]> git.proxmox.com Git - ceph.git/blob - ceph/src/tools/ceph_objectstore_tool.cc
update sources to v12.2.4
[ceph.git] / ceph / src / tools / ceph_objectstore_tool.cc
1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
3 /*
4 * Ceph - scalable distributed file system
5 *
6 * Copyright (C) 2013 Inktank
7 *
8 * This is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License version 2.1, as published by the Free Software
11 * Foundation. See file COPYING.
12 *
13 */
14
15 #include <boost/program_options/variables_map.hpp>
16 #include <boost/program_options/parsers.hpp>
17 #include <boost/scoped_ptr.hpp>
18 #include <boost/optional.hpp>
19
20 #include <stdlib.h>
21
22 #include "common/Formatter.h"
23 #include "common/errno.h"
24 #include "common/ceph_argparse.h"
25
26 #include "global/global_init.h"
27
28 #include "os/ObjectStore.h"
29 #include "os/filestore/FileJournal.h"
30 #include "os/filestore/FileStore.h"
31 #ifdef HAVE_LIBFUSE
32 #include "os/FuseStore.h"
33 #endif
34
35 #include "osd/PGLog.h"
36 #include "osd/OSD.h"
37 #include "osd/PG.h"
38
39 #include "json_spirit/json_spirit_value.h"
40 #include "json_spirit/json_spirit_reader.h"
41
42 #include "rebuild_mondb.h"
43 #include "ceph_objectstore_tool.h"
44 #include "include/compat.h"
45 #include "include/util.h"
46
47 namespace po = boost::program_options;
48 using namespace std;
49
50 #ifdef INTERNAL_TEST
51 CompatSet get_test_compat_set() {
52 CompatSet::FeatureSet ceph_osd_feature_compat;
53 CompatSet::FeatureSet ceph_osd_feature_ro_compat;
54 CompatSet::FeatureSet ceph_osd_feature_incompat;
55 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE);
56 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_PGINFO);
57 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_OLOC);
58 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_LEC);
59 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_CATEGORIES);
60 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_HOBJECTPOOL);
61 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_BIGINFO);
62 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_LEVELDBINFO);
63 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_LEVELDBLOG);
64 #ifdef INTERNAL_TEST2
65 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_SNAPMAPPER);
66 ceph_osd_feature_incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS);
67 #endif
68 return CompatSet(ceph_osd_feature_compat, ceph_osd_feature_ro_compat,
69 ceph_osd_feature_incompat);
70 }
71 #endif
72
73 const ssize_t max_read = 1024 * 1024;
74 const int fd_none = INT_MIN;
75 bool outistty;
76 bool dry_run;
77
78 struct action_on_object_t {
79 virtual ~action_on_object_t() {}
80 virtual int call(ObjectStore *store, coll_t coll, ghobject_t &ghobj, object_info_t &oi) = 0;
81 };
82
83 int _action_on_all_objects_in_pg(ObjectStore *store, coll_t coll, action_on_object_t &action, bool debug)
84 {
85 unsigned LIST_AT_A_TIME = 100;
86 ghobject_t next;
87 while (!next.is_max()) {
88 vector<ghobject_t> list;
89 int r = store->collection_list(
90 coll,
91 next,
92 ghobject_t::get_max(),
93 LIST_AT_A_TIME,
94 &list,
95 &next);
96 if (r < 0) {
97 cerr << "Error listing collection: " << coll << ", "
98 << cpp_strerror(r) << std::endl;
99 return r;
100 }
101 for (vector<ghobject_t>::iterator obj = list.begin();
102 obj != list.end();
103 ++obj) {
104 if (obj->is_pgmeta())
105 continue;
106 object_info_t oi;
107 if (coll != coll_t::meta()) {
108 bufferlist attr;
109 r = store->getattr(coll, *obj, OI_ATTR, attr);
110 if (r < 0) {
111 cerr << "Error getting attr on : " << make_pair(coll, *obj) << ", "
112 << cpp_strerror(r) << std::endl;
113 continue;
114 }
115 bufferlist::iterator bp = attr.begin();
116 try {
117 ::decode(oi, bp);
118 } catch (...) {
119 r = -EINVAL;
120 cerr << "Error getting attr on : " << make_pair(coll, *obj) << ", "
121 << cpp_strerror(r) << std::endl;
122 continue;
123 }
124 }
125 r = action.call(store, coll, *obj, oi);
126 if (r < 0)
127 return r;
128 }
129 }
130 return 0;
131 }
132
133 int action_on_all_objects_in_pg(ObjectStore *store, string pgidstr, action_on_object_t &action, bool debug)
134 {
135 spg_t pgid;
136 // Scan collections in case this is an ec pool but no shard specified
137 unsigned scanned = 0;
138 int r = 0;
139 vector<coll_t> colls_to_check;
140 vector<coll_t> candidates;
141 r = store->list_collections(candidates);
142 if (r < 0) {
143 cerr << "Error listing collections: " << cpp_strerror(r) << std::endl;
144 return r;
145 }
146 pgid.parse(pgidstr.c_str());
147 for (vector<coll_t>::iterator i = candidates.begin();
148 i != candidates.end();
149 ++i) {
150 spg_t cand_pgid;
151 if (!i->is_pg(&cand_pgid))
152 continue;
153
154 // If an exact match or treat no shard as any shard
155 if (cand_pgid == pgid ||
156 (pgid.is_no_shard() && pgid.pgid == cand_pgid.pgid)) {
157 colls_to_check.push_back(*i);
158 }
159 }
160
161 if (debug)
162 cerr << colls_to_check.size() << " pgs to scan" << std::endl;
163 for (vector<coll_t>::iterator i = colls_to_check.begin();
164 i != colls_to_check.end();
165 ++i, ++scanned) {
166 if (debug)
167 cerr << "Scanning " << *i << ", " << scanned << "/"
168 << colls_to_check.size() << " completed" << std::endl;
169 r = _action_on_all_objects_in_pg(store, *i, action, debug);
170 if (r < 0)
171 break;
172 }
173 return r;
174 }
175
176 int action_on_all_objects_in_exact_pg(ObjectStore *store, coll_t coll, action_on_object_t &action, bool debug)
177 {
178 int r = _action_on_all_objects_in_pg(store, coll, action, debug);
179 return r;
180 }
181
182 int _action_on_all_objects(ObjectStore *store, action_on_object_t &action, bool debug)
183 {
184 unsigned scanned = 0;
185 int r = 0;
186 vector<coll_t> colls_to_check;
187 vector<coll_t> candidates;
188 r = store->list_collections(candidates);
189 if (r < 0) {
190 cerr << "Error listing collections: " << cpp_strerror(r) << std::endl;
191 return r;
192 }
193 for (vector<coll_t>::iterator i = candidates.begin();
194 i != candidates.end();
195 ++i) {
196 if (i->is_pg()) {
197 colls_to_check.push_back(*i);
198 }
199 }
200
201 if (debug)
202 cerr << colls_to_check.size() << " pgs to scan" << std::endl;
203 for (vector<coll_t>::iterator i = colls_to_check.begin();
204 i != colls_to_check.end();
205 ++i, ++scanned) {
206 if (debug)
207 cerr << "Scanning " << *i << ", " << scanned << "/"
208 << colls_to_check.size() << " completed" << std::endl;
209 r = _action_on_all_objects_in_pg(store, *i, action, debug);
210 if (r < 0)
211 return r;
212 }
213 return 0;
214 }
215
216 int action_on_all_objects(ObjectStore *store, action_on_object_t &action, bool debug)
217 {
218 int r = _action_on_all_objects(store, action, debug);
219 return r;
220 }
221
222 struct pgid_object_list {
223 list<pair<coll_t, ghobject_t> > _objects;
224
225 void insert(coll_t coll, ghobject_t &ghobj) {
226 _objects.push_back(make_pair(coll, ghobj));
227 }
228
229 void dump(Formatter *f, bool human_readable) const {
230 if (!human_readable)
231 f->open_array_section("pgid_objects");
232 for (list<pair<coll_t, ghobject_t> >::const_iterator i = _objects.begin();
233 i != _objects.end();
234 ++i) {
235 f->open_array_section("pgid_object");
236 spg_t pgid;
237 bool is_pg = i->first.is_pg(&pgid);
238 if (is_pg)
239 f->dump_string("pgid", stringify(pgid));
240 if (!is_pg || !human_readable)
241 f->dump_string("coll", i->first.to_str());
242 f->open_object_section("ghobject");
243 i->second.dump(f);
244 f->close_section();
245 f->close_section();
246 if (human_readable) {
247 f->flush(cout);
248 cout << std::endl;
249 }
250 }
251 if (!human_readable) {
252 f->close_section();
253 f->flush(cout);
254 cout << std::endl;
255 }
256 }
257 };
258
259 struct lookup_ghobject : public action_on_object_t {
260 pgid_object_list _objects;
261 const string _name;
262 const boost::optional<std::string> _namespace;
263 bool _need_snapset;
264
265 lookup_ghobject(const string& name, const boost::optional<std::string>& nspace, bool need_snapset = false) : _name(name),
266 _namespace(nspace), _need_snapset(need_snapset) { }
267
268 int call(ObjectStore *store, coll_t coll, ghobject_t &ghobj, object_info_t &oi) override {
269 if (_need_snapset && !ghobj.hobj.has_snapset())
270 return 0;
271 if ((_name.length() == 0 || ghobj.hobj.oid.name == _name) &&
272 (!_namespace || ghobj.hobj.nspace == _namespace))
273 _objects.insert(coll, ghobj);
274 return 0;
275 }
276
277 int size() const {
278 return _objects._objects.size();
279 }
280
281 pair<coll_t, ghobject_t> pop() {
282 pair<coll_t, ghobject_t> front = _objects._objects.front();
283 _objects._objects.pop_front();
284 return front;
285 }
286
287 void dump(Formatter *f, bool human_readable) const {
288 _objects.dump(f, human_readable);
289 }
290 };
291
292 ghobject_t infos_oid = OSD::make_infos_oid();
293 ghobject_t log_oid;
294 ghobject_t biginfo_oid;
295
296 int file_fd = fd_none;
297 bool debug;
298 super_header sh;
299
300 static int get_fd_data(int fd, bufferlist &bl)
301 {
302 uint64_t total = 0;
303 do {
304 ssize_t bytes = bl.read_fd(fd, max_read);
305 if (bytes < 0) {
306 cerr << "read_fd error " << cpp_strerror(bytes) << std::endl;
307 return bytes;
308 }
309
310 if (bytes == 0)
311 break;
312
313 total += bytes;
314 } while(true);
315
316 assert(bl.length() == total);
317 return 0;
318 }
319
320 int get_log(ObjectStore *fs, __u8 struct_ver,
321 coll_t coll, spg_t pgid, const pg_info_t &info,
322 PGLog::IndexedLog &log, pg_missing_t &missing)
323 {
324 try {
325 ostringstream oss;
326 assert(struct_ver > 0);
327 PGLog::read_log_and_missing(fs, coll,
328 struct_ver >= 8 ? coll : coll_t::meta(),
329 struct_ver >= 8 ? pgid.make_pgmeta_oid() : log_oid,
330 info, log, missing,
331 struct_ver < 9,
332 oss,
333 g_ceph_context->_conf->osd_ignore_stale_divergent_priors);
334 if (debug && oss.str().size())
335 cerr << oss.str() << std::endl;
336 }
337 catch (const buffer::error &e) {
338 cerr << "read_log_and_missing threw exception error " << e.what() << std::endl;
339 return -EFAULT;
340 }
341 return 0;
342 }
343
344 void dump_log(Formatter *formatter, ostream &out, pg_log_t &log,
345 pg_missing_t &missing)
346 {
347 formatter->open_object_section("op_log");
348 formatter->open_object_section("pg_log_t");
349 log.dump(formatter);
350 formatter->close_section();
351 formatter->flush(out);
352 formatter->open_object_section("pg_missing_t");
353 missing.dump(formatter);
354 formatter->close_section();
355 formatter->close_section();
356 formatter->flush(out);
357 }
358
359 //Based on part of OSD::load_pgs()
360 int finish_remove_pgs(ObjectStore *store)
361 {
362 vector<coll_t> ls;
363 int r = store->list_collections(ls);
364 if (r < 0) {
365 cerr << "finish_remove_pgs: failed to list pgs: " << cpp_strerror(r)
366 << std::endl;
367 return r;
368 }
369
370 for (vector<coll_t>::iterator it = ls.begin();
371 it != ls.end();
372 ++it) {
373 spg_t pgid;
374
375 if (it->is_temp(&pgid) ||
376 (it->is_pg(&pgid) && PG::_has_removal_flag(store, pgid))) {
377 cout << "finish_remove_pgs " << *it << " removing " << pgid << std::endl;
378 OSD::recursive_remove_collection(g_ceph_context, store, pgid, *it);
379 continue;
380 }
381
382 //cout << "finish_remove_pgs ignoring unrecognized " << *it << std::endl;
383 }
384 return 0;
385 }
386
387 #pragma GCC diagnostic ignored "-Wpragmas"
388 #pragma GCC diagnostic push
389 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
390
391 int mark_pg_for_removal(ObjectStore *fs, spg_t pgid, ObjectStore::Transaction *t)
392 {
393 pg_info_t info(pgid);
394 coll_t coll(pgid);
395 ghobject_t pgmeta_oid(info.pgid.make_pgmeta_oid());
396
397 bufferlist bl;
398 epoch_t map_epoch = 0;
399 int r = PG::peek_map_epoch(fs, pgid, &map_epoch, &bl);
400 if (r < 0)
401 cerr << __func__ << " warning: peek_map_epoch reported error" << std::endl;
402 PastIntervals past_intervals;
403 __u8 struct_v;
404 r = PG::read_info(fs, pgid, coll, bl, info, past_intervals, struct_v);
405 if (r < 0) {
406 cerr << __func__ << " error on read_info " << cpp_strerror(r) << std::endl;
407 return r;
408 }
409 assert(struct_v >= 8);
410 // new omap key
411 cout << "setting '_remove' omap key" << std::endl;
412 map<string,bufferlist> values;
413 ::encode((char)1, values["_remove"]);
414 t->omap_setkeys(coll, pgmeta_oid, values);
415 return 0;
416 }
417
418 #pragma GCC diagnostic pop
419 #pragma GCC diagnostic warning "-Wpragmas"
420
421 int initiate_new_remove_pg(ObjectStore *store, spg_t r_pgid,
422 ObjectStore::Sequencer &osr)
423 {
424 if (!dry_run)
425 finish_remove_pgs(store);
426 if (!store->collection_exists(coll_t(r_pgid)))
427 return -ENOENT;
428
429 cout << " marking collection for removal" << std::endl;
430 if (dry_run)
431 return 0;
432 ObjectStore::Transaction rmt;
433 int r = mark_pg_for_removal(store, r_pgid, &rmt);
434 if (r < 0) {
435 return r;
436 }
437 store->apply_transaction(&osr, std::move(rmt));
438 finish_remove_pgs(store);
439 return r;
440 }
441
442 int write_info(ObjectStore::Transaction &t, epoch_t epoch, pg_info_t &info,
443 PastIntervals &past_intervals)
444 {
445 //Empty for this
446 coll_t coll(info.pgid);
447 ghobject_t pgmeta_oid(info.pgid.make_pgmeta_oid());
448 map<string,bufferlist> km;
449 pg_info_t last_written_info;
450 int ret = PG::_prepare_write_info(
451 g_ceph_context,
452 &km, epoch,
453 info,
454 last_written_info,
455 past_intervals,
456 true, true, false);
457 if (ret) cerr << "Failed to write info" << std::endl;
458 t.omap_setkeys(coll, pgmeta_oid, km);
459 return ret;
460 }
461
462 typedef map<eversion_t, hobject_t> divergent_priors_t;
463
464 int write_pg(ObjectStore::Transaction &t, epoch_t epoch, pg_info_t &info,
465 pg_log_t &log, PastIntervals &past_intervals,
466 divergent_priors_t &divergent,
467 pg_missing_t &missing)
468 {
469 int ret = write_info(t, epoch, info, past_intervals);
470 if (ret)
471 return ret;
472 coll_t coll(info.pgid);
473 map<string,bufferlist> km;
474
475 if (!divergent.empty()) {
476 assert(missing.get_items().empty());
477 PGLog::write_log_and_missing_wo_missing(
478 t, &km, log, coll, info.pgid.make_pgmeta_oid(), divergent, true);
479 } else {
480 pg_missing_tracker_t tmissing(missing);
481 bool rebuilt_missing_set_with_deletes = missing.may_include_deletes;
482 PGLog::write_log_and_missing(
483 t, &km, log, coll, info.pgid.make_pgmeta_oid(), tmissing, true,
484 &rebuilt_missing_set_with_deletes);
485 }
486 t.omap_setkeys(coll, info.pgid.make_pgmeta_oid(), km);
487 return 0;
488 }
489
490 const int OMAP_BATCH_SIZE = 25;
491 void get_omap_batch(ObjectMap::ObjectMapIterator &iter, map<string, bufferlist> &oset)
492 {
493 oset.clear();
494 for (int count = OMAP_BATCH_SIZE; count && iter->valid(); --count, iter->next()) {
495 oset.insert(pair<string, bufferlist>(iter->key(), iter->value()));
496 }
497 }
498
499 int ObjectStoreTool::export_file(ObjectStore *store, coll_t cid, ghobject_t &obj)
500 {
501 struct stat st;
502 mysize_t total;
503 footer ft;
504
505 int ret = store->stat(cid, obj, &st);
506 if (ret < 0)
507 return ret;
508
509 cerr << "Read " << obj << std::endl;
510
511 total = st.st_size;
512 if (debug)
513 cerr << "size=" << total << std::endl;
514
515 object_begin objb(obj);
516
517 {
518 bufferptr bp;
519 bufferlist bl;
520 ret = store->getattr(cid, obj, OI_ATTR, bp);
521 if (ret < 0) {
522 cerr << "getattr failure object_info " << ret << std::endl;
523 return ret;
524 }
525 bl.push_back(bp);
526 decode(objb.oi, bl);
527 if (debug)
528 cerr << "object_info: " << objb.oi << std::endl;
529 }
530
531 // NOTE: we include whiteouts, lost, etc.
532
533 ret = write_section(TYPE_OBJECT_BEGIN, objb, file_fd);
534 if (ret < 0)
535 return ret;
536
537 uint64_t offset = 0;
538 bufferlist rawdatabl;
539 while(total > 0) {
540 rawdatabl.clear();
541 mysize_t len = max_read;
542 if (len > total)
543 len = total;
544
545 ret = store->read(cid, obj, offset, len, rawdatabl);
546 if (ret < 0)
547 return ret;
548 if (ret == 0)
549 return -EINVAL;
550
551 data_section dblock(offset, len, rawdatabl);
552 if (debug)
553 cerr << "data section offset=" << offset << " len=" << len << std::endl;
554
555 total -= ret;
556 offset += ret;
557
558 ret = write_section(TYPE_DATA, dblock, file_fd);
559 if (ret) return ret;
560 }
561
562 //Handle attrs for this object
563 map<string,bufferptr> aset;
564 ret = store->getattrs(cid, obj, aset);
565 if (ret) return ret;
566 attr_section as(aset);
567 ret = write_section(TYPE_ATTRS, as, file_fd);
568 if (ret)
569 return ret;
570
571 if (debug) {
572 cerr << "attrs size " << aset.size() << std::endl;
573 }
574
575 //Handle omap information
576 bufferlist hdrbuf;
577 ret = store->omap_get_header(cid, obj, &hdrbuf, true);
578 if (ret < 0) {
579 cerr << "omap_get_header: " << cpp_strerror(ret) << std::endl;
580 return ret;
581 }
582
583 omap_hdr_section ohs(hdrbuf);
584 ret = write_section(TYPE_OMAP_HDR, ohs, file_fd);
585 if (ret)
586 return ret;
587
588 ObjectMap::ObjectMapIterator iter = store->get_omap_iterator(cid, obj);
589 if (!iter) {
590 ret = -ENOENT;
591 cerr << "omap_get_iterator: " << cpp_strerror(ret) << std::endl;
592 return ret;
593 }
594 iter->seek_to_first();
595 int mapcount = 0;
596 map<string, bufferlist> out;
597 while(iter->valid()) {
598 get_omap_batch(iter, out);
599
600 if (out.empty()) break;
601
602 mapcount += out.size();
603 omap_section oms(out);
604 ret = write_section(TYPE_OMAP, oms, file_fd);
605 if (ret)
606 return ret;
607 }
608 if (debug)
609 cerr << "omap map size " << mapcount << std::endl;
610
611 ret = write_simple(TYPE_OBJECT_END, file_fd);
612 if (ret)
613 return ret;
614
615 return 0;
616 }
617
618 int ObjectStoreTool::export_files(ObjectStore *store, coll_t coll)
619 {
620 ghobject_t next;
621
622 while (!next.is_max()) {
623 vector<ghobject_t> objects;
624 int r = store->collection_list(coll, next, ghobject_t::get_max(), 300,
625 &objects, &next);
626 if (r < 0)
627 return r;
628 for (vector<ghobject_t>::iterator i = objects.begin();
629 i != objects.end();
630 ++i) {
631 assert(!i->hobj.is_meta());
632 if (i->is_pgmeta() || i->hobj.is_temp()) {
633 continue;
634 }
635 r = export_file(store, coll, *i);
636 if (r < 0)
637 return r;
638 }
639 }
640 return 0;
641 }
642
643 int set_inc_osdmap(ObjectStore *store, epoch_t e, bufferlist& bl, bool force,
644 ObjectStore::Sequencer &osr) {
645 OSDMap::Incremental inc;
646 bufferlist::iterator it = bl.begin();
647 inc.decode(it);
648 if (e == 0) {
649 e = inc.epoch;
650 } else if (e != inc.epoch) {
651 cerr << "incremental.epoch mismatch: "
652 << inc.epoch << " != " << e << std::endl;
653 if (force) {
654 cerr << "But will continue anyway." << std::endl;
655 } else {
656 return -EINVAL;
657 }
658 }
659 const ghobject_t inc_oid = OSD::get_inc_osdmap_pobject_name(e);
660 if (!store->exists(coll_t::meta(), inc_oid)) {
661 cerr << "inc-osdmap (" << inc_oid << ") does not exist." << std::endl;
662 if (!force) {
663 return -ENOENT;
664 }
665 cout << "Creating a new epoch." << std::endl;
666 }
667 if (dry_run)
668 return 0;
669 ObjectStore::Transaction t;
670 t.write(coll_t::meta(), inc_oid, 0, bl.length(), bl);
671 t.truncate(coll_t::meta(), inc_oid, bl.length());
672 int ret = store->apply_transaction(&osr, std::move(t));
673 if (ret) {
674 cerr << "Failed to set inc-osdmap (" << inc_oid << "): " << ret << std::endl;
675 } else {
676 cout << "Wrote inc-osdmap." << inc.epoch << std::endl;
677 }
678 return ret;
679 }
680
681 int get_inc_osdmap(ObjectStore *store, epoch_t e, bufferlist& bl)
682 {
683 if (store->read(coll_t::meta(),
684 OSD::get_inc_osdmap_pobject_name(e),
685 0, 0, bl) < 0) {
686 return -ENOENT;
687 }
688 return 0;
689 }
690
691 int set_osdmap(ObjectStore *store, epoch_t e, bufferlist& bl, bool force,
692 ObjectStore::Sequencer &osr) {
693 OSDMap osdmap;
694 osdmap.decode(bl);
695 if (e == 0) {
696 e = osdmap.get_epoch();
697 } else if (e != osdmap.get_epoch()) {
698 cerr << "osdmap.epoch mismatch: "
699 << e << " != " << osdmap.get_epoch() << std::endl;
700 if (force) {
701 cerr << "But will continue anyway." << std::endl;
702 } else {
703 return -EINVAL;
704 }
705 }
706 const ghobject_t full_oid = OSD::get_osdmap_pobject_name(e);
707 if (!store->exists(coll_t::meta(), full_oid)) {
708 cerr << "osdmap (" << full_oid << ") does not exist." << std::endl;
709 if (!force) {
710 return -ENOENT;
711 }
712 cout << "Creating a new epoch." << std::endl;
713 }
714 if (dry_run)
715 return 0;
716 ObjectStore::Transaction t;
717 t.write(coll_t::meta(), full_oid, 0, bl.length(), bl);
718 t.truncate(coll_t::meta(), full_oid, bl.length());
719 int ret = store->apply_transaction(&osr, std::move(t));
720 if (ret) {
721 cerr << "Failed to set osdmap (" << full_oid << "): " << ret << std::endl;
722 } else {
723 cout << "Wrote osdmap." << osdmap.get_epoch() << std::endl;
724 }
725 return ret;
726 }
727
728 int get_osdmap(ObjectStore *store, epoch_t e, OSDMap &osdmap, bufferlist& bl)
729 {
730 bool found = store->read(
731 coll_t::meta(), OSD::get_osdmap_pobject_name(e), 0, 0, bl) >= 0;
732 if (!found) {
733 cerr << "Can't find OSDMap for pg epoch " << e << std::endl;
734 return -ENOENT;
735 }
736 osdmap.decode(bl);
737 if (debug)
738 cerr << osdmap << std::endl;
739 return 0;
740 }
741
742 int add_osdmap(ObjectStore *store, metadata_section &ms)
743 {
744 return get_osdmap(store, ms.map_epoch, ms.osdmap, ms.osdmap_bl);
745 }
746
747 int ObjectStoreTool::do_export(ObjectStore *fs, coll_t coll, spg_t pgid,
748 pg_info_t &info, epoch_t map_epoch, __u8 struct_ver,
749 const OSDSuperblock& superblock,
750 PastIntervals &past_intervals)
751 {
752 PGLog::IndexedLog log;
753 pg_missing_t missing;
754
755 cerr << "Exporting " << pgid << std::endl;
756
757 int ret = get_log(fs, struct_ver, coll, pgid, info, log, missing);
758 if (ret > 0)
759 return ret;
760
761 if (debug) {
762 Formatter *formatter = Formatter::create("json-pretty");
763 assert(formatter);
764 dump_log(formatter, cerr, log, missing);
765 delete formatter;
766 }
767 write_super();
768
769 pg_begin pgb(pgid, superblock);
770 // Special case: If replicated pg don't require the importing OSD to have shard feature
771 if (pgid.is_no_shard()) {
772 pgb.superblock.compat_features.incompat.remove(CEPH_OSD_FEATURE_INCOMPAT_SHARDS);
773 }
774 ret = write_section(TYPE_PG_BEGIN, pgb, file_fd);
775 if (ret)
776 return ret;
777
778 // The metadata_section is now before files, so import can detect
779 // errors and abort without wasting time.
780 metadata_section ms(
781 struct_ver,
782 map_epoch,
783 info,
784 log,
785 past_intervals,
786 missing);
787 ret = add_osdmap(fs, ms);
788 if (ret)
789 return ret;
790 ret = write_section(TYPE_PG_METADATA, ms, file_fd);
791 if (ret)
792 return ret;
793
794 ret = export_files(fs, coll);
795 if (ret) {
796 cerr << "export_files error " << ret << std::endl;
797 return ret;
798 }
799
800 ret = write_simple(TYPE_PG_END, file_fd);
801 if (ret)
802 return ret;
803
804 return 0;
805 }
806
807 int dump_data(Formatter *formatter, bufferlist &bl)
808 {
809 bufferlist::iterator ebliter = bl.begin();
810 data_section ds;
811 ds.decode(ebliter);
812
813 formatter->open_object_section("data_block");
814 formatter->dump_unsigned("offset", ds.offset);
815 formatter->dump_unsigned("len", ds.len);
816 // XXX: Add option to dump data like od -cx ?
817 formatter->close_section();
818 formatter->flush(cout);
819 return 0;
820 }
821
822 int get_data(ObjectStore *store, coll_t coll, ghobject_t hoid,
823 ObjectStore::Transaction *t, bufferlist &bl)
824 {
825 bufferlist::iterator ebliter = bl.begin();
826 data_section ds;
827 ds.decode(ebliter);
828
829 if (debug)
830 cerr << "\tdata: offset " << ds.offset << " len " << ds.len << std::endl;
831 t->write(coll, hoid, ds.offset, ds.len, ds.databl);
832 return 0;
833 }
834
835 int dump_attrs(
836 Formatter *formatter, ghobject_t hoid,
837 bufferlist &bl)
838 {
839 bufferlist::iterator ebliter = bl.begin();
840 attr_section as;
841 as.decode(ebliter);
842
843 // This could have been handled in the caller if we didn't need to
844 // support exports that didn't include object_info_t in object_begin.
845 if (hoid.generation == ghobject_t::NO_GEN &&
846 hoid.hobj.is_head()) {
847 map<string,bufferlist>::iterator mi = as.data.find(SS_ATTR);
848 if (mi != as.data.end()) {
849 SnapSet snapset;
850 auto p = mi->second.begin();
851 snapset.decode(p);
852 formatter->open_object_section("snapset");
853 snapset.dump(formatter);
854 formatter->close_section();
855 } else {
856 formatter->open_object_section("snapset");
857 formatter->dump_string("error", "missing SS_ATTR");
858 formatter->close_section();
859 }
860 }
861
862 formatter->open_object_section("attrs");
863 formatter->open_array_section("user");
864 for (auto kv : as.data) {
865 // Skip system attributes
866 if (('_' != kv.first.at(0)) || kv.first.size() == 1)
867 continue;
868 formatter->open_object_section("user_attr");
869 formatter->dump_string("name", kv.first.substr(1));
870 bool b64;
871 formatter->dump_string("value", cleanbin(kv.second, b64));
872 formatter->dump_bool("Base64", b64);
873 formatter->close_section();
874 }
875 formatter->close_section();
876 formatter->open_array_section("system");
877 for (auto kv : as.data) {
878 // Skip user attributes
879 if (('_' == kv.first.at(0)) && kv.first.size() != 1)
880 continue;
881 formatter->open_object_section("sys_attr");
882 formatter->dump_string("name", kv.first);
883 formatter->close_section();
884 }
885 formatter->close_section();
886 formatter->close_section();
887 formatter->flush(cout);
888
889 return 0;
890 }
891
892 int get_attrs(
893 ObjectStore *store, coll_t coll, ghobject_t hoid,
894 ObjectStore::Transaction *t, bufferlist &bl,
895 OSDriver &driver, SnapMapper &snap_mapper)
896 {
897 bufferlist::iterator ebliter = bl.begin();
898 attr_section as;
899 as.decode(ebliter);
900
901 if (debug)
902 cerr << "\tattrs: len " << as.data.size() << std::endl;
903 t->setattrs(coll, hoid, as.data);
904
905 // This could have been handled in the caller if we didn't need to
906 // support exports that didn't include object_info_t in object_begin.
907 if (hoid.generation == ghobject_t::NO_GEN) {
908 if (hoid.hobj.snap < CEPH_MAXSNAP) {
909 map<string,bufferlist>::iterator mi = as.data.find(OI_ATTR);
910 if (mi != as.data.end()) {
911 object_info_t oi(mi->second);
912
913 if (debug)
914 cerr << "object_info " << oi << std::endl;
915
916 OSDriver::OSTransaction _t(driver.get_transaction(t));
917 set<snapid_t> oi_snaps(oi.legacy_snaps.begin(), oi.legacy_snaps.end());
918 if (!oi_snaps.empty()) {
919 if (debug)
920 cerr << "\tsetting legacy snaps " << oi_snaps << std::endl;
921 snap_mapper.add_oid(hoid.hobj, oi_snaps, &_t);
922 }
923 }
924 } else {
925 if (hoid.hobj.is_head()) {
926 map<string,bufferlist>::iterator mi = as.data.find(SS_ATTR);
927 if (mi != as.data.end()) {
928 SnapSet snapset;
929 auto p = mi->second.begin();
930 snapset.decode(p);
931 cout << "snapset " << snapset << std::endl;
932 if (!snapset.is_legacy()) {
933 for (auto& p : snapset.clone_snaps) {
934 ghobject_t clone = hoid;
935 clone.hobj.snap = p.first;
936 set<snapid_t> snaps(p.second.begin(), p.second.end());
937 if (!store->exists(coll, clone)) {
938 // no clone, skip. this is probably a cache pool. this works
939 // because we use a separate transaction per object and clones
940 // come before head in the archive.
941 if (debug)
942 cerr << "\tskipping missing " << clone << " (snaps "
943 << snaps << ")" << std::endl;
944 continue;
945 }
946 if (debug)
947 cerr << "\tsetting " << clone.hobj << " snaps " << snaps
948 << std::endl;
949 OSDriver::OSTransaction _t(driver.get_transaction(t));
950 assert(!snaps.empty());
951 snap_mapper.add_oid(clone.hobj, snaps, &_t);
952 }
953 }
954 } else {
955 cerr << "missing SS_ATTR on " << hoid << std::endl;
956 }
957 }
958 }
959 }
960
961 return 0;
962 }
963
964 int dump_omap_hdr(Formatter *formatter, bufferlist &bl)
965 {
966 bufferlist::iterator ebliter = bl.begin();
967 omap_hdr_section oh;
968 oh.decode(ebliter);
969
970 formatter->open_object_section("omap_header");
971 formatter->dump_string("value", string(oh.hdr.c_str(), oh.hdr.length()));
972 formatter->close_section();
973 formatter->flush(cout);
974 return 0;
975 }
976
977 int get_omap_hdr(ObjectStore *store, coll_t coll, ghobject_t hoid,
978 ObjectStore::Transaction *t, bufferlist &bl)
979 {
980 bufferlist::iterator ebliter = bl.begin();
981 omap_hdr_section oh;
982 oh.decode(ebliter);
983
984 if (debug)
985 cerr << "\tomap header: " << string(oh.hdr.c_str(), oh.hdr.length())
986 << std::endl;
987 t->omap_setheader(coll, hoid, oh.hdr);
988 return 0;
989 }
990
991 int dump_omap(Formatter *formatter, bufferlist &bl)
992 {
993 bufferlist::iterator ebliter = bl.begin();
994 omap_section os;
995 os.decode(ebliter);
996
997 formatter->open_object_section("omaps");
998 formatter->dump_unsigned("count", os.omap.size());
999 formatter->open_array_section("data");
1000 for (auto o : os.omap) {
1001 formatter->open_object_section("omap");
1002 formatter->dump_string("name", o.first);
1003 bool b64;
1004 formatter->dump_string("value", cleanbin(o.second, b64));
1005 formatter->dump_bool("Base64", b64);
1006 formatter->close_section();
1007 }
1008 formatter->close_section();
1009 formatter->close_section();
1010 formatter->flush(cout);
1011 return 0;
1012 }
1013
1014 int get_omap(ObjectStore *store, coll_t coll, ghobject_t hoid,
1015 ObjectStore::Transaction *t, bufferlist &bl)
1016 {
1017 bufferlist::iterator ebliter = bl.begin();
1018 omap_section os;
1019 os.decode(ebliter);
1020
1021 if (debug)
1022 cerr << "\tomap: size " << os.omap.size() << std::endl;
1023 t->omap_setkeys(coll, hoid, os.omap);
1024 return 0;
1025 }
1026
1027 int ObjectStoreTool::dump_object(Formatter *formatter,
1028 bufferlist &bl)
1029 {
1030 bufferlist::iterator ebliter = bl.begin();
1031 object_begin ob;
1032 ob.decode(ebliter);
1033
1034 if (ob.hoid.hobj.is_temp()) {
1035 cerr << "ERROR: Export contains temporary object '" << ob.hoid << "'" << std::endl;
1036 return -EFAULT;
1037 }
1038
1039 formatter->open_object_section("object");
1040 formatter->open_object_section("oid");
1041 ob.hoid.dump(formatter);
1042 formatter->close_section();
1043 formatter->open_object_section("object_info");
1044 ob.oi.dump(formatter);
1045 formatter->close_section();
1046
1047 bufferlist ebl;
1048 bool done = false;
1049 while(!done) {
1050 sectiontype_t type;
1051 int ret = read_section(&type, &ebl);
1052 if (ret)
1053 return ret;
1054
1055 //cout << "\tdo_object: Section type " << hex << type << dec << std::endl;
1056 //cout << "\t\tsection size " << ebl.length() << std::endl;
1057 if (type >= END_OF_TYPES) {
1058 cout << "Skipping unknown object section type" << std::endl;
1059 continue;
1060 }
1061 switch(type) {
1062 case TYPE_DATA:
1063 if (dry_run) break;
1064 ret = dump_data(formatter, ebl);
1065 if (ret) return ret;
1066 break;
1067 case TYPE_ATTRS:
1068 if (dry_run) break;
1069 ret = dump_attrs(formatter, ob.hoid, ebl);
1070 if (ret) return ret;
1071 break;
1072 case TYPE_OMAP_HDR:
1073 if (dry_run) break;
1074 ret = dump_omap_hdr(formatter, ebl);
1075 if (ret) return ret;
1076 break;
1077 case TYPE_OMAP:
1078 if (dry_run) break;
1079 ret = dump_omap(formatter, ebl);
1080 if (ret) return ret;
1081 break;
1082 case TYPE_OBJECT_END:
1083 done = true;
1084 break;
1085 default:
1086 cerr << "Unknown section type " << type << std::endl;
1087 return -EFAULT;
1088 }
1089 }
1090 formatter->close_section();
1091 return 0;
1092 }
1093
1094 int ObjectStoreTool::get_object(ObjectStore *store, coll_t coll,
1095 bufferlist &bl, OSDMap &curmap,
1096 bool *skipped_objects,
1097 ObjectStore::Sequencer &osr)
1098 {
1099 ObjectStore::Transaction tran;
1100 ObjectStore::Transaction *t = &tran;
1101 bufferlist::iterator ebliter = bl.begin();
1102 object_begin ob;
1103 ob.decode(ebliter);
1104 OSDriver driver(
1105 store,
1106 coll_t(),
1107 OSD::make_snapmapper_oid());
1108 spg_t pg;
1109 coll.is_pg_prefix(&pg);
1110 SnapMapper mapper(g_ceph_context, &driver, 0, 0, 0, pg.shard);
1111
1112 if (ob.hoid.hobj.is_temp()) {
1113 cerr << "ERROR: Export contains temporary object '" << ob.hoid << "'" << std::endl;
1114 return -EFAULT;
1115 }
1116 assert(g_ceph_context);
1117 if (ob.hoid.hobj.nspace != g_ceph_context->_conf->osd_hit_set_namespace) {
1118 object_t oid = ob.hoid.hobj.oid;
1119 object_locator_t loc(ob.hoid.hobj);
1120 pg_t raw_pgid = curmap.object_locator_to_pg(oid, loc);
1121 pg_t pgid = curmap.raw_pg_to_pg(raw_pgid);
1122
1123 spg_t coll_pgid;
1124 if (coll.is_pg(&coll_pgid) == false) {
1125 cerr << "INTERNAL ERROR: Bad collection during import" << std::endl;
1126 return -EFAULT;
1127 }
1128 if (coll_pgid.shard != ob.hoid.shard_id) {
1129 cerr << "INTERNAL ERROR: Importing shard " << coll_pgid.shard
1130 << " but object shard is " << ob.hoid.shard_id << std::endl;
1131 return -EFAULT;
1132 }
1133
1134 if (coll_pgid.pgid != pgid) {
1135 cerr << "Skipping object '" << ob.hoid << "' which belongs in pg " << pgid << std::endl;
1136 *skipped_objects = true;
1137 skip_object(bl);
1138 return 0;
1139 }
1140 }
1141
1142 if (!dry_run)
1143 t->touch(coll, ob.hoid);
1144
1145 cout << "Write " << ob.hoid << std::endl;
1146
1147 bufferlist ebl;
1148 bool done = false;
1149 while(!done) {
1150 sectiontype_t type;
1151 int ret = read_section(&type, &ebl);
1152 if (ret)
1153 return ret;
1154
1155 //cout << "\tdo_object: Section type " << hex << type << dec << std::endl;
1156 //cout << "\t\tsection size " << ebl.length() << std::endl;
1157 if (type >= END_OF_TYPES) {
1158 cout << "Skipping unknown object section type" << std::endl;
1159 continue;
1160 }
1161 switch(type) {
1162 case TYPE_DATA:
1163 if (dry_run) break;
1164 ret = get_data(store, coll, ob.hoid, t, ebl);
1165 if (ret) return ret;
1166 break;
1167 case TYPE_ATTRS:
1168 if (dry_run) break;
1169 ret = get_attrs(store, coll, ob.hoid, t, ebl, driver, mapper);
1170 if (ret) return ret;
1171 break;
1172 case TYPE_OMAP_HDR:
1173 if (dry_run) break;
1174 ret = get_omap_hdr(store, coll, ob.hoid, t, ebl);
1175 if (ret) return ret;
1176 break;
1177 case TYPE_OMAP:
1178 if (dry_run) break;
1179 ret = get_omap(store, coll, ob.hoid, t, ebl);
1180 if (ret) return ret;
1181 break;
1182 case TYPE_OBJECT_END:
1183 done = true;
1184 break;
1185 default:
1186 cerr << "Unknown section type " << type << std::endl;
1187 return -EFAULT;
1188 }
1189 }
1190 if (!dry_run)
1191 store->apply_transaction(&osr, std::move(*t));
1192 return 0;
1193 }
1194
1195 int dump_pg_metadata(Formatter *formatter, bufferlist &bl, metadata_section &ms)
1196 {
1197 bufferlist::iterator ebliter = bl.begin();
1198 ms.decode(ebliter);
1199
1200 formatter->open_object_section("metadata_section");
1201
1202 formatter->dump_unsigned("pg_disk_version", (int)ms.struct_ver);
1203 formatter->dump_unsigned("map_epoch", ms.map_epoch);
1204
1205 formatter->open_object_section("OSDMap");
1206 ms.osdmap.dump(formatter);
1207 formatter->close_section();
1208 formatter->flush(cout);
1209 cout << std::endl;
1210
1211 formatter->open_object_section("info");
1212 ms.info.dump(formatter);
1213 formatter->close_section();
1214 formatter->flush(cout);
1215
1216 formatter->open_object_section("log");
1217 ms.log.dump(formatter);
1218 formatter->close_section();
1219 formatter->flush(cout);
1220
1221 formatter->open_object_section("pg_missing_t");
1222 ms.missing.dump(formatter);
1223 formatter->close_section();
1224
1225 // XXX: ms.past_intervals?
1226
1227 formatter->close_section();
1228 formatter->flush(cout);
1229
1230 if (ms.osdmap.get_epoch() != 0 && ms.map_epoch != ms.osdmap.get_epoch()) {
1231 cerr << "FATAL: Invalid OSDMap epoch in export data" << std::endl;
1232 return -EFAULT;
1233 }
1234
1235 return 0;
1236 }
1237
1238 int get_pg_metadata(ObjectStore *store, bufferlist &bl, metadata_section &ms,
1239 const OSDSuperblock& sb, OSDMap& curmap, spg_t pgid)
1240 {
1241 bufferlist::iterator ebliter = bl.begin();
1242 ms.decode(ebliter);
1243 spg_t old_pgid = ms.info.pgid;
1244 ms.info.pgid = pgid;
1245
1246 #if DIAGNOSTIC
1247 Formatter *formatter = new JSONFormatter(true);
1248 cout << "export pgid " << old_pgid << std::endl;
1249 cout << "struct_v " << (int)ms.struct_ver << std::endl;
1250 cout << "map epoch " << ms.map_epoch << std::endl;
1251
1252 formatter->open_object_section("importing OSDMap");
1253 ms.osdmap.dump(formatter);
1254 formatter->close_section();
1255 formatter->flush(cout);
1256 cout << std::endl;
1257
1258 cout << "osd current epoch " << sb.current_epoch << std::endl;
1259 formatter->open_object_section("current OSDMap");
1260 curmap.dump(formatter);
1261 formatter->close_section();
1262 formatter->flush(cout);
1263 cout << std::endl;
1264
1265 formatter->open_object_section("info");
1266 ms.info.dump(formatter);
1267 formatter->close_section();
1268 formatter->flush(cout);
1269 cout << std::endl;
1270
1271 formatter->open_object_section("log");
1272 ms.log.dump(formatter);
1273 formatter->close_section();
1274 formatter->flush(cout);
1275 cout << std::endl;
1276
1277 formatter->flush(cout);
1278 cout << std::endl;
1279 #endif
1280
1281 if (ms.osdmap.get_epoch() != 0 && ms.map_epoch != ms.osdmap.get_epoch()) {
1282 cerr << "FATAL: Invalid OSDMap epoch in export data" << std::endl;
1283 return -EFAULT;
1284 }
1285
1286 if (ms.map_epoch > sb.current_epoch) {
1287 cerr << "ERROR: Export PG's map_epoch " << ms.map_epoch << " > OSD's epoch " << sb.current_epoch << std::endl;
1288 cerr << "The OSD you are using is older than the exported PG" << std::endl;
1289 cerr << "Either use another OSD or join selected OSD to cluster to update it first" << std::endl;
1290 return -EINVAL;
1291 }
1292
1293 // Pool verified to exist for call to get_pg_num().
1294 unsigned new_pg_num = curmap.get_pg_num(pgid.pgid.pool());
1295
1296 if (pgid.pgid.ps() >= new_pg_num) {
1297 cerr << "Illegal pgid, the seed is larger than current pg_num" << std::endl;
1298 return -EINVAL;
1299 }
1300
1301 // Old exports didn't include OSDMap, see if we have a copy locally
1302 if (ms.osdmap.get_epoch() == 0) {
1303 OSDMap findmap;
1304 bufferlist findmap_bl;
1305 int ret = get_osdmap(store, ms.map_epoch, findmap, findmap_bl);
1306 if (ret == 0) {
1307 ms.osdmap.deepish_copy_from(findmap);
1308 } else {
1309 cerr << "WARNING: No OSDMap in old export,"
1310 " some objects may be ignored due to a split" << std::endl;
1311 }
1312 }
1313
1314 // Make sure old_pg_num is 0 in the unusual case that OSDMap not in export
1315 // nor can we find a local copy.
1316 unsigned old_pg_num = 0;
1317 if (ms.osdmap.get_epoch() != 0)
1318 old_pg_num = ms.osdmap.get_pg_num(pgid.pgid.pool());
1319
1320 if (debug) {
1321 cerr << "old_pg_num " << old_pg_num << std::endl;
1322 cerr << "new_pg_num " << new_pg_num << std::endl;
1323 cerr << ms.osdmap << std::endl;
1324 cerr << curmap << std::endl;
1325 }
1326
1327 // If we have managed to have a good OSDMap we can do these checks
1328 if (old_pg_num) {
1329 if (old_pgid.pgid.ps() >= old_pg_num) {
1330 cerr << "FATAL: pgid invalid for original map epoch" << std::endl;
1331 return -EFAULT;
1332 }
1333 if (pgid.pgid.ps() >= old_pg_num) {
1334 cout << "NOTICE: Post split pgid specified" << std::endl;
1335 } else {
1336 spg_t parent(pgid);
1337 if (parent.is_split(old_pg_num, new_pg_num, NULL)) {
1338 cerr << "WARNING: Split occurred, some objects may be ignored" << std::endl;
1339 }
1340 }
1341 }
1342
1343 if (debug) {
1344 cerr << "Import pgid " << ms.info.pgid << std::endl;
1345 cerr << "Previous past_intervals " << ms.past_intervals << std::endl;
1346 cerr << "history.same_interval_since " << ms.info.history.same_interval_since << std::endl;
1347 }
1348
1349 // Let osd recompute past_intervals and same_interval_since
1350 ms.past_intervals.clear();
1351 ms.info.history.same_interval_since = 0;
1352
1353 if (debug)
1354 cerr << "Changing pg epoch " << ms.map_epoch << " to " << sb.current_epoch << std::endl;
1355
1356 ms.map_epoch = sb.current_epoch;
1357
1358 return 0;
1359 }
1360
1361 // out: pg_log_t that only has entries that apply to import_pgid using curmap
1362 // reject: Entries rejected from "in" are in the reject.log. Other fields not set.
1363 void filter_divergent_priors(spg_t import_pgid, const OSDMap &curmap,
1364 const string &hit_set_namespace, const divergent_priors_t &in,
1365 divergent_priors_t &out, divergent_priors_t &reject)
1366 {
1367 out.clear();
1368 reject.clear();
1369
1370 for (divergent_priors_t::const_iterator i = in.begin();
1371 i != in.end(); ++i) {
1372
1373 // Reject divergent priors for temporary objects
1374 if (i->second.is_temp()) {
1375 reject.insert(*i);
1376 continue;
1377 }
1378
1379 if (i->second.nspace != hit_set_namespace) {
1380 object_t oid = i->second.oid;
1381 object_locator_t loc(i->second);
1382 pg_t raw_pgid = curmap.object_locator_to_pg(oid, loc);
1383 pg_t pgid = curmap.raw_pg_to_pg(raw_pgid);
1384
1385 if (import_pgid.pgid == pgid) {
1386 out.insert(*i);
1387 } else {
1388 reject.insert(*i);
1389 }
1390 } else {
1391 out.insert(*i);
1392 }
1393 }
1394 }
1395
1396 int ObjectStoreTool::dump_import(Formatter *formatter)
1397 {
1398 bufferlist ebl;
1399 pg_info_t info;
1400 PGLog::IndexedLog log;
1401 //bool skipped_objects = false;
1402
1403 int ret = read_super();
1404 if (ret)
1405 return ret;
1406
1407 if (sh.magic != super_header::super_magic) {
1408 cerr << "Invalid magic number" << std::endl;
1409 return -EFAULT;
1410 }
1411
1412 if (sh.version > super_header::super_ver) {
1413 cerr << "Can't handle export format version=" << sh.version << std::endl;
1414 return -EINVAL;
1415 }
1416
1417 formatter->open_object_section("Export");
1418
1419 //First section must be TYPE_PG_BEGIN
1420 sectiontype_t type;
1421 ret = read_section(&type, &ebl);
1422 if (ret)
1423 return ret;
1424 if (type == TYPE_POOL_BEGIN) {
1425 cerr << "Dump of pool exports not supported" << std::endl;
1426 return -EINVAL;
1427 } else if (type != TYPE_PG_BEGIN) {
1428 cerr << "Invalid first section type " << std::to_string(type) << std::endl;
1429 return -EFAULT;
1430 }
1431
1432 bufferlist::iterator ebliter = ebl.begin();
1433 pg_begin pgb;
1434 pgb.decode(ebliter);
1435 spg_t pgid = pgb.pgid;
1436
1437 formatter->dump_string("pgid", stringify(pgid));
1438 formatter->dump_string("cluster_fsid", stringify(pgb.superblock.cluster_fsid));
1439 formatter->dump_string("features", stringify(pgb.superblock.compat_features));
1440
1441 bool done = false;
1442 bool found_metadata = false;
1443 metadata_section ms;
1444 bool objects_started = false;
1445 while(!done) {
1446 ret = read_section(&type, &ebl);
1447 if (ret)
1448 return ret;
1449
1450 if (debug) {
1451 cerr << "dump_import: Section type " << std::to_string(type) << std::endl;
1452 }
1453 if (type >= END_OF_TYPES) {
1454 cerr << "Skipping unknown section type" << std::endl;
1455 continue;
1456 }
1457 switch(type) {
1458 case TYPE_OBJECT_BEGIN:
1459 if (!objects_started) {
1460 formatter->open_array_section("objects");
1461 objects_started = true;
1462 }
1463 ret = dump_object(formatter, ebl);
1464 if (ret) return ret;
1465 break;
1466 case TYPE_PG_METADATA:
1467 if (objects_started)
1468 cerr << "WARNING: metadata_section out of order" << std::endl;
1469 ret = dump_pg_metadata(formatter, ebl, ms);
1470 if (ret) return ret;
1471 found_metadata = true;
1472 break;
1473 case TYPE_PG_END:
1474 if (objects_started) {
1475 formatter->close_section();
1476 }
1477 done = true;
1478 break;
1479 default:
1480 cerr << "Unknown section type " << std::to_string(type) << std::endl;
1481 return -EFAULT;
1482 }
1483 }
1484
1485 if (!found_metadata) {
1486 cerr << "Missing metadata section" << std::endl;
1487 return -EFAULT;
1488 }
1489
1490 formatter->close_section();
1491 formatter->flush(cout);
1492
1493 return 0;
1494 }
1495
1496 int ObjectStoreTool::do_import(ObjectStore *store, OSDSuperblock& sb,
1497 bool force, std::string pgidstr,
1498 ObjectStore::Sequencer &osr)
1499 {
1500 bufferlist ebl;
1501 pg_info_t info;
1502 PGLog::IndexedLog log;
1503 bool skipped_objects = false;
1504
1505 if (!dry_run)
1506 finish_remove_pgs(store);
1507
1508 int ret = read_super();
1509 if (ret)
1510 return ret;
1511
1512 if (sh.magic != super_header::super_magic) {
1513 cerr << "Invalid magic number" << std::endl;
1514 return -EFAULT;
1515 }
1516
1517 if (sh.version > super_header::super_ver) {
1518 cerr << "Can't handle export format version=" << sh.version << std::endl;
1519 return -EINVAL;
1520 }
1521
1522 //First section must be TYPE_PG_BEGIN
1523 sectiontype_t type;
1524 ret = read_section(&type, &ebl);
1525 if (ret)
1526 return ret;
1527 if (type == TYPE_POOL_BEGIN) {
1528 cerr << "Pool exports cannot be imported into a PG" << std::endl;
1529 return -EINVAL;
1530 } else if (type != TYPE_PG_BEGIN) {
1531 cerr << "Invalid first section type " << std::to_string(type) << std::endl;
1532 return -EFAULT;
1533 }
1534
1535 bufferlist::iterator ebliter = ebl.begin();
1536 pg_begin pgb;
1537 pgb.decode(ebliter);
1538 spg_t pgid = pgb.pgid;
1539 spg_t orig_pgid = pgid;
1540
1541 if (pgidstr.length()) {
1542 spg_t user_pgid;
1543
1544 bool ok = user_pgid.parse(pgidstr.c_str());
1545 // This succeeded in main() already
1546 assert(ok);
1547 if (pgid != user_pgid) {
1548 if (pgid.pool() != user_pgid.pool()) {
1549 cerr << "Can't specify a different pgid pool, must be " << pgid.pool() << std::endl;
1550 return -EINVAL;
1551 }
1552 if (pgid.is_no_shard() && !user_pgid.is_no_shard()) {
1553 cerr << "Can't specify a sharded pgid with a non-sharded export" << std::endl;
1554 return -EINVAL;
1555 }
1556 // Get shard from export information if not specified
1557 if (!pgid.is_no_shard() && user_pgid.is_no_shard()) {
1558 user_pgid.shard = pgid.shard;
1559 }
1560 if (pgid.shard != user_pgid.shard) {
1561 cerr << "Can't specify a different shard, must be " << pgid.shard << std::endl;
1562 return -EINVAL;
1563 }
1564 pgid = user_pgid;
1565 }
1566 }
1567
1568 if (!pgb.superblock.cluster_fsid.is_zero()
1569 && pgb.superblock.cluster_fsid != sb.cluster_fsid) {
1570 cerr << "Export came from different cluster with fsid "
1571 << pgb.superblock.cluster_fsid << std::endl;
1572 return -EINVAL;
1573 }
1574
1575 if (debug) {
1576 cerr << "Exported features: " << pgb.superblock.compat_features << std::endl;
1577 }
1578
1579 // Special case: Old export has SHARDS incompat feature on replicated pg, remove it
1580 if (pgid.is_no_shard())
1581 pgb.superblock.compat_features.incompat.remove(CEPH_OSD_FEATURE_INCOMPAT_SHARDS);
1582
1583 if (sb.compat_features.compare(pgb.superblock.compat_features) == -1) {
1584 CompatSet unsupported = sb.compat_features.unsupported(pgb.superblock.compat_features);
1585
1586 cerr << "Export has incompatible features set " << unsupported << std::endl;
1587
1588 // Let them import if they specify the --force option
1589 if (!force)
1590 return 11; // Positive return means exit status
1591 }
1592
1593 // Don't import if pool no longer exists
1594 OSDMap curmap;
1595 bufferlist bl;
1596 ret = get_osdmap(store, sb.current_epoch, curmap, bl);
1597 if (ret) {
1598 cerr << "Can't find local OSDMap" << std::endl;
1599 return ret;
1600 }
1601 if (!curmap.have_pg_pool(pgid.pgid.m_pool)) {
1602 cerr << "Pool " << pgid.pgid.m_pool << " no longer exists" << std::endl;
1603 // Special exit code for this error, used by test code
1604 return 10; // Positive return means exit status
1605 }
1606
1607 ghobject_t pgmeta_oid = pgid.make_pgmeta_oid();
1608 log_oid = OSD::make_pg_log_oid(pgid);
1609 biginfo_oid = OSD::make_pg_biginfo_oid(pgid);
1610
1611 //Check for PG already present.
1612 coll_t coll(pgid);
1613 if (store->collection_exists(coll)) {
1614 cerr << "pgid " << pgid << " already exists" << std::endl;
1615 return -EEXIST;
1616 }
1617
1618 if (!dry_run) {
1619 ObjectStore::Transaction t;
1620 PG::_create(t, pgid,
1621 pgid.get_split_bits(curmap.get_pg_pool(pgid.pool())->get_pg_num()));
1622 PG::_init(t, pgid, NULL);
1623
1624 // mark this coll for removal until we're done
1625 map<string,bufferlist> values;
1626 ::encode((char)1, values["_remove"]);
1627 t.omap_setkeys(coll, pgid.make_pgmeta_oid(), values);
1628
1629 store->apply_transaction(&osr, std::move(t));
1630 }
1631
1632 cout << "Importing pgid " << pgid;
1633 if (orig_pgid != pgid) {
1634 cout << " exported as " << orig_pgid;
1635 }
1636 cout << std::endl;
1637
1638 bool done = false;
1639 bool found_metadata = false;
1640 metadata_section ms;
1641 while(!done) {
1642 ret = read_section(&type, &ebl);
1643 if (ret)
1644 return ret;
1645
1646 if (debug) {
1647 cout << __func__ << ": Section type " << std::to_string(type) << std::endl;
1648 }
1649 if (type >= END_OF_TYPES) {
1650 cout << "Skipping unknown section type" << std::endl;
1651 continue;
1652 }
1653 switch(type) {
1654 case TYPE_OBJECT_BEGIN:
1655 ret = get_object(store, coll, ebl, curmap, &skipped_objects, osr);
1656 if (ret) return ret;
1657 break;
1658 case TYPE_PG_METADATA:
1659 ret = get_pg_metadata(store, ebl, ms, sb, curmap, pgid);
1660 if (ret) return ret;
1661 found_metadata = true;
1662 break;
1663 case TYPE_PG_END:
1664 done = true;
1665 break;
1666 default:
1667 cerr << "Unknown section type " << std::to_string(type) << std::endl;
1668 return -EFAULT;
1669 }
1670 }
1671
1672 if (!found_metadata) {
1673 cerr << "Missing metadata section" << std::endl;
1674 return -EFAULT;
1675 }
1676
1677 ObjectStore::Transaction t;
1678 if (!dry_run) {
1679 pg_log_t newlog, reject;
1680 pg_log_t::filter_log(pgid, curmap, g_ceph_context->_conf->osd_hit_set_namespace,
1681 ms.log, newlog, reject);
1682 if (debug) {
1683 for (list<pg_log_entry_t>::iterator i = newlog.log.begin();
1684 i != newlog.log.end(); ++i)
1685 cerr << "Keeping log entry " << *i << std::endl;
1686 for (list<pg_log_entry_t>::iterator i = reject.log.begin();
1687 i != reject.log.end(); ++i)
1688 cerr << "Skipping log entry " << *i << std::endl;
1689 }
1690
1691 divergent_priors_t newdp, rejectdp;
1692 filter_divergent_priors(pgid, curmap, g_ceph_context->_conf->osd_hit_set_namespace,
1693 ms.divergent_priors, newdp, rejectdp);
1694 ms.divergent_priors = newdp;
1695 if (debug) {
1696 for (divergent_priors_t::iterator i = newdp.begin();
1697 i != newdp.end(); ++i)
1698 cerr << "Keeping divergent_prior " << *i << std::endl;
1699 for (divergent_priors_t::iterator i = rejectdp.begin();
1700 i != rejectdp.end(); ++i)
1701 cerr << "Skipping divergent_prior " << *i << std::endl;
1702 }
1703
1704 ms.missing.filter_objects([&](const hobject_t &obj) {
1705 if (obj.nspace == g_ceph_context->_conf->osd_hit_set_namespace)
1706 return false;
1707 assert(!obj.is_temp());
1708 object_t oid = obj.oid;
1709 object_locator_t loc(obj);
1710 pg_t raw_pgid = curmap.object_locator_to_pg(oid, loc);
1711 pg_t _pgid = curmap.raw_pg_to_pg(raw_pgid);
1712
1713 return pgid.pgid != _pgid;
1714 });
1715
1716
1717 if (debug) {
1718 pg_missing_t missing;
1719 Formatter *formatter = Formatter::create("json-pretty");
1720 dump_log(formatter, cerr, newlog, ms.missing);
1721 delete formatter;
1722 }
1723
1724 // Just like a split invalidate stats since the object count is changed
1725 if (skipped_objects)
1726 ms.info.stats.stats_invalid = true;
1727
1728 ret = write_pg(
1729 t,
1730 ms.map_epoch,
1731 ms.info,
1732 newlog,
1733 ms.past_intervals,
1734 ms.divergent_priors,
1735 ms.missing);
1736 if (ret) return ret;
1737 }
1738
1739 // done, clear removal flag
1740 if (debug)
1741 cerr << "done, clearing removal flag" << std::endl;
1742
1743 if (!dry_run) {
1744 set<string> remove;
1745 remove.insert("_remove");
1746 t.omap_rmkeys(coll, pgid.make_pgmeta_oid(), remove);
1747 store->apply_transaction(&osr, std::move(t));
1748 }
1749
1750 return 0;
1751 }
1752
1753 int do_list(ObjectStore *store, string pgidstr, string object, boost::optional<std::string> nspace,
1754 Formatter *formatter, bool debug, bool human_readable, bool head)
1755 {
1756 int r;
1757 lookup_ghobject lookup(object, nspace, head);
1758 if (pgidstr.length() > 0) {
1759 r = action_on_all_objects_in_pg(store, pgidstr, lookup, debug);
1760 } else {
1761 r = action_on_all_objects(store, lookup, debug);
1762 }
1763 if (r)
1764 return r;
1765 lookup.dump(formatter, human_readable);
1766 formatter->flush(cout);
1767 return 0;
1768 }
1769
1770 int do_meta(ObjectStore *store, string object, Formatter *formatter, bool debug, bool human_readable)
1771 {
1772 int r;
1773 boost::optional<std::string> nspace; // Not specified
1774 lookup_ghobject lookup(object, nspace);
1775 r = action_on_all_objects_in_exact_pg(store, coll_t::meta(), lookup, debug);
1776 if (r)
1777 return r;
1778 lookup.dump(formatter, human_readable);
1779 formatter->flush(cout);
1780 return 0;
1781 }
1782
1783 int remove_object(coll_t coll, ghobject_t &ghobj,
1784 SnapMapper &mapper,
1785 MapCacher::Transaction<std::string, bufferlist> *_t,
1786 ObjectStore::Transaction *t)
1787 {
1788 int r = mapper.remove_oid(ghobj.hobj, _t);
1789 if (r < 0 && r != -ENOENT) {
1790 cerr << "remove_oid returned " << cpp_strerror(r) << std::endl;
1791 return r;
1792 }
1793
1794 t->remove(coll, ghobj);
1795 return 0;
1796 }
1797
1798 int get_snapset(ObjectStore *store, coll_t coll, ghobject_t &ghobj, SnapSet &ss, bool silent);
1799
1800 int do_remove_object(ObjectStore *store, coll_t coll,
1801 ghobject_t &ghobj, bool all, bool force,
1802 ObjectStore::Sequencer &osr)
1803 {
1804 spg_t pg;
1805 coll.is_pg_prefix(&pg);
1806 OSDriver driver(
1807 store,
1808 coll_t(),
1809 OSD::make_snapmapper_oid());
1810 SnapMapper mapper(g_ceph_context, &driver, 0, 0, 0, pg.shard);
1811 struct stat st;
1812
1813 int r = store->stat(coll, ghobj, &st);
1814 if (r < 0) {
1815 cerr << "remove: " << cpp_strerror(r) << std::endl;
1816 return r;
1817 }
1818
1819 SnapSet ss;
1820 if (ghobj.hobj.has_snapset()) {
1821 r = get_snapset(store, coll, ghobj, ss, false);
1822 if (r < 0) {
1823 cerr << "Can't get snapset error " << cpp_strerror(r) << std::endl;
1824 return r;
1825 }
1826 if (!ss.snaps.empty() && !all) {
1827 if (force) {
1828 cout << "WARNING: only removing "
1829 << (ghobj.hobj.is_head() ? "head" : "snapdir")
1830 << " with snapshots present" << std::endl;
1831 ss.snaps.clear();
1832 } else {
1833 cerr << "Snapshots are present, use removeall to delete everything" << std::endl;
1834 return -EINVAL;
1835 }
1836 }
1837 }
1838
1839 ObjectStore::Transaction t;
1840 OSDriver::OSTransaction _t(driver.get_transaction(&t));
1841
1842 cout << "remove " << ghobj << std::endl;
1843
1844 if (!dry_run) {
1845 r = remove_object(coll, ghobj, mapper, &_t, &t);
1846 if (r < 0)
1847 return r;
1848 }
1849
1850 ghobject_t snapobj = ghobj;
1851 for (vector<snapid_t>::iterator i = ss.snaps.begin() ;
1852 i != ss.snaps.end() ; ++i) {
1853 snapobj.hobj.snap = *i;
1854 cout << "remove " << snapobj << std::endl;
1855 if (!dry_run) {
1856 r = remove_object(coll, snapobj, mapper, &_t, &t);
1857 if (r < 0)
1858 return r;
1859 }
1860 }
1861
1862 if (!dry_run)
1863 store->apply_transaction(&osr, std::move(t));
1864
1865 return 0;
1866 }
1867
1868 int do_list_attrs(ObjectStore *store, coll_t coll, ghobject_t &ghobj)
1869 {
1870 map<string,bufferptr> aset;
1871 int r = store->getattrs(coll, ghobj, aset);
1872 if (r < 0) {
1873 cerr << "getattrs: " << cpp_strerror(r) << std::endl;
1874 return r;
1875 }
1876
1877 for (map<string,bufferptr>::iterator i = aset.begin();i != aset.end(); ++i) {
1878 string key(i->first);
1879 if (outistty)
1880 key = cleanbin(key);
1881 cout << key << std::endl;
1882 }
1883 return 0;
1884 }
1885
1886 int do_list_omap(ObjectStore *store, coll_t coll, ghobject_t &ghobj)
1887 {
1888 ObjectMap::ObjectMapIterator iter = store->get_omap_iterator(coll, ghobj);
1889 if (!iter) {
1890 cerr << "omap_get_iterator: " << cpp_strerror(ENOENT) << std::endl;
1891 return -ENOENT;
1892 }
1893 iter->seek_to_first();
1894 map<string, bufferlist> oset;
1895 while(iter->valid()) {
1896 get_omap_batch(iter, oset);
1897
1898 for (map<string,bufferlist>::iterator i = oset.begin();i != oset.end(); ++i) {
1899 string key(i->first);
1900 if (outistty)
1901 key = cleanbin(key);
1902 cout << key << std::endl;
1903 }
1904 }
1905 return 0;
1906 }
1907
1908 int do_get_bytes(ObjectStore *store, coll_t coll, ghobject_t &ghobj, int fd)
1909 {
1910 struct stat st;
1911 mysize_t total;
1912
1913 int ret = store->stat(coll, ghobj, &st);
1914 if (ret < 0) {
1915 cerr << "get-bytes: " << cpp_strerror(ret) << std::endl;
1916 return ret;
1917 }
1918
1919 total = st.st_size;
1920 if (debug)
1921 cerr << "size=" << total << std::endl;
1922
1923 uint64_t offset = 0;
1924 bufferlist rawdatabl;
1925 while(total > 0) {
1926 rawdatabl.clear();
1927 mysize_t len = max_read;
1928 if (len > total)
1929 len = total;
1930
1931 ret = store->read(coll, ghobj, offset, len, rawdatabl);
1932 if (ret < 0)
1933 return ret;
1934 if (ret == 0)
1935 return -EINVAL;
1936
1937 if (debug)
1938 cerr << "data section offset=" << offset << " len=" << len << std::endl;
1939
1940 total -= ret;
1941 offset += ret;
1942
1943 ret = write(fd, rawdatabl.c_str(), ret);
1944 if (ret == -1) {
1945 perror("write");
1946 return -errno;
1947 }
1948 }
1949
1950 return 0;
1951 }
1952
1953 int do_set_bytes(ObjectStore *store, coll_t coll,
1954 ghobject_t &ghobj, int fd,
1955 ObjectStore::Sequencer &osr)
1956 {
1957 ObjectStore::Transaction tran;
1958 ObjectStore::Transaction *t = &tran;
1959
1960 if (debug)
1961 cerr << "Write " << ghobj << std::endl;
1962
1963 if (!dry_run) {
1964 t->touch(coll, ghobj);
1965 t->truncate(coll, ghobj, 0);
1966 }
1967
1968 uint64_t offset = 0;
1969 bufferlist rawdatabl;
1970 do {
1971 rawdatabl.clear();
1972 ssize_t bytes = rawdatabl.read_fd(fd, max_read);
1973 if (bytes < 0) {
1974 cerr << "read_fd error " << cpp_strerror(bytes) << std::endl;
1975 return bytes;
1976 }
1977
1978 if (bytes == 0)
1979 break;
1980
1981 if (debug)
1982 cerr << "\tdata: offset " << offset << " bytes " << bytes << std::endl;
1983 if (!dry_run)
1984 t->write(coll, ghobj, offset, bytes, rawdatabl);
1985
1986 offset += bytes;
1987 // XXX: Should we apply_transaction() every once in a while for very large files
1988 } while(true);
1989
1990 if (!dry_run)
1991 store->apply_transaction(&osr, std::move(*t));
1992 return 0;
1993 }
1994
1995 int do_get_attr(ObjectStore *store, coll_t coll, ghobject_t &ghobj, string key)
1996 {
1997 bufferptr bp;
1998
1999 int r = store->getattr(coll, ghobj, key.c_str(), bp);
2000 if (r < 0) {
2001 cerr << "getattr: " << cpp_strerror(r) << std::endl;
2002 return r;
2003 }
2004
2005 string value(bp.c_str(), bp.length());
2006 if (outistty) {
2007 value = cleanbin(value);
2008 value.push_back('\n');
2009 }
2010 cout << value;
2011
2012 return 0;
2013 }
2014
2015 int do_set_attr(ObjectStore *store, coll_t coll,
2016 ghobject_t &ghobj, string key, int fd,
2017 ObjectStore::Sequencer &osr)
2018 {
2019 ObjectStore::Transaction tran;
2020 ObjectStore::Transaction *t = &tran;
2021 bufferlist bl;
2022
2023 if (debug)
2024 cerr << "Setattr " << ghobj << std::endl;
2025
2026 int ret = get_fd_data(fd, bl);
2027 if (ret < 0)
2028 return ret;
2029
2030 if (dry_run)
2031 return 0;
2032
2033 t->touch(coll, ghobj);
2034
2035 t->setattr(coll, ghobj, key, bl);
2036
2037 store->apply_transaction(&osr, std::move(*t));
2038 return 0;
2039 }
2040
2041 int do_rm_attr(ObjectStore *store, coll_t coll,
2042 ghobject_t &ghobj, string key,
2043 ObjectStore::Sequencer &osr)
2044 {
2045 ObjectStore::Transaction tran;
2046 ObjectStore::Transaction *t = &tran;
2047
2048 if (debug)
2049 cerr << "Rmattr " << ghobj << std::endl;
2050
2051 if (dry_run)
2052 return 0;
2053
2054 t->rmattr(coll, ghobj, key);
2055
2056 store->apply_transaction(&osr, std::move(*t));
2057 return 0;
2058 }
2059
2060 int do_get_omap(ObjectStore *store, coll_t coll, ghobject_t &ghobj, string key)
2061 {
2062 set<string> keys;
2063 map<string, bufferlist> out;
2064
2065 keys.insert(key);
2066
2067 int r = store->omap_get_values(coll, ghobj, keys, &out);
2068 if (r < 0) {
2069 cerr << "omap_get_values: " << cpp_strerror(r) << std::endl;
2070 return r;
2071 }
2072
2073 if (out.empty()) {
2074 cerr << "Key not found" << std::endl;
2075 return -ENOENT;
2076 }
2077
2078 assert(out.size() == 1);
2079
2080 bufferlist bl = out.begin()->second;
2081 string value(bl.c_str(), bl.length());
2082 if (outistty) {
2083 value = cleanbin(value);
2084 value.push_back('\n');
2085 }
2086 cout << value;
2087
2088 return 0;
2089 }
2090
2091 int do_set_omap(ObjectStore *store, coll_t coll,
2092 ghobject_t &ghobj, string key, int fd,
2093 ObjectStore::Sequencer &osr)
2094 {
2095 ObjectStore::Transaction tran;
2096 ObjectStore::Transaction *t = &tran;
2097 map<string, bufferlist> attrset;
2098 bufferlist valbl;
2099
2100 if (debug)
2101 cerr << "Set_omap " << ghobj << std::endl;
2102
2103 int ret = get_fd_data(fd, valbl);
2104 if (ret < 0)
2105 return ret;
2106
2107 attrset.insert(pair<string, bufferlist>(key, valbl));
2108
2109 if (dry_run)
2110 return 0;
2111
2112 t->touch(coll, ghobj);
2113
2114 t->omap_setkeys(coll, ghobj, attrset);
2115
2116 store->apply_transaction(&osr, std::move(*t));
2117 return 0;
2118 }
2119
2120 int do_rm_omap(ObjectStore *store, coll_t coll,
2121 ghobject_t &ghobj, string key,
2122 ObjectStore::Sequencer &osr)
2123 {
2124 ObjectStore::Transaction tran;
2125 ObjectStore::Transaction *t = &tran;
2126 set<string> keys;
2127
2128 keys.insert(key);
2129
2130 if (debug)
2131 cerr << "Rm_omap " << ghobj << std::endl;
2132
2133 if (dry_run)
2134 return 0;
2135
2136 t->omap_rmkeys(coll, ghobj, keys);
2137
2138 store->apply_transaction(&osr, std::move(*t));
2139 return 0;
2140 }
2141
2142 int do_get_omaphdr(ObjectStore *store, coll_t coll, ghobject_t &ghobj)
2143 {
2144 bufferlist hdrbl;
2145
2146 int r = store->omap_get_header(coll, ghobj, &hdrbl, true);
2147 if (r < 0) {
2148 cerr << "omap_get_header: " << cpp_strerror(r) << std::endl;
2149 return r;
2150 }
2151
2152 string header(hdrbl.c_str(), hdrbl.length());
2153 if (outistty) {
2154 header = cleanbin(header);
2155 header.push_back('\n');
2156 }
2157 cout << header;
2158
2159 return 0;
2160 }
2161
2162 int do_set_omaphdr(ObjectStore *store, coll_t coll,
2163 ghobject_t &ghobj, int fd,
2164 ObjectStore::Sequencer &osr)
2165 {
2166 ObjectStore::Transaction tran;
2167 ObjectStore::Transaction *t = &tran;
2168 bufferlist hdrbl;
2169
2170 if (debug)
2171 cerr << "Omap_setheader " << ghobj << std::endl;
2172
2173 int ret = get_fd_data(fd, hdrbl);
2174 if (ret)
2175 return ret;
2176
2177 if (dry_run)
2178 return 0;
2179
2180 t->touch(coll, ghobj);
2181
2182 t->omap_setheader(coll, ghobj, hdrbl);
2183
2184 store->apply_transaction(&osr, std::move(*t));
2185 return 0;
2186 }
2187
2188 struct do_fix_lost : public action_on_object_t {
2189 ObjectStore::Sequencer *osr;
2190
2191 explicit do_fix_lost(ObjectStore::Sequencer *_osr) : osr(_osr) {}
2192
2193 int call(ObjectStore *store, coll_t coll,
2194 ghobject_t &ghobj, object_info_t &oi) override {
2195 if (oi.is_lost()) {
2196 cout << coll << "/" << ghobj << " is lost";
2197 if (!dry_run)
2198 cout << ", fixing";
2199 cout << std::endl;
2200 if (dry_run)
2201 return 0;
2202 oi.clear_flag(object_info_t::FLAG_LOST);
2203 bufferlist bl;
2204 ::encode(oi, bl, -1); /* fixme: using full features */
2205 ObjectStore::Transaction t;
2206 t.setattr(coll, ghobj, OI_ATTR, bl);
2207 int r = store->apply_transaction(osr, std::move(t));
2208 if (r < 0) {
2209 cerr << "Error getting fixing attr on : " << make_pair(coll, ghobj)
2210 << ", "
2211 << cpp_strerror(r) << std::endl;
2212 return r;
2213 }
2214 }
2215 return 0;
2216 }
2217 };
2218
2219 int get_snapset(ObjectStore *store, coll_t coll, ghobject_t &ghobj, SnapSet &ss, bool silent = false)
2220 {
2221 bufferlist attr;
2222 int r = store->getattr(coll, ghobj, SS_ATTR, attr);
2223 if (r < 0) {
2224 if (!silent)
2225 cerr << "Error getting snapset on : " << make_pair(coll, ghobj) << ", "
2226 << cpp_strerror(r) << std::endl;
2227 return r;
2228 }
2229 bufferlist::iterator bp = attr.begin();
2230 try {
2231 ::decode(ss, bp);
2232 } catch (...) {
2233 r = -EINVAL;
2234 cerr << "Error decoding snapset on : " << make_pair(coll, ghobj) << ", "
2235 << cpp_strerror(r) << std::endl;
2236 return r;
2237 }
2238 return 0;
2239 }
2240
2241 int print_obj_info(ObjectStore *store, coll_t coll, ghobject_t &ghobj, Formatter* formatter)
2242 {
2243 int r = 0;
2244 formatter->open_object_section("obj");
2245 formatter->open_object_section("id");
2246 ghobj.dump(formatter);
2247 formatter->close_section();
2248
2249 bufferlist attr;
2250 int gr = store->getattr(coll, ghobj, OI_ATTR, attr);
2251 if (gr < 0) {
2252 r = gr;
2253 cerr << "Error getting attr on : " << make_pair(coll, ghobj) << ", "
2254 << cpp_strerror(r) << std::endl;
2255 } else {
2256 object_info_t oi;
2257 bufferlist::iterator bp = attr.begin();
2258 try {
2259 ::decode(oi, bp);
2260 formatter->open_object_section("info");
2261 oi.dump(formatter);
2262 formatter->close_section();
2263 } catch (...) {
2264 r = -EINVAL;
2265 cerr << "Error decoding attr on : " << make_pair(coll, ghobj) << ", "
2266 << cpp_strerror(r) << std::endl;
2267 }
2268 }
2269 struct stat st;
2270 int sr = store->stat(coll, ghobj, &st, true);
2271 if (sr < 0) {
2272 r = sr;
2273 cerr << "Error stat on : " << make_pair(coll, ghobj) << ", "
2274 << cpp_strerror(r) << std::endl;
2275 } else {
2276 formatter->open_object_section("stat");
2277 formatter->dump_int("size", st.st_size);
2278 formatter->dump_int("blksize", st.st_blksize);
2279 formatter->dump_int("blocks", st.st_blocks);
2280 formatter->dump_int("nlink", st.st_nlink);
2281 formatter->close_section();
2282 }
2283
2284 if (ghobj.hobj.has_snapset()) {
2285 SnapSet ss;
2286 int snr = get_snapset(store, coll, ghobj, ss);
2287 if (snr < 0) {
2288 r = snr;
2289 } else {
2290 formatter->open_object_section("SnapSet");
2291 ss.dump(formatter);
2292 formatter->close_section();
2293 }
2294 }
2295 formatter->close_section();
2296 formatter->flush(cout);
2297 cout << std::endl;
2298 return r;
2299 }
2300
2301 int set_size(ObjectStore *store, coll_t coll, ghobject_t &ghobj, uint64_t setsize, Formatter* formatter,
2302 ObjectStore::Sequencer &osr, bool corrupt)
2303 {
2304 if (ghobj.hobj.is_snapdir()) {
2305 cerr << "Can't set the size of a snapdir" << std::endl;
2306 return -EINVAL;
2307 }
2308 bufferlist attr;
2309 int r = store->getattr(coll, ghobj, OI_ATTR, attr);
2310 if (r < 0) {
2311 cerr << "Error getting attr on : " << make_pair(coll, ghobj) << ", "
2312 << cpp_strerror(r) << std::endl;
2313 return r;
2314 }
2315 object_info_t oi;
2316 bufferlist::iterator bp = attr.begin();
2317 try {
2318 ::decode(oi, bp);
2319 } catch (...) {
2320 r = -EINVAL;
2321 cerr << "Error getting attr on : " << make_pair(coll, ghobj) << ", "
2322 << cpp_strerror(r) << std::endl;
2323 return r;
2324 }
2325 struct stat st;
2326 r = store->stat(coll, ghobj, &st, true);
2327 if (r < 0) {
2328 cerr << "Error stat on : " << make_pair(coll, ghobj) << ", "
2329 << cpp_strerror(r) << std::endl;
2330 }
2331 ghobject_t head(ghobj);
2332 SnapSet ss;
2333 bool found_head = true;
2334 map<snapid_t, uint64_t>::iterator csi;
2335 bool is_snap = ghobj.hobj.is_snap();
2336 if (is_snap) {
2337 head.hobj = head.hobj.get_head();
2338 r = get_snapset(store, coll, head, ss, true);
2339 if (r < 0 && r != -ENOENT) {
2340 // Requested get_snapset() silent, so if not -ENOENT show error
2341 cerr << "Error getting snapset on : " << make_pair(coll, head) << ", "
2342 << cpp_strerror(r) << std::endl;
2343 return r;
2344 }
2345 if (r == -ENOENT) {
2346 head.hobj = head.hobj.get_snapdir();
2347 r = get_snapset(store, coll, head, ss);
2348 if (r < 0)
2349 return r;
2350 found_head = false;
2351 } else {
2352 found_head = true;
2353 }
2354 csi = ss.clone_size.find(ghobj.hobj.snap);
2355 if (csi == ss.clone_size.end()) {
2356 cerr << "SnapSet is missing clone_size for snap " << ghobj.hobj.snap << std::endl;
2357 return -EINVAL;
2358 }
2359 }
2360 if ((uint64_t)st.st_size == setsize && oi.size == setsize
2361 && (!is_snap || csi->second == setsize)) {
2362 cout << "Size of object is already " << setsize << std::endl;
2363 return 0;
2364 }
2365 cout << "Setting size to " << setsize << ", stat size " << st.st_size
2366 << ", obj info size " << oi.size;
2367 if (is_snap) {
2368 cout << ", " << (found_head ? "head" : "snapdir")
2369 << " clone_size " << csi->second;
2370 csi->second = setsize;
2371 }
2372 cout << std::endl;
2373 if (!dry_run) {
2374 attr.clear();
2375 oi.size = setsize;
2376 ObjectStore::Transaction t;
2377 // Only modify object info if we want to corrupt it
2378 if (!corrupt && (uint64_t)st.st_size != setsize) {
2379 t.truncate(coll, ghobj, setsize);
2380 // Changing objectstore size will invalidate data_digest, so clear it.
2381 oi.clear_data_digest();
2382 }
2383 ::encode(oi, attr, -1); /* fixme: using full features */
2384 t.setattr(coll, ghobj, OI_ATTR, attr);
2385 if (is_snap) {
2386 bufferlist snapattr;
2387 snapattr.clear();
2388 ::encode(ss, snapattr);
2389 t.setattr(coll, head, SS_ATTR, snapattr);
2390 }
2391 r = store->apply_transaction(&osr, std::move(t));
2392 if (r < 0) {
2393 cerr << "Error writing object info: " << make_pair(coll, ghobj) << ", "
2394 << cpp_strerror(r) << std::endl;
2395 return r;
2396 }
2397 }
2398 return 0;
2399 }
2400
2401 int clear_snapset(ObjectStore *store, coll_t coll, ghobject_t &ghobj,
2402 string arg, ObjectStore::Sequencer &osr)
2403 {
2404 SnapSet ss;
2405 int ret = get_snapset(store, coll, ghobj, ss);
2406 if (ret < 0)
2407 return ret;
2408
2409 // Use "head" to set head_exists incorrectly
2410 if (arg == "corrupt" || arg == "head")
2411 ss.head_exists = !ghobj.hobj.is_head();
2412 else if (ss.head_exists != ghobj.hobj.is_head()) {
2413 cerr << "Correcting head_exists, set to "
2414 << (ghobj.hobj.is_head() ? "true" : "false") << std::endl;
2415 ss.head_exists = ghobj.hobj.is_head();
2416 }
2417 // Use "corrupt" to clear entire SnapSet
2418 // Use "seq" to just corrupt SnapSet.seq
2419 if (arg == "corrupt" || arg == "seq")
2420 ss.seq = 0;
2421 // Use "snaps" to just clear SnapSet.snaps
2422 if (arg == "corrupt" || arg == "snaps")
2423 ss.snaps.clear();
2424 // By default just clear clone, clone_overlap and clone_size
2425 if (arg == "corrupt")
2426 arg = "";
2427 if (arg == "" || arg == "clones")
2428 ss.clones.clear();
2429 if (arg == "" || arg == "clone_overlap")
2430 ss.clone_overlap.clear();
2431 if (arg == "" || arg == "clone_size")
2432 ss.clone_size.clear();
2433 // Break all clone sizes by adding 1
2434 if (arg == "size") {
2435 for (map<snapid_t, uint64_t>::iterator i = ss.clone_size.begin();
2436 i != ss.clone_size.end(); ++i)
2437 ++(i->second);
2438 }
2439
2440 if (!dry_run) {
2441 bufferlist bl;
2442 ::encode(ss, bl);
2443 ObjectStore::Transaction t;
2444 t.setattr(coll, ghobj, SS_ATTR, bl);
2445 int r = store->apply_transaction(&osr, std::move(t));
2446 if (r < 0) {
2447 cerr << "Error setting snapset on : " << make_pair(coll, ghobj) << ", "
2448 << cpp_strerror(r) << std::endl;
2449 return r;
2450 }
2451 }
2452 return 0;
2453 }
2454
2455 vector<snapid_t>::iterator find(vector<snapid_t> &v, snapid_t clid)
2456 {
2457 return std::find(v.begin(), v.end(), clid);
2458 }
2459
2460 map<snapid_t, interval_set<uint64_t> >::iterator
2461 find(map<snapid_t, interval_set<uint64_t> > &m, snapid_t clid)
2462 {
2463 return m.find(clid);
2464 }
2465
2466 map<snapid_t, uint64_t>::iterator find(map<snapid_t, uint64_t> &m,
2467 snapid_t clid)
2468 {
2469 return m.find(clid);
2470 }
2471
2472 template<class T>
2473 int remove_from(T &mv, string name, snapid_t cloneid, bool force)
2474 {
2475 typename T::iterator i = find(mv, cloneid);
2476 if (i != mv.end()) {
2477 mv.erase(i);
2478 } else {
2479 cerr << "Clone " << cloneid << " doesn't exist in " << name;
2480 if (force) {
2481 cerr << " (ignored)" << std::endl;
2482 return 0;
2483 }
2484 cerr << std::endl;
2485 return -EINVAL;
2486 }
2487 return 0;
2488 }
2489
2490 int remove_clone(ObjectStore *store, coll_t coll, ghobject_t &ghobj, snapid_t cloneid, bool force,
2491 ObjectStore::Sequencer &osr)
2492 {
2493 // XXX: Don't allow this if in a cache tier or former cache tier
2494 // bool allow_incomplete_clones() const {
2495 // return cache_mode != CACHEMODE_NONE || has_flag(FLAG_INCOMPLETE_CLONES);
2496
2497 SnapSet snapset;
2498 int ret = get_snapset(store, coll, ghobj, snapset);
2499 if (ret < 0)
2500 return ret;
2501
2502 // Derived from trim_object()
2503 // ...from snapset
2504 vector<snapid_t>::iterator p;
2505 for (p = snapset.clones.begin(); p != snapset.clones.end(); ++p)
2506 if (*p == cloneid)
2507 break;
2508 if (p == snapset.clones.end()) {
2509 cerr << "Clone " << cloneid << " not present";
2510 return -ENOENT;
2511 }
2512 if (p != snapset.clones.begin()) {
2513 // not the oldest... merge overlap into next older clone
2514 vector<snapid_t>::iterator n = p - 1;
2515 hobject_t prev_coid = ghobj.hobj;
2516 prev_coid.snap = *n;
2517 //bool adjust_prev_bytes = is_present_clone(prev_coid);
2518
2519 //if (adjust_prev_bytes)
2520 // ctx->delta_stats.num_bytes -= snapset.get_clone_bytes(*n);
2521
2522 snapset.clone_overlap[*n].intersection_of(
2523 snapset.clone_overlap[*p]);
2524
2525 //if (adjust_prev_bytes)
2526 // ctx->delta_stats.num_bytes += snapset.get_clone_bytes(*n);
2527 }
2528
2529 ret = remove_from(snapset.clones, "clones", cloneid, force);
2530 if (ret) return ret;
2531 ret = remove_from(snapset.clone_overlap, "clone_overlap", cloneid, force);
2532 if (ret) return ret;
2533 ret = remove_from(snapset.clone_size, "clone_size", cloneid, force);
2534 if (ret) return ret;
2535
2536 if (dry_run)
2537 return 0;
2538
2539 bufferlist bl;
2540 ::encode(snapset, bl);
2541 ObjectStore::Transaction t;
2542 t.setattr(coll, ghobj, SS_ATTR, bl);
2543 int r = store->apply_transaction(&osr, std::move(t));
2544 if (r < 0) {
2545 cerr << "Error setting snapset on : " << make_pair(coll, ghobj) << ", "
2546 << cpp_strerror(r) << std::endl;
2547 return r;
2548 }
2549 cout << "Removal of clone " << cloneid << " complete" << std::endl;
2550 cout << "Use pg repair after OSD restarted to correct stat information" << std::endl;
2551 return 0;
2552 }
2553
2554 int dup(string srcpath, ObjectStore *src, string dstpath, ObjectStore *dst)
2555 {
2556 cout << "dup from " << src->get_type() << ": " << srcpath << "\n"
2557 << " to " << dst->get_type() << ": " << dstpath
2558 << std::endl;
2559 ObjectStore::Sequencer osr("dup");
2560 int num, i;
2561 vector<coll_t> collections;
2562 int r;
2563
2564 r = src->mount();
2565 if (r < 0) {
2566 cerr << "failed to mount src: " << cpp_strerror(r) << std::endl;
2567 return r;
2568 }
2569 r = dst->mount();
2570 if (r < 0) {
2571 cerr << "failed to mount dst: " << cpp_strerror(r) << std::endl;
2572 goto out_src;
2573 }
2574
2575 if (src->get_fsid() != dst->get_fsid()) {
2576 cerr << "src fsid " << src->get_fsid() << " != dest " << dst->get_fsid()
2577 << std::endl;
2578 goto out;
2579 }
2580 cout << "fsid " << src->get_fsid() << std::endl;
2581
2582 // make sure dst is empty
2583 r = dst->list_collections(collections);
2584 if (r < 0) {
2585 cerr << "error listing collections on dst: " << cpp_strerror(r) << std::endl;
2586 goto out;
2587 }
2588 if (!collections.empty()) {
2589 cerr << "destination store is not empty" << std::endl;
2590 goto out;
2591 }
2592
2593 r = src->list_collections(collections);
2594 if (r < 0) {
2595 cerr << "error listing collections on src: " << cpp_strerror(r) << std::endl;
2596 goto out;
2597 }
2598
2599 num = collections.size();
2600 cout << num << " collections" << std::endl;
2601 i = 1;
2602 for (auto cid : collections) {
2603 cout << i++ << "/" << num << " " << cid << std::endl;
2604 {
2605 ObjectStore::Transaction t;
2606 int bits = src->collection_bits(cid);
2607 if (bits < 0) {
2608 if (src->get_type() == "filestore" && cid.is_meta()) {
2609 bits = 0;
2610 } else {
2611 cerr << "cannot get bit count for collection " << cid << ": "
2612 << cpp_strerror(bits) << std::endl;
2613 goto out;
2614 }
2615 }
2616 t.create_collection(cid, bits);
2617 dst->apply_transaction(&osr, std::move(t));
2618 }
2619
2620 ghobject_t pos;
2621 uint64_t n = 0;
2622 uint64_t bytes = 0, keys = 0;
2623 while (true) {
2624 vector<ghobject_t> ls;
2625 r = src->collection_list(cid, pos, ghobject_t::get_max(), 1000, &ls, &pos);
2626 if (r < 0) {
2627 cerr << "collection_list on " << cid << " from " << pos << " got: "
2628 << cpp_strerror(r) << std::endl;
2629 goto out;
2630 }
2631 if (ls.empty()) {
2632 break;
2633 }
2634
2635 for (auto& oid : ls) {
2636 //cout << " " << cid << " " << oid << std::endl;
2637 if (n % 100 == 0) {
2638 cout << " " << std::setw(16) << n << " objects, "
2639 << std::setw(16) << bytes << " bytes, "
2640 << std::setw(16) << keys << " keys"
2641 << std::setw(1) << "\r" << std::flush;
2642 }
2643 n++;
2644
2645 ObjectStore::Transaction t;
2646 t.touch(cid, oid);
2647
2648 map<string,bufferptr> attrs;
2649 src->getattrs(cid, oid, attrs);
2650 if (!attrs.empty()) {
2651 t.setattrs(cid, oid, attrs);
2652 }
2653
2654 bufferlist bl;
2655 src->read(cid, oid, 0, 0, bl);
2656 if (bl.length()) {
2657 t.write(cid, oid, 0, bl.length(), bl);
2658 bytes += bl.length();
2659 }
2660
2661 bufferlist header;
2662 map<string,bufferlist> omap;
2663 src->omap_get(cid, oid, &header, &omap);
2664 if (header.length()) {
2665 t.omap_setheader(cid, oid, header);
2666 ++keys;
2667 }
2668 if (!omap.empty()) {
2669 keys += omap.size();
2670 t.omap_setkeys(cid, oid, omap);
2671 }
2672
2673 dst->apply_transaction(&osr, std::move(t));
2674 }
2675 }
2676 cout << " " << std::setw(16) << n << " objects, "
2677 << std::setw(16) << bytes << " bytes, "
2678 << std::setw(16) << keys << " keys"
2679 << std::setw(1) << std::endl;
2680 }
2681
2682 // keyring
2683 cout << "keyring" << std::endl;
2684 {
2685 bufferlist bl;
2686 string s = srcpath + "/keyring";
2687 string err;
2688 r = bl.read_file(s.c_str(), &err);
2689 if (r < 0) {
2690 cerr << "failed to copy " << s << ": " << err << std::endl;
2691 } else {
2692 string d = dstpath + "/keyring";
2693 bl.write_file(d.c_str(), 0600);
2694 }
2695 }
2696
2697 // osd metadata
2698 cout << "duping osd metadata" << std::endl;
2699 {
2700 for (auto k : {"magic", "whoami", "ceph_fsid", "fsid"}) {
2701 string val;
2702 src->read_meta(k, &val);
2703 dst->write_meta(k, val);
2704 }
2705 }
2706
2707 dst->write_meta("ready", "ready");
2708
2709 cout << "done." << std::endl;
2710 r = 0;
2711 out:
2712 dst->umount();
2713 out_src:
2714 src->umount();
2715 return r;
2716 }
2717
2718 void usage(po::options_description &desc)
2719 {
2720 cerr << std::endl;
2721 cerr << desc << std::endl;
2722 cerr << std::endl;
2723 cerr << "Positional syntax:" << std::endl;
2724 cerr << std::endl;
2725 cerr << "ceph-objectstore-tool ... <object> (get|set)-bytes [file]" << std::endl;
2726 cerr << "ceph-objectstore-tool ... <object> set-(attr|omap) <key> [file]" << std::endl;
2727 cerr << "ceph-objectstore-tool ... <object> (get|rm)-(attr|omap) <key>" << std::endl;
2728 cerr << "ceph-objectstore-tool ... <object> get-omaphdr" << std::endl;
2729 cerr << "ceph-objectstore-tool ... <object> set-omaphdr [file]" << std::endl;
2730 cerr << "ceph-objectstore-tool ... <object> list-attrs" << std::endl;
2731 cerr << "ceph-objectstore-tool ... <object> list-omap" << std::endl;
2732 cerr << "ceph-objectstore-tool ... <object> remove|removeall" << std::endl;
2733 cerr << "ceph-objectstore-tool ... <object> dump" << std::endl;
2734 cerr << "ceph-objectstore-tool ... <object> set-size" << std::endl;
2735 cerr << "ceph-objectstore-tool ... <object> remove-clone-metadata <cloneid>" << std::endl;
2736 cerr << std::endl;
2737 cerr << "<object> can be a JSON object description as displayed" << std::endl;
2738 cerr << "by --op list." << std::endl;
2739 cerr << "<object> can be an object name which will be looked up in all" << std::endl;
2740 cerr << "the OSD's PGs." << std::endl;
2741 cerr << "<object> can be the empty string ('') which with a provided pgid " << std::endl;
2742 cerr << "specifies the pgmeta object" << std::endl;
2743 cerr << std::endl;
2744 cerr << "The optional [file] argument will read stdin or write stdout" << std::endl;
2745 cerr << "if not specified or if '-' specified." << std::endl;
2746 }
2747
2748 bool ends_with(const string& check, const string& ending)
2749 {
2750 return check.size() >= ending.size() && check.rfind(ending) == (check.size() - ending.size());
2751 }
2752
2753 // Based on FileStore::dump_journal(), set-up enough to only dump
2754 int mydump_journal(Formatter *f, string journalpath, bool m_journal_dio)
2755 {
2756 int r;
2757
2758 if (!journalpath.length())
2759 return -EINVAL;
2760
2761 FileJournal *journal = new FileJournal(g_ceph_context, uuid_d(), NULL, NULL,
2762 journalpath.c_str(), m_journal_dio);
2763 r = journal->_fdump(*f, false);
2764 delete journal;
2765 return r;
2766 }
2767
2768 int apply_layout_settings(ObjectStore *os, const OSDSuperblock &superblock,
2769 const string &pool_name, const spg_t &pgid, bool dry_run)
2770 {
2771 int r = 0;
2772
2773 FileStore *fs = dynamic_cast<FileStore*>(os);
2774 if (!fs) {
2775 cerr << "Nothing to do for non-filestore backend" << std::endl;
2776 return 0; // making this return success makes testing easier
2777 }
2778
2779 OSDMap curmap;
2780 bufferlist bl;
2781 r = get_osdmap(os, superblock.current_epoch, curmap, bl);
2782 if (r) {
2783 cerr << "Can't find local OSDMap: " << cpp_strerror(r) << std::endl;
2784 return r;
2785 }
2786
2787 int64_t poolid = -1;
2788 if (pool_name.length()) {
2789 poolid = curmap.lookup_pg_pool_name(pool_name);
2790 if (poolid < 0) {
2791 cerr << "Couldn't find pool " << pool_name << ": " << cpp_strerror(poolid)
2792 << std::endl;
2793 return poolid;
2794 }
2795 }
2796
2797 vector<coll_t> collections, filtered_colls;
2798 r = os->list_collections(collections);
2799 if (r < 0) {
2800 cerr << "Error listing collections: " << cpp_strerror(r) << std::endl;
2801 return r;
2802 }
2803
2804 for (auto const &coll : collections) {
2805 spg_t coll_pgid;
2806 if (coll.is_pg(&coll_pgid) &&
2807 ((poolid >= 0 && coll_pgid.pool() == (uint64_t)poolid) ||
2808 coll_pgid == pgid)) {
2809 filtered_colls.push_back(coll);
2810 }
2811 }
2812
2813 size_t done = 0, total = filtered_colls.size();
2814 for (auto const &coll : filtered_colls) {
2815 if (dry_run) {
2816 cerr << "Would apply layout settings to " << coll << std::endl;
2817 } else {
2818 cerr << "Finished " << done << "/" << total << " collections" << "\r";
2819 r = fs->apply_layout_settings(coll);
2820 if (r < 0) {
2821 cerr << "Error applying layout settings to " << coll << std::endl;
2822 return r;
2823 }
2824 }
2825 ++done;
2826 }
2827
2828 cerr << "Finished " << total << "/" << total << " collections" << "\r" << std::endl;
2829 return r;
2830 }
2831
2832 int main(int argc, char **argv)
2833 {
2834 string dpath, jpath, pgidstr, op, file, mountpoint, mon_store_path, object;
2835 string target_data_path, fsid;
2836 string objcmd, arg1, arg2, type, format, argnspace, pool;
2837 boost::optional<std::string> nspace;
2838 spg_t pgid;
2839 unsigned epoch = 0;
2840 ghobject_t ghobj;
2841 bool human_readable;
2842 bool force;
2843 Formatter *formatter;
2844 bool head;
2845
2846 po::options_description desc("Allowed options");
2847 desc.add_options()
2848 ("help", "produce help message")
2849 ("type", po::value<string>(&type),
2850 "Arg is one of [bluestore, filestore (default), memstore]")
2851 ("data-path", po::value<string>(&dpath),
2852 "path to object store, mandatory")
2853 ("journal-path", po::value<string>(&jpath),
2854 "path to journal, use if tool can't find it")
2855 ("pgid", po::value<string>(&pgidstr),
2856 "PG id, mandatory for info, log, remove, export, export-remove, rm-past-intervals, mark-complete, and mandatory for apply-layout-settings if --pool is not specified")
2857 ("pool", po::value<string>(&pool),
2858 "Pool name, mandatory for apply-layout-settings if --pgid is not specified")
2859 ("op", po::value<string>(&op),
2860 "Arg is one of [info, log, remove, mkfs, fsck, repair, fuse, dup, export, export-remove, import, list, fix-lost, list-pgs, rm-past-intervals, dump-journal, dump-super, meta-list, "
2861 "get-osdmap, set-osdmap, get-inc-osdmap, set-inc-osdmap, mark-complete, apply-layout-settings, update-mon-db, dump-import]")
2862 ("epoch", po::value<unsigned>(&epoch),
2863 "epoch# for get-osdmap and get-inc-osdmap, the current epoch in use if not specified")
2864 ("file", po::value<string>(&file),
2865 "path of file to export, export-remove, import, get-osdmap, set-osdmap, get-inc-osdmap or set-inc-osdmap")
2866 ("mon-store-path", po::value<string>(&mon_store_path),
2867 "path of monstore to update-mon-db")
2868 ("fsid", po::value<string>(&fsid),
2869 "fsid for new store created by mkfs")
2870 ("target-data-path", po::value<string>(&target_data_path),
2871 "path of target object store (for --op dup)")
2872 ("mountpoint", po::value<string>(&mountpoint),
2873 "fuse mountpoint")
2874 ("format", po::value<string>(&format)->default_value("json-pretty"),
2875 "Output format which may be json, json-pretty, xml, xml-pretty")
2876 ("debug", "Enable diagnostic output to stderr")
2877 ("force", "Ignore some types of errors and proceed with operation - USE WITH CAUTION: CORRUPTION POSSIBLE NOW OR IN THE FUTURE")
2878 ("skip-journal-replay", "Disable journal replay")
2879 ("skip-mount-omap", "Disable mounting of omap")
2880 ("head", "Find head/snapdir when searching for objects by name")
2881 ("dry-run", "Don't modify the objectstore")
2882 ("namespace", po::value<string>(&argnspace), "Specify namespace when searching for objects")
2883 ;
2884
2885 po::options_description positional("Positional options");
2886 positional.add_options()
2887 ("object", po::value<string>(&object), "'' for pgmeta_oid, object name or ghobject in json")
2888 ("objcmd", po::value<string>(&objcmd), "command [(get|set)-bytes, (get|set|rm)-(attr|omap), (get|set)-omaphdr, list-attrs, list-omap, remove]")
2889 ("arg1", po::value<string>(&arg1), "arg1 based on cmd")
2890 ("arg2", po::value<string>(&arg2), "arg2 based on cmd")
2891 ;
2892
2893 po::options_description all;
2894 all.add(desc).add(positional);
2895
2896 po::positional_options_description pd;
2897 pd.add("object", 1).add("objcmd", 1).add("arg1", 1).add("arg2", 1);
2898
2899 vector<string> ceph_option_strings;
2900 po::variables_map vm;
2901 try {
2902 po::parsed_options parsed =
2903 po::command_line_parser(argc, argv).options(all).allow_unregistered().positional(pd).run();
2904 po::store( parsed, vm);
2905 po::notify(vm);
2906 ceph_option_strings = po::collect_unrecognized(parsed.options,
2907 po::include_positional);
2908 } catch(po::error &e) {
2909 std::cerr << e.what() << std::endl;
2910 return 1;
2911 }
2912
2913 if (vm.count("help")) {
2914 usage(desc);
2915 return 1;
2916 }
2917
2918 debug = (vm.count("debug") > 0);
2919
2920 force = (vm.count("force") > 0);
2921
2922 if (vm.count("namespace"))
2923 nspace = argnspace;
2924
2925 dry_run = (vm.count("dry-run") > 0);
2926
2927 osflagbits_t flags = 0;
2928 if (dry_run || vm.count("skip-journal-replay"))
2929 flags |= SKIP_JOURNAL_REPLAY;
2930 if (vm.count("skip-mount-omap"))
2931 flags |= SKIP_MOUNT_OMAP;
2932 if (op == "update-mon-db")
2933 flags |= SKIP_JOURNAL_REPLAY;
2934
2935 head = (vm.count("head") > 0);
2936
2937 vector<const char *> ceph_options;
2938 env_to_vec(ceph_options);
2939 ceph_options.reserve(ceph_options.size() + ceph_option_strings.size());
2940 for (vector<string>::iterator i = ceph_option_strings.begin();
2941 i != ceph_option_strings.end();
2942 ++i) {
2943 ceph_options.push_back(i->c_str());
2944 }
2945
2946 char fn[PATH_MAX];
2947 snprintf(fn, sizeof(fn), "%s/type", dpath.c_str());
2948 int fd = ::open(fn, O_RDONLY);
2949 if (fd >= 0) {
2950 bufferlist bl;
2951 bl.read_fd(fd, 64);
2952 if (bl.length()) {
2953 string dp_type = string(bl.c_str(), bl.length() - 1); // drop \n
2954 if (vm.count("type") && dp_type != "" && type != dp_type)
2955 cerr << "WARNING: Ignoring type \"" << type << "\" - found data-path type \""
2956 << dp_type << "\"" << std::endl;
2957 type = dp_type;
2958 //cout << "object store type is " << type << std::endl;
2959 }
2960 ::close(fd);
2961 }
2962 if (!vm.count("type") && type == "") {
2963 type = "filestore";
2964 }
2965 if (!vm.count("data-path") &&
2966 op != "dump-import" &&
2967 !(op == "dump-journal" && type == "filestore")) {
2968 cerr << "Must provide --data-path" << std::endl;
2969 usage(desc);
2970 return 1;
2971 }
2972 if (type == "filestore" && !vm.count("journal-path")) {
2973 jpath = dpath + "/journal";
2974 }
2975 if (!vm.count("op") && !vm.count("object")) {
2976 cerr << "Must provide --op or object command..." << std::endl;
2977 usage(desc);
2978 return 1;
2979 }
2980 if (op != "list" &&
2981 vm.count("op") && vm.count("object")) {
2982 cerr << "Can't specify both --op and object command syntax" << std::endl;
2983 usage(desc);
2984 return 1;
2985 }
2986 if (op == "apply-layout-settings" && !(vm.count("pool") ^ vm.count("pgid"))) {
2987 cerr << "apply-layout-settings requires either --pool or --pgid"
2988 << std::endl;
2989 usage(desc);
2990 return 1;
2991 }
2992 if (op != "list" && vm.count("object") && !vm.count("objcmd")) {
2993 cerr << "Invalid syntax, missing command" << std::endl;
2994 usage(desc);
2995 return 1;
2996 }
2997 if (op == "fuse" && mountpoint.length() == 0) {
2998 cerr << "Missing fuse mountpoint" << std::endl;
2999 usage(desc);
3000 return 1;
3001 }
3002 outistty = isatty(STDOUT_FILENO);
3003
3004 file_fd = fd_none;
3005 if ((op == "export" || op == "export-remove" || op == "get-osdmap" || op == "get-inc-osdmap") && !dry_run) {
3006 if (!vm.count("file") || file == "-") {
3007 if (outistty) {
3008 cerr << "stdout is a tty and no --file filename specified" << std::endl;
3009 return 1;
3010 }
3011 file_fd = STDOUT_FILENO;
3012 } else {
3013 file_fd = open(file.c_str(), O_WRONLY|O_CREAT|O_TRUNC, 0666);
3014 }
3015 } else if (op == "import" || op == "dump-import" || op == "set-osdmap" || op == "set-inc-osdmap") {
3016 if (!vm.count("file") || file == "-") {
3017 if (isatty(STDIN_FILENO)) {
3018 cerr << "stdin is a tty and no --file filename specified" << std::endl;
3019 return 1;
3020 }
3021 file_fd = STDIN_FILENO;
3022 } else {
3023 file_fd = open(file.c_str(), O_RDONLY);
3024 }
3025 }
3026
3027 ObjectStoreTool tool = ObjectStoreTool(file_fd, dry_run);
3028
3029 if (vm.count("file") && file_fd == fd_none && !dry_run) {
3030 cerr << "--file option only applies to import, dump-import, export, export-remove, "
3031 << "get-osdmap, set-osdmap, get-inc-osdmap or set-inc-osdmap" << std::endl;
3032 return 1;
3033 }
3034
3035 if (file_fd != fd_none && file_fd < 0) {
3036 string err = string("file: ") + file;
3037 perror(err.c_str());
3038 return 1;
3039 }
3040
3041 auto cct = global_init(
3042 NULL, ceph_options, CEPH_ENTITY_TYPE_OSD,
3043 CODE_ENVIRONMENT_UTILITY_NODOUT, 0);
3044 //CINIT_FLAG_NO_DEFAULT_CONFIG_FILE);
3045 common_init_finish(g_ceph_context);
3046 g_conf = g_ceph_context->_conf;
3047 if (debug) {
3048 g_conf->set_val_or_die("log_to_stderr", "true");
3049 g_conf->set_val_or_die("err_to_stderr", "true");
3050 }
3051 g_conf->apply_changes(NULL);
3052
3053 // Special list handling. Treating pretty_format as human readable,
3054 // with one object per line and not an enclosing array.
3055 human_readable = ends_with(format, "-pretty");
3056 if ((op == "list" || op == "meta-list") && human_readable) {
3057 // Remove -pretty from end of format which we know is there
3058 format = format.substr(0, format.size() - strlen("-pretty"));
3059 }
3060
3061 formatter = Formatter::create(format);
3062 if (formatter == NULL) {
3063 cerr << "unrecognized format: " << format << std::endl;
3064 return 1;
3065 }
3066
3067 // Special handling for filestore journal, so we can dump it without mounting
3068 if (op == "dump-journal" && type == "filestore") {
3069 int ret = mydump_journal(formatter, jpath, g_conf->journal_dio);
3070 if (ret < 0) {
3071 cerr << "journal-path: " << jpath << ": "
3072 << cpp_strerror(ret) << std::endl;
3073 return 1;
3074 }
3075 formatter->flush(cout);
3076 return 0;
3077 }
3078
3079 if (op == "dump-import") {
3080 int ret = tool.dump_import(formatter);
3081 if (ret < 0) {
3082 cerr << "dump-import: "
3083 << cpp_strerror(ret) << std::endl;
3084 return 1;
3085 }
3086 return 0;
3087 }
3088
3089 //Verify that data-path really exists
3090 struct stat st;
3091 if (::stat(dpath.c_str(), &st) == -1) {
3092 string err = string("data-path: ") + dpath;
3093 perror(err.c_str());
3094 return 1;
3095 }
3096
3097 if (pgidstr.length() && !pgid.parse(pgidstr.c_str())) {
3098 cerr << "Invalid pgid '" << pgidstr << "' specified" << std::endl;
3099 return 1;
3100 }
3101
3102 //Verify that the journal-path really exists
3103 if (type == "filestore") {
3104 if (::stat(jpath.c_str(), &st) == -1) {
3105 string err = string("journal-path: ") + jpath;
3106 perror(err.c_str());
3107 return 1;
3108 }
3109 if (S_ISDIR(st.st_mode)) {
3110 cerr << "journal-path: " << jpath << ": "
3111 << cpp_strerror(EISDIR) << std::endl;
3112 return 1;
3113 }
3114 }
3115
3116 ObjectStore *fs = ObjectStore::create(g_ceph_context, type, dpath, jpath, flags);
3117 if (fs == NULL) {
3118 cerr << "Unable to create store of type " << type << std::endl;
3119 return 1;
3120 }
3121
3122 if (op == "fsck" || op == "fsck-deep") {
3123 int r = fs->fsck(op == "fsck-deep");
3124 if (r < 0) {
3125 cerr << "fsck failed: " << cpp_strerror(r) << std::endl;
3126 return 1;
3127 }
3128 if (r > 0) {
3129 cerr << "fsck found " << r << " errors" << std::endl;
3130 return 1;
3131 }
3132 cout << "fsck found no errors" << std::endl;
3133 return 0;
3134 }
3135 if (op == "repair" || op == "repair-deep") {
3136 int r = fs->repair(op == "repair-deep");
3137 if (r < 0) {
3138 cerr << "repair failed: " << cpp_strerror(r) << std::endl;
3139 return 1;
3140 }
3141 if (r > 0) {
3142 cerr << "repair found " << r << " errors" << std::endl;
3143 return 1;
3144 }
3145 cout << "repair found no errors" << std::endl;
3146 return 0;
3147 }
3148 if (op == "mkfs") {
3149 if (fsid.length()) {
3150 uuid_d f;
3151 bool r = f.parse(fsid.c_str());
3152 if (!r) {
3153 cerr << "failed to parse uuid '" << fsid << "'" << std::endl;
3154 return 1;
3155 }
3156 fs->set_fsid(f);
3157 }
3158 int r = fs->mkfs();
3159 if (r < 0) {
3160 cerr << "mkfs failed: " << cpp_strerror(r) << std::endl;
3161 return 1;
3162 }
3163 return 0;
3164 }
3165 if (op == "dup") {
3166 string target_type;
3167 char fn[PATH_MAX];
3168 snprintf(fn, sizeof(fn), "%s/type", target_data_path.c_str());
3169 int fd = ::open(fn, O_RDONLY);
3170 if (fd < 0) {
3171 cerr << "Unable to open " << target_data_path << "/type" << std::endl;
3172 exit(1);
3173 }
3174 bufferlist bl;
3175 bl.read_fd(fd, 64);
3176 if (bl.length()) {
3177 target_type = string(bl.c_str(), bl.length() - 1); // drop \n
3178 }
3179 ::close(fd);
3180 ObjectStore *targetfs = ObjectStore::create(
3181 g_ceph_context, target_type,
3182 target_data_path, "", 0);
3183 if (targetfs == NULL) {
3184 cerr << "Unable to open store of type " << target_type << std::endl;
3185 return 1;
3186 }
3187 int r = dup(dpath, fs, target_data_path, targetfs);
3188 if (r < 0) {
3189 cerr << "dup failed: " << cpp_strerror(r) << std::endl;
3190 return 1;
3191 }
3192 return 0;
3193 }
3194
3195 ObjectStore::Sequencer *osr = new ObjectStore::Sequencer(__func__);
3196 int ret = fs->mount();
3197 if (ret < 0) {
3198 if (ret == -EBUSY) {
3199 cerr << "OSD has the store locked" << std::endl;
3200 } else {
3201 cerr << "Mount failed with '" << cpp_strerror(ret) << "'" << std::endl;
3202 }
3203 return 1;
3204 }
3205
3206 if (op == "fuse") {
3207 #ifdef HAVE_LIBFUSE
3208 FuseStore fuse(fs, mountpoint);
3209 cout << "mounting fuse at " << mountpoint << " ..." << std::endl;
3210 int r = fuse.main();
3211 if (r < 0) {
3212 cerr << "failed to mount fuse: " << cpp_strerror(r) << std::endl;
3213 return 1;
3214 }
3215 #else
3216 cerr << "fuse support not enabled" << std::endl;
3217 #endif
3218 return 0;
3219 }
3220
3221 vector<coll_t> ls;
3222 vector<coll_t>::iterator it;
3223 CompatSet supported;
3224
3225 #ifdef INTERNAL_TEST
3226 supported = get_test_compat_set();
3227 #else
3228 supported = OSD::get_osd_compat_set();
3229 #endif
3230
3231 bufferlist bl;
3232 OSDSuperblock superblock;
3233 bufferlist::iterator p;
3234 ret = fs->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT, 0, 0, bl);
3235 if (ret < 0) {
3236 cerr << "Failure to read OSD superblock: " << cpp_strerror(ret) << std::endl;
3237 goto out;
3238 }
3239
3240 p = bl.begin();
3241 ::decode(superblock, p);
3242
3243 if (debug) {
3244 cerr << "Cluster fsid=" << superblock.cluster_fsid << std::endl;
3245 }
3246
3247 if (debug) {
3248 cerr << "Supported features: " << supported << std::endl;
3249 cerr << "On-disk features: " << superblock.compat_features << std::endl;
3250 }
3251 if (supported.compare(superblock.compat_features) == -1) {
3252 CompatSet unsupported = supported.unsupported(superblock.compat_features);
3253 cerr << "On-disk OSD incompatible features set "
3254 << unsupported << std::endl;
3255 ret = -EINVAL;
3256 goto out;
3257 }
3258
3259 if (op == "apply-layout-settings") {
3260 ret = apply_layout_settings(fs, superblock, pool, pgid, dry_run);
3261 goto out;
3262 }
3263
3264 if (op != "list" && vm.count("object")) {
3265 // Special case: Create pgmeta_oid if empty string specified
3266 // This can't conflict with any actual object names.
3267 if (object == "") {
3268 ghobj = pgid.make_pgmeta_oid();
3269 } else {
3270 json_spirit::Value v;
3271 try {
3272 if (!json_spirit::read(object, v) ||
3273 (v.type() != json_spirit::array_type && v.type() != json_spirit::obj_type)) {
3274 // Special: Need head/snapdir so set even if user didn't specify
3275 if (vm.count("objcmd") && (objcmd == "remove-clone-metadata"))
3276 head = true;
3277 lookup_ghobject lookup(object, nspace, head);
3278 if (pgidstr.length())
3279 ret = action_on_all_objects_in_exact_pg(fs, coll_t(pgid), lookup, debug);
3280 else
3281 ret = action_on_all_objects(fs, lookup, debug);
3282 if (ret) {
3283 throw std::runtime_error("Internal error");
3284 } else {
3285 if (lookup.size() != 1) {
3286 stringstream ss;
3287 if (lookup.size() == 0)
3288 ss << "No object id '" << object << "' found or invalid JSON specified";
3289 else
3290 ss << "Found " << lookup.size() << " objects with id '" << object
3291 << "', please use a JSON spec from --op list instead";
3292 throw std::runtime_error(ss.str());
3293 }
3294 pair<coll_t, ghobject_t> found = lookup.pop();
3295 pgidstr = found.first.to_str();
3296 pgid.parse(pgidstr.c_str());
3297 ghobj = found.second;
3298 }
3299 } else {
3300 stringstream ss;
3301 if (pgidstr.length() == 0 && v.type() != json_spirit::array_type) {
3302 ss << "Without --pgid the object '" << object
3303 << "' must be a JSON array";
3304 throw std::runtime_error(ss.str());
3305 }
3306 if (v.type() == json_spirit::array_type) {
3307 json_spirit::Array array = v.get_array();
3308 if (array.size() != 2) {
3309 ss << "Object '" << object
3310 << "' must be a JSON array with 2 elements";
3311 throw std::runtime_error(ss.str());
3312 }
3313 vector<json_spirit::Value>::iterator i = array.begin();
3314 assert(i != array.end());
3315 if (i->type() != json_spirit::str_type) {
3316 ss << "Object '" << object
3317 << "' must be a JSON array with the first element a string";
3318 throw std::runtime_error(ss.str());
3319 }
3320 string object_pgidstr = i->get_str();
3321 if (object_pgidstr != "meta") {
3322 spg_t object_pgid;
3323 object_pgid.parse(object_pgidstr.c_str());
3324 if (pgidstr.length() > 0) {
3325 if (object_pgid != pgid) {
3326 ss << "object '" << object
3327 << "' has a pgid different from the --pgid="
3328 << pgidstr << " option";
3329 throw std::runtime_error(ss.str());
3330 }
3331 } else {
3332 pgidstr = object_pgidstr;
3333 pgid = object_pgid;
3334 }
3335 } else {
3336 pgidstr = object_pgidstr;
3337 }
3338 ++i;
3339 v = *i;
3340 }
3341 try {
3342 ghobj.decode(v);
3343 } catch (std::runtime_error& e) {
3344 ss << "Decode object JSON error: " << e.what();
3345 throw std::runtime_error(ss.str());
3346 }
3347 if (pgidstr != "meta" && (uint64_t)pgid.pgid.m_pool != (uint64_t)ghobj.hobj.pool) {
3348 cerr << "Object pool and pgid pool don't match" << std::endl;
3349 ret = 1;
3350 goto out;
3351 }
3352 }
3353 } catch (std::runtime_error& e) {
3354 cerr << e.what() << std::endl;
3355 ret = 1;
3356 goto out;
3357 }
3358 }
3359 }
3360
3361 // The ops which require --pgid option are checked here and
3362 // mentioned in the usage for --pgid.
3363 if ((op == "info" || op == "log" || op == "remove" || op == "export"
3364 || op == "export-remove" || op == "rm-past-intervals" || op == "mark-complete") &&
3365 pgidstr.length() == 0) {
3366 cerr << "Must provide pgid" << std::endl;
3367 usage(desc);
3368 ret = 1;
3369 goto out;
3370 }
3371
3372 if (op == "import") {
3373
3374 try {
3375 ret = tool.do_import(fs, superblock, force, pgidstr, *osr);
3376 }
3377 catch (const buffer::error &e) {
3378 cerr << "do_import threw exception error " << e.what() << std::endl;
3379 ret = -EFAULT;
3380 }
3381 if (ret == -EFAULT) {
3382 cerr << "Corrupt input for import" << std::endl;
3383 }
3384 if (ret == 0)
3385 cout << "Import successful" << std::endl;
3386 goto out;
3387 } else if (op == "dump-journal-mount") {
3388 // Undocumented feature to dump journal with mounted fs
3389 // This doesn't support the format option, but it uses the
3390 // ObjectStore::dump_journal() and mounts to get replay to run.
3391 ret = fs->dump_journal(cout);
3392 if (ret) {
3393 if (ret == -EOPNOTSUPP) {
3394 cerr << "Object store type \"" << type << "\" doesn't support journal dump" << std::endl;
3395 } else {
3396 cerr << "Journal dump failed with error " << cpp_strerror(ret) << std::endl;
3397 }
3398 }
3399 goto out;
3400 } else if (op == "get-osdmap") {
3401 bufferlist bl;
3402 OSDMap osdmap;
3403 if (epoch == 0) {
3404 epoch = superblock.current_epoch;
3405 }
3406 ret = get_osdmap(fs, epoch, osdmap, bl);
3407 if (ret) {
3408 cerr << "Failed to get osdmap#" << epoch << ": "
3409 << cpp_strerror(ret) << std::endl;
3410 goto out;
3411 }
3412 ret = bl.write_fd(file_fd);
3413 if (ret) {
3414 cerr << "Failed to write to " << file << ": " << cpp_strerror(ret) << std::endl;
3415 } else {
3416 cout << "osdmap#" << epoch << " exported." << std::endl;
3417 }
3418 goto out;
3419 } else if (op == "set-osdmap") {
3420 bufferlist bl;
3421 ret = get_fd_data(file_fd, bl);
3422 if (ret < 0) {
3423 cerr << "Failed to read osdmap " << cpp_strerror(ret) << std::endl;
3424 } else {
3425 ret = set_osdmap(fs, epoch, bl, force, *osr);
3426 }
3427 goto out;
3428 } else if (op == "get-inc-osdmap") {
3429 bufferlist bl;
3430 if (epoch == 0) {
3431 epoch = superblock.current_epoch;
3432 }
3433 ret = get_inc_osdmap(fs, epoch, bl);
3434 if (ret < 0) {
3435 cerr << "Failed to get incremental osdmap# " << epoch << ": "
3436 << cpp_strerror(ret) << std::endl;
3437 goto out;
3438 }
3439 ret = bl.write_fd(file_fd);
3440 if (ret) {
3441 cerr << "Failed to write to " << file << ": " << cpp_strerror(ret) << std::endl;
3442 } else {
3443 cout << "inc-osdmap#" << epoch << " exported." << std::endl;
3444 }
3445 goto out;
3446 } else if (op == "set-inc-osdmap") {
3447 bufferlist bl;
3448 ret = get_fd_data(file_fd, bl);
3449 if (ret < 0) {
3450 cerr << "Failed to read incremental osdmap " << cpp_strerror(ret) << std::endl;
3451 goto out;
3452 } else {
3453 ret = set_inc_osdmap(fs, epoch, bl, force, *osr);
3454 }
3455 goto out;
3456 } else if (op == "update-mon-db") {
3457 if (!vm.count("mon-store-path")) {
3458 cerr << "Please specify the path to monitor db to update" << std::endl;
3459 ret = -EINVAL;
3460 } else {
3461 ret = update_mon_db(*fs, superblock, dpath + "/keyring", mon_store_path);
3462 }
3463 goto out;
3464 }
3465
3466 log_oid = OSD::make_pg_log_oid(pgid);
3467 biginfo_oid = OSD::make_pg_biginfo_oid(pgid);
3468
3469 if (op == "remove") {
3470 if (!force && !dry_run) {
3471 cerr << "Please use export-remove or you must use --force option" << std::endl;
3472 ret = -EINVAL;
3473 goto out;
3474 }
3475 ret = initiate_new_remove_pg(fs, pgid, *osr);
3476 if (ret < 0) {
3477 cerr << "PG '" << pgid << "' not found" << std::endl;
3478 goto out;
3479 }
3480 cout << "Remove successful" << std::endl;
3481 goto out;
3482 }
3483
3484 if (op == "fix-lost") {
3485 boost::scoped_ptr<action_on_object_t> action;
3486 action.reset(new do_fix_lost(osr));
3487 if (pgidstr.length())
3488 ret = action_on_all_objects_in_exact_pg(fs, coll_t(pgid), *action, debug);
3489 else
3490 ret = action_on_all_objects(fs, *action, debug);
3491 goto out;
3492 }
3493
3494 if (op == "list") {
3495 ret = do_list(fs, pgidstr, object, nspace, formatter, debug,
3496 human_readable, head);
3497 if (ret < 0) {
3498 cerr << "do_list failed: " << cpp_strerror(ret) << std::endl;
3499 }
3500 goto out;
3501 }
3502
3503 if (op == "dump-super") {
3504 formatter->open_object_section("superblock");
3505 superblock.dump(formatter);
3506 formatter->close_section();
3507 formatter->flush(cout);
3508 cout << std::endl;
3509 goto out;
3510 }
3511
3512 if (op == "meta-list") {
3513 ret = do_meta(fs, object, formatter, debug, human_readable);
3514 if (ret < 0) {
3515 cerr << "do_meta failed: " << cpp_strerror(ret) << std::endl;
3516 }
3517 goto out;
3518 }
3519
3520 ret = fs->list_collections(ls);
3521 if (ret < 0) {
3522 cerr << "failed to list pgs: " << cpp_strerror(ret) << std::endl;
3523 goto out;
3524 }
3525
3526 if (debug && op == "list-pgs")
3527 cout << "Performing list-pgs operation" << std::endl;
3528
3529 // Find pg
3530 for (it = ls.begin(); it != ls.end(); ++it) {
3531 spg_t tmppgid;
3532
3533 if (pgidstr == "meta") {
3534 if (it->to_str() == "meta")
3535 break;
3536 else
3537 continue;
3538 }
3539
3540 if (!it->is_pg(&tmppgid)) {
3541 continue;
3542 }
3543
3544 if (it->is_temp(&tmppgid)) {
3545 continue;
3546 }
3547
3548 if (op != "list-pgs" && tmppgid != pgid) {
3549 continue;
3550 }
3551
3552 if (op != "list-pgs") {
3553 //Found!
3554 break;
3555 }
3556
3557 cout << tmppgid << std::endl;
3558 }
3559
3560 if (op == "list-pgs") {
3561 ret = 0;
3562 goto out;
3563 }
3564
3565 // If not an object command nor any of the ops handled below, then output this usage
3566 // before complaining about a bad pgid
3567 if (!vm.count("objcmd") && op != "export" && op != "export-remove" && op != "info" && op != "log" && op != "rm-past-intervals" && op != "mark-complete") {
3568 cerr << "Must provide --op (info, log, remove, mkfs, fsck, repair, export, export-remove, import, list, fix-lost, list-pgs, rm-past-intervals, dump-journal, dump-super, meta-list, "
3569 "get-osdmap, set-osdmap, get-inc-osdmap, set-inc-osdmap, mark-complete, dump-import)"
3570 << std::endl;
3571 usage(desc);
3572 ret = 1;
3573 goto out;
3574 }
3575 epoch_t map_epoch;
3576 // The following code for export, info, log require omap or !skip-mount-omap
3577 if (it != ls.end()) {
3578
3579 coll_t coll = *it;
3580
3581 if (vm.count("objcmd")) {
3582 ret = 0;
3583 if (objcmd == "remove" || objcmd == "removeall") {
3584 bool all = (objcmd == "removeall");
3585 ret = do_remove_object(fs, coll, ghobj, all, force, *osr);
3586 goto out;
3587 } else if (objcmd == "list-attrs") {
3588 ret = do_list_attrs(fs, coll, ghobj);
3589 goto out;
3590 } else if (objcmd == "list-omap") {
3591 ret = do_list_omap(fs, coll, ghobj);
3592 goto out;
3593 } else if (objcmd == "get-bytes" || objcmd == "set-bytes") {
3594 if (objcmd == "get-bytes") {
3595 int fd;
3596 if (vm.count("arg1") == 0 || arg1 == "-") {
3597 fd = STDOUT_FILENO;
3598 } else {
3599 fd = open(arg1.c_str(), O_WRONLY|O_TRUNC|O_CREAT|O_EXCL|O_LARGEFILE, 0666);
3600 if (fd == -1) {
3601 cerr << "open " << arg1 << " " << cpp_strerror(errno) << std::endl;
3602 ret = 1;
3603 goto out;
3604 }
3605 }
3606 ret = do_get_bytes(fs, coll, ghobj, fd);
3607 if (fd != STDOUT_FILENO)
3608 close(fd);
3609 } else {
3610 int fd;
3611 if (vm.count("arg1") == 0 || arg1 == "-") {
3612 // Since read_fd() doesn't handle ^D from a tty stdin, don't allow it.
3613 if (isatty(STDIN_FILENO)) {
3614 cerr << "stdin is a tty and no file specified" << std::endl;
3615 ret = 1;
3616 goto out;
3617 }
3618 fd = STDIN_FILENO;
3619 } else {
3620 fd = open(arg1.c_str(), O_RDONLY|O_LARGEFILE, 0666);
3621 if (fd == -1) {
3622 cerr << "open " << arg1 << " " << cpp_strerror(errno) << std::endl;
3623 ret = 1;
3624 goto out;
3625 }
3626 }
3627 ret = do_set_bytes(fs, coll, ghobj, fd, *osr);
3628 if (fd != STDIN_FILENO)
3629 close(fd);
3630 }
3631 goto out;
3632 } else if (objcmd == "get-attr") {
3633 if (vm.count("arg1") == 0) {
3634 usage(desc);
3635 ret = 1;
3636 goto out;
3637 }
3638 ret = do_get_attr(fs, coll, ghobj, arg1);
3639 goto out;
3640 } else if (objcmd == "set-attr") {
3641 if (vm.count("arg1") == 0) {
3642 usage(desc);
3643 ret = 1;
3644 }
3645
3646 int fd;
3647 if (vm.count("arg2") == 0 || arg2 == "-") {
3648 // Since read_fd() doesn't handle ^D from a tty stdin, don't allow it.
3649 if (isatty(STDIN_FILENO)) {
3650 cerr << "stdin is a tty and no file specified" << std::endl;
3651 ret = 1;
3652 goto out;
3653 }
3654 fd = STDIN_FILENO;
3655 } else {
3656 fd = open(arg2.c_str(), O_RDONLY|O_LARGEFILE, 0666);
3657 if (fd == -1) {
3658 cerr << "open " << arg2 << " " << cpp_strerror(errno) << std::endl;
3659 ret = 1;
3660 goto out;
3661 }
3662 }
3663 ret = do_set_attr(fs, coll, ghobj, arg1, fd, *osr);
3664 if (fd != STDIN_FILENO)
3665 close(fd);
3666 goto out;
3667 } else if (objcmd == "rm-attr") {
3668 if (vm.count("arg1") == 0) {
3669 usage(desc);
3670 ret = 1;
3671 goto out;
3672 }
3673 ret = do_rm_attr(fs, coll, ghobj, arg1, *osr);
3674 goto out;
3675 } else if (objcmd == "get-omap") {
3676 if (vm.count("arg1") == 0) {
3677 usage(desc);
3678 ret = 1;
3679 goto out;
3680 }
3681 ret = do_get_omap(fs, coll, ghobj, arg1);
3682 goto out;
3683 } else if (objcmd == "set-omap") {
3684 if (vm.count("arg1") == 0) {
3685 usage(desc);
3686 ret = 1;
3687 goto out;
3688 }
3689 int fd;
3690 if (vm.count("arg2") == 0 || arg2 == "-") {
3691 // Since read_fd() doesn't handle ^D from a tty stdin, don't allow it.
3692 if (isatty(STDIN_FILENO)) {
3693 cerr << "stdin is a tty and no file specified" << std::endl;
3694 ret = 1;
3695 goto out;
3696 }
3697 fd = STDIN_FILENO;
3698 } else {
3699 fd = open(arg2.c_str(), O_RDONLY|O_LARGEFILE, 0666);
3700 if (fd == -1) {
3701 cerr << "open " << arg2 << " " << cpp_strerror(errno) << std::endl;
3702 ret = 1;
3703 goto out;
3704 }
3705 }
3706 ret = do_set_omap(fs, coll, ghobj, arg1, fd, *osr);
3707 if (fd != STDIN_FILENO)
3708 close(fd);
3709 goto out;
3710 } else if (objcmd == "rm-omap") {
3711 if (vm.count("arg1") == 0) {
3712 usage(desc);
3713 ret = 1;
3714 goto out;
3715 }
3716 ret = do_rm_omap(fs, coll, ghobj, arg1, *osr);
3717 goto out;
3718 } else if (objcmd == "get-omaphdr") {
3719 if (vm.count("arg1")) {
3720 usage(desc);
3721 ret = 1;
3722 goto out;
3723 }
3724 ret = do_get_omaphdr(fs, coll, ghobj);
3725 goto out;
3726 } else if (objcmd == "set-omaphdr") {
3727 // Extra arg
3728 if (vm.count("arg2")) {
3729 usage(desc);
3730 ret = 1;
3731 goto out;
3732 }
3733 int fd;
3734 if (vm.count("arg1") == 0 || arg1 == "-") {
3735 // Since read_fd() doesn't handle ^D from a tty stdin, don't allow it.
3736 if (isatty(STDIN_FILENO)) {
3737 cerr << "stdin is a tty and no file specified" << std::endl;
3738 ret = 1;
3739 goto out;
3740 }
3741 fd = STDIN_FILENO;
3742 } else {
3743 fd = open(arg1.c_str(), O_RDONLY|O_LARGEFILE, 0666);
3744 if (fd == -1) {
3745 cerr << "open " << arg1 << " " << cpp_strerror(errno) << std::endl;
3746 ret = 1;
3747 goto out;
3748 }
3749 }
3750 ret = do_set_omaphdr(fs, coll, ghobj, fd, *osr);
3751 if (fd != STDIN_FILENO)
3752 close(fd);
3753 goto out;
3754 } else if (objcmd == "dump") {
3755 // There should not be any other arguments
3756 if (vm.count("arg1") || vm.count("arg2")) {
3757 usage(desc);
3758 ret = 1;
3759 goto out;
3760 }
3761 ret = print_obj_info(fs, coll, ghobj, formatter);
3762 goto out;
3763 } else if (objcmd == "set-size" || objcmd == "corrupt-size") {
3764 // Undocumented testing feature
3765 bool corrupt = (objcmd == "corrupt-size");
3766 // Extra arg
3767 if (vm.count("arg1") == 0 || vm.count("arg2")) {
3768 usage(desc);
3769 ret = 1;
3770 goto out;
3771 }
3772 if (arg1.length() == 0 || !isdigit(arg1.c_str()[0])) {
3773 cerr << "Invalid size '" << arg1 << "' specified" << std::endl;
3774 ret = 1;
3775 goto out;
3776 }
3777 uint64_t size = atoll(arg1.c_str());
3778 ret = set_size(fs, coll, ghobj, size, formatter, *osr, corrupt);
3779 goto out;
3780 } else if (objcmd == "clear-snapset") {
3781 // UNDOCUMENTED: For testing zap SnapSet
3782 // IGNORE extra args since not in usage anyway
3783 if (!ghobj.hobj.has_snapset()) {
3784 cerr << "'" << objcmd << "' requires a head or snapdir object" << std::endl;
3785 ret = 1;
3786 goto out;
3787 }
3788 ret = clear_snapset(fs, coll, ghobj, arg1, *osr);
3789 goto out;
3790 } else if (objcmd == "remove-clone-metadata") {
3791 // Extra arg
3792 if (vm.count("arg1") == 0 || vm.count("arg2")) {
3793 usage(desc);
3794 ret = 1;
3795 goto out;
3796 }
3797 if (!ghobj.hobj.has_snapset()) {
3798 cerr << "'" << objcmd << "' requires a head or snapdir object" << std::endl;
3799 ret = 1;
3800 goto out;
3801 }
3802 if (arg1.length() == 0 || !isdigit(arg1.c_str()[0])) {
3803 cerr << "Invalid cloneid '" << arg1 << "' specified" << std::endl;
3804 ret = 1;
3805 goto out;
3806 }
3807 snapid_t cloneid = atoi(arg1.c_str());
3808 ret = remove_clone(fs, coll, ghobj, cloneid, force, *osr);
3809 goto out;
3810 }
3811 cerr << "Unknown object command '" << objcmd << "'" << std::endl;
3812 usage(desc);
3813 ret = 1;
3814 goto out;
3815 }
3816
3817 bufferlist bl;
3818 map_epoch = 0;
3819 ret = PG::peek_map_epoch(fs, pgid, &map_epoch, &bl);
3820 if (ret < 0)
3821 cerr << "peek_map_epoch reports error" << std::endl;
3822 if (debug)
3823 cerr << "map_epoch " << map_epoch << std::endl;
3824
3825 pg_info_t info(pgid);
3826 PastIntervals past_intervals;
3827 __u8 struct_ver;
3828 ret = PG::read_info(fs, pgid, coll, bl, info, past_intervals,
3829 struct_ver);
3830 if (ret < 0) {
3831 cerr << "read_info error " << cpp_strerror(ret) << std::endl;
3832 goto out;
3833 }
3834 if (struct_ver < PG::compat_struct_v) {
3835 cerr << "PG is too old to upgrade, use older Ceph version" << std::endl;
3836 ret = -EFAULT;
3837 goto out;
3838 }
3839 if (debug)
3840 cerr << "struct_v " << (int)struct_ver << std::endl;
3841
3842 if (op == "export" || op == "export-remove") {
3843 ret = tool.do_export(fs, coll, pgid, info, map_epoch, struct_ver, superblock, past_intervals);
3844 if (ret == 0) {
3845 cerr << "Export successful" << std::endl;
3846 if (op == "export-remove") {
3847 ret = initiate_new_remove_pg(fs, pgid, *osr);
3848 // Export succeeded, so pgid is there
3849 assert(ret == 0);
3850 cerr << "Remove successful" << std::endl;
3851 }
3852 }
3853 } else if (op == "info") {
3854 formatter->open_object_section("info");
3855 info.dump(formatter);
3856 formatter->close_section();
3857 formatter->flush(cout);
3858 cout << std::endl;
3859 } else if (op == "log") {
3860 PGLog::IndexedLog log;
3861 pg_missing_t missing;
3862 ret = get_log(fs, struct_ver, coll, pgid, info, log, missing);
3863 if (ret < 0)
3864 goto out;
3865
3866 dump_log(formatter, cout, log, missing);
3867 } else if (op == "rm-past-intervals") {
3868 ObjectStore::Transaction tran;
3869 ObjectStore::Transaction *t = &tran;
3870
3871 if (struct_ver < PG::compat_struct_v) {
3872 cerr << "Can't remove past-intervals, version mismatch " << (int)struct_ver
3873 << " (pg) < compat " << (int)PG::compat_struct_v << " (tool)"
3874 << std::endl;
3875 ret = -EFAULT;
3876 goto out;
3877 }
3878
3879 cout << "Remove past-intervals " << past_intervals << std::endl;
3880
3881 past_intervals.clear();
3882 if (dry_run) {
3883 ret = 0;
3884 goto out;
3885 }
3886 ret = write_info(*t, map_epoch, info, past_intervals);
3887
3888 if (ret == 0) {
3889 fs->apply_transaction(osr, std::move(*t));
3890 cout << "Removal succeeded" << std::endl;
3891 }
3892 } else if (op == "mark-complete") {
3893 ObjectStore::Transaction tran;
3894 ObjectStore::Transaction *t = &tran;
3895
3896 if (struct_ver < PG::compat_struct_v) {
3897 cerr << "Can't mark-complete, version mismatch " << (int)struct_ver
3898 << " (pg) < compat " << (int)PG::compat_struct_v << " (tool)"
3899 << std::endl;
3900 ret = 1;
3901 goto out;
3902 }
3903
3904 cout << "Marking complete " << std::endl;
3905
3906 info.last_update = eversion_t(superblock.current_epoch, info.last_update.version + 1);
3907 info.last_backfill = hobject_t::get_max();
3908 info.last_epoch_started = superblock.current_epoch;
3909 info.history.last_epoch_started = superblock.current_epoch;
3910 info.history.last_epoch_clean = superblock.current_epoch;
3911 past_intervals.clear();
3912
3913 if (!dry_run) {
3914 ret = write_info(*t, map_epoch, info, past_intervals);
3915 if (ret != 0)
3916 goto out;
3917 fs->apply_transaction(osr, std::move(*t));
3918 }
3919 cout << "Marking complete succeeded" << std::endl;
3920 } else {
3921 assert(!"Should have already checked for valid --op");
3922 }
3923 } else {
3924 cerr << "PG '" << pgid << "' not found" << std::endl;
3925 ret = -ENOENT;
3926 }
3927
3928 out:
3929 int r = fs->umount();
3930 delete osr;
3931 if (r < 0) {
3932 cerr << "umount failed: " << cpp_strerror(r) << std::endl;
3933 // If no previous error, then use umount() error
3934 if (ret == 0)
3935 ret = r;
3936 }
3937
3938 if (dry_run) {
3939 // Export output can go to stdout, so put this message on stderr
3940 if (op == "export")
3941 cerr << "dry-run: Nothing changed" << std::endl;
3942 else
3943 cout << "dry-run: Nothing changed" << std::endl;
3944 }
3945
3946 if (ret < 0)
3947 ret = 1;
3948 return ret;
3949 }