xref: /linux/fs/bcachefs/journal.c (revision 24168c5e6dfbdd5b414f048f47f75d64533296ca)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * bcachefs journalling code, for btree insertions
4  *
5  * Copyright 2012 Google, Inc.
6  */
7 
8 #include "bcachefs.h"
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
11 #include "btree_gc.h"
12 #include "btree_update.h"
13 #include "btree_write_buffer.h"
14 #include "buckets.h"
15 #include "error.h"
16 #include "journal.h"
17 #include "journal_io.h"
18 #include "journal_reclaim.h"
19 #include "journal_sb.h"
20 #include "journal_seq_blacklist.h"
21 #include "trace.h"
22 
23 static const char * const bch2_journal_errors[] = {
24 #define x(n)	#n,
25 	JOURNAL_ERRORS()
26 #undef x
27 	NULL
28 };
29 
30 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
31 {
32 	return seq > j->seq_ondisk;
33 }
34 
35 static bool __journal_entry_is_open(union journal_res_state state)
36 {
37 	return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
38 }
39 
40 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
41 {
42 	return atomic64_read(&j->seq) - j->seq_ondisk;
43 }
44 
45 static bool journal_entry_is_open(struct journal *j)
46 {
47 	return __journal_entry_is_open(j->reservations);
48 }
49 
50 static void bch2_journal_buf_to_text(struct printbuf *out, struct journal *j, u64 seq)
51 {
52 	union journal_res_state s = READ_ONCE(j->reservations);
53 	unsigned i = seq & JOURNAL_BUF_MASK;
54 	struct journal_buf *buf = j->buf + i;
55 
56 	prt_printf(out, "seq:\t%llu\n", seq);
57 	printbuf_indent_add(out, 2);
58 
59 	prt_printf(out, "refcount:\t%u\n", journal_state_count(s, i));
60 
61 	prt_printf(out, "size:\t");
62 	prt_human_readable_u64(out, vstruct_bytes(buf->data));
63 	prt_newline(out);
64 
65 	prt_printf(out, "expires:\t");
66 	prt_printf(out, "%li jiffies\n", buf->expires - jiffies);
67 
68 	prt_printf(out, "flags:\t");
69 	if (buf->noflush)
70 		prt_str(out, "noflush ");
71 	if (buf->must_flush)
72 		prt_str(out, "must_flush ");
73 	if (buf->separate_flush)
74 		prt_str(out, "separate_flush ");
75 	if (buf->need_flush_to_write_buffer)
76 		prt_str(out, "need_flush_to_write_buffer ");
77 	if (buf->write_started)
78 		prt_str(out, "write_started ");
79 	if (buf->write_allocated)
80 		prt_str(out, "write_allocated ");
81 	if (buf->write_done)
82 		prt_str(out, "write_done");
83 	prt_newline(out);
84 
85 	printbuf_indent_sub(out, 2);
86 }
87 
88 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j)
89 {
90 	if (!out->nr_tabstops)
91 		printbuf_tabstop_push(out, 24);
92 
93 	for (u64 seq = journal_last_unwritten_seq(j);
94 	     seq <= journal_cur_seq(j);
95 	     seq++)
96 		bch2_journal_buf_to_text(out, j, seq);
97 	prt_printf(out, "last buf %s\n", journal_entry_is_open(j) ? "open" : "closed");
98 }
99 
100 static inline struct journal_buf *
101 journal_seq_to_buf(struct journal *j, u64 seq)
102 {
103 	struct journal_buf *buf = NULL;
104 
105 	EBUG_ON(seq > journal_cur_seq(j));
106 
107 	if (journal_seq_unwritten(j, seq)) {
108 		buf = j->buf + (seq & JOURNAL_BUF_MASK);
109 		EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
110 	}
111 	return buf;
112 }
113 
114 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
115 {
116 	unsigned i;
117 
118 	for (i = 0; i < ARRAY_SIZE(p->list); i++)
119 		INIT_LIST_HEAD(&p->list[i]);
120 	INIT_LIST_HEAD(&p->flushed);
121 	atomic_set(&p->count, count);
122 	p->devs.nr = 0;
123 }
124 
125 /*
126  * Detect stuck journal conditions and trigger shutdown. Technically the journal
127  * can end up stuck for a variety of reasons, such as a blocked I/O, journal
128  * reservation lockup, etc. Since this is a fatal error with potentially
129  * unpredictable characteristics, we want to be fairly conservative before we
130  * decide to shut things down.
131  *
132  * Consider the journal stuck when it appears full with no ability to commit
133  * btree transactions, to discard journal buckets, nor acquire priority
134  * (reserved watermark) reservation.
135  */
136 static inline bool
137 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
138 {
139 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
140 	bool stuck = false;
141 	struct printbuf buf = PRINTBUF;
142 
143 	if (!(error == JOURNAL_ERR_journal_full ||
144 	      error == JOURNAL_ERR_journal_pin_full) ||
145 	    nr_unwritten_journal_entries(j) ||
146 	    (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
147 		return stuck;
148 
149 	spin_lock(&j->lock);
150 
151 	if (j->can_discard) {
152 		spin_unlock(&j->lock);
153 		return stuck;
154 	}
155 
156 	stuck = true;
157 
158 	/*
159 	 * The journal shutdown path will set ->err_seq, but do it here first to
160 	 * serialize against concurrent failures and avoid duplicate error
161 	 * reports.
162 	 */
163 	if (j->err_seq) {
164 		spin_unlock(&j->lock);
165 		return stuck;
166 	}
167 	j->err_seq = journal_cur_seq(j);
168 	spin_unlock(&j->lock);
169 
170 	bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
171 		bch2_journal_errors[error]);
172 	bch2_journal_debug_to_text(&buf, j);
173 	bch_err(c, "%s", buf.buf);
174 
175 	printbuf_reset(&buf);
176 	bch2_journal_pins_to_text(&buf, j);
177 	bch_err(c, "Journal pins:\n%s", buf.buf);
178 	printbuf_exit(&buf);
179 
180 	bch2_fatal_error(c);
181 	dump_stack();
182 
183 	return stuck;
184 }
185 
186 void bch2_journal_do_writes(struct journal *j)
187 {
188 	for (u64 seq = journal_last_unwritten_seq(j);
189 	     seq <= journal_cur_seq(j);
190 	     seq++) {
191 		unsigned idx = seq & JOURNAL_BUF_MASK;
192 		struct journal_buf *w = j->buf + idx;
193 
194 		if (w->write_started && !w->write_allocated)
195 			break;
196 		if (w->write_started)
197 			continue;
198 
199 		if (!journal_state_count(j->reservations, idx)) {
200 			w->write_started = true;
201 			closure_call(&w->io, bch2_journal_write, j->wq, NULL);
202 		}
203 
204 		break;
205 	}
206 }
207 
208 /*
209  * Final processing when the last reference of a journal buffer has been
210  * dropped. Drop the pin list reference acquired at journal entry open and write
211  * the buffer, if requested.
212  */
213 void bch2_journal_buf_put_final(struct journal *j, u64 seq)
214 {
215 	lockdep_assert_held(&j->lock);
216 
217 	if (__bch2_journal_pin_put(j, seq))
218 		bch2_journal_reclaim_fast(j);
219 	bch2_journal_do_writes(j);
220 }
221 
222 /*
223  * Returns true if journal entry is now closed:
224  *
225  * We don't close a journal_buf until the next journal_buf is finished writing,
226  * and can be opened again - this also initializes the next journal_buf:
227  */
228 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace)
229 {
230 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
231 	struct journal_buf *buf = journal_cur_buf(j);
232 	union journal_res_state old, new;
233 	u64 v = atomic64_read(&j->reservations.counter);
234 	unsigned sectors;
235 
236 	BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
237 	       closed_val != JOURNAL_ENTRY_ERROR_VAL);
238 
239 	lockdep_assert_held(&j->lock);
240 
241 	do {
242 		old.v = new.v = v;
243 		new.cur_entry_offset = closed_val;
244 
245 		if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
246 		    old.cur_entry_offset == new.cur_entry_offset)
247 			return;
248 	} while ((v = atomic64_cmpxchg(&j->reservations.counter,
249 				       old.v, new.v)) != old.v);
250 
251 	if (!__journal_entry_is_open(old))
252 		return;
253 
254 	/* Close out old buffer: */
255 	buf->data->u64s		= cpu_to_le32(old.cur_entry_offset);
256 
257 	if (trace_journal_entry_close_enabled() && trace) {
258 		struct printbuf pbuf = PRINTBUF;
259 		pbuf.atomic++;
260 
261 		prt_str(&pbuf, "entry size: ");
262 		prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
263 		prt_newline(&pbuf);
264 		bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT);
265 		trace_journal_entry_close(c, pbuf.buf);
266 		printbuf_exit(&pbuf);
267 	}
268 
269 	sectors = vstruct_blocks_plus(buf->data, c->block_bits,
270 				      buf->u64s_reserved) << c->block_bits;
271 	BUG_ON(sectors > buf->sectors);
272 	buf->sectors = sectors;
273 
274 	/*
275 	 * We have to set last_seq here, _before_ opening a new journal entry:
276 	 *
277 	 * A threads may replace an old pin with a new pin on their current
278 	 * journal reservation - the expectation being that the journal will
279 	 * contain either what the old pin protected or what the new pin
280 	 * protects.
281 	 *
282 	 * After the old pin is dropped journal_last_seq() won't include the old
283 	 * pin, so we can only write the updated last_seq on the entry that
284 	 * contains whatever the new pin protects.
285 	 *
286 	 * Restated, we can _not_ update last_seq for a given entry if there
287 	 * could be a newer entry open with reservations/pins that have been
288 	 * taken against it.
289 	 *
290 	 * Hence, we want update/set last_seq on the current journal entry right
291 	 * before we open a new one:
292 	 */
293 	buf->last_seq		= journal_last_seq(j);
294 	buf->data->last_seq	= cpu_to_le64(buf->last_seq);
295 	BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
296 
297 	cancel_delayed_work(&j->write_work);
298 
299 	bch2_journal_space_available(j);
300 
301 	__bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
302 }
303 
304 void bch2_journal_halt(struct journal *j)
305 {
306 	spin_lock(&j->lock);
307 	__journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true);
308 	if (!j->err_seq)
309 		j->err_seq = journal_cur_seq(j);
310 	journal_wake(j);
311 	spin_unlock(&j->lock);
312 }
313 
314 static bool journal_entry_want_write(struct journal *j)
315 {
316 	bool ret = !journal_entry_is_open(j) ||
317 		journal_cur_seq(j) == journal_last_unwritten_seq(j);
318 
319 	/* Don't close it yet if we already have a write in flight: */
320 	if (ret)
321 		__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
322 	else if (nr_unwritten_journal_entries(j)) {
323 		struct journal_buf *buf = journal_cur_buf(j);
324 
325 		if (!buf->flush_time) {
326 			buf->flush_time	= local_clock() ?: 1;
327 			buf->expires = jiffies;
328 		}
329 	}
330 
331 	return ret;
332 }
333 
334 bool bch2_journal_entry_close(struct journal *j)
335 {
336 	bool ret;
337 
338 	spin_lock(&j->lock);
339 	ret = journal_entry_want_write(j);
340 	spin_unlock(&j->lock);
341 
342 	return ret;
343 }
344 
345 /*
346  * should _only_ called from journal_res_get() - when we actually want a
347  * journal reservation - journal entry is open means journal is dirty:
348  */
349 static int journal_entry_open(struct journal *j)
350 {
351 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
352 	struct journal_buf *buf = j->buf +
353 		((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
354 	union journal_res_state old, new;
355 	int u64s;
356 	u64 v;
357 
358 	lockdep_assert_held(&j->lock);
359 	BUG_ON(journal_entry_is_open(j));
360 	BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
361 
362 	if (j->blocked)
363 		return JOURNAL_ERR_blocked;
364 
365 	if (j->cur_entry_error)
366 		return j->cur_entry_error;
367 
368 	if (bch2_journal_error(j))
369 		return JOURNAL_ERR_insufficient_devices; /* -EROFS */
370 
371 	if (!fifo_free(&j->pin))
372 		return JOURNAL_ERR_journal_pin_full;
373 
374 	if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
375 		return JOURNAL_ERR_max_in_flight;
376 
377 	BUG_ON(!j->cur_entry_sectors);
378 
379 	buf->expires		=
380 		(journal_cur_seq(j) == j->flushed_seq_ondisk
381 		 ? jiffies
382 		 : j->last_flush_write) +
383 		msecs_to_jiffies(c->opts.journal_flush_delay);
384 
385 	buf->u64s_reserved	= j->entry_u64s_reserved;
386 	buf->disk_sectors	= j->cur_entry_sectors;
387 	buf->sectors		= min(buf->disk_sectors, buf->buf_size >> 9);
388 
389 	u64s = (int) (buf->sectors << 9) / sizeof(u64) -
390 		journal_entry_overhead(j);
391 	u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
392 
393 	if (u64s <= (ssize_t) j->early_journal_entries.nr)
394 		return JOURNAL_ERR_journal_full;
395 
396 	if (fifo_empty(&j->pin) && j->reclaim_thread)
397 		wake_up_process(j->reclaim_thread);
398 
399 	/*
400 	 * The fifo_push() needs to happen at the same time as j->seq is
401 	 * incremented for journal_last_seq() to be calculated correctly
402 	 */
403 	atomic64_inc(&j->seq);
404 	journal_pin_list_init(fifo_push_ref(&j->pin), 1);
405 
406 	BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
407 
408 	BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
409 
410 	bkey_extent_init(&buf->key);
411 	buf->noflush		= false;
412 	buf->must_flush		= false;
413 	buf->separate_flush	= false;
414 	buf->flush_time		= 0;
415 	buf->need_flush_to_write_buffer = true;
416 	buf->write_started	= false;
417 	buf->write_allocated	= false;
418 	buf->write_done		= false;
419 
420 	memset(buf->data, 0, sizeof(*buf->data));
421 	buf->data->seq	= cpu_to_le64(journal_cur_seq(j));
422 	buf->data->u64s	= 0;
423 
424 	if (j->early_journal_entries.nr) {
425 		memcpy(buf->data->_data, j->early_journal_entries.data,
426 		       j->early_journal_entries.nr * sizeof(u64));
427 		le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
428 	}
429 
430 	/*
431 	 * Must be set before marking the journal entry as open:
432 	 */
433 	j->cur_entry_u64s = u64s;
434 
435 	v = atomic64_read(&j->reservations.counter);
436 	do {
437 		old.v = new.v = v;
438 
439 		BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
440 
441 		new.idx++;
442 		BUG_ON(journal_state_count(new, new.idx));
443 		BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
444 
445 		journal_state_inc(&new);
446 
447 		/* Handle any already added entries */
448 		new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
449 	} while ((v = atomic64_cmpxchg(&j->reservations.counter,
450 				       old.v, new.v)) != old.v);
451 
452 	if (nr_unwritten_journal_entries(j) == 1)
453 		mod_delayed_work(j->wq,
454 				 &j->write_work,
455 				 msecs_to_jiffies(c->opts.journal_flush_delay));
456 	journal_wake(j);
457 
458 	if (j->early_journal_entries.nr)
459 		darray_exit(&j->early_journal_entries);
460 	return 0;
461 }
462 
463 static bool journal_quiesced(struct journal *j)
464 {
465 	bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
466 
467 	if (!ret)
468 		bch2_journal_entry_close(j);
469 	return ret;
470 }
471 
472 static void journal_quiesce(struct journal *j)
473 {
474 	wait_event(j->wait, journal_quiesced(j));
475 }
476 
477 static void journal_write_work(struct work_struct *work)
478 {
479 	struct journal *j = container_of(work, struct journal, write_work.work);
480 
481 	spin_lock(&j->lock);
482 	if (__journal_entry_is_open(j->reservations)) {
483 		long delta = journal_cur_buf(j)->expires - jiffies;
484 
485 		if (delta > 0)
486 			mod_delayed_work(j->wq, &j->write_work, delta);
487 		else
488 			__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
489 	}
490 	spin_unlock(&j->lock);
491 }
492 
493 static int __journal_res_get(struct journal *j, struct journal_res *res,
494 			     unsigned flags)
495 {
496 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
497 	struct journal_buf *buf;
498 	bool can_discard;
499 	int ret;
500 retry:
501 	if (journal_res_get_fast(j, res, flags))
502 		return 0;
503 
504 	if (bch2_journal_error(j))
505 		return -BCH_ERR_erofs_journal_err;
506 
507 	if (j->blocked)
508 		return -BCH_ERR_journal_res_get_blocked;
509 
510 	if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
511 		ret = JOURNAL_ERR_journal_full;
512 		can_discard = j->can_discard;
513 		goto out;
514 	}
515 
516 	if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) {
517 		ret = JOURNAL_ERR_max_in_flight;
518 		goto out;
519 	}
520 
521 	spin_lock(&j->lock);
522 
523 	/*
524 	 * Recheck after taking the lock, so we don't race with another thread
525 	 * that just did journal_entry_open() and call bch2_journal_entry_close()
526 	 * unnecessarily
527 	 */
528 	if (journal_res_get_fast(j, res, flags)) {
529 		ret = 0;
530 		goto unlock;
531 	}
532 
533 	/*
534 	 * If we couldn't get a reservation because the current buf filled up,
535 	 * and we had room for a bigger entry on disk, signal that we want to
536 	 * realloc the journal bufs:
537 	 */
538 	buf = journal_cur_buf(j);
539 	if (journal_entry_is_open(j) &&
540 	    buf->buf_size >> 9 < buf->disk_sectors &&
541 	    buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
542 		j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
543 
544 	__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
545 	ret = journal_entry_open(j) ?: JOURNAL_ERR_retry;
546 unlock:
547 	can_discard = j->can_discard;
548 	spin_unlock(&j->lock);
549 out:
550 	if (ret == JOURNAL_ERR_retry)
551 		goto retry;
552 	if (!ret)
553 		return 0;
554 
555 	if (journal_error_check_stuck(j, ret, flags))
556 		ret = -BCH_ERR_journal_res_get_blocked;
557 
558 	if (ret == JOURNAL_ERR_max_in_flight &&
559 	    track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true)) {
560 
561 		struct printbuf buf = PRINTBUF;
562 		prt_printf(&buf, "seq %llu\n", journal_cur_seq(j));
563 		bch2_journal_bufs_to_text(&buf, j);
564 		trace_journal_entry_full(c, buf.buf);
565 		printbuf_exit(&buf);
566 		count_event(c, journal_entry_full);
567 	}
568 
569 	/*
570 	 * Journal is full - can't rely on reclaim from work item due to
571 	 * freezing:
572 	 */
573 	if ((ret == JOURNAL_ERR_journal_full ||
574 	     ret == JOURNAL_ERR_journal_pin_full) &&
575 	    !(flags & JOURNAL_RES_GET_NONBLOCK)) {
576 		if (can_discard) {
577 			bch2_journal_do_discards(j);
578 			goto retry;
579 		}
580 
581 		if (mutex_trylock(&j->reclaim_lock)) {
582 			bch2_journal_reclaim(j);
583 			mutex_unlock(&j->reclaim_lock);
584 		}
585 	}
586 
587 	return ret == JOURNAL_ERR_insufficient_devices
588 		? -BCH_ERR_erofs_journal_err
589 		: -BCH_ERR_journal_res_get_blocked;
590 }
591 
592 /*
593  * Essentially the entry function to the journaling code. When bcachefs is doing
594  * a btree insert, it calls this function to get the current journal write.
595  * Journal write is the structure used set up journal writes. The calling
596  * function will then add its keys to the structure, queuing them for the next
597  * write.
598  *
599  * To ensure forward progress, the current task must not be holding any
600  * btree node write locks.
601  */
602 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
603 				  unsigned flags)
604 {
605 	int ret;
606 
607 	closure_wait_event(&j->async_wait,
608 		   (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
609 		   (flags & JOURNAL_RES_GET_NONBLOCK));
610 	return ret;
611 }
612 
613 /* journal_entry_res: */
614 
615 void bch2_journal_entry_res_resize(struct journal *j,
616 				   struct journal_entry_res *res,
617 				   unsigned new_u64s)
618 {
619 	union journal_res_state state;
620 	int d = new_u64s - res->u64s;
621 
622 	spin_lock(&j->lock);
623 
624 	j->entry_u64s_reserved += d;
625 	if (d <= 0)
626 		goto out;
627 
628 	j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
629 	smp_mb();
630 	state = READ_ONCE(j->reservations);
631 
632 	if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
633 	    state.cur_entry_offset > j->cur_entry_u64s) {
634 		j->cur_entry_u64s += d;
635 		/*
636 		 * Not enough room in current journal entry, have to flush it:
637 		 */
638 		__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
639 	} else {
640 		journal_cur_buf(j)->u64s_reserved += d;
641 	}
642 out:
643 	spin_unlock(&j->lock);
644 	res->u64s += d;
645 }
646 
647 /* journal flushing: */
648 
649 /**
650  * bch2_journal_flush_seq_async - wait for a journal entry to be written
651  * @j:		journal object
652  * @seq:	seq to flush
653  * @parent:	closure object to wait with
654  * Returns:	1 if @seq has already been flushed, 0 if @seq is being flushed,
655  *		-EIO if @seq will never be flushed
656  *
657  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
658  * necessary
659  */
660 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
661 				 struct closure *parent)
662 {
663 	struct journal_buf *buf;
664 	int ret = 0;
665 
666 	if (seq <= j->flushed_seq_ondisk)
667 		return 1;
668 
669 	spin_lock(&j->lock);
670 
671 	if (WARN_ONCE(seq > journal_cur_seq(j),
672 		      "requested to flush journal seq %llu, but currently at %llu",
673 		      seq, journal_cur_seq(j)))
674 		goto out;
675 
676 	/* Recheck under lock: */
677 	if (j->err_seq && seq >= j->err_seq) {
678 		ret = -EIO;
679 		goto out;
680 	}
681 
682 	if (seq <= j->flushed_seq_ondisk) {
683 		ret = 1;
684 		goto out;
685 	}
686 
687 	/* if seq was written, but not flushed - flush a newer one instead */
688 	seq = max(seq, journal_last_unwritten_seq(j));
689 
690 recheck_need_open:
691 	if (seq > journal_cur_seq(j)) {
692 		struct journal_res res = { 0 };
693 
694 		if (journal_entry_is_open(j))
695 			__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
696 
697 		spin_unlock(&j->lock);
698 
699 		/*
700 		 * We're called from bch2_journal_flush_seq() -> wait_event();
701 		 * but this might block. We won't usually block, so we won't
702 		 * livelock:
703 		 */
704 		sched_annotate_sleep();
705 		ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
706 		if (ret)
707 			return ret;
708 
709 		seq = res.seq;
710 		buf = journal_seq_to_buf(j, seq);
711 		buf->must_flush = true;
712 
713 		if (!buf->flush_time) {
714 			buf->flush_time	= local_clock() ?: 1;
715 			buf->expires = jiffies;
716 		}
717 
718 		if (parent && !closure_wait(&buf->wait, parent))
719 			BUG();
720 
721 		bch2_journal_res_put(j, &res);
722 
723 		spin_lock(&j->lock);
724 		goto want_write;
725 	}
726 
727 	/*
728 	 * if write was kicked off without a flush, or if we promised it
729 	 * wouldn't be a flush, flush the next sequence number instead
730 	 */
731 	buf = journal_seq_to_buf(j, seq);
732 	if (buf->noflush) {
733 		seq++;
734 		goto recheck_need_open;
735 	}
736 
737 	buf->must_flush = true;
738 
739 	if (parent && !closure_wait(&buf->wait, parent))
740 		BUG();
741 want_write:
742 	if (seq == journal_cur_seq(j))
743 		journal_entry_want_write(j);
744 out:
745 	spin_unlock(&j->lock);
746 	return ret;
747 }
748 
749 int bch2_journal_flush_seq(struct journal *j, u64 seq)
750 {
751 	u64 start_time = local_clock();
752 	int ret, ret2;
753 
754 	/*
755 	 * Don't update time_stats when @seq is already flushed:
756 	 */
757 	if (seq <= j->flushed_seq_ondisk)
758 		return 0;
759 
760 	ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
761 
762 	if (!ret)
763 		bch2_time_stats_update(j->flush_seq_time, start_time);
764 
765 	return ret ?: ret2 < 0 ? ret2 : 0;
766 }
767 
768 /*
769  * bch2_journal_flush_async - if there is an open journal entry, or a journal
770  * still being written, write it and wait for the write to complete
771  */
772 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
773 {
774 	bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
775 }
776 
777 int bch2_journal_flush(struct journal *j)
778 {
779 	return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
780 }
781 
782 /*
783  * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
784  * @seq
785  */
786 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
787 {
788 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
789 	u64 unwritten_seq;
790 	bool ret = false;
791 
792 	if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
793 		return false;
794 
795 	if (seq <= c->journal.flushed_seq_ondisk)
796 		return false;
797 
798 	spin_lock(&j->lock);
799 	if (seq <= c->journal.flushed_seq_ondisk)
800 		goto out;
801 
802 	for (unwritten_seq = journal_last_unwritten_seq(j);
803 	     unwritten_seq < seq;
804 	     unwritten_seq++) {
805 		struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
806 
807 		/* journal flush already in flight, or flush requseted */
808 		if (buf->must_flush)
809 			goto out;
810 
811 		buf->noflush = true;
812 	}
813 
814 	ret = true;
815 out:
816 	spin_unlock(&j->lock);
817 	return ret;
818 }
819 
820 int bch2_journal_meta(struct journal *j)
821 {
822 	struct journal_buf *buf;
823 	struct journal_res res;
824 	int ret;
825 
826 	memset(&res, 0, sizeof(res));
827 
828 	ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
829 	if (ret)
830 		return ret;
831 
832 	buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
833 	buf->must_flush = true;
834 
835 	if (!buf->flush_time) {
836 		buf->flush_time	= local_clock() ?: 1;
837 		buf->expires = jiffies;
838 	}
839 
840 	bch2_journal_res_put(j, &res);
841 
842 	return bch2_journal_flush_seq(j, res.seq);
843 }
844 
845 /* block/unlock the journal: */
846 
847 void bch2_journal_unblock(struct journal *j)
848 {
849 	spin_lock(&j->lock);
850 	j->blocked--;
851 	spin_unlock(&j->lock);
852 
853 	journal_wake(j);
854 }
855 
856 void bch2_journal_block(struct journal *j)
857 {
858 	spin_lock(&j->lock);
859 	j->blocked++;
860 	spin_unlock(&j->lock);
861 
862 	journal_quiesce(j);
863 }
864 
865 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
866 {
867 	struct journal_buf *ret = NULL;
868 
869 	/* We're inside wait_event(), but using mutex_lock(: */
870 	sched_annotate_sleep();
871 	mutex_lock(&j->buf_lock);
872 	spin_lock(&j->lock);
873 	max_seq = min(max_seq, journal_cur_seq(j));
874 
875 	for (u64 seq = journal_last_unwritten_seq(j);
876 	     seq <= max_seq;
877 	     seq++) {
878 		unsigned idx = seq & JOURNAL_BUF_MASK;
879 		struct journal_buf *buf = j->buf + idx;
880 
881 		if (buf->need_flush_to_write_buffer) {
882 			if (seq == journal_cur_seq(j))
883 				__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
884 
885 			union journal_res_state s;
886 			s.v = atomic64_read_acquire(&j->reservations.counter);
887 
888 			ret = journal_state_count(s, idx)
889 				? ERR_PTR(-EAGAIN)
890 				: buf;
891 			break;
892 		}
893 	}
894 
895 	spin_unlock(&j->lock);
896 	if (IS_ERR_OR_NULL(ret))
897 		mutex_unlock(&j->buf_lock);
898 	return ret;
899 }
900 
901 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
902 {
903 	struct journal_buf *ret;
904 
905 	wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
906 	return ret;
907 }
908 
909 /* allocate journal on a device: */
910 
911 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
912 					 bool new_fs, struct closure *cl)
913 {
914 	struct bch_fs *c = ca->fs;
915 	struct journal_device *ja = &ca->journal;
916 	u64 *new_bucket_seq = NULL, *new_buckets = NULL;
917 	struct open_bucket **ob = NULL;
918 	long *bu = NULL;
919 	unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
920 	int ret = 0;
921 
922 	BUG_ON(nr <= ja->nr);
923 
924 	bu		= kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
925 	ob		= kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
926 	new_buckets	= kcalloc(nr, sizeof(u64), GFP_KERNEL);
927 	new_bucket_seq	= kcalloc(nr, sizeof(u64), GFP_KERNEL);
928 	if (!bu || !ob || !new_buckets || !new_bucket_seq) {
929 		ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
930 		goto err_free;
931 	}
932 
933 	for (nr_got = 0; nr_got < nr_want; nr_got++) {
934 		if (new_fs) {
935 			bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
936 			if (bu[nr_got] < 0) {
937 				ret = -BCH_ERR_ENOSPC_bucket_alloc;
938 				break;
939 			}
940 		} else {
941 			ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal,
942 						       BCH_DATA_journal, cl);
943 			ret = PTR_ERR_OR_ZERO(ob[nr_got]);
944 			if (ret)
945 				break;
946 
947 			ret = bch2_trans_run(c,
948 				bch2_trans_mark_metadata_bucket(trans, ca,
949 						ob[nr_got]->bucket, BCH_DATA_journal,
950 						ca->mi.bucket_size, BTREE_TRIGGER_transactional));
951 			if (ret) {
952 				bch2_open_bucket_put(c, ob[nr_got]);
953 				bch_err_msg(c, ret, "marking new journal buckets");
954 				break;
955 			}
956 
957 			bu[nr_got] = ob[nr_got]->bucket;
958 		}
959 	}
960 
961 	if (!nr_got)
962 		goto err_free;
963 
964 	/* Don't return an error if we successfully allocated some buckets: */
965 	ret = 0;
966 
967 	if (c) {
968 		bch2_journal_flush_all_pins(&c->journal);
969 		bch2_journal_block(&c->journal);
970 		mutex_lock(&c->sb_lock);
971 	}
972 
973 	memcpy(new_buckets,	ja->buckets,	ja->nr * sizeof(u64));
974 	memcpy(new_bucket_seq,	ja->bucket_seq,	ja->nr * sizeof(u64));
975 
976 	BUG_ON(ja->discard_idx > ja->nr);
977 
978 	pos = ja->discard_idx ?: ja->nr;
979 
980 	memmove(new_buckets + pos + nr_got,
981 		new_buckets + pos,
982 		sizeof(new_buckets[0]) * (ja->nr - pos));
983 	memmove(new_bucket_seq + pos + nr_got,
984 		new_bucket_seq + pos,
985 		sizeof(new_bucket_seq[0]) * (ja->nr - pos));
986 
987 	for (i = 0; i < nr_got; i++) {
988 		new_buckets[pos + i] = bu[i];
989 		new_bucket_seq[pos + i] = 0;
990 	}
991 
992 	nr = ja->nr + nr_got;
993 
994 	ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
995 	if (ret)
996 		goto err_unblock;
997 
998 	if (!new_fs)
999 		bch2_write_super(c);
1000 
1001 	/* Commit: */
1002 	if (c)
1003 		spin_lock(&c->journal.lock);
1004 
1005 	swap(new_buckets,	ja->buckets);
1006 	swap(new_bucket_seq,	ja->bucket_seq);
1007 	ja->nr = nr;
1008 
1009 	if (pos <= ja->discard_idx)
1010 		ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
1011 	if (pos <= ja->dirty_idx_ondisk)
1012 		ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
1013 	if (pos <= ja->dirty_idx)
1014 		ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
1015 	if (pos <= ja->cur_idx)
1016 		ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
1017 
1018 	if (c)
1019 		spin_unlock(&c->journal.lock);
1020 err_unblock:
1021 	if (c) {
1022 		bch2_journal_unblock(&c->journal);
1023 		mutex_unlock(&c->sb_lock);
1024 	}
1025 
1026 	if (ret && !new_fs)
1027 		for (i = 0; i < nr_got; i++)
1028 			bch2_trans_run(c,
1029 				bch2_trans_mark_metadata_bucket(trans, ca,
1030 						bu[i], BCH_DATA_free, 0,
1031 						BTREE_TRIGGER_transactional));
1032 err_free:
1033 	if (!new_fs)
1034 		for (i = 0; i < nr_got; i++)
1035 			bch2_open_bucket_put(c, ob[i]);
1036 
1037 	kfree(new_bucket_seq);
1038 	kfree(new_buckets);
1039 	kfree(ob);
1040 	kfree(bu);
1041 	return ret;
1042 }
1043 
1044 /*
1045  * Allocate more journal space at runtime - not currently making use if it, but
1046  * the code works:
1047  */
1048 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1049 				unsigned nr)
1050 {
1051 	struct journal_device *ja = &ca->journal;
1052 	struct closure cl;
1053 	int ret = 0;
1054 
1055 	closure_init_stack(&cl);
1056 
1057 	down_write(&c->state_lock);
1058 
1059 	/* don't handle reducing nr of buckets yet: */
1060 	if (nr < ja->nr)
1061 		goto unlock;
1062 
1063 	while (ja->nr < nr) {
1064 		struct disk_reservation disk_res = { 0, 0, 0 };
1065 
1066 		/*
1067 		 * note: journal buckets aren't really counted as _sectors_ used yet, so
1068 		 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1069 		 * when space used goes up without a reservation - but we do need the
1070 		 * reservation to ensure we'll actually be able to allocate:
1071 		 *
1072 		 * XXX: that's not right, disk reservations only ensure a
1073 		 * filesystem-wide allocation will succeed, this is a device
1074 		 * specific allocation - we can hang here:
1075 		 */
1076 
1077 		ret = bch2_disk_reservation_get(c, &disk_res,
1078 						bucket_to_sector(ca, nr - ja->nr), 1, 0);
1079 		if (ret)
1080 			break;
1081 
1082 		ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1083 
1084 		bch2_disk_reservation_put(c, &disk_res);
1085 
1086 		closure_sync(&cl);
1087 
1088 		if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1089 			break;
1090 	}
1091 
1092 	bch_err_fn(c, ret);
1093 unlock:
1094 	up_write(&c->state_lock);
1095 	return ret;
1096 }
1097 
1098 int bch2_dev_journal_alloc(struct bch_dev *ca)
1099 {
1100 	unsigned nr;
1101 	int ret;
1102 
1103 	if (dynamic_fault("bcachefs:add:journal_alloc")) {
1104 		ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1105 		goto err;
1106 	}
1107 
1108 	/* 1/128th of the device by default: */
1109 	nr = ca->mi.nbuckets >> 7;
1110 
1111 	/*
1112 	 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1113 	 * is smaller:
1114 	 */
1115 	nr = clamp_t(unsigned, nr,
1116 		     BCH_JOURNAL_BUCKETS_MIN,
1117 		     min(1 << 13,
1118 			 (1 << 24) / ca->mi.bucket_size));
1119 
1120 	ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1121 err:
1122 	bch_err_fn(ca, ret);
1123 	return ret;
1124 }
1125 
1126 int bch2_fs_journal_alloc(struct bch_fs *c)
1127 {
1128 	for_each_online_member(c, ca) {
1129 		if (ca->journal.nr)
1130 			continue;
1131 
1132 		int ret = bch2_dev_journal_alloc(ca);
1133 		if (ret) {
1134 			percpu_ref_put(&ca->io_ref);
1135 			return ret;
1136 		}
1137 	}
1138 
1139 	return 0;
1140 }
1141 
1142 /* startup/shutdown: */
1143 
1144 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1145 {
1146 	bool ret = false;
1147 	u64 seq;
1148 
1149 	spin_lock(&j->lock);
1150 	for (seq = journal_last_unwritten_seq(j);
1151 	     seq <= journal_cur_seq(j) && !ret;
1152 	     seq++) {
1153 		struct journal_buf *buf = journal_seq_to_buf(j, seq);
1154 
1155 		if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1156 			ret = true;
1157 	}
1158 	spin_unlock(&j->lock);
1159 
1160 	return ret;
1161 }
1162 
1163 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1164 {
1165 	wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1166 }
1167 
1168 void bch2_fs_journal_stop(struct journal *j)
1169 {
1170 	bch2_journal_reclaim_stop(j);
1171 	bch2_journal_flush_all_pins(j);
1172 
1173 	wait_event(j->wait, bch2_journal_entry_close(j));
1174 
1175 	/*
1176 	 * Always write a new journal entry, to make sure the clock hands are up
1177 	 * to date (and match the superblock)
1178 	 */
1179 	bch2_journal_meta(j);
1180 
1181 	journal_quiesce(j);
1182 	cancel_delayed_work_sync(&j->write_work);
1183 
1184 	BUG_ON(!bch2_journal_error(j) &&
1185 	       test_bit(JOURNAL_replay_done, &j->flags) &&
1186 	       j->last_empty_seq != journal_cur_seq(j));
1187 
1188 	if (!bch2_journal_error(j))
1189 		clear_bit(JOURNAL_running, &j->flags);
1190 }
1191 
1192 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1193 {
1194 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
1195 	struct journal_entry_pin_list *p;
1196 	struct journal_replay *i, **_i;
1197 	struct genradix_iter iter;
1198 	bool had_entries = false;
1199 	u64 last_seq = cur_seq, nr, seq;
1200 
1201 	genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1202 		i = *_i;
1203 
1204 		if (journal_replay_ignore(i))
1205 			continue;
1206 
1207 		last_seq = le64_to_cpu(i->j.last_seq);
1208 		break;
1209 	}
1210 
1211 	nr = cur_seq - last_seq;
1212 
1213 	if (nr + 1 > j->pin.size) {
1214 		free_fifo(&j->pin);
1215 		init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1216 		if (!j->pin.data) {
1217 			bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1218 			return -BCH_ERR_ENOMEM_journal_pin_fifo;
1219 		}
1220 	}
1221 
1222 	j->replay_journal_seq	= last_seq;
1223 	j->replay_journal_seq_end = cur_seq;
1224 	j->last_seq_ondisk	= last_seq;
1225 	j->flushed_seq_ondisk	= cur_seq - 1;
1226 	j->seq_ondisk		= cur_seq - 1;
1227 	j->pin.front		= last_seq;
1228 	j->pin.back		= cur_seq;
1229 	atomic64_set(&j->seq, cur_seq - 1);
1230 
1231 	fifo_for_each_entry_ptr(p, &j->pin, seq)
1232 		journal_pin_list_init(p, 1);
1233 
1234 	genradix_for_each(&c->journal_entries, iter, _i) {
1235 		i = *_i;
1236 
1237 		if (journal_replay_ignore(i))
1238 			continue;
1239 
1240 		seq = le64_to_cpu(i->j.seq);
1241 		BUG_ON(seq >= cur_seq);
1242 
1243 		if (seq < last_seq)
1244 			continue;
1245 
1246 		if (journal_entry_empty(&i->j))
1247 			j->last_empty_seq = le64_to_cpu(i->j.seq);
1248 
1249 		p = journal_seq_pin(j, seq);
1250 
1251 		p->devs.nr = 0;
1252 		darray_for_each(i->ptrs, ptr)
1253 			bch2_dev_list_add_dev(&p->devs, ptr->dev);
1254 
1255 		had_entries = true;
1256 	}
1257 
1258 	if (!had_entries)
1259 		j->last_empty_seq = cur_seq;
1260 
1261 	spin_lock(&j->lock);
1262 
1263 	set_bit(JOURNAL_running, &j->flags);
1264 	j->last_flush_write = jiffies;
1265 
1266 	j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1267 	j->reservations.unwritten_idx++;
1268 
1269 	c->last_bucket_seq_cleanup = journal_cur_seq(j);
1270 
1271 	bch2_journal_space_available(j);
1272 	spin_unlock(&j->lock);
1273 
1274 	return bch2_journal_reclaim_start(j);
1275 }
1276 
1277 /* init/exit: */
1278 
1279 void bch2_dev_journal_exit(struct bch_dev *ca)
1280 {
1281 	struct journal_device *ja = &ca->journal;
1282 
1283 	for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1284 		kfree(ja->bio[i]);
1285 		ja->bio[i] = NULL;
1286 	}
1287 
1288 	kfree(ja->buckets);
1289 	kfree(ja->bucket_seq);
1290 	ja->buckets	= NULL;
1291 	ja->bucket_seq	= NULL;
1292 }
1293 
1294 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1295 {
1296 	struct journal_device *ja = &ca->journal;
1297 	struct bch_sb_field_journal *journal_buckets =
1298 		bch2_sb_field_get(sb, journal);
1299 	struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1300 		bch2_sb_field_get(sb, journal_v2);
1301 
1302 	ja->nr = 0;
1303 
1304 	if (journal_buckets_v2) {
1305 		unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1306 
1307 		for (unsigned i = 0; i < nr; i++)
1308 			ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1309 	} else if (journal_buckets) {
1310 		ja->nr = bch2_nr_journal_buckets(journal_buckets);
1311 	}
1312 
1313 	ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1314 	if (!ja->bucket_seq)
1315 		return -BCH_ERR_ENOMEM_dev_journal_init;
1316 
1317 	unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1318 
1319 	for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1320 		ja->bio[i] = kmalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
1321 				     nr_bvecs), GFP_KERNEL);
1322 		if (!ja->bio[i])
1323 			return -BCH_ERR_ENOMEM_dev_journal_init;
1324 
1325 		ja->bio[i]->ca = ca;
1326 		ja->bio[i]->buf_idx = i;
1327 		bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0);
1328 	}
1329 
1330 	ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1331 	if (!ja->buckets)
1332 		return -BCH_ERR_ENOMEM_dev_journal_init;
1333 
1334 	if (journal_buckets_v2) {
1335 		unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1336 		unsigned dst = 0;
1337 
1338 		for (unsigned i = 0; i < nr; i++)
1339 			for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1340 				ja->buckets[dst++] =
1341 					le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1342 	} else if (journal_buckets) {
1343 		for (unsigned i = 0; i < ja->nr; i++)
1344 			ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1345 	}
1346 
1347 	return 0;
1348 }
1349 
1350 void bch2_fs_journal_exit(struct journal *j)
1351 {
1352 	if (j->wq)
1353 		destroy_workqueue(j->wq);
1354 
1355 	darray_exit(&j->early_journal_entries);
1356 
1357 	for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++)
1358 		kvfree(j->buf[i].data);
1359 	free_fifo(&j->pin);
1360 }
1361 
1362 int bch2_fs_journal_init(struct journal *j)
1363 {
1364 	static struct lock_class_key res_key;
1365 
1366 	mutex_init(&j->buf_lock);
1367 	spin_lock_init(&j->lock);
1368 	spin_lock_init(&j->err_lock);
1369 	init_waitqueue_head(&j->wait);
1370 	INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1371 	init_waitqueue_head(&j->reclaim_wait);
1372 	init_waitqueue_head(&j->pin_flush_wait);
1373 	mutex_init(&j->reclaim_lock);
1374 	mutex_init(&j->discard_lock);
1375 
1376 	lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1377 
1378 	atomic64_set(&j->reservations.counter,
1379 		((union journal_res_state)
1380 		 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1381 
1382 	if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1383 		return -BCH_ERR_ENOMEM_journal_pin_fifo;
1384 
1385 	for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) {
1386 		j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1387 		j->buf[i].data = kvmalloc(j->buf[i].buf_size, GFP_KERNEL);
1388 		if (!j->buf[i].data)
1389 			return -BCH_ERR_ENOMEM_journal_buf;
1390 		j->buf[i].idx = i;
1391 	}
1392 
1393 	j->pin.front = j->pin.back = 1;
1394 
1395 	j->wq = alloc_workqueue("bcachefs_journal",
1396 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512);
1397 	if (!j->wq)
1398 		return -BCH_ERR_ENOMEM_fs_other_alloc;
1399 	return 0;
1400 }
1401 
1402 /* debug: */
1403 
1404 static const char * const bch2_journal_flags_strs[] = {
1405 #define x(n)	#n,
1406 	JOURNAL_FLAGS()
1407 #undef x
1408 	NULL
1409 };
1410 
1411 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1412 {
1413 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
1414 	union journal_res_state s;
1415 	unsigned long now = jiffies;
1416 	u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1417 
1418 	if (!out->nr_tabstops)
1419 		printbuf_tabstop_push(out, 28);
1420 	out->atomic++;
1421 
1422 	rcu_read_lock();
1423 	s = READ_ONCE(j->reservations);
1424 
1425 	prt_printf(out, "flags:\t");
1426 	prt_bitflags(out, bch2_journal_flags_strs, j->flags);
1427 	prt_newline(out);
1428 	prt_printf(out, "dirty journal entries:\t%llu/%llu\n",	fifo_used(&j->pin), j->pin.size);
1429 	prt_printf(out, "seq:\t%llu\n",				journal_cur_seq(j));
1430 	prt_printf(out, "seq_ondisk:\t%llu\n",			j->seq_ondisk);
1431 	prt_printf(out, "last_seq:\t%llu\n",			journal_last_seq(j));
1432 	prt_printf(out, "last_seq_ondisk:\t%llu\n",		j->last_seq_ondisk);
1433 	prt_printf(out, "flushed_seq_ondisk:\t%llu\n",		j->flushed_seq_ondisk);
1434 	prt_printf(out, "watermark:\t%s\n",			bch2_watermarks[j->watermark]);
1435 	prt_printf(out, "each entry reserved:\t%u\n",		j->entry_u64s_reserved);
1436 	prt_printf(out, "nr flush writes:\t%llu\n",		j->nr_flush_writes);
1437 	prt_printf(out, "nr noflush writes:\t%llu\n",		j->nr_noflush_writes);
1438 	prt_printf(out, "average write size:\t");
1439 	prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1440 	prt_newline(out);
1441 	prt_printf(out, "nr direct reclaim:\t%llu\n",		j->nr_direct_reclaim);
1442 	prt_printf(out, "nr background reclaim:\t%llu\n",	j->nr_background_reclaim);
1443 	prt_printf(out, "reclaim kicked:\t%u\n",		j->reclaim_kicked);
1444 	prt_printf(out, "reclaim runs in:\t%u ms\n",		time_after(j->next_reclaim, now)
1445 	       ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1446 	prt_printf(out, "blocked:\t%u\n",			j->blocked);
1447 	prt_printf(out, "current entry sectors:\t%u\n",		j->cur_entry_sectors);
1448 	prt_printf(out, "current entry error:\t%s\n",		bch2_journal_errors[j->cur_entry_error]);
1449 	prt_printf(out, "current entry:\t");
1450 
1451 	switch (s.cur_entry_offset) {
1452 	case JOURNAL_ENTRY_ERROR_VAL:
1453 		prt_printf(out, "error\n");
1454 		break;
1455 	case JOURNAL_ENTRY_CLOSED_VAL:
1456 		prt_printf(out, "closed\n");
1457 		break;
1458 	default:
1459 		prt_printf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s);
1460 		break;
1461 	}
1462 
1463 	prt_printf(out, "unwritten entries:\n");
1464 	bch2_journal_bufs_to_text(out, j);
1465 
1466 	prt_printf(out, "space:\n");
1467 	printbuf_indent_add(out, 2);
1468 	prt_printf(out, "discarded\t%u:%u\n",
1469 	       j->space[journal_space_discarded].next_entry,
1470 	       j->space[journal_space_discarded].total);
1471 	prt_printf(out, "clean ondisk\t%u:%u\n",
1472 	       j->space[journal_space_clean_ondisk].next_entry,
1473 	       j->space[journal_space_clean_ondisk].total);
1474 	prt_printf(out, "clean\t%u:%u\n",
1475 	       j->space[journal_space_clean].next_entry,
1476 	       j->space[journal_space_clean].total);
1477 	prt_printf(out, "total\t%u:%u\n",
1478 	       j->space[journal_space_total].next_entry,
1479 	       j->space[journal_space_total].total);
1480 	printbuf_indent_sub(out, 2);
1481 
1482 	for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1483 		struct journal_device *ja = &ca->journal;
1484 
1485 		if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1486 			continue;
1487 
1488 		if (!ja->nr)
1489 			continue;
1490 
1491 		prt_printf(out, "dev %u:\n",			ca->dev_idx);
1492 		printbuf_indent_add(out, 2);
1493 		prt_printf(out, "nr\t%u\n",			ja->nr);
1494 		prt_printf(out, "bucket size\t%u\n",		ca->mi.bucket_size);
1495 		prt_printf(out, "available\t%u:%u\n",		bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1496 		prt_printf(out, "discard_idx\t%u\n",		ja->discard_idx);
1497 		prt_printf(out, "dirty_ondisk\t%u (seq %llu)\n",ja->dirty_idx_ondisk,	ja->bucket_seq[ja->dirty_idx_ondisk]);
1498 		prt_printf(out, "dirty_idx\t%u (seq %llu)\n",	ja->dirty_idx,		ja->bucket_seq[ja->dirty_idx]);
1499 		prt_printf(out, "cur_idx\t%u (seq %llu)\n",	ja->cur_idx,		ja->bucket_seq[ja->cur_idx]);
1500 		printbuf_indent_sub(out, 2);
1501 	}
1502 
1503 	rcu_read_unlock();
1504 
1505 	--out->atomic;
1506 }
1507 
1508 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1509 {
1510 	spin_lock(&j->lock);
1511 	__bch2_journal_debug_to_text(out, j);
1512 	spin_unlock(&j->lock);
1513 }
1514 
1515 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1516 {
1517 	struct journal_entry_pin_list *pin_list;
1518 	struct journal_entry_pin *pin;
1519 
1520 	spin_lock(&j->lock);
1521 	*seq = max(*seq, j->pin.front);
1522 
1523 	if (*seq >= j->pin.back) {
1524 		spin_unlock(&j->lock);
1525 		return true;
1526 	}
1527 
1528 	out->atomic++;
1529 
1530 	pin_list = journal_seq_pin(j, *seq);
1531 
1532 	prt_printf(out, "%llu: count %u\n", *seq, atomic_read(&pin_list->count));
1533 	printbuf_indent_add(out, 2);
1534 
1535 	for (unsigned i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1536 		list_for_each_entry(pin, &pin_list->list[i], list)
1537 			prt_printf(out, "\t%px %ps\n", pin, pin->flush);
1538 
1539 	if (!list_empty(&pin_list->flushed))
1540 		prt_printf(out, "flushed:\n");
1541 
1542 	list_for_each_entry(pin, &pin_list->flushed, list)
1543 		prt_printf(out, "\t%px %ps\n", pin, pin->flush);
1544 
1545 	printbuf_indent_sub(out, 2);
1546 
1547 	--out->atomic;
1548 	spin_unlock(&j->lock);
1549 
1550 	return false;
1551 }
1552 
1553 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1554 {
1555 	u64 seq = 0;
1556 
1557 	while (!bch2_journal_seq_pins_to_text(out, j, &seq))
1558 		seq++;
1559 }
1560