ceph/src/cls/rgw/cls_rgw.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3
   4 #include "include/types.h"
   5
   6 #include <errno.h>
   7
   8 #include "objclass/objclass.h"
   9 #include "cls/rgw/cls_rgw_ops.h"
  10 #include "cls/rgw/cls_rgw_const.h"
  11 #include "common/Clock.h"
  12 #include "common/strtol.h"
  13 #include "common/escape.h"
  14
  15 #include "include/compat.h"
  16
  17 CLS_VER(1,0)
  18 CLS_NAME(rgw)
  19
  20
  21 #define BI_PREFIX_CHAR 0x80
  22
  23 #define BI_BUCKET_OBJS_INDEX          0
  24 #define BI_BUCKET_LOG_INDEX           1
  25 #define BI_BUCKET_OBJ_INSTANCE_INDEX  2
  26 #define BI_BUCKET_OLH_DATA_INDEX      3
  27
  28 #define BI_BUCKET_LAST_INDEX          4
  29
  30 static std::string bucket_index_prefixes[] = { "", /* special handling for the objs list index */
  31                                           "0_",     /* bucket log index */
  32                                           "1000_",  /* obj instance index */
  33                                           "1001_",  /* olh data index */
  34
  35                                           /* this must be the last index */
  36                                           "9999_",};
  37
  38 static bool bi_is_objs_index(const string& s) {
  39   return ((unsigned char)s[0] != BI_PREFIX_CHAR);
  40 }
  41
  42 int bi_entry_type(const string& s)
  43 {
  44   if (bi_is_objs_index(s)) {
  45     return BI_BUCKET_OBJS_INDEX;
  46   }
  47
  48   for (size_t i = 1;
  49        i < sizeof(bucket_index_prefixes) / sizeof(bucket_index_prefixes[0]);
  50        ++i) {
  51     const string& t = bucket_index_prefixes[i];
  52
  53     if (s.compare(1, t.size(), t) == 0) {
  54       return i;
  55     }
  56   }
  57
  58   return -EINVAL;
  59 }
  60
  61 static bool bi_entry_gt(const string& first, const string& second)
  62 {
  63   int fi = bi_entry_type(first);
  64   int si = bi_entry_type(second);
  65
  66   if (fi > si) {
  67     return true;
  68   } else if (fi < si) {
  69     return false;
  70   }
  71
  72   return first > second;
  73 }
  74
  75 static void get_time_key(real_time& ut, string *key)
  76 {
  77   char buf[32];
  78   ceph_timespec ts = ceph::real_clock::to_ceph_timespec(ut);
  79   snprintf(buf, 32, "%011llu.%09u", (unsigned long long)ts.tv_sec, (unsigned int)ts.tv_nsec);
  80   *key = buf;
  81 }
  82
  83 static void get_index_ver_key(cls_method_context_t hctx, uint64_t index_ver, string *key)
  84 {
  85   char buf[48];
  86   snprintf(buf, sizeof(buf), "%011llu.%llu.%d", (unsigned long long)index_ver,
  87            (unsigned long long)cls_current_version(hctx),
  88            cls_current_subop_num(hctx));
  89   *key = buf;
  90 }
  91
  92 static void bi_log_prefix(string& key)
  93 {
  94   key = BI_PREFIX_CHAR;
  95   key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
  96 }
  97
  98 static void bi_log_index_key(cls_method_context_t hctx, string& key, string& id, uint64_t index_ver)
  99 {
 100   bi_log_prefix(key);
 101   get_index_ver_key(hctx, index_ver, &id);
 102   key.append(id);
 103 }
 104
 105 static int log_index_operation(cls_method_context_t hctx, cls_rgw_obj_key& obj_key, RGWModifyOp op,
 106                                string& tag, real_time& timestamp,
 107                                rgw_bucket_entry_ver& ver, RGWPendingState state, uint64_t index_ver,
 108                                string& max_marker, uint16_t bilog_flags, string *owner, string *owner_display_name, rgw_zone_set *zones_trace)
 109 {
 110   bufferlist bl;
 111
 112   rgw_bi_log_entry entry;
 113
 114   entry.object = obj_key.name;
 115   entry.instance = obj_key.instance;
 116   entry.timestamp = timestamp;
 117   entry.op = op;
 118   entry.ver = ver;
 119   entry.state = state;
 120   entry.index_ver = index_ver;
 121   entry.tag = tag;
 122   entry.bilog_flags = bilog_flags;
 123   if (owner) {
 124     entry.owner = *owner;
 125   }
 126   if (owner_display_name) {
 127     entry.owner_display_name = *owner_display_name;
 128   }
 129   if (zones_trace) {
 130     entry.zones_trace = std::move(*zones_trace);
 131   }
 132
 133   string key;
 134   bi_log_index_key(hctx, key, entry.id, index_ver);
 135
 136   encode(entry, bl);
 137
 138   if (entry.id > max_marker)
 139     max_marker = entry.id;
 140
 141   return cls_cxx_map_set_val(hctx, key, &bl);
 142 }
 143
 144 /*
 145  * read list of objects, skips objects in the ugly namespace
 146  */
 147 static int get_obj_vals(cls_method_context_t hctx, const string& start, const string& filter_prefix,
 148                         int num_entries, map<string, bufferlist> *pkeys, bool *pmore)
 149 {
 150   int ret = cls_cxx_map_get_vals(hctx, start, filter_prefix, num_entries, pkeys, pmore);
 151   if (ret < 0)
 152     return ret;
 153
 154   if (pkeys->empty())
 155     return 0;
 156
 157   auto last_element = pkeys->rbegin();
 158   if ((unsigned char)last_element->first[0] < BI_PREFIX_CHAR) {
 159     /* nothing to see here, move along */
 160     return 0;
 161   }
 162
 163   auto first_element = pkeys->begin();
 164   if ((unsigned char)first_element->first[0] > BI_PREFIX_CHAR) {
 165     return 0;
 166   }
 167
 168   /* let's rebuild the list, only keep entries we're interested in */
 169   auto comp = [](const pair<string, bufferlist>& l, const string &r) { return l.first < r; };
 170   string new_start = {static_cast<char>(BI_PREFIX_CHAR + 1)};
 171
 172   auto lower = pkeys->lower_bound(string{static_cast<char>(BI_PREFIX_CHAR)});
 173   auto upper = std::lower_bound(lower, pkeys->end(), new_start, comp);
 174   pkeys->erase(lower, upper);
 175
 176   if (num_entries == (int)pkeys->size() || !(*pmore))
 177     return 0;
 178
 179   if (pkeys->size() && new_start < pkeys->rbegin()->first) {
 180     new_start = pkeys->rbegin()->first;
 181   }
 182
 183   map<string, bufferlist> new_keys;
 184
 185   /* now get some more keys */
 186   ret = cls_cxx_map_get_vals(hctx, new_start, filter_prefix, num_entries - pkeys->size(), &new_keys, pmore);
 187   if (ret < 0)
 188     return ret;
 189
 190   pkeys->insert(std::make_move_iterator(new_keys.begin()),
 191                 std::make_move_iterator(new_keys.end()));
 192   return 0;
 193 }
 194
 195 /*
 196  * get a monotonically decreasing string representation.
 197  * For num = x, num = y, where x > y, str(x) < str(y)
 198  * Another property is that string size starts short and grows as num increases
 199  */
 200 static void decreasing_str(uint64_t num, string *str)
 201 {
 202   char buf[32];
 203   if (num < 0x10) { /* 16 */
 204     snprintf(buf, sizeof(buf), "9%02lld", 15 - (long long)num);
 205   } else if (num < 0x100) { /* 256 */
 206     snprintf(buf, sizeof(buf), "8%03lld", 255 - (long long)num);
 207   } else if (num < 0x1000) /* 4096 */ {
 208     snprintf(buf, sizeof(buf), "7%04lld", 4095 - (long long)num);
 209   } else if (num < 0x10000) /* 65536 */ {
 210     snprintf(buf, sizeof(buf), "6%05lld", 65535 - (long long)num);
 211   } else if (num < 0x100000000) /* 4G */ {
 212     snprintf(buf, sizeof(buf), "5%010lld", 0xFFFFFFFF - (long long)num);
 213   } else {
 214     snprintf(buf, sizeof(buf), "4%020lld",  (long long)-num);
 215   }
 216
 217   *str = buf;
 218 }
 219
 220 /*
 221  * we now hold two different indexes for objects. The first one holds the list of objects in the
 222  * order that we want them to be listed. The second one only holds the objects instances (for
 223  * versioned objects), and they're not arranged in any particular order.
 224  * When listing objects we'll use the first index, when doing operations on the objects themselves
 225  * we'll use the second index. Note that regular objects only map to the first index anyway
 226  */
 227
 228 static void get_list_index_key(rgw_bucket_dir_entry& entry, string *index_key)
 229 {
 230   *index_key = entry.key.name;
 231
 232   string ver_str;
 233   decreasing_str(entry.versioned_epoch, &ver_str);
 234   string instance_delim("\0i", 2);
 235   string ver_delim("\0v", 2);
 236
 237   index_key->append(ver_delim);
 238   index_key->append(ver_str);
 239   index_key->append(instance_delim);
 240   index_key->append(entry.key.instance);
 241 }
 242
 243 static void encode_obj_versioned_data_key(const cls_rgw_obj_key& key, string *index_key, bool append_delete_marker_suffix = false)
 244 {
 245   *index_key = BI_PREFIX_CHAR;
 246   index_key->append(bucket_index_prefixes[BI_BUCKET_OBJ_INSTANCE_INDEX]);
 247   index_key->append(key.name);
 248   string delim("\0i", 2);
 249   index_key->append(delim);
 250   index_key->append(key.instance);
 251   if (append_delete_marker_suffix) {
 252     string dm("\0d", 2);
 253     index_key->append(dm);
 254   }
 255 }
 256
 257 static void encode_obj_index_key(const cls_rgw_obj_key& key, string *index_key)
 258 {
 259   if (key.instance.empty()) {
 260     *index_key = key.name;
 261   } else {
 262     encode_obj_versioned_data_key(key, index_key);
 263   }
 264 }
 265
 266 static void encode_olh_data_key(const cls_rgw_obj_key& key, string *index_key)
 267 {
 268   *index_key = BI_PREFIX_CHAR;
 269   index_key->append(bucket_index_prefixes[BI_BUCKET_OLH_DATA_INDEX]);
 270   index_key->append(key.name);
 271 }
 272
 273 template <class T>
 274 static int read_index_entry(cls_method_context_t hctx, string& name, T *entry);
 275
 276 static int encode_list_index_key(cls_method_context_t hctx, const cls_rgw_obj_key& key, string *index_key)
 277 {
 278   if (key.instance.empty()) {
 279     *index_key = key.name;
 280     return 0;
 281   }
 282
 283   string obj_index_key;
 284   encode_obj_index_key(key, &obj_index_key);
 285
 286   rgw_bucket_dir_entry entry;
 287
 288   int ret = read_index_entry(hctx, obj_index_key, &entry);
 289   if (ret == -ENOENT) {
 290    /* couldn't find the entry, set key value after the current object */
 291     char buf[2] = { 0x1, 0 };
 292     string s(buf);
 293     *index_key  = key.name + s;
 294     return 0;
 295   }
 296   if (ret < 0) {
 297     CLS_LOG(1, "ERROR: encode_list_index_key(): cls_cxx_map_get_val returned %d\n", ret);
 298     return ret;
 299   }
 300
 301   get_list_index_key(entry, index_key);
 302
 303   return 0;
 304 }
 305
 306 static void split_key(const string& key, list<string>& vals)
 307 {
 308   size_t pos = 0;
 309   const char *p = key.c_str();
 310   while (pos < key.size()) {
 311     size_t len = strlen(p);
 312     vals.push_back(p);
 313     pos += len + 1;
 314     p += len + 1;
 315   }
 316 }
 317
 318 static string escape_str(const string& s)
 319 {
 320   int len = escape_json_attr_len(s.c_str(), s.size());
 321   std::string escaped(len, 0);
 322   escape_json_attr(s.c_str(), s.size(), escaped.data());
 323   return escaped;
 324 }
 325
 326 /*
 327  * list index key structure:
 328  *
 329  * <obj name>\0[v<ver>\0i<instance id>]
 330  */
 331 static int decode_list_index_key(const string& index_key, cls_rgw_obj_key *key, uint64_t *ver)
 332 {
 333   size_t len = strlen(index_key.c_str());
 334
 335   key->instance.clear();
 336   *ver = 0;
 337
 338   if (len == index_key.size()) {
 339     key->name = index_key;
 340     return 0;
 341   }
 342
 343   list<string> vals;
 344   split_key(index_key, vals);
 345
 346   if (vals.empty()) {
 347     CLS_LOG(0, "ERROR: %s(): bad index_key (%s): split_key() returned empty vals", __func__, escape_str(index_key).c_str());
 348     return -EIO;
 349   }
 350
 351   list<string>::iterator iter = vals.begin();
 352   key->name = *iter;
 353   ++iter;
 354
 355   if (iter == vals.end()) {
 356     CLS_LOG(0, "ERROR: %s(): bad index_key (%s): no vals", __func__, escape_str(index_key).c_str());
 357     return -EIO;
 358   }
 359
 360   for (; iter != vals.end(); ++iter) {
 361     string& val = *iter;
 362     if (val[0] == 'i') {
 363       key->instance = val.substr(1);
 364     } else if (val[0] == 'v') {
 365       string err;
 366       const char *s = val.c_str() + 1;
 367       *ver = strict_strtoll(s, 10, &err);
 368       if (!err.empty()) {
 369         CLS_LOG(0, "ERROR: %s(): bad index_key (%s): could not parse val (v=%s)", __func__, escape_str(index_key).c_str(), s);
 370         return -EIO;
 371       }
 372     }
 373   }
 374
 375   return 0;
 376 }
 377
 378 static int read_bucket_header(cls_method_context_t hctx,
 379                               rgw_bucket_dir_header *header)
 380 {
 381   bufferlist bl;
 382   int rc = cls_cxx_map_read_header(hctx, &bl);
 383   if (rc < 0)
 384     return rc;
 385
 386   if (bl.length() == 0) {
 387       *header = rgw_bucket_dir_header();
 388       return 0;
 389   }
 390   auto iter = bl.cbegin();
 391   try {
 392     decode(*header, iter);
 393   } catch (buffer::error& err) {
 394     CLS_LOG(1, "ERROR: read_bucket_header(): failed to decode header\n");
 395     return -EIO;
 396   }
 397
 398   return 0;
 399 }
 400
 401 int rgw_bucket_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 402 {
 403   auto iter = in->cbegin();
 404
 405   rgw_cls_list_op op;
 406   try {
 407     decode(op, iter);
 408   } catch (buffer::error& err) {
 409     CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode request\n");
 410     return -EINVAL;
 411   }
 412
 413   rgw_cls_list_ret ret;
 414   rgw_bucket_dir& new_dir = ret.dir;
 415   int rc = read_bucket_header(hctx, &new_dir.header);
 416   if (rc < 0) {
 417     CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to read header\n");
 418     return rc;
 419   }
 420
 421   map<string, bufferlist> keys;
 422   std::map<string, bufferlist>::iterator kiter;
 423   string start_key;
 424   encode_list_index_key(hctx, op.start_obj, &start_key);
 425   bool done = false;
 426   uint32_t left_to_read = op.num_entries;
 427   bool more;
 428
 429   do {
 430     rc = get_obj_vals(hctx, start_key, op.filter_prefix, left_to_read, &keys, &more);
 431     if (rc < 0)
 432       return rc;
 433
 434     std::map<string, rgw_bucket_dir_entry>& m = new_dir.m;
 435
 436     done = keys.empty();
 437
 438     for (kiter = keys.begin(); kiter != keys.end(); ++kiter) {
 439       rgw_bucket_dir_entry entry;
 440
 441       if (!bi_is_objs_index(kiter->first)) {
 442         done = true;
 443         break;
 444       }
 445
 446       bufferlist& entrybl = kiter->second;
 447       auto eiter = entrybl.cbegin();
 448       try {
 449         decode(entry, eiter);
 450       } catch (buffer::error& err) {
 451         CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode entry, key=%s\n", kiter->first.c_str());
 452         return -EINVAL;
 453       }
 454
 455       cls_rgw_obj_key key;
 456       uint64_t ver;
 457
 458       start_key = kiter->first;
 459       CLS_LOG(20, "start_key=%s len=%zu", start_key.c_str(), start_key.size());
 460
 461       int ret = decode_list_index_key(kiter->first, &key, &ver);
 462       if (ret < 0) {
 463         CLS_LOG(0, "ERROR: failed to decode list index key (%s)\n", escape_str(kiter->first).c_str());
 464         continue;
 465       }
 466
 467       if (!entry.is_valid()) {
 468         CLS_LOG(20, "entry %s[%s] is not valid\n", key.name.c_str(), key.instance.c_str());
 469         continue;
 470       }
 471
 472       // filter out noncurrent versions, delete markers, and initial marker
 473       if (!op.list_versions && (!entry.is_visible() || op.start_obj.name == key.name)) {
 474         CLS_LOG(20, "entry %s[%s] is not visible\n", key.name.c_str(), key.instance.c_str());
 475         continue;
 476       }
 477       if (m.size() < op.num_entries) {
 478         m[kiter->first] = entry;
 479       }
 480       left_to_read--;
 481
 482       CLS_LOG(20, "got entry %s[%s] m.size()=%d\n", key.name.c_str(), key.instance.c_str(), (int)m.size());
 483     }
 484   } while (left_to_read > 0 && !done);
 485
 486   ret.is_truncated = more && !done;
 487
 488   encode(ret, *out);
 489   return 0;
 490 }
 491
 492 static int check_index(cls_method_context_t hctx,
 493                        rgw_bucket_dir_header *existing_header,
 494                        rgw_bucket_dir_header *calc_header)
 495 {
 496   int rc = read_bucket_header(hctx, existing_header);
 497   if (rc < 0) {
 498     CLS_LOG(1, "ERROR: check_index(): failed to read header\n");
 499     return rc;
 500   }
 501
 502   calc_header->tag_timeout = existing_header->tag_timeout;
 503   calc_header->ver = existing_header->ver;
 504
 505   map<string, bufferlist> keys;
 506   string start_obj;
 507   string filter_prefix;
 508
 509 #define CHECK_CHUNK_SIZE 1000
 510   bool done = false;
 511   bool more;
 512
 513   do {
 514     rc = get_obj_vals(hctx, start_obj, filter_prefix, CHECK_CHUNK_SIZE, &keys, &more);
 515     if (rc < 0)
 516       return rc;
 517
 518     std::map<string, bufferlist>::iterator kiter = keys.begin();
 519     for (; kiter != keys.end(); ++kiter) {
 520       if (!bi_is_objs_index(kiter->first)) {
 521         done = true;
 522         break;
 523       }
 524
 525       rgw_bucket_dir_entry entry;
 526       auto eiter = kiter->second.cbegin();
 527       try {
 528         decode(entry, eiter);
 529       } catch (buffer::error& err) {
 530         CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode entry, key=%s\n", kiter->first.c_str());
 531         return -EIO;
 532       }
 533       rgw_bucket_category_stats& stats = calc_header->stats[entry.meta.category];
 534       stats.num_entries++;
 535       stats.total_size += entry.meta.accounted_size;
 536       stats.total_size_rounded += cls_rgw_get_rounded_size(entry.meta.accounted_size);
 537       stats.actual_size += entry.meta.size;
 538
 539       start_obj = kiter->first;
 540     }
 541   } while (keys.size() == CHECK_CHUNK_SIZE && !done);
 542
 543   return 0;
 544 }
 545
 546 int rgw_bucket_check_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 547 {
 548   rgw_cls_check_index_ret ret;
 549
 550   int rc = check_index(hctx, &ret.existing_header, &ret.calculated_header);
 551   if (rc < 0)
 552     return rc;
 553
 554   encode(ret, *out);
 555
 556   return 0;
 557 }
 558
 559 static int write_bucket_header(cls_method_context_t hctx, rgw_bucket_dir_header *header)
 560 {
 561   header->ver++;
 562
 563   bufferlist header_bl;
 564   encode(*header, header_bl);
 565   return cls_cxx_map_write_header(hctx, &header_bl);
 566 }
 567
 568
 569 int rgw_bucket_rebuild_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 570 {
 571   rgw_bucket_dir_header existing_header;
 572   rgw_bucket_dir_header calc_header;
 573   int rc = check_index(hctx, &existing_header, &calc_header);
 574   if (rc < 0)
 575     return rc;
 576
 577   return write_bucket_header(hctx, &calc_header);
 578 }
 579
 580 int rgw_bucket_update_stats(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 581 {
 582   // decode request
 583   rgw_cls_bucket_update_stats_op op;
 584   auto iter = in->cbegin();
 585   try {
 586     decode(op, iter);
 587   } catch (buffer::error& err) {
 588     CLS_LOG(1, "ERROR: %s(): failed to decode request\n", __func__);
 589     return -EINVAL;
 590   }
 591
 592   rgw_bucket_dir_header header;
 593   int rc = read_bucket_header(hctx, &header);
 594   if (rc < 0) {
 595     CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
 596     return rc;
 597   }
 598
 599   for (auto& s : op.stats) {
 600     auto& dest = header.stats[s.first];
 601     if (op.absolute) {
 602       dest = s.second;
 603     } else {
 604       dest.total_size += s.second.total_size;
 605       dest.total_size_rounded += s.second.total_size_rounded;
 606       dest.num_entries += s.second.num_entries;
 607       dest.actual_size += s.second.actual_size;
 608     }
 609   }
 610
 611   return write_bucket_header(hctx, &header);
 612 }
 613
 614 int rgw_bucket_init_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 615 {
 616   bufferlist header_bl;
 617   int rc = cls_cxx_map_read_header(hctx, &header_bl);
 618   if (rc < 0) {
 619     switch (rc) {
 620     case -ENODATA:
 621     case -ENOENT:
 622       break;
 623     default:
 624       return rc;
 625     }
 626   }
 627
 628   if (header_bl.length() != 0) {
 629     CLS_LOG(1, "ERROR: index already initialized\n");
 630     return -EINVAL;
 631   }
 632
 633   rgw_bucket_dir dir;
 634
 635   return write_bucket_header(hctx, &dir.header);
 636 }
 637
 638 int rgw_bucket_set_tag_timeout(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 639 {
 640   // decode request
 641   rgw_cls_tag_timeout_op op;
 642   auto iter = in->cbegin();
 643   try {
 644     decode(op, iter);
 645   } catch (buffer::error& err) {
 646     CLS_LOG(1, "ERROR: rgw_bucket_set_tag_timeout(): failed to decode request\n");
 647     return -EINVAL;
 648   }
 649
 650   rgw_bucket_dir_header header;
 651   int rc = read_bucket_header(hctx, &header);
 652   if (rc < 0) {
 653     CLS_LOG(1, "ERROR: rgw_bucket_set_tag_timeout(): failed to read header\n");
 654     return rc;
 655   }
 656
 657   header.tag_timeout = op.tag_timeout;
 658
 659   return write_bucket_header(hctx, &header);
 660 }
 661
 662 static int read_key_entry(cls_method_context_t hctx, cls_rgw_obj_key& key,
 663                           string *idx, rgw_bucket_dir_entry *entry,
 664                           bool special_delete_marker_name = false);
 665
 666 int rgw_bucket_prepare_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 667 {
 668   // decode request
 669   rgw_cls_obj_prepare_op op;
 670   auto iter = in->cbegin();
 671   try {
 672     decode(op, iter);
 673   } catch (buffer::error& err) {
 674     CLS_LOG(1, "ERROR: rgw_bucket_prepare_op(): failed to decode request\n");
 675     return -EINVAL;
 676   }
 677
 678   if (op.tag.empty()) {
 679     CLS_LOG(1, "ERROR: tag is empty\n");
 680     return -EINVAL;
 681   }
 682
 683   CLS_LOG(1, "rgw_bucket_prepare_op(): request: op=%d name=%s instance=%s tag=%s\n",
 684           op.op, op.key.name.c_str(), op.key.instance.c_str(), op.tag.c_str());
 685
 686   // get on-disk state
 687   string idx;
 688
 689   rgw_bucket_dir_entry entry;
 690   int rc = read_key_entry(hctx, op.key, &idx, &entry);
 691   if (rc < 0 && rc != -ENOENT)
 692     return rc;
 693
 694   bool noent = (rc == -ENOENT);
 695
 696   rc = 0;
 697
 698   if (noent) { // no entry, initialize fields
 699     entry.key = op.key;
 700     entry.ver = rgw_bucket_entry_ver();
 701     entry.exists = false;
 702     entry.locator = op.locator;
 703   }
 704
 705   // fill in proper state
 706   rgw_bucket_pending_info info;
 707   info.timestamp = real_clock::now();
 708   info.state = CLS_RGW_STATE_PENDING_MODIFY;
 709   info.op = op.op;
 710   entry.pending_map.insert(pair<string, rgw_bucket_pending_info>(op.tag, info));
 711
 712   rgw_bucket_dir_header header;
 713   rc = read_bucket_header(hctx, &header);
 714   if (rc < 0) {
 715     CLS_LOG(1, "ERROR: rgw_bucket_prepare_op(): failed to read header\n");
 716     return rc;
 717   }
 718
 719   if (op.log_op && !header.syncstopped) {
 720     rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime,
 721                              entry.ver, info.state, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
 722     if (rc < 0)
 723       return rc;
 724   }
 725
 726   // write out new key to disk
 727   bufferlist info_bl;
 728   encode(entry, info_bl);
 729   rc = cls_cxx_map_set_val(hctx, idx, &info_bl);
 730   if (rc < 0)
 731     return rc;
 732
 733   if (op.log_op && !header.syncstopped)
 734     return write_bucket_header(hctx, &header);
 735   return 0;
 736 }
 737
 738 static void unaccount_entry(rgw_bucket_dir_header& header,
 739                             rgw_bucket_dir_entry& entry)
 740 {
 741   rgw_bucket_category_stats& stats = header.stats[entry.meta.category];
 742   stats.num_entries--;
 743   stats.total_size -= entry.meta.accounted_size;
 744   stats.total_size_rounded -= cls_rgw_get_rounded_size(entry.meta.accounted_size);
 745   stats.actual_size -= entry.meta.size;
 746 }
 747
 748 static void log_entry(const char *func, const char *str, rgw_bucket_dir_entry *entry)
 749 {
 750   CLS_LOG(1, "%s(): %s: ver=%ld:%llu name=%s instance=%s locator=%s\n", func, str,
 751           (long)entry->ver.pool, (unsigned long long)entry->ver.epoch,
 752           entry->key.name.c_str(), entry->key.instance.c_str(), entry->locator.c_str());
 753 }
 754
 755 static void log_entry(const char *func, const char *str, rgw_bucket_olh_entry *entry)
 756 {
 757   CLS_LOG(1, "%s(): %s: epoch=%llu name=%s instance=%s tag=%s\n", func, str,
 758           (unsigned long long)entry->epoch, entry->key.name.c_str(), entry->key.instance.c_str(),
 759           entry->tag.c_str());
 760 }
 761
 762 template <class T>
 763 static int read_omap_entry(cls_method_context_t hctx, const std::string& name,
 764                            T* entry)
 765 {
 766   bufferlist current_entry;
 767   int rc = cls_cxx_map_get_val(hctx, name, &current_entry);
 768   if (rc < 0) {
 769     return rc;
 770   }
 771
 772   auto cur_iter = current_entry.cbegin();
 773   try {
 774     decode(*entry, cur_iter);
 775   } catch (buffer::error& err) {
 776     CLS_LOG(1, "ERROR: %s(): failed to decode entry\n", __func__);
 777     return -EIO;
 778   }
 779   return 0;
 780 }
 781
 782 template <class T>
 783 static int read_index_entry(cls_method_context_t hctx, string& name, T* entry)
 784 {
 785   int ret = read_omap_entry(hctx, name, entry);
 786   if (ret < 0) {
 787     return ret;
 788   }
 789
 790   log_entry(__func__, "existing entry", entry);
 791   return 0;
 792 }
 793
 794 static int read_key_entry(cls_method_context_t hctx, cls_rgw_obj_key& key,
 795                           string *idx, rgw_bucket_dir_entry *entry,
 796                           bool special_delete_marker_name)
 797 {
 798   encode_obj_index_key(key, idx);
 799   int rc = read_index_entry(hctx, *idx, entry);
 800   if (rc < 0) {
 801     return rc;
 802   }
 803
 804   if (key.instance.empty() &&
 805       entry->flags & RGW_BUCKET_DIRENT_FLAG_VER_MARKER) {
 806     /* we only do it where key.instance is empty. In this case the delete marker will have a
 807      * separate entry in the index to avoid collisions with the actual object, as it's mutable
 808      */
 809     if (special_delete_marker_name) {
 810       encode_obj_versioned_data_key(key, idx, true);
 811       rc = read_index_entry(hctx, *idx, entry);
 812       if (rc == 0) {
 813         return 0;
 814       }
 815     }
 816     encode_obj_versioned_data_key(key, idx);
 817     rc = read_index_entry(hctx, *idx, entry);
 818     if (rc < 0) {
 819       *entry = rgw_bucket_dir_entry(); /* need to reset entry because we initialized it earlier */
 820       return rc;
 821     }
 822   }
 823
 824   return 0;
 825 }
 826
 827 int rgw_bucket_complete_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
 828 {
 829   // decode request
 830   rgw_cls_obj_complete_op op;
 831   auto iter = in->cbegin();
 832   try {
 833     decode(op, iter);
 834   } catch (buffer::error& err) {
 835     CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to decode request\n");
 836     return -EINVAL;
 837   }
 838   CLS_LOG(1, "rgw_bucket_complete_op(): request: op=%d name=%s instance=%s ver=%lu:%llu tag=%s\n",
 839           op.op, op.key.name.c_str(), op.key.instance.c_str(),
 840           (unsigned long)op.ver.pool, (unsigned long long)op.ver.epoch,
 841           op.tag.c_str());
 842
 843   rgw_bucket_dir_header header;
 844   int rc = read_bucket_header(hctx, &header);
 845   if (rc < 0) {
 846     CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n");
 847     return -EINVAL;
 848   }
 849
 850   rgw_bucket_dir_entry entry;
 851   bool ondisk = true;
 852
 853   string idx;
 854   rc = read_key_entry(hctx, op.key, &idx, &entry);
 855   if (rc == -ENOENT) {
 856     entry.key = op.key;
 857     entry.ver = op.ver;
 858     entry.meta = op.meta;
 859     entry.locator = op.locator;
 860     ondisk = false;
 861   } else if (rc < 0) {
 862     return rc;
 863   }
 864
 865   entry.index_ver = header.ver;
 866   entry.flags = (entry.key.instance.empty() ? 0 : RGW_BUCKET_DIRENT_FLAG_VER); /* resetting entry flags, entry might have been previously a delete marker */
 867
 868   if (op.tag.size()) {
 869     map<string, rgw_bucket_pending_info>::iterator pinter = entry.pending_map.find(op.tag);
 870     if (pinter == entry.pending_map.end()) {
 871       CLS_LOG(1, "ERROR: couldn't find tag for pending operation\n");
 872       return -EINVAL;
 873     }
 874     entry.pending_map.erase(pinter);
 875   }
 876
 877   bool cancel = false;
 878   bufferlist update_bl;
 879
 880   if (op.tag.size() && op.op == CLS_RGW_OP_CANCEL) {
 881     CLS_LOG(1, "rgw_bucket_complete_op(): cancel requested\n");
 882     cancel = true;
 883   } else if (op.ver.pool == entry.ver.pool &&
 884              op.ver.epoch && op.ver.epoch <= entry.ver.epoch) {
 885     CLS_LOG(1, "rgw_bucket_complete_op(): skipping request, old epoch\n");
 886     cancel = true;
 887   }
 888
 889   bufferlist op_bl;
 890   if (cancel) {
 891     if (op.log_op && !header.syncstopped) {
 892       rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, entry.ver,
 893                                CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
 894       if (rc < 0)
 895         return rc;
 896     }
 897
 898     if (op.tag.size()) {
 899       bufferlist new_key_bl;
 900       encode(entry, new_key_bl);
 901       rc = cls_cxx_map_set_val(hctx, idx, &new_key_bl);
 902       if (rc < 0)
 903         return rc;
 904     }
 905
 906     if (op.log_op && !header.syncstopped) {
 907       return write_bucket_header(hctx, &header);
 908     }
 909     return 0;
 910   }
 911
 912   if (entry.exists) {
 913     unaccount_entry(header, entry);
 914   }
 915
 916   entry.ver = op.ver;
 917   switch ((int)op.op) {
 918   case CLS_RGW_OP_DEL:
 919     entry.meta = op.meta;
 920     if (ondisk) {
 921       if (!entry.pending_map.size()) {
 922         int ret = cls_cxx_map_remove_key(hctx, idx);
 923         if (ret < 0)
 924           return ret;
 925       } else {
 926         entry.exists = false;
 927         bufferlist new_key_bl;
 928         encode(entry, new_key_bl);
 929         int ret = cls_cxx_map_set_val(hctx, idx, &new_key_bl);
 930         if (ret < 0)
 931           return ret;
 932       }
 933     } else {
 934       return -ENOENT;
 935     }
 936     break;
 937   case CLS_RGW_OP_ADD:
 938     {
 939       rgw_bucket_dir_entry_meta& meta = op.meta;
 940       rgw_bucket_category_stats& stats = header.stats[meta.category];
 941       entry.meta = meta;
 942       entry.key = op.key;
 943       entry.exists = true;
 944       entry.tag = op.tag;
 945       stats.num_entries++;
 946       stats.total_size += meta.accounted_size;
 947       stats.total_size_rounded += cls_rgw_get_rounded_size(meta.accounted_size);
 948       stats.actual_size += meta.size;
 949       bufferlist new_key_bl;
 950       encode(entry, new_key_bl);
 951       int ret = cls_cxx_map_set_val(hctx, idx, &new_key_bl);
 952       if (ret < 0)
 953         return ret;
 954     }
 955     break;
 956   }
 957
 958   if (op.log_op && !header.syncstopped) {
 959     rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, entry.ver,
 960                              CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
 961     if (rc < 0)
 962       return rc;
 963   }
 964
 965   list<cls_rgw_obj_key>::iterator remove_iter;
 966   CLS_LOG(20, "rgw_bucket_complete_op(): remove_objs.size()=%d\n", (int)op.remove_objs.size());
 967   for (remove_iter = op.remove_objs.begin(); remove_iter != op.remove_objs.end(); ++remove_iter) {
 968     cls_rgw_obj_key& remove_key = *remove_iter;
 969     CLS_LOG(1, "rgw_bucket_complete_op(): removing entries, read_index_entry name=%s instance=%s\n",
 970             remove_key.name.c_str(), remove_key.instance.c_str());
 971     rgw_bucket_dir_entry remove_entry;
 972     string k;
 973     int ret = read_key_entry(hctx, remove_key, &k, &remove_entry);
 974     if (ret < 0) {
 975       CLS_LOG(1, "rgw_bucket_complete_op(): removing entries, read_index_entry name=%s instance=%s ret=%d\n",
 976             remove_key.name.c_str(), remove_key.instance.c_str(), ret);
 977       continue;
 978     }
 979     CLS_LOG(0,
 980             "rgw_bucket_complete_op(): entry.name=%s entry.instance=%s entry.meta.category=%d\n",
 981             remove_entry.key.name.c_str(),
 982             remove_entry.key.instance.c_str(),
 983             int(remove_entry.meta.category));
 984     unaccount_entry(header, remove_entry);
 985
 986     if (op.log_op && !header.syncstopped) {
 987       ++header.ver; // increment index version, or we'll overwrite keys previously written
 988       rc = log_index_operation(hctx, remove_key, CLS_RGW_OP_DEL, op.tag, remove_entry.meta.mtime,
 989                                remove_entry.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
 990       if (rc < 0)
 991         continue;
 992     }
 993
 994     ret = cls_cxx_map_remove_key(hctx, k);
 995     if (ret < 0) {
 996       CLS_LOG(1, "rgw_bucket_complete_op(): cls_cxx_map_remove_key, failed to remove entry, name=%s instance=%s read_index_entry ret=%d\n", remove_key.name.c_str(), remove_key.instance.c_str(), rc);
 997       continue;
 998     }
 999   }
1000
1001   return write_bucket_header(hctx, &header);
1002 }
1003
1004 template <class T>
1005 static int write_entry(cls_method_context_t hctx, T& entry, const string& key)
1006 {
1007   bufferlist bl;
1008   encode(entry, bl);
1009   return cls_cxx_map_set_val(hctx, key, &bl);
1010 }
1011
1012 static int read_olh(cls_method_context_t hctx,cls_rgw_obj_key& obj_key, rgw_bucket_olh_entry *olh_data_entry, string *index_key, bool *found)
1013 {
1014   cls_rgw_obj_key olh_key;
1015   olh_key.name = obj_key.name;
1016
1017   encode_olh_data_key(olh_key, index_key);
1018   int ret = read_index_entry(hctx, *index_key, olh_data_entry);
1019   if (ret < 0 && ret != -ENOENT) {
1020     CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_key.name.c_str(), ret);
1021     return ret;
1022   }
1023   if (found) {
1024     *found = (ret != -ENOENT);
1025   }
1026   return 0;
1027 }
1028
1029 static void update_olh_log(rgw_bucket_olh_entry& olh_data_entry, OLHLogOp op, const string& op_tag,
1030                            cls_rgw_obj_key& key, bool delete_marker, uint64_t epoch)
1031 {
1032   vector<rgw_bucket_olh_log_entry>& log = olh_data_entry.pending_log[olh_data_entry.epoch];
1033   rgw_bucket_olh_log_entry log_entry;
1034   log_entry.epoch = epoch;
1035   log_entry.op = op;
1036   log_entry.op_tag = op_tag;
1037   log_entry.key = key;
1038   log_entry.delete_marker = delete_marker;
1039   log.push_back(log_entry);
1040 }
1041
1042 static int write_obj_instance_entry(cls_method_context_t hctx, rgw_bucket_dir_entry& instance_entry, const string& instance_idx)
1043 {
1044   CLS_LOG(20, "write_entry() instance=%s idx=%s flags=%d", escape_str(instance_entry.key.instance).c_str(), instance_idx.c_str(), instance_entry.flags);
1045   /* write the instance entry */
1046   int ret = write_entry(hctx, instance_entry, instance_idx);
1047   if (ret < 0) {
1048     CLS_LOG(0, "ERROR: write_entry() instance_key=%s ret=%d", escape_str(instance_idx).c_str(), ret);
1049     return ret;
1050   }
1051   return 0;
1052 }
1053
1054 /*
1055  * write object instance entry, and if needed also the list entry
1056  */
1057 static int write_obj_entries(cls_method_context_t hctx, rgw_bucket_dir_entry& instance_entry, const string& instance_idx)
1058 {
1059   int ret = write_obj_instance_entry(hctx, instance_entry, instance_idx);
1060   if (ret < 0) {
1061     return ret;
1062   }
1063   string instance_list_idx;
1064   get_list_index_key(instance_entry, &instance_list_idx);
1065
1066   if (instance_idx != instance_list_idx) {
1067     CLS_LOG(20, "write_entry() idx=%s flags=%d", escape_str(instance_list_idx).c_str(), instance_entry.flags);
1068     /* write a new list entry for the object instance */
1069     ret = write_entry(hctx, instance_entry, instance_list_idx);
1070     if (ret < 0) {
1071       CLS_LOG(0, "ERROR: write_entry() instance=%s instance_list_idx=%s ret=%d", instance_entry.key.instance.c_str(), instance_list_idx.c_str(), ret);
1072       return ret;
1073     }
1074   }
1075   return 0;
1076 }
1077
1078
1079 class BIVerObjEntry {
1080   cls_method_context_t hctx;
1081   cls_rgw_obj_key key;
1082   string instance_idx;
1083
1084   rgw_bucket_dir_entry instance_entry;
1085
1086   bool initialized;
1087
1088 public:
1089   BIVerObjEntry(cls_method_context_t& _hctx, const cls_rgw_obj_key& _key) : hctx(_hctx), key(_key), initialized(false) {
1090   }
1091
1092   int init(bool check_delete_marker = true) {
1093     int ret = read_key_entry(hctx, key, &instance_idx, &instance_entry,
1094                              check_delete_marker && key.instance.empty()); /* this is potentially a delete marker, for null objects we
1095                                                                               keep separate instance entry for the delete markers */
1096
1097     if (ret < 0) {
1098       CLS_LOG(0, "ERROR: read_key_entry() idx=%s ret=%d", instance_idx.c_str(), ret);
1099       return ret;
1100     }
1101     initialized = true;
1102     CLS_LOG(20, "read instance_entry key.name=%s key.instance=%s flags=%d", instance_entry.key.name.c_str(), instance_entry.key.instance.c_str(), instance_entry.flags);
1103     return 0;
1104   }
1105
1106   rgw_bucket_dir_entry& get_dir_entry() {
1107     return instance_entry;
1108   }
1109
1110   void init_as_delete_marker(rgw_bucket_dir_entry_meta& meta) {
1111     /* a deletion marker, need to initialize it, there's no instance entry for it yet */
1112     instance_entry.key = key;
1113     instance_entry.flags = RGW_BUCKET_DIRENT_FLAG_DELETE_MARKER;
1114     instance_entry.meta = meta;
1115     instance_entry.tag = "delete-marker";
1116
1117     initialized = true;
1118   }
1119
1120   void set_epoch(uint64_t epoch) {
1121     instance_entry.versioned_epoch = epoch;
1122   }
1123
1124   int unlink_list_entry() {
1125     string list_idx;
1126     /* this instance has a previous list entry, remove that entry */
1127     get_list_index_key(instance_entry, &list_idx);
1128     CLS_LOG(20, "unlink_list_entry() list_idx=%s", escape_str(list_idx).c_str());
1129     int ret = cls_cxx_map_remove_key(hctx, list_idx);
1130     if (ret < 0) {
1131       CLS_LOG(0, "ERROR: cls_cxx_map_remove_key() list_idx=%s ret=%d", list_idx.c_str(), ret);
1132       return ret;
1133     }
1134     return 0;
1135   }
1136
1137   int unlink() {
1138     /* remove the instance entry */
1139     CLS_LOG(20, "unlink() idx=%s", escape_str(instance_idx).c_str());
1140     int ret = cls_cxx_map_remove_key(hctx, instance_idx);
1141     if (ret < 0) {
1142       CLS_LOG(0, "ERROR: cls_cxx_map_remove_key() instance_idx=%s ret=%d", instance_idx.c_str(), ret);
1143       return ret;
1144     }
1145     return 0;
1146   }
1147
1148   int write_entries(uint64_t flags_set, uint64_t flags_reset) {
1149     if (!initialized) {
1150       int ret = init();
1151       if (ret < 0) {
1152         return ret;
1153       }
1154     }
1155     instance_entry.flags &= ~flags_reset;
1156     instance_entry.flags |= flags_set;
1157
1158     /* write the instance and list entries */
1159     bool special_delete_marker_key = (instance_entry.is_delete_marker() && instance_entry.key.instance.empty());
1160     encode_obj_versioned_data_key(key, &instance_idx, special_delete_marker_key);
1161     int ret = write_obj_entries(hctx, instance_entry, instance_idx);
1162     if (ret < 0) {
1163       CLS_LOG(0, "ERROR: write_obj_entries() instance_idx=%s ret=%d", instance_idx.c_str(), ret);
1164       return ret;
1165     }
1166
1167     return 0;
1168   }
1169
1170   int write(uint64_t epoch, bool current) {
1171     if (instance_entry.versioned_epoch > 0) {
1172       CLS_LOG(20, "%s(): instance_entry.versioned_epoch=%d epoch=%d", __func__, (int)instance_entry.versioned_epoch, (int)epoch);
1173       /* this instance has a previous list entry, remove that entry */
1174       int ret = unlink_list_entry();
1175       if (ret < 0) {
1176         return ret;
1177       }
1178     }
1179
1180     uint64_t flags = RGW_BUCKET_DIRENT_FLAG_VER;
1181     if (current) {
1182       flags |= RGW_BUCKET_DIRENT_FLAG_CURRENT;
1183     }
1184
1185     instance_entry.versioned_epoch = epoch;
1186     return write_entries(flags, 0);
1187   }
1188
1189   int demote_current() {
1190     return write_entries(0, RGW_BUCKET_DIRENT_FLAG_CURRENT);
1191   }
1192
1193   bool is_delete_marker() {
1194     return instance_entry.is_delete_marker();
1195   }
1196
1197   int find_next_key(cls_rgw_obj_key *next_key, bool *found) {
1198     string list_idx;
1199     /* this instance has a previous list entry, remove that entry */
1200     get_list_index_key(instance_entry, &list_idx);
1201     /* this is the current head, need to update! */
1202     map<string, bufferlist> keys;
1203     bool more;
1204     string filter = key.name; /* list key starts with key name, filter it to avoid a case where we cross to
1205                                  different namespace */
1206     int ret = cls_cxx_map_get_vals(hctx, list_idx, filter, 1, &keys, &more);
1207     if (ret < 0) {
1208       return ret;
1209     }
1210
1211     if (keys.size() < 1) {
1212       *found = false;
1213       return 0;
1214     }
1215
1216     rgw_bucket_dir_entry next_entry;
1217
1218     map<string, bufferlist>::reverse_iterator last = keys.rbegin();
1219     try {
1220       auto iter = last->second.cbegin();
1221       decode(next_entry, iter);
1222     } catch (buffer::error& err) {
1223       CLS_LOG(0, "ERROR; failed to decode entry: %s", last->first.c_str());
1224       return -EIO;
1225     }
1226
1227     *found = (key.name == next_entry.key.name);
1228     if (*found) {
1229       *next_key = next_entry.key;
1230     }
1231
1232     return 0;
1233   }
1234
1235   real_time mtime() {
1236     return instance_entry.meta.mtime;
1237   }
1238 };
1239
1240
1241 class BIOLHEntry {
1242   cls_method_context_t hctx;
1243   cls_rgw_obj_key key;
1244
1245   string olh_data_idx;
1246   rgw_bucket_olh_entry olh_data_entry;
1247
1248   bool initialized;
1249 public:
1250   BIOLHEntry(cls_method_context_t& _hctx, const cls_rgw_obj_key& _key) : hctx(_hctx), key(_key), initialized(false) { }
1251
1252   int init(bool *exists) {
1253     /* read olh */
1254     int ret = read_olh(hctx, key, &olh_data_entry, &olh_data_idx, exists);
1255     if (ret < 0) {
1256       return ret;
1257     }
1258
1259     initialized = true;
1260     return 0;
1261   }
1262
1263   bool start_modify(uint64_t candidate_epoch) {
1264     if (candidate_epoch) {
1265       if (candidate_epoch < olh_data_entry.epoch) {
1266         return false; /* olh cannot be modified, old epoch */
1267       }
1268       olh_data_entry.epoch = candidate_epoch;
1269     } else {
1270       if (olh_data_entry.epoch == 0) {
1271         olh_data_entry.epoch = 2; /* versioned epoch should start with 2, 1 is reserved to converted plain entries */
1272       } else {
1273         olh_data_entry.epoch++;
1274       }
1275     }
1276     return true;
1277   }
1278
1279   uint64_t get_epoch() {
1280     return olh_data_entry.epoch;
1281   }
1282
1283   rgw_bucket_olh_entry& get_entry() {
1284     return olh_data_entry;
1285   }
1286
1287   void update(cls_rgw_obj_key& key, bool delete_marker) {
1288     olh_data_entry.delete_marker = delete_marker;
1289     olh_data_entry.key = key;
1290   }
1291
1292   int write() {
1293     /* write the olh data entry */
1294     int ret = write_entry(hctx, olh_data_entry, olh_data_idx);
1295     if (ret < 0) {
1296       CLS_LOG(0, "ERROR: write_entry() olh_key=%s ret=%d", olh_data_idx.c_str(), ret);
1297       return ret;
1298     }
1299
1300     return 0;
1301   }
1302
1303   void update_log(OLHLogOp op, const string& op_tag, cls_rgw_obj_key& key, bool delete_marker, uint64_t epoch = 0) {
1304     if (epoch == 0) {
1305       epoch = olh_data_entry.epoch;
1306     }
1307     update_olh_log(olh_data_entry, op, op_tag, key, delete_marker, epoch);
1308   }
1309
1310   bool exists() { return olh_data_entry.exists; }
1311
1312   void set_exists(bool exists) {
1313     olh_data_entry.exists = exists;
1314   }
1315
1316   bool pending_removal() { return olh_data_entry.pending_removal; }
1317
1318   void set_pending_removal(bool pending_removal) {
1319     olh_data_entry.pending_removal = pending_removal;
1320   }
1321
1322   const string& get_tag() { return olh_data_entry.tag; }
1323   void set_tag(const string& tag) {
1324     olh_data_entry.tag = tag;
1325   }
1326 };
1327
1328 static int write_version_marker(cls_method_context_t hctx, cls_rgw_obj_key& key)
1329 {
1330   rgw_bucket_dir_entry entry;
1331   entry.key = key;
1332   entry.flags = RGW_BUCKET_DIRENT_FLAG_VER_MARKER;
1333   int ret = write_entry(hctx, entry, key.name);
1334   if (ret < 0) {
1335     CLS_LOG(0, "ERROR: write_entry returned ret=%d", ret);
1336     return ret;
1337   }
1338   return 0;
1339 }
1340
1341 /*
1342  * plain entries are the ones who were created when bucket was not versioned,
1343  * if we override these objects, we need to convert these to versioned entries -- ones that have
1344  * both data entry, and listing key. Their version is going to be empty though
1345  */
1346 static int convert_plain_entry_to_versioned(cls_method_context_t hctx, cls_rgw_obj_key& key, bool demote_current, bool instance_only)
1347 {
1348   if (!key.instance.empty()) {
1349     return -EINVAL;
1350   }
1351
1352   rgw_bucket_dir_entry entry;
1353
1354   string orig_idx;
1355   int ret = read_key_entry(hctx, key, &orig_idx, &entry);
1356   if (ret != -ENOENT) {
1357     if (ret < 0) {
1358       CLS_LOG(0, "ERROR: read_key_entry() returned ret=%d", ret);
1359       return ret;
1360     }
1361
1362     entry.versioned_epoch = 1; /* converted entries are always 1 */
1363     entry.flags |= RGW_BUCKET_DIRENT_FLAG_VER;
1364
1365     if (demote_current) {
1366       entry.flags &= ~RGW_BUCKET_DIRENT_FLAG_CURRENT;
1367     }
1368
1369     string new_idx;
1370     encode_obj_versioned_data_key(key, &new_idx);
1371
1372     if (instance_only) {
1373       ret = write_obj_instance_entry(hctx, entry, new_idx);
1374     } else {
1375       ret = write_obj_entries(hctx, entry, new_idx);
1376     }
1377     if (ret < 0) {
1378       CLS_LOG(0, "ERROR: write_obj_entries new_idx=%s returned %d", new_idx.c_str(), ret);
1379       return ret;
1380     }
1381   }
1382
1383   ret = write_version_marker(hctx, key);
1384   if (ret < 0) {
1385     return ret;
1386   }
1387
1388   return 0;
1389 }
1390
1391 /*
1392  * link an object version to an olh, update the relevant index entries. It will also handle the
1393  * deletion marker case. We have a few entries that we need to take care of. For object 'foo',
1394  * instance BAR, we'd update the following (not actual encoding):
1395  *  - olh data: [BI_BUCKET_OLH_DATA_INDEX]foo
1396  *  - object instance data: [BI_BUCKET_OBJ_INSTANCE_INDEX]foo,BAR
1397  *  - object instance list entry: foo,123,BAR
1398  *
1399  *  The instance list entry needs to be ordered by newer to older, so we generate an appropriate
1400  *  number string that follows the name.
1401  *  The top instance for each object is marked appropriately.
1402  *  We generate instance entry for deletion markers here, as they are not created prior.
1403  */
1404 static int rgw_bucket_link_olh(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1405 {
1406   string olh_data_idx;
1407   string instance_idx;
1408
1409   // decode request
1410   rgw_cls_link_olh_op op;
1411   auto iter = in->cbegin();
1412   try {
1413     decode(op, iter);
1414   } catch (buffer::error& err) {
1415     CLS_LOG(0, "ERROR: rgw_bucket_link_olh_op(): failed to decode request\n");
1416     return -EINVAL;
1417   }
1418
1419   BIVerObjEntry obj(hctx, op.key);
1420   BIOLHEntry olh(hctx, op.key);
1421
1422   /* read instance entry */
1423   int ret = obj.init(op.delete_marker);
1424   bool existed = (ret == 0);
1425   if (ret == -ENOENT && op.delete_marker) {
1426     ret = 0;
1427   }
1428   if (ret < 0) {
1429     return ret;
1430   }
1431
1432   if (existed && !real_clock::is_zero(op.unmod_since)) {
1433     timespec mtime = ceph::real_clock::to_timespec(obj.mtime());
1434     timespec unmod = ceph::real_clock::to_timespec(op.unmod_since);
1435     if (!op.high_precision_time) {
1436       mtime.tv_nsec = 0;
1437       unmod.tv_nsec = 0;
1438     }
1439     if (mtime >= unmod) {
1440       return 0; /* no need to set error, we just return 0 and avoid writing to the bi log */
1441     }
1442   }
1443
1444   bool removing;
1445
1446   /*
1447    * Special handling for null instance object / delete-marker. For these objects we're going to
1448    * have separate instances for a data object vs. delete-marker to avoid collisions. We now check
1449    * if we got to overwrite a previous entry, and in that case we'll remove its list entry.
1450    */
1451   if (op.key.instance.empty()) {
1452     BIVerObjEntry other_obj(hctx, op.key);
1453     ret = other_obj.init(!op.delete_marker); /* try reading the other null versioned entry */
1454     existed = (ret >= 0 && !other_obj.is_delete_marker());
1455     if (ret >= 0 && other_obj.is_delete_marker() != op.delete_marker) {
1456       ret = other_obj.unlink_list_entry();
1457       if (ret < 0) {
1458         return ret;
1459       }
1460     }
1461
1462     removing = existed && op.delete_marker;
1463     if (!removing) {
1464       ret = other_obj.unlink();
1465       if (ret < 0) {
1466         return ret;
1467       }
1468     }
1469   } else {
1470     removing = (existed && !obj.is_delete_marker() && op.delete_marker);
1471   }
1472
1473   if (op.delete_marker) {
1474     /* a deletion marker, need to initialize entry as such */
1475     obj.init_as_delete_marker(op.meta);
1476   }
1477
1478   /* read olh */
1479   bool olh_found;
1480   ret = olh.init(&olh_found);
1481   if (ret < 0) {
1482     return ret;
1483   }
1484
1485   if (!olh.start_modify(op.olh_epoch)) {
1486     ret = obj.write(op.olh_epoch, false);
1487     if (ret < 0) {
1488       return ret;
1489     }
1490     if (removing) {
1491       olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false, op.olh_epoch);
1492     }
1493     return 0;
1494   }
1495
1496   if (olh_found) {
1497     const string& olh_tag = olh.get_tag();
1498     if (op.olh_tag != olh_tag) {
1499       if (!olh.pending_removal()) {
1500         CLS_LOG(5, "NOTICE: op.olh_tag (%s) != olh.tag (%s)", op.olh_tag.c_str(), olh_tag.c_str());
1501         return -ECANCELED;
1502       }
1503       /* if pending removal, this is a new olh instance */
1504       olh.set_tag(op.olh_tag);
1505     }
1506     if (olh.exists()) {
1507       rgw_bucket_olh_entry& olh_entry = olh.get_entry();
1508       /* found olh, previous instance is no longer the latest, need to update */
1509       if (!(olh_entry.key == op.key)) {
1510         BIVerObjEntry old_obj(hctx, olh_entry.key);
1511
1512         ret = old_obj.demote_current();
1513         if (ret < 0) {
1514           CLS_LOG(0, "ERROR: could not demote current on previous key ret=%d", ret);
1515           return ret;
1516         }
1517       }
1518     }
1519     olh.set_pending_removal(false);
1520   } else {
1521     bool instance_only = (op.key.instance.empty() && op.delete_marker);
1522     cls_rgw_obj_key key(op.key.name);
1523     ret = convert_plain_entry_to_versioned(hctx, key, true, instance_only);
1524     if (ret < 0) {
1525       CLS_LOG(0, "ERROR: convert_plain_entry_to_versioned ret=%d", ret);
1526       return ret;
1527     }
1528     olh.set_tag(op.olh_tag);
1529   }
1530
1531   /* update the olh log */
1532   olh.update_log(CLS_RGW_OLH_OP_LINK_OLH, op.op_tag, op.key, op.delete_marker);
1533   if (removing) {
1534     olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false);
1535   }
1536
1537   olh.update(op.key, op.delete_marker);
1538
1539   olh.set_exists(true);
1540
1541   ret = olh.write();
1542   if (ret < 0) {
1543     CLS_LOG(0, "ERROR: failed to update olh ret=%d", ret);
1544     return ret;
1545   }
1546
1547   /* write the instance and list entries */
1548   ret = obj.write(olh.get_epoch(), true);
1549   if (ret < 0) {
1550     return ret;
1551   }
1552
1553   rgw_bucket_dir_header header;
1554   ret = read_bucket_header(hctx, &header);
1555   if (ret < 0) {
1556     CLS_LOG(1, "ERROR: rgw_bucket_link_olh(): failed to read header\n");
1557     return ret;
1558   }
1559
1560   if (op.log_op && !header.syncstopped) {
1561     rgw_bucket_dir_entry& entry = obj.get_dir_entry();
1562
1563     rgw_bucket_entry_ver ver;
1564     ver.epoch = (op.olh_epoch ? op.olh_epoch : olh.get_epoch());
1565
1566     string *powner = NULL;
1567     string *powner_display_name = NULL;
1568
1569     if (op.delete_marker) {
1570       powner = &entry.meta.owner;
1571       powner_display_name = &entry.meta.owner_display_name;
1572     }
1573
1574     RGWModifyOp operation = (op.delete_marker ? CLS_RGW_OP_LINK_OLH_DM : CLS_RGW_OP_LINK_OLH);
1575     ret = log_index_operation(hctx, op.key, operation, op.op_tag,
1576                               entry.meta.mtime, ver,
1577                               CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags | RGW_BILOG_FLAG_VERSIONED_OP,
1578                               powner, powner_display_name, &op.zones_trace);
1579     if (ret < 0)
1580       return ret;
1581
1582     return write_bucket_header(hctx, &header); /* updates header version */
1583   }
1584
1585   return 0;
1586 }
1587
1588 static int rgw_bucket_unlink_instance(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1589 {
1590   string olh_data_idx;
1591   string instance_idx;
1592
1593   // decode request
1594   rgw_cls_unlink_instance_op op;
1595   auto iter = in->cbegin();
1596   try {
1597     decode(op, iter);
1598   } catch (buffer::error& err) {
1599     CLS_LOG(0, "ERROR: rgw_bucket_rm_obj_instance_op(): failed to decode request\n");
1600     return -EINVAL;
1601   }
1602
1603   cls_rgw_obj_key dest_key = op.key;
1604   if (dest_key.instance == "null") {
1605     dest_key.instance.clear();
1606   }
1607
1608   BIVerObjEntry obj(hctx, dest_key);
1609   BIOLHEntry olh(hctx, dest_key);
1610
1611   int ret = obj.init();
1612   if (ret == -ENOENT) {
1613     return 0; /* already removed */
1614   }
1615   if (ret < 0) {
1616     CLS_LOG(0, "ERROR: obj.init() returned ret=%d", ret);
1617     return ret;
1618   }
1619
1620   bool olh_found;
1621   ret = olh.init(&olh_found);
1622   if (ret < 0) {
1623     CLS_LOG(0, "ERROR: olh.init() returned ret=%d", ret);
1624     return ret;
1625   }
1626
1627   if (!olh_found) {
1628     bool instance_only = false;
1629     cls_rgw_obj_key key(dest_key.name);
1630     ret = convert_plain_entry_to_versioned(hctx, key, true, instance_only);
1631     if (ret < 0) {
1632       CLS_LOG(0, "ERROR: convert_plain_entry_to_versioned ret=%d", ret);
1633       return ret;
1634     }
1635     olh.update(dest_key, false);
1636     olh.set_tag(op.olh_tag);
1637
1638     obj.set_epoch(1);
1639   }
1640
1641   if (!olh.start_modify(op.olh_epoch)) {
1642     ret = obj.unlink_list_entry();
1643     if (ret < 0) {
1644       return ret;
1645     }
1646
1647     if (!obj.is_delete_marker()) {
1648       olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false, op.olh_epoch);
1649     }
1650
1651     return 0;
1652   }
1653
1654   rgw_bucket_olh_entry& olh_entry = olh.get_entry();
1655   cls_rgw_obj_key& olh_key = olh_entry.key;
1656   CLS_LOG(20, "%s(): updating olh log: existing olh entry: %s[%s] (delete_marker=%d)", __func__,
1657              olh_key.name.c_str(), olh_key.instance.c_str(), olh_entry.delete_marker);
1658
1659   if (olh_key == dest_key) {
1660     /* this is the current head, need to update! */
1661     cls_rgw_obj_key next_key;
1662     bool found = false;
1663     ret = obj.find_next_key(&next_key, &found);
1664     if (ret < 0) {
1665       CLS_LOG(0, "ERROR: obj.find_next_key() returned ret=%d", ret);
1666       return ret;
1667     }
1668
1669     if (found) {
1670       BIVerObjEntry next(hctx, next_key);
1671       ret = next.write(olh.get_epoch(), true);
1672       if (ret < 0) {
1673         CLS_LOG(0, "ERROR: next.write() returned ret=%d", ret);
1674         return ret;
1675       }
1676
1677       CLS_LOG(20, "%s(): updating olh log: link olh -> %s[%s] (is_delete=%d)", __func__,
1678               next_key.name.c_str(), next_key.instance.c_str(), (int)next.is_delete_marker());
1679
1680       olh.update(next_key, next.is_delete_marker());
1681       olh.update_log(CLS_RGW_OLH_OP_LINK_OLH, op.op_tag, next_key, next.is_delete_marker());
1682     } else {
1683       /* next_key is empty */
1684       olh.update(next_key, false);
1685       olh.update_log(CLS_RGW_OLH_OP_UNLINK_OLH, op.op_tag, next_key, false);
1686       olh.set_exists(false);
1687       olh.set_pending_removal(true);
1688     }
1689   }
1690
1691   if (!obj.is_delete_marker()) {
1692     olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false);
1693   } else {
1694     /* this is a delete marker, it's our responsibility to remove its instance entry */
1695     ret = obj.unlink();
1696     if (ret < 0) {
1697       return ret;
1698     }
1699   }
1700
1701   ret = obj.unlink_list_entry();
1702   if (ret < 0) {
1703     return ret;
1704   }
1705
1706   ret = olh.write();
1707   if (ret < 0) {
1708     return ret;
1709   }
1710
1711   rgw_bucket_dir_header header;
1712   ret = read_bucket_header(hctx, &header);
1713   if (ret < 0) {
1714     CLS_LOG(1, "ERROR: rgw_bucket_unlink_instance(): failed to read header\n");
1715     return ret;
1716   }
1717
1718   if (op.log_op && !header.syncstopped) {
1719     rgw_bucket_entry_ver ver;
1720     ver.epoch = (op.olh_epoch ? op.olh_epoch : olh.get_epoch());
1721
1722     real_time mtime = obj.mtime(); /* mtime has no real meaning in instance removal context */
1723     ret = log_index_operation(hctx, op.key, CLS_RGW_OP_UNLINK_INSTANCE, op.op_tag,
1724                               mtime, ver,
1725                               CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker,
1726                               op.bilog_flags | RGW_BILOG_FLAG_VERSIONED_OP, NULL, NULL, &op.zones_trace);
1727     if (ret < 0)
1728       return ret;
1729
1730     return write_bucket_header(hctx, &header); /* updates header version */
1731   }
1732
1733   return 0;
1734 }
1735
1736 static int rgw_bucket_read_olh_log(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1737 {
1738   // decode request
1739   rgw_cls_read_olh_log_op op;
1740   auto iter = in->cbegin();
1741   try {
1742     decode(op, iter);
1743   } catch (buffer::error& err) {
1744     CLS_LOG(0, "ERROR: rgw_bucket_read_olh_log(): failed to decode request\n");
1745     return -EINVAL;
1746   }
1747
1748   if (!op.olh.instance.empty()) {
1749     CLS_LOG(1, "bad key passed in (non empty instance)");
1750     return -EINVAL;
1751   }
1752
1753   rgw_bucket_olh_entry olh_data_entry;
1754   string olh_data_key;
1755   encode_olh_data_key(op.olh, &olh_data_key);
1756   int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry);
1757   if (ret < 0 && ret != -ENOENT) {
1758     CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1759     return ret;
1760   }
1761
1762   if (olh_data_entry.tag != op.olh_tag) {
1763     CLS_LOG(1, "NOTICE: %s(): olh_tag_mismatch olh_data_entry.tag=%s op.olh_tag=%s", __func__, olh_data_entry.tag.c_str(), op.olh_tag.c_str());
1764     return -ECANCELED;
1765   }
1766
1767   rgw_cls_read_olh_log_ret op_ret;
1768
1769 #define MAX_OLH_LOG_ENTRIES 1000
1770   map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log = olh_data_entry.pending_log;
1771
1772   if (log.begin()->first > op.ver_marker && log.size() <= MAX_OLH_LOG_ENTRIES) {
1773     op_ret.log = log;
1774     op_ret.is_truncated = false;
1775   } else {
1776     map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator iter = log.upper_bound(op.ver_marker);
1777
1778     for (int i = 0; i < MAX_OLH_LOG_ENTRIES && iter != log.end(); ++i, ++iter) {
1779       op_ret.log[iter->first] = iter->second;
1780     }
1781     op_ret.is_truncated = (iter != log.end());
1782   }
1783
1784   encode(op_ret, *out);
1785
1786   return 0;
1787 }
1788
1789 static int rgw_bucket_trim_olh_log(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1790 {
1791   // decode request
1792   rgw_cls_trim_olh_log_op op;
1793   auto iter = in->cbegin();
1794   try {
1795     decode(op, iter);
1796   } catch (buffer::error& err) {
1797     CLS_LOG(0, "ERROR: rgw_bucket_trim_olh_log(): failed to decode request\n");
1798     return -EINVAL;
1799   }
1800
1801   if (!op.olh.instance.empty()) {
1802     CLS_LOG(1, "bad key passed in (non empty instance)");
1803     return -EINVAL;
1804   }
1805
1806   /* read olh entry */
1807   rgw_bucket_olh_entry olh_data_entry;
1808   string olh_data_key;
1809   encode_olh_data_key(op.olh, &olh_data_key);
1810   int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry);
1811   if (ret < 0 && ret != -ENOENT) {
1812     CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1813     return ret;
1814   }
1815
1816   if (olh_data_entry.tag != op.olh_tag) {
1817     CLS_LOG(1, "NOTICE: %s(): olh_tag_mismatch olh_data_entry.tag=%s op.olh_tag=%s", __func__, olh_data_entry.tag.c_str(), op.olh_tag.c_str());
1818     return -ECANCELED;
1819   }
1820
1821   /* remove all versions up to and including ver from the pending map */
1822   map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log = olh_data_entry.pending_log;
1823   map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator liter = log.begin();
1824   while (liter != log.end() && liter->first <= op.ver) {
1825     map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator rm_iter = liter;
1826     ++liter;
1827     log.erase(rm_iter);
1828   }
1829
1830   /* write the olh data entry */
1831   ret = write_entry(hctx, olh_data_entry, olh_data_key);
1832   if (ret < 0) {
1833     CLS_LOG(0, "ERROR: write_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1834     return ret;
1835   }
1836
1837   return 0;
1838 }
1839
1840 static int rgw_bucket_clear_olh(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1841 {
1842   // decode request
1843   rgw_cls_bucket_clear_olh_op op;
1844   auto iter = in->cbegin();
1845   try {
1846     decode(op, iter);
1847   } catch (buffer::error& err) {
1848     CLS_LOG(0, "ERROR: rgw_bucket_clear_olh(): failed to decode request\n");
1849     return -EINVAL;
1850   }
1851
1852   if (!op.key.instance.empty()) {
1853     CLS_LOG(1, "bad key passed in (non empty instance)");
1854     return -EINVAL;
1855   }
1856
1857   /* read olh entry */
1858   rgw_bucket_olh_entry olh_data_entry;
1859   string olh_data_key;
1860   encode_olh_data_key(op.key, &olh_data_key);
1861   int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry);
1862   if (ret < 0 && ret != -ENOENT) {
1863     CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1864     return ret;
1865   }
1866
1867   if (olh_data_entry.tag != op.olh_tag) {
1868     CLS_LOG(1, "NOTICE: %s(): olh_tag_mismatch olh_data_entry.tag=%s op.olh_tag=%s", __func__, olh_data_entry.tag.c_str(), op.olh_tag.c_str());
1869     return -ECANCELED;
1870   }
1871
1872   ret = cls_cxx_map_remove_key(hctx, olh_data_key);
1873   if (ret < 0) {
1874     CLS_LOG(1, "NOTICE: %s(): can't remove key %s ret=%d", __func__, olh_data_key.c_str(), ret);
1875     return ret;
1876   }
1877
1878   rgw_bucket_dir_entry plain_entry;
1879
1880   /* read plain entry, make sure it's a versioned place holder */
1881   ret = read_index_entry(hctx, op.key.name, &plain_entry);
1882   if (ret == -ENOENT) {
1883     /* we're done, no entry existing */
1884     return 0;
1885   }
1886   if (ret < 0) {
1887     CLS_LOG(0, "ERROR: read_index_entry key=%s ret=%d", op.key.name.c_str(), ret);
1888     return ret;
1889   }
1890
1891   if ((plain_entry.flags & RGW_BUCKET_DIRENT_FLAG_VER_MARKER) == 0) {
1892     /* it's not a version marker, don't remove it */
1893     return 0;
1894   }
1895
1896   ret = cls_cxx_map_remove_key(hctx, op.key.name);
1897   if (ret < 0) {
1898     CLS_LOG(1, "NOTICE: %s(): can't remove key %s ret=%d", __func__, op.key.name.c_str(), ret);
1899     return ret;
1900   }
1901
1902   return 0;
1903 }
1904
1905 int rgw_dir_suggest_changes(cls_method_context_t hctx,
1906                             bufferlist *in, bufferlist *out)
1907 {
1908   CLS_LOG(1, "rgw_dir_suggest_changes()");
1909
1910   bufferlist header_bl;
1911   rgw_bucket_dir_header header;
1912   bool header_changed = false;
1913
1914   int rc = read_bucket_header(hctx, &header);
1915   if (rc < 0) {
1916     CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to read header\n");
1917     return rc;
1918   }
1919
1920   timespan tag_timeout(
1921     std::chrono::seconds(
1922       header.tag_timeout ? header.tag_timeout : CEPH_RGW_TAG_TIMEOUT));
1923
1924   auto in_iter = in->cbegin();
1925
1926   while (!in_iter.end()) {
1927     __u8 op;
1928     rgw_bucket_dir_entry cur_change;
1929     rgw_bucket_dir_entry cur_disk;
1930     try {
1931       decode(op, in_iter);
1932       decode(cur_change, in_iter);
1933     } catch (buffer::error& err) {
1934       CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to decode request\n");
1935       return -EINVAL;
1936     }
1937
1938     bufferlist cur_disk_bl;
1939     string cur_change_key;
1940     encode_obj_index_key(cur_change.key, &cur_change_key);
1941     int ret = cls_cxx_map_get_val(hctx, cur_change_key, &cur_disk_bl);
1942     if (ret < 0 && ret != -ENOENT)
1943       return -EINVAL;
1944
1945     if (ret == -ENOENT) {
1946       continue;
1947     }
1948
1949     if (cur_disk_bl.length()) {
1950       auto cur_disk_iter = cur_disk_bl.cbegin();
1951       try {
1952         decode(cur_disk, cur_disk_iter);
1953       } catch (buffer::error& error) {
1954         CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to decode cur_disk\n");
1955         return -EINVAL;
1956       }
1957
1958       real_time cur_time = real_clock::now();
1959       map<string, rgw_bucket_pending_info>::iterator iter =
1960                 cur_disk.pending_map.begin();
1961       while(iter != cur_disk.pending_map.end()) {
1962         map<string, rgw_bucket_pending_info>::iterator cur_iter=iter++;
1963         if (cur_time > (cur_iter->second.timestamp + timespan(tag_timeout))) {
1964           cur_disk.pending_map.erase(cur_iter);
1965         }
1966       }
1967     }
1968
1969     CLS_LOG(20, "cur_disk.pending_map.empty()=%d op=%d cur_disk.exists=%d cur_change.pending_map.size()=%d cur_change.exists=%d\n",
1970             cur_disk.pending_map.empty(), (int)op, cur_disk.exists,
1971             (int)cur_change.pending_map.size(), cur_change.exists);
1972
1973     if (cur_disk.pending_map.empty()) {
1974       if (cur_disk.exists) {
1975         rgw_bucket_category_stats& old_stats = header.stats[cur_disk.meta.category];
1976         CLS_LOG(10, "total_entries: %" PRId64 " -> %" PRId64 "\n", old_stats.num_entries, old_stats.num_entries - 1);
1977         old_stats.num_entries--;
1978         old_stats.total_size -= cur_disk.meta.accounted_size;
1979         old_stats.total_size_rounded -= cls_rgw_get_rounded_size(cur_disk.meta.accounted_size);
1980         old_stats.actual_size -= cur_disk.meta.size;
1981         header_changed = true;
1982       }
1983       rgw_bucket_category_stats& stats = header.stats[cur_change.meta.category];
1984       bool log_op = (op & CEPH_RGW_DIR_SUGGEST_LOG_OP) != 0;
1985       op &= CEPH_RGW_DIR_SUGGEST_OP_MASK;
1986       switch(op) {
1987       case CEPH_RGW_REMOVE:
1988         CLS_LOG(10, "CEPH_RGW_REMOVE name=%s instance=%s\n", cur_change.key.name.c_str(), cur_change.key.instance.c_str());
1989         ret = cls_cxx_map_remove_key(hctx, cur_change_key);
1990         if (ret < 0)
1991           return ret;
1992         if (log_op && cur_disk.exists && !header.syncstopped) {
1993           ret = log_index_operation(hctx, cur_disk.key, CLS_RGW_OP_DEL, cur_disk.tag, cur_disk.meta.mtime,
1994                                     cur_disk.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, 0, NULL, NULL, NULL);
1995           if (ret < 0) {
1996             CLS_LOG(0, "ERROR: %s(): failed to log operation ret=%d", __func__, ret);
1997             return ret;
1998           }
1999         }
2000         break;
2001       case CEPH_RGW_UPDATE:
2002         CLS_LOG(10, "CEPH_RGW_UPDATE name=%s instance=%s total_entries: %" PRId64 " -> %" PRId64 "\n",
2003                 cur_change.key.name.c_str(), cur_change.key.instance.c_str(), stats.num_entries, stats.num_entries + 1);
2004
2005         stats.num_entries++;
2006         stats.total_size += cur_change.meta.accounted_size;
2007         stats.total_size_rounded += cls_rgw_get_rounded_size(cur_change.meta.accounted_size);
2008         stats.actual_size += cur_change.meta.size;
2009         header_changed = true;
2010         cur_change.index_ver = header.ver;
2011         bufferlist cur_state_bl;
2012         encode(cur_change, cur_state_bl);
2013         ret = cls_cxx_map_set_val(hctx, cur_change_key, &cur_state_bl);
2014         if (ret < 0)
2015           return ret;
2016         if (log_op && !header.syncstopped) {
2017           ret = log_index_operation(hctx, cur_change.key, CLS_RGW_OP_ADD, cur_change.tag, cur_change.meta.mtime,
2018                                     cur_change.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, 0, NULL, NULL, NULL);
2019           if (ret < 0) {
2020             CLS_LOG(0, "ERROR: %s(): failed to log operation ret=%d", __func__, ret);
2021             return ret;
2022           }
2023         }
2024         break;
2025       } // switch(op)
2026     } // if (cur_disk.pending_map.empty())
2027   } // while (!in_iter.end())
2028
2029   if (header_changed) {
2030     return write_bucket_header(hctx, &header);
2031   }
2032   return 0;
2033 }
2034
2035 static int rgw_obj_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2036 {
2037   // decode request
2038   rgw_cls_obj_remove_op op;
2039   auto iter = in->cbegin();
2040   try {
2041     decode(op, iter);
2042   } catch (buffer::error& err) {
2043     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2044     return -EINVAL;
2045   }
2046
2047   if (op.keep_attr_prefixes.empty()) {
2048     return cls_cxx_remove(hctx);
2049   }
2050
2051   map<string, bufferlist> attrset;
2052   int ret = cls_cxx_getxattrs(hctx, &attrset);
2053   if (ret < 0 && ret != -ENOENT) {
2054     CLS_LOG(0, "ERROR: %s(): cls_cxx_getxattrs() returned %d", __func__, ret);
2055     return ret;
2056   }
2057
2058   map<string, bufferlist> new_attrs;
2059   for (list<string>::iterator iter = op.keep_attr_prefixes.begin();
2060        iter != op.keep_attr_prefixes.end(); ++iter) {
2061     string& check_prefix = *iter;
2062
2063     for (map<string, bufferlist>::iterator aiter = attrset.lower_bound(check_prefix);
2064          aiter != attrset.end(); ++aiter) {
2065       const string& attr = aiter->first;
2066
2067       if (attr.substr(0, check_prefix.size()) > check_prefix) {
2068         break;
2069       }
2070
2071       new_attrs[attr] = aiter->second;
2072     }
2073   }
2074
2075   CLS_LOG(20, "%s(): removing object", __func__);
2076   ret = cls_cxx_remove(hctx);
2077   if (ret < 0) {
2078     CLS_LOG(0, "ERROR: %s(): cls_cxx_remove returned %d", __func__, ret);
2079     return ret;
2080   }
2081
2082   if (new_attrs.empty()) {
2083     /* no data to keep */
2084     return 0;
2085   }
2086
2087   ret = cls_cxx_create(hctx, false);
2088   if (ret < 0) {
2089     CLS_LOG(0, "ERROR: %s(): cls_cxx_create returned %d", __func__, ret);
2090     return ret;
2091   }
2092
2093   for (map<string, bufferlist>::iterator aiter = new_attrs.begin();
2094        aiter != new_attrs.end(); ++aiter) {
2095     const string& attr = aiter->first;
2096
2097     ret = cls_cxx_setxattr(hctx, attr.c_str(), &aiter->second);
2098     CLS_LOG(20, "%s(): setting attr: %s", __func__, attr.c_str());
2099     if (ret < 0) {
2100       CLS_LOG(0, "ERROR: %s(): cls_cxx_setxattr (attr=%s) returned %d", __func__, attr.c_str(), ret);
2101       return ret;
2102     }
2103   }
2104
2105   return 0;
2106 }
2107
2108 static int rgw_obj_store_pg_ver(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2109 {
2110   // decode request
2111   rgw_cls_obj_store_pg_ver_op op;
2112   auto iter = in->cbegin();
2113   try {
2114     decode(op, iter);
2115   } catch (buffer::error& err) {
2116     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2117     return -EINVAL;
2118   }
2119
2120   bufferlist bl;
2121   uint64_t ver = cls_current_version(hctx);
2122   encode(ver, bl);
2123   int ret = cls_cxx_setxattr(hctx, op.attr.c_str(), &bl);
2124   if (ret < 0) {
2125     CLS_LOG(0, "ERROR: %s(): cls_cxx_setxattr (attr=%s) returned %d", __func__, op.attr.c_str(), ret);
2126     return ret;
2127   }
2128
2129   return 0;
2130 }
2131
2132 static int rgw_obj_check_attrs_prefix(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2133 {
2134   // decode request
2135   rgw_cls_obj_check_attrs_prefix op;
2136   auto iter = in->cbegin();
2137   try {
2138     decode(op, iter);
2139   } catch (buffer::error& err) {
2140     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2141     return -EINVAL;
2142   }
2143
2144   if (op.check_prefix.empty()) {
2145     return -EINVAL;
2146   }
2147
2148   map<string, bufferlist> attrset;
2149   int ret = cls_cxx_getxattrs(hctx, &attrset);
2150   if (ret < 0 && ret != -ENOENT) {
2151     CLS_LOG(0, "ERROR: %s(): cls_cxx_getxattrs() returned %d", __func__, ret);
2152     return ret;
2153   }
2154
2155   bool exist = false;
2156
2157   for (map<string, bufferlist>::iterator aiter = attrset.lower_bound(op.check_prefix);
2158        aiter != attrset.end(); ++aiter) {
2159     const string& attr = aiter->first;
2160
2161     if (attr.substr(0, op.check_prefix.size()) > op.check_prefix) {
2162       break;
2163     }
2164
2165     exist = true;
2166   }
2167
2168   if (exist == op.fail_if_exist) {
2169     return -ECANCELED;
2170   }
2171
2172   return 0;
2173 }
2174
2175 static int rgw_obj_check_mtime(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2176 {
2177   // decode request
2178   rgw_cls_obj_check_mtime op;
2179   auto iter = in->cbegin();
2180   try {
2181     decode(op, iter);
2182   } catch (buffer::error& err) {
2183     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2184     return -EINVAL;
2185   }
2186
2187   real_time obj_ut;
2188   int ret = cls_cxx_stat2(hctx, NULL, &obj_ut);
2189   if (ret < 0 && ret != -ENOENT) {
2190     CLS_LOG(0, "ERROR: %s(): cls_cxx_stat() returned %d", __func__, ret);
2191     return ret;
2192   }
2193   if (ret == -ENOENT) {
2194     CLS_LOG(10, "object does not exist, skipping check");
2195   }
2196
2197   ceph_timespec obj_ts = ceph::real_clock::to_ceph_timespec(obj_ut);
2198   ceph_timespec op_ts = ceph::real_clock::to_ceph_timespec(op.mtime);
2199
2200   if (!op.high_precision_time) {
2201     obj_ts.tv_nsec = 0;
2202     op_ts.tv_nsec = 0;
2203   }
2204
2205   CLS_LOG(10, "%s: obj_ut=%lld.%06lld op.mtime=%lld.%06lld", __func__,
2206           (long long)obj_ts.tv_sec, (long long)obj_ts.tv_nsec,
2207           (long long)op_ts.tv_sec, (long long)op_ts.tv_nsec);
2208
2209   bool check;
2210
2211   switch (op.type) {
2212   case CLS_RGW_CHECK_TIME_MTIME_EQ:
2213     check = (obj_ts == op_ts);
2214     break;
2215   case CLS_RGW_CHECK_TIME_MTIME_LT:
2216     check = (obj_ts < op_ts);
2217     break;
2218   case CLS_RGW_CHECK_TIME_MTIME_LE:
2219     check = (obj_ts <= op_ts);
2220     break;
2221   case CLS_RGW_CHECK_TIME_MTIME_GT:
2222     check = (obj_ts > op_ts);
2223     break;
2224   case CLS_RGW_CHECK_TIME_MTIME_GE:
2225     check = (obj_ts >= op_ts);
2226     break;
2227   default:
2228     return -EINVAL;
2229   };
2230
2231   if (!check) {
2232     return -ECANCELED;
2233   }
2234
2235   return 0;
2236 }
2237
2238 static int rgw_bi_get_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2239 {
2240   // decode request
2241   rgw_cls_bi_get_op op;
2242   auto iter = in->cbegin();
2243   try {
2244     decode(op, iter);
2245   } catch (buffer::error& err) {
2246     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2247     return -EINVAL;
2248   }
2249
2250   string idx;
2251
2252   switch (op.type) {
2253     case BIIndexType::Plain:
2254       idx = op.key.name;
2255       break;
2256     case BIIndexType::Instance:
2257       encode_obj_index_key(op.key, &idx);
2258       break;
2259     case BIIndexType::OLH:
2260       encode_olh_data_key(op.key, &idx);
2261       break;
2262     default:
2263       CLS_LOG(10, "%s(): invalid key type encoding: %d",
2264               __func__, int(op.type));
2265       return -EINVAL;
2266   }
2267
2268   rgw_cls_bi_get_ret op_ret;
2269
2270   rgw_cls_bi_entry& entry = op_ret.entry;
2271
2272   entry.type = op.type;
2273   entry.idx = idx;
2274
2275   int r = cls_cxx_map_get_val(hctx, idx, &entry.data);
2276   if (r < 0) {
2277       CLS_LOG(10, "%s(): cls_cxx_map_get_val() returned %d", __func__, r);
2278       return r;
2279   }
2280
2281   encode(op_ret, *out);
2282
2283   return 0;
2284 }
2285
2286 static int rgw_bi_put_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2287 {
2288   // decode request
2289   rgw_cls_bi_put_op op;
2290   auto iter = in->cbegin();
2291   try {
2292     decode(op, iter);
2293   } catch (buffer::error& err) {
2294     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2295     return -EINVAL;
2296   }
2297
2298   rgw_cls_bi_entry& entry = op.entry;
2299
2300   int r = cls_cxx_map_set_val(hctx, entry.idx, &entry.data);
2301   if (r < 0) {
2302     CLS_LOG(0, "ERROR: %s(): cls_cxx_map_set_val() returned r=%d", __func__, r);
2303   }
2304
2305   return 0;
2306 }
2307
2308 static int list_plain_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max,
2309                               list<rgw_cls_bi_entry> *entries, bool *pmore)
2310 {
2311   string filter = name;
2312   string start_key = marker;
2313
2314   string end_key; // stop listing at bi_log_prefix
2315   bi_log_prefix(end_key);
2316
2317   int count = 0;
2318   map<string, bufferlist> keys;
2319   int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max, &keys, pmore);
2320   if (ret < 0) {
2321     return ret;
2322   }
2323
2324   map<string, bufferlist>::iterator iter;
2325   for (iter = keys.begin(); iter != keys.end(); ++iter) {
2326     if (iter->first >= end_key) {
2327       /* past the end of plain namespace */
2328       if (pmore) {
2329         *pmore = false;
2330       }
2331       return count;
2332     }
2333
2334     rgw_cls_bi_entry entry;
2335     entry.type = BIIndexType::Plain;
2336     entry.idx = iter->first;
2337     entry.data = iter->second;
2338
2339     auto biter = entry.data.cbegin();
2340
2341     rgw_bucket_dir_entry e;
2342     try {
2343       decode(e, biter);
2344     } catch (buffer::error& err) {
2345       CLS_LOG(0, "ERROR: %s(): failed to decode buffer", __func__);
2346       return -EIO;
2347     }
2348
2349     CLS_LOG(20, "%s(): entry.idx=%s e.key.name=%s", __func__, escape_str(entry.idx).c_str(), escape_str(e.key.name).c_str());
2350
2351     if (!name.empty() && e.key.name != name) {
2352       /* we are skipping the rest of the entries */
2353       if (pmore) {
2354         *pmore = false;
2355       }
2356       return count;
2357     }
2358
2359     entries->push_back(entry);
2360     count++;
2361     if (count >= (int)max) {
2362       return count;
2363     }
2364     start_key = entry.idx;
2365   }
2366
2367   return count;
2368 }
2369
2370 static int list_instance_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max,
2371                                  list<rgw_cls_bi_entry> *entries, bool *pmore)
2372 {
2373   cls_rgw_obj_key key(name);
2374   string first_instance_idx;
2375   encode_obj_versioned_data_key(key, &first_instance_idx);
2376   string start_key;
2377
2378   if (!name.empty()) {
2379     start_key = first_instance_idx;
2380   } else {
2381     start_key = BI_PREFIX_CHAR;
2382     start_key.append(bucket_index_prefixes[BI_BUCKET_OBJ_INSTANCE_INDEX]);
2383   }
2384   string filter = start_key;
2385   if (bi_entry_gt(marker, start_key)) {
2386     start_key = marker;
2387   }
2388   int count = 0;
2389   map<string, bufferlist> keys;
2390   bufferlist k;
2391   int ret = cls_cxx_map_get_val(hctx, start_key, &k);
2392   if (ret < 0 && ret != -ENOENT) {
2393     return ret;
2394   }
2395   bool found_first = (ret == 0);
2396   if (found_first) {
2397     --max;
2398   }
2399   if (max > 0) {
2400     ret = cls_cxx_map_get_vals(hctx, start_key, string(), max, &keys, pmore);
2401     CLS_LOG(20, "%s(): start_key=%s first_instance_idx=%s keys.size()=%d", __func__, escape_str(start_key).c_str(), escape_str(first_instance_idx).c_str(), (int)keys.size());
2402     if (ret < 0) {
2403       return ret;
2404     }
2405   }
2406   if (found_first) {
2407     keys[start_key].claim(k);
2408   }
2409
2410   map<string, bufferlist>::iterator iter;
2411   for (iter = keys.begin(); iter != keys.end(); ++iter) {
2412     rgw_cls_bi_entry entry;
2413     entry.type = BIIndexType::Instance;
2414     entry.idx = iter->first;
2415     entry.data = iter->second;
2416
2417     if (!filter.empty() && entry.idx.compare(0, filter.size(), filter) != 0) {
2418       /* we are skipping the rest of the entries */
2419       if (pmore) {
2420         *pmore = false;
2421       }
2422       return count;
2423     }
2424
2425     CLS_LOG(20, "%s(): entry.idx=%s", __func__, escape_str(entry.idx).c_str());
2426
2427     auto biter = entry.data.cbegin();
2428
2429     rgw_bucket_dir_entry e;
2430     try {
2431       decode(e, biter);
2432     } catch (buffer::error& err) {
2433       CLS_LOG(0, "ERROR: %s(): failed to decode buffer (size=%d)", __func__, entry.data.length());
2434       return -EIO;
2435     }
2436
2437     if (!name.empty() && e.key.name != name) {
2438       /* we are skipping the rest of the entries */
2439       if (pmore) {
2440         *pmore = false;
2441       }
2442       return count;
2443     }
2444
2445     entries->push_back(entry);
2446     count++;
2447     start_key = entry.idx;
2448   }
2449
2450   return count;
2451 }
2452
2453 static int list_olh_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max,
2454                             list<rgw_cls_bi_entry> *entries, bool *pmore)
2455 {
2456   cls_rgw_obj_key key(name);
2457   string first_instance_idx;
2458   encode_olh_data_key(key, &first_instance_idx);
2459   string start_key;
2460
2461   if (!name.empty()) {
2462     start_key = first_instance_idx;
2463   } else {
2464     start_key = BI_PREFIX_CHAR;
2465     start_key.append(bucket_index_prefixes[BI_BUCKET_OLH_DATA_INDEX]);
2466   }
2467   string filter = start_key;
2468   if (bi_entry_gt(marker, start_key)) {
2469     start_key = marker;
2470   }
2471   int count = 0;
2472   map<string, bufferlist> keys;
2473   int ret;
2474   bufferlist k;
2475   ret = cls_cxx_map_get_val(hctx, start_key, &k);
2476   if (ret < 0 && ret != -ENOENT) {
2477     return ret;
2478   }
2479   bool found_first = (ret == 0);
2480   if (found_first) {
2481     --max;
2482   }
2483   if (max > 0) {
2484     ret = cls_cxx_map_get_vals(hctx, start_key, string(), max, &keys, pmore);
2485     CLS_LOG(20, "%s(): start_key=%s first_instance_idx=%s keys.size()=%d", __func__, escape_str(start_key).c_str(), escape_str(first_instance_idx).c_str(), (int)keys.size());
2486     if (ret < 0) {
2487       return ret;
2488     }
2489   }
2490
2491   if (found_first) {
2492     keys[start_key].claim(k);
2493   }
2494
2495   map<string, bufferlist>::iterator iter;
2496   for (iter = keys.begin(); iter != keys.end(); ++iter) {
2497     rgw_cls_bi_entry entry;
2498     entry.type = BIIndexType::OLH;
2499     entry.idx = iter->first;
2500     entry.data = iter->second;
2501
2502     if (!filter.empty() && entry.idx.compare(0, filter.size(), filter) != 0) {
2503       /* we are skipping the rest of the entries */
2504       if (pmore) {
2505         *pmore = false;
2506       }
2507       return count;
2508     }
2509
2510     CLS_LOG(20, "%s(): entry.idx=%s", __func__, escape_str(entry.idx).c_str());
2511
2512     auto biter = entry.data.cbegin();
2513
2514     rgw_bucket_olh_entry e;
2515     try {
2516       decode(e, biter);
2517     } catch (buffer::error& err) {
2518       CLS_LOG(0, "ERROR: %s(): failed to decode buffer (size=%d)", __func__, entry.data.length());
2519       return -EIO;
2520     }
2521
2522     if (!name.empty() && e.key.name != name) {
2523       /* we are skipping the rest of the entries */
2524       if (pmore) {
2525         *pmore = false;
2526       }
2527       return count;
2528     }
2529
2530     entries->push_back(entry);
2531     count++;
2532     start_key = entry.idx;
2533   }
2534
2535   return count;
2536 }
2537
2538 static int rgw_bi_list_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2539 {
2540   // decode request
2541   rgw_cls_bi_list_op op;
2542   auto iter = in->cbegin();
2543   try {
2544     decode(op, iter);
2545   } catch (buffer::error& err) {
2546     CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2547     return -EINVAL;
2548   }
2549
2550   rgw_cls_bi_list_ret op_ret;
2551
2552   string filter = op.name;
2553 #define MAX_BI_LIST_ENTRIES 1000
2554   int32_t max = (op.max < MAX_BI_LIST_ENTRIES ? op.max : MAX_BI_LIST_ENTRIES);
2555   string start_key = op.marker;
2556   bool more;
2557   int ret = list_plain_entries(hctx, op.name, op.marker, max, &op_ret.entries, &more);
2558   if (ret < 0) {
2559     CLS_LOG(0, "ERROR: %s(): list_plain_entries returned ret=%d", __func__, ret);
2560     return ret;
2561   }
2562   int count = ret;
2563
2564   CLS_LOG(20, "found %d plain entries", count);
2565
2566   if (!more) {
2567     ret = list_instance_entries(hctx, op.name, op.marker, max - count, &op_ret.entries, &more);
2568     if (ret < 0) {
2569       CLS_LOG(0, "ERROR: %s(): list_instance_entries returned ret=%d", __func__, ret);
2570       return ret;
2571     }
2572
2573     count += ret;
2574   }
2575
2576   if (!more) {
2577     ret = list_olh_entries(hctx, op.name, op.marker, max - count, &op_ret.entries, &more);
2578     if (ret < 0) {
2579       CLS_LOG(0, "ERROR: %s(): list_olh_entries returned ret=%d", __func__, ret);
2580       return ret;
2581     }
2582
2583     count += ret;
2584   }
2585
2586   op_ret.is_truncated = (count >= max) || more;
2587   while (count > max) {
2588     op_ret.entries.pop_back();
2589     count--;
2590   }
2591
2592   encode(op_ret, *out);
2593
2594   return 0;
2595 }
2596
2597 int bi_log_record_decode(bufferlist& bl, rgw_bi_log_entry& e)
2598 {
2599   auto iter = bl.cbegin();
2600   try {
2601     decode(e, iter);
2602   } catch (buffer::error& err) {
2603     CLS_LOG(0, "ERROR: failed to decode rgw_bi_log_entry");
2604     return -EIO;
2605   }
2606   return 0;
2607 }
2608
2609 static int bi_log_iterate_entries(cls_method_context_t hctx, const string& marker, const string& end_marker,
2610                               string& key_iter, uint32_t max_entries, bool *truncated,
2611                               int (*cb)(cls_method_context_t, const string&, rgw_bi_log_entry&, void *),
2612                               void *param)
2613 {
2614   CLS_LOG(10, "bi_log_iterate_range");
2615
2616   map<string, bufferlist> keys;
2617   string filter_prefix, end_key;
2618   uint32_t i = 0;
2619   string key;
2620
2621   if (truncated)
2622     *truncated = false;
2623
2624   string start_key;
2625   if (key_iter.empty()) {
2626     key = BI_PREFIX_CHAR;
2627     key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
2628     key.append(marker);
2629
2630     start_key = key;
2631   } else {
2632     start_key = key_iter;
2633   }
2634
2635   if (end_marker.empty()) {
2636     end_key = BI_PREFIX_CHAR;
2637     end_key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX + 1]);
2638   } else {
2639     end_key = BI_PREFIX_CHAR;
2640     end_key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
2641     end_key.append(end_marker);
2642   }
2643
2644   CLS_LOG(10, "bi_log_iterate_entries start_key=%s end_key=%s\n", start_key.c_str(), end_key.c_str());
2645
2646   string filter;
2647
2648   int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max_entries, &keys, truncated);
2649   if (ret < 0)
2650     return ret;
2651
2652   map<string, bufferlist>::iterator iter = keys.begin();
2653   if (iter == keys.end())
2654     return 0;
2655
2656   uint32_t num_keys = keys.size();
2657
2658   for (; iter != keys.end(); ++iter,++i) {
2659     const string& key = iter->first;
2660     rgw_bi_log_entry e;
2661
2662     CLS_LOG(10, "bi_log_iterate_entries key=%s bl.length=%d\n", key.c_str(), (int)iter->second.length());
2663
2664     if (key.compare(end_key) > 0) {
2665       key_iter = key;
2666       if (truncated) {
2667         *truncated = false;
2668       }
2669       return 0;
2670     }
2671
2672     ret = bi_log_record_decode(iter->second, e);
2673     if (ret < 0)
2674       return ret;
2675
2676     ret = cb(hctx, key, e, param);
2677     if (ret < 0)
2678       return ret;
2679
2680     if (i == num_keys - 1) {
2681       key_iter = key;
2682     }
2683   }
2684
2685   return 0;
2686 }
2687
2688 static int bi_log_list_cb(cls_method_context_t hctx, const string& key, rgw_bi_log_entry& info, void *param)
2689 {
2690   list<rgw_bi_log_entry> *l = (list<rgw_bi_log_entry> *)param;
2691   l->push_back(info);
2692   return 0;
2693 }
2694
2695 static int bi_log_list_entries(cls_method_context_t hctx, const string& marker,
2696                            uint32_t max, list<rgw_bi_log_entry>& entries, bool *truncated)
2697 {
2698   string key_iter;
2699   string end_marker;
2700   int ret = bi_log_iterate_entries(hctx, marker, end_marker,
2701                               key_iter, max, truncated,
2702                               bi_log_list_cb, &entries);
2703   return ret;
2704 }
2705
2706 static int rgw_bi_log_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2707 {
2708   auto in_iter = in->cbegin();
2709
2710   cls_rgw_bi_log_list_op op;
2711   try {
2712     decode(op, in_iter);
2713   } catch (buffer::error& err) {
2714     CLS_LOG(1, "ERROR: rgw_bi_log_list(): failed to decode entry\n");
2715     return -EINVAL;
2716   }
2717
2718   cls_rgw_bi_log_list_ret op_ret;
2719   int ret = bi_log_list_entries(hctx, op.marker, op.max, op_ret.entries, &op_ret.truncated);
2720   if (ret < 0)
2721     return ret;
2722
2723   encode(op_ret, *out);
2724
2725   return 0;
2726 }
2727
2728 static int bi_log_list_trim_cb(cls_method_context_t hctx, const string& key, rgw_bi_log_entry& info, void *param)
2729 {
2730   list<rgw_bi_log_entry> *entries = (list<rgw_bi_log_entry> *)param;
2731
2732   entries->push_back(info);
2733   return 0;
2734 }
2735
2736 static int bi_log_remove_entry(cls_method_context_t hctx, rgw_bi_log_entry& entry)
2737 {
2738   string key;
2739   key = BI_PREFIX_CHAR;
2740   key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
2741   key.append(entry.id);
2742   return cls_cxx_map_remove_key(hctx, key);
2743 }
2744
2745 static int bi_log_list_trim_entries(cls_method_context_t hctx,
2746                                     const string& start_marker, const string& end_marker,
2747                                     list<rgw_bi_log_entry>& entries, bool *truncated)
2748 {
2749   string key_iter;
2750 #define MAX_TRIM_ENTRIES 1000 /* max entries to trim in a single operation */
2751   int ret = bi_log_iterate_entries(hctx, start_marker, end_marker,
2752                               key_iter, MAX_TRIM_ENTRIES, truncated,
2753                               bi_log_list_trim_cb, &entries);
2754   return ret;
2755 }
2756
2757 static int rgw_bi_log_trim(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2758 {
2759   auto in_iter = in->cbegin();
2760
2761   cls_rgw_bi_log_trim_op op;
2762   try {
2763     decode(op, in_iter);
2764   } catch (buffer::error& err) {
2765     CLS_LOG(1, "ERROR: rgw_bi_log_list(): failed to decode entry\n");
2766     return -EINVAL;
2767   }
2768
2769   cls_rgw_bi_log_list_ret op_ret;
2770   list<rgw_bi_log_entry> entries;
2771 #define MAX_TRIM_ENTRIES 1000 /* don't do more than that in a single operation */
2772   bool truncated;
2773   int ret = bi_log_list_trim_entries(hctx, op.start_marker, op.end_marker, entries, &truncated);
2774   if (ret < 0)
2775     return ret;
2776
2777   if (entries.empty())
2778     return -ENODATA;
2779
2780   list<rgw_bi_log_entry>::iterator iter;
2781   for (iter = entries.begin(); iter != entries.end(); ++iter) {
2782     rgw_bi_log_entry& entry = *iter;
2783
2784     ret = bi_log_remove_entry(hctx, entry);
2785     if (ret < 0)
2786       return ret;
2787   }
2788
2789   return 0;
2790 }
2791
2792 static int rgw_bi_log_resync(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2793 {
2794   rgw_bucket_dir_header header;
2795   int rc = read_bucket_header(hctx, &header);
2796   if (rc < 0) {
2797     CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n");
2798     return rc;
2799   }
2800
2801   bufferlist bl;
2802
2803   rgw_bi_log_entry entry;
2804
2805   entry.timestamp = real_clock::now();
2806   entry.op = RGWModifyOp::CLS_RGW_OP_RESYNC;
2807   entry.state = RGWPendingState::CLS_RGW_STATE_COMPLETE;
2808
2809   string key;
2810   bi_log_index_key(hctx, key, entry.id, header.ver);
2811
2812   encode(entry, bl);
2813
2814   if (entry.id > header.max_marker)
2815     header.max_marker = entry.id;
2816
2817   header.syncstopped = false;
2818
2819   rc = cls_cxx_map_set_val(hctx, key, &bl);
2820   if (rc < 0)
2821     return rc;
2822
2823   return write_bucket_header(hctx, &header);
2824 }
2825
2826 static int rgw_bi_log_stop(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2827 {
2828   rgw_bucket_dir_header header;
2829   int rc = read_bucket_header(hctx, &header);
2830   if (rc < 0) {
2831     CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n");
2832     return rc;
2833   }
2834
2835   bufferlist bl;
2836
2837   rgw_bi_log_entry entry;
2838
2839   entry.timestamp = real_clock::now();
2840   entry.op = RGWModifyOp::CLS_RGW_OP_SYNCSTOP;
2841   entry.state = RGWPendingState::CLS_RGW_STATE_COMPLETE;
2842
2843   string key;
2844   bi_log_index_key(hctx, key, entry.id, header.ver);
2845
2846   encode(entry, bl);
2847
2848   if (entry.id > header.max_marker)
2849     header.max_marker = entry.id;
2850   header.syncstopped = true;
2851
2852   rc = cls_cxx_map_set_val(hctx, key, &bl);
2853   if (rc < 0)
2854     return rc;
2855
2856   return write_bucket_header(hctx, &header);
2857 }
2858
2859
2860 static void usage_record_prefix_by_time(uint64_t epoch, string& key)
2861 {
2862   char buf[32];
2863   snprintf(buf, sizeof(buf), "%011llu", (long long unsigned)epoch);
2864   key = buf;
2865 }
2866
2867 static void usage_record_prefix_by_user(const string& user, uint64_t epoch, string& key)
2868 {
2869   char buf[user.size() + 32];
2870   snprintf(buf, sizeof(buf), "%s_%011llu_", user.c_str(), (long long unsigned)epoch);
2871   key = buf;
2872 }
2873
2874 static void usage_record_name_by_time(uint64_t epoch, const string& user, const string& bucket, string& key)
2875 {
2876   char buf[32 + user.size() + bucket.size()];
2877   snprintf(buf, sizeof(buf), "%011llu_%s_%s", (long long unsigned)epoch, user.c_str(), bucket.c_str());
2878   key = buf;
2879 }
2880
2881 static void usage_record_name_by_user(const string& user, uint64_t epoch, const string& bucket, string& key)
2882 {
2883   char buf[32 + user.size() + bucket.size()];
2884   snprintf(buf, sizeof(buf), "%s_%011llu_%s", user.c_str(), (long long unsigned)epoch, bucket.c_str());
2885   key = buf;
2886 }
2887
2888 static int usage_record_decode(bufferlist& record_bl, rgw_usage_log_entry& e)
2889 {
2890   auto kiter = record_bl.cbegin();
2891   try {
2892     decode(e, kiter);
2893   } catch (buffer::error& err) {
2894     CLS_LOG(1, "ERROR: usage_record_decode(): failed to decode record_bl\n");
2895     return -EINVAL;
2896   }
2897
2898   return 0;
2899 }
2900
2901 int rgw_user_usage_log_add(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2902 {
2903   CLS_LOG(10, "rgw_user_usage_log_add()");
2904
2905   auto in_iter = in->cbegin();
2906   rgw_cls_usage_log_add_op op;
2907
2908   try {
2909     decode(op, in_iter);
2910   } catch (buffer::error& err) {
2911     CLS_LOG(1, "ERROR: rgw_user_usage_log_add(): failed to decode request\n");
2912     return -EINVAL;
2913   }
2914
2915   rgw_usage_log_info& info = op.info;
2916   vector<rgw_usage_log_entry>::iterator iter;
2917
2918   for (iter = info.entries.begin(); iter != info.entries.end(); ++iter) {
2919     rgw_usage_log_entry& entry = *iter;
2920     string key_by_time;
2921
2922     rgw_user *puser = (entry.payer.empty() ? &entry.owner : &entry.payer);
2923
2924     usage_record_name_by_time(entry.epoch, puser->to_str(), entry.bucket, key_by_time);
2925
2926     CLS_LOG(10, "rgw_user_usage_log_add user=%s bucket=%s\n", puser->to_str().c_str(), entry.bucket.c_str());
2927
2928     bufferlist record_bl;
2929     int ret = cls_cxx_map_get_val(hctx, key_by_time, &record_bl);
2930     if (ret < 0 && ret != -ENOENT) {
2931       CLS_LOG(1, "ERROR: rgw_user_usage_log_add(): cls_cxx_map_read_key returned %d\n", ret);
2932       return -EINVAL;
2933     }
2934     if (ret >= 0) {
2935       rgw_usage_log_entry e;
2936       ret = usage_record_decode(record_bl, e);
2937       if (ret < 0)
2938         return ret;
2939       CLS_LOG(10, "rgw_user_usage_log_add aggregating existing bucket\n");
2940       entry.aggregate(e);
2941     }
2942
2943     bufferlist new_record_bl;
2944     encode(entry, new_record_bl);
2945     ret = cls_cxx_map_set_val(hctx, key_by_time, &new_record_bl);
2946     if (ret < 0)
2947       return ret;
2948
2949     string key_by_user;
2950     usage_record_name_by_user(puser->to_str(), entry.epoch, entry.bucket, key_by_user);
2951     ret = cls_cxx_map_set_val(hctx, key_by_user, &new_record_bl);
2952     if (ret < 0)
2953       return ret;
2954   }
2955
2956   return 0;
2957 }
2958
2959 static int usage_iterate_range(cls_method_context_t hctx, uint64_t start, uint64_t end, const string& user,
2960                             const string& bucket, string& key_iter, uint32_t max_entries, bool *truncated,
2961                             int (*cb)(cls_method_context_t, const string&, rgw_usage_log_entry&, void *),
2962                             void *param)
2963 {
2964   CLS_LOG(10, "usage_iterate_range");
2965
2966   map<string, bufferlist> keys;
2967 #define NUM_KEYS 32
2968   string filter_prefix;
2969   string start_key, end_key;
2970   bool by_user = !user.empty();
2971   uint32_t i = 0;
2972   string user_key;
2973   bool truncated_status = false;
2974
2975   ceph_assert(truncated != nullptr);
2976
2977   if (!by_user) {
2978     usage_record_prefix_by_time(end, end_key);
2979   } else {
2980     user_key = user;
2981     user_key.append("_");
2982   }
2983
2984   if (key_iter.empty()) {
2985     if (by_user) {
2986       usage_record_prefix_by_user(user, start, start_key);
2987     } else {
2988       usage_record_prefix_by_time(start, start_key);
2989     }
2990   } else {
2991     start_key = key_iter;
2992   }
2993
2994   CLS_LOG(20, "usage_iterate_range start_key=%s", start_key.c_str());
2995   int ret = cls_cxx_map_get_vals(hctx, start_key, filter_prefix, max_entries, &keys, &truncated_status);
2996   if (ret < 0)
2997     return ret;
2998
2999   *truncated = truncated_status;
3000
3001   map<string, bufferlist>::iterator iter = keys.begin();
3002   if (iter == keys.end())
3003     return 0;
3004
3005   uint32_t num_keys = keys.size();
3006
3007   for (; iter != keys.end(); ++iter,++i) {
3008     const string& key = iter->first;
3009     rgw_usage_log_entry e;
3010
3011     if (!by_user && key.compare(end_key) >= 0) {
3012       CLS_LOG(20, "usage_iterate_range reached key=%s, done", key.c_str());
3013       *truncated = false;
3014       key_iter = key;
3015       return 0;
3016     }
3017
3018     if (by_user && key.compare(0, user_key.size(), user_key) != 0) {
3019       CLS_LOG(20, "usage_iterate_range reached key=%s, done", key.c_str());
3020       *truncated = false;
3021       key_iter = key;
3022       return 0;
3023     }
3024
3025     ret = usage_record_decode(iter->second, e);
3026     if (ret < 0)
3027       return ret;
3028
3029     if (!bucket.empty() && bucket.compare(e.bucket))
3030       continue;
3031
3032     if (e.epoch < start)
3033       continue;
3034
3035     /* keys are sorted by epoch, so once we're past end we're done */
3036     if (e.epoch >= end) {
3037       *truncated = false;
3038       return 0;
3039     }
3040
3041     ret = cb(hctx, key, e, param);
3042     if (ret < 0)
3043       return ret;
3044
3045
3046     if (i == num_keys - 1) {
3047       key_iter = key;
3048       return 0;
3049     }
3050   }
3051   return 0;
3052 }
3053
3054 static int usage_log_read_cb(cls_method_context_t hctx, const string& key, rgw_usage_log_entry& entry, void *param)
3055 {
3056   map<rgw_user_bucket, rgw_usage_log_entry> *usage = (map<rgw_user_bucket, rgw_usage_log_entry> *)param;
3057   rgw_user *puser;
3058   if (!entry.payer.empty()) {
3059     puser = &entry.payer;
3060   } else {
3061     puser = &entry.owner;
3062   }
3063   rgw_user_bucket ub(puser->to_str(), entry.bucket);
3064   rgw_usage_log_entry& le = (*usage)[ub];
3065   le.aggregate(entry);
3066
3067   return 0;
3068 }
3069
3070 int rgw_user_usage_log_read(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3071 {
3072   CLS_LOG(10, "rgw_user_usage_log_read()");
3073
3074   auto in_iter = in->cbegin();
3075   rgw_cls_usage_log_read_op op;
3076
3077   try {
3078     decode(op, in_iter);
3079   } catch (buffer::error& err) {
3080     CLS_LOG(1, "ERROR: rgw_user_usage_log_read(): failed to decode request\n");
3081     return -EINVAL;
3082   }
3083
3084   rgw_cls_usage_log_read_ret ret_info;
3085   map<rgw_user_bucket, rgw_usage_log_entry> *usage = &ret_info.usage;
3086   string iter = op.iter;
3087 #define MAX_ENTRIES 1000
3088   uint32_t max_entries = (op.max_entries ? op.max_entries : MAX_ENTRIES);
3089   int ret = usage_iterate_range(hctx, op.start_epoch, op.end_epoch, op.owner, op.bucket, iter, max_entries, &ret_info.truncated, usage_log_read_cb, (void *)usage);
3090   if (ret < 0)
3091     return ret;
3092
3093   if (ret_info.truncated)
3094     ret_info.next_iter = iter;
3095
3096   encode(ret_info, *out);
3097   return 0;
3098 }
3099
3100 static int usage_log_trim_cb(cls_method_context_t hctx, const string& key, rgw_usage_log_entry& entry, void *param)
3101 {
3102   bool *found = (bool *)param;
3103   if (found) {
3104     *found = true;
3105   }
3106   string key_by_time;
3107   string key_by_user;
3108
3109   string o = entry.owner.to_str();
3110   usage_record_name_by_time(entry.epoch, o, entry.bucket, key_by_time);
3111   usage_record_name_by_user(o, entry.epoch, entry.bucket, key_by_user);
3112
3113   int ret = cls_cxx_map_remove_key(hctx, key_by_time);
3114   if (ret < 0)
3115     return ret;
3116
3117   return cls_cxx_map_remove_key(hctx, key_by_user);
3118 }
3119
3120 int rgw_user_usage_log_trim(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3121 {
3122   CLS_LOG(10, "rgw_user_usage_log_trim()");
3123
3124   /* only continue if object exists! */
3125   int ret = cls_cxx_stat(hctx, NULL, NULL);
3126   if (ret < 0)
3127     return ret;
3128
3129   auto in_iter = in->cbegin();
3130   rgw_cls_usage_log_trim_op op;
3131
3132   try {
3133     decode(op, in_iter);
3134   } catch (buffer::error& err) {
3135     CLS_LOG(1, "ERROR: rgw_user_log_usage_log_trim(): failed to decode request\n");
3136     return -EINVAL;
3137   }
3138
3139   string iter;
3140   bool more;
3141   bool found = false;
3142 #define MAX_USAGE_TRIM_ENTRIES 128
3143   ret = usage_iterate_range(hctx, op.start_epoch, op.end_epoch, op.user, op.bucket, iter, MAX_USAGE_TRIM_ENTRIES, &more, usage_log_trim_cb, (void *)&found);
3144   if (ret < 0)
3145     return ret;
3146
3147   if (!more && !found)
3148     return -ENODATA;
3149
3150   return 0;
3151 }
3152
3153 int rgw_usage_log_clear(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3154 {
3155   CLS_LOG(10,"%s", __func__);
3156
3157   int ret = cls_cxx_map_clear(hctx);
3158   /* if object doesn't exist all the logs are cleared anyway */
3159   if (ret == -ENOENT)
3160     ret = 0;
3161
3162   return ret;
3163 }
3164 /*
3165  * We hold the garbage collection chain data under two different indexes: the first 'name' index
3166  * keeps them under a unique tag that represents the chains, and a second 'time' index keeps
3167  * them by their expiration timestamp
3168  */
3169 #define GC_OBJ_NAME_INDEX 0
3170 #define GC_OBJ_TIME_INDEX 1
3171
3172 static string gc_index_prefixes[] = { "0_",
3173                                       "1_" };
3174
3175 static void prepend_index_prefix(const string& src, int index, string *dest)
3176 {
3177   *dest = gc_index_prefixes[index];
3178   dest->append(src);
3179 }
3180
3181 static int gc_omap_get(cls_method_context_t hctx, int type, const string& key, cls_rgw_gc_obj_info *info)
3182 {
3183   string index;
3184   prepend_index_prefix(key, type, &index);
3185
3186   int ret = read_omap_entry(hctx, index, info);
3187   if (ret < 0)
3188     return ret;
3189
3190   return 0;
3191 }
3192
3193 static int gc_omap_set(cls_method_context_t hctx, int type, const string& key, const cls_rgw_gc_obj_info *info)
3194 {
3195   bufferlist bl;
3196   encode(*info, bl);
3197
3198   string index = gc_index_prefixes[type];
3199   index.append(key);
3200
3201   int ret = cls_cxx_map_set_val(hctx, index, &bl);
3202   if (ret < 0)
3203     return ret;
3204
3205   return 0;
3206 }
3207
3208 static int gc_omap_remove(cls_method_context_t hctx, int type, const string& key)
3209 {
3210   string index = gc_index_prefixes[type];
3211   index.append(key);
3212
3213   int ret = cls_cxx_map_remove_key(hctx, index);
3214   if (ret < 0)
3215     return ret;
3216
3217   return 0;
3218 }
3219
3220 static bool key_in_index(const string& key, int index_type)
3221 {
3222   const string& prefix = gc_index_prefixes[index_type];
3223   return (key.compare(0, prefix.size(), prefix) == 0);
3224 }
3225
3226
3227 static int gc_update_entry(cls_method_context_t hctx, uint32_t expiration_secs,
3228                            cls_rgw_gc_obj_info& info)
3229 {
3230   cls_rgw_gc_obj_info old_info;
3231   int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, info.tag, &old_info);
3232   if (ret == 0) {
3233     string key;
3234     get_time_key(old_info.time, &key);
3235     ret = gc_omap_remove(hctx, GC_OBJ_TIME_INDEX, key);
3236     if (ret < 0 && ret != -ENOENT) {
3237       CLS_LOG(0, "ERROR: failed to remove key=%s\n", key.c_str());
3238       return ret;
3239     }
3240   }
3241
3242   // calculate time and time key
3243   info.time = ceph::real_clock::now();
3244   info.time += make_timespan(expiration_secs);
3245   string time_key;
3246   get_time_key(info.time, &time_key);
3247
3248   if (info.chain.objs.empty()) {
3249     CLS_LOG(0,
3250             "WARNING: %s setting GC log entry with zero-length chain, "
3251             "tag='%s', timekey='%s'",
3252             __func__, info.tag.c_str(), time_key.c_str());
3253   }
3254
3255   ret = gc_omap_set(hctx, GC_OBJ_NAME_INDEX, info.tag, &info);
3256   if (ret < 0)
3257     return ret;
3258
3259   ret = gc_omap_set(hctx, GC_OBJ_TIME_INDEX, time_key, &info);
3260   if (ret < 0)
3261     goto done_err;
3262
3263   return 0;
3264
3265 done_err:
3266
3267   CLS_LOG(0, "ERROR: gc_set_entry error info.tag=%s, ret=%d\n",
3268           info.tag.c_str(), ret);
3269   gc_omap_remove(hctx, GC_OBJ_NAME_INDEX, info.tag);
3270
3271   return ret;
3272 }
3273
3274 static int gc_defer_entry(cls_method_context_t hctx, const string& tag, uint32_t expiration_secs)
3275 {
3276   cls_rgw_gc_obj_info info;
3277   int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, tag, &info);
3278   if (ret == -ENOENT)
3279     return 0;
3280   if (ret < 0)
3281     return ret;
3282   return gc_update_entry(hctx, expiration_secs, info);
3283 }
3284
3285 int gc_record_decode(bufferlist& bl, cls_rgw_gc_obj_info& e)
3286 {
3287   auto iter = bl.cbegin();
3288   try {
3289     decode(e, iter);
3290   } catch (buffer::error& err) {
3291     CLS_LOG(0, "ERROR: failed to decode cls_rgw_gc_obj_info");
3292     return -EIO;
3293   }
3294   return 0;
3295 }
3296
3297 static int rgw_cls_gc_set_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3298 {
3299   auto in_iter = in->cbegin();
3300
3301   cls_rgw_gc_set_entry_op op;
3302   try {
3303     decode(op, in_iter);
3304   } catch (buffer::error& err) {
3305     CLS_LOG(1, "ERROR: rgw_cls_gc_set_entry(): failed to decode entry\n");
3306     return -EINVAL;
3307   }
3308
3309   return gc_update_entry(hctx, op.expiration_secs, op.info);
3310 }
3311
3312 static int rgw_cls_gc_defer_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3313 {
3314   auto in_iter = in->cbegin();
3315
3316   cls_rgw_gc_defer_entry_op op;
3317   try {
3318     decode(op, in_iter);
3319   } catch (buffer::error& err) {
3320     CLS_LOG(1, "ERROR: rgw_cls_gc_defer_entry(): failed to decode entry\n");
3321     return -EINVAL;
3322   }
3323
3324   return gc_defer_entry(hctx, op.tag, op.expiration_secs);
3325 }
3326
3327 static int gc_iterate_entries(cls_method_context_t hctx,
3328                               const string& marker,
3329                               bool expired_only,
3330                               string& out_marker,
3331                               uint32_t max_entries,
3332                               bool *truncated,
3333                               int (*cb)(cls_method_context_t,
3334                                         const string&,
3335                                         cls_rgw_gc_obj_info&,
3336                                         void *),
3337                               void *param)
3338 {
3339   CLS_LOG(10, "gc_iterate_entries");
3340
3341   map<string, bufferlist> keys;
3342   string filter_prefix, end_key;
3343   string key;
3344
3345   if (truncated)
3346     *truncated = false;
3347
3348   string start_key;
3349   if (marker.empty()) {
3350     prepend_index_prefix(marker, GC_OBJ_TIME_INDEX, &start_key);
3351   } else {
3352     start_key = marker;
3353   }
3354
3355   if (expired_only) {
3356     real_time now = ceph::real_clock::now();
3357     string now_str;
3358     get_time_key(now, &now_str);
3359     prepend_index_prefix(now_str, GC_OBJ_TIME_INDEX, &end_key);
3360
3361     CLS_LOG(10, "gc_iterate_entries end_key=%s\n", end_key.c_str());
3362   }
3363
3364   string filter;
3365
3366   int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max_entries,
3367                                  &keys, truncated);
3368   if (ret < 0)
3369     return ret;
3370
3371   map<string, bufferlist>::iterator iter = keys.begin();
3372   if (iter == keys.end()) {
3373     // if keys empty must not come back as truncated
3374     ceph_assert(!truncated || !(*truncated));
3375     return 0;
3376   }
3377
3378   const string* last_key = nullptr; // last key processed, for end-marker
3379   for (; iter != keys.end(); ++iter) {
3380     const string& key = iter->first;
3381     cls_rgw_gc_obj_info e;
3382
3383     CLS_LOG(10, "gc_iterate_entries key=%s\n", key.c_str());
3384
3385     if (!end_key.empty() && key.compare(end_key) >= 0) {
3386       if (truncated)
3387         *truncated = false;
3388       return 0;
3389     }
3390
3391     if (!key_in_index(key, GC_OBJ_TIME_INDEX)) {
3392       if (truncated)
3393         *truncated = false;
3394       return 0;
3395     }
3396
3397     ret = gc_record_decode(iter->second, e);
3398     if (ret < 0)
3399       return ret;
3400
3401     ret = cb(hctx, key, e, param);
3402     if (ret < 0)
3403       return ret;
3404     last_key = &(iter->first); // update when callback successful
3405   }
3406
3407   // set the out marker if either caller does not capture truncated or
3408   // if they do capture and we are truncated
3409   if (!truncated || *truncated) {
3410     assert(last_key);
3411     out_marker = *last_key;
3412   }
3413
3414   return 0;
3415 }
3416
3417 static int gc_list_cb(cls_method_context_t hctx, const string& key, cls_rgw_gc_obj_info& info, void *param)
3418 {
3419   list<cls_rgw_gc_obj_info> *l = (list<cls_rgw_gc_obj_info> *)param;
3420   l->push_back(info);
3421   return 0;
3422 }
3423
3424 static int gc_list_entries(cls_method_context_t hctx, const string& marker,
3425                            uint32_t max, bool expired_only,
3426                            list<cls_rgw_gc_obj_info>& entries, bool *truncated, string& next_marker)
3427 {
3428   int ret = gc_iterate_entries(hctx, marker, expired_only,
3429                               next_marker, max, truncated,
3430                               gc_list_cb, &entries);
3431   return ret;
3432 }
3433
3434 static int rgw_cls_gc_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3435 {
3436   auto in_iter = in->cbegin();
3437
3438   cls_rgw_gc_list_op op;
3439   try {
3440     decode(op, in_iter);
3441   } catch (buffer::error& err) {
3442     CLS_LOG(1, "ERROR: rgw_cls_gc_list(): failed to decode entry\n");
3443     return -EINVAL;
3444   }
3445
3446   cls_rgw_gc_list_ret op_ret;
3447 #define GC_LIST_ENTRIES_DEFAULT 128
3448   int ret = gc_list_entries(hctx, op.marker, (op.max ? op.max : GC_LIST_ENTRIES_DEFAULT), op.expired_only,
3449    op_ret.entries, &op_ret.truncated, op_ret.next_marker);
3450   if (ret < 0)
3451     return ret;
3452
3453   encode(op_ret, *out);
3454
3455   return 0;
3456 }
3457
3458 static int gc_remove(cls_method_context_t hctx, vector<string>& tags)
3459 {
3460   for (auto iter = tags.begin(); iter != tags.end(); ++iter) {
3461     string& tag = *iter;
3462     cls_rgw_gc_obj_info info;
3463     int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, tag, &info);
3464     if (ret == -ENOENT) {
3465       CLS_LOG(0, "couldn't find tag in name index tag=%s\n", tag.c_str());
3466       continue;
3467     }
3468
3469     if (ret < 0)
3470       return ret;
3471
3472     string time_key;
3473     get_time_key(info.time, &time_key);
3474     ret = gc_omap_remove(hctx, GC_OBJ_TIME_INDEX, time_key);
3475     if (ret < 0 && ret != -ENOENT)
3476       return ret;
3477     if (ret == -ENOENT) {
3478       CLS_LOG(0, "couldn't find key in time index key=%s\n", time_key.c_str());
3479     }
3480
3481     ret = gc_omap_remove(hctx, GC_OBJ_NAME_INDEX, tag);
3482     if (ret < 0 && ret != -ENOENT)
3483       return ret;
3484   }
3485
3486   return 0;
3487 }
3488
3489 static int rgw_cls_gc_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3490 {
3491   auto in_iter = in->cbegin();
3492
3493   cls_rgw_gc_remove_op op;
3494   try {
3495     decode(op, in_iter);
3496   } catch (buffer::error& err) {
3497     CLS_LOG(1, "ERROR: rgw_cls_gc_remove(): failed to decode entry\n");
3498     return -EINVAL;
3499   }
3500
3501   return gc_remove(hctx, op.tags);
3502 }
3503
3504 static int rgw_cls_lc_get_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3505 {
3506   auto in_iter = in->cbegin();
3507
3508   cls_rgw_lc_get_entry_op op;
3509   try {
3510     decode(op, in_iter);
3511   } catch (buffer::error& err) {
3512     CLS_LOG(1, "ERROR: rgw_cls_lc_set_entry(): failed to decode entry\n");
3513     return -EINVAL;
3514   }
3515
3516   rgw_lc_entry_t lc_entry;
3517   int ret = read_omap_entry(hctx, op.marker, &lc_entry);
3518   if (ret < 0)
3519     return ret;
3520
3521   cls_rgw_lc_get_entry_ret op_ret(std::move(lc_entry));
3522   encode(op_ret, *out);
3523   return 0;
3524 }
3525
3526
3527 static int rgw_cls_lc_set_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3528 {
3529   auto in_iter = in->cbegin();
3530
3531   cls_rgw_lc_set_entry_op op;
3532   try {
3533     decode(op, in_iter);
3534   } catch (buffer::error& err) {
3535     CLS_LOG(1, "ERROR: rgw_cls_lc_set_entry(): failed to decode entry\n");
3536     return -EINVAL;
3537   }
3538
3539   bufferlist bl;
3540   encode(op.entry, bl);
3541
3542   int ret = cls_cxx_map_set_val(hctx, op.entry.first, &bl);
3543   return ret;
3544 }
3545
3546 static int rgw_cls_lc_rm_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3547 {
3548   auto in_iter = in->cbegin();
3549
3550   cls_rgw_lc_rm_entry_op op;
3551   try {
3552     decode(op, in_iter);
3553   } catch (buffer::error& err) {
3554     CLS_LOG(1, "ERROR: rgw_cls_lc_rm_entry(): failed to decode entry\n");
3555     return -EINVAL;
3556   }
3557
3558   int ret = cls_cxx_map_remove_key(hctx, op.entry.first);
3559   return ret;
3560 }
3561
3562 static int rgw_cls_lc_get_next_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3563 {
3564   auto in_iter = in->cbegin();
3565   cls_rgw_lc_get_next_entry_ret op_ret;
3566   cls_rgw_lc_get_next_entry_op op;
3567   try {
3568     decode(op, in_iter);
3569   } catch (buffer::error& err) {
3570     CLS_LOG(1, "ERROR: rgw_cls_lc_get_next_entry: failed to decode op\n");
3571     return -EINVAL;
3572   }
3573
3574   map<string, bufferlist> vals;
3575   string filter_prefix;
3576   bool more;
3577   int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, 1, &vals, &more);
3578   if (ret < 0)
3579     return ret;
3580   map<string, bufferlist>::iterator it;
3581   pair<string, int> entry;
3582   if (!vals.empty()) {
3583     it=vals.begin();
3584     in_iter = it->second.begin();
3585     try {
3586       decode(entry, in_iter);
3587     } catch (buffer::error& err) {
3588       CLS_LOG(1, "ERROR: rgw_cls_lc_get_next_entry(): failed to decode entry\n");
3589       return -EIO;
3590     }
3591   }
3592   op_ret.entry = entry;
3593   encode(op_ret, *out);
3594   return 0;
3595 }
3596
3597 static int rgw_cls_lc_list_entries(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3598 {
3599   cls_rgw_lc_list_entries_op op;
3600   auto in_iter = in->cbegin();
3601   try {
3602     decode(op, in_iter);
3603   } catch (buffer::error& err) {
3604     CLS_LOG(1, "ERROR: rgw_cls_lc_list_entries(): failed to decode op\n");
3605     return -EINVAL;
3606   }
3607
3608   cls_rgw_lc_list_entries_ret op_ret;
3609   bufferlist::const_iterator iter;
3610   map<string, bufferlist> vals;
3611   string filter_prefix;
3612   int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, op.max_entries, &vals, &op_ret.is_truncated);
3613   if (ret < 0)
3614     return ret;
3615   map<string, bufferlist>::iterator it;
3616   pair<string, int> entry;
3617   for (it = vals.begin(); it != vals.end(); ++it) {
3618     iter = it->second.cbegin();
3619     try {
3620     decode(entry, iter);
3621     } catch (buffer::error& err) {
3622     CLS_LOG(1, "ERROR: rgw_cls_lc_list_entries(): failed to decode entry\n");
3623     return -EIO;
3624    }
3625    op_ret.entries.insert(entry);
3626   }
3627   encode(op_ret, *out);
3628   return 0;
3629 }
3630
3631 static int rgw_cls_lc_put_head(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3632 {
3633   auto in_iter = in->cbegin();
3634
3635   cls_rgw_lc_put_head_op op;
3636   try {
3637     decode(op, in_iter);
3638   } catch (buffer::error& err) {
3639     CLS_LOG(1, "ERROR: rgw_cls_lc_put_head(): failed to decode entry\n");
3640     return -EINVAL;
3641   }
3642
3643   bufferlist bl;
3644   encode(op.head, bl);
3645   int ret = cls_cxx_map_write_header(hctx,&bl);
3646   return ret;
3647 }
3648
3649 static int rgw_cls_lc_get_head(cls_method_context_t hctx, bufferlist *in,  bufferlist *out)
3650 {
3651   bufferlist bl;
3652   int ret = cls_cxx_map_read_header(hctx, &bl);
3653   if (ret < 0)
3654     return ret;
3655   cls_rgw_lc_obj_head head;
3656   if (bl.length() != 0) {
3657     auto iter = bl.cbegin();
3658     try {
3659       decode(head, iter);
3660     } catch (buffer::error& err) {
3661       CLS_LOG(0, "ERROR: rgw_cls_lc_get_head(): failed to decode entry %s\n",err.what());
3662       return -EINVAL;
3663     }
3664   } else {
3665     head.start_date = 0;
3666     head.marker.clear();
3667   }
3668   cls_rgw_lc_get_head_ret op_ret;
3669   op_ret.head = head;
3670   encode(op_ret, *out);
3671   return 0;
3672 }
3673
3674 static int rgw_reshard_add(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3675 {
3676   auto in_iter = in->cbegin();
3677
3678   cls_rgw_reshard_add_op op;
3679   try {
3680     decode(op, in_iter);
3681   } catch (buffer::error& err) {
3682     CLS_LOG(1, "ERROR: rgw_reshard_add: failed to decode entry\n");
3683     return -EINVAL;
3684   }
3685
3686
3687   string key;
3688   op.entry.get_key(&key);
3689
3690   bufferlist bl;
3691   encode(op.entry, bl);
3692   int ret = cls_cxx_map_set_val(hctx, key, &bl);
3693   if (ret < 0) {
3694     CLS_ERR("error adding reshard job for bucket %s with key %s",op.entry.bucket_name.c_str(), key.c_str());
3695     return ret;
3696   }
3697
3698   return ret;
3699 }
3700
3701 static int rgw_reshard_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3702 {
3703   cls_rgw_reshard_list_op op;
3704   auto in_iter = in->cbegin();
3705   try {
3706     decode(op, in_iter);
3707   } catch (buffer::error& err) {
3708     CLS_LOG(1, "ERROR: rgw_cls_rehard_list(): failed to decode entry\n");
3709     return -EINVAL;
3710   }
3711   cls_rgw_reshard_list_ret op_ret;
3712   bufferlist::const_iterator iter;
3713   map<string, bufferlist> vals;
3714   string filter_prefix;
3715 #define MAX_RESHARD_LIST_ENTRIES 1000
3716   /* one extra entry for identifying truncation */
3717   int32_t max = (op.max && (op.max < MAX_RESHARD_LIST_ENTRIES) ? op.max : MAX_RESHARD_LIST_ENTRIES);
3718   int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, max, &vals, &op_ret.is_truncated);
3719   if (ret < 0)
3720     return ret;
3721   map<string, bufferlist>::iterator it;
3722   cls_rgw_reshard_entry entry;
3723   int i = 0;
3724   for (it = vals.begin(); i < (int)op.max && it != vals.end(); ++it, ++i) {
3725     iter = it->second.cbegin();
3726     try {
3727       decode(entry, iter);
3728     } catch (buffer::error& err) {
3729       CLS_LOG(1, "ERROR: rgw_cls_rehard_list(): failed to decode entry\n");
3730       return -EIO;
3731    }
3732     op_ret.entries.push_back(entry);
3733   }
3734   encode(op_ret, *out);
3735   return 0;
3736 }
3737
3738 static int rgw_reshard_get(cls_method_context_t hctx, bufferlist *in,  bufferlist *out)
3739 {
3740   auto in_iter = in->cbegin();
3741
3742   cls_rgw_reshard_get_op op;
3743   try {
3744     decode(op, in_iter);
3745   } catch (buffer::error& err) {
3746     CLS_LOG(1, "ERROR: rgw_reshard_get: failed to decode entry\n");
3747     return -EINVAL;
3748   }
3749
3750   string key;
3751   cls_rgw_reshard_entry  entry;
3752   op.entry.get_key(&key);
3753   int ret = read_omap_entry(hctx, key, &entry);
3754   if (ret < 0) {
3755     return ret;
3756   }
3757
3758   cls_rgw_reshard_get_ret op_ret;
3759   op_ret.entry = entry;
3760   encode(op_ret, *out);
3761   return 0;
3762 }
3763
3764 static int rgw_reshard_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3765 {
3766   auto in_iter = in->cbegin();
3767
3768   cls_rgw_reshard_remove_op op;
3769   try {
3770     decode(op, in_iter);
3771   } catch (buffer::error& err) {
3772     CLS_LOG(1, "ERROR: rgw_cls_rehard_remove: failed to decode entry\n");
3773     return -EINVAL;
3774   }
3775
3776   string key;
3777   cls_rgw_reshard_entry  entry;
3778   cls_rgw_reshard_entry::generate_key(op.tenant, op.bucket_name, &key);
3779   int ret = read_omap_entry(hctx, key, &entry);
3780   if (ret < 0) {
3781     return ret;
3782   }
3783
3784   if (!op.bucket_id.empty() &&
3785       entry.bucket_id != op.bucket_id) {
3786     return 0;
3787   }
3788
3789   ret = cls_cxx_map_remove_key(hctx, key);
3790   if (ret < 0) {
3791     CLS_LOG(0, "ERROR: failed to remove key: key=%s ret=%d", key.c_str(), ret);
3792     return 0;
3793   }
3794   return ret;
3795 }
3796
3797 static int rgw_set_bucket_resharding(cls_method_context_t hctx, bufferlist *in,  bufferlist *out)
3798 {
3799   cls_rgw_set_bucket_resharding_op op;
3800
3801   auto in_iter = in->cbegin();
3802   try {
3803     decode(op, in_iter);
3804   } catch (buffer::error& err) {
3805     CLS_LOG(1, "ERROR: cls_rgw_set_bucket_resharding: failed to decode entry\n");
3806     return -EINVAL;
3807   }
3808
3809   rgw_bucket_dir_header header;
3810   int rc = read_bucket_header(hctx, &header);
3811   if (rc < 0) {
3812     CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3813     return rc;
3814   }
3815
3816   header.new_instance.set_status(op.entry.new_bucket_instance_id, op.entry.num_shards, op.entry.reshard_status);
3817
3818   return write_bucket_header(hctx, &header);
3819 }
3820
3821 static int rgw_clear_bucket_resharding(cls_method_context_t hctx, bufferlist *in,  bufferlist *out)
3822 {
3823   cls_rgw_clear_bucket_resharding_op op;
3824
3825   auto in_iter = in->cbegin();
3826   try {
3827     decode(op, in_iter);
3828   } catch (buffer::error& err) {
3829     CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n");
3830     return -EINVAL;
3831   }
3832
3833   rgw_bucket_dir_header header;
3834   int rc = read_bucket_header(hctx, &header);
3835   if (rc < 0) {
3836     CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3837     return rc;
3838   }
3839   header.new_instance.clear();
3840
3841   return write_bucket_header(hctx, &header);
3842 }
3843
3844 static int rgw_guard_bucket_resharding(cls_method_context_t hctx, bufferlist *in,  bufferlist *out)
3845 {
3846   cls_rgw_guard_bucket_resharding_op op;
3847
3848   auto in_iter = in->cbegin();
3849   try {
3850     decode(op, in_iter);
3851   } catch (buffer::error& err) {
3852     CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n");
3853     return -EINVAL;
3854   }
3855
3856   rgw_bucket_dir_header header;
3857   int rc = read_bucket_header(hctx, &header);
3858   if (rc < 0) {
3859     CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3860     return rc;
3861   }
3862
3863   if (header.resharding()) {
3864     return op.ret_err;
3865   }
3866
3867   return 0;
3868 }
3869
3870 static int rgw_get_bucket_resharding(cls_method_context_t hctx,
3871                                      bufferlist *in, bufferlist *out)
3872 {
3873   cls_rgw_get_bucket_resharding_op op;
3874
3875   auto in_iter = in->cbegin();
3876   try {
3877     decode(op, in_iter);
3878   } catch (buffer::error& err) {
3879     CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n");
3880     return -EINVAL;
3881   }
3882
3883   rgw_bucket_dir_header header;
3884   int rc = read_bucket_header(hctx, &header);
3885   if (rc < 0) {
3886     CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3887     return rc;
3888   }
3889
3890   cls_rgw_get_bucket_resharding_ret op_ret;
3891   op_ret.new_instance = header.new_instance;
3892
3893   encode(op_ret, *out);
3894
3895   return 0;
3896 }
3897
3898 CLS_INIT(rgw)
3899 {
3900   CLS_LOG(1, "Loaded rgw class!");
3901
3902   cls_handle_t h_class;
3903   cls_method_handle_t h_rgw_bucket_init_index;
3904   cls_method_handle_t h_rgw_bucket_set_tag_timeout;
3905   cls_method_handle_t h_rgw_bucket_list;
3906   cls_method_handle_t h_rgw_bucket_check_index;
3907   cls_method_handle_t h_rgw_bucket_rebuild_index;
3908   cls_method_handle_t h_rgw_bucket_update_stats;
3909   cls_method_handle_t h_rgw_bucket_prepare_op;
3910   cls_method_handle_t h_rgw_bucket_complete_op;
3911   cls_method_handle_t h_rgw_bucket_link_olh;
3912   cls_method_handle_t h_rgw_bucket_unlink_instance_op;
3913   cls_method_handle_t h_rgw_bucket_read_olh_log;
3914   cls_method_handle_t h_rgw_bucket_trim_olh_log;
3915   cls_method_handle_t h_rgw_bucket_clear_olh;
3916   cls_method_handle_t h_rgw_obj_remove;
3917   cls_method_handle_t h_rgw_obj_store_pg_ver;
3918   cls_method_handle_t h_rgw_obj_check_attrs_prefix;
3919   cls_method_handle_t h_rgw_obj_check_mtime;
3920   cls_method_handle_t h_rgw_bi_get_op;
3921   cls_method_handle_t h_rgw_bi_put_op;
3922   cls_method_handle_t h_rgw_bi_list_op;
3923   cls_method_handle_t h_rgw_bi_log_list_op;
3924   cls_method_handle_t h_rgw_bi_log_resync_op;
3925   cls_method_handle_t h_rgw_bi_log_stop_op;
3926   cls_method_handle_t h_rgw_dir_suggest_changes;
3927   cls_method_handle_t h_rgw_user_usage_log_add;
3928   cls_method_handle_t h_rgw_user_usage_log_read;
3929   cls_method_handle_t h_rgw_user_usage_log_trim;
3930   cls_method_handle_t h_rgw_usage_log_clear;
3931   cls_method_handle_t h_rgw_gc_set_entry;
3932   cls_method_handle_t h_rgw_gc_list;
3933   cls_method_handle_t h_rgw_gc_remove;
3934   cls_method_handle_t h_rgw_lc_get_entry;
3935   cls_method_handle_t h_rgw_lc_set_entry;
3936   cls_method_handle_t h_rgw_lc_rm_entry;
3937   cls_method_handle_t h_rgw_lc_get_next_entry;
3938   cls_method_handle_t h_rgw_lc_put_head;
3939   cls_method_handle_t h_rgw_lc_get_head;
3940   cls_method_handle_t h_rgw_lc_list_entries;
3941   cls_method_handle_t h_rgw_reshard_add;
3942   cls_method_handle_t h_rgw_reshard_list;
3943   cls_method_handle_t h_rgw_reshard_get;
3944   cls_method_handle_t h_rgw_reshard_remove;
3945   cls_method_handle_t h_rgw_set_bucket_resharding;
3946   cls_method_handle_t h_rgw_clear_bucket_resharding;
3947   cls_method_handle_t h_rgw_guard_bucket_resharding;
3948   cls_method_handle_t h_rgw_get_bucket_resharding;
3949
3950
3951   cls_register(RGW_CLASS, &h_class);
3952
3953   /* bucket index */
3954   cls_register_cxx_method(h_class, RGW_BUCKET_INIT_INDEX, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_init_index, &h_rgw_bucket_init_index);
3955   cls_register_cxx_method(h_class, RGW_BUCKET_SET_TAG_TIMEOUT, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_set_tag_timeout, &h_rgw_bucket_set_tag_timeout);
3956   cls_register_cxx_method(h_class, RGW_BUCKET_LIST, CLS_METHOD_RD, rgw_bucket_list, &h_rgw_bucket_list);
3957   cls_register_cxx_method(h_class, RGW_BUCKET_CHECK_INDEX, CLS_METHOD_RD, rgw_bucket_check_index, &h_rgw_bucket_check_index);
3958   cls_register_cxx_method(h_class, RGW_BUCKET_REBUILD_INDEX, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_rebuild_index, &h_rgw_bucket_rebuild_index);
3959   cls_register_cxx_method(h_class, RGW_BUCKET_UPDATE_STATS, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_update_stats, &h_rgw_bucket_update_stats);
3960   cls_register_cxx_method(h_class, RGW_BUCKET_PREPARE_OP, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_prepare_op, &h_rgw_bucket_prepare_op);
3961   cls_register_cxx_method(h_class, RGW_BUCKET_COMPLETE_OP, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_complete_op, &h_rgw_bucket_complete_op);
3962   cls_register_cxx_method(h_class, RGW_BUCKET_LINK_OLH, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_link_olh, &h_rgw_bucket_link_olh);
3963   cls_register_cxx_method(h_class, RGW_BUCKET_UNLINK_INSTANCE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_unlink_instance, &h_rgw_bucket_unlink_instance_op);
3964   cls_register_cxx_method(h_class, RGW_BUCKET_READ_OLH_LOG, CLS_METHOD_RD, rgw_bucket_read_olh_log, &h_rgw_bucket_read_olh_log);
3965   cls_register_cxx_method(h_class, RGW_BUCKET_TRIM_OLH_LOG, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_trim_olh_log, &h_rgw_bucket_trim_olh_log);
3966   cls_register_cxx_method(h_class, RGW_BUCKET_CLEAR_OLH, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_clear_olh, &h_rgw_bucket_clear_olh);
3967
3968   cls_register_cxx_method(h_class, RGW_OBJ_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_obj_remove, &h_rgw_obj_remove);
3969   cls_register_cxx_method(h_class, RGW_OBJ_STORE_PG_VER, CLS_METHOD_WR, rgw_obj_store_pg_ver, &h_rgw_obj_store_pg_ver);
3970   cls_register_cxx_method(h_class, RGW_OBJ_CHECK_ATTRS_PREFIX, CLS_METHOD_RD, rgw_obj_check_attrs_prefix, &h_rgw_obj_check_attrs_prefix);
3971   cls_register_cxx_method(h_class, RGW_OBJ_CHECK_MTIME, CLS_METHOD_RD, rgw_obj_check_mtime, &h_rgw_obj_check_mtime);
3972
3973   cls_register_cxx_method(h_class, RGW_BI_GET, CLS_METHOD_RD, rgw_bi_get_op, &h_rgw_bi_get_op);
3974   cls_register_cxx_method(h_class, RGW_BI_PUT, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_put_op, &h_rgw_bi_put_op);
3975   cls_register_cxx_method(h_class, RGW_BI_LIST, CLS_METHOD_RD, rgw_bi_list_op, &h_rgw_bi_list_op);
3976
3977   cls_register_cxx_method(h_class, RGW_BI_LOG_LIST, CLS_METHOD_RD, rgw_bi_log_list, &h_rgw_bi_log_list_op);
3978   cls_register_cxx_method(h_class, RGW_BI_LOG_TRIM, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_trim, &h_rgw_bi_log_list_op);
3979   cls_register_cxx_method(h_class, RGW_DIR_SUGGEST_CHANGES, CLS_METHOD_RD | CLS_METHOD_WR, rgw_dir_suggest_changes, &h_rgw_dir_suggest_changes);
3980
3981   cls_register_cxx_method(h_class, RGW_BI_LOG_RESYNC, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_resync, &h_rgw_bi_log_resync_op);
3982   cls_register_cxx_method(h_class, RGW_BI_LOG_STOP, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_stop, &h_rgw_bi_log_stop_op);
3983
3984   /* usage logging */
3985   cls_register_cxx_method(h_class, RGW_USER_USAGE_LOG_ADD, CLS_METHOD_RD | CLS_METHOD_WR, rgw_user_usage_log_add, &h_rgw_user_usage_log_add);
3986   cls_register_cxx_method(h_class, RGW_USER_USAGE_LOG_READ, CLS_METHOD_RD, rgw_user_usage_log_read, &h_rgw_user_usage_log_read);
3987   cls_register_cxx_method(h_class, RGW_USER_USAGE_LOG_TRIM, CLS_METHOD_RD | CLS_METHOD_WR, rgw_user_usage_log_trim, &h_rgw_user_usage_log_trim);
3988   cls_register_cxx_method(h_class, RGW_USAGE_LOG_CLEAR, CLS_METHOD_WR, rgw_usage_log_clear, &h_rgw_usage_log_clear);
3989
3990   /* garbage collection */
3991   cls_register_cxx_method(h_class, RGW_GC_SET_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_set_entry, &h_rgw_gc_set_entry);
3992   cls_register_cxx_method(h_class, RGW_GC_DEFER_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_defer_entry, &h_rgw_gc_set_entry);
3993   cls_register_cxx_method(h_class, RGW_GC_LIST, CLS_METHOD_RD, rgw_cls_gc_list, &h_rgw_gc_list);
3994   cls_register_cxx_method(h_class, RGW_GC_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_remove, &h_rgw_gc_remove);
3995
3996   /* lifecycle bucket list */
3997   cls_register_cxx_method(h_class, RGW_LC_GET_ENTRY, CLS_METHOD_RD, rgw_cls_lc_get_entry, &h_rgw_lc_get_entry);
3998   cls_register_cxx_method(h_class, RGW_LC_SET_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_lc_set_entry, &h_rgw_lc_set_entry);
3999   cls_register_cxx_method(h_class, RGW_LC_RM_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_lc_rm_entry, &h_rgw_lc_rm_entry);
4000   cls_register_cxx_method(h_class, RGW_LC_GET_NEXT_ENTRY, CLS_METHOD_RD, rgw_cls_lc_get_next_entry, &h_rgw_lc_get_next_entry);
4001   cls_register_cxx_method(h_class, RGW_LC_PUT_HEAD, CLS_METHOD_RD| CLS_METHOD_WR, rgw_cls_lc_put_head, &h_rgw_lc_put_head);
4002   cls_register_cxx_method(h_class, RGW_LC_GET_HEAD, CLS_METHOD_RD, rgw_cls_lc_get_head, &h_rgw_lc_get_head);
4003   cls_register_cxx_method(h_class, RGW_LC_LIST_ENTRIES, CLS_METHOD_RD, rgw_cls_lc_list_entries, &h_rgw_lc_list_entries);
4004
4005   /* resharding */
4006   cls_register_cxx_method(h_class, RGW_RESHARD_ADD, CLS_METHOD_RD | CLS_METHOD_WR, rgw_reshard_add, &h_rgw_reshard_add);
4007   cls_register_cxx_method(h_class, RGW_RESHARD_LIST, CLS_METHOD_RD, rgw_reshard_list, &h_rgw_reshard_list);
4008   cls_register_cxx_method(h_class, RGW_RESHARD_GET, CLS_METHOD_RD,rgw_reshard_get, &h_rgw_reshard_get);
4009   cls_register_cxx_method(h_class, RGW_RESHARD_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_reshard_remove, &h_rgw_reshard_remove);
4010
4011   /* resharding attribute  */
4012   cls_register_cxx_method(h_class, RGW_SET_BUCKET_RESHARDING, CLS_METHOD_RD | CLS_METHOD_WR,
4013                           rgw_set_bucket_resharding, &h_rgw_set_bucket_resharding);
4014   cls_register_cxx_method(h_class, RGW_CLEAR_BUCKET_RESHARDING, CLS_METHOD_RD | CLS_METHOD_WR,
4015                           rgw_clear_bucket_resharding, &h_rgw_clear_bucket_resharding);
4016   cls_register_cxx_method(h_class, RGW_GUARD_BUCKET_RESHARDING, CLS_METHOD_RD ,
4017                           rgw_guard_bucket_resharding, &h_rgw_guard_bucket_resharding);
4018   cls_register_cxx_method(h_class, RGW_GET_BUCKET_RESHARDING, CLS_METHOD_RD ,
4019                           rgw_get_bucket_resharding, &h_rgw_get_bucket_resharding);
4020
4021   return;
4022 }
4023