ceph/src/mds/FSMap.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) 2004-2006 Sage Weil <sage@newdream.net>
   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
  16 #include "FSMap.h"
  17
  18 #include <sstream>
  19 using std::stringstream;
  20
  21 #include "mon/health_check.h"
  22
  23
  24 void Filesystem::dump(Formatter *f) const
  25 {
  26   f->open_object_section("mdsmap");
  27   mds_map.dump(f);
  28   f->close_section();
  29   f->dump_int("id", fscid);
  30 }
  31
  32 void FSMap::dump(Formatter *f) const
  33 {
  34   f->dump_int("epoch", epoch);
  35
  36   f->open_object_section("compat");
  37   compat.dump(f);
  38   f->close_section();
  39
  40   f->open_object_section("feature_flags");
  41   f->dump_bool("enable_multiple", enable_multiple);
  42   f->dump_bool("ever_enabled_multiple", ever_enabled_multiple);
  43   f->close_section();
  44
  45   f->open_array_section("standbys");
  46   for (const auto &i : standby_daemons) {
  47     f->open_object_section("info");
  48     i.second.dump(f);
  49     f->dump_int("epoch", standby_epochs.at(i.first));
  50     f->close_section();
  51   }
  52   f->close_section();
  53
  54   f->open_array_section("filesystems");
  55   for (const auto &fs : filesystems) {
  56     f->open_object_section("filesystem");
  57     fs.second->dump(f);
  58     f->close_section();
  59   }
  60   f->close_section();
  61 }
  62
  63 void FSMap::generate_test_instances(list<FSMap*>& ls)
  64 {
  65   FSMap *m = new FSMap();
  66
  67   std::list<MDSMap*> mds_map_instances;
  68   MDSMap::generate_test_instances(mds_map_instances);
  69
  70   int k = 20;
  71   for (auto i : mds_map_instances) {
  72     auto fs = std::make_shared<Filesystem>();
  73     fs->fscid = k++;
  74     fs->mds_map = *i;
  75     delete i;
  76     m->filesystems[fs->fscid] = fs;
  77   }
  78   mds_map_instances.clear();
  79
  80   ls.push_back(m);
  81 }
  82
  83 void FSMap::print(ostream& out) const
  84 {
  85   out << "e" << epoch << std::endl;
  86   out << "enable_multiple, ever_enabled_multiple: " << enable_multiple << ","
  87       << ever_enabled_multiple << std::endl;
  88   out << "compat: " << compat << std::endl;
  89   out << "legacy client fscid: " << legacy_client_fscid << std::endl;
  90   out << " " << std::endl;
  91
  92   if (filesystems.empty()) {
  93     out << "No filesystems configured" << std::endl;
  94     return;
  95   }
  96
  97   for (const auto &fs : filesystems) {
  98     fs.second->print(out);
  99     out << " " << std::endl << " " << std::endl;  // Space out a bit
 100   }
 101
 102   if (!standby_daemons.empty()) {
 103     out << "Standby daemons:" << std::endl << " " << std::endl;
 104   }
 105
 106   for (const auto &p : standby_daemons) {
 107     p.second.print_summary(out);
 108     out << std::endl;
 109   }
 110 }
 111
 112
 113
 114 void FSMap::print_summary(Formatter *f, ostream *out) const
 115 {
 116   map<mds_role_t,string> by_rank;
 117   map<string,int> by_state;
 118
 119   if (f) {
 120     f->dump_unsigned("epoch", get_epoch());
 121     for (auto i : filesystems) {
 122       auto fs = i.second;
 123       f->dump_unsigned("id", fs->fscid);
 124       f->dump_unsigned("up", fs->mds_map.up.size());
 125       f->dump_unsigned("in", fs->mds_map.in.size());
 126       f->dump_unsigned("max", fs->mds_map.max_mds);
 127     }
 128   } else {
 129     for (auto i : filesystems) {
 130       auto fs = i.second;
 131       *out << fs->mds_map.fs_name << "-" << fs->mds_map.up.size() << "/"
 132            << fs->mds_map.in.size() << "/" << fs->mds_map.max_mds << " up ";
 133     }
 134   }
 135
 136   if (f) {
 137     f->open_array_section("by_rank");
 138   }
 139
 140   const auto all_info = get_mds_info();
 141   for (const auto &p : all_info) {
 142     const auto &info = p.second;
 143     string s = ceph_mds_state_name(info.state);
 144     if (info.laggy()) {
 145       s += "(laggy or crashed)";
 146     }
 147
 148     const fs_cluster_id_t fscid = mds_roles.at(info.global_id);
 149
 150     if (info.rank != MDS_RANK_NONE &&
 151         info.state != MDSMap::STATE_STANDBY_REPLAY) {
 152       if (f) {
 153         f->open_object_section("mds");
 154         f->dump_unsigned("filesystem_id", fscid);
 155         f->dump_unsigned("rank", info.rank);
 156         f->dump_string("name", info.name);
 157         f->dump_string("status", s);
 158         f->close_section();
 159       } else {
 160         by_rank[mds_role_t(fscid, info.rank)] = info.name + "=" + s;
 161       }
 162     } else {
 163       by_state[s]++;
 164     }
 165   }
 166
 167   if (f) {
 168     f->close_section();
 169   } else {
 170     if (!by_rank.empty()) {
 171       if (filesystems.size() > 1) {
 172         // Disambiguate filesystems
 173         std::map<std::string, std::string> pretty;
 174         for (auto i : by_rank) {
 175           const auto &fs_name = filesystems.at(i.first.fscid)->mds_map.fs_name;
 176           std::ostringstream o;
 177           o << "[" << fs_name << ":" << i.first.rank << "]";
 178           pretty[o.str()] = i.second;
 179         }
 180         *out << " " << pretty;
 181       } else {
 182         // Omit FSCID in output when only one filesystem exists
 183         std::map<mds_rank_t, std::string> shortened;
 184         for (auto i : by_rank) {
 185           shortened[i.first.rank] = i.second;
 186         }
 187         *out << " " << shortened;
 188       }
 189     }
 190   }
 191
 192   for (map<string,int>::reverse_iterator p = by_state.rbegin(); p != by_state.rend(); ++p) {
 193     if (f) {
 194       f->dump_unsigned(p->first.c_str(), p->second);
 195     } else {
 196       *out << ", " << p->second << " " << p->first;
 197     }
 198   }
 199
 200   size_t failed = 0;
 201   size_t damaged = 0;
 202   for (auto i : filesystems) {
 203     auto fs = i.second;
 204     failed += fs->mds_map.failed.size();
 205     damaged += fs->mds_map.damaged.size();
 206   }
 207
 208   if (failed > 0) {
 209     if (f) {
 210       f->dump_unsigned("failed", failed);
 211     } else {
 212       *out << ", " << failed << " failed";
 213     }
 214   }
 215
 216   if (damaged > 0) {
 217     if (f) {
 218       f->dump_unsigned("damaged", damaged);
 219     } else {
 220       *out << ", " << damaged << " damaged";
 221     }
 222   }
 223   //if (stopped.size())
 224   //out << ", " << stopped.size() << " stopped";
 225 }
 226
 227
 228 void FSMap::create_filesystem(const std::string &name,
 229                               int64_t metadata_pool, int64_t data_pool,
 230                               uint64_t features)
 231 {
 232   auto fs = std::make_shared<Filesystem>();
 233   fs->mds_map.fs_name = name;
 234   fs->mds_map.max_mds = 1;
 235   fs->mds_map.data_pools.push_back(data_pool);
 236   fs->mds_map.metadata_pool = metadata_pool;
 237   fs->mds_map.cas_pool = -1;
 238   fs->mds_map.max_file_size = g_conf->mds_max_file_size;
 239   fs->mds_map.compat = compat;
 240   fs->mds_map.created = ceph_clock_now();
 241   fs->mds_map.modified = ceph_clock_now();
 242   fs->mds_map.session_timeout = g_conf->mds_session_timeout;
 243   fs->mds_map.session_autoclose = g_conf->mds_session_autoclose;
 244   fs->mds_map.enabled = true;
 245   if (features & CEPH_FEATURE_SERVER_JEWEL) {
 246     fs->fscid = next_filesystem_id++;
 247     // ANONYMOUS is only for upgrades from legacy mdsmaps, we should
 248     // have initialized next_filesystem_id such that it's never used here.
 249     assert(fs->fscid != FS_CLUSTER_ID_ANONYMOUS);
 250   } else {
 251     // Use anon fscid because this will get thrown away when encoding
 252     // as legacy MDSMap for legacy mons.
 253     assert(filesystems.empty());
 254     fs->fscid = FS_CLUSTER_ID_ANONYMOUS;
 255   }
 256   filesystems[fs->fscid] = fs;
 257
 258   // Created first filesystem?  Set it as the one
 259   // for legacy clients to use
 260   if (filesystems.size() == 1) {
 261     legacy_client_fscid = fs->fscid;
 262   }
 263 }
 264
 265 void FSMap::reset_filesystem(fs_cluster_id_t fscid)
 266 {
 267   auto fs = get_filesystem(fscid);
 268   auto new_fs = std::make_shared<Filesystem>();
 269
 270   // Populate rank 0 as existing (so don't go into CREATING)
 271   // but failed (so that next available MDS is assigned the rank)
 272   new_fs->mds_map.in.insert(mds_rank_t(0));
 273   new_fs->mds_map.failed.insert(mds_rank_t(0));
 274
 275   // Carry forward what makes sense
 276   new_fs->fscid = fs->fscid;
 277   new_fs->mds_map.inline_data_enabled = fs->mds_map.inline_data_enabled;
 278   new_fs->mds_map.max_mds = 1;
 279   new_fs->mds_map.data_pools = fs->mds_map.data_pools;
 280   new_fs->mds_map.metadata_pool = fs->mds_map.metadata_pool;
 281   new_fs->mds_map.cas_pool = fs->mds_map.cas_pool;
 282   new_fs->mds_map.fs_name = fs->mds_map.fs_name;
 283   new_fs->mds_map.max_file_size = g_conf->mds_max_file_size;
 284   new_fs->mds_map.compat = compat;
 285   new_fs->mds_map.created = ceph_clock_now();
 286   new_fs->mds_map.modified = ceph_clock_now();
 287   new_fs->mds_map.session_timeout = g_conf->mds_session_timeout;
 288   new_fs->mds_map.session_autoclose = g_conf->mds_session_autoclose;
 289   new_fs->mds_map.standby_count_wanted = fs->mds_map.standby_count_wanted;
 290   new_fs->mds_map.enabled = true;
 291
 292   // Remember mds ranks that have ever started. (They should load old inotable
 293   // instead of creating new one if they start again.)
 294   new_fs->mds_map.stopped.insert(fs->mds_map.in.begin(), fs->mds_map.in.end());
 295   new_fs->mds_map.stopped.insert(fs->mds_map.stopped.begin(), fs->mds_map.stopped.end());
 296   new_fs->mds_map.stopped.erase(mds_rank_t(0));
 297
 298   // Persist the new FSMap
 299   filesystems[new_fs->fscid] = new_fs;
 300 }
 301
 302 void FSMap::get_health(list<pair<health_status_t,string> >& summary,
 303                         list<pair<health_status_t,string> > *detail) const
 304 {
 305   mds_rank_t standby_count_wanted = 0;
 306   for (const auto &i : filesystems) {
 307     const auto &fs = i.second;
 308
 309     // TODO: move get_health up into here so that we can qualify
 310     // all the messages with what filesystem they're talking about
 311     fs->mds_map.get_health(summary, detail);
 312
 313     standby_count_wanted = std::max(standby_count_wanted, fs->mds_map.get_standby_count_wanted((mds_rank_t)standby_daemons.size()));
 314   }
 315
 316   if (standby_count_wanted) {
 317     std::ostringstream oss;
 318     oss << "insufficient standby daemons available: have " << standby_daemons.size() << "; want " << standby_count_wanted << " more";
 319     summary.push_back(make_pair(HEALTH_WARN, oss.str()));
 320   }
 321 }
 322
 323 bool FSMap::check_health(void)
 324 {
 325   bool changed = false;
 326   for (auto &i : filesystems) {
 327     changed |= i.second->mds_map.check_health((mds_rank_t)standby_daemons.size());
 328   }
 329   return changed;
 330 }
 331
 332 void FSMap::get_health_checks(health_check_map_t *checks) const
 333 {
 334   mds_rank_t standby_count_wanted = 0;
 335   for (const auto &i : filesystems) {
 336     const auto &fs = i.second;
 337     health_check_map_t fschecks;
 338
 339     fs->mds_map.get_health_checks(&fschecks);
 340
 341     // Some of the failed ranks might be transient (i.e. there are standbys
 342     // ready to replace them).  We will report only on "stuck" failed, i.e.
 343     // ranks which are failed and have no standby replacement available.
 344     std::set<mds_rank_t> stuck_failed;
 345
 346     for (const auto &rank : fs->mds_map.failed) {
 347       const mds_gid_t replacement = find_replacement_for(
 348           {fs->fscid, rank}, {}, g_conf->mon_force_standby_active);
 349       if (replacement == MDS_GID_NONE) {
 350         stuck_failed.insert(rank);
 351       }
 352     }
 353
 354     // FS_WITH_FAILED_MDS
 355     if (!stuck_failed.empty()) {
 356       health_check_t& fscheck = checks->get_or_add(
 357         "FS_WITH_FAILED_MDS", HEALTH_WARN,
 358         "%num% filesystem%plurals% %isorare% have a failed mds daemon");
 359       ostringstream ss;
 360       ss << "fs " << fs->mds_map.fs_name << " has " << stuck_failed.size()
 361          << " failed mds" << (stuck_failed.size() > 1 ? "s" : "");
 362       fscheck.detail.push_back(ss.str()); }
 363
 364     checks->merge(fschecks);
 365     standby_count_wanted = std::max(
 366       standby_count_wanted,
 367       fs->mds_map.get_standby_count_wanted((mds_rank_t)standby_daemons.size()));
 368   }
 369
 370   // MDS_INSUFFICIENT_STANDBY
 371   if (standby_count_wanted) {
 372     std::ostringstream oss, dss;
 373     oss << "insufficient standby MDS daemons available";
 374     auto& d = checks->get_or_add("MDS_INSUFFICIENT_STANDBY", HEALTH_WARN, oss.str());
 375     dss << "have " << standby_daemons.size() << "; want " << standby_count_wanted
 376         << " more";
 377     d.detail.push_back(dss.str());
 378   }
 379 }
 380
 381 void FSMap::encode(bufferlist& bl, uint64_t features) const
 382 {
 383   if (features & CEPH_FEATURE_SERVER_JEWEL) {
 384     ENCODE_START(7, 6, bl);
 385     ::encode(epoch, bl);
 386     ::encode(next_filesystem_id, bl);
 387     ::encode(legacy_client_fscid, bl);
 388     ::encode(compat, bl);
 389     ::encode(enable_multiple, bl);
 390     std::vector<Filesystem> fs_list;
 391     for (auto i : filesystems) {
 392       fs_list.push_back(*(i.second));
 393     }
 394     ::encode(fs_list, bl, features);
 395     ::encode(mds_roles, bl);
 396     ::encode(standby_daemons, bl, features);
 397     ::encode(standby_epochs, bl);
 398     ::encode(ever_enabled_multiple, bl);
 399     ENCODE_FINISH(bl);
 400   } else {
 401     if (filesystems.empty()) {
 402       MDSMap disabled_map;
 403       disabled_map.epoch = epoch;
 404       disabled_map.encode(bl, features);
 405     } else {
 406       // MDSMonitor should never have created multiple filesystems
 407       // until the quorum features indicated Jewel
 408       assert(filesystems.size() == 1);
 409       auto fs = filesystems.begin()->second;
 410
 411       // Take the MDSMap for the enabled filesystem, and populated its
 412       // mds_info with the standbys to get a pre-jewel-style mon MDSMap.
 413       MDSMap full_mdsmap = fs->mds_map;
 414       full_mdsmap.epoch = epoch;
 415       for (const auto &p : standby_daemons) {
 416         full_mdsmap.mds_info[p.first] = p.second;
 417       }
 418
 419       // Old MDSMaps don't set rank on standby replay daemons
 420       for (auto &i : full_mdsmap.mds_info) {
 421         auto &info = i.second;
 422         if (info.state == MDSMap::STATE_STANDBY_REPLAY) {
 423           info.rank = MDS_RANK_NONE;
 424         }
 425       }
 426
 427       full_mdsmap.encode(bl, features);
 428     }
 429   }
 430 }
 431
 432 void FSMap::decode(bufferlist::iterator& p)
 433 {
 434   // Because the mon used to store an MDSMap where we now
 435   // store an FSMap, FSMap knows how to decode the legacy
 436   // MDSMap format (it never needs to encode it though).
 437   MDSMap legacy_mds_map;
 438
 439   // The highest MDSMap encoding version before we changed the
 440   // MDSMonitor to store an FSMap instead of an MDSMap was
 441   // 5, so anything older than 6 is decoded as an MDSMap,
 442   // and anything newer is decoded as an FSMap.
 443   DECODE_START_LEGACY_COMPAT_LEN_16(7, 4, 4, p);
 444   if (struct_v < 6) {
 445     // Decoding an MDSMap (upgrade)
 446     ::decode(epoch, p);
 447     ::decode(legacy_mds_map.flags, p);
 448     ::decode(legacy_mds_map.last_failure, p);
 449     ::decode(legacy_mds_map.root, p);
 450     ::decode(legacy_mds_map.session_timeout, p);
 451     ::decode(legacy_mds_map.session_autoclose, p);
 452     ::decode(legacy_mds_map.max_file_size, p);
 453     ::decode(legacy_mds_map.max_mds, p);
 454     ::decode(legacy_mds_map.mds_info, p);
 455     if (struct_v < 3) {
 456       __u32 n;
 457       ::decode(n, p);
 458       while (n--) {
 459         __u32 m;
 460         ::decode(m, p);
 461         legacy_mds_map.data_pools.push_back(m);
 462       }
 463       __s32 s;
 464       ::decode(s, p);
 465       legacy_mds_map.cas_pool = s;
 466     } else {
 467       ::decode(legacy_mds_map.data_pools, p);
 468       ::decode(legacy_mds_map.cas_pool, p);
 469     }
 470
 471     // kclient ignores everything from here
 472     __u16 ev = 1;
 473     if (struct_v >= 2)
 474       ::decode(ev, p);
 475     if (ev >= 3)
 476       ::decode(legacy_mds_map.compat, p);
 477     else
 478       legacy_mds_map.compat = get_mdsmap_compat_set_base();
 479     if (ev < 5) {
 480       __u32 n;
 481       ::decode(n, p);
 482       legacy_mds_map.metadata_pool = n;
 483     } else {
 484       ::decode(legacy_mds_map.metadata_pool, p);
 485     }
 486     ::decode(legacy_mds_map.created, p);
 487     ::decode(legacy_mds_map.modified, p);
 488     ::decode(legacy_mds_map.tableserver, p);
 489     ::decode(legacy_mds_map.in, p);
 490     std::map<mds_rank_t,int32_t> inc;  // Legacy field, parse and drop
 491     ::decode(inc, p);
 492     ::decode(legacy_mds_map.up, p);
 493     ::decode(legacy_mds_map.failed, p);
 494     ::decode(legacy_mds_map.stopped, p);
 495     if (ev >= 4)
 496       ::decode(legacy_mds_map.last_failure_osd_epoch, p);
 497     if (ev >= 6) {
 498       if (ev < 10) {
 499         // previously this was a bool about snaps, not a flag map
 500         bool flag;
 501         ::decode(flag, p);
 502         legacy_mds_map.ever_allowed_features = flag ?
 503           CEPH_MDSMAP_ALLOW_SNAPS : 0;
 504         ::decode(flag, p);
 505         legacy_mds_map.explicitly_allowed_features = flag ?
 506           CEPH_MDSMAP_ALLOW_SNAPS : 0;
 507         if (legacy_mds_map.max_mds > 1) {
 508           legacy_mds_map.set_multimds_allowed();
 509         }
 510       } else {
 511         ::decode(legacy_mds_map.ever_allowed_features, p);
 512         ::decode(legacy_mds_map.explicitly_allowed_features, p);
 513       }
 514     } else {
 515       legacy_mds_map.ever_allowed_features = CEPH_MDSMAP_ALLOW_CLASSICS;
 516       legacy_mds_map.explicitly_allowed_features = 0;
 517       if (legacy_mds_map.max_mds > 1) {
 518         legacy_mds_map.set_multimds_allowed();
 519       }
 520     }
 521     if (ev >= 7)
 522       ::decode(legacy_mds_map.inline_data_enabled, p);
 523
 524     if (ev >= 8) {
 525       assert(struct_v >= 5);
 526       ::decode(legacy_mds_map.enabled, p);
 527       ::decode(legacy_mds_map.fs_name, p);
 528     } else {
 529       legacy_mds_map.fs_name = "default";
 530       if (epoch > 1) {
 531         // If an MDS has ever been started, epoch will be greater than 1,
 532         // assume filesystem is enabled.
 533         legacy_mds_map.enabled = true;
 534       } else {
 535         // Upgrading from a cluster that never used an MDS, switch off
 536         // filesystem until it's explicitly enabled.
 537         legacy_mds_map.enabled = false;
 538       }
 539     }
 540
 541     if (ev >= 9) {
 542       ::decode(legacy_mds_map.damaged, p);
 543     }
 544
 545     // We're upgrading, populate filesystems from the legacy fields
 546     filesystems.clear();
 547     standby_daemons.clear();
 548     standby_epochs.clear();
 549     mds_roles.clear();
 550     compat = legacy_mds_map.compat;
 551     enable_multiple = false;
 552
 553     // Synthesise a Filesystem from legacy_mds_map, if enabled
 554     if (legacy_mds_map.enabled) {
 555       // Construct a Filesystem from the legacy MDSMap
 556       auto migrate_fs = std::make_shared<Filesystem>();
 557       migrate_fs->fscid = FS_CLUSTER_ID_ANONYMOUS;
 558       migrate_fs->mds_map = legacy_mds_map;
 559       migrate_fs->mds_map.epoch = epoch;
 560       filesystems[migrate_fs->fscid] = migrate_fs;
 561
 562       // List of GIDs that had invalid states
 563       std::set<mds_gid_t> drop_gids;
 564
 565       // Construct mds_roles, standby_daemons, and remove
 566       // standbys from the MDSMap in the Filesystem.
 567       for (auto &p : migrate_fs->mds_map.mds_info) {
 568         if (p.second.state == MDSMap::STATE_STANDBY_REPLAY) {
 569           // In legacy MDSMap, standby replay daemons don't have
 570           // rank set, but since FSMap they do.
 571           p.second.rank = p.second.standby_for_rank;
 572         }
 573         if (p.second.rank == MDS_RANK_NONE) {
 574           if (p.second.state != MDSMap::STATE_STANDBY) {
 575             // Old MDSMaps can have down:dne here, which
 576             // is invalid in an FSMap (#17837)
 577             drop_gids.insert(p.first);
 578           } else {
 579             insert(p.second); // into standby_daemons
 580           }
 581         } else {
 582           mds_roles[p.first] = migrate_fs->fscid;
 583         }
 584       }
 585       for (const auto &p : standby_daemons) {
 586         // Erase from this Filesystem's MDSMap, because it has
 587         // been copied into FSMap::Standby_daemons above
 588         migrate_fs->mds_map.mds_info.erase(p.first);
 589       }
 590       for (const auto &gid : drop_gids) {
 591         // Throw away all info for this MDS because it was identified
 592         // as having invalid state above.
 593         migrate_fs->mds_map.mds_info.erase(gid);
 594       }
 595
 596       legacy_client_fscid = migrate_fs->fscid;
 597     } else {
 598       legacy_client_fscid = FS_CLUSTER_ID_NONE;
 599     }
 600   } else {
 601     ::decode(epoch, p);
 602     ::decode(next_filesystem_id, p);
 603     ::decode(legacy_client_fscid, p);
 604     ::decode(compat, p);
 605     ::decode(enable_multiple, p);
 606     std::vector<Filesystem> fs_list;
 607     ::decode(fs_list, p);
 608     filesystems.clear();
 609     for (std::vector<Filesystem>::const_iterator fs = fs_list.begin(); fs != fs_list.end(); ++fs) {
 610       filesystems[fs->fscid] = std::make_shared<Filesystem>(*fs);
 611     }
 612
 613     ::decode(mds_roles, p);
 614     ::decode(standby_daemons, p);
 615     ::decode(standby_epochs, p);
 616     if (struct_v >= 7) {
 617       ::decode(ever_enabled_multiple, p);
 618     }
 619   }
 620
 621   DECODE_FINISH(p);
 622 }
 623
 624
 625 void Filesystem::encode(bufferlist& bl, uint64_t features) const
 626 {
 627   ENCODE_START(1, 1, bl);
 628   ::encode(fscid, bl);
 629   bufferlist mdsmap_bl;
 630   mds_map.encode(mdsmap_bl, features);
 631   ::encode(mdsmap_bl, bl);
 632   ENCODE_FINISH(bl);
 633 }
 634
 635 void Filesystem::decode(bufferlist::iterator& p)
 636 {
 637   DECODE_START(1, p);
 638   ::decode(fscid, p);
 639   bufferlist mdsmap_bl;
 640   ::decode(mdsmap_bl, p);
 641   bufferlist::iterator mdsmap_bl_iter = mdsmap_bl.begin();
 642   mds_map.decode(mdsmap_bl_iter);
 643   DECODE_FINISH(p);
 644 }
 645
 646 int FSMap::parse_filesystem(
 647       std::string const &ns_str,
 648       std::shared_ptr<const Filesystem> *result
 649       ) const
 650 {
 651   std::string ns_err;
 652   fs_cluster_id_t fscid = strict_strtol(ns_str.c_str(), 10, &ns_err);
 653   if (!ns_err.empty() || filesystems.count(fscid) == 0) {
 654     for (auto &fs : filesystems) {
 655       if (fs.second->mds_map.fs_name == ns_str) {
 656         *result = std::const_pointer_cast<const Filesystem>(fs.second);
 657         return 0;
 658       }
 659     }
 660     return -ENOENT;
 661   } else {
 662     *result = get_filesystem(fscid);
 663     return 0;
 664   }
 665 }
 666
 667 void Filesystem::print(std::ostream &out) const
 668 {
 669   out << "Filesystem '" << mds_map.fs_name
 670       << "' (" << fscid << ")" << std::endl;
 671   mds_map.print(out);
 672 }
 673
 674 mds_gid_t FSMap::find_standby_for(mds_role_t role, const std::string& name) const
 675 {
 676   mds_gid_t result = MDS_GID_NONE;
 677
 678   // First see if we have a STANDBY_REPLAY
 679   auto fs = get_filesystem(role.fscid);
 680   for (const auto &i : fs->mds_map.mds_info) {
 681     const auto &info = i.second;
 682     if (info.rank == role.rank && info.state == MDSMap::STATE_STANDBY_REPLAY) {
 683       return info.global_id;
 684     }
 685   }
 686
 687   // See if there are any STANDBY daemons available
 688   for (const auto &i : standby_daemons) {
 689     const auto &gid = i.first;
 690     const auto &info = i.second;
 691     assert(info.state == MDSMap::STATE_STANDBY);
 692     assert(info.rank == MDS_RANK_NONE);
 693
 694     if (info.laggy()) {
 695       continue;
 696     }
 697
 698     // The mds_info_t may or may not tell us exactly which filesystem
 699     // the standby_for_rank refers to: lookup via legacy_client_fscid
 700     mds_role_t target_role = {
 701       info.standby_for_fscid == FS_CLUSTER_ID_NONE ?
 702         legacy_client_fscid : info.standby_for_fscid,
 703       info.standby_for_rank};
 704
 705     if ((target_role.rank == role.rank && target_role.fscid == role.fscid)
 706         || (name.length() && info.standby_for_name == name)) {
 707       // It's a named standby for *me*, use it.
 708       return gid;
 709     } else if (
 710         info.standby_for_rank < 0 && info.standby_for_name.length() == 0 &&
 711         (info.standby_for_fscid == FS_CLUSTER_ID_NONE ||
 712          info.standby_for_fscid == role.fscid)) {
 713         // It's not a named standby for anyone, use it if we don't find
 714         // a named standby for me later, unless it targets another FSCID.
 715         result = gid;
 716       }
 717   }
 718
 719   return result;
 720 }
 721
 722 mds_gid_t FSMap::find_unused_for(mds_role_t role,
 723                                  bool force_standby_active) const {
 724   for (const auto &i : standby_daemons) {
 725     const auto &gid = i.first;
 726     const auto &info = i.second;
 727     assert(info.state == MDSMap::STATE_STANDBY);
 728
 729     if (info.laggy() || info.rank >= 0)
 730       continue;
 731
 732     if (info.standby_for_fscid != FS_CLUSTER_ID_NONE &&
 733         info.standby_for_fscid != role.fscid)
 734       continue;
 735     if (info.standby_for_rank != MDS_RANK_NONE &&
 736         info.standby_for_rank != role.rank)
 737       continue;
 738
 739     // To be considered 'unused' a daemon must either not
 740     // be selected for standby-replay or the force_standby_active
 741     // setting must be enabled to use replay daemons anyway.
 742     if (!info.standby_replay || force_standby_active) {
 743       return gid;
 744     }
 745   }
 746   return MDS_GID_NONE;
 747 }
 748
 749 mds_gid_t FSMap::find_replacement_for(mds_role_t role, const std::string& name,
 750                                bool force_standby_active) const {
 751   const mds_gid_t standby = find_standby_for(role, name);
 752   if (standby)
 753     return standby;
 754   else
 755     return find_unused_for(role, force_standby_active);
 756 }
 757
 758 void FSMap::sanity() const
 759 {
 760   if (legacy_client_fscid != FS_CLUSTER_ID_NONE) {
 761     assert(filesystems.count(legacy_client_fscid) == 1);
 762   }
 763
 764   for (const auto &i : filesystems) {
 765     auto fs = i.second;
 766     assert(fs->mds_map.compat.compare(compat) == 0);
 767     assert(fs->fscid == i.first);
 768     for (const auto &j : fs->mds_map.mds_info) {
 769       assert(j.second.rank != MDS_RANK_NONE);
 770       assert(mds_roles.count(j.first) == 1);
 771       assert(standby_daemons.count(j.first) == 0);
 772       assert(standby_epochs.count(j.first) == 0);
 773       assert(mds_roles.at(j.first) == i.first);
 774       if (j.second.state != MDSMap::STATE_STANDBY_REPLAY) {
 775         assert(fs->mds_map.up.at(j.second.rank) == j.first);
 776         assert(fs->mds_map.failed.count(j.second.rank) == 0);
 777         assert(fs->mds_map.damaged.count(j.second.rank) == 0);
 778       }
 779     }
 780
 781     for (const auto &j : fs->mds_map.up) {
 782       mds_rank_t rank = j.first;
 783       assert(fs->mds_map.in.count(rank) == 1);
 784       mds_gid_t gid = j.second;
 785       assert(fs->mds_map.mds_info.count(gid) == 1);
 786     }
 787   }
 788
 789   for (const auto &i : standby_daemons) {
 790     assert(i.second.state == MDSMap::STATE_STANDBY);
 791     assert(i.second.rank == MDS_RANK_NONE);
 792     assert(i.second.global_id == i.first);
 793     assert(standby_epochs.count(i.first) == 1);
 794     assert(mds_roles.count(i.first) == 1);
 795     assert(mds_roles.at(i.first) == FS_CLUSTER_ID_NONE);
 796   }
 797
 798   for (const auto &i : standby_epochs) {
 799     assert(standby_daemons.count(i.first) == 1);
 800   }
 801
 802   for (const auto &i : mds_roles) {
 803     if (i.second == FS_CLUSTER_ID_NONE) {
 804       assert(standby_daemons.count(i.first) == 1);
 805     } else {
 806       assert(filesystems.count(i.second) == 1);
 807       assert(filesystems.at(i.second)->mds_map.mds_info.count(i.first) == 1);
 808     }
 809   }
 810 }
 811
 812 void FSMap::promote(
 813     mds_gid_t standby_gid,
 814     const std::shared_ptr<Filesystem> &filesystem,
 815     mds_rank_t assigned_rank)
 816 {
 817   assert(gid_exists(standby_gid));
 818   bool is_standby_replay = mds_roles.at(standby_gid) != FS_CLUSTER_ID_NONE;
 819   if (!is_standby_replay) {
 820     assert(standby_daemons.count(standby_gid));
 821     assert(standby_daemons.at(standby_gid).state == MDSMap::STATE_STANDBY);
 822   }
 823
 824   MDSMap &mds_map = filesystem->mds_map;
 825
 826   // Insert daemon state to Filesystem
 827   if (!is_standby_replay) {
 828     mds_map.mds_info[standby_gid] = standby_daemons.at(standby_gid);
 829   } else {
 830     assert(mds_map.mds_info.count(standby_gid));
 831     assert(mds_map.mds_info.at(standby_gid).state == MDSMap::STATE_STANDBY_REPLAY);
 832     assert(mds_map.mds_info.at(standby_gid).rank == assigned_rank);
 833   }
 834   MDSMap::mds_info_t &info = mds_map.mds_info[standby_gid];
 835
 836   if (mds_map.stopped.erase(assigned_rank)) {
 837     // The cluster is being expanded with a stopped rank
 838     info.state = MDSMap::STATE_STARTING;
 839   } else if (!mds_map.is_in(assigned_rank)) {
 840     // The cluster is being expanded with a new rank
 841     info.state = MDSMap::STATE_CREATING;
 842   } else {
 843     // An existing rank is being assigned to a replacement
 844     info.state = MDSMap::STATE_REPLAY;
 845     mds_map.failed.erase(assigned_rank);
 846   }
 847   info.rank = assigned_rank;
 848   info.inc = epoch;
 849   mds_roles[standby_gid] = filesystem->fscid;
 850
 851   // Update the rank state in Filesystem
 852   mds_map.in.insert(assigned_rank);
 853   mds_map.up[assigned_rank] = standby_gid;
 854
 855   // Remove from the list of standbys
 856   if (!is_standby_replay) {
 857     standby_daemons.erase(standby_gid);
 858     standby_epochs.erase(standby_gid);
 859   }
 860
 861   // Indicate that Filesystem has been modified
 862   mds_map.epoch = epoch;
 863 }
 864
 865 void FSMap::assign_standby_replay(
 866     const mds_gid_t standby_gid,
 867     const fs_cluster_id_t leader_ns,
 868     const mds_rank_t leader_rank)
 869 {
 870   assert(mds_roles.at(standby_gid) == FS_CLUSTER_ID_NONE);
 871   assert(gid_exists(standby_gid));
 872   assert(!gid_has_rank(standby_gid));
 873   assert(standby_daemons.count(standby_gid));
 874
 875   // Insert to the filesystem
 876   auto fs = filesystems.at(leader_ns);
 877   fs->mds_map.mds_info[standby_gid] = standby_daemons.at(standby_gid);
 878   fs->mds_map.mds_info[standby_gid].rank = leader_rank;
 879   fs->mds_map.mds_info[standby_gid].state = MDSMap::STATE_STANDBY_REPLAY;
 880   mds_roles[standby_gid] = leader_ns;
 881
 882   // Remove from the list of standbys
 883   standby_daemons.erase(standby_gid);
 884   standby_epochs.erase(standby_gid);
 885
 886   // Indicate that Filesystem has been modified
 887   fs->mds_map.epoch = epoch;
 888 }
 889
 890 void FSMap::erase(mds_gid_t who, epoch_t blacklist_epoch)
 891 {
 892   if (mds_roles.at(who) == FS_CLUSTER_ID_NONE) {
 893     standby_daemons.erase(who);
 894     standby_epochs.erase(who);
 895   } else {
 896     auto &fs = filesystems.at(mds_roles.at(who));
 897     const auto &info = fs->mds_map.mds_info.at(who);
 898     if (info.state != MDSMap::STATE_STANDBY_REPLAY) {
 899       if (info.state == MDSMap::STATE_CREATING) {
 900         // If this gid didn't make it past CREATING, then forget
 901         // the rank ever existed so that next time it's handed out
 902         // to a gid it'll go back into CREATING.
 903         fs->mds_map.in.erase(info.rank);
 904       } else {
 905         // Put this rank into the failed list so that the next available
 906         // STANDBY will pick it up.
 907         fs->mds_map.failed.insert(info.rank);
 908       }
 909       assert(fs->mds_map.up.at(info.rank) == info.global_id);
 910       fs->mds_map.up.erase(info.rank);
 911     }
 912     fs->mds_map.mds_info.erase(who);
 913     fs->mds_map.last_failure_osd_epoch = blacklist_epoch;
 914     fs->mds_map.epoch = epoch;
 915   }
 916
 917   mds_roles.erase(who);
 918 }
 919
 920 void FSMap::damaged(mds_gid_t who, epoch_t blacklist_epoch)
 921 {
 922   assert(mds_roles.at(who) != FS_CLUSTER_ID_NONE);
 923   auto fs = filesystems.at(mds_roles.at(who));
 924   mds_rank_t rank = fs->mds_map.mds_info[who].rank;
 925
 926   erase(who, blacklist_epoch);
 927   fs->mds_map.failed.erase(rank);
 928   fs->mds_map.damaged.insert(rank);
 929
 930   assert(fs->mds_map.epoch == epoch);
 931 }
 932
 933 /**
 934  * Update to indicate that the rank `rank` is to be removed
 935  * from the damaged list of the filesystem `fscid`
 936  */
 937 bool FSMap::undamaged(const fs_cluster_id_t fscid, const mds_rank_t rank)
 938 {
 939   auto fs = filesystems.at(fscid);
 940
 941   if (fs->mds_map.damaged.erase(rank)) {
 942     fs->mds_map.failed.insert(rank);
 943     fs->mds_map.epoch = epoch;
 944     return true;
 945   } else {
 946     return false;
 947   }
 948 }
 949
 950 void FSMap::insert(const MDSMap::mds_info_t &new_info)
 951 {
 952   assert(new_info.state == MDSMap::STATE_STANDBY);
 953   assert(new_info.rank == MDS_RANK_NONE);
 954   mds_roles[new_info.global_id] = FS_CLUSTER_ID_NONE;
 955   standby_daemons[new_info.global_id] = new_info;
 956   standby_epochs[new_info.global_id] = epoch;
 957 }
 958
 959 std::list<mds_gid_t> FSMap::stop(mds_gid_t who)
 960 {
 961   assert(mds_roles.at(who) != FS_CLUSTER_ID_NONE);
 962   auto fs = filesystems.at(mds_roles.at(who));
 963   const auto &info = fs->mds_map.mds_info.at(who);
 964   fs->mds_map.up.erase(info.rank);
 965   fs->mds_map.in.erase(info.rank);
 966   fs->mds_map.stopped.insert(info.rank);
 967
 968   // Also drop any standby replays that were following this rank
 969   std::list<mds_gid_t> standbys;
 970   for (const auto &i : fs->mds_map.mds_info) {
 971     const auto &other_gid = i.first;
 972     const auto &other_info = i.second;
 973     if (other_info.rank == info.rank
 974         && other_info.state == MDSMap::STATE_STANDBY_REPLAY) {
 975       standbys.push_back(other_gid);
 976       erase(other_gid, 0);
 977     }
 978   }
 979
 980   fs->mds_map.mds_info.erase(who);
 981   mds_roles.erase(who);
 982
 983   fs->mds_map.epoch = epoch;
 984
 985   return standbys;
 986 }
 987
 988
 989 /**
 990  * Given one of the following forms:
 991  *   <fs name>:<rank>
 992  *   <fs id>:<rank>
 993  *   <rank>
 994  *
 995  * Parse into a mds_role_t.  The rank-only form is only valid
 996  * if legacy_client_ns is set.
 997  */
 998 int FSMap::parse_role(
 999     const std::string &role_str,
1000     mds_role_t *role,
1001     std::ostream &ss) const
1002 {
1003   size_t colon_pos = role_str.find(":");
1004   size_t rank_pos;
1005   std::shared_ptr<const Filesystem> fs;
1006   if (colon_pos == std::string::npos) {
1007     if (legacy_client_fscid == FS_CLUSTER_ID_NONE) {
1008       ss << "No filesystem selected";
1009       return -ENOENT;
1010     }
1011     fs = get_filesystem(legacy_client_fscid);
1012     rank_pos = 0;
1013   } else {
1014     if (parse_filesystem(role_str.substr(0, colon_pos), &fs) < 0) {
1015       ss << "Invalid filesystem";
1016       return -ENOENT;
1017     }
1018     rank_pos = colon_pos+1;
1019   }
1020
1021   mds_rank_t rank;
1022   std::string err;
1023   std::string rank_str = role_str.substr(rank_pos);
1024   long rank_i = strict_strtol(rank_str.c_str(), 10, &err);
1025   if (rank_i < 0 || !err.empty()) {
1026     ss << "Invalid rank '" << rank_str << "'";
1027     return -EINVAL;
1028   } else {
1029     rank = rank_i;
1030   }
1031
1032   if (fs->mds_map.in.count(rank) == 0) {
1033     ss << "Rank '" << rank << "' not found";
1034     return -ENOENT;
1035   }
1036
1037   *role = {fs->fscid, rank};
1038
1039   return 0;
1040 }