ceph/src/msg/async/AsyncMessenger.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2014 UnitedStack <haomai@unitedstack.com>
   7  *
   8  * Author: Haomai Wang <haomaiwang@gmail.com>
   9  *
  10  * This is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License version 2.1, as published by the Free Software
  13  * Foundation.  See file COPYING.
  14  *
  15  */
  16
  17 #include "acconfig.h"
  18
  19 #include <iostream>
  20 #include <fstream>
  21
  22 #include "AsyncMessenger.h"
  23
  24 #include "common/config.h"
  25 #include "common/Timer.h"
  26 #include "common/errno.h"
  27
  28 #include "messages/MOSDOp.h"
  29 #include "messages/MOSDOpReply.h"
  30 #include "common/EventTrace.h"
  31
  32 #define dout_subsys ceph_subsys_ms
  33 #undef dout_prefix
  34 #define dout_prefix _prefix(_dout, this)
  35 static ostream& _prefix(std::ostream *_dout, AsyncMessenger *m) {
  36   return *_dout << "-- " << m->get_myaddrs() << " ";
  37 }
  38
  39 static ostream& _prefix(std::ostream *_dout, Processor *p) {
  40   return *_dout << " Processor -- ";
  41 }
  42
  43
  44 /*******************
  45  * Processor
  46  */
  47
  48 class Processor::C_processor_accept : public EventCallback {
  49   Processor *pro;
  50
  51  public:
  52   explicit C_processor_accept(Processor *p): pro(p) {}
  53   void do_request(uint64_t id) override {
  54     pro->accept();
  55   }
  56 };
  57
  58 Processor::Processor(AsyncMessenger *r, Worker *w, CephContext *c)
  59   : msgr(r), net(c), worker(w),
  60     listen_handler(new C_processor_accept(this)) {}
  61
  62 int Processor::bind(const entity_addrvec_t &bind_addrs,
  63                     const set<int>& avoid_ports,
  64                     entity_addrvec_t* bound_addrs)
  65 {
  66   const auto& conf = msgr->cct->_conf;
  67   // bind to socket(s)
  68   ldout(msgr->cct, 10) << __func__ << " " << bind_addrs << dendl;
  69
  70   SocketOptions opts;
  71   opts.nodelay = msgr->cct->_conf->ms_tcp_nodelay;
  72   opts.rcbuf_size = msgr->cct->_conf->ms_tcp_rcvbuf;
  73
  74   listen_sockets.resize(bind_addrs.v.size());
  75   *bound_addrs = bind_addrs;
  76
  77   for (unsigned k = 0; k < bind_addrs.v.size(); ++k) {
  78     auto& listen_addr = bound_addrs->v[k];
  79
  80     /* bind to port */
  81     int r = -1;
  82
  83     for (int i = 0; i < conf->ms_bind_retry_count; i++) {
  84       if (i > 0) {
  85         lderr(msgr->cct) << __func__ << " was unable to bind. Trying again in "
  86                          << conf->ms_bind_retry_delay << " seconds " << dendl;
  87         sleep(conf->ms_bind_retry_delay);
  88       }
  89
  90       if (listen_addr.get_port()) {
  91         worker->center.submit_to(
  92           worker->center.get_id(),
  93           [this, k, &listen_addr, &opts, &r]() {
  94             r = worker->listen(listen_addr, k, opts, &listen_sockets[k]);
  95           }, false);
  96         if (r < 0) {
  97           lderr(msgr->cct) << __func__ << " unable to bind to " << listen_addr
  98                            << ": " << cpp_strerror(r) << dendl;
  99           continue;
 100         }
 101       } else {
 102         // try a range of ports
 103         for (int port = msgr->cct->_conf->ms_bind_port_min;
 104              port <= msgr->cct->_conf->ms_bind_port_max;
 105              port++) {
 106           if (avoid_ports.count(port))
 107             continue;
 108
 109           listen_addr.set_port(port);
 110           worker->center.submit_to(
 111             worker->center.get_id(),
 112             [this, k, &listen_addr, &opts, &r]() {
 113               r = worker->listen(listen_addr, k, opts, &listen_sockets[k]);
 114             }, false);
 115           if (r == 0)
 116             break;
 117         }
 118         if (r < 0) {
 119           lderr(msgr->cct) << __func__ << " unable to bind to " << listen_addr
 120                            << " on any port in range "
 121                            << msgr->cct->_conf->ms_bind_port_min
 122                            << "-" << msgr->cct->_conf->ms_bind_port_max << ": "
 123                            << cpp_strerror(r) << dendl;
 124           listen_addr.set_port(0); // Clear port before retry, otherwise we shall fail again.
 125           continue;
 126         }
 127         ldout(msgr->cct, 10) << __func__ << " bound on random port "
 128                              << listen_addr << dendl;
 129       }
 130       if (r == 0) {
 131         break;
 132       }
 133     }
 134
 135     // It seems that binding completely failed, return with that exit status
 136     if (r < 0) {
 137       lderr(msgr->cct) << __func__ << " was unable to bind after "
 138                        << conf->ms_bind_retry_count
 139                        << " attempts: " << cpp_strerror(r) << dendl;
 140       for (unsigned j = 0; j < k; ++j) {
 141         // clean up previous bind
 142         listen_sockets[j].abort_accept();
 143       }
 144       return r;
 145     }
 146   }
 147
 148   ldout(msgr->cct, 10) << __func__ << " bound to " << *bound_addrs << dendl;
 149   return 0;
 150 }
 151
 152 void Processor::start()
 153 {
 154   ldout(msgr->cct, 1) << __func__ << dendl;
 155
 156   // start thread
 157   worker->center.submit_to(worker->center.get_id(), [this]() {
 158       for (auto& listen_socket : listen_sockets) {
 159         if (listen_socket) {
 160           if (listen_socket.fd() == -1) {
 161             ldout(msgr->cct, 1) << __func__
 162                 << " Error: processor restart after listen_socket.fd closed. "
 163                 << this << dendl;
 164             return;
 165           }
 166           worker->center.create_file_event(listen_socket.fd(), EVENT_READABLE,
 167                                            listen_handler); }
 168       }
 169     }, false);
 170 }
 171
 172 void Processor::accept()
 173 {
 174   SocketOptions opts;
 175   opts.nodelay = msgr->cct->_conf->ms_tcp_nodelay;
 176   opts.rcbuf_size = msgr->cct->_conf->ms_tcp_rcvbuf;
 177   opts.priority = msgr->get_socket_priority();
 178
 179   for (auto& listen_socket : listen_sockets) {
 180     ldout(msgr->cct, 10) << __func__ << " listen_fd=" << listen_socket.fd()
 181                          << dendl;
 182     unsigned accept_error_num = 0;
 183
 184     while (true) {
 185       entity_addr_t addr;
 186       ConnectedSocket cli_socket;
 187       Worker *w = worker;
 188       if (!msgr->get_stack()->support_local_listen_table())
 189         w = msgr->get_stack()->get_worker();
 190       else
 191         ++w->references;
 192       int r = listen_socket.accept(&cli_socket, opts, &addr, w);
 193       if (r == 0) {
 194         ldout(msgr->cct, 10) << __func__ << " accepted incoming on sd "
 195                              << cli_socket.fd() << dendl;
 196
 197         msgr->add_accept(
 198           w, std::move(cli_socket),
 199           msgr->get_myaddrs().v[listen_socket.get_addr_slot()],
 200           addr);
 201         accept_error_num = 0;
 202         continue;
 203       } else {
 204         --w->references;
 205         if (r == -EINTR) {
 206           continue;
 207         } else if (r == -EAGAIN) {
 208           break;
 209         } else if (r == -EMFILE || r == -ENFILE) {
 210           lderr(msgr->cct) << __func__ << " open file descriptions limit reached sd = " << listen_socket.fd()
 211                            << " errno " << r << " " << cpp_strerror(r) << dendl;
 212           if (++accept_error_num > msgr->cct->_conf->ms_max_accept_failures) {
 213             lderr(msgr->cct) << "Proccessor accept has encountered enough error numbers, just do ceph_abort()." << dendl;
 214             ceph_abort();
 215           }
 216           continue;
 217         } else if (r == -ECONNABORTED) {
 218           ldout(msgr->cct, 0) << __func__ << " it was closed because of rst arrived sd = " << listen_socket.fd()
 219                               << " errno " << r << " " << cpp_strerror(r) << dendl;
 220           continue;
 221         } else {
 222           lderr(msgr->cct) << __func__ << " no incoming connection?"
 223                            << " errno " << r << " " << cpp_strerror(r) << dendl;
 224           if (++accept_error_num > msgr->cct->_conf->ms_max_accept_failures) {
 225             lderr(msgr->cct) << "Proccessor accept has encountered enough error numbers, just do ceph_abort()." << dendl;
 226             ceph_abort();
 227           }
 228           continue;
 229         }
 230       }
 231     }
 232   }
 233 }
 234
 235 void Processor::stop()
 236 {
 237   ldout(msgr->cct,10) << __func__ << dendl;
 238
 239   worker->center.submit_to(worker->center.get_id(), [this]() {
 240       for (auto& listen_socket : listen_sockets) {
 241         if (listen_socket) {
 242           worker->center.delete_file_event(listen_socket.fd(), EVENT_READABLE);
 243           listen_socket.abort_accept();
 244         }
 245       }
 246     }, false);
 247 }
 248
 249
 250 struct StackSingleton {
 251   CephContext *cct;
 252   std::shared_ptr<NetworkStack> stack;
 253
 254   explicit StackSingleton(CephContext *c): cct(c) {}
 255   void ready(std::string &type) {
 256     if (!stack)
 257       stack = NetworkStack::create(cct, type);
 258   }
 259   ~StackSingleton() {
 260     stack->stop();
 261   }
 262 };
 263
 264
 265 class C_handle_reap : public EventCallback {
 266   AsyncMessenger *msgr;
 267
 268   public:
 269   explicit C_handle_reap(AsyncMessenger *m): msgr(m) {}
 270   void do_request(uint64_t id) override {
 271     // judge whether is a time event
 272     msgr->reap_dead();
 273   }
 274 };
 275
 276 /*******************
 277  * AsyncMessenger
 278  */
 279
 280 AsyncMessenger::AsyncMessenger(CephContext *cct, entity_name_t name,
 281                                const std::string &type, string mname, uint64_t _nonce)
 282   : SimplePolicyMessenger(cct, name),
 283     dispatch_queue(cct, this, mname),
 284     nonce(_nonce)
 285 {
 286   std::string transport_type = "posix";
 287   if (type.find("rdma") != std::string::npos)
 288     transport_type = "rdma";
 289   else if (type.find("dpdk") != std::string::npos)
 290     transport_type = "dpdk";
 291
 292   auto single = &cct->lookup_or_create_singleton_object<StackSingleton>(
 293     "AsyncMessenger::NetworkStack::" + transport_type, true, cct);
 294   single->ready(transport_type);
 295   stack = single->stack.get();
 296   stack->start();
 297   local_worker = stack->get_worker();
 298   local_connection = ceph::make_ref<AsyncConnection>(cct, this, &dispatch_queue,
 299                                          local_worker, true, true);
 300   init_local_connection();
 301   reap_handler = new C_handle_reap(this);
 302   unsigned processor_num = 1;
 303   if (stack->support_local_listen_table())
 304     processor_num = stack->get_num_worker();
 305   for (unsigned i = 0; i < processor_num; ++i)
 306     processors.push_back(new Processor(this, stack->get_worker(i), cct));
 307 }
 308
 309 /**
 310  * Destroy the AsyncMessenger. Pretty simple since all the work is done
 311  * elsewhere.
 312  */
 313 AsyncMessenger::~AsyncMessenger()
 314 {
 315   delete reap_handler;
 316   ceph_assert(!did_bind); // either we didn't bind or we shut down the Processor
 317   for (auto &&p : processors)
 318     delete p;
 319 }
 320
 321 void AsyncMessenger::ready()
 322 {
 323   ldout(cct,10) << __func__ << " " << get_myaddrs() << dendl;
 324
 325   stack->ready();
 326   if (pending_bind) {
 327     int err = bindv(pending_bind_addrs);
 328     if (err) {
 329       lderr(cct) << __func__ << " postponed bind failed" << dendl;
 330       ceph_abort();
 331     }
 332   }
 333
 334   std::lock_guard l{lock};
 335   for (auto &&p : processors)
 336     p->start();
 337   dispatch_queue.start();
 338 }
 339
 340 int AsyncMessenger::shutdown()
 341 {
 342   ldout(cct,10) << __func__ << " " << get_myaddrs() << dendl;
 343
 344   // done!  clean up.
 345   for (auto &&p : processors)
 346     p->stop();
 347   mark_down_all();
 348   // break ref cycles on the loopback connection
 349   local_connection->clear_priv();
 350   local_connection->mark_down();
 351   did_bind = false;
 352   lock.lock();
 353   stop_cond.notify_all();
 354   stopped = true;
 355   lock.unlock();
 356   stack->drain();
 357   return 0;
 358 }
 359
 360 int AsyncMessenger::bind(const entity_addr_t &bind_addr)
 361 {
 362   ldout(cct,10) << __func__ << " " << bind_addr << dendl;
 363   // old bind() can take entity_addr_t(). new bindv() can take a
 364   // 0.0.0.0-like address but needs type and family to be set.
 365   auto a = bind_addr;
 366   if (a == entity_addr_t()) {
 367     a.set_type(entity_addr_t::TYPE_LEGACY);
 368     if (cct->_conf->ms_bind_ipv6) {
 369       a.set_family(AF_INET6);
 370     } else {
 371       a.set_family(AF_INET);
 372     }
 373   }
 374   return bindv(entity_addrvec_t(a));
 375 }
 376
 377 int AsyncMessenger::bindv(const entity_addrvec_t &bind_addrs)
 378 {
 379   lock.lock();
 380
 381   if (!pending_bind && started) {
 382     ldout(cct,10) << __func__ << " already started" << dendl;
 383     lock.unlock();
 384     return -1;
 385   }
 386
 387   ldout(cct,10) << __func__ << " " << bind_addrs << dendl;
 388
 389   if (!stack->is_ready()) {
 390     ldout(cct, 10) << __func__ << " Network Stack is not ready for bind yet - postponed" << dendl;
 391     pending_bind_addrs = bind_addrs;
 392     pending_bind = true;
 393     lock.unlock();
 394     return 0;
 395   }
 396
 397   lock.unlock();
 398
 399   // bind to a socket
 400   set<int> avoid_ports;
 401   entity_addrvec_t bound_addrs;
 402   unsigned i = 0;
 403   for (auto &&p : processors) {
 404     int r = p->bind(bind_addrs, avoid_ports, &bound_addrs);
 405     if (r) {
 406       // Note: this is related to local tcp listen table problem.
 407       // Posix(default kernel implementation) backend shares listen table
 408       // in the kernel, so all threads can use the same listen table naturally
 409       // and only one thread need to bind. But other backends(like dpdk) uses local
 410       // listen table, we need to bind/listen tcp port for each worker. So if the
 411       // first worker failed to bind, it could be think the normal error then handle
 412       // it, like port is used case. But if the first worker successfully to bind
 413       // but the second worker failed, it's not expected and we need to assert
 414       // here
 415       ceph_assert(i == 0);
 416       return r;
 417     }
 418     ++i;
 419   }
 420   _finish_bind(bind_addrs, bound_addrs);
 421   return 0;
 422 }
 423
 424 int AsyncMessenger::rebind(const set<int>& avoid_ports)
 425 {
 426   ldout(cct,1) << __func__ << " rebind avoid " << avoid_ports << dendl;
 427   ceph_assert(did_bind);
 428
 429   for (auto &&p : processors)
 430     p->stop();
 431   mark_down_all();
 432
 433   // adjust the nonce; we want our entity_addr_t to be truly unique.
 434   nonce += 1000000;
 435   ldout(cct, 10) << __func__ << " new nonce " << nonce
 436                  << " and addr " << get_myaddrs() << dendl;
 437
 438   entity_addrvec_t bound_addrs;
 439   entity_addrvec_t bind_addrs = get_myaddrs();
 440   set<int> new_avoid(avoid_ports);
 441   for (auto& a : bind_addrs.v) {
 442     new_avoid.insert(a.get_port());
 443     a.set_port(0);
 444   }
 445   ldout(cct, 10) << __func__ << " will try " << bind_addrs
 446                  << " and avoid ports " << new_avoid << dendl;
 447   unsigned i = 0;
 448   for (auto &&p : processors) {
 449     int r = p->bind(bind_addrs, avoid_ports, &bound_addrs);
 450     if (r) {
 451       ceph_assert(i == 0);
 452       return r;
 453     }
 454     ++i;
 455   }
 456   _finish_bind(bind_addrs, bound_addrs);
 457   for (auto &&p : processors) {
 458     p->start();
 459   }
 460   return 0;
 461 }
 462
 463 int AsyncMessenger::client_bind(const entity_addr_t &bind_addr)
 464 {
 465   if (!cct->_conf->ms_bind_before_connect)
 466     return 0;
 467   std::lock_guard l{lock};
 468   if (did_bind) {
 469     return 0;
 470   }
 471   if (started) {
 472     ldout(cct, 10) << __func__ << " already started" << dendl;
 473     return -1;
 474   }
 475   ldout(cct, 10) << __func__ << " " << bind_addr << dendl;
 476
 477   set_myaddrs(entity_addrvec_t(bind_addr));
 478   return 0;
 479 }
 480
 481 void AsyncMessenger::_finish_bind(const entity_addrvec_t& bind_addrs,
 482                                   const entity_addrvec_t& listen_addrs)
 483 {
 484   set_myaddrs(bind_addrs);
 485   for (auto& a : bind_addrs.v) {
 486     if (!a.is_blank_ip()) {
 487       learned_addr(a);
 488     }
 489   }
 490
 491   if (get_myaddrs().front().get_port() == 0) {
 492     set_myaddrs(listen_addrs);
 493   }
 494   entity_addrvec_t newaddrs = *my_addrs;
 495   for (auto& a : newaddrs.v) {
 496     a.set_nonce(nonce);
 497   }
 498   set_myaddrs(newaddrs);
 499
 500   init_local_connection();
 501
 502   ldout(cct,1) << __func__ << " bind my_addrs is " << get_myaddrs() << dendl;
 503   did_bind = true;
 504 }
 505
 506 int AsyncMessenger::start()
 507 {
 508   std::scoped_lock l{lock};
 509   ldout(cct,1) << __func__ << " start" << dendl;
 510
 511   // register at least one entity, first!
 512   ceph_assert(my_name.type() >= 0);
 513
 514   ceph_assert(!started);
 515   started = true;
 516   stopped = false;
 517
 518   if (!did_bind) {
 519     entity_addrvec_t newaddrs = *my_addrs;
 520     for (auto& a : newaddrs.v) {
 521       a.nonce = nonce;
 522     }
 523     set_myaddrs(newaddrs);
 524     _init_local_connection();
 525   }
 526
 527   return 0;
 528 }
 529
 530 void AsyncMessenger::wait()
 531 {
 532   {
 533     std::unique_lock locker{lock};
 534     if (!started) {
 535       return;
 536     }
 537     if (!stopped)
 538       stop_cond.wait(locker);
 539   }
 540   dispatch_queue.shutdown();
 541   if (dispatch_queue.is_started()) {
 542     ldout(cct, 10) << __func__ << ": waiting for dispatch queue" << dendl;
 543     dispatch_queue.wait();
 544     dispatch_queue.discard_local();
 545     ldout(cct, 10) << __func__ << ": dispatch queue is stopped" << dendl;
 546   }
 547
 548   // close all connections
 549   shutdown_connections(false);
 550   stack->drain();
 551
 552   ldout(cct, 10) << __func__ << ": done." << dendl;
 553   ldout(cct, 1) << __func__ << " complete." << dendl;
 554   started = false;
 555 }
 556
 557 void AsyncMessenger::add_accept(Worker *w, ConnectedSocket cli_socket,
 558                                 const entity_addr_t &listen_addr,
 559                                 const entity_addr_t &peer_addr)
 560 {
 561   std::lock_guard l{lock};
 562   auto conn = ceph::make_ref<AsyncConnection>(cct, this, &dispatch_queue, w,
 563                                                 listen_addr.is_msgr2(), false);
 564   conn->accept(std::move(cli_socket), listen_addr, peer_addr);
 565   accepting_conns.insert(conn);
 566 }
 567
 568 AsyncConnectionRef AsyncMessenger::create_connect(
 569   const entity_addrvec_t& addrs, int type, bool anon)
 570 {
 571   ceph_assert(ceph_mutex_is_locked(lock));
 572
 573   ldout(cct, 10) << __func__ << " " << addrs
 574       << ", creating connection and registering" << dendl;
 575
 576   // here is where we decide which of the addrs to connect to.  always prefer
 577   // the first one, if we support it.
 578   entity_addr_t target;
 579   for (auto& a : addrs.v) {
 580     if (!a.is_msgr2() && !a.is_legacy()) {
 581       continue;
 582     }
 583     // FIXME: for ipv4 vs ipv6, check whether local host can handle ipv6 before
 584     // trying it?  for now, just pick whichever is listed first.
 585     target = a;
 586     break;
 587   }
 588
 589   // create connection
 590   Worker *w = stack->get_worker();
 591   auto conn = ceph::make_ref<AsyncConnection>(cct, this, &dispatch_queue, w,
 592                                                 target.is_msgr2(), false);
 593   conn->anon = anon;
 594   conn->connect(addrs, type, target);
 595   if (anon) {
 596     anon_conns.insert(conn);
 597   } else {
 598     ceph_assert(!conns.count(addrs));
 599     ldout(cct, 10) << __func__ << " " << conn << " " << addrs << " "
 600                    << *conn->peer_addrs << dendl;
 601     conns[addrs] = conn;
 602   }
 603   w->get_perf_counter()->inc(l_msgr_active_connections);
 604
 605   return conn;
 606 }
 607
 608
 609 ConnectionRef AsyncMessenger::get_loopback_connection()
 610 {
 611   return local_connection;
 612 }
 613
 614 bool AsyncMessenger::should_use_msgr2()
 615 {
 616   // if we are bound to v1 only, and we are connecting to a v2 peer,
 617   // we cannot use the peer's v2 address. otherwise the connection
 618   // is assymetrical, because they would have to use v1 to connect
 619   // to us, and we would use v2, and connection race detection etc
 620   // would totally break down (among other things).  or, the other
 621   // end will be confused that we advertise ourselve with a v1
 622   // address only (that we bound to) but connected with protocol v2.
 623   return !did_bind || get_myaddrs().has_msgr2();
 624 }
 625
 626 entity_addrvec_t AsyncMessenger::_filter_addrs(const entity_addrvec_t& addrs)
 627 {
 628   if (!should_use_msgr2()) {
 629     ldout(cct, 10) << __func__ << " " << addrs << " limiting to v1 ()" << dendl;
 630     entity_addrvec_t r;
 631     for (auto& i : addrs.v) {
 632       if (i.is_msgr2()) {
 633         continue;
 634       }
 635       r.v.push_back(i);
 636     }
 637     return r;
 638   } else {
 639     return addrs;
 640   }
 641 }
 642
 643 int AsyncMessenger::send_to(Message *m, int type, const entity_addrvec_t& addrs)
 644 {
 645   FUNCTRACE(cct);
 646   ceph_assert(m);
 647
 648 #if defined(WITH_EVENTTRACE)
 649   if (m->get_type() == CEPH_MSG_OSD_OP)
 650     OID_EVENT_TRACE(((MOSDOp *)m)->get_oid().name.c_str(), "SEND_MSG_OSD_OP");
 651   else if (m->get_type() == CEPH_MSG_OSD_OPREPLY)
 652     OID_EVENT_TRACE(((MOSDOpReply *)m)->get_oid().name.c_str(), "SEND_MSG_OSD_OP_REPLY");
 653 #endif
 654
 655   ldout(cct, 1) << __func__ << "--> " << ceph_entity_type_name(type) << " "
 656       << addrs << " -- " << *m << " -- ?+"
 657       << m->get_data().length() << " " << m << dendl;
 658
 659   if (addrs.empty()) {
 660     ldout(cct,0) << __func__ <<  " message " << *m
 661         << " with empty dest " << addrs << dendl;
 662     m->put();
 663     return -EINVAL;
 664   }
 665
 666   if (cct->_conf->ms_dump_on_send) {
 667     m->encode(-1, MSG_CRC_ALL);
 668     ldout(cct, 0) << __func__ << " submit_message " << *m << "\n";
 669     m->get_payload().hexdump(*_dout);
 670     if (m->get_data().length() > 0) {
 671       *_dout << " data:\n";
 672       m->get_data().hexdump(*_dout);
 673     }
 674     *_dout << dendl;
 675     m->clear_payload();
 676   }
 677
 678   connect_to(type, addrs, false)->send_message(m);
 679   return 0;
 680 }
 681
 682 ConnectionRef AsyncMessenger::connect_to(int type,
 683                                          const entity_addrvec_t& addrs,
 684                                          bool anon, bool not_local_dest)
 685 {
 686   if (!not_local_dest) {
 687     if (*my_addrs == addrs ||
 688         (addrs.v.size() == 1 &&
 689          my_addrs->contains(addrs.front()))) {
 690       // local
 691       return local_connection;
 692     }
 693   }
 694
 695   auto av = _filter_addrs(addrs);
 696   std::lock_guard l{lock};
 697   if (anon) {
 698     return create_connect(av, type, anon);
 699   }
 700
 701   AsyncConnectionRef conn = _lookup_conn(av);
 702   if (conn) {
 703     ldout(cct, 10) << __func__ << " " << av << " existing " << conn << dendl;
 704   } else {
 705     conn = create_connect(av, type, false);
 706     ldout(cct, 10) << __func__ << " " << av << " new " << conn << dendl;
 707   }
 708
 709   return conn;
 710 }
 711
 712 /**
 713  * If my_addr doesn't have an IP set, this function
 714  * will fill it in from the passed addr. Otherwise it does nothing and returns.
 715  */
 716 bool AsyncMessenger::set_addr_unknowns(const entity_addrvec_t &addrs)
 717 {
 718   ldout(cct,1) << __func__ << " " << addrs << dendl;
 719   bool ret = false;
 720   std::lock_guard l{lock};
 721
 722   entity_addrvec_t newaddrs = *my_addrs;
 723   for (auto& a : newaddrs.v) {
 724     if (a.is_blank_ip()) {
 725       int type = a.get_type();
 726       int port = a.get_port();
 727       uint32_t nonce = a.get_nonce();
 728       for (auto& b : addrs.v) {
 729         if (a.get_family() == b.get_family()) {
 730           ldout(cct,1) << __func__ << " assuming my addr " << a
 731                        << " matches provided addr " << b << dendl;
 732           a = b;
 733           a.set_nonce(nonce);
 734           a.set_type(type);
 735           a.set_port(port);
 736           ret = true;
 737           break;
 738         }
 739       }
 740     }
 741   }
 742   set_myaddrs(newaddrs);
 743   if (ret) {
 744     _init_local_connection();
 745   }
 746   ldout(cct,1) << __func__ << " now " << *my_addrs << dendl;
 747   return ret;
 748 }
 749
 750 void AsyncMessenger::set_addrs(const entity_addrvec_t &addrs)
 751 {
 752   std::lock_guard l{lock};
 753   auto t = addrs;
 754   for (auto& a : t.v) {
 755     a.set_nonce(nonce);
 756   }
 757   set_myaddrs(t);
 758   _init_local_connection();
 759 }
 760
 761 void AsyncMessenger::shutdown_connections(bool queue_reset)
 762 {
 763   ldout(cct,1) << __func__ << " " << dendl;
 764   std::lock_guard l{lock};
 765   for (const auto& c : accepting_conns) {
 766     ldout(cct, 5) << __func__ << " accepting_conn " << c << dendl;
 767     c->stop(queue_reset);
 768   }
 769   accepting_conns.clear();
 770
 771   for (const auto& [e, c] : conns) {
 772     ldout(cct, 5) << __func__ << " mark down " << e << " " << c << dendl;
 773     c->get_perf_counter()->dec(l_msgr_active_connections);
 774     c->stop(queue_reset);
 775   }
 776   conns.clear();
 777
 778   for (const auto& c : anon_conns) {
 779     ldout(cct, 5) << __func__ << " mark down " << c << dendl;
 780     c->get_perf_counter()->dec(l_msgr_active_connections);
 781     c->stop(queue_reset);
 782   }
 783   anon_conns.clear();
 784
 785   {
 786     std::lock_guard l{deleted_lock};
 787     if (cct->_conf->subsys.should_gather<ceph_subsys_ms, 5>()) {
 788       for (const auto& c : deleted_conns) {
 789         ldout(cct, 5) << __func__ << " delete " << c << dendl;
 790       }
 791     }
 792     deleted_conns.clear();
 793   }
 794 }
 795
 796 void AsyncMessenger::mark_down_addrs(const entity_addrvec_t& addrs)
 797 {
 798   std::lock_guard l{lock};
 799   const AsyncConnectionRef& conn = _lookup_conn(addrs);
 800   if (conn) {
 801     ldout(cct, 1) << __func__ << " " << addrs << " -- " << conn << dendl;
 802     conn->stop(true);
 803   } else {
 804     ldout(cct, 1) << __func__ << " " << addrs << " -- connection dne" << dendl;
 805   }
 806 }
 807
 808 int AsyncMessenger::get_proto_version(int peer_type, bool connect) const
 809 {
 810   int my_type = my_name.type();
 811
 812   // set reply protocol version
 813   if (peer_type == my_type) {
 814     // internal
 815     return cluster_protocol;
 816   } else {
 817     // public
 818     switch (connect ? peer_type : my_type) {
 819       case CEPH_ENTITY_TYPE_OSD: return CEPH_OSDC_PROTOCOL;
 820       case CEPH_ENTITY_TYPE_MDS: return CEPH_MDSC_PROTOCOL;
 821       case CEPH_ENTITY_TYPE_MON: return CEPH_MONC_PROTOCOL;
 822     }
 823   }
 824   return 0;
 825 }
 826
 827 int AsyncMessenger::accept_conn(const AsyncConnectionRef& conn)
 828 {
 829   std::lock_guard l{lock};
 830   if (conn->policy.server &&
 831       conn->policy.lossy &&
 832       !conn->policy.register_lossy_clients) {
 833     anon_conns.insert(conn);
 834     conn->get_perf_counter()->inc(l_msgr_active_connections);
 835     return 0;
 836   }
 837   auto it = conns.find(*conn->peer_addrs);
 838   if (it != conns.end()) {
 839     auto& existing = it->second;
 840
 841     // lazy delete, see "deleted_conns"
 842     // If conn already in, we will return 0
 843     std::lock_guard l{deleted_lock};
 844     if (deleted_conns.erase(existing)) {
 845       conns.erase(it);
 846     } else if (conn != existing) {
 847       return -1;
 848     }
 849   }
 850   ldout(cct, 10) << __func__ << " " << conn << " " << *conn->peer_addrs << dendl;
 851   conns[*conn->peer_addrs] = conn;
 852   conn->get_perf_counter()->inc(l_msgr_active_connections);
 853   accepting_conns.erase(conn);
 854   return 0;
 855 }
 856
 857
 858 bool AsyncMessenger::learned_addr(const entity_addr_t &peer_addr_for_me)
 859 {
 860   // be careful here: multiple threads may block here, and readers of
 861   // my_addr do NOT hold any lock.
 862
 863   // this always goes from true -> false under the protection of the
 864   // mutex.  if it is already false, we need not retake the mutex at
 865   // all.
 866   if (!need_addr)
 867     return false;
 868   std::lock_guard l(lock);
 869   if (need_addr) {
 870     if (my_addrs->empty()) {
 871       auto a = peer_addr_for_me;
 872       a.set_type(entity_addr_t::TYPE_ANY);
 873       a.set_nonce(nonce);
 874       if (!did_bind) {
 875         a.set_port(0);
 876       }
 877       set_myaddrs(entity_addrvec_t(a));
 878       ldout(cct,10) << __func__ << " had no addrs" << dendl;
 879     } else {
 880       // fix all addrs of the same family, regardless of type (msgr2 vs legacy)
 881       entity_addrvec_t newaddrs = *my_addrs;
 882       for (auto& a : newaddrs.v) {
 883         if (a.is_blank_ip() &&
 884             a.get_family() == peer_addr_for_me.get_family()) {
 885           entity_addr_t t = peer_addr_for_me;
 886           if (!did_bind) {
 887             t.set_type(entity_addr_t::TYPE_ANY);
 888             t.set_port(0);
 889           } else {
 890             t.set_type(a.get_type());
 891             t.set_port(a.get_port());
 892           }
 893           t.set_nonce(a.get_nonce());
 894           ldout(cct,10) << __func__ << " " << a << " -> " << t << dendl;
 895           a = t;
 896         }
 897       }
 898       set_myaddrs(newaddrs);
 899     }
 900     ldout(cct, 1) << __func__ << " learned my addr " << *my_addrs
 901                   << " (peer_addr_for_me " << peer_addr_for_me << ")" << dendl;
 902     _init_local_connection();
 903     need_addr = false;
 904     return true;
 905   }
 906   return false;
 907 }
 908
 909 void AsyncMessenger::reap_dead()
 910 {
 911   ldout(cct, 1) << __func__ << " start" << dendl;
 912
 913   std::lock_guard l1{lock};
 914
 915   {
 916     std::lock_guard l2{deleted_lock};
 917     for (auto& c : deleted_conns) {
 918       ldout(cct, 5) << __func__ << " delete " << c << dendl;
 919       auto conns_it = conns.find(*c->peer_addrs);
 920       if (conns_it != conns.end() && conns_it->second == c)
 921         conns.erase(conns_it);
 922       accepting_conns.erase(c);
 923       anon_conns.erase(c);
 924     }
 925     deleted_conns.clear();
 926   }
 927 }