Merge pull request #10757: jewel: mon/osdmonitor: decouple adjust_hea…

…rtbeat_grace and min_down_reporters

Reviewed-by: Loic Dachary <ldachary@redhat.com>

src/mon/OSDMonitor.cc

@@ -1733,9 +1733,10 @@ bool OSDMonitor::check_failure(utime_t now, int target_osd, failure_info_t& fi)
 
   utime_t grace = orig_grace;
   double my_grace = 0, peer_grace = 0;
+  double decay_k = 0;
   if (g_conf->mon_osd_adjust_heartbeat_grace) {
     double halflife = (double)g_conf->mon_osd_laggy_halflife;
-    double decay_k = ::log(.5) / halflife;
+    decay_k = ::log(.5) / halflife;
 
     // scale grace period based on historical probability of 'lagginess'
     // (false positive failures due to slowness).
@@ -1745,31 +1746,35 @@ bool OSDMonitor::check_failure(utime_t now, int target_osd, failure_info_t& fi)
 	     << " failed_for " << failed_for << " decay " << decay << dendl;
     my_grace = decay * (double)xi.laggy_interval * xi.laggy_probability;
     grace += my_grace;
+  }
 
-    // consider the peers reporting a failure a proxy for a potential
-    // 'subcluster' over the overall cluster that is similarly
-    // laggy.  this is clearly not true in all cases, but will sometimes
-    // help us localize the grace correction to a subset of the system
-    // (say, a rack with a bad switch) that is unhappy.
-    assert(fi.reporters.size());
-    for (map<int,failure_reporter_t>::iterator p = fi.reporters.begin();
-	 p != fi.reporters.end();
-	 ++p) {
-      // get the parent bucket whose type matches with "reporter_subtree_level".
-      // fall back to OSD if the level doesn't exist.
-      map<string, string> reporter_loc = osdmap.crush->get_full_location(p->first);
-      map<string, string>::iterator iter = reporter_loc.find(reporter_subtree_level);
-      if (iter == reporter_loc.end()) {
-	reporters_by_subtree.insert("osd." + to_string(p->first));
-      } else {
-	reporters_by_subtree.insert(iter->second);
-      }
-
+  // consider the peers reporting a failure a proxy for a potential
+  // 'subcluster' over the overall cluster that is similarly
+  // laggy.  this is clearly not true in all cases, but will sometimes
+  // help us localize the grace correction to a subset of the system
+  // (say, a rack with a bad switch) that is unhappy.
+  assert(fi.reporters.size());
+  for (map<int,failure_reporter_t>::iterator p = fi.reporters.begin();
+	p != fi.reporters.end();
+	++p) {
+    // get the parent bucket whose type matches with "reporter_subtree_level".
+    // fall back to OSD if the level doesn't exist.
+    map<string, string> reporter_loc = osdmap.crush->get_full_location(p->first);
+    map<string, string>::iterator iter = reporter_loc.find(reporter_subtree_level);
+    if (iter == reporter_loc.end()) {
+      reporters_by_subtree.insert("osd." + to_string(p->first));
+    } else {
+      reporters_by_subtree.insert(iter->second);
+    }
+    if (g_conf->mon_osd_adjust_heartbeat_grace) {
       const osd_xinfo_t& xi = osdmap.get_xinfo(p->first);
       utime_t elapsed = now - xi.down_stamp;
       double decay = exp((double)elapsed * decay_k);
       peer_grace += decay * (double)xi.laggy_interval * xi.laggy_probability;
     }
+  }
+
+  if (g_conf->mon_osd_adjust_heartbeat_grace) {
     peer_grace /= (double)fi.reporters.size();
     grace += peer_grace;
   }