1 // SPDX-License-Identifier: GPL-2.0
6 #include "alloc_foreground.h"
7 #include "backpointers.h"
11 #include "btree_update.h"
12 #include "btree_write_buffer.h"
14 #include "disk_groups.h"
24 #include <linux/sort.h>
28 #include <linux/raid/pq.h>
29 #include <linux/raid/xor.h>
31 static void raid5_recov(unsigned disks, unsigned failed_idx,
32 size_t size, void **data)
36 BUG_ON(failed_idx >= disks);
38 swap(data[0], data[failed_idx]);
39 memcpy(data[0], data[1], size);
42 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
43 xor_blocks(nr, size, data[0], data + i);
47 swap(data[0], data[failed_idx]);
50 static void raid_gen(int nd, int np, size_t size, void **v)
53 raid5_recov(nd + np, nd, size, v);
55 raid6_call.gen_syndrome(nd + np, size, v);
59 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
66 raid5_recov(nd + 1, ir[0], size, v);
68 raid6_call.gen_syndrome(nd + np, size, v);
72 /* data+data failure. */
73 raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
74 } else if (ir[0] < nd) {
75 /* data + p/q failure */
77 if (ir[1] == nd) /* data + p failure */
78 raid6_datap_recov(nd + np, size, ir[0], v);
79 else { /* data + q failure */
80 raid5_recov(nd + 1, ir[0], size, v);
81 raid6_call.gen_syndrome(nd + np, size, v);
84 raid_gen(nd, np, size, v);
94 #include <raid/raid.h>
100 struct ec_stripe_buf *buf;
105 /* Stripes btree keys: */
107 int bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k,
108 unsigned flags, struct printbuf *err)
110 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
112 if (bkey_eq(k.k->p, POS_MIN)) {
113 prt_printf(err, "stripe at POS_MIN");
114 return -BCH_ERR_invalid_bkey;
118 prt_printf(err, "nonzero inode field");
119 return -BCH_ERR_invalid_bkey;
122 if (bkey_val_bytes(k.k) < sizeof(*s)) {
123 prt_printf(err, "incorrect value size (%zu < %zu)",
124 bkey_val_bytes(k.k), sizeof(*s));
125 return -BCH_ERR_invalid_bkey;
128 if (bkey_val_u64s(k.k) < stripe_val_u64s(s)) {
129 prt_printf(err, "incorrect value size (%zu < %u)",
130 bkey_val_u64s(k.k), stripe_val_u64s(s));
131 return -BCH_ERR_invalid_bkey;
134 return bch2_bkey_ptrs_invalid(c, k, flags, err);
137 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
140 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
141 unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
143 prt_printf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
145 le16_to_cpu(s->sectors),
149 1U << s->csum_granularity_bits);
151 for (i = 0; i < s->nr_blocks; i++) {
152 const struct bch_extent_ptr *ptr = s->ptrs + i;
153 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
155 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
157 prt_printf(out, " %u:%llu:%u", ptr->dev, b, offset);
159 prt_printf(out, "#%u", stripe_blockcount_get(s, i));
160 if (ptr_stale(ca, ptr))
161 prt_printf(out, " stale");
165 /* returns blocknr in stripe that we matched: */
166 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
167 struct bkey_s_c k, unsigned *block)
169 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
170 const struct bch_extent_ptr *ptr;
171 unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
173 bkey_for_each_ptr(ptrs, ptr)
174 for (i = 0; i < nr_data; i++)
175 if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
176 le16_to_cpu(s->sectors))) {
184 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
187 case KEY_TYPE_extent: {
188 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
189 const union bch_extent_entry *entry;
191 extent_for_each_entry(e, entry)
192 if (extent_entry_type(entry) ==
193 BCH_EXTENT_ENTRY_stripe_ptr &&
194 entry->stripe_ptr.idx == idx)
206 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
210 for (i = 0; i < buf->key.v.nr_blocks; i++) {
211 kvpfree(buf->data[i], buf->size << 9);
216 /* XXX: this is a non-mempoolified memory allocation: */
217 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
218 unsigned offset, unsigned size)
220 struct bch_stripe *v = &buf->key.v;
221 unsigned csum_granularity = 1U << v->csum_granularity_bits;
222 unsigned end = offset + size;
225 BUG_ON(end > le16_to_cpu(v->sectors));
227 offset = round_down(offset, csum_granularity);
228 end = min_t(unsigned, le16_to_cpu(v->sectors),
229 round_up(end, csum_granularity));
231 buf->offset = offset;
232 buf->size = end - offset;
234 memset(buf->valid, 0xFF, sizeof(buf->valid));
236 for (i = 0; i < buf->key.v.nr_blocks; i++) {
237 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
244 ec_stripe_buf_exit(buf);
245 return -BCH_ERR_ENOMEM_stripe_buf;
250 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
251 unsigned block, unsigned offset)
253 struct bch_stripe *v = &buf->key.v;
254 unsigned csum_granularity = 1 << v->csum_granularity_bits;
255 unsigned end = buf->offset + buf->size;
256 unsigned len = min(csum_granularity, end - offset);
258 BUG_ON(offset >= end);
259 BUG_ON(offset < buf->offset);
260 BUG_ON(offset & (csum_granularity - 1));
261 BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
262 (len & (csum_granularity - 1)));
264 return bch2_checksum(NULL, v->csum_type,
266 buf->data[block] + ((offset - buf->offset) << 9),
270 static void ec_generate_checksums(struct ec_stripe_buf *buf)
272 struct bch_stripe *v = &buf->key.v;
273 unsigned i, j, csums_per_device = stripe_csums_per_device(v);
279 BUG_ON(buf->size != le16_to_cpu(v->sectors));
281 for (i = 0; i < v->nr_blocks; i++)
282 for (j = 0; j < csums_per_device; j++)
283 stripe_csum_set(v, i, j,
284 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
287 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
289 struct bch_stripe *v = &buf->key.v;
290 unsigned csum_granularity = 1 << v->csum_granularity_bits;
296 for (i = 0; i < v->nr_blocks; i++) {
297 unsigned offset = buf->offset;
298 unsigned end = buf->offset + buf->size;
300 if (!test_bit(i, buf->valid))
303 while (offset < end) {
304 unsigned j = offset >> v->csum_granularity_bits;
305 unsigned len = min(csum_granularity, end - offset);
306 struct bch_csum want = stripe_csum_get(v, i, j);
307 struct bch_csum got = ec_block_checksum(buf, i, offset);
309 if (bch2_crc_cmp(want, got)) {
310 struct printbuf buf2 = PRINTBUF;
312 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
314 bch_err_ratelimited(c,
315 "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
316 (void *) _RET_IP_, i, j, v->csum_type,
317 want.lo, got.lo, buf2.buf);
318 printbuf_exit(&buf2);
319 clear_bit(i, buf->valid);
328 /* Erasure coding: */
330 static void ec_generate_ec(struct ec_stripe_buf *buf)
332 struct bch_stripe *v = &buf->key.v;
333 unsigned nr_data = v->nr_blocks - v->nr_redundant;
334 unsigned bytes = le16_to_cpu(v->sectors) << 9;
336 raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
339 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
341 return buf->key.v.nr_blocks -
342 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
345 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
347 struct bch_stripe *v = &buf->key.v;
348 unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
349 unsigned nr_data = v->nr_blocks - v->nr_redundant;
350 unsigned bytes = buf->size << 9;
352 if (ec_nr_failed(buf) > v->nr_redundant) {
353 bch_err_ratelimited(c,
354 "error doing reconstruct read: unable to read enough blocks");
358 for (i = 0; i < nr_data; i++)
359 if (!test_bit(i, buf->valid))
360 failed[nr_failed++] = i;
362 raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
368 static void ec_block_endio(struct bio *bio)
370 struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
371 struct bch_stripe *v = &ec_bio->buf->key.v;
372 struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
373 struct bch_dev *ca = ec_bio->ca;
374 struct closure *cl = bio->bi_private;
376 if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
377 bio_data_dir(bio) ? "write" : "read",
378 bch2_blk_status_to_str(bio->bi_status)))
379 clear_bit(ec_bio->idx, ec_bio->buf->valid);
381 if (ptr_stale(ca, ptr)) {
382 bch_err_ratelimited(ca->fs,
383 "error %s stripe: stale pointer after io",
384 bio_data_dir(bio) == READ ? "reading from" : "writing to");
385 clear_bit(ec_bio->idx, ec_bio->buf->valid);
388 bio_put(&ec_bio->bio);
389 percpu_ref_put(&ca->io_ref);
393 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
394 unsigned rw, unsigned idx, struct closure *cl)
396 struct bch_stripe *v = &buf->key.v;
397 unsigned offset = 0, bytes = buf->size << 9;
398 struct bch_extent_ptr *ptr = &v->ptrs[idx];
399 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
400 enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
404 if (ptr_stale(ca, ptr)) {
405 bch_err_ratelimited(c,
406 "error %s stripe: stale pointer",
407 rw == READ ? "reading from" : "writing to");
408 clear_bit(idx, buf->valid);
412 if (!bch2_dev_get_ioref(ca, rw)) {
413 clear_bit(idx, buf->valid);
417 this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
419 while (offset < bytes) {
420 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
421 DIV_ROUND_UP(bytes, PAGE_SIZE));
422 unsigned b = min_t(size_t, bytes - offset,
423 nr_iovecs << PAGE_SHIFT);
424 struct ec_bio *ec_bio;
426 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
437 ec_bio->bio.bi_iter.bi_sector = ptr->offset + buf->offset + (offset >> 9);
438 ec_bio->bio.bi_end_io = ec_block_endio;
439 ec_bio->bio.bi_private = cl;
441 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
444 percpu_ref_get(&ca->io_ref);
446 submit_bio(&ec_bio->bio);
451 percpu_ref_put(&ca->io_ref);
454 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
455 struct ec_stripe_buf *stripe)
457 struct btree_iter iter;
461 bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
462 POS(0, idx), BTREE_ITER_SLOTS);
463 k = bch2_btree_iter_peek_slot(&iter);
467 if (k.k->type != KEY_TYPE_stripe) {
471 bkey_reassemble(&stripe->key.k_i, k);
473 bch2_trans_iter_exit(trans, &iter);
477 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
479 return bch2_trans_run(c, get_stripe_key_trans(&trans, idx, stripe));
482 /* recovery read path: */
483 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
485 struct ec_stripe_buf *buf;
487 struct bch_stripe *v;
491 closure_init_stack(&cl);
493 BUG_ON(!rbio->pick.has_ec);
495 buf = kzalloc(sizeof(*buf), GFP_NOIO);
497 return -BCH_ERR_ENOMEM_ec_read_extent;
499 ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
501 bch_err_ratelimited(c,
502 "error doing reconstruct read: error %i looking up stripe", ret);
509 if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
510 bch_err_ratelimited(c,
511 "error doing reconstruct read: pointer doesn't match stripe");
516 offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
517 if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
518 bch_err_ratelimited(c,
519 "error doing reconstruct read: read is bigger than stripe");
524 ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
528 for (i = 0; i < v->nr_blocks; i++)
529 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
533 if (ec_nr_failed(buf) > v->nr_redundant) {
534 bch_err_ratelimited(c,
535 "error doing reconstruct read: unable to read enough blocks");
540 ec_validate_checksums(c, buf);
542 ret = ec_do_recov(c, buf);
546 memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
547 buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
549 ec_stripe_buf_exit(buf);
554 /* stripe bucket accounting: */
556 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
558 ec_stripes_heap n, *h = &c->ec_stripes_heap;
560 if (idx >= h->size) {
561 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
562 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
564 mutex_lock(&c->ec_stripes_heap_lock);
565 if (n.size > h->size) {
566 memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
570 mutex_unlock(&c->ec_stripes_heap_lock);
575 if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
576 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
578 if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
579 !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
580 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
585 static int ec_stripe_mem_alloc(struct btree_trans *trans,
586 struct btree_iter *iter)
588 size_t idx = iter->pos.offset;
590 if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
593 bch2_trans_unlock(trans);
595 return __ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL) ?:
596 bch2_trans_relock(trans);
600 * Hash table of open stripes:
601 * Stripes that are being created or modified are kept in a hash table, so that
602 * stripe deletion can skip them.
605 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
607 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
608 struct ec_stripe_new *s;
610 hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
616 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
620 spin_lock(&c->ec_stripes_new_lock);
621 ret = __bch2_stripe_is_open(c, idx);
622 spin_unlock(&c->ec_stripes_new_lock);
627 static bool bch2_try_open_stripe(struct bch_fs *c,
628 struct ec_stripe_new *s,
633 spin_lock(&c->ec_stripes_new_lock);
634 ret = !__bch2_stripe_is_open(c, idx);
636 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
639 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
641 spin_unlock(&c->ec_stripes_new_lock);
646 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
650 spin_lock(&c->ec_stripes_new_lock);
651 hlist_del_init(&s->hash);
652 spin_unlock(&c->ec_stripes_new_lock);
657 /* Heap of all existing stripes, ordered by blocks_nonempty */
659 static u64 stripe_idx_to_delete(struct bch_fs *c)
661 ec_stripes_heap *h = &c->ec_stripes_heap;
663 lockdep_assert_held(&c->ec_stripes_heap_lock);
666 h->data[0].blocks_nonempty == 0 &&
667 !bch2_stripe_is_open(c, h->data[0].idx))
668 return h->data[0].idx;
673 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
674 struct ec_stripe_heap_entry l,
675 struct ec_stripe_heap_entry r)
677 return ((l.blocks_nonempty > r.blocks_nonempty) -
678 (l.blocks_nonempty < r.blocks_nonempty));
681 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
684 struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
686 genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
689 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
691 ec_stripes_heap *h = &c->ec_stripes_heap;
692 struct stripe *m = genradix_ptr(&c->stripes, idx);
694 BUG_ON(m->heap_idx >= h->used);
695 BUG_ON(h->data[m->heap_idx].idx != idx);
698 void bch2_stripes_heap_del(struct bch_fs *c,
699 struct stripe *m, size_t idx)
701 mutex_lock(&c->ec_stripes_heap_lock);
702 heap_verify_backpointer(c, idx);
704 heap_del(&c->ec_stripes_heap, m->heap_idx,
706 ec_stripes_heap_set_backpointer);
707 mutex_unlock(&c->ec_stripes_heap_lock);
710 void bch2_stripes_heap_insert(struct bch_fs *c,
711 struct stripe *m, size_t idx)
713 mutex_lock(&c->ec_stripes_heap_lock);
714 BUG_ON(heap_full(&c->ec_stripes_heap));
716 heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
718 .blocks_nonempty = m->blocks_nonempty,
721 ec_stripes_heap_set_backpointer);
723 heap_verify_backpointer(c, idx);
724 mutex_unlock(&c->ec_stripes_heap_lock);
727 void bch2_stripes_heap_update(struct bch_fs *c,
728 struct stripe *m, size_t idx)
730 ec_stripes_heap *h = &c->ec_stripes_heap;
734 mutex_lock(&c->ec_stripes_heap_lock);
735 heap_verify_backpointer(c, idx);
737 h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
740 heap_sift_up(h, i, ec_stripes_heap_cmp,
741 ec_stripes_heap_set_backpointer);
742 heap_sift_down(h, i, ec_stripes_heap_cmp,
743 ec_stripes_heap_set_backpointer);
745 heap_verify_backpointer(c, idx);
747 do_deletes = stripe_idx_to_delete(c) != 0;
748 mutex_unlock(&c->ec_stripes_heap_lock);
751 bch2_do_stripe_deletes(c);
754 /* stripe deletion */
756 static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
758 struct bch_fs *c = trans->c;
759 struct btree_iter iter;
761 struct bkey_s_c_stripe s;
764 bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, idx),
766 k = bch2_btree_iter_peek_slot(&iter);
771 if (k.k->type != KEY_TYPE_stripe) {
772 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
777 s = bkey_s_c_to_stripe(k);
778 for (unsigned i = 0; i < s.v->nr_blocks; i++)
779 if (stripe_blockcount_get(s.v, i)) {
780 struct printbuf buf = PRINTBUF;
782 bch2_bkey_val_to_text(&buf, c, k);
783 bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
789 ret = bch2_btree_delete_at(trans, &iter, 0);
791 bch2_trans_iter_exit(trans, &iter);
795 static void ec_stripe_delete_work(struct work_struct *work)
798 container_of(work, struct bch_fs, ec_stripe_delete_work);
799 struct btree_trans trans;
803 bch2_trans_init(&trans, c, 0, 0);
806 mutex_lock(&c->ec_stripes_heap_lock);
807 idx = stripe_idx_to_delete(c);
808 mutex_unlock(&c->ec_stripes_heap_lock);
813 ret = commit_do(&trans, NULL, NULL, BTREE_INSERT_NOFAIL,
814 ec_stripe_delete(&trans, idx));
816 bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
821 bch2_trans_exit(&trans);
823 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
826 void bch2_do_stripe_deletes(struct bch_fs *c)
828 if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
829 !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work))
830 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
833 /* stripe creation: */
835 static int ec_stripe_key_update(struct btree_trans *trans,
836 struct bkey_i_stripe *new,
839 struct bch_fs *c = trans->c;
840 struct btree_iter iter;
844 bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
845 new->k.p, BTREE_ITER_INTENT);
846 k = bch2_btree_iter_peek_slot(&iter);
851 if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
852 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
853 create ? "creating" : "updating",
854 bch2_bkey_types[k.k->type]);
859 if (k.k->type == KEY_TYPE_stripe) {
860 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
863 if (old->nr_blocks != new->v.nr_blocks) {
864 bch_err(c, "error updating stripe: nr_blocks does not match");
869 for (i = 0; i < new->v.nr_blocks; i++) {
870 unsigned v = stripe_blockcount_get(old, i);
873 (old->ptrs[i].dev != new->v.ptrs[i].dev ||
874 old->ptrs[i].gen != new->v.ptrs[i].gen ||
875 old->ptrs[i].offset != new->v.ptrs[i].offset));
877 stripe_blockcount_set(&new->v, i, v);
881 ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
883 bch2_trans_iter_exit(trans, &iter);
887 static int ec_stripe_update_extent(struct btree_trans *trans,
888 struct bpos bucket, u8 gen,
889 struct ec_stripe_buf *s,
892 struct bch_fs *c = trans->c;
893 struct bch_backpointer bp;
894 struct btree_iter iter;
896 const struct bch_extent_ptr *ptr_c;
897 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
898 struct bch_extent_stripe_ptr stripe_ptr;
902 ret = bch2_get_next_backpointer(trans, bucket, gen,
903 bp_pos, &bp, BTREE_ITER_CACHED);
906 if (bpos_eq(*bp_pos, SPOS_MAX))
910 struct printbuf buf = PRINTBUF;
911 struct btree_iter node_iter;
914 b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
915 bch2_trans_iter_exit(trans, &node_iter);
920 prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
921 bch2_backpointer_to_text(&buf, &bp);
923 bch2_fs_inconsistent(c, "%s", buf.buf);
928 k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT);
934 * extent no longer exists - we could flush the btree
935 * write buffer and retry to verify, but no need:
940 if (extent_has_stripe_ptr(k, s->key.k.p.offset))
943 ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
945 * It doesn't generally make sense to erasure code cached ptrs:
946 * XXX: should we be incrementing a counter?
948 if (!ptr_c || ptr_c->cached)
951 dev = s->key.v.ptrs[block].dev;
953 n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
954 ret = PTR_ERR_OR_ZERO(n);
958 bkey_reassemble(n, k);
960 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
961 ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev);
964 stripe_ptr = (struct bch_extent_stripe_ptr) {
965 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
967 .redundancy = s->key.v.nr_redundant,
968 .idx = s->key.k.p.offset,
971 __extent_entry_insert(n,
972 (union bch_extent_entry *) ec_ptr,
973 (union bch_extent_entry *) &stripe_ptr);
975 ret = bch2_trans_update(trans, &iter, n, 0);
977 bch2_trans_iter_exit(trans, &iter);
981 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
984 struct bch_fs *c = trans->c;
985 struct bch_extent_ptr bucket = s->key.v.ptrs[block];
986 struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
987 struct bpos bp_pos = POS_MIN;
991 ret = commit_do(trans, NULL, NULL,
992 BTREE_INSERT_NOCHECK_RW|
994 ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
998 if (bkey_eq(bp_pos, POS_MAX))
1001 bp_pos = bpos_nosnap_successor(bp_pos);
1007 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
1009 struct btree_trans trans;
1010 struct bch_stripe *v = &s->key.v;
1011 unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1014 bch2_trans_init(&trans, c, 0, 0);
1016 ret = bch2_btree_write_buffer_flush(&trans);
1020 for (i = 0; i < nr_data; i++) {
1021 ret = ec_stripe_update_bucket(&trans, s, i);
1026 bch2_trans_exit(&trans);
1031 static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1032 struct ec_stripe_new *s,
1034 struct open_bucket *ob)
1036 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1037 unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1040 if (!bch2_dev_get_ioref(ca, WRITE)) {
1045 memset(s->new_stripe.data[block] + (offset << 9),
1047 ob->sectors_free << 9);
1049 ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
1050 ob->bucket * ca->mi.bucket_size + offset,
1054 percpu_ref_put(&ca->io_ref);
1060 void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s)
1063 bch2_stripe_close(c, s);
1068 * data buckets of new stripe all written: create the stripe
1070 static void ec_stripe_create(struct ec_stripe_new *s)
1072 struct bch_fs *c = s->c;
1073 struct open_bucket *ob;
1074 struct bch_stripe *v = &s->new_stripe.key.v;
1075 unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1078 BUG_ON(s->h->s == s);
1080 closure_sync(&s->iodone);
1083 for (i = 0; i < nr_data; i++)
1085 ob = c->open_buckets + s->blocks[i];
1087 if (ob->sectors_free)
1088 zero_out_rest_of_ec_bucket(c, s, i, ob);
1093 if (!bch2_err_matches(s->err, EROFS))
1094 bch_err(c, "error creating stripe: error writing data buckets");
1098 if (s->have_existing_stripe) {
1099 ec_validate_checksums(c, &s->existing_stripe);
1101 if (ec_do_recov(c, &s->existing_stripe)) {
1102 bch_err(c, "error creating stripe: error reading existing stripe");
1106 for (i = 0; i < nr_data; i++)
1107 if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
1108 swap(s->new_stripe.data[i],
1109 s->existing_stripe.data[i]);
1111 ec_stripe_buf_exit(&s->existing_stripe);
1114 BUG_ON(!s->allocated);
1117 ec_generate_ec(&s->new_stripe);
1119 ec_generate_checksums(&s->new_stripe);
1122 for (i = nr_data; i < v->nr_blocks; i++)
1123 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
1124 closure_sync(&s->iodone);
1126 if (ec_nr_failed(&s->new_stripe)) {
1127 bch_err(c, "error creating stripe: error writing redundancy buckets");
1131 ret = bch2_trans_do(c, &s->res, NULL,
1132 BTREE_INSERT_NOCHECK_RW|
1133 BTREE_INSERT_NOFAIL,
1134 ec_stripe_key_update(&trans, &s->new_stripe.key,
1135 !s->have_existing_stripe));
1137 bch_err(c, "error creating stripe: error creating stripe key");
1141 ret = ec_stripe_update_extents(c, &s->new_stripe);
1143 bch_err(c, "error creating stripe: error updating pointers: %s",
1148 bch2_disk_reservation_put(c, &s->res);
1150 for (i = 0; i < v->nr_blocks; i++)
1152 ob = c->open_buckets + s->blocks[i];
1156 __bch2_open_bucket_put(c, ob);
1158 bch2_open_bucket_put(c, ob);
1162 mutex_lock(&c->ec_stripe_new_lock);
1164 mutex_unlock(&c->ec_stripe_new_lock);
1166 ec_stripe_buf_exit(&s->existing_stripe);
1167 ec_stripe_buf_exit(&s->new_stripe);
1168 closure_debug_destroy(&s->iodone);
1170 ec_stripe_new_put(c, s, STRIPE_REF_stripe);
1173 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1175 struct ec_stripe_new *s;
1177 mutex_lock(&c->ec_stripe_new_lock);
1178 list_for_each_entry(s, &c->ec_stripe_new_list, list)
1179 if (!atomic_read(&s->ref[STRIPE_REF_io]))
1183 mutex_unlock(&c->ec_stripe_new_lock);
1188 static void ec_stripe_create_work(struct work_struct *work)
1190 struct bch_fs *c = container_of(work,
1191 struct bch_fs, ec_stripe_create_work);
1192 struct ec_stripe_new *s;
1194 while ((s = get_pending_stripe(c)))
1195 ec_stripe_create(s);
1197 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1200 void bch2_ec_do_stripe_creates(struct bch_fs *c)
1202 bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
1204 if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
1205 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1208 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1210 struct ec_stripe_new *s = h->s;
1212 BUG_ON(!s->allocated && !s->err);
1217 mutex_lock(&c->ec_stripe_new_lock);
1218 list_add(&s->list, &c->ec_stripe_new_list);
1219 mutex_unlock(&c->ec_stripe_new_lock);
1221 ec_stripe_new_put(c, s, STRIPE_REF_io);
1224 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1226 struct ec_stripe_new *s = ob->ec;
1231 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1233 struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1240 BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
1242 ca = bch_dev_bkey_exists(c, ob->dev);
1243 offset = ca->mi.bucket_size - ob->sectors_free;
1245 return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1248 static int unsigned_cmp(const void *_l, const void *_r)
1250 unsigned l = *((const unsigned *) _l);
1251 unsigned r = *((const unsigned *) _r);
1253 return cmp_int(l, r);
1256 /* pick most common bucket size: */
1257 static unsigned pick_blocksize(struct bch_fs *c,
1258 struct bch_devs_mask *devs)
1261 unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1264 } cur = { 0, 0 }, best = { 0, 0 };
1266 for_each_member_device_rcu(ca, c, i, devs)
1267 sizes[nr++] = ca->mi.bucket_size;
1269 sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1271 for (i = 0; i < nr; i++) {
1272 if (sizes[i] != cur.size) {
1273 if (cur.nr > best.nr)
1277 cur.size = sizes[i];
1283 if (cur.nr > best.nr)
1289 static bool may_create_new_stripe(struct bch_fs *c)
1294 static void ec_stripe_key_init(struct bch_fs *c,
1295 struct bkey_i_stripe *s,
1298 unsigned stripe_size)
1302 bkey_stripe_init(&s->k_i);
1303 s->v.sectors = cpu_to_le16(stripe_size);
1305 s->v.nr_blocks = nr_data + nr_parity;
1306 s->v.nr_redundant = nr_parity;
1307 s->v.csum_granularity_bits = ilog2(c->opts.encoded_extent_max >> 9);
1308 s->v.csum_type = BCH_CSUM_crc32c;
1311 while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1312 BUG_ON(1 << s->v.csum_granularity_bits >=
1313 le16_to_cpu(s->v.sectors) ||
1314 s->v.csum_granularity_bits == U8_MAX);
1315 s->v.csum_granularity_bits++;
1318 set_bkey_val_u64s(&s->k, u64s);
1321 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1323 struct ec_stripe_new *s;
1325 lockdep_assert_held(&h->lock);
1327 s = kzalloc(sizeof(*s), GFP_KERNEL);
1329 return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
1331 mutex_init(&s->lock);
1332 closure_init(&s->iodone, NULL);
1333 atomic_set(&s->ref[STRIPE_REF_stripe], 1);
1334 atomic_set(&s->ref[STRIPE_REF_io], 1);
1337 s->nr_data = min_t(unsigned, h->nr_active_devs,
1338 BCH_BKEY_PTRS_MAX) - h->redundancy;
1339 s->nr_parity = h->redundancy;
1341 ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1342 s->nr_parity, h->blocksize);
1348 static struct ec_stripe_head *
1349 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1350 unsigned algo, unsigned redundancy,
1351 enum alloc_reserve reserve)
1353 struct ec_stripe_head *h;
1357 h = kzalloc(sizeof(*h), GFP_KERNEL);
1361 mutex_init(&h->lock);
1362 BUG_ON(!mutex_trylock(&h->lock));
1366 h->redundancy = redundancy;
1367 h->reserve = reserve;
1370 h->devs = target_rw_devs(c, BCH_DATA_user, target);
1372 for_each_member_device_rcu(ca, c, i, &h->devs)
1373 if (!ca->mi.durability)
1374 __clear_bit(i, h->devs.d);
1376 h->blocksize = pick_blocksize(c, &h->devs);
1378 for_each_member_device_rcu(ca, c, i, &h->devs)
1379 if (ca->mi.bucket_size == h->blocksize)
1380 h->nr_active_devs++;
1383 list_add(&h->list, &c->ec_stripe_head_list);
1387 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1391 bitmap_weight(h->s->blocks_allocated,
1392 h->s->nr_data) == h->s->nr_data)
1393 ec_stripe_set_pending(c, h);
1395 mutex_unlock(&h->lock);
1398 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
1401 unsigned redundancy,
1402 enum alloc_reserve reserve)
1404 struct bch_fs *c = trans->c;
1405 struct ec_stripe_head *h;
1411 ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
1413 return ERR_PTR(ret);
1415 if (test_bit(BCH_FS_GOING_RO, &c->flags)) {
1416 h = ERR_PTR(-EROFS);
1420 list_for_each_entry(h, &c->ec_stripe_head_list, list)
1421 if (h->target == target &&
1423 h->redundancy == redundancy &&
1424 h->reserve == reserve) {
1425 ret = bch2_trans_mutex_lock(trans, &h->lock);
1431 h = ec_new_stripe_head_alloc(c, target, algo, redundancy, reserve);
1433 mutex_unlock(&c->ec_stripe_head_lock);
1437 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1438 enum alloc_reserve reserve, struct closure *cl)
1440 struct bch_fs *c = trans->c;
1441 struct bch_devs_mask devs = h->devs;
1442 struct open_bucket *ob;
1443 struct open_buckets buckets;
1444 unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1445 bool have_cache = true;
1448 BUG_ON(h->s->new_stripe.key.v.nr_blocks != h->s->nr_data + h->s->nr_parity);
1449 BUG_ON(h->s->new_stripe.key.v.nr_redundant != h->s->nr_parity);
1451 for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
1452 __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1453 if (i < h->s->nr_data)
1459 BUG_ON(nr_have_data > h->s->nr_data);
1460 BUG_ON(nr_have_parity > h->s->nr_parity);
1463 if (nr_have_parity < h->s->nr_parity) {
1464 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1474 open_bucket_for_each(c, &buckets, ob, i) {
1475 j = find_next_zero_bit(h->s->blocks_gotten,
1476 h->s->nr_data + h->s->nr_parity,
1478 BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1480 h->s->blocks[j] = buckets.v[i];
1481 h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1482 __set_bit(j, h->s->blocks_gotten);
1490 if (nr_have_data < h->s->nr_data) {
1491 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1501 open_bucket_for_each(c, &buckets, ob, i) {
1502 j = find_next_zero_bit(h->s->blocks_gotten,
1504 BUG_ON(j >= h->s->nr_data);
1506 h->s->blocks[j] = buckets.v[i];
1507 h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1508 __set_bit(j, h->s->blocks_gotten);
1518 /* XXX: doesn't obey target: */
1519 static s64 get_existing_stripe(struct bch_fs *c,
1520 struct ec_stripe_head *head)
1522 ec_stripes_heap *h = &c->ec_stripes_heap;
1528 if (may_create_new_stripe(c))
1531 mutex_lock(&c->ec_stripes_heap_lock);
1532 for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1533 /* No blocks worth reusing, stripe will just be deleted: */
1534 if (!h->data[heap_idx].blocks_nonempty)
1537 stripe_idx = h->data[heap_idx].idx;
1539 m = genradix_ptr(&c->stripes, stripe_idx);
1541 if (m->algorithm == head->algo &&
1542 m->nr_redundant == head->redundancy &&
1543 m->sectors == head->blocksize &&
1544 m->blocks_nonempty < m->nr_blocks - m->nr_redundant &&
1545 bch2_try_open_stripe(c, head->s, stripe_idx)) {
1550 mutex_unlock(&c->ec_stripes_heap_lock);
1554 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1556 struct bch_fs *c = trans->c;
1562 * If we can't allocate a new stripe, and there's no stripes with empty
1563 * blocks for us to reuse, that means we have to wait on copygc:
1565 idx = get_existing_stripe(c, h);
1567 return -BCH_ERR_stripe_alloc_blocked;
1569 ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
1571 bch2_stripe_close(c, h->s);
1572 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
1573 bch2_fs_fatal_error(c, "error reading stripe key: %s", bch2_err_str(ret));
1577 BUG_ON(h->s->existing_stripe.key.v.nr_redundant != h->s->nr_parity);
1578 h->s->nr_data = h->s->existing_stripe.key.v.nr_blocks -
1579 h->s->existing_stripe.key.v.nr_redundant;
1581 ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
1583 bch2_stripe_close(c, h->s);
1587 BUG_ON(h->s->existing_stripe.size != h->blocksize);
1588 BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1591 * Free buckets we initially allocated - they might conflict with
1592 * blocks from the stripe we're reusing:
1594 for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
1595 bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
1596 h->s->blocks[i] = 0;
1598 memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
1599 memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
1601 for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1602 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1603 __set_bit(i, h->s->blocks_gotten);
1604 __set_bit(i, h->s->blocks_allocated);
1607 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1610 bkey_copy(&h->s->new_stripe.key.k_i, &h->s->existing_stripe.key.k_i);
1611 h->s->have_existing_stripe = true;
1616 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1618 struct bch_fs *c = trans->c;
1619 struct btree_iter iter;
1621 struct bpos min_pos = POS(0, 1);
1622 struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
1625 if (!h->s->res.sectors) {
1626 ret = bch2_disk_reservation_get(c, &h->s->res,
1629 BCH_DISK_RESERVATION_NOFAIL);
1634 for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
1635 BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
1636 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
1637 if (start_pos.offset) {
1638 start_pos = min_pos;
1639 bch2_btree_iter_set_pos(&iter, start_pos);
1643 ret = -BCH_ERR_ENOSPC_stripe_create;
1647 if (bkey_deleted(k.k) &&
1648 bch2_try_open_stripe(c, h->s, k.k->p.offset))
1652 c->ec_stripe_hint = iter.pos.offset;
1657 ret = ec_stripe_mem_alloc(trans, &iter);
1659 bch2_stripe_close(c, h->s);
1663 h->s->new_stripe.key.k.p = iter.pos;
1665 bch2_trans_iter_exit(trans, &iter);
1668 bch2_disk_reservation_put(c, &h->s->res);
1672 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
1675 unsigned redundancy,
1676 enum alloc_reserve reserve,
1679 struct bch_fs *c = trans->c;
1680 struct ec_stripe_head *h;
1681 bool waiting = false;
1684 h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, reserve);
1686 bch_err(c, "no stripe head");
1687 if (IS_ERR_OR_NULL(h))
1691 ret = ec_new_stripe_alloc(c, h);
1693 bch_err(c, "failed to allocate new stripe");
1698 if (h->s->allocated)
1701 if (h->s->have_existing_stripe)
1702 goto alloc_existing;
1704 /* First, try to allocate a full stripe: */
1705 ret = new_stripe_alloc_buckets(trans, h, RESERVE_stripe, NULL) ?:
1706 __bch2_ec_stripe_head_reserve(trans, h);
1709 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1710 bch2_err_matches(ret, ENOMEM))
1714 * Not enough buckets available for a full stripe: we must reuse an
1718 ret = __bch2_ec_stripe_head_reuse(trans, h);
1721 if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
1724 if (reserve == RESERVE_movinggc) {
1725 ret = new_stripe_alloc_buckets(trans, h, reserve, NULL) ?:
1726 __bch2_ec_stripe_head_reserve(trans, h);
1732 /* XXX freelist_wait? */
1733 closure_wait(&c->freelist_wait, cl);
1738 closure_wake_up(&c->freelist_wait);
1741 * Retry allocating buckets, with the reserve watermark for this
1744 ret = new_stripe_alloc_buckets(trans, h, reserve, cl);
1749 ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
1753 h->s->allocated = true;
1756 BUG_ON(!h->s->new_stripe.data[0]);
1757 BUG_ON(trans->restarted);
1760 bch2_ec_stripe_head_put(c, h);
1761 return ERR_PTR(ret);
1764 static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
1766 struct ec_stripe_head *h;
1767 struct open_bucket *ob;
1770 mutex_lock(&c->ec_stripe_head_lock);
1771 list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1772 mutex_lock(&h->lock);
1779 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1780 if (!h->s->blocks[i])
1783 ob = c->open_buckets + h->s->blocks[i];
1784 if (ob->dev == ca->dev_idx)
1790 ec_stripe_set_pending(c, h);
1792 mutex_unlock(&h->lock);
1794 mutex_unlock(&c->ec_stripe_head_lock);
1797 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1799 __bch2_ec_stop(c, ca);
1802 void bch2_fs_ec_stop(struct bch_fs *c)
1804 __bch2_ec_stop(c, NULL);
1807 static bool bch2_fs_ec_flush_done(struct bch_fs *c)
1811 mutex_lock(&c->ec_stripe_new_lock);
1812 ret = list_empty(&c->ec_stripe_new_list);
1813 mutex_unlock(&c->ec_stripe_new_lock);
1818 void bch2_fs_ec_flush(struct bch_fs *c)
1820 wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c));
1823 int bch2_stripes_read(struct bch_fs *c)
1825 struct btree_trans trans;
1826 struct btree_iter iter;
1828 const struct bch_stripe *s;
1833 bch2_trans_init(&trans, c, 0, 0);
1835 for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1836 BTREE_ITER_PREFETCH, k, ret) {
1837 if (k.k->type != KEY_TYPE_stripe)
1840 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1844 s = bkey_s_c_to_stripe(k).v;
1846 m = genradix_ptr(&c->stripes, k.k->p.offset);
1847 m->sectors = le16_to_cpu(s->sectors);
1848 m->algorithm = s->algorithm;
1849 m->nr_blocks = s->nr_blocks;
1850 m->nr_redundant = s->nr_redundant;
1851 m->blocks_nonempty = 0;
1853 for (i = 0; i < s->nr_blocks; i++)
1854 m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1856 bch2_stripes_heap_insert(c, m, k.k->p.offset);
1858 bch2_trans_iter_exit(&trans, &iter);
1860 bch2_trans_exit(&trans);
1863 bch_err(c, "error reading stripes: %i", ret);
1868 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1870 ec_stripes_heap *h = &c->ec_stripes_heap;
1874 mutex_lock(&c->ec_stripes_heap_lock);
1875 for (i = 0; i < min_t(size_t, h->used, 50); i++) {
1876 m = genradix_ptr(&c->stripes, h->data[i].idx);
1878 prt_printf(out, "%zu %u/%u+%u", h->data[i].idx,
1879 h->data[i].blocks_nonempty,
1880 m->nr_blocks - m->nr_redundant,
1882 if (bch2_stripe_is_open(c, h->data[i].idx))
1883 prt_str(out, " open");
1886 mutex_unlock(&c->ec_stripes_heap_lock);
1889 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1891 struct ec_stripe_head *h;
1892 struct ec_stripe_new *s;
1894 mutex_lock(&c->ec_stripe_head_lock);
1895 list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1896 prt_printf(out, "target %u algo %u redundancy %u %s:\n",
1897 h->target, h->algo, h->redundancy,
1898 bch2_alloc_reserves[h->reserve]);
1901 prt_printf(out, "\tidx %llu blocks %u+%u allocated %u\n",
1902 h->s->idx, h->s->nr_data, h->s->nr_parity,
1903 bitmap_weight(h->s->blocks_allocated,
1906 mutex_unlock(&c->ec_stripe_head_lock);
1908 prt_printf(out, "in flight:\n");
1910 mutex_lock(&c->ec_stripe_new_lock);
1911 list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1912 prt_printf(out, "\tidx %llu blocks %u+%u ref %u %u %s\n",
1913 s->idx, s->nr_data, s->nr_parity,
1914 atomic_read(&s->ref[STRIPE_REF_io]),
1915 atomic_read(&s->ref[STRIPE_REF_stripe]),
1916 bch2_alloc_reserves[s->h->reserve]);
1918 mutex_unlock(&c->ec_stripe_new_lock);
1921 void bch2_fs_ec_exit(struct bch_fs *c)
1923 struct ec_stripe_head *h;
1927 mutex_lock(&c->ec_stripe_head_lock);
1928 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1929 struct ec_stripe_head, list);
1932 mutex_unlock(&c->ec_stripe_head_lock);
1937 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++)
1938 BUG_ON(h->s->blocks[i]);
1945 BUG_ON(!list_empty(&c->ec_stripe_new_list));
1947 free_heap(&c->ec_stripes_heap);
1948 genradix_free(&c->stripes);
1949 bioset_exit(&c->ec_bioset);
1952 void bch2_fs_ec_init_early(struct bch_fs *c)
1954 spin_lock_init(&c->ec_stripes_new_lock);
1955 mutex_init(&c->ec_stripes_heap_lock);
1957 INIT_LIST_HEAD(&c->ec_stripe_head_list);
1958 mutex_init(&c->ec_stripe_head_lock);
1960 INIT_LIST_HEAD(&c->ec_stripe_new_list);
1961 mutex_init(&c->ec_stripe_new_lock);
1962 init_waitqueue_head(&c->ec_stripe_new_wait);
1964 INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1965 INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1968 int bch2_fs_ec_init(struct bch_fs *c)
1970 return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),