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);
135 /* journal entry close/open: */
137 void __bch2_journal_buf_put(struct journal *j)
139 struct bch_fs *c = container_of(j, struct bch_fs, journal);
141 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
145 * Returns true if journal entry is now closed:
147 * We don't close a journal_buf until the next journal_buf is finished writing,
148 * and can be opened again - this also initializes the next journal_buf:
150 static void __journal_entry_close(struct journal *j, unsigned closed_val)
152 struct bch_fs *c = container_of(j, struct bch_fs, journal);
153 struct journal_buf *buf = journal_cur_buf(j);
154 union journal_res_state old, new;
155 u64 v = atomic64_read(&j->reservations.counter);
158 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
159 closed_val != JOURNAL_ENTRY_ERROR_VAL);
161 lockdep_assert_held(&j->lock);
165 new.cur_entry_offset = closed_val;
167 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
168 old.cur_entry_offset == new.cur_entry_offset)
170 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
171 old.v, new.v)) != old.v);
173 if (!__journal_entry_is_open(old))
176 /* Close out old buffer: */
177 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
179 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
180 buf->u64s_reserved) << c->block_bits;
181 BUG_ON(sectors > buf->sectors);
182 buf->sectors = sectors;
185 * We have to set last_seq here, _before_ opening a new journal entry:
187 * A threads may replace an old pin with a new pin on their current
188 * journal reservation - the expectation being that the journal will
189 * contain either what the old pin protected or what the new pin
192 * After the old pin is dropped journal_last_seq() won't include the old
193 * pin, so we can only write the updated last_seq on the entry that
194 * contains whatever the new pin protects.
196 * Restated, we can _not_ update last_seq for a given entry if there
197 * could be a newer entry open with reservations/pins that have been
200 * Hence, we want update/set last_seq on the current journal entry right
201 * before we open a new one:
203 buf->last_seq = journal_last_seq(j);
204 buf->data->last_seq = cpu_to_le64(buf->last_seq);
205 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
207 __bch2_journal_pin_put(j, le64_to_cpu(buf->data->seq));
209 cancel_delayed_work(&j->write_work);
211 bch2_journal_space_available(j);
213 bch2_journal_buf_put(j, old.idx);
216 void bch2_journal_halt(struct journal *j)
219 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL);
221 j->err_seq = journal_cur_seq(j);
223 spin_unlock(&j->lock);
226 static bool journal_entry_want_write(struct journal *j)
228 bool ret = !journal_entry_is_open(j) ||
229 journal_cur_seq(j) == journal_last_unwritten_seq(j);
231 /* Don't close it yet if we already have a write in flight: */
233 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
234 else if (nr_unwritten_journal_entries(j)) {
235 struct journal_buf *buf = journal_cur_buf(j);
237 if (!buf->flush_time) {
238 buf->flush_time = local_clock() ?: 1;
239 buf->expires = jiffies;
246 static bool journal_entry_close(struct journal *j)
251 ret = journal_entry_want_write(j);
252 spin_unlock(&j->lock);
258 * should _only_ called from journal_res_get() - when we actually want a
259 * journal reservation - journal entry is open means journal is dirty:
261 static int journal_entry_open(struct journal *j)
263 struct bch_fs *c = container_of(j, struct bch_fs, journal);
264 struct journal_buf *buf = j->buf +
265 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
266 union journal_res_state old, new;
270 lockdep_assert_held(&j->lock);
271 BUG_ON(journal_entry_is_open(j));
272 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
275 return JOURNAL_ERR_blocked;
277 if (j->cur_entry_error)
278 return j->cur_entry_error;
280 if (bch2_journal_error(j))
281 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
283 if (!fifo_free(&j->pin))
284 return JOURNAL_ERR_journal_pin_full;
286 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
287 return JOURNAL_ERR_max_in_flight;
289 BUG_ON(!j->cur_entry_sectors);
292 (journal_cur_seq(j) == j->flushed_seq_ondisk
294 : j->last_flush_write) +
295 msecs_to_jiffies(c->opts.journal_flush_delay);
297 buf->u64s_reserved = j->entry_u64s_reserved;
298 buf->disk_sectors = j->cur_entry_sectors;
299 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
301 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
302 journal_entry_overhead(j);
303 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
305 if (u64s <= (ssize_t) j->early_journal_entries.nr)
306 return JOURNAL_ERR_journal_full;
308 if (fifo_empty(&j->pin) && j->reclaim_thread)
309 wake_up_process(j->reclaim_thread);
312 * The fifo_push() needs to happen at the same time as j->seq is
313 * incremented for journal_last_seq() to be calculated correctly
315 atomic64_inc(&j->seq);
316 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
318 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
320 bkey_extent_init(&buf->key);
321 buf->noflush = false;
322 buf->must_flush = false;
323 buf->separate_flush = false;
326 memset(buf->data, 0, sizeof(*buf->data));
327 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
330 if (j->early_journal_entries.nr) {
331 memcpy(buf->data->_data, j->early_journal_entries.data,
332 j->early_journal_entries.nr * sizeof(u64));
333 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
337 * Must be set before marking the journal entry as open:
339 j->cur_entry_u64s = u64s;
341 v = atomic64_read(&j->reservations.counter);
345 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
348 BUG_ON(journal_state_count(new, new.idx));
349 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
351 journal_state_inc(&new);
353 /* Handle any already added entries */
354 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
355 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
356 old.v, new.v)) != old.v);
358 if (j->res_get_blocked_start)
359 bch2_time_stats_update(j->blocked_time,
360 j->res_get_blocked_start);
361 j->res_get_blocked_start = 0;
363 mod_delayed_work(c->io_complete_wq,
365 msecs_to_jiffies(c->opts.journal_flush_delay));
368 if (j->early_journal_entries.nr)
369 darray_exit(&j->early_journal_entries);
373 static bool journal_quiesced(struct journal *j)
375 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
378 journal_entry_close(j);
382 static void journal_quiesce(struct journal *j)
384 wait_event(j->wait, journal_quiesced(j));
387 static void journal_write_work(struct work_struct *work)
389 struct journal *j = container_of(work, struct journal, write_work.work);
390 struct bch_fs *c = container_of(j, struct bch_fs, journal);
394 if (!__journal_entry_is_open(j->reservations))
397 delta = journal_cur_buf(j)->expires - jiffies;
400 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
402 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
404 spin_unlock(&j->lock);
407 static int __journal_res_get(struct journal *j, struct journal_res *res,
410 struct bch_fs *c = container_of(j, struct bch_fs, journal);
411 struct journal_buf *buf;
415 if (journal_res_get_fast(j, res, flags))
418 if (bch2_journal_error(j))
419 return -BCH_ERR_erofs_journal_err;
423 /* check once more in case somebody else shut things down... */
424 if (bch2_journal_error(j)) {
425 spin_unlock(&j->lock);
426 return -BCH_ERR_erofs_journal_err;
430 * Recheck after taking the lock, so we don't race with another thread
431 * that just did journal_entry_open() and call journal_entry_close()
434 if (journal_res_get_fast(j, res, flags)) {
435 spin_unlock(&j->lock);
439 if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
441 * Don't want to close current journal entry, just need to
444 ret = JOURNAL_ERR_journal_full;
449 * If we couldn't get a reservation because the current buf filled up,
450 * and we had room for a bigger entry on disk, signal that we want to
451 * realloc the journal bufs:
453 buf = journal_cur_buf(j);
454 if (journal_entry_is_open(j) &&
455 buf->buf_size >> 9 < buf->disk_sectors &&
456 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
457 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
459 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
460 ret = journal_entry_open(j);
462 if (ret == JOURNAL_ERR_max_in_flight)
463 trace_and_count(c, journal_entry_full, c);
465 if ((ret && ret != JOURNAL_ERR_insufficient_devices) &&
466 !j->res_get_blocked_start) {
467 j->res_get_blocked_start = local_clock() ?: 1;
468 trace_and_count(c, journal_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_preres: */
525 static bool journal_preres_available(struct journal *j,
526 struct journal_preres *res,
530 bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags, true);
532 if (!ret && mutex_trylock(&j->reclaim_lock)) {
533 bch2_journal_reclaim(j);
534 mutex_unlock(&j->reclaim_lock);
540 int __bch2_journal_preres_get(struct journal *j,
541 struct journal_preres *res,
547 closure_wait_event(&j->preres_wait,
548 (ret = bch2_journal_error(j)) ||
549 journal_preres_available(j, res, new_u64s, flags));
553 /* journal_entry_res: */
555 void bch2_journal_entry_res_resize(struct journal *j,
556 struct journal_entry_res *res,
559 union journal_res_state state;
560 int d = new_u64s - res->u64s;
564 j->entry_u64s_reserved += d;
568 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
570 state = READ_ONCE(j->reservations);
572 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
573 state.cur_entry_offset > j->cur_entry_u64s) {
574 j->cur_entry_u64s += d;
576 * Not enough room in current journal entry, have to flush it:
578 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
580 journal_cur_buf(j)->u64s_reserved += d;
583 spin_unlock(&j->lock);
587 /* journal flushing: */
590 * bch2_journal_flush_seq_async - wait for a journal entry to be written
592 * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
595 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
596 struct closure *parent)
598 struct journal_buf *buf;
601 if (seq <= j->flushed_seq_ondisk)
606 if (WARN_ONCE(seq > journal_cur_seq(j),
607 "requested to flush journal seq %llu, but currently at %llu",
608 seq, journal_cur_seq(j)))
611 /* Recheck under lock: */
612 if (j->err_seq && seq >= j->err_seq) {
617 if (seq <= j->flushed_seq_ondisk) {
622 /* if seq was written, but not flushed - flush a newer one instead */
623 seq = max(seq, journal_last_unwritten_seq(j));
626 if (seq > journal_cur_seq(j)) {
627 struct journal_res res = { 0 };
629 if (journal_entry_is_open(j))
630 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
632 spin_unlock(&j->lock);
634 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
639 buf = j->buf + (seq & JOURNAL_BUF_MASK);
640 buf->must_flush = true;
642 if (!buf->flush_time) {
643 buf->flush_time = local_clock() ?: 1;
644 buf->expires = jiffies;
647 if (parent && !closure_wait(&buf->wait, parent))
650 bch2_journal_res_put(j, &res);
657 * if write was kicked off without a flush, flush the next sequence
660 buf = journal_seq_to_buf(j, seq);
663 goto recheck_need_open;
666 buf->must_flush = true;
668 if (parent && !closure_wait(&buf->wait, parent))
671 if (seq == journal_cur_seq(j))
672 journal_entry_want_write(j);
674 spin_unlock(&j->lock);
678 int bch2_journal_flush_seq(struct journal *j, u64 seq)
680 u64 start_time = local_clock();
684 * Don't update time_stats when @seq is already flushed:
686 if (seq <= j->flushed_seq_ondisk)
689 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
692 bch2_time_stats_update(j->flush_seq_time, start_time);
694 return ret ?: ret2 < 0 ? ret2 : 0;
698 * bch2_journal_flush_async - if there is an open journal entry, or a journal
699 * still being written, write it and wait for the write to complete
701 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
703 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
706 int bch2_journal_flush(struct journal *j)
708 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
712 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
715 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
717 struct bch_fs *c = container_of(j, struct bch_fs, journal);
721 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
724 if (seq <= c->journal.flushed_seq_ondisk)
728 if (seq <= c->journal.flushed_seq_ondisk)
731 for (unwritten_seq = journal_last_unwritten_seq(j);
734 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
736 /* journal write is already in flight, and was a flush write: */
737 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
745 spin_unlock(&j->lock);
749 int bch2_journal_meta(struct journal *j)
751 struct journal_buf *buf;
752 struct journal_res res;
755 memset(&res, 0, sizeof(res));
757 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
761 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
762 buf->must_flush = true;
764 if (!buf->flush_time) {
765 buf->flush_time = local_clock() ?: 1;
766 buf->expires = jiffies;
769 bch2_journal_res_put(j, &res);
771 return bch2_journal_flush_seq(j, res.seq);
774 /* block/unlock the journal: */
776 void bch2_journal_unblock(struct journal *j)
780 spin_unlock(&j->lock);
785 void bch2_journal_block(struct journal *j)
789 spin_unlock(&j->lock);
794 /* allocate journal on a device: */
796 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
797 bool new_fs, struct closure *cl)
799 struct bch_fs *c = ca->fs;
800 struct journal_device *ja = &ca->journal;
801 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
802 struct open_bucket **ob = NULL;
804 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
807 BUG_ON(nr <= ja->nr);
809 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
810 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
811 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
812 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
813 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
814 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
818 for (nr_got = 0; nr_got < nr_want; nr_got++) {
820 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
821 if (bu[nr_got] < 0) {
822 ret = -BCH_ERR_ENOSPC_bucket_alloc;
826 ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
827 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
831 ret = bch2_trans_run(c,
832 bch2_trans_mark_metadata_bucket(&trans, ca,
833 ob[nr_got]->bucket, BCH_DATA_journal,
834 ca->mi.bucket_size));
836 bch2_open_bucket_put(c, ob[nr_got]);
837 bch_err(c, "error marking new journal buckets: %s", bch2_err_str(ret));
841 bu[nr_got] = ob[nr_got]->bucket;
848 /* Don't return an error if we successfully allocated some buckets: */
852 bch2_journal_flush_all_pins(&c->journal);
853 bch2_journal_block(&c->journal);
854 mutex_lock(&c->sb_lock);
857 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
858 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
860 BUG_ON(ja->discard_idx > ja->nr);
862 pos = ja->discard_idx ?: ja->nr;
864 memmove(new_buckets + pos + nr_got,
866 sizeof(new_buckets[0]) * (ja->nr - pos));
867 memmove(new_bucket_seq + pos + nr_got,
868 new_bucket_seq + pos,
869 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
871 for (i = 0; i < nr_got; i++) {
872 new_buckets[pos + i] = bu[i];
873 new_bucket_seq[pos + i] = 0;
876 nr = ja->nr + nr_got;
878 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
887 spin_lock(&c->journal.lock);
889 swap(new_buckets, ja->buckets);
890 swap(new_bucket_seq, ja->bucket_seq);
893 if (pos <= ja->discard_idx)
894 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
895 if (pos <= ja->dirty_idx_ondisk)
896 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
897 if (pos <= ja->dirty_idx)
898 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
899 if (pos <= ja->cur_idx)
900 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
903 spin_unlock(&c->journal.lock);
906 bch2_journal_unblock(&c->journal);
907 mutex_unlock(&c->sb_lock);
911 for (i = 0; i < nr_got; i++)
913 bch2_trans_mark_metadata_bucket(&trans, ca,
914 bu[i], BCH_DATA_free, 0));
917 for (i = 0; i < nr_got; i++)
918 bch2_open_bucket_put(c, ob[i]);
920 kfree(new_bucket_seq);
928 * Allocate more journal space at runtime - not currently making use if it, but
931 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
934 struct journal_device *ja = &ca->journal;
938 closure_init_stack(&cl);
940 down_write(&c->state_lock);
942 /* don't handle reducing nr of buckets yet: */
946 while (ja->nr < nr) {
947 struct disk_reservation disk_res = { 0, 0 };
950 * note: journal buckets aren't really counted as _sectors_ used yet, so
951 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
952 * when space used goes up without a reservation - but we do need the
953 * reservation to ensure we'll actually be able to allocate:
955 * XXX: that's not right, disk reservations only ensure a
956 * filesystem-wide allocation will succeed, this is a device
957 * specific allocation - we can hang here:
960 ret = bch2_disk_reservation_get(c, &disk_res,
961 bucket_to_sector(ca, nr - ja->nr), 1, 0);
965 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
967 bch2_disk_reservation_put(c, &disk_res);
971 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
978 up_write(&c->state_lock);
982 int bch2_dev_journal_alloc(struct bch_dev *ca)
987 if (dynamic_fault("bcachefs:add:journal_alloc")) {
988 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
992 /* 1/128th of the device by default: */
993 nr = ca->mi.nbuckets >> 7;
996 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
999 nr = clamp_t(unsigned, nr,
1000 BCH_JOURNAL_BUCKETS_MIN,
1002 (1 << 24) / ca->mi.bucket_size));
1004 ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1007 bch_err_fn(ca, ret);
1011 /* startup/shutdown: */
1013 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1018 spin_lock(&j->lock);
1019 for (seq = journal_last_unwritten_seq(j);
1020 seq <= journal_cur_seq(j) && !ret;
1022 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1024 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1027 spin_unlock(&j->lock);
1032 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1034 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1037 void bch2_fs_journal_stop(struct journal *j)
1039 bch2_journal_reclaim_stop(j);
1040 bch2_journal_flush_all_pins(j);
1042 wait_event(j->wait, journal_entry_close(j));
1045 * Always write a new journal entry, to make sure the clock hands are up
1046 * to date (and match the superblock)
1048 bch2_journal_meta(j);
1052 BUG_ON(!bch2_journal_error(j) &&
1053 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1054 j->last_empty_seq != journal_cur_seq(j));
1056 cancel_delayed_work_sync(&j->write_work);
1059 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1061 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1062 struct journal_entry_pin_list *p;
1063 struct journal_replay *i, **_i;
1064 struct genradix_iter iter;
1065 bool had_entries = false;
1067 u64 last_seq = cur_seq, nr, seq;
1069 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1072 if (!i || i->ignore)
1075 last_seq = le64_to_cpu(i->j.last_seq);
1079 nr = cur_seq - last_seq;
1081 if (nr + 1 > j->pin.size) {
1083 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1085 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1086 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1090 j->replay_journal_seq = last_seq;
1091 j->replay_journal_seq_end = cur_seq;
1092 j->last_seq_ondisk = last_seq;
1093 j->flushed_seq_ondisk = cur_seq - 1;
1094 j->seq_ondisk = cur_seq - 1;
1095 j->pin.front = last_seq;
1096 j->pin.back = cur_seq;
1097 atomic64_set(&j->seq, cur_seq - 1);
1099 fifo_for_each_entry_ptr(p, &j->pin, seq)
1100 journal_pin_list_init(p, 1);
1102 genradix_for_each(&c->journal_entries, iter, _i) {
1105 if (!i || i->ignore)
1108 seq = le64_to_cpu(i->j.seq);
1109 BUG_ON(seq >= cur_seq);
1114 if (journal_entry_empty(&i->j))
1115 j->last_empty_seq = le64_to_cpu(i->j.seq);
1117 p = journal_seq_pin(j, seq);
1120 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1121 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1127 j->last_empty_seq = cur_seq;
1129 spin_lock(&j->lock);
1131 set_bit(JOURNAL_STARTED, &j->flags);
1132 j->last_flush_write = jiffies;
1134 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1135 j->reservations.unwritten_idx++;
1137 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1139 bch2_journal_space_available(j);
1140 spin_unlock(&j->lock);
1142 return bch2_journal_reclaim_start(j);
1147 void bch2_dev_journal_exit(struct bch_dev *ca)
1149 kfree(ca->journal.bio);
1150 kfree(ca->journal.buckets);
1151 kfree(ca->journal.bucket_seq);
1153 ca->journal.bio = NULL;
1154 ca->journal.buckets = NULL;
1155 ca->journal.bucket_seq = NULL;
1158 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1160 struct journal_device *ja = &ca->journal;
1161 struct bch_sb_field_journal *journal_buckets =
1162 bch2_sb_get_journal(sb);
1163 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1164 bch2_sb_get_journal_v2(sb);
1165 unsigned i, nr_bvecs;
1169 if (journal_buckets_v2) {
1170 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1172 for (i = 0; i < nr; i++)
1173 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1174 } else if (journal_buckets) {
1175 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1178 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1179 if (!ja->bucket_seq)
1180 return -BCH_ERR_ENOMEM_dev_journal_init;
1182 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1184 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1185 if (!ca->journal.bio)
1186 return -BCH_ERR_ENOMEM_dev_journal_init;
1188 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1190 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1192 return -BCH_ERR_ENOMEM_dev_journal_init;
1194 if (journal_buckets_v2) {
1195 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1196 unsigned j, dst = 0;
1198 for (i = 0; i < nr; i++)
1199 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1200 ja->buckets[dst++] =
1201 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1202 } else if (journal_buckets) {
1203 for (i = 0; i < ja->nr; i++)
1204 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1210 void bch2_fs_journal_exit(struct journal *j)
1214 darray_exit(&j->early_journal_entries);
1216 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1217 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1221 int bch2_fs_journal_init(struct journal *j)
1223 static struct lock_class_key res_key;
1226 spin_lock_init(&j->lock);
1227 spin_lock_init(&j->err_lock);
1228 init_waitqueue_head(&j->wait);
1229 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1230 init_waitqueue_head(&j->reclaim_wait);
1231 init_waitqueue_head(&j->pin_flush_wait);
1232 mutex_init(&j->reclaim_lock);
1233 mutex_init(&j->discard_lock);
1235 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1237 atomic64_set(&j->reservations.counter,
1238 ((union journal_res_state)
1239 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1241 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1242 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1244 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1245 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1246 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1247 if (!j->buf[i].data)
1248 return -BCH_ERR_ENOMEM_journal_buf;
1251 j->pin.front = j->pin.back = 1;
1257 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1259 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1260 union journal_res_state s;
1262 unsigned long now = jiffies;
1266 if (!out->nr_tabstops)
1267 printbuf_tabstop_push(out, 24);
1271 s = READ_ONCE(j->reservations);
1273 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1274 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1275 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1276 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1277 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1278 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1279 prt_printf(out, "prereserved:\t\t%u/%u\n", j->prereserved.reserved, j->prereserved.remaining);
1280 prt_printf(out, "watermark:\t\t%s\n", bch2_watermarks[j->watermark]);
1281 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1282 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1283 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1284 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1285 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1286 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1287 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1288 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1289 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1290 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1291 prt_printf(out, "current entry:\t\t");
1293 switch (s.cur_entry_offset) {
1294 case JOURNAL_ENTRY_ERROR_VAL:
1295 prt_printf(out, "error");
1297 case JOURNAL_ENTRY_CLOSED_VAL:
1298 prt_printf(out, "closed");
1301 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1307 for (seq = journal_cur_seq(j);
1308 seq >= journal_last_unwritten_seq(j);
1310 i = seq & JOURNAL_BUF_MASK;
1312 prt_printf(out, "unwritten entry:");
1314 prt_printf(out, "%llu", seq);
1316 printbuf_indent_add(out, 2);
1318 prt_printf(out, "refcount:");
1320 prt_printf(out, "%u", journal_state_count(s, i));
1323 prt_printf(out, "sectors:");
1325 prt_printf(out, "%u", j->buf[i].sectors);
1328 prt_printf(out, "expires");
1330 prt_printf(out, "%li jiffies", j->buf[i].expires - jiffies);
1333 printbuf_indent_sub(out, 2);
1337 "replay done:\t\t%i\n",
1338 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1340 prt_printf(out, "space:\n");
1341 prt_printf(out, "\tdiscarded\t%u:%u\n",
1342 j->space[journal_space_discarded].next_entry,
1343 j->space[journal_space_discarded].total);
1344 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1345 j->space[journal_space_clean_ondisk].next_entry,
1346 j->space[journal_space_clean_ondisk].total);
1347 prt_printf(out, "\tclean\t\t%u:%u\n",
1348 j->space[journal_space_clean].next_entry,
1349 j->space[journal_space_clean].total);
1350 prt_printf(out, "\ttotal\t\t%u:%u\n",
1351 j->space[journal_space_total].next_entry,
1352 j->space[journal_space_total].total);
1354 for_each_member_device_rcu(ca, c, i,
1355 &c->rw_devs[BCH_DATA_journal]) {
1356 struct journal_device *ja = &ca->journal;
1358 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1364 prt_printf(out, "dev %u:\n", i);
1365 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1366 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1367 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1368 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1369 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1370 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1371 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1379 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1381 spin_lock(&j->lock);
1382 __bch2_journal_debug_to_text(out, j);
1383 spin_unlock(&j->lock);
1386 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1388 struct journal_entry_pin_list *pin_list;
1389 struct journal_entry_pin *pin;
1392 spin_lock(&j->lock);
1393 *seq = max(*seq, j->pin.front);
1395 if (*seq >= j->pin.back) {
1396 spin_unlock(&j->lock);
1402 pin_list = journal_seq_pin(j, *seq);
1404 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1406 printbuf_indent_add(out, 2);
1408 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1409 list_for_each_entry(pin, &pin_list->list[i], list) {
1410 prt_printf(out, "\t%px %ps", pin, pin->flush);
1414 if (!list_empty(&pin_list->flushed)) {
1415 prt_printf(out, "flushed:");
1419 list_for_each_entry(pin, &pin_list->flushed, list) {
1420 prt_printf(out, "\t%px %ps", pin, pin->flush);
1424 printbuf_indent_sub(out, 2);
1427 spin_unlock(&j->lock);
1432 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1436 while (!bch2_journal_seq_pins_to_text(out, j, &seq))