Commit | Line | Data |
---|---|---|
b411b363 PR |
1 | /* |
2 | drbd_actlog.c | |
3 | ||
4 | This file is part of DRBD by Philipp Reisner and Lars Ellenberg. | |
5 | ||
6 | Copyright (C) 2003-2008, LINBIT Information Technologies GmbH. | |
7 | Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>. | |
8 | Copyright (C) 2003-2008, 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 | ||
26 | #include <linux/slab.h> | |
27 | #include <linux/drbd.h> | |
28 | #include "drbd_int.h" | |
b411b363 PR |
29 | #include "drbd_wrappers.h" |
30 | ||
31 | /* We maintain a trivial check sum in our on disk activity log. | |
32 | * With that we can ensure correct operation even when the storage | |
33 | * device might do a partial (last) sector write while loosing power. | |
34 | */ | |
35 | struct __packed al_transaction { | |
36 | u32 magic; | |
37 | u32 tr_number; | |
38 | struct __packed { | |
39 | u32 pos; | |
40 | u32 extent; } updates[1 + AL_EXTENTS_PT]; | |
41 | u32 xor_sum; | |
42 | }; | |
43 | ||
44 | struct update_odbm_work { | |
45 | struct drbd_work w; | |
46 | unsigned int enr; | |
47 | }; | |
48 | ||
49 | struct update_al_work { | |
50 | struct drbd_work w; | |
51 | struct lc_element *al_ext; | |
52 | struct completion event; | |
53 | unsigned int enr; | |
54 | /* if old_enr != LC_FREE, write corresponding bitmap sector, too */ | |
55 | unsigned int old_enr; | |
56 | }; | |
57 | ||
58 | struct drbd_atodb_wait { | |
59 | atomic_t count; | |
60 | struct completion io_done; | |
61 | struct drbd_conf *mdev; | |
62 | int error; | |
63 | }; | |
64 | ||
65 | ||
66 | int w_al_write_transaction(struct drbd_conf *, struct drbd_work *, int); | |
67 | ||
b411b363 PR |
68 | static int _drbd_md_sync_page_io(struct drbd_conf *mdev, |
69 | struct drbd_backing_dev *bdev, | |
70 | struct page *page, sector_t sector, | |
71 | int rw, int size) | |
72 | { | |
73 | struct bio *bio; | |
74 | struct drbd_md_io md_io; | |
75 | int ok; | |
76 | ||
77 | md_io.mdev = mdev; | |
78 | init_completion(&md_io.event); | |
79 | md_io.error = 0; | |
80 | ||
a8a4e51e PR |
81 | if ((rw & WRITE) && !test_bit(MD_NO_FUA, &mdev->flags)) |
82 | rw |= REQ_FUA; | |
721a9602 | 83 | rw |= REQ_SYNC; |
b411b363 | 84 | |
b411b363 PR |
85 | bio = bio_alloc(GFP_NOIO, 1); |
86 | bio->bi_bdev = bdev->md_bdev; | |
87 | bio->bi_sector = sector; | |
88 | ok = (bio_add_page(bio, page, size, 0) == size); | |
89 | if (!ok) | |
90 | goto out; | |
91 | bio->bi_private = &md_io; | |
92 | bio->bi_end_io = drbd_md_io_complete; | |
93 | bio->bi_rw = rw; | |
94 | ||
b411b363 PR |
95 | if (FAULT_ACTIVE(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) |
96 | bio_endio(bio, -EIO); | |
97 | else | |
98 | submit_bio(rw, bio); | |
99 | wait_for_completion(&md_io.event); | |
100 | ok = bio_flagged(bio, BIO_UPTODATE) && md_io.error == 0; | |
101 | ||
b411b363 PR |
102 | out: |
103 | bio_put(bio); | |
104 | return ok; | |
105 | } | |
106 | ||
107 | int drbd_md_sync_page_io(struct drbd_conf *mdev, struct drbd_backing_dev *bdev, | |
108 | sector_t sector, int rw) | |
109 | { | |
110 | int logical_block_size, mask, ok; | |
111 | int offset = 0; | |
112 | struct page *iop = mdev->md_io_page; | |
113 | ||
114 | D_ASSERT(mutex_is_locked(&mdev->md_io_mutex)); | |
115 | ||
116 | BUG_ON(!bdev->md_bdev); | |
117 | ||
118 | logical_block_size = bdev_logical_block_size(bdev->md_bdev); | |
119 | if (logical_block_size == 0) | |
120 | logical_block_size = MD_SECTOR_SIZE; | |
121 | ||
122 | /* in case logical_block_size != 512 [ s390 only? ] */ | |
123 | if (logical_block_size != MD_SECTOR_SIZE) { | |
124 | mask = (logical_block_size / MD_SECTOR_SIZE) - 1; | |
125 | D_ASSERT(mask == 1 || mask == 3 || mask == 7); | |
126 | D_ASSERT(logical_block_size == (mask+1) * MD_SECTOR_SIZE); | |
127 | offset = sector & mask; | |
128 | sector = sector & ~mask; | |
129 | iop = mdev->md_io_tmpp; | |
130 | ||
131 | if (rw & WRITE) { | |
132 | /* these are GFP_KERNEL pages, pre-allocated | |
133 | * on device initialization */ | |
134 | void *p = page_address(mdev->md_io_page); | |
135 | void *hp = page_address(mdev->md_io_tmpp); | |
136 | ||
137 | ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, | |
138 | READ, logical_block_size); | |
139 | ||
140 | if (unlikely(!ok)) { | |
141 | dev_err(DEV, "drbd_md_sync_page_io(,%llus," | |
142 | "READ [logical_block_size!=512]) failed!\n", | |
143 | (unsigned long long)sector); | |
144 | return 0; | |
145 | } | |
146 | ||
147 | memcpy(hp + offset*MD_SECTOR_SIZE, p, MD_SECTOR_SIZE); | |
148 | } | |
149 | } | |
150 | ||
151 | if (sector < drbd_md_first_sector(bdev) || | |
152 | sector > drbd_md_last_sector(bdev)) | |
153 | dev_alert(DEV, "%s [%d]:%s(,%llus,%s) out of range md access!\n", | |
154 | current->comm, current->pid, __func__, | |
155 | (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ"); | |
156 | ||
157 | ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, rw, logical_block_size); | |
158 | if (unlikely(!ok)) { | |
159 | dev_err(DEV, "drbd_md_sync_page_io(,%llus,%s) failed!\n", | |
160 | (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ"); | |
161 | return 0; | |
162 | } | |
163 | ||
164 | if (logical_block_size != MD_SECTOR_SIZE && !(rw & WRITE)) { | |
165 | void *p = page_address(mdev->md_io_page); | |
166 | void *hp = page_address(mdev->md_io_tmpp); | |
167 | ||
168 | memcpy(p, hp + offset*MD_SECTOR_SIZE, MD_SECTOR_SIZE); | |
169 | } | |
170 | ||
171 | return ok; | |
172 | } | |
173 | ||
174 | static struct lc_element *_al_get(struct drbd_conf *mdev, unsigned int enr) | |
175 | { | |
176 | struct lc_element *al_ext; | |
177 | struct lc_element *tmp; | |
178 | unsigned long al_flags = 0; | |
179 | ||
180 | spin_lock_irq(&mdev->al_lock); | |
181 | tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT); | |
182 | if (unlikely(tmp != NULL)) { | |
183 | struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce); | |
184 | if (test_bit(BME_NO_WRITES, &bm_ext->flags)) { | |
185 | spin_unlock_irq(&mdev->al_lock); | |
186 | return NULL; | |
187 | } | |
188 | } | |
189 | al_ext = lc_get(mdev->act_log, enr); | |
190 | al_flags = mdev->act_log->flags; | |
191 | spin_unlock_irq(&mdev->al_lock); | |
192 | ||
193 | /* | |
194 | if (!al_ext) { | |
195 | if (al_flags & LC_STARVING) | |
196 | dev_warn(DEV, "Have to wait for LRU element (AL too small?)\n"); | |
197 | if (al_flags & LC_DIRTY) | |
198 | dev_warn(DEV, "Ongoing AL update (AL device too slow?)\n"); | |
199 | } | |
200 | */ | |
201 | ||
202 | return al_ext; | |
203 | } | |
204 | ||
205 | void drbd_al_begin_io(struct drbd_conf *mdev, sector_t sector) | |
206 | { | |
207 | unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9)); | |
208 | struct lc_element *al_ext; | |
209 | struct update_al_work al_work; | |
210 | ||
211 | D_ASSERT(atomic_read(&mdev->local_cnt) > 0); | |
212 | ||
b411b363 PR |
213 | wait_event(mdev->al_wait, (al_ext = _al_get(mdev, enr))); |
214 | ||
215 | if (al_ext->lc_number != enr) { | |
216 | /* drbd_al_write_transaction(mdev,al_ext,enr); | |
217 | * recurses into generic_make_request(), which | |
218 | * disallows recursion, bios being serialized on the | |
219 | * current->bio_tail list now. | |
220 | * we have to delegate updates to the activity log | |
221 | * to the worker thread. */ | |
222 | init_completion(&al_work.event); | |
223 | al_work.al_ext = al_ext; | |
224 | al_work.enr = enr; | |
225 | al_work.old_enr = al_ext->lc_number; | |
226 | al_work.w.cb = w_al_write_transaction; | |
227 | drbd_queue_work_front(&mdev->data.work, &al_work.w); | |
228 | wait_for_completion(&al_work.event); | |
229 | ||
230 | mdev->al_writ_cnt++; | |
231 | ||
232 | spin_lock_irq(&mdev->al_lock); | |
233 | lc_changed(mdev->act_log, al_ext); | |
234 | spin_unlock_irq(&mdev->al_lock); | |
235 | wake_up(&mdev->al_wait); | |
236 | } | |
237 | } | |
238 | ||
239 | void drbd_al_complete_io(struct drbd_conf *mdev, sector_t sector) | |
240 | { | |
241 | unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9)); | |
242 | struct lc_element *extent; | |
243 | unsigned long flags; | |
244 | ||
b411b363 PR |
245 | spin_lock_irqsave(&mdev->al_lock, flags); |
246 | ||
247 | extent = lc_find(mdev->act_log, enr); | |
248 | ||
249 | if (!extent) { | |
250 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
251 | dev_err(DEV, "al_complete_io() called on inactive extent %u\n", enr); | |
252 | return; | |
253 | } | |
254 | ||
255 | if (lc_put(mdev->act_log, extent) == 0) | |
256 | wake_up(&mdev->al_wait); | |
257 | ||
258 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
259 | } | |
260 | ||
261 | int | |
262 | w_al_write_transaction(struct drbd_conf *mdev, struct drbd_work *w, int unused) | |
263 | { | |
264 | struct update_al_work *aw = container_of(w, struct update_al_work, w); | |
265 | struct lc_element *updated = aw->al_ext; | |
266 | const unsigned int new_enr = aw->enr; | |
267 | const unsigned int evicted = aw->old_enr; | |
268 | struct al_transaction *buffer; | |
269 | sector_t sector; | |
270 | int i, n, mx; | |
271 | unsigned int extent_nr; | |
272 | u32 xor_sum = 0; | |
273 | ||
274 | if (!get_ldev(mdev)) { | |
6719fb03 LE |
275 | dev_err(DEV, |
276 | "disk is %s, cannot start al transaction (-%d +%d)\n", | |
277 | drbd_disk_str(mdev->state.disk), evicted, new_enr); | |
b411b363 PR |
278 | complete(&((struct update_al_work *)w)->event); |
279 | return 1; | |
280 | } | |
281 | /* do we have to do a bitmap write, first? | |
282 | * TODO reduce maximum latency: | |
283 | * submit both bios, then wait for both, | |
6719fb03 LE |
284 | * instead of doing two synchronous sector writes. |
285 | * For now, we must not write the transaction, | |
286 | * if we cannot write out the bitmap of the evicted extent. */ | |
b411b363 PR |
287 | if (mdev->state.conn < C_CONNECTED && evicted != LC_FREE) |
288 | drbd_bm_write_sect(mdev, evicted/AL_EXT_PER_BM_SECT); | |
289 | ||
6719fb03 LE |
290 | /* The bitmap write may have failed, causing a state change. */ |
291 | if (mdev->state.disk < D_INCONSISTENT) { | |
292 | dev_err(DEV, | |
293 | "disk is %s, cannot write al transaction (-%d +%d)\n", | |
294 | drbd_disk_str(mdev->state.disk), evicted, new_enr); | |
295 | complete(&((struct update_al_work *)w)->event); | |
296 | put_ldev(mdev); | |
297 | return 1; | |
298 | } | |
299 | ||
300 | mutex_lock(&mdev->md_io_mutex); /* protects md_io_buffer, al_tr_cycle, ... */ | |
b411b363 PR |
301 | buffer = (struct al_transaction *)page_address(mdev->md_io_page); |
302 | ||
303 | buffer->magic = __constant_cpu_to_be32(DRBD_MAGIC); | |
304 | buffer->tr_number = cpu_to_be32(mdev->al_tr_number); | |
305 | ||
306 | n = lc_index_of(mdev->act_log, updated); | |
307 | ||
308 | buffer->updates[0].pos = cpu_to_be32(n); | |
309 | buffer->updates[0].extent = cpu_to_be32(new_enr); | |
310 | ||
311 | xor_sum ^= new_enr; | |
312 | ||
313 | mx = min_t(int, AL_EXTENTS_PT, | |
314 | mdev->act_log->nr_elements - mdev->al_tr_cycle); | |
315 | for (i = 0; i < mx; i++) { | |
316 | unsigned idx = mdev->al_tr_cycle + i; | |
317 | extent_nr = lc_element_by_index(mdev->act_log, idx)->lc_number; | |
318 | buffer->updates[i+1].pos = cpu_to_be32(idx); | |
319 | buffer->updates[i+1].extent = cpu_to_be32(extent_nr); | |
320 | xor_sum ^= extent_nr; | |
321 | } | |
322 | for (; i < AL_EXTENTS_PT; i++) { | |
323 | buffer->updates[i+1].pos = __constant_cpu_to_be32(-1); | |
324 | buffer->updates[i+1].extent = __constant_cpu_to_be32(LC_FREE); | |
325 | xor_sum ^= LC_FREE; | |
326 | } | |
327 | mdev->al_tr_cycle += AL_EXTENTS_PT; | |
328 | if (mdev->al_tr_cycle >= mdev->act_log->nr_elements) | |
329 | mdev->al_tr_cycle = 0; | |
330 | ||
331 | buffer->xor_sum = cpu_to_be32(xor_sum); | |
332 | ||
333 | sector = mdev->ldev->md.md_offset | |
334 | + mdev->ldev->md.al_offset + mdev->al_tr_pos; | |
335 | ||
336 | if (!drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE)) | |
337 | drbd_chk_io_error(mdev, 1, TRUE); | |
338 | ||
339 | if (++mdev->al_tr_pos > | |
340 | div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT)) | |
341 | mdev->al_tr_pos = 0; | |
342 | ||
343 | D_ASSERT(mdev->al_tr_pos < MD_AL_MAX_SIZE); | |
344 | mdev->al_tr_number++; | |
345 | ||
346 | mutex_unlock(&mdev->md_io_mutex); | |
347 | ||
348 | complete(&((struct update_al_work *)w)->event); | |
349 | put_ldev(mdev); | |
350 | ||
351 | return 1; | |
352 | } | |
353 | ||
354 | /** | |
355 | * drbd_al_read_tr() - Read a single transaction from the on disk activity log | |
356 | * @mdev: DRBD device. | |
357 | * @bdev: Block device to read form. | |
358 | * @b: pointer to an al_transaction. | |
359 | * @index: On disk slot of the transaction to read. | |
360 | * | |
361 | * Returns -1 on IO error, 0 on checksum error and 1 upon success. | |
362 | */ | |
363 | static int drbd_al_read_tr(struct drbd_conf *mdev, | |
364 | struct drbd_backing_dev *bdev, | |
365 | struct al_transaction *b, | |
366 | int index) | |
367 | { | |
368 | sector_t sector; | |
369 | int rv, i; | |
370 | u32 xor_sum = 0; | |
371 | ||
372 | sector = bdev->md.md_offset + bdev->md.al_offset + index; | |
373 | ||
374 | /* Dont process error normally, | |
375 | * as this is done before disk is attached! */ | |
376 | if (!drbd_md_sync_page_io(mdev, bdev, sector, READ)) | |
377 | return -1; | |
378 | ||
379 | rv = (be32_to_cpu(b->magic) == DRBD_MAGIC); | |
380 | ||
381 | for (i = 0; i < AL_EXTENTS_PT + 1; i++) | |
382 | xor_sum ^= be32_to_cpu(b->updates[i].extent); | |
383 | rv &= (xor_sum == be32_to_cpu(b->xor_sum)); | |
384 | ||
385 | return rv; | |
386 | } | |
387 | ||
388 | /** | |
389 | * drbd_al_read_log() - Restores the activity log from its on disk representation. | |
390 | * @mdev: DRBD device. | |
391 | * @bdev: Block device to read form. | |
392 | * | |
393 | * Returns 1 on success, returns 0 when reading the log failed due to IO errors. | |
394 | */ | |
395 | int drbd_al_read_log(struct drbd_conf *mdev, struct drbd_backing_dev *bdev) | |
396 | { | |
397 | struct al_transaction *buffer; | |
398 | int i; | |
399 | int rv; | |
400 | int mx; | |
401 | int active_extents = 0; | |
402 | int transactions = 0; | |
403 | int found_valid = 0; | |
404 | int from = 0; | |
405 | int to = 0; | |
406 | u32 from_tnr = 0; | |
407 | u32 to_tnr = 0; | |
408 | u32 cnr; | |
409 | ||
410 | mx = div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT); | |
411 | ||
412 | /* lock out all other meta data io for now, | |
413 | * and make sure the page is mapped. | |
414 | */ | |
415 | mutex_lock(&mdev->md_io_mutex); | |
416 | buffer = page_address(mdev->md_io_page); | |
417 | ||
418 | /* Find the valid transaction in the log */ | |
419 | for (i = 0; i <= mx; i++) { | |
420 | rv = drbd_al_read_tr(mdev, bdev, buffer, i); | |
421 | if (rv == 0) | |
422 | continue; | |
423 | if (rv == -1) { | |
424 | mutex_unlock(&mdev->md_io_mutex); | |
425 | return 0; | |
426 | } | |
427 | cnr = be32_to_cpu(buffer->tr_number); | |
428 | ||
429 | if (++found_valid == 1) { | |
430 | from = i; | |
431 | to = i; | |
432 | from_tnr = cnr; | |
433 | to_tnr = cnr; | |
434 | continue; | |
435 | } | |
436 | if ((int)cnr - (int)from_tnr < 0) { | |
437 | D_ASSERT(from_tnr - cnr + i - from == mx+1); | |
438 | from = i; | |
439 | from_tnr = cnr; | |
440 | } | |
441 | if ((int)cnr - (int)to_tnr > 0) { | |
442 | D_ASSERT(cnr - to_tnr == i - to); | |
443 | to = i; | |
444 | to_tnr = cnr; | |
445 | } | |
446 | } | |
447 | ||
448 | if (!found_valid) { | |
449 | dev_warn(DEV, "No usable activity log found.\n"); | |
450 | mutex_unlock(&mdev->md_io_mutex); | |
451 | return 1; | |
452 | } | |
453 | ||
454 | /* Read the valid transactions. | |
455 | * dev_info(DEV, "Reading from %d to %d.\n",from,to); */ | |
456 | i = from; | |
457 | while (1) { | |
458 | int j, pos; | |
459 | unsigned int extent_nr; | |
460 | unsigned int trn; | |
461 | ||
462 | rv = drbd_al_read_tr(mdev, bdev, buffer, i); | |
463 | ERR_IF(rv == 0) goto cancel; | |
464 | if (rv == -1) { | |
465 | mutex_unlock(&mdev->md_io_mutex); | |
466 | return 0; | |
467 | } | |
468 | ||
469 | trn = be32_to_cpu(buffer->tr_number); | |
470 | ||
471 | spin_lock_irq(&mdev->al_lock); | |
472 | ||
473 | /* This loop runs backwards because in the cyclic | |
474 | elements there might be an old version of the | |
475 | updated element (in slot 0). So the element in slot 0 | |
476 | can overwrite old versions. */ | |
477 | for (j = AL_EXTENTS_PT; j >= 0; j--) { | |
478 | pos = be32_to_cpu(buffer->updates[j].pos); | |
479 | extent_nr = be32_to_cpu(buffer->updates[j].extent); | |
480 | ||
481 | if (extent_nr == LC_FREE) | |
482 | continue; | |
483 | ||
484 | lc_set(mdev->act_log, extent_nr, pos); | |
485 | active_extents++; | |
486 | } | |
487 | spin_unlock_irq(&mdev->al_lock); | |
488 | ||
489 | transactions++; | |
490 | ||
491 | cancel: | |
492 | if (i == to) | |
493 | break; | |
494 | i++; | |
495 | if (i > mx) | |
496 | i = 0; | |
497 | } | |
498 | ||
499 | mdev->al_tr_number = to_tnr+1; | |
500 | mdev->al_tr_pos = to; | |
501 | if (++mdev->al_tr_pos > | |
502 | div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT)) | |
503 | mdev->al_tr_pos = 0; | |
504 | ||
505 | /* ok, we are done with it */ | |
506 | mutex_unlock(&mdev->md_io_mutex); | |
507 | ||
508 | dev_info(DEV, "Found %d transactions (%d active extents) in activity log.\n", | |
509 | transactions, active_extents); | |
510 | ||
511 | return 1; | |
512 | } | |
513 | ||
514 | static void atodb_endio(struct bio *bio, int error) | |
515 | { | |
516 | struct drbd_atodb_wait *wc = bio->bi_private; | |
517 | struct drbd_conf *mdev = wc->mdev; | |
518 | struct page *page; | |
519 | int uptodate = bio_flagged(bio, BIO_UPTODATE); | |
520 | ||
521 | /* strange behavior of some lower level drivers... | |
522 | * fail the request by clearing the uptodate flag, | |
523 | * but do not return any error?! */ | |
524 | if (!error && !uptodate) | |
525 | error = -EIO; | |
526 | ||
527 | drbd_chk_io_error(mdev, error, TRUE); | |
528 | if (error && wc->error == 0) | |
529 | wc->error = error; | |
530 | ||
531 | if (atomic_dec_and_test(&wc->count)) | |
532 | complete(&wc->io_done); | |
533 | ||
534 | page = bio->bi_io_vec[0].bv_page; | |
535 | put_page(page); | |
536 | bio_put(bio); | |
537 | mdev->bm_writ_cnt++; | |
538 | put_ldev(mdev); | |
539 | } | |
540 | ||
39ad2bbb | 541 | /* sector to word */ |
b411b363 | 542 | #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) |
39ad2bbb | 543 | |
b411b363 PR |
544 | /* activity log to on disk bitmap -- prepare bio unless that sector |
545 | * is already covered by previously prepared bios */ | |
546 | static int atodb_prepare_unless_covered(struct drbd_conf *mdev, | |
547 | struct bio **bios, | |
548 | unsigned int enr, | |
549 | struct drbd_atodb_wait *wc) __must_hold(local) | |
550 | { | |
551 | struct bio *bio; | |
552 | struct page *page; | |
39ad2bbb | 553 | sector_t on_disk_sector; |
b411b363 PR |
554 | unsigned int page_offset = PAGE_SIZE; |
555 | int offset; | |
556 | int i = 0; | |
557 | int err = -ENOMEM; | |
558 | ||
39ad2bbb LE |
559 | /* We always write aligned, full 4k blocks, |
560 | * so we can ignore the logical_block_size (for now) */ | |
561 | enr &= ~7U; | |
562 | on_disk_sector = enr + mdev->ldev->md.md_offset | |
563 | + mdev->ldev->md.bm_offset; | |
564 | ||
565 | D_ASSERT(!(on_disk_sector & 7U)); | |
566 | ||
b411b363 PR |
567 | /* Check if that enr is already covered by an already created bio. |
568 | * Caution, bios[] is not NULL terminated, | |
569 | * but only initialized to all NULL. | |
570 | * For completely scattered activity log, | |
571 | * the last invocation iterates over all bios, | |
572 | * and finds the last NULL entry. | |
573 | */ | |
574 | while ((bio = bios[i])) { | |
575 | if (bio->bi_sector == on_disk_sector) | |
576 | return 0; | |
577 | i++; | |
578 | } | |
579 | /* bios[i] == NULL, the next not yet used slot */ | |
580 | ||
581 | /* GFP_KERNEL, we are not in the write-out path */ | |
582 | bio = bio_alloc(GFP_KERNEL, 1); | |
583 | if (bio == NULL) | |
584 | return -ENOMEM; | |
585 | ||
586 | if (i > 0) { | |
587 | const struct bio_vec *prev_bv = bios[i-1]->bi_io_vec; | |
588 | page_offset = prev_bv->bv_offset + prev_bv->bv_len; | |
589 | page = prev_bv->bv_page; | |
590 | } | |
591 | if (page_offset == PAGE_SIZE) { | |
592 | page = alloc_page(__GFP_HIGHMEM); | |
593 | if (page == NULL) | |
594 | goto out_bio_put; | |
595 | page_offset = 0; | |
596 | } else { | |
597 | get_page(page); | |
598 | } | |
599 | ||
600 | offset = S2W(enr); | |
601 | drbd_bm_get_lel(mdev, offset, | |
39ad2bbb | 602 | min_t(size_t, S2W(8), drbd_bm_words(mdev) - offset), |
b411b363 PR |
603 | kmap(page) + page_offset); |
604 | kunmap(page); | |
605 | ||
606 | bio->bi_private = wc; | |
607 | bio->bi_end_io = atodb_endio; | |
608 | bio->bi_bdev = mdev->ldev->md_bdev; | |
609 | bio->bi_sector = on_disk_sector; | |
610 | ||
39ad2bbb | 611 | if (bio_add_page(bio, page, 4096, page_offset) != 4096) |
b411b363 PR |
612 | goto out_put_page; |
613 | ||
614 | atomic_inc(&wc->count); | |
615 | /* we already know that we may do this... | |
616 | * get_ldev_if_state(mdev,D_ATTACHING); | |
617 | * just get the extra reference, so that the local_cnt reflects | |
618 | * the number of pending IO requests DRBD at its backing device. | |
619 | */ | |
620 | atomic_inc(&mdev->local_cnt); | |
621 | ||
622 | bios[i] = bio; | |
623 | ||
624 | return 0; | |
625 | ||
626 | out_put_page: | |
627 | err = -EINVAL; | |
628 | put_page(page); | |
629 | out_bio_put: | |
630 | bio_put(bio); | |
631 | return err; | |
632 | } | |
633 | ||
634 | /** | |
635 | * drbd_al_to_on_disk_bm() - * Writes bitmap parts covered by active AL extents | |
636 | * @mdev: DRBD device. | |
637 | * | |
638 | * Called when we detach (unconfigure) local storage, | |
639 | * or when we go from R_PRIMARY to R_SECONDARY role. | |
640 | */ | |
641 | void drbd_al_to_on_disk_bm(struct drbd_conf *mdev) | |
642 | { | |
643 | int i, nr_elements; | |
644 | unsigned int enr; | |
645 | struct bio **bios; | |
646 | struct drbd_atodb_wait wc; | |
647 | ||
648 | ERR_IF (!get_ldev_if_state(mdev, D_ATTACHING)) | |
649 | return; /* sorry, I don't have any act_log etc... */ | |
650 | ||
651 | wait_event(mdev->al_wait, lc_try_lock(mdev->act_log)); | |
652 | ||
653 | nr_elements = mdev->act_log->nr_elements; | |
654 | ||
655 | /* GFP_KERNEL, we are not in anyone's write-out path */ | |
656 | bios = kzalloc(sizeof(struct bio *) * nr_elements, GFP_KERNEL); | |
657 | if (!bios) | |
658 | goto submit_one_by_one; | |
659 | ||
660 | atomic_set(&wc.count, 0); | |
661 | init_completion(&wc.io_done); | |
662 | wc.mdev = mdev; | |
663 | wc.error = 0; | |
664 | ||
665 | for (i = 0; i < nr_elements; i++) { | |
666 | enr = lc_element_by_index(mdev->act_log, i)->lc_number; | |
667 | if (enr == LC_FREE) | |
668 | continue; | |
669 | /* next statement also does atomic_inc wc.count and local_cnt */ | |
670 | if (atodb_prepare_unless_covered(mdev, bios, | |
671 | enr/AL_EXT_PER_BM_SECT, | |
672 | &wc)) | |
673 | goto free_bios_submit_one_by_one; | |
674 | } | |
675 | ||
676 | /* unnecessary optimization? */ | |
677 | lc_unlock(mdev->act_log); | |
678 | wake_up(&mdev->al_wait); | |
679 | ||
680 | /* all prepared, submit them */ | |
681 | for (i = 0; i < nr_elements; i++) { | |
682 | if (bios[i] == NULL) | |
683 | break; | |
684 | if (FAULT_ACTIVE(mdev, DRBD_FAULT_MD_WR)) { | |
685 | bios[i]->bi_rw = WRITE; | |
686 | bio_endio(bios[i], -EIO); | |
687 | } else { | |
688 | submit_bio(WRITE, bios[i]); | |
689 | } | |
690 | } | |
691 | ||
b411b363 PR |
692 | /* always (try to) flush bitmap to stable storage */ |
693 | drbd_md_flush(mdev); | |
694 | ||
695 | /* In case we did not submit a single IO do not wait for | |
696 | * them to complete. ( Because we would wait forever here. ) | |
697 | * | |
698 | * In case we had IOs and they are already complete, there | |
699 | * is not point in waiting anyways. | |
700 | * Therefore this if () ... */ | |
701 | if (atomic_read(&wc.count)) | |
702 | wait_for_completion(&wc.io_done); | |
703 | ||
704 | put_ldev(mdev); | |
705 | ||
706 | kfree(bios); | |
707 | return; | |
708 | ||
709 | free_bios_submit_one_by_one: | |
710 | /* free everything by calling the endio callback directly. */ | |
711 | for (i = 0; i < nr_elements && bios[i]; i++) | |
712 | bio_endio(bios[i], 0); | |
713 | ||
714 | kfree(bios); | |
715 | ||
716 | submit_one_by_one: | |
717 | dev_warn(DEV, "Using the slow drbd_al_to_on_disk_bm()\n"); | |
718 | ||
719 | for (i = 0; i < mdev->act_log->nr_elements; i++) { | |
720 | enr = lc_element_by_index(mdev->act_log, i)->lc_number; | |
721 | if (enr == LC_FREE) | |
722 | continue; | |
723 | /* Really slow: if we have al-extents 16..19 active, | |
724 | * sector 4 will be written four times! Synchronous! */ | |
725 | drbd_bm_write_sect(mdev, enr/AL_EXT_PER_BM_SECT); | |
726 | } | |
727 | ||
728 | lc_unlock(mdev->act_log); | |
729 | wake_up(&mdev->al_wait); | |
730 | put_ldev(mdev); | |
731 | } | |
732 | ||
733 | /** | |
734 | * drbd_al_apply_to_bm() - Sets the bitmap to diry(1) where covered ba active AL extents | |
735 | * @mdev: DRBD device. | |
736 | */ | |
737 | void drbd_al_apply_to_bm(struct drbd_conf *mdev) | |
738 | { | |
739 | unsigned int enr; | |
740 | unsigned long add = 0; | |
741 | char ppb[10]; | |
6719fb03 | 742 | int i, tmp; |
b411b363 PR |
743 | |
744 | wait_event(mdev->al_wait, lc_try_lock(mdev->act_log)); | |
745 | ||
746 | for (i = 0; i < mdev->act_log->nr_elements; i++) { | |
747 | enr = lc_element_by_index(mdev->act_log, i)->lc_number; | |
748 | if (enr == LC_FREE) | |
749 | continue; | |
6719fb03 LE |
750 | tmp = drbd_bm_ALe_set_all(mdev, enr); |
751 | dynamic_dev_dbg(DEV, "AL: set %d bits in extent %u\n", tmp, enr); | |
752 | add += tmp; | |
b411b363 PR |
753 | } |
754 | ||
755 | lc_unlock(mdev->act_log); | |
756 | wake_up(&mdev->al_wait); | |
757 | ||
758 | dev_info(DEV, "Marked additional %s as out-of-sync based on AL.\n", | |
759 | ppsize(ppb, Bit2KB(add))); | |
760 | } | |
761 | ||
762 | static int _try_lc_del(struct drbd_conf *mdev, struct lc_element *al_ext) | |
763 | { | |
764 | int rv; | |
765 | ||
766 | spin_lock_irq(&mdev->al_lock); | |
767 | rv = (al_ext->refcnt == 0); | |
768 | if (likely(rv)) | |
769 | lc_del(mdev->act_log, al_ext); | |
770 | spin_unlock_irq(&mdev->al_lock); | |
771 | ||
772 | return rv; | |
773 | } | |
774 | ||
775 | /** | |
776 | * drbd_al_shrink() - Removes all active extents form the activity log | |
777 | * @mdev: DRBD device. | |
778 | * | |
779 | * Removes all active extents form the activity log, waiting until | |
780 | * the reference count of each entry dropped to 0 first, of course. | |
781 | * | |
782 | * You need to lock mdev->act_log with lc_try_lock() / lc_unlock() | |
783 | */ | |
784 | void drbd_al_shrink(struct drbd_conf *mdev) | |
785 | { | |
786 | struct lc_element *al_ext; | |
787 | int i; | |
788 | ||
789 | D_ASSERT(test_bit(__LC_DIRTY, &mdev->act_log->flags)); | |
790 | ||
791 | for (i = 0; i < mdev->act_log->nr_elements; i++) { | |
792 | al_ext = lc_element_by_index(mdev->act_log, i); | |
793 | if (al_ext->lc_number == LC_FREE) | |
794 | continue; | |
795 | wait_event(mdev->al_wait, _try_lc_del(mdev, al_ext)); | |
796 | } | |
797 | ||
798 | wake_up(&mdev->al_wait); | |
799 | } | |
800 | ||
801 | static int w_update_odbm(struct drbd_conf *mdev, struct drbd_work *w, int unused) | |
802 | { | |
803 | struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w); | |
804 | ||
805 | if (!get_ldev(mdev)) { | |
806 | if (__ratelimit(&drbd_ratelimit_state)) | |
807 | dev_warn(DEV, "Can not update on disk bitmap, local IO disabled.\n"); | |
808 | kfree(udw); | |
809 | return 1; | |
810 | } | |
811 | ||
812 | drbd_bm_write_sect(mdev, udw->enr); | |
813 | put_ldev(mdev); | |
814 | ||
815 | kfree(udw); | |
816 | ||
817 | if (drbd_bm_total_weight(mdev) <= mdev->rs_failed) { | |
818 | switch (mdev->state.conn) { | |
819 | case C_SYNC_SOURCE: case C_SYNC_TARGET: | |
820 | case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T: | |
821 | drbd_resync_finished(mdev); | |
822 | default: | |
823 | /* nothing to do */ | |
824 | break; | |
825 | } | |
826 | } | |
827 | drbd_bcast_sync_progress(mdev); | |
828 | ||
829 | return 1; | |
830 | } | |
831 | ||
832 | ||
833 | /* ATTENTION. The AL's extents are 4MB each, while the extents in the | |
834 | * resync LRU-cache are 16MB each. | |
835 | * The caller of this function has to hold an get_ldev() reference. | |
836 | * | |
837 | * TODO will be obsoleted once we have a caching lru of the on disk bitmap | |
838 | */ | |
839 | static void drbd_try_clear_on_disk_bm(struct drbd_conf *mdev, sector_t sector, | |
840 | int count, int success) | |
841 | { | |
842 | struct lc_element *e; | |
843 | struct update_odbm_work *udw; | |
844 | ||
845 | unsigned int enr; | |
846 | ||
847 | D_ASSERT(atomic_read(&mdev->local_cnt)); | |
848 | ||
849 | /* I simply assume that a sector/size pair never crosses | |
850 | * a 16 MB extent border. (Currently this is true...) */ | |
851 | enr = BM_SECT_TO_EXT(sector); | |
852 | ||
853 | e = lc_get(mdev->resync, enr); | |
854 | if (e) { | |
855 | struct bm_extent *ext = lc_entry(e, struct bm_extent, lce); | |
856 | if (ext->lce.lc_number == enr) { | |
857 | if (success) | |
858 | ext->rs_left -= count; | |
859 | else | |
860 | ext->rs_failed += count; | |
861 | if (ext->rs_left < ext->rs_failed) { | |
862 | dev_err(DEV, "BAD! sector=%llus enr=%u rs_left=%d " | |
863 | "rs_failed=%d count=%d\n", | |
864 | (unsigned long long)sector, | |
865 | ext->lce.lc_number, ext->rs_left, | |
866 | ext->rs_failed, count); | |
867 | dump_stack(); | |
868 | ||
869 | lc_put(mdev->resync, &ext->lce); | |
870 | drbd_force_state(mdev, NS(conn, C_DISCONNECTING)); | |
871 | return; | |
872 | } | |
873 | } else { | |
874 | /* Normally this element should be in the cache, | |
875 | * since drbd_rs_begin_io() pulled it already in. | |
876 | * | |
877 | * But maybe an application write finished, and we set | |
878 | * something outside the resync lru_cache in sync. | |
879 | */ | |
880 | int rs_left = drbd_bm_e_weight(mdev, enr); | |
881 | if (ext->flags != 0) { | |
882 | dev_warn(DEV, "changing resync lce: %d[%u;%02lx]" | |
883 | " -> %d[%u;00]\n", | |
884 | ext->lce.lc_number, ext->rs_left, | |
885 | ext->flags, enr, rs_left); | |
886 | ext->flags = 0; | |
887 | } | |
888 | if (ext->rs_failed) { | |
889 | dev_warn(DEV, "Kicking resync_lru element enr=%u " | |
890 | "out with rs_failed=%d\n", | |
891 | ext->lce.lc_number, ext->rs_failed); | |
892 | set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags); | |
893 | } | |
894 | ext->rs_left = rs_left; | |
895 | ext->rs_failed = success ? 0 : count; | |
896 | lc_changed(mdev->resync, &ext->lce); | |
897 | } | |
898 | lc_put(mdev->resync, &ext->lce); | |
899 | /* no race, we are within the al_lock! */ | |
900 | ||
901 | if (ext->rs_left == ext->rs_failed) { | |
902 | ext->rs_failed = 0; | |
903 | ||
904 | udw = kmalloc(sizeof(*udw), GFP_ATOMIC); | |
905 | if (udw) { | |
906 | udw->enr = ext->lce.lc_number; | |
907 | udw->w.cb = w_update_odbm; | |
908 | drbd_queue_work_front(&mdev->data.work, &udw->w); | |
909 | } else { | |
910 | dev_warn(DEV, "Could not kmalloc an udw\n"); | |
911 | set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags); | |
912 | } | |
913 | } | |
914 | } else { | |
915 | dev_err(DEV, "lc_get() failed! locked=%d/%d flags=%lu\n", | |
916 | mdev->resync_locked, | |
917 | mdev->resync->nr_elements, | |
918 | mdev->resync->flags); | |
919 | } | |
920 | } | |
921 | ||
922 | /* clear the bit corresponding to the piece of storage in question: | |
923 | * size byte of data starting from sector. Only clear a bits of the affected | |
924 | * one ore more _aligned_ BM_BLOCK_SIZE blocks. | |
925 | * | |
926 | * called by worker on C_SYNC_TARGET and receiver on SyncSource. | |
927 | * | |
928 | */ | |
929 | void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, int size, | |
930 | const char *file, const unsigned int line) | |
931 | { | |
932 | /* Is called from worker and receiver context _only_ */ | |
933 | unsigned long sbnr, ebnr, lbnr; | |
934 | unsigned long count = 0; | |
935 | sector_t esector, nr_sectors; | |
936 | int wake_up = 0; | |
937 | unsigned long flags; | |
938 | ||
939 | if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) { | |
940 | dev_err(DEV, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n", | |
941 | (unsigned long long)sector, size); | |
942 | return; | |
943 | } | |
944 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
945 | esector = sector + (size >> 9) - 1; | |
946 | ||
947 | ERR_IF(sector >= nr_sectors) return; | |
948 | ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1); | |
949 | ||
950 | lbnr = BM_SECT_TO_BIT(nr_sectors-1); | |
951 | ||
952 | /* we clear it (in sync). | |
953 | * round up start sector, round down end sector. we make sure we only | |
954 | * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */ | |
955 | if (unlikely(esector < BM_SECT_PER_BIT-1)) | |
956 | return; | |
957 | if (unlikely(esector == (nr_sectors-1))) | |
958 | ebnr = lbnr; | |
959 | else | |
960 | ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1)); | |
961 | sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1); | |
962 | ||
b411b363 PR |
963 | if (sbnr > ebnr) |
964 | return; | |
965 | ||
966 | /* | |
967 | * ok, (capacity & 7) != 0 sometimes, but who cares... | |
968 | * we count rs_{total,left} in bits, not sectors. | |
969 | */ | |
b411b363 | 970 | count = drbd_bm_clear_bits(mdev, sbnr, ebnr); |
1d7734a0 LE |
971 | if (count && get_ldev(mdev)) { |
972 | unsigned long now = jiffies; | |
973 | unsigned long last = mdev->rs_mark_time[mdev->rs_last_mark]; | |
974 | int next = (mdev->rs_last_mark + 1) % DRBD_SYNC_MARKS; | |
975 | if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) { | |
976 | unsigned long tw = drbd_bm_total_weight(mdev); | |
977 | if (mdev->rs_mark_left[mdev->rs_last_mark] != tw && | |
b411b363 PR |
978 | mdev->state.conn != C_PAUSED_SYNC_T && |
979 | mdev->state.conn != C_PAUSED_SYNC_S) { | |
1d7734a0 LE |
980 | mdev->rs_mark_time[next] = now; |
981 | mdev->rs_mark_left[next] = tw; | |
982 | mdev->rs_last_mark = next; | |
b411b363 PR |
983 | } |
984 | } | |
1d7734a0 LE |
985 | spin_lock_irqsave(&mdev->al_lock, flags); |
986 | drbd_try_clear_on_disk_bm(mdev, sector, count, TRUE); | |
987 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
988 | ||
b411b363 PR |
989 | /* just wake_up unconditional now, various lc_chaged(), |
990 | * lc_put() in drbd_try_clear_on_disk_bm(). */ | |
991 | wake_up = 1; | |
1d7734a0 | 992 | put_ldev(mdev); |
b411b363 | 993 | } |
b411b363 PR |
994 | if (wake_up) |
995 | wake_up(&mdev->al_wait); | |
996 | } | |
997 | ||
998 | /* | |
999 | * this is intended to set one request worth of data out of sync. | |
1000 | * affects at least 1 bit, | |
1001 | * and at most 1+DRBD_MAX_SEGMENT_SIZE/BM_BLOCK_SIZE bits. | |
1002 | * | |
1003 | * called by tl_clear and drbd_send_dblock (==drbd_make_request). | |
1004 | * so this can be _any_ process. | |
1005 | */ | |
1006 | void __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, int size, | |
1007 | const char *file, const unsigned int line) | |
1008 | { | |
1009 | unsigned long sbnr, ebnr, lbnr, flags; | |
1010 | sector_t esector, nr_sectors; | |
1011 | unsigned int enr, count; | |
1012 | struct lc_element *e; | |
1013 | ||
1014 | if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) { | |
1015 | dev_err(DEV, "sector: %llus, size: %d\n", | |
1016 | (unsigned long long)sector, size); | |
1017 | return; | |
1018 | } | |
1019 | ||
1020 | if (!get_ldev(mdev)) | |
1021 | return; /* no disk, no metadata, no bitmap to set bits in */ | |
1022 | ||
1023 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
1024 | esector = sector + (size >> 9) - 1; | |
1025 | ||
1026 | ERR_IF(sector >= nr_sectors) | |
1027 | goto out; | |
1028 | ERR_IF(esector >= nr_sectors) | |
1029 | esector = (nr_sectors-1); | |
1030 | ||
1031 | lbnr = BM_SECT_TO_BIT(nr_sectors-1); | |
1032 | ||
1033 | /* we set it out of sync, | |
1034 | * we do not need to round anything here */ | |
1035 | sbnr = BM_SECT_TO_BIT(sector); | |
1036 | ebnr = BM_SECT_TO_BIT(esector); | |
1037 | ||
b411b363 PR |
1038 | /* ok, (capacity & 7) != 0 sometimes, but who cares... |
1039 | * we count rs_{total,left} in bits, not sectors. */ | |
1040 | spin_lock_irqsave(&mdev->al_lock, flags); | |
1041 | count = drbd_bm_set_bits(mdev, sbnr, ebnr); | |
1042 | ||
1043 | enr = BM_SECT_TO_EXT(sector); | |
1044 | e = lc_find(mdev->resync, enr); | |
1045 | if (e) | |
1046 | lc_entry(e, struct bm_extent, lce)->rs_left += count; | |
1047 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1048 | ||
1049 | out: | |
1050 | put_ldev(mdev); | |
1051 | } | |
1052 | ||
1053 | static | |
1054 | struct bm_extent *_bme_get(struct drbd_conf *mdev, unsigned int enr) | |
1055 | { | |
1056 | struct lc_element *e; | |
1057 | struct bm_extent *bm_ext; | |
1058 | int wakeup = 0; | |
1059 | unsigned long rs_flags; | |
1060 | ||
1061 | spin_lock_irq(&mdev->al_lock); | |
1062 | if (mdev->resync_locked > mdev->resync->nr_elements/2) { | |
1063 | spin_unlock_irq(&mdev->al_lock); | |
1064 | return NULL; | |
1065 | } | |
1066 | e = lc_get(mdev->resync, enr); | |
1067 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1068 | if (bm_ext) { | |
1069 | if (bm_ext->lce.lc_number != enr) { | |
1070 | bm_ext->rs_left = drbd_bm_e_weight(mdev, enr); | |
1071 | bm_ext->rs_failed = 0; | |
1072 | lc_changed(mdev->resync, &bm_ext->lce); | |
1073 | wakeup = 1; | |
1074 | } | |
1075 | if (bm_ext->lce.refcnt == 1) | |
1076 | mdev->resync_locked++; | |
1077 | set_bit(BME_NO_WRITES, &bm_ext->flags); | |
1078 | } | |
1079 | rs_flags = mdev->resync->flags; | |
1080 | spin_unlock_irq(&mdev->al_lock); | |
1081 | if (wakeup) | |
1082 | wake_up(&mdev->al_wait); | |
1083 | ||
1084 | if (!bm_ext) { | |
1085 | if (rs_flags & LC_STARVING) | |
1086 | dev_warn(DEV, "Have to wait for element" | |
1087 | " (resync LRU too small?)\n"); | |
1088 | BUG_ON(rs_flags & LC_DIRTY); | |
1089 | } | |
1090 | ||
1091 | return bm_ext; | |
1092 | } | |
1093 | ||
1094 | static int _is_in_al(struct drbd_conf *mdev, unsigned int enr) | |
1095 | { | |
1096 | struct lc_element *al_ext; | |
1097 | int rv = 0; | |
1098 | ||
1099 | spin_lock_irq(&mdev->al_lock); | |
1100 | if (unlikely(enr == mdev->act_log->new_number)) | |
1101 | rv = 1; | |
1102 | else { | |
1103 | al_ext = lc_find(mdev->act_log, enr); | |
1104 | if (al_ext) { | |
1105 | if (al_ext->refcnt) | |
1106 | rv = 1; | |
1107 | } | |
1108 | } | |
1109 | spin_unlock_irq(&mdev->al_lock); | |
1110 | ||
1111 | /* | |
1112 | if (unlikely(rv)) { | |
1113 | dev_info(DEV, "Delaying sync read until app's write is done\n"); | |
1114 | } | |
1115 | */ | |
1116 | return rv; | |
1117 | } | |
1118 | ||
1119 | /** | |
1120 | * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED | |
1121 | * @mdev: DRBD device. | |
1122 | * @sector: The sector number. | |
1123 | * | |
80a40e43 | 1124 | * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted. |
b411b363 PR |
1125 | */ |
1126 | int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector) | |
1127 | { | |
1128 | unsigned int enr = BM_SECT_TO_EXT(sector); | |
1129 | struct bm_extent *bm_ext; | |
1130 | int i, sig; | |
1131 | ||
b411b363 PR |
1132 | sig = wait_event_interruptible(mdev->al_wait, |
1133 | (bm_ext = _bme_get(mdev, enr))); | |
1134 | if (sig) | |
80a40e43 | 1135 | return -EINTR; |
b411b363 PR |
1136 | |
1137 | if (test_bit(BME_LOCKED, &bm_ext->flags)) | |
80a40e43 | 1138 | return 0; |
b411b363 PR |
1139 | |
1140 | for (i = 0; i < AL_EXT_PER_BM_SECT; i++) { | |
1141 | sig = wait_event_interruptible(mdev->al_wait, | |
1142 | !_is_in_al(mdev, enr * AL_EXT_PER_BM_SECT + i)); | |
1143 | if (sig) { | |
1144 | spin_lock_irq(&mdev->al_lock); | |
1145 | if (lc_put(mdev->resync, &bm_ext->lce) == 0) { | |
1146 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1147 | mdev->resync_locked--; | |
1148 | wake_up(&mdev->al_wait); | |
1149 | } | |
1150 | spin_unlock_irq(&mdev->al_lock); | |
80a40e43 | 1151 | return -EINTR; |
b411b363 PR |
1152 | } |
1153 | } | |
b411b363 | 1154 | set_bit(BME_LOCKED, &bm_ext->flags); |
80a40e43 | 1155 | return 0; |
b411b363 PR |
1156 | } |
1157 | ||
1158 | /** | |
1159 | * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep | |
1160 | * @mdev: DRBD device. | |
1161 | * @sector: The sector number. | |
1162 | * | |
1163 | * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then | |
1164 | * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN | |
1165 | * if there is still application IO going on in this area. | |
1166 | */ | |
1167 | int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector) | |
1168 | { | |
1169 | unsigned int enr = BM_SECT_TO_EXT(sector); | |
1170 | const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT; | |
1171 | struct lc_element *e; | |
1172 | struct bm_extent *bm_ext; | |
1173 | int i; | |
1174 | ||
b411b363 PR |
1175 | spin_lock_irq(&mdev->al_lock); |
1176 | if (mdev->resync_wenr != LC_FREE && mdev->resync_wenr != enr) { | |
1177 | /* in case you have very heavy scattered io, it may | |
1178 | * stall the syncer undefined if we give up the ref count | |
1179 | * when we try again and requeue. | |
1180 | * | |
1181 | * if we don't give up the refcount, but the next time | |
1182 | * we are scheduled this extent has been "synced" by new | |
1183 | * application writes, we'd miss the lc_put on the | |
1184 | * extent we keep the refcount on. | |
1185 | * so we remembered which extent we had to try again, and | |
1186 | * if the next requested one is something else, we do | |
1187 | * the lc_put here... | |
1188 | * we also have to wake_up | |
1189 | */ | |
b411b363 PR |
1190 | e = lc_find(mdev->resync, mdev->resync_wenr); |
1191 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1192 | if (bm_ext) { | |
1193 | D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags)); | |
1194 | D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags)); | |
1195 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1196 | mdev->resync_wenr = LC_FREE; | |
1197 | if (lc_put(mdev->resync, &bm_ext->lce) == 0) | |
1198 | mdev->resync_locked--; | |
1199 | wake_up(&mdev->al_wait); | |
1200 | } else { | |
1201 | dev_alert(DEV, "LOGIC BUG\n"); | |
1202 | } | |
1203 | } | |
1204 | /* TRY. */ | |
1205 | e = lc_try_get(mdev->resync, enr); | |
1206 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1207 | if (bm_ext) { | |
1208 | if (test_bit(BME_LOCKED, &bm_ext->flags)) | |
1209 | goto proceed; | |
1210 | if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) { | |
1211 | mdev->resync_locked++; | |
1212 | } else { | |
1213 | /* we did set the BME_NO_WRITES, | |
1214 | * but then could not set BME_LOCKED, | |
1215 | * so we tried again. | |
1216 | * drop the extra reference. */ | |
b411b363 PR |
1217 | bm_ext->lce.refcnt--; |
1218 | D_ASSERT(bm_ext->lce.refcnt > 0); | |
1219 | } | |
1220 | goto check_al; | |
1221 | } else { | |
1222 | /* do we rather want to try later? */ | |
6a0afdf5 | 1223 | if (mdev->resync_locked > mdev->resync->nr_elements-3) |
b411b363 | 1224 | goto try_again; |
b411b363 PR |
1225 | /* Do or do not. There is no try. -- Yoda */ |
1226 | e = lc_get(mdev->resync, enr); | |
1227 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1228 | if (!bm_ext) { | |
1229 | const unsigned long rs_flags = mdev->resync->flags; | |
1230 | if (rs_flags & LC_STARVING) | |
1231 | dev_warn(DEV, "Have to wait for element" | |
1232 | " (resync LRU too small?)\n"); | |
1233 | BUG_ON(rs_flags & LC_DIRTY); | |
1234 | goto try_again; | |
1235 | } | |
1236 | if (bm_ext->lce.lc_number != enr) { | |
1237 | bm_ext->rs_left = drbd_bm_e_weight(mdev, enr); | |
1238 | bm_ext->rs_failed = 0; | |
1239 | lc_changed(mdev->resync, &bm_ext->lce); | |
1240 | wake_up(&mdev->al_wait); | |
1241 | D_ASSERT(test_bit(BME_LOCKED, &bm_ext->flags) == 0); | |
1242 | } | |
1243 | set_bit(BME_NO_WRITES, &bm_ext->flags); | |
1244 | D_ASSERT(bm_ext->lce.refcnt == 1); | |
1245 | mdev->resync_locked++; | |
1246 | goto check_al; | |
1247 | } | |
1248 | check_al: | |
b411b363 PR |
1249 | for (i = 0; i < AL_EXT_PER_BM_SECT; i++) { |
1250 | if (unlikely(al_enr+i == mdev->act_log->new_number)) | |
1251 | goto try_again; | |
1252 | if (lc_is_used(mdev->act_log, al_enr+i)) | |
1253 | goto try_again; | |
1254 | } | |
1255 | set_bit(BME_LOCKED, &bm_ext->flags); | |
1256 | proceed: | |
1257 | mdev->resync_wenr = LC_FREE; | |
1258 | spin_unlock_irq(&mdev->al_lock); | |
1259 | return 0; | |
1260 | ||
1261 | try_again: | |
b411b363 PR |
1262 | if (bm_ext) |
1263 | mdev->resync_wenr = enr; | |
1264 | spin_unlock_irq(&mdev->al_lock); | |
1265 | return -EAGAIN; | |
1266 | } | |
1267 | ||
1268 | void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector) | |
1269 | { | |
1270 | unsigned int enr = BM_SECT_TO_EXT(sector); | |
1271 | struct lc_element *e; | |
1272 | struct bm_extent *bm_ext; | |
1273 | unsigned long flags; | |
1274 | ||
b411b363 PR |
1275 | spin_lock_irqsave(&mdev->al_lock, flags); |
1276 | e = lc_find(mdev->resync, enr); | |
1277 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1278 | if (!bm_ext) { | |
1279 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1280 | if (__ratelimit(&drbd_ratelimit_state)) | |
1281 | dev_err(DEV, "drbd_rs_complete_io() called, but extent not found\n"); | |
1282 | return; | |
1283 | } | |
1284 | ||
1285 | if (bm_ext->lce.refcnt == 0) { | |
1286 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1287 | dev_err(DEV, "drbd_rs_complete_io(,%llu [=%u]) called, " | |
1288 | "but refcnt is 0!?\n", | |
1289 | (unsigned long long)sector, enr); | |
1290 | return; | |
1291 | } | |
1292 | ||
1293 | if (lc_put(mdev->resync, &bm_ext->lce) == 0) { | |
1294 | clear_bit(BME_LOCKED, &bm_ext->flags); | |
1295 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1296 | mdev->resync_locked--; | |
1297 | wake_up(&mdev->al_wait); | |
1298 | } | |
1299 | ||
1300 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1301 | } | |
1302 | ||
1303 | /** | |
1304 | * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED) | |
1305 | * @mdev: DRBD device. | |
1306 | */ | |
1307 | void drbd_rs_cancel_all(struct drbd_conf *mdev) | |
1308 | { | |
b411b363 PR |
1309 | spin_lock_irq(&mdev->al_lock); |
1310 | ||
1311 | if (get_ldev_if_state(mdev, D_FAILED)) { /* Makes sure ->resync is there. */ | |
1312 | lc_reset(mdev->resync); | |
1313 | put_ldev(mdev); | |
1314 | } | |
1315 | mdev->resync_locked = 0; | |
1316 | mdev->resync_wenr = LC_FREE; | |
1317 | spin_unlock_irq(&mdev->al_lock); | |
1318 | wake_up(&mdev->al_wait); | |
1319 | } | |
1320 | ||
1321 | /** | |
1322 | * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU | |
1323 | * @mdev: DRBD device. | |
1324 | * | |
1325 | * Returns 0 upon success, -EAGAIN if at least one reference count was | |
1326 | * not zero. | |
1327 | */ | |
1328 | int drbd_rs_del_all(struct drbd_conf *mdev) | |
1329 | { | |
1330 | struct lc_element *e; | |
1331 | struct bm_extent *bm_ext; | |
1332 | int i; | |
1333 | ||
b411b363 PR |
1334 | spin_lock_irq(&mdev->al_lock); |
1335 | ||
1336 | if (get_ldev_if_state(mdev, D_FAILED)) { | |
1337 | /* ok, ->resync is there. */ | |
1338 | for (i = 0; i < mdev->resync->nr_elements; i++) { | |
1339 | e = lc_element_by_index(mdev->resync, i); | |
b2b163dd | 1340 | bm_ext = lc_entry(e, struct bm_extent, lce); |
b411b363 PR |
1341 | if (bm_ext->lce.lc_number == LC_FREE) |
1342 | continue; | |
1343 | if (bm_ext->lce.lc_number == mdev->resync_wenr) { | |
1344 | dev_info(DEV, "dropping %u in drbd_rs_del_all, apparently" | |
1345 | " got 'synced' by application io\n", | |
1346 | mdev->resync_wenr); | |
1347 | D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags)); | |
1348 | D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags)); | |
1349 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1350 | mdev->resync_wenr = LC_FREE; | |
1351 | lc_put(mdev->resync, &bm_ext->lce); | |
1352 | } | |
1353 | if (bm_ext->lce.refcnt != 0) { | |
1354 | dev_info(DEV, "Retrying drbd_rs_del_all() later. " | |
1355 | "refcnt=%d\n", bm_ext->lce.refcnt); | |
1356 | put_ldev(mdev); | |
1357 | spin_unlock_irq(&mdev->al_lock); | |
1358 | return -EAGAIN; | |
1359 | } | |
1360 | D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags)); | |
1361 | D_ASSERT(!test_bit(BME_NO_WRITES, &bm_ext->flags)); | |
1362 | lc_del(mdev->resync, &bm_ext->lce); | |
1363 | } | |
1364 | D_ASSERT(mdev->resync->used == 0); | |
1365 | put_ldev(mdev); | |
1366 | } | |
1367 | spin_unlock_irq(&mdev->al_lock); | |
1368 | ||
1369 | return 0; | |
1370 | } | |
1371 | ||
1372 | /** | |
1373 | * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks | |
1374 | * @mdev: DRBD device. | |
1375 | * @sector: The sector number. | |
1376 | * @size: Size of failed IO operation, in byte. | |
1377 | */ | |
1378 | void drbd_rs_failed_io(struct drbd_conf *mdev, sector_t sector, int size) | |
1379 | { | |
1380 | /* Is called from worker and receiver context _only_ */ | |
1381 | unsigned long sbnr, ebnr, lbnr; | |
1382 | unsigned long count; | |
1383 | sector_t esector, nr_sectors; | |
1384 | int wake_up = 0; | |
1385 | ||
b411b363 PR |
1386 | if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) { |
1387 | dev_err(DEV, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n", | |
1388 | (unsigned long long)sector, size); | |
1389 | return; | |
1390 | } | |
1391 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
1392 | esector = sector + (size >> 9) - 1; | |
1393 | ||
1394 | ERR_IF(sector >= nr_sectors) return; | |
1395 | ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1); | |
1396 | ||
1397 | lbnr = BM_SECT_TO_BIT(nr_sectors-1); | |
1398 | ||
1399 | /* | |
1400 | * round up start sector, round down end sector. we make sure we only | |
1401 | * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */ | |
1402 | if (unlikely(esector < BM_SECT_PER_BIT-1)) | |
1403 | return; | |
1404 | if (unlikely(esector == (nr_sectors-1))) | |
1405 | ebnr = lbnr; | |
1406 | else | |
1407 | ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1)); | |
1408 | sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1); | |
1409 | ||
1410 | if (sbnr > ebnr) | |
1411 | return; | |
1412 | ||
1413 | /* | |
1414 | * ok, (capacity & 7) != 0 sometimes, but who cares... | |
1415 | * we count rs_{total,left} in bits, not sectors. | |
1416 | */ | |
1417 | spin_lock_irq(&mdev->al_lock); | |
1418 | count = drbd_bm_count_bits(mdev, sbnr, ebnr); | |
1419 | if (count) { | |
1420 | mdev->rs_failed += count; | |
1421 | ||
1422 | if (get_ldev(mdev)) { | |
1423 | drbd_try_clear_on_disk_bm(mdev, sector, count, FALSE); | |
1424 | put_ldev(mdev); | |
1425 | } | |
1426 | ||
1427 | /* just wake_up unconditional now, various lc_chaged(), | |
1428 | * lc_put() in drbd_try_clear_on_disk_bm(). */ | |
1429 | wake_up = 1; | |
1430 | } | |
1431 | spin_unlock_irq(&mdev->al_lock); | |
1432 | if (wake_up) | |
1433 | wake_up(&mdev->al_wait); | |
1434 | } |