]>
Commit | Line | Data |
---|---|---|
c6ae4c04 | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
b411b363 PR |
2 | /* |
3 | drbd_bitmap.c | |
4 | ||
5 | This file is part of DRBD by Philipp Reisner and Lars Ellenberg. | |
6 | ||
7 | Copyright (C) 2004-2008, LINBIT Information Technologies GmbH. | |
8 | Copyright (C) 2004-2008, Philipp Reisner <philipp.reisner@linbit.com>. | |
9 | Copyright (C) 2004-2008, Lars Ellenberg <lars.ellenberg@linbit.com>. | |
10 | ||
b411b363 PR |
11 | */ |
12 | ||
f88c5d90 LE |
13 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
14 | ||
5fb3bc4d | 15 | #include <linux/bitmap.h> |
b411b363 PR |
16 | #include <linux/vmalloc.h> |
17 | #include <linux/string.h> | |
18 | #include <linux/drbd.h> | |
5a0e3ad6 | 19 | #include <linux/slab.h> |
dbcbdc43 | 20 | #include <linux/highmem.h> |
f0ff1357 | 21 | |
b411b363 PR |
22 | #include "drbd_int.h" |
23 | ||
95a0f10c | 24 | |
b411b363 PR |
25 | /* OPAQUE outside this file! |
26 | * interface defined in drbd_int.h | |
27 | ||
28 | * convention: | |
29 | * function name drbd_bm_... => used elsewhere, "public". | |
30 | * function name bm_... => internal to implementation, "private". | |
4b0715f0 LE |
31 | */ |
32 | ||
b411b363 | 33 | |
4b0715f0 LE |
34 | /* |
35 | * LIMITATIONS: | |
36 | * We want to support >= peta byte of backend storage, while for now still using | |
37 | * a granularity of one bit per 4KiB of storage. | |
38 | * 1 << 50 bytes backend storage (1 PiB) | |
39 | * 1 << (50 - 12) bits needed | |
40 | * 38 --> we need u64 to index and count bits | |
41 | * 1 << (38 - 3) bitmap bytes needed | |
42 | * 35 --> we still need u64 to index and count bytes | |
43 | * (that's 32 GiB of bitmap for 1 PiB storage) | |
44 | * 1 << (35 - 2) 32bit longs needed | |
45 | * 33 --> we'd even need u64 to index and count 32bit long words. | |
46 | * 1 << (35 - 3) 64bit longs needed | |
47 | * 32 --> we could get away with a 32bit unsigned int to index and count | |
48 | * 64bit long words, but I rather stay with unsigned long for now. | |
49 | * We probably should neither count nor point to bytes or long words | |
50 | * directly, but either by bitnumber, or by page index and offset. | |
51 | * 1 << (35 - 12) | |
52 | * 22 --> we need that much 4KiB pages of bitmap. | |
53 | * 1 << (22 + 3) --> on a 64bit arch, | |
54 | * we need 32 MiB to store the array of page pointers. | |
55 | * | |
56 | * Because I'm lazy, and because the resulting patch was too large, too ugly | |
57 | * and still incomplete, on 32bit we still "only" support 16 TiB (minus some), | |
58 | * (1 << 32) bits * 4k storage. | |
59 | * | |
60 | ||
61 | * bitmap storage and IO: | |
62 | * Bitmap is stored little endian on disk, and is kept little endian in | |
63 | * core memory. Currently we still hold the full bitmap in core as long | |
64 | * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage | |
65 | * seems excessive. | |
66 | * | |
24c4830c | 67 | * We plan to reduce the amount of in-core bitmap pages by paging them in |
4b0715f0 LE |
68 | * and out against their on-disk location as necessary, but need to make |
69 | * sure we don't cause too much meta data IO, and must not deadlock in | |
70 | * tight memory situations. This needs some more work. | |
b411b363 PR |
71 | */ |
72 | ||
73 | /* | |
74 | * NOTE | |
75 | * Access to the *bm_pages is protected by bm_lock. | |
76 | * It is safe to read the other members within the lock. | |
77 | * | |
78 | * drbd_bm_set_bits is called from bio_endio callbacks, | |
79 | * We may be called with irq already disabled, | |
80 | * so we need spin_lock_irqsave(). | |
81 | * And we need the kmap_atomic. | |
82 | */ | |
83 | struct drbd_bitmap { | |
84 | struct page **bm_pages; | |
85 | spinlock_t bm_lock; | |
4b0715f0 | 86 | |
27ea1d87 LE |
87 | /* exclusively to be used by __al_write_transaction(), |
88 | * drbd_bm_mark_for_writeout() and | |
89 | * and drbd_bm_write_hinted() -> bm_rw() called from there. | |
90 | */ | |
91 | unsigned int n_bitmap_hints; | |
92 | unsigned int al_bitmap_hints[AL_UPDATES_PER_TRANSACTION]; | |
93 | ||
4b0715f0 LE |
94 | /* see LIMITATIONS: above */ |
95 | ||
b411b363 PR |
96 | unsigned long bm_set; /* nr of set bits; THINK maybe atomic_t? */ |
97 | unsigned long bm_bits; | |
98 | size_t bm_words; | |
99 | size_t bm_number_of_pages; | |
100 | sector_t bm_dev_capacity; | |
8a03ae2a | 101 | struct mutex bm_change; /* serializes resize operations */ |
b411b363 | 102 | |
19f843aa | 103 | wait_queue_head_t bm_io_wait; /* used to serialize IO of single pages */ |
b411b363 | 104 | |
20ceb2b2 | 105 | enum bm_flag bm_flags; |
b411b363 PR |
106 | |
107 | /* debugging aid, in case we are still racy somewhere */ | |
108 | char *bm_why; | |
109 | struct task_struct *bm_task; | |
110 | }; | |
111 | ||
b411b363 | 112 | #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__) |
b30ab791 | 113 | static void __bm_print_lock_info(struct drbd_device *device, const char *func) |
b411b363 | 114 | { |
b30ab791 | 115 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
116 | if (!__ratelimit(&drbd_ratelimit_state)) |
117 | return; | |
c60b0251 AG |
118 | drbd_err(device, "FIXME %s[%d] in %s, bitmap locked for '%s' by %s[%d]\n", |
119 | current->comm, task_pid_nr(current), | |
120 | func, b->bm_why ?: "?", | |
121 | b->bm_task->comm, task_pid_nr(b->bm_task)); | |
b411b363 PR |
122 | } |
123 | ||
b30ab791 | 124 | void drbd_bm_lock(struct drbd_device *device, char *why, enum bm_flag flags) |
b411b363 | 125 | { |
b30ab791 | 126 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
127 | int trylock_failed; |
128 | ||
129 | if (!b) { | |
d0180171 | 130 | drbd_err(device, "FIXME no bitmap in drbd_bm_lock!?\n"); |
b411b363 PR |
131 | return; |
132 | } | |
133 | ||
8a03ae2a | 134 | trylock_failed = !mutex_trylock(&b->bm_change); |
b411b363 PR |
135 | |
136 | if (trylock_failed) { | |
c60b0251 AG |
137 | drbd_warn(device, "%s[%d] going to '%s' but bitmap already locked for '%s' by %s[%d]\n", |
138 | current->comm, task_pid_nr(current), | |
139 | why, b->bm_why ?: "?", | |
140 | b->bm_task->comm, task_pid_nr(b->bm_task)); | |
8a03ae2a | 141 | mutex_lock(&b->bm_change); |
b411b363 | 142 | } |
20ceb2b2 | 143 | if (BM_LOCKED_MASK & b->bm_flags) |
d0180171 | 144 | drbd_err(device, "FIXME bitmap already locked in bm_lock\n"); |
20ceb2b2 | 145 | b->bm_flags |= flags & BM_LOCKED_MASK; |
b411b363 PR |
146 | |
147 | b->bm_why = why; | |
148 | b->bm_task = current; | |
149 | } | |
150 | ||
b30ab791 | 151 | void drbd_bm_unlock(struct drbd_device *device) |
b411b363 | 152 | { |
b30ab791 | 153 | struct drbd_bitmap *b = device->bitmap; |
b411b363 | 154 | if (!b) { |
d0180171 | 155 | drbd_err(device, "FIXME no bitmap in drbd_bm_unlock!?\n"); |
b411b363 PR |
156 | return; |
157 | } | |
158 | ||
b30ab791 | 159 | if (!(BM_LOCKED_MASK & device->bitmap->bm_flags)) |
d0180171 | 160 | drbd_err(device, "FIXME bitmap not locked in bm_unlock\n"); |
b411b363 | 161 | |
20ceb2b2 | 162 | b->bm_flags &= ~BM_LOCKED_MASK; |
b411b363 PR |
163 | b->bm_why = NULL; |
164 | b->bm_task = NULL; | |
8a03ae2a | 165 | mutex_unlock(&b->bm_change); |
b411b363 PR |
166 | } |
167 | ||
19f843aa LE |
168 | /* we store some "meta" info about our pages in page->private */ |
169 | /* at a granularity of 4k storage per bitmap bit: | |
170 | * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks | |
171 | * 1<<38 bits, | |
172 | * 1<<23 4k bitmap pages. | |
173 | * Use 24 bits as page index, covers 2 peta byte storage | |
174 | * at a granularity of 4k per bit. | |
175 | * Used to report the failed page idx on io error from the endio handlers. | |
176 | */ | |
177 | #define BM_PAGE_IDX_MASK ((1UL<<24)-1) | |
178 | /* this page is currently read in, or written back */ | |
179 | #define BM_PAGE_IO_LOCK 31 | |
180 | /* if there has been an IO error for this page */ | |
181 | #define BM_PAGE_IO_ERROR 30 | |
182 | /* this is to be able to intelligently skip disk IO, | |
183 | * set if bits have been set since last IO. */ | |
184 | #define BM_PAGE_NEED_WRITEOUT 29 | |
185 | /* to mark for lazy writeout once syncer cleared all clearable bits, | |
186 | * we if bits have been cleared since last IO. */ | |
187 | #define BM_PAGE_LAZY_WRITEOUT 28 | |
45dfffeb LE |
188 | /* pages marked with this "HINT" will be considered for writeout |
189 | * on activity log transactions */ | |
190 | #define BM_PAGE_HINT_WRITEOUT 27 | |
19f843aa | 191 | |
24c4830c | 192 | /* store_page_idx uses non-atomic assignment. It is only used directly after |
19f843aa LE |
193 | * allocating the page. All other bm_set_page_* and bm_clear_page_* need to |
194 | * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap | |
195 | * changes) may happen from various contexts, and wait_on_bit/wake_up_bit | |
196 | * requires it all to be atomic as well. */ | |
197 | static void bm_store_page_idx(struct page *page, unsigned long idx) | |
95a0f10c | 198 | { |
19f843aa | 199 | BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK)); |
0c7db279 | 200 | set_page_private(page, idx); |
95a0f10c LE |
201 | } |
202 | ||
19f843aa | 203 | static unsigned long bm_page_to_idx(struct page *page) |
b411b363 | 204 | { |
19f843aa LE |
205 | return page_private(page) & BM_PAGE_IDX_MASK; |
206 | } | |
207 | ||
208 | /* As is very unlikely that the same page is under IO from more than one | |
209 | * context, we can get away with a bit per page and one wait queue per bitmap. | |
210 | */ | |
b30ab791 | 211 | static void bm_page_lock_io(struct drbd_device *device, int page_nr) |
19f843aa | 212 | { |
b30ab791 | 213 | struct drbd_bitmap *b = device->bitmap; |
19f843aa LE |
214 | void *addr = &page_private(b->bm_pages[page_nr]); |
215 | wait_event(b->bm_io_wait, !test_and_set_bit(BM_PAGE_IO_LOCK, addr)); | |
216 | } | |
217 | ||
b30ab791 | 218 | static void bm_page_unlock_io(struct drbd_device *device, int page_nr) |
19f843aa | 219 | { |
b30ab791 | 220 | struct drbd_bitmap *b = device->bitmap; |
19f843aa | 221 | void *addr = &page_private(b->bm_pages[page_nr]); |
4738fa16 | 222 | clear_bit_unlock(BM_PAGE_IO_LOCK, addr); |
b30ab791 | 223 | wake_up(&device->bitmap->bm_io_wait); |
19f843aa LE |
224 | } |
225 | ||
226 | /* set _before_ submit_io, so it may be reset due to being changed | |
227 | * while this page is in flight... will get submitted later again */ | |
228 | static void bm_set_page_unchanged(struct page *page) | |
229 | { | |
230 | /* use cmpxchg? */ | |
231 | clear_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
232 | clear_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
233 | } | |
234 | ||
235 | static void bm_set_page_need_writeout(struct page *page) | |
236 | { | |
237 | set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); | |
238 | } | |
239 | ||
27ea1d87 LE |
240 | void drbd_bm_reset_al_hints(struct drbd_device *device) |
241 | { | |
242 | device->bitmap->n_bitmap_hints = 0; | |
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 | 255 | { |
27ea1d87 | 256 | struct drbd_bitmap *b = device->bitmap; |
45dfffeb | 257 | struct page *page; |
b30ab791 | 258 | if (page_nr >= device->bitmap->bm_number_of_pages) { |
d0180171 | 259 | drbd_warn(device, "BAD: page_nr: %u, number_of_pages: %u\n", |
b30ab791 | 260 | page_nr, (int)device->bitmap->bm_number_of_pages); |
45dfffeb LE |
261 | return; |
262 | } | |
b30ab791 | 263 | page = device->bitmap->bm_pages[page_nr]; |
27ea1d87 LE |
264 | BUG_ON(b->n_bitmap_hints >= ARRAY_SIZE(b->al_bitmap_hints)); |
265 | if (!test_and_set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page))) | |
266 | b->al_bitmap_hints[b->n_bitmap_hints++] = page_nr; | |
45dfffeb LE |
267 | } |
268 | ||
19f843aa LE |
269 | static int bm_test_page_unchanged(struct page *page) |
270 | { | |
271 | volatile const unsigned long *addr = &page_private(page); | |
272 | return (*addr & ((1UL<<BM_PAGE_NEED_WRITEOUT)|(1UL<<BM_PAGE_LAZY_WRITEOUT))) == 0; | |
273 | } | |
b411b363 | 274 | |
19f843aa LE |
275 | static void bm_set_page_io_err(struct page *page) |
276 | { | |
277 | set_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
278 | } | |
279 | ||
280 | static void bm_clear_page_io_err(struct page *page) | |
281 | { | |
282 | clear_bit(BM_PAGE_IO_ERROR, &page_private(page)); | |
283 | } | |
284 | ||
285 | static void bm_set_page_lazy_writeout(struct page *page) | |
286 | { | |
287 | set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
288 | } | |
289 | ||
290 | static int bm_test_page_lazy_writeout(struct page *page) | |
291 | { | |
292 | return test_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); | |
293 | } | |
294 | ||
295 | /* on a 32bit box, this would allow for exactly (2<<38) bits. */ | |
296 | static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr) | |
297 | { | |
b411b363 | 298 | /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */ |
19f843aa | 299 | unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3); |
b411b363 | 300 | BUG_ON(page_nr >= b->bm_number_of_pages); |
19f843aa LE |
301 | return page_nr; |
302 | } | |
b411b363 | 303 | |
19f843aa LE |
304 | static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr) |
305 | { | |
306 | /* page_nr = (bitnr/8) >> PAGE_SHIFT; */ | |
307 | unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3); | |
308 | BUG_ON(page_nr >= b->bm_number_of_pages); | |
309 | return page_nr; | |
b411b363 PR |
310 | } |
311 | ||
589973a7 | 312 | static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) |
95a0f10c LE |
313 | { |
314 | struct page *page = b->bm_pages[idx]; | |
cfd8005c | 315 | return (unsigned long *) kmap_atomic(page); |
95a0f10c LE |
316 | } |
317 | ||
318 | static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) | |
319 | { | |
cfd8005c | 320 | return __bm_map_pidx(b, idx); |
95a0f10c LE |
321 | } |
322 | ||
cfd8005c | 323 | static void __bm_unmap(unsigned long *p_addr) |
b411b363 | 324 | { |
cfd8005c | 325 | kunmap_atomic(p_addr); |
b411b363 PR |
326 | }; |
327 | ||
328 | static void bm_unmap(unsigned long *p_addr) | |
329 | { | |
cfd8005c | 330 | return __bm_unmap(p_addr); |
b411b363 PR |
331 | } |
332 | ||
333 | /* long word offset of _bitmap_ sector */ | |
334 | #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) | |
335 | /* word offset from start of bitmap to word number _in_page_ | |
336 | * modulo longs per page | |
337 | #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long)) | |
24c4830c | 338 | hm, well, Philipp thinks gcc might not optimize the % into & (... - 1) |
b411b363 PR |
339 | so do it explicitly: |
340 | */ | |
341 | #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1)) | |
342 | ||
343 | /* Long words per page */ | |
344 | #define LWPP (PAGE_SIZE/sizeof(long)) | |
345 | ||
346 | /* | |
347 | * actually most functions herein should take a struct drbd_bitmap*, not a | |
b30ab791 | 348 | * struct drbd_device*, but for the debug macros I like to have the device around |
b411b363 PR |
349 | * to be able to report device specific. |
350 | */ | |
351 | ||
19f843aa | 352 | |
b411b363 PR |
353 | static void bm_free_pages(struct page **pages, unsigned long number) |
354 | { | |
355 | unsigned long i; | |
356 | if (!pages) | |
357 | return; | |
358 | ||
359 | for (i = 0; i < number; i++) { | |
360 | if (!pages[i]) { | |
f88c5d90 LE |
361 | pr_alert("bm_free_pages tried to free a NULL pointer; i=%lu n=%lu\n", |
362 | i, number); | |
b411b363 PR |
363 | continue; |
364 | } | |
365 | __free_page(pages[i]); | |
366 | pages[i] = NULL; | |
367 | } | |
368 | } | |
369 | ||
1d5cfdb0 | 370 | static inline void bm_vk_free(void *ptr) |
b411b363 | 371 | { |
1d5cfdb0 | 372 | kvfree(ptr); |
b411b363 PR |
373 | } |
374 | ||
375 | /* | |
376 | * "have" and "want" are NUMBER OF PAGES. | |
377 | */ | |
378 | static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want) | |
379 | { | |
380 | struct page **old_pages = b->bm_pages; | |
381 | struct page **new_pages, *page; | |
1d5cfdb0 | 382 | unsigned int i, bytes; |
b411b363 PR |
383 | unsigned long have = b->bm_number_of_pages; |
384 | ||
385 | BUG_ON(have == 0 && old_pages != NULL); | |
386 | BUG_ON(have != 0 && old_pages == NULL); | |
387 | ||
388 | if (have == want) | |
389 | return old_pages; | |
390 | ||
391 | /* Trying kmalloc first, falling back to vmalloc. | |
0b143d43 LE |
392 | * GFP_NOIO, as this is called while drbd IO is "suspended", |
393 | * and during resize or attach on diskless Primary, | |
394 | * we must not block on IO to ourselves. | |
bc891c9a | 395 | * Context is receiver thread or dmsetup. */ |
b411b363 | 396 | bytes = sizeof(struct page *)*want; |
8be04b93 | 397 | new_pages = kzalloc(bytes, GFP_NOIO | __GFP_NOWARN); |
b411b363 | 398 | if (!new_pages) { |
0b143d43 | 399 | new_pages = __vmalloc(bytes, |
19809c2d | 400 | GFP_NOIO | __GFP_ZERO, |
0b143d43 | 401 | PAGE_KERNEL); |
b411b363 PR |
402 | if (!new_pages) |
403 | return NULL; | |
b411b363 PR |
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); | |
1d5cfdb0 | 413 | bm_vk_free(new_pages); |
b411b363 PR |
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 | ||
b411b363 PR |
429 | return new_pages; |
430 | } | |
431 | ||
432 | /* | |
7e5fec31 | 433 | * allocates the drbd_bitmap and stores it in device->bitmap. |
b411b363 | 434 | */ |
b30ab791 | 435 | int drbd_bm_init(struct drbd_device *device) |
b411b363 | 436 | { |
b30ab791 | 437 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
438 | WARN_ON(b != NULL); |
439 | b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL); | |
440 | if (!b) | |
441 | return -ENOMEM; | |
442 | spin_lock_init(&b->bm_lock); | |
8a03ae2a | 443 | mutex_init(&b->bm_change); |
b411b363 PR |
444 | init_waitqueue_head(&b->bm_io_wait); |
445 | ||
b30ab791 | 446 | device->bitmap = b; |
b411b363 PR |
447 | |
448 | return 0; | |
449 | } | |
450 | ||
b30ab791 | 451 | sector_t drbd_bm_capacity(struct drbd_device *device) |
b411b363 | 452 | { |
b30ab791 | 453 | if (!expect(device->bitmap)) |
841ce241 | 454 | return 0; |
b30ab791 | 455 | return device->bitmap->bm_dev_capacity; |
b411b363 PR |
456 | } |
457 | ||
458 | /* called on driver unload. TODO: call when a device is destroyed. | |
459 | */ | |
b30ab791 | 460 | void drbd_bm_cleanup(struct drbd_device *device) |
b411b363 | 461 | { |
b30ab791 | 462 | if (!expect(device->bitmap)) |
841ce241 | 463 | return; |
b30ab791 | 464 | bm_free_pages(device->bitmap->bm_pages, device->bitmap->bm_number_of_pages); |
1d5cfdb0 | 465 | bm_vk_free(device->bitmap->bm_pages); |
b30ab791 AG |
466 | kfree(device->bitmap); |
467 | device->bitmap = NULL; | |
b411b363 PR |
468 | } |
469 | ||
470 | /* | |
471 | * since (b->bm_bits % BITS_PER_LONG) != 0, | |
472 | * this masks out the remaining bits. | |
473 | * Returns the number of bits cleared. | |
474 | */ | |
2630628b | 475 | #ifndef BITS_PER_PAGE |
95a0f10c LE |
476 | #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3)) |
477 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1) | |
2630628b LE |
478 | #else |
479 | # if BITS_PER_PAGE != (1UL << (PAGE_SHIFT + 3)) | |
480 | # error "ambiguous BITS_PER_PAGE" | |
481 | # endif | |
482 | #endif | |
95a0f10c | 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; |
5fb3bc4d | 561 | int idx, 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); |
5fb3bc4d | 566 | bits += bitmap_weight(p_addr, BITS_PER_PAGE); |
cfd8005c | 567 | __bm_unmap(p_addr); |
b411b363 PR |
568 | cond_resched(); |
569 | } | |
4b0715f0 LE |
570 | /* last (or only) page */ |
571 | last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL; | |
589973a7 | 572 | p_addr = __bm_map_pidx(b, idx); |
5fb3bc4d | 573 | bits += bitmap_weight(p_addr, last_word * BITS_PER_LONG); |
4b0715f0 LE |
574 | p_addr[last_word] &= cpu_to_lel(mask); |
575 | bits += hweight_long(p_addr[last_word]); | |
576 | /* 32bit arch, may have an unused padding long */ | |
577 | if (BITS_PER_LONG == 32 && (last_word & 1) == 0) | |
578 | p_addr[last_word+1] = 0; | |
589973a7 | 579 | __bm_unmap(p_addr); |
b411b363 PR |
580 | return bits; |
581 | } | |
582 | ||
b411b363 PR |
583 | /* offset and len in long words.*/ |
584 | static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len) | |
585 | { | |
586 | unsigned long *p_addr, *bm; | |
19f843aa | 587 | unsigned int idx; |
b411b363 PR |
588 | size_t do_now, end; |
589 | ||
b411b363 PR |
590 | end = offset + len; |
591 | ||
592 | if (end > b->bm_words) { | |
f88c5d90 | 593 | pr_alert("bm_memset end > bm_words\n"); |
b411b363 PR |
594 | return; |
595 | } | |
596 | ||
597 | while (offset < end) { | |
598 | do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset; | |
19f843aa LE |
599 | idx = bm_word_to_page_idx(b, offset); |
600 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
601 | bm = p_addr + MLPP(offset); |
602 | if (bm+do_now > p_addr + LWPP) { | |
f88c5d90 | 603 | pr_alert("BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n", |
b411b363 | 604 | p_addr, bm, (int)do_now); |
84e7c0f7 LE |
605 | } else |
606 | memset(bm, c, do_now * sizeof(long)); | |
b411b363 | 607 | bm_unmap(p_addr); |
19f843aa | 608 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
609 | offset += do_now; |
610 | } | |
611 | } | |
612 | ||
ae8bf312 LE |
613 | /* For the layout, see comment above drbd_md_set_sector_offsets(). */ |
614 | static u64 drbd_md_on_disk_bits(struct drbd_backing_dev *ldev) | |
615 | { | |
616 | u64 bitmap_sectors; | |
617 | if (ldev->md.al_offset == 8) | |
618 | bitmap_sectors = ldev->md.md_size_sect - ldev->md.bm_offset; | |
619 | else | |
620 | bitmap_sectors = ldev->md.al_offset - ldev->md.bm_offset; | |
621 | return bitmap_sectors << (9 + 3); | |
622 | } | |
623 | ||
b411b363 PR |
624 | /* |
625 | * make sure the bitmap has enough room for the attached storage, | |
626 | * if necessary, resize. | |
627 | * called whenever we may have changed the device size. | |
628 | * returns -ENOMEM if we could not allocate enough memory, 0 on success. | |
629 | * In case this is actually a resize, we copy the old bitmap into the new one. | |
630 | * Otherwise, the bitmap is initialized to all bits set. | |
631 | */ | |
b30ab791 | 632 | int drbd_bm_resize(struct drbd_device *device, sector_t capacity, int set_new_bits) |
b411b363 | 633 | { |
b30ab791 | 634 | struct drbd_bitmap *b = device->bitmap; |
6850c442 | 635 | unsigned long bits, words, owords, obits; |
b411b363 PR |
636 | unsigned long want, have, onpages; /* number of pages */ |
637 | struct page **npages, **opages = NULL; | |
7e5fec31 FF |
638 | int err = 0; |
639 | bool growing; | |
b411b363 | 640 | |
841ce241 AG |
641 | if (!expect(b)) |
642 | return -ENOMEM; | |
b411b363 | 643 | |
b30ab791 | 644 | drbd_bm_lock(device, "resize", BM_LOCKED_MASK); |
b411b363 | 645 | |
d0180171 | 646 | drbd_info(device, "drbd_bm_resize called with capacity == %llu\n", |
b411b363 PR |
647 | (unsigned long long)capacity); |
648 | ||
649 | if (capacity == b->bm_dev_capacity) | |
650 | goto out; | |
651 | ||
b411b363 PR |
652 | if (capacity == 0) { |
653 | spin_lock_irq(&b->bm_lock); | |
654 | opages = b->bm_pages; | |
655 | onpages = b->bm_number_of_pages; | |
656 | owords = b->bm_words; | |
657 | b->bm_pages = NULL; | |
658 | b->bm_number_of_pages = | |
659 | b->bm_set = | |
660 | b->bm_bits = | |
661 | b->bm_words = | |
662 | b->bm_dev_capacity = 0; | |
663 | spin_unlock_irq(&b->bm_lock); | |
664 | bm_free_pages(opages, onpages); | |
1d5cfdb0 | 665 | bm_vk_free(opages); |
b411b363 PR |
666 | goto out; |
667 | } | |
668 | bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT)); | |
669 | ||
670 | /* if we would use | |
671 | words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL; | |
672 | a 32bit host could present the wrong number of words | |
673 | to a 64bit host. | |
674 | */ | |
675 | words = ALIGN(bits, 64) >> LN2_BPL; | |
676 | ||
b30ab791 AG |
677 | if (get_ldev(device)) { |
678 | u64 bits_on_disk = drbd_md_on_disk_bits(device->ldev); | |
679 | put_ldev(device); | |
4b0715f0 | 680 | if (bits > bits_on_disk) { |
d0180171 AG |
681 | drbd_info(device, "bits = %lu\n", bits); |
682 | drbd_info(device, "bits_on_disk = %llu\n", bits_on_disk); | |
4b0715f0 LE |
683 | err = -ENOSPC; |
684 | goto out; | |
685 | } | |
b411b363 PR |
686 | } |
687 | ||
6850c442 | 688 | want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT; |
b411b363 PR |
689 | have = b->bm_number_of_pages; |
690 | if (want == have) { | |
0b0ba1ef | 691 | D_ASSERT(device, b->bm_pages != NULL); |
b411b363 PR |
692 | npages = b->bm_pages; |
693 | } else { | |
b30ab791 | 694 | if (drbd_insert_fault(device, DRBD_FAULT_BM_ALLOC)) |
b411b363 PR |
695 | npages = NULL; |
696 | else | |
697 | npages = bm_realloc_pages(b, want); | |
698 | } | |
699 | ||
700 | if (!npages) { | |
701 | err = -ENOMEM; | |
702 | goto out; | |
703 | } | |
704 | ||
705 | spin_lock_irq(&b->bm_lock); | |
706 | opages = b->bm_pages; | |
707 | owords = b->bm_words; | |
708 | obits = b->bm_bits; | |
709 | ||
710 | growing = bits > obits; | |
5223671b | 711 | if (opages && growing && set_new_bits) |
b411b363 PR |
712 | bm_set_surplus(b); |
713 | ||
714 | b->bm_pages = npages; | |
715 | b->bm_number_of_pages = want; | |
716 | b->bm_bits = bits; | |
717 | b->bm_words = words; | |
718 | b->bm_dev_capacity = capacity; | |
719 | ||
720 | if (growing) { | |
02d9a94b PR |
721 | if (set_new_bits) { |
722 | bm_memset(b, owords, 0xff, words-owords); | |
723 | b->bm_set += bits - obits; | |
724 | } else | |
725 | bm_memset(b, owords, 0x00, words-owords); | |
726 | ||
b411b363 PR |
727 | } |
728 | ||
729 | if (want < have) { | |
730 | /* implicit: (opages != NULL) && (opages != npages) */ | |
731 | bm_free_pages(opages + want, have - want); | |
732 | } | |
733 | ||
b411b363 PR |
734 | (void)bm_clear_surplus(b); |
735 | ||
736 | spin_unlock_irq(&b->bm_lock); | |
737 | if (opages != npages) | |
1d5cfdb0 | 738 | bm_vk_free(opages); |
b411b363 PR |
739 | if (!growing) |
740 | b->bm_set = bm_count_bits(b); | |
d0180171 | 741 | drbd_info(device, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want); |
b411b363 PR |
742 | |
743 | out: | |
b30ab791 | 744 | drbd_bm_unlock(device); |
b411b363 PR |
745 | return err; |
746 | } | |
747 | ||
748 | /* inherently racy: | |
749 | * if not protected by other means, return value may be out of date when | |
750 | * leaving this function... | |
751 | * we still need to lock it, since it is important that this returns | |
752 | * bm_set == 0 precisely. | |
753 | * | |
754 | * maybe bm_set should be atomic_t ? | |
755 | */ | |
b30ab791 | 756 | unsigned long _drbd_bm_total_weight(struct drbd_device *device) |
b411b363 | 757 | { |
b30ab791 | 758 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
759 | unsigned long s; |
760 | unsigned long flags; | |
761 | ||
841ce241 AG |
762 | if (!expect(b)) |
763 | return 0; | |
764 | if (!expect(b->bm_pages)) | |
765 | return 0; | |
b411b363 PR |
766 | |
767 | spin_lock_irqsave(&b->bm_lock, flags); | |
768 | s = b->bm_set; | |
769 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
770 | ||
771 | return s; | |
772 | } | |
773 | ||
b30ab791 | 774 | unsigned long drbd_bm_total_weight(struct drbd_device *device) |
b411b363 PR |
775 | { |
776 | unsigned long s; | |
777 | /* if I don't have a disk, I don't know about out-of-sync status */ | |
b30ab791 | 778 | if (!get_ldev_if_state(device, D_NEGOTIATING)) |
b411b363 | 779 | return 0; |
b30ab791 AG |
780 | s = _drbd_bm_total_weight(device); |
781 | put_ldev(device); | |
b411b363 PR |
782 | return s; |
783 | } | |
784 | ||
b30ab791 | 785 | size_t drbd_bm_words(struct drbd_device *device) |
b411b363 | 786 | { |
b30ab791 | 787 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
788 | if (!expect(b)) |
789 | return 0; | |
790 | if (!expect(b->bm_pages)) | |
791 | return 0; | |
b411b363 PR |
792 | |
793 | return b->bm_words; | |
794 | } | |
795 | ||
b30ab791 | 796 | unsigned long drbd_bm_bits(struct drbd_device *device) |
b411b363 | 797 | { |
b30ab791 | 798 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
799 | if (!expect(b)) |
800 | return 0; | |
b411b363 PR |
801 | |
802 | return b->bm_bits; | |
803 | } | |
804 | ||
805 | /* merge number words from buffer into the bitmap starting at offset. | |
806 | * buffer[i] is expected to be little endian unsigned long. | |
807 | * bitmap must be locked by drbd_bm_lock. | |
808 | * currently only used from receive_bitmap. | |
809 | */ | |
b30ab791 | 810 | void drbd_bm_merge_lel(struct drbd_device *device, size_t offset, size_t number, |
b411b363 PR |
811 | unsigned long *buffer) |
812 | { | |
b30ab791 | 813 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
814 | unsigned long *p_addr, *bm; |
815 | unsigned long word, bits; | |
19f843aa | 816 | unsigned int idx; |
b411b363 PR |
817 | size_t end, do_now; |
818 | ||
819 | end = offset + number; | |
820 | ||
841ce241 AG |
821 | if (!expect(b)) |
822 | return; | |
823 | if (!expect(b->bm_pages)) | |
824 | return; | |
b411b363 PR |
825 | if (number == 0) |
826 | return; | |
827 | WARN_ON(offset >= b->bm_words); | |
828 | WARN_ON(end > b->bm_words); | |
829 | ||
830 | spin_lock_irq(&b->bm_lock); | |
831 | while (offset < end) { | |
832 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa LE |
833 | idx = bm_word_to_page_idx(b, offset); |
834 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
835 | bm = p_addr + MLPP(offset); |
836 | offset += do_now; | |
837 | while (do_now--) { | |
838 | bits = hweight_long(*bm); | |
95a0f10c | 839 | word = *bm | *buffer++; |
b411b363 PR |
840 | *bm++ = word; |
841 | b->bm_set += hweight_long(word) - bits; | |
842 | } | |
843 | bm_unmap(p_addr); | |
19f843aa | 844 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
845 | } |
846 | /* with 32bit <-> 64bit cross-platform connect | |
847 | * this is only correct for current usage, | |
848 | * where we _know_ that we are 64 bit aligned, | |
849 | * and know that this function is used in this way, too... | |
850 | */ | |
851 | if (end == b->bm_words) | |
852 | b->bm_set -= bm_clear_surplus(b); | |
b411b363 PR |
853 | spin_unlock_irq(&b->bm_lock); |
854 | } | |
855 | ||
856 | /* copy number words from the bitmap starting at offset into the buffer. | |
857 | * buffer[i] will be little endian unsigned long. | |
858 | */ | |
b30ab791 | 859 | void drbd_bm_get_lel(struct drbd_device *device, size_t offset, size_t number, |
b411b363 PR |
860 | unsigned long *buffer) |
861 | { | |
b30ab791 | 862 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
863 | unsigned long *p_addr, *bm; |
864 | size_t end, do_now; | |
865 | ||
866 | end = offset + number; | |
867 | ||
841ce241 AG |
868 | if (!expect(b)) |
869 | return; | |
870 | if (!expect(b->bm_pages)) | |
871 | return; | |
b411b363 PR |
872 | |
873 | spin_lock_irq(&b->bm_lock); | |
874 | if ((offset >= b->bm_words) || | |
875 | (end > b->bm_words) || | |
876 | (number <= 0)) | |
d0180171 | 877 | drbd_err(device, "offset=%lu number=%lu bm_words=%lu\n", |
b411b363 PR |
878 | (unsigned long) offset, |
879 | (unsigned long) number, | |
880 | (unsigned long) b->bm_words); | |
881 | else { | |
882 | while (offset < end) { | |
883 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa | 884 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset)); |
b411b363 PR |
885 | bm = p_addr + MLPP(offset); |
886 | offset += do_now; | |
887 | while (do_now--) | |
95a0f10c | 888 | *buffer++ = *bm++; |
b411b363 PR |
889 | bm_unmap(p_addr); |
890 | } | |
891 | } | |
892 | spin_unlock_irq(&b->bm_lock); | |
893 | } | |
894 | ||
895 | /* set all bits in the bitmap */ | |
b30ab791 | 896 | void drbd_bm_set_all(struct drbd_device *device) |
b411b363 | 897 | { |
b30ab791 | 898 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
899 | if (!expect(b)) |
900 | return; | |
901 | if (!expect(b->bm_pages)) | |
902 | return; | |
b411b363 PR |
903 | |
904 | spin_lock_irq(&b->bm_lock); | |
905 | bm_memset(b, 0, 0xff, b->bm_words); | |
906 | (void)bm_clear_surplus(b); | |
907 | b->bm_set = b->bm_bits; | |
908 | spin_unlock_irq(&b->bm_lock); | |
909 | } | |
910 | ||
911 | /* clear all bits in the bitmap */ | |
b30ab791 | 912 | void drbd_bm_clear_all(struct drbd_device *device) |
b411b363 | 913 | { |
b30ab791 | 914 | struct drbd_bitmap *b = device->bitmap; |
841ce241 AG |
915 | if (!expect(b)) |
916 | return; | |
917 | if (!expect(b->bm_pages)) | |
918 | return; | |
b411b363 PR |
919 | |
920 | spin_lock_irq(&b->bm_lock); | |
921 | bm_memset(b, 0, 0, b->bm_words); | |
922 | b->bm_set = 0; | |
923 | spin_unlock_irq(&b->bm_lock); | |
924 | } | |
925 | ||
4ce49266 | 926 | static void drbd_bm_aio_ctx_destroy(struct kref *kref) |
d1f3779b | 927 | { |
4ce49266 LE |
928 | struct drbd_bm_aio_ctx *ctx = container_of(kref, struct drbd_bm_aio_ctx, kref); |
929 | unsigned long flags; | |
d1f3779b | 930 | |
4ce49266 LE |
931 | spin_lock_irqsave(&ctx->device->resource->req_lock, flags); |
932 | list_del(&ctx->list); | |
933 | spin_unlock_irqrestore(&ctx->device->resource->req_lock, flags); | |
b30ab791 | 934 | put_ldev(ctx->device); |
d1f3779b PR |
935 | kfree(ctx); |
936 | } | |
937 | ||
19f843aa | 938 | /* bv_page may be a copy, or may be the original */ |
4246a0b6 | 939 | static void drbd_bm_endio(struct bio *bio) |
b411b363 | 940 | { |
4ce49266 | 941 | struct drbd_bm_aio_ctx *ctx = bio->bi_private; |
b30ab791 AG |
942 | struct drbd_device *device = ctx->device; |
943 | struct drbd_bitmap *b = device->bitmap; | |
263663cd | 944 | unsigned int idx = bm_page_to_idx(bio_first_page_all(bio)); |
b411b363 | 945 | |
7648cdfe LE |
946 | if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 && |
947 | !bm_test_page_unchanged(b->bm_pages[idx])) | |
d0180171 | 948 | drbd_warn(device, "bitmap page idx %u changed during IO!\n", idx); |
19f843aa | 949 | |
4e4cbee9 | 950 | if (bio->bi_status) { |
19f843aa LE |
951 | /* ctx error will hold the completed-last non-zero error code, |
952 | * in case error codes differ. */ | |
4e4cbee9 | 953 | ctx->error = blk_status_to_errno(bio->bi_status); |
19f843aa LE |
954 | bm_set_page_io_err(b->bm_pages[idx]); |
955 | /* Not identical to on disk version of it. | |
956 | * Is BM_PAGE_IO_ERROR enough? */ | |
957 | if (__ratelimit(&drbd_ratelimit_state)) | |
d0180171 | 958 | drbd_err(device, "IO ERROR %d on bitmap page idx %u\n", |
4e4cbee9 | 959 | bio->bi_status, idx); |
19f843aa LE |
960 | } else { |
961 | bm_clear_page_io_err(b->bm_pages[idx]); | |
d0180171 | 962 | dynamic_drbd_dbg(device, "bitmap page idx %u completed\n", idx); |
b411b363 | 963 | } |
19f843aa | 964 | |
b30ab791 | 965 | bm_page_unlock_io(device, idx); |
19f843aa | 966 | |
19f843aa | 967 | if (ctx->flags & BM_AIO_COPY_PAGES) |
0892fac8 | 968 | mempool_free(bio->bi_io_vec[0].bv_page, &drbd_md_io_page_pool); |
b411b363 PR |
969 | |
970 | bio_put(bio); | |
19f843aa | 971 | |
d1f3779b | 972 | if (atomic_dec_and_test(&ctx->in_flight)) { |
9e58c4da | 973 | ctx->done = 1; |
b30ab791 | 974 | wake_up(&device->misc_wait); |
4ce49266 | 975 | kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); |
d1f3779b | 976 | } |
b411b363 PR |
977 | } |
978 | ||
4ce49266 | 979 | static void bm_page_io_async(struct drbd_bm_aio_ctx *ctx, int page_nr) __must_hold(local) |
b411b363 | 980 | { |
9476f39d | 981 | struct bio *bio = bio_alloc_drbd(GFP_NOIO); |
b30ab791 AG |
982 | struct drbd_device *device = ctx->device; |
983 | struct drbd_bitmap *b = device->bitmap; | |
19f843aa | 984 | struct page *page; |
b411b363 | 985 | unsigned int len; |
bb3cc85e | 986 | unsigned int op = (ctx->flags & BM_AIO_READ) ? REQ_OP_READ : REQ_OP_WRITE; |
19f843aa | 987 | |
b411b363 | 988 | sector_t on_disk_sector = |
b30ab791 | 989 | device->ldev->md.md_offset + device->ldev->md.bm_offset; |
b411b363 PR |
990 | on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9); |
991 | ||
992 | /* this might happen with very small | |
19f843aa LE |
993 | * flexible external meta data device, |
994 | * or with PAGE_SIZE > 4k */ | |
b411b363 | 995 | len = min_t(unsigned int, PAGE_SIZE, |
b30ab791 | 996 | (drbd_md_last_sector(device->ldev) - on_disk_sector + 1)<<9); |
b411b363 | 997 | |
19f843aa | 998 | /* serialize IO on this page */ |
b30ab791 | 999 | bm_page_lock_io(device, page_nr); |
19f843aa LE |
1000 | /* before memcpy and submit, |
1001 | * so it can be redirtied any time */ | |
1002 | bm_set_page_unchanged(b->bm_pages[page_nr]); | |
1003 | ||
1004 | if (ctx->flags & BM_AIO_COPY_PAGES) { | |
0892fac8 | 1005 | page = mempool_alloc(&drbd_md_io_page_pool, |
0eb0b63c | 1006 | GFP_NOIO | __GFP_HIGHMEM); |
f1d6a328 | 1007 | copy_highpage(page, b->bm_pages[page_nr]); |
19f843aa LE |
1008 | bm_store_page_idx(page, page_nr); |
1009 | } else | |
1010 | page = b->bm_pages[page_nr]; | |
74d46992 | 1011 | bio_set_dev(bio, device->ldev->md_bdev); |
4f024f37 | 1012 | bio->bi_iter.bi_sector = on_disk_sector; |
4d95a10f LE |
1013 | /* bio_add_page of a single page to an empty bio will always succeed, |
1014 | * according to api. Do we want to assert that? */ | |
19f843aa LE |
1015 | bio_add_page(bio, page, len, 0); |
1016 | bio->bi_private = ctx; | |
ed15b795 | 1017 | bio->bi_end_io = drbd_bm_endio; |
bb3cc85e | 1018 | bio_set_op_attrs(bio, op, 0); |
b411b363 | 1019 | |
bb3cc85e | 1020 | if (drbd_insert_fault(device, (op == REQ_OP_WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) { |
4246a0b6 | 1021 | bio_io_error(bio); |
b411b363 | 1022 | } else { |
4e49ea4a | 1023 | submit_bio(bio); |
5a8b4242 LE |
1024 | /* this should not count as user activity and cause the |
1025 | * resync to throttle -- see drbd_rs_should_slow_down(). */ | |
b30ab791 | 1026 | atomic_add(len >> 9, &device->rs_sect_ev); |
b411b363 PR |
1027 | } |
1028 | } | |
1029 | ||
b411b363 PR |
1030 | /* |
1031 | * bm_rw: read/write the whole bitmap from/to its on disk location. | |
1032 | */ | |
4ce49266 | 1033 | static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned lazy_writeout_upper_idx) __must_hold(local) |
b411b363 | 1034 | { |
4ce49266 | 1035 | struct drbd_bm_aio_ctx *ctx; |
b30ab791 | 1036 | struct drbd_bitmap *b = device->bitmap; |
27ea1d87 | 1037 | unsigned int num_pages, i, count = 0; |
b411b363 PR |
1038 | unsigned long now; |
1039 | char ppb[10]; | |
1040 | int err = 0; | |
1041 | ||
19f843aa LE |
1042 | /* |
1043 | * We are protected against bitmap disappearing/resizing by holding an | |
1044 | * ldev reference (caller must have called get_ldev()). | |
1045 | * For read/write, we are protected against changes to the bitmap by | |
1046 | * the bitmap lock (see drbd_bitmap_io). | |
1047 | * For lazy writeout, we don't care for ongoing changes to the bitmap, | |
1048 | * as we submit copies of pages anyways. | |
1049 | */ | |
d1f3779b | 1050 | |
4ce49266 | 1051 | ctx = kmalloc(sizeof(struct drbd_bm_aio_ctx), GFP_NOIO); |
d1f3779b PR |
1052 | if (!ctx) |
1053 | return -ENOMEM; | |
1054 | ||
4ce49266 | 1055 | *ctx = (struct drbd_bm_aio_ctx) { |
b30ab791 | 1056 | .device = device, |
4ce49266 | 1057 | .start_jif = jiffies, |
d1f3779b | 1058 | .in_flight = ATOMIC_INIT(1), |
9e58c4da | 1059 | .done = 0, |
0e8488ad | 1060 | .flags = flags, |
d1f3779b | 1061 | .error = 0, |
1e24edca | 1062 | .kref = KREF_INIT(2), |
d1f3779b PR |
1063 | }; |
1064 | ||
4ce49266 | 1065 | if (!get_ldev_if_state(device, D_ATTACHING)) { /* put is in drbd_bm_aio_ctx_destroy() */ |
d0180171 | 1066 | drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n"); |
9e58c4da PR |
1067 | kfree(ctx); |
1068 | return -ENODEV; | |
1069 | } | |
8fe39aac PR |
1070 | /* Here D_ATTACHING is sufficient since drbd_bm_read() is called only from |
1071 | drbd_adm_attach(), after device->ldev was assigned. */ | |
9e58c4da | 1072 | |
4ce49266 | 1073 | if (0 == (ctx->flags & ~BM_AIO_READ)) |
20ceb2b2 | 1074 | WARN_ON(!(BM_LOCKED_MASK & b->bm_flags)); |
b411b363 | 1075 | |
4ce49266 LE |
1076 | spin_lock_irq(&device->resource->req_lock); |
1077 | list_add_tail(&ctx->list, &device->pending_bitmap_io); | |
1078 | spin_unlock_irq(&device->resource->req_lock); | |
1079 | ||
6850c442 | 1080 | num_pages = b->bm_number_of_pages; |
b411b363 | 1081 | |
b411b363 | 1082 | now = jiffies; |
b411b363 PR |
1083 | |
1084 | /* let the layers below us try to merge these bios... */ | |
fef45d29 | 1085 | |
27ea1d87 LE |
1086 | if (flags & BM_AIO_READ) { |
1087 | for (i = 0; i < num_pages; i++) { | |
1088 | atomic_inc(&ctx->in_flight); | |
1089 | bm_page_io_async(ctx, i); | |
1090 | ++count; | |
1091 | cond_resched(); | |
1092 | } | |
1093 | } else if (flags & BM_AIO_WRITE_HINTED) { | |
1094 | /* ASSERT: BM_AIO_WRITE_ALL_PAGES is not set. */ | |
1095 | unsigned int hint; | |
1096 | for (hint = 0; hint < b->n_bitmap_hints; hint++) { | |
1097 | i = b->al_bitmap_hints[hint]; | |
1098 | if (i >= num_pages) /* == -1U: no hint here. */ | |
1099 | continue; | |
1100 | /* Several AL-extents may point to the same page. */ | |
1101 | if (!test_and_clear_bit(BM_PAGE_HINT_WRITEOUT, | |
1102 | &page_private(b->bm_pages[i]))) | |
1103 | continue; | |
1104 | /* Has it even changed? */ | |
1105 | if (bm_test_page_unchanged(b->bm_pages[i])) | |
1106 | continue; | |
1107 | atomic_inc(&ctx->in_flight); | |
1108 | bm_page_io_async(ctx, i); | |
1109 | ++count; | |
1110 | } | |
1111 | } else { | |
1112 | for (i = 0; i < num_pages; i++) { | |
1113 | /* ignore completely unchanged pages */ | |
1114 | if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx) | |
1115 | break; | |
4ce49266 | 1116 | if (!(flags & BM_AIO_WRITE_ALL_PAGES) && |
d1aa4d04 | 1117 | bm_test_page_unchanged(b->bm_pages[i])) { |
d0180171 | 1118 | dynamic_drbd_dbg(device, "skipped bm write for idx %u\n", i); |
19f843aa LE |
1119 | continue; |
1120 | } | |
1121 | /* during lazy writeout, | |
1122 | * ignore those pages not marked for lazy writeout. */ | |
1123 | if (lazy_writeout_upper_idx && | |
1124 | !bm_test_page_lazy_writeout(b->bm_pages[i])) { | |
d0180171 | 1125 | dynamic_drbd_dbg(device, "skipped bm lazy write for idx %u\n", i); |
19f843aa LE |
1126 | continue; |
1127 | } | |
27ea1d87 LE |
1128 | atomic_inc(&ctx->in_flight); |
1129 | bm_page_io_async(ctx, i); | |
1130 | ++count; | |
1131 | cond_resched(); | |
19f843aa | 1132 | } |
19f843aa | 1133 | } |
b411b363 | 1134 | |
725a97e4 | 1135 | /* |
ed15b795 | 1136 | * We initialize ctx->in_flight to one to make sure drbd_bm_endio |
9e58c4da | 1137 | * will not set ctx->done early, and decrement / test it here. If there |
725a97e4 | 1138 | * are still some bios in flight, we need to wait for them here. |
9e58c4da PR |
1139 | * If all IO is done already (or nothing had been submitted), there is |
1140 | * no need to wait. Still, we need to put the kref associated with the | |
1141 | * "in_flight reached zero, all done" event. | |
725a97e4 | 1142 | */ |
d1f3779b | 1143 | if (!atomic_dec_and_test(&ctx->in_flight)) |
b30ab791 | 1144 | wait_until_done_or_force_detached(device, device->ldev, &ctx->done); |
9e58c4da | 1145 | else |
4ce49266 | 1146 | kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); |
d1f3779b | 1147 | |
c9d963a4 | 1148 | /* summary for global bitmap IO */ |
13c2088d LE |
1149 | if (flags == 0) { |
1150 | unsigned int ms = jiffies_to_msecs(jiffies - now); | |
1151 | if (ms > 5) { | |
1152 | drbd_info(device, "bitmap %s of %u pages took %u ms\n", | |
1153 | (flags & BM_AIO_READ) ? "READ" : "WRITE", | |
1154 | count, ms); | |
1155 | } | |
1156 | } | |
b411b363 | 1157 | |
d1f3779b | 1158 | if (ctx->error) { |
d0180171 | 1159 | drbd_alert(device, "we had at least one MD IO ERROR during bitmap IO\n"); |
b30ab791 | 1160 | drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR); |
d1f3779b | 1161 | err = -EIO; /* ctx->error ? */ |
b411b363 PR |
1162 | } |
1163 | ||
9e58c4da | 1164 | if (atomic_read(&ctx->in_flight)) |
44edfb0d | 1165 | err = -EIO; /* Disk timeout/force-detach during IO... */ |
9e58c4da | 1166 | |
b411b363 | 1167 | now = jiffies; |
4ce49266 | 1168 | if (flags & BM_AIO_READ) { |
95a0f10c | 1169 | b->bm_set = bm_count_bits(b); |
d0180171 | 1170 | drbd_info(device, "recounting of set bits took additional %lu jiffies\n", |
b411b363 PR |
1171 | jiffies - now); |
1172 | } | |
1173 | now = b->bm_set; | |
1174 | ||
4ce49266 | 1175 | if ((flags & ~BM_AIO_READ) == 0) |
d0180171 | 1176 | drbd_info(device, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n", |
c9d963a4 | 1177 | ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now); |
b411b363 | 1178 | |
4ce49266 | 1179 | kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); |
b411b363 PR |
1180 | return err; |
1181 | } | |
1182 | ||
1183 | /** | |
1184 | * drbd_bm_read() - Read the whole bitmap from its on disk location. | |
b30ab791 | 1185 | * @device: DRBD device. |
b411b363 | 1186 | */ |
b30ab791 | 1187 | int drbd_bm_read(struct drbd_device *device) __must_hold(local) |
b411b363 | 1188 | { |
4ce49266 | 1189 | return bm_rw(device, BM_AIO_READ, 0); |
b411b363 PR |
1190 | } |
1191 | ||
1192 | /** | |
1193 | * drbd_bm_write() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1194 | * @device: DRBD device. |
19f843aa LE |
1195 | * |
1196 | * Will only write pages that have changed since last IO. | |
b411b363 | 1197 | */ |
b30ab791 | 1198 | int drbd_bm_write(struct drbd_device *device) __must_hold(local) |
b411b363 | 1199 | { |
4ce49266 | 1200 | return bm_rw(device, 0, 0); |
b411b363 PR |
1201 | } |
1202 | ||
d1aa4d04 PR |
1203 | /** |
1204 | * drbd_bm_write_all() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1205 | * @device: DRBD device. |
d1aa4d04 PR |
1206 | * |
1207 | * Will write all pages. | |
1208 | */ | |
b30ab791 | 1209 | int drbd_bm_write_all(struct drbd_device *device) __must_hold(local) |
d1aa4d04 | 1210 | { |
4ce49266 | 1211 | return bm_rw(device, BM_AIO_WRITE_ALL_PAGES, 0); |
d1aa4d04 PR |
1212 | } |
1213 | ||
c7a58db4 LE |
1214 | /** |
1215 | * drbd_bm_write_lazy() - Write bitmap pages 0 to @upper_idx-1, if they have changed. | |
1216 | * @device: DRBD device. | |
1217 | * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages | |
1218 | */ | |
1219 | int drbd_bm_write_lazy(struct drbd_device *device, unsigned upper_idx) __must_hold(local) | |
1220 | { | |
1221 | return bm_rw(device, BM_AIO_COPY_PAGES, upper_idx); | |
1222 | } | |
1223 | ||
0e8488ad LE |
1224 | /** |
1225 | * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location. | |
b30ab791 | 1226 | * @device: DRBD device. |
0e8488ad LE |
1227 | * |
1228 | * Will only write pages that have changed since last IO. | |
1229 | * In contrast to drbd_bm_write(), this will copy the bitmap pages | |
1230 | * to temporary writeout pages. It is intended to trigger a full write-out | |
1231 | * while still allowing the bitmap to change, for example if a resync or online | |
1232 | * verify is aborted due to a failed peer disk, while local IO continues, or | |
1233 | * pending resync acks are still being processed. | |
1234 | */ | |
b30ab791 | 1235 | int drbd_bm_write_copy_pages(struct drbd_device *device) __must_hold(local) |
0e8488ad | 1236 | { |
4ce49266 | 1237 | return bm_rw(device, BM_AIO_COPY_PAGES, 0); |
19f843aa LE |
1238 | } |
1239 | ||
45dfffeb LE |
1240 | /** |
1241 | * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed. | |
b30ab791 | 1242 | * @device: DRBD device. |
45dfffeb | 1243 | */ |
b30ab791 | 1244 | int drbd_bm_write_hinted(struct drbd_device *device) __must_hold(local) |
45dfffeb | 1245 | { |
4ce49266 | 1246 | return bm_rw(device, BM_AIO_WRITE_HINTED | BM_AIO_COPY_PAGES, 0); |
45dfffeb | 1247 | } |
19f843aa | 1248 | |
b411b363 PR |
1249 | /* NOTE |
1250 | * find_first_bit returns int, we return unsigned long. | |
4b0715f0 LE |
1251 | * For this to work on 32bit arch with bitnumbers > (1<<32), |
1252 | * we'd need to return u64, and get a whole lot of other places | |
1253 | * fixed where we still use unsigned long. | |
b411b363 PR |
1254 | * |
1255 | * this returns a bit number, NOT a sector! | |
1256 | */ | |
b30ab791 | 1257 | static unsigned long __bm_find_next(struct drbd_device *device, unsigned long bm_fo, |
cfd8005c | 1258 | const int find_zero_bit) |
b411b363 | 1259 | { |
b30ab791 | 1260 | struct drbd_bitmap *b = device->bitmap; |
b411b363 | 1261 | unsigned long *p_addr; |
4b0715f0 LE |
1262 | unsigned long bit_offset; |
1263 | unsigned i; | |
1264 | ||
b411b363 PR |
1265 | |
1266 | if (bm_fo > b->bm_bits) { | |
d0180171 | 1267 | drbd_err(device, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits); |
4b0715f0 | 1268 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1269 | } else { |
1270 | while (bm_fo < b->bm_bits) { | |
19f843aa | 1271 | /* bit offset of the first bit in the page */ |
4b0715f0 | 1272 | bit_offset = bm_fo & ~BITS_PER_PAGE_MASK; |
cfd8005c | 1273 | p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo)); |
b411b363 PR |
1274 | |
1275 | if (find_zero_bit) | |
7e599e6e | 1276 | i = find_next_zero_bit_le(p_addr, |
4b0715f0 | 1277 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1278 | else |
7e599e6e | 1279 | i = find_next_bit_le(p_addr, |
4b0715f0 | 1280 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1281 | |
cfd8005c | 1282 | __bm_unmap(p_addr); |
b411b363 | 1283 | if (i < PAGE_SIZE*8) { |
4b0715f0 LE |
1284 | bm_fo = bit_offset + i; |
1285 | if (bm_fo >= b->bm_bits) | |
b411b363 PR |
1286 | break; |
1287 | goto found; | |
1288 | } | |
1289 | bm_fo = bit_offset + PAGE_SIZE*8; | |
1290 | } | |
4b0715f0 | 1291 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1292 | } |
1293 | found: | |
4b0715f0 | 1294 | return bm_fo; |
b411b363 PR |
1295 | } |
1296 | ||
b30ab791 | 1297 | static unsigned long bm_find_next(struct drbd_device *device, |
b411b363 PR |
1298 | unsigned long bm_fo, const int find_zero_bit) |
1299 | { | |
b30ab791 | 1300 | struct drbd_bitmap *b = device->bitmap; |
4b0715f0 | 1301 | unsigned long i = DRBD_END_OF_BITMAP; |
b411b363 | 1302 | |
841ce241 AG |
1303 | if (!expect(b)) |
1304 | return i; | |
1305 | if (!expect(b->bm_pages)) | |
1306 | return i; | |
b411b363 PR |
1307 | |
1308 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1309 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1310 | bm_print_lock_info(device); |
b411b363 | 1311 | |
b30ab791 | 1312 | i = __bm_find_next(device, bm_fo, find_zero_bit); |
b411b363 PR |
1313 | |
1314 | spin_unlock_irq(&b->bm_lock); | |
1315 | return i; | |
1316 | } | |
1317 | ||
b30ab791 | 1318 | unsigned long drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1319 | { |
b30ab791 | 1320 | return bm_find_next(device, bm_fo, 0); |
b411b363 PR |
1321 | } |
1322 | ||
1323 | #if 0 | |
1324 | /* not yet needed for anything. */ | |
b30ab791 | 1325 | unsigned long drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1326 | { |
b30ab791 | 1327 | return bm_find_next(device, bm_fo, 1); |
b411b363 PR |
1328 | } |
1329 | #endif | |
1330 | ||
1331 | /* does not spin_lock_irqsave. | |
1332 | * you must take drbd_bm_lock() first */ | |
b30ab791 | 1333 | unsigned long _drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1334 | { |
b30ab791 AG |
1335 | /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ |
1336 | return __bm_find_next(device, bm_fo, 0); | |
b411b363 PR |
1337 | } |
1338 | ||
b30ab791 | 1339 | unsigned long _drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) |
b411b363 | 1340 | { |
b30ab791 AG |
1341 | /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ |
1342 | return __bm_find_next(device, bm_fo, 1); | |
b411b363 PR |
1343 | } |
1344 | ||
1345 | /* returns number of bits actually changed. | |
1346 | * for val != 0, we change 0 -> 1, return code positive | |
1347 | * for val == 0, we change 1 -> 0, return code negative | |
1348 | * wants bitnr, not sector. | |
1349 | * expected to be called for only a few bits (e - s about BITS_PER_LONG). | |
1350 | * Must hold bitmap lock already. */ | |
b30ab791 | 1351 | static int __bm_change_bits_to(struct drbd_device *device, const unsigned long s, |
829c6087 | 1352 | unsigned long e, int val) |
b411b363 | 1353 | { |
b30ab791 | 1354 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1355 | unsigned long *p_addr = NULL; |
1356 | unsigned long bitnr; | |
19f843aa | 1357 | unsigned int last_page_nr = -1U; |
b411b363 | 1358 | int c = 0; |
19f843aa | 1359 | int changed_total = 0; |
b411b363 PR |
1360 | |
1361 | if (e >= b->bm_bits) { | |
d0180171 | 1362 | drbd_err(device, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n", |
b411b363 PR |
1363 | s, e, b->bm_bits); |
1364 | e = b->bm_bits ? b->bm_bits -1 : 0; | |
1365 | } | |
1366 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa | 1367 | unsigned int page_nr = bm_bit_to_page_idx(b, bitnr); |
b411b363 PR |
1368 | if (page_nr != last_page_nr) { |
1369 | if (p_addr) | |
cfd8005c | 1370 | __bm_unmap(p_addr); |
19f843aa LE |
1371 | if (c < 0) |
1372 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1373 | else if (c > 0) | |
1374 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1375 | changed_total += c; | |
1376 | c = 0; | |
cfd8005c | 1377 | p_addr = __bm_map_pidx(b, page_nr); |
b411b363 PR |
1378 | last_page_nr = page_nr; |
1379 | } | |
1380 | if (val) | |
7e599e6e | 1381 | c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 | 1382 | else |
7e599e6e | 1383 | c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 PR |
1384 | } |
1385 | if (p_addr) | |
cfd8005c | 1386 | __bm_unmap(p_addr); |
19f843aa LE |
1387 | if (c < 0) |
1388 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1389 | else if (c > 0) | |
1390 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1391 | changed_total += c; | |
1392 | b->bm_set += changed_total; | |
1393 | return changed_total; | |
b411b363 PR |
1394 | } |
1395 | ||
1396 | /* returns number of bits actually changed. | |
1397 | * for val != 0, we change 0 -> 1, return code positive | |
1398 | * for val == 0, we change 1 -> 0, return code negative | |
1399 | * wants bitnr, not sector */ | |
b30ab791 | 1400 | static int bm_change_bits_to(struct drbd_device *device, const unsigned long s, |
b411b363 PR |
1401 | const unsigned long e, int val) |
1402 | { | |
1403 | unsigned long flags; | |
b30ab791 | 1404 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1405 | int c = 0; |
1406 | ||
841ce241 AG |
1407 | if (!expect(b)) |
1408 | return 1; | |
1409 | if (!expect(b->bm_pages)) | |
1410 | return 0; | |
b411b363 PR |
1411 | |
1412 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1413 | if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags) |
b30ab791 | 1414 | bm_print_lock_info(device); |
b411b363 | 1415 | |
b30ab791 | 1416 | c = __bm_change_bits_to(device, s, e, val); |
b411b363 PR |
1417 | |
1418 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1419 | return c; | |
1420 | } | |
1421 | ||
1422 | /* returns number of bits changed 0 -> 1 */ | |
b30ab791 | 1423 | int drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 | 1424 | { |
b30ab791 | 1425 | return bm_change_bits_to(device, s, e, 1); |
b411b363 PR |
1426 | } |
1427 | ||
1428 | /* returns number of bits changed 1 -> 0 */ | |
b30ab791 | 1429 | int drbd_bm_clear_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 | 1430 | { |
b30ab791 | 1431 | return -bm_change_bits_to(device, s, e, 0); |
b411b363 PR |
1432 | } |
1433 | ||
1434 | /* sets all bits in full words, | |
1435 | * from first_word up to, but not including, last_word */ | |
1436 | static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b, | |
1437 | int page_nr, int first_word, int last_word) | |
1438 | { | |
1439 | int i; | |
1440 | int bits; | |
22d81140 | 1441 | int changed = 0; |
cfd8005c | 1442 | unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr]); |
5fb3bc4d LE |
1443 | |
1444 | /* I think it is more cache line friendly to hweight_long then set to ~0UL, | |
1445 | * than to first bitmap_weight() all words, then bitmap_fill() all words */ | |
b411b363 PR |
1446 | for (i = first_word; i < last_word; i++) { |
1447 | bits = hweight_long(paddr[i]); | |
1448 | paddr[i] = ~0UL; | |
22d81140 | 1449 | changed += BITS_PER_LONG - bits; |
b411b363 | 1450 | } |
cfd8005c | 1451 | kunmap_atomic(paddr); |
22d81140 LE |
1452 | if (changed) { |
1453 | /* We only need lazy writeout, the information is still in the | |
1454 | * remote bitmap as well, and is reconstructed during the next | |
1455 | * bitmap exchange, if lost locally due to a crash. */ | |
1456 | bm_set_page_lazy_writeout(b->bm_pages[page_nr]); | |
1457 | b->bm_set += changed; | |
1458 | } | |
b411b363 PR |
1459 | } |
1460 | ||
829c6087 LE |
1461 | /* Same thing as drbd_bm_set_bits, |
1462 | * but more efficient for a large bit range. | |
b411b363 PR |
1463 | * You must first drbd_bm_lock(). |
1464 | * Can be called to set the whole bitmap in one go. | |
1465 | * Sets bits from s to e _inclusive_. */ | |
b30ab791 | 1466 | void _drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 PR |
1467 | { |
1468 | /* First set_bit from the first bit (s) | |
1469 | * up to the next long boundary (sl), | |
1470 | * then assign full words up to the last long boundary (el), | |
1471 | * then set_bit up to and including the last bit (e). | |
1472 | * | |
1473 | * Do not use memset, because we must account for changes, | |
1474 | * so we need to loop over the words with hweight() anyways. | |
1475 | */ | |
b30ab791 | 1476 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1477 | unsigned long sl = ALIGN(s,BITS_PER_LONG); |
1478 | unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1); | |
1479 | int first_page; | |
1480 | int last_page; | |
1481 | int page_nr; | |
1482 | int first_word; | |
1483 | int last_word; | |
1484 | ||
1485 | if (e - s <= 3*BITS_PER_LONG) { | |
1486 | /* don't bother; el and sl may even be wrong. */ | |
829c6087 | 1487 | spin_lock_irq(&b->bm_lock); |
b30ab791 | 1488 | __bm_change_bits_to(device, s, e, 1); |
829c6087 | 1489 | spin_unlock_irq(&b->bm_lock); |
b411b363 PR |
1490 | return; |
1491 | } | |
1492 | ||
1493 | /* difference is large enough that we can trust sl and el */ | |
1494 | ||
829c6087 LE |
1495 | spin_lock_irq(&b->bm_lock); |
1496 | ||
b411b363 PR |
1497 | /* bits filling the current long */ |
1498 | if (sl) | |
b30ab791 | 1499 | __bm_change_bits_to(device, s, sl-1, 1); |
b411b363 PR |
1500 | |
1501 | first_page = sl >> (3 + PAGE_SHIFT); | |
1502 | last_page = el >> (3 + PAGE_SHIFT); | |
1503 | ||
1504 | /* MLPP: modulo longs per page */ | |
1505 | /* LWPP: long words per page */ | |
1506 | first_word = MLPP(sl >> LN2_BPL); | |
1507 | last_word = LWPP; | |
1508 | ||
1509 | /* first and full pages, unless first page == last page */ | |
1510 | for (page_nr = first_page; page_nr < last_page; page_nr++) { | |
b30ab791 | 1511 | bm_set_full_words_within_one_page(device->bitmap, page_nr, first_word, last_word); |
8ccee20e LE |
1512 | spin_unlock_irq(&b->bm_lock); |
1513 | cond_resched(); | |
b411b363 | 1514 | first_word = 0; |
8ccee20e | 1515 | spin_lock_irq(&b->bm_lock); |
b411b363 | 1516 | } |
b411b363 PR |
1517 | /* last page (respectively only page, for first page == last page) */ |
1518 | last_word = MLPP(el >> LN2_BPL); | |
4eccc579 LE |
1519 | |
1520 | /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples). | |
1521 | * ==> e = 32767, el = 32768, last_page = 2, | |
1522 | * and now last_word = 0. | |
1523 | * We do not want to touch last_page in this case, | |
1524 | * as we did not allocate it, it is not present in bitmap->bm_pages. | |
1525 | */ | |
1526 | if (last_word) | |
b30ab791 | 1527 | bm_set_full_words_within_one_page(device->bitmap, last_page, first_word, last_word); |
b411b363 PR |
1528 | |
1529 | /* possibly trailing bits. | |
1530 | * example: (e & 63) == 63, el will be e+1. | |
1531 | * if that even was the very last bit, | |
1532 | * it would trigger an assert in __bm_change_bits_to() | |
1533 | */ | |
1534 | if (el <= e) | |
b30ab791 | 1535 | __bm_change_bits_to(device, el, e, 1); |
829c6087 | 1536 | spin_unlock_irq(&b->bm_lock); |
b411b363 PR |
1537 | } |
1538 | ||
1539 | /* returns bit state | |
1540 | * wants bitnr, NOT sector. | |
1541 | * inherently racy... area needs to be locked by means of {al,rs}_lru | |
1542 | * 1 ... bit set | |
1543 | * 0 ... bit not set | |
1544 | * -1 ... first out of bounds access, stop testing for bits! | |
1545 | */ | |
b30ab791 | 1546 | int drbd_bm_test_bit(struct drbd_device *device, const unsigned long bitnr) |
b411b363 PR |
1547 | { |
1548 | unsigned long flags; | |
b30ab791 | 1549 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1550 | unsigned long *p_addr; |
1551 | int i; | |
1552 | ||
841ce241 AG |
1553 | if (!expect(b)) |
1554 | return 0; | |
1555 | if (!expect(b->bm_pages)) | |
1556 | return 0; | |
b411b363 PR |
1557 | |
1558 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1559 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1560 | bm_print_lock_info(device); |
b411b363 | 1561 | if (bitnr < b->bm_bits) { |
19f843aa | 1562 | p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr)); |
7e599e6e | 1563 | i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0; |
b411b363 PR |
1564 | bm_unmap(p_addr); |
1565 | } else if (bitnr == b->bm_bits) { | |
1566 | i = -1; | |
1567 | } else { /* (bitnr > b->bm_bits) */ | |
d0180171 | 1568 | drbd_err(device, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits); |
b411b363 PR |
1569 | i = 0; |
1570 | } | |
1571 | ||
1572 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1573 | return i; | |
1574 | } | |
1575 | ||
1576 | /* returns number of bits set in the range [s, e] */ | |
b30ab791 | 1577 | int drbd_bm_count_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) |
b411b363 PR |
1578 | { |
1579 | unsigned long flags; | |
b30ab791 | 1580 | struct drbd_bitmap *b = device->bitmap; |
19f843aa | 1581 | unsigned long *p_addr = NULL; |
b411b363 | 1582 | unsigned long bitnr; |
19f843aa | 1583 | unsigned int page_nr = -1U; |
b411b363 | 1584 | int c = 0; |
b411b363 PR |
1585 | |
1586 | /* If this is called without a bitmap, that is a bug. But just to be | |
1587 | * robust in case we screwed up elsewhere, in that case pretend there | |
1588 | * was one dirty bit in the requested area, so we won't try to do a | |
1589 | * local read there (no bitmap probably implies no disk) */ | |
841ce241 AG |
1590 | if (!expect(b)) |
1591 | return 1; | |
1592 | if (!expect(b->bm_pages)) | |
1593 | return 1; | |
b411b363 PR |
1594 | |
1595 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1596 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1597 | bm_print_lock_info(device); |
b411b363 | 1598 | for (bitnr = s; bitnr <= e; bitnr++) { |
19f843aa LE |
1599 | unsigned int idx = bm_bit_to_page_idx(b, bitnr); |
1600 | if (page_nr != idx) { | |
1601 | page_nr = idx; | |
b411b363 PR |
1602 | if (p_addr) |
1603 | bm_unmap(p_addr); | |
19f843aa | 1604 | p_addr = bm_map_pidx(b, idx); |
b411b363 | 1605 | } |
841ce241 | 1606 | if (expect(bitnr < b->bm_bits)) |
7e599e6e | 1607 | c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr)); |
841ce241 | 1608 | else |
d0180171 | 1609 | drbd_err(device, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits); |
b411b363 PR |
1610 | } |
1611 | if (p_addr) | |
1612 | bm_unmap(p_addr); | |
1613 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1614 | return c; | |
1615 | } | |
1616 | ||
1617 | ||
1618 | /* inherently racy... | |
1619 | * return value may be already out-of-date when this function returns. | |
1620 | * but the general usage is that this is only use during a cstate when bits are | |
1621 | * only cleared, not set, and typically only care for the case when the return | |
1622 | * value is zero, or we already "locked" this "bitmap extent" by other means. | |
1623 | * | |
1624 | * enr is bm-extent number, since we chose to name one sector (512 bytes) | |
1625 | * worth of the bitmap a "bitmap extent". | |
1626 | * | |
1627 | * TODO | |
1628 | * I think since we use it like a reference count, we should use the real | |
1629 | * reference count of some bitmap extent element from some lru instead... | |
1630 | * | |
1631 | */ | |
b30ab791 | 1632 | int drbd_bm_e_weight(struct drbd_device *device, unsigned long enr) |
b411b363 | 1633 | { |
b30ab791 | 1634 | struct drbd_bitmap *b = device->bitmap; |
b411b363 PR |
1635 | int count, s, e; |
1636 | unsigned long flags; | |
1637 | unsigned long *p_addr, *bm; | |
1638 | ||
841ce241 AG |
1639 | if (!expect(b)) |
1640 | return 0; | |
1641 | if (!expect(b->bm_pages)) | |
1642 | return 0; | |
b411b363 PR |
1643 | |
1644 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1645 | if (BM_DONT_TEST & b->bm_flags) |
b30ab791 | 1646 | bm_print_lock_info(device); |
b411b363 PR |
1647 | |
1648 | s = S2W(enr); | |
1649 | e = min((size_t)S2W(enr+1), b->bm_words); | |
1650 | count = 0; | |
1651 | if (s < b->bm_words) { | |
1652 | int n = e-s; | |
19f843aa | 1653 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); |
b411b363 | 1654 | bm = p_addr + MLPP(s); |
5fb3bc4d | 1655 | count += bitmap_weight(bm, n * BITS_PER_LONG); |
b411b363 PR |
1656 | bm_unmap(p_addr); |
1657 | } else { | |
d0180171 | 1658 | drbd_err(device, "start offset (%d) too large in drbd_bm_e_weight\n", s); |
b411b363 PR |
1659 | } |
1660 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1661 | return count; | |
1662 | } |