ceph/src/rgw/rgw_rados.h

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3
   4 #ifndef CEPH_RGWRADOS_H
   5 #define CEPH_RGWRADOS_H
   6
   7 #include <functional>
   8
   9 #include "include/rados/librados.hpp"
  10 #include "include/Context.h"
  11 #include "common/RefCountedObj.h"
  12 #include "common/RWLock.h"
  13 #include "common/ceph_time.h"
  14 #include "common/lru_map.h"
  15 #include "rgw_common.h"
  16 #include "cls/rgw/cls_rgw_types.h"
  17 #include "cls/version/cls_version_types.h"
  18 #include "cls/log/cls_log_types.h"
  19 #include "cls/statelog/cls_statelog_types.h"
  20 #include "cls/timeindex/cls_timeindex_types.h"
  21 #include "rgw_log.h"
  22 #include "rgw_metadata.h"
  23 #include "rgw_meta_sync_status.h"
  24 #include "rgw_period_puller.h"
  25 #include "rgw_sync_module.h"
  26
  27 class RGWWatcher;
  28 class SafeTimer;
  29 class ACLOwner;
  30 class RGWGC;
  31 class RGWMetaNotifier;
  32 class RGWDataNotifier;
  33 class RGWLC;
  34 class RGWObjectExpirer;
  35 class RGWMetaSyncProcessorThread;
  36 class RGWDataSyncProcessorThread;
  37 class RGWSyncLogTrimThread;
  38 class RGWRESTConn;
  39 struct RGWZoneGroup;
  40 struct RGWZoneParams;
  41 class RGWReshard;
  42 class RGWReshardWait;
  43
  44 /* flags for put_obj_meta() */
  45 #define PUT_OBJ_CREATE      0x01
  46 #define PUT_OBJ_EXCL        0x02
  47 #define PUT_OBJ_CREATE_EXCL (PUT_OBJ_CREATE | PUT_OBJ_EXCL)
  48
  49 #define RGW_OBJ_NS_MULTIPART "multipart"
  50 #define RGW_OBJ_NS_SHADOW    "shadow"
  51
  52 #define RGW_BUCKET_INSTANCE_MD_PREFIX ".bucket.meta."
  53
  54 #define RGW_NO_SHARD -1
  55
  56 #define RGW_SHARDS_PRIME_0 7877
  57 #define RGW_SHARDS_PRIME_1 65521
  58
  59 static inline int rgw_shards_mod(unsigned hval, int max_shards)
  60 {
  61   if (max_shards <= RGW_SHARDS_PRIME_0) {
  62     return hval % RGW_SHARDS_PRIME_0 % max_shards;
  63   }
  64   return hval % RGW_SHARDS_PRIME_1 % max_shards;
  65 }
  66
  67 static inline int rgw_shards_hash(const string& key, int max_shards)
  68 {
  69   return rgw_shards_mod(ceph_str_hash_linux(key.c_str(), key.size()), max_shards);
  70 }
  71
  72 static inline int rgw_shards_max()
  73 {
  74   return RGW_SHARDS_PRIME_1;
  75 }
  76
  77 static inline void prepend_bucket_marker(const rgw_bucket& bucket, const string& orig_oid, string& oid)
  78 {
  79   if (bucket.marker.empty() || orig_oid.empty()) {
  80     oid = orig_oid;
  81   } else {
  82     oid = bucket.marker;
  83     oid.append("_");
  84     oid.append(orig_oid);
  85   }
  86 }
  87
  88 static inline void get_obj_bucket_and_oid_loc(const rgw_obj& obj, string& oid, string& locator)
  89 {
  90   const rgw_bucket& bucket = obj.bucket;
  91   prepend_bucket_marker(bucket, obj.get_oid(), oid);
  92   const string& loc = obj.key.get_loc();
  93   if (!loc.empty()) {
  94     prepend_bucket_marker(bucket, loc, locator);
  95   } else {
  96     locator.clear();
  97   }
  98 }
  99
 100 int rgw_init_ioctx(librados::Rados *rados, const rgw_pool& pool, librados::IoCtx& ioctx, bool create = false);
 101
 102 int rgw_policy_from_attrset(CephContext *cct, map<string, bufferlist>& attrset, RGWAccessControlPolicy *policy);
 103
 104 static inline bool rgw_raw_obj_to_obj(const rgw_bucket& bucket, const rgw_raw_obj& raw_obj, rgw_obj *obj)
 105 {
 106   ssize_t pos = raw_obj.oid.find('_');
 107   if (pos < 0) {
 108     return false;
 109   }
 110
 111   if (!rgw_obj_key::parse_raw_oid(raw_obj.oid.substr(pos + 1), &obj->key)) {
 112     return false;
 113   }
 114   obj->bucket = bucket;
 115
 116   return true;
 117 }
 118
 119 struct rgw_bucket_placement {
 120   string placement_rule;
 121   rgw_bucket bucket;
 122
 123   void dump(Formatter *f) const;
 124 };
 125
 126 class rgw_obj_select {
 127   string placement_rule;
 128   rgw_obj obj;
 129   rgw_raw_obj raw_obj;
 130   bool is_raw;
 131
 132 public:
 133   rgw_obj_select() : is_raw(false) {}
 134   rgw_obj_select(const rgw_obj& _obj) : obj(_obj), is_raw(false) {}
 135   rgw_obj_select(const rgw_raw_obj& _raw_obj) : raw_obj(_raw_obj), is_raw(true) {}
 136   rgw_obj_select(const rgw_obj_select& rhs) {
 137     placement_rule = rhs.placement_rule;
 138     is_raw = rhs.is_raw;
 139     if (is_raw) {
 140       raw_obj = rhs.raw_obj;
 141     } else {
 142       obj = rhs.obj;
 143     }
 144   }
 145
 146   rgw_raw_obj get_raw_obj(const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params) const;
 147   rgw_raw_obj get_raw_obj(RGWRados *store) const;
 148
 149   rgw_obj_select& operator=(const rgw_obj& rhs) {
 150     obj = rhs;
 151     is_raw = false;
 152     return *this;
 153   }
 154
 155   rgw_obj_select& operator=(const rgw_raw_obj& rhs) {
 156     raw_obj = rhs;
 157     is_raw = true;
 158     return *this;
 159   }
 160
 161   void set_placement_rule(const string& rule) {
 162     placement_rule = rule;
 163   }
 164 };
 165
 166 struct compression_block {
 167   uint64_t old_ofs;
 168   uint64_t new_ofs;
 169   uint64_t len;
 170
 171   void encode(bufferlist& bl) const {
 172     ENCODE_START(1, 1, bl);
 173     ::encode(old_ofs, bl);
 174     ::encode(new_ofs, bl);
 175     ::encode(len, bl);
 176     ENCODE_FINISH(bl);
 177   }
 178
 179   void decode(bufferlist::iterator& bl) {
 180      DECODE_START(1, bl);
 181      ::decode(old_ofs, bl);
 182      ::decode(new_ofs, bl);
 183      ::decode(len, bl);
 184      DECODE_FINISH(bl);
 185   }
 186 };
 187 WRITE_CLASS_ENCODER(compression_block)
 188
 189 struct RGWCompressionInfo {
 190   string compression_type;
 191   uint64_t orig_size;
 192   vector<compression_block> blocks;
 193
 194   RGWCompressionInfo() : compression_type("none"), orig_size(0) {}
 195   RGWCompressionInfo(const RGWCompressionInfo& cs_info) : compression_type(cs_info.compression_type),
 196                                                           orig_size(cs_info.orig_size),
 197                                                           blocks(cs_info.blocks) {}
 198
 199   void encode(bufferlist& bl) const {
 200     ENCODE_START(1, 1, bl);
 201     ::encode(compression_type, bl);
 202     ::encode(orig_size, bl);
 203     ::encode(blocks, bl);
 204     ENCODE_FINISH(bl);
 205   }
 206
 207   void decode(bufferlist::iterator& bl) {
 208      DECODE_START(1, bl);
 209      ::decode(compression_type, bl);
 210      ::decode(orig_size, bl);
 211      ::decode(blocks, bl);
 212      DECODE_FINISH(bl);
 213   }
 214 };
 215 WRITE_CLASS_ENCODER(RGWCompressionInfo)
 216
 217 int rgw_compression_info_from_attrset(map<string, bufferlist>& attrs, bool& need_decompress, RGWCompressionInfo& cs_info);
 218
 219 struct RGWOLHInfo {
 220   rgw_obj target;
 221   bool removed;
 222
 223   RGWOLHInfo() : removed(false) {}
 224
 225   void encode(bufferlist& bl) const {
 226     ENCODE_START(1, 1, bl);
 227     ::encode(target, bl);
 228     ::encode(removed, bl);
 229     ENCODE_FINISH(bl);
 230   }
 231
 232   void decode(bufferlist::iterator& bl) {
 233      DECODE_START(1, bl);
 234      ::decode(target, bl);
 235      ::decode(removed, bl);
 236      DECODE_FINISH(bl);
 237   }
 238   static void generate_test_instances(list<RGWOLHInfo*>& o);
 239   void dump(Formatter *f) const;
 240 };
 241 WRITE_CLASS_ENCODER(RGWOLHInfo)
 242
 243 struct RGWOLHPendingInfo {
 244   ceph::real_time time;
 245
 246   RGWOLHPendingInfo() {}
 247
 248   void encode(bufferlist& bl) const {
 249     ENCODE_START(1, 1, bl);
 250     ::encode(time, bl);
 251     ENCODE_FINISH(bl);
 252   }
 253
 254   void decode(bufferlist::iterator& bl) {
 255      DECODE_START(1, bl);
 256      ::decode(time, bl);
 257      DECODE_FINISH(bl);
 258   }
 259
 260   void dump(Formatter *f) const;
 261 };
 262 WRITE_CLASS_ENCODER(RGWOLHPendingInfo)
 263
 264 struct RGWUsageBatch {
 265   map<ceph::real_time, rgw_usage_log_entry> m;
 266
 267   void insert(ceph::real_time& t, rgw_usage_log_entry& entry, bool *account) {
 268     bool exists = m.find(t) != m.end();
 269     *account = !exists;
 270     m[t].aggregate(entry);
 271   }
 272 };
 273
 274 struct RGWUsageIter {
 275   string read_iter;
 276   uint32_t index;
 277
 278   RGWUsageIter() : index(0) {}
 279 };
 280
 281 class RGWGetDataCB {
 282 protected:
 283   uint64_t extra_data_len;
 284 public:
 285   virtual int handle_data(bufferlist& bl, off_t bl_ofs, off_t bl_len) = 0;
 286   RGWGetDataCB() : extra_data_len(0) {}
 287   virtual ~RGWGetDataCB() {}
 288   virtual void set_extra_data_len(uint64_t len) {
 289     extra_data_len = len;
 290   }
 291   /**
 292    * Flushes any cached data. Used by RGWGetObjFilter.
 293    * Return logic same as handle_data.
 294    */
 295   virtual int flush() {
 296     return 0;
 297   }
 298   /**
 299    * Allows to extend fetch range of RGW object. Used by RGWGetObjFilter.
 300    */
 301   virtual int fixup_range(off_t& bl_ofs, off_t& bl_end) {
 302     return 0;
 303   }
 304 };
 305
 306 class RGWAccessListFilter {
 307 public:
 308   virtual ~RGWAccessListFilter() {}
 309   virtual bool filter(string& name, string& key) = 0;
 310 };
 311
 312 struct RGWCloneRangeInfo {
 313   rgw_obj src;
 314   off_t src_ofs;
 315   off_t dst_ofs;
 316   uint64_t len;
 317 };
 318
 319 struct RGWObjManifestPart {
 320   rgw_obj loc;   /* the object where the data is located */
 321   uint64_t loc_ofs;  /* the offset at that object where the data is located */
 322   uint64_t size;     /* the part size */
 323
 324   RGWObjManifestPart() : loc_ofs(0), size(0) {}
 325
 326   void encode(bufferlist& bl) const {
 327     ENCODE_START(2, 2, bl);
 328     ::encode(loc, bl);
 329     ::encode(loc_ofs, bl);
 330     ::encode(size, bl);
 331     ENCODE_FINISH(bl);
 332   }
 333
 334   void decode(bufferlist::iterator& bl) {
 335      DECODE_START_LEGACY_COMPAT_LEN_32(2, 2, 2, bl);
 336      ::decode(loc, bl);
 337      ::decode(loc_ofs, bl);
 338      ::decode(size, bl);
 339      DECODE_FINISH(bl);
 340   }
 341
 342   void dump(Formatter *f) const;
 343   static void generate_test_instances(list<RGWObjManifestPart*>& o);
 344 };
 345 WRITE_CLASS_ENCODER(RGWObjManifestPart)
 346
 347 /*
 348  The manifest defines a set of rules for structuring the object parts.
 349  There are a few terms to note:
 350      - head: the head part of the object, which is the part that contains
 351        the first chunk of data. An object might not have a head (as in the
 352        case of multipart-part objects).
 353      - stripe: data portion of a single rgw object that resides on a single
 354        rados object.
 355      - part: a collection of stripes that make a contiguous part of an
 356        object. A regular object will only have one part (although might have
 357        many stripes), a multipart object might have many parts. Each part
 358        has a fixed stripe size, although the last stripe of a part might
 359        be smaller than that. Consecutive parts may be merged if their stripe
 360        value is the same.
 361 */
 362
 363 struct RGWObjManifestRule {
 364   uint32_t start_part_num;
 365   uint64_t start_ofs;
 366   uint64_t part_size; /* each part size, 0 if there's no part size, meaning it's unlimited */
 367   uint64_t stripe_max_size; /* underlying obj max size */
 368   string override_prefix;
 369
 370   RGWObjManifestRule() : start_part_num(0), start_ofs(0), part_size(0), stripe_max_size(0) {}
 371   RGWObjManifestRule(uint32_t _start_part_num, uint64_t _start_ofs, uint64_t _part_size, uint64_t _stripe_max_size) :
 372                        start_part_num(_start_part_num), start_ofs(_start_ofs), part_size(_part_size), stripe_max_size(_stripe_max_size) {}
 373
 374   void encode(bufferlist& bl) const {
 375     ENCODE_START(2, 1, bl);
 376     ::encode(start_part_num, bl);
 377     ::encode(start_ofs, bl);
 378     ::encode(part_size, bl);
 379     ::encode(stripe_max_size, bl);
 380     ::encode(override_prefix, bl);
 381     ENCODE_FINISH(bl);
 382   }
 383
 384   void decode(bufferlist::iterator& bl) {
 385     DECODE_START(2, bl);
 386     ::decode(start_part_num, bl);
 387     ::decode(start_ofs, bl);
 388     ::decode(part_size, bl);
 389     ::decode(stripe_max_size, bl);
 390     if (struct_v >= 2)
 391       ::decode(override_prefix, bl);
 392     DECODE_FINISH(bl);
 393   }
 394   void dump(Formatter *f) const;
 395 };
 396 WRITE_CLASS_ENCODER(RGWObjManifestRule)
 397
 398 class RGWObjManifest {
 399 protected:
 400   bool explicit_objs; /* old manifest? */
 401   map<uint64_t, RGWObjManifestPart> objs;
 402
 403   uint64_t obj_size;
 404
 405   rgw_obj obj;
 406   uint64_t head_size;
 407   string head_placement_rule;
 408
 409   uint64_t max_head_size;
 410   string prefix;
 411   rgw_bucket_placement tail_placement; /* might be different than the original bucket,
 412                                        as object might have been copied across pools */
 413   map<uint64_t, RGWObjManifestRule> rules;
 414
 415   string tail_instance; /* tail object's instance */
 416
 417   void convert_to_explicit(const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params);
 418   int append_explicit(RGWObjManifest& m, const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params);
 419   void append_rules(RGWObjManifest& m, map<uint64_t, RGWObjManifestRule>::iterator& iter, string *override_prefix);
 420
 421   void update_iterators() {
 422     begin_iter.seek(0);
 423     end_iter.seek(obj_size);
 424   }
 425 public:
 426
 427   RGWObjManifest() : explicit_objs(false), obj_size(0), head_size(0), max_head_size(0),
 428                      begin_iter(this), end_iter(this) {}
 429   RGWObjManifest(const RGWObjManifest& rhs) {
 430     *this = rhs;
 431   }
 432   RGWObjManifest& operator=(const RGWObjManifest& rhs) {
 433     explicit_objs = rhs.explicit_objs;
 434     objs = rhs.objs;
 435     obj_size = rhs.obj_size;
 436     obj = rhs.obj;
 437     head_size = rhs.head_size;
 438     max_head_size = rhs.max_head_size;
 439     prefix = rhs.prefix;
 440     tail_placement = rhs.tail_placement;
 441     rules = rhs.rules;
 442     tail_instance = rhs.tail_instance;
 443
 444     begin_iter.set_manifest(this);
 445     end_iter.set_manifest(this);
 446
 447     begin_iter.seek(rhs.begin_iter.get_ofs());
 448     end_iter.seek(rhs.end_iter.get_ofs());
 449
 450     return *this;
 451   }
 452
 453   map<uint64_t, RGWObjManifestPart>& get_explicit_objs() {
 454     return objs;
 455   }
 456
 457
 458   void set_explicit(uint64_t _size, map<uint64_t, RGWObjManifestPart>& _objs) {
 459     explicit_objs = true;
 460     obj_size = _size;
 461     objs.swap(_objs);
 462   }
 463
 464   void get_implicit_location(uint64_t cur_part_id, uint64_t cur_stripe, uint64_t ofs, string *override_prefix, rgw_obj_select *location);
 465
 466   void set_trivial_rule(uint64_t tail_ofs, uint64_t stripe_max_size) {
 467     RGWObjManifestRule rule(0, tail_ofs, 0, stripe_max_size);
 468     rules[0] = rule;
 469     max_head_size = tail_ofs;
 470   }
 471
 472   void set_multipart_part_rule(uint64_t stripe_max_size, uint64_t part_num) {
 473     RGWObjManifestRule rule(0, 0, 0, stripe_max_size);
 474     rule.start_part_num = part_num;
 475     rules[0] = rule;
 476     max_head_size = 0;
 477   }
 478
 479   void encode(bufferlist& bl) const {
 480     ENCODE_START(7, 6, bl);
 481     ::encode(obj_size, bl);
 482     ::encode(objs, bl);
 483     ::encode(explicit_objs, bl);
 484     ::encode(obj, bl);
 485     ::encode(head_size, bl);
 486     ::encode(max_head_size, bl);
 487     ::encode(prefix, bl);
 488     ::encode(rules, bl);
 489     bool encode_tail_bucket = !(tail_placement.bucket == obj.bucket);
 490     ::encode(encode_tail_bucket, bl);
 491     if (encode_tail_bucket) {
 492       ::encode(tail_placement.bucket, bl);
 493     }
 494     bool encode_tail_instance = (tail_instance != obj.key.instance);
 495     ::encode(encode_tail_instance, bl);
 496     if (encode_tail_instance) {
 497       ::encode(tail_instance, bl);
 498     }
 499     ::encode(head_placement_rule, bl);
 500     ::encode(tail_placement.placement_rule, bl);
 501     ENCODE_FINISH(bl);
 502   }
 503
 504   void decode(bufferlist::iterator& bl) {
 505     DECODE_START_LEGACY_COMPAT_LEN_32(7, 2, 2, bl);
 506     ::decode(obj_size, bl);
 507     ::decode(objs, bl);
 508     if (struct_v >= 3) {
 509       ::decode(explicit_objs, bl);
 510       ::decode(obj, bl);
 511       ::decode(head_size, bl);
 512       ::decode(max_head_size, bl);
 513       ::decode(prefix, bl);
 514       ::decode(rules, bl);
 515     } else {
 516       explicit_objs = true;
 517       if (!objs.empty()) {
 518         map<uint64_t, RGWObjManifestPart>::iterator iter = objs.begin();
 519         obj = iter->second.loc;
 520         head_size = iter->second.size;
 521         max_head_size = head_size;
 522       }
 523     }
 524
 525     if (explicit_objs && head_size > 0 && !objs.empty()) {
 526       /* patch up manifest due to issue 16435:
 527        * the first object in the explicit objs list might not be the one we need to access, use the
 528        * head object instead if set. This would happen if we had an old object that was created
 529        * when the explicit objs manifest was around, and it got copied.
 530        */
 531       rgw_obj& obj_0 = objs[0].loc;
 532       if (!obj_0.get_oid().empty() && obj_0.key.ns.empty()) {
 533         objs[0].loc = obj;
 534         objs[0].size = head_size;
 535       }
 536     }
 537
 538     if (struct_v >= 4) {
 539       if (struct_v < 6) {
 540         ::decode(tail_placement.bucket, bl);
 541       } else {
 542         bool need_to_decode;
 543         ::decode(need_to_decode, bl);
 544         if (need_to_decode) {
 545           ::decode(tail_placement.bucket, bl);
 546         } else {
 547           tail_placement.bucket = obj.bucket;
 548         }
 549       }
 550     }
 551
 552     if (struct_v >= 5) {
 553       if (struct_v < 6) {
 554         ::decode(tail_instance, bl);
 555       } else {
 556         bool need_to_decode;
 557         ::decode(need_to_decode, bl);
 558         if (need_to_decode) {
 559           ::decode(tail_instance, bl);
 560         } else {
 561           tail_instance = obj.key.instance;
 562         }
 563       }
 564     } else { // old object created before 'tail_instance' field added to manifest
 565       tail_instance = obj.key.instance;
 566     }
 567
 568     if (struct_v >= 7) {
 569       ::decode(head_placement_rule, bl);
 570       ::decode(tail_placement.placement_rule, bl);
 571     }
 572
 573     update_iterators();
 574     DECODE_FINISH(bl);
 575   }
 576
 577   void dump(Formatter *f) const;
 578   static void generate_test_instances(list<RGWObjManifest*>& o);
 579
 580   int append(RGWObjManifest& m, RGWZoneGroup& zonegroup, RGWZoneParams& zone_params);
 581   int append(RGWObjManifest& m, RGWRados *store);
 582
 583   bool get_rule(uint64_t ofs, RGWObjManifestRule *rule);
 584
 585   bool empty() {
 586     if (explicit_objs)
 587       return objs.empty();
 588     return rules.empty();
 589   }
 590
 591   bool has_explicit_objs() {
 592     return explicit_objs;
 593   }
 594
 595   bool has_tail() {
 596     if (explicit_objs) {
 597       if (objs.size() == 1) {
 598         map<uint64_t, RGWObjManifestPart>::iterator iter = objs.begin();
 599         rgw_obj& o = iter->second.loc;
 600         return !(obj == o);
 601       }
 602       return (objs.size() >= 2);
 603     }
 604     return (obj_size > head_size);
 605   }
 606
 607   void set_head(const string& placement_rule, const rgw_obj& _o, uint64_t _s) {
 608     head_placement_rule = placement_rule;
 609     obj = _o;
 610     head_size = _s;
 611
 612     if (explicit_objs && head_size > 0) {
 613       objs[0].loc = obj;
 614       objs[0].size = head_size;
 615     }
 616   }
 617
 618   const rgw_obj& get_obj() {
 619     return obj;
 620   }
 621
 622   void set_tail_placement(const string& placement_rule, const rgw_bucket& _b) {
 623     tail_placement.placement_rule = placement_rule;
 624     tail_placement.bucket = _b;
 625   }
 626
 627   const rgw_bucket_placement& get_tail_placement() {
 628     return tail_placement;
 629   }
 630
 631   const string& get_head_placement_rule() {
 632     return head_placement_rule;
 633   }
 634
 635   void set_prefix(const string& _p) {
 636     prefix = _p;
 637   }
 638
 639   const string& get_prefix() {
 640     return prefix;
 641   }
 642
 643   void set_tail_instance(const string& _ti) {
 644     tail_instance = _ti;
 645   }
 646
 647   const string& get_tail_instance() {
 648     return tail_instance;
 649   }
 650
 651   void set_head_size(uint64_t _s) {
 652     head_size = _s;
 653   }
 654
 655   void set_obj_size(uint64_t s) {
 656     obj_size = s;
 657
 658     update_iterators();
 659   }
 660
 661   uint64_t get_obj_size() {
 662     return obj_size;
 663   }
 664
 665   uint64_t get_head_size() {
 666     return head_size;
 667   }
 668
 669   void set_max_head_size(uint64_t s) {
 670     max_head_size = s;
 671   }
 672
 673   uint64_t get_max_head_size() {
 674     return max_head_size;
 675   }
 676
 677   class obj_iterator {
 678     RGWObjManifest *manifest;
 679     uint64_t part_ofs; /* where current part starts */
 680     uint64_t stripe_ofs; /* where current stripe starts */
 681     uint64_t ofs;       /* current position within the object */
 682     uint64_t stripe_size;      /* current part size */
 683
 684     int cur_part_id;
 685     int cur_stripe;
 686     string cur_override_prefix;
 687
 688     rgw_obj_select location;
 689
 690     map<uint64_t, RGWObjManifestRule>::iterator rule_iter;
 691     map<uint64_t, RGWObjManifestRule>::iterator next_rule_iter;
 692
 693     map<uint64_t, RGWObjManifestPart>::iterator explicit_iter;
 694
 695     void init() {
 696       part_ofs = 0;
 697       stripe_ofs = 0;
 698       ofs = 0;
 699       stripe_size = 0;
 700       cur_part_id = 0;
 701       cur_stripe = 0;
 702     }
 703
 704     void update_explicit_pos();
 705
 706
 707   protected:
 708
 709     void set_manifest(RGWObjManifest *m) {
 710       manifest = m;
 711     }
 712
 713   public:
 714     obj_iterator() : manifest(NULL) {
 715       init();
 716     }
 717     explicit obj_iterator(RGWObjManifest *_m) : manifest(_m) {
 718       init();
 719       if (!manifest->empty()) {
 720         seek(0);
 721       }
 722     }
 723     obj_iterator(RGWObjManifest *_m, uint64_t _ofs) : manifest(_m) {
 724       init();
 725       if (!manifest->empty()) {
 726         seek(_ofs);
 727       }
 728     }
 729     void seek(uint64_t ofs);
 730
 731     void operator++();
 732     bool operator==(const obj_iterator& rhs) {
 733       return (ofs == rhs.ofs);
 734     }
 735     bool operator!=(const obj_iterator& rhs) {
 736       return (ofs != rhs.ofs);
 737     }
 738     const rgw_obj_select& get_location() {
 739       return location;
 740     }
 741
 742     /* start of current stripe */
 743     uint64_t get_stripe_ofs() {
 744       if (manifest->explicit_objs) {
 745         return explicit_iter->first;
 746       }
 747       return stripe_ofs;
 748     }
 749
 750     /* current ofs relative to start of rgw object */
 751     uint64_t get_ofs() const {
 752       return ofs;
 753     }
 754
 755     /* stripe number */
 756     int get_cur_stripe() const {
 757       return cur_stripe;
 758     }
 759
 760     /* current stripe size */
 761     uint64_t get_stripe_size() {
 762       if (manifest->explicit_objs) {
 763         return explicit_iter->second.size;
 764       }
 765       return stripe_size;
 766     }
 767
 768     /* offset where data starts within current stripe */
 769     uint64_t location_ofs() {
 770       if (manifest->explicit_objs) {
 771         return explicit_iter->second.loc_ofs;
 772       }
 773       return 0; /* all stripes start at zero offset */
 774     }
 775
 776     void update_location();
 777
 778     friend class RGWObjManifest;
 779   };
 780
 781   const obj_iterator& obj_begin();
 782   const obj_iterator& obj_end();
 783   obj_iterator obj_find(uint64_t ofs);
 784
 785   obj_iterator begin_iter;
 786   obj_iterator end_iter;
 787
 788   /*
 789    * simple object generator. Using a simple single rule manifest.
 790    */
 791   class generator {
 792     RGWObjManifest *manifest;
 793     uint64_t last_ofs;
 794     uint64_t cur_part_ofs;
 795     int cur_part_id;
 796     int cur_stripe;
 797     uint64_t cur_stripe_size;
 798     string cur_oid;
 799
 800     string oid_prefix;
 801
 802     rgw_obj_select cur_obj;
 803
 804     RGWObjManifestRule rule;
 805
 806   public:
 807     generator() : manifest(NULL), last_ofs(0), cur_part_ofs(0), cur_part_id(0),
 808                   cur_stripe(0), cur_stripe_size(0) {}
 809     int create_begin(CephContext *cct, RGWObjManifest *manifest, const string& placement_rule, rgw_bucket& bucket, rgw_obj& obj);
 810
 811     int create_next(uint64_t ofs);
 812
 813     rgw_raw_obj get_cur_obj(RGWZoneGroup& zonegroup, RGWZoneParams& zone_params) { return cur_obj.get_raw_obj(zonegroup, zone_params); }
 814     rgw_raw_obj get_cur_obj(RGWRados *store) { return cur_obj.get_raw_obj(store); }
 815
 816     /* total max size of current stripe (including head obj) */
 817     uint64_t cur_stripe_max_size() {
 818       return cur_stripe_size;
 819     }
 820   };
 821 };
 822 WRITE_CLASS_ENCODER(RGWObjManifest)
 823
 824 struct RGWUploadPartInfo {
 825   uint32_t num;
 826   uint64_t size;
 827   uint64_t accounted_size{0};
 828   string etag;
 829   ceph::real_time modified;
 830   RGWObjManifest manifest;
 831   RGWCompressionInfo cs_info;
 832
 833   RGWUploadPartInfo() : num(0), size(0) {}
 834
 835   void encode(bufferlist& bl) const {
 836     ENCODE_START(4, 2, bl);
 837     ::encode(num, bl);
 838     ::encode(size, bl);
 839     ::encode(etag, bl);
 840     ::encode(modified, bl);
 841     ::encode(manifest, bl);
 842     ::encode(cs_info, bl);
 843     ::encode(accounted_size, bl);
 844     ENCODE_FINISH(bl);
 845   }
 846   void decode(bufferlist::iterator& bl) {
 847     DECODE_START_LEGACY_COMPAT_LEN(4, 2, 2, bl);
 848     ::decode(num, bl);
 849     ::decode(size, bl);
 850     ::decode(etag, bl);
 851     ::decode(modified, bl);
 852     if (struct_v >= 3)
 853       ::decode(manifest, bl);
 854     if (struct_v >= 4) {
 855       ::decode(cs_info, bl);
 856       ::decode(accounted_size, bl);
 857     } else {
 858       accounted_size = size;
 859     }
 860     DECODE_FINISH(bl);
 861   }
 862   void dump(Formatter *f) const;
 863   static void generate_test_instances(list<RGWUploadPartInfo*>& o);
 864 };
 865 WRITE_CLASS_ENCODER(RGWUploadPartInfo)
 866
 867 struct RGWObjState {
 868   rgw_obj obj;
 869   bool is_atomic;
 870   bool has_attrs;
 871   bool exists;
 872   uint64_t size; //< size of raw object
 873   uint64_t accounted_size{0}; //< size before compression, encryption
 874   ceph::real_time mtime;
 875   uint64_t epoch;
 876   bufferlist obj_tag;
 877   string write_tag;
 878   bool fake_tag;
 879   RGWObjManifest manifest;
 880   bool has_manifest;
 881   string shadow_obj;
 882   bool has_data;
 883   bufferlist data;
 884   bool prefetch_data;
 885   bool keep_tail;
 886   bool is_olh;
 887   bufferlist olh_tag;
 888   uint64_t pg_ver;
 889   uint32_t zone_short_id;
 890
 891   /* important! don't forget to update copy constructor */
 892
 893   RGWObjVersionTracker objv_tracker;
 894
 895   map<string, bufferlist> attrset;
 896   RGWObjState() : is_atomic(false), has_attrs(0), exists(false),
 897                   size(0), epoch(0), fake_tag(false), has_manifest(false),
 898                   has_data(false), prefetch_data(false), keep_tail(false), is_olh(false),
 899                   pg_ver(0), zone_short_id(0) {}
 900   RGWObjState(const RGWObjState& rhs) : obj (rhs.obj) {
 901     is_atomic = rhs.is_atomic;
 902     has_attrs = rhs.has_attrs;
 903     exists = rhs.exists;
 904     size = rhs.size;
 905     accounted_size = rhs.accounted_size;
 906     mtime = rhs.mtime;
 907     epoch = rhs.epoch;
 908     if (rhs.obj_tag.length()) {
 909       obj_tag = rhs.obj_tag;
 910     }
 911     write_tag = rhs.write_tag;
 912     fake_tag = rhs.fake_tag;
 913     if (rhs.has_manifest) {
 914       manifest = rhs.manifest;
 915     }
 916     has_manifest = rhs.has_manifest;
 917     shadow_obj = rhs.shadow_obj;
 918     has_data = rhs.has_data;
 919     if (rhs.data.length()) {
 920       data = rhs.data;
 921     }
 922     prefetch_data = rhs.prefetch_data;
 923     keep_tail = rhs.keep_tail;
 924     is_olh = rhs.is_olh;
 925     objv_tracker = rhs.objv_tracker;
 926     pg_ver = rhs.pg_ver;
 927   }
 928
 929   bool get_attr(string name, bufferlist& dest) {
 930     map<string, bufferlist>::iterator iter = attrset.find(name);
 931     if (iter != attrset.end()) {
 932       dest = iter->second;
 933       return true;
 934     }
 935     return false;
 936   }
 937 };
 938
 939 struct RGWRawObjState {
 940   rgw_raw_obj obj;
 941   bool has_attrs{false};
 942   bool exists{false};
 943   uint64_t size{0};
 944   ceph::real_time mtime;
 945   uint64_t epoch;
 946   bufferlist obj_tag;
 947   bool has_data{false};
 948   bufferlist data;
 949   bool prefetch_data{false};
 950   uint64_t pg_ver{0};
 951
 952   /* important! don't forget to update copy constructor */
 953
 954   RGWObjVersionTracker objv_tracker;
 955
 956   map<string, bufferlist> attrset;
 957   RGWRawObjState() {}
 958   RGWRawObjState(const RGWRawObjState& rhs) : obj (rhs.obj) {
 959     has_attrs = rhs.has_attrs;
 960     exists = rhs.exists;
 961     size = rhs.size;
 962     mtime = rhs.mtime;
 963     epoch = rhs.epoch;
 964     if (rhs.obj_tag.length()) {
 965       obj_tag = rhs.obj_tag;
 966     }
 967     has_data = rhs.has_data;
 968     if (rhs.data.length()) {
 969       data = rhs.data;
 970     }
 971     prefetch_data = rhs.prefetch_data;
 972     pg_ver = rhs.pg_ver;
 973     objv_tracker = rhs.objv_tracker;
 974   }
 975 };
 976
 977 struct RGWPoolIterCtx {
 978   librados::IoCtx io_ctx;
 979   librados::NObjectIterator iter;
 980 };
 981
 982 struct RGWListRawObjsCtx {
 983   bool initialized;
 984   RGWPoolIterCtx iter_ctx;
 985
 986   RGWListRawObjsCtx() : initialized(false) {}
 987 };
 988
 989 struct RGWDefaultSystemMetaObjInfo {
 990   string default_id;
 991
 992   void encode(bufferlist& bl) const {
 993     ENCODE_START(1, 1, bl);
 994     ::encode(default_id, bl);
 995     ENCODE_FINISH(bl);
 996   }
 997
 998   void decode(bufferlist::iterator& bl) {
 999     DECODE_START(1, bl);
1000     ::decode(default_id, bl);
1001     DECODE_FINISH(bl);
1002   }
1003
1004   void dump(Formatter *f) const;
1005   void decode_json(JSONObj *obj);
1006 };
1007 WRITE_CLASS_ENCODER(RGWDefaultSystemMetaObjInfo)
1008
1009 struct RGWNameToId {
1010   string obj_id;
1011
1012   void encode(bufferlist& bl) const {
1013     ENCODE_START(1, 1, bl);
1014     ::encode(obj_id, bl);
1015     ENCODE_FINISH(bl);
1016   }
1017
1018   void decode(bufferlist::iterator& bl) {
1019     DECODE_START(1, bl);
1020     ::decode(obj_id, bl);
1021     DECODE_FINISH(bl);
1022   }
1023
1024   void dump(Formatter *f) const;
1025   void decode_json(JSONObj *obj);
1026 };
1027 WRITE_CLASS_ENCODER(RGWNameToId)
1028
1029 class RGWSystemMetaObj {
1030 protected:
1031   string id;
1032   string name;
1033
1034   CephContext *cct;
1035   RGWRados *store;
1036
1037   int store_name(bool exclusive);
1038   int store_info(bool exclusive);
1039   int read_info(const string& obj_id, bool old_format = false);
1040   int read_id(const string& obj_name, string& obj_id);
1041   int read_default(RGWDefaultSystemMetaObjInfo& default_info,
1042                    const string& oid);
1043   /* read and use default id */
1044   int use_default(bool old_format = false);
1045
1046 public:
1047   RGWSystemMetaObj() : cct(NULL), store(NULL) {}
1048   RGWSystemMetaObj(const string& _name): name(_name), cct(NULL), store(NULL)  {}
1049   RGWSystemMetaObj(const string& _id, const string& _name) : id(_id), name(_name), cct(NULL), store(NULL) {}
1050   RGWSystemMetaObj(CephContext *_cct, RGWRados *_store): cct(_cct), store(_store){}
1051   RGWSystemMetaObj(const string& _name, CephContext *_cct, RGWRados *_store): name(_name), cct(_cct), store(_store){}
1052   const string& get_name() const { return name; }
1053   const string& get_id() const { return id; }
1054
1055   void set_name(const string& _name) { name = _name;}
1056   void set_id(const string& _id) { id = _id;}
1057   void clear_id() { id.clear(); }
1058
1059   virtual ~RGWSystemMetaObj() {}
1060
1061   virtual void encode(bufferlist& bl) const {
1062     ENCODE_START(1, 1, bl);
1063     ::encode(id, bl);
1064     ::encode(name, bl);
1065     ENCODE_FINISH(bl);
1066   }
1067
1068   virtual void decode(bufferlist::iterator& bl) {
1069     DECODE_START(1, bl);
1070     ::decode(id, bl);
1071     ::decode(name, bl);
1072     DECODE_FINISH(bl);
1073   }
1074
1075   void reinit_instance(CephContext *_cct, RGWRados *_store) {
1076     cct = _cct;
1077     store = _store;
1078   }
1079   int init(CephContext *_cct, RGWRados *_store, bool setup_obj = true, bool old_format = false);
1080   virtual int read_default_id(string& default_id, bool old_format = false);
1081   virtual int set_as_default(bool exclusive = false);
1082   int delete_default();
1083   virtual int create(bool exclusive = true);
1084   int delete_obj(bool old_format = false);
1085   int rename(const string& new_name);
1086   int update() { return store_info(false);}
1087   int update_name() { return store_name(false);}
1088   int read();
1089   int write(bool exclusive);
1090
1091   virtual rgw_pool get_pool(CephContext *cct) = 0;
1092   virtual const string get_default_oid(bool old_format = false) = 0;
1093   virtual const string& get_names_oid_prefix() = 0;
1094   virtual const string& get_info_oid_prefix(bool old_format = false) = 0;
1095   virtual const string& get_predefined_name(CephContext *cct) = 0;
1096
1097   void dump(Formatter *f) const;
1098   void decode_json(JSONObj *obj);
1099 };
1100 WRITE_CLASS_ENCODER(RGWSystemMetaObj)
1101
1102 struct RGWZonePlacementInfo {
1103   rgw_pool index_pool;
1104   rgw_pool data_pool;
1105   rgw_pool data_extra_pool; /* if not set we should use data_pool */
1106   RGWBucketIndexType index_type;
1107   std::string compression_type;
1108
1109   RGWZonePlacementInfo() : index_type(RGWBIType_Normal) {}
1110
1111   void encode(bufferlist& bl) const {
1112     ENCODE_START(6, 1, bl);
1113     ::encode(index_pool.to_str(), bl);
1114     ::encode(data_pool.to_str(), bl);
1115     ::encode(data_extra_pool.to_str(), bl);
1116     ::encode((uint32_t)index_type, bl);
1117     ::encode(compression_type, bl);
1118     ENCODE_FINISH(bl);
1119   }
1120
1121   void decode(bufferlist::iterator& bl) {
1122     DECODE_START(6, bl);
1123     string index_pool_str;
1124     string data_pool_str;
1125     ::decode(index_pool_str, bl);
1126     index_pool = rgw_pool(index_pool_str);
1127     ::decode(data_pool_str, bl);
1128     data_pool = rgw_pool(data_pool_str);
1129     if (struct_v >= 4) {
1130       string data_extra_pool_str;
1131       ::decode(data_extra_pool_str, bl);
1132       data_extra_pool = rgw_pool(data_extra_pool_str);
1133     }
1134     if (struct_v >= 5) {
1135       uint32_t it;
1136       ::decode(it, bl);
1137       index_type = (RGWBucketIndexType)it;
1138     }
1139     if (struct_v >= 6) {
1140       ::decode(compression_type, bl);
1141     }
1142     DECODE_FINISH(bl);
1143   }
1144   const rgw_pool& get_data_extra_pool() const {
1145     if (data_extra_pool.empty()) {
1146       return data_pool;
1147     }
1148     return data_extra_pool;
1149   }
1150   void dump(Formatter *f) const;
1151   void decode_json(JSONObj *obj);
1152 };
1153 WRITE_CLASS_ENCODER(RGWZonePlacementInfo)
1154
1155 struct RGWZoneParams : RGWSystemMetaObj {
1156   rgw_pool domain_root;
1157   rgw_pool metadata_heap;
1158   rgw_pool control_pool;
1159   rgw_pool gc_pool;
1160   rgw_pool lc_pool;
1161   rgw_pool log_pool;
1162   rgw_pool intent_log_pool;
1163   rgw_pool usage_log_pool;
1164
1165   rgw_pool user_keys_pool;
1166   rgw_pool user_email_pool;
1167   rgw_pool user_swift_pool;
1168   rgw_pool user_uid_pool;
1169   rgw_pool roles_pool;
1170   rgw_pool reshard_pool;
1171
1172   RGWAccessKey system_key;
1173
1174   map<string, RGWZonePlacementInfo> placement_pools;
1175
1176   string realm_id;
1177
1178   map<string, string, ltstr_nocase> tier_config;
1179
1180   RGWZoneParams() : RGWSystemMetaObj() {}
1181   RGWZoneParams(const string& name) : RGWSystemMetaObj(name){}
1182   RGWZoneParams(const string& id, const string& name) : RGWSystemMetaObj(id, name) {}
1183   RGWZoneParams(const string& id, const string& name, const string& _realm_id)
1184     : RGWSystemMetaObj(id, name), realm_id(_realm_id) {}
1185
1186   rgw_pool get_pool(CephContext *cct);
1187   const string get_default_oid(bool old_format = false) override;
1188   const string& get_names_oid_prefix() override;
1189   const string& get_info_oid_prefix(bool old_format = false) override;
1190   const string& get_predefined_name(CephContext *cct) override;
1191
1192   int init(CephContext *_cct, RGWRados *_store, bool setup_obj = true,
1193            bool old_format = false);
1194   using RGWSystemMetaObj::init;
1195   int read_default_id(string& default_id, bool old_format = false) override;
1196   int set_as_default(bool exclusive = false) override;
1197   int create_default(bool old_format = false);
1198   int create(bool exclusive = true) override;
1199   int fix_pool_names();
1200
1201   const string& get_compression_type(const string& placement_rule) const;
1202
1203   void encode(bufferlist& bl) const override {
1204     ENCODE_START(10, 1, bl);
1205     ::encode(domain_root, bl);
1206     ::encode(control_pool, bl);
1207     ::encode(gc_pool, bl);
1208     ::encode(log_pool, bl);
1209     ::encode(intent_log_pool, bl);
1210     ::encode(usage_log_pool, bl);
1211     ::encode(user_keys_pool, bl);
1212     ::encode(user_email_pool, bl);
1213     ::encode(user_swift_pool, bl);
1214     ::encode(user_uid_pool, bl);
1215     RGWSystemMetaObj::encode(bl);
1216     ::encode(system_key, bl);
1217     ::encode(placement_pools, bl);
1218     ::encode(metadata_heap, bl);
1219     ::encode(realm_id, bl);
1220     ::encode(lc_pool, bl);
1221     ::encode(tier_config, bl);
1222     ::encode(roles_pool, bl);
1223     ::encode(reshard_pool, bl);
1224     ENCODE_FINISH(bl);
1225   }
1226
1227   void decode(bufferlist::iterator& bl) override {
1228     DECODE_START(10, bl);
1229     ::decode(domain_root, bl);
1230     ::decode(control_pool, bl);
1231     ::decode(gc_pool, bl);
1232     ::decode(log_pool, bl);
1233     ::decode(intent_log_pool, bl);
1234     ::decode(usage_log_pool, bl);
1235     ::decode(user_keys_pool, bl);
1236     ::decode(user_email_pool, bl);
1237     ::decode(user_swift_pool, bl);
1238     ::decode(user_uid_pool, bl);
1239     if (struct_v >= 6) {
1240       RGWSystemMetaObj::decode(bl);
1241     } else if (struct_v >= 2) {
1242       ::decode(name, bl);
1243       id = name;
1244     }
1245     if (struct_v >= 3)
1246       ::decode(system_key, bl);
1247     if (struct_v >= 4)
1248       ::decode(placement_pools, bl);
1249     if (struct_v >= 5)
1250       ::decode(metadata_heap, bl);
1251     if (struct_v >= 6) {
1252       ::decode(realm_id, bl);
1253     }
1254     if (struct_v >= 7) {
1255       ::decode(lc_pool, bl);
1256     } else {
1257       lc_pool.init(name + ".rgw.lc");
1258     }
1259     if (struct_v >= 8) {
1260       ::decode(tier_config, bl);
1261     }
1262     if (struct_v >= 9) {
1263       ::decode(roles_pool, bl);
1264     } else {
1265       roles_pool = name + ".rgw.roles";
1266     }
1267     if (struct_v >= 10) {
1268       ::decode(reshard_pool, bl);
1269     } else {
1270       reshard_pool = log_pool.name + ":reshard";
1271     }
1272     DECODE_FINISH(bl);
1273   }
1274   void dump(Formatter *f) const;
1275   void decode_json(JSONObj *obj);
1276   static void generate_test_instances(list<RGWZoneParams*>& o);
1277
1278   bool get_placement(const string& placement_id, RGWZonePlacementInfo *placement) const {
1279     auto iter = placement_pools.find(placement_id);
1280     if (iter == placement_pools.end()) {
1281       return false;
1282     }
1283     *placement = iter->second;
1284     return true;
1285   }
1286
1287   /*
1288    * return data pool of the head object
1289    */
1290   bool get_head_data_pool(const string& placement_id, const rgw_obj& obj, rgw_pool *pool) const {
1291     const rgw_data_placement_target& explicit_placement = obj.bucket.explicit_placement;
1292     if (!explicit_placement.data_pool.empty()) {
1293       if (!obj.in_extra_data) {
1294         *pool = explicit_placement.data_pool;
1295       } else {
1296         *pool = explicit_placement.get_data_extra_pool();
1297       }
1298       return true;
1299     }
1300     if (placement_id.empty()) {
1301       return false;
1302     }
1303     auto iter = placement_pools.find(placement_id);
1304     if (iter == placement_pools.end()) {
1305       return false;
1306     }
1307     if (!obj.in_extra_data) {
1308       *pool = iter->second.data_pool;
1309     } else {
1310       *pool = iter->second.get_data_extra_pool();
1311     }
1312     return true;
1313   }
1314 };
1315 WRITE_CLASS_ENCODER(RGWZoneParams)
1316
1317 struct RGWZone {
1318   string id;
1319   string name;
1320   list<string> endpoints;
1321   bool log_meta;
1322   bool log_data;
1323   bool read_only;
1324   string tier_type;
1325
1326 /**
1327  * Represents the number of shards for the bucket index object, a value of zero
1328  * indicates there is no sharding. By default (no sharding, the name of the object
1329  * is '.dir.{marker}', with sharding, the name is '.dir.{marker}.{sharding_id}',
1330  * sharding_id is zero-based value. It is not recommended to set a too large value
1331  * (e.g. thousand) as it increases the cost for bucket listing.
1332  */
1333   uint32_t bucket_index_max_shards;
1334
1335   bool sync_from_all;
1336   set<string> sync_from; /* list of zones to sync from */
1337
1338   RGWZone() : log_meta(false), log_data(false), read_only(false), bucket_index_max_shards(0),
1339               sync_from_all(true) {}
1340
1341   void encode(bufferlist& bl) const {
1342     ENCODE_START(6, 1, bl);
1343     ::encode(name, bl);
1344     ::encode(endpoints, bl);
1345     ::encode(log_meta, bl);
1346     ::encode(log_data, bl);
1347     ::encode(bucket_index_max_shards, bl);
1348     ::encode(id, bl);
1349     ::encode(read_only, bl);
1350     ::encode(tier_type, bl);
1351     ::encode(sync_from_all, bl);
1352     ::encode(sync_from, bl);
1353     ENCODE_FINISH(bl);
1354   }
1355
1356   void decode(bufferlist::iterator& bl) {
1357     DECODE_START(6, bl);
1358     ::decode(name, bl);
1359     if (struct_v < 4) {
1360       id = name;
1361     }
1362     ::decode(endpoints, bl);
1363     if (struct_v >= 2) {
1364       ::decode(log_meta, bl);
1365       ::decode(log_data, bl);
1366     }
1367     if (struct_v >= 3) {
1368       ::decode(bucket_index_max_shards, bl);
1369     }
1370     if (struct_v >= 4) {
1371       ::decode(id, bl);
1372       ::decode(read_only, bl);
1373     }
1374     if (struct_v >= 5) {
1375       ::decode(tier_type, bl);
1376     }
1377     if (struct_v >= 6) {
1378       ::decode(sync_from_all, bl);
1379       ::decode(sync_from, bl);
1380     }
1381     DECODE_FINISH(bl);
1382   }
1383   void dump(Formatter *f) const;
1384   void decode_json(JSONObj *obj);
1385   static void generate_test_instances(list<RGWZone*>& o);
1386
1387   bool is_read_only() { return read_only; }
1388
1389   bool syncs_from(const string& zone_id) {
1390     return (sync_from_all || sync_from.find(zone_id) != sync_from.end());
1391   }
1392 };
1393 WRITE_CLASS_ENCODER(RGWZone)
1394
1395 struct RGWDefaultZoneGroupInfo {
1396   string default_zonegroup;
1397
1398   void encode(bufferlist& bl) const {
1399     ENCODE_START(1, 1, bl);
1400     ::encode(default_zonegroup, bl);
1401     ENCODE_FINISH(bl);
1402   }
1403
1404   void decode(bufferlist::iterator& bl) {
1405     DECODE_START(1, bl);
1406     ::decode(default_zonegroup, bl);
1407     DECODE_FINISH(bl);
1408   }
1409   void dump(Formatter *f) const;
1410   void decode_json(JSONObj *obj);
1411   //todo: implement ceph-dencoder
1412 };
1413 WRITE_CLASS_ENCODER(RGWDefaultZoneGroupInfo)
1414
1415 struct RGWZoneGroupPlacementTarget {
1416   string name;
1417   set<string> tags;
1418
1419   bool user_permitted(list<string>& user_tags) const {
1420     if (tags.empty()) {
1421       return true;
1422     }
1423     for (auto& rule : user_tags) {
1424       if (tags.find(rule) != tags.end()) {
1425         return true;
1426       }
1427     }
1428     return false;
1429   }
1430
1431   void encode(bufferlist& bl) const {
1432     ENCODE_START(1, 1, bl);
1433     ::encode(name, bl);
1434     ::encode(tags, bl);
1435     ENCODE_FINISH(bl);
1436   }
1437
1438   void decode(bufferlist::iterator& bl) {
1439     DECODE_START(1, bl);
1440     ::decode(name, bl);
1441     ::decode(tags, bl);
1442     DECODE_FINISH(bl);
1443   }
1444   void dump(Formatter *f) const;
1445   void decode_json(JSONObj *obj);
1446 };
1447 WRITE_CLASS_ENCODER(RGWZoneGroupPlacementTarget)
1448
1449
1450 struct RGWZoneGroup : public RGWSystemMetaObj {
1451   string api_name;
1452   list<string> endpoints;
1453   bool is_master;
1454
1455   string master_zone;
1456   map<string, RGWZone> zones;
1457
1458   map<string, RGWZoneGroupPlacementTarget> placement_targets;
1459   string default_placement;
1460
1461   list<string> hostnames;
1462   list<string> hostnames_s3website;
1463   // TODO: Maybe convert hostnames to a map<string,list<string>> for
1464   // endpoint_type->hostnames
1465 /*
1466 20:05 < _robbat21irssi> maybe I do someting like: if (hostname_map.empty()) { populate all map keys from hostnames; };
1467 20:05 < _robbat21irssi> but that's a later compatability migration planning bit
1468 20:06 < yehudasa> more like if (!hostnames.empty()) {
1469 20:06 < yehudasa> for (list<string>::iterator iter = hostnames.begin(); iter != hostnames.end(); ++iter) {
1470 20:06 < yehudasa> hostname_map["s3"].append(iter->second);
1471 20:07 < yehudasa> hostname_map["s3website"].append(iter->second);
1472 20:07 < yehudasa> s/append/push_back/g
1473 20:08 < _robbat21irssi> inner loop over APIs
1474 20:08 < yehudasa> yeah, probably
1475 20:08 < _robbat21irssi> s3, s3website, swift, swith_auth, swift_website
1476 */
1477   map<string, list<string> > api_hostname_map;
1478   map<string, list<string> > api_endpoints_map;
1479
1480   string realm_id;
1481
1482   RGWZoneGroup(): is_master(false){}
1483   RGWZoneGroup(const std::string &id, const std::string &name):RGWSystemMetaObj(id, name) {}
1484   RGWZoneGroup(const std::string &_name):RGWSystemMetaObj(_name) {}
1485   RGWZoneGroup(const std::string &_name, bool _is_master, CephContext *cct, RGWRados* store,
1486                const string& _realm_id, const list<string>& _endpoints)
1487     : RGWSystemMetaObj(_name, cct , store), endpoints(_endpoints), is_master(_is_master),
1488       realm_id(_realm_id) {}
1489
1490   bool is_master_zonegroup() const { return is_master;}
1491   void update_master(bool _is_master) {
1492     is_master = _is_master;
1493     post_process_params();
1494   }
1495   void post_process_params();
1496
1497   void encode(bufferlist& bl) const override {
1498     ENCODE_START(4, 1, bl);
1499     ::encode(name, bl);
1500     ::encode(api_name, bl);
1501     ::encode(is_master, bl);
1502     ::encode(endpoints, bl);
1503     ::encode(master_zone, bl);
1504     ::encode(zones, bl);
1505     ::encode(placement_targets, bl);
1506     ::encode(default_placement, bl);
1507     ::encode(hostnames, bl);
1508     ::encode(hostnames_s3website, bl);
1509     RGWSystemMetaObj::encode(bl);
1510     ::encode(realm_id, bl);
1511     ENCODE_FINISH(bl);
1512   }
1513
1514   void decode(bufferlist::iterator& bl) override {
1515     DECODE_START(4, bl);
1516     ::decode(name, bl);
1517     ::decode(api_name, bl);
1518     ::decode(is_master, bl);
1519     ::decode(endpoints, bl);
1520     ::decode(master_zone, bl);
1521     ::decode(zones, bl);
1522     ::decode(placement_targets, bl);
1523     ::decode(default_placement, bl);
1524     if (struct_v >= 2) {
1525       ::decode(hostnames, bl);
1526     }
1527     if (struct_v >= 3) {
1528       ::decode(hostnames_s3website, bl);
1529     }
1530     if (struct_v >= 4) {
1531       RGWSystemMetaObj::decode(bl);
1532       ::decode(realm_id, bl);
1533     } else {
1534       id = name;
1535     }
1536     DECODE_FINISH(bl);
1537   }
1538
1539   int read_default_id(string& default_id, bool old_format = false) override;
1540   int set_as_default(bool exclusive = false) override;
1541   int create_default(bool old_format = false);
1542   int equals(const string& other_zonegroup) const;
1543   int add_zone(const RGWZoneParams& zone_params, bool *is_master, bool *read_only,
1544                const list<string>& endpoints, const string *ptier_type,
1545                bool *psync_from_all, list<string>& sync_from, list<string>& sync_from_rm);
1546   int remove_zone(const std::string& zone_id);
1547   int rename_zone(const RGWZoneParams& zone_params);
1548   rgw_pool get_pool(CephContext *cct);
1549   const string get_default_oid(bool old_region_format = false) override;
1550   const string& get_info_oid_prefix(bool old_region_format = false) override;
1551   const string& get_names_oid_prefix() override;
1552   const string& get_predefined_name(CephContext *cct) override;
1553
1554   void dump(Formatter *f) const;
1555   void decode_json(JSONObj *obj);
1556   static void generate_test_instances(list<RGWZoneGroup*>& o);
1557 };
1558 WRITE_CLASS_ENCODER(RGWZoneGroup)
1559
1560 struct RGWPeriodMap
1561 {
1562   string id;
1563   map<string, RGWZoneGroup> zonegroups;
1564   map<string, RGWZoneGroup> zonegroups_by_api;
1565   map<string, uint32_t> short_zone_ids;
1566
1567   string master_zonegroup;
1568
1569   void encode(bufferlist& bl) const;
1570   void decode(bufferlist::iterator& bl);
1571
1572   int update(const RGWZoneGroup& zonegroup, CephContext *cct);
1573
1574   void dump(Formatter *f) const;
1575   void decode_json(JSONObj *obj);
1576
1577   void reset() {
1578     zonegroups.clear();
1579     zonegroups_by_api.clear();
1580     master_zonegroup.clear();
1581   }
1582
1583   uint32_t get_zone_short_id(const string& zone_id) const;
1584 };
1585 WRITE_CLASS_ENCODER(RGWPeriodMap)
1586
1587 struct RGWPeriodConfig
1588 {
1589   RGWQuotaInfo bucket_quota;
1590   RGWQuotaInfo user_quota;
1591
1592   void encode(bufferlist& bl) const {
1593     ENCODE_START(1, 1, bl);
1594     ::encode(bucket_quota, bl);
1595     ::encode(user_quota, bl);
1596     ENCODE_FINISH(bl);
1597   }
1598
1599   void decode(bufferlist::iterator& bl) {
1600     DECODE_START(1, bl);
1601     ::decode(bucket_quota, bl);
1602     ::decode(user_quota, bl);
1603     DECODE_FINISH(bl);
1604   }
1605
1606   void dump(Formatter *f) const;
1607   void decode_json(JSONObj *obj);
1608
1609   // the period config must be stored in a local object outside of the period,
1610   // so that it can be used in a default configuration where no realm/period
1611   // exists
1612   int read(RGWRados *store, const std::string& realm_id);
1613   int write(RGWRados *store, const std::string& realm_id);
1614
1615   static std::string get_oid(const std::string& realm_id);
1616   static rgw_pool get_pool(CephContext *cct);
1617 };
1618 WRITE_CLASS_ENCODER(RGWPeriodConfig)
1619
1620 /* for backward comaptability */
1621 struct RGWRegionMap {
1622
1623   map<string, RGWZoneGroup> regions;
1624
1625   string master_region;
1626
1627   RGWQuotaInfo bucket_quota;
1628   RGWQuotaInfo user_quota;
1629
1630   void encode(bufferlist& bl) const;
1631   void decode(bufferlist::iterator& bl);
1632
1633   void dump(Formatter *f) const;
1634   void decode_json(JSONObj *obj);
1635 };
1636 WRITE_CLASS_ENCODER(RGWRegionMap)
1637
1638 struct RGWZoneGroupMap {
1639
1640   map<string, RGWZoneGroup> zonegroups;
1641   map<string, RGWZoneGroup> zonegroups_by_api;
1642
1643   string master_zonegroup;
1644
1645   RGWQuotaInfo bucket_quota;
1646   RGWQuotaInfo user_quota;
1647
1648   /* constract the map */
1649   int read(CephContext *cct, RGWRados *store);
1650
1651   void encode(bufferlist& bl) const;
1652   void decode(bufferlist::iterator& bl);
1653
1654   void dump(Formatter *f) const;
1655   void decode_json(JSONObj *obj);
1656 };
1657 WRITE_CLASS_ENCODER(RGWZoneGroupMap)
1658
1659 class RGWRealm;
1660
1661 struct objexp_hint_entry {
1662   string tenant;
1663   string bucket_name;
1664   string bucket_id;
1665   rgw_obj_key obj_key;
1666   ceph::real_time exp_time;
1667
1668   void encode(bufferlist& bl) const {
1669     ENCODE_START(2, 1, bl);
1670     ::encode(bucket_name, bl);
1671     ::encode(bucket_id, bl);
1672     ::encode(obj_key, bl);
1673     ::encode(exp_time, bl);
1674     ::encode(tenant, bl);
1675     ENCODE_FINISH(bl);
1676   }
1677
1678   void decode(bufferlist::iterator& bl) {
1679     // XXX Do we want DECODE_START_LEGACY_COMPAT_LEN(2, 1, 1, bl); ?
1680     DECODE_START(2, bl);
1681     ::decode(bucket_name, bl);
1682     ::decode(bucket_id, bl);
1683     ::decode(obj_key, bl);
1684     ::decode(exp_time, bl);
1685     if (struct_v >= 2) {
1686       ::decode(tenant, bl);
1687     } else {
1688       tenant.clear();
1689     }
1690     DECODE_FINISH(bl);
1691   }
1692 };
1693 WRITE_CLASS_ENCODER(objexp_hint_entry)
1694
1695 class RGWPeriod;
1696
1697 class RGWRealm : public RGWSystemMetaObj
1698 {
1699   string current_period;
1700   epoch_t epoch{0}; //< realm epoch, incremented for each new period
1701
1702   int create_control(bool exclusive);
1703   int delete_control();
1704 public:
1705   RGWRealm() {}
1706   RGWRealm(const string& _id, const string& _name = "") : RGWSystemMetaObj(_id, _name) {}
1707   RGWRealm(CephContext *_cct, RGWRados *_store): RGWSystemMetaObj(_cct, _store) {}
1708   RGWRealm(const string& _name, CephContext *_cct, RGWRados *_store): RGWSystemMetaObj(_name, _cct, _store){}
1709
1710   void encode(bufferlist& bl) const override {
1711     ENCODE_START(1, 1, bl);
1712     RGWSystemMetaObj::encode(bl);
1713     ::encode(current_period, bl);
1714     ::encode(epoch, bl);
1715     ENCODE_FINISH(bl);
1716   }
1717
1718   void decode(bufferlist::iterator& bl) override {
1719     DECODE_START(1, bl);
1720     RGWSystemMetaObj::decode(bl);
1721     ::decode(current_period, bl);
1722     ::decode(epoch, bl);
1723     DECODE_FINISH(bl);
1724   }
1725
1726   int create(bool exclusive = true) override;
1727   int delete_obj();
1728   rgw_pool get_pool(CephContext *cct);
1729   const string get_default_oid(bool old_format = false) override;
1730   const string& get_names_oid_prefix() override;
1731   const string& get_info_oid_prefix(bool old_format = false) override;
1732   const string& get_predefined_name(CephContext *cct) override;
1733
1734   using RGWSystemMetaObj::read_id; // expose as public for radosgw-admin
1735
1736   void dump(Formatter *f) const;
1737   void decode_json(JSONObj *obj);
1738
1739   const string& get_current_period() const {
1740     return current_period;
1741   }
1742   int set_current_period(RGWPeriod& period);
1743   void clear_current_period_and_epoch() {
1744     current_period.clear();
1745     epoch = 0;
1746   }
1747   epoch_t get_epoch() const { return epoch; }
1748
1749   string get_control_oid();
1750   /// send a notify on the realm control object
1751   int notify_zone(bufferlist& bl);
1752   /// notify the zone of a new period
1753   int notify_new_period(const RGWPeriod& period);
1754 };
1755 WRITE_CLASS_ENCODER(RGWRealm)
1756
1757 struct RGWPeriodLatestEpochInfo {
1758   epoch_t epoch;
1759
1760   void encode(bufferlist& bl) const {
1761     ENCODE_START(1, 1, bl);
1762     ::encode(epoch, bl);
1763     ENCODE_FINISH(bl);
1764   }
1765
1766   void decode(bufferlist::iterator& bl) {
1767     DECODE_START(1, bl);
1768     ::decode(epoch, bl);
1769     DECODE_FINISH(bl);
1770   }
1771
1772   void dump(Formatter *f) const;
1773   void decode_json(JSONObj *obj);
1774 };
1775 WRITE_CLASS_ENCODER(RGWPeriodLatestEpochInfo)
1776
1777 class RGWPeriod
1778 {
1779   string id;
1780   epoch_t epoch;
1781   string predecessor_uuid;
1782   std::vector<std::string> sync_status;
1783   RGWPeriodMap period_map;
1784   RGWPeriodConfig period_config;
1785   string master_zonegroup;
1786   string master_zone;
1787
1788   string realm_id;
1789   string realm_name;
1790   epoch_t realm_epoch{1}; //< realm epoch when period was made current
1791
1792   CephContext *cct;
1793   RGWRados *store;
1794
1795   int read_info();
1796   int read_latest_epoch(RGWPeriodLatestEpochInfo& epoch_info,
1797                         RGWObjVersionTracker *objv = nullptr);
1798   int use_latest_epoch();
1799   int use_current_period();
1800
1801   const string get_period_oid();
1802   const string get_period_oid_prefix();
1803
1804   // gather the metadata sync status for each shard; only for use on master zone
1805   int update_sync_status(const RGWPeriod &current_period,
1806                          std::ostream& error_stream, bool force_if_stale);
1807
1808 public:
1809   RGWPeriod() : epoch(0), cct(NULL), store(NULL) {}
1810
1811   RGWPeriod(const string& period_id, epoch_t _epoch = 0)
1812     : id(period_id), epoch(_epoch),
1813       cct(NULL), store(NULL) {}
1814
1815   const string& get_id() const { return id; }
1816   epoch_t get_epoch() const { return epoch; }
1817   epoch_t get_realm_epoch() const { return realm_epoch; }
1818   const string& get_predecessor() const { return predecessor_uuid; }
1819   const string& get_master_zone() const { return master_zone; }
1820   const string& get_master_zonegroup() const { return master_zonegroup; }
1821   const string& get_realm() const { return realm_id; }
1822   const RGWPeriodMap& get_map() const { return period_map; }
1823   RGWPeriodConfig& get_config() { return period_config; }
1824   const RGWPeriodConfig& get_config() const { return period_config; }
1825   const std::vector<std::string>& get_sync_status() const { return sync_status; }
1826   rgw_pool get_pool(CephContext *cct);
1827   const string& get_latest_epoch_oid();
1828   const string& get_info_oid_prefix();
1829
1830   void set_user_quota(RGWQuotaInfo& user_quota) {
1831     period_config.user_quota = user_quota;
1832   }
1833
1834   void set_bucket_quota(RGWQuotaInfo& bucket_quota) {
1835     period_config.bucket_quota = bucket_quota;
1836   }
1837
1838   void set_id(const string& id) {
1839     this->id = id;
1840     period_map.id = id;
1841   }
1842   void set_epoch(epoch_t epoch) { this->epoch = epoch; }
1843   void set_realm_epoch(epoch_t epoch) { realm_epoch = epoch; }
1844
1845   void set_predecessor(const string& predecessor)
1846   {
1847     predecessor_uuid = predecessor;
1848   }
1849
1850   void set_realm_id(const string& _realm_id) {
1851     realm_id = _realm_id;
1852   }
1853
1854   int reflect();
1855
1856   int get_zonegroup(RGWZoneGroup& zonegroup,
1857                     const string& zonegroup_id);
1858
1859   bool is_single_zonegroup()
1860   {
1861       return (period_map.zonegroups.size() == 1);
1862   }
1863
1864   /*
1865     returns true if there are several zone groups with a least one zone
1866    */
1867   bool is_multi_zonegroups_with_zones()
1868   {
1869     int count = 0;
1870     for (const auto& zg:  period_map.zonegroups) {
1871       if (zg.second.zones.size() > 0) {
1872         if (count++ > 0) {
1873           return true;
1874         }
1875       }
1876     }
1877     return false;
1878   }
1879
1880   int get_latest_epoch(epoch_t& epoch);
1881   int set_latest_epoch(epoch_t epoch, bool exclusive = false,
1882                        RGWObjVersionTracker *objv = nullptr);
1883   // update latest_epoch if the given epoch is higher, else return -EEXIST
1884   int update_latest_epoch(epoch_t epoch);
1885
1886   int init(CephContext *_cct, RGWRados *_store, const string &period_realm_id, const string &period_realm_name = "",
1887            bool setup_obj = true);
1888   int init(CephContext *_cct, RGWRados *_store, bool setup_obj = true);
1889
1890   int create(bool exclusive = true);
1891   int delete_obj();
1892   int store_info(bool exclusive);
1893   int add_zonegroup(const RGWZoneGroup& zonegroup);
1894
1895   void fork();
1896   int update();
1897
1898   // commit a staging period; only for use on master zone
1899   int commit(RGWRealm& realm, const RGWPeriod &current_period,
1900              std::ostream& error_stream, bool force_if_stale = false);
1901
1902   void encode(bufferlist& bl) const {
1903     ENCODE_START(1, 1, bl);
1904     ::encode(id, bl);
1905     ::encode(epoch, bl);
1906     ::encode(realm_epoch, bl);
1907     ::encode(predecessor_uuid, bl);
1908     ::encode(sync_status, bl);
1909     ::encode(period_map, bl);
1910     ::encode(master_zone, bl);
1911     ::encode(master_zonegroup, bl);
1912     ::encode(period_config, bl);
1913     ::encode(realm_id, bl);
1914     ::encode(realm_name, bl);
1915     ENCODE_FINISH(bl);
1916   }
1917
1918   void decode(bufferlist::iterator& bl) {
1919     DECODE_START(1, bl);
1920     ::decode(id, bl);
1921     ::decode(epoch, bl);
1922     ::decode(realm_epoch, bl);
1923     ::decode(predecessor_uuid, bl);
1924     ::decode(sync_status, bl);
1925     ::decode(period_map, bl);
1926     ::decode(master_zone, bl);
1927     ::decode(master_zonegroup, bl);
1928     ::decode(period_config, bl);
1929     ::decode(realm_id, bl);
1930     ::decode(realm_name, bl);
1931     DECODE_FINISH(bl);
1932   }
1933   void dump(Formatter *f) const;
1934   void decode_json(JSONObj *obj);
1935
1936   static string get_staging_id(const string& realm_id) {
1937     return realm_id + ":staging";
1938   }
1939 };
1940 WRITE_CLASS_ENCODER(RGWPeriod)
1941
1942 class RGWDataChangesLog;
1943 class RGWMetaSyncStatusManager;
1944 class RGWDataSyncStatusManager;
1945 class RGWReplicaLogger;
1946 class RGWCoroutinesManagerRegistry;
1947
1948 class RGWStateLog {
1949   RGWRados *store;
1950   int num_shards;
1951   string module_name;
1952
1953   void oid_str(int shard, string& oid);
1954   int get_shard_num(const string& object);
1955   string get_oid(const string& object);
1956   int open_ioctx(librados::IoCtx& ioctx);
1957
1958   struct list_state {
1959     int cur_shard;
1960     int max_shard;
1961     string marker;
1962     string client_id;
1963     string op_id;
1964     string object;
1965
1966     list_state() : cur_shard(0), max_shard(0) {}
1967   };
1968
1969 protected:
1970   virtual bool dump_entry_internal(const cls_statelog_entry& entry, Formatter *f) {
1971     return false;
1972   }
1973
1974 public:
1975   RGWStateLog(RGWRados *_store, int _num_shards, const string& _module_name) :
1976               store(_store), num_shards(_num_shards), module_name(_module_name) {}
1977   virtual ~RGWStateLog() {}
1978
1979   int store_entry(const string& client_id, const string& op_id, const string& object,
1980                   uint32_t state, bufferlist *bl, uint32_t *check_state);
1981
1982   int remove_entry(const string& client_id, const string& op_id, const string& object);
1983
1984   void init_list_entries(const string& client_id, const string& op_id, const string& object,
1985                          void **handle);
1986
1987   int list_entries(void *handle, int max_entries, list<cls_statelog_entry>& entries, bool *done);
1988
1989   void finish_list_entries(void *handle);
1990
1991   virtual void dump_entry(const cls_statelog_entry& entry, Formatter *f);
1992 };
1993
1994 /*
1995  * state transitions:
1996  *
1997  * unknown -> in-progress -> complete
1998  *                        -> error
1999  *
2000  * user can try setting the 'abort' state, and it can only succeed if state is
2001  * in-progress.
2002  *
2003  * state renewal cannot switch state (stays in the same state)
2004  *
2005  * rgw can switch from in-progress to complete
2006  * rgw can switch from in-progress to error
2007  *
2008  * rgw can switch from abort to cancelled
2009  *
2010  */
2011
2012 class RGWOpState : public RGWStateLog {
2013 protected:
2014   bool dump_entry_internal(const cls_statelog_entry& entry, Formatter *f) override;
2015 public:
2016
2017   enum OpState {
2018     OPSTATE_UNKNOWN     = 0,
2019     OPSTATE_IN_PROGRESS = 1,
2020     OPSTATE_COMPLETE    = 2,
2021     OPSTATE_ERROR       = 3,
2022     OPSTATE_ABORT       = 4,
2023     OPSTATE_CANCELLED   = 5,
2024   };
2025
2026   explicit RGWOpState(RGWRados *_store);
2027
2028   int state_from_str(const string& s, OpState *state);
2029   int set_state(const string& client_id, const string& op_id, const string& object, OpState state);
2030   int renew_state(const string& client_id, const string& op_id, const string& object, OpState state);
2031 };
2032
2033 class RGWOpStateSingleOp
2034 {
2035   RGWOpState os;
2036   string client_id;
2037   string op_id;
2038   string object;
2039
2040   CephContext *cct;
2041
2042   RGWOpState::OpState cur_state;
2043   ceph::real_time last_update;
2044
2045 public:
2046   RGWOpStateSingleOp(RGWRados *store, const string& cid, const string& oid, const string& obj);
2047
2048   int set_state(RGWOpState::OpState state);
2049   int renew_state();
2050 };
2051
2052 class RGWGetBucketStats_CB : public RefCountedObject {
2053 protected:
2054   rgw_bucket bucket;
2055   map<RGWObjCategory, RGWStorageStats> *stats;
2056 public:
2057   explicit RGWGetBucketStats_CB(const rgw_bucket& _bucket) : bucket(_bucket), stats(NULL) {}
2058   ~RGWGetBucketStats_CB() override {}
2059   virtual void handle_response(int r) = 0;
2060   virtual void set_response(map<RGWObjCategory, RGWStorageStats> *_stats) {
2061     stats = _stats;
2062   }
2063 };
2064
2065 class RGWGetUserStats_CB : public RefCountedObject {
2066 protected:
2067   rgw_user user;
2068   RGWStorageStats stats;
2069 public:
2070   explicit RGWGetUserStats_CB(const rgw_user& _user) : user(_user) {}
2071   ~RGWGetUserStats_CB() override {}
2072   virtual void handle_response(int r) = 0;
2073   virtual void set_response(RGWStorageStats& _stats) {
2074     stats = _stats;
2075   }
2076 };
2077
2078 class RGWGetDirHeader_CB;
2079 class RGWGetUserHeader_CB;
2080
2081 struct rgw_rados_ref {
2082   rgw_pool pool;
2083   string oid;
2084   string key;
2085   librados::IoCtx ioctx;
2086 };
2087
2088 class RGWChainedCache {
2089 public:
2090   virtual ~RGWChainedCache() {}
2091   virtual void chain_cb(const string& key, void *data) = 0;
2092   virtual void invalidate(const string& key) = 0;
2093   virtual void invalidate_all() = 0;
2094
2095   struct Entry {
2096     RGWChainedCache *cache;
2097     const string& key;
2098     void *data;
2099
2100     Entry(RGWChainedCache *_c, const string& _k, void *_d) : cache(_c), key(_k), data(_d) {}
2101   };
2102 };
2103
2104 template <class T, class S>
2105 class RGWObjectCtxImpl {
2106   RGWRados *store;
2107   std::map<T, S> objs_state;
2108   RWLock lock;
2109
2110 public:
2111   RGWObjectCtxImpl(RGWRados *_store) : store(_store), lock("RGWObjectCtxImpl") {}
2112
2113   S *get_state(const T& obj) {
2114     S *result;
2115     typename std::map<T, S>::iterator iter;
2116     lock.get_read();
2117     assert (!obj.empty());
2118     iter = objs_state.find(obj);
2119     if (iter != objs_state.end()) {
2120       result = &iter->second;
2121       lock.unlock();
2122     } else {
2123       lock.unlock();
2124       lock.get_write();
2125       result = &objs_state[obj];
2126       lock.unlock();
2127     }
2128     return result;
2129   }
2130
2131   void set_atomic(T& obj) {
2132     RWLock::WLocker wl(lock);
2133     assert (!obj.empty());
2134     objs_state[obj].is_atomic = true;
2135   }
2136   void set_prefetch_data(T& obj) {
2137     RWLock::WLocker wl(lock);
2138     assert (!obj.empty());
2139     objs_state[obj].prefetch_data = true;
2140   }
2141   void invalidate(T& obj) {
2142     RWLock::WLocker wl(lock);
2143     auto iter = objs_state.find(obj);
2144     if (iter == objs_state.end()) {
2145       return;
2146     }
2147     bool is_atomic = iter->second.is_atomic;
2148     bool prefetch_data = iter->second.prefetch_data;
2149
2150     objs_state.erase(iter);
2151
2152     if (is_atomic || prefetch_data) {
2153       auto& s = objs_state[obj];
2154       s.is_atomic = is_atomic;
2155       s.prefetch_data = prefetch_data;
2156     }
2157   }
2158 };
2159
2160 template<>
2161 void RGWObjectCtxImpl<rgw_obj, RGWObjState>::invalidate(rgw_obj& obj);
2162
2163 template<>
2164 void RGWObjectCtxImpl<rgw_raw_obj, RGWRawObjState>::invalidate(rgw_raw_obj& obj);
2165
2166 struct RGWObjectCtx {
2167   RGWRados *store;
2168   void *user_ctx;
2169
2170   RGWObjectCtxImpl<rgw_obj, RGWObjState> obj;
2171   RGWObjectCtxImpl<rgw_raw_obj, RGWRawObjState> raw;
2172
2173   explicit RGWObjectCtx(RGWRados *_store) : store(_store), user_ctx(NULL), obj(store), raw(store) { }
2174   RGWObjectCtx(RGWRados *_store, void *_user_ctx) : store(_store), user_ctx(_user_ctx), obj(store), raw(store) { }
2175 };
2176
2177 class Finisher;
2178 class RGWAsyncRadosProcessor;
2179
2180 template <class T>
2181 class RGWChainedCacheImpl;
2182
2183 struct bucket_info_entry {
2184   RGWBucketInfo info;
2185   real_time mtime;
2186   map<string, bufferlist> attrs;
2187 };
2188
2189 struct tombstone_entry {
2190   ceph::real_time mtime;
2191   uint32_t zone_short_id;
2192   uint64_t pg_ver;
2193
2194   tombstone_entry() = default;
2195   tombstone_entry(const RGWObjState& state)
2196     : mtime(state.mtime), zone_short_id(state.zone_short_id),
2197       pg_ver(state.pg_ver) {}
2198 };
2199
2200 class RGWIndexCompletionManager;
2201
2202 class RGWRados
2203 {
2204   friend class RGWGC;
2205   friend class RGWMetaNotifier;
2206   friend class RGWDataNotifier;
2207   friend class RGWLC;
2208   friend class RGWObjectExpirer;
2209   friend class RGWMetaSyncProcessorThread;
2210   friend class RGWDataSyncProcessorThread;
2211   friend class RGWStateLog;
2212   friend class RGWReplicaLogger;
2213   friend class RGWReshard;
2214   friend class RGWBucketReshard;
2215   friend class BucketIndexLockGuard;
2216
2217   /** Open the pool used as root for this gateway */
2218   int open_root_pool_ctx();
2219   int open_gc_pool_ctx();
2220   int open_lc_pool_ctx();
2221   int open_objexp_pool_ctx();
2222   int open_reshard_pool_ctx();
2223
2224   int open_pool_ctx(const rgw_pool& pool, librados::IoCtx&  io_ctx);
2225   int open_bucket_index_ctx(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx);
2226   int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx&  index_ctx, string& bucket_oid);
2227   int open_bucket_index_base(const RGWBucketInfo& bucket_info, librados::IoCtx&  index_ctx,
2228       string& bucket_oid_base);
2229   int open_bucket_index_shard(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx,
2230       const string& obj_key, string *bucket_obj, int *shard_id);
2231   int open_bucket_index_shard(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx,
2232                               int shard_id, string *bucket_obj);
2233   int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx,
2234       map<int, string>& bucket_objs, int shard_id = -1, map<int, string> *bucket_instance_ids = NULL);
2235   template<typename T>
2236   int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx,
2237                         map<int, string>& oids, map<int, T>& bucket_objs,
2238                         int shard_id = -1, map<int, string> *bucket_instance_ids = NULL);
2239   void build_bucket_index_marker(const string& shard_id_str, const string& shard_marker,
2240       string *marker);
2241
2242   void get_bucket_instance_ids(const RGWBucketInfo& bucket_info, int shard_id, map<int, string> *result);
2243
2244   std::atomic<int64_t> max_req_id = { 0 };
2245   Mutex lock;
2246   Mutex watchers_lock;
2247   SafeTimer *timer;
2248
2249   RGWGC *gc;
2250   RGWLC *lc;
2251   RGWObjectExpirer *obj_expirer;
2252   bool use_gc_thread;
2253   bool use_lc_thread;
2254   bool quota_threads;
2255   bool run_sync_thread;
2256   bool run_reshard_thread;
2257
2258   RGWAsyncRadosProcessor* async_rados;
2259
2260   RGWMetaNotifier *meta_notifier;
2261   RGWDataNotifier *data_notifier;
2262   RGWMetaSyncProcessorThread *meta_sync_processor_thread;
2263   map<string, RGWDataSyncProcessorThread *> data_sync_processor_threads;
2264
2265   RGWSyncLogTrimThread *sync_log_trimmer{nullptr};
2266
2267   Mutex meta_sync_thread_lock;
2268   Mutex data_sync_thread_lock;
2269
2270   int num_watchers;
2271   RGWWatcher **watchers;
2272   std::set<int> watchers_set;
2273   librados::IoCtx root_pool_ctx;      // .rgw
2274   librados::IoCtx control_pool_ctx;   // .rgw.control
2275   bool watch_initialized;
2276
2277   friend class RGWWatcher;
2278
2279   Mutex bucket_id_lock;
2280
2281   // This field represents the number of bucket index object shards
2282   uint32_t bucket_index_max_shards;
2283
2284   int get_obj_head_ioctx(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::IoCtx *ioctx);
2285   int get_obj_head_ref(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_rados_ref *ref);
2286   int get_system_obj_ref(const rgw_raw_obj& obj, rgw_rados_ref *ref);
2287   uint64_t max_bucket_id;
2288
2289   int get_olh_target_state(RGWObjectCtx& rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj,
2290                            RGWObjState *olh_state, RGWObjState **target_state);
2291   int get_system_obj_state_impl(RGWObjectCtx *rctx, rgw_raw_obj& obj, RGWRawObjState **state, RGWObjVersionTracker *objv_tracker);
2292   int get_obj_state_impl(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state,
2293                          bool follow_olh, bool assume_noent = false);
2294   int append_atomic_test(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj,
2295                          librados::ObjectOperation& op, RGWObjState **state);
2296
2297   int update_placement_map();
2298   int store_bucket_info(RGWBucketInfo& info, map<string, bufferlist> *pattrs, RGWObjVersionTracker *objv_tracker, bool exclusive);
2299
2300   void remove_rgw_head_obj(librados::ObjectWriteOperation& op);
2301   void cls_obj_check_prefix_exist(librados::ObjectOperation& op, const string& prefix, bool fail_if_exist);
2302   void cls_obj_check_mtime(librados::ObjectOperation& op, const real_time& mtime, bool high_precision_time, RGWCheckMTimeType type);
2303 protected:
2304   CephContext *cct;
2305
2306   std::vector<librados::Rados> rados;
2307   uint32_t next_rados_handle;
2308   RWLock handle_lock;
2309   std::map<pthread_t, int> rados_map;
2310
2311   using RGWChainedCacheImpl_bucket_info_entry = RGWChainedCacheImpl<bucket_info_entry>;
2312   RGWChainedCacheImpl_bucket_info_entry *binfo_cache;
2313
2314   using tombstone_cache_t = lru_map<rgw_obj, tombstone_entry>;
2315   tombstone_cache_t *obj_tombstone_cache;
2316
2317   librados::IoCtx gc_pool_ctx;        // .rgw.gc
2318   librados::IoCtx lc_pool_ctx;        // .rgw.lc
2319   librados::IoCtx objexp_pool_ctx;
2320   librados::IoCtx reshard_pool_ctx;
2321
2322   bool pools_initialized;
2323
2324   string trans_id_suffix;
2325
2326   RGWQuotaHandler *quota_handler;
2327
2328   Finisher *finisher;
2329
2330   RGWCoroutinesManagerRegistry *cr_registry;
2331
2332   RGWSyncModulesManager *sync_modules_manager{nullptr};
2333   RGWSyncModuleInstanceRef sync_module;
2334   bool writeable_zone{false};
2335
2336   RGWZoneGroup zonegroup;
2337   RGWZone zone_public_config; /* external zone params, e.g., entrypoints, log flags, etc. */
2338   RGWZoneParams zone_params; /* internal zone params, e.g., rados pools */
2339   uint32_t zone_short_id;
2340
2341   RGWPeriod current_period;
2342
2343   RGWIndexCompletionManager *index_completion_manager{nullptr};
2344 public:
2345   RGWRados() : lock("rados_timer_lock"), watchers_lock("watchers_lock"), timer(NULL),
2346                gc(NULL), lc(NULL), obj_expirer(NULL), use_gc_thread(false), use_lc_thread(false), quota_threads(false),
2347                run_sync_thread(false), run_reshard_thread(false), async_rados(nullptr), meta_notifier(NULL),
2348                data_notifier(NULL), meta_sync_processor_thread(NULL),
2349                meta_sync_thread_lock("meta_sync_thread_lock"), data_sync_thread_lock("data_sync_thread_lock"),
2350                num_watchers(0), watchers(NULL),
2351                watch_initialized(false),
2352                bucket_id_lock("rados_bucket_id"),
2353                bucket_index_max_shards(0),
2354                max_bucket_id(0), cct(NULL),
2355                next_rados_handle(0),
2356                handle_lock("rados_handle_lock"),
2357                binfo_cache(NULL), obj_tombstone_cache(nullptr),
2358                pools_initialized(false),
2359                quota_handler(NULL),
2360                finisher(NULL),
2361                cr_registry(NULL),
2362                zone_short_id(0),
2363                rest_master_conn(NULL),
2364                meta_mgr(NULL), data_log(NULL), reshard(NULL) {}
2365
2366   uint64_t get_new_req_id() {
2367     return ++max_req_id;
2368   }
2369
2370   librados::IoCtx* get_lc_pool_ctx() {
2371     return &lc_pool_ctx;
2372   }
2373   void set_context(CephContext *_cct) {
2374     cct = _cct;
2375   }
2376
2377   /**
2378    * AmazonS3 errors contain a HostId string, but is an opaque base64 blob; we
2379    * try to be more transparent. This has a wrapper so we can update it when zonegroup/zone are changed.
2380    */
2381   void init_host_id() {
2382     /* uint64_t needs 16, two '-' separators and a trailing null */
2383     const string& zone_name = get_zone().name;
2384     const string& zonegroup_name = zonegroup.get_name();
2385     char charbuf[16 + zone_name.size() + zonegroup_name.size() + 2 + 1];
2386     snprintf(charbuf, sizeof(charbuf), "%llx-%s-%s", (unsigned long long)instance_id(), zone_name.c_str(), zonegroup_name.c_str());
2387     string s(charbuf);
2388     host_id = s;
2389   }
2390
2391   string host_id;
2392
2393   RGWRealm realm;
2394
2395   RGWRESTConn *rest_master_conn;
2396   map<string, RGWRESTConn *> zone_conn_map;
2397   map<string, RGWRESTConn *> zone_data_sync_from_map;
2398   map<string, RGWRESTConn *> zone_data_notify_to_map;
2399   map<string, RGWRESTConn *> zonegroup_conn_map;
2400
2401   map<string, string> zone_id_by_name;
2402   map<string, RGWZone> zone_by_id;
2403
2404   RGWRESTConn *get_zone_conn_by_id(const string& id) {
2405     auto citer = zone_conn_map.find(id);
2406     if (citer == zone_conn_map.end()) {
2407       return NULL;
2408     }
2409
2410     return citer->second;
2411   }
2412
2413   RGWRESTConn *get_zone_conn_by_name(const string& name) {
2414     auto i = zone_id_by_name.find(name);
2415     if (i == zone_id_by_name.end()) {
2416       return NULL;
2417     }
2418
2419     return get_zone_conn_by_id(i->second);
2420   }
2421
2422   bool find_zone_id_by_name(const string& name, string *id) {
2423     auto i = zone_id_by_name.find(name);
2424     if (i == zone_id_by_name.end()) {
2425       return false;
2426     }
2427     *id = i->second;
2428     return true;
2429   }
2430
2431   int get_zonegroup(const string& id, RGWZoneGroup& zonegroup) {
2432     int ret = 0;
2433     if (id == get_zonegroup().get_id()) {
2434       zonegroup = get_zonegroup();
2435     } else if (!current_period.get_id().empty()) {
2436       ret = current_period.get_zonegroup(zonegroup, id);
2437     }
2438     return ret;
2439   }
2440
2441   RGWRealm& get_realm() {
2442     return realm;
2443   }
2444
2445   RGWZoneParams& get_zone_params() { return zone_params; }
2446   RGWZoneGroup& get_zonegroup() {
2447     return zonegroup;
2448   }
2449   RGWZone& get_zone() {
2450     return zone_public_config;
2451   }
2452
2453   bool zone_is_writeable() {
2454     return writeable_zone && !get_zone().is_read_only();
2455   }
2456
2457   uint32_t get_zone_short_id() const {
2458     return zone_short_id;
2459   }
2460
2461   bool zone_syncs_from(RGWZone& target_zone, RGWZone& source_zone);
2462
2463   const RGWQuotaInfo& get_bucket_quota() {
2464     return current_period.get_config().bucket_quota;
2465   }
2466
2467   const RGWQuotaInfo& get_user_quota() {
2468     return current_period.get_config().user_quota;
2469   }
2470
2471   const string& get_current_period_id() {
2472     return current_period.get_id();
2473   }
2474
2475   bool has_zonegroup_api(const std::string& api) const {
2476     if (!current_period.get_id().empty()) {
2477       const auto& zonegroups_by_api = current_period.get_map().zonegroups_by_api;
2478       if (zonegroups_by_api.find(api) != zonegroups_by_api.end())
2479         return true;
2480     }
2481     return false;
2482   }
2483
2484   // pulls missing periods for period_history
2485   std::unique_ptr<RGWPeriodPuller> period_puller;
2486   // maintains a connected history of periods
2487   std::unique_ptr<RGWPeriodHistory> period_history;
2488
2489   RGWAsyncRadosProcessor* get_async_rados() const { return async_rados; };
2490
2491   RGWMetadataManager *meta_mgr;
2492
2493   RGWDataChangesLog *data_log;
2494
2495   RGWReshard *reshard;
2496   std::shared_ptr<RGWReshardWait> reshard_wait;
2497
2498   virtual ~RGWRados() = default;
2499
2500   tombstone_cache_t *get_tombstone_cache() {
2501     return obj_tombstone_cache;
2502   }
2503
2504   RGWSyncModulesManager *get_sync_modules_manager() {
2505     return sync_modules_manager;
2506   }
2507   const RGWSyncModuleInstanceRef& get_sync_module() {
2508     return sync_module;
2509   }
2510
2511   int get_required_alignment(const rgw_pool& pool, uint64_t *alignment);
2512   int get_max_chunk_size(const rgw_pool& pool, uint64_t *max_chunk_size);
2513   int get_max_chunk_size(const string& placement_rule, const rgw_obj& obj, uint64_t *max_chunk_size);
2514
2515   uint32_t get_max_bucket_shards() {
2516     return rgw_shards_max();
2517   }
2518
2519   int get_raw_obj_ref(const rgw_raw_obj& obj, rgw_rados_ref *ref);
2520
2521   int list_raw_objects(const rgw_pool& pool, const string& prefix_filter, int max,
2522                        RGWListRawObjsCtx& ctx, list<string>& oids,
2523                        bool *is_truncated);
2524
2525   int list_raw_prefixed_objs(const rgw_pool& pool, const string& prefix, list<string>& result);
2526   int list_zonegroups(list<string>& zonegroups);
2527   int list_regions(list<string>& regions);
2528   int list_zones(list<string>& zones);
2529   int list_realms(list<string>& realms);
2530   int list_periods(list<string>& periods);
2531   int list_periods(const string& current_period, list<string>& periods);
2532   void tick();
2533
2534   CephContext *ctx() { return cct; }
2535   /** do all necessary setup of the storage device */
2536   int initialize(CephContext *_cct, bool _use_gc_thread, bool _use_lc_thread, bool _quota_threads, bool _run_sync_thread, bool _run_reshard_thread) {
2537     set_context(_cct);
2538     use_gc_thread = _use_gc_thread;
2539     use_lc_thread = _use_lc_thread;
2540     quota_threads = _quota_threads;
2541     run_sync_thread = _run_sync_thread;
2542     run_reshard_thread = _run_reshard_thread;
2543     return initialize();
2544   }
2545   /** Initialize the RADOS instance and prepare to do other ops */
2546   virtual int init_rados();
2547   int init_zg_from_period(bool *initialized);
2548   int init_zg_from_local(bool *creating_defaults);
2549   int init_complete();
2550   int replace_region_with_zonegroup();
2551   int convert_regionmap();
2552   int initialize();
2553   void finalize();
2554
2555   int register_to_service_map(const string& daemon_type, const map<string, string>& meta);
2556
2557   void schedule_context(Context *c);
2558
2559   /** set up a bucket listing. handle is filled in. */
2560   int list_buckets_init(RGWAccessHandle *handle);
2561   /**
2562    * get the next bucket in the listing. obj is filled in,
2563    * handle is updated.
2564    */
2565   int list_buckets_next(rgw_bucket_dir_entry& obj, RGWAccessHandle *handle);
2566
2567   /// list logs
2568   int log_list_init(const string& prefix, RGWAccessHandle *handle);
2569   int log_list_next(RGWAccessHandle handle, string *name);
2570
2571   /// remove log
2572   int log_remove(const string& name);
2573
2574   /// show log
2575   int log_show_init(const string& name, RGWAccessHandle *handle);
2576   int log_show_next(RGWAccessHandle handle, rgw_log_entry *entry);
2577
2578   // log bandwidth info
2579   int log_usage(map<rgw_user_bucket, RGWUsageBatch>& usage_info);
2580   int read_usage(const rgw_user& user, uint64_t start_epoch, uint64_t end_epoch, uint32_t max_entries,
2581                  bool *is_truncated, RGWUsageIter& read_iter, map<rgw_user_bucket, rgw_usage_log_entry>& usage);
2582   int trim_usage(rgw_user& user, uint64_t start_epoch, uint64_t end_epoch);
2583
2584   int create_pool(const rgw_pool& pool);
2585
2586   int init_bucket_index(RGWBucketInfo& bucket_info, int num_shards);
2587   int select_bucket_placement(RGWUserInfo& user_info, const string& zonegroup_id, const string& rule,
2588                               string *pselected_rule_name, RGWZonePlacementInfo *rule_info);
2589   int select_legacy_bucket_placement(RGWZonePlacementInfo *rule_info);
2590   int select_new_bucket_location(RGWUserInfo& user_info, const string& zonegroup_id, const string& rule,
2591                                  string *pselected_rule_name, RGWZonePlacementInfo *rule_info);
2592   int select_bucket_location_by_rule(const string& location_rule, RGWZonePlacementInfo *rule_info);
2593   void create_bucket_id(string *bucket_id);
2594
2595   bool get_obj_data_pool(const string& placement_rule, const rgw_obj& obj, rgw_pool *pool);
2596   bool obj_to_raw(const string& placement_rule, const rgw_obj& obj, rgw_raw_obj *raw_obj);
2597
2598   int create_bucket(RGWUserInfo& owner, rgw_bucket& bucket,
2599                             const string& zonegroup_id,
2600                             const string& placement_rule,
2601                             const string& swift_ver_location,
2602                             const RGWQuotaInfo * pquota_info,
2603                             map<std::string,bufferlist>& attrs,
2604                             RGWBucketInfo& bucket_info,
2605                             obj_version *pobjv,
2606                             obj_version *pep_objv,
2607                             ceph::real_time creation_time,
2608                             rgw_bucket *master_bucket,
2609                             uint32_t *master_num_shards,
2610                             bool exclusive = true);
2611   int add_bucket_placement(const rgw_pool& new_pool);
2612   int remove_bucket_placement(const rgw_pool& new_pool);
2613   int list_placement_set(set<rgw_pool>& names);
2614   int create_pools(vector<rgw_pool>& pools, vector<int>& retcodes);
2615
2616   RGWCoroutinesManagerRegistry *get_cr_registry() { return cr_registry; }
2617
2618   class SystemObject {
2619     RGWRados *store;
2620     RGWObjectCtx& ctx;
2621     rgw_raw_obj obj;
2622
2623     RGWObjState *state;
2624
2625   protected:
2626     int get_state(RGWRawObjState **pstate, RGWObjVersionTracker *objv_tracker);
2627
2628   public:
2629     SystemObject(RGWRados *_store, RGWObjectCtx& _ctx, rgw_raw_obj& _obj) : store(_store), ctx(_ctx), obj(_obj), state(NULL) {}
2630
2631     void invalidate_state();
2632
2633     RGWRados *get_store() { return store; }
2634     rgw_raw_obj& get_obj() { return obj; }
2635     RGWObjectCtx& get_ctx() { return ctx; }
2636
2637     struct Read {
2638       RGWRados::SystemObject *source;
2639
2640       struct GetObjState {
2641         rgw_rados_ref ref;
2642         bool has_ref{false};
2643         uint64_t last_ver{0};
2644
2645         GetObjState() {}
2646
2647         int get_ref(RGWRados *store, rgw_raw_obj& obj, rgw_rados_ref **pref);
2648       } state;
2649
2650       struct StatParams {
2651         ceph::real_time *lastmod;
2652         uint64_t *obj_size;
2653         map<string, bufferlist> *attrs;
2654
2655         StatParams() : lastmod(NULL), obj_size(NULL), attrs(NULL) {}
2656       } stat_params;
2657
2658       struct ReadParams {
2659         rgw_cache_entry_info *cache_info{nullptr};
2660         map<string, bufferlist> *attrs;
2661
2662         ReadParams() : attrs(NULL) {}
2663       } read_params;
2664
2665       explicit Read(RGWRados::SystemObject *_source) : source(_source) {}
2666
2667       int stat(RGWObjVersionTracker *objv_tracker);
2668       int read(int64_t ofs, int64_t end, bufferlist& bl, RGWObjVersionTracker *objv_tracker);
2669       int get_attr(const char *name, bufferlist& dest);
2670     };
2671   };
2672
2673   struct BucketShard {
2674     RGWRados *store;
2675     rgw_bucket bucket;
2676     int shard_id;
2677     librados::IoCtx index_ctx;
2678     string bucket_obj;
2679
2680     explicit BucketShard(RGWRados *_store) : store(_store), shard_id(-1) {}
2681     int init(const rgw_bucket& _bucket, const rgw_obj& obj);
2682     int init(const rgw_bucket& _bucket, int sid);
2683   };
2684
2685   class Object {
2686     RGWRados *store;
2687     RGWBucketInfo bucket_info;
2688     RGWObjectCtx& ctx;
2689     rgw_obj obj;
2690
2691     BucketShard bs;
2692
2693     RGWObjState *state;
2694
2695     bool versioning_disabled;
2696
2697     bool bs_initialized;
2698
2699   protected:
2700     int get_state(RGWObjState **pstate, bool follow_olh, bool assume_noent = false);
2701     void invalidate_state();
2702
2703     int prepare_atomic_modification(librados::ObjectWriteOperation& op, bool reset_obj, const string *ptag,
2704                                     const char *ifmatch, const char *ifnomatch, bool removal_op);
2705     int complete_atomic_modification();
2706
2707   public:
2708     Object(RGWRados *_store, const RGWBucketInfo& _bucket_info, RGWObjectCtx& _ctx, const rgw_obj& _obj) : store(_store), bucket_info(_bucket_info),
2709                                                                                                ctx(_ctx), obj(_obj), bs(store),
2710                                                                                                state(NULL), versioning_disabled(false),
2711                                                                                                bs_initialized(false) {}
2712
2713     RGWRados *get_store() { return store; }
2714     rgw_obj& get_obj() { return obj; }
2715     RGWObjectCtx& get_ctx() { return ctx; }
2716     RGWBucketInfo& get_bucket_info() { return bucket_info; }
2717     int get_manifest(RGWObjManifest **pmanifest);
2718
2719     int get_bucket_shard(BucketShard **pbs) {
2720       if (!bs_initialized) {
2721         int r = bs.init(bucket_info.bucket, obj);
2722         if (r < 0) {
2723           return r;
2724         }
2725         bs_initialized = true;
2726       }
2727       *pbs = &bs;
2728       return 0;
2729     }
2730
2731     void set_versioning_disabled(bool status) {
2732       versioning_disabled = status;
2733     }
2734
2735     bool versioning_enabled() {
2736       return (!versioning_disabled && bucket_info.versioning_enabled());
2737     }
2738
2739     struct Read {
2740       RGWRados::Object *source;
2741
2742       struct GetObjState {
2743         librados::IoCtx io_ctx;
2744         rgw_obj obj;
2745         rgw_raw_obj head_obj;
2746       } state;
2747
2748       struct ConditionParams {
2749         const ceph::real_time *mod_ptr;
2750         const ceph::real_time *unmod_ptr;
2751         bool high_precision_time;
2752         uint32_t mod_zone_id;
2753         uint64_t mod_pg_ver;
2754         const char *if_match;
2755         const char *if_nomatch;
2756
2757         ConditionParams() :
2758                  mod_ptr(NULL), unmod_ptr(NULL), high_precision_time(false), mod_zone_id(0), mod_pg_ver(0),
2759                  if_match(NULL), if_nomatch(NULL) {}
2760       } conds;
2761
2762       struct Params {
2763         ceph::real_time *lastmod;
2764         uint64_t *obj_size;
2765         map<string, bufferlist> *attrs;
2766
2767         Params() : lastmod(NULL), obj_size(NULL), attrs(NULL) {}
2768       } params;
2769
2770       explicit Read(RGWRados::Object *_source) : source(_source) {}
2771
2772       int prepare();
2773       static int range_to_ofs(uint64_t obj_size, int64_t &ofs, int64_t &end);
2774       int read(int64_t ofs, int64_t end, bufferlist& bl);
2775       int iterate(int64_t ofs, int64_t end, RGWGetDataCB *cb);
2776       int get_attr(const char *name, bufferlist& dest);
2777     };
2778
2779     struct Write {
2780       RGWRados::Object *target;
2781
2782       struct MetaParams {
2783         ceph::real_time *mtime;
2784         map<std::string, bufferlist>* rmattrs;
2785         const bufferlist *data;
2786         RGWObjManifest *manifest;
2787         const string *ptag;
2788         list<rgw_obj_index_key> *remove_objs;
2789         ceph::real_time set_mtime;
2790         rgw_user owner;
2791         RGWObjCategory category;
2792         int flags;
2793         const char *if_match;
2794         const char *if_nomatch;
2795         uint64_t olh_epoch;
2796         ceph::real_time delete_at;
2797         bool canceled;
2798         const string *user_data;
2799         rgw_zone_set *zones_trace;
2800
2801         MetaParams() : mtime(NULL), rmattrs(NULL), data(NULL), manifest(NULL), ptag(NULL),
2802                  remove_objs(NULL), category(RGW_OBJ_CATEGORY_MAIN), flags(0),
2803                  if_match(NULL), if_nomatch(NULL), olh_epoch(0), canceled(false), user_data(nullptr), zones_trace(nullptr) {}
2804       } meta;
2805
2806       explicit Write(RGWRados::Object *_target) : target(_target) {}
2807
2808       int _do_write_meta(uint64_t size, uint64_t accounted_size,
2809                      map<std::string, bufferlist>& attrs,
2810                      bool assume_noent,
2811                      void *index_op);
2812       int write_meta(uint64_t size, uint64_t accounted_size,
2813                      map<std::string, bufferlist>& attrs);
2814       int write_data(const char *data, uint64_t ofs, uint64_t len, bool exclusive);
2815     };
2816
2817     struct Delete {
2818       RGWRados::Object *target;
2819
2820       struct DeleteParams {
2821         rgw_user bucket_owner;
2822         int versioning_status;
2823         ACLOwner obj_owner; /* needed for creation of deletion marker */
2824         uint64_t olh_epoch;
2825         string marker_version_id;
2826         uint32_t bilog_flags;
2827         list<rgw_obj_index_key> *remove_objs;
2828         ceph::real_time expiration_time;
2829         ceph::real_time unmod_since;
2830         ceph::real_time mtime; /* for setting delete marker mtime */
2831         bool high_precision_time;
2832         rgw_zone_set *zones_trace;
2833
2834         DeleteParams() : versioning_status(0), olh_epoch(0), bilog_flags(0), remove_objs(NULL), high_precision_time(false), zones_trace(nullptr) {}
2835       } params;
2836
2837       struct DeleteResult {
2838         bool delete_marker;
2839         string version_id;
2840
2841         DeleteResult() : delete_marker(false) {}
2842       } result;
2843
2844       explicit Delete(RGWRados::Object *_target) : target(_target) {}
2845
2846       int delete_obj();
2847     };
2848
2849     struct Stat {
2850       RGWRados::Object *source;
2851
2852       struct Result {
2853         rgw_obj obj;
2854         RGWObjManifest manifest;
2855         bool has_manifest;
2856         uint64_t size;
2857         struct timespec mtime;
2858         map<string, bufferlist> attrs;
2859
2860         Result() : has_manifest(false), size(0) {}
2861       } result;
2862
2863       struct State {
2864         librados::IoCtx io_ctx;
2865         librados::AioCompletion *completion;
2866         int ret;
2867
2868         State() : completion(NULL), ret(0) {}
2869       } state;
2870
2871
2872       explicit Stat(RGWRados::Object *_source) : source(_source) {}
2873
2874       int stat_async();
2875       int wait();
2876       int stat();
2877     private:
2878       int finish();
2879     };
2880   };
2881
2882   class Bucket {
2883     RGWRados *store;
2884     RGWBucketInfo bucket_info;
2885     rgw_bucket& bucket;
2886     int shard_id;
2887
2888   public:
2889     Bucket(RGWRados *_store, const RGWBucketInfo& _bucket_info) : store(_store), bucket_info(_bucket_info), bucket(bucket_info.bucket),
2890                                                             shard_id(RGW_NO_SHARD) {}
2891     RGWRados *get_store() { return store; }
2892     rgw_bucket& get_bucket() { return bucket; }
2893     RGWBucketInfo& get_bucket_info() { return bucket_info; }
2894
2895     int update_bucket_id(const string& new_bucket_id);
2896
2897     int get_shard_id() { return shard_id; }
2898     void set_shard_id(int id) {
2899       shard_id = id;
2900     }
2901
2902     class UpdateIndex {
2903       RGWRados::Bucket *target;
2904       string optag;
2905       rgw_obj obj;
2906       uint16_t bilog_flags{0};
2907       BucketShard bs;
2908       bool bs_initialized{false};
2909       bool blind;
2910       bool prepared{false};
2911       rgw_zone_set *zones_trace{nullptr};
2912
2913       int init_bs() {
2914         int r = bs.init(target->get_bucket(), obj);
2915         if (r < 0) {
2916           return r;
2917         }
2918         bs_initialized = true;
2919         return 0;
2920       }
2921
2922       void invalidate_bs() {
2923         bs_initialized = false;
2924       }
2925
2926       int guard_reshard(BucketShard **pbs, std::function<int(BucketShard *)> call);
2927     public:
2928
2929       UpdateIndex(RGWRados::Bucket *_target, const rgw_obj& _obj) : target(_target), obj(_obj),
2930                                                               bs(target->get_store()) {
2931                                                                 blind = (target->get_bucket_info().index_type == RGWBIType_Indexless);
2932                                                               }
2933
2934       int get_bucket_shard(BucketShard **pbs) {
2935         if (!bs_initialized) {
2936           int r = init_bs();
2937           if (r < 0) {
2938             return r;
2939           }
2940         }
2941         *pbs = &bs;
2942         return 0;
2943       }
2944
2945       void set_bilog_flags(uint16_t flags) {
2946         bilog_flags = flags;
2947       }
2948
2949       void set_zones_trace(rgw_zone_set *_zones_trace) {
2950         zones_trace = _zones_trace;
2951       }
2952
2953       int prepare(RGWModifyOp, const string *write_tag);
2954       int complete(int64_t poolid, uint64_t epoch, uint64_t size,
2955                    uint64_t accounted_size, ceph::real_time& ut,
2956                    const string& etag, const string& content_type,
2957                    bufferlist *acl_bl, RGWObjCategory category,
2958                    list<rgw_obj_index_key> *remove_objs, const string *user_data = nullptr);
2959       int complete_del(int64_t poolid, uint64_t epoch,
2960                        ceph::real_time& removed_mtime, /* mtime of removed object */
2961                        list<rgw_obj_index_key> *remove_objs);
2962       int cancel();
2963
2964       const string *get_optag() { return &optag; }
2965
2966       bool is_prepared() { return prepared; }
2967     };
2968
2969     struct List {
2970       RGWRados::Bucket *target;
2971       rgw_obj_key next_marker;
2972
2973       struct Params {
2974         string prefix;
2975         string delim;
2976         rgw_obj_key marker;
2977         rgw_obj_key end_marker;
2978         string ns;
2979         bool enforce_ns;
2980         RGWAccessListFilter *filter;
2981         bool list_versions;
2982
2983         Params() : enforce_ns(true), filter(NULL), list_versions(false) {}
2984       } params;
2985
2986     public:
2987       explicit List(RGWRados::Bucket *_target) : target(_target) {}
2988
2989       int list_objects(int64_t max, vector<rgw_bucket_dir_entry> *result, map<string, bool> *common_prefixes, bool *is_truncated);
2990       rgw_obj_key& get_next_marker() {
2991         return next_marker;
2992       }
2993     };
2994   };
2995
2996   /** Write/overwrite an object to the bucket storage. */
2997   virtual int put_system_obj_impl(rgw_raw_obj& obj, uint64_t size, ceph::real_time *mtime,
2998               map<std::string, bufferlist>& attrs, int flags,
2999               bufferlist& data,
3000               RGWObjVersionTracker *objv_tracker,
3001               ceph::real_time set_mtime /* 0 for don't set */);
3002
3003   virtual int put_system_obj_data(void *ctx, rgw_raw_obj& obj, bufferlist& bl,
3004               off_t ofs, bool exclusive,
3005               RGWObjVersionTracker *objv_tracker = nullptr);
3006   int aio_put_obj_data(void *ctx, rgw_raw_obj& obj, bufferlist& bl,
3007                         off_t ofs, bool exclusive, void **handle);
3008
3009   int put_system_obj(void *ctx, rgw_raw_obj& obj, const char *data, size_t len, bool exclusive,
3010               ceph::real_time *mtime, map<std::string, bufferlist>& attrs, RGWObjVersionTracker *objv_tracker,
3011               ceph::real_time set_mtime) {
3012     bufferlist bl;
3013     bl.append(data, len);
3014     int flags = PUT_OBJ_CREATE;
3015     if (exclusive)
3016       flags |= PUT_OBJ_EXCL;
3017
3018     return put_system_obj_impl(obj, len, mtime, attrs, flags, bl, objv_tracker, set_mtime);
3019   }
3020   int aio_wait(void *handle);
3021   bool aio_completed(void *handle);
3022
3023   int on_last_entry_in_listing(RGWBucketInfo& bucket_info,
3024                                const std::string& obj_prefix,
3025                                const std::string& obj_delim,
3026                                std::function<int(const rgw_bucket_dir_entry&)> handler);
3027
3028   bool swift_versioning_enabled(const RGWBucketInfo& bucket_info) const {
3029     return bucket_info.has_swift_versioning() &&
3030         bucket_info.swift_ver_location.size();
3031   }
3032
3033   int swift_versioning_copy(RGWObjectCtx& obj_ctx,              /* in/out */
3034                             const rgw_user& user,               /* in */
3035                             RGWBucketInfo& bucket_info,         /* in */
3036                             rgw_obj& obj);                      /* in */
3037   int swift_versioning_restore(RGWObjectCtx& obj_ctx,           /* in/out */
3038                                const rgw_user& user,            /* in */
3039                                RGWBucketInfo& bucket_info,      /* in */
3040                                rgw_obj& obj,                    /* in */
3041                                bool& restored);                 /* out */
3042   int copy_obj_to_remote_dest(RGWObjState *astate,
3043                               map<string, bufferlist>& src_attrs,
3044                               RGWRados::Object::Read& read_op,
3045                               const rgw_user& user_id,
3046                               rgw_obj& dest_obj,
3047                               ceph::real_time *mtime);
3048
3049   enum AttrsMod {
3050     ATTRSMOD_NONE    = 0,
3051     ATTRSMOD_REPLACE = 1,
3052     ATTRSMOD_MERGE   = 2
3053   };
3054
3055   int rewrite_obj(RGWBucketInfo& dest_bucket_info, rgw_obj& obj);
3056
3057   int stat_remote_obj(RGWObjectCtx& obj_ctx,
3058                const rgw_user& user_id,
3059                const string& client_id,
3060                req_info *info,
3061                const string& source_zone,
3062                rgw_obj& src_obj,
3063                RGWBucketInfo& src_bucket_info,
3064                real_time *src_mtime,
3065                uint64_t *psize,
3066                const real_time *mod_ptr,
3067                const real_time *unmod_ptr,
3068                bool high_precision_time,
3069                const char *if_match,
3070                const char *if_nomatch,
3071                map<string, bufferlist> *pattrs,
3072                string *version_id,
3073                string *ptag,
3074                string *petag);
3075
3076   int fetch_remote_obj(RGWObjectCtx& obj_ctx,
3077                        const rgw_user& user_id,
3078                        const string& client_id,
3079                        const string& op_id,
3080                        bool record_op_state,
3081                        req_info *info,
3082                        const string& source_zone,
3083                        rgw_obj& dest_obj,
3084                        rgw_obj& src_obj,
3085                        RGWBucketInfo& dest_bucket_info,
3086                        RGWBucketInfo& src_bucket_info,
3087                        ceph::real_time *src_mtime,
3088                        ceph::real_time *mtime,
3089                        const ceph::real_time *mod_ptr,
3090                        const ceph::real_time *unmod_ptr,
3091                        bool high_precision_time,
3092                        const char *if_match,
3093                        const char *if_nomatch,
3094                        AttrsMod attrs_mod,
3095                        bool copy_if_newer,
3096                        map<string, bufferlist>& attrs,
3097                        RGWObjCategory category,
3098                        uint64_t olh_epoch,
3099                        ceph::real_time delete_at,
3100                        string *version_id,
3101                        string *ptag,
3102                        ceph::buffer::list *petag,
3103                        void (*progress_cb)(off_t, void *),
3104                        void *progress_data,
3105                        rgw_zone_set *zones_trace= nullptr);
3106   /**
3107    * Copy an object.
3108    * dest_obj: the object to copy into
3109    * src_obj: the object to copy from
3110    * attrs: usage depends on attrs_mod parameter
3111    * attrs_mod: the modification mode of the attrs, may have the following values:
3112    *            ATTRSMOD_NONE - the attributes of the source object will be
3113    *                            copied without modifications, attrs parameter is ignored;
3114    *            ATTRSMOD_REPLACE - new object will have the attributes provided by attrs
3115    *                               parameter, source object attributes are not copied;
3116    *            ATTRSMOD_MERGE - any conflicting meta keys on the source object's attributes
3117    *                             are overwritten by values contained in attrs parameter.
3118    * Returns: 0 on success, -ERR# otherwise.
3119    */
3120   int copy_obj(RGWObjectCtx& obj_ctx,
3121                const rgw_user& user_id,
3122                const string& client_id,
3123                const string& op_id,
3124                req_info *info,
3125                const string& source_zone,
3126                rgw_obj& dest_obj,
3127                rgw_obj& src_obj,
3128                RGWBucketInfo& dest_bucket_info,
3129                RGWBucketInfo& src_bucket_info,
3130                ceph::real_time *src_mtime,
3131                ceph::real_time *mtime,
3132                const ceph::real_time *mod_ptr,
3133                const ceph::real_time *unmod_ptr,
3134                bool high_precision_time,
3135                const char *if_match,
3136                const char *if_nomatch,
3137                AttrsMod attrs_mod,
3138                bool copy_if_newer,
3139                map<std::string, bufferlist>& attrs,
3140                RGWObjCategory category,
3141                uint64_t olh_epoch,
3142                ceph::real_time delete_at,
3143                string *version_id,
3144                string *ptag,
3145                ceph::buffer::list *petag,
3146                void (*progress_cb)(off_t, void *),
3147                void *progress_data);
3148
3149   int copy_obj_data(RGWObjectCtx& obj_ctx,
3150                RGWBucketInfo& dest_bucket_info,
3151                RGWRados::Object::Read& read_op, off_t end,
3152                rgw_obj& dest_obj,
3153                rgw_obj& src_obj,
3154                uint64_t max_chunk_size,
3155                ceph::real_time *mtime,
3156                ceph::real_time set_mtime,
3157                map<string, bufferlist>& attrs,
3158                RGWObjCategory category,
3159                uint64_t olh_epoch,
3160                ceph::real_time delete_at,
3161                string *version_id,
3162                string *ptag,
3163                ceph::buffer::list *petag);
3164
3165   int check_bucket_empty(RGWBucketInfo& bucket_info);
3166
3167   /**
3168    * Delete a bucket.
3169    * bucket: the name of the bucket to delete
3170    * Returns 0 on success, -ERR# otherwise.
3171    */
3172   int delete_bucket(RGWBucketInfo& bucket_info, RGWObjVersionTracker& objv_tracker, bool check_empty = true);
3173
3174   bool is_meta_master();
3175
3176   /**
3177    * Check to see if the bucket metadata is synced
3178    */
3179   bool is_syncing_bucket_meta(const rgw_bucket& bucket);
3180   void wakeup_meta_sync_shards(set<int>& shard_ids);
3181   void wakeup_data_sync_shards(const string& source_zone, map<int, set<string> >& shard_ids);
3182
3183   RGWMetaSyncStatusManager* get_meta_sync_manager();
3184   RGWDataSyncStatusManager* get_data_sync_manager(const std::string& source_zone);
3185
3186   int set_bucket_owner(rgw_bucket& bucket, ACLOwner& owner);
3187   int set_buckets_enabled(std::vector<rgw_bucket>& buckets, bool enabled);
3188   int bucket_suspended(rgw_bucket& bucket, bool *suspended);
3189
3190   /** Delete an object.*/
3191   int delete_obj(RGWObjectCtx& obj_ctx,
3192                          const RGWBucketInfo& bucket_owner,
3193                          const rgw_obj& src_obj,
3194                          int versioning_status,
3195                          uint16_t bilog_flags = 0,
3196                          const ceph::real_time& expiration_time = ceph::real_time(),
3197                          rgw_zone_set *zones_trace = nullptr);
3198
3199   /** Delete a raw object.*/
3200   int delete_raw_obj(const rgw_raw_obj& obj);
3201
3202   /* Delete a system object */
3203   virtual int delete_system_obj(rgw_raw_obj& src_obj, RGWObjVersionTracker *objv_tracker = NULL);
3204
3205   /** Remove an object from the bucket index */
3206   int delete_obj_index(const rgw_obj& obj);
3207
3208   /**
3209    * Get an attribute for a system object.
3210    * obj: the object to get attr
3211    * name: name of the attr to retrieve
3212    * dest: bufferlist to store the result in
3213    * Returns: 0 on success, -ERR# otherwise.
3214    */
3215   virtual int system_obj_get_attr(rgw_raw_obj& obj, const char *name, bufferlist& dest);
3216
3217   int system_obj_set_attr(void *ctx, rgw_raw_obj& obj, const char *name, bufferlist& bl,
3218                           RGWObjVersionTracker *objv_tracker);
3219   virtual int system_obj_set_attrs(void *ctx, rgw_raw_obj& obj,
3220                                    map<string, bufferlist>& attrs,
3221                                    map<string, bufferlist>* rmattrs,
3222                                    RGWObjVersionTracker *objv_tracker);
3223
3224   /**
3225    * Set an attr on an object.
3226    * bucket: name of the bucket holding the object
3227    * obj: name of the object to set the attr on
3228    * name: the attr to set
3229    * bl: the contents of the attr
3230    * Returns: 0 on success, -ERR# otherwise.
3231    */
3232   int set_attr(void *ctx, const RGWBucketInfo& bucket_info, rgw_obj& obj, const char *name, bufferlist& bl);
3233
3234   int set_attrs(void *ctx, const RGWBucketInfo& bucket_info, rgw_obj& obj,
3235                         map<string, bufferlist>& attrs,
3236                         map<string, bufferlist>* rmattrs);
3237
3238   int get_system_obj_state(RGWObjectCtx *rctx, rgw_raw_obj& obj, RGWRawObjState **state, RGWObjVersionTracker *objv_tracker);
3239   int get_obj_state(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state,
3240                     bool follow_olh, bool assume_noent = false);
3241   int get_obj_state(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state) {
3242     return get_obj_state(rctx, bucket_info, obj, state, true);
3243   }
3244
3245   virtual int stat_system_obj(RGWObjectCtx& obj_ctx,
3246                               RGWRados::SystemObject::Read::GetObjState& state,
3247                               rgw_raw_obj& obj,
3248                               map<string, bufferlist> *attrs,
3249                               ceph::real_time *lastmod,
3250                               uint64_t *obj_size,
3251                               RGWObjVersionTracker *objv_tracker);
3252
3253   virtual int get_system_obj(RGWObjectCtx& obj_ctx, RGWRados::SystemObject::Read::GetObjState& read_state,
3254                              RGWObjVersionTracker *objv_tracker, rgw_raw_obj& obj,
3255                              bufferlist& bl, off_t ofs, off_t end,
3256                              map<string, bufferlist> *attrs,
3257                              rgw_cache_entry_info *cache_info);
3258
3259   virtual void register_chained_cache(RGWChainedCache *cache) {}
3260   virtual bool chain_cache_entry(list<rgw_cache_entry_info *>& cache_info_entries, RGWChainedCache::Entry *chained_entry) { return false; }
3261
3262   int iterate_obj(RGWObjectCtx& ctx,
3263                   const RGWBucketInfo& bucket_info, const rgw_obj& obj,
3264                   off_t ofs, off_t end,
3265                   uint64_t max_chunk_size,
3266                   int (*iterate_obj_cb)(const RGWBucketInfo& bucket_info, const rgw_obj& obj, const rgw_raw_obj&, off_t, off_t, off_t, bool, RGWObjState *, void *),
3267                   void *arg);
3268
3269   int flush_read_list(struct get_obj_data *d);
3270
3271   int get_obj_iterate_cb(RGWObjectCtx *ctx, RGWObjState *astate,
3272                          const RGWBucketInfo& bucket_info, const rgw_obj& obj,
3273                          const rgw_raw_obj& read_obj,
3274                          off_t obj_ofs, off_t read_ofs, off_t len,
3275                          bool is_head_obj, void *arg);
3276
3277   void get_obj_aio_completion_cb(librados::completion_t cb, void *arg);
3278
3279   /**
3280    * a simple object read without keeping state
3281    */
3282
3283   virtual int raw_obj_stat(rgw_raw_obj& obj, uint64_t *psize, ceph::real_time *pmtime, uint64_t *epoch,
3284                        map<string, bufferlist> *attrs, bufferlist *first_chunk,
3285                        RGWObjVersionTracker *objv_tracker);
3286
3287   int obj_operate(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::ObjectWriteOperation *op);
3288   int obj_operate(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::ObjectReadOperation *op);
3289
3290   int guard_reshard(BucketShard *bs, const rgw_obj& obj_instance, std::function<int(BucketShard *)> call);
3291   int block_while_resharding(RGWRados::BucketShard *bs, string *new_bucket_id);
3292
3293   void bucket_index_guard_olh_op(RGWObjState& olh_state, librados::ObjectOperation& op);
3294   int olh_init_modification(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, string *op_tag);
3295   int olh_init_modification_impl(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, string *op_tag);
3296   int bucket_index_link_olh(const RGWBucketInfo& bucket_info, RGWObjState& olh_state,
3297                             const rgw_obj& obj_instance, bool delete_marker,
3298                             const string& op_tag, struct rgw_bucket_dir_entry_meta *meta,
3299                             uint64_t olh_epoch,
3300                             ceph::real_time unmod_since, bool high_precision_time, rgw_zone_set *zones_trace = nullptr);
3301   int bucket_index_unlink_instance(const RGWBucketInfo& bucket_info, const rgw_obj& obj_instance, const string& op_tag, const string& olh_tag, uint64_t olh_epoch, rgw_zone_set *zones_trace = nullptr);
3302   int bucket_index_read_olh_log(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& obj_instance, uint64_t ver_marker,
3303                                 map<uint64_t, vector<rgw_bucket_olh_log_entry> > *log, bool *is_truncated);
3304   int bucket_index_trim_olh_log(const RGWBucketInfo& bucket_info, RGWObjState& obj_state, const rgw_obj& obj_instance, uint64_t ver);
3305   int bucket_index_clear_olh(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& obj_instance);
3306   int apply_olh_log(RGWObjectCtx& ctx, RGWObjState& obj_state, const RGWBucketInfo& bucket_info, const rgw_obj& obj,
3307                     bufferlist& obj_tag, map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log,
3308                     uint64_t *plast_ver, rgw_zone_set *zones_trace = nullptr);
3309   int update_olh(RGWObjectCtx& obj_ctx, RGWObjState *state, const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_zone_set *zones_trace = nullptr);
3310   int set_olh(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, const rgw_obj& target_obj, bool delete_marker, rgw_bucket_dir_entry_meta *meta,
3311               uint64_t olh_epoch, ceph::real_time unmod_since, bool high_precision_time, rgw_zone_set *zones_trace = nullptr);
3312   int unlink_obj_instance(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, const rgw_obj& target_obj,
3313                           uint64_t olh_epoch, rgw_zone_set *zones_trace = nullptr);
3314
3315   void check_pending_olh_entries(map<string, bufferlist>& pending_entries, map<string, bufferlist> *rm_pending_entries);
3316   int remove_olh_pending_entries(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, map<string, bufferlist>& pending_attrs);
3317   int follow_olh(const RGWBucketInfo& bucket_info, RGWObjectCtx& ctx, RGWObjState *state, const rgw_obj& olh_obj, rgw_obj *target);
3318   int get_olh(const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWOLHInfo *olh);
3319
3320   void gen_rand_obj_instance_name(rgw_obj *target);
3321
3322   int omap_get_vals(rgw_raw_obj& obj, bufferlist& header, const std::string& marker, uint64_t count, std::map<string, bufferlist>& m);
3323   int omap_get_all(rgw_raw_obj& obj, bufferlist& header, std::map<string, bufferlist>& m);
3324   int omap_set(rgw_raw_obj& obj, const std::string& key, bufferlist& bl);
3325   int omap_set(rgw_raw_obj& obj, map<std::string, bufferlist>& m);
3326   int omap_del(rgw_raw_obj& obj, const std::string& key);
3327   int update_containers_stats(map<string, RGWBucketEnt>& m);
3328   int append_async(rgw_raw_obj& obj, size_t size, bufferlist& bl);
3329
3330   int watch(const string& oid, uint64_t *watch_handle, librados::WatchCtx2 *ctx);
3331   int unwatch(uint64_t watch_handle);
3332   void add_watcher(int i);
3333   void remove_watcher(int i);
3334   virtual bool need_watch_notify() { return false; }
3335   int init_watch();
3336   void finalize_watch();
3337   int distribute(const string& key, bufferlist& bl);
3338   virtual int watch_cb(uint64_t notify_id,
3339                        uint64_t cookie,
3340                        uint64_t notifier_id,
3341                        bufferlist& bl) { return 0; }
3342   void pick_control_oid(const string& key, string& notify_oid);
3343
3344   virtual void set_cache_enabled(bool state) {}
3345
3346   void set_atomic(void *ctx, rgw_obj& obj) {
3347     RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx);
3348     rctx->obj.set_atomic(obj);
3349   }
3350   void set_prefetch_data(void *ctx, rgw_obj& obj) {
3351     RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx);
3352     rctx->obj.set_prefetch_data(obj);
3353   }
3354   void set_prefetch_data(void *ctx, rgw_raw_obj& obj) {
3355     RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx);
3356     rctx->raw.set_prefetch_data(obj);
3357   }
3358
3359   int decode_policy(bufferlist& bl, ACLOwner *owner);
3360   int get_bucket_stats(RGWBucketInfo& bucket_info, int shard_id, string *bucket_ver, string *master_ver,
3361       map<RGWObjCategory, RGWStorageStats>& stats, string *max_marker, bool* syncstopped = NULL);
3362   int get_bucket_stats_async(RGWBucketInfo& bucket_info, int shard_id, RGWGetBucketStats_CB *cb);
3363   int get_user_stats(const rgw_user& user, RGWStorageStats& stats);
3364   int get_user_stats_async(const rgw_user& user, RGWGetUserStats_CB *cb);
3365   void get_bucket_instance_obj(const rgw_bucket& bucket, rgw_raw_obj& obj);
3366   void get_bucket_meta_oid(const rgw_bucket& bucket, string& oid);
3367
3368   int put_bucket_entrypoint_info(const string& tenant_name, const string& bucket_name, RGWBucketEntryPoint& entry_point,
3369                                  bool exclusive, RGWObjVersionTracker& objv_tracker, ceph::real_time mtime,
3370                                  map<string, bufferlist> *pattrs);
3371   int put_bucket_instance_info(RGWBucketInfo& info, bool exclusive, ceph::real_time mtime, map<string, bufferlist> *pattrs);
3372   int get_bucket_entrypoint_info(RGWObjectCtx& obj_ctx, const string& tenant_name, const string& bucket_name,
3373                                  RGWBucketEntryPoint& entry_point, RGWObjVersionTracker *objv_tracker,
3374                                  ceph::real_time *pmtime, map<string, bufferlist> *pattrs, rgw_cache_entry_info *cache_info = NULL);
3375   int get_bucket_instance_info(RGWObjectCtx& obj_ctx, const string& meta_key, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs);
3376   int get_bucket_instance_info(RGWObjectCtx& obj_ctx, const rgw_bucket& bucket, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs);
3377   int get_bucket_instance_from_oid(RGWObjectCtx& obj_ctx, const string& oid, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs,
3378                                    rgw_cache_entry_info *cache_info = NULL);
3379
3380   int convert_old_bucket_info(RGWObjectCtx& obj_ctx, const string& tenant_name, const string& bucket_name);
3381   static void make_bucket_entry_name(const string& tenant_name, const string& bucket_name, string& bucket_entry);
3382   int get_bucket_info(RGWObjectCtx& obj_ctx,
3383                               const string& tenant_name, const string& bucket_name,
3384                               RGWBucketInfo& info,
3385                               ceph::real_time *pmtime, map<string, bufferlist> *pattrs = NULL);
3386   int put_linked_bucket_info(RGWBucketInfo& info, bool exclusive, ceph::real_time mtime, obj_version *pep_objv,
3387                                      map<string, bufferlist> *pattrs, bool create_entry_point);
3388
3389   int cls_rgw_init_index(librados::IoCtx& io_ctx, librados::ObjectWriteOperation& op, string& oid);
3390   int cls_obj_prepare_op(BucketShard& bs, RGWModifyOp op, string& tag, rgw_obj& obj, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr);
3391   int cls_obj_complete_op(BucketShard& bs, const rgw_obj& obj, RGWModifyOp op, string& tag, int64_t pool, uint64_t epoch,
3392                           rgw_bucket_dir_entry& ent, RGWObjCategory category, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr);
3393   int cls_obj_complete_add(BucketShard& bs, const rgw_obj& obj, string& tag, int64_t pool, uint64_t epoch, rgw_bucket_dir_entry& ent,
3394                            RGWObjCategory category, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr);
3395   int cls_obj_complete_del(BucketShard& bs, string& tag, int64_t pool, uint64_t epoch, rgw_obj& obj,
3396                            ceph::real_time& removed_mtime, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr);
3397   int cls_obj_complete_cancel(BucketShard& bs, string& tag, rgw_obj& obj, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr);
3398   int cls_obj_set_bucket_tag_timeout(RGWBucketInfo& bucket_info, uint64_t timeout);
3399   int cls_bucket_list(RGWBucketInfo& bucket_info, int shard_id, rgw_obj_index_key& start, const string& prefix,
3400                       uint32_t num_entries, bool list_versions, map<string, rgw_bucket_dir_entry>& m,
3401                       bool *is_truncated, rgw_obj_index_key *last_entry,
3402                       bool (*force_check_filter)(const string&  name) = NULL);
3403   int cls_bucket_head(const RGWBucketInfo& bucket_info, int shard_id, map<string, struct rgw_bucket_dir_header>& headers, map<int, string> *bucket_instance_ids = NULL);
3404   int cls_bucket_head_async(const RGWBucketInfo& bucket_info, int shard_id, RGWGetDirHeader_CB *ctx, int *num_aio);
3405   int list_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id, string& marker, uint32_t max, std::list<rgw_bi_log_entry>& result, bool *truncated);
3406   int trim_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id, string& marker, string& end_marker);
3407   int resync_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id);
3408   int stop_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id);
3409   int get_bi_log_status(RGWBucketInfo& bucket_info, int shard_id, map<int, string>& max_marker);
3410
3411   int bi_get_instance(const RGWBucketInfo& bucket_info, rgw_obj& obj, rgw_bucket_dir_entry *dirent);
3412   int bi_get(rgw_bucket& bucket, rgw_obj& obj, BIIndexType index_type, rgw_cls_bi_entry *entry);
3413   void bi_put(librados::ObjectWriteOperation& op, BucketShard& bs, rgw_cls_bi_entry& entry);
3414   int bi_put(BucketShard& bs, rgw_cls_bi_entry& entry);
3415   int bi_put(rgw_bucket& bucket, rgw_obj& obj, rgw_cls_bi_entry& entry);
3416   int bi_list(rgw_bucket& bucket, int shard_id, const string& filter_obj, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated);
3417   int bi_list(BucketShard& bs, const string& filter_obj, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated);
3418   int bi_list(rgw_bucket& bucket, const string& obj_name, const string& marker, uint32_t max,
3419               list<rgw_cls_bi_entry> *entries, bool *is_truncated);
3420   int bi_remove(BucketShard& bs);
3421
3422   int cls_obj_usage_log_add(const string& oid, rgw_usage_log_info& info);
3423   int cls_obj_usage_log_read(string& oid, string& user, uint64_t start_epoch, uint64_t end_epoch, uint32_t max_entries,
3424                              string& read_iter, map<rgw_user_bucket, rgw_usage_log_entry>& usage, bool *is_truncated);
3425   int cls_obj_usage_log_trim(string& oid, string& user, uint64_t start_epoch, uint64_t end_epoch);
3426
3427   int key_to_shard_id(const string& key, int max_shards);
3428   void shard_name(const string& prefix, unsigned max_shards, const string& key, string& name, int *shard_id);
3429   void shard_name(const string& prefix, unsigned max_shards, const string& section, const string& key, string& name);
3430   void shard_name(const string& prefix, unsigned shard_id, string& name);
3431   int get_target_shard_id(const RGWBucketInfo& bucket_info, const string& obj_key, int *shard_id);
3432   void time_log_prepare_entry(cls_log_entry& entry, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl);
3433   int time_log_add_init(librados::IoCtx& io_ctx);
3434   int time_log_add(const string& oid, list<cls_log_entry>& entries,
3435                    librados::AioCompletion *completion, bool monotonic_inc = true);
3436   int time_log_add(const string& oid, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl);
3437   int time_log_list(const string& oid, const ceph::real_time& start_time, const ceph::real_time& end_time,
3438                     int max_entries, list<cls_log_entry>& entries,
3439                     const string& marker, string *out_marker, bool *truncated);
3440   int time_log_info(const string& oid, cls_log_header *header);
3441   int time_log_info_async(librados::IoCtx& io_ctx, const string& oid, cls_log_header *header, librados::AioCompletion *completion);
3442   int time_log_trim(const string& oid, const ceph::real_time& start_time, const ceph::real_time& end_time,
3443                     const string& from_marker, const string& to_marker,
3444                     librados::AioCompletion *completion = nullptr);
3445
3446   string objexp_hint_get_shardname(int shard_num);
3447   int objexp_key_shard(const rgw_obj_index_key& key);
3448   void objexp_get_shard(int shard_num,
3449                         string& shard);                       /* out */
3450   int objexp_hint_add(const ceph::real_time& delete_at,
3451                       const string& tenant_name,
3452                       const string& bucket_name,
3453                       const string& bucket_id,
3454                       const rgw_obj_index_key& obj_key);
3455   int objexp_hint_list(const string& oid,
3456                        const ceph::real_time& start_time,
3457                        const ceph::real_time& end_time,
3458                        const int max_entries,
3459                        const string& marker,
3460                        list<cls_timeindex_entry>& entries, /* out */
3461                        string *out_marker,                 /* out */
3462                        bool *truncated);                   /* out */
3463   int objexp_hint_parse(cls_timeindex_entry &ti_entry,
3464                         objexp_hint_entry& hint_entry);    /* out */
3465   int objexp_hint_trim(const string& oid,
3466                        const ceph::real_time& start_time,
3467                        const ceph::real_time& end_time,
3468                        const string& from_marker = std::string(),
3469                        const string& to_marker   = std::string());
3470
3471   int lock_exclusive(rgw_pool& pool, const string& oid, ceph::timespan& duration, string& zone_id, string& owner_id);
3472   int unlock(rgw_pool& pool, const string& oid, string& zone_id, string& owner_id);
3473
3474   void update_gc_chain(rgw_obj& head_obj, RGWObjManifest& manifest, cls_rgw_obj_chain *chain);
3475   int send_chain_to_gc(cls_rgw_obj_chain& chain, const string& tag, bool sync);
3476   int gc_operate(string& oid, librados::ObjectWriteOperation *op);
3477   int gc_aio_operate(string& oid, librados::ObjectWriteOperation *op);
3478   int gc_operate(string& oid, librados::ObjectReadOperation *op, bufferlist *pbl);
3479
3480   int list_gc_objs(int *index, string& marker, uint32_t max, bool expired_only, std::list<cls_rgw_gc_obj_info>& result, bool *truncated);
3481   int process_gc();
3482   int process_expire_objects();
3483   int defer_gc(void *ctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj);
3484
3485   int process_lc();
3486   int list_lc_progress(const string& marker, uint32_t max_entries, map<string, int> *progress_map);
3487
3488   int bucket_check_index(RGWBucketInfo& bucket_info,
3489                          map<RGWObjCategory, RGWStorageStats> *existing_stats,
3490                          map<RGWObjCategory, RGWStorageStats> *calculated_stats);
3491   int bucket_rebuild_index(RGWBucketInfo& bucket_info);
3492   int bucket_set_reshard(RGWBucketInfo& bucket_info, const cls_rgw_bucket_instance_entry& entry);
3493   int remove_objs_from_index(RGWBucketInfo& bucket_info, list<rgw_obj_index_key>& oid_list);
3494   int move_rados_obj(librados::IoCtx& src_ioctx,
3495                      const string& src_oid, const string& src_locator,
3496                      librados::IoCtx& dst_ioctx,
3497                      const string& dst_oid, const string& dst_locator);
3498   int fix_head_obj_locator(const RGWBucketInfo& bucket_info, bool copy_obj, bool remove_bad, rgw_obj_key& key);
3499   int fix_tail_obj_locator(const RGWBucketInfo& bucket_info, rgw_obj_key& key, bool fix, bool *need_fix);
3500
3501   int cls_user_get_header(const string& user_id, cls_user_header *header);
3502   int cls_user_get_header_async(const string& user_id, RGWGetUserHeader_CB *ctx);
3503   int cls_user_sync_bucket_stats(rgw_raw_obj& user_obj, const RGWBucketInfo& bucket_info);
3504   int cls_user_list_buckets(rgw_raw_obj& obj,
3505                             const string& in_marker,
3506                             const string& end_marker,
3507                             int max_entries,
3508                             list<cls_user_bucket_entry>& entries,
3509                             string *out_marker,
3510                             bool *truncated);
3511   int cls_user_add_bucket(rgw_raw_obj& obj, const cls_user_bucket_entry& entry);
3512   int cls_user_update_buckets(rgw_raw_obj& obj, list<cls_user_bucket_entry>& entries, bool add);
3513   int cls_user_complete_stats_sync(rgw_raw_obj& obj);
3514   int complete_sync_user_stats(const rgw_user& user_id);
3515   int cls_user_add_bucket(rgw_raw_obj& obj, list<cls_user_bucket_entry>& entries);
3516   int cls_user_remove_bucket(rgw_raw_obj& obj, const cls_user_bucket& bucket);
3517   int cls_user_get_bucket_stats(const rgw_bucket& bucket, cls_user_bucket_entry& entry);
3518
3519   int check_quota(const rgw_user& bucket_owner, rgw_bucket& bucket,
3520                   RGWQuotaInfo& user_quota, RGWQuotaInfo& bucket_quota, uint64_t obj_size);
3521
3522   int check_bucket_shards(const RGWBucketInfo& bucket_info, const rgw_bucket& bucket,
3523                           RGWQuotaInfo& bucket_quota);
3524
3525   int add_bucket_to_reshard(const RGWBucketInfo& bucket_info, uint32_t new_num_shards);
3526
3527   uint64_t instance_id();
3528   const string& zone_name() {
3529     return get_zone_params().get_name();
3530   }
3531   const string& zone_id() {
3532     return get_zone_params().get_id();
3533   }
3534   string unique_id(uint64_t unique_num) {
3535     char buf[32];
3536     snprintf(buf, sizeof(buf), ".%llu.%llu", (unsigned long long)instance_id(), (unsigned long long)unique_num);
3537     string s = get_zone_params().get_id() + buf;
3538     return s;
3539   }
3540
3541   void init_unique_trans_id_deps() {
3542     char buf[16 + 2 + 1]; /* uint64_t needs 16, 2 hyphens add further 2 */
3543
3544     snprintf(buf, sizeof(buf), "-%llx-", (unsigned long long)instance_id());
3545     url_encode(string(buf) + get_zone_params().get_name(), trans_id_suffix);
3546   }
3547
3548   /* In order to preserve compability with Swift API, transaction ID
3549    * should contain at least 32 characters satisfying following spec:
3550    *  - first 21 chars must be in range [0-9a-f]. Swift uses this
3551    *    space for storing fragment of UUID obtained through a call to
3552    *    uuid4() function of Python's uuid module;
3553    *  - char no. 22 must be a hyphen;
3554    *  - at least 10 next characters constitute hex-formatted timestamp
3555    *    padded with zeroes if necessary. All bytes must be in [0-9a-f]
3556    *    range;
3557    *  - last, optional part of transaction ID is any url-encoded string
3558    *    without restriction on length. */
3559   string unique_trans_id(const uint64_t unique_num) {
3560     char buf[41]; /* 2 + 21 + 1 + 16 (timestamp can consume up to 16) + 1 */
3561     time_t timestamp = time(NULL);
3562
3563     snprintf(buf, sizeof(buf), "tx%021llx-%010llx",
3564              (unsigned long long)unique_num,
3565              (unsigned long long)timestamp);
3566
3567     return string(buf) + trans_id_suffix;
3568   }
3569
3570   void get_log_pool(rgw_pool& pool) {
3571     pool = get_zone_params().log_pool;
3572   }
3573
3574   bool need_to_log_data() {
3575     return get_zone().log_data;
3576   }
3577
3578   bool need_to_log_metadata() {
3579     return is_meta_master() &&
3580       (get_zonegroup().zones.size() > 1 || current_period.is_multi_zonegroups_with_zones());
3581   }
3582
3583   librados::Rados* get_rados_handle();
3584
3585   int delete_raw_obj_aio(const rgw_raw_obj& obj, list<librados::AioCompletion *>& handles);
3586   int delete_obj_aio(const rgw_obj& obj, RGWBucketInfo& info, RGWObjState *astate,
3587                      list<librados::AioCompletion *>& handles, bool keep_index_consistent);
3588  private:
3589   /**
3590    * This is a helper method, it generates a list of bucket index objects with the given
3591    * bucket base oid and number of shards.
3592    *
3593    * bucket_oid_base [in] - base name of the bucket index object;
3594    * num_shards [in] - number of bucket index object shards.
3595    * bucket_objs [out] - filled by this method, a list of bucket index objects.
3596    */
3597   void get_bucket_index_objects(const string& bucket_oid_base, uint32_t num_shards,
3598       map<int, string>& bucket_objs, int shard_id = -1);
3599
3600   /**
3601    * Get the bucket index object with the given base bucket index object and object key,
3602    * and the number of bucket index shards.
3603    *
3604    * bucket_oid_base [in] - bucket object base name.
3605    * obj_key [in] - object key.
3606    * num_shards [in] - number of bucket index shards.
3607    * hash_type [in] - type of hash to find the shard ID.
3608    * bucket_obj [out] - the bucket index object for the given object.
3609    *
3610    * Return 0 on success, a failure code otherwise.
3611    */
3612   int get_bucket_index_object(const string& bucket_oid_base, const string& obj_key,
3613       uint32_t num_shards, RGWBucketInfo::BIShardsHashType hash_type, string *bucket_obj, int *shard);
3614
3615   void get_bucket_index_object(const string& bucket_oid_base, uint32_t num_shards,
3616                                int shard_id, string *bucket_obj);
3617
3618   /**
3619    * Check the actual on-disk state of the object specified
3620    * by list_state, and fill in the time and size of object.
3621    * Then append any changes to suggested_updates for
3622    * the rgw class' dir_suggest_changes function.
3623    *
3624    * Note that this can maul list_state; don't use it afterwards. Also
3625    * it expects object to already be filled in from list_state; it only
3626    * sets the size and mtime.
3627    *
3628    * Returns 0 on success, -ENOENT if the object doesn't exist on disk,
3629    * and -errno on other failures. (-ENOENT is not a failure, and it
3630    * will encode that info as a suggested update.)
3631    */
3632   int check_disk_state(librados::IoCtx io_ctx,
3633                        const RGWBucketInfo& bucket_info,
3634                        rgw_bucket_dir_entry& list_state,
3635                        rgw_bucket_dir_entry& object,
3636                        bufferlist& suggested_updates);
3637
3638   /**
3639    * Init pool iteration
3640    * pool: pool to use for the ctx initialization
3641    * ctx: context object to use for the iteration
3642    * Returns: 0 on success, -ERR# otherwise.
3643    */
3644   int pool_iterate_begin(const rgw_pool& pool, RGWPoolIterCtx& ctx);
3645
3646   /**
3647    * Iterate over pool return object names, use optional filter
3648    * ctx: iteration context, initialized with pool_iterate_begin()
3649    * num: max number of objects to return
3650    * objs: a vector that the results will append into
3651    * is_truncated: if not NULL, will hold true iff iteration is complete
3652    * filter: if not NULL, will be used to filter returned objects
3653    * Returns: 0 on success, -ERR# otherwise.
3654    */
3655   int pool_iterate(RGWPoolIterCtx& ctx, uint32_t num, vector<rgw_bucket_dir_entry>& objs,
3656                    bool *is_truncated, RGWAccessListFilter *filter);
3657
3658   uint64_t next_bucket_id();
3659 };
3660
3661 class RGWStoreManager {
3662 public:
3663   RGWStoreManager() {}
3664   static RGWRados *get_storage(CephContext *cct, bool use_gc_thread, bool use_lc_thread, bool quota_threads, bool run_sync_thread, bool run_reshard_thread) {
3665     RGWRados *store = init_storage_provider(cct, use_gc_thread, use_lc_thread, quota_threads, run_sync_thread,
3666                                             run_reshard_thread);
3667     return store;
3668   }
3669   static RGWRados *get_raw_storage(CephContext *cct) {
3670     RGWRados *store = init_raw_storage_provider(cct);
3671     return store;
3672   }
3673   static RGWRados *init_storage_provider(CephContext *cct, bool use_gc_thread, bool use_lc_thread, bool quota_threads, bool run_sync_thread, bool run_reshard_thread);
3674   static RGWRados *init_raw_storage_provider(CephContext *cct);
3675   static void close_storage(RGWRados *store);
3676
3677 };
3678
3679 template <class T>
3680 class RGWChainedCacheImpl : public RGWChainedCache {
3681   RWLock lock;
3682
3683   map<string, T> entries;
3684
3685 public:
3686   RGWChainedCacheImpl() : lock("RGWChainedCacheImpl::lock") {}
3687
3688   void init(RGWRados *store) {
3689     store->register_chained_cache(this);
3690   }
3691
3692   bool find(const string& key, T *entry) {
3693     RWLock::RLocker rl(lock);
3694     typename map<string, T>::iterator iter = entries.find(key);
3695     if (iter == entries.end()) {
3696       return false;
3697     }
3698
3699     *entry = iter->second;
3700     return true;
3701   }
3702
3703   bool put(RGWRados *store, const string& key, T *entry, list<rgw_cache_entry_info *>& cache_info_entries) {
3704     Entry chain_entry(this, key, entry);
3705
3706     /* we need the store cache to call us under its lock to maintain lock ordering */
3707     return store->chain_cache_entry(cache_info_entries, &chain_entry);
3708   }
3709
3710   void chain_cb(const string& key, void *data) override {
3711     T *entry = static_cast<T *>(data);
3712     RWLock::WLocker wl(lock);
3713     entries[key] = *entry;
3714   }
3715
3716   void invalidate(const string& key) override {
3717     RWLock::WLocker wl(lock);
3718     entries.erase(key);
3719   }
3720
3721   void invalidate_all() override {
3722     RWLock::WLocker wl(lock);
3723     entries.clear();
3724   }
3725 }; /* RGWChainedCacheImpl */
3726
3727 /**
3728  * Base of PUT operation.
3729  * Allow to create chained data transformers like compresors and encryptors.
3730  */
3731 class RGWPutObjDataProcessor
3732 {
3733 public:
3734   RGWPutObjDataProcessor(){}
3735   virtual ~RGWPutObjDataProcessor(){}
3736   virtual int handle_data(bufferlist& bl, off_t ofs, void **phandle, rgw_raw_obj *pobj, bool *again) = 0;
3737   virtual int throttle_data(void *handle, const rgw_raw_obj& obj, uint64_t size, bool need_to_wait) = 0;
3738 }; /* RGWPutObjDataProcessor */
3739
3740
3741 class RGWPutObjProcessor : public RGWPutObjDataProcessor
3742 {
3743 protected:
3744   RGWRados *store;
3745   RGWObjectCtx& obj_ctx;
3746   bool is_complete;
3747   RGWBucketInfo bucket_info;
3748   bool canceled;
3749
3750   virtual int do_complete(size_t accounted_size, const string& etag,
3751                           ceph::real_time *mtime, ceph::real_time set_mtime,
3752                           map<string, bufferlist>& attrs, ceph::real_time delete_at,
3753                           const char *if_match, const char *if_nomatch, const string *user_data,
3754                           rgw_zone_set* zones_trace = nullptr) = 0;
3755
3756 public:
3757   RGWPutObjProcessor(RGWObjectCtx& _obj_ctx, RGWBucketInfo& _bi) : store(NULL),
3758                                                                    obj_ctx(_obj_ctx),
3759                                                                    is_complete(false),
3760                                                                    bucket_info(_bi),
3761                                                                    canceled(false) {}
3762   ~RGWPutObjProcessor() override {}
3763   virtual int prepare(RGWRados *_store, string *oid_rand) {
3764     store = _store;
3765     return 0;
3766   }
3767
3768   int complete(size_t accounted_size, const string& etag,
3769                ceph::real_time *mtime, ceph::real_time set_mtime,
3770                map<string, bufferlist>& attrs, ceph::real_time delete_at,
3771                const char *if_match = NULL, const char *if_nomatch = NULL, const string *user_data = nullptr,
3772                rgw_zone_set *zones_trace = nullptr);
3773
3774   CephContext *ctx();
3775
3776   bool is_canceled() { return canceled; }
3777 }; /* RGWPutObjProcessor */
3778
3779 struct put_obj_aio_info {
3780   void *handle;
3781   rgw_raw_obj obj;
3782   uint64_t size;
3783 };
3784
3785 #define RGW_PUT_OBJ_MIN_WINDOW_SIZE_DEFAULT (16 * 1024 * 1024)
3786
3787 class RGWPutObjProcessor_Aio : public RGWPutObjProcessor
3788 {
3789   list<struct put_obj_aio_info> pending;
3790   uint64_t window_size{RGW_PUT_OBJ_MIN_WINDOW_SIZE_DEFAULT};
3791   uint64_t pending_size{0};
3792
3793   struct put_obj_aio_info pop_pending();
3794   int wait_pending_front();
3795   bool pending_has_completed();
3796
3797   rgw_raw_obj last_written_obj;
3798
3799 protected:
3800   uint64_t obj_len{0};
3801
3802   set<rgw_raw_obj> written_objs;
3803   rgw_obj head_obj;
3804
3805   void add_written_obj(const rgw_raw_obj& obj) {
3806     written_objs.insert(obj);
3807   }
3808
3809   int drain_pending();
3810   int handle_obj_data(rgw_raw_obj& obj, bufferlist& bl, off_t ofs, off_t abs_ofs, void **phandle, bool exclusive);
3811
3812 public:
3813   int prepare(RGWRados *store, string *oid_rand) override;
3814   int throttle_data(void *handle, const rgw_raw_obj& obj, uint64_t size, bool need_to_wait) override;
3815
3816   RGWPutObjProcessor_Aio(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info) : RGWPutObjProcessor(obj_ctx, bucket_info) {}
3817   ~RGWPutObjProcessor_Aio() override;
3818 }; /* RGWPutObjProcessor_Aio */
3819
3820 class RGWPutObjProcessor_Atomic : public RGWPutObjProcessor_Aio
3821 {
3822   bufferlist first_chunk;
3823   uint64_t part_size;
3824   off_t cur_part_ofs;
3825   off_t next_part_ofs;
3826   int cur_part_id;
3827   off_t data_ofs;
3828
3829   bufferlist pending_data_bl;
3830   uint64_t max_chunk_size;
3831
3832   bool versioned_object;
3833   uint64_t olh_epoch;
3834   string version_id;
3835
3836 protected:
3837   rgw_bucket bucket;
3838   string obj_str;
3839
3840   string unique_tag;
3841
3842   rgw_raw_obj cur_obj;
3843   RGWObjManifest manifest;
3844   RGWObjManifest::generator manifest_gen;
3845
3846   int write_data(bufferlist& bl, off_t ofs, void **phandle, rgw_raw_obj *pobj, bool exclusive);
3847   int do_complete(size_t accounted_size, const string& etag,
3848                   ceph::real_time *mtime, ceph::real_time set_mtime,
3849                   map<string, bufferlist>& attrs, ceph::real_time delete_at,
3850                   const char *if_match, const char *if_nomatch, const string *user_data, rgw_zone_set *zones_trace) override;
3851
3852   int prepare_next_part(off_t ofs);
3853   int complete_parts();
3854   int complete_writing_data();
3855
3856   int prepare_init(RGWRados *store, string *oid_rand);
3857
3858 public:
3859   ~RGWPutObjProcessor_Atomic() override {}
3860   RGWPutObjProcessor_Atomic(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info,
3861                             rgw_bucket& _b, const string& _o, uint64_t _p, const string& _t, bool versioned) :
3862                                 RGWPutObjProcessor_Aio(obj_ctx, bucket_info),
3863                                 part_size(_p),
3864                                 cur_part_ofs(0),
3865                                 next_part_ofs(_p),
3866                                 cur_part_id(0),
3867                                 data_ofs(0),
3868                                 max_chunk_size(0),
3869                                 versioned_object(versioned),
3870                                 olh_epoch(0),
3871                                 bucket(_b),
3872                                 obj_str(_o),
3873                                 unique_tag(_t) {}
3874   int prepare(RGWRados *store, string *oid_rand) override;
3875   virtual bool immutable_head() { return false; }
3876   int handle_data(bufferlist& bl, off_t ofs, void **phandle, rgw_raw_obj *pobj, bool *again) override;
3877
3878   void set_olh_epoch(uint64_t epoch) {
3879     olh_epoch = epoch;
3880   }
3881
3882   void set_version_id(const string& vid) {
3883     version_id = vid;
3884   }
3885 }; /* RGWPutObjProcessor_Atomic */
3886
3887 #define MP_META_SUFFIX ".meta"
3888
3889 class RGWMPObj {
3890   string oid;
3891   string prefix;
3892   string meta;
3893   string upload_id;
3894 public:
3895   RGWMPObj() {}
3896   RGWMPObj(const string& _oid, const string& _upload_id) {
3897     init(_oid, _upload_id, _upload_id);
3898   }
3899   void init(const string& _oid, const string& _upload_id) {
3900     init(_oid, _upload_id, _upload_id);
3901   }
3902   void init(const string& _oid, const string& _upload_id, const string& part_unique_str) {
3903     if (_oid.empty()) {
3904       clear();
3905       return;
3906     }
3907     oid = _oid;
3908     upload_id = _upload_id;
3909     prefix = oid + ".";
3910     meta = prefix + upload_id + MP_META_SUFFIX;
3911     prefix.append(part_unique_str);
3912   }
3913   string& get_meta() { return meta; }
3914   string get_part(int num) {
3915     char buf[16];
3916     snprintf(buf, 16, ".%d", num);
3917     string s = prefix;
3918     s.append(buf);
3919     return s;
3920   }
3921   string get_part(string& part) {
3922     string s = prefix;
3923     s.append(".");
3924     s.append(part);
3925     return s;
3926   }
3927   string& get_upload_id() {
3928     return upload_id;
3929   }
3930   string& get_key() {
3931     return oid;
3932   }
3933   bool from_meta(string& meta) {
3934     int end_pos = meta.rfind('.'); // search for ".meta"
3935     if (end_pos < 0)
3936       return false;
3937     int mid_pos = meta.rfind('.', end_pos - 1); // <key>.<upload_id>
3938     if (mid_pos < 0)
3939       return false;
3940     oid = meta.substr(0, mid_pos);
3941     upload_id = meta.substr(mid_pos + 1, end_pos - mid_pos - 1);
3942     init(oid, upload_id, upload_id);
3943     return true;
3944   }
3945   void clear() {
3946     oid = "";
3947     prefix = "";
3948     meta = "";
3949     upload_id = "";
3950   }
3951 };
3952
3953 class RGWPutObjProcessor_Multipart : public RGWPutObjProcessor_Atomic
3954 {
3955   string part_num;
3956   RGWMPObj mp;
3957   req_state *s;
3958   string upload_id;
3959
3960 protected:
3961   int prepare(RGWRados *store, string *oid_rand);
3962   int do_complete(size_t accounted_size, const string& etag,
3963                   ceph::real_time *mtime, ceph::real_time set_mtime,
3964                   map<string, bufferlist>& attrs, ceph::real_time delete_at,
3965                   const char *if_match, const char *if_nomatch, const string *user_data,
3966                   rgw_zone_set *zones_trace) override;
3967 public:
3968   bool immutable_head() { return true; }
3969   RGWPutObjProcessor_Multipart(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, uint64_t _p, req_state *_s) :
3970                    RGWPutObjProcessor_Atomic(obj_ctx, bucket_info, _s->bucket, _s->object.name, _p, _s->req_id, false), s(_s) {}
3971   void get_mp(RGWMPObj** _mp);
3972 }; /* RGWPutObjProcessor_Multipart */
3973 #endif