xref: /linux/fs/bcachefs/btree_write_buffer.c (revision 4abcd80f23357808b0444d261ed08e5a77dbaa9a)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "bkey_buf.h"
5 #include "btree_locking.h"
6 #include "btree_update.h"
7 #include "btree_update_interior.h"
8 #include "btree_write_buffer.h"
9 #include "disk_accounting.h"
10 #include "error.h"
11 #include "extents.h"
12 #include "journal.h"
13 #include "journal_io.h"
14 #include "journal_reclaim.h"
15 
16 #include <linux/prefetch.h>
17 #include <linux/sort.h>
18 
19 static int bch2_btree_write_buffer_journal_flush(struct journal *,
20 				struct journal_entry_pin *, u64);
21 
22 static int bch2_journal_keys_to_write_buffer(struct bch_fs *, struct journal_buf *);
23 
__wb_key_ref_cmp(const struct wb_key_ref * l,const struct wb_key_ref * r)24 static inline bool __wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
25 {
26 	return (cmp_int(l->hi, r->hi) ?:
27 		cmp_int(l->mi, r->mi) ?:
28 		cmp_int(l->lo, r->lo)) >= 0;
29 }
30 
wb_key_ref_cmp(const struct wb_key_ref * l,const struct wb_key_ref * r)31 static inline bool wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
32 {
33 #ifdef CONFIG_X86_64
34 	int cmp;
35 
36 	asm("mov   (%[l]), %%rax;"
37 	    "sub   (%[r]), %%rax;"
38 	    "mov  8(%[l]), %%rax;"
39 	    "sbb  8(%[r]), %%rax;"
40 	    "mov 16(%[l]), %%rax;"
41 	    "sbb 16(%[r]), %%rax;"
42 	    : "=@ccae" (cmp)
43 	    : [l] "r" (l), [r] "r" (r)
44 	    : "rax", "cc");
45 
46 	EBUG_ON(cmp != __wb_key_ref_cmp(l, r));
47 	return cmp;
48 #else
49 	return __wb_key_ref_cmp(l, r);
50 #endif
51 }
52 
wb_key_seq_cmp(const void * _l,const void * _r)53 static int wb_key_seq_cmp(const void *_l, const void *_r)
54 {
55 	const struct btree_write_buffered_key *l = _l;
56 	const struct btree_write_buffered_key *r = _r;
57 
58 	return cmp_int(l->journal_seq, r->journal_seq);
59 }
60 
61 /* Compare excluding idx, the low 24 bits: */
wb_key_eq(const void * _l,const void * _r)62 static inline bool wb_key_eq(const void *_l, const void *_r)
63 {
64 	const struct wb_key_ref *l = _l;
65 	const struct wb_key_ref *r = _r;
66 
67 	return !((l->hi ^ r->hi)|
68 		 (l->mi ^ r->mi)|
69 		 ((l->lo >> 24) ^ (r->lo >> 24)));
70 }
71 
wb_sort(struct wb_key_ref * base,size_t num)72 static noinline void wb_sort(struct wb_key_ref *base, size_t num)
73 {
74 	size_t n = num, a = num / 2;
75 
76 	if (!a)		/* num < 2 || size == 0 */
77 		return;
78 
79 	for (;;) {
80 		size_t b, c, d;
81 
82 		if (a)			/* Building heap: sift down --a */
83 			--a;
84 		else if (--n)		/* Sorting: Extract root to --n */
85 			swap(base[0], base[n]);
86 		else			/* Sort complete */
87 			break;
88 
89 		/*
90 		 * Sift element at "a" down into heap.  This is the
91 		 * "bottom-up" variant, which significantly reduces
92 		 * calls to cmp_func(): we find the sift-down path all
93 		 * the way to the leaves (one compare per level), then
94 		 * backtrack to find where to insert the target element.
95 		 *
96 		 * Because elements tend to sift down close to the leaves,
97 		 * this uses fewer compares than doing two per level
98 		 * on the way down.  (A bit more than half as many on
99 		 * average, 3/4 worst-case.)
100 		 */
101 		for (b = a; c = 2*b + 1, (d = c + 1) < n;)
102 			b = wb_key_ref_cmp(base + c, base + d) ? c : d;
103 		if (d == n)		/* Special case last leaf with no sibling */
104 			b = c;
105 
106 		/* Now backtrack from "b" to the correct location for "a" */
107 		while (b != a && wb_key_ref_cmp(base + a, base + b))
108 			b = (b - 1) / 2;
109 		c = b;			/* Where "a" belongs */
110 		while (b != a) {	/* Shift it into place */
111 			b = (b - 1) / 2;
112 			swap(base[b], base[c]);
113 		}
114 	}
115 }
116 
wb_flush_one_slowpath(struct btree_trans * trans,struct btree_iter * iter,struct btree_write_buffered_key * wb)117 static noinline int wb_flush_one_slowpath(struct btree_trans *trans,
118 					  struct btree_iter *iter,
119 					  struct btree_write_buffered_key *wb)
120 {
121 	struct btree_path *path = btree_iter_path(trans, iter);
122 
123 	bch2_btree_node_unlock_write(trans, path, path->l[0].b);
124 
125 	trans->journal_res.seq = wb->journal_seq;
126 
127 	return bch2_trans_update(trans, iter, &wb->k,
128 				 BTREE_UPDATE_internal_snapshot_node) ?:
129 		bch2_trans_commit(trans, NULL, NULL,
130 				  BCH_TRANS_COMMIT_no_enospc|
131 				  BCH_TRANS_COMMIT_no_check_rw|
132 				  BCH_TRANS_COMMIT_no_journal_res|
133 				  BCH_TRANS_COMMIT_journal_reclaim);
134 }
135 
wb_flush_one(struct btree_trans * trans,struct btree_iter * iter,struct btree_write_buffered_key * wb,bool * write_locked,bool * accounting_accumulated,size_t * fast)136 static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
137 			       struct btree_write_buffered_key *wb,
138 			       bool *write_locked,
139 			       bool *accounting_accumulated,
140 			       size_t *fast)
141 {
142 	struct btree_path *path;
143 	int ret;
144 
145 	EBUG_ON(!wb->journal_seq);
146 	EBUG_ON(!trans->c->btree_write_buffer.flushing.pin.seq);
147 	EBUG_ON(trans->c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);
148 
149 	ret = bch2_btree_iter_traverse(iter);
150 	if (ret)
151 		return ret;
152 
153 	if (!*accounting_accumulated && wb->k.k.type == KEY_TYPE_accounting) {
154 		struct bkey u;
155 		struct bkey_s_c k = bch2_btree_path_peek_slot_exact(btree_iter_path(trans, iter), &u);
156 
157 		if (k.k->type == KEY_TYPE_accounting)
158 			bch2_accounting_accumulate(bkey_i_to_accounting(&wb->k),
159 						   bkey_s_c_to_accounting(k));
160 	}
161 	*accounting_accumulated = true;
162 
163 	/*
164 	 * We can't clone a path that has write locks: unshare it now, before
165 	 * set_pos and traverse():
166 	 */
167 	if (btree_iter_path(trans, iter)->ref > 1)
168 		iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);
169 
170 	path = btree_iter_path(trans, iter);
171 
172 	if (!*write_locked) {
173 		ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
174 		if (ret)
175 			return ret;
176 
177 		bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
178 		*write_locked = true;
179 	}
180 
181 	if (unlikely(!bch2_btree_node_insert_fits(path->l[0].b, wb->k.k.u64s))) {
182 		*write_locked = false;
183 		return wb_flush_one_slowpath(trans, iter, wb);
184 	}
185 
186 	bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
187 	(*fast)++;
188 	return 0;
189 }
190 
191 /*
192  * Update a btree with a write buffered key using the journal seq of the
193  * original write buffer insert.
194  *
195  * It is not safe to rejournal the key once it has been inserted into the write
196  * buffer because that may break recovery ordering. For example, the key may
197  * have already been modified in the active write buffer in a seq that comes
198  * before the current transaction. If we were to journal this key again and
199  * crash, recovery would process updates in the wrong order.
200  */
201 static int
btree_write_buffered_insert(struct btree_trans * trans,struct btree_write_buffered_key * wb)202 btree_write_buffered_insert(struct btree_trans *trans,
203 			  struct btree_write_buffered_key *wb)
204 {
205 	struct btree_iter iter;
206 	int ret;
207 
208 	bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
209 			     BTREE_ITER_cached|BTREE_ITER_intent);
210 
211 	trans->journal_res.seq = wb->journal_seq;
212 
213 	ret   = bch2_btree_iter_traverse(&iter) ?:
214 		bch2_trans_update(trans, &iter, &wb->k,
215 				  BTREE_UPDATE_internal_snapshot_node);
216 	bch2_trans_iter_exit(trans, &iter);
217 	return ret;
218 }
219 
move_keys_from_inc_to_flushing(struct btree_write_buffer * wb)220 static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
221 {
222 	struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
223 	struct journal *j = &c->journal;
224 
225 	if (!wb->inc.keys.nr)
226 		return;
227 
228 	bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
229 			     bch2_btree_write_buffer_journal_flush);
230 
231 	darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
232 	darray_resize(&wb->sorted, wb->flushing.keys.size);
233 
234 	if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
235 		swap(wb->flushing.keys, wb->inc.keys);
236 		goto out;
237 	}
238 
239 	size_t nr = min(darray_room(wb->flushing.keys),
240 			wb->sorted.size - wb->flushing.keys.nr);
241 	nr = min(nr, wb->inc.keys.nr);
242 
243 	memcpy(&darray_top(wb->flushing.keys),
244 	       wb->inc.keys.data,
245 	       sizeof(wb->inc.keys.data[0]) * nr);
246 
247 	memmove(wb->inc.keys.data,
248 		wb->inc.keys.data + nr,
249 	       sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));
250 
251 	wb->flushing.keys.nr	+= nr;
252 	wb->inc.keys.nr		-= nr;
253 out:
254 	if (!wb->inc.keys.nr)
255 		bch2_journal_pin_drop(j, &wb->inc.pin);
256 	else
257 		bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
258 					bch2_btree_write_buffer_journal_flush);
259 
260 	if (j->watermark) {
261 		spin_lock(&j->lock);
262 		bch2_journal_set_watermark(j);
263 		spin_unlock(&j->lock);
264 	}
265 
266 	BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
267 }
268 
bch2_btree_write_buffer_flush_locked(struct btree_trans * trans)269 static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
270 {
271 	struct bch_fs *c = trans->c;
272 	struct journal *j = &c->journal;
273 	struct btree_write_buffer *wb = &c->btree_write_buffer;
274 	struct btree_iter iter = { NULL };
275 	size_t overwritten = 0, fast = 0, slowpath = 0, could_not_insert = 0;
276 	bool write_locked = false;
277 	bool accounting_replay_done = test_bit(BCH_FS_accounting_replay_done, &c->flags);
278 	int ret = 0;
279 
280 	ret = bch2_journal_error(&c->journal);
281 	if (ret)
282 		return ret;
283 
284 	bch2_trans_unlock(trans);
285 	bch2_trans_begin(trans);
286 
287 	mutex_lock(&wb->inc.lock);
288 	move_keys_from_inc_to_flushing(wb);
289 	mutex_unlock(&wb->inc.lock);
290 
291 	for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
292 		wb->sorted.data[i].idx = i;
293 		wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
294 		memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
295 	}
296 	wb->sorted.nr = wb->flushing.keys.nr;
297 
298 	/*
299 	 * We first sort so that we can detect and skip redundant updates, and
300 	 * then we attempt to flush in sorted btree order, as this is most
301 	 * efficient.
302 	 *
303 	 * However, since we're not flushing in the order they appear in the
304 	 * journal we won't be able to drop our journal pin until everything is
305 	 * flushed - which means this could deadlock the journal if we weren't
306 	 * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
307 	 * if it would block taking a journal reservation.
308 	 *
309 	 * If that happens, simply skip the key so we can optimistically insert
310 	 * as many keys as possible in the fast path.
311 	 */
312 	wb_sort(wb->sorted.data, wb->sorted.nr);
313 
314 	darray_for_each(wb->sorted, i) {
315 		struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
316 
317 		for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
318 			prefetch(&wb->flushing.keys.data[n->idx]);
319 
320 		BUG_ON(!k->journal_seq);
321 
322 		if (!accounting_replay_done &&
323 		    k->k.k.type == KEY_TYPE_accounting) {
324 			slowpath++;
325 			continue;
326 		}
327 
328 		if (i + 1 < &darray_top(wb->sorted) &&
329 		    wb_key_eq(i, i + 1)) {
330 			struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];
331 
332 			if (k->k.k.type == KEY_TYPE_accounting &&
333 			    n->k.k.type == KEY_TYPE_accounting)
334 				bch2_accounting_accumulate(bkey_i_to_accounting(&n->k),
335 							   bkey_i_to_s_c_accounting(&k->k));
336 
337 			overwritten++;
338 			n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
339 			k->journal_seq = 0;
340 			continue;
341 		}
342 
343 		if (write_locked) {
344 			struct btree_path *path = btree_iter_path(trans, &iter);
345 
346 			if (path->btree_id != i->btree ||
347 			    bpos_gt(k->k.k.p, path->l[0].b->key.k.p)) {
348 				bch2_btree_node_unlock_write(trans, path, path->l[0].b);
349 				write_locked = false;
350 
351 				ret = lockrestart_do(trans,
352 					bch2_btree_iter_traverse(&iter) ?:
353 					bch2_foreground_maybe_merge(trans, iter.path, 0,
354 							BCH_WATERMARK_reclaim|
355 							BCH_TRANS_COMMIT_journal_reclaim|
356 							BCH_TRANS_COMMIT_no_check_rw|
357 							BCH_TRANS_COMMIT_no_enospc));
358 				if (ret)
359 					goto err;
360 			}
361 		}
362 
363 		if (!iter.path || iter.btree_id != k->btree) {
364 			bch2_trans_iter_exit(trans, &iter);
365 			bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
366 					     BTREE_ITER_intent|BTREE_ITER_all_snapshots);
367 		}
368 
369 		bch2_btree_iter_set_pos(&iter, k->k.k.p);
370 		btree_iter_path(trans, &iter)->preserve = false;
371 
372 		bool accounting_accumulated = false;
373 		do {
374 			if (race_fault()) {
375 				ret = -BCH_ERR_journal_reclaim_would_deadlock;
376 				break;
377 			}
378 
379 			ret = wb_flush_one(trans, &iter, k, &write_locked,
380 					   &accounting_accumulated, &fast);
381 			if (!write_locked)
382 				bch2_trans_begin(trans);
383 		} while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
384 
385 		if (!ret) {
386 			k->journal_seq = 0;
387 		} else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
388 			slowpath++;
389 			ret = 0;
390 		} else
391 			break;
392 	}
393 
394 	if (write_locked) {
395 		struct btree_path *path = btree_iter_path(trans, &iter);
396 		bch2_btree_node_unlock_write(trans, path, path->l[0].b);
397 	}
398 	bch2_trans_iter_exit(trans, &iter);
399 
400 	if (ret)
401 		goto err;
402 
403 	if (slowpath) {
404 		/*
405 		 * Flush in the order they were present in the journal, so that
406 		 * we can release journal pins:
407 		 * The fastpath zapped the seq of keys that were successfully flushed so
408 		 * we can skip those here.
409 		 */
410 		trace_and_count(c, write_buffer_flush_slowpath, trans, slowpath, wb->flushing.keys.nr);
411 
412 		sort(wb->flushing.keys.data,
413 		     wb->flushing.keys.nr,
414 		     sizeof(wb->flushing.keys.data[0]),
415 		     wb_key_seq_cmp, NULL);
416 
417 		darray_for_each(wb->flushing.keys, i) {
418 			if (!i->journal_seq)
419 				continue;
420 
421 			if (!accounting_replay_done &&
422 			    i->k.k.type == KEY_TYPE_accounting) {
423 				could_not_insert++;
424 				continue;
425 			}
426 
427 			if (!could_not_insert)
428 				bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
429 							bch2_btree_write_buffer_journal_flush);
430 
431 			bch2_trans_begin(trans);
432 
433 			ret = commit_do(trans, NULL, NULL,
434 					BCH_WATERMARK_reclaim|
435 					BCH_TRANS_COMMIT_journal_reclaim|
436 					BCH_TRANS_COMMIT_no_check_rw|
437 					BCH_TRANS_COMMIT_no_enospc|
438 					BCH_TRANS_COMMIT_no_journal_res ,
439 					btree_write_buffered_insert(trans, i));
440 			if (ret)
441 				goto err;
442 
443 			i->journal_seq = 0;
444 		}
445 
446 		/*
447 		 * If journal replay hasn't finished with accounting keys we
448 		 * can't flush accounting keys at all - condense them and leave
449 		 * them for next time.
450 		 *
451 		 * Q: Can the write buffer overflow?
452 		 * A Shouldn't be any actual risk. It's just new accounting
453 		 * updates that the write buffer can't flush, and those are only
454 		 * going to be generated by interior btree node updates as
455 		 * journal replay has to split/rewrite nodes to make room for
456 		 * its updates.
457 		 *
458 		 * And for those new acounting updates, updates to the same
459 		 * counters get accumulated as they're flushed from the journal
460 		 * to the write buffer - see the patch for eytzingcer tree
461 		 * accumulated. So we could only overflow if the number of
462 		 * distinct counters touched somehow was very large.
463 		 */
464 		if (could_not_insert) {
465 			struct btree_write_buffered_key *dst = wb->flushing.keys.data;
466 
467 			darray_for_each(wb->flushing.keys, i)
468 				if (i->journal_seq)
469 					*dst++ = *i;
470 			wb->flushing.keys.nr = dst - wb->flushing.keys.data;
471 		}
472 	}
473 err:
474 	if (ret || !could_not_insert) {
475 		bch2_journal_pin_drop(j, &wb->flushing.pin);
476 		wb->flushing.keys.nr = 0;
477 	}
478 
479 	bch2_fs_fatal_err_on(ret, c, "%s", bch2_err_str(ret));
480 	trace_write_buffer_flush(trans, wb->flushing.keys.nr, overwritten, fast, 0);
481 	return ret;
482 }
483 
fetch_wb_keys_from_journal(struct bch_fs * c,u64 seq)484 static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
485 {
486 	struct journal *j = &c->journal;
487 	struct journal_buf *buf;
488 	int ret = 0;
489 
490 	while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, seq))) {
491 		ret = bch2_journal_keys_to_write_buffer(c, buf);
492 		mutex_unlock(&j->buf_lock);
493 	}
494 
495 	return ret;
496 }
497 
btree_write_buffer_flush_seq(struct btree_trans * trans,u64 seq,bool * did_work)498 static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq,
499 					bool *did_work)
500 {
501 	struct bch_fs *c = trans->c;
502 	struct btree_write_buffer *wb = &c->btree_write_buffer;
503 	int ret = 0, fetch_from_journal_err;
504 
505 	do {
506 		bch2_trans_unlock(trans);
507 
508 		fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
509 
510 		*did_work |= wb->inc.keys.nr || wb->flushing.keys.nr;
511 
512 		/*
513 		 * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
514 		 * is not guaranteed to empty wb->inc:
515 		 */
516 		mutex_lock(&wb->flushing.lock);
517 		ret = bch2_btree_write_buffer_flush_locked(trans);
518 		mutex_unlock(&wb->flushing.lock);
519 	} while (!ret &&
520 		 (fetch_from_journal_err ||
521 		  (wb->inc.pin.seq && wb->inc.pin.seq <= seq) ||
522 		  (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq)));
523 
524 	return ret;
525 }
526 
bch2_btree_write_buffer_journal_flush(struct journal * j,struct journal_entry_pin * _pin,u64 seq)527 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
528 				struct journal_entry_pin *_pin, u64 seq)
529 {
530 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
531 	bool did_work = false;
532 
533 	return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq, &did_work));
534 }
535 
bch2_btree_write_buffer_flush_sync(struct btree_trans * trans)536 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
537 {
538 	struct bch_fs *c = trans->c;
539 	bool did_work = false;
540 
541 	trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);
542 
543 	return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal), &did_work);
544 }
545 
546 /*
547  * The write buffer requires flushing when going RO: keys in the journal for the
548  * write buffer don't have a journal pin yet
549  */
bch2_btree_write_buffer_flush_going_ro(struct bch_fs * c)550 bool bch2_btree_write_buffer_flush_going_ro(struct bch_fs *c)
551 {
552 	if (bch2_journal_error(&c->journal))
553 		return false;
554 
555 	bool did_work = false;
556 	bch2_trans_run(c, btree_write_buffer_flush_seq(trans,
557 				journal_cur_seq(&c->journal), &did_work));
558 	return did_work;
559 }
560 
bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans * trans)561 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
562 {
563 	struct bch_fs *c = trans->c;
564 	struct btree_write_buffer *wb = &c->btree_write_buffer;
565 	int ret = 0;
566 
567 	if (mutex_trylock(&wb->flushing.lock)) {
568 		ret = bch2_btree_write_buffer_flush_locked(trans);
569 		mutex_unlock(&wb->flushing.lock);
570 	}
571 
572 	return ret;
573 }
574 
bch2_btree_write_buffer_tryflush(struct btree_trans * trans)575 int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
576 {
577 	struct bch_fs *c = trans->c;
578 
579 	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer))
580 		return -BCH_ERR_erofs_no_writes;
581 
582 	int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
583 	bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
584 	return ret;
585 }
586 
587 /*
588  * In check and repair code, when checking references to write buffer btrees we
589  * need to issue a flush before we have a definitive error: this issues a flush
590  * if this is a key we haven't yet checked.
591  */
bch2_btree_write_buffer_maybe_flush(struct btree_trans * trans,struct bkey_s_c referring_k,struct bkey_buf * last_flushed)592 int bch2_btree_write_buffer_maybe_flush(struct btree_trans *trans,
593 					struct bkey_s_c referring_k,
594 					struct bkey_buf *last_flushed)
595 {
596 	struct bch_fs *c = trans->c;
597 	struct bkey_buf tmp;
598 	int ret = 0;
599 
600 	bch2_bkey_buf_init(&tmp);
601 
602 	if (!bkey_and_val_eq(referring_k, bkey_i_to_s_c(last_flushed->k))) {
603 		bch2_bkey_buf_reassemble(&tmp, c, referring_k);
604 
605 		if (bkey_is_btree_ptr(referring_k.k)) {
606 			bch2_trans_unlock(trans);
607 			bch2_btree_interior_updates_flush(c);
608 		}
609 
610 		ret = bch2_btree_write_buffer_flush_sync(trans);
611 		if (ret)
612 			goto err;
613 
614 		bch2_bkey_buf_copy(last_flushed, c, tmp.k);
615 		ret = -BCH_ERR_transaction_restart_write_buffer_flush;
616 	}
617 err:
618 	bch2_bkey_buf_exit(&tmp, c);
619 	return ret;
620 }
621 
bch2_btree_write_buffer_flush_work(struct work_struct * work)622 static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
623 {
624 	struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
625 	struct btree_write_buffer *wb = &c->btree_write_buffer;
626 	int ret;
627 
628 	mutex_lock(&wb->flushing.lock);
629 	do {
630 		ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
631 	} while (!ret && bch2_btree_write_buffer_should_flush(c));
632 	mutex_unlock(&wb->flushing.lock);
633 
634 	bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
635 }
636 
wb_accounting_sort(struct btree_write_buffer * wb)637 static void wb_accounting_sort(struct btree_write_buffer *wb)
638 {
639 	eytzinger0_sort(wb->accounting.data, wb->accounting.nr,
640 			sizeof(wb->accounting.data[0]),
641 			wb_key_cmp, NULL);
642 }
643 
bch2_accounting_key_to_wb_slowpath(struct bch_fs * c,enum btree_id btree,struct bkey_i_accounting * k)644 int bch2_accounting_key_to_wb_slowpath(struct bch_fs *c, enum btree_id btree,
645 				       struct bkey_i_accounting *k)
646 {
647 	struct btree_write_buffer *wb = &c->btree_write_buffer;
648 	struct btree_write_buffered_key new = { .btree = btree };
649 
650 	bkey_copy(&new.k, &k->k_i);
651 
652 	int ret = darray_push(&wb->accounting, new);
653 	if (ret)
654 		return ret;
655 
656 	wb_accounting_sort(wb);
657 	return 0;
658 }
659 
bch2_journal_key_to_wb_slowpath(struct bch_fs * c,struct journal_keys_to_wb * dst,enum btree_id btree,struct bkey_i * k)660 int bch2_journal_key_to_wb_slowpath(struct bch_fs *c,
661 			     struct journal_keys_to_wb *dst,
662 			     enum btree_id btree, struct bkey_i *k)
663 {
664 	struct btree_write_buffer *wb = &c->btree_write_buffer;
665 	int ret;
666 retry:
667 	ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
668 	if (!ret && dst->wb == &wb->flushing)
669 		ret = darray_resize(&wb->sorted, wb->flushing.keys.size);
670 
671 	if (unlikely(ret)) {
672 		if (dst->wb == &c->btree_write_buffer.flushing) {
673 			mutex_unlock(&dst->wb->lock);
674 			dst->wb = &c->btree_write_buffer.inc;
675 			bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
676 					     bch2_btree_write_buffer_journal_flush);
677 			goto retry;
678 		}
679 
680 		return ret;
681 	}
682 
683 	dst->room = darray_room(dst->wb->keys);
684 	if (dst->wb == &wb->flushing)
685 		dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
686 	BUG_ON(!dst->room);
687 	BUG_ON(!dst->seq);
688 
689 	struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
690 	wb_k->journal_seq	= dst->seq;
691 	wb_k->btree		= btree;
692 	bkey_copy(&wb_k->k, k);
693 	dst->wb->keys.nr++;
694 	dst->room--;
695 	return 0;
696 }
697 
bch2_journal_keys_to_write_buffer_start(struct bch_fs * c,struct journal_keys_to_wb * dst,u64 seq)698 void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
699 {
700 	struct btree_write_buffer *wb = &c->btree_write_buffer;
701 
702 	if (mutex_trylock(&wb->flushing.lock)) {
703 		mutex_lock(&wb->inc.lock);
704 		move_keys_from_inc_to_flushing(wb);
705 
706 		/*
707 		 * Attempt to skip wb->inc, and add keys directly to
708 		 * wb->flushing, saving us a copy later:
709 		 */
710 
711 		if (!wb->inc.keys.nr) {
712 			dst->wb = &wb->flushing;
713 		} else {
714 			mutex_unlock(&wb->flushing.lock);
715 			dst->wb = &wb->inc;
716 		}
717 	} else {
718 		mutex_lock(&wb->inc.lock);
719 		dst->wb = &wb->inc;
720 	}
721 
722 	dst->room = darray_room(dst->wb->keys);
723 	if (dst->wb == &wb->flushing)
724 		dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
725 	dst->seq = seq;
726 
727 	bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
728 			     bch2_btree_write_buffer_journal_flush);
729 
730 	darray_for_each(wb->accounting, i)
731 		memset(&i->k.v, 0, bkey_val_bytes(&i->k.k));
732 }
733 
bch2_journal_keys_to_write_buffer_end(struct bch_fs * c,struct journal_keys_to_wb * dst)734 int bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
735 {
736 	struct btree_write_buffer *wb = &c->btree_write_buffer;
737 	unsigned live_accounting_keys = 0;
738 	int ret = 0;
739 
740 	darray_for_each(wb->accounting, i)
741 		if (!bch2_accounting_key_is_zero(bkey_i_to_s_c_accounting(&i->k))) {
742 			i->journal_seq = dst->seq;
743 			live_accounting_keys++;
744 			ret = __bch2_journal_key_to_wb(c, dst, i->btree, &i->k);
745 			if (ret)
746 				break;
747 		}
748 
749 	if (live_accounting_keys * 2 < wb->accounting.nr) {
750 		struct btree_write_buffered_key *dst = wb->accounting.data;
751 
752 		darray_for_each(wb->accounting, src)
753 			if (!bch2_accounting_key_is_zero(bkey_i_to_s_c_accounting(&src->k)))
754 				*dst++ = *src;
755 		wb->accounting.nr = dst - wb->accounting.data;
756 		wb_accounting_sort(wb);
757 	}
758 
759 	if (!dst->wb->keys.nr)
760 		bch2_journal_pin_drop(&c->journal, &dst->wb->pin);
761 
762 	if (bch2_btree_write_buffer_should_flush(c) &&
763 	    __bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer) &&
764 	    !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
765 		bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
766 
767 	if (dst->wb == &wb->flushing)
768 		mutex_unlock(&wb->flushing.lock);
769 	mutex_unlock(&wb->inc.lock);
770 
771 	return ret;
772 }
773 
bch2_journal_keys_to_write_buffer(struct bch_fs * c,struct journal_buf * buf)774 static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
775 {
776 	struct journal_keys_to_wb dst;
777 	int ret = 0;
778 
779 	bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
780 
781 	for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
782 		jset_entry_for_each_key(entry, k) {
783 			ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
784 			if (ret)
785 				goto out;
786 		}
787 
788 		entry->type = BCH_JSET_ENTRY_btree_keys;
789 	}
790 
791 	spin_lock(&c->journal.lock);
792 	buf->need_flush_to_write_buffer = false;
793 	spin_unlock(&c->journal.lock);
794 out:
795 	ret = bch2_journal_keys_to_write_buffer_end(c, &dst) ?: ret;
796 	return ret;
797 }
798 
wb_keys_resize(struct btree_write_buffer_keys * wb,size_t new_size)799 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
800 {
801 	if (wb->keys.size >= new_size)
802 		return 0;
803 
804 	if (!mutex_trylock(&wb->lock))
805 		return -EINTR;
806 
807 	int ret = darray_resize(&wb->keys, new_size);
808 	mutex_unlock(&wb->lock);
809 	return ret;
810 }
811 
bch2_btree_write_buffer_resize(struct bch_fs * c,size_t new_size)812 int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
813 {
814 	struct btree_write_buffer *wb = &c->btree_write_buffer;
815 
816 	return wb_keys_resize(&wb->flushing, new_size) ?:
817 		wb_keys_resize(&wb->inc, new_size);
818 }
819 
bch2_fs_btree_write_buffer_exit(struct bch_fs * c)820 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
821 {
822 	struct btree_write_buffer *wb = &c->btree_write_buffer;
823 
824 	BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
825 	       !bch2_journal_error(&c->journal));
826 
827 	darray_exit(&wb->accounting);
828 	darray_exit(&wb->sorted);
829 	darray_exit(&wb->flushing.keys);
830 	darray_exit(&wb->inc.keys);
831 }
832 
bch2_fs_btree_write_buffer_init(struct bch_fs * c)833 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
834 {
835 	struct btree_write_buffer *wb = &c->btree_write_buffer;
836 
837 	mutex_init(&wb->inc.lock);
838 	mutex_init(&wb->flushing.lock);
839 	INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
840 
841 	/* Will be resized by journal as needed: */
842 	unsigned initial_size = 1 << 16;
843 
844 	return  darray_make_room(&wb->inc.keys, initial_size) ?:
845 		darray_make_room(&wb->flushing.keys, initial_size) ?:
846 		darray_make_room(&wb->sorted, initial_size);
847 }
848