mon/OSDMonitor: account for PG merging in epoch_by_pg accounting

After a pool has merged PGs, the epoch_by_pg accounting will refer
to osdmap epochs of PGs that no longer exist. We'll never again get
OSD beacons for these PGs, so the min epoch in epoch_by_pg will not
advance until the mon leader has restarted. The effect of this is
that osdmaps are not trimmed after a pool has undergone PG merging,
until the mon leader restarts. To fix, we unconditionally resize
epoch_by_pg to the pg_num of the pool during each beacon report.

Fixes: https://tracker.ceph.com/issues/48212
Signed-off-by: Dan van der Ster <daniel.vanderster@cern.ch>
(cherry picked from commit cf5ea22cc0b10560c9fa3fbd5d93431f874d38b9)

diff --git a/src/mon/OSDMonitor.cc b/src/mon/OSDMonitor.cc
index 877ce53a6fd4a288d2e9d6ae1421119563274161..d69b8f542b726f41bbf5a800e2fab6edb948ed61 100644 (file)
--- a/src/mon/OSDMonitor.cc
+++ b/src/mon/OSDMonitor.cc
@@ -318,11 +318,14 @@ bool is_unmanaged_snap_op_permitted(CephContext* cct,
 
 } // anonymous namespace
 
-void LastEpochClean::Lec::report(ps_t ps, epoch_t last_epoch_clean)
+void LastEpochClean::Lec::report(unsigned pg_num, ps_t ps,
+                                epoch_t last_epoch_clean)
 {
-  if (epoch_by_pg.size() <= ps) {
-    epoch_by_pg.resize(ps + 1, 0);
+  if (ps >= pg_num) {
+    // removed PG
+    return;
   }
+  epoch_by_pg.resize(pg_num, 0);
   const auto old_lec = epoch_by_pg[ps];
   if (old_lec >= last_epoch_clean) {
     // stale lec
@@ -354,10 +357,11 @@ void LastEpochClean::remove_pool(uint64_t pool)
   report_by_pool.erase(pool);
 }
 
-void LastEpochClean::report(const pg_t& pg, epoch_t last_epoch_clean)
+void LastEpochClean::report(unsigned pg_num, const pg_t& pg,
+                           epoch_t last_epoch_clean)
 {
   auto& lec = report_by_pool[pg.pool()];
-  return lec.report(pg.ps(), last_epoch_clean);
+  return lec.report(pg_num, pg.ps(), last_epoch_clean);
 }
 
 epoch_t LastEpochClean::get_lower_bound(const OSDMap& latest) const
@@ -4322,7 +4326,10 @@ bool OSDMonitor::prepare_beacon(MonOpRequestRef op)
   osd_epochs[from] = beacon->version;
 
   for (const auto& pg : beacon->pgs) {
-    last_epoch_clean.report(pg, beacon->min_last_epoch_clean);
+    if (auto* pool = osdmap.get_pg_pool(pg.pool()); pool != nullptr) {
+      unsigned pg_num = pool->get_pg_num();
+      last_epoch_clean.report(pg_num, pg, beacon->min_last_epoch_clean);
+    }
   }
 
   if (osdmap.osd_xinfo[from].last_purged_snaps_scrub <

diff --git a/src/mon/OSDMonitor.h b/src/mon/OSDMonitor.h
index fb941a6a0e56ef740a4dd75f514b92e7a7491a13..7986eb02d703b63aab216fafd7b08ce3568f4590 100644 (file)
--- a/src/mon/OSDMonitor.h
+++ b/src/mon/OSDMonitor.h
@@ -111,11 +111,11 @@ class LastEpochClean {
     vector<epoch_t> epoch_by_pg;
     ps_t next_missing = 0;
     epoch_t floor = std::numeric_limits<epoch_t>::max();
-    void report(ps_t pg, epoch_t last_epoch_clean);
+    void report(unsigned pg_num, ps_t pg, epoch_t last_epoch_clean);
   };
   std::map<uint64_t, Lec> report_by_pool;
 public:
-  void report(const pg_t& pg, epoch_t last_epoch_clean);
+  void report(unsigned pg_num, const pg_t& pg, epoch_t last_epoch_clean);
   void remove_pool(uint64_t pool);
   epoch_t get_lower_bound(const OSDMap& latest) const;