1 // -*- mode:C; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 #include "include/types.h"
8 #include "objclass/objclass.h"
9 #include "cls/rgw/cls_rgw_ops.h"
10 #include "cls/rgw/cls_rgw_const.h"
11 #include "common/Clock.h"
12 #include "common/strtol.h"
13 #include "common/escape.h"
15 #include "include/compat.h"
21 #define BI_PREFIX_CHAR 0x80
23 #define BI_BUCKET_OBJS_INDEX 0
24 #define BI_BUCKET_LOG_INDEX 1
25 #define BI_BUCKET_OBJ_INSTANCE_INDEX 2
26 #define BI_BUCKET_OLH_DATA_INDEX 3
28 #define BI_BUCKET_LAST_INDEX 4
30 static string bucket_index_prefixes[] = { "", /* special handling for the objs list index */
31 "0_", /* bucket log index */
32 "1000_", /* obj instance index */
33 "1001_", /* olh data index */
35 /* this must be the last index */
38 static bool bi_is_objs_index(const string& s) {
39 return ((unsigned char)s[0] != BI_PREFIX_CHAR);
42 int bi_entry_type(const string& s)
44 if (bi_is_objs_index(s)) {
45 return BI_BUCKET_OBJS_INDEX;
49 i < sizeof(bucket_index_prefixes) / sizeof(bucket_index_prefixes[0]);
51 const string& t = bucket_index_prefixes[i];
53 if (s.compare(1, t.size(), t) == 0) {
61 static bool bi_entry_gt(const string& first, const string& second)
63 int fi = bi_entry_type(first);
64 int si = bi_entry_type(second);
72 return first > second;
75 static void get_time_key(real_time& ut, string *key)
78 ceph_timespec ts = ceph::real_clock::to_ceph_timespec(ut);
79 snprintf(buf, 32, "%011llu.%09u", (unsigned long long)ts.tv_sec, (unsigned int)ts.tv_nsec);
83 static void get_index_ver_key(cls_method_context_t hctx, uint64_t index_ver, string *key)
86 snprintf(buf, sizeof(buf), "%011llu.%llu.%d", (unsigned long long)index_ver,
87 (unsigned long long)cls_current_version(hctx),
88 cls_current_subop_num(hctx));
92 static void bi_log_prefix(string& key)
95 key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
98 static void bi_log_index_key(cls_method_context_t hctx, string& key, string& id, uint64_t index_ver)
101 get_index_ver_key(hctx, index_ver, &id);
105 static int log_index_operation(cls_method_context_t hctx, cls_rgw_obj_key& obj_key, RGWModifyOp op,
106 string& tag, real_time& timestamp,
107 rgw_bucket_entry_ver& ver, RGWPendingState state, uint64_t index_ver,
108 string& max_marker, uint16_t bilog_flags, string *owner, string *owner_display_name, rgw_zone_set *zones_trace)
112 struct rgw_bi_log_entry entry;
114 entry.object = obj_key.name;
115 entry.instance = obj_key.instance;
116 entry.timestamp = timestamp;
120 entry.index_ver = index_ver;
122 entry.bilog_flags = bilog_flags;
124 entry.owner = *owner;
126 if (owner_display_name) {
127 entry.owner_display_name = *owner_display_name;
130 entry.zones_trace = std::move(*zones_trace);
134 bi_log_index_key(hctx, key, entry.id, index_ver);
138 if (entry.id > max_marker)
139 max_marker = entry.id;
141 return cls_cxx_map_set_val(hctx, key, &bl);
145 * read list of objects, skips objects in the ugly namespace
147 static int get_obj_vals(cls_method_context_t hctx, const string& start, const string& filter_prefix,
148 int num_entries, map<string, bufferlist> *pkeys, bool *pmore)
150 int ret = cls_cxx_map_get_vals(hctx, start, filter_prefix, num_entries, pkeys, pmore);
157 map<string, bufferlist>::reverse_iterator last_element = pkeys->rbegin();
158 if ((unsigned char)last_element->first[0] < BI_PREFIX_CHAR) {
159 /* nothing to see here, move along */
163 map<string, bufferlist>::iterator first_element = pkeys->begin();
164 if ((unsigned char)first_element->first[0] > BI_PREFIX_CHAR) {
168 /* let's rebuild the list, only keep entries we're interested in */
169 map<string, bufferlist> old_keys;
170 old_keys.swap(*pkeys);
172 for (map<string, bufferlist>::iterator iter = old_keys.begin(); iter != old_keys.end(); ++iter) {
173 if ((unsigned char)iter->first[0] != BI_PREFIX_CHAR) {
174 (*pkeys)[iter->first] = iter->second;
178 if (num_entries == (int)pkeys->size())
181 map<string, bufferlist> new_keys;
182 char c[] = { (char)(BI_PREFIX_CHAR + 1), 0 };
183 string new_start = c;
185 /* now get some more keys */
186 ret = cls_cxx_map_get_vals(hctx, new_start, filter_prefix, num_entries - pkeys->size(), &new_keys, pmore);
190 for (map<string, bufferlist>::iterator iter = new_keys.begin(); iter != new_keys.end(); ++iter) {
191 (*pkeys)[iter->first] = iter->second;
198 * get a monotonically decreasing string representation.
199 * For num = x, num = y, where x > y, str(x) < str(y)
200 * Another property is that string size starts short and grows as num increases
202 static void decreasing_str(uint64_t num, string *str)
205 if (num < 0x10) { /* 16 */
206 snprintf(buf, sizeof(buf), "9%02lld", 15 - (long long)num);
207 } else if (num < 0x100) { /* 256 */
208 snprintf(buf, sizeof(buf), "8%03lld", 255 - (long long)num);
209 } else if (num < 0x1000) /* 4096 */ {
210 snprintf(buf, sizeof(buf), "7%04lld", 4095 - (long long)num);
211 } else if (num < 0x10000) /* 65536 */ {
212 snprintf(buf, sizeof(buf), "6%05lld", 65535 - (long long)num);
213 } else if (num < 0x100000000) /* 4G */ {
214 snprintf(buf, sizeof(buf), "5%010lld", 0xFFFFFFFF - (long long)num);
216 snprintf(buf, sizeof(buf), "4%020lld", (long long)-num);
223 * we now hold two different indexes for objects. The first one holds the list of objects in the
224 * order that we want them to be listed. The second one only holds the objects instances (for
225 * versioned objects), and they're not arranged in any particular order.
226 * When listing objects we'll use the first index, when doing operations on the objects themselves
227 * we'll use the second index. Note that regular objects only map to the first index anyway
230 static void get_list_index_key(struct rgw_bucket_dir_entry& entry, string *index_key)
232 *index_key = entry.key.name;
235 decreasing_str(entry.versioned_epoch, &ver_str);
236 string instance_delim("\0i", 2);
237 string ver_delim("\0v", 2);
239 index_key->append(ver_delim);
240 index_key->append(ver_str);
241 index_key->append(instance_delim);
242 index_key->append(entry.key.instance);
245 static void encode_obj_versioned_data_key(const cls_rgw_obj_key& key, string *index_key, bool append_delete_marker_suffix = false)
247 *index_key = BI_PREFIX_CHAR;
248 index_key->append(bucket_index_prefixes[BI_BUCKET_OBJ_INSTANCE_INDEX]);
249 index_key->append(key.name);
250 string delim("\0i", 2);
251 index_key->append(delim);
252 index_key->append(key.instance);
253 if (append_delete_marker_suffix) {
255 index_key->append(dm);
259 static void encode_obj_index_key(const cls_rgw_obj_key& key, string *index_key)
261 if (key.instance.empty()) {
262 *index_key = key.name;
264 encode_obj_versioned_data_key(key, index_key);
268 static void encode_olh_data_key(const cls_rgw_obj_key& key, string *index_key)
270 *index_key = BI_PREFIX_CHAR;
271 index_key->append(bucket_index_prefixes[BI_BUCKET_OLH_DATA_INDEX]);
272 index_key->append(key.name);
276 static int read_index_entry(cls_method_context_t hctx, string& name, T *entry);
278 static int encode_list_index_key(cls_method_context_t hctx, const cls_rgw_obj_key& key, string *index_key)
280 if (key.instance.empty()) {
281 *index_key = key.name;
285 string obj_index_key;
286 encode_obj_index_key(key, &obj_index_key);
288 rgw_bucket_dir_entry entry;
290 int ret = read_index_entry(hctx, obj_index_key, &entry);
291 if (ret == -ENOENT) {
292 /* couldn't find the entry, set key value after the current object */
293 char buf[2] = { 0x1, 0 };
295 *index_key = key.name + s;
299 CLS_LOG(1, "ERROR: encode_list_index_key(): cls_cxx_map_get_val returned %d\n", ret);
303 get_list_index_key(entry, index_key);
308 static void split_key(const string& key, list<string>& vals)
311 const char *p = key.c_str();
312 while (pos < key.size()) {
313 size_t len = strlen(p);
321 * list index key structure:
323 * <obj name>\0[v<ver>\0i<instance id>]
325 static void decode_list_index_key(const string& index_key, cls_rgw_obj_key *key, uint64_t *ver)
327 size_t len = strlen(index_key.c_str());
329 key->instance.clear();
332 if (len == index_key.size()) {
333 key->name = index_key;
338 split_key(index_key, vals);
340 assert(!vals.empty());
342 list<string>::iterator iter = vals.begin();
346 assert(iter != vals.end());
348 for (; iter != vals.end(); ++iter) {
351 key->instance = val.substr(1);
352 } else if (val[0] == 'v') {
354 const char *s = val.c_str() + 1;
355 *ver = strict_strtoll(s, 10, &err);
361 static int read_bucket_header(cls_method_context_t hctx, struct rgw_bucket_dir_header *header)
364 int rc = cls_cxx_map_read_header(hctx, &bl);
368 if (bl.length() == 0) {
369 *header = rgw_bucket_dir_header();
372 bufferlist::iterator iter = bl.begin();
374 ::decode(*header, iter);
375 } catch (buffer::error& err) {
376 CLS_LOG(1, "ERROR: read_bucket_header(): failed to decode header\n");
383 int rgw_bucket_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
385 bufferlist::iterator iter = in->begin();
387 struct rgw_cls_list_op op;
390 } catch (buffer::error& err) {
391 CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode request\n");
395 struct rgw_cls_list_ret ret;
396 struct rgw_bucket_dir& new_dir = ret.dir;
397 int rc = read_bucket_header(hctx, &new_dir.header);
399 CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to read header\n");
403 map<string, bufferlist> keys;
404 std::map<string, bufferlist>::iterator kiter;
406 encode_list_index_key(hctx, op.start_obj, &start_key);
408 uint32_t left_to_read = op.num_entries;
412 rc = get_obj_vals(hctx, start_key, op.filter_prefix, left_to_read, &keys, &more);
416 std::map<string, struct rgw_bucket_dir_entry>& m = new_dir.m;
420 for (kiter = keys.begin(); kiter != keys.end(); ++kiter) {
421 struct rgw_bucket_dir_entry entry;
423 if (!bi_is_objs_index(kiter->first)) {
428 bufferlist& entrybl = kiter->second;
429 bufferlist::iterator eiter = entrybl.begin();
431 ::decode(entry, eiter);
432 } catch (buffer::error& err) {
433 CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode entry, key=%s\n", kiter->first.c_str());
439 decode_list_index_key(kiter->first, &key, &ver);
441 start_key = kiter->first;
442 CLS_LOG(20, "start_key=%s len=%zu", start_key.c_str(), start_key.size());
444 if (!entry.is_valid()) {
445 CLS_LOG(20, "entry %s[%s] is not valid\n", key.name.c_str(), key.instance.c_str());
449 // filter out noncurrent versions, delete markers, and initial marker
450 if (!op.list_versions && (!entry.is_visible() || op.start_obj.name == key.name)) {
451 CLS_LOG(20, "entry %s[%s] is not visible\n", key.name.c_str(), key.instance.c_str());
454 if (m.size() < op.num_entries) {
455 m[kiter->first] = entry;
459 CLS_LOG(20, "got entry %s[%s] m.size()=%d\n", key.name.c_str(), key.instance.c_str(), (int)m.size());
461 } while (left_to_read > 0 && !done);
463 ret.is_truncated = more && !done;
469 static int check_index(cls_method_context_t hctx, struct rgw_bucket_dir_header *existing_header, struct rgw_bucket_dir_header *calc_header)
471 int rc = read_bucket_header(hctx, existing_header);
473 CLS_LOG(1, "ERROR: check_index(): failed to read header\n");
477 calc_header->tag_timeout = existing_header->tag_timeout;
478 calc_header->ver = existing_header->ver;
480 map<string, bufferlist> keys;
482 string filter_prefix;
484 #define CHECK_CHUNK_SIZE 1000
489 rc = get_obj_vals(hctx, start_obj, filter_prefix, CHECK_CHUNK_SIZE, &keys, &more);
493 std::map<string, bufferlist>::iterator kiter = keys.begin();
494 for (; kiter != keys.end(); ++kiter) {
495 if (!bi_is_objs_index(kiter->first)) {
500 struct rgw_bucket_dir_entry entry;
501 bufferlist::iterator eiter = kiter->second.begin();
503 ::decode(entry, eiter);
504 } catch (buffer::error& err) {
505 CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode entry, key=%s\n", kiter->first.c_str());
508 struct rgw_bucket_category_stats& stats = calc_header->stats[entry.meta.category];
510 stats.total_size += entry.meta.accounted_size;
511 stats.total_size_rounded += cls_rgw_get_rounded_size(entry.meta.accounted_size);
512 stats.actual_size += entry.meta.size;
514 start_obj = kiter->first;
516 } while (keys.size() == CHECK_CHUNK_SIZE && !done);
521 int rgw_bucket_check_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
523 struct rgw_cls_check_index_ret ret;
525 int rc = check_index(hctx, &ret.existing_header, &ret.calculated_header);
534 static int write_bucket_header(cls_method_context_t hctx, struct rgw_bucket_dir_header *header)
538 bufferlist header_bl;
539 ::encode(*header, header_bl);
540 return cls_cxx_map_write_header(hctx, &header_bl);
544 int rgw_bucket_rebuild_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
546 struct rgw_bucket_dir_header existing_header;
547 struct rgw_bucket_dir_header calc_header;
548 int rc = check_index(hctx, &existing_header, &calc_header);
552 return write_bucket_header(hctx, &calc_header);
555 int rgw_bucket_update_stats(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
558 rgw_cls_bucket_update_stats_op op;
559 auto iter = in->begin();
562 } catch (buffer::error& err) {
563 CLS_LOG(1, "ERROR: %s(): failed to decode request\n", __func__);
567 struct rgw_bucket_dir_header header;
568 int rc = read_bucket_header(hctx, &header);
570 CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
574 for (auto& s : op.stats) {
575 auto& dest = header.stats[s.first];
579 dest.total_size += s.second.total_size;
580 dest.total_size_rounded += s.second.total_size_rounded;
581 dest.num_entries += s.second.num_entries;
585 return write_bucket_header(hctx, &header);
588 int rgw_bucket_init_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
590 bufferlist::iterator iter;
592 bufferlist header_bl;
593 int rc = cls_cxx_map_read_header(hctx, &header_bl);
604 if (header_bl.length() != 0) {
605 CLS_LOG(1, "ERROR: index already initialized\n");
611 return write_bucket_header(hctx, &dir.header);
614 int rgw_bucket_set_tag_timeout(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
617 rgw_cls_tag_timeout_op op;
618 bufferlist::iterator iter = in->begin();
621 } catch (buffer::error& err) {
622 CLS_LOG(1, "ERROR: rgw_bucket_set_tag_timeout(): failed to decode request\n");
626 struct rgw_bucket_dir_header header;
627 int rc = read_bucket_header(hctx, &header);
629 CLS_LOG(1, "ERROR: rgw_bucket_set_tag_timeout(): failed to read header\n");
633 header.tag_timeout = op.tag_timeout;
635 return write_bucket_header(hctx, &header);
638 static int read_key_entry(cls_method_context_t hctx, cls_rgw_obj_key& key, string *idx, struct rgw_bucket_dir_entry *entry,
639 bool special_delete_marker_name = false);
641 int rgw_bucket_prepare_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
644 rgw_cls_obj_prepare_op op;
645 bufferlist::iterator iter = in->begin();
648 } catch (buffer::error& err) {
649 CLS_LOG(1, "ERROR: rgw_bucket_prepare_op(): failed to decode request\n");
653 if (op.tag.empty()) {
654 CLS_LOG(1, "ERROR: tag is empty\n");
658 CLS_LOG(1, "rgw_bucket_prepare_op(): request: op=%d name=%s instance=%s tag=%s\n",
659 op.op, op.key.name.c_str(), op.key.instance.c_str(), op.tag.c_str());
664 struct rgw_bucket_dir_entry entry;
665 int rc = read_key_entry(hctx, op.key, &idx, &entry);
666 if (rc < 0 && rc != -ENOENT)
669 bool noent = (rc == -ENOENT);
673 if (noent) { // no entry, initialize fields
675 entry.ver = rgw_bucket_entry_ver();
676 entry.exists = false;
677 entry.locator = op.locator;
680 // fill in proper state
681 struct rgw_bucket_pending_info info;
682 info.timestamp = real_clock::now();
683 info.state = CLS_RGW_STATE_PENDING_MODIFY;
685 entry.pending_map.insert(pair<string, rgw_bucket_pending_info>(op.tag, info));
687 struct rgw_bucket_dir_header header;
688 rc = read_bucket_header(hctx, &header);
690 CLS_LOG(1, "ERROR: rgw_bucket_prepare_op(): failed to read header\n");
694 if (op.log_op && !header.syncstopped) {
695 rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime,
696 entry.ver, info.state, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
701 // write out new key to disk
703 ::encode(entry, info_bl);
704 rc = cls_cxx_map_set_val(hctx, idx, &info_bl);
708 return write_bucket_header(hctx, &header);
711 static void unaccount_entry(struct rgw_bucket_dir_header& header, struct rgw_bucket_dir_entry& entry)
713 struct rgw_bucket_category_stats& stats = header.stats[entry.meta.category];
715 stats.total_size -= entry.meta.accounted_size;
716 stats.total_size_rounded -= cls_rgw_get_rounded_size(entry.meta.accounted_size);
717 stats.actual_size -= entry.meta.size;
720 static void log_entry(const char *func, const char *str, struct rgw_bucket_dir_entry *entry)
722 CLS_LOG(1, "%s(): %s: ver=%ld:%llu name=%s instance=%s locator=%s\n", func, str,
723 (long)entry->ver.pool, (unsigned long long)entry->ver.epoch,
724 entry->key.name.c_str(), entry->key.instance.c_str(), entry->locator.c_str());
727 static void log_entry(const char *func, const char *str, struct rgw_bucket_olh_entry *entry)
729 CLS_LOG(1, "%s(): %s: epoch=%llu name=%s instance=%s tag=%s\n", func, str,
730 (unsigned long long)entry->epoch, entry->key.name.c_str(), entry->key.instance.c_str(),
735 static int read_index_entry(cls_method_context_t hctx, string& name, T *entry)
737 bufferlist current_entry;
738 int rc = cls_cxx_map_get_val(hctx, name, ¤t_entry);
743 bufferlist::iterator cur_iter = current_entry.begin();
745 ::decode(*entry, cur_iter);
746 } catch (buffer::error& err) {
747 CLS_LOG(1, "ERROR: read_index_entry(): failed to decode entry\n");
751 log_entry(__func__, "existing entry", entry);
755 static int read_key_entry(cls_method_context_t hctx, cls_rgw_obj_key& key, string *idx, struct rgw_bucket_dir_entry *entry,
756 bool special_delete_marker_name)
758 encode_obj_index_key(key, idx);
759 int rc = read_index_entry(hctx, *idx, entry);
764 if (key.instance.empty() &&
765 entry->flags & RGW_BUCKET_DIRENT_FLAG_VER_MARKER) {
766 /* we only do it where key.instance is empty. In this case the delete marker will have a
767 * separate entry in the index to avoid collisions with the actual object, as it's mutable
769 if (special_delete_marker_name) {
770 encode_obj_versioned_data_key(key, idx, true);
771 rc = read_index_entry(hctx, *idx, entry);
776 encode_obj_versioned_data_key(key, idx);
777 rc = read_index_entry(hctx, *idx, entry);
779 *entry = rgw_bucket_dir_entry(); /* need to reset entry because we initialized it earlier */
787 int rgw_bucket_complete_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
790 rgw_cls_obj_complete_op op;
791 bufferlist::iterator iter = in->begin();
794 } catch (buffer::error& err) {
795 CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to decode request\n");
798 CLS_LOG(1, "rgw_bucket_complete_op(): request: op=%d name=%s instance=%s ver=%lu:%llu tag=%s\n",
799 op.op, op.key.name.c_str(), op.key.instance.c_str(),
800 (unsigned long)op.ver.pool, (unsigned long long)op.ver.epoch,
803 struct rgw_bucket_dir_header header;
804 int rc = read_bucket_header(hctx, &header);
806 CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n");
810 struct rgw_bucket_dir_entry entry;
814 rc = read_key_entry(hctx, op.key, &idx, &entry);
818 entry.meta = op.meta;
819 entry.locator = op.locator;
825 entry.index_ver = header.ver;
826 entry.flags = (entry.key.instance.empty() ? 0 : RGW_BUCKET_DIRENT_FLAG_VER); /* resetting entry flags, entry might have been previously a delete marker */
829 map<string, struct rgw_bucket_pending_info>::iterator pinter = entry.pending_map.find(op.tag);
830 if (pinter == entry.pending_map.end()) {
831 CLS_LOG(1, "ERROR: couldn't find tag for pending operation\n");
834 entry.pending_map.erase(pinter);
838 bufferlist update_bl;
840 if (op.tag.size() && op.op == CLS_RGW_OP_CANCEL) {
841 CLS_LOG(1, "rgw_bucket_complete_op(): cancel requested\n");
843 } else if (op.ver.pool == entry.ver.pool &&
844 op.ver.epoch && op.ver.epoch <= entry.ver.epoch) {
845 CLS_LOG(1, "rgw_bucket_complete_op(): skipping request, old epoch\n");
851 if (op.log_op && !header.syncstopped) {
852 rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, entry.ver,
853 CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
859 bufferlist new_key_bl;
860 ::encode(entry, new_key_bl);
861 return cls_cxx_map_set_val(hctx, idx, &new_key_bl);
868 unaccount_entry(header, entry);
872 switch ((int)op.op) {
874 entry.meta = op.meta;
876 if (!entry.pending_map.size()) {
877 int ret = cls_cxx_map_remove_key(hctx, idx);
881 entry.exists = false;
882 bufferlist new_key_bl;
883 ::encode(entry, new_key_bl);
884 int ret = cls_cxx_map_set_val(hctx, idx, &new_key_bl);
894 struct rgw_bucket_dir_entry_meta& meta = op.meta;
895 struct rgw_bucket_category_stats& stats = header.stats[meta.category];
901 stats.total_size += meta.accounted_size;
902 stats.total_size_rounded += cls_rgw_get_rounded_size(meta.accounted_size);
903 stats.actual_size += meta.size;
904 bufferlist new_key_bl;
905 ::encode(entry, new_key_bl);
906 int ret = cls_cxx_map_set_val(hctx, idx, &new_key_bl);
913 if (op.log_op && !header.syncstopped) {
914 rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, entry.ver,
915 CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
920 list<cls_rgw_obj_key>::iterator remove_iter;
921 CLS_LOG(20, "rgw_bucket_complete_op(): remove_objs.size()=%d\n", (int)op.remove_objs.size());
922 for (remove_iter = op.remove_objs.begin(); remove_iter != op.remove_objs.end(); ++remove_iter) {
923 cls_rgw_obj_key& remove_key = *remove_iter;
924 CLS_LOG(1, "rgw_bucket_complete_op(): removing entries, read_index_entry name=%s instance=%s\n",
925 remove_key.name.c_str(), remove_key.instance.c_str());
926 struct rgw_bucket_dir_entry remove_entry;
928 int ret = read_key_entry(hctx, remove_key, &k, &remove_entry);
930 CLS_LOG(1, "rgw_bucket_complete_op(): removing entries, read_index_entry name=%s instance=%s ret=%d\n",
931 remove_key.name.c_str(), remove_key.instance.c_str(), ret);
934 CLS_LOG(0, "rgw_bucket_complete_op(): entry.name=%s entry.instance=%s entry.meta.category=%d\n",
935 remove_entry.key.name.c_str(), remove_entry.key.instance.c_str(), remove_entry.meta.category);
936 unaccount_entry(header, remove_entry);
938 if (op.log_op && !header.syncstopped) {
939 ++header.ver; // increment index version, or we'll overwrite keys previously written
940 rc = log_index_operation(hctx, remove_key, CLS_RGW_OP_DEL, op.tag, remove_entry.meta.mtime,
941 remove_entry.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace);
946 ret = cls_cxx_map_remove_key(hctx, k);
948 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);
953 return write_bucket_header(hctx, &header);
957 static int write_entry(cls_method_context_t hctx, T& entry, const string& key)
961 return cls_cxx_map_set_val(hctx, key, &bl);
964 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)
966 cls_rgw_obj_key olh_key;
967 olh_key.name = obj_key.name;
969 encode_olh_data_key(olh_key, index_key);
970 int ret = read_index_entry(hctx, *index_key, olh_data_entry);
971 if (ret < 0 && ret != -ENOENT) {
972 CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_key.name.c_str(), ret);
976 *found = (ret != -ENOENT);
981 static void update_olh_log(struct rgw_bucket_olh_entry& olh_data_entry, OLHLogOp op, const string& op_tag,
982 cls_rgw_obj_key& key, bool delete_marker, uint64_t epoch)
984 vector<rgw_bucket_olh_log_entry>& log = olh_data_entry.pending_log[olh_data_entry.epoch];
985 rgw_bucket_olh_log_entry log_entry;
986 log_entry.epoch = epoch;
988 log_entry.op_tag = op_tag;
990 log_entry.delete_marker = delete_marker;
991 log.push_back(log_entry);
994 static string escape_str(const string& s)
996 int len = escape_json_attr_len(s.c_str(), s.size());
998 escape_json_attr(s.c_str(), s.size(), escaped);
999 return string(escaped);
1002 static int write_obj_instance_entry(cls_method_context_t hctx, struct rgw_bucket_dir_entry& instance_entry, const string& instance_idx)
1004 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);
1005 /* write the instance entry */
1006 int ret = write_entry(hctx, instance_entry, instance_idx);
1008 CLS_LOG(0, "ERROR: write_entry() instance_key=%s ret=%d", escape_str(instance_idx).c_str(), ret);
1015 * write object instance entry, and if needed also the list entry
1017 static int write_obj_entries(cls_method_context_t hctx, struct rgw_bucket_dir_entry& instance_entry, const string& instance_idx)
1019 int ret = write_obj_instance_entry(hctx, instance_entry, instance_idx);
1023 string instance_list_idx;
1024 get_list_index_key(instance_entry, &instance_list_idx);
1026 if (instance_idx != instance_list_idx) {
1027 CLS_LOG(20, "write_entry() idx=%s flags=%d", escape_str(instance_list_idx).c_str(), instance_entry.flags);
1028 /* write a new list entry for the object instance */
1029 ret = write_entry(hctx, instance_entry, instance_list_idx);
1031 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);
1039 class BIVerObjEntry {
1040 cls_method_context_t hctx;
1041 cls_rgw_obj_key key;
1042 string instance_idx;
1044 struct rgw_bucket_dir_entry instance_entry;
1049 BIVerObjEntry(cls_method_context_t& _hctx, const cls_rgw_obj_key& _key) : hctx(_hctx), key(_key), initialized(false) {
1052 int init(bool check_delete_marker = true) {
1053 int ret = read_key_entry(hctx, key, &instance_idx, &instance_entry,
1054 check_delete_marker && key.instance.empty()); /* this is potentially a delete marker, for null objects we
1055 keep separate instance entry for the delete markers */
1058 CLS_LOG(0, "ERROR: read_key_entry() idx=%s ret=%d", instance_idx.c_str(), ret);
1062 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);
1066 rgw_bucket_dir_entry& get_dir_entry() {
1067 return instance_entry;
1070 void init_as_delete_marker(rgw_bucket_dir_entry_meta& meta) {
1071 /* a deletion marker, need to initialize it, there's no instance entry for it yet */
1072 instance_entry.key = key;
1073 instance_entry.flags = RGW_BUCKET_DIRENT_FLAG_DELETE_MARKER;
1074 instance_entry.meta = meta;
1075 instance_entry.tag = "delete-marker";
1080 void set_epoch(uint64_t epoch) {
1081 instance_entry.versioned_epoch = epoch;
1084 int unlink_list_entry() {
1086 /* this instance has a previous list entry, remove that entry */
1087 get_list_index_key(instance_entry, &list_idx);
1088 CLS_LOG(20, "unlink_list_entry() list_idx=%s", escape_str(list_idx).c_str());
1089 int ret = cls_cxx_map_remove_key(hctx, list_idx);
1091 CLS_LOG(0, "ERROR: cls_cxx_map_remove_key() list_idx=%s ret=%d", list_idx.c_str(), ret);
1098 /* remove the instance entry */
1099 CLS_LOG(20, "unlink() idx=%s", escape_str(instance_idx).c_str());
1100 int ret = cls_cxx_map_remove_key(hctx, instance_idx);
1102 CLS_LOG(0, "ERROR: cls_cxx_map_remove_key() instance_idx=%s ret=%d", instance_idx.c_str(), ret);
1108 int write_entries(uint64_t flags_set, uint64_t flags_reset) {
1115 instance_entry.flags &= ~flags_reset;
1116 instance_entry.flags |= flags_set;
1118 /* write the instance and list entries */
1119 bool special_delete_marker_key = (instance_entry.is_delete_marker() && instance_entry.key.instance.empty());
1120 encode_obj_versioned_data_key(key, &instance_idx, special_delete_marker_key);
1121 int ret = write_obj_entries(hctx, instance_entry, instance_idx);
1123 CLS_LOG(0, "ERROR: write_obj_entries() instance_idx=%s ret=%d", instance_idx.c_str(), ret);
1130 int write(uint64_t epoch, bool current) {
1131 if (instance_entry.versioned_epoch > 0) {
1132 CLS_LOG(20, "%s(): instance_entry.versioned_epoch=%d epoch=%d", __func__, (int)instance_entry.versioned_epoch, (int)epoch);
1133 /* this instance has a previous list entry, remove that entry */
1134 int ret = unlink_list_entry();
1140 uint64_t flags = RGW_BUCKET_DIRENT_FLAG_VER;
1142 flags |= RGW_BUCKET_DIRENT_FLAG_CURRENT;
1145 instance_entry.versioned_epoch = epoch;
1146 return write_entries(flags, 0);
1149 int demote_current() {
1150 return write_entries(0, RGW_BUCKET_DIRENT_FLAG_CURRENT);
1153 bool is_delete_marker() {
1154 return instance_entry.is_delete_marker();
1157 int find_next_key(cls_rgw_obj_key *next_key, bool *found) {
1159 /* this instance has a previous list entry, remove that entry */
1160 get_list_index_key(instance_entry, &list_idx);
1161 /* this is the current head, need to update! */
1162 map<string, bufferlist> keys;
1164 string filter = key.name; /* list key starts with key name, filter it to avoid a case where we cross to
1165 different namespace */
1166 int ret = cls_cxx_map_get_vals(hctx, list_idx, filter, 1, &keys, &more);
1171 if (keys.size() < 1) {
1176 rgw_bucket_dir_entry next_entry;
1178 map<string, bufferlist>::reverse_iterator last = keys.rbegin();
1180 bufferlist::iterator iter = last->second.begin();
1181 ::decode(next_entry, iter);
1182 } catch (buffer::error& err) {
1183 CLS_LOG(0, "ERROR; failed to decode entry: %s", last->first.c_str());
1187 *found = (key.name == next_entry.key.name);
1189 *next_key = next_entry.key;
1196 return instance_entry.meta.mtime;
1202 cls_method_context_t hctx;
1203 cls_rgw_obj_key key;
1205 string olh_data_idx;
1206 struct rgw_bucket_olh_entry olh_data_entry;
1210 BIOLHEntry(cls_method_context_t& _hctx, const cls_rgw_obj_key& _key) : hctx(_hctx), key(_key), initialized(false) { }
1212 int init(bool *exists) {
1214 int ret = read_olh(hctx, key, &olh_data_entry, &olh_data_idx, exists);
1223 bool apply_epoch(uint64_t candidate_epoch) {
1224 if (candidate_epoch < olh_data_entry.epoch) {
1228 olh_data_entry.epoch = candidate_epoch;
1232 bool start_modify(uint64_t candidate_epoch) {
1233 if (candidate_epoch) {
1234 if (candidate_epoch < olh_data_entry.epoch) {
1235 return false; /* olh cannot be modified, old epoch */
1237 olh_data_entry.epoch = candidate_epoch;
1239 if (olh_data_entry.epoch == 0) {
1240 olh_data_entry.epoch = 2; /* versioned epoch should start with 2, 1 is reserved to converted plain entries */
1242 olh_data_entry.epoch++;
1248 uint64_t get_epoch() {
1249 return olh_data_entry.epoch;
1252 rgw_bucket_olh_entry& get_entry() {
1253 return olh_data_entry;
1256 void update(cls_rgw_obj_key& key, bool delete_marker) {
1257 olh_data_entry.delete_marker = delete_marker;
1258 olh_data_entry.key = key;
1262 /* write the olh data entry */
1263 int ret = write_entry(hctx, olh_data_entry, olh_data_idx);
1265 CLS_LOG(0, "ERROR: write_entry() olh_key=%s ret=%d", olh_data_idx.c_str(), ret);
1272 void update_log(OLHLogOp op, const string& op_tag, cls_rgw_obj_key& key, bool delete_marker, uint64_t epoch = 0) {
1274 epoch = olh_data_entry.epoch;
1276 update_olh_log(olh_data_entry, op, op_tag, key, delete_marker, epoch);
1279 bool exists() { return olh_data_entry.exists; }
1281 void set_exists(bool exists) {
1282 olh_data_entry.exists = exists;
1285 bool pending_removal() { return olh_data_entry.pending_removal; }
1287 void set_pending_removal(bool pending_removal) {
1288 olh_data_entry.pending_removal = pending_removal;
1291 const string& get_tag() { return olh_data_entry.tag; }
1292 void set_tag(const string& tag) {
1293 olh_data_entry.tag = tag;
1297 static int write_version_marker(cls_method_context_t hctx, cls_rgw_obj_key& key)
1299 struct rgw_bucket_dir_entry entry;
1301 entry.flags = RGW_BUCKET_DIRENT_FLAG_VER_MARKER;
1302 int ret = write_entry(hctx, entry, key.name);
1304 CLS_LOG(0, "ERROR: write_entry returned ret=%d", ret);
1311 * plain entries are the ones who were created when bucket was not versioned,
1312 * if we override these objects, we need to convert these to versioned entries -- ones that have
1313 * both data entry, and listing key. Their version is going to be empty though
1315 static int convert_plain_entry_to_versioned(cls_method_context_t hctx, cls_rgw_obj_key& key, bool demote_current, bool instance_only)
1317 if (!key.instance.empty()) {
1321 struct rgw_bucket_dir_entry entry;
1324 int ret = read_key_entry(hctx, key, &orig_idx, &entry);
1325 if (ret != -ENOENT) {
1327 CLS_LOG(0, "ERROR: read_key_entry() returned ret=%d", ret);
1331 entry.versioned_epoch = 1; /* converted entries are always 1 */
1332 entry.flags |= RGW_BUCKET_DIRENT_FLAG_VER;
1334 if (demote_current) {
1335 entry.flags &= ~RGW_BUCKET_DIRENT_FLAG_CURRENT;
1339 encode_obj_versioned_data_key(key, &new_idx);
1341 if (instance_only) {
1342 ret = write_obj_instance_entry(hctx, entry, new_idx);
1344 ret = write_obj_entries(hctx, entry, new_idx);
1347 CLS_LOG(0, "ERROR: write_obj_entries new_idx=%s returned %d", new_idx.c_str(), ret);
1352 ret = write_version_marker(hctx, key);
1361 * link an object version to an olh, update the relevant index entries. It will also handle the
1362 * deletion marker case. We have a few entries that we need to take care of. For object 'foo',
1363 * instance BAR, we'd update the following (not actual encoding):
1364 * - olh data: [BI_BUCKET_OLH_DATA_INDEX]foo
1365 * - object instance data: [BI_BUCKET_OBJ_INSTANCE_INDEX]foo,BAR
1366 * - object instance list entry: foo,123,BAR
1368 * The instance list entry needs to be ordered by newer to older, so we generate an appropriate
1369 * number string that follows the name.
1370 * The top instance for each object is marked appropriately.
1371 * We generate instance entry for deletion markers here, as they are not created prior.
1373 static int rgw_bucket_link_olh(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1375 string olh_data_idx;
1376 string instance_idx;
1379 rgw_cls_link_olh_op op;
1380 bufferlist::iterator iter = in->begin();
1383 } catch (buffer::error& err) {
1384 CLS_LOG(0, "ERROR: rgw_bucket_link_olh_op(): failed to decode request\n");
1388 BIVerObjEntry obj(hctx, op.key);
1389 BIOLHEntry olh(hctx, op.key);
1391 /* read instance entry */
1392 int ret = obj.init(op.delete_marker);
1393 bool existed = (ret == 0);
1394 if (ret == -ENOENT && op.delete_marker) {
1401 if (existed && !real_clock::is_zero(op.unmod_since)) {
1402 struct timespec mtime = ceph::real_clock::to_timespec(obj.mtime());
1403 struct timespec unmod = ceph::real_clock::to_timespec(op.unmod_since);
1404 if (!op.high_precision_time) {
1408 if (mtime >= unmod) {
1409 return 0; /* no need to set error, we just return 0 and avoid writing to the bi log */
1416 * Special handling for null instance object / delete-marker. For these objects we're going to
1417 * have separate instances for a data object vs. delete-marker to avoid collisions. We now check
1418 * if we got to overwrite a previous entry, and in that case we'll remove its list entry.
1420 if (op.key.instance.empty()) {
1421 BIVerObjEntry other_obj(hctx, op.key);
1422 ret = other_obj.init(!op.delete_marker); /* try reading the other null versioned entry */
1423 existed = (ret >= 0 && !other_obj.is_delete_marker());
1424 if (ret >= 0 && other_obj.is_delete_marker() != op.delete_marker) {
1425 ret = other_obj.unlink_list_entry();
1429 ret = other_obj.unlink();
1435 removing = existed && op.delete_marker;
1437 removing = (existed && !obj.is_delete_marker() && op.delete_marker);
1440 if (op.delete_marker) {
1441 /* a deletion marker, need to initialize entry as such */
1442 obj.init_as_delete_marker(op.meta);
1447 ret = olh.init(&olh_found);
1452 if (!olh.start_modify(op.olh_epoch)) {
1453 ret = obj.write(op.olh_epoch, false);
1458 olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false, op.olh_epoch);
1464 const string& olh_tag = olh.get_tag();
1465 if (op.olh_tag != olh_tag) {
1466 if (!olh.pending_removal()) {
1467 CLS_LOG(5, "NOTICE: op.olh_tag (%s) != olh.tag (%s)", op.olh_tag.c_str(), olh_tag.c_str());
1470 /* if pending removal, this is a new olh instance */
1471 olh.set_tag(op.olh_tag);
1474 rgw_bucket_olh_entry& olh_entry = olh.get_entry();
1475 /* found olh, previous instance is no longer the latest, need to update */
1476 if (!(olh_entry.key == op.key)) {
1477 BIVerObjEntry old_obj(hctx, olh_entry.key);
1479 ret = old_obj.demote_current();
1481 CLS_LOG(0, "ERROR: could not demote current on previous key ret=%d", ret);
1486 olh.set_pending_removal(false);
1488 bool instance_only = (op.key.instance.empty() && op.delete_marker);
1489 cls_rgw_obj_key key(op.key.name);
1490 ret = convert_plain_entry_to_versioned(hctx, key, true, instance_only);
1492 CLS_LOG(0, "ERROR: convert_plain_entry_to_versioned ret=%d", ret);
1495 olh.set_tag(op.olh_tag);
1498 /* update the olh log */
1499 olh.update_log(CLS_RGW_OLH_OP_LINK_OLH, op.op_tag, op.key, op.delete_marker);
1501 olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false);
1504 olh.update(op.key, op.delete_marker);
1506 olh.set_exists(true);
1510 CLS_LOG(0, "ERROR: failed to update olh ret=%d", ret);
1514 /* write the instance and list entries */
1515 ret = obj.write(olh.get_epoch(), true);
1520 struct rgw_bucket_dir_header header;
1521 ret = read_bucket_header(hctx, &header);
1523 CLS_LOG(1, "ERROR: rgw_bucket_unlink_instance(): failed to read header\n");
1527 if (op.log_op && !header.syncstopped) {
1528 rgw_bucket_dir_entry& entry = obj.get_dir_entry();
1530 rgw_bucket_entry_ver ver;
1531 ver.epoch = (op.olh_epoch ? op.olh_epoch : olh.get_epoch());
1533 string *powner = NULL;
1534 string *powner_display_name = NULL;
1536 if (op.delete_marker) {
1537 powner = &entry.meta.owner;
1538 powner_display_name = &entry.meta.owner_display_name;
1541 RGWModifyOp operation = (op.delete_marker ? CLS_RGW_OP_LINK_OLH_DM : CLS_RGW_OP_LINK_OLH);
1542 ret = log_index_operation(hctx, op.key, operation, op.op_tag,
1543 entry.meta.mtime, ver,
1544 CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags | RGW_BILOG_FLAG_VERSIONED_OP,
1545 powner, powner_display_name, &op.zones_trace);
1550 return write_bucket_header(hctx, &header); /* updates header version */
1553 static int rgw_bucket_unlink_instance(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1555 string olh_data_idx;
1556 string instance_idx;
1559 rgw_cls_unlink_instance_op op;
1560 bufferlist::iterator iter = in->begin();
1563 } catch (buffer::error& err) {
1564 CLS_LOG(0, "ERROR: rgw_bucket_rm_obj_instance_op(): failed to decode request\n");
1568 cls_rgw_obj_key dest_key = op.key;
1569 if (dest_key.instance == "null") {
1570 dest_key.instance.clear();
1573 BIVerObjEntry obj(hctx, dest_key);
1574 BIOLHEntry olh(hctx, dest_key);
1576 int ret = obj.init();
1577 if (ret == -ENOENT) {
1578 return 0; /* already removed */
1581 CLS_LOG(0, "ERROR: obj.init() returned ret=%d", ret);
1586 ret = olh.init(&olh_found);
1588 CLS_LOG(0, "ERROR: olh.init() returned ret=%d", ret);
1593 bool instance_only = false;
1594 cls_rgw_obj_key key(dest_key.name);
1595 ret = convert_plain_entry_to_versioned(hctx, key, true, instance_only);
1597 CLS_LOG(0, "ERROR: convert_plain_entry_to_versioned ret=%d", ret);
1600 olh.update(dest_key, false);
1601 olh.set_tag(op.olh_tag);
1606 if (!olh.start_modify(op.olh_epoch)) {
1607 ret = obj.unlink_list_entry();
1612 if (!obj.is_delete_marker()) {
1613 olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false, op.olh_epoch);
1619 rgw_bucket_olh_entry& olh_entry = olh.get_entry();
1620 cls_rgw_obj_key& olh_key = olh_entry.key;
1621 CLS_LOG(20, "%s(): updating olh log: existing olh entry: %s[%s] (delete_marker=%d)", __func__,
1622 olh_key.name.c_str(), olh_key.instance.c_str(), olh_entry.delete_marker);
1624 if (olh_key == dest_key) {
1625 /* this is the current head, need to update! */
1626 cls_rgw_obj_key next_key;
1628 ret = obj.find_next_key(&next_key, &found);
1630 CLS_LOG(0, "ERROR: obj.find_next_key() returned ret=%d", ret);
1635 BIVerObjEntry next(hctx, next_key);
1636 ret = next.write(olh.get_epoch(), true);
1638 CLS_LOG(0, "ERROR: next.write() returned ret=%d", ret);
1642 CLS_LOG(20, "%s(): updating olh log: link olh -> %s[%s] (is_delete=%d)", __func__,
1643 next_key.name.c_str(), next_key.instance.c_str(), (int)next.is_delete_marker());
1645 olh.update(next_key, next.is_delete_marker());
1646 olh.update_log(CLS_RGW_OLH_OP_LINK_OLH, op.op_tag, next_key, next.is_delete_marker());
1648 /* next_key is empty */
1649 olh.update(next_key, false);
1650 olh.update_log(CLS_RGW_OLH_OP_UNLINK_OLH, op.op_tag, next_key, false);
1651 olh.set_exists(false);
1652 olh.set_pending_removal(true);
1656 if (!obj.is_delete_marker()) {
1657 olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false);
1659 /* this is a delete marker, it's our responsibility to remove its instance entry */
1666 ret = obj.unlink_list_entry();
1676 struct rgw_bucket_dir_header header;
1677 ret = read_bucket_header(hctx, &header);
1679 CLS_LOG(1, "ERROR: rgw_bucket_unlink_instance(): failed to read header\n");
1683 if (op.log_op && !header.syncstopped) {
1684 rgw_bucket_entry_ver ver;
1685 ver.epoch = (op.olh_epoch ? op.olh_epoch : olh.get_epoch());
1687 real_time mtime = real_clock::now(); /* mtime has no real meaning in instance removal context */
1688 ret = log_index_operation(hctx, op.key, CLS_RGW_OP_UNLINK_INSTANCE, op.op_tag,
1690 CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker,
1691 op.bilog_flags | RGW_BILOG_FLAG_VERSIONED_OP, NULL, NULL, &op.zones_trace);
1696 return write_bucket_header(hctx, &header); /* updates header version */
1699 static int rgw_bucket_read_olh_log(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1702 rgw_cls_read_olh_log_op op;
1703 bufferlist::iterator iter = in->begin();
1706 } catch (buffer::error& err) {
1707 CLS_LOG(0, "ERROR: rgw_bucket_read_olh_log(): failed to decode request\n");
1711 if (!op.olh.instance.empty()) {
1712 CLS_LOG(1, "bad key passed in (non empty instance)");
1716 struct rgw_bucket_olh_entry olh_data_entry;
1717 string olh_data_key;
1718 encode_olh_data_key(op.olh, &olh_data_key);
1719 int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry);
1720 if (ret < 0 && ret != -ENOENT) {
1721 CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1725 if (olh_data_entry.tag != op.olh_tag) {
1726 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());
1730 rgw_cls_read_olh_log_ret op_ret;
1732 #define MAX_OLH_LOG_ENTRIES 1000
1733 map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log = olh_data_entry.pending_log;
1735 if (log.begin()->first > op.ver_marker && log.size() <= MAX_OLH_LOG_ENTRIES) {
1737 op_ret.is_truncated = false;
1739 map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator iter = log.upper_bound(op.ver_marker);
1741 for (int i = 0; i < MAX_OLH_LOG_ENTRIES && iter != log.end(); ++i, ++iter) {
1742 op_ret.log[iter->first] = iter->second;
1744 op_ret.is_truncated = (iter != log.end());
1747 ::encode(op_ret, *out);
1752 static int rgw_bucket_trim_olh_log(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1755 rgw_cls_trim_olh_log_op op;
1756 bufferlist::iterator iter = in->begin();
1759 } catch (buffer::error& err) {
1760 CLS_LOG(0, "ERROR: rgw_bucket_trim_olh_log(): failed to decode request\n");
1764 if (!op.olh.instance.empty()) {
1765 CLS_LOG(1, "bad key passed in (non empty instance)");
1769 /* read olh entry */
1770 struct rgw_bucket_olh_entry olh_data_entry;
1771 string olh_data_key;
1772 encode_olh_data_key(op.olh, &olh_data_key);
1773 int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry);
1774 if (ret < 0 && ret != -ENOENT) {
1775 CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1779 if (olh_data_entry.tag != op.olh_tag) {
1780 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());
1784 /* remove all versions up to and including ver from the pending map */
1785 map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log = olh_data_entry.pending_log;
1786 map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator liter = log.begin();
1787 while (liter != log.end() && liter->first <= op.ver) {
1788 map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator rm_iter = liter;
1793 /* write the olh data entry */
1794 ret = write_entry(hctx, olh_data_entry, olh_data_key);
1796 CLS_LOG(0, "ERROR: write_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1803 static int rgw_bucket_clear_olh(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
1806 rgw_cls_bucket_clear_olh_op op;
1807 bufferlist::iterator iter = in->begin();
1810 } catch (buffer::error& err) {
1811 CLS_LOG(0, "ERROR: rgw_bucket_clear_olh(): failed to decode request\n");
1815 if (!op.key.instance.empty()) {
1816 CLS_LOG(1, "bad key passed in (non empty instance)");
1820 /* read olh entry */
1821 struct rgw_bucket_olh_entry olh_data_entry;
1822 string olh_data_key;
1823 encode_olh_data_key(op.key, &olh_data_key);
1824 int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry);
1825 if (ret < 0 && ret != -ENOENT) {
1826 CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret);
1830 if (olh_data_entry.tag != op.olh_tag) {
1831 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());
1835 ret = cls_cxx_map_remove_key(hctx, olh_data_key);
1837 CLS_LOG(1, "NOTICE: %s(): can't remove key %s ret=%d", __func__, olh_data_key.c_str(), ret);
1841 rgw_bucket_dir_entry plain_entry;
1843 /* read plain entry, make sure it's a versioned place holder */
1844 ret = read_index_entry(hctx, op.key.name, &plain_entry);
1845 if (ret == -ENOENT) {
1846 /* we're done, no entry existing */
1850 CLS_LOG(0, "ERROR: read_index_entry key=%s ret=%d", op.key.name.c_str(), ret);
1854 if ((plain_entry.flags & RGW_BUCKET_DIRENT_FLAG_VER_MARKER) == 0) {
1855 /* it's not a version marker, don't remove it */
1859 ret = cls_cxx_map_remove_key(hctx, op.key.name);
1861 CLS_LOG(1, "NOTICE: %s(): can't remove key %s ret=%d", __func__, op.key.name.c_str(), ret);
1868 int rgw_dir_suggest_changes(cls_method_context_t hctx,
1869 bufferlist *in, bufferlist *out)
1871 CLS_LOG(1, "rgw_dir_suggest_changes()");
1873 bufferlist header_bl;
1874 struct rgw_bucket_dir_header header;
1875 bool header_changed = false;
1877 int rc = read_bucket_header(hctx, &header);
1879 CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to read header\n");
1883 timespan tag_timeout(
1884 std::chrono::seconds(
1885 header.tag_timeout ? header.tag_timeout : CEPH_RGW_TAG_TIMEOUT));
1887 bufferlist::iterator in_iter = in->begin();
1889 while (!in_iter.end()) {
1891 rgw_bucket_dir_entry cur_change;
1892 rgw_bucket_dir_entry cur_disk;
1894 ::decode(op, in_iter);
1895 ::decode(cur_change, in_iter);
1896 } catch (buffer::error& err) {
1897 CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to decode request\n");
1901 bufferlist cur_disk_bl;
1902 string cur_change_key;
1903 encode_obj_index_key(cur_change.key, &cur_change_key);
1904 int ret = cls_cxx_map_get_val(hctx, cur_change_key, &cur_disk_bl);
1905 if (ret < 0 && ret != -ENOENT)
1908 if (cur_disk_bl.length()) {
1909 bufferlist::iterator cur_disk_iter = cur_disk_bl.begin();
1911 ::decode(cur_disk, cur_disk_iter);
1912 } catch (buffer::error& error) {
1913 CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to decode cur_disk\n");
1917 real_time cur_time = real_clock::now();
1918 map<string, struct rgw_bucket_pending_info>::iterator iter =
1919 cur_disk.pending_map.begin();
1920 while(iter != cur_disk.pending_map.end()) {
1921 map<string, struct rgw_bucket_pending_info>::iterator cur_iter=iter++;
1922 if (cur_time > (cur_iter->second.timestamp + timespan(tag_timeout))) {
1923 cur_disk.pending_map.erase(cur_iter);
1928 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",
1929 cur_disk.pending_map.empty(), (int)op, cur_disk.exists,
1930 (int)cur_change.pending_map.size(), cur_change.exists);
1932 if (cur_disk.pending_map.empty()) {
1933 if (cur_disk.exists) {
1934 struct rgw_bucket_category_stats& old_stats = header.stats[cur_disk.meta.category];
1935 CLS_LOG(10, "total_entries: %" PRId64 " -> %" PRId64 "\n", old_stats.num_entries, old_stats.num_entries - 1);
1936 old_stats.num_entries--;
1937 old_stats.total_size -= cur_disk.meta.accounted_size;
1938 old_stats.total_size_rounded -= cls_rgw_get_rounded_size(cur_disk.meta.accounted_size);
1939 old_stats.actual_size -= cur_disk.meta.size;
1940 header_changed = true;
1942 struct rgw_bucket_category_stats& stats =
1943 header.stats[cur_change.meta.category];
1944 bool log_op = (op & CEPH_RGW_DIR_SUGGEST_LOG_OP) != 0;
1945 op &= CEPH_RGW_DIR_SUGGEST_OP_MASK;
1947 case CEPH_RGW_REMOVE:
1948 CLS_LOG(10, "CEPH_RGW_REMOVE name=%s instance=%s\n", cur_change.key.name.c_str(), cur_change.key.instance.c_str());
1949 ret = cls_cxx_map_remove_key(hctx, cur_change_key);
1952 if (log_op && cur_disk.exists && !header.syncstopped) {
1953 ret = log_index_operation(hctx, cur_disk.key, CLS_RGW_OP_DEL, cur_disk.tag, cur_disk.meta.mtime,
1954 cur_disk.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, 0, NULL, NULL, NULL);
1956 CLS_LOG(0, "ERROR: %s(): failed to log operation ret=%d", __func__, ret);
1961 case CEPH_RGW_UPDATE:
1962 if (!cur_disk.exists) {
1963 // this update would only have been sent by the rgw client
1964 // if the rgw_bucket_dir_entry existed, however between that
1965 // check and now the entry has diappeared, so we were likely
1966 // in the midst of a delete op, and we will not recreate the
1969 "CEPH_RGW_UPDATE not applied because rgw_bucket_dir_entry"
1970 " no longer exists\n");
1974 CLS_LOG(10, "CEPH_RGW_UPDATE name=%s instance=%s total_entries: %" PRId64 " -> %" PRId64 "\n",
1975 cur_change.key.name.c_str(), cur_change.key.instance.c_str(), stats.num_entries, stats.num_entries + 1);
1977 stats.num_entries++;
1978 stats.total_size += cur_change.meta.accounted_size;
1979 stats.total_size_rounded += cls_rgw_get_rounded_size(cur_change.meta.accounted_size);
1980 stats.actual_size += cur_change.meta.size;
1981 header_changed = true;
1982 cur_change.index_ver = header.ver;
1983 bufferlist cur_state_bl;
1984 ::encode(cur_change, cur_state_bl);
1985 ret = cls_cxx_map_set_val(hctx, cur_change_key, &cur_state_bl);
1988 if (log_op && !header.syncstopped) {
1989 ret = log_index_operation(hctx, cur_change.key, CLS_RGW_OP_ADD, cur_change.tag, cur_change.meta.mtime,
1990 cur_change.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, 0, NULL, NULL, NULL);
1992 CLS_LOG(0, "ERROR: %s(): failed to log operation ret=%d", __func__, ret);
1998 } // if (cur_disk.pending_map.empty())
1999 } // while (!in_iter.end())
2001 if (header_changed) {
2002 return write_bucket_header(hctx, &header);
2007 static int rgw_obj_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2010 rgw_cls_obj_remove_op op;
2011 bufferlist::iterator iter = in->begin();
2014 } catch (buffer::error& err) {
2015 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2019 if (op.keep_attr_prefixes.empty()) {
2020 return cls_cxx_remove(hctx);
2023 map<string, bufferlist> attrset;
2024 int ret = cls_cxx_getxattrs(hctx, &attrset);
2025 if (ret < 0 && ret != -ENOENT) {
2026 CLS_LOG(0, "ERROR: %s(): cls_cxx_getxattrs() returned %d", __func__, ret);
2030 map<string, bufferlist> new_attrs;
2031 for (list<string>::iterator iter = op.keep_attr_prefixes.begin();
2032 iter != op.keep_attr_prefixes.end(); ++iter) {
2033 string& check_prefix = *iter;
2035 for (map<string, bufferlist>::iterator aiter = attrset.lower_bound(check_prefix);
2036 aiter != attrset.end(); ++aiter) {
2037 const string& attr = aiter->first;
2039 if (attr.substr(0, check_prefix.size()) > check_prefix) {
2043 new_attrs[attr] = aiter->second;
2047 CLS_LOG(20, "%s(): removing object", __func__);
2048 ret = cls_cxx_remove(hctx);
2050 CLS_LOG(0, "ERROR: %s(): cls_cxx_remove returned %d", __func__, ret);
2054 if (new_attrs.empty()) {
2055 /* no data to keep */
2059 ret = cls_cxx_create(hctx, false);
2061 CLS_LOG(0, "ERROR: %s(): cls_cxx_create returned %d", __func__, ret);
2065 for (map<string, bufferlist>::iterator aiter = new_attrs.begin();
2066 aiter != new_attrs.end(); ++aiter) {
2067 const string& attr = aiter->first;
2069 ret = cls_cxx_setxattr(hctx, attr.c_str(), &aiter->second);
2070 CLS_LOG(20, "%s(): setting attr: %s", __func__, attr.c_str());
2072 CLS_LOG(0, "ERROR: %s(): cls_cxx_setxattr (attr=%s) returned %d", __func__, attr.c_str(), ret);
2080 static int rgw_obj_store_pg_ver(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2083 rgw_cls_obj_store_pg_ver_op op;
2084 bufferlist::iterator iter = in->begin();
2087 } catch (buffer::error& err) {
2088 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2093 uint64_t ver = cls_current_version(hctx);
2095 int ret = cls_cxx_setxattr(hctx, op.attr.c_str(), &bl);
2097 CLS_LOG(0, "ERROR: %s(): cls_cxx_setxattr (attr=%s) returned %d", __func__, op.attr.c_str(), ret);
2104 static int rgw_obj_check_attrs_prefix(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2107 rgw_cls_obj_check_attrs_prefix op;
2108 bufferlist::iterator iter = in->begin();
2111 } catch (buffer::error& err) {
2112 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2116 if (op.check_prefix.empty()) {
2120 map<string, bufferlist> attrset;
2121 int ret = cls_cxx_getxattrs(hctx, &attrset);
2122 if (ret < 0 && ret != -ENOENT) {
2123 CLS_LOG(0, "ERROR: %s(): cls_cxx_getxattrs() returned %d", __func__, ret);
2129 for (map<string, bufferlist>::iterator aiter = attrset.lower_bound(op.check_prefix);
2130 aiter != attrset.end(); ++aiter) {
2131 const string& attr = aiter->first;
2133 if (attr.substr(0, op.check_prefix.size()) > op.check_prefix) {
2140 if (exist == op.fail_if_exist) {
2147 static int rgw_obj_check_mtime(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2150 rgw_cls_obj_check_mtime op;
2151 bufferlist::iterator iter = in->begin();
2154 } catch (buffer::error& err) {
2155 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2160 int ret = cls_cxx_stat2(hctx, NULL, &obj_ut);
2161 if (ret < 0 && ret != -ENOENT) {
2162 CLS_LOG(0, "ERROR: %s(): cls_cxx_stat() returned %d", __func__, ret);
2165 if (ret == -ENOENT) {
2166 CLS_LOG(10, "object does not exist, skipping check");
2169 ceph_timespec obj_ts = ceph::real_clock::to_ceph_timespec(obj_ut);
2170 ceph_timespec op_ts = ceph::real_clock::to_ceph_timespec(op.mtime);
2172 if (!op.high_precision_time) {
2177 CLS_LOG(10, "%s: obj_ut=%lld.%06lld op.mtime=%lld.%06lld", __func__,
2178 (long long)obj_ts.tv_sec, (long long)obj_ts.tv_nsec,
2179 (long long)op_ts.tv_sec, (long long)op_ts.tv_nsec);
2184 case CLS_RGW_CHECK_TIME_MTIME_EQ:
2185 check = (obj_ts == op_ts);
2187 case CLS_RGW_CHECK_TIME_MTIME_LT:
2188 check = (obj_ts < op_ts);
2190 case CLS_RGW_CHECK_TIME_MTIME_LE:
2191 check = (obj_ts <= op_ts);
2193 case CLS_RGW_CHECK_TIME_MTIME_GT:
2194 check = (obj_ts > op_ts);
2196 case CLS_RGW_CHECK_TIME_MTIME_GE:
2197 check = (obj_ts >= op_ts);
2210 static int rgw_bi_get_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2213 rgw_cls_bi_get_op op;
2214 bufferlist::iterator iter = in->begin();
2217 } catch (buffer::error& err) {
2218 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2229 encode_obj_index_key(op.key, &idx);
2232 encode_olh_data_key(op.key, &idx);
2235 CLS_LOG(10, "%s(): invalid key type encoding: %d", __func__, op.type);
2239 rgw_cls_bi_get_ret op_ret;
2241 rgw_cls_bi_entry& entry = op_ret.entry;
2243 entry.type = op.type;
2246 int r = cls_cxx_map_get_val(hctx, idx, &entry.data);
2248 CLS_LOG(10, "%s(): cls_cxx_map_get_val() returned %d", __func__, r);
2252 ::encode(op_ret, *out);
2257 static int rgw_bi_put_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2260 rgw_cls_bi_put_op op;
2261 bufferlist::iterator iter = in->begin();
2264 } catch (buffer::error& err) {
2265 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2269 rgw_cls_bi_entry& entry = op.entry;
2271 int r = cls_cxx_map_set_val(hctx, entry.idx, &entry.data);
2273 CLS_LOG(0, "ERROR: %s(): cls_cxx_map_set_val() returned r=%d", __func__, r);
2279 static int list_plain_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max,
2280 list<rgw_cls_bi_entry> *entries, bool *pmore)
2282 string filter = name;
2283 string start_key = marker;
2285 string end_key; // stop listing at bi_log_prefix
2286 bi_log_prefix(end_key);
2289 map<string, bufferlist> keys;
2290 int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max, &keys, pmore);
2295 map<string, bufferlist>::iterator iter;
2296 for (iter = keys.begin(); iter != keys.end(); ++iter) {
2297 if (iter->first >= end_key) {
2298 /* past the end of plain namespace */
2302 rgw_cls_bi_entry entry;
2303 entry.type = PlainIdx;
2304 entry.idx = iter->first;
2305 entry.data = iter->second;
2307 bufferlist::iterator biter = entry.data.begin();
2309 rgw_bucket_dir_entry e;
2312 } catch (buffer::error& err) {
2313 CLS_LOG(0, "ERROR: %s(): failed to decode buffer", __func__);
2317 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());
2319 if (!name.empty() && e.key.name != name) {
2323 entries->push_back(entry);
2325 if (count >= (int)max) {
2328 start_key = entry.idx;
2334 static int list_instance_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max,
2335 list<rgw_cls_bi_entry> *entries, bool *pmore)
2337 cls_rgw_obj_key key(name);
2338 string first_instance_idx;
2339 encode_obj_versioned_data_key(key, &first_instance_idx);
2342 if (!name.empty()) {
2343 start_key = first_instance_idx;
2345 start_key = BI_PREFIX_CHAR;
2346 start_key.append(bucket_index_prefixes[BI_BUCKET_OBJ_INSTANCE_INDEX]);
2348 string filter = start_key;
2349 if (bi_entry_gt(marker, start_key)) {
2353 map<string, bufferlist> keys;
2355 int ret = cls_cxx_map_get_val(hctx, start_key, &k);
2356 if (ret < 0 && ret != -ENOENT) {
2359 bool found_first = (ret == 0);
2364 ret = cls_cxx_map_get_vals(hctx, start_key, string(), max, &keys, pmore);
2365 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());
2371 keys[start_key].claim(k);
2374 map<string, bufferlist>::iterator iter;
2375 for (iter = keys.begin(); iter != keys.end(); ++iter) {
2376 rgw_cls_bi_entry entry;
2377 entry.type = InstanceIdx;
2378 entry.idx = iter->first;
2379 entry.data = iter->second;
2381 if (!filter.empty() && entry.idx.compare(0, filter.size(), filter) != 0) {
2385 CLS_LOG(20, "%s(): entry.idx=%s", __func__, escape_str(entry.idx).c_str());
2387 bufferlist::iterator biter = entry.data.begin();
2389 rgw_bucket_dir_entry e;
2392 } catch (buffer::error& err) {
2393 CLS_LOG(0, "ERROR: %s(): failed to decode buffer (size=%d)", __func__, entry.data.length());
2397 if (!name.empty() && e.key.name != name) {
2401 entries->push_back(entry);
2403 start_key = entry.idx;
2409 static int list_olh_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max,
2410 list<rgw_cls_bi_entry> *entries, bool *pmore)
2412 cls_rgw_obj_key key(name);
2413 string first_instance_idx;
2414 encode_olh_data_key(key, &first_instance_idx);
2417 if (!name.empty()) {
2418 start_key = first_instance_idx;
2420 start_key = BI_PREFIX_CHAR;
2421 start_key.append(bucket_index_prefixes[BI_BUCKET_OLH_DATA_INDEX]);
2423 string filter = start_key;
2424 if (bi_entry_gt(marker, start_key)) {
2428 map<string, bufferlist> keys;
2431 ret = cls_cxx_map_get_val(hctx, start_key, &k);
2432 if (ret < 0 && ret != -ENOENT) {
2435 bool found_first = (ret == 0);
2440 ret = cls_cxx_map_get_vals(hctx, start_key, string(), max, &keys, pmore);
2441 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());
2448 keys[start_key].claim(k);
2451 map<string, bufferlist>::iterator iter;
2452 for (iter = keys.begin(); iter != keys.end(); ++iter) {
2453 rgw_cls_bi_entry entry;
2454 entry.type = OLHIdx;
2455 entry.idx = iter->first;
2456 entry.data = iter->second;
2458 if (!filter.empty() && entry.idx.compare(0, filter.size(), filter) != 0) {
2462 CLS_LOG(20, "%s(): entry.idx=%s", __func__, escape_str(entry.idx).c_str());
2464 bufferlist::iterator biter = entry.data.begin();
2466 rgw_bucket_olh_entry e;
2469 } catch (buffer::error& err) {
2470 CLS_LOG(0, "ERROR: %s(): failed to decode buffer (size=%d)", __func__, entry.data.length());
2474 if (!name.empty() && e.key.name != name) {
2478 entries->push_back(entry);
2480 start_key = entry.idx;
2486 static int rgw_bi_list_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2489 rgw_cls_bi_list_op op;
2490 bufferlist::iterator iter = in->begin();
2493 } catch (buffer::error& err) {
2494 CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__);
2498 rgw_cls_bi_list_ret op_ret;
2500 string filter = op.name;
2501 #define MAX_BI_LIST_ENTRIES 1000
2502 int32_t max = (op.max < MAX_BI_LIST_ENTRIES ? op.max : MAX_BI_LIST_ENTRIES);
2503 string start_key = op.marker;
2505 int ret = list_plain_entries(hctx, op.name, op.marker, max, &op_ret.entries, &more);
2507 CLS_LOG(0, "ERROR: %s(): list_plain_entries retured ret=%d", __func__, ret);
2512 CLS_LOG(20, "found %d plain entries", count);
2515 ret = list_instance_entries(hctx, op.name, op.marker, max - count, &op_ret.entries, &more);
2517 CLS_LOG(0, "ERROR: %s(): list_instance_entries retured ret=%d", __func__, ret);
2525 ret = list_olh_entries(hctx, op.name, op.marker, max - count, &op_ret.entries, &more);
2527 CLS_LOG(0, "ERROR: %s(): list_olh_entries retured ret=%d", __func__, ret);
2534 op_ret.is_truncated = (count >= max) || more;
2535 while (count >= max) {
2536 op_ret.entries.pop_back();
2540 ::encode(op_ret, *out);
2545 int bi_log_record_decode(bufferlist& bl, rgw_bi_log_entry& e)
2547 bufferlist::iterator iter = bl.begin();
2550 } catch (buffer::error& err) {
2551 CLS_LOG(0, "ERROR: failed to decode rgw_bi_log_entry");
2557 static int bi_log_iterate_entries(cls_method_context_t hctx, const string& marker, const string& end_marker,
2558 string& key_iter, uint32_t max_entries, bool *truncated,
2559 int (*cb)(cls_method_context_t, const string&, rgw_bi_log_entry&, void *),
2562 CLS_LOG(10, "bi_log_iterate_range");
2564 map<string, bufferlist> keys;
2565 string filter_prefix, end_key;
2573 if (key_iter.empty()) {
2574 key = BI_PREFIX_CHAR;
2575 key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
2580 start_key = key_iter;
2583 if (end_marker.empty()) {
2584 end_key = BI_PREFIX_CHAR;
2585 end_key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX + 1]);
2587 end_key = BI_PREFIX_CHAR;
2588 end_key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
2589 end_key.append(end_marker);
2592 CLS_LOG(0, "bi_log_iterate_entries start_key=%s end_key=%s\n", start_key.c_str(), end_key.c_str());
2596 int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max_entries, &keys, truncated);
2600 map<string, bufferlist>::iterator iter = keys.begin();
2601 if (iter == keys.end())
2604 uint32_t num_keys = keys.size();
2606 for (; iter != keys.end(); ++iter,++i) {
2607 const string& key = iter->first;
2610 CLS_LOG(0, "bi_log_iterate_entries key=%s bl.length=%d\n", key.c_str(), (int)iter->second.length());
2612 if (key.compare(end_key) > 0) {
2617 ret = bi_log_record_decode(iter->second, e);
2621 ret = cb(hctx, key, e, param);
2625 if (i == num_keys - 1) {
2633 static int bi_log_list_cb(cls_method_context_t hctx, const string& key, rgw_bi_log_entry& info, void *param)
2635 list<rgw_bi_log_entry> *l = (list<rgw_bi_log_entry> *)param;
2640 static int bi_log_list_entries(cls_method_context_t hctx, const string& marker,
2641 uint32_t max, list<rgw_bi_log_entry>& entries, bool *truncated)
2645 int ret = bi_log_iterate_entries(hctx, marker, end_marker,
2646 key_iter, max, truncated,
2647 bi_log_list_cb, &entries);
2651 static int rgw_bi_log_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2653 bufferlist::iterator in_iter = in->begin();
2655 cls_rgw_bi_log_list_op op;
2657 ::decode(op, in_iter);
2658 } catch (buffer::error& err) {
2659 CLS_LOG(1, "ERROR: rgw_bi_log_list(): failed to decode entry\n");
2663 cls_rgw_bi_log_list_ret op_ret;
2664 int ret = bi_log_list_entries(hctx, op.marker, op.max, op_ret.entries, &op_ret.truncated);
2668 ::encode(op_ret, *out);
2673 static int bi_log_list_trim_cb(cls_method_context_t hctx, const string& key, rgw_bi_log_entry& info, void *param)
2675 list<rgw_bi_log_entry> *entries = (list<rgw_bi_log_entry> *)param;
2677 entries->push_back(info);
2681 static int bi_log_remove_entry(cls_method_context_t hctx, rgw_bi_log_entry& entry)
2684 key = BI_PREFIX_CHAR;
2685 key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]);
2686 key.append(entry.id);
2687 return cls_cxx_map_remove_key(hctx, key);
2690 static int bi_log_list_trim_entries(cls_method_context_t hctx,
2691 const string& start_marker, const string& end_marker,
2692 list<rgw_bi_log_entry>& entries, bool *truncated)
2695 #define MAX_TRIM_ENTRIES 1000 /* max entries to trim in a single operation */
2696 int ret = bi_log_iterate_entries(hctx, start_marker, end_marker,
2697 key_iter, MAX_TRIM_ENTRIES, truncated,
2698 bi_log_list_trim_cb, &entries);
2702 static int rgw_bi_log_trim(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2704 bufferlist::iterator in_iter = in->begin();
2706 cls_rgw_bi_log_trim_op op;
2708 ::decode(op, in_iter);
2709 } catch (buffer::error& err) {
2710 CLS_LOG(1, "ERROR: rgw_bi_log_list(): failed to decode entry\n");
2714 cls_rgw_bi_log_list_ret op_ret;
2715 list<rgw_bi_log_entry> entries;
2716 #define MAX_TRIM_ENTRIES 1000 /* don't do more than that in a single operation */
2718 int ret = bi_log_list_trim_entries(hctx, op.start_marker, op.end_marker, entries, &truncated);
2722 if (entries.empty())
2725 list<rgw_bi_log_entry>::iterator iter;
2726 for (iter = entries.begin(); iter != entries.end(); ++iter) {
2727 rgw_bi_log_entry& entry = *iter;
2729 ret = bi_log_remove_entry(hctx, entry);
2737 static int rgw_bi_log_resync(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2739 struct rgw_bucket_dir_header header;
2740 int rc = read_bucket_header(hctx, &header);
2742 CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n");
2748 struct rgw_bi_log_entry entry;
2750 entry.timestamp = real_clock::now();
2751 entry.op = RGWModifyOp::CLS_RGW_OP_RESYNC;
2752 entry.state = RGWPendingState::CLS_RGW_STATE_COMPLETE;
2755 bi_log_index_key(hctx, key, entry.id, header.ver);
2757 ::encode(entry, bl);
2759 if (entry.id > header.max_marker)
2760 header.max_marker = entry.id;
2762 header.syncstopped = false;
2764 rc = cls_cxx_map_set_val(hctx, key, &bl);
2768 return write_bucket_header(hctx, &header);
2771 static int rgw_bi_log_stop(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2773 struct rgw_bucket_dir_header header;
2774 int rc = read_bucket_header(hctx, &header);
2776 CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n");
2782 struct rgw_bi_log_entry entry;
2784 entry.timestamp = real_clock::now();
2785 entry.op = RGWModifyOp::CLS_RGW_OP_SYNCSTOP;
2786 entry.state = RGWPendingState::CLS_RGW_STATE_COMPLETE;
2789 bi_log_index_key(hctx, key, entry.id, header.ver);
2791 ::encode(entry, bl);
2793 if (entry.id > header.max_marker)
2794 header.max_marker = entry.id;
2795 header.syncstopped = true;
2797 rc = cls_cxx_map_set_val(hctx, key, &bl);
2801 return write_bucket_header(hctx, &header);
2805 static void usage_record_prefix_by_time(uint64_t epoch, string& key)
2808 snprintf(buf, sizeof(buf), "%011llu", (long long unsigned)epoch);
2812 static void usage_record_prefix_by_user(string& user, uint64_t epoch, string& key)
2814 char buf[user.size() + 32];
2815 snprintf(buf, sizeof(buf), "%s_%011llu_", user.c_str(), (long long unsigned)epoch);
2819 static void usage_record_name_by_time(uint64_t epoch, const string& user, string& bucket, string& key)
2821 char buf[32 + user.size() + bucket.size()];
2822 snprintf(buf, sizeof(buf), "%011llu_%s_%s", (long long unsigned)epoch, user.c_str(), bucket.c_str());
2826 static void usage_record_name_by_user(const string& user, uint64_t epoch, string& bucket, string& key)
2828 char buf[32 + user.size() + bucket.size()];
2829 snprintf(buf, sizeof(buf), "%s_%011llu_%s", user.c_str(), (long long unsigned)epoch, bucket.c_str());
2833 static int usage_record_decode(bufferlist& record_bl, rgw_usage_log_entry& e)
2835 bufferlist::iterator kiter = record_bl.begin();
2838 } catch (buffer::error& err) {
2839 CLS_LOG(1, "ERROR: usage_record_decode(): failed to decode record_bl\n");
2846 int rgw_user_usage_log_add(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
2848 CLS_LOG(10, "rgw_user_usage_log_add()");
2850 bufferlist::iterator in_iter = in->begin();
2851 rgw_cls_usage_log_add_op op;
2854 ::decode(op, in_iter);
2855 } catch (buffer::error& err) {
2856 CLS_LOG(1, "ERROR: rgw_user_usage_log_add(): failed to decode request\n");
2860 rgw_usage_log_info& info = op.info;
2861 vector<rgw_usage_log_entry>::iterator iter;
2863 for (iter = info.entries.begin(); iter != info.entries.end(); ++iter) {
2864 rgw_usage_log_entry& entry = *iter;
2867 rgw_user *puser = (entry.payer.empty() ? &entry.owner : &entry.payer);
2869 usage_record_name_by_time(entry.epoch, puser->to_str(), entry.bucket, key_by_time);
2871 CLS_LOG(10, "rgw_user_usage_log_add user=%s bucket=%s\n", puser->to_str().c_str(), entry.bucket.c_str());
2873 bufferlist record_bl;
2874 int ret = cls_cxx_map_get_val(hctx, key_by_time, &record_bl);
2875 if (ret < 0 && ret != -ENOENT) {
2876 CLS_LOG(1, "ERROR: rgw_user_usage_log_add(): cls_cxx_map_read_key returned %d\n", ret);
2880 rgw_usage_log_entry e;
2881 ret = usage_record_decode(record_bl, e);
2884 CLS_LOG(10, "rgw_user_usage_log_add aggregating existing bucket\n");
2888 bufferlist new_record_bl;
2889 ::encode(entry, new_record_bl);
2890 ret = cls_cxx_map_set_val(hctx, key_by_time, &new_record_bl);
2895 usage_record_name_by_user(puser->to_str(), entry.epoch, entry.bucket, key_by_user);
2896 ret = cls_cxx_map_set_val(hctx, key_by_user, &new_record_bl);
2904 static int usage_iterate_range(cls_method_context_t hctx, uint64_t start, uint64_t end,
2905 string& user, string& key_iter, uint32_t max_entries, bool *truncated,
2906 int (*cb)(cls_method_context_t, const string&, rgw_usage_log_entry&, void *),
2909 CLS_LOG(10, "usage_iterate_range");
2911 map<string, bufferlist> keys;
2913 string filter_prefix;
2914 string start_key, end_key;
2915 bool by_user = !user.empty();
2918 bool truncated_status = false;
2921 usage_record_prefix_by_time(end, end_key);
2924 user_key.append("_");
2927 if (key_iter.empty()) {
2929 usage_record_prefix_by_user(user, start, start_key);
2931 usage_record_prefix_by_time(start, start_key);
2934 start_key = key_iter;
2937 CLS_LOG(20, "usage_iterate_range start_key=%s", start_key.c_str());
2938 int ret = cls_cxx_map_get_vals(hctx, start_key, filter_prefix, max_entries, &keys, &truncated_status);
2943 *truncated = truncated_status;
2946 map<string, bufferlist>::iterator iter = keys.begin();
2947 if (iter == keys.end())
2950 uint32_t num_keys = keys.size();
2952 for (; iter != keys.end(); ++iter,++i) {
2953 const string& key = iter->first;
2954 rgw_usage_log_entry e;
2956 if (!by_user && key.compare(end_key) >= 0) {
2957 CLS_LOG(20, "usage_iterate_range reached key=%s, done", key.c_str());
2958 if (truncated_status) {
2964 if (by_user && key.compare(0, user_key.size(), user_key) != 0) {
2965 CLS_LOG(20, "usage_iterate_range reached key=%s, done", key.c_str());
2966 if (truncated_status) {
2972 ret = usage_record_decode(iter->second, e);
2976 if (e.epoch < start)
2979 /* keys are sorted by epoch, so once we're past end we're done */
2983 ret = cb(hctx, key, e, param);
2988 if (i == num_keys - 1) {
2996 static int usage_log_read_cb(cls_method_context_t hctx, const string& key, rgw_usage_log_entry& entry, void *param)
2998 map<rgw_user_bucket, rgw_usage_log_entry> *usage = (map<rgw_user_bucket, rgw_usage_log_entry> *)param;
3000 if (!entry.payer.empty()) {
3001 puser = &entry.payer;
3003 puser = &entry.owner;
3005 rgw_user_bucket ub(puser->to_str(), entry.bucket);
3006 rgw_usage_log_entry& le = (*usage)[ub];
3007 le.aggregate(entry);
3012 int rgw_user_usage_log_read(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3014 CLS_LOG(10, "rgw_user_usage_log_read()");
3016 bufferlist::iterator in_iter = in->begin();
3017 rgw_cls_usage_log_read_op op;
3020 ::decode(op, in_iter);
3021 } catch (buffer::error& err) {
3022 CLS_LOG(1, "ERROR: rgw_user_usage_log_read(): failed to decode request\n");
3026 rgw_cls_usage_log_read_ret ret_info;
3027 map<rgw_user_bucket, rgw_usage_log_entry> *usage = &ret_info.usage;
3028 string iter = op.iter;
3029 #define MAX_ENTRIES 1000
3030 uint32_t max_entries = (op.max_entries ? op.max_entries : MAX_ENTRIES);
3031 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);
3035 if (ret_info.truncated)
3036 ret_info.next_iter = iter;
3038 ::encode(ret_info, *out);
3042 static int usage_log_trim_cb(cls_method_context_t hctx, const string& key, rgw_usage_log_entry& entry, void *param)
3047 string o = entry.owner.to_str();
3048 usage_record_name_by_time(entry.epoch, o, entry.bucket, key_by_time);
3049 usage_record_name_by_user(o, entry.epoch, entry.bucket, key_by_user);
3051 int ret = cls_cxx_map_remove_key(hctx, key_by_time);
3055 return cls_cxx_map_remove_key(hctx, key_by_user);
3058 int rgw_user_usage_log_trim(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3060 CLS_LOG(10, "rgw_user_usage_log_trim()");
3062 /* only continue if object exists! */
3063 int ret = cls_cxx_stat(hctx, NULL, NULL);
3067 bufferlist::iterator in_iter = in->begin();
3068 rgw_cls_usage_log_trim_op op;
3071 ::decode(op, in_iter);
3072 } catch (buffer::error& err) {
3073 CLS_LOG(1, "ERROR: rgw_user_log_usage_log_trim(): failed to decode request\n");
3079 #define MAX_USAGE_TRIM_ENTRIES 128
3080 ret = usage_iterate_range(hctx, op.start_epoch, op.end_epoch, op.user, iter, MAX_USAGE_TRIM_ENTRIES, &more, usage_log_trim_cb, NULL);
3088 * We hold the garbage collection chain data under two different indexes: the first 'name' index
3089 * keeps them under a unique tag that represents the chains, and a second 'time' index keeps
3090 * them by their expiration timestamp
3092 #define GC_OBJ_NAME_INDEX 0
3093 #define GC_OBJ_TIME_INDEX 1
3095 static string gc_index_prefixes[] = { "0_",
3098 static void prepend_index_prefix(const string& src, int index, string *dest)
3100 *dest = gc_index_prefixes[index];
3104 static int gc_omap_get(cls_method_context_t hctx, int type, const string& key, cls_rgw_gc_obj_info *info)
3107 prepend_index_prefix(key, type, &index);
3110 int ret = cls_cxx_map_get_val(hctx, index, &bl);
3115 bufferlist::iterator iter = bl.begin();
3116 ::decode(*info, iter);
3117 } catch (buffer::error& err) {
3118 CLS_LOG(0, "ERROR: rgw_cls_gc_omap_get(): failed to decode index=%s\n", index.c_str());
3124 static int gc_omap_set(cls_method_context_t hctx, int type, const string& key, const cls_rgw_gc_obj_info *info)
3127 ::encode(*info, bl);
3129 string index = gc_index_prefixes[type];
3132 int ret = cls_cxx_map_set_val(hctx, index, &bl);
3139 static int gc_omap_remove(cls_method_context_t hctx, int type, const string& key)
3141 string index = gc_index_prefixes[type];
3144 int ret = cls_cxx_map_remove_key(hctx, index);
3151 static bool key_in_index(const string& key, int index_type)
3153 const string& prefix = gc_index_prefixes[index_type];
3154 return (key.compare(0, prefix.size(), prefix) == 0);
3158 static int gc_update_entry(cls_method_context_t hctx, uint32_t expiration_secs,
3159 cls_rgw_gc_obj_info& info)
3161 cls_rgw_gc_obj_info old_info;
3162 int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, info.tag, &old_info);
3165 get_time_key(old_info.time, &key);
3166 ret = gc_omap_remove(hctx, GC_OBJ_TIME_INDEX, key);
3167 if (ret < 0 && ret != -ENOENT) {
3168 CLS_LOG(0, "ERROR: failed to remove key=%s\n", key.c_str());
3172 info.time = ceph::real_clock::now();
3173 info.time += make_timespan(expiration_secs);
3174 ret = gc_omap_set(hctx, GC_OBJ_NAME_INDEX, info.tag, &info);
3179 get_time_key(info.time, &key);
3180 ret = gc_omap_set(hctx, GC_OBJ_TIME_INDEX, key, &info);
3187 CLS_LOG(0, "ERROR: gc_set_entry error info.tag=%s, ret=%d\n", info.tag.c_str(), ret);
3188 gc_omap_remove(hctx, GC_OBJ_NAME_INDEX, info.tag);
3192 static int gc_defer_entry(cls_method_context_t hctx, const string& tag, uint32_t expiration_secs)
3194 cls_rgw_gc_obj_info info;
3195 int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, tag, &info);
3200 return gc_update_entry(hctx, expiration_secs, info);
3203 int gc_record_decode(bufferlist& bl, cls_rgw_gc_obj_info& e)
3205 bufferlist::iterator iter = bl.begin();
3208 } catch (buffer::error& err) {
3209 CLS_LOG(0, "ERROR: failed to decode cls_rgw_gc_obj_info");
3215 static int rgw_cls_gc_set_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3217 bufferlist::iterator in_iter = in->begin();
3219 cls_rgw_gc_set_entry_op op;
3221 ::decode(op, in_iter);
3222 } catch (buffer::error& err) {
3223 CLS_LOG(1, "ERROR: rgw_cls_gc_set_entry(): failed to decode entry\n");
3227 return gc_update_entry(hctx, op.expiration_secs, op.info);
3230 static int rgw_cls_gc_defer_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3232 bufferlist::iterator in_iter = in->begin();
3234 cls_rgw_gc_defer_entry_op op;
3236 ::decode(op, in_iter);
3237 } catch (buffer::error& err) {
3238 CLS_LOG(1, "ERROR: rgw_cls_gc_defer_entry(): failed to decode entry\n");
3242 return gc_defer_entry(hctx, op.tag, op.expiration_secs);
3245 static int gc_iterate_entries(cls_method_context_t hctx, const string& marker, bool expired_only,
3246 string& key_iter, uint32_t max_entries, bool *truncated,
3247 int (*cb)(cls_method_context_t, const string&, cls_rgw_gc_obj_info&, void *),
3250 CLS_LOG(10, "gc_iterate_range");
3252 map<string, bufferlist> keys;
3253 string filter_prefix, end_key;
3261 if (marker.empty()) {
3262 prepend_index_prefix(marker, GC_OBJ_TIME_INDEX, &start_key);
3268 real_time now = ceph::real_clock::now();
3270 get_time_key(now, &now_str);
3271 prepend_index_prefix(now_str, GC_OBJ_TIME_INDEX, &end_key);
3273 CLS_LOG(0, "gc_iterate_entries end_key=%s\n", end_key.c_str());
3278 int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max_entries, &keys, truncated);
3283 map<string, bufferlist>::iterator iter = keys.begin();
3284 if (iter == keys.end())
3287 uint32_t num_keys = keys.size();
3289 for (; iter != keys.end(); ++iter, ++i) {
3290 const string& key = iter->first;
3291 cls_rgw_gc_obj_info e;
3293 CLS_LOG(10, "gc_iterate_entries key=%s\n", key.c_str());
3295 if (!end_key.empty() && key.compare(end_key) >= 0)
3298 if (!key_in_index(key, GC_OBJ_TIME_INDEX))
3301 ret = gc_record_decode(iter->second, e);
3305 ret = cb(hctx, key, e, param);
3309 if (i == num_keys - 1) {
3317 static int gc_list_cb(cls_method_context_t hctx, const string& key, cls_rgw_gc_obj_info& info, void *param)
3319 list<cls_rgw_gc_obj_info> *l = (list<cls_rgw_gc_obj_info> *)param;
3324 static int gc_list_entries(cls_method_context_t hctx, const string& marker,
3325 uint32_t max, bool expired_only,
3326 list<cls_rgw_gc_obj_info>& entries, bool *truncated, string& next_marker)
3328 int ret = gc_iterate_entries(hctx, marker, expired_only,
3329 next_marker, max, truncated,
3330 gc_list_cb, &entries);
3334 static int rgw_cls_gc_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3336 bufferlist::iterator in_iter = in->begin();
3338 cls_rgw_gc_list_op op;
3340 ::decode(op, in_iter);
3341 } catch (buffer::error& err) {
3342 CLS_LOG(1, "ERROR: rgw_cls_gc_list(): failed to decode entry\n");
3346 cls_rgw_gc_list_ret op_ret;
3347 #define GC_LIST_ENTRIES_DEFAULT 128
3348 int ret = gc_list_entries(hctx, op.marker, (op.max ? op.max : GC_LIST_ENTRIES_DEFAULT), op.expired_only,
3349 op_ret.entries, &op_ret.truncated, op_ret.next_marker);
3353 ::encode(op_ret, *out);
3358 static int gc_remove(cls_method_context_t hctx, list<string>& tags)
3360 list<string>::iterator iter;
3362 for (iter = tags.begin(); iter != tags.end(); ++iter) {
3363 string& tag = *iter;
3364 cls_rgw_gc_obj_info info;
3365 int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, tag, &info);
3366 if (ret == -ENOENT) {
3367 CLS_LOG(0, "couldn't find tag in name index tag=%s\n", tag.c_str());
3375 get_time_key(info.time, &time_key);
3376 ret = gc_omap_remove(hctx, GC_OBJ_TIME_INDEX, time_key);
3377 if (ret < 0 && ret != -ENOENT)
3379 if (ret == -ENOENT) {
3380 CLS_LOG(0, "couldn't find key in time index key=%s\n", time_key.c_str());
3383 ret = gc_omap_remove(hctx, GC_OBJ_NAME_INDEX, tag);
3384 if (ret < 0 && ret != -ENOENT)
3391 static int rgw_cls_gc_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3393 bufferlist::iterator in_iter = in->begin();
3395 cls_rgw_gc_remove_op op;
3397 ::decode(op, in_iter);
3398 } catch (buffer::error& err) {
3399 CLS_LOG(1, "ERROR: rgw_cls_gc_remove(): failed to decode entry\n");
3403 return gc_remove(hctx, op.tags);
3406 static int rgw_cls_lc_set_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3408 bufferlist::iterator in_iter = in->begin();
3410 cls_rgw_lc_set_entry_op op;
3412 ::decode(op, in_iter);
3413 } catch (buffer::error& err) {
3414 CLS_LOG(1, "ERROR: rgw_cls_lc_set_entry(): failed to decode entry\n");
3419 ::encode(op.entry, bl);
3421 int ret = cls_cxx_map_set_val(hctx, op.entry.first, &bl);
3425 static int rgw_cls_lc_rm_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3427 bufferlist::iterator in_iter = in->begin();
3429 cls_rgw_lc_rm_entry_op op;
3431 ::decode(op, in_iter);
3432 } catch (buffer::error& err) {
3433 CLS_LOG(1, "ERROR: rgw_cls_lc_rm_entry(): failed to decode entry\n");
3438 ::encode(op.entry, bl);
3440 int ret = cls_cxx_map_remove_key(hctx, op.entry.first);
3444 static int rgw_cls_lc_get_next_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3446 bufferlist::iterator in_iter = in->begin();
3447 cls_rgw_lc_get_next_entry_ret op_ret;
3448 cls_rgw_lc_get_next_entry_op op;
3450 ::decode(op, in_iter);
3451 } catch (buffer::error& err) {
3452 CLS_LOG(1, "ERROR: rgw_cls_lc_get_next_entry: failed to decode op\n");
3456 map<string, bufferlist> vals;
3457 string filter_prefix;
3459 int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, 1, &vals, &more);
3462 map<string, bufferlist>::iterator it;
3463 pair<string, int> entry;
3464 if (!vals.empty()) {
3466 in_iter = it->second.begin();
3468 ::decode(entry, in_iter);
3469 } catch (buffer::error& err) {
3470 CLS_LOG(1, "ERROR: rgw_cls_lc_get_next_entry(): failed to decode entry\n");
3474 op_ret.entry = entry;
3475 ::encode(op_ret, *out);
3479 static int rgw_cls_lc_list_entries(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3481 cls_rgw_lc_list_entries_op op;
3482 bufferlist::iterator in_iter = in->begin();
3484 ::decode(op, in_iter);
3485 } catch (buffer::error& err) {
3486 CLS_LOG(1, "ERROR: rgw_cls_lc_list_entries(): failed to decode op\n");
3490 cls_rgw_lc_list_entries_ret op_ret;
3491 bufferlist::iterator iter;
3492 map<string, bufferlist> vals;
3493 string filter_prefix;
3494 int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, op.max_entries, &vals, &op_ret.is_truncated);
3497 map<string, bufferlist>::iterator it;
3498 pair<string, int> entry;
3499 for (it = vals.begin(); it != vals.end(); ++it) {
3500 iter = it->second.begin();
3502 ::decode(entry, iter);
3503 } catch (buffer::error& err) {
3504 CLS_LOG(1, "ERROR: rgw_cls_lc_list_entries(): failed to decode entry\n");
3507 op_ret.entries.insert(entry);
3509 ::encode(op_ret, *out);
3513 static int rgw_cls_lc_put_head(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3515 bufferlist::iterator in_iter = in->begin();
3517 cls_rgw_lc_put_head_op op;
3519 ::decode(op, in_iter);
3520 } catch (buffer::error& err) {
3521 CLS_LOG(1, "ERROR: rgw_cls_lc_put_head(): failed to decode entry\n");
3526 ::encode(op.head, bl);
3527 int ret = cls_cxx_map_write_header(hctx,&bl);
3531 static int rgw_cls_lc_get_head(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3534 int ret = cls_cxx_map_read_header(hctx, &bl);
3537 cls_rgw_lc_obj_head head;
3538 if (bl.length() != 0) {
3539 bufferlist::iterator iter = bl.begin();
3541 ::decode(head, iter);
3542 } catch (buffer::error& err) {
3543 CLS_LOG(0, "ERROR: rgw_cls_lc_get_head(): failed to decode entry %s\n",err.what());
3547 head.start_date = 0;
3548 head.marker.clear();
3550 cls_rgw_lc_get_head_ret op_ret;
3552 ::encode(op_ret, *out);
3556 static int rgw_reshard_add(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3558 bufferlist::iterator in_iter = in->begin();
3560 cls_rgw_reshard_add_op op;
3562 ::decode(op, in_iter);
3563 } catch (buffer::error& err) {
3564 CLS_LOG(1, "ERROR: rgw_reshard_add: failed to decode entry\n");
3570 op.entry.get_key(&key);
3573 ::encode(op.entry, bl);
3574 int ret = cls_cxx_map_set_val(hctx, key, &bl);
3576 CLS_ERR("error adding reshard job for bucket %s with key %s",op.entry.bucket_name.c_str(), key.c_str());
3583 static int rgw_reshard_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3585 cls_rgw_reshard_list_op op;
3586 bufferlist::iterator in_iter = in->begin();
3588 ::decode(op, in_iter);
3589 } catch (buffer::error& err) {
3590 CLS_LOG(1, "ERROR: rgw_cls_rehard_list(): failed to decode entry\n");
3593 cls_rgw_reshard_list_ret op_ret;
3594 bufferlist::iterator iter;
3595 map<string, bufferlist> vals;
3596 string filter_prefix;
3597 #define MAX_RESHARD_LIST_ENTRIES 1000
3598 /* one extra entry for identifying truncation */
3599 int32_t max = (op.max && (op.max < MAX_RESHARD_LIST_ENTRIES) ? op.max : MAX_RESHARD_LIST_ENTRIES);
3600 int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, max, &vals, &op_ret.is_truncated);
3603 map<string, bufferlist>::iterator it;
3604 cls_rgw_reshard_entry entry;
3606 for (it = vals.begin(); i < (int)op.max && it != vals.end(); ++it, ++i) {
3607 iter = it->second.begin();
3609 ::decode(entry, iter);
3610 } catch (buffer::error& err) {
3611 CLS_LOG(1, "ERROR: rgw_cls_rehard_list(): failed to decode entry\n");
3614 op_ret.entries.push_back(entry);
3616 ::encode(op_ret, *out);
3620 static int get_reshard_entry(cls_method_context_t hctx, const string& key, cls_rgw_reshard_entry *entry)
3623 int ret = cls_cxx_map_get_val(hctx, key, &bl);
3626 bufferlist::iterator iter = bl.begin();
3628 ::decode(*entry, iter);
3629 } catch (buffer::error& err) {
3630 CLS_LOG(0, "ERROR: %s : failed to decode entry %s\n", __func__, err.what());
3636 static int rgw_reshard_get(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3638 bufferlist::iterator in_iter = in->begin();
3640 cls_rgw_reshard_get_op op;
3642 ::decode(op, in_iter);
3643 } catch (buffer::error& err) {
3644 CLS_LOG(1, "ERROR: rgw_reshard_get: failed to decode entry\n");
3649 cls_rgw_reshard_entry entry;
3650 op.entry.get_key(&key);
3651 int ret = get_reshard_entry(hctx, key, &entry);
3656 cls_rgw_reshard_get_ret op_ret;
3657 op_ret.entry = entry;
3658 ::encode(op_ret, *out);
3662 static int rgw_reshard_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3664 bufferlist::iterator in_iter = in->begin();
3666 cls_rgw_reshard_remove_op op;
3668 ::decode(op, in_iter);
3669 } catch (buffer::error& err) {
3670 CLS_LOG(1, "ERROR: rgw_cls_rehard_remove: failed to decode entry\n");
3675 cls_rgw_reshard_entry entry;
3676 cls_rgw_reshard_entry::generate_key(op.tenant, op.bucket_name, &key);
3677 int ret = get_reshard_entry(hctx, key, &entry);
3682 if (!op.bucket_id.empty() &&
3683 entry.bucket_id != op.bucket_id) {
3687 ret = cls_cxx_map_remove_key(hctx, key);
3689 CLS_LOG(0, "ERROR: failed to remove key: key=%s ret=%d", key.c_str(), ret);
3695 static int rgw_set_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3697 cls_rgw_set_bucket_resharding_op op;
3699 bufferlist::iterator in_iter = in->begin();
3701 ::decode(op, in_iter);
3702 } catch (buffer::error& err) {
3703 CLS_LOG(1, "ERROR: cls_rgw_set_bucket_resharding: failed to decode entry\n");
3707 struct rgw_bucket_dir_header header;
3708 int rc = read_bucket_header(hctx, &header);
3710 CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3714 header.new_instance.set_status(op.entry.new_bucket_instance_id, op.entry.num_shards, op.entry.reshard_status);
3716 return write_bucket_header(hctx, &header);
3719 static int rgw_clear_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3721 cls_rgw_set_bucket_resharding_op op;
3723 bufferlist::iterator in_iter = in->begin();
3725 ::decode(op, in_iter);
3726 } catch (buffer::error& err) {
3727 CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n");
3731 struct rgw_bucket_dir_header header;
3732 int rc = read_bucket_header(hctx, &header);
3734 CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3737 header.new_instance.clear();
3739 return write_bucket_header(hctx, &header);
3742 static int rgw_guard_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3744 cls_rgw_guard_bucket_resharding_op op;
3746 bufferlist::iterator in_iter = in->begin();
3748 ::decode(op, in_iter);
3749 } catch (buffer::error& err) {
3750 CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n");
3754 struct rgw_bucket_dir_header header;
3755 int rc = read_bucket_header(hctx, &header);
3757 CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3761 if (header.resharding()) {
3768 static int rgw_get_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
3770 cls_rgw_get_bucket_resharding_op op;
3772 bufferlist::iterator in_iter = in->begin();
3774 ::decode(op, in_iter);
3775 } catch (buffer::error& err) {
3776 CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n");
3780 struct rgw_bucket_dir_header header;
3781 int rc = read_bucket_header(hctx, &header);
3783 CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__);
3787 cls_rgw_get_bucket_resharding_ret op_ret;
3788 op_ret.new_instance = header.new_instance;
3790 ::encode(op_ret, *out);
3797 CLS_LOG(1, "Loaded rgw class!");
3799 cls_handle_t h_class;
3800 cls_method_handle_t h_rgw_bucket_init_index;
3801 cls_method_handle_t h_rgw_bucket_set_tag_timeout;
3802 cls_method_handle_t h_rgw_bucket_list;
3803 cls_method_handle_t h_rgw_bucket_check_index;
3804 cls_method_handle_t h_rgw_bucket_rebuild_index;
3805 cls_method_handle_t h_rgw_bucket_update_stats;
3806 cls_method_handle_t h_rgw_bucket_prepare_op;
3807 cls_method_handle_t h_rgw_bucket_complete_op;
3808 cls_method_handle_t h_rgw_bucket_link_olh;
3809 cls_method_handle_t h_rgw_bucket_unlink_instance_op;
3810 cls_method_handle_t h_rgw_bucket_read_olh_log;
3811 cls_method_handle_t h_rgw_bucket_trim_olh_log;
3812 cls_method_handle_t h_rgw_bucket_clear_olh;
3813 cls_method_handle_t h_rgw_obj_remove;
3814 cls_method_handle_t h_rgw_obj_store_pg_ver;
3815 cls_method_handle_t h_rgw_obj_check_attrs_prefix;
3816 cls_method_handle_t h_rgw_obj_check_mtime;
3817 cls_method_handle_t h_rgw_bi_get_op;
3818 cls_method_handle_t h_rgw_bi_put_op;
3819 cls_method_handle_t h_rgw_bi_list_op;
3820 cls_method_handle_t h_rgw_bi_log_list_op;
3821 cls_method_handle_t h_rgw_bi_log_resync_op;
3822 cls_method_handle_t h_rgw_bi_log_stop_op;
3823 cls_method_handle_t h_rgw_dir_suggest_changes;
3824 cls_method_handle_t h_rgw_user_usage_log_add;
3825 cls_method_handle_t h_rgw_user_usage_log_read;
3826 cls_method_handle_t h_rgw_user_usage_log_trim;
3827 cls_method_handle_t h_rgw_gc_set_entry;
3828 cls_method_handle_t h_rgw_gc_list;
3829 cls_method_handle_t h_rgw_gc_remove;
3830 cls_method_handle_t h_rgw_lc_set_entry;
3831 cls_method_handle_t h_rgw_lc_rm_entry;
3832 cls_method_handle_t h_rgw_lc_get_next_entry;
3833 cls_method_handle_t h_rgw_lc_put_head;
3834 cls_method_handle_t h_rgw_lc_get_head;
3835 cls_method_handle_t h_rgw_lc_list_entries;
3836 cls_method_handle_t h_rgw_reshard_add;
3837 cls_method_handle_t h_rgw_reshard_list;
3838 cls_method_handle_t h_rgw_reshard_get;
3839 cls_method_handle_t h_rgw_reshard_remove;
3840 cls_method_handle_t h_rgw_set_bucket_resharding;
3841 cls_method_handle_t h_rgw_clear_bucket_resharding;
3842 cls_method_handle_t h_rgw_guard_bucket_resharding;
3843 cls_method_handle_t h_rgw_get_bucket_resharding;
3846 cls_register(RGW_CLASS, &h_class);
3849 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);
3850 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);
3851 cls_register_cxx_method(h_class, RGW_BUCKET_LIST, CLS_METHOD_RD, rgw_bucket_list, &h_rgw_bucket_list);
3852 cls_register_cxx_method(h_class, RGW_BUCKET_CHECK_INDEX, CLS_METHOD_RD, rgw_bucket_check_index, &h_rgw_bucket_check_index);
3853 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);
3854 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);
3855 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);
3856 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);
3857 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);
3858 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);
3859 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);
3860 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);
3861 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);
3863 cls_register_cxx_method(h_class, RGW_OBJ_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_obj_remove, &h_rgw_obj_remove);
3864 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);
3865 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);
3866 cls_register_cxx_method(h_class, RGW_OBJ_CHECK_MTIME, CLS_METHOD_RD, rgw_obj_check_mtime, &h_rgw_obj_check_mtime);
3868 cls_register_cxx_method(h_class, RGW_BI_GET, CLS_METHOD_RD, rgw_bi_get_op, &h_rgw_bi_get_op);
3869 cls_register_cxx_method(h_class, RGW_BI_PUT, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_put_op, &h_rgw_bi_put_op);
3870 cls_register_cxx_method(h_class, RGW_BI_LIST, CLS_METHOD_RD, rgw_bi_list_op, &h_rgw_bi_list_op);
3872 cls_register_cxx_method(h_class, RGW_BI_LOG_LIST, CLS_METHOD_RD, rgw_bi_log_list, &h_rgw_bi_log_list_op);
3873 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);
3874 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);
3876 cls_register_cxx_method(h_class, "bi_log_resync", CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_resync, &h_rgw_bi_log_resync_op);
3877 cls_register_cxx_method(h_class, "bi_log_stop", CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_stop, &h_rgw_bi_log_stop_op);
3880 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);
3881 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);
3882 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);
3884 /* garbage collection */
3885 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);
3886 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);
3887 cls_register_cxx_method(h_class, RGW_GC_LIST, CLS_METHOD_RD, rgw_cls_gc_list, &h_rgw_gc_list);
3888 cls_register_cxx_method(h_class, RGW_GC_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_remove, &h_rgw_gc_remove);
3890 /* lifecycle bucket list */
3891 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);
3892 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);
3893 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);
3894 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);
3895 cls_register_cxx_method(h_class, RGW_LC_GET_HEAD, CLS_METHOD_RD, rgw_cls_lc_get_head, &h_rgw_lc_get_head);
3896 cls_register_cxx_method(h_class, RGW_LC_LIST_ENTRIES, CLS_METHOD_RD, rgw_cls_lc_list_entries, &h_rgw_lc_list_entries);
3897 cls_register_cxx_method(h_class, "reshard_add", CLS_METHOD_RD | CLS_METHOD_WR, rgw_reshard_add, &h_rgw_reshard_add);
3898 cls_register_cxx_method(h_class, "reshard_list", CLS_METHOD_RD, rgw_reshard_list, &h_rgw_reshard_list);
3899 cls_register_cxx_method(h_class, "reshard_get", CLS_METHOD_RD,rgw_reshard_get, &h_rgw_reshard_get);
3900 cls_register_cxx_method(h_class, "reshard_remove", CLS_METHOD_RD | CLS_METHOD_WR, rgw_reshard_remove, &h_rgw_reshard_remove);
3901 cls_register_cxx_method(h_class, "set_bucket_resharding", CLS_METHOD_RD | CLS_METHOD_WR,
3902 rgw_set_bucket_resharding, &h_rgw_set_bucket_resharding);
3903 cls_register_cxx_method(h_class, "clear_bucket_resharding", CLS_METHOD_RD | CLS_METHOD_WR,
3904 rgw_clear_bucket_resharding, &h_rgw_clear_bucket_resharding);
3905 cls_register_cxx_method(h_class, "guard_bucket_resharding", CLS_METHOD_RD ,
3906 rgw_guard_bucket_resharding, &h_rgw_guard_bucket_resharding);
3907 cls_register_cxx_method(h_class, "get_bucket_resharding", CLS_METHOD_RD ,
3908 rgw_get_bucket_resharding, &h_rgw_get_bucket_resharding);