ceph/src/rbd_replay/Replayer.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 Adam Crume <adamcrume@gmail.com>
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15 #include "Replayer.hpp"
  16 #include "common/errno.h"
  17 #include "rbd_replay/ActionTypes.h"
  18 #include "rbd_replay/BufferReader.h"
  19 #include <boost/foreach.hpp>
  20 #include <boost/thread/thread.hpp>
  21 #include <boost/scope_exit.hpp>
  22 #include <fstream>
  23 #include "global/global_context.h"
  24 #include "rbd_replay_debug.hpp"
  25
  26 #define dout_context g_ceph_context
  27
  28 using namespace std;
  29 using namespace rbd_replay;
  30
  31 namespace {
  32
  33 bool is_versioned_replay(BufferReader &buffer_reader) {
  34   bufferlist::iterator *it;
  35   int r = buffer_reader.fetch(&it);
  36   if (r < 0) {
  37     return false;
  38   }
  39
  40   if (it->get_remaining() < action::BANNER.size()) {
  41     return false;
  42   }
  43
  44   std::string banner;
  45   it->copy(action::BANNER.size(), banner);
  46   bool versioned = (banner == action::BANNER);
  47   if (!versioned) {
  48     it->seek(0);
  49   }
  50   return versioned;
  51 }
  52
  53 } // anonymous namespace
  54
  55 Worker::Worker(Replayer &replayer)
  56   : m_replayer(replayer),
  57     m_buffer(100),
  58     m_done(false) {
  59 }
  60
  61 void Worker::start() {
  62   m_thread = boost::shared_ptr<boost::thread>(new boost::thread(boost::bind(&Worker::run, this)));
  63 }
  64
  65 // Should only be called by StopThreadAction
  66 void Worker::stop() {
  67   m_done = true;
  68 }
  69
  70 void Worker::join() {
  71   m_thread->join();
  72 }
  73
  74 void Worker::send(Action::ptr action) {
  75   assert(action);
  76   m_buffer.push_front(action);
  77 }
  78
  79 void Worker::add_pending(PendingIO::ptr io) {
  80   assert(io);
  81   boost::mutex::scoped_lock lock(m_pending_ios_mutex);
  82   assertf(m_pending_ios.count(io->id()) == 0, "id = %d", io->id());
  83   m_pending_ios[io->id()] = io;
  84 }
  85
  86 void Worker::run() {
  87   dout(THREAD_LEVEL) << "Worker thread started" << dendl;
  88   while (!m_done) {
  89     Action::ptr action;
  90     m_buffer.pop_back(&action);
  91     m_replayer.wait_for_actions(action->predecessors());
  92     action->perform(*this);
  93     m_replayer.set_action_complete(action->id());
  94   }
  95   {
  96     boost::mutex::scoped_lock lock(m_pending_ios_mutex);
  97     bool first_time = true;
  98     while (!m_pending_ios.empty()) {
  99       if (!first_time) {
 100         dout(THREAD_LEVEL) << "Worker thread trying to stop, still waiting for " << m_pending_ios.size() << " pending IOs to complete:" << dendl;
 101         pair<action_id_t, PendingIO::ptr> p;
 102         BOOST_FOREACH(p, m_pending_ios) {
 103           dout(THREAD_LEVEL) << "> " << p.first << dendl;
 104         }
 105       }
 106       m_pending_ios_empty.timed_wait(lock, boost::posix_time::seconds(1));
 107       first_time = false;
 108     }
 109   }
 110   dout(THREAD_LEVEL) << "Worker thread stopped" << dendl;
 111 }
 112
 113
 114 void Worker::remove_pending(PendingIO::ptr io) {
 115   assert(io);
 116   m_replayer.set_action_complete(io->id());
 117   boost::mutex::scoped_lock lock(m_pending_ios_mutex);
 118   size_t num_erased = m_pending_ios.erase(io->id());
 119   assertf(num_erased == 1, "id = %d", io->id());
 120   if (m_pending_ios.empty()) {
 121     m_pending_ios_empty.notify_all();
 122   }
 123 }
 124
 125
 126 librbd::Image* Worker::get_image(imagectx_id_t imagectx_id) {
 127   return m_replayer.get_image(imagectx_id);
 128 }
 129
 130
 131 void Worker::put_image(imagectx_id_t imagectx_id, librbd::Image* image) {
 132   assert(image);
 133   m_replayer.put_image(imagectx_id, image);
 134 }
 135
 136
 137 void Worker::erase_image(imagectx_id_t imagectx_id) {
 138   m_replayer.erase_image(imagectx_id);
 139 }
 140
 141
 142 librbd::RBD* Worker::rbd() {
 143   return m_replayer.get_rbd();
 144 }
 145
 146
 147 librados::IoCtx* Worker::ioctx() {
 148   return m_replayer.get_ioctx();
 149 }
 150
 151 void Worker::set_action_complete(action_id_t id) {
 152   m_replayer.set_action_complete(id);
 153 }
 154
 155 bool Worker::readonly() const {
 156   return m_replayer.readonly();
 157 }
 158
 159 rbd_loc Worker::map_image_name(string image_name, string snap_name) const {
 160   return m_replayer.image_name_map().map(rbd_loc("", image_name, snap_name));
 161 }
 162
 163
 164 Replayer::Replayer(int num_action_trackers)
 165   : m_rbd(NULL), m_ioctx(0),
 166     m_pool_name("rbd"), m_latency_multiplier(1.0),
 167     m_readonly(false), m_dump_perf_counters(false),
 168     m_num_action_trackers(num_action_trackers),
 169     m_action_trackers(new action_tracker_d[m_num_action_trackers]) {
 170   assertf(num_action_trackers > 0, "num_action_trackers = %d", num_action_trackers);
 171 }
 172
 173 Replayer::~Replayer() {
 174   delete[] m_action_trackers;
 175 }
 176
 177 Replayer::action_tracker_d &Replayer::tracker_for(action_id_t id) {
 178   return m_action_trackers[id % m_num_action_trackers];
 179 }
 180
 181 void Replayer::run(const std::string& replay_file) {
 182   {
 183     librados::Rados rados;
 184     rados.init(NULL);
 185     int r = rados.init_with_context(g_ceph_context);
 186     if (r) {
 187       cerr << "Failed to initialize RADOS: " << cpp_strerror(r) << std::endl;
 188       goto out;
 189     }
 190     r = rados.connect();
 191     if (r) {
 192       cerr << "Failed to connect to cluster: " << cpp_strerror(r) << std::endl;
 193       goto out;
 194     }
 195     m_ioctx = new librados::IoCtx();
 196     {
 197       r = rados.ioctx_create(m_pool_name.c_str(), *m_ioctx);
 198       if (r) {
 199         cerr << "Failed to open pool " << m_pool_name << ": "
 200              << cpp_strerror(r) << std::endl;
 201         goto out2;
 202       }
 203       m_rbd = new librbd::RBD();
 204       map<thread_id_t, Worker*> workers;
 205
 206       int fd = open(replay_file.c_str(), O_RDONLY);
 207       if (fd < 0) {
 208         std::cerr << "Failed to open " << replay_file << ": "
 209                   << cpp_strerror(errno) << std::endl;
 210         exit(1);
 211       }
 212       BOOST_SCOPE_EXIT( (fd) ) {
 213         close(fd);
 214       } BOOST_SCOPE_EXIT_END;
 215
 216       BufferReader buffer_reader(fd);
 217       bool versioned = is_versioned_replay(buffer_reader);
 218       while (true) {
 219         action::ActionEntry action_entry;
 220         try {
 221           bufferlist::iterator *it;
 222           int r = buffer_reader.fetch(&it);
 223           if (r < 0) {
 224             std::cerr << "Failed to read from trace file: " << cpp_strerror(r)
 225                       << std::endl;
 226             exit(-r);
 227           }
 228           if (it->get_remaining() == 0) {
 229             break;
 230           }
 231
 232           if (versioned) {
 233             action_entry.decode(*it);
 234           } else {
 235             action_entry.decode_unversioned(*it);
 236           }
 237         } catch (const buffer::error &err) {
 238           std::cerr << "Failed to decode trace action: " << err.what() << std::endl;
 239           exit(1);
 240         }
 241
 242         Action::ptr action = Action::construct(action_entry);
 243         if (!action) {
 244           // unknown / unsupported action
 245           continue;
 246         }
 247
 248         if (action->is_start_thread()) {
 249           Worker *worker = new Worker(*this);
 250           workers[action->thread_id()] = worker;
 251           worker->start();
 252         } else {
 253           workers[action->thread_id()]->send(action);
 254         }
 255       }
 256
 257       dout(THREAD_LEVEL) << "Waiting for workers to die" << dendl;
 258       pair<thread_id_t, Worker*> w;
 259       BOOST_FOREACH(w, workers) {
 260         w.second->join();
 261         delete w.second;
 262       }
 263       clear_images();
 264       delete m_rbd;
 265       m_rbd = NULL;
 266     }
 267   out2:
 268     delete m_ioctx;
 269     m_ioctx = NULL;
 270   }
 271  out:
 272   ;
 273 }
 274
 275
 276 librbd::Image* Replayer::get_image(imagectx_id_t imagectx_id) {
 277   boost::shared_lock<boost::shared_mutex> lock(m_images_mutex);
 278   return m_images[imagectx_id];
 279 }
 280
 281 void Replayer::put_image(imagectx_id_t imagectx_id, librbd::Image *image) {
 282   assert(image);
 283   boost::unique_lock<boost::shared_mutex> lock(m_images_mutex);
 284   assert(m_images.count(imagectx_id) == 0);
 285   m_images[imagectx_id] = image;
 286 }
 287
 288 void Replayer::erase_image(imagectx_id_t imagectx_id) {
 289   boost::unique_lock<boost::shared_mutex> lock(m_images_mutex);
 290   librbd::Image* image = m_images[imagectx_id];
 291   if (m_dump_perf_counters) {
 292     string command = "perf dump";
 293     cmdmap_t cmdmap;
 294     string format = "json-pretty";
 295     bufferlist out;
 296     g_ceph_context->do_command(command, cmdmap, format, &out);
 297     out.write_stream(cout);
 298     cout << std::endl;
 299     cout.flush();
 300   }
 301   delete image;
 302   m_images.erase(imagectx_id);
 303 }
 304
 305 void Replayer::set_action_complete(action_id_t id) {
 306   dout(DEPGRAPH_LEVEL) << "ActionTracker::set_complete(" << id << ")" << dendl;
 307   boost::system_time now(boost::get_system_time());
 308   action_tracker_d &tracker = tracker_for(id);
 309   boost::unique_lock<boost::shared_mutex> lock(tracker.mutex);
 310   assert(tracker.actions.count(id) == 0);
 311   tracker.actions[id] = now;
 312   tracker.condition.notify_all();
 313 }
 314
 315 bool Replayer::is_action_complete(action_id_t id) {
 316   action_tracker_d &tracker = tracker_for(id);
 317   boost::shared_lock<boost::shared_mutex> lock(tracker.mutex);
 318   return tracker.actions.count(id) > 0;
 319 }
 320
 321 void Replayer::wait_for_actions(const action::Dependencies &deps) {
 322   boost::posix_time::ptime release_time(boost::posix_time::neg_infin);
 323   BOOST_FOREACH(const action::Dependency &dep, deps) {
 324     dout(DEPGRAPH_LEVEL) << "Waiting for " << dep.id << dendl;
 325     boost::system_time start_time(boost::get_system_time());
 326     action_tracker_d &tracker = tracker_for(dep.id);
 327     boost::shared_lock<boost::shared_mutex> lock(tracker.mutex);
 328     bool first_time = true;
 329     while (tracker.actions.count(dep.id) == 0) {
 330       if (!first_time) {
 331         dout(DEPGRAPH_LEVEL) << "Still waiting for " << dep.id << dendl;
 332       }
 333       tracker.condition.timed_wait(lock, boost::posix_time::seconds(1));
 334       first_time = false;
 335     }
 336     boost::system_time action_completed_time(tracker.actions[dep.id]);
 337     lock.unlock();
 338     boost::system_time end_time(boost::get_system_time());
 339     long long micros = (end_time - start_time).total_microseconds();
 340     dout(DEPGRAPH_LEVEL) << "Finished waiting for " << dep.id << " after " << micros << " microseconds" << dendl;
 341     // Apparently the nanoseconds constructor is optional:
 342     // http://www.boost.org/doc/libs/1_46_0/doc/html/date_time/details.html#compile_options
 343     boost::system_time sub_release_time(action_completed_time + boost::posix_time::microseconds(dep.time_delta * m_latency_multiplier / 1000));
 344     if (sub_release_time > release_time) {
 345       release_time = sub_release_time;
 346     }
 347   }
 348   if (release_time > boost::get_system_time()) {
 349     dout(SLEEP_LEVEL) << "Sleeping for " << (release_time - boost::get_system_time()).total_microseconds() << " microseconds" << dendl;
 350     boost::this_thread::sleep(release_time);
 351   }
 352 }
 353
 354 void Replayer::clear_images() {
 355   boost::unique_lock<boost::shared_mutex> lock(m_images_mutex);
 356   if (m_dump_perf_counters && !m_images.empty()) {
 357     string command = "perf dump";
 358     cmdmap_t cmdmap;
 359     string format = "json-pretty";
 360     bufferlist out;
 361     g_ceph_context->do_command(command, cmdmap, format, &out);
 362     out.write_stream(cout);
 363     cout << std::endl;
 364     cout.flush();
 365   }
 366   pair<imagectx_id_t, librbd::Image*> p;
 367   BOOST_FOREACH(p, m_images) {
 368     delete p.second;
 369   }
 370   m_images.clear();
 371 }
 372
 373 void Replayer::set_latency_multiplier(float f) {
 374   assertf(f >= 0, "f = %f", f);
 375   m_latency_multiplier = f;
 376 }
 377
 378 bool Replayer::readonly() const {
 379   return m_readonly;
 380 }
 381
 382 void Replayer::set_readonly(bool readonly) {
 383   m_readonly = readonly;
 384 }
 385
 386 string Replayer::pool_name() const {
 387   return m_pool_name;
 388 }
 389
 390 void Replayer::set_pool_name(string pool_name) {
 391   m_pool_name = pool_name;
 392 }