ceph/src/mon/MgrMonitor.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) 2016 John Spray <john.spray@redhat.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/MMgrBeacon.h"
  15 #include "messages/MMgrMap.h"
  16 #include "messages/MMgrDigest.h"
  17
  18 #include "PGMap.h"
  19 #include "PGMonitor.h"
  20 #include "include/stringify.h"
  21 #include "mgr/MgrContext.h"
  22 #include "OSDMonitor.h"
  23
  24 #include "MgrMonitor.h"
  25
  26 #define dout_subsys ceph_subsys_mon
  27 #undef dout_prefix
  28 #define dout_prefix _prefix(_dout, mon, map)
  29 static ostream& _prefix(std::ostream *_dout, Monitor *mon,
  30                         const MgrMap& mgrmap) {
  31   return *_dout << "mon." << mon->name << "@" << mon->rank
  32                 << "(" << mon->get_state_name()
  33                 << ").mgr e" << mgrmap.get_epoch() << " ";
  34 }
  35
  36 void MgrMonitor::create_initial()
  37 {
  38 }
  39
  40 void MgrMonitor::update_from_paxos(bool *need_bootstrap)
  41 {
  42   version_t version = get_last_committed();
  43   if (version != map.epoch) {
  44     dout(4) << "loading version " << version << dendl;
  45
  46     bufferlist bl;
  47     int err = get_version(version, bl);
  48     assert(err == 0);
  49
  50     bufferlist::iterator p = bl.begin();
  51     map.decode(p);
  52
  53     dout(4) << "active server: " << map.active_addr
  54             << "(" << map.active_gid << ")" << dendl;
  55
  56     if (map.available) {
  57       first_seen_inactive = utime_t();
  58     } else {
  59       first_seen_inactive = ceph_clock_now();
  60     }
  61
  62     check_subs();
  63   }
  64
  65   // feed our pet MgrClient
  66   mon->mgr_client.ms_dispatch(new MMgrMap(map));
  67 }
  68
  69 void MgrMonitor::create_pending()
  70 {
  71   pending_map = map;
  72   pending_map.epoch++;
  73 }
  74
  75 void MgrMonitor::encode_pending(MonitorDBStore::TransactionRef t)
  76 {
  77   dout(10) << __func__ << " " << pending_map << dendl;
  78   bufferlist bl;
  79   pending_map.encode(bl, mon->get_quorum_con_features());
  80   put_version(t, pending_map.epoch, bl);
  81   put_last_committed(t, pending_map.epoch);
  82 }
  83
  84 bool MgrMonitor::check_caps(MonOpRequestRef op, const uuid_d& fsid)
  85 {
  86   // check permissions
  87   MonSession *session = op->get_session();
  88   if (!session)
  89     return false;
  90   if (!session->is_capable("mgr", MON_CAP_X)) {
  91     dout(1) << __func__ << " insufficient caps " << session->caps << dendl;
  92     return false;
  93   }
  94   if (fsid != mon->monmap->fsid) {
  95     dout(1) << __func__ << " op fsid " << fsid
  96             << " != " << mon->monmap->fsid << dendl;
  97     return false;
  98   }
  99   return true;
 100 }
 101
 102 bool MgrMonitor::preprocess_query(MonOpRequestRef op)
 103 {
 104   PaxosServiceMessage *m = static_cast<PaxosServiceMessage*>(op->get_req());
 105   switch (m->get_type()) {
 106     case MSG_MGR_BEACON:
 107       return preprocess_beacon(op);
 108     case MSG_MON_COMMAND:
 109       return preprocess_command(op);
 110     default:
 111       mon->no_reply(op);
 112       derr << "Unhandled message type " << m->get_type() << dendl;
 113       return true;
 114   }
 115 }
 116
 117 bool MgrMonitor::prepare_update(MonOpRequestRef op)
 118 {
 119   PaxosServiceMessage *m = static_cast<PaxosServiceMessage*>(op->get_req());
 120   switch (m->get_type()) {
 121     case MSG_MGR_BEACON:
 122       return prepare_beacon(op);
 123
 124     case MSG_MON_COMMAND:
 125       return prepare_command(op);
 126
 127     default:
 128       mon->no_reply(op);
 129       derr << "Unhandled message type " << m->get_type() << dendl;
 130       return true;
 131   }
 132 }
 133
 134
 135
 136 class C_Updated : public Context {
 137   MgrMonitor *mm;
 138   MonOpRequestRef op;
 139 public:
 140   C_Updated(MgrMonitor *a, MonOpRequestRef c) :
 141     mm(a), op(c) {}
 142   void finish(int r) override {
 143     if (r >= 0) {
 144       // Success
 145     } else if (r == -ECANCELED) {
 146       mm->mon->no_reply(op);
 147     } else {
 148       mm->dispatch(op);        // try again
 149     }
 150   }
 151 };
 152
 153 bool MgrMonitor::preprocess_beacon(MonOpRequestRef op)
 154 {
 155   MMgrBeacon *m = static_cast<MMgrBeacon*>(op->get_req());
 156   dout(4) << "beacon from " << m->get_gid() << dendl;
 157
 158   if (!check_caps(op, m->get_fsid())) {
 159     // drop it on the floor
 160     return true;
 161   }
 162
 163   // always send this to the leader's prepare_beacon()
 164   return false;
 165 }
 166
 167 bool MgrMonitor::prepare_beacon(MonOpRequestRef op)
 168 {
 169   MMgrBeacon *m = static_cast<MMgrBeacon*>(op->get_req());
 170   dout(4) << "beacon from " << m->get_gid() << dendl;
 171
 172   // See if we are seeing same name, new GID for the active daemon
 173   if (m->get_name() == pending_map.active_name
 174       && m->get_gid() != pending_map.active_gid)
 175   {
 176     dout(4) << "Active daemon restart (mgr." << m->get_name() << ")" << dendl;
 177     drop_active();
 178   }
 179
 180   // See if we are seeing same name, new GID for any standbys
 181   for (const auto &i : pending_map.standbys) {
 182     const StandbyInfo &s = i.second;
 183     if (s.name == m->get_name() && s.gid != m->get_gid()) {
 184       dout(4) << "Standby daemon restart (mgr." << m->get_name() << ")" << dendl;
 185       drop_standby(i.first);
 186       break;
 187     }
 188   }
 189
 190   last_beacon[m->get_gid()] = ceph_clock_now();
 191
 192   // Track whether we modified pending_map
 193   bool updated = false;
 194
 195   if (pending_map.active_gid == m->get_gid()) {
 196     // A beacon from the currently active daemon
 197     if (pending_map.active_addr != m->get_server_addr()) {
 198       dout(4) << "learned address " << m->get_server_addr()
 199               << " (was " << pending_map.active_addr << ")" << dendl;
 200       pending_map.active_addr = m->get_server_addr();
 201       updated = true;
 202     }
 203
 204     if (pending_map.get_available() != m->get_available()) {
 205       dout(4) << "available " << m->get_gid() << dendl;
 206       pending_map.available = m->get_available();
 207       updated = true;
 208     }
 209   } else if (pending_map.active_gid == 0) {
 210     // There is no currently active daemon, select this one.
 211     if (pending_map.standbys.count(m->get_gid())) {
 212       drop_standby(m->get_gid());
 213     }
 214     dout(4) << "selecting new active " << m->get_gid()
 215             << " " << m->get_name()
 216             << " (was " << pending_map.active_gid << " "
 217             << pending_map.active_name << ")" << dendl;
 218     pending_map.active_gid = m->get_gid();
 219     pending_map.active_name = m->get_name();
 220
 221     updated = true;
 222   } else {
 223     if (pending_map.standbys.count(m->get_gid()) > 0) {
 224       dout(10) << "from existing standby " << m->get_gid() << dendl;
 225     } else {
 226       dout(10) << "new standby " << m->get_gid() << dendl;
 227       pending_map.standbys[m->get_gid()] = {m->get_gid(), m->get_name()};
 228       updated = true;
 229     }
 230   }
 231
 232   if (updated) {
 233     dout(4) << "updating map" << dendl;
 234     wait_for_finished_proposal(op, new C_Updated(this, op));
 235   } else {
 236     dout(10) << "no change" << dendl;
 237   }
 238
 239   return updated;
 240 }
 241
 242 void MgrMonitor::check_subs()
 243 {
 244   const std::string type = "mgrmap";
 245   if (mon->session_map.subs.count(type) == 0)
 246     return;
 247   for (auto sub : *(mon->session_map.subs[type])) {
 248     check_sub(sub);
 249   }
 250 }
 251
 252 void MgrMonitor::check_sub(Subscription *sub)
 253 {
 254   if (sub->type == "mgrmap") {
 255     if (sub->next <= map.get_epoch()) {
 256       dout(20) << "Sending map to subscriber " << sub->session->con << dendl;
 257       sub->session->con->send_message(new MMgrMap(map));
 258       if (sub->onetime) {
 259         mon->session_map.remove_sub(sub);
 260       } else {
 261         sub->next = map.get_epoch() + 1;
 262       }
 263     }
 264   } else {
 265     assert(sub->type == "mgrdigest");
 266     if (digest_callback == nullptr) {
 267       send_digests();
 268     }
 269   }
 270 }
 271
 272 /**
 273  * Handle digest subscriptions separately (outside of check_sub) because
 274  * they are going to be periodic rather than version-driven.
 275  */
 276 void MgrMonitor::send_digests()
 277 {
 278   digest_callback = nullptr;
 279
 280   const std::string type = "mgrdigest";
 281   if (mon->session_map.subs.count(type) == 0)
 282     return;
 283
 284   for (auto sub : *(mon->session_map.subs[type])) {
 285     MMgrDigest *mdigest = new MMgrDigest;
 286
 287     JSONFormatter f;
 288     std::list<std::string> health_strs;
 289     mon->get_health(health_strs, nullptr, &f);
 290     f.flush(mdigest->health_json);
 291     f.reset();
 292
 293     std::ostringstream ss;
 294     mon->get_mon_status(&f, ss);
 295     f.flush(mdigest->mon_status_json);
 296     f.reset();
 297
 298     sub->session->con->send_message(mdigest);
 299   }
 300
 301   digest_callback = new C_MonContext(mon, [this](int){
 302       send_digests();
 303   });
 304   mon->timer.add_event_after(g_conf->mon_mgr_digest_period, digest_callback);
 305 }
 306
 307 void MgrMonitor::on_active()
 308 {
 309   if (mon->is_leader())
 310     mon->clog->info() << "mgrmap e" << map.epoch << ": " << map;
 311 }
 312
 313 void MgrMonitor::get_health(
 314   list<pair<health_status_t,string> >& summary,
 315   list<pair<health_status_t,string> > *detail,
 316   CephContext *cct) const
 317 {
 318   // start mgr warnings as soon as the mons and osds are all upgraded,
 319   // but before the require_luminous osdmap flag is set.  this way the
 320   // user gets some warning before the osd flag is set and mgr is
 321   // actually *required*.
 322   if (!mon->monmap->get_required_features().contains_all(
 323         ceph::features::mon::FEATURE_LUMINOUS) ||
 324       !HAVE_FEATURE(mon->osdmon()->osdmap.get_up_osd_features(),
 325                     SERVER_LUMINOUS)) {
 326     return;
 327   }
 328
 329   if (!map.available) {
 330     auto level = HEALTH_WARN;
 331     // do not escalate to ERR if they are still upgrading to jewel.
 332     if (mon->osdmon()->osdmap.test_flag(CEPH_OSDMAP_REQUIRE_LUMINOUS)) {
 333       utime_t now = ceph_clock_now();
 334       if (first_seen_inactive != utime_t() &&
 335           now - first_seen_inactive > g_conf->mon_mgr_inactive_grace) {
 336         level = HEALTH_ERR;
 337       }
 338     }
 339     summary.push_back(make_pair(level, "no active mgr"));
 340   }
 341 }
 342
 343 void MgrMonitor::tick()
 344 {
 345   if (!is_active() || !mon->is_leader())
 346     return;
 347
 348   const utime_t now = ceph_clock_now();
 349   utime_t cutoff = now;
 350   cutoff -= g_conf->mon_mgr_beacon_grace;
 351
 352   // Populate any missing beacons (i.e. no beacon since MgrMonitor
 353   // instantiation) with the current time, so that they will
 354   // eventually look laggy if they fail to give us a beacon.
 355   if (pending_map.active_gid != 0
 356       && last_beacon.count(pending_map.active_gid) == 0) {
 357     last_beacon[pending_map.active_gid] = now;
 358   }
 359   for (auto s : pending_map.standbys) {
 360     if (last_beacon.count(s.first) == 0) {
 361       last_beacon[s.first] = now;
 362     }
 363   }
 364
 365   // Cull standbys first so that any remaining standbys
 366   // will be eligible to take over from the active if we cull him.
 367   std::list<uint64_t> dead_standbys;
 368   for (const auto &i : pending_map.standbys) {
 369     auto last_beacon_time = last_beacon.at(i.first);
 370     if (last_beacon_time < cutoff) {
 371       dead_standbys.push_back(i.first);
 372     }
 373   }
 374
 375   bool propose = false;
 376
 377   for (auto i : dead_standbys) {
 378     dout(4) << "Dropping laggy standby " << i << dendl;
 379     drop_standby(i);
 380     propose = true;
 381   }
 382
 383   if (pending_map.active_gid != 0
 384       && last_beacon.at(pending_map.active_gid) < cutoff) {
 385
 386     drop_active();
 387     propose = true;
 388     dout(4) << "Dropping active" << pending_map.active_gid << dendl;
 389     if (promote_standby()) {
 390       dout(4) << "Promoted standby " << pending_map.active_gid << dendl;
 391     } else {
 392       dout(4) << "Active is laggy but have no standbys to replace it" << dendl;
 393     }
 394   } else if (pending_map.active_gid == 0) {
 395     if (promote_standby()) {
 396       dout(4) << "Promoted standby " << pending_map.active_gid << dendl;
 397       propose = true;
 398     }
 399   }
 400
 401   if (propose) {
 402     propose_pending();
 403   }
 404 }
 405
 406 bool MgrMonitor::promote_standby()
 407 {
 408   assert(pending_map.active_gid == 0);
 409   if (pending_map.standbys.size()) {
 410     // Promote a replacement (arbitrary choice of standby)
 411     auto replacement_gid = pending_map.standbys.begin()->first;
 412     pending_map.active_gid = replacement_gid;
 413     pending_map.active_name = pending_map.standbys.at(replacement_gid).name;
 414     pending_map.available = false;
 415     pending_map.active_addr = entity_addr_t();
 416
 417     drop_standby(replacement_gid);
 418     return true;
 419   } else {
 420     return false;
 421   }
 422 }
 423
 424 void MgrMonitor::drop_active()
 425 {
 426   if (last_beacon.count(pending_map.active_gid) > 0) {
 427     last_beacon.erase(pending_map.active_gid);
 428   }
 429
 430   pending_map.active_name = "";
 431   pending_map.active_gid = 0;
 432   pending_map.available = false;
 433   pending_map.active_addr = entity_addr_t();
 434 }
 435
 436 void MgrMonitor::drop_standby(uint64_t gid)
 437 {
 438   pending_map.standbys.erase(gid);
 439   if (last_beacon.count(gid) > 0) {
 440     last_beacon.erase(gid);
 441   }
 442
 443 }
 444
 445 bool MgrMonitor::preprocess_command(MonOpRequestRef op)
 446 {
 447   return false;
 448
 449 }
 450
 451 bool MgrMonitor::prepare_command(MonOpRequestRef op)
 452 {
 453   MMonCommand *m = static_cast<MMonCommand*>(op->get_req());
 454
 455   std::stringstream ss;
 456   bufferlist rdata;
 457
 458   std::map<std::string, cmd_vartype> cmdmap;
 459   if (!cmdmap_from_json(m->cmd, &cmdmap, ss)) {
 460     string rs = ss.str();
 461     mon->reply_command(op, -EINVAL, rs, rdata, get_last_committed());
 462     return true;
 463   }
 464
 465   MonSession *session = m->get_session();
 466   if (!session) {
 467     mon->reply_command(op, -EACCES, "access denied", rdata, get_last_committed());
 468     return true;
 469   }
 470
 471   string prefix;
 472   cmd_getval(g_ceph_context, cmdmap, "prefix", prefix);
 473
 474   int r = 0;
 475
 476   if (prefix == "mgr fail") {
 477     string who;
 478     cmd_getval(g_ceph_context, cmdmap, "who", who);
 479
 480     std::string err;
 481     uint64_t gid = strict_strtol(who.c_str(), 10, &err);
 482     bool changed = false;
 483     if (!err.empty()) {
 484       // Does not parse as a gid, treat it as a name
 485       if (pending_map.active_name == who) {
 486         drop_active();
 487         changed = true;
 488       } else {
 489         gid = 0;
 490         for (const auto &i : pending_map.standbys) {
 491           if (i.second.name == who) {
 492             gid = i.first;
 493             break;
 494           }
 495         }
 496         if (gid != 0) {
 497           drop_standby(gid);
 498           changed = true;
 499         } else {
 500           ss << "Daemon not found '" << who << "', already failed?";
 501         }
 502       }
 503     } else {
 504       if (pending_map.active_gid == gid) {
 505         drop_active();
 506         changed = true;
 507       } else if (pending_map.standbys.count(gid) > 0) {
 508         drop_standby(gid);
 509         changed = true;
 510       } else {
 511         ss << "Daemon not found '" << gid << "', already failed?";
 512       }
 513     }
 514
 515     if (changed && pending_map.active_gid == 0) {
 516       promote_standby();
 517     }
 518   } else {
 519     r = -ENOSYS;
 520   }
 521
 522   dout(4) << __func__ << " done, r=" << r << dendl;
 523   /* Compose response */
 524   string rs;
 525   getline(ss, rs);
 526
 527   if (r >= 0) {
 528     // success.. delay reply
 529     wait_for_finished_proposal(op, new Monitor::C_Command(mon, op, r, rs,
 530                                               get_last_committed() + 1));
 531     return true;
 532   } else {
 533     // reply immediately
 534     mon->reply_command(op, r, rs, rdata, get_last_committed());
 535     return false;
 536   }
 537 }
 538
 539 void MgrMonitor::init()
 540 {
 541   if (digest_callback == nullptr) {
 542     send_digests();  // To get it to schedule its own event
 543   }
 544 }
 545
 546 void MgrMonitor::on_shutdown()
 547 {
 548   if (digest_callback) {
 549     mon->timer.cancel_event(digest_callback);
 550   }
 551 }
 552