4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
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>.
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)
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.
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.
26 #include <linux/slab.h>
27 #include <linux/crc32c.h>
28 #include <linux/drbd.h>
29 #include <linux/drbd_limits.h>
30 #include <linux/dynamic_debug.h>
34 enum al_transaction_types {
36 AL_TR_INITIALIZED = 0xffff
38 /* all fields on disc in big endian */
39 struct __packed al_transaction_on_disk {
40 /* don't we all like magic */
43 /* to identify the most recent transaction block
44 * in the on disk ring buffer */
47 /* checksum on the full 4k block, with this field set to 0. */
50 /* type of transaction, special transaction types like:
51 * purge-all, set-all-idle, set-all-active, ... to-be-defined
52 * see also enum al_transaction_types */
53 __be16 transaction_type;
55 /* we currently allow only a few thousand extents,
56 * so 16bit will be enough for the slot number. */
58 /* how many updates in this transaction */
61 /* maximum slot number, "al-extents" in drbd.conf speak.
62 * Having this in each transaction should make reconfiguration
63 * of that parameter easier. */
66 /* slot number the context starts with */
67 __be16 context_start_slot_nr;
69 /* Some reserved bytes. Expected usage is a 64bit counter of
70 * sectors-written since device creation, and other data generation tag
74 /* --- 36 byte used --- */
76 /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
77 * in one transaction, then use the remaining byte in the 4k block for
78 * context information. "Flexible" number of updates per transaction
79 * does not help, as we have to account for the case when all update
80 * slots are used anyways, so it would only complicate code without
83 __be16 update_slot_nr[AL_UPDATES_PER_TRANSACTION];
85 /* but the extent number is 32bit, which at an extent size of 4 MiB
86 * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
87 __be32 update_extent_nr[AL_UPDATES_PER_TRANSACTION];
89 /* --- 420 bytes used (36 + 64*6) --- */
91 /* 4096 - 420 = 3676 = 919 * 4 */
92 __be32 context[AL_CONTEXT_PER_TRANSACTION];
95 void *drbd_md_get_buffer(struct drbd_device *device, const char *intent)
99 wait_event(device->misc_wait,
100 (r = atomic_cmpxchg(&device->md_io.in_use, 0, 1)) == 0 ||
101 device->state.disk <= D_FAILED);
106 device->md_io.current_use = intent;
107 device->md_io.start_jif = jiffies;
108 device->md_io.submit_jif = device->md_io.start_jif - 1;
109 return page_address(device->md_io.page);
112 void drbd_md_put_buffer(struct drbd_device *device)
114 if (atomic_dec_and_test(&device->md_io.in_use))
115 wake_up(&device->misc_wait);
118 void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
124 dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
128 dt = MAX_SCHEDULE_TIMEOUT;
130 dt = wait_event_timeout(device->misc_wait,
131 *done || test_bit(FORCE_DETACH, &device->flags), dt);
133 drbd_err(device, "meta-data IO operation timed out\n");
134 drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
138 static int _drbd_md_sync_page_io(struct drbd_device *device,
139 struct drbd_backing_dev *bdev,
140 struct page *page, sector_t sector,
146 device->md_io.done = 0;
147 device->md_io.error = -ENODEV;
149 if ((rw & WRITE) && !test_bit(MD_NO_FUA, &device->flags))
150 rw |= REQ_FUA | REQ_FLUSH;
151 rw |= REQ_SYNC | REQ_NOIDLE;
153 bio = bio_alloc_drbd(GFP_NOIO);
154 bio->bi_bdev = bdev->md_bdev;
155 bio->bi_iter.bi_sector = sector;
157 if (bio_add_page(bio, page, size, 0) != size)
159 bio->bi_private = device;
160 bio->bi_end_io = drbd_md_io_complete;
163 if (!(rw & WRITE) && device->state.disk == D_DISKLESS && device->ldev == NULL)
164 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
166 else if (!get_ldev_if_state(device, D_ATTACHING)) {
167 /* Corresponding put_ldev in drbd_md_io_complete() */
168 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
173 bio_get(bio); /* one bio_put() is in the completion handler */
174 atomic_inc(&device->md_io.in_use); /* drbd_md_put_buffer() is in the completion handler */
175 device->md_io.submit_jif = jiffies;
176 if (drbd_insert_fault(device, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
177 bio_endio(bio, -EIO);
180 wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
181 if (bio_flagged(bio, BIO_UPTODATE))
182 err = device->md_io.error;
189 int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
190 sector_t sector, int rw)
193 struct page *iop = device->md_io.page;
195 D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
197 BUG_ON(!bdev->md_bdev);
199 dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
200 current->comm, current->pid, __func__,
201 (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ",
204 if (sector < drbd_md_first_sector(bdev) ||
205 sector + 7 > drbd_md_last_sector(bdev))
206 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
207 current->comm, current->pid, __func__,
208 (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
210 /* we do all our meta data IO in aligned 4k blocks. */
211 err = _drbd_md_sync_page_io(device, bdev, iop, sector, rw, 4096);
213 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
214 (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ", err);
219 static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
221 struct lc_element *tmp;
222 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
223 if (unlikely(tmp != NULL)) {
224 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
225 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
231 static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
233 struct lc_element *al_ext;
234 struct bm_extent *bm_ext;
237 spin_lock_irq(&device->al_lock);
238 bm_ext = find_active_resync_extent(device, enr);
240 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
241 spin_unlock_irq(&device->al_lock);
243 wake_up(&device->al_wait);
247 al_ext = lc_try_get(device->act_log, enr);
249 al_ext = lc_get(device->act_log, enr);
250 spin_unlock_irq(&device->al_lock);
254 bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
256 /* for bios crossing activity log extent boundaries,
257 * we may need to activate two extents in one go */
258 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
259 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
261 D_ASSERT(device, (unsigned)(last - first) <= 1);
262 D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
264 /* FIXME figure out a fast path for bios crossing AL extent boundaries */
268 return _al_get(device, first, true);
271 bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
273 /* for bios crossing activity log extent boundaries,
274 * we may need to activate two extents in one go */
275 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
276 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
278 bool need_transaction = false;
280 D_ASSERT(device, first <= last);
281 D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
283 for (enr = first; enr <= last; enr++) {
284 struct lc_element *al_ext;
285 wait_event(device->al_wait,
286 (al_ext = _al_get(device, enr, false)) != NULL);
287 if (al_ext->lc_number != enr)
288 need_transaction = true;
290 return need_transaction;
293 static int al_write_transaction(struct drbd_device *device);
295 void drbd_al_begin_io_commit(struct drbd_device *device)
299 /* Serialize multiple transactions.
300 * This uses test_and_set_bit, memory barrier is implicit.
302 wait_event(device->al_wait,
303 device->act_log->pending_changes == 0 ||
304 (locked = lc_try_lock_for_transaction(device->act_log)));
307 /* Double check: it may have been committed by someone else,
308 * while we have been waiting for the lock. */
309 if (device->act_log->pending_changes) {
310 bool write_al_updates;
313 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
316 if (write_al_updates)
317 al_write_transaction(device);
318 spin_lock_irq(&device->al_lock);
321 we need an "lc_cancel" here;
323 lc_committed(device->act_log);
324 spin_unlock_irq(&device->al_lock);
326 lc_unlock(device->act_log);
327 wake_up(&device->al_wait);
332 * @delegate: delegate activity log I/O to the worker thread
334 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
336 if (drbd_al_begin_io_prepare(device, i))
337 drbd_al_begin_io_commit(device);
340 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
342 struct lru_cache *al = device->act_log;
343 /* for bios crossing activity log extent boundaries,
344 * we may need to activate two extents in one go */
345 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
346 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
347 unsigned nr_al_extents;
348 unsigned available_update_slots;
351 D_ASSERT(device, first <= last);
353 nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
354 available_update_slots = min(al->nr_elements - al->used,
355 al->max_pending_changes - al->pending_changes);
357 /* We want all necessary updates for a given request within the same transaction
358 * We could first check how many updates are *actually* needed,
359 * and use that instead of the worst-case nr_al_extents */
360 if (available_update_slots < nr_al_extents)
363 /* Is resync active in this area? */
364 for (enr = first; enr <= last; enr++) {
365 struct lc_element *tmp;
366 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
367 if (unlikely(tmp != NULL)) {
368 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
369 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
370 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
377 /* Checkout the refcounts.
378 * Given that we checked for available elements and update slots above,
379 * this has to be successful. */
380 for (enr = first; enr <= last; enr++) {
381 struct lc_element *al_ext;
382 al_ext = lc_get_cumulative(device->act_log, enr);
384 drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
389 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
391 /* for bios crossing activity log extent boundaries,
392 * we may need to activate two extents in one go */
393 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
394 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
396 struct lc_element *extent;
399 D_ASSERT(device, first <= last);
400 spin_lock_irqsave(&device->al_lock, flags);
402 for (enr = first; enr <= last; enr++) {
403 extent = lc_find(device->act_log, enr);
405 drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
408 lc_put(device->act_log, extent);
410 spin_unlock_irqrestore(&device->al_lock, flags);
411 wake_up(&device->al_wait);
414 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
415 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
416 * are still coupled, or assume too much about their relation.
417 * Code below will not work if this is violated.
418 * Will be cleaned up with some followup patch.
423 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
428 /* al extent number to bit */
429 (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
432 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
434 const unsigned int stripes = device->ldev->md.al_stripes;
435 const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
437 /* transaction number, modulo on-disk ring buffer wrap around */
438 unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
440 /* ... to aligned 4k on disk block */
441 t = ((t % stripes) * stripe_size_4kB) + t/stripes;
443 /* ... to 512 byte sector in activity log */
446 /* ... plus offset to the on disk position */
447 return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
450 int al_write_transaction(struct drbd_device *device)
452 struct al_transaction_on_disk *buffer;
453 struct lc_element *e;
460 if (!get_ldev(device)) {
461 drbd_err(device, "disk is %s, cannot start al transaction\n",
462 drbd_disk_str(device->state.disk));
466 /* The bitmap write may have failed, causing a state change. */
467 if (device->state.disk < D_INCONSISTENT) {
469 "disk is %s, cannot write al transaction\n",
470 drbd_disk_str(device->state.disk));
475 /* protects md_io_buffer, al_tr_cycle, ... */
476 buffer = drbd_md_get_buffer(device, __func__);
478 drbd_err(device, "disk failed while waiting for md_io buffer\n");
483 memset(buffer, 0, sizeof(*buffer));
484 buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
485 buffer->tr_number = cpu_to_be32(device->al_tr_number);
489 /* Even though no one can start to change this list
490 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
491 * lc_try_lock_for_transaction() --, someone may still
492 * be in the process of changing it. */
493 spin_lock_irq(&device->al_lock);
494 list_for_each_entry(e, &device->act_log->to_be_changed, list) {
495 if (i == AL_UPDATES_PER_TRANSACTION) {
499 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
500 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
501 if (e->lc_number != LC_FREE)
502 drbd_bm_mark_for_writeout(device,
503 al_extent_to_bm_page(e->lc_number));
506 spin_unlock_irq(&device->al_lock);
507 BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
509 buffer->n_updates = cpu_to_be16(i);
510 for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
511 buffer->update_slot_nr[i] = cpu_to_be16(-1);
512 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
515 buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
516 buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
518 mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
519 device->act_log->nr_elements - device->al_tr_cycle);
520 for (i = 0; i < mx; i++) {
521 unsigned idx = device->al_tr_cycle + i;
522 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
523 buffer->context[i] = cpu_to_be32(extent_nr);
525 for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
526 buffer->context[i] = cpu_to_be32(LC_FREE);
528 device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
529 if (device->al_tr_cycle >= device->act_log->nr_elements)
530 device->al_tr_cycle = 0;
532 sector = al_tr_number_to_on_disk_sector(device);
534 crc = crc32c(0, buffer, 4096);
535 buffer->crc32c = cpu_to_be32(crc);
537 if (drbd_bm_write_hinted(device))
540 bool write_al_updates;
542 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
544 if (write_al_updates) {
545 if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
547 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
549 device->al_tr_number++;
550 device->al_writ_cnt++;
555 drbd_md_put_buffer(device);
561 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
565 spin_lock_irq(&device->al_lock);
566 rv = (al_ext->refcnt == 0);
568 lc_del(device->act_log, al_ext);
569 spin_unlock_irq(&device->al_lock);
575 * drbd_al_shrink() - Removes all active extents form the activity log
576 * @device: DRBD device.
578 * Removes all active extents form the activity log, waiting until
579 * the reference count of each entry dropped to 0 first, of course.
581 * You need to lock device->act_log with lc_try_lock() / lc_unlock()
583 void drbd_al_shrink(struct drbd_device *device)
585 struct lc_element *al_ext;
588 D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
590 for (i = 0; i < device->act_log->nr_elements; i++) {
591 al_ext = lc_element_by_index(device->act_log, i);
592 if (al_ext->lc_number == LC_FREE)
594 wait_event(device->al_wait, _try_lc_del(device, al_ext));
597 wake_up(&device->al_wait);
600 int drbd_initialize_al(struct drbd_device *device, void *buffer)
602 struct al_transaction_on_disk *al = buffer;
603 struct drbd_md *md = &device->ldev->md;
604 sector_t al_base = md->md_offset + md->al_offset;
605 int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
609 al->magic = cpu_to_be32(DRBD_AL_MAGIC);
610 al->transaction_type = cpu_to_be16(AL_TR_INITIALIZED);
611 al->crc32c = cpu_to_be32(crc32c(0, al, 4096));
613 for (i = 0; i < al_size_4k; i++) {
614 int err = drbd_md_sync_page_io(device, device->ldev, al_base + i * 8, WRITE);
621 static const char *drbd_change_sync_fname[] = {
622 [RECORD_RS_FAILED] = "drbd_rs_failed_io",
623 [SET_IN_SYNC] = "drbd_set_in_sync",
624 [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
627 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
628 * resync LRU-cache are 16MB each.
629 * The caller of this function has to hold an get_ldev() reference.
631 * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
632 * potentially pulling in (and recounting the corresponding bits)
633 * this resync extent into the resync extent lru cache.
635 * Returns whether all bits have been cleared for this resync extent,
636 * precisely: (rs_left <= rs_failed)
638 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
640 static bool update_rs_extent(struct drbd_device *device,
641 unsigned int enr, int count,
642 enum update_sync_bits_mode mode)
644 struct lc_element *e;
646 D_ASSERT(device, atomic_read(&device->local_cnt));
648 /* When setting out-of-sync bits,
649 * we don't need it cached (lc_find).
650 * But if it is present in the cache,
651 * we should update the cached bit count.
652 * Otherwise, that extent should be in the resync extent lru cache
653 * already -- or we want to pull it in if necessary -- (lc_get),
654 * then update and check rs_left and rs_failed. */
655 if (mode == SET_OUT_OF_SYNC)
656 e = lc_find(device->resync, enr);
658 e = lc_get(device->resync, enr);
660 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
661 if (ext->lce.lc_number == enr) {
662 if (mode == SET_IN_SYNC)
663 ext->rs_left -= count;
664 else if (mode == SET_OUT_OF_SYNC)
665 ext->rs_left += count;
667 ext->rs_failed += count;
668 if (ext->rs_left < ext->rs_failed) {
669 drbd_warn(device, "BAD! enr=%u rs_left=%d "
670 "rs_failed=%d count=%d cstate=%s\n",
671 ext->lce.lc_number, ext->rs_left,
672 ext->rs_failed, count,
673 drbd_conn_str(device->state.conn));
675 /* We don't expect to be able to clear more bits
676 * than have been set when we originally counted
677 * the set bits to cache that value in ext->rs_left.
678 * Whatever the reason (disconnect during resync,
679 * delayed local completion of an application write),
680 * try to fix it up by recounting here. */
681 ext->rs_left = drbd_bm_e_weight(device, enr);
684 /* Normally this element should be in the cache,
685 * since drbd_rs_begin_io() pulled it already in.
687 * But maybe an application write finished, and we set
688 * something outside the resync lru_cache in sync.
690 int rs_left = drbd_bm_e_weight(device, enr);
691 if (ext->flags != 0) {
692 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
694 ext->lce.lc_number, ext->rs_left,
695 ext->flags, enr, rs_left);
698 if (ext->rs_failed) {
699 drbd_warn(device, "Kicking resync_lru element enr=%u "
700 "out with rs_failed=%d\n",
701 ext->lce.lc_number, ext->rs_failed);
703 ext->rs_left = rs_left;
704 ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
705 /* we don't keep a persistent log of the resync lru,
706 * we can commit any change right away. */
707 lc_committed(device->resync);
709 if (mode != SET_OUT_OF_SYNC)
710 lc_put(device->resync, &ext->lce);
711 /* no race, we are within the al_lock! */
713 if (ext->rs_left <= ext->rs_failed) {
717 } else if (mode != SET_OUT_OF_SYNC) {
718 /* be quiet if lc_find() did not find it. */
719 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
720 device->resync_locked,
721 device->resync->nr_elements,
722 device->resync->flags);
727 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
729 unsigned long now = jiffies;
730 unsigned long last = device->rs_mark_time[device->rs_last_mark];
731 int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
732 if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
733 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
734 device->state.conn != C_PAUSED_SYNC_T &&
735 device->state.conn != C_PAUSED_SYNC_S) {
736 device->rs_mark_time[next] = now;
737 device->rs_mark_left[next] = still_to_go;
738 device->rs_last_mark = next;
743 /* It is called lazy update, so don't do write-out too often. */
744 static bool lazy_bitmap_update_due(struct drbd_device *device)
746 return time_after(jiffies, device->rs_last_bcast + 2*HZ);
749 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
752 set_bit(RS_DONE, &device->flags);
753 /* and also set RS_PROGRESS below */
754 else if (!lazy_bitmap_update_due(device))
757 drbd_device_post_work(device, RS_PROGRESS);
760 static int update_sync_bits(struct drbd_device *device,
761 unsigned long sbnr, unsigned long ebnr,
762 enum update_sync_bits_mode mode)
765 * We keep a count of set bits per resync-extent in the ->rs_left
766 * caching member, so we need to loop and work within the resync extent
767 * alignment. Typically this loop will execute exactly once.
770 unsigned long count = 0;
771 unsigned int cleared = 0;
772 while (sbnr <= ebnr) {
773 /* set temporary boundary bit number to last bit number within
774 * the resync extent of the current start bit number,
775 * but cap at provided end bit number */
776 unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
779 if (mode == RECORD_RS_FAILED)
780 /* Only called from drbd_rs_failed_io(), bits
781 * supposedly still set. Recount, maybe some
782 * of the bits have been successfully cleared
783 * by application IO meanwhile.
785 c = drbd_bm_count_bits(device, sbnr, tbnr);
786 else if (mode == SET_IN_SYNC)
787 c = drbd_bm_clear_bits(device, sbnr, tbnr);
788 else /* if (mode == SET_OUT_OF_SYNC) */
789 c = drbd_bm_set_bits(device, sbnr, tbnr);
792 spin_lock_irqsave(&device->al_lock, flags);
793 cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
794 spin_unlock_irqrestore(&device->al_lock, flags);
800 if (mode == SET_IN_SYNC) {
801 unsigned long still_to_go = drbd_bm_total_weight(device);
802 bool rs_is_done = (still_to_go <= device->rs_failed);
803 drbd_advance_rs_marks(device, still_to_go);
804 if (cleared || rs_is_done)
805 maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
806 } else if (mode == RECORD_RS_FAILED)
807 device->rs_failed += count;
808 wake_up(&device->al_wait);
813 /* clear the bit corresponding to the piece of storage in question:
814 * size byte of data starting from sector. Only clear a bits of the affected
815 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
817 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
820 int __drbd_change_sync(struct drbd_device *device, sector_t sector, int size,
821 enum update_sync_bits_mode mode,
822 const char *file, const unsigned int line)
824 /* Is called from worker and receiver context _only_ */
825 unsigned long sbnr, ebnr, lbnr;
826 unsigned long count = 0;
827 sector_t esector, nr_sectors;
829 /* This would be an empty REQ_FLUSH, be silent. */
830 if ((mode == SET_OUT_OF_SYNC) && size == 0)
833 if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
834 drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
835 drbd_change_sync_fname[mode],
836 (unsigned long long)sector, size);
840 if (!get_ldev(device))
841 return 0; /* no disk, no metadata, no bitmap to manipulate bits in */
843 nr_sectors = drbd_get_capacity(device->this_bdev);
844 esector = sector + (size >> 9) - 1;
846 if (!expect(sector < nr_sectors))
848 if (!expect(esector < nr_sectors))
849 esector = nr_sectors - 1;
851 lbnr = BM_SECT_TO_BIT(nr_sectors-1);
853 if (mode == SET_IN_SYNC) {
854 /* Round up start sector, round down end sector. We make sure
855 * we only clear full, aligned, BM_BLOCK_SIZE blocks. */
856 if (unlikely(esector < BM_SECT_PER_BIT-1))
858 if (unlikely(esector == (nr_sectors-1)))
861 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
862 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
864 /* We set it out of sync, or record resync failure.
865 * Should not round anything here. */
866 sbnr = BM_SECT_TO_BIT(sector);
867 ebnr = BM_SECT_TO_BIT(esector);
870 count = update_sync_bits(device, sbnr, ebnr, mode);
877 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
879 struct lc_element *e;
880 struct bm_extent *bm_ext;
882 unsigned long rs_flags;
884 spin_lock_irq(&device->al_lock);
885 if (device->resync_locked > device->resync->nr_elements/2) {
886 spin_unlock_irq(&device->al_lock);
889 e = lc_get(device->resync, enr);
890 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
892 if (bm_ext->lce.lc_number != enr) {
893 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
894 bm_ext->rs_failed = 0;
895 lc_committed(device->resync);
898 if (bm_ext->lce.refcnt == 1)
899 device->resync_locked++;
900 set_bit(BME_NO_WRITES, &bm_ext->flags);
902 rs_flags = device->resync->flags;
903 spin_unlock_irq(&device->al_lock);
905 wake_up(&device->al_wait);
908 if (rs_flags & LC_STARVING)
909 drbd_warn(device, "Have to wait for element"
910 " (resync LRU too small?)\n");
911 BUG_ON(rs_flags & LC_LOCKED);
917 static int _is_in_al(struct drbd_device *device, unsigned int enr)
921 spin_lock_irq(&device->al_lock);
922 rv = lc_is_used(device->act_log, enr);
923 spin_unlock_irq(&device->al_lock);
929 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
930 * @device: DRBD device.
931 * @sector: The sector number.
933 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
935 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
937 unsigned int enr = BM_SECT_TO_EXT(sector);
938 struct bm_extent *bm_ext;
943 sig = wait_event_interruptible(device->al_wait,
944 (bm_ext = _bme_get(device, enr)));
948 if (test_bit(BME_LOCKED, &bm_ext->flags))
951 /* step aside only while we are above c-min-rate; unless disabled. */
952 sa = drbd_rs_c_min_rate_throttle(device);
954 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
955 sig = wait_event_interruptible(device->al_wait,
956 !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
957 (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
959 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
960 spin_lock_irq(&device->al_lock);
961 if (lc_put(device->resync, &bm_ext->lce) == 0) {
962 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
963 device->resync_locked--;
964 wake_up(&device->al_wait);
966 spin_unlock_irq(&device->al_lock);
969 if (schedule_timeout_interruptible(HZ/10))
974 set_bit(BME_LOCKED, &bm_ext->flags);
979 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
980 * @device: DRBD device.
981 * @sector: The sector number.
983 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
984 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
985 * if there is still application IO going on in this area.
987 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
989 unsigned int enr = BM_SECT_TO_EXT(sector);
990 const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
991 struct lc_element *e;
992 struct bm_extent *bm_ext;
995 spin_lock_irq(&device->al_lock);
996 if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
997 /* in case you have very heavy scattered io, it may
998 * stall the syncer undefined if we give up the ref count
999 * when we try again and requeue.
1001 * if we don't give up the refcount, but the next time
1002 * we are scheduled this extent has been "synced" by new
1003 * application writes, we'd miss the lc_put on the
1004 * extent we keep the refcount on.
1005 * so we remembered which extent we had to try again, and
1006 * if the next requested one is something else, we do
1007 * the lc_put here...
1008 * we also have to wake_up
1010 e = lc_find(device->resync, device->resync_wenr);
1011 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1013 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1014 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1015 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1016 device->resync_wenr = LC_FREE;
1017 if (lc_put(device->resync, &bm_ext->lce) == 0)
1018 device->resync_locked--;
1019 wake_up(&device->al_wait);
1021 drbd_alert(device, "LOGIC BUG\n");
1025 e = lc_try_get(device->resync, enr);
1026 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1028 if (test_bit(BME_LOCKED, &bm_ext->flags))
1030 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1031 device->resync_locked++;
1033 /* we did set the BME_NO_WRITES,
1034 * but then could not set BME_LOCKED,
1035 * so we tried again.
1036 * drop the extra reference. */
1037 bm_ext->lce.refcnt--;
1038 D_ASSERT(device, bm_ext->lce.refcnt > 0);
1042 /* do we rather want to try later? */
1043 if (device->resync_locked > device->resync->nr_elements-3)
1045 /* Do or do not. There is no try. -- Yoda */
1046 e = lc_get(device->resync, enr);
1047 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1049 const unsigned long rs_flags = device->resync->flags;
1050 if (rs_flags & LC_STARVING)
1051 drbd_warn(device, "Have to wait for element"
1052 " (resync LRU too small?)\n");
1053 BUG_ON(rs_flags & LC_LOCKED);
1056 if (bm_ext->lce.lc_number != enr) {
1057 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1058 bm_ext->rs_failed = 0;
1059 lc_committed(device->resync);
1060 wake_up(&device->al_wait);
1061 D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1063 set_bit(BME_NO_WRITES, &bm_ext->flags);
1064 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1065 device->resync_locked++;
1069 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1070 if (lc_is_used(device->act_log, al_enr+i))
1073 set_bit(BME_LOCKED, &bm_ext->flags);
1075 device->resync_wenr = LC_FREE;
1076 spin_unlock_irq(&device->al_lock);
1081 device->resync_wenr = enr;
1082 spin_unlock_irq(&device->al_lock);
1086 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1088 unsigned int enr = BM_SECT_TO_EXT(sector);
1089 struct lc_element *e;
1090 struct bm_extent *bm_ext;
1091 unsigned long flags;
1093 spin_lock_irqsave(&device->al_lock, flags);
1094 e = lc_find(device->resync, enr);
1095 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1097 spin_unlock_irqrestore(&device->al_lock, flags);
1098 if (__ratelimit(&drbd_ratelimit_state))
1099 drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1103 if (bm_ext->lce.refcnt == 0) {
1104 spin_unlock_irqrestore(&device->al_lock, flags);
1105 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1106 "but refcnt is 0!?\n",
1107 (unsigned long long)sector, enr);
1111 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1112 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1113 device->resync_locked--;
1114 wake_up(&device->al_wait);
1117 spin_unlock_irqrestore(&device->al_lock, flags);
1121 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1122 * @device: DRBD device.
1124 void drbd_rs_cancel_all(struct drbd_device *device)
1126 spin_lock_irq(&device->al_lock);
1128 if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1129 lc_reset(device->resync);
1132 device->resync_locked = 0;
1133 device->resync_wenr = LC_FREE;
1134 spin_unlock_irq(&device->al_lock);
1135 wake_up(&device->al_wait);
1139 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1140 * @device: DRBD device.
1142 * Returns 0 upon success, -EAGAIN if at least one reference count was
1145 int drbd_rs_del_all(struct drbd_device *device)
1147 struct lc_element *e;
1148 struct bm_extent *bm_ext;
1151 spin_lock_irq(&device->al_lock);
1153 if (get_ldev_if_state(device, D_FAILED)) {
1154 /* ok, ->resync is there. */
1155 for (i = 0; i < device->resync->nr_elements; i++) {
1156 e = lc_element_by_index(device->resync, i);
1157 bm_ext = lc_entry(e, struct bm_extent, lce);
1158 if (bm_ext->lce.lc_number == LC_FREE)
1160 if (bm_ext->lce.lc_number == device->resync_wenr) {
1161 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1162 " got 'synced' by application io\n",
1163 device->resync_wenr);
1164 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1165 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1166 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1167 device->resync_wenr = LC_FREE;
1168 lc_put(device->resync, &bm_ext->lce);
1170 if (bm_ext->lce.refcnt != 0) {
1171 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1172 "refcnt=%d\n", bm_ext->lce.refcnt);
1174 spin_unlock_irq(&device->al_lock);
1177 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1178 D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1179 lc_del(device->resync, &bm_ext->lce);
1181 D_ASSERT(device, device->resync->used == 0);
1184 spin_unlock_irq(&device->al_lock);
1185 wake_up(&device->al_wait);