mgr/DaemonServer: prevent pgp_num reductions from outpacing pg_num merges

If we are merging lots of pgs down to a much smaller number of pgs, and
the pgs are able to move quickly (faster than the merges happen), we can
end up with too many pgs on a small number of osds, triggering the max
pgs per osd limits.

Avoid this by preventing the pgp_num reductions from getting too far
out in front of the merges themselves.  Basically, cap the delta between
pgp_num and pg_num to the max_misplaced ratio.  We are already limiting
the movement caused by pgp_num by max_misplaced; this effectively just
makes sure that the actual merging (and pg_num reductions) are keeping
up.

Fixes: http://tracker.ceph.com/issues/38786
Signed-off-by: Sage Weil <sage@redhat.com>
(cherry picked from commit 76503a1438fa1f166d2c230c73ca8d7b67e6468d)

diff --git a/src/mgr/DaemonServer.cc b/src/mgr/DaemonServer.cc
index cf1d79bf5ef9af6f667c1770474cbb7565d3a655..38143ecc227f8c36e1c0f625919539018840eca8 100644 (file)
--- a/src/mgr/DaemonServer.cc
+++ b/src/mgr/DaemonServer.cc
@@ -2601,6 +2601,22 @@ void DaemonServer::adjust_pgs()
              }
              dout(20) << " room " << room << " estmax " << estmax
                       << " delta " << delta << " next " << next << dendl;
+             if (p.get_pgp_num_target() == p.get_pg_num_target()) {
+               // since pgp_num is tracking pg_num, ceph is handling
+               // pgp_num.  so, be responsible: don't let pgp_num get
+               // too far out ahead of merges (if we are merging).
+               // this avoids moving lots of unmerged pgs onto a
+               // small number of OSDs where we might blow out the
+               // per-osd pg max.
+               unsigned max_outpace_merges =
+                 std::max<unsigned>(8, p.get_pg_num() * max_misplaced);
+               if (next + max_outpace_merges < p.get_pg_num()) {
+                 next = p.get_pg_num() - max_outpace_merges;
+                 dout(10) << "  using next " << next
+                          << " to avoid outpacing merges (max_outpace_merges "
+                          << max_outpace_merges << ")" << dendl;
+               }
+             }
            }
            dout(10) << "pool " << i.first
                     << " pgp_num_target " << p.get_pgp_num_target()