1 // SPDX-License-Identifier: GPL-2.0
3 * bcachefs journalling code, for btree insertions
5 * Copyright 2012 Google, Inc.
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
12 #include "btree_update.h"
16 #include "journal_io.h"
17 #include "journal_reclaim.h"
18 #include "journal_sb.h"
19 #include "journal_seq_blacklist.h"
22 static const char * const bch2_journal_errors[] = {
29 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
31 return seq > j->seq_ondisk;
34 static bool __journal_entry_is_open(union journal_res_state state)
36 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
39 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
41 return atomic64_read(&j->seq) - j->seq_ondisk;
44 static bool journal_entry_is_open(struct journal *j)
46 return __journal_entry_is_open(j->reservations);
49 static inline struct journal_buf *
50 journal_seq_to_buf(struct journal *j, u64 seq)
52 struct journal_buf *buf = NULL;
54 EBUG_ON(seq > journal_cur_seq(j));
56 if (journal_seq_unwritten(j, seq)) {
57 buf = j->buf + (seq & JOURNAL_BUF_MASK);
58 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
63 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
67 for (i = 0; i < ARRAY_SIZE(p->list); i++)
68 INIT_LIST_HEAD(&p->list[i]);
69 INIT_LIST_HEAD(&p->flushed);
70 atomic_set(&p->count, count);
75 * Detect stuck journal conditions and trigger shutdown. Technically the journal
76 * can end up stuck for a variety of reasons, such as a blocked I/O, journal
77 * reservation lockup, etc. Since this is a fatal error with potentially
78 * unpredictable characteristics, we want to be fairly conservative before we
79 * decide to shut things down.
81 * Consider the journal stuck when it appears full with no ability to commit
82 * btree transactions, to discard journal buckets, nor acquire priority
83 * (reserved watermark) reservation.
86 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
88 struct bch_fs *c = container_of(j, struct bch_fs, journal);
90 struct printbuf buf = PRINTBUF;
92 if (!(error == JOURNAL_ERR_journal_full ||
93 error == JOURNAL_ERR_journal_pin_full) ||
94 nr_unwritten_journal_entries(j) ||
95 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
100 if (j->can_discard) {
101 spin_unlock(&j->lock);
108 * The journal shutdown path will set ->err_seq, but do it here first to
109 * serialize against concurrent failures and avoid duplicate error
113 spin_unlock(&j->lock);
116 j->err_seq = journal_cur_seq(j);
117 spin_unlock(&j->lock);
119 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
120 bch2_journal_errors[error]);
121 bch2_journal_debug_to_text(&buf, j);
122 bch_err(c, "%s", buf.buf);
124 printbuf_reset(&buf);
125 bch2_journal_pins_to_text(&buf, j);
126 bch_err(c, "Journal pins:\n%s", buf.buf);
136 * Final processing when the last reference of a journal buffer has been
137 * dropped. Drop the pin list reference acquired at journal entry open and write
138 * the buffer, if requested.
140 void bch2_journal_buf_put_final(struct journal *j, u64 seq, bool write)
142 struct bch_fs *c = container_of(j, struct bch_fs, journal);
144 lockdep_assert_held(&j->lock);
146 if (__bch2_journal_pin_put(j, seq))
147 bch2_journal_reclaim_fast(j);
149 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
153 * Returns true if journal entry is now closed:
155 * We don't close a journal_buf until the next journal_buf is finished writing,
156 * and can be opened again - this also initializes the next journal_buf:
158 static void __journal_entry_close(struct journal *j, unsigned closed_val)
160 struct bch_fs *c = container_of(j, struct bch_fs, journal);
161 struct journal_buf *buf = journal_cur_buf(j);
162 union journal_res_state old, new;
163 u64 v = atomic64_read(&j->reservations.counter);
166 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
167 closed_val != JOURNAL_ENTRY_ERROR_VAL);
169 lockdep_assert_held(&j->lock);
173 new.cur_entry_offset = closed_val;
175 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
176 old.cur_entry_offset == new.cur_entry_offset)
178 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
179 old.v, new.v)) != old.v);
181 if (!__journal_entry_is_open(old))
184 /* Close out old buffer: */
185 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
187 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
188 buf->u64s_reserved) << c->block_bits;
189 BUG_ON(sectors > buf->sectors);
190 buf->sectors = sectors;
193 * We have to set last_seq here, _before_ opening a new journal entry:
195 * A threads may replace an old pin with a new pin on their current
196 * journal reservation - the expectation being that the journal will
197 * contain either what the old pin protected or what the new pin
200 * After the old pin is dropped journal_last_seq() won't include the old
201 * pin, so we can only write the updated last_seq on the entry that
202 * contains whatever the new pin protects.
204 * Restated, we can _not_ update last_seq for a given entry if there
205 * could be a newer entry open with reservations/pins that have been
208 * Hence, we want update/set last_seq on the current journal entry right
209 * before we open a new one:
211 buf->last_seq = journal_last_seq(j);
212 buf->data->last_seq = cpu_to_le64(buf->last_seq);
213 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
215 cancel_delayed_work(&j->write_work);
217 bch2_journal_space_available(j);
219 __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
222 void bch2_journal_halt(struct journal *j)
225 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL);
227 j->err_seq = journal_cur_seq(j);
229 spin_unlock(&j->lock);
232 static bool journal_entry_want_write(struct journal *j)
234 bool ret = !journal_entry_is_open(j) ||
235 journal_cur_seq(j) == journal_last_unwritten_seq(j);
237 /* Don't close it yet if we already have a write in flight: */
239 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
240 else if (nr_unwritten_journal_entries(j)) {
241 struct journal_buf *buf = journal_cur_buf(j);
243 if (!buf->flush_time) {
244 buf->flush_time = local_clock() ?: 1;
245 buf->expires = jiffies;
252 static bool journal_entry_close(struct journal *j)
257 ret = journal_entry_want_write(j);
258 spin_unlock(&j->lock);
264 * should _only_ called from journal_res_get() - when we actually want a
265 * journal reservation - journal entry is open means journal is dirty:
267 static int journal_entry_open(struct journal *j)
269 struct bch_fs *c = container_of(j, struct bch_fs, journal);
270 struct journal_buf *buf = j->buf +
271 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
272 union journal_res_state old, new;
276 lockdep_assert_held(&j->lock);
277 BUG_ON(journal_entry_is_open(j));
278 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
281 return JOURNAL_ERR_blocked;
283 if (j->cur_entry_error)
284 return j->cur_entry_error;
286 if (bch2_journal_error(j))
287 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
289 if (!fifo_free(&j->pin))
290 return JOURNAL_ERR_journal_pin_full;
292 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
293 return JOURNAL_ERR_max_in_flight;
295 BUG_ON(!j->cur_entry_sectors);
298 (journal_cur_seq(j) == j->flushed_seq_ondisk
300 : j->last_flush_write) +
301 msecs_to_jiffies(c->opts.journal_flush_delay);
303 buf->u64s_reserved = j->entry_u64s_reserved;
304 buf->disk_sectors = j->cur_entry_sectors;
305 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
307 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
308 journal_entry_overhead(j);
309 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
311 if (u64s <= (ssize_t) j->early_journal_entries.nr)
312 return JOURNAL_ERR_journal_full;
314 if (fifo_empty(&j->pin) && j->reclaim_thread)
315 wake_up_process(j->reclaim_thread);
318 * The fifo_push() needs to happen at the same time as j->seq is
319 * incremented for journal_last_seq() to be calculated correctly
321 atomic64_inc(&j->seq);
322 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
324 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
326 bkey_extent_init(&buf->key);
327 buf->noflush = false;
328 buf->must_flush = false;
329 buf->separate_flush = false;
332 memset(buf->data, 0, sizeof(*buf->data));
333 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
336 if (j->early_journal_entries.nr) {
337 memcpy(buf->data->_data, j->early_journal_entries.data,
338 j->early_journal_entries.nr * sizeof(u64));
339 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
343 * Must be set before marking the journal entry as open:
345 j->cur_entry_u64s = u64s;
347 v = atomic64_read(&j->reservations.counter);
351 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
354 BUG_ON(journal_state_count(new, new.idx));
355 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
357 journal_state_inc(&new);
359 /* Handle any already added entries */
360 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
361 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
362 old.v, new.v)) != old.v);
364 mod_delayed_work(c->io_complete_wq,
366 msecs_to_jiffies(c->opts.journal_flush_delay));
369 if (j->early_journal_entries.nr)
370 darray_exit(&j->early_journal_entries);
374 static bool journal_quiesced(struct journal *j)
376 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
379 journal_entry_close(j);
383 static void journal_quiesce(struct journal *j)
385 wait_event(j->wait, journal_quiesced(j));
388 static void journal_write_work(struct work_struct *work)
390 struct journal *j = container_of(work, struct journal, write_work.work);
391 struct bch_fs *c = container_of(j, struct bch_fs, journal);
395 if (!__journal_entry_is_open(j->reservations))
398 delta = journal_cur_buf(j)->expires - jiffies;
401 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
403 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
405 spin_unlock(&j->lock);
408 static int __journal_res_get(struct journal *j, struct journal_res *res,
411 struct bch_fs *c = container_of(j, struct bch_fs, journal);
412 struct journal_buf *buf;
416 if (journal_res_get_fast(j, res, flags))
419 if (bch2_journal_error(j))
420 return -BCH_ERR_erofs_journal_err;
424 /* check once more in case somebody else shut things down... */
425 if (bch2_journal_error(j)) {
426 spin_unlock(&j->lock);
427 return -BCH_ERR_erofs_journal_err;
431 * Recheck after taking the lock, so we don't race with another thread
432 * that just did journal_entry_open() and call journal_entry_close()
435 if (journal_res_get_fast(j, res, flags)) {
436 spin_unlock(&j->lock);
440 if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
442 * Don't want to close current journal entry, just need to
445 ret = JOURNAL_ERR_journal_full;
450 * If we couldn't get a reservation because the current buf filled up,
451 * and we had room for a bigger entry on disk, signal that we want to
452 * realloc the journal bufs:
454 buf = journal_cur_buf(j);
455 if (journal_entry_is_open(j) &&
456 buf->buf_size >> 9 < buf->disk_sectors &&
457 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
458 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
460 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
461 ret = journal_entry_open(j);
463 if (ret == JOURNAL_ERR_max_in_flight) {
464 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
465 &j->max_in_flight_start, true);
466 trace_and_count(c, journal_entry_full, c);
469 can_discard = j->can_discard;
470 spin_unlock(&j->lock);
474 if (journal_error_check_stuck(j, ret, flags))
475 ret = -BCH_ERR_journal_res_get_blocked;
478 * Journal is full - can't rely on reclaim from work item due to
481 if ((ret == JOURNAL_ERR_journal_full ||
482 ret == JOURNAL_ERR_journal_pin_full) &&
483 !(flags & JOURNAL_RES_GET_NONBLOCK)) {
485 bch2_journal_do_discards(j);
489 if (mutex_trylock(&j->reclaim_lock)) {
490 bch2_journal_reclaim(j);
491 mutex_unlock(&j->reclaim_lock);
495 return ret == JOURNAL_ERR_insufficient_devices
496 ? -BCH_ERR_erofs_journal_err
497 : -BCH_ERR_journal_res_get_blocked;
501 * Essentially the entry function to the journaling code. When bcachefs is doing
502 * a btree insert, it calls this function to get the current journal write.
503 * Journal write is the structure used set up journal writes. The calling
504 * function will then add its keys to the structure, queuing them for the next
507 * To ensure forward progress, the current task must not be holding any
508 * btree node write locks.
510 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
515 closure_wait_event(&j->async_wait,
516 (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
517 (flags & JOURNAL_RES_GET_NONBLOCK));
521 /* journal_entry_res: */
523 void bch2_journal_entry_res_resize(struct journal *j,
524 struct journal_entry_res *res,
527 union journal_res_state state;
528 int d = new_u64s - res->u64s;
532 j->entry_u64s_reserved += d;
536 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
538 state = READ_ONCE(j->reservations);
540 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
541 state.cur_entry_offset > j->cur_entry_u64s) {
542 j->cur_entry_u64s += d;
544 * Not enough room in current journal entry, have to flush it:
546 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
548 journal_cur_buf(j)->u64s_reserved += d;
551 spin_unlock(&j->lock);
555 /* journal flushing: */
558 * bch2_journal_flush_seq_async - wait for a journal entry to be written
561 * @parent: closure object to wait with
562 * Returns: 1 if @seq has already been flushed, 0 if @seq is being flushed,
563 * -EIO if @seq will never be flushed
565 * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
568 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
569 struct closure *parent)
571 struct journal_buf *buf;
574 if (seq <= j->flushed_seq_ondisk)
579 if (WARN_ONCE(seq > journal_cur_seq(j),
580 "requested to flush journal seq %llu, but currently at %llu",
581 seq, journal_cur_seq(j)))
584 /* Recheck under lock: */
585 if (j->err_seq && seq >= j->err_seq) {
590 if (seq <= j->flushed_seq_ondisk) {
595 /* if seq was written, but not flushed - flush a newer one instead */
596 seq = max(seq, journal_last_unwritten_seq(j));
599 if (seq > journal_cur_seq(j)) {
600 struct journal_res res = { 0 };
602 if (journal_entry_is_open(j))
603 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
605 spin_unlock(&j->lock);
607 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
612 buf = j->buf + (seq & JOURNAL_BUF_MASK);
613 buf->must_flush = true;
615 if (!buf->flush_time) {
616 buf->flush_time = local_clock() ?: 1;
617 buf->expires = jiffies;
620 if (parent && !closure_wait(&buf->wait, parent))
623 bch2_journal_res_put(j, &res);
630 * if write was kicked off without a flush, flush the next sequence
633 buf = journal_seq_to_buf(j, seq);
636 goto recheck_need_open;
639 buf->must_flush = true;
641 if (parent && !closure_wait(&buf->wait, parent))
644 if (seq == journal_cur_seq(j))
645 journal_entry_want_write(j);
647 spin_unlock(&j->lock);
651 int bch2_journal_flush_seq(struct journal *j, u64 seq)
653 u64 start_time = local_clock();
657 * Don't update time_stats when @seq is already flushed:
659 if (seq <= j->flushed_seq_ondisk)
662 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
665 bch2_time_stats_update(j->flush_seq_time, start_time);
667 return ret ?: ret2 < 0 ? ret2 : 0;
671 * bch2_journal_flush_async - if there is an open journal entry, or a journal
672 * still being written, write it and wait for the write to complete
674 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
676 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
679 int bch2_journal_flush(struct journal *j)
681 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
685 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
688 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
690 struct bch_fs *c = container_of(j, struct bch_fs, journal);
694 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
697 if (seq <= c->journal.flushed_seq_ondisk)
701 if (seq <= c->journal.flushed_seq_ondisk)
704 for (unwritten_seq = journal_last_unwritten_seq(j);
707 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
709 /* journal write is already in flight, and was a flush write: */
710 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
718 spin_unlock(&j->lock);
722 int bch2_journal_meta(struct journal *j)
724 struct journal_buf *buf;
725 struct journal_res res;
728 memset(&res, 0, sizeof(res));
730 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
734 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
735 buf->must_flush = true;
737 if (!buf->flush_time) {
738 buf->flush_time = local_clock() ?: 1;
739 buf->expires = jiffies;
742 bch2_journal_res_put(j, &res);
744 return bch2_journal_flush_seq(j, res.seq);
747 /* block/unlock the journal: */
749 void bch2_journal_unblock(struct journal *j)
753 spin_unlock(&j->lock);
758 void bch2_journal_block(struct journal *j)
762 spin_unlock(&j->lock);
767 /* allocate journal on a device: */
769 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
770 bool new_fs, struct closure *cl)
772 struct bch_fs *c = ca->fs;
773 struct journal_device *ja = &ca->journal;
774 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
775 struct open_bucket **ob = NULL;
777 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
780 BUG_ON(nr <= ja->nr);
782 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
783 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
784 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
785 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
786 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
787 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
791 for (nr_got = 0; nr_got < nr_want; nr_got++) {
793 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
794 if (bu[nr_got] < 0) {
795 ret = -BCH_ERR_ENOSPC_bucket_alloc;
799 ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
800 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
804 ret = bch2_trans_run(c,
805 bch2_trans_mark_metadata_bucket(trans, ca,
806 ob[nr_got]->bucket, BCH_DATA_journal,
807 ca->mi.bucket_size));
809 bch2_open_bucket_put(c, ob[nr_got]);
810 bch_err_msg(c, ret, "marking new journal buckets");
814 bu[nr_got] = ob[nr_got]->bucket;
821 /* Don't return an error if we successfully allocated some buckets: */
825 bch2_journal_flush_all_pins(&c->journal);
826 bch2_journal_block(&c->journal);
827 mutex_lock(&c->sb_lock);
830 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
831 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
833 BUG_ON(ja->discard_idx > ja->nr);
835 pos = ja->discard_idx ?: ja->nr;
837 memmove(new_buckets + pos + nr_got,
839 sizeof(new_buckets[0]) * (ja->nr - pos));
840 memmove(new_bucket_seq + pos + nr_got,
841 new_bucket_seq + pos,
842 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
844 for (i = 0; i < nr_got; i++) {
845 new_buckets[pos + i] = bu[i];
846 new_bucket_seq[pos + i] = 0;
849 nr = ja->nr + nr_got;
851 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
860 spin_lock(&c->journal.lock);
862 swap(new_buckets, ja->buckets);
863 swap(new_bucket_seq, ja->bucket_seq);
866 if (pos <= ja->discard_idx)
867 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
868 if (pos <= ja->dirty_idx_ondisk)
869 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
870 if (pos <= ja->dirty_idx)
871 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
872 if (pos <= ja->cur_idx)
873 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
876 spin_unlock(&c->journal.lock);
879 bch2_journal_unblock(&c->journal);
880 mutex_unlock(&c->sb_lock);
884 for (i = 0; i < nr_got; i++)
886 bch2_trans_mark_metadata_bucket(trans, ca,
887 bu[i], BCH_DATA_free, 0));
890 for (i = 0; i < nr_got; i++)
891 bch2_open_bucket_put(c, ob[i]);
893 kfree(new_bucket_seq);
901 * Allocate more journal space at runtime - not currently making use if it, but
904 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
907 struct journal_device *ja = &ca->journal;
911 closure_init_stack(&cl);
913 down_write(&c->state_lock);
915 /* don't handle reducing nr of buckets yet: */
919 while (ja->nr < nr) {
920 struct disk_reservation disk_res = { 0, 0, 0 };
923 * note: journal buckets aren't really counted as _sectors_ used yet, so
924 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
925 * when space used goes up without a reservation - but we do need the
926 * reservation to ensure we'll actually be able to allocate:
928 * XXX: that's not right, disk reservations only ensure a
929 * filesystem-wide allocation will succeed, this is a device
930 * specific allocation - we can hang here:
933 ret = bch2_disk_reservation_get(c, &disk_res,
934 bucket_to_sector(ca, nr - ja->nr), 1, 0);
938 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
940 bch2_disk_reservation_put(c, &disk_res);
944 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
951 up_write(&c->state_lock);
955 int bch2_dev_journal_alloc(struct bch_dev *ca)
960 if (dynamic_fault("bcachefs:add:journal_alloc")) {
961 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
965 /* 1/128th of the device by default: */
966 nr = ca->mi.nbuckets >> 7;
969 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
972 nr = clamp_t(unsigned, nr,
973 BCH_JOURNAL_BUCKETS_MIN,
975 (1 << 24) / ca->mi.bucket_size));
977 ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
984 int bch2_fs_journal_alloc(struct bch_fs *c)
989 for_each_online_member(ca, c, i) {
993 int ret = bch2_dev_journal_alloc(ca);
995 percpu_ref_put(&ca->io_ref);
1003 /* startup/shutdown: */
1005 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1010 spin_lock(&j->lock);
1011 for (seq = journal_last_unwritten_seq(j);
1012 seq <= journal_cur_seq(j) && !ret;
1014 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1016 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1019 spin_unlock(&j->lock);
1024 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1026 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1029 void bch2_fs_journal_stop(struct journal *j)
1031 bch2_journal_reclaim_stop(j);
1032 bch2_journal_flush_all_pins(j);
1034 wait_event(j->wait, journal_entry_close(j));
1037 * Always write a new journal entry, to make sure the clock hands are up
1038 * to date (and match the superblock)
1040 bch2_journal_meta(j);
1044 BUG_ON(!bch2_journal_error(j) &&
1045 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1046 j->last_empty_seq != journal_cur_seq(j));
1048 cancel_delayed_work_sync(&j->write_work);
1051 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1053 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1054 struct journal_entry_pin_list *p;
1055 struct journal_replay *i, **_i;
1056 struct genradix_iter iter;
1057 bool had_entries = false;
1059 u64 last_seq = cur_seq, nr, seq;
1061 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1064 if (!i || i->ignore)
1067 last_seq = le64_to_cpu(i->j.last_seq);
1071 nr = cur_seq - last_seq;
1073 if (nr + 1 > j->pin.size) {
1075 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1077 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1078 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1082 j->replay_journal_seq = last_seq;
1083 j->replay_journal_seq_end = cur_seq;
1084 j->last_seq_ondisk = last_seq;
1085 j->flushed_seq_ondisk = cur_seq - 1;
1086 j->seq_ondisk = cur_seq - 1;
1087 j->pin.front = last_seq;
1088 j->pin.back = cur_seq;
1089 atomic64_set(&j->seq, cur_seq - 1);
1091 fifo_for_each_entry_ptr(p, &j->pin, seq)
1092 journal_pin_list_init(p, 1);
1094 genradix_for_each(&c->journal_entries, iter, _i) {
1097 if (!i || i->ignore)
1100 seq = le64_to_cpu(i->j.seq);
1101 BUG_ON(seq >= cur_seq);
1106 if (journal_entry_empty(&i->j))
1107 j->last_empty_seq = le64_to_cpu(i->j.seq);
1109 p = journal_seq_pin(j, seq);
1112 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1113 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1119 j->last_empty_seq = cur_seq;
1121 spin_lock(&j->lock);
1123 set_bit(JOURNAL_STARTED, &j->flags);
1124 j->last_flush_write = jiffies;
1126 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1127 j->reservations.unwritten_idx++;
1129 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1131 bch2_journal_space_available(j);
1132 spin_unlock(&j->lock);
1134 return bch2_journal_reclaim_start(j);
1139 void bch2_dev_journal_exit(struct bch_dev *ca)
1141 kfree(ca->journal.bio);
1142 kfree(ca->journal.buckets);
1143 kfree(ca->journal.bucket_seq);
1145 ca->journal.bio = NULL;
1146 ca->journal.buckets = NULL;
1147 ca->journal.bucket_seq = NULL;
1150 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1152 struct journal_device *ja = &ca->journal;
1153 struct bch_sb_field_journal *journal_buckets =
1154 bch2_sb_field_get(sb, journal);
1155 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1156 bch2_sb_field_get(sb, journal_v2);
1157 unsigned i, nr_bvecs;
1161 if (journal_buckets_v2) {
1162 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1164 for (i = 0; i < nr; i++)
1165 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1166 } else if (journal_buckets) {
1167 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1170 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1171 if (!ja->bucket_seq)
1172 return -BCH_ERR_ENOMEM_dev_journal_init;
1174 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1176 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1177 if (!ca->journal.bio)
1178 return -BCH_ERR_ENOMEM_dev_journal_init;
1180 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1182 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1184 return -BCH_ERR_ENOMEM_dev_journal_init;
1186 if (journal_buckets_v2) {
1187 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1188 unsigned j, dst = 0;
1190 for (i = 0; i < nr; i++)
1191 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1192 ja->buckets[dst++] =
1193 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1194 } else if (journal_buckets) {
1195 for (i = 0; i < ja->nr; i++)
1196 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1202 void bch2_fs_journal_exit(struct journal *j)
1206 darray_exit(&j->early_journal_entries);
1208 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1209 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1213 int bch2_fs_journal_init(struct journal *j)
1215 static struct lock_class_key res_key;
1218 spin_lock_init(&j->lock);
1219 spin_lock_init(&j->err_lock);
1220 init_waitqueue_head(&j->wait);
1221 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1222 init_waitqueue_head(&j->reclaim_wait);
1223 init_waitqueue_head(&j->pin_flush_wait);
1224 mutex_init(&j->reclaim_lock);
1225 mutex_init(&j->discard_lock);
1227 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1229 atomic64_set(&j->reservations.counter,
1230 ((union journal_res_state)
1231 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1233 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1234 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1236 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1237 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1238 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1239 if (!j->buf[i].data)
1240 return -BCH_ERR_ENOMEM_journal_buf;
1243 j->pin.front = j->pin.back = 1;
1249 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1251 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1252 union journal_res_state s;
1254 unsigned long now = jiffies;
1255 u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1259 if (!out->nr_tabstops)
1260 printbuf_tabstop_push(out, 24);
1264 s = READ_ONCE(j->reservations);
1266 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1267 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1268 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1269 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1270 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1271 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1272 prt_printf(out, "watermark:\t\t%s\n", bch2_watermarks[j->watermark]);
1273 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1274 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1275 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1276 prt_printf(out, "average write size:\t");
1277 prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1279 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1280 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1281 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1282 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1283 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1284 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1285 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1286 prt_printf(out, "current entry:\t\t");
1288 switch (s.cur_entry_offset) {
1289 case JOURNAL_ENTRY_ERROR_VAL:
1290 prt_printf(out, "error");
1292 case JOURNAL_ENTRY_CLOSED_VAL:
1293 prt_printf(out, "closed");
1296 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1302 for (seq = journal_cur_seq(j);
1303 seq >= journal_last_unwritten_seq(j);
1305 i = seq & JOURNAL_BUF_MASK;
1307 prt_printf(out, "unwritten entry:");
1309 prt_printf(out, "%llu", seq);
1311 printbuf_indent_add(out, 2);
1313 prt_printf(out, "refcount:");
1315 prt_printf(out, "%u", journal_state_count(s, i));
1318 prt_printf(out, "sectors:");
1320 prt_printf(out, "%u", j->buf[i].sectors);
1323 prt_printf(out, "expires");
1325 prt_printf(out, "%li jiffies", j->buf[i].expires - jiffies);
1328 printbuf_indent_sub(out, 2);
1332 "replay done:\t\t%i\n",
1333 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1335 prt_printf(out, "space:\n");
1336 prt_printf(out, "\tdiscarded\t%u:%u\n",
1337 j->space[journal_space_discarded].next_entry,
1338 j->space[journal_space_discarded].total);
1339 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1340 j->space[journal_space_clean_ondisk].next_entry,
1341 j->space[journal_space_clean_ondisk].total);
1342 prt_printf(out, "\tclean\t\t%u:%u\n",
1343 j->space[journal_space_clean].next_entry,
1344 j->space[journal_space_clean].total);
1345 prt_printf(out, "\ttotal\t\t%u:%u\n",
1346 j->space[journal_space_total].next_entry,
1347 j->space[journal_space_total].total);
1349 for_each_member_device_rcu(ca, c, i,
1350 &c->rw_devs[BCH_DATA_journal]) {
1351 struct journal_device *ja = &ca->journal;
1353 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1359 prt_printf(out, "dev %u:\n", i);
1360 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1361 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1362 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1363 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1364 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1365 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1366 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1374 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1376 spin_lock(&j->lock);
1377 __bch2_journal_debug_to_text(out, j);
1378 spin_unlock(&j->lock);
1381 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1383 struct journal_entry_pin_list *pin_list;
1384 struct journal_entry_pin *pin;
1387 spin_lock(&j->lock);
1388 *seq = max(*seq, j->pin.front);
1390 if (*seq >= j->pin.back) {
1391 spin_unlock(&j->lock);
1397 pin_list = journal_seq_pin(j, *seq);
1399 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1401 printbuf_indent_add(out, 2);
1403 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1404 list_for_each_entry(pin, &pin_list->list[i], list) {
1405 prt_printf(out, "\t%px %ps", pin, pin->flush);
1409 if (!list_empty(&pin_list->flushed)) {
1410 prt_printf(out, "flushed:");
1414 list_for_each_entry(pin, &pin_list->flushed, list) {
1415 prt_printf(out, "\t%px %ps", pin, pin->flush);
1419 printbuf_indent_sub(out, 2);
1422 spin_unlock(&j->lock);
1427 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1431 while (!bch2_journal_seq_pins_to_text(out, j, &seq))