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"
21 #include <trace/events/bcachefs.h>
24 static const char * const bch2_journal_watermarks[] = {
29 static const char * const bch2_journal_errors[] = {
35 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
37 return seq > j->seq_ondisk;
40 static bool __journal_entry_is_open(union journal_res_state state)
42 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
45 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
47 return atomic64_read(&j->seq) - j->seq_ondisk;
50 static bool journal_entry_is_open(struct journal *j)
52 return __journal_entry_is_open(j->reservations);
55 static inline struct journal_buf *
56 journal_seq_to_buf(struct journal *j, u64 seq)
58 struct journal_buf *buf = NULL;
60 EBUG_ON(seq > journal_cur_seq(j));
62 if (journal_seq_unwritten(j, seq)) {
63 buf = j->buf + (seq & JOURNAL_BUF_MASK);
64 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
69 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
72 for (i = 0; i < ARRAY_SIZE(p->list); i++)
73 INIT_LIST_HEAD(&p->list[i]);
74 INIT_LIST_HEAD(&p->flushed);
75 atomic_set(&p->count, count);
80 * Detect stuck journal conditions and trigger shutdown. Technically the journal
81 * can end up stuck for a variety of reasons, such as a blocked I/O, journal
82 * reservation lockup, etc. Since this is a fatal error with potentially
83 * unpredictable characteristics, we want to be fairly conservative before we
84 * decide to shut things down.
86 * Consider the journal stuck when it appears full with no ability to commit
87 * btree transactions, to discard journal buckets, nor acquire priority
88 * (reserved watermark) reservation.
91 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
93 struct bch_fs *c = container_of(j, struct bch_fs, journal);
95 struct printbuf buf = PRINTBUF;
97 if (!(error == JOURNAL_ERR_journal_full ||
98 error == JOURNAL_ERR_journal_pin_full) ||
99 nr_unwritten_journal_entries(j) ||
100 (flags & JOURNAL_WATERMARK_MASK) != JOURNAL_WATERMARK_reserved)
105 if (j->can_discard) {
106 spin_unlock(&j->lock);
113 * The journal shutdown path will set ->err_seq, but do it here first to
114 * serialize against concurrent failures and avoid duplicate error
118 spin_unlock(&j->lock);
121 j->err_seq = journal_cur_seq(j);
122 spin_unlock(&j->lock);
124 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
125 bch2_journal_errors[error]);
126 bch2_journal_debug_to_text(&buf, j);
127 bch_err(c, "%s", buf.buf);
129 printbuf_reset(&buf);
130 bch2_journal_pins_to_text(&buf, j);
131 bch_err(c, "Journal pins:\n%s", buf.buf);
140 /* journal entry close/open: */
142 void __bch2_journal_buf_put(struct journal *j)
144 struct bch_fs *c = container_of(j, struct bch_fs, journal);
146 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
150 * Returns true if journal entry is now closed:
152 * We don't close a journal_buf until the next journal_buf is finished writing,
153 * and can be opened again - this also initializes the next journal_buf:
155 static void __journal_entry_close(struct journal *j, unsigned closed_val)
157 struct bch_fs *c = container_of(j, struct bch_fs, journal);
158 struct journal_buf *buf = journal_cur_buf(j);
159 union journal_res_state old, new;
160 u64 v = atomic64_read(&j->reservations.counter);
163 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
164 closed_val != JOURNAL_ENTRY_ERROR_VAL);
166 lockdep_assert_held(&j->lock);
170 new.cur_entry_offset = closed_val;
172 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
173 old.cur_entry_offset == new.cur_entry_offset)
175 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
176 old.v, new.v)) != old.v);
178 if (!__journal_entry_is_open(old))
181 /* Close out old buffer: */
182 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
184 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
185 buf->u64s_reserved) << c->block_bits;
186 BUG_ON(sectors > buf->sectors);
187 buf->sectors = sectors;
190 * We have to set last_seq here, _before_ opening a new journal entry:
192 * A threads may replace an old pin with a new pin on their current
193 * journal reservation - the expectation being that the journal will
194 * contain either what the old pin protected or what the new pin
197 * After the old pin is dropped journal_last_seq() won't include the old
198 * pin, so we can only write the updated last_seq on the entry that
199 * contains whatever the new pin protects.
201 * Restated, we can _not_ update last_seq for a given entry if there
202 * could be a newer entry open with reservations/pins that have been
205 * Hence, we want update/set last_seq on the current journal entry right
206 * before we open a new one:
208 buf->last_seq = journal_last_seq(j);
209 buf->data->last_seq = cpu_to_le64(buf->last_seq);
210 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
212 __bch2_journal_pin_put(j, le64_to_cpu(buf->data->seq));
214 cancel_delayed_work(&j->write_work);
216 bch2_journal_space_available(j);
218 bch2_journal_buf_put(j, old.idx);
221 void bch2_journal_halt(struct journal *j)
224 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL);
226 j->err_seq = journal_cur_seq(j);
228 spin_unlock(&j->lock);
231 static bool journal_entry_want_write(struct journal *j)
233 bool ret = !journal_entry_is_open(j) ||
234 journal_cur_seq(j) == journal_last_unwritten_seq(j);
236 /* Don't close it yet if we already have a write in flight: */
238 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
239 else if (nr_unwritten_journal_entries(j)) {
240 struct journal_buf *buf = journal_cur_buf(j);
242 if (!buf->flush_time) {
243 buf->flush_time = local_clock() ?: 1;
244 buf->expires = jiffies;
251 static bool journal_entry_close(struct journal *j)
256 ret = journal_entry_want_write(j);
257 spin_unlock(&j->lock);
263 * should _only_ called from journal_res_get() - when we actually want a
264 * journal reservation - journal entry is open means journal is dirty:
266 static int journal_entry_open(struct journal *j)
268 struct bch_fs *c = container_of(j, struct bch_fs, journal);
269 struct journal_buf *buf = j->buf +
270 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
271 union journal_res_state old, new;
275 lockdep_assert_held(&j->lock);
276 BUG_ON(journal_entry_is_open(j));
277 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
280 return JOURNAL_ERR_blocked;
282 if (j->cur_entry_error)
283 return j->cur_entry_error;
285 if (bch2_journal_error(j))
286 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
288 if (!fifo_free(&j->pin))
289 return JOURNAL_ERR_journal_pin_full;
291 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
292 return JOURNAL_ERR_max_in_flight;
294 BUG_ON(!j->cur_entry_sectors);
297 (journal_cur_seq(j) == j->flushed_seq_ondisk
299 : j->last_flush_write) +
300 msecs_to_jiffies(c->opts.journal_flush_delay);
302 buf->u64s_reserved = j->entry_u64s_reserved;
303 buf->disk_sectors = j->cur_entry_sectors;
304 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
306 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
307 journal_entry_overhead(j);
308 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
310 if (u64s <= (ssize_t) j->early_journal_entries.nr)
311 return JOURNAL_ERR_journal_full;
313 if (fifo_empty(&j->pin) && j->reclaim_thread)
314 wake_up_process(j->reclaim_thread);
317 * The fifo_push() needs to happen at the same time as j->seq is
318 * incremented for journal_last_seq() to be calculated correctly
320 atomic64_inc(&j->seq);
321 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
323 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
325 bkey_extent_init(&buf->key);
326 buf->noflush = false;
327 buf->must_flush = false;
328 buf->separate_flush = false;
331 memset(buf->data, 0, sizeof(*buf->data));
332 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
335 if (j->early_journal_entries.nr) {
336 memcpy(buf->data->_data, j->early_journal_entries.data,
337 j->early_journal_entries.nr * sizeof(u64));
338 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
342 * Must be set before marking the journal entry as open:
344 j->cur_entry_u64s = u64s;
346 v = atomic64_read(&j->reservations.counter);
350 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
353 BUG_ON(journal_state_count(new, new.idx));
354 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
356 journal_state_inc(&new);
358 /* Handle any already added entries */
359 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
360 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
361 old.v, new.v)) != old.v);
363 if (j->res_get_blocked_start)
364 bch2_time_stats_update(j->blocked_time,
365 j->res_get_blocked_start);
366 j->res_get_blocked_start = 0;
368 mod_delayed_work(c->io_complete_wq,
370 msecs_to_jiffies(c->opts.journal_flush_delay));
373 if (j->early_journal_entries.nr)
374 darray_exit(&j->early_journal_entries);
378 static bool journal_quiesced(struct journal *j)
380 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
383 journal_entry_close(j);
387 static void journal_quiesce(struct journal *j)
389 wait_event(j->wait, journal_quiesced(j));
392 static void journal_write_work(struct work_struct *work)
394 struct journal *j = container_of(work, struct journal, write_work.work);
395 struct bch_fs *c = container_of(j, struct bch_fs, journal);
399 if (!__journal_entry_is_open(j->reservations))
402 delta = journal_cur_buf(j)->expires - jiffies;
405 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
407 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
409 spin_unlock(&j->lock);
412 static int __journal_res_get(struct journal *j, struct journal_res *res,
415 struct bch_fs *c = container_of(j, struct bch_fs, journal);
416 struct journal_buf *buf;
420 if (journal_res_get_fast(j, res, flags))
423 if (bch2_journal_error(j))
424 return -BCH_ERR_erofs_journal_err;
428 /* check once more in case somebody else shut things down... */
429 if (bch2_journal_error(j)) {
430 spin_unlock(&j->lock);
431 return -BCH_ERR_erofs_journal_err;
435 * Recheck after taking the lock, so we don't race with another thread
436 * that just did journal_entry_open() and call journal_entry_close()
439 if (journal_res_get_fast(j, res, flags)) {
440 spin_unlock(&j->lock);
444 if ((flags & JOURNAL_WATERMARK_MASK) < j->watermark) {
446 * Don't want to close current journal entry, just need to
449 ret = JOURNAL_ERR_journal_full;
454 * If we couldn't get a reservation because the current buf filled up,
455 * and we had room for a bigger entry on disk, signal that we want to
456 * realloc the journal bufs:
458 buf = journal_cur_buf(j);
459 if (journal_entry_is_open(j) &&
460 buf->buf_size >> 9 < buf->disk_sectors &&
461 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
462 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
464 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
465 ret = journal_entry_open(j);
467 if (ret == JOURNAL_ERR_max_in_flight)
468 trace_and_count(c, journal_entry_full, c);
470 if ((ret && ret != JOURNAL_ERR_insufficient_devices) &&
471 !j->res_get_blocked_start) {
472 j->res_get_blocked_start = local_clock() ?: 1;
473 trace_and_count(c, journal_full, c);
476 can_discard = j->can_discard;
477 spin_unlock(&j->lock);
481 if (journal_error_check_stuck(j, ret, flags))
482 ret = -BCH_ERR_journal_res_get_blocked;
485 * Journal is full - can't rely on reclaim from work item due to
488 if ((ret == JOURNAL_ERR_journal_full ||
489 ret == JOURNAL_ERR_journal_pin_full) &&
490 !(flags & JOURNAL_RES_GET_NONBLOCK)) {
492 bch2_journal_do_discards(j);
496 if (mutex_trylock(&j->reclaim_lock)) {
497 bch2_journal_reclaim(j);
498 mutex_unlock(&j->reclaim_lock);
502 return ret == JOURNAL_ERR_insufficient_devices
504 : -BCH_ERR_journal_res_get_blocked;
508 * Essentially the entry function to the journaling code. When bcachefs is doing
509 * a btree insert, it calls this function to get the current journal write.
510 * Journal write is the structure used set up journal writes. The calling
511 * function will then add its keys to the structure, queuing them for the next
514 * To ensure forward progress, the current task must not be holding any
515 * btree node write locks.
517 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
522 closure_wait_event(&j->async_wait,
523 (ret = __journal_res_get(j, res, flags)) !=
524 -BCH_ERR_journal_res_get_blocked||
525 (flags & JOURNAL_RES_GET_NONBLOCK));
529 /* journal_preres: */
531 static bool journal_preres_available(struct journal *j,
532 struct journal_preres *res,
536 bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags, true);
538 if (!ret && mutex_trylock(&j->reclaim_lock)) {
539 bch2_journal_reclaim(j);
540 mutex_unlock(&j->reclaim_lock);
546 int __bch2_journal_preres_get(struct journal *j,
547 struct journal_preres *res,
553 closure_wait_event(&j->preres_wait,
554 (ret = bch2_journal_error(j)) ||
555 journal_preres_available(j, res, new_u64s, flags));
559 /* journal_entry_res: */
561 void bch2_journal_entry_res_resize(struct journal *j,
562 struct journal_entry_res *res,
565 union journal_res_state state;
566 int d = new_u64s - res->u64s;
570 j->entry_u64s_reserved += d;
574 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
576 state = READ_ONCE(j->reservations);
578 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
579 state.cur_entry_offset > j->cur_entry_u64s) {
580 j->cur_entry_u64s += d;
582 * Not enough room in current journal entry, have to flush it:
584 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
586 journal_cur_buf(j)->u64s_reserved += d;
589 spin_unlock(&j->lock);
593 /* journal flushing: */
596 * bch2_journal_flush_seq_async - wait for a journal entry to be written
598 * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
601 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
602 struct closure *parent)
604 struct journal_buf *buf;
607 if (seq <= j->flushed_seq_ondisk)
612 if (WARN_ONCE(seq > journal_cur_seq(j),
613 "requested to flush journal seq %llu, but currently at %llu",
614 seq, journal_cur_seq(j)))
617 /* Recheck under lock: */
618 if (j->err_seq && seq >= j->err_seq) {
623 if (seq <= j->flushed_seq_ondisk) {
628 /* if seq was written, but not flushed - flush a newer one instead */
629 seq = max(seq, journal_last_unwritten_seq(j));
632 if (seq > journal_cur_seq(j)) {
633 struct journal_res res = { 0 };
635 if (journal_entry_is_open(j))
636 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
638 spin_unlock(&j->lock);
640 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
645 buf = j->buf + (seq & JOURNAL_BUF_MASK);
646 buf->must_flush = true;
648 if (!buf->flush_time) {
649 buf->flush_time = local_clock() ?: 1;
650 buf->expires = jiffies;
653 if (parent && !closure_wait(&buf->wait, parent))
656 bch2_journal_res_put(j, &res);
663 * if write was kicked off without a flush, flush the next sequence
666 buf = journal_seq_to_buf(j, seq);
669 goto recheck_need_open;
672 buf->must_flush = true;
674 if (parent && !closure_wait(&buf->wait, parent))
677 if (seq == journal_cur_seq(j))
678 journal_entry_want_write(j);
680 spin_unlock(&j->lock);
684 int bch2_journal_flush_seq(struct journal *j, u64 seq)
686 u64 start_time = local_clock();
690 * Don't update time_stats when @seq is already flushed:
692 if (seq <= j->flushed_seq_ondisk)
695 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
698 bch2_time_stats_update(j->flush_seq_time, start_time);
700 return ret ?: ret2 < 0 ? ret2 : 0;
704 * bch2_journal_flush_async - if there is an open journal entry, or a journal
705 * still being written, write it and wait for the write to complete
707 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
709 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
712 int bch2_journal_flush(struct journal *j)
714 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
718 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
721 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
723 struct bch_fs *c = container_of(j, struct bch_fs, journal);
727 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
730 if (seq <= c->journal.flushed_seq_ondisk)
734 if (seq <= c->journal.flushed_seq_ondisk)
737 for (unwritten_seq = journal_last_unwritten_seq(j);
740 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
742 /* journal write is already in flight, and was a flush write: */
743 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
751 spin_unlock(&j->lock);
755 int bch2_journal_meta(struct journal *j)
757 struct journal_buf *buf;
758 struct journal_res res;
761 memset(&res, 0, sizeof(res));
763 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
767 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
768 buf->must_flush = true;
770 if (!buf->flush_time) {
771 buf->flush_time = local_clock() ?: 1;
772 buf->expires = jiffies;
775 bch2_journal_res_put(j, &res);
777 return bch2_journal_flush_seq(j, res.seq);
780 /* block/unlock the journal: */
782 void bch2_journal_unblock(struct journal *j)
786 spin_unlock(&j->lock);
791 void bch2_journal_block(struct journal *j)
795 spin_unlock(&j->lock);
800 /* allocate journal on a device: */
802 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
803 bool new_fs, struct closure *cl)
805 struct bch_fs *c = ca->fs;
806 struct journal_device *ja = &ca->journal;
807 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
808 struct open_bucket **ob = NULL;
810 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
813 BUG_ON(nr <= ja->nr);
815 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
816 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
817 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
818 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
819 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
820 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
824 for (nr_got = 0; nr_got < nr_want; nr_got++) {
826 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
827 if (bu[nr_got] < 0) {
828 ret = -BCH_ERR_ENOSPC_bucket_alloc;
832 ob[nr_got] = bch2_bucket_alloc(c, ca, RESERVE_none, cl);
833 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
837 ret = bch2_trans_run(c,
838 bch2_trans_mark_metadata_bucket(&trans, ca,
839 ob[nr_got]->bucket, BCH_DATA_journal,
840 ca->mi.bucket_size));
842 bch2_open_bucket_put(c, ob[nr_got]);
843 bch_err(c, "error marking new journal buckets: %s", bch2_err_str(ret));
847 bu[nr_got] = ob[nr_got]->bucket;
854 /* Don't return an error if we successfully allocated some buckets: */
858 bch2_journal_flush_all_pins(&c->journal);
859 bch2_journal_block(&c->journal);
860 mutex_lock(&c->sb_lock);
863 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
864 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
866 BUG_ON(ja->discard_idx > ja->nr);
868 pos = ja->discard_idx ?: ja->nr;
870 memmove(new_buckets + pos + nr_got,
872 sizeof(new_buckets[0]) * (ja->nr - pos));
873 memmove(new_bucket_seq + pos + nr_got,
874 new_bucket_seq + pos,
875 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
877 for (i = 0; i < nr_got; i++) {
878 new_buckets[pos + i] = bu[i];
879 new_bucket_seq[pos + i] = 0;
882 nr = ja->nr + nr_got;
884 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
893 spin_lock(&c->journal.lock);
895 swap(new_buckets, ja->buckets);
896 swap(new_bucket_seq, ja->bucket_seq);
899 if (pos <= ja->discard_idx)
900 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
901 if (pos <= ja->dirty_idx_ondisk)
902 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
903 if (pos <= ja->dirty_idx)
904 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
905 if (pos <= ja->cur_idx)
906 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
909 spin_unlock(&c->journal.lock);
912 bch2_journal_unblock(&c->journal);
913 mutex_unlock(&c->sb_lock);
917 for (i = 0; i < nr_got; i++)
919 bch2_trans_mark_metadata_bucket(&trans, ca,
920 bu[i], BCH_DATA_free, 0));
923 for (i = 0; i < nr_got; i++)
924 bch2_open_bucket_put(c, ob[i]);
926 kfree(new_bucket_seq);
934 * Allocate more journal space at runtime - not currently making use if it, but
937 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
940 struct journal_device *ja = &ca->journal;
944 closure_init_stack(&cl);
946 down_write(&c->state_lock);
948 /* don't handle reducing nr of buckets yet: */
952 while (ja->nr < nr) {
953 struct disk_reservation disk_res = { 0, 0 };
956 * note: journal buckets aren't really counted as _sectors_ used yet, so
957 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
958 * when space used goes up without a reservation - but we do need the
959 * reservation to ensure we'll actually be able to allocate:
961 * XXX: that's not right, disk reservations only ensure a
962 * filesystem-wide allocation will succeed, this is a device
963 * specific allocation - we can hang here:
966 ret = bch2_disk_reservation_get(c, &disk_res,
967 bucket_to_sector(ca, nr - ja->nr), 1, 0);
971 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
973 bch2_disk_reservation_put(c, &disk_res);
977 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
982 bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
984 up_write(&c->state_lock);
988 int bch2_dev_journal_alloc(struct bch_dev *ca)
992 if (dynamic_fault("bcachefs:add:journal_alloc"))
993 return -BCH_ERR_ENOMEM_set_nr_journal_buckets;
995 /* 1/128th of the device by default: */
996 nr = ca->mi.nbuckets >> 7;
999 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1002 nr = clamp_t(unsigned, nr,
1003 BCH_JOURNAL_BUCKETS_MIN,
1005 (1 << 24) / ca->mi.bucket_size));
1007 return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1010 /* startup/shutdown: */
1012 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1017 spin_lock(&j->lock);
1018 for (seq = journal_last_unwritten_seq(j);
1019 seq <= journal_cur_seq(j) && !ret;
1021 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1023 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1026 spin_unlock(&j->lock);
1031 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1033 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1036 void bch2_fs_journal_stop(struct journal *j)
1038 bch2_journal_reclaim_stop(j);
1039 bch2_journal_flush_all_pins(j);
1041 wait_event(j->wait, journal_entry_close(j));
1044 * Always write a new journal entry, to make sure the clock hands are up
1045 * to date (and match the superblock)
1047 bch2_journal_meta(j);
1051 BUG_ON(!bch2_journal_error(j) &&
1052 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1053 j->last_empty_seq != journal_cur_seq(j));
1055 cancel_delayed_work_sync(&j->write_work);
1058 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1060 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1061 struct journal_entry_pin_list *p;
1062 struct journal_replay *i, **_i;
1063 struct genradix_iter iter;
1064 bool had_entries = false;
1066 u64 last_seq = cur_seq, nr, seq;
1068 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1071 if (!i || i->ignore)
1074 last_seq = le64_to_cpu(i->j.last_seq);
1078 nr = cur_seq - last_seq;
1080 if (nr + 1 > j->pin.size) {
1082 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1084 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1085 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1089 j->replay_journal_seq = last_seq;
1090 j->replay_journal_seq_end = cur_seq;
1091 j->last_seq_ondisk = last_seq;
1092 j->flushed_seq_ondisk = cur_seq - 1;
1093 j->seq_ondisk = cur_seq - 1;
1094 j->pin.front = last_seq;
1095 j->pin.back = cur_seq;
1096 atomic64_set(&j->seq, cur_seq - 1);
1098 fifo_for_each_entry_ptr(p, &j->pin, seq)
1099 journal_pin_list_init(p, 1);
1101 genradix_for_each(&c->journal_entries, iter, _i) {
1104 if (!i || i->ignore)
1107 seq = le64_to_cpu(i->j.seq);
1108 BUG_ON(seq >= cur_seq);
1113 if (journal_entry_empty(&i->j))
1114 j->last_empty_seq = le64_to_cpu(i->j.seq);
1116 p = journal_seq_pin(j, seq);
1119 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1120 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1126 j->last_empty_seq = cur_seq;
1128 spin_lock(&j->lock);
1130 set_bit(JOURNAL_STARTED, &j->flags);
1131 j->last_flush_write = jiffies;
1133 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1134 j->reservations.unwritten_idx++;
1136 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1138 bch2_journal_space_available(j);
1139 spin_unlock(&j->lock);
1141 return bch2_journal_reclaim_start(j);
1146 void bch2_dev_journal_exit(struct bch_dev *ca)
1148 kfree(ca->journal.bio);
1149 kfree(ca->journal.buckets);
1150 kfree(ca->journal.bucket_seq);
1152 ca->journal.bio = NULL;
1153 ca->journal.buckets = NULL;
1154 ca->journal.bucket_seq = NULL;
1157 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1159 struct journal_device *ja = &ca->journal;
1160 struct bch_sb_field_journal *journal_buckets =
1161 bch2_sb_get_journal(sb);
1162 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1163 bch2_sb_get_journal_v2(sb);
1164 unsigned i, nr_bvecs;
1168 if (journal_buckets_v2) {
1169 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1171 for (i = 0; i < nr; i++)
1172 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1173 } else if (journal_buckets) {
1174 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1177 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1178 if (!ja->bucket_seq)
1179 return -BCH_ERR_ENOMEM_dev_journal_init;
1181 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1183 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1184 if (!ca->journal.bio)
1185 return -BCH_ERR_ENOMEM_dev_journal_init;
1187 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1189 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1191 return -BCH_ERR_ENOMEM_dev_journal_init;
1193 if (journal_buckets_v2) {
1194 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1195 unsigned j, dst = 0;
1197 for (i = 0; i < nr; i++)
1198 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1199 ja->buckets[dst++] =
1200 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1201 } else if (journal_buckets) {
1202 for (i = 0; i < ja->nr; i++)
1203 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1209 void bch2_fs_journal_exit(struct journal *j)
1213 darray_exit(&j->early_journal_entries);
1215 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1216 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1220 int bch2_fs_journal_init(struct journal *j)
1222 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1223 static struct lock_class_key res_key;
1227 pr_verbose_init(c->opts, "");
1229 spin_lock_init(&j->lock);
1230 spin_lock_init(&j->err_lock);
1231 init_waitqueue_head(&j->wait);
1232 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1233 init_waitqueue_head(&j->reclaim_wait);
1234 init_waitqueue_head(&j->pin_flush_wait);
1235 mutex_init(&j->reclaim_lock);
1236 mutex_init(&j->discard_lock);
1238 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1240 atomic64_set(&j->reservations.counter,
1241 ((union journal_res_state)
1242 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1244 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) {
1245 ret = -BCH_ERR_ENOMEM_journal_pin_fifo;
1249 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1250 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1251 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1252 if (!j->buf[i].data) {
1253 ret = -BCH_ERR_ENOMEM_journal_buf;
1258 j->pin.front = j->pin.back = 1;
1260 pr_verbose_init(c->opts, "ret %i", ret);
1266 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1268 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1269 union journal_res_state s;
1271 unsigned long now = jiffies;
1275 if (!out->nr_tabstops)
1276 printbuf_tabstop_push(out, 24);
1280 s = READ_ONCE(j->reservations);
1282 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1283 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1284 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1285 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1286 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1287 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1288 prt_printf(out, "prereserved:\t\t%u/%u\n", j->prereserved.reserved, j->prereserved.remaining);
1289 prt_printf(out, "watermark:\t\t%s\n", bch2_journal_watermarks[j->watermark]);
1290 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1291 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1292 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1293 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1294 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1295 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1296 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1297 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1298 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1299 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1300 prt_printf(out, "current entry:\t\t");
1302 switch (s.cur_entry_offset) {
1303 case JOURNAL_ENTRY_ERROR_VAL:
1304 prt_printf(out, "error");
1306 case JOURNAL_ENTRY_CLOSED_VAL:
1307 prt_printf(out, "closed");
1310 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1316 for (seq = journal_cur_seq(j);
1317 seq >= journal_last_unwritten_seq(j);
1319 i = seq & JOURNAL_BUF_MASK;
1321 prt_printf(out, "unwritten entry:");
1323 prt_printf(out, "%llu", seq);
1325 printbuf_indent_add(out, 2);
1327 prt_printf(out, "refcount:");
1329 prt_printf(out, "%u", journal_state_count(s, i));
1332 prt_printf(out, "sectors:");
1334 prt_printf(out, "%u", j->buf[i].sectors);
1337 prt_printf(out, "expires");
1339 prt_printf(out, "%li jiffies", j->buf[i].expires - jiffies);
1342 printbuf_indent_sub(out, 2);
1346 "replay done:\t\t%i\n",
1347 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1349 prt_printf(out, "space:\n");
1350 prt_printf(out, "\tdiscarded\t%u:%u\n",
1351 j->space[journal_space_discarded].next_entry,
1352 j->space[journal_space_discarded].total);
1353 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1354 j->space[journal_space_clean_ondisk].next_entry,
1355 j->space[journal_space_clean_ondisk].total);
1356 prt_printf(out, "\tclean\t\t%u:%u\n",
1357 j->space[journal_space_clean].next_entry,
1358 j->space[journal_space_clean].total);
1359 prt_printf(out, "\ttotal\t\t%u:%u\n",
1360 j->space[journal_space_total].next_entry,
1361 j->space[journal_space_total].total);
1363 for_each_member_device_rcu(ca, c, i,
1364 &c->rw_devs[BCH_DATA_journal]) {
1365 struct journal_device *ja = &ca->journal;
1367 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1373 prt_printf(out, "dev %u:\n", i);
1374 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1375 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1376 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1377 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1378 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1379 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1380 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1388 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1390 spin_lock(&j->lock);
1391 __bch2_journal_debug_to_text(out, j);
1392 spin_unlock(&j->lock);
1395 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1397 struct journal_entry_pin_list *pin_list;
1398 struct journal_entry_pin *pin;
1401 spin_lock(&j->lock);
1402 *seq = max(*seq, j->pin.front);
1404 if (*seq >= j->pin.back) {
1405 spin_unlock(&j->lock);
1411 pin_list = journal_seq_pin(j, *seq);
1413 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1415 printbuf_indent_add(out, 2);
1417 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1418 list_for_each_entry(pin, &pin_list->list[i], list) {
1419 prt_printf(out, "\t%px %ps", pin, pin->flush);
1423 if (!list_empty(&pin_list->flushed)) {
1424 prt_printf(out, "flushed:");
1428 list_for_each_entry(pin, &pin_list->flushed, list) {
1429 prt_printf(out, "\t%px %ps", pin, pin->flush);
1433 printbuf_indent_sub(out, 2);
1436 spin_unlock(&j->lock);
1441 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1445 while (!bch2_journal_seq_pins_to_text(out, j, &seq))