; |
845 | b->bm_set += hweight_long(word) - bits; | |
846 | } | |
847 | bm_unmap(p_addr); | |
19f843aa | 848 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
849 | } |
850 | /* with 32bit <-> 64bit cross-platform connect | |
851 | * this is only correct for current usage, | |
852 | * where we _know_ that we are 64 bit aligned, | |
853 | * and know that this function is used in this way, too... | |
854 | */ | |
855 | if (end == b->bm_words) | |
856 | b->bm_set -= bm_clear_surplus(b); | |
b411b363 PR |
857 | spin_unlock_irq(&b->bm_lock); |
858 | } | |
859 | ||
860 | /* copy number words from the bitmap starting at offset into the buffer. | |
861 | * buffer[i] will be little endian unsigned long. | |
862 | */ | |
b30ab791 | 863 | void drbd_bm_get_lel(struct drbd_device *device, size_t offset, size_t number, |
b411b363 PR |
864 | unsigned long *buffer) |
865 | { | |
b30ab791 | 866 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
867 | unsigned long *p_addr, *bm; |
868 | size_t end, do_now; | |
869 | ||
870 | end = offset + number; | |
871 | ||
841ce241 AG |
872 | if (!expect(b)) |
873 | return; | |
874 | if (!expect(b->bm_pages)) | |
875 | return; | |
b411b363 PR |
876 | |
877 | spin_lock_irq(&b->bm_lock); | |
878 | if ((offset >= b->bm_words) || | |
879 | (end > b->bm_words) || | |
880 | (number <= 0)) | |
d0180171 | 881 | drbd_err(device, "offset=%lu number=%lu bm_words=%lu\n", |
b411b363 PR |
882 | (unsigned long) offset, |
883 | (unsigned long) number, | |
884 | (unsigned long) b->bm_words); | |
885 | else { | |
886 | while (offset < end) { | |
887 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa | 888 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset)); |
b411b363 PR |
889 | bm = p_addr + MLPP(offset); |
890 | offset += do_now; | |
891 | while (do_now--) | |
95a0f10c | 892 | *buffer++ = *bm++; |
b411b363 PR |
893 | bm_unmap(p_addr); |
894 | } | |
895 | } | |
896 | spin_unlock_irq(&b->bm_lock); | |
897 | } | |
898 | ||
899 | /* set all bits in the bitmap */ | |
b30ab791 | 900 | void drbd_bm_set_all(struct drbd_device *device) |
b411b363 | 901 | { |
b30ab791 | 902 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
903 | if (!expect(b)) |
904 | return; | |
905 | if (!expect(b->bm_pages)) | |
906 | return; | |
b411b363 PR |
907 | |
908 | spin_lock_irq(&b->bm_lock); | |
909 | bm_memset(b, 0, 0xff, b->bm_words); | |
910 | (void)bm_clear_surplus(b); | |
911 | b->bm_set = b->bm_bits; | |
912 | spin_unlock_irq(&b->bm_lock); | |
913 | } | |
914 | ||
915 | /* clear all bits in the bitmap */ | |
b30ab791 | 916 | void drbd_bm_clear_all(struct drbd_device *device) |
b411b363 | 917 | { |
b30ab791 | 918 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
919 | if (!expect(b)) |
920 | return; | |
921 | if (!expect(b->bm_pages)) | |
922 | return; | |
b411b363 PR |
923 | |
924 | spin_lock_irq(&b->bm_lock); | |
925 | bm_memset(b, 0, 0, b->bm_words); | |
926 | b->bm_set = 0; | |
927 | spin_unlock_irq(&b->bm_lock); | |
928 | } | |
929 | ||
19f843aa | 930 | struct bm_aio_ctx { |
b30ab791 | 931 | struct drbd_device *device; |
19f843aa | 932 | atomic_t in_flight; |
9e58c4da | 933 | unsigned int done; |
19f843aa LE |
934 | unsigned flags; |
935 | #define BM_AIO_COPY_PAGES 1 | |
45dfffeb | 936 | #define BM_AIO_WRITE_HINTED 2 |
fef45d29 | 937 | #define BM_WRITE_ALL_PAGES 4 |
19f843aa | 938 | int error; |
d1f3779b | 939 | struct kref kref; |
19f843aa LE |
940 | }; |
941 | ||
d1f3779b PR |
942 | static void bm_aio_ctx_destroy(struct kref *kref) |
943 | { | |
944 | struct bm_aio_ctx *ctx = container_of(kref, struct bm_aio_ctx, kref); | |
945 | ||
b30ab791 | 946 | put_ldev(ctx->device); |
d1f3779b PR |
947 | kfree(ctx); |
948 | } | |
949 | ||
19f843aa | 950 | /* bv_page may be a copy, or may be the original */ |
b411b363 PR |
951 | static void bm_async_io_complete(struct bio *bio, int error) |
952 | { | |
19f843aa | 953 | struct bm_aio_ctx *ctx = bio->bi_private; |
b30ab791 AG |
954 | struct drbd_device *device = ctx->device; |
955 | struct drbd_bitmap *b = device->bitmap; | |
19f843aa | 956 | unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page); |
b411b363 PR |
957 | int uptodate = bio_flagged(bio, BIO_UPTODATE); |
958 | ||
959 | ||
960 | /* strange behavior of some lower level drivers... | |
961 | * fail the request by clearing the uptodate flag, | |
962 | * but do not return any error?! | |
963 | * do we want to WARN() on this? */ | |
964 | if (!error && !uptodate) | |
965 | error = -EIO; | |
966 | ||
7648cdfe LE |
967 | if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 && |
968 | !bm_test_page_unchanged(b->bm_pages[idx])) | |
d0180171 | 969 | drbd_warn(device, "bitmap page idx %u changed during IO!\n", idx); |
19f843aa | 970 | |
b411b363 | 971 | if (error) { |
19f843aa LE |
972 | /* ctx error will hold the completed-last non-zero error code, |
973 | * in case error codes differ. */ | |
974 | ctx->error = error; | |
975 | bm_set_page_io_err(b->bm_pages[idx]); | |
976 | /* Not identical to on disk version of it. | |
977 | * Is BM_PAGE_IO_ERROR enough? */ | |
978 | if (__ratelimit(&drbd_ratelimit_state)) | |
d0180171 | 979 | drbd_err(device, "IO ERROR %d on bitmap page idx %u\n", |
19f843aa LE |
980 | error, idx); |
981 | } else { | |
982 | bm_clear_page_io_err(b->bm_pages[idx]); | |
d0180171 | 983 | dynamic_drbd_dbg(device, "bitmap page idx %u completed\n", idx); |
b411b363 | 984 | } |
19f843aa | 985 | |
b30ab791 | 986 | bm_page_unlock_io(device, idx); |
19f843aa | 987 | |
19f843aa | 988 | if (ctx->flags & BM_AIO_COPY_PAGES) |
4d95a10f | 989 | mempool_free(bio->bi_io_vec[0].bv_page, drbd_md_io_page_pool); |
b411b363 PR |
990 | |
991 | bio_put(bio); | |
19f843aa | 992 | |
d1f3779b | 993 | if (atomic_dec_and_test(&ctx->in_flight)) { |
9e58c4da | 994 | ctx->done = 1; |
b30ab791 | 995 | wake_up(&device->misc_wait); |
d1f3779b PR |
996 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); |
997 | } | |
b411b363 PR |
998 | } |
999 | ||
19f843aa | 1000 | static void bm_page_io_async(struct bm_aio_ctx *ctx, int page_nr, int rw) __must_hold(local) |
b411b363 | 1001 | { |
9476f39d | 1002 | struct bio *bio = bio_alloc_drbd(GFP_NOIO); |
b30ab791 AG |
1003 | struct drbd_device *device = ctx->device; |
1004 | struct drbd_bitmap *b = device->bitmap; | |
19f843aa | 1005 | struct page *page; |
b411b363 | 1006 | unsigned int len; |
19f843aa | 1007 | |
b411b363 | 1008 | sector_t on_disk_sector = |
b30ab791 | 1009 | device->ldev->md.md_offset + device->ldev->md.bm_offset; |
b411b363 PR |
1010 | on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9); |
1011 | ||
1012 | /* this might happen with very small | |
19f843aa LE |
1013 | * flexible external meta data device, |
1014 | * or with PAGE_SIZE > 4k */ | |
b411b363 | 1015 | len = min_t(unsigned int, PAGE_SIZE, |
b30ab791 | 1016 | (drbd_md_last_sector(device->ldev) - on_disk_sector + 1)<<9); |
b411b363 | 1017 | |
19f843aa | 1018 | /* serialize IO on this page */ |
b30ab791 | 1019 | bm_page_lock_io(device, page_nr); |
19f843aa LE |
1020 | /* before memcpy and submit, |
1021 | * so it can be redirtied any time */ | |
1022 | bm_set_page_unchanged(b->bm_pages[page_nr]); | |
1023 | ||
1024 | if (ctx->flags & BM_AIO_COPY_PAGES) { | |
4d95a10f | 1025 | page = mempool_alloc(drbd_md_io_page_pool, __GFP_HIGHMEM|__GFP_WAIT); |
f1d6a328 | 1026 | copy_highpage(page, b->bm_pages[page_nr]); |
19f843aa LE |
1027 | bm_store_page_idx(page, page_nr); |
1028 | } else | |
1029 | page = b->bm_pages[page_nr]; | |
b30ab791 | 1030 | bio->bi_bdev = device->ldev->md_bdev; |
4f024f37 | 1031 | bio->bi_iter.bi_sector = on_disk_sector; |
4d95a10f LE |
1032 | /* bio_add_page of a single page to an empty bio will always succeed, |
1033 | * according to api. Do we want to assert that? */ | |
19f843aa LE |
1034 | bio_add_page(bio, page, len, 0); |
1035 | bio->bi_private = ctx; | |
b411b363 PR |
1036 | bio->bi_end_io = bm_async_io_complete; |
1037 | ||
b30ab791 | 1038 | if (drbd_insert_fault(device, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) { |
b411b363 PR |
1039 | bio->bi_rw |= rw; |
1040 | bio_endio(bio, -EIO); | |
1041 | } else { | |
1042 | submit_bio(rw, bio); | |
5a8b4242 LE |
1043 | /* this should not count as user activity and cause the |
1044 | * resync to throttle -- see drbd_rs_should_slow_down(). */ | |
b30ab791 | 1045 | atomic_add(len >> 9, &device->rs_sect_ev); |
b411b363 PR |
1046 | } |
1047 | } | |
1048 | ||
b411b363 PR |
1049 | /* |
1050 | * bm_rw: read/write the whole bitmap from/to its on disk location. | |
1051 | */ | |
b30ab791 | 1052 | static int bm_rw(struct drbd_device *device, int rw, unsigned flags, unsigned lazy_writeout_upper_idx) __must_hold(local) |
b411b363 | 1053 | { |
d1f3779b | 1054 | struct bm_aio_ctx *ctx; |
b30ab791 | 1055 | struct drbd_bitmap *b = device->bitmap; |
6850c442 | 1056 | int num_pages, i, count = 0; |
b411b363 PR |
1057 | unsigned long now; |
1058 | char ppb[10]; | |
1059 | int err = 0; | |
1060 | ||
19f843aa LE |
1061 | /* |
1062 | * We are protected against bitmap disappearing/resizing by holding an | |
1063 | * ldev reference (caller must have called get_ldev()). | |
1064 | * For read/write, we are protected against changes to the bitmap by | |
1065 | * the bitmap lock (see drbd_bitmap_io). | |
1066 | * For lazy writeout, we don't care for ongoing changes to the bitmap, | |
1067 | * as we submit copies of pages anyways. | |
1068 | */ | |
d1f3779b | 1069 | |
22f46ce2 | 1070 | ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO); |
d1f3779b PR |
1071 | if (!ctx) |
1072 | return -ENOMEM; | |
1073 | ||
1074 | *ctx = (struct bm_aio_ctx) { | |
b30ab791 | 1075 | .device = device, |
d1f3779b | 1076 | .in_flight = ATOMIC_INIT(1), |
9e58c4da | 1077 | .done = 0, |
0e8488ad | 1078 | .flags = flags, |
d1f3779b PR |
1079 | .error = 0, |
1080 | .kref = { ATOMIC_INIT(2) }, | |
1081 | }; | |
1082 | ||
b30ab791 | 1083 | if (!get_ldev_if_state(device, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */ |
d0180171 | 1084 | drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n"); |
9e58c4da PR |
1085 | kfree(ctx); |
1086 | return -ENODEV; | |
1087 | } | |
1088 | ||
d1f3779b | 1089 | if (!ctx->flags) |
20ceb2b2 | 1090 | WARN_ON(!(BM_LOCKED_MASK & b->bm_flags)); |
b411b363 | 1091 | |
6850c442 | 1092 | num_pages = b->bm_number_of_pages; |
b411b363 | 1093 | |
b411b363 | 1094 | now = jiffies; |
b411b363 PR |
1095 | |
1096 | /* let the layers below us try to merge these bios... */ | |
6850c442 | 1097 | for (i = 0; i < num_pages; i++) { |
19f843aa LE |
1098 | /* ignore completely unchanged pages */ |
1099 | if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx) | |
1100 | break; | |
1101 | if (rw & WRITE) { | |
45dfffeb LE |
1102 | if ((flags & BM_AIO_WRITE_HINTED) && |
1103 | !test_and_clear_bit(BM_PAGE_HINT_WRITEOUT, | |
1104 | &page_private(b->bm_pages[i]))) | |
1105 | continue; | |
fef45d29 | 1106 | |
d1aa4d04 PR |
1107 | if (!(flags & BM_WRITE_ALL_PAGES) && |
1108 | bm_test_page_unchanged(b->bm_pages[i])) { | |
d0180171 | 1109 | dynamic_drbd_dbg(device, "skipped bm write for idx %u\n", i); |
19f843aa LE |
1110 | continue; |
1111 | } | |
1112 | /* during lazy writeout, | |
1113 | * ignore those pages not marked for lazy writeout. */ | |
1114 | if (lazy_writeout_upper_idx && | |
1115 | !bm_test_page_lazy_writeout(b->bm_pages[i])) { | |
d0180171 | 1116 | dynamic_drbd_dbg(device, "skipped bm lazy write for idx %u\n", i); |
19f843aa LE |
1117 | continue; |
1118 | } | |
1119 | } | |
d1f3779b PR |
1120 | atomic_inc(&ctx->in_flight); |
1121 | bm_page_io_async(ctx, i, rw); | |
19f843aa LE |
1122 | ++count; |
1123 | cond_resched(); | |
1124 | } | |
b411b363 | 1125 | |
725a97e4 | 1126 | /* |
d1f3779b | 1127 | * We initialize ctx->in_flight to one to make sure bm_async_io_complete |
9e58c4da | 1128 | * will not set ctx->done early, and decrement / test it here. If there |
725a97e4 | 1129 | * are still some bios in flight, we need to wait for them here. |
9e58c4da PR |
1130 | * If all IO is done already (or nothing had been submitted), there is |
1131 | * no need to wait. Still, we need to put the kref associated with the | |
1132 | * "in_flight reached zero, all done" event. | |
725a97e4 | 1133 | */ |
d1f3779b | 1134 | if (!atomic_dec_and_test(&ctx->in_flight)) |
b30ab791 | 1135 | wait_until_done_or_force_detached(device, device->ldev, &ctx->done); |
9e58c4da PR |
1136 | else |
1137 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); | |
d1f3779b | 1138 | |
c9d963a4 LE |
1139 | /* summary for global bitmap IO */ |
1140 | if (flags == 0) | |
d0180171 | 1141 | drbd_info(device, "bitmap %s of %u pages took %lu jiffies\n", |
cdfda633 PR |
1142 | rw == WRITE ? "WRITE" : "READ", |
1143 | count, jiffies - now); | |
b411b363 | 1144 | |
d1f3779b | 1145 | if (ctx->error) { |
d0180171 | 1146 | drbd_alert(device, "we had at least one MD IO ERROR during bitmap IO\n"); |
b30ab791 | 1147 | drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR); |
d1f3779b | 1148 | err = -EIO; /* ctx->error ? */ |
b411b363 PR |
1149 | } |
1150 | ||
9e58c4da | 1151 | if (atomic_read(&ctx->in_flight)) |
44edfb0d | 1152 | err = -EIO; /* Disk timeout/force-detach during IO... */ |
9e58c4da | 1153 | |
b411b363 PR |
1154 | now = jiffies; |
1155 | if (rw == WRITE) { | |
b30ab791 | 1156 | drbd_md_flush(device); |
b411b363 | 1157 | } else /* rw == READ */ { |
95a0f10c | 1158 | b->bm_set = bm_count_bits(b); |
d0180171 | 1159 | drbd_info(device, "recounting of set bits took additional %lu jiffies\n", |
b411b363 PR |
1160 | jiffies - now); |
1161 | } | |
1162 | now = b->bm_set; | |
1163 | ||
c9d963a4 | 1164 | if (flags == 0) |
d0180171 | 1165 | drbd_info(device, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n", |
c9d963a4 | 1166 | ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now); |
b411b363 | 1167 | |
d1f3779b | 1168 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); |
b411b363 PR |
1169 | return err; |
1170 | } | |
1171 | ||
1172 | /** | |
1173 | * drbd_bm_read() - Read the whole bitmap from its on disk location. | |
b30ab791 | 1174 | * @device: DRBD device. |
b411b363 | 1175 | */ |
b30ab791 | 1176 | int drbd_bm_read(struct drbd_device *device) __must_hold(local) |
b411b363 | 1177 | { |
b30ab791 | 1178 | return bm_rw(device, READ, 0, 0); |
b411b363 PR |
1179 | } |
1180 | ||
1181 | /** | |
1182 | * drbd_bm_write() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1183 | * @device: DRBD device. |
19f843aa LE |
1184 | * |
1185 | * Will only write pages that have changed since last IO. | |
b411b363 | 1186 | */ |
b30ab791 | 1187 | int drbd_bm_write(struct drbd_device *device) __must_hold(local) |
b411b363 | 1188 | { |
b30ab791 | 1189 | return bm_rw(device, WRITE, 0, 0); |
b411b363 PR |
1190 | } |
1191 | ||
d1aa4d04 PR |
1192 | /** |
1193 | * drbd_bm_write_all() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1194 | * @device: DRBD device. |
d1aa4d04 PR |
1195 | * |
1196 | * Will write all pages. | |
1197 | */ | |
b30ab791 | 1198 | int drbd_bm_write_all(struct drbd_device *device) __must_hold(local) |
d1aa4d04 | 1199 | { |
b30ab791 | 1200 | return bm_rw(device, WRITE, BM_WRITE_ALL_PAGES, 0); |
d1aa4d04 PR |
1201 | } |
1202 | ||
0e8488ad LE |
1203 | /** |
1204 | * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1205 | * @device: DRBD device. |
0e8488ad LE |
1206 | * |
1207 | * Will only write pages that have changed since last IO. | |
1208 | * In contrast to drbd_bm_write(), this will copy the bitmap pages | |
1209 | * to temporary writeout pages. It is intended to trigger a full write-out | |
1210 | * while still allowing the bitmap to change, for example if a resync or online | |
1211 | * verify is aborted due to a failed peer disk, while local IO continues, or | |
1212 | * pending resync acks are still being processed. | |
1213 | */ | |
b30ab791 | 1214 | int drbd_bm_write_copy_pages(struct drbd_device *device) __must_hold(local) |
0e8488ad | 1215 | { |
b30ab791 | 1216 | return bm_rw(device, WRITE, BM_AIO_COPY_PAGES, 0); |
19f843aa LE |
1217 | } |
1218 | ||
45dfffeb LE |
1219 | /** |
1220 | * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed. | |
b30ab791 | 1221 | * @device: DRBD device. |
45dfffeb | 1222 | */ |
b30ab791 | 1223 | int drbd_bm_write_hinted(struct drbd_device *device) __must_hold(local) |
45dfffeb | 1224 | { |
b30ab791 | 1225 | return bm_rw(device, WRITE, BM_AIO_WRITE_HINTED | BM_AIO_COPY_PAGES, 0); |
45dfffeb | 1226 | } |
19f843aa LE |
1227 | |
1228 | /** | |
45dfffeb | 1229 | * drbd_bm_write_page() - Writes a PAGE_SIZE aligned piece of bitmap |
b30ab791 | 1230 | * @device: DRBD device. |
19f843aa | 1231 | * @idx: bitmap page index |
b411b363 | 1232 | * |
4b0715f0 LE |
1233 | * We don't want to special case on logical_block_size of the backend device, |
1234 | * so we submit PAGE_SIZE aligned pieces. | |
19f843aa | 1235 | * Note that on "most" systems, PAGE_SIZE is 4k. |
4b0715f0 LE |
1236 | * |
1237 | * In case this becomes an issue on systems with larger PAGE_SIZE, | |
1238 | * we may want to change this again to write 4k aligned 4k pieces. | |
b411b363 | 1239 | */ |
b30ab791 | 1240 | int drbd_bm_write_page(struct drbd_device *device, unsigned int idx) __must_hold(local) |
b411b363 | 1241 | { |
d1f3779b PR |
1242 | struct bm_aio_ctx *ctx; |
1243 | int err; | |
b411b363 | 1244 | |
b30ab791 | 1245 | if (bm_test_page_unchanged(device->bitmap->bm_pages[idx])) { |
d0180171 | 1246 | dynamic_drbd_dbg(device, "skipped bm page write for idx %u\n", idx); |
19f843aa | 1247 | return 0; |
b411b363 | 1248 | } |
19f843aa | 1249 | |
22f46ce2 | 1250 | ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO); |
d1f3779b PR |
1251 | if (!ctx) |
1252 | return -ENOMEM; | |
1253 | ||
1254 | *ctx = (struct bm_aio_ctx) { | |
b30ab791 | 1255 | .device = device, |
d1f3779b | 1256 | .in_flight = ATOMIC_INIT(1), |
9e58c4da | 1257 | .done = 0, |
d1f3779b PR |
1258 | .flags = BM_AIO_COPY_PAGES, |
1259 | .error = 0, | |
1260 | .kref = { ATOMIC_INIT(2) }, | |
1261 | }; | |
1262 | ||
b30ab791 | 1263 | if (!get_ldev_if_state(device, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */ |
d0180171 | 1264 | drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in drbd_bm_write_page()\n"); |
9e58c4da PR |
1265 | kfree(ctx); |
1266 | return -ENODEV; | |
1267 | } | |
1268 | ||
d1f3779b | 1269 | bm_page_io_async(ctx, idx, WRITE_SYNC); |
b30ab791 | 1270 | wait_until_done_or_force_detached(device, device->ldev, &ctx->done); |
19f843aa | 1271 | |
d1f3779b | 1272 | if (ctx->error) |
b30ab791 | 1273 | drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR); |
44edfb0d | 1274 | /* that causes us to detach, so the in memory bitmap will be |
19f843aa LE |
1275 | * gone in a moment as well. */ |
1276 | ||
b30ab791 | 1277 | device->bm_writ_cnt++; |
9e58c4da | 1278 | err = atomic_read(&ctx->in_flight) ? -EIO : ctx->error; |
d1f3779b | 1279 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); |
d1f3779b | 1280 | return err; |
b411b363 PR |
1281 | } |
1282 | ||
1283 | /* NOTE | |
1284 | * find_first_bit returns int, we return unsigned long. | |
4b0715f0 LE |
1285 | * For this to work on 32bit arch with bitnumbers > (1<<32), |
1286 | * we'd need to return u64, and get a whole lot of other places | |
1287 | * fixed where we still use unsigned long. | |
b411b363 PR |
1288 | * |
1289 | * this returns a bit number, NOT a sector! | |
1290 | */ | |
b30ab791 | 1291 | static unsigned long __bm_find_next(struct drbd_device *device, unsigned long bm_fo, |
cfd8005c | 1292 | const int find_zero_bit) |
b411b363 | 1293 | { |
b30ab791 | 1294 | struct drbd_bitmap *b = device->bitmap; |
b411b363 | 1295 | unsigned long *p_addr; |
4b0715f0 LE |
1296 | unsigned long bit_offset; |
1297 | unsigned i; | |
1298 | ||
b411b363 PR |
1299 | |
1300 | if (bm_fo > b->bm_bits) { | |
d0180171 | 1301 | drbd_err(device, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits); |
4b0715f0 | 1302 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1303 | } else { |
1304 | while (bm_fo < b->bm_bits) { | |
19f843aa | 1305 | /* bit offset of the first bit in the page */ |
4b0715f0 | 1306 | bit_offset = bm_fo & ~BITS_PER_PAGE_MASK; |
cfd8005c | 1307 | p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo)); |
b411b363 PR |
1308 | |
1309 | if (find_zero_bit) | |
7e599e6e | 1310 | i = find_next_zero_bit_le(p_addr, |
4b0715f0 | 1311 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1312 | else |
7e599e6e | 1313 | i = find_next_bit_le(p_addr, |
4b0715f0 | 1314 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1315 | |
cfd8005c | 1316 | __bm_unmap(p_addr); |
b411b363 | 1317 | if (i < PAGE_SIZE*8) { |
4b0715f0 LE |
1318 | bm_fo = bit_offset + i; |
1319 | if (bm_fo >= b->bm_bits) | |
b411b363 PR |
1320 | break; |
1321 | goto found; | |
1322 | } | |
1323 | bm_fo = bit_offset + PAGE_SIZE*8; | |
1324 | } | |
4b0715f0 | 1325 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1326 | } |
1327 | found: | |
4b0715f0 | 1328 | return bm_fo; |
b411b363 PR |
1329 | } |
1330 | ||
b30ab791 | 1331 | static unsigned long bm_find_next(struct drbd_device *device, |
b411b363 PR |
1332 | unsigned long bm_fo, const int find_zero_bit) |
1333 | { | |
b30ab791 | 1334 | struct drbd_bitmap *b = device->bitmap; |
4b0715f0 | 1335 | unsigned long i = DRBD_END_OF_BITMAP; |
b411b363 | 1336 | |
841ce241 AG |
1337 | if (!expect(b)) |
1338 | return i; | |
1339 | if (!expect(b->bm_pages)) | |
1340 | return i; | |
b411b363 PR |
1341 | |
1342 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1343 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1344 | bm_print_lock_info(device); |
b411b363 | 1345 | |
b30ab791 | 1346 | i = __bm_find_next(device, bm_fo, find_zero_bit); |
b411b363 PR |
1347 | |
1348 | spin_unlock_irq(&b->bm_lock); | |
1349 | return i; | |
1350 | } | |
1351 | ||
b30ab791 | 1352 | unsigned long drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1353 | { |
b30ab791 | 1354 | return bm_find_next(device, bm_fo, 0); |
b411b363 PR |
1355 | } |
1356 | ||
1357 | #if 0 | |
1358 | /* not yet needed for anything. */ | |
b30ab791 | 1359 | unsigned long drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1360 | { |
b30ab791 | 1361 | return bm_find_next(device, bm_fo, 1); |
b411b363 PR |
1362 | } |
1363 | #endif | |
1364 | ||
1365 | /* does not spin_lock_irqsave. | |
1366 | * you must take drbd_bm_lock() first */ | |
b30ab791 | 1367 | unsigned long _drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1368 | { |
b30ab791 AG |
1369 | /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ |
1370 | return __bm_find_next(device, bm_fo, 0); | |
b411b363 PR |
1371 | } |
1372 | ||
b30ab791 | 1373 | unsigned long _drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1374 | { |
b30ab791 AG |
1375 | /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ |
1376 | return __bm_find_next(device, bm_fo, 1); | |
b411b363 PR |
1377 | } |
1378 | ||
1379 | /* returns number of bits actually changed. | |
1380 | * for val != 0, we change 0 -> 1, return code positive | |
1381 | * for val == 0, we change 1 -> 0, return code negative | |
1382 | * wants bitnr, not sector. | |
1383 | * expected to be called for only a few bits (e - s about BITS_PER_LONG). | |
1384 | * Must hold bitmap lock already. */ | |
b30ab791 | 1385 | static int __bm_change_bits_to(struct drbd_device *device, const unsigned long s, |
829c6087 | 1386 | unsigned long e, int val) |
b411b363 | 1387 | { |
b30ab791 | 1388 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1389 | unsigned long *p_addr = NULL; |
1390 | unsigned long bitnr; | |
19f843aa | 1391 | unsigned int last_page_nr = -1U; |
b411b363 | 1392 | int c = 0; |
19f843aa | 1393 | int changed_total = 0; |
b411b363 PR |
1394 | |
1395 | if (e >= b->bm_bits) { | |
d0180171 | 1396 | drbd_err(device, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n", |
b411b363 PR |
1397 | s, e, b->bm_bits); |
1398 | e = b->bm_bits ? b->bm_bits -1 : 0; | |
1399 | } | |
1400 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa | 1401 | unsigned int page_nr = bm_bit_to_page_idx(b, bitnr); |
b411b363 PR |
1402 | if (page_nr != last_page_nr) { |
1403 | if (p_addr) | |
cfd8005c | 1404 | __bm_unmap(p_addr); |
19f843aa LE |
1405 | if (c < 0) |
1406 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1407 | else if (c > 0) | |
1408 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1409 | changed_total += c; | |
1410 | c = 0; | |
cfd8005c | 1411 | p_addr = __bm_map_pidx(b, page_nr); |
b411b363 PR |
1412 | last_page_nr = page_nr; |
1413 | } | |
1414 | if (val) | |
7e599e6e | 1415 | c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 | 1416 | else |
7e599e6e | 1417 | c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 PR |
1418 | } |
1419 | if (p_addr) | |
cfd8005c | 1420 | __bm_unmap(p_addr); |
19f843aa LE |
1421 | if (c < 0) |
1422 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1423 | else if (c > 0) | |
1424 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1425 | changed_total += c; | |
1426 | b->bm_set += changed_total; | |
1427 | return changed_total; | |
b411b363 PR |
1428 | } |
1429 | ||
1430 | /* returns number of bits actually changed. | |
1431 | * for val != 0, we change 0 -> 1, return code positive | |
1432 | * for val == 0, we change 1 -> 0, return code negative | |
1433 | * wants bitnr, not sector */ | |
b30ab791 | 1434 | static int bm_change_bits_to(struct drbd_device *device, const unsigned long s, |
b411b363 PR |
1435 | const unsigned long e, int val) |
1436 | { | |
1437 | unsigned long flags; | |
b30ab791 | 1438 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1439 | int c = 0; |
1440 | ||
841ce241 AG |
1441 | if (!expect(b)) |
1442 | return 1; | |
1443 | if (!expect(b->bm_pages)) | |
1444 | return 0; | |
b411b363 PR |
1445 | |
1446 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1447 | if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags) |
b30ab791 | 1448 | bm_print_lock_info(device); |
b411b363 | 1449 | |
b30ab791 | 1450 | c = __bm_change_bits_to(device, s, e, val); |
b411b363 PR |
1451 | |
1452 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1453 | return c; | |
1454 | } | |
1455 | ||
1456 | /* returns number of bits changed 0 -> 1 */ | |
b30ab791 | 1457 | int drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 | 1458 | { |
b30ab791 | 1459 | return bm_change_bits_to(device, s, e, 1); |
b411b363 PR |
1460 | } |
1461 | ||
1462 | /* returns number of bits changed 1 -> 0 */ | |
b30ab791 | 1463 | int drbd_bm_clear_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 | 1464 | { |
b30ab791 | 1465 | return -bm_change_bits_to(device, s, e, 0); |
b411b363 PR |
1466 | } |
1467 | ||
1468 | /* sets all bits in full words, | |
1469 | * from first_word up to, but not including, last_word */ | |
1470 | static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b, | |
1471 | int page_nr, int first_word, int last_word) | |
1472 | { | |
1473 | int i; | |
1474 | int bits; | |
22d81140 | 1475 | int changed = 0; |
cfd8005c | 1476 | unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr]); |
b411b363 PR |
1477 | for (i = first_word; i < last_word; i++) { |
1478 | bits = hweight_long(paddr[i]); | |
1479 | paddr[i] = ~0UL; | |
22d81140 | 1480 | changed += BITS_PER_LONG - bits; |
b411b363 | 1481 | } |
cfd8005c | 1482 | kunmap_atomic(paddr); |
22d81140 LE |
1483 | if (changed) { |
1484 | /* We only need lazy writeout, the information is still in the | |
1485 | * remote bitmap as well, and is reconstructed during the next | |
1486 | * bitmap exchange, if lost locally due to a crash. */ | |
1487 | bm_set_page_lazy_writeout(b->bm_pages[page_nr]); | |
1488 | b->bm_set += changed; | |
1489 | } | |
b411b363 PR |
1490 | } |
1491 | ||
829c6087 LE |
1492 | /* Same thing as drbd_bm_set_bits, |
1493 | * but more efficient for a large bit range. | |
b411b363 PR |
1494 | * You must first drbd_bm_lock(). |
1495 | * Can be called to set the whole bitmap in one go. | |
1496 | * Sets bits from s to e _inclusive_. */ | |
b30ab791 | 1497 | void _drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 PR |
1498 | { |
1499 | /* First set_bit from the first bit (s) | |
1500 | * up to the next long boundary (sl), | |
1501 | * then assign full words up to the last long boundary (el), | |
1502 | * then set_bit up to and including the last bit (e). | |
1503 | * | |
1504 | * Do not use memset, because we must account for changes, | |
1505 | * so we need to loop over the words with hweight() anyways. | |
1506 | */ | |
b30ab791 | 1507 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1508 | unsigned long sl = ALIGN(s,BITS_PER_LONG); |
1509 | unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1); | |
1510 | int first_page; | |
1511 | int last_page; | |
1512 | int page_nr; | |
1513 | int first_word; | |
1514 | int last_word; | |
1515 | ||
1516 | if (e - s <= 3*BITS_PER_LONG) { | |
1517 | /* don't bother; el and sl may even be wrong. */ | |
829c6087 | 1518 | spin_lock_irq(&b->bm_lock); |
b30ab791 | 1519 | __bm_change_bits_to(device, s, e, 1); |
829c6087 | 1520 | spin_unlock_irq(&b->bm_lock); |
b411b363 PR |
1521 | return; |
1522 | } | |
1523 | ||
1524 | /* difference is large enough that we can trust sl and el */ | |
1525 | ||
829c6087 LE |
1526 | spin_lock_irq(&b->bm_lock); |
1527 | ||
b411b363 PR |
1528 | /* bits filling the current long */ |
1529 | if (sl) | |
b30ab791 | 1530 | __bm_change_bits_to(device, s, sl-1, 1); |
b411b363 PR |
1531 | |
1532 | first_page = sl >> (3 + PAGE_SHIFT); | |
1533 | last_page = el >> (3 + PAGE_SHIFT); | |
1534 | ||
1535 | /* MLPP: modulo longs per page */ | |
1536 | /* LWPP: long words per page */ | |
1537 | first_word = MLPP(sl >> LN2_BPL); | |
1538 | last_word = LWPP; | |
1539 | ||
1540 | /* first and full pages, unless first page == last page */ | |
1541 | for (page_nr = first_page; page_nr < last_page; page_nr++) { | |
b30ab791 | 1542 | bm_set_full_words_within_one_page(device->bitmap, page_nr, first_word, last_word); |
8ccee20e LE |
1543 | spin_unlock_irq(&b->bm_lock); |
1544 | cond_resched(); | |
b411b363 | 1545 | first_word = 0; |
8ccee20e | 1546 | spin_lock_irq(&b->bm_lock); |
b411b363 | 1547 | } |
b411b363 PR |
1548 | /* last page (respectively only page, for first page == last page) */ |
1549 | last_word = MLPP(el >> LN2_BPL); | |
4eccc579 LE |
1550 | |
1551 | /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples). | |
1552 | * ==> e = 32767, el = 32768, last_page = 2, | |
1553 | * and now last_word = 0. | |
1554 | * We do not want to touch last_page in this case, | |
1555 | * as we did not allocate it, it is not present in bitmap->bm_pages. | |
1556 | */ | |
1557 | if (last_word) | |
b30ab791 | 1558 | bm_set_full_words_within_one_page(device->bitmap, last_page, first_word, last_word); |
b411b363 PR |
1559 | |
1560 | /* possibly trailing bits. | |
1561 | * example: (e & 63) == 63, el will be e+1. | |
1562 | * if that even was the very last bit, | |
1563 | * it would trigger an assert in __bm_change_bits_to() | |
1564 | */ | |
1565 | if (el <= e) | |
b30ab791 | 1566 | __bm_change_bits_to(device, el, e, 1); |
829c6087 | 1567 | spin_unlock_irq(&b->bm_lock); |
b411b363 PR |
1568 | } |
1569 | ||
1570 | /* returns bit state | |
1571 | * wants bitnr, NOT sector. | |
1572 | * inherently racy... area needs to be locked by means of {al,rs}_lru | |
1573 | * 1 ... bit set | |
1574 | * 0 ... bit not set | |
1575 | * -1 ... first out of bounds access, stop testing for bits! | |
1576 | */ | |
b30ab791 | 1577 | int drbd_bm_test_bit(struct drbd_device *device, const unsigned long bitnr) |
b411b363 PR |
1578 | { |
1579 | unsigned long flags; | |
b30ab791 | 1580 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1581 | unsigned long *p_addr; |
1582 | int i; | |
1583 | ||
841ce241 AG |
1584 | if (!expect(b)) |
1585 | return 0; | |
1586 | if (!expect(b->bm_pages)) | |
1587 | return 0; | |
b411b363 PR |
1588 | |
1589 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1590 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1591 | bm_print_lock_info(device); |
b411b363 | 1592 | if (bitnr < b->bm_bits) { |
19f843aa | 1593 | p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr)); |
7e599e6e | 1594 | i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0; |
b411b363 PR |
1595 | bm_unmap(p_addr); |
1596 | } else if (bitnr == b->bm_bits) { | |
1597 | i = -1; | |
1598 | } else { /* (bitnr > b->bm_bits) */ | |
d0180171 | 1599 | drbd_err(device, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits); |
b411b363 PR |
1600 | i = 0; |
1601 | } | |
1602 | ||
1603 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1604 | return i; | |
1605 | } | |
1606 | ||
1607 | /* returns number of bits set in the range [s, e] */ | |
b30ab791 | 1608 | int drbd_bm_count_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 PR |
1609 | { |
1610 | unsigned long flags; | |
b30ab791 | 1611 | struct drbd_bitmap *b = device->bitmap; |
19f843aa | 1612 | unsigned long *p_addr = NULL; |
b411b363 | 1613 | unsigned long bitnr; |
19f843aa | 1614 | unsigned int page_nr = -1U; |
b411b363 | 1615 | int c = 0; |
b411b363 PR |
1616 | |
1617 | /* If this is called without a bitmap, that is a bug. But just to be | |
1618 | * robust in case we screwed up elsewhere, in that case pretend there | |
1619 | * was one dirty bit in the requested area, so we won't try to do a | |
1620 | * local read there (no bitmap probably implies no disk) */ | |
841ce241 AG |
1621 | if (!expect(b)) |
1622 | return 1; | |
1623 | if (!expect(b->bm_pages)) | |
1624 | return 1; | |
b411b363 PR |
1625 | |
1626 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1627 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1628 | bm_print_lock_info(device); |
b411b363 | 1629 | for (bitnr = s; bitnr <= e; bitnr++) { |
19f843aa LE |
1630 | unsigned int idx = bm_bit_to_page_idx(b, bitnr); |
1631 | if (page_nr != idx) { | |
1632 | page_nr = idx; | |
b411b363 PR |
1633 | if (p_addr) |
1634 | bm_unmap(p_addr); | |
19f843aa | 1635 | p_addr = bm_map_pidx(b, idx); |
b411b363 | 1636 | } |
841ce241 | 1637 | if (expect(bitnr < b->bm_bits)) |
7e599e6e | 1638 | c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr)); |
841ce241 | 1639 | else |
d0180171 | 1640 | drbd_err(device, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits); |
b411b363 PR |
1641 | } |
1642 | if (p_addr) | |
1643 | bm_unmap(p_addr); | |
1644 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1645 | return c; | |
1646 | } | |
1647 | ||
1648 | ||
1649 | /* inherently racy... | |
1650 | * return value may be already out-of-date when this function returns. | |
1651 | * but the general usage is that this is only use during a cstate when bits are | |
1652 | * only cleared, not set, and typically only care for the case when the return | |
1653 | * value is zero, or we already "locked" this "bitmap extent" by other means. | |
1654 | * | |
1655 | * enr is bm-extent number, since we chose to name one sector (512 bytes) | |
1656 | * worth of the bitmap a "bitmap extent". | |
1657 | * | |
1658 | * TODO | |
1659 | * I think since we use it like a reference count, we should use the real | |
1660 | * reference count of some bitmap extent element from some lru instead... | |
1661 | * | |
1662 | */ | |
b30ab791 | 1663 | int drbd_bm_e_weight(struct drbd_device *device, unsigned long enr) |
b411b363 | 1664 | { |
b30ab791 | 1665 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1666 | int count, s, e; |
1667 | unsigned long flags; | |
1668 | unsigned long *p_addr, *bm; | |
1669 | ||
841ce241 AG |
1670 | if (!expect(b)) |
1671 | return 0; | |
1672 | if (!expect(b->bm_pages)) | |
1673 | return 0; | |
b411b363 PR |
1674 | |
1675 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1676 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1677 | bm_print_lock_info(device); |
b411b363 PR |
1678 | |
1679 | s = S2W(enr); | |
1680 | e = min((size_t)S2W(enr+1), b->bm_words); | |
1681 | count = 0; | |
1682 | if (s < b->bm_words) { | |
1683 | int n = e-s; | |
19f843aa | 1684 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); |
b411b363 PR |
1685 | bm = p_addr + MLPP(s); |
1686 | while (n--) | |
1687 | count += hweight_long(*bm++); | |
1688 | bm_unmap(p_addr); | |
1689 | } else { | |
d0180171 | 1690 | drbd_err(device, "start offset (%d) too large in drbd_bm_e_weight\n", s); |
b411b363 PR |
1691 | } |
1692 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1693 | return count; | |
1694 | } |