]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - libbcachefs/ec.c
Update bcachefs sources to 2115a2ffde bcachefs: Kill bch2_verify_bucket_evacuated()
[bcachefs-tools-debian] / libbcachefs / ec.c
index 7bd68880995f28bebb9ac3efc1391b87920ae53f..1855d08efd4be2b190398a57250b091ec3afa8a5 100644 (file)
@@ -213,8 +213,9 @@ static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
        }
 }
 
+/* XXX: this is a non-mempoolified memory allocation: */
 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
-                              unsigned offset, unsigned size)
+                             unsigned offset, unsigned size)
 {
        struct bch_stripe *v = &buf->key.v;
        unsigned csum_granularity = 1U << v->csum_granularity_bits;
@@ -241,7 +242,7 @@ static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
        return 0;
 err:
        ec_stripe_buf_exit(buf);
-       return -ENOMEM;
+       return -BCH_ERR_ENOMEM_stripe_buf;
 }
 
 /* Checksumming: */
@@ -493,7 +494,7 @@ int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 
        buf = kzalloc(sizeof(*buf), GFP_NOIO);
        if (!buf)
-               return -ENOMEM;
+               return -BCH_ERR_ENOMEM_ec_read_extent;
 
        ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
        if (ret) {
@@ -558,7 +559,7 @@ 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;
 
                mutex_lock(&c->ec_stripes_heap_lock);
                if (n.size > h->size) {
@@ -572,11 +573,11 @@ static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
        }
 
        if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
-               return -ENOMEM;
+               return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 
        if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
            !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
-               return -ENOMEM;
+               return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 
        return 0;
 }
@@ -647,7 +648,7 @@ 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(&s->hash);
+       hlist_del_init(&s->hash);
        spin_unlock(&c->ec_stripes_new_lock);
 
        s->idx = 0;
@@ -658,14 +659,13 @@ static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
 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;
+       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;
 }
@@ -826,7 +826,7 @@ static void ec_stripe_delete_work(struct work_struct *work)
 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))
+           !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work))
                bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 }
 
@@ -866,8 +866,16 @@ static int ec_stripe_key_update(struct btree_trans *trans,
                        goto err;
                }
 
-               for (i = 0; i < new->v.nr_blocks; i++)
-                       stripe_blockcount_set(&new->v, i, stripe_blockcount_get(old, i));
+               for (i = 0; i < new->v.nr_blocks; i++) {
+                       unsigned v = stripe_blockcount_get(old, i);
+
+                       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);
@@ -879,7 +887,7 @@ err:
 static int ec_stripe_update_extent(struct btree_trans *trans,
                                   struct bpos bucket, u8 gen,
                                   struct ec_stripe_buf *s,
-                                  u64 *bp_offset)
+                                  struct bpos *bp_pos)
 {
        struct bch_fs *c = trans->c;
        struct bch_backpointer bp;
@@ -892,16 +900,32 @@ static int ec_stripe_update_extent(struct btree_trans *trans,
        int ret, dev, block;
 
        ret = bch2_get_next_backpointer(trans, bucket, gen,
-                               bp_offset, &bp, BTREE_ITER_CACHED);
+                               bp_pos, &bp, BTREE_ITER_CACHED);
        if (ret)
                return ret;
-       if (*bp_offset == U64_MAX)
+       if (bpos_eq(*bp_pos, SPOS_MAX))
                return 0;
 
-       if (bch2_fs_inconsistent_on(bp.level, c, "found btree node in erasure coded bucket!?"))
+       if (bp.level) {
+               struct printbuf buf = PRINTBUF;
+               struct btree_iter node_iter;
+               struct btree *b;
+
+               b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
+               bch2_trans_iter_exit(trans, &node_iter);
+
+               if (!b)
+                       return 0;
+
+               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);
+       k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT);
        ret = bkey_err(k);
        if (ret)
                return ret;
@@ -934,7 +958,7 @@ static int ec_stripe_update_extent(struct btree_trans *trans,
        bkey_reassemble(n, k);
 
        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);
+       ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev);
        BUG_ON(!ec_ptr);
 
        stripe_ptr = (struct bch_extent_stripe_ptr) {
@@ -960,20 +984,21 @@ static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_b
        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;
+       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_offset));
+                                               s, &bp_pos));
                if (ret)
                        break;
-               if (bp_offset == U64_MAX)
+               if (bkey_eq(bp_pos, POS_MAX))
                        break;
 
-               bp_offset++;
+               bp_pos = bpos_nosnap_successor(bp_pos);
        }
 
        return ret;
@@ -1032,6 +1057,13 @@ static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
                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
  */
@@ -1047,13 +1079,15 @@ 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 (!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 (ob->sectors_free)
+                                       zero_out_rest_of_ec_bucket(c, s, i, ob);
+                       }
+       }
 
        if (s->err) {
                if (!bch2_err_matches(s->err, EROFS))
@@ -1078,6 +1112,7 @@ static void ec_stripe_create(struct ec_stripe_new *s)
        }
 
        BUG_ON(!s->allocated);
+       BUG_ON(!s->idx);
 
        ec_generate_ec(&s->new_stripe);
 
@@ -1093,7 +1128,9 @@ static void ec_stripe_create(struct ec_stripe_new *s)
                goto err;
        }
 
-       ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
+       ret = bch2_trans_do(c, &s->res, NULL,
+                           BTREE_INSERT_NOCHECK_RW|
+                           BTREE_INSERT_NOFAIL,
                            ec_stripe_key_update(&trans, &s->new_stripe.key,
                                                 !s->have_existing_stripe));
        if (ret) {
@@ -1107,8 +1144,6 @@ static void ec_stripe_create(struct ec_stripe_new *s)
                        bch2_err_str(ret));
                goto err;
        }
-
-       bch2_stripe_close(c, s);
 err:
        bch2_disk_reservation_put(c, &s->res);
 
@@ -1124,10 +1159,15 @@ err:
                        }
                }
 
+       mutex_lock(&c->ec_stripe_new_lock);
+       list_del(&s->list);
+       mutex_unlock(&c->ec_stripe_new_lock);
+
        ec_stripe_buf_exit(&s->existing_stripe);
        ec_stripe_buf_exit(&s->new_stripe);
        closure_debug_destroy(&s->iodone);
-       kfree(s);
+
+       ec_stripe_new_put(c, s, STRIPE_REF_stripe);
 }
 
 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
@@ -1136,10 +1176,8 @@ static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
 
        mutex_lock(&c->ec_stripe_new_lock);
        list_for_each_entry(s, &c->ec_stripe_new_list, list)
-               if (!atomic_read(&s->pin)) {
-                       list_del(&s->list);
+               if (!atomic_read(&s->ref[STRIPE_REF_io]))
                        goto out;
-               }
        s = NULL;
 out:
        mutex_unlock(&c->ec_stripe_new_lock);
@@ -1167,14 +1205,6 @@ void bch2_ec_do_stripe_creates(struct bch_fs *c)
                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))
-               bch2_ec_do_stripe_creates(c);
-}
-
 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
 {
        struct ec_stripe_new *s = h->s;
@@ -1188,15 +1218,7 @@ static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
        list_add(&s->list, &c->ec_stripe_new_list);
        mutex_unlock(&c->ec_stripe_new_lock);
 
-       ec_stripe_new_put(c, s);
-}
-
-/* have a full bucket - hand it off to be erasure coded: */
-void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
-{
-       struct ec_stripe_new *s = ob->ec;
-
-       ec_stripe_new_put(c, s);
+       ec_stripe_new_put(c, s, STRIPE_REF_io);
 }
 
 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
@@ -1215,6 +1237,8 @@ void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
        if (!ob)
                return NULL;
 
+       BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
+
        ca      = bch_dev_bkey_exists(c, ob->dev);
        offset  = ca->mi.bucket_size - ob->sectors_free;
 
@@ -1302,11 +1326,12 @@ static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 
        s = kzalloc(sizeof(*s), GFP_KERNEL);
        if (!s)
-               return -ENOMEM;
+               return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
 
        mutex_init(&s->lock);
        closure_init(&s->iodone, NULL);
-       atomic_set(&s->pin, 1);
+       atomic_set(&s->ref[STRIPE_REF_stripe], 1);
+       atomic_set(&s->ref[STRIPE_REF_io], 1);
        s->c            = c;
        s->h            = h;
        s->nr_data      = min_t(unsigned, h->nr_active_devs,
@@ -1323,7 +1348,7 @@ static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 static struct ec_stripe_head *
 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
                         unsigned algo, unsigned redundancy,
-                        bool copygc)
+                        enum alloc_reserve reserve)
 {
        struct ec_stripe_head *h;
        struct bch_dev *ca;
@@ -1339,7 +1364,7 @@ ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
        h->target       = target;
        h->algo         = algo;
        h->redundancy   = redundancy;
-       h->copygc       = copygc;
+       h->reserve      = reserve;
 
        rcu_read_lock();
        h->devs = target_rw_devs(c, BCH_DATA_user, target);
@@ -1374,7 +1399,7 @@ struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
                                                 unsigned target,
                                                 unsigned algo,
                                                 unsigned redundancy,
-                                                bool copygc)
+                                                enum alloc_reserve reserve)
 {
        struct bch_fs *c = trans->c;
        struct ec_stripe_head *h;
@@ -1387,25 +1412,30 @@ struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
        if (ret)
                return ERR_PTR(ret);
 
+       if (test_bit(BCH_FS_GOING_RO, &c->flags)) {
+               h = ERR_PTR(-EROFS);
+               goto found;
+       }
+
        list_for_each_entry(h, &c->ec_stripe_head_list, list)
                if (h->target           == target &&
                    h->algo             == algo &&
                    h->redundancy       == redundancy &&
-                   h->copygc           == copygc) {
+                   h->reserve          == reserve) {
                        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, copygc);
+       h = ec_new_stripe_head_alloc(c, target, algo, redundancy, reserve);
 found:
        mutex_unlock(&c->ec_stripe_head_lock);
        return h;
 }
 
 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
-                                   struct closure *cl)
+                                   enum alloc_reserve reserve, struct closure *cl)
 {
        struct bch_fs *c = trans->c;
        struct bch_devs_mask devs = h->devs;
@@ -1415,14 +1445,15 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
        bool have_cache = true;
        int ret = 0;
 
-       for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
-               if (test_bit(i, h->s->blocks_gotten)) {
-                       __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
-                       if (i < h->s->nr_data)
-                               nr_have_data++;
-                       else
-                               nr_have_parity++;
-               }
+       BUG_ON(h->s->new_stripe.key.v.nr_blocks         != h->s->nr_data + h->s->nr_parity);
+       BUG_ON(h->s->new_stripe.key.v.nr_redundant      != h->s->nr_parity);
+
+       for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
+               __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
+               if (i < h->s->nr_data)
+                       nr_have_data++;
+               else
+                       nr_have_parity++;
        }
 
        BUG_ON(nr_have_data     > h->s->nr_data);
@@ -1435,11 +1466,9 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
                                            &devs,
                                            h->s->nr_parity,
                                            &nr_have_parity,
-                                           &have_cache,
+                                           &have_cache, 0,
                                            BCH_DATA_parity,
-                                           h->copygc
-                                           ? RESERVE_movinggc
-                                           : RESERVE_none,
+                                           reserve,
                                            cl);
 
                open_bucket_for_each(c, &buckets, ob, i) {
@@ -1464,11 +1493,9 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
                                            &devs,
                                            h->s->nr_data,
                                            &nr_have_data,
-                                           &have_cache,
+                                           &have_cache, 0,
                                            BCH_DATA_user,
-                                           h->copygc
-                                           ? RESERVE_movinggc
-                                           : RESERVE_none,
+                                           reserve,
                                            cl);
 
                open_bucket_for_each(c, &buckets, ob, i) {
@@ -1531,28 +1558,46 @@ static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stri
        s64 idx;
        int ret;
 
+       /*
+        * 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_ENOSPC_stripe_reuse;
+               return -BCH_ERR_stripe_alloc_blocked;
 
-       h->s->have_existing_stripe = true;
        ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
        if (ret) {
-               bch2_fs_fatal_error(c, "error reading stripe key: %i", 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 (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
-               /*
-                * this is a problem: we have deleted from the
-                * stripes heap already
-                */
-               BUG();
+       BUG_ON(h->s->existing_stripe.key.v.nr_redundant != h->s->nr_parity);
+       h->s->nr_data = h->s->existing_stripe.key.v.nr_blocks -
+               h->s->existing_stripe.key.v.nr_redundant;
+
+       ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
+       if (ret) {
+               bch2_stripe_close(c, h->s);
+               return ret;
        }
 
        BUG_ON(h->s->existing_stripe.size != h->blocksize);
        BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
 
+       /*
+        * 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, h->s->new_stripe.key.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));
+
        for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
                if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
                        __set_bit(i, h->s->blocks_gotten);
@@ -1562,8 +1607,8 @@ static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stri
                ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
        }
 
-       bkey_copy(&h->s->new_stripe.key.k_i,
-                 &h->s->existing_stripe.key.k_i);
+       bkey_copy(&h->s->new_stripe.key.k_i, &h->s->existing_stripe.key.k_i);
+       h->s->have_existing_stripe = true;
 
        return 0;
 }
@@ -1577,13 +1622,14 @@ static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_st
        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,
+       if (!h->s->res.sectors) {
+               ret = bch2_disk_reservation_get(c, &h->s->res,
                                        h->blocksize,
-                                       h->s->nr_parity, 0);
-       if (ret)
-               return ret;
+                                       h->s->nr_parity,
+                                       BCH_DISK_RESERVATION_NOFAIL);
+               if (ret)
+                       return ret;
+       }
 
        for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
                           BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
@@ -1627,58 +1673,87 @@ struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
                                               unsigned target,
                                               unsigned algo,
                                               unsigned redundancy,
-                                              bool copygc,
+                                              enum alloc_reserve reserve,
                                               struct closure *cl)
 {
        struct bch_fs *c = trans->c;
        struct ec_stripe_head *h;
+       bool waiting = false;
        int ret;
-       bool needs_stripe_new;
 
-       h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, copygc);
+       h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, reserve);
        if (!h)
                bch_err(c, "no stripe head");
        if (IS_ERR_OR_NULL(h))
                return h;
 
-       needs_stripe_new = !h->s;
-       if (needs_stripe_new) {
-               if (ec_new_stripe_alloc(c, h)) {
-                       ret = -ENOMEM;
+       if (!h->s) {
+               ret = ec_new_stripe_alloc(c, h);
+               if (ret) {
                        bch_err(c, "failed to allocate new stripe");
                        goto err;
                }
-
-               if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
-                       BUG();
        }
 
-       /*
-        * Try reserve a new stripe before reusing an
-        * existing stripe. This will prevent unnecessary
-        * read amplification during write oriented workloads.
-        */
-       ret = 0;
-       if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
-               ret = __bch2_ec_stripe_head_reserve(trans, h);
-       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
-               goto err;
+       if (h->s->allocated)
+               goto allocated;
 
-       if (ret && needs_stripe_new)
-               ret = __bch2_ec_stripe_head_reuse(trans, h);
-       if (ret) {
-               bch_err_ratelimited(c, "failed to get stripe: %s", bch2_err_str(ret));
+       if (h->s->have_existing_stripe)
+               goto alloc_existing;
+
+       /* First, try to allocate a full stripe: */
+       ret =   new_stripe_alloc_buckets(trans, h, RESERVE_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;
-       }
 
-       if (!h->s->allocated) {
-               ret = new_stripe_alloc_buckets(trans, h, cl);
-               if (ret)
+       /*
+        * 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;
 
-               h->s->allocated = true;
+               if (reserve == RESERVE_movinggc) {
+                       ret =   new_stripe_alloc_buckets(trans, h, reserve, 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 reserve watermark for this
+        * particular write:
+        */
+       ret = new_stripe_alloc_buckets(trans, h, reserve, cl);
+       if (ret)
+               goto err;
+
+allocate_buf:
+       ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
+       if (ret)
+               goto err;
+
+       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:
@@ -1686,7 +1761,7 @@ err:
        return ERR_PTR(ret);
 }
 
-void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
+static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
 {
        struct ec_stripe_head *h;
        struct open_bucket *ob;
@@ -1694,11 +1769,13 @@ void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
 
        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 < h->s->new_stripe.key.v.nr_blocks; i++) {
                        if (!h->s->blocks[i])
                                continue;
@@ -1717,6 +1794,32 @@ unlock:
        mutex_unlock(&c->ec_stripe_head_lock);
 }
 
+void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
+{
+       __bch2_ec_stop(c, ca);
+}
+
+void bch2_fs_ec_stop(struct bch_fs *c)
+{
+       __bch2_ec_stop(c, NULL);
+}
+
+static bool bch2_fs_ec_flush_done(struct bch_fs *c)
+{
+       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;
@@ -1769,13 +1872,16 @@ void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
        size_t i;
 
        mutex_lock(&c->ec_stripes_heap_lock);
-       for (i = 0; i < min_t(size_t, h->used, 20); i++) {
+       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\n", 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);
 }
@@ -1787,22 +1893,27 @@ 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) {
-               prt_printf(out, "target %u algo %u redundancy %u:\n",
-                      h->target, h->algo, h->redundancy);
+               prt_printf(out, "target %u algo %u redundancy %u %s:\n",
+                      h->target, h->algo, h->redundancy,
+                      bch2_alloc_reserves[h->reserve]);
 
                if (h->s)
-                       prt_printf(out, "\tpending: blocks %u+%u allocated %u\n",
-                              h->s->nr_data, h->s->nr_parity,
+                       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, "\tin flight: blocks %u+%u pin %u\n",
-                      s->nr_data, s->nr_parity,
-                      atomic_read(&s->pin));
+               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_alloc_reserves[s->h->reserve]);
        }
        mutex_unlock(&c->ec_stripe_new_lock);
 }
@@ -1840,14 +1951,22 @@ void bch2_fs_ec_exit(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)
 {
-       spin_lock_init(&c->ec_stripes_new_lock);
-
        return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
                           BIOSET_NEED_BVECS);
 }