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);
79 /* journal entry close/open: */
81 void __bch2_journal_buf_put(struct journal *j)
83 struct bch_fs *c = container_of(j, struct bch_fs, journal);
85 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
89 * Returns true if journal entry is now closed:
91 * We don't close a journal_buf until the next journal_buf is finished writing,
92 * and can be opened again - this also initializes the next journal_buf:
94 static void __journal_entry_close(struct journal *j, unsigned closed_val)
96 struct bch_fs *c = container_of(j, struct bch_fs, journal);
97 struct journal_buf *buf = journal_cur_buf(j);
98 union journal_res_state old, new;
99 u64 v = atomic64_read(&j->reservations.counter);
102 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
103 closed_val != JOURNAL_ENTRY_ERROR_VAL);
105 lockdep_assert_held(&j->lock);
109 new.cur_entry_offset = closed_val;
111 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
112 old.cur_entry_offset == new.cur_entry_offset)
114 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
115 old.v, new.v)) != old.v);
117 if (!__journal_entry_is_open(old))
120 /* Close out old buffer: */
121 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
123 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
124 buf->u64s_reserved) << c->block_bits;
125 BUG_ON(sectors > buf->sectors);
126 buf->sectors = sectors;
129 * We have to set last_seq here, _before_ opening a new journal entry:
131 * A threads may replace an old pin with a new pin on their current
132 * journal reservation - the expectation being that the journal will
133 * contain either what the old pin protected or what the new pin
136 * After the old pin is dropped journal_last_seq() won't include the old
137 * pin, so we can only write the updated last_seq on the entry that
138 * contains whatever the new pin protects.
140 * Restated, we can _not_ update last_seq for a given entry if there
141 * could be a newer entry open with reservations/pins that have been
144 * Hence, we want update/set last_seq on the current journal entry right
145 * before we open a new one:
147 buf->last_seq = journal_last_seq(j);
148 buf->data->last_seq = cpu_to_le64(buf->last_seq);
149 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
151 __bch2_journal_pin_put(j, le64_to_cpu(buf->data->seq));
153 cancel_delayed_work(&j->write_work);
155 bch2_journal_space_available(j);
157 bch2_journal_buf_put(j, old.idx);
160 void bch2_journal_halt(struct journal *j)
163 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL);
165 j->err_seq = journal_cur_seq(j);
166 spin_unlock(&j->lock);
169 static bool journal_entry_want_write(struct journal *j)
171 bool ret = !journal_entry_is_open(j) ||
172 journal_cur_seq(j) == journal_last_unwritten_seq(j);
174 /* Don't close it yet if we already have a write in flight: */
176 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
177 else if (nr_unwritten_journal_entries(j)) {
178 struct journal_buf *buf = journal_cur_buf(j);
180 if (!buf->flush_time) {
181 buf->flush_time = local_clock() ?: 1;
182 buf->expires = jiffies;
189 static bool journal_entry_close(struct journal *j)
194 ret = journal_entry_want_write(j);
195 spin_unlock(&j->lock);
201 * should _only_ called from journal_res_get() - when we actually want a
202 * journal reservation - journal entry is open means journal is dirty:
204 static int journal_entry_open(struct journal *j)
206 struct bch_fs *c = container_of(j, struct bch_fs, journal);
207 struct journal_buf *buf = j->buf +
208 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
209 union journal_res_state old, new;
213 lockdep_assert_held(&j->lock);
214 BUG_ON(journal_entry_is_open(j));
215 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
218 return JOURNAL_ERR_blocked;
220 if (j->cur_entry_error)
221 return j->cur_entry_error;
223 if (bch2_journal_error(j))
224 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
226 if (!fifo_free(&j->pin))
227 return JOURNAL_ERR_journal_pin_full;
229 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
230 return JOURNAL_ERR_max_in_flight;
232 BUG_ON(!j->cur_entry_sectors);
235 (journal_cur_seq(j) == j->flushed_seq_ondisk
237 : j->last_flush_write) +
238 msecs_to_jiffies(c->opts.journal_flush_delay);
240 buf->u64s_reserved = j->entry_u64s_reserved;
241 buf->disk_sectors = j->cur_entry_sectors;
242 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
244 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
245 journal_entry_overhead(j);
246 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
248 if (u64s <= (ssize_t) j->early_journal_entries.nr)
249 return JOURNAL_ERR_journal_full;
251 if (fifo_empty(&j->pin) && j->reclaim_thread)
252 wake_up_process(j->reclaim_thread);
255 * The fifo_push() needs to happen at the same time as j->seq is
256 * incremented for journal_last_seq() to be calculated correctly
258 atomic64_inc(&j->seq);
259 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
261 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
263 bkey_extent_init(&buf->key);
264 buf->noflush = false;
265 buf->must_flush = false;
266 buf->separate_flush = false;
269 memset(buf->data, 0, sizeof(*buf->data));
270 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
273 if (j->early_journal_entries.nr) {
274 memcpy(buf->data->_data, j->early_journal_entries.data,
275 j->early_journal_entries.nr * sizeof(u64));
276 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
280 * Must be set before marking the journal entry as open:
282 j->cur_entry_u64s = u64s;
284 v = atomic64_read(&j->reservations.counter);
288 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
291 BUG_ON(journal_state_count(new, new.idx));
292 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
294 journal_state_inc(&new);
296 /* Handle any already added entries */
297 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
298 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
299 old.v, new.v)) != old.v);
301 if (j->res_get_blocked_start)
302 bch2_time_stats_update(j->blocked_time,
303 j->res_get_blocked_start);
304 j->res_get_blocked_start = 0;
306 mod_delayed_work(c->io_complete_wq,
308 msecs_to_jiffies(c->opts.journal_flush_delay));
311 if (j->early_journal_entries.nr)
312 darray_exit(&j->early_journal_entries);
316 static bool journal_quiesced(struct journal *j)
318 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
321 journal_entry_close(j);
325 static void journal_quiesce(struct journal *j)
327 wait_event(j->wait, journal_quiesced(j));
330 static void journal_write_work(struct work_struct *work)
332 struct journal *j = container_of(work, struct journal, write_work.work);
333 struct bch_fs *c = container_of(j, struct bch_fs, journal);
337 if (!__journal_entry_is_open(j->reservations))
340 delta = journal_cur_buf(j)->expires - jiffies;
343 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
345 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
347 spin_unlock(&j->lock);
350 static int __journal_res_get(struct journal *j, struct journal_res *res,
353 struct bch_fs *c = container_of(j, struct bch_fs, journal);
354 struct journal_buf *buf;
358 if (journal_res_get_fast(j, res, flags))
361 if (bch2_journal_error(j))
362 return -BCH_ERR_erofs_journal_err;
367 * Recheck after taking the lock, so we don't race with another thread
368 * that just did journal_entry_open() and call journal_entry_close()
371 if (journal_res_get_fast(j, res, flags)) {
372 spin_unlock(&j->lock);
376 if ((flags & JOURNAL_WATERMARK_MASK) < j->watermark) {
378 * Don't want to close current journal entry, just need to
381 ret = JOURNAL_ERR_journal_full;
386 * If we couldn't get a reservation because the current buf filled up,
387 * and we had room for a bigger entry on disk, signal that we want to
388 * realloc the journal bufs:
390 buf = journal_cur_buf(j);
391 if (journal_entry_is_open(j) &&
392 buf->buf_size >> 9 < buf->disk_sectors &&
393 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
394 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
396 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
397 ret = journal_entry_open(j);
399 if (ret == JOURNAL_ERR_max_in_flight)
400 trace_and_count(c, journal_entry_full, c);
402 if ((ret && ret != JOURNAL_ERR_insufficient_devices) &&
403 !j->res_get_blocked_start) {
404 j->res_get_blocked_start = local_clock() ?: 1;
405 trace_and_count(c, journal_full, c);
408 can_discard = j->can_discard;
409 spin_unlock(&j->lock);
414 if ((ret == JOURNAL_ERR_journal_full ||
415 ret == JOURNAL_ERR_journal_pin_full) &&
417 !nr_unwritten_journal_entries(j) &&
418 (flags & JOURNAL_WATERMARK_MASK) == JOURNAL_WATERMARK_reserved) {
419 struct printbuf buf = PRINTBUF;
421 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (ret %s)",
422 bch2_journal_errors[ret]);
424 bch2_journal_debug_to_text(&buf, j);
425 bch_err(c, "%s", buf.buf);
427 printbuf_reset(&buf);
428 bch2_journal_pins_to_text(&buf, j);
429 bch_err(c, "Journal pins:\n%s", buf.buf);
437 * Journal is full - can't rely on reclaim from work item due to
440 if ((ret == JOURNAL_ERR_journal_full ||
441 ret == JOURNAL_ERR_journal_pin_full) &&
442 !(flags & JOURNAL_RES_GET_NONBLOCK)) {
444 bch2_journal_do_discards(j);
448 if (mutex_trylock(&j->reclaim_lock)) {
449 bch2_journal_reclaim(j);
450 mutex_unlock(&j->reclaim_lock);
454 return ret == JOURNAL_ERR_insufficient_devices
456 : -BCH_ERR_journal_res_get_blocked;
460 * Essentially the entry function to the journaling code. When bcachefs is doing
461 * a btree insert, it calls this function to get the current journal write.
462 * Journal write is the structure used set up journal writes. The calling
463 * function will then add its keys to the structure, queuing them for the next
466 * To ensure forward progress, the current task must not be holding any
467 * btree node write locks.
469 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
474 closure_wait_event(&j->async_wait,
475 (ret = __journal_res_get(j, res, flags)) !=
476 -BCH_ERR_journal_res_get_blocked||
477 (flags & JOURNAL_RES_GET_NONBLOCK));
481 /* journal_preres: */
483 static bool journal_preres_available(struct journal *j,
484 struct journal_preres *res,
488 bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags, true);
490 if (!ret && mutex_trylock(&j->reclaim_lock)) {
491 bch2_journal_reclaim(j);
492 mutex_unlock(&j->reclaim_lock);
498 int __bch2_journal_preres_get(struct journal *j,
499 struct journal_preres *res,
505 closure_wait_event(&j->preres_wait,
506 (ret = bch2_journal_error(j)) ||
507 journal_preres_available(j, res, new_u64s, flags));
511 /* journal_entry_res: */
513 void bch2_journal_entry_res_resize(struct journal *j,
514 struct journal_entry_res *res,
517 union journal_res_state state;
518 int d = new_u64s - res->u64s;
522 j->entry_u64s_reserved += d;
526 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
528 state = READ_ONCE(j->reservations);
530 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
531 state.cur_entry_offset > j->cur_entry_u64s) {
532 j->cur_entry_u64s += d;
534 * Not enough room in current journal entry, have to flush it:
536 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
538 journal_cur_buf(j)->u64s_reserved += d;
541 spin_unlock(&j->lock);
545 /* journal flushing: */
548 * bch2_journal_flush_seq_async - wait for a journal entry to be written
550 * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
553 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
554 struct closure *parent)
556 struct journal_buf *buf;
559 if (seq <= j->flushed_seq_ondisk)
564 if (WARN_ONCE(seq > journal_cur_seq(j),
565 "requested to flush journal seq %llu, but currently at %llu",
566 seq, journal_cur_seq(j)))
569 /* Recheck under lock: */
570 if (j->err_seq && seq >= j->err_seq) {
575 if (seq <= j->flushed_seq_ondisk) {
580 /* if seq was written, but not flushed - flush a newer one instead */
581 seq = max(seq, journal_last_unwritten_seq(j));
584 if (seq > journal_cur_seq(j)) {
585 struct journal_res res = { 0 };
587 if (journal_entry_is_open(j))
588 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
590 spin_unlock(&j->lock);
592 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
597 buf = j->buf + (seq & JOURNAL_BUF_MASK);
598 buf->must_flush = true;
600 if (!buf->flush_time) {
601 buf->flush_time = local_clock() ?: 1;
602 buf->expires = jiffies;
605 if (parent && !closure_wait(&buf->wait, parent))
608 bch2_journal_res_put(j, &res);
615 * if write was kicked off without a flush, flush the next sequence
618 buf = journal_seq_to_buf(j, seq);
621 goto recheck_need_open;
624 buf->must_flush = true;
626 if (parent && !closure_wait(&buf->wait, parent))
629 if (seq == journal_cur_seq(j))
630 journal_entry_want_write(j);
632 spin_unlock(&j->lock);
636 int bch2_journal_flush_seq(struct journal *j, u64 seq)
638 u64 start_time = local_clock();
642 * Don't update time_stats when @seq is already flushed:
644 if (seq <= j->flushed_seq_ondisk)
647 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
650 bch2_time_stats_update(j->flush_seq_time, start_time);
652 return ret ?: ret2 < 0 ? ret2 : 0;
656 * bch2_journal_flush_async - if there is an open journal entry, or a journal
657 * still being written, write it and wait for the write to complete
659 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
661 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
664 int bch2_journal_flush(struct journal *j)
666 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
670 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
673 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
675 struct bch_fs *c = container_of(j, struct bch_fs, journal);
679 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
682 if (seq <= c->journal.flushed_seq_ondisk)
686 if (seq <= c->journal.flushed_seq_ondisk)
689 for (unwritten_seq = journal_last_unwritten_seq(j);
692 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
694 /* journal write is already in flight, and was a flush write: */
695 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
703 spin_unlock(&j->lock);
707 int bch2_journal_meta(struct journal *j)
709 struct journal_buf *buf;
710 struct journal_res res;
713 memset(&res, 0, sizeof(res));
715 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
719 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
720 buf->must_flush = true;
722 if (!buf->flush_time) {
723 buf->flush_time = local_clock() ?: 1;
724 buf->expires = jiffies;
727 bch2_journal_res_put(j, &res);
729 return bch2_journal_flush_seq(j, res.seq);
732 /* block/unlock the journal: */
734 void bch2_journal_unblock(struct journal *j)
738 spin_unlock(&j->lock);
743 void bch2_journal_block(struct journal *j)
747 spin_unlock(&j->lock);
752 /* allocate journal on a device: */
754 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
755 bool new_fs, struct closure *cl)
757 struct bch_fs *c = ca->fs;
758 struct journal_device *ja = &ca->journal;
759 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
760 struct open_bucket **ob = NULL;
762 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
765 BUG_ON(nr <= ja->nr);
767 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
768 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
769 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
770 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
771 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
776 for (nr_got = 0; nr_got < nr_want; nr_got++) {
778 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
779 if (bu[nr_got] < 0) {
780 ret = -BCH_ERR_ENOSPC_bucket_alloc;
784 ob[nr_got] = bch2_bucket_alloc(c, ca, RESERVE_none, cl);
785 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
789 ret = bch2_trans_run(c,
790 bch2_trans_mark_metadata_bucket(&trans, ca,
791 ob[nr_got]->bucket, BCH_DATA_journal,
792 ca->mi.bucket_size));
794 bch2_open_bucket_put(c, ob[nr_got]);
795 bch_err(c, "error marking new journal buckets: %s", bch2_err_str(ret));
799 bu[nr_got] = ob[nr_got]->bucket;
806 /* Don't return an error if we successfully allocated some buckets: */
810 bch2_journal_flush_all_pins(&c->journal);
811 bch2_journal_block(&c->journal);
812 mutex_lock(&c->sb_lock);
815 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
816 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
818 BUG_ON(ja->discard_idx > ja->nr);
820 pos = ja->discard_idx ?: ja->nr;
822 memmove(new_buckets + pos + nr_got,
824 sizeof(new_buckets[0]) * (ja->nr - pos));
825 memmove(new_bucket_seq + pos + nr_got,
826 new_bucket_seq + pos,
827 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
829 for (i = 0; i < nr_got; i++) {
830 new_buckets[pos + i] = bu[i];
831 new_bucket_seq[pos + i] = 0;
834 nr = ja->nr + nr_got;
836 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
845 spin_lock(&c->journal.lock);
847 swap(new_buckets, ja->buckets);
848 swap(new_bucket_seq, ja->bucket_seq);
851 if (pos <= ja->discard_idx)
852 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
853 if (pos <= ja->dirty_idx_ondisk)
854 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
855 if (pos <= ja->dirty_idx)
856 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
857 if (pos <= ja->cur_idx)
858 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
861 spin_unlock(&c->journal.lock);
864 bch2_journal_unblock(&c->journal);
865 mutex_unlock(&c->sb_lock);
869 for (i = 0; i < nr_got; i++)
871 bch2_trans_mark_metadata_bucket(&trans, ca,
872 bu[i], BCH_DATA_free, 0));
875 for (i = 0; i < nr_got; i++)
876 bch2_open_bucket_put(c, ob[i]);
878 kfree(new_bucket_seq);
886 * Allocate more journal space at runtime - not currently making use if it, but
889 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
892 struct journal_device *ja = &ca->journal;
896 closure_init_stack(&cl);
898 down_write(&c->state_lock);
900 /* don't handle reducing nr of buckets yet: */
904 while (ja->nr < nr) {
905 struct disk_reservation disk_res = { 0, 0 };
908 * note: journal buckets aren't really counted as _sectors_ used yet, so
909 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
910 * when space used goes up without a reservation - but we do need the
911 * reservation to ensure we'll actually be able to allocate:
913 * XXX: that's not right, disk reservations only ensure a
914 * filesystem-wide allocation will succeed, this is a device
915 * specific allocation - we can hang here:
918 ret = bch2_disk_reservation_get(c, &disk_res,
919 bucket_to_sector(ca, nr - ja->nr), 1, 0);
923 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
925 bch2_disk_reservation_put(c, &disk_res);
929 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
934 bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
936 up_write(&c->state_lock);
940 int bch2_dev_journal_alloc(struct bch_dev *ca)
944 if (dynamic_fault("bcachefs:add:journal_alloc"))
947 /* 1/128th of the device by default: */
948 nr = ca->mi.nbuckets >> 7;
951 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
954 nr = clamp_t(unsigned, nr,
955 BCH_JOURNAL_BUCKETS_MIN,
957 (1 << 24) / ca->mi.bucket_size));
959 return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
962 /* startup/shutdown: */
964 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
970 for (seq = journal_last_unwritten_seq(j);
971 seq <= journal_cur_seq(j) && !ret;
973 struct journal_buf *buf = journal_seq_to_buf(j, seq);
975 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
978 spin_unlock(&j->lock);
983 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
985 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
988 void bch2_fs_journal_stop(struct journal *j)
990 bch2_journal_reclaim_stop(j);
991 bch2_journal_flush_all_pins(j);
993 wait_event(j->wait, journal_entry_close(j));
996 * Always write a new journal entry, to make sure the clock hands are up
997 * to date (and match the superblock)
999 bch2_journal_meta(j);
1003 BUG_ON(!bch2_journal_error(j) &&
1004 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1005 j->last_empty_seq != journal_cur_seq(j));
1007 cancel_delayed_work_sync(&j->write_work);
1010 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1012 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1013 struct journal_entry_pin_list *p;
1014 struct journal_replay *i, **_i;
1015 struct genradix_iter iter;
1016 bool had_entries = false;
1018 u64 last_seq = cur_seq, nr, seq;
1020 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1023 if (!i || i->ignore)
1026 last_seq = le64_to_cpu(i->j.last_seq);
1030 nr = cur_seq - last_seq;
1032 if (nr + 1 > j->pin.size) {
1034 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1036 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1041 j->replay_journal_seq = last_seq;
1042 j->replay_journal_seq_end = cur_seq;
1043 j->last_seq_ondisk = last_seq;
1044 j->flushed_seq_ondisk = cur_seq - 1;
1045 j->seq_ondisk = cur_seq - 1;
1046 j->pin.front = last_seq;
1047 j->pin.back = cur_seq;
1048 atomic64_set(&j->seq, cur_seq - 1);
1050 fifo_for_each_entry_ptr(p, &j->pin, seq)
1051 journal_pin_list_init(p, 1);
1053 genradix_for_each(&c->journal_entries, iter, _i) {
1056 if (!i || i->ignore)
1059 seq = le64_to_cpu(i->j.seq);
1060 BUG_ON(seq >= cur_seq);
1065 if (journal_entry_empty(&i->j))
1066 j->last_empty_seq = le64_to_cpu(i->j.seq);
1068 p = journal_seq_pin(j, seq);
1071 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1072 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1078 j->last_empty_seq = cur_seq;
1080 spin_lock(&j->lock);
1082 set_bit(JOURNAL_STARTED, &j->flags);
1083 j->last_flush_write = jiffies;
1085 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1086 j->reservations.unwritten_idx++;
1088 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1090 bch2_journal_space_available(j);
1091 spin_unlock(&j->lock);
1093 return bch2_journal_reclaim_start(j);
1098 void bch2_dev_journal_exit(struct bch_dev *ca)
1100 kfree(ca->journal.bio);
1101 kfree(ca->journal.buckets);
1102 kfree(ca->journal.bucket_seq);
1104 ca->journal.bio = NULL;
1105 ca->journal.buckets = NULL;
1106 ca->journal.bucket_seq = NULL;
1109 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1111 struct journal_device *ja = &ca->journal;
1112 struct bch_sb_field_journal *journal_buckets =
1113 bch2_sb_get_journal(sb);
1114 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1115 bch2_sb_get_journal_v2(sb);
1116 unsigned i, nr_bvecs;
1120 if (journal_buckets_v2) {
1121 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1123 for (i = 0; i < nr; i++)
1124 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1125 } else if (journal_buckets) {
1126 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1129 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1130 if (!ja->bucket_seq)
1133 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1135 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1136 if (!ca->journal.bio)
1139 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1141 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1145 if (journal_buckets_v2) {
1146 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1147 unsigned j, dst = 0;
1149 for (i = 0; i < nr; i++)
1150 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1151 ja->buckets[dst++] =
1152 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1153 } else if (journal_buckets) {
1154 for (i = 0; i < ja->nr; i++)
1155 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1161 void bch2_fs_journal_exit(struct journal *j)
1165 darray_exit(&j->early_journal_entries);
1167 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1168 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1172 int bch2_fs_journal_init(struct journal *j)
1174 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1175 static struct lock_class_key res_key;
1179 pr_verbose_init(c->opts, "");
1181 spin_lock_init(&j->lock);
1182 spin_lock_init(&j->err_lock);
1183 init_waitqueue_head(&j->wait);
1184 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1185 init_waitqueue_head(&j->reclaim_wait);
1186 init_waitqueue_head(&j->pin_flush_wait);
1187 mutex_init(&j->reclaim_lock);
1188 mutex_init(&j->discard_lock);
1190 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1192 atomic64_set(&j->reservations.counter,
1193 ((union journal_res_state)
1194 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1196 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) {
1201 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1202 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1203 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1204 if (!j->buf[i].data) {
1210 j->pin.front = j->pin.back = 1;
1212 pr_verbose_init(c->opts, "ret %i", ret);
1218 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1220 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1221 union journal_res_state s;
1223 unsigned long now = jiffies;
1227 if (!out->nr_tabstops)
1228 printbuf_tabstop_push(out, 24);
1232 s = READ_ONCE(j->reservations);
1234 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1235 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1236 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1237 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1238 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1239 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1240 prt_printf(out, "prereserved:\t\t%u/%u\n", j->prereserved.reserved, j->prereserved.remaining);
1241 prt_printf(out, "watermark:\t\t%s\n", bch2_journal_watermarks[j->watermark]);
1242 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1243 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1244 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1245 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1246 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1247 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1248 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1249 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1250 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1251 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1252 prt_printf(out, "current entry:\t\t");
1254 switch (s.cur_entry_offset) {
1255 case JOURNAL_ENTRY_ERROR_VAL:
1256 prt_printf(out, "error");
1258 case JOURNAL_ENTRY_CLOSED_VAL:
1259 prt_printf(out, "closed");
1262 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1268 for (seq = journal_cur_seq(j);
1269 seq >= journal_last_unwritten_seq(j);
1271 i = seq & JOURNAL_BUF_MASK;
1273 prt_printf(out, "unwritten entry:");
1275 prt_printf(out, "%llu", seq);
1277 printbuf_indent_add(out, 2);
1279 prt_printf(out, "refcount:");
1281 prt_printf(out, "%u", journal_state_count(s, i));
1284 prt_printf(out, "sectors:");
1286 prt_printf(out, "%u", j->buf[i].sectors);
1289 prt_printf(out, "expires");
1291 prt_printf(out, "%li jiffies", j->buf[i].expires - jiffies);
1294 printbuf_indent_sub(out, 2);
1298 "replay done:\t\t%i\n",
1299 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1301 prt_printf(out, "space:\n");
1302 prt_printf(out, "\tdiscarded\t%u:%u\n",
1303 j->space[journal_space_discarded].next_entry,
1304 j->space[journal_space_discarded].total);
1305 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1306 j->space[journal_space_clean_ondisk].next_entry,
1307 j->space[journal_space_clean_ondisk].total);
1308 prt_printf(out, "\tclean\t\t%u:%u\n",
1309 j->space[journal_space_clean].next_entry,
1310 j->space[journal_space_clean].total);
1311 prt_printf(out, "\ttotal\t\t%u:%u\n",
1312 j->space[journal_space_total].next_entry,
1313 j->space[journal_space_total].total);
1315 for_each_member_device_rcu(ca, c, i,
1316 &c->rw_devs[BCH_DATA_journal]) {
1317 struct journal_device *ja = &ca->journal;
1319 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1325 prt_printf(out, "dev %u:\n", i);
1326 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1327 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1328 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1329 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1330 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1331 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1332 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1340 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1342 spin_lock(&j->lock);
1343 __bch2_journal_debug_to_text(out, j);
1344 spin_unlock(&j->lock);
1347 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1349 struct journal_entry_pin_list *pin_list;
1350 struct journal_entry_pin *pin;
1353 spin_lock(&j->lock);
1354 *seq = max(*seq, j->pin.front);
1356 if (*seq >= j->pin.back) {
1357 spin_unlock(&j->lock);
1363 pin_list = journal_seq_pin(j, *seq);
1365 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1367 printbuf_indent_add(out, 2);
1369 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1370 list_for_each_entry(pin, &pin_list->list[i], list) {
1371 prt_printf(out, "\t%px %ps", pin, pin->flush);
1375 if (!list_empty(&pin_list->flushed)) {
1376 prt_printf(out, "flushed:");
1380 list_for_each_entry(pin, &pin_list->flushed, list) {
1381 prt_printf(out, "\t%px %ps", pin, pin->flush);
1385 printbuf_indent_sub(out, 2);
1388 spin_unlock(&j->lock);
1393 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1397 while (!bch2_journal_seq_pins_to_text(out, j, &seq))