ceph/src/osd/PGBackend.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) 2013,2014 Inktank Storage, Inc.
   7  * Copyright (C) 2013,2014 Cloudwatt <libre.licensing@cloudwatt.com>
   8  *
   9  * Author: Loic Dachary <loic@dachary.org>
  10  *
  11  * This is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License version 2.1, as published by the Free Software
  14  * Foundation.  See file COPYING.
  15  *
  16  */
  17
  18
  19 #include "common/errno.h"
  20 #include "common/scrub_types.h"
  21 #include "ReplicatedBackend.h"
  22 #include "ScrubStore.h"
  23 #include "ECBackend.h"
  24 #include "PGBackend.h"
  25 #include "OSD.h"
  26 #include "erasure-code/ErasureCodePlugin.h"
  27 #include "OSDMap.h"
  28 #include "PGLog.h"
  29 #include "common/LogClient.h"
  30 #include "messages/MOSDPGRecoveryDelete.h"
  31 #include "messages/MOSDPGRecoveryDeleteReply.h"
  32
  33 #define dout_context cct
  34 #define dout_subsys ceph_subsys_osd
  35 #define DOUT_PREFIX_ARGS this
  36 #undef dout_prefix
  37 #define dout_prefix _prefix(_dout, this)
  38 static ostream& _prefix(std::ostream *_dout, PGBackend *pgb) {
  39   return *_dout << pgb->get_parent()->gen_dbg_prefix();
  40 }
  41
  42 void PGBackend::recover_delete_object(const hobject_t &oid, eversion_t v,
  43                                       RecoveryHandle *h)
  44 {
  45   assert(get_parent()->get_actingbackfill_shards().size() > 0);
  46   for (const auto& shard : get_parent()->get_actingbackfill_shards()) {
  47     if (shard == get_parent()->whoami_shard())
  48       continue;
  49     if (get_parent()->get_shard_missing(shard).is_missing(oid)) {
  50       dout(20) << __func__ << " will remove " << oid << " " << v << " from "
  51                << shard << dendl;
  52       h->deletes[shard].push_back(make_pair(oid, v));
  53       get_parent()->begin_peer_recover(shard, oid);
  54     }
  55   }
  56 }
  57
  58 void PGBackend::send_recovery_deletes(int prio,
  59                                       const map<pg_shard_t, vector<pair<hobject_t, eversion_t> > > &deletes)
  60 {
  61   epoch_t min_epoch = get_parent()->get_last_peering_reset_epoch();
  62   for (const auto& p : deletes) {
  63     const auto& shard = p.first;
  64     const auto& objects = p.second;
  65     ConnectionRef con = get_parent()->get_con_osd_cluster(
  66       shard.osd,
  67       get_osdmap()->get_epoch());
  68     if (!con)
  69       continue;
  70     auto it = objects.begin();
  71     while (it != objects.end()) {
  72       uint64_t cost = 0;
  73       uint64_t deletes = 0;
  74       spg_t target_pg = spg_t(get_parent()->get_info().pgid.pgid, shard.shard);
  75       MOSDPGRecoveryDelete *msg =
  76         new MOSDPGRecoveryDelete(get_parent()->whoami_shard(),
  77                                  target_pg,
  78                                  get_osdmap()->get_epoch(),
  79                                  min_epoch);
  80       msg->set_priority(prio);
  81
  82       while (it != objects.end() &&
  83              cost < cct->_conf->osd_max_push_cost &&
  84              deletes < cct->_conf->osd_max_push_objects) {
  85         dout(20) << __func__ << ": sending recovery delete << " << it->first
  86                  << " " << it->second << " to osd." << shard << dendl;
  87         msg->objects.push_back(*it);
  88         cost += cct->_conf->osd_push_per_object_cost;
  89         ++deletes;
  90         ++it;
  91       }
  92
  93       msg->set_cost(cost);
  94       get_parent()->send_message_osd_cluster(msg, con);
  95     }
  96   }
  97 }
  98
  99 bool PGBackend::handle_message(OpRequestRef op)
 100 {
 101   switch (op->get_req()->get_type()) {
 102   case MSG_OSD_PG_RECOVERY_DELETE:
 103     handle_recovery_delete(op);
 104     return true;
 105
 106   case MSG_OSD_PG_RECOVERY_DELETE_REPLY:
 107     handle_recovery_delete_reply(op);
 108     return true;
 109
 110   default:
 111     break;
 112   }
 113
 114   return _handle_message(op);
 115 }
 116
 117 void PGBackend::handle_recovery_delete(OpRequestRef op)
 118 {
 119   const MOSDPGRecoveryDelete *m = static_cast<const MOSDPGRecoveryDelete *>(op->get_req());
 120   assert(m->get_type() == MSG_OSD_PG_RECOVERY_DELETE);
 121   dout(20) << __func__ << " " << op << dendl;
 122
 123   op->mark_started();
 124
 125   C_GatherBuilder gather(cct);
 126   for (const auto &p : m->objects) {
 127     get_parent()->remove_missing_object(p.first, p.second, gather.new_sub());
 128   }
 129
 130   MOSDPGRecoveryDeleteReply *reply = new MOSDPGRecoveryDeleteReply;
 131   reply->from = get_parent()->whoami_shard();
 132   reply->set_priority(m->get_priority());
 133   reply->pgid = spg_t(get_parent()->get_info().pgid.pgid, m->from.shard);
 134   reply->map_epoch = m->map_epoch;
 135   reply->min_epoch = m->min_epoch;
 136   reply->objects = m->objects;
 137   ConnectionRef conn = m->get_connection();
 138
 139   gather.set_finisher(new FunctionContext(
 140     [=](int r) {
 141       if (r != -EAGAIN) {
 142         get_parent()->send_message_osd_cluster(reply, conn.get());
 143       } else {
 144         reply->put();
 145       }
 146     }));
 147   gather.activate();
 148 }
 149
 150 void PGBackend::handle_recovery_delete_reply(OpRequestRef op)
 151 {
 152   const MOSDPGRecoveryDeleteReply *m = static_cast<const MOSDPGRecoveryDeleteReply *>(op->get_req());
 153   assert(m->get_type() == MSG_OSD_PG_RECOVERY_DELETE_REPLY);
 154   dout(20) << __func__ << " " << op << dendl;
 155
 156   for (const auto &p : m->objects) {
 157     ObjectRecoveryInfo recovery_info;
 158     hobject_t oid = p.first;
 159     recovery_info.version = p.second;
 160     get_parent()->on_peer_recover(m->from, oid, recovery_info);
 161     bool peers_recovered = true;
 162     for (const auto& shard : get_parent()->get_actingbackfill_shards()) {
 163       if (shard == get_parent()->whoami_shard())
 164         continue;
 165       if (get_parent()->get_shard_missing(shard).is_missing(oid)) {
 166         dout(20) << __func__ << " " << oid << " still missing on at least "
 167                  << shard << dendl;
 168         peers_recovered = false;
 169         break;
 170       }
 171     }
 172     if (peers_recovered && !get_parent()->get_local_missing().is_missing(oid)) {
 173       dout(20) << __func__ << " completed recovery, local_missing = "
 174                << get_parent()->get_local_missing() << dendl;
 175       object_stat_sum_t stat_diff;
 176       stat_diff.num_objects_recovered = 1;
 177       get_parent()->on_global_recover(p.first, stat_diff, true);
 178     }
 179   }
 180 }
 181
 182 void PGBackend::rollback(
 183   const pg_log_entry_t &entry,
 184   ObjectStore::Transaction *t)
 185 {
 186
 187   struct RollbackVisitor : public ObjectModDesc::Visitor {
 188     const hobject_t &hoid;
 189     PGBackend *pg;
 190     ObjectStore::Transaction t;
 191     RollbackVisitor(
 192       const hobject_t &hoid,
 193       PGBackend *pg) : hoid(hoid), pg(pg) {}
 194     void append(uint64_t old_size) override {
 195       ObjectStore::Transaction temp;
 196       pg->rollback_append(hoid, old_size, &temp);
 197       temp.append(t);
 198       temp.swap(t);
 199     }
 200     void setattrs(map<string, boost::optional<bufferlist> > &attrs) override {
 201       ObjectStore::Transaction temp;
 202       pg->rollback_setattrs(hoid, attrs, &temp);
 203       temp.append(t);
 204       temp.swap(t);
 205     }
 206     void rmobject(version_t old_version) override {
 207       ObjectStore::Transaction temp;
 208       pg->rollback_stash(hoid, old_version, &temp);
 209       temp.append(t);
 210       temp.swap(t);
 211     }
 212     void try_rmobject(version_t old_version) override {
 213       ObjectStore::Transaction temp;
 214       pg->rollback_try_stash(hoid, old_version, &temp);
 215       temp.append(t);
 216       temp.swap(t);
 217     }
 218     void create() override {
 219       ObjectStore::Transaction temp;
 220       pg->rollback_create(hoid, &temp);
 221       temp.append(t);
 222       temp.swap(t);
 223     }
 224     void update_snaps(const set<snapid_t> &snaps) override {
 225       ObjectStore::Transaction temp;
 226       pg->get_parent()->pgb_set_object_snap_mapping(hoid, snaps, &temp);
 227       temp.append(t);
 228       temp.swap(t);
 229     }
 230     void rollback_extents(
 231       version_t gen,
 232       const vector<pair<uint64_t, uint64_t> > &extents) override {
 233       ObjectStore::Transaction temp;
 234       pg->rollback_extents(gen, extents, hoid, &temp);
 235       temp.append(t);
 236       temp.swap(t);
 237     }
 238   };
 239
 240   assert(entry.mod_desc.can_rollback());
 241   RollbackVisitor vis(entry.soid, this);
 242   entry.mod_desc.visit(&vis);
 243   t->append(vis.t);
 244 }
 245
 246 struct Trimmer : public ObjectModDesc::Visitor {
 247   const hobject_t &soid;
 248   PGBackend *pg;
 249   ObjectStore::Transaction *t;
 250   Trimmer(
 251     const hobject_t &soid,
 252     PGBackend *pg,
 253     ObjectStore::Transaction *t)
 254     : soid(soid), pg(pg), t(t) {}
 255   void rmobject(version_t old_version) override {
 256     pg->trim_rollback_object(
 257       soid,
 258       old_version,
 259       t);
 260   }
 261   // try_rmobject defaults to rmobject
 262   void rollback_extents(
 263     version_t gen,
 264     const vector<pair<uint64_t, uint64_t> > &extents) override {
 265     pg->trim_rollback_object(
 266       soid,
 267       gen,
 268       t);
 269   }
 270 };
 271
 272 void PGBackend::rollforward(
 273   const pg_log_entry_t &entry,
 274   ObjectStore::Transaction *t)
 275 {
 276   auto dpp = get_parent()->get_dpp();
 277   ldpp_dout(dpp, 20) << __func__ << ": entry=" << entry << dendl;
 278   if (!entry.can_rollback())
 279     return;
 280   Trimmer trimmer(entry.soid, this, t);
 281   entry.mod_desc.visit(&trimmer);
 282 }
 283
 284 void PGBackend::trim(
 285   const pg_log_entry_t &entry,
 286   ObjectStore::Transaction *t)
 287 {
 288   if (!entry.can_rollback())
 289     return;
 290   Trimmer trimmer(entry.soid, this, t);
 291   entry.mod_desc.visit(&trimmer);
 292 }
 293
 294 void PGBackend::try_stash(
 295   const hobject_t &hoid,
 296   version_t v,
 297   ObjectStore::Transaction *t)
 298 {
 299   t->try_rename(
 300     coll,
 301     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 302     ghobject_t(hoid, v, get_parent()->whoami_shard().shard));
 303 }
 304
 305 void PGBackend::remove(
 306   const hobject_t &hoid,
 307   ObjectStore::Transaction *t) {
 308   assert(!hoid.is_temp());
 309   t->remove(
 310     coll,
 311     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard));
 312   get_parent()->pgb_clear_object_snap_mapping(hoid, t);
 313 }
 314
 315 void PGBackend::on_change_cleanup(ObjectStore::Transaction *t)
 316 {
 317   dout(10) << __func__ << dendl;
 318   // clear temp
 319   for (set<hobject_t>::iterator i = temp_contents.begin();
 320        i != temp_contents.end();
 321        ++i) {
 322     dout(10) << __func__ << ": Removing oid "
 323              << *i << " from the temp collection" << dendl;
 324     t->remove(
 325       coll,
 326       ghobject_t(*i, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard));
 327   }
 328   temp_contents.clear();
 329 }
 330
 331 int PGBackend::objects_list_partial(
 332   const hobject_t &begin,
 333   int min,
 334   int max,
 335   vector<hobject_t> *ls,
 336   hobject_t *next)
 337 {
 338   assert(ls);
 339   // Starts with the smallest generation to make sure the result list
 340   // has the marker object (it might have multiple generations
 341   // though, which would be filtered).
 342   ghobject_t _next;
 343   if (!begin.is_min())
 344     _next = ghobject_t(begin, 0, get_parent()->whoami_shard().shard);
 345   ls->reserve(max);
 346   int r = 0;
 347
 348   if (min > max)
 349     min = max;
 350
 351   while (!_next.is_max() && ls->size() < (unsigned)min) {
 352     vector<ghobject_t> objects;
 353     r = store->collection_list(
 354       ch,
 355       _next,
 356       ghobject_t::get_max(),
 357       max - ls->size(),
 358       &objects,
 359       &_next);
 360     if (r != 0) {
 361       derr << __func__ << " list collection " << ch << " got: " << cpp_strerror(r) << dendl;
 362       break;
 363     }
 364     for (vector<ghobject_t>::iterator i = objects.begin();
 365          i != objects.end();
 366          ++i) {
 367       if (i->is_pgmeta() || i->hobj.is_temp()) {
 368         continue;
 369       }
 370       if (i->is_no_gen()) {
 371         ls->push_back(i->hobj);
 372       }
 373     }
 374   }
 375   if (r == 0)
 376     *next = _next.hobj;
 377   return r;
 378 }
 379
 380 int PGBackend::objects_list_range(
 381   const hobject_t &start,
 382   const hobject_t &end,
 383   snapid_t seq,
 384   vector<hobject_t> *ls,
 385   vector<ghobject_t> *gen_obs)
 386 {
 387   assert(ls);
 388   vector<ghobject_t> objects;
 389   int r = store->collection_list(
 390     ch,
 391     ghobject_t(start, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 392     ghobject_t(end, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 393     INT_MAX,
 394     &objects,
 395     NULL);
 396   ls->reserve(objects.size());
 397   for (vector<ghobject_t>::iterator i = objects.begin();
 398        i != objects.end();
 399        ++i) {
 400     if (i->is_pgmeta() || i->hobj.is_temp()) {
 401       continue;
 402     }
 403     if (i->is_no_gen()) {
 404       ls->push_back(i->hobj);
 405     } else if (gen_obs) {
 406       gen_obs->push_back(*i);
 407     }
 408   }
 409   return r;
 410 }
 411
 412 int PGBackend::objects_get_attr(
 413   const hobject_t &hoid,
 414   const string &attr,
 415   bufferlist *out)
 416 {
 417   bufferptr bp;
 418   int r = store->getattr(
 419     ch,
 420     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 421     attr.c_str(),
 422     bp);
 423   if (r >= 0 && out) {
 424     out->clear();
 425     out->push_back(std::move(bp));
 426   }
 427   return r;
 428 }
 429
 430 int PGBackend::objects_get_attrs(
 431   const hobject_t &hoid,
 432   map<string, bufferlist> *out)
 433 {
 434   return store->getattrs(
 435     ch,
 436     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 437     *out);
 438 }
 439
 440 void PGBackend::rollback_setattrs(
 441   const hobject_t &hoid,
 442   map<string, boost::optional<bufferlist> > &old_attrs,
 443   ObjectStore::Transaction *t) {
 444   map<string, bufferlist> to_set;
 445   assert(!hoid.is_temp());
 446   for (map<string, boost::optional<bufferlist> >::iterator i = old_attrs.begin();
 447        i != old_attrs.end();
 448        ++i) {
 449     if (i->second) {
 450       to_set[i->first] = i->second.get();
 451     } else {
 452       t->rmattr(
 453         coll,
 454         ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 455         i->first);
 456     }
 457   }
 458   t->setattrs(
 459     coll,
 460     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 461     to_set);
 462 }
 463
 464 void PGBackend::rollback_append(
 465   const hobject_t &hoid,
 466   uint64_t old_size,
 467   ObjectStore::Transaction *t) {
 468   assert(!hoid.is_temp());
 469   t->truncate(
 470     coll,
 471     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 472     old_size);
 473 }
 474
 475 void PGBackend::rollback_stash(
 476   const hobject_t &hoid,
 477   version_t old_version,
 478   ObjectStore::Transaction *t) {
 479   assert(!hoid.is_temp());
 480   t->remove(
 481     coll,
 482     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard));
 483   t->collection_move_rename(
 484     coll,
 485     ghobject_t(hoid, old_version, get_parent()->whoami_shard().shard),
 486     coll,
 487     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard));
 488 }
 489
 490 void PGBackend::rollback_try_stash(
 491   const hobject_t &hoid,
 492   version_t old_version,
 493   ObjectStore::Transaction *t) {
 494   assert(!hoid.is_temp());
 495   t->remove(
 496     coll,
 497     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard));
 498   t->try_rename(
 499     coll,
 500     ghobject_t(hoid, old_version, get_parent()->whoami_shard().shard),
 501     ghobject_t(hoid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard));
 502 }
 503
 504 void PGBackend::rollback_extents(
 505   version_t gen,
 506   const vector<pair<uint64_t, uint64_t> > &extents,
 507   const hobject_t &hoid,
 508   ObjectStore::Transaction *t) {
 509   auto shard = get_parent()->whoami_shard().shard;
 510   for (auto &&extent: extents) {
 511     t->clone_range(
 512       coll,
 513       ghobject_t(hoid, gen, shard),
 514       ghobject_t(hoid, ghobject_t::NO_GEN, shard),
 515       extent.first,
 516       extent.second,
 517       extent.first);
 518   }
 519   t->remove(
 520     coll,
 521     ghobject_t(hoid, gen, shard));
 522 }
 523
 524 void PGBackend::trim_rollback_object(
 525   const hobject_t &hoid,
 526   version_t old_version,
 527   ObjectStore::Transaction *t) {
 528   assert(!hoid.is_temp());
 529   t->remove(
 530     coll, ghobject_t(hoid, old_version, get_parent()->whoami_shard().shard));
 531 }
 532
 533 PGBackend *PGBackend::build_pg_backend(
 534   const pg_pool_t &pool,
 535   const OSDMapRef curmap,
 536   Listener *l,
 537   coll_t coll,
 538   ObjectStore::CollectionHandle &ch,
 539   ObjectStore *store,
 540   CephContext *cct)
 541 {
 542   switch (pool.type) {
 543   case pg_pool_t::TYPE_REPLICATED: {
 544     return new ReplicatedBackend(l, coll, ch, store, cct);
 545   }
 546   case pg_pool_t::TYPE_ERASURE: {
 547     ErasureCodeInterfaceRef ec_impl;
 548     ErasureCodeProfile profile = curmap->get_erasure_code_profile(pool.erasure_code_profile);
 549     assert(profile.count("plugin"));
 550     stringstream ss;
 551     ceph::ErasureCodePluginRegistry::instance().factory(
 552       profile.find("plugin")->second,
 553       cct->_conf->get_val<std::string>("erasure_code_dir"),
 554       profile,
 555       &ec_impl,
 556       &ss);
 557     assert(ec_impl);
 558     return new ECBackend(
 559       l,
 560       coll,
 561       ch,
 562       store,
 563       cct,
 564       ec_impl,
 565       pool.stripe_width);
 566   }
 567   default:
 568     ceph_abort();
 569     return NULL;
 570   }
 571 }
 572
 573 int PGBackend::be_scan_list(
 574   ScrubMap &map,
 575   ScrubMapBuilder &pos)
 576 {
 577   dout(10) << __func__ << " " << pos << dendl;
 578   assert(!pos.done());
 579   assert(pos.pos < pos.ls.size());
 580   hobject_t& poid = pos.ls[pos.pos];
 581
 582   struct stat st;
 583   int r = store->stat(
 584     ch,
 585     ghobject_t(
 586       poid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 587     &st,
 588     true);
 589   if (r == 0) {
 590     ScrubMap::object &o = map.objects[poid];
 591     o.size = st.st_size;
 592     assert(!o.negative);
 593     store->getattrs(
 594       ch,
 595       ghobject_t(
 596         poid, ghobject_t::NO_GEN, get_parent()->whoami_shard().shard),
 597       o.attrs);
 598
 599     if (pos.deep) {
 600       r = be_deep_scrub(poid, map, pos, o);
 601     }
 602     dout(25) << __func__ << "  " << poid << dendl;
 603   } else if (r == -ENOENT) {
 604     dout(25) << __func__ << "  " << poid << " got " << r
 605              << ", skipping" << dendl;
 606   } else if (r == -EIO) {
 607     dout(25) << __func__ << "  " << poid << " got " << r
 608              << ", stat_error" << dendl;
 609     ScrubMap::object &o = map.objects[poid];
 610     o.stat_error = true;
 611   } else {
 612     derr << __func__ << " got: " << cpp_strerror(r) << dendl;
 613     ceph_abort();
 614   }
 615   if (r == -EINPROGRESS) {
 616     return -EINPROGRESS;
 617   }
 618   pos.next_object();
 619   return 0;
 620 }
 621
 622 bool PGBackend::be_compare_scrub_objects(
 623   pg_shard_t auth_shard,
 624   const ScrubMap::object &auth,
 625   const object_info_t& auth_oi,
 626   const ScrubMap::object &candidate,
 627   shard_info_wrapper &shard_result,
 628   inconsistent_obj_wrapper &obj_result,
 629   ostream &errorstream)
 630 {
 631   enum { CLEAN, FOUND_ERROR } error = CLEAN;
 632   if (candidate.stat_error) {
 633     assert(shard_result.has_stat_error());
 634     error = FOUND_ERROR;
 635     errorstream << "candidate had a stat error";
 636   }
 637   if (candidate.read_error || candidate.ec_hash_mismatch || candidate.ec_size_mismatch) {
 638     error = FOUND_ERROR;
 639     errorstream << "candidate had a read error";
 640   }
 641   if (auth.digest_present && candidate.digest_present) {
 642     if (auth.digest != candidate.digest) {
 643       if (error != CLEAN)
 644         errorstream << ", ";
 645       error = FOUND_ERROR;
 646       errorstream << "data_digest 0x" << std::hex << candidate.digest
 647                   << " != data_digest 0x" << auth.digest << std::dec
 648                   << " from shard " << auth_shard;
 649       obj_result.set_data_digest_mismatch();
 650     }
 651   }
 652   if (auth.omap_digest_present && candidate.omap_digest_present) {
 653     if (auth.omap_digest != candidate.omap_digest) {
 654       if (error != CLEAN)
 655         errorstream << ", ";
 656       error = FOUND_ERROR;
 657       errorstream << "omap_digest 0x" << std::hex << candidate.omap_digest
 658                   << " != omap_digest 0x" << auth.omap_digest << std::dec
 659                   << " from shard " << auth_shard;
 660       obj_result.set_omap_digest_mismatch();
 661     }
 662   }
 663   if (parent->get_pool().is_replicated()) {
 664     if (auth_oi.is_data_digest() && candidate.digest_present) {
 665       if (auth_oi.data_digest != candidate.digest) {
 666         if (error != CLEAN)
 667           errorstream << ", ";
 668         error = FOUND_ERROR;
 669         errorstream << "data_digest 0x" << std::hex << candidate.digest
 670                     << " != data_digest 0x" << auth_oi.data_digest << std::dec
 671                     << " from auth oi " << auth_oi;
 672         shard_result.set_data_digest_mismatch_info();
 673       }
 674     }
 675     if (auth_oi.is_omap_digest() && candidate.omap_digest_present) {
 676       if (auth_oi.omap_digest != candidate.omap_digest) {
 677         if (error != CLEAN)
 678           errorstream << ", ";
 679         error = FOUND_ERROR;
 680         errorstream << "omap_digest 0x" << std::hex << candidate.omap_digest
 681                     << " != omap_digest 0x" << auth_oi.omap_digest << std::dec
 682                     << " from auth oi " << auth_oi;
 683         shard_result.set_omap_digest_mismatch_info();
 684       }
 685     }
 686   }
 687   if (candidate.stat_error)
 688     return error == FOUND_ERROR;
 689   uint64_t oi_size = be_get_ondisk_size(auth_oi.size);
 690   if (oi_size != candidate.size) {
 691     if (error != CLEAN)
 692       errorstream << ", ";
 693     error = FOUND_ERROR;
 694     errorstream << "size " << candidate.size
 695                 << " != size " << oi_size
 696                 << " from auth oi " << auth_oi;
 697     shard_result.set_size_mismatch_info();
 698   }
 699   if (auth.size != candidate.size) {
 700     if (error != CLEAN)
 701       errorstream << ", ";
 702     error = FOUND_ERROR;
 703     errorstream << "size " << candidate.size
 704                 << " != size " << auth.size
 705                 << " from shard " << auth_shard;
 706     obj_result.set_size_mismatch();
 707   }
 708   for (map<string,bufferptr>::const_iterator i = auth.attrs.begin();
 709        i != auth.attrs.end();
 710        ++i) {
 711     // We check system keys seperately
 712     if (i->first == OI_ATTR || i->first[0] != '_')
 713       continue;
 714     if (!candidate.attrs.count(i->first)) {
 715       if (error != CLEAN)
 716         errorstream << ", ";
 717       error = FOUND_ERROR;
 718       errorstream << "attr name mismatch '" << i->first << "'";
 719       obj_result.set_attr_name_mismatch();
 720     } else if (candidate.attrs.find(i->first)->second.cmp(i->second)) {
 721       if (error != CLEAN)
 722         errorstream << ", ";
 723       error = FOUND_ERROR;
 724       errorstream << "attr value mismatch '" << i->first << "'";
 725       obj_result.set_attr_value_mismatch();
 726     }
 727   }
 728   for (map<string,bufferptr>::const_iterator i = candidate.attrs.begin();
 729        i != candidate.attrs.end();
 730        ++i) {
 731     // We check system keys seperately
 732     if (i->first == OI_ATTR || i->first[0] != '_')
 733       continue;
 734     if (!auth.attrs.count(i->first)) {
 735       if (error != CLEAN)
 736         errorstream << ", ";
 737       error = FOUND_ERROR;
 738       errorstream << "attr name mismatch '" << i->first << "'";
 739       obj_result.set_attr_name_mismatch();
 740     }
 741   }
 742   return error == FOUND_ERROR;
 743 }
 744
 745 static int dcount(const object_info_t &oi)
 746 {
 747   int count = 0;
 748   if (oi.is_data_digest())
 749     count++;
 750   if (oi.is_omap_digest())
 751     count++;
 752   return count;
 753 }
 754
 755 map<pg_shard_t, ScrubMap *>::const_iterator
 756   PGBackend::be_select_auth_object(
 757   const hobject_t &obj,
 758   const map<pg_shard_t,ScrubMap*> &maps,
 759   object_info_t *auth_oi,
 760   map<pg_shard_t, shard_info_wrapper> &shard_map,
 761   inconsistent_obj_wrapper &object_error)
 762 {
 763   eversion_t auth_version;
 764   bufferlist first_oi_bl, first_ss_bl, first_hk_bl;
 765
 766   // Create list of shards with primary first so it will be auth copy all
 767   // other things being equal.
 768   list<pg_shard_t> shards;
 769   for (map<pg_shard_t, ScrubMap *>::const_iterator j = maps.begin();
 770        j != maps.end();
 771        ++j) {
 772     if (j->first == get_parent()->whoami_shard())
 773       continue;
 774     shards.push_back(j->first);
 775   }
 776   shards.push_front(get_parent()->whoami_shard());
 777
 778   map<pg_shard_t, ScrubMap *>::const_iterator auth = maps.end();
 779   for (auto &l : shards) {
 780     map<pg_shard_t, ScrubMap *>::const_iterator j = maps.find(l);
 781     map<hobject_t, ScrubMap::object>::iterator i =
 782       j->second->objects.find(obj);
 783     if (i == j->second->objects.end()) {
 784       continue;
 785     }
 786     string error_string;
 787     auto& shard_info = shard_map[j->first];
 788     if (j->first == get_parent()->whoami_shard())
 789       shard_info.primary = true;
 790     if (i->second.read_error) {
 791       shard_info.set_read_error();
 792       error_string += " read_error";
 793     }
 794     if (i->second.ec_hash_mismatch) {
 795       shard_info.set_ec_hash_mismatch();
 796       error_string += " ec_hash_mismatch";
 797     }
 798     if (i->second.ec_size_mismatch) {
 799       shard_info.set_ec_size_mismatch();
 800       error_string += " ec_size_mismatch";
 801     }
 802
 803     object_info_t oi;
 804     bufferlist bl;
 805     map<string, bufferptr>::iterator k;
 806     SnapSet ss;
 807     bufferlist ss_bl, hk_bl;
 808
 809     if (i->second.stat_error) {
 810       shard_info.set_stat_error();
 811       error_string += " stat_error";
 812       // With stat_error no further checking
 813       // We don't need to also see a missing_object_info_attr
 814       goto out;
 815     }
 816
 817     // We won't pick an auth copy if the snapset is missing or won't decode.
 818     if (obj.is_head() || obj.is_snapdir()) {
 819       k = i->second.attrs.find(SS_ATTR);
 820       if (k == i->second.attrs.end()) {
 821         shard_info.set_snapset_missing();
 822         error_string += " snapset_missing";
 823       } else {
 824         ss_bl.push_back(k->second);
 825         try {
 826           bufferlist::iterator bliter = ss_bl.begin();
 827           ::decode(ss, bliter);
 828           if (first_ss_bl.length() == 0) {
 829             first_ss_bl.append(ss_bl);
 830           } else if (!object_error.has_snapset_inconsistency() && !ss_bl.contents_equal(first_ss_bl)) {
 831             object_error.set_snapset_inconsistency();
 832             error_string += " snapset_inconsistency";
 833           }
 834         } catch (...) {
 835           // invalid snapset, probably corrupt
 836           shard_info.set_snapset_corrupted();
 837           error_string += " snapset_corrupted";
 838         }
 839       }
 840     }
 841
 842     if (parent->get_pool().is_erasure()) {
 843       ECUtil::HashInfo hi;
 844       k = i->second.attrs.find(ECUtil::get_hinfo_key());
 845       if (k == i->second.attrs.end()) {
 846         shard_info.set_hinfo_missing();
 847         error_string += " hinfo_key_missing";
 848       } else {
 849         hk_bl.push_back(k->second);
 850         try {
 851           bufferlist::iterator bliter = hk_bl.begin();
 852           decode(hi, bliter);
 853           if (first_hk_bl.length() == 0) {
 854             first_hk_bl.append(hk_bl);
 855           } else if (!object_error.has_hinfo_inconsistency() && !hk_bl.contents_equal(first_hk_bl)) {
 856             object_error.set_hinfo_inconsistency();
 857             error_string += " hinfo_inconsistency";
 858           }
 859         } catch (...) {
 860           // invalid snapset, probably corrupt
 861           shard_info.set_hinfo_corrupted();
 862           error_string += " hinfo_corrupted";
 863         }
 864       }
 865     }
 866
 867     k = i->second.attrs.find(OI_ATTR);
 868     if (k == i->second.attrs.end()) {
 869       // no object info on object, probably corrupt
 870       shard_info.set_info_missing();
 871       error_string += " info_missing";
 872       goto out;
 873     }
 874     bl.push_back(k->second);
 875     try {
 876       bufferlist::iterator bliter = bl.begin();
 877       ::decode(oi, bliter);
 878     } catch (...) {
 879       // invalid object info, probably corrupt
 880       shard_info.set_info_corrupted();
 881       error_string += " info_corrupted";
 882       goto out;
 883     }
 884
 885     // This is automatically corrected in PG::_repair_oinfo_oid()
 886     assert(oi.soid == obj);
 887
 888     if (first_oi_bl.length() == 0) {
 889       first_oi_bl.append(bl);
 890     } else if (!object_error.has_object_info_inconsistency() && !bl.contents_equal(first_oi_bl)) {
 891       object_error.set_object_info_inconsistency();
 892       error_string += " object_info_inconsistency";
 893     }
 894
 895     if (i->second.size != be_get_ondisk_size(oi.size)) {
 896       dout(5) << __func__ << " size " << i->second.size << " oi size " << oi.size << dendl;
 897       shard_info.set_obj_size_info_mismatch();
 898       error_string += " obj_size_info_mismatch";
 899     }
 900
 901     // Don't use this particular shard due to previous errors
 902     // XXX: For now we can't pick one shard for repair and another's object info or snapset
 903     if (shard_info.errors)
 904       goto out;
 905
 906     if (auth_version == eversion_t() || oi.version > auth_version ||
 907         (oi.version == auth_version && dcount(oi) > dcount(*auth_oi))) {
 908       auth = j;
 909       *auth_oi = oi;
 910       auth_version = oi.version;
 911     }
 912
 913 out:
 914     // Check error_string because some errors already generated messages
 915     if (error_string != "") {
 916       dout(10) << __func__ << ": error(s) osd " << j->first
 917                << " for obj " << obj
 918                << "," << error_string
 919                << dendl;
 920     }
 921     // Keep scanning other shards
 922   }
 923   dout(10) << __func__ << ": selecting osd " << auth->first
 924            << " for obj " << obj
 925            << " with oi " << *auth_oi
 926            << dendl;
 927   return auth;
 928 }
 929
 930 void PGBackend::be_compare_scrubmaps(
 931   const map<pg_shard_t,ScrubMap*> &maps,
 932   const set<hobject_t> &master_set,
 933   bool repair,
 934   map<hobject_t, set<pg_shard_t>> &missing,
 935   map<hobject_t, set<pg_shard_t>> &inconsistent,
 936   map<hobject_t, list<pg_shard_t>> &authoritative,
 937   map<hobject_t, pair<boost::optional<uint32_t>,
 938                       boost::optional<uint32_t>>> &missing_digest,
 939   int &shallow_errors, int &deep_errors,
 940   Scrub::Store *store,
 941   const spg_t& pgid,
 942   const vector<int> &acting,
 943   ostream &errorstream)
 944 {
 945   utime_t now = ceph_clock_now();
 946
 947   // Check maps against master set and each other
 948   for (set<hobject_t>::const_iterator k = master_set.begin();
 949        k != master_set.end();
 950        ++k) {
 951     object_info_t auth_oi;
 952     map<pg_shard_t, shard_info_wrapper> shard_map;
 953
 954     inconsistent_obj_wrapper object_error{*k};
 955
 956     map<pg_shard_t, ScrubMap *>::const_iterator auth =
 957       be_select_auth_object(*k, maps, &auth_oi, shard_map, object_error);
 958
 959     list<pg_shard_t> auth_list;
 960     set<pg_shard_t> object_errors;
 961     if (auth == maps.end()) {
 962       object_error.set_version(0);
 963       object_error.set_auth_missing(*k, maps, shard_map, shallow_errors,
 964         deep_errors, get_parent()->whoami_shard());
 965       if (object_error.has_deep_errors())
 966         ++deep_errors;
 967       else if (object_error.has_shallow_errors())
 968         ++shallow_errors;
 969       store->add_object_error(k->pool, object_error);
 970       errorstream << pgid.pgid << " soid " << *k
 971                   << ": failed to pick suitable object info\n";
 972       continue;
 973     }
 974     object_error.set_version(auth_oi.user_version);
 975     ScrubMap::object& auth_object = auth->second->objects[*k];
 976     set<pg_shard_t> cur_missing;
 977     set<pg_shard_t> cur_inconsistent;
 978
 979     for (auto  j = maps.cbegin(); j != maps.cend(); ++j) {
 980       if (j == auth)
 981         shard_map[auth->first].selected_oi = true;
 982       if (j->second->objects.count(*k)) {
 983         shard_map[j->first].set_object(j->second->objects[*k]);
 984         // Compare
 985         stringstream ss;
 986         bool found = be_compare_scrub_objects(auth->first,
 987                                    auth_object,
 988                                    auth_oi,
 989                                    j->second->objects[*k],
 990                                    shard_map[j->first],
 991                                    object_error,
 992                                    ss);
 993         // Some errors might have already been set in be_select_auth_object()
 994         if (shard_map[j->first].errors != 0) {
 995           cur_inconsistent.insert(j->first);
 996           if (shard_map[j->first].has_deep_errors())
 997             ++deep_errors;
 998           else
 999             ++shallow_errors;
1000           // Only true if be_compare_scrub_objects() found errors and put something
1001           // in ss.
1002           if (found)
1003             errorstream << pgid << " shard " << j->first << ": soid " << *k
1004                       << " " << ss.str() << "\n";
1005         } else if (found) {
1006           // Track possible shard to use as authoritative, if needed
1007           // There are errors, without identifying the shard
1008           object_errors.insert(j->first);
1009           errorstream << pgid << " : soid " << *k << " " << ss.str() << "\n";
1010         } else {
1011           // XXX: The auth shard might get here that we don't know
1012           // that it has the "correct" data.
1013           auth_list.push_back(j->first);
1014         }
1015       } else {
1016         cur_missing.insert(j->first);
1017         shard_map[j->first].set_missing();
1018         shard_map[j->first].primary = (j->first == get_parent()->whoami_shard());
1019         // Can't have any other errors if there is no information available
1020         ++shallow_errors;
1021         errorstream << pgid << " shard " << j->first << " missing " << *k
1022                     << "\n";
1023       }
1024       object_error.add_shard(j->first, shard_map[j->first]);
1025     }
1026
1027     if (auth_list.empty()) {
1028       if (object_errors.empty()) {
1029         errorstream << pgid.pgid << " soid " << *k
1030                   << ": failed to pick suitable auth object\n";
1031         goto out;
1032       }
1033       // Object errors exist and nothing in auth_list
1034       // Prefer the auth shard otherwise take first from list.
1035       pg_shard_t shard;
1036       if (object_errors.count(auth->first)) {
1037         shard = auth->first;
1038       } else {
1039         shard = *(object_errors.begin());
1040       }
1041       auth_list.push_back(shard);
1042       object_errors.erase(shard);
1043     }
1044     // At this point auth_list is populated, so we add the object errors shards
1045     // as inconsistent.
1046     cur_inconsistent.insert(object_errors.begin(), object_errors.end());
1047     if (!cur_missing.empty()) {
1048       missing[*k] = cur_missing;
1049     }
1050     if (!cur_inconsistent.empty()) {
1051       inconsistent[*k] = cur_inconsistent;
1052     }
1053     if (!cur_inconsistent.empty() || !cur_missing.empty()) {
1054       authoritative[*k] = auth_list;
1055     } else if (parent->get_pool().is_replicated()) {
1056       enum {
1057         NO = 0,
1058         MAYBE = 1,
1059         FORCE = 2,
1060       } update = NO;
1061
1062       if (auth_object.digest_present && !auth_oi.is_data_digest()) {
1063         dout(20) << __func__ << " missing data digest on " << *k << dendl;
1064         update = MAYBE;
1065       }
1066       if (auth_object.omap_digest_present && !auth_oi.is_omap_digest()) {
1067         dout(20) << __func__ << " missing omap digest on " << *k << dendl;
1068         update = MAYBE;
1069       }
1070
1071       // recorded digest != actual digest?
1072       if (auth_oi.is_data_digest() && auth_object.digest_present &&
1073           auth_oi.data_digest != auth_object.digest) {
1074         assert(shard_map[auth->first].has_data_digest_mismatch_info());
1075         errorstream << pgid << " recorded data digest 0x"
1076                     << std::hex << auth_oi.data_digest << " != on disk 0x"
1077                     << auth_object.digest << std::dec << " on " << auth_oi.soid
1078                     << "\n";
1079         if (repair)
1080           update = FORCE;
1081       }
1082       if (auth_oi.is_omap_digest() && auth_object.omap_digest_present &&
1083           auth_oi.omap_digest != auth_object.omap_digest) {
1084         assert(shard_map[auth->first].has_omap_digest_mismatch_info());
1085         errorstream << pgid << " recorded omap digest 0x"
1086                     << std::hex << auth_oi.omap_digest << " != on disk 0x"
1087                     << auth_object.omap_digest << std::dec
1088                     << " on " << auth_oi.soid << "\n";
1089         if (repair)
1090           update = FORCE;
1091       }
1092
1093       if (update != NO) {
1094         utime_t age = now - auth_oi.local_mtime;
1095         if (update == FORCE ||
1096             age > cct->_conf->osd_deep_scrub_update_digest_min_age) {
1097           boost::optional<uint32_t> data_digest, omap_digest;
1098           if (auth_object.digest_present) {
1099             data_digest = auth_object.digest;
1100             dout(20) << __func__ << " will update data digest on " << *k << dendl;
1101           }
1102           if (auth_object.omap_digest_present) {
1103             omap_digest = auth_object.omap_digest;
1104             dout(20) << __func__ << " will update omap digest on " << *k << dendl;
1105           }
1106           missing_digest[*k] = make_pair(data_digest, omap_digest);
1107         } else {
1108           dout(20) << __func__ << " missing digest but age " << age
1109                    << " < " << cct->_conf->osd_deep_scrub_update_digest_min_age
1110                    << " on " << *k << dendl;
1111         }
1112       }
1113     }
1114 out:
1115     if (object_error.has_deep_errors())
1116       ++deep_errors;
1117     else if (object_error.has_shallow_errors())
1118       ++shallow_errors;
1119     if (object_error.errors || object_error.union_shards.errors) {
1120       store->add_object_error(k->pool, object_error);
1121     }
1122   }
1123 }
1124
1125 void PGBackend::be_large_omap_check(const map<pg_shard_t,ScrubMap*> &maps,
1126   const set<hobject_t> &master_set,
1127   int& large_omap_objects,
1128   ostream &warnstream) const
1129 {
1130   bool needs_check = false;
1131   for (const auto& map : maps) {
1132     if (map.second->has_large_omap_object_errors) {
1133       needs_check = true;
1134       break;
1135     }
1136   }
1137
1138   if (!needs_check) {
1139     return;
1140   }
1141
1142   // Iterate through objects and check large omap object flag
1143   for (const auto& k : master_set) {
1144     for (const auto& map : maps) {
1145       ScrubMap::object& obj = map.second->objects[k];
1146       if (obj.large_omap_object_found) {
1147         large_omap_objects++;
1148         warnstream << "Large omap object found. Object: " << k << " Key count: "
1149                    << obj.large_omap_object_key_count << " Size (bytes): "
1150                    << obj.large_omap_object_value_size << '\n';
1151         break;
1152       }
1153     }
1154   }
1155 }