crimson/osd/pg: Handle peer replies handling once received

Instead of updating last complete on disk and calling complete_write
in PG::submit_transacion (chained to all_completed).

Move the completion handling earlier as soon as all the peers acked.
This essentially means that we move the handling to
ReplicatedBackend::get_rep_op_reply, unless replication size is 1 -
then we would mark completion in ReplicatedBackend::submit_transaction.

Fixes: https://tracker.ceph.com/issues/69439
Signed-off-by: Matan Breizman <mbreizma@redhat.com>
Signed-off-by: Xuehan Xu <xuxuehan@qianxin.com>

diff --git a/src/crimson/osd/pg.cc b/src/crimson/osd/pg.cc
index 43cd46611640b588d8db478018ede379a83d3e1b..d49e056881afa40bebf9e1df6d0009d399839259 100644 (file)
--- a/src/crimson/osd/pg.cc
+++ b/src/crimson/osd/pg.cc
@@ -930,16 +930,7 @@ PG::submit_transaction(
       std::move(log_entries));
   co_return std::make_tuple(
     std::move(submitted),
-    all_completed.then_interruptible(
-      [this, last_complete=peering_state.get_info().last_complete, at_version]
-      (auto acked) {
-      for (const auto& peer : acked) {
-        peering_state.update_peer_last_complete_ondisk(
-          peer.shard, peer.last_complete_ondisk);
-      }
-      peering_state.complete_write(at_version, last_complete);
-      return seastar::now();
-    })
+    std::move(all_completed)
   );
 }
 

diff --git a/src/crimson/osd/replicated_backend.cc b/src/crimson/osd/replicated_backend.cc
index 987432e08dbd295c150a0e299856488554a01a3b..ebb34a52a6f070bf35666c365a78288a933b788c 100644 (file)
--- a/src/crimson/osd/replicated_backend.cc
+++ b/src/crimson/osd/replicated_backend.cc
@@ -166,10 +166,13 @@ ReplicatedBackend::submit_transaction(
       assert(0 == "impossible");
     }
     if (--peers->pending == 0) {
+      // no peers other than me, replication size is 1
+      pg.complete_write(peers->at_version, peers->last_complete);
       peers->all_committed.set_value();
       peers->all_committed = {};
       return seastar::now();
     }
+    // wait for all peers to ack (ReplicatedBackend::got_rep_op_reply)
     return peers->all_committed.get_shared_future();
   }).then_interruptible([pending_txn, this, _new_clone, &hoid,
                        to_push_delete=std::move(to_push_delete),
@@ -216,7 +219,10 @@ void ReplicatedBackend::got_rep_op_reply(const MOSDRepOpReply& reply)
   for (auto& peer : peers.acked_peers) {
     if (peer.shard == reply.from) {
       peer.last_complete_ondisk = reply.get_last_complete_ondisk();
+      pg.update_peer_last_complete_ondisk(
+       peer.shard, peer.last_complete_ondisk);
       if (--peers.pending == 0) {
+        pg.complete_write(peers.at_version, peers.last_complete);
         peers.all_committed.set_value();
         peers.all_committed = {};
       }