ceph/src/rgw/rgw_gc.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab ft=cpp
   3
   4 #include "rgw_gc.h"
   5
   6 #include "rgw_tools.h"
   7 #include "include/scope_guard.h"
   8 #include "include/rados/librados.hpp"
   9 #include "cls/rgw/cls_rgw_client.h"
  10 #include "cls/rgw_gc/cls_rgw_gc_client.h"
  11 #include "cls/refcount/cls_refcount_client.h"
  12 #include "cls/version/cls_version_client.h"
  13 #include "rgw_perf_counters.h"
  14 #include "cls/lock/cls_lock_client.h"
  15 #include "include/random.h"
  16 #include "rgw_gc_log.h"
  17
  18 #include <list> // XXX
  19 #include <sstream>
  20
  21 #define dout_context g_ceph_context
  22 #define dout_subsys ceph_subsys_rgw
  23
  24 using namespace librados;
  25
  26 static string gc_oid_prefix = "gc";
  27 static string gc_index_lock_name = "gc_process";
  28
  29 void RGWGC::initialize(CephContext *_cct, RGWRados *_store) {
  30   cct = _cct;
  31   store = _store;
  32
  33   max_objs = min(static_cast<int>(cct->_conf->rgw_gc_max_objs), rgw_shards_max());
  34
  35   obj_names = new string[max_objs];
  36
  37   for (int i = 0; i < max_objs; i++) {
  38     obj_names[i] = gc_oid_prefix;
  39     char buf[32];
  40     snprintf(buf, 32, ".%d", i);
  41     obj_names[i].append(buf);
  42
  43     auto it = transitioned_objects_cache.begin() + i;
  44     transitioned_objects_cache.insert(it, false);
  45
  46     //version = 0 -> not ready for transition
  47     //version = 1 -> marked ready for transition
  48     librados::ObjectWriteOperation op;
  49     op.create(false);
  50     const uint64_t queue_size = cct->_conf->rgw_gc_max_queue_size, num_deferred_entries = cct->_conf->rgw_gc_max_deferred;
  51     gc_log_init2(op, queue_size, num_deferred_entries);
  52     store->gc_operate(obj_names[i], &op);
  53   }
  54 }
  55
  56 void RGWGC::finalize()
  57 {
  58   delete[] obj_names;
  59 }
  60
  61 int RGWGC::tag_index(const string& tag)
  62 {
  63   return rgw_shard_id(tag, max_objs);
  64 }
  65
  66 int RGWGC::send_chain(cls_rgw_obj_chain& chain, const string& tag)
  67 {
  68   ObjectWriteOperation op;
  69   cls_rgw_gc_obj_info info;
  70   info.chain = chain;
  71   info.tag = tag;
  72   gc_log_enqueue2(op, cct->_conf->rgw_gc_obj_min_wait, info);
  73
  74   int i = tag_index(tag);
  75
  76   ldpp_dout(this, 20) << "RGWGC::send_chain - on object name: " << obj_names[i] << "tag is: " << tag << dendl;
  77
  78   auto ret = store->gc_operate(obj_names[i], &op);
  79   if (ret != -ECANCELED && ret != -EPERM) {
  80     return ret;
  81   }
  82   ObjectWriteOperation set_entry_op;
  83   cls_rgw_gc_set_entry(set_entry_op, cct->_conf->rgw_gc_obj_min_wait, info);
  84   return store->gc_operate(obj_names[i], &set_entry_op);
  85 }
  86
  87 struct defer_chain_state {
  88   librados::AioCompletion* completion = nullptr;
  89   // TODO: hold a reference on the state in RGWGC to avoid use-after-free if
  90   // RGWGC destructs before this completion fires
  91   RGWGC* gc = nullptr;
  92   cls_rgw_gc_obj_info info;
  93
  94   ~defer_chain_state() {
  95     if (completion) {
  96       completion->release();
  97     }
  98   }
  99 };
 100
 101 static void async_defer_callback(librados::completion_t, void* arg)
 102 {
 103   std::unique_ptr<defer_chain_state> state{static_cast<defer_chain_state*>(arg)};
 104   if (state->completion->get_return_value() == -ECANCELED) {
 105     state->gc->on_defer_canceled(state->info);
 106   }
 107 }
 108
 109 void RGWGC::on_defer_canceled(const cls_rgw_gc_obj_info& info)
 110 {
 111   const std::string& tag = info.tag;
 112   const int i = tag_index(tag);
 113
 114   // ECANCELED from cls_version_check() tells us that we've transitioned
 115   transitioned_objects_cache[i] = true;
 116
 117   ObjectWriteOperation op;
 118   cls_rgw_gc_queue_defer_entry(op, cct->_conf->rgw_gc_obj_min_wait, info);
 119   cls_rgw_gc_remove(op, {tag});
 120
 121   auto c = librados::Rados::aio_create_completion(nullptr, nullptr);
 122   store->gc_aio_operate(obj_names[i], c, &op);
 123   c->release();
 124 }
 125
 126 int RGWGC::async_defer_chain(const string& tag, const cls_rgw_obj_chain& chain)
 127 {
 128   const int i = tag_index(tag);
 129   cls_rgw_gc_obj_info info;
 130   info.chain = chain;
 131   info.tag = tag;
 132
 133   // if we've transitioned this shard object, we can rely on the cls_rgw_gc queue
 134   if (transitioned_objects_cache[i]) {
 135     ObjectWriteOperation op;
 136     cls_rgw_gc_queue_defer_entry(op, cct->_conf->rgw_gc_obj_min_wait, info);
 137
 138     // this tag may still be present in omap, so remove it once the cls_rgw_gc
 139     // enqueue succeeds
 140     cls_rgw_gc_remove(op, {tag});
 141
 142     auto c = librados::Rados::aio_create_completion(nullptr, nullptr);
 143     int ret = store->gc_aio_operate(obj_names[i], c, &op);
 144     c->release();
 145     return ret;
 146   }
 147
 148   // if we haven't seen the transition yet, write the defer to omap with cls_rgw
 149   ObjectWriteOperation op;
 150
 151   // assert that we haven't initialized cls_rgw_gc queue. this prevents us
 152   // from writing new entries to omap after the transition
 153   gc_log_defer1(op, cct->_conf->rgw_gc_obj_min_wait, info);
 154
 155   // prepare a callback to detect the transition via ECANCELED from cls_version_check()
 156   auto state = std::make_unique<defer_chain_state>();
 157   state->gc = this;
 158   state->info.chain = chain;
 159   state->info.tag = tag;
 160   state->completion = librados::Rados::aio_create_completion(
 161       state.get(), async_defer_callback);
 162
 163   int ret = store->gc_aio_operate(obj_names[i], state->completion, &op);
 164   if (ret == 0) {
 165     state.release(); // release ownership until async_defer_callback()
 166   }
 167   return ret;
 168 }
 169
 170 int RGWGC::remove(int index, const std::vector<string>& tags, AioCompletion **pc)
 171 {
 172   ObjectWriteOperation op;
 173   cls_rgw_gc_remove(op, tags);
 174
 175   auto c = librados::Rados::aio_create_completion(nullptr, nullptr);
 176   int ret = store->gc_aio_operate(obj_names[index], c, &op);
 177   if (ret < 0) {
 178     c->release();
 179   } else {
 180     *pc = c;
 181   }
 182   return ret;
 183 }
 184
 185 int RGWGC::remove(int index, int num_entries)
 186 {
 187   ObjectWriteOperation op;
 188   cls_rgw_gc_queue_remove_entries(op, num_entries);
 189
 190   return store->gc_operate(obj_names[index], &op);
 191 }
 192
 193 int RGWGC::list(int *index, string& marker, uint32_t max, bool expired_only, std::list<cls_rgw_gc_obj_info>& result, bool *truncated, bool& processing_queue)
 194 {
 195   result.clear();
 196   string next_marker;
 197   bool check_queue = false;
 198
 199   for (; *index < max_objs && result.size() < max; (*index)++, marker.clear(), check_queue = false) {
 200     std::list<cls_rgw_gc_obj_info> entries, queue_entries;
 201     int ret = 0;
 202
 203     //processing_queue is set to true from previous iteration if the queue was under process and probably has more elements in it.
 204     if (! transitioned_objects_cache[*index] && ! check_queue && ! processing_queue) {
 205       ret = cls_rgw_gc_list(store->gc_pool_ctx, obj_names[*index], marker, max - result.size(), expired_only, entries, truncated, next_marker);
 206       if (ret != -ENOENT && ret < 0) {
 207         return ret;
 208       }
 209       obj_version objv;
 210       cls_version_read(store->gc_pool_ctx, obj_names[*index], &objv);
 211       if (ret == -ENOENT) {
 212         if (objv.ver == 0) {
 213           continue;
 214         } else {
 215           if (! expired_only) {
 216             transitioned_objects_cache[*index] = true;
 217             marker.clear();
 218           } else {
 219             std::list<cls_rgw_gc_obj_info> non_expired_entries;
 220             ret = cls_rgw_gc_list(store->gc_pool_ctx, obj_names[*index], marker, 1, false, non_expired_entries, truncated, next_marker);
 221             if (non_expired_entries.size() == 0) {
 222               transitioned_objects_cache[*index] = true;
 223               marker.clear();
 224             }
 225           }
 226         }
 227       }
 228       if ((objv.ver == 1) && (entries.size() < max - result.size())) {
 229         check_queue = true;
 230         marker.clear();
 231       }
 232     }
 233     if (transitioned_objects_cache[*index] || check_queue || processing_queue) {
 234       processing_queue = false;
 235       ret = cls_rgw_gc_queue_list_entries(store->gc_pool_ctx, obj_names[*index], marker, (max - result.size()) - entries.size(), expired_only, queue_entries, truncated, next_marker);
 236       if (ret < 0) {
 237         return ret;
 238       }
 239     }
 240     if (entries.size() == 0 && queue_entries.size() == 0)
 241       continue;
 242
 243     std::list<cls_rgw_gc_obj_info>::iterator iter;
 244     for (iter = entries.begin(); iter != entries.end(); ++iter) {
 245       result.push_back(*iter);
 246     }
 247
 248     for (iter = queue_entries.begin(); iter != queue_entries.end(); ++iter) {
 249       result.push_back(*iter);
 250     }
 251
 252     marker = next_marker;
 253
 254     if (*index == max_objs - 1) {
 255       /* we cut short here, truncated will hold the correct value */
 256       return 0;
 257     }
 258
 259     if (result.size() == max) {
 260       if (queue_entries.size() > 0 && *truncated) {
 261         processing_queue = true;
 262       } else {
 263         processing_queue = false;
 264         *index += 1; //move to next gc object
 265       }
 266
 267       /* close approximation, it might be that the next of the objects don't hold
 268        * anything, in this case truncated should have been false, but we can find
 269        * that out on the next iteration
 270        */
 271       *truncated = true;
 272       return 0;
 273     }
 274   }
 275   *truncated = false;
 276   processing_queue = false;
 277
 278   return 0;
 279 }
 280
 281 class RGWGCIOManager {
 282   const DoutPrefixProvider* dpp;
 283   CephContext *cct;
 284   RGWGC *gc;
 285
 286   struct IO {
 287     enum Type {
 288       UnknownIO = 0,
 289       TailIO = 1,
 290       IndexIO = 2,
 291     } type{UnknownIO};
 292     librados::AioCompletion *c{nullptr};
 293     string oid;
 294     int index{-1};
 295     string tag;
 296   };
 297
 298   deque<IO> ios;
 299   vector<std::vector<string> > remove_tags;
 300   /* tracks the number of remaining shadow objects for a given tag in order to
 301    * only remove the tag once all shadow objects have themselves been removed
 302    */
 303   vector<map<string, size_t> > tag_io_size;
 304
 305 #define MAX_AIO_DEFAULT 10
 306   size_t max_aio{MAX_AIO_DEFAULT};
 307
 308 public:
 309   RGWGCIOManager(const DoutPrefixProvider* _dpp, CephContext *_cct, RGWGC *_gc) : dpp(_dpp),
 310                                                   cct(_cct),
 311                                                   gc(_gc),
 312                                                   remove_tags(cct->_conf->rgw_gc_max_objs),
 313                                                   tag_io_size(cct->_conf->rgw_gc_max_objs) {
 314     max_aio = cct->_conf->rgw_gc_max_concurrent_io;
 315   }
 316
 317   ~RGWGCIOManager() {
 318     for (auto io : ios) {
 319       io.c->release();
 320     }
 321   }
 322
 323   int schedule_io(IoCtx *ioctx, const string& oid, ObjectWriteOperation *op,
 324                   int index, const string& tag) {
 325     while (ios.size() > max_aio) {
 326       if (gc->going_down()) {
 327         return 0;
 328       }
 329       auto ret = handle_next_completion();
 330       //Return error if we are using queue, else ignore it
 331       if (gc->transitioned_objects_cache[index] && ret < 0) {
 332         return ret;
 333       }
 334     }
 335
 336     auto c = librados::Rados::aio_create_completion(nullptr, nullptr);
 337     int ret = ioctx->aio_operate(oid, c, op);
 338     if (ret < 0) {
 339       return ret;
 340     }
 341     ios.push_back(IO{IO::TailIO, c, oid, index, tag});
 342
 343     return 0;
 344   }
 345
 346   int handle_next_completion() {
 347     ceph_assert(!ios.empty());
 348     IO& io = ios.front();
 349     io.c->wait_for_complete();
 350     int ret = io.c->get_return_value();
 351     io.c->release();
 352
 353     if (ret == -ENOENT) {
 354       ret = 0;
 355     }
 356
 357     if (io.type == IO::IndexIO && ! gc->transitioned_objects_cache[io.index]) {
 358       if (ret < 0) {
 359         ldpp_dout(dpp, 0) << "WARNING: gc cleanup of tags on gc shard index=" <<
 360           io.index << " returned error, ret=" << ret << dendl;
 361       }
 362       goto done;
 363     }
 364
 365     if (ret < 0) {
 366       ldpp_dout(dpp, 0) << "WARNING: gc could not remove oid=" << io.oid <<
 367         ", ret=" << ret << dendl;
 368       goto done;
 369     }
 370
 371     if (! gc->transitioned_objects_cache[io.index]) {
 372       schedule_tag_removal(io.index, io.tag);
 373     }
 374
 375   done:
 376     ios.pop_front();
 377     return ret;
 378   }
 379
 380   /* This is a request to schedule a tag removal. It will be called once when
 381    * there are no shadow objects. But it will also be called for every shadow
 382    * object when there are any. Since we do not want the tag to be removed
 383    * until all shadow objects have been successfully removed, the scheduling
 384    * will not happen until the shadow object count goes down to zero
 385    */
 386   void schedule_tag_removal(int index, string tag) {
 387     auto& ts = tag_io_size[index];
 388     auto ts_it = ts.find(tag);
 389     if (ts_it != ts.end()) {
 390       auto& size = ts_it->second;
 391       --size;
 392       // wait all shadow obj delete return
 393       if (size != 0)
 394         return;
 395
 396       ts.erase(ts_it);
 397     }
 398
 399     auto& rt = remove_tags[index];
 400
 401     rt.push_back(tag);
 402     if (rt.size() >= (size_t)cct->_conf->rgw_gc_max_trim_chunk) {
 403       flush_remove_tags(index, rt);
 404     }
 405   }
 406
 407   void add_tag_io_size(int index, string tag, size_t size) {
 408     auto& ts = tag_io_size[index];
 409     ts.emplace(tag, size);
 410   }
 411
 412   int drain_ios() {
 413     int ret_val = 0;
 414     while (!ios.empty()) {
 415       if (gc->going_down()) {
 416         return -EAGAIN;
 417       }
 418       auto ret = handle_next_completion();
 419       if (ret < 0) {
 420         ret_val = ret;
 421       }
 422     }
 423     return ret_val;
 424   }
 425
 426   void drain() {
 427     drain_ios();
 428     flush_remove_tags();
 429     /* the tags draining might have generated more ios, drain those too */
 430     drain_ios();
 431   }
 432
 433   void flush_remove_tags(int index, vector<string>& rt) {
 434     IO index_io;
 435     index_io.type = IO::IndexIO;
 436     index_io.index = index;
 437
 438     ldpp_dout(dpp, 20) << __func__ <<
 439       " removing entries from gc log shard index=" << index << ", size=" <<
 440       rt.size() << ", entries=" << rt << dendl;
 441
 442     auto rt_guard = make_scope_guard(
 443       [&]
 444         {
 445           rt.clear();
 446         }
 447       );
 448
 449     int ret = gc->remove(index, rt, &index_io.c);
 450     if (ret < 0) {
 451       /* we already cleared list of tags, this prevents us from
 452        * ballooning in case of a persistent problem
 453        */
 454       ldpp_dout(dpp, 0) << "WARNING: failed to remove tags on gc shard index=" <<
 455         index << " ret=" << ret << dendl;
 456       return;
 457     }
 458     if (perfcounter) {
 459       /* log the count of tags retired for rate estimation */
 460       perfcounter->inc(l_rgw_gc_retire, rt.size());
 461     }
 462     ios.push_back(index_io);
 463   }
 464
 465   void flush_remove_tags() {
 466     int index = 0;
 467     for (auto& rt : remove_tags) {
 468       if (! gc->transitioned_objects_cache[index]) {
 469         flush_remove_tags(index, rt);
 470       }
 471       ++index;
 472     }
 473   }
 474
 475   int remove_queue_entries(int index, int num_entries) {
 476     int ret = gc->remove(index, num_entries);
 477     if (ret < 0) {
 478       ldpp_dout(dpp, 0) << "ERROR: failed to remove queue entries on index=" <<
 479             index << " ret=" << ret << dendl;
 480       return ret;
 481     }
 482     return 0;
 483   }
 484 }; // class RGWGCIOManger
 485
 486 int RGWGC::process(int index, int max_secs, bool expired_only,
 487                    RGWGCIOManager& io_manager)
 488 {
 489   ldpp_dout(this, 20) << "RGWGC::process entered with GC index_shard=" <<
 490     index << ", max_secs=" << max_secs << ", expired_only=" <<
 491     expired_only << dendl;
 492
 493   rados::cls::lock::Lock l(gc_index_lock_name);
 494   utime_t end = ceph_clock_now();
 495
 496   /* max_secs should be greater than zero. We don't want a zero max_secs
 497    * to be translated as no timeout, since we'd then need to break the
 498    * lock and that would require a manual intervention. In this case
 499    * we can just wait it out. */
 500   if (max_secs <= 0)
 501     return -EAGAIN;
 502
 503   end += max_secs;
 504   utime_t time(max_secs, 0);
 505   l.set_duration(time);
 506
 507   int ret = l.lock_exclusive(&store->gc_pool_ctx, obj_names[index]);
 508   if (ret == -EBUSY) { /* already locked by another gc processor */
 509     ldpp_dout(this, 10) << "RGWGC::process failed to acquire lock on " <<
 510       obj_names[index] << dendl;
 511     return 0;
 512   }
 513   if (ret < 0)
 514     return ret;
 515
 516   string marker;
 517   string next_marker;
 518   bool truncated;
 519   IoCtx *ctx = new IoCtx;
 520   do {
 521     int max = 100;
 522     std::list<cls_rgw_gc_obj_info> entries;
 523
 524     int ret = 0;
 525
 526     if (! transitioned_objects_cache[index]) {
 527       ret = cls_rgw_gc_list(store->gc_pool_ctx, obj_names[index], marker, max, expired_only, entries, &truncated, next_marker);
 528       ldpp_dout(this, 20) <<
 529       "RGWGC::process cls_rgw_gc_list returned with returned:" << ret <<
 530       ", entries.size=" << entries.size() << ", truncated=" << truncated <<
 531       ", next_marker='" << next_marker << "'" << dendl;
 532       obj_version objv;
 533       cls_version_read(store->gc_pool_ctx, obj_names[index], &objv);
 534       if ((objv.ver == 1) && entries.size() == 0) {
 535         std::list<cls_rgw_gc_obj_info> non_expired_entries;
 536         ret = cls_rgw_gc_list(store->gc_pool_ctx, obj_names[index], marker, 1, false, non_expired_entries, &truncated, next_marker);
 537         if (non_expired_entries.size() == 0) {
 538           transitioned_objects_cache[index] = true;
 539           marker.clear();
 540           ldpp_dout(this, 20) << "RGWGC::process cls_rgw_gc_list returned ENOENT for non expired entries, so setting cache entry to TRUE" << dendl;
 541         } else {
 542           ret = 0;
 543           goto done;
 544         }
 545       }
 546       if ((objv.ver == 0) && (ret == -ENOENT)) {
 547         ret = 0;
 548         goto done;
 549       }
 550     }
 551
 552     if (transitioned_objects_cache[index]) {
 553       ret = cls_rgw_gc_queue_list_entries(store->gc_pool_ctx, obj_names[index], marker, max, expired_only, entries, &truncated, next_marker);
 554       ldpp_dout(this, 20) <<
 555       "RGWGC::process cls_rgw_gc_queue_list_entries returned with return value:" << ret <<
 556       ", entries.size=" << entries.size() << ", truncated=" << truncated <<
 557       ", next_marker='" << next_marker << "'" << dendl;
 558       if (entries.size() == 0) {
 559         ret = 0;
 560         goto done;
 561       }
 562     }
 563
 564     if (ret < 0)
 565       goto done;
 566
 567     marker = next_marker;
 568
 569     string last_pool;
 570     std::list<cls_rgw_gc_obj_info>::iterator iter;
 571     for (iter = entries.begin(); iter != entries.end(); ++iter) {
 572       cls_rgw_gc_obj_info& info = *iter;
 573
 574       ldpp_dout(this, 20) << "RGWGC::process iterating over entry tag='" <<
 575         info.tag << "', time=" << info.time << ", chain.objs.size()=" <<
 576         info.chain.objs.size() << dendl;
 577
 578       std::list<cls_rgw_obj>::iterator liter;
 579       cls_rgw_obj_chain& chain = info.chain;
 580
 581       utime_t now = ceph_clock_now();
 582       if (now >= end) {
 583         goto done;
 584       }
 585       if (! transitioned_objects_cache[index]) {
 586         if (chain.objs.empty()) {
 587           io_manager.schedule_tag_removal(index, info.tag);
 588         } else {
 589           io_manager.add_tag_io_size(index, info.tag, chain.objs.size());
 590         }
 591       }
 592       if (! chain.objs.empty()) {
 593         for (liter = chain.objs.begin(); liter != chain.objs.end(); ++liter) {
 594           cls_rgw_obj& obj = *liter;
 595
 596           if (obj.pool != last_pool) {
 597             delete ctx;
 598             ctx = new IoCtx;
 599             ret = rgw_init_ioctx(store->get_rados_handle(), obj.pool, *ctx);
 600             if (ret < 0) {
 601         if (transitioned_objects_cache[index]) {
 602           goto done;
 603         }
 604               last_pool = "";
 605               ldpp_dout(this, 0) << "ERROR: failed to create ioctx pool=" <<
 606                 obj.pool << dendl;
 607               continue;
 608             }
 609             last_pool = obj.pool;
 610           }
 611
 612           ctx->locator_set_key(obj.loc);
 613
 614           const string& oid = obj.key.name; /* just stored raw oid there */
 615
 616           ldpp_dout(this, 5) << "RGWGC::process removing " << obj.pool <<
 617             ":" << obj.key.name << dendl;
 618           ObjectWriteOperation op;
 619           cls_refcount_put(op, info.tag, true);
 620
 621           ret = io_manager.schedule_io(ctx, oid, &op, index, info.tag);
 622           if (ret < 0) {
 623             ldpp_dout(this, 0) <<
 624               "WARNING: failed to schedule deletion for oid=" << oid << dendl;
 625       if (transitioned_objects_cache[index]) {
 626         //If deleting oid failed for any of them, we will not delete queue entries
 627         goto done;
 628       }
 629           }
 630           if (going_down()) {
 631             // leave early, even if tag isn't removed, it's ok since it
 632             // will be picked up next time around
 633             goto done;
 634           }
 635         } // chains loop
 636       } // else -- chains not empty
 637     } // entries loop
 638     if (transitioned_objects_cache[index] && entries.size() > 0) {
 639       ret = io_manager.drain_ios();
 640       if (ret < 0) {
 641         goto done;
 642       }
 643       //Remove the entries from the queue
 644       ldpp_dout(this, 5) << "RGWGC::process removing entries, marker: " << marker << dendl;
 645       ret = io_manager.remove_queue_entries(index, entries.size());
 646       if (ret < 0) {
 647         ldpp_dout(this, 0) <<
 648           "WARNING: failed to remove queue entries" << dendl;
 649         goto done;
 650       }
 651     }
 652   } while (truncated);
 653
 654 done:
 655   /* we don't drain here, because if we're going down we don't want to
 656    * hold the system if backend is unresponsive
 657    */
 658   l.unlock(&store->gc_pool_ctx, obj_names[index]);
 659   delete ctx;
 660
 661   return 0;
 662 }
 663
 664 int RGWGC::process(bool expired_only)
 665 {
 666   int max_secs = cct->_conf->rgw_gc_processor_max_time;
 667
 668   const int start = ceph::util::generate_random_number(0, max_objs - 1);
 669
 670   RGWGCIOManager io_manager(this, store->ctx(), this);
 671
 672   for (int i = 0; i < max_objs; i++) {
 673     int index = (i + start) % max_objs;
 674     int ret = process(index, max_secs, expired_only, io_manager);
 675     if (ret < 0)
 676       return ret;
 677   }
 678   if (!going_down()) {
 679     io_manager.drain();
 680   }
 681
 682   return 0;
 683 }
 684
 685 bool RGWGC::going_down()
 686 {
 687   return down_flag;
 688 }
 689
 690 void RGWGC::start_processor()
 691 {
 692   worker = new GCWorker(this, cct, this);
 693   worker->create("rgw_gc");
 694 }
 695
 696 void RGWGC::stop_processor()
 697 {
 698   down_flag = true;
 699   if (worker) {
 700     worker->stop();
 701     worker->join();
 702   }
 703   delete worker;
 704   worker = NULL;
 705 }
 706
 707 unsigned RGWGC::get_subsys() const
 708 {
 709   return dout_subsys;
 710 }
 711
 712 std::ostream& RGWGC::gen_prefix(std::ostream& out) const
 713 {
 714   return out << "garbage collection: ";
 715 }
 716
 717 void *RGWGC::GCWorker::entry() {
 718   do {
 719     utime_t start = ceph_clock_now();
 720     ldpp_dout(dpp, 2) << "garbage collection: start" << dendl;
 721     int r = gc->process(true);
 722     if (r < 0) {
 723       ldpp_dout(dpp, 0) << "ERROR: garbage collection process() returned error r=" << r << dendl;
 724     }
 725     ldpp_dout(dpp, 2) << "garbage collection: stop" << dendl;
 726
 727     if (gc->going_down())
 728       break;
 729
 730     utime_t end = ceph_clock_now();
 731     end -= start;
 732     int secs = cct->_conf->rgw_gc_processor_period;
 733
 734     if (secs <= end.sec())
 735       continue; // next round
 736
 737     secs -= end.sec();
 738
 739     std::unique_lock locker{lock};
 740     cond.wait_for(locker, std::chrono::seconds(secs));
 741   } while (!gc->going_down());
 742
 743   return NULL;
 744 }
 745
 746 void RGWGC::GCWorker::stop()
 747 {
 748   std::lock_guard l{lock};
 749   cond.notify_all();
 750 }