OSDMonitor: use a different approach to prevent extreme multipliers on PG splits

Signed-off-by: Greg Farnum <greg@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
(cherry picked from commit d8ccd73968fbd0753ca08916ebf1062cdb4d5ac1)

Conflicts:

	src/mon/OSDMonitor.cc

diff --git a/src/common/config_opts.h b/src/common/config_opts.h
index f83338195682102ed22576e0c250e019b08b3761..90f85c2baf9cad0b43d747726e3a7c6a5dcaab86 100644 (file)
--- a/src/common/config_opts.h
+++ b/src/common/config_opts.h
@@ -147,7 +147,7 @@ OPTION(mon_osd_down_out_subtree_limit, OPT_STR, "rack")   // smallest crush unit
 OPTION(mon_osd_min_up_ratio, OPT_DOUBLE, .3)    // min osds required to be up to mark things down
 OPTION(mon_osd_min_in_ratio, OPT_DOUBLE, .3)   // min osds required to be in to mark things out
 OPTION(mon_osd_max_op_age, OPT_DOUBLE, 32)     // max op age before we get concerned (make it a power of 2)
-OPTION(mon_osd_max_split_ratio, OPT_INT, 8) // largest multiple allowed when doing PG split
+OPTION(mon_osd_max_split_count, OPT_INT, 32) // largest number of PGs per "involved" OSD to let split create
 OPTION(mon_stat_smooth_intervals, OPT_INT, 2)  // smooth stats over last N PGMap maps
 OPTION(mon_lease, OPT_FLOAT, 5)       // lease interval
 OPTION(mon_lease_renew_interval, OPT_FLOAT, 3) // on leader, to renew the lease

diff --git a/src/mon/OSDMonitor.cc b/src/mon/OSDMonitor.cc
index e6d6983ec9675ea44121977f4ceff341bda8de7b..d98d37192b00c5e5e1627e6037b90729d435e8ae 100644 (file)
--- a/src/mon/OSDMonitor.cc
+++ b/src/mon/OSDMonitor.cc
@@ -3608,11 +3608,16 @@ done:
        pending_inc.new_pools[pool].crash_replay_interval = n;
        ss << "set pool " << pool << " to crash_replay_interval to " << n;
       } else if (var == "pg_num") {
+       int expected_osds = MIN(p->get_pg_num(), osdmap.get_num_osds());
+       int64_t new_pgs = n - p->get_pg_num();
+       int64_t pgs_per_osd = new_pgs / expected_osds;
        if (n <= p->get_pg_num()) {
          ss << "specified pg_num " << n << " <= current " << p->get_pg_num();
-       } else if (n > (int)p->get_pg_num() * g_conf->mon_osd_max_split_ratio) {
-         ss << "specified pg_num " << n << " is too large (> current "
-            << p->get_pg_num() << '*' << g_conf->mon_osd_max_split_ratio << ')';
+       } else if (pgs_per_osd > g_conf->mon_osd_max_split_count) {
+         ss << "specified pg_num " << n << " is too large (creating "
+            << new_pgs << " new PGs on ~" << expected_osds
+            << " OSDs exceeds per-OSD max of" << g_conf->mon_osd_max_split_count
+            << ')';
          err = -E2BIG;
        } else if (!mon->pgmon()->pg_map.creating_pgs.empty()) {
          ss << "currently creating pgs, wait";