ceph/src/mon/HealthMonitor.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2013 Inktank, Inc
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15 #include <stdlib.h>
  16 #include <limits.h>
  17 #include <sstream>
  18 #include <boost/regex.hpp>
  19
  20 #include "include/assert.h"
  21 #include "include/stringify.h"
  22
  23 #include "mon/Monitor.h"
  24 #include "mon/HealthService.h"
  25 #include "mon/HealthMonitor.h"
  26 #include "mon/DataHealthService.h"
  27
  28 #include "messages/MMonHealth.h"
  29 #include "messages/MMonHealthChecks.h"
  30
  31 #include "common/Formatter.h"
  32
  33 #define dout_subsys ceph_subsys_mon
  34 #undef dout_prefix
  35 #define dout_prefix _prefix(_dout, mon, this)
  36 static ostream& _prefix(std::ostream *_dout, const Monitor *mon,
  37                         const HealthMonitor *hmon) {
  38   return *_dout << "mon." << mon->name << "@" << mon->rank
  39                 << "(" << mon->get_state_name() << ").health ";
  40 }
  41
  42 HealthMonitor::HealthMonitor(Monitor *m, Paxos *p, const string& service_name)
  43   : PaxosService(m, p, service_name) {
  44 }
  45
  46 void HealthMonitor::init()
  47 {
  48   dout(10) << __func__ << dendl;
  49 }
  50
  51 void HealthMonitor::create_initial()
  52 {
  53   dout(10) << __func__ << dendl;
  54 }
  55
  56 void HealthMonitor::update_from_paxos(bool *need_bootstrap)
  57 {
  58   version = get_last_committed();
  59   dout(10) << __func__ << dendl;
  60   load_health();
  61
  62   bufferlist qbl;
  63   mon->store->get(service_name, "quorum", qbl);
  64   if (qbl.length()) {
  65     auto p = qbl.begin();
  66     ::decode(quorum_checks, p);
  67   } else {
  68     quorum_checks.clear();
  69   }
  70
  71   bufferlist lbl;
  72   mon->store->get(service_name, "leader", lbl);
  73   if (lbl.length()) {
  74     auto p = lbl.begin();
  75     ::decode(leader_checks, p);
  76   } else {
  77     leader_checks.clear();
  78   }
  79
  80   dout(20) << "dump:";
  81   JSONFormatter jf(true);
  82   jf.open_object_section("health");
  83   jf.open_object_section("quorum_health");
  84   for (auto& p : quorum_checks) {
  85     string s = string("mon.") + stringify(p.first);
  86     jf.dump_object(s.c_str(), p.second);
  87   }
  88   jf.close_section();
  89   jf.dump_object("leader_health", leader_checks);
  90   jf.close_section();
  91   jf.flush(*_dout);
  92   *_dout << dendl;
  93 }
  94
  95 void HealthMonitor::create_pending()
  96 {
  97   dout(10) << " " << version << dendl;
  98 }
  99
 100 void HealthMonitor::encode_pending(MonitorDBStore::TransactionRef t)
 101 {
 102   ++version;
 103   dout(10) << " " << version << dendl;
 104   put_last_committed(t, version);
 105
 106   bufferlist qbl;
 107   ::encode(quorum_checks, qbl);
 108   t->put(service_name, "quorum", qbl);
 109   bufferlist lbl;
 110   ::encode(leader_checks, lbl);
 111   t->put(service_name, "leader", lbl);
 112
 113   health_check_map_t pending_health;
 114
 115   // combine per-mon details carefully...
 116   map<string,set<string>> names; // code -> <mon names>
 117   for (auto p : quorum_checks) {
 118     for (auto q : p.second.checks) {
 119       names[q.first].insert(mon->monmap->get_name(p.first));
 120     }
 121     pending_health.merge(p.second);
 122   }
 123   for (auto p : pending_health.checks) {
 124     p.second.summary = boost::regex_replace(
 125       p.second.summary,
 126       boost::regex("%num%"), stringify(names[p.first].size()));
 127     p.second.summary = boost::regex_replace(
 128       p.second.summary,
 129       boost::regex("%names%"), stringify(names[p.first]));
 130     p.second.summary = boost::regex_replace(
 131       p.second.summary,
 132       boost::regex("%plurals%"),
 133       names[p.first].size() > 1 ? "s" : "");
 134     p.second.summary = boost::regex_replace(
 135       p.second.summary,
 136       boost::regex("%isorare%"),
 137       names[p.first].size() > 1 ? "are" : "is");
 138   }
 139
 140   pending_health.merge(leader_checks);
 141   encode_health(pending_health, t);
 142 }
 143
 144 version_t HealthMonitor::get_trim_to()
 145 {
 146   // we don't actually need *any* old states, but keep a few.
 147   if (version > 5) {
 148     return version - 5;
 149   }
 150   return 0;
 151 }
 152
 153 bool HealthMonitor::preprocess_query(MonOpRequestRef op)
 154 {
 155   switch (op->get_req()->get_type()) {
 156   case MSG_MON_HEALTH:
 157     {
 158       MMonHealth *hm = static_cast<MMonHealth*>(op->get_req());
 159       int service_type = hm->get_service_type();
 160       if (services.count(service_type) == 0) {
 161         dout(1) << __func__ << " service type " << service_type
 162                 << " not registered -- drop message!" << dendl;
 163         return false;
 164       }
 165       return services[service_type]->service_dispatch(op);
 166     }
 167
 168   case MSG_MON_HEALTH_CHECKS:
 169     return preprocess_health_checks(op);
 170   }
 171   return false;
 172 }
 173
 174 bool HealthMonitor::prepare_update(MonOpRequestRef op)
 175 {
 176   return false;
 177 }
 178
 179 bool HealthMonitor::preprocess_health_checks(MonOpRequestRef op)
 180 {
 181   MMonHealthChecks *m = static_cast<MMonHealthChecks*>(op->get_req());
 182   quorum_checks[m->get_source().num()] = m->health_checks;
 183   return true;
 184 }
 185
 186 void HealthMonitor::tick()
 187 {
 188   if (!is_active()) {
 189     return;
 190   }
 191   dout(10) << __func__ << dendl;
 192   bool changed = false;
 193   if (check_member_health()) {
 194     changed = true;
 195   }
 196   if (mon->is_leader()) {
 197     if (check_leader_health()) {
 198       changed = true;
 199     }
 200   }
 201   if (changed) {
 202     propose_pending();
 203   }
 204 }
 205
 206 bool HealthMonitor::check_member_health()
 207 {
 208   dout(20) << __func__ << dendl;
 209   bool changed = false;
 210
 211   // snapshot of usage
 212   DataStats stats;
 213   get_fs_stats(stats.fs_stats, g_conf->mon_data.c_str());
 214   map<string,uint64_t> extra;
 215   uint64_t store_size = mon->store->get_estimated_size(extra);
 216   assert(store_size > 0);
 217   stats.store_stats.bytes_total = store_size;
 218   stats.store_stats.bytes_sst = extra["sst"];
 219   stats.store_stats.bytes_log = extra["log"];
 220   stats.store_stats.bytes_misc = extra["misc"];
 221   stats.last_update = ceph_clock_now();
 222   dout(10) << __func__ << " avail " << stats.fs_stats.avail_percent << "%"
 223            << " total " << prettybyte_t(stats.fs_stats.byte_total)
 224            << ", used " << prettybyte_t(stats.fs_stats.byte_used)
 225            << ", avail " << prettybyte_t(stats.fs_stats.byte_avail) << dendl;
 226
 227   // MON_DISK_{LOW,CRIT,BIG}
 228   health_check_map_t next;
 229   if (stats.fs_stats.avail_percent <= g_conf->mon_data_avail_crit) {
 230     stringstream ss, ss2;
 231     ss << "mon%plurals% %names% %isorare% very low on available space";
 232     auto& d = next.add("MON_DISK_CRIT", HEALTH_ERR, ss.str());
 233     ss2 << "mon." << mon->name << " has " << stats.fs_stats.avail_percent
 234         << "% avail";
 235     d.detail.push_back(ss2.str());
 236   } else if (stats.fs_stats.avail_percent <= g_conf->mon_data_avail_warn) {
 237     stringstream ss, ss2;
 238     ss << "mon%plurals% %names% %isorare% low on available space";
 239     auto& d = next.add("MON_DISK_LOW", HEALTH_ERR, ss.str());
 240     ss2 << "mon." << mon->name << " has " << stats.fs_stats.avail_percent
 241         << "% avail";
 242     d.detail.push_back(ss2.str());
 243   }
 244   if (stats.store_stats.bytes_total >= g_conf->mon_data_size_warn) {
 245     stringstream ss, ss2;
 246     ss << "mon%plurals% %names% %isorare% using a lot of disk space";
 247     auto& d = next.add("MON_DISK_BIG", HEALTH_WARN, ss.str());
 248     ss2 << "mon." << mon->name << " is "
 249         << prettybyte_t(stats.store_stats.bytes_total)
 250         << " >= mon_data_size_warn ("
 251         << prettybyte_t(g_conf->mon_data_size_warn) << ")";
 252     d.detail.push_back(ss2.str());
 253   }
 254
 255   auto p = quorum_checks.find(mon->rank);
 256   if (p == quorum_checks.end() ||
 257       p->second != next) {
 258     if (mon->is_leader()) {
 259       // prepare to propose
 260       quorum_checks[mon->rank] = next;
 261       changed = true;
 262     } else {
 263       // tell the leader
 264       mon->messenger->send_message(new MMonHealthChecks(next),
 265                                    mon->monmap->get_inst(mon->get_leader()));
 266     }
 267   }
 268
 269   // OSD_NO_DOWN_OUT_INTERVAL
 270   {
 271     // Warn if 'mon_osd_down_out_interval' is set to zero.
 272     // Having this option set to zero on the leader acts much like the
 273     // 'noout' flag.  It's hard to figure out what's going wrong with clusters
 274     // without the 'noout' flag set but acting like that just the same, so
 275     // we report a HEALTH_WARN in case this option is set to zero.
 276     // This is an ugly hack to get the warning out, but until we find a way
 277     // to spread global options throughout the mon cluster and have all mons
 278     // using a base set of the same options, we need to work around this sort
 279     // of things.
 280     // There's also the obvious drawback that if this is set on a single
 281     // monitor on a 3-monitor cluster, this warning will only be shown every
 282     // third monitor connection.
 283     if (g_conf->mon_warn_on_osd_down_out_interval_zero &&
 284         g_conf->mon_osd_down_out_interval == 0) {
 285       ostringstream ss, ds;
 286       ss << "mon%plurals% %names %hasorhave% mon_osd_down_out_interval set to 0";
 287       auto& d = next.add("OSD_NO_DOWN_OUT_INTERVAL", HEALTH_WARN, ss.str());
 288       ds << "mon." << mon->name << " has mon_osd_down_out_interval set to 0";
 289       d.detail.push_back(ds.str());
 290     }
 291   }
 292
 293   return changed;
 294 }
 295
 296 bool HealthMonitor::check_leader_health()
 297 {
 298   dout(20) << __func__ << dendl;
 299   bool changed = false;
 300
 301   // prune quorum_health
 302   {
 303     auto& qset = mon->get_quorum();
 304     auto p = quorum_checks.begin();
 305     while (p != quorum_checks.end()) {
 306       if (qset.count(p->first) == 0) {
 307         p = quorum_checks.erase(p);
 308         changed = true;
 309       } else {
 310         ++p;
 311       }
 312     }
 313   }
 314
 315   health_check_map_t next;
 316
 317   // MON_DOWN
 318   {
 319     int max = mon->monmap->size();
 320     int actual = mon->get_quorum().size();
 321     if (actual < max) {
 322       ostringstream ss;
 323       ss << (max-actual) << "/" << max << " mons down, quorum "
 324          << mon->get_quorum_names();
 325       auto& d = next.add("MON_DOWN", HEALTH_WARN, ss.str());
 326       set<int> q = mon->get_quorum();
 327       for (int i=0; i<max; i++) {
 328         if (q.count(i) == 0) {
 329           ostringstream ss;
 330           ss << "mon." << mon->monmap->get_name(i) << " (rank " << i
 331              << ") addr " << mon->monmap->get_addr(i)
 332              << " is down (out of quorum)";
 333           d.detail.push_back(ss.str());
 334         }
 335       }
 336     }
 337   }
 338
 339   // MON_CLOCK_SKEW
 340   if (!mon->timecheck_skews.empty()) {
 341     list<string> warns;
 342     list<string> details;
 343     for (map<entity_inst_t,double>::iterator i = mon->timecheck_skews.begin();
 344          i != mon->timecheck_skews.end(); ++i) {
 345       entity_inst_t inst = i->first;
 346       double skew = i->second;
 347       double latency = mon->timecheck_latencies[inst];
 348       string name = mon->monmap->get_name(inst.addr);
 349       ostringstream tcss;
 350       health_status_t tcstatus = mon->timecheck_status(tcss, skew, latency);
 351       if (tcstatus != HEALTH_OK) {
 352         warns.push_back(name);
 353         ostringstream tmp_ss;
 354         tmp_ss << "mon." << name
 355                << " addr " << inst.addr << " " << tcss.str()
 356                << " (latency " << latency << "s)";
 357         details.push_back(tmp_ss.str());
 358       }
 359     }
 360     if (!warns.empty()) {
 361       ostringstream ss;
 362       ss << "clock skew detected on";
 363       while (!warns.empty()) {
 364         ss << " mon." << warns.front();
 365         warns.pop_front();
 366         if (!warns.empty())
 367           ss << ",";
 368       }
 369       auto& d = next.add("MON_CLOCK_SKEW", HEALTH_WARN,
 370                          "monitor clock skew detected");
 371       d.detail.swap(details);
 372     }
 373   }
 374
 375   if (next != leader_checks) {
 376     changed = true;
 377     leader_checks = next;
 378   }
 379   return changed;
 380 }