From 842b7992ef9bbc5bb835cf83628f4a9e95a30da4 Mon Sep 17 00:00:00 2001
From: Noah Watkins <nwatkins@redhat.com>
Date: Thu, 26 Jul 2018 18:36:17 -0700
Subject: [PATCH] mon: timely health updates to manager

When the health of a monitor service changes immediately dispatch a
digest to manager daemons with the new health checks rather than waiting
for the next scheduled event. This has the added benefit that health
checks that are set by the manager daemon are echoed back and observable
quickly.

Signed-off-by: Noah Watkins <nwatkins@redhat.com>
---
 src/mon/MgrMonitor.cc | 34 +++++++++++++++++++++++++++++++++-
 src/mon/MgrMonitor.h  |  3 +++
 2 files changed, 36 insertions(+), 1 deletion(-)

diff --git a/src/mon/MgrMonitor.cc b/src/mon/MgrMonitor.cc
index a7af4241c00..ccbcb072ccb 100644
--- a/src/mon/MgrMonitor.cc
+++ b/src/mon/MgrMonitor.cc
@@ -149,6 +149,36 @@ health_status_t MgrMonitor::should_warn_about_mgr_down()
   return HEALTH_OK;
 }
 
+void MgrMonitor::post_paxos_update()
+{
+  // are we handling digest subscribers?
+  if (digest_event) {
+    bool send = false;
+    if (prev_health_checks.empty()) {
+      prev_health_checks.resize(mon->paxos_service.size());
+      send = true;
+    }
+    assert(prev_health_checks.size() == mon->paxos_service.size());
+    for (auto i = 0u; i < prev_health_checks.size(); i++) {
+      const auto& curr = mon->paxos_service[i]->get_health_checks();
+      if (!send && curr != prev_health_checks[i]) {
+        send = true;
+      }
+      prev_health_checks[i] = curr;
+    }
+    if (send) {
+      if (is_active()) {
+        send_digests();
+      } else {
+        cancel_timer();
+        wait_for_active_ctx(new C_MonContext(mon, [this](int) {
+          send_digests();
+        }));
+      }
+    }
+  }
+}
+
 void MgrMonitor::encode_pending(MonitorDBStore::TransactionRef t)
 {
   dout(10) << __func__ << " " << pending_map << dendl;
@@ -453,8 +483,10 @@ void MgrMonitor::send_digests()
   cancel_timer();
 
   const std::string type = "mgrdigest";
-  if (mon->session_map.subs.count(type) == 0)
+  if (mon->session_map.subs.count(type) == 0) {
+    prev_health_checks.clear();
     return;
+  }
 
   if (!is_active()) {
     // if paxos is currently not active, don't send a digest but reenable timer
diff --git a/src/mon/MgrMonitor.h b/src/mon/MgrMonitor.h
index d73b1f26dd9..b295e403009 100644
--- a/src/mon/MgrMonitor.h
+++ b/src/mon/MgrMonitor.h
@@ -57,6 +57,8 @@ class MgrMonitor: public PaxosService
   Context *digest_event = nullptr;
   void cancel_timer();
 
+  std::vector<health_check_map_t> prev_health_checks;
+
   bool check_caps(MonOpRequestRef op, const uuid_d& fsid);
 
   health_status_t should_warn_about_mgr_down();
@@ -83,6 +85,7 @@ public:
   void create_initial() override;
   void get_store_prefixes(std::set<string>& s) const override;
   void update_from_paxos(bool *need_bootstrap) override;
+  void post_paxos_update() override;
   void create_pending() override;
   void encode_pending(MonitorDBStore::TransactionRef t) override;
 
-- 
2.47.3