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" | |
29 | #include "drbd_tracing.h" | |
30 | #include "drbd_wrappers.h" | |
31 | ||
32 | /* We maintain a trivial check sum in our on disk activity log. | |
33 | * With that we can ensure correct operation even when the storage | |
34 | * device might do a partial (last) sector write while loosing power. | |
35 | */ | |
36 | struct __packed al_transaction { | |
37 | u32 magic; | |
38 | u32 tr_number; | |
39 | struct __packed { | |
40 | u32 pos; | |
41 | u32 extent; } updates[1 + AL_EXTENTS_PT]; | |
42 | u32 xor_sum; | |
43 | }; | |
44 | ||
45 | struct update_odbm_work { | |
46 | struct drbd_work w; | |
47 | unsigned int enr; | |
48 | }; | |
49 | ||
50 | struct update_al_work { | |
51 | struct drbd_work w; | |
52 | struct lc_element *al_ext; | |
53 | struct completion event; | |
54 | unsigned int enr; | |
55 | /* if old_enr != LC_FREE, write corresponding bitmap sector, too */ | |
56 | unsigned int old_enr; | |
57 | }; | |
58 | ||
59 | struct drbd_atodb_wait { | |
60 | atomic_t count; | |
61 | struct completion io_done; | |
62 | struct drbd_conf *mdev; | |
63 | int error; | |
64 | }; | |
65 | ||
66 | ||
67 | int w_al_write_transaction(struct drbd_conf *, struct drbd_work *, int); | |
68 | ||
69 | /* The actual tracepoint needs to have constant number of known arguments... | |
70 | */ | |
71 | void trace_drbd_resync(struct drbd_conf *mdev, int level, const char *fmt, ...) | |
72 | { | |
73 | va_list ap; | |
74 | ||
75 | va_start(ap, fmt); | |
76 | trace__drbd_resync(mdev, level, fmt, ap); | |
77 | va_end(ap); | |
78 | } | |
79 | ||
80 | static int _drbd_md_sync_page_io(struct drbd_conf *mdev, | |
81 | struct drbd_backing_dev *bdev, | |
82 | struct page *page, sector_t sector, | |
83 | int rw, int size) | |
84 | { | |
85 | struct bio *bio; | |
86 | struct drbd_md_io md_io; | |
87 | int ok; | |
88 | ||
89 | md_io.mdev = mdev; | |
90 | init_completion(&md_io.event); | |
91 | md_io.error = 0; | |
92 | ||
93 | if ((rw & WRITE) && !test_bit(MD_NO_BARRIER, &mdev->flags)) | |
94 | rw |= (1 << BIO_RW_BARRIER); | |
95 | rw |= ((1<<BIO_RW_UNPLUG) | (1<<BIO_RW_SYNCIO)); | |
96 | ||
97 | retry: | |
98 | bio = bio_alloc(GFP_NOIO, 1); | |
99 | bio->bi_bdev = bdev->md_bdev; | |
100 | bio->bi_sector = sector; | |
101 | ok = (bio_add_page(bio, page, size, 0) == size); | |
102 | if (!ok) | |
103 | goto out; | |
104 | bio->bi_private = &md_io; | |
105 | bio->bi_end_io = drbd_md_io_complete; | |
106 | bio->bi_rw = rw; | |
107 | ||
108 | trace_drbd_bio(mdev, "Md", bio, 0, NULL); | |
109 | ||
110 | if (FAULT_ACTIVE(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) | |
111 | bio_endio(bio, -EIO); | |
112 | else | |
113 | submit_bio(rw, bio); | |
114 | wait_for_completion(&md_io.event); | |
115 | ok = bio_flagged(bio, BIO_UPTODATE) && md_io.error == 0; | |
116 | ||
117 | /* check for unsupported barrier op. | |
118 | * would rather check on EOPNOTSUPP, but that is not reliable. | |
119 | * don't try again for ANY return value != 0 */ | |
120 | if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER) && !ok)) { | |
121 | /* Try again with no barrier */ | |
122 | dev_warn(DEV, "Barriers not supported on meta data device - disabling\n"); | |
123 | set_bit(MD_NO_BARRIER, &mdev->flags); | |
124 | rw &= ~(1 << BIO_RW_BARRIER); | |
125 | bio_put(bio); | |
126 | goto retry; | |
127 | } | |
128 | out: | |
129 | bio_put(bio); | |
130 | return ok; | |
131 | } | |
132 | ||
133 | int drbd_md_sync_page_io(struct drbd_conf *mdev, struct drbd_backing_dev *bdev, | |
134 | sector_t sector, int rw) | |
135 | { | |
136 | int logical_block_size, mask, ok; | |
137 | int offset = 0; | |
138 | struct page *iop = mdev->md_io_page; | |
139 | ||
140 | D_ASSERT(mutex_is_locked(&mdev->md_io_mutex)); | |
141 | ||
142 | BUG_ON(!bdev->md_bdev); | |
143 | ||
144 | logical_block_size = bdev_logical_block_size(bdev->md_bdev); | |
145 | if (logical_block_size == 0) | |
146 | logical_block_size = MD_SECTOR_SIZE; | |
147 | ||
148 | /* in case logical_block_size != 512 [ s390 only? ] */ | |
149 | if (logical_block_size != MD_SECTOR_SIZE) { | |
150 | mask = (logical_block_size / MD_SECTOR_SIZE) - 1; | |
151 | D_ASSERT(mask == 1 || mask == 3 || mask == 7); | |
152 | D_ASSERT(logical_block_size == (mask+1) * MD_SECTOR_SIZE); | |
153 | offset = sector & mask; | |
154 | sector = sector & ~mask; | |
155 | iop = mdev->md_io_tmpp; | |
156 | ||
157 | if (rw & WRITE) { | |
158 | /* these are GFP_KERNEL pages, pre-allocated | |
159 | * on device initialization */ | |
160 | void *p = page_address(mdev->md_io_page); | |
161 | void *hp = page_address(mdev->md_io_tmpp); | |
162 | ||
163 | ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, | |
164 | READ, logical_block_size); | |
165 | ||
166 | if (unlikely(!ok)) { | |
167 | dev_err(DEV, "drbd_md_sync_page_io(,%llus," | |
168 | "READ [logical_block_size!=512]) failed!\n", | |
169 | (unsigned long long)sector); | |
170 | return 0; | |
171 | } | |
172 | ||
173 | memcpy(hp + offset*MD_SECTOR_SIZE, p, MD_SECTOR_SIZE); | |
174 | } | |
175 | } | |
176 | ||
177 | if (sector < drbd_md_first_sector(bdev) || | |
178 | sector > drbd_md_last_sector(bdev)) | |
179 | dev_alert(DEV, "%s [%d]:%s(,%llus,%s) out of range md access!\n", | |
180 | current->comm, current->pid, __func__, | |
181 | (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ"); | |
182 | ||
183 | ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, rw, logical_block_size); | |
184 | if (unlikely(!ok)) { | |
185 | dev_err(DEV, "drbd_md_sync_page_io(,%llus,%s) failed!\n", | |
186 | (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ"); | |
187 | return 0; | |
188 | } | |
189 | ||
190 | if (logical_block_size != MD_SECTOR_SIZE && !(rw & WRITE)) { | |
191 | void *p = page_address(mdev->md_io_page); | |
192 | void *hp = page_address(mdev->md_io_tmpp); | |
193 | ||
194 | memcpy(p, hp + offset*MD_SECTOR_SIZE, MD_SECTOR_SIZE); | |
195 | } | |
196 | ||
197 | return ok; | |
198 | } | |
199 | ||
200 | static struct lc_element *_al_get(struct drbd_conf *mdev, unsigned int enr) | |
201 | { | |
202 | struct lc_element *al_ext; | |
203 | struct lc_element *tmp; | |
204 | unsigned long al_flags = 0; | |
205 | ||
206 | spin_lock_irq(&mdev->al_lock); | |
207 | tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT); | |
208 | if (unlikely(tmp != NULL)) { | |
209 | struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce); | |
210 | if (test_bit(BME_NO_WRITES, &bm_ext->flags)) { | |
211 | spin_unlock_irq(&mdev->al_lock); | |
212 | return NULL; | |
213 | } | |
214 | } | |
215 | al_ext = lc_get(mdev->act_log, enr); | |
216 | al_flags = mdev->act_log->flags; | |
217 | spin_unlock_irq(&mdev->al_lock); | |
218 | ||
219 | /* | |
220 | if (!al_ext) { | |
221 | if (al_flags & LC_STARVING) | |
222 | dev_warn(DEV, "Have to wait for LRU element (AL too small?)\n"); | |
223 | if (al_flags & LC_DIRTY) | |
224 | dev_warn(DEV, "Ongoing AL update (AL device too slow?)\n"); | |
225 | } | |
226 | */ | |
227 | ||
228 | return al_ext; | |
229 | } | |
230 | ||
231 | void drbd_al_begin_io(struct drbd_conf *mdev, sector_t sector) | |
232 | { | |
233 | unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9)); | |
234 | struct lc_element *al_ext; | |
235 | struct update_al_work al_work; | |
236 | ||
237 | D_ASSERT(atomic_read(&mdev->local_cnt) > 0); | |
238 | ||
239 | trace_drbd_actlog(mdev, sector, "al_begin_io"); | |
240 | ||
241 | wait_event(mdev->al_wait, (al_ext = _al_get(mdev, enr))); | |
242 | ||
243 | if (al_ext->lc_number != enr) { | |
244 | /* drbd_al_write_transaction(mdev,al_ext,enr); | |
245 | * recurses into generic_make_request(), which | |
246 | * disallows recursion, bios being serialized on the | |
247 | * current->bio_tail list now. | |
248 | * we have to delegate updates to the activity log | |
249 | * to the worker thread. */ | |
250 | init_completion(&al_work.event); | |
251 | al_work.al_ext = al_ext; | |
252 | al_work.enr = enr; | |
253 | al_work.old_enr = al_ext->lc_number; | |
254 | al_work.w.cb = w_al_write_transaction; | |
255 | drbd_queue_work_front(&mdev->data.work, &al_work.w); | |
256 | wait_for_completion(&al_work.event); | |
257 | ||
258 | mdev->al_writ_cnt++; | |
259 | ||
260 | spin_lock_irq(&mdev->al_lock); | |
261 | lc_changed(mdev->act_log, al_ext); | |
262 | spin_unlock_irq(&mdev->al_lock); | |
263 | wake_up(&mdev->al_wait); | |
264 | } | |
265 | } | |
266 | ||
267 | void drbd_al_complete_io(struct drbd_conf *mdev, sector_t sector) | |
268 | { | |
269 | unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9)); | |
270 | struct lc_element *extent; | |
271 | unsigned long flags; | |
272 | ||
273 | trace_drbd_actlog(mdev, sector, "al_complete_io"); | |
274 | ||
275 | spin_lock_irqsave(&mdev->al_lock, flags); | |
276 | ||
277 | extent = lc_find(mdev->act_log, enr); | |
278 | ||
279 | if (!extent) { | |
280 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
281 | dev_err(DEV, "al_complete_io() called on inactive extent %u\n", enr); | |
282 | return; | |
283 | } | |
284 | ||
285 | if (lc_put(mdev->act_log, extent) == 0) | |
286 | wake_up(&mdev->al_wait); | |
287 | ||
288 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
289 | } | |
290 | ||
291 | int | |
292 | w_al_write_transaction(struct drbd_conf *mdev, struct drbd_work *w, int unused) | |
293 | { | |
294 | struct update_al_work *aw = container_of(w, struct update_al_work, w); | |
295 | struct lc_element *updated = aw->al_ext; | |
296 | const unsigned int new_enr = aw->enr; | |
297 | const unsigned int evicted = aw->old_enr; | |
298 | struct al_transaction *buffer; | |
299 | sector_t sector; | |
300 | int i, n, mx; | |
301 | unsigned int extent_nr; | |
302 | u32 xor_sum = 0; | |
303 | ||
304 | if (!get_ldev(mdev)) { | |
305 | dev_err(DEV, "get_ldev() failed in w_al_write_transaction\n"); | |
306 | complete(&((struct update_al_work *)w)->event); | |
307 | return 1; | |
308 | } | |
309 | /* do we have to do a bitmap write, first? | |
310 | * TODO reduce maximum latency: | |
311 | * submit both bios, then wait for both, | |
312 | * instead of doing two synchronous sector writes. */ | |
313 | if (mdev->state.conn < C_CONNECTED && evicted != LC_FREE) | |
314 | drbd_bm_write_sect(mdev, evicted/AL_EXT_PER_BM_SECT); | |
315 | ||
316 | mutex_lock(&mdev->md_io_mutex); /* protects md_io_page, al_tr_cycle, ... */ | |
317 | buffer = (struct al_transaction *)page_address(mdev->md_io_page); | |
318 | ||
319 | buffer->magic = __constant_cpu_to_be32(DRBD_MAGIC); | |
320 | buffer->tr_number = cpu_to_be32(mdev->al_tr_number); | |
321 | ||
322 | n = lc_index_of(mdev->act_log, updated); | |
323 | ||
324 | buffer->updates[0].pos = cpu_to_be32(n); | |
325 | buffer->updates[0].extent = cpu_to_be32(new_enr); | |
326 | ||
327 | xor_sum ^= new_enr; | |
328 | ||
329 | mx = min_t(int, AL_EXTENTS_PT, | |
330 | mdev->act_log->nr_elements - mdev->al_tr_cycle); | |
331 | for (i = 0; i < mx; i++) { | |
332 | unsigned idx = mdev->al_tr_cycle + i; | |
333 | extent_nr = lc_element_by_index(mdev->act_log, idx)->lc_number; | |
334 | buffer->updates[i+1].pos = cpu_to_be32(idx); | |
335 | buffer->updates[i+1].extent = cpu_to_be32(extent_nr); | |
336 | xor_sum ^= extent_nr; | |
337 | } | |
338 | for (; i < AL_EXTENTS_PT; i++) { | |
339 | buffer->updates[i+1].pos = __constant_cpu_to_be32(-1); | |
340 | buffer->updates[i+1].extent = __constant_cpu_to_be32(LC_FREE); | |
341 | xor_sum ^= LC_FREE; | |
342 | } | |
343 | mdev->al_tr_cycle += AL_EXTENTS_PT; | |
344 | if (mdev->al_tr_cycle >= mdev->act_log->nr_elements) | |
345 | mdev->al_tr_cycle = 0; | |
346 | ||
347 | buffer->xor_sum = cpu_to_be32(xor_sum); | |
348 | ||
349 | sector = mdev->ldev->md.md_offset | |
350 | + mdev->ldev->md.al_offset + mdev->al_tr_pos; | |
351 | ||
352 | if (!drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE)) | |
353 | drbd_chk_io_error(mdev, 1, TRUE); | |
354 | ||
355 | if (++mdev->al_tr_pos > | |
356 | div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT)) | |
357 | mdev->al_tr_pos = 0; | |
358 | ||
359 | D_ASSERT(mdev->al_tr_pos < MD_AL_MAX_SIZE); | |
360 | mdev->al_tr_number++; | |
361 | ||
362 | mutex_unlock(&mdev->md_io_mutex); | |
363 | ||
364 | complete(&((struct update_al_work *)w)->event); | |
365 | put_ldev(mdev); | |
366 | ||
367 | return 1; | |
368 | } | |
369 | ||
370 | /** | |
371 | * drbd_al_read_tr() - Read a single transaction from the on disk activity log | |
372 | * @mdev: DRBD device. | |
373 | * @bdev: Block device to read form. | |
374 | * @b: pointer to an al_transaction. | |
375 | * @index: On disk slot of the transaction to read. | |
376 | * | |
377 | * Returns -1 on IO error, 0 on checksum error and 1 upon success. | |
378 | */ | |
379 | static int drbd_al_read_tr(struct drbd_conf *mdev, | |
380 | struct drbd_backing_dev *bdev, | |
381 | struct al_transaction *b, | |
382 | int index) | |
383 | { | |
384 | sector_t sector; | |
385 | int rv, i; | |
386 | u32 xor_sum = 0; | |
387 | ||
388 | sector = bdev->md.md_offset + bdev->md.al_offset + index; | |
389 | ||
390 | /* Dont process error normally, | |
391 | * as this is done before disk is attached! */ | |
392 | if (!drbd_md_sync_page_io(mdev, bdev, sector, READ)) | |
393 | return -1; | |
394 | ||
395 | rv = (be32_to_cpu(b->magic) == DRBD_MAGIC); | |
396 | ||
397 | for (i = 0; i < AL_EXTENTS_PT + 1; i++) | |
398 | xor_sum ^= be32_to_cpu(b->updates[i].extent); | |
399 | rv &= (xor_sum == be32_to_cpu(b->xor_sum)); | |
400 | ||
401 | return rv; | |
402 | } | |
403 | ||
404 | /** | |
405 | * drbd_al_read_log() - Restores the activity log from its on disk representation. | |
406 | * @mdev: DRBD device. | |
407 | * @bdev: Block device to read form. | |
408 | * | |
409 | * Returns 1 on success, returns 0 when reading the log failed due to IO errors. | |
410 | */ | |
411 | int drbd_al_read_log(struct drbd_conf *mdev, struct drbd_backing_dev *bdev) | |
412 | { | |
413 | struct al_transaction *buffer; | |
414 | int i; | |
415 | int rv; | |
416 | int mx; | |
417 | int active_extents = 0; | |
418 | int transactions = 0; | |
419 | int found_valid = 0; | |
420 | int from = 0; | |
421 | int to = 0; | |
422 | u32 from_tnr = 0; | |
423 | u32 to_tnr = 0; | |
424 | u32 cnr; | |
425 | ||
426 | mx = div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT); | |
427 | ||
428 | /* lock out all other meta data io for now, | |
429 | * and make sure the page is mapped. | |
430 | */ | |
431 | mutex_lock(&mdev->md_io_mutex); | |
432 | buffer = page_address(mdev->md_io_page); | |
433 | ||
434 | /* Find the valid transaction in the log */ | |
435 | for (i = 0; i <= mx; i++) { | |
436 | rv = drbd_al_read_tr(mdev, bdev, buffer, i); | |
437 | if (rv == 0) | |
438 | continue; | |
439 | if (rv == -1) { | |
440 | mutex_unlock(&mdev->md_io_mutex); | |
441 | return 0; | |
442 | } | |
443 | cnr = be32_to_cpu(buffer->tr_number); | |
444 | ||
445 | if (++found_valid == 1) { | |
446 | from = i; | |
447 | to = i; | |
448 | from_tnr = cnr; | |
449 | to_tnr = cnr; | |
450 | continue; | |
451 | } | |
452 | if ((int)cnr - (int)from_tnr < 0) { | |
453 | D_ASSERT(from_tnr - cnr + i - from == mx+1); | |
454 | from = i; | |
455 | from_tnr = cnr; | |
456 | } | |
457 | if ((int)cnr - (int)to_tnr > 0) { | |
458 | D_ASSERT(cnr - to_tnr == i - to); | |
459 | to = i; | |
460 | to_tnr = cnr; | |
461 | } | |
462 | } | |
463 | ||
464 | if (!found_valid) { | |
465 | dev_warn(DEV, "No usable activity log found.\n"); | |
466 | mutex_unlock(&mdev->md_io_mutex); | |
467 | return 1; | |
468 | } | |
469 | ||
470 | /* Read the valid transactions. | |
471 | * dev_info(DEV, "Reading from %d to %d.\n",from,to); */ | |
472 | i = from; | |
473 | while (1) { | |
474 | int j, pos; | |
475 | unsigned int extent_nr; | |
476 | unsigned int trn; | |
477 | ||
478 | rv = drbd_al_read_tr(mdev, bdev, buffer, i); | |
479 | ERR_IF(rv == 0) goto cancel; | |
480 | if (rv == -1) { | |
481 | mutex_unlock(&mdev->md_io_mutex); | |
482 | return 0; | |
483 | } | |
484 | ||
485 | trn = be32_to_cpu(buffer->tr_number); | |
486 | ||
487 | spin_lock_irq(&mdev->al_lock); | |
488 | ||
489 | /* This loop runs backwards because in the cyclic | |
490 | elements there might be an old version of the | |
491 | updated element (in slot 0). So the element in slot 0 | |
492 | can overwrite old versions. */ | |
493 | for (j = AL_EXTENTS_PT; j >= 0; j--) { | |
494 | pos = be32_to_cpu(buffer->updates[j].pos); | |
495 | extent_nr = be32_to_cpu(buffer->updates[j].extent); | |
496 | ||
497 | if (extent_nr == LC_FREE) | |
498 | continue; | |
499 | ||
500 | lc_set(mdev->act_log, extent_nr, pos); | |
501 | active_extents++; | |
502 | } | |
503 | spin_unlock_irq(&mdev->al_lock); | |
504 | ||
505 | transactions++; | |
506 | ||
507 | cancel: | |
508 | if (i == to) | |
509 | break; | |
510 | i++; | |
511 | if (i > mx) | |
512 | i = 0; | |
513 | } | |
514 | ||
515 | mdev->al_tr_number = to_tnr+1; | |
516 | mdev->al_tr_pos = to; | |
517 | if (++mdev->al_tr_pos > | |
518 | div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT)) | |
519 | mdev->al_tr_pos = 0; | |
520 | ||
521 | /* ok, we are done with it */ | |
522 | mutex_unlock(&mdev->md_io_mutex); | |
523 | ||
524 | dev_info(DEV, "Found %d transactions (%d active extents) in activity log.\n", | |
525 | transactions, active_extents); | |
526 | ||
527 | return 1; | |
528 | } | |
529 | ||
530 | static void atodb_endio(struct bio *bio, int error) | |
531 | { | |
532 | struct drbd_atodb_wait *wc = bio->bi_private; | |
533 | struct drbd_conf *mdev = wc->mdev; | |
534 | struct page *page; | |
535 | int uptodate = bio_flagged(bio, BIO_UPTODATE); | |
536 | ||
537 | /* strange behavior of some lower level drivers... | |
538 | * fail the request by clearing the uptodate flag, | |
539 | * but do not return any error?! */ | |
540 | if (!error && !uptodate) | |
541 | error = -EIO; | |
542 | ||
543 | drbd_chk_io_error(mdev, error, TRUE); | |
544 | if (error && wc->error == 0) | |
545 | wc->error = error; | |
546 | ||
547 | if (atomic_dec_and_test(&wc->count)) | |
548 | complete(&wc->io_done); | |
549 | ||
550 | page = bio->bi_io_vec[0].bv_page; | |
551 | put_page(page); | |
552 | bio_put(bio); | |
553 | mdev->bm_writ_cnt++; | |
554 | put_ldev(mdev); | |
555 | } | |
556 | ||
557 | #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) | |
558 | /* activity log to on disk bitmap -- prepare bio unless that sector | |
559 | * is already covered by previously prepared bios */ | |
560 | static int atodb_prepare_unless_covered(struct drbd_conf *mdev, | |
561 | struct bio **bios, | |
562 | unsigned int enr, | |
563 | struct drbd_atodb_wait *wc) __must_hold(local) | |
564 | { | |
565 | struct bio *bio; | |
566 | struct page *page; | |
567 | sector_t on_disk_sector = enr + mdev->ldev->md.md_offset | |
568 | + mdev->ldev->md.bm_offset; | |
569 | unsigned int page_offset = PAGE_SIZE; | |
570 | int offset; | |
571 | int i = 0; | |
572 | int err = -ENOMEM; | |
573 | ||
574 | /* Check if that enr is already covered by an already created bio. | |
575 | * Caution, bios[] is not NULL terminated, | |
576 | * but only initialized to all NULL. | |
577 | * For completely scattered activity log, | |
578 | * the last invocation iterates over all bios, | |
579 | * and finds the last NULL entry. | |
580 | */ | |
581 | while ((bio = bios[i])) { | |
582 | if (bio->bi_sector == on_disk_sector) | |
583 | return 0; | |
584 | i++; | |
585 | } | |
586 | /* bios[i] == NULL, the next not yet used slot */ | |
587 | ||
588 | /* GFP_KERNEL, we are not in the write-out path */ | |
589 | bio = bio_alloc(GFP_KERNEL, 1); | |
590 | if (bio == NULL) | |
591 | return -ENOMEM; | |
592 | ||
593 | if (i > 0) { | |
594 | const struct bio_vec *prev_bv = bios[i-1]->bi_io_vec; | |
595 | page_offset = prev_bv->bv_offset + prev_bv->bv_len; | |
596 | page = prev_bv->bv_page; | |
597 | } | |
598 | if (page_offset == PAGE_SIZE) { | |
599 | page = alloc_page(__GFP_HIGHMEM); | |
600 | if (page == NULL) | |
601 | goto out_bio_put; | |
602 | page_offset = 0; | |
603 | } else { | |
604 | get_page(page); | |
605 | } | |
606 | ||
607 | offset = S2W(enr); | |
608 | drbd_bm_get_lel(mdev, offset, | |
609 | min_t(size_t, S2W(1), drbd_bm_words(mdev) - offset), | |
610 | kmap(page) + page_offset); | |
611 | kunmap(page); | |
612 | ||
613 | bio->bi_private = wc; | |
614 | bio->bi_end_io = atodb_endio; | |
615 | bio->bi_bdev = mdev->ldev->md_bdev; | |
616 | bio->bi_sector = on_disk_sector; | |
617 | ||
618 | if (bio_add_page(bio, page, MD_SECTOR_SIZE, page_offset) != MD_SECTOR_SIZE) | |
619 | goto out_put_page; | |
620 | ||
621 | atomic_inc(&wc->count); | |
622 | /* we already know that we may do this... | |
623 | * get_ldev_if_state(mdev,D_ATTACHING); | |
624 | * just get the extra reference, so that the local_cnt reflects | |
625 | * the number of pending IO requests DRBD at its backing device. | |
626 | */ | |
627 | atomic_inc(&mdev->local_cnt); | |
628 | ||
629 | bios[i] = bio; | |
630 | ||
631 | return 0; | |
632 | ||
633 | out_put_page: | |
634 | err = -EINVAL; | |
635 | put_page(page); | |
636 | out_bio_put: | |
637 | bio_put(bio); | |
638 | return err; | |
639 | } | |
640 | ||
641 | /** | |
642 | * drbd_al_to_on_disk_bm() - * Writes bitmap parts covered by active AL extents | |
643 | * @mdev: DRBD device. | |
644 | * | |
645 | * Called when we detach (unconfigure) local storage, | |
646 | * or when we go from R_PRIMARY to R_SECONDARY role. | |
647 | */ | |
648 | void drbd_al_to_on_disk_bm(struct drbd_conf *mdev) | |
649 | { | |
650 | int i, nr_elements; | |
651 | unsigned int enr; | |
652 | struct bio **bios; | |
653 | struct drbd_atodb_wait wc; | |
654 | ||
655 | ERR_IF (!get_ldev_if_state(mdev, D_ATTACHING)) | |
656 | return; /* sorry, I don't have any act_log etc... */ | |
657 | ||
658 | wait_event(mdev->al_wait, lc_try_lock(mdev->act_log)); | |
659 | ||
660 | nr_elements = mdev->act_log->nr_elements; | |
661 | ||
662 | /* GFP_KERNEL, we are not in anyone's write-out path */ | |
663 | bios = kzalloc(sizeof(struct bio *) * nr_elements, GFP_KERNEL); | |
664 | if (!bios) | |
665 | goto submit_one_by_one; | |
666 | ||
667 | atomic_set(&wc.count, 0); | |
668 | init_completion(&wc.io_done); | |
669 | wc.mdev = mdev; | |
670 | wc.error = 0; | |
671 | ||
672 | for (i = 0; i < nr_elements; i++) { | |
673 | enr = lc_element_by_index(mdev->act_log, i)->lc_number; | |
674 | if (enr == LC_FREE) | |
675 | continue; | |
676 | /* next statement also does atomic_inc wc.count and local_cnt */ | |
677 | if (atodb_prepare_unless_covered(mdev, bios, | |
678 | enr/AL_EXT_PER_BM_SECT, | |
679 | &wc)) | |
680 | goto free_bios_submit_one_by_one; | |
681 | } | |
682 | ||
683 | /* unnecessary optimization? */ | |
684 | lc_unlock(mdev->act_log); | |
685 | wake_up(&mdev->al_wait); | |
686 | ||
687 | /* all prepared, submit them */ | |
688 | for (i = 0; i < nr_elements; i++) { | |
689 | if (bios[i] == NULL) | |
690 | break; | |
691 | if (FAULT_ACTIVE(mdev, DRBD_FAULT_MD_WR)) { | |
692 | bios[i]->bi_rw = WRITE; | |
693 | bio_endio(bios[i], -EIO); | |
694 | } else { | |
695 | submit_bio(WRITE, bios[i]); | |
696 | } | |
697 | } | |
698 | ||
699 | drbd_blk_run_queue(bdev_get_queue(mdev->ldev->md_bdev)); | |
700 | ||
701 | /* always (try to) flush bitmap to stable storage */ | |
702 | drbd_md_flush(mdev); | |
703 | ||
704 | /* In case we did not submit a single IO do not wait for | |
705 | * them to complete. ( Because we would wait forever here. ) | |
706 | * | |
707 | * In case we had IOs and they are already complete, there | |
708 | * is not point in waiting anyways. | |
709 | * Therefore this if () ... */ | |
710 | if (atomic_read(&wc.count)) | |
711 | wait_for_completion(&wc.io_done); | |
712 | ||
713 | put_ldev(mdev); | |
714 | ||
715 | kfree(bios); | |
716 | return; | |
717 | ||
718 | free_bios_submit_one_by_one: | |
719 | /* free everything by calling the endio callback directly. */ | |
720 | for (i = 0; i < nr_elements && bios[i]; i++) | |
721 | bio_endio(bios[i], 0); | |
722 | ||
723 | kfree(bios); | |
724 | ||
725 | submit_one_by_one: | |
726 | dev_warn(DEV, "Using the slow drbd_al_to_on_disk_bm()\n"); | |
727 | ||
728 | for (i = 0; i < mdev->act_log->nr_elements; i++) { | |
729 | enr = lc_element_by_index(mdev->act_log, i)->lc_number; | |
730 | if (enr == LC_FREE) | |
731 | continue; | |
732 | /* Really slow: if we have al-extents 16..19 active, | |
733 | * sector 4 will be written four times! Synchronous! */ | |
734 | drbd_bm_write_sect(mdev, enr/AL_EXT_PER_BM_SECT); | |
735 | } | |
736 | ||
737 | lc_unlock(mdev->act_log); | |
738 | wake_up(&mdev->al_wait); | |
739 | put_ldev(mdev); | |
740 | } | |
741 | ||
742 | /** | |
743 | * drbd_al_apply_to_bm() - Sets the bitmap to diry(1) where covered ba active AL extents | |
744 | * @mdev: DRBD device. | |
745 | */ | |
746 | void drbd_al_apply_to_bm(struct drbd_conf *mdev) | |
747 | { | |
748 | unsigned int enr; | |
749 | unsigned long add = 0; | |
750 | char ppb[10]; | |
751 | int i; | |
752 | ||
753 | wait_event(mdev->al_wait, lc_try_lock(mdev->act_log)); | |
754 | ||
755 | for (i = 0; i < mdev->act_log->nr_elements; i++) { | |
756 | enr = lc_element_by_index(mdev->act_log, i)->lc_number; | |
757 | if (enr == LC_FREE) | |
758 | continue; | |
759 | add += drbd_bm_ALe_set_all(mdev, enr); | |
760 | } | |
761 | ||
762 | lc_unlock(mdev->act_log); | |
763 | wake_up(&mdev->al_wait); | |
764 | ||
765 | dev_info(DEV, "Marked additional %s as out-of-sync based on AL.\n", | |
766 | ppsize(ppb, Bit2KB(add))); | |
767 | } | |
768 | ||
769 | static int _try_lc_del(struct drbd_conf *mdev, struct lc_element *al_ext) | |
770 | { | |
771 | int rv; | |
772 | ||
773 | spin_lock_irq(&mdev->al_lock); | |
774 | rv = (al_ext->refcnt == 0); | |
775 | if (likely(rv)) | |
776 | lc_del(mdev->act_log, al_ext); | |
777 | spin_unlock_irq(&mdev->al_lock); | |
778 | ||
779 | return rv; | |
780 | } | |
781 | ||
782 | /** | |
783 | * drbd_al_shrink() - Removes all active extents form the activity log | |
784 | * @mdev: DRBD device. | |
785 | * | |
786 | * Removes all active extents form the activity log, waiting until | |
787 | * the reference count of each entry dropped to 0 first, of course. | |
788 | * | |
789 | * You need to lock mdev->act_log with lc_try_lock() / lc_unlock() | |
790 | */ | |
791 | void drbd_al_shrink(struct drbd_conf *mdev) | |
792 | { | |
793 | struct lc_element *al_ext; | |
794 | int i; | |
795 | ||
796 | D_ASSERT(test_bit(__LC_DIRTY, &mdev->act_log->flags)); | |
797 | ||
798 | for (i = 0; i < mdev->act_log->nr_elements; i++) { | |
799 | al_ext = lc_element_by_index(mdev->act_log, i); | |
800 | if (al_ext->lc_number == LC_FREE) | |
801 | continue; | |
802 | wait_event(mdev->al_wait, _try_lc_del(mdev, al_ext)); | |
803 | } | |
804 | ||
805 | wake_up(&mdev->al_wait); | |
806 | } | |
807 | ||
808 | static int w_update_odbm(struct drbd_conf *mdev, struct drbd_work *w, int unused) | |
809 | { | |
810 | struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w); | |
811 | ||
812 | if (!get_ldev(mdev)) { | |
813 | if (__ratelimit(&drbd_ratelimit_state)) | |
814 | dev_warn(DEV, "Can not update on disk bitmap, local IO disabled.\n"); | |
815 | kfree(udw); | |
816 | return 1; | |
817 | } | |
818 | ||
819 | drbd_bm_write_sect(mdev, udw->enr); | |
820 | put_ldev(mdev); | |
821 | ||
822 | kfree(udw); | |
823 | ||
824 | if (drbd_bm_total_weight(mdev) <= mdev->rs_failed) { | |
825 | switch (mdev->state.conn) { | |
826 | case C_SYNC_SOURCE: case C_SYNC_TARGET: | |
827 | case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T: | |
828 | drbd_resync_finished(mdev); | |
829 | default: | |
830 | /* nothing to do */ | |
831 | break; | |
832 | } | |
833 | } | |
834 | drbd_bcast_sync_progress(mdev); | |
835 | ||
836 | return 1; | |
837 | } | |
838 | ||
839 | ||
840 | /* ATTENTION. The AL's extents are 4MB each, while the extents in the | |
841 | * resync LRU-cache are 16MB each. | |
842 | * The caller of this function has to hold an get_ldev() reference. | |
843 | * | |
844 | * TODO will be obsoleted once we have a caching lru of the on disk bitmap | |
845 | */ | |
846 | static void drbd_try_clear_on_disk_bm(struct drbd_conf *mdev, sector_t sector, | |
847 | int count, int success) | |
848 | { | |
849 | struct lc_element *e; | |
850 | struct update_odbm_work *udw; | |
851 | ||
852 | unsigned int enr; | |
853 | ||
854 | D_ASSERT(atomic_read(&mdev->local_cnt)); | |
855 | ||
856 | /* I simply assume that a sector/size pair never crosses | |
857 | * a 16 MB extent border. (Currently this is true...) */ | |
858 | enr = BM_SECT_TO_EXT(sector); | |
859 | ||
860 | e = lc_get(mdev->resync, enr); | |
861 | if (e) { | |
862 | struct bm_extent *ext = lc_entry(e, struct bm_extent, lce); | |
863 | if (ext->lce.lc_number == enr) { | |
864 | if (success) | |
865 | ext->rs_left -= count; | |
866 | else | |
867 | ext->rs_failed += count; | |
868 | if (ext->rs_left < ext->rs_failed) { | |
869 | dev_err(DEV, "BAD! sector=%llus enr=%u rs_left=%d " | |
870 | "rs_failed=%d count=%d\n", | |
871 | (unsigned long long)sector, | |
872 | ext->lce.lc_number, ext->rs_left, | |
873 | ext->rs_failed, count); | |
874 | dump_stack(); | |
875 | ||
876 | lc_put(mdev->resync, &ext->lce); | |
877 | drbd_force_state(mdev, NS(conn, C_DISCONNECTING)); | |
878 | return; | |
879 | } | |
880 | } else { | |
881 | /* Normally this element should be in the cache, | |
882 | * since drbd_rs_begin_io() pulled it already in. | |
883 | * | |
884 | * But maybe an application write finished, and we set | |
885 | * something outside the resync lru_cache in sync. | |
886 | */ | |
887 | int rs_left = drbd_bm_e_weight(mdev, enr); | |
888 | if (ext->flags != 0) { | |
889 | dev_warn(DEV, "changing resync lce: %d[%u;%02lx]" | |
890 | " -> %d[%u;00]\n", | |
891 | ext->lce.lc_number, ext->rs_left, | |
892 | ext->flags, enr, rs_left); | |
893 | ext->flags = 0; | |
894 | } | |
895 | if (ext->rs_failed) { | |
896 | dev_warn(DEV, "Kicking resync_lru element enr=%u " | |
897 | "out with rs_failed=%d\n", | |
898 | ext->lce.lc_number, ext->rs_failed); | |
899 | set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags); | |
900 | } | |
901 | ext->rs_left = rs_left; | |
902 | ext->rs_failed = success ? 0 : count; | |
903 | lc_changed(mdev->resync, &ext->lce); | |
904 | } | |
905 | lc_put(mdev->resync, &ext->lce); | |
906 | /* no race, we are within the al_lock! */ | |
907 | ||
908 | if (ext->rs_left == ext->rs_failed) { | |
909 | ext->rs_failed = 0; | |
910 | ||
911 | udw = kmalloc(sizeof(*udw), GFP_ATOMIC); | |
912 | if (udw) { | |
913 | udw->enr = ext->lce.lc_number; | |
914 | udw->w.cb = w_update_odbm; | |
915 | drbd_queue_work_front(&mdev->data.work, &udw->w); | |
916 | } else { | |
917 | dev_warn(DEV, "Could not kmalloc an udw\n"); | |
918 | set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags); | |
919 | } | |
920 | } | |
921 | } else { | |
922 | dev_err(DEV, "lc_get() failed! locked=%d/%d flags=%lu\n", | |
923 | mdev->resync_locked, | |
924 | mdev->resync->nr_elements, | |
925 | mdev->resync->flags); | |
926 | } | |
927 | } | |
928 | ||
929 | /* clear the bit corresponding to the piece of storage in question: | |
930 | * size byte of data starting from sector. Only clear a bits of the affected | |
931 | * one ore more _aligned_ BM_BLOCK_SIZE blocks. | |
932 | * | |
933 | * called by worker on C_SYNC_TARGET and receiver on SyncSource. | |
934 | * | |
935 | */ | |
936 | void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, int size, | |
937 | const char *file, const unsigned int line) | |
938 | { | |
939 | /* Is called from worker and receiver context _only_ */ | |
940 | unsigned long sbnr, ebnr, lbnr; | |
941 | unsigned long count = 0; | |
942 | sector_t esector, nr_sectors; | |
943 | int wake_up = 0; | |
944 | unsigned long flags; | |
945 | ||
946 | if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) { | |
947 | dev_err(DEV, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n", | |
948 | (unsigned long long)sector, size); | |
949 | return; | |
950 | } | |
951 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
952 | esector = sector + (size >> 9) - 1; | |
953 | ||
954 | ERR_IF(sector >= nr_sectors) return; | |
955 | ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1); | |
956 | ||
957 | lbnr = BM_SECT_TO_BIT(nr_sectors-1); | |
958 | ||
959 | /* we clear it (in sync). | |
960 | * round up start sector, round down end sector. we make sure we only | |
961 | * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */ | |
962 | if (unlikely(esector < BM_SECT_PER_BIT-1)) | |
963 | return; | |
964 | if (unlikely(esector == (nr_sectors-1))) | |
965 | ebnr = lbnr; | |
966 | else | |
967 | ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1)); | |
968 | sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1); | |
969 | ||
970 | trace_drbd_resync(mdev, TRACE_LVL_METRICS, | |
971 | "drbd_set_in_sync: sector=%llus size=%u sbnr=%lu ebnr=%lu\n", | |
972 | (unsigned long long)sector, size, sbnr, ebnr); | |
973 | ||
974 | if (sbnr > ebnr) | |
975 | return; | |
976 | ||
977 | /* | |
978 | * ok, (capacity & 7) != 0 sometimes, but who cares... | |
979 | * we count rs_{total,left} in bits, not sectors. | |
980 | */ | |
981 | spin_lock_irqsave(&mdev->al_lock, flags); | |
982 | count = drbd_bm_clear_bits(mdev, sbnr, ebnr); | |
983 | if (count) { | |
984 | /* we need the lock for drbd_try_clear_on_disk_bm */ | |
985 | if (jiffies - mdev->rs_mark_time > HZ*10) { | |
986 | /* should be rolling marks, | |
987 | * but we estimate only anyways. */ | |
988 | if (mdev->rs_mark_left != drbd_bm_total_weight(mdev) && | |
989 | mdev->state.conn != C_PAUSED_SYNC_T && | |
990 | mdev->state.conn != C_PAUSED_SYNC_S) { | |
991 | mdev->rs_mark_time = jiffies; | |
992 | mdev->rs_mark_left = drbd_bm_total_weight(mdev); | |
993 | } | |
994 | } | |
995 | if (get_ldev(mdev)) { | |
996 | drbd_try_clear_on_disk_bm(mdev, sector, count, TRUE); | |
997 | put_ldev(mdev); | |
998 | } | |
999 | /* just wake_up unconditional now, various lc_chaged(), | |
1000 | * lc_put() in drbd_try_clear_on_disk_bm(). */ | |
1001 | wake_up = 1; | |
1002 | } | |
1003 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1004 | if (wake_up) | |
1005 | wake_up(&mdev->al_wait); | |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * this is intended to set one request worth of data out of sync. | |
1010 | * affects at least 1 bit, | |
1011 | * and at most 1+DRBD_MAX_SEGMENT_SIZE/BM_BLOCK_SIZE bits. | |
1012 | * | |
1013 | * called by tl_clear and drbd_send_dblock (==drbd_make_request). | |
1014 | * so this can be _any_ process. | |
1015 | */ | |
1016 | void __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, int size, | |
1017 | const char *file, const unsigned int line) | |
1018 | { | |
1019 | unsigned long sbnr, ebnr, lbnr, flags; | |
1020 | sector_t esector, nr_sectors; | |
1021 | unsigned int enr, count; | |
1022 | struct lc_element *e; | |
1023 | ||
1024 | if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) { | |
1025 | dev_err(DEV, "sector: %llus, size: %d\n", | |
1026 | (unsigned long long)sector, size); | |
1027 | return; | |
1028 | } | |
1029 | ||
1030 | if (!get_ldev(mdev)) | |
1031 | return; /* no disk, no metadata, no bitmap to set bits in */ | |
1032 | ||
1033 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
1034 | esector = sector + (size >> 9) - 1; | |
1035 | ||
1036 | ERR_IF(sector >= nr_sectors) | |
1037 | goto out; | |
1038 | ERR_IF(esector >= nr_sectors) | |
1039 | esector = (nr_sectors-1); | |
1040 | ||
1041 | lbnr = BM_SECT_TO_BIT(nr_sectors-1); | |
1042 | ||
1043 | /* we set it out of sync, | |
1044 | * we do not need to round anything here */ | |
1045 | sbnr = BM_SECT_TO_BIT(sector); | |
1046 | ebnr = BM_SECT_TO_BIT(esector); | |
1047 | ||
1048 | trace_drbd_resync(mdev, TRACE_LVL_METRICS, | |
1049 | "drbd_set_out_of_sync: sector=%llus size=%u sbnr=%lu ebnr=%lu\n", | |
1050 | (unsigned long long)sector, size, sbnr, ebnr); | |
1051 | ||
1052 | /* ok, (capacity & 7) != 0 sometimes, but who cares... | |
1053 | * we count rs_{total,left} in bits, not sectors. */ | |
1054 | spin_lock_irqsave(&mdev->al_lock, flags); | |
1055 | count = drbd_bm_set_bits(mdev, sbnr, ebnr); | |
1056 | ||
1057 | enr = BM_SECT_TO_EXT(sector); | |
1058 | e = lc_find(mdev->resync, enr); | |
1059 | if (e) | |
1060 | lc_entry(e, struct bm_extent, lce)->rs_left += count; | |
1061 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1062 | ||
1063 | out: | |
1064 | put_ldev(mdev); | |
1065 | } | |
1066 | ||
1067 | static | |
1068 | struct bm_extent *_bme_get(struct drbd_conf *mdev, unsigned int enr) | |
1069 | { | |
1070 | struct lc_element *e; | |
1071 | struct bm_extent *bm_ext; | |
1072 | int wakeup = 0; | |
1073 | unsigned long rs_flags; | |
1074 | ||
1075 | spin_lock_irq(&mdev->al_lock); | |
1076 | if (mdev->resync_locked > mdev->resync->nr_elements/2) { | |
1077 | spin_unlock_irq(&mdev->al_lock); | |
1078 | return NULL; | |
1079 | } | |
1080 | e = lc_get(mdev->resync, enr); | |
1081 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1082 | if (bm_ext) { | |
1083 | if (bm_ext->lce.lc_number != enr) { | |
1084 | bm_ext->rs_left = drbd_bm_e_weight(mdev, enr); | |
1085 | bm_ext->rs_failed = 0; | |
1086 | lc_changed(mdev->resync, &bm_ext->lce); | |
1087 | wakeup = 1; | |
1088 | } | |
1089 | if (bm_ext->lce.refcnt == 1) | |
1090 | mdev->resync_locked++; | |
1091 | set_bit(BME_NO_WRITES, &bm_ext->flags); | |
1092 | } | |
1093 | rs_flags = mdev->resync->flags; | |
1094 | spin_unlock_irq(&mdev->al_lock); | |
1095 | if (wakeup) | |
1096 | wake_up(&mdev->al_wait); | |
1097 | ||
1098 | if (!bm_ext) { | |
1099 | if (rs_flags & LC_STARVING) | |
1100 | dev_warn(DEV, "Have to wait for element" | |
1101 | " (resync LRU too small?)\n"); | |
1102 | BUG_ON(rs_flags & LC_DIRTY); | |
1103 | } | |
1104 | ||
1105 | return bm_ext; | |
1106 | } | |
1107 | ||
1108 | static int _is_in_al(struct drbd_conf *mdev, unsigned int enr) | |
1109 | { | |
1110 | struct lc_element *al_ext; | |
1111 | int rv = 0; | |
1112 | ||
1113 | spin_lock_irq(&mdev->al_lock); | |
1114 | if (unlikely(enr == mdev->act_log->new_number)) | |
1115 | rv = 1; | |
1116 | else { | |
1117 | al_ext = lc_find(mdev->act_log, enr); | |
1118 | if (al_ext) { | |
1119 | if (al_ext->refcnt) | |
1120 | rv = 1; | |
1121 | } | |
1122 | } | |
1123 | spin_unlock_irq(&mdev->al_lock); | |
1124 | ||
1125 | /* | |
1126 | if (unlikely(rv)) { | |
1127 | dev_info(DEV, "Delaying sync read until app's write is done\n"); | |
1128 | } | |
1129 | */ | |
1130 | return rv; | |
1131 | } | |
1132 | ||
1133 | /** | |
1134 | * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED | |
1135 | * @mdev: DRBD device. | |
1136 | * @sector: The sector number. | |
1137 | * | |
1138 | * This functions sleeps on al_wait. Returns 1 on success, 0 if interrupted. | |
1139 | */ | |
1140 | int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector) | |
1141 | { | |
1142 | unsigned int enr = BM_SECT_TO_EXT(sector); | |
1143 | struct bm_extent *bm_ext; | |
1144 | int i, sig; | |
1145 | ||
1146 | trace_drbd_resync(mdev, TRACE_LVL_ALL, | |
1147 | "drbd_rs_begin_io: sector=%llus (rs_end=%d)\n", | |
1148 | (unsigned long long)sector, enr); | |
1149 | ||
1150 | sig = wait_event_interruptible(mdev->al_wait, | |
1151 | (bm_ext = _bme_get(mdev, enr))); | |
1152 | if (sig) | |
1153 | return 0; | |
1154 | ||
1155 | if (test_bit(BME_LOCKED, &bm_ext->flags)) | |
1156 | return 1; | |
1157 | ||
1158 | for (i = 0; i < AL_EXT_PER_BM_SECT; i++) { | |
1159 | sig = wait_event_interruptible(mdev->al_wait, | |
1160 | !_is_in_al(mdev, enr * AL_EXT_PER_BM_SECT + i)); | |
1161 | if (sig) { | |
1162 | spin_lock_irq(&mdev->al_lock); | |
1163 | if (lc_put(mdev->resync, &bm_ext->lce) == 0) { | |
1164 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1165 | mdev->resync_locked--; | |
1166 | wake_up(&mdev->al_wait); | |
1167 | } | |
1168 | spin_unlock_irq(&mdev->al_lock); | |
1169 | return 0; | |
1170 | } | |
1171 | } | |
1172 | ||
1173 | set_bit(BME_LOCKED, &bm_ext->flags); | |
1174 | ||
1175 | return 1; | |
1176 | } | |
1177 | ||
1178 | /** | |
1179 | * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep | |
1180 | * @mdev: DRBD device. | |
1181 | * @sector: The sector number. | |
1182 | * | |
1183 | * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then | |
1184 | * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN | |
1185 | * if there is still application IO going on in this area. | |
1186 | */ | |
1187 | int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector) | |
1188 | { | |
1189 | unsigned int enr = BM_SECT_TO_EXT(sector); | |
1190 | const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT; | |
1191 | struct lc_element *e; | |
1192 | struct bm_extent *bm_ext; | |
1193 | int i; | |
1194 | ||
1195 | trace_drbd_resync(mdev, TRACE_LVL_ALL, "drbd_try_rs_begin_io: sector=%llus\n", | |
1196 | (unsigned long long)sector); | |
1197 | ||
1198 | spin_lock_irq(&mdev->al_lock); | |
1199 | if (mdev->resync_wenr != LC_FREE && mdev->resync_wenr != enr) { | |
1200 | /* in case you have very heavy scattered io, it may | |
1201 | * stall the syncer undefined if we give up the ref count | |
1202 | * when we try again and requeue. | |
1203 | * | |
1204 | * if we don't give up the refcount, but the next time | |
1205 | * we are scheduled this extent has been "synced" by new | |
1206 | * application writes, we'd miss the lc_put on the | |
1207 | * extent we keep the refcount on. | |
1208 | * so we remembered which extent we had to try again, and | |
1209 | * if the next requested one is something else, we do | |
1210 | * the lc_put here... | |
1211 | * we also have to wake_up | |
1212 | */ | |
1213 | ||
1214 | trace_drbd_resync(mdev, TRACE_LVL_ALL, | |
1215 | "dropping %u, apparently got 'synced' by application io\n", | |
1216 | mdev->resync_wenr); | |
1217 | ||
1218 | e = lc_find(mdev->resync, mdev->resync_wenr); | |
1219 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1220 | if (bm_ext) { | |
1221 | D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags)); | |
1222 | D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags)); | |
1223 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1224 | mdev->resync_wenr = LC_FREE; | |
1225 | if (lc_put(mdev->resync, &bm_ext->lce) == 0) | |
1226 | mdev->resync_locked--; | |
1227 | wake_up(&mdev->al_wait); | |
1228 | } else { | |
1229 | dev_alert(DEV, "LOGIC BUG\n"); | |
1230 | } | |
1231 | } | |
1232 | /* TRY. */ | |
1233 | e = lc_try_get(mdev->resync, enr); | |
1234 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1235 | if (bm_ext) { | |
1236 | if (test_bit(BME_LOCKED, &bm_ext->flags)) | |
1237 | goto proceed; | |
1238 | if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) { | |
1239 | mdev->resync_locked++; | |
1240 | } else { | |
1241 | /* we did set the BME_NO_WRITES, | |
1242 | * but then could not set BME_LOCKED, | |
1243 | * so we tried again. | |
1244 | * drop the extra reference. */ | |
1245 | trace_drbd_resync(mdev, TRACE_LVL_ALL, | |
1246 | "dropping extra reference on %u\n", enr); | |
1247 | ||
1248 | bm_ext->lce.refcnt--; | |
1249 | D_ASSERT(bm_ext->lce.refcnt > 0); | |
1250 | } | |
1251 | goto check_al; | |
1252 | } else { | |
1253 | /* do we rather want to try later? */ | |
1254 | if (mdev->resync_locked > mdev->resync->nr_elements-3) { | |
1255 | trace_drbd_resync(mdev, TRACE_LVL_ALL, | |
1256 | "resync_locked = %u!\n", mdev->resync_locked); | |
1257 | ||
1258 | goto try_again; | |
1259 | } | |
1260 | /* Do or do not. There is no try. -- Yoda */ | |
1261 | e = lc_get(mdev->resync, enr); | |
1262 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1263 | if (!bm_ext) { | |
1264 | const unsigned long rs_flags = mdev->resync->flags; | |
1265 | if (rs_flags & LC_STARVING) | |
1266 | dev_warn(DEV, "Have to wait for element" | |
1267 | " (resync LRU too small?)\n"); | |
1268 | BUG_ON(rs_flags & LC_DIRTY); | |
1269 | goto try_again; | |
1270 | } | |
1271 | if (bm_ext->lce.lc_number != enr) { | |
1272 | bm_ext->rs_left = drbd_bm_e_weight(mdev, enr); | |
1273 | bm_ext->rs_failed = 0; | |
1274 | lc_changed(mdev->resync, &bm_ext->lce); | |
1275 | wake_up(&mdev->al_wait); | |
1276 | D_ASSERT(test_bit(BME_LOCKED, &bm_ext->flags) == 0); | |
1277 | } | |
1278 | set_bit(BME_NO_WRITES, &bm_ext->flags); | |
1279 | D_ASSERT(bm_ext->lce.refcnt == 1); | |
1280 | mdev->resync_locked++; | |
1281 | goto check_al; | |
1282 | } | |
1283 | check_al: | |
1284 | trace_drbd_resync(mdev, TRACE_LVL_ALL, "checking al for %u\n", enr); | |
1285 | ||
1286 | for (i = 0; i < AL_EXT_PER_BM_SECT; i++) { | |
1287 | if (unlikely(al_enr+i == mdev->act_log->new_number)) | |
1288 | goto try_again; | |
1289 | if (lc_is_used(mdev->act_log, al_enr+i)) | |
1290 | goto try_again; | |
1291 | } | |
1292 | set_bit(BME_LOCKED, &bm_ext->flags); | |
1293 | proceed: | |
1294 | mdev->resync_wenr = LC_FREE; | |
1295 | spin_unlock_irq(&mdev->al_lock); | |
1296 | return 0; | |
1297 | ||
1298 | try_again: | |
1299 | trace_drbd_resync(mdev, TRACE_LVL_ALL, "need to try again for %u\n", enr); | |
1300 | if (bm_ext) | |
1301 | mdev->resync_wenr = enr; | |
1302 | spin_unlock_irq(&mdev->al_lock); | |
1303 | return -EAGAIN; | |
1304 | } | |
1305 | ||
1306 | void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector) | |
1307 | { | |
1308 | unsigned int enr = BM_SECT_TO_EXT(sector); | |
1309 | struct lc_element *e; | |
1310 | struct bm_extent *bm_ext; | |
1311 | unsigned long flags; | |
1312 | ||
1313 | trace_drbd_resync(mdev, TRACE_LVL_ALL, | |
1314 | "drbd_rs_complete_io: sector=%llus (rs_enr=%d)\n", | |
1315 | (long long)sector, enr); | |
1316 | ||
1317 | spin_lock_irqsave(&mdev->al_lock, flags); | |
1318 | e = lc_find(mdev->resync, enr); | |
1319 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1320 | if (!bm_ext) { | |
1321 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1322 | if (__ratelimit(&drbd_ratelimit_state)) | |
1323 | dev_err(DEV, "drbd_rs_complete_io() called, but extent not found\n"); | |
1324 | return; | |
1325 | } | |
1326 | ||
1327 | if (bm_ext->lce.refcnt == 0) { | |
1328 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1329 | dev_err(DEV, "drbd_rs_complete_io(,%llu [=%u]) called, " | |
1330 | "but refcnt is 0!?\n", | |
1331 | (unsigned long long)sector, enr); | |
1332 | return; | |
1333 | } | |
1334 | ||
1335 | if (lc_put(mdev->resync, &bm_ext->lce) == 0) { | |
1336 | clear_bit(BME_LOCKED, &bm_ext->flags); | |
1337 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1338 | mdev->resync_locked--; | |
1339 | wake_up(&mdev->al_wait); | |
1340 | } | |
1341 | ||
1342 | spin_unlock_irqrestore(&mdev->al_lock, flags); | |
1343 | } | |
1344 | ||
1345 | /** | |
1346 | * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED) | |
1347 | * @mdev: DRBD device. | |
1348 | */ | |
1349 | void drbd_rs_cancel_all(struct drbd_conf *mdev) | |
1350 | { | |
1351 | trace_drbd_resync(mdev, TRACE_LVL_METRICS, "drbd_rs_cancel_all\n"); | |
1352 | ||
1353 | spin_lock_irq(&mdev->al_lock); | |
1354 | ||
1355 | if (get_ldev_if_state(mdev, D_FAILED)) { /* Makes sure ->resync is there. */ | |
1356 | lc_reset(mdev->resync); | |
1357 | put_ldev(mdev); | |
1358 | } | |
1359 | mdev->resync_locked = 0; | |
1360 | mdev->resync_wenr = LC_FREE; | |
1361 | spin_unlock_irq(&mdev->al_lock); | |
1362 | wake_up(&mdev->al_wait); | |
1363 | } | |
1364 | ||
1365 | /** | |
1366 | * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU | |
1367 | * @mdev: DRBD device. | |
1368 | * | |
1369 | * Returns 0 upon success, -EAGAIN if at least one reference count was | |
1370 | * not zero. | |
1371 | */ | |
1372 | int drbd_rs_del_all(struct drbd_conf *mdev) | |
1373 | { | |
1374 | struct lc_element *e; | |
1375 | struct bm_extent *bm_ext; | |
1376 | int i; | |
1377 | ||
1378 | trace_drbd_resync(mdev, TRACE_LVL_METRICS, "drbd_rs_del_all\n"); | |
1379 | ||
1380 | spin_lock_irq(&mdev->al_lock); | |
1381 | ||
1382 | if (get_ldev_if_state(mdev, D_FAILED)) { | |
1383 | /* ok, ->resync is there. */ | |
1384 | for (i = 0; i < mdev->resync->nr_elements; i++) { | |
1385 | e = lc_element_by_index(mdev->resync, i); | |
1386 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL; | |
1387 | if (bm_ext->lce.lc_number == LC_FREE) | |
1388 | continue; | |
1389 | if (bm_ext->lce.lc_number == mdev->resync_wenr) { | |
1390 | dev_info(DEV, "dropping %u in drbd_rs_del_all, apparently" | |
1391 | " got 'synced' by application io\n", | |
1392 | mdev->resync_wenr); | |
1393 | D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags)); | |
1394 | D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags)); | |
1395 | clear_bit(BME_NO_WRITES, &bm_ext->flags); | |
1396 | mdev->resync_wenr = LC_FREE; | |
1397 | lc_put(mdev->resync, &bm_ext->lce); | |
1398 | } | |
1399 | if (bm_ext->lce.refcnt != 0) { | |
1400 | dev_info(DEV, "Retrying drbd_rs_del_all() later. " | |
1401 | "refcnt=%d\n", bm_ext->lce.refcnt); | |
1402 | put_ldev(mdev); | |
1403 | spin_unlock_irq(&mdev->al_lock); | |
1404 | return -EAGAIN; | |
1405 | } | |
1406 | D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags)); | |
1407 | D_ASSERT(!test_bit(BME_NO_WRITES, &bm_ext->flags)); | |
1408 | lc_del(mdev->resync, &bm_ext->lce); | |
1409 | } | |
1410 | D_ASSERT(mdev->resync->used == 0); | |
1411 | put_ldev(mdev); | |
1412 | } | |
1413 | spin_unlock_irq(&mdev->al_lock); | |
1414 | ||
1415 | return 0; | |
1416 | } | |
1417 | ||
1418 | /** | |
1419 | * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks | |
1420 | * @mdev: DRBD device. | |
1421 | * @sector: The sector number. | |
1422 | * @size: Size of failed IO operation, in byte. | |
1423 | */ | |
1424 | void drbd_rs_failed_io(struct drbd_conf *mdev, sector_t sector, int size) | |
1425 | { | |
1426 | /* Is called from worker and receiver context _only_ */ | |
1427 | unsigned long sbnr, ebnr, lbnr; | |
1428 | unsigned long count; | |
1429 | sector_t esector, nr_sectors; | |
1430 | int wake_up = 0; | |
1431 | ||
1432 | trace_drbd_resync(mdev, TRACE_LVL_SUMMARY, | |
1433 | "drbd_rs_failed_io: sector=%llus, size=%u\n", | |
1434 | (unsigned long long)sector, size); | |
1435 | ||
1436 | if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) { | |
1437 | dev_err(DEV, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n", | |
1438 | (unsigned long long)sector, size); | |
1439 | return; | |
1440 | } | |
1441 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
1442 | esector = sector + (size >> 9) - 1; | |
1443 | ||
1444 | ERR_IF(sector >= nr_sectors) return; | |
1445 | ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1); | |
1446 | ||
1447 | lbnr = BM_SECT_TO_BIT(nr_sectors-1); | |
1448 | ||
1449 | /* | |
1450 | * round up start sector, round down end sector. we make sure we only | |
1451 | * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */ | |
1452 | if (unlikely(esector < BM_SECT_PER_BIT-1)) | |
1453 | return; | |
1454 | if (unlikely(esector == (nr_sectors-1))) | |
1455 | ebnr = lbnr; | |
1456 | else | |
1457 | ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1)); | |
1458 | sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1); | |
1459 | ||
1460 | if (sbnr > ebnr) | |
1461 | return; | |
1462 | ||
1463 | /* | |
1464 | * ok, (capacity & 7) != 0 sometimes, but who cares... | |
1465 | * we count rs_{total,left} in bits, not sectors. | |
1466 | */ | |
1467 | spin_lock_irq(&mdev->al_lock); | |
1468 | count = drbd_bm_count_bits(mdev, sbnr, ebnr); | |
1469 | if (count) { | |
1470 | mdev->rs_failed += count; | |
1471 | ||
1472 | if (get_ldev(mdev)) { | |
1473 | drbd_try_clear_on_disk_bm(mdev, sector, count, FALSE); | |
1474 | put_ldev(mdev); | |
1475 | } | |
1476 | ||
1477 | /* just wake_up unconditional now, various lc_chaged(), | |
1478 | * lc_put() in drbd_try_clear_on_disk_bm(). */ | |
1479 | wake_up = 1; | |
1480 | } | |
1481 | spin_unlock_irq(&mdev->al_lock); | |
1482 | if (wake_up) | |
1483 | wake_up(&mdev->al_wait); | |
1484 | } |