bcachefs: Erasure coding

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

1 files changed, 1265 insertions, 0 deletions

diff --git a/fs/bcachefs/ec.c b/fs/bcachefs/ec.c
new file mode 100644
index 000000000000..f6314aa6a0f1
--- /dev/null
+++ b/fs/bcachefs/ec.c
@@ -0,0 +1,1265 @@
+// SPDX-License-Identifier: GPL-2.0
+
+/* erasure coding */
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "bset.h"
+#include "btree_gc.h"
+#include "btree_update.h"
+#include "buckets.h"
+#include "disk_groups.h"
+#include "ec.h"
+#include "error.h"
+#include "io.h"
+#include "keylist.h"
+#include "super-io.h"
+#include "util.h"
+
+#include <linux/raid/pq.h>
+#include <linux/raid/xor.h>
+#include <linux/sort.h>
+
+struct ec_bio {
+	struct bch_dev		*ca;
+	struct ec_stripe_buf	*buf;
+	size_t			idx;
+	struct bio		bio;
+};
+
+/* Stripes btree keys: */
+
+static unsigned stripe_csums_per_device(const struct bch_stripe *s)
+{
+	return DIV_ROUND_UP(le16_to_cpu(s->sectors),
+			    1 << s->csum_granularity_bits);
+}
+
+static unsigned stripe_val_u64s(const struct bch_stripe *s)
+{
+	unsigned bytes = sizeof(struct bch_stripe) +
+		sizeof(struct bch_extent_ptr) * s->nr_blocks +
+		bch_crc_bytes[s->csum_type] * s->nr_blocks * stripe_csums_per_device(s);
+	return DIV_ROUND_UP(bytes, sizeof(u64));
+}
+
+static void *stripe_csum(struct bch_stripe *s, unsigned dev, unsigned csum_idx)
+{
+	unsigned csum_bytes = bch_crc_bytes[s->csum_type];
+	void *csums = s->ptrs + s->nr_blocks;
+
+	BUG_ON(!csum_bytes);
+
+	return csums + (dev * stripe_csums_per_device(s) + csum_idx) * csum_bytes;
+}
+
+const char *bch2_ec_key_invalid(const struct bch_fs *c, struct bkey_s_c k)
+{
+	if (k.k->p.inode)
+		return "invalid stripe key";
+
+	switch (k.k->type) {
+	case BCH_STRIPE: {
+		const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
+
+		if (bkey_val_bytes(k.k) < sizeof(*s))
+			return "incorrect value size";
+
+		if (bkey_val_u64s(k.k) != stripe_val_u64s(s))
+			return "incorrect value size";
+
+		return NULL;
+	}
+	default:
+		return "invalid type";
+	}
+}
+
+void bch2_ec_key_to_text(struct printbuf *out, struct bch_fs *c,
+			 struct bkey_s_c k)
+{
+	switch (k.k->type) {
+	case BCH_STRIPE: {
+		const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
+		unsigned i;
+
+		pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
+		       s->algorithm,
+		       le16_to_cpu(s->sectors),
+		       s->nr_blocks - s->nr_redundant,
+		       s->nr_redundant,
+		       s->csum_type,
+		       1U << s->csum_granularity_bits);
+
+		for (i = 0; i < s->nr_blocks; i++)
+			pr_buf(out, " %u:%llu", s->ptrs[i].dev,
+			       (u64) s->ptrs[i].offset);
+	}
+	}
+}
+
+static int ptr_matches_stripe(struct bch_fs *c,
+			      struct bch_stripe *v,
+			      const struct bch_extent_ptr *ptr)
+{
+	unsigned i;
+
+	for (i = 0; i < v->nr_blocks - v->nr_redundant; i++) {
+		const struct bch_extent_ptr *ptr2 = v->ptrs + i;
+
+		if (ptr->dev == ptr2->dev &&
+		    ptr->gen == ptr2->gen &&
+		    ptr->offset >= ptr2->offset &&
+		    ptr->offset <  ptr2->offset + le16_to_cpu(v->sectors))
+			return i;
+	}
+
+	return -1;
+}
+
+static int extent_matches_stripe(struct bch_fs *c,
+				 struct bch_stripe *v,
+				 struct bkey_s_c k)
+{
+	struct bkey_s_c_extent e;
+	const struct bch_extent_ptr *ptr;
+	int idx;
+
+	if (!bkey_extent_is_data(k.k))
+		return -1;
+
+	e = bkey_s_c_to_extent(k);
+
+	extent_for_each_ptr(e, ptr) {
+		idx = ptr_matches_stripe(c, v, ptr);
+		if (idx >= 0)
+			return idx;
+	}
+
+	return -1;
+}
+
+static void ec_stripe_key_init(struct bch_fs *c,
+			       struct bkey_i_stripe *s,
+			       struct open_buckets *blocks,
+			       struct open_buckets *parity,
+			       unsigned stripe_size)
+{
+	struct open_bucket *ob;
+	unsigned i, u64s;
+
+	bkey_stripe_init(&s->k_i);
+	s->v.sectors			= cpu_to_le16(stripe_size);
+	s->v.algorithm			= 0;
+	s->v.nr_blocks			= parity->nr + blocks->nr;
+	s->v.nr_redundant		= parity->nr;
+	s->v.csum_granularity_bits	= ilog2(c->sb.encoded_extent_max);
+	s->v.csum_type			= BCH_CSUM_CRC32C;
+	s->v.pad			= 0;
+
+	open_bucket_for_each(c, blocks, ob, i)
+		s->v.ptrs[i]			= ob->ptr;
+
+	open_bucket_for_each(c, parity, ob, i)
+		s->v.ptrs[blocks->nr + i]	= ob->ptr;
+
+	while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
+		BUG_ON(1 << s->v.csum_granularity_bits >=
+		       le16_to_cpu(s->v.sectors) ||
+		       s->v.csum_granularity_bits == U8_MAX);
+		s->v.csum_granularity_bits++;
+	}
+
+	set_bkey_val_u64s(&s->k, u64s);
+}
+
+/* Checksumming: */
+
+static void ec_generate_checksums(struct ec_stripe_buf *buf)
+{
+	struct bch_stripe *v = &buf->key.v;
+	unsigned csum_granularity = 1 << v->csum_granularity_bits;
+	unsigned csums_per_device = stripe_csums_per_device(v);
+	unsigned csum_bytes = bch_crc_bytes[v->csum_type];
+	unsigned i, j;
+
+	if (!csum_bytes)
+		return;
+
+	BUG_ON(buf->offset);
+	BUG_ON(buf->size != le16_to_cpu(v->sectors));
+
+	for (i = 0; i < v->nr_blocks; i++) {
+		for (j = 0; j < csums_per_device; j++) {
+			unsigned offset = j << v->csum_granularity_bits;
+			unsigned len = min(csum_granularity, buf->size - offset);
+
+			struct bch_csum csum =
+				bch2_checksum(NULL, v->csum_type,
+					      null_nonce(),
+					      buf->data[i] + (offset << 9),
+					      len << 9);
+
+			memcpy(stripe_csum(v, i, j), &csum, csum_bytes);
+		}
+	}
+}
+
+static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
+{
+	struct bch_stripe *v = &buf->key.v;
+	unsigned csum_granularity = 1 << v->csum_granularity_bits;
+	unsigned csum_bytes = bch_crc_bytes[v->csum_type];
+	unsigned i;
+
+	if (!csum_bytes)
+		return;
+
+	for (i = 0; i < v->nr_blocks; i++) {
+		unsigned offset = buf->offset;
+		unsigned end = buf->offset + buf->size;
+
+		if (!test_bit(i, buf->valid))
+			continue;
+
+		while (offset < end) {
+			unsigned j = offset >> v->csum_granularity_bits;
+			unsigned len = min(csum_granularity, end - offset);
+			struct bch_csum csum;
+
+			BUG_ON(offset & (csum_granularity - 1));
+			BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
+			       ((offset + len) & (csum_granularity - 1)));
+
+			csum = bch2_checksum(NULL, v->csum_type,
+					     null_nonce(),
+					     buf->data[i] + ((offset - buf->offset) << 9),
+					     len << 9);
+
+			if (memcmp(stripe_csum(v, i, j), &csum, csum_bytes)) {
+				__bcache_io_error(c,
+					"checksum error while doing reconstruct read (%u:%u)",
+					i, j);
+				clear_bit(i, buf->valid);
+				break;
+			}
+
+			offset += len;
+		}
+	}
+}
+
+/* Erasure coding: */
+
+static void raid5_recov(unsigned disks, unsigned bytes,
+			unsigned failed, void **data)
+{
+	unsigned i = 2, nr;
+
+	BUG_ON(failed >= disks);
+
+	swap(data[0], data[failed]);
+	memcpy(data[0], data[1], bytes);
+
+	while (i < disks) {
+		nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
+		xor_blocks(nr, bytes, data[0], data + i);
+		i += nr;
+	}
+
+	swap(data[0], data[failed]);
+}
+
+static void ec_generate_ec(struct ec_stripe_buf *buf)
+{
+	struct bch_stripe *v = &buf->key.v;
+	unsigned nr_data = v->nr_blocks - v->nr_redundant;
+	unsigned bytes = le16_to_cpu(v->sectors) << 9;
+
+	switch (v->nr_redundant) {
+	case 2:
+		raid6_call.gen_syndrome(v->nr_blocks, bytes, buf->data);
+		fallthrough;
+	case 1:
+		raid5_recov(v->nr_blocks, bytes, nr_data, buf->data);
+		break;
+	default:
+		BUG();
+	}
+}
+
+static unsigned __ec_nr_failed(struct ec_stripe_buf *buf, unsigned nr)
+{
+	return nr - bitmap_weight(buf->valid, nr);
+}
+
+static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
+{
+	return __ec_nr_failed(buf, buf->key.v.nr_blocks);
+}
+
+static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
+{
+	struct bch_stripe *v = &buf->key.v;
+	unsigned i, failed[EC_STRIPE_MAX], nr_failed = 0;
+	unsigned nr_data = v->nr_blocks - v->nr_redundant;
+	unsigned bytes = buf->size << 9;
+
+	if (ec_nr_failed(buf) > v->nr_redundant) {
+		__bcache_io_error(c,
+			"error doing reconstruct read: unable to read enough blocks");
+		return -1;
+	}
+
+	for (i = 0; i < nr_data; i++)
+		if (!test_bit(i, buf->valid))
+			failed[nr_failed++] = i;
+
+	switch (nr_failed) {
+	case 0:
+		break;
+	case 1:
+		if (test_bit(nr_data, buf->valid))
+			raid5_recov(nr_data + 1, bytes, failed[0], buf->data);
+		else
+			raid6_datap_recov(v->nr_blocks, bytes, failed[0], buf->data);
+		break;
+	case 2:
+		/* data+data failure. */
+		raid6_2data_recov(v->nr_blocks, bytes, failed[0], failed[1], buf->data);
+		break;
+
+	default:
+		BUG();
+	}
+
+	for (i = nr_data; i < v->nr_blocks; i++)
+		if (!test_bit(i, buf->valid)) {
+			ec_generate_ec(buf);
+			break;
+		}
+
+	return 0;
+}
+
+/* IO: */
+
+static void ec_block_endio(struct bio *bio)
+{
+	struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
+	struct bch_dev *ca = ec_bio->ca;
+	struct closure *cl = bio->bi_private;
+
+	if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding"))
+		clear_bit(ec_bio->idx, ec_bio->buf->valid);
+
+	bio_put(&ec_bio->bio);
+	percpu_ref_put(&ca->io_ref);
+	closure_put(cl);
+}
+
+static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
+			unsigned rw, unsigned idx, struct closure *cl)
+{
+	struct bch_stripe *v = &buf->key.v;
+	unsigned offset = 0, bytes = buf->size << 9;
+	struct bch_extent_ptr *ptr = &v->ptrs[idx];
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+
+	if (!bch2_dev_get_ioref(ca, rw)) {
+		clear_bit(idx, buf->valid);
+		return;
+	}
+
+	while (offset < bytes) {
+		unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
+					   DIV_ROUND_UP(bytes, PAGE_SIZE));
+		unsigned b = min_t(size_t, bytes - offset,
+				   nr_iovecs << PAGE_SHIFT);
+		struct ec_bio *ec_bio;
+
+		ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
+						       nr_iovecs,
+						       rw,
+						       GFP_KERNEL,
+						       &c->ec_bioset),
+				      struct ec_bio, bio);
+
+		ec_bio->ca			= ca;
+		ec_bio->buf			= buf;
+		ec_bio->idx			= idx;
+
+		ec_bio->bio.bi_iter.bi_sector	= ptr->offset + buf->offset + (offset >> 9);
+		ec_bio->bio.bi_iter.bi_size	= b;
+		ec_bio->bio.bi_end_io		= ec_block_endio;
+		ec_bio->bio.bi_private		= cl;
+
+		bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset);
+
+		closure_get(cl);
+		percpu_ref_get(&ca->io_ref);
+
+		submit_bio(&ec_bio->bio);
+
+		offset += b;
+	}
+
+	percpu_ref_put(&ca->io_ref);
+}
+
+/* recovery read path: */
+int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
+{
+	struct btree_iter iter;
+	struct ec_stripe_buf *buf;
+	struct closure cl;
+	struct bkey_s_c k;
+	struct bch_stripe *v;
+	unsigned stripe_idx;
+	unsigned offset, end;
+	unsigned i, nr_data, csum_granularity;
+	int ret = 0, idx;
+
+	closure_init_stack(&cl);
+
+	BUG_ON(!rbio->pick.idx ||
+	       rbio->pick.idx - 1 >= rbio->pick.ec_nr);
+
+	stripe_idx = rbio->pick.ec[rbio->pick.idx - 1].idx;
+
+	buf = kzalloc(sizeof(*buf), GFP_NOIO);
+	if (!buf)
+		return -ENOMEM;
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_EC,
+			     POS(0, stripe_idx),
+			     BTREE_ITER_SLOTS);
+	k = bch2_btree_iter_peek_slot(&iter);
+	if (btree_iter_err(k) || k.k->type != BCH_STRIPE) {
+		__bcache_io_error(c,
+			"error doing reconstruct read: stripe not found");
+		kfree(buf);
+		return bch2_btree_iter_unlock(&iter) ?: -EIO;
+	}
+
+	bkey_reassemble(&buf->key.k_i, k);
+	bch2_btree_iter_unlock(&iter);
+
+	v = &buf->key.v;
+
+	nr_data = v->nr_blocks - v->nr_redundant;
+
+	idx = ptr_matches_stripe(c, v, &rbio->pick.ptr);
+	BUG_ON(idx < 0);
+
+	csum_granularity = 1U << v->csum_granularity_bits;
+
+	offset	= rbio->bio.bi_iter.bi_sector - v->ptrs[idx].offset;
+	end	= offset + bio_sectors(&rbio->bio);
+
+	BUG_ON(end > le16_to_cpu(v->sectors));
+
+	buf->offset	= round_down(offset, csum_granularity);
+	buf->size	= min_t(unsigned, le16_to_cpu(v->sectors),
+				round_up(end, csum_granularity)) - buf->offset;
+
+	for (i = 0; i < v->nr_blocks; i++) {
+		buf->data[i] = kmalloc(buf->size << 9, GFP_NOIO);
+		if (!buf->data[i]) {
+			ret = -ENOMEM;
+			goto err;
+		}
+	}
+
+	memset(buf->valid, 0xFF, sizeof(buf->valid));
+
+	for (i = 0; i < v->nr_blocks; i++) {
+		struct bch_extent_ptr *ptr = v->ptrs + i;
+		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+
+		if (ptr_stale(ca, ptr)) {
+			__bcache_io_error(c,
+					  "error doing reconstruct read: stale pointer");
+			clear_bit(i, buf->valid);
+			continue;
+		}
+
+		ec_block_io(c, buf, REQ_OP_READ, i, &cl);
+	}
+
+	closure_sync(&cl);
+
+	if (ec_nr_failed(buf) > v->nr_redundant) {
+		__bcache_io_error(c,
+			"error doing reconstruct read: unable to read enough blocks");
+		ret = -EIO;
+		goto err;
+	}
+
+	ec_validate_checksums(c, buf);
+
+	ret = ec_do_recov(c, buf);
+	if (ret)
+		goto err;
+
+	memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
+		      buf->data[idx] + ((offset - buf->offset) << 9));
+err:
+	for (i = 0; i < v->nr_blocks; i++)
+		kfree(buf->data[i]);
+	kfree(buf);
+	return ret;
+}
+
+/* ec_stripe bucket accounting: */
+
+static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
+{
+	ec_stripes_heap n, *h = &c->ec_stripes_heap;
+
+	if (idx >= h->size) {
+		if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
+			return -ENOMEM;
+
+		spin_lock(&c->ec_stripes_heap_lock);
+		if (n.size > h->size) {
+			memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
+			n.used = h->used;
+			swap(*h, n);
+		}
+		spin_unlock(&c->ec_stripes_heap_lock);
+
+		free_heap(&n);
+	}
+
+	if (!genradix_ptr_alloc(&c->ec_stripes, idx, gfp))
+		return -ENOMEM;
+
+	return 0;
+}
+
+static int ec_stripe_mem_alloc(struct bch_fs *c,
+			       struct btree_iter *iter)
+{
+	size_t idx = iter->pos.offset;
+
+	if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT|__GFP_NOWARN))
+		return 0;
+
+	bch2_btree_iter_unlock(iter);
+
+	if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
+		return -EINTR;
+	return -ENOMEM;
+}
+
+static ssize_t stripe_idx_to_delete(struct bch_fs *c)
+{
+	ec_stripes_heap *h = &c->ec_stripes_heap;
+
+	return h->data[0].blocks_nonempty == 0 ? h->data[0].idx : -1;
+}
+
+static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
+				      struct ec_stripe_heap_entry l,
+				      struct ec_stripe_heap_entry r)
+{
+	return ((l.blocks_nonempty > r.blocks_nonempty) -
+		(l.blocks_nonempty < r.blocks_nonempty));
+}
+
+static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
+						   size_t i)
+{
+	struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
+
+	genradix_ptr(&c->ec_stripes, h->data[i].idx)->heap_idx = i;
+}
+
+static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
+{
+	ec_stripes_heap *h = &c->ec_stripes_heap;
+	struct ec_stripe *m = genradix_ptr(&c->ec_stripes, idx);
+
+	BUG_ON(!m->alive);
+	BUG_ON(m->heap_idx >= h->used);
+	BUG_ON(h->data[m->heap_idx].idx != idx);
+}
+
+static inline unsigned stripe_entry_blocks(struct ec_stripe *m)
+{
+	return atomic_read(&m->pin)
+		? UINT_MAX : atomic_read(&m->blocks_nonempty);
+}
+
+void bch2_stripes_heap_update(struct bch_fs *c,
+			      struct ec_stripe *m, size_t idx)
+{
+	ec_stripes_heap *h = &c->ec_stripes_heap;
+	bool queue_delete;
+	size_t i;
+
+	spin_lock(&c->ec_stripes_heap_lock);
+
+	if (!m->alive) {
+		spin_unlock(&c->ec_stripes_heap_lock);
+		return;
+	}
+
+	heap_verify_backpointer(c, idx);
+
+	h->data[m->heap_idx].blocks_nonempty =
+		stripe_entry_blocks(m);
+
+	i = m->heap_idx;
+	heap_sift_up(h,	  i, ec_stripes_heap_cmp,
+		     ec_stripes_heap_set_backpointer);
+	heap_sift_down(h, i, ec_stripes_heap_cmp,
+		       ec_stripes_heap_set_backpointer);
+
+	heap_verify_backpointer(c, idx);
+
+	queue_delete = stripe_idx_to_delete(c) >= 0;
+	spin_unlock(&c->ec_stripes_heap_lock);
+
+	if (queue_delete)
+		schedule_work(&c->ec_stripe_delete_work);
+}
+
+void bch2_stripes_heap_del(struct bch_fs *c,
+			   struct ec_stripe *m, size_t idx)
+{
+	spin_lock(&c->ec_stripes_heap_lock);
+	heap_verify_backpointer(c, idx);
+
+	m->alive = false;
+	heap_del(&c->ec_stripes_heap, m->heap_idx,
+		 ec_stripes_heap_cmp,
+		 ec_stripes_heap_set_backpointer);
+	spin_unlock(&c->ec_stripes_heap_lock);
+}
+
+void bch2_stripes_heap_insert(struct bch_fs *c,
+			      struct ec_stripe *m, size_t idx)
+{
+	spin_lock(&c->ec_stripes_heap_lock);
+
+	BUG_ON(heap_full(&c->ec_stripes_heap));
+
+	heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
+			.idx = idx,
+			.blocks_nonempty = stripe_entry_blocks(m),
+		}),
+		 ec_stripes_heap_cmp,
+		 ec_stripes_heap_set_backpointer);
+	m->alive = true;
+
+	heap_verify_backpointer(c, idx);
+
+	spin_unlock(&c->ec_stripes_heap_lock);
+}
+
+static void ec_stripe_delete(struct bch_fs *c, unsigned idx)
+{
+	struct btree_iter iter;
+	struct bch_stripe *v = NULL;
+	struct bkey_s_c k;
+	struct bkey_i delete;
+	u64 journal_seq = 0;
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_EC,
+			     POS(0, idx),
+			     BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+	k = bch2_btree_iter_peek_slot(&iter);
+	if (btree_iter_err(k) || k.k->type != BCH_STRIPE)
+		goto out;
+
+	v = kmalloc(bkey_val_bytes(k.k), GFP_KERNEL);
+	BUG_ON(!v);
+	memcpy(v, bkey_s_c_to_stripe(k).v, bkey_val_bytes(k.k));
+
+	bkey_init(&delete.k);
+	delete.k.p = iter.pos;
+
+	bch2_btree_insert_at(c, NULL, &journal_seq,
+			     BTREE_INSERT_NOFAIL|
+			     BTREE_INSERT_USE_RESERVE|
+			     BTREE_INSERT_NOUNLOCK,
+			     BTREE_INSERT_ENTRY(&iter, &delete));
+out:
+	bch2_btree_iter_unlock(&iter);
+	kfree(v);
+}
+
+static void ec_stripe_delete_work(struct work_struct *work)
+{
+	struct bch_fs *c =
+		container_of(work, struct bch_fs, ec_stripe_delete_work);
+	ssize_t idx;
+
+	down_read(&c->gc_lock);
+
+	while (1) {
+		spin_lock(&c->ec_stripes_heap_lock);
+		idx = stripe_idx_to_delete(c);
+		spin_unlock(&c->ec_stripes_heap_lock);
+
+		if (idx < 0)
+			break;
+
+		ec_stripe_delete(c, idx);
+	}
+
+	up_read(&c->gc_lock);
+}
+
+static int ec_stripe_bkey_insert(struct bch_fs *c,
+				 struct bkey_i_stripe *stripe)
+{
+	struct ec_stripe *m;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	/* XXX: start pos hint */
+retry:
+	for_each_btree_key(&iter, c, BTREE_ID_EC, POS_MIN,
+			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k) {
+		if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
+			bch2_btree_iter_unlock(&iter);
+			return -ENOSPC;
+		}
+
+		if (bkey_deleted(k.k))
+			goto found_slot;
+	}
+
+	return bch2_btree_iter_unlock(&iter) ?: -ENOSPC;
+found_slot:
+	mutex_lock(&c->ec_stripes_lock);
+	ret = ec_stripe_mem_alloc(c, &iter);
+	mutex_unlock(&c->ec_stripes_lock);
+
+	if (ret == -EINTR)
+		goto retry;
+	if (ret)
+		return ret;
+
+	m = genradix_ptr(&c->ec_stripes, iter.pos.offset);
+	atomic_inc(&m->pin);
+
+	stripe->k.p = iter.pos;
+
+	ret = bch2_btree_insert_at(c, NULL, NULL,
+				   BTREE_INSERT_NOFAIL|
+				   BTREE_INSERT_USE_RESERVE,
+				   BTREE_INSERT_ENTRY(&iter, &stripe->k_i));
+	bch2_btree_iter_unlock(&iter);
+
+	if (ret)
+		atomic_dec(&m->pin);
+
+	return ret;
+}
+
+/* stripe creation: */
+
+static void extent_stripe_ptr_add(struct bkey_s_extent e,
+				  struct ec_stripe_buf *s,
+				  struct bch_extent_ptr *ptr,
+				  unsigned block)
+{
+	struct bch_extent_stripe_ptr *dst = (void *) ptr;
+	union bch_extent_entry *end = extent_entry_last(e);
+
+	memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
+	e.k->u64s += sizeof(*dst) / sizeof(u64);
+
+	*dst = (struct bch_extent_stripe_ptr) {
+		.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
+		.block		= block,
+		.idx		= s->key.k.p.offset,
+	};
+}
+
+static int ec_stripe_update_ptrs(struct bch_fs *c,
+				 struct ec_stripe_buf *s,
+				 struct bkey *pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	struct bkey_s_extent e;
+	struct bch_extent_ptr *ptr;
+	BKEY_PADDED(k) tmp;
+	int ret = 0, dev, idx;
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS,
+			     bkey_start_pos(pos),
+			     BTREE_ITER_INTENT);
+
+	while ((k = bch2_btree_iter_peek(&iter)).k &&
+	       !btree_iter_err(k) &&
+	       bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
+		idx = extent_matches_stripe(c, &s->key.v, k);
+		if (idx < 0) {
+			bch2_btree_iter_next(&iter);
+			continue;
+		}
+
+		dev = s->key.v.ptrs[idx].dev;
+
+		bkey_reassemble(&tmp.k, k);
+		e = bkey_i_to_s_extent(&tmp.k);
+
+		extent_for_each_ptr(e, ptr)
+			if (ptr->dev != dev)
+				ptr->cached = true;
+
+		ptr = (void *) bch2_extent_has_device(e.c, dev);
+		BUG_ON(!ptr);
+
+		extent_stripe_ptr_add(e, s, ptr, idx);
+
+		ret = bch2_btree_insert_at(c, NULL, NULL,
+				BTREE_INSERT_ATOMIC|
+				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_USE_RESERVE,
+				BTREE_INSERT_ENTRY(&iter, &tmp.k));
+		if (ret == -EINTR)
+			ret = 0;
+		if (ret)
+			break;
+	}
+
+	return bch2_btree_iter_unlock(&iter) ?: ret;
+}
+
+/*
+ * data buckets of new stripe all written: create the stripe
+ */
+static void ec_stripe_create(struct ec_stripe_new *s)
+{
+	struct ec_stripe *ec_stripe;
+	struct bch_fs *c = s->c;
+	struct open_bucket *ob;
+	struct bkey_i *k;
+	struct bch_stripe *v = &s->stripe.key.v;
+	unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
+	struct closure cl;
+	int ret;
+
+	BUG_ON(s->h->s == s);
+
+	closure_init_stack(&cl);
+
+	if (s->err) {
+		bch_err(c, "error creating stripe: error writing data buckets");
+		goto err;
+	}
+
+	if (!percpu_ref_tryget(&c->writes))
+		goto err;
+
+	BUG_ON(bitmap_weight(s->blocks_allocated,
+			     s->blocks.nr) != s->blocks.nr);
+
+	ec_generate_ec(&s->stripe);
+
+	ec_generate_checksums(&s->stripe);
+
+	/* write p/q: */
+	for (i = nr_data; i < v->nr_blocks; i++)
+		ec_block_io(c, &s->stripe, REQ_OP_WRITE, i, &cl);
+
+	closure_sync(&cl);
+
+	for (i = nr_data; i < v->nr_blocks; i++)
+		if (!test_bit(i, s->stripe.valid)) {
+			bch_err(c, "error creating stripe: error writing redundancy buckets");
+			goto err_put_writes;
+		}
+
+	ret = ec_stripe_bkey_insert(c, &s->stripe.key);
+	if (ret) {
+		bch_err(c, "error creating stripe: error creating stripe key");
+		goto err_put_writes;
+	}
+
+	for_each_keylist_key(&s->keys, k) {
+		ret = ec_stripe_update_ptrs(c, &s->stripe, &k->k);
+		if (ret)
+			break;
+	}
+
+	ec_stripe = genradix_ptr(&c->ec_stripes, s->stripe.key.k.p.offset);
+
+	atomic_dec(&ec_stripe->pin);
+	bch2_stripes_heap_update(c, ec_stripe,
+				 s->stripe.key.k.p.offset);
+
+err_put_writes:
+	percpu_ref_put(&c->writes);
+err:
+	open_bucket_for_each(c, &s->blocks, ob, i) {
+		ob->ec = NULL;
+		__bch2_open_bucket_put(c, ob);
+	}
+
+	bch2_open_buckets_put(c, &s->parity);
+
+	bch2_keylist_free(&s->keys, s->inline_keys);
+
+	mutex_lock(&s->h->lock);
+	list_del(&s->list);
+	mutex_unlock(&s->h->lock);
+
+	for (i = 0; i < s->stripe.key.v.nr_blocks; i++)
+		kvpfree(s->stripe.data[i], s->stripe.size << 9);
+	kfree(s);
+}
+
+static struct ec_stripe_new *ec_stripe_set_pending(struct ec_stripe_head *h)
+{
+	struct ec_stripe_new *s = h->s;
+
+	list_add(&s->list, &h->stripes);
+	h->s = NULL;
+
+	return s;
+}
+
+static void ec_stripe_new_put(struct ec_stripe_new *s)
+{
+	BUG_ON(atomic_read(&s->pin) <= 0);
+	if (atomic_dec_and_test(&s->pin))
+		ec_stripe_create(s);
+}
+
+/* have a full bucket - hand it off to be erasure coded: */
+void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
+{
+	struct ec_stripe_new *s = ob->ec;
+
+	if (ob->sectors_free)
+		s->err = -1;
+
+	ec_stripe_new_put(s);
+}
+
+void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
+{
+	struct ec_stripe_new *s = ob->ec;
+
+	s->err = -EIO;
+}
+
+void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
+{
+	struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
+	struct bch_dev *ca;
+	unsigned offset;
+
+	if (!ob)
+		return NULL;
+
+	ca	= bch_dev_bkey_exists(c, ob->ptr.dev);
+	offset	= ca->mi.bucket_size - ob->sectors_free;
+
+	return ob->ec->stripe.data[ob->ec_idx] + (offset << 9);
+}
+
+void bch2_ec_add_backpointer(struct bch_fs *c, struct write_point *wp,
+			     struct bpos pos, unsigned sectors)
+{
+	struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
+	struct ec_stripe_new *ec;
+
+	if (!ob)
+		return;
+
+	ec = ob->ec;
+	mutex_lock(&ec->lock);
+
+	if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
+				 ARRAY_SIZE(ec->inline_keys),
+				 BKEY_U64s)) {
+		BUG();
+	}
+
+	bkey_init(&ec->keys.top->k);
+	ec->keys.top->k.p	= pos;
+	bch2_key_resize(&ec->keys.top->k, sectors);
+	bch2_keylist_push(&ec->keys);
+
+	mutex_unlock(&ec->lock);
+}
+
+static int unsigned_cmp(const void *_l, const void *_r)
+{
+	unsigned l = *((const unsigned *) _l);
+	unsigned r = *((const unsigned *) _r);
+
+	return (l > r) - (l < r);
+}
+
+/* pick most common bucket size: */
+static unsigned pick_blocksize(struct bch_fs *c,
+			       struct bch_devs_mask *devs)
+{
+	struct bch_dev *ca;
+	unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
+	struct {
+		unsigned nr, size;
+	} cur = { 0, 0 }, best = { 0, 0 };
+
+	for_each_member_device_rcu(ca, c, i, devs)
+		sizes[nr++] = ca->mi.bucket_size;
+
+	sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
+
+	for (i = 0; i < nr; i++) {
+		if (sizes[i] != cur.size) {
+			if (cur.nr > best.nr)
+				best = cur;
+
+			cur.nr = 0;
+			cur.size = sizes[i];
+		}
+
+		cur.nr++;
+	}
+
+	if (cur.nr > best.nr)
+		best = cur;
+
+	return best.size;
+}
+
+int bch2_ec_stripe_new_alloc(struct bch_fs *c, struct ec_stripe_head *h)
+{
+	struct ec_stripe_new *s;
+	unsigned i;
+
+	BUG_ON(h->parity.nr != h->redundancy);
+	BUG_ON(!h->blocks.nr);
+	BUG_ON(h->parity.nr + h->blocks.nr > EC_STRIPE_MAX);
+	lockdep_assert_held(&h->lock);
+
+	s = kzalloc(sizeof(*s), GFP_KERNEL);
+	if (!s)
+		return -ENOMEM;
+
+	mutex_init(&s->lock);
+	atomic_set(&s->pin, 1);
+	s->c		= c;
+	s->h		= h;
+	s->blocks	= h->blocks;
+	s->parity	= h->parity;
+
+	memset(&h->blocks, 0, sizeof(h->blocks));
+	memset(&h->parity, 0, sizeof(h->parity));
+
+	bch2_keylist_init(&s->keys, s->inline_keys);
+
+	s->stripe.offset	= 0;
+	s->stripe.size		= h->blocksize;
+	memset(s->stripe.valid, 0xFF, sizeof(s->stripe.valid));
+
+	ec_stripe_key_init(c, &s->stripe.key,
+			   &s->blocks, &s->parity,
+			   h->blocksize);
+
+	for (i = 0; i < s->stripe.key.v.nr_blocks; i++) {
+		s->stripe.data[i] = kvpmalloc(s->stripe.size << 9, GFP_KERNEL);
+		if (!s->stripe.data[i])
+			goto err;
+	}
+
+	h->s = s;
+
+	return 0;
+err:
+	for (i = 0; i < s->stripe.key.v.nr_blocks; i++)
+		kvpfree(s->stripe.data[i], s->stripe.size << 9);
+	kfree(s);
+	return -ENOMEM;
+}
+
+static struct ec_stripe_head *
+ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
+			 unsigned algo, unsigned redundancy)
+{
+	struct ec_stripe_head *h;
+	struct bch_dev *ca;
+	unsigned i;
+
+	h = kzalloc(sizeof(*h), GFP_KERNEL);
+	if (!h)
+		return NULL;
+
+	mutex_init(&h->lock);
+	mutex_lock(&h->lock);
+	INIT_LIST_HEAD(&h->stripes);
+
+	h->target	= target;
+	h->algo		= algo;
+	h->redundancy	= redundancy;
+
+	rcu_read_lock();
+	h->devs = target_rw_devs(c, BCH_DATA_USER, target);
+
+	for_each_member_device_rcu(ca, c, i, &h->devs)
+		if (!ca->mi.durability)
+			__clear_bit(i, h->devs.d);
+
+	h->blocksize = pick_blocksize(c, &h->devs);
+
+	for_each_member_device_rcu(ca, c, i, &h->devs)
+		if (ca->mi.bucket_size == h->blocksize)
+			h->nr_active_devs++;
+
+	rcu_read_unlock();
+	list_add(&h->list, &c->ec_new_stripe_list);
+	return h;
+}
+
+void bch2_ec_stripe_head_put(struct ec_stripe_head *h)
+{
+	struct ec_stripe_new *s = NULL;
+
+	if (h->s &&
+	    bitmap_weight(h->s->blocks_allocated,
+			  h->s->blocks.nr) == h->s->blocks.nr)
+		s = ec_stripe_set_pending(h);
+
+	mutex_unlock(&h->lock);
+
+	if (s)
+		ec_stripe_new_put(s);
+}
+
+struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
+					       unsigned target,
+					       unsigned algo,
+					       unsigned redundancy)
+{
+	struct ec_stripe_head *h;
+
+	if (!redundancy)
+		return NULL;
+
+	mutex_lock(&c->ec_new_stripe_lock);
+	list_for_each_entry(h, &c->ec_new_stripe_list, list)
+		if (h->target		== target &&
+		    h->algo		== algo &&
+		    h->redundancy	== redundancy) {
+			mutex_lock(&h->lock);
+			goto found;
+		}
+
+	h = ec_new_stripe_head_alloc(c, target, algo, redundancy);
+found:
+	mutex_unlock(&c->ec_new_stripe_lock);
+	return h;
+}
+
+void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
+{
+	struct ec_stripe_head *h;
+	struct open_bucket *ob;
+	unsigned i;
+
+	mutex_lock(&c->ec_new_stripe_lock);
+	list_for_each_entry(h, &c->ec_new_stripe_list, list) {
+		struct ec_stripe_new *s = NULL;
+
+		mutex_lock(&h->lock);
+		bch2_open_buckets_stop_dev(c, ca,
+					   &h->blocks,
+					   BCH_DATA_USER);
+		bch2_open_buckets_stop_dev(c, ca,
+					   &h->parity,
+					   BCH_DATA_USER);
+
+		if (!h->s)
+			goto unlock;
+
+		open_bucket_for_each(c, &h->s->blocks, ob, i)
+			if (ob->ptr.dev == ca->dev_idx)
+				goto found;
+		open_bucket_for_each(c, &h->s->parity, ob, i)
+			if (ob->ptr.dev == ca->dev_idx)
+				goto found;
+		goto unlock;
+found:
+		h->s->err = -1;
+		s = ec_stripe_set_pending(h);
+unlock:
+		mutex_unlock(&h->lock);
+
+		if (s)
+			ec_stripe_new_put(s);
+	}
+	mutex_unlock(&c->ec_new_stripe_lock);
+}
+
+int bch2_fs_ec_start(struct bch_fs *c)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	size_t i, idx = 0;
+	int ret = 0;
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_EC, POS(0, U64_MAX), 0);
+
+	k = bch2_btree_iter_prev(&iter);
+	if (!IS_ERR_OR_NULL(k.k))
+		idx = k.k->p.offset + 1;
+	ret = bch2_btree_iter_unlock(&iter);
+	if (ret)
+		return ret;
+
+	if (!init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
+		       GFP_KERNEL))
+		return -ENOMEM;
+#if 0
+	ret = genradix_prealloc(&c->ec_stripes, idx, GFP_KERNEL);
+#else
+	for (i = 0; i < idx; i++)
+		if (!genradix_ptr_alloc(&c->ec_stripes, i, GFP_KERNEL))
+			return -ENOMEM;
+#endif
+	return 0;
+}
+
+void bch2_fs_ec_exit(struct bch_fs *c)
+{
+	struct ec_stripe_head *h;
+
+	while (1) {
+		mutex_lock(&c->ec_new_stripe_lock);
+		h = list_first_entry_or_null(&c->ec_new_stripe_list,
+					     struct ec_stripe_head, list);
+		if (h)
+			list_del(&h->list);
+		mutex_unlock(&c->ec_new_stripe_lock);
+		if (!h)
+			break;
+
+		BUG_ON(h->s);
+		BUG_ON(!list_empty(&h->stripes));
+		kfree(h);
+	}
+
+	free_heap(&c->ec_stripes_heap);
+	genradix_free(&c->ec_stripes);
+	bioset_exit(&c->ec_bioset);
+}
+
+int bch2_fs_ec_init(struct bch_fs *c)
+{
+	INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
+
+	return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
+			   BIOSET_NEED_BVECS);
+}