1 // SPDX-License-Identifier: GPL-2.0
6 #include "alloc_foreground.h"
10 #include "btree_update.h"
12 #include "disk_groups.h"
21 #include <linux/sort.h>
25 #include <linux/raid/pq.h>
26 #include <linux/raid/xor.h>
28 static void raid5_recov(unsigned disks, unsigned failed_idx,
29 size_t size, void **data)
33 BUG_ON(failed_idx >= disks);
35 swap(data[0], data[failed_idx]);
36 memcpy(data[0], data[1], size);
39 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
40 xor_blocks(nr, size, data[0], data + i);
44 swap(data[0], data[failed_idx]);
47 static void raid_gen(int nd, int np, size_t size, void **v)
50 raid5_recov(nd + np, nd, size, v);
52 raid6_call.gen_syndrome(nd + np, size, v);
56 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
63 raid5_recov(nd + 1, ir[0], size, v);
65 raid6_call.gen_syndrome(nd + np, size, v);
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 */
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);
81 raid_gen(nd, np, size, v);
91 #include <raid/raid.h>
97 struct ec_stripe_buf *buf;
102 /* Stripes btree keys: */
104 const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
106 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
108 if (!bkey_cmp(k.k->p, POS_MIN))
109 return "stripe at pos 0";
112 return "invalid stripe key";
114 if (bkey_val_bytes(k.k) < sizeof(*s))
115 return "incorrect value size";
117 if (bkey_val_bytes(k.k) < sizeof(*s) ||
118 bkey_val_u64s(k.k) < stripe_val_u64s(s))
119 return "incorrect value size";
121 return bch2_bkey_ptrs_invalid(c, k);
124 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
127 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
130 pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
132 le16_to_cpu(s->sectors),
133 s->nr_blocks - s->nr_redundant,
136 1U << s->csum_granularity_bits);
138 for (i = 0; i < s->nr_blocks; i++)
139 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
140 (u64) s->ptrs[i].offset,
141 stripe_blockcount_get(s, i));
144 /* returns blocknr in stripe that we matched: */
145 static int bkey_matches_stripe(struct bch_stripe *s,
148 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
149 const struct bch_extent_ptr *ptr;
150 unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
152 bkey_for_each_ptr(ptrs, ptr)
153 for (i = 0; i < nr_data; i++)
154 if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
155 le16_to_cpu(s->sectors)))
161 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
164 case KEY_TYPE_extent: {
165 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
166 const union bch_extent_entry *entry;
168 extent_for_each_entry(e, entry)
169 if (extent_entry_type(entry) ==
170 BCH_EXTENT_ENTRY_stripe_ptr &&
171 entry->stripe_ptr.idx == idx)
183 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
187 for (i = 0; i < buf->key.v.nr_blocks; i++) {
188 kvpfree(buf->data[i], buf->size << 9);
193 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
194 unsigned offset, unsigned size)
196 struct bch_stripe *v = &buf->key.v;
197 unsigned csum_granularity = 1U << v->csum_granularity_bits;
198 unsigned end = offset + size;
201 BUG_ON(end > le16_to_cpu(v->sectors));
203 offset = round_down(offset, csum_granularity);
204 end = min_t(unsigned, le16_to_cpu(v->sectors),
205 round_up(end, csum_granularity));
207 buf->offset = offset;
208 buf->size = end - offset;
210 memset(buf->valid, 0xFF, sizeof(buf->valid));
212 for (i = 0; i < buf->key.v.nr_blocks; i++) {
213 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
220 ec_stripe_buf_exit(buf);
226 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
227 unsigned block, unsigned offset)
229 struct bch_stripe *v = &buf->key.v;
230 unsigned csum_granularity = 1 << v->csum_granularity_bits;
231 unsigned end = buf->offset + buf->size;
232 unsigned len = min(csum_granularity, end - offset);
234 BUG_ON(offset >= end);
235 BUG_ON(offset < buf->offset);
236 BUG_ON(offset & (csum_granularity - 1));
237 BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
238 (len & (csum_granularity - 1)));
240 return bch2_checksum(NULL, v->csum_type,
242 buf->data[block] + ((offset - buf->offset) << 9),
246 static void ec_generate_checksums(struct ec_stripe_buf *buf)
248 struct bch_stripe *v = &buf->key.v;
249 unsigned i, j, csums_per_device = stripe_csums_per_device(v);
255 BUG_ON(buf->size != le16_to_cpu(v->sectors));
257 for (i = 0; i < v->nr_blocks; i++)
258 for (j = 0; j < csums_per_device; j++)
259 stripe_csum_set(v, i, j,
260 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
263 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
265 struct bch_stripe *v = &buf->key.v;
266 unsigned csum_granularity = 1 << v->csum_granularity_bits;
272 for (i = 0; i < v->nr_blocks; i++) {
273 unsigned offset = buf->offset;
274 unsigned end = buf->offset + buf->size;
276 if (!test_bit(i, buf->valid))
279 while (offset < end) {
280 unsigned j = offset >> v->csum_granularity_bits;
281 unsigned len = min(csum_granularity, end - offset);
282 struct bch_csum want = stripe_csum_get(v, i, j);
283 struct bch_csum got = ec_block_checksum(buf, i, offset);
285 if (bch2_crc_cmp(want, got)) {
288 bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&buf->key.k_i));
290 bch_err_ratelimited(c,
291 "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
292 (void *) _RET_IP_, i, j, v->csum_type,
293 want.lo, got.lo, buf2);
294 clear_bit(i, buf->valid);
303 /* Erasure coding: */
305 static void ec_generate_ec(struct ec_stripe_buf *buf)
307 struct bch_stripe *v = &buf->key.v;
308 unsigned nr_data = v->nr_blocks - v->nr_redundant;
309 unsigned bytes = le16_to_cpu(v->sectors) << 9;
311 raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
314 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
316 return buf->key.v.nr_blocks -
317 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
320 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
322 struct bch_stripe *v = &buf->key.v;
323 unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
324 unsigned nr_data = v->nr_blocks - v->nr_redundant;
325 unsigned bytes = buf->size << 9;
327 if (ec_nr_failed(buf) > v->nr_redundant) {
328 bch_err_ratelimited(c,
329 "error doing reconstruct read: unable to read enough blocks");
333 for (i = 0; i < nr_data; i++)
334 if (!test_bit(i, buf->valid))
335 failed[nr_failed++] = i;
337 raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
343 static void ec_block_endio(struct bio *bio)
345 struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
346 struct bch_stripe *v = &ec_bio->buf->key.v;
347 struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
348 struct bch_dev *ca = ec_bio->ca;
349 struct closure *cl = bio->bi_private;
351 if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
352 bio_data_dir(bio) ? "write" : "read",
353 bch2_blk_status_to_str(bio->bi_status)))
354 clear_bit(ec_bio->idx, ec_bio->buf->valid);
356 if (ptr_stale(ca, ptr)) {
357 bch_err_ratelimited(ca->fs,
358 "error %s stripe: stale pointer after io",
359 bio_data_dir(bio) == READ ? "reading from" : "writing to");
360 clear_bit(ec_bio->idx, ec_bio->buf->valid);
363 bio_put(&ec_bio->bio);
364 percpu_ref_put(&ca->io_ref);
368 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
369 unsigned rw, unsigned idx, struct closure *cl)
371 struct bch_stripe *v = &buf->key.v;
372 unsigned offset = 0, bytes = buf->size << 9;
373 struct bch_extent_ptr *ptr = &v->ptrs[idx];
374 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
375 enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
379 if (ptr_stale(ca, ptr)) {
380 bch_err_ratelimited(c,
381 "error %s stripe: stale pointer",
382 rw == READ ? "reading from" : "writing to");
383 clear_bit(idx, buf->valid);
387 if (!bch2_dev_get_ioref(ca, rw)) {
388 clear_bit(idx, buf->valid);
392 this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
394 while (offset < bytes) {
395 unsigned nr_iovecs = min_t(size_t, BIO_MAX_PAGES,
396 DIV_ROUND_UP(bytes, PAGE_SIZE));
397 unsigned b = min_t(size_t, bytes - offset,
398 nr_iovecs << PAGE_SHIFT);
399 struct ec_bio *ec_bio;
401 ec_bio = container_of(bio_alloc_bioset(GFP_KERNEL, nr_iovecs,
409 bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
410 bio_set_op_attrs(&ec_bio->bio, rw, 0);
412 ec_bio->bio.bi_iter.bi_sector = ptr->offset + buf->offset + (offset >> 9);
413 ec_bio->bio.bi_end_io = ec_block_endio;
414 ec_bio->bio.bi_private = cl;
416 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
419 percpu_ref_get(&ca->io_ref);
421 submit_bio(&ec_bio->bio);
426 percpu_ref_put(&ca->io_ref);
429 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
431 struct btree_trans trans;
432 struct btree_iter *iter;
436 bch2_trans_init(&trans, c, 0, 0);
437 iter = bch2_trans_get_iter(&trans, BTREE_ID_stripes, POS(0, idx), BTREE_ITER_SLOTS);
438 k = bch2_btree_iter_peek_slot(iter);
442 if (k.k->type != KEY_TYPE_stripe) {
446 bkey_reassemble(&stripe->key.k_i, k);
448 bch2_trans_exit(&trans);
452 /* recovery read path: */
453 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
455 struct ec_stripe_buf *buf;
457 struct bch_stripe *v;
461 closure_init_stack(&cl);
463 BUG_ON(!rbio->pick.has_ec);
465 buf = kzalloc(sizeof(*buf), GFP_NOIO);
469 ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
471 bch_err_ratelimited(c,
472 "error doing reconstruct read: error %i looking up stripe", ret);
479 if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
480 bch_err_ratelimited(c,
481 "error doing reconstruct read: pointer doesn't match stripe");
486 offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
487 if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
488 bch_err_ratelimited(c,
489 "error doing reconstruct read: read is bigger than stripe");
494 ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
498 for (i = 0; i < v->nr_blocks; i++)
499 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
503 if (ec_nr_failed(buf) > v->nr_redundant) {
504 bch_err_ratelimited(c,
505 "error doing reconstruct read: unable to read enough blocks");
510 ec_validate_checksums(c, buf);
512 ret = ec_do_recov(c, buf);
516 memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
517 buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
519 ec_stripe_buf_exit(buf);
524 /* stripe bucket accounting: */
526 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
528 ec_stripes_heap n, *h = &c->ec_stripes_heap;
530 if (idx >= h->size) {
531 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
534 spin_lock(&c->ec_stripes_heap_lock);
535 if (n.size > h->size) {
536 memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
540 spin_unlock(&c->ec_stripes_heap_lock);
545 if (!genradix_ptr_alloc(&c->stripes[0], idx, gfp))
548 if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
549 !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
555 static int ec_stripe_mem_alloc(struct bch_fs *c,
556 struct btree_iter *iter)
558 size_t idx = iter->pos.offset;
561 if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT|__GFP_NOWARN))
564 bch2_trans_unlock(iter->trans);
567 if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
573 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
575 ec_stripes_heap *h = &c->ec_stripes_heap;
577 return h->used && h->data[0].blocks_nonempty == 0
578 ? h->data[0].idx : -1;
581 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
582 struct ec_stripe_heap_entry l,
583 struct ec_stripe_heap_entry r)
585 return ((l.blocks_nonempty > r.blocks_nonempty) -
586 (l.blocks_nonempty < r.blocks_nonempty));
589 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
592 struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
594 genradix_ptr(&c->stripes[0], h->data[i].idx)->heap_idx = i;
597 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
599 ec_stripes_heap *h = &c->ec_stripes_heap;
600 struct stripe *m = genradix_ptr(&c->stripes[0], idx);
603 BUG_ON(m->heap_idx >= h->used);
604 BUG_ON(h->data[m->heap_idx].idx != idx);
607 void bch2_stripes_heap_del(struct bch_fs *c,
608 struct stripe *m, size_t idx)
615 heap_verify_backpointer(c, idx);
617 heap_del(&c->ec_stripes_heap, m->heap_idx,
619 ec_stripes_heap_set_backpointer);
622 void bch2_stripes_heap_insert(struct bch_fs *c,
623 struct stripe *m, size_t idx)
628 BUG_ON(heap_full(&c->ec_stripes_heap));
632 heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
634 .blocks_nonempty = m->blocks_nonempty,
637 ec_stripes_heap_set_backpointer);
639 heap_verify_backpointer(c, idx);
642 void bch2_stripes_heap_update(struct bch_fs *c,
643 struct stripe *m, size_t idx)
645 ec_stripes_heap *h = &c->ec_stripes_heap;
651 heap_verify_backpointer(c, idx);
653 h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
656 heap_sift_up(h, i, ec_stripes_heap_cmp,
657 ec_stripes_heap_set_backpointer);
658 heap_sift_down(h, i, ec_stripes_heap_cmp,
659 ec_stripes_heap_set_backpointer);
661 heap_verify_backpointer(c, idx);
663 if (stripe_idx_to_delete(c) >= 0 &&
664 !percpu_ref_is_dying(&c->writes))
665 schedule_work(&c->ec_stripe_delete_work);
668 /* stripe deletion */
670 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
672 return bch2_btree_delete_range(c, BTREE_ID_stripes,
678 static void ec_stripe_delete_work(struct work_struct *work)
681 container_of(work, struct bch_fs, ec_stripe_delete_work);
685 spin_lock(&c->ec_stripes_heap_lock);
686 idx = stripe_idx_to_delete(c);
688 spin_unlock(&c->ec_stripes_heap_lock);
692 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes[0], idx), idx);
693 spin_unlock(&c->ec_stripes_heap_lock);
695 if (ec_stripe_delete(c, idx))
700 /* stripe creation: */
702 static int ec_stripe_bkey_insert(struct bch_fs *c,
703 struct bkey_i_stripe *stripe,
704 struct disk_reservation *res)
706 struct btree_trans trans;
707 struct btree_iter *iter;
709 struct bpos min_pos = POS(0, 1);
710 struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
713 bch2_trans_init(&trans, c, 0, 0);
715 bch2_trans_begin(&trans);
717 for_each_btree_key(&trans, iter, BTREE_ID_stripes, start_pos,
718 BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
719 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
720 if (start_pos.offset) {
722 bch2_btree_iter_set_pos(iter, start_pos);
730 if (bkey_deleted(k.k))
736 start_pos = iter->pos;
738 ret = ec_stripe_mem_alloc(c, iter);
742 stripe->k.p = iter->pos;
744 bch2_trans_update(&trans, iter, &stripe->k_i, 0);
746 ret = bch2_trans_commit(&trans, res, NULL,
747 BTREE_INSERT_NOFAIL);
749 bch2_trans_iter_put(&trans, iter);
754 c->ec_stripe_hint = ret ? start_pos.offset : start_pos.offset + 1;
755 bch2_trans_exit(&trans);
760 static int ec_stripe_bkey_update(struct btree_trans *trans,
761 struct bkey_i_stripe *new)
763 struct btree_iter *iter;
765 const struct bch_stripe *existing;
769 iter = bch2_trans_get_iter(trans, BTREE_ID_stripes,
770 new->k.p, BTREE_ITER_INTENT);
771 k = bch2_btree_iter_peek_slot(iter);
776 if (!k.k || k.k->type != KEY_TYPE_stripe) {
777 bch_err(trans->c, "error updating stripe: not found");
782 existing = bkey_s_c_to_stripe(k).v;
784 if (existing->nr_blocks != new->v.nr_blocks) {
785 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
790 for (i = 0; i < new->v.nr_blocks; i++)
791 stripe_blockcount_set(&new->v, i,
792 stripe_blockcount_get(existing, i));
794 bch2_trans_update(trans, iter, &new->k_i, 0);
796 bch2_trans_iter_put(trans, iter);
800 static void extent_stripe_ptr_add(struct bkey_s_extent e,
801 struct ec_stripe_buf *s,
802 struct bch_extent_ptr *ptr,
805 struct bch_extent_stripe_ptr *dst = (void *) ptr;
806 union bch_extent_entry *end = extent_entry_last(e);
808 memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
809 e.k->u64s += sizeof(*dst) / sizeof(u64);
811 *dst = (struct bch_extent_stripe_ptr) {
812 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
814 .redundancy = s->key.v.nr_redundant,
815 .idx = s->key.k.p.offset,
819 static int ec_stripe_update_ptrs(struct bch_fs *c,
820 struct ec_stripe_buf *s,
823 struct btree_trans trans;
824 struct btree_iter *iter;
826 struct bkey_s_extent e;
828 int ret = 0, dev, block;
830 bch2_bkey_buf_init(&sk);
831 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
833 /* XXX this doesn't support the reflink btree */
835 iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
839 while ((k = bch2_btree_iter_peek(iter)).k &&
840 !(ret = bkey_err(k)) &&
841 bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
842 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
844 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
845 bch2_btree_iter_next(iter);
849 block = bkey_matches_stripe(&s->key.v, k);
851 bch2_btree_iter_next(iter);
855 dev = s->key.v.ptrs[block].dev;
857 bch2_bkey_buf_reassemble(&sk, c, k);
858 e = bkey_i_to_s_extent(sk.k);
860 bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
861 ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
864 extent_stripe_ptr_add(e, s, ec_ptr, block);
866 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
867 bch2_trans_update(&trans, iter, sk.k, 0);
869 ret = bch2_trans_commit(&trans, NULL, NULL,
870 BTREE_INSERT_NOFAIL);
877 bch2_trans_exit(&trans);
878 bch2_bkey_buf_exit(&sk, c);
884 * data buckets of new stripe all written: create the stripe
886 static void ec_stripe_create(struct ec_stripe_new *s)
888 struct bch_fs *c = s->c;
889 struct open_bucket *ob;
892 struct bch_stripe *v = &s->new_stripe.key.v;
893 unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
896 BUG_ON(s->h->s == s);
898 closure_sync(&s->iodone);
901 if (s->err != -EROFS)
902 bch_err(c, "error creating stripe: error writing data buckets");
906 if (s->have_existing_stripe) {
907 ec_validate_checksums(c, &s->existing_stripe);
909 if (ec_do_recov(c, &s->existing_stripe)) {
910 bch_err(c, "error creating stripe: error reading existing stripe");
914 for (i = 0; i < nr_data; i++)
915 if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
916 swap(s->new_stripe.data[i],
917 s->existing_stripe.data[i]);
919 ec_stripe_buf_exit(&s->existing_stripe);
922 BUG_ON(!s->allocated);
924 if (!percpu_ref_tryget(&c->writes))
927 ec_generate_ec(&s->new_stripe);
929 ec_generate_checksums(&s->new_stripe);
932 for (i = nr_data; i < v->nr_blocks; i++)
933 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
934 closure_sync(&s->iodone);
936 if (ec_nr_failed(&s->new_stripe)) {
937 bch_err(c, "error creating stripe: error writing redundancy buckets");
941 ret = s->have_existing_stripe
942 ? bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
943 ec_stripe_bkey_update(&trans, &s->new_stripe.key))
944 : ec_stripe_bkey_insert(c, &s->new_stripe.key, &s->res);
946 bch_err(c, "error creating stripe: error creating stripe key");
950 for_each_keylist_key(&s->keys, k) {
951 ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
953 bch_err(c, "error creating stripe: error %i updating pointers", ret);
958 spin_lock(&c->ec_stripes_heap_lock);
959 m = genradix_ptr(&c->stripes[0], s->new_stripe.key.k.p.offset);
962 bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
963 spin_unlock(&c->ec_stripes_heap_lock);
965 percpu_ref_put(&c->writes);
967 bch2_disk_reservation_put(c, &s->res);
969 for (i = 0; i < v->nr_blocks; i++)
971 ob = c->open_buckets + s->blocks[i];
975 __bch2_open_bucket_put(c, ob);
977 bch2_open_bucket_put(c, ob);
981 bch2_keylist_free(&s->keys, s->inline_keys);
983 ec_stripe_buf_exit(&s->existing_stripe);
984 ec_stripe_buf_exit(&s->new_stripe);
985 closure_debug_destroy(&s->iodone);
989 static void ec_stripe_create_work(struct work_struct *work)
991 struct bch_fs *c = container_of(work,
992 struct bch_fs, ec_stripe_create_work);
993 struct ec_stripe_new *s, *n;
995 mutex_lock(&c->ec_stripe_new_lock);
996 list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
997 if (!atomic_read(&s->pin)) {
999 mutex_unlock(&c->ec_stripe_new_lock);
1000 ec_stripe_create(s);
1003 mutex_unlock(&c->ec_stripe_new_lock);
1006 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1008 BUG_ON(atomic_read(&s->pin) <= 0);
1010 if (atomic_dec_and_test(&s->pin)) {
1011 BUG_ON(!s->pending);
1012 queue_work(system_long_wq, &c->ec_stripe_create_work);
1016 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1018 struct ec_stripe_new *s = h->s;
1020 BUG_ON(!s->allocated && !s->err);
1025 mutex_lock(&c->ec_stripe_new_lock);
1026 list_add(&s->list, &c->ec_stripe_new_list);
1027 mutex_unlock(&c->ec_stripe_new_lock);
1029 ec_stripe_new_put(c, s);
1032 /* have a full bucket - hand it off to be erasure coded: */
1033 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1035 struct ec_stripe_new *s = ob->ec;
1037 if (ob->sectors_free)
1040 ec_stripe_new_put(c, s);
1043 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1045 struct ec_stripe_new *s = ob->ec;
1050 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1052 struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1059 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1060 offset = ca->mi.bucket_size - ob->sectors_free;
1062 return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1065 void bch2_ec_add_backpointer(struct bch_fs *c, struct write_point *wp,
1066 struct bpos pos, unsigned sectors)
1068 struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1069 struct ec_stripe_new *ec;
1075 mutex_lock(&ec->lock);
1077 if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1078 ARRAY_SIZE(ec->inline_keys),
1083 bkey_init(&ec->keys.top->k);
1084 ec->keys.top->k.p = pos;
1085 bch2_key_resize(&ec->keys.top->k, sectors);
1086 bch2_keylist_push(&ec->keys);
1088 mutex_unlock(&ec->lock);
1091 static int unsigned_cmp(const void *_l, const void *_r)
1093 unsigned l = *((const unsigned *) _l);
1094 unsigned r = *((const unsigned *) _r);
1096 return cmp_int(l, r);
1099 /* pick most common bucket size: */
1100 static unsigned pick_blocksize(struct bch_fs *c,
1101 struct bch_devs_mask *devs)
1104 unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1107 } cur = { 0, 0 }, best = { 0, 0 };
1109 for_each_member_device_rcu(ca, c, i, devs)
1110 sizes[nr++] = ca->mi.bucket_size;
1112 sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1114 for (i = 0; i < nr; i++) {
1115 if (sizes[i] != cur.size) {
1116 if (cur.nr > best.nr)
1120 cur.size = sizes[i];
1126 if (cur.nr > best.nr)
1132 static bool may_create_new_stripe(struct bch_fs *c)
1137 static void ec_stripe_key_init(struct bch_fs *c,
1138 struct bkey_i_stripe *s,
1141 unsigned stripe_size)
1145 bkey_stripe_init(&s->k_i);
1146 s->v.sectors = cpu_to_le16(stripe_size);
1148 s->v.nr_blocks = nr_data + nr_parity;
1149 s->v.nr_redundant = nr_parity;
1150 s->v.csum_granularity_bits = ilog2(c->sb.encoded_extent_max);
1151 s->v.csum_type = BCH_CSUM_CRC32C;
1154 while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1155 BUG_ON(1 << s->v.csum_granularity_bits >=
1156 le16_to_cpu(s->v.sectors) ||
1157 s->v.csum_granularity_bits == U8_MAX);
1158 s->v.csum_granularity_bits++;
1161 set_bkey_val_u64s(&s->k, u64s);
1164 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1166 struct ec_stripe_new *s;
1168 lockdep_assert_held(&h->lock);
1170 s = kzalloc(sizeof(*s), GFP_KERNEL);
1174 mutex_init(&s->lock);
1175 closure_init(&s->iodone, NULL);
1176 atomic_set(&s->pin, 1);
1179 s->nr_data = min_t(unsigned, h->nr_active_devs,
1180 BCH_BKEY_PTRS_MAX) - h->redundancy;
1181 s->nr_parity = h->redundancy;
1183 bch2_keylist_init(&s->keys, s->inline_keys);
1185 ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1186 s->nr_parity, h->blocksize);
1192 static struct ec_stripe_head *
1193 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1194 unsigned algo, unsigned redundancy,
1197 struct ec_stripe_head *h;
1201 h = kzalloc(sizeof(*h), GFP_KERNEL);
1205 mutex_init(&h->lock);
1206 mutex_lock(&h->lock);
1210 h->redundancy = redundancy;
1214 h->devs = target_rw_devs(c, BCH_DATA_user, target);
1216 for_each_member_device_rcu(ca, c, i, &h->devs)
1217 if (!ca->mi.durability)
1218 __clear_bit(i, h->devs.d);
1220 h->blocksize = pick_blocksize(c, &h->devs);
1222 for_each_member_device_rcu(ca, c, i, &h->devs)
1223 if (ca->mi.bucket_size == h->blocksize)
1224 h->nr_active_devs++;
1227 list_add(&h->list, &c->ec_stripe_head_list);
1231 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1235 bitmap_weight(h->s->blocks_allocated,
1236 h->s->nr_data) == h->s->nr_data)
1237 ec_stripe_set_pending(c, h);
1239 mutex_unlock(&h->lock);
1242 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1245 unsigned redundancy,
1248 struct ec_stripe_head *h;
1253 mutex_lock(&c->ec_stripe_head_lock);
1254 list_for_each_entry(h, &c->ec_stripe_head_list, list)
1255 if (h->target == target &&
1257 h->redundancy == redundancy &&
1258 h->copygc == copygc) {
1259 mutex_lock(&h->lock);
1263 h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1265 mutex_unlock(&c->ec_stripe_head_lock);
1269 static enum bucket_alloc_ret
1270 new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1273 struct bch_devs_mask devs = h->devs;
1274 struct open_bucket *ob;
1275 struct open_buckets buckets;
1276 unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1277 bool have_cache = true;
1278 enum bucket_alloc_ret ret = ALLOC_SUCCESS;
1280 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1281 if (test_bit(i, h->s->blocks_gotten)) {
1282 __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1283 if (i < h->s->nr_data)
1290 BUG_ON(nr_have_data > h->s->nr_data);
1291 BUG_ON(nr_have_parity > h->s->nr_parity);
1293 percpu_down_read(&c->mark_lock);
1297 if (nr_have_parity < h->s->nr_parity) {
1298 ret = bch2_bucket_alloc_set(c, &buckets,
1310 open_bucket_for_each(c, &buckets, ob, i) {
1311 j = find_next_zero_bit(h->s->blocks_gotten,
1312 h->s->nr_data + h->s->nr_parity,
1314 BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1316 h->s->blocks[j] = buckets.v[i];
1317 h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1318 __set_bit(j, h->s->blocks_gotten);
1326 if (nr_have_data < h->s->nr_data) {
1327 ret = bch2_bucket_alloc_set(c, &buckets,
1339 open_bucket_for_each(c, &buckets, ob, i) {
1340 j = find_next_zero_bit(h->s->blocks_gotten,
1342 BUG_ON(j >= h->s->nr_data);
1344 h->s->blocks[j] = buckets.v[i];
1345 h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1346 __set_bit(j, h->s->blocks_gotten);
1354 percpu_up_read(&c->mark_lock);
1358 /* XXX: doesn't obey target: */
1359 static s64 get_existing_stripe(struct bch_fs *c,
1360 struct ec_stripe_head *head)
1362 ec_stripes_heap *h = &c->ec_stripes_heap;
1368 if (may_create_new_stripe(c))
1371 spin_lock(&c->ec_stripes_heap_lock);
1372 for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1373 /* No blocks worth reusing, stripe will just be deleted: */
1374 if (!h->data[heap_idx].blocks_nonempty)
1377 stripe_idx = h->data[heap_idx].idx;
1378 m = genradix_ptr(&c->stripes[0], stripe_idx);
1380 if (m->algorithm == head->algo &&
1381 m->nr_redundant == head->redundancy &&
1382 m->sectors == head->blocksize &&
1383 m->blocks_nonempty < m->nr_blocks - m->nr_redundant) {
1384 bch2_stripes_heap_del(c, m, stripe_idx);
1389 spin_unlock(&c->ec_stripes_heap_lock);
1393 static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
1394 struct ec_stripe_head *h)
1400 idx = get_existing_stripe(c, h);
1402 bch_err(c, "failed to find an existing stripe");
1406 h->s->have_existing_stripe = true;
1407 ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1409 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1413 if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1415 * this is a problem: we have deleted from the
1416 * stripes heap already
1421 BUG_ON(h->s->existing_stripe.size != h->blocksize);
1422 BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1424 for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1425 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1426 __set_bit(i, h->s->blocks_gotten);
1427 __set_bit(i, h->s->blocks_allocated);
1430 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1433 bkey_copy(&h->s->new_stripe.key.k_i,
1434 &h->s->existing_stripe.key.k_i);
1439 static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
1440 struct ec_stripe_head *h)
1444 ret = bch2_disk_reservation_get(c, &h->s->res,
1446 h->s->nr_parity, 0);
1450 * This means we need to wait for copygc to
1451 * empty out buckets from existing stripes:
1453 bch_err(c, "failed to reserve stripe");
1459 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1462 unsigned redundancy,
1466 struct ec_stripe_head *h;
1468 bool needs_stripe_new;
1470 h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1472 bch_err(c, "no stripe head");
1476 needs_stripe_new = !h->s;
1477 if (needs_stripe_new) {
1478 if (ec_new_stripe_alloc(c, h)) {
1480 bch_err(c, "failed to allocate new stripe");
1484 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1489 * Try reserve a new stripe before reusing an
1490 * existing stripe. This will prevent unnecessary
1491 * read amplification during write oriented workloads.
1494 if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1495 ret = __bch2_ec_stripe_head_reserve(c, h);
1496 if (ret && needs_stripe_new)
1497 ret = __bch2_ec_stripe_head_reuse(c, h);
1501 if (!h->s->allocated) {
1502 ret = new_stripe_alloc_buckets(c, h, cl);
1506 h->s->allocated = true;
1512 bch2_ec_stripe_head_put(c, h);
1513 return ERR_PTR(-ret);
1516 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1518 struct ec_stripe_head *h;
1519 struct open_bucket *ob;
1522 mutex_lock(&c->ec_stripe_head_lock);
1523 list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1525 mutex_lock(&h->lock);
1529 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1530 if (!h->s->blocks[i])
1533 ob = c->open_buckets + h->s->blocks[i];
1534 if (ob->ptr.dev == ca->dev_idx)
1540 ec_stripe_set_pending(c, h);
1542 mutex_unlock(&h->lock);
1544 mutex_unlock(&c->ec_stripe_head_lock);
1547 void bch2_stripes_heap_start(struct bch_fs *c)
1549 struct genradix_iter iter;
1552 genradix_for_each(&c->stripes[0], iter, m)
1554 bch2_stripes_heap_insert(c, m, iter.pos);
1557 static int __bch2_stripe_write_key(struct btree_trans *trans,
1558 struct btree_iter *iter,
1561 struct bkey_i_stripe *new_key)
1563 const struct bch_stripe *v;
1568 bch2_btree_iter_set_pos(iter, POS(0, idx));
1570 k = bch2_btree_iter_peek_slot(iter);
1575 if (k.k->type != KEY_TYPE_stripe)
1578 v = bkey_s_c_to_stripe(k).v;
1579 for (i = 0; i < v->nr_blocks; i++)
1580 if (m->block_sectors[i] != stripe_blockcount_get(v, i))
1584 bkey_reassemble(&new_key->k_i, k);
1586 for (i = 0; i < new_key->v.nr_blocks; i++)
1587 stripe_blockcount_set(&new_key->v, i,
1588 m->block_sectors[i]);
1590 bch2_trans_update(trans, iter, &new_key->k_i, 0);
1594 int bch2_stripes_write(struct bch_fs *c, unsigned flags)
1596 struct btree_trans trans;
1597 struct btree_iter *iter;
1598 struct genradix_iter giter;
1599 struct bkey_i_stripe *new_key;
1603 new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
1606 bch2_trans_init(&trans, c, 0, 0);
1608 iter = bch2_trans_get_iter(&trans, BTREE_ID_stripes, POS_MIN,
1609 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1611 genradix_for_each(&c->stripes[0], giter, m) {
1615 ret = __bch2_trans_do(&trans, NULL, NULL,
1616 BTREE_INSERT_NOFAIL|flags,
1617 __bch2_stripe_write_key(&trans, iter, m,
1618 giter.pos, new_key));
1624 bch2_trans_exit(&trans);
1631 static int bch2_stripes_read_fn(struct bch_fs *c, enum btree_id id,
1632 unsigned level, struct bkey_s_c k)
1636 if (k.k->type == KEY_TYPE_stripe) {
1637 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL) ?:
1638 bch2_mark_key(c, k, 0, 0, NULL, 0,
1639 BTREE_TRIGGER_NOATOMIC);
1647 int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
1649 int ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_stripes,
1650 NULL, bch2_stripes_read_fn);
1652 bch_err(c, "error reading stripes: %i", ret);
1657 int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
1659 struct btree_trans trans;
1660 struct btree_iter *iter;
1665 bch2_trans_init(&trans, c, 0, 0);
1666 iter = bch2_trans_get_iter(&trans, BTREE_ID_stripes, POS(0, U64_MAX), 0);
1668 k = bch2_btree_iter_prev(iter);
1669 if (!IS_ERR_OR_NULL(k.k))
1670 idx = k.k->p.offset + 1;
1672 bch2_trans_iter_put(&trans, iter);
1673 ret = bch2_trans_exit(&trans);
1681 !init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
1685 ret = genradix_prealloc(&c->stripes[gc], idx, GFP_KERNEL);
1687 for (i = 0; i < idx; i++)
1688 if (!genradix_ptr_alloc(&c->stripes[gc], i, GFP_KERNEL))
1694 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1696 ec_stripes_heap *h = &c->ec_stripes_heap;
1700 spin_lock(&c->ec_stripes_heap_lock);
1701 for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1702 m = genradix_ptr(&c->stripes[0], h->data[i].idx);
1704 pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1705 h->data[i].blocks_nonempty,
1706 m->nr_blocks - m->nr_redundant,
1709 spin_unlock(&c->ec_stripes_heap_lock);
1712 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1714 struct ec_stripe_head *h;
1715 struct ec_stripe_new *s;
1717 mutex_lock(&c->ec_stripe_head_lock);
1718 list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1719 pr_buf(out, "target %u algo %u redundancy %u:\n",
1720 h->target, h->algo, h->redundancy);
1723 pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
1724 h->s->nr_data, h->s->nr_parity,
1725 bitmap_weight(h->s->blocks_allocated,
1728 mutex_unlock(&c->ec_stripe_head_lock);
1730 mutex_lock(&c->ec_stripe_new_lock);
1731 list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1732 pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
1733 s->nr_data, s->nr_parity,
1734 atomic_read(&s->pin));
1736 mutex_unlock(&c->ec_stripe_new_lock);
1739 void bch2_fs_ec_exit(struct bch_fs *c)
1741 struct ec_stripe_head *h;
1744 mutex_lock(&c->ec_stripe_head_lock);
1745 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1746 struct ec_stripe_head, list);
1749 mutex_unlock(&c->ec_stripe_head_lock);
1757 BUG_ON(!list_empty(&c->ec_stripe_new_list));
1759 free_heap(&c->ec_stripes_heap);
1760 genradix_free(&c->stripes[0]);
1761 bioset_exit(&c->ec_bioset);
1764 int bch2_fs_ec_init(struct bch_fs *c)
1766 INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1767 INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1769 return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),