xref: /linux/fs/bcachefs/journal.c (revision 55d0969c451159cff86949b38c39171cab962069)
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 	unsigned sectors;
234 
235 	BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
236 	       closed_val != JOURNAL_ENTRY_ERROR_VAL);
237 
238 	lockdep_assert_held(&j->lock);
239 
240 	old.v = atomic64_read(&j->reservations.counter);
241 	do {
242 		new.v = old.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 (!atomic64_try_cmpxchg(&j->reservations.counter,
249 				       &old.v, new.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 
357 	lockdep_assert_held(&j->lock);
358 	BUG_ON(journal_entry_is_open(j));
359 	BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
360 
361 	if (j->blocked)
362 		return JOURNAL_ERR_blocked;
363 
364 	if (j->cur_entry_error)
365 		return j->cur_entry_error;
366 
367 	if (bch2_journal_error(j))
368 		return JOURNAL_ERR_insufficient_devices; /* -EROFS */
369 
370 	if (!fifo_free(&j->pin))
371 		return JOURNAL_ERR_journal_pin_full;
372 
373 	if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
374 		return JOURNAL_ERR_max_in_flight;
375 
376 	BUG_ON(!j->cur_entry_sectors);
377 
378 	buf->expires		=
379 		(journal_cur_seq(j) == j->flushed_seq_ondisk
380 		 ? jiffies
381 		 : j->last_flush_write) +
382 		msecs_to_jiffies(c->opts.journal_flush_delay);
383 
384 	buf->u64s_reserved	= j->entry_u64s_reserved;
385 	buf->disk_sectors	= j->cur_entry_sectors;
386 	buf->sectors		= min(buf->disk_sectors, buf->buf_size >> 9);
387 
388 	u64s = (int) (buf->sectors << 9) / sizeof(u64) -
389 		journal_entry_overhead(j);
390 	u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
391 
392 	if (u64s <= (ssize_t) j->early_journal_entries.nr)
393 		return JOURNAL_ERR_journal_full;
394 
395 	if (fifo_empty(&j->pin) && j->reclaim_thread)
396 		wake_up_process(j->reclaim_thread);
397 
398 	/*
399 	 * The fifo_push() needs to happen at the same time as j->seq is
400 	 * incremented for journal_last_seq() to be calculated correctly
401 	 */
402 	atomic64_inc(&j->seq);
403 	journal_pin_list_init(fifo_push_ref(&j->pin), 1);
404 
405 	BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
406 
407 	BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
408 
409 	bkey_extent_init(&buf->key);
410 	buf->noflush		= false;
411 	buf->must_flush		= false;
412 	buf->separate_flush	= false;
413 	buf->flush_time		= 0;
414 	buf->need_flush_to_write_buffer = true;
415 	buf->write_started	= false;
416 	buf->write_allocated	= false;
417 	buf->write_done		= false;
418 
419 	memset(buf->data, 0, sizeof(*buf->data));
420 	buf->data->seq	= cpu_to_le64(journal_cur_seq(j));
421 	buf->data->u64s	= 0;
422 
423 	if (j->early_journal_entries.nr) {
424 		memcpy(buf->data->_data, j->early_journal_entries.data,
425 		       j->early_journal_entries.nr * sizeof(u64));
426 		le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
427 	}
428 
429 	/*
430 	 * Must be set before marking the journal entry as open:
431 	 */
432 	j->cur_entry_u64s = u64s;
433 
434 	old.v = atomic64_read(&j->reservations.counter);
435 	do {
436 		new.v = old.v;
437 
438 		BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
439 
440 		new.idx++;
441 		BUG_ON(journal_state_count(new, new.idx));
442 		BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
443 
444 		journal_state_inc(&new);
445 
446 		/* Handle any already added entries */
447 		new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
448 	} while (!atomic64_try_cmpxchg(&j->reservations.counter,
449 				       &old.v, new.v));
450 
451 	if (nr_unwritten_journal_entries(j) == 1)
452 		mod_delayed_work(j->wq,
453 				 &j->write_work,
454 				 msecs_to_jiffies(c->opts.journal_flush_delay));
455 	journal_wake(j);
456 
457 	if (j->early_journal_entries.nr)
458 		darray_exit(&j->early_journal_entries);
459 	return 0;
460 }
461 
462 static bool journal_quiesced(struct journal *j)
463 {
464 	bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
465 
466 	if (!ret)
467 		bch2_journal_entry_close(j);
468 	return ret;
469 }
470 
471 static void journal_quiesce(struct journal *j)
472 {
473 	wait_event(j->wait, journal_quiesced(j));
474 }
475 
476 static void journal_write_work(struct work_struct *work)
477 {
478 	struct journal *j = container_of(work, struct journal, write_work.work);
479 
480 	spin_lock(&j->lock);
481 	if (__journal_entry_is_open(j->reservations)) {
482 		long delta = journal_cur_buf(j)->expires - jiffies;
483 
484 		if (delta > 0)
485 			mod_delayed_work(j->wq, &j->write_work, delta);
486 		else
487 			__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
488 	}
489 	spin_unlock(&j->lock);
490 }
491 
492 static int __journal_res_get(struct journal *j, struct journal_res *res,
493 			     unsigned flags)
494 {
495 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
496 	struct journal_buf *buf;
497 	bool can_discard;
498 	int ret;
499 retry:
500 	if (journal_res_get_fast(j, res, flags))
501 		return 0;
502 
503 	if (bch2_journal_error(j))
504 		return -BCH_ERR_erofs_journal_err;
505 
506 	if (j->blocked)
507 		return -BCH_ERR_journal_res_get_blocked;
508 
509 	if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
510 		ret = JOURNAL_ERR_journal_full;
511 		can_discard = j->can_discard;
512 		goto out;
513 	}
514 
515 	if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) {
516 		ret = JOURNAL_ERR_max_in_flight;
517 		goto out;
518 	}
519 
520 	spin_lock(&j->lock);
521 
522 	/*
523 	 * Recheck after taking the lock, so we don't race with another thread
524 	 * that just did journal_entry_open() and call bch2_journal_entry_close()
525 	 * unnecessarily
526 	 */
527 	if (journal_res_get_fast(j, res, flags)) {
528 		ret = 0;
529 		goto unlock;
530 	}
531 
532 	/*
533 	 * If we couldn't get a reservation because the current buf filled up,
534 	 * and we had room for a bigger entry on disk, signal that we want to
535 	 * realloc the journal bufs:
536 	 */
537 	buf = journal_cur_buf(j);
538 	if (journal_entry_is_open(j) &&
539 	    buf->buf_size >> 9 < buf->disk_sectors &&
540 	    buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
541 		j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
542 
543 	__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
544 	ret = journal_entry_open(j) ?: JOURNAL_ERR_retry;
545 unlock:
546 	can_discard = j->can_discard;
547 	spin_unlock(&j->lock);
548 out:
549 	if (ret == JOURNAL_ERR_retry)
550 		goto retry;
551 	if (!ret)
552 		return 0;
553 
554 	if (journal_error_check_stuck(j, ret, flags))
555 		ret = -BCH_ERR_journal_res_get_blocked;
556 
557 	if (ret == JOURNAL_ERR_max_in_flight &&
558 	    track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true)) {
559 
560 		struct printbuf buf = PRINTBUF;
561 		prt_printf(&buf, "seq %llu\n", journal_cur_seq(j));
562 		bch2_journal_bufs_to_text(&buf, j);
563 		trace_journal_entry_full(c, buf.buf);
564 		printbuf_exit(&buf);
565 		count_event(c, journal_entry_full);
566 	}
567 
568 	/*
569 	 * Journal is full - can't rely on reclaim from work item due to
570 	 * freezing:
571 	 */
572 	if ((ret == JOURNAL_ERR_journal_full ||
573 	     ret == JOURNAL_ERR_journal_pin_full) &&
574 	    !(flags & JOURNAL_RES_GET_NONBLOCK)) {
575 		if (can_discard) {
576 			bch2_journal_do_discards(j);
577 			goto retry;
578 		}
579 
580 		if (mutex_trylock(&j->reclaim_lock)) {
581 			bch2_journal_reclaim(j);
582 			mutex_unlock(&j->reclaim_lock);
583 		}
584 	}
585 
586 	return ret == JOURNAL_ERR_insufficient_devices
587 		? -BCH_ERR_erofs_journal_err
588 		: -BCH_ERR_journal_res_get_blocked;
589 }
590 
591 /*
592  * Essentially the entry function to the journaling code. When bcachefs is doing
593  * a btree insert, it calls this function to get the current journal write.
594  * Journal write is the structure used set up journal writes. The calling
595  * function will then add its keys to the structure, queuing them for the next
596  * write.
597  *
598  * To ensure forward progress, the current task must not be holding any
599  * btree node write locks.
600  */
601 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
602 				  unsigned flags)
603 {
604 	int ret;
605 
606 	if (closure_wait_event_timeout(&j->async_wait,
607 		   (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
608 		   (flags & JOURNAL_RES_GET_NONBLOCK),
609 		   HZ * 10))
610 		return ret;
611 
612 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
613 	struct printbuf buf = PRINTBUF;
614 	bch2_journal_debug_to_text(&buf, j);
615 	bch_err(c, "Journal stuck? Waited for 10 seconds...\n%s",
616 		buf.buf);
617 	printbuf_exit(&buf);
618 
619 	closure_wait_event(&j->async_wait,
620 		   (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
621 		   (flags & JOURNAL_RES_GET_NONBLOCK));
622 	return ret;
623 }
624 
625 /* journal_entry_res: */
626 
627 void bch2_journal_entry_res_resize(struct journal *j,
628 				   struct journal_entry_res *res,
629 				   unsigned new_u64s)
630 {
631 	union journal_res_state state;
632 	int d = new_u64s - res->u64s;
633 
634 	spin_lock(&j->lock);
635 
636 	j->entry_u64s_reserved += d;
637 	if (d <= 0)
638 		goto out;
639 
640 	j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
641 	smp_mb();
642 	state = READ_ONCE(j->reservations);
643 
644 	if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
645 	    state.cur_entry_offset > j->cur_entry_u64s) {
646 		j->cur_entry_u64s += d;
647 		/*
648 		 * Not enough room in current journal entry, have to flush it:
649 		 */
650 		__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
651 	} else {
652 		journal_cur_buf(j)->u64s_reserved += d;
653 	}
654 out:
655 	spin_unlock(&j->lock);
656 	res->u64s += d;
657 }
658 
659 /* journal flushing: */
660 
661 /**
662  * bch2_journal_flush_seq_async - wait for a journal entry to be written
663  * @j:		journal object
664  * @seq:	seq to flush
665  * @parent:	closure object to wait with
666  * Returns:	1 if @seq has already been flushed, 0 if @seq is being flushed,
667  *		-EIO if @seq will never be flushed
668  *
669  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
670  * necessary
671  */
672 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
673 				 struct closure *parent)
674 {
675 	struct journal_buf *buf;
676 	int ret = 0;
677 
678 	if (seq <= j->flushed_seq_ondisk)
679 		return 1;
680 
681 	spin_lock(&j->lock);
682 
683 	if (WARN_ONCE(seq > journal_cur_seq(j),
684 		      "requested to flush journal seq %llu, but currently at %llu",
685 		      seq, journal_cur_seq(j)))
686 		goto out;
687 
688 	/* Recheck under lock: */
689 	if (j->err_seq && seq >= j->err_seq) {
690 		ret = -EIO;
691 		goto out;
692 	}
693 
694 	if (seq <= j->flushed_seq_ondisk) {
695 		ret = 1;
696 		goto out;
697 	}
698 
699 	/* if seq was written, but not flushed - flush a newer one instead */
700 	seq = max(seq, journal_last_unwritten_seq(j));
701 
702 recheck_need_open:
703 	if (seq > journal_cur_seq(j)) {
704 		struct journal_res res = { 0 };
705 
706 		if (journal_entry_is_open(j))
707 			__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
708 
709 		spin_unlock(&j->lock);
710 
711 		/*
712 		 * We're called from bch2_journal_flush_seq() -> wait_event();
713 		 * but this might block. We won't usually block, so we won't
714 		 * livelock:
715 		 */
716 		sched_annotate_sleep();
717 		ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
718 		if (ret)
719 			return ret;
720 
721 		seq = res.seq;
722 		buf = journal_seq_to_buf(j, seq);
723 		buf->must_flush = true;
724 
725 		if (!buf->flush_time) {
726 			buf->flush_time	= local_clock() ?: 1;
727 			buf->expires = jiffies;
728 		}
729 
730 		if (parent && !closure_wait(&buf->wait, parent))
731 			BUG();
732 
733 		bch2_journal_res_put(j, &res);
734 
735 		spin_lock(&j->lock);
736 		goto want_write;
737 	}
738 
739 	/*
740 	 * if write was kicked off without a flush, or if we promised it
741 	 * wouldn't be a flush, flush the next sequence number instead
742 	 */
743 	buf = journal_seq_to_buf(j, seq);
744 	if (buf->noflush) {
745 		seq++;
746 		goto recheck_need_open;
747 	}
748 
749 	buf->must_flush = true;
750 
751 	if (parent && !closure_wait(&buf->wait, parent))
752 		BUG();
753 want_write:
754 	if (seq == journal_cur_seq(j))
755 		journal_entry_want_write(j);
756 out:
757 	spin_unlock(&j->lock);
758 	return ret;
759 }
760 
761 int bch2_journal_flush_seq(struct journal *j, u64 seq)
762 {
763 	u64 start_time = local_clock();
764 	int ret, ret2;
765 
766 	/*
767 	 * Don't update time_stats when @seq is already flushed:
768 	 */
769 	if (seq <= j->flushed_seq_ondisk)
770 		return 0;
771 
772 	ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
773 
774 	if (!ret)
775 		bch2_time_stats_update(j->flush_seq_time, start_time);
776 
777 	return ret ?: ret2 < 0 ? ret2 : 0;
778 }
779 
780 /*
781  * bch2_journal_flush_async - if there is an open journal entry, or a journal
782  * still being written, write it and wait for the write to complete
783  */
784 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
785 {
786 	bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
787 }
788 
789 int bch2_journal_flush(struct journal *j)
790 {
791 	return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
792 }
793 
794 /*
795  * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
796  * @seq
797  */
798 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
799 {
800 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
801 	u64 unwritten_seq;
802 	bool ret = false;
803 
804 	if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
805 		return false;
806 
807 	if (seq <= c->journal.flushed_seq_ondisk)
808 		return false;
809 
810 	spin_lock(&j->lock);
811 	if (seq <= c->journal.flushed_seq_ondisk)
812 		goto out;
813 
814 	for (unwritten_seq = journal_last_unwritten_seq(j);
815 	     unwritten_seq < seq;
816 	     unwritten_seq++) {
817 		struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
818 
819 		/* journal flush already in flight, or flush requseted */
820 		if (buf->must_flush)
821 			goto out;
822 
823 		buf->noflush = true;
824 	}
825 
826 	ret = true;
827 out:
828 	spin_unlock(&j->lock);
829 	return ret;
830 }
831 
832 int bch2_journal_meta(struct journal *j)
833 {
834 	struct journal_buf *buf;
835 	struct journal_res res;
836 	int ret;
837 
838 	memset(&res, 0, sizeof(res));
839 
840 	ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
841 	if (ret)
842 		return ret;
843 
844 	buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
845 	buf->must_flush = true;
846 
847 	if (!buf->flush_time) {
848 		buf->flush_time	= local_clock() ?: 1;
849 		buf->expires = jiffies;
850 	}
851 
852 	bch2_journal_res_put(j, &res);
853 
854 	return bch2_journal_flush_seq(j, res.seq);
855 }
856 
857 /* block/unlock the journal: */
858 
859 void bch2_journal_unblock(struct journal *j)
860 {
861 	spin_lock(&j->lock);
862 	j->blocked--;
863 	spin_unlock(&j->lock);
864 
865 	journal_wake(j);
866 }
867 
868 void bch2_journal_block(struct journal *j)
869 {
870 	spin_lock(&j->lock);
871 	j->blocked++;
872 	spin_unlock(&j->lock);
873 
874 	journal_quiesce(j);
875 }
876 
877 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
878 {
879 	struct journal_buf *ret = NULL;
880 
881 	/* We're inside wait_event(), but using mutex_lock(: */
882 	sched_annotate_sleep();
883 	mutex_lock(&j->buf_lock);
884 	spin_lock(&j->lock);
885 	max_seq = min(max_seq, journal_cur_seq(j));
886 
887 	for (u64 seq = journal_last_unwritten_seq(j);
888 	     seq <= max_seq;
889 	     seq++) {
890 		unsigned idx = seq & JOURNAL_BUF_MASK;
891 		struct journal_buf *buf = j->buf + idx;
892 
893 		if (buf->need_flush_to_write_buffer) {
894 			if (seq == journal_cur_seq(j))
895 				__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
896 
897 			union journal_res_state s;
898 			s.v = atomic64_read_acquire(&j->reservations.counter);
899 
900 			ret = journal_state_count(s, idx)
901 				? ERR_PTR(-EAGAIN)
902 				: buf;
903 			break;
904 		}
905 	}
906 
907 	spin_unlock(&j->lock);
908 	if (IS_ERR_OR_NULL(ret))
909 		mutex_unlock(&j->buf_lock);
910 	return ret;
911 }
912 
913 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
914 {
915 	struct journal_buf *ret;
916 
917 	wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
918 	return ret;
919 }
920 
921 /* allocate journal on a device: */
922 
923 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
924 					 bool new_fs, struct closure *cl)
925 {
926 	struct bch_fs *c = ca->fs;
927 	struct journal_device *ja = &ca->journal;
928 	u64 *new_bucket_seq = NULL, *new_buckets = NULL;
929 	struct open_bucket **ob = NULL;
930 	long *bu = NULL;
931 	unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
932 	int ret = 0;
933 
934 	BUG_ON(nr <= ja->nr);
935 
936 	bu		= kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
937 	ob		= kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
938 	new_buckets	= kcalloc(nr, sizeof(u64), GFP_KERNEL);
939 	new_bucket_seq	= kcalloc(nr, sizeof(u64), GFP_KERNEL);
940 	if (!bu || !ob || !new_buckets || !new_bucket_seq) {
941 		ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
942 		goto err_free;
943 	}
944 
945 	for (nr_got = 0; nr_got < nr_want; nr_got++) {
946 		if (new_fs) {
947 			bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
948 			if (bu[nr_got] < 0) {
949 				ret = -BCH_ERR_ENOSPC_bucket_alloc;
950 				break;
951 			}
952 		} else {
953 			ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal,
954 						       BCH_DATA_journal, cl);
955 			ret = PTR_ERR_OR_ZERO(ob[nr_got]);
956 			if (ret)
957 				break;
958 
959 			ret = bch2_trans_run(c,
960 				bch2_trans_mark_metadata_bucket(trans, ca,
961 						ob[nr_got]->bucket, BCH_DATA_journal,
962 						ca->mi.bucket_size, BTREE_TRIGGER_transactional));
963 			if (ret) {
964 				bch2_open_bucket_put(c, ob[nr_got]);
965 				bch_err_msg(c, ret, "marking new journal buckets");
966 				break;
967 			}
968 
969 			bu[nr_got] = ob[nr_got]->bucket;
970 		}
971 	}
972 
973 	if (!nr_got)
974 		goto err_free;
975 
976 	/* Don't return an error if we successfully allocated some buckets: */
977 	ret = 0;
978 
979 	if (c) {
980 		bch2_journal_flush_all_pins(&c->journal);
981 		bch2_journal_block(&c->journal);
982 		mutex_lock(&c->sb_lock);
983 	}
984 
985 	memcpy(new_buckets,	ja->buckets,	ja->nr * sizeof(u64));
986 	memcpy(new_bucket_seq,	ja->bucket_seq,	ja->nr * sizeof(u64));
987 
988 	BUG_ON(ja->discard_idx > ja->nr);
989 
990 	pos = ja->discard_idx ?: ja->nr;
991 
992 	memmove(new_buckets + pos + nr_got,
993 		new_buckets + pos,
994 		sizeof(new_buckets[0]) * (ja->nr - pos));
995 	memmove(new_bucket_seq + pos + nr_got,
996 		new_bucket_seq + pos,
997 		sizeof(new_bucket_seq[0]) * (ja->nr - pos));
998 
999 	for (i = 0; i < nr_got; i++) {
1000 		new_buckets[pos + i] = bu[i];
1001 		new_bucket_seq[pos + i] = 0;
1002 	}
1003 
1004 	nr = ja->nr + nr_got;
1005 
1006 	ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
1007 	if (ret)
1008 		goto err_unblock;
1009 
1010 	if (!new_fs)
1011 		bch2_write_super(c);
1012 
1013 	/* Commit: */
1014 	if (c)
1015 		spin_lock(&c->journal.lock);
1016 
1017 	swap(new_buckets,	ja->buckets);
1018 	swap(new_bucket_seq,	ja->bucket_seq);
1019 	ja->nr = nr;
1020 
1021 	if (pos <= ja->discard_idx)
1022 		ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
1023 	if (pos <= ja->dirty_idx_ondisk)
1024 		ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
1025 	if (pos <= ja->dirty_idx)
1026 		ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
1027 	if (pos <= ja->cur_idx)
1028 		ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
1029 
1030 	if (c)
1031 		spin_unlock(&c->journal.lock);
1032 err_unblock:
1033 	if (c) {
1034 		bch2_journal_unblock(&c->journal);
1035 		mutex_unlock(&c->sb_lock);
1036 	}
1037 
1038 	if (ret && !new_fs)
1039 		for (i = 0; i < nr_got; i++)
1040 			bch2_trans_run(c,
1041 				bch2_trans_mark_metadata_bucket(trans, ca,
1042 						bu[i], BCH_DATA_free, 0,
1043 						BTREE_TRIGGER_transactional));
1044 err_free:
1045 	if (!new_fs)
1046 		for (i = 0; i < nr_got; i++)
1047 			bch2_open_bucket_put(c, ob[i]);
1048 
1049 	kfree(new_bucket_seq);
1050 	kfree(new_buckets);
1051 	kfree(ob);
1052 	kfree(bu);
1053 	return ret;
1054 }
1055 
1056 /*
1057  * Allocate more journal space at runtime - not currently making use if it, but
1058  * the code works:
1059  */
1060 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1061 				unsigned nr)
1062 {
1063 	struct journal_device *ja = &ca->journal;
1064 	struct closure cl;
1065 	int ret = 0;
1066 
1067 	closure_init_stack(&cl);
1068 
1069 	down_write(&c->state_lock);
1070 
1071 	/* don't handle reducing nr of buckets yet: */
1072 	if (nr < ja->nr)
1073 		goto unlock;
1074 
1075 	while (ja->nr < nr) {
1076 		struct disk_reservation disk_res = { 0, 0, 0 };
1077 
1078 		/*
1079 		 * note: journal buckets aren't really counted as _sectors_ used yet, so
1080 		 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1081 		 * when space used goes up without a reservation - but we do need the
1082 		 * reservation to ensure we'll actually be able to allocate:
1083 		 *
1084 		 * XXX: that's not right, disk reservations only ensure a
1085 		 * filesystem-wide allocation will succeed, this is a device
1086 		 * specific allocation - we can hang here:
1087 		 */
1088 
1089 		ret = bch2_disk_reservation_get(c, &disk_res,
1090 						bucket_to_sector(ca, nr - ja->nr), 1, 0);
1091 		if (ret)
1092 			break;
1093 
1094 		ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1095 
1096 		bch2_disk_reservation_put(c, &disk_res);
1097 
1098 		closure_sync(&cl);
1099 
1100 		if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1101 			break;
1102 	}
1103 
1104 	bch_err_fn(c, ret);
1105 unlock:
1106 	up_write(&c->state_lock);
1107 	return ret;
1108 }
1109 
1110 int bch2_dev_journal_alloc(struct bch_dev *ca, bool new_fs)
1111 {
1112 	unsigned nr;
1113 	int ret;
1114 
1115 	if (dynamic_fault("bcachefs:add:journal_alloc")) {
1116 		ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1117 		goto err;
1118 	}
1119 
1120 	/* 1/128th of the device by default: */
1121 	nr = ca->mi.nbuckets >> 7;
1122 
1123 	/*
1124 	 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1125 	 * is smaller:
1126 	 */
1127 	nr = clamp_t(unsigned, nr,
1128 		     BCH_JOURNAL_BUCKETS_MIN,
1129 		     min(1 << 13,
1130 			 (1 << 24) / ca->mi.bucket_size));
1131 
1132 	ret = __bch2_set_nr_journal_buckets(ca, nr, new_fs, NULL);
1133 err:
1134 	bch_err_fn(ca, ret);
1135 	return ret;
1136 }
1137 
1138 int bch2_fs_journal_alloc(struct bch_fs *c)
1139 {
1140 	for_each_online_member(c, ca) {
1141 		if (ca->journal.nr)
1142 			continue;
1143 
1144 		int ret = bch2_dev_journal_alloc(ca, true);
1145 		if (ret) {
1146 			percpu_ref_put(&ca->io_ref);
1147 			return ret;
1148 		}
1149 	}
1150 
1151 	return 0;
1152 }
1153 
1154 /* startup/shutdown: */
1155 
1156 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1157 {
1158 	bool ret = false;
1159 	u64 seq;
1160 
1161 	spin_lock(&j->lock);
1162 	for (seq = journal_last_unwritten_seq(j);
1163 	     seq <= journal_cur_seq(j) && !ret;
1164 	     seq++) {
1165 		struct journal_buf *buf = journal_seq_to_buf(j, seq);
1166 
1167 		if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1168 			ret = true;
1169 	}
1170 	spin_unlock(&j->lock);
1171 
1172 	return ret;
1173 }
1174 
1175 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1176 {
1177 	wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1178 }
1179 
1180 void bch2_fs_journal_stop(struct journal *j)
1181 {
1182 	if (!test_bit(JOURNAL_running, &j->flags))
1183 		return;
1184 
1185 	bch2_journal_reclaim_stop(j);
1186 	bch2_journal_flush_all_pins(j);
1187 
1188 	wait_event(j->wait, bch2_journal_entry_close(j));
1189 
1190 	/*
1191 	 * Always write a new journal entry, to make sure the clock hands are up
1192 	 * to date (and match the superblock)
1193 	 */
1194 	bch2_journal_meta(j);
1195 
1196 	journal_quiesce(j);
1197 	cancel_delayed_work_sync(&j->write_work);
1198 
1199 	WARN(!bch2_journal_error(j) &&
1200 	     test_bit(JOURNAL_replay_done, &j->flags) &&
1201 	     j->last_empty_seq != journal_cur_seq(j),
1202 	     "journal shutdown error: cur seq %llu but last empty seq %llu",
1203 	     journal_cur_seq(j), j->last_empty_seq);
1204 
1205 	if (!bch2_journal_error(j))
1206 		clear_bit(JOURNAL_running, &j->flags);
1207 }
1208 
1209 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1210 {
1211 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
1212 	struct journal_entry_pin_list *p;
1213 	struct journal_replay *i, **_i;
1214 	struct genradix_iter iter;
1215 	bool had_entries = false;
1216 	u64 last_seq = cur_seq, nr, seq;
1217 
1218 	genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1219 		i = *_i;
1220 
1221 		if (journal_replay_ignore(i))
1222 			continue;
1223 
1224 		last_seq = le64_to_cpu(i->j.last_seq);
1225 		break;
1226 	}
1227 
1228 	nr = cur_seq - last_seq;
1229 
1230 	if (nr + 1 > j->pin.size) {
1231 		free_fifo(&j->pin);
1232 		init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1233 		if (!j->pin.data) {
1234 			bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1235 			return -BCH_ERR_ENOMEM_journal_pin_fifo;
1236 		}
1237 	}
1238 
1239 	j->replay_journal_seq	= last_seq;
1240 	j->replay_journal_seq_end = cur_seq;
1241 	j->last_seq_ondisk	= last_seq;
1242 	j->flushed_seq_ondisk	= cur_seq - 1;
1243 	j->seq_ondisk		= cur_seq - 1;
1244 	j->pin.front		= last_seq;
1245 	j->pin.back		= cur_seq;
1246 	atomic64_set(&j->seq, cur_seq - 1);
1247 
1248 	fifo_for_each_entry_ptr(p, &j->pin, seq)
1249 		journal_pin_list_init(p, 1);
1250 
1251 	genradix_for_each(&c->journal_entries, iter, _i) {
1252 		i = *_i;
1253 
1254 		if (journal_replay_ignore(i))
1255 			continue;
1256 
1257 		seq = le64_to_cpu(i->j.seq);
1258 		BUG_ON(seq >= cur_seq);
1259 
1260 		if (seq < last_seq)
1261 			continue;
1262 
1263 		if (journal_entry_empty(&i->j))
1264 			j->last_empty_seq = le64_to_cpu(i->j.seq);
1265 
1266 		p = journal_seq_pin(j, seq);
1267 
1268 		p->devs.nr = 0;
1269 		darray_for_each(i->ptrs, ptr)
1270 			bch2_dev_list_add_dev(&p->devs, ptr->dev);
1271 
1272 		had_entries = true;
1273 	}
1274 
1275 	if (!had_entries)
1276 		j->last_empty_seq = cur_seq - 1; /* to match j->seq */
1277 
1278 	spin_lock(&j->lock);
1279 
1280 	set_bit(JOURNAL_running, &j->flags);
1281 	j->last_flush_write = jiffies;
1282 
1283 	j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1284 	j->reservations.unwritten_idx++;
1285 
1286 	c->last_bucket_seq_cleanup = journal_cur_seq(j);
1287 
1288 	bch2_journal_space_available(j);
1289 	spin_unlock(&j->lock);
1290 
1291 	return bch2_journal_reclaim_start(j);
1292 }
1293 
1294 /* init/exit: */
1295 
1296 void bch2_dev_journal_exit(struct bch_dev *ca)
1297 {
1298 	struct journal_device *ja = &ca->journal;
1299 
1300 	for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1301 		kfree(ja->bio[i]);
1302 		ja->bio[i] = NULL;
1303 	}
1304 
1305 	kfree(ja->buckets);
1306 	kfree(ja->bucket_seq);
1307 	ja->buckets	= NULL;
1308 	ja->bucket_seq	= NULL;
1309 }
1310 
1311 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1312 {
1313 	struct journal_device *ja = &ca->journal;
1314 	struct bch_sb_field_journal *journal_buckets =
1315 		bch2_sb_field_get(sb, journal);
1316 	struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1317 		bch2_sb_field_get(sb, journal_v2);
1318 
1319 	ja->nr = 0;
1320 
1321 	if (journal_buckets_v2) {
1322 		unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1323 
1324 		for (unsigned i = 0; i < nr; i++)
1325 			ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1326 	} else if (journal_buckets) {
1327 		ja->nr = bch2_nr_journal_buckets(journal_buckets);
1328 	}
1329 
1330 	ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1331 	if (!ja->bucket_seq)
1332 		return -BCH_ERR_ENOMEM_dev_journal_init;
1333 
1334 	unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1335 
1336 	for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1337 		ja->bio[i] = kmalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
1338 				     nr_bvecs), GFP_KERNEL);
1339 		if (!ja->bio[i])
1340 			return -BCH_ERR_ENOMEM_dev_journal_init;
1341 
1342 		ja->bio[i]->ca = ca;
1343 		ja->bio[i]->buf_idx = i;
1344 		bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0);
1345 	}
1346 
1347 	ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1348 	if (!ja->buckets)
1349 		return -BCH_ERR_ENOMEM_dev_journal_init;
1350 
1351 	if (journal_buckets_v2) {
1352 		unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1353 		unsigned dst = 0;
1354 
1355 		for (unsigned i = 0; i < nr; i++)
1356 			for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1357 				ja->buckets[dst++] =
1358 					le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1359 	} else if (journal_buckets) {
1360 		for (unsigned i = 0; i < ja->nr; i++)
1361 			ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1362 	}
1363 
1364 	return 0;
1365 }
1366 
1367 void bch2_fs_journal_exit(struct journal *j)
1368 {
1369 	if (j->wq)
1370 		destroy_workqueue(j->wq);
1371 
1372 	darray_exit(&j->early_journal_entries);
1373 
1374 	for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++)
1375 		kvfree(j->buf[i].data);
1376 	free_fifo(&j->pin);
1377 }
1378 
1379 int bch2_fs_journal_init(struct journal *j)
1380 {
1381 	static struct lock_class_key res_key;
1382 
1383 	mutex_init(&j->buf_lock);
1384 	spin_lock_init(&j->lock);
1385 	spin_lock_init(&j->err_lock);
1386 	init_waitqueue_head(&j->wait);
1387 	INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1388 	init_waitqueue_head(&j->reclaim_wait);
1389 	init_waitqueue_head(&j->pin_flush_wait);
1390 	mutex_init(&j->reclaim_lock);
1391 	mutex_init(&j->discard_lock);
1392 
1393 	lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1394 
1395 	atomic64_set(&j->reservations.counter,
1396 		((union journal_res_state)
1397 		 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1398 
1399 	if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1400 		return -BCH_ERR_ENOMEM_journal_pin_fifo;
1401 
1402 	for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) {
1403 		j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1404 		j->buf[i].data = kvmalloc(j->buf[i].buf_size, GFP_KERNEL);
1405 		if (!j->buf[i].data)
1406 			return -BCH_ERR_ENOMEM_journal_buf;
1407 		j->buf[i].idx = i;
1408 	}
1409 
1410 	j->pin.front = j->pin.back = 1;
1411 
1412 	j->wq = alloc_workqueue("bcachefs_journal",
1413 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512);
1414 	if (!j->wq)
1415 		return -BCH_ERR_ENOMEM_fs_other_alloc;
1416 	return 0;
1417 }
1418 
1419 /* debug: */
1420 
1421 static const char * const bch2_journal_flags_strs[] = {
1422 #define x(n)	#n,
1423 	JOURNAL_FLAGS()
1424 #undef x
1425 	NULL
1426 };
1427 
1428 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1429 {
1430 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
1431 	union journal_res_state s;
1432 	unsigned long now = jiffies;
1433 	u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1434 
1435 	printbuf_tabstops_reset(out);
1436 	printbuf_tabstop_push(out, 28);
1437 	out->atomic++;
1438 
1439 	rcu_read_lock();
1440 	s = READ_ONCE(j->reservations);
1441 
1442 	prt_printf(out, "flags:\t");
1443 	prt_bitflags(out, bch2_journal_flags_strs, j->flags);
1444 	prt_newline(out);
1445 	prt_printf(out, "dirty journal entries:\t%llu/%llu\n",	fifo_used(&j->pin), j->pin.size);
1446 	prt_printf(out, "seq:\t%llu\n",				journal_cur_seq(j));
1447 	prt_printf(out, "seq_ondisk:\t%llu\n",			j->seq_ondisk);
1448 	prt_printf(out, "last_seq:\t%llu\n",			journal_last_seq(j));
1449 	prt_printf(out, "last_seq_ondisk:\t%llu\n",		j->last_seq_ondisk);
1450 	prt_printf(out, "flushed_seq_ondisk:\t%llu\n",		j->flushed_seq_ondisk);
1451 	prt_printf(out, "watermark:\t%s\n",			bch2_watermarks[j->watermark]);
1452 	prt_printf(out, "each entry reserved:\t%u\n",		j->entry_u64s_reserved);
1453 	prt_printf(out, "nr flush writes:\t%llu\n",		j->nr_flush_writes);
1454 	prt_printf(out, "nr noflush writes:\t%llu\n",		j->nr_noflush_writes);
1455 	prt_printf(out, "average write size:\t");
1456 	prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1457 	prt_newline(out);
1458 	prt_printf(out, "nr direct reclaim:\t%llu\n",		j->nr_direct_reclaim);
1459 	prt_printf(out, "nr background reclaim:\t%llu\n",	j->nr_background_reclaim);
1460 	prt_printf(out, "reclaim kicked:\t%u\n",		j->reclaim_kicked);
1461 	prt_printf(out, "reclaim runs in:\t%u ms\n",		time_after(j->next_reclaim, now)
1462 	       ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1463 	prt_printf(out, "blocked:\t%u\n",			j->blocked);
1464 	prt_printf(out, "current entry sectors:\t%u\n",		j->cur_entry_sectors);
1465 	prt_printf(out, "current entry error:\t%s\n",		bch2_journal_errors[j->cur_entry_error]);
1466 	prt_printf(out, "current entry:\t");
1467 
1468 	switch (s.cur_entry_offset) {
1469 	case JOURNAL_ENTRY_ERROR_VAL:
1470 		prt_printf(out, "error\n");
1471 		break;
1472 	case JOURNAL_ENTRY_CLOSED_VAL:
1473 		prt_printf(out, "closed\n");
1474 		break;
1475 	default:
1476 		prt_printf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s);
1477 		break;
1478 	}
1479 
1480 	prt_printf(out, "unwritten entries:\n");
1481 	bch2_journal_bufs_to_text(out, j);
1482 
1483 	prt_printf(out, "space:\n");
1484 	printbuf_indent_add(out, 2);
1485 	prt_printf(out, "discarded\t%u:%u\n",
1486 	       j->space[journal_space_discarded].next_entry,
1487 	       j->space[journal_space_discarded].total);
1488 	prt_printf(out, "clean ondisk\t%u:%u\n",
1489 	       j->space[journal_space_clean_ondisk].next_entry,
1490 	       j->space[journal_space_clean_ondisk].total);
1491 	prt_printf(out, "clean\t%u:%u\n",
1492 	       j->space[journal_space_clean].next_entry,
1493 	       j->space[journal_space_clean].total);
1494 	prt_printf(out, "total\t%u:%u\n",
1495 	       j->space[journal_space_total].next_entry,
1496 	       j->space[journal_space_total].total);
1497 	printbuf_indent_sub(out, 2);
1498 
1499 	for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1500 		struct journal_device *ja = &ca->journal;
1501 
1502 		if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1503 			continue;
1504 
1505 		if (!ja->nr)
1506 			continue;
1507 
1508 		prt_printf(out, "dev %u:\n",			ca->dev_idx);
1509 		printbuf_indent_add(out, 2);
1510 		prt_printf(out, "nr\t%u\n",			ja->nr);
1511 		prt_printf(out, "bucket size\t%u\n",		ca->mi.bucket_size);
1512 		prt_printf(out, "available\t%u:%u\n",		bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1513 		prt_printf(out, "discard_idx\t%u\n",		ja->discard_idx);
1514 		prt_printf(out, "dirty_ondisk\t%u (seq %llu)\n",ja->dirty_idx_ondisk,	ja->bucket_seq[ja->dirty_idx_ondisk]);
1515 		prt_printf(out, "dirty_idx\t%u (seq %llu)\n",	ja->dirty_idx,		ja->bucket_seq[ja->dirty_idx]);
1516 		prt_printf(out, "cur_idx\t%u (seq %llu)\n",	ja->cur_idx,		ja->bucket_seq[ja->cur_idx]);
1517 		printbuf_indent_sub(out, 2);
1518 	}
1519 
1520 	rcu_read_unlock();
1521 
1522 	--out->atomic;
1523 }
1524 
1525 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1526 {
1527 	spin_lock(&j->lock);
1528 	__bch2_journal_debug_to_text(out, j);
1529 	spin_unlock(&j->lock);
1530 }
1531 
1532 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1533 {
1534 	struct journal_entry_pin_list *pin_list;
1535 	struct journal_entry_pin *pin;
1536 
1537 	spin_lock(&j->lock);
1538 	if (!test_bit(JOURNAL_running, &j->flags)) {
1539 		spin_unlock(&j->lock);
1540 		return true;
1541 	}
1542 
1543 	*seq = max(*seq, j->pin.front);
1544 
1545 	if (*seq >= j->pin.back) {
1546 		spin_unlock(&j->lock);
1547 		return true;
1548 	}
1549 
1550 	out->atomic++;
1551 
1552 	pin_list = journal_seq_pin(j, *seq);
1553 
1554 	prt_printf(out, "%llu: count %u\n", *seq, atomic_read(&pin_list->count));
1555 	printbuf_indent_add(out, 2);
1556 
1557 	for (unsigned i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1558 		list_for_each_entry(pin, &pin_list->list[i], list)
1559 			prt_printf(out, "\t%px %ps\n", pin, pin->flush);
1560 
1561 	if (!list_empty(&pin_list->flushed))
1562 		prt_printf(out, "flushed:\n");
1563 
1564 	list_for_each_entry(pin, &pin_list->flushed, list)
1565 		prt_printf(out, "\t%px %ps\n", pin, pin->flush);
1566 
1567 	printbuf_indent_sub(out, 2);
1568 
1569 	--out->atomic;
1570 	spin_unlock(&j->lock);
1571 
1572 	return false;
1573 }
1574 
1575 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1576 {
1577 	u64 seq = 0;
1578 
1579 	while (!bch2_journal_seq_pins_to_text(out, j, &seq))
1580 		seq++;
1581 }
1582