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 sector_t sector, int op)
143 /* we do all our meta data IO in aligned 4k blocks. */
144 const int size = 4096;
145 int err, op_flags = 0;
147 device->md_io.done = 0;
148 device->md_io.error = -ENODEV;
150 if ((op == REQ_OP_WRITE) && !test_bit(MD_NO_FUA, &device->flags))
151 op_flags |= REQ_FUA | REQ_PREFLUSH;
152 op_flags |= REQ_SYNC | REQ_NOIDLE;
154 bio = bio_alloc_drbd(GFP_NOIO);
155 bio->bi_bdev = bdev->md_bdev;
156 bio->bi_iter.bi_sector = sector;
158 if (bio_add_page(bio, device->md_io.page, size, 0) != size)
160 bio->bi_private = device;
161 bio->bi_end_io = drbd_md_endio;
162 bio_set_op_attrs(bio, op, op_flags);
164 if (op != REQ_OP_WRITE && device->state.disk == D_DISKLESS && device->ldev == NULL)
165 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
167 else if (!get_ldev_if_state(device, D_ATTACHING)) {
168 /* Corresponding put_ldev in drbd_md_endio() */
169 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
174 bio_get(bio); /* one bio_put() is in the completion handler */
175 atomic_inc(&device->md_io.in_use); /* drbd_md_put_buffer() is in the completion handler */
176 device->md_io.submit_jif = jiffies;
177 if (drbd_insert_fault(device, (op == REQ_OP_WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
181 wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
183 err = device->md_io.error;
190 int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
191 sector_t sector, int op)
194 D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
196 BUG_ON(!bdev->md_bdev);
198 dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
199 current->comm, current->pid, __func__,
200 (unsigned long long)sector, (op == REQ_OP_WRITE) ? "WRITE" : "READ",
203 if (sector < drbd_md_first_sector(bdev) ||
204 sector + 7 > drbd_md_last_sector(bdev))
205 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
206 current->comm, current->pid, __func__,
207 (unsigned long long)sector,
208 (op == REQ_OP_WRITE) ? "WRITE" : "READ");
210 err = _drbd_md_sync_page_io(device, bdev, sector, op);
212 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
213 (unsigned long long)sector,
214 (op == REQ_OP_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 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
294 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
295 * are still coupled, or assume too much about their relation.
296 * Code below will not work if this is violated.
297 * Will be cleaned up with some followup patch.
302 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
307 /* al extent number to bit */
308 (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
311 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
313 const unsigned int stripes = device->ldev->md.al_stripes;
314 const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
316 /* transaction number, modulo on-disk ring buffer wrap around */
317 unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
319 /* ... to aligned 4k on disk block */
320 t = ((t % stripes) * stripe_size_4kB) + t/stripes;
322 /* ... to 512 byte sector in activity log */
325 /* ... plus offset to the on disk position */
326 return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
329 static int __al_write_transaction(struct drbd_device *device, struct al_transaction_on_disk *buffer)
331 struct lc_element *e;
338 memset(buffer, 0, sizeof(*buffer));
339 buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
340 buffer->tr_number = cpu_to_be32(device->al_tr_number);
344 /* Even though no one can start to change this list
345 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
346 * lc_try_lock_for_transaction() --, someone may still
347 * be in the process of changing it. */
348 spin_lock_irq(&device->al_lock);
349 list_for_each_entry(e, &device->act_log->to_be_changed, list) {
350 if (i == AL_UPDATES_PER_TRANSACTION) {
354 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
355 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
356 if (e->lc_number != LC_FREE)
357 drbd_bm_mark_for_writeout(device,
358 al_extent_to_bm_page(e->lc_number));
361 spin_unlock_irq(&device->al_lock);
362 BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
364 buffer->n_updates = cpu_to_be16(i);
365 for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
366 buffer->update_slot_nr[i] = cpu_to_be16(-1);
367 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
370 buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
371 buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
373 mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
374 device->act_log->nr_elements - device->al_tr_cycle);
375 for (i = 0; i < mx; i++) {
376 unsigned idx = device->al_tr_cycle + i;
377 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
378 buffer->context[i] = cpu_to_be32(extent_nr);
380 for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
381 buffer->context[i] = cpu_to_be32(LC_FREE);
383 device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
384 if (device->al_tr_cycle >= device->act_log->nr_elements)
385 device->al_tr_cycle = 0;
387 sector = al_tr_number_to_on_disk_sector(device);
389 crc = crc32c(0, buffer, 4096);
390 buffer->crc32c = cpu_to_be32(crc);
392 if (drbd_bm_write_hinted(device))
395 bool write_al_updates;
397 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
399 if (write_al_updates) {
400 if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
402 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
404 device->al_tr_number++;
405 device->al_writ_cnt++;
413 static int al_write_transaction(struct drbd_device *device)
415 struct al_transaction_on_disk *buffer;
418 if (!get_ldev(device)) {
419 drbd_err(device, "disk is %s, cannot start al transaction\n",
420 drbd_disk_str(device->state.disk));
424 /* The bitmap write may have failed, causing a state change. */
425 if (device->state.disk < D_INCONSISTENT) {
427 "disk is %s, cannot write al transaction\n",
428 drbd_disk_str(device->state.disk));
433 /* protects md_io_buffer, al_tr_cycle, ... */
434 buffer = drbd_md_get_buffer(device, __func__);
436 drbd_err(device, "disk failed while waiting for md_io buffer\n");
441 err = __al_write_transaction(device, buffer);
443 drbd_md_put_buffer(device);
450 void drbd_al_begin_io_commit(struct drbd_device *device)
454 /* Serialize multiple transactions.
455 * This uses test_and_set_bit, memory barrier is implicit.
457 wait_event(device->al_wait,
458 device->act_log->pending_changes == 0 ||
459 (locked = lc_try_lock_for_transaction(device->act_log)));
462 /* Double check: it may have been committed by someone else,
463 * while we have been waiting for the lock. */
464 if (device->act_log->pending_changes) {
465 bool write_al_updates;
468 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
471 if (write_al_updates)
472 al_write_transaction(device);
473 spin_lock_irq(&device->al_lock);
476 we need an "lc_cancel" here;
478 lc_committed(device->act_log);
479 spin_unlock_irq(&device->al_lock);
481 lc_unlock(device->act_log);
482 wake_up(&device->al_wait);
487 * @delegate: delegate activity log I/O to the worker thread
489 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
491 if (drbd_al_begin_io_prepare(device, i))
492 drbd_al_begin_io_commit(device);
495 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
497 struct lru_cache *al = device->act_log;
498 /* for bios crossing activity log extent boundaries,
499 * we may need to activate two extents in one go */
500 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
501 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
502 unsigned nr_al_extents;
503 unsigned available_update_slots;
506 D_ASSERT(device, first <= last);
508 nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
509 available_update_slots = min(al->nr_elements - al->used,
510 al->max_pending_changes - al->pending_changes);
512 /* We want all necessary updates for a given request within the same transaction
513 * We could first check how many updates are *actually* needed,
514 * and use that instead of the worst-case nr_al_extents */
515 if (available_update_slots < nr_al_extents) {
516 /* Too many activity log extents are currently "hot".
518 * If we have accumulated pending changes already,
521 * If we cannot get even a single pending change through,
522 * stop the fast path until we made some progress,
523 * or requests to "cold" extents could be starved. */
524 if (!al->pending_changes)
525 __set_bit(__LC_STARVING, &device->act_log->flags);
529 /* Is resync active in this area? */
530 for (enr = first; enr <= last; enr++) {
531 struct lc_element *tmp;
532 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
533 if (unlikely(tmp != NULL)) {
534 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
535 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
536 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
543 /* Checkout the refcounts.
544 * Given that we checked for available elements and update slots above,
545 * this has to be successful. */
546 for (enr = first; enr <= last; enr++) {
547 struct lc_element *al_ext;
548 al_ext = lc_get_cumulative(device->act_log, enr);
550 drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
555 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
557 /* for bios crossing activity log extent boundaries,
558 * we may need to activate two extents in one go */
559 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
560 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
562 struct lc_element *extent;
565 D_ASSERT(device, first <= last);
566 spin_lock_irqsave(&device->al_lock, flags);
568 for (enr = first; enr <= last; enr++) {
569 extent = lc_find(device->act_log, enr);
571 drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
574 lc_put(device->act_log, extent);
576 spin_unlock_irqrestore(&device->al_lock, flags);
577 wake_up(&device->al_wait);
580 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
584 spin_lock_irq(&device->al_lock);
585 rv = (al_ext->refcnt == 0);
587 lc_del(device->act_log, al_ext);
588 spin_unlock_irq(&device->al_lock);
594 * drbd_al_shrink() - Removes all active extents form the activity log
595 * @device: DRBD device.
597 * Removes all active extents form the activity log, waiting until
598 * the reference count of each entry dropped to 0 first, of course.
600 * You need to lock device->act_log with lc_try_lock() / lc_unlock()
602 void drbd_al_shrink(struct drbd_device *device)
604 struct lc_element *al_ext;
607 D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
609 for (i = 0; i < device->act_log->nr_elements; i++) {
610 al_ext = lc_element_by_index(device->act_log, i);
611 if (al_ext->lc_number == LC_FREE)
613 wait_event(device->al_wait, _try_lc_del(device, al_ext));
616 wake_up(&device->al_wait);
619 int drbd_al_initialize(struct drbd_device *device, void *buffer)
621 struct al_transaction_on_disk *al = buffer;
622 struct drbd_md *md = &device->ldev->md;
623 int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
626 __al_write_transaction(device, al);
627 /* There may or may not have been a pending transaction. */
628 spin_lock_irq(&device->al_lock);
629 lc_committed(device->act_log);
630 spin_unlock_irq(&device->al_lock);
632 /* The rest of the transactions will have an empty "updates" list, and
633 * are written out only to provide the context, and to initialize the
634 * on-disk ring buffer. */
635 for (i = 1; i < al_size_4k; i++) {
636 int err = __al_write_transaction(device, al);
643 static const char *drbd_change_sync_fname[] = {
644 [RECORD_RS_FAILED] = "drbd_rs_failed_io",
645 [SET_IN_SYNC] = "drbd_set_in_sync",
646 [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
649 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
650 * resync LRU-cache are 16MB each.
651 * The caller of this function has to hold an get_ldev() reference.
653 * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
654 * potentially pulling in (and recounting the corresponding bits)
655 * this resync extent into the resync extent lru cache.
657 * Returns whether all bits have been cleared for this resync extent,
658 * precisely: (rs_left <= rs_failed)
660 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
662 static bool update_rs_extent(struct drbd_device *device,
663 unsigned int enr, int count,
664 enum update_sync_bits_mode mode)
666 struct lc_element *e;
668 D_ASSERT(device, atomic_read(&device->local_cnt));
670 /* When setting out-of-sync bits,
671 * we don't need it cached (lc_find).
672 * But if it is present in the cache,
673 * we should update the cached bit count.
674 * Otherwise, that extent should be in the resync extent lru cache
675 * already -- or we want to pull it in if necessary -- (lc_get),
676 * then update and check rs_left and rs_failed. */
677 if (mode == SET_OUT_OF_SYNC)
678 e = lc_find(device->resync, enr);
680 e = lc_get(device->resync, enr);
682 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
683 if (ext->lce.lc_number == enr) {
684 if (mode == SET_IN_SYNC)
685 ext->rs_left -= count;
686 else if (mode == SET_OUT_OF_SYNC)
687 ext->rs_left += count;
689 ext->rs_failed += count;
690 if (ext->rs_left < ext->rs_failed) {
691 drbd_warn(device, "BAD! enr=%u rs_left=%d "
692 "rs_failed=%d count=%d cstate=%s\n",
693 ext->lce.lc_number, ext->rs_left,
694 ext->rs_failed, count,
695 drbd_conn_str(device->state.conn));
697 /* We don't expect to be able to clear more bits
698 * than have been set when we originally counted
699 * the set bits to cache that value in ext->rs_left.
700 * Whatever the reason (disconnect during resync,
701 * delayed local completion of an application write),
702 * try to fix it up by recounting here. */
703 ext->rs_left = drbd_bm_e_weight(device, enr);
706 /* Normally this element should be in the cache,
707 * since drbd_rs_begin_io() pulled it already in.
709 * But maybe an application write finished, and we set
710 * something outside the resync lru_cache in sync.
712 int rs_left = drbd_bm_e_weight(device, enr);
713 if (ext->flags != 0) {
714 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
716 ext->lce.lc_number, ext->rs_left,
717 ext->flags, enr, rs_left);
720 if (ext->rs_failed) {
721 drbd_warn(device, "Kicking resync_lru element enr=%u "
722 "out with rs_failed=%d\n",
723 ext->lce.lc_number, ext->rs_failed);
725 ext->rs_left = rs_left;
726 ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
727 /* we don't keep a persistent log of the resync lru,
728 * we can commit any change right away. */
729 lc_committed(device->resync);
731 if (mode != SET_OUT_OF_SYNC)
732 lc_put(device->resync, &ext->lce);
733 /* no race, we are within the al_lock! */
735 if (ext->rs_left <= ext->rs_failed) {
739 } else if (mode != SET_OUT_OF_SYNC) {
740 /* be quiet if lc_find() did not find it. */
741 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
742 device->resync_locked,
743 device->resync->nr_elements,
744 device->resync->flags);
749 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
751 unsigned long now = jiffies;
752 unsigned long last = device->rs_mark_time[device->rs_last_mark];
753 int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
754 if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
755 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
756 device->state.conn != C_PAUSED_SYNC_T &&
757 device->state.conn != C_PAUSED_SYNC_S) {
758 device->rs_mark_time[next] = now;
759 device->rs_mark_left[next] = still_to_go;
760 device->rs_last_mark = next;
765 /* It is called lazy update, so don't do write-out too often. */
766 static bool lazy_bitmap_update_due(struct drbd_device *device)
768 return time_after(jiffies, device->rs_last_bcast + 2*HZ);
771 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
774 set_bit(RS_DONE, &device->flags);
775 /* and also set RS_PROGRESS below */
776 else if (!lazy_bitmap_update_due(device))
779 drbd_device_post_work(device, RS_PROGRESS);
782 static int update_sync_bits(struct drbd_device *device,
783 unsigned long sbnr, unsigned long ebnr,
784 enum update_sync_bits_mode mode)
787 * We keep a count of set bits per resync-extent in the ->rs_left
788 * caching member, so we need to loop and work within the resync extent
789 * alignment. Typically this loop will execute exactly once.
792 unsigned long count = 0;
793 unsigned int cleared = 0;
794 while (sbnr <= ebnr) {
795 /* set temporary boundary bit number to last bit number within
796 * the resync extent of the current start bit number,
797 * but cap at provided end bit number */
798 unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
801 if (mode == RECORD_RS_FAILED)
802 /* Only called from drbd_rs_failed_io(), bits
803 * supposedly still set. Recount, maybe some
804 * of the bits have been successfully cleared
805 * by application IO meanwhile.
807 c = drbd_bm_count_bits(device, sbnr, tbnr);
808 else if (mode == SET_IN_SYNC)
809 c = drbd_bm_clear_bits(device, sbnr, tbnr);
810 else /* if (mode == SET_OUT_OF_SYNC) */
811 c = drbd_bm_set_bits(device, sbnr, tbnr);
814 spin_lock_irqsave(&device->al_lock, flags);
815 cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
816 spin_unlock_irqrestore(&device->al_lock, flags);
822 if (mode == SET_IN_SYNC) {
823 unsigned long still_to_go = drbd_bm_total_weight(device);
824 bool rs_is_done = (still_to_go <= device->rs_failed);
825 drbd_advance_rs_marks(device, still_to_go);
826 if (cleared || rs_is_done)
827 maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
828 } else if (mode == RECORD_RS_FAILED)
829 device->rs_failed += count;
830 wake_up(&device->al_wait);
835 /* clear the bit corresponding to the piece of storage in question:
836 * size byte of data starting from sector. Only clear a bits of the affected
837 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
839 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
842 int __drbd_change_sync(struct drbd_device *device, sector_t sector, int size,
843 enum update_sync_bits_mode mode)
845 /* Is called from worker and receiver context _only_ */
846 unsigned long sbnr, ebnr, lbnr;
847 unsigned long count = 0;
848 sector_t esector, nr_sectors;
850 /* This would be an empty REQ_PREFLUSH, be silent. */
851 if ((mode == SET_OUT_OF_SYNC) && size == 0)
854 if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
855 drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
856 drbd_change_sync_fname[mode],
857 (unsigned long long)sector, size);
861 if (!get_ldev(device))
862 return 0; /* no disk, no metadata, no bitmap to manipulate bits in */
864 nr_sectors = drbd_get_capacity(device->this_bdev);
865 esector = sector + (size >> 9) - 1;
867 if (!expect(sector < nr_sectors))
869 if (!expect(esector < nr_sectors))
870 esector = nr_sectors - 1;
872 lbnr = BM_SECT_TO_BIT(nr_sectors-1);
874 if (mode == SET_IN_SYNC) {
875 /* Round up start sector, round down end sector. We make sure
876 * we only clear full, aligned, BM_BLOCK_SIZE blocks. */
877 if (unlikely(esector < BM_SECT_PER_BIT-1))
879 if (unlikely(esector == (nr_sectors-1)))
882 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
883 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
885 /* We set it out of sync, or record resync failure.
886 * Should not round anything here. */
887 sbnr = BM_SECT_TO_BIT(sector);
888 ebnr = BM_SECT_TO_BIT(esector);
891 count = update_sync_bits(device, sbnr, ebnr, mode);
898 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
900 struct lc_element *e;
901 struct bm_extent *bm_ext;
903 unsigned long rs_flags;
905 spin_lock_irq(&device->al_lock);
906 if (device->resync_locked > device->resync->nr_elements/2) {
907 spin_unlock_irq(&device->al_lock);
910 e = lc_get(device->resync, enr);
911 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
913 if (bm_ext->lce.lc_number != enr) {
914 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
915 bm_ext->rs_failed = 0;
916 lc_committed(device->resync);
919 if (bm_ext->lce.refcnt == 1)
920 device->resync_locked++;
921 set_bit(BME_NO_WRITES, &bm_ext->flags);
923 rs_flags = device->resync->flags;
924 spin_unlock_irq(&device->al_lock);
926 wake_up(&device->al_wait);
929 if (rs_flags & LC_STARVING)
930 drbd_warn(device, "Have to wait for element"
931 " (resync LRU too small?)\n");
932 BUG_ON(rs_flags & LC_LOCKED);
938 static int _is_in_al(struct drbd_device *device, unsigned int enr)
942 spin_lock_irq(&device->al_lock);
943 rv = lc_is_used(device->act_log, enr);
944 spin_unlock_irq(&device->al_lock);
950 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
951 * @device: DRBD device.
952 * @sector: The sector number.
954 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
956 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
958 unsigned int enr = BM_SECT_TO_EXT(sector);
959 struct bm_extent *bm_ext;
964 sig = wait_event_interruptible(device->al_wait,
965 (bm_ext = _bme_get(device, enr)));
969 if (test_bit(BME_LOCKED, &bm_ext->flags))
972 /* step aside only while we are above c-min-rate; unless disabled. */
973 sa = drbd_rs_c_min_rate_throttle(device);
975 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
976 sig = wait_event_interruptible(device->al_wait,
977 !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
978 (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
980 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
981 spin_lock_irq(&device->al_lock);
982 if (lc_put(device->resync, &bm_ext->lce) == 0) {
983 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
984 device->resync_locked--;
985 wake_up(&device->al_wait);
987 spin_unlock_irq(&device->al_lock);
990 if (schedule_timeout_interruptible(HZ/10))
995 set_bit(BME_LOCKED, &bm_ext->flags);
1000 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1001 * @device: DRBD device.
1002 * @sector: The sector number.
1004 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1005 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1006 * if there is still application IO going on in this area.
1008 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
1010 unsigned int enr = BM_SECT_TO_EXT(sector);
1011 const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1012 struct lc_element *e;
1013 struct bm_extent *bm_ext;
1015 bool throttle = drbd_rs_should_slow_down(device, sector, true);
1017 /* If we need to throttle, a half-locked (only marked BME_NO_WRITES,
1018 * not yet BME_LOCKED) extent needs to be kicked out explicitly if we
1019 * need to throttle. There is at most one such half-locked extent,
1020 * which is remembered in resync_wenr. */
1022 if (throttle && device->resync_wenr != enr)
1025 spin_lock_irq(&device->al_lock);
1026 if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
1027 /* in case you have very heavy scattered io, it may
1028 * stall the syncer undefined if we give up the ref count
1029 * when we try again and requeue.
1031 * if we don't give up the refcount, but the next time
1032 * we are scheduled this extent has been "synced" by new
1033 * application writes, we'd miss the lc_put on the
1034 * extent we keep the refcount on.
1035 * so we remembered which extent we had to try again, and
1036 * if the next requested one is something else, we do
1037 * the lc_put here...
1038 * we also have to wake_up
1040 e = lc_find(device->resync, device->resync_wenr);
1041 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1043 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1044 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1045 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1046 device->resync_wenr = LC_FREE;
1047 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1049 device->resync_locked--;
1051 wake_up(&device->al_wait);
1053 drbd_alert(device, "LOGIC BUG\n");
1057 e = lc_try_get(device->resync, enr);
1058 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1060 if (test_bit(BME_LOCKED, &bm_ext->flags))
1062 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1063 device->resync_locked++;
1065 /* we did set the BME_NO_WRITES,
1066 * but then could not set BME_LOCKED,
1067 * so we tried again.
1068 * drop the extra reference. */
1069 bm_ext->lce.refcnt--;
1070 D_ASSERT(device, bm_ext->lce.refcnt > 0);
1074 /* do we rather want to try later? */
1075 if (device->resync_locked > device->resync->nr_elements-3)
1077 /* Do or do not. There is no try. -- Yoda */
1078 e = lc_get(device->resync, enr);
1079 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1081 const unsigned long rs_flags = device->resync->flags;
1082 if (rs_flags & LC_STARVING)
1083 drbd_warn(device, "Have to wait for element"
1084 " (resync LRU too small?)\n");
1085 BUG_ON(rs_flags & LC_LOCKED);
1088 if (bm_ext->lce.lc_number != enr) {
1089 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1090 bm_ext->rs_failed = 0;
1091 lc_committed(device->resync);
1092 wake_up(&device->al_wait);
1093 D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1095 set_bit(BME_NO_WRITES, &bm_ext->flags);
1096 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1097 device->resync_locked++;
1101 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1102 if (lc_is_used(device->act_log, al_enr+i))
1105 set_bit(BME_LOCKED, &bm_ext->flags);
1107 device->resync_wenr = LC_FREE;
1108 spin_unlock_irq(&device->al_lock);
1114 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1115 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1116 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1117 device->resync_wenr = LC_FREE;
1118 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1120 device->resync_locked--;
1122 wake_up(&device->al_wait);
1124 device->resync_wenr = enr;
1126 spin_unlock_irq(&device->al_lock);
1130 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1132 unsigned int enr = BM_SECT_TO_EXT(sector);
1133 struct lc_element *e;
1134 struct bm_extent *bm_ext;
1135 unsigned long flags;
1137 spin_lock_irqsave(&device->al_lock, flags);
1138 e = lc_find(device->resync, enr);
1139 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1141 spin_unlock_irqrestore(&device->al_lock, flags);
1142 if (__ratelimit(&drbd_ratelimit_state))
1143 drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1147 if (bm_ext->lce.refcnt == 0) {
1148 spin_unlock_irqrestore(&device->al_lock, flags);
1149 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1150 "but refcnt is 0!?\n",
1151 (unsigned long long)sector, enr);
1155 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1156 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1157 device->resync_locked--;
1158 wake_up(&device->al_wait);
1161 spin_unlock_irqrestore(&device->al_lock, flags);
1165 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1166 * @device: DRBD device.
1168 void drbd_rs_cancel_all(struct drbd_device *device)
1170 spin_lock_irq(&device->al_lock);
1172 if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1173 lc_reset(device->resync);
1176 device->resync_locked = 0;
1177 device->resync_wenr = LC_FREE;
1178 spin_unlock_irq(&device->al_lock);
1179 wake_up(&device->al_wait);
1183 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1184 * @device: DRBD device.
1186 * Returns 0 upon success, -EAGAIN if at least one reference count was
1189 int drbd_rs_del_all(struct drbd_device *device)
1191 struct lc_element *e;
1192 struct bm_extent *bm_ext;
1195 spin_lock_irq(&device->al_lock);
1197 if (get_ldev_if_state(device, D_FAILED)) {
1198 /* ok, ->resync is there. */
1199 for (i = 0; i < device->resync->nr_elements; i++) {
1200 e = lc_element_by_index(device->resync, i);
1201 bm_ext = lc_entry(e, struct bm_extent, lce);
1202 if (bm_ext->lce.lc_number == LC_FREE)
1204 if (bm_ext->lce.lc_number == device->resync_wenr) {
1205 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1206 " got 'synced' by application io\n",
1207 device->resync_wenr);
1208 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1209 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1210 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1211 device->resync_wenr = LC_FREE;
1212 lc_put(device->resync, &bm_ext->lce);
1214 if (bm_ext->lce.refcnt != 0) {
1215 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1216 "refcnt=%d\n", bm_ext->lce.refcnt);
1218 spin_unlock_irq(&device->al_lock);
1221 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1222 D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1223 lc_del(device->resync, &bm_ext->lce);
1225 D_ASSERT(device, device->resync->used == 0);
1228 spin_unlock_irq(&device->al_lock);
1229 wake_up(&device->al_wait);