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