1 // SPDX-License-Identifier: GPL-2.0
3 * bcachefs journalling code, for btree insertions
5 * Copyright 2012 Google, Inc.
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
12 #include "btree_update.h"
13 #include "btree_write_buffer.h"
17 #include "journal_io.h"
18 #include "journal_reclaim.h"
19 #include "journal_sb.h"
20 #include "journal_seq_blacklist.h"
23 static const char * const bch2_journal_errors[] = {
30 static void bch2_journal_buf_to_text(struct printbuf *out, struct journal *j, u64 seq)
32 union journal_res_state s = READ_ONCE(j->reservations);
33 unsigned i = seq & JOURNAL_BUF_MASK;
34 struct journal_buf *buf = j->buf + i;
36 prt_printf(out, "seq:");
38 prt_printf(out, "%llu", seq);
40 printbuf_indent_add(out, 2);
42 prt_printf(out, "refcount:");
44 prt_printf(out, "%u", journal_state_count(s, i));
47 prt_printf(out, "size:");
49 prt_human_readable_u64(out, vstruct_bytes(buf->data));
52 prt_printf(out, "expires");
54 prt_printf(out, "%li jiffies", buf->expires - jiffies);
57 printbuf_indent_sub(out, 2);
60 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j)
62 if (!out->nr_tabstops)
63 printbuf_tabstop_push(out, 24);
65 for (u64 seq = journal_last_unwritten_seq(j);
66 seq <= journal_cur_seq(j);
68 bch2_journal_buf_to_text(out, j, seq);
71 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
73 return seq > j->seq_ondisk;
76 static bool __journal_entry_is_open(union journal_res_state state)
78 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
81 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
83 return atomic64_read(&j->seq) - j->seq_ondisk;
86 static bool journal_entry_is_open(struct journal *j)
88 return __journal_entry_is_open(j->reservations);
91 static inline struct journal_buf *
92 journal_seq_to_buf(struct journal *j, u64 seq)
94 struct journal_buf *buf = NULL;
96 EBUG_ON(seq > journal_cur_seq(j));
98 if (journal_seq_unwritten(j, seq)) {
99 buf = j->buf + (seq & JOURNAL_BUF_MASK);
100 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
105 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
109 for (i = 0; i < ARRAY_SIZE(p->list); i++)
110 INIT_LIST_HEAD(&p->list[i]);
111 INIT_LIST_HEAD(&p->flushed);
112 atomic_set(&p->count, count);
117 * Detect stuck journal conditions and trigger shutdown. Technically the journal
118 * can end up stuck for a variety of reasons, such as a blocked I/O, journal
119 * reservation lockup, etc. Since this is a fatal error with potentially
120 * unpredictable characteristics, we want to be fairly conservative before we
121 * decide to shut things down.
123 * Consider the journal stuck when it appears full with no ability to commit
124 * btree transactions, to discard journal buckets, nor acquire priority
125 * (reserved watermark) reservation.
128 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
130 struct bch_fs *c = container_of(j, struct bch_fs, journal);
132 struct printbuf buf = PRINTBUF;
134 if (!(error == JOURNAL_ERR_journal_full ||
135 error == JOURNAL_ERR_journal_pin_full) ||
136 nr_unwritten_journal_entries(j) ||
137 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
142 if (j->can_discard) {
143 spin_unlock(&j->lock);
150 * The journal shutdown path will set ->err_seq, but do it here first to
151 * serialize against concurrent failures and avoid duplicate error
155 spin_unlock(&j->lock);
158 j->err_seq = journal_cur_seq(j);
159 spin_unlock(&j->lock);
161 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
162 bch2_journal_errors[error]);
163 bch2_journal_debug_to_text(&buf, j);
164 bch_err(c, "%s", buf.buf);
166 printbuf_reset(&buf);
167 bch2_journal_pins_to_text(&buf, j);
168 bch_err(c, "Journal pins:\n%s", buf.buf);
178 * Final processing when the last reference of a journal buffer has been
179 * dropped. Drop the pin list reference acquired at journal entry open and write
180 * the buffer, if requested.
182 void bch2_journal_buf_put_final(struct journal *j, u64 seq, bool write)
184 struct bch_fs *c = container_of(j, struct bch_fs, journal);
186 lockdep_assert_held(&j->lock);
188 if (__bch2_journal_pin_put(j, seq))
189 bch2_journal_reclaim_fast(j);
191 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
195 * Returns true if journal entry is now closed:
197 * We don't close a journal_buf until the next journal_buf is finished writing,
198 * and can be opened again - this also initializes the next journal_buf:
200 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace)
202 struct bch_fs *c = container_of(j, struct bch_fs, journal);
203 struct journal_buf *buf = journal_cur_buf(j);
204 union journal_res_state old, new;
205 u64 v = atomic64_read(&j->reservations.counter);
208 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
209 closed_val != JOURNAL_ENTRY_ERROR_VAL);
211 lockdep_assert_held(&j->lock);
215 new.cur_entry_offset = closed_val;
217 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
218 old.cur_entry_offset == new.cur_entry_offset)
220 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
221 old.v, new.v)) != old.v);
223 if (!__journal_entry_is_open(old))
226 /* Close out old buffer: */
227 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
229 if (trace_journal_entry_close_enabled() && trace) {
230 struct printbuf pbuf = PRINTBUF;
233 prt_str(&pbuf, "entry size: ");
234 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
236 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT);
237 trace_journal_entry_close(c, pbuf.buf);
238 printbuf_exit(&pbuf);
241 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
242 buf->u64s_reserved) << c->block_bits;
243 BUG_ON(sectors > buf->sectors);
244 buf->sectors = sectors;
247 * We have to set last_seq here, _before_ opening a new journal entry:
249 * A threads may replace an old pin with a new pin on their current
250 * journal reservation - the expectation being that the journal will
251 * contain either what the old pin protected or what the new pin
254 * After the old pin is dropped journal_last_seq() won't include the old
255 * pin, so we can only write the updated last_seq on the entry that
256 * contains whatever the new pin protects.
258 * Restated, we can _not_ update last_seq for a given entry if there
259 * could be a newer entry open with reservations/pins that have been
262 * Hence, we want update/set last_seq on the current journal entry right
263 * before we open a new one:
265 buf->last_seq = journal_last_seq(j);
266 buf->data->last_seq = cpu_to_le64(buf->last_seq);
267 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
269 cancel_delayed_work(&j->write_work);
271 bch2_journal_space_available(j);
273 __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
276 void bch2_journal_halt(struct journal *j)
279 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true);
281 j->err_seq = journal_cur_seq(j);
283 spin_unlock(&j->lock);
286 static bool journal_entry_want_write(struct journal *j)
288 bool ret = !journal_entry_is_open(j) ||
289 journal_cur_seq(j) == journal_last_unwritten_seq(j);
291 /* Don't close it yet if we already have a write in flight: */
293 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
294 else if (nr_unwritten_journal_entries(j)) {
295 struct journal_buf *buf = journal_cur_buf(j);
297 if (!buf->flush_time) {
298 buf->flush_time = local_clock() ?: 1;
299 buf->expires = jiffies;
306 bool bch2_journal_entry_close(struct journal *j)
311 ret = journal_entry_want_write(j);
312 spin_unlock(&j->lock);
318 * should _only_ called from journal_res_get() - when we actually want a
319 * journal reservation - journal entry is open means journal is dirty:
321 static int journal_entry_open(struct journal *j)
323 struct bch_fs *c = container_of(j, struct bch_fs, journal);
324 struct journal_buf *buf = j->buf +
325 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
326 union journal_res_state old, new;
330 lockdep_assert_held(&j->lock);
331 BUG_ON(journal_entry_is_open(j));
332 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
335 return JOURNAL_ERR_blocked;
337 if (j->cur_entry_error)
338 return j->cur_entry_error;
340 if (bch2_journal_error(j))
341 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
343 if (!fifo_free(&j->pin))
344 return JOURNAL_ERR_journal_pin_full;
346 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
347 return JOURNAL_ERR_max_in_flight;
349 BUG_ON(!j->cur_entry_sectors);
352 (journal_cur_seq(j) == j->flushed_seq_ondisk
354 : j->last_flush_write) +
355 msecs_to_jiffies(c->opts.journal_flush_delay);
357 buf->u64s_reserved = j->entry_u64s_reserved;
358 buf->disk_sectors = j->cur_entry_sectors;
359 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
361 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
362 journal_entry_overhead(j);
363 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
365 if (u64s <= (ssize_t) j->early_journal_entries.nr)
366 return JOURNAL_ERR_journal_full;
368 if (fifo_empty(&j->pin) && j->reclaim_thread)
369 wake_up_process(j->reclaim_thread);
372 * The fifo_push() needs to happen at the same time as j->seq is
373 * incremented for journal_last_seq() to be calculated correctly
375 atomic64_inc(&j->seq);
376 journal_pin_list_init(fifo_push_ref(&j->pin), 1);
378 BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
380 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
382 bkey_extent_init(&buf->key);
383 buf->noflush = false;
384 buf->must_flush = false;
385 buf->separate_flush = false;
387 buf->need_flush_to_write_buffer = true;
389 memset(buf->data, 0, sizeof(*buf->data));
390 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
393 if (j->early_journal_entries.nr) {
394 memcpy(buf->data->_data, j->early_journal_entries.data,
395 j->early_journal_entries.nr * sizeof(u64));
396 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
400 * Must be set before marking the journal entry as open:
402 j->cur_entry_u64s = u64s;
404 v = atomic64_read(&j->reservations.counter);
408 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
411 BUG_ON(journal_state_count(new, new.idx));
412 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
414 journal_state_inc(&new);
416 /* Handle any already added entries */
417 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
418 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
419 old.v, new.v)) != old.v);
421 mod_delayed_work(c->io_complete_wq,
423 msecs_to_jiffies(c->opts.journal_flush_delay));
426 if (j->early_journal_entries.nr)
427 darray_exit(&j->early_journal_entries);
431 static bool journal_quiesced(struct journal *j)
433 bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
436 bch2_journal_entry_close(j);
440 static void journal_quiesce(struct journal *j)
442 wait_event(j->wait, journal_quiesced(j));
445 static void journal_write_work(struct work_struct *work)
447 struct journal *j = container_of(work, struct journal, write_work.work);
448 struct bch_fs *c = container_of(j, struct bch_fs, journal);
452 if (!__journal_entry_is_open(j->reservations))
455 delta = journal_cur_buf(j)->expires - jiffies;
458 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
460 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
462 spin_unlock(&j->lock);
465 static int __journal_res_get(struct journal *j, struct journal_res *res,
468 struct bch_fs *c = container_of(j, struct bch_fs, journal);
469 struct journal_buf *buf;
473 if (journal_res_get_fast(j, res, flags))
476 if (bch2_journal_error(j))
477 return -BCH_ERR_erofs_journal_err;
481 /* check once more in case somebody else shut things down... */
482 if (bch2_journal_error(j)) {
483 spin_unlock(&j->lock);
484 return -BCH_ERR_erofs_journal_err;
488 * Recheck after taking the lock, so we don't race with another thread
489 * that just did journal_entry_open() and call bch2_journal_entry_close()
492 if (journal_res_get_fast(j, res, flags)) {
493 spin_unlock(&j->lock);
497 if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
499 * Don't want to close current journal entry, just need to
502 ret = JOURNAL_ERR_journal_full;
507 * If we couldn't get a reservation because the current buf filled up,
508 * and we had room for a bigger entry on disk, signal that we want to
509 * realloc the journal bufs:
511 buf = journal_cur_buf(j);
512 if (journal_entry_is_open(j) &&
513 buf->buf_size >> 9 < buf->disk_sectors &&
514 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
515 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
517 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
518 ret = journal_entry_open(j);
520 if (ret == JOURNAL_ERR_max_in_flight) {
521 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
522 &j->max_in_flight_start, true);
523 if (trace_journal_entry_full_enabled()) {
524 struct printbuf buf = PRINTBUF;
527 bch2_journal_bufs_to_text(&buf, j);
528 trace_journal_entry_full(c, buf.buf);
531 count_event(c, journal_entry_full);
534 can_discard = j->can_discard;
535 spin_unlock(&j->lock);
539 if (journal_error_check_stuck(j, ret, flags))
540 ret = -BCH_ERR_journal_res_get_blocked;
543 * Journal is full - can't rely on reclaim from work item due to
546 if ((ret == JOURNAL_ERR_journal_full ||
547 ret == JOURNAL_ERR_journal_pin_full) &&
548 !(flags & JOURNAL_RES_GET_NONBLOCK)) {
550 bch2_journal_do_discards(j);
554 if (mutex_trylock(&j->reclaim_lock)) {
555 bch2_journal_reclaim(j);
556 mutex_unlock(&j->reclaim_lock);
560 return ret == JOURNAL_ERR_insufficient_devices
561 ? -BCH_ERR_erofs_journal_err
562 : -BCH_ERR_journal_res_get_blocked;
566 * Essentially the entry function to the journaling code. When bcachefs is doing
567 * a btree insert, it calls this function to get the current journal write.
568 * Journal write is the structure used set up journal writes. The calling
569 * function will then add its keys to the structure, queuing them for the next
572 * To ensure forward progress, the current task must not be holding any
573 * btree node write locks.
575 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
580 closure_wait_event(&j->async_wait,
581 (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
582 (flags & JOURNAL_RES_GET_NONBLOCK));
586 /* journal_entry_res: */
588 void bch2_journal_entry_res_resize(struct journal *j,
589 struct journal_entry_res *res,
592 union journal_res_state state;
593 int d = new_u64s - res->u64s;
597 j->entry_u64s_reserved += d;
601 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
603 state = READ_ONCE(j->reservations);
605 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
606 state.cur_entry_offset > j->cur_entry_u64s) {
607 j->cur_entry_u64s += d;
609 * Not enough room in current journal entry, have to flush it:
611 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
613 journal_cur_buf(j)->u64s_reserved += d;
616 spin_unlock(&j->lock);
620 /* journal flushing: */
623 * bch2_journal_flush_seq_async - wait for a journal entry to be written
626 * @parent: closure object to wait with
627 * Returns: 1 if @seq has already been flushed, 0 if @seq is being flushed,
628 * -EIO if @seq will never be flushed
630 * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
633 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
634 struct closure *parent)
636 struct journal_buf *buf;
639 if (seq <= j->flushed_seq_ondisk)
644 if (WARN_ONCE(seq > journal_cur_seq(j),
645 "requested to flush journal seq %llu, but currently at %llu",
646 seq, journal_cur_seq(j)))
649 /* Recheck under lock: */
650 if (j->err_seq && seq >= j->err_seq) {
655 if (seq <= j->flushed_seq_ondisk) {
660 /* if seq was written, but not flushed - flush a newer one instead */
661 seq = max(seq, journal_last_unwritten_seq(j));
664 if (seq > journal_cur_seq(j)) {
665 struct journal_res res = { 0 };
667 if (journal_entry_is_open(j))
668 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
670 spin_unlock(&j->lock);
672 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
677 buf = j->buf + (seq & JOURNAL_BUF_MASK);
678 buf->must_flush = true;
680 if (!buf->flush_time) {
681 buf->flush_time = local_clock() ?: 1;
682 buf->expires = jiffies;
685 if (parent && !closure_wait(&buf->wait, parent))
688 bch2_journal_res_put(j, &res);
695 * if write was kicked off without a flush, flush the next sequence
698 buf = journal_seq_to_buf(j, seq);
701 goto recheck_need_open;
704 buf->must_flush = true;
706 if (parent && !closure_wait(&buf->wait, parent))
709 if (seq == journal_cur_seq(j))
710 journal_entry_want_write(j);
712 spin_unlock(&j->lock);
716 int bch2_journal_flush_seq(struct journal *j, u64 seq)
718 u64 start_time = local_clock();
722 * Don't update time_stats when @seq is already flushed:
724 if (seq <= j->flushed_seq_ondisk)
727 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
730 bch2_time_stats_update(j->flush_seq_time, start_time);
732 return ret ?: ret2 < 0 ? ret2 : 0;
736 * bch2_journal_flush_async - if there is an open journal entry, or a journal
737 * still being written, write it and wait for the write to complete
739 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
741 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
744 int bch2_journal_flush(struct journal *j)
746 return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
750 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
753 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
755 struct bch_fs *c = container_of(j, struct bch_fs, journal);
759 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
762 if (seq <= c->journal.flushed_seq_ondisk)
766 if (seq <= c->journal.flushed_seq_ondisk)
769 for (unwritten_seq = journal_last_unwritten_seq(j);
772 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
774 /* journal write is already in flight, and was a flush write: */
775 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
783 spin_unlock(&j->lock);
787 int bch2_journal_meta(struct journal *j)
789 struct journal_buf *buf;
790 struct journal_res res;
793 memset(&res, 0, sizeof(res));
795 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
799 buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
800 buf->must_flush = true;
802 if (!buf->flush_time) {
803 buf->flush_time = local_clock() ?: 1;
804 buf->expires = jiffies;
807 bch2_journal_res_put(j, &res);
809 return bch2_journal_flush_seq(j, res.seq);
812 /* block/unlock the journal: */
814 void bch2_journal_unblock(struct journal *j)
818 spin_unlock(&j->lock);
823 void bch2_journal_block(struct journal *j)
827 spin_unlock(&j->lock);
832 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
834 struct journal_buf *ret = NULL;
836 mutex_lock(&j->buf_lock);
838 max_seq = min(max_seq, journal_cur_seq(j));
840 for (u64 seq = journal_last_unwritten_seq(j);
843 unsigned idx = seq & JOURNAL_BUF_MASK;
844 struct journal_buf *buf = j->buf + idx;
846 if (buf->need_flush_to_write_buffer) {
847 if (seq == journal_cur_seq(j))
848 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
850 union journal_res_state s;
851 s.v = atomic64_read_acquire(&j->reservations.counter);
853 ret = journal_state_count(s, idx)
860 spin_unlock(&j->lock);
861 if (IS_ERR_OR_NULL(ret))
862 mutex_unlock(&j->buf_lock);
866 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
868 struct journal_buf *ret;
870 wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
874 /* allocate journal on a device: */
876 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
877 bool new_fs, struct closure *cl)
879 struct bch_fs *c = ca->fs;
880 struct journal_device *ja = &ca->journal;
881 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
882 struct open_bucket **ob = NULL;
884 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
887 BUG_ON(nr <= ja->nr);
889 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
890 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
891 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL);
892 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL);
893 if (!bu || !ob || !new_buckets || !new_bucket_seq) {
894 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
898 for (nr_got = 0; nr_got < nr_want; nr_got++) {
900 bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
901 if (bu[nr_got] < 0) {
902 ret = -BCH_ERR_ENOSPC_bucket_alloc;
906 ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
907 ret = PTR_ERR_OR_ZERO(ob[nr_got]);
911 ret = bch2_trans_run(c,
912 bch2_trans_mark_metadata_bucket(trans, ca,
913 ob[nr_got]->bucket, BCH_DATA_journal,
914 ca->mi.bucket_size));
916 bch2_open_bucket_put(c, ob[nr_got]);
917 bch_err_msg(c, ret, "marking new journal buckets");
921 bu[nr_got] = ob[nr_got]->bucket;
928 /* Don't return an error if we successfully allocated some buckets: */
932 bch2_journal_flush_all_pins(&c->journal);
933 bch2_journal_block(&c->journal);
934 mutex_lock(&c->sb_lock);
937 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
938 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
940 BUG_ON(ja->discard_idx > ja->nr);
942 pos = ja->discard_idx ?: ja->nr;
944 memmove(new_buckets + pos + nr_got,
946 sizeof(new_buckets[0]) * (ja->nr - pos));
947 memmove(new_bucket_seq + pos + nr_got,
948 new_bucket_seq + pos,
949 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
951 for (i = 0; i < nr_got; i++) {
952 new_buckets[pos + i] = bu[i];
953 new_bucket_seq[pos + i] = 0;
956 nr = ja->nr + nr_got;
958 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
967 spin_lock(&c->journal.lock);
969 swap(new_buckets, ja->buckets);
970 swap(new_bucket_seq, ja->bucket_seq);
973 if (pos <= ja->discard_idx)
974 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
975 if (pos <= ja->dirty_idx_ondisk)
976 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
977 if (pos <= ja->dirty_idx)
978 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
979 if (pos <= ja->cur_idx)
980 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
983 spin_unlock(&c->journal.lock);
986 bch2_journal_unblock(&c->journal);
987 mutex_unlock(&c->sb_lock);
991 for (i = 0; i < nr_got; i++)
993 bch2_trans_mark_metadata_bucket(trans, ca,
994 bu[i], BCH_DATA_free, 0));
997 for (i = 0; i < nr_got; i++)
998 bch2_open_bucket_put(c, ob[i]);
1000 kfree(new_bucket_seq);
1008 * Allocate more journal space at runtime - not currently making use if it, but
1011 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1014 struct journal_device *ja = &ca->journal;
1018 closure_init_stack(&cl);
1020 down_write(&c->state_lock);
1022 /* don't handle reducing nr of buckets yet: */
1026 while (ja->nr < nr) {
1027 struct disk_reservation disk_res = { 0, 0, 0 };
1030 * note: journal buckets aren't really counted as _sectors_ used yet, so
1031 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1032 * when space used goes up without a reservation - but we do need the
1033 * reservation to ensure we'll actually be able to allocate:
1035 * XXX: that's not right, disk reservations only ensure a
1036 * filesystem-wide allocation will succeed, this is a device
1037 * specific allocation - we can hang here:
1040 ret = bch2_disk_reservation_get(c, &disk_res,
1041 bucket_to_sector(ca, nr - ja->nr), 1, 0);
1045 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1047 bch2_disk_reservation_put(c, &disk_res);
1051 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1057 up_write(&c->state_lock);
1061 int bch2_dev_journal_alloc(struct bch_dev *ca)
1066 if (dynamic_fault("bcachefs:add:journal_alloc")) {
1067 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1071 /* 1/128th of the device by default: */
1072 nr = ca->mi.nbuckets >> 7;
1075 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1078 nr = clamp_t(unsigned, nr,
1079 BCH_JOURNAL_BUCKETS_MIN,
1081 (1 << 24) / ca->mi.bucket_size));
1083 ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1085 bch_err_fn(ca, ret);
1089 int bch2_fs_journal_alloc(struct bch_fs *c)
1091 for_each_online_member(c, ca) {
1095 int ret = bch2_dev_journal_alloc(ca);
1097 percpu_ref_put(&ca->io_ref);
1105 /* startup/shutdown: */
1107 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1112 spin_lock(&j->lock);
1113 for (seq = journal_last_unwritten_seq(j);
1114 seq <= journal_cur_seq(j) && !ret;
1116 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1118 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1121 spin_unlock(&j->lock);
1126 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1128 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1131 void bch2_fs_journal_stop(struct journal *j)
1133 bch2_journal_reclaim_stop(j);
1134 bch2_journal_flush_all_pins(j);
1136 wait_event(j->wait, bch2_journal_entry_close(j));
1139 * Always write a new journal entry, to make sure the clock hands are up
1140 * to date (and match the superblock)
1142 bch2_journal_meta(j);
1146 BUG_ON(!bch2_journal_error(j) &&
1147 test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1148 j->last_empty_seq != journal_cur_seq(j));
1150 cancel_delayed_work_sync(&j->write_work);
1153 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1155 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1156 struct journal_entry_pin_list *p;
1157 struct journal_replay *i, **_i;
1158 struct genradix_iter iter;
1159 bool had_entries = false;
1161 u64 last_seq = cur_seq, nr, seq;
1163 genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1166 if (!i || i->ignore)
1169 last_seq = le64_to_cpu(i->j.last_seq);
1173 nr = cur_seq - last_seq;
1175 if (nr + 1 > j->pin.size) {
1177 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1179 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1180 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1184 j->replay_journal_seq = last_seq;
1185 j->replay_journal_seq_end = cur_seq;
1186 j->last_seq_ondisk = last_seq;
1187 j->flushed_seq_ondisk = cur_seq - 1;
1188 j->seq_ondisk = cur_seq - 1;
1189 j->pin.front = last_seq;
1190 j->pin.back = cur_seq;
1191 atomic64_set(&j->seq, cur_seq - 1);
1193 fifo_for_each_entry_ptr(p, &j->pin, seq)
1194 journal_pin_list_init(p, 1);
1196 genradix_for_each(&c->journal_entries, iter, _i) {
1199 if (!i || i->ignore)
1202 seq = le64_to_cpu(i->j.seq);
1203 BUG_ON(seq >= cur_seq);
1208 if (journal_entry_empty(&i->j))
1209 j->last_empty_seq = le64_to_cpu(i->j.seq);
1211 p = journal_seq_pin(j, seq);
1214 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1215 bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1221 j->last_empty_seq = cur_seq;
1223 spin_lock(&j->lock);
1225 set_bit(JOURNAL_STARTED, &j->flags);
1226 j->last_flush_write = jiffies;
1228 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1229 j->reservations.unwritten_idx++;
1231 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1233 bch2_journal_space_available(j);
1234 spin_unlock(&j->lock);
1236 return bch2_journal_reclaim_start(j);
1241 void bch2_dev_journal_exit(struct bch_dev *ca)
1243 kfree(ca->journal.bio);
1244 kfree(ca->journal.buckets);
1245 kfree(ca->journal.bucket_seq);
1247 ca->journal.bio = NULL;
1248 ca->journal.buckets = NULL;
1249 ca->journal.bucket_seq = NULL;
1252 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1254 struct journal_device *ja = &ca->journal;
1255 struct bch_sb_field_journal *journal_buckets =
1256 bch2_sb_field_get(sb, journal);
1257 struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1258 bch2_sb_field_get(sb, journal_v2);
1259 unsigned i, nr_bvecs;
1263 if (journal_buckets_v2) {
1264 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1266 for (i = 0; i < nr; i++)
1267 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1268 } else if (journal_buckets) {
1269 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1272 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1273 if (!ja->bucket_seq)
1274 return -BCH_ERR_ENOMEM_dev_journal_init;
1276 nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1278 ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1279 if (!ca->journal.bio)
1280 return -BCH_ERR_ENOMEM_dev_journal_init;
1282 bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1284 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1286 return -BCH_ERR_ENOMEM_dev_journal_init;
1288 if (journal_buckets_v2) {
1289 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1290 unsigned j, dst = 0;
1292 for (i = 0; i < nr; i++)
1293 for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1294 ja->buckets[dst++] =
1295 le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1296 } else if (journal_buckets) {
1297 for (i = 0; i < ja->nr; i++)
1298 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1304 void bch2_fs_journal_exit(struct journal *j)
1308 darray_exit(&j->early_journal_entries);
1310 for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1311 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1315 int bch2_fs_journal_init(struct journal *j)
1317 static struct lock_class_key res_key;
1320 mutex_init(&j->buf_lock);
1321 spin_lock_init(&j->lock);
1322 spin_lock_init(&j->err_lock);
1323 init_waitqueue_head(&j->wait);
1324 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1325 init_waitqueue_head(&j->reclaim_wait);
1326 init_waitqueue_head(&j->pin_flush_wait);
1327 mutex_init(&j->reclaim_lock);
1328 mutex_init(&j->discard_lock);
1330 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1332 atomic64_set(&j->reservations.counter,
1333 ((union journal_res_state)
1334 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1336 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1337 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1339 for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1340 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1341 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1342 if (!j->buf[i].data)
1343 return -BCH_ERR_ENOMEM_journal_buf;
1346 j->pin.front = j->pin.back = 1;
1352 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1354 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1355 union journal_res_state s;
1356 unsigned long now = jiffies;
1357 u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1359 if (!out->nr_tabstops)
1360 printbuf_tabstop_push(out, 24);
1364 s = READ_ONCE(j->reservations);
1366 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size);
1367 prt_printf(out, "seq:\t\t\t%llu\n", journal_cur_seq(j));
1368 prt_printf(out, "seq_ondisk:\t\t%llu\n", j->seq_ondisk);
1369 prt_printf(out, "last_seq:\t\t%llu\n", journal_last_seq(j));
1370 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk);
1371 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk);
1372 prt_printf(out, "watermark:\t\t%s\n", bch2_watermarks[j->watermark]);
1373 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved);
1374 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes);
1375 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes);
1376 prt_printf(out, "average write size:\t");
1377 prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1379 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim);
1380 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim);
1381 prt_printf(out, "reclaim kicked:\t\t%u\n", j->reclaim_kicked);
1382 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now)
1383 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1384 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors);
1385 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]);
1386 prt_printf(out, "current entry:\t\t");
1388 switch (s.cur_entry_offset) {
1389 case JOURNAL_ENTRY_ERROR_VAL:
1390 prt_printf(out, "error");
1392 case JOURNAL_ENTRY_CLOSED_VAL:
1393 prt_printf(out, "closed");
1396 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1401 prt_printf(out, "unwritten entries:");
1403 bch2_journal_bufs_to_text(out, j);
1406 "replay done:\t\t%i\n",
1407 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1409 prt_printf(out, "space:\n");
1410 prt_printf(out, "\tdiscarded\t%u:%u\n",
1411 j->space[journal_space_discarded].next_entry,
1412 j->space[journal_space_discarded].total);
1413 prt_printf(out, "\tclean ondisk\t%u:%u\n",
1414 j->space[journal_space_clean_ondisk].next_entry,
1415 j->space[journal_space_clean_ondisk].total);
1416 prt_printf(out, "\tclean\t\t%u:%u\n",
1417 j->space[journal_space_clean].next_entry,
1418 j->space[journal_space_clean].total);
1419 prt_printf(out, "\ttotal\t\t%u:%u\n",
1420 j->space[journal_space_total].next_entry,
1421 j->space[journal_space_total].total);
1423 for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1424 struct journal_device *ja = &ca->journal;
1426 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1432 prt_printf(out, "dev %u:\n", ca->dev_idx);
1433 prt_printf(out, "\tnr\t\t%u\n", ja->nr);
1434 prt_printf(out, "\tbucket size\t%u\n", ca->mi.bucket_size);
1435 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1436 prt_printf(out, "\tdiscard_idx\t%u\n", ja->discard_idx);
1437 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]);
1438 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]);
1439 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1447 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1449 spin_lock(&j->lock);
1450 __bch2_journal_debug_to_text(out, j);
1451 spin_unlock(&j->lock);
1454 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1456 struct journal_entry_pin_list *pin_list;
1457 struct journal_entry_pin *pin;
1460 spin_lock(&j->lock);
1461 *seq = max(*seq, j->pin.front);
1463 if (*seq >= j->pin.back) {
1464 spin_unlock(&j->lock);
1470 pin_list = journal_seq_pin(j, *seq);
1472 prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1474 printbuf_indent_add(out, 2);
1476 for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1477 list_for_each_entry(pin, &pin_list->list[i], list) {
1478 prt_printf(out, "\t%px %ps", pin, pin->flush);
1482 if (!list_empty(&pin_list->flushed)) {
1483 prt_printf(out, "flushed:");
1487 list_for_each_entry(pin, &pin_list->flushed, list) {
1488 prt_printf(out, "\t%px %ps", pin, pin->flush);
1492 printbuf_indent_sub(out, 2);
1495 spin_unlock(&j->lock);
1500 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1504 while (!bch2_journal_seq_pins_to_text(out, j, &seq))