]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - libbcachefs/ec.c
Update bcachefs sources to 2272c5f5b7 bcachefs: Mark stripe buckets with correct...
[bcachefs-tools-debian] / libbcachefs / ec.c
index 71d85c9347414e4f5e864a5a8c457b643a03f50f..a09b39c7e0a8569124ce26041a225c7d4984a62f 100644 (file)
@@ -4,10 +4,12 @@
 
 #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 "disk_groups.h"
 #include "ec.h"
@@ -102,49 +104,67 @@ 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,
+                       unsigned flags, struct printbuf *err)
 {
        const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 
-       if (!bkey_cmp(k.k->p, POS_MIN))
-               return "stripe at pos 0";
+       if (bkey_eq(k.k->p, POS_MIN)) {
+               prt_printf(err, "stripe at POS_MIN");
+               return -BCH_ERR_invalid_bkey;
+       }
 
-       if (k.k->p.inode)
-               return "invalid stripe key";
+       if (k.k->p.inode) {
+               prt_printf(err, "nonzero inode field");
+               return -BCH_ERR_invalid_bkey;
+       }
 
-       if (bkey_val_bytes(k.k) < sizeof(*s))
-               return "incorrect value size";
+       if (bkey_val_bytes(k.k) < sizeof(*s)) {
+               prt_printf(err, "incorrect value size (%zu < %zu)",
+                      bkey_val_bytes(k.k), sizeof(*s));
+               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));
+       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");
+       }
 }
 
 /* returns blocknr in stripe that we matched: */
-static int bkey_matches_stripe(struct bch_stripe *s,
-                              struct bkey_s_c k)
+static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
+                                               struct bkey_s_c k, unsigned *block)
 {
        struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
        const struct bch_extent_ptr *ptr;
@@ -153,10 +173,12 @@ static int bkey_matches_stripe(struct bch_stripe *s,
        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)))
-                               return i;
+                                                     le16_to_cpu(s->sectors))) {
+                               *block = i;
+                               return ptr;
+                       }
 
-       return -1;
+       return NULL;
 }
 
 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
@@ -284,14 +306,15 @@ static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
                        struct bch_csum got = ec_block_checksum(buf, i, offset);
 
                        if (bch2_crc_cmp(want, got)) {
-                               char buf2[200];
+                               struct printbuf buf2 = PRINTBUF;
 
-                               bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&buf->key.k_i));
+                               bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
 
                                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);
+                                       want.lo, got.lo, buf2.buf);
+                               printbuf_exit(&buf2);
                                clear_bit(i, buf->valid);
                                break;
                        }
@@ -399,7 +422,10 @@ static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
                                   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,
+                                                      rw,
+                                                      GFP_KERNEL,
                                                       &c->ec_bioset),
                                      struct ec_bio, bio);
 
@@ -407,9 +433,6 @@ 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;
@@ -427,15 +450,14 @@ 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(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
+static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
+                               struct ec_stripe_buf *stripe)
 {
-       struct btree_trans trans;
        struct btree_iter iter;
        struct bkey_s_c k;
        int ret;
 
-       bch2_trans_init(&trans, c, 0, 0);
-       bch2_trans_iter_init(&trans, &iter, BTREE_ID_stripes,
+       bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
                             POS(0, idx), BTREE_ITER_SLOTS);
        k = bch2_btree_iter_peek_slot(&iter);
        ret = bkey_err(k);
@@ -447,11 +469,15 @@ static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *strip
        }
        bkey_reassemble(&stripe->key.k_i, k);
 err:
-       bch2_trans_iter_exit(&trans, &iter);
-       bch2_trans_exit(&trans);
+       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)
 {
@@ -534,22 +560,22 @@ static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
                if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
                        return -ENOMEM;
 
-               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))
+       if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
                return -ENOMEM;
 
        if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
-           !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
+           !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
                return -ENOMEM;
 
        return 0;
@@ -559,26 +585,89 @@ static int ec_stripe_mem_alloc(struct btree_trans *trans,
                               struct btree_iter *iter)
 {
        size_t idx = iter->pos.offset;
-       int ret = 0;
 
        if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
-               return ret;
+               return 0;
 
        bch2_trans_unlock(trans);
-       ret = -EINTR;
 
-       if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL))
-               return ret;
+       return   __ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL) ?:
+               bch2_trans_relock(trans);
+}
 
-       return -ENOMEM;
+/*
+ * 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 ssize_t stripe_idx_to_delete(struct bch_fs *c)
+static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
+{
+       bool ret = false;
+
+       spin_lock(&c->ec_stripes_new_lock);
+       ret = __bch2_stripe_is_open(c, idx);
+       spin_unlock(&c->ec_stripes_new_lock);
+
+       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 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;
+}
+
+/* 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;
+       size_t heap_idx;
+
+       lockdep_assert_held(&c->ec_stripes_heap_lock);
+
+       for (heap_idx = 0; heap_idx < h->used; heap_idx++)
+               if (h->data[heap_idx].blocks_nonempty == 0 &&
+                   !bch2_stripe_is_open(c, h->data[heap_idx].idx))
+                       return h->data[heap_idx].idx;
 
-       return h->used && h->data[0].blocks_nonempty == 0
-               ? h->data[0].idx : -1;
+       return 0;
 }
 
 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
@@ -594,15 +683,14 @@ 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);
 }
@@ -610,28 +698,21 @@ static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 void bch2_stripes_heap_del(struct bch_fs *c,
                           struct stripe *m, size_t idx)
 {
-       if (!m->on_heap)
-               return;
-
-       m->on_heap = false;
-
+       mutex_lock(&c->ec_stripes_heap_lock);
        heap_verify_backpointer(c, idx);
 
        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)
 {
-       if (m->on_heap)
-               return;
-
+       mutex_lock(&c->ec_stripes_heap_lock);
        BUG_ON(heap_full(&c->ec_stripes_heap));
 
-       m->on_heap = true;
-
        heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
                        .idx = idx,
                        .blocks_nonempty = m->blocks_nonempty,
@@ -640,17 +721,17 @@ void bch2_stripes_heap_insert(struct bch_fs *c,
                 ec_stripes_heap_set_backpointer);
 
        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;
 
-       if (!m->on_heap)
-               return;
-
+       mutex_lock(&c->ec_stripes_heap_lock);
        heap_verify_backpointer(c, idx);
 
        h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
@@ -663,109 +744,101 @@ void bch2_stripes_heap_update(struct bch_fs *c,
 
        heap_verify_backpointer(c, idx);
 
-       if (stripe_idx_to_delete(c) >= 0 &&
-           !percpu_ref_is_dying(&c->writes))
-               schedule_work(&c->ec_stripe_delete_work);
+       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_stripes,
-                                      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;
 
-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;
+       bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, idx),
+                            BTREE_ITER_INTENT);
+       k = bch2_btree_iter_peek_slot(&iter);
+       ret = bkey_err(k);
+       if (ret)
+               goto err;
 
-       while (1) {
-               spin_lock(&c->ec_stripes_heap_lock);
-               idx = stripe_idx_to_delete(c);
-               if (idx < 0) {
-                       spin_unlock(&c->ec_stripes_heap_lock);
-                       break;
-               }
+       if (k.k->type != KEY_TYPE_stripe) {
+               bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
+               ret = -EINVAL;
+               goto err;
+       }
 
-               bch2_stripes_heap_del(c, genradix_ptr(&c->stripes[0], idx), idx);
-               spin_unlock(&c->ec_stripes_heap_lock);
+       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;
 
-               if (ec_stripe_delete(c, idx))
-                       break;
-       }
-}
+                       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;
+               }
 
-/* stripe creation: */
+       ret = bch2_btree_delete_at(trans, &iter, 0);
+err:
+       bch2_trans_iter_exit(trans, &iter);
+       return ret;
+}
 
-static int ec_stripe_bkey_insert(struct bch_fs *c,
-                                struct bkey_i_stripe *stripe,
-                                struct disk_reservation *res)
+static void ec_stripe_delete_work(struct work_struct *work)
 {
+       struct bch_fs *c =
+               container_of(work, struct bch_fs, ec_stripe_delete_work);
        struct btree_trans trans;
-       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;
+       u64 idx;
 
        bch2_trans_init(&trans, c, 0, 0);
-retry:
-       bch2_trans_begin(&trans);
 
-       for_each_btree_key(&trans, iter, BTREE_ID_stripes, start_pos,
-                          BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
-               if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
-                       if (start_pos.offset) {
-                               start_pos = min_pos;
-                               bch2_btree_iter_set_pos(&iter, start_pos);
-                               continue;
-                       }
+       while (1) {
+               mutex_lock(&c->ec_stripes_heap_lock);
+               idx = stripe_idx_to_delete(c);
+               mutex_unlock(&c->ec_stripes_heap_lock);
 
-                       ret = -ENOSPC;
+               if (!idx)
                        break;
-               }
 
-               if (bkey_deleted(k.k))
-                       goto found_slot;
+               ret = commit_do(&trans, NULL, NULL, BTREE_INSERT_NOFAIL,
+                               ec_stripe_delete(&trans, idx));
+               if (ret) {
+                       bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
+                       break;
+               }
        }
 
-       goto err;
-found_slot:
-       start_pos = iter.pos;
-
-       ret = ec_stripe_mem_alloc(&trans, &iter);
-       if (ret)
-               goto err;
-
-       stripe->k.p = iter.pos;
-
-       ret   = bch2_trans_update(&trans, &iter, &stripe->k_i, 0) ?:
-               bch2_trans_commit(&trans, res, NULL,
-                               BTREE_INSERT_NOFAIL);
-err:
-       bch2_trans_iter_exit(&trans, &iter);
-
-       if (ret == -EINTR)
-               goto retry;
-
-       c->ec_stripe_hint = ret ? start_pos.offset : start_pos.offset + 1;
        bch2_trans_exit(&trans);
 
-       return ret;
+       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) &&
+           !schedule_work(&c->ec_stripe_delete_work))
+               bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 }
 
-static int ec_stripe_bkey_update(struct btree_trans *trans,
-                                struct bkey_i_stripe *new)
+/* stripe creation: */
+
+static int ec_stripe_key_update(struct btree_trans *trans,
+                               struct bkey_i_stripe *new,
+                               bool create)
 {
+       struct bch_fs *c = trans->c;
        struct btree_iter iter;
        struct bkey_s_c k;
-       const struct bch_stripe *existing;
-       unsigned i;
        int ret;
 
        bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
@@ -775,23 +848,27 @@ static int ec_stripe_bkey_update(struct btree_trans *trans,
        if (ret)
                goto err;
 
-       if (!k.k || k.k->type != KEY_TYPE_stripe) {
-               bch_err(trans->c, "error updating stripe: not found");
-               ret = -ENOENT;
+       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;
        }
 
-       existing = bkey_s_c_to_stripe(k).v;
+       if (k.k->type == KEY_TYPE_stripe) {
+               const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
+               unsigned i;
 
-       if (existing->nr_blocks != new->v.nr_blocks) {
-               bch_err(trans->c, "error updating stripe: nr_blocks does not match");
-               ret = -EINVAL;
-               goto err;
-       }
+               if (old->nr_blocks != new->v.nr_blocks) {
+                       bch_err(c, "error updating stripe: nr_blocks does not match");
+                       ret = -EINVAL;
+                       goto err;
+               }
 
-       for (i = 0; i < new->v.nr_blocks; i++)
-               stripe_blockcount_set(&new->v, i,
-                       stripe_blockcount_get(existing, i));
+               for (i = 0; i < new->v.nr_blocks; i++)
+                       stripe_blockcount_set(&new->v, i, stripe_blockcount_get(old, i));
+       }
 
        ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 err:
@@ -799,96 +876,175 @@ err:
        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,
+                                  u64 *bp_offset)
 {
-       struct bch_extent_stripe_ptr *dst = (void *) ptr;
-       union bch_extent_entry *end = extent_entry_last(e);
+       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_offset, &bp, BTREE_ITER_CACHED);
+       if (ret)
+               return ret;
+       if (*bp_offset == U64_MAX)
+               return 0;
+
+       if (bp.level) {
+               struct printbuf buf = PRINTBUF;
+               struct btree_iter node_iter;
+               struct btree *b;
+
+               b = bch2_backpointer_get_node(trans, &node_iter, bucket, *bp_offset, bp);
+               bch2_trans_iter_exit(trans, &node_iter);
+
+               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;
+       }
+
+       k = bch2_backpointer_get_key(trans, &iter, bucket, *bp_offset, bp);
+       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;
+       }
+
+       if (extent_has_stripe_ptr(k, s->key.k.p.offset))
+               goto out;
+
+       ptr_c = bkey_matches_stripe(&s->key.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;
+
+       dev = s->key.v.ptrs[block].dev;
+
+       n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
+       ret = PTR_ERR_OR_ZERO(n);
+       if (ret)
+               goto out;
 
-       memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
-       e.k->u64s += sizeof(*dst) / sizeof(u64);
+       bkey_reassemble(n, k);
 
-       *dst = (struct bch_extent_stripe_ptr) {
+       bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
+       ec_ptr = (void *) bch2_bkey_has_device(bkey_i_to_s_c(n), dev);
+       BUG_ON(!ec_ptr);
+
+       stripe_ptr = (struct bch_extent_stripe_ptr) {
                .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
                .block          = block,
                .redundancy     = s->key.v.nr_redundant,
                .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_trans trans;
-       struct btree_iter iter;
-       struct bkey_s_c k;
-       struct bkey_s_extent e;
-       struct bkey_buf sk;
-       struct bpos next_pos;
-       int ret = 0, dev, block;
-
-       bch2_bkey_buf_init(&sk);
-       bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+       __extent_entry_insert(n,
+                       (union bch_extent_entry *) ec_ptr,
+                       (union bch_extent_entry *) &stripe_ptr);
 
-       /* XXX this doesn't support the reflink btree */
+       ret = bch2_trans_update(trans, &iter, n, 0);
+out:
+       bch2_trans_iter_exit(trans, &iter);
+       return ret;
+}
 
-       bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
-                            bkey_start_pos(pos),
-                            BTREE_ITER_INTENT);
-retry:
-       while (bch2_trans_begin(&trans),
-              (k = bch2_btree_iter_peek(&iter)).k &&
-              !(ret = bkey_err(k)) &&
-              bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
-               struct bch_extent_ptr *ptr, *ec_ptr = NULL;
-
-               if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
-                       bch2_btree_iter_advance(&iter);
-                       continue;
-               }
+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_extent_ptr bucket = s->key.v.ptrs[block];
+       struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
+       u64 bp_offset = 0;
+       int ret = 0;
 
-               block = bkey_matches_stripe(&s->key.v, k);
-               if (block < 0) {
-                       bch2_btree_iter_advance(&iter);
-                       continue;
-               }
+       while (1) {
+               ret = commit_do(trans, NULL, NULL,
+                               BTREE_INSERT_NOFAIL,
+                       ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
+                                               s, &bp_offset));
+               if (ret)
+                       break;
+               if (bp_offset == U64_MAX)
+                       break;
 
-               dev = s->key.v.ptrs[block].dev;
+               bp_offset++;
+       }
 
-               bch2_bkey_buf_reassemble(&sk, c, k);
-               e = bkey_i_to_s_extent(sk.k);
+       return ret;
+}
 
-               bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
-               ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
-               BUG_ON(!ec_ptr);
+static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
+{
+       struct btree_trans trans;
+       struct bch_stripe *v = &s->key.v;
+       unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
+       int ret = 0;
 
-               extent_stripe_ptr_add(e, s, ec_ptr, block);
+       bch2_trans_init(&trans, c, 0, 0);
 
-               bch2_btree_iter_set_pos(&iter, bkey_start_pos(&sk.k->k));
-               next_pos = sk.k->k.p;
+       ret = bch2_btree_write_buffer_flush(&trans);
+       if (ret)
+               goto err;
 
-               ret   = bch2_btree_iter_traverse(&iter) ?:
-                       bch2_trans_update(&trans, &iter, sk.k, 0) ?:
-                       bch2_trans_commit(&trans, NULL, NULL,
-                                       BTREE_INSERT_NOFAIL);
-               if (!ret)
-                       bch2_btree_iter_set_pos(&iter, next_pos);
+       for (i = 0; i < nr_data; i++) {
+               ret = ec_stripe_update_bucket(&trans, s, i);
                if (ret)
                        break;
        }
-       if (ret == -EINTR)
-               goto retry;
-       bch2_trans_iter_exit(&trans, &iter);
-
+err:
        bch2_trans_exit(&trans);
-       bch2_bkey_buf_exit(&sk, c);
 
        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 = -EROFS;
+               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;
+}
+
 /*
  * data buckets of new stripe all written: create the stripe
  */
@@ -896,8 +1052,6 @@ 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 stripe *m;
        struct bch_stripe *v = &s->new_stripe.key.v;
        unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
        int ret;
@@ -906,8 +1060,16 @@ static void ec_stripe_create(struct ec_stripe_new *s)
 
        closure_sync(&s->iodone);
 
+       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) {
-               if (s->err != -EROFS)
+               if (!bch2_err_matches(s->err, EROFS))
                        bch_err(c, "error creating stripe: error writing data buckets");
                goto err;
        }
@@ -930,9 +1092,6 @@ static void ec_stripe_create(struct ec_stripe_new *s)
 
        BUG_ON(!s->allocated);
 
-       if (!percpu_ref_tryget(&c->writes))
-               goto err;
-
        ec_generate_ec(&s->new_stripe);
 
        ec_generate_checksums(&s->new_stripe);
@@ -944,34 +1103,23 @@ static void ec_stripe_create(struct ec_stripe_new *s)
 
        if (ec_nr_failed(&s->new_stripe)) {
                bch_err(c, "error creating stripe: error writing redundancy buckets");
-               goto err_put_writes;
+               goto err;
        }
 
-       ret = s->have_existing_stripe
-               ? bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
-                               ec_stripe_bkey_update(&trans, &s->new_stripe.key))
-               : ec_stripe_bkey_insert(c, &s->new_stripe.key, &s->res);
+       ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
+                           ec_stripe_key_update(&trans, &s->new_stripe.key,
+                                                !s->have_existing_stripe));
        if (ret) {
                bch_err(c, "error creating stripe: error creating stripe key");
-               goto err_put_writes;
+               goto err;
        }
 
-       for_each_keylist_key(&s->keys, k) {
-               ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
-               if (ret) {
-                       bch_err(c, "error creating stripe: error %i updating pointers", ret);
-                       break;
-               }
+       ret = ec_stripe_update_extents(c, &s->new_stripe);
+       if (ret) {
+               bch_err(c, "error creating stripe: error updating pointers: %s",
+                       bch2_err_str(ret));
+               goto err;
        }
-
-       spin_lock(&c->ec_stripes_heap_lock);
-       m = genradix_ptr(&c->stripes[0], s->new_stripe.key.k.p.offset);
-
-       BUG_ON(m->on_heap);
-       bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
-       spin_unlock(&c->ec_stripes_heap_lock);
-err_put_writes:
-       percpu_ref_put(&c->writes);
 err:
        bch2_disk_reservation_put(c, &s->res);
 
@@ -987,7 +1135,7 @@ err:
                        }
                }
 
-       bch2_keylist_free(&s->keys, s->inline_keys);
+       bch2_stripe_close(c, s);
 
        ec_stripe_buf_exit(&s->existing_stripe);
        ec_stripe_buf_exit(&s->new_stripe);
@@ -995,31 +1143,49 @@ err:
        kfree(s);
 }
 
-static void ec_stripe_create_work(struct work_struct *work)
+static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
 {
-       struct bch_fs *c = container_of(work,
-               struct bch_fs, ec_stripe_create_work);
-       struct ec_stripe_new *s, *n;
-restart:
+       struct ec_stripe_new *s;
+
        mutex_lock(&c->ec_stripe_new_lock);
-       list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
+       list_for_each_entry(s, &c->ec_stripe_new_list, list)
                if (!atomic_read(&s->pin)) {
                        list_del(&s->list);
-                       mutex_unlock(&c->ec_stripe_new_lock);
-                       ec_stripe_create(s);
-                       goto restart;
+                       goto out;
                }
+       s = NULL;
+out:
        mutex_unlock(&c->ec_stripe_new_lock);
+
+       return s;
+}
+
+static void ec_stripe_create_work(struct work_struct *work)
+{
+       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);
+}
+
+void bch2_ec_do_stripe_creates(struct bch_fs *c)
+{
+       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_new_put(struct bch_fs *c, struct ec_stripe_new *s)
 {
        BUG_ON(atomic_read(&s->pin) <= 0);
 
-       if (atomic_dec_and_test(&s->pin)) {
-               BUG_ON(!s->pending);
-               queue_work(system_long_wq, &c->ec_stripe_create_work);
-       }
+       if (atomic_dec_and_test(&s->pin))
+               bch2_ec_do_stripe_creates(c);
 }
 
 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
@@ -1043,9 +1209,6 @@ 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(c, s);
 }
 
@@ -1065,36 +1228,12 @@ 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);
+       ca      = bch_dev_bkey_exists(c, ob->dev);
        offset  = ca->mi.bucket_size - ob->sectors_free;
 
        return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
 }
 
-void bch2_ob_add_backpointer(struct bch_fs *c, struct open_bucket *ob,
-                            struct bkey *k)
-{
-       struct ec_stripe_new *ec = ob->ec;
-
-       if (!ec)
-               return;
-
-       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       = k->p;
-       ec->keys.top->k.size    = k->size;
-       bch2_keylist_push(&ec->keys);
-
-       mutex_unlock(&ec->lock);
-}
-
 static int unsigned_cmp(const void *_l, const void *_r)
 {
        unsigned l = *((const unsigned *) _l);
@@ -1154,7 +1293,7 @@ static void ec_stripe_key_init(struct bch_fs *c,
        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->sb.encoded_extent_max);
+       s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
        s->v.csum_type                  = BCH_CSUM_crc32c;
        s->v.pad                        = 0;
 
@@ -1187,8 +1326,6 @@ static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
                                BCH_BKEY_PTRS_MAX) - h->redundancy;
        s->nr_parity    = h->redundancy;
 
-       bch2_keylist_init(&s->keys, s->inline_keys);
-
        ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
                           s->nr_parity, h->blocksize);
 
@@ -1210,7 +1347,7 @@ ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
                return NULL;
 
        mutex_init(&h->lock);
-       mutex_lock(&h->lock);
+       BUG_ON(!mutex_trylock(&h->lock));
 
        h->target       = target;
        h->algo         = algo;
@@ -1246,24 +1383,31 @@ void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
        mutex_unlock(&h->lock);
 }
 
-struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
+struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
                                                 unsigned target,
                                                 unsigned algo,
                                                 unsigned redundancy,
                                                 bool copygc)
 {
+       struct bch_fs *c = trans->c;
        struct ec_stripe_head *h;
+       int ret;
 
        if (!redundancy)
                return NULL;
 
-       mutex_lock(&c->ec_stripe_head_lock);
+       ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
+       if (ret)
+               return ERR_PTR(ret);
+
        list_for_each_entry(h, &c->ec_stripe_head_list, list)
                if (h->target           == target &&
                    h->algo             == algo &&
                    h->redundancy       == redundancy &&
                    h->copygc           == copygc) {
-                       mutex_lock(&h->lock);
+                       ret = bch2_trans_mutex_lock(trans, &h->lock);
+                       if (ret)
+                               h = ERR_PTR(ret);
                        goto found;
                }
 
@@ -1273,9 +1417,10 @@ found:
        return h;
 }
 
-static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
+static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
                                    struct closure *cl)
 {
+       struct bch_fs *c = trans->c;
        struct bch_devs_mask devs = h->devs;
        struct open_bucket *ob;
        struct open_buckets buckets;
@@ -1296,21 +1441,18 @@ static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
        BUG_ON(nr_have_data     > h->s->nr_data);
        BUG_ON(nr_have_parity   > h->s->nr_parity);
 
-       percpu_down_read(&c->mark_lock);
-       rcu_read_lock();
-
        buckets.nr = 0;
        if (nr_have_parity < h->s->nr_parity) {
-               ret = bch2_bucket_alloc_set(c, &buckets,
+               ret = bch2_bucket_alloc_set_trans(trans, &buckets,
                                            &h->parity_stripe,
                                            &devs,
                                            h->s->nr_parity,
                                            &nr_have_parity,
                                            &have_cache,
+                                           BCH_DATA_parity,
                                            h->copygc
-                                           ? RESERVE_MOVINGGC
-                                           : RESERVE_NONE,
-                                           0,
+                                           ? RESERVE_movinggc
+                                           : RESERVE_none,
                                            cl);
 
                open_bucket_for_each(c, &buckets, ob, i) {
@@ -1320,26 +1462,26 @@ static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
                        BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
 
                        h->s->blocks[j] = buckets.v[i];
-                       h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
+                       h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
                        __set_bit(j, h->s->blocks_gotten);
                }
 
                if (ret)
-                       goto err;
+                       return ret;
        }
 
        buckets.nr = 0;
        if (nr_have_data < h->s->nr_data) {
-               ret = bch2_bucket_alloc_set(c, &buckets,
+               ret = bch2_bucket_alloc_set_trans(trans, &buckets,
                                            &h->block_stripe,
                                            &devs,
                                            h->s->nr_data,
                                            &nr_have_data,
                                            &have_cache,
+                                           BCH_DATA_user,
                                            h->copygc
-                                           ? RESERVE_MOVINGGC
-                                           : RESERVE_NONE,
-                                           0,
+                                           ? RESERVE_movinggc
+                                           : RESERVE_none,
                                            cl);
 
                open_bucket_for_each(c, &buckets, ob, i) {
@@ -1348,17 +1490,15 @@ static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
                        BUG_ON(j >= h->s->nr_data);
 
                        h->s->blocks[j] = buckets.v[i];
-                       h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
+                       h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
                        __set_bit(j, h->s->blocks_gotten);
                }
 
                if (ret)
-                       goto err;
+                       return ret;
        }
-err:
-       rcu_read_unlock();
-       percpu_up_read(&c->mark_lock);
-       return ret;
+
+       return 0;
 }
 
 /* XXX: doesn't obey target: */
@@ -1374,43 +1514,42 @@ static s64 get_existing_stripe(struct bch_fs *c,
        if (may_create_new_stripe(c))
                return -1;
 
-       spin_lock(&c->ec_stripes_heap_lock);
+       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[0], stripe_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_stripes_heap_del(c, m, stripe_idx);
+                   m->blocks_nonempty  < m->nr_blocks - m->nr_redundant &&
+                   bch2_try_open_stripe(c, head->s, stripe_idx)) {
                        ret = stripe_idx;
                        break;
                }
        }
-       spin_unlock(&c->ec_stripes_heap_lock);
+       mutex_unlock(&c->ec_stripes_heap_lock);
        return ret;
 }
 
-static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
-                                                  struct ec_stripe_head *h)
+static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
 {
+       struct bch_fs *c = trans->c;
        unsigned i;
        s64 idx;
        int ret;
 
        idx = get_existing_stripe(c, h);
-       if (idx < 0) {
-               bch_err(c, "failed to find an existing stripe");
-               return -ENOSPC;
-       }
+       if (idx < 0)
+               return -BCH_ERR_ENOSPC_stripe_reuse;
 
        h->s->have_existing_stripe = true;
-       ret = get_stripe_key(c, idx, &h->s->existing_stripe);
+       ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
        if (ret) {
                bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
                return ret;
@@ -1437,47 +1576,83 @@ static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
        }
 
        bkey_copy(&h->s->new_stripe.key.k_i,
-                       &h->s->existing_stripe.key.k_i);
+                 &h->s->existing_stripe.key.k_i);
 
        return 0;
 }
 
-static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
-                                                       struct ec_stripe_head *h)
+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;
 
+       BUG_ON(h->s->res.sectors);
+
        ret = bch2_disk_reservation_get(c, &h->s->res,
-                       h->blocksize,
-                       h->s->nr_parity, 0);
+                                       h->blocksize,
+                                       h->s->nr_parity, 0);
+       if (ret)
+               return ret;
+
+       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 = -BCH_ERR_ENOSPC_stripe_create;
+                       break;
+               }
 
+               if (bkey_deleted(k.k) &&
+                   bch2_try_open_stripe(c, h->s, k.k->p.offset))
+                       break;
+       }
+
+       c->ec_stripe_hint = iter.pos.offset;
+
+       if (ret)
+               goto err;
+
+       ret = ec_stripe_mem_alloc(trans, &iter);
        if (ret) {
-               /*
-                * This means we need to wait for copygc to
-                * empty out buckets from existing stripes:
-                */
-               bch_err(c, "failed to reserve stripe");
+               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;
 }
 
-struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
+struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
                                               unsigned target,
                                               unsigned algo,
                                               unsigned redundancy,
                                               bool copygc,
                                               struct closure *cl)
 {
+       struct bch_fs *c = trans->c;
        struct ec_stripe_head *h;
        int ret;
        bool needs_stripe_new;
 
-       h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
-       if (!h) {
+       h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, copygc);
+       if (!h)
                bch_err(c, "no stripe head");
-               return NULL;
-       }
+       if (IS_ERR_OR_NULL(h))
+               return h;
 
        needs_stripe_new = !h->s;
        if (needs_stripe_new) {
@@ -1498,22 +1673,27 @@ struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
         */
        ret = 0;
        if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
-               ret = __bch2_ec_stripe_head_reserve(c, h);
+               ret = __bch2_ec_stripe_head_reserve(trans, h);
+       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+               goto err;
+
        if (ret && needs_stripe_new)
-               ret = __bch2_ec_stripe_head_reuse(c, h);
-       if (ret)
+               ret = __bch2_ec_stripe_head_reuse(trans, h);
+       if (ret) {
+               bch_err_ratelimited(c, "failed to get stripe: %s", bch2_err_str(ret));
                goto err;
+       }
 
        if (!h->s->allocated) {
-               ret = new_stripe_alloc_buckets(c, h, cl);
+               ret = new_stripe_alloc_buckets(trans, h, cl);
                if (ret)
                        goto err;
 
                h->s->allocated = true;
        }
 
+       BUG_ON(trans->restarted);
        return h;
-
 err:
        bch2_ec_stripe_head_put(c, h);
        return ERR_PTR(ret);
@@ -1537,7 +1717,7 @@ void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
                                continue;
 
                        ob = c->open_buckets + h->s->blocks[i];
-                       if (ob->ptr.dev == ca->dev_idx)
+                       if (ob->dev == ca->dev_idx)
                                goto found;
                }
                goto unlock;
@@ -1550,200 +1730,67 @@ unlock:
        mutex_unlock(&c->ec_stripe_head_lock);
 }
 
-void bch2_stripes_heap_start(struct bch_fs *c)
-{
-       struct genradix_iter iter;
-       struct stripe *m;
-
-       genradix_for_each(&c->stripes[0], iter, m)
-               if (m->alive)
-                       bch2_stripes_heap_insert(c, m, iter.pos);
-}
-
-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)
-{
-       const struct bch_stripe *v;
-       struct bkey_s_c k;
-       unsigned i;
-       int ret;
-
-       bch2_btree_iter_set_pos(iter, POS(0, idx));
-
-       k = bch2_btree_iter_peek_slot(iter);
-       ret = bkey_err(k);
-       if (ret)
-               return ret;
-
-       if (k.k->type != KEY_TYPE_stripe)
-               return -EIO;
-
-       v = bkey_s_c_to_stripe(k).v;
-       for (i = 0; i < v->nr_blocks; i++)
-               if (m->block_sectors[i] != stripe_blockcount_get(v, i))
-                       goto write;
-       return 0;
-write:
-       bkey_reassemble(&new_key->k_i, k);
-
-       for (i = 0; i < new_key->v.nr_blocks; i++)
-               stripe_blockcount_set(&new_key->v, i,
-                                     m->block_sectors[i]);
-
-       return bch2_trans_update(trans, iter, &new_key->k_i, 0);
-}
-
-int bch2_stripes_write(struct bch_fs *c, unsigned flags)
+int bch2_stripes_read(struct bch_fs *c)
 {
        struct btree_trans trans;
        struct btree_iter iter;
-       struct genradix_iter giter;
-       struct bkey_i_stripe *new_key;
+       struct bkey_s_c k;
+       const struct bch_stripe *s;
        struct stripe *m;
-       int ret = 0;
-
-       new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
-       BUG_ON(!new_key);
+       unsigned i;
+       int ret;
 
        bch2_trans_init(&trans, c, 0, 0);
 
-       bch2_trans_iter_init(&trans, &iter, BTREE_ID_stripes, POS_MIN,
-                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
-
-       genradix_for_each(&c->stripes[0], giter, m) {
-               if (!m->alive)
+       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 = __bch2_trans_do(&trans, NULL, NULL,
-                                     BTREE_INSERT_NOFAIL|flags,
-                       __bch2_stripe_write_key(&trans, &iter, m,
-                                       giter.pos, new_key));
-
+               ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
                if (ret)
                        break;
-       }
-       bch2_trans_iter_exit(&trans, &iter);
-
-       bch2_trans_exit(&trans);
-
-       kfree(new_key);
-
-       return ret;
-}
-
-static int bch2_stripes_read_fn(struct btree_trans *trans, struct bkey_s_c k)
-{
-       const struct bch_stripe *s;
-       struct bch_fs *c = trans->c;
-       struct stripe *m;
-       unsigned i;
-       int ret = 0;
 
-       if (k.k->type != KEY_TYPE_stripe)
-               return 0;
-
-       ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
-       if (ret)
-               return ret;
+               s = bkey_s_c_to_stripe(k).v;
 
-       s = bkey_s_c_to_stripe(k).v;
+               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;
 
-       m = genradix_ptr(&c->stripes[0], k.k->p.offset);
-       m->alive        = true;
-       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);
 
-       for (i = 0; i < s->nr_blocks; i++) {
-               m->block_sectors[i] =
-                       stripe_blockcount_get(s, i);
-               m->blocks_nonempty += !!m->block_sectors[i];
-               m->ptrs[i] = s->ptrs[i];
+               bch2_stripes_heap_insert(c, m, k.k->p.offset);
        }
+       bch2_trans_iter_exit(&trans, &iter);
 
-       bch2_bkey_to_replicas(&m->r.e, k);
-
-       spin_lock(&c->ec_stripes_heap_lock);
-       bch2_stripes_heap_update(c, m, k.k->p.offset);
-       spin_unlock(&c->ec_stripes_heap_lock);
-
-       return ret;
-}
-
-int bch2_stripes_read(struct bch_fs *c)
-{
-       struct btree_trans trans;
-       int ret;
-
-       bch2_trans_init(&trans, c, 0, 0);
-       ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_stripes,
-                                         bch2_stripes_read_fn);
        bch2_trans_exit(&trans);
+
        if (ret)
                bch_err(c, "error reading stripes: %i", ret);
 
        return ret;
 }
 
-int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
-{
-       struct btree_trans trans;
-       struct btree_iter iter;
-       struct bkey_s_c k;
-       size_t i, idx = 0;
-       int ret = 0;
-
-       bch2_trans_init(&trans, c, 0, 0);
-       bch2_trans_iter_init(&trans, &iter, BTREE_ID_stripes, POS(0, U64_MAX), 0);
-
-       k = bch2_btree_iter_prev(&iter);
-       ret = bkey_err(k);
-       if (!ret && k.k)
-               idx = k.k->p.offset + 1;
-
-       bch2_trans_iter_exit(&trans, &iter);
-       bch2_trans_exit(&trans);
-       if (ret)
-               return ret;
-
-       if (!idx)
-               return 0;
-
-       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;
-}
-
 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
 {
        ec_stripes_heap *h = &c->ec_stripes_heap;
        struct stripe *m;
        size_t i;
 
-       spin_lock(&c->ec_stripes_heap_lock);
+       mutex_lock(&c->ec_stripes_heap_lock);
        for (i = 0; i < min_t(size_t, h->used, 20); i++) {
-               m = genradix_ptr(&c->stripes[0], h->data[i].idx);
+               m = genradix_ptr(&c->stripes, h->data[i].idx);
 
-               pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
+               prt_printf(out, "%zu %u/%u+%u\n", h->data[i].idx,
                       h->data[i].blocks_nonempty,
                       m->nr_blocks - m->nr_redundant,
                       m->nr_redundant);
        }
-       spin_unlock(&c->ec_stripes_heap_lock);
+       mutex_unlock(&c->ec_stripes_heap_lock);
 }
 
 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
@@ -1753,11 +1800,11 @@ void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
 
        mutex_lock(&c->ec_stripe_head_lock);
        list_for_each_entry(h, &c->ec_stripe_head_list, list) {
-               pr_buf(out, "target %u algo %u redundancy %u:\n",
+               prt_printf(out, "target %u algo %u redundancy %u:\n",
                       h->target, h->algo, h->redundancy);
 
                if (h->s)
-                       pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
+                       prt_printf(out, "\tpending: blocks %u+%u allocated %u\n",
                               h->s->nr_data, h->s->nr_parity,
                               bitmap_weight(h->s->blocks_allocated,
                                             h->s->nr_data));
@@ -1766,7 +1813,7 @@ void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
 
        mutex_lock(&c->ec_stripe_new_lock);
        list_for_each_entry(s, &c->ec_stripe_new_list, list) {
-               pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
+               prt_printf(out, "\tin flight: blocks %u+%u pin %u\n",
                       s->nr_data, s->nr_parity,
                       atomic_read(&s->pin));
        }
@@ -1776,6 +1823,7 @@ void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
 void bch2_fs_ec_exit(struct bch_fs *c)
 {
        struct ec_stripe_head *h;
+       unsigned i;
 
        while (1) {
                mutex_lock(&c->ec_stripe_head_lock);
@@ -1787,21 +1835,31 @@ void bch2_fs_ec_exit(struct bch_fs *c)
                if (!h)
                        break;
 
-               BUG_ON(h->s);
+               if (h->s) {
+                       for (i = 0; i < 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)
 {
        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)
+{
+       spin_lock_init(&c->ec_stripes_new_lock);
 
        return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
                           BIOSET_NEED_BVECS);