1 // SPDX-License-Identifier: GPL-2.0
3 * bcachefs journalling code, for btree insertions
5 * Copyright 2012 Google, Inc.
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
12 #include "btree_update.h"
16 #include "journal_io.h"
17 #include "journal_reclaim.h"
18 #include "journal_sb.h"
19 #include "journal_seq_blacklist.h"
22 static const char * const bch2_journal_errors[] = {
29 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
31 return seq > j->seq_ondisk;
34 static bool __journal_entry_is_open(union journal_res_state state)
36 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
39 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
41 return atomic64_read(&j->seq) - j->seq_ondisk;
44 static bool journal_entry_is_open(struct journal *j)
46 return __journal_entry_is_open(j->reservations);
49 static inline struct journal_buf *
50 journal_seq_to_buf(struct journal *j, u64 seq)
52 struct journal_buf *buf = NULL;
54 EBUG_ON(seq > journal_cur_seq(j));
56 if (journal_seq_unwritten(j, seq)) {
57 buf = j->buf + (seq & JOURNAL_BUF_MASK);
58 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
63 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
67 for (i = 0; i < ARRAY_SIZE(p->list); i++)
68 INIT_LIST_HEAD(&p->list[i]);
69 INIT_LIST_HEAD(&p->flushed);
70 atomic_set(&p->count, count);
75 * Detect stuck journal conditions and trigger shutdown. Technically the journal
76 * can end up stuck for a variety of reasons, such as a blocked I/O, journal
77 * reservation lockup, etc. Since this is a fatal error with potentially
78 * unpredictable characteristics, we want to be fairly conservative before we
79 * decide to shut things down.
81 * Consider the journal stuck when it appears full with no ability to commit
82 * btree transactions, to discard journal buckets, nor acquire priority
83 * (reserved watermark) reservation.
86 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
88 struct bch_fs *c = container_of(j, struct bch_fs, journal);
90 struct printbuf buf = PRINTBUF;
92 if (!(error == JOURNAL_ERR_journal_full ||
93 error == JOURNAL_ERR_journal_pin_full) ||
94 nr_unwritten_journal_entries(j) ||
95 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
100 if (j->can_discard) {
101 spin_unlock(&j->lock);
108 * The journal shutdown path will set ->err_seq, but do it here first to
109 * serialize against concurrent failures and avoid duplicate error
113 spin_unlock(&j->lock);
116 j->err_seq = journal_cur_seq(j);
117 spin_unlock(&j->lock);
119 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
120 bch2_journal_errors[error]);
121 bch2_journal_debug_to_text(&buf, j);
122 bch_err(c, "%s", buf.buf);
124 printbuf_reset(&buf);
125 bch2_journal_pins_to_text(&buf, j);
126 bch_err(c, "Journal pins:\n%s", buf.buf);
136 * Final processing when the last reference of a journal buffer has been
137 * dropped. Drop the pin list reference acquired at journal entry open and write
138 * the buffer, if requested.
140 void bch2_journal_buf_put_final(struct journal *j, u64 seq, bool write)
142 struct bch_fs *c = container_of(j, struct bch_fs, journal);
144 lockdep_assert_held(&j->lock);
146 if (__bch2_journal_pin_put(j, seq))
147 bch2_journal_reclaim_fast(j);
149 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
153 * Returns true if journal entry is now closed:
155 * We don't close a journal_buf until the next journal_buf is finished writing,
156 * and can be opened again - this also initializes the next journal_buf:
158 static void __journal_entry_close(struct journal *j, unsigned closed_val)
160 struct bch_fs *c = container_of(j, struct bch_fs, journal);
161 struct journal_buf *buf = journal_cur_buf(j);
162 union journal_res_state old, new;
163 u64 v = atomic64_read(&j->reservations.counter);
166 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
167 closed_val != JOURNAL_ENTRY_ERROR_VAL);
169 lockdep_assert_held(&j->lock);
173 new.cur_entry_offset = closed_val;
175 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
176 old.cur_entry_offset == new.cur_entry_offset)
178 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
179 old.v, new.v)) != old.v);
181 if (!__journal_entry_is_open(old))
184 /* Close out old buffer: */
185 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
187 trace_journal_entry_close(c, vstruct_bytes(buf->data));
189 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
190 buf->u64s_reserved) << c->block_bits;
191 BUG_ON(sectors > buf->sectors);
192 buf->sectors = sectors;
195 * We have to set last_seq here, _before_ opening a new journal entry:
197 * A threads may replace an old pin with a new pin on their current
198 * journal reservation - the expectation being that the journal will
199 * contain either what the old pin protected or what the new pin
202 * After the old pin is dropped journal_last_seq() won't include the old
203 * pin, so we can only write the updated last_seq on the entry that
204 * contains whatever the new pin protects.
206 * Restated, we can _not_ update last_seq for a given entry if there
207 * could be a newer entry open with reservations/pins that have been
210 * Hence, we want update/set last_seq on the current journal entry right
211 * before we open a new one:
213 buf->last_seq = journal_last_seq(j);
214 buf->data->last_seq = cpu_to_le64(buf->last_seq);
215 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
217 cancel_delayed_work(&j->write_work);
219 bch2_journal_space_available(j);
221 __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
224 void bch2_journal_halt(struct journal *j)
227 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL);
229 j->err_seq = journal_cur_seq(j);
231 spin_unlock(&j->lock);
234 static bool journal_entry_want_write(struct journal *j)
236 bool ret = !journal_entry_is_open(j) ||
237 journal_cur_seq(j) == journal_last_unwritten_seq(j);
239 /* Don't close it yet if we already have a write in flight: */
241 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
242 else if (nr_unwritten_journal_entries(j)) {
243 struct journal_buf *buf = journal_cur_buf(j);
245 if (!buf->flush_time) {
246 buf->flush_time = local_clock() ?: 1;
247 buf->expires = jiffies;
254 static bool journal_entry_close(struct journal *j)
259 ret = journal_entry_want_write(j);
260 spin_unlock(&j->lock);
266 * should _only_ called from journal_res_get() - when we actually want a
267 * journal reservation - journal entry is open means journal is dirty:
269 static int journal_entry_open(struct journal *j)
271 struct bch_fs *c = container_of(j, struct bch_fs, journal);
272 struct journal_buf *buf = j->buf +
273 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
274 union journal_res_state old, new;
278 lockdep_assert_held(&j->lock);
279 BUG_ON(journal_entry_is_open(j));
280 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
283 return JOURNAL_ERR_blocked;
285 if (j->cur_entry_error)
286 return j->cur_entry_error;
288 if (bch2_journal_error(j))
289 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
291 if (!fifo_free(&j->pin))
292 return JOURNAL_ERR_journal_pin_full;
294 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
295 return JOURNAL_ERR_max_in_flight;
297 BUG_ON(!j->cur_entry_sectors);
300 (journal_cur_seq(j) == j->flushed_seq_ondisk
302 : j->last_flush_write) +
303 msecs_to_jiffies(c->opts.journal_flush_delay);
305 buf->u64s_reserved = j->entry_u64s_reserved;
306 buf->disk_sectors = j->cur_entry_sectors;
307 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
309 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
310 journal_entry_overhead(j);
311 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
313 if (u64s <= (ssize_t) j->early_journal_entries.nr)
314 return JOURNAL_ERR_journal_full;
316 if (fifo_empty(&j->pin) && j->reclaim_thread)
317 wake_up_process(j->reclaim_thread);
320 * The fifo_push() needs to happen at the same time as j->seq is
321 * incremented for journal_last_seq() to be calculated correctly
323 atomic64_inc(&j->seq);
324 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
326 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
328 bkey_extent_init(&buf->key);
329 buf->noflush = false;
330 buf->must_flush = false;
331 buf->separate_flush = false;
334 memset(buf->data, 0, sizeof(*buf->data));
335 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
338 if (j->early_journal_entries.nr) {
339 memcpy(buf->data->_data, j->early_journal_entries.data,
340 j->early_journal_entries.nr * sizeof(u64));
341 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
345 * Must be set before marking the journal entry as open:
347 j->cur_entry_u64s = u64s;
349 v = atomic64_read(&j->reservations.counter);
353 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
356 BUG_ON(journal_state_count(new, new.idx));
357 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
359 journal_state_inc(&new);
361 /* Handle any already added entries */
362 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
363 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
364 old.v, new.v)) != old.v);
366 mod_delayed_work(c->io_complete_wq,
368 msecs_to_jiffies(c->opts.journal_flush_delay));
371 if (j->early_journal_entries.nr)
372 darray_exit(&j->early_journal_entries);
376 static bool journal_quiesced(struct journal *j)
378 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
381 journal_entry_close(j);
385 static void journal_quiesce(struct journal *j)
387 wait_event(j->wait, journal_quiesced(j));
390 static void journal_write_work(struct work_struct *work)
392 struct journal *j = container_of(work, struct journal, write_work.work);
393 struct bch_fs *c = container_of(j, struct bch_fs, journal);
397 if (!__journal_entry_is_open(j->reservations))
400 delta = journal_cur_buf(j)->expires - jiffies;
403 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
405 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
407 spin_unlock(&j->lock);
410 static int __journal_res_get(struct journal *j, struct journal_res *res,
413 struct bch_fs *c = container_of(j, struct bch_fs, journal);
414 struct journal_buf *buf;
418 if (journal_res_get_fast(j, res, flags))
421 if (bch2_journal_error(j))
422 return -BCH_ERR_erofs_journal_err;
426 /* check once more in case somebody else shut things down... */
427 if (bch2_journal_error(j)) {
428 spin_unlock(&j->lock);
429 return -BCH_ERR_erofs_journal_err;
433 * Recheck after taking the lock, so we don't race with another thread
434 * that just did journal_entry_open() and call journal_entry_close()
437 if (journal_res_get_fast(j, res, flags)) {
438 spin_unlock(&j->lock);
442 if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
444 * Don't want to close current journal entry, just need to
447 ret = JOURNAL_ERR_journal_full;
452 * If we couldn't get a reservation because the current buf filled up,
453 * and we had room for a bigger entry on disk, signal that we want to
454 * realloc the journal bufs:
456 buf = journal_cur_buf(j);
457 if (journal_entry_is_open(j) &&
458 buf->buf_size >> 9 < buf->disk_sectors &&
459 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
460 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
462 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
463 ret = journal_entry_open(j);
465 if (ret == JOURNAL_ERR_max_in_flight) {
466 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
467 &j->max_in_flight_start, true);
468 trace_and_count(c, journal_entry_full, c);
471 can_discard = j->can_discard;
472 spin_unlock(&j->lock);
476 if (journal_error_check_stuck(j, ret, flags))
477 ret = -BCH_ERR_journal_res_get_blocked;
480 * Journal is full - can't rely on reclaim from work item due to
483 if ((ret == JOURNAL_ERR_journal_full ||
484 ret == JOURNAL_ERR_journal_pin_full) &&
485 !(flags & JOURNAL_RES_GET_NONBLOCK)) {
487 bch2_journal_do_discards(j);
491 if (mutex_trylock(&j->reclaim_lock)) {
492 bch2_journal_reclaim(j);
493 mutex_unlock(&j->reclaim_lock);
497 return ret == JOURNAL_ERR_insufficient_devices
498 ? -BCH_ERR_erofs_journal_err
499 : -BCH_ERR_journal_res_get_blocked;
503 * Essentially the entry function to the journaling code. When bcachefs is doing
504 * a btree insert, it calls this function to get the current journal write.
505 * Journal write is the structure used set up journal writes. The calling
506 * function will then add its keys to the structure, queuing them for the next
509 * To ensure forward progress, the current task must not be holding any
510 * btree node write locks.
512 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
517 closure_wait_event(&j->async_wait,
518 (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
519 (flags & JOURNAL_RES_GET_NONBLOCK));
523 /* journal_entry_res: */
525 void bch2_journal_entry_res_resize(struct journal *j,
526 struct journal_entry_res *res,
529 union journal_res_state state;
530 int d = new_u64s - res->u64s;
534 j->entry_u64s_reserved += d;
538 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
540 state = READ_ONCE(j->reservations);
542 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
543 state.cur_entry_offset > j->cur_entry_u64s) {
544 j->cur_entry_u64s += d;
546 * Not enough room in current journal entry, have to flush it:
548 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
550 journal_cur_buf(j)->u64s_reserved += d;
553 spin_unlock(&j->lock);
557 /* journal flushing: */
560 * bch2_journal_flush_seq_async - wait for a journal entry to be written
563 * @parent: closure object to wait with
564 * Returns: 1 if @seq has already been flushed, 0 if @seq is being flushed,
565 * -EIO if @seq will never be flushed
567 * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
570 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
571 struct closure *parent)
573 struct journal_buf *buf;
576 if (seq <= j->flushed_seq_ondisk)
581 if (WARN_ONCE(seq > journal_cur_seq(j),
582 "requested to flush journal seq %llu, but currently at %llu",
583 seq, journal_cur_seq(j)))
586 /* Recheck under lock: */
587 if (j->err_seq && seq >= j->err_seq) {
592 if (seq <= j->flushed_seq_ondisk) {
597 /* if seq was written, but not flushed - flush a newer one instead */
598 seq = max(seq, journal_last_unwritten_seq(j));
601 if (seq > journal_cur_seq(j)) {
602 struct journal_res res = { 0 };
604 if (journal_entry_is_open(j))
605 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
607 spin_unlock(&j->lock);
609 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
614 buf = j->buf + (seq & JOURNAL_BUF_MASK);
615 buf->must_flush = true;
617 if (!buf->flush_time) {
618 buf->flush_time = local_clock() ?: 1;
619 buf->expires = jiffies;
622 if (parent && !closure_wait(&buf->wait, parent))
625 bch2_journal_res_put(j, &res);
632 * if write was kicked off without a flush, flush the next sequence
635 buf = journal_seq_to_buf(j, seq);
638 goto recheck_need_open;
641 buf->must_flush = true;
643 if (parent && !closure_wait(&buf->wait, parent))
646 if (seq == journal_cur_seq(j))
647 journal_entry_want_write(j);
649 spin_unlock(&j->lock);
653 int bch2_journal_flush_seq(struct journal *j, u64 seq)
655 u64 start_time = local_clock();
659 * Don't update time_stats when @seq is already flushed:
661 if (seq <= j->flushed_seq_ondisk)
664 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
667 bch2_time_stats_update(j->flush_seq_time, start_time);
669 return ret ?: ret2 < 0 ? ret2 : 0;
673 * bch2_journal_flush_async - if there is an open journal entry, or a journal
674 * still being written, write it and wait for the write to complete
676 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
678 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
681 int bch2_journal_flush(struct journal *j)
683 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
687 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
690 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
692 struct bch_fs *c = container_of(j, struct bch_fs, journal);
696 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
699 if (seq <= c->journal.flushed_seq_ondisk)
703 if (seq <= c->journal.flushed_seq_ondisk)
706 for (unwritten_seq = journal_last_unwritten_seq(j);
709 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
711 /* journal write is already in flight, and was a flush write: */
712 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
720 spin_unlock(&j->lock);
724 int bch2_journal_meta(struct journal *j)
726 struct journal_buf *buf;
727 struct journal_res res;
730 memset(&res, 0, sizeof(res));
732 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
736 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
737 buf->must_flush = true;
739 if (!buf->flush_time) {
740 buf->flush_time = local_clock() ?: 1;
741 buf->expires = jiffies;
744 bch2_journal_res_put(j, &res);
746 return bch2_journal_flush_seq(j, res.seq);
749 /* block/unlock the journal: */
751 void bch2_journal_unblock(struct journal *j)
755 spin_unlock(&j->lock);
760 void bch2_journal_block(struct journal *j)
764 spin_unlock(&j->lock);
769 /* allocate journal on a device: */
771 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
772 bool new_fs, struct closure *cl)
774 struct bch_fs *c = ca->fs;
775 struct journal_device *ja = &ca->journal;
776 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
777 struct open_bucket **ob = NULL;
779 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
782 BUG_ON(nr <= ja->nr);
784 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
785 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
786 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
787 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
788 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
789 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
793 for (nr_got = 0; nr_got < nr_want; nr_got++) {
795 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
796 if (bu[nr_got] < 0) {
797 ret = -BCH_ERR_ENOSPC_bucket_alloc;
801 ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
802 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
806 ret = bch2_trans_run(c,
807 bch2_trans_mark_metadata_bucket(trans, ca,
808 ob[nr_got]->bucket, BCH_DATA_journal,
809 ca->mi.bucket_size));
811 bch2_open_bucket_put(c, ob[nr_got]);
812 bch_err_msg(c, ret, "marking new journal buckets");
816 bu[nr_got] = ob[nr_got]->bucket;
823 /* Don't return an error if we successfully allocated some buckets: */
827 bch2_journal_flush_all_pins(&c->journal);
828 bch2_journal_block(&c->journal);
829 mutex_lock(&c->sb_lock);
832 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
833 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
835 BUG_ON(ja->discard_idx > ja->nr);
837 pos = ja->discard_idx ?: ja->nr;
839 memmove(new_buckets + pos + nr_got,
841 sizeof(new_buckets[0]) * (ja->nr - pos));
842 memmove(new_bucket_seq + pos + nr_got,
843 new_bucket_seq + pos,
844 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
846 for (i = 0; i < nr_got; i++) {
847 new_buckets[pos + i] = bu[i];
848 new_bucket_seq[pos + i] = 0;
851 nr = ja->nr + nr_got;
853 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
862 spin_lock(&c->journal.lock);
864 swap(new_buckets, ja->buckets);
865 swap(new_bucket_seq, ja->bucket_seq);
868 if (pos <= ja->discard_idx)
869 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
870 if (pos <= ja->dirty_idx_ondisk)
871 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
872 if (pos <= ja->dirty_idx)
873 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
874 if (pos <= ja->cur_idx)
875 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
878 spin_unlock(&c->journal.lock);
881 bch2_journal_unblock(&c->journal);
882 mutex_unlock(&c->sb_lock);
886 for (i = 0; i < nr_got; i++)
888 bch2_trans_mark_metadata_bucket(trans, ca,
889 bu[i], BCH_DATA_free, 0));
892 for (i = 0; i < nr_got; i++)
893 bch2_open_bucket_put(c, ob[i]);
895 kfree(new_bucket_seq);
903 * Allocate more journal space at runtime - not currently making use if it, but
906 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
909 struct journal_device *ja = &ca->journal;
913 closure_init_stack(&cl);
915 down_write(&c->state_lock);
917 /* don't handle reducing nr of buckets yet: */
921 while (ja->nr < nr) {
922 struct disk_reservation disk_res = { 0, 0, 0 };
925 * note: journal buckets aren't really counted as _sectors_ used yet, so
926 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
927 * when space used goes up without a reservation - but we do need the
928 * reservation to ensure we'll actually be able to allocate:
930 * XXX: that's not right, disk reservations only ensure a
931 * filesystem-wide allocation will succeed, this is a device
932 * specific allocation - we can hang here:
935 ret = bch2_disk_reservation_get(c, &disk_res,
936 bucket_to_sector(ca, nr - ja->nr), 1, 0);
940 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
942 bch2_disk_reservation_put(c, &disk_res);
946 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
953 up_write(&c->state_lock);
957 int bch2_dev_journal_alloc(struct bch_dev *ca)
962 if (dynamic_fault("bcachefs:add:journal_alloc")) {
963 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
967 /* 1/128th of the device by default: */
968 nr = ca->mi.nbuckets >> 7;
971 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
974 nr = clamp_t(unsigned, nr,
975 BCH_JOURNAL_BUCKETS_MIN,
977 (1 << 24) / ca->mi.bucket_size));
979 ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
986 int bch2_fs_journal_alloc(struct bch_fs *c)
991 for_each_online_member(ca, c, i) {
995 int ret = bch2_dev_journal_alloc(ca);
997 percpu_ref_put(&ca->io_ref);
1005 /* startup/shutdown: */
1007 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1012 spin_lock(&j->lock);
1013 for (seq = journal_last_unwritten_seq(j);
1014 seq <= journal_cur_seq(j) && !ret;
1016 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1018 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1021 spin_unlock(&j->lock);
1026 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1028 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1031 void bch2_fs_journal_stop(struct journal *j)
1033 bch2_journal_reclaim_stop(j);
1034 bch2_journal_flush_all_pins(j);
1036 wait_event(j->wait, journal_entry_close(j));
1039 * Always write a new journal entry, to make sure the clock hands are up
1040 * to date (and match the superblock)
1042 bch2_journal_meta(j);
1046 BUG_ON(!bch2_journal_error(j) &&
1047 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1048 j->last_empty_seq != journal_cur_seq(j));
1050 cancel_delayed_work_sync(&j->write_work);
1053 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1055 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1056 struct journal_entry_pin_list *p;
1057 struct journal_replay *i, **_i;
1058 struct genradix_iter iter;
1059 bool had_entries = false;
1061 u64 last_seq = cur_seq, nr, seq;
1063 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1066 if (!i || i->ignore)
1069 last_seq = le64_to_cpu(i->j.last_seq);
1073 nr = cur_seq - last_seq;
1075 if (nr + 1 > j->pin.size) {
1077 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1079 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1080 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1084 j->replay_journal_seq = last_seq;
1085 j->replay_journal_seq_end = cur_seq;
1086 j->last_seq_ondisk = last_seq;
1087 j->flushed_seq_ondisk = cur_seq - 1;
1088 j->seq_ondisk = cur_seq - 1;
1089 j->pin.front = last_seq;
1090 j->pin.back = cur_seq;
1091 atomic64_set(&j->seq, cur_seq - 1);
1093 fifo_for_each_entry_ptr(p, &j->pin, seq)
1094 journal_pin_list_init(p, 1);
1096 genradix_for_each(&c->journal_entries, iter, _i) {
1099 if (!i || i->ignore)
1102 seq = le64_to_cpu(i->j.seq);
1103 BUG_ON(seq >= cur_seq);
1108 if (journal_entry_empty(&i->j))
1109 j->last_empty_seq = le64_to_cpu(i->j.seq);
1111 p = journal_seq_pin(j, seq);
1114 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1115 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1121 j->last_empty_seq = cur_seq;
1123 spin_lock(&j->lock);
1125 set_bit(JOURNAL_STARTED, &j->flags);
1126 j->last_flush_write = jiffies;
1128 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1129 j->reservations.unwritten_idx++;
1131 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1133 bch2_journal_space_available(j);
1134 spin_unlock(&j->lock);
1136 return bch2_journal_reclaim_start(j);
1141 void bch2_dev_journal_exit(struct bch_dev *ca)
1143 kfree(ca->journal.bio);
1144 kfree(ca->journal.buckets);
1145 kfree(ca->journal.bucket_seq);
1147 ca->journal.bio = NULL;
1148 ca->journal.buckets = NULL;
1149 ca->journal.bucket_seq = NULL;
1152 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1154 struct journal_device *ja = &ca->journal;
1155 struct bch_sb_field_journal *journal_buckets =
1156 bch2_sb_field_get(sb, journal);
1157 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1158 bch2_sb_field_get(sb, journal_v2);
1159 unsigned i, nr_bvecs;
1163 if (journal_buckets_v2) {
1164 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1166 for (i = 0; i < nr; i++)
1167 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1168 } else if (journal_buckets) {
1169 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1172 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1173 if (!ja->bucket_seq)
1174 return -BCH_ERR_ENOMEM_dev_journal_init;
1176 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1178 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1179 if (!ca->journal.bio)
1180 return -BCH_ERR_ENOMEM_dev_journal_init;
1182 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1184 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1186 return -BCH_ERR_ENOMEM_dev_journal_init;
1188 if (journal_buckets_v2) {
1189 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1190 unsigned j, dst = 0;
1192 for (i = 0; i < nr; i++)
1193 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1194 ja->buckets[dst++] =
1195 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1196 } else if (journal_buckets) {
1197 for (i = 0; i < ja->nr; i++)
1198 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1204 void bch2_fs_journal_exit(struct journal *j)
1208 darray_exit(&j->early_journal_entries);
1210 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1211 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1215 int bch2_fs_journal_init(struct journal *j)
1217 static struct lock_class_key res_key;
1220 spin_lock_init(&j->lock);
1221 spin_lock_init(&j->err_lock);
1222 init_waitqueue_head(&j->wait);
1223 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1224 init_waitqueue_head(&j->reclaim_wait);
1225 init_waitqueue_head(&j->pin_flush_wait);
1226 mutex_init(&j->reclaim_lock);
1227 mutex_init(&j->discard_lock);
1229 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1231 atomic64_set(&j->reservations.counter,
1232 ((union journal_res_state)
1233 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1235 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1236 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1238 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1239 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1240 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1241 if (!j->buf[i].data)
1242 return -BCH_ERR_ENOMEM_journal_buf;
1245 j->pin.front = j->pin.back = 1;
1251 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1253 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1254 union journal_res_state s;
1256 unsigned long now = jiffies;
1257 u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1261 if (!out->nr_tabstops)
1262 printbuf_tabstop_push(out, 24);
1266 s = READ_ONCE(j->reservations);
1268 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1269 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1270 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1271 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1272 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1273 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1274 prt_printf(out, "watermark:\t\t%s\n", bch2_watermarks[j->watermark]);
1275 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1276 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1277 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1278 prt_printf(out, "average write size:\t");
1279 prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1281 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1282 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1283 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1284 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1285 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1286 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1287 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1288 prt_printf(out, "current entry:\t\t");
1290 switch (s.cur_entry_offset) {
1291 case JOURNAL_ENTRY_ERROR_VAL:
1292 prt_printf(out, "error");
1294 case JOURNAL_ENTRY_CLOSED_VAL:
1295 prt_printf(out, "closed");
1298 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1304 for (seq = journal_cur_seq(j);
1305 seq >= journal_last_unwritten_seq(j);
1307 i = seq & JOURNAL_BUF_MASK;
1309 prt_printf(out, "unwritten entry:");
1311 prt_printf(out, "%llu", seq);
1313 printbuf_indent_add(out, 2);
1315 prt_printf(out, "refcount:");
1317 prt_printf(out, "%u", journal_state_count(s, i));
1320 prt_printf(out, "sectors:");
1322 prt_printf(out, "%u", j->buf[i].sectors);
1325 prt_printf(out, "expires");
1327 prt_printf(out, "%li jiffies", j->buf[i].expires - jiffies);
1330 printbuf_indent_sub(out, 2);
1334 "replay done:\t\t%i\n",
1335 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1337 prt_printf(out, "space:\n");
1338 prt_printf(out, "\tdiscarded\t%u:%u\n",
1339 j->space[journal_space_discarded].next_entry,
1340 j->space[journal_space_discarded].total);
1341 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1342 j->space[journal_space_clean_ondisk].next_entry,
1343 j->space[journal_space_clean_ondisk].total);
1344 prt_printf(out, "\tclean\t\t%u:%u\n",
1345 j->space[journal_space_clean].next_entry,
1346 j->space[journal_space_clean].total);
1347 prt_printf(out, "\ttotal\t\t%u:%u\n",
1348 j->space[journal_space_total].next_entry,
1349 j->space[journal_space_total].total);
1351 for_each_member_device_rcu(ca, c, i,
1352 &c->rw_devs[BCH_DATA_journal]) {
1353 struct journal_device *ja = &ca->journal;
1355 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1361 prt_printf(out, "dev %u:\n", i);
1362 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1363 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1364 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1365 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1366 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1367 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1368 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1376 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1378 spin_lock(&j->lock);
1379 __bch2_journal_debug_to_text(out, j);
1380 spin_unlock(&j->lock);
1383 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1385 struct journal_entry_pin_list *pin_list;
1386 struct journal_entry_pin *pin;
1389 spin_lock(&j->lock);
1390 *seq = max(*seq, j->pin.front);
1392 if (*seq >= j->pin.back) {
1393 spin_unlock(&j->lock);
1399 pin_list = journal_seq_pin(j, *seq);
1401 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1403 printbuf_indent_add(out, 2);
1405 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1406 list_for_each_entry(pin, &pin_list->list[i], list) {
1407 prt_printf(out, "\t%px %ps", pin, pin->flush);
1411 if (!list_empty(&pin_list->flushed)) {
1412 prt_printf(out, "flushed:");
1416 list_for_each_entry(pin, &pin_list->flushed, list) {
1417 prt_printf(out, "\t%px %ps", pin, pin->flush);
1421 printbuf_indent_sub(out, 2);
1424 spin_unlock(&j->lock);
1429 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1433 while (!bch2_journal_seq_pins_to_text(out, j, &seq))