ceph/src/osdc/Objecter.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 #include "Objecter.h"
  16 #include "osd/OSDMap.h"
  17 #include "Filer.h"
  18
  19 #include "mon/MonClient.h"
  20
  21 #include "msg/Messenger.h"
  22 #include "msg/Message.h"
  23
  24 #include "messages/MPing.h"
  25 #include "messages/MOSDOp.h"
  26 #include "messages/MOSDOpReply.h"
  27 #include "messages/MOSDBackoff.h"
  28 #include "messages/MOSDMap.h"
  29
  30 #include "messages/MPoolOp.h"
  31 #include "messages/MPoolOpReply.h"
  32
  33 #include "messages/MGetPoolStats.h"
  34 #include "messages/MGetPoolStatsReply.h"
  35 #include "messages/MStatfs.h"
  36 #include "messages/MStatfsReply.h"
  37
  38 #include "messages/MMonCommand.h"
  39
  40 #include "messages/MCommand.h"
  41 #include "messages/MCommandReply.h"
  42
  43 #include "messages/MWatchNotify.h"
  44
  45 #include <errno.h>
  46
  47 #include "common/config.h"
  48 #include "common/perf_counters.h"
  49 #include "common/scrub_types.h"
  50 #include "include/str_list.h"
  51 #include "common/errno.h"
  52 #include "common/EventTrace.h"
  53
  54 using ceph::real_time;
  55 using ceph::real_clock;
  56
  57 using ceph::mono_clock;
  58 using ceph::mono_time;
  59
  60 using ceph::timespan;
  61
  62
  63 #define dout_subsys ceph_subsys_objecter
  64 #undef dout_prefix
  65 #define dout_prefix *_dout << messenger->get_myname() << ".objecter "
  66
  67
  68 enum {
  69   l_osdc_first = 123200,
  70   l_osdc_op_active,
  71   l_osdc_op_laggy,
  72   l_osdc_op_send,
  73   l_osdc_op_send_bytes,
  74   l_osdc_op_resend,
  75   l_osdc_op_reply,
  76
  77   l_osdc_op,
  78   l_osdc_op_r,
  79   l_osdc_op_w,
  80   l_osdc_op_rmw,
  81   l_osdc_op_pg,
  82
  83   l_osdc_osdop_stat,
  84   l_osdc_osdop_create,
  85   l_osdc_osdop_read,
  86   l_osdc_osdop_write,
  87   l_osdc_osdop_writefull,
  88   l_osdc_osdop_writesame,
  89   l_osdc_osdop_append,
  90   l_osdc_osdop_zero,
  91   l_osdc_osdop_truncate,
  92   l_osdc_osdop_delete,
  93   l_osdc_osdop_mapext,
  94   l_osdc_osdop_sparse_read,
  95   l_osdc_osdop_clonerange,
  96   l_osdc_osdop_getxattr,
  97   l_osdc_osdop_setxattr,
  98   l_osdc_osdop_cmpxattr,
  99   l_osdc_osdop_rmxattr,
 100   l_osdc_osdop_resetxattrs,
 101   l_osdc_osdop_tmap_up,
 102   l_osdc_osdop_tmap_put,
 103   l_osdc_osdop_tmap_get,
 104   l_osdc_osdop_call,
 105   l_osdc_osdop_watch,
 106   l_osdc_osdop_notify,
 107   l_osdc_osdop_src_cmpxattr,
 108   l_osdc_osdop_pgls,
 109   l_osdc_osdop_pgls_filter,
 110   l_osdc_osdop_other,
 111
 112   l_osdc_linger_active,
 113   l_osdc_linger_send,
 114   l_osdc_linger_resend,
 115   l_osdc_linger_ping,
 116
 117   l_osdc_poolop_active,
 118   l_osdc_poolop_send,
 119   l_osdc_poolop_resend,
 120
 121   l_osdc_poolstat_active,
 122   l_osdc_poolstat_send,
 123   l_osdc_poolstat_resend,
 124
 125   l_osdc_statfs_active,
 126   l_osdc_statfs_send,
 127   l_osdc_statfs_resend,
 128
 129   l_osdc_command_active,
 130   l_osdc_command_send,
 131   l_osdc_command_resend,
 132
 133   l_osdc_map_epoch,
 134   l_osdc_map_full,
 135   l_osdc_map_inc,
 136
 137   l_osdc_osd_sessions,
 138   l_osdc_osd_session_open,
 139   l_osdc_osd_session_close,
 140   l_osdc_osd_laggy,
 141
 142   l_osdc_osdop_omap_wr,
 143   l_osdc_osdop_omap_rd,
 144   l_osdc_osdop_omap_del,
 145
 146   l_osdc_last,
 147 };
 148
 149
 150 // config obs ----------------------------
 151
 152 static const char *config_keys[] = {
 153   "crush_location",
 154   NULL
 155 };
 156
 157 class Objecter::RequestStateHook : public AdminSocketHook {
 158   Objecter *m_objecter;
 159 public:
 160   explicit RequestStateHook(Objecter *objecter);
 161   bool call(std::string command, cmdmap_t& cmdmap, std::string format,
 162             bufferlist& out) override;
 163 };
 164
 165 /**
 166  * This is a more limited form of C_Contexts, but that requires
 167  * a ceph_context which we don't have here.
 168  */
 169 class ObjectOperation::C_TwoContexts : public Context {
 170   Context *first;
 171   Context *second;
 172 public:
 173   C_TwoContexts(Context *first, Context *second) :
 174     first(first), second(second) {}
 175   void finish(int r) override {
 176     first->complete(r);
 177     second->complete(r);
 178     first = NULL;
 179     second = NULL;
 180   }
 181
 182   ~C_TwoContexts() override {
 183     delete first;
 184     delete second;
 185   }
 186 };
 187
 188 void ObjectOperation::add_handler(Context *extra) {
 189   size_t last = out_handler.size() - 1;
 190   Context *orig = out_handler[last];
 191   if (orig) {
 192     Context *wrapper = new C_TwoContexts(orig, extra);
 193     out_handler[last] = wrapper;
 194   } else {
 195     out_handler[last] = extra;
 196   }
 197 }
 198
 199 Objecter::OSDSession::unique_completion_lock Objecter::OSDSession::get_lock(
 200   object_t& oid)
 201 {
 202   if (oid.name.empty())
 203     return unique_completion_lock();
 204
 205   static constexpr uint32_t HASH_PRIME = 1021;
 206   uint32_t h = ceph_str_hash_linux(oid.name.c_str(), oid.name.size())
 207     % HASH_PRIME;
 208
 209   return unique_completion_lock(completion_locks[h % num_locks],
 210                                 std::defer_lock);
 211 }
 212
 213 const char** Objecter::get_tracked_conf_keys() const
 214 {
 215   return config_keys;
 216 }
 217
 218
 219 void Objecter::handle_conf_change(const struct md_config_t *conf,
 220                                   const std::set <std::string> &changed)
 221 {
 222   if (changed.count("crush_location")) {
 223     update_crush_location();
 224   }
 225 }
 226
 227 void Objecter::update_crush_location()
 228 {
 229   unique_lock wl(rwlock);
 230   crush_location = cct->crush_location.get_location();
 231 }
 232
 233 // messages ------------------------------
 234
 235 /*
 236  * initialize only internal data structures, don't initiate cluster interaction
 237  */
 238 void Objecter::init()
 239 {
 240   assert(!initialized);
 241
 242   if (!logger) {
 243     PerfCountersBuilder pcb(cct, "objecter", l_osdc_first, l_osdc_last);
 244
 245     pcb.add_u64(l_osdc_op_active, "op_active", "Operations active", "actv",
 246                 PerfCountersBuilder::PRIO_CRITICAL);
 247     pcb.add_u64(l_osdc_op_laggy, "op_laggy", "Laggy operations");
 248     pcb.add_u64_counter(l_osdc_op_send, "op_send", "Sent operations");
 249     pcb.add_u64_counter(l_osdc_op_send_bytes, "op_send_bytes", "Sent data");
 250     pcb.add_u64_counter(l_osdc_op_resend, "op_resend", "Resent operations");
 251     pcb.add_u64_counter(l_osdc_op_reply, "op_reply", "Operation reply");
 252
 253     pcb.add_u64_counter(l_osdc_op, "op", "Operations");
 254     pcb.add_u64_counter(l_osdc_op_r, "op_r", "Read operations", "rd",
 255                         PerfCountersBuilder::PRIO_CRITICAL);
 256     pcb.add_u64_counter(l_osdc_op_w, "op_w", "Write operations", "wr",
 257                         PerfCountersBuilder::PRIO_CRITICAL);
 258     pcb.add_u64_counter(l_osdc_op_rmw, "op_rmw", "Read-modify-write operations",
 259                         "rdwr", PerfCountersBuilder::PRIO_INTERESTING);
 260     pcb.add_u64_counter(l_osdc_op_pg, "op_pg", "PG operation");
 261
 262     pcb.add_u64_counter(l_osdc_osdop_stat, "osdop_stat", "Stat operations");
 263     pcb.add_u64_counter(l_osdc_osdop_create, "osdop_create",
 264                         "Create object operations");
 265     pcb.add_u64_counter(l_osdc_osdop_read, "osdop_read", "Read operations");
 266     pcb.add_u64_counter(l_osdc_osdop_write, "osdop_write", "Write operations");
 267     pcb.add_u64_counter(l_osdc_osdop_writefull, "osdop_writefull",
 268                         "Write full object operations");
 269     pcb.add_u64_counter(l_osdc_osdop_writesame, "osdop_writesame",
 270                         "Write same operations");
 271     pcb.add_u64_counter(l_osdc_osdop_append, "osdop_append",
 272                         "Append operation");
 273     pcb.add_u64_counter(l_osdc_osdop_zero, "osdop_zero",
 274                         "Set object to zero operations");
 275     pcb.add_u64_counter(l_osdc_osdop_truncate, "osdop_truncate",
 276                         "Truncate object operations");
 277     pcb.add_u64_counter(l_osdc_osdop_delete, "osdop_delete",
 278                         "Delete object operations");
 279     pcb.add_u64_counter(l_osdc_osdop_mapext, "osdop_mapext",
 280                         "Map extent operations");
 281     pcb.add_u64_counter(l_osdc_osdop_sparse_read, "osdop_sparse_read",
 282                         "Sparse read operations");
 283     pcb.add_u64_counter(l_osdc_osdop_clonerange, "osdop_clonerange",
 284                         "Clone range operations");
 285     pcb.add_u64_counter(l_osdc_osdop_getxattr, "osdop_getxattr",
 286                         "Get xattr operations");
 287     pcb.add_u64_counter(l_osdc_osdop_setxattr, "osdop_setxattr",
 288                         "Set xattr operations");
 289     pcb.add_u64_counter(l_osdc_osdop_cmpxattr, "osdop_cmpxattr",
 290                         "Xattr comparison operations");
 291     pcb.add_u64_counter(l_osdc_osdop_rmxattr, "osdop_rmxattr",
 292                         "Remove xattr operations");
 293     pcb.add_u64_counter(l_osdc_osdop_resetxattrs, "osdop_resetxattrs",
 294                         "Reset xattr operations");
 295     pcb.add_u64_counter(l_osdc_osdop_tmap_up, "osdop_tmap_up",
 296                         "TMAP update operations");
 297     pcb.add_u64_counter(l_osdc_osdop_tmap_put, "osdop_tmap_put",
 298                         "TMAP put operations");
 299     pcb.add_u64_counter(l_osdc_osdop_tmap_get, "osdop_tmap_get",
 300                         "TMAP get operations");
 301     pcb.add_u64_counter(l_osdc_osdop_call, "osdop_call",
 302                         "Call (execute) operations");
 303     pcb.add_u64_counter(l_osdc_osdop_watch, "osdop_watch",
 304                         "Watch by object operations");
 305     pcb.add_u64_counter(l_osdc_osdop_notify, "osdop_notify",
 306                         "Notify about object operations");
 307     pcb.add_u64_counter(l_osdc_osdop_src_cmpxattr, "osdop_src_cmpxattr",
 308                         "Extended attribute comparison in multi operations");
 309     pcb.add_u64_counter(l_osdc_osdop_pgls, "osdop_pgls");
 310     pcb.add_u64_counter(l_osdc_osdop_pgls_filter, "osdop_pgls_filter");
 311     pcb.add_u64_counter(l_osdc_osdop_other, "osdop_other", "Other operations");
 312
 313     pcb.add_u64(l_osdc_linger_active, "linger_active",
 314                 "Active lingering operations");
 315     pcb.add_u64_counter(l_osdc_linger_send, "linger_send",
 316                         "Sent lingering operations");
 317     pcb.add_u64_counter(l_osdc_linger_resend, "linger_resend",
 318                         "Resent lingering operations");
 319     pcb.add_u64_counter(l_osdc_linger_ping, "linger_ping",
 320                         "Sent pings to lingering operations");
 321
 322     pcb.add_u64(l_osdc_poolop_active, "poolop_active",
 323                 "Active pool operations");
 324     pcb.add_u64_counter(l_osdc_poolop_send, "poolop_send",
 325                         "Sent pool operations");
 326     pcb.add_u64_counter(l_osdc_poolop_resend, "poolop_resend",
 327                         "Resent pool operations");
 328
 329     pcb.add_u64(l_osdc_poolstat_active, "poolstat_active",
 330                 "Active get pool stat operations");
 331     pcb.add_u64_counter(l_osdc_poolstat_send, "poolstat_send",
 332                         "Pool stat operations sent");
 333     pcb.add_u64_counter(l_osdc_poolstat_resend, "poolstat_resend",
 334                         "Resent pool stats");
 335
 336     pcb.add_u64(l_osdc_statfs_active, "statfs_active", "Statfs operations");
 337     pcb.add_u64_counter(l_osdc_statfs_send, "statfs_send", "Sent FS stats");
 338     pcb.add_u64_counter(l_osdc_statfs_resend, "statfs_resend",
 339                         "Resent FS stats");
 340
 341     pcb.add_u64(l_osdc_command_active, "command_active", "Active commands");
 342     pcb.add_u64_counter(l_osdc_command_send, "command_send",
 343                         "Sent commands");
 344     pcb.add_u64_counter(l_osdc_command_resend, "command_resend",
 345                         "Resent commands");
 346
 347     pcb.add_u64(l_osdc_map_epoch, "map_epoch", "OSD map epoch");
 348     pcb.add_u64_counter(l_osdc_map_full, "map_full",
 349                         "Full OSD maps received");
 350     pcb.add_u64_counter(l_osdc_map_inc, "map_inc",
 351                         "Incremental OSD maps received");
 352
 353     pcb.add_u64(l_osdc_osd_sessions, "osd_sessions",
 354                 "Open sessions");  // open sessions
 355     pcb.add_u64_counter(l_osdc_osd_session_open, "osd_session_open",
 356                         "Sessions opened");
 357     pcb.add_u64_counter(l_osdc_osd_session_close, "osd_session_close",
 358                         "Sessions closed");
 359     pcb.add_u64(l_osdc_osd_laggy, "osd_laggy", "Laggy OSD sessions");
 360
 361     pcb.add_u64_counter(l_osdc_osdop_omap_wr, "omap_wr",
 362                         "OSD OMAP write operations");
 363     pcb.add_u64_counter(l_osdc_osdop_omap_rd, "omap_rd",
 364                         "OSD OMAP read operations");
 365     pcb.add_u64_counter(l_osdc_osdop_omap_del, "omap_del",
 366                         "OSD OMAP delete operations");
 367
 368     logger = pcb.create_perf_counters();
 369     cct->get_perfcounters_collection()->add(logger);
 370   }
 371
 372   m_request_state_hook = new RequestStateHook(this);
 373   AdminSocket* admin_socket = cct->get_admin_socket();
 374   int ret = admin_socket->register_command("objecter_requests",
 375                                            "objecter_requests",
 376                                            m_request_state_hook,
 377                                            "show in-progress osd requests");
 378
 379   /* Don't warn on EEXIST, happens if multiple ceph clients
 380    * are instantiated from one process */
 381   if (ret < 0 && ret != -EEXIST) {
 382     lderr(cct) << "error registering admin socket command: "
 383                << cpp_strerror(ret) << dendl;
 384   }
 385
 386   update_crush_location();
 387
 388   cct->_conf->add_observer(this);
 389
 390   initialized = true;
 391 }
 392
 393 /*
 394  * ok, cluster interaction can happen
 395  */
 396 void Objecter::start(const OSDMap* o)
 397 {
 398   shared_lock rl(rwlock);
 399
 400   start_tick();
 401   if (o) {
 402     osdmap->deepish_copy_from(*o);
 403   } else if (osdmap->get_epoch() == 0) {
 404     _maybe_request_map();
 405   }
 406 }
 407
 408 void Objecter::shutdown()
 409 {
 410   assert(initialized);
 411
 412   unique_lock wl(rwlock);
 413
 414   initialized = false;
 415
 416   cct->_conf->remove_observer(this);
 417
 418   map<int,OSDSession*>::iterator p;
 419   while (!osd_sessions.empty()) {
 420     p = osd_sessions.begin();
 421     close_session(p->second);
 422   }
 423
 424   while(!check_latest_map_lingers.empty()) {
 425     map<uint64_t, LingerOp*>::iterator i = check_latest_map_lingers.begin();
 426     i->second->put();
 427     check_latest_map_lingers.erase(i->first);
 428   }
 429
 430   while(!check_latest_map_ops.empty()) {
 431     map<ceph_tid_t, Op*>::iterator i = check_latest_map_ops.begin();
 432     i->second->put();
 433     check_latest_map_ops.erase(i->first);
 434   }
 435
 436   while(!check_latest_map_commands.empty()) {
 437     map<ceph_tid_t, CommandOp*>::iterator i
 438       = check_latest_map_commands.begin();
 439     i->second->put();
 440     check_latest_map_commands.erase(i->first);
 441   }
 442
 443   while(!poolstat_ops.empty()) {
 444     map<ceph_tid_t,PoolStatOp*>::iterator i = poolstat_ops.begin();
 445     delete i->second;
 446     poolstat_ops.erase(i->first);
 447   }
 448
 449   while(!statfs_ops.empty()) {
 450     map<ceph_tid_t, StatfsOp*>::iterator i = statfs_ops.begin();
 451     delete i->second;
 452     statfs_ops.erase(i->first);
 453   }
 454
 455   while(!pool_ops.empty()) {
 456     map<ceph_tid_t, PoolOp*>::iterator i = pool_ops.begin();
 457     delete i->second;
 458     pool_ops.erase(i->first);
 459   }
 460
 461   ldout(cct, 20) << __func__ << " clearing up homeless session..." << dendl;
 462   while(!homeless_session->linger_ops.empty()) {
 463     std::map<uint64_t, LingerOp*>::iterator i
 464       = homeless_session->linger_ops.begin();
 465     ldout(cct, 10) << " linger_op " << i->first << dendl;
 466     LingerOp *lop = i->second;
 467     {
 468       OSDSession::unique_lock swl(homeless_session->lock);
 469       _session_linger_op_remove(homeless_session, lop);
 470     }
 471     linger_ops.erase(lop->linger_id);
 472     linger_ops_set.erase(lop);
 473     lop->put();
 474   }
 475
 476   while(!homeless_session->ops.empty()) {
 477     std::map<ceph_tid_t, Op*>::iterator i = homeless_session->ops.begin();
 478     ldout(cct, 10) << " op " << i->first << dendl;
 479     Op *op = i->second;
 480     {
 481       OSDSession::unique_lock swl(homeless_session->lock);
 482       _session_op_remove(homeless_session, op);
 483     }
 484     op->put();
 485   }
 486
 487   while(!homeless_session->command_ops.empty()) {
 488     std::map<ceph_tid_t, CommandOp*>::iterator i
 489       = homeless_session->command_ops.begin();
 490     ldout(cct, 10) << " command_op " << i->first << dendl;
 491     CommandOp *cop = i->second;
 492     {
 493       OSDSession::unique_lock swl(homeless_session->lock);
 494       _session_command_op_remove(homeless_session, cop);
 495     }
 496     cop->put();
 497   }
 498
 499   if (tick_event) {
 500     if (timer.cancel_event(tick_event)) {
 501       ldout(cct, 10) <<  " successfully canceled tick" << dendl;
 502     }
 503     tick_event = 0;
 504   }
 505
 506   if (logger) {
 507     cct->get_perfcounters_collection()->remove(logger);
 508     delete logger;
 509     logger = NULL;
 510   }
 511
 512   // Let go of Objecter write lock so timer thread can shutdown
 513   wl.unlock();
 514
 515   // Outside of lock to avoid cycle WRT calls to RequestStateHook
 516   // This is safe because we guarantee no concurrent calls to
 517   // shutdown() with the ::initialized check at start.
 518   if (m_request_state_hook) {
 519     AdminSocket* admin_socket = cct->get_admin_socket();
 520     admin_socket->unregister_command("objecter_requests");
 521     delete m_request_state_hook;
 522     m_request_state_hook = NULL;
 523   }
 524 }
 525
 526 void Objecter::_send_linger(LingerOp *info,
 527                             shunique_lock& sul)
 528 {
 529   assert(sul.owns_lock() && sul.mutex() == &rwlock);
 530
 531   vector<OSDOp> opv;
 532   Context *oncommit = NULL;
 533   LingerOp::shared_lock watchl(info->watch_lock);
 534   bufferlist *poutbl = NULL;
 535   if (info->registered && info->is_watch) {
 536     ldout(cct, 15) << "send_linger " << info->linger_id << " reconnect"
 537                    << dendl;
 538     opv.push_back(OSDOp());
 539     opv.back().op.op = CEPH_OSD_OP_WATCH;
 540     opv.back().op.watch.cookie = info->get_cookie();
 541     opv.back().op.watch.op = CEPH_OSD_WATCH_OP_RECONNECT;
 542     opv.back().op.watch.gen = ++info->register_gen;
 543     oncommit = new C_Linger_Reconnect(this, info);
 544   } else {
 545     ldout(cct, 15) << "send_linger " << info->linger_id << " register"
 546                    << dendl;
 547     opv = info->ops;
 548     C_Linger_Commit *c = new C_Linger_Commit(this, info);
 549     if (!info->is_watch) {
 550       info->notify_id = 0;
 551       poutbl = &c->outbl;
 552     }
 553     oncommit = c;
 554   }
 555   watchl.unlock();
 556   Op *o = new Op(info->target.base_oid, info->target.base_oloc,
 557                  opv, info->target.flags | CEPH_OSD_FLAG_READ,
 558                  oncommit, info->pobjver);
 559   o->outbl = poutbl;
 560   o->snapid = info->snap;
 561   o->snapc = info->snapc;
 562   o->mtime = info->mtime;
 563
 564   o->target = info->target;
 565   o->tid = ++last_tid;
 566
 567   // do not resend this; we will send a new op to reregister
 568   o->should_resend = false;
 569
 570   if (info->register_tid) {
 571     // repeat send.  cancel old registeration op, if any.
 572     OSDSession::unique_lock sl(info->session->lock);
 573     if (info->session->ops.count(info->register_tid)) {
 574       Op *o = info->session->ops[info->register_tid];
 575       _op_cancel_map_check(o);
 576       _cancel_linger_op(o);
 577     }
 578     sl.unlock();
 579
 580     _op_submit(o, sul, &info->register_tid);
 581   } else {
 582     // first send
 583     _op_submit_with_budget(o, sul, &info->register_tid);
 584   }
 585
 586   logger->inc(l_osdc_linger_send);
 587 }
 588
 589 void Objecter::_linger_commit(LingerOp *info, int r, bufferlist& outbl)
 590 {
 591   LingerOp::unique_lock wl(info->watch_lock);
 592   ldout(cct, 10) << "_linger_commit " << info->linger_id << dendl;
 593   if (info->on_reg_commit) {
 594     info->on_reg_commit->complete(r);
 595     info->on_reg_commit = NULL;
 596   }
 597
 598   // only tell the user the first time we do this
 599   info->registered = true;
 600   info->pobjver = NULL;
 601
 602   if (!info->is_watch) {
 603     // make note of the notify_id
 604     bufferlist::iterator p = outbl.begin();
 605     try {
 606       ::decode(info->notify_id, p);
 607       ldout(cct, 10) << "_linger_commit  notify_id=" << info->notify_id
 608                      << dendl;
 609     }
 610     catch (buffer::error& e) {
 611     }
 612   }
 613 }
 614
 615 struct C_DoWatchError : public Context {
 616   Objecter *objecter;
 617   Objecter::LingerOp *info;
 618   int err;
 619   C_DoWatchError(Objecter *o, Objecter::LingerOp *i, int r)
 620     : objecter(o), info(i), err(r) {
 621     info->get();
 622     info->_queued_async();
 623   }
 624   void finish(int r) override {
 625     Objecter::unique_lock wl(objecter->rwlock);
 626     bool canceled = info->canceled;
 627     wl.unlock();
 628
 629     if (!canceled) {
 630       info->watch_context->handle_error(info->get_cookie(), err);
 631     }
 632
 633     info->finished_async();
 634     info->put();
 635   }
 636 };
 637
 638 int Objecter::_normalize_watch_error(int r)
 639 {
 640   // translate ENOENT -> ENOTCONN so that a delete->disconnection
 641   // notification and a failure to reconnect becuase we raced with
 642   // the delete appear the same to the user.
 643   if (r == -ENOENT)
 644     r = -ENOTCONN;
 645   return r;
 646 }
 647
 648 void Objecter::_linger_reconnect(LingerOp *info, int r)
 649 {
 650   ldout(cct, 10) << __func__ << " " << info->linger_id << " = " << r
 651                  << " (last_error " << info->last_error << ")" << dendl;
 652   if (r < 0) {
 653     LingerOp::unique_lock wl(info->watch_lock);
 654     if (!info->last_error) {
 655       r = _normalize_watch_error(r);
 656       info->last_error = r;
 657       if (info->watch_context) {
 658         finisher->queue(new C_DoWatchError(this, info, r));
 659       }
 660     }
 661     wl.unlock();
 662   }
 663 }
 664
 665 void Objecter::_send_linger_ping(LingerOp *info)
 666 {
 667   // rwlock is locked unique
 668   // info->session->lock is locked
 669
 670   if (cct->_conf->objecter_inject_no_watch_ping) {
 671     ldout(cct, 10) << __func__ << " " << info->linger_id << " SKIPPING"
 672                    << dendl;
 673     return;
 674   }
 675   if (osdmap->test_flag(CEPH_OSDMAP_PAUSERD)) {
 676     ldout(cct, 10) << __func__ << " PAUSERD" << dendl;
 677     return;
 678   }
 679
 680   ceph::mono_time now = ceph::mono_clock::now();
 681   ldout(cct, 10) << __func__ << " " << info->linger_id << " now " << now
 682                  << dendl;
 683
 684   vector<OSDOp> opv(1);
 685   opv[0].op.op = CEPH_OSD_OP_WATCH;
 686   opv[0].op.watch.cookie = info->get_cookie();
 687   opv[0].op.watch.op = CEPH_OSD_WATCH_OP_PING;
 688   opv[0].op.watch.gen = info->register_gen;
 689   C_Linger_Ping *onack = new C_Linger_Ping(this, info);
 690   Op *o = new Op(info->target.base_oid, info->target.base_oloc,
 691                  opv, info->target.flags | CEPH_OSD_FLAG_READ,
 692                  onack, NULL, NULL);
 693   o->target = info->target;
 694   o->should_resend = false;
 695   _send_op_account(o);
 696   MOSDOp *m = _prepare_osd_op(o);
 697   o->tid = ++last_tid;
 698   _session_op_assign(info->session, o);
 699   _send_op(o, m);
 700   info->ping_tid = o->tid;
 701
 702   onack->sent = now;
 703   logger->inc(l_osdc_linger_ping);
 704 }
 705
 706 void Objecter::_linger_ping(LingerOp *info, int r, mono_time sent,
 707                             uint32_t register_gen)
 708 {
 709   LingerOp::unique_lock l(info->watch_lock);
 710   ldout(cct, 10) << __func__ << " " << info->linger_id
 711                  << " sent " << sent << " gen " << register_gen << " = " << r
 712                  << " (last_error " << info->last_error
 713                  << " register_gen " << info->register_gen << ")" << dendl;
 714   if (info->register_gen == register_gen) {
 715     if (r == 0) {
 716       info->watch_valid_thru = sent;
 717     } else if (r < 0 && !info->last_error) {
 718       r = _normalize_watch_error(r);
 719       info->last_error = r;
 720       if (info->watch_context) {
 721         finisher->queue(new C_DoWatchError(this, info, r));
 722       }
 723     }
 724   } else {
 725     ldout(cct, 20) << " ignoring old gen" << dendl;
 726   }
 727 }
 728
 729 int Objecter::linger_check(LingerOp *info)
 730 {
 731   LingerOp::shared_lock l(info->watch_lock);
 732
 733   mono_time stamp = info->watch_valid_thru;
 734   if (!info->watch_pending_async.empty())
 735     stamp = MIN(info->watch_valid_thru, info->watch_pending_async.front());
 736   auto age = mono_clock::now() - stamp;
 737
 738   ldout(cct, 10) << __func__ << " " << info->linger_id
 739                  << " err " << info->last_error
 740                  << " age " << age << dendl;
 741   if (info->last_error)
 742     return info->last_error;
 743   // return a safe upper bound (we are truncating to ms)
 744   return
 745     1 + std::chrono::duration_cast<std::chrono::milliseconds>(age).count();
 746 }
 747
 748 void Objecter::linger_cancel(LingerOp *info)
 749 {
 750   unique_lock wl(rwlock);
 751   _linger_cancel(info);
 752   info->put();
 753 }
 754
 755 void Objecter::_linger_cancel(LingerOp *info)
 756 {
 757   // rwlock is locked unique
 758   ldout(cct, 20) << __func__ << " linger_id=" << info->linger_id << dendl;
 759   if (!info->canceled) {
 760     OSDSession *s = info->session;
 761     OSDSession::unique_lock sl(s->lock);
 762     _session_linger_op_remove(s, info);
 763     sl.unlock();
 764
 765     linger_ops.erase(info->linger_id);
 766     linger_ops_set.erase(info);
 767     assert(linger_ops.size() == linger_ops_set.size());
 768
 769     info->canceled = true;
 770     info->put();
 771
 772     logger->dec(l_osdc_linger_active);
 773   }
 774 }
 775
 776
 777
 778 Objecter::LingerOp *Objecter::linger_register(const object_t& oid,
 779                                               const object_locator_t& oloc,
 780                                               int flags)
 781 {
 782   LingerOp *info = new LingerOp;
 783   info->target.base_oid = oid;
 784   info->target.base_oloc = oloc;
 785   if (info->target.base_oloc.key == oid)
 786     info->target.base_oloc.key.clear();
 787   info->target.flags = flags;
 788   info->watch_valid_thru = mono_clock::now();
 789
 790   unique_lock l(rwlock);
 791
 792   // Acquire linger ID
 793   info->linger_id = ++max_linger_id;
 794   ldout(cct, 10) << __func__ << " info " << info
 795                  << " linger_id " << info->linger_id
 796                  << " cookie " << info->get_cookie()
 797                  << dendl;
 798   linger_ops[info->linger_id] = info;
 799   linger_ops_set.insert(info);
 800   assert(linger_ops.size() == linger_ops_set.size());
 801
 802   info->get(); // for the caller
 803   return info;
 804 }
 805
 806 ceph_tid_t Objecter::linger_watch(LingerOp *info,
 807                                   ObjectOperation& op,
 808                                   const SnapContext& snapc,
 809                                   real_time mtime,
 810                                   bufferlist& inbl,
 811                                   Context *oncommit,
 812                                   version_t *objver)
 813 {
 814   info->is_watch = true;
 815   info->snapc = snapc;
 816   info->mtime = mtime;
 817   info->target.flags |= CEPH_OSD_FLAG_WRITE;
 818   info->ops = op.ops;
 819   info->inbl = inbl;
 820   info->poutbl = NULL;
 821   info->pobjver = objver;
 822   info->on_reg_commit = oncommit;
 823
 824   shunique_lock sul(rwlock, ceph::acquire_unique);
 825   _linger_submit(info, sul);
 826   logger->inc(l_osdc_linger_active);
 827
 828   return info->linger_id;
 829 }
 830
 831 ceph_tid_t Objecter::linger_notify(LingerOp *info,
 832                                    ObjectOperation& op,
 833                                    snapid_t snap, bufferlist& inbl,
 834                                    bufferlist *poutbl,
 835                                    Context *onfinish,
 836                                    version_t *objver)
 837 {
 838   info->snap = snap;
 839   info->target.flags |= CEPH_OSD_FLAG_READ;
 840   info->ops = op.ops;
 841   info->inbl = inbl;
 842   info->poutbl = poutbl;
 843   info->pobjver = objver;
 844   info->on_reg_commit = onfinish;
 845
 846   shunique_lock sul(rwlock, ceph::acquire_unique);
 847   _linger_submit(info, sul);
 848   logger->inc(l_osdc_linger_active);
 849
 850   return info->linger_id;
 851 }
 852
 853 void Objecter::_linger_submit(LingerOp *info, shunique_lock& sul)
 854 {
 855   assert(sul.owns_lock() && sul.mutex() == &rwlock);
 856   assert(info->linger_id);
 857
 858   // Populate Op::target
 859   OSDSession *s = NULL;
 860   _calc_target(&info->target, nullptr);
 861
 862   // Create LingerOp<->OSDSession relation
 863   int r = _get_session(info->target.osd, &s, sul);
 864   assert(r == 0);
 865   OSDSession::unique_lock sl(s->lock);
 866   _session_linger_op_assign(s, info);
 867   sl.unlock();
 868   put_session(s);
 869
 870   _send_linger(info, sul);
 871 }
 872
 873 struct C_DoWatchNotify : public Context {
 874   Objecter *objecter;
 875   Objecter::LingerOp *info;
 876   MWatchNotify *msg;
 877   C_DoWatchNotify(Objecter *o, Objecter::LingerOp *i, MWatchNotify *m)
 878     : objecter(o), info(i), msg(m) {
 879     info->get();
 880     info->_queued_async();
 881     msg->get();
 882   }
 883   void finish(int r) override {
 884     objecter->_do_watch_notify(info, msg);
 885   }
 886 };
 887
 888 void Objecter::handle_watch_notify(MWatchNotify *m)
 889 {
 890   shared_lock l(rwlock);
 891   if (!initialized) {
 892     return;
 893   }
 894
 895   LingerOp *info = reinterpret_cast<LingerOp*>(m->cookie);
 896   if (linger_ops_set.count(info) == 0) {
 897     ldout(cct, 7) << __func__ << " cookie " << m->cookie << " dne" << dendl;
 898     return;
 899   }
 900   LingerOp::unique_lock wl(info->watch_lock);
 901   if (m->opcode == CEPH_WATCH_EVENT_DISCONNECT) {
 902     if (!info->last_error) {
 903       info->last_error = -ENOTCONN;
 904       if (info->watch_context) {
 905         finisher->queue(new C_DoWatchError(this, info, -ENOTCONN));
 906       }
 907     }
 908   } else if (!info->is_watch) {
 909     // we have CEPH_WATCH_EVENT_NOTIFY_COMPLETE; we can do this inline
 910     // since we know the only user (librados) is safe to call in
 911     // fast-dispatch context
 912     if (info->notify_id &&
 913         info->notify_id != m->notify_id) {
 914       ldout(cct, 10) << __func__ << " reply notify " << m->notify_id
 915                      << " != " << info->notify_id << ", ignoring" << dendl;
 916     } else if (info->on_notify_finish) {
 917       info->notify_result_bl->claim(m->get_data());
 918       info->on_notify_finish->complete(m->return_code);
 919
 920       // if we race with reconnect we might get a second notify; only
 921       // notify the caller once!
 922       info->on_notify_finish = NULL;
 923     }
 924   } else {
 925     finisher->queue(new C_DoWatchNotify(this, info, m));
 926   }
 927 }
 928
 929 void Objecter::_do_watch_notify(LingerOp *info, MWatchNotify *m)
 930 {
 931   ldout(cct, 10) << __func__ << " " << *m << dendl;
 932
 933   shared_lock l(rwlock);
 934   assert(initialized);
 935
 936   if (info->canceled) {
 937     l.unlock();
 938     goto out;
 939   }
 940
 941   // notify completion?
 942   assert(info->is_watch);
 943   assert(info->watch_context);
 944   assert(m->opcode != CEPH_WATCH_EVENT_DISCONNECT);
 945
 946   l.unlock();
 947
 948   switch (m->opcode) {
 949   case CEPH_WATCH_EVENT_NOTIFY:
 950     info->watch_context->handle_notify(m->notify_id, m->cookie,
 951                                        m->notifier_gid, m->bl);
 952     break;
 953   }
 954
 955  out:
 956   info->finished_async();
 957   info->put();
 958   m->put();
 959 }
 960
 961 bool Objecter::ms_dispatch(Message *m)
 962 {
 963   ldout(cct, 10) << __func__ << " " << cct << " " << *m << dendl;
 964   if (!initialized)
 965     return false;
 966
 967   switch (m->get_type()) {
 968     // these we exlusively handle
 969   case CEPH_MSG_OSD_OPREPLY:
 970     handle_osd_op_reply(static_cast<MOSDOpReply*>(m));
 971     return true;
 972
 973   case CEPH_MSG_OSD_BACKOFF:
 974     handle_osd_backoff(static_cast<MOSDBackoff*>(m));
 975     return true;
 976
 977   case CEPH_MSG_WATCH_NOTIFY:
 978     handle_watch_notify(static_cast<MWatchNotify*>(m));
 979     m->put();
 980     return true;
 981
 982   case MSG_COMMAND_REPLY:
 983     if (m->get_source().type() == CEPH_ENTITY_TYPE_OSD) {
 984       handle_command_reply(static_cast<MCommandReply*>(m));
 985       return true;
 986     } else {
 987       return false;
 988     }
 989
 990   case MSG_GETPOOLSTATSREPLY:
 991     handle_get_pool_stats_reply(static_cast<MGetPoolStatsReply*>(m));
 992     return true;
 993
 994   case CEPH_MSG_POOLOP_REPLY:
 995     handle_pool_op_reply(static_cast<MPoolOpReply*>(m));
 996     return true;
 997
 998   case CEPH_MSG_STATFS_REPLY:
 999     handle_fs_stats_reply(static_cast<MStatfsReply*>(m));
1000     return true;
1001
1002     // these we give others a chance to inspect
1003
1004     // MDS, OSD
1005   case CEPH_MSG_OSD_MAP:
1006     handle_osd_map(static_cast<MOSDMap*>(m));
1007     return false;
1008   }
1009   return false;
1010 }
1011
1012 void Objecter::_scan_requests(OSDSession *s,
1013                               bool force_resend,
1014                               bool cluster_full,
1015                               map<int64_t, bool> *pool_full_map,
1016                               map<ceph_tid_t, Op*>& need_resend,
1017                               list<LingerOp*>& need_resend_linger,
1018                               map<ceph_tid_t, CommandOp*>& need_resend_command,
1019                               shunique_lock& sul)
1020 {
1021   assert(sul.owns_lock() && sul.mutex() == &rwlock);
1022
1023   list<LingerOp*> unregister_lingers;
1024
1025   OSDSession::unique_lock sl(s->lock);
1026
1027   // check for changed linger mappings (_before_ regular ops)
1028   map<ceph_tid_t,LingerOp*>::iterator lp = s->linger_ops.begin();
1029   while (lp != s->linger_ops.end()) {
1030     LingerOp *op = lp->second;
1031     assert(op->session == s);
1032     // check_linger_pool_dne() may touch linger_ops; prevent iterator
1033     // invalidation
1034     ++lp;
1035     ldout(cct, 10) << " checking linger op " << op->linger_id << dendl;
1036     bool unregister, force_resend_writes = cluster_full;
1037     int r = _recalc_linger_op_target(op, sul);
1038     if (pool_full_map)
1039       force_resend_writes = force_resend_writes ||
1040         (*pool_full_map)[op->target.base_oloc.pool];
1041     switch (r) {
1042     case RECALC_OP_TARGET_NO_ACTION:
1043       if (!force_resend && !force_resend_writes)
1044         break;
1045       // -- fall-thru --
1046     case RECALC_OP_TARGET_NEED_RESEND:
1047       need_resend_linger.push_back(op);
1048       _linger_cancel_map_check(op);
1049       break;
1050     case RECALC_OP_TARGET_POOL_DNE:
1051       _check_linger_pool_dne(op, &unregister);
1052       if (unregister) {
1053         ldout(cct, 10) << " need to unregister linger op "
1054                        << op->linger_id << dendl;
1055         op->get();
1056         unregister_lingers.push_back(op);
1057       }
1058       break;
1059     }
1060   }
1061
1062   // check for changed request mappings
1063   map<ceph_tid_t,Op*>::iterator p = s->ops.begin();
1064   while (p != s->ops.end()) {
1065     Op *op = p->second;
1066     ++p;   // check_op_pool_dne() may touch ops; prevent iterator invalidation
1067     ldout(cct, 10) << " checking op " << op->tid << dendl;
1068     bool force_resend_writes = cluster_full;
1069     if (pool_full_map)
1070       force_resend_writes = force_resend_writes ||
1071         (*pool_full_map)[op->target.base_oloc.pool];
1072     int r = _calc_target(&op->target,
1073                          op->session ? op->session->con.get() : nullptr);
1074     switch (r) {
1075     case RECALC_OP_TARGET_NO_ACTION:
1076       if (!force_resend && !(force_resend_writes && op->respects_full()))
1077         break;
1078       // -- fall-thru --
1079     case RECALC_OP_TARGET_NEED_RESEND:
1080       if (op->session) {
1081         _session_op_remove(op->session, op);
1082       }
1083       need_resend[op->tid] = op;
1084       _op_cancel_map_check(op);
1085       break;
1086     case RECALC_OP_TARGET_POOL_DNE:
1087       _check_op_pool_dne(op, &sl);
1088       break;
1089     }
1090   }
1091
1092   // commands
1093   map<ceph_tid_t,CommandOp*>::iterator cp = s->command_ops.begin();
1094   while (cp != s->command_ops.end()) {
1095     CommandOp *c = cp->second;
1096     ++cp;
1097     ldout(cct, 10) << " checking command " << c->tid << dendl;
1098     bool force_resend_writes = cluster_full;
1099     if (pool_full_map)
1100       force_resend_writes = force_resend_writes ||
1101         (*pool_full_map)[c->target_pg.pool()];
1102     int r = _calc_command_target(c, sul);
1103     switch (r) {
1104     case RECALC_OP_TARGET_NO_ACTION:
1105       // resend if skipped map; otherwise do nothing.
1106       if (!force_resend && !force_resend_writes)
1107         break;
1108       // -- fall-thru --
1109     case RECALC_OP_TARGET_NEED_RESEND:
1110       need_resend_command[c->tid] = c;
1111       if (c->session) {
1112         _session_command_op_remove(c->session, c);
1113       }
1114       _command_cancel_map_check(c);
1115       break;
1116     case RECALC_OP_TARGET_POOL_DNE:
1117     case RECALC_OP_TARGET_OSD_DNE:
1118     case RECALC_OP_TARGET_OSD_DOWN:
1119       _check_command_map_dne(c);
1120       break;
1121     }
1122   }
1123
1124   sl.unlock();
1125
1126   for (list<LingerOp*>::iterator iter = unregister_lingers.begin();
1127        iter != unregister_lingers.end();
1128        ++iter) {
1129     _linger_cancel(*iter);
1130     (*iter)->put();
1131   }
1132 }
1133
1134 void Objecter::handle_osd_map(MOSDMap *m)
1135 {
1136   shunique_lock sul(rwlock, acquire_unique);
1137   if (!initialized)
1138     return;
1139
1140   assert(osdmap);
1141
1142   if (m->fsid != monc->get_fsid()) {
1143     ldout(cct, 0) << "handle_osd_map fsid " << m->fsid
1144                   << " != " << monc->get_fsid() << dendl;
1145     return;
1146   }
1147
1148   bool was_pauserd = osdmap->test_flag(CEPH_OSDMAP_PAUSERD);
1149   bool cluster_full = _osdmap_full_flag();
1150   bool was_pausewr = osdmap->test_flag(CEPH_OSDMAP_PAUSEWR) || cluster_full ||
1151     _osdmap_has_pool_full();
1152   map<int64_t, bool> pool_full_map;
1153   for (map<int64_t, pg_pool_t>::const_iterator it
1154          = osdmap->get_pools().begin();
1155        it != osdmap->get_pools().end(); ++it)
1156     pool_full_map[it->first] = _osdmap_pool_full(it->second);
1157
1158
1159   list<LingerOp*> need_resend_linger;
1160   map<ceph_tid_t, Op*> need_resend;
1161   map<ceph_tid_t, CommandOp*> need_resend_command;
1162
1163   if (m->get_last() <= osdmap->get_epoch()) {
1164     ldout(cct, 3) << "handle_osd_map ignoring epochs ["
1165                   << m->get_first() << "," << m->get_last()
1166                   << "] <= " << osdmap->get_epoch() << dendl;
1167   } else {
1168     ldout(cct, 3) << "handle_osd_map got epochs ["
1169                   << m->get_first() << "," << m->get_last()
1170                   << "] > " << osdmap->get_epoch() << dendl;
1171
1172     if (osdmap->get_epoch()) {
1173       bool skipped_map = false;
1174       // we want incrementals
1175       for (epoch_t e = osdmap->get_epoch() + 1;
1176            e <= m->get_last();
1177            e++) {
1178
1179         if (osdmap->get_epoch() == e-1 &&
1180             m->incremental_maps.count(e)) {
1181           ldout(cct, 3) << "handle_osd_map decoding incremental epoch " << e
1182                         << dendl;
1183           OSDMap::Incremental inc(m->incremental_maps[e]);
1184           osdmap->apply_incremental(inc);
1185
1186           emit_blacklist_events(inc);
1187
1188           logger->inc(l_osdc_map_inc);
1189         }
1190         else if (m->maps.count(e)) {
1191           ldout(cct, 3) << "handle_osd_map decoding full epoch " << e << dendl;
1192           OSDMap *new_osdmap = new OSDMap();
1193           new_osdmap->decode(m->maps[e]);
1194
1195           emit_blacklist_events(*osdmap, *new_osdmap);
1196
1197           osdmap = new_osdmap;
1198
1199           logger->inc(l_osdc_map_full);
1200         }
1201         else {
1202           if (e >= m->get_oldest()) {
1203             ldout(cct, 3) << "handle_osd_map requesting missing epoch "
1204                           << osdmap->get_epoch()+1 << dendl;
1205             _maybe_request_map();
1206             break;
1207           }
1208           ldout(cct, 3) << "handle_osd_map missing epoch "
1209                         << osdmap->get_epoch()+1
1210                         << ", jumping to " << m->get_oldest() << dendl;
1211           e = m->get_oldest() - 1;
1212           skipped_map = true;
1213           continue;
1214         }
1215         logger->set(l_osdc_map_epoch, osdmap->get_epoch());
1216
1217         cluster_full = cluster_full || _osdmap_full_flag();
1218         update_pool_full_map(pool_full_map);
1219
1220         // check all outstanding requests on every epoch
1221         _scan_requests(homeless_session, skipped_map, cluster_full,
1222                        &pool_full_map, need_resend,
1223                        need_resend_linger, need_resend_command, sul);
1224         for (map<int,OSDSession*>::iterator p = osd_sessions.begin();
1225              p != osd_sessions.end(); ) {
1226           OSDSession *s = p->second;
1227           _scan_requests(s, skipped_map, cluster_full,
1228                          &pool_full_map, need_resend,
1229                          need_resend_linger, need_resend_command, sul);
1230           ++p;
1231           // osd down or addr change?
1232           if (!osdmap->is_up(s->osd) ||
1233               (s->con &&
1234                s->con->get_peer_addr() != osdmap->get_inst(s->osd).addr)) {
1235             close_session(s);
1236           }
1237         }
1238
1239         assert(e == osdmap->get_epoch());
1240       }
1241
1242     } else {
1243       // first map.  we want the full thing.
1244       if (m->maps.count(m->get_last())) {
1245         for (map<int,OSDSession*>::iterator p = osd_sessions.begin();
1246              p != osd_sessions.end(); ++p) {
1247           OSDSession *s = p->second;
1248           _scan_requests(s, false, false, NULL, need_resend,
1249                          need_resend_linger, need_resend_command, sul);
1250         }
1251         ldout(cct, 3) << "handle_osd_map decoding full epoch "
1252                       << m->get_last() << dendl;
1253         osdmap->decode(m->maps[m->get_last()]);
1254
1255         _scan_requests(homeless_session, false, false, NULL,
1256                        need_resend, need_resend_linger,
1257                        need_resend_command, sul);
1258       } else {
1259         ldout(cct, 3) << "handle_osd_map hmm, i want a full map, requesting"
1260                       << dendl;
1261         monc->sub_want("osdmap", 0, CEPH_SUBSCRIBE_ONETIME);
1262         monc->renew_subs();
1263       }
1264     }
1265   }
1266
1267   // make sure need_resend targets reflect latest map
1268   for (auto p = need_resend.begin(); p != need_resend.end(); ) {
1269     Op *op = p->second;
1270     if (op->target.epoch < osdmap->get_epoch()) {
1271       ldout(cct, 10) << __func__ << "  checking op " << p->first << dendl;
1272       int r = _calc_target(&op->target, nullptr);
1273       if (r == RECALC_OP_TARGET_POOL_DNE) {
1274         p = need_resend.erase(p);
1275         _check_op_pool_dne(op, nullptr);
1276       } else {
1277         ++p;
1278       }
1279     } else {
1280       ++p;
1281     }
1282   }
1283
1284   bool pauserd = osdmap->test_flag(CEPH_OSDMAP_PAUSERD);
1285   bool pausewr = osdmap->test_flag(CEPH_OSDMAP_PAUSEWR) || _osdmap_full_flag()
1286     || _osdmap_has_pool_full();
1287
1288   // was/is paused?
1289   if (was_pauserd || was_pausewr || pauserd || pausewr ||
1290       osdmap->get_epoch() < epoch_barrier) {
1291     _maybe_request_map();
1292   }
1293
1294   // resend requests
1295   for (map<ceph_tid_t, Op*>::iterator p = need_resend.begin();
1296        p != need_resend.end(); ++p) {
1297     Op *op = p->second;
1298     OSDSession *s = op->session;
1299     bool mapped_session = false;
1300     if (!s) {
1301       int r = _map_session(&op->target, &s, sul);
1302       assert(r == 0);
1303       mapped_session = true;
1304     } else {
1305       get_session(s);
1306     }
1307     OSDSession::unique_lock sl(s->lock);
1308     if (mapped_session) {
1309       _session_op_assign(s, op);
1310     }
1311     if (op->should_resend) {
1312       if (!op->session->is_homeless() && !op->target.paused) {
1313         logger->inc(l_osdc_op_resend);
1314         _send_op(op);
1315       }
1316     } else {
1317       _op_cancel_map_check(op);
1318       _cancel_linger_op(op);
1319     }
1320     sl.unlock();
1321     put_session(s);
1322   }
1323   for (list<LingerOp*>::iterator p = need_resend_linger.begin();
1324        p != need_resend_linger.end(); ++p) {
1325     LingerOp *op = *p;
1326     if (!op->session) {
1327       _calc_target(&op->target, nullptr);
1328       OSDSession *s = NULL;
1329       int const r = _get_session(op->target.osd, &s, sul);
1330       assert(r == 0);
1331       assert(s != NULL);
1332       op->session = s;
1333       put_session(s);
1334     }
1335     if (!op->session->is_homeless()) {
1336       logger->inc(l_osdc_linger_resend);
1337       _send_linger(op, sul);
1338     }
1339   }
1340   for (map<ceph_tid_t,CommandOp*>::iterator p = need_resend_command.begin();
1341        p != need_resend_command.end(); ++p) {
1342     CommandOp *c = p->second;
1343     if (c->target.osd >= 0) {
1344       _assign_command_session(c, sul);
1345       if (c->session && !c->session->is_homeless()) {
1346         _send_command(c);
1347       }
1348     }
1349   }
1350
1351   _dump_active();
1352
1353   // finish any Contexts that were waiting on a map update
1354   map<epoch_t,list< pair< Context*, int > > >::iterator p =
1355     waiting_for_map.begin();
1356   while (p != waiting_for_map.end() &&
1357          p->first <= osdmap->get_epoch()) {
1358     //go through the list and call the onfinish methods
1359     for (list<pair<Context*, int> >::iterator i = p->second.begin();
1360          i != p->second.end(); ++i) {
1361       i->first->complete(i->second);
1362     }
1363     waiting_for_map.erase(p++);
1364   }
1365
1366   monc->sub_got("osdmap", osdmap->get_epoch());
1367
1368   if (!waiting_for_map.empty()) {
1369     _maybe_request_map();
1370   }
1371 }
1372
1373 void Objecter::enable_blacklist_events()
1374 {
1375   unique_lock wl(rwlock);
1376
1377   blacklist_events_enabled = true;
1378 }
1379
1380 void Objecter::consume_blacklist_events(std::set<entity_addr_t> *events)
1381 {
1382   unique_lock wl(rwlock);
1383
1384   if (events->empty()) {
1385     events->swap(blacklist_events);
1386   } else {
1387     for (const auto &i : blacklist_events) {
1388       events->insert(i);
1389     }
1390     blacklist_events.clear();
1391   }
1392 }
1393
1394 void Objecter::emit_blacklist_events(const OSDMap::Incremental &inc)
1395 {
1396   if (!blacklist_events_enabled) {
1397     return;
1398   }
1399
1400   for (const auto &i : inc.new_blacklist) {
1401     blacklist_events.insert(i.first);
1402   }
1403 }
1404
1405 void Objecter::emit_blacklist_events(const OSDMap &old_osd_map,
1406                                      const OSDMap &new_osd_map)
1407 {
1408   if (!blacklist_events_enabled) {
1409     return;
1410   }
1411
1412   std::set<entity_addr_t> old_set;
1413   std::set<entity_addr_t> new_set;
1414
1415   old_osd_map.get_blacklist(&old_set);
1416   new_osd_map.get_blacklist(&new_set);
1417
1418   std::set<entity_addr_t> delta_set;
1419   std::set_difference(
1420       new_set.begin(), new_set.end(), old_set.begin(), old_set.end(),
1421       std::inserter(delta_set, delta_set.begin()));
1422   blacklist_events.insert(delta_set.begin(), delta_set.end());
1423 }
1424
1425 // op pool check
1426
1427 void Objecter::C_Op_Map_Latest::finish(int r)
1428 {
1429   if (r == -EAGAIN || r == -ECANCELED)
1430     return;
1431
1432   lgeneric_subdout(objecter->cct, objecter, 10)
1433     << "op_map_latest r=" << r << " tid=" << tid
1434     << " latest " << latest << dendl;
1435
1436   Objecter::unique_lock wl(objecter->rwlock);
1437
1438   map<ceph_tid_t, Op*>::iterator iter =
1439     objecter->check_latest_map_ops.find(tid);
1440   if (iter == objecter->check_latest_map_ops.end()) {
1441     lgeneric_subdout(objecter->cct, objecter, 10)
1442       << "op_map_latest op "<< tid << " not found" << dendl;
1443     return;
1444   }
1445
1446   Op *op = iter->second;
1447   objecter->check_latest_map_ops.erase(iter);
1448
1449   lgeneric_subdout(objecter->cct, objecter, 20)
1450     << "op_map_latest op "<< op << dendl;
1451
1452   if (op->map_dne_bound == 0)
1453     op->map_dne_bound = latest;
1454
1455   OSDSession::unique_lock sl(op->session->lock, defer_lock);
1456   objecter->_check_op_pool_dne(op, &sl);
1457
1458   op->put();
1459 }
1460
1461 int Objecter::pool_snap_by_name(int64_t poolid, const char *snap_name,
1462                                 snapid_t *snap) const
1463 {
1464   shared_lock rl(rwlock);
1465
1466   auto& pools = osdmap->get_pools();
1467   auto iter = pools.find(poolid);
1468   if (iter == pools.end()) {
1469     return -ENOENT;
1470   }
1471   const pg_pool_t& pg_pool = iter->second;
1472   for (auto p = pg_pool.snaps.begin();
1473        p != pg_pool.snaps.end();
1474        ++p) {
1475     if (p->second.name == snap_name) {
1476       *snap = p->first;
1477       return 0;
1478     }
1479   }
1480   return -ENOENT;
1481 }
1482
1483 int Objecter::pool_snap_get_info(int64_t poolid, snapid_t snap,
1484                                  pool_snap_info_t *info) const
1485 {
1486   shared_lock rl(rwlock);
1487
1488   auto& pools = osdmap->get_pools();
1489   auto iter = pools.find(poolid);
1490   if (iter == pools.end()) {
1491     return -ENOENT;
1492   }
1493   const pg_pool_t& pg_pool = iter->second;
1494   auto p = pg_pool.snaps.find(snap);
1495   if (p == pg_pool.snaps.end())
1496     return -ENOENT;
1497   *info = p->second;
1498
1499   return 0;
1500 }
1501
1502 int Objecter::pool_snap_list(int64_t poolid, vector<uint64_t> *snaps)
1503 {
1504   shared_lock rl(rwlock);
1505
1506   const pg_pool_t *pi = osdmap->get_pg_pool(poolid);
1507   if (!pi)
1508     return -ENOENT;
1509   for (map<snapid_t,pool_snap_info_t>::const_iterator p = pi->snaps.begin();
1510        p != pi->snaps.end();
1511        ++p) {
1512     snaps->push_back(p->first);
1513   }
1514   return 0;
1515 }
1516
1517 // sl may be unlocked.
1518 void Objecter::_check_op_pool_dne(Op *op, unique_lock *sl)
1519 {
1520   // rwlock is locked unique
1521
1522   if (op->target.pool_ever_existed) {
1523     // the pool previously existed and now it does not, which means it
1524     // was deleted.
1525     op->map_dne_bound = osdmap->get_epoch();
1526     ldout(cct, 10) << "check_op_pool_dne tid " << op->tid
1527                    << " pool previously exists but now does not"
1528                    << dendl;
1529   } else {
1530     ldout(cct, 10) << "check_op_pool_dne tid " << op->tid
1531                    << " current " << osdmap->get_epoch()
1532                    << " map_dne_bound " << op->map_dne_bound
1533                    << dendl;
1534   }
1535   if (op->map_dne_bound > 0) {
1536     if (osdmap->get_epoch() >= op->map_dne_bound) {
1537       // we had a new enough map
1538       ldout(cct, 10) << "check_op_pool_dne tid " << op->tid
1539                      << " concluding pool " << op->target.base_pgid.pool()
1540                      << " dne" << dendl;
1541       if (op->onfinish) {
1542         op->onfinish->complete(-ENOENT);
1543       }
1544
1545       OSDSession *s = op->session;
1546       if (s) {
1547         assert(s != NULL);
1548         assert(sl->mutex() == &s->lock);
1549         bool session_locked = sl->owns_lock();
1550         if (!session_locked) {
1551           sl->lock();
1552         }
1553         _finish_op(op, 0);
1554         if (!session_locked) {
1555           sl->unlock();
1556         }
1557       } else {
1558         _finish_op(op, 0);      // no session
1559       }
1560     }
1561   } else {
1562     _send_op_map_check(op);
1563   }
1564 }
1565
1566 void Objecter::_send_op_map_check(Op *op)
1567 {
1568   // rwlock is locked unique
1569   // ask the monitor
1570   if (check_latest_map_ops.count(op->tid) == 0) {
1571     op->get();
1572     check_latest_map_ops[op->tid] = op;
1573     C_Op_Map_Latest *c = new C_Op_Map_Latest(this, op->tid);
1574     monc->get_version("osdmap", &c->latest, NULL, c);
1575   }
1576 }
1577
1578 void Objecter::_op_cancel_map_check(Op *op)
1579 {
1580   // rwlock is locked unique
1581   map<ceph_tid_t, Op*>::iterator iter =
1582     check_latest_map_ops.find(op->tid);
1583   if (iter != check_latest_map_ops.end()) {
1584     Op *op = iter->second;
1585     op->put();
1586     check_latest_map_ops.erase(iter);
1587   }
1588 }
1589
1590 // linger pool check
1591
1592 void Objecter::C_Linger_Map_Latest::finish(int r)
1593 {
1594   if (r == -EAGAIN || r == -ECANCELED) {
1595     // ignore callback; we will retry in resend_mon_ops()
1596     return;
1597   }
1598
1599   unique_lock wl(objecter->rwlock);
1600
1601   map<uint64_t, LingerOp*>::iterator iter =
1602     objecter->check_latest_map_lingers.find(linger_id);
1603   if (iter == objecter->check_latest_map_lingers.end()) {
1604     return;
1605   }
1606
1607   LingerOp *op = iter->second;
1608   objecter->check_latest_map_lingers.erase(iter);
1609
1610   if (op->map_dne_bound == 0)
1611     op->map_dne_bound = latest;
1612
1613   bool unregister;
1614   objecter->_check_linger_pool_dne(op, &unregister);
1615
1616   if (unregister) {
1617     objecter->_linger_cancel(op);
1618   }
1619
1620   op->put();
1621 }
1622
1623 void Objecter::_check_linger_pool_dne(LingerOp *op, bool *need_unregister)
1624 {
1625   // rwlock is locked unique
1626
1627   *need_unregister = false;
1628
1629   if (op->register_gen > 0) {
1630     ldout(cct, 10) << "_check_linger_pool_dne linger_id " << op->linger_id
1631                    << " pool previously existed but now does not"
1632                    << dendl;
1633     op->map_dne_bound = osdmap->get_epoch();
1634   } else {
1635     ldout(cct, 10) << "_check_linger_pool_dne linger_id " << op->linger_id
1636                    << " current " << osdmap->get_epoch()
1637                    << " map_dne_bound " << op->map_dne_bound
1638                    << dendl;
1639   }
1640   if (op->map_dne_bound > 0) {
1641     if (osdmap->get_epoch() >= op->map_dne_bound) {
1642       if (op->on_reg_commit) {
1643         op->on_reg_commit->complete(-ENOENT);
1644       }
1645       *need_unregister = true;
1646     }
1647   } else {
1648     _send_linger_map_check(op);
1649   }
1650 }
1651
1652 void Objecter::_send_linger_map_check(LingerOp *op)
1653 {
1654   // ask the monitor
1655   if (check_latest_map_lingers.count(op->linger_id) == 0) {
1656     op->get();
1657     check_latest_map_lingers[op->linger_id] = op;
1658     C_Linger_Map_Latest *c = new C_Linger_Map_Latest(this, op->linger_id);
1659     monc->get_version("osdmap", &c->latest, NULL, c);
1660   }
1661 }
1662
1663 void Objecter::_linger_cancel_map_check(LingerOp *op)
1664 {
1665   // rwlock is locked unique
1666
1667   map<uint64_t, LingerOp*>::iterator iter =
1668     check_latest_map_lingers.find(op->linger_id);
1669   if (iter != check_latest_map_lingers.end()) {
1670     LingerOp *op = iter->second;
1671     op->put();
1672     check_latest_map_lingers.erase(iter);
1673   }
1674 }
1675
1676 // command pool check
1677
1678 void Objecter::C_Command_Map_Latest::finish(int r)
1679 {
1680   if (r == -EAGAIN || r == -ECANCELED) {
1681     // ignore callback; we will retry in resend_mon_ops()
1682     return;
1683   }
1684
1685   unique_lock wl(objecter->rwlock);
1686
1687   map<uint64_t, CommandOp*>::iterator iter =
1688     objecter->check_latest_map_commands.find(tid);
1689   if (iter == objecter->check_latest_map_commands.end()) {
1690     return;
1691   }
1692
1693   CommandOp *c = iter->second;
1694   objecter->check_latest_map_commands.erase(iter);
1695
1696   if (c->map_dne_bound == 0)
1697     c->map_dne_bound = latest;
1698
1699   objecter->_check_command_map_dne(c);
1700
1701   c->put();
1702 }
1703
1704 void Objecter::_check_command_map_dne(CommandOp *c)
1705 {
1706   // rwlock is locked unique
1707
1708   ldout(cct, 10) << "_check_command_map_dne tid " << c->tid
1709                  << " current " << osdmap->get_epoch()
1710                  << " map_dne_bound " << c->map_dne_bound
1711                  << dendl;
1712   if (c->map_dne_bound > 0) {
1713     if (osdmap->get_epoch() >= c->map_dne_bound) {
1714       _finish_command(c, c->map_check_error, c->map_check_error_str);
1715     }
1716   } else {
1717     _send_command_map_check(c);
1718   }
1719 }
1720
1721 void Objecter::_send_command_map_check(CommandOp *c)
1722 {
1723   // rwlock is locked unique
1724
1725   // ask the monitor
1726   if (check_latest_map_commands.count(c->tid) == 0) {
1727     c->get();
1728     check_latest_map_commands[c->tid] = c;
1729     C_Command_Map_Latest *f = new C_Command_Map_Latest(this, c->tid);
1730     monc->get_version("osdmap", &f->latest, NULL, f);
1731   }
1732 }
1733
1734 void Objecter::_command_cancel_map_check(CommandOp *c)
1735 {
1736   // rwlock is locked uniqe
1737
1738   map<uint64_t, CommandOp*>::iterator iter =
1739     check_latest_map_commands.find(c->tid);
1740   if (iter != check_latest_map_commands.end()) {
1741     CommandOp *c = iter->second;
1742     c->put();
1743     check_latest_map_commands.erase(iter);
1744   }
1745 }
1746
1747
1748 /**
1749  * Look up OSDSession by OSD id.
1750  *
1751  * @returns 0 on success, or -EAGAIN if the lock context requires
1752  * promotion to write.
1753  */
1754 int Objecter::_get_session(int osd, OSDSession **session, shunique_lock& sul)
1755 {
1756   assert(sul && sul.mutex() == &rwlock);
1757
1758   if (osd < 0) {
1759     *session = homeless_session;
1760     ldout(cct, 20) << __func__ << " osd=" << osd << " returning homeless"
1761                    << dendl;
1762     return 0;
1763   }
1764
1765   map<int,OSDSession*>::iterator p = osd_sessions.find(osd);
1766   if (p != osd_sessions.end()) {
1767     OSDSession *s = p->second;
1768     s->get();
1769     *session = s;
1770     ldout(cct, 20) << __func__ << " s=" << s << " osd=" << osd << " "
1771                    << s->get_nref() << dendl;
1772     return 0;
1773   }
1774   if (!sul.owns_lock()) {
1775     return -EAGAIN;
1776   }
1777   OSDSession *s = new OSDSession(cct, osd);
1778   osd_sessions[osd] = s;
1779   s->con = messenger->get_connection(osdmap->get_inst(osd));
1780   s->con->set_priv(s->get());
1781   logger->inc(l_osdc_osd_session_open);
1782   logger->set(l_osdc_osd_sessions, osd_sessions.size());
1783   s->get();
1784   *session = s;
1785   ldout(cct, 20) << __func__ << " s=" << s << " osd=" << osd << " "
1786                  << s->get_nref() << dendl;
1787   return 0;
1788 }
1789
1790 void Objecter::put_session(Objecter::OSDSession *s)
1791 {
1792   if (s && !s->is_homeless()) {
1793     ldout(cct, 20) << __func__ << " s=" << s << " osd=" << s->osd << " "
1794                    << s->get_nref() << dendl;
1795     s->put();
1796   }
1797 }
1798
1799 void Objecter::get_session(Objecter::OSDSession *s)
1800 {
1801   assert(s != NULL);
1802
1803   if (!s->is_homeless()) {
1804     ldout(cct, 20) << __func__ << " s=" << s << " osd=" << s->osd << " "
1805                    << s->get_nref() << dendl;
1806     s->get();
1807   }
1808 }
1809
1810 void Objecter::_reopen_session(OSDSession *s)
1811 {
1812   // s->lock is locked
1813
1814   entity_inst_t inst = osdmap->get_inst(s->osd);
1815   ldout(cct, 10) << "reopen_session osd." << s->osd << " session, addr now "
1816                  << inst << dendl;
1817   if (s->con) {
1818     s->con->set_priv(NULL);
1819     s->con->mark_down();
1820     logger->inc(l_osdc_osd_session_close);
1821   }
1822   s->con = messenger->get_connection(inst);
1823   s->con->set_priv(s->get());
1824   s->incarnation++;
1825   logger->inc(l_osdc_osd_session_open);
1826 }
1827
1828 void Objecter::close_session(OSDSession *s)
1829 {
1830   // rwlock is locked unique
1831
1832   ldout(cct, 10) << "close_session for osd." << s->osd << dendl;
1833   if (s->con) {
1834     s->con->set_priv(NULL);
1835     s->con->mark_down();
1836     logger->inc(l_osdc_osd_session_close);
1837   }
1838   OSDSession::unique_lock sl(s->lock);
1839
1840   std::list<LingerOp*> homeless_lingers;
1841   std::list<CommandOp*> homeless_commands;
1842   std::list<Op*> homeless_ops;
1843
1844   while (!s->linger_ops.empty()) {
1845     std::map<uint64_t, LingerOp*>::iterator i = s->linger_ops.begin();
1846     ldout(cct, 10) << " linger_op " << i->first << dendl;
1847     homeless_lingers.push_back(i->second);
1848     _session_linger_op_remove(s, i->second);
1849   }
1850
1851   while (!s->ops.empty()) {
1852     std::map<ceph_tid_t, Op*>::iterator i = s->ops.begin();
1853     ldout(cct, 10) << " op " << i->first << dendl;
1854     homeless_ops.push_back(i->second);
1855     _session_op_remove(s, i->second);
1856   }
1857
1858   while (!s->command_ops.empty()) {
1859     std::map<ceph_tid_t, CommandOp*>::iterator i = s->command_ops.begin();
1860     ldout(cct, 10) << " command_op " << i->first << dendl;
1861     homeless_commands.push_back(i->second);
1862     _session_command_op_remove(s, i->second);
1863   }
1864
1865   osd_sessions.erase(s->osd);
1866   sl.unlock();
1867   put_session(s);
1868
1869   // Assign any leftover ops to the homeless session
1870   {
1871     OSDSession::unique_lock hsl(homeless_session->lock);
1872     for (std::list<LingerOp*>::iterator i = homeless_lingers.begin();
1873          i != homeless_lingers.end(); ++i) {
1874       _session_linger_op_assign(homeless_session, *i);
1875     }
1876     for (std::list<Op*>::iterator i = homeless_ops.begin();
1877          i != homeless_ops.end(); ++i) {
1878       _session_op_assign(homeless_session, *i);
1879     }
1880     for (std::list<CommandOp*>::iterator i = homeless_commands.begin();
1881          i != homeless_commands.end(); ++i) {
1882       _session_command_op_assign(homeless_session, *i);
1883     }
1884   }
1885
1886   logger->set(l_osdc_osd_sessions, osd_sessions.size());
1887 }
1888
1889 void Objecter::wait_for_osd_map()
1890 {
1891   unique_lock l(rwlock);
1892   if (osdmap->get_epoch()) {
1893     l.unlock();
1894     return;
1895   }
1896
1897   // Leave this since it goes with C_SafeCond
1898   Mutex lock("");
1899   Cond cond;
1900   bool done;
1901   lock.Lock();
1902   C_SafeCond *context = new C_SafeCond(&lock, &cond, &done, NULL);
1903   waiting_for_map[0].push_back(pair<Context*, int>(context, 0));
1904   l.unlock();
1905   while (!done)
1906     cond.Wait(lock);
1907   lock.Unlock();
1908 }
1909
1910 struct C_Objecter_GetVersion : public Context {
1911   Objecter *objecter;
1912   uint64_t oldest, newest;
1913   Context *fin;
1914   C_Objecter_GetVersion(Objecter *o, Context *c)
1915     : objecter(o), oldest(0), newest(0), fin(c) {}
1916   void finish(int r) override {
1917     if (r >= 0) {
1918       objecter->get_latest_version(oldest, newest, fin);
1919     } else if (r == -EAGAIN) { // try again as instructed
1920       objecter->wait_for_latest_osdmap(fin);
1921     } else {
1922       // it doesn't return any other error codes!
1923       ceph_abort();
1924     }
1925   }
1926 };
1927
1928 void Objecter::wait_for_latest_osdmap(Context *fin)
1929 {
1930   ldout(cct, 10) << __func__ << dendl;
1931   C_Objecter_GetVersion *c = new C_Objecter_GetVersion(this, fin);
1932   monc->get_version("osdmap", &c->newest, &c->oldest, c);
1933 }
1934
1935 void Objecter::get_latest_version(epoch_t oldest, epoch_t newest, Context *fin)
1936 {
1937   unique_lock wl(rwlock);
1938   _get_latest_version(oldest, newest, fin);
1939 }
1940
1941 void Objecter::_get_latest_version(epoch_t oldest, epoch_t newest,
1942                                    Context *fin)
1943 {
1944   // rwlock is locked unique
1945   if (osdmap->get_epoch() >= newest) {
1946   ldout(cct, 10) << __func__ << " latest " << newest << ", have it" << dendl;
1947     if (fin)
1948       fin->complete(0);
1949     return;
1950   }
1951
1952   ldout(cct, 10) << __func__ << " latest " << newest << ", waiting" << dendl;
1953   _wait_for_new_map(fin, newest, 0);
1954 }
1955
1956 void Objecter::maybe_request_map()
1957 {
1958   shared_lock rl(rwlock);
1959   _maybe_request_map();
1960 }
1961
1962 void Objecter::_maybe_request_map()
1963 {
1964   // rwlock is locked
1965   int flag = 0;
1966   if (_osdmap_full_flag()
1967       || osdmap->test_flag(CEPH_OSDMAP_PAUSERD)
1968       || osdmap->test_flag(CEPH_OSDMAP_PAUSEWR)) {
1969     ldout(cct, 10) << "_maybe_request_map subscribing (continuous) to next "
1970       "osd map (FULL flag is set)" << dendl;
1971   } else {
1972     ldout(cct, 10)
1973       << "_maybe_request_map subscribing (onetime) to next osd map" << dendl;
1974     flag = CEPH_SUBSCRIBE_ONETIME;
1975   }
1976   epoch_t epoch = osdmap->get_epoch() ? osdmap->get_epoch()+1 : 0;
1977   if (monc->sub_want("osdmap", epoch, flag)) {
1978     monc->renew_subs();
1979   }
1980 }
1981
1982 void Objecter::_wait_for_new_map(Context *c, epoch_t epoch, int err)
1983 {
1984   // rwlock is locked unique
1985   waiting_for_map[epoch].push_back(pair<Context *, int>(c, err));
1986   _maybe_request_map();
1987 }
1988
1989
1990 /**
1991  * Use this together with wait_for_map: this is a pre-check to avoid
1992  * allocating a Context for wait_for_map if we can see that we
1993  * definitely already have the epoch.
1994  *
1995  * This does *not* replace the need to handle the return value of
1996  * wait_for_map: just because we don't have it in this pre-check
1997  * doesn't mean we won't have it when calling back into wait_for_map,
1998  * since the objecter lock is dropped in between.
1999  */
2000 bool Objecter::have_map(const epoch_t epoch)
2001 {
2002   shared_lock rl(rwlock);
2003   if (osdmap->get_epoch() >= epoch) {
2004     return true;
2005   } else {
2006     return false;
2007   }
2008 }
2009
2010 bool Objecter::wait_for_map(epoch_t epoch, Context *c, int err)
2011 {
2012   unique_lock wl(rwlock);
2013   if (osdmap->get_epoch() >= epoch) {
2014     return true;
2015   }
2016   _wait_for_new_map(c, epoch, err);
2017   return false;
2018 }
2019
2020 void Objecter::kick_requests(OSDSession *session)
2021 {
2022   ldout(cct, 10) << "kick_requests for osd." << session->osd << dendl;
2023
2024   map<uint64_t, LingerOp *> lresend;
2025   unique_lock wl(rwlock);
2026
2027   OSDSession::unique_lock sl(session->lock);
2028   _kick_requests(session, lresend);
2029   sl.unlock();
2030
2031   _linger_ops_resend(lresend, wl);
2032 }
2033
2034 void Objecter::_kick_requests(OSDSession *session,
2035                               map<uint64_t, LingerOp *>& lresend)
2036 {
2037   // rwlock is locked unique
2038
2039   // clear backoffs
2040   session->backoffs.clear();
2041   session->backoffs_by_id.clear();
2042
2043   // resend ops
2044   map<ceph_tid_t,Op*> resend;  // resend in tid order
2045   for (map<ceph_tid_t, Op*>::iterator p = session->ops.begin();
2046        p != session->ops.end();) {
2047     Op *op = p->second;
2048     ++p;
2049     logger->inc(l_osdc_op_resend);
2050     if (op->should_resend) {
2051       if (!op->target.paused)
2052         resend[op->tid] = op;
2053     } else {
2054       _op_cancel_map_check(op);
2055       _cancel_linger_op(op);
2056     }
2057   }
2058
2059   while (!resend.empty()) {
2060     _send_op(resend.begin()->second);
2061     resend.erase(resend.begin());
2062   }
2063
2064   // resend lingers
2065   for (map<ceph_tid_t, LingerOp*>::iterator j = session->linger_ops.begin();
2066        j != session->linger_ops.end(); ++j) {
2067     LingerOp *op = j->second;
2068     op->get();
2069     logger->inc(l_osdc_linger_resend);
2070     assert(lresend.count(j->first) == 0);
2071     lresend[j->first] = op;
2072   }
2073
2074   // resend commands
2075   map<uint64_t,CommandOp*> cresend;  // resend in order
2076   for (map<ceph_tid_t, CommandOp*>::iterator k = session->command_ops.begin();
2077        k != session->command_ops.end(); ++k) {
2078     logger->inc(l_osdc_command_resend);
2079     cresend[k->first] = k->second;
2080   }
2081   while (!cresend.empty()) {
2082     _send_command(cresend.begin()->second);
2083     cresend.erase(cresend.begin());
2084   }
2085 }
2086
2087 void Objecter::_linger_ops_resend(map<uint64_t, LingerOp *>& lresend,
2088                                   unique_lock& ul)
2089 {
2090   assert(ul.owns_lock());
2091   shunique_lock sul(std::move(ul));
2092   while (!lresend.empty()) {
2093     LingerOp *op = lresend.begin()->second;
2094     if (!op->canceled) {
2095       _send_linger(op, sul);
2096     }
2097     op->put();
2098     lresend.erase(lresend.begin());
2099   }
2100   ul = unique_lock(sul.release_to_unique());
2101 }
2102
2103 void Objecter::start_tick()
2104 {
2105   assert(tick_event == 0);
2106   tick_event =
2107     timer.add_event(ceph::make_timespan(cct->_conf->objecter_tick_interval),
2108                     &Objecter::tick, this);
2109 }
2110
2111 void Objecter::tick()
2112 {
2113   shared_lock rl(rwlock);
2114
2115   ldout(cct, 10) << "tick" << dendl;
2116
2117   // we are only called by C_Tick
2118   tick_event = 0;
2119
2120   if (!initialized) {
2121     // we raced with shutdown
2122     ldout(cct, 10) << __func__ << " raced with shutdown" << dendl;
2123     return;
2124   }
2125
2126   set<OSDSession*> toping;
2127
2128
2129   // look for laggy requests
2130   auto cutoff = ceph::mono_clock::now();
2131   cutoff -= ceph::make_timespan(cct->_conf->objecter_timeout);  // timeout
2132
2133   unsigned laggy_ops = 0;
2134
2135   for (map<int,OSDSession*>::iterator siter = osd_sessions.begin();
2136        siter != osd_sessions.end(); ++siter) {
2137     OSDSession *s = siter->second;
2138     OSDSession::lock_guard l(s->lock);
2139     bool found = false;
2140     for (map<ceph_tid_t,Op*>::iterator p = s->ops.begin();
2141         p != s->ops.end();
2142         ++p) {
2143       Op *op = p->second;
2144       assert(op->session);
2145       if (op->stamp < cutoff) {
2146         ldout(cct, 2) << " tid " << p->first << " on osd." << op->session->osd
2147                       << " is laggy" << dendl;
2148         found = true;
2149         ++laggy_ops;
2150       }
2151     }
2152     for (map<uint64_t,LingerOp*>::iterator p = s->linger_ops.begin();
2153         p != s->linger_ops.end();
2154         ++p) {
2155       LingerOp *op = p->second;
2156       LingerOp::unique_lock wl(op->watch_lock);
2157       assert(op->session);
2158       ldout(cct, 10) << " pinging osd that serves lingering tid " << p->first
2159                      << " (osd." << op->session->osd << ")" << dendl;
2160       found = true;
2161       if (op->is_watch && op->registered && !op->last_error)
2162         _send_linger_ping(op);
2163     }
2164     for (map<uint64_t,CommandOp*>::iterator p = s->command_ops.begin();
2165         p != s->command_ops.end();
2166         ++p) {
2167       CommandOp *op = p->second;
2168       assert(op->session);
2169       ldout(cct, 10) << " pinging osd that serves command tid " << p->first
2170                      << " (osd." << op->session->osd << ")" << dendl;
2171       found = true;
2172     }
2173     if (found)
2174       toping.insert(s);
2175   }
2176   if (num_homeless_ops || !toping.empty()) {
2177     _maybe_request_map();
2178   }
2179
2180   logger->set(l_osdc_op_laggy, laggy_ops);
2181   logger->set(l_osdc_osd_laggy, toping.size());
2182
2183   if (!toping.empty()) {
2184     // send a ping to these osds, to ensure we detect any session resets
2185     // (osd reply message policy is lossy)
2186     for (set<OSDSession*>::const_iterator i = toping.begin();
2187          i != toping.end();
2188          ++i) {
2189       (*i)->con->send_message(new MPing);
2190     }
2191   }
2192
2193   // Make sure we don't resechedule if we wake up after shutdown
2194   if (initialized) {
2195     tick_event = timer.reschedule_me(ceph::make_timespan(
2196                                        cct->_conf->objecter_tick_interval));
2197   }
2198 }
2199
2200 void Objecter::resend_mon_ops()
2201 {
2202   unique_lock wl(rwlock);
2203
2204   ldout(cct, 10) << "resend_mon_ops" << dendl;
2205
2206   for (map<ceph_tid_t,PoolStatOp*>::iterator p = poolstat_ops.begin();
2207        p != poolstat_ops.end();
2208        ++p) {
2209     _poolstat_submit(p->second);
2210     logger->inc(l_osdc_poolstat_resend);
2211   }
2212
2213   for (map<ceph_tid_t,StatfsOp*>::iterator p = statfs_ops.begin();
2214        p != statfs_ops.end();
2215        ++p) {
2216     _fs_stats_submit(p->second);
2217     logger->inc(l_osdc_statfs_resend);
2218   }
2219
2220   for (map<ceph_tid_t,PoolOp*>::iterator p = pool_ops.begin();
2221        p != pool_ops.end();
2222        ++p) {
2223     _pool_op_submit(p->second);
2224     logger->inc(l_osdc_poolop_resend);
2225   }
2226
2227   for (map<ceph_tid_t, Op*>::iterator p = check_latest_map_ops.begin();
2228        p != check_latest_map_ops.end();
2229        ++p) {
2230     C_Op_Map_Latest *c = new C_Op_Map_Latest(this, p->second->tid);
2231     monc->get_version("osdmap", &c->latest, NULL, c);
2232   }
2233
2234   for (map<uint64_t, LingerOp*>::iterator p = check_latest_map_lingers.begin();
2235        p != check_latest_map_lingers.end();
2236        ++p) {
2237     C_Linger_Map_Latest *c
2238       = new C_Linger_Map_Latest(this, p->second->linger_id);
2239     monc->get_version("osdmap", &c->latest, NULL, c);
2240   }
2241
2242   for (map<uint64_t, CommandOp*>::iterator p
2243          = check_latest_map_commands.begin();
2244        p != check_latest_map_commands.end();
2245        ++p) {
2246     C_Command_Map_Latest *c = new C_Command_Map_Latest(this, p->second->tid);
2247     monc->get_version("osdmap", &c->latest, NULL, c);
2248   }
2249 }
2250
2251 // read | write ---------------------------
2252
2253 void Objecter::op_submit(Op *op, ceph_tid_t *ptid, int *ctx_budget)
2254 {
2255   shunique_lock rl(rwlock, ceph::acquire_shared);
2256   ceph_tid_t tid = 0;
2257   if (!ptid)
2258     ptid = &tid;
2259   op->trace.event("op submit");
2260   _op_submit_with_budget(op, rl, ptid, ctx_budget);
2261 }
2262
2263 void Objecter::_op_submit_with_budget(Op *op, shunique_lock& sul,
2264                                       ceph_tid_t *ptid,
2265                                       int *ctx_budget)
2266 {
2267   assert(initialized);
2268
2269   assert(op->ops.size() == op->out_bl.size());
2270   assert(op->ops.size() == op->out_rval.size());
2271   assert(op->ops.size() == op->out_handler.size());
2272
2273   // throttle.  before we look at any state, because
2274   // _take_op_budget() may drop our lock while it blocks.
2275   if (!op->ctx_budgeted || (ctx_budget && (*ctx_budget == -1))) {
2276     int op_budget = _take_op_budget(op, sul);
2277     // take and pass out the budget for the first OP
2278     // in the context session
2279     if (ctx_budget && (*ctx_budget == -1)) {
2280       *ctx_budget = op_budget;
2281     }
2282   }
2283
2284   if (osd_timeout > timespan(0)) {
2285     if (op->tid == 0)
2286       op->tid = ++last_tid;
2287     auto tid = op->tid;
2288     op->ontimeout = timer.add_event(osd_timeout,
2289                                     [this, tid]() {
2290                                       op_cancel(tid, -ETIMEDOUT); });
2291   }
2292
2293   _op_submit(op, sul, ptid);
2294 }
2295
2296 void Objecter::_send_op_account(Op *op)
2297 {
2298   inflight_ops++;
2299
2300   // add to gather set(s)
2301   if (op->onfinish) {
2302     num_in_flight++;
2303   } else {
2304     ldout(cct, 20) << " note: not requesting reply" << dendl;
2305   }
2306
2307   logger->inc(l_osdc_op_active);
2308   logger->inc(l_osdc_op);
2309
2310   if ((op->target.flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE)) ==
2311       (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE))
2312     logger->inc(l_osdc_op_rmw);
2313   else if (op->target.flags & CEPH_OSD_FLAG_WRITE)
2314     logger->inc(l_osdc_op_w);
2315   else if (op->target.flags & CEPH_OSD_FLAG_READ)
2316     logger->inc(l_osdc_op_r);
2317
2318   if (op->target.flags & CEPH_OSD_FLAG_PGOP)
2319     logger->inc(l_osdc_op_pg);
2320
2321   for (vector<OSDOp>::iterator p = op->ops.begin(); p != op->ops.end(); ++p) {
2322     int code = l_osdc_osdop_other;
2323     switch (p->op.op) {
2324     case CEPH_OSD_OP_STAT: code = l_osdc_osdop_stat; break;
2325     case CEPH_OSD_OP_CREATE: code = l_osdc_osdop_create; break;
2326     case CEPH_OSD_OP_READ: code = l_osdc_osdop_read; break;
2327     case CEPH_OSD_OP_WRITE: code = l_osdc_osdop_write; break;
2328     case CEPH_OSD_OP_WRITEFULL: code = l_osdc_osdop_writefull; break;
2329     case CEPH_OSD_OP_WRITESAME: code = l_osdc_osdop_writesame; break;
2330     case CEPH_OSD_OP_APPEND: code = l_osdc_osdop_append; break;
2331     case CEPH_OSD_OP_ZERO: code = l_osdc_osdop_zero; break;
2332     case CEPH_OSD_OP_TRUNCATE: code = l_osdc_osdop_truncate; break;
2333     case CEPH_OSD_OP_DELETE: code = l_osdc_osdop_delete; break;
2334     case CEPH_OSD_OP_MAPEXT: code = l_osdc_osdop_mapext; break;
2335     case CEPH_OSD_OP_SPARSE_READ: code = l_osdc_osdop_sparse_read; break;
2336     case CEPH_OSD_OP_GETXATTR: code = l_osdc_osdop_getxattr; break;
2337     case CEPH_OSD_OP_SETXATTR: code = l_osdc_osdop_setxattr; break;
2338     case CEPH_OSD_OP_CMPXATTR: code = l_osdc_osdop_cmpxattr; break;
2339     case CEPH_OSD_OP_RMXATTR: code = l_osdc_osdop_rmxattr; break;
2340     case CEPH_OSD_OP_RESETXATTRS: code = l_osdc_osdop_resetxattrs; break;
2341     case CEPH_OSD_OP_TMAPUP: code = l_osdc_osdop_tmap_up; break;
2342     case CEPH_OSD_OP_TMAPPUT: code = l_osdc_osdop_tmap_put; break;
2343     case CEPH_OSD_OP_TMAPGET: code = l_osdc_osdop_tmap_get; break;
2344
2345     // OMAP read operations
2346     case CEPH_OSD_OP_OMAPGETVALS:
2347     case CEPH_OSD_OP_OMAPGETKEYS:
2348     case CEPH_OSD_OP_OMAPGETHEADER:
2349     case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
2350     case CEPH_OSD_OP_OMAP_CMP: code = l_osdc_osdop_omap_rd; break;
2351
2352     // OMAP write operations
2353     case CEPH_OSD_OP_OMAPSETVALS:
2354     case CEPH_OSD_OP_OMAPSETHEADER: code = l_osdc_osdop_omap_wr; break;
2355
2356     // OMAP del operations
2357     case CEPH_OSD_OP_OMAPCLEAR:
2358     case CEPH_OSD_OP_OMAPRMKEYS: code = l_osdc_osdop_omap_del; break;
2359
2360     case CEPH_OSD_OP_CALL: code = l_osdc_osdop_call; break;
2361     case CEPH_OSD_OP_WATCH: code = l_osdc_osdop_watch; break;
2362     case CEPH_OSD_OP_NOTIFY: code = l_osdc_osdop_notify; break;
2363     }
2364     if (code)
2365       logger->inc(code);
2366   }
2367 }
2368
2369 void Objecter::_op_submit(Op *op, shunique_lock& sul, ceph_tid_t *ptid)
2370 {
2371   // rwlock is locked
2372
2373   ldout(cct, 10) << __func__ << " op " << op << dendl;
2374
2375   // pick target
2376   assert(op->session == NULL);
2377   OSDSession *s = NULL;
2378
2379   bool check_for_latest_map = _calc_target(&op->target, nullptr)
2380     == RECALC_OP_TARGET_POOL_DNE;
2381
2382   // Try to get a session, including a retry if we need to take write lock
2383   int r = _get_session(op->target.osd, &s, sul);
2384   if (r == -EAGAIN ||
2385       (check_for_latest_map && sul.owns_lock_shared())) {
2386     epoch_t orig_epoch = osdmap->get_epoch();
2387     sul.unlock();
2388     if (cct->_conf->objecter_debug_inject_relock_delay) {
2389       sleep(1);
2390     }
2391     sul.lock();
2392     if (orig_epoch != osdmap->get_epoch()) {
2393       // map changed; recalculate mapping
2394       ldout(cct, 10) << __func__ << " relock raced with osdmap, recalc target"
2395                      << dendl;
2396       check_for_latest_map = _calc_target(&op->target, nullptr)
2397         == RECALC_OP_TARGET_POOL_DNE;
2398       if (s) {
2399         put_session(s);
2400         s = NULL;
2401         r = -EAGAIN;
2402       }
2403     }
2404   }
2405   if (r == -EAGAIN) {
2406     assert(s == NULL);
2407     r = _get_session(op->target.osd, &s, sul);
2408   }
2409   assert(r == 0);
2410   assert(s);  // may be homeless
2411
2412   _send_op_account(op);
2413
2414   // send?
2415
2416   assert(op->target.flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE));
2417
2418   if (osdmap_full_try) {
2419     op->target.flags |= CEPH_OSD_FLAG_FULL_TRY;
2420   }
2421
2422   bool need_send = false;
2423
2424   if (osdmap->get_epoch() < epoch_barrier) {
2425     ldout(cct, 10) << " barrier, paused " << op << " tid " << op->tid
2426                    << dendl;
2427     op->target.paused = true;
2428     _maybe_request_map();
2429   } else if ((op->target.flags & CEPH_OSD_FLAG_WRITE) &&
2430              osdmap->test_flag(CEPH_OSDMAP_PAUSEWR)) {
2431     ldout(cct, 10) << " paused modify " << op << " tid " << op->tid
2432                    << dendl;
2433     op->target.paused = true;
2434     _maybe_request_map();
2435   } else if ((op->target.flags & CEPH_OSD_FLAG_READ) &&
2436              osdmap->test_flag(CEPH_OSDMAP_PAUSERD)) {
2437     ldout(cct, 10) << " paused read " << op << " tid " << op->tid
2438                    << dendl;
2439     op->target.paused = true;
2440     _maybe_request_map();
2441   } else if (op->respects_full() &&
2442              (_osdmap_full_flag() ||
2443               _osdmap_pool_full(op->target.base_oloc.pool))) {
2444     ldout(cct, 0) << " FULL, paused modify " << op << " tid "
2445                   << op->tid << dendl;
2446     op->target.paused = true;
2447     _maybe_request_map();
2448   } else if (!s->is_homeless()) {
2449     need_send = true;
2450   } else {
2451     _maybe_request_map();
2452   }
2453
2454   MOSDOp *m = NULL;
2455   if (need_send) {
2456     m = _prepare_osd_op(op);
2457   }
2458
2459   OSDSession::unique_lock sl(s->lock);
2460   if (op->tid == 0)
2461     op->tid = ++last_tid;
2462
2463   ldout(cct, 10) << "_op_submit oid " << op->target.base_oid
2464                  << " '" << op->target.base_oloc << "' '"
2465                  << op->target.target_oloc << "' " << op->ops << " tid "
2466                  << op->tid << " osd." << (!s->is_homeless() ? s->osd : -1)
2467                  << dendl;
2468
2469   _session_op_assign(s, op);
2470
2471   if (need_send) {
2472     _send_op(op, m);
2473   }
2474
2475   // Last chance to touch Op here, after giving up session lock it can
2476   // be freed at any time by response handler.
2477   ceph_tid_t tid = op->tid;
2478   if (check_for_latest_map) {
2479     _send_op_map_check(op);
2480   }
2481   if (ptid)
2482     *ptid = tid;
2483   op = NULL;
2484
2485   sl.unlock();
2486   put_session(s);
2487
2488   ldout(cct, 5) << num_in_flight << " in flight" << dendl;
2489 }
2490
2491 int Objecter::op_cancel(OSDSession *s, ceph_tid_t tid, int r)
2492 {
2493   assert(initialized);
2494
2495   OSDSession::unique_lock sl(s->lock);
2496
2497   map<ceph_tid_t, Op*>::iterator p = s->ops.find(tid);
2498   if (p == s->ops.end()) {
2499     ldout(cct, 10) << __func__ << " tid " << tid << " dne in session "
2500                    << s->osd << dendl;
2501     return -ENOENT;
2502   }
2503
2504   if (s->con) {
2505     ldout(cct, 20) << " revoking rx buffer for " << tid
2506                    << " on " << s->con << dendl;
2507     s->con->revoke_rx_buffer(tid);
2508   }
2509
2510   ldout(cct, 10) << __func__ << " tid " << tid << " in session " << s->osd
2511                  << dendl;
2512   Op *op = p->second;
2513   if (op->onfinish) {
2514     num_in_flight--;
2515     op->onfinish->complete(r);
2516     op->onfinish = NULL;
2517   }
2518   _op_cancel_map_check(op);
2519   _finish_op(op, r);
2520   sl.unlock();
2521
2522   return 0;
2523 }
2524
2525 int Objecter::op_cancel(ceph_tid_t tid, int r)
2526 {
2527   int ret = 0;
2528
2529   unique_lock wl(rwlock);
2530   ret = _op_cancel(tid, r);
2531
2532   return ret;
2533 }
2534
2535 int Objecter::_op_cancel(ceph_tid_t tid, int r)
2536 {
2537   int ret = 0;
2538
2539   ldout(cct, 5) << __func__ << ": cancelling tid " << tid << " r=" << r
2540                 << dendl;
2541
2542 start:
2543
2544   for (map<int, OSDSession *>::iterator siter = osd_sessions.begin();
2545        siter != osd_sessions.end(); ++siter) {
2546     OSDSession *s = siter->second;
2547     OSDSession::shared_lock sl(s->lock);
2548     if (s->ops.find(tid) != s->ops.end()) {
2549       sl.unlock();
2550       ret = op_cancel(s, tid, r);
2551       if (ret == -ENOENT) {
2552         /* oh no! raced, maybe tid moved to another session, restarting */
2553         goto start;
2554       }
2555       return ret;
2556     }
2557   }
2558
2559   ldout(cct, 5) << __func__ << ": tid " << tid
2560                 << " not found in live sessions" << dendl;
2561
2562   // Handle case where the op is in homeless session
2563   OSDSession::shared_lock sl(homeless_session->lock);
2564   if (homeless_session->ops.find(tid) != homeless_session->ops.end()) {
2565     sl.unlock();
2566     ret = op_cancel(homeless_session, tid, r);
2567     if (ret == -ENOENT) {
2568       /* oh no! raced, maybe tid moved to another session, restarting */
2569       goto start;
2570     } else {
2571       return ret;
2572     }
2573   } else {
2574     sl.unlock();
2575   }
2576
2577   ldout(cct, 5) << __func__ << ": tid " << tid
2578                 << " not found in homeless session" << dendl;
2579
2580   return ret;
2581 }
2582
2583
2584 epoch_t Objecter::op_cancel_writes(int r, int64_t pool)
2585 {
2586   unique_lock wl(rwlock);
2587
2588   std::vector<ceph_tid_t> to_cancel;
2589   bool found = false;
2590
2591   for (map<int, OSDSession *>::iterator siter = osd_sessions.begin();
2592        siter != osd_sessions.end(); ++siter) {
2593     OSDSession *s = siter->second;
2594     OSDSession::shared_lock sl(s->lock);
2595     for (map<ceph_tid_t, Op*>::iterator op_i = s->ops.begin();
2596          op_i != s->ops.end(); ++op_i) {
2597       if (op_i->second->target.flags & CEPH_OSD_FLAG_WRITE
2598           && (pool == -1 || op_i->second->target.target_oloc.pool == pool)) {
2599         to_cancel.push_back(op_i->first);
2600       }
2601     }
2602     sl.unlock();
2603
2604     for (std::vector<ceph_tid_t>::iterator titer = to_cancel.begin();
2605          titer != to_cancel.end();
2606          ++titer) {
2607       int cancel_result = op_cancel(s, *titer, r);
2608       // We hold rwlock across search and cancellation, so cancels
2609       // should always succeed
2610       assert(cancel_result == 0);
2611     }
2612     if (!found && to_cancel.size())
2613       found = true;
2614     to_cancel.clear();
2615   }
2616
2617   const epoch_t epoch = osdmap->get_epoch();
2618
2619   wl.unlock();
2620
2621   if (found) {
2622     return epoch;
2623   } else {
2624     return -1;
2625   }
2626 }
2627
2628 bool Objecter::is_pg_changed(
2629   int oldprimary,
2630   const vector<int>& oldacting,
2631   int newprimary,
2632   const vector<int>& newacting,
2633   bool any_change)
2634 {
2635   if (OSDMap::primary_changed(
2636         oldprimary,
2637         oldacting,
2638         newprimary,
2639         newacting))
2640     return true;
2641   if (any_change && oldacting != newacting)
2642     return true;
2643   return false;      // same primary (tho replicas may have changed)
2644 }
2645
2646 bool Objecter::target_should_be_paused(op_target_t *t)
2647 {
2648   const pg_pool_t *pi = osdmap->get_pg_pool(t->base_oloc.pool);
2649   bool pauserd = osdmap->test_flag(CEPH_OSDMAP_PAUSERD);
2650   bool pausewr = osdmap->test_flag(CEPH_OSDMAP_PAUSEWR) ||
2651     _osdmap_full_flag() || _osdmap_pool_full(*pi);
2652
2653   return (t->flags & CEPH_OSD_FLAG_READ && pauserd) ||
2654     (t->flags & CEPH_OSD_FLAG_WRITE && pausewr) ||
2655     (osdmap->get_epoch() < epoch_barrier);
2656 }
2657
2658 /**
2659  * Locking public accessor for _osdmap_full_flag
2660  */
2661 bool Objecter::osdmap_full_flag() const
2662 {
2663   shared_lock rl(rwlock);
2664
2665   return _osdmap_full_flag();
2666 }
2667
2668 bool Objecter::osdmap_pool_full(const int64_t pool_id) const
2669 {
2670   shared_lock rl(rwlock);
2671
2672   if (_osdmap_full_flag()) {
2673     return true;
2674   }
2675
2676   return _osdmap_pool_full(pool_id);
2677 }
2678
2679 bool Objecter::_osdmap_pool_full(const int64_t pool_id) const
2680 {
2681   const pg_pool_t *pool = osdmap->get_pg_pool(pool_id);
2682   if (pool == NULL) {
2683     ldout(cct, 4) << __func__ << ": DNE pool " << pool_id << dendl;
2684     return false;
2685   }
2686
2687   return _osdmap_pool_full(*pool);
2688 }
2689
2690 bool Objecter::_osdmap_has_pool_full() const
2691 {
2692   for (map<int64_t, pg_pool_t>::const_iterator it
2693          = osdmap->get_pools().begin();
2694        it != osdmap->get_pools().end(); ++it) {
2695     if (_osdmap_pool_full(it->second))
2696       return true;
2697   }
2698   return false;
2699 }
2700
2701 bool Objecter::_osdmap_pool_full(const pg_pool_t &p) const
2702 {
2703   return p.has_flag(pg_pool_t::FLAG_FULL) && honor_osdmap_full;
2704 }
2705
2706 /**
2707  * Wrapper around osdmap->test_flag for special handling of the FULL flag.
2708  */
2709 bool Objecter::_osdmap_full_flag() const
2710 {
2711   // Ignore the FULL flag if the caller has honor_osdmap_full
2712   return osdmap->test_flag(CEPH_OSDMAP_FULL) && honor_osdmap_full;
2713 }
2714
2715 void Objecter::update_pool_full_map(map<int64_t, bool>& pool_full_map)
2716 {
2717   for (map<int64_t, pg_pool_t>::const_iterator it
2718          = osdmap->get_pools().begin();
2719        it != osdmap->get_pools().end(); ++it) {
2720     if (pool_full_map.find(it->first) == pool_full_map.end()) {
2721       pool_full_map[it->first] = _osdmap_pool_full(it->second);
2722     } else {
2723       pool_full_map[it->first] = _osdmap_pool_full(it->second) ||
2724         pool_full_map[it->first];
2725     }
2726   }
2727 }
2728
2729 int64_t Objecter::get_object_hash_position(int64_t pool, const string& key,
2730                                            const string& ns)
2731 {
2732   shared_lock rl(rwlock);
2733   const pg_pool_t *p = osdmap->get_pg_pool(pool);
2734   if (!p)
2735     return -ENOENT;
2736   return p->hash_key(key, ns);
2737 }
2738
2739 int64_t Objecter::get_object_pg_hash_position(int64_t pool, const string& key,
2740                                               const string& ns)
2741 {
2742   shared_lock rl(rwlock);
2743   const pg_pool_t *p = osdmap->get_pg_pool(pool);
2744   if (!p)
2745     return -ENOENT;
2746   return p->raw_hash_to_pg(p->hash_key(key, ns));
2747 }
2748
2749 int Objecter::_calc_target(op_target_t *t, Connection *con, bool any_change)
2750 {
2751   // rwlock is locked
2752   bool is_read = t->flags & CEPH_OSD_FLAG_READ;
2753   bool is_write = t->flags & CEPH_OSD_FLAG_WRITE;
2754   t->epoch = osdmap->get_epoch();
2755   ldout(cct,20) << __func__ << " epoch " << t->epoch
2756                 << " base " << t->base_oid << " " << t->base_oloc
2757                 << " precalc_pgid " << (int)t->precalc_pgid
2758                 << " pgid " << t->base_pgid
2759                 << (is_read ? " is_read" : "")
2760                 << (is_write ? " is_write" : "")
2761                 << dendl;
2762
2763   const pg_pool_t *pi = osdmap->get_pg_pool(t->base_oloc.pool);
2764   if (!pi) {
2765     t->osd = -1;
2766     return RECALC_OP_TARGET_POOL_DNE;
2767   }
2768   ldout(cct,30) << __func__ << "  base pi " << pi
2769                 << " pg_num " << pi->get_pg_num() << dendl;
2770
2771   bool force_resend = false;
2772   if (osdmap->get_epoch() == pi->last_force_op_resend) {
2773     if (t->last_force_resend < pi->last_force_op_resend) {
2774       t->last_force_resend = pi->last_force_op_resend;
2775       force_resend = true;
2776     } else if (t->last_force_resend == 0) {
2777       force_resend = true;
2778     }
2779   }
2780
2781   // apply tiering
2782   t->target_oid = t->base_oid;
2783   t->target_oloc = t->base_oloc;
2784   if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
2785     if (is_read && pi->has_read_tier())
2786       t->target_oloc.pool = pi->read_tier;
2787     if (is_write && pi->has_write_tier())
2788       t->target_oloc.pool = pi->write_tier;
2789     pi = osdmap->get_pg_pool(t->target_oloc.pool);
2790     if (!pi) {
2791       t->osd = -1;
2792       return RECALC_OP_TARGET_POOL_DNE;
2793     }
2794   }
2795
2796   pg_t pgid;
2797   if (t->precalc_pgid) {
2798     assert(t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY);
2799     assert(t->base_oid.name.empty()); // make sure this is a pg op
2800     assert(t->base_oloc.pool == (int64_t)t->base_pgid.pool());
2801     pgid = t->base_pgid;
2802   } else {
2803     int ret = osdmap->object_locator_to_pg(t->target_oid, t->target_oloc,
2804                                            pgid);
2805     if (ret == -ENOENT) {
2806       t->osd = -1;
2807       return RECALC_OP_TARGET_POOL_DNE;
2808     }
2809   }
2810   ldout(cct,20) << __func__ << " target " << t->target_oid << " "
2811                 << t->target_oloc << " -> pgid " << pgid << dendl;
2812   ldout(cct,30) << __func__ << "  target pi " << pi
2813                 << " pg_num " << pi->get_pg_num() << dendl;
2814   t->pool_ever_existed = true;
2815
2816   int size = pi->size;
2817   int min_size = pi->min_size;
2818   unsigned pg_num = pi->get_pg_num();
2819   int up_primary, acting_primary;
2820   vector<int> up, acting;
2821   osdmap->pg_to_up_acting_osds(pgid, &up, &up_primary,
2822                                &acting, &acting_primary);
2823   bool sort_bitwise = osdmap->test_flag(CEPH_OSDMAP_SORTBITWISE);
2824   unsigned prev_seed = ceph_stable_mod(pgid.ps(), t->pg_num, t->pg_num_mask);
2825   pg_t prev_pgid(prev_seed, pgid.pool());
2826   if (any_change && PastIntervals::is_new_interval(
2827         t->acting_primary,
2828         acting_primary,
2829         t->acting,
2830         acting,
2831         t->up_primary,
2832         up_primary,
2833         t->up,
2834         up,
2835         t->size,
2836         size,
2837         t->min_size,
2838         min_size,
2839         t->pg_num,
2840         pg_num,
2841         t->sort_bitwise,
2842         sort_bitwise,
2843         prev_pgid)) {
2844     force_resend = true;
2845   }
2846
2847   bool unpaused = false;
2848   if (t->paused && !target_should_be_paused(t)) {
2849     t->paused = false;
2850     unpaused = true;
2851   }
2852
2853   bool legacy_change =
2854     t->pgid != pgid ||
2855       is_pg_changed(
2856         t->acting_primary, t->acting, acting_primary, acting,
2857         t->used_replica || any_change);
2858   bool split = false;
2859   if (t->pg_num) {
2860     split = prev_pgid.is_split(t->pg_num, pg_num, nullptr);
2861   }
2862
2863   if (legacy_change || split || force_resend) {
2864     t->pgid = pgid;
2865     t->acting = acting;
2866     t->acting_primary = acting_primary;
2867     t->up_primary = up_primary;
2868     t->up = up;
2869     t->size = size;
2870     t->min_size = min_size;
2871     t->pg_num = pg_num;
2872     t->pg_num_mask = pi->get_pg_num_mask();
2873     osdmap->get_primary_shard(
2874       pg_t(ceph_stable_mod(pgid.ps(), t->pg_num, t->pg_num_mask), pgid.pool()),
2875       &t->actual_pgid);
2876     t->sort_bitwise = sort_bitwise;
2877     ldout(cct, 10) << __func__ << " "
2878                    << " raw pgid " << pgid << " -> actual " << t->actual_pgid
2879                    << " acting " << acting
2880                    << " primary " << acting_primary << dendl;
2881     t->used_replica = false;
2882     if (acting_primary == -1) {
2883       t->osd = -1;
2884     } else {
2885       int osd;
2886       bool read = is_read && !is_write;
2887       if (read && (t->flags & CEPH_OSD_FLAG_BALANCE_READS)) {
2888         int p = rand() % acting.size();
2889         if (p)
2890           t->used_replica = true;
2891         osd = acting[p];
2892         ldout(cct, 10) << " chose random osd." << osd << " of " << acting
2893                        << dendl;
2894       } else if (read && (t->flags & CEPH_OSD_FLAG_LOCALIZE_READS) &&
2895                  acting.size() > 1) {
2896         // look for a local replica.  prefer the primary if the
2897         // distance is the same.
2898         int best = -1;
2899         int best_locality = 0;
2900         for (unsigned i = 0; i < acting.size(); ++i) {
2901           int locality = osdmap->crush->get_common_ancestor_distance(
2902                  cct, acting[i], crush_location);
2903           ldout(cct, 20) << __func__ << " localize: rank " << i
2904                          << " osd." << acting[i]
2905                          << " locality " << locality << dendl;
2906           if (i == 0 ||
2907               (locality >= 0 && best_locality >= 0 &&
2908                locality < best_locality) ||
2909               (best_locality < 0 && locality >= 0)) {
2910             best = i;
2911             best_locality = locality;
2912             if (i)
2913               t->used_replica = true;
2914           }
2915         }
2916         assert(best >= 0);
2917         osd = acting[best];
2918       } else {
2919         osd = acting_primary;
2920       }
2921       t->osd = osd;
2922     }
2923   }
2924   if (legacy_change || unpaused || force_resend) {
2925     return RECALC_OP_TARGET_NEED_RESEND;
2926   }
2927   if (split && con && con->has_features(CEPH_FEATUREMASK_RESEND_ON_SPLIT)) {
2928     return RECALC_OP_TARGET_NEED_RESEND;
2929   }
2930   return RECALC_OP_TARGET_NO_ACTION;
2931 }
2932
2933 int Objecter::_map_session(op_target_t *target, OSDSession **s,
2934                            shunique_lock& sul)
2935 {
2936   _calc_target(target, nullptr);
2937   return _get_session(target->osd, s, sul);
2938 }
2939
2940 void Objecter::_session_op_assign(OSDSession *to, Op *op)
2941 {
2942   // to->lock is locked
2943   assert(op->session == NULL);
2944   assert(op->tid);
2945
2946   get_session(to);
2947   op->session = to;
2948   to->ops[op->tid] = op;
2949
2950   if (to->is_homeless()) {
2951     num_homeless_ops++;
2952   }
2953
2954   ldout(cct, 15) << __func__ << " " << to->osd << " " << op->tid << dendl;
2955 }
2956
2957 void Objecter::_session_op_remove(OSDSession *from, Op *op)
2958 {
2959   assert(op->session == from);
2960   // from->lock is locked
2961
2962   if (from->is_homeless()) {
2963     num_homeless_ops--;
2964   }
2965
2966   from->ops.erase(op->tid);
2967   put_session(from);
2968   op->session = NULL;
2969
2970   ldout(cct, 15) << __func__ << " " << from->osd << " " << op->tid << dendl;
2971 }
2972
2973 void Objecter::_session_linger_op_assign(OSDSession *to, LingerOp *op)
2974 {
2975   // to lock is locked unique
2976   assert(op->session == NULL);
2977
2978   if (to->is_homeless()) {
2979     num_homeless_ops++;
2980   }
2981
2982   get_session(to);
2983   op->session = to;
2984   to->linger_ops[op->linger_id] = op;
2985
2986   ldout(cct, 15) << __func__ << " " << to->osd << " " << op->linger_id
2987                  << dendl;
2988 }
2989
2990 void Objecter::_session_linger_op_remove(OSDSession *from, LingerOp *op)
2991 {
2992   assert(from == op->session);
2993   // from->lock is locked unique
2994
2995   if (from->is_homeless()) {
2996     num_homeless_ops--;
2997   }
2998
2999   from->linger_ops.erase(op->linger_id);
3000   put_session(from);
3001   op->session = NULL;
3002
3003   ldout(cct, 15) << __func__ << " " << from->osd << " " << op->linger_id
3004                  << dendl;
3005 }
3006
3007 void Objecter::_session_command_op_remove(OSDSession *from, CommandOp *op)
3008 {
3009   assert(from == op->session);
3010   // from->lock is locked
3011
3012   if (from->is_homeless()) {
3013     num_homeless_ops--;
3014   }
3015
3016   from->command_ops.erase(op->tid);
3017   put_session(from);
3018   op->session = NULL;
3019
3020   ldout(cct, 15) << __func__ << " " << from->osd << " " << op->tid << dendl;
3021 }
3022
3023 void Objecter::_session_command_op_assign(OSDSession *to, CommandOp *op)
3024 {
3025   // to->lock is locked
3026   assert(op->session == NULL);
3027   assert(op->tid);
3028
3029   if (to->is_homeless()) {
3030     num_homeless_ops++;
3031   }
3032
3033   get_session(to);
3034   op->session = to;
3035   to->command_ops[op->tid] = op;
3036
3037   ldout(cct, 15) << __func__ << " " << to->osd << " " << op->tid << dendl;
3038 }
3039
3040 int Objecter::_recalc_linger_op_target(LingerOp *linger_op,
3041                                        shunique_lock& sul)
3042 {
3043   // rwlock is locked unique
3044
3045   int r = _calc_target(&linger_op->target, nullptr, true);
3046   if (r == RECALC_OP_TARGET_NEED_RESEND) {
3047     ldout(cct, 10) << "recalc_linger_op_target tid " << linger_op->linger_id
3048                    << " pgid " << linger_op->target.pgid
3049                    << " acting " << linger_op->target.acting << dendl;
3050
3051     OSDSession *s = NULL;
3052     r = _get_session(linger_op->target.osd, &s, sul);
3053     assert(r == 0);
3054
3055     if (linger_op->session != s) {
3056       // NB locking two sessions (s and linger_op->session) at the
3057       // same time here is only safe because we are the only one that
3058       // takes two, and we are holding rwlock for write.  Disable
3059       // lockdep because it doesn't know that.
3060       OSDSession::unique_lock sl(s->lock);
3061       _session_linger_op_remove(linger_op->session, linger_op);
3062       _session_linger_op_assign(s, linger_op);
3063     }
3064
3065     put_session(s);
3066     return RECALC_OP_TARGET_NEED_RESEND;
3067   }
3068   return r;
3069 }
3070
3071 void Objecter::_cancel_linger_op(Op *op)
3072 {
3073   ldout(cct, 15) << "cancel_op " << op->tid << dendl;
3074
3075   assert(!op->should_resend);
3076   if (op->onfinish) {
3077     delete op->onfinish;
3078     num_in_flight--;
3079   }
3080
3081   _finish_op(op, 0);
3082 }
3083
3084 void Objecter::_finish_op(Op *op, int r)
3085 {
3086   ldout(cct, 15) << "finish_op " << op->tid << dendl;
3087
3088   // op->session->lock is locked unique or op->session is null
3089
3090   if (!op->ctx_budgeted && op->budgeted)
3091     put_op_budget(op);
3092
3093   if (op->ontimeout && r != -ETIMEDOUT)
3094     timer.cancel_event(op->ontimeout);
3095
3096   if (op->session) {
3097     _session_op_remove(op->session, op);
3098   }
3099
3100   logger->dec(l_osdc_op_active);
3101
3102   assert(check_latest_map_ops.find(op->tid) == check_latest_map_ops.end());
3103
3104   inflight_ops--;
3105
3106   op->put();
3107 }
3108
3109 void Objecter::finish_op(OSDSession *session, ceph_tid_t tid)
3110 {
3111   ldout(cct, 15) << "finish_op " << tid << dendl;
3112   shared_lock rl(rwlock);
3113
3114   OSDSession::unique_lock wl(session->lock);
3115
3116   map<ceph_tid_t, Op *>::iterator iter = session->ops.find(tid);
3117   if (iter == session->ops.end())
3118     return;
3119
3120   Op *op = iter->second;
3121
3122   _finish_op(op, 0);
3123 }
3124
3125 MOSDOp *Objecter::_prepare_osd_op(Op *op)
3126 {
3127   // rwlock is locked
3128
3129   int flags = op->target.flags;
3130   flags |= CEPH_OSD_FLAG_KNOWN_REDIR;
3131
3132   // Nothing checks this any longer, but needed for compatibility with
3133   // pre-luminous osds
3134   flags |= CEPH_OSD_FLAG_ONDISK;
3135
3136   if (!honor_osdmap_full)
3137     flags |= CEPH_OSD_FLAG_FULL_FORCE;
3138
3139   op->target.paused = false;
3140   op->stamp = ceph::mono_clock::now();
3141
3142   hobject_t hobj = op->target.get_hobj();
3143   MOSDOp *m = new MOSDOp(client_inc, op->tid,
3144                          hobj, op->target.actual_pgid,
3145                          osdmap->get_epoch(),
3146                          flags, op->features);
3147
3148   m->set_snapid(op->snapid);
3149   m->set_snap_seq(op->snapc.seq);
3150   m->set_snaps(op->snapc.snaps);
3151
3152   m->ops = op->ops;
3153   m->set_mtime(op->mtime);
3154   m->set_retry_attempt(op->attempts++);
3155
3156   if (!op->trace.valid() && cct->_conf->osdc_blkin_trace_all) {
3157     op->trace.init("op", &trace_endpoint);
3158   }
3159
3160   if (op->priority)
3161     m->set_priority(op->priority);
3162   else
3163     m->set_priority(cct->_conf->osd_client_op_priority);
3164
3165   if (op->reqid != osd_reqid_t()) {
3166     m->set_reqid(op->reqid);
3167   }
3168
3169   logger->inc(l_osdc_op_send);
3170   logger->inc(l_osdc_op_send_bytes, m->get_data().length());
3171
3172   return m;
3173 }
3174
3175 void Objecter::_send_op(Op *op, MOSDOp *m)
3176 {
3177   // rwlock is locked
3178   // op->session->lock is locked
3179
3180   // backoff?
3181   hobject_t hoid = op->target.get_hobj();
3182   auto p = op->session->backoffs.find(op->target.actual_pgid);
3183   if (p != op->session->backoffs.end()) {
3184     auto q = p->second.lower_bound(hoid);
3185     if (q != p->second.begin()) {
3186       --q;
3187       if (hoid >= q->second.end) {
3188         ++q;
3189       }
3190     }
3191     if (q != p->second.end()) {
3192       ldout(cct, 20) << __func__ << " ? " << q->first << " [" << q->second.begin
3193                      << "," << q->second.end << ")" << dendl;
3194       int r = cmp(hoid, q->second.begin);
3195       if (r == 0 || (r > 0 && hoid < q->second.end)) {
3196         ldout(cct, 10) << __func__ << " backoff " << op->target.actual_pgid
3197                        << " id " << q->second.id << " on " << hoid
3198                        << ", queuing " << op << " tid " << op->tid << dendl;
3199         return;
3200       }
3201     }
3202   }
3203
3204   if (!m) {
3205     assert(op->tid > 0);
3206     m = _prepare_osd_op(op);
3207   }
3208
3209   if (op->target.actual_pgid != m->get_spg()) {
3210     ldout(cct, 10) << __func__ << " " << op->tid << " pgid change from "
3211                    << m->get_spg() << " to " << op->target.actual_pgid
3212                    << ", updating and reencoding" << dendl;
3213     m->set_spg(op->target.actual_pgid);
3214     m->clear_payload();  // reencode
3215   }
3216
3217   ldout(cct, 15) << "_send_op " << op->tid << " to "
3218                  << op->target.actual_pgid << " on osd." << op->session->osd
3219                  << dendl;
3220
3221   ConnectionRef con = op->session->con;
3222   assert(con);
3223
3224   // preallocated rx buffer?
3225   if (op->con) {
3226     ldout(cct, 20) << " revoking rx buffer for " << op->tid << " on "
3227                    << op->con << dendl;
3228     op->con->revoke_rx_buffer(op->tid);
3229   }
3230   if (op->outbl &&
3231       op->ontimeout == 0 &&  // only post rx_buffer if no timeout; see #9582
3232       op->outbl->length()) {
3233     ldout(cct, 20) << " posting rx buffer for " << op->tid << " on " << con
3234                    << dendl;
3235     op->con = con;
3236     op->con->post_rx_buffer(op->tid, *op->outbl);
3237   }
3238
3239   op->incarnation = op->session->incarnation;
3240
3241   m->set_tid(op->tid);
3242
3243   if (op->trace.valid()) {
3244     m->trace.init("op msg", nullptr, &op->trace);
3245   }
3246   op->session->con->send_message(m);
3247 }
3248
3249 int Objecter::calc_op_budget(Op *op)
3250 {
3251   int op_budget = 0;
3252   for (vector<OSDOp>::iterator i = op->ops.begin();
3253        i != op->ops.end();
3254        ++i) {
3255     if (i->op.op & CEPH_OSD_OP_MODE_WR) {
3256       op_budget += i->indata.length();
3257     } else if (ceph_osd_op_mode_read(i->op.op)) {
3258       if (ceph_osd_op_type_data(i->op.op)) {
3259         if ((int64_t)i->op.extent.length > 0)
3260           op_budget += (int64_t)i->op.extent.length;
3261       } else if (ceph_osd_op_type_attr(i->op.op)) {
3262         op_budget += i->op.xattr.name_len + i->op.xattr.value_len;
3263       }
3264     }
3265   }
3266   return op_budget;
3267 }
3268
3269 void Objecter::_throttle_op(Op *op,
3270                             shunique_lock& sul,
3271                             int op_budget)
3272 {
3273   assert(sul && sul.mutex() == &rwlock);
3274   bool locked_for_write = sul.owns_lock();
3275
3276   if (!op_budget)
3277     op_budget = calc_op_budget(op);
3278   if (!op_throttle_bytes.get_or_fail(op_budget)) { //couldn't take right now
3279     sul.unlock();
3280     op_throttle_bytes.get(op_budget);
3281     if (locked_for_write)
3282       sul.lock();
3283     else
3284       sul.lock_shared();
3285   }
3286   if (!op_throttle_ops.get_or_fail(1)) { //couldn't take right now
3287     sul.unlock();
3288     op_throttle_ops.get(1);
3289     if (locked_for_write)
3290       sul.lock();
3291     else
3292       sul.lock_shared();
3293   }
3294 }
3295
3296 void Objecter::unregister_op(Op *op)
3297 {
3298   OSDSession::unique_lock sl(op->session->lock);
3299   op->session->ops.erase(op->tid);
3300   sl.unlock();
3301   put_session(op->session);
3302   op->session = NULL;
3303
3304   inflight_ops--;
3305 }
3306
3307 /* This function DOES put the passed message before returning */
3308 void Objecter::handle_osd_op_reply(MOSDOpReply *m)
3309 {
3310   ldout(cct, 10) << "in handle_osd_op_reply" << dendl;
3311
3312   // get pio
3313   ceph_tid_t tid = m->get_tid();
3314
3315   shunique_lock sul(rwlock, ceph::acquire_shared);
3316   if (!initialized) {
3317     m->put();
3318     return;
3319   }
3320
3321   ConnectionRef con = m->get_connection();
3322   OSDSession *s = static_cast<OSDSession*>(con->get_priv());
3323   if (!s || s->con != con) {
3324     ldout(cct, 7) << __func__ << " no session on con " << con << dendl;
3325     if (s) {
3326       s->put();
3327     }
3328     m->put();
3329     return;
3330   }
3331
3332   OSDSession::unique_lock sl(s->lock);
3333
3334   map<ceph_tid_t, Op *>::iterator iter = s->ops.find(tid);
3335   if (iter == s->ops.end()) {
3336     ldout(cct, 7) << "handle_osd_op_reply " << tid
3337                   << (m->is_ondisk() ? " ondisk" : (m->is_onnvram() ?
3338                                                     " onnvram" : " ack"))
3339                   << " ... stray" << dendl;
3340     sl.unlock();
3341     put_session(s);
3342     m->put();
3343     return;
3344   }
3345
3346   ldout(cct, 7) << "handle_osd_op_reply " << tid
3347                 << (m->is_ondisk() ? " ondisk" :
3348                     (m->is_onnvram() ? " onnvram" : " ack"))
3349                 << " uv " << m->get_user_version()
3350                 << " in " << m->get_pg()
3351                 << " attempt " << m->get_retry_attempt()
3352                 << dendl;
3353   Op *op = iter->second;
3354   op->trace.event("osd op reply");
3355
3356   if (retry_writes_after_first_reply && op->attempts == 1 &&
3357       (op->target.flags & CEPH_OSD_FLAG_WRITE)) {
3358     ldout(cct, 7) << "retrying write after first reply: " << tid << dendl;
3359     if (op->onfinish) {
3360       num_in_flight--;
3361     }
3362     _session_op_remove(s, op);
3363     sl.unlock();
3364     put_session(s);
3365
3366     _op_submit(op, sul, NULL);
3367     m->put();
3368     return;
3369   }
3370
3371   if (m->get_retry_attempt() >= 0) {
3372     if (m->get_retry_attempt() != (op->attempts - 1)) {
3373       ldout(cct, 7) << " ignoring reply from attempt "
3374                     << m->get_retry_attempt()
3375                     << " from " << m->get_source_inst()
3376                     << "; last attempt " << (op->attempts - 1) << " sent to "
3377                     << op->session->con->get_peer_addr() << dendl;
3378       m->put();
3379       sl.unlock();
3380       put_session(s);
3381       return;
3382     }
3383   } else {
3384     // we don't know the request attempt because the server is old, so
3385     // just accept this one.  we may do ACK callbacks we shouldn't
3386     // have, but that is better than doing callbacks out of order.
3387   }
3388
3389   Context *onfinish = 0;
3390
3391   int rc = m->get_result();
3392
3393   if (m->is_redirect_reply()) {
3394     ldout(cct, 5) << " got redirect reply; redirecting" << dendl;
3395     if (op->onfinish)
3396       num_in_flight--;
3397     _session_op_remove(s, op);
3398     sl.unlock();
3399     put_session(s);
3400
3401     // FIXME: two redirects could race and reorder
3402
3403     op->tid = 0;
3404     m->get_redirect().combine_with_locator(op->target.target_oloc,
3405                                            op->target.target_oid.name);
3406     op->target.flags |= CEPH_OSD_FLAG_REDIRECTED;
3407     _op_submit(op, sul, NULL);
3408     m->put();
3409     return;
3410   }
3411
3412   if (rc == -EAGAIN) {
3413     ldout(cct, 7) << " got -EAGAIN, resubmitting" << dendl;
3414
3415     // new tid
3416     s->ops.erase(op->tid);
3417     op->tid = ++last_tid;
3418
3419     _send_op(op);
3420     sl.unlock();
3421     put_session(s);
3422     m->put();
3423     return;
3424   }
3425
3426   sul.unlock();
3427
3428   if (op->objver)
3429     *op->objver = m->get_user_version();
3430   if (op->reply_epoch)
3431     *op->reply_epoch = m->get_map_epoch();
3432   if (op->data_offset)
3433     *op->data_offset = m->get_header().data_off;
3434
3435   // got data?
3436   if (op->outbl) {
3437     if (op->con)
3438       op->con->revoke_rx_buffer(op->tid);
3439     m->claim_data(*op->outbl);
3440     op->outbl = 0;
3441   }
3442
3443   // per-op result demuxing
3444   vector<OSDOp> out_ops;
3445   m->claim_ops(out_ops);
3446
3447   if (out_ops.size() != op->ops.size())
3448     ldout(cct, 0) << "WARNING: tid " << op->tid << " reply ops " << out_ops
3449                   << " != request ops " << op->ops
3450                   << " from " << m->get_source_inst() << dendl;
3451
3452   vector<bufferlist*>::iterator pb = op->out_bl.begin();
3453   vector<int*>::iterator pr = op->out_rval.begin();
3454   vector<Context*>::iterator ph = op->out_handler.begin();
3455   assert(op->out_bl.size() == op->out_rval.size());
3456   assert(op->out_bl.size() == op->out_handler.size());
3457   vector<OSDOp>::iterator p = out_ops.begin();
3458   for (unsigned i = 0;
3459        p != out_ops.end() && pb != op->out_bl.end();
3460        ++i, ++p, ++pb, ++pr, ++ph) {
3461     ldout(cct, 10) << " op " << i << " rval " << p->rval
3462                    << " len " << p->outdata.length() << dendl;
3463     if (*pb)
3464       **pb = p->outdata;
3465     // set rval before running handlers so that handlers
3466     // can change it if e.g. decoding fails
3467     if (*pr)
3468       **pr = ceph_to_hostos_errno(p->rval);
3469     if (*ph) {
3470       ldout(cct, 10) << " op " << i << " handler " << *ph << dendl;
3471       (*ph)->complete(ceph_to_hostos_errno(p->rval));
3472       *ph = NULL;
3473     }
3474   }
3475
3476   // NOTE: we assume that since we only request ONDISK ever we will
3477   // only ever get back one (type of) ack ever.
3478
3479   if (op->onfinish) {
3480     num_in_flight--;
3481     onfinish = op->onfinish;
3482     op->onfinish = NULL;
3483   }
3484   logger->inc(l_osdc_op_reply);
3485
3486   /* get it before we call _finish_op() */
3487   auto completion_lock = s->get_lock(op->target.base_oid);
3488
3489   ldout(cct, 15) << "handle_osd_op_reply completed tid " << tid << dendl;
3490   _finish_op(op, 0);
3491
3492   ldout(cct, 5) << num_in_flight << " in flight" << dendl;
3493
3494   // serialize completions
3495   if (completion_lock.mutex()) {
3496     completion_lock.lock();
3497   }
3498   sl.unlock();
3499
3500   // do callbacks
3501   if (onfinish) {
3502     onfinish->complete(rc);
3503   }
3504   if (completion_lock.mutex()) {
3505     completion_lock.unlock();
3506   }
3507
3508   m->put();
3509   put_session(s);
3510 }
3511
3512 void Objecter::handle_osd_backoff(MOSDBackoff *m)
3513 {
3514   ldout(cct, 10) << __func__ << " " << *m << dendl;
3515   shunique_lock sul(rwlock, ceph::acquire_shared);
3516   if (!initialized) {
3517     m->put();
3518     return;
3519   }
3520
3521   ConnectionRef con = m->get_connection();
3522   OSDSession *s = static_cast<OSDSession*>(con->get_priv());
3523   if (!s || s->con != con) {
3524     ldout(cct, 7) << __func__ << " no session on con " << con << dendl;
3525     if (s)
3526       s->put();
3527     m->put();
3528     return;
3529   }
3530
3531   get_session(s);
3532   s->put();  // from get_priv() above
3533
3534   OSDSession::unique_lock sl(s->lock);
3535
3536   switch (m->op) {
3537   case CEPH_OSD_BACKOFF_OP_BLOCK:
3538     {
3539       // register
3540       OSDBackoff& b = s->backoffs[m->pgid][m->begin];
3541       s->backoffs_by_id.insert(make_pair(m->id, &b));
3542       b.pgid = m->pgid;
3543       b.id = m->id;
3544       b.begin = m->begin;
3545       b.end = m->end;
3546
3547       // ack with original backoff's epoch so that the osd can discard this if
3548       // there was a pg split.
3549       Message *r = new MOSDBackoff(m->pgid,
3550                                    m->map_epoch,
3551                                    CEPH_OSD_BACKOFF_OP_ACK_BLOCK,
3552                                    m->id, m->begin, m->end);
3553       // this priority must match the MOSDOps from _prepare_osd_op
3554       r->set_priority(cct->_conf->osd_client_op_priority);
3555       con->send_message(r);
3556     }
3557     break;
3558
3559   case CEPH_OSD_BACKOFF_OP_UNBLOCK:
3560     {
3561       auto p = s->backoffs_by_id.find(m->id);
3562       if (p != s->backoffs_by_id.end()) {
3563         OSDBackoff *b = p->second;
3564         if (b->begin != m->begin &&
3565             b->end != m->end) {
3566           lderr(cct) << __func__ << " got " << m->pgid << " id " << m->id
3567                      << " unblock on ["
3568                      << m->begin << "," << m->end << ") but backoff is ["
3569                      << b->begin << "," << b->end << ")" << dendl;
3570           // hrmpf, unblock it anyway.
3571         }
3572         ldout(cct, 10) << __func__ << " unblock backoff " << b->pgid
3573                        << " id " << b->id
3574                        << " [" << b->begin << "," << b->end
3575                        << ")" << dendl;
3576         auto spgp = s->backoffs.find(b->pgid);
3577         assert(spgp != s->backoffs.end());
3578         spgp->second.erase(b->begin);
3579         if (spgp->second.empty()) {
3580           s->backoffs.erase(spgp);
3581         }
3582         s->backoffs_by_id.erase(p);
3583
3584         // check for any ops to resend
3585         for (auto& q : s->ops) {
3586           if (q.second->target.actual_pgid == m->pgid) {
3587             int r = q.second->target.contained_by(m->begin, m->end);
3588             ldout(cct, 20) << __func__ <<  " contained_by " << r << " on "
3589                            << q.second->target.get_hobj() << dendl;
3590             if (r) {
3591               _send_op(q.second);
3592             }
3593           }
3594         }
3595       } else {
3596         lderr(cct) << __func__ << " " << m->pgid << " id " << m->id
3597                    << " unblock on ["
3598                    << m->begin << "," << m->end << ") but backoff dne" << dendl;
3599       }
3600     }
3601     break;
3602
3603   default:
3604     ldout(cct, 10) << __func__ << " unrecognized op " << (int)m->op << dendl;
3605   }
3606
3607   sul.unlock();
3608   sl.unlock();
3609
3610   m->put();
3611   put_session(s);
3612 }
3613
3614 uint32_t Objecter::list_nobjects_seek(NListContext *list_context,
3615                                       uint32_t pos)
3616 {
3617   shared_lock rl(rwlock);
3618   list_context->pos = hobject_t(object_t(), string(), CEPH_NOSNAP,
3619                                 pos, list_context->pool_id, string());
3620   ldout(cct, 10) << __func__ << list_context
3621                  << " pos " << pos << " -> " << list_context->pos << dendl;
3622   pg_t actual = osdmap->raw_pg_to_pg(pg_t(pos, list_context->pool_id));
3623   list_context->current_pg = actual.ps();
3624   list_context->at_end_of_pool = false;
3625   return pos;
3626 }
3627
3628 uint32_t Objecter::list_nobjects_seek(NListContext *list_context,
3629                                       const hobject_t& cursor)
3630 {
3631   shared_lock rl(rwlock);
3632   ldout(cct, 10) << "list_nobjects_seek " << list_context << dendl;
3633   list_context->pos = cursor;
3634   list_context->at_end_of_pool = false;
3635   pg_t actual = osdmap->raw_pg_to_pg(pg_t(cursor.get_hash(), list_context->pool_id));
3636   list_context->current_pg = actual.ps();
3637   list_context->sort_bitwise = true;
3638   return list_context->current_pg;
3639 }
3640
3641 void Objecter::list_nobjects_get_cursor(NListContext *list_context,
3642                                         hobject_t *cursor)
3643 {
3644   shared_lock rl(rwlock);
3645   if (list_context->list.empty()) {
3646     *cursor = list_context->pos;
3647   } else {
3648     const librados::ListObjectImpl& entry = list_context->list.front();
3649     const string *key = (entry.locator.empty() ? &entry.oid : &entry.locator);
3650     uint32_t h = osdmap->get_pg_pool(list_context->pool_id)->hash_key(*key, entry.nspace);
3651     *cursor = hobject_t(entry.oid, entry.locator, list_context->pool_snap_seq, h, list_context->pool_id, entry.nspace);
3652   }
3653 }
3654
3655 void Objecter::list_nobjects(NListContext *list_context, Context *onfinish)
3656 {
3657   ldout(cct, 10) << __func__ << " pool_id " << list_context->pool_id
3658                  << " pool_snap_seq " << list_context->pool_snap_seq
3659                  << " max_entries " << list_context->max_entries
3660                  << " list_context " << list_context
3661                  << " onfinish " << onfinish
3662                  << " current_pg " << list_context->current_pg
3663                  << " pos " << list_context->pos << dendl;
3664
3665   shared_lock rl(rwlock);
3666   const pg_pool_t *pool = osdmap->get_pg_pool(list_context->pool_id);
3667   if (!pool) { // pool is gone
3668     rl.unlock();
3669     put_nlist_context_budget(list_context);
3670     onfinish->complete(-ENOENT);
3671     return;
3672   }
3673   int pg_num = pool->get_pg_num();
3674   bool sort_bitwise = osdmap->test_flag(CEPH_OSDMAP_SORTBITWISE);
3675
3676   if (list_context->pos.is_min()) {
3677     list_context->starting_pg_num = 0;
3678     list_context->sort_bitwise = sort_bitwise;
3679     list_context->starting_pg_num = pg_num;
3680   }
3681   if (list_context->sort_bitwise != sort_bitwise) {
3682     list_context->pos = hobject_t(
3683       object_t(), string(), CEPH_NOSNAP,
3684       list_context->current_pg, list_context->pool_id, string());
3685     list_context->sort_bitwise = sort_bitwise;
3686     ldout(cct, 10) << " hobject sort order changed, restarting this pg at "
3687                    << list_context->pos << dendl;
3688   }
3689   if (list_context->starting_pg_num != pg_num) {
3690     if (!sort_bitwise) {
3691       // start reading from the beginning; the pgs have changed
3692       ldout(cct, 10) << " pg_num changed; restarting with " << pg_num << dendl;
3693       list_context->pos = collection_list_handle_t();
3694     }
3695     list_context->starting_pg_num = pg_num;
3696   }
3697
3698   if (list_context->pos.is_max()) {
3699     ldout(cct, 20) << __func__ << " end of pool, list "
3700                    << list_context->list << dendl;
3701     if (list_context->list.empty()) {
3702       list_context->at_end_of_pool = true;
3703     }
3704     // release the listing context's budget once all
3705     // OPs (in the session) are finished
3706     put_nlist_context_budget(list_context);
3707     onfinish->complete(0);
3708     return;
3709   }
3710
3711   ObjectOperation op;
3712   op.pg_nls(list_context->max_entries, list_context->filter,
3713             list_context->pos, osdmap->get_epoch());
3714   list_context->bl.clear();
3715   C_NList *onack = new C_NList(list_context, onfinish, this);
3716   object_locator_t oloc(list_context->pool_id, list_context->nspace);
3717
3718   // note current_pg in case we don't have (or lose) SORTBITWISE
3719   list_context->current_pg = pool->raw_hash_to_pg(list_context->pos.get_hash());
3720   rl.unlock();
3721
3722   pg_read(list_context->current_pg, oloc, op,
3723           &list_context->bl, 0, onack, &onack->epoch,
3724           &list_context->ctx_budget);
3725 }
3726
3727 void Objecter::_nlist_reply(NListContext *list_context, int r,
3728                             Context *final_finish, epoch_t reply_epoch)
3729 {
3730   ldout(cct, 10) << __func__ << " " << list_context << dendl;
3731
3732   bufferlist::iterator iter = list_context->bl.begin();
3733   pg_nls_response_t response;
3734   bufferlist extra_info;
3735   ::decode(response, iter);
3736   if (!iter.end()) {
3737     ::decode(extra_info, iter);
3738   }
3739
3740   // if the osd returns 1 (newer code), or handle MAX, it means we
3741   // hit the end of the pg.
3742   if ((response.handle.is_max() || r == 1) &&
3743       !list_context->sort_bitwise) {
3744     // legacy OSD and !sortbitwise, figure out the next PG on our own
3745     ++list_context->current_pg;
3746     if (list_context->current_pg == list_context->starting_pg_num) {
3747       // end of pool
3748       list_context->pos = hobject_t::get_max();
3749     } else {
3750       // next pg
3751       list_context->pos = hobject_t(object_t(), string(), CEPH_NOSNAP,
3752                                     list_context->current_pg,
3753                                     list_context->pool_id, string());
3754     }
3755   } else {
3756     list_context->pos = response.handle;
3757   }
3758
3759   int response_size = response.entries.size();
3760   ldout(cct, 20) << " response.entries.size " << response_size
3761                  << ", response.entries " << response.entries
3762                  << ", handle " << response.handle
3763                  << ", tentative new pos " << list_context->pos << dendl;
3764   list_context->extra_info.append(extra_info);
3765   if (response_size) {
3766     list_context->list.splice(list_context->list.end(), response.entries);
3767   }
3768
3769   if (list_context->list.size() >= list_context->max_entries) {
3770     ldout(cct, 20) << " hit max, returning results so far, "
3771                    << list_context->list << dendl;
3772     // release the listing context's budget once all
3773     // OPs (in the session) are finished
3774     put_nlist_context_budget(list_context);
3775     final_finish->complete(0);
3776     return;
3777   }
3778
3779   // continue!
3780   list_nobjects(list_context, final_finish);
3781 }
3782
3783 void Objecter::put_nlist_context_budget(NListContext *list_context)
3784 {
3785   if (list_context->ctx_budget >= 0) {
3786     ldout(cct, 10) << " release listing context's budget " <<
3787       list_context->ctx_budget << dendl;
3788     put_op_budget_bytes(list_context->ctx_budget);
3789     list_context->ctx_budget = -1;
3790   }
3791 }
3792
3793 // snapshots
3794
3795 int Objecter::create_pool_snap(int64_t pool, string& snap_name,
3796                                Context *onfinish)
3797 {
3798   unique_lock wl(rwlock);
3799   ldout(cct, 10) << "create_pool_snap; pool: " << pool << "; snap: "
3800                  << snap_name << dendl;
3801
3802   const pg_pool_t *p = osdmap->get_pg_pool(pool);
3803   if (!p)
3804     return -EINVAL;
3805   if (p->snap_exists(snap_name.c_str()))
3806     return -EEXIST;
3807
3808   PoolOp *op = new PoolOp;
3809   if (!op)
3810     return -ENOMEM;
3811   op->tid = ++last_tid;
3812   op->pool = pool;
3813   op->name = snap_name;
3814   op->onfinish = onfinish;
3815   op->pool_op = POOL_OP_CREATE_SNAP;
3816   pool_ops[op->tid] = op;
3817
3818   pool_op_submit(op);
3819
3820   return 0;
3821 }
3822
3823 struct C_SelfmanagedSnap : public Context {
3824   bufferlist bl;
3825   snapid_t *psnapid;
3826   Context *fin;
3827   C_SelfmanagedSnap(snapid_t *ps, Context *f) : psnapid(ps), fin(f) {}
3828   void finish(int r) override {
3829     if (r == 0) {
3830       bufferlist::iterator p = bl.begin();
3831       ::decode(*psnapid, p);
3832     }
3833     fin->complete(r);
3834   }
3835 };
3836
3837 int Objecter::allocate_selfmanaged_snap(int64_t pool, snapid_t *psnapid,
3838                                         Context *onfinish)
3839 {
3840   unique_lock wl(rwlock);
3841   ldout(cct, 10) << "allocate_selfmanaged_snap; pool: " << pool << dendl;
3842   PoolOp *op = new PoolOp;
3843   if (!op) return -ENOMEM;
3844   op->tid = ++last_tid;
3845   op->pool = pool;
3846   C_SelfmanagedSnap *fin = new C_SelfmanagedSnap(psnapid, onfinish);
3847   op->onfinish = fin;
3848   op->blp = &fin->bl;
3849   op->pool_op = POOL_OP_CREATE_UNMANAGED_SNAP;
3850   pool_ops[op->tid] = op;
3851
3852   pool_op_submit(op);
3853   return 0;
3854 }
3855
3856 int Objecter::delete_pool_snap(int64_t pool, string& snap_name,
3857                                Context *onfinish)
3858 {
3859   unique_lock wl(rwlock);
3860   ldout(cct, 10) << "delete_pool_snap; pool: " << pool << "; snap: "
3861                  << snap_name << dendl;
3862
3863   const pg_pool_t *p = osdmap->get_pg_pool(pool);
3864   if (!p)
3865     return -EINVAL;
3866   if (!p->snap_exists(snap_name.c_str()))
3867     return -ENOENT;
3868
3869   PoolOp *op = new PoolOp;
3870   if (!op)
3871     return -ENOMEM;
3872   op->tid = ++last_tid;
3873   op->pool = pool;
3874   op->name = snap_name;
3875   op->onfinish = onfinish;
3876   op->pool_op = POOL_OP_DELETE_SNAP;
3877   pool_ops[op->tid] = op;
3878
3879   pool_op_submit(op);
3880
3881   return 0;
3882 }
3883
3884 int Objecter::delete_selfmanaged_snap(int64_t pool, snapid_t snap,
3885                                       Context *onfinish)
3886 {
3887   unique_lock wl(rwlock);
3888   ldout(cct, 10) << "delete_selfmanaged_snap; pool: " << pool << "; snap: "
3889                  << snap << dendl;
3890   PoolOp *op = new PoolOp;
3891   if (!op) return -ENOMEM;
3892   op->tid = ++last_tid;
3893   op->pool = pool;
3894   op->onfinish = onfinish;
3895   op->pool_op = POOL_OP_DELETE_UNMANAGED_SNAP;
3896   op->snapid = snap;
3897   pool_ops[op->tid] = op;
3898
3899   pool_op_submit(op);
3900
3901   return 0;
3902 }
3903
3904 int Objecter::create_pool(string& name, Context *onfinish, uint64_t auid,
3905                           int crush_rule)
3906 {
3907   unique_lock wl(rwlock);
3908   ldout(cct, 10) << "create_pool name=" << name << dendl;
3909
3910   if (osdmap->lookup_pg_pool_name(name) >= 0)
3911     return -EEXIST;
3912
3913   PoolOp *op = new PoolOp;
3914   if (!op)
3915     return -ENOMEM;
3916   op->tid = ++last_tid;
3917   op->pool = 0;
3918   op->name = name;
3919   op->onfinish = onfinish;
3920   op->pool_op = POOL_OP_CREATE;
3921   pool_ops[op->tid] = op;
3922   op->auid = auid;
3923   op->crush_rule = crush_rule;
3924
3925   pool_op_submit(op);
3926
3927   return 0;
3928 }
3929
3930 int Objecter::delete_pool(int64_t pool, Context *onfinish)
3931 {
3932   unique_lock wl(rwlock);
3933   ldout(cct, 10) << "delete_pool " << pool << dendl;
3934
3935   if (!osdmap->have_pg_pool(pool))
3936     return -ENOENT;
3937
3938   _do_delete_pool(pool, onfinish);
3939   return 0;
3940 }
3941
3942 int Objecter::delete_pool(const string &pool_name, Context *onfinish)
3943 {
3944   unique_lock wl(rwlock);
3945   ldout(cct, 10) << "delete_pool " << pool_name << dendl;
3946
3947   int64_t pool = osdmap->lookup_pg_pool_name(pool_name);
3948   if (pool < 0)
3949     return pool;
3950
3951   _do_delete_pool(pool, onfinish);
3952   return 0;
3953 }
3954
3955 void Objecter::_do_delete_pool(int64_t pool, Context *onfinish)
3956 {
3957   PoolOp *op = new PoolOp;
3958   op->tid = ++last_tid;
3959   op->pool = pool;
3960   op->name = "delete";
3961   op->onfinish = onfinish;
3962   op->pool_op = POOL_OP_DELETE;
3963   pool_ops[op->tid] = op;
3964   pool_op_submit(op);
3965 }
3966
3967 /**
3968  * change the auid owner of a pool by contacting the monitor.
3969  * This requires the current connection to have write permissions
3970  * on both the pool's current auid and the new (parameter) auid.
3971  * Uses the standard Context callback when done.
3972  */
3973 int Objecter::change_pool_auid(int64_t pool, Context *onfinish, uint64_t auid)
3974 {
3975   unique_lock wl(rwlock);
3976   ldout(cct, 10) << "change_pool_auid " << pool << " to " << auid << dendl;
3977   PoolOp *op = new PoolOp;
3978   if (!op) return -ENOMEM;
3979   op->tid = ++last_tid;
3980   op->pool = pool;
3981   op->name = "change_pool_auid";
3982   op->onfinish = onfinish;
3983   op->pool_op = POOL_OP_AUID_CHANGE;
3984   op->auid = auid;
3985   pool_ops[op->tid] = op;
3986
3987   logger->set(l_osdc_poolop_active, pool_ops.size());
3988
3989   pool_op_submit(op);
3990   return 0;
3991 }
3992
3993 void Objecter::pool_op_submit(PoolOp *op)
3994 {
3995   // rwlock is locked
3996   if (mon_timeout > timespan(0)) {
3997     op->ontimeout = timer.add_event(mon_timeout,
3998                                     [this, op]() {
3999                                       pool_op_cancel(op->tid, -ETIMEDOUT); });
4000   }
4001   _pool_op_submit(op);
4002 }
4003
4004 void Objecter::_pool_op_submit(PoolOp *op)
4005 {
4006   // rwlock is locked unique
4007
4008   ldout(cct, 10) << "pool_op_submit " << op->tid << dendl;
4009   MPoolOp *m = new MPoolOp(monc->get_fsid(), op->tid, op->pool,
4010                            op->name, op->pool_op,
4011                            op->auid, last_seen_osdmap_version);
4012   if (op->snapid) m->snapid = op->snapid;
4013   if (op->crush_rule) m->crush_rule = op->crush_rule;
4014   monc->send_mon_message(m);
4015   op->last_submit = ceph::mono_clock::now();
4016
4017   logger->inc(l_osdc_poolop_send);
4018 }
4019
4020 /**
4021  * Handle a reply to a PoolOp message. Check that we sent the message
4022  * and give the caller responsibility for the returned bufferlist.
4023  * Then either call the finisher or stash the PoolOp, depending on if we
4024  * have a new enough map.
4025  * Lastly, clean up the message and PoolOp.
4026  */
4027 void Objecter::handle_pool_op_reply(MPoolOpReply *m)
4028 {
4029   FUNCTRACE();
4030   shunique_lock sul(rwlock, acquire_shared);
4031   if (!initialized) {
4032     sul.unlock();
4033     m->put();
4034     return;
4035   }
4036
4037   ldout(cct, 10) << "handle_pool_op_reply " << *m << dendl;
4038   ceph_tid_t tid = m->get_tid();
4039   map<ceph_tid_t, PoolOp *>::iterator iter = pool_ops.find(tid);
4040   if (iter != pool_ops.end()) {
4041     PoolOp *op = iter->second;
4042     ldout(cct, 10) << "have request " << tid << " at " << op << " Op: "
4043                    << ceph_pool_op_name(op->pool_op) << dendl;
4044     if (op->blp)
4045       op->blp->claim(m->response_data);
4046     if (m->version > last_seen_osdmap_version)
4047       last_seen_osdmap_version = m->version;
4048     if (osdmap->get_epoch() < m->epoch) {
4049       sul.unlock();
4050       sul.lock();
4051       // recheck op existence since we have let go of rwlock
4052       // (for promotion) above.
4053       iter = pool_ops.find(tid);
4054       if (iter == pool_ops.end())
4055         goto done; // op is gone.
4056       if (osdmap->get_epoch() < m->epoch) {
4057         ldout(cct, 20) << "waiting for client to reach epoch " << m->epoch
4058                        << " before calling back" << dendl;
4059         _wait_for_new_map(op->onfinish, m->epoch, m->replyCode);
4060       } else {
4061         // map epoch changed, probably because a MOSDMap message
4062         // sneaked in. Do caller-specified callback now or else
4063         // we lose it forever.
4064         assert(op->onfinish);
4065         op->onfinish->complete(m->replyCode);
4066       }
4067     } else {
4068       assert(op->onfinish);
4069       op->onfinish->complete(m->replyCode);
4070     }
4071     op->onfinish = NULL;
4072     if (!sul.owns_lock()) {
4073       sul.unlock();
4074       sul.lock();
4075     }
4076     iter = pool_ops.find(tid);
4077     if (iter != pool_ops.end()) {
4078       _finish_pool_op(op, 0);
4079     }
4080   } else {
4081     ldout(cct, 10) << "unknown request " << tid << dendl;
4082   }
4083
4084 done:
4085   // Not strictly necessary, since we'll release it on return.
4086   sul.unlock();
4087
4088   ldout(cct, 10) << "done" << dendl;
4089   m->put();
4090 }
4091
4092 int Objecter::pool_op_cancel(ceph_tid_t tid, int r)
4093 {
4094   assert(initialized);
4095
4096   unique_lock wl(rwlock);
4097
4098   map<ceph_tid_t, PoolOp*>::iterator it = pool_ops.find(tid);
4099   if (it == pool_ops.end()) {
4100     ldout(cct, 10) << __func__ << " tid " << tid << " dne" << dendl;
4101     return -ENOENT;
4102   }
4103
4104   ldout(cct, 10) << __func__ << " tid " << tid << dendl;
4105
4106   PoolOp *op = it->second;
4107   if (op->onfinish)
4108     op->onfinish->complete(r);
4109
4110   _finish_pool_op(op, r);
4111   return 0;
4112 }
4113
4114 void Objecter::_finish_pool_op(PoolOp *op, int r)
4115 {
4116   // rwlock is locked unique
4117   pool_ops.erase(op->tid);
4118   logger->set(l_osdc_poolop_active, pool_ops.size());
4119
4120   if (op->ontimeout && r != -ETIMEDOUT) {
4121     timer.cancel_event(op->ontimeout);
4122   }
4123
4124   delete op;
4125 }
4126
4127 // pool stats
4128
4129 void Objecter::get_pool_stats(list<string>& pools,
4130                               map<string,pool_stat_t> *result,
4131                               Context *onfinish)
4132 {
4133   ldout(cct, 10) << "get_pool_stats " << pools << dendl;
4134
4135   PoolStatOp *op = new PoolStatOp;
4136   op->tid = ++last_tid;
4137   op->pools = pools;
4138   op->pool_stats = result;
4139   op->onfinish = onfinish;
4140   if (mon_timeout > timespan(0)) {
4141     op->ontimeout = timer.add_event(mon_timeout,
4142                                     [this, op]() {
4143                                       pool_stat_op_cancel(op->tid,
4144                                                           -ETIMEDOUT); });
4145   } else {
4146     op->ontimeout = 0;
4147   }
4148
4149   unique_lock wl(rwlock);
4150
4151   poolstat_ops[op->tid] = op;
4152
4153   logger->set(l_osdc_poolstat_active, poolstat_ops.size());
4154
4155   _poolstat_submit(op);
4156 }
4157
4158 void Objecter::_poolstat_submit(PoolStatOp *op)
4159 {
4160   ldout(cct, 10) << "_poolstat_submit " << op->tid << dendl;
4161   monc->send_mon_message(new MGetPoolStats(monc->get_fsid(), op->tid,
4162                                            op->pools,
4163                                            last_seen_pgmap_version));
4164   op->last_submit = ceph::mono_clock::now();
4165
4166   logger->inc(l_osdc_poolstat_send);
4167 }
4168
4169 void Objecter::handle_get_pool_stats_reply(MGetPoolStatsReply *m)
4170 {
4171   ldout(cct, 10) << "handle_get_pool_stats_reply " << *m << dendl;
4172   ceph_tid_t tid = m->get_tid();
4173
4174   unique_lock wl(rwlock);
4175   if (!initialized) {
4176     m->put();
4177     return;
4178   }
4179
4180   map<ceph_tid_t, PoolStatOp *>::iterator iter = poolstat_ops.find(tid);
4181   if (iter != poolstat_ops.end()) {
4182     PoolStatOp *op = poolstat_ops[tid];
4183     ldout(cct, 10) << "have request " << tid << " at " << op << dendl;
4184     *op->pool_stats = m->pool_stats;
4185     if (m->version > last_seen_pgmap_version) {
4186       last_seen_pgmap_version = m->version;
4187     }
4188     op->onfinish->complete(0);
4189     _finish_pool_stat_op(op, 0);
4190   } else {
4191     ldout(cct, 10) << "unknown request " << tid << dendl;
4192   }
4193   ldout(cct, 10) << "done" << dendl;
4194   m->put();
4195 }
4196
4197 int Objecter::pool_stat_op_cancel(ceph_tid_t tid, int r)
4198 {
4199   assert(initialized);
4200
4201   unique_lock wl(rwlock);
4202
4203   map<ceph_tid_t, PoolStatOp*>::iterator it = poolstat_ops.find(tid);
4204   if (it == poolstat_ops.end()) {
4205     ldout(cct, 10) << __func__ << " tid " << tid << " dne" << dendl;
4206     return -ENOENT;
4207   }
4208
4209   ldout(cct, 10) << __func__ << " tid " << tid << dendl;
4210
4211   PoolStatOp *op = it->second;
4212   if (op->onfinish)
4213     op->onfinish->complete(r);
4214   _finish_pool_stat_op(op, r);
4215   return 0;
4216 }
4217
4218 void Objecter::_finish_pool_stat_op(PoolStatOp *op, int r)
4219 {
4220   // rwlock is locked unique
4221
4222   poolstat_ops.erase(op->tid);
4223   logger->set(l_osdc_poolstat_active, poolstat_ops.size());
4224
4225   if (op->ontimeout && r != -ETIMEDOUT)
4226     timer.cancel_event(op->ontimeout);
4227
4228   delete op;
4229 }
4230
4231 void Objecter::get_fs_stats(ceph_statfs& result, Context *onfinish)
4232 {
4233   ldout(cct, 10) << "get_fs_stats" << dendl;
4234   unique_lock l(rwlock);
4235
4236   StatfsOp *op = new StatfsOp;
4237   op->tid = ++last_tid;
4238   op->stats = &result;
4239   op->onfinish = onfinish;
4240   if (mon_timeout > timespan(0)) {
4241     op->ontimeout = timer.add_event(mon_timeout,
4242                                     [this, op]() {
4243                                       statfs_op_cancel(op->tid,
4244                                                        -ETIMEDOUT); });
4245   } else {
4246     op->ontimeout = 0;
4247   }
4248   statfs_ops[op->tid] = op;
4249
4250   logger->set(l_osdc_statfs_active, statfs_ops.size());
4251
4252   _fs_stats_submit(op);
4253 }
4254
4255 void Objecter::_fs_stats_submit(StatfsOp *op)
4256 {
4257   // rwlock is locked unique
4258
4259   ldout(cct, 10) << "fs_stats_submit" << op->tid << dendl;
4260   monc->send_mon_message(new MStatfs(monc->get_fsid(), op->tid,
4261                                      last_seen_pgmap_version));
4262   op->last_submit = ceph::mono_clock::now();
4263
4264   logger->inc(l_osdc_statfs_send);
4265 }
4266
4267 void Objecter::handle_fs_stats_reply(MStatfsReply *m)
4268 {
4269   unique_lock wl(rwlock);
4270   if (!initialized) {
4271     m->put();
4272     return;
4273   }
4274
4275   ldout(cct, 10) << "handle_fs_stats_reply " << *m << dendl;
4276   ceph_tid_t tid = m->get_tid();
4277
4278   if (statfs_ops.count(tid)) {
4279     StatfsOp *op = statfs_ops[tid];
4280     ldout(cct, 10) << "have request " << tid << " at " << op << dendl;
4281     *(op->stats) = m->h.st;
4282     if (m->h.version > last_seen_pgmap_version)
4283       last_seen_pgmap_version = m->h.version;
4284     op->onfinish->complete(0);
4285     _finish_statfs_op(op, 0);
4286   } else {
4287     ldout(cct, 10) << "unknown request " << tid << dendl;
4288   }
4289   m->put();
4290   ldout(cct, 10) << "done" << dendl;
4291 }
4292
4293 int Objecter::statfs_op_cancel(ceph_tid_t tid, int r)
4294 {
4295   assert(initialized);
4296
4297   unique_lock wl(rwlock);
4298
4299   map<ceph_tid_t, StatfsOp*>::iterator it = statfs_ops.find(tid);
4300   if (it == statfs_ops.end()) {
4301     ldout(cct, 10) << __func__ << " tid " << tid << " dne" << dendl;
4302     return -ENOENT;
4303   }
4304
4305   ldout(cct, 10) << __func__ << " tid " << tid << dendl;
4306
4307   StatfsOp *op = it->second;
4308   if (op->onfinish)
4309     op->onfinish->complete(r);
4310   _finish_statfs_op(op, r);
4311   return 0;
4312 }
4313
4314 void Objecter::_finish_statfs_op(StatfsOp *op, int r)
4315 {
4316   // rwlock is locked unique
4317
4318   statfs_ops.erase(op->tid);
4319   logger->set(l_osdc_statfs_active, statfs_ops.size());
4320
4321   if (op->ontimeout && r != -ETIMEDOUT)
4322     timer.cancel_event(op->ontimeout);
4323
4324   delete op;
4325 }
4326
4327 // scatter/gather
4328
4329 void Objecter::_sg_read_finish(vector<ObjectExtent>& extents,
4330                                vector<bufferlist>& resultbl,
4331                                bufferlist *bl, Context *onfinish)
4332 {
4333   // all done
4334   ldout(cct, 15) << "_sg_read_finish" << dendl;
4335
4336   if (extents.size() > 1) {
4337     Striper::StripedReadResult r;
4338     vector<bufferlist>::iterator bit = resultbl.begin();
4339     for (vector<ObjectExtent>::iterator eit = extents.begin();
4340          eit != extents.end();
4341          ++eit, ++bit) {
4342       r.add_partial_result(cct, *bit, eit->buffer_extents);
4343     }
4344     bl->clear();
4345     r.assemble_result(cct, *bl, false);
4346   } else {
4347     ldout(cct, 15) << "  only one frag" << dendl;
4348     bl->claim(resultbl[0]);
4349   }
4350
4351   // done
4352   uint64_t bytes_read = bl->length();
4353   ldout(cct, 7) << "_sg_read_finish " << bytes_read << " bytes" << dendl;
4354
4355   if (onfinish) {
4356     onfinish->complete(bytes_read);// > 0 ? bytes_read:m->get_result());
4357   }
4358 }
4359
4360
4361 void Objecter::ms_handle_connect(Connection *con)
4362 {
4363   ldout(cct, 10) << "ms_handle_connect " << con << dendl;
4364   if (!initialized)
4365     return;
4366
4367   if (con->get_peer_type() == CEPH_ENTITY_TYPE_MON)
4368     resend_mon_ops();
4369 }
4370
4371 bool Objecter::ms_handle_reset(Connection *con)
4372 {
4373   if (!initialized)
4374     return false;
4375   if (con->get_peer_type() == CEPH_ENTITY_TYPE_OSD) {
4376     OSDSession *session = static_cast<OSDSession*>(con->get_priv());
4377     if (session) {
4378       ldout(cct, 1) << "ms_handle_reset " << con << " session " << session
4379                     << " osd." << session->osd << dendl;
4380       unique_lock wl(rwlock);
4381       if (!initialized) {
4382         wl.unlock();
4383         return false;
4384       }
4385       map<uint64_t, LingerOp *> lresend;
4386       OSDSession::unique_lock sl(session->lock);
4387       _reopen_session(session);
4388       _kick_requests(session, lresend);
4389       sl.unlock();
4390       _linger_ops_resend(lresend, wl);
4391       wl.unlock();
4392       maybe_request_map();
4393       session->put();
4394     }
4395     return true;
4396   }
4397   return false;
4398 }
4399
4400 void Objecter::ms_handle_remote_reset(Connection *con)
4401 {
4402   /*
4403    * treat these the same.
4404    */
4405   ms_handle_reset(con);
4406 }
4407
4408 bool Objecter::ms_handle_refused(Connection *con)
4409 {
4410   // just log for now
4411   if (osdmap && (con->get_peer_type() == CEPH_ENTITY_TYPE_OSD)) {
4412     int osd = osdmap->identify_osd(con->get_peer_addr());
4413     if (osd >= 0) {
4414       ldout(cct, 1) << "ms_handle_refused on osd." << osd << dendl;
4415     }
4416   }
4417   return false;
4418 }
4419
4420 bool Objecter::ms_get_authorizer(int dest_type,
4421                                  AuthAuthorizer **authorizer,
4422                                  bool force_new)
4423 {
4424   if (!initialized)
4425     return false;
4426   if (dest_type == CEPH_ENTITY_TYPE_MON)
4427     return true;
4428   *authorizer = monc->build_authorizer(dest_type);
4429   return *authorizer != NULL;
4430 }
4431
4432 void Objecter::op_target_t::dump(Formatter *f) const
4433 {
4434   f->dump_stream("pg") << pgid;
4435   f->dump_int("osd", osd);
4436   f->dump_stream("object_id") << base_oid;
4437   f->dump_stream("object_locator") << base_oloc;
4438   f->dump_stream("target_object_id") << target_oid;
4439   f->dump_stream("target_object_locator") << target_oloc;
4440   f->dump_int("paused", (int)paused);
4441   f->dump_int("used_replica", (int)used_replica);
4442   f->dump_int("precalc_pgid", (int)precalc_pgid);
4443 }
4444
4445 void Objecter::_dump_active(OSDSession *s)
4446 {
4447   for (map<ceph_tid_t,Op*>::iterator p = s->ops.begin();
4448        p != s->ops.end();
4449        ++p) {
4450     Op *op = p->second;
4451     ldout(cct, 20) << op->tid << "\t" << op->target.pgid
4452                    << "\tosd." << (op->session ? op->session->osd : -1)
4453                    << "\t" << op->target.base_oid
4454                    << "\t" << op->ops << dendl;
4455   }
4456 }
4457
4458 void Objecter::_dump_active()
4459 {
4460   ldout(cct, 20) << "dump_active .. " << num_homeless_ops << " homeless"
4461                  << dendl;
4462   for (map<int, OSDSession *>::iterator siter = osd_sessions.begin();
4463        siter != osd_sessions.end(); ++siter) {
4464     OSDSession *s = siter->second;
4465     OSDSession::shared_lock sl(s->lock);
4466     _dump_active(s);
4467     sl.unlock();
4468   }
4469   _dump_active(homeless_session);
4470 }
4471
4472 void Objecter::dump_active()
4473 {
4474   shared_lock rl(rwlock);
4475   _dump_active();
4476   rl.unlock();
4477 }
4478
4479 void Objecter::dump_requests(Formatter *fmt)
4480 {
4481   // Read-lock on Objecter held here
4482   fmt->open_object_section("requests");
4483   dump_ops(fmt);
4484   dump_linger_ops(fmt);
4485   dump_pool_ops(fmt);
4486   dump_pool_stat_ops(fmt);
4487   dump_statfs_ops(fmt);
4488   dump_command_ops(fmt);
4489   fmt->close_section(); // requests object
4490 }
4491
4492 void Objecter::_dump_ops(const OSDSession *s, Formatter *fmt)
4493 {
4494   for (map<ceph_tid_t,Op*>::const_iterator p = s->ops.begin();
4495        p != s->ops.end();
4496        ++p) {
4497     Op *op = p->second;
4498     fmt->open_object_section("op");
4499     fmt->dump_unsigned("tid", op->tid);
4500     op->target.dump(fmt);
4501     fmt->dump_stream("last_sent") << op->stamp;
4502     fmt->dump_int("attempts", op->attempts);
4503     fmt->dump_stream("snapid") << op->snapid;
4504     fmt->dump_stream("snap_context") << op->snapc;
4505     fmt->dump_stream("mtime") << op->mtime;
4506
4507     fmt->open_array_section("osd_ops");
4508     for (vector<OSDOp>::const_iterator it = op->ops.begin();
4509          it != op->ops.end();
4510          ++it) {
4511       fmt->dump_stream("osd_op") << *it;
4512     }
4513     fmt->close_section(); // osd_ops array
4514
4515     fmt->close_section(); // op object
4516   }
4517 }
4518
4519 void Objecter::dump_ops(Formatter *fmt)
4520 {
4521   // Read-lock on Objecter held
4522   fmt->open_array_section("ops");
4523   for (map<int, OSDSession *>::const_iterator siter = osd_sessions.begin();
4524        siter != osd_sessions.end(); ++siter) {
4525     OSDSession *s = siter->second;
4526     OSDSession::shared_lock sl(s->lock);
4527     _dump_ops(s, fmt);
4528     sl.unlock();
4529   }
4530   _dump_ops(homeless_session, fmt);
4531   fmt->close_section(); // ops array
4532 }
4533
4534 void Objecter::_dump_linger_ops(const OSDSession *s, Formatter *fmt)
4535 {
4536   for (map<uint64_t, LingerOp*>::const_iterator p = s->linger_ops.begin();
4537        p != s->linger_ops.end();
4538        ++p) {
4539     LingerOp *op = p->second;
4540     fmt->open_object_section("linger_op");
4541     fmt->dump_unsigned("linger_id", op->linger_id);
4542     op->target.dump(fmt);
4543     fmt->dump_stream("snapid") << op->snap;
4544     fmt->dump_stream("registered") << op->registered;
4545     fmt->close_section(); // linger_op object
4546   }
4547 }
4548
4549 void Objecter::dump_linger_ops(Formatter *fmt)
4550 {
4551   // We have a read-lock on the objecter
4552   fmt->open_array_section("linger_ops");
4553   for (map<int, OSDSession *>::const_iterator siter = osd_sessions.begin();
4554        siter != osd_sessions.end(); ++siter) {
4555     OSDSession *s = siter->second;
4556     OSDSession::shared_lock sl(s->lock);
4557     _dump_linger_ops(s, fmt);
4558     sl.unlock();
4559   }
4560   _dump_linger_ops(homeless_session, fmt);
4561   fmt->close_section(); // linger_ops array
4562 }
4563
4564 void Objecter::_dump_command_ops(const OSDSession *s, Formatter *fmt)
4565 {
4566   for (map<uint64_t, CommandOp*>::const_iterator p = s->command_ops.begin();
4567        p != s->command_ops.end();
4568        ++p) {
4569     CommandOp *op = p->second;
4570     fmt->open_object_section("command_op");
4571     fmt->dump_unsigned("command_id", op->tid);
4572     fmt->dump_int("osd", op->session ? op->session->osd : -1);
4573     fmt->open_array_section("command");
4574     for (vector<string>::const_iterator q = op->cmd.begin();
4575          q != op->cmd.end(); ++q)
4576       fmt->dump_string("word", *q);
4577     fmt->close_section();
4578     if (op->target_osd >= 0)
4579       fmt->dump_int("target_osd", op->target_osd);
4580     else
4581       fmt->dump_stream("target_pg") << op->target_pg;
4582     fmt->close_section(); // command_op object
4583   }
4584 }
4585
4586 void Objecter::dump_command_ops(Formatter *fmt)
4587 {
4588   // We have a read-lock on the Objecter here
4589   fmt->open_array_section("command_ops");
4590   for (map<int, OSDSession *>::const_iterator siter = osd_sessions.begin();
4591        siter != osd_sessions.end(); ++siter) {
4592     OSDSession *s = siter->second;
4593     OSDSession::shared_lock sl(s->lock);
4594     _dump_command_ops(s, fmt);
4595     sl.unlock();
4596   }
4597   _dump_command_ops(homeless_session, fmt);
4598   fmt->close_section(); // command_ops array
4599 }
4600
4601 void Objecter::dump_pool_ops(Formatter *fmt) const
4602 {
4603   fmt->open_array_section("pool_ops");
4604   for (map<ceph_tid_t, PoolOp*>::const_iterator p = pool_ops.begin();
4605        p != pool_ops.end();
4606        ++p) {
4607     PoolOp *op = p->second;
4608     fmt->open_object_section("pool_op");
4609     fmt->dump_unsigned("tid", op->tid);
4610     fmt->dump_int("pool", op->pool);
4611     fmt->dump_string("name", op->name);
4612     fmt->dump_int("operation_type", op->pool_op);
4613     fmt->dump_unsigned("auid", op->auid);
4614     fmt->dump_unsigned("crush_rule", op->crush_rule);
4615     fmt->dump_stream("snapid") << op->snapid;
4616     fmt->dump_stream("last_sent") << op->last_submit;
4617     fmt->close_section(); // pool_op object
4618   }
4619   fmt->close_section(); // pool_ops array
4620 }
4621
4622 void Objecter::dump_pool_stat_ops(Formatter *fmt) const
4623 {
4624   fmt->open_array_section("pool_stat_ops");
4625   for (map<ceph_tid_t, PoolStatOp*>::const_iterator p = poolstat_ops.begin();
4626        p != poolstat_ops.end();
4627        ++p) {
4628     PoolStatOp *op = p->second;
4629     fmt->open_object_section("pool_stat_op");
4630     fmt->dump_unsigned("tid", op->tid);
4631     fmt->dump_stream("last_sent") << op->last_submit;
4632
4633     fmt->open_array_section("pools");
4634     for (list<string>::const_iterator it = op->pools.begin();
4635          it != op->pools.end();
4636          ++it) {
4637       fmt->dump_string("pool", *it);
4638     }
4639     fmt->close_section(); // pools array
4640
4641     fmt->close_section(); // pool_stat_op object
4642   }
4643   fmt->close_section(); // pool_stat_ops array
4644 }
4645
4646 void Objecter::dump_statfs_ops(Formatter *fmt) const
4647 {
4648   fmt->open_array_section("statfs_ops");
4649   for (map<ceph_tid_t, StatfsOp*>::const_iterator p = statfs_ops.begin();
4650        p != statfs_ops.end();
4651        ++p) {
4652     StatfsOp *op = p->second;
4653     fmt->open_object_section("statfs_op");
4654     fmt->dump_unsigned("tid", op->tid);
4655     fmt->dump_stream("last_sent") << op->last_submit;
4656     fmt->close_section(); // statfs_op object
4657   }
4658   fmt->close_section(); // statfs_ops array
4659 }
4660
4661 Objecter::RequestStateHook::RequestStateHook(Objecter *objecter) :
4662   m_objecter(objecter)
4663 {
4664 }
4665
4666 bool Objecter::RequestStateHook::call(std::string command, cmdmap_t& cmdmap,
4667                                       std::string format, bufferlist& out)
4668 {
4669   Formatter *f = Formatter::create(format, "json-pretty", "json-pretty");
4670   shared_lock rl(m_objecter->rwlock);
4671   m_objecter->dump_requests(f);
4672   f->flush(out);
4673   delete f;
4674   return true;
4675 }
4676
4677 void Objecter::blacklist_self(bool set)
4678 {
4679   ldout(cct, 10) << "blacklist_self " << (set ? "add" : "rm") << dendl;
4680
4681   vector<string> cmd;
4682   cmd.push_back("{\"prefix\":\"osd blacklist\", ");
4683   if (set)
4684     cmd.push_back("\"blacklistop\":\"add\",");
4685   else
4686     cmd.push_back("\"blacklistop\":\"rm\",");
4687   stringstream ss;
4688   ss << messenger->get_myaddr();
4689   cmd.push_back("\"addr\":\"" + ss.str() + "\"");
4690
4691   MMonCommand *m = new MMonCommand(monc->get_fsid());
4692   m->cmd = cmd;
4693
4694   monc->send_mon_message(m);
4695 }
4696
4697 // commands
4698
4699 void Objecter::handle_command_reply(MCommandReply *m)
4700 {
4701   unique_lock wl(rwlock);
4702   if (!initialized) {
4703     m->put();
4704     return;
4705   }
4706
4707   ConnectionRef con = m->get_connection();
4708   OSDSession *s = static_cast<OSDSession*>(con->get_priv());
4709   if (!s || s->con != con) {
4710     ldout(cct, 7) << __func__ << " no session on con " << con << dendl;
4711     m->put();
4712     if (s)
4713       s->put();
4714     return;
4715   }
4716
4717   OSDSession::shared_lock sl(s->lock);
4718   map<ceph_tid_t,CommandOp*>::iterator p = s->command_ops.find(m->get_tid());
4719   if (p == s->command_ops.end()) {
4720     ldout(cct, 10) << "handle_command_reply tid " << m->get_tid()
4721                    << " not found" << dendl;
4722     m->put();
4723     sl.unlock();
4724     if (s)
4725       s->put();
4726     return;
4727   }
4728
4729   CommandOp *c = p->second;
4730   if (!c->session ||
4731       m->get_connection() != c->session->con) {
4732     ldout(cct, 10) << "handle_command_reply tid " << m->get_tid()
4733                    << " got reply from wrong connection "
4734                    << m->get_connection() << " " << m->get_source_inst()
4735                    << dendl;
4736     m->put();
4737     sl.unlock();
4738     if (s)
4739       s->put();
4740     return;
4741   }
4742   if (c->poutbl) {
4743     c->poutbl->claim(m->get_data());
4744   }
4745
4746   sl.unlock();
4747
4748
4749   _finish_command(c, m->r, m->rs);
4750   m->put();
4751   if (s)
4752     s->put();
4753 }
4754
4755 void Objecter::submit_command(CommandOp *c, ceph_tid_t *ptid)
4756 {
4757   shunique_lock sul(rwlock, ceph::acquire_unique);
4758
4759   ceph_tid_t tid = ++last_tid;
4760   ldout(cct, 10) << "_submit_command " << tid << " " << c->cmd << dendl;
4761   c->tid = tid;
4762
4763   {
4764     OSDSession::unique_lock hs_wl(homeless_session->lock);
4765     _session_command_op_assign(homeless_session, c);
4766   }
4767
4768   _calc_command_target(c, sul);
4769   _assign_command_session(c, sul);
4770   if (osd_timeout > timespan(0)) {
4771     c->ontimeout = timer.add_event(osd_timeout,
4772                                    [this, c, tid]() {
4773                                      command_op_cancel(c->session, tid,
4774                                                        -ETIMEDOUT); });
4775   }
4776
4777   if (!c->session->is_homeless()) {
4778     _send_command(c);
4779   } else {
4780     _maybe_request_map();
4781   }
4782   if (c->map_check_error)
4783     _send_command_map_check(c);
4784   *ptid = tid;
4785
4786   logger->inc(l_osdc_command_active);
4787 }
4788
4789 int Objecter::_calc_command_target(CommandOp *c, shunique_lock& sul)
4790 {
4791   assert(sul.owns_lock() && sul.mutex() == &rwlock);
4792
4793   c->map_check_error = 0;
4794
4795   // ignore overlays, just like we do with pg ops
4796   c->target.flags |= CEPH_OSD_FLAG_IGNORE_OVERLAY;
4797
4798   if (c->target_osd >= 0) {
4799     if (!osdmap->exists(c->target_osd)) {
4800       c->map_check_error = -ENOENT;
4801       c->map_check_error_str = "osd dne";
4802       c->target.osd = -1;
4803       return RECALC_OP_TARGET_OSD_DNE;
4804     }
4805     if (osdmap->is_down(c->target_osd)) {
4806       c->map_check_error = -ENXIO;
4807       c->map_check_error_str = "osd down";
4808       c->target.osd = -1;
4809       return RECALC_OP_TARGET_OSD_DOWN;
4810     }
4811     c->target.osd = c->target_osd;
4812   } else {
4813     int ret = _calc_target(&(c->target), nullptr, true);
4814     if (ret == RECALC_OP_TARGET_POOL_DNE) {
4815       c->map_check_error = -ENOENT;
4816       c->map_check_error_str = "pool dne";
4817       c->target.osd = -1;
4818       return ret;
4819     } else if (ret == RECALC_OP_TARGET_OSD_DOWN) {
4820       c->map_check_error = -ENXIO;
4821       c->map_check_error_str = "osd down";
4822       c->target.osd = -1;
4823       return ret;
4824     }
4825   }
4826
4827   OSDSession *s;
4828   int r = _get_session(c->target.osd, &s, sul);
4829   assert(r != -EAGAIN); /* shouldn't happen as we're holding the write lock */
4830
4831   if (c->session != s) {
4832     put_session(s);
4833     return RECALC_OP_TARGET_NEED_RESEND;
4834   }
4835
4836   put_session(s);
4837
4838   ldout(cct, 20) << "_recalc_command_target " << c->tid << " no change, "
4839                  << c->session << dendl;
4840
4841   return RECALC_OP_TARGET_NO_ACTION;
4842 }
4843
4844 void Objecter::_assign_command_session(CommandOp *c,
4845                                        shunique_lock& sul)
4846 {
4847   assert(sul.owns_lock() && sul.mutex() == &rwlock);
4848
4849   OSDSession *s;
4850   int r = _get_session(c->target.osd, &s, sul);
4851   assert(r != -EAGAIN); /* shouldn't happen as we're holding the write lock */
4852
4853   if (c->session != s) {
4854     if (c->session) {
4855       OSDSession *cs = c->session;
4856       OSDSession::unique_lock csl(cs->lock);
4857       _session_command_op_remove(c->session, c);
4858       csl.unlock();
4859     }
4860     OSDSession::unique_lock sl(s->lock);
4861     _session_command_op_assign(s, c);
4862   }
4863
4864   put_session(s);
4865 }
4866
4867 void Objecter::_send_command(CommandOp *c)
4868 {
4869   ldout(cct, 10) << "_send_command " << c->tid << dendl;
4870   assert(c->session);
4871   assert(c->session->con);
4872   MCommand *m = new MCommand(monc->monmap.fsid);
4873   m->cmd = c->cmd;
4874   m->set_data(c->inbl);
4875   m->set_tid(c->tid);
4876   c->session->con->send_message(m);
4877   logger->inc(l_osdc_command_send);
4878 }
4879
4880 int Objecter::command_op_cancel(OSDSession *s, ceph_tid_t tid, int r)
4881 {
4882   assert(initialized);
4883
4884   unique_lock wl(rwlock);
4885
4886   map<ceph_tid_t, CommandOp*>::iterator it = s->command_ops.find(tid);
4887   if (it == s->command_ops.end()) {
4888     ldout(cct, 10) << __func__ << " tid " << tid << " dne" << dendl;
4889     return -ENOENT;
4890   }
4891
4892   ldout(cct, 10) << __func__ << " tid " << tid << dendl;
4893
4894   CommandOp *op = it->second;
4895   _command_cancel_map_check(op);
4896   _finish_command(op, r, "");
4897   return 0;
4898 }
4899
4900 void Objecter::_finish_command(CommandOp *c, int r, string rs)
4901 {
4902   // rwlock is locked unique
4903
4904   ldout(cct, 10) << "_finish_command " << c->tid << " = " << r << " "
4905                  << rs << dendl;
4906   if (c->prs)
4907     *c->prs = rs;
4908   if (c->onfinish)
4909     c->onfinish->complete(r);
4910
4911   if (c->ontimeout && r != -ETIMEDOUT)
4912     timer.cancel_event(c->ontimeout);
4913
4914   OSDSession *s = c->session;
4915   OSDSession::unique_lock sl(s->lock);
4916   _session_command_op_remove(c->session, c);
4917   sl.unlock();
4918
4919   c->put();
4920
4921   logger->dec(l_osdc_command_active);
4922 }
4923
4924 Objecter::OSDSession::~OSDSession()
4925 {
4926   // Caller is responsible for re-assigning or
4927   // destroying any ops that were assigned to us
4928   assert(ops.empty());
4929   assert(linger_ops.empty());
4930   assert(command_ops.empty());
4931 }
4932
4933 Objecter::~Objecter()
4934 {
4935   delete osdmap;
4936
4937   assert(homeless_session->get_nref() == 1);
4938   assert(num_homeless_ops == 0);
4939   homeless_session->put();
4940
4941   assert(osd_sessions.empty());
4942   assert(poolstat_ops.empty());
4943   assert(statfs_ops.empty());
4944   assert(pool_ops.empty());
4945   assert(waiting_for_map.empty());
4946   assert(linger_ops.empty());
4947   assert(check_latest_map_lingers.empty());
4948   assert(check_latest_map_ops.empty());
4949   assert(check_latest_map_commands.empty());
4950
4951   assert(!m_request_state_hook);
4952   assert(!logger);
4953 }
4954
4955 /**
4956  * Wait until this OSD map epoch is received before
4957  * sending any more operations to OSDs.  Use this
4958  * when it is known that the client can't trust
4959  * anything from before this epoch (e.g. due to
4960  * client blacklist at this epoch).
4961  */
4962 void Objecter::set_epoch_barrier(epoch_t epoch)
4963 {
4964   unique_lock wl(rwlock);
4965
4966   ldout(cct, 7) << __func__ << ": barrier " << epoch << " (was "
4967                 << epoch_barrier << ") current epoch " << osdmap->get_epoch()
4968                 << dendl;
4969   if (epoch > epoch_barrier) {
4970     epoch_barrier = epoch;
4971     _maybe_request_map();
4972   }
4973 }
4974
4975
4976
4977 hobject_t Objecter::enumerate_objects_begin()
4978 {
4979   return hobject_t();
4980 }
4981
4982 hobject_t Objecter::enumerate_objects_end()
4983 {
4984   return hobject_t::get_max();
4985 }
4986
4987 struct C_EnumerateReply : public Context {
4988   bufferlist bl;
4989
4990   Objecter *objecter;
4991   hobject_t *next;
4992   std::list<librados::ListObjectImpl> *result;
4993   const hobject_t end;
4994   const int64_t pool_id;
4995   Context *on_finish;
4996
4997   epoch_t epoch;
4998   int budget;
4999
5000   C_EnumerateReply(Objecter *objecter_, hobject_t *next_,
5001       std::list<librados::ListObjectImpl> *result_,
5002       const hobject_t end_, const int64_t pool_id_, Context *on_finish_) :
5003     objecter(objecter_), next(next_), result(result_),
5004     end(end_), pool_id(pool_id_), on_finish(on_finish_),
5005     epoch(0), budget(0)
5006   {}
5007
5008   void finish(int r) override {
5009     objecter->_enumerate_reply(
5010       bl, r, end, pool_id, budget, epoch, result, next, on_finish);
5011   }
5012 };
5013
5014 void Objecter::enumerate_objects(
5015     int64_t pool_id,
5016     const std::string &ns,
5017     const hobject_t &start,
5018     const hobject_t &end,
5019     const uint32_t max,
5020     const bufferlist &filter_bl,
5021     std::list<librados::ListObjectImpl> *result,
5022     hobject_t *next,
5023     Context *on_finish)
5024 {
5025   assert(result);
5026
5027   if (!end.is_max() && start > end) {
5028     lderr(cct) << __func__ << ": start " << start << " > end " << end << dendl;
5029     on_finish->complete(-EINVAL);
5030     return;
5031   }
5032
5033   if (max < 1) {
5034     lderr(cct) << __func__ << ": result size may not be zero" << dendl;
5035     on_finish->complete(-EINVAL);
5036     return;
5037   }
5038
5039   if (start.is_max()) {
5040     on_finish->complete(0);
5041     return;
5042   }
5043
5044   shared_lock rl(rwlock);
5045   assert(osdmap->get_epoch());
5046   if (!osdmap->test_flag(CEPH_OSDMAP_SORTBITWISE)) {
5047     rl.unlock();
5048     lderr(cct) << __func__ << ": SORTBITWISE cluster flag not set" << dendl;
5049     on_finish->complete(-EOPNOTSUPP);
5050     return;
5051   }
5052   const pg_pool_t *p = osdmap->get_pg_pool(pool_id);
5053   if (!p) {
5054     lderr(cct) << __func__ << ": pool " << pool_id << " DNE in osd epoch "
5055                << osdmap->get_epoch() << dendl;
5056     rl.unlock();
5057     on_finish->complete(-ENOENT);
5058     return;
5059   } else {
5060     rl.unlock();
5061   }
5062
5063   ldout(cct, 20) << __func__ << ": start=" << start << " end=" << end << dendl;
5064
5065   // Stash completion state
5066   C_EnumerateReply *on_ack = new C_EnumerateReply(
5067       this, next, result, end, pool_id, on_finish);
5068
5069   ObjectOperation op;
5070   op.pg_nls(max, filter_bl, start, 0);
5071
5072   // Issue.  See you later in _enumerate_reply
5073   object_locator_t oloc(pool_id, ns);
5074   pg_read(start.get_hash(), oloc, op,
5075           &on_ack->bl, 0, on_ack, &on_ack->epoch, &on_ack->budget);
5076 }
5077
5078 void Objecter::_enumerate_reply(
5079     bufferlist &bl,
5080     int r,
5081     const hobject_t &end,
5082     const int64_t pool_id,
5083     int budget,
5084     epoch_t reply_epoch,
5085     std::list<librados::ListObjectImpl> *result,
5086     hobject_t *next,
5087     Context *on_finish)
5088 {
5089   if (budget > 0) {
5090     put_op_budget_bytes(budget);
5091   }
5092
5093   if (r < 0) {
5094     ldout(cct, 4) << __func__ << ": remote error " << r << dendl;
5095     on_finish->complete(r);
5096     return;
5097   }
5098
5099   assert(next != NULL);
5100
5101   // Decode the results
5102   bufferlist::iterator iter = bl.begin();
5103   pg_nls_response_t response;
5104
5105   // XXX extra_info doesn't seem used anywhere?
5106   bufferlist extra_info;
5107   ::decode(response, iter);
5108   if (!iter.end()) {
5109     ::decode(extra_info, iter);
5110   }
5111
5112   ldout(cct, 10) << __func__ << ": got " << response.entries.size()
5113                  << " handle " << response.handle
5114                  << " reply_epoch " << reply_epoch << dendl;
5115   ldout(cct, 20) << __func__ << ": response.entries.size "
5116                  << response.entries.size() << ", response.entries "
5117                  << response.entries << dendl;
5118   if (response.handle <= end) {
5119     *next = response.handle;
5120   } else {
5121     ldout(cct, 10) << __func__ << ": adjusted next down to end " << end
5122                    << dendl;
5123     *next = end;
5124
5125     // drop anything after 'end'
5126     shared_lock rl(rwlock);
5127     const pg_pool_t *pool = osdmap->get_pg_pool(pool_id);
5128     if (!pool) {
5129       // pool is gone, drop any results which are now meaningless.
5130       rl.unlock();
5131       on_finish->complete(-ENOENT);
5132       return;
5133     }
5134     while (!response.entries.empty()) {
5135       uint32_t hash = response.entries.back().locator.empty() ?
5136         pool->hash_key(response.entries.back().oid,
5137                        response.entries.back().nspace) :
5138         pool->hash_key(response.entries.back().locator,
5139                        response.entries.back().nspace);
5140       hobject_t last(response.entries.back().oid,
5141                      response.entries.back().locator,
5142                      CEPH_NOSNAP,
5143                      hash,
5144                      pool_id,
5145                      response.entries.back().nspace);
5146       if (last < end)
5147         break;
5148       ldout(cct, 20) << __func__ << " dropping item " << last
5149                      << " >= end " << end << dendl;
5150       response.entries.pop_back();
5151     }
5152     rl.unlock();
5153   }
5154   if (!response.entries.empty()) {
5155     result->merge(response.entries);
5156   }
5157
5158   // release the listing context's budget once all
5159   // OPs (in the session) are finished
5160 #if 0
5161   put_nlist_context_budget(list_context);
5162 #endif
5163   on_finish->complete(r);
5164   return;
5165 }
5166
5167 namespace {
5168   using namespace librados;
5169
5170   template <typename T>
5171   void do_decode(std::vector<T>& items, std::vector<bufferlist>& bls)
5172   {
5173     for (auto bl : bls) {
5174       auto p = bl.begin();
5175       T t;
5176       decode(t, p);
5177       items.push_back(t);
5178     }
5179   }
5180
5181   struct C_ObjectOperation_scrub_ls : public Context {
5182     bufferlist bl;
5183     uint32_t *interval;
5184     std::vector<inconsistent_obj_t> *objects = nullptr;
5185     std::vector<inconsistent_snapset_t> *snapsets = nullptr;
5186     int *rval;
5187
5188     C_ObjectOperation_scrub_ls(uint32_t *interval,
5189                                std::vector<inconsistent_obj_t> *objects,
5190                                int *rval)
5191       : interval(interval), objects(objects), rval(rval) {}
5192     C_ObjectOperation_scrub_ls(uint32_t *interval,
5193                                std::vector<inconsistent_snapset_t> *snapsets,
5194                                int *rval)
5195       : interval(interval), snapsets(snapsets), rval(rval) {}
5196     void finish(int r) override {
5197       if (r < 0 && r != -EAGAIN) {
5198         if (rval)
5199           *rval = r;
5200         return;
5201       }
5202
5203       if (rval)
5204         *rval = 0;
5205
5206       try {
5207         decode();
5208       } catch (buffer::error&) {
5209         if (rval)
5210           *rval = -EIO;
5211       }
5212     }
5213   private:
5214     void decode() {
5215       scrub_ls_result_t result;
5216       auto p = bl.begin();
5217       result.decode(p);
5218       *interval = result.interval;
5219       if (objects) {
5220         do_decode(*objects, result.vals);
5221       } else {
5222         do_decode(*snapsets, result.vals);
5223       }
5224     }
5225   };
5226
5227   template <typename T>
5228   void do_scrub_ls(::ObjectOperation *op,
5229                    const scrub_ls_arg_t& arg,
5230                    std::vector<T> *items,
5231                    uint32_t *interval,
5232                    int *rval)
5233   {
5234     OSDOp& osd_op = op->add_op(CEPH_OSD_OP_SCRUBLS);
5235     op->flags |= CEPH_OSD_FLAG_PGOP;
5236     assert(interval);
5237     arg.encode(osd_op.indata);
5238     unsigned p = op->ops.size() - 1;
5239     auto *h = new C_ObjectOperation_scrub_ls{interval, items, rval};
5240     op->out_handler[p] = h;
5241     op->out_bl[p] = &h->bl;
5242     op->out_rval[p] = rval;
5243   }
5244 }
5245
5246 void ::ObjectOperation::scrub_ls(const librados::object_id_t& start_after,
5247                                  uint64_t max_to_get,
5248                                  std::vector<librados::inconsistent_obj_t> *objects,
5249                                  uint32_t *interval,
5250                                  int *rval)
5251 {
5252   scrub_ls_arg_t arg = {*interval, 0, start_after, max_to_get};
5253   do_scrub_ls(this, arg, objects, interval, rval);
5254 }
5255
5256 void ::ObjectOperation::scrub_ls(const librados::object_id_t& start_after,
5257                                  uint64_t max_to_get,
5258                                  std::vector<librados::inconsistent_snapset_t> *snapsets,
5259                                  uint32_t *interval,
5260                                  int *rval)
5261 {
5262   scrub_ls_arg_t arg = {*interval, 1, start_after, max_to_get};
5263   do_scrub_ls(this, arg, snapsets, interval, rval);
5264 }