ceph/src/mgr/ClusterState.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) 2014 John Spray <john.spray@inktank.com>
   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 #include "messages/MMgrDigest.h"
  15 #include "messages/MMonMgrReport.h"
  16 #include "messages/MPGStats.h"
  17
  18 #include "mgr/ClusterState.h"
  19 #include <time.h>
  20 #include <boost/range/adaptor/reversed.hpp>
  21
  22 #define dout_context g_ceph_context
  23 #define dout_subsys ceph_subsys_mgr
  24 #undef dout_prefix
  25 #define dout_prefix *_dout << "mgr " << __func__ << " "
  26
  27 ClusterState::ClusterState(
  28   MonClient *monc_,
  29   Objecter *objecter_,
  30   const MgrMap& mgrmap)
  31   : monc(monc_),
  32     objecter(objecter_),
  33     mgr_map(mgrmap),
  34     asok_hook(NULL)
  35 {}
  36
  37 void ClusterState::set_objecter(Objecter *objecter_)
  38 {
  39   std::lock_guard l(lock);
  40
  41   objecter = objecter_;
  42 }
  43
  44 void ClusterState::set_fsmap(FSMap const &new_fsmap)
  45 {
  46   std::lock_guard l(lock);
  47
  48   fsmap = new_fsmap;
  49 }
  50
  51 void ClusterState::set_mgr_map(MgrMap const &new_mgrmap)
  52 {
  53   std::lock_guard l(lock);
  54   mgr_map = new_mgrmap;
  55 }
  56
  57 void ClusterState::set_service_map(ServiceMap const &new_service_map)
  58 {
  59   std::lock_guard l(lock);
  60   servicemap = new_service_map;
  61 }
  62
  63 void ClusterState::load_digest(MMgrDigest *m)
  64 {
  65   std::lock_guard l(lock);
  66   health_json = std::move(m->health_json);
  67   mon_status_json = std::move(m->mon_status_json);
  68 }
  69
  70 void ClusterState::ingest_pgstats(ref_t<MPGStats> stats)
  71 {
  72   std::lock_guard l(lock);
  73
  74   const int from = stats->get_orig_source().num();
  75   pending_inc.update_stat(from, std::move(stats->osd_stat));
  76
  77   for (auto p : stats->pg_stat) {
  78     pg_t pgid = p.first;
  79     const auto &pg_stats = p.second;
  80
  81     // In case we're hearing about a PG that according to last
  82     // OSDMap update should not exist
  83     auto r = existing_pools.find(pgid.pool());
  84     if (r == existing_pools.end()) {
  85       dout(15) << " got " << pgid
  86                << " reported at " << pg_stats.reported_epoch << ":"
  87                << pg_stats.reported_seq
  88                << " state " << pg_state_string(pg_stats.state)
  89                << " but pool not in " << existing_pools
  90                << dendl;
  91       continue;
  92     }
  93     if (pgid.ps() >= r->second) {
  94       dout(15) << " got " << pgid
  95                << " reported at " << pg_stats.reported_epoch << ":"
  96                << pg_stats.reported_seq
  97                << " state " << pg_state_string(pg_stats.state)
  98                << " but > pg_num " << r->second
  99                << dendl;
 100       continue;
 101     }
 102     // In case we already heard about more recent stats from this PG
 103     // from another OSD
 104     const auto q = pg_map.pg_stat.find(pgid);
 105     if (q != pg_map.pg_stat.end() &&
 106         q->second.get_version_pair() > pg_stats.get_version_pair()) {
 107       dout(15) << " had " << pgid << " from "
 108                << q->second.reported_epoch << ":"
 109                << q->second.reported_seq << dendl;
 110       continue;
 111     }
 112
 113     pending_inc.pg_stat_updates[pgid] = pg_stats;
 114   }
 115   for (auto p : stats->pool_stat) {
 116     pending_inc.pool_statfs_updates[std::make_pair(p.first, from)] = p.second;
 117   }
 118 }
 119
 120 void ClusterState::update_delta_stats()
 121 {
 122   pending_inc.stamp = ceph_clock_now();
 123   pending_inc.version = pg_map.version + 1; // to make apply_incremental happy
 124   dout(10) << " v" << pending_inc.version << dendl;
 125
 126   dout(30) << " pg_map before:\n";
 127   JSONFormatter jf(true);
 128   jf.dump_object("pg_map", pg_map);
 129   jf.flush(*_dout);
 130   *_dout << dendl;
 131   dout(30) << " incremental:\n";
 132   JSONFormatter jf(true);
 133   jf.dump_object("pending_inc", pending_inc);
 134   jf.flush(*_dout);
 135   *_dout << dendl;
 136   pg_map.apply_incremental(g_ceph_context, pending_inc);
 137   pending_inc = PGMap::Incremental();
 138 }
 139
 140 void ClusterState::notify_osdmap(const OSDMap &osd_map)
 141 {
 142   assert(ceph_mutex_is_locked(lock));
 143
 144   pending_inc.stamp = ceph_clock_now();
 145   pending_inc.version = pg_map.version + 1; // to make apply_incremental happy
 146   dout(10) << " v" << pending_inc.version << dendl;
 147
 148   PGMapUpdater::check_osd_map(g_ceph_context, osd_map, pg_map, &pending_inc);
 149
 150   // update our list of pools that exist, so that we can filter pg_map updates
 151   // in synchrony with this OSDMap.
 152   existing_pools.clear();
 153   for (auto& p : osd_map.get_pools()) {
 154     existing_pools[p.first] = p.second.get_pg_num();
 155   }
 156
 157   // brute force this for now (don't bother being clever by only
 158   // checking osds that went up/down)
 159   set<int> need_check_down_pg_osds;
 160   PGMapUpdater::check_down_pgs(osd_map, pg_map, true,
 161                                need_check_down_pg_osds, &pending_inc);
 162
 163   dout(30) << " pg_map before:\n";
 164   JSONFormatter jf(true);
 165   jf.dump_object("pg_map", pg_map);
 166   jf.flush(*_dout);
 167   *_dout << dendl;
 168   dout(30) << " incremental:\n";
 169   JSONFormatter jf(true);
 170   jf.dump_object("pending_inc", pending_inc);
 171   jf.flush(*_dout);
 172   *_dout << dendl;
 173
 174   pg_map.apply_incremental(g_ceph_context, pending_inc);
 175   pending_inc = PGMap::Incremental();
 176   // TODO: Complete the separation of PG state handling so
 177   // that a cut-down set of functionality remains in PGMonitor
 178   // while the full-blown PGMap lives only here.
 179 }
 180
 181 class ClusterSocketHook : public AdminSocketHook {
 182   ClusterState *cluster_state;
 183 public:
 184   explicit ClusterSocketHook(ClusterState *o) : cluster_state(o) {}
 185   int call(std::string_view admin_command, const cmdmap_t& cmdmap,
 186            Formatter *f,
 187            std::ostream& errss,
 188            bufferlist& out) override {
 189     stringstream outss;
 190     int r = 0;
 191     try {
 192       r = cluster_state->asok_command(admin_command, cmdmap, f, outss);
 193       out.append(outss);
 194     } catch (const TOPNSPC::common::bad_cmd_get& e) {
 195       errss << e.what();
 196       r = -EINVAL;
 197     }
 198     return r;
 199   }
 200 };
 201
 202 void ClusterState::final_init()
 203 {
 204   AdminSocket *admin_socket = g_ceph_context->get_admin_socket();
 205   asok_hook = new ClusterSocketHook(this);
 206   int r = admin_socket->register_command(
 207     "dump_osd_network name=value,type=CephInt,req=false", asok_hook,
 208     "Dump osd heartbeat network ping times");
 209   ceph_assert(r == 0);
 210 }
 211
 212 void ClusterState::shutdown()
 213 {
 214   // unregister commands
 215   g_ceph_context->get_admin_socket()->unregister_commands(asok_hook);
 216   delete asok_hook;
 217   asok_hook = NULL;
 218 }
 219
 220 bool ClusterState::asok_command(
 221   std::string_view admin_command,
 222   const cmdmap_t& cmdmap,
 223   Formatter *f,
 224   ostream& ss)
 225 {
 226   std::lock_guard l(lock);
 227   if (admin_command == "dump_osd_network") {
 228     int64_t value = 0;
 229     // Default to health warning level if nothing specified
 230     if (!(TOPNSPC::common::cmd_getval(cmdmap, "value", value))) {
 231       // Convert milliseconds to microseconds
 232       value = static_cast<int64_t>(g_ceph_context->_conf.get_val<double>("mon_warn_on_slow_ping_time")) * 1000;
 233       if (value == 0) {
 234         double ratio = g_conf().get_val<double>("mon_warn_on_slow_ping_ratio");
 235         value = g_conf().get_val<int64_t>("osd_heartbeat_grace");
 236         value *= 1000000 * ratio; // Seconds of grace to microseconds at ratio
 237       }
 238     } else {
 239       // Convert user input to microseconds
 240       value *= 1000;
 241     }
 242     if (value < 0)
 243       value = 0;
 244
 245     struct mgr_ping_time_t {
 246       uint32_t pingtime;
 247       int from;
 248       int to;
 249       bool back;
 250       std::array<uint32_t,3> times;
 251       std::array<uint32_t,3> min;
 252       std::array<uint32_t,3> max;
 253       uint32_t last;
 254       uint32_t last_update;
 255
 256       bool operator<(const mgr_ping_time_t& rhs) const {
 257         if (pingtime < rhs.pingtime)
 258           return true;
 259         if (pingtime > rhs.pingtime)
 260           return false;
 261         if (from < rhs.from)
 262           return true;
 263         if (from > rhs.from)
 264           return false;
 265         if (to < rhs.to)
 266           return true;
 267         if (to > rhs.to)
 268           return false;
 269         return back;
 270       }
 271     };
 272
 273     set<mgr_ping_time_t> sorted;
 274     utime_t now = ceph_clock_now();
 275     for (auto i : pg_map.osd_stat) {
 276       for (auto j : i.second.hb_pingtime) {
 277
 278         if (j.second.last_update == 0)
 279           continue;
 280         auto stale_time = g_ceph_context->_conf.get_val<int64_t>("osd_mon_heartbeat_stat_stale");
 281         if (now.sec() - j.second.last_update > stale_time) {
 282           dout(20) << __func__ << " time out heartbeat for osd " << i.first
 283                    << " last_update " << j.second.last_update << dendl;
 284            continue;
 285         }
 286         mgr_ping_time_t item;
 287         item.pingtime = std::max(j.second.back_pingtime[0], j.second.back_pingtime[1]);
 288         item.pingtime = std::max(item.pingtime, j.second.back_pingtime[2]);
 289         if (!value || item.pingtime >= value) {
 290           item.from = i.first;
 291           item.to = j.first;
 292           item.times[0] = j.second.back_pingtime[0];
 293           item.times[1] = j.second.back_pingtime[1];
 294           item.times[2] = j.second.back_pingtime[2];
 295           item.min[0] = j.second.back_min[0];
 296           item.min[1] = j.second.back_min[1];
 297           item.min[2] = j.second.back_min[2];
 298           item.max[0] = j.second.back_max[0];
 299           item.max[1] = j.second.back_max[1];
 300           item.max[2] = j.second.back_max[2];
 301           item.last = j.second.back_last;
 302           item.back = true;
 303           item.last_update = j.second.last_update;
 304           sorted.emplace(item);
 305         }
 306
 307         if (j.second.front_last == 0)
 308           continue;
 309         item.pingtime = std::max(j.second.front_pingtime[0], j.second.front_pingtime[1]);
 310         item.pingtime = std::max(item.pingtime, j.second.front_pingtime[2]);
 311         if (!value || item.pingtime >= value) {
 312           item.from = i.first;
 313           item.to = j.first;
 314           item.times[0] = j.second.front_pingtime[0];
 315           item.times[1] = j.second.front_pingtime[1];
 316           item.times[2] = j.second.front_pingtime[2];
 317           item.min[0] = j.second.front_min[0];
 318           item.min[1] = j.second.front_min[1];
 319           item.min[2] = j.second.front_min[2];
 320           item.max[0] = j.second.front_max[0];
 321           item.max[1] = j.second.front_max[1];
 322           item.max[2] = j.second.front_max[2];
 323           item.last = j.second.front_last;
 324           item.back = false;
 325           item.last_update = j.second.last_update;
 326           sorted.emplace(item);
 327         }
 328       }
 329     }
 330
 331     // Network ping times (1min 5min 15min)
 332     f->open_object_section("network_ping_times");
 333     f->dump_int("threshold", value / 1000);
 334     f->open_array_section("entries");
 335     for (auto &sitem : boost::adaptors::reverse(sorted)) {
 336       ceph_assert(!value || sitem.pingtime >= value);
 337
 338       f->open_object_section("entry");
 339
 340       const time_t lu(sitem.last_update);
 341       char buffer[26];
 342       string lustr(ctime_r(&lu, buffer));
 343       lustr.pop_back();   // Remove trailing \n
 344       auto stale = g_ceph_context->_conf.get_val<int64_t>("osd_heartbeat_stale");
 345       f->dump_string("last update", lustr);
 346       f->dump_bool("stale", ceph_clock_now().sec() - sitem.last_update > stale);
 347       f->dump_int("from osd", sitem.from);
 348       f->dump_int("to osd", sitem.to);
 349       f->dump_string("interface", (sitem.back ? "back" : "front"));
 350       f->open_object_section("average");
 351       f->dump_format_unquoted("1min", "%s", fixed_u_to_string(sitem.times[0],3).c_str());
 352       f->dump_format_unquoted("5min", "%s", fixed_u_to_string(sitem.times[1],3).c_str());
 353       f->dump_format_unquoted("15min", "%s", fixed_u_to_string(sitem.times[2],3).c_str());
 354       f->close_section(); // average
 355       f->open_object_section("min");
 356       f->dump_format_unquoted("1min", "%s", fixed_u_to_string(sitem.min[0],3).c_str());
 357       f->dump_format_unquoted("5min", "%s", fixed_u_to_string(sitem.min[1],3).c_str());
 358       f->dump_format_unquoted("15min", "%s", fixed_u_to_string(sitem.min[2],3).c_str());
 359       f->close_section(); // min
 360       f->open_object_section("max");
 361       f->dump_format_unquoted("1min", "%s", fixed_u_to_string(sitem.max[0],3).c_str());
 362       f->dump_format_unquoted("5min", "%s", fixed_u_to_string(sitem.max[1],3).c_str());
 363       f->dump_format_unquoted("15min", "%s", fixed_u_to_string(sitem.max[2],3).c_str());
 364       f->close_section(); // max
 365       f->dump_format_unquoted("last", "%s", fixed_u_to_string(sitem.last,3).c_str());
 366       f->close_section(); // entry
 367     }
 368     f->close_section(); // entries
 369     f->close_section(); // network_ping_times
 370   } else {
 371     ceph_abort_msg("broken asok registration");
 372   }
 373   return true;
 374 }