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 struct update_odbm_work {
97 struct drbd_device *device;
101 struct update_al_work {
103 struct drbd_device *device;
104 struct completion event;
109 void *drbd_md_get_buffer(struct drbd_device *device)
113 wait_event(device->misc_wait,
114 (r = atomic_cmpxchg(&device->md_io_in_use, 0, 1)) == 0 ||
115 device->state.disk <= D_FAILED);
117 return r ? NULL : page_address(device->md_io_page);
120 void drbd_md_put_buffer(struct drbd_device *device)
122 if (atomic_dec_and_test(&device->md_io_in_use))
123 wake_up(&device->misc_wait);
126 void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
132 dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
136 dt = MAX_SCHEDULE_TIMEOUT;
138 dt = wait_event_timeout(device->misc_wait,
139 *done || test_bit(FORCE_DETACH, &device->flags), dt);
141 drbd_err(device, "meta-data IO operation timed out\n");
142 drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
146 static int _drbd_md_sync_page_io(struct drbd_device *device,
147 struct drbd_backing_dev *bdev,
148 struct page *page, sector_t sector,
154 device->md_io.done = 0;
155 device->md_io.error = -ENODEV;
157 if ((rw & WRITE) && !test_bit(MD_NO_FUA, &device->flags))
158 rw |= REQ_FUA | REQ_FLUSH;
161 bio = bio_alloc_drbd(GFP_NOIO);
162 bio->bi_bdev = bdev->md_bdev;
163 bio->bi_iter.bi_sector = sector;
165 if (bio_add_page(bio, page, size, 0) != size)
167 bio->bi_private = &device->md_io;
168 bio->bi_end_io = drbd_md_io_complete;
171 if (!(rw & WRITE) && device->state.disk == D_DISKLESS && device->ldev == NULL)
172 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
174 else if (!get_ldev_if_state(device, D_ATTACHING)) {
175 /* Corresponding put_ldev in drbd_md_io_complete() */
176 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
181 bio_get(bio); /* one bio_put() is in the completion handler */
182 atomic_inc(&device->md_io_in_use); /* drbd_md_put_buffer() is in the completion handler */
183 if (drbd_insert_fault(device, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
184 bio_endio(bio, -EIO);
187 wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
188 if (bio_flagged(bio, BIO_UPTODATE))
189 err = device->md_io.error;
196 int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
197 sector_t sector, int rw)
200 struct page *iop = device->md_io_page;
202 D_ASSERT(device, atomic_read(&device->md_io_in_use) == 1);
204 BUG_ON(!bdev->md_bdev);
206 dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
207 current->comm, current->pid, __func__,
208 (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ",
211 if (sector < drbd_md_first_sector(bdev) ||
212 sector + 7 > drbd_md_last_sector(bdev))
213 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
214 current->comm, current->pid, __func__,
215 (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
217 /* we do all our meta data IO in aligned 4k blocks. */
218 err = _drbd_md_sync_page_io(device, bdev, iop, sector, rw, 4096);
220 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
221 (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ", err);
226 static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
228 struct lc_element *tmp;
229 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
230 if (unlikely(tmp != NULL)) {
231 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
232 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
238 static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
240 struct lc_element *al_ext;
241 struct bm_extent *bm_ext;
244 spin_lock_irq(&device->al_lock);
245 bm_ext = find_active_resync_extent(device, enr);
247 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
248 spin_unlock_irq(&device->al_lock);
250 wake_up(&device->al_wait);
254 al_ext = lc_try_get(device->act_log, enr);
256 al_ext = lc_get(device->act_log, enr);
257 spin_unlock_irq(&device->al_lock);
261 bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
263 /* for bios crossing activity log extent boundaries,
264 * we may need to activate two extents in one go */
265 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
266 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
268 D_ASSERT(device, (unsigned)(last - first) <= 1);
269 D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
271 /* FIXME figure out a fast path for bios crossing AL extent boundaries */
275 return _al_get(device, first, true);
278 bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
280 /* for bios crossing activity log extent boundaries,
281 * we may need to activate two extents in one go */
282 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
283 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
285 bool need_transaction = false;
287 D_ASSERT(device, first <= last);
288 D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
290 for (enr = first; enr <= last; enr++) {
291 struct lc_element *al_ext;
292 wait_event(device->al_wait,
293 (al_ext = _al_get(device, enr, false)) != NULL);
294 if (al_ext->lc_number != enr)
295 need_transaction = true;
297 return need_transaction;
300 static int al_write_transaction(struct drbd_device *device, bool delegate);
302 /* When called through generic_make_request(), we must delegate
303 * activity log I/O to the worker thread: a further request
304 * submitted via generic_make_request() within the same task
305 * would be queued on current->bio_list, and would only start
306 * after this function returns (see generic_make_request()).
308 * However, if we *are* the worker, we must not delegate to ourselves.
312 * @delegate: delegate activity log I/O to the worker thread
314 void drbd_al_begin_io_commit(struct drbd_device *device, bool delegate)
318 BUG_ON(delegate && current == first_peer_device(device)->connection->worker.task);
320 /* Serialize multiple transactions.
321 * This uses test_and_set_bit, memory barrier is implicit.
323 wait_event(device->al_wait,
324 device->act_log->pending_changes == 0 ||
325 (locked = lc_try_lock_for_transaction(device->act_log)));
328 /* Double check: it may have been committed by someone else,
329 * while we have been waiting for the lock. */
330 if (device->act_log->pending_changes) {
331 bool write_al_updates;
334 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
337 if (write_al_updates)
338 al_write_transaction(device, delegate);
339 spin_lock_irq(&device->al_lock);
342 we need an "lc_cancel" here;
344 lc_committed(device->act_log);
345 spin_unlock_irq(&device->al_lock);
347 lc_unlock(device->act_log);
348 wake_up(&device->al_wait);
353 * @delegate: delegate activity log I/O to the worker thread
355 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i, bool delegate)
357 BUG_ON(delegate && current == first_peer_device(device)->connection->worker.task);
359 if (drbd_al_begin_io_prepare(device, i))
360 drbd_al_begin_io_commit(device, delegate);
363 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
365 struct lru_cache *al = device->act_log;
366 /* for bios crossing activity log extent boundaries,
367 * we may need to activate two extents in one go */
368 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
369 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
370 unsigned nr_al_extents;
371 unsigned available_update_slots;
374 D_ASSERT(device, first <= last);
376 nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
377 available_update_slots = min(al->nr_elements - al->used,
378 al->max_pending_changes - al->pending_changes);
380 /* We want all necessary updates for a given request within the same transaction
381 * We could first check how many updates are *actually* needed,
382 * and use that instead of the worst-case nr_al_extents */
383 if (available_update_slots < nr_al_extents)
386 /* Is resync active in this area? */
387 for (enr = first; enr <= last; enr++) {
388 struct lc_element *tmp;
389 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
390 if (unlikely(tmp != NULL)) {
391 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
392 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
393 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
400 /* Checkout the refcounts.
401 * Given that we checked for available elements and update slots above,
402 * this has to be successful. */
403 for (enr = first; enr <= last; enr++) {
404 struct lc_element *al_ext;
405 al_ext = lc_get_cumulative(device->act_log, enr);
407 drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
412 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
414 /* for bios crossing activity log extent boundaries,
415 * we may need to activate two extents in one go */
416 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
417 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
419 struct lc_element *extent;
422 D_ASSERT(device, first <= last);
423 spin_lock_irqsave(&device->al_lock, flags);
425 for (enr = first; enr <= last; enr++) {
426 extent = lc_find(device->act_log, enr);
428 drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
431 lc_put(device->act_log, extent);
433 spin_unlock_irqrestore(&device->al_lock, flags);
434 wake_up(&device->al_wait);
437 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
438 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
439 * are still coupled, or assume too much about their relation.
440 * Code below will not work if this is violated.
441 * Will be cleaned up with some followup patch.
446 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
451 /* al extent number to bit */
452 (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
455 static unsigned int rs_extent_to_bm_page(unsigned int rs_enr)
460 /* resync extent number to bit */
461 (BM_EXT_SHIFT - BM_BLOCK_SHIFT));
464 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
466 const unsigned int stripes = device->ldev->md.al_stripes;
467 const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
469 /* transaction number, modulo on-disk ring buffer wrap around */
470 unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
472 /* ... to aligned 4k on disk block */
473 t = ((t % stripes) * stripe_size_4kB) + t/stripes;
475 /* ... to 512 byte sector in activity log */
478 /* ... plus offset to the on disk position */
479 return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
483 _al_write_transaction(struct drbd_device *device)
485 struct al_transaction_on_disk *buffer;
486 struct lc_element *e;
493 if (!get_ldev(device)) {
494 drbd_err(device, "disk is %s, cannot start al transaction\n",
495 drbd_disk_str(device->state.disk));
499 /* The bitmap write may have failed, causing a state change. */
500 if (device->state.disk < D_INCONSISTENT) {
502 "disk is %s, cannot write al transaction\n",
503 drbd_disk_str(device->state.disk));
508 buffer = drbd_md_get_buffer(device); /* protects md_io_buffer, al_tr_cycle, ... */
510 drbd_err(device, "disk failed while waiting for md_io buffer\n");
515 memset(buffer, 0, sizeof(*buffer));
516 buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
517 buffer->tr_number = cpu_to_be32(device->al_tr_number);
521 /* Even though no one can start to change this list
522 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
523 * lc_try_lock_for_transaction() --, someone may still
524 * be in the process of changing it. */
525 spin_lock_irq(&device->al_lock);
526 list_for_each_entry(e, &device->act_log->to_be_changed, list) {
527 if (i == AL_UPDATES_PER_TRANSACTION) {
531 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
532 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
533 if (e->lc_number != LC_FREE)
534 drbd_bm_mark_for_writeout(device,
535 al_extent_to_bm_page(e->lc_number));
538 spin_unlock_irq(&device->al_lock);
539 BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
541 buffer->n_updates = cpu_to_be16(i);
542 for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
543 buffer->update_slot_nr[i] = cpu_to_be16(-1);
544 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
547 buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
548 buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
550 mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
551 device->act_log->nr_elements - device->al_tr_cycle);
552 for (i = 0; i < mx; i++) {
553 unsigned idx = device->al_tr_cycle + i;
554 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
555 buffer->context[i] = cpu_to_be32(extent_nr);
557 for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
558 buffer->context[i] = cpu_to_be32(LC_FREE);
560 device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
561 if (device->al_tr_cycle >= device->act_log->nr_elements)
562 device->al_tr_cycle = 0;
564 sector = al_tr_number_to_on_disk_sector(device);
566 crc = crc32c(0, buffer, 4096);
567 buffer->crc32c = cpu_to_be32(crc);
569 if (drbd_bm_write_hinted(device))
572 bool write_al_updates;
574 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
576 if (write_al_updates) {
577 if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
579 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
581 device->al_tr_number++;
582 device->al_writ_cnt++;
587 drbd_md_put_buffer(device);
594 static int w_al_write_transaction(struct drbd_work *w, int unused)
596 struct update_al_work *aw = container_of(w, struct update_al_work, w);
597 struct drbd_device *device = aw->device;
600 err = _al_write_transaction(device);
602 complete(&aw->event);
604 return err != -EIO ? err : 0;
607 /* Calls from worker context (see w_restart_disk_io()) need to write the
608 transaction directly. Others came through generic_make_request(),
609 those need to delegate it to the worker. */
610 static int al_write_transaction(struct drbd_device *device, bool delegate)
613 struct update_al_work al_work;
614 init_completion(&al_work.event);
615 al_work.w.cb = w_al_write_transaction;
616 al_work.device = device;
617 drbd_queue_work_front(&first_peer_device(device)->connection->sender_work,
619 wait_for_completion(&al_work.event);
622 return _al_write_transaction(device);
625 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
629 spin_lock_irq(&device->al_lock);
630 rv = (al_ext->refcnt == 0);
632 lc_del(device->act_log, al_ext);
633 spin_unlock_irq(&device->al_lock);
639 * drbd_al_shrink() - Removes all active extents form the activity log
640 * @device: DRBD device.
642 * Removes all active extents form the activity log, waiting until
643 * the reference count of each entry dropped to 0 first, of course.
645 * You need to lock device->act_log with lc_try_lock() / lc_unlock()
647 void drbd_al_shrink(struct drbd_device *device)
649 struct lc_element *al_ext;
652 D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
654 for (i = 0; i < device->act_log->nr_elements; i++) {
655 al_ext = lc_element_by_index(device->act_log, i);
656 if (al_ext->lc_number == LC_FREE)
658 wait_event(device->al_wait, _try_lc_del(device, al_ext));
661 wake_up(&device->al_wait);
664 int drbd_initialize_al(struct drbd_device *device, void *buffer)
666 struct al_transaction_on_disk *al = buffer;
667 struct drbd_md *md = &device->ldev->md;
668 sector_t al_base = md->md_offset + md->al_offset;
669 int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
673 al->magic = cpu_to_be32(DRBD_AL_MAGIC);
674 al->transaction_type = cpu_to_be16(AL_TR_INITIALIZED);
675 al->crc32c = cpu_to_be32(crc32c(0, al, 4096));
677 for (i = 0; i < al_size_4k; i++) {
678 int err = drbd_md_sync_page_io(device, device->ldev, al_base + i * 8, WRITE);
685 static int w_update_odbm(struct drbd_work *w, int unused)
687 struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w);
688 struct drbd_device *device = udw->device;
689 struct sib_info sib = { .sib_reason = SIB_SYNC_PROGRESS, };
691 if (!get_ldev(device)) {
692 if (__ratelimit(&drbd_ratelimit_state))
693 drbd_warn(device, "Can not update on disk bitmap, local IO disabled.\n");
698 drbd_bm_write_page(device, rs_extent_to_bm_page(udw->enr));
703 if (drbd_bm_total_weight(device) <= device->rs_failed) {
704 switch (device->state.conn) {
705 case C_SYNC_SOURCE: case C_SYNC_TARGET:
706 case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T:
707 drbd_resync_finished(device);
713 drbd_bcast_event(device, &sib);
719 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
720 * resync LRU-cache are 16MB each.
721 * The caller of this function has to hold an get_ldev() reference.
723 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
725 static void drbd_try_clear_on_disk_bm(struct drbd_device *device, sector_t sector,
726 int count, int success)
728 struct lc_element *e;
729 struct update_odbm_work *udw;
733 D_ASSERT(device, atomic_read(&device->local_cnt));
735 /* I simply assume that a sector/size pair never crosses
736 * a 16 MB extent border. (Currently this is true...) */
737 enr = BM_SECT_TO_EXT(sector);
739 e = lc_get(device->resync, enr);
741 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
742 if (ext->lce.lc_number == enr) {
744 ext->rs_left -= count;
746 ext->rs_failed += count;
747 if (ext->rs_left < ext->rs_failed) {
748 drbd_warn(device, "BAD! sector=%llus enr=%u rs_left=%d "
749 "rs_failed=%d count=%d cstate=%s\n",
750 (unsigned long long)sector,
751 ext->lce.lc_number, ext->rs_left,
752 ext->rs_failed, count,
753 drbd_conn_str(device->state.conn));
755 /* We don't expect to be able to clear more bits
756 * than have been set when we originally counted
757 * the set bits to cache that value in ext->rs_left.
758 * Whatever the reason (disconnect during resync,
759 * delayed local completion of an application write),
760 * try to fix it up by recounting here. */
761 ext->rs_left = drbd_bm_e_weight(device, enr);
764 /* Normally this element should be in the cache,
765 * since drbd_rs_begin_io() pulled it already in.
767 * But maybe an application write finished, and we set
768 * something outside the resync lru_cache in sync.
770 int rs_left = drbd_bm_e_weight(device, enr);
771 if (ext->flags != 0) {
772 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
774 ext->lce.lc_number, ext->rs_left,
775 ext->flags, enr, rs_left);
778 if (ext->rs_failed) {
779 drbd_warn(device, "Kicking resync_lru element enr=%u "
780 "out with rs_failed=%d\n",
781 ext->lce.lc_number, ext->rs_failed);
783 ext->rs_left = rs_left;
784 ext->rs_failed = success ? 0 : count;
785 /* we don't keep a persistent log of the resync lru,
786 * we can commit any change right away. */
787 lc_committed(device->resync);
789 lc_put(device->resync, &ext->lce);
790 /* no race, we are within the al_lock! */
792 if (ext->rs_left == ext->rs_failed) {
795 udw = kmalloc(sizeof(*udw), GFP_ATOMIC);
797 udw->enr = ext->lce.lc_number;
798 udw->w.cb = w_update_odbm;
799 udw->device = device;
800 drbd_queue_work_front(&first_peer_device(device)->connection->sender_work,
803 drbd_warn(device, "Could not kmalloc an udw\n");
807 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
808 device->resync_locked,
809 device->resync->nr_elements,
810 device->resync->flags);
814 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
816 unsigned long now = jiffies;
817 unsigned long last = device->rs_mark_time[device->rs_last_mark];
818 int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
819 if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
820 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
821 device->state.conn != C_PAUSED_SYNC_T &&
822 device->state.conn != C_PAUSED_SYNC_S) {
823 device->rs_mark_time[next] = now;
824 device->rs_mark_left[next] = still_to_go;
825 device->rs_last_mark = next;
830 /* clear the bit corresponding to the piece of storage in question:
831 * size byte of data starting from sector. Only clear a bits of the affected
832 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
834 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
837 void __drbd_set_in_sync(struct drbd_device *device, sector_t sector, int size,
838 const char *file, const unsigned int line)
840 /* Is called from worker and receiver context _only_ */
841 unsigned long sbnr, ebnr, lbnr;
842 unsigned long count = 0;
843 sector_t esector, nr_sectors;
847 if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
848 drbd_err(device, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
849 (unsigned long long)sector, size);
853 if (!get_ldev(device))
854 return; /* no disk, no metadata, no bitmap to clear bits in */
856 nr_sectors = drbd_get_capacity(device->this_bdev);
857 esector = sector + (size >> 9) - 1;
859 if (!expect(sector < nr_sectors))
861 if (!expect(esector < nr_sectors))
862 esector = nr_sectors - 1;
864 lbnr = BM_SECT_TO_BIT(nr_sectors-1);
866 /* we clear it (in sync).
867 * round up start sector, round down end sector. we make sure we only
868 * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
869 if (unlikely(esector < BM_SECT_PER_BIT-1))
871 if (unlikely(esector == (nr_sectors-1)))
874 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
875 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
881 * ok, (capacity & 7) != 0 sometimes, but who cares...
882 * we count rs_{total,left} in bits, not sectors.
884 count = drbd_bm_clear_bits(device, sbnr, ebnr);
886 drbd_advance_rs_marks(device, drbd_bm_total_weight(device));
887 spin_lock_irqsave(&device->al_lock, flags);
888 drbd_try_clear_on_disk_bm(device, sector, count, true);
889 spin_unlock_irqrestore(&device->al_lock, flags);
891 /* just wake_up unconditional now, various lc_chaged(),
892 * lc_put() in drbd_try_clear_on_disk_bm(). */
898 wake_up(&device->al_wait);
902 * this is intended to set one request worth of data out of sync.
903 * affects at least 1 bit,
904 * and at most 1+DRBD_MAX_BIO_SIZE/BM_BLOCK_SIZE bits.
906 * called by tl_clear and drbd_send_dblock (==drbd_make_request).
907 * so this can be _any_ process.
909 int __drbd_set_out_of_sync(struct drbd_device *device, sector_t sector, int size,
910 const char *file, const unsigned int line)
912 unsigned long sbnr, ebnr, flags;
913 sector_t esector, nr_sectors;
914 unsigned int enr, count = 0;
915 struct lc_element *e;
917 /* this should be an empty REQ_FLUSH */
921 if (size < 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
922 drbd_err(device, "sector: %llus, size: %d\n",
923 (unsigned long long)sector, size);
927 if (!get_ldev(device))
928 return 0; /* no disk, no metadata, no bitmap to set bits in */
930 nr_sectors = drbd_get_capacity(device->this_bdev);
931 esector = sector + (size >> 9) - 1;
933 if (!expect(sector < nr_sectors))
935 if (!expect(esector < nr_sectors))
936 esector = nr_sectors - 1;
938 /* we set it out of sync,
939 * we do not need to round anything here */
940 sbnr = BM_SECT_TO_BIT(sector);
941 ebnr = BM_SECT_TO_BIT(esector);
943 /* ok, (capacity & 7) != 0 sometimes, but who cares...
944 * we count rs_{total,left} in bits, not sectors. */
945 spin_lock_irqsave(&device->al_lock, flags);
946 count = drbd_bm_set_bits(device, sbnr, ebnr);
948 enr = BM_SECT_TO_EXT(sector);
949 e = lc_find(device->resync, enr);
951 lc_entry(e, struct bm_extent, lce)->rs_left += count;
952 spin_unlock_irqrestore(&device->al_lock, flags);
961 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
963 struct lc_element *e;
964 struct bm_extent *bm_ext;
966 unsigned long rs_flags;
968 spin_lock_irq(&device->al_lock);
969 if (device->resync_locked > device->resync->nr_elements/2) {
970 spin_unlock_irq(&device->al_lock);
973 e = lc_get(device->resync, enr);
974 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
976 if (bm_ext->lce.lc_number != enr) {
977 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
978 bm_ext->rs_failed = 0;
979 lc_committed(device->resync);
982 if (bm_ext->lce.refcnt == 1)
983 device->resync_locked++;
984 set_bit(BME_NO_WRITES, &bm_ext->flags);
986 rs_flags = device->resync->flags;
987 spin_unlock_irq(&device->al_lock);
989 wake_up(&device->al_wait);
992 if (rs_flags & LC_STARVING)
993 drbd_warn(device, "Have to wait for element"
994 " (resync LRU too small?)\n");
995 BUG_ON(rs_flags & LC_LOCKED);
1001 static int _is_in_al(struct drbd_device *device, unsigned int enr)
1005 spin_lock_irq(&device->al_lock);
1006 rv = lc_is_used(device->act_log, enr);
1007 spin_unlock_irq(&device->al_lock);
1013 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
1014 * @device: DRBD device.
1015 * @sector: The sector number.
1017 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
1019 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
1021 unsigned int enr = BM_SECT_TO_EXT(sector);
1022 struct bm_extent *bm_ext;
1027 sig = wait_event_interruptible(device->al_wait,
1028 (bm_ext = _bme_get(device, enr)));
1032 if (test_bit(BME_LOCKED, &bm_ext->flags))
1035 /* step aside only while we are above c-min-rate; unless disabled. */
1036 sa = drbd_rs_c_min_rate_throttle(device);
1038 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1039 sig = wait_event_interruptible(device->al_wait,
1040 !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
1041 (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
1043 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
1044 spin_lock_irq(&device->al_lock);
1045 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1046 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
1047 device->resync_locked--;
1048 wake_up(&device->al_wait);
1050 spin_unlock_irq(&device->al_lock);
1053 if (schedule_timeout_interruptible(HZ/10))
1058 set_bit(BME_LOCKED, &bm_ext->flags);
1063 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1064 * @device: DRBD device.
1065 * @sector: The sector number.
1067 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1068 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1069 * if there is still application IO going on in this area.
1071 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
1073 unsigned int enr = BM_SECT_TO_EXT(sector);
1074 const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1075 struct lc_element *e;
1076 struct bm_extent *bm_ext;
1079 spin_lock_irq(&device->al_lock);
1080 if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
1081 /* in case you have very heavy scattered io, it may
1082 * stall the syncer undefined if we give up the ref count
1083 * when we try again and requeue.
1085 * if we don't give up the refcount, but the next time
1086 * we are scheduled this extent has been "synced" by new
1087 * application writes, we'd miss the lc_put on the
1088 * extent we keep the refcount on.
1089 * so we remembered which extent we had to try again, and
1090 * if the next requested one is something else, we do
1091 * the lc_put here...
1092 * we also have to wake_up
1094 e = lc_find(device->resync, device->resync_wenr);
1095 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1097 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1098 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1099 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1100 device->resync_wenr = LC_FREE;
1101 if (lc_put(device->resync, &bm_ext->lce) == 0)
1102 device->resync_locked--;
1103 wake_up(&device->al_wait);
1105 drbd_alert(device, "LOGIC BUG\n");
1109 e = lc_try_get(device->resync, enr);
1110 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1112 if (test_bit(BME_LOCKED, &bm_ext->flags))
1114 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1115 device->resync_locked++;
1117 /* we did set the BME_NO_WRITES,
1118 * but then could not set BME_LOCKED,
1119 * so we tried again.
1120 * drop the extra reference. */
1121 bm_ext->lce.refcnt--;
1122 D_ASSERT(device, bm_ext->lce.refcnt > 0);
1126 /* do we rather want to try later? */
1127 if (device->resync_locked > device->resync->nr_elements-3)
1129 /* Do or do not. There is no try. -- Yoda */
1130 e = lc_get(device->resync, enr);
1131 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1133 const unsigned long rs_flags = device->resync->flags;
1134 if (rs_flags & LC_STARVING)
1135 drbd_warn(device, "Have to wait for element"
1136 " (resync LRU too small?)\n");
1137 BUG_ON(rs_flags & LC_LOCKED);
1140 if (bm_ext->lce.lc_number != enr) {
1141 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1142 bm_ext->rs_failed = 0;
1143 lc_committed(device->resync);
1144 wake_up(&device->al_wait);
1145 D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1147 set_bit(BME_NO_WRITES, &bm_ext->flags);
1148 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1149 device->resync_locked++;
1153 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1154 if (lc_is_used(device->act_log, al_enr+i))
1157 set_bit(BME_LOCKED, &bm_ext->flags);
1159 device->resync_wenr = LC_FREE;
1160 spin_unlock_irq(&device->al_lock);
1165 device->resync_wenr = enr;
1166 spin_unlock_irq(&device->al_lock);
1170 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1172 unsigned int enr = BM_SECT_TO_EXT(sector);
1173 struct lc_element *e;
1174 struct bm_extent *bm_ext;
1175 unsigned long flags;
1177 spin_lock_irqsave(&device->al_lock, flags);
1178 e = lc_find(device->resync, enr);
1179 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1181 spin_unlock_irqrestore(&device->al_lock, flags);
1182 if (__ratelimit(&drbd_ratelimit_state))
1183 drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1187 if (bm_ext->lce.refcnt == 0) {
1188 spin_unlock_irqrestore(&device->al_lock, flags);
1189 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1190 "but refcnt is 0!?\n",
1191 (unsigned long long)sector, enr);
1195 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1196 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1197 device->resync_locked--;
1198 wake_up(&device->al_wait);
1201 spin_unlock_irqrestore(&device->al_lock, flags);
1205 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1206 * @device: DRBD device.
1208 void drbd_rs_cancel_all(struct drbd_device *device)
1210 spin_lock_irq(&device->al_lock);
1212 if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1213 lc_reset(device->resync);
1216 device->resync_locked = 0;
1217 device->resync_wenr = LC_FREE;
1218 spin_unlock_irq(&device->al_lock);
1219 wake_up(&device->al_wait);
1223 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1224 * @device: DRBD device.
1226 * Returns 0 upon success, -EAGAIN if at least one reference count was
1229 int drbd_rs_del_all(struct drbd_device *device)
1231 struct lc_element *e;
1232 struct bm_extent *bm_ext;
1235 spin_lock_irq(&device->al_lock);
1237 if (get_ldev_if_state(device, D_FAILED)) {
1238 /* ok, ->resync is there. */
1239 for (i = 0; i < device->resync->nr_elements; i++) {
1240 e = lc_element_by_index(device->resync, i);
1241 bm_ext = lc_entry(e, struct bm_extent, lce);
1242 if (bm_ext->lce.lc_number == LC_FREE)
1244 if (bm_ext->lce.lc_number == device->resync_wenr) {
1245 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1246 " got 'synced' by application io\n",
1247 device->resync_wenr);
1248 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1249 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1250 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1251 device->resync_wenr = LC_FREE;
1252 lc_put(device->resync, &bm_ext->lce);
1254 if (bm_ext->lce.refcnt != 0) {
1255 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1256 "refcnt=%d\n", bm_ext->lce.refcnt);
1258 spin_unlock_irq(&device->al_lock);
1261 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1262 D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1263 lc_del(device->resync, &bm_ext->lce);
1265 D_ASSERT(device, device->resync->used == 0);
1268 spin_unlock_irq(&device->al_lock);
1269 wake_up(&device->al_wait);
1275 * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1276 * @device: DRBD device.
1277 * @sector: The sector number.
1278 * @size: Size of failed IO operation, in byte.
1280 void drbd_rs_failed_io(struct drbd_device *device, sector_t sector, int size)
1282 /* Is called from worker and receiver context _only_ */
1283 unsigned long sbnr, ebnr, lbnr;
1284 unsigned long count;
1285 sector_t esector, nr_sectors;
1288 if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
1289 drbd_err(device, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1290 (unsigned long long)sector, size);
1293 nr_sectors = drbd_get_capacity(device->this_bdev);
1294 esector = sector + (size >> 9) - 1;
1296 if (!expect(sector < nr_sectors))
1298 if (!expect(esector < nr_sectors))
1299 esector = nr_sectors - 1;
1301 lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1304 * round up start sector, round down end sector. we make sure we only
1305 * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1306 if (unlikely(esector < BM_SECT_PER_BIT-1))
1308 if (unlikely(esector == (nr_sectors-1)))
1311 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
1312 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
1318 * ok, (capacity & 7) != 0 sometimes, but who cares...
1319 * we count rs_{total,left} in bits, not sectors.
1321 spin_lock_irq(&device->al_lock);
1322 count = drbd_bm_count_bits(device, sbnr, ebnr);
1324 device->rs_failed += count;
1326 if (get_ldev(device)) {
1327 drbd_try_clear_on_disk_bm(device, sector, count, false);
1331 /* just wake_up unconditional now, various lc_chaged(),
1332 * lc_put() in drbd_try_clear_on_disk_bm(). */
1335 spin_unlock_irq(&device->al_lock);
1337 wake_up(&device->al_wait);