1 // SPDX-License-Identifier: GPL-2.0-or-later
5 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
7 Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
8 Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
9 Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
14 #include <linux/slab.h>
15 #include <linux/crc32c.h>
16 #include <linux/drbd.h>
17 #include <linux/drbd_limits.h>
21 enum al_transaction_types {
23 AL_TR_INITIALIZED = 0xffff
25 /* all fields on disc in big endian */
26 struct __packed al_transaction_on_disk {
27 /* don't we all like magic */
30 /* to identify the most recent transaction block
31 * in the on disk ring buffer */
34 /* checksum on the full 4k block, with this field set to 0. */
37 /* type of transaction, special transaction types like:
38 * purge-all, set-all-idle, set-all-active, ... to-be-defined
39 * see also enum al_transaction_types */
40 __be16 transaction_type;
42 /* we currently allow only a few thousand extents,
43 * so 16bit will be enough for the slot number. */
45 /* how many updates in this transaction */
48 /* maximum slot number, "al-extents" in drbd.conf speak.
49 * Having this in each transaction should make reconfiguration
50 * of that parameter easier. */
53 /* slot number the context starts with */
54 __be16 context_start_slot_nr;
56 /* Some reserved bytes. Expected usage is a 64bit counter of
57 * sectors-written since device creation, and other data generation tag
61 /* --- 36 byte used --- */
63 /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
64 * in one transaction, then use the remaining byte in the 4k block for
65 * context information. "Flexible" number of updates per transaction
66 * does not help, as we have to account for the case when all update
67 * slots are used anyways, so it would only complicate code without
70 __be16 update_slot_nr[AL_UPDATES_PER_TRANSACTION];
72 /* but the extent number is 32bit, which at an extent size of 4 MiB
73 * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
74 __be32 update_extent_nr[AL_UPDATES_PER_TRANSACTION];
76 /* --- 420 bytes used (36 + 64*6) --- */
78 /* 4096 - 420 = 3676 = 919 * 4 */
79 __be32 context[AL_CONTEXT_PER_TRANSACTION];
82 void *drbd_md_get_buffer(struct drbd_device *device, const char *intent)
86 wait_event(device->misc_wait,
87 (r = atomic_cmpxchg(&device->md_io.in_use, 0, 1)) == 0 ||
88 device->state.disk <= D_FAILED);
93 device->md_io.current_use = intent;
94 device->md_io.start_jif = jiffies;
95 device->md_io.submit_jif = device->md_io.start_jif - 1;
96 return page_address(device->md_io.page);
99 void drbd_md_put_buffer(struct drbd_device *device)
101 if (atomic_dec_and_test(&device->md_io.in_use))
102 wake_up(&device->misc_wait);
105 void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
111 dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
115 dt = MAX_SCHEDULE_TIMEOUT;
117 dt = wait_event_timeout(device->misc_wait,
118 *done || test_bit(FORCE_DETACH, &device->flags), dt);
120 drbd_err(device, "meta-data IO operation timed out\n");
121 drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
125 static int _drbd_md_sync_page_io(struct drbd_device *device,
126 struct drbd_backing_dev *bdev,
127 sector_t sector, int op)
130 /* we do all our meta data IO in aligned 4k blocks. */
131 const int size = 4096;
132 int err, op_flags = 0;
134 device->md_io.done = 0;
135 device->md_io.error = -ENODEV;
137 if ((op == REQ_OP_WRITE) && !test_bit(MD_NO_FUA, &device->flags))
138 op_flags |= REQ_FUA | REQ_PREFLUSH;
139 op_flags |= REQ_SYNC;
141 bio = bio_alloc_bioset(bdev->md_bdev, 1, op | op_flags, GFP_NOIO,
142 &drbd_md_io_bio_set);
143 bio->bi_iter.bi_sector = sector;
145 if (bio_add_page(bio, device->md_io.page, size, 0) != size)
147 bio->bi_private = device;
148 bio->bi_end_io = drbd_md_endio;
150 if (op != REQ_OP_WRITE && device->state.disk == D_DISKLESS && device->ldev == NULL)
151 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
153 else if (!get_ldev_if_state(device, D_ATTACHING)) {
154 /* Corresponding put_ldev in drbd_md_endio() */
155 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
160 bio_get(bio); /* one bio_put() is in the completion handler */
161 atomic_inc(&device->md_io.in_use); /* drbd_md_put_buffer() is in the completion handler */
162 device->md_io.submit_jif = jiffies;
163 if (drbd_insert_fault(device, (op == REQ_OP_WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
167 wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
169 err = device->md_io.error;
176 int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
177 sector_t sector, int op)
180 D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
182 BUG_ON(!bdev->md_bdev);
184 dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
185 current->comm, current->pid, __func__,
186 (unsigned long long)sector, (op == REQ_OP_WRITE) ? "WRITE" : "READ",
189 if (sector < drbd_md_first_sector(bdev) ||
190 sector + 7 > drbd_md_last_sector(bdev))
191 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
192 current->comm, current->pid, __func__,
193 (unsigned long long)sector,
194 (op == REQ_OP_WRITE) ? "WRITE" : "READ");
196 err = _drbd_md_sync_page_io(device, bdev, sector, op);
198 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
199 (unsigned long long)sector,
200 (op == REQ_OP_WRITE) ? "WRITE" : "READ", err);
205 static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
207 struct lc_element *tmp;
208 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
209 if (unlikely(tmp != NULL)) {
210 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
211 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
217 static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
219 struct lc_element *al_ext;
220 struct bm_extent *bm_ext;
223 spin_lock_irq(&device->al_lock);
224 bm_ext = find_active_resync_extent(device, enr);
226 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
227 spin_unlock_irq(&device->al_lock);
229 wake_up(&device->al_wait);
233 al_ext = lc_try_get(device->act_log, enr);
235 al_ext = lc_get(device->act_log, enr);
236 spin_unlock_irq(&device->al_lock);
240 bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
242 /* for bios crossing activity log extent boundaries,
243 * we may need to activate two extents in one go */
244 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
245 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
247 D_ASSERT(device, first <= last);
248 D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
250 /* FIXME figure out a fast path for bios crossing AL extent boundaries */
254 return _al_get(device, first, true);
257 bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
259 /* for bios crossing activity log extent boundaries,
260 * we may need to activate two extents in one go */
261 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
262 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
264 bool need_transaction = false;
266 D_ASSERT(device, first <= last);
267 D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
269 for (enr = first; enr <= last; enr++) {
270 struct lc_element *al_ext;
271 wait_event(device->al_wait,
272 (al_ext = _al_get(device, enr, false)) != NULL);
273 if (al_ext->lc_number != enr)
274 need_transaction = true;
276 return need_transaction;
279 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
280 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
281 * are still coupled, or assume too much about their relation.
282 * Code below will not work if this is violated.
283 * Will be cleaned up with some followup patch.
288 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
293 /* al extent number to bit */
294 (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
297 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
299 const unsigned int stripes = device->ldev->md.al_stripes;
300 const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
302 /* transaction number, modulo on-disk ring buffer wrap around */
303 unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
305 /* ... to aligned 4k on disk block */
306 t = ((t % stripes) * stripe_size_4kB) + t/stripes;
308 /* ... to 512 byte sector in activity log */
311 /* ... plus offset to the on disk position */
312 return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
315 static int __al_write_transaction(struct drbd_device *device, struct al_transaction_on_disk *buffer)
317 struct lc_element *e;
324 memset(buffer, 0, sizeof(*buffer));
325 buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
326 buffer->tr_number = cpu_to_be32(device->al_tr_number);
330 drbd_bm_reset_al_hints(device);
332 /* Even though no one can start to change this list
333 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
334 * lc_try_lock_for_transaction() --, someone may still
335 * be in the process of changing it. */
336 spin_lock_irq(&device->al_lock);
337 list_for_each_entry(e, &device->act_log->to_be_changed, list) {
338 if (i == AL_UPDATES_PER_TRANSACTION) {
342 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
343 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
344 if (e->lc_number != LC_FREE)
345 drbd_bm_mark_for_writeout(device,
346 al_extent_to_bm_page(e->lc_number));
349 spin_unlock_irq(&device->al_lock);
350 BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
352 buffer->n_updates = cpu_to_be16(i);
353 for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
354 buffer->update_slot_nr[i] = cpu_to_be16(-1);
355 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
358 buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
359 buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
361 mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
362 device->act_log->nr_elements - device->al_tr_cycle);
363 for (i = 0; i < mx; i++) {
364 unsigned idx = device->al_tr_cycle + i;
365 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
366 buffer->context[i] = cpu_to_be32(extent_nr);
368 for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
369 buffer->context[i] = cpu_to_be32(LC_FREE);
371 device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
372 if (device->al_tr_cycle >= device->act_log->nr_elements)
373 device->al_tr_cycle = 0;
375 sector = al_tr_number_to_on_disk_sector(device);
377 crc = crc32c(0, buffer, 4096);
378 buffer->crc32c = cpu_to_be32(crc);
380 if (drbd_bm_write_hinted(device))
383 bool write_al_updates;
385 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
387 if (write_al_updates) {
388 if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
390 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
392 device->al_tr_number++;
393 device->al_writ_cnt++;
401 static int al_write_transaction(struct drbd_device *device)
403 struct al_transaction_on_disk *buffer;
406 if (!get_ldev(device)) {
407 drbd_err(device, "disk is %s, cannot start al transaction\n",
408 drbd_disk_str(device->state.disk));
412 /* The bitmap write may have failed, causing a state change. */
413 if (device->state.disk < D_INCONSISTENT) {
415 "disk is %s, cannot write al transaction\n",
416 drbd_disk_str(device->state.disk));
421 /* protects md_io_buffer, al_tr_cycle, ... */
422 buffer = drbd_md_get_buffer(device, __func__);
424 drbd_err(device, "disk failed while waiting for md_io buffer\n");
429 err = __al_write_transaction(device, buffer);
431 drbd_md_put_buffer(device);
438 void drbd_al_begin_io_commit(struct drbd_device *device)
442 /* Serialize multiple transactions.
443 * This uses test_and_set_bit, memory barrier is implicit.
445 wait_event(device->al_wait,
446 device->act_log->pending_changes == 0 ||
447 (locked = lc_try_lock_for_transaction(device->act_log)));
450 /* Double check: it may have been committed by someone else,
451 * while we have been waiting for the lock. */
452 if (device->act_log->pending_changes) {
453 bool write_al_updates;
456 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
459 if (write_al_updates)
460 al_write_transaction(device);
461 spin_lock_irq(&device->al_lock);
464 we need an "lc_cancel" here;
466 lc_committed(device->act_log);
467 spin_unlock_irq(&device->al_lock);
469 lc_unlock(device->act_log);
470 wake_up(&device->al_wait);
475 * @delegate: delegate activity log I/O to the worker thread
477 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
479 if (drbd_al_begin_io_prepare(device, i))
480 drbd_al_begin_io_commit(device);
483 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
485 struct lru_cache *al = device->act_log;
486 /* for bios crossing activity log extent boundaries,
487 * we may need to activate two extents in one go */
488 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
489 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
490 unsigned nr_al_extents;
491 unsigned available_update_slots;
494 D_ASSERT(device, first <= last);
496 nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
497 available_update_slots = min(al->nr_elements - al->used,
498 al->max_pending_changes - al->pending_changes);
500 /* We want all necessary updates for a given request within the same transaction
501 * We could first check how many updates are *actually* needed,
502 * and use that instead of the worst-case nr_al_extents */
503 if (available_update_slots < nr_al_extents) {
504 /* Too many activity log extents are currently "hot".
506 * If we have accumulated pending changes already,
509 * If we cannot get even a single pending change through,
510 * stop the fast path until we made some progress,
511 * or requests to "cold" extents could be starved. */
512 if (!al->pending_changes)
513 __set_bit(__LC_STARVING, &device->act_log->flags);
517 /* Is resync active in this area? */
518 for (enr = first; enr <= last; enr++) {
519 struct lc_element *tmp;
520 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
521 if (unlikely(tmp != NULL)) {
522 struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
523 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
524 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
531 /* Checkout the refcounts.
532 * Given that we checked for available elements and update slots above,
533 * this has to be successful. */
534 for (enr = first; enr <= last; enr++) {
535 struct lc_element *al_ext;
536 al_ext = lc_get_cumulative(device->act_log, enr);
538 drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
543 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
545 /* for bios crossing activity log extent boundaries,
546 * we may need to activate two extents in one go */
547 unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
548 unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
550 struct lc_element *extent;
553 D_ASSERT(device, first <= last);
554 spin_lock_irqsave(&device->al_lock, flags);
556 for (enr = first; enr <= last; enr++) {
557 extent = lc_find(device->act_log, enr);
559 drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
562 lc_put(device->act_log, extent);
564 spin_unlock_irqrestore(&device->al_lock, flags);
565 wake_up(&device->al_wait);
568 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
572 spin_lock_irq(&device->al_lock);
573 rv = (al_ext->refcnt == 0);
575 lc_del(device->act_log, al_ext);
576 spin_unlock_irq(&device->al_lock);
582 * drbd_al_shrink() - Removes all active extents form the activity log
583 * @device: DRBD device.
585 * Removes all active extents form the activity log, waiting until
586 * the reference count of each entry dropped to 0 first, of course.
588 * You need to lock device->act_log with lc_try_lock() / lc_unlock()
590 void drbd_al_shrink(struct drbd_device *device)
592 struct lc_element *al_ext;
595 D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
597 for (i = 0; i < device->act_log->nr_elements; i++) {
598 al_ext = lc_element_by_index(device->act_log, i);
599 if (al_ext->lc_number == LC_FREE)
601 wait_event(device->al_wait, _try_lc_del(device, al_ext));
604 wake_up(&device->al_wait);
607 int drbd_al_initialize(struct drbd_device *device, void *buffer)
609 struct al_transaction_on_disk *al = buffer;
610 struct drbd_md *md = &device->ldev->md;
611 int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
614 __al_write_transaction(device, al);
615 /* There may or may not have been a pending transaction. */
616 spin_lock_irq(&device->al_lock);
617 lc_committed(device->act_log);
618 spin_unlock_irq(&device->al_lock);
620 /* The rest of the transactions will have an empty "updates" list, and
621 * are written out only to provide the context, and to initialize the
622 * on-disk ring buffer. */
623 for (i = 1; i < al_size_4k; i++) {
624 int err = __al_write_transaction(device, al);
631 static const char *drbd_change_sync_fname[] = {
632 [RECORD_RS_FAILED] = "drbd_rs_failed_io",
633 [SET_IN_SYNC] = "drbd_set_in_sync",
634 [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
637 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
638 * resync LRU-cache are 16MB each.
639 * The caller of this function has to hold an get_ldev() reference.
641 * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
642 * potentially pulling in (and recounting the corresponding bits)
643 * this resync extent into the resync extent lru cache.
645 * Returns whether all bits have been cleared for this resync extent,
646 * precisely: (rs_left <= rs_failed)
648 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
650 static bool update_rs_extent(struct drbd_device *device,
651 unsigned int enr, int count,
652 enum update_sync_bits_mode mode)
654 struct lc_element *e;
656 D_ASSERT(device, atomic_read(&device->local_cnt));
658 /* When setting out-of-sync bits,
659 * we don't need it cached (lc_find).
660 * But if it is present in the cache,
661 * we should update the cached bit count.
662 * Otherwise, that extent should be in the resync extent lru cache
663 * already -- or we want to pull it in if necessary -- (lc_get),
664 * then update and check rs_left and rs_failed. */
665 if (mode == SET_OUT_OF_SYNC)
666 e = lc_find(device->resync, enr);
668 e = lc_get(device->resync, enr);
670 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
671 if (ext->lce.lc_number == enr) {
672 if (mode == SET_IN_SYNC)
673 ext->rs_left -= count;
674 else if (mode == SET_OUT_OF_SYNC)
675 ext->rs_left += count;
677 ext->rs_failed += count;
678 if (ext->rs_left < ext->rs_failed) {
679 drbd_warn(device, "BAD! enr=%u rs_left=%d "
680 "rs_failed=%d count=%d cstate=%s\n",
681 ext->lce.lc_number, ext->rs_left,
682 ext->rs_failed, count,
683 drbd_conn_str(device->state.conn));
685 /* We don't expect to be able to clear more bits
686 * than have been set when we originally counted
687 * the set bits to cache that value in ext->rs_left.
688 * Whatever the reason (disconnect during resync,
689 * delayed local completion of an application write),
690 * try to fix it up by recounting here. */
691 ext->rs_left = drbd_bm_e_weight(device, enr);
694 /* Normally this element should be in the cache,
695 * since drbd_rs_begin_io() pulled it already in.
697 * But maybe an application write finished, and we set
698 * something outside the resync lru_cache in sync.
700 int rs_left = drbd_bm_e_weight(device, enr);
701 if (ext->flags != 0) {
702 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
704 ext->lce.lc_number, ext->rs_left,
705 ext->flags, enr, rs_left);
708 if (ext->rs_failed) {
709 drbd_warn(device, "Kicking resync_lru element enr=%u "
710 "out with rs_failed=%d\n",
711 ext->lce.lc_number, ext->rs_failed);
713 ext->rs_left = rs_left;
714 ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
715 /* we don't keep a persistent log of the resync lru,
716 * we can commit any change right away. */
717 lc_committed(device->resync);
719 if (mode != SET_OUT_OF_SYNC)
720 lc_put(device->resync, &ext->lce);
721 /* no race, we are within the al_lock! */
723 if (ext->rs_left <= ext->rs_failed) {
727 } else if (mode != SET_OUT_OF_SYNC) {
728 /* be quiet if lc_find() did not find it. */
729 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
730 device->resync_locked,
731 device->resync->nr_elements,
732 device->resync->flags);
737 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
739 unsigned long now = jiffies;
740 unsigned long last = device->rs_mark_time[device->rs_last_mark];
741 int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
742 if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
743 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
744 device->state.conn != C_PAUSED_SYNC_T &&
745 device->state.conn != C_PAUSED_SYNC_S) {
746 device->rs_mark_time[next] = now;
747 device->rs_mark_left[next] = still_to_go;
748 device->rs_last_mark = next;
753 /* It is called lazy update, so don't do write-out too often. */
754 static bool lazy_bitmap_update_due(struct drbd_device *device)
756 return time_after(jiffies, device->rs_last_bcast + 2*HZ);
759 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
762 struct drbd_connection *connection = first_peer_device(device)->connection;
763 if (connection->agreed_pro_version <= 95 ||
764 is_sync_target_state(device->state.conn))
765 set_bit(RS_DONE, &device->flags);
766 /* and also set RS_PROGRESS below */
768 /* Else: rather wait for explicit notification via receive_state,
769 * to avoid uuids-rotated-too-fast causing full resync
770 * in next handshake, in case the replication link breaks
771 * at the most unfortunate time... */
772 } else if (!lazy_bitmap_update_due(device))
775 drbd_device_post_work(device, RS_PROGRESS);
778 static int update_sync_bits(struct drbd_device *device,
779 unsigned long sbnr, unsigned long ebnr,
780 enum update_sync_bits_mode mode)
783 * We keep a count of set bits per resync-extent in the ->rs_left
784 * caching member, so we need to loop and work within the resync extent
785 * alignment. Typically this loop will execute exactly once.
788 unsigned long count = 0;
789 unsigned int cleared = 0;
790 while (sbnr <= ebnr) {
791 /* set temporary boundary bit number to last bit number within
792 * the resync extent of the current start bit number,
793 * but cap at provided end bit number */
794 unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
797 if (mode == RECORD_RS_FAILED)
798 /* Only called from drbd_rs_failed_io(), bits
799 * supposedly still set. Recount, maybe some
800 * of the bits have been successfully cleared
801 * by application IO meanwhile.
803 c = drbd_bm_count_bits(device, sbnr, tbnr);
804 else if (mode == SET_IN_SYNC)
805 c = drbd_bm_clear_bits(device, sbnr, tbnr);
806 else /* if (mode == SET_OUT_OF_SYNC) */
807 c = drbd_bm_set_bits(device, sbnr, tbnr);
810 spin_lock_irqsave(&device->al_lock, flags);
811 cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
812 spin_unlock_irqrestore(&device->al_lock, flags);
818 if (mode == SET_IN_SYNC) {
819 unsigned long still_to_go = drbd_bm_total_weight(device);
820 bool rs_is_done = (still_to_go <= device->rs_failed);
821 drbd_advance_rs_marks(device, still_to_go);
822 if (cleared || rs_is_done)
823 maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
824 } else if (mode == RECORD_RS_FAILED)
825 device->rs_failed += count;
826 wake_up(&device->al_wait);
831 static bool plausible_request_size(int size)
834 && size <= DRBD_MAX_BATCH_BIO_SIZE
835 && IS_ALIGNED(size, 512);
838 /* clear the bit corresponding to the piece of storage in question:
839 * size byte of data starting from sector. Only clear a bits of the affected
840 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
842 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
845 int __drbd_change_sync(struct drbd_device *device, sector_t sector, int size,
846 enum update_sync_bits_mode mode)
848 /* Is called from worker and receiver context _only_ */
849 unsigned long sbnr, ebnr, lbnr;
850 unsigned long count = 0;
851 sector_t esector, nr_sectors;
853 /* This would be an empty REQ_PREFLUSH, be silent. */
854 if ((mode == SET_OUT_OF_SYNC) && size == 0)
857 if (!plausible_request_size(size)) {
858 drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
859 drbd_change_sync_fname[mode],
860 (unsigned long long)sector, size);
864 if (!get_ldev(device))
865 return 0; /* no disk, no metadata, no bitmap to manipulate bits in */
867 nr_sectors = get_capacity(device->vdisk);
868 esector = sector + (size >> 9) - 1;
870 if (!expect(sector < nr_sectors))
872 if (!expect(esector < nr_sectors))
873 esector = nr_sectors - 1;
875 lbnr = BM_SECT_TO_BIT(nr_sectors-1);
877 if (mode == SET_IN_SYNC) {
878 /* Round up start sector, round down end sector. We make sure
879 * we only clear full, aligned, BM_BLOCK_SIZE blocks. */
880 if (unlikely(esector < BM_SECT_PER_BIT-1))
882 if (unlikely(esector == (nr_sectors-1)))
885 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
886 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
888 /* We set it out of sync, or record resync failure.
889 * Should not round anything here. */
890 sbnr = BM_SECT_TO_BIT(sector);
891 ebnr = BM_SECT_TO_BIT(esector);
894 count = update_sync_bits(device, sbnr, ebnr, mode);
901 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
903 struct lc_element *e;
904 struct bm_extent *bm_ext;
906 unsigned long rs_flags;
908 spin_lock_irq(&device->al_lock);
909 if (device->resync_locked > device->resync->nr_elements/2) {
910 spin_unlock_irq(&device->al_lock);
913 e = lc_get(device->resync, enr);
914 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
916 if (bm_ext->lce.lc_number != enr) {
917 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
918 bm_ext->rs_failed = 0;
919 lc_committed(device->resync);
922 if (bm_ext->lce.refcnt == 1)
923 device->resync_locked++;
924 set_bit(BME_NO_WRITES, &bm_ext->flags);
926 rs_flags = device->resync->flags;
927 spin_unlock_irq(&device->al_lock);
929 wake_up(&device->al_wait);
932 if (rs_flags & LC_STARVING)
933 drbd_warn(device, "Have to wait for element"
934 " (resync LRU too small?)\n");
935 BUG_ON(rs_flags & LC_LOCKED);
941 static int _is_in_al(struct drbd_device *device, unsigned int enr)
945 spin_lock_irq(&device->al_lock);
946 rv = lc_is_used(device->act_log, enr);
947 spin_unlock_irq(&device->al_lock);
953 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
954 * @device: DRBD device.
955 * @sector: The sector number.
957 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
959 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
961 unsigned int enr = BM_SECT_TO_EXT(sector);
962 struct bm_extent *bm_ext;
967 sig = wait_event_interruptible(device->al_wait,
968 (bm_ext = _bme_get(device, enr)));
972 if (test_bit(BME_LOCKED, &bm_ext->flags))
975 /* step aside only while we are above c-min-rate; unless disabled. */
976 sa = drbd_rs_c_min_rate_throttle(device);
978 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
979 sig = wait_event_interruptible(device->al_wait,
980 !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
981 (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
983 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
984 spin_lock_irq(&device->al_lock);
985 if (lc_put(device->resync, &bm_ext->lce) == 0) {
986 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
987 device->resync_locked--;
988 wake_up(&device->al_wait);
990 spin_unlock_irq(&device->al_lock);
993 if (schedule_timeout_interruptible(HZ/10))
998 set_bit(BME_LOCKED, &bm_ext->flags);
1003 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1004 * @device: DRBD device.
1005 * @sector: The sector number.
1007 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1008 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1009 * if there is still application IO going on in this area.
1011 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
1013 unsigned int enr = BM_SECT_TO_EXT(sector);
1014 const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1015 struct lc_element *e;
1016 struct bm_extent *bm_ext;
1018 bool throttle = drbd_rs_should_slow_down(device, sector, true);
1020 /* If we need to throttle, a half-locked (only marked BME_NO_WRITES,
1021 * not yet BME_LOCKED) extent needs to be kicked out explicitly if we
1022 * need to throttle. There is at most one such half-locked extent,
1023 * which is remembered in resync_wenr. */
1025 if (throttle && device->resync_wenr != enr)
1028 spin_lock_irq(&device->al_lock);
1029 if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
1030 /* in case you have very heavy scattered io, it may
1031 * stall the syncer undefined if we give up the ref count
1032 * when we try again and requeue.
1034 * if we don't give up the refcount, but the next time
1035 * we are scheduled this extent has been "synced" by new
1036 * application writes, we'd miss the lc_put on the
1037 * extent we keep the refcount on.
1038 * so we remembered which extent we had to try again, and
1039 * if the next requested one is something else, we do
1040 * the lc_put here...
1041 * we also have to wake_up
1043 e = lc_find(device->resync, device->resync_wenr);
1044 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1046 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1047 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1048 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1049 device->resync_wenr = LC_FREE;
1050 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1052 device->resync_locked--;
1054 wake_up(&device->al_wait);
1056 drbd_alert(device, "LOGIC BUG\n");
1060 e = lc_try_get(device->resync, enr);
1061 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1063 if (test_bit(BME_LOCKED, &bm_ext->flags))
1065 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1066 device->resync_locked++;
1068 /* we did set the BME_NO_WRITES,
1069 * but then could not set BME_LOCKED,
1070 * so we tried again.
1071 * drop the extra reference. */
1072 bm_ext->lce.refcnt--;
1073 D_ASSERT(device, bm_ext->lce.refcnt > 0);
1077 /* do we rather want to try later? */
1078 if (device->resync_locked > device->resync->nr_elements-3)
1080 /* Do or do not. There is no try. -- Yoda */
1081 e = lc_get(device->resync, enr);
1082 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1084 const unsigned long rs_flags = device->resync->flags;
1085 if (rs_flags & LC_STARVING)
1086 drbd_warn(device, "Have to wait for element"
1087 " (resync LRU too small?)\n");
1088 BUG_ON(rs_flags & LC_LOCKED);
1091 if (bm_ext->lce.lc_number != enr) {
1092 bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1093 bm_ext->rs_failed = 0;
1094 lc_committed(device->resync);
1095 wake_up(&device->al_wait);
1096 D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1098 set_bit(BME_NO_WRITES, &bm_ext->flags);
1099 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1100 device->resync_locked++;
1104 for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1105 if (lc_is_used(device->act_log, al_enr+i))
1108 set_bit(BME_LOCKED, &bm_ext->flags);
1110 device->resync_wenr = LC_FREE;
1111 spin_unlock_irq(&device->al_lock);
1117 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1118 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1119 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1120 device->resync_wenr = LC_FREE;
1121 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1123 device->resync_locked--;
1125 wake_up(&device->al_wait);
1127 device->resync_wenr = enr;
1129 spin_unlock_irq(&device->al_lock);
1133 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1135 unsigned int enr = BM_SECT_TO_EXT(sector);
1136 struct lc_element *e;
1137 struct bm_extent *bm_ext;
1138 unsigned long flags;
1140 spin_lock_irqsave(&device->al_lock, flags);
1141 e = lc_find(device->resync, enr);
1142 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1144 spin_unlock_irqrestore(&device->al_lock, flags);
1145 if (__ratelimit(&drbd_ratelimit_state))
1146 drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1150 if (bm_ext->lce.refcnt == 0) {
1151 spin_unlock_irqrestore(&device->al_lock, flags);
1152 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1153 "but refcnt is 0!?\n",
1154 (unsigned long long)sector, enr);
1158 if (lc_put(device->resync, &bm_ext->lce) == 0) {
1159 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1160 device->resync_locked--;
1161 wake_up(&device->al_wait);
1164 spin_unlock_irqrestore(&device->al_lock, flags);
1168 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1169 * @device: DRBD device.
1171 void drbd_rs_cancel_all(struct drbd_device *device)
1173 spin_lock_irq(&device->al_lock);
1175 if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1176 lc_reset(device->resync);
1179 device->resync_locked = 0;
1180 device->resync_wenr = LC_FREE;
1181 spin_unlock_irq(&device->al_lock);
1182 wake_up(&device->al_wait);
1186 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1187 * @device: DRBD device.
1189 * Returns 0 upon success, -EAGAIN if at least one reference count was
1192 int drbd_rs_del_all(struct drbd_device *device)
1194 struct lc_element *e;
1195 struct bm_extent *bm_ext;
1198 spin_lock_irq(&device->al_lock);
1200 if (get_ldev_if_state(device, D_FAILED)) {
1201 /* ok, ->resync is there. */
1202 for (i = 0; i < device->resync->nr_elements; i++) {
1203 e = lc_element_by_index(device->resync, i);
1204 bm_ext = lc_entry(e, struct bm_extent, lce);
1205 if (bm_ext->lce.lc_number == LC_FREE)
1207 if (bm_ext->lce.lc_number == device->resync_wenr) {
1208 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1209 " got 'synced' by application io\n",
1210 device->resync_wenr);
1211 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1212 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1213 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1214 device->resync_wenr = LC_FREE;
1215 lc_put(device->resync, &bm_ext->lce);
1217 if (bm_ext->lce.refcnt != 0) {
1218 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1219 "refcnt=%d\n", bm_ext->lce.refcnt);
1221 spin_unlock_irq(&device->al_lock);
1224 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1225 D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1226 lc_del(device->resync, &bm_ext->lce);
1228 D_ASSERT(device, device->resync->used == 0);
1231 spin_unlock_irq(&device->al_lock);
1232 wake_up(&device->al_wait);