1	// SPDX-License-Identifier: GPL-2.0
2
3	/* erasure coding */
4
5	#include "bcachefs.h"
6	#include "alloc_background.h"
7	#include "alloc_foreground.h"
8	#include "backpointers.h"
9	#include "bkey_buf.h"
10	#include "bset.h"
11	#include "btree_gc.h"
12	#include "btree_update.h"
13	#include "btree_write_buffer.h"
14	#include "buckets.h"
15	#include "checksum.h"
16	#include "disk_groups.h"
17	#include "ec.h"
18	#include "error.h"
19	#include "io_read.h"
20	#include "keylist.h"
21	#include "recovery.h"
22	#include "replicas.h"
23	#include "super-io.h"
24	#include "util.h"
25
26	#include <linux/sort.h>
27
28	#ifdef __KERNEL__
29
30	#include <linux/raid/pq.h>
31	#include <linux/raid/xor.h>
32
33	static void raid5_recov(unsigned disks, unsigned failed_idx,
34	size_t size, void **data)
35	{
36	unsigned i = 2, nr;
37
38	BUG_ON(failed_idx >= disks);
39
40	swap(data[0], data[failed_idx]);
41	memcpy(data[0], data[1], size);
42
43	while (i < disks) {
44	nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
45	xor_blocks(count: nr, bytes: size, dest: data[0], srcs: data + i);
46	i += nr;
47	}
48
49	swap(data[0], data[failed_idx]);
50	}
51
52	static void raid_gen(int nd, int np, size_t size, void **v)
53	{
54	if (np >= 1)
55	raid5_recov(disks: nd + np, failed_idx: nd, size, data: v);
56	if (np >= 2)
57	raid6_call.gen_syndrome(nd + np, size, v);
58	BUG_ON(np > 2);
59	}
60
61	static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
62	{
63	switch (nr) {
64	case 0:
65	break;
66	case 1:
67	if (ir[0] < nd + 1)
68	raid5_recov(disks: nd + 1, failed_idx: ir[0], size, data: v);
69	else
70	raid6_call.gen_syndrome(nd + np, size, v);
71	break;
72	case 2:
73	if (ir[1] < nd) {
74	/* data+data failure. */
75	raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
76	} else if (ir[0] < nd) {
77	/* data + p/q failure */
78
79	if (ir[1] == nd) /* data + p failure */
80	raid6_datap_recov(nd + np, size, ir[0], v);
81	else { /* data + q failure */
82	raid5_recov(disks: nd + 1, failed_idx: ir[0], size, data: v);
83	raid6_call.gen_syndrome(nd + np, size, v);
84	}
85	} else {
86	raid_gen(nd, np, size, v);
87	}
88	break;
89	default:
90	BUG();
91	}
92	}
93
94	#else
95
96	#include <raid/raid.h>
97
98	#endif
99
100	struct ec_bio {
101	struct bch_dev *ca;
102	struct ec_stripe_buf *buf;
103	size_t idx;
104	struct bio bio;
105	};
106
107	/* Stripes btree keys: */
108
109	int bch2_stripe_invalid(struct bch_fs *c, struct bkey_s_c k,
110	enum bkey_invalid_flags flags,
111	struct printbuf *err)
112	{
113	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
114	int ret = 0;
115
116	bkey_fsck_err_on(bkey_eq(k.k->p, POS_MIN) ||
117	bpos_gt(k.k->p, POS(0, U32_MAX)), c, err,
118	stripe_pos_bad,
119	"stripe at bad pos");
120
121	bkey_fsck_err_on(bkey_val_u64s(k.k) < stripe_val_u64s(s), c, err,
122	stripe_val_size_bad,
123	"incorrect value size (%zu < %u)",
124	bkey_val_u64s(k.k), stripe_val_u64s(s));
125
126	ret = bch2_bkey_ptrs_invalid(c, k, flags, err);
127	fsck_err:
128	return ret;
129	}
130
131	void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
132	struct bkey_s_c k)
133	{
134	const struct bch_stripe *sp = bkey_s_c_to_stripe(k).v;
135	struct bch_stripe s = {};
136
137	memcpy(&s, sp, min(sizeof(s), bkey_val_bytes(k.k)));
138
139	unsigned nr_data = s.nr_blocks - s.nr_redundant;
140
141	prt_printf(out, "algo %u sectors %u blocks %u:%u csum ",
142	s.algorithm,
143	le16_to_cpu(s.sectors),
144	nr_data,
145	s.nr_redundant);
146	bch2_prt_csum_type(out, s.csum_type);
147	prt_printf(out, " gran %u", 1U << s.csum_granularity_bits);
148
149	for (unsigned i = 0; i < s.nr_blocks; i++) {
150	const struct bch_extent_ptr *ptr = sp->ptrs + i;
151
152	if ((void *) ptr >= bkey_val_end(k))
153	break;
154
155	bch2_extent_ptr_to_text(out, c, ptr);
156
157	if (s.csum_type < BCH_CSUM_NR &&
158	i < nr_data &&
159	stripe_blockcount_offset(s: &s, idx: i) < bkey_val_bytes(k: k.k))
160	prt_printf(out, "#%u", stripe_blockcount_get(sp, i));
161	}
162	}
163
164	/* Triggers: */
165
166	static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
167	struct bkey_s_c_stripe s,
168	unsigned idx, bool deleting)
169	{
170	struct bch_fs *c = trans->c;
171	const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
172	struct btree_iter iter;
173	struct bkey_i_alloc_v4 *a;
174	enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
175	? BCH_DATA_parity : 0;
176	s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
177	int ret = 0;
178
179	if (deleting)
180	sectors = -sectors;
181
182	a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
183	if (IS_ERR(ptr: a))
184	return PTR_ERR(ptr: a);
185
186	ret = bch2_check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
187	a->v.gen, a->v.data_type,
188	a->v.dirty_sectors);
189	if (ret)
190	goto err;
191
192	if (!deleting) {
193	if (bch2_trans_inconsistent_on(a->v.stripe ||
194	a->v.stripe_redundancy, trans,
195	"bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
196	iter.pos.inode, iter.pos.offset, a->v.gen,
197	bch2_data_type_str(a->v.data_type),
198	a->v.dirty_sectors,
199	a->v.stripe, s.k->p.offset)) {
200	ret = -EIO;
201	goto err;
202	}
203
204	if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
205	"bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
206	iter.pos.inode, iter.pos.offset, a->v.gen,
207	bch2_data_type_str(a->v.data_type),
208	a->v.dirty_sectors,
209	s.k->p.offset)) {
210	ret = -EIO;
211	goto err;
212	}
213
214	a->v.stripe = s.k->p.offset;
215	a->v.stripe_redundancy = s.v->nr_redundant;
216	a->v.data_type = BCH_DATA_stripe;
217	} else {
218	if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
219	a->v.stripe_redundancy != s.v->nr_redundant, trans,
220	"bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
221	iter.pos.inode, iter.pos.offset, a->v.gen,
222	s.k->p.offset, a->v.stripe)) {
223	ret = -EIO;
224	goto err;
225	}
226
227	a->v.stripe = 0;
228	a->v.stripe_redundancy = 0;
229	a->v.data_type = alloc_data_type(a: a->v, data_type: BCH_DATA_user);
230	}
231
232	a->v.dirty_sectors += sectors;
233	if (data_type)
234	a->v.data_type = !deleting ? data_type : 0;
235
236	ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
237	if (ret)
238	goto err;
239	err:
240	bch2_trans_iter_exit(trans, &iter);
241	return ret;
242	}
243
244	static int mark_stripe_bucket(struct btree_trans *trans,
245	struct bkey_s_c k,
246	unsigned ptr_idx,
247	unsigned flags)
248	{
249	struct bch_fs *c = trans->c;
250	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
251	unsigned nr_data = s->nr_blocks - s->nr_redundant;
252	bool parity = ptr_idx >= nr_data;
253	enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
254	s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
255	const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
256	struct bch_dev *ca = bch_dev_bkey_exists(c, idx: ptr->dev);
257	struct bucket old, new, *g;
258	struct printbuf buf = PRINTBUF;
259	int ret = 0;
260
261	BUG_ON(!(flags & BTREE_TRIGGER_GC));
262
263	/* * XXX doesn't handle deletion */
264
265	percpu_down_read(sem: &c->mark_lock);
266	g = PTR_GC_BUCKET(ca, ptr);
267
268	if (g->dirty_sectors ||
269	(g->stripe && g->stripe != k.k->p.offset)) {
270	bch2_fs_inconsistent(c,
271	"bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
272	ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
273	(bch2_bkey_val_to_text(&buf, c, k), buf.buf));
274	ret = -EINVAL;
275	goto err;
276	}
277
278	bucket_lock(b: g);
279	old = *g;
280
281	ret = bch2_check_bucket_ref(trans, k, ptr, sectors, data_type,
282	g->gen, g->data_type,
283	g->dirty_sectors);
284	if (ret)
285	goto err;
286
287	g->data_type = data_type;
288	g->dirty_sectors += sectors;
289
290	g->stripe = k.k->p.offset;
291	g->stripe_redundancy = s->nr_redundant;
292	new = *g;
293	err:
294	bucket_unlock(b: g);
295	if (!ret)
296	bch2_dev_usage_update_m(c, ca, &old, &new);
297	percpu_up_read(sem: &c->mark_lock);
298	printbuf_exit(&buf);
299	return ret;
300	}
301
302	int bch2_trigger_stripe(struct btree_trans *trans,
303	enum btree_id btree_id, unsigned level,
304	struct bkey_s_c old, struct bkey_s _new,
305	unsigned flags)
306	{
307	struct bkey_s_c new = _new.s_c;
308	struct bch_fs *c = trans->c;
309	u64 idx = new.k->p.offset;
310	const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
311	? bkey_s_c_to_stripe(k: old).v : NULL;
312	const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
313	? bkey_s_c_to_stripe(k: new).v : NULL;
314
315	if (flags & BTREE_TRIGGER_TRANSACTIONAL) {
316	/*
317	* If the pointers aren't changing, we don't need to do anything:
318	*/
319	if (new_s && old_s &&
320	new_s->nr_blocks == old_s->nr_blocks &&
321	new_s->nr_redundant == old_s->nr_redundant &&
322	!memcmp(p: old_s->ptrs, q: new_s->ptrs,
323	size: new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
324	return 0;
325
326	BUG_ON(new_s && old_s &&
327	(new_s->nr_blocks != old_s->nr_blocks ||
328	new_s->nr_redundant != old_s->nr_redundant));
329
330	if (new_s) {
331	s64 sectors = le16_to_cpu(new_s->sectors);
332
333	struct bch_replicas_padded r;
334	bch2_bkey_to_replicas(&r.e, new);
335	int ret = bch2_update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
336	if (ret)
337	return ret;
338	}
339
340	if (old_s) {
341	s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
342
343	struct bch_replicas_padded r;
344	bch2_bkey_to_replicas(&r.e, old);
345	int ret = bch2_update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
346	if (ret)
347	return ret;
348	}
349
350	unsigned nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
351	for (unsigned i = 0; i < nr_blocks; i++) {
352	if (new_s && old_s &&
353	!memcmp(p: &new_s->ptrs[i],
354	q: &old_s->ptrs[i],
355	size: sizeof(new_s->ptrs[i])))
356	continue;
357
358	if (new_s) {
359	int ret = bch2_trans_mark_stripe_bucket(trans,
360	s: bkey_s_c_to_stripe(k: new), idx: i, deleting: false);
361	if (ret)
362	return ret;
363	}
364
365	if (old_s) {
366	int ret = bch2_trans_mark_stripe_bucket(trans,
367	s: bkey_s_c_to_stripe(k: old), idx: i, deleting: true);
368	if (ret)
369	return ret;
370	}
371	}
372	}
373
374	if (flags & BTREE_TRIGGER_ATOMIC) {
375	struct stripe *m = genradix_ptr(&c->stripes, idx);
376
377	if (!m) {
378	struct printbuf buf1 = PRINTBUF;
379	struct printbuf buf2 = PRINTBUF;
380
381	bch2_bkey_val_to_text(&buf1, c, old);
382	bch2_bkey_val_to_text(&buf2, c, new);
383	bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
384	"old %s\n"
385	"new %s", idx, buf1.buf, buf2.buf);
386	printbuf_exit(&buf2);
387	printbuf_exit(&buf1);
388	bch2_inconsistent_error(c);
389	return -1;
390	}
391
392	if (!new_s) {
393	bch2_stripes_heap_del(c, m, idx);
394
395	memset(m, 0, sizeof(*m));
396	} else {
397	m->sectors = le16_to_cpu(new_s->sectors);
398	m->algorithm = new_s->algorithm;
399	m->nr_blocks = new_s->nr_blocks;
400	m->nr_redundant = new_s->nr_redundant;
401	m->blocks_nonempty = 0;
402
403	for (unsigned i = 0; i < new_s->nr_blocks; i++)
404	m->blocks_nonempty += !!stripe_blockcount_get(s: new_s, idx: i);
405
406	if (!old_s)
407	bch2_stripes_heap_insert(c, m, idx);
408	else
409	bch2_stripes_heap_update(c, m, idx);
410	}
411	}
412
413	if (flags & BTREE_TRIGGER_GC) {
414	struct gc_stripe *m =
415	genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
416
417	if (!m) {
418	bch_err(c, "error allocating memory for gc_stripes, idx %llu",
419	idx);
420	return -BCH_ERR_ENOMEM_mark_stripe;
421	}
422	/*
423	* This will be wrong when we bring back runtime gc: we should
424	* be unmarking the old key and then marking the new key
425	*/
426	m->alive = true;
427	m->sectors = le16_to_cpu(new_s->sectors);
428	m->nr_blocks = new_s->nr_blocks;
429	m->nr_redundant = new_s->nr_redundant;
430
431	for (unsigned i = 0; i < new_s->nr_blocks; i++)
432	m->ptrs[i] = new_s->ptrs[i];
433
434	bch2_bkey_to_replicas(&m->r.e, new);
435
436	/*
437	* gc recalculates this field from stripe ptr
438	* references:
439	*/
440	memset(m->block_sectors, 0, sizeof(m->block_sectors));
441
442	for (unsigned i = 0; i < new_s->nr_blocks; i++) {
443	int ret = mark_stripe_bucket(trans, k: new, ptr_idx: i, flags);
444	if (ret)
445	return ret;
446	}
447
448	int ret = bch2_update_replicas(c, new, &m->r.e,
449	((s64) m->sectors * m->nr_redundant),
450	0, true);
451	if (ret) {
452	struct printbuf buf = PRINTBUF;
453
454	bch2_bkey_val_to_text(&buf, c, new);
455	bch2_fs_fatal_error(c, ": no replicas entry for %s", buf.buf);
456	printbuf_exit(&buf);
457	return ret;
458	}
459	}
460
461	return 0;
462	}
463
464	/* returns blocknr in stripe that we matched: */
465	static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
466	struct bkey_s_c k, unsigned *block)
467	{
468	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
469	unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
470
471	bkey_for_each_ptr(ptrs, ptr)
472	for (i = 0; i < nr_data; i++)
473	if (__bch2_ptr_matches_stripe(stripe_ptr: &s->ptrs[i], data_ptr: ptr,
474	le16_to_cpu(s->sectors))) {
475	*block = i;
476	return ptr;
477	}
478
479	return NULL;
480	}
481
482	static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
483	{
484	switch (k.k->type) {
485	case KEY_TYPE_extent: {
486	struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
487	const union bch_extent_entry *entry;
488
489	extent_for_each_entry(e, entry)
490	if (extent_entry_type(e: entry) ==
491	BCH_EXTENT_ENTRY_stripe_ptr &&
492	entry->stripe_ptr.idx == idx)
493	return true;
494
495	break;
496	}
497	}
498
499	return false;
500	}
501
502	/* Stripe bufs: */
503
504	static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
505	{
506	if (buf->key.k.type == KEY_TYPE_stripe) {
507	struct bkey_i_stripe *s = bkey_i_to_stripe(k: &buf->key);
508	unsigned i;
509
510	for (i = 0; i < s->v.nr_blocks; i++) {
511	kvfree(addr: buf->data[i]);
512	buf->data[i] = NULL;
513	}
514	}
515	}
516
517	/* XXX: this is a non-mempoolified memory allocation: */
518	static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
519	unsigned offset, unsigned size)
520	{
521	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
522	unsigned csum_granularity = 1U << v->csum_granularity_bits;
523	unsigned end = offset + size;
524	unsigned i;
525
526	BUG_ON(end > le16_to_cpu(v->sectors));
527
528	offset = round_down(offset, csum_granularity);
529	end = min_t(unsigned, le16_to_cpu(v->sectors),
530	round_up(end, csum_granularity));
531
532	buf->offset = offset;
533	buf->size = end - offset;
534
535	memset(buf->valid, 0xFF, sizeof(buf->valid));
536
537	for (i = 0; i < v->nr_blocks; i++) {
538	buf->data[i] = kvmalloc(size: buf->size << 9, GFP_KERNEL);
539	if (!buf->data[i])
540	goto err;
541	}
542
543	return 0;
544	err:
545	ec_stripe_buf_exit(buf);
546	return -BCH_ERR_ENOMEM_stripe_buf;
547	}
548
549	/* Checksumming: */
550
551	static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
552	unsigned block, unsigned offset)
553	{
554	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
555	unsigned csum_granularity = 1 << v->csum_granularity_bits;
556	unsigned end = buf->offset + buf->size;
557	unsigned len = min(csum_granularity, end - offset);
558
559	BUG_ON(offset >= end);
560	BUG_ON(offset < buf->offset);
561	BUG_ON(offset & (csum_granularity - 1));
562	BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
563	(len & (csum_granularity - 1)));
564
565	return bch2_checksum(NULL, v->csum_type,
566	null_nonce(),
567	buf->data[block] + ((offset - buf->offset) << 9),
568	len << 9);
569	}
570
571	static void ec_generate_checksums(struct ec_stripe_buf *buf)
572	{
573	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
574	unsigned i, j, csums_per_device = stripe_csums_per_device(s: v);
575
576	if (!v->csum_type)
577	return;
578
579	BUG_ON(buf->offset);
580	BUG_ON(buf->size != le16_to_cpu(v->sectors));
581
582	for (i = 0; i < v->nr_blocks; i++)
583	for (j = 0; j < csums_per_device; j++)
584	stripe_csum_set(s: v, block: i, csum_idx: j,
585	csum: ec_block_checksum(buf, block: i, offset: j << v->csum_granularity_bits));
586	}
587
588	static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
589	{
590	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
591	unsigned csum_granularity = 1 << v->csum_granularity_bits;
592	unsigned i;
593
594	if (!v->csum_type)
595	return;
596
597	for (i = 0; i < v->nr_blocks; i++) {
598	unsigned offset = buf->offset;
599	unsigned end = buf->offset + buf->size;
600
601	if (!test_bit(i, buf->valid))
602	continue;
603
604	while (offset < end) {
605	unsigned j = offset >> v->csum_granularity_bits;
606	unsigned len = min(csum_granularity, end - offset);
607	struct bch_csum want = stripe_csum_get(s: v, block: i, csum_idx: j);
608	struct bch_csum got = ec_block_checksum(buf, block: i, offset);
609
610	if (bch2_crc_cmp(l: want, r: got)) {
611	struct printbuf err = PRINTBUF;
612	struct bch_dev *ca = bch_dev_bkey_exists(c, idx: v->ptrs[i].dev);
613
614	prt_str(out: &err, str: "stripe ");
615	bch2_csum_err_msg(out: &err, type: v->csum_type, expected: want, got);
616	prt_printf(&err, " for %ps at %u of\n ", (void *) _RET_IP_, i);
617	bch2_bkey_val_to_text(&err, c, bkey_i_to_s_c(k: &buf->key));
618	bch_err_ratelimited(ca, "%s", err.buf);
619	printbuf_exit(&err);
620
621	clear_bit(nr: i, addr: buf->valid);
622
623	bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
624	break;
625	}
626
627	offset += len;
628	}
629	}
630	}
631
632	/* Erasure coding: */
633
634	static void ec_generate_ec(struct ec_stripe_buf *buf)
635	{
636	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
637	unsigned nr_data = v->nr_blocks - v->nr_redundant;
638	unsigned bytes = le16_to_cpu(v->sectors) << 9;
639
640	raid_gen(nd: nr_data, np: v->nr_redundant, size: bytes, v: buf->data);
641	}
642
643	static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
644	{
645	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
646
647	return v->nr_blocks - bitmap_weight(src: buf->valid, nbits: v->nr_blocks);
648	}
649
650	static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
651	{
652	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
653	unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
654	unsigned nr_data = v->nr_blocks - v->nr_redundant;
655	unsigned bytes = buf->size << 9;
656
657	if (ec_nr_failed(buf) > v->nr_redundant) {
658	bch_err_ratelimited(c,
659	"error doing reconstruct read: unable to read enough blocks");
660	return -1;
661	}
662
663	for (i = 0; i < nr_data; i++)
664	if (!test_bit(i, buf->valid))
665	failed[nr_failed++] = i;
666
667	raid_rec(nr: nr_failed, ir: failed, nd: nr_data, np: v->nr_redundant, size: bytes, v: buf->data);
668	return 0;
669	}
670
671	/* IO: */
672
673	static void ec_block_endio(struct bio *bio)
674	{
675	struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
676	struct bch_stripe *v = &bkey_i_to_stripe(k: &ec_bio->buf->key)->v;
677	struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
678	struct bch_dev *ca = ec_bio->ca;
679	struct closure *cl = bio->bi_private;
680
681	if (bch2_dev_io_err_on(bio->bi_status, ca,
682	bio_data_dir(bio)
683	? BCH_MEMBER_ERROR_write
684	: BCH_MEMBER_ERROR_read,
685	"erasure coding %s error: %s",
686	bio_data_dir(bio) ? "write" : "read",
687	bch2_blk_status_to_str(bio->bi_status)))
688	clear_bit(nr: ec_bio->idx, addr: ec_bio->buf->valid);
689
690	if (ptr_stale(ca, ptr)) {
691	bch_err_ratelimited(ca->fs,
692	"error %s stripe: stale pointer after io",
693	bio_data_dir(bio) == READ ? "reading from" : "writing to");
694	clear_bit(nr: ec_bio->idx, addr: ec_bio->buf->valid);
695	}
696
697	bio_put(&ec_bio->bio);
698	percpu_ref_put(ref: &ca->io_ref);
699	closure_put(cl);
700	}
701
702	static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
703	blk_opf_t opf, unsigned idx, struct closure *cl)
704	{
705	struct bch_stripe *v = &bkey_i_to_stripe(k: &buf->key)->v;
706	unsigned offset = 0, bytes = buf->size << 9;
707	struct bch_extent_ptr *ptr = &v->ptrs[idx];
708	struct bch_dev *ca = bch_dev_bkey_exists(c, idx: ptr->dev);
709	enum bch_data_type data_type = idx < v->nr_blocks - v->nr_redundant
710	? BCH_DATA_user
711	: BCH_DATA_parity;
712	int rw = op_is_write(op: opf);
713
714	if (ptr_stale(ca, ptr)) {
715	bch_err_ratelimited(c,
716	"error %s stripe: stale pointer",
717	rw == READ ? "reading from" : "writing to");
718	clear_bit(nr: idx, addr: buf->valid);
719	return;
720	}
721
722	if (!bch2_dev_get_ioref(ca, rw)) {
723	clear_bit(nr: idx, addr: buf->valid);
724	return;
725	}
726
727	this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
728
729	while (offset < bytes) {
730	unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
731	DIV_ROUND_UP(bytes, PAGE_SIZE));
732	unsigned b = min_t(size_t, bytes - offset,
733	nr_iovecs << PAGE_SHIFT);
734	struct ec_bio *ec_bio;
735
736	ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
737	nr_iovecs,
738	opf,
739	GFP_KERNEL,
740	&c->ec_bioset),
741	struct ec_bio, bio);
742
743	ec_bio->ca = ca;
744	ec_bio->buf = buf;
745	ec_bio->idx = idx;
746
747	ec_bio->bio.bi_iter.bi_sector = ptr->offset + buf->offset + (offset >> 9);
748	ec_bio->bio.bi_end_io = ec_block_endio;
749	ec_bio->bio.bi_private = cl;
750
751	bch2_bio_map(bio: &ec_bio->bio, base: buf->data[idx] + offset, b);
752
753	closure_get(cl);
754	percpu_ref_get(ref: &ca->io_ref);
755
756	submit_bio(bio: &ec_bio->bio);
757
758	offset += b;
759	}
760
761	percpu_ref_put(ref: &ca->io_ref);
762	}
763
764	static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
765	struct ec_stripe_buf *stripe)
766	{
767	struct btree_iter iter;
768	struct bkey_s_c k;
769	int ret;
770
771	k = bch2_bkey_get_iter(trans, iter: &iter, btree_id: BTREE_ID_stripes,
772	POS(0, idx), flags: BTREE_ITER_SLOTS);
773	ret = bkey_err(k);
774	if (ret)
775	goto err;
776	if (k.k->type != KEY_TYPE_stripe) {
777	ret = -ENOENT;
778	goto err;
779	}
780	bkey_reassemble(dst: &stripe->key, src: k);
781	err:
782	bch2_trans_iter_exit(trans, &iter);
783	return ret;
784	}
785
786	/* recovery read path: */
787	int bch2_ec_read_extent(struct btree_trans *trans, struct bch_read_bio *rbio)
788	{
789	struct bch_fs *c = trans->c;
790	struct ec_stripe_buf *buf;
791	struct closure cl;
792	struct bch_stripe *v;
793	unsigned i, offset;
794	int ret = 0;
795
796	closure_init_stack(cl: &cl);
797
798	BUG_ON(!rbio->pick.has_ec);
799
800	buf = kzalloc(size: sizeof(*buf), GFP_NOFS);
801	if (!buf)
802	return -BCH_ERR_ENOMEM_ec_read_extent;
803
804	ret = lockrestart_do(trans, get_stripe_key_trans(trans, rbio->pick.ec.idx, buf));
805	if (ret) {
806	bch_err_ratelimited(c,
807	"error doing reconstruct read: error %i looking up stripe", ret);
808	kfree(objp: buf);
809	return -EIO;
810	}
811
812	v = &bkey_i_to_stripe(k: &buf->key)->v;
813
814	if (!bch2_ptr_matches_stripe(s: v, p: rbio->pick)) {
815	bch_err_ratelimited(c,
816	"error doing reconstruct read: pointer doesn't match stripe");
817	ret = -EIO;
818	goto err;
819	}
820
821	offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
822	if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
823	bch_err_ratelimited(c,
824	"error doing reconstruct read: read is bigger than stripe");
825	ret = -EIO;
826	goto err;
827	}
828
829	ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
830	if (ret)
831	goto err;
832
833	for (i = 0; i < v->nr_blocks; i++)
834	ec_block_io(c, buf, opf: REQ_OP_READ, idx: i, cl: &cl);
835
836	closure_sync(cl: &cl);
837
838	if (ec_nr_failed(buf) > v->nr_redundant) {
839	bch_err_ratelimited(c,
840	"error doing reconstruct read: unable to read enough blocks");
841	ret = -EIO;
842	goto err;
843	}
844
845	ec_validate_checksums(c, buf);
846
847	ret = ec_do_recov(c, buf);
848	if (ret)
849	goto err;
850
851	memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
852	buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
853	err:
854	ec_stripe_buf_exit(buf);
855	kfree(objp: buf);
856	return ret;
857	}
858
859	/* stripe bucket accounting: */
860
861	static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
862	{
863	ec_stripes_heap n, *h = &c->ec_stripes_heap;
864
865	if (idx >= h->size) {
866	if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
867	return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
868
869	mutex_lock(&c->ec_stripes_heap_lock);
870	if (n.size > h->size) {
871	memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
872	n.used = h->used;
873	swap(*h, n);
874	}
875	mutex_unlock(lock: &c->ec_stripes_heap_lock);
876
877	free_heap(&n);
878	}
879
880	if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
881	return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
882
883	if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
884	!genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
885	return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
886
887	return 0;
888	}
889
890	static int ec_stripe_mem_alloc(struct btree_trans *trans,
891	struct btree_iter *iter)
892	{
893	return allocate_dropping_locks_errcode(trans,
894	__ec_stripe_mem_alloc(trans->c, iter->pos.offset, _gfp));
895	}
896
897	/*
898	* Hash table of open stripes:
899	* Stripes that are being created or modified are kept in a hash table, so that
900	* stripe deletion can skip them.
901	*/
902
903	static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
904	{
905	unsigned hash = hash_64(val: idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
906	struct ec_stripe_new *s;
907
908	hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
909	if (s->idx == idx)
910	return true;
911	return false;
912	}
913
914	static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
915	{
916	bool ret = false;
917
918	spin_lock(lock: &c->ec_stripes_new_lock);
919	ret = __bch2_stripe_is_open(c, idx);
920	spin_unlock(lock: &c->ec_stripes_new_lock);
921
922	return ret;
923	}
924
925	static bool bch2_try_open_stripe(struct bch_fs *c,
926	struct ec_stripe_new *s,
927	u64 idx)
928	{
929	bool ret;
930
931	spin_lock(lock: &c->ec_stripes_new_lock);
932	ret = !__bch2_stripe_is_open(c, idx);
933	if (ret) {
934	unsigned hash = hash_64(val: idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
935
936	s->idx = idx;
937	hlist_add_head(n: &s->hash, h: &c->ec_stripes_new[hash]);
938	}
939	spin_unlock(lock: &c->ec_stripes_new_lock);
940
941	return ret;
942	}
943
944	static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
945	{
946	BUG_ON(!s->idx);
947
948	spin_lock(lock: &c->ec_stripes_new_lock);
949	hlist_del_init(n: &s->hash);
950	spin_unlock(lock: &c->ec_stripes_new_lock);
951
952	s->idx = 0;
953	}
954
955	/* Heap of all existing stripes, ordered by blocks_nonempty */
956
957	static u64 stripe_idx_to_delete(struct bch_fs *c)
958	{
959	ec_stripes_heap *h = &c->ec_stripes_heap;
960
961	lockdep_assert_held(&c->ec_stripes_heap_lock);
962
963	if (h->used &&
964	h->data[0].blocks_nonempty == 0 &&
965	!bch2_stripe_is_open(c, idx: h->data[0].idx))
966	return h->data[0].idx;
967
968	return 0;
969	}
970
971	static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
972	struct ec_stripe_heap_entry l,
973	struct ec_stripe_heap_entry r)
974	{
975	return ((l.blocks_nonempty > r.blocks_nonempty) -
976	(l.blocks_nonempty < r.blocks_nonempty));
977	}
978
979	static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
980	size_t i)
981	{
982	struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
983
984	genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
985	}
986
987	static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
988	{
989	ec_stripes_heap *h = &c->ec_stripes_heap;
990	struct stripe *m = genradix_ptr(&c->stripes, idx);
991
992	BUG_ON(m->heap_idx >= h->used);
993	BUG_ON(h->data[m->heap_idx].idx != idx);
994	}
995
996	void bch2_stripes_heap_del(struct bch_fs *c,
997	struct stripe *m, size_t idx)
998	{
999	mutex_lock(&c->ec_stripes_heap_lock);
1000	heap_verify_backpointer(c, idx);
1001
1002	heap_del(&c->ec_stripes_heap, m->heap_idx,
1003	ec_stripes_heap_cmp,
1004	ec_stripes_heap_set_backpointer);
1005	mutex_unlock(lock: &c->ec_stripes_heap_lock);
1006	}
1007
1008	void bch2_stripes_heap_insert(struct bch_fs *c,
1009	struct stripe *m, size_t idx)
1010	{
1011	mutex_lock(&c->ec_stripes_heap_lock);
1012	BUG_ON(heap_full(&c->ec_stripes_heap));
1013
1014	heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
1015	.idx = idx,
1016	.blocks_nonempty = m->blocks_nonempty,
1017	}),
1018	ec_stripes_heap_cmp,
1019	ec_stripes_heap_set_backpointer);
1020
1021	heap_verify_backpointer(c, idx);
1022	mutex_unlock(lock: &c->ec_stripes_heap_lock);
1023	}
1024
1025	void bch2_stripes_heap_update(struct bch_fs *c,
1026	struct stripe *m, size_t idx)
1027	{
1028	ec_stripes_heap *h = &c->ec_stripes_heap;
1029	bool do_deletes;
1030	size_t i;
1031
1032	mutex_lock(&c->ec_stripes_heap_lock);
1033	heap_verify_backpointer(c, idx);
1034
1035	h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
1036
1037	i = m->heap_idx;
1038	heap_sift_up(h, i, ec_stripes_heap_cmp,
1039	ec_stripes_heap_set_backpointer);
1040	heap_sift_down(h, i, ec_stripes_heap_cmp,
1041	ec_stripes_heap_set_backpointer);
1042
1043	heap_verify_backpointer(c, idx);
1044
1045	do_deletes = stripe_idx_to_delete(c) != 0;
1046	mutex_unlock(lock: &c->ec_stripes_heap_lock);
1047
1048	if (do_deletes)
1049	bch2_do_stripe_deletes(c);
1050	}
1051
1052	/* stripe deletion */
1053
1054	static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
1055	{
1056	struct bch_fs *c = trans->c;
1057	struct btree_iter iter;
1058	struct bkey_s_c k;
1059	struct bkey_s_c_stripe s;
1060	int ret;
1061
1062	k = bch2_bkey_get_iter(trans, iter: &iter, btree_id: BTREE_ID_stripes, POS(0, idx),
1063	flags: BTREE_ITER_INTENT);
1064	ret = bkey_err(k);
1065	if (ret)
1066	goto err;
1067
1068	if (k.k->type != KEY_TYPE_stripe) {
1069	bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
1070	ret = -EINVAL;
1071	goto err;
1072	}
1073
1074	s = bkey_s_c_to_stripe(k);
1075	for (unsigned i = 0; i < s.v->nr_blocks; i++)
1076	if (stripe_blockcount_get(s: s.v, idx: i)) {
1077	struct printbuf buf = PRINTBUF;
1078
1079	bch2_bkey_val_to_text(&buf, c, k);
1080	bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
1081	printbuf_exit(&buf);
1082	ret = -EINVAL;
1083	goto err;
1084	}
1085
1086	ret = bch2_btree_delete_at(trans, &iter, 0);
1087	err:
1088	bch2_trans_iter_exit(trans, &iter);
1089	return ret;
1090	}
1091
1092	static void ec_stripe_delete_work(struct work_struct *work)
1093	{
1094	struct bch_fs *c =
1095	container_of(work, struct bch_fs, ec_stripe_delete_work);
1096
1097	while (1) {
1098	mutex_lock(&c->ec_stripes_heap_lock);
1099	u64 idx = stripe_idx_to_delete(c);
1100	mutex_unlock(lock: &c->ec_stripes_heap_lock);
1101
1102	if (!idx)
1103	break;
1104
1105	int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1106	ec_stripe_delete(trans, idx));
1107	bch_err_fn(c, ret);
1108	if (ret)
1109	break;
1110	}
1111
1112	bch2_write_ref_put(c, ref: BCH_WRITE_REF_stripe_delete);
1113	}
1114
1115	void bch2_do_stripe_deletes(struct bch_fs *c)
1116	{
1117	if (bch2_write_ref_tryget(c, ref: BCH_WRITE_REF_stripe_delete) &&
1118	!queue_work(wq: c->write_ref_wq, work: &c->ec_stripe_delete_work))
1119	bch2_write_ref_put(c, ref: BCH_WRITE_REF_stripe_delete);
1120	}
1121
1122	/* stripe creation: */
1123
1124	static int ec_stripe_key_update(struct btree_trans *trans,
1125	struct bkey_i_stripe *new,
1126	bool create)
1127	{
1128	struct bch_fs *c = trans->c;
1129	struct btree_iter iter;
1130	struct bkey_s_c k;
1131	int ret;
1132
1133	k = bch2_bkey_get_iter(trans, iter: &iter, btree_id: BTREE_ID_stripes,
1134	pos: new->k.p, flags: BTREE_ITER_INTENT);
1135	ret = bkey_err(k);
1136	if (ret)
1137	goto err;
1138
1139	if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
1140	bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
1141	create ? "creating" : "updating",
1142	bch2_bkey_types[k.k->type]);
1143	ret = -EINVAL;
1144	goto err;
1145	}
1146
1147	if (k.k->type == KEY_TYPE_stripe) {
1148	const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
1149	unsigned i;
1150
1151	if (old->nr_blocks != new->v.nr_blocks) {
1152	bch_err(c, "error updating stripe: nr_blocks does not match");
1153	ret = -EINVAL;
1154	goto err;
1155	}
1156
1157	for (i = 0; i < new->v.nr_blocks; i++) {
1158	unsigned v = stripe_blockcount_get(s: old, idx: i);
1159
1160	BUG_ON(v &&
1161	(old->ptrs[i].dev != new->v.ptrs[i].dev ||
1162	old->ptrs[i].gen != new->v.ptrs[i].gen ||
1163	old->ptrs[i].offset != new->v.ptrs[i].offset));
1164
1165	stripe_blockcount_set(s: &new->v, idx: i, v);
1166	}
1167	}
1168
1169	ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
1170	err:
1171	bch2_trans_iter_exit(trans, &iter);
1172	return ret;
1173	}
1174
1175	static int ec_stripe_update_extent(struct btree_trans *trans,
1176	struct bpos bucket, u8 gen,
1177	struct ec_stripe_buf *s,
1178	struct bpos *bp_pos)
1179	{
1180	struct bch_stripe *v = &bkey_i_to_stripe(k: &s->key)->v;
1181	struct bch_fs *c = trans->c;
1182	struct bch_backpointer bp;
1183	struct btree_iter iter;
1184	struct bkey_s_c k;
1185	const struct bch_extent_ptr *ptr_c;
1186	struct bch_extent_ptr *ptr, *ec_ptr = NULL;
1187	struct bch_extent_stripe_ptr stripe_ptr;
1188	struct bkey_i *n;
1189	int ret, dev, block;
1190
1191	ret = bch2_get_next_backpointer(trans, bucket, gen,
1192	bp_pos, &bp, BTREE_ITER_CACHED);
1193	if (ret)
1194	return ret;
1195	if (bpos_eq(l: *bp_pos, SPOS_MAX))
1196	return 0;
1197
1198	if (bp.level) {
1199	struct printbuf buf = PRINTBUF;
1200	struct btree_iter node_iter;
1201	struct btree *b;
1202
1203	b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
1204	bch2_trans_iter_exit(trans, &node_iter);
1205
1206	if (!b)
1207	return 0;
1208
1209	prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
1210	bch2_backpointer_to_text(&buf, &bp);
1211
1212	bch2_fs_inconsistent(c, "%s", buf.buf);
1213	printbuf_exit(&buf);
1214	return -EIO;
1215	}
1216
1217	k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT);
1218	ret = bkey_err(k);
1219	if (ret)
1220	return ret;
1221	if (!k.k) {
1222	/*
1223	* extent no longer exists - we could flush the btree
1224	* write buffer and retry to verify, but no need:
1225	*/
1226	return 0;
1227	}
1228
1229	if (extent_has_stripe_ptr(k, idx: s->key.k.p.offset))
1230	goto out;
1231
1232	ptr_c = bkey_matches_stripe(s: v, k, block: &block);
1233	/*
1234	* It doesn't generally make sense to erasure code cached ptrs:
1235	* XXX: should we be incrementing a counter?
1236	*/
1237	if (!ptr_c || ptr_c->cached)
1238	goto out;
1239
1240	dev = v->ptrs[block].dev;
1241
1242	n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
1243	ret = PTR_ERR_OR_ZERO(ptr: n);
1244	if (ret)
1245	goto out;
1246
1247	bkey_reassemble(dst: n, src: k);
1248
1249	bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
1250	ec_ptr = bch2_bkey_has_device(k: bkey_i_to_s(k: n), dev);
1251	BUG_ON(!ec_ptr);
1252
1253	stripe_ptr = (struct bch_extent_stripe_ptr) {
1254	.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
1255	.block = block,
1256	.redundancy = v->nr_redundant,
1257	.idx = s->key.k.p.offset,
1258	};
1259
1260	__extent_entry_insert(k: n,
1261	dst: (union bch_extent_entry *) ec_ptr,
1262	new: (union bch_extent_entry *) &stripe_ptr);
1263
1264	ret = bch2_trans_update(trans, &iter, n, 0);
1265	out:
1266	bch2_trans_iter_exit(trans, &iter);
1267	return ret;
1268	}
1269
1270	static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
1271	unsigned block)
1272	{
1273	struct bch_fs *c = trans->c;
1274	struct bch_stripe *v = &bkey_i_to_stripe(k: &s->key)->v;
1275	struct bch_extent_ptr bucket = v->ptrs[block];
1276	struct bpos bucket_pos = PTR_BUCKET_POS(c, ptr: &bucket);
1277	struct bpos bp_pos = POS_MIN;
1278	int ret = 0;
1279
1280	while (1) {
1281	ret = commit_do(trans, NULL, NULL,
1282	BCH_TRANS_COMMIT_no_check_rw|
1283	BCH_TRANS_COMMIT_no_enospc,
1284	ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
1285	s, &bp_pos));
1286	if (ret)
1287	break;
1288	if (bkey_eq(l: bp_pos, POS_MAX))
1289	break;
1290
1291	bp_pos = bpos_nosnap_successor(p: bp_pos);
1292	}
1293
1294	return ret;
1295	}
1296
1297	static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
1298	{
1299	struct btree_trans *trans = bch2_trans_get(c);
1300	struct bch_stripe *v = &bkey_i_to_stripe(k: &s->key)->v;
1301	unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1302	int ret = 0;
1303
1304	ret = bch2_btree_write_buffer_flush_sync(trans);
1305	if (ret)
1306	goto err;
1307
1308	for (i = 0; i < nr_data; i++) {
1309	ret = ec_stripe_update_bucket(trans, s, block: i);
1310	if (ret)
1311	break;
1312	}
1313	err:
1314	bch2_trans_put(trans);
1315
1316	return ret;
1317	}
1318
1319	static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1320	struct ec_stripe_new *s,
1321	unsigned block,
1322	struct open_bucket *ob)
1323	{
1324	struct bch_dev *ca = bch_dev_bkey_exists(c, idx: ob->dev);
1325	unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1326	int ret;
1327
1328	if (!bch2_dev_get_ioref(ca, WRITE)) {
1329	s->err = -BCH_ERR_erofs_no_writes;
1330	return;
1331	}
1332
1333	memset(s->new_stripe.data[block] + (offset << 9),
1334	0,
1335	ob->sectors_free << 9);
1336
1337	ret = blkdev_issue_zeroout(bdev: ca->disk_sb.bdev,
1338	sector: ob->bucket * ca->mi.bucket_size + offset,
1339	nr_sects: ob->sectors_free,
1340	GFP_KERNEL, flags: 0);
1341
1342	percpu_ref_put(ref: &ca->io_ref);
1343
1344	if (ret)
1345	s->err = ret;
1346	}
1347
1348	void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s)
1349	{
1350	if (s->idx)
1351	bch2_stripe_close(c, s);
1352	kfree(objp: s);
1353	}
1354
1355	/*
1356	* data buckets of new stripe all written: create the stripe
1357	*/
1358	static void ec_stripe_create(struct ec_stripe_new *s)
1359	{
1360	struct bch_fs *c = s->c;
1361	struct open_bucket *ob;
1362	struct bch_stripe *v = &bkey_i_to_stripe(k: &s->new_stripe.key)->v;
1363	unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1364	int ret;
1365
1366	BUG_ON(s->h->s == s);
1367
1368	closure_sync(cl: &s->iodone);
1369
1370	if (!s->err) {
1371	for (i = 0; i < nr_data; i++)
1372	if (s->blocks[i]) {
1373	ob = c->open_buckets + s->blocks[i];
1374
1375	if (ob->sectors_free)
1376	zero_out_rest_of_ec_bucket(c, s, block: i, ob);
1377	}
1378	}
1379
1380	if (s->err) {
1381	if (!bch2_err_matches(s->err, EROFS))
1382	bch_err(c, "error creating stripe: error writing data buckets");
1383	goto err;
1384	}
1385
1386	if (s->have_existing_stripe) {
1387	ec_validate_checksums(c, buf: &s->existing_stripe);
1388
1389	if (ec_do_recov(c, buf: &s->existing_stripe)) {
1390	bch_err(c, "error creating stripe: error reading existing stripe");
1391	goto err;
1392	}
1393
1394	for (i = 0; i < nr_data; i++)
1395	if (stripe_blockcount_get(s: &bkey_i_to_stripe(k: &s->existing_stripe.key)->v, idx: i))
1396	swap(s->new_stripe.data[i],
1397	s->existing_stripe.data[i]);
1398
1399	ec_stripe_buf_exit(buf: &s->existing_stripe);
1400	}
1401
1402	BUG_ON(!s->allocated);
1403	BUG_ON(!s->idx);
1404
1405	ec_generate_ec(buf: &s->new_stripe);
1406
1407	ec_generate_checksums(buf: &s->new_stripe);
1408
1409	/* write p/q: */
1410	for (i = nr_data; i < v->nr_blocks; i++)
1411	ec_block_io(c, buf: &s->new_stripe, opf: REQ_OP_WRITE, idx: i, cl: &s->iodone);
1412	closure_sync(cl: &s->iodone);
1413
1414	if (ec_nr_failed(buf: &s->new_stripe)) {
1415	bch_err(c, "error creating stripe: error writing redundancy buckets");
1416	goto err;
1417	}
1418
1419	ret = bch2_trans_do(c, &s->res, NULL,
1420	BCH_TRANS_COMMIT_no_check_rw|
1421	BCH_TRANS_COMMIT_no_enospc,
1422	ec_stripe_key_update(trans,
1423	bkey_i_to_stripe(&s->new_stripe.key),
1424	!s->have_existing_stripe));
1425	bch_err_msg(c, ret, "creating stripe key");
1426	if (ret) {
1427	goto err;
1428	}
1429
1430	ret = ec_stripe_update_extents(c, s: &s->new_stripe);
1431	bch_err_msg(c, ret, "error updating extents");
1432	if (ret)
1433	goto err;
1434	err:
1435	bch2_disk_reservation_put(c, res: &s->res);
1436
1437	for (i = 0; i < v->nr_blocks; i++)
1438	if (s->blocks[i]) {
1439	ob = c->open_buckets + s->blocks[i];
1440
1441	if (i < nr_data) {
1442	ob->ec = NULL;
1443	__bch2_open_bucket_put(c, ob);
1444	} else {
1445	bch2_open_bucket_put(c, ob);
1446	}
1447	}
1448
1449	mutex_lock(&c->ec_stripe_new_lock);
1450	list_del(entry: &s->list);
1451	mutex_unlock(lock: &c->ec_stripe_new_lock);
1452	wake_up(&c->ec_stripe_new_wait);
1453
1454	ec_stripe_buf_exit(buf: &s->existing_stripe);
1455	ec_stripe_buf_exit(buf: &s->new_stripe);
1456	closure_debug_destroy(cl: &s->iodone);
1457
1458	ec_stripe_new_put(c, s, ref: STRIPE_REF_stripe);
1459	}
1460
1461	static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1462	{
1463	struct ec_stripe_new *s;
1464
1465	mutex_lock(&c->ec_stripe_new_lock);
1466	list_for_each_entry(s, &c->ec_stripe_new_list, list)
1467	if (!atomic_read(v: &s->ref[STRIPE_REF_io]))
1468	goto out;
1469	s = NULL;
1470	out:
1471	mutex_unlock(lock: &c->ec_stripe_new_lock);
1472
1473	return s;
1474	}
1475
1476	static void ec_stripe_create_work(struct work_struct *work)
1477	{
1478	struct bch_fs *c = container_of(work,
1479	struct bch_fs, ec_stripe_create_work);
1480	struct ec_stripe_new *s;
1481
1482	while ((s = get_pending_stripe(c)))
1483	ec_stripe_create(s);
1484
1485	bch2_write_ref_put(c, ref: BCH_WRITE_REF_stripe_create);
1486	}
1487
1488	void bch2_ec_do_stripe_creates(struct bch_fs *c)
1489	{
1490	bch2_write_ref_get(c, ref: BCH_WRITE_REF_stripe_create);
1491
1492	if (!queue_work(wq: system_long_wq, work: &c->ec_stripe_create_work))
1493	bch2_write_ref_put(c, ref: BCH_WRITE_REF_stripe_create);
1494	}
1495
1496	static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1497	{
1498	struct ec_stripe_new *s = h->s;
1499
1500	BUG_ON(!s->allocated && !s->err);
1501
1502	h->s = NULL;
1503	s->pending = true;
1504
1505	mutex_lock(&c->ec_stripe_new_lock);
1506	list_add(new: &s->list, head: &c->ec_stripe_new_list);
1507	mutex_unlock(lock: &c->ec_stripe_new_lock);
1508
1509	ec_stripe_new_put(c, s, ref: STRIPE_REF_io);
1510	}
1511
1512	void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1513	{
1514	struct ec_stripe_new *s = ob->ec;
1515
1516	s->err = -EIO;
1517	}
1518
1519	void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1520	{
1521	struct open_bucket *ob = ec_open_bucket(c, obs: &wp->ptrs);
1522	struct bch_dev *ca;
1523	unsigned offset;
1524
1525	if (!ob)
1526	return NULL;
1527
1528	BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
1529
1530	ca = bch_dev_bkey_exists(c, idx: ob->dev);
1531	offset = ca->mi.bucket_size - ob->sectors_free;
1532
1533	return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1534	}
1535
1536	static int unsigned_cmp(const void *_l, const void *_r)
1537	{
1538	unsigned l = *((const unsigned *) _l);
1539	unsigned r = *((const unsigned *) _r);
1540
1541	return cmp_int(l, r);
1542	}
1543
1544	/* pick most common bucket size: */
1545	static unsigned pick_blocksize(struct bch_fs *c,
1546	struct bch_devs_mask *devs)
1547	{
1548	unsigned nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1549	struct {
1550	unsigned nr, size;
1551	} cur = { 0, 0 }, best = { 0, 0 };
1552
1553	for_each_member_device_rcu(c, ca, devs)
1554	sizes[nr++] = ca->mi.bucket_size;
1555
1556	sort(base: sizes, num: nr, size: sizeof(unsigned), cmp_func: unsigned_cmp, NULL);
1557
1558	for (unsigned i = 0; i < nr; i++) {
1559	if (sizes[i] != cur.size) {
1560	if (cur.nr > best.nr)
1561	best = cur;
1562
1563	cur.nr = 0;
1564	cur.size = sizes[i];
1565	}
1566
1567	cur.nr++;
1568	}
1569
1570	if (cur.nr > best.nr)
1571	best = cur;
1572
1573	return best.size;
1574	}
1575
1576	static bool may_create_new_stripe(struct bch_fs *c)
1577	{
1578	return false;
1579	}
1580
1581	static void ec_stripe_key_init(struct bch_fs *c,
1582	struct bkey_i *k,
1583	unsigned nr_data,
1584	unsigned nr_parity,
1585	unsigned stripe_size)
1586	{
1587	struct bkey_i_stripe *s = bkey_stripe_init(k: k);
1588	unsigned u64s;
1589
1590	s->v.sectors = cpu_to_le16(stripe_size);
1591	s->v.algorithm = 0;
1592	s->v.nr_blocks = nr_data + nr_parity;
1593	s->v.nr_redundant = nr_parity;
1594	s->v.csum_granularity_bits = ilog2(c->opts.encoded_extent_max >> 9);
1595	s->v.csum_type = BCH_CSUM_crc32c;
1596	s->v.pad = 0;
1597
1598	while ((u64s = stripe_val_u64s(s: &s->v)) > BKEY_VAL_U64s_MAX) {
1599	BUG_ON(1 << s->v.csum_granularity_bits >=
1600	le16_to_cpu(s->v.sectors) ||
1601	s->v.csum_granularity_bits == U8_MAX);
1602	s->v.csum_granularity_bits++;
1603	}
1604
1605	set_bkey_val_u64s(k: &s->k, val_u64s: u64s);
1606	}
1607
1608	static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1609	{
1610	struct ec_stripe_new *s;
1611
1612	lockdep_assert_held(&h->lock);
1613
1614	s = kzalloc(size: sizeof(*s), GFP_KERNEL);
1615	if (!s)
1616	return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
1617
1618	mutex_init(&s->lock);
1619	closure_init(cl: &s->iodone, NULL);
1620	atomic_set(v: &s->ref[STRIPE_REF_stripe], i: 1);
1621	atomic_set(v: &s->ref[STRIPE_REF_io], i: 1);
1622	s->c = c;
1623	s->h = h;
1624	s->nr_data = min_t(unsigned, h->nr_active_devs,
1625	BCH_BKEY_PTRS_MAX) - h->redundancy;
1626	s->nr_parity = h->redundancy;
1627
1628	ec_stripe_key_init(c, k: &s->new_stripe.key,
1629	nr_data: s->nr_data, nr_parity: s->nr_parity, stripe_size: h->blocksize);
1630
1631	h->s = s;
1632	return 0;
1633	}
1634
1635	static struct ec_stripe_head *
1636	ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1637	unsigned algo, unsigned redundancy,
1638	enum bch_watermark watermark)
1639	{
1640	struct ec_stripe_head *h;
1641
1642	h = kzalloc(size: sizeof(*h), GFP_KERNEL);
1643	if (!h)
1644	return NULL;
1645
1646	mutex_init(&h->lock);
1647	BUG_ON(!mutex_trylock(&h->lock));
1648
1649	h->target = target;
1650	h->algo = algo;
1651	h->redundancy = redundancy;
1652	h->watermark = watermark;
1653
1654	rcu_read_lock();
1655	h->devs = target_rw_devs(c, data_type: BCH_DATA_user, target);
1656
1657	for_each_member_device_rcu(c, ca, &h->devs)
1658	if (!ca->mi.durability)
1659	__clear_bit(ca->dev_idx, h->devs.d);
1660
1661	h->blocksize = pick_blocksize(c, devs: &h->devs);
1662
1663	for_each_member_device_rcu(c, ca, &h->devs)
1664	if (ca->mi.bucket_size == h->blocksize)
1665	h->nr_active_devs++;
1666
1667	rcu_read_unlock();
1668
1669	/*
1670	* If we only have redundancy + 1 devices, we're better off with just
1671	* replication:
1672	*/
1673	if (h->nr_active_devs < h->redundancy + 2)
1674	bch_err(c, "insufficient devices available to create stripe (have %u, need %u) - mismatched bucket sizes?",
1675	h->nr_active_devs, h->redundancy + 2);
1676
1677	list_add(new: &h->list, head: &c->ec_stripe_head_list);
1678	return h;
1679	}
1680
1681	void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1682	{
1683	if (h->s &&
1684	h->s->allocated &&
1685	bitmap_weight(src: h->s->blocks_allocated,
1686	nbits: h->s->nr_data) == h->s->nr_data)
1687	ec_stripe_set_pending(c, h);
1688
1689	mutex_unlock(lock: &h->lock);
1690	}
1691
1692	static struct ec_stripe_head *
1693	__bch2_ec_stripe_head_get(struct btree_trans *trans,
1694	unsigned target,
1695	unsigned algo,
1696	unsigned redundancy,
1697	enum bch_watermark watermark)
1698	{
1699	struct bch_fs *c = trans->c;
1700	struct ec_stripe_head *h;
1701	int ret;
1702
1703	if (!redundancy)
1704	return NULL;
1705
1706	ret = bch2_trans_mutex_lock(trans, lock: &c->ec_stripe_head_lock);
1707	if (ret)
1708	return ERR_PTR(error: ret);
1709
1710	if (test_bit(BCH_FS_going_ro, &c->flags)) {
1711	h = ERR_PTR(error: -BCH_ERR_erofs_no_writes);
1712	goto found;
1713	}
1714
1715	list_for_each_entry(h, &c->ec_stripe_head_list, list)
1716	if (h->target == target &&
1717	h->algo == algo &&
1718	h->redundancy == redundancy &&
1719	h->watermark == watermark) {
1720	ret = bch2_trans_mutex_lock(trans, lock: &h->lock);
1721	if (ret)
1722	h = ERR_PTR(error: ret);
1723	goto found;
1724	}
1725
1726	h = ec_new_stripe_head_alloc(c, target, algo, redundancy, watermark);
1727	found:
1728	if (!IS_ERR_OR_NULL(ptr: h) &&
1729	h->nr_active_devs < h->redundancy + 2) {
1730	mutex_unlock(lock: &h->lock);
1731	h = NULL;
1732	}
1733	mutex_unlock(lock: &c->ec_stripe_head_lock);
1734	return h;
1735	}
1736
1737	static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1738	enum bch_watermark watermark, struct closure *cl)
1739	{
1740	struct bch_fs *c = trans->c;
1741	struct bch_devs_mask devs = h->devs;
1742	struct open_bucket *ob;
1743	struct open_buckets buckets;
1744	struct bch_stripe *v = &bkey_i_to_stripe(k: &h->s->new_stripe.key)->v;
1745	unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1746	bool have_cache = true;
1747	int ret = 0;
1748
1749	BUG_ON(v->nr_blocks != h->s->nr_data + h->s->nr_parity);
1750	BUG_ON(v->nr_redundant != h->s->nr_parity);
1751
1752	for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) {
1753	__clear_bit(v->ptrs[i].dev, devs.d);
1754	if (i < h->s->nr_data)
1755	nr_have_data++;
1756	else
1757	nr_have_parity++;
1758	}
1759
1760	BUG_ON(nr_have_data > h->s->nr_data);
1761	BUG_ON(nr_have_parity > h->s->nr_parity);
1762
1763	buckets.nr = 0;
1764	if (nr_have_parity < h->s->nr_parity) {
1765	ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1766	&h->parity_stripe,
1767	&devs,
1768	h->s->nr_parity,
1769	&nr_have_parity,
1770	&have_cache, 0,
1771	BCH_DATA_parity,
1772	watermark,
1773	cl);
1774
1775	open_bucket_for_each(c, &buckets, ob, i) {
1776	j = find_next_zero_bit(addr: h->s->blocks_gotten,
1777	size: h->s->nr_data + h->s->nr_parity,
1778	offset: h->s->nr_data);
1779	BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1780
1781	h->s->blocks[j] = buckets.v[i];
1782	v->ptrs[j] = bch2_ob_ptr(c, ob);
1783	__set_bit(j, h->s->blocks_gotten);
1784	}
1785
1786	if (ret)
1787	return ret;
1788	}
1789
1790	buckets.nr = 0;
1791	if (nr_have_data < h->s->nr_data) {
1792	ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1793	&h->block_stripe,
1794	&devs,
1795	h->s->nr_data,
1796	&nr_have_data,
1797	&have_cache, 0,
1798	BCH_DATA_user,
1799	watermark,
1800	cl);
1801
1802	open_bucket_for_each(c, &buckets, ob, i) {
1803	j = find_next_zero_bit(addr: h->s->blocks_gotten,
1804	size: h->s->nr_data, offset: 0);
1805	BUG_ON(j >= h->s->nr_data);
1806
1807	h->s->blocks[j] = buckets.v[i];
1808	v->ptrs[j] = bch2_ob_ptr(c, ob);
1809	__set_bit(j, h->s->blocks_gotten);
1810	}
1811
1812	if (ret)
1813	return ret;
1814	}
1815
1816	return 0;
1817	}
1818
1819	/* XXX: doesn't obey target: */
1820	static s64 get_existing_stripe(struct bch_fs *c,
1821	struct ec_stripe_head *head)
1822	{
1823	ec_stripes_heap *h = &c->ec_stripes_heap;
1824	struct stripe *m;
1825	size_t heap_idx;
1826	u64 stripe_idx;
1827	s64 ret = -1;
1828
1829	if (may_create_new_stripe(c))
1830	return -1;
1831
1832	mutex_lock(&c->ec_stripes_heap_lock);
1833	for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1834	/* No blocks worth reusing, stripe will just be deleted: */
1835	if (!h->data[heap_idx].blocks_nonempty)
1836	continue;
1837
1838	stripe_idx = h->data[heap_idx].idx;
1839
1840	m = genradix_ptr(&c->stripes, stripe_idx);
1841
1842	if (m->algorithm == head->algo &&
1843	m->nr_redundant == head->redundancy &&
1844	m->sectors == head->blocksize &&
1845	m->blocks_nonempty < m->nr_blocks - m->nr_redundant &&
1846	bch2_try_open_stripe(c, s: head->s, idx: stripe_idx)) {
1847	ret = stripe_idx;
1848	break;
1849	}
1850	}
1851	mutex_unlock(lock: &c->ec_stripes_heap_lock);
1852	return ret;
1853	}
1854
1855	static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1856	{
1857	struct bch_fs *c = trans->c;
1858	struct bch_stripe *new_v = &bkey_i_to_stripe(k: &h->s->new_stripe.key)->v;
1859	struct bch_stripe *existing_v;
1860	unsigned i;
1861	s64 idx;
1862	int ret;
1863
1864	/*
1865	* If we can't allocate a new stripe, and there's no stripes with empty
1866	* blocks for us to reuse, that means we have to wait on copygc:
1867	*/
1868	idx = get_existing_stripe(c, head: h);
1869	if (idx < 0)
1870	return -BCH_ERR_stripe_alloc_blocked;
1871
1872	ret = get_stripe_key_trans(trans, idx, stripe: &h->s->existing_stripe);
1873	bch2_fs_fatal_err_on(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart), c,
1874	"reading stripe key: %s", bch2_err_str(ret));
1875	if (ret) {
1876	bch2_stripe_close(c, s: h->s);
1877	return ret;
1878	}
1879
1880	existing_v = &bkey_i_to_stripe(k: &h->s->existing_stripe.key)->v;
1881
1882	BUG_ON(existing_v->nr_redundant != h->s->nr_parity);
1883	h->s->nr_data = existing_v->nr_blocks -
1884	existing_v->nr_redundant;
1885
1886	ret = ec_stripe_buf_init(buf: &h->s->existing_stripe, offset: 0, size: h->blocksize);
1887	if (ret) {
1888	bch2_stripe_close(c, s: h->s);
1889	return ret;
1890	}
1891
1892	BUG_ON(h->s->existing_stripe.size != h->blocksize);
1893	BUG_ON(h->s->existing_stripe.size != le16_to_cpu(existing_v->sectors));
1894
1895	/*
1896	* Free buckets we initially allocated - they might conflict with
1897	* blocks from the stripe we're reusing:
1898	*/
1899	for_each_set_bit(i, h->s->blocks_gotten, new_v->nr_blocks) {
1900	bch2_open_bucket_put(c, ob: c->open_buckets + h->s->blocks[i]);
1901	h->s->blocks[i] = 0;
1902	}
1903	memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
1904	memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
1905
1906	for (i = 0; i < existing_v->nr_blocks; i++) {
1907	if (stripe_blockcount_get(s: existing_v, idx: i)) {
1908	__set_bit(i, h->s->blocks_gotten);
1909	__set_bit(i, h->s->blocks_allocated);
1910	}
1911
1912	ec_block_io(c, buf: &h->s->existing_stripe, READ, idx: i, cl: &h->s->iodone);
1913	}
1914
1915	bkey_copy(dst: &h->s->new_stripe.key, src: &h->s->existing_stripe.key);
1916	h->s->have_existing_stripe = true;
1917
1918	return 0;
1919	}
1920
1921	static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1922	{
1923	struct bch_fs *c = trans->c;
1924	struct btree_iter iter;
1925	struct bkey_s_c k;
1926	struct bpos min_pos = POS(0, 1);
1927	struct bpos start_pos = bpos_max(l: min_pos, POS(0, c->ec_stripe_hint));
1928	int ret;
1929
1930	if (!h->s->res.sectors) {
1931	ret = bch2_disk_reservation_get(c, res: &h->s->res,
1932	sectors: h->blocksize,
1933	nr_replicas: h->s->nr_parity,
1934	BCH_DISK_RESERVATION_NOFAIL);
1935	if (ret)
1936	return ret;
1937	}
1938
1939	for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
1940	BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
1941	if (bkey_gt(l: k.k->p, POS(0, U32_MAX))) {
1942	if (start_pos.offset) {
1943	start_pos = min_pos;
1944	bch2_btree_iter_set_pos(iter: &iter, new_pos: start_pos);
1945	continue;
1946	}
1947
1948	ret = -BCH_ERR_ENOSPC_stripe_create;
1949	break;
1950	}
1951
1952	if (bkey_deleted(k.k) &&
1953	bch2_try_open_stripe(c, s: h->s, idx: k.k->p.offset))
1954	break;
1955	}
1956
1957	c->ec_stripe_hint = iter.pos.offset;
1958
1959	if (ret)
1960	goto err;
1961
1962	ret = ec_stripe_mem_alloc(trans, iter: &iter);
1963	if (ret) {
1964	bch2_stripe_close(c, s: h->s);
1965	goto err;
1966	}
1967
1968	h->s->new_stripe.key.k.p = iter.pos;
1969	out:
1970	bch2_trans_iter_exit(trans, &iter);
1971	return ret;
1972	err:
1973	bch2_disk_reservation_put(c, res: &h->s->res);
1974	goto out;
1975	}
1976
1977	struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
1978	unsigned target,
1979	unsigned algo,
1980	unsigned redundancy,
1981	enum bch_watermark watermark,
1982	struct closure *cl)
1983	{
1984	struct bch_fs *c = trans->c;
1985	struct ec_stripe_head *h;
1986	bool waiting = false;
1987	int ret;
1988
1989	h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, watermark);
1990	if (IS_ERR_OR_NULL(ptr: h))
1991	return h;
1992
1993	if (!h->s) {
1994	ret = ec_new_stripe_alloc(c, h);
1995	if (ret) {
1996	bch_err(c, "failed to allocate new stripe");
1997	goto err;
1998	}
1999	}
2000
2001	if (h->s->allocated)
2002	goto allocated;
2003
2004	if (h->s->have_existing_stripe)
2005	goto alloc_existing;
2006
2007	/* First, try to allocate a full stripe: */
2008	ret = new_stripe_alloc_buckets(trans, h, watermark: BCH_WATERMARK_stripe, NULL) ?:
2009	__bch2_ec_stripe_head_reserve(trans, h);
2010	if (!ret)
2011	goto allocate_buf;
2012	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
2013	bch2_err_matches(ret, ENOMEM))
2014	goto err;
2015
2016	/*
2017	* Not enough buckets available for a full stripe: we must reuse an
2018	* existing stripe:
2019	*/
2020	while (1) {
2021	ret = __bch2_ec_stripe_head_reuse(trans, h);
2022	if (!ret)
2023	break;
2024	if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
2025	goto err;
2026
2027	if (watermark == BCH_WATERMARK_copygc) {
2028	ret = new_stripe_alloc_buckets(trans, h, watermark, NULL) ?:
2029	__bch2_ec_stripe_head_reserve(trans, h);
2030	if (ret)
2031	goto err;
2032	goto allocate_buf;
2033	}
2034
2035	/* XXX freelist_wait? */
2036	closure_wait(list: &c->freelist_wait, cl);
2037	waiting = true;
2038	}
2039
2040	if (waiting)
2041	closure_wake_up(list: &c->freelist_wait);
2042	alloc_existing:
2043	/*
2044	* Retry allocating buckets, with the watermark for this
2045	* particular write:
2046	*/
2047	ret = new_stripe_alloc_buckets(trans, h, watermark, cl);
2048	if (ret)
2049	goto err;
2050
2051	allocate_buf:
2052	ret = ec_stripe_buf_init(buf: &h->s->new_stripe, offset: 0, size: h->blocksize);
2053	if (ret)
2054	goto err;
2055
2056	h->s->allocated = true;
2057	allocated:
2058	BUG_ON(!h->s->idx);
2059	BUG_ON(!h->s->new_stripe.data[0]);
2060	BUG_ON(trans->restarted);
2061	return h;
2062	err:
2063	bch2_ec_stripe_head_put(c, h);
2064	return ERR_PTR(error: ret);
2065	}
2066
2067	static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
2068	{
2069	struct ec_stripe_head *h;
2070	struct open_bucket *ob;
2071	unsigned i;
2072
2073	mutex_lock(&c->ec_stripe_head_lock);
2074	list_for_each_entry(h, &c->ec_stripe_head_list, list) {
2075	mutex_lock(&h->lock);
2076	if (!h->s)
2077	goto unlock;
2078
2079	if (!ca)
2080	goto found;
2081
2082	for (i = 0; i < bkey_i_to_stripe(k: &h->s->new_stripe.key)->v.nr_blocks; i++) {
2083	if (!h->s->blocks[i])
2084	continue;
2085
2086	ob = c->open_buckets + h->s->blocks[i];
2087	if (ob->dev == ca->dev_idx)
2088	goto found;
2089	}
2090	goto unlock;
2091	found:
2092	h->s->err = -BCH_ERR_erofs_no_writes;
2093	ec_stripe_set_pending(c, h);
2094	unlock:
2095	mutex_unlock(lock: &h->lock);
2096	}
2097	mutex_unlock(lock: &c->ec_stripe_head_lock);
2098	}
2099
2100	void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
2101	{
2102	__bch2_ec_stop(c, ca);
2103	}
2104
2105	void bch2_fs_ec_stop(struct bch_fs *c)
2106	{
2107	__bch2_ec_stop(c, NULL);
2108	}
2109
2110	static bool bch2_fs_ec_flush_done(struct bch_fs *c)
2111	{
2112	bool ret;
2113
2114	mutex_lock(&c->ec_stripe_new_lock);
2115	ret = list_empty(head: &c->ec_stripe_new_list);
2116	mutex_unlock(lock: &c->ec_stripe_new_lock);
2117
2118	return ret;
2119	}
2120
2121	void bch2_fs_ec_flush(struct bch_fs *c)
2122	{
2123	wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c));
2124	}
2125
2126	int bch2_stripes_read(struct bch_fs *c)
2127	{
2128	int ret = bch2_trans_run(c,
2129	for_each_btree_key(trans, iter, BTREE_ID_stripes, POS_MIN,
2130	BTREE_ITER_PREFETCH, k, ({
2131	if (k.k->type != KEY_TYPE_stripe)
2132	continue;
2133
2134	ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
2135	if (ret)
2136	break;
2137
2138	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
2139
2140	struct stripe *m = genradix_ptr(&c->stripes, k.k->p.offset);
2141	m->sectors = le16_to_cpu(s->sectors);
2142	m->algorithm = s->algorithm;
2143	m->nr_blocks = s->nr_blocks;
2144	m->nr_redundant = s->nr_redundant;
2145	m->blocks_nonempty = 0;
2146
2147	for (unsigned i = 0; i < s->nr_blocks; i++)
2148	m->blocks_nonempty += !!stripe_blockcount_get(s, i);
2149
2150	bch2_stripes_heap_insert(c, m, k.k->p.offset);
2151	0;
2152	})));
2153	bch_err_fn(c, ret);
2154	return ret;
2155	}
2156
2157	void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
2158	{
2159	ec_stripes_heap *h = &c->ec_stripes_heap;
2160	struct stripe *m;
2161	size_t i;
2162
2163	mutex_lock(&c->ec_stripes_heap_lock);
2164	for (i = 0; i < min_t(size_t, h->used, 50); i++) {
2165	m = genradix_ptr(&c->stripes, h->data[i].idx);
2166
2167	prt_printf(out, "%zu %u/%u+%u", h->data[i].idx,
2168	h->data[i].blocks_nonempty,
2169	m->nr_blocks - m->nr_redundant,
2170	m->nr_redundant);
2171	if (bch2_stripe_is_open(c, idx: h->data[i].idx))
2172	prt_str(out, str: " open");
2173	prt_newline(out);
2174	}
2175	mutex_unlock(lock: &c->ec_stripes_heap_lock);
2176	}
2177
2178	void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
2179	{
2180	struct ec_stripe_head *h;
2181	struct ec_stripe_new *s;
2182
2183	mutex_lock(&c->ec_stripe_head_lock);
2184	list_for_each_entry(h, &c->ec_stripe_head_list, list) {
2185	prt_printf(out, "target %u algo %u redundancy %u %s:\n",
2186	h->target, h->algo, h->redundancy,
2187	bch2_watermarks[h->watermark]);
2188
2189	if (h->s)
2190	prt_printf(out, "\tidx %llu blocks %u+%u allocated %u\n",
2191	h->s->idx, h->s->nr_data, h->s->nr_parity,
2192	bitmap_weight(h->s->blocks_allocated,
2193	h->s->nr_data));
2194	}
2195	mutex_unlock(lock: &c->ec_stripe_head_lock);
2196
2197	prt_printf(out, "in flight:\n");
2198
2199	mutex_lock(&c->ec_stripe_new_lock);
2200	list_for_each_entry(s, &c->ec_stripe_new_list, list) {
2201	prt_printf(out, "\tidx %llu blocks %u+%u ref %u %u %s\n",
2202	s->idx, s->nr_data, s->nr_parity,
2203	atomic_read(&s->ref[STRIPE_REF_io]),
2204	atomic_read(&s->ref[STRIPE_REF_stripe]),
2205	bch2_watermarks[s->h->watermark]);
2206	}
2207	mutex_unlock(lock: &c->ec_stripe_new_lock);
2208	}
2209
2210	void bch2_fs_ec_exit(struct bch_fs *c)
2211	{
2212	struct ec_stripe_head *h;
2213	unsigned i;
2214
2215	while (1) {
2216	mutex_lock(&c->ec_stripe_head_lock);
2217	h = list_first_entry_or_null(&c->ec_stripe_head_list,
2218	struct ec_stripe_head, list);
2219	if (h)
2220	list_del(entry: &h->list);
2221	mutex_unlock(lock: &c->ec_stripe_head_lock);
2222	if (!h)
2223	break;
2224
2225	if (h->s) {
2226	for (i = 0; i < bkey_i_to_stripe(k: &h->s->new_stripe.key)->v.nr_blocks; i++)
2227	BUG_ON(h->s->blocks[i]);
2228
2229	kfree(objp: h->s);
2230	}
2231	kfree(objp: h);
2232	}
2233
2234	BUG_ON(!list_empty(&c->ec_stripe_new_list));
2235
2236	free_heap(&c->ec_stripes_heap);
2237	genradix_free(&c->stripes);
2238	bioset_exit(&c->ec_bioset);
2239	}
2240
2241	void bch2_fs_ec_init_early(struct bch_fs *c)
2242	{
2243	spin_lock_init(&c->ec_stripes_new_lock);
2244	mutex_init(&c->ec_stripes_heap_lock);
2245
2246	INIT_LIST_HEAD(list: &c->ec_stripe_head_list);
2247	mutex_init(&c->ec_stripe_head_lock);
2248
2249	INIT_LIST_HEAD(list: &c->ec_stripe_new_list);
2250	mutex_init(&c->ec_stripe_new_lock);
2251	init_waitqueue_head(&c->ec_stripe_new_wait);
2252
2253	INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
2254	INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
2255	}
2256
2257	int bch2_fs_ec_init(struct bch_fs *c)
2258	{
2259	return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
2260	flags: BIOSET_NEED_BVECS);
2261	}
2262