update sources to v12.1.3

[ceph.git] / ceph / src / osd / PG.cc

diff --git a/ceph/src/osd/PG.cc b/ceph/src/osd/PG.cc
index 37ae7d797831fc015084cd0fd9d7747f0bdcb8db..e0707efe2d581bf8ce1ca81ff5c3ab2391e8c01c 100644 (file)
--- a/ceph/src/osd/PG.cc
+++ b/ceph/src/osd/PG.cc
@@ -1644,6 +1644,7 @@ void PG::activate(ObjectStore::Transaction& t,
       dout(10) << "activate peer osd." << peer << " " << pi << dendl;
 
       MOSDPGLog *m = 0;
+      assert(peer_missing.count(peer));
       pg_missing_t& pm = peer_missing[peer];
 
       bool needs_past_intervals = pi.dne();
@@ -2041,17 +2042,17 @@ void PG::mark_clean()
   kick_snap_trim();
 }
 
-void PG::change_recovery_force_mode(int new_mode, bool clear)
+void PG::_change_recovery_force_mode(int new_mode, bool clear)
 {
-  lock(true);
-  if (clear) {
-    state_clear(new_mode);
-  } else {
-    state_set(new_mode);
+  if (!deleting) {
+    // we can't and shouldn't do anything if the PG is being deleted locally
+    if (clear) {
+      state_clear(new_mode);
+    } else {
+      state_set(new_mode);
+    }
+    publish_stats_to_osd();
   }
-  publish_stats_to_osd();
-
-  unlock();
 }
 
 inline int PG::clamp_recovery_priority(int priority)
@@ -2229,6 +2230,10 @@ void PG::split_into(pg_t child_pgid, PG *child, unsigned split_bits)
     // in the future).
     info.set_last_backfill(hobject_t());
     child->info.set_last_backfill(hobject_t());
+    // restarting backfill implies that the missing set is empty,
+    // since it is only used for objects prior to last_backfill
+    pg_log.reset_backfill();
+    child->pg_log.reset_backfill();
   }
 
   child->info.stats = info.stats;
@@ -3310,6 +3315,17 @@ void PG::read_state(ObjectStore *store, bufferlist &bl)
 
   last_written_info = info;
 
+  // if we are upgrading from jewel, we need to force rebuild of
+  // missing set.  v9 was fastinfo, added v11.0.2-331-g1d5dc29a13
+  // (before kraken).  persisted missing set was circa
+  // v11.0.0-866-gb0e239da95 (a bit earlier, also before kraken).
+  // v8 was pre-jewel (per-pg meta object).
+  bool force_rebuild_missing = info_struct_v < 9;
+  if (force_rebuild_missing) {
+    dout(10) << __func__ << " detected upgrade from jewel, force_rebuild_missing"
+            << dendl;
+  }
+
   ostringstream oss;
   pg_log.read_log_and_missing(
     store,
@@ -3317,12 +3333,19 @@ void PG::read_state(ObjectStore *store, bufferlist &bl)
     info_struct_v < 8 ? coll_t::meta() : coll,
     ghobject_t(info_struct_v < 8 ? OSD::make_pg_log_oid(pg_id) : pgmeta_oid),
     info,
+    force_rebuild_missing,
     oss,
     cct->_conf->osd_ignore_stale_divergent_priors,
     cct->_conf->osd_debug_verify_missing_on_start);
   if (oss.tellp())
     osd->clog->error() << oss.rdbuf();
 
+  if (force_rebuild_missing) {
+    dout(10) << __func__ << " forced rebuild of missing got "
+            << pg_log.get_missing()
+            << dendl;
+  }
+
   // log any weirdness
   log_weirdness();
 }
@@ -4492,7 +4515,7 @@ void PG::chunky_scrub(ThreadPool::TPHandle &handle)
          const char *mode = (repair ? "repair": (deep_scrub ? "deep-scrub" : "scrub"));
          stringstream oss;
          oss << info.pgid.pgid << " " << mode << " starts" << std::endl;
-         osd->clog->info(oss);
+         osd->clog->debug(oss);
        }
 
        scrubber.seed = -1;
@@ -4931,7 +4954,7 @@ void PG::scrub_finish()
     if (total_errors)
       osd->clog->error(oss);
     else
-      osd->clog->info(oss);
+      osd->clog->debug(oss);
   }
 
   // finish up
@@ -5069,6 +5092,7 @@ void PG::merge_new_log_entries(
     assert(peer_missing.count(peer));
     assert(peer_info.count(peer));
     pg_missing_t& pmissing(peer_missing[peer]);
+    dout(20) << __func__ << " peer_missing for " << peer << " = " << pmissing << dendl;
     pg_info_t& pinfo(peer_info[peer]);
     bool invalidate_stats = PGLog::append_log_entries_update_missing(
       pinfo.last_backfill,
@@ -8003,18 +8027,23 @@ PG::RecoveryState::GetMissing::GetMissing(my_context ctx)
        ++i) {
     if (*i == pg->get_primary()) continue;
     const pg_info_t& pi = pg->peer_info[*i];
+    // reset this so to make sure the pg_missing_t is initialized and
+    // has the correct semantics even if we don't need to get a
+    // missing set from a shard. This way later additions due to
+    // lost+unfound delete work properly.
+    pg->peer_missing[*i].may_include_deletes = !pg->perform_deletes_during_peering();
 
     if (pi.is_empty())
       continue;                                // no pg data, nothing divergent
 
     if (pi.last_update < pg->pg_log.get_tail()) {
       ldout(pg->cct, 10) << " osd." << *i << " is not contiguous, will restart backfill" << dendl;
-      pg->peer_missing[*i];
+      pg->peer_missing[*i].clear();
       continue;
     }
     if (pi.last_backfill == hobject_t()) {
       ldout(pg->cct, 10) << " osd." << *i << " will fully backfill; can infer empty missing set" << dendl;
-      pg->peer_missing[*i];
+      pg->peer_missing[*i].clear();
       continue;
     }
 
@@ -8025,7 +8054,7 @@ PG::RecoveryState::GetMissing::GetMissing(my_context ctx)
       // FIXME: we can do better here.  if last_update==last_complete we
       //        can infer the rest!
       ldout(pg->cct, 10) << " osd." << *i << " has no missing, identical log" << dendl;
-      pg->peer_missing[*i];
+      pg->peer_missing[*i].clear();
       continue;
     }