]> git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/ec.c
Update bcachefs sources to ed6b7f81a7 six locks: Disable percpu read lock mode in...
[bcachefs-tools-debian] / libbcachefs / ec.c
1 // SPDX-License-Identifier: GPL-2.0
2
3 /* erasure coding */
4
5 #include "bcachefs.h"
6 #include "alloc_foreground.h"
7 #include "backpointers.h"
8 #include "bkey_buf.h"
9 #include "bset.h"
10 #include "btree_gc.h"
11 #include "btree_update.h"
12 #include "btree_write_buffer.h"
13 #include "buckets.h"
14 #include "disk_groups.h"
15 #include "ec.h"
16 #include "error.h"
17 #include "io.h"
18 #include "keylist.h"
19 #include "recovery.h"
20 #include "replicas.h"
21 #include "super-io.h"
22 #include "util.h"
23
24 #include <linux/sort.h>
25
26 #ifdef __KERNEL__
27
28 #include <linux/raid/pq.h>
29 #include <linux/raid/xor.h>
30
31 static void raid5_recov(unsigned disks, unsigned failed_idx,
32                         size_t size, void **data)
33 {
34         unsigned i = 2, nr;
35
36         BUG_ON(failed_idx >= disks);
37
38         swap(data[0], data[failed_idx]);
39         memcpy(data[0], data[1], size);
40
41         while (i < disks) {
42                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
43                 xor_blocks(nr, size, data[0], data + i);
44                 i += nr;
45         }
46
47         swap(data[0], data[failed_idx]);
48 }
49
50 static void raid_gen(int nd, int np, size_t size, void **v)
51 {
52         if (np >= 1)
53                 raid5_recov(nd + np, nd, size, v);
54         if (np >= 2)
55                 raid6_call.gen_syndrome(nd + np, size, v);
56         BUG_ON(np > 2);
57 }
58
59 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
60 {
61         switch (nr) {
62         case 0:
63                 break;
64         case 1:
65                 if (ir[0] < nd + 1)
66                         raid5_recov(nd + 1, ir[0], size, v);
67                 else
68                         raid6_call.gen_syndrome(nd + np, size, v);
69                 break;
70         case 2:
71                 if (ir[1] < nd) {
72                         /* data+data failure. */
73                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
74                 } else if (ir[0] < nd) {
75                         /* data + p/q failure */
76
77                         if (ir[1] == nd) /* data + p failure */
78                                 raid6_datap_recov(nd + np, size, ir[0], v);
79                         else { /* data + q failure */
80                                 raid5_recov(nd + 1, ir[0], size, v);
81                                 raid6_call.gen_syndrome(nd + np, size, v);
82                         }
83                 } else {
84                         raid_gen(nd, np, size, v);
85                 }
86                 break;
87         default:
88                 BUG();
89         }
90 }
91
92 #else
93
94 #include <raid/raid.h>
95
96 #endif
97
98 struct ec_bio {
99         struct bch_dev          *ca;
100         struct ec_stripe_buf    *buf;
101         size_t                  idx;
102         struct bio              bio;
103 };
104
105 /* Stripes btree keys: */
106
107 int bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k,
108                         unsigned flags, struct printbuf *err)
109 {
110         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
111
112         if (bkey_eq(k.k->p, POS_MIN)) {
113                 prt_printf(err, "stripe at POS_MIN");
114                 return -BCH_ERR_invalid_bkey;
115         }
116
117         if (k.k->p.inode) {
118                 prt_printf(err, "nonzero inode field");
119                 return -BCH_ERR_invalid_bkey;
120         }
121
122         if (bkey_val_u64s(k.k) < stripe_val_u64s(s)) {
123                 prt_printf(err, "incorrect value size (%zu < %u)",
124                        bkey_val_u64s(k.k), stripe_val_u64s(s));
125                 return -BCH_ERR_invalid_bkey;
126         }
127
128         return bch2_bkey_ptrs_invalid(c, k, flags, err);
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 *s = bkey_s_c_to_stripe(k).v;
135         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
136
137         prt_printf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
138                s->algorithm,
139                le16_to_cpu(s->sectors),
140                nr_data,
141                s->nr_redundant,
142                s->csum_type,
143                1U << s->csum_granularity_bits);
144
145         for (i = 0; i < s->nr_blocks; i++) {
146                 const struct bch_extent_ptr *ptr = s->ptrs + i;
147                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
148                 u32 offset;
149                 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
150
151                 prt_printf(out, " %u:%llu:%u", ptr->dev, b, offset);
152                 if (i < nr_data)
153                         prt_printf(out, "#%u", stripe_blockcount_get(s, i));
154                 if (ptr_stale(ca, ptr))
155                         prt_printf(out, " stale");
156         }
157 }
158
159 /* returns blocknr in stripe that we matched: */
160 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
161                                                 struct bkey_s_c k, unsigned *block)
162 {
163         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
164         const struct bch_extent_ptr *ptr;
165         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
166
167         bkey_for_each_ptr(ptrs, ptr)
168                 for (i = 0; i < nr_data; i++)
169                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
170                                                       le16_to_cpu(s->sectors))) {
171                                 *block = i;
172                                 return ptr;
173                         }
174
175         return NULL;
176 }
177
178 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
179 {
180         switch (k.k->type) {
181         case KEY_TYPE_extent: {
182                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
183                 const union bch_extent_entry *entry;
184
185                 extent_for_each_entry(e, entry)
186                         if (extent_entry_type(entry) ==
187                             BCH_EXTENT_ENTRY_stripe_ptr &&
188                             entry->stripe_ptr.idx == idx)
189                                 return true;
190
191                 break;
192         }
193         }
194
195         return false;
196 }
197
198 /* Stripe bufs: */
199
200 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
201 {
202         unsigned i;
203
204         for (i = 0; i < buf->key.v.nr_blocks; i++) {
205                 kvpfree(buf->data[i], buf->size << 9);
206                 buf->data[i] = NULL;
207         }
208 }
209
210 /* XXX: this is a non-mempoolified memory allocation: */
211 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
212                               unsigned offset, unsigned size)
213 {
214         struct bch_stripe *v = &buf->key.v;
215         unsigned csum_granularity = 1U << v->csum_granularity_bits;
216         unsigned end = offset + size;
217         unsigned i;
218
219         BUG_ON(end > le16_to_cpu(v->sectors));
220
221         offset  = round_down(offset, csum_granularity);
222         end     = min_t(unsigned, le16_to_cpu(v->sectors),
223                         round_up(end, csum_granularity));
224
225         buf->offset     = offset;
226         buf->size       = end - offset;
227
228         memset(buf->valid, 0xFF, sizeof(buf->valid));
229
230         for (i = 0; i < buf->key.v.nr_blocks; i++) {
231                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
232                 if (!buf->data[i])
233                         goto err;
234         }
235
236         return 0;
237 err:
238         ec_stripe_buf_exit(buf);
239         return -BCH_ERR_ENOMEM_stripe_buf;
240 }
241
242 /* Checksumming: */
243
244 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
245                                          unsigned block, unsigned offset)
246 {
247         struct bch_stripe *v = &buf->key.v;
248         unsigned csum_granularity = 1 << v->csum_granularity_bits;
249         unsigned end = buf->offset + buf->size;
250         unsigned len = min(csum_granularity, end - offset);
251
252         BUG_ON(offset >= end);
253         BUG_ON(offset <  buf->offset);
254         BUG_ON(offset & (csum_granularity - 1));
255         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
256                (len & (csum_granularity - 1)));
257
258         return bch2_checksum(NULL, v->csum_type,
259                              null_nonce(),
260                              buf->data[block] + ((offset - buf->offset) << 9),
261                              len << 9);
262 }
263
264 static void ec_generate_checksums(struct ec_stripe_buf *buf)
265 {
266         struct bch_stripe *v = &buf->key.v;
267         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
268
269         if (!v->csum_type)
270                 return;
271
272         BUG_ON(buf->offset);
273         BUG_ON(buf->size != le16_to_cpu(v->sectors));
274
275         for (i = 0; i < v->nr_blocks; i++)
276                 for (j = 0; j < csums_per_device; j++)
277                         stripe_csum_set(v, i, j,
278                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
279 }
280
281 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
282 {
283         struct bch_stripe *v = &buf->key.v;
284         unsigned csum_granularity = 1 << v->csum_granularity_bits;
285         unsigned i;
286
287         if (!v->csum_type)
288                 return;
289
290         for (i = 0; i < v->nr_blocks; i++) {
291                 unsigned offset = buf->offset;
292                 unsigned end = buf->offset + buf->size;
293
294                 if (!test_bit(i, buf->valid))
295                         continue;
296
297                 while (offset < end) {
298                         unsigned j = offset >> v->csum_granularity_bits;
299                         unsigned len = min(csum_granularity, end - offset);
300                         struct bch_csum want = stripe_csum_get(v, i, j);
301                         struct bch_csum got = ec_block_checksum(buf, i, offset);
302
303                         if (bch2_crc_cmp(want, got)) {
304                                 struct printbuf buf2 = PRINTBUF;
305
306                                 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
307
308                                 bch_err_ratelimited(c,
309                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
310                                         (void *) _RET_IP_, i, j, v->csum_type,
311                                         want.lo, got.lo, buf2.buf);
312                                 printbuf_exit(&buf2);
313                                 clear_bit(i, buf->valid);
314                                 break;
315                         }
316
317                         offset += len;
318                 }
319         }
320 }
321
322 /* Erasure coding: */
323
324 static void ec_generate_ec(struct ec_stripe_buf *buf)
325 {
326         struct bch_stripe *v = &buf->key.v;
327         unsigned nr_data = v->nr_blocks - v->nr_redundant;
328         unsigned bytes = le16_to_cpu(v->sectors) << 9;
329
330         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
331 }
332
333 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
334 {
335         return buf->key.v.nr_blocks -
336                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
337 }
338
339 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
340 {
341         struct bch_stripe *v = &buf->key.v;
342         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
343         unsigned nr_data = v->nr_blocks - v->nr_redundant;
344         unsigned bytes = buf->size << 9;
345
346         if (ec_nr_failed(buf) > v->nr_redundant) {
347                 bch_err_ratelimited(c,
348                         "error doing reconstruct read: unable to read enough blocks");
349                 return -1;
350         }
351
352         for (i = 0; i < nr_data; i++)
353                 if (!test_bit(i, buf->valid))
354                         failed[nr_failed++] = i;
355
356         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
357         return 0;
358 }
359
360 /* IO: */
361
362 static void ec_block_endio(struct bio *bio)
363 {
364         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
365         struct bch_stripe *v = &ec_bio->buf->key.v;
366         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
367         struct bch_dev *ca = ec_bio->ca;
368         struct closure *cl = bio->bi_private;
369
370         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
371                                bio_data_dir(bio) ? "write" : "read",
372                                bch2_blk_status_to_str(bio->bi_status)))
373                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
374
375         if (ptr_stale(ca, ptr)) {
376                 bch_err_ratelimited(ca->fs,
377                                     "error %s stripe: stale pointer after io",
378                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
379                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
380         }
381
382         bio_put(&ec_bio->bio);
383         percpu_ref_put(&ca->io_ref);
384         closure_put(cl);
385 }
386
387 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
388                         unsigned rw, unsigned idx, struct closure *cl)
389 {
390         struct bch_stripe *v = &buf->key.v;
391         unsigned offset = 0, bytes = buf->size << 9;
392         struct bch_extent_ptr *ptr = &v->ptrs[idx];
393         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
394         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
395                 ? BCH_DATA_user
396                 : BCH_DATA_parity;
397
398         if (ptr_stale(ca, ptr)) {
399                 bch_err_ratelimited(c,
400                                     "error %s stripe: stale pointer",
401                                     rw == READ ? "reading from" : "writing to");
402                 clear_bit(idx, buf->valid);
403                 return;
404         }
405
406         if (!bch2_dev_get_ioref(ca, rw)) {
407                 clear_bit(idx, buf->valid);
408                 return;
409         }
410
411         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
412
413         while (offset < bytes) {
414                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
415                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
416                 unsigned b = min_t(size_t, bytes - offset,
417                                    nr_iovecs << PAGE_SHIFT);
418                 struct ec_bio *ec_bio;
419
420                 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
421                                                        nr_iovecs,
422                                                        rw,
423                                                        GFP_KERNEL,
424                                                        &c->ec_bioset),
425                                       struct ec_bio, bio);
426
427                 ec_bio->ca                      = ca;
428                 ec_bio->buf                     = buf;
429                 ec_bio->idx                     = idx;
430
431                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
432                 ec_bio->bio.bi_end_io           = ec_block_endio;
433                 ec_bio->bio.bi_private          = cl;
434
435                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
436
437                 closure_get(cl);
438                 percpu_ref_get(&ca->io_ref);
439
440                 submit_bio(&ec_bio->bio);
441
442                 offset += b;
443         }
444
445         percpu_ref_put(&ca->io_ref);
446 }
447
448 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
449                                 struct ec_stripe_buf *stripe)
450 {
451         struct btree_iter iter;
452         struct bkey_s_c k;
453         int ret;
454
455         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
456                                POS(0, idx), BTREE_ITER_SLOTS);
457         ret = bkey_err(k);
458         if (ret)
459                 goto err;
460         if (k.k->type != KEY_TYPE_stripe) {
461                 ret = -ENOENT;
462                 goto err;
463         }
464         bkey_reassemble(&stripe->key.k_i, k);
465 err:
466         bch2_trans_iter_exit(trans, &iter);
467         return ret;
468 }
469
470 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
471 {
472         return bch2_trans_run(c, get_stripe_key_trans(&trans, idx, stripe));
473 }
474
475 /* recovery read path: */
476 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
477 {
478         struct ec_stripe_buf *buf;
479         struct closure cl;
480         struct bch_stripe *v;
481         unsigned i, offset;
482         int ret = 0;
483
484         closure_init_stack(&cl);
485
486         BUG_ON(!rbio->pick.has_ec);
487
488         buf = kzalloc(sizeof(*buf), GFP_NOFS);
489         if (!buf)
490                 return -BCH_ERR_ENOMEM_ec_read_extent;
491
492         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
493         if (ret) {
494                 bch_err_ratelimited(c,
495                         "error doing reconstruct read: error %i looking up stripe", ret);
496                 kfree(buf);
497                 return -EIO;
498         }
499
500         v = &buf->key.v;
501
502         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
503                 bch_err_ratelimited(c,
504                         "error doing reconstruct read: pointer doesn't match stripe");
505                 ret = -EIO;
506                 goto err;
507         }
508
509         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
510         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
511                 bch_err_ratelimited(c,
512                         "error doing reconstruct read: read is bigger than stripe");
513                 ret = -EIO;
514                 goto err;
515         }
516
517         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
518         if (ret)
519                 goto err;
520
521         for (i = 0; i < v->nr_blocks; i++)
522                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
523
524         closure_sync(&cl);
525
526         if (ec_nr_failed(buf) > v->nr_redundant) {
527                 bch_err_ratelimited(c,
528                         "error doing reconstruct read: unable to read enough blocks");
529                 ret = -EIO;
530                 goto err;
531         }
532
533         ec_validate_checksums(c, buf);
534
535         ret = ec_do_recov(c, buf);
536         if (ret)
537                 goto err;
538
539         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
540                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
541 err:
542         ec_stripe_buf_exit(buf);
543         kfree(buf);
544         return ret;
545 }
546
547 /* stripe bucket accounting: */
548
549 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
550 {
551         ec_stripes_heap n, *h = &c->ec_stripes_heap;
552
553         if (idx >= h->size) {
554                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
555                         return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
556
557                 mutex_lock(&c->ec_stripes_heap_lock);
558                 if (n.size > h->size) {
559                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
560                         n.used = h->used;
561                         swap(*h, n);
562                 }
563                 mutex_unlock(&c->ec_stripes_heap_lock);
564
565                 free_heap(&n);
566         }
567
568         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
569                 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
570
571         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
572             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
573                 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
574
575         return 0;
576 }
577
578 static int ec_stripe_mem_alloc(struct btree_trans *trans,
579                                struct btree_iter *iter)
580 {
581         return allocate_dropping_locks_errcode(trans,
582                         __ec_stripe_mem_alloc(trans->c, iter->pos.offset, _gfp));
583 }
584
585 /*
586  * Hash table of open stripes:
587  * Stripes that are being created or modified are kept in a hash table, so that
588  * stripe deletion can skip them.
589  */
590
591 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
592 {
593         unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
594         struct ec_stripe_new *s;
595
596         hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
597                 if (s->idx == idx)
598                         return true;
599         return false;
600 }
601
602 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
603 {
604         bool ret = false;
605
606         spin_lock(&c->ec_stripes_new_lock);
607         ret = __bch2_stripe_is_open(c, idx);
608         spin_unlock(&c->ec_stripes_new_lock);
609
610         return ret;
611 }
612
613 static bool bch2_try_open_stripe(struct bch_fs *c,
614                                  struct ec_stripe_new *s,
615                                  u64 idx)
616 {
617         bool ret;
618
619         spin_lock(&c->ec_stripes_new_lock);
620         ret = !__bch2_stripe_is_open(c, idx);
621         if (ret) {
622                 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
623
624                 s->idx = idx;
625                 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
626         }
627         spin_unlock(&c->ec_stripes_new_lock);
628
629         return ret;
630 }
631
632 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
633 {
634         BUG_ON(!s->idx);
635
636         spin_lock(&c->ec_stripes_new_lock);
637         hlist_del_init(&s->hash);
638         spin_unlock(&c->ec_stripes_new_lock);
639
640         s->idx = 0;
641 }
642
643 /* Heap of all existing stripes, ordered by blocks_nonempty */
644
645 static u64 stripe_idx_to_delete(struct bch_fs *c)
646 {
647         ec_stripes_heap *h = &c->ec_stripes_heap;
648
649         lockdep_assert_held(&c->ec_stripes_heap_lock);
650
651         if (h->used &&
652             h->data[0].blocks_nonempty == 0 &&
653             !bch2_stripe_is_open(c, h->data[0].idx))
654                 return h->data[0].idx;
655
656         return 0;
657 }
658
659 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
660                                       struct ec_stripe_heap_entry l,
661                                       struct ec_stripe_heap_entry r)
662 {
663         return ((l.blocks_nonempty > r.blocks_nonempty) -
664                 (l.blocks_nonempty < r.blocks_nonempty));
665 }
666
667 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
668                                                    size_t i)
669 {
670         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
671
672         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
673 }
674
675 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
676 {
677         ec_stripes_heap *h = &c->ec_stripes_heap;
678         struct stripe *m = genradix_ptr(&c->stripes, idx);
679
680         BUG_ON(m->heap_idx >= h->used);
681         BUG_ON(h->data[m->heap_idx].idx != idx);
682 }
683
684 void bch2_stripes_heap_del(struct bch_fs *c,
685                            struct stripe *m, size_t idx)
686 {
687         mutex_lock(&c->ec_stripes_heap_lock);
688         heap_verify_backpointer(c, idx);
689
690         heap_del(&c->ec_stripes_heap, m->heap_idx,
691                  ec_stripes_heap_cmp,
692                  ec_stripes_heap_set_backpointer);
693         mutex_unlock(&c->ec_stripes_heap_lock);
694 }
695
696 void bch2_stripes_heap_insert(struct bch_fs *c,
697                               struct stripe *m, size_t idx)
698 {
699         mutex_lock(&c->ec_stripes_heap_lock);
700         BUG_ON(heap_full(&c->ec_stripes_heap));
701
702         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
703                         .idx = idx,
704                         .blocks_nonempty = m->blocks_nonempty,
705                 }),
706                  ec_stripes_heap_cmp,
707                  ec_stripes_heap_set_backpointer);
708
709         heap_verify_backpointer(c, idx);
710         mutex_unlock(&c->ec_stripes_heap_lock);
711 }
712
713 void bch2_stripes_heap_update(struct bch_fs *c,
714                               struct stripe *m, size_t idx)
715 {
716         ec_stripes_heap *h = &c->ec_stripes_heap;
717         bool do_deletes;
718         size_t i;
719
720         mutex_lock(&c->ec_stripes_heap_lock);
721         heap_verify_backpointer(c, idx);
722
723         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
724
725         i = m->heap_idx;
726         heap_sift_up(h,   i, ec_stripes_heap_cmp,
727                      ec_stripes_heap_set_backpointer);
728         heap_sift_down(h, i, ec_stripes_heap_cmp,
729                        ec_stripes_heap_set_backpointer);
730
731         heap_verify_backpointer(c, idx);
732
733         do_deletes = stripe_idx_to_delete(c) != 0;
734         mutex_unlock(&c->ec_stripes_heap_lock);
735
736         if (do_deletes)
737                 bch2_do_stripe_deletes(c);
738 }
739
740 /* stripe deletion */
741
742 static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
743 {
744         struct bch_fs *c = trans->c;
745         struct btree_iter iter;
746         struct bkey_s_c k;
747         struct bkey_s_c_stripe s;
748         int ret;
749
750         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, POS(0, idx),
751                                BTREE_ITER_INTENT);
752         ret = bkey_err(k);
753         if (ret)
754                 goto err;
755
756         if (k.k->type != KEY_TYPE_stripe) {
757                 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
758                 ret = -EINVAL;
759                 goto err;
760         }
761
762         s = bkey_s_c_to_stripe(k);
763         for (unsigned i = 0; i < s.v->nr_blocks; i++)
764                 if (stripe_blockcount_get(s.v, i)) {
765                         struct printbuf buf = PRINTBUF;
766
767                         bch2_bkey_val_to_text(&buf, c, k);
768                         bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
769                         printbuf_exit(&buf);
770                         ret = -EINVAL;
771                         goto err;
772                 }
773
774         ret = bch2_btree_delete_at(trans, &iter, 0);
775 err:
776         bch2_trans_iter_exit(trans, &iter);
777         return ret;
778 }
779
780 static void ec_stripe_delete_work(struct work_struct *work)
781 {
782         struct bch_fs *c =
783                 container_of(work, struct bch_fs, ec_stripe_delete_work);
784         struct btree_trans trans;
785         int ret;
786         u64 idx;
787
788         bch2_trans_init(&trans, c, 0, 0);
789
790         while (1) {
791                 mutex_lock(&c->ec_stripes_heap_lock);
792                 idx = stripe_idx_to_delete(c);
793                 mutex_unlock(&c->ec_stripes_heap_lock);
794
795                 if (!idx)
796                         break;
797
798                 ret = commit_do(&trans, NULL, NULL, BTREE_INSERT_NOFAIL,
799                                 ec_stripe_delete(&trans, idx));
800                 if (ret) {
801                         bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
802                         break;
803                 }
804         }
805
806         bch2_trans_exit(&trans);
807
808         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
809 }
810
811 void bch2_do_stripe_deletes(struct bch_fs *c)
812 {
813         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
814             !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work))
815                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
816 }
817
818 /* stripe creation: */
819
820 static int ec_stripe_key_update(struct btree_trans *trans,
821                                 struct bkey_i_stripe *new,
822                                 bool create)
823 {
824         struct bch_fs *c = trans->c;
825         struct btree_iter iter;
826         struct bkey_s_c k;
827         int ret;
828
829         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
830                                new->k.p, BTREE_ITER_INTENT);
831         ret = bkey_err(k);
832         if (ret)
833                 goto err;
834
835         if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
836                 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
837                                      create ? "creating" : "updating",
838                                      bch2_bkey_types[k.k->type]);
839                 ret = -EINVAL;
840                 goto err;
841         }
842
843         if (k.k->type == KEY_TYPE_stripe) {
844                 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
845                 unsigned i;
846
847                 if (old->nr_blocks != new->v.nr_blocks) {
848                         bch_err(c, "error updating stripe: nr_blocks does not match");
849                         ret = -EINVAL;
850                         goto err;
851                 }
852
853                 for (i = 0; i < new->v.nr_blocks; i++) {
854                         unsigned v = stripe_blockcount_get(old, i);
855
856                         BUG_ON(v &&
857                                (old->ptrs[i].dev != new->v.ptrs[i].dev ||
858                                 old->ptrs[i].gen != new->v.ptrs[i].gen ||
859                                 old->ptrs[i].offset != new->v.ptrs[i].offset));
860
861                         stripe_blockcount_set(&new->v, i, v);
862                 }
863         }
864
865         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
866 err:
867         bch2_trans_iter_exit(trans, &iter);
868         return ret;
869 }
870
871 static int ec_stripe_update_extent(struct btree_trans *trans,
872                                    struct bpos bucket, u8 gen,
873                                    struct ec_stripe_buf *s,
874                                    struct bpos *bp_pos)
875 {
876         struct bch_fs *c = trans->c;
877         struct bch_backpointer bp;
878         struct btree_iter iter;
879         struct bkey_s_c k;
880         const struct bch_extent_ptr *ptr_c;
881         struct bch_extent_ptr *ptr, *ec_ptr = NULL;
882         struct bch_extent_stripe_ptr stripe_ptr;
883         struct bkey_i *n;
884         int ret, dev, block;
885
886         ret = bch2_get_next_backpointer(trans, bucket, gen,
887                                 bp_pos, &bp, BTREE_ITER_CACHED);
888         if (ret)
889                 return ret;
890         if (bpos_eq(*bp_pos, SPOS_MAX))
891                 return 0;
892
893         if (bp.level) {
894                 struct printbuf buf = PRINTBUF;
895                 struct btree_iter node_iter;
896                 struct btree *b;
897
898                 b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
899                 bch2_trans_iter_exit(trans, &node_iter);
900
901                 if (!b)
902                         return 0;
903
904                 prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
905                 bch2_backpointer_to_text(&buf, &bp);
906
907                 bch2_fs_inconsistent(c, "%s", buf.buf);
908                 printbuf_exit(&buf);
909                 return -EIO;
910         }
911
912         k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT);
913         ret = bkey_err(k);
914         if (ret)
915                 return ret;
916         if (!k.k) {
917                 /*
918                  * extent no longer exists - we could flush the btree
919                  * write buffer and retry to verify, but no need:
920                  */
921                 return 0;
922         }
923
924         if (extent_has_stripe_ptr(k, s->key.k.p.offset))
925                 goto out;
926
927         ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
928         /*
929          * It doesn't generally make sense to erasure code cached ptrs:
930          * XXX: should we be incrementing a counter?
931          */
932         if (!ptr_c || ptr_c->cached)
933                 goto out;
934
935         dev = s->key.v.ptrs[block].dev;
936
937         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
938         ret = PTR_ERR_OR_ZERO(n);
939         if (ret)
940                 goto out;
941
942         bkey_reassemble(n, k);
943
944         bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
945         ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev);
946         BUG_ON(!ec_ptr);
947
948         stripe_ptr = (struct bch_extent_stripe_ptr) {
949                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
950                 .block          = block,
951                 .redundancy     = s->key.v.nr_redundant,
952                 .idx            = s->key.k.p.offset,
953         };
954
955         __extent_entry_insert(n,
956                         (union bch_extent_entry *) ec_ptr,
957                         (union bch_extent_entry *) &stripe_ptr);
958
959         ret = bch2_trans_update(trans, &iter, n, 0);
960 out:
961         bch2_trans_iter_exit(trans, &iter);
962         return ret;
963 }
964
965 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
966                                    unsigned block)
967 {
968         struct bch_fs *c = trans->c;
969         struct bch_extent_ptr bucket = s->key.v.ptrs[block];
970         struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
971         struct bpos bp_pos = POS_MIN;
972         int ret = 0;
973
974         while (1) {
975                 ret = commit_do(trans, NULL, NULL,
976                                 BTREE_INSERT_NOCHECK_RW|
977                                 BTREE_INSERT_NOFAIL,
978                         ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
979                                                 s, &bp_pos));
980                 if (ret)
981                         break;
982                 if (bkey_eq(bp_pos, POS_MAX))
983                         break;
984
985                 bp_pos = bpos_nosnap_successor(bp_pos);
986         }
987
988         return ret;
989 }
990
991 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
992 {
993         struct btree_trans trans;
994         struct bch_stripe *v = &s->key.v;
995         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
996         int ret = 0;
997
998         bch2_trans_init(&trans, c, 0, 0);
999
1000         ret = bch2_btree_write_buffer_flush(&trans);
1001         if (ret)
1002                 goto err;
1003
1004         for (i = 0; i < nr_data; i++) {
1005                 ret = ec_stripe_update_bucket(&trans, s, i);
1006                 if (ret)
1007                         break;
1008         }
1009 err:
1010         bch2_trans_exit(&trans);
1011
1012         return ret;
1013 }
1014
1015 static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1016                                        struct ec_stripe_new *s,
1017                                        unsigned block,
1018                                        struct open_bucket *ob)
1019 {
1020         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1021         unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1022         int ret;
1023
1024         if (!bch2_dev_get_ioref(ca, WRITE)) {
1025                 s->err = -EROFS;
1026                 return;
1027         }
1028
1029         memset(s->new_stripe.data[block] + (offset << 9),
1030                0,
1031                ob->sectors_free << 9);
1032
1033         ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
1034                         ob->bucket * ca->mi.bucket_size + offset,
1035                         ob->sectors_free,
1036                         GFP_KERNEL, 0);
1037
1038         percpu_ref_put(&ca->io_ref);
1039
1040         if (ret)
1041                 s->err = ret;
1042 }
1043
1044 void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s)
1045 {
1046         if (s->idx)
1047                 bch2_stripe_close(c, s);
1048         kfree(s);
1049 }
1050
1051 /*
1052  * data buckets of new stripe all written: create the stripe
1053  */
1054 static void ec_stripe_create(struct ec_stripe_new *s)
1055 {
1056         struct bch_fs *c = s->c;
1057         struct open_bucket *ob;
1058         struct bch_stripe *v = &s->new_stripe.key.v;
1059         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1060         int ret;
1061
1062         BUG_ON(s->h->s == s);
1063
1064         closure_sync(&s->iodone);
1065
1066         if (!s->err) {
1067                 for (i = 0; i < nr_data; i++)
1068                         if (s->blocks[i]) {
1069                                 ob = c->open_buckets + s->blocks[i];
1070
1071                                 if (ob->sectors_free)
1072                                         zero_out_rest_of_ec_bucket(c, s, i, ob);
1073                         }
1074         }
1075
1076         if (s->err) {
1077                 if (!bch2_err_matches(s->err, EROFS))
1078                         bch_err(c, "error creating stripe: error writing data buckets");
1079                 goto err;
1080         }
1081
1082         if (s->have_existing_stripe) {
1083                 ec_validate_checksums(c, &s->existing_stripe);
1084
1085                 if (ec_do_recov(c, &s->existing_stripe)) {
1086                         bch_err(c, "error creating stripe: error reading existing stripe");
1087                         goto err;
1088                 }
1089
1090                 for (i = 0; i < nr_data; i++)
1091                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
1092                                 swap(s->new_stripe.data[i],
1093                                      s->existing_stripe.data[i]);
1094
1095                 ec_stripe_buf_exit(&s->existing_stripe);
1096         }
1097
1098         BUG_ON(!s->allocated);
1099         BUG_ON(!s->idx);
1100
1101         ec_generate_ec(&s->new_stripe);
1102
1103         ec_generate_checksums(&s->new_stripe);
1104
1105         /* write p/q: */
1106         for (i = nr_data; i < v->nr_blocks; i++)
1107                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
1108         closure_sync(&s->iodone);
1109
1110         if (ec_nr_failed(&s->new_stripe)) {
1111                 bch_err(c, "error creating stripe: error writing redundancy buckets");
1112                 goto err;
1113         }
1114
1115         ret = bch2_trans_do(c, &s->res, NULL,
1116                             BTREE_INSERT_NOCHECK_RW|
1117                             BTREE_INSERT_NOFAIL,
1118                             ec_stripe_key_update(&trans, &s->new_stripe.key,
1119                                                  !s->have_existing_stripe));
1120         if (ret) {
1121                 bch_err(c, "error creating stripe: error creating stripe key");
1122                 goto err;
1123         }
1124
1125         ret = ec_stripe_update_extents(c, &s->new_stripe);
1126         if (ret) {
1127                 bch_err(c, "error creating stripe: error updating pointers: %s",
1128                         bch2_err_str(ret));
1129                 goto err;
1130         }
1131 err:
1132         bch2_disk_reservation_put(c, &s->res);
1133
1134         for (i = 0; i < v->nr_blocks; i++)
1135                 if (s->blocks[i]) {
1136                         ob = c->open_buckets + s->blocks[i];
1137
1138                         if (i < nr_data) {
1139                                 ob->ec = NULL;
1140                                 __bch2_open_bucket_put(c, ob);
1141                         } else {
1142                                 bch2_open_bucket_put(c, ob);
1143                         }
1144                 }
1145
1146         mutex_lock(&c->ec_stripe_new_lock);
1147         list_del(&s->list);
1148         mutex_unlock(&c->ec_stripe_new_lock);
1149         wake_up(&c->ec_stripe_new_wait);
1150
1151         ec_stripe_buf_exit(&s->existing_stripe);
1152         ec_stripe_buf_exit(&s->new_stripe);
1153         closure_debug_destroy(&s->iodone);
1154
1155         ec_stripe_new_put(c, s, STRIPE_REF_stripe);
1156 }
1157
1158 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1159 {
1160         struct ec_stripe_new *s;
1161
1162         mutex_lock(&c->ec_stripe_new_lock);
1163         list_for_each_entry(s, &c->ec_stripe_new_list, list)
1164                 if (!atomic_read(&s->ref[STRIPE_REF_io]))
1165                         goto out;
1166         s = NULL;
1167 out:
1168         mutex_unlock(&c->ec_stripe_new_lock);
1169
1170         return s;
1171 }
1172
1173 static void ec_stripe_create_work(struct work_struct *work)
1174 {
1175         struct bch_fs *c = container_of(work,
1176                 struct bch_fs, ec_stripe_create_work);
1177         struct ec_stripe_new *s;
1178
1179         while ((s = get_pending_stripe(c)))
1180                 ec_stripe_create(s);
1181
1182         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1183 }
1184
1185 void bch2_ec_do_stripe_creates(struct bch_fs *c)
1186 {
1187         bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
1188
1189         if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
1190                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1191 }
1192
1193 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1194 {
1195         struct ec_stripe_new *s = h->s;
1196
1197         BUG_ON(!s->allocated && !s->err);
1198
1199         h->s            = NULL;
1200         s->pending      = true;
1201
1202         mutex_lock(&c->ec_stripe_new_lock);
1203         list_add(&s->list, &c->ec_stripe_new_list);
1204         mutex_unlock(&c->ec_stripe_new_lock);
1205
1206         ec_stripe_new_put(c, s, STRIPE_REF_io);
1207 }
1208
1209 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1210 {
1211         struct ec_stripe_new *s = ob->ec;
1212
1213         s->err = -EIO;
1214 }
1215
1216 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1217 {
1218         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1219         struct bch_dev *ca;
1220         unsigned offset;
1221
1222         if (!ob)
1223                 return NULL;
1224
1225         BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
1226
1227         ca      = bch_dev_bkey_exists(c, ob->dev);
1228         offset  = ca->mi.bucket_size - ob->sectors_free;
1229
1230         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1231 }
1232
1233 static int unsigned_cmp(const void *_l, const void *_r)
1234 {
1235         unsigned l = *((const unsigned *) _l);
1236         unsigned r = *((const unsigned *) _r);
1237
1238         return cmp_int(l, r);
1239 }
1240
1241 /* pick most common bucket size: */
1242 static unsigned pick_blocksize(struct bch_fs *c,
1243                                struct bch_devs_mask *devs)
1244 {
1245         struct bch_dev *ca;
1246         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1247         struct {
1248                 unsigned nr, size;
1249         } cur = { 0, 0 }, best = { 0, 0 };
1250
1251         for_each_member_device_rcu(ca, c, i, devs)
1252                 sizes[nr++] = ca->mi.bucket_size;
1253
1254         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1255
1256         for (i = 0; i < nr; i++) {
1257                 if (sizes[i] != cur.size) {
1258                         if (cur.nr > best.nr)
1259                                 best = cur;
1260
1261                         cur.nr = 0;
1262                         cur.size = sizes[i];
1263                 }
1264
1265                 cur.nr++;
1266         }
1267
1268         if (cur.nr > best.nr)
1269                 best = cur;
1270
1271         return best.size;
1272 }
1273
1274 static bool may_create_new_stripe(struct bch_fs *c)
1275 {
1276         return false;
1277 }
1278
1279 static void ec_stripe_key_init(struct bch_fs *c,
1280                                struct bkey_i_stripe *s,
1281                                unsigned nr_data,
1282                                unsigned nr_parity,
1283                                unsigned stripe_size)
1284 {
1285         unsigned u64s;
1286
1287         bkey_stripe_init(&s->k_i);
1288         s->v.sectors                    = cpu_to_le16(stripe_size);
1289         s->v.algorithm                  = 0;
1290         s->v.nr_blocks                  = nr_data + nr_parity;
1291         s->v.nr_redundant               = nr_parity;
1292         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1293         s->v.csum_type                  = BCH_CSUM_crc32c;
1294         s->v.pad                        = 0;
1295
1296         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1297                 BUG_ON(1 << s->v.csum_granularity_bits >=
1298                        le16_to_cpu(s->v.sectors) ||
1299                        s->v.csum_granularity_bits == U8_MAX);
1300                 s->v.csum_granularity_bits++;
1301         }
1302
1303         set_bkey_val_u64s(&s->k, u64s);
1304 }
1305
1306 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1307 {
1308         struct ec_stripe_new *s;
1309
1310         lockdep_assert_held(&h->lock);
1311
1312         s = kzalloc(sizeof(*s), GFP_KERNEL);
1313         if (!s)
1314                 return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
1315
1316         mutex_init(&s->lock);
1317         closure_init(&s->iodone, NULL);
1318         atomic_set(&s->ref[STRIPE_REF_stripe], 1);
1319         atomic_set(&s->ref[STRIPE_REF_io], 1);
1320         s->c            = c;
1321         s->h            = h;
1322         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1323                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1324         s->nr_parity    = h->redundancy;
1325
1326         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1327                            s->nr_parity, h->blocksize);
1328
1329         h->s = s;
1330         return 0;
1331 }
1332
1333 static struct ec_stripe_head *
1334 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1335                          unsigned algo, unsigned redundancy,
1336                          enum alloc_reserve reserve)
1337 {
1338         struct ec_stripe_head *h;
1339         struct bch_dev *ca;
1340         unsigned i;
1341
1342         h = kzalloc(sizeof(*h), GFP_KERNEL);
1343         if (!h)
1344                 return NULL;
1345
1346         mutex_init(&h->lock);
1347         BUG_ON(!mutex_trylock(&h->lock));
1348
1349         h->target       = target;
1350         h->algo         = algo;
1351         h->redundancy   = redundancy;
1352         h->reserve      = reserve;
1353
1354         rcu_read_lock();
1355         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1356
1357         for_each_member_device_rcu(ca, c, i, &h->devs)
1358                 if (!ca->mi.durability)
1359                         __clear_bit(i, h->devs.d);
1360
1361         h->blocksize = pick_blocksize(c, &h->devs);
1362
1363         for_each_member_device_rcu(ca, c, i, &h->devs)
1364                 if (ca->mi.bucket_size == h->blocksize)
1365                         h->nr_active_devs++;
1366
1367         rcu_read_unlock();
1368         list_add(&h->list, &c->ec_stripe_head_list);
1369         return h;
1370 }
1371
1372 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1373 {
1374         if (h->s &&
1375             h->s->allocated &&
1376             bitmap_weight(h->s->blocks_allocated,
1377                           h->s->nr_data) == h->s->nr_data)
1378                 ec_stripe_set_pending(c, h);
1379
1380         mutex_unlock(&h->lock);
1381 }
1382
1383 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
1384                                                  unsigned target,
1385                                                  unsigned algo,
1386                                                  unsigned redundancy,
1387                                                  enum alloc_reserve reserve)
1388 {
1389         struct bch_fs *c = trans->c;
1390         struct ec_stripe_head *h;
1391         int ret;
1392
1393         if (!redundancy)
1394                 return NULL;
1395
1396         ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
1397         if (ret)
1398                 return ERR_PTR(ret);
1399
1400         if (test_bit(BCH_FS_GOING_RO, &c->flags)) {
1401                 h = ERR_PTR(-EROFS);
1402                 goto found;
1403         }
1404
1405         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1406                 if (h->target           == target &&
1407                     h->algo             == algo &&
1408                     h->redundancy       == redundancy &&
1409                     h->reserve          == reserve) {
1410                         ret = bch2_trans_mutex_lock(trans, &h->lock);
1411                         if (ret)
1412                                 h = ERR_PTR(ret);
1413                         goto found;
1414                 }
1415
1416         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, reserve);
1417 found:
1418         mutex_unlock(&c->ec_stripe_head_lock);
1419         return h;
1420 }
1421
1422 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1423                                     enum alloc_reserve reserve, struct closure *cl)
1424 {
1425         struct bch_fs *c = trans->c;
1426         struct bch_devs_mask devs = h->devs;
1427         struct open_bucket *ob;
1428         struct open_buckets buckets;
1429         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1430         bool have_cache = true;
1431         int ret = 0;
1432
1433         BUG_ON(h->s->new_stripe.key.v.nr_blocks         != h->s->nr_data + h->s->nr_parity);
1434         BUG_ON(h->s->new_stripe.key.v.nr_redundant      != h->s->nr_parity);
1435
1436         for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
1437                 __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1438                 if (i < h->s->nr_data)
1439                         nr_have_data++;
1440                 else
1441                         nr_have_parity++;
1442         }
1443
1444         BUG_ON(nr_have_data     > h->s->nr_data);
1445         BUG_ON(nr_have_parity   > h->s->nr_parity);
1446
1447         buckets.nr = 0;
1448         if (nr_have_parity < h->s->nr_parity) {
1449                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1450                                             &h->parity_stripe,
1451                                             &devs,
1452                                             h->s->nr_parity,
1453                                             &nr_have_parity,
1454                                             &have_cache, 0,
1455                                             BCH_DATA_parity,
1456                                             reserve,
1457                                             cl);
1458
1459                 open_bucket_for_each(c, &buckets, ob, i) {
1460                         j = find_next_zero_bit(h->s->blocks_gotten,
1461                                                h->s->nr_data + h->s->nr_parity,
1462                                                h->s->nr_data);
1463                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1464
1465                         h->s->blocks[j] = buckets.v[i];
1466                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1467                         __set_bit(j, h->s->blocks_gotten);
1468                 }
1469
1470                 if (ret)
1471                         return ret;
1472         }
1473
1474         buckets.nr = 0;
1475         if (nr_have_data < h->s->nr_data) {
1476                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1477                                             &h->block_stripe,
1478                                             &devs,
1479                                             h->s->nr_data,
1480                                             &nr_have_data,
1481                                             &have_cache, 0,
1482                                             BCH_DATA_user,
1483                                             reserve,
1484                                             cl);
1485
1486                 open_bucket_for_each(c, &buckets, ob, i) {
1487                         j = find_next_zero_bit(h->s->blocks_gotten,
1488                                                h->s->nr_data, 0);
1489                         BUG_ON(j >= h->s->nr_data);
1490
1491                         h->s->blocks[j] = buckets.v[i];
1492                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1493                         __set_bit(j, h->s->blocks_gotten);
1494                 }
1495
1496                 if (ret)
1497                         return ret;
1498         }
1499
1500         return 0;
1501 }
1502
1503 /* XXX: doesn't obey target: */
1504 static s64 get_existing_stripe(struct bch_fs *c,
1505                                struct ec_stripe_head *head)
1506 {
1507         ec_stripes_heap *h = &c->ec_stripes_heap;
1508         struct stripe *m;
1509         size_t heap_idx;
1510         u64 stripe_idx;
1511         s64 ret = -1;
1512
1513         if (may_create_new_stripe(c))
1514                 return -1;
1515
1516         mutex_lock(&c->ec_stripes_heap_lock);
1517         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1518                 /* No blocks worth reusing, stripe will just be deleted: */
1519                 if (!h->data[heap_idx].blocks_nonempty)
1520                         continue;
1521
1522                 stripe_idx = h->data[heap_idx].idx;
1523
1524                 m = genradix_ptr(&c->stripes, stripe_idx);
1525
1526                 if (m->algorithm        == head->algo &&
1527                     m->nr_redundant     == head->redundancy &&
1528                     m->sectors          == head->blocksize &&
1529                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant &&
1530                     bch2_try_open_stripe(c, head->s, stripe_idx)) {
1531                         ret = stripe_idx;
1532                         break;
1533                 }
1534         }
1535         mutex_unlock(&c->ec_stripes_heap_lock);
1536         return ret;
1537 }
1538
1539 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1540 {
1541         struct bch_fs *c = trans->c;
1542         unsigned i;
1543         s64 idx;
1544         int ret;
1545
1546         /*
1547          * If we can't allocate a new stripe, and there's no stripes with empty
1548          * blocks for us to reuse, that means we have to wait on copygc:
1549          */
1550         idx = get_existing_stripe(c, h);
1551         if (idx < 0)
1552                 return -BCH_ERR_stripe_alloc_blocked;
1553
1554         ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
1555         if (ret) {
1556                 bch2_stripe_close(c, h->s);
1557                 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
1558                         bch2_fs_fatal_error(c, "error reading stripe key: %s", bch2_err_str(ret));
1559                 return ret;
1560         }
1561
1562         BUG_ON(h->s->existing_stripe.key.v.nr_redundant != h->s->nr_parity);
1563         h->s->nr_data = h->s->existing_stripe.key.v.nr_blocks -
1564                 h->s->existing_stripe.key.v.nr_redundant;
1565
1566         ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
1567         if (ret) {
1568                 bch2_stripe_close(c, h->s);
1569                 return ret;
1570         }
1571
1572         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1573         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1574
1575         /*
1576          * Free buckets we initially allocated - they might conflict with
1577          * blocks from the stripe we're reusing:
1578          */
1579         for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
1580                 bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
1581                 h->s->blocks[i] = 0;
1582         }
1583         memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
1584         memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
1585
1586         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1587                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1588                         __set_bit(i, h->s->blocks_gotten);
1589                         __set_bit(i, h->s->blocks_allocated);
1590                 }
1591
1592                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1593         }
1594
1595         bkey_copy(&h->s->new_stripe.key.k_i, &h->s->existing_stripe.key.k_i);
1596         h->s->have_existing_stripe = true;
1597
1598         return 0;
1599 }
1600
1601 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1602 {
1603         struct bch_fs *c = trans->c;
1604         struct btree_iter iter;
1605         struct bkey_s_c k;
1606         struct bpos min_pos = POS(0, 1);
1607         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
1608         int ret;
1609
1610         if (!h->s->res.sectors) {
1611                 ret = bch2_disk_reservation_get(c, &h->s->res,
1612                                         h->blocksize,
1613                                         h->s->nr_parity,
1614                                         BCH_DISK_RESERVATION_NOFAIL);
1615                 if (ret)
1616                         return ret;
1617         }
1618
1619         for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
1620                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
1621                 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
1622                         if (start_pos.offset) {
1623                                 start_pos = min_pos;
1624                                 bch2_btree_iter_set_pos(&iter, start_pos);
1625                                 continue;
1626                         }
1627
1628                         ret = -BCH_ERR_ENOSPC_stripe_create;
1629                         break;
1630                 }
1631
1632                 if (bkey_deleted(k.k) &&
1633                     bch2_try_open_stripe(c, h->s, k.k->p.offset))
1634                         break;
1635         }
1636
1637         c->ec_stripe_hint = iter.pos.offset;
1638
1639         if (ret)
1640                 goto err;
1641
1642         ret = ec_stripe_mem_alloc(trans, &iter);
1643         if (ret) {
1644                 bch2_stripe_close(c, h->s);
1645                 goto err;
1646         }
1647
1648         h->s->new_stripe.key.k.p = iter.pos;
1649 out:
1650         bch2_trans_iter_exit(trans, &iter);
1651         return ret;
1652 err:
1653         bch2_disk_reservation_put(c, &h->s->res);
1654         goto out;
1655 }
1656
1657 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
1658                                                unsigned target,
1659                                                unsigned algo,
1660                                                unsigned redundancy,
1661                                                enum alloc_reserve reserve,
1662                                                struct closure *cl)
1663 {
1664         struct bch_fs *c = trans->c;
1665         struct ec_stripe_head *h;
1666         bool waiting = false;
1667         int ret;
1668
1669         h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, reserve);
1670         if (!h)
1671                 bch_err(c, "no stripe head");
1672         if (IS_ERR_OR_NULL(h))
1673                 return h;
1674
1675         if (!h->s) {
1676                 ret = ec_new_stripe_alloc(c, h);
1677                 if (ret) {
1678                         bch_err(c, "failed to allocate new stripe");
1679                         goto err;
1680                 }
1681         }
1682
1683         if (h->s->allocated)
1684                 goto allocated;
1685
1686         if (h->s->have_existing_stripe)
1687                 goto alloc_existing;
1688
1689         /* First, try to allocate a full stripe: */
1690         ret =   new_stripe_alloc_buckets(trans, h, RESERVE_stripe, NULL) ?:
1691                 __bch2_ec_stripe_head_reserve(trans, h);
1692         if (!ret)
1693                 goto allocate_buf;
1694         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1695             bch2_err_matches(ret, ENOMEM))
1696                 goto err;
1697
1698         /*
1699          * Not enough buckets available for a full stripe: we must reuse an
1700          * existing stripe:
1701          */
1702         while (1) {
1703                 ret = __bch2_ec_stripe_head_reuse(trans, h);
1704                 if (!ret)
1705                         break;
1706                 if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
1707                         goto err;
1708
1709                 if (reserve == RESERVE_movinggc) {
1710                         ret =   new_stripe_alloc_buckets(trans, h, reserve, NULL) ?:
1711                                 __bch2_ec_stripe_head_reserve(trans, h);
1712                         if (ret)
1713                                 goto err;
1714                         goto allocate_buf;
1715                 }
1716
1717                 /* XXX freelist_wait? */
1718                 closure_wait(&c->freelist_wait, cl);
1719                 waiting = true;
1720         }
1721
1722         if (waiting)
1723                 closure_wake_up(&c->freelist_wait);
1724 alloc_existing:
1725         /*
1726          * Retry allocating buckets, with the reserve watermark for this
1727          * particular write:
1728          */
1729         ret = new_stripe_alloc_buckets(trans, h, reserve, cl);
1730         if (ret)
1731                 goto err;
1732
1733 allocate_buf:
1734         ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
1735         if (ret)
1736                 goto err;
1737
1738         h->s->allocated = true;
1739 allocated:
1740         BUG_ON(!h->s->idx);
1741         BUG_ON(!h->s->new_stripe.data[0]);
1742         BUG_ON(trans->restarted);
1743         return h;
1744 err:
1745         bch2_ec_stripe_head_put(c, h);
1746         return ERR_PTR(ret);
1747 }
1748
1749 static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
1750 {
1751         struct ec_stripe_head *h;
1752         struct open_bucket *ob;
1753         unsigned i;
1754
1755         mutex_lock(&c->ec_stripe_head_lock);
1756         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1757                 mutex_lock(&h->lock);
1758                 if (!h->s)
1759                         goto unlock;
1760
1761                 if (!ca)
1762                         goto found;
1763
1764                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1765                         if (!h->s->blocks[i])
1766                                 continue;
1767
1768                         ob = c->open_buckets + h->s->blocks[i];
1769                         if (ob->dev == ca->dev_idx)
1770                                 goto found;
1771                 }
1772                 goto unlock;
1773 found:
1774                 h->s->err = -EROFS;
1775                 ec_stripe_set_pending(c, h);
1776 unlock:
1777                 mutex_unlock(&h->lock);
1778         }
1779         mutex_unlock(&c->ec_stripe_head_lock);
1780 }
1781
1782 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1783 {
1784         __bch2_ec_stop(c, ca);
1785 }
1786
1787 void bch2_fs_ec_stop(struct bch_fs *c)
1788 {
1789         __bch2_ec_stop(c, NULL);
1790 }
1791
1792 static bool bch2_fs_ec_flush_done(struct bch_fs *c)
1793 {
1794         bool ret;
1795
1796         mutex_lock(&c->ec_stripe_new_lock);
1797         ret = list_empty(&c->ec_stripe_new_list);
1798         mutex_unlock(&c->ec_stripe_new_lock);
1799
1800         return ret;
1801 }
1802
1803 void bch2_fs_ec_flush(struct bch_fs *c)
1804 {
1805         wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c));
1806 }
1807
1808 int bch2_stripes_read(struct bch_fs *c)
1809 {
1810         struct btree_trans trans;
1811         struct btree_iter iter;
1812         struct bkey_s_c k;
1813         const struct bch_stripe *s;
1814         struct stripe *m;
1815         unsigned i;
1816         int ret;
1817
1818         bch2_trans_init(&trans, c, 0, 0);
1819
1820         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1821                            BTREE_ITER_PREFETCH, k, ret) {
1822                 if (k.k->type != KEY_TYPE_stripe)
1823                         continue;
1824
1825                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1826                 if (ret)
1827                         break;
1828
1829                 s = bkey_s_c_to_stripe(k).v;
1830
1831                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1832                 m->sectors      = le16_to_cpu(s->sectors);
1833                 m->algorithm    = s->algorithm;
1834                 m->nr_blocks    = s->nr_blocks;
1835                 m->nr_redundant = s->nr_redundant;
1836                 m->blocks_nonempty = 0;
1837
1838                 for (i = 0; i < s->nr_blocks; i++)
1839                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1840
1841                 bch2_stripes_heap_insert(c, m, k.k->p.offset);
1842         }
1843         bch2_trans_iter_exit(&trans, &iter);
1844
1845         bch2_trans_exit(&trans);
1846
1847         if (ret)
1848                 bch_err(c, "error reading stripes: %i", ret);
1849
1850         return ret;
1851 }
1852
1853 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1854 {
1855         ec_stripes_heap *h = &c->ec_stripes_heap;
1856         struct stripe *m;
1857         size_t i;
1858
1859         mutex_lock(&c->ec_stripes_heap_lock);
1860         for (i = 0; i < min_t(size_t, h->used, 50); i++) {
1861                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1862
1863                 prt_printf(out, "%zu %u/%u+%u", h->data[i].idx,
1864                        h->data[i].blocks_nonempty,
1865                        m->nr_blocks - m->nr_redundant,
1866                        m->nr_redundant);
1867                 if (bch2_stripe_is_open(c, h->data[i].idx))
1868                         prt_str(out, " open");
1869                 prt_newline(out);
1870         }
1871         mutex_unlock(&c->ec_stripes_heap_lock);
1872 }
1873
1874 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1875 {
1876         struct ec_stripe_head *h;
1877         struct ec_stripe_new *s;
1878
1879         mutex_lock(&c->ec_stripe_head_lock);
1880         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1881                 prt_printf(out, "target %u algo %u redundancy %u %s:\n",
1882                        h->target, h->algo, h->redundancy,
1883                        bch2_alloc_reserves[h->reserve]);
1884
1885                 if (h->s)
1886                         prt_printf(out, "\tidx %llu blocks %u+%u allocated %u\n",
1887                                h->s->idx, h->s->nr_data, h->s->nr_parity,
1888                                bitmap_weight(h->s->blocks_allocated,
1889                                              h->s->nr_data));
1890         }
1891         mutex_unlock(&c->ec_stripe_head_lock);
1892
1893         prt_printf(out, "in flight:\n");
1894
1895         mutex_lock(&c->ec_stripe_new_lock);
1896         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1897                 prt_printf(out, "\tidx %llu blocks %u+%u ref %u %u %s\n",
1898                            s->idx, s->nr_data, s->nr_parity,
1899                            atomic_read(&s->ref[STRIPE_REF_io]),
1900                            atomic_read(&s->ref[STRIPE_REF_stripe]),
1901                            bch2_alloc_reserves[s->h->reserve]);
1902         }
1903         mutex_unlock(&c->ec_stripe_new_lock);
1904 }
1905
1906 void bch2_fs_ec_exit(struct bch_fs *c)
1907 {
1908         struct ec_stripe_head *h;
1909         unsigned i;
1910
1911         while (1) {
1912                 mutex_lock(&c->ec_stripe_head_lock);
1913                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1914                                              struct ec_stripe_head, list);
1915                 if (h)
1916                         list_del(&h->list);
1917                 mutex_unlock(&c->ec_stripe_head_lock);
1918                 if (!h)
1919                         break;
1920
1921                 if (h->s) {
1922                         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++)
1923                                 BUG_ON(h->s->blocks[i]);
1924
1925                         kfree(h->s);
1926                 }
1927                 kfree(h);
1928         }
1929
1930         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1931
1932         free_heap(&c->ec_stripes_heap);
1933         genradix_free(&c->stripes);
1934         bioset_exit(&c->ec_bioset);
1935 }
1936
1937 void bch2_fs_ec_init_early(struct bch_fs *c)
1938 {
1939         spin_lock_init(&c->ec_stripes_new_lock);
1940         mutex_init(&c->ec_stripes_heap_lock);
1941
1942         INIT_LIST_HEAD(&c->ec_stripe_head_list);
1943         mutex_init(&c->ec_stripe_head_lock);
1944
1945         INIT_LIST_HEAD(&c->ec_stripe_new_list);
1946         mutex_init(&c->ec_stripe_new_lock);
1947         init_waitqueue_head(&c->ec_stripe_new_wait);
1948
1949         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1950         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1951 }
1952
1953 int bch2_fs_ec_init(struct bch_fs *c)
1954 {
1955         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1956                            BIOSET_NEED_BVECS);
1957 }