if (unlikely(trans->journal_transaction_names))
                        journal_transaction_name(trans);
-       } else {
-               trans->journal_res.seq = c->journal.replay_journal_seq;
        }
 
        /*
         * Drop journal reservation after dropping write locks, since dropping
         * the journal reservation may kick off a journal write:
         */
-       bch2_journal_res_put(&c->journal, &trans->journal_res);
+       if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
+               bch2_journal_res_put(&c->journal, &trans->journal_res);
 
        return ret;
 }
        }
 retry:
        bch2_trans_verify_not_in_restart(trans);
-       memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+       if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
+               memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 
        ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
 
        if (ret)
                goto out;
 
+       /*
+        * We might have done another transaction commit in the error path -
+        * i.e. btree write buffer flush - which will have made use of
+        * trans->journal_res, but with BTREE_INSERT_JOURNAL_REPLAY that is how
+        * the journal sequence number to pin is passed in - so we must restart:
+        */
+       if (flags & BTREE_INSERT_JOURNAL_REPLAY) {
+               ret = -BCH_ERR_transaction_restart_nested;
+               goto out;
+       }
+
        goto retry;
 }
 
        unsigned update_flags = BTREE_TRIGGER_NORUN;
        int ret;
 
+       trans->journal_res.seq = k->journal_seq;
+
        /*
         * BTREE_UPDATE_KEY_CACHE_RECLAIM disables key cache lookup/update to
         * keep the key cache coherent with the underlying btree. Nothing