]> git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/ec.c
Update bcachefs sources to bee34d805c bcachefs: Repair bad data pointers
[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 "bkey_buf.h"
8 #include "bset.h"
9 #include "btree_gc.h"
10 #include "btree_update.h"
11 #include "buckets.h"
12 #include "disk_groups.h"
13 #include "ec.h"
14 #include "error.h"
15 #include "io.h"
16 #include "keylist.h"
17 #include "recovery.h"
18 #include "super-io.h"
19 #include "util.h"
20
21 #include <linux/sort.h>
22
23 #ifdef __KERNEL__
24
25 #include <linux/raid/pq.h>
26 #include <linux/raid/xor.h>
27
28 static void raid5_recov(unsigned disks, unsigned failed_idx,
29                         size_t size, void **data)
30 {
31         unsigned i = 2, nr;
32
33         BUG_ON(failed_idx >= disks);
34
35         swap(data[0], data[failed_idx]);
36         memcpy(data[0], data[1], size);
37
38         while (i < disks) {
39                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
40                 xor_blocks(nr, size, data[0], data + i);
41                 i += nr;
42         }
43
44         swap(data[0], data[failed_idx]);
45 }
46
47 static void raid_gen(int nd, int np, size_t size, void **v)
48 {
49         if (np >= 1)
50                 raid5_recov(nd + np, nd, size, v);
51         if (np >= 2)
52                 raid6_call.gen_syndrome(nd + np, size, v);
53         BUG_ON(np > 2);
54 }
55
56 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
57 {
58         switch (nr) {
59         case 0:
60                 break;
61         case 1:
62                 if (ir[0] < nd + 1)
63                         raid5_recov(nd + 1, ir[0], size, v);
64                 else
65                         raid6_call.gen_syndrome(nd + np, size, v);
66                 break;
67         case 2:
68                 if (ir[1] < nd) {
69                         /* data+data failure. */
70                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
71                 } else if (ir[0] < nd) {
72                         /* data + p/q failure */
73
74                         if (ir[1] == nd) /* data + p failure */
75                                 raid6_datap_recov(nd + np, size, ir[0], v);
76                         else { /* data + q failure */
77                                 raid5_recov(nd + 1, ir[0], size, v);
78                                 raid6_call.gen_syndrome(nd + np, size, v);
79                         }
80                 } else {
81                         raid_gen(nd, np, size, v);
82                 }
83                 break;
84         default:
85                 BUG();
86         }
87 }
88
89 #else
90
91 #include <raid/raid.h>
92
93 #endif
94
95 struct ec_bio {
96         struct bch_dev          *ca;
97         struct ec_stripe_buf    *buf;
98         size_t                  idx;
99         struct bio              bio;
100 };
101
102 /* Stripes btree keys: */
103
104 const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
105 {
106         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
107
108         if (k.k->p.inode)
109                 return "invalid stripe key";
110
111         if (bkey_val_bytes(k.k) < sizeof(*s))
112                 return "incorrect value size";
113
114         if (bkey_val_bytes(k.k) < sizeof(*s) ||
115             bkey_val_u64s(k.k) < stripe_val_u64s(s))
116                 return "incorrect value size";
117
118         return bch2_bkey_ptrs_invalid(c, k);
119 }
120
121 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
122                          struct bkey_s_c k)
123 {
124         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
125         unsigned i;
126
127         pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
128                s->algorithm,
129                le16_to_cpu(s->sectors),
130                s->nr_blocks - s->nr_redundant,
131                s->nr_redundant,
132                s->csum_type,
133                1U << s->csum_granularity_bits);
134
135         for (i = 0; i < s->nr_blocks; i++)
136                 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
137                        (u64) s->ptrs[i].offset,
138                        stripe_blockcount_get(s, i));
139 }
140
141 /* returns blocknr in stripe that we matched: */
142 static int bkey_matches_stripe(struct bch_stripe *s,
143                                struct bkey_s_c k)
144 {
145         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
146         const struct bch_extent_ptr *ptr;
147         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
148
149         bkey_for_each_ptr(ptrs, ptr)
150                 for (i = 0; i < nr_data; i++)
151                         if (__bch2_ptr_matches_stripe(s, ptr, i))
152                                 return i;
153
154         return -1;
155 }
156
157 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
158 {
159         switch (k.k->type) {
160         case KEY_TYPE_extent: {
161                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
162                 const union bch_extent_entry *entry;
163
164                 extent_for_each_entry(e, entry)
165                         if (extent_entry_type(entry) ==
166                             BCH_EXTENT_ENTRY_stripe_ptr &&
167                             entry->stripe_ptr.idx == idx)
168                                 return true;
169
170                 break;
171         }
172         }
173
174         return false;
175 }
176
177 /* Stripe bufs: */
178
179 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
180 {
181         unsigned i;
182
183         for (i = 0; i < buf->key.v.nr_blocks; i++) {
184                 kvpfree(buf->data[i], buf->size << 9);
185                 buf->data[i] = NULL;
186         }
187 }
188
189 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
190                                unsigned offset, unsigned size)
191 {
192         struct bch_stripe *v = &buf->key.v;
193         unsigned csum_granularity = 1U << v->csum_granularity_bits;
194         unsigned end = offset + size;
195         unsigned i;
196
197         BUG_ON(end > le16_to_cpu(v->sectors));
198
199         offset  = round_down(offset, csum_granularity);
200         end     = min_t(unsigned, le16_to_cpu(v->sectors),
201                         round_up(end, csum_granularity));
202
203         buf->offset     = offset;
204         buf->size       = end - offset;
205
206         memset(buf->valid, 0xFF, sizeof(buf->valid));
207
208         for (i = 0; i < buf->key.v.nr_blocks; i++) {
209                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
210                 if (!buf->data[i])
211                         goto err;
212         }
213
214         return 0;
215 err:
216         ec_stripe_buf_exit(buf);
217         return -ENOMEM;
218 }
219
220 /* Checksumming: */
221
222 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
223                                          unsigned block, unsigned offset)
224 {
225         struct bch_stripe *v = &buf->key.v;
226         unsigned csum_granularity = 1 << v->csum_granularity_bits;
227         unsigned end = buf->offset + buf->size;
228         unsigned len = min(csum_granularity, end - offset);
229
230         BUG_ON(offset >= end);
231         BUG_ON(offset <  buf->offset);
232         BUG_ON(offset & (csum_granularity - 1));
233         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
234                (len & (csum_granularity - 1)));
235
236         return bch2_checksum(NULL, v->csum_type,
237                              null_nonce(),
238                              buf->data[block] + ((offset - buf->offset) << 9),
239                              len << 9);
240 }
241
242 static void ec_generate_checksums(struct ec_stripe_buf *buf)
243 {
244         struct bch_stripe *v = &buf->key.v;
245         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
246
247         if (!v->csum_type)
248                 return;
249
250         BUG_ON(buf->offset);
251         BUG_ON(buf->size != le16_to_cpu(v->sectors));
252
253         for (i = 0; i < v->nr_blocks; i++)
254                 for (j = 0; j < csums_per_device; j++)
255                         stripe_csum_set(v, i, j,
256                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
257 }
258
259 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
260 {
261         struct bch_stripe *v = &buf->key.v;
262         unsigned csum_granularity = 1 << v->csum_granularity_bits;
263         unsigned i;
264
265         if (!v->csum_type)
266                 return;
267
268         for (i = 0; i < v->nr_blocks; i++) {
269                 unsigned offset = buf->offset;
270                 unsigned end = buf->offset + buf->size;
271
272                 if (!test_bit(i, buf->valid))
273                         continue;
274
275                 while (offset < end) {
276                         unsigned j = offset >> v->csum_granularity_bits;
277                         unsigned len = min(csum_granularity, end - offset);
278                         struct bch_csum want = stripe_csum_get(v, i, j);
279                         struct bch_csum got = ec_block_checksum(buf, i, offset);
280
281                         if (bch2_crc_cmp(want, got)) {
282                                 bch_err_ratelimited(c,
283                                         "stripe checksum error at %u:%u: csum type %u, expected %llx got %llx",
284                                         i, j, v->csum_type,
285                                         want.lo, got.lo);
286                                 clear_bit(i, buf->valid);
287                                 break;
288                         }
289
290                         offset += len;
291                 }
292         }
293 }
294
295 /* Erasure coding: */
296
297 static void ec_generate_ec(struct ec_stripe_buf *buf)
298 {
299         struct bch_stripe *v = &buf->key.v;
300         unsigned nr_data = v->nr_blocks - v->nr_redundant;
301         unsigned bytes = le16_to_cpu(v->sectors) << 9;
302
303         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
304 }
305
306 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
307 {
308         return buf->key.v.nr_blocks -
309                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
310 }
311
312 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
313 {
314         struct bch_stripe *v = &buf->key.v;
315         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
316         unsigned nr_data = v->nr_blocks - v->nr_redundant;
317         unsigned bytes = buf->size << 9;
318
319         if (ec_nr_failed(buf) > v->nr_redundant) {
320                 bch_err_ratelimited(c,
321                         "error doing reconstruct read: unable to read enough blocks");
322                 return -1;
323         }
324
325         for (i = 0; i < nr_data; i++)
326                 if (!test_bit(i, buf->valid))
327                         failed[nr_failed++] = i;
328
329         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
330         return 0;
331 }
332
333 /* IO: */
334
335 static void ec_block_endio(struct bio *bio)
336 {
337         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
338         struct bch_dev *ca = ec_bio->ca;
339         struct closure *cl = bio->bi_private;
340
341         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
342                                bio_data_dir(bio) ? "write" : "read",
343                                bch2_blk_status_to_str(bio->bi_status)))
344                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
345
346         bio_put(&ec_bio->bio);
347         percpu_ref_put(&ca->io_ref);
348         closure_put(cl);
349 }
350
351 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
352                         unsigned rw, unsigned idx, struct closure *cl)
353 {
354         struct bch_stripe *v = &buf->key.v;
355         unsigned offset = 0, bytes = buf->size << 9;
356         struct bch_extent_ptr *ptr = &v->ptrs[idx];
357         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
358         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
359                 ? BCH_DATA_user
360                 : BCH_DATA_parity;
361
362         if (ptr_stale(ca, ptr)) {
363                 bch_err_ratelimited(c,
364                                     "error %s stripe: stale pointer",
365                                     rw == READ ? "reading from" : "writing to");
366                 clear_bit(idx, buf->valid);
367                 return;
368         }
369
370         if (!bch2_dev_get_ioref(ca, rw)) {
371                 clear_bit(idx, buf->valid);
372                 return;
373         }
374
375         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
376
377         while (offset < bytes) {
378                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_PAGES,
379                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
380                 unsigned b = min_t(size_t, bytes - offset,
381                                    nr_iovecs << PAGE_SHIFT);
382                 struct ec_bio *ec_bio;
383
384                 ec_bio = container_of(bio_alloc_bioset(GFP_KERNEL, nr_iovecs,
385                                                        &c->ec_bioset),
386                                       struct ec_bio, bio);
387
388                 ec_bio->ca                      = ca;
389                 ec_bio->buf                     = buf;
390                 ec_bio->idx                     = idx;
391
392                 bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
393                 bio_set_op_attrs(&ec_bio->bio, rw, 0);
394
395                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
396                 ec_bio->bio.bi_end_io           = ec_block_endio;
397                 ec_bio->bio.bi_private          = cl;
398
399                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
400
401                 closure_get(cl);
402                 percpu_ref_get(&ca->io_ref);
403
404                 submit_bio(&ec_bio->bio);
405
406                 offset += b;
407         }
408
409         percpu_ref_put(&ca->io_ref);
410 }
411
412 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
413 {
414         struct btree_trans trans;
415         struct btree_iter *iter;
416         struct bkey_s_c k;
417         int ret;
418
419         bch2_trans_init(&trans, c, 0, 0);
420         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS(0, idx), BTREE_ITER_SLOTS);
421         k = bch2_btree_iter_peek_slot(iter);
422         ret = bkey_err(k);
423         if (ret)
424                 goto err;
425         if (k.k->type != KEY_TYPE_stripe) {
426                 ret = -ENOENT;
427                 goto err;
428         }
429         bkey_reassemble(&stripe->key.k_i, k);
430 err:
431         bch2_trans_exit(&trans);
432         return ret;
433 }
434
435 /* recovery read path: */
436 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
437 {
438         struct ec_stripe_buf *buf;
439         struct closure cl;
440         struct bch_stripe *v;
441         unsigned i, offset;
442         int ret = 0;
443
444         closure_init_stack(&cl);
445
446         BUG_ON(!rbio->pick.has_ec);
447
448         buf = kzalloc(sizeof(*buf), GFP_NOIO);
449         if (!buf)
450                 return -ENOMEM;
451
452         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
453         if (ret) {
454                 bch_err_ratelimited(c,
455                         "error doing reconstruct read: error %i looking up stripe", ret);
456                 kfree(buf);
457                 return -EIO;
458         }
459
460         v = &buf->key.v;
461
462         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
463                 bch_err_ratelimited(c,
464                         "error doing reconstruct read: pointer doesn't match stripe");
465                 ret = -EIO;
466                 goto err;
467         }
468
469         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
470         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
471                 bch_err_ratelimited(c,
472                         "error doing reconstruct read: read is bigger than stripe");
473                 ret = -EIO;
474                 goto err;
475         }
476
477         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
478         if (ret)
479                 goto err;
480
481         for (i = 0; i < v->nr_blocks; i++)
482                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
483
484         closure_sync(&cl);
485
486         if (ec_nr_failed(buf) > v->nr_redundant) {
487                 bch_err_ratelimited(c,
488                         "error doing reconstruct read: unable to read enough blocks");
489                 ret = -EIO;
490                 goto err;
491         }
492
493         ec_validate_checksums(c, buf);
494
495         ret = ec_do_recov(c, buf);
496         if (ret)
497                 goto err;
498
499         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
500                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
501 err:
502         ec_stripe_buf_exit(buf);
503         kfree(buf);
504         return ret;
505 }
506
507 /* stripe bucket accounting: */
508
509 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
510 {
511         ec_stripes_heap n, *h = &c->ec_stripes_heap;
512
513         if (idx >= h->size) {
514                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
515                         return -ENOMEM;
516
517                 spin_lock(&c->ec_stripes_heap_lock);
518                 if (n.size > h->size) {
519                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
520                         n.used = h->used;
521                         swap(*h, n);
522                 }
523                 spin_unlock(&c->ec_stripes_heap_lock);
524
525                 free_heap(&n);
526         }
527
528         if (!genradix_ptr_alloc(&c->stripes[0], idx, gfp))
529                 return -ENOMEM;
530
531         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
532             !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
533                 return -ENOMEM;
534
535         return 0;
536 }
537
538 static int ec_stripe_mem_alloc(struct bch_fs *c,
539                                struct btree_iter *iter)
540 {
541         size_t idx = iter->pos.offset;
542         int ret = 0;
543
544         if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT|__GFP_NOWARN))
545                 return ret;
546
547         bch2_trans_unlock(iter->trans);
548         ret = -EINTR;
549
550         if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
551                 return ret;
552
553         return -ENOMEM;
554 }
555
556 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
557 {
558         ec_stripes_heap *h = &c->ec_stripes_heap;
559
560         return h->used && h->data[0].blocks_nonempty == 0
561                 ? h->data[0].idx : -1;
562 }
563
564 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
565                                       struct ec_stripe_heap_entry l,
566                                       struct ec_stripe_heap_entry r)
567 {
568         return ((l.blocks_nonempty > r.blocks_nonempty) -
569                 (l.blocks_nonempty < r.blocks_nonempty));
570 }
571
572 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
573                                                    size_t i)
574 {
575         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
576
577         genradix_ptr(&c->stripes[0], h->data[i].idx)->heap_idx = i;
578 }
579
580 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
581 {
582         ec_stripes_heap *h = &c->ec_stripes_heap;
583         struct stripe *m = genradix_ptr(&c->stripes[0], idx);
584
585         BUG_ON(!m->alive);
586         BUG_ON(m->heap_idx >= h->used);
587         BUG_ON(h->data[m->heap_idx].idx != idx);
588 }
589
590 void bch2_stripes_heap_del(struct bch_fs *c,
591                            struct stripe *m, size_t idx)
592 {
593         if (!m->on_heap)
594                 return;
595
596         m->on_heap = false;
597
598         heap_verify_backpointer(c, idx);
599
600         heap_del(&c->ec_stripes_heap, m->heap_idx,
601                  ec_stripes_heap_cmp,
602                  ec_stripes_heap_set_backpointer);
603 }
604
605 void bch2_stripes_heap_insert(struct bch_fs *c,
606                               struct stripe *m, size_t idx)
607 {
608         if (m->on_heap)
609                 return;
610
611         BUG_ON(heap_full(&c->ec_stripes_heap));
612
613         m->on_heap = true;
614
615         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
616                         .idx = idx,
617                         .blocks_nonempty = m->blocks_nonempty,
618                 }),
619                  ec_stripes_heap_cmp,
620                  ec_stripes_heap_set_backpointer);
621
622         heap_verify_backpointer(c, idx);
623 }
624
625 void bch2_stripes_heap_update(struct bch_fs *c,
626                               struct stripe *m, size_t idx)
627 {
628         ec_stripes_heap *h = &c->ec_stripes_heap;
629         size_t i;
630
631         if (!m->on_heap)
632                 return;
633
634         heap_verify_backpointer(c, idx);
635
636         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
637
638         i = m->heap_idx;
639         heap_sift_up(h,   i, ec_stripes_heap_cmp,
640                      ec_stripes_heap_set_backpointer);
641         heap_sift_down(h, i, ec_stripes_heap_cmp,
642                        ec_stripes_heap_set_backpointer);
643
644         heap_verify_backpointer(c, idx);
645
646         if (stripe_idx_to_delete(c) >= 0 &&
647             !percpu_ref_is_dying(&c->writes))
648                 schedule_work(&c->ec_stripe_delete_work);
649 }
650
651 /* stripe deletion */
652
653 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
654 {
655         //pr_info("deleting stripe %zu", idx);
656         return bch2_btree_delete_range(c, BTREE_ID_EC,
657                                        POS(0, idx),
658                                        POS(0, idx + 1),
659                                        NULL);
660 }
661
662 static void ec_stripe_delete_work(struct work_struct *work)
663 {
664         struct bch_fs *c =
665                 container_of(work, struct bch_fs, ec_stripe_delete_work);
666         ssize_t idx;
667
668         while (1) {
669                 spin_lock(&c->ec_stripes_heap_lock);
670                 idx = stripe_idx_to_delete(c);
671                 if (idx < 0) {
672                         spin_unlock(&c->ec_stripes_heap_lock);
673                         break;
674                 }
675
676                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes[0], idx), idx);
677                 spin_unlock(&c->ec_stripes_heap_lock);
678
679                 if (ec_stripe_delete(c, idx))
680                         break;
681         }
682 }
683
684 /* stripe creation: */
685
686 static int ec_stripe_bkey_insert(struct bch_fs *c,
687                                  struct bkey_i_stripe *stripe,
688                                  struct disk_reservation *res)
689 {
690         struct btree_trans trans;
691         struct btree_iter *iter;
692         struct bkey_s_c k;
693         struct bpos min_pos = POS(0, 1);
694         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
695         int ret;
696
697         bch2_trans_init(&trans, c, 0, 0);
698 retry:
699         bch2_trans_begin(&trans);
700
701         for_each_btree_key(&trans, iter, BTREE_ID_EC, start_pos,
702                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
703                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
704                         if (start_pos.offset) {
705                                 start_pos = min_pos;
706                                 bch2_btree_iter_set_pos(iter, start_pos);
707                                 continue;
708                         }
709
710                         ret = -ENOSPC;
711                         break;
712                 }
713
714                 if (bkey_deleted(k.k))
715                         goto found_slot;
716         }
717
718         goto err;
719 found_slot:
720         start_pos = iter->pos;
721
722         ret = ec_stripe_mem_alloc(c, iter);
723         if (ret)
724                 goto err;
725
726         stripe->k.p = iter->pos;
727
728         bch2_trans_update(&trans, iter, &stripe->k_i, 0);
729
730         ret = bch2_trans_commit(&trans, res, NULL,
731                                 BTREE_INSERT_NOFAIL);
732 err:
733         bch2_trans_iter_put(&trans, iter);
734
735         if (ret == -EINTR)
736                 goto retry;
737
738         c->ec_stripe_hint = ret ? start_pos.offset : start_pos.offset + 1;
739         bch2_trans_exit(&trans);
740
741         return ret;
742 }
743
744 static int ec_stripe_bkey_update(struct btree_trans *trans,
745                                  struct bkey_i_stripe *new)
746 {
747         struct btree_iter *iter;
748         struct bkey_s_c k;
749         const struct bch_stripe *existing;
750         unsigned i;
751         int ret;
752
753         iter = bch2_trans_get_iter(trans, BTREE_ID_EC,
754                                    new->k.p, BTREE_ITER_INTENT);
755         k = bch2_btree_iter_peek_slot(iter);
756         ret = bkey_err(k);
757         if (ret)
758                 goto err;
759
760         if (!k.k || k.k->type != KEY_TYPE_stripe) {
761                 bch_err(trans->c, "error updating stripe: not found");
762                 ret = -ENOENT;
763                 goto err;
764         }
765
766         existing = bkey_s_c_to_stripe(k).v;
767
768         if (existing->nr_blocks != new->v.nr_blocks) {
769                 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
770                 ret = -EINVAL;
771                 goto err;
772         }
773
774         for (i = 0; i < new->v.nr_blocks; i++)
775                 stripe_blockcount_set(&new->v, i,
776                         stripe_blockcount_get(existing, i));
777
778         bch2_trans_update(trans, iter, &new->k_i, 0);
779 err:
780         bch2_trans_iter_put(trans, iter);
781         return ret;
782 }
783
784 static void extent_stripe_ptr_add(struct bkey_s_extent e,
785                                   struct ec_stripe_buf *s,
786                                   struct bch_extent_ptr *ptr,
787                                   unsigned block)
788 {
789         struct bch_extent_stripe_ptr *dst = (void *) ptr;
790         union bch_extent_entry *end = extent_entry_last(e);
791
792         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
793         e.k->u64s += sizeof(*dst) / sizeof(u64);
794
795         *dst = (struct bch_extent_stripe_ptr) {
796                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
797                 .block          = block,
798                 .idx            = s->key.k.p.offset,
799         };
800 }
801
802 static int ec_stripe_update_ptrs(struct bch_fs *c,
803                                  struct ec_stripe_buf *s,
804                                  struct bkey *pos)
805 {
806         struct btree_trans trans;
807         struct btree_iter *iter;
808         struct bkey_s_c k;
809         struct bkey_s_extent e;
810         struct bkey_buf sk;
811         int ret = 0, dev, block;
812
813         bch2_bkey_buf_init(&sk);
814         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
815
816         /* XXX this doesn't support the reflink btree */
817
818         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
819                                    bkey_start_pos(pos),
820                                    BTREE_ITER_INTENT);
821
822         while ((k = bch2_btree_iter_peek(iter)).k &&
823                !(ret = bkey_err(k)) &&
824                bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
825                 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
826
827                 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
828                         bch2_btree_iter_next(iter);
829                         continue;
830                 }
831
832                 block = bkey_matches_stripe(&s->key.v, k);
833                 if (block < 0) {
834                         bch2_btree_iter_next(iter);
835                         continue;
836                 }
837
838                 dev = s->key.v.ptrs[block].dev;
839
840                 bch2_bkey_buf_reassemble(&sk, c, k);
841                 e = bkey_i_to_s_extent(sk.k);
842
843                 bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
844                 ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
845                 BUG_ON(!ec_ptr);
846
847                 extent_stripe_ptr_add(e, s, ec_ptr, block);
848
849                 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
850                 bch2_trans_update(&trans, iter, sk.k, 0);
851
852                 ret = bch2_trans_commit(&trans, NULL, NULL,
853                                         BTREE_INSERT_NOFAIL);
854                 if (ret == -EINTR)
855                         ret = 0;
856                 if (ret)
857                         break;
858         }
859
860         bch2_trans_exit(&trans);
861         bch2_bkey_buf_exit(&sk, c);
862
863         return ret;
864 }
865
866 /*
867  * data buckets of new stripe all written: create the stripe
868  */
869 static void ec_stripe_create(struct ec_stripe_new *s)
870 {
871         struct bch_fs *c = s->c;
872         struct open_bucket *ob;
873         struct bkey_i *k;
874         struct stripe *m;
875         struct bch_stripe *v = &s->new_stripe.key.v;
876         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
877         int ret;
878
879         BUG_ON(s->h->s == s);
880
881         closure_sync(&s->iodone);
882
883         if (s->err) {
884                 if (s->err != -EROFS)
885                         bch_err(c, "error creating stripe: error writing data buckets");
886                 goto err;
887         }
888
889         if (s->have_existing_stripe) {
890                 ec_validate_checksums(c, &s->existing_stripe);
891
892                 if (ec_do_recov(c, &s->existing_stripe)) {
893                         bch_err(c, "error creating stripe: error reading existing stripe");
894                         goto err;
895                 }
896
897                 for (i = 0; i < nr_data; i++)
898                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
899                                 swap(s->new_stripe.data[i],
900                                      s->existing_stripe.data[i]);
901
902                 ec_stripe_buf_exit(&s->existing_stripe);
903         }
904
905         BUG_ON(!s->allocated);
906
907         if (!percpu_ref_tryget(&c->writes))
908                 goto err;
909
910         ec_generate_ec(&s->new_stripe);
911
912         ec_generate_checksums(&s->new_stripe);
913
914         /* write p/q: */
915         for (i = nr_data; i < v->nr_blocks; i++)
916                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
917         closure_sync(&s->iodone);
918
919         if (ec_nr_failed(&s->new_stripe)) {
920                 bch_err(c, "error creating stripe: error writing redundancy buckets");
921                 goto err_put_writes;
922         }
923
924         ret = s->have_existing_stripe
925                 ? bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
926                                 ec_stripe_bkey_update(&trans, &s->new_stripe.key))
927                 : ec_stripe_bkey_insert(c, &s->new_stripe.key, &s->res);
928         if (ret) {
929                 bch_err(c, "error creating stripe: error creating stripe key");
930                 goto err_put_writes;
931         }
932
933         for_each_keylist_key(&s->keys, k) {
934                 ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
935                 if (ret) {
936                         bch_err(c, "error creating stripe: error %i updating pointers", ret);
937                         break;
938                 }
939         }
940
941         spin_lock(&c->ec_stripes_heap_lock);
942         m = genradix_ptr(&c->stripes[0], s->new_stripe.key.k.p.offset);
943
944         BUG_ON(m->on_heap);
945         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
946         spin_unlock(&c->ec_stripes_heap_lock);
947 err_put_writes:
948         percpu_ref_put(&c->writes);
949 err:
950         bch2_disk_reservation_put(c, &s->res);
951
952         for (i = 0; i < v->nr_blocks; i++)
953                 if (s->blocks[i]) {
954                         ob = c->open_buckets + s->blocks[i];
955
956                         if (i < nr_data) {
957                                 ob->ec = NULL;
958                                 __bch2_open_bucket_put(c, ob);
959                         } else {
960                                 bch2_open_bucket_put(c, ob);
961                         }
962                 }
963
964         bch2_keylist_free(&s->keys, s->inline_keys);
965
966         ec_stripe_buf_exit(&s->existing_stripe);
967         ec_stripe_buf_exit(&s->new_stripe);
968         closure_debug_destroy(&s->iodone);
969         kfree(s);
970 }
971
972 static void ec_stripe_create_work(struct work_struct *work)
973 {
974         struct bch_fs *c = container_of(work,
975                 struct bch_fs, ec_stripe_create_work);
976         struct ec_stripe_new *s, *n;
977 restart:
978         mutex_lock(&c->ec_stripe_new_lock);
979         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
980                 if (!atomic_read(&s->pin)) {
981                         list_del(&s->list);
982                         mutex_unlock(&c->ec_stripe_new_lock);
983                         ec_stripe_create(s);
984                         goto restart;
985                 }
986         mutex_unlock(&c->ec_stripe_new_lock);
987 }
988
989 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
990 {
991         BUG_ON(atomic_read(&s->pin) <= 0);
992
993         if (atomic_dec_and_test(&s->pin)) {
994                 BUG_ON(!s->pending);
995                 queue_work(system_long_wq, &c->ec_stripe_create_work);
996         }
997 }
998
999 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1000 {
1001         struct ec_stripe_new *s = h->s;
1002
1003         BUG_ON(!s->allocated && !s->err);
1004
1005         h->s            = NULL;
1006         s->pending      = true;
1007
1008         mutex_lock(&c->ec_stripe_new_lock);
1009         list_add(&s->list, &c->ec_stripe_new_list);
1010         mutex_unlock(&c->ec_stripe_new_lock);
1011
1012         ec_stripe_new_put(c, s);
1013 }
1014
1015 /* have a full bucket - hand it off to be erasure coded: */
1016 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1017 {
1018         struct ec_stripe_new *s = ob->ec;
1019
1020         if (ob->sectors_free)
1021                 s->err = -1;
1022
1023         ec_stripe_new_put(c, s);
1024 }
1025
1026 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1027 {
1028         struct ec_stripe_new *s = ob->ec;
1029
1030         s->err = -EIO;
1031 }
1032
1033 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1034 {
1035         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1036         struct bch_dev *ca;
1037         unsigned offset;
1038
1039         if (!ob)
1040                 return NULL;
1041
1042         ca      = bch_dev_bkey_exists(c, ob->ptr.dev);
1043         offset  = ca->mi.bucket_size - ob->sectors_free;
1044
1045         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1046 }
1047
1048 void bch2_ec_add_backpointer(struct bch_fs *c, struct write_point *wp,
1049                              struct bpos pos, unsigned sectors)
1050 {
1051         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1052         struct ec_stripe_new *ec;
1053
1054         if (!ob)
1055                 return;
1056
1057         //pr_info("adding backpointer at %llu:%llu", pos.inode, pos.offset);
1058
1059         ec = ob->ec;
1060         mutex_lock(&ec->lock);
1061
1062         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1063                                  ARRAY_SIZE(ec->inline_keys),
1064                                  BKEY_U64s)) {
1065                 BUG();
1066         }
1067
1068         bkey_init(&ec->keys.top->k);
1069         ec->keys.top->k.p       = pos;
1070         bch2_key_resize(&ec->keys.top->k, sectors);
1071         bch2_keylist_push(&ec->keys);
1072
1073         mutex_unlock(&ec->lock);
1074 }
1075
1076 static int unsigned_cmp(const void *_l, const void *_r)
1077 {
1078         unsigned l = *((const unsigned *) _l);
1079         unsigned r = *((const unsigned *) _r);
1080
1081         return cmp_int(l, r);
1082 }
1083
1084 /* pick most common bucket size: */
1085 static unsigned pick_blocksize(struct bch_fs *c,
1086                                struct bch_devs_mask *devs)
1087 {
1088         struct bch_dev *ca;
1089         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1090         struct {
1091                 unsigned nr, size;
1092         } cur = { 0, 0 }, best = { 0, 0 };
1093
1094         for_each_member_device_rcu(ca, c, i, devs)
1095                 sizes[nr++] = ca->mi.bucket_size;
1096
1097         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1098
1099         for (i = 0; i < nr; i++) {
1100                 if (sizes[i] != cur.size) {
1101                         if (cur.nr > best.nr)
1102                                 best = cur;
1103
1104                         cur.nr = 0;
1105                         cur.size = sizes[i];
1106                 }
1107
1108                 cur.nr++;
1109         }
1110
1111         if (cur.nr > best.nr)
1112                 best = cur;
1113
1114         return best.size;
1115 }
1116
1117 static bool may_create_new_stripe(struct bch_fs *c)
1118 {
1119         return false;
1120 }
1121
1122 static void ec_stripe_key_init(struct bch_fs *c,
1123                                struct bkey_i_stripe *s,
1124                                unsigned nr_data,
1125                                unsigned nr_parity,
1126                                unsigned stripe_size)
1127 {
1128         unsigned u64s;
1129
1130         bkey_stripe_init(&s->k_i);
1131         s->v.sectors                    = cpu_to_le16(stripe_size);
1132         s->v.algorithm                  = 0;
1133         s->v.nr_blocks                  = nr_data + nr_parity;
1134         s->v.nr_redundant               = nr_parity;
1135         s->v.csum_granularity_bits      = ilog2(c->sb.encoded_extent_max);
1136         s->v.csum_type                  = BCH_CSUM_CRC32C;
1137         s->v.pad                        = 0;
1138
1139         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1140                 BUG_ON(1 << s->v.csum_granularity_bits >=
1141                        le16_to_cpu(s->v.sectors) ||
1142                        s->v.csum_granularity_bits == U8_MAX);
1143                 s->v.csum_granularity_bits++;
1144         }
1145
1146         set_bkey_val_u64s(&s->k, u64s);
1147 }
1148
1149 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1150 {
1151         struct ec_stripe_new *s;
1152
1153         lockdep_assert_held(&h->lock);
1154
1155         s = kzalloc(sizeof(*s), GFP_KERNEL);
1156         if (!s)
1157                 return -ENOMEM;
1158
1159         mutex_init(&s->lock);
1160         closure_init(&s->iodone, NULL);
1161         atomic_set(&s->pin, 1);
1162         s->c            = c;
1163         s->h            = h;
1164         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1165                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1166         s->nr_parity    = h->redundancy;
1167
1168         bch2_keylist_init(&s->keys, s->inline_keys);
1169
1170         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1171                            s->nr_parity, h->blocksize);
1172
1173         h->s = s;
1174         return 0;
1175 }
1176
1177 static struct ec_stripe_head *
1178 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1179                          unsigned algo, unsigned redundancy,
1180                          bool copygc)
1181 {
1182         struct ec_stripe_head *h;
1183         struct bch_dev *ca;
1184         unsigned i;
1185
1186         h = kzalloc(sizeof(*h), GFP_KERNEL);
1187         if (!h)
1188                 return NULL;
1189
1190         mutex_init(&h->lock);
1191         mutex_lock(&h->lock);
1192
1193         h->target       = target;
1194         h->algo         = algo;
1195         h->redundancy   = redundancy;
1196         h->copygc       = copygc;
1197
1198         rcu_read_lock();
1199         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1200
1201         for_each_member_device_rcu(ca, c, i, &h->devs)
1202                 if (!ca->mi.durability)
1203                         __clear_bit(i, h->devs.d);
1204
1205         h->blocksize = pick_blocksize(c, &h->devs);
1206
1207         for_each_member_device_rcu(ca, c, i, &h->devs)
1208                 if (ca->mi.bucket_size == h->blocksize)
1209                         h->nr_active_devs++;
1210
1211         rcu_read_unlock();
1212         list_add(&h->list, &c->ec_stripe_head_list);
1213         return h;
1214 }
1215
1216 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1217 {
1218         if (h->s &&
1219             h->s->allocated &&
1220             bitmap_weight(h->s->blocks_allocated,
1221                           h->s->nr_data) == h->s->nr_data)
1222                 ec_stripe_set_pending(c, h);
1223
1224         mutex_unlock(&h->lock);
1225 }
1226
1227 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1228                                                  unsigned target,
1229                                                  unsigned algo,
1230                                                  unsigned redundancy,
1231                                                  bool copygc)
1232 {
1233         struct ec_stripe_head *h;
1234
1235         if (!redundancy)
1236                 return NULL;
1237
1238         mutex_lock(&c->ec_stripe_head_lock);
1239         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1240                 if (h->target           == target &&
1241                     h->algo             == algo &&
1242                     h->redundancy       == redundancy &&
1243                     h->copygc           == copygc) {
1244                         mutex_lock(&h->lock);
1245                         goto found;
1246                 }
1247
1248         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1249 found:
1250         mutex_unlock(&c->ec_stripe_head_lock);
1251         return h;
1252 }
1253
1254 static enum bucket_alloc_ret
1255 new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1256                          struct closure *cl)
1257 {
1258         struct bch_devs_mask devs = h->devs;
1259         struct open_bucket *ob;
1260         struct open_buckets buckets;
1261         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1262         bool have_cache = true;
1263         enum bucket_alloc_ret ret = ALLOC_SUCCESS;
1264
1265         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1266                 if (test_bit(i, h->s->blocks_gotten)) {
1267                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1268                         if (i < h->s->nr_data)
1269                                 nr_have_data++;
1270                         else
1271                                 nr_have_parity++;
1272                 }
1273         }
1274
1275         BUG_ON(nr_have_data     > h->s->nr_data);
1276         BUG_ON(nr_have_parity   > h->s->nr_parity);
1277
1278         percpu_down_read(&c->mark_lock);
1279         rcu_read_lock();
1280
1281         buckets.nr = 0;
1282         if (nr_have_parity < h->s->nr_parity) {
1283                 ret = bch2_bucket_alloc_set(c, &buckets,
1284                                             &h->parity_stripe,
1285                                             &devs,
1286                                             h->s->nr_parity,
1287                                             &nr_have_parity,
1288                                             &have_cache,
1289                                             h->copygc
1290                                             ? RESERVE_MOVINGGC
1291                                             : RESERVE_NONE,
1292                                             0,
1293                                             cl);
1294
1295                 open_bucket_for_each(c, &buckets, ob, i) {
1296                         j = find_next_zero_bit(h->s->blocks_gotten,
1297                                                h->s->nr_data + h->s->nr_parity,
1298                                                h->s->nr_data);
1299                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1300
1301                         h->s->blocks[j] = buckets.v[i];
1302                         h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1303                         __set_bit(j, h->s->blocks_gotten);
1304                 }
1305
1306                 if (ret)
1307                         goto err;
1308         }
1309
1310         buckets.nr = 0;
1311         if (nr_have_data < h->s->nr_data) {
1312                 ret = bch2_bucket_alloc_set(c, &buckets,
1313                                             &h->block_stripe,
1314                                             &devs,
1315                                             h->s->nr_data,
1316                                             &nr_have_data,
1317                                             &have_cache,
1318                                             h->copygc
1319                                             ? RESERVE_MOVINGGC
1320                                             : RESERVE_NONE,
1321                                             0,
1322                                             cl);
1323
1324                 open_bucket_for_each(c, &buckets, ob, i) {
1325                         j = find_next_zero_bit(h->s->blocks_gotten,
1326                                                h->s->nr_data, 0);
1327                         BUG_ON(j >= h->s->nr_data);
1328
1329                         h->s->blocks[j] = buckets.v[i];
1330                         h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1331                         __set_bit(j, h->s->blocks_gotten);
1332                 }
1333
1334                 if (ret)
1335                         goto err;
1336         }
1337 err:
1338         rcu_read_unlock();
1339         percpu_up_read(&c->mark_lock);
1340         return ret;
1341 }
1342
1343 /* XXX: doesn't obey target: */
1344 static s64 get_existing_stripe(struct bch_fs *c,
1345                                struct ec_stripe_head *head)
1346 {
1347         ec_stripes_heap *h = &c->ec_stripes_heap;
1348         struct stripe *m;
1349         size_t heap_idx;
1350         u64 stripe_idx;
1351
1352         if (may_create_new_stripe(c))
1353                 return -1;
1354
1355         spin_lock(&c->ec_stripes_heap_lock);
1356         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1357                 if (!h->data[heap_idx].blocks_nonempty)
1358                         continue;
1359
1360                 stripe_idx = h->data[heap_idx].idx;
1361                 m = genradix_ptr(&c->stripes[0], stripe_idx);
1362
1363                 if (m->algorithm        == head->algo &&
1364                     m->nr_redundant     == head->redundancy &&
1365                     m->sectors          == head->blocksize &&
1366                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1367                         bch2_stripes_heap_del(c, m, stripe_idx);
1368                         spin_unlock(&c->ec_stripes_heap_lock);
1369                         return stripe_idx;
1370                 }
1371         }
1372
1373         spin_unlock(&c->ec_stripes_heap_lock);
1374         return -1;
1375 }
1376
1377 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1378                                                unsigned target,
1379                                                unsigned algo,
1380                                                unsigned redundancy,
1381                                                bool copygc,
1382                                                struct closure *cl)
1383 {
1384         struct ec_stripe_head *h;
1385         unsigned i;
1386         s64 idx;
1387         int ret;
1388
1389         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1390         if (!h) {
1391                 bch_err(c, "no stripe head");
1392                 return NULL;
1393         }
1394
1395         if (!h->s) {
1396                 if (ec_new_stripe_alloc(c, h)) {
1397                         bch2_ec_stripe_head_put(c, h);
1398                         bch_err(c, "failed to allocate new stripe");
1399                         return NULL;
1400                 }
1401
1402                 idx = get_existing_stripe(c, h);
1403                 if (idx >= 0) {
1404                         h->s->have_existing_stripe = true;
1405                         ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1406                         if (ret) {
1407                                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1408                                 bch2_ec_stripe_head_put(c, h);
1409                                 return NULL;
1410                         }
1411
1412                         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1413                                 /*
1414                                  * this is a problem: we have deleted from the
1415                                  * stripes heap already
1416                                  */
1417                                 BUG();
1418                         }
1419
1420                         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1421                         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1422
1423                         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1424                                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1425                                         __set_bit(i, h->s->blocks_gotten);
1426                                         __set_bit(i, h->s->blocks_allocated);
1427                                 }
1428
1429                                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1430                         }
1431
1432                         bkey_copy(&h->s->new_stripe.key.k_i,
1433                                   &h->s->existing_stripe.key.k_i);
1434                 }
1435
1436                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize)) {
1437                         BUG();
1438                 }
1439         }
1440
1441         if (!h->s->allocated) {
1442                 if (!h->s->have_existing_stripe &&
1443                     !h->s->res.sectors) {
1444                         ret = bch2_disk_reservation_get(c, &h->s->res,
1445                                         h->blocksize,
1446                                         h->s->nr_parity, 0);
1447                         if (ret) {
1448                                 /*
1449                                  * This means we need to wait for copygc to
1450                                  * empty out buckets from existing stripes:
1451                                  */
1452                                 bch2_ec_stripe_head_put(c, h);
1453                                 h = NULL;
1454                                 goto out;
1455                         }
1456                 }
1457
1458                 ret = new_stripe_alloc_buckets(c, h, cl);
1459                 if (ret) {
1460                         bch2_ec_stripe_head_put(c, h);
1461                         h = ERR_PTR(-ret);
1462                         goto out;
1463                 }
1464
1465                 h->s->allocated = true;
1466         }
1467 out:
1468         return h;
1469 }
1470
1471 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1472 {
1473         struct ec_stripe_head *h;
1474         struct open_bucket *ob;
1475         unsigned i;
1476
1477         mutex_lock(&c->ec_stripe_head_lock);
1478         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1479
1480                 mutex_lock(&h->lock);
1481                 if (!h->s)
1482                         goto unlock;
1483
1484                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1485                         if (!h->s->blocks[i])
1486                                 continue;
1487
1488                         ob = c->open_buckets + h->s->blocks[i];
1489                         if (ob->ptr.dev == ca->dev_idx)
1490                                 goto found;
1491                 }
1492                 goto unlock;
1493 found:
1494                 h->s->err = -EROFS;
1495                 ec_stripe_set_pending(c, h);
1496 unlock:
1497                 mutex_unlock(&h->lock);
1498         }
1499         mutex_unlock(&c->ec_stripe_head_lock);
1500 }
1501
1502 void bch2_stripes_heap_start(struct bch_fs *c)
1503 {
1504         struct genradix_iter iter;
1505         struct stripe *m;
1506
1507         genradix_for_each(&c->stripes[0], iter, m)
1508                 if (m->alive)
1509                         bch2_stripes_heap_insert(c, m, iter.pos);
1510 }
1511
1512 static int __bch2_stripe_write_key(struct btree_trans *trans,
1513                                    struct btree_iter *iter,
1514                                    struct stripe *m,
1515                                    size_t idx,
1516                                    struct bkey_i_stripe *new_key)
1517 {
1518         const struct bch_stripe *v;
1519         struct bkey_s_c k;
1520         unsigned i;
1521         int ret;
1522
1523         bch2_btree_iter_set_pos(iter, POS(0, idx));
1524
1525         k = bch2_btree_iter_peek_slot(iter);
1526         ret = bkey_err(k);
1527         if (ret)
1528                 return ret;
1529
1530         if (k.k->type != KEY_TYPE_stripe)
1531                 return -EIO;
1532
1533         v = bkey_s_c_to_stripe(k).v;
1534         for (i = 0; i < v->nr_blocks; i++)
1535                 if (m->block_sectors[i] != stripe_blockcount_get(v, i))
1536                         goto write;
1537         return 0;
1538 write:
1539         bkey_reassemble(&new_key->k_i, k);
1540
1541         for (i = 0; i < new_key->v.nr_blocks; i++)
1542                 stripe_blockcount_set(&new_key->v, i,
1543                                       m->block_sectors[i]);
1544
1545         bch2_trans_update(trans, iter, &new_key->k_i, 0);
1546         return 0;
1547 }
1548
1549 int bch2_stripes_write(struct bch_fs *c, unsigned flags)
1550 {
1551         struct btree_trans trans;
1552         struct btree_iter *iter;
1553         struct genradix_iter giter;
1554         struct bkey_i_stripe *new_key;
1555         struct stripe *m;
1556         int ret = 0;
1557
1558         new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
1559         BUG_ON(!new_key);
1560
1561         bch2_trans_init(&trans, c, 0, 0);
1562
1563         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS_MIN,
1564                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1565
1566         genradix_for_each(&c->stripes[0], giter, m) {
1567                 if (!m->alive)
1568                         continue;
1569
1570                 ret = __bch2_trans_do(&trans, NULL, NULL,
1571                                       BTREE_INSERT_NOFAIL|flags,
1572                         __bch2_stripe_write_key(&trans, iter, m,
1573                                         giter.pos, new_key));
1574
1575                 if (ret)
1576                         break;
1577         }
1578
1579         bch2_trans_exit(&trans);
1580
1581         kfree(new_key);
1582
1583         return ret;
1584 }
1585
1586 static int bch2_stripes_read_fn(struct bch_fs *c, enum btree_id id,
1587                               unsigned level, struct bkey_s_c k)
1588 {
1589         int ret = 0;
1590
1591         if (k.k->type == KEY_TYPE_stripe) {
1592                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL) ?:
1593                         bch2_mark_key(c, k, 0, 0, NULL, 0,
1594                                       BTREE_TRIGGER_NOATOMIC);
1595                 if (ret)
1596                         return ret;
1597         }
1598
1599         return ret;
1600 }
1601
1602 int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
1603 {
1604         int ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_EC,
1605                                           NULL, bch2_stripes_read_fn);
1606         if (ret)
1607                 bch_err(c, "error reading stripes: %i", ret);
1608
1609         return ret;
1610 }
1611
1612 int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
1613 {
1614         struct btree_trans trans;
1615         struct btree_iter *iter;
1616         struct bkey_s_c k;
1617         size_t i, idx = 0;
1618         int ret = 0;
1619
1620         bch2_trans_init(&trans, c, 0, 0);
1621
1622         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS(0, U64_MAX), 0);
1623
1624         k = bch2_btree_iter_prev(iter);
1625         if (!IS_ERR_OR_NULL(k.k))
1626                 idx = k.k->p.offset + 1;
1627         ret = bch2_trans_exit(&trans);
1628         if (ret)
1629                 return ret;
1630
1631         if (!idx)
1632                 return 0;
1633
1634         if (!gc &&
1635             !init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
1636                        GFP_KERNEL))
1637                 return -ENOMEM;
1638 #if 0
1639         ret = genradix_prealloc(&c->stripes[gc], idx, GFP_KERNEL);
1640 #else
1641         for (i = 0; i < idx; i++)
1642                 if (!genradix_ptr_alloc(&c->stripes[gc], i, GFP_KERNEL))
1643                         return -ENOMEM;
1644 #endif
1645         return 0;
1646 }
1647
1648 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1649 {
1650         ec_stripes_heap *h = &c->ec_stripes_heap;
1651         struct stripe *m;
1652         size_t i;
1653
1654         spin_lock(&c->ec_stripes_heap_lock);
1655         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1656                 m = genradix_ptr(&c->stripes[0], h->data[i].idx);
1657
1658                 pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1659                        h->data[i].blocks_nonempty,
1660                        m->nr_blocks - m->nr_redundant,
1661                        m->nr_redundant);
1662         }
1663         spin_unlock(&c->ec_stripes_heap_lock);
1664 }
1665
1666 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1667 {
1668         struct ec_stripe_head *h;
1669         struct ec_stripe_new *s;
1670
1671         mutex_lock(&c->ec_stripe_head_lock);
1672         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1673                 pr_buf(out, "target %u algo %u redundancy %u:\n",
1674                        h->target, h->algo, h->redundancy);
1675
1676                 if (h->s)
1677                         pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
1678                                h->s->nr_data, h->s->nr_parity,
1679                                bitmap_weight(h->s->blocks_allocated,
1680                                              h->s->nr_data));
1681         }
1682         mutex_unlock(&c->ec_stripe_head_lock);
1683
1684         mutex_lock(&c->ec_stripe_new_lock);
1685         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1686                 pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
1687                        s->nr_data, s->nr_parity,
1688                        atomic_read(&s->pin));
1689         }
1690         mutex_unlock(&c->ec_stripe_new_lock);
1691 }
1692
1693 void bch2_fs_ec_exit(struct bch_fs *c)
1694 {
1695         struct ec_stripe_head *h;
1696
1697         while (1) {
1698                 mutex_lock(&c->ec_stripe_head_lock);
1699                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1700                                              struct ec_stripe_head, list);
1701                 if (h)
1702                         list_del(&h->list);
1703                 mutex_unlock(&c->ec_stripe_head_lock);
1704                 if (!h)
1705                         break;
1706
1707                 BUG_ON(h->s);
1708                 kfree(h);
1709         }
1710
1711         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1712
1713         free_heap(&c->ec_stripes_heap);
1714         genradix_free(&c->stripes[0]);
1715         bioset_exit(&c->ec_bioset);
1716 }
1717
1718 int bch2_fs_ec_init(struct bch_fs *c)
1719 {
1720         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1721         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1722
1723         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1724                            BIOSET_NEED_BVECS);
1725 }