Commit | Line | Data |
---|---|---|
b411b363 PR |
1 | /* |
2 | drbd_bitmap.c | |
3 | ||
4 | This file is part of DRBD by Philipp Reisner and Lars Ellenberg. | |
5 | ||
6 | Copyright (C) 2004-2008, LINBIT Information Technologies GmbH. | |
7 | Copyright (C) 2004-2008, Philipp Reisner <philipp.reisner@linbit.com>. | |
8 | Copyright (C) 2004-2008, Lars Ellenberg <lars.ellenberg@linbit.com>. | |
9 | ||
10 | drbd is free software; you can redistribute it and/or modify | |
11 | it under the terms of the GNU General Public License as published by | |
12 | the Free Software Foundation; either version 2, or (at your option) | |
13 | any later version. | |
14 | ||
15 | drbd is distributed in the hope that it will be useful, | |
16 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
18 | GNU General Public License for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
21 | along with drbd; see the file COPYING. If not, write to | |
22 | the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. | |
23 | */ | |
24 | ||
25 | #include <linux/bitops.h> | |
26 | #include <linux/vmalloc.h> | |
27 | #include <linux/string.h> | |
28 | #include <linux/drbd.h> | |
5a0e3ad6 | 29 | #include <linux/slab.h> |
b411b363 | 30 | #include <asm/kmap_types.h> |
f0ff1357 | 31 | |
b411b363 PR |
32 | #include "drbd_int.h" |
33 | ||
95a0f10c | 34 | |
b411b363 PR |
35 | /* OPAQUE outside this file! |
36 | * interface defined in drbd_int.h | |
37 | ||
38 | * convention: | |
39 | * function name drbd_bm_... => used elsewhere, "public". | |
40 | * function name bm_... => internal to implementation, "private". | |
4b0715f0 LE |
41 | */ |
42 | ||
b411b363 | 43 | |
4b0715f0 LE |
44 | /* |
45 | * LIMITATIONS: | |
46 | * We want to support >= peta byte of backend storage, while for now still using | |
47 | * a granularity of one bit per 4KiB of storage. | |
48 | * 1 << 50 bytes backend storage (1 PiB) | |
49 | * 1 << (50 - 12) bits needed | |
50 | * 38 --> we need u64 to index and count bits | |
51 | * 1 << (38 - 3) bitmap bytes needed | |
52 | * 35 --> we still need u64 to index and count bytes | |
53 | * (that's 32 GiB of bitmap for 1 PiB storage) | |
54 | * 1 << (35 - 2) 32bit longs needed | |
55 | * 33 --> we'd even need u64 to index and count 32bit long words. | |
56 | * 1 << (35 - 3) 64bit longs needed | |
57 | * 32 --> we could get away with a 32bit unsigned int to index and count | |
58 | * 64bit long words, but I rather stay with unsigned long for now. | |
59 | * We probably should neither count nor point to bytes or long words | |
60 | * directly, but either by bitnumber, or by page index and offset. | |
61 | * 1 << (35 - 12) | |
62 | * 22 --> we need that much 4KiB pages of bitmap. | |
63 | * 1 << (22 + 3) --> on a 64bit arch, | |
64 | * we need 32 MiB to store the array of page pointers. | |
65 | * | |
66 | * Because I'm lazy, and because the resulting patch was too large, too ugly | |
67 | * and still incomplete, on 32bit we still "only" support 16 TiB (minus some), | |
68 | * (1 << 32) bits * 4k storage. | |
69 | * | |
70 | ||
71 | * bitmap storage and IO: | |
72 | * Bitmap is stored little endian on disk, and is kept little endian in | |
73 | * core memory. Currently we still hold the full bitmap in core as long | |
74 | * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage | |
75 | * seems excessive. | |
76 | * | |
24c4830c | 77 | * We plan to reduce the amount of in-core bitmap pages by paging them in |
4b0715f0 LE |
78 | * and out against their on-disk location as necessary, but need to make |
79 | * sure we don't cause too much meta data IO, and must not deadlock in | |
80 | * tight memory situations. This needs some more work. | |
b411b363 PR |
81 | */ |
82 | ||
83 | /* | |
84 | * NOTE | |
85 | * Access to the *bm_pages is protected by bm_lock. | |
86 | * It is safe to read the other members within the lock. | |
87 | * | |
88 | * drbd_bm_set_bits is called from bio_endio callbacks, | |
89 | * We may be called with irq already disabled, | |
90 | * so we need spin_lock_irqsave(). | |
91 | * And we need the kmap_atomic. | |
92 | */ | |
93 | struct drbd_bitmap { | |
94 | struct page **bm_pages; | |
95 | spinlock_t bm_lock; | |
4b0715f0 LE |
96 | |
97 | /* see LIMITATIONS: above */ | |
98 | ||
b411b363 PR |
99 | unsigned long bm_set; /* nr of set bits; THINK maybe atomic_t? */ |
100 | unsigned long bm_bits; | |
101 | size_t bm_words; | |
102 | size_t bm_number_of_pages; | |
103 | sector_t bm_dev_capacity; | |
8a03ae2a | 104 | struct mutex bm_change; /* serializes resize operations */ |
b411b363 | 105 | |
19f843aa | 106 | wait_queue_head_t bm_io_wait; /* used to serialize IO of single pages */ |
b411b363 | 107 | |
20ceb2b2 | 108 | enum bm_flag bm_flags; |
b411b363 PR |
109 | |
110 | /* debugging aid, in case we are still racy somewhere */ | |
111 | char *bm_why; | |
112 | struct task_struct *bm_task; | |
113 | }; | |
114 | ||
b411b363 PR |
115 | #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__) |
116 | static void __bm_print_lock_info(struct drbd_conf *mdev, const char *func) | |
117 | { | |
118 | struct drbd_bitmap *b = mdev->bitmap; | |
119 | if (!__ratelimit(&drbd_ratelimit_state)) | |
120 | return; | |
121 | dev_err(DEV, "FIXME %s in %s, bitmap locked for '%s' by %s\n", | |
392c8801 | 122 | drbd_task_to_thread_name(mdev->tconn, current), |
bed879ae | 123 | func, b->bm_why ?: "?", |
392c8801 | 124 | drbd_task_to_thread_name(mdev->tconn, b->bm_task)); |
b411b363 PR |
125 | } |
126 | ||
20ceb2b2 | 127 | void drbd_bm_lock(struct drbd_conf *mdev, char *why, enum bm_flag flags) |
b411b363 PR |
128 | { |
129 | struct drbd_bitmap *b = mdev->bitmap; | |
130 | int trylock_failed; | |
131 | ||
132 | if (!b) { | |
133 | dev_err(DEV, "FIXME no bitmap in drbd_bm_lock!?\n"); | |
134 | return; | |
135 | } | |
136 | ||
8a03ae2a | 137 | trylock_failed = !mutex_trylock(&b->bm_change); |
b411b363 PR |
138 | |
139 | if (trylock_failed) { | |
140 | dev_warn(DEV, "%s going to '%s' but bitmap already locked for '%s' by %s\n", | |
392c8801 | 141 | drbd_task_to_thread_name(mdev->tconn, current), |
bed879ae | 142 | why, b->bm_why ?: "?", |
392c8801 | 143 | drbd_task_to_thread_name(mdev->tconn, b->bm_task)); |
8a03ae2a | 144 | mutex_lock(&b->bm_change); |
b411b363 | 145 | } |
20ceb2b2 | 146 | if (BM_LOCKED_MASK & b->bm_flags) |
b411b363 | 147 | dev_err(DEV, "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 | ||
154 | void drbd_bm_unlock(struct drbd_conf *mdev) | |
155 | { | |
156 | struct drbd_bitmap *b = mdev->bitmap; | |
157 | if (!b) { | |
158 | dev_err(DEV, "FIXME no bitmap in drbd_bm_unlock!?\n"); | |
159 | return; | |
160 | } | |
161 | ||
20ceb2b2 | 162 | if (!(BM_LOCKED_MASK & mdev->bitmap->bm_flags)) |
b411b363 PR |
163 | dev_err(DEV, "FIXME bitmap not locked in bm_unlock\n"); |
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 LE |
202 | BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK)); |
203 | 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 | */ | |
214 | static void bm_page_lock_io(struct drbd_conf *mdev, int page_nr) | |
215 | { | |
216 | struct drbd_bitmap *b = mdev->bitmap; | |
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 | ||
221 | static void bm_page_unlock_io(struct drbd_conf *mdev, int page_nr) | |
222 | { | |
223 | struct drbd_bitmap *b = mdev->bitmap; | |
224 | void *addr = &page_private(b->bm_pages[page_nr]); | |
4738fa16 | 225 | clear_bit_unlock(BM_PAGE_IO_LOCK, addr); |
19f843aa LE |
226 | wake_up(&mdev->bitmap->bm_io_wait); |
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 | |
245 | * @mdev: DRBD device. | |
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 | */ | |
252 | void drbd_bm_mark_for_writeout(struct drbd_conf *mdev, int page_nr) | |
253 | { | |
254 | struct page *page; | |
255 | if (page_nr >= mdev->bitmap->bm_number_of_pages) { | |
256 | dev_warn(DEV, "BAD: page_nr: %u, number_of_pages: %u\n", | |
257 | page_nr, (int)mdev->bitmap->bm_number_of_pages); | |
258 | return; | |
259 | } | |
260 | page = mdev->bitmap->bm_pages[page_nr]; | |
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 | ||
95a0f10c LE |
307 | static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx, const enum km_type km) |
308 | { | |
309 | struct page *page = b->bm_pages[idx]; | |
310 | return (unsigned long *) kmap_atomic(page, km); | |
311 | } | |
312 | ||
313 | static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) | |
314 | { | |
315 | return __bm_map_pidx(b, idx, KM_IRQ1); | |
316 | } | |
317 | ||
b411b363 PR |
318 | static void __bm_unmap(unsigned long *p_addr, const enum km_type km) |
319 | { | |
320 | kunmap_atomic(p_addr, km); | |
321 | }; | |
322 | ||
323 | static void bm_unmap(unsigned long *p_addr) | |
324 | { | |
325 | return __bm_unmap(p_addr, KM_IRQ1); | |
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 | |
343 | * struct drbd_conf*, but for the debug macros I like to have the mdev around | |
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. | |
391 | * GFP_KERNEL is ok, as this is done when a lower level disk is | |
a209b4ae AG |
392 | * "attached" to the drbd. Context is receiver thread or drbdsetup / |
393 | * netlink process. As we have no disk yet, we are not in the IO path, | |
b411b363 PR |
394 | * not even the IO path of the peer. */ |
395 | bytes = sizeof(struct page *)*want; | |
396 | new_pages = kmalloc(bytes, GFP_KERNEL); | |
397 | if (!new_pages) { | |
398 | new_pages = vmalloc(bytes); | |
399 | if (!new_pages) | |
400 | return NULL; | |
401 | vmalloced = 1; | |
402 | } | |
403 | ||
404 | memset(new_pages, 0, bytes); | |
405 | if (want >= have) { | |
406 | for (i = 0; i < have; i++) | |
407 | new_pages[i] = old_pages[i]; | |
408 | for (; i < want; i++) { | |
409 | page = alloc_page(GFP_HIGHUSER); | |
410 | if (!page) { | |
411 | bm_free_pages(new_pages + have, i - have); | |
412 | bm_vk_free(new_pages, vmalloced); | |
413 | return NULL; | |
414 | } | |
19f843aa LE |
415 | /* we want to know which page it is |
416 | * from the endio handlers */ | |
417 | bm_store_page_idx(page, i); | |
b411b363 PR |
418 | new_pages[i] = page; |
419 | } | |
420 | } else { | |
421 | for (i = 0; i < want; i++) | |
422 | new_pages[i] = old_pages[i]; | |
423 | /* NOT HERE, we are outside the spinlock! | |
424 | bm_free_pages(old_pages + want, have - want); | |
425 | */ | |
426 | } | |
427 | ||
428 | if (vmalloced) | |
20ceb2b2 | 429 | b->bm_flags |= BM_P_VMALLOCED; |
b411b363 | 430 | else |
20ceb2b2 | 431 | b->bm_flags &= ~BM_P_VMALLOCED; |
b411b363 PR |
432 | |
433 | return new_pages; | |
434 | } | |
435 | ||
436 | /* | |
437 | * called on driver init only. TODO call when a device is created. | |
438 | * allocates the drbd_bitmap, and stores it in mdev->bitmap. | |
439 | */ | |
440 | int drbd_bm_init(struct drbd_conf *mdev) | |
441 | { | |
442 | struct drbd_bitmap *b = mdev->bitmap; | |
443 | WARN_ON(b != NULL); | |
444 | b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL); | |
445 | if (!b) | |
446 | return -ENOMEM; | |
447 | spin_lock_init(&b->bm_lock); | |
8a03ae2a | 448 | mutex_init(&b->bm_change); |
b411b363 PR |
449 | init_waitqueue_head(&b->bm_io_wait); |
450 | ||
451 | mdev->bitmap = b; | |
452 | ||
453 | return 0; | |
454 | } | |
455 | ||
456 | sector_t drbd_bm_capacity(struct drbd_conf *mdev) | |
457 | { | |
841ce241 AG |
458 | if (!expect(mdev->bitmap)) |
459 | return 0; | |
b411b363 PR |
460 | return mdev->bitmap->bm_dev_capacity; |
461 | } | |
462 | ||
463 | /* called on driver unload. TODO: call when a device is destroyed. | |
464 | */ | |
465 | void drbd_bm_cleanup(struct drbd_conf *mdev) | |
466 | { | |
841ce241 AG |
467 | if (!expect(mdev->bitmap)) |
468 | return; | |
b411b363 | 469 | bm_free_pages(mdev->bitmap->bm_pages, mdev->bitmap->bm_number_of_pages); |
20ceb2b2 | 470 | bm_vk_free(mdev->bitmap->bm_pages, (BM_P_VMALLOCED & mdev->bitmap->bm_flags)); |
b411b363 PR |
471 | kfree(mdev->bitmap); |
472 | mdev->bitmap = NULL; | |
473 | } | |
474 | ||
475 | /* | |
476 | * since (b->bm_bits % BITS_PER_LONG) != 0, | |
477 | * this masks out the remaining bits. | |
478 | * Returns the number of bits cleared. | |
479 | */ | |
95a0f10c LE |
480 | #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3)) |
481 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1) | |
482 | #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1) | |
b411b363 PR |
483 | static int bm_clear_surplus(struct drbd_bitmap *b) |
484 | { | |
95a0f10c | 485 | unsigned long mask; |
b411b363 | 486 | unsigned long *p_addr, *bm; |
95a0f10c LE |
487 | int tmp; |
488 | int cleared = 0; | |
b411b363 | 489 | |
95a0f10c LE |
490 | /* number of bits modulo bits per page */ |
491 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
492 | /* mask the used bits of the word containing the last bit */ | |
493 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
494 | /* bitmap is always stored little endian, | |
495 | * on disk and in core memory alike */ | |
496 | mask = cpu_to_lel(mask); | |
497 | ||
6850c442 | 498 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
499 | bm = p_addr + (tmp/BITS_PER_LONG); |
500 | if (mask) { | |
501 | /* If mask != 0, we are not exactly aligned, so bm now points | |
502 | * to the long containing the last bit. | |
503 | * If mask == 0, bm already points to the word immediately | |
504 | * after the last (long word aligned) bit. */ | |
b411b363 PR |
505 | cleared = hweight_long(*bm & ~mask); |
506 | *bm &= mask; | |
95a0f10c | 507 | bm++; |
b411b363 PR |
508 | } |
509 | ||
95a0f10c LE |
510 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
511 | /* on a 32bit arch, we may need to zero out | |
512 | * a padding long to align with a 64bit remote */ | |
b411b363 PR |
513 | cleared += hweight_long(*bm); |
514 | *bm = 0; | |
515 | } | |
516 | bm_unmap(p_addr); | |
517 | return cleared; | |
518 | } | |
519 | ||
520 | static void bm_set_surplus(struct drbd_bitmap *b) | |
521 | { | |
95a0f10c | 522 | unsigned long mask; |
b411b363 | 523 | unsigned long *p_addr, *bm; |
95a0f10c LE |
524 | int tmp; |
525 | ||
526 | /* number of bits modulo bits per page */ | |
527 | tmp = (b->bm_bits & BITS_PER_PAGE_MASK); | |
528 | /* mask the used bits of the word containing the last bit */ | |
529 | mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; | |
530 | /* bitmap is always stored little endian, | |
531 | * on disk and in core memory alike */ | |
532 | mask = cpu_to_lel(mask); | |
533 | ||
6850c442 | 534 | p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); |
95a0f10c LE |
535 | bm = p_addr + (tmp/BITS_PER_LONG); |
536 | if (mask) { | |
537 | /* If mask != 0, we are not exactly aligned, so bm now points | |
538 | * to the long containing the last bit. | |
539 | * If mask == 0, bm already points to the word immediately | |
540 | * after the last (long word aligned) bit. */ | |
b411b363 | 541 | *bm |= ~mask; |
95a0f10c | 542 | bm++; |
b411b363 PR |
543 | } |
544 | ||
95a0f10c LE |
545 | if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { |
546 | /* on a 32bit arch, we may need to zero out | |
547 | * a padding long to align with a 64bit remote */ | |
548 | *bm = ~0UL; | |
b411b363 PR |
549 | } |
550 | bm_unmap(p_addr); | |
551 | } | |
552 | ||
4b0715f0 LE |
553 | /* you better not modify the bitmap while this is running, |
554 | * or its results will be stale */ | |
95a0f10c | 555 | static unsigned long bm_count_bits(struct drbd_bitmap *b) |
b411b363 | 556 | { |
4b0715f0 | 557 | unsigned long *p_addr; |
b411b363 | 558 | unsigned long bits = 0; |
4b0715f0 | 559 | unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1; |
6850c442 | 560 | int idx, i, last_word; |
4b0715f0 LE |
561 | |
562 | /* all but last page */ | |
6850c442 | 563 | for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) { |
4b0715f0 LE |
564 | p_addr = __bm_map_pidx(b, idx, KM_USER0); |
565 | for (i = 0; i < LWPP; i++) | |
566 | bits += hweight_long(p_addr[i]); | |
7777a8ba | 567 | __bm_unmap(p_addr, KM_USER0); |
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; | |
572 | p_addr = __bm_map_pidx(b, idx, KM_USER0); | |
573 | for (i = 0; i < last_word; i++) | |
574 | bits += hweight_long(p_addr[i]); | |
575 | p_addr[last_word] &= cpu_to_lel(mask); | |
576 | bits += hweight_long(p_addr[last_word]); | |
577 | /* 32bit arch, may have an unused padding long */ | |
578 | if (BITS_PER_LONG == 32 && (last_word & 1) == 0) | |
579 | p_addr[last_word+1] = 0; | |
580 | __bm_unmap(p_addr, KM_USER0); | |
b411b363 PR |
581 | return bits; |
582 | } | |
583 | ||
b411b363 PR |
584 | /* offset and len in long words.*/ |
585 | static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len) | |
586 | { | |
587 | unsigned long *p_addr, *bm; | |
19f843aa | 588 | unsigned int idx; |
b411b363 PR |
589 | size_t do_now, end; |
590 | ||
b411b363 PR |
591 | end = offset + len; |
592 | ||
593 | if (end > b->bm_words) { | |
594 | printk(KERN_ALERT "drbd: bm_memset end > bm_words\n"); | |
595 | return; | |
596 | } | |
597 | ||
598 | while (offset < end) { | |
599 | do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset; | |
19f843aa LE |
600 | idx = bm_word_to_page_idx(b, offset); |
601 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
602 | bm = p_addr + MLPP(offset); |
603 | if (bm+do_now > p_addr + LWPP) { | |
604 | printk(KERN_ALERT "drbd: BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n", | |
605 | p_addr, bm, (int)do_now); | |
84e7c0f7 LE |
606 | } else |
607 | memset(bm, c, do_now * sizeof(long)); | |
b411b363 | 608 | bm_unmap(p_addr); |
19f843aa | 609 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
610 | offset += do_now; |
611 | } | |
612 | } | |
613 | ||
614 | /* | |
615 | * make sure the bitmap has enough room for the attached storage, | |
616 | * if necessary, resize. | |
617 | * called whenever we may have changed the device size. | |
618 | * returns -ENOMEM if we could not allocate enough memory, 0 on success. | |
619 | * In case this is actually a resize, we copy the old bitmap into the new one. | |
620 | * Otherwise, the bitmap is initialized to all bits set. | |
621 | */ | |
02d9a94b | 622 | int drbd_bm_resize(struct drbd_conf *mdev, sector_t capacity, int set_new_bits) |
b411b363 PR |
623 | { |
624 | struct drbd_bitmap *b = mdev->bitmap; | |
6850c442 | 625 | unsigned long bits, words, owords, obits; |
b411b363 PR |
626 | unsigned long want, have, onpages; /* number of pages */ |
627 | struct page **npages, **opages = NULL; | |
628 | int err = 0, growing; | |
629 | int opages_vmalloced; | |
630 | ||
841ce241 AG |
631 | if (!expect(b)) |
632 | return -ENOMEM; | |
b411b363 | 633 | |
20ceb2b2 | 634 | drbd_bm_lock(mdev, "resize", BM_LOCKED_MASK); |
b411b363 PR |
635 | |
636 | dev_info(DEV, "drbd_bm_resize called with capacity == %llu\n", | |
637 | (unsigned long long)capacity); | |
638 | ||
639 | if (capacity == b->bm_dev_capacity) | |
640 | goto out; | |
641 | ||
20ceb2b2 | 642 | opages_vmalloced = (BM_P_VMALLOCED & b->bm_flags); |
b411b363 PR |
643 | |
644 | if (capacity == 0) { | |
645 | spin_lock_irq(&b->bm_lock); | |
646 | opages = b->bm_pages; | |
647 | onpages = b->bm_number_of_pages; | |
648 | owords = b->bm_words; | |
649 | b->bm_pages = NULL; | |
650 | b->bm_number_of_pages = | |
651 | b->bm_set = | |
652 | b->bm_bits = | |
653 | b->bm_words = | |
654 | b->bm_dev_capacity = 0; | |
655 | spin_unlock_irq(&b->bm_lock); | |
656 | bm_free_pages(opages, onpages); | |
657 | bm_vk_free(opages, opages_vmalloced); | |
658 | goto out; | |
659 | } | |
660 | bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT)); | |
661 | ||
662 | /* if we would use | |
663 | words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL; | |
664 | a 32bit host could present the wrong number of words | |
665 | to a 64bit host. | |
666 | */ | |
667 | words = ALIGN(bits, 64) >> LN2_BPL; | |
668 | ||
669 | if (get_ldev(mdev)) { | |
4b0715f0 | 670 | u64 bits_on_disk = ((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12; |
b411b363 | 671 | put_ldev(mdev); |
4b0715f0 LE |
672 | if (bits > bits_on_disk) { |
673 | dev_info(DEV, "bits = %lu\n", bits); | |
674 | dev_info(DEV, "bits_on_disk = %llu\n", bits_on_disk); | |
675 | err = -ENOSPC; | |
676 | goto out; | |
677 | } | |
b411b363 PR |
678 | } |
679 | ||
6850c442 | 680 | want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT; |
b411b363 PR |
681 | have = b->bm_number_of_pages; |
682 | if (want == have) { | |
683 | D_ASSERT(b->bm_pages != NULL); | |
684 | npages = b->bm_pages; | |
685 | } else { | |
0cf9d27e | 686 | if (drbd_insert_fault(mdev, DRBD_FAULT_BM_ALLOC)) |
b411b363 PR |
687 | npages = NULL; |
688 | else | |
689 | npages = bm_realloc_pages(b, want); | |
690 | } | |
691 | ||
692 | if (!npages) { | |
693 | err = -ENOMEM; | |
694 | goto out; | |
695 | } | |
696 | ||
697 | spin_lock_irq(&b->bm_lock); | |
698 | opages = b->bm_pages; | |
699 | owords = b->bm_words; | |
700 | obits = b->bm_bits; | |
701 | ||
702 | growing = bits > obits; | |
5223671b | 703 | if (opages && growing && set_new_bits) |
b411b363 PR |
704 | bm_set_surplus(b); |
705 | ||
706 | b->bm_pages = npages; | |
707 | b->bm_number_of_pages = want; | |
708 | b->bm_bits = bits; | |
709 | b->bm_words = words; | |
710 | b->bm_dev_capacity = capacity; | |
711 | ||
712 | if (growing) { | |
02d9a94b PR |
713 | if (set_new_bits) { |
714 | bm_memset(b, owords, 0xff, words-owords); | |
715 | b->bm_set += bits - obits; | |
716 | } else | |
717 | bm_memset(b, owords, 0x00, words-owords); | |
718 | ||
b411b363 PR |
719 | } |
720 | ||
721 | if (want < have) { | |
722 | /* implicit: (opages != NULL) && (opages != npages) */ | |
723 | bm_free_pages(opages + want, have - want); | |
724 | } | |
725 | ||
b411b363 PR |
726 | (void)bm_clear_surplus(b); |
727 | ||
728 | spin_unlock_irq(&b->bm_lock); | |
729 | if (opages != npages) | |
730 | bm_vk_free(opages, opages_vmalloced); | |
731 | if (!growing) | |
732 | b->bm_set = bm_count_bits(b); | |
19f843aa | 733 | dev_info(DEV, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want); |
b411b363 PR |
734 | |
735 | out: | |
736 | drbd_bm_unlock(mdev); | |
737 | return err; | |
738 | } | |
739 | ||
740 | /* inherently racy: | |
741 | * if not protected by other means, return value may be out of date when | |
742 | * leaving this function... | |
743 | * we still need to lock it, since it is important that this returns | |
744 | * bm_set == 0 precisely. | |
745 | * | |
746 | * maybe bm_set should be atomic_t ? | |
747 | */ | |
0778286a | 748 | unsigned long _drbd_bm_total_weight(struct drbd_conf *mdev) |
b411b363 PR |
749 | { |
750 | struct drbd_bitmap *b = mdev->bitmap; | |
751 | unsigned long s; | |
752 | unsigned long flags; | |
753 | ||
841ce241 AG |
754 | if (!expect(b)) |
755 | return 0; | |
756 | if (!expect(b->bm_pages)) | |
757 | return 0; | |
b411b363 PR |
758 | |
759 | spin_lock_irqsave(&b->bm_lock, flags); | |
760 | s = b->bm_set; | |
761 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
762 | ||
763 | return s; | |
764 | } | |
765 | ||
766 | unsigned long drbd_bm_total_weight(struct drbd_conf *mdev) | |
767 | { | |
768 | unsigned long s; | |
769 | /* if I don't have a disk, I don't know about out-of-sync status */ | |
770 | if (!get_ldev_if_state(mdev, D_NEGOTIATING)) | |
771 | return 0; | |
772 | s = _drbd_bm_total_weight(mdev); | |
773 | put_ldev(mdev); | |
774 | return s; | |
775 | } | |
776 | ||
777 | size_t drbd_bm_words(struct drbd_conf *mdev) | |
778 | { | |
779 | struct drbd_bitmap *b = mdev->bitmap; | |
841ce241 AG |
780 | if (!expect(b)) |
781 | return 0; | |
782 | if (!expect(b->bm_pages)) | |
783 | return 0; | |
b411b363 PR |
784 | |
785 | return b->bm_words; | |
786 | } | |
787 | ||
788 | unsigned long drbd_bm_bits(struct drbd_conf *mdev) | |
789 | { | |
790 | struct drbd_bitmap *b = mdev->bitmap; | |
841ce241 AG |
791 | if (!expect(b)) |
792 | return 0; | |
b411b363 PR |
793 | |
794 | return b->bm_bits; | |
795 | } | |
796 | ||
797 | /* merge number words from buffer into the bitmap starting at offset. | |
798 | * buffer[i] is expected to be little endian unsigned long. | |
799 | * bitmap must be locked by drbd_bm_lock. | |
800 | * currently only used from receive_bitmap. | |
801 | */ | |
802 | void drbd_bm_merge_lel(struct drbd_conf *mdev, size_t offset, size_t number, | |
803 | unsigned long *buffer) | |
804 | { | |
805 | struct drbd_bitmap *b = mdev->bitmap; | |
806 | unsigned long *p_addr, *bm; | |
807 | unsigned long word, bits; | |
19f843aa | 808 | unsigned int idx; |
b411b363 PR |
809 | size_t end, do_now; |
810 | ||
811 | end = offset + number; | |
812 | ||
841ce241 AG |
813 | if (!expect(b)) |
814 | return; | |
815 | if (!expect(b->bm_pages)) | |
816 | return; | |
b411b363 PR |
817 | if (number == 0) |
818 | return; | |
819 | WARN_ON(offset >= b->bm_words); | |
820 | WARN_ON(end > b->bm_words); | |
821 | ||
822 | spin_lock_irq(&b->bm_lock); | |
823 | while (offset < end) { | |
824 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa LE |
825 | idx = bm_word_to_page_idx(b, offset); |
826 | p_addr = bm_map_pidx(b, idx); | |
b411b363 PR |
827 | bm = p_addr + MLPP(offset); |
828 | offset += do_now; | |
829 | while (do_now--) { | |
830 | bits = hweight_long(*bm); | |
95a0f10c | 831 | word = *bm | *buffer++; |
b411b363 PR |
832 | *bm++ = word; |
833 | b->bm_set += hweight_long(word) - bits; | |
834 | } | |
835 | bm_unmap(p_addr); | |
19f843aa | 836 | bm_set_page_need_writeout(b->bm_pages[idx]); |
b411b363 PR |
837 | } |
838 | /* with 32bit <-> 64bit cross-platform connect | |
839 | * this is only correct for current usage, | |
840 | * where we _know_ that we are 64 bit aligned, | |
841 | * and know that this function is used in this way, too... | |
842 | */ | |
843 | if (end == b->bm_words) | |
844 | b->bm_set -= bm_clear_surplus(b); | |
b411b363 PR |
845 | spin_unlock_irq(&b->bm_lock); |
846 | } | |
847 | ||
848 | /* copy number words from the bitmap starting at offset into the buffer. | |
849 | * buffer[i] will be little endian unsigned long. | |
850 | */ | |
851 | void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset, size_t number, | |
852 | unsigned long *buffer) | |
853 | { | |
854 | struct drbd_bitmap *b = mdev->bitmap; | |
855 | unsigned long *p_addr, *bm; | |
856 | size_t end, do_now; | |
857 | ||
858 | end = offset + number; | |
859 | ||
841ce241 AG |
860 | if (!expect(b)) |
861 | return; | |
862 | if (!expect(b->bm_pages)) | |
863 | return; | |
b411b363 PR |
864 | |
865 | spin_lock_irq(&b->bm_lock); | |
866 | if ((offset >= b->bm_words) || | |
867 | (end > b->bm_words) || | |
868 | (number <= 0)) | |
869 | dev_err(DEV, "offset=%lu number=%lu bm_words=%lu\n", | |
870 | (unsigned long) offset, | |
871 | (unsigned long) number, | |
872 | (unsigned long) b->bm_words); | |
873 | else { | |
874 | while (offset < end) { | |
875 | do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; | |
19f843aa | 876 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset)); |
b411b363 PR |
877 | bm = p_addr + MLPP(offset); |
878 | offset += do_now; | |
879 | while (do_now--) | |
95a0f10c | 880 | *buffer++ = *bm++; |
b411b363 PR |
881 | bm_unmap(p_addr); |
882 | } | |
883 | } | |
884 | spin_unlock_irq(&b->bm_lock); | |
885 | } | |
886 | ||
887 | /* set all bits in the bitmap */ | |
888 | void drbd_bm_set_all(struct drbd_conf *mdev) | |
889 | { | |
890 | struct drbd_bitmap *b = mdev->bitmap; | |
841ce241 AG |
891 | if (!expect(b)) |
892 | return; | |
893 | if (!expect(b->bm_pages)) | |
894 | return; | |
b411b363 PR |
895 | |
896 | spin_lock_irq(&b->bm_lock); | |
897 | bm_memset(b, 0, 0xff, b->bm_words); | |
898 | (void)bm_clear_surplus(b); | |
899 | b->bm_set = b->bm_bits; | |
900 | spin_unlock_irq(&b->bm_lock); | |
901 | } | |
902 | ||
903 | /* clear all bits in the bitmap */ | |
904 | void drbd_bm_clear_all(struct drbd_conf *mdev) | |
905 | { | |
906 | struct drbd_bitmap *b = mdev->bitmap; | |
841ce241 AG |
907 | if (!expect(b)) |
908 | return; | |
909 | if (!expect(b->bm_pages)) | |
910 | return; | |
b411b363 PR |
911 | |
912 | spin_lock_irq(&b->bm_lock); | |
913 | bm_memset(b, 0, 0, b->bm_words); | |
914 | b->bm_set = 0; | |
915 | spin_unlock_irq(&b->bm_lock); | |
916 | } | |
917 | ||
19f843aa LE |
918 | struct bm_aio_ctx { |
919 | struct drbd_conf *mdev; | |
920 | atomic_t in_flight; | |
cdfda633 | 921 | unsigned int done; |
19f843aa LE |
922 | unsigned flags; |
923 | #define BM_AIO_COPY_PAGES 1 | |
45dfffeb | 924 | #define BM_AIO_WRITE_HINTED 2 |
19f843aa | 925 | int error; |
cdfda633 | 926 | struct kref kref; |
19f843aa LE |
927 | }; |
928 | ||
cdfda633 PR |
929 | static void bm_aio_ctx_destroy(struct kref *kref) |
930 | { | |
931 | struct bm_aio_ctx *ctx = container_of(kref, struct bm_aio_ctx, kref); | |
932 | ||
933 | put_ldev(ctx->mdev); | |
934 | kfree(ctx); | |
935 | } | |
936 | ||
19f843aa | 937 | /* bv_page may be a copy, or may be the original */ |
b411b363 PR |
938 | static void bm_async_io_complete(struct bio *bio, int error) |
939 | { | |
19f843aa LE |
940 | struct bm_aio_ctx *ctx = bio->bi_private; |
941 | struct drbd_conf *mdev = ctx->mdev; | |
942 | struct drbd_bitmap *b = mdev->bitmap; | |
943 | unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page); | |
b411b363 PR |
944 | int uptodate = bio_flagged(bio, BIO_UPTODATE); |
945 | ||
946 | ||
947 | /* strange behavior of some lower level drivers... | |
948 | * fail the request by clearing the uptodate flag, | |
949 | * but do not return any error?! | |
950 | * do we want to WARN() on this? */ | |
951 | if (!error && !uptodate) | |
952 | error = -EIO; | |
953 | ||
7648cdfe LE |
954 | if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 && |
955 | !bm_test_page_unchanged(b->bm_pages[idx])) | |
956 | dev_warn(DEV, "bitmap page idx %u changed during IO!\n", idx); | |
19f843aa | 957 | |
b411b363 | 958 | if (error) { |
19f843aa LE |
959 | /* ctx error will hold the completed-last non-zero error code, |
960 | * in case error codes differ. */ | |
961 | ctx->error = error; | |
962 | bm_set_page_io_err(b->bm_pages[idx]); | |
963 | /* Not identical to on disk version of it. | |
964 | * Is BM_PAGE_IO_ERROR enough? */ | |
965 | if (__ratelimit(&drbd_ratelimit_state)) | |
966 | dev_err(DEV, "IO ERROR %d on bitmap page idx %u\n", | |
967 | error, idx); | |
968 | } else { | |
969 | bm_clear_page_io_err(b->bm_pages[idx]); | |
970 | dynamic_dev_dbg(DEV, "bitmap page idx %u completed\n", idx); | |
b411b363 | 971 | } |
19f843aa LE |
972 | |
973 | bm_page_unlock_io(mdev, idx); | |
974 | ||
19f843aa | 975 | if (ctx->flags & BM_AIO_COPY_PAGES) |
9db4e77f | 976 | mempool_free(bio->bi_io_vec[0].bv_page, drbd_md_io_page_pool); |
b411b363 PR |
977 | |
978 | bio_put(bio); | |
19f843aa | 979 | |
cdfda633 PR |
980 | if (atomic_dec_and_test(&ctx->in_flight)) { |
981 | ctx->done = 1; | |
982 | wake_up(&mdev->misc_wait); | |
983 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); | |
984 | } | |
b411b363 PR |
985 | } |
986 | ||
19f843aa | 987 | static void bm_page_io_async(struct bm_aio_ctx *ctx, int page_nr, int rw) __must_hold(local) |
b411b363 | 988 | { |
cdfda633 | 989 | struct bio *bio = bio_alloc_drbd(GFP_NOIO); |
19f843aa LE |
990 | struct drbd_conf *mdev = ctx->mdev; |
991 | struct drbd_bitmap *b = mdev->bitmap; | |
992 | struct page *page; | |
b411b363 | 993 | unsigned int len; |
19f843aa | 994 | |
b411b363 PR |
995 | sector_t on_disk_sector = |
996 | mdev->ldev->md.md_offset + mdev->ldev->md.bm_offset; | |
997 | on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9); | |
998 | ||
999 | /* this might happen with very small | |
19f843aa LE |
1000 | * flexible external meta data device, |
1001 | * or with PAGE_SIZE > 4k */ | |
b411b363 PR |
1002 | len = min_t(unsigned int, PAGE_SIZE, |
1003 | (drbd_md_last_sector(mdev->ldev) - on_disk_sector + 1)<<9); | |
1004 | ||
19f843aa LE |
1005 | /* serialize IO on this page */ |
1006 | bm_page_lock_io(mdev, page_nr); | |
1007 | /* before memcpy and submit, | |
1008 | * so it can be redirtied any time */ | |
1009 | bm_set_page_unchanged(b->bm_pages[page_nr]); | |
1010 | ||
1011 | if (ctx->flags & BM_AIO_COPY_PAGES) { | |
19f843aa | 1012 | void *src, *dest; |
9db4e77f | 1013 | page = mempool_alloc(drbd_md_io_page_pool, __GFP_HIGHMEM|__GFP_WAIT); |
19f843aa LE |
1014 | dest = kmap_atomic(page, KM_USER0); |
1015 | src = kmap_atomic(b->bm_pages[page_nr], KM_USER1); | |
1016 | memcpy(dest, src, PAGE_SIZE); | |
1017 | kunmap_atomic(src, KM_USER1); | |
1018 | kunmap_atomic(dest, KM_USER0); | |
1019 | bm_store_page_idx(page, page_nr); | |
1020 | } else | |
1021 | page = b->bm_pages[page_nr]; | |
1022 | ||
b411b363 PR |
1023 | bio->bi_bdev = mdev->ldev->md_bdev; |
1024 | bio->bi_sector = on_disk_sector; | |
9db4e77f LE |
1025 | /* bio_add_page of a single page to an empty bio will always succeed, |
1026 | * according to api. Do we want to assert that? */ | |
19f843aa LE |
1027 | bio_add_page(bio, page, len, 0); |
1028 | bio->bi_private = ctx; | |
b411b363 PR |
1029 | bio->bi_end_io = bm_async_io_complete; |
1030 | ||
0cf9d27e | 1031 | if (drbd_insert_fault(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) { |
b411b363 PR |
1032 | bio->bi_rw |= rw; |
1033 | bio_endio(bio, -EIO); | |
1034 | } else { | |
1035 | submit_bio(rw, bio); | |
5a8b4242 LE |
1036 | /* this should not count as user activity and cause the |
1037 | * resync to throttle -- see drbd_rs_should_slow_down(). */ | |
1038 | atomic_add(len >> 9, &mdev->rs_sect_ev); | |
b411b363 PR |
1039 | } |
1040 | } | |
1041 | ||
b411b363 PR |
1042 | /* |
1043 | * bm_rw: read/write the whole bitmap from/to its on disk location. | |
1044 | */ | |
45dfffeb | 1045 | static int bm_rw(struct drbd_conf *mdev, int rw, unsigned flags, unsigned lazy_writeout_upper_idx) __must_hold(local) |
b411b363 | 1046 | { |
cdfda633 | 1047 | struct bm_aio_ctx *ctx; |
b411b363 | 1048 | struct drbd_bitmap *b = mdev->bitmap; |
6850c442 | 1049 | int num_pages, i, count = 0; |
b411b363 PR |
1050 | unsigned long now; |
1051 | char ppb[10]; | |
1052 | int err = 0; | |
1053 | ||
19f843aa LE |
1054 | /* |
1055 | * We are protected against bitmap disappearing/resizing by holding an | |
1056 | * ldev reference (caller must have called get_ldev()). | |
1057 | * For read/write, we are protected against changes to the bitmap by | |
1058 | * the bitmap lock (see drbd_bitmap_io). | |
1059 | * For lazy writeout, we don't care for ongoing changes to the bitmap, | |
1060 | * as we submit copies of pages anyways. | |
1061 | */ | |
cdfda633 PR |
1062 | |
1063 | ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO); | |
1064 | if (!ctx) | |
1065 | return -ENOMEM; | |
1066 | ||
1067 | *ctx = (struct bm_aio_ctx) { | |
1068 | .mdev = mdev, | |
1069 | .in_flight = ATOMIC_INIT(1), | |
1070 | .done = 0, | |
1071 | .flags = flags, | |
1072 | .error = 0, | |
1073 | .kref = { ATOMIC_INIT(2) }, | |
1074 | }; | |
1075 | ||
1076 | if (!get_ldev_if_state(mdev, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */ | |
1077 | dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n"); | |
1078 | err = -ENODEV; | |
1079 | goto out; | |
1080 | } | |
1081 | ||
1082 | if (!ctx->flags) | |
20ceb2b2 | 1083 | WARN_ON(!(BM_LOCKED_MASK & b->bm_flags)); |
b411b363 | 1084 | |
6850c442 | 1085 | num_pages = b->bm_number_of_pages; |
b411b363 | 1086 | |
b411b363 | 1087 | now = jiffies; |
b411b363 PR |
1088 | |
1089 | /* let the layers below us try to merge these bios... */ | |
6850c442 | 1090 | for (i = 0; i < num_pages; i++) { |
19f843aa LE |
1091 | /* ignore completely unchanged pages */ |
1092 | if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx) | |
1093 | break; | |
1094 | if (rw & WRITE) { | |
45dfffeb LE |
1095 | if ((flags & BM_AIO_WRITE_HINTED) && |
1096 | !test_and_clear_bit(BM_PAGE_HINT_WRITEOUT, | |
1097 | &page_private(b->bm_pages[i]))) | |
1098 | continue; | |
19f843aa LE |
1099 | if (bm_test_page_unchanged(b->bm_pages[i])) { |
1100 | dynamic_dev_dbg(DEV, "skipped bm write for idx %u\n", i); | |
1101 | continue; | |
1102 | } | |
1103 | /* during lazy writeout, | |
1104 | * ignore those pages not marked for lazy writeout. */ | |
1105 | if (lazy_writeout_upper_idx && | |
1106 | !bm_test_page_lazy_writeout(b->bm_pages[i])) { | |
1107 | dynamic_dev_dbg(DEV, "skipped bm lazy write for idx %u\n", i); | |
1108 | continue; | |
1109 | } | |
1110 | } | |
cdfda633 PR |
1111 | atomic_inc(&ctx->in_flight); |
1112 | bm_page_io_async(ctx, i, rw); | |
19f843aa LE |
1113 | ++count; |
1114 | cond_resched(); | |
1115 | } | |
b411b363 | 1116 | |
725a97e4 | 1117 | /* |
cdfda633 PR |
1118 | * We initialize ctx->in_flight to one to make sure bm_async_io_complete |
1119 | * will not set ctx->done early, and decrement / test it here. If there | |
725a97e4 | 1120 | * are still some bios in flight, we need to wait for them here. |
cdfda633 PR |
1121 | * If all IO is done already (or nothing had been submitted), there is |
1122 | * no need to wait. Still, we need to put the kref associated with the | |
1123 | * "in_flight reached zero, all done" event. | |
725a97e4 | 1124 | */ |
cdfda633 PR |
1125 | if (!atomic_dec_and_test(&ctx->in_flight)) |
1126 | wait_until_done_or_disk_failure(mdev, &ctx->done); | |
1127 | else | |
1128 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); | |
c9d963a4 LE |
1129 | |
1130 | /* summary for global bitmap IO */ | |
1131 | if (flags == 0) | |
1132 | dev_info(DEV, "bitmap %s of %u pages took %lu jiffies\n", | |
cdfda633 PR |
1133 | rw == WRITE ? "WRITE" : "READ", |
1134 | count, jiffies - now); | |
b411b363 | 1135 | |
cdfda633 | 1136 | if (ctx->error) { |
b411b363 | 1137 | dev_alert(DEV, "we had at least one MD IO ERROR during bitmap IO\n"); |
81e84650 | 1138 | drbd_chk_io_error(mdev, 1, true); |
cdfda633 | 1139 | err = -EIO; /* ctx->error ? */ |
b411b363 PR |
1140 | } |
1141 | ||
cdfda633 PR |
1142 | if (atomic_read(&ctx->in_flight)) |
1143 | err = -EIO; /* Disk failed during IO... */ | |
1144 | ||
b411b363 PR |
1145 | now = jiffies; |
1146 | if (rw == WRITE) { | |
b411b363 PR |
1147 | drbd_md_flush(mdev); |
1148 | } else /* rw == READ */ { | |
95a0f10c | 1149 | b->bm_set = bm_count_bits(b); |
b411b363 PR |
1150 | dev_info(DEV, "recounting of set bits took additional %lu jiffies\n", |
1151 | jiffies - now); | |
1152 | } | |
1153 | now = b->bm_set; | |
1154 | ||
c9d963a4 LE |
1155 | if (flags == 0) |
1156 | dev_info(DEV, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n", | |
1157 | ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now); | |
b411b363 | 1158 | |
cdfda633 PR |
1159 | out: |
1160 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); | |
b411b363 PR |
1161 | return err; |
1162 | } | |
1163 | ||
1164 | /** | |
1165 | * drbd_bm_read() - Read the whole bitmap from its on disk location. | |
1166 | * @mdev: DRBD device. | |
1167 | */ | |
1168 | int drbd_bm_read(struct drbd_conf *mdev) __must_hold(local) | |
1169 | { | |
45dfffeb | 1170 | return bm_rw(mdev, READ, 0, 0); |
b411b363 PR |
1171 | } |
1172 | ||
1173 | /** | |
1174 | * drbd_bm_write() - Write the whole bitmap to its on disk location. | |
1175 | * @mdev: DRBD device. | |
19f843aa LE |
1176 | * |
1177 | * Will only write pages that have changed since last IO. | |
b411b363 PR |
1178 | */ |
1179 | int drbd_bm_write(struct drbd_conf *mdev) __must_hold(local) | |
1180 | { | |
45dfffeb | 1181 | return bm_rw(mdev, WRITE, 0, 0); |
b411b363 PR |
1182 | } |
1183 | ||
1184 | /** | |
19f843aa | 1185 | * drbd_bm_lazy_write_out() - Write bitmap pages 0 to @upper_idx-1, if they have changed. |
b411b363 | 1186 | * @mdev: DRBD device. |
19f843aa LE |
1187 | * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages |
1188 | */ | |
1189 | int drbd_bm_write_lazy(struct drbd_conf *mdev, unsigned upper_idx) __must_hold(local) | |
1190 | { | |
45dfffeb | 1191 | return bm_rw(mdev, WRITE, BM_AIO_COPY_PAGES, upper_idx); |
19f843aa LE |
1192 | } |
1193 | ||
45dfffeb LE |
1194 | /** |
1195 | * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed. | |
1196 | * @mdev: DRBD device. | |
1197 | */ | |
1198 | int drbd_bm_write_hinted(struct drbd_conf *mdev) __must_hold(local) | |
1199 | { | |
1200 | return bm_rw(mdev, WRITE, BM_AIO_WRITE_HINTED | BM_AIO_COPY_PAGES, 0); | |
1201 | } | |
19f843aa LE |
1202 | |
1203 | /** | |
45dfffeb | 1204 | * drbd_bm_write_page() - Writes a PAGE_SIZE aligned piece of bitmap |
19f843aa LE |
1205 | * @mdev: DRBD device. |
1206 | * @idx: bitmap page index | |
b411b363 | 1207 | * |
4b0715f0 LE |
1208 | * We don't want to special case on logical_block_size of the backend device, |
1209 | * so we submit PAGE_SIZE aligned pieces. | |
19f843aa | 1210 | * Note that on "most" systems, PAGE_SIZE is 4k. |
4b0715f0 LE |
1211 | * |
1212 | * In case this becomes an issue on systems with larger PAGE_SIZE, | |
1213 | * we may want to change this again to write 4k aligned 4k pieces. | |
b411b363 | 1214 | */ |
19f843aa | 1215 | int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local) |
b411b363 | 1216 | { |
cdfda633 PR |
1217 | struct bm_aio_ctx *ctx; |
1218 | int err; | |
1219 | ||
1220 | if (bm_test_page_unchanged(mdev->bitmap->bm_pages[idx])) { | |
1221 | dynamic_dev_dbg(DEV, "skipped bm page write for idx %u\n", idx); | |
1222 | return 0; | |
1223 | } | |
1224 | ||
1225 | ctx = kmalloc(sizeof(struct bm_aio_ctx), GFP_NOIO); | |
1226 | if (!ctx) | |
1227 | return -ENOMEM; | |
1228 | ||
1229 | *ctx = (struct bm_aio_ctx) { | |
725a97e4 LE |
1230 | .mdev = mdev, |
1231 | .in_flight = ATOMIC_INIT(1), | |
cdfda633 | 1232 | .done = 0, |
725a97e4 | 1233 | .flags = BM_AIO_COPY_PAGES, |
cdfda633 PR |
1234 | .error = 0, |
1235 | .kref = { ATOMIC_INIT(2) }, | |
725a97e4 | 1236 | }; |
b411b363 | 1237 | |
cdfda633 PR |
1238 | if (!get_ldev_if_state(mdev, D_ATTACHING)) { /* put is in bm_aio_ctx_destroy() */ |
1239 | dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in drbd_bm_write_page()\n"); | |
1240 | err = -ENODEV; | |
1241 | goto out; | |
b411b363 | 1242 | } |
19f843aa | 1243 | |
cdfda633 PR |
1244 | bm_page_io_async(ctx, idx, WRITE_SYNC); |
1245 | wait_until_done_or_disk_failure(mdev, &ctx->done); | |
19f843aa | 1246 | |
cdfda633 | 1247 | if (ctx->error) |
19f843aa LE |
1248 | drbd_chk_io_error(mdev, 1, true); |
1249 | /* that should force detach, so the in memory bitmap will be | |
1250 | * gone in a moment as well. */ | |
1251 | ||
b411b363 | 1252 | mdev->bm_writ_cnt++; |
cdfda633 PR |
1253 | err = atomic_read(&ctx->in_flight) ? -EIO : ctx->error; |
1254 | out: | |
1255 | kref_put(&ctx->kref, &bm_aio_ctx_destroy); | |
1256 | return err; | |
b411b363 PR |
1257 | } |
1258 | ||
1259 | /* NOTE | |
1260 | * find_first_bit returns int, we return unsigned long. | |
4b0715f0 LE |
1261 | * For this to work on 32bit arch with bitnumbers > (1<<32), |
1262 | * we'd need to return u64, and get a whole lot of other places | |
1263 | * fixed where we still use unsigned long. | |
b411b363 PR |
1264 | * |
1265 | * this returns a bit number, NOT a sector! | |
1266 | */ | |
b411b363 PR |
1267 | static unsigned long __bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo, |
1268 | const int find_zero_bit, const enum km_type km) | |
1269 | { | |
1270 | struct drbd_bitmap *b = mdev->bitmap; | |
b411b363 | 1271 | unsigned long *p_addr; |
4b0715f0 LE |
1272 | unsigned long bit_offset; |
1273 | unsigned i; | |
1274 | ||
b411b363 PR |
1275 | |
1276 | if (bm_fo > b->bm_bits) { | |
1277 | dev_err(DEV, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits); | |
4b0715f0 | 1278 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1279 | } else { |
1280 | while (bm_fo < b->bm_bits) { | |
19f843aa | 1281 | /* bit offset of the first bit in the page */ |
4b0715f0 | 1282 | bit_offset = bm_fo & ~BITS_PER_PAGE_MASK; |
19f843aa | 1283 | p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo), km); |
b411b363 PR |
1284 | |
1285 | if (find_zero_bit) | |
7e599e6e | 1286 | i = find_next_zero_bit_le(p_addr, |
4b0715f0 | 1287 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 | 1288 | else |
7e599e6e | 1289 | i = find_next_bit_le(p_addr, |
4b0715f0 | 1290 | PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); |
b411b363 PR |
1291 | |
1292 | __bm_unmap(p_addr, km); | |
1293 | if (i < PAGE_SIZE*8) { | |
4b0715f0 LE |
1294 | bm_fo = bit_offset + i; |
1295 | if (bm_fo >= b->bm_bits) | |
b411b363 PR |
1296 | break; |
1297 | goto found; | |
1298 | } | |
1299 | bm_fo = bit_offset + PAGE_SIZE*8; | |
1300 | } | |
4b0715f0 | 1301 | bm_fo = DRBD_END_OF_BITMAP; |
b411b363 PR |
1302 | } |
1303 | found: | |
4b0715f0 | 1304 | return bm_fo; |
b411b363 PR |
1305 | } |
1306 | ||
1307 | static unsigned long bm_find_next(struct drbd_conf *mdev, | |
1308 | unsigned long bm_fo, const int find_zero_bit) | |
1309 | { | |
1310 | struct drbd_bitmap *b = mdev->bitmap; | |
4b0715f0 | 1311 | unsigned long i = DRBD_END_OF_BITMAP; |
b411b363 | 1312 | |
841ce241 AG |
1313 | if (!expect(b)) |
1314 | return i; | |
1315 | if (!expect(b->bm_pages)) | |
1316 | return i; | |
b411b363 PR |
1317 | |
1318 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1319 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1320 | bm_print_lock_info(mdev); |
1321 | ||
1322 | i = __bm_find_next(mdev, bm_fo, find_zero_bit, KM_IRQ1); | |
1323 | ||
1324 | spin_unlock_irq(&b->bm_lock); | |
1325 | return i; | |
1326 | } | |
1327 | ||
1328 | unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo) | |
1329 | { | |
1330 | return bm_find_next(mdev, bm_fo, 0); | |
1331 | } | |
1332 | ||
1333 | #if 0 | |
1334 | /* not yet needed for anything. */ | |
1335 | unsigned long drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo) | |
1336 | { | |
1337 | return bm_find_next(mdev, bm_fo, 1); | |
1338 | } | |
1339 | #endif | |
1340 | ||
1341 | /* does not spin_lock_irqsave. | |
1342 | * you must take drbd_bm_lock() first */ | |
1343 | unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo) | |
1344 | { | |
20ceb2b2 | 1345 | /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */ |
b411b363 PR |
1346 | return __bm_find_next(mdev, bm_fo, 0, KM_USER1); |
1347 | } | |
1348 | ||
1349 | unsigned long _drbd_bm_find_next_zero(struct drbd_conf *mdev, unsigned long bm_fo) | |
1350 | { | |
20ceb2b2 | 1351 | /* WARN_ON(!(BM_DONT_SET & mdev->b->bm_flags)); */ |
b411b363 PR |
1352 | return __bm_find_next(mdev, bm_fo, 1, KM_USER1); |
1353 | } | |
1354 | ||
1355 | /* returns number of bits actually changed. | |
1356 | * for val != 0, we change 0 -> 1, return code positive | |
1357 | * for val == 0, we change 1 -> 0, return code negative | |
1358 | * wants bitnr, not sector. | |
1359 | * expected to be called for only a few bits (e - s about BITS_PER_LONG). | |
1360 | * Must hold bitmap lock already. */ | |
b4ee79da | 1361 | static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s, |
829c6087 | 1362 | unsigned long e, int val) |
b411b363 PR |
1363 | { |
1364 | struct drbd_bitmap *b = mdev->bitmap; | |
1365 | unsigned long *p_addr = NULL; | |
1366 | unsigned long bitnr; | |
19f843aa | 1367 | unsigned int last_page_nr = -1U; |
b411b363 | 1368 | int c = 0; |
19f843aa | 1369 | int changed_total = 0; |
b411b363 PR |
1370 | |
1371 | if (e >= b->bm_bits) { | |
1372 | dev_err(DEV, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n", | |
1373 | s, e, b->bm_bits); | |
1374 | e = b->bm_bits ? b->bm_bits -1 : 0; | |
1375 | } | |
1376 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa | 1377 | unsigned int page_nr = bm_bit_to_page_idx(b, bitnr); |
b411b363 PR |
1378 | if (page_nr != last_page_nr) { |
1379 | if (p_addr) | |
829c6087 | 1380 | __bm_unmap(p_addr, KM_IRQ1); |
19f843aa LE |
1381 | if (c < 0) |
1382 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1383 | else if (c > 0) | |
1384 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1385 | changed_total += c; | |
1386 | c = 0; | |
829c6087 | 1387 | p_addr = __bm_map_pidx(b, page_nr, KM_IRQ1); |
b411b363 PR |
1388 | last_page_nr = page_nr; |
1389 | } | |
1390 | if (val) | |
7e599e6e | 1391 | c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 | 1392 | else |
7e599e6e | 1393 | c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); |
b411b363 PR |
1394 | } |
1395 | if (p_addr) | |
829c6087 | 1396 | __bm_unmap(p_addr, KM_IRQ1); |
19f843aa LE |
1397 | if (c < 0) |
1398 | bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); | |
1399 | else if (c > 0) | |
1400 | bm_set_page_need_writeout(b->bm_pages[last_page_nr]); | |
1401 | changed_total += c; | |
1402 | b->bm_set += changed_total; | |
1403 | return changed_total; | |
b411b363 PR |
1404 | } |
1405 | ||
1406 | /* returns number of bits actually changed. | |
1407 | * for val != 0, we change 0 -> 1, return code positive | |
1408 | * for val == 0, we change 1 -> 0, return code negative | |
1409 | * wants bitnr, not sector */ | |
b4ee79da | 1410 | static int bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s, |
b411b363 PR |
1411 | const unsigned long e, int val) |
1412 | { | |
1413 | unsigned long flags; | |
1414 | struct drbd_bitmap *b = mdev->bitmap; | |
1415 | int c = 0; | |
1416 | ||
841ce241 AG |
1417 | if (!expect(b)) |
1418 | return 1; | |
1419 | if (!expect(b->bm_pages)) | |
1420 | return 0; | |
b411b363 PR |
1421 | |
1422 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1423 | if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags) |
b411b363 PR |
1424 | bm_print_lock_info(mdev); |
1425 | ||
829c6087 | 1426 | c = __bm_change_bits_to(mdev, s, e, val); |
b411b363 PR |
1427 | |
1428 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1429 | return c; | |
1430 | } | |
1431 | ||
1432 | /* returns number of bits changed 0 -> 1 */ | |
1433 | int drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1434 | { | |
1435 | return bm_change_bits_to(mdev, s, e, 1); | |
1436 | } | |
1437 | ||
1438 | /* returns number of bits changed 1 -> 0 */ | |
1439 | int drbd_bm_clear_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1440 | { | |
1441 | return -bm_change_bits_to(mdev, s, e, 0); | |
1442 | } | |
1443 | ||
1444 | /* sets all bits in full words, | |
1445 | * from first_word up to, but not including, last_word */ | |
1446 | static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b, | |
1447 | int page_nr, int first_word, int last_word) | |
1448 | { | |
1449 | int i; | |
1450 | int bits; | |
829c6087 | 1451 | unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr], KM_IRQ1); |
b411b363 PR |
1452 | for (i = first_word; i < last_word; i++) { |
1453 | bits = hweight_long(paddr[i]); | |
1454 | paddr[i] = ~0UL; | |
1455 | b->bm_set += BITS_PER_LONG - bits; | |
1456 | } | |
829c6087 | 1457 | kunmap_atomic(paddr, KM_IRQ1); |
b411b363 PR |
1458 | } |
1459 | ||
829c6087 LE |
1460 | /* Same thing as drbd_bm_set_bits, |
1461 | * but more efficient for a large bit range. | |
b411b363 PR |
1462 | * You must first drbd_bm_lock(). |
1463 | * Can be called to set the whole bitmap in one go. | |
1464 | * Sets bits from s to e _inclusive_. */ | |
1465 | void _drbd_bm_set_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1466 | { | |
1467 | /* First set_bit from the first bit (s) | |
1468 | * up to the next long boundary (sl), | |
1469 | * then assign full words up to the last long boundary (el), | |
1470 | * then set_bit up to and including the last bit (e). | |
1471 | * | |
1472 | * Do not use memset, because we must account for changes, | |
1473 | * so we need to loop over the words with hweight() anyways. | |
1474 | */ | |
829c6087 | 1475 | struct drbd_bitmap *b = mdev->bitmap; |
b411b363 PR |
1476 | unsigned long sl = ALIGN(s,BITS_PER_LONG); |
1477 | unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1); | |
1478 | int first_page; | |
1479 | int last_page; | |
1480 | int page_nr; | |
1481 | int first_word; | |
1482 | int last_word; | |
1483 | ||
1484 | if (e - s <= 3*BITS_PER_LONG) { | |
1485 | /* don't bother; el and sl may even be wrong. */ | |
829c6087 LE |
1486 | spin_lock_irq(&b->bm_lock); |
1487 | __bm_change_bits_to(mdev, s, e, 1); | |
1488 | spin_unlock_irq(&b->bm_lock); | |
b411b363 PR |
1489 | return; |
1490 | } | |
1491 | ||
1492 | /* difference is large enough that we can trust sl and el */ | |
1493 | ||
829c6087 LE |
1494 | spin_lock_irq(&b->bm_lock); |
1495 | ||
b411b363 PR |
1496 | /* bits filling the current long */ |
1497 | if (sl) | |
829c6087 | 1498 | __bm_change_bits_to(mdev, s, sl-1, 1); |
b411b363 PR |
1499 | |
1500 | first_page = sl >> (3 + PAGE_SHIFT); | |
1501 | last_page = el >> (3 + PAGE_SHIFT); | |
1502 | ||
1503 | /* MLPP: modulo longs per page */ | |
1504 | /* LWPP: long words per page */ | |
1505 | first_word = MLPP(sl >> LN2_BPL); | |
1506 | last_word = LWPP; | |
1507 | ||
1508 | /* first and full pages, unless first page == last page */ | |
1509 | for (page_nr = first_page; page_nr < last_page; page_nr++) { | |
1510 | bm_set_full_words_within_one_page(mdev->bitmap, page_nr, first_word, last_word); | |
8ccee20e LE |
1511 | spin_unlock_irq(&b->bm_lock); |
1512 | cond_resched(); | |
b411b363 | 1513 | first_word = 0; |
8ccee20e | 1514 | spin_lock_irq(&b->bm_lock); |
b411b363 PR |
1515 | } |
1516 | ||
1517 | /* last page (respectively only page, for first page == last page) */ | |
1518 | last_word = MLPP(el >> LN2_BPL); | |
1519 | bm_set_full_words_within_one_page(mdev->bitmap, last_page, first_word, last_word); | |
1520 | ||
1521 | /* possibly trailing bits. | |
1522 | * example: (e & 63) == 63, el will be e+1. | |
1523 | * if that even was the very last bit, | |
1524 | * it would trigger an assert in __bm_change_bits_to() | |
1525 | */ | |
1526 | if (el <= e) | |
829c6087 LE |
1527 | __bm_change_bits_to(mdev, el, e, 1); |
1528 | spin_unlock_irq(&b->bm_lock); | |
b411b363 PR |
1529 | } |
1530 | ||
1531 | /* returns bit state | |
1532 | * wants bitnr, NOT sector. | |
1533 | * inherently racy... area needs to be locked by means of {al,rs}_lru | |
1534 | * 1 ... bit set | |
1535 | * 0 ... bit not set | |
1536 | * -1 ... first out of bounds access, stop testing for bits! | |
1537 | */ | |
1538 | int drbd_bm_test_bit(struct drbd_conf *mdev, const unsigned long bitnr) | |
1539 | { | |
1540 | unsigned long flags; | |
1541 | struct drbd_bitmap *b = mdev->bitmap; | |
1542 | unsigned long *p_addr; | |
1543 | int i; | |
1544 | ||
841ce241 AG |
1545 | if (!expect(b)) |
1546 | return 0; | |
1547 | if (!expect(b->bm_pages)) | |
1548 | return 0; | |
b411b363 PR |
1549 | |
1550 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1551 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1552 | bm_print_lock_info(mdev); |
1553 | if (bitnr < b->bm_bits) { | |
19f843aa | 1554 | p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr)); |
7e599e6e | 1555 | i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0; |
b411b363 PR |
1556 | bm_unmap(p_addr); |
1557 | } else if (bitnr == b->bm_bits) { | |
1558 | i = -1; | |
1559 | } else { /* (bitnr > b->bm_bits) */ | |
1560 | dev_err(DEV, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits); | |
1561 | i = 0; | |
1562 | } | |
1563 | ||
1564 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1565 | return i; | |
1566 | } | |
1567 | ||
1568 | /* returns number of bits set in the range [s, e] */ | |
1569 | int drbd_bm_count_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e) | |
1570 | { | |
1571 | unsigned long flags; | |
1572 | struct drbd_bitmap *b = mdev->bitmap; | |
19f843aa | 1573 | unsigned long *p_addr = NULL; |
b411b363 | 1574 | unsigned long bitnr; |
19f843aa | 1575 | unsigned int page_nr = -1U; |
b411b363 | 1576 | int c = 0; |
b411b363 PR |
1577 | |
1578 | /* If this is called without a bitmap, that is a bug. But just to be | |
1579 | * robust in case we screwed up elsewhere, in that case pretend there | |
1580 | * was one dirty bit in the requested area, so we won't try to do a | |
1581 | * local read there (no bitmap probably implies no disk) */ | |
841ce241 AG |
1582 | if (!expect(b)) |
1583 | return 1; | |
1584 | if (!expect(b->bm_pages)) | |
1585 | return 1; | |
b411b363 PR |
1586 | |
1587 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1588 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1589 | bm_print_lock_info(mdev); |
1590 | for (bitnr = s; bitnr <= e; bitnr++) { | |
19f843aa LE |
1591 | unsigned int idx = bm_bit_to_page_idx(b, bitnr); |
1592 | if (page_nr != idx) { | |
1593 | page_nr = idx; | |
b411b363 PR |
1594 | if (p_addr) |
1595 | bm_unmap(p_addr); | |
19f843aa | 1596 | p_addr = bm_map_pidx(b, idx); |
b411b363 | 1597 | } |
841ce241 | 1598 | if (expect(bitnr < b->bm_bits)) |
7e599e6e | 1599 | c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr)); |
841ce241 AG |
1600 | else |
1601 | dev_err(DEV, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits); | |
b411b363 PR |
1602 | } |
1603 | if (p_addr) | |
1604 | bm_unmap(p_addr); | |
1605 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1606 | return c; | |
1607 | } | |
1608 | ||
1609 | ||
1610 | /* inherently racy... | |
1611 | * return value may be already out-of-date when this function returns. | |
1612 | * but the general usage is that this is only use during a cstate when bits are | |
1613 | * only cleared, not set, and typically only care for the case when the return | |
1614 | * value is zero, or we already "locked" this "bitmap extent" by other means. | |
1615 | * | |
1616 | * enr is bm-extent number, since we chose to name one sector (512 bytes) | |
1617 | * worth of the bitmap a "bitmap extent". | |
1618 | * | |
1619 | * TODO | |
1620 | * I think since we use it like a reference count, we should use the real | |
1621 | * reference count of some bitmap extent element from some lru instead... | |
1622 | * | |
1623 | */ | |
1624 | int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr) | |
1625 | { | |
1626 | struct drbd_bitmap *b = mdev->bitmap; | |
1627 | int count, s, e; | |
1628 | unsigned long flags; | |
1629 | unsigned long *p_addr, *bm; | |
1630 | ||
841ce241 AG |
1631 | if (!expect(b)) |
1632 | return 0; | |
1633 | if (!expect(b->bm_pages)) | |
1634 | return 0; | |
b411b363 PR |
1635 | |
1636 | spin_lock_irqsave(&b->bm_lock, flags); | |
20ceb2b2 | 1637 | if (BM_DONT_TEST & b->bm_flags) |
b411b363 PR |
1638 | bm_print_lock_info(mdev); |
1639 | ||
1640 | s = S2W(enr); | |
1641 | e = min((size_t)S2W(enr+1), b->bm_words); | |
1642 | count = 0; | |
1643 | if (s < b->bm_words) { | |
1644 | int n = e-s; | |
19f843aa | 1645 | p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); |
b411b363 PR |
1646 | bm = p_addr + MLPP(s); |
1647 | while (n--) | |
1648 | count += hweight_long(*bm++); | |
1649 | bm_unmap(p_addr); | |
1650 | } else { | |
1651 | dev_err(DEV, "start offset (%d) too large in drbd_bm_e_weight\n", s); | |
1652 | } | |
1653 | spin_unlock_irqrestore(&b->bm_lock, flags); | |
1654 | return count; | |
1655 | } | |
1656 | ||
4b0715f0 LE |
1657 | /* Set all bits covered by the AL-extent al_enr. |
1658 | * Returns number of bits changed. */ | |
b411b363 PR |
1659 | unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev, unsigned long al_enr) |
1660 | { | |
1661 | struct drbd_bitmap *b = mdev->bitmap; | |
1662 | unsigned long *p_addr, *bm; | |
1663 | unsigned long weight; | |
4b0715f0 LE |
1664 | unsigned long s, e; |
1665 | int count, i, do_now; | |
841ce241 AG |
1666 | if (!expect(b)) |
1667 | return 0; | |
1668 | if (!expect(b->bm_pages)) | |
1669 | return 0; | |
b411b363 PR |
1670 | |
1671 | spin_lock_irq(&b->bm_lock); | |
20ceb2b2 | 1672 | if (BM_DONT_SET & b->bm_flags) |
b411b363 PR |
1673 | bm_print_lock_info(mdev); |
1674 | weight = b->bm_set; | |
1675 | ||
1676 | s = al_enr * BM_WORDS_PER_AL_EXT; | |
1677 | e = min_t(size_t, s + BM_WORDS_PER_AL_EXT, b->bm_words); | |
1678 | /* assert that s and e are on the same page */ | |
1679 | D_ASSERT((e-1) >> (PAGE_SHIFT - LN2_BPL + 3) | |
1680 | == s >> (PAGE_SHIFT - LN2_BPL + 3)); | |
1681 | count = 0; | |
1682 | if (s < b->bm_words) { | |
1683 | i = do_now = 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 (i--) { | |
1687 | count += hweight_long(*bm); | |
1688 | *bm = -1UL; | |
1689 | bm++; | |
1690 | } | |
1691 | bm_unmap(p_addr); | |
1692 | b->bm_set += do_now*BITS_PER_LONG - count; | |
1693 | if (e == b->bm_words) | |
1694 | b->bm_set -= bm_clear_surplus(b); | |
1695 | } else { | |
4b0715f0 | 1696 | dev_err(DEV, "start offset (%lu) too large in drbd_bm_ALe_set_all\n", s); |
b411b363 PR |
1697 | } |
1698 | weight = b->bm_set - weight; | |
1699 | spin_unlock_irq(&b->bm_lock); | |
1700 | return weight; | |
1701 | } |