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"
13 #include "btree_write_buffer.h"
17 #include "journal_io.h"
18 #include "journal_reclaim.h"
19 #include "journal_sb.h"
20 #include "journal_seq_blacklist.h"
23 static const char * const bch2_journal_errors[] = {
30 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
32 return seq > j->seq_ondisk;
35 static bool __journal_entry_is_open(union journal_res_state state)
37 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
40 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
42 return atomic64_read(&j->seq) - j->seq_ondisk;
45 static bool journal_entry_is_open(struct journal *j)
47 return __journal_entry_is_open(j->reservations);
50 static inline struct journal_buf *
51 journal_seq_to_buf(struct journal *j, u64 seq)
53 struct journal_buf *buf = NULL;
55 EBUG_ON(seq > journal_cur_seq(j));
57 if (journal_seq_unwritten(j, seq)) {
58 buf = j->buf + (seq & JOURNAL_BUF_MASK);
59 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
64 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
68 for (i = 0; i < ARRAY_SIZE(p->list); i++)
69 INIT_LIST_HEAD(&p->list[i]);
70 INIT_LIST_HEAD(&p->flushed);
71 atomic_set(&p->count, count);
76 * Detect stuck journal conditions and trigger shutdown. Technically the journal
77 * can end up stuck for a variety of reasons, such as a blocked I/O, journal
78 * reservation lockup, etc. Since this is a fatal error with potentially
79 * unpredictable characteristics, we want to be fairly conservative before we
80 * decide to shut things down.
82 * Consider the journal stuck when it appears full with no ability to commit
83 * btree transactions, to discard journal buckets, nor acquire priority
84 * (reserved watermark) reservation.
87 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
89 struct bch_fs *c = container_of(j, struct bch_fs, journal);
91 struct printbuf buf = PRINTBUF;
93 if (!(error == JOURNAL_ERR_journal_full ||
94 error == JOURNAL_ERR_journal_pin_full) ||
95 nr_unwritten_journal_entries(j) ||
96 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
101 if (j->can_discard) {
102 spin_unlock(&j->lock);
109 * The journal shutdown path will set ->err_seq, but do it here first to
110 * serialize against concurrent failures and avoid duplicate error
114 spin_unlock(&j->lock);
117 j->err_seq = journal_cur_seq(j);
118 spin_unlock(&j->lock);
120 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
121 bch2_journal_errors[error]);
122 bch2_journal_debug_to_text(&buf, j);
123 bch_err(c, "%s", buf.buf);
125 printbuf_reset(&buf);
126 bch2_journal_pins_to_text(&buf, j);
127 bch_err(c, "Journal pins:\n%s", buf.buf);
137 * Final processing when the last reference of a journal buffer has been
138 * dropped. Drop the pin list reference acquired at journal entry open and write
139 * the buffer, if requested.
141 void bch2_journal_buf_put_final(struct journal *j, u64 seq, bool write)
143 struct bch_fs *c = container_of(j, struct bch_fs, journal);
145 lockdep_assert_held(&j->lock);
147 if (__bch2_journal_pin_put(j, seq))
148 bch2_journal_reclaim_fast(j);
150 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
155 * Returns true if journal entry is now closed:
157 * We don't close a journal_buf until the next journal_buf is finished writing,
158 * and can be opened again - this also initializes the next journal_buf:
160 static void __journal_entry_close(struct journal *j, unsigned closed_val)
162 struct bch_fs *c = container_of(j, struct bch_fs, journal);
163 struct journal_buf *buf = journal_cur_buf(j);
164 union journal_res_state old, new;
165 u64 v = atomic64_read(&j->reservations.counter);
168 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
169 closed_val != JOURNAL_ENTRY_ERROR_VAL);
171 lockdep_assert_held(&j->lock);
175 new.cur_entry_offset = closed_val;
177 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
178 old.cur_entry_offset == new.cur_entry_offset)
180 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
181 old.v, new.v)) != old.v);
183 if (!__journal_entry_is_open(old))
186 /* Close out old buffer: */
187 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
189 trace_journal_entry_close(c, vstruct_bytes(buf->data));
191 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
192 buf->u64s_reserved) << c->block_bits;
193 BUG_ON(sectors > buf->sectors);
194 buf->sectors = sectors;
197 * We have to set last_seq here, _before_ opening a new journal entry:
199 * A threads may replace an old pin with a new pin on their current
200 * journal reservation - the expectation being that the journal will
201 * contain either what the old pin protected or what the new pin
204 * After the old pin is dropped journal_last_seq() won't include the old
205 * pin, so we can only write the updated last_seq on the entry that
206 * contains whatever the new pin protects.
208 * Restated, we can _not_ update last_seq for a given entry if there
209 * could be a newer entry open with reservations/pins that have been
212 * Hence, we want update/set last_seq on the current journal entry right
213 * before we open a new one:
215 buf->last_seq = journal_last_seq(j);
216 buf->data->last_seq = cpu_to_le64(buf->last_seq);
217 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
219 cancel_delayed_work(&j->write_work);
221 bch2_journal_space_available(j);
223 __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
226 void bch2_journal_halt(struct journal *j)
229 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL);
231 j->err_seq = journal_cur_seq(j);
233 spin_unlock(&j->lock);
236 static bool journal_entry_want_write(struct journal *j)
238 bool ret = !journal_entry_is_open(j) ||
239 journal_cur_seq(j) == journal_last_unwritten_seq(j);
241 /* Don't close it yet if we already have a write in flight: */
243 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
244 else if (nr_unwritten_journal_entries(j)) {
245 struct journal_buf *buf = journal_cur_buf(j);
247 if (!buf->flush_time) {
248 buf->flush_time = local_clock() ?: 1;
249 buf->expires = jiffies;
256 static bool journal_entry_close(struct journal *j)
261 ret = journal_entry_want_write(j);
262 spin_unlock(&j->lock);
268 * should _only_ called from journal_res_get() - when we actually want a
269 * journal reservation - journal entry is open means journal is dirty:
271 static int journal_entry_open(struct journal *j)
273 struct bch_fs *c = container_of(j, struct bch_fs, journal);
274 struct journal_buf *buf = j->buf +
275 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
276 union journal_res_state old, new;
280 lockdep_assert_held(&j->lock);
281 BUG_ON(journal_entry_is_open(j));
282 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
285 return JOURNAL_ERR_blocked;
287 if (j->cur_entry_error)
288 return j->cur_entry_error;
290 if (bch2_journal_error(j))
291 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
293 if (!fifo_free(&j->pin))
294 return JOURNAL_ERR_journal_pin_full;
296 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
297 return JOURNAL_ERR_max_in_flight;
299 BUG_ON(!j->cur_entry_sectors);
302 (journal_cur_seq(j) == j->flushed_seq_ondisk
304 : j->last_flush_write) +
305 msecs_to_jiffies(c->opts.journal_flush_delay);
307 buf->u64s_reserved = j->entry_u64s_reserved;
308 buf->disk_sectors = j->cur_entry_sectors;
309 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
311 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
312 journal_entry_overhead(j);
313 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
315 if (u64s <= (ssize_t) j->early_journal_entries.nr)
316 return JOURNAL_ERR_journal_full;
318 if (fifo_empty(&j->pin) && j->reclaim_thread)
319 wake_up_process(j->reclaim_thread);
322 * The fifo_push() needs to happen at the same time as j->seq is
323 * incremented for journal_last_seq() to be calculated correctly
325 atomic64_inc(&j->seq);
326 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
328 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
330 bkey_extent_init(&buf->key);
331 buf->noflush = false;
332 buf->must_flush = false;
333 buf->separate_flush = false;
335 buf->need_flush_to_write_buffer = true;
337 memset(buf->data, 0, sizeof(*buf->data));
338 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
341 if (j->early_journal_entries.nr) {
342 memcpy(buf->data->_data, j->early_journal_entries.data,
343 j->early_journal_entries.nr * sizeof(u64));
344 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
348 * Must be set before marking the journal entry as open:
350 j->cur_entry_u64s = u64s;
352 v = atomic64_read(&j->reservations.counter);
356 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
359 BUG_ON(journal_state_count(new, new.idx));
360 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
362 journal_state_inc(&new);
364 /* Handle any already added entries */
365 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
366 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
367 old.v, new.v)) != old.v);
369 mod_delayed_work(c->io_complete_wq,
371 msecs_to_jiffies(c->opts.journal_flush_delay));
374 if (j->early_journal_entries.nr)
375 darray_exit(&j->early_journal_entries);
379 static bool journal_quiesced(struct journal *j)
381 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
384 journal_entry_close(j);
388 static void journal_quiesce(struct journal *j)
390 wait_event(j->wait, journal_quiesced(j));
393 static void journal_write_work(struct work_struct *work)
395 struct journal *j = container_of(work, struct journal, write_work.work);
396 struct bch_fs *c = container_of(j, struct bch_fs, journal);
400 if (!__journal_entry_is_open(j->reservations))
403 delta = journal_cur_buf(j)->expires - jiffies;
406 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
408 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
410 spin_unlock(&j->lock);
413 static int __journal_res_get(struct journal *j, struct journal_res *res,
416 struct bch_fs *c = container_of(j, struct bch_fs, journal);
417 struct journal_buf *buf;
421 if (journal_res_get_fast(j, res, flags))
424 if (bch2_journal_error(j))
425 return -BCH_ERR_erofs_journal_err;
429 /* check once more in case somebody else shut things down... */
430 if (bch2_journal_error(j)) {
431 spin_unlock(&j->lock);
432 return -BCH_ERR_erofs_journal_err;
436 * Recheck after taking the lock, so we don't race with another thread
437 * that just did journal_entry_open() and call journal_entry_close()
440 if (journal_res_get_fast(j, res, flags)) {
441 spin_unlock(&j->lock);
445 if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
447 * Don't want to close current journal entry, just need to
450 ret = JOURNAL_ERR_journal_full;
455 * If we couldn't get a reservation because the current buf filled up,
456 * and we had room for a bigger entry on disk, signal that we want to
457 * realloc the journal bufs:
459 buf = journal_cur_buf(j);
460 if (journal_entry_is_open(j) &&
461 buf->buf_size >> 9 < buf->disk_sectors &&
462 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
463 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
465 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
466 ret = journal_entry_open(j);
468 if (ret == JOURNAL_ERR_max_in_flight) {
469 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
470 &j->max_in_flight_start, true);
471 trace_and_count(c, journal_entry_full, c);
474 can_discard = j->can_discard;
475 spin_unlock(&j->lock);
479 if (journal_error_check_stuck(j, ret, flags))
480 ret = -BCH_ERR_journal_res_get_blocked;
483 * Journal is full - can't rely on reclaim from work item due to
486 if ((ret == JOURNAL_ERR_journal_full ||
487 ret == JOURNAL_ERR_journal_pin_full) &&
488 !(flags & JOURNAL_RES_GET_NONBLOCK)) {
490 bch2_journal_do_discards(j);
494 if (mutex_trylock(&j->reclaim_lock)) {
495 bch2_journal_reclaim(j);
496 mutex_unlock(&j->reclaim_lock);
500 return ret == JOURNAL_ERR_insufficient_devices
501 ? -BCH_ERR_erofs_journal_err
502 : -BCH_ERR_journal_res_get_blocked;
506 * Essentially the entry function to the journaling code. When bcachefs is doing
507 * a btree insert, it calls this function to get the current journal write.
508 * Journal write is the structure used set up journal writes. The calling
509 * function will then add its keys to the structure, queuing them for the next
512 * To ensure forward progress, the current task must not be holding any
513 * btree node write locks.
515 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
520 closure_wait_event(&j->async_wait,
521 (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
522 (flags & JOURNAL_RES_GET_NONBLOCK));
526 /* journal_entry_res: */
528 void bch2_journal_entry_res_resize(struct journal *j,
529 struct journal_entry_res *res,
532 union journal_res_state state;
533 int d = new_u64s - res->u64s;
537 j->entry_u64s_reserved += d;
541 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
543 state = READ_ONCE(j->reservations);
545 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
546 state.cur_entry_offset > j->cur_entry_u64s) {
547 j->cur_entry_u64s += d;
549 * Not enough room in current journal entry, have to flush it:
551 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
553 journal_cur_buf(j)->u64s_reserved += d;
556 spin_unlock(&j->lock);
560 /* journal flushing: */
563 * bch2_journal_flush_seq_async - wait for a journal entry to be written
566 * @parent: closure object to wait with
567 * Returns: 1 if @seq has already been flushed, 0 if @seq is being flushed,
568 * -EIO if @seq will never be flushed
570 * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
573 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
574 struct closure *parent)
576 struct journal_buf *buf;
579 if (seq <= j->flushed_seq_ondisk)
584 if (WARN_ONCE(seq > journal_cur_seq(j),
585 "requested to flush journal seq %llu, but currently at %llu",
586 seq, journal_cur_seq(j)))
589 /* Recheck under lock: */
590 if (j->err_seq && seq >= j->err_seq) {
595 if (seq <= j->flushed_seq_ondisk) {
600 /* if seq was written, but not flushed - flush a newer one instead */
601 seq = max(seq, journal_last_unwritten_seq(j));
604 if (seq > journal_cur_seq(j)) {
605 struct journal_res res = { 0 };
607 if (journal_entry_is_open(j))
608 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
610 spin_unlock(&j->lock);
612 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
617 buf = j->buf + (seq & JOURNAL_BUF_MASK);
618 buf->must_flush = true;
620 if (!buf->flush_time) {
621 buf->flush_time = local_clock() ?: 1;
622 buf->expires = jiffies;
625 if (parent && !closure_wait(&buf->wait, parent))
628 bch2_journal_res_put(j, &res);
635 * if write was kicked off without a flush, flush the next sequence
638 buf = journal_seq_to_buf(j, seq);
641 goto recheck_need_open;
644 buf->must_flush = true;
646 if (parent && !closure_wait(&buf->wait, parent))
649 if (seq == journal_cur_seq(j))
650 journal_entry_want_write(j);
652 spin_unlock(&j->lock);
656 int bch2_journal_flush_seq(struct journal *j, u64 seq)
658 u64 start_time = local_clock();
662 * Don't update time_stats when @seq is already flushed:
664 if (seq <= j->flushed_seq_ondisk)
667 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
670 bch2_time_stats_update(j->flush_seq_time, start_time);
672 return ret ?: ret2 < 0 ? ret2 : 0;
676 * bch2_journal_flush_async - if there is an open journal entry, or a journal
677 * still being written, write it and wait for the write to complete
679 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
681 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
684 int bch2_journal_flush(struct journal *j)
686 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
690 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
693 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
695 struct bch_fs *c = container_of(j, struct bch_fs, journal);
699 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
702 if (seq <= c->journal.flushed_seq_ondisk)
706 if (seq <= c->journal.flushed_seq_ondisk)
709 for (unwritten_seq = journal_last_unwritten_seq(j);
712 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
714 /* journal write is already in flight, and was a flush write: */
715 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
723 spin_unlock(&j->lock);
727 int bch2_journal_meta(struct journal *j)
729 struct journal_buf *buf;
730 struct journal_res res;
733 memset(&res, 0, sizeof(res));
735 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
739 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
740 buf->must_flush = true;
742 if (!buf->flush_time) {
743 buf->flush_time = local_clock() ?: 1;
744 buf->expires = jiffies;
747 bch2_journal_res_put(j, &res);
749 return bch2_journal_flush_seq(j, res.seq);
752 /* block/unlock the journal: */
754 void bch2_journal_unblock(struct journal *j)
758 spin_unlock(&j->lock);
763 void bch2_journal_block(struct journal *j)
767 spin_unlock(&j->lock);
773 * XXX: ideally this would not be closing the current journal entry, but
774 * otherwise we do not have a way to avoid racing with res_get() - j->blocked
777 static bool journal_reservations_stopped(struct journal *j)
779 union journal_res_state s;
781 journal_entry_close(j);
783 s.v = atomic64_read_acquire(&j->reservations.counter);
785 return s.buf0_count == 0 &&
791 void bch2_journal_block_reservations(struct journal *j)
795 spin_unlock(&j->lock);
797 wait_event(j->wait, journal_reservations_stopped(j));
800 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
803 max_seq = min(max_seq, journal_cur_seq(j));
805 for (u64 seq = journal_last_unwritten_seq(j);
808 unsigned idx = seq & JOURNAL_BUF_MASK;
809 struct journal_buf *buf = j->buf + idx;
810 union journal_res_state s;
812 if (!buf->need_flush_to_write_buffer)
815 if (seq == journal_cur_seq(j))
816 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
818 s.v = atomic64_read_acquire(&j->reservations.counter);
820 if (journal_state_count(s, idx)) {
821 spin_unlock(&j->lock);
822 return ERR_PTR(-EAGAIN);
825 spin_unlock(&j->lock);
829 spin_unlock(&j->lock);
833 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
835 struct journal_buf *ret;
837 wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
841 /* allocate journal on a device: */
843 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
844 bool new_fs, struct closure *cl)
846 struct bch_fs *c = ca->fs;
847 struct journal_device *ja = &ca->journal;
848 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
849 struct open_bucket **ob = NULL;
851 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
854 BUG_ON(nr <= ja->nr);
856 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
857 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
858 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
859 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
860 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
861 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
865 for (nr_got = 0; nr_got < nr_want; nr_got++) {
867 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
868 if (bu[nr_got] < 0) {
869 ret = -BCH_ERR_ENOSPC_bucket_alloc;
873 ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
874 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
878 ret = bch2_trans_run(c,
879 bch2_trans_mark_metadata_bucket(trans, ca,
880 ob[nr_got]->bucket, BCH_DATA_journal,
881 ca->mi.bucket_size));
883 bch2_open_bucket_put(c, ob[nr_got]);
884 bch_err_msg(c, ret, "marking new journal buckets");
888 bu[nr_got] = ob[nr_got]->bucket;
895 /* Don't return an error if we successfully allocated some buckets: */
899 bch2_journal_flush_all_pins(&c->journal);
900 bch2_journal_block(&c->journal);
901 mutex_lock(&c->sb_lock);
904 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
905 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
907 BUG_ON(ja->discard_idx > ja->nr);
909 pos = ja->discard_idx ?: ja->nr;
911 memmove(new_buckets + pos + nr_got,
913 sizeof(new_buckets[0]) * (ja->nr - pos));
914 memmove(new_bucket_seq + pos + nr_got,
915 new_bucket_seq + pos,
916 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
918 for (i = 0; i < nr_got; i++) {
919 new_buckets[pos + i] = bu[i];
920 new_bucket_seq[pos + i] = 0;
923 nr = ja->nr + nr_got;
925 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
934 spin_lock(&c->journal.lock);
936 swap(new_buckets, ja->buckets);
937 swap(new_bucket_seq, ja->bucket_seq);
940 if (pos <= ja->discard_idx)
941 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
942 if (pos <= ja->dirty_idx_ondisk)
943 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
944 if (pos <= ja->dirty_idx)
945 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
946 if (pos <= ja->cur_idx)
947 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
950 spin_unlock(&c->journal.lock);
953 bch2_journal_unblock(&c->journal);
954 mutex_unlock(&c->sb_lock);
958 for (i = 0; i < nr_got; i++)
960 bch2_trans_mark_metadata_bucket(trans, ca,
961 bu[i], BCH_DATA_free, 0));
964 for (i = 0; i < nr_got; i++)
965 bch2_open_bucket_put(c, ob[i]);
967 kfree(new_bucket_seq);
975 * Allocate more journal space at runtime - not currently making use if it, but
978 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
981 struct journal_device *ja = &ca->journal;
985 closure_init_stack(&cl);
987 down_write(&c->state_lock);
989 /* don't handle reducing nr of buckets yet: */
993 while (ja->nr < nr) {
994 struct disk_reservation disk_res = { 0, 0, 0 };
997 * note: journal buckets aren't really counted as _sectors_ used yet, so
998 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
999 * when space used goes up without a reservation - but we do need the
1000 * reservation to ensure we'll actually be able to allocate:
1002 * XXX: that's not right, disk reservations only ensure a
1003 * filesystem-wide allocation will succeed, this is a device
1004 * specific allocation - we can hang here:
1007 ret = bch2_disk_reservation_get(c, &disk_res,
1008 bucket_to_sector(ca, nr - ja->nr), 1, 0);
1012 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1014 bch2_disk_reservation_put(c, &disk_res);
1018 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1025 up_write(&c->state_lock);
1029 int bch2_dev_journal_alloc(struct bch_dev *ca)
1034 if (dynamic_fault("bcachefs:add:journal_alloc")) {
1035 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1039 /* 1/128th of the device by default: */
1040 nr = ca->mi.nbuckets >> 7;
1043 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1046 nr = clamp_t(unsigned, nr,
1047 BCH_JOURNAL_BUCKETS_MIN,
1049 (1 << 24) / ca->mi.bucket_size));
1051 ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1054 bch_err_fn(ca, ret);
1058 int bch2_fs_journal_alloc(struct bch_fs *c)
1063 for_each_online_member(ca, c, i) {
1067 int ret = bch2_dev_journal_alloc(ca);
1069 percpu_ref_put(&ca->io_ref);
1077 /* startup/shutdown: */
1079 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1084 spin_lock(&j->lock);
1085 for (seq = journal_last_unwritten_seq(j);
1086 seq <= journal_cur_seq(j) && !ret;
1088 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1090 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1093 spin_unlock(&j->lock);
1098 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1100 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1103 void bch2_fs_journal_stop(struct journal *j)
1105 bch2_journal_reclaim_stop(j);
1106 bch2_journal_flush_all_pins(j);
1108 wait_event(j->wait, journal_entry_close(j));
1111 * Always write a new journal entry, to make sure the clock hands are up
1112 * to date (and match the superblock)
1114 bch2_journal_meta(j);
1118 BUG_ON(!bch2_journal_error(j) &&
1119 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1120 j->last_empty_seq != journal_cur_seq(j));
1122 cancel_delayed_work_sync(&j->write_work);
1125 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1127 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1128 struct journal_entry_pin_list *p;
1129 struct journal_replay *i, **_i;
1130 struct genradix_iter iter;
1131 bool had_entries = false;
1133 u64 last_seq = cur_seq, nr, seq;
1135 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1138 if (!i || i->ignore)
1141 last_seq = le64_to_cpu(i->j.last_seq);
1145 nr = cur_seq - last_seq;
1147 if (nr + 1 > j->pin.size) {
1149 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1151 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1152 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1156 j->replay_journal_seq = last_seq;
1157 j->replay_journal_seq_end = cur_seq;
1158 j->last_seq_ondisk = last_seq;
1159 j->flushed_seq_ondisk = cur_seq - 1;
1160 j->seq_ondisk = cur_seq - 1;
1161 j->pin.front = last_seq;
1162 j->pin.back = cur_seq;
1163 atomic64_set(&j->seq, cur_seq - 1);
1165 fifo_for_each_entry_ptr(p, &j->pin, seq)
1166 journal_pin_list_init(p, 1);
1168 genradix_for_each(&c->journal_entries, iter, _i) {
1171 if (!i || i->ignore)
1174 seq = le64_to_cpu(i->j.seq);
1175 BUG_ON(seq >= cur_seq);
1180 if (journal_entry_empty(&i->j))
1181 j->last_empty_seq = le64_to_cpu(i->j.seq);
1183 p = journal_seq_pin(j, seq);
1186 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1187 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1193 j->last_empty_seq = cur_seq;
1195 spin_lock(&j->lock);
1197 set_bit(JOURNAL_STARTED, &j->flags);
1198 j->last_flush_write = jiffies;
1200 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1201 j->reservations.unwritten_idx++;
1203 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1205 bch2_journal_space_available(j);
1206 spin_unlock(&j->lock);
1208 return bch2_journal_reclaim_start(j);
1213 void bch2_dev_journal_exit(struct bch_dev *ca)
1215 kfree(ca->journal.bio);
1216 kfree(ca->journal.buckets);
1217 kfree(ca->journal.bucket_seq);
1219 ca->journal.bio = NULL;
1220 ca->journal.buckets = NULL;
1221 ca->journal.bucket_seq = NULL;
1224 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1226 struct journal_device *ja = &ca->journal;
1227 struct bch_sb_field_journal *journal_buckets =
1228 bch2_sb_field_get(sb, journal);
1229 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1230 bch2_sb_field_get(sb, journal_v2);
1231 unsigned i, nr_bvecs;
1235 if (journal_buckets_v2) {
1236 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1238 for (i = 0; i < nr; i++)
1239 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1240 } else if (journal_buckets) {
1241 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1244 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1245 if (!ja->bucket_seq)
1246 return -BCH_ERR_ENOMEM_dev_journal_init;
1248 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1250 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1251 if (!ca->journal.bio)
1252 return -BCH_ERR_ENOMEM_dev_journal_init;
1254 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1256 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1258 return -BCH_ERR_ENOMEM_dev_journal_init;
1260 if (journal_buckets_v2) {
1261 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1262 unsigned j, dst = 0;
1264 for (i = 0; i < nr; i++)
1265 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1266 ja->buckets[dst++] =
1267 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1268 } else if (journal_buckets) {
1269 for (i = 0; i < ja->nr; i++)
1270 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1276 void bch2_fs_journal_exit(struct journal *j)
1280 darray_exit(&j->early_journal_entries);
1282 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1283 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1287 int bch2_fs_journal_init(struct journal *j)
1289 static struct lock_class_key res_key;
1292 mutex_init(&j->buf_lock);
1293 spin_lock_init(&j->lock);
1294 spin_lock_init(&j->err_lock);
1295 init_waitqueue_head(&j->wait);
1296 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1297 init_waitqueue_head(&j->reclaim_wait);
1298 init_waitqueue_head(&j->pin_flush_wait);
1299 mutex_init(&j->reclaim_lock);
1300 mutex_init(&j->discard_lock);
1302 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1304 atomic64_set(&j->reservations.counter,
1305 ((union journal_res_state)
1306 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1308 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1309 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1311 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1312 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1313 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1314 if (!j->buf[i].data)
1315 return -BCH_ERR_ENOMEM_journal_buf;
1318 j->pin.front = j->pin.back = 1;
1324 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1326 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1327 union journal_res_state s;
1329 unsigned long now = jiffies;
1330 u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1334 if (!out->nr_tabstops)
1335 printbuf_tabstop_push(out, 24);
1339 s = READ_ONCE(j->reservations);
1341 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1342 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1343 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1344 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1345 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1346 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1347 prt_printf(out, "watermark:\t\t%s\n", bch2_watermarks[j->watermark]);
1348 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1349 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1350 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1351 prt_printf(out, "average write size:\t");
1352 prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1354 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1355 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1356 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1357 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1358 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1359 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1360 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1361 prt_printf(out, "current entry:\t\t");
1363 switch (s.cur_entry_offset) {
1364 case JOURNAL_ENTRY_ERROR_VAL:
1365 prt_printf(out, "error");
1367 case JOURNAL_ENTRY_CLOSED_VAL:
1368 prt_printf(out, "closed");
1371 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1377 for (seq = journal_cur_seq(j);
1378 seq >= journal_last_unwritten_seq(j);
1380 i = seq & JOURNAL_BUF_MASK;
1382 prt_printf(out, "unwritten entry:");
1384 prt_printf(out, "%llu", seq);
1386 printbuf_indent_add(out, 2);
1388 prt_printf(out, "refcount:");
1390 prt_printf(out, "%u", journal_state_count(s, i));
1393 prt_printf(out, "sectors:");
1395 prt_printf(out, "%u", j->buf[i].sectors);
1398 prt_printf(out, "expires");
1400 prt_printf(out, "%li jiffies", j->buf[i].expires - jiffies);
1403 printbuf_indent_sub(out, 2);
1407 "replay done:\t\t%i\n",
1408 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1410 prt_printf(out, "space:\n");
1411 prt_printf(out, "\tdiscarded\t%u:%u\n",
1412 j->space[journal_space_discarded].next_entry,
1413 j->space[journal_space_discarded].total);
1414 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1415 j->space[journal_space_clean_ondisk].next_entry,
1416 j->space[journal_space_clean_ondisk].total);
1417 prt_printf(out, "\tclean\t\t%u:%u\n",
1418 j->space[journal_space_clean].next_entry,
1419 j->space[journal_space_clean].total);
1420 prt_printf(out, "\ttotal\t\t%u:%u\n",
1421 j->space[journal_space_total].next_entry,
1422 j->space[journal_space_total].total);
1424 for_each_member_device_rcu(ca, c, i,
1425 &c->rw_devs[BCH_DATA_journal]) {
1426 struct journal_device *ja = &ca->journal;
1428 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1434 prt_printf(out, "dev %u:\n", i);
1435 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1436 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1437 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1438 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1439 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1440 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1441 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1449 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1451 spin_lock(&j->lock);
1452 __bch2_journal_debug_to_text(out, j);
1453 spin_unlock(&j->lock);
1456 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1458 struct journal_entry_pin_list *pin_list;
1459 struct journal_entry_pin *pin;
1462 spin_lock(&j->lock);
1463 *seq = max(*seq, j->pin.front);
1465 if (*seq >= j->pin.back) {
1466 spin_unlock(&j->lock);
1472 pin_list = journal_seq_pin(j, *seq);
1474 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1476 printbuf_indent_add(out, 2);
1478 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1479 list_for_each_entry(pin, &pin_list->list[i], list) {
1480 prt_printf(out, "\t%px %ps", pin, pin->flush);
1484 if (!list_empty(&pin_list->flushed)) {
1485 prt_printf(out, "flushed:");
1489 list_for_each_entry(pin, &pin_list->flushed, list) {
1490 prt_printf(out, "\t%px %ps", pin, pin->flush);
1494 printbuf_indent_sub(out, 2);
1497 spin_unlock(&j->lock);
1502 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1506 while (!bch2_journal_seq_pins_to_text(out, j, &seq))