X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fec.c;h=8646856e4539eae46fed2634154ebb340f841a46;hb=bd9e0153342c51390ec655b4e78eda1aa1c32a84;hp=aa67af8b98160b0571ebc2bbbec8985880924e6d;hpb=33c91e2ff4e228cb618ca22d642a34ec1c2cf0ef;p=bcachefs-tools-debian

diff --git a/libbcachefs/ec.c b/libbcachefs/ec.c
index aa67af8..8646856 100644
--- a/libbcachefs/ec.c
+++ b/libbcachefs/ec.c
@@ -4,16 +4,21 @@
 
 #include "bcachefs.h"
 #include "alloc_foreground.h"
+#include "backpointers.h"
+#include "bkey_buf.h"
 #include "bset.h"
 #include "btree_gc.h"
 #include "btree_update.h"
+#include "btree_write_buffer.h"
 #include "buckets.h"
+#include "checksum.h"
 #include "disk_groups.h"
 #include "ec.h"
 #include "error.h"
-#include "io.h"
+#include "io_read.h"
 #include "keylist.h"
 #include "recovery.h"
+#include "replicas.h"
 #include "super-io.h"
 #include "util.h"
 
@@ -100,179 +105,191 @@ struct ec_bio {
 
 /* Stripes btree keys: */
 
-const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
+int bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k,
+			enum bkey_invalid_flags flags,
+			struct printbuf *err)
 {
 	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 
-	if (k.k->p.inode)
-		return "invalid stripe key";
+	if (bkey_eq(k.k->p, POS_MIN)) {
+		prt_printf(err, "stripe at POS_MIN");
+		return -BCH_ERR_invalid_bkey;
+	}
 
-	if (bkey_val_bytes(k.k) < sizeof(*s))
-		return "incorrect value size";
+	if (k.k->p.inode) {
+		prt_printf(err, "nonzero inode field");
+		return -BCH_ERR_invalid_bkey;
+	}
 
-	if (bkey_val_bytes(k.k) < sizeof(*s) ||
-	    bkey_val_u64s(k.k) < stripe_val_u64s(s))
-		return "incorrect value size";
+	if (bkey_val_u64s(k.k) < stripe_val_u64s(s)) {
+		prt_printf(err, "incorrect value size (%zu < %u)",
+		       bkey_val_u64s(k.k), stripe_val_u64s(s));
+		return -BCH_ERR_invalid_bkey;
+	}
 
-	return bch2_bkey_ptrs_invalid(c, k);
+	return bch2_bkey_ptrs_invalid(c, k, flags, err);
 }
 
 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 			 struct bkey_s_c k)
 {
 	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
-	unsigned i;
+	unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 
-	pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
+	prt_printf(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,
+	       nr_data,
 	       s->nr_redundant,
 	       s->csum_type,
 	       1U << s->csum_granularity_bits);
 
-	for (i = 0; i < s->nr_blocks; i++)
-		pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
-		       (u64) s->ptrs[i].offset,
-		       stripe_blockcount_get(s, i));
-
-	bch2_bkey_ptrs_to_text(out, c, k);
+	for (i = 0; i < s->nr_blocks; i++) {
+		const struct bch_extent_ptr *ptr = s->ptrs + i;
+		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+		u32 offset;
+		u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
+
+		prt_printf(out, " %u:%llu:%u", ptr->dev, b, offset);
+		if (i < nr_data)
+			prt_printf(out, "#%u", stripe_blockcount_get(s, i));
+		if (ptr_stale(ca, ptr))
+			prt_printf(out, " stale");
+	}
 }
 
-static int ptr_matches_stripe(struct bch_fs *c,
-			      struct bch_stripe *v,
-			      const struct bch_extent_ptr *ptr)
+/* returns blocknr in stripe that we matched: */
+static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
+						struct bkey_s_c k, unsigned *block)
 {
-	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;
-	}
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
+	unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
+
+	bkey_for_each_ptr(ptrs, ptr)
+		for (i = 0; i < nr_data; i++)
+			if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
+						      le16_to_cpu(s->sectors))) {
+				*block = i;
+				return ptr;
+			}
 
-	return -1;
+	return NULL;
 }
 
-static int extent_matches_stripe(struct bch_fs *c,
-				 struct bch_stripe *v,
-				 struct bkey_s_c k)
+static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 {
-	struct bkey_s_c_extent e;
-	const struct bch_extent_ptr *ptr;
-	int idx;
-
-	if (!bkey_extent_is_data(k.k))
-		return -1;
+	switch (k.k->type) {
+	case KEY_TYPE_extent: {
+		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+		const union bch_extent_entry *entry;
 
-	e = bkey_s_c_to_extent(k);
+		extent_for_each_entry(e, entry)
+			if (extent_entry_type(entry) ==
+			    BCH_EXTENT_ENTRY_stripe_ptr &&
+			    entry->stripe_ptr.idx == idx)
+				return true;
 
-	extent_for_each_ptr(e, ptr) {
-		idx = ptr_matches_stripe(c, v, ptr);
-		if (idx >= 0)
-			return idx;
+		break;
+	}
 	}
 
-	return -1;
+	return false;
 }
 
-static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
-{
-	struct bkey_s_c_extent e;
-	const union bch_extent_entry *entry;
-
-	if (!bkey_extent_is_data(k.k))
-		return false;
-
-	e = bkey_s_c_to_extent(k);
+/* Stripe bufs: */
 
-	extent_for_each_entry(e, entry)
-		if (extent_entry_type(entry) ==
-		    BCH_EXTENT_ENTRY_stripe_ptr &&
-		    entry->stripe_ptr.idx == idx)
-			return true;
+static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
+{
+	if (buf->key.k.type == KEY_TYPE_stripe) {
+		struct bkey_i_stripe *s = bkey_i_to_stripe(&buf->key);
+		unsigned i;
 
-	return false;
+		for (i = 0; i < s->v.nr_blocks; i++) {
+			kvpfree(buf->data[i], buf->size << 9);
+			buf->data[i] = NULL;
+		}
+	}
 }
 
-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)
+/* XXX: this is a non-mempoolified memory allocation: */
+static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
+			      unsigned offset, unsigned size)
 {
-	struct open_bucket *ob;
-	unsigned i, u64s;
+	struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
+	unsigned csum_granularity = 1U << v->csum_granularity_bits;
+	unsigned end = offset + size;
+	unsigned i;
 
-	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;
+	BUG_ON(end > le16_to_cpu(v->sectors));
 
-	open_bucket_for_each(c, blocks, ob, i)
-		s->v.ptrs[i]			= ob->ptr;
+	offset	= round_down(offset, csum_granularity);
+	end	= min_t(unsigned, le16_to_cpu(v->sectors),
+			round_up(end, csum_granularity));
 
-	open_bucket_for_each(c, parity, ob, i)
-		s->v.ptrs[blocks->nr + i]	= ob->ptr;
+	buf->offset	= offset;
+	buf->size	= end - offset;
 
-	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++;
+	memset(buf->valid, 0xFF, sizeof(buf->valid));
+
+	for (i = 0; i < v->nr_blocks; i++) {
+		buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
+		if (!buf->data[i])
+			goto err;
 	}
 
-	set_bkey_val_u64s(&s->k, u64s);
+	return 0;
+err:
+	ec_stripe_buf_exit(buf);
+	return -BCH_ERR_ENOMEM_stripe_buf;
 }
 
 /* Checksumming: */
 
-static void ec_generate_checksums(struct ec_stripe_buf *buf)
+static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
+					 unsigned block, unsigned offset)
 {
-	struct bch_stripe *v = &buf->key.v;
+	struct bch_stripe *v = &bkey_i_to_stripe(&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;
+	unsigned end = buf->offset + buf->size;
+	unsigned len = min(csum_granularity, end - offset);
+
+	BUG_ON(offset >= end);
+	BUG_ON(offset <  buf->offset);
+	BUG_ON(offset & (csum_granularity - 1));
+	BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
+	       (len & (csum_granularity - 1)));
+
+	return bch2_checksum(NULL, v->csum_type,
+			     null_nonce(),
+			     buf->data[block] + ((offset - buf->offset) << 9),
+			     len << 9);
+}
+
+static void ec_generate_checksums(struct ec_stripe_buf *buf)
+{
+	struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
+	unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 
-	if (!csum_bytes)
+	if (!v->csum_type)
 		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);
-		}
-	}
+	for (i = 0; i < v->nr_blocks; i++)
+		for (j = 0; j < csums_per_device; j++)
+			stripe_csum_set(v, i, j,
+				ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 }
 
 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 {
-	struct bch_stripe *v = &buf->key.v;
+	struct bch_stripe *v = &bkey_i_to_stripe(&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)
+	if (!v->csum_type)
 		return;
 
 	for (i = 0; i < v->nr_blocks; i++) {
@@ -285,21 +302,19 @@ static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 		while (offset < end) {
 			unsigned j = offset >> v->csum_granularity_bits;
 			unsigned len = min(csum_granularity, end - offset);
-			struct bch_csum csum;
+			struct bch_csum want = stripe_csum_get(v, i, j);
+			struct bch_csum got = ec_block_checksum(buf, i, offset);
 
-			BUG_ON(offset & (csum_granularity - 1));
-			BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
-			       ((offset + len) & (csum_granularity - 1)));
+			if (bch2_crc_cmp(want, got)) {
+				struct printbuf buf2 = PRINTBUF;
 
-			csum = bch2_checksum(NULL, v->csum_type,
-					     null_nonce(),
-					     buf->data[i] + ((offset - buf->offset) << 9),
-					     len << 9);
+				bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key));
 
-			if (memcmp(stripe_csum(v, i, j), &csum, csum_bytes)) {
-				__bcache_io_error(c,
-					"checksum error while doing reconstruct read (%u:%u)",
-					i, j);
+				bch_err_ratelimited(c,
+					"stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
+					(void *) _RET_IP_, i, j, v->csum_type,
+					want.lo, got.lo, buf2.buf);
+				printbuf_exit(&buf2);
 				clear_bit(i, buf->valid);
 				break;
 			}
@@ -313,32 +328,29 @@ static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 
 static void ec_generate_ec(struct ec_stripe_buf *buf)
 {
-	struct bch_stripe *v = &buf->key.v;
+	struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 	unsigned nr_data = v->nr_blocks - v->nr_redundant;
 	unsigned bytes = le16_to_cpu(v->sectors) << 9;
 
 	raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 }
 
-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);
+	struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
+
+	return v->nr_blocks - bitmap_weight(buf->valid, 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;
+	struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
+	unsigned i, failed[BCH_BKEY_PTRS_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,
+		bch_err_ratelimited(c,
 			"error doing reconstruct read: unable to read enough blocks");
 		return -1;
 	}
@@ -356,38 +368,66 @@ static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 static void ec_block_endio(struct bio *bio)
 {
 	struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
+	struct bch_stripe *v = &bkey_i_to_stripe(&ec_bio->buf->key)->v;
+	struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 	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"))
+	if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
+			       bio_data_dir(bio) ? "write" : "read",
+			       bch2_blk_status_to_str(bio->bi_status)))
 		clear_bit(ec_bio->idx, ec_bio->buf->valid);
 
+	if (ptr_stale(ca, ptr)) {
+		bch_err_ratelimited(ca->fs,
+				    "error %s stripe: stale pointer after io",
+				    bio_data_dir(bio) == READ ? "reading from" : "writing to");
+		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)
+			blk_opf_t opf, unsigned idx, struct closure *cl)
 {
-	struct bch_stripe *v = &buf->key.v;
+	struct bch_stripe *v = &bkey_i_to_stripe(&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);
+	enum bch_data_type data_type = idx < v->nr_blocks - v->nr_redundant
+		? BCH_DATA_user
+		: BCH_DATA_parity;
+	int rw = op_is_write(opf);
+
+	if (ptr_stale(ca, ptr)) {
+		bch_err_ratelimited(c,
+				    "error %s stripe: stale pointer",
+				    rw == READ ? "reading from" : "writing to");
+		clear_bit(idx, buf->valid);
+		return;
+	}
 
 	if (!bch2_dev_get_ioref(ca, rw)) {
 		clear_bit(idx, buf->valid);
 		return;
 	}
 
+	this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
+
 	while (offset < bytes) {
-		unsigned nr_iovecs = min_t(size_t, BIO_MAX_PAGES,
+		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(GFP_KERNEL, nr_iovecs,
+		ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
+						       nr_iovecs,
+						       opf,
+						       GFP_KERNEL,
 						       &c->ec_bioset),
 				      struct ec_bio, bio);
 
@@ -395,9 +435,6 @@ static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 		ec_bio->buf			= buf;
 		ec_bio->idx			= idx;
 
-		bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
-		bio_set_op_attrs(&ec_bio->bio, rw, 0);
-
 		ec_bio->bio.bi_iter.bi_sector	= ptr->offset + buf->offset + (offset >> 9);
 		ec_bio->bio.bi_end_io		= ec_block_endio;
 		ec_bio->bio.bi_private		= cl;
@@ -415,93 +452,86 @@ static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 	percpu_ref_put(&ca->io_ref);
 }
 
+static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
+				struct ec_stripe_buf *stripe)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
+			       POS(0, idx), BTREE_ITER_SLOTS);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+	if (k.k->type != KEY_TYPE_stripe) {
+		ret = -ENOENT;
+		goto err;
+	}
+	bkey_reassemble(&stripe->key, k);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
+{
+	return bch2_trans_run(c, get_stripe_key_trans(trans, idx, stripe));
+}
+
 /* recovery read path: */
 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 {
-	struct btree_trans trans;
-	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;
+	unsigned i, offset;
+	int ret = 0;
 
 	closure_init_stack(&cl);
 
-	BUG_ON(!rbio->pick.idx ||
-	       rbio->pick.idx - 1 >= rbio->pick.ec_nr);
+	BUG_ON(!rbio->pick.has_ec);
 
-	stripe_idx = rbio->pick.ec[rbio->pick.idx - 1].idx;
-
-	buf = kzalloc(sizeof(*buf), GFP_NOIO);
+	buf = kzalloc(sizeof(*buf), GFP_NOFS);
 	if (!buf)
-		return -ENOMEM;
-
-	bch2_trans_init(&trans, c, 0, 0);
+		return -BCH_ERR_ENOMEM_ec_read_extent;
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_EC,
-				   POS(0, stripe_idx),
-				   BTREE_ITER_SLOTS);
-	k = bch2_btree_iter_peek_slot(iter);
-	if (bkey_err(k) || k.k->type != KEY_TYPE_stripe) {
-		__bcache_io_error(c,
-			"error doing reconstruct read: stripe not found");
+	ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
+	if (ret) {
+		bch_err_ratelimited(c,
+			"error doing reconstruct read: error %i looking up stripe", ret);
 		kfree(buf);
-		return bch2_trans_exit(&trans) ?: -EIO;
+		return -EIO;
 	}
 
-	bkey_reassemble(&buf->key.k_i, k);
-	bch2_trans_exit(&trans);
-
-	v = &buf->key.v;
-
-	nr_data = v->nr_blocks - v->nr_redundant;
+	v = &bkey_i_to_stripe(&buf->key)->v;
 
-	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;
-		}
+	if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
+		bch_err_ratelimited(c,
+			"error doing reconstruct read: pointer doesn't match stripe");
+		ret = -EIO;
+		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);
+	offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
+	if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
+		bch_err_ratelimited(c,
+			"error doing reconstruct read: read is bigger than stripe");
+		ret = -EIO;
+		goto err;
+	}
 
-		if (ptr_stale(ca, ptr)) {
-			__bcache_io_error(c,
-					  "error doing reconstruct read: stale pointer");
-			clear_bit(i, buf->valid);
-			continue;
-		}
+	ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
+	if (ret)
+		goto err;
 
+	for (i = 0; i < v->nr_blocks; i++)
 		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,
+		bch_err_ratelimited(c,
 			"error doing reconstruct read: unable to read enough blocks");
 		ret = -EIO;
 		goto err;
@@ -514,10 +544,9 @@ int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 		goto err;
 
 	memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
-		      buf->data[idx] + ((offset - buf->offset) << 9));
+		      buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 err:
-	for (i = 0; i < v->nr_blocks; i++)
-		kfree(buf->data[i]);
+	ec_stripe_buf_exit(buf);
 	kfree(buf);
 	return ret;
 }
@@ -530,53 +559,108 @@ static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 
 	if (idx >= h->size) {
 		if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
-			return -ENOMEM;
+			return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 
-		spin_lock(&c->ec_stripes_heap_lock);
+		mutex_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);
+		mutex_unlock(&c->ec_stripes_heap_lock);
 
 		free_heap(&n);
 	}
 
-	if (!genradix_ptr_alloc(&c->stripes[0], idx, gfp))
-		return -ENOMEM;
+	if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
+		return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 
 	if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
-	    !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
-		return -ENOMEM;
+	    !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
+		return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 
 	return 0;
 }
 
-static int ec_stripe_mem_alloc(struct bch_fs *c,
+static int ec_stripe_mem_alloc(struct btree_trans *trans,
 			       struct btree_iter *iter)
 {
-	size_t idx = iter->pos.offset;
-	int ret = 0;
+	return allocate_dropping_locks_errcode(trans,
+			__ec_stripe_mem_alloc(trans->c, iter->pos.offset, _gfp));
+}
 
-	if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT))
-		return ret;
+/*
+ * Hash table of open stripes:
+ * Stripes that are being created or modified are kept in a hash table, so that
+ * stripe deletion can skip them.
+ */
+
+static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
+{
+	unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
+	struct ec_stripe_new *s;
+
+	hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
+		if (s->idx == idx)
+			return true;
+	return false;
+}
+
+static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
+{
+	bool ret = false;
 
-	bch2_trans_unlock(iter->trans);
-	ret = -EINTR;
+	spin_lock(&c->ec_stripes_new_lock);
+	ret = __bch2_stripe_is_open(c, idx);
+	spin_unlock(&c->ec_stripes_new_lock);
 
-	if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
-		return ret;
+	return ret;
+}
+
+static bool bch2_try_open_stripe(struct bch_fs *c,
+				 struct ec_stripe_new *s,
+				 u64 idx)
+{
+	bool ret;
+
+	spin_lock(&c->ec_stripes_new_lock);
+	ret = !__bch2_stripe_is_open(c, idx);
+	if (ret) {
+		unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
+
+		s->idx = idx;
+		hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
+	}
+	spin_unlock(&c->ec_stripes_new_lock);
 
-	return -ENOMEM;
+	return ret;
+}
+
+static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
+{
+	BUG_ON(!s->idx);
+
+	spin_lock(&c->ec_stripes_new_lock);
+	hlist_del_init(&s->hash);
+	spin_unlock(&c->ec_stripes_new_lock);
+
+	s->idx = 0;
 }
 
-static ssize_t stripe_idx_to_delete(struct bch_fs *c)
+/* Heap of all existing stripes, ordered by blocks_nonempty */
+
+static u64 stripe_idx_to_delete(struct bch_fs *c)
 {
 	ec_stripes_heap *h = &c->ec_stripes_heap;
 
-	return h->used && h->data[0].blocks_nonempty == 0
-		? h->data[0].idx : -1;
+	lockdep_assert_held(&c->ec_stripes_heap_lock);
+
+	if (h->used &&
+	    h->data[0].blocks_nonempty == 0 &&
+	    !bch2_stripe_is_open(c, h->data[0].idx))
+		return h->data[0].idx;
+
+	return 0;
 }
 
 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
@@ -592,60 +676,34 @@ static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 {
 	struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 
-	genradix_ptr(&c->stripes[0], h->data[i].idx)->heap_idx = i;
+	genradix_ptr(&c->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 stripe *m = genradix_ptr(&c->stripes[0], idx);
+	struct stripe *m = genradix_ptr(&c->stripes, idx);
 
-	BUG_ON(!m->alive);
 	BUG_ON(m->heap_idx >= h->used);
 	BUG_ON(h->data[m->heap_idx].idx != idx);
 }
 
-void bch2_stripes_heap_update(struct bch_fs *c,
-			      struct stripe *m, size_t idx)
-{
-	ec_stripes_heap *h = &c->ec_stripes_heap;
-	size_t i;
-
-	if (m->alive) {
-		heap_verify_backpointer(c, idx);
-
-		h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
-
-		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);
-	} else {
-		bch2_stripes_heap_insert(c, m, idx);
-	}
-
-	if (stripe_idx_to_delete(c) >= 0 &&
-	    !percpu_ref_is_dying(&c->writes))
-		schedule_work(&c->ec_stripe_delete_work);
-}
-
 void bch2_stripes_heap_del(struct bch_fs *c,
 			   struct stripe *m, size_t idx)
 {
+	mutex_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);
+	mutex_unlock(&c->ec_stripes_heap_lock);
 }
 
 void bch2_stripes_heap_insert(struct bch_fs *c,
 			      struct stripe *m, size_t idx)
 {
+	mutex_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) {
@@ -654,283 +712,503 @@ void bch2_stripes_heap_insert(struct bch_fs *c,
 		}),
 		 ec_stripes_heap_cmp,
 		 ec_stripes_heap_set_backpointer);
-	m->alive = true;
 
 	heap_verify_backpointer(c, idx);
+	mutex_unlock(&c->ec_stripes_heap_lock);
+}
+
+void bch2_stripes_heap_update(struct bch_fs *c,
+			      struct stripe *m, size_t idx)
+{
+	ec_stripes_heap *h = &c->ec_stripes_heap;
+	bool do_deletes;
+	size_t i;
+
+	mutex_lock(&c->ec_stripes_heap_lock);
+	heap_verify_backpointer(c, idx);
+
+	h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
+
+	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);
+
+	do_deletes = stripe_idx_to_delete(c) != 0;
+	mutex_unlock(&c->ec_stripes_heap_lock);
+
+	if (do_deletes)
+		bch2_do_stripe_deletes(c);
 }
 
 /* stripe deletion */
 
-static int ec_stripe_delete(struct bch_fs *c, size_t idx)
+static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
 {
-	return bch2_btree_delete_range(c, BTREE_ID_EC,
-				       POS(0, idx),
-				       POS(0, idx + 1),
-				       NULL);
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	struct bkey_s_c_stripe s;
+	int ret;
+
+	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, POS(0, idx),
+			       BTREE_ITER_INTENT);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	if (k.k->type != KEY_TYPE_stripe) {
+		bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
+		ret = -EINVAL;
+		goto err;
+	}
+
+	s = bkey_s_c_to_stripe(k);
+	for (unsigned i = 0; i < s.v->nr_blocks; i++)
+		if (stripe_blockcount_get(s.v, i)) {
+			struct printbuf buf = PRINTBUF;
+
+			bch2_bkey_val_to_text(&buf, c, k);
+			bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
+			printbuf_exit(&buf);
+			ret = -EINVAL;
+			goto err;
+		}
+
+	ret = bch2_btree_delete_at(trans, &iter, 0);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
 }
 
 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);
-	mutex_lock(&c->ec_stripe_create_lock);
+	struct btree_trans *trans = bch2_trans_get(c);
+	int ret;
+	u64 idx;
 
 	while (1) {
-		spin_lock(&c->ec_stripes_heap_lock);
+		mutex_lock(&c->ec_stripes_heap_lock);
 		idx = stripe_idx_to_delete(c);
-		spin_unlock(&c->ec_stripes_heap_lock);
+		mutex_unlock(&c->ec_stripes_heap_lock);
 
-		if (idx < 0)
+		if (!idx)
 			break;
 
-		if (ec_stripe_delete(c, idx))
+		ret = commit_do(trans, NULL, NULL, BTREE_INSERT_NOFAIL,
+				ec_stripe_delete(trans, idx));
+		if (ret) {
+			bch_err_fn(c, ret);
 			break;
+		}
 	}
 
-	mutex_unlock(&c->ec_stripe_create_lock);
-	up_read(&c->gc_lock);
+	bch2_trans_put(trans);
+
+	bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
+}
+
+void bch2_do_stripe_deletes(struct bch_fs *c)
+{
+	if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
+	    !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work))
+		bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 }
 
 /* stripe creation: */
 
-static int ec_stripe_bkey_insert(struct bch_fs *c,
-				 struct bkey_i_stripe *stripe)
+static int ec_stripe_key_update(struct btree_trans *trans,
+				struct bkey_i_stripe *new,
+				bool create)
 {
-	struct btree_trans trans;
-	struct btree_iter *iter;
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
 	struct bkey_s_c k;
 	int ret;
 
-	bch2_trans_init(&trans, c, 0, 0);
-retry:
-	bch2_trans_begin(&trans);
-
-	/* XXX: start pos hint */
-	for_each_btree_key(&trans, iter, BTREE_ID_EC, POS_MIN,
-			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
-		if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0)
-			break;
+	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
+			       new->k.p, BTREE_ITER_INTENT);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
 
-		if (bkey_deleted(k.k))
-			goto found_slot;
+	if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
+		bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
+				     create ? "creating" : "updating",
+				     bch2_bkey_types[k.k->type]);
+		ret = -EINVAL;
+		goto err;
 	}
 
-	if (!ret)
-		ret = -ENOSPC;
-	goto err;
-found_slot:
-	ret = ec_stripe_mem_alloc(c, iter);
-	if (ret)
-		goto err;
+	if (k.k->type == KEY_TYPE_stripe) {
+		const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
+		unsigned i;
 
-	stripe->k.p = iter->pos;
+		if (old->nr_blocks != new->v.nr_blocks) {
+			bch_err(c, "error updating stripe: nr_blocks does not match");
+			ret = -EINVAL;
+			goto err;
+		}
 
-	bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &stripe->k_i));
+		for (i = 0; i < new->v.nr_blocks; i++) {
+			unsigned v = stripe_blockcount_get(old, i);
 
-	ret = bch2_trans_commit(&trans, NULL, NULL,
-				BTREE_INSERT_ATOMIC|
-				BTREE_INSERT_NOFAIL);
-err:
-	if (ret == -EINTR)
-		goto retry;
-	bch2_trans_exit(&trans);
+			BUG_ON(v &&
+			       (old->ptrs[i].dev != new->v.ptrs[i].dev ||
+				old->ptrs[i].gen != new->v.ptrs[i].gen ||
+				old->ptrs[i].offset != new->v.ptrs[i].offset));
 
+			stripe_blockcount_set(&new->v, i, v);
+		}
+	}
+
+	ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
+err:
+	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
 
-static void extent_stripe_ptr_add(struct bkey_s_extent e,
-				  struct ec_stripe_buf *s,
-				  struct bch_extent_ptr *ptr,
-				  unsigned block)
+static int ec_stripe_update_extent(struct btree_trans *trans,
+				   struct bpos bucket, u8 gen,
+				   struct ec_stripe_buf *s,
+				   struct bpos *bp_pos)
 {
-	struct bch_extent_stripe_ptr *dst = (void *) ptr;
-	union bch_extent_entry *end = extent_entry_last(e);
+	struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
+	struct bch_fs *c = trans->c;
+	struct bch_backpointer bp;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	const struct bch_extent_ptr *ptr_c;
+	struct bch_extent_ptr *ptr, *ec_ptr = NULL;
+	struct bch_extent_stripe_ptr stripe_ptr;
+	struct bkey_i *n;
+	int ret, dev, block;
+
+	ret = bch2_get_next_backpointer(trans, bucket, gen,
+				bp_pos, &bp, BTREE_ITER_CACHED);
+	if (ret)
+		return ret;
+	if (bpos_eq(*bp_pos, SPOS_MAX))
+		return 0;
 
-	memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
-	e.k->u64s += sizeof(*dst) / sizeof(u64);
+	if (bp.level) {
+		struct printbuf buf = PRINTBUF;
+		struct btree_iter node_iter;
+		struct btree *b;
 
-	*dst = (struct bch_extent_stripe_ptr) {
-		.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
-		.block		= block,
-		.idx		= s->key.k.p.offset,
-	};
-}
+		b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
+		bch2_trans_iter_exit(trans, &node_iter);
 
-static int ec_stripe_update_ptrs(struct bch_fs *c,
-				 struct ec_stripe_buf *s,
-				 struct bkey *pos)
-{
-	struct btree_trans trans;
-	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_trans_init(&trans, c, BTREE_ITER_MAX, 0);
-
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
-				   bkey_start_pos(pos),
-				   BTREE_ITER_INTENT);
-
-	while ((k = bch2_btree_iter_peek(iter)).k &&
-	       !(ret = bkey_err(k)) &&
-	       bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
-		if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
-			bch2_btree_iter_next(iter);
-			continue;
-		}
+		if (!b)
+			return 0;
 
-		idx = extent_matches_stripe(c, &s->key.v, k);
-		if (idx < 0) {
-			bch2_btree_iter_next(iter);
-			continue;
-		}
+		prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
+		bch2_backpointer_to_text(&buf, &bp);
+
+		bch2_fs_inconsistent(c, "%s", buf.buf);
+		printbuf_exit(&buf);
+		return -EIO;
+	}
 
-		bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
+	k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT);
+	ret = bkey_err(k);
+	if (ret)
+		return ret;
+	if (!k.k) {
+		/*
+		 * extent no longer exists - we could flush the btree
+		 * write buffer and retry to verify, but no need:
+		 */
+		return 0;
+	}
 
-		dev = s->key.v.ptrs[idx].dev;
+	if (extent_has_stripe_ptr(k, s->key.k.p.offset))
+		goto out;
 
-		bkey_reassemble(&tmp.k, k);
-		e = bkey_i_to_s_extent(&tmp.k);
+	ptr_c = bkey_matches_stripe(v, k, &block);
+	/*
+	 * It doesn't generally make sense to erasure code cached ptrs:
+	 * XXX: should we be incrementing a counter?
+	 */
+	if (!ptr_c || ptr_c->cached)
+		goto out;
 
-		extent_for_each_ptr(e, ptr)
-			if (ptr->dev != dev)
-				ptr->cached = true;
+	dev = v->ptrs[block].dev;
 
-		ptr = (void *) bch2_extent_has_device(e.c, dev);
-		BUG_ON(!ptr);
+	n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
+	ret = PTR_ERR_OR_ZERO(n);
+	if (ret)
+		goto out;
 
-		extent_stripe_ptr_add(e, s, ptr, idx);
+	bkey_reassemble(n, k);
 
-		bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &tmp.k));
+	bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
+	ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev);
+	BUG_ON(!ec_ptr);
 
-		ret = bch2_trans_commit(&trans, NULL, NULL,
-					BTREE_INSERT_ATOMIC|
-					BTREE_INSERT_NOFAIL|
-					BTREE_INSERT_USE_RESERVE);
-		if (ret == -EINTR)
-			ret = 0;
-		if (ret)
-			break;
-	}
+	stripe_ptr = (struct bch_extent_stripe_ptr) {
+		.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
+		.block		= block,
+		.redundancy	= v->nr_redundant,
+		.idx		= s->key.k.p.offset,
+	};
 
-	bch2_trans_exit(&trans);
+	__extent_entry_insert(n,
+			(union bch_extent_entry *) ec_ptr,
+			(union bch_extent_entry *) &stripe_ptr);
 
+	ret = bch2_trans_update(trans, &iter, n, 0);
+out:
+	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
 
-/*
- * data buckets of new stripe all written: create the stripe
+static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
+				   unsigned block)
+{
+	struct bch_fs *c = trans->c;
+	struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
+	struct bch_extent_ptr bucket = v->ptrs[block];
+	struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
+	struct bpos bp_pos = POS_MIN;
+	int ret = 0;
+
+	while (1) {
+		ret = commit_do(trans, NULL, NULL,
+				BTREE_INSERT_NOCHECK_RW|
+				BTREE_INSERT_NOFAIL,
+			ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
+						s, &bp_pos));
+		if (ret)
+			break;
+		if (bkey_eq(bp_pos, POS_MAX))
+			break;
+
+		bp_pos = bpos_nosnap_successor(bp_pos);
+	}
+
+	return ret;
+}
+
+static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
+{
+	struct btree_trans *trans = bch2_trans_get(c);
+	struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
+	unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
+	int ret = 0;
+
+	ret = bch2_btree_write_buffer_flush(trans);
+	if (ret)
+		goto err;
+
+	for (i = 0; i < nr_data; i++) {
+		ret = ec_stripe_update_bucket(trans, s, i);
+		if (ret)
+			break;
+	}
+err:
+	bch2_trans_put(trans);
+
+	return ret;
+}
+
+static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
+				       struct ec_stripe_new *s,
+				       unsigned block,
+				       struct open_bucket *ob)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
+	unsigned offset = ca->mi.bucket_size - ob->sectors_free;
+	int ret;
+
+	if (!bch2_dev_get_ioref(ca, WRITE)) {
+		s->err = -BCH_ERR_erofs_no_writes;
+		return;
+	}
+
+	memset(s->new_stripe.data[block] + (offset << 9),
+	       0,
+	       ob->sectors_free << 9);
+
+	ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
+			ob->bucket * ca->mi.bucket_size + offset,
+			ob->sectors_free,
+			GFP_KERNEL, 0);
+
+	percpu_ref_put(&ca->io_ref);
+
+	if (ret)
+		s->err = ret;
+}
+
+void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s)
+{
+	if (s->idx)
+		bch2_stripe_close(c, s);
+	kfree(s);
+}
+
+/*
+ * data buckets of new stripe all written: create the stripe
  */
 static void ec_stripe_create(struct ec_stripe_new *s)
 {
 	struct bch_fs *c = s->c;
 	struct open_bucket *ob;
-	struct bkey_i *k;
-	struct bch_stripe *v = &s->stripe.key.v;
+	struct bch_stripe *v = &bkey_i_to_stripe(&s->new_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);
+	closure_sync(&s->iodone);
+
+	if (!s->err) {
+		for (i = 0; i < nr_data; i++)
+			if (s->blocks[i]) {
+				ob = c->open_buckets + s->blocks[i];
+
+				if (ob->sectors_free)
+					zero_out_rest_of_ec_bucket(c, s, i, ob);
+			}
+	}
 
 	if (s->err) {
-		bch_err(c, "error creating stripe: error writing data buckets");
+		if (!bch2_err_matches(s->err, EROFS))
+			bch_err(c, "error creating stripe: error writing data buckets");
 		goto err;
 	}
 
-	if (!percpu_ref_tryget(&c->writes))
-		goto err;
+	if (s->have_existing_stripe) {
+		ec_validate_checksums(c, &s->existing_stripe);
+
+		if (ec_do_recov(c, &s->existing_stripe)) {
+			bch_err(c, "error creating stripe: error reading existing stripe");
+			goto err;
+		}
 
-	BUG_ON(bitmap_weight(s->blocks_allocated,
-			     s->blocks.nr) != s->blocks.nr);
+		for (i = 0; i < nr_data; i++)
+			if (stripe_blockcount_get(&bkey_i_to_stripe(&s->existing_stripe.key)->v, i))
+				swap(s->new_stripe.data[i],
+				     s->existing_stripe.data[i]);
 
-	ec_generate_ec(&s->stripe);
+		ec_stripe_buf_exit(&s->existing_stripe);
+	}
 
-	ec_generate_checksums(&s->stripe);
+	BUG_ON(!s->allocated);
+	BUG_ON(!s->idx);
 
-	/* write p/q: */
-	for (i = nr_data; i < v->nr_blocks; i++)
-		ec_block_io(c, &s->stripe, REQ_OP_WRITE, i, &cl);
+	ec_generate_ec(&s->new_stripe);
 
-	closure_sync(&cl);
+	ec_generate_checksums(&s->new_stripe);
 
+	/* write p/q: */
 	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;
-		}
+		ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
+	closure_sync(&s->iodone);
 
-	mutex_lock(&c->ec_stripe_create_lock);
+	if (ec_nr_failed(&s->new_stripe)) {
+		bch_err(c, "error creating stripe: error writing redundancy buckets");
+		goto err;
+	}
 
-	ret = ec_stripe_bkey_insert(c, &s->stripe.key);
+	ret = bch2_trans_do(c, &s->res, NULL,
+			    BTREE_INSERT_NOCHECK_RW|
+			    BTREE_INSERT_NOFAIL,
+			    ec_stripe_key_update(trans,
+					bkey_i_to_stripe(&s->new_stripe.key),
+					!s->have_existing_stripe));
 	if (ret) {
 		bch_err(c, "error creating stripe: error creating stripe key");
-		goto err_unlock;
+		goto err;
 	}
 
-	for_each_keylist_key(&s->keys, k) {
-		ret = ec_stripe_update_ptrs(c, &s->stripe, &k->k);
-		if (ret)
-			break;
+	ret = ec_stripe_update_extents(c, &s->new_stripe);
+	if (ret) {
+		bch_err_msg(c, ret, "creating stripe: error updating pointers");
+		goto err;
 	}
-
-err_unlock:
-	mutex_unlock(&c->ec_stripe_create_lock);
-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_disk_reservation_put(c, &s->res);
 
-	bch2_keylist_free(&s->keys, s->inline_keys);
+	for (i = 0; i < v->nr_blocks; i++)
+		if (s->blocks[i]) {
+			ob = c->open_buckets + s->blocks[i];
+
+			if (i < nr_data) {
+				ob->ec = NULL;
+				__bch2_open_bucket_put(c, ob);
+			} else {
+				bch2_open_bucket_put(c, ob);
+			}
+		}
 
-	mutex_lock(&s->h->lock);
+	mutex_lock(&c->ec_stripe_new_lock);
 	list_del(&s->list);
-	mutex_unlock(&s->h->lock);
+	mutex_unlock(&c->ec_stripe_new_lock);
+	wake_up(&c->ec_stripe_new_wait);
 
-	for (i = 0; i < s->stripe.key.v.nr_blocks; i++)
-		kvpfree(s->stripe.data[i], s->stripe.size << 9);
-	kfree(s);
+	ec_stripe_buf_exit(&s->existing_stripe);
+	ec_stripe_buf_exit(&s->new_stripe);
+	closure_debug_destroy(&s->iodone);
+
+	ec_stripe_new_put(c, s, STRIPE_REF_stripe);
 }
 
-static struct ec_stripe_new *ec_stripe_set_pending(struct ec_stripe_head *h)
+static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
 {
-	struct ec_stripe_new *s = h->s;
+	struct ec_stripe_new *s;
 
-	list_add(&s->list, &h->stripes);
-	h->s = NULL;
+	mutex_lock(&c->ec_stripe_new_lock);
+	list_for_each_entry(s, &c->ec_stripe_new_list, list)
+		if (!atomic_read(&s->ref[STRIPE_REF_io]))
+			goto out;
+	s = NULL;
+out:
+	mutex_unlock(&c->ec_stripe_new_lock);
 
 	return s;
 }
 
-static void ec_stripe_new_put(struct ec_stripe_new *s)
+static void ec_stripe_create_work(struct work_struct *work)
 {
-	BUG_ON(atomic_read(&s->pin) <= 0);
-	if (atomic_dec_and_test(&s->pin))
+	struct bch_fs *c = container_of(work,
+		struct bch_fs, ec_stripe_create_work);
+	struct ec_stripe_new *s;
+
+	while ((s = get_pending_stripe(c)))
 		ec_stripe_create(s);
+
+	bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
 }
 
-/* have a full bucket - hand it off to be erasure coded: */
-void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
+void bch2_ec_do_stripe_creates(struct bch_fs *c)
 {
-	struct ec_stripe_new *s = ob->ec;
+	bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
+
+	if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
+		bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
+}
+
+static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
+{
+	struct ec_stripe_new *s = h->s;
 
-	if (ob->sectors_free)
-		s->err = -1;
+	BUG_ON(!s->allocated && !s->err);
 
-	ec_stripe_new_put(s);
+	h->s		= NULL;
+	s->pending	= true;
+
+	mutex_lock(&c->ec_stripe_new_lock);
+	list_add(&s->list, &c->ec_stripe_new_list);
+	mutex_unlock(&c->ec_stripe_new_lock);
+
+	ec_stripe_new_put(c, s, STRIPE_REF_io);
 }
 
 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
@@ -949,36 +1227,12 @@ void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
 	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();
-	}
+	BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
 
-	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);
+	ca	= bch_dev_bkey_exists(c, ob->dev);
+	offset	= ca->mi.bucket_size - ob->sectors_free;
 
-	mutex_unlock(&ec->lock);
+	return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
 }
 
 static int unsigned_cmp(const void *_l, const void *_r)
@@ -1022,59 +1276,69 @@ static unsigned pick_blocksize(struct bch_fs *c,
 	return best.size;
 }
 
-int bch2_ec_stripe_new_alloc(struct bch_fs *c, struct ec_stripe_head *h)
+static bool may_create_new_stripe(struct bch_fs *c)
+{
+	return false;
+}
+
+static void ec_stripe_key_init(struct bch_fs *c,
+			       struct bkey_i *k,
+			       unsigned nr_data,
+			       unsigned nr_parity,
+			       unsigned stripe_size)
+{
+	struct bkey_i_stripe *s = bkey_stripe_init(k);
+	unsigned u64s;
+
+	s->v.sectors			= cpu_to_le16(stripe_size);
+	s->v.algorithm			= 0;
+	s->v.nr_blocks			= nr_data + nr_parity;
+	s->v.nr_redundant		= nr_parity;
+	s->v.csum_granularity_bits	= ilog2(c->opts.encoded_extent_max >> 9);
+	s->v.csum_type			= BCH_CSUM_crc32c;
+	s->v.pad			= 0;
+
+	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);
+}
+
+static int ec_new_stripe_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;
+		return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
 
 	mutex_init(&s->lock);
-	atomic_set(&s->pin, 1);
+	closure_init(&s->iodone, NULL);
+	atomic_set(&s->ref[STRIPE_REF_stripe], 1);
+	atomic_set(&s->ref[STRIPE_REF_io], 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));
+	s->nr_data	= min_t(unsigned, h->nr_active_devs,
+				BCH_BKEY_PTRS_MAX) - h->redundancy;
+	s->nr_parity	= h->redundancy;
 
-	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;
-	}
+	ec_stripe_key_init(c, &s->new_stripe.key,
+			   s->nr_data, s->nr_parity, h->blocksize);
 
 	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)
+			 unsigned algo, unsigned redundancy,
+			 enum bch_watermark watermark)
 {
 	struct ec_stripe_head *h;
 	struct bch_dev *ca;
@@ -1085,15 +1349,15 @@ ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
 		return NULL;
 
 	mutex_init(&h->lock);
-	mutex_lock(&h->lock);
-	INIT_LIST_HEAD(&h->stripes);
+	BUG_ON(!mutex_trylock(&h->lock));
 
 	h->target	= target;
 	h->algo		= algo;
 	h->redundancy	= redundancy;
+	h->watermark	= watermark;
 
 	rcu_read_lock();
-	h->devs = target_rw_devs(c, BCH_DATA_USER, target);
+	h->devs = target_rw_devs(c, BCH_DATA_user, target);
 
 	for_each_member_device_rcu(ca, c, i, &h->devs)
 		if (!ca->mi.durability)
@@ -1106,266 +1370,597 @@ ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
 			h->nr_active_devs++;
 
 	rcu_read_unlock();
-	list_add(&h->list, &c->ec_new_stripe_list);
+	list_add(&h->list, &c->ec_stripe_head_list);
 	return h;
 }
 
-void bch2_ec_stripe_head_put(struct ec_stripe_head *h)
+void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
 {
-	struct ec_stripe_new *s = NULL;
-
 	if (h->s &&
+	    h->s->allocated &&
 	    bitmap_weight(h->s->blocks_allocated,
-			  h->s->blocks.nr) == h->s->blocks.nr)
-		s = ec_stripe_set_pending(h);
+			  h->s->nr_data) == h->s->nr_data)
+		ec_stripe_set_pending(c, 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)
+static struct ec_stripe_head *
+__bch2_ec_stripe_head_get(struct btree_trans *trans,
+			  unsigned target,
+			  unsigned algo,
+			  unsigned redundancy,
+			  enum bch_watermark watermark)
 {
+	struct bch_fs *c = trans->c;
 	struct ec_stripe_head *h;
+	int ret;
 
 	if (!redundancy)
 		return NULL;
 
-	mutex_lock(&c->ec_new_stripe_lock);
-	list_for_each_entry(h, &c->ec_new_stripe_list, list)
+	ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
+	if (ret)
+		return ERR_PTR(ret);
+
+	if (test_bit(BCH_FS_GOING_RO, &c->flags)) {
+		h = ERR_PTR(-BCH_ERR_erofs_no_writes);
+		goto found;
+	}
+
+	list_for_each_entry(h, &c->ec_stripe_head_list, list)
 		if (h->target		== target &&
 		    h->algo		== algo &&
-		    h->redundancy	== redundancy) {
-			mutex_lock(&h->lock);
+		    h->redundancy	== redundancy &&
+		    h->watermark	== watermark) {
+			ret = bch2_trans_mutex_lock(trans, &h->lock);
+			if (ret)
+				h = ERR_PTR(ret);
 			goto found;
 		}
 
-	h = ec_new_stripe_head_alloc(c, target, algo, redundancy);
+	h = ec_new_stripe_head_alloc(c, target, algo, redundancy, watermark);
 found:
-	mutex_unlock(&c->ec_new_stripe_lock);
+	mutex_unlock(&c->ec_stripe_head_lock);
 	return h;
 }
 
-void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
+static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
+				    enum bch_watermark watermark, struct closure *cl)
 {
-	struct ec_stripe_head *h;
+	struct bch_fs *c = trans->c;
+	struct bch_devs_mask devs = h->devs;
 	struct open_bucket *ob;
-	unsigned i;
+	struct open_buckets buckets;
+	struct bch_stripe *v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
+	unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
+	bool have_cache = true;
+	int ret = 0;
 
-	mutex_lock(&c->ec_new_stripe_lock);
-	list_for_each_entry(h, &c->ec_new_stripe_list, list) {
-		struct ec_stripe_new *s = NULL;
+	BUG_ON(v->nr_blocks	!= h->s->nr_data + h->s->nr_parity);
+	BUG_ON(v->nr_redundant	!= h->s->nr_parity);
 
-		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);
+	for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) {
+		__clear_bit(v->ptrs[i].dev, devs.d);
+		if (i < h->s->nr_data)
+			nr_have_data++;
+		else
+			nr_have_parity++;
+	}
 
-		if (!h->s)
-			goto unlock;
+	BUG_ON(nr_have_data	> h->s->nr_data);
+	BUG_ON(nr_have_parity	> h->s->nr_parity);
+
+	buckets.nr = 0;
+	if (nr_have_parity < h->s->nr_parity) {
+		ret = bch2_bucket_alloc_set_trans(trans, &buckets,
+					    &h->parity_stripe,
+					    &devs,
+					    h->s->nr_parity,
+					    &nr_have_parity,
+					    &have_cache, 0,
+					    BCH_DATA_parity,
+					    watermark,
+					    cl);
+
+		open_bucket_for_each(c, &buckets, ob, i) {
+			j = find_next_zero_bit(h->s->blocks_gotten,
+					       h->s->nr_data + h->s->nr_parity,
+					       h->s->nr_data);
+			BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
+
+			h->s->blocks[j] = buckets.v[i];
+			v->ptrs[j] = bch2_ob_ptr(c, ob);
+			__set_bit(j, h->s->blocks_gotten);
+		}
 
-		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 (ret)
+			return ret;
+	}
 
-		if (s)
-			ec_stripe_new_put(s);
+	buckets.nr = 0;
+	if (nr_have_data < h->s->nr_data) {
+		ret = bch2_bucket_alloc_set_trans(trans, &buckets,
+					    &h->block_stripe,
+					    &devs,
+					    h->s->nr_data,
+					    &nr_have_data,
+					    &have_cache, 0,
+					    BCH_DATA_user,
+					    watermark,
+					    cl);
+
+		open_bucket_for_each(c, &buckets, ob, i) {
+			j = find_next_zero_bit(h->s->blocks_gotten,
+					       h->s->nr_data, 0);
+			BUG_ON(j >= h->s->nr_data);
+
+			h->s->blocks[j] = buckets.v[i];
+			v->ptrs[j] = bch2_ob_ptr(c, ob);
+			__set_bit(j, h->s->blocks_gotten);
+		}
+
+		if (ret)
+			return ret;
 	}
-	mutex_unlock(&c->ec_new_stripe_lock);
+
+	return 0;
+}
+
+/* XXX: doesn't obey target: */
+static s64 get_existing_stripe(struct bch_fs *c,
+			       struct ec_stripe_head *head)
+{
+	ec_stripes_heap *h = &c->ec_stripes_heap;
+	struct stripe *m;
+	size_t heap_idx;
+	u64 stripe_idx;
+	s64 ret = -1;
+
+	if (may_create_new_stripe(c))
+		return -1;
+
+	mutex_lock(&c->ec_stripes_heap_lock);
+	for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
+		/* No blocks worth reusing, stripe will just be deleted: */
+		if (!h->data[heap_idx].blocks_nonempty)
+			continue;
+
+		stripe_idx = h->data[heap_idx].idx;
+
+		m = genradix_ptr(&c->stripes, stripe_idx);
+
+		if (m->algorithm	== head->algo &&
+		    m->nr_redundant	== head->redundancy &&
+		    m->sectors		== head->blocksize &&
+		    m->blocks_nonempty	< m->nr_blocks - m->nr_redundant &&
+		    bch2_try_open_stripe(c, head->s, stripe_idx)) {
+			ret = stripe_idx;
+			break;
+		}
+	}
+	mutex_unlock(&c->ec_stripes_heap_lock);
+	return ret;
 }
 
-static int __bch2_stripe_write_key(struct btree_trans *trans,
-				   struct btree_iter *iter,
-				   struct stripe *m,
-				   size_t idx,
-				   struct bkey_i_stripe *new_key,
-				   unsigned flags)
+static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
 {
 	struct bch_fs *c = trans->c;
-	struct bkey_s_c k;
+	struct bch_stripe *new_v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
+	struct bch_stripe *existing_v;
 	unsigned i;
+	s64 idx;
 	int ret;
 
-	bch2_btree_iter_set_pos(iter, POS(0, idx));
+	/*
+	 * If we can't allocate a new stripe, and there's no stripes with empty
+	 * blocks for us to reuse, that means we have to wait on copygc:
+	 */
+	idx = get_existing_stripe(c, h);
+	if (idx < 0)
+		return -BCH_ERR_stripe_alloc_blocked;
 
-	k = bch2_btree_iter_peek_slot(iter);
-	ret = bkey_err(k);
-	if (ret)
+	ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
+	if (ret) {
+		bch2_stripe_close(c, h->s);
+		if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
+			bch2_fs_fatal_error(c, "error reading stripe key: %s", bch2_err_str(ret));
 		return ret;
+	}
 
-	if (k.k->type != KEY_TYPE_stripe)
-		return -EIO;
+	existing_v = &bkey_i_to_stripe(&h->s->existing_stripe.key)->v;
 
-	bkey_reassemble(&new_key->k_i, k);
+	BUG_ON(existing_v->nr_redundant != h->s->nr_parity);
+	h->s->nr_data = existing_v->nr_blocks -
+		existing_v->nr_redundant;
 
-	spin_lock(&c->ec_stripes_heap_lock);
+	ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
+	if (ret) {
+		bch2_stripe_close(c, h->s);
+		return ret;
+	}
 
-	for (i = 0; i < new_key->v.nr_blocks; i++)
-		stripe_blockcount_set(&new_key->v, i,
-				      m->block_sectors[i]);
-	m->dirty = false;
+	BUG_ON(h->s->existing_stripe.size != h->blocksize);
+	BUG_ON(h->s->existing_stripe.size != le16_to_cpu(existing_v->sectors));
 
-	spin_unlock(&c->ec_stripes_heap_lock);
+	/*
+	 * Free buckets we initially allocated - they might conflict with
+	 * blocks from the stripe we're reusing:
+	 */
+	for_each_set_bit(i, h->s->blocks_gotten, new_v->nr_blocks) {
+		bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
+		h->s->blocks[i] = 0;
+	}
+	memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
+	memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
 
-	bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &new_key->k_i));
+	for (i = 0; i < existing_v->nr_blocks; i++) {
+		if (stripe_blockcount_get(existing_v, i)) {
+			__set_bit(i, h->s->blocks_gotten);
+			__set_bit(i, h->s->blocks_allocated);
+		}
 
-	return bch2_trans_commit(trans, NULL, NULL,
-				 BTREE_INSERT_NOFAIL|flags);
-}
+		ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
+	}
 
-int bch2_stripes_write(struct bch_fs *c, unsigned flags, bool *wrote)
-{
-	struct btree_trans trans;
-	struct btree_iter *iter;
-	struct genradix_iter giter;
-	struct bkey_i_stripe *new_key;
-	struct stripe *m;
-	int ret = 0;
+	bkey_copy(&h->s->new_stripe.key, &h->s->existing_stripe.key);
+	h->s->have_existing_stripe = true;
 
-	new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
-	BUG_ON(!new_key);
+	return 0;
+}
 
-	bch2_trans_init(&trans, c, 0, 0);
+static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	struct bpos min_pos = POS(0, 1);
+	struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
+	int ret;
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS_MIN,
-				   BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+	if (!h->s->res.sectors) {
+		ret = bch2_disk_reservation_get(c, &h->s->res,
+					h->blocksize,
+					h->s->nr_parity,
+					BCH_DISK_RESERVATION_NOFAIL);
+		if (ret)
+			return ret;
+	}
 
-	genradix_for_each(&c->stripes[0], giter, m) {
-		if (!m->dirty)
-			continue;
+	for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
+			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
+		if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
+			if (start_pos.offset) {
+				start_pos = min_pos;
+				bch2_btree_iter_set_pos(&iter, start_pos);
+				continue;
+			}
 
-		ret = __bch2_stripe_write_key(&trans, iter, m, giter.pos,
-					      new_key, flags);
-		if (ret)
+			ret = -BCH_ERR_ENOSPC_stripe_create;
 			break;
+		}
 
-		*wrote = true;
+		if (bkey_deleted(k.k) &&
+		    bch2_try_open_stripe(c, h->s, k.k->p.offset))
+			break;
 	}
 
-	bch2_trans_exit(&trans);
+	c->ec_stripe_hint = iter.pos.offset;
 
-	kfree(new_key);
+	if (ret)
+		goto err;
 
+	ret = ec_stripe_mem_alloc(trans, &iter);
+	if (ret) {
+		bch2_stripe_close(c, h->s);
+		goto err;
+	}
+
+	h->s->new_stripe.key.k.p = iter.pos;
+out:
+	bch2_trans_iter_exit(trans, &iter);
 	return ret;
+err:
+	bch2_disk_reservation_put(c, &h->s->res);
+	goto out;
 }
 
-int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
+struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
+					       unsigned target,
+					       unsigned algo,
+					       unsigned redundancy,
+					       enum bch_watermark watermark,
+					       struct closure *cl)
 {
-	struct journal_key *i;
-	struct btree_trans trans;
-	struct btree_iter *iter;
-	struct bkey_s_c k;
+	struct bch_fs *c = trans->c;
+	struct ec_stripe_head *h;
+	bool waiting = false;
 	int ret;
 
-	ret = bch2_fs_ec_start(c);
+	h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, watermark);
+	if (!h)
+		bch_err(c, "no stripe head");
+	if (IS_ERR_OR_NULL(h))
+		return h;
+
+	if (!h->s) {
+		ret = ec_new_stripe_alloc(c, h);
+		if (ret) {
+			bch_err(c, "failed to allocate new stripe");
+			goto err;
+		}
+	}
+
+	if (h->s->allocated)
+		goto allocated;
+
+	if (h->s->have_existing_stripe)
+		goto alloc_existing;
+
+	/* First, try to allocate a full stripe: */
+	ret =   new_stripe_alloc_buckets(trans, h, BCH_WATERMARK_stripe, NULL) ?:
+		__bch2_ec_stripe_head_reserve(trans, h);
+	if (!ret)
+		goto allocate_buf;
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
+	    bch2_err_matches(ret, ENOMEM))
+		goto err;
+
+	/*
+	 * Not enough buckets available for a full stripe: we must reuse an
+	 * existing stripe:
+	 */
+	while (1) {
+		ret = __bch2_ec_stripe_head_reuse(trans, h);
+		if (!ret)
+			break;
+		if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
+			goto err;
+
+		if (watermark == BCH_WATERMARK_copygc) {
+			ret =   new_stripe_alloc_buckets(trans, h, watermark, NULL) ?:
+				__bch2_ec_stripe_head_reserve(trans, h);
+			if (ret)
+				goto err;
+			goto allocate_buf;
+		}
+
+		/* XXX freelist_wait? */
+		closure_wait(&c->freelist_wait, cl);
+		waiting = true;
+	}
+
+	if (waiting)
+		closure_wake_up(&c->freelist_wait);
+alloc_existing:
+	/*
+	 * Retry allocating buckets, with the watermark for this
+	 * particular write:
+	 */
+	ret = new_stripe_alloc_buckets(trans, h, watermark, cl);
 	if (ret)
-		return ret;
+		goto err;
 
-	bch2_trans_init(&trans, c, 0, 0);
+allocate_buf:
+	ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
+	if (ret)
+		goto err;
 
-	for_each_btree_key(&trans, iter, BTREE_ID_EC, POS_MIN, 0, k, ret)
-		bch2_mark_key(c, k, 0, NULL, 0,
-			      BCH_BUCKET_MARK_ALLOC_READ|
-			      BCH_BUCKET_MARK_NOATOMIC);
+	h->s->allocated = true;
+allocated:
+	BUG_ON(!h->s->idx);
+	BUG_ON(!h->s->new_stripe.data[0]);
+	BUG_ON(trans->restarted);
+	return h;
+err:
+	bch2_ec_stripe_head_put(c, h);
+	return ERR_PTR(ret);
+}
 
-	ret = bch2_trans_exit(&trans) ?: ret;
-	if (ret) {
-		bch_err(c, "error reading stripes: %i", ret);
-		return ret;
+static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
+{
+	struct ec_stripe_head *h;
+	struct open_bucket *ob;
+	unsigned i;
+
+	mutex_lock(&c->ec_stripe_head_lock);
+	list_for_each_entry(h, &c->ec_stripe_head_list, list) {
+		mutex_lock(&h->lock);
+		if (!h->s)
+			goto unlock;
+
+		if (!ca)
+			goto found;
+
+		for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++) {
+			if (!h->s->blocks[i])
+				continue;
+
+			ob = c->open_buckets + h->s->blocks[i];
+			if (ob->dev == ca->dev_idx)
+				goto found;
+		}
+		goto unlock;
+found:
+		h->s->err = -BCH_ERR_erofs_no_writes;
+		ec_stripe_set_pending(c, h);
+unlock:
+		mutex_unlock(&h->lock);
 	}
+	mutex_unlock(&c->ec_stripe_head_lock);
+}
 
-	for_each_journal_key(*journal_keys, i)
-		if (i->btree_id == BTREE_ID_EC)
-			bch2_mark_key(c, bkey_i_to_s_c(i->k),
-				      0, NULL, 0,
-				      BCH_BUCKET_MARK_ALLOC_READ|
-				      BCH_BUCKET_MARK_NOATOMIC);
+void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
+{
+	__bch2_ec_stop(c, ca);
+}
 
-	return 0;
+void bch2_fs_ec_stop(struct bch_fs *c)
+{
+	__bch2_ec_stop(c, NULL);
 }
 
-int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
+static bool bch2_fs_ec_flush_done(struct bch_fs *c)
 {
-	struct btree_trans trans;
-	struct btree_iter *iter;
+	bool ret;
+
+	mutex_lock(&c->ec_stripe_new_lock);
+	ret = list_empty(&c->ec_stripe_new_list);
+	mutex_unlock(&c->ec_stripe_new_lock);
+
+	return ret;
+}
+
+void bch2_fs_ec_flush(struct bch_fs *c)
+{
+	wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c));
+}
+
+int bch2_stripes_read(struct bch_fs *c)
+{
+	struct btree_trans *trans = bch2_trans_get(c);
+	struct btree_iter iter;
 	struct bkey_s_c k;
-	size_t i, idx = 0;
-	int ret = 0;
+	const struct bch_stripe *s;
+	struct stripe *m;
+	unsigned i;
+	int ret;
+
+	for_each_btree_key(trans, iter, BTREE_ID_stripes, POS_MIN,
+			   BTREE_ITER_PREFETCH, k, ret) {
+		if (k.k->type != KEY_TYPE_stripe)
+			continue;
+
+		ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
+		if (ret)
+			break;
 
-	bch2_trans_init(&trans, c, 0, 0);
+		s = bkey_s_c_to_stripe(k).v;
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS(0, U64_MAX), 0);
+		m = genradix_ptr(&c->stripes, k.k->p.offset);
+		m->sectors	= le16_to_cpu(s->sectors);
+		m->algorithm	= s->algorithm;
+		m->nr_blocks	= s->nr_blocks;
+		m->nr_redundant	= s->nr_redundant;
+		m->blocks_nonempty = 0;
+
+		for (i = 0; i < s->nr_blocks; i++)
+			m->blocks_nonempty += !!stripe_blockcount_get(s, i);
+
+		bch2_stripes_heap_insert(c, m, k.k->p.offset);
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	bch2_trans_put(trans);
 
-	k = bch2_btree_iter_prev(iter);
-	if (!IS_ERR_OR_NULL(k.k))
-		idx = k.k->p.offset + 1;
-	ret = bch2_trans_exit(&trans);
 	if (ret)
-		return ret;
+		bch_err_fn(c, ret);
 
-	if (!gc &&
-	    !init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
-		       GFP_KERNEL))
-		return -ENOMEM;
-#if 0
-	ret = genradix_prealloc(&c->stripes[gc], idx, GFP_KERNEL);
-#else
-	for (i = 0; i < idx; i++)
-		if (!genradix_ptr_alloc(&c->stripes[gc], i, GFP_KERNEL))
-			return -ENOMEM;
-#endif
-	return 0;
+	return ret;
 }
 
-int bch2_fs_ec_start(struct bch_fs *c)
+void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
 {
-	return bch2_ec_mem_alloc(c, false);
+	ec_stripes_heap *h = &c->ec_stripes_heap;
+	struct stripe *m;
+	size_t i;
+
+	mutex_lock(&c->ec_stripes_heap_lock);
+	for (i = 0; i < min_t(size_t, h->used, 50); i++) {
+		m = genradix_ptr(&c->stripes, h->data[i].idx);
+
+		prt_printf(out, "%zu %u/%u+%u", h->data[i].idx,
+		       h->data[i].blocks_nonempty,
+		       m->nr_blocks - m->nr_redundant,
+		       m->nr_redundant);
+		if (bch2_stripe_is_open(c, h->data[i].idx))
+			prt_str(out, " open");
+		prt_newline(out);
+	}
+	mutex_unlock(&c->ec_stripes_heap_lock);
+}
+
+void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
+{
+	struct ec_stripe_head *h;
+	struct ec_stripe_new *s;
+
+	mutex_lock(&c->ec_stripe_head_lock);
+	list_for_each_entry(h, &c->ec_stripe_head_list, list) {
+		prt_printf(out, "target %u algo %u redundancy %u %s:\n",
+		       h->target, h->algo, h->redundancy,
+		       bch2_watermarks[h->watermark]);
+
+		if (h->s)
+			prt_printf(out, "\tidx %llu blocks %u+%u allocated %u\n",
+			       h->s->idx, h->s->nr_data, h->s->nr_parity,
+			       bitmap_weight(h->s->blocks_allocated,
+					     h->s->nr_data));
+	}
+	mutex_unlock(&c->ec_stripe_head_lock);
+
+	prt_printf(out, "in flight:\n");
+
+	mutex_lock(&c->ec_stripe_new_lock);
+	list_for_each_entry(s, &c->ec_stripe_new_list, list) {
+		prt_printf(out, "\tidx %llu blocks %u+%u ref %u %u %s\n",
+			   s->idx, s->nr_data, s->nr_parity,
+			   atomic_read(&s->ref[STRIPE_REF_io]),
+			   atomic_read(&s->ref[STRIPE_REF_stripe]),
+			   bch2_watermarks[s->h->watermark]);
+	}
+	mutex_unlock(&c->ec_stripe_new_lock);
 }
 
 void bch2_fs_ec_exit(struct bch_fs *c)
 {
 	struct ec_stripe_head *h;
+	unsigned i;
 
 	while (1) {
-		mutex_lock(&c->ec_new_stripe_lock);
-		h = list_first_entry_or_null(&c->ec_new_stripe_list,
+		mutex_lock(&c->ec_stripe_head_lock);
+		h = list_first_entry_or_null(&c->ec_stripe_head_list,
 					     struct ec_stripe_head, list);
 		if (h)
 			list_del(&h->list);
-		mutex_unlock(&c->ec_new_stripe_lock);
+		mutex_unlock(&c->ec_stripe_head_lock);
 		if (!h)
 			break;
 
-		BUG_ON(h->s);
-		BUG_ON(!list_empty(&h->stripes));
+		if (h->s) {
+			for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++)
+				BUG_ON(h->s->blocks[i]);
+
+			kfree(h->s);
+		}
 		kfree(h);
 	}
 
+	BUG_ON(!list_empty(&c->ec_stripe_new_list));
+
 	free_heap(&c->ec_stripes_heap);
-	genradix_free(&c->stripes[0]);
+	genradix_free(&c->stripes);
 	bioset_exit(&c->ec_bioset);
 }
 
-int bch2_fs_ec_init(struct bch_fs *c)
+void bch2_fs_ec_init_early(struct bch_fs *c)
 {
+	spin_lock_init(&c->ec_stripes_new_lock);
+	mutex_init(&c->ec_stripes_heap_lock);
+
+	INIT_LIST_HEAD(&c->ec_stripe_head_list);
+	mutex_init(&c->ec_stripe_head_lock);
+
+	INIT_LIST_HEAD(&c->ec_stripe_new_list);
+	mutex_init(&c->ec_stripe_new_lock);
+	init_waitqueue_head(&c->ec_stripe_new_wait);
+
+	INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
 	INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
+}
 
+int bch2_fs_ec_init(struct bch_fs *c)
+{
 	return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
 			   BIOSET_NEED_BVECS);
 }