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_accounting.h"
17 #include "disk_groups.h"
18 #include "ec.h"
19 #include "error.h"
20 #include "io_read.h"
21 #include "keylist.h"
22 #include "recovery.h"
23 #include "replicas.h"
24 #include "super-io.h"
25 #include "util.h"
26
27 #include <linux/sort.h>
28
29 #ifdef __KERNEL__
30
31 #include <linux/raid/pq.h>
32 #include <linux/raid/xor.h>
33
34 static void raid5_recov(unsigned disks, unsigned failed_idx,
35 size_t size, void **data)
36 {
37 unsigned i = 2, nr;
38
39 BUG_ON(failed_idx >= disks);
40
41 swap(data[0], data[failed_idx]);
42 memcpy(data[0], data[1], size);
43
44 while (i < disks) {
45 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
46 xor_blocks(nr, size, data[0], data + i);
47 i += nr;
48 }
49
50 swap(data[0], data[failed_idx]);
51 }
52
53 static void raid_gen(int nd, int np, size_t size, void **v)
54 {
55 if (np >= 1)
56 raid5_recov(nd + np, nd, size, v);
57 if (np >= 2)
58 raid6_call.gen_syndrome(nd + np, size, v);
59 BUG_ON(np > 2);
60 }
61
62 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
63 {
64 switch (nr) {
65 case 0:
66 break;
67 case 1:
68 if (ir[0] < nd + 1)
69 raid5_recov(nd + 1, ir[0], size, v);
70 else
71 raid6_call.gen_syndrome(nd + np, size, v);
72 break;
73 case 2:
74 if (ir[1] < nd) {
75 /* data+data failure. */
76 raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
77 } else if (ir[0] < nd) {
78 /* data + p/q failure */
79
80 if (ir[1] == nd) /* data + p failure */
81 raid6_datap_recov(nd + np, size, ir[0], v);
82 else { /* data + q failure */
83 raid5_recov(nd + 1, ir[0], size, v);
84 raid6_call.gen_syndrome(nd + np, size, v);
85 }
86 } else {
87 raid_gen(nd, np, size, v);
88 }
89 break;
90 default:
91 BUG();
92 }
93 }
94
95 #else
96
97 #include <raid/raid.h>
98
99 #endif
100
101 struct ec_bio {
102 struct bch_dev *ca;
103 struct ec_stripe_buf *buf;
104 size_t idx;
105 struct bio bio;
106 };
107
108 /* Stripes btree keys: */
109
110 int bch2_stripe_validate(struct bch_fs *c, struct bkey_s_c k,
111 enum bch_validate_flags flags)
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)),
118 c, stripe_pos_bad,
119 "stripe at bad pos");
120
121 bkey_fsck_err_on(bkey_val_u64s(k.k) < stripe_val_u64s(s),
122 c, 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_validate(c, k, flags);
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, i) < bkey_val_bytes(k.k))
160 prt_printf(out, "#%u", stripe_blockcount_get(sp, i));
161 }
162 }
163
164 /* Triggers: */
165
166 static int __mark_stripe_bucket(struct btree_trans *trans,
167 struct bch_dev *ca,
168 struct bkey_s_c_stripe s,
169 unsigned ptr_idx, bool deleting,
170 struct bpos bucket,
171 struct bch_alloc_v4 *a,
172 enum btree_iter_update_trigger_flags flags)
173 {
174 const struct bch_extent_ptr *ptr = s.v->ptrs + ptr_idx;
175 unsigned nr_data = s.v->nr_blocks - s.v->nr_redundant;
176 bool parity = ptr_idx >= nr_data;
177 enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
178 s64 sectors = parity ? le16_to_cpu(s.v->sectors) : 0;
179 struct printbuf buf = PRINTBUF;
180 int ret = 0;
181
182 struct bch_fs *c = trans->c;
183 if (deleting)
184 sectors = -sectors;
185
186 if (!deleting) {
187 if (bch2_trans_inconsistent_on(a->stripe ||
188 a->stripe_redundancy, trans,
189 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)\n%s",
190 bucket.inode, bucket.offset, a->gen,
191 bch2_data_type_str(a->data_type),
192 a->dirty_sectors,
193 a->stripe, s.k->p.offset,
194 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
195 ret = -EIO;
196 goto err;
197 }
198
199 if (bch2_trans_inconsistent_on(parity && bch2_bucket_sectors_total(*a), trans,
200 "bucket %llu:%llu gen %u data type %s dirty_sectors %u cached_sectors %u: data already in parity bucket\n%s",
201 bucket.inode, bucket.offset, a->gen,
202 bch2_data_type_str(a->data_type),
203 a->dirty_sectors,
204 a->cached_sectors,
205 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
206 ret = -EIO;
207 goto err;
208 }
209 } else {
210 if (bch2_trans_inconsistent_on(a->stripe != s.k->p.offset ||
211 a->stripe_redundancy != s.v->nr_redundant, trans,
212 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe (got %u)\n%s",
213 bucket.inode, bucket.offset, a->gen,
214 a->stripe,
215 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
216 ret = -EIO;
217 goto err;
218 }
219
220 if (bch2_trans_inconsistent_on(a->data_type != data_type, trans,
221 "bucket %llu:%llu gen %u data type %s: wrong data type when stripe, should be %s\n%s",
222 bucket.inode, bucket.offset, a->gen,
223 bch2_data_type_str(a->data_type),
224 bch2_data_type_str(data_type),
225 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
226 ret = -EIO;
227 goto err;
228 }
229
230 if (bch2_trans_inconsistent_on(parity &&
231 (a->dirty_sectors != -sectors ||
232 a->cached_sectors), trans,
233 "bucket %llu:%llu gen %u dirty_sectors %u cached_sectors %u: wrong sectors when deleting parity block of stripe\n%s",
234 bucket.inode, bucket.offset, a->gen,
235 a->dirty_sectors,
236 a->cached_sectors,
237 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
238 ret = -EIO;
239 goto err;
240 }
241 }
242
243 if (sectors) {
244 ret = bch2_bucket_ref_update(trans, ca, s.s_c, ptr, sectors, data_type,
245 a->gen, a->data_type, &a->dirty_sectors);
246 if (ret)
247 goto err;
248 }
249
250 if (!deleting) {
251 a->stripe = s.k->p.offset;
252 a->stripe_redundancy = s.v->nr_redundant;
253 } else {
254 a->stripe = 0;
255 a->stripe_redundancy = 0;
256 }
257
258 alloc_data_type_set(a, data_type);
259 err:
260 printbuf_exit(&buf);
261 return ret;
262 }
263
264 static int mark_stripe_bucket(struct btree_trans *trans,
265 struct bkey_s_c_stripe s,
266 unsigned ptr_idx, bool deleting,
267 enum btree_iter_update_trigger_flags flags)
268 {
269 struct bch_fs *c = trans->c;
270 const struct bch_extent_ptr *ptr = s.v->ptrs + ptr_idx;
271 struct printbuf buf = PRINTBUF;
272 int ret = 0;
273
274 struct bch_dev *ca = bch2_dev_tryget(c, ptr->dev);
275 if (unlikely(!ca)) {
276 if (!(flags & BTREE_TRIGGER_overwrite))
277 ret = -EIO;
278 goto err;
279 }
280
281 struct bpos bucket = PTR_BUCKET_POS(ca, ptr);
282
283 if (flags & BTREE_TRIGGER_transactional) {
284 struct bkey_i_alloc_v4 *a =
285 bch2_trans_start_alloc_update(trans, bucket, 0);
286 ret = PTR_ERR_OR_ZERO(a) ?:
287 __mark_stripe_bucket(trans, ca, s, ptr_idx, deleting, bucket, &a->v, flags);
288 }
289
290 if (flags & BTREE_TRIGGER_gc) {
291 percpu_down_read(&c->mark_lock);
292 struct bucket *g = gc_bucket(ca, bucket.offset);
293 if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u\n %s",
294 ptr->dev,
295 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
296 ret = -EIO;
297 goto err_unlock;
298 }
299
300 bucket_lock(g);
301 struct bch_alloc_v4 old = bucket_m_to_alloc(*g), new = old;
302 ret = __mark_stripe_bucket(trans, ca, s, ptr_idx, deleting, bucket, &new, flags);
303 alloc_to_bucket(g, new);
304 bucket_unlock(g);
305 err_unlock:
306 percpu_up_read(&c->mark_lock);
307 if (!ret)
308 ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
309 }
310 err:
311 bch2_dev_put(ca);
312 printbuf_exit(&buf);
313 return ret;
314 }
315
316 static int mark_stripe_buckets(struct btree_trans *trans,
317 struct bkey_s_c old, struct bkey_s_c new,
318 enum btree_iter_update_trigger_flags flags)
319 {
320 const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
321 ? bkey_s_c_to_stripe(old).v : NULL;
322 const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
323 ? bkey_s_c_to_stripe(new).v : NULL;
324
325 BUG_ON(old_s && new_s && old_s->nr_blocks != new_s->nr_blocks);
326
327 unsigned nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
328
329 for (unsigned i = 0; i < nr_blocks; i++) {
330 if (new_s && old_s &&
331 !memcmp(&new_s->ptrs[i],
332 &old_s->ptrs[i],
333 sizeof(new_s->ptrs[i])))
334 continue;
335
336 if (new_s) {
337 int ret = mark_stripe_bucket(trans,
338 bkey_s_c_to_stripe(new), i, false, flags);
339 if (ret)
340 return ret;
341 }
342
343 if (old_s) {
344 int ret = mark_stripe_bucket(trans,
345 bkey_s_c_to_stripe(old), i, true, flags);
346 if (ret)
347 return ret;
348 }
349 }
350
351 return 0;
352 }
353
354 int bch2_trigger_stripe(struct btree_trans *trans,
355 enum btree_id btree, unsigned level,
356 struct bkey_s_c old, struct bkey_s _new,
357 enum btree_iter_update_trigger_flags flags)
358 {
359 struct bkey_s_c new = _new.s_c;
360 struct bch_fs *c = trans->c;
361 u64 idx = new.k->p.offset;
362 const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
363 ? bkey_s_c_to_stripe(old).v : NULL;
364 const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
365 ? bkey_s_c_to_stripe(new).v : NULL;
366
367 if (unlikely(flags & BTREE_TRIGGER_check_repair))
368 return bch2_check_fix_ptrs(trans, btree, level, _new.s_c, flags);
369
370 BUG_ON(new_s && old_s &&
371 (new_s->nr_blocks != old_s->nr_blocks ||
372 new_s->nr_redundant != old_s->nr_redundant));
373
374
375 if (flags & (BTREE_TRIGGER_transactional|BTREE_TRIGGER_gc)) {
376 /*
377 * If the pointers aren't changing, we don't need to do anything:
378 */
379 if (new_s && old_s &&
380 new_s->nr_blocks == old_s->nr_blocks &&
381 new_s->nr_redundant == old_s->nr_redundant &&
382 !memcmp(old_s->ptrs, new_s->ptrs,
383 new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
384 return 0;
385
386 struct gc_stripe *gc = NULL;
387 if (flags & BTREE_TRIGGER_gc) {
388 gc = genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
389 if (!gc) {
390 bch_err(c, "error allocating memory for gc_stripes, idx %llu", idx);
391 return -BCH_ERR_ENOMEM_mark_stripe;
392 }
393
394 /*
395 * This will be wrong when we bring back runtime gc: we should
396 * be unmarking the old key and then marking the new key
397 *
398 * Also: when we bring back runtime gc, locking
399 */
400 gc->alive = true;
401 gc->sectors = le16_to_cpu(new_s->sectors);
402 gc->nr_blocks = new_s->nr_blocks;
403 gc->nr_redundant = new_s->nr_redundant;
404
405 for (unsigned i = 0; i < new_s->nr_blocks; i++)
406 gc->ptrs[i] = new_s->ptrs[i];
407
408 /*
409 * gc recalculates this field from stripe ptr
410 * references:
411 */
412 memset(gc->block_sectors, 0, sizeof(gc->block_sectors));
413 }
414
415 if (new_s) {
416 s64 sectors = (u64) le16_to_cpu(new_s->sectors) * new_s->nr_redundant;
417
418 struct disk_accounting_pos acc = {
419 .type = BCH_DISK_ACCOUNTING_replicas,
420 };
421 bch2_bkey_to_replicas(&acc.replicas, new);
422 int ret = bch2_disk_accounting_mod(trans, &acc, §ors, 1, gc);
423 if (ret)
424 return ret;
425
426 if (gc)
427 memcpy(&gc->r.e, &acc.replicas, replicas_entry_bytes(&acc.replicas));
428 }
429
430 if (old_s) {
431 s64 sectors = -((s64) le16_to_cpu(old_s->sectors)) * old_s->nr_redundant;
432
433 struct disk_accounting_pos acc = {
434 .type = BCH_DISK_ACCOUNTING_replicas,
435 };
436 bch2_bkey_to_replicas(&acc.replicas, old);
437 int ret = bch2_disk_accounting_mod(trans, &acc, §ors, 1, gc);
438 if (ret)
439 return ret;
440 }
441
442 int ret = mark_stripe_buckets(trans, old, new, flags);
443 if (ret)
444 return ret;
445 }
446
447 if (flags & BTREE_TRIGGER_atomic) {
448 struct stripe *m = genradix_ptr(&c->stripes, idx);
449
450 if (!m) {
451 struct printbuf buf1 = PRINTBUF;
452 struct printbuf buf2 = PRINTBUF;
453
454 bch2_bkey_val_to_text(&buf1, c, old);
455 bch2_bkey_val_to_text(&buf2, c, new);
456 bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
457 "old %s\n"
458 "new %s", idx, buf1.buf, buf2.buf);
459 printbuf_exit(&buf2);
460 printbuf_exit(&buf1);
461 bch2_inconsistent_error(c);
462 return -1;
463 }
464
465 if (!new_s) {
466 bch2_stripes_heap_del(c, m, idx);
467
468 memset(m, 0, sizeof(*m));
469 } else {
470 m->sectors = le16_to_cpu(new_s->sectors);
471 m->algorithm = new_s->algorithm;
472 m->nr_blocks = new_s->nr_blocks;
473 m->nr_redundant = new_s->nr_redundant;
474 m->blocks_nonempty = 0;
475
476 for (unsigned i = 0; i < new_s->nr_blocks; i++)
477 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
478
479 if (!old_s)
480 bch2_stripes_heap_insert(c, m, idx);
481 else
482 bch2_stripes_heap_update(c, m, idx);
483 }
484 }
485
486 return 0;
487 }
488
489 /* returns blocknr in stripe that we matched: */
490 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
491 struct bkey_s_c k, unsigned *block)
492 {
493 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
494 unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
495
496 bkey_for_each_ptr(ptrs, ptr)
497 for (i = 0; i < nr_data; i++)
498 if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
499 le16_to_cpu(s->sectors))) {
500 *block = i;
501 return ptr;
502 }
503
504 return NULL;
505 }
506
507 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
508 {
509 switch (k.k->type) {
510 case KEY_TYPE_extent: {
511 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
512 const union bch_extent_entry *entry;
513
514 extent_for_each_entry(e, entry)
515 if (extent_entry_type(entry) ==
516 BCH_EXTENT_ENTRY_stripe_ptr &&
517 entry->stripe_ptr.idx == idx)
518 return true;
519
520 break;
521 }
522 }
523
524 return false;
525 }
526
527 /* Stripe bufs: */
528
529 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
530 {
531 if (buf->key.k.type == KEY_TYPE_stripe) {
532 struct bkey_i_stripe *s = bkey_i_to_stripe(&buf->key);
533 unsigned i;
534
535 for (i = 0; i < s->v.nr_blocks; i++) {
536 kvfree(buf->data[i]);
537 buf->data[i] = NULL;
538 }
539 }
540 }
541
542 /* XXX: this is a non-mempoolified memory allocation: */
543 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
544 unsigned offset, unsigned size)
545 {
546 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
547 unsigned csum_granularity = 1U << v->csum_granularity_bits;
548 unsigned end = offset + size;
549 unsigned i;
550
551 BUG_ON(end > le16_to_cpu(v->sectors));
552
553 offset = round_down(offset, csum_granularity);
554 end = min_t(unsigned, le16_to_cpu(v->sectors),
555 round_up(end, csum_granularity));
556
557 buf->offset = offset;
558 buf->size = end - offset;
559
560 memset(buf->valid, 0xFF, sizeof(buf->valid));
561
562 for (i = 0; i < v->nr_blocks; i++) {
563 buf->data[i] = kvmalloc(buf->size << 9, GFP_KERNEL);
564 if (!buf->data[i])
565 goto err;
566 }
567
568 return 0;
569 err:
570 ec_stripe_buf_exit(buf);
571 return -BCH_ERR_ENOMEM_stripe_buf;
572 }
573
574 /* Checksumming: */
575
576 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
577 unsigned block, unsigned offset)
578 {
579 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
580 unsigned csum_granularity = 1 << v->csum_granularity_bits;
581 unsigned end = buf->offset + buf->size;
582 unsigned len = min(csum_granularity, end - offset);
583
584 BUG_ON(offset >= end);
585 BUG_ON(offset < buf->offset);
586 BUG_ON(offset & (csum_granularity - 1));
587 BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
588 (len & (csum_granularity - 1)));
589
590 return bch2_checksum(NULL, v->csum_type,
591 null_nonce(),
592 buf->data[block] + ((offset - buf->offset) << 9),
593 len << 9);
594 }
595
596 static void ec_generate_checksums(struct ec_stripe_buf *buf)
597 {
598 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
599 unsigned i, j, csums_per_device = stripe_csums_per_device(v);
600
601 if (!v->csum_type)
602 return;
603
604 BUG_ON(buf->offset);
605 BUG_ON(buf->size != le16_to_cpu(v->sectors));
606
607 for (i = 0; i < v->nr_blocks; i++)
608 for (j = 0; j < csums_per_device; j++)
609 stripe_csum_set(v, i, j,
610 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
611 }
612
613 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
614 {
615 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
616 unsigned csum_granularity = 1 << v->csum_granularity_bits;
617 unsigned i;
618
619 if (!v->csum_type)
620 return;
621
622 for (i = 0; i < v->nr_blocks; i++) {
623 unsigned offset = buf->offset;
624 unsigned end = buf->offset + buf->size;
625
626 if (!test_bit(i, buf->valid))
627 continue;
628
629 while (offset < end) {
630 unsigned j = offset >> v->csum_granularity_bits;
631 unsigned len = min(csum_granularity, end - offset);
632 struct bch_csum want = stripe_csum_get(v, i, j);
633 struct bch_csum got = ec_block_checksum(buf, i, offset);
634
635 if (bch2_crc_cmp(want, got)) {
636 struct bch_dev *ca = bch2_dev_tryget(c, v->ptrs[i].dev);
637 if (ca) {
638 struct printbuf err = PRINTBUF;
639
640 prt_str(&err, "stripe ");
641 bch2_csum_err_msg(&err, v->csum_type, want, got);
642 prt_printf(&err, " for %ps at %u of\n ", (void *) _RET_IP_, i);
643 bch2_bkey_val_to_text(&err, c, bkey_i_to_s_c(&buf->key));
644 bch_err_ratelimited(ca, "%s", err.buf);
645 printbuf_exit(&err);
646
647 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
648 }
649
650 clear_bit(i, buf->valid);
651 break;
652 }
653
654 offset += len;
655 }
656 }
657 }
658
659 /* Erasure coding: */
660
661 static void ec_generate_ec(struct ec_stripe_buf *buf)
662 {
663 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
664 unsigned nr_data = v->nr_blocks - v->nr_redundant;
665 unsigned bytes = le16_to_cpu(v->sectors) << 9;
666
667 raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
668 }
669
670 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
671 {
672 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
673
674 return v->nr_blocks - bitmap_weight(buf->valid, v->nr_blocks);
675 }
676
677 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
678 {
679 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
680 unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
681 unsigned nr_data = v->nr_blocks - v->nr_redundant;
682 unsigned bytes = buf->size << 9;
683
684 if (ec_nr_failed(buf) > v->nr_redundant) {
685 bch_err_ratelimited(c,
686 "error doing reconstruct read: unable to read enough blocks");
687 return -1;
688 }
689
690 for (i = 0; i < nr_data; i++)
691 if (!test_bit(i, buf->valid))
692 failed[nr_failed++] = i;
693
694 raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
695 return 0;
696 }
697
698 /* IO: */
699
700 static void ec_block_endio(struct bio *bio)
701 {
702 struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
703 struct bch_stripe *v = &bkey_i_to_stripe(&ec_bio->buf->key)->v;
704 struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
705 struct bch_dev *ca = ec_bio->ca;
706 struct closure *cl = bio->bi_private;
707
708 if (bch2_dev_io_err_on(bio->bi_status, ca,
709 bio_data_dir(bio)
710 ? BCH_MEMBER_ERROR_write
711 : BCH_MEMBER_ERROR_read,
712 "erasure coding %s error: %s",
713 bio_data_dir(bio) ? "write" : "read",
714 bch2_blk_status_to_str(bio->bi_status)))
715 clear_bit(ec_bio->idx, ec_bio->buf->valid);
716
717 int stale = dev_ptr_stale(ca, ptr);
718 if (stale) {
719 bch_err_ratelimited(ca->fs,
720 "error %s stripe: stale/invalid pointer (%i) after io",
721 bio_data_dir(bio) == READ ? "reading from" : "writing to",
722 stale);
723 clear_bit(ec_bio->idx, ec_bio->buf->valid);
724 }
725
726 bio_put(&ec_bio->bio);
727 percpu_ref_put(&ca->io_ref);
728 closure_put(cl);
729 }
730
731 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
732 blk_opf_t opf, unsigned idx, struct closure *cl)
733 {
734 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
735 unsigned offset = 0, bytes = buf->size << 9;
736 struct bch_extent_ptr *ptr = &v->ptrs[idx];
737 enum bch_data_type data_type = idx < v->nr_blocks - v->nr_redundant
738 ? BCH_DATA_user
739 : BCH_DATA_parity;
740 int rw = op_is_write(opf);
741
742 struct bch_dev *ca = bch2_dev_get_ioref(c, ptr->dev, rw);
743 if (!ca) {
744 clear_bit(idx, buf->valid);
745 return;
746 }
747
748 int stale = dev_ptr_stale(ca, ptr);
749 if (stale) {
750 bch_err_ratelimited(c,
751 "error %s stripe: stale pointer (%i)",
752 rw == READ ? "reading from" : "writing to",
753 stale);
754 clear_bit(idx, buf->valid);
755 return;
756 }
757
758
759 this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
760
761 while (offset < bytes) {
762 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
763 DIV_ROUND_UP(bytes, PAGE_SIZE));
764 unsigned b = min_t(size_t, bytes - offset,
765 nr_iovecs << PAGE_SHIFT);
766 struct ec_bio *ec_bio;
767
768 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
769 nr_iovecs,
770 opf,
771 GFP_KERNEL,
772 &c->ec_bioset),
773 struct ec_bio, bio);
774
775 ec_bio->ca = ca;
776 ec_bio->buf = buf;
777 ec_bio->idx = idx;
778
779 ec_bio->bio.bi_iter.bi_sector = ptr->offset + buf->offset + (offset >> 9);
780 ec_bio->bio.bi_end_io = ec_block_endio;
781 ec_bio->bio.bi_private = cl;
782
783 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
784
785 closure_get(cl);
786 percpu_ref_get(&ca->io_ref);
787
788 submit_bio(&ec_bio->bio);
789
790 offset += b;
791 }
792
793 percpu_ref_put(&ca->io_ref);
794 }
795
796 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
797 struct ec_stripe_buf *stripe)
798 {
799 struct btree_iter iter;
800 struct bkey_s_c k;
801 int ret;
802
803 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
804 POS(0, idx), BTREE_ITER_slots);
805 ret = bkey_err(k);
806 if (ret)
807 goto err;
808 if (k.k->type != KEY_TYPE_stripe) {
809 ret = -ENOENT;
810 goto err;
811 }
812 bkey_reassemble(&stripe->key, k);
813 err:
814 bch2_trans_iter_exit(trans, &iter);
815 return ret;
816 }
817
818 /* recovery read path: */
819 int bch2_ec_read_extent(struct btree_trans *trans, struct bch_read_bio *rbio)
820 {
821 struct bch_fs *c = trans->c;
822 struct ec_stripe_buf *buf;
823 struct closure cl;
824 struct bch_stripe *v;
825 unsigned i, offset;
826 int ret = 0;
827
828 closure_init_stack(&cl);
829
830 BUG_ON(!rbio->pick.has_ec);
831
832 buf = kzalloc(sizeof(*buf), GFP_NOFS);
833 if (!buf)
834 return -BCH_ERR_ENOMEM_ec_read_extent;
835
836 ret = lockrestart_do(trans, get_stripe_key_trans(trans, rbio->pick.ec.idx, buf));
837 if (ret) {
838 bch_err_ratelimited(c,
839 "error doing reconstruct read: error %i looking up stripe", ret);
840 kfree(buf);
841 return -EIO;
842 }
843
844 v = &bkey_i_to_stripe(&buf->key)->v;
845
846 if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
847 bch_err_ratelimited(c,
848 "error doing reconstruct read: pointer doesn't match stripe");
849 ret = -EIO;
850 goto err;
851 }
852
853 offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
854 if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
855 bch_err_ratelimited(c,
856 "error doing reconstruct read: read is bigger than stripe");
857 ret = -EIO;
858 goto err;
859 }
860
861 ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
862 if (ret)
863 goto err;
864
865 for (i = 0; i < v->nr_blocks; i++)
866 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
867
868 closure_sync(&cl);
869
870 if (ec_nr_failed(buf) > v->nr_redundant) {
871 bch_err_ratelimited(c,
872 "error doing reconstruct read: unable to read enough blocks");
873 ret = -EIO;
874 goto err;
875 }
876
877 ec_validate_checksums(c, buf);
878
879 ret = ec_do_recov(c, buf);
880 if (ret)
881 goto err;
882
883 memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
884 buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
885 err:
886 ec_stripe_buf_exit(buf);
887 kfree(buf);
888 return ret;
889 }
890
891 /* stripe bucket accounting: */
892
893 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
894 {
895 ec_stripes_heap n, *h = &c->ec_stripes_heap;
896
897 if (idx >= h->size) {
898 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
899 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
900
901 mutex_lock(&c->ec_stripes_heap_lock);
902 if (n.size > h->size) {
903 memcpy(n.data, h->data, h->nr * sizeof(h->data[0]));
904 n.nr = h->nr;
905 swap(*h, n);
906 }
907 mutex_unlock(&c->ec_stripes_heap_lock);
908
909 free_heap(&n);
910 }
911
912 if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
913 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
914
915 if (c->gc_pos.phase != GC_PHASE_not_running &&
916 !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
917 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
918
919 return 0;
920 }
921
922 static int ec_stripe_mem_alloc(struct btree_trans *trans,
923 struct btree_iter *iter)
924 {
925 return allocate_dropping_locks_errcode(trans,
926 __ec_stripe_mem_alloc(trans->c, iter->pos.offset, _gfp));
927 }
928
929 /*
930 * Hash table of open stripes:
931 * Stripes that are being created or modified are kept in a hash table, so that
932 * stripe deletion can skip them.
933 */
934
935 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
936 {
937 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
938 struct ec_stripe_new *s;
939
940 hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
941 if (s->idx == idx)
942 return true;
943 return false;
944 }
945
946 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
947 {
948 bool ret = false;
949
950 spin_lock(&c->ec_stripes_new_lock);
951 ret = __bch2_stripe_is_open(c, idx);
952 spin_unlock(&c->ec_stripes_new_lock);
953
954 return ret;
955 }
956
957 static bool bch2_try_open_stripe(struct bch_fs *c,
958 struct ec_stripe_new *s,
959 u64 idx)
960 {
961 bool ret;
962
963 spin_lock(&c->ec_stripes_new_lock);
964 ret = !__bch2_stripe_is_open(c, idx);
965 if (ret) {
966 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
967
968 s->idx = idx;
969 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
970 }
971 spin_unlock(&c->ec_stripes_new_lock);
972
973 return ret;
974 }
975
976 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
977 {
978 BUG_ON(!s->idx);
979
980 spin_lock(&c->ec_stripes_new_lock);
981 hlist_del_init(&s->hash);
982 spin_unlock(&c->ec_stripes_new_lock);
983
984 s->idx = 0;
985 }
986
987 /* Heap of all existing stripes, ordered by blocks_nonempty */
988
989 static u64 stripe_idx_to_delete(struct bch_fs *c)
990 {
991 ec_stripes_heap *h = &c->ec_stripes_heap;
992
993 lockdep_assert_held(&c->ec_stripes_heap_lock);
994
995 if (h->nr &&
996 h->data[0].blocks_nonempty == 0 &&
997 !bch2_stripe_is_open(c, h->data[0].idx))
998 return h->data[0].idx;
999
1000 return 0;
1001 }
1002
1003 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
1004 size_t i)
1005 {
1006 struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
1007
1008 genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
1009 }
1010
1011 static inline bool ec_stripes_heap_cmp(const void *l, const void *r, void __always_unused *args)
1012 {
1013 struct ec_stripe_heap_entry *_l = (struct ec_stripe_heap_entry *)l;
1014 struct ec_stripe_heap_entry *_r = (struct ec_stripe_heap_entry *)r;
1015
1016 return ((_l->blocks_nonempty > _r->blocks_nonempty) <
1017 (_l->blocks_nonempty < _r->blocks_nonempty));
1018 }
1019
1020 static inline void ec_stripes_heap_swap(void *l, void *r, void *h)
1021 {
1022 struct ec_stripe_heap_entry *_l = (struct ec_stripe_heap_entry *)l;
1023 struct ec_stripe_heap_entry *_r = (struct ec_stripe_heap_entry *)r;
1024 ec_stripes_heap *_h = (ec_stripes_heap *)h;
1025 size_t i = _l - _h->data;
1026 size_t j = _r - _h->data;
1027
1028 swap(*_l, *_r);
1029
1030 ec_stripes_heap_set_backpointer(_h, i);
1031 ec_stripes_heap_set_backpointer(_h, j);
1032 }
1033
1034 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
1035 {
1036 ec_stripes_heap *h = &c->ec_stripes_heap;
1037 struct stripe *m = genradix_ptr(&c->stripes, idx);
1038
1039 BUG_ON(m->heap_idx >= h->nr);
1040 BUG_ON(h->data[m->heap_idx].idx != idx);
1041 }
1042
1043 void bch2_stripes_heap_del(struct bch_fs *c,
1044 struct stripe *m, size_t idx)
1045 {
1046 const struct min_heap_callbacks callbacks = {
1047 .less = ec_stripes_heap_cmp,
1048 .swp = ec_stripes_heap_swap,
1049 };
1050
1051 mutex_lock(&c->ec_stripes_heap_lock);
1052 heap_verify_backpointer(c, idx);
1053
1054 min_heap_del(&c->ec_stripes_heap, m->heap_idx, &callbacks, &c->ec_stripes_heap);
1055 mutex_unlock(&c->ec_stripes_heap_lock);
1056 }
1057
1058 void bch2_stripes_heap_insert(struct bch_fs *c,
1059 struct stripe *m, size_t idx)
1060 {
1061 const struct min_heap_callbacks callbacks = {
1062 .less = ec_stripes_heap_cmp,
1063 .swp = ec_stripes_heap_swap,
1064 };
1065
1066 mutex_lock(&c->ec_stripes_heap_lock);
1067 BUG_ON(min_heap_full(&c->ec_stripes_heap));
1068
1069 genradix_ptr(&c->stripes, idx)->heap_idx = c->ec_stripes_heap.nr;
1070 min_heap_push(&c->ec_stripes_heap, &((struct ec_stripe_heap_entry) {
1071 .idx = idx,
1072 .blocks_nonempty = m->blocks_nonempty,
1073 }),
1074 &callbacks,
1075 &c->ec_stripes_heap);
1076
1077 heap_verify_backpointer(c, idx);
1078 mutex_unlock(&c->ec_stripes_heap_lock);
1079 }
1080
1081 void bch2_stripes_heap_update(struct bch_fs *c,
1082 struct stripe *m, size_t idx)
1083 {
1084 const struct min_heap_callbacks callbacks = {
1085 .less = ec_stripes_heap_cmp,
1086 .swp = ec_stripes_heap_swap,
1087 };
1088 ec_stripes_heap *h = &c->ec_stripes_heap;
1089 bool do_deletes;
1090 size_t i;
1091
1092 mutex_lock(&c->ec_stripes_heap_lock);
1093 heap_verify_backpointer(c, idx);
1094
1095 h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
1096
1097 i = m->heap_idx;
1098 min_heap_sift_up(h, i, &callbacks, &c->ec_stripes_heap);
1099 min_heap_sift_down(h, i, &callbacks, &c->ec_stripes_heap);
1100
1101 heap_verify_backpointer(c, idx);
1102
1103 do_deletes = stripe_idx_to_delete(c) != 0;
1104 mutex_unlock(&c->ec_stripes_heap_lock);
1105
1106 if (do_deletes)
1107 bch2_do_stripe_deletes(c);
1108 }
1109
1110 /* stripe deletion */
1111
1112 static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
1113 {
1114 struct bch_fs *c = trans->c;
1115 struct btree_iter iter;
1116 struct bkey_s_c k;
1117 struct bkey_s_c_stripe s;
1118 int ret;
1119
1120 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, POS(0, idx),
1121 BTREE_ITER_intent);
1122 ret = bkey_err(k);
1123 if (ret)
1124 goto err;
1125
1126 if (k.k->type != KEY_TYPE_stripe) {
1127 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
1128 ret = -EINVAL;
1129 goto err;
1130 }
1131
1132 s = bkey_s_c_to_stripe(k);
1133 for (unsigned i = 0; i < s.v->nr_blocks; i++)
1134 if (stripe_blockcount_get(s.v, i)) {
1135 struct printbuf buf = PRINTBUF;
1136
1137 bch2_bkey_val_to_text(&buf, c, k);
1138 bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
1139 printbuf_exit(&buf);
1140 ret = -EINVAL;
1141 goto err;
1142 }
1143
1144 ret = bch2_btree_delete_at(trans, &iter, 0);
1145 err:
1146 bch2_trans_iter_exit(trans, &iter);
1147 return ret;
1148 }
1149
1150 static void ec_stripe_delete_work(struct work_struct *work)
1151 {
1152 struct bch_fs *c =
1153 container_of(work, struct bch_fs, ec_stripe_delete_work);
1154
1155 while (1) {
1156 mutex_lock(&c->ec_stripes_heap_lock);
1157 u64 idx = stripe_idx_to_delete(c);
1158 mutex_unlock(&c->ec_stripes_heap_lock);
1159
1160 if (!idx)
1161 break;
1162
1163 int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1164 ec_stripe_delete(trans, idx));
1165 bch_err_fn(c, ret);
1166 if (ret)
1167 break;
1168 }
1169
1170 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
1171 }
1172
1173 void bch2_do_stripe_deletes(struct bch_fs *c)
1174 {
1175 if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
1176 !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work))
1177 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
1178 }
1179
1180 /* stripe creation: */
1181
1182 static int ec_stripe_key_update(struct btree_trans *trans,
1183 struct bkey_i_stripe *new,
1184 bool create)
1185 {
1186 struct bch_fs *c = trans->c;
1187 struct btree_iter iter;
1188 struct bkey_s_c k;
1189 int ret;
1190
1191 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
1192 new->k.p, BTREE_ITER_intent);
1193 ret = bkey_err(k);
1194 if (ret)
1195 goto err;
1196
1197 if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
1198 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
1199 create ? "creating" : "updating",
1200 bch2_bkey_types[k.k->type]);
1201 ret = -EINVAL;
1202 goto err;
1203 }
1204
1205 if (k.k->type == KEY_TYPE_stripe) {
1206 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
1207 unsigned i;
1208
1209 if (old->nr_blocks != new->v.nr_blocks) {
1210 bch_err(c, "error updating stripe: nr_blocks does not match");
1211 ret = -EINVAL;
1212 goto err;
1213 }
1214
1215 for (i = 0; i < new->v.nr_blocks; i++) {
1216 unsigned v = stripe_blockcount_get(old, i);
1217
1218 BUG_ON(v &&
1219 (old->ptrs[i].dev != new->v.ptrs[i].dev ||
1220 old->ptrs[i].gen != new->v.ptrs[i].gen ||
1221 old->ptrs[i].offset != new->v.ptrs[i].offset));
1222
1223 stripe_blockcount_set(&new->v, i, v);
1224 }
1225 }
1226
1227 ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
1228 err:
1229 bch2_trans_iter_exit(trans, &iter);
1230 return ret;
1231 }
1232
1233 static int ec_stripe_update_extent(struct btree_trans *trans,
1234 struct bch_dev *ca,
1235 struct bpos bucket, u8 gen,
1236 struct ec_stripe_buf *s,
1237 struct bpos *bp_pos)
1238 {
1239 struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
1240 struct bch_fs *c = trans->c;
1241 struct bch_backpointer bp;
1242 struct btree_iter iter;
1243 struct bkey_s_c k;
1244 const struct bch_extent_ptr *ptr_c;
1245 struct bch_extent_ptr *ec_ptr = NULL;
1246 struct bch_extent_stripe_ptr stripe_ptr;
1247 struct bkey_i *n;
1248 int ret, dev, block;
1249
1250 ret = bch2_get_next_backpointer(trans, ca, bucket, gen,
1251 bp_pos, &bp, BTREE_ITER_cached);
1252 if (ret)
1253 return ret;
1254 if (bpos_eq(*bp_pos, SPOS_MAX))
1255 return 0;
1256
1257 if (bp.level) {
1258 struct printbuf buf = PRINTBUF;
1259 struct btree_iter node_iter;
1260 struct btree *b;
1261
1262 b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
1263 bch2_trans_iter_exit(trans, &node_iter);
1264
1265 if (!b)
1266 return 0;
1267
1268 prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
1269 bch2_backpointer_to_text(&buf, &bp);
1270
1271 bch2_fs_inconsistent(c, "%s", buf.buf);
1272 printbuf_exit(&buf);
1273 return -EIO;
1274 }
1275
1276 k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_intent);
1277 ret = bkey_err(k);
1278 if (ret)
1279 return ret;
1280 if (!k.k) {
1281 /*
1282 * extent no longer exists - we could flush the btree
1283 * write buffer and retry to verify, but no need:
1284 */
1285 return 0;
1286 }
1287
1288 if (extent_has_stripe_ptr(k, s->key.k.p.offset))
1289 goto out;
1290
1291 ptr_c = bkey_matches_stripe(v, k, &block);
1292 /*
1293 * It doesn't generally make sense to erasure code cached ptrs:
1294 * XXX: should we be incrementing a counter?
1295 */
1296 if (!ptr_c || ptr_c->cached)
1297 goto out;
1298
1299 dev = v->ptrs[block].dev;
1300
1301 n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
1302 ret = PTR_ERR_OR_ZERO(n);
1303 if (ret)
1304 goto out;
1305
1306 bkey_reassemble(n, k);
1307
1308 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
1309 ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev);
1310 BUG_ON(!ec_ptr);
1311
1312 stripe_ptr = (struct bch_extent_stripe_ptr) {
1313 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
1314 .block = block,
1315 .redundancy = v->nr_redundant,
1316 .idx = s->key.k.p.offset,
1317 };
1318
1319 __extent_entry_insert(n,
1320 (union bch_extent_entry *) ec_ptr,
1321 (union bch_extent_entry *) &stripe_ptr);
1322
1323 ret = bch2_trans_update(trans, &iter, n, 0);
1324 out:
1325 bch2_trans_iter_exit(trans, &iter);
1326 return ret;
1327 }
1328
1329 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
1330 unsigned block)
1331 {
1332 struct bch_fs *c = trans->c;
1333 struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
1334 struct bch_extent_ptr ptr = v->ptrs[block];
1335 struct bpos bp_pos = POS_MIN;
1336 int ret = 0;
1337
1338 struct bch_dev *ca = bch2_dev_tryget(c, ptr.dev);
1339 if (!ca)
1340 return -EIO;
1341
1342 struct bpos bucket_pos = PTR_BUCKET_POS(ca, &ptr);
1343
1344 while (1) {
1345 ret = commit_do(trans, NULL, NULL,
1346 BCH_TRANS_COMMIT_no_check_rw|
1347 BCH_TRANS_COMMIT_no_enospc,
1348 ec_stripe_update_extent(trans, ca, bucket_pos, ptr.gen, s, &bp_pos));
1349 if (ret)
1350 break;
1351 if (bkey_eq(bp_pos, POS_MAX))
1352 break;
1353
1354 bp_pos = bpos_nosnap_successor(bp_pos);
1355 }
1356
1357 bch2_dev_put(ca);
1358 return ret;
1359 }
1360
1361 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
1362 {
1363 struct btree_trans *trans = bch2_trans_get(c);
1364 struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
1365 unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1366 int ret = 0;
1367
1368 ret = bch2_btree_write_buffer_flush_sync(trans);
1369 if (ret)
1370 goto err;
1371
1372 for (i = 0; i < nr_data; i++) {
1373 ret = ec_stripe_update_bucket(trans, s, i);
1374 if (ret)
1375 break;
1376 }
1377 err:
1378 bch2_trans_put(trans);
1379
1380 return ret;
1381 }
1382
1383 static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1384 struct ec_stripe_new *s,
1385 unsigned block,
1386 struct open_bucket *ob)
1387 {
1388 struct bch_dev *ca = bch2_dev_get_ioref(c, ob->dev, WRITE);
1389 if (!ca) {
1390 s->err = -BCH_ERR_erofs_no_writes;
1391 return;
1392 }
1393
1394 unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1395 memset(s->new_stripe.data[block] + (offset << 9),
1396 0,
1397 ob->sectors_free << 9);
1398
1399 int ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
1400 ob->bucket * ca->mi.bucket_size + offset,
1401 ob->sectors_free,
1402 GFP_KERNEL, 0);
1403
1404 percpu_ref_put(&ca->io_ref);
1405
1406 if (ret)
1407 s->err = ret;
1408 }
1409
1410 void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s)
1411 {
1412 if (s->idx)
1413 bch2_stripe_close(c, s);
1414 kfree(s);
1415 }
1416
1417 /*
1418 * data buckets of new stripe all written: create the stripe
1419 */
1420 static void ec_stripe_create(struct ec_stripe_new *s)
1421 {
1422 struct bch_fs *c = s->c;
1423 struct open_bucket *ob;
1424 struct bch_stripe *v = &bkey_i_to_stripe(&s->new_stripe.key)->v;
1425 unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1426 int ret;
1427
1428 BUG_ON(s->h->s == s);
1429
1430 closure_sync(&s->iodone);
1431
1432 if (!s->err) {
1433 for (i = 0; i < nr_data; i++)
1434 if (s->blocks[i]) {
1435 ob = c->open_buckets + s->blocks[i];
1436
1437 if (ob->sectors_free)
1438 zero_out_rest_of_ec_bucket(c, s, i, ob);
1439 }
1440 }
1441
1442 if (s->err) {
1443 if (!bch2_err_matches(s->err, EROFS))
1444 bch_err(c, "error creating stripe: error writing data buckets");
1445 goto err;
1446 }
1447
1448 if (s->have_existing_stripe) {
1449 ec_validate_checksums(c, &s->existing_stripe);
1450
1451 if (ec_do_recov(c, &s->existing_stripe)) {
1452 bch_err(c, "error creating stripe: error reading existing stripe");
1453 goto err;
1454 }
1455
1456 for (i = 0; i < nr_data; i++)
1457 if (stripe_blockcount_get(&bkey_i_to_stripe(&s->existing_stripe.key)->v, i))
1458 swap(s->new_stripe.data[i],
1459 s->existing_stripe.data[i]);
1460
1461 ec_stripe_buf_exit(&s->existing_stripe);
1462 }
1463
1464 BUG_ON(!s->allocated);
1465 BUG_ON(!s->idx);
1466
1467 ec_generate_ec(&s->new_stripe);
1468
1469 ec_generate_checksums(&s->new_stripe);
1470
1471 /* write p/q: */
1472 for (i = nr_data; i < v->nr_blocks; i++)
1473 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
1474 closure_sync(&s->iodone);
1475
1476 if (ec_nr_failed(&s->new_stripe)) {
1477 bch_err(c, "error creating stripe: error writing redundancy buckets");
1478 goto err;
1479 }
1480
1481 ret = bch2_trans_do(c, &s->res, NULL,
1482 BCH_TRANS_COMMIT_no_check_rw|
1483 BCH_TRANS_COMMIT_no_enospc,
1484 ec_stripe_key_update(trans,
1485 bkey_i_to_stripe(&s->new_stripe.key),
1486 !s->have_existing_stripe));
1487 bch_err_msg(c, ret, "creating stripe key");
1488 if (ret) {
1489 goto err;
1490 }
1491
1492 ret = ec_stripe_update_extents(c, &s->new_stripe);
1493 bch_err_msg(c, ret, "error updating extents");
1494 if (ret)
1495 goto err;
1496 err:
1497 bch2_disk_reservation_put(c, &s->res);
1498
1499 for (i = 0; i < v->nr_blocks; i++)
1500 if (s->blocks[i]) {
1501 ob = c->open_buckets + s->blocks[i];
1502
1503 if (i < nr_data) {
1504 ob->ec = NULL;
1505 __bch2_open_bucket_put(c, ob);
1506 } else {
1507 bch2_open_bucket_put(c, ob);
1508 }
1509 }
1510
1511 mutex_lock(&c->ec_stripe_new_lock);
1512 list_del(&s->list);
1513 mutex_unlock(&c->ec_stripe_new_lock);
1514 wake_up(&c->ec_stripe_new_wait);
1515
1516 ec_stripe_buf_exit(&s->existing_stripe);
1517 ec_stripe_buf_exit(&s->new_stripe);
1518 closure_debug_destroy(&s->iodone);
1519
1520 ec_stripe_new_put(c, s, STRIPE_REF_stripe);
1521 }
1522
1523 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1524 {
1525 struct ec_stripe_new *s;
1526
1527 mutex_lock(&c->ec_stripe_new_lock);
1528 list_for_each_entry(s, &c->ec_stripe_new_list, list)
1529 if (!atomic_read(&s->ref[STRIPE_REF_io]))
1530 goto out;
1531 s = NULL;
1532 out:
1533 mutex_unlock(&c->ec_stripe_new_lock);
1534
1535 return s;
1536 }
1537
1538 static void ec_stripe_create_work(struct work_struct *work)
1539 {
1540 struct bch_fs *c = container_of(work,
1541 struct bch_fs, ec_stripe_create_work);
1542 struct ec_stripe_new *s;
1543
1544 while ((s = get_pending_stripe(c)))
1545 ec_stripe_create(s);
1546
1547 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1548 }
1549
1550 void bch2_ec_do_stripe_creates(struct bch_fs *c)
1551 {
1552 bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
1553
1554 if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
1555 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1556 }
1557
1558 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1559 {
1560 struct ec_stripe_new *s = h->s;
1561
1562 BUG_ON(!s->allocated && !s->err);
1563
1564 h->s = NULL;
1565 s->pending = true;
1566
1567 mutex_lock(&c->ec_stripe_new_lock);
1568 list_add(&s->list, &c->ec_stripe_new_list);
1569 mutex_unlock(&c->ec_stripe_new_lock);
1570
1571 ec_stripe_new_put(c, s, STRIPE_REF_io);
1572 }
1573
1574 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1575 {
1576 struct ec_stripe_new *s = ob->ec;
1577
1578 s->err = -EIO;
1579 }
1580
1581 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1582 {
1583 struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1584 if (!ob)
1585 return NULL;
1586
1587 BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
1588
1589 struct bch_dev *ca = ob_dev(c, ob);
1590 unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1591
1592 return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1593 }
1594
1595 static int unsigned_cmp(const void *_l, const void *_r)
1596 {
1597 unsigned l = *((const unsigned *) _l);
1598 unsigned r = *((const unsigned *) _r);
1599
1600 return cmp_int(l, r);
1601 }
1602
1603 /* pick most common bucket size: */
1604 static unsigned pick_blocksize(struct bch_fs *c,
1605 struct bch_devs_mask *devs)
1606 {
1607 unsigned nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1608 struct {
1609 unsigned nr, size;
1610 } cur = { 0, 0 }, best = { 0, 0 };
1611
1612 for_each_member_device_rcu(c, ca, devs)
1613 sizes[nr++] = ca->mi.bucket_size;
1614
1615 sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1616
1617 for (unsigned i = 0; i < nr; i++) {
1618 if (sizes[i] != cur.size) {
1619 if (cur.nr > best.nr)
1620 best = cur;
1621
1622 cur.nr = 0;
1623 cur.size = sizes[i];
1624 }
1625
1626 cur.nr++;
1627 }
1628
1629 if (cur.nr > best.nr)
1630 best = cur;
1631
1632 return best.size;
1633 }
1634
1635 static bool may_create_new_stripe(struct bch_fs *c)
1636 {
1637 return false;
1638 }
1639
1640 static void ec_stripe_key_init(struct bch_fs *c,
1641 struct bkey_i *k,
1642 unsigned nr_data,
1643 unsigned nr_parity,
1644 unsigned stripe_size)
1645 {
1646 struct bkey_i_stripe *s = bkey_stripe_init(k);
1647 unsigned u64s;
1648
1649 s->v.sectors = cpu_to_le16(stripe_size);
1650 s->v.algorithm = 0;
1651 s->v.nr_blocks = nr_data + nr_parity;
1652 s->v.nr_redundant = nr_parity;
1653 s->v.csum_granularity_bits = ilog2(c->opts.encoded_extent_max >> 9);
1654 s->v.csum_type = BCH_CSUM_crc32c;
1655 s->v.pad = 0;
1656
1657 while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1658 BUG_ON(1 << s->v.csum_granularity_bits >=
1659 le16_to_cpu(s->v.sectors) ||
1660 s->v.csum_granularity_bits == U8_MAX);
1661 s->v.csum_granularity_bits++;
1662 }
1663
1664 set_bkey_val_u64s(&s->k, u64s);
1665 }
1666
1667 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1668 {
1669 struct ec_stripe_new *s;
1670
1671 lockdep_assert_held(&h->lock);
1672
1673 s = kzalloc(sizeof(*s), GFP_KERNEL);
1674 if (!s)
1675 return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
1676
1677 mutex_init(&s->lock);
1678 closure_init(&s->iodone, NULL);
1679 atomic_set(&s->ref[STRIPE_REF_stripe], 1);
1680 atomic_set(&s->ref[STRIPE_REF_io], 1);
1681 s->c = c;
1682 s->h = h;
1683 s->nr_data = min_t(unsigned, h->nr_active_devs,
1684 BCH_BKEY_PTRS_MAX) - h->redundancy;
1685 s->nr_parity = h->redundancy;
1686
1687 ec_stripe_key_init(c, &s->new_stripe.key,
1688 s->nr_data, s->nr_parity, h->blocksize);
1689
1690 h->s = s;
1691 return 0;
1692 }
1693
1694 static struct ec_stripe_head *
1695 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1696 unsigned algo, unsigned redundancy,
1697 enum bch_watermark watermark)
1698 {
1699 struct ec_stripe_head *h;
1700
1701 h = kzalloc(sizeof(*h), GFP_KERNEL);
1702 if (!h)
1703 return NULL;
1704
1705 mutex_init(&h->lock);
1706 BUG_ON(!mutex_trylock(&h->lock));
1707
1708 h->target = target;
1709 h->algo = algo;
1710 h->redundancy = redundancy;
1711 h->watermark = watermark;
1712
1713 rcu_read_lock();
1714 h->devs = target_rw_devs(c, BCH_DATA_user, target);
1715
1716 for_each_member_device_rcu(c, ca, &h->devs)
1717 if (!ca->mi.durability)
1718 __clear_bit(ca->dev_idx, h->devs.d);
1719
1720 h->blocksize = pick_blocksize(c, &h->devs);
1721
1722 for_each_member_device_rcu(c, ca, &h->devs)
1723 if (ca->mi.bucket_size == h->blocksize)
1724 h->nr_active_devs++;
1725
1726 rcu_read_unlock();
1727
1728 /*
1729 * If we only have redundancy + 1 devices, we're better off with just
1730 * replication:
1731 */
1732 if (h->nr_active_devs < h->redundancy + 2)
1733 bch_err(c, "insufficient devices available to create stripe (have %u, need %u) - mismatched bucket sizes?",
1734 h->nr_active_devs, h->redundancy + 2);
1735
1736 list_add(&h->list, &c->ec_stripe_head_list);
1737 return h;
1738 }
1739
1740 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1741 {
1742 if (h->s &&
1743 h->s->allocated &&
1744 bitmap_weight(h->s->blocks_allocated,
1745 h->s->nr_data) == h->s->nr_data)
1746 ec_stripe_set_pending(c, h);
1747
1748 mutex_unlock(&h->lock);
1749 }
1750
1751 static struct ec_stripe_head *
1752 __bch2_ec_stripe_head_get(struct btree_trans *trans,
1753 unsigned target,
1754 unsigned algo,
1755 unsigned redundancy,
1756 enum bch_watermark watermark)
1757 {
1758 struct bch_fs *c = trans->c;
1759 struct ec_stripe_head *h;
1760 int ret;
1761
1762 if (!redundancy)
1763 return NULL;
1764
1765 ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
1766 if (ret)
1767 return ERR_PTR(ret);
1768
1769 if (test_bit(BCH_FS_going_ro, &c->flags)) {
1770 h = ERR_PTR(-BCH_ERR_erofs_no_writes);
1771 goto found;
1772 }
1773
1774 list_for_each_entry(h, &c->ec_stripe_head_list, list)
1775 if (h->target == target &&
1776 h->algo == algo &&
1777 h->redundancy == redundancy &&
1778 h->watermark == watermark) {
1779 ret = bch2_trans_mutex_lock(trans, &h->lock);
1780 if (ret)
1781 h = ERR_PTR(ret);
1782 goto found;
1783 }
1784
1785 h = ec_new_stripe_head_alloc(c, target, algo, redundancy, watermark);
1786 found:
1787 if (!IS_ERR_OR_NULL(h) &&
1788 h->nr_active_devs < h->redundancy + 2) {
1789 mutex_unlock(&h->lock);
1790 h = NULL;
1791 }
1792 mutex_unlock(&c->ec_stripe_head_lock);
1793 return h;
1794 }
1795
1796 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1797 enum bch_watermark watermark, struct closure *cl)
1798 {
1799 struct bch_fs *c = trans->c;
1800 struct bch_devs_mask devs = h->devs;
1801 struct open_bucket *ob;
1802 struct open_buckets buckets;
1803 struct bch_stripe *v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
1804 unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1805 bool have_cache = true;
1806 int ret = 0;
1807
1808 BUG_ON(v->nr_blocks != h->s->nr_data + h->s->nr_parity);
1809 BUG_ON(v->nr_redundant != h->s->nr_parity);
1810
1811 /* * We bypass the sector allocator which normally does this: */
1812 bitmap_and(devs.d, devs.d, c->rw_devs[BCH_DATA_user].d, BCH_SB_MEMBERS_MAX);
1813
1814 for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) {
1815 __clear_bit(v->ptrs[i].dev, devs.d);
1816 if (i < h->s->nr_data)
1817 nr_have_data++;
1818 else
1819 nr_have_parity++;
1820 }
1821
1822 BUG_ON(nr_have_data > h->s->nr_data);
1823 BUG_ON(nr_have_parity > h->s->nr_parity);
1824
1825 buckets.nr = 0;
1826 if (nr_have_parity < h->s->nr_parity) {
1827 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1828 &h->parity_stripe,
1829 &devs,
1830 h->s->nr_parity,
1831 &nr_have_parity,
1832 &have_cache, 0,
1833 BCH_DATA_parity,
1834 watermark,
1835 cl);
1836
1837 open_bucket_for_each(c, &buckets, ob, i) {
1838 j = find_next_zero_bit(h->s->blocks_gotten,
1839 h->s->nr_data + h->s->nr_parity,
1840 h->s->nr_data);
1841 BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1842
1843 h->s->blocks[j] = buckets.v[i];
1844 v->ptrs[j] = bch2_ob_ptr(c, ob);
1845 __set_bit(j, h->s->blocks_gotten);
1846 }
1847
1848 if (ret)
1849 return ret;
1850 }
1851
1852 buckets.nr = 0;
1853 if (nr_have_data < h->s->nr_data) {
1854 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1855 &h->block_stripe,
1856 &devs,
1857 h->s->nr_data,
1858 &nr_have_data,
1859 &have_cache, 0,
1860 BCH_DATA_user,
1861 watermark,
1862 cl);
1863
1864 open_bucket_for_each(c, &buckets, ob, i) {
1865 j = find_next_zero_bit(h->s->blocks_gotten,
1866 h->s->nr_data, 0);
1867 BUG_ON(j >= h->s->nr_data);
1868
1869 h->s->blocks[j] = buckets.v[i];
1870 v->ptrs[j] = bch2_ob_ptr(c, ob);
1871 __set_bit(j, h->s->blocks_gotten);
1872 }
1873
1874 if (ret)
1875 return ret;
1876 }
1877
1878 return 0;
1879 }
1880
1881 /* XXX: doesn't obey target: */
1882 static s64 get_existing_stripe(struct bch_fs *c,
1883 struct ec_stripe_head *head)
1884 {
1885 ec_stripes_heap *h = &c->ec_stripes_heap;
1886 struct stripe *m;
1887 size_t heap_idx;
1888 u64 stripe_idx;
1889 s64 ret = -1;
1890
1891 if (may_create_new_stripe(c))
1892 return -1;
1893
1894 mutex_lock(&c->ec_stripes_heap_lock);
1895 for (heap_idx = 0; heap_idx < h->nr; heap_idx++) {
1896 /* No blocks worth reusing, stripe will just be deleted: */
1897 if (!h->data[heap_idx].blocks_nonempty)
1898 continue;
1899
1900 stripe_idx = h->data[heap_idx].idx;
1901
1902 m = genradix_ptr(&c->stripes, stripe_idx);
1903
1904 if (m->algorithm == head->algo &&
1905 m->nr_redundant == head->redundancy &&
1906 m->sectors == head->blocksize &&
1907 m->blocks_nonempty < m->nr_blocks - m->nr_redundant &&
1908 bch2_try_open_stripe(c, head->s, stripe_idx)) {
1909 ret = stripe_idx;
1910 break;
1911 }
1912 }
1913 mutex_unlock(&c->ec_stripes_heap_lock);
1914 return ret;
1915 }
1916
1917 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1918 {
1919 struct bch_fs *c = trans->c;
1920 struct bch_stripe *new_v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
1921 struct bch_stripe *existing_v;
1922 unsigned i;
1923 s64 idx;
1924 int ret;
1925
1926 /*
1927 * If we can't allocate a new stripe, and there's no stripes with empty
1928 * blocks for us to reuse, that means we have to wait on copygc:
1929 */
1930 idx = get_existing_stripe(c, h);
1931 if (idx < 0)
1932 return -BCH_ERR_stripe_alloc_blocked;
1933
1934 ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
1935 bch2_fs_fatal_err_on(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart), c,
1936 "reading stripe key: %s", bch2_err_str(ret));
1937 if (ret) {
1938 bch2_stripe_close(c, h->s);
1939 return ret;
1940 }
1941
1942 existing_v = &bkey_i_to_stripe(&h->s->existing_stripe.key)->v;
1943
1944 BUG_ON(existing_v->nr_redundant != h->s->nr_parity);
1945 h->s->nr_data = existing_v->nr_blocks -
1946 existing_v->nr_redundant;
1947
1948 ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
1949 if (ret) {
1950 bch2_stripe_close(c, h->s);
1951 return ret;
1952 }
1953
1954 BUG_ON(h->s->existing_stripe.size != h->blocksize);
1955 BUG_ON(h->s->existing_stripe.size != le16_to_cpu(existing_v->sectors));
1956
1957 /*
1958 * Free buckets we initially allocated - they might conflict with
1959 * blocks from the stripe we're reusing:
1960 */
1961 for_each_set_bit(i, h->s->blocks_gotten, new_v->nr_blocks) {
1962 bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
1963 h->s->blocks[i] = 0;
1964 }
1965 memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
1966 memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
1967
1968 for (i = 0; i < existing_v->nr_blocks; i++) {
1969 if (stripe_blockcount_get(existing_v, i)) {
1970 __set_bit(i, h->s->blocks_gotten);
1971 __set_bit(i, h->s->blocks_allocated);
1972 }
1973
1974 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1975 }
1976
1977 bkey_copy(&h->s->new_stripe.key, &h->s->existing_stripe.key);
1978 h->s->have_existing_stripe = true;
1979
1980 return 0;
1981 }
1982
1983 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1984 {
1985 struct bch_fs *c = trans->c;
1986 struct btree_iter iter;
1987 struct bkey_s_c k;
1988 struct bpos min_pos = POS(0, 1);
1989 struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
1990 int ret;
1991
1992 if (!h->s->res.sectors) {
1993 ret = bch2_disk_reservation_get(c, &h->s->res,
1994 h->blocksize,
1995 h->s->nr_parity,
1996 BCH_DISK_RESERVATION_NOFAIL);
1997 if (ret)
1998 return ret;
1999 }
2000
2001 for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
2002 BTREE_ITER_slots|BTREE_ITER_intent, k, ret) {
2003 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
2004 if (start_pos.offset) {
2005 start_pos = min_pos;
2006 bch2_btree_iter_set_pos(&iter, start_pos);
2007 continue;
2008 }
2009
2010 ret = -BCH_ERR_ENOSPC_stripe_create;
2011 break;
2012 }
2013
2014 if (bkey_deleted(k.k) &&
2015 bch2_try_open_stripe(c, h->s, k.k->p.offset))
2016 break;
2017 }
2018
2019 c->ec_stripe_hint = iter.pos.offset;
2020
2021 if (ret)
2022 goto err;
2023
2024 ret = ec_stripe_mem_alloc(trans, &iter);
2025 if (ret) {
2026 bch2_stripe_close(c, h->s);
2027 goto err;
2028 }
2029
2030 h->s->new_stripe.key.k.p = iter.pos;
2031 out:
2032 bch2_trans_iter_exit(trans, &iter);
2033 return ret;
2034 err:
2035 bch2_disk_reservation_put(c, &h->s->res);
2036 goto out;
2037 }
2038
2039 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
2040 unsigned target,
2041 unsigned algo,
2042 unsigned redundancy,
2043 enum bch_watermark watermark,
2044 struct closure *cl)
2045 {
2046 struct bch_fs *c = trans->c;
2047 struct ec_stripe_head *h;
2048 bool waiting = false;
2049 int ret;
2050
2051 h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, watermark);
2052 if (IS_ERR_OR_NULL(h))
2053 return h;
2054
2055 if (!h->s) {
2056 ret = ec_new_stripe_alloc(c, h);
2057 if (ret) {
2058 bch_err(c, "failed to allocate new stripe");
2059 goto err;
2060 }
2061 }
2062
2063 if (h->s->allocated)
2064 goto allocated;
2065
2066 if (h->s->have_existing_stripe)
2067 goto alloc_existing;
2068
2069 /* First, try to allocate a full stripe: */
2070 ret = new_stripe_alloc_buckets(trans, h, BCH_WATERMARK_stripe, NULL) ?:
2071 __bch2_ec_stripe_head_reserve(trans, h);
2072 if (!ret)
2073 goto allocate_buf;
2074 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
2075 bch2_err_matches(ret, ENOMEM))
2076 goto err;
2077
2078 /*
2079 * Not enough buckets available for a full stripe: we must reuse an
2080 * existing stripe:
2081 */
2082 while (1) {
2083 ret = __bch2_ec_stripe_head_reuse(trans, h);
2084 if (!ret)
2085 break;
2086 if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
2087 goto err;
2088
2089 if (watermark == BCH_WATERMARK_copygc) {
2090 ret = new_stripe_alloc_buckets(trans, h, watermark, NULL) ?:
2091 __bch2_ec_stripe_head_reserve(trans, h);
2092 if (ret)
2093 goto err;
2094 goto allocate_buf;
2095 }
2096
2097 /* XXX freelist_wait? */
2098 closure_wait(&c->freelist_wait, cl);
2099 waiting = true;
2100 }
2101
2102 if (waiting)
2103 closure_wake_up(&c->freelist_wait);
2104 alloc_existing:
2105 /*
2106 * Retry allocating buckets, with the watermark for this
2107 * particular write:
2108 */
2109 ret = new_stripe_alloc_buckets(trans, h, watermark, cl);
2110 if (ret)
2111 goto err;
2112
2113 allocate_buf:
2114 ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
2115 if (ret)
2116 goto err;
2117
2118 h->s->allocated = true;
2119 allocated:
2120 BUG_ON(!h->s->idx);
2121 BUG_ON(!h->s->new_stripe.data[0]);
2122 BUG_ON(trans->restarted);
2123 return h;
2124 err:
2125 bch2_ec_stripe_head_put(c, h);
2126 return ERR_PTR(ret);
2127 }
2128
2129 static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
2130 {
2131 struct ec_stripe_head *h;
2132 struct open_bucket *ob;
2133 unsigned i;
2134
2135 mutex_lock(&c->ec_stripe_head_lock);
2136 list_for_each_entry(h, &c->ec_stripe_head_list, list) {
2137 mutex_lock(&h->lock);
2138 if (!h->s)
2139 goto unlock;
2140
2141 if (!ca)
2142 goto found;
2143
2144 for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++) {
2145 if (!h->s->blocks[i])
2146 continue;
2147
2148 ob = c->open_buckets + h->s->blocks[i];
2149 if (ob->dev == ca->dev_idx)
2150 goto found;
2151 }
2152 goto unlock;
2153 found:
2154 h->s->err = -BCH_ERR_erofs_no_writes;
2155 ec_stripe_set_pending(c, h);
2156 unlock:
2157 mutex_unlock(&h->lock);
2158 }
2159 mutex_unlock(&c->ec_stripe_head_lock);
2160 }
2161
2162 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
2163 {
2164 __bch2_ec_stop(c, ca);
2165 }
2166
2167 void bch2_fs_ec_stop(struct bch_fs *c)
2168 {
2169 __bch2_ec_stop(c, NULL);
2170 }
2171
2172 static bool bch2_fs_ec_flush_done(struct bch_fs *c)
2173 {
2174 bool ret;
2175
2176 mutex_lock(&c->ec_stripe_new_lock);
2177 ret = list_empty(&c->ec_stripe_new_list);
2178 mutex_unlock(&c->ec_stripe_new_lock);
2179
2180 return ret;
2181 }
2182
2183 void bch2_fs_ec_flush(struct bch_fs *c)
2184 {
2185 wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c));
2186 }
2187
2188 int bch2_stripes_read(struct bch_fs *c)
2189 {
2190 int ret = bch2_trans_run(c,
2191 for_each_btree_key(trans, iter, BTREE_ID_stripes, POS_MIN,
2192 BTREE_ITER_prefetch, k, ({
2193 if (k.k->type != KEY_TYPE_stripe)
2194 continue;
2195
2196 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
2197 if (ret)
2198 break;
2199
2200 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
2201
2202 struct stripe *m = genradix_ptr(&c->stripes, k.k->p.offset);
2203 m->sectors = le16_to_cpu(s->sectors);
2204 m->algorithm = s->algorithm;
2205 m->nr_blocks = s->nr_blocks;
2206 m->nr_redundant = s->nr_redundant;
2207 m->blocks_nonempty = 0;
2208
2209 for (unsigned i = 0; i < s->nr_blocks; i++)
2210 m->blocks_nonempty += !!stripe_blockcount_get(s, i);
2211
2212 bch2_stripes_heap_insert(c, m, k.k->p.offset);
2213 0;
2214 })));
2215 bch_err_fn(c, ret);
2216 return ret;
2217 }
2218
2219 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
2220 {
2221 ec_stripes_heap *h = &c->ec_stripes_heap;
2222 struct stripe *m;
2223 size_t i;
2224
2225 mutex_lock(&c->ec_stripes_heap_lock);
2226 for (i = 0; i < min_t(size_t, h->nr, 50); i++) {
2227 m = genradix_ptr(&c->stripes, h->data[i].idx);
2228
2229 prt_printf(out, "%zu %u/%u+%u", h->data[i].idx,
2230 h->data[i].blocks_nonempty,
2231 m->nr_blocks - m->nr_redundant,
2232 m->nr_redundant);
2233 if (bch2_stripe_is_open(c, h->data[i].idx))
2234 prt_str(out, " open");
2235 prt_newline(out);
2236 }
2237 mutex_unlock(&c->ec_stripes_heap_lock);
2238 }
2239
2240 static void bch2_new_stripe_to_text(struct printbuf *out, struct bch_fs *c,
2241 struct ec_stripe_new *s)
2242 {
2243 prt_printf(out, "\tidx %llu blocks %u+%u allocated %u ref %u %u %s obs",
2244 s->idx, s->nr_data, s->nr_parity,
2245 bitmap_weight(s->blocks_allocated, s->nr_data),
2246 atomic_read(&s->ref[STRIPE_REF_io]),
2247 atomic_read(&s->ref[STRIPE_REF_stripe]),
2248 bch2_watermarks[s->h->watermark]);
2249
2250 struct bch_stripe *v = &bkey_i_to_stripe(&s->new_stripe.key)->v;
2251 unsigned i;
2252 for_each_set_bit(i, s->blocks_gotten, v->nr_blocks)
2253 prt_printf(out, " %u", s->blocks[i]);
2254 prt_newline(out);
2255 }
2256
2257 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
2258 {
2259 struct ec_stripe_head *h;
2260 struct ec_stripe_new *s;
2261
2262 mutex_lock(&c->ec_stripe_head_lock);
2263 list_for_each_entry(h, &c->ec_stripe_head_list, list) {
2264 prt_printf(out, "target %u algo %u redundancy %u %s:\n",
2265 h->target, h->algo, h->redundancy,
2266 bch2_watermarks[h->watermark]);
2267
2268 if (h->s)
2269 bch2_new_stripe_to_text(out, c, h->s);
2270 }
2271 mutex_unlock(&c->ec_stripe_head_lock);
2272
2273 prt_printf(out, "in flight:\n");
2274
2275 mutex_lock(&c->ec_stripe_new_lock);
2276 list_for_each_entry(s, &c->ec_stripe_new_list, list)
2277 bch2_new_stripe_to_text(out, c, s);
2278 mutex_unlock(&c->ec_stripe_new_lock);
2279 }
2280
2281 void bch2_fs_ec_exit(struct bch_fs *c)
2282 {
2283 struct ec_stripe_head *h;
2284 unsigned i;
2285
2286 while (1) {
2287 mutex_lock(&c->ec_stripe_head_lock);
2288 h = list_first_entry_or_null(&c->ec_stripe_head_list,
2289 struct ec_stripe_head, list);
2290 if (h)
2291 list_del(&h->list);
2292 mutex_unlock(&c->ec_stripe_head_lock);
2293 if (!h)
2294 break;
2295
2296 if (h->s) {
2297 for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++)
2298 BUG_ON(h->s->blocks[i]);
2299
2300 kfree(h->s);
2301 }
2302 kfree(h);
2303 }
2304
2305 BUG_ON(!list_empty(&c->ec_stripe_new_list));
2306
2307 free_heap(&c->ec_stripes_heap);
2308 genradix_free(&c->stripes);
2309 bioset_exit(&c->ec_bioset);
2310 }
2311
2312 void bch2_fs_ec_init_early(struct bch_fs *c)
2313 {
2314 spin_lock_init(&c->ec_stripes_new_lock);
2315 mutex_init(&c->ec_stripes_heap_lock);
2316
2317 INIT_LIST_HEAD(&c->ec_stripe_head_list);
2318 mutex_init(&c->ec_stripe_head_lock);
2319
2320 INIT_LIST_HEAD(&c->ec_stripe_new_list);
2321 mutex_init(&c->ec_stripe_new_lock);
2322 init_waitqueue_head(&c->ec_stripe_new_wait);
2323
2324 INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
2325 INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
2326 }
2327
2328 int bch2_fs_ec_init(struct bch_fs *c)
2329 {
2330 return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
2331 BIOSET_NEED_BVECS);
2332 }
2333
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.